| 1 |
/* ------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* decode.c |
| 4 |
* This module decodes WAL records read using xlogreader.h's APIs for the |
| 5 |
* purpose of logical decoding by passing information to the |
| 6 |
* reorderbuffer module (containing the actual changes) and to the |
| 7 |
* snapbuild module to build a fitting catalog snapshot (to be able to |
| 8 |
* properly decode the changes in the reorderbuffer). |
| 9 |
* |
| 10 |
* NOTE: |
| 11 |
* This basically tries to handle all low level xlog stuff for |
| 12 |
* reorderbuffer.c and snapbuild.c. There's some minor leakage where a |
| 13 |
* specific record's struct is used to pass data along, but those just |
| 14 |
* happen to contain the right amount of data in a convenient |
| 15 |
* format. There isn't and shouldn't be much intelligence about the |
| 16 |
* contents of records in here except turning them into a more usable |
| 17 |
* format. |
| 18 |
* |
| 19 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 20 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 21 |
* |
| 22 |
* IDENTIFICATION |
| 23 |
* src/backend/replication/logical/decode.c |
| 24 |
* |
| 25 |
* ------------------------------------------------------------------------- |
| 26 |
*/ |
| 27 |
#include "postgres.h" |
| 28 |
|
| 29 |
#include "access/heapam_xlog.h" |
| 30 |
#include "access/transam.h" |
| 31 |
#include "access/xact.h" |
| 32 |
#include "access/xlog_internal.h" |
| 33 |
#include "access/xlogreader.h" |
| 34 |
#include "access/xlogrecord.h" |
| 35 |
#include "catalog/pg_control.h" |
| 36 |
#include "replication/decode.h" |
| 37 |
#include "replication/logical.h" |
| 38 |
#include "replication/message.h" |
| 39 |
#include "replication/reorderbuffer.h" |
| 40 |
#include "replication/snapbuild.h" |
| 41 |
#include "storage/standbydefs.h" |
| 42 |
|
| 43 |
/* individual record(group)'s handlers */ |
| 44 |
static void DecodeInsert(LogicalDecodingContext *ctx, XLogRecordBuffer *buf); |
| 45 |
static void DecodeUpdate(LogicalDecodingContext *ctx, XLogRecordBuffer *buf); |
| 46 |
static void DecodeDelete(LogicalDecodingContext *ctx, XLogRecordBuffer *buf); |
| 47 |
static void DecodeTruncate(LogicalDecodingContext *ctx, XLogRecordBuffer *buf); |
| 48 |
static void DecodeMultiInsert(LogicalDecodingContext *ctx, XLogRecordBuffer *buf); |
| 49 |
static void DecodeSpecConfirm(LogicalDecodingContext *ctx, XLogRecordBuffer *buf); |
| 50 |
|
| 51 |
static void DecodeCommit(LogicalDecodingContext *ctx, XLogRecordBuffer *buf, |
| 52 |
xl_xact_parsed_commit *parsed, TransactionId xid, |
| 53 |
bool two_phase); |
| 54 |
static void DecodeAbort(LogicalDecodingContext *ctx, XLogRecordBuffer *buf, |
| 55 |
xl_xact_parsed_abort *parsed, TransactionId xid, |
| 56 |
bool two_phase); |
| 57 |
static void DecodePrepare(LogicalDecodingContext *ctx, XLogRecordBuffer *buf, |
| 58 |
xl_xact_parsed_prepare *parsed); |
| 59 |
|
| 60 |
|
| 61 |
/* common function to decode tuples */ |
| 62 |
static void DecodeXLogTuple(char *data, Size len, HeapTuple tuple); |
| 63 |
|
| 64 |
/* helper functions for decoding transactions */ |
| 65 |
static inline bool FilterPrepare(LogicalDecodingContext *ctx, |
| 66 |
TransactionId xid, const char *gid); |
| 67 |
static bool DecodeTXNNeedSkip(LogicalDecodingContext *ctx, |
| 68 |
XLogRecordBuffer *buf, Oid txn_dbid, |
| 69 |
RepOriginId origin_id); |
| 70 |
|
| 71 |
/* |
| 72 |
* Take every XLogReadRecord()ed record and perform the actions required to |
| 73 |
* decode it using the output plugin already setup in the logical decoding |
| 74 |
* context. |
| 75 |
* |
| 76 |
* NB: Note that every record's xid needs to be processed by reorderbuffer |
| 77 |
* (xids contained in the content of records are not relevant for this rule). |
| 78 |
* That means that for records which'd otherwise not go through the |
| 79 |
* reorderbuffer ReorderBufferProcessXid() has to be called. We don't want to |
| 80 |
* call ReorderBufferProcessXid for each record type by default, because |
| 81 |
* e.g. empty xacts can be handled more efficiently if there's no previous |
| 82 |
* state for them. |
| 83 |
* |
| 84 |
* We also support the ability to fast forward thru records, skipping some |
| 85 |
* record types completely - see individual record types for details. |
| 86 |
*/ |
| 87 |
void |
| 88 |
LogicalDecodingProcessRecord(LogicalDecodingContext *ctx, XLogReaderState *record) |
| 89 |
{ |
| 90 |
XLogRecordBuffer buf; |
| 91 |
TransactionId txid; |
| 92 |
RmgrData rmgr; |
| 93 |
|
| 94 |
buf.origptr = ctx->reader->ReadRecPtr; |
| 95 |
buf.endptr = ctx->reader->EndRecPtr; |
| 96 |
buf.record = record; |
| 97 |
|
| 98 |
txid = XLogRecGetTopXid(record); |
| 99 |
|
| 100 |
/* |
| 101 |
* If the top-level xid is valid, we need to assign the subxact to the |
| 102 |
* top-level xact. We need to do this for all records, hence we do it |
| 103 |
* before the switch. |
| 104 |
*/ |
| 105 |
if (TransactionIdIsValid(txid)) |
| 106 |
{ |
| 107 |
ReorderBufferAssignChild(ctx->reorder, |
| 108 |
txid, |
| 109 |
XLogRecGetXid(record), |
| 110 |
buf.origptr); |
| 111 |
} |
| 112 |
|
| 113 |
rmgr = GetRmgr(XLogRecGetRmid(record)); |
| 114 |
|
| 115 |
if (rmgr.rm_decode != NULL) |
| 116 |
rmgr.rm_decode(ctx, &buf); |
| 117 |
else |
| 118 |
{ |
| 119 |
/* just deal with xid, and done */ |
| 120 |
ReorderBufferProcessXid(ctx->reorder, XLogRecGetXid(record), |
| 121 |
buf.origptr); |
| 122 |
} |
| 123 |
} |
| 124 |
|
| 125 |
/* |
| 126 |
* Handle rmgr XLOG_ID records for LogicalDecodingProcessRecord(). |
| 127 |
*/ |
| 128 |
void |
| 129 |
xlog_decode(LogicalDecodingContext *ctx, XLogRecordBuffer *buf) |
| 130 |
{ |
| 131 |
SnapBuild *builder = ctx->snapshot_builder; |
| 132 |
uint8 info = XLogRecGetInfo(buf->record) & ~XLR_INFO_MASK; |
| 133 |
|
| 134 |
ReorderBufferProcessXid(ctx->reorder, XLogRecGetXid(buf->record), |
| 135 |
buf->origptr); |
| 136 |
|
| 137 |
switch (info) |
| 138 |
{ |
| 139 |
/* this is also used in END_OF_RECOVERY checkpoints */ |
| 140 |
case XLOG_CHECKPOINT_SHUTDOWN: |
| 141 |
case XLOG_END_OF_RECOVERY: |
| 142 |
SnapBuildSerializationPoint(builder, buf->origptr); |
| 143 |
|
| 144 |
break; |
| 145 |
case XLOG_CHECKPOINT_ONLINE: |
| 146 |
|
| 147 |
/* |
| 148 |
* a RUNNING_XACTS record will have been logged near to this, we |
| 149 |
* can restart from there. |
| 150 |
*/ |
