| blob | e3a5c7b660c5b852d6c58b18d11d8151a23b59aa |
1 /*-------------------------------------------------------------------------
2 *
3 * reorderbuffer.c
4 * PostgreSQL logical replay/reorder buffer management
5 *
6 *
7 * Copyright (c) 2012-2024, PostgreSQL Global Development Group
8 *
9 *
10 * IDENTIFICATION
11 * src/backend/replication/logical/reorderbuffer.c
12 *
13 * NOTES
14 * This module gets handed individual pieces of transactions in the order
15 * they are written to the WAL and is responsible to reassemble them into
16 * toplevel transaction sized pieces. When a transaction is completely
17 * reassembled - signaled by reading the transaction commit record - it
18 * will then call the output plugin (cf. ReorderBufferCommit()) with the
19 * individual changes. The output plugins rely on snapshots built by
20 * snapbuild.c which hands them to us.
21 *
22 * Transactions and subtransactions/savepoints in postgres are not
23 * immediately linked to each other from outside the performing
24 * backend. Only at commit/abort (or special xact_assignment records) they
25 * are linked together. Which means that we will have to splice together a
26 * toplevel transaction from its subtransactions. To do that efficiently we
27 * build a binary heap indexed by the smallest current lsn of the individual
28 * subtransactions' changestreams. As the individual streams are inherently
29 * ordered by LSN - since that is where we build them from - the transaction
30 * can easily be reassembled by always using the subtransaction with the
31 * smallest current LSN from the heap.
32 *
33 * In order to cope with large transactions - which can be several times as
34 * big as the available memory - this module supports spooling the contents
35 * of large transactions to disk. When the transaction is replayed the
36 * contents of individual (sub-)transactions will be read from disk in
37 * chunks.
38 *
39 * This module also has to deal with reassembling toast records from the
40 * individual chunks stored in WAL. When a new (or initial) version of a
41 * tuple is stored in WAL it will always be preceded by the toast chunks
42 * emitted for the columns stored out of line. Within a single toplevel
43 * transaction there will be no other data carrying records between a row's
44 * toast chunks and the row data itself. See ReorderBufferToast* for
45 * details.
46 *
47 * ReorderBuffer uses two special memory context types - SlabContext for
48 * allocations of fixed-length structures (changes and transactions), and
49 * GenerationContext for the variable-length transaction data (allocated
50 * and freed in groups with similar lifespans).
51 *
52 * To limit the amount of memory used by decoded changes, we track memory
53 * used at the reorder buffer level (i.e. total amount of memory), and for
54 * each transaction. When the total amount of used memory exceeds the
55 * limit, the transaction consuming the most memory is then serialized to
56 * disk.
57 *
58 * Only decoded changes are evicted from memory (spilled to disk), not the
59 * transaction records. The number of toplevel transactions is limited,
60 * but a transaction with many subtransactions may still consume significant
61 * amounts of memory. However, the transaction records are fairly small and
62 * are not included in the memory limit.
63 *
64 * The current eviction algorithm is very simple - the transaction is
65 * picked merely by size, while it might be useful to also consider age
66 * (LSN) of the changes for example. With the new Generational memory
67 * allocator, evicting the oldest changes would make it more likely the
68 * memory gets actually freed.
69 *
70 * We use a max-heap with transaction size as the key to efficiently find
71 * the largest transaction. We update the max-heap whenever the memory
72 * counter is updated; however transactions with size 0 are not stored in
73 * the heap, because they have no changes to evict.
74 *
75 * We still rely on max_changes_in_memory when loading serialized changes
76 * back into memory. At that point we can't use the memory limit directly
77 * as we load the subxacts independently. One option to deal with this
78 * would be to count the subxacts, and allow each to allocate 1/N of the
79 * memory limit. That however does not seem very appealing, because with
80 * many subtransactions it may easily cause thrashing (short cycles of
81 * deserializing and applying very few changes). We probably should give
82 * a bit more memory to the oldest subtransactions, because it's likely
83 * they are the source for the next sequence of changes.
84 *
85 * -------------------------------------------------------------------------
86 */
89 #include <unistd.h>
90 #include <sys/stat.h>
115 /* entry for a hash table we use to map from xid to our transaction state */
117 {
118 TransactionId xid;
122 /* data structures for (relfilelocator, ctid) => (cmin, cmax) mapping */
124 {
125 RelFileLocator rlocator;
126 ItemPointerData tid;
130 {
131 ReorderBufferTupleCidKey key;
132 CommandId cmin;
133 CommandId cmax;
137 /* Virtual file descriptor with file offset tracking */
139 {
142 * when vfd is opened. */
145 /* k-way in-order change iteration support structures */
147 {
148 XLogRecPtr lsn;
151 TXNEntryFile file;
152 XLogSegNo segno;
156 {
158 Size nr_txns;
159 dlist_head old_change;
163 /* toast datastructures */
165 {
168 * have seen */
173 * main tup */
176 /* Disk serialization support datastructures */
178 {
179 Size size;
180 ReorderBufferChange change;
181 /* data follows */
184 #define IsSpecInsert(action) \
185 ( \
186 ((action) == REORDER_BUFFER_CHANGE_INTERNAL_SPEC_INSERT) \
187 )
188 #define IsSpecConfirmOrAbort(action) \
189 ( \
190 (((action) == REORDER_BUFFER_CHANGE_INTERNAL_SPEC_CONFIRM) || \
191 ((action) == REORDER_BUFFER_CHANGE_INTERNAL_SPEC_ABORT)) \
192 )
193 #define IsInsertOrUpdate(action) \
194 ( \
195 (((action) == REORDER_BUFFER_CHANGE_INSERT) || \
196 ((action) == REORDER_BUFFER_CHANGE_UPDATE) || \
197 ((action) == REORDER_BUFFER_CHANGE_INTERNAL_SPEC_INSERT)) \
198 )
200 /*
201 * Maximum number of changes kept in memory, per transaction. After that,
202 * changes are spooled to disk.
203 *
204 * The current value should be sufficient to decode the entire transaction
205 * without hitting disk in OLTP workloads, while starting to spool to disk in
206 * other workloads reasonably fast.
207 *
208 * At some point in the future it probably makes sense to have a more elaborate
209 * resource management here, but it's not entirely clear what that would look
210 * like.
211 */
215 /* GUC variable */
218 /* ---------------------------------------
219 * primary reorderbuffer support routines
220 * ---------------------------------------
221 */
232 /* ---------------------------------------
233 * support functions for lsn-order iterating over the ->changes of a
234 * transaction and its subtransactions
235 *
236 * used for iteration over the k-way heap merge of a transaction and its
237 * subtransactions
238 * ---------------------------------------
239 */
242 static ReorderBufferChange *ReorderBufferIterTXNNext(ReorderBuffer *rb, ReorderBufferIterTXNState *state);
247 /*
248 * ---------------------------------------
249 * Disk serialization support functions
250 * ---------------------------------------
251 */
266 static int ReorderBufferTXNSizeCompare(const pairingheap_node *a, const pairingheap_node *b, void *arg);
272 /*
273 * ---------------------------------------
274 * Streaming support functions
275 * ---------------------------------------
276 */
282 /* ---------------------------------------
283 * toast reassembly support
284 * ---------------------------------------
285 */
293 /*
294 * ---------------------------------------
295 * memory accounting
296 * ---------------------------------------
297 */
304 /*
305 * Allocate a new ReorderBuffer and clean out any old serialized state from
306 * prior ReorderBuffer instances for the same slot.
307 */
308 ReorderBuffer *
310 {
312 HASHCTL hash_ctl;
313 MemoryContext new_ctx;
317 /* allocate memory in own context, to have better accountability */
320 ALLOCSET_DEFAULT_SIZES);
322 buffer =
331 SLAB_DEFAULT_BLOCK_SIZE,
336 SLAB_DEFAULT_BLOCK_SIZE,
339 /*
340 * To minimize memory fragmentation caused by long-running transactions
341 * with changes spanning multiple memory blocks, we use a single
342 * fixed-size memory block for decoded tuple storage. The performance
343 * testing showed that the default memory block size maintains logical
344 * decoding performance without causing fragmentation due to concurrent
345 * transactions. One might think that we can use the max size as
346 * SLAB_LARGE_BLOCK_SIZE but the test also showed it doesn't help resolve
347 * the memory fragmentation.
348 */
351 SLAB_DEFAULT_BLOCK_SIZE,
352 SLAB_DEFAULT_BLOCK_SIZE,
353 SLAB_DEFAULT_BLOCK_SIZE);
369 /* txn_heap is ordered by transaction size */
387 /*
388 * Ensure there's no stale data from prior uses of this slot, in case some
389 * prior exit avoided calling ReorderBufferFree. Failure to do this can
390 * produce duplicated txns, and it's very cheap if there's nothing there.
391 */
395 }
397 /*
398 * Free a ReorderBuffer
399 */
400 void
402 {
405 /*
406 * We free separately allocated data by entirely scrapping reorderbuffer's
407 * memory context.
408 */
411 /* Free disk space used by unconsumed reorder buffers */
413 }
415 /*
416 * Get an unused, possibly preallocated, ReorderBufferTXN.
417 */
420 {
432 /* InvalidCommandId is not zero, so set it explicitly */
437 }
439 /*
440 * Free a ReorderBufferTXN.
441 */
442 static void
444 {
445 /* clean the lookup cache if we were cached (quite likely) */
447 {
450 }
452 /* free data that's contained */
455 {
458 }
461 {
464 }
467 {
470 }
472 /* Reset the toast hash */
475 /* All changes must be deallocated */
479 }
481 /*
482 * Get a fresh ReorderBufferChange.
483 */
484 ReorderBufferChange *
486 {
494 }
496 /*
497 * Free a ReorderBufferChange and update memory accounting, if requested.
498 */
499 void
502 {
503 /* update memory accounting info */
508 /* free contained data */
510 {
516 {
519 }
522 {
525 }
542 {
545 }
547 /* no data in addition to the struct itself */
550 {
553 }
560 }
563 }
565 /*
566 * Get a fresh HeapTuple fitting a tuple of size tuple_len (excluding header
567 * overhead).
