1 /*-------------------------------------------------------------------------
4 * PostgreSQL logical replay/reorder buffer management
7 * Copyright (c) 2012-2016, PostgreSQL Global Development Group
11 * src/backend/replication/reorderbuffer.c
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 - signalled by reading the transaction commit record - it
18 * will then call the output plugin (c.f. ReorderBufferCommit()) with the
19 * individual changes. The output plugins rely on snapshots built by
20 * snapbuild.c which hands them to us.
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.
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 a large transactions to disk. When the transaction is replayed the
36 * contents of individual (sub-)transactions will be read from disk in
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
46 * -------------------------------------------------------------------------
53 #include "access/rewriteheap.h"
54 #include "access/transam.h"
55 #include "access/tuptoaster.h"
56 #include "access/xact.h"
57 #include "access/xlog_internal.h"
58 #include "catalog/catalog.h"
59 #include "lib/binaryheap.h"
60 #include "miscadmin.h"
61 #include "replication/logical.h"
62 #include "replication/reorderbuffer.h"
63 #include "replication/slot.h"
64 #include "replication/snapbuild.h" /* just for SnapBuildSnapDecRefcount */
65 #include "storage/bufmgr.h"
66 #include "storage/fd.h"
67 #include "storage/sinval.h"
68 #include "utils/builtins.h"
69 #include "utils/combocid.h"
70 #include "utils/memdebug.h"
71 #include "utils/memutils.h"
72 #include "utils/rel.h"
73 #include "utils/relfilenodemap.h"
74 #include "utils/tqual.h"
77 /* entry for a hash table we use to map from xid to our transaction state */
78 typedef struct ReorderBufferTXNByIdEnt
81 ReorderBufferTXN *txn;
82 } ReorderBufferTXNByIdEnt;
84 /* data structures for (relfilenode, ctid) => (cmin, cmax) mapping */
85 typedef struct ReorderBufferTupleCidKey
89 } ReorderBufferTupleCidKey;
91 typedef struct ReorderBufferTupleCidEnt
93 ReorderBufferTupleCidKey key;
96 CommandId combocid; /* just for debugging */
97 } ReorderBufferTupleCidEnt;
99 /* k-way in-order change iteration support structures */
100 typedef struct ReorderBufferIterTXNEntry
103 ReorderBufferChange *change;
104 ReorderBufferTXN *txn;
107 } ReorderBufferIterTXNEntry;
109 typedef struct ReorderBufferIterTXNState
113 dlist_head old_change;
114 ReorderBufferIterTXNEntry entries[FLEXIBLE_ARRAY_MEMBER];
115 } ReorderBufferIterTXNState;
117 /* toast datastructures */
118 typedef struct ReorderBufferToastEnt
120 Oid chunk_id; /* toast_table.chunk_id */
121 int32 last_chunk_seq; /* toast_table.chunk_seq of the last chunk we
123 Size num_chunks; /* number of chunks we've already seen */
124 Size size; /* combined size of chunks seen */
125 dlist_head chunks; /* linked list of chunks */
126 struct varlena *reconstructed; /* reconstructed varlena now pointed
128 } ReorderBufferToastEnt;
130 /* Disk serialization support datastructures */
131 typedef struct ReorderBufferDiskChange
134 ReorderBufferChange change;
136 } ReorderBufferDiskChange;
139 * Maximum number of changes kept in memory, per transaction. After that,
140 * changes are spooled to disk.
142 * The current value should be sufficient to decode the entire transaction
143 * without hitting disk in OLTP workloads, while starting to spool to disk in
144 * other workloads reasonably fast.
