| 1 | /*------------------------------------------------------------------------- |
| 2 | * |
| 3 | * twophase.c |
| 4 | * Two-phase commit support functions. |
| 5 | * |
| 6 | * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group |
| 7 | * Portions Copyright (c) 1994, Regents of the University of California |
| 8 | * |
| 9 | * IDENTIFICATION |
| 10 | * src/backend/access/transam/twophase.c |
| 11 | * |
| 12 | * NOTES |
| 13 | * Each global transaction is associated with a global transaction |
| 14 | * identifier (GID). The client assigns a GID to a postgres |
| 15 | * transaction with the PREPARE TRANSACTION command. |
| 16 | * |
| 17 | * We keep all active global transactions in a shared memory array. |
| 18 | * When the PREPARE TRANSACTION command is issued, the GID is |
| 19 | * reserved for the transaction in the array. This is done before |
| 20 | * a WAL entry is made, because the reservation checks for duplicate |
| 21 | * GIDs and aborts the transaction if there already is a global |
| 22 | * transaction in prepared state with the same GID. |
| 23 | * |
| 24 | * A global transaction (gxact) also has dummy PGXACT and PGPROC; this is |
| 25 | * what keeps the XID considered running by TransactionIdIsInProgress. |
| 26 | * It is also convenient as a PGPROC to hook the gxact's locks to. |
| 27 | * |
| 28 | * Information to recover prepared transactions in case of crash is |
| 29 | * now stored in WAL for the common case. In some cases there will be |
| 30 | * an extended period between preparing a GXACT and commit/abort, in |
| 31 | * which case we need to separately record prepared transaction data |
| 32 | * in permanent storage. This includes locking information, pending |
| 33 | * notifications etc. All that state information is written to the |
| 34 | * per-transaction state file in the pg_twophase directory. |
| 35 | * All prepared transactions will be written prior to shutdown. |
| 36 | * |
| 37 | * Life track of state data is following: |
| 38 | * |
| 39 | * * On PREPARE TRANSACTION backend writes state data only to the WAL and |
| 40 | * stores pointer to the start of the WAL record in |
| 41 | * gxact->prepare_start_lsn. |
| 42 | * * If COMMIT occurs before checkpoint then backend reads data from WAL |
| 43 | * using prepare_start_lsn. |
| 44 | * * On checkpoint state data copied to files in pg_twophase directory and |
| 45 | * fsynced |
| 46 | * * If COMMIT happens after checkpoint then backend reads state data from |
| 47 | * files |
| 48 | * |
| 49 | * During replay and replication, TwoPhaseState also holds information |
| 50 | * about active prepared transactions that haven't been moved to disk yet. |
| 51 | * |
| 52 | * Replay of twophase records happens by the following rules: |
| 53 | * |
| 54 | * * At the beginning of recovery, pg_twophase is scanned once, filling |
| 55 | * TwoPhaseState with entries marked with gxact->inredo and |
| 56 | * gxact->ondisk. Two-phase file data older than the XID horizon of |
| 57 | * the redo position are discarded. |
| 58 | * * On PREPARE redo, the transaction is added to TwoPhaseState->prepXacts. |
| 59 | * gxact->inredo is set to true for such entries. |
| 60 | * * On Checkpoint we iterate through TwoPhaseState->prepXacts entries |
| 61 | * that have gxact->inredo set and are behind the redo_horizon. We |
| 62 | * save them to disk and then switch gxact->ondisk to true. |
| 63 | * * On COMMIT/ABORT we delete the entry from TwoPhaseState->prepXacts. |
| 64 | * If gxact->ondisk is true, the corresponding entry from the disk |
| 65 | * is additionally deleted. |
| 66 | * * RecoverPreparedTransactions(), StandbyRecoverPreparedTransactions() |
| 67 | * and PrescanPreparedTransactions() have been modified to go through |
| 68 | * gxact->inredo entries that have not made it to disk. |
| 69 | * |
| 70 | *------------------------------------------------------------------------- |
| 71 | */ |
| 72 | #include "postgres.h" |
| 73 | |
| 74 | #include <fcntl.h> |
| 75 | #include <sys/stat.h> |
| 76 | #include <time.h> |
| 77 | #include <unistd.h> |
| 78 | |
| 79 | #include "access/commit_ts.h" |
| 80 | #include "access/htup_details.h" |
| 81 | #include "access/subtrans.h" |
| 82 | #include "access/transam.h" |
| 83 | #include "access/twophase.h" |
| 84 | #include "access/twophase_rmgr.h" |
| 85 | #include "access/xact.h" |
| 86 | #include "access/xlog.h" |
| 87 | #include "access/xloginsert.h" |
| 88 | #include "access/xlogutils.h" |
| 89 | #include "access/xlogreader.h" |
| 90 | #include "catalog/pg_type.h" |
| 91 | #include "catalog/storage.h" |
| 92 | #include "funcapi.h" |
| 93 | #include "miscadmin.h" |
| 94 | #include "pg_trace.h" |
| 95 | #include "pgstat.h" |
| 96 | #include "replication/origin.h" |
| 97 | #include "replication/syncrep.h" |
| 98 | #include "replication/walsender.h" |
| 99 | #include "storage/fd.h" |
| 100 | #include "storage/ipc.h" |
| 101 | #include "storage/md.h" |
| 102 | #include "storage/predicate.h" |
| 103 | #include "storage/proc.h" |
| 104 | #include "storage/procarray.h" |
| 105 | #include "storage/sinvaladt.h" |
| 106 | #include "storage/smgr.h" |
| 107 | #include "utils/builtins.h" |
| 108 | #include "utils/memutils.h" |
| 109 | #include "utils/timestamp.h" |
| 110 | |
| 111 | |
| 112 | /* |
| 113 | * Directory where Two-phase commit files reside within PGDATA |
| 114 | */ |
| 115 | #define TWOPHASE_DIR "pg_twophase" |
| 116 | |
| 117 | /* GUC variable, can't be changed after startup */ |
| 118 | int max_prepared_xacts = 0; |
| 119 | |
| 120 | /* |
| 121 | * This struct describes one global transaction that is in prepared state |
| 122 | * or attempting to become prepared. |
| 123 | * |
| 124 | * The lifecycle of a global transaction is: |
| 125 | * |
| 126 | * 1. After checking that the requested GID is not in use, set up an entry in |
| 127 | * the TwoPhaseState->prepXacts array with the correct GID and valid = false, |
| 128 | * and mark it as locked by my backend. |
| 129 | * |
| 130 | * 2. After successfully completing prepare, set valid = true and enter the |
| 131 | * referenced PGPROC into the global ProcArray. |
| 132 | * |
| 133 | * 3. To begin COMMIT PREPARED or ROLLBACK PREPARED, check that the entry is |
| 134 | * valid and not locked, then mark the entry as locked by storing my current |
| 135 | * backend ID into locking_backend. This prevents concurrent attempts to |
| 136 | * commit or rollback the same prepared xact. |
| 137 | * |
| 138 | * 4. On completion of COMMIT PREPARED or ROLLBACK PREPARED, remove the entry |
| 139 | * from the ProcArray and the TwoPhaseState->prepXacts array and return it to |
| 140 | * the freelist. |
| 141 | * |
| 142 | * Note that if the preparing transaction fails between steps 1 and 2, the |
| 143 | * entry must be removed so that the GID and the GlobalTransaction struct |
| 144 | * can be reused. See AtAbort_Twophase(). |
| 145 | * |
| 146 | * typedef struct GlobalTransactionData *GlobalTransaction appears in |
| 147 | * twophase.h |
| 148 | */ |
| 149 | |
| 150 | typedef struct GlobalTransactionData |
| 151 | { |
| 152 | GlobalTransaction next; /* list link for free list */ |
| 153 | int pgprocno; /* ID of associated dummy PGPROC */ |
| 154 | BackendId dummyBackendId; /* similar to backend id for backends */ |
| 155 | TimestampTz prepared_at; /* time of preparation */ |
| 156 | |
| 157 | /* |
| 158 | * Note that we need to keep track of two LSNs for each GXACT. We keep |
| 159 | * track of the start LSN because this is the address we must use to read |
| 160 | * state data back from WAL when committing a prepared GXACT. We keep |
| 161 | * track of the end LSN because that is the LSN we need to wait for prior |
| 162 | * to commit. |
| 163 | */ |
| 164 | XLogRecPtr prepare_start_lsn; /* XLOG offset of prepare record start */ |
| 165 | XLogRecPtr prepare_end_lsn; /* XLOG offset of prepare record end */ |
| 166 | TransactionId xid; /* The GXACT id */ |
| 167 | |
| 168 | Oid owner; /* ID of user that executed the xact */ |
| 169 | BackendId locking_backend; /* backend currently working on the xact */ |
| 170 | bool valid; /* true if PGPROC entry is in proc array */ |
| 171 | bool ondisk; /* true if prepare state file is on disk */ |
| 172 | bool inredo; /* true if entry was added via xlog_redo */ |
| 173 | char gid[GIDSIZE]; /* The GID assigned to the prepared xact */ |
| 174 | } GlobalTransactionData; |
| 175 | |
| 176 | /* |
| 177 | * Two Phase Commit shared state. Access to this struct is protected |
| 178 | * by TwoPhaseStateLock. |
| 179 | */ |
| 180 | typedef struct TwoPhaseStateData |
| 181 | { |
| 182 | /* Head of linked list of free GlobalTransactionData structs */ |
| 183 | GlobalTransaction freeGXacts; |
| 184 | |
| 185 | /* Number of valid prepXacts entries. */ |
| 186 | int numPrepXacts; |
| 187 | |
| 188 | /* There are max_prepared_xacts items in this array */ |
| 189 | GlobalTransaction prepXacts[FLEXIBLE_ARRAY_MEMBER]; |
| 190 | } TwoPhaseStateData; |
| 191 | |
| 192 | static TwoPhaseStateData *TwoPhaseState; |
| 193 | |
| 194 | /* |
| 195 | * Global transaction entry currently locked by us, if any. Note that any |
| 196 | * access to the entry pointed to by this variable must be protected by |
| 197 | * TwoPhaseStateLock, though obviously the pointer itself doesn't need to be |
| 198 | * (since it's just local memory). |
| 199 | */ |
| 200 | static GlobalTransaction MyLockedGxact = NULL; |
| 201 | |
| 202 | static bool twophaseExitRegistered = false; |
| 203 | |
| 204 | static void RecordTransactionCommitPrepared(TransactionId xid, |
| 205 | int nchildren, |
| 206 | TransactionId *children, |
| 207 | int nrels, |
| 208 | RelFileNode *rels, |
| 209 | int ninvalmsgs, |
| 210 | SharedInvalidationMessage *invalmsgs, |
| 211 | bool initfileinval, |
| 212 | const char *gid); |
| 213 | static void RecordTransactionAbortPrepared(TransactionId xid, |
| 214 | int nchildren, |
| 215 | TransactionId *children, |
| 216 | int nrels, |
| 217 | RelFileNode *rels, |
| 218 | const char *gid); |
| 219 | static void ProcessRecords(char *bufptr, TransactionId xid, |
| 220 | const TwoPhaseCallback callbacks[]); |
| 221 | static void RemoveGXact(GlobalTransaction gxact); |
| 222 | |
| 223 | static void XlogReadTwoPhaseData(XLogRecPtr lsn, char **buf, int *len); |
| 224 | static char *ProcessTwoPhaseBuffer(TransactionId xid, |
| 225 | XLogRecPtr prepare_start_lsn, |
| 226 | bool fromdisk, bool setParent, bool setNextXid); |
| 227 | static void MarkAsPreparingGuts(GlobalTransaction gxact, TransactionId xid, |
| 228 | const char *gid, TimestampTz prepared_at, Oid owner, |
| 229 | Oid databaseid); |
| 230 | static void RemoveTwoPhaseFile(TransactionId xid, bool giveWarning); |
| 231 | static void RecreateTwoPhaseFile(TransactionId xid, void *content, int len); |
| 232 | |
| 233 | /* |
| 234 | * Initialization of shared memory |
| 235 | */ |
| 236 | Size |
| 237 | TwoPhaseShmemSize(void) |
| 238 | { |
| 239 | Size size; |
| 240 | |
| 241 | /* Need the fixed struct, the array of pointers, and the GTD structs */ |
| 242 | size = offsetof(TwoPhaseStateData, prepXacts); |
| 243 | size = add_size(size, mul_size(max_prepared_xacts, |
| 244 | sizeof(GlobalTransaction))); |
| 245 | size = MAXALIGN(size); |
| 246 | size = add_size(size, mul_size(max_prepared_xacts, |
| 247 | sizeof(GlobalTransactionData))); |
| 248 | |
| 249 | return size; |
| 250 | } |
| 251 | |
| 252 | void |
| 253 | TwoPhaseShmemInit(void) |
| 254 | { |
| 255 | bool found; |
| 256 | |
| 257 | TwoPhaseState = ShmemInitStruct("Prepared Transaction Table" , |
| 258 | TwoPhaseShmemSize(), |
| 259 | &found); |
| 260 | if (!IsUnderPostmaster) |
| 261 | { |
| 262 | GlobalTransaction gxacts; |
| 263 | int i; |
| 264 | |
| 265 | Assert(!found); |
| 266 | TwoPhaseState->freeGXacts = NULL; |
| 267 | TwoPhaseState->numPrepXacts = 0; |
| 268 | |
| 269 | /* |