568 */
569 HeapTuple
571 {
572 HeapTuple tuple;
573 Size alloc_len;
582 }
584 /*
585 * Free a HeapTuple returned by ReorderBufferGetTupleBuf().
586 */
587 void
589 {
591 }
593 /*
594 * Get an array for relids of truncated relations.
595 *
596 * We use the global memory context (for the whole reorder buffer), because
597 * none of the existing ones seems like a good match (some are SLAB, so we
598 * can't use those, and tup_context is meant for tuple data, not relids). We
599 * could add yet another context, but it seems like an overkill - TRUNCATE is
600 * not particularly common operation, so it does not seem worth it.
601 */
602 Oid *
604 {
606 Size alloc_len;
613 }
615 /*
616 * Free an array of relids.
617 */
618 void
620 {
622 }
624 /*
625 * Return the ReorderBufferTXN from the given buffer, specified by Xid.
626 * If create is true, and a transaction doesn't already exist, create it
627 * (with the given LSN, and as top transaction if that's specified);
628 * when this happens, is_new is set to true.
629 */
633 {
640 /*
641 * Check the one-entry lookup cache first
642 */
645 {
649 {
650 /* found it, and it's valid */
654 }
656 /*
657 * cached as non-existent, and asked not to create? Then nothing else
658 * to do.
659 */
662 /* otherwise fall through to create it */
663 }
665 /*
666 * If the cache wasn't hit or it yielded a "does-not-exist" and we want to
667 * create an entry.
668 */
670 /* search the lookup table */
679 {
680 /* initialize the new entry, if creation was requested */
691 {
694 }
695 }
696 else
699 /* update cache */
708 }
710 /*
711 * Record the partial change for the streaming of in-progress transactions. We
712 * can stream only complete changes so if we have a partial change like toast
713 * table insert or speculative insert then we mark such a 'txn' so that it
714 * can't be streamed. We also ensure that if the changes in such a 'txn' can
715 * be streamed and are above logical_decoding_work_mem threshold then we stream
716 * them as soon as we have a complete change.
717 */
718 static void
722 {
725 /*
726 * The partial changes need to be processed only while streaming
727 * in-progress transactions.
728 */
732 /* Get the top transaction. */
735 /*
736 * Indicate a partial change for toast inserts. The change will be
737 * considered as complete once we get the insert or update on the main
738 * table and we are sure that the pending toast chunks are not required
739 * anymore.
740 *
741 * If we allow streaming when there are pending toast chunks then such
742 * chunks won't be released till the insert (multi_insert) is complete and
743 * we expect the txn to have streamed all changes after streaming. This
744 * restriction is mainly to ensure the correctness of streamed
745 * transactions and it doesn't seem worth uplifting such a restriction
746 * just to allow this case because anyway we will stream the transaction
747 * once such an insert is complete.
748 */
756 /*
757 * Indicate a partial change for speculative inserts. The change will be
758 * considered as complete once we get the speculative confirm or abort
759 * token.
760 */
767 /*
768 * Stream the transaction if it is serialized before and the changes are
769 * now complete in the top-level transaction.
770 *
771 * The reason for doing the streaming of such a transaction as soon as we
772 * get the complete change for it is that previously it would have reached
773 * the memory threshold and wouldn't get streamed because of incomplete
774 * changes. Delaying such transactions would increase apply lag for them.
775 */
781 }
783 /*
784 * Queue a change into a transaction so it can be replayed upon commit or will be
785 * streamed when we reach logical_decoding_work_mem threshold.
786 */
787 void
790 {
795 /*
796 * While streaming the previous changes we have detected that the
797 * transaction is aborted. So there is no point in collecting further
798 * changes for it.
799 */
801 {
802 /*
803 * We don't need to update memory accounting for this change as we
804 * have not added it to the queue yet.
805 */
808 }
810 /*
811 * The changes that are sent downstream are considered streamable. We
812 * remember such transactions so that only those will later be considered
813 * for streaming.
814 */
821 {
825 }
835 /* update memory accounting information */
839 /* process partial change */
842 /* check the memory limits and evict something if needed */
844 }
846 /*
847 * A transactional message is queued to be processed upon commit and a
848 * non-transactional message gets processed immediately.
849 */
850 void
855 {
857 {
858 MemoryContext oldcontext;
863 /*
864 * We don't expect snapshots for transactional changes - we'll use the
865 * snapshot derived later during apply (unless the change gets
866 * skipped).
867 */
882 }
883 else
884 {
888 /* Non-transactional changes require a valid snapshot. */
894 /* setup snapshot to allow catalog access */
897 {
901 }
903 {
906 }
908 }
909 }
911 /*
912 * AssertTXNLsnOrder
913 * Verify LSN ordering of transaction lists in the reorderbuffer
914 *
915 * Other LSN-related invariants are checked too.
916 *
917 * No-op if assertions are not in use.
918 */
919 static void
921 {
922 #ifdef USE_ASSERT_CHECKING
924 dlist_iter iter;
928 /*
929 * Skip the verification if we don't reach the LSN at which we start
930 * decoding the contents of transactions yet because until we reach the
931 * LSN, we could have transactions that don't have the association between
932 * the top-level transaction and subtransaction yet and consequently have
933 * the same LSN. We don't guarantee this association until we try to
934 * decode the actual contents of transaction. The ordering of the records
935 * prior to the start_decoding_at LSN should have been checked before the
936 * restart.
937 */
942 {
946 /* start LSN must be set */
949 /* If there is an end LSN, it must be higher than start LSN */
953 /* Current initial LSN must be strictly higher than previous */
957 /* known-as-subtxn txns must not be listed */
961 }
964 {
966 base_snapshot_node,
969 /* base snapshot (and its LSN) must be set */
973 /* current LSN must be strictly higher than previous */
977 /* known-as-subtxn txns must not be listed */
981 }
982 #endif
983 }
985 /*
986 * AssertChangeLsnOrder
987 *
988 * Check ordering of changes in the (sub)transaction.
989 */
990 static void
992 {
993 #ifdef USE_ASSERT_CHECKING
994 dlist_iter iter;
998 {
1013 }
1014 #endif
1015 }
1017 /*
1018 * ReorderBufferGetOldestTXN
1019 * Return oldest transaction in reorderbuffer
1020 */
1021 ReorderBufferTXN *
1023 {
1036 }
1038 /*
1039 * ReorderBufferGetOldestXmin
1040 * Return oldest Xmin in reorderbuffer
1041 *
1042 * Returns oldest possibly running Xid from the point of view of snapshots
1043 * used in the transactions kept by reorderbuffer, or InvalidTransactionId if
1044 * there are none.
1045 *
1046 * Since snapshots are assigned monotonically, this equals the Xmin of the
1047 * base snapshot with minimal base_snapshot_lsn.
1048 */
1049 TransactionId
1051 {
1062 }
1064 void
1066 {
1068 }
1070 /*
1071 * ReorderBufferAssignChild
1072 *
1073 * Make note that we know that subxid is a subtransaction of xid, seen as of
1074 * the given lsn.
1075 */
1076 void
1079 {
1089 {
1091 {
1092 /* already associated, nothing to do */
1094 }
1095 else
1096 {
1097 /*
1098 * We already saw this transaction, but initially added it to the
1099 * list of top-level txns. Now that we know it's not top-level,
1100 * remove it from there.
1101 */
1103 }
1104 }
1110 /* set the reference to top-level transaction */
1113 /* add to subtransaction list */
1117 /* Possibly transfer the subtxn's snapshot to its top-level txn. */
1120 /* Verify LSN-ordering invariant */
1122 }
1124 /*
1125 * ReorderBufferTransferSnapToParent
1126 * Transfer base snapshot from subtxn to top-level txn, if needed
1127 *
1128 * This is done if the top-level txn doesn't have a base snapshot, or if the
1129 * subtxn's base snapshot has an earlier LSN than the top-level txn's base
1130 * snapshot's LSN. This can happen if there are no changes in the toplevel
1131 * txn but there are some in the subtxn, or the first change in subtxn has
1132 * earlier LSN than first change in the top-level txn and we learned about
1133 * their kinship only now.
1134 *
1135 * The subtransaction's snapshot is cleared regardless of the transfer
1136 * happening, since it's not needed anymore in either case.
1137 *
1138 * We do this as soon as we become aware of their kinship, to avoid queueing
1139 * extra snapshots to txns known-as-subtxns -- only top-level txns will
1140 * receive further snapshots.
1141 */
1142 static void
1145 {
1149 {
1152 {
1153 /*
1154 * If the toplevel transaction already has a base snapshot but
1155 * it's newer than the subxact's, purge it.
1156 */
1158 {
1161 }
1163 /*
1164 * The snapshot is now the top transaction's; transfer it, and
1165 * adjust the list position of the top transaction in the list by
1166 * moving it to where the subtransaction is.
1167 */
1173 /*
1174 * The subtransaction doesn't have a snapshot anymore (so it
1175 * mustn't be in the list.)
1176 */
1180 }
1181 else
1182 {
1183 /* Base snap of toplevel is fine, so subxact's is not needed */
1188 }
1189 }
1190 }
1192 /*
1193 * Associate a subtransaction with its toplevel transaction at commit
1194 * time. There may be no further changes added after this.
1195 */
1196 void
1199 XLogRecPtr end_lsn)
1200 {
1206 /*
1207 * No need to do anything if that subtxn didn't contain any changes
1208 */
1215 /*
1216 * Assign this subxact as a child of the toplevel xact (no-op if already
1217 * done.)
1218 */
1220 }
1223 /*
1224 * Support for efficiently iterating over a transaction's and its
1225 * subtransactions' changes.
1226 *
1227 * We do by doing a k-way merge between transactions/subtransactions. For that
1228 * we model the current heads of the different transactions as a binary heap
1229 * so we easily know which (sub-)transaction has the change with the smallest
1230 * lsn next.