146 * At some point in the future it probably makes sense to have a more elaborate
147 * resource management here, but it's not entirely clear what that would look
150 static const Size max_changes_in_memory = 4096;
153 * We use a very simple form of a slab allocator for frequently allocated
154 * objects, simply keeping a fixed number in a linked list when unused,
155 * instead pfree()ing them. Without that in many workloads aset.c becomes a
156 * major bottleneck, especially when spilling to disk while decoding batch
159 static const Size max_cached_changes = 4096 * 2;
160 static const Size max_cached_tuplebufs = 4096 * 2; /* ~8MB */
161 static const Size max_cached_transactions = 512;
164 /* ---------------------------------------
165 * primary reorderbuffer support routines
166 * ---------------------------------------
168 static ReorderBufferTXN *ReorderBufferGetTXN(ReorderBuffer *rb);
169 static void ReorderBufferReturnTXN(ReorderBuffer *rb, ReorderBufferTXN *txn);
170 static ReorderBufferTXN *ReorderBufferTXNByXid(ReorderBuffer *rb,
171 TransactionId xid, bool create, bool *is_new,
172 XLogRecPtr lsn, bool create_as_top);
174 static void AssertTXNLsnOrder(ReorderBuffer *rb);
176 /* ---------------------------------------
177 * support functions for lsn-order iterating over the ->changes of a
178 * transaction and its subtransactions
180 * used for iteration over the k-way heap merge of a transaction and its
182 * ---------------------------------------
184 static ReorderBufferIterTXNState *ReorderBufferIterTXNInit(ReorderBuffer *rb, ReorderBufferTXN *txn);
185 static ReorderBufferChange *
186 ReorderBufferIterTXNNext(ReorderBuffer *rb, ReorderBufferIterTXNState *state);
187 static void ReorderBufferIterTXNFinish(ReorderBuffer *rb,
188 ReorderBufferIterTXNState *state);
189 static void ReorderBufferExecuteInvalidations(ReorderBuffer *rb, ReorderBufferTXN *txn);
192 * ---------------------------------------
193 * Disk serialization support functions
194 * ---------------------------------------
196 static void ReorderBufferCheckSerializeTXN(ReorderBuffer *rb, ReorderBufferTXN *txn);
197 static void ReorderBufferSerializeTXN(ReorderBuffer *rb, ReorderBufferTXN *txn);
198 static void ReorderBufferSerializeChange(ReorderBuffer *rb, ReorderBufferTXN *txn,
199 int fd, ReorderBufferChange *change);
200 static Size ReorderBufferRestoreChanges(ReorderBuffer *rb, ReorderBufferTXN *txn,
201 int *fd, XLogSegNo *segno);
202 static void ReorderBufferRestoreChange(ReorderBuffer *rb, ReorderBufferTXN *txn,
204 static void ReorderBufferRestoreCleanup(ReorderBuffer *rb, ReorderBufferTXN *txn);
206 static void ReorderBufferFreeSnap(ReorderBuffer *rb, Snapshot snap);
207 static Snapshot ReorderBufferCopySnap(ReorderBuffer *rb, Snapshot orig_snap,
208 ReorderBufferTXN *txn, CommandId cid);
210 /* ---------------------------------------
211 * toast reassembly support
212 * ---------------------------------------
214 static void ReorderBufferToastInitHash(ReorderBuffer *rb, ReorderBufferTXN *txn);
215 static void ReorderBufferToastReset(ReorderBuffer *rb, ReorderBufferTXN *txn);
216 static void ReorderBufferToastReplace(ReorderBuffer *rb, ReorderBufferTXN *txn,
217 Relation relation, ReorderBufferChange *change);
218 static void ReorderBufferToastAppendChunk(ReorderBuffer *rb, ReorderBufferTXN *txn,
219 Relation relation, ReorderBufferChange *change);
223 * Allocate a new ReorderBuffer
226 ReorderBufferAllocate(void)
228 ReorderBuffer *buffer;
230 MemoryContext new_ctx;
232 /* allocate memory in own context, to have better accountability */
233 new_ctx = AllocSetContextCreate(CurrentMemoryContext,
235 ALLOCSET_DEFAULT_MINSIZE,
236 ALLOCSET_DEFAULT_INITSIZE,
237 ALLOCSET_DEFAULT_MAXSIZE);
240 (ReorderBuffer *) MemoryContextAlloc(new_ctx, sizeof(ReorderBuffer));
242 memset(&hash_ctl, 0, sizeof(hash_ctl));
244 buffer->context = new_ctx;
246 hash_ctl.keysize = sizeof(TransactionId);
247 hash_ctl.entrysize = sizeof(ReorderBufferTXNByIdEnt);
248 hash_ctl.hcxt = buffer->context;
250 buffer->by_txn = hash_create("ReorderBufferByXid", 1000, &hash_ctl,
251 HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
253 buffer->by_txn_last_xid = InvalidTransactionId;
254 buffer->by_txn_last_txn = NULL;
256 buffer->nr_cached_transactions = 0;
257 buffer->nr_cached_changes = 0;
258 buffer->nr_cached_tuplebufs = 0;
260 buffer->outbuf = NULL;
261 buffer->outbufsize = 0;
263 buffer->current_restart_decoding_lsn = InvalidXLogRecPtr;
265 dlist_init(&buffer->toplevel_by_lsn);
266 dlist_init(&buffer->cached_transactions);
267 dlist_init(&buffer->cached_changes);
268 slist_init(&buffer->cached_tuplebufs);
274 * Free a ReorderBuffer
277 ReorderBufferFree(ReorderBuffer *rb)
279 MemoryContext context = rb->context;
282 * We free separately allocated data by entirely scrapping reorderbuffer's
285 MemoryContextDelete(context);
289 * Get an unused, possibly preallocated, ReorderBufferTXN.
291 static ReorderBufferTXN *
292 ReorderBufferGetTXN(ReorderBuffer *rb)
294 ReorderBufferTXN *txn;
296 /* check the slab cache */
297 if (rb->nr_cached_transactions > 0)
299 rb->nr_cached_transactions--;
300 txn = (ReorderBufferTXN *)
301 dlist_container(ReorderBufferTXN, node,
302 dlist_pop_head_node(&rb->cached_transactions));
306 txn = (ReorderBufferTXN *)
307 MemoryContextAlloc(rb->context, sizeof(ReorderBufferTXN));
310 memset(txn, 0, sizeof(ReorderBufferTXN));
312 dlist_init(&txn->changes);
313 dlist_init(&txn->tuplecids);
314 dlist_init(&txn->subtxns);
320 * Free a ReorderBufferTXN.