| 270 | * Initialize the linked list of free GlobalTransactionData structs |
| 271 | */ |
| 272 | gxacts = (GlobalTransaction) |
| 273 | ((char *) TwoPhaseState + |
| 274 | MAXALIGN(offsetof(TwoPhaseStateData, prepXacts) + |
| 275 | sizeof(GlobalTransaction) * max_prepared_xacts)); |
| 276 | for (i = 0; i < max_prepared_xacts; i++) |
| 277 | { |
| 278 | /* insert into linked list */ |
| 279 | gxacts[i].next = TwoPhaseState->freeGXacts; |
| 280 | TwoPhaseState->freeGXacts = &gxacts[i]; |
| 281 | |
| 282 | /* associate it with a PGPROC assigned by InitProcGlobal */ |
| 283 | gxacts[i].pgprocno = PreparedXactProcs[i].pgprocno; |
| 284 | |
| 285 | /* |
| 286 | * Assign a unique ID for each dummy proc, so that the range of |
| 287 | * dummy backend IDs immediately follows the range of normal |
| 288 | * backend IDs. We don't dare to assign a real backend ID to dummy |
| 289 | * procs, because prepared transactions don't take part in cache |
| 290 | * invalidation like a real backend ID would imply, but having a |
| 291 | * unique ID for them is nevertheless handy. This arrangement |
| 292 | * allows you to allocate an array of size (MaxBackends + |
| 293 | * max_prepared_xacts + 1), and have a slot for every backend and |
| 294 | * prepared transaction. Currently multixact.c uses that |
| 295 | * technique. |
| 296 | */ |
| 297 | gxacts[i].dummyBackendId = MaxBackends + 1 + i; |
| 298 | } |
| 299 | } |
| 300 | else |
| 301 | Assert(found); |
| 302 | } |
| 303 | |
| 304 | /* |
| 305 | * Exit hook to unlock the global transaction entry we're working on. |
| 306 | */ |
| 307 | static void |
| 308 | AtProcExit_Twophase(int code, Datum arg) |
| 309 | { |
| 310 | /* same logic as abort */ |
| 311 | AtAbort_Twophase(); |
| 312 | } |
| 313 | |
| 314 | /* |
| 315 | * Abort hook to unlock the global transaction entry we're working on. |
| 316 | */ |
| 317 | void |
| 318 | AtAbort_Twophase(void) |
| 319 | { |
| 320 | if (MyLockedGxact == NULL) |
| 321 | return; |
| 322 | |
| 323 | /* |
| 324 | * What to do with the locked global transaction entry? If we were in the |
| 325 | * process of preparing the transaction, but haven't written the WAL |
| 326 | * record and state file yet, the transaction must not be considered as |
| 327 | * prepared. Likewise, if we are in the process of finishing an |
| 328 | * already-prepared transaction, and fail after having already written the |
| 329 | * 2nd phase commit or rollback record to the WAL, the transaction should |
| 330 | * not be considered as prepared anymore. In those cases, just remove the |
| 331 | * entry from shared memory. |
| 332 | * |
| 333 | * Otherwise, the entry must be left in place so that the transaction can |
| 334 | * be finished later, so just unlock it. |
| 335 | * |
| 336 | * If we abort during prepare, after having written the WAL record, we |
| 337 | * might not have transferred all locks and other state to the prepared |
| 338 | * transaction yet. Likewise, if we abort during commit or rollback, |
| 339 | * after having written the WAL record, we might not have released all the |
| 340 | * resources held by the transaction yet. In those cases, the in-memory |
| 341 | * state can be wrong, but it's too late to back out. |
| 342 | */ |
| 343 | LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE); |
| 344 | if (!MyLockedGxact->valid) |
| 345 | RemoveGXact(MyLockedGxact); |
| 346 | else |
| 347 | MyLockedGxact->locking_backend = InvalidBackendId; |
| 348 | LWLockRelease(TwoPhaseStateLock); |
| 349 | |
| 350 | MyLockedGxact = NULL; |
| 351 | } |
| 352 | |
| 353 | /* |
| 354 | * This is called after we have finished transferring state to the prepared |
| 355 | * PGXACT entry. |
| 356 | */ |
| 357 | void |
| 358 | PostPrepare_Twophase(void) |
| 359 | { |
| 360 | LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE); |
| 361 | MyLockedGxact->locking_backend = InvalidBackendId; |
| 362 | LWLockRelease(TwoPhaseStateLock); |
| 363 | |
| 364 | MyLockedGxact = NULL; |
| 365 | } |
| 366 | |
| 367 | |
| 368 | /* |
| 369 | * MarkAsPreparing |
| 370 | * Reserve the GID for the given transaction. |
| 371 | */ |
| 372 | GlobalTransaction |
| 373 | MarkAsPreparing(TransactionId xid, const char *gid, |
| 374 | TimestampTz prepared_at, Oid owner, Oid databaseid) |
| 375 | { |
| 376 | GlobalTransaction gxact; |
| 377 | int i; |
| 378 | |
| 379 | if (strlen(gid) >= GIDSIZE) |
| 380 | ereport(ERROR, |
| 381 | (errcode(ERRCODE_INVALID_PARAMETER_VALUE), |
| 382 | errmsg("transaction identifier \"%s\" is too long" , |
| 383 | gid))); |
| 384 | |
| 385 | /* fail immediately if feature is disabled */ |
| 386 | if (max_prepared_xacts == 0) |
| 387 | ereport(ERROR, |
| 388 | (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE), |
| 389 | errmsg("prepared transactions are disabled" ), |
| 390 | errhint("Set max_prepared_transactions to a nonzero value." ))); |
| 391 | |
| 392 | /* on first call, register the exit hook */ |
| 393 | if (!twophaseExitRegistered) |
| 394 | { |
| 395 | before_shmem_exit(AtProcExit_Twophase, 0); |
| 396 | twophaseExitRegistered = true; |
| 397 | } |
| 398 | |
| 399 | LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE); |
| 400 | |
| 401 | /* Check for conflicting GID */ |
| 402 | for (i = 0; i < TwoPhaseState->numPrepXacts; i++) |
| 403 | { |
| 404 | gxact = TwoPhaseState->prepXacts[i]; |
| 405 | if (strcmp(gxact->gid, gid) == 0) |
| 406 | { |
| 407 | ereport(ERROR, |
| 408 | (errcode(ERRCODE_DUPLICATE_OBJECT), |
| 409 | errmsg("transaction identifier \"%s\" is already in use" , |
| 410 | gid))); |
| 411 | } |
| 412 | } |
| 413 | |
| 414 | /* Get a free gxact from the freelist */ |
| 415 | if (TwoPhaseState->freeGXacts == NULL) |
| 416 | ereport(ERROR, |
| 417 | (errcode(ERRCODE_OUT_OF_MEMORY), |
| 418 | errmsg("maximum number of prepared transactions reached" ), |
| 419 | errhint("Increase max_prepared_transactions (currently %d)." , |
| 420 | max_prepared_xacts))); |
| 421 | gxact = TwoPhaseState->freeGXacts; |
| 422 | TwoPhaseState->freeGXacts = gxact->next; |
| 423 | |
| 424 | MarkAsPreparingGuts(gxact, xid, gid, prepared_at, owner, databaseid); |
| 425 | |
| 426 | gxact->ondisk = false; |
| 427 | |
| 428 | /* And insert it into the active array */ |
| 429 | Assert(TwoPhaseState->numPrepXacts < max_prepared_xacts); |
| 430 | TwoPhaseState->prepXacts[TwoPhaseState->numPrepXacts++] = gxact; |
| 431 | |
| 432 | LWLockRelease(TwoPhaseStateLock); |
| 433 | |
| 434 | return gxact; |
| 435 | } |
| 436 | |
| 437 | /* |
| 438 | * MarkAsPreparingGuts |
| 439 | * |
| 440 | * This uses a gxact struct and puts it into the active array. |
| 441 | * NOTE: this is also used when reloading a gxact after a crash; so avoid |
| 442 | * assuming that we can use very much backend context. |
| 443 | * |
| 444 | * Note: This function should be called with appropriate locks held. |
| 445 | */ |
| 446 | static void |
| 447 | MarkAsPreparingGuts(GlobalTransaction gxact, TransactionId xid, const char *gid, |
| 448 | TimestampTz prepared_at, Oid owner, Oid databaseid) |
| 449 | { |
| 450 | PGPROC *proc; |
| 451 | PGXACT *pgxact; |
| 452 | int i; |
| 453 | |
| 454 | Assert(LWLockHeldByMeInMode(TwoPhaseStateLock, LW_EXCLUSIVE)); |
| 455 | |
| 456 | Assert(gxact != NULL); |
| 457 | proc = &ProcGlobal->allProcs[gxact->pgprocno]; |
| 458 | pgxact = &ProcGlobal->allPgXact[gxact->pgprocno]; |
| 459 | |
| 460 | /* Initialize the PGPROC entry */ |
| 461 | MemSet(proc, 0, sizeof(PGPROC)); |
| 462 | proc->pgprocno = gxact->pgprocno; |
| 463 | SHMQueueElemInit(&(proc->links)); |
| 464 | proc->waitStatus = STATUS_OK; |
| 465 | /* We set up the gxact's VXID as InvalidBackendId/XID */ |
| 466 | proc->lxid = (LocalTransactionId) xid; |
| 467 | pgxact->xid = xid; |
| 468 | pgxact->xmin = InvalidTransactionId; |
| 469 | pgxact->delayChkpt = false; |
| 470 | pgxact->vacuumFlags = 0; |
| 471 | proc->pid = 0; |
| 472 | proc->backendId = InvalidBackendId; |
| 473 | proc->databaseId = databaseid; |
| 474 | proc->roleId = owner; |
| 475 | proc->tempNamespaceId = InvalidOid; |
| 476 | proc->isBackgroundWorker = false; |
| 477 | proc->lwWaiting = false; |
| 478 | proc->lwWaitMode = 0; |
| 479 | proc->waitLock = NULL; |
| 480 | proc->waitProcLock = NULL; |
| 481 | for (i = 0; i < NUM_LOCK_PARTITIONS; i++) |
| 482 | SHMQueueInit(&(proc->myProcLocks[i])); |
| 483 | /* subxid data must be filled later by GXactLoadSubxactData */ |
| 484 | pgxact->overflowed = false; |
| 485 | pgxact->nxids = 0; |
| 486 | |
| 487 | gxact->prepared_at = prepared_at; |
| 488 | gxact->xid = xid; |
| 489 | gxact->owner = owner; |
| 490 | gxact->locking_backend = MyBackendId; |
| 491 | gxact->valid = false; |
| 492 | gxact->inredo = false; |
| 493 | strcpy(gxact->gid, gid); |
| 494 | |
| 495 | /* |
| 496 | * Remember that we have this GlobalTransaction entry locked for us. If we |
| 497 | * abort after this, we must release it. |
| 498 | */ |
| 499 | MyLockedGxact = gxact; |
| 500 | } |
| 501 | |
| 502 | /* |
| 503 | * GXactLoadSubxactData |
| 504 | * |
| 505 | * If the transaction being persisted had any subtransactions, this must |
| 506 | * be called before MarkAsPrepared() to load information into the dummy |
| 507 | * PGPROC. |
| 508 | */ |
| 509 | static void |
| 510 | GXactLoadSubxactData(GlobalTransaction gxact, int nsubxacts, |
| 511 | TransactionId *children) |
| 512 | { |
| 513 | PGPROC *proc = &ProcGlobal->allProcs[gxact->pgprocno]; |
| 514 | PGXACT *pgxact = &ProcGlobal->allPgXact[gxact->pgprocno]; |
| 515 | |
| 516 | /* We need no extra lock since the GXACT isn't valid yet */ |
| 517 | if (nsubxacts > PGPROC_MAX_CACHED_SUBXIDS) |
| 518 | { |
| 519 | pgxact->overflowed = true; |
| 520 | nsubxacts = PGPROC_MAX_CACHED_SUBXIDS; |
| 521 | } |
| 522 | if (nsubxacts > 0) |
| 523 | { |
| 524 | memcpy(proc->subxids.xids, children, |
| 525 | nsubxacts * sizeof(TransactionId)); |
| 526 | pgxact->nxids = nsubxacts; |
| 527 | } |
| 528 | } |
| 529 | |
| 530 | /* |
| 531 | * MarkAsPrepared |
| 532 | * Mark the GXACT as fully valid, and enter it into the global ProcArray. |
| 533 | * |
| 534 | * lock_held indicates whether caller already holds TwoPhaseStateLock. |
| 535 | */ |
| 536 | static void |
| 537 | MarkAsPrepared(GlobalTransaction gxact, bool lock_held) |
| 538 | { |
| 539 | /* Lock here may be overkill, but I'm not convinced of that ... */ |
| 540 | if (!lock_held) |
| 541 | LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE); |
| 542 | Assert(!gxact->valid); |
| 543 | gxact->valid = true; |
| 544 | if (!lock_held) |
| 545 | LWLockRelease(TwoPhaseStateLock); |
| 546 | |
| 547 | /* |
| 548 | * Put it into the global ProcArray so TransactionIdIsInProgress considers |
| 549 | * the XID as still running. |
| 550 | */ |
| 551 | ProcArrayAdd(&ProcGlobal->allProcs[gxact->pgprocno]); |
| 552 | } |
| 553 | |
| 554 | /* |
| 555 | * LockGXact |
| 556 | * Locate the prepared transaction and mark it busy for COMMIT or PREPARE. |
| 557 | */ |
| 558 | static GlobalTransaction |
| 559 | LockGXact(const char *gid, Oid user) |
| 560 | { |
| 561 | int i; |
| 562 | |
| 563 | /* on first call, register the exit hook */ |
| 564 | if (!twophaseExitRegistered) |
| 565 | { |
| 566 | before_shmem_exit(AtProcExit_Twophase, 0); |
| 567 | twophaseExitRegistered = true; |
| 568 | } |
| 569 | |
| 570 | LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE); |
| 571 | |
| 572 | for (i = 0; i < TwoPhaseState->numPrepXacts; i++) |
| 573 | { |
| 574 | GlobalTransaction gxact = TwoPhaseState->prepXacts[i]; |
| 575 | PGPROC *proc = &ProcGlobal->allProcs[gxact->pgprocno]; |
| 576 | |
| 577 | /* Ignore not-yet-valid GIDs */ |
| 578 | if (!gxact->valid) |
| 579 | continue; |
| 580 | if (strcmp(gxact->gid, gid) != 0) |
| 581 | continue; |
| 582 | |
| 583 | /* Found it, but has someone else got it locked? */ |
| 584 | if (gxact->locking_backend != InvalidBackendId) |
| 585 | ereport(ERROR, |