1231 *
1232 * We assume the changes in individual transactions are already sorted by LSN.
1233 */
1235 /*
1236 * Binary heap comparison function.
1237 */
1238 static int
1240 {
1250 }
1252 /*
1253 * Allocate & initialize an iterator which iterates in lsn order over a
1254 * transaction and all its subtransactions.
1255 *
1256 * Note: The iterator state is returned through iter_state parameter rather
1257 * than the function's return value. This is because the state gets cleaned up
1258 * in a PG_CATCH block in the caller, so we want to make sure the caller gets
1259 * back the state even if this function throws an exception.
1260 */
1261 static void
1264 {
1267 dlist_iter cur_txn_i;
1268 int32 off;
1272 /* Check ordering of changes in the toplevel transaction. */
1275 /*
1276 * Calculate the size of our heap: one element for every transaction that
1277 * contains changes. (Besides the transactions already in the reorder
1278 * buffer, we count the one we were directly passed.)
1279 */
1281 nr_txns++;
1284 {
1289 /* Check ordering of changes in this subtransaction. */
1293 nr_txns++;
1294 }
1296 /* allocate iteration state */
1306 {
1309 }
1311 /* allocate heap */
1313 ReorderBufferIterCompare,
1314 state);
1316 /* Now that the state fields are initialized, it is safe to return it. */
1319 /*
1320 * Now insert items into the binary heap, in an unordered fashion. (We
1321 * will run a heap assembly step at the end; this is more efficient.)
1322 */
1326 /* add toplevel transaction if it contains changes */
1328 {
1332 {
1333 /* serialize remaining changes */
1337 }
1347 }
1349 /* add subtransactions if they contain changes */
1351 {
1357 {
1361 {
1362 /* serialize remaining changes */
1367 }
1376 }
1377 }
1379 /* assemble a valid binary heap */
1381 }
1383 /*
1384 * Return the next change when iterating over a transaction and its
1385 * subtransactions.
1386 *
1387 * Returns NULL when no further changes exist.
1388 */
1391 {
1394 int32 off;
1396 /* nothing there anymore */
1403 /* free memory we might have "leaked" in the previous *Next call */
1405 {
1410 }
1414 /*
1415 * update heap with information about which transaction has the next
1416 * relevant change in LSN order
1417 */
1419 /* there are in-memory changes */
1421 {
1426 /* txn stays the same */
1432 }
1434 /* try to load changes from disk */
1436 {
1437 /*
1438 * Ugly: restoring changes will reuse *Change records, thus delete the
1439 * current one from the per-tx list and only free in the next call.
1440 */
1444 /*
1445 * Update the total bytes processed by the txn for which we are
1446 * releasing the current set of changes and restoring the new set of
1447 * changes.
1448 */
1452 {
1453 /* successfully restored changes from disk */
1463 /* txn stays the same */
1469 }
1470 }
1472 /* ok, no changes there anymore, remove */
1476 }
1478 /*
1479 * Deallocate the iterator
1480 */
1481 static void
1484 {
1485 int32 off;
1488 {
1491 }
1493 /* free memory we might have "leaked" in the last *Next call */
1495 {
1502 }
1506 }
1508 /*
1509 * Cleanup the contents of a transaction, usually after the transaction
1510 * committed or aborted.
1511 */
1512 static void
1514 {
1516 dlist_mutable_iter iter;
1519 /* cleanup subtransactions & their changes */
1521 {
1526 /*
1527 * Subtransactions are always associated to the toplevel TXN, even if
1528 * they originally were happening inside another subtxn, so we won't
1529 * ever recurse more than one level deep here.
1530 */
1535 }
1537 /* cleanup changes in the txn */
1539 {
1544 /* Check we're not mixing changes from different transactions. */
1547 /*
1548 * Instead of updating the memory counter for individual changes, we
1549 * sum up the size of memory to free so we can update the memory
1550 * counter all together below. This saves costs of maintaining the
1551 * max-heap.
1552 */
1556 }
1558 /* Update the memory counter */
1561 /*
1562 * Cleanup the tuplecids we stored for decoding catalog snapshot access.
1563 * They are always stored in the toplevel transaction.
1564 */
1566 {
1571 /* Check we're not mixing changes from different transactions. */
1576 }
1578 /*
1579 * Cleanup the base snapshot, if set.
1580 */
1582 {
1585 }
1587 /*
1588 * Cleanup the snapshot for the last streamed run.
1589 */
1591 {
1594 }
1596 /*
1597 * Remove TXN from its containing lists.
1598 *
1599 * Note: if txn is known as subxact, we are deleting the TXN from its
1600 * parent's list of known subxacts; this leaves the parent's nsubxacts
1601 * count too high, but we don't care. Otherwise, we are deleting the TXN
1602 * from the LSN-ordered list of toplevel TXNs. We remove the TXN from the
1603 * list of catalog modifying transactions as well.
1604 */
1609 /* now remove reference from buffer */
1613 /* remove entries spilled to disk */
1617 /* deallocate */
1619 }
1621 /*
1622 * Discard changes from a transaction (and subtransactions), either after
1623 * streaming or decoding them at PREPARE. Keep the remaining info -
1624 * transactions, tuplecids, invalidations and snapshots.
1625 *
1626 * We additionally remove tuplecids after decoding the transaction at prepare
1627 * time as we only need to perform invalidation at rollback or commit prepared.
1628 *
1629 * 'txn_prepared' indicates that we have decoded the transaction at prepare
1630 * time.
1631 */
1632 static void
1634 {
1635 dlist_mutable_iter iter;
1638 /* cleanup subtransactions & their changes */
1640 {
1645 /*
1646 * Subtransactions are always associated to the toplevel TXN, even if
1647 * they originally were happening inside another subtxn, so we won't
1648 * ever recurse more than one level deep here.
1649 */
1654 }
1656 /* cleanup changes in the txn */
1658 {
1663 /* Check we're not mixing changes from different transactions. */
1666 /* remove the change from it's containing list */
1669 /*
1670 * Instead of updating the memory counter for individual changes, we
1671 * sum up the size of memory to free so we can update the memory
1672 * counter all together below. This saves costs of maintaining the
1673 * max-heap.
1674 */
1678 }
1680 /* Update the memory counter */
1683 /*
1684 * Mark the transaction as streamed.
1685 *
1686 * The top-level transaction, is marked as streamed always, even if it
1687 * does not contain any changes (that is, when all the changes are in
1688 * subtransactions).
1689 *
1690 * For subtransactions, we only mark them as streamed when there are
1691 * changes in them.
1692 *
1693 * We do it this way because of aborts - we don't want to send aborts for
1694 * XIDs the downstream is not aware of. And of course, it always knows
1695 * about the toplevel xact (we send the XID in all messages), but we never
1696 * stream XIDs of empty subxacts.
1697 */
1702 {
1703 /*
1704 * If this is a prepared txn, cleanup the tuplecids we stored for
1705 * decoding catalog snapshot access. They are always stored in the
1706 * toplevel transaction.
1707 */
1709 {
1714 /* Check we're not mixing changes from different transactions. */
1718 /* Remove the change from its containing list. */
1722 }
1723 }
1725 /*
1726 * Destroy the (relfilelocator, ctid) hashtable, so that we don't leak any
1727 * memory. We could also keep the hash table and update it with new ctid
1728 * values, but this seems simpler and good enough for now.
1729 */
1731 {
1734 }
1736 /* If this txn is serialized then clean the disk space. */
1738 {
1742 /*
1743 * We set this flag to indicate if the transaction is ever serialized.
1744 * We need this to accurately update the stats as otherwise the same
1745 * transaction can be counted as serialized multiple times.
1746 */
1748 }
1750 /* also reset the number of entries in the transaction */
1753 }
1755 /*
1756 * Build a hash with a (relfilelocator, ctid) -> (cmin, cmax) mapping for use by
1757 * HeapTupleSatisfiesHistoricMVCC.
1758 */
1759 static void
1761 {
1762 dlist_iter iter;
1763 HASHCTL hash_ctl;
1772 /*
1773 * create the hash with the exact number of to-be-stored tuplecids from
1774 * the start
1775 */
1781 {
1782 ReorderBufferTupleCidKey key;
1791 /* be careful about padding */
1802 {
1806 }
1807 else
1808 {
1809 /*
1810 * Maybe we already saw this tuple before in this transaction, but
1811 * if so it must have the same cmin.
1812 */
1815 /*
1816 * cmax may be initially invalid, but once set it can only grow,
1817 * and never become invalid again.
1818 */
1823 }
1824 }
1825 }
1827 /*
1828 * Copy a provided snapshot so we can modify it privately. This is needed so
1829 * that catalog modifying transactions can look into intermediate catalog
1830 * states.
1831 */
1832 static Snapshot
1835 {
1836 Snapshot snap;
1837 dlist_iter iter;
1839 Size size;
1855 /*
1856 * snap->subxip contains all txids that belong to our transaction which we
1857 * need to check via cmin/cmax. That's why we store the toplevel
1858 * transaction in there as well.
1859 */
1863 /*
1864 * subxcnt isn't decreased when subtransactions abort, so count manually.
1865 * Since it's an upper boundary it is safe to use it for the allocation
1866 * above.
1867 */
1871 {
1877 }
1879 /* sort so we can bsearch() later */
1882 /* store the specified current CommandId */
1886 }
1888 /*
1889 * Free a previously ReorderBufferCopySnap'ed snapshot
1890 */
1891 static void
1893 {
1896 else
1898 }
1900 /*
1901 * If the transaction was (partially) streamed, we need to prepare or commit
1902 * it in a 'streamed' way. That is, we first stream the remaining part of the
1903 * transaction, and then invoke stream_prepare or stream_commit message as per
1904 * the case.
1905 */
1906 static void
1908 {
1909 /* we should only call this for previously streamed transactions */
1915 {
1916 /*
1917 * Note, we send stream prepare even if a concurrent abort is
1918 * detected. See DecodePrepare for more information.