322 * Deallocation might be delayed for efficiency purposes, for details check
323 * the comments above max_cached_changes's definition.
326 ReorderBufferReturnTXN(ReorderBuffer *rb, ReorderBufferTXN *txn)
328 /* clean the lookup cache if we were cached (quite likely) */
329 if (rb->by_txn_last_xid == txn->xid)
331 rb->by_txn_last_xid = InvalidTransactionId;
332 rb->by_txn_last_txn = NULL;
335 /* free data that's contained */
337 if (txn->tuplecid_hash != NULL)
339 hash_destroy(txn->tuplecid_hash);
340 txn->tuplecid_hash = NULL;
343 if (txn->invalidations)
345 pfree(txn->invalidations);
346 txn->invalidations = NULL;
349 /* check whether to put into the slab cache */
350 if (rb->nr_cached_transactions < max_cached_transactions)
352 rb->nr_cached_transactions++;
353 dlist_push_head(&rb->cached_transactions, &txn->node);
354 VALGRIND_MAKE_MEM_UNDEFINED(txn, sizeof(ReorderBufferTXN));
355 VALGRIND_MAKE_MEM_DEFINED(&txn->node, sizeof(txn->node));
364 * Get an unused, possibly preallocated, ReorderBufferChange.
366 ReorderBufferChange *
367 ReorderBufferGetChange(ReorderBuffer *rb)
369 ReorderBufferChange *change;
371 /* check the slab cache */
372 if (rb->nr_cached_changes)
374 rb->nr_cached_changes--;
375 change = (ReorderBufferChange *)
376 dlist_container(ReorderBufferChange, node,
377 dlist_pop_head_node(&rb->cached_changes));
381 change = (ReorderBufferChange *)
382 MemoryContextAlloc(rb->context, sizeof(ReorderBufferChange));
385 memset(change, 0, sizeof(ReorderBufferChange));
390 * Free an ReorderBufferChange.
392 * Deallocation might be delayed for efficiency purposes, for details check
393 * the comments above max_cached_changes's definition.
396 ReorderBufferReturnChange(ReorderBuffer *rb, ReorderBufferChange *change)
398 /* free contained data */
399 switch (change->action)
401 case REORDER_BUFFER_CHANGE_INSERT:
402 case REORDER_BUFFER_CHANGE_UPDATE:
403 case REORDER_BUFFER_CHANGE_DELETE:
404 case REORDER_BUFFER_CHANGE_INTERNAL_SPEC_INSERT:
405 if (change->data.tp.newtuple)
407 ReorderBufferReturnTupleBuf(rb, change->data.tp.newtuple);
408 change->data.tp.newtuple = NULL;
411 if (change->data.tp.oldtuple)
413 ReorderBufferReturnTupleBuf(rb, change->data.tp.oldtuple);
414 change->data.tp.oldtuple = NULL;
417 case REORDER_BUFFER_CHANGE_INTERNAL_SNAPSHOT:
418 if (change->data.snapshot)
420 ReorderBufferFreeSnap(rb, change->data.snapshot);
421 change->data.snapshot = NULL;
424 /* no data in addition to the struct itself */
425 case REORDER_BUFFER_CHANGE_INTERNAL_SPEC_CONFIRM:
426 case REORDER_BUFFER_CHANGE_INTERNAL_COMMAND_ID:
427 case REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID:
431 /* check whether to put into the slab cache */
432 if (rb->nr_cached_changes < max_cached_changes)
434 rb->nr_cached_changes++;
435 dlist_push_head(&rb->cached_changes, &change->node);
436 VALGRIND_MAKE_MEM_UNDEFINED(change, sizeof(ReorderBufferChange));
437 VALGRIND_MAKE_MEM_DEFINED(&change->node, sizeof(change->node));
447 * Get an unused, possibly preallocated, ReorderBufferTupleBuf fitting at
448 * least a tuple of size tuple_len (excluding header overhead).
450 ReorderBufferTupleBuf *
451 ReorderBufferGetTupleBuf(ReorderBuffer *rb, Size tuple_len)
453 ReorderBufferTupleBuf *tuple;
456 alloc_len = tuple_len + SizeofHeapTupleHeader;
459 * Most tuples are below MaxHeapTupleSize, so we use a slab allocator for
460 * those. Thus always allocate at least MaxHeapTupleSize. Note that tuples
461 * tuples generated for oldtuples can be bigger, as they don't have
462 * out-of-line toast columns.