| 586 | (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE), |
| 587 | errmsg("prepared transaction with identifier \"%s\" is busy" , |
| 588 | gid))); |
| 589 | |
| 590 | if (user != gxact->owner && !superuser_arg(user)) |
| 591 | ereport(ERROR, |
| 592 | (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE), |
| 593 | errmsg("permission denied to finish prepared transaction" ), |
| 594 | errhint("Must be superuser or the user that prepared the transaction." ))); |
| 595 | |
| 596 | /* |
| 597 | * Note: it probably would be possible to allow committing from |
| 598 | * another database; but at the moment NOTIFY is known not to work and |
| 599 | * there may be some other issues as well. Hence disallow until |
| 600 | * someone gets motivated to make it work. |
| 601 | */ |
| 602 | if (MyDatabaseId != proc->databaseId) |
| 603 | ereport(ERROR, |
| 604 | (errcode(ERRCODE_FEATURE_NOT_SUPPORTED), |
| 605 | errmsg("prepared transaction belongs to another database" ), |
| 606 | errhint("Connect to the database where the transaction was prepared to finish it." ))); |
| 607 | |
| 608 | /* OK for me to lock it */ |
| 609 | gxact->locking_backend = MyBackendId; |
| 610 | MyLockedGxact = gxact; |
| 611 | |
| 612 | LWLockRelease(TwoPhaseStateLock); |
| 613 | |
| 614 | return gxact; |
| 615 | } |
| 616 | |
| 617 | LWLockRelease(TwoPhaseStateLock); |
| 618 | |
| 619 | ereport(ERROR, |
| 620 | (errcode(ERRCODE_UNDEFINED_OBJECT), |
| 621 | errmsg("prepared transaction with identifier \"%s\" does not exist" , |
| 622 | gid))); |
| 623 | |
| 624 | /* NOTREACHED */ |
| 625 | return NULL; |
| 626 | } |
| 627 | |
| 628 | /* |
| 629 | * RemoveGXact |
| 630 | * Remove the prepared transaction from the shared memory array. |
| 631 | * |
| 632 | * NB: caller should have already removed it from ProcArray |
| 633 | */ |
| 634 | static void |
| 635 | RemoveGXact(GlobalTransaction gxact) |
| 636 | { |
| 637 | int i; |
| 638 | |
| 639 | Assert(LWLockHeldByMeInMode(TwoPhaseStateLock, LW_EXCLUSIVE)); |
| 640 | |
| 641 | for (i = 0; i < TwoPhaseState->numPrepXacts; i++) |
| 642 | { |
| 643 | if (gxact == TwoPhaseState->prepXacts[i]) |
| 644 | { |
| 645 | /* remove from the active array */ |
| 646 | TwoPhaseState->numPrepXacts--; |
| 647 | TwoPhaseState->prepXacts[i] = TwoPhaseState->prepXacts[TwoPhaseState->numPrepXacts]; |
| 648 | |
| 649 | /* and put it back in the freelist */ |
| 650 | gxact->next = TwoPhaseState->freeGXacts; |
| 651 | TwoPhaseState->freeGXacts = gxact; |
| 652 | |
| 653 | return; |
| 654 | } |
| 655 | } |
| 656 | |
| 657 | elog(ERROR, "failed to find %p in GlobalTransaction array" , gxact); |
| 658 | } |
| 659 | |
| 660 | /* |
| 661 | * Returns an array of all prepared transactions for the user-level |
| 662 | * function pg_prepared_xact. |
| 663 | * |
| 664 | * The returned array and all its elements are copies of internal data |
| 665 | * structures, to minimize the time we need to hold the TwoPhaseStateLock. |
| 666 | * |
| 667 | * WARNING -- we return even those transactions that are not fully prepared |
| 668 | * yet. The caller should filter them out if he doesn't want them. |
| 669 | * |
| 670 | * The returned array is palloc'd. |
| 671 | */ |
| 672 | static int |
| 673 | GetPreparedTransactionList(GlobalTransaction *gxacts) |
| 674 | { |
| 675 | GlobalTransaction array; |
| 676 | int num; |
| 677 | int i; |
| 678 | |
| 679 | LWLockAcquire(TwoPhaseStateLock, LW_SHARED); |
| 680 | |
| 681 | if (TwoPhaseState->numPrepXacts == 0) |
| 682 | { |
| 683 | LWLockRelease(TwoPhaseStateLock); |
| 684 | |
| 685 | *gxacts = NULL; |
| 686 | return 0; |
| 687 | } |
| 688 | |
| 689 | num = TwoPhaseState->numPrepXacts; |
| 690 | array = (GlobalTransaction) palloc(sizeof(GlobalTransactionData) * num); |
| 691 | *gxacts = array; |
| 692 | for (i = 0; i < num; i++) |
| 693 | memcpy(array + i, TwoPhaseState->prepXacts[i], |
| 694 | sizeof(GlobalTransactionData)); |
| 695 | |
| 696 | LWLockRelease(TwoPhaseStateLock); |
| 697 | |
| 698 | return num; |
| 699 | } |
| 700 | |
| 701 | |
| 702 | /* Working status for pg_prepared_xact */ |
| 703 | typedef struct |
| 704 | { |
| 705 | GlobalTransaction array; |
| 706 | int ngxacts; |
| 707 | int currIdx; |
| 708 | } Working_State; |
| 709 | |
| 710 | /* |
| 711 | * pg_prepared_xact |
| 712 | * Produce a view with one row per prepared transaction. |
| 713 | * |
| 714 | * This function is here so we don't have to export the |
| 715 | * GlobalTransactionData struct definition. |
| 716 | */ |
| 717 | Datum |
| 718 | pg_prepared_xact(PG_FUNCTION_ARGS) |
| 719 | { |
| 720 | FuncCallContext *funcctx; |
| 721 | Working_State *status; |
| 722 | |
| 723 | if (SRF_IS_FIRSTCALL()) |
| 724 | { |
| 725 | TupleDesc tupdesc; |
| 726 | MemoryContext oldcontext; |
| 727 | |
| 728 | /* create a function context for cross-call persistence */ |
| 729 | funcctx = SRF_FIRSTCALL_INIT(); |
| 730 | |
| 731 | /* |
| 732 | * Switch to memory context appropriate for multiple function calls |
| 733 | */ |
| 734 | oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx); |
| 735 | |
| 736 | /* build tupdesc for result tuples */ |
| 737 | /* this had better match pg_prepared_xacts view in system_views.sql */ |
| 738 | tupdesc = CreateTemplateTupleDesc(5); |
| 739 | TupleDescInitEntry(tupdesc, (AttrNumber) 1, "transaction" , |
| 740 | XIDOID, -1, 0); |
| 741 | TupleDescInitEntry(tupdesc, (AttrNumber) 2, "gid" , |
| 742 | TEXTOID, -1, 0); |
| 743 | TupleDescInitEntry(tupdesc, (AttrNumber) 3, "prepared" , |
| 744 | TIMESTAMPTZOID, -1, 0); |
| 745 | TupleDescInitEntry(tupdesc, (AttrNumber) 4, "ownerid" , |
| 746 | OIDOID, -1, 0); |
| 747 | TupleDescInitEntry(tupdesc, (AttrNumber) 5, "dbid" , |
| 748 | OIDOID, -1, 0); |
| 749 | |
| 750 | funcctx->tuple_desc = BlessTupleDesc(tupdesc); |
| 751 | |
| 752 | /* |
| 753 | * Collect all the 2PC status information that we will format and send |
| 754 | * out as a result set. |
| 755 | */ |
| 756 | status = (Working_State *) palloc(sizeof(Working_State)); |
| 757 | funcctx->user_fctx = (void *) status; |
| 758 | |
| 759 | status->ngxacts = GetPreparedTransactionList(&status->array); |
| 760 | status->currIdx = 0; |
| 761 | |
| 762 | MemoryContextSwitchTo(oldcontext); |
| 763 | } |
| 764 | |
| 765 | funcctx = SRF_PERCALL_SETUP(); |
| 766 | status = (Working_State *) funcctx->user_fctx; |
| 767 | |
| 768 | while (status->array != NULL && status->currIdx < status->ngxacts) |
| 769 | { |
| 770 | GlobalTransaction gxact = &status->array[status->currIdx++]; |
| 771 | PGPROC *proc = &ProcGlobal->allProcs[gxact->pgprocno]; |
| 772 | PGXACT *pgxact = &ProcGlobal->allPgXact[gxact->pgprocno]; |
| 773 | Datum values[5]; |
| 774 | bool nulls[5]; |
| 775 | HeapTuple tuple; |
| 776 | Datum result; |
| 777 | |
| 778 | if (!gxact->valid) |
| 779 | continue; |
| 780 | |
| 781 | /* |
| 782 | * Form tuple with appropriate data. |
| 783 | */ |
| 784 | MemSet(values, 0, sizeof(values)); |
| 785 | MemSet(nulls, 0, sizeof(nulls)); |
| 786 | |
| 787 | values[0] = TransactionIdGetDatum(pgxact->xid); |
| 788 | values[1] = CStringGetTextDatum(gxact->gid); |
| 789 | values[2] = TimestampTzGetDatum(gxact->prepared_at); |
| 790 | values[3] = ObjectIdGetDatum(gxact->owner); |
| 791 | values[4] = ObjectIdGetDatum(proc->databaseId); |
| 792 | |
| 793 | tuple = heap_form_tuple(funcctx->tuple_desc, values, nulls); |
| 794 | result = HeapTupleGetDatum(tuple); |
| 795 | SRF_RETURN_NEXT(funcctx, result); |
| 796 | } |
| 797 | |
| 798 | SRF_RETURN_DONE(funcctx); |
| 799 | } |
| 800 | |
| 801 | /* |
| 802 | * TwoPhaseGetGXact |
| 803 | * Get the GlobalTransaction struct for a prepared transaction |
| 804 | * specified by XID |
| 805 | * |
| 806 | * If lock_held is set to true, TwoPhaseStateLock will not be taken, so the |
| 807 | * caller had better hold it. |
| 808 | */ |
| 809 | static GlobalTransaction |
| 810 | TwoPhaseGetGXact(TransactionId xid, bool lock_held) |
| 811 | { |
| 812 | GlobalTransaction result = NULL; |
| 813 | int i; |
| 814 | |
| 815 | static TransactionId cached_xid = InvalidTransactionId; |
| 816 | static GlobalTransaction cached_gxact = NULL; |
| 817 | |
| 818 | Assert(!lock_held || LWLockHeldByMe(TwoPhaseStateLock)); |
| 819 | |
| 820 | /* |
| 821 | * During a recovery, COMMIT PREPARED, or ABORT PREPARED, we'll be called |
| 822 | * repeatedly for the same XID. We can save work with a simple cache. |
| 823 | */ |
| 824 | if (xid == cached_xid) |
| 825 | return cached_gxact; |
| 826 | |
| 827 | if (!lock_held) |
| 828 | LWLockAcquire(TwoPhaseStateLock, LW_SHARED); |
| 829 | |
| 830 | for (i = 0; i < TwoPhaseState->numPrepXacts; i++) |
| 831 | { |
| 832 | GlobalTransaction gxact = TwoPhaseState->prepXacts[i]; |
| 833 | PGXACT *pgxact = &ProcGlobal->allPgXact[gxact->pgprocno]; |
| 834 | |
| 835 | if (pgxact->xid == xid) |
| 836 | { |
| 837 | result = gxact; |
| 838 | break; |
| 839 | } |
| 840 | } |
| 841 | |
| 842 | if (!lock_held) |
| 843 | LWLockRelease(TwoPhaseStateLock); |
| 844 | |
| 845 | if (result == NULL) /* should not happen */ |
| 846 | elog(ERROR, "failed to find GlobalTransaction for xid %u" , xid); |
| 847 | |
| 848 | cached_xid = xid; |
| 849 | cached_gxact = result; |
| 850 | |
| 851 | return result; |
| 852 | } |
| 853 | |
| 854 | /* |
| 855 | * TwoPhaseGetDummyBackendId |
| 856 | * Get the dummy backend ID for prepared transaction specified by XID |
| 857 | * |
| 858 | * Dummy backend IDs are similar to real backend IDs of real backends. |
| 859 | * They start at MaxBackends + 1, and are unique across all currently active |
| 860 | * real backends and prepared transactions. If lock_held is set to true, |
| 861 | * TwoPhaseStateLock will not be taken, so the caller had better hold it. |
| 862 | */ |
| 863 | BackendId |
| 864 | TwoPhaseGetDummyBackendId(TransactionId xid, bool lock_held) |
| 865 | { |
| 866 | GlobalTransaction gxact = TwoPhaseGetGXact(xid, lock_held); |
| 867 | |
| 868 | return gxact->dummyBackendId; |
| 869 | } |
| 870 | |
| 871 | /* |
| 872 | * TwoPhaseGetDummyProc |
| 873 | * Get the PGPROC that represents a prepared transaction specified by XID |
| 874 | * |
| 875 | * If lock_held is set to true, TwoPhaseStateLock will not be taken, so the |
| 876 | * caller had better hold it. |
| 877 | */ |
| 878 | PGPROC * |
| 879 | TwoPhaseGetDummyProc(TransactionId xid, bool lock_held) |
| 880 | { |
| 881 | GlobalTransaction gxact = TwoPhaseGetGXact(xid, lock_held); |
| 882 | |
| 883 | return &ProcGlobal->allProcs[gxact->pgprocno]; |
| 884 | } |
| 885 | |
| 886 | /************************************************************************/ |
| 887 | /* State file support */ |
| 888 | /************************************************************************/ |
| 889 | |
| 890 | #define TwoPhaseFilePath(path, xid) \ |
| 891 | snprintf(path, MAXPGPATH, TWOPHASE_DIR "/%08X", xid) |
| 892 | |
| 893 | /* |
| 894 | * 2PC state file format: |
| 895 | * |
| 896 | * 1. TwoPhaseFileHeader |
| 897 | * 2. TransactionId[] (subtransactions) |
| 898 | * 3. RelFileNode[] (files to be deleted at commit) |
| 899 | * 4. RelFileNode[] (files to be deleted at abort) |
| 900 | * 5. SharedInvalidationMessage[] (inval messages to be sent at commit) |
| 901 | * 6. TwoPhaseRecordOnDisk |
| 902 | * 7. ... |
| 903 | * 8. TwoPhaseRecordOnDisk (end sentinel, rmid == TWOPHASE_RM_END_ID) |
| 904 | * 9. checksum (CRC-32C) |
| 905 | * |
| 906 | * Each segment except the final checksum is MAXALIGN'd. |
| 907 | */ |
| 908 | |
| 909 | /* |
| 910 | * Header for a 2PC state file |
| 911 | */ |
| 912 | #define TWOPHASE_MAGIC 0x57F94534 /* format identifier */ |
| 913 | |
| 914 | typedef struct |
| 915 | { |
| 916 | uint32 ; /* format identifier */ |
| 917 | uint32 ; /* actual file length */ |
| 918 | TransactionId ; /* original transaction XID */ |
| 919 | Oid ; /* OID of database it was in */ |
| 920 | TimestampTz ; /* time of preparation */ |
| 921 | Oid ; /* user running the transaction */ |