| 151 |
break; |
| 152 |
case XLOG_PARAMETER_CHANGE: |
| 153 |
{ |
| 154 |
xl_parameter_change *xlrec = |
| 155 |
(xl_parameter_change *) XLogRecGetData(buf->record); |
| 156 |
|
| 157 |
/* |
| 158 |
* If wal_level on the primary is reduced to less than |
| 159 |
* logical, we want to prevent existing logical slots from |
| 160 |
* being used. Existing logical slots on the standby get |
| 161 |
* invalidated when this WAL record is replayed; and further, |
| 162 |
* slot creation fails when wal_level is not sufficient; but |
| 163 |
* all these operations are not synchronized, so a logical |
| 164 |
* slot may creep in while the wal_level is being reduced. |
| 165 |
* Hence this extra check. |
| 166 |
*/ |
| 167 |
if (xlrec->wal_level < WAL_LEVEL_LOGICAL) |
| 168 |
{ |
| 169 |
/* |
| 170 |
* This can occur only on a standby, as a primary would |
| 171 |
* not allow to restart after changing wal_level < logical |
| 172 |
* if there is pre-existing logical slot. |
| 173 |
*/ |
| 174 |
Assert(RecoveryInProgress()); |
| 175 |
ereport(ERROR, |
| 176 |
(errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE), |
| 177 |
errmsg("logical decoding on standby requires \"wal_level\" >= \"logical\" on the primary"))); |
| 178 |
} |
| 179 |
break; |
| 180 |
} |
| 181 |
case XLOG_NOOP: |
| 182 |
case XLOG_NEXTOID: |
| 183 |
case XLOG_SWITCH: |
| 184 |
case XLOG_BACKUP_END: |
| 185 |
case XLOG_RESTORE_POINT: |
| 186 |
case XLOG_FPW_CHANGE: |
| 187 |
case XLOG_FPI_FOR_HINT: |
| 188 |
case XLOG_FPI: |
| 189 |
case XLOG_OVERWRITE_CONTRECORD: |
| 190 |
case XLOG_CHECKPOINT_REDO: |
| 191 |
break; |
| 192 |
default: |
| 193 |
elog(ERROR, "unexpected RM_XLOG_ID record type: %u", info); |
| 194 |
} |
| 195 |
} |
| 196 |
|
| 197 |
/* |
| 198 |
* Handle rmgr XACT_ID records for LogicalDecodingProcessRecord(). |
| 199 |
*/ |
| 200 |
void |
| 201 |
xact_decode(LogicalDecodingContext *ctx, XLogRecordBuffer *buf) |
| 202 |
{ |
| 203 |
SnapBuild *builder = ctx->snapshot_builder; |
| 204 |
ReorderBuffer *reorder = ctx->reorder; |
| 205 |
XLogReaderState *r = buf->record; |
| 206 |
uint8 info = XLogRecGetInfo(r) & XLOG_XACT_OPMASK; |
| 207 |
|
| 208 |
/* |
| 209 |
* If the snapshot isn't yet fully built, we cannot decode anything, so |
| 210 |
* bail out. |
| 211 |
*/ |
| 212 |
if (SnapBuildCurrentState(builder) < SNAPBUILD_FULL_SNAPSHOT) |
| 213 |
return; |
| 214 |
|
| 215 |
switch (info) |
| 216 |
{ |
| 217 |
case XLOG_XACT_COMMIT: |
| 218 |
case XLOG_XACT_COMMIT_PREPARED: |
| 219 |
{ |
| 220 |
xl_xact_commit *xlrec; |
| 221 |
xl_xact_parsed_commit parsed; |
| 222 |
TransactionId xid; |
| 223 |
bool two_phase = false; |
| 224 |
|
| 225 |
xlrec = (xl_xact_commit *) XLogRecGetData(r); |
| 226 |
ParseCommitRecord(XLogRecGetInfo(buf->record), xlrec, &parsed); |
| 227 |
|
| 228 |
if (!TransactionIdIsValid(parsed.twophase_xid)) |
| 229 |
xid = XLogRecGetXid(r); |
| 230 |
else |
| 231 |
xid = parsed.twophase_xid; |
| 232 |
|
| 233 |
/* |
| 234 |
* We would like to process the transaction in a two-phase |
| 235 |
* manner iff output plugin supports two-phase commits and |
| 236 |
* doesn't filter the transaction at prepare time. |
| 237 |
*/ |
| 238 |
if (info == XLOG_XACT_COMMIT_PREPARED) |
| 239 |
two_phase = !(FilterPrepare(ctx, xid, |
| 240 |
parsed.twophase_gid)); |
| 241 |
|
| 242 |
DecodeCommit(ctx, buf, &parsed, xid, two_phase); |
| 243 |
break; |
| 244 |
} |
| 245 |
case XLOG_XACT_ABORT: |
| 246 |
case XLOG_XACT_ABORT_PREPARED: |
| 247 |
{ |
| 248 |
xl_xact_abort *xlrec; |
| 249 |
xl_xact_parsed_abort parsed; |
| 250 |
TransactionId xid; |
| 251 |
bool two_phase = false; |
| 252 |
|
| 253 |
xlrec = (xl_xact_abort *) XLogRecGetData(r); |
| 254 |
ParseAbortRecord(XLogRecGetInfo(buf->record), xlrec, &parsed); |
| 255 |
|
| 256 |
if (!TransactionIdIsValid(parsed.twophase_xid)) |
| 257 |
xid = XLogRecGetXid(r); |
| 258 |
else |
| 259 |
xid = parsed.twophase_xid; |
| 260 |
|
| 261 |
/* |
| 262 |
* We would like to process the transaction in a two-phase |
| 263 |
* manner iff output plugin supports two-phase commits and |
| 264 |
* doesn't filter the transaction at prepare time. |
| 265 |
*/ |
| 266 |
if (info == XLOG_XACT_ABORT_PREPARED) |
| 267 |
two_phase = !(FilterPrepare(ctx, xid, |
| 268 |
parsed.twophase_gid)); |
| 269 |
|
| 270 |
DecodeAbort(ctx, buf, &parsed, xid, two_phase); |
| 271 |
break; |
| 272 |
} |
| 273 |
case XLOG_XACT_ASSIGNMENT: |
| 274 |
|
| 275 |
/* |
| 276 |
* We assign subxact to the toplevel xact while processing each |
| 277 |
* record if required. So, we don't need to do anything here. See |
| 278 |
* LogicalDecodingProcessRecord. |
| 279 |
*/ |
| 280 |
break; |
| 281 |
case XLOG_XACT_INVALIDATIONS: |
| 282 |
{ |
| 283 |
TransactionId xid; |
| 284 |
xl_xact_invals *invals; |
| 285 |
|
| 286 |
xid = XLogRecGetXid(r); |
| 287 |
invals = (xl_xact_invals *) XLogRecGetData(r); |
| 288 |
|
| 289 |
/* |
| 290 |
* Execute the invalidations for xid-less transactions, |
| 291 |
* otherwise, accumulate them so that they can be processed at |
| 292 |
* the commit time. |
| 293 |
*/ |
| 294 |
if (TransactionIdIsValid(xid)) |
| 295 |
{ |
| 296 |
if (!ctx->fast_forward) |
| 297 |
ReorderBufferAddInvalidations(reorder, xid, |
| 298 |
buf->origptr, |
| 299 |
invals->nmsgs, |
| 300 |
invals->msgs); |
| 301 |
ReorderBufferXidSetCatalogChanges(ctx->reorder, xid, |
| 302 |
buf->origptr); |
| 303 |
} |
| 304 |
else if (!ctx->fast_forward) |
| 305 |
ReorderBufferImmediateInvalidation(ctx->reorder, |
| 306 |
invals->nmsgs, |
| 307 |
invals->msgs); |
| 308 |
|
| 309 |
break; |
| 310 |
} |
| 311 |
case XLOG_XACT_PREPARE: |
| 312 |
{ |
| 313 |
xl_xact_parsed_prepare parsed; |
| 314 |
xl_xact_prepare *xlrec; |
| 315 |
|
| 316 |
/* ok, parse it */ |
| 317 |
xlrec = (xl_xact_prepare *) XLogRecGetData(r); |
| 318 |
ParsePrepareRecord(XLogRecGetInfo(buf->record), |
| 319 |
xlrec, &parsed); |
| 320 |
|
| 321 |
/* |
| 322 |
* We would like to process the transaction in a two-phase |
| 323 |
* manner iff output plugin supports two-phase commits and |
| 324 |
* doesn't filter the transaction at prepare time. |
| 325 |
*/ |
| 326 |
if (FilterPrepare(ctx, parsed.twophase_xid, |
| 327 |
parsed.twophase_gid)) |
| 328 |
{ |
| 329 |
ReorderBufferProcessXid(reorder, parsed.twophase_xid, |
| 330 |
buf->origptr); |
| 331 |
break; |
| 332 |
} |
| 333 |
|
| 334 |
/* |
| 335 |