464 if (alloc_len < MaxHeapTupleSize)
465 alloc_len = MaxHeapTupleSize;
468 /* if small enough, check the slab cache */
469 if (alloc_len <= MaxHeapTupleSize && rb->nr_cached_tuplebufs)
471 rb->nr_cached_tuplebufs--;
472 tuple = slist_container(ReorderBufferTupleBuf, node,
473 slist_pop_head_node(&rb->cached_tuplebufs));
474 Assert(tuple->alloc_tuple_size == MaxHeapTupleSize);
475 #ifdef USE_ASSERT_CHECKING
476 memset(&tuple->tuple, 0xa9, sizeof(HeapTupleData));
477 VALGRIND_MAKE_MEM_UNDEFINED(&tuple->tuple, sizeof(HeapTupleData));
479 tuple->tuple.t_data = ReorderBufferTupleBufData(tuple);
480 #ifdef USE_ASSERT_CHECKING
481 memset(tuple->tuple.t_data, 0xa8, tuple->alloc_tuple_size);
482 VALGRIND_MAKE_MEM_UNDEFINED(tuple->tuple.t_data, tuple->alloc_tuple_size);
487 tuple = (ReorderBufferTupleBuf *)
488 MemoryContextAlloc(rb->context,
489 sizeof(ReorderBufferTupleBuf) +
490 MAXIMUM_ALIGNOF + alloc_len);
491 tuple->alloc_tuple_size = alloc_len;
492 tuple->tuple.t_data = ReorderBufferTupleBufData(tuple);
499 * Free an ReorderBufferTupleBuf.
501 * Deallocation might be delayed for efficiency purposes, for details check
502 * the comments above max_cached_changes's definition.
505 ReorderBufferReturnTupleBuf(ReorderBuffer *rb, ReorderBufferTupleBuf *tuple)
507 /* check whether to put into the slab cache, oversized tuples never are */
508 if (tuple->alloc_tuple_size == MaxHeapTupleSize &&
509 rb->nr_cached_tuplebufs < max_cached_tuplebufs)
511 rb->nr_cached_tuplebufs++;
512 slist_push_head(&rb->cached_tuplebufs, &tuple->node);
513 VALGRIND_MAKE_MEM_UNDEFINED(tuple->tuple.t_data, tuple->alloc_tuple_size);
514 VALGRIND_MAKE_MEM_UNDEFINED(tuple, sizeof(ReorderBufferTupleBuf));
515 VALGRIND_MAKE_MEM_DEFINED(&tuple->node, sizeof(tuple->node));
516 VALGRIND_MAKE_MEM_DEFINED(&tuple->alloc_tuple_size, sizeof(tuple->alloc_tuple_size));
525 * Return the ReorderBufferTXN from the given buffer, specified by Xid.
526 * If create is true, and a transaction doesn't already exist, create it
527 * (with the given LSN, and as top transaction if that's specified);
528 * when this happens, is_new is set to true.
530 static ReorderBufferTXN *
531 ReorderBufferTXNByXid(ReorderBuffer *rb, TransactionId xid, bool create,
532 bool *is_new, XLogRecPtr lsn, bool create_as_top)
534 ReorderBufferTXN *txn;
535 ReorderBufferTXNByIdEnt *ent;
538 Assert(TransactionIdIsValid(xid));
539 Assert(!create || lsn != InvalidXLogRecPtr);
542 * Check the one-entry lookup cache first
544 if (TransactionIdIsValid(rb->by_txn_last_xid) &&
545 rb->by_txn_last_xid == xid)
547 txn = rb->by_txn_last_txn;
551 /* found it, and it's valid */
558 * cached as non-existant, and asked not to create? Then nothing else
563 /* otherwise fall through to create it */
567 * If the cache wasn't hit or it yielded an "does-not-exist" and we want
568 * to create an entry.
571 /* search the lookup table */
572 ent = (ReorderBufferTXNByIdEnt *)
573 hash_search(rb->by_txn,
575 create ? HASH_ENTER : HASH_FIND,
581 /* initialize the new entry, if creation was requested */
584 ent->txn = ReorderBufferGetTXN(rb);
587 txn->first_lsn = lsn;
588 txn->restart_decoding_lsn = rb->current_restart_decoding_lsn;
592 dlist_push_tail(&rb->toplevel_by_lsn, &txn->node);
593 AssertTXNLsnOrder(rb);
597 txn = NULL; /* not found and not asked to create */
600 rb->by_txn_last_xid = xid;
601 rb->by_txn_last_txn = txn;
606 Assert(!create || !!txn);
611 * Queue a change into a transaction so it can be replayed upon commit.