| 922 | int32 ; /* number of following subxact XIDs */ |
| 923 | int32 ; /* number of delete-on-commit rels */ |
| 924 | int32 ; /* number of delete-on-abort rels */ |
| 925 | int32 ; /* number of cache invalidation messages */ |
| 926 | bool ; /* does relcache init file need invalidation? */ |
| 927 | uint16 ; /* length of the GID - GID follows the header */ |
| 928 | XLogRecPtr ; /* lsn of this record at origin node */ |
| 929 | TimestampTz ; /* time of prepare at origin node */ |
| 930 | } ; |
| 931 | |
| 932 | /* |
| 933 | * Header for each record in a state file |
| 934 | * |
| 935 | * NOTE: len counts only the rmgr data, not the TwoPhaseRecordOnDisk header. |
| 936 | * The rmgr data will be stored starting on a MAXALIGN boundary. |
| 937 | */ |
| 938 | typedef struct TwoPhaseRecordOnDisk |
| 939 | { |
| 940 | uint32 len; /* length of rmgr data */ |
| 941 | TwoPhaseRmgrId rmid; /* resource manager for this record */ |
| 942 | uint16 info; /* flag bits for use by rmgr */ |
| 943 | } TwoPhaseRecordOnDisk; |
| 944 | |
| 945 | /* |
| 946 | * During prepare, the state file is assembled in memory before writing it |
| 947 | * to WAL and the actual state file. We use a chain of StateFileChunk blocks |
| 948 | * for that. |
| 949 | */ |
| 950 | typedef struct StateFileChunk |
| 951 | { |
| 952 | char *data; |
| 953 | uint32 len; |
| 954 | struct StateFileChunk *next; |
| 955 | } StateFileChunk; |
| 956 | |
| 957 | static struct xllist |
| 958 | { |
| 959 | StateFileChunk *head; /* first data block in the chain */ |
| 960 | StateFileChunk *tail; /* last block in chain */ |
| 961 | uint32 num_chunks; |
| 962 | uint32 bytes_free; /* free bytes left in tail block */ |
| 963 | uint32 total_len; /* total data bytes in chain */ |
| 964 | } records; |
| 965 | |
| 966 | |
| 967 | /* |
| 968 | * Append a block of data to records data structure. |
| 969 | * |
| 970 | * NB: each block is padded to a MAXALIGN multiple. This must be |
| 971 | * accounted for when the file is later read! |
| 972 | * |
| 973 | * The data is copied, so the caller is free to modify it afterwards. |
| 974 | */ |
| 975 | static void |
| 976 | save_state_data(const void *data, uint32 len) |
| 977 | { |
| 978 | uint32 padlen = MAXALIGN(len); |
| 979 | |
| 980 | if (padlen > records.bytes_free) |
| 981 | { |
| 982 | records.tail->next = palloc0(sizeof(StateFileChunk)); |
| 983 | records.tail = records.tail->next; |
| 984 | records.tail->len = 0; |
| 985 | records.tail->next = NULL; |
| 986 | records.num_chunks++; |
| 987 | |
| 988 | records.bytes_free = Max(padlen, 512); |
| 989 | records.tail->data = palloc(records.bytes_free); |
| 990 | } |
| 991 | |
| 992 | memcpy(((char *) records.tail->data) + records.tail->len, data, len); |
| 993 | records.tail->len += padlen; |
| 994 | records.bytes_free -= padlen; |
| 995 | records.total_len += padlen; |
| 996 | } |
| 997 | |
| 998 | /* |
| 999 | * Start preparing a state file. |
| 1000 | * |
| 1001 | * Initializes data structure and inserts the 2PC file header record. |
| 1002 | */ |
| 1003 | void |
| 1004 | StartPrepare(GlobalTransaction gxact) |
| 1005 | { |
| 1006 | PGPROC *proc = &ProcGlobal->allProcs[gxact->pgprocno]; |
| 1007 | PGXACT *pgxact = &ProcGlobal->allPgXact[gxact->pgprocno]; |
| 1008 | TransactionId xid = pgxact->xid; |
| 1009 | TwoPhaseFileHeader hdr; |
| 1010 | TransactionId *children; |
| 1011 | RelFileNode *commitrels; |
| 1012 | RelFileNode *abortrels; |
| 1013 | SharedInvalidationMessage *invalmsgs; |
| 1014 | |
| 1015 | /* Initialize linked list */ |
| 1016 | records.head = palloc0(sizeof(StateFileChunk)); |
| 1017 | records.head->len = 0; |
| 1018 | records.head->next = NULL; |
| 1019 | |
| 1020 | records.bytes_free = Max(sizeof(TwoPhaseFileHeader), 512); |
| 1021 | records.head->data = palloc(records.bytes_free); |
| 1022 | |
| 1023 | records.tail = records.head; |
| 1024 | records.num_chunks = 1; |
| 1025 | |
| 1026 | records.total_len = 0; |
| 1027 | |
| 1028 | /* Create header */ |
| 1029 | hdr.magic = TWOPHASE_MAGIC; |
| 1030 | hdr.total_len = 0; /* EndPrepare will fill this in */ |
| 1031 | hdr.xid = xid; |
| 1032 | hdr.database = proc->databaseId; |
| 1033 | hdr.prepared_at = gxact->prepared_at; |
| 1034 | hdr.owner = gxact->owner; |
| 1035 | hdr.nsubxacts = xactGetCommittedChildren(&children); |
| 1036 | hdr.ncommitrels = smgrGetPendingDeletes(true, &commitrels); |
| 1037 | hdr.nabortrels = smgrGetPendingDeletes(false, &abortrels); |
| 1038 | hdr.ninvalmsgs = xactGetCommittedInvalidationMessages(&invalmsgs, |
| 1039 | &hdr.initfileinval); |
| 1040 | hdr.gidlen = strlen(gxact->gid) + 1; /* Include '\0' */ |
| 1041 | |
| 1042 | save_state_data(&hdr, sizeof(TwoPhaseFileHeader)); |
| 1043 | save_state_data(gxact->gid, hdr.gidlen); |
| 1044 | |
| 1045 | /* |
| 1046 | * Add the additional info about subxacts, deletable files and cache |
| 1047 | * invalidation messages. |
| 1048 | */ |
| 1049 | if (hdr.nsubxacts > 0) |
| 1050 | { |
| 1051 | save_state_data(children, hdr.nsubxacts * sizeof(TransactionId)); |
| 1052 | /* While we have the child-xact data, stuff it in the gxact too */ |
| 1053 | GXactLoadSubxactData(gxact, hdr.nsubxacts, children); |
| 1054 | } |
| 1055 | if (hdr.ncommitrels > 0) |
| 1056 | { |
| 1057 | save_state_data(commitrels, hdr.ncommitrels * sizeof(RelFileNode)); |
| 1058 | pfree(commitrels); |
| 1059 | } |
| 1060 | if (hdr.nabortrels > 0) |
| 1061 | { |
| 1062 | save_state_data(abortrels, hdr.nabortrels * sizeof(RelFileNode)); |
| 1063 | pfree(abortrels); |
| 1064 | } |
| 1065 | if (hdr.ninvalmsgs > 0) |
| 1066 | { |
| 1067 | save_state_data(invalmsgs, |
| 1068 | hdr.ninvalmsgs * sizeof(SharedInvalidationMessage)); |
| 1069 | pfree(invalmsgs); |
| 1070 | } |
| 1071 | } |
| 1072 | |
| 1073 | /* |
| 1074 | * Finish preparing state data and writing it to WAL. |
| 1075 | */ |
| 1076 | void |
| 1077 | EndPrepare(GlobalTransaction gxact) |
| 1078 | { |
| 1079 | TwoPhaseFileHeader *hdr; |
| 1080 | StateFileChunk *record; |
| 1081 | bool replorigin; |
| 1082 | |
| 1083 | /* Add the end sentinel to the list of 2PC records */ |
| 1084 | RegisterTwoPhaseRecord(TWOPHASE_RM_END_ID, 0, |
| 1085 | NULL, 0); |
| 1086 | |
| 1087 | /* Go back and fill in total_len in the file header record */ |
| 1088 | hdr = (TwoPhaseFileHeader *) records.head->data; |
| 1089 | Assert(hdr->magic == TWOPHASE_MAGIC); |
| 1090 | hdr->total_len = records.total_len + sizeof(pg_crc32c); |
| 1091 | |
| 1092 | replorigin = (replorigin_session_origin != InvalidRepOriginId && |
| 1093 | replorigin_session_origin != DoNotReplicateId); |
| 1094 | |
| 1095 | if (replorigin) |
| 1096 | { |
| 1097 | Assert(replorigin_session_origin_lsn != InvalidXLogRecPtr); |
| 1098 | hdr->origin_lsn = replorigin_session_origin_lsn; |
| 1099 | hdr->origin_timestamp = replorigin_session_origin_timestamp; |
| 1100 | } |
| 1101 | else |
| 1102 | { |
| 1103 | hdr->origin_lsn = InvalidXLogRecPtr; |
| 1104 | hdr->origin_timestamp = 0; |
| 1105 | } |
| 1106 | |
| 1107 | /* |
| 1108 | * If the data size exceeds MaxAllocSize, we won't be able to read it in |
| 1109 | * ReadTwoPhaseFile. Check for that now, rather than fail in the case |
| 1110 | * where we write data to file and then re-read at commit time. |
| 1111 | */ |
| 1112 | if (hdr->total_len > MaxAllocSize) |
| 1113 | ereport(ERROR, |
| 1114 | (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED), |
| 1115 | errmsg("two-phase state file maximum length exceeded" ))); |
| 1116 | |
| 1117 | /* |
| 1118 | * Now writing 2PC state data to WAL. We let the WAL's CRC protection |
| 1119 | * cover us, so no need to calculate a separate CRC. |
| 1120 | * |
| 1121 | * We have to set delayChkpt here, too; otherwise a checkpoint starting |
| 1122 | * immediately after the WAL record is inserted could complete without |
| 1123 | * fsync'ing our state file. (This is essentially the same kind of race |
| 1124 | * condition as the COMMIT-to-clog-write case that RecordTransactionCommit |
| 1125 | * uses delayChkpt for; see notes there.) |
| 1126 | * |
| 1127 | * We save the PREPARE record's location in the gxact for later use by |
| 1128 | * CheckPointTwoPhase. |
| 1129 | */ |
| 1130 | XLogEnsureRecordSpace(0, records.num_chunks); |
| 1131 | |
| 1132 | START_CRIT_SECTION(); |
| 1133 | |
| 1134 | MyPgXact->delayChkpt = true; |
| 1135 | |
| 1136 | XLogBeginInsert(); |
| 1137 | for (record = records.head; record != NULL; record = record->next) |
| 1138 | XLogRegisterData(record->data, record->len); |
| 1139 | |
| 1140 | XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN); |
| 1141 | |
| 1142 | gxact->prepare_end_lsn = XLogInsert(RM_XACT_ID, XLOG_XACT_PREPARE); |
| 1143 | |
| 1144 | if (replorigin) |
| 1145 | { |
| 1146 | /* Move LSNs forward for this replication origin */ |
| 1147 | replorigin_session_advance(replorigin_session_origin_lsn, |
| 1148 | gxact->prepare_end_lsn); |
| 1149 | } |
| 1150 | |
| 1151 | XLogFlush(gxact->prepare_end_lsn); |
| 1152 | |
| 1153 | /* If we crash now, we have prepared: WAL replay will fix things */ |
| 1154 | |
| 1155 | /* Store record's start location to read that later on Commit */ |
| 1156 | gxact->prepare_start_lsn = ProcLastRecPtr; |
| 1157 | |
| 1158 | /* |
| 1159 | * Mark the prepared transaction as valid. As soon as xact.c marks |
| 1160 | * MyPgXact as not running our XID (which it will do immediately after |
| 1161 | * this function returns), others can commit/rollback the xact. |
| 1162 | * |
| 1163 | * NB: a side effect of this is to make a dummy ProcArray entry for the |
| 1164 | * prepared XID. This must happen before we clear the XID from MyPgXact, |
| 1165 | * else there is a window where the XID is not running according to |
| 1166 | * TransactionIdIsInProgress, and onlookers would be entitled to assume |
| 1167 | * the xact crashed. Instead we have a window where the same XID appears |
| 1168 | * twice in ProcArray, which is OK. |
| 1169 | */ |
| 1170 | MarkAsPrepared(gxact, false); |
| 1171 | |
| 1172 | /* |
| 1173 | * Now we can mark ourselves as out of the commit critical section: a |
| 1174 | * checkpoint starting after this will certainly see the gxact as a |
| 1175 | * candidate for fsyncing. |
| 1176 | */ |
| 1177 | MyPgXact->delayChkpt = false; |
| 1178 | |
| 1179 | /* |
| 1180 | * Remember that we have this GlobalTransaction entry locked for us. If |
| 1181 | * we crash after this point, it's too late to abort, but we must unlock |
| 1182 | * it so that the prepared transaction can be committed or rolled back. |
| 1183 | */ |
| 1184 | MyLockedGxact = gxact; |
| 1185 | |
| 1186 | END_CRIT_SECTION(); |
| 1187 | |
| 1188 | /* |
| 1189 | * Wait for synchronous replication, if required. |
| 1190 | * |
| 1191 | * Note that at this stage we have marked the prepare, but still show as |
| 1192 | * running in the procarray (twice!) and continue to hold locks. |
| 1193 | */ |
| 1194 | SyncRepWaitForLSN(gxact->prepare_end_lsn, false); |
| 1195 | |
| 1196 | records.tail = records.head = NULL; |
| 1197 | records.num_chunks = 0; |
| 1198 | } |
| 1199 | |
| 1200 | /* |
| 1201 | * Register a 2PC record to be written to state file. |
| 1202 | */ |
| 1203 | void |
| 1204 | RegisterTwoPhaseRecord(TwoPhaseRmgrId rmid, uint16 info, |
| 1205 | const void *data, uint32 len) |
| 1206 | { |
| 1207 | TwoPhaseRecordOnDisk record; |
| 1208 | |
| 1209 | record.rmid = rmid; |
| 1210 | record.info = info; |
| 1211 | record.len = len; |
| 1212 | save_state_data(&record, sizeof(TwoPhaseRecordOnDisk)); |
| 1213 | if (len > 0) |
| 1214 | save_state_data(data, len); |
| 1215 | } |
| 1216 | |
| 1217 | |
| 1218 | /* |
| 1219 | * Read and validate the state file for xid. |
| 1220 | * |
| 1221 | * If it looks OK (has a valid magic number and CRC), return the palloc'd |
| 1222 | * contents of the file, issuing an error when finding corrupted data. If |