1919 */
1922 /*
1923 * This is a PREPARED transaction, part of a two-phase commit. The
1924 * full cleanup will happen as part of the COMMIT PREPAREDs, so now
1925 * just truncate txn by removing changes and tuplecids.
1926 */
1928 /* Reset the CheckXidAlive */
1930 }
1931 else
1932 {
1935 }
1936 }
1938 /*
1939 * Set xid to detect concurrent aborts.
1940 *
1941 * While streaming an in-progress transaction or decoding a prepared
1942 * transaction there is a possibility that the (sub)transaction might get
1943 * aborted concurrently. In such case if the (sub)transaction has catalog
1944 * update then we might decode the tuple using wrong catalog version. For
1945 * example, suppose there is one catalog tuple with (xmin: 500, xmax: 0). Now,
1946 * the transaction 501 updates the catalog tuple and after that we will have
1947 * two tuples (xmin: 500, xmax: 501) and (xmin: 501, xmax: 0). Now, if 501 is
1948 * aborted and some other transaction say 502 updates the same catalog tuple
1949 * then the first tuple will be changed to (xmin: 500, xmax: 502). So, the
1950 * problem is that when we try to decode the tuple inserted/updated in 501
1951 * after the catalog update, we will see the catalog tuple with (xmin: 500,
1952 * xmax: 502) as visible because it will consider that the tuple is deleted by
1953 * xid 502 which is not visible to our snapshot. And when we will try to
1954 * decode with that catalog tuple, it can lead to a wrong result or a crash.
1955 * So, it is necessary to detect concurrent aborts to allow streaming of
1956 * in-progress transactions or decoding of prepared transactions.
1957 *
1958 * For detecting the concurrent abort we set CheckXidAlive to the current
1959 * (sub)transaction's xid for which this change belongs to. And, during
1960 * catalog scan we can check the status of the xid and if it is aborted we will
1961 * report a specific error so that we can stop streaming current transaction
1962 * and discard the already streamed changes on such an error. We might have
1963 * already streamed some of the changes for the aborted (sub)transaction, but
1964 * that is fine because when we decode the abort we will stream abort message
1965 * to truncate the changes in the subscriber. Similarly, for prepared
1966 * transactions, we stop decoding if concurrent abort is detected and then
1967 * rollback the changes when rollback prepared is encountered. See
1968 * DecodePrepare.
1969 */
1972 {
1973 /*
1974 * If the input transaction id is already set as a CheckXidAlive then
1975 * nothing to do.
1976 */
1980 /*
1981 * setup CheckXidAlive if it's not committed yet. We don't check if the
1982 * xid is aborted. That will happen during catalog access.
1983 */
1986 else
1988 }
1990 /*
1991 * Helper function for ReorderBufferProcessTXN for applying change.
1992 */
1997 {
2000 else
2002 }
2004 /*
2005 * Helper function for ReorderBufferProcessTXN for applying the truncate.
2006 */
2011 {
2014 else
2016 }
2018 /*
2019 * Helper function for ReorderBufferProcessTXN for applying the message.
2020 */
2024 {
2030 else
2035 }
2037 /*
2038 * Function to store the command id and snapshot at the end of the current
2039 * stream so that we can reuse the same while sending the next stream.
2040 */
2044 {
2047 /* Avoid copying if it's already copied. */
2050 else
2053 }
2055 /*
2056 * Helper function for ReorderBufferProcessTXN to handle the concurrent
2057 * abort of the streaming transaction. This resets the TXN such that it
2058 * can be used to stream the remaining data of transaction being processed.
2059 * This can happen when the subtransaction is aborted and we still want to
2060 * continue processing the main or other subtransactions data.
2061 */
2062 static void
2064 Snapshot snapshot_now,
2065 CommandId command_id,
2066 XLogRecPtr last_lsn,
2068 {
2069 /* Discard the changes that we just streamed */
2072 /* Free all resources allocated for toast reconstruction */
2075 /* Return the spec insert change if it is not NULL */
2077 {
2080 }
2082 /*
2083 * For the streaming case, stop the stream and remember the command ID and
2084 * snapshot for the streaming run.
2085 */
2087 {
2090 }
2092 /* All changes must be deallocated */
2094 }
2096 /*
2097 * Helper function for ReorderBufferReplay and ReorderBufferStreamTXN.
2098 *
2099 * Send data of a transaction (and its subtransactions) to the
2100 * output plugin. We iterate over the top and subtransactions (using a k-way
2101 * merge) and replay the changes in lsn order.
2102 *
2103 * If streaming is true then data will be sent using stream API.
2104 *
2105 * Note: "volatile" markers on some parameters are to avoid trouble with
2106 * PG_TRY inside the function.
2107 */
2108 static void
2110 XLogRecPtr commit_lsn,
2114 {
2123 /* build data to be able to lookup the CommandIds of catalog tuples */
2126 /* setup the initial snapshot */
2129 /*
2130 * Decoding needs access to syscaches et al., which in turn use
2131 * heavyweight locks and such. Thus we need to have enough state around to
2132 * keep track of those. The easiest way is to simply use a transaction
2133 * internally. That also allows us to easily enforce that nothing writes
2134 * to the database by checking for xid assignments.
2135 *
2136 * When we're called via the SQL SRF there's already a transaction
2137 * started, so start an explicit subtransaction there.
2138 */
2142 {
2145 * changes */
2149 else
2152 /*
2153 * We only need to send begin/begin-prepare for non-streamed
2154 * transactions.
2155 */
2157 {
2160 else
2162 }
2166 {
2168 Oid reloid;
2172 /*
2173 * We can't call start stream callback before processing first
2174 * change.
2175 */
2177 {
2179 {
2183 }
2184 }
2186 /*
2187 * Enforce correct ordering of changes, merged from multiple
2188 * subtransactions. The changes may have the same LSN due to
2189 * MULTI_INSERT xlog records.
2190 */
2195 /*
2196 * Set the current xid to detect concurrent aborts. This is
2197 * required for the cases when we decode the changes before the
2198 * COMMIT record is processed.
2199 */
2201 {
2204 }
2207 {
2210 /*
2211 * Confirmation for speculative insertion arrived. Simply
2212 * use as a normal record. It'll be cleaned up at the end
2213 * of INSERT processing.
2214 */
2221 /* intentionally fall through */
2230 /*
2231 * Mapped catalog tuple without data, emitted while
2232 * catalog table was in the process of being rewritten. We
2233 * can fail to look up the relfilenumber, because the
2234 * relmapper has no "historic" view, in contrast to the
2235 * normal catalog during decoding. Thus repeated rewrites
2236 * can cause a lookup failure. That's OK because we do not
2237 * decode catalog changes anyway. Normally such tuples
2238 * would be skipped over below, but we can't identify
2239 * whether the table should be logically logged without
2240 * mapping the relfilenumber to the oid.
2241 */
2249 MAIN_FORKNUM));
2255 reloid,
2257 MAIN_FORKNUM));
2262 /*
2263 * Ignore temporary heaps created during DDL unless the
2264 * plugin has asked for them.
2265 */
2269 /*
2270 * For now ignore sequence changes entirely. Most of the
2271 * time they don't log changes using records we
2272 * understand, so it doesn't make sense to handle the few
2273 * cases we do.
2274 */
2278 /* user-triggered change */
2280 {
2283 streaming);
2285 /*
2286 * Only clear reassembled toast chunks if we're sure
2287 * they're not required anymore. The creator of the
2288 * tuple tells us.
2289 */
2292 }
2293 /* we're not interested in toast deletions */
2295 {
2296 /*
2297 * Need to reassemble the full toasted Datum in
2298 * memory, to ensure the chunks don't get reused till
2299 * we're done remove it from the list of this
2300 * transaction's changes. Otherwise it will get
2301 * freed/reused while restoring spooled data from
2302 * disk.
2303 */
2308 change);
2309 }
2311 change_done:
2313 /*
2314 * If speculative insertion was confirmed, the record
2315 * isn't needed anymore.
2316 */
2318 {
2321 }
2324 {
2327 }
2332 /*
2333 * Speculative insertions are dealt with by delaying the
2334 * processing of the insert until the confirmation record
2335 * arrives. For that we simply unlink the record from the
2336 * chain, so it does not get freed/reused while restoring
2337 * spooled data from disk.
2338 *
2339 * This is safe in the face of concurrent catalog changes
2340 * because the relevant relation can't be changed between
2341 * speculative insertion and confirmation due to
2342 * CheckTableNotInUse() and locking.
2343 */
2345 /* clear out a pending (and thus failed) speculation */
2347 {
2350 }
2352 /* and memorize the pending insertion */
2359 /*
2360 * Abort for speculative insertion arrived. So cleanup the
2361 * specinsert tuple and toast hash.
2362 *
2363 * Note that we get the spec abort change for each toast
2364 * entry but we need to perform the cleanup only the first
2365 * time we get it for the main table.
2366 */
2368 {
2369 /*
2370 * We must clean the toast hash before processing a
2371 * completely new tuple to avoid confusion about the
2372 * previous tuple's toast chunks.
2373 */
2377 /* We don't need this record anymore. */
2380 }
2384 {
2392 {
2394 Relation rel;
2405 }
2407 /* Apply the truncate. */
2410 streaming);
2416 }
2423 /* Execute the invalidation messages locally */
2429 /* get rid of the old */
2433 {
2435 snapshot_now =
2438 }
2440 /*
2441 * Restored from disk, need to be careful not to double
2442 * free. We could introduce refcounting for that, but for
2443 * now this seems infrequent enough not to care.
2444 */
2446 {
2447 snapshot_now =
2450 }
2451 else
2452 {
2454 }
2456 /* and continue with the new one */
2464 {
2468 {
2469 /* we don't use the global one anymore */
2472 }
2478 }
2485 }
2487 /*
2488 * It is possible that the data is not sent to downstream for a
2489 * long time either because the output plugin filtered it or there
2490 * is a DDL that generates a lot of data that is not processed by
2491 * the plugin. So, in such cases, the downstream can timeout. To
2492 * avoid that we try to send a keepalive message if required.