614 ReorderBufferQueueChange(ReorderBuffer *rb, TransactionId xid, XLogRecPtr lsn,
615 ReorderBufferChange *change)
617 ReorderBufferTXN *txn;
619 txn = ReorderBufferTXNByXid(rb, xid, true, NULL, lsn, true);
622 Assert(InvalidXLogRecPtr != lsn);
623 dlist_push_tail(&txn->changes, &change->node);
627 ReorderBufferCheckSerializeTXN(rb, txn);
631 AssertTXNLsnOrder(ReorderBuffer *rb)
633 #ifdef USE_ASSERT_CHECKING
635 XLogRecPtr prev_first_lsn = InvalidXLogRecPtr;
637 dlist_foreach(iter, &rb->toplevel_by_lsn)
639 ReorderBufferTXN *cur_txn;
641 cur_txn = dlist_container(ReorderBufferTXN, node, iter.cur);
642 Assert(cur_txn->first_lsn != InvalidXLogRecPtr);
644 if (cur_txn->end_lsn != InvalidXLogRecPtr)
645 Assert(cur_txn->first_lsn <= cur_txn->end_lsn);
647 if (prev_first_lsn != InvalidXLogRecPtr)
648 Assert(prev_first_lsn < cur_txn->first_lsn);
650 Assert(!cur_txn->is_known_as_subxact);
651 prev_first_lsn = cur_txn->first_lsn;
657 ReorderBufferGetOldestTXN(ReorderBuffer *rb)
659 ReorderBufferTXN *txn;
661 if (dlist_is_empty(&rb->toplevel_by_lsn))
664 AssertTXNLsnOrder(rb);
666 txn = dlist_head_element(ReorderBufferTXN, node, &rb->toplevel_by_lsn);
668 Assert(!txn->is_known_as_subxact);
669 Assert(txn->first_lsn != InvalidXLogRecPtr);
674 ReorderBufferSetRestartPoint(ReorderBuffer *rb, XLogRecPtr ptr)
676 rb->current_restart_decoding_lsn = ptr;
680 ReorderBufferAssignChild(ReorderBuffer *rb, TransactionId xid,
681 TransactionId subxid, XLogRecPtr lsn)
683 ReorderBufferTXN *txn;
684 ReorderBufferTXN *subtxn;
688 txn = ReorderBufferTXNByXid(rb, xid, true, &new_top, lsn, true);
689 subtxn = ReorderBufferTXNByXid(rb, subxid, true, &new_sub, lsn, false);
694 * we assign subtransactions to top level transaction even if we don't
695 * have data for it yet, assignment records frequently reference xids
696 * that have not yet produced any records. Knowing those aren't top
697 * level xids allows us to make processing cheaper in some places.
699 dlist_push_tail(&txn->subtxns, &subtxn->node);
702 else if (!subtxn->is_known_as_subxact)
704 subtxn->is_known_as_subxact = true;
705 Assert(subtxn->nsubtxns == 0);
707 /* remove from lsn order list of top-level transactions */
708 dlist_delete(&subtxn->node);
710 /* add to toplevel transaction */
711 dlist_push_tail(&txn->subtxns, &subtxn->node);
716 elog(ERROR, "existing subxact assigned to unknown toplevel xact");
721 * Associate a subtransaction with its toplevel transaction at commit
722 * time. There may be no further changes added after this.
725 ReorderBufferCommitChild(ReorderBuffer *rb, TransactionId xid,
726 TransactionId subxid, XLogRecPtr commit_lsn,
729 ReorderBufferTXN *txn;
730 ReorderBufferTXN *subtxn;
732 subtxn = ReorderBufferTXNByXid(rb, subxid, false, NULL,
733 InvalidXLogRecPtr, false);
736 * No need to do anything if that subtxn didn't contain any changes
741 txn = ReorderBufferTXNByXid(rb, xid, false, NULL, commit_lsn, true);
744 elog(ERROR, "subxact logged without previous toplevel record");
747 * Pass the our base snapshot to the parent transaction if it doesn't have
748 * one, or ours is older. That can happen if there are no changes in the
749 * toplevel transaction but in one of the child transactions. This allows
750 * the parent to simply use it's base snapshot initially.
752 if (txn->base_snapshot == NULL ||
753 txn->base_snapshot_lsn > subtxn->base_snapshot_lsn)
755 txn->base_snapshot = subtxn->base_snapshot;
756 txn->base_snapshot_lsn = subtxn->base_snapshot_lsn;
757 subtxn->base_snapshot = NULL;
758 subtxn->base_snapshot_lsn = InvalidXLogRecPtr;
761 subtxn->final_lsn = commit_lsn;
762 subtxn->end_lsn = end_lsn;
764 if (!subtxn->is_known_as_subxact)
766 subtxn->is_known_as_subxact = true;
767 Assert(subtxn->nsubtxns == 0);
769 /* remove from lsn order list of top-level transactions */
770 dlist_delete(&subtxn->node);
772 /* add to subtransaction list */
773 dlist_push_tail(&txn->subtxns, &subtxn->node);
780 * Support for efficiently iterating over a transaction's and its
781 * subtransactions' changes.
783 * We do by doing a k-way merge between transactions/subtransactions. For that
784 * we model the current heads of the different transactions as a binary heap
785 * so we easily know which (sub-)transaction has the change with the smallest
788 * We assume the changes in individual transactions are already sorted by LSN.
792 * Binary heap comparison function.