* Note that if the prepared transaction has locked [user] |
| 336 |
* catalog tables exclusively then decoding prepare can block |
| 337 |
* till the main transaction is committed because it needs to |
| 338 |
* lock the catalog tables. |
| 339 |
* |
| 340 |
* XXX Now, this can even lead to a deadlock if the prepare |
| 341 |
* transaction is waiting to get it logically replicated for |
| 342 |
* distributed 2PC. This can be avoided by disallowing |
| 343 |
* preparing transactions that have locked [user] catalog |
| 344 |
* tables exclusively but as of now, we ask users not to do |
| 345 |
* such an operation. |
| 346 |
*/ |
| 347 |
DecodePrepare(ctx, buf, &parsed); |
| 348 |
break; |
| 349 |
} |
| 350 |
default: |
| 351 |
elog(ERROR, "unexpected RM_XACT_ID record type: %u", info); |
| 352 |
} |
| 353 |
} |
| 354 |
|
| 355 |
/* |
| 356 |
* Handle rmgr STANDBY_ID records for LogicalDecodingProcessRecord(). |
| 357 |
*/ |
| 358 |
void |
| 359 |
standby_decode(LogicalDecodingContext *ctx, XLogRecordBuffer *buf) |
| 360 |
{ |
| 361 |
SnapBuild *builder = ctx->snapshot_builder; |
| 362 |
XLogReaderState *r = buf->record; |
| 363 |
uint8 info = XLogRecGetInfo(r) & ~XLR_INFO_MASK; |
| 364 |
|
| 365 |
ReorderBufferProcessXid(ctx->reorder, XLogRecGetXid(r), buf->origptr); |
| 366 |
|
| 367 |
switch (info) |
| 368 |
{ |
| 369 |
case XLOG_RUNNING_XACTS: |
| 370 |
{ |
| 371 |
xl_running_xacts *running = (xl_running_xacts *) XLogRecGetData(r); |
| 372 |
|
| 373 |
SnapBuildProcessRunningXacts(builder, buf->origptr, running); |
| 374 |
|
| 375 |
/* |
| 376 |
* Abort all transactions that we keep track of, that are |
| 377 |
* older than the record's oldestRunningXid. This is the most |
| 378 |
* convenient spot for doing so since, in contrast to shutdown |
| 379 |
* or end-of-recovery checkpoints, we have information about |
| 380 |
* all running transactions which includes prepared ones, |
| 381 |
* while shutdown checkpoints just know that no non-prepared |
| 382 |
* transactions are in progress. |
| 383 |
*/ |
| 384 |
ReorderBufferAbortOld(ctx->reorder, running->oldestRunningXid); |
| 385 |
} |
| 386 |
break; |
| 387 |
case XLOG_STANDBY_LOCK: |
| 388 |
break; |
| 389 |
case XLOG_INVALIDATIONS: |
| 390 |
|
| 391 |
/* |
| 392 |
* We are processing the invalidations at the command level via |
| 393 |
* XLOG_XACT_INVALIDATIONS. So we don't need to do anything here. |
| 394 |
*/ |
| 395 |
break; |
| 396 |
default: |
| 397 |
elog(ERROR, "unexpected RM_STANDBY_ID record type: %u", info); |
| 398 |
} |
| 399 |
} |
| 400 |
|
| 401 |
/* |
| 402 |
* Handle rmgr HEAP2_ID records for LogicalDecodingProcessRecord(). |
| 403 |
*/ |
| 404 |
void |
| 405 |
heap2_decode(LogicalDecodingContext *ctx, XLogRecordBuffer *buf) |
| 406 |
{ |
| 407 |
uint8 info = XLogRecGetInfo(buf->record) & XLOG_HEAP_OPMASK; |
| 408 |
TransactionId xid = XLogRecGetXid(buf->record); |
| 409 |
SnapBuild *builder = ctx->snapshot_builder; |
| 410 |
|
| 411 |
ReorderBufferProcessXid(ctx->reorder, xid, buf->origptr); |
| 412 |
|
| 413 |
/* |
| 414 |
* If we don't have snapshot or we are just fast-forwarding, there is no |
| 415 |
* point in decoding data changes. However, it's crucial to build the base |
| 416 |
* snapshot during fast-forward mode (as is done in |
| 417 |
* SnapBuildProcessChange()) because we require the snapshot's xmin when |
| 418 |
* determining the candidate catalog_xmin for the replication slot. See |
| 419 |
* SnapBuildProcessRunningXacts(). |
| 420 |
*/ |
| 421 |
if (SnapBuildCurrentState(builder) < SNAPBUILD_FULL_SNAPSHOT) |
| 422 |
return; |
| 423 |
|
| 424 |
switch (info) |
| 425 |
{ |
| 426 |
case XLOG_HEAP2_MULTI_INSERT: |
| 427 |
if (SnapBuildProcessChange(builder, xid, buf->origptr) && |
| 428 |
!ctx->fast_forward) |
| 429 |
DecodeMultiInsert(ctx, buf); |
| 430 |
break; |
| 431 |
case XLOG_HEAP2_NEW_CID: |
| 432 |
if (!ctx->fast_forward) |
| 433 |
{ |
| 434 |
xl_heap_new_cid *xlrec; |
| 435 |
|
| 436 |
xlrec = (xl_heap_new_cid *) XLogRecGetData(buf->record); |
| 437 |
SnapBuildProcessNewCid(builder, xid, buf->origptr, xlrec); |
| 438 |
|
| 439 |
break; |
| 440 |
} |
| 441 |
case XLOG_HEAP2_REWRITE: |
| 442 |
|
| 443 |
/* |
| 444 |
* Although these records only exist to serve the needs of logical |
| 445 |
* decoding, all the work happens as part of crash or archive |
| 446 |
* recovery, so we don't need to do anything here. |
| 447 |
*/ |
| 448 |
break; |
| 449 |
|
| 450 |
/* |
| 451 |
* Everything else here is just low level physical stuff we're not |
| 452 |
* interested in. |
| 453 |
*/ |
| 454 |
case XLOG_HEAP2_PRUNE_ON_ACCESS: |
| 455 |
case XLOG_HEAP2_PRUNE_VACUUM_SCAN: |
| 456 |
case XLOG_HEAP2_PRUNE_VACUUM_CLEANUP: |
| 457 |
case XLOG_HEAP2_VISIBLE: |
| 458 |
case XLOG_HEAP2_LOCK_UPDATED: |
| 459 |
break; |
| 460 |
default: |
| 461 |
elog(ERROR, "unexpected RM_HEAP2_ID record type: %u", info); |
| 462 |
} |
| 463 |
} |
| 464 |
|
| 465 |
/* |
| 466 |
* Handle rmgr HEAP_ID records for LogicalDecodingProcessRecord(). |
| 467 |
*/ |
| 468 |
void |
| 469 |
heap_decode(LogicalDecodingContext *ctx, XLogRecordBuffer *buf) |
| 470 |
{ |
| 471 |
uint8 info = XLogRecGetInfo(buf->record) & XLOG_HEAP_OPMASK; |
| 472 |
TransactionId xid = XLogRecGetXid(buf->record); |
| 473 |
SnapBuild *builder = ctx->snapshot_builder; |
| 474 |
|
| 475 |
ReorderBufferProcessXid(ctx->reorder, xid, buf->origptr); |
| 476 |
|
| 477 |
/* |
| 478 |
* If we don't have snapshot or we are just fast-forwarding, there is no |
| 479 |
* point in decoding data changes. However, it's crucial to build the base |
| 480 |
* snapshot during fast-forward mode (as is done in |
| 481 |
* SnapBuildProcessChange()) because we require the snapshot's xmin when |
| 482 |
* determining the candidate catalog_xmin for the replication slot. See |
| 483 |
* SnapBuildProcessRunningXacts(). |
| 484 |
*/ |
| 485 |
if (SnapBuildCurrentState(builder) < SNAPBUILD_FULL_SNAPSHOT) |
| 486 |
return; |
| 487 |
|
| 488 |
switch (info) |
| 489 |
{ |
| 490 |
case XLOG_HEAP_INSERT: |
| 491 |
if (SnapBuildProcessChange(builder, xid, buf->origptr) && |
| 492 |
!ctx->fast_forward) |
| 493 |
DecodeInsert(ctx, buf); |
| 494 |
break; |
| 495 |
|
| 496 |
/* |
| 497 |
* Treat HOT update as normal updates. There is no useful |
| 498 |
* information in the fact that we could make it a HOT update |
| 499 |
* locally and the WAL layout is compatible. |