| 1223 | * missing_ok is true, which indicates that missing files can be safely |
| 1224 | * ignored, then return NULL. This state can be reached when doing recovery. |
| 1225 | */ |
| 1226 | static char * |
| 1227 | ReadTwoPhaseFile(TransactionId xid, bool missing_ok) |
| 1228 | { |
| 1229 | char path[MAXPGPATH]; |
| 1230 | char *buf; |
| 1231 | TwoPhaseFileHeader *hdr; |
| 1232 | int fd; |
| 1233 | struct stat stat; |
| 1234 | uint32 crc_offset; |
| 1235 | pg_crc32c calc_crc, |
| 1236 | file_crc; |
| 1237 | int r; |
| 1238 | |
| 1239 | TwoPhaseFilePath(path, xid); |
| 1240 | |
| 1241 | fd = OpenTransientFile(path, O_RDONLY | PG_BINARY); |
| 1242 | if (fd < 0) |
| 1243 | { |
| 1244 | if (missing_ok && errno == ENOENT) |
| 1245 | return NULL; |
| 1246 | |
| 1247 | ereport(ERROR, |
| 1248 | (errcode_for_file_access(), |
| 1249 | errmsg("could not open file \"%s\": %m" , path))); |
| 1250 | } |
| 1251 | |
| 1252 | /* |
| 1253 | * Check file length. We can determine a lower bound pretty easily. We |
| 1254 | * set an upper bound to avoid palloc() failure on a corrupt file, though |
| 1255 | * we can't guarantee that we won't get an out of memory error anyway, |
| 1256 | * even on a valid file. |
| 1257 | */ |
| 1258 | if (fstat(fd, &stat)) |
| 1259 | ereport(ERROR, |
| 1260 | (errcode_for_file_access(), |
| 1261 | errmsg("could not stat file \"%s\": %m" , path))); |
| 1262 | |
| 1263 | if (stat.st_size < (MAXALIGN(sizeof(TwoPhaseFileHeader)) + |
| 1264 | MAXALIGN(sizeof(TwoPhaseRecordOnDisk)) + |
| 1265 | sizeof(pg_crc32c)) || |
| 1266 | stat.st_size > MaxAllocSize) |
| 1267 | ereport(ERROR, |
| 1268 | (errcode(ERRCODE_DATA_CORRUPTED), |
| 1269 | errmsg_plural("incorrect size of file \"%s\": %zu byte" , |
| 1270 | "incorrect size of file \"%s\": %zu bytes" , |
| 1271 | (Size) stat.st_size, path, |
| 1272 | (Size) stat.st_size))); |
| 1273 | |
| 1274 | crc_offset = stat.st_size - sizeof(pg_crc32c); |
| 1275 | if (crc_offset != MAXALIGN(crc_offset)) |
| 1276 | ereport(ERROR, |
| 1277 | (errcode(ERRCODE_DATA_CORRUPTED), |
| 1278 | errmsg("incorrect alignment of CRC offset for file \"%s\"" , |
| 1279 | path))); |
| 1280 | |
| 1281 | /* |
| 1282 | * OK, slurp in the file. |
| 1283 | */ |
| 1284 | buf = (char *) palloc(stat.st_size); |
| 1285 | |
| 1286 | pgstat_report_wait_start(WAIT_EVENT_TWOPHASE_FILE_READ); |
| 1287 | r = read(fd, buf, stat.st_size); |
| 1288 | if (r != stat.st_size) |
| 1289 | { |
| 1290 | if (r < 0) |
| 1291 | ereport(ERROR, |
| 1292 | (errcode_for_file_access(), |
| 1293 | errmsg("could not read file \"%s\": %m" , path))); |
| 1294 | else |
| 1295 | ereport(ERROR, |
| 1296 | (errmsg("could not read file \"%s\": read %d of %zu" , |
| 1297 | path, r, (Size) stat.st_size))); |
| 1298 | } |
| 1299 | |
| 1300 | pgstat_report_wait_end(); |
| 1301 | |
| 1302 | if (CloseTransientFile(fd)) |
| 1303 | ereport(ERROR, |
| 1304 | (errcode_for_file_access(), |
| 1305 | errmsg("could not close file \"%s\": %m" , path))); |
| 1306 | |
| 1307 | hdr = (TwoPhaseFileHeader *) buf; |
| 1308 | if (hdr->magic != TWOPHASE_MAGIC) |
| 1309 | ereport(ERROR, |
| 1310 | (errcode(ERRCODE_DATA_CORRUPTED), |
| 1311 | errmsg("invalid magic number stored in file \"%s\"" , |
| 1312 | path))); |
| 1313 | |
| 1314 | if (hdr->total_len != stat.st_size) |
| 1315 | ereport(ERROR, |
| 1316 | (errcode(ERRCODE_DATA_CORRUPTED), |
| 1317 | errmsg("invalid size stored in file \"%s\"" , |
| 1318 | path))); |
| 1319 | |
| 1320 | INIT_CRC32C(calc_crc); |
| 1321 | COMP_CRC32C(calc_crc, buf, crc_offset); |
| 1322 | FIN_CRC32C(calc_crc); |
| 1323 | |
| 1324 | file_crc = *((pg_crc32c *) (buf + crc_offset)); |
| 1325 | |
| 1326 | if (!EQ_CRC32C(calc_crc, file_crc)) |
| 1327 | ereport(ERROR, |
| 1328 | (errcode(ERRCODE_DATA_CORRUPTED), |
| 1329 | errmsg("calculated CRC checksum does not match value stored in file \"%s\"" , |
| 1330 | path))); |
| 1331 | |
| 1332 | return buf; |
| 1333 | } |
| 1334 | |
| 1335 | /* |
| 1336 | * ParsePrepareRecord |
| 1337 | */ |
| 1338 | void |
| 1339 | ParsePrepareRecord(uint8 info, char *xlrec, xl_xact_parsed_prepare *parsed) |
| 1340 | { |
| 1341 | TwoPhaseFileHeader *hdr; |
| 1342 | char *bufptr; |
| 1343 | |
| 1344 | hdr = (TwoPhaseFileHeader *) xlrec; |
| 1345 | bufptr = xlrec + MAXALIGN(sizeof(TwoPhaseFileHeader)); |
| 1346 | |
| 1347 | parsed->origin_lsn = hdr->origin_lsn; |
| 1348 | parsed->origin_timestamp = hdr->origin_timestamp; |
| 1349 | parsed->twophase_xid = hdr->xid; |
| 1350 | parsed->dbId = hdr->database; |
| 1351 | parsed->nsubxacts = hdr->nsubxacts; |
| 1352 | parsed->nrels = hdr->ncommitrels; |
| 1353 | parsed->nabortrels = hdr->nabortrels; |
| 1354 | parsed->nmsgs = hdr->ninvalmsgs; |
| 1355 | |
| 1356 | strncpy(parsed->twophase_gid, bufptr, hdr->gidlen); |
| 1357 | bufptr += MAXALIGN(hdr->gidlen); |
| 1358 | |
| 1359 | parsed->subxacts = (TransactionId *) bufptr; |
| 1360 | bufptr += MAXALIGN(hdr->nsubxacts * sizeof(TransactionId)); |
| 1361 | |
| 1362 | parsed->xnodes = (RelFileNode *) bufptr; |
| 1363 | bufptr += MAXALIGN(hdr->ncommitrels * sizeof(RelFileNode)); |
| 1364 | |
| 1365 | parsed->abortnodes = (RelFileNode *) bufptr; |
| 1366 | bufptr += MAXALIGN(hdr->nabortrels * sizeof(RelFileNode)); |
| 1367 | |
| 1368 | parsed->msgs = (SharedInvalidationMessage *) bufptr; |
| 1369 | bufptr += MAXALIGN(hdr->ninvalmsgs * sizeof(SharedInvalidationMessage)); |
| 1370 | } |
| 1371 | |
| 1372 | |
| 1373 | |
| 1374 | /* |
| 1375 | * Reads 2PC data from xlog. During checkpoint this data will be moved to |
| 1376 | * twophase files and ReadTwoPhaseFile should be used instead. |
| 1377 | * |
| 1378 | * Note clearly that this function can access WAL during normal operation, |
| 1379 | * similarly to the way WALSender or Logical Decoding would do. |
| 1380 | * |
| 1381 | */ |
| 1382 | static void |
| 1383 | XlogReadTwoPhaseData(XLogRecPtr lsn, char **buf, int *len) |
| 1384 | { |
| 1385 | XLogRecord *record; |
| 1386 | XLogReaderState *xlogreader; |
| 1387 | char *errormsg; |
| 1388 | |
| 1389 | xlogreader = XLogReaderAllocate(wal_segment_size, &read_local_xlog_page, |
| 1390 | NULL); |
| 1391 | if (!xlogreader) |
| 1392 | ereport(ERROR, |
| 1393 | (errcode(ERRCODE_OUT_OF_MEMORY), |
| 1394 | errmsg("out of memory" ), |
| 1395 | errdetail("Failed while allocating a WAL reading processor." ))); |
| 1396 | |
| 1397 | record = XLogReadRecord(xlogreader, lsn, &errormsg); |
| 1398 | if (record == NULL) |
| 1399 | ereport(ERROR, |
| 1400 | (errcode_for_file_access(), |
| 1401 | errmsg("could not read two-phase state from WAL at %X/%X" , |
| 1402 | (uint32) (lsn >> 32), |
| 1403 | (uint32) lsn))); |
| 1404 | |
| 1405 | if (XLogRecGetRmid(xlogreader) != RM_XACT_ID || |
| 1406 | (XLogRecGetInfo(xlogreader) & XLOG_XACT_OPMASK) != XLOG_XACT_PREPARE) |
| 1407 | ereport(ERROR, |
| 1408 | (errcode_for_file_access(), |
| 1409 | errmsg("expected two-phase state data is not present in WAL at %X/%X" , |
| 1410 | (uint32) (lsn >> 32), |
| 1411 | (uint32) lsn))); |
| 1412 | |
| 1413 | if (len != NULL) |
| 1414 | *len = XLogRecGetDataLen(xlogreader); |
| 1415 | |
| 1416 | *buf = palloc(sizeof(char) * XLogRecGetDataLen(xlogreader)); |
| 1417 | memcpy(*buf, XLogRecGetData(xlogreader), sizeof(char) * XLogRecGetDataLen(xlogreader)); |
| 1418 | |
| 1419 | XLogReaderFree(xlogreader); |
| 1420 | } |
| 1421 | |
| 1422 | |
| 1423 | /* |
| 1424 | * Confirms an xid is prepared, during recovery |
| 1425 | */ |
| 1426 | bool |
| 1427 | StandbyTransactionIdIsPrepared(TransactionId xid) |
| 1428 | { |
| 1429 | char *buf; |
| 1430 | TwoPhaseFileHeader *hdr; |
| 1431 | bool result; |
| 1432 | |
| 1433 | Assert(TransactionIdIsValid(xid)); |
| 1434 | |
| 1435 | if (max_prepared_xacts <= 0) |
| 1436 | return false; /* nothing to do */ |
| 1437 | |
| 1438 | /* Read and validate file */ |
| 1439 | buf = ReadTwoPhaseFile(xid, true); |
| 1440 | if (buf == NULL) |
| 1441 | return false; |
| 1442 | |
| 1443 | /* Check header also */ |
| 1444 | hdr = (TwoPhaseFileHeader *) buf; |
| 1445 | result = TransactionIdEquals(hdr->xid, xid); |
| 1446 | pfree(buf); |
| 1447 | |
| 1448 | return result; |
| 1449 | } |
| 1450 | |
| 1451 | /* |
| 1452 | * FinishPreparedTransaction: execute COMMIT PREPARED or ROLLBACK PREPARED |
| 1453 | */ |
| 1454 | void |
| 1455 | FinishPreparedTransaction(const char *gid, bool isCommit) |
| 1456 | { |
| 1457 | GlobalTransaction gxact; |
| 1458 | PGPROC *proc; |
| 1459 | PGXACT *pgxact; |
| 1460 | TransactionId xid; |
| 1461 | char *buf; |
| 1462 | char *bufptr; |
| 1463 | TwoPhaseFileHeader *hdr; |
| 1464 | TransactionId latestXid; |
| 1465 | TransactionId *children; |
| 1466 | RelFileNode *commitrels; |
| 1467 | RelFileNode *abortrels; |
| 1468 | RelFileNode *delrels; |
| 1469 | int ndelrels; |
| 1470 | SharedInvalidationMessage *invalmsgs; |
| 1471 | |
| 1472 | /* |
| 1473 | * Validate the GID, and lock the GXACT to ensure that two backends do not |
| 1474 | * try to commit the same GID at once. |
| 1475 | */ |
| 1476 | gxact = LockGXact(gid, GetUserId()); |
| 1477 | proc = &ProcGlobal->allProcs[gxact->pgprocno]; |
| 1478 | pgxact = &ProcGlobal->allPgXact[gxact->pgprocno]; |
| 1479 | xid = pgxact->xid; |
| 1480 | |
| 1481 | /* |
| 1482 | * Read and validate 2PC state data. State data will typically be stored |
| 1483 | * in WAL files if the LSN is after the last checkpoint record, or moved |
| 1484 | * to disk if for some reason they have lived for a long time. |
| 1485 | */ |
| 1486 | if (gxact->ondisk) |
| 1487 | buf = ReadTwoPhaseFile(xid, false); |
| 1488 | else |
| 1489 | XlogReadTwoPhaseData(gxact->prepare_start_lsn, &buf, NULL); |
| 1490 | |
| 1491 | |
| 1492 | /* |
| 1493 | * Disassemble the header area |
| 1494 | */ |
| 1495 | hdr = (TwoPhaseFileHeader *) buf; |
| 1496 | Assert(TransactionIdEquals(hdr->xid, xid)); |
| 1497 | bufptr = buf + MAXALIGN(sizeof(TwoPhaseFileHeader)); |
| 1498 | bufptr += MAXALIGN(hdr->gidlen); |
| 1499 | children = (TransactionId *) bufptr; |
| 1500 | bufptr += MAXALIGN(hdr->nsubxacts * sizeof(TransactionId)); |
| 1501 | commitrels = (RelFileNode *) bufptr; |
| 1502 | bufptr += MAXALIGN(hdr->ncommitrels * sizeof(RelFileNode)); |
| 1503 | abortrels = (RelFileNode *) bufptr; |
| 1504 | bufptr += MAXALIGN(hdr->nabortrels * sizeof(RelFileNode)); |
| 1505 | invalmsgs = (SharedInvalidationMessage *) bufptr; |
| 1506 | bufptr += MAXALIGN(hdr->ninvalmsgs * sizeof(SharedInvalidationMessage)); |
| 1507 | |
| 1508 | /* compute latestXid among all children */ |
| 1509 | latestXid = TransactionIdLatest(xid, hdr->nsubxacts, children); |
| 1510 | |
| 1511 | /* Prevent cancel/die interrupt while cleaning up */ |
| 1512 | HOLD_INTERRUPTS(); |
| 1513 | |
| 1514 | /* |
| 1515 | * The order of operations here is critical: make the XLOG entry for |
| 1516 | * commit or abort, then mark the transaction committed or aborted in |
| 1517 | * pg_xact, then remove its PGPROC from the global ProcArray (which means |
| 1518 | * TransactionIdIsInProgress will stop saying the prepared xact is in |
| 1519 | * progress), then run the post-commit or post-abort callbacks. The |
| 1520 | * callbacks will release the locks the transaction held. |
| 1521 | */ |
| 1522 | if (isCommit) |
| 1523 | RecordTransactionCommitPrepared(xid, |
| 1524 | hdr->nsubxacts, children, |
| 1525 | hdr->ncommitrels, commitrels, |
| 1526 | hdr->ninvalmsgs, invalmsgs, |
| 1527 | hdr->initfileinval, gid); |
| 1528 | else |
| 1529 | RecordTransactionAbortPrepared(xid, |
| 1530 | hdr->nsubxacts, children, |
| 1531 | hdr->nabortrels, abortrels, |
| 1532 | gid); |
| 1533 | |
| 1534 | ProcArrayRemove(proc, latestXid); |
| 1535 | |
| 1536 | /* |
| 1537 | * In case we fail while running the callbacks, mark the gxact invalid so |
| 1538 | * no one else will try to commit/rollback, and so it will be recycled if |
| 1539 | * we fail after this point. It is still locked by our backend so it |
| 1540 | * won't go away yet. |
| 1541 | * |