795 ReorderBufferIterCompare(Datum a, Datum b, void *arg)
797 ReorderBufferIterTXNState *state = (ReorderBufferIterTXNState *) arg;
798 XLogRecPtr pos_a = state->entries[DatumGetInt32(a)].lsn;
799 XLogRecPtr pos_b = state->entries[DatumGetInt32(b)].lsn;
803 else if (pos_a == pos_b)
809 * Allocate & initialize an iterator which iterates in lsn order over a
810 * transaction and all its subtransactions.
812 static ReorderBufferIterTXNState *
813 ReorderBufferIterTXNInit(ReorderBuffer *rb, ReorderBufferTXN *txn)
816 ReorderBufferIterTXNState *state;
817 dlist_iter cur_txn_i;
821 * Calculate the size of our heap: one element for every transaction that
822 * contains changes. (Besides the transactions already in the reorder
823 * buffer, we count the one we were directly passed.)
825 if (txn->nentries > 0)
828 dlist_foreach(cur_txn_i, &txn->subtxns)
830 ReorderBufferTXN *cur_txn;
832 cur_txn = dlist_container(ReorderBufferTXN, node, cur_txn_i.cur);
834 if (cur_txn->nentries > 0)
839 * TODO: Consider adding fastpath for the rather common nr_txns=1 case, no
840 * need to allocate/build a heap then.
843 /* allocate iteration state */
844 state = (ReorderBufferIterTXNState *)
845 MemoryContextAllocZero(rb->context,
846 sizeof(ReorderBufferIterTXNState) +
847 sizeof(ReorderBufferIterTXNEntry) * nr_txns);
849 state->nr_txns = nr_txns;
850 dlist_init(&state->old_change);
852 for (off = 0; off < state->nr_txns; off++)
854 state->entries[off].fd = -1;
855 state->entries[off].segno = 0;
859 state->heap = binaryheap_allocate(state->nr_txns,
860 ReorderBufferIterCompare,
864 * Now insert items into the binary heap, in an unordered fashion. (We
865 * will run a heap assembly step at the end; this is more efficient.)
870 /* add toplevel transaction if it contains changes */
871 if (txn->nentries > 0)
873 ReorderBufferChange *cur_change;
875 if (txn->nentries != txn->nentries_mem)
876 ReorderBufferRestoreChanges(rb, txn, &state->entries[off].fd,
877 &state->entries[off].segno);
879 cur_change = dlist_head_element(ReorderBufferChange, node,
882 state->entries[off].lsn = cur_change->lsn;
883 state->entries[off].change = cur_change;
884 state->entries[off].txn = txn;
886 binaryheap_add_unordered(state->heap, Int32GetDatum(off++));
889 /* add subtransactions if they contain changes */
890 dlist_foreach(cur_txn_i, &txn->subtxns)
892 ReorderBufferTXN *cur_txn;
894 cur_txn = dlist_container(ReorderBufferTXN, node, cur_txn_i.cur);
896 if (cur_txn->nentries > 0)
898 ReorderBufferChange *cur_change;
900 if (txn->nentries != txn->nentries_mem)
901 ReorderBufferRestoreChanges(rb, cur_txn,
902 &state->entries[off].fd,
903 &state->entries[off].segno);
905 cur_change = dlist_head_element(ReorderBufferChange, node,
908 state->entries[off].lsn = cur_change->lsn;
909 state->entries[off].change = cur_change;
910 state->entries[off].txn = cur_txn;
912 binaryheap_add_unordered(state->heap, Int32GetDatum(off++));
916 /* assemble a valid binary heap */
917 binaryheap_build(state->heap);
923 * Return the next change when iterating over a transaction and its
926 * Returns NULL when no further changes exist.
928 static ReorderBufferChange *
929 ReorderBufferIterTXNNext(ReorderBuffer *rb, ReorderBufferIterTXNState *state)
931 ReorderBufferChange *change;
932 ReorderBufferIterTXNEntry *entry;
935 /* nothing there anymore */
936 if (state->heap->bh_size == 0)
939 off = DatumGetInt32(binaryheap_first(state->heap));
940 entry = &state->entries[off];
942 /* free memory we might have "leaked" in the previous *Next call */
943 if (!dlist_is_empty(&state->old_change))
945 change = dlist_container(ReorderBufferChange, node,
946 dlist_pop_head_node(&state->old_change));
947 ReorderBufferReturnChange(rb, change);
948 Assert(dlist_is_empty(&state->old_change));
951 change = entry->change;
954 * update heap with information about which transaction has the next
955 * relevant change in LSN order
958 /* there are in-memory changes */
959 if (dlist_has_next(&entry->txn->changes, &entry->change->node))
961 dlist_node *next = dlist_next_node(&entry->txn->changes, &change->node);
962 ReorderBufferChange *next_change =
963 dlist_container(ReorderBufferChange, node, next);
965 /* txn stays the same */
966 state->entries[off].lsn = next_change->lsn;
967 state->entries[off].change = next_change;
969 binaryheap_replace_first(state->heap, Int32GetDatum(off));
973 /* try to load changes from disk */
974 if (entry->txn->nentries != entry->txn->nentries_mem)
977 * Ugly: restoring changes will reuse *Change records, thus delete the
978 * current one from the per-tx list and only free in the next call.