| 500 |
*/ |
| 501 |
case XLOG_HEAP_HOT_UPDATE: |
| 502 |
case XLOG_HEAP_UPDATE: |
| 503 |
if (SnapBuildProcessChange(builder, xid, buf->origptr) && |
| 504 |
!ctx->fast_forward) |
| 505 |
DecodeUpdate(ctx, buf); |
| 506 |
break; |
| 507 |
|
| 508 |
case XLOG_HEAP_DELETE: |
| 509 |
if (SnapBuildProcessChange(builder, xid, buf->origptr) && |
| 510 |
!ctx->fast_forward) |
| 511 |
DecodeDelete(ctx, buf); |
| 512 |
break; |
| 513 |
|
| 514 |
case XLOG_HEAP_TRUNCATE: |
| 515 |
if (SnapBuildProcessChange(builder, xid, buf->origptr) && |
| 516 |
!ctx->fast_forward) |
| 517 |
DecodeTruncate(ctx, buf); |
| 518 |
break; |
| 519 |
|
| 520 |
case XLOG_HEAP_INPLACE: |
| 521 |
|
| 522 |
/* |
| 523 |
* Inplace updates are only ever performed on catalog tuples and |
| 524 |
* can, per definition, not change tuple visibility. Since we |
| 525 |
* also don't decode catalog tuples, we're not interested in the |
| 526 |
* record's contents. |
| 527 |
*/ |
| 528 |
break; |
| 529 |
|
| 530 |
case XLOG_HEAP_CONFIRM: |
| 531 |
if (SnapBuildProcessChange(builder, xid, buf->origptr) && |
| 532 |
!ctx->fast_forward) |
| 533 |
DecodeSpecConfirm(ctx, buf); |
| 534 |
break; |
| 535 |
|
| 536 |
case XLOG_HEAP_LOCK: |
| 537 |
/* we don't care about row level locks for now */ |
| 538 |
break; |
| 539 |
|
| 540 |
default: |
| 541 |
elog(ERROR, "unexpected RM_HEAP_ID record type: %u", info); |
| 542 |
break; |
| 543 |
} |
| 544 |
} |
| 545 |
|
| 546 |
/* |
| 547 |
* Ask output plugin whether we want to skip this PREPARE and send |
| 548 |
* this transaction as a regular commit later. |
| 549 |
*/ |
| 550 |
static inline bool |
| 551 |
FilterPrepare(LogicalDecodingContext *ctx, TransactionId xid, |
| 552 |
const char *gid) |
| 553 |
{ |
| 554 |
/* |
| 555 |
* Skip if decoding of two-phase transactions at PREPARE time is not |
| 556 |
* enabled. In that case, all two-phase transactions are considered |
| 557 |
* filtered out and will be applied as regular transactions at COMMIT |
| 558 |
* PREPARED. |
| 559 |
*/ |
| 560 |
if (!ctx->twophase) |
| 561 |
return true; |
| 562 |
|
| 563 |
/* |
| 564 |
* The filter_prepare callback is optional. When not supplied, all |
| 565 |
* prepared transactions should go through. |
| 566 |
*/ |
| 567 |
if (ctx->callbacks.filter_prepare_cb == NULL) |
| 568 |
return false; |
| 569 |
|
| 570 |
return filter_prepare_cb_wrapper(ctx, xid, gid); |
| 571 |
} |
| 572 |
|
| 573 |
static inline bool |
| 574 |
FilterByOrigin(LogicalDecodingContext *ctx, RepOriginId origin_id) |
| 575 |
{ |
| 576 |
if (ctx->callbacks.filter_by_origin_cb == NULL) |
| 577 |
return false; |
| 578 |
|
| 579 |
return filter_by_origin_cb_wrapper(ctx, origin_id); |
| 580 |
} |
| 581 |
|
| 582 |
/* |
| 583 |
* Handle rmgr LOGICALMSG_ID records for LogicalDecodingProcessRecord(). |
| 584 |
*/ |
| 585 |
void |
| 586 |
logicalmsg_decode(LogicalDecodingContext *ctx, XLogRecordBuffer *buf) |
| 587 |
{ |
| 588 |
SnapBuild *builder = ctx->snapshot_builder; |
| 589 |
XLogReaderState *r = buf->record; |
| 590 |
TransactionId xid = XLogRecGetXid(r); |
| 591 |
uint8 info = XLogRecGetInfo(r) & ~XLR_INFO_MASK; |
| 592 |
RepOriginId origin_id = XLogRecGetOrigin(r); |
| 593 |
Snapshot snapshot = NULL; |
| 594 |
xl_logical_message *message; |
| 595 |
|
| 596 |
if (info != XLOG_LOGICAL_MESSAGE) |
| 597 |
elog(ERROR, "unexpected RM_LOGICALMSG_ID record type: %u", info); |
| 598 |
|
| 599 |
ReorderBufferProcessXid(ctx->reorder, XLogRecGetXid(r), buf->origptr); |
| 600 |
|
| 601 |
/* If we don't have snapshot, there is no point in decoding messages */ |
| 602 |
if (SnapBuildCurrentState(builder) < SNAPBUILD_FULL_SNAPSHOT) |
| 603 |
return; |
| 604 |
|
| 605 |
message = (xl_logical_message *) XLogRecGetData(r); |
| 606 |
|
| 607 |
if (message->dbId != ctx->slot->data.database || |
| 608 |
FilterByOrigin(ctx, origin_id)) |
| 609 |
return; |
| 610 |
|
| 611 |
if (message->transactional && |
| 612 |
!SnapBuildProcessChange(builder, xid, buf->origptr)) |
| 613 |
return; |
| 614 |
else if (!message->transactional && |
| 615 |
(SnapBuildCurrentState(builder) != SNAPBUILD_CONSISTENT || |
| 616 |
SnapBuildXactNeedsSkip(builder, buf->origptr))) |
| 617 |
return; |
| 618 |
|
| 619 |
/* |
| 620 |
* We also skip decoding in fast_forward mode. This check must be last |
| 621 |
* because we don't want to set the processing_required flag unless we |
| 622 |
* have a decodable message. |
| 623 |
*/ |
| 624 |
if (ctx->fast_forward) |
| 625 |
{ |
| 626 |
/* |
| 627 |
* We need to set processing_required flag to notify the message's |
| 628 |
* existence to the caller. Usually, the flag is set when either the |
| 629 |
* COMMIT or ABORT records are decoded, but this must be turned on |
| 630 |
* here because the non-transactional logical message is decoded |
| 631 |
* without waiting for these records. |
| 632 |
*/ |
| 633 |
if (!message->transactional) |
| 634 |
ctx->processing_required = true; |
| 635 |
|
| 636 |
return; |
| 637 |
} |
| 638 |
|
| 639 |
/* |
| 640 |
* If this is a non-transactional change, get the snapshot we're expected |
| 641 |
* to use. We only get here when the snapshot is consistent, and the |
| 642 |
* change is not meant to be skipped. |
| 643 |
* |
| 644 |
* For transactional changes we don't need a snapshot, we'll use the |
| 645 |
* regular snapshot maintained by ReorderBuffer. We just leave it NULL. |
| 646 |
*/ |
| 647 |
if (!message->transactional) |
| 648 |
snapshot = SnapBuildGetOrBuildSnapshot(builder); |
| 649 |
|
| 650 |
ReorderBufferQueueMessage(ctx->reorder, xid, snapshot, buf->endptr, |
| 651 |
message->transactional, |
| 652 |
message->message, /* first part of message is |
| 653 |
* prefix */ |
| 654 |
message->message_size, |
| 655 |
message->message + message->prefix_size); |
| 656 |
} |
| 657 |
|
| 658 |
/* |
| 659 |
* Consolidated commit record handling between the different form of commit |
| 660 |
* records. |
| 661 |
* |
| 662 |
* 'two_phase' indicates that caller wants to process the transaction in two |
| 663 |
* phases, first process prepare if not already done and then process |
| 664 |
* commit_prepared. |
| 665 |
*/ |
| 666 |
static void |
| 667 |
DecodeCommit(LogicalDecodingContext *ctx, XLogRecordBuffer *buf, |
| 668 |
xl_xact_parsed_commit *parsed, TransactionId xid, |
| 669 |
bool two_phase) |
| 670 |
{ |
| 671 |
XLogRecPtr origin_lsn = InvalidXLogRecPtr; |