| 1542 | * (We assume it's safe to do this without taking TwoPhaseStateLock.) |
| 1543 | */ |
| 1544 | gxact->valid = false; |
| 1545 | |
| 1546 | /* |
| 1547 | * We have to remove any files that were supposed to be dropped. For |
| 1548 | * consistency with the regular xact.c code paths, must do this before |
| 1549 | * releasing locks, so do it before running the callbacks. |
| 1550 | * |
| 1551 | * NB: this code knows that we couldn't be dropping any temp rels ... |
| 1552 | */ |
| 1553 | if (isCommit) |
| 1554 | { |
| 1555 | delrels = commitrels; |
| 1556 | ndelrels = hdr->ncommitrels; |
| 1557 | } |
| 1558 | else |
| 1559 | { |
| 1560 | delrels = abortrels; |
| 1561 | ndelrels = hdr->nabortrels; |
| 1562 | } |
| 1563 | |
| 1564 | /* Make sure files supposed to be dropped are dropped */ |
| 1565 | DropRelationFiles(delrels, ndelrels, false); |
| 1566 | |
| 1567 | /* |
| 1568 | * Handle cache invalidation messages. |
| 1569 | * |
| 1570 | * Relcache init file invalidation requires processing both before and |
| 1571 | * after we send the SI messages. See AtEOXact_Inval() |
| 1572 | */ |
| 1573 | if (hdr->initfileinval) |
| 1574 | RelationCacheInitFilePreInvalidate(); |
| 1575 | SendSharedInvalidMessages(invalmsgs, hdr->ninvalmsgs); |
| 1576 | if (hdr->initfileinval) |
| 1577 | RelationCacheInitFilePostInvalidate(); |
| 1578 | |
| 1579 | /* |
| 1580 | * Acquire the two-phase lock. We want to work on the two-phase callbacks |
| 1581 | * while holding it to avoid potential conflicts with other transactions |
| 1582 | * attempting to use the same GID, so the lock is released once the shared |
| 1583 | * memory state is cleared. |
| 1584 | */ |
| 1585 | LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE); |
| 1586 | |
| 1587 | /* And now do the callbacks */ |
| 1588 | if (isCommit) |
| 1589 | ProcessRecords(bufptr, xid, twophase_postcommit_callbacks); |
| 1590 | else |
| 1591 | ProcessRecords(bufptr, xid, twophase_postabort_callbacks); |
| 1592 | |
| 1593 | PredicateLockTwoPhaseFinish(xid, isCommit); |
| 1594 | |
| 1595 | /* Clear shared memory state */ |
| 1596 | RemoveGXact(gxact); |
| 1597 | |
| 1598 | /* |
| 1599 | * Release the lock as all callbacks are called and shared memory cleanup |
| 1600 | * is done. |
| 1601 | */ |
| 1602 | LWLockRelease(TwoPhaseStateLock); |
| 1603 | |
| 1604 | /* Count the prepared xact as committed or aborted */ |
| 1605 | AtEOXact_PgStat(isCommit, false); |
| 1606 | |
| 1607 | /* |
| 1608 | * And now we can clean up any files we may have left. |
| 1609 | */ |
| 1610 | if (gxact->ondisk) |
| 1611 | RemoveTwoPhaseFile(xid, true); |
| 1612 | |
| 1613 | MyLockedGxact = NULL; |
| 1614 | |
| 1615 | RESUME_INTERRUPTS(); |
| 1616 | |
| 1617 | pfree(buf); |
| 1618 | } |
| 1619 | |
| 1620 | /* |
| 1621 | * Scan 2PC state data in memory and call the indicated callbacks for each 2PC record. |
| 1622 | */ |
| 1623 | static void |
| 1624 | ProcessRecords(char *bufptr, TransactionId xid, |
| 1625 | const TwoPhaseCallback callbacks[]) |
| 1626 | { |
| 1627 | for (;;) |
| 1628 | { |
| 1629 | TwoPhaseRecordOnDisk *record = (TwoPhaseRecordOnDisk *) bufptr; |
| 1630 | |
| 1631 | Assert(record->rmid <= TWOPHASE_RM_MAX_ID); |
| 1632 | if (record->rmid == TWOPHASE_RM_END_ID) |
| 1633 | break; |
| 1634 | |
| 1635 | bufptr += MAXALIGN(sizeof(TwoPhaseRecordOnDisk)); |
| 1636 | |
| 1637 | if (callbacks[record->rmid] != NULL) |
| 1638 | callbacks[record->rmid] (xid, record->info, |
| 1639 | (void *) bufptr, record->len); |
| 1640 | |
| 1641 | bufptr += MAXALIGN(record->len); |
| 1642 | } |
| 1643 | } |
| 1644 | |
| 1645 | /* |
| 1646 | * Remove the 2PC file for the specified XID. |
| 1647 | * |
| 1648 | * If giveWarning is false, do not complain about file-not-present; |
| 1649 | * this is an expected case during WAL replay. |
| 1650 | */ |
| 1651 | static void |
| 1652 | RemoveTwoPhaseFile(TransactionId xid, bool giveWarning) |
| 1653 | { |
| 1654 | char path[MAXPGPATH]; |
| 1655 | |
| 1656 | TwoPhaseFilePath(path, xid); |
| 1657 | if (unlink(path)) |
| 1658 | if (errno != ENOENT || giveWarning) |
| 1659 | ereport(WARNING, |
| 1660 | (errcode_for_file_access(), |
| 1661 | errmsg("could not remove file \"%s\": %m" , path))); |
| 1662 | } |
| 1663 | |
| 1664 | /* |
| 1665 | * Recreates a state file. This is used in WAL replay and during |
| 1666 | * checkpoint creation. |
| 1667 | * |
| 1668 | * Note: content and len don't include CRC. |
| 1669 | */ |
| 1670 | static void |
| 1671 | RecreateTwoPhaseFile(TransactionId xid, void *content, int len) |
| 1672 | { |
| 1673 | char path[MAXPGPATH]; |
| 1674 | pg_crc32c statefile_crc; |
| 1675 | int fd; |
| 1676 | |
| 1677 | /* Recompute CRC */ |
| 1678 | INIT_CRC32C(statefile_crc); |
| 1679 | COMP_CRC32C(statefile_crc, content, len); |
| 1680 | FIN_CRC32C(statefile_crc); |
| 1681 | |
| 1682 | TwoPhaseFilePath(path, xid); |
| 1683 | |
| 1684 | fd = OpenTransientFile(path, |
| 1685 | O_CREAT | O_TRUNC | O_WRONLY | PG_BINARY); |
| 1686 | if (fd < 0) |
| 1687 | ereport(ERROR, |
| 1688 | (errcode_for_file_access(), |
| 1689 | errmsg("could not recreate file \"%s\": %m" , path))); |
| 1690 | |
| 1691 | /* Write content and CRC */ |
| 1692 | errno = 0; |
| 1693 | pgstat_report_wait_start(WAIT_EVENT_TWOPHASE_FILE_WRITE); |
| 1694 | if (write(fd, content, len) != len) |
| 1695 | { |
| 1696 | /* if write didn't set errno, assume problem is no disk space */ |
| 1697 | if (errno == 0) |
| 1698 | errno = ENOSPC; |
| 1699 | ereport(ERROR, |
| 1700 | (errcode_for_file_access(), |
| 1701 | errmsg("could not write file \"%s\": %m" , path))); |
| 1702 | } |
| 1703 | if (write(fd, &statefile_crc, sizeof(pg_crc32c)) != sizeof(pg_crc32c)) |
| 1704 | { |
| 1705 | /* if write didn't set errno, assume problem is no disk space */ |
| 1706 | if (errno == 0) |
| 1707 | errno = ENOSPC; |
| 1708 | ereport(ERROR, |
| 1709 | (errcode_for_file_access(), |
| 1710 | errmsg("could not write file \"%s\": %m" , path))); |
| 1711 | } |
| 1712 | pgstat_report_wait_end(); |
| 1713 | |
| 1714 | /* |
| 1715 | * We must fsync the file because the end-of-replay checkpoint will not do |
| 1716 | * so, there being no GXACT in shared memory yet to tell it to. |
| 1717 | */ |
| 1718 | pgstat_report_wait_start(WAIT_EVENT_TWOPHASE_FILE_SYNC); |
| 1719 | if (pg_fsync(fd) != 0) |
| 1720 | ereport(ERROR, |
| 1721 | (errcode_for_file_access(), |
| 1722 | errmsg("could not fsync file \"%s\": %m" , path))); |
| 1723 | pgstat_report_wait_end(); |
| 1724 | |
| 1725 | if (CloseTransientFile(fd) != 0) |
| 1726 | ereport(ERROR, |
| 1727 | (errcode_for_file_access(), |
| 1728 | errmsg("could not close file \"%s\": %m" , path))); |
| 1729 | } |
| 1730 | |
| 1731 | /* |
| 1732 | * CheckPointTwoPhase -- handle 2PC component of checkpointing. |
| 1733 | * |
| 1734 | * We must fsync the state file of any GXACT that is valid or has been |
| 1735 | * generated during redo and has a PREPARE LSN <= the checkpoint's redo |
| 1736 | * horizon. (If the gxact isn't valid yet, has not been generated in |
| 1737 | * redo, or has a later LSN, this checkpoint is not responsible for |
| 1738 | * fsyncing it.) |
| 1739 | * |
| 1740 | * This is deliberately run as late as possible in the checkpoint sequence, |
| 1741 | * because GXACTs ordinarily have short lifespans, and so it is quite |
| 1742 | * possible that GXACTs that were valid at checkpoint start will no longer |
| 1743 | * exist if we wait a little bit. With typical checkpoint settings this |
| 1744 | * will be about 3 minutes for an online checkpoint, so as a result we |
| 1745 | * expect that there will be no GXACTs that need to be copied to disk. |
| 1746 | * |
| 1747 | * If a GXACT remains valid across multiple checkpoints, it will already |
| 1748 | * be on disk so we don't bother to repeat that write. |
| 1749 | */ |
| 1750 | void |
| 1751 | CheckPointTwoPhase(XLogRecPtr redo_horizon) |
| 1752 | { |
| 1753 | int i; |
| 1754 | int serialized_xacts = 0; |
| 1755 | |
| 1756 | if (max_prepared_xacts <= 0) |
| 1757 | return; /* nothing to do */ |
| 1758 | |
| 1759 | TRACE_POSTGRESQL_TWOPHASE_CHECKPOINT_START(); |
| 1760 | |
| 1761 | /* |
| 1762 | * We are expecting there to be zero GXACTs that need to be copied to |
| 1763 | * disk, so we perform all I/O while holding TwoPhaseStateLock for |
| 1764 | * simplicity. This prevents any new xacts from preparing while this |
| 1765 | * occurs, which shouldn't be a problem since the presence of long-lived |
| 1766 | * prepared xacts indicates the transaction manager isn't active. |
| 1767 | * |
| 1768 | * It's also possible to move I/O out of the lock, but on every error we |
| 1769 | * should check whether somebody committed our transaction in different |
| 1770 | * backend. Let's leave this optimization for future, if somebody will |
| 1771 | * spot that this place cause bottleneck. |
| 1772 | * |
| 1773 | * Note that it isn't possible for there to be a GXACT with a |
| 1774 | * prepare_end_lsn set prior to the last checkpoint yet is marked invalid, |
| 1775 | * because of the efforts with delayChkpt. |
| 1776 | */ |
| 1777 | LWLockAcquire(TwoPhaseStateLock, LW_SHARED); |
| 1778 | for (i = 0; i < TwoPhaseState->numPrepXacts; i++) |
| 1779 | { |
| 1780 | /* |
| 1781 | * Note that we are using gxact not pgxact so this works in recovery |
| 1782 | * also |
| 1783 | */ |
| 1784 | GlobalTransaction gxact = TwoPhaseState->prepXacts[i]; |
| 1785 | |
| 1786 | if ((gxact->valid || gxact->inredo) && |
| 1787 | !gxact->ondisk && |
| 1788 | gxact->prepare_end_lsn <= redo_horizon) |
| 1789 | { |
| 1790 | char *buf; |
| 1791 | int len; |
| 1792 | |
| 1793 | XlogReadTwoPhaseData(gxact->prepare_start_lsn, &buf, &len); |
| 1794 | RecreateTwoPhaseFile(gxact->xid, buf, len); |
| 1795 | gxact->ondisk = true; |
| 1796 | gxact->prepare_start_lsn = InvalidXLogRecPtr; |
| 1797 | gxact->prepare_end_lsn = InvalidXLogRecPtr; |
| 1798 | pfree(buf); |
| 1799 | serialized_xacts++; |
| 1800 | } |
| 1801 | } |
| 1802 | LWLockRelease(TwoPhaseStateLock); |
| 1803 | |
| 1804 | /* |
| 1805 | * Flush unconditionally the parent directory to make any information |
| 1806 | * durable on disk. Two-phase files could have been removed and those |
| 1807 | * removals need to be made persistent as well as any files newly created |
| 1808 | * previously since the last checkpoint. |
| 1809 | */ |
| 1810 | fsync_fname(TWOPHASE_DIR, true); |
| 1811 | |
| 1812 | TRACE_POSTGRESQL_TWOPHASE_CHECKPOINT_DONE(); |
| 1813 | |
| 1814 | if (log_checkpoints && serialized_xacts > 0) |
| 1815 | ereport(LOG, |
| 1816 | (errmsg_plural("%u two-phase state file was written " |
| 1817 | "for a long-running prepared transaction" , |
| 1818 | "%u two-phase state files were written " |
| 1819 | "for long-running prepared transactions" , |
| 1820 | serialized_xacts, |
| 1821 | serialized_xacts))); |
| 1822 | } |
| 1823 | |
| 1824 | /* |
| 1825 | * restoreTwoPhaseData |
| 1826 | * |
| 1827 | * Scan pg_twophase and fill TwoPhaseState depending on the on-disk data. |
| 1828 | * This is called once at the beginning of recovery, saving any extra |
| 1829 | * lookups in the future. Two-phase files that are newer than the |
| 1830 | * minimum XID horizon are discarded on the way. |
| 1831 | */ |
| 1832 | void |
| 1833 | restoreTwoPhaseData(void) |
| 1834 | { |
| 1835 | DIR *cldir; |
| 1836 | struct dirent *clde; |
| 1837 | |
| 1838 | LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE); |
| 1839 | cldir = AllocateDir(TWOPHASE_DIR); |
| 1840 | while ((clde = ReadDir(cldir, TWOPHASE_DIR)) != NULL) |
| 1841 | { |
| 1842 | if (strlen(clde->d_name) == 8 && |
| 1843 | strspn(clde->d_name, "0123456789ABCDEF" ) == 8) |
| 1844 | { |
| 1845 | TransactionId xid; |
| 1846 | char *buf; |
| 1847 | |
| 1848 | xid = (TransactionId) strtoul(clde->d_name, NULL, 16); |
| 1849 | |
| 1850 | buf = ProcessTwoPhaseBuffer(xid, InvalidXLogRecPtr, |
| 1851 | true, false, false); |