2493 * Trying to send a keepalive message after every change has some
2494 * overhead, but testing showed there is no noticeable overhead if
2495 * we do it after every ~100 changes.
2496 */
2497 #define CHANGES_THRESHOLD 100
2500 {
2503 }
2504 }
2506 /* speculative insertion record must be freed by now */
2509 /* clean up the iterator */
2513 /*
2514 * Update total transaction count and total bytes processed by the
2515 * transaction and its subtransactions. Ensure to not count the
2516 * streamed transaction multiple times.
2517 *
2518 * Note that the statistics computation has to be done after
2519 * ReorderBufferIterTXNFinish as it releases the serialized change
2520 * which we have already accounted in ReorderBufferIterTXNNext.
2521 */
2527 /*
2528 * Done with current changes, send the last message for this set of
2529 * changes depending upon streaming mode.
2530 */
2532 {
2534 {
2537 }
2538 }
2539 else
2540 {
2541 /*
2542 * Call either PREPARE (for two-phase transactions) or COMMIT (for
2543 * regular ones).
2544 */
2547 else
2549 }
2551 /* this is just a sanity check against bad output plugin behaviour */
2556 /*
2557 * Remember the command ID and snapshot for the next set of changes in
2558 * streaming mode.
2559 */
2565 /* cleanup */
2568 /*
2569 * Aborting the current (sub-)transaction as a whole has the right
2570 * semantics. We want all locks acquired in here to be released, not
2571 * reassigned to the parent and we do not want any database access
2572 * have persistent effects.
2573 */
2576 /* make sure there's no cache pollution */
2582 /*
2583 * We are here due to one of the four reasons: 1. Decoding an
2584 * in-progress txn. 2. Decoding a prepared txn. 3. Decoding of a
2585 * prepared txn that was (partially) streamed. 4. Decoding a committed
2586 * txn.
2587 *
2588 * For 1, we allow truncation of txn data by removing the changes
2589 * already streamed but still keeping other things like invalidations,
2590 * snapshot, and tuplecids. For 2 and 3, we indicate
2591 * ReorderBufferTruncateTXN to do more elaborate truncation of txn
2592 * data as the entire transaction has been decoded except for commit.
2593 * For 4, as the entire txn has been decoded, we can fully clean up
2594 * the TXN reorder buffer.
2595 */
2597 {
2599 /* Reset the CheckXidAlive */
2601 }
2602 else
2604 }
2606 {
2610 /* TODO: Encapsulate cleanup from the PG_TRY and PG_CATCH blocks */
2616 /*
2617 * Force cache invalidation to happen outside of a valid transaction
2618 * to prevent catalog access as we just caught an error.
2619 */
2622 /* make sure there's no cache pollution */
2629 /*
2630 * The error code ERRCODE_TRANSACTION_ROLLBACK indicates a concurrent
2631 * abort of the (sub)transaction we are streaming or preparing. We
2632 * need to do the cleanup and return gracefully on this error, see
2633 * SetupCheckXidLive.
2634 *
2635 * This error code can be thrown by one of the callbacks we call
2636 * during decoding so we need to ensure that we return gracefully only
2637 * when we are sending the data in streaming mode and the streaming is
2638 * not finished yet or when we are sending the data out on a PREPARE
2639 * during a two-phase commit.
2640 */
2643 {
2644 /* curtxn must be set for streaming or prepared transactions */
2647 /* Cleanup the temporary error state. */
2653 /* Reset the TXN so that it is allowed to stream remaining data. */
2656 specinsert);
2657 }
2658 else
2659 {
2663 }
2664 }
2666 }
2668 /*
2669 * Perform the replay of a transaction and its non-aborted subtransactions.
2670 *
2671 * Subtransactions previously have to be processed by
2672 * ReorderBufferCommitChild(), even if previously assigned to the toplevel
2673 * transaction with ReorderBufferAssignChild.
2674 *
2675 * This interface is called once a prepare or toplevel commit is read for both
2676 * streamed as well as non-streamed transactions.
2677 */
2678 static void
2682 TimestampTz commit_time,
2684 {
2685 Snapshot snapshot_now;
2694 /*
2695 * If the transaction was (partially) streamed, we need to commit it in a
2696 * 'streamed' way. That is, we first stream the remaining part of the
2697 * transaction, and then invoke stream_commit message.
2698 *
2699 * Called after everything (origin ID, LSN, ...) is stored in the
2700 * transaction to avoid passing that information directly.
2701 */
2703 {
2706 }
2708 /*
2709 * If this transaction has no snapshot, it didn't make any changes to the
2710 * database, so there's nothing to decode. Note that
2711 * ReorderBufferCommitChild will have transferred any snapshots from
2712 * subtransactions if there were any.
2713 */
2715 {
2718 /*
2719 * Removing this txn before a commit might result in the computation
2720 * of an incorrect restart_lsn. See SnapBuildProcessRunningXacts.
2721 */
2725 }
2729 /* Process and send the changes to output plugin. */
2732 }
2734 /*
2735 * Commit a transaction.
2736 *
2737 * See comments for ReorderBufferReplay().
2738 */
2739 void
2742 TimestampTz commit_time,
2744 {
2750 /* unknown transaction, nothing to replay */
2756 }
2758 /*
2759 * Record the prepare information for a transaction.
2760 */
2761 bool
2764 TimestampTz prepare_time,
2766 {
2771 /* unknown transaction, nothing to do */
2775 /*
2776 * Remember the prepare information to be later used by commit prepared in
2777 * case we skip doing prepare.
2778 */
2786 }
2788 /* Remember that we have skipped prepare */
2789 void
2791 {
2796 /* unknown transaction, nothing to do */
2801 }
2803 /*
2804 * Prepare a two-phase transaction.
2805 *
2806 * See comments for ReorderBufferReplay().
2807 */
2808 void
2811 {
2817 /* unknown transaction, nothing to replay */
2824 /* The prepare info must have been updated in txn by now. */
2830 /*
2831 * We send the prepare for the concurrently aborted xacts so that later
2832 * when rollback prepared is decoded and sent, the downstream should be
2833 * able to rollback such a xact. See comments atop DecodePrepare.
2834 *
2835 * Note, for the concurrent_abort + streaming case a stream_prepare was
2836 * already sent within the ReorderBufferReplay call above.
2837 */
2840 }
2842 /*
2843 * This is used to handle COMMIT/ROLLBACK PREPARED.
2844 */
2845 void
2848 XLogRecPtr two_phase_at,
2851 {
2853 XLogRecPtr prepare_end_lsn;
2854 TimestampTz prepare_time;
2858 /* unknown transaction, nothing to do */
2862 /*
2863 * By this time the txn has the prepare record information, remember it to
2864 * be later used for rollback.
2865 */
2869 /* add the gid in the txn */
2872 /*
2873 * It is possible that this transaction is not decoded at prepare time
2874 * either because by that time we didn't have a consistent snapshot, or
2875 * two_phase was not enabled, or it was decoded earlier but we have
2876 * restarted. We only need to send the prepare if it was not decoded
2877 * earlier. We don't need to decode the xact for aborts if it is not done
2878 * already.
2879 */
2881 {
2884 /*
2885 * The prepare info must have been updated in txn even if we skip
2886 * prepare.
2887 */
2890 /*
2891 * By this time the txn has the prepare record information and it is
2892 * important to use that so that downstream gets the accurate
2893 * information. If instead, we have passed commit information here
2894 * then downstream can behave as it has already replayed commit
2895 * prepared after the restart.
2896 */
2899 }
2909 else
2912 /* cleanup: make sure there's no cache pollution */
2916 }
2918 /*
2919 * Abort a transaction that possibly has previous changes. Needs to be first
2920 * called for subtransactions and then for the toplevel xid.
2921 *
2922 * NB: Transactions handled here have to have actively aborted (i.e. have
2923 * produced an abort record). Implicitly aborted transactions are handled via
2924 * ReorderBufferAbortOld(); transactions we're just not interested in, but
2925 * which have committed are handled in ReorderBufferForget().
2926 *
2927 * This function purges this transaction and its contents from memory and
2928 * disk.
2929 */
2930 void
2932 TimestampTz abort_time)
2933 {
2939 /* unknown, nothing to remove */
2945 /* For streamed transactions notify the remote node about the abort. */
2947 {
2950 /*
2951 * We might have decoded changes for this transaction that could load
2952 * the cache as per the current transaction's view (consider DDL's
2953 * happened in this transaction). We don't want the decoding of future
2954 * transactions to use those cache entries so execute invalidations.
2955 */
2959 }
2961 /* cosmetic... */
2964 /* remove potential on-disk data, and deallocate */
2966 }
2968 /*
2969 * Abort all transactions that aren't actually running anymore because the
2970 * server restarted.
2971 *
2972 * NB: These really have to be transactions that have aborted due to a server
2973 * crash/immediate restart, as we don't deal with invalidations here.
2974 */
2975 void
2977 {
2978 dlist_mutable_iter it;
2980 /*
2981 * Iterate through all (potential) toplevel TXNs and abort all that are
2982 * older than what possibly can be running. Once we've found the first
2983 * that is alive we stop, there might be some that acquired an xid earlier
2984 * but started writing later, but it's unlikely and they will be cleaned
2985 * up in a later call to this function.
2986 */
2988 {
2994 {
2997 /* Notify the remote node about the crash/immediate restart. */
3001 /* remove potential on-disk data, and deallocate this tx */
3003 }
3004 else
3006 }
3007 }
3009 /*
3010 * Forget the contents of a transaction if we aren't interested in its
3011 * contents. Needs to be first called for subtransactions and then for the
3012 * toplevel xid.
3013 *
3014 * This is significantly different to ReorderBufferAbort() because
3015 * transactions that have committed need to be treated differently from aborted
3016 * ones since they may have modified the catalog.
3017 *
3018 * Note that this is only allowed to be called in the moment a transaction
3019 * commit has just been read, not earlier; otherwise later records referring
3020 * to this xid might re-create the transaction incompletely.