| 672 |
TimestampTz commit_time = parsed->xact_time; |
| 673 |
RepOriginId origin_id = XLogRecGetOrigin(buf->record); |
| 674 |
int i; |
| 675 |
|
| 676 |
if (parsed->xinfo & XACT_XINFO_HAS_ORIGIN) |
| 677 |
{ |
| 678 |
origin_lsn = parsed->origin_lsn; |
| 679 |
commit_time = parsed->origin_timestamp; |
| 680 |
} |
| 681 |
|
| 682 |
SnapBuildCommitTxn(ctx->snapshot_builder, buf->origptr, xid, |
| 683 |
parsed->nsubxacts, parsed->subxacts, |
| 684 |
parsed->xinfo); |
| 685 |
|
| 686 |
/* ---- |
| 687 |
* Check whether we are interested in this specific transaction, and tell |
| 688 |
* the reorderbuffer to forget the content of the (sub-)transactions |
| 689 |
* if not. |
| 690 |
* |
| 691 |
* We can't just use ReorderBufferAbort() here, because we need to execute |
| 692 |
* the transaction's invalidations. This currently won't be needed if |
| 693 |
* we're just skipping over the transaction because currently we only do |
| 694 |
* so during startup, to get to the first transaction the client needs. As |
| 695 |
* we have reset the catalog caches before starting to read WAL, and we |
| 696 |
* haven't yet touched any catalogs, there can't be anything to invalidate. |
| 697 |
* But if we're "forgetting" this commit because it happened in another |
| 698 |
* database, the invalidations might be important, because they could be |
| 699 |
* for shared catalogs and we might have loaded data into the relevant |
| 700 |
* syscaches. |
| 701 |
* --- |
| 702 |
*/ |
| 703 |
if (DecodeTXNNeedSkip(ctx, buf, parsed->dbId, origin_id)) |
| 704 |
{ |
| 705 |
for (i = 0; i < parsed->nsubxacts; i++) |
| 706 |
{ |
| 707 |
ReorderBufferForget(ctx->reorder, parsed->subxacts[i], buf->origptr); |
| 708 |
} |
| 709 |
ReorderBufferForget(ctx->reorder, xid, buf->origptr); |
| 710 |
|
| 711 |
return; |
| 712 |
} |
| 713 |
|
| 714 |
/* tell the reorderbuffer about the surviving subtransactions */ |
| 715 |
for (i = 0; i < parsed->nsubxacts; i++) |
| 716 |
{ |
| 717 |
ReorderBufferCommitChild(ctx->reorder, xid, parsed->subxacts[i], |
| 718 |
buf->origptr, buf->endptr); |
| 719 |
} |
| 720 |
|
| 721 |
/* |
| 722 |
* Send the final commit record if the transaction data is already |
| 723 |
* decoded, otherwise, process the entire transaction. |
| 724 |
*/ |
| 725 |
if (two_phase) |
| 726 |
{ |
| 727 |
ReorderBufferFinishPrepared(ctx->reorder, xid, buf->origptr, buf->endptr, |
| 728 |
SnapBuildGetTwoPhaseAt(ctx->snapshot_builder), |
| 729 |
commit_time, origin_id, origin_lsn, |
| 730 |
parsed->twophase_gid, true); |
| 731 |
} |
| 732 |
else |
| 733 |
{ |
| 734 |
ReorderBufferCommit(ctx->reorder, xid, buf->origptr, buf->endptr, |
| 735 |
commit_time, origin_id, origin_lsn); |
| 736 |
} |
| 737 |
|
| 738 |
/* |
| 739 |
* Update the decoding stats at transaction prepare/commit/abort. |
| 740 |
* Additionally we send the stats when we spill or stream the changes to |
| 741 |
* avoid losing them in case the decoding is interrupted. It is not clear |
| 742 |
* that sending more or less frequently than this would be better. |
| 743 |
*/ |
| 744 |
UpdateDecodingStats(ctx); |
| 745 |
} |
| 746 |
|
| 747 |
/* |
| 748 |
* Decode PREPARE record. Similar logic as in DecodeCommit. |
| 749 |
* |
| 750 |
* Note that we don't skip prepare even if have detected concurrent abort |
| 751 |
* because it is quite possible that we had already sent some changes before we |
| 752 |
* detect abort in which case we need to abort those changes in the subscriber. |
| 753 |
* To abort such changes, we do send the prepare and then the rollback prepared |
| 754 |
* which is what happened on the publisher-side as well. Now, we can invent a |
| 755 |
* new abort API wherein in such cases we send abort and skip sending prepared |
| 756 |
* and rollback prepared but then it is not that straightforward because we |
| 757 |
* might have streamed this transaction by that time in which case it is |
| 758 |
* handled when the rollback is encountered. It is not impossible to optimize |
| 759 |
* the concurrent abort case but it can introduce design complexity w.r.t |
| 760 |
* handling different cases so leaving it for now as it doesn't seem worth it. |
| 761 |
*/ |
| 762 |
static void |
| 763 |
DecodePrepare(LogicalDecodingContext *ctx, XLogRecordBuffer *buf, |
| 764 |
xl_xact_parsed_prepare *parsed) |
| 765 |
{ |
| 766 |
SnapBuild *builder = ctx->snapshot_builder; |
| 767 |
XLogRecPtr origin_lsn = parsed->origin_lsn; |
| 768 |
TimestampTz prepare_time = parsed->xact_time; |
| 769 |
RepOriginId origin_id = XLogRecGetOrigin(buf->record); |
| 770 |
int i; |
| 771 |
TransactionId xid = parsed->twophase_xid; |
| 772 |
|
| 773 |
if (parsed->origin_timestamp != 0) |
| 774 |
prepare_time = parsed->origin_timestamp; |
| 775 |
|
| 776 |
/* |
| 777 |
* Remember the prepare info for a txn so that it can be used later in |
| 778 |
* commit prepared if required. See ReorderBufferFinishPrepared. |
| 779 |
*/ |
| 780 |
if (!ReorderBufferRememberPrepareInfo(ctx->reorder, xid, buf->origptr, |
| 781 |
buf->endptr, prepare_time, origin_id, |
| 782 |
origin_lsn)) |
| 783 |
return; |
| 784 |
|
| 785 |
/* We can't start streaming unless a consistent state is reached. */ |
| 786 |
if (SnapBuildCurrentState(builder) < SNAPBUILD_CONSISTENT) |
| 787 |
{ |
| 788 |
ReorderBufferSkipPrepare(ctx->reorder, xid); |
| 789 |
return; |
| 790 |
} |
| 791 |
|
| 792 |
/* |
| 793 |
* Check whether we need to process this transaction. See |
| 794 |
* DecodeTXNNeedSkip for the reasons why we sometimes want to skip the |
| 795 |
* transaction. |
| 796 |
* |
| 797 |
* We can't call ReorderBufferForget as we did in DecodeCommit as the txn |
| 798 |
* hasn't yet been committed, removing this txn before a commit might |
| 799 |
* result in the computation of an incorrect restart_lsn. See |
| 800 |
* SnapBuildProcessRunningXacts. But we need to process cache |
| 801 |
* invalidations if there are any for the reasons mentioned in |
| 802 |
* DecodeCommit. |
| 803 |
*/ |
| 804 |
if (DecodeTXNNeedSkip(ctx, buf, parsed->dbId, origin_id)) |
| 805 |
{ |
| 806 |
ReorderBufferSkipPrepare(ctx->reorder, xid); |
| 807 |
ReorderBufferInvalidate(ctx->reorder, xid, buf->origptr); |
| 808 |
return; |
| 809 |
} |
| 810 |
|
| 811 |
/* Tell the reorderbuffer about the surviving subtransactions. */ |
| 812 |
for (i = 0; i < parsed->nsubxacts; i++) |
| 813 |
{ |
| 814 |