980 dlist_delete(&change->node);
981 dlist_push_tail(&state->old_change, &change->node);
983 if (ReorderBufferRestoreChanges(rb, entry->txn, &entry->fd,
984 &state->entries[off].segno))
986 /* successfully restored changes from disk */
987 ReorderBufferChange *next_change =
988 dlist_head_element(ReorderBufferChange, node,
989 &entry->txn->changes);
991 elog(DEBUG2, "restored %u/%u changes from disk",
992 (uint32) entry->txn->nentries_mem,
993 (uint32) entry->txn->nentries);
995 Assert(entry->txn->nentries_mem);
996 /* txn stays the same */
997 state->entries[off].lsn = next_change->lsn;
998 state->entries[off].change = next_change;
999 binaryheap_replace_first(state->heap, Int32GetDatum(off));
1005 /* ok, no changes there anymore, remove */
1006 binaryheap_remove_first(state->heap);
1012 * Deallocate the iterator
1015 ReorderBufferIterTXNFinish(ReorderBuffer *rb,
1016 ReorderBufferIterTXNState *state)
1020 for (off = 0; off < state->nr_txns; off++)
1022 if (state->entries[off].fd != -1)
1023 CloseTransientFile(state->entries[off].fd);
1026 /* free memory we might have "leaked" in the last *Next call */
1027 if (!dlist_is_empty(&state->old_change))
1029 ReorderBufferChange *change;
1031 change = dlist_container(ReorderBufferChange, node,
1032 dlist_pop_head_node(&state->old_change));
1033 ReorderBufferReturnChange(rb, change);
1034 Assert(dlist_is_empty(&state->old_change));
1037 binaryheap_free(state->heap);
1042 * Cleanup the contents of a transaction, usually after the transaction
1043 * committed or aborted.
1046 ReorderBufferCleanupTXN(ReorderBuffer *rb, ReorderBufferTXN *txn)
1049 dlist_mutable_iter iter;
1051 /* cleanup subtransactions & their changes */
1052 dlist_foreach_modify(iter, &txn->subtxns)
1054 ReorderBufferTXN *subtxn;
1056 subtxn = dlist_container(ReorderBufferTXN, node, iter.cur);
1059 * Subtransactions are always associated to the toplevel TXN, even if
1060 * they originally were happening inside another subtxn, so we won't
1061 * ever recurse more than one level deep here.
1063 Assert(subtxn->is_known_as_subxact);
1064 Assert(subtxn->nsubtxns == 0);
1066 ReorderBufferCleanupTXN(rb, subtxn);
1069 /* cleanup changes in the toplevel txn */
1070 dlist_foreach_modify(iter, &txn->changes)
1072 ReorderBufferChange *change;
1074 change = dlist_container(ReorderBufferChange, node, iter.cur);
1076 ReorderBufferReturnChange(rb, change);
1080 * Cleanup the tuplecids we stored for decoding catalog snapshot access.
1081 * They are always stored in the toplevel transaction.
1083 dlist_foreach_modify(iter, &txn->tuplecids)
1085 ReorderBufferChange *change;
1087 change = dlist_container(ReorderBufferChange, node, iter.cur);
1088 Assert(change->action == REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID);
1089 ReorderBufferReturnChange(rb, change);
1092 if (txn->base_snapshot != NULL)
1094 SnapBuildSnapDecRefcount(txn->base_snapshot);
1095 txn->base_snapshot = NULL;
1096 txn->base_snapshot_lsn = InvalidXLogRecPtr;
1099 /* delete from list of known subxacts */
1100 if (txn->is_known_as_subxact)
1102 /* NB: nsubxacts count of parent will be too high now */
1103 dlist_delete(&txn->node);
1105 /* delete from LSN ordered list of toplevel TXNs */
1108 dlist_delete(&txn->node);
1111 /* now remove reference from buffer */
1112 hash_search(rb->by_txn,
1118 /* remove entries spilled to disk */
1119 if (txn->nentries != txn->nentries_mem)
1120 ReorderBufferRestoreCleanup(rb, txn);
1123 ReorderBufferReturnTXN(rb, txn);
1127 * Build a hash with a (relfilenode, ctid) -> (cmin, cmax) mapping for use by
1128 * tqual.c's HeapTupleSatisfiesHistoricMVCC.