3021 */
3022 void
3024 {
3030 /* unknown, nothing to forget */
3034 /* this transaction mustn't be streamed */
3037 /* cosmetic... */
3040 /*
3041 * Process cache invalidation messages if there are any. Even if we're not
3042 * interested in the transaction's contents, it could have manipulated the
3043 * catalog and we need to update the caches according to that.
3044 */
3048 else
3051 /* remove potential on-disk data, and deallocate */
3053 }
3055 /*
3056 * Invalidate cache for those transactions that need to be skipped just in case
3057 * catalogs were manipulated as part of the transaction.
3058 *
3059 * Note that this is a special-purpose function for prepared transactions where
3060 * we don't want to clean up the TXN even when we decide to skip it. See
3061 * DecodePrepare.
3062 */
3063 void
3065 {
3071 /* unknown, nothing to do */
3075 /*
3076 * Process cache invalidation messages if there are any. Even if we're not
3077 * interested in the transaction's contents, it could have manipulated the
3078 * catalog and we need to update the caches according to that.
3079 */
3083 else
3085 }
3088 /*
3089 * Execute invalidations happening outside the context of a decoded
3090 * transaction. That currently happens either for xid-less commits
3091 * (cf. RecordTransactionCommit()) or for invalidations in uninteresting
3092 * transactions (via ReorderBufferForget()).
3093 */
3094 void
3097 {
3104 /*
3105 * Force invalidations to happen outside of a valid transaction - that way
3106 * entries will just be marked as invalid without accessing the catalog.
3107 * That's advantageous because we don't need to setup the full state
3108 * necessary for catalog access.
3109 */
3118 }
3120 /*
3121 * Tell reorderbuffer about an xid seen in the WAL stream. Has to be called at
3122 * least once for every xid in XLogRecord->xl_xid (other places in records
3123 * may, but do not have to be passed through here).
3124 *
3125 * Reorderbuffer keeps some data structures about transactions in LSN order,
3126 * for efficiency. To do that it has to know about when transactions are seen
3127 * first in the WAL. As many types of records are not actually interesting for
3128 * logical decoding, they do not necessarily pass through here.
3129 */
3130 void
3132 {
3133 /* many records won't have an xid assigned, centralize check here */
3136 }
3138 /*
3139 * Add a new snapshot to this transaction that may only used after lsn 'lsn'
3140 * because the previous snapshot doesn't describe the catalog correctly for
3141 * following rows.
3142 */
3143 void
3146 {
3153 }
3155 /*
3156 * Set up the transaction's base snapshot.
3157 *
3158 * If we know that xid is a subtransaction, set the base snapshot on the
3159 * top-level transaction instead.
3160 */
3161 void
3164 {
3170 /*
3171 * Fetch the transaction to operate on. If we know it's a subtransaction,
3172 * operate on its top-level transaction instead.
3173 */
3185 }
3187 /*
3188 * Access the catalog with this CommandId at this point in the changestream.
3189 *
3190 * May only be called for command ids > 1
3191 */
3192 void
3195 {
3202 }
3204 /*
3205 * Update memory counters to account for the new or removed change.
3206 *
3207 * We update two counters - in the reorder buffer, and in the transaction
3208 * containing the change. The reorder buffer counter allows us to quickly
3209 * decide if we reached the memory limit, the transaction counter allows
3210 * us to quickly pick the largest transaction for eviction.
3211 *
3212 * Either txn or change must be non-NULL at least. We update the memory
3213 * counter of txn if it's non-NULL, otherwise change->txn.
3214 *
3215 * When streaming is enabled, we need to update the toplevel transaction
3216 * counters instead - we don't really care about subtransactions as we
3217 * can't stream them individually anyway, and we only pick toplevel
3218 * transactions for eviction. So only toplevel transactions matter.
3219 */
3220 static void
3225 {
3230 /*
3231 * Ignore tuple CID changes, because those are not evicted when reaching
3232 * memory limit. So we just don't count them, because it might easily
3233 * trigger a pointless attempt to spill.
3234 */
3245 /*
3246 * Update the total size in top level as well. This is later used to
3247 * compute the decoding stats.
3248 */
3252 {
3258 /* Update the total size in the top transaction. */
3261 /* Update the max-heap */
3265 }
3266 else
3267 {
3272 /* Update the total size in the top transaction. */
3275 /* Update the max-heap */
3279 }
3282 }
3284 /*
3285 * Add new (relfilelocator, tid) -> (cmin, cmax) mappings.
3286 *
3287 * We do not include this change type in memory accounting, because we
3288 * keep CIDs in a separate list and do not evict them when reaching
3289 * the memory limit.
3290 */
3291 void
3296 {
3313 }
3315 /*
3316 * Accumulate the invalidations for executing them later.
3317 *
3318 * This needs to be called for each XLOG_XACT_INVALIDATIONS message and
3319 * accumulates all the invalidation messages in the toplevel transaction, if
3320 * available, otherwise in the current transaction, as well as in the form of
3321 * change in reorder buffer. We require to record it in form of the change
3322 * so that we can execute only the required invalidations instead of executing
3323 * all the invalidations on each CommandId increment. We also need to
3324 * accumulate these in the txn buffer because in some cases where we skip
3325 * processing the transaction (see ReorderBufferForget), we need to execute
3326 * all the invalidations together.
3327 */
3328 void
3332 {
3334 MemoryContext oldcontext;
3341 /*
3342 * Collect all the invalidations under the top transaction, if available,
3343 * so that we can execute them all together. See comments atop this
3344 * function.
3345 */
3350 /* Accumulate invalidations. */
3352 {
3358 }
3359 else
3360 {
3368 }
3381 }
3383 /*
3384 * Apply all invalidations we know. Possibly we only need parts at this point
3385 * in the changestream but we don't know which those are.
3386 */
3387 static void
3389 {
3394 }
3396 /*
3397 * Mark a transaction as containing catalog changes
3398 */
3399 void
3401 XLogRecPtr lsn)
3402 {
3408 {
3411 }
3413 /*
3414 * Mark top-level transaction as having catalog changes too if one of its
3415 * children has so that the ReorderBufferBuildTupleCidHash can
3416 * conveniently check just top-level transaction and decide whether to
3417 * build the hash table or not.
3418 */
3420 {
3424 {
3427 }
3428 }
3429 }
3431 /*
3432 * Return palloc'ed array of the transactions that have changed catalogs.
3433 * The returned array is sorted in xidComparator order.
3434 *
3435 * The caller must free the returned array when done with it.
3436 */
3437 TransactionId *
3439 {
3440 dlist_iter iter;
3444 /* Quick return if the list is empty */
3448 /* Initialize XID array */
3452 {
3454 catchange_node,
3460 }
3466 }
3468 /*
3469 * Query whether a transaction is already *known* to contain catalog
3470 * changes. This can be wrong until directly before the commit!
3471 */
3472 bool
3474 {
3483 }
3485 /*
3486 * ReorderBufferXidHasBaseSnapshot
3487 * Have we already set the base snapshot for the given txn/subtxn?
3488 */
3489 bool
3491 {
3497 /* transaction isn't known yet, ergo no snapshot */
3501 /* a known subtxn? operate on top-level txn instead */
3507 }
3510 /*
3511 * ---------------------------------------
3512 * Disk serialization support
3513 * ---------------------------------------
3514 */
3516 /*
3517 * Ensure the IO buffer is >= sz.
3518 */
3519 static void
3521 {
3523 {
3526 }
3528 {
3531 }
3532 }
3535 /* Compare two transactions by size */
3536 static int
3538 {
3547 }
3549 /*
3550 * Find the largest transaction (toplevel or subxact) to evict (spill to disk).
3551 */
3554 {
3557 /* Get the largest transaction from the max-heap */
3566 }
3568 /*
3569 * Find the largest streamable toplevel transaction to evict (by streaming).
3570 *
3571 * This can be seen as an optimized version of ReorderBufferLargestTXN, which
3572 * should give us the same transaction (because we don't update memory account
3573 * for subtransaction with streaming, so it's always 0). But we can simply
3574 * iterate over the limited number of toplevel transactions that have a base
3575 * snapshot. There is no use of selecting a transaction that doesn't have base
3576 * snapshot because we don't decode such transactions. Also, we do not select
3577 * the transaction which doesn't have any streamable change.
3578 *
3579 * Note that, we skip transactions that contain incomplete changes. There
3580 * is a scope of optimization here such that we can select the largest
3581 * transaction which has incomplete changes. But that will make the code and
3582 * design quite complex and that might not be worth the benefit. If we plan to
3583 * stream the transactions that contain incomplete changes then we need to
3584 * find a way to partially stream/truncate the transaction changes in-memory
3585 * and build a mechanism to partially truncate the spilled files.
3586 * Additionally, whenever we partially stream the transaction we need to
3587 * maintain the last streamed lsn and next time we need to restore from that
3588 * segment and the offset in WAL. As we stream the changes from the top
3589 * transaction and restore them subtransaction wise, we need to even remember
3590 * the subxact from where we streamed the last change.
3591 */
3594 {
3595 dlist_iter iter;
3599 /* Find the largest top-level transaction having a base snapshot. */
3601 {
3606 /* must not be a subtxn */
3608 /* base_snapshot must be set */
3614 {
3617 }
3618 }
3621 }
3623 /*
3624 * Check whether the logical_decoding_work_mem limit was reached, and if yes
3625 * pick the largest (sub)transaction at-a-time to evict and spill its changes to
3626 * disk or send to the output plugin until we reach under the memory limit.
3627 *
3628 * If debug_logical_replication_streaming is set to "immediate", stream or
3629 * serialize the changes immediately.
3630 *
3631 * XXX At this point we select the transactions until we reach under the memory
3632 * limit, but we might also adapt a more elaborate eviction strategy - for example
3633 * evicting enough transactions to free certain fraction (e.g. 50%) of the memory
3634 * limit.