ReorderBufferCommitChild(ctx->reorder, xid, parsed->subxacts[i], |
| 815 |
buf->origptr, buf->endptr); |
| 816 |
} |
| 817 |
|
| 818 |
/* replay actions of all transaction + subtransactions in order */ |
| 819 |
ReorderBufferPrepare(ctx->reorder, xid, parsed->twophase_gid); |
| 820 |
|
| 821 |
/* |
| 822 |
* Update the decoding stats at transaction prepare/commit/abort. |
| 823 |
* Additionally we send the stats when we spill or stream the changes to |
| 824 |
* avoid losing them in case the decoding is interrupted. It is not clear |
| 825 |
* that sending more or less frequently than this would be better. |
| 826 |
*/ |
| 827 |
UpdateDecodingStats(ctx); |
| 828 |
} |
| 829 |
|
| 830 |
|
| 831 |
/* |
| 832 |
* Get the data from the various forms of abort records and pass it on to |
| 833 |
* snapbuild.c and reorderbuffer.c. |
| 834 |
* |
| 835 |
* 'two_phase' indicates to finish prepared transaction. |
| 836 |
*/ |
| 837 |
static void |
| 838 |
DecodeAbort(LogicalDecodingContext *ctx, XLogRecordBuffer *buf, |
| 839 |
xl_xact_parsed_abort *parsed, TransactionId xid, |
| 840 |
bool two_phase) |
| 841 |
{ |
| 842 |
int i; |
| 843 |
XLogRecPtr origin_lsn = InvalidXLogRecPtr; |
| 844 |
TimestampTz abort_time = parsed->xact_time; |
| 845 |
RepOriginId origin_id = XLogRecGetOrigin(buf->record); |
| 846 |
bool skip_xact; |
| 847 |
|
| 848 |
if (parsed->xinfo & XACT_XINFO_HAS_ORIGIN) |
| 849 |
{ |
| 850 |
origin_lsn = parsed->origin_lsn; |
| 851 |
abort_time = parsed->origin_timestamp; |
| 852 |
} |
| 853 |
|
| 854 |
/* |
| 855 |
* Check whether we need to process this transaction. See |
| 856 |
* DecodeTXNNeedSkip for the reasons why we sometimes want to skip the |
| 857 |
* transaction. |
| 858 |
*/ |
| 859 |
skip_xact = DecodeTXNNeedSkip(ctx, buf, parsed->dbId, origin_id); |
| 860 |
|
| 861 |
/* |
| 862 |
* Send the final rollback record for a prepared transaction unless we |
| 863 |
* need to skip it. For non-two-phase xacts, simply forget the xact. |
| 864 |
*/ |
| 865 |
if (two_phase && !skip_xact) |
| 866 |
{ |
| 867 |
ReorderBufferFinishPrepared(ctx->reorder, xid, buf->origptr, buf->endptr, |
| 868 |
InvalidXLogRecPtr, |
| 869 |
abort_time, origin_id, origin_lsn, |
| 870 |
parsed->twophase_gid, false); |
| 871 |
} |
| 872 |
else |
| 873 |
{ |
| 874 |
for (i = 0; i < parsed->nsubxacts; i++) |
| 875 |
{ |
| 876 |
ReorderBufferAbort(ctx->reorder, parsed->subxacts[i], |
| 877 |
buf->record->EndRecPtr, abort_time); |
| 878 |
} |
| 879 |
|
| 880 |
ReorderBufferAbort(ctx->reorder, xid, buf->record->EndRecPtr, |
| 881 |
abort_time); |
| 882 |
} |
| 883 |
|
| 884 |
/* update the decoding stats */ |
| 885 |
UpdateDecodingStats(ctx); |
| 886 |
} |
| 887 |
|
| 888 |
/* |
| 889 |
* Parse XLOG_HEAP_INSERT (not MULTI_INSERT!) records into tuplebufs. |
| 890 |
* |
| 891 |
* Inserts can contain the new tuple. |
| 892 |
*/ |
| 893 |
static void |
| 894 |
DecodeInsert(LogicalDecodingContext *ctx, XLogRecordBuffer *buf) |
| 895 |
{ |
| 896 |
Size datalen; |
| 897 |
char *tupledata; |
| 898 |
Size tuplelen; |
| 899 |
XLogReaderState *r = buf->record; |
| 900 |
xl_heap_insert *xlrec; |
| 901 |
ReorderBufferChange *change; |
| 902 |
RelFileLocator target_locator; |
| 903 |
|
| 904 |
xlrec = (xl_heap_insert *) XLogRecGetData(r); |
| 905 |
|
| 906 |
/* |
| 907 |
* Ignore insert records without new tuples (this does happen when |
| 908 |
* raw_heap_insert marks the TOAST record as HEAP_INSERT_NO_LOGICAL). |
| 909 |
*/ |
| 910 |
if (!(xlrec->flags & XLH_INSERT_CONTAINS_NEW_TUPLE)) |
| 911 |
return; |
| 912 |
|
| 913 |
/* only interested in our database */ |
| 914 |
XLogRecGetBlockTag(r, 0, &target_locator, NULL, NULL); |
| 915 |
if (target_locator.dbOid != ctx->slot->data.database) |
| 916 |
return; |
| 917 |
|
| 918 |
/* output plugin doesn't look for this origin, no need to queue */ |
| 919 |
if (FilterByOrigin(ctx, XLogRecGetOrigin(r))) |
| 920 |
return; |
| 921 |
|
| 922 |
change = ReorderBufferAllocChange(ctx->reorder); |
| 923 |
if (!(xlrec->flags & XLH_INSERT_IS_SPECULATIVE)) |
| 924 |
change->action = REORDER_BUFFER_CHANGE_INSERT; |
| 925 |
else |
| 926 |
change->action = REORDER_BUFFER_CHANGE_INTERNAL_SPEC_INSERT; |
| 927 |
change->origin_id = XLogRecGetOrigin(r); |
| 928 |
|
| 929 |
memcpy(&change->data.tp.rlocator, &target_locator, sizeof(RelFileLocator)); |
| 930 |
|
| 931 |
tupledata = XLogRecGetBlockData(r, 0, &datalen); |
| 932 |
tuplelen = datalen - SizeOfHeapHeader; |
| 933 |
|
| 934 |
change->data.tp.newtuple = |
| 935 |
ReorderBufferAllocTupleBuf(ctx->reorder, tuplelen); |
| 936 |
|
| 937 |
DecodeXLogTuple(tupledata, datalen, change->data.tp.newtuple); |
| 938 |
|
| 939 |
change->data.tp.clear_toast_afterwards = true; |
| 940 |
|
| 941 |
ReorderBufferQueueChange(ctx->reorder, XLogRecGetXid(r), buf->origptr, |
| 942 |
change, |
| 943 |
xlrec->flags & XLH_INSERT_ON_TOAST_RELATION); |
| 944 |
} |
| 945 |
|
| 946 |
/* |
| 947 |
* Parse XLOG_HEAP_UPDATE and XLOG_HEAP_HOT_UPDATE, which have the same layout |
| 948 |
* in the record, from wal into proper tuplebufs. |
| 949 |
* |
| 950 |
* Updates can possibly contain a new tuple and the old primary key. |
| 951 |
*/ |
| 952 |
static void |
| 953 |
DecodeUpdate(LogicalDecodingContext *ctx, XLogRecordBuffer *buf) |
| 954 |
{ |
| 955 |
XLogReaderState *r = buf->record; |
| 956 |
xl_heap_update *xlrec; |
| 957 |
ReorderBufferChange *change; |
| 958 |
char *data; |
| 959 |
RelFileLocator target_locator; |
| 960 |
|
| 961 |
xlrec = (xl_heap_update *) XLogRecGetData(r); |
| 962 |
|
| 963 |
/* only interested in our database */ |
| 964 |
XLogRecGetBlockTag(r, 0, &target_locator, NULL, NULL); |
| 965 |
if (target_locator.dbOid != ctx->slot->data.database) |
| 966 |
return; |
| 967 |
|
| 968 |
/* output plugin doesn't look for this origin, no need to queue */ |
| 969 |
if (FilterByOrigin(ctx, XLogRecGetOrigin(r))) |
| 970 |
return; |
| 971 |
|
| 972 |
change = ReorderBufferAllocChange(ctx->reorder); |
| 973 |
change->action = REORDER_BUFFER_CHANGE_UPDATE; |
| 974 |
change->origin_id = XLogRecGetOrigin(r); |
| 975 |
memcpy(&change->data.tp.rlocator, &target_locator, sizeof(RelFileLocator)); |
| 976 |
|
| 977 |
if (xlrec->flags & XLH_UPDATE_CONTAINS_NEW_TUPLE) |
| 978 |
{ |
| 979 |
Size datalen; |
| 980 |
Size tuplelen; |
| 981 |
|
| 982 |
data = XLogRecGetBlockData(r, 0, &datalen); |
| 983 |
|
| 984 |