1131 ReorderBufferBuildTupleCidHash(ReorderBuffer *rb, ReorderBufferTXN *txn)
1136 if (!txn->has_catalog_changes || dlist_is_empty(&txn->tuplecids))
1139 memset(&hash_ctl, 0, sizeof(hash_ctl));
1141 hash_ctl.keysize = sizeof(ReorderBufferTupleCidKey);
1142 hash_ctl.entrysize = sizeof(ReorderBufferTupleCidEnt);
1143 hash_ctl.hcxt = rb->context;
1146 * create the hash with the exact number of to-be-stored tuplecids from
1149 txn->tuplecid_hash =
1150 hash_create("ReorderBufferTupleCid", txn->ntuplecids, &hash_ctl,
1151 HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1153 dlist_foreach(iter, &txn->tuplecids)
1155 ReorderBufferTupleCidKey key;
1156 ReorderBufferTupleCidEnt *ent;
1158 ReorderBufferChange *change;
1160 change = dlist_container(ReorderBufferChange, node, iter.cur);
1162 Assert(change->action == REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID);
1164 /* be careful about padding */
1165 memset(&key, 0, sizeof(ReorderBufferTupleCidKey));
1167 key.relnode = change->data.tuplecid.node;
1169 ItemPointerCopy(&change->data.tuplecid.tid,
1172 ent = (ReorderBufferTupleCidEnt *)
1173 hash_search(txn->tuplecid_hash,
1175 HASH_ENTER | HASH_FIND,
1179 ent->cmin = change->data.tuplecid.cmin;
1180 ent->cmax = change->data.tuplecid.cmax;
1181 ent->combocid = change->data.tuplecid.combocid;
1185 Assert(ent->cmin == change->data.tuplecid.cmin);
1186 Assert(ent->cmax == InvalidCommandId ||
1187 ent->cmax == change->data.tuplecid.cmax);
1190 * if the tuple got valid in this transaction and now got deleted
1191 * we already have a valid cmin stored. The cmax will be
1192 * InvalidCommandId though.
1194 ent->cmax = change->data.tuplecid.cmax;
1200 * Copy a provided snapshot so we can modify it privately. This is needed so
1201 * that catalog modifying transactions can look into intermediate catalog
1205 ReorderBufferCopySnap(ReorderBuffer *rb, Snapshot orig_snap,
1206 ReorderBufferTXN *txn, CommandId cid)
1213 size = sizeof(SnapshotData) +
1214 sizeof(TransactionId) * orig_snap->xcnt +
1215 sizeof(TransactionId) * (txn->nsubtxns + 1);
1217 snap = MemoryContextAllocZero(rb->context, size);
1218 memcpy(snap, orig_snap, sizeof(SnapshotData));
1220 snap->copied = true;
1221 snap->active_count = 1; /* mark as active so nobody frees it */
1222 snap->regd_count = 0;
1223 snap->xip = (TransactionId *) (snap + 1);
1225 memcpy(snap->xip, orig_snap->xip, sizeof(TransactionId) * snap->xcnt);
1228 * snap->subxip contains all txids that belong to our transaction which we
1229 * need to check via cmin/cmax. Thats why we store the toplevel
1230 * transaction in there as well.
1232 snap->subxip = snap->xip + snap->xcnt;
1233 snap->subxip[i++] = txn->xid;
1236 * nsubxcnt isn't decreased when subtransactions abort, so count manually.
1237 * Since it's an upper boundary it is safe to use it for the allocation
1242 dlist_foreach(iter, &txn->subtxns)
1244 ReorderBufferTXN *sub_txn;
1246 sub_txn = dlist_container(ReorderBufferTXN, node, iter.cur);
1247 snap->subxip[i++] = sub_txn->xid;
1251 /* sort so we can bsearch() later */
1252 qsort(snap->subxip, snap->subxcnt, sizeof(TransactionId), xidComparator);
1254 /* store the specified current CommandId */
1261 * Free a previously ReorderBufferCopySnap'ed snapshot
1264 ReorderBufferFreeSnap(ReorderBuffer *rb, Snapshot snap)
1269 SnapBuildSnapDecRefcount(snap);
1273 * Perform the replay of a transaction and it's non-aborted subtransactions.
1275 * Subtransactions previously have to be processed by
1276 * ReorderBufferCommitChild(), even if previously assigned to the toplevel
1277 * transaction with ReorderBufferAssignChild.
1279 * We currently can only decode a transaction's contents in when their commit
1280 * record is read because that's currently the only place where we know about
1281 * cache invalidations. Thus, once a toplevel commit is read, we iterate over
1282 * the top and subtransactions (using a k-way merge) and replay the changes in
1286 ReorderBufferCommit(ReorderBuffer *rb, TransactionId xid,
1287 XLogRecPtr commit_lsn, XLogRecPtr end_lsn,
1288 TimestampTz commit_time,
1289 RepOriginId origin_id, XLogRecPtr origin_lsn)
1291 ReorderBufferTXN *txn;
1292 volatile Snapshot snapshot_now;
1293 volatile CommandId command_id = FirstCommandId;
1295 ReorderBufferIterTXNState *volatile iterstate = NULL;
1297 txn = ReorderBufferTXNByXid(rb, xid, false, NULL, InvalidXLogRecPtr,
1300 /* unknown transaction, nothing to replay */
1304 txn->final