| 1852 | if (buf == NULL) |
| 1853 | continue; |
| 1854 | |
| 1855 | PrepareRedoAdd(buf, InvalidXLogRecPtr, |
| 1856 | InvalidXLogRecPtr, InvalidRepOriginId); |
| 1857 | } |
| 1858 | } |
| 1859 | LWLockRelease(TwoPhaseStateLock); |
| 1860 | FreeDir(cldir); |
| 1861 | } |
| 1862 | |
| 1863 | /* |
| 1864 | * PrescanPreparedTransactions |
| 1865 | * |
| 1866 | * Scan the shared memory entries of TwoPhaseState and determine the range |
| 1867 | * of valid XIDs present. This is run during database startup, after we |
| 1868 | * have completed reading WAL. ShmemVariableCache->nextFullXid has been set to |
| 1869 | * one more than the highest XID for which evidence exists in WAL. |
| 1870 | * |
| 1871 | * We throw away any prepared xacts with main XID beyond nextFullXid --- if any |
| 1872 | * are present, it suggests that the DBA has done a PITR recovery to an |
| 1873 | * earlier point in time without cleaning out pg_twophase. We dare not |
| 1874 | * try to recover such prepared xacts since they likely depend on database |
| 1875 | * state that doesn't exist now. |
| 1876 | * |
| 1877 | * However, we will advance nextFullXid beyond any subxact XIDs belonging to |
| 1878 | * valid prepared xacts. We need to do this since subxact commit doesn't |
| 1879 | * write a WAL entry, and so there might be no evidence in WAL of those |
| 1880 | * subxact XIDs. |
| 1881 | * |
| 1882 | * On corrupted two-phase files, fail immediately. Keeping around broken |
| 1883 | * entries and let replay continue causes harm on the system, and a new |
| 1884 | * backup should be rolled in. |
| 1885 | * |
| 1886 | * Our other responsibility is to determine and return the oldest valid XID |
| 1887 | * among the prepared xacts (if none, return ShmemVariableCache->nextFullXid). |
| 1888 | * This is needed to synchronize pg_subtrans startup properly. |
| 1889 | * |
| 1890 | * If xids_p and nxids_p are not NULL, pointer to a palloc'd array of all |
| 1891 | * top-level xids is stored in *xids_p. The number of entries in the array |
| 1892 | * is returned in *nxids_p. |
| 1893 | */ |
| 1894 | TransactionId |
| 1895 | PrescanPreparedTransactions(TransactionId **xids_p, int *nxids_p) |
| 1896 | { |
| 1897 | FullTransactionId nextFullXid = ShmemVariableCache->nextFullXid; |
| 1898 | TransactionId origNextXid = XidFromFullTransactionId(nextFullXid); |
| 1899 | TransactionId result = origNextXid; |
| 1900 | TransactionId *xids = NULL; |
| 1901 | int nxids = 0; |
| 1902 | int allocsize = 0; |
| 1903 | int i; |
| 1904 | |
| 1905 | LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE); |
| 1906 | for (i = 0; i < TwoPhaseState->numPrepXacts; i++) |
| 1907 | { |
| 1908 | TransactionId xid; |
| 1909 | char *buf; |
| 1910 | GlobalTransaction gxact = TwoPhaseState->prepXacts[i]; |
| 1911 | |
| 1912 | Assert(gxact->inredo); |
| 1913 | |
| 1914 | xid = gxact->xid; |
| 1915 | |
| 1916 | buf = ProcessTwoPhaseBuffer(xid, |
| 1917 | gxact->prepare_start_lsn, |
| 1918 | gxact->ondisk, false, true); |
| 1919 | |
| 1920 | if (buf == NULL) |
| 1921 | continue; |
| 1922 | |
| 1923 | /* |
| 1924 | * OK, we think this file is valid. Incorporate xid into the |
| 1925 | * running-minimum result. |
| 1926 | */ |
| 1927 | if (TransactionIdPrecedes(xid, result)) |
| 1928 | result = xid; |
| 1929 | |
| 1930 | if (xids_p) |
| 1931 | { |
| 1932 | if (nxids == allocsize) |
| 1933 | { |
| 1934 | if (nxids == 0) |
| 1935 | { |
| 1936 | allocsize = 10; |
| 1937 | xids = palloc(allocsize * sizeof(TransactionId)); |
| 1938 | } |
| 1939 | else |
| 1940 | { |
| 1941 | allocsize = allocsize * 2; |
| 1942 | xids = repalloc(xids, allocsize * sizeof(TransactionId)); |
| 1943 | } |
| 1944 | } |
| 1945 | xids[nxids++] = xid; |
| 1946 | } |
| 1947 | |
| 1948 | pfree(buf); |
| 1949 | } |
| 1950 | LWLockRelease(TwoPhaseStateLock); |
| 1951 | |
| 1952 | if (xids_p) |
| 1953 | { |
| 1954 | *xids_p = xids; |
| 1955 | *nxids_p = nxids; |
| 1956 | } |
| 1957 | |
| 1958 | return result; |
| 1959 | } |
| 1960 | |
| 1961 | /* |
| 1962 | * StandbyRecoverPreparedTransactions |
| 1963 | * |
| 1964 | * Scan the shared memory entries of TwoPhaseState and setup all the required |
| 1965 | * information to allow standby queries to treat prepared transactions as still |
| 1966 | * active. |
| 1967 | * |
| 1968 | * This is never called at the end of recovery - we use |
| 1969 | * RecoverPreparedTransactions() at that point. |
| 1970 | * |
| 1971 | * The lack of calls to SubTransSetParent() calls here is by design; |
| 1972 | * those calls are made by RecoverPreparedTransactions() at the end of recovery |
| 1973 | * for those xacts that need this. |
| 1974 | */ |
| 1975 | void |
| 1976 | StandbyRecoverPreparedTransactions(void) |
| 1977 | { |
| 1978 | int i; |
| 1979 | |
| 1980 | LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE); |
| 1981 | for (i = 0; i < TwoPhaseState->numPrepXacts; i++) |
| 1982 | { |
| 1983 | TransactionId xid; |
| 1984 | char *buf; |
| 1985 | GlobalTransaction gxact = TwoPhaseState->prepXacts[i]; |
| 1986 | |
| 1987 | Assert(gxact->inredo); |
| 1988 | |
| 1989 | xid = gxact->xid; |
| 1990 | |
| 1991 | buf = ProcessTwoPhaseBuffer(xid, |
| 1992 | gxact->prepare_start_lsn, |
| 1993 | gxact->ondisk, false, false); |
| 1994 | if (buf != NULL) |
| 1995 | pfree(buf); |
| 1996 | } |
| 1997 | LWLockRelease(TwoPhaseStateLock); |
| 1998 | } |
| 1999 | |
| 2000 | /* |
| 2001 | * RecoverPreparedTransactions |
| 2002 | * |
| 2003 | * Scan the shared memory entries of TwoPhaseState and reload the state for |
| 2004 | * each prepared transaction (reacquire locks, etc). |
| 2005 | * |
| 2006 | * This is run at the end of recovery, but before we allow backends to write |
| 2007 | * WAL. |
| 2008 | * |
| 2009 | * At the end of recovery the way we take snapshots will change. We now need |
| 2010 | * to mark all running transactions with their full SubTransSetParent() info |
| 2011 | * to allow normal snapshots to work correctly if snapshots overflow. |
| 2012 | * We do this here because by definition prepared transactions are the only |
| 2013 | * type of write transaction still running, so this is necessary and |
| 2014 | * complete. |
| 2015 | */ |
| 2016 | void |
| 2017 | RecoverPreparedTransactions(void) |
| 2018 | { |
| 2019 | int i; |
| 2020 | |
| 2021 | LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE); |
| 2022 | for (i = 0; i < TwoPhaseState->numPrepXacts; i++) |
| 2023 | { |
| 2024 | TransactionId xid; |
| 2025 | char *buf; |
| 2026 | GlobalTransaction gxact = TwoPhaseState->prepXacts[i]; |
| 2027 | char *bufptr; |
| 2028 | TwoPhaseFileHeader *hdr; |
| 2029 | TransactionId *subxids; |
| 2030 | const char *gid; |
| 2031 | |
| 2032 | xid = gxact->xid; |
| 2033 | |
| 2034 | /* |
| 2035 | * Reconstruct subtrans state for the transaction --- needed because |
| 2036 | * pg_subtrans is not preserved over a restart. Note that we are |
| 2037 | * linking all the subtransactions directly to the top-level XID; |
| 2038 | * there may originally have been a more complex hierarchy, but |
| 2039 | * there's no need to restore that exactly. It's possible that |
| 2040 | * SubTransSetParent has been set before, if the prepared transaction |
| 2041 | * generated xid assignment records. |
| 2042 | */ |
| 2043 | buf = ProcessTwoPhaseBuffer(xid, |
| 2044 | gxact->prepare_start_lsn, |
| 2045 | gxact->ondisk, true, false); |
| 2046 | if (buf == NULL) |
| 2047 | continue; |
| 2048 | |
| 2049 | ereport(LOG, |
| 2050 | (errmsg("recovering prepared transaction %u from shared memory" , xid))); |
| 2051 | |
| 2052 | hdr = (TwoPhaseFileHeader *) buf; |
| 2053 | Assert(TransactionIdEquals(hdr->xid, xid)); |
| 2054 | bufptr = buf + MAXALIGN(sizeof(TwoPhaseFileHeader)); |
| 2055 | gid = (const char *) bufptr; |
| 2056 | bufptr += MAXALIGN(hdr->gidlen); |
| 2057 | subxids = (TransactionId *) bufptr; |
| 2058 | bufptr += MAXALIGN(hdr->nsubxacts * sizeof(TransactionId)); |
| 2059 | bufptr += MAXALIGN(hdr->ncommitrels * sizeof(RelFileNode)); |
| 2060 | bufptr += MAXALIGN(hdr->nabortrels * sizeof(RelFileNode)); |
| 2061 | bufptr += MAXALIGN(hdr->ninvalmsgs * sizeof(SharedInvalidationMessage)); |
| 2062 | |
| 2063 | /* |
| 2064 | * Recreate its GXACT and dummy PGPROC. But, check whether it was |
| 2065 | * added in redo and already has a shmem entry for it. |
| 2066 | */ |
| 2067 | MarkAsPreparingGuts(gxact, xid, gid, |
| 2068 | hdr->prepared_at, |
| 2069 | hdr->owner, hdr->database); |
| 2070 | |
| 2071 | /* recovered, so reset the flag for entries generated by redo */ |
| 2072 | gxact->inredo = false; |
| 2073 | |
| 2074 | GXactLoadSubxactData(gxact, hdr->nsubxacts, subxids); |
| 2075 | MarkAsPrepared(gxact, true); |
| 2076 | |
| 2077 | LWLockRelease(TwoPhaseStateLock); |
| 2078 | |
| 2079 | /* |
| 2080 | * Recover other state (notably locks) using resource managers. |
| 2081 | */ |
| 2082 | ProcessRecords(bufptr, xid, twophase_recover_callbacks); |
| 2083 | |
| 2084 | /* |
| 2085 | * Release locks held by the standby process after we process each |
| 2086 | * prepared transaction. As a result, we don't need too many |
| 2087 | * additional locks at any one time. |
| 2088 | */ |
| 2089 | if (InHotStandby) |
| 2090 | StandbyReleaseLockTree(xid, hdr->nsubxacts, subxids); |
| 2091 | |
| 2092 | /* |
| 2093 | * We're done with recovering this transaction. Clear MyLockedGxact, |
| 2094 | * like we do in PrepareTransaction() during normal operation. |
| 2095 | */ |
| 2096 | PostPrepare_Twophase(); |
| 2097 | |
| 2098 | pfree(buf); |
| 2099 | |
| 2100 | LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE); |
| 2101 | } |
| 2102 | |
| 2103 | LWLockRelease(TwoPhaseStateLock); |
| 2104 | } |
| 2105 | |
| 2106 | /* |
| 2107 | * ProcessTwoPhaseBuffer |
| 2108 | * |
| 2109 | * Given a transaction id, read it either from disk or read it directly |
| 2110 | * via shmem xlog record pointer using the provided "prepare_start_lsn". |
| 2111 | * |
| 2112 | * If setParent is true, set up subtransaction parent linkages. |
| 2113 | * |
| 2114 | * If setNextXid is true, set ShmemVariableCache->nextFullXid to the newest |
| 2115 | * value scanned. |
| 2116 | */ |
| 2117 | static char * |
| 2118 | ProcessTwoPhaseBuffer(TransactionId xid, |
| 2119 | XLogRecPtr prepare_start_lsn, |
| 2120 | bool fromdisk, |
| 2121 | bool setParent, bool setNextXid) |
| 2122 | { |
| 2123 | FullTransactionId nextFullXid = ShmemVariableCache->nextFullXid; |
| 2124 | TransactionId origNextXid = XidFromFullTransactionId(nextFullXid); |
| 2125 | TransactionId *subxids; |
| 2126 | char *buf; |
| 2127 | TwoPhaseFileHeader *hdr; |
| 2128 | int i; |
| 2129 | |
| 2130 | Assert(LWLockHeldByMeInMode(TwoPhaseStateLock, LW_EXCLUSIVE)); |
| 2131 | |
| 2132 | if (!fromdisk) |
| 2133 | Assert(prepare_start_lsn != InvalidXLogRecPtr); |
| 2134 | |
| 2135 | /* Already processed? */ |
| 2136 | if (TransactionIdDidCommit(xid) || TransactionIdDidAbort(xid)) |
| 2137 | { |
| 2138 | if (fromdisk) |
| 2139 | { |
| 2140 | ereport(WARNING, |
| 2141 | (errmsg("removing stale two-phase state file for transaction %u" , |
| 2142 | xid))); |
| 2143 | RemoveTwoPhaseFile(xid, true); |
| 2144 | } |
| 2145 | else |
| 2146 | { |
| 2147 | ereport(WARNING, |
| 2148 | (errmsg("removing stale two-phase state from memory for transaction %u" , |
| 2149 | xid))); |
| 2150 | PrepareRedoRemove(xid, true); |
| 2151 | } |
| 2152 | return NULL; |
| 2153 | } |
| 2154 | |
| 2155 | /* Reject XID if too new */ |
| 2156 | if (TransactionIdFollowsOrEquals(xid, origNextXid)) |
| 2157 | { |
| 2158 | if (fromdisk) |
| 2159 | { |
| 2160 | ereport(WARNING, |
| 2161 | (errmsg("removing future two-phase state file for transaction %u" , |
| 2162 | xid))); |
| 2163 | RemoveTwoPhaseFile(xid, true); |
| 2164 | } |
| 2165 | else |
| 2166 | { |
| 2167 | ereport(WARNING, |
| 2168 | (errmsg("removing future two-phase state from memory for transaction %u" , |