tuplelen = datalen - SizeOfHeapHeader; |
| 985 |
|
| 986 |
change->data.tp.newtuple = |
| 987 |
ReorderBufferAllocTupleBuf(ctx->reorder, tuplelen); |
| 988 |
|
| 989 |
DecodeXLogTuple(data, datalen, change->data.tp.newtuple); |
| 990 |
} |
| 991 |
|
| 992 |
if (xlrec->flags & XLH_UPDATE_CONTAINS_OLD) |
| 993 |
{ |
| 994 |
Size datalen; |
| 995 |
Size tuplelen; |
| 996 |
|
| 997 |
/* caution, remaining data in record is not aligned */ |
| 998 |
data = XLogRecGetData(r) + SizeOfHeapUpdate; |
| 999 |
datalen = XLogRecGetDataLen(r) - SizeOfHeapUpdate; |
| 1000 |
tuplelen = datalen - SizeOfHeapHeader; |
| 1001 |
|
| 1002 |
change->data.tp.oldtuple = |
| 1003 |
ReorderBufferAllocTupleBuf(ctx->reorder, tuplelen); |
| 1004 |
|
| 1005 |
DecodeXLogTuple(data, datalen, change->data.tp.oldtuple); |
| 1006 |
} |
| 1007 |
|
| 1008 |
change->data.tp.clear_toast_afterwards = true; |
| 1009 |
|
| 1010 |
ReorderBufferQueueChange(ctx->reorder, XLogRecGetXid(r), buf->origptr, |
| 1011 |
change, false); |
| 1012 |
} |
| 1013 |
|
| 1014 |
/* |
| 1015 |
* Parse XLOG_HEAP_DELETE from wal into proper tuplebufs. |
| 1016 |
* |
| 1017 |
* Deletes can possibly contain the old primary key. |
| 1018 |
*/ |
| 1019 |
static void |
| 1020 |
DecodeDelete(LogicalDecodingContext *ctx, XLogRecordBuffer *buf) |
| 1021 |
{ |
| 1022 |
XLogReaderState *r = buf->record; |
| 1023 |
xl_heap_delete *xlrec; |
| 1024 |
ReorderBufferChange *change; |
| 1025 |
RelFileLocator target_locator; |
| 1026 |
|
| 1027 |
xlrec = (xl_heap_delete *) XLogRecGetData(r); |
| 1028 |
|
| 1029 |
/* only interested in our database */ |
| 1030 |
XLogRecGetBlockTag(r, 0, &target_locator, NULL, NULL); |
| 1031 |
if (target_locator.dbOid != ctx->slot->data.database) |
| 1032 |
return; |
| 1033 |
|
| 1034 |
/* output plugin doesn't look for this origin, no need to queue */ |
| 1035 |
if (FilterByOrigin(ctx, XLogRecGetOrigin(r))) |
| 1036 |
return; |
| 1037 |
|
| 1038 |
change = ReorderBufferAllocChange(ctx->reorder); |
| 1039 |
|
| 1040 |
if (xlrec->flags & XLH_DELETE_IS_SUPER) |
| 1041 |
change->action = REORDER_BUFFER_CHANGE_INTERNAL_SPEC_ABORT; |
| 1042 |
else |
| 1043 |
change->action = REORDER_BUFFER_CHANGE_DELETE; |
| 1044 |
|
| 1045 |
change->origin_id = XLogRecGetOrigin(r); |
| 1046 |
|
| 1047 |
memcpy(&change->data.tp.rlocator, &target_locator, sizeof(RelFileLocator)); |
| 1048 |
|
| 1049 |
/* old primary key stored */ |
| 1050 |
if (xlrec->flags & XLH_DELETE_CONTAINS_OLD) |
| 1051 |
{ |
| 1052 |
Size datalen = XLogRecGetDataLen(r) - SizeOfHeapDelete; |
| 1053 |
Size tuplelen = datalen - SizeOfHeapHeader; |
| 1054 |
|
| 1055 |
Assert(XLogRecGetDataLen(r) > (SizeOfHeapDelete + SizeOfHeapHeader)); |
| 1056 |
|
| 1057 |
change->data.tp.oldtuple = |
| 1058 |
ReorderBufferAllocTupleBuf(ctx->reorder, tuplelen); |
| 1059 |
|
| 1060 |
DecodeXLogTuple((char *) xlrec + SizeOfHeapDelete, |
| 1061 |
datalen, change->data.tp.oldtuple); |
| 1062 |
} |
| 1063 |
|
| 1064 |
change->data.tp.clear_toast_afterwards = true; |
| 1065 |
|
| 1066 |
ReorderBufferQueueChange(ctx->reorder, XLogRecGetXid(r), buf->origptr, |
| 1067 |
change, false); |
| 1068 |
} |
| 1069 |
|
| 1070 |
/* |
| 1071 |
* Parse XLOG_HEAP_TRUNCATE from wal |
| 1072 |
*/ |
| 1073 |
static void |
| 1074 |
DecodeTruncate(LogicalDecodingContext *ctx, XLogRecordBuffer *buf) |
| 1075 |
{ |
| 1076 |
XLogReaderState *r = buf->record; |
| 1077 |
xl_heap_truncate *xlrec; |
| 1078 |
ReorderBufferChange *change; |
| 1079 |
|
| 1080 |
xlrec = (xl_heap_truncate *) XLogRecGetData(r); |
| 1081 |
|
| 1082 |
/* only interested in our database */ |
| 1083 |
if (xlrec->dbId != ctx->slot->data.database) |
| 1084 |
return; |
| 1085 |
|
| 1086 |
/* output plugin doesn't look for this origin, no need to queue */ |
| 1087 |
if (FilterByOrigin(ctx, XLogRecGetOrigin(r))) |
| 1088 |
return; |
| 1089 |
|
| 1090 |
change = ReorderBufferAllocChange(ctx->reorder); |
| 1091 |
change->action = REORDER_BUFFER_CHANGE_TRUNCATE; |
| 1092 |
change->origin_id = XLogRecGetOrigin(r); |
| 1093 |
if (xlrec->flags & XLH_TRUNCATE_CASCADE) |
| 1094 |
change->data.truncate.cascade = true; |
| 1095 |
if (xlrec->flags & XLH_TRUNCATE_RESTART_SEQS) |
| 1096 |
change->data.truncate.restart_seqs = true; |
| 1097 |
change->data.truncate.nrelids = xlrec->nrelids; |
| 1098 |
change->data.truncate.relids = ReorderBufferAllocRelids(ctx->reorder, |
| 1099 |
xlrec->nrelids); |
| 1100 |
memcpy(change->data.truncate.relids, xlrec->relids, |
| 1101 |
xlrec->nrelids * sizeof(Oid)); |
| 1102 |
ReorderBufferQueueChange(ctx->reorder, XLogRecGetXid(r), |
| 1103 |
buf->origptr, change, false); |
| 1104 |
} |
| 1105 |
|
| 1106 |
/* |
| 1107 |
* Decode XLOG_HEAP2_MULTI_INSERT record into multiple tuplebufs. |
| 1108 |
* |
| 1109 |
* Currently MULTI_INSERT will always contain the full tuples. |
| 1110 |
*/ |
| 1111 |
static void |
| 1112 |
DecodeMultiInsert(LogicalDecodingContext *ctx, XLogRecordBuffer *buf) |
| 1113 |
{ |
| 1114 |
XLogReaderState *r = buf->record; |
| 1115 |
xl_heap_multi_insert *xlrec; |
| 1116 |
int i; |
| 1117 |
char *data; |
| 1118 |
char *tupledata; |
| 1119 |
Size tuplelen; |
| 1120 |
RelFileLocator rlocator; |
| 1121 |
|
| 1122 |
xlrec = (xl_heap_multi_insert *) XLogRecGetData(r); |
| 1123 |
|
| 1124 |
/* |
| 1125 |
* Ignore insert records without new tuples. This happens when a |
| 1126 |
* multi_insert is done on a catalog or on a non-persistent relation. |
| 1127 |
*/ |
| 1128 |
if (!(xlrec->flags & XLH_INSERT_CONTAINS_NEW_TUPLE)) |
| 1129 |
return; |
| 1130 |
|
| 1131 |
/* only interested in our database */ |
| 1132 |
XLogRecGetBlockTag(r, 0, &rlocator, NULL, NULL); |
| 1133 |
if (rlocator.dbOid != ctx->slot->data.database) |
| 1134 |
return; |
| 1135 |
|
| 1136 |
/* output plugin doesn't look for this origin, no need to queue */ |
| 1137 |
if (FilterByOrigin(ctx, XLogRecGetOrigin(r))) |
| 1138 |
return; |
| 1139 |
|
| 1140 |
/* |
| 1141 |
* We know that this multi_insert isn't for a catalog, so the block should |
| 1142 |
* always have data even if a full-page write of it is taken. |
| 1143 |
*/ |
| 1144 |
tupledata = XLogRecGetBlockData(r, 0, &tuplelen); |
| 1145 |
Assert(tupledata != NULL); |
| 1146 |
|
| 1147 |
data = tupledata; |
| 1148 |
for (i = 0; i < xlrec->ntuples; i++) |
| 1149 |
{ |
| 1150 |
ReorderBufferChange *change; |
| 1151 |
xl_multi_insert_tuple *xlhdr; |
| 1152 |
int datalen; |
| 1153 |