3635 */
3636 static void
3638 {
3641 /*
3642 * Bail out if debug_logical_replication_streaming is buffered and we
3643 * haven't exceeded the memory limit.
3644 */
3649 /*
3650 * If debug_logical_replication_streaming is immediate, loop until there's
3651 * no change. Otherwise, loop until we reach under the memory limit. One
3652 * might think that just by evicting the largest (sub)transaction we will
3653 * come under the memory limit based on assumption that the selected
3654 * transaction is at least as large as the most recent change (which
3655 * caused us to go over the memory limit). However, that is not true
3656 * because a user can reduce the logical_decoding_work_mem to a smaller
3657 * value before the most recent change.
3658 */
3662 {
3663 /*
3664 * Pick the largest transaction and evict it from memory by streaming,
3665 * if possible. Otherwise, spill to disk.
3666 */
3669 {
3670 /* we know there has to be one, because the size is not zero */
3676 }
3677 else
3678 {
3679 /*
3680 * Pick the largest transaction (or subtransaction) and evict it
3681 * from memory by serializing it to disk.
3682 */
3685 /* we know there has to be one, because the size is not zero */
3691 }
3693 /*
3694 * After eviction, the transaction should have no entries in memory,
3695 * and should use 0 bytes for changes.
3696 */
3699 }
3701 /* We must be under the memory limit now. */
3704 }
3706 /*
3707 * Spill data of a large transaction (and its subtransactions) to disk.
3708 */
3709 static void
3711 {
3712 dlist_iter subtxn_i;
3713 dlist_mutable_iter change_i;
3722 /* do the same to all child TXs */
3724 {
3729 }
3731 /* serialize changestream */
3733 {
3738 /*
3739 * store in segment in which it belongs by start lsn, don't split over
3740 * multiple segments tho
3741 */
3744 {
3752 /*
3753 * No need to care about TLIs here, only used during a single run,
3754 * so each LSN only maps to a specific WAL record.
3755 */
3757 curOpenSegNo);
3759 /* open segment, create it if necessary */
3767 }
3773 spilled++;
3774 }
3776 /* Update the memory counter */
3779 /* update the statistics iff we have spilled anything */
3781 {
3785 /* don't consider already serialized transactions */
3788 /* update the decoding stats */
3790 }
3799 }
3801 /*
3802 * Serialize individual change to disk.
3803 */
3804 static void
3807 {
3817 {
3818 /* fall through these, they're all similar enough */
3823 {
3825 HeapTuple oldtup,
3826 newtup;
3834 {
3838 }
3841 {
3845 }
3847 /* make sure we have enough space */
3851 /* might have been reallocated above */
3855 {
3861 }
3864 {
3870 }
3872 }
3874 {
3884 /* might have been reallocated above */
3887 /* write the prefix including the size */
3891 prefix_size);
3894 /* write the message including the size */
3902 }
3904 {
3914 /* might have been reallocated above */
3920 }
3922 {
3923 Snapshot snap;
3932 /* make sure we have enough space */
3935 /* might have been reallocated above */
3942 {
3946 }
3949 {
3953 }
3955 }
3957 {
3958 Size size;
3961 /* account for the OIDs of truncated relations */
3965 /* make sure we have enough space */
3969 /* might have been reallocated above */
3976 }
3981 /* ReorderBufferChange contains everything important */
3983 }
3990 {
3995 /* if write didn't set errno, assume problem is no disk space */
4001 }
4004 /*
4005 * Keep the transaction's final_lsn up to date with each change we send to
4006 * disk, so that ReorderBufferRestoreCleanup works correctly. (We used to
4007 * only do this on commit and abort records, but that doesn't work if a
4008 * system crash leaves a transaction without its abort record).
4009 *
4010 * Make sure not to move it backwards.
4011 */
4016 }
4018 /* Returns true, if the output plugin supports streaming, false, otherwise. */
4021 {
4025 }
4027 /* Returns true, if the streaming can be started now, false, otherwise. */
4030 {
4034 /* We can't start streaming unless a consistent state is reached. */
4038 /*
4039 * We can't start streaming immediately even if the streaming is enabled
4040 * because we previously decoded this transaction and now just are
4041 * restarting.
4042 */
4048 }
4050 /*
4051 * Send data of a large transaction (and its subtransactions) to the
4052 * output plugin, but using the stream API.
4053 */
4054 static void
4056 {
4057 Snapshot snapshot_now;
4058 CommandId command_id;
4059 Size stream_bytes;
4062 /* We can never reach here for a subtransaction. */
4065 /*
4066 * We can't make any assumptions about base snapshot here, similar to what
4067 * ReorderBufferCommit() does. That relies on base_snapshot getting
4068 * transferred from subxact in ReorderBufferCommitChild(), but that was
4069 * not yet called as the transaction is in-progress.
4070 *
4071 * So just walk the subxacts and use the same logic here. But we only need
4072 * to do that once, when the transaction is streamed for the first time.
4073 * After that we need to reuse the snapshot from the previous run.
4074 *
4075 * Unlike DecodeCommit which adds xids of all the subtransactions in
4076 * snapshot's xip array via SnapBuildCommitTxn, we can't do that here but
4077 * we do add them to subxip array instead via ReorderBufferCopySnap. This
4078 * allows the catalog changes made in subtransactions decoded till now to
4079 * be visible.
4080 */
4082 {
4083 dlist_iter subxact_i;
4085 /* make sure this transaction is streamed for the first time */
4088 /* at the beginning we should have invalid command ID */
4092 {
4097 }
4099 /*
4100 * If this transaction has no snapshot, it didn't make any changes to
4101 * the database till now, so there's nothing to decode.
4102 */
4104 {
4107 }
4112 }
4113 else
4114 {
4115 /* the transaction must have been already streamed */
4118 /*
4119 * Nah, we already have snapshot from the previous streaming run. We
4120 * assume new subxacts can't move the LSN backwards, and so can't beat
4121 * the LSN condition in the previous branch (so no need to walk
4122 * through subxacts again). In fact, we must not do that as we may be
4123 * using snapshot half-way through the subxact.
4124 */
4127 /*
4128 * We can't use txn->snapshot_now directly because after the last
4129 * streaming run, we might have got some new sub-transactions. So we
4130 * need to add them to the snapshot.
4131 */
4135 /* Free the previously copied snapshot. */
4139 }
4141 /*
4142 * Remember this information to be used later to update stats. We can't
4143 * update the stats here as an error while processing the changes would
4144 * lead to the accumulation of stats even though we haven't streamed all
4145 * the changes.
4146 */
4150 /* Process and send the changes to output plugin. */
4157 /* Don't consider already streamed transaction. */
4160 /* update the decoding stats */
4166 }
4168 /*
4169 * Size of a change in memory.
4170 */
4171 static Size
4173 {
4177 {
4178 /* fall through these, they're all similar enough */
4183 {
4184 HeapTuple oldtup,
4185 newtup;
4193 {
4197 }
4200 {
4204 }
4207 }
4209 {
4216 }
4218 {
4222 }
4224 {
4225 Snapshot snap;
4234 }
4236 {
4240 }
4245 /* ReorderBufferChange contains everything important */
4247 }
4250 }
4253 /*
4254 * Restore a number of changes spilled to disk back into memory.
4255 */
4256 static Size
4259 {
4261 XLogSegNo last_segno;
4262 dlist_mutable_iter cleanup_iter;
4268 /* free current entries, so we have memory for more */
4270 {
4276 }
4283 {
4290 {
4293 /* first time in */
4299 /*
4300 * No need to care about TLIs here, only used during a single run,
4301 * so each LSN only maps to a specific WAL record.
4302 */
4308 /* No harm in resetting the offset even in case of failure */
4312 {
4316 }
4321 path)));
4322 }
4324 /*
4325 * Read the statically sized part of a change which has information
4326 * about the total size. If we couldn't read a record, we're at the
4327 * end of this file.
4328 */
4334 /* eof */
4336 {
4341 }
4350 readBytes,
4365 WAIT_EVENT_REORDER_BUFFER_READ);
4375 readBytes,
4380 /*
4381 * ok, read a full change from disk, now restore it into proper
4382 * in-memory format
4383 */
4385 restored++;
4386 }
4389 }
4391 /*
4392 * Convert change from its on-disk format to in-memory format and queue it onto
4393 * the TXN's ->changes list.
4394 *
4395 * Note: although "data" is declared char*, at entry it points to a
4396 * maxalign'd buffer, making it safe in most of this function to assume
4397 * that the pointed-to data is suitably aligned for direct access.
4398 */
4399 static void
4402 {
4410 /* copy static part */
4415 /* restore individual stuff */
4417 {
4418 /* fall through these, they're all similar enough */
4424 {
4430 /* restore ->tuple */
4435 /* reset t_data pointer into the new tuplebuf */
4439 /* restore tuple data itself */
4442 }
4445 {
4446 /* here, data might not be suitably aligned! */
4447 uint32 tuplelen;
4455 /* restore ->tuple */
4460 /* reset t_data pointer into the new tuplebuf */
4464 /* restore tuple data itself */
4467 }
4471 {
4472 Size prefix_size;
4474 /* read prefix */
4478 prefix_size);
4483 /* read the message */
4493 }
4495 {
4502 /* read the message */
4506 }
4508 {
4509 Snapshot oldsnap;
4510 Snapshot newsnap;
4511 Size size;
4529 }
4530 /* the base struct contains all the data, easy peasy */
4532 {
4541 }
4547 }
4552 /*
4553 * Update memory accounting for the restored change. We need to do this
4554 * although we don't check the memory limit when restoring the changes in
4555 * this branch (we only do that when initially queueing the changes after
4556 * decoding), because we will release the changes later, and that will
4557 * update the accounting too (subtracting the size from the counters). And
4558 * we don't want to underflow there.
4559 */
4562 }
4564 /*
4565 * Remove all on-disk stored for the passed in transaction.