| 2169 | xid))); |
| 2170 | PrepareRedoRemove(xid, true); |
| 2171 | } |
| 2172 | return NULL; |
| 2173 | } |
| 2174 | |
| 2175 | if (fromdisk) |
| 2176 | { |
| 2177 | /* Read and validate file */ |
| 2178 | buf = ReadTwoPhaseFile(xid, false); |
| 2179 | } |
| 2180 | else |
| 2181 | { |
| 2182 | /* Read xlog data */ |
| 2183 | XlogReadTwoPhaseData(prepare_start_lsn, &buf, NULL); |
| 2184 | } |
| 2185 | |
| 2186 | /* Deconstruct header */ |
| 2187 | hdr = (TwoPhaseFileHeader *) buf; |
| 2188 | if (!TransactionIdEquals(hdr->xid, xid)) |
| 2189 | { |
| 2190 | if (fromdisk) |
| 2191 | ereport(ERROR, |
| 2192 | (errcode(ERRCODE_DATA_CORRUPTED), |
| 2193 | errmsg("corrupted two-phase state file for transaction %u" , |
| 2194 | xid))); |
| 2195 | else |
| 2196 | ereport(ERROR, |
| 2197 | (errcode(ERRCODE_DATA_CORRUPTED), |
| 2198 | errmsg("corrupted two-phase state in memory for transaction %u" , |
| 2199 | xid))); |
| 2200 | } |
| 2201 | |
| 2202 | /* |
| 2203 | * Examine subtransaction XIDs ... they should all follow main XID, and |
| 2204 | * they may force us to advance nextFullXid. |
| 2205 | */ |
| 2206 | subxids = (TransactionId *) (buf + |
| 2207 | MAXALIGN(sizeof(TwoPhaseFileHeader)) + |
| 2208 | MAXALIGN(hdr->gidlen)); |
| 2209 | for (i = 0; i < hdr->nsubxacts; i++) |
| 2210 | { |
| 2211 | TransactionId subxid = subxids[i]; |
| 2212 | |
| 2213 | Assert(TransactionIdFollows(subxid, xid)); |
| 2214 | |
| 2215 | /* update nextFullXid if needed */ |
| 2216 | if (setNextXid) |
| 2217 | AdvanceNextFullTransactionIdPastXid(subxid); |
| 2218 | |
| 2219 | if (setParent) |
| 2220 | SubTransSetParent(subxid, xid); |
| 2221 | } |
| 2222 | |
| 2223 | return buf; |
| 2224 | } |
| 2225 | |
| 2226 | |
| 2227 | /* |
| 2228 | * RecordTransactionCommitPrepared |
| 2229 | * |
| 2230 | * This is basically the same as RecordTransactionCommit (q.v. if you change |
| 2231 | * this function): in particular, we must set the delayChkpt flag to avoid a |
| 2232 | * race condition. |
| 2233 | * |
| 2234 | * We know the transaction made at least one XLOG entry (its PREPARE), |
| 2235 | * so it is never possible to optimize out the commit record. |
| 2236 | */ |
| 2237 | static void |
| 2238 | RecordTransactionCommitPrepared(TransactionId xid, |
| 2239 | int nchildren, |
| 2240 | TransactionId *children, |
| 2241 | int nrels, |
| 2242 | RelFileNode *rels, |
| 2243 | int ninvalmsgs, |
| 2244 | SharedInvalidationMessage *invalmsgs, |
| 2245 | bool initfileinval, |
| 2246 | const char *gid) |
| 2247 | { |
| 2248 | XLogRecPtr recptr; |
| 2249 | TimestampTz committs = GetCurrentTimestamp(); |
| 2250 | bool replorigin; |
| 2251 | |
| 2252 | /* |
| 2253 | * Are we using the replication origins feature? Or, in other words, are |
| 2254 | * we replaying remote actions? |
| 2255 | */ |
| 2256 | replorigin = (replorigin_session_origin != InvalidRepOriginId && |
| 2257 | replorigin_session_origin != DoNotReplicateId); |
| 2258 | |
| 2259 | START_CRIT_SECTION(); |
| 2260 | |
| 2261 | /* See notes in RecordTransactionCommit */ |
| 2262 | MyPgXact->delayChkpt = true; |
| 2263 | |
| 2264 | /* |
| 2265 | * Emit the XLOG commit record. Note that we mark 2PC commits as |
| 2266 | * potentially having AccessExclusiveLocks since we don't know whether or |
| 2267 | * not they do. |
| 2268 | */ |
| 2269 | recptr = XactLogCommitRecord(committs, |
| 2270 | nchildren, children, nrels, rels, |
| 2271 | ninvalmsgs, invalmsgs, |
| 2272 | initfileinval, false, |
| 2273 | MyXactFlags | XACT_FLAGS_ACQUIREDACCESSEXCLUSIVELOCK, |
| 2274 | xid, gid); |
| 2275 | |
| 2276 | |
| 2277 | if (replorigin) |
| 2278 | /* Move LSNs forward for this replication origin */ |
| 2279 | replorigin_session_advance(replorigin_session_origin_lsn, |
| 2280 | XactLastRecEnd); |
| 2281 | |
| 2282 | /* |
| 2283 | * Record commit timestamp. The value comes from plain commit timestamp |
| 2284 | * if replorigin is not enabled, or replorigin already set a value for us |
| 2285 | * in replorigin_session_origin_timestamp otherwise. |
| 2286 | * |
| 2287 | * We don't need to WAL-log anything here, as the commit record written |
| 2288 | * above already contains the data. |
| 2289 | */ |
| 2290 | if (!replorigin || replorigin_session_origin_timestamp == 0) |
| 2291 | replorigin_session_origin_timestamp = committs; |
| 2292 | |
| 2293 | TransactionTreeSetCommitTsData(xid, nchildren, children, |
| 2294 | replorigin_session_origin_timestamp, |
| 2295 | replorigin_session_origin, false); |
| 2296 | |
| 2297 | /* |
| 2298 | * We don't currently try to sleep before flush here ... nor is there any |
| 2299 | * support for async commit of a prepared xact (the very idea is probably |
| 2300 | * a contradiction) |
| 2301 | */ |
| 2302 | |
| 2303 | /* Flush XLOG to disk */ |
| 2304 | XLogFlush(recptr); |
| 2305 | |
| 2306 | /* Mark the transaction committed in pg_xact */ |
| 2307 | TransactionIdCommitTree(xid, nchildren, children); |
| 2308 | |
| 2309 | /* Checkpoint can proceed now */ |
| 2310 | MyPgXact->delayChkpt = false; |
| 2311 | |
| 2312 | END_CRIT_SECTION(); |
| 2313 | |
| 2314 | /* |
| 2315 | * Wait for synchronous replication, if required. |
| 2316 | * |
| 2317 | * Note that at this stage we have marked clog, but still show as running |
| 2318 | * in the procarray and continue to hold locks. |
| 2319 | */ |
| 2320 | SyncRepWaitForLSN(recptr, true); |
| 2321 | } |
| 2322 | |
| 2323 | /* |
| 2324 | * RecordTransactionAbortPrepared |
| 2325 | * |
| 2326 | * This is basically the same as RecordTransactionAbort. |
| 2327 | * |
| 2328 | * We know the transaction made at least one XLOG entry (its PREPARE), |
| 2329 | * so it is never possible to optimize out the abort record. |
| 2330 | */ |
| 2331 | static void |
| 2332 | RecordTransactionAbortPrepared(TransactionId xid, |
| 2333 | int nchildren, |
| 2334 | TransactionId *children, |
| 2335 | int nrels, |
| 2336 | RelFileNode *rels, |
| 2337 | const char *gid) |
| 2338 | { |
| 2339 | XLogRecPtr recptr; |
| 2340 | |
| 2341 | /* |
| 2342 | * Catch the scenario where we aborted partway through |
| 2343 | * RecordTransactionCommitPrepared ... |
| 2344 | */ |
| 2345 | if (TransactionIdDidCommit(xid)) |
| 2346 | elog(PANIC, "cannot abort transaction %u, it was already committed" , |
| 2347 | xid); |
| 2348 | |
| 2349 | START_CRIT_SECTION(); |
| 2350 | |
| 2351 | /* |
| 2352 | * Emit the XLOG commit record. Note that we mark 2PC aborts as |
| 2353 | * potentially having AccessExclusiveLocks since we don't know whether or |
| 2354 | * not they do. |
| 2355 | */ |
| 2356 | recptr = XactLogAbortRecord(GetCurrentTimestamp(), |
| 2357 | nchildren, children, |
| 2358 | nrels, rels, |
| 2359 | MyXactFlags | XACT_FLAGS_ACQUIREDACCESSEXCLUSIVELOCK, |
| 2360 | xid, gid); |
| 2361 | |
| 2362 | /* Always flush, since we're about to remove the 2PC state file */ |
| 2363 | XLogFlush(recptr); |
| 2364 | |
| 2365 | /* |
| 2366 | * Mark the transaction aborted in clog. This is not absolutely necessary |
| 2367 | * but we may as well do it while we are here. |
| 2368 | */ |
| 2369 | TransactionIdAbortTree(xid, nchildren, children); |
| 2370 | |
| 2371 | END_CRIT_SECTION(); |
| 2372 | |
| 2373 | /* |
| 2374 | * Wait for synchronous replication, if required. |
| 2375 | * |
| 2376 | * Note that at this stage we have marked clog, but still show as running |
| 2377 | * in the procarray and continue to hold locks. |
| 2378 | */ |
| 2379 | SyncRepWaitForLSN(recptr, false); |
| 2380 | } |
| 2381 | |
| 2382 | /* |
| 2383 | * PrepareRedoAdd |
| 2384 | * |
| 2385 | * Store pointers to the start/end of the WAL record along with the xid in |
| 2386 | * a gxact entry in shared memory TwoPhaseState structure. If caller |
| 2387 | * specifies InvalidXLogRecPtr as WAL location to fetch the two-phase |
| 2388 | * data, the entry is marked as located on disk. |
| 2389 | */ |
| 2390 | void |
| 2391 | PrepareRedoAdd(char *buf, XLogRecPtr start_lsn, |
| 2392 | XLogRecPtr end_lsn, RepOriginId origin_id) |
| 2393 | { |
| 2394 | TwoPhaseFileHeader *hdr = (TwoPhaseFileHeader *) buf; |
| 2395 | char *bufptr; |
| 2396 | const char *gid; |
| 2397 | GlobalTransaction gxact; |
| 2398 | |
| 2399 | Assert(LWLockHeldByMeInMode(TwoPhaseStateLock, LW_EXCLUSIVE)); |
| 2400 | Assert(RecoveryInProgress()); |
| 2401 | |
| 2402 | bufptr = buf + MAXALIGN(sizeof(TwoPhaseFileHeader)); |
| 2403 | gid = (const char *) bufptr; |
| 2404 | |
| 2405 | /* |
| 2406 | * Reserve the GID for the given transaction in the redo code path. |
| 2407 | * |
| 2408 | * This creates a gxact struct and puts it into the active array. |
| 2409 | * |
| 2410 | * In redo, this struct is mainly used to track PREPARE/COMMIT entries in |
| 2411 | * shared memory. Hence, we only fill up the bare minimum contents here. |
| 2412 | * The gxact also gets marked with gxact->inredo set to true to indicate |
| 2413 | * that it got added in the redo phase |
| 2414 | */ |
| 2415 | |
| 2416 | /* Get a free gxact from the freelist */ |
| 2417 | if (TwoPhaseState->freeGXacts == NULL) |
| 2418 | ereport(ERROR, |
| 2419 | (errcode(ERRCODE_OUT_OF_MEMORY), |
| 2420 | errmsg("maximum number of prepared transactions reached" ), |
| 2421 | errhint("Increase max_prepared_transactions (currently %d)." , |
| 2422 | max_prepared_xacts))); |
| 2423 | gxact = TwoPhaseState->freeGXacts; |
| 2424 | TwoPhaseState->freeGXacts = gxact->next; |
| 2425 | |
| 2426 | gxact->prepared_at = hdr->prepared_at; |
| 2427 | gxact->prepare_start_lsn = start_lsn; |
| 2428 | gxact->prepare_end_lsn = end_lsn; |
| 2429 | gxact->xid = hdr->xid; |
| 2430 | gxact->owner = hdr->owner; |
| 2431 | gxact->locking_backend = InvalidBackendId; |
| 2432 | gxact->valid = false; |
| 2433 | gxact->ondisk = XLogRecPtrIsInvalid(start_lsn); |
| 2434 | gxact->inredo = true; /* yes, added in redo */ |
| 2435 | strcpy(gxact->gid, gid); |
| 2436 | |
| 2437 | /* And insert it into the active array */ |
| 2438 | Assert(TwoPhaseState->numPrepXacts < max_prepared_xacts); |
| 2439 | TwoPhaseState->prepXacts[TwoPhaseState->numPrepXacts++] = gxact; |
| 2440 | |
| 2441 | if (origin_id != InvalidRepOriginId) |
| 2442 | { |
| 2443 | /* recover apply progress */ |
| 2444 | replorigin_advance(origin_id, hdr->origin_lsn, end_lsn, |
| 2445 | false /* backward */ , false /* WAL */ ); |
| 2446 | } |
| 2447 | |
| 2448 | elog(DEBUG2, "added 2PC data in shared memory for transaction %u" , gxact->xid); |
| 2449 | } |
| 2450 | |
| 2451 | /* |
| 2452 | * PrepareRedoRemove |
| 2453 | * |
| 2454 | * Remove the corresponding gxact entry from TwoPhaseState. Also remove |
| 2455 | * the 2PC file if a prepared transaction was saved via an earlier checkpoint. |
| 2456 | * |
| 2457 | * Caller must hold TwoPhaseStateLock in exclusive mode, because TwoPhaseState |
| 2458 | * is updated. |
| 2459 | */ |
| 2460 | void |
| 2461 | PrepareRedoRemove(TransactionId xid, bool giveWarning) |
| 2462 | { |
| 2463 | GlobalTransaction gxact = NULL; |
| 2464 | int i; |
| 2465 | bool found = false; |
| 2466 | |
| 2467 | Assert(LWLockHeldByMeInMode(TwoPhaseStateLock, LW_EXCLUSIVE)); |
| 2468 | Assert(RecoveryInProgress()); |
| 2469 | |
| 2470 | for (i = 0; i < TwoPhaseState->numPrepXacts; i++) |
| 2471 | { |
| 2472 | gxact = TwoPhaseState->prepXacts[i]; |
| 2473 | |
| 2474 | if (gxact->xid == xid) |
| 2475 | { |
| 2476 | Assert(gxact->inredo); |
| 2477 | found = true; |
| 2478 | break; |
| 2479 | } |
| 2480 | } |
| 2481 | |
| 2482 | /* |
| 2483 | * Just leave if there is nothing, this is expected during WAL replay. |
| 2484 | */ |
| 2485 | if (!found) |
| 2486 | return; |
| 2487 | |
| 2488 | /* |
| 2489 | * And now we can clean up any files we may have left. |
| 2490 | */ |
| 2491 | elog(DEBUG2, "removing 2PC data for transaction %u" , xid); |
| 2492 | if (gxact->ondisk) |
| 2493 | RemoveTwoPhaseFile(xid, giveWarning); |
| 2494 | RemoveGXact(gxact); |
| 2495 | |
| 2496 | return; |
| 2497 | } |
| 2498 | |