HeapTuple tuple; |
| 1154 |
HeapTupleHeader header; |
| 1155 |
|
| 1156 |
change = ReorderBufferAllocChange(ctx->reorder); |
| 1157 |
change->action = REORDER_BUFFER_CHANGE_INSERT; |
| 1158 |
change->origin_id = XLogRecGetOrigin(r); |
| 1159 |
|
| 1160 |
memcpy(&change->data.tp.rlocator, &rlocator, sizeof(RelFileLocator)); |
| 1161 |
|
| 1162 |
xlhdr = (xl_multi_insert_tuple *) SHORTALIGN(data); |
| 1163 |
data = ((char *) xlhdr) + SizeOfMultiInsertTuple; |
| 1164 |
datalen = xlhdr->datalen; |
| 1165 |
|
| 1166 |
change->data.tp.newtuple = |
| 1167 |
ReorderBufferAllocTupleBuf(ctx->reorder, datalen); |
| 1168 |
|
| 1169 |
tuple = change->data.tp.newtuple; |
| 1170 |
header = tuple->t_data; |
| 1171 |
|
| 1172 |
/* not a disk based tuple */ |
| 1173 |
ItemPointerSetInvalid(&tuple->t_self); |
| 1174 |
|
| 1175 |
/* |
| 1176 |
* We can only figure this out after reassembling the transactions. |
| 1177 |
*/ |
| 1178 |
tuple->t_tableOid = InvalidOid; |
| 1179 |
|
| 1180 |
tuple->t_len = datalen + SizeofHeapTupleHeader; |
| 1181 |
|
| 1182 |
memset(header, 0, SizeofHeapTupleHeader); |
| 1183 |
|
| 1184 |
memcpy((char *) tuple->t_data + SizeofHeapTupleHeader, data, datalen); |
| 1185 |
header->t_infomask = xlhdr->t_infomask; |
| 1186 |
header->t_infomask2 = xlhdr->t_infomask2; |
| 1187 |
header->t_hoff = xlhdr->t_hoff; |
| 1188 |
|
| 1189 |
/* |
| 1190 |
* Reset toast reassembly state only after the last row in the last |
| 1191 |
* xl_multi_insert_tuple record emitted by one heap_multi_insert() |
| 1192 |
* call. |
| 1193 |
*/ |
| 1194 |
if (xlrec->flags & XLH_INSERT_LAST_IN_MULTI && |
| 1195 |
(i + 1) == xlrec->ntuples) |
| 1196 |
change->data.tp.clear_toast_afterwards = true; |
| 1197 |
else |
| 1198 |
change->data.tp.clear_toast_afterwards = false; |
| 1199 |
|
| 1200 |
ReorderBufferQueueChange(ctx->reorder, XLogRecGetXid(r), |
| 1201 |
buf->origptr, change, false); |
| 1202 |
|
| 1203 |
/* move to the next xl_multi_insert_tuple entry */ |
| 1204 |
data += datalen; |
| 1205 |
} |
| 1206 |
Assert(data == tupledata + tuplelen); |
| 1207 |
} |
| 1208 |
|
| 1209 |
/* |
| 1210 |
* Parse XLOG_HEAP_CONFIRM from wal into a confirmation change. |
| 1211 |
* |
| 1212 |
* This is pretty trivial, all the state essentially already setup by the |
| 1213 |
* speculative insertion. |
| 1214 |
*/ |
| 1215 |
static void |
| 1216 |
DecodeSpecConfirm(LogicalDecodingContext *ctx, XLogRecordBuffer *buf) |
| 1217 |
{ |
| 1218 |
XLogReaderState *r = buf->record; |
| 1219 |
ReorderBufferChange *change; |
| 1220 |
RelFileLocator target_locator; |
| 1221 |
|
| 1222 |
/* only interested in our database */ |
| 1223 |
XLogRecGetBlockTag(r, 0, &target_locator, NULL, NULL); |
| 1224 |
if (target_locator.dbOid != ctx->slot->data.database) |
| 1225 |
return; |
| 1226 |
|
| 1227 |
/* output plugin doesn't look for this origin, no need to queue */ |
| 1228 |
if (FilterByOrigin(ctx, XLogRecGetOrigin(r))) |
| 1229 |
return; |
| 1230 |
|
| 1231 |
change = ReorderBufferAllocChange(ctx->reorder); |
| 1232 |
change->action = REORDER_BUFFER_CHANGE_INTERNAL_SPEC_CONFIRM; |
| 1233 |
change->origin_id = XLogRecGetOrigin(r); |
| 1234 |
|
| 1235 |
memcpy(&change->data.tp.rlocator, &target_locator, sizeof(RelFileLocator)); |
| 1236 |
|
| 1237 |
change->data.tp.clear_toast_afterwards = true; |
| 1238 |
|
| 1239 |
ReorderBufferQueueChange(ctx->reorder, XLogRecGetXid(r), buf->origptr, |
| 1240 |
change, false); |
| 1241 |
} |
| 1242 |
|
| 1243 |
|
| 1244 |
/* |
| 1245 |
* Read a HeapTuple as WAL logged by heap_insert, heap_update and heap_delete |
| 1246 |
* (but not by heap_multi_insert) into a tuplebuf. |
| 1247 |
* |
| 1248 |
* The size 'len' and the pointer 'data' in the record need to be |
| 1249 |
* computed outside as they are record specific. |
| 1250 |
*/ |
| 1251 |
static void |
| 1252 |
DecodeXLogTuple(char *data, Size len, HeapTuple tuple) |
| 1253 |
{ |
| 1254 |
xl_heap_header xlhdr; |
| 1255 |
int datalen = len - SizeOfHeapHeader; |
| 1256 |
HeapTupleHeader header; |
| 1257 |
|
| 1258 |
Assert(datalen >= 0); |
| 1259 |
|
| 1260 |
tuple->t_len = datalen + SizeofHeapTupleHeader; |
| 1261 |
header = tuple->t_data; |
| 1262 |
|
| 1263 |
/* not a disk based tuple */ |
| 1264 |
ItemPointerSetInvalid(&tuple->t_self); |
| 1265 |
|
| 1266 |
/* we can only figure this out after reassembling the transactions */ |
| 1267 |
tuple->t_tableOid = InvalidOid; |
| 1268 |
|
| 1269 |
/* data is not stored aligned, copy to aligned storage */ |
| 1270 |
memcpy(&xlhdr, data, SizeOfHeapHeader); |
| 1271 |
|
| 1272 |
memset(header, 0, SizeofHeapTupleHeader); |
| 1273 |
|
| 1274 |
memcpy(((char *) tuple->t_data) + SizeofHeapTupleHeader, |
| 1275 |
data + SizeOfHeapHeader, |
| 1276 |
datalen); |
| 1277 |
|
| 1278 |
header->t_infomask = xlhdr.t_infomask; |
| 1279 |
header->t_infomask2 = xlhdr.t_infomask2; |
| 1280 |
header->t_hoff = xlhdr.t_hoff; |
| 1281 |
} |
| 1282 |
|
| 1283 |
/* |
| 1284 |
* Check whether we are interested in this specific transaction. |
| 1285 |
* |
| 1286 |
* There can be several reasons we might not be interested in this |
| 1287 |
* transaction: |
| 1288 |
* 1) We might not be interested in decoding transactions up to this |
| 1289 |
* LSN. This can happen because we previously decoded it and now just |
| 1290 |
* are restarting or if we haven't assembled a consistent snapshot yet. |
| 1291 |
* 2) The transaction happened in another database. |
| 1292 |
* 3) The output plugin is not interested in the origin. |
| 1293 |
* 4) We are doing fast-forwarding |
| 1294 |
*/ |
| 1295 |
static bool |
| 1296 |
DecodeTXNNeedSkip(LogicalDecodingContext *ctx, XLogRecordBuffer *buf, |
| 1297 |
Oid txn_dbid, RepOriginId origin_id) |
| 1298 |
{ |
| 1299 |
if (SnapBuildXactNeedsSkip(ctx->snapshot_builder, buf->origptr) || |
| 1300 |
(txn_dbid != InvalidOid && txn_dbid != ctx->slot->data.database) || |
| 1301 |
FilterByOrigin(ctx, origin_id)) |
| 1302 |
return true; |
| 1303 |
|
| 1304 |
/* |
| 1305 |
* We also skip decoding in fast_forward mode. In passing set the |
| 1306 |
* processing_required flag to indicate that if it were not for |
| 1307 |
* fast_forward mode, processing would have been required. |
| 1308 |
*/ |
| 1309 |
if (ctx->fast_forward) |
| 1310 |
{ |
| 1311 |
ctx->processing_required = true; |
| 1312 |
return true; |
| 1313 |
} |
| 1314 |
|
| 1315 |
return false; |
| 1316 |
} |