4566 */
4567 static void
4569 {
4570 XLogSegNo first;
4571 XLogSegNo cur;
4572 XLogSegNo last;
4580 /* iterate over all possible filenames, and delete them */
4582 {
4590 }
4591 }
4593 /*
4594 * Remove any leftover serialized reorder buffers from a slot directory after a
4595 * prior crash or decoding session exit.
4596 */
4597 static void
4599 {
4607 /* we're only handling directories here, skip if it's not ours */
4613 {
4614 /* only look at names that can be ours */
4616 {
4626 }
4627 }
4629 }
4631 /*
4632 * Given a replication slot, transaction ID and segment number, fill in the
4633 * corresponding spill file into 'path', which is a caller-owned buffer of size
4634 * at least MAXPGPATH.
4635 */
4636 static void
4638 XLogSegNo segno)
4639 {
4640 XLogRecPtr recptr;
4645 PG_REPLSLOT_DIR,
4648 }
4650 /*
4651 * Delete all data spilled to disk after we've restarted/crashed. It will be
4652 * recreated when the respective slots are reused.
4653 */
4654 void
4656 {
4662 {
4667 /* if it cannot be a slot, skip the directory */
4671 /*
4672 * ok, has to be a surviving logical slot, iterate and delete
4673 * everything starting with xid-*
4674 */
4676 }
4678 }
4680 /* ---------------------------------------
4681 * toast reassembly support
4682 * ---------------------------------------
4683 */
4685 /*
4686 * Initialize per tuple toast reconstruction support.
4687 */
4688 static void
4690 {
4691 HASHCTL hash_ctl;
4700 }
4702 /*
4703 * Per toast-chunk handling for toast reconstruction
4704 *
4705 * Appends a toast chunk so we can reconstruct it when the tuple "owning" the
4706 * toasted Datum comes along.
4707 */
4708 static void
4711 {
4713 HeapTuple newtup;
4715 int32 chunksize;
4717 Pointer chunk;
4719 Oid chunk_id;
4720 int32 chunk_seq;
4737 {
4748 }
4756 /* calculate size so we can allocate the right size at once later */
4760 /* could happen due to heap_form_tuple doing its thing */
4762 else
4769 }
4771 /*
4772 * Rejigger change->newtuple to point to in-memory toast tuples instead of
4773 * on-disk toast tuples that may no longer exist (think DROP TABLE or VACUUM).
4774 *
4775 * We cannot replace unchanged toast tuples though, so those will still point
4776 * to on-disk toast data.
4777 *
4778 * While updating the existing change with detoasted tuple data, we need to
4779 * update the memory accounting info, because the change size will differ.
4780 * Otherwise the accounting may get out of sync, triggering serialization
4781 * at unexpected times.
4782 *
4783 * We simply subtract size of the change before rejiggering the tuple, and
4784 * then add the new size. This makes it look like the change was removed
4785 * and then added back, except it only tweaks the accounting info.
4786 *
4787 * In particular it can't trigger serialization, which would be pointless
4788 * anyway as it happens during commit processing right before handing
4789 * the change to the output plugin.
4790 */
4791 static void
4794 {
4795 TupleDesc desc;
4800 HeapTuple tmphtup;
4801 Relation toast_rel;
4802 TupleDesc toast_desc;
4803 MemoryContext oldcontext;
4804 HeapTuple newtup;
4805 Size old_size;
4807 /* no toast tuples changed */
4811 /*
4812 * We're going to modify the size of the change. So, to make sure the
4813 * accounting is correct we record the current change size and then after
4814 * re-computing the change we'll subtract the recorded size and then
4815 * re-add the new change size at the end. We don't immediately subtract
4816 * the old size because if there is any error before we add the new size,
4817 * we will release the changes and that will update the accounting info
4818 * (subtracting the size from the counters). And we don't want to
4819 * underflow there.
4820 */
4825 /* we should only have toast tuples in an INSERT or UPDATE */
4837 /* should we allocate from stack instead? */
4847 {
4852 /* va_rawsize is the size of the original datum -- including header */
4857 dlist_iter it;
4860 /* system columns aren't toasted */
4867 /* not a varlena datatype */
4871 /* no data */
4875 /* ok, we know we have a toast datum */
4878 /* no need to do anything if the tuple isn't external */
4884 /*
4885 * Check whether the toast tuple changed, replace if so.
4886 */
4890 HASH_FIND,
4891 NULL);
4895 new_datum =
4904 /* stitch toast tuple back together from its parts */
4906 {
4909 HeapTuple ctup;
4910 Pointer chunk;
4924 }
4927 /* make sure its marked as compressed or not */
4930 else
4941 }
4943 /*
4944 * Build tuple in separate memory & copy tuple back into the tuplebuf
4945 * passed to the output plugin. We can't directly heap_fill_tuple() into
4946 * the tuplebuf because attrs[] will point back into the current content.
4947 */
4955 /*
4956 * free resources we won't further need, more persistent stuff will be
4957 * free'd in ReorderBufferToastReset().
4958 */
4962 {
4965 }
4972 /* subtract the old change size */
4974 /* now add the change back, with the correct size */
4977 }
4979 /*
4980 * Free all resources allocated for toast reconstruction.
4981 */
4982 static void
4984 {
4985 HASH_SEQ_STATUS hstat;
4991 /* sequentially walk over the hash and free everything */
4994 {
4995 dlist_mutable_iter it;
5001 {
5007 }
5008 }
5012 }
5015 /* ---------------------------------------
5016 * Visibility support for logical decoding
5017 *
5018 *
5019 * Lookup actual cmin/cmax values when using decoding snapshot. We can't
5020 * always rely on stored cmin/cmax values because of two scenarios:
5021 *
5022 * * A tuple got changed multiple times during a single transaction and thus
5023 * has got a combo CID. Combo CIDs are only valid for the duration of a
5024 * single transaction.
5025 * * A tuple with a cmin but no cmax (and thus no combo CID) got
5026 * deleted/updated in another transaction than the one which created it
5027 * which we are looking at right now. As only one of cmin, cmax or combo CID
5028 * is actually stored in the heap we don't have access to the value we
5029 * need anymore.
5030 *
5031 * To resolve those problems we have a per-transaction hash of (cmin,
5032 * cmax) tuples keyed by (relfilelocator, ctid) which contains the actual
5033 * (cmin, cmax) values. That also takes care of combo CIDs by simply
5034 * not caring about them at all. As we have the real cmin/cmax values
5035 * combo CIDs aren't interesting.
5036 *
5037 * As we only care about catalog tuples here the overhead of this
5038 * hashtable should be acceptable.
5039 *
5040 * Heap rewrites complicate this a bit, check rewriteheap.c for
5041 * details.
5042 * -------------------------------------------------------------------------
5043 */
5045 /* struct for sorting mapping files by LSN efficiently */
5047 {
5048 XLogRecPtr lsn;
5052 #ifdef NOT_USED
5053 static void
5055 {
5056 HASH_SEQ_STATUS hstat;
5061 {
5069 ent->cmax
5070 );
5071 }
5072 }
5073 #endif
5075 /*
5076 * Apply a single mapping file to tuplecid_data.
5077 *
5078 * The mapping file has to have been verified to be a) committed b) for our
5079 * transaction c) applied in LSN order.
5080 */
5081 static void
5083 {
5087 LogicalRewriteMappingData map;
5097 {
5098 ReorderBufferTupleCidKey key;
5103 /* be careful about padding */
5106 /* read all mappings till the end of the file */
5115 path)));
5133 /* no existing mapping, no need to update */
5145 {
5146 /*
5147 * Make sure the existing mapping makes sense. We sometime update
5148 * old records that did not yet have a cmax (e.g. pg_class' own
5149 * entry while rewriting it) during rewrites, so allow that.
5150 */
5153 }
5154 else
5155 {
5156 /* update mapping */
5160 }
5161 }
5167 }
5170 /*
5171 * Check whether the TransactionId 'xid' is in the pre-sorted array 'xip'.
5172 */
5173 static bool
5175 {
5178 }
5180 /*
5181 * list_sort() comparator for sorting RewriteMappingFiles in LSN order.
5182 */
5183 static int
5185 {
5190 }
5192 /*
5193 * Apply any existing logical remapping files if there are any targeted at our
5194 * transaction for relid.
5195 */
5196 static void
5198 {
5207 {
5208 Oid f_dboid;
5209 Oid f_relid;
5210 TransactionId f_mapped_xid;
5211 TransactionId f_create_xid;
5212 XLogRecPtr f_lsn;
5213 uint32 f_hi,
5214 f_lo;
5221 /* Ignore files that aren't ours */
5232 /* mapping for another database */
5236 /* mapping for another relation */
5240 /* did the creating transaction abort? */
5244 /* not for our transaction */
5248 /* ok, relevant, queue for apply */
5253 }
5256 /* sort files so we apply them in LSN order */
5260 {
5267 }
5268 }
5270 /*
5271 * Lookup cmin/cmax of a tuple, during logical decoding where we can't rely on
5272 * combo CIDs.
5273 */
5274 bool
5276 Snapshot snapshot,
5279 {
5280 ReorderBufferTupleCidKey key;
5282 ForkNumber forkno;
5283 BlockNumber blockno;
5286 /*
5287 * Return unresolved if tuplecid_data is not valid. That's because when
5288 * streaming in-progress transactions we may run into tuples with the CID
5289 * before actually decoding them. Think e.g. about INSERT followed by
5290 * TRUNCATE, where the TRUNCATE may not be decoded yet when applying the
5291 * INSERT. So in such cases, we assume the CID is from the future
5292 * command.
5293 */
5297 /* be careful about padding */
5302 /*
5303 * get relfilelocator from the buffer, no convenient way to access it
5304 * other than that.
5305 */
5308 /* tuples can only be in the main fork */
5315 restart:
5319 /*
5320 * failed to find a mapping, check whether the table was rewritten and
5321 * apply mapping if so, but only do that once - there can be no new
5322 * mappings while we are in here since we have to hold a lock on the
5323 * relation.
5324 */
5326 {
5328 /* now check but don't update for a mapping again */
5331 }
5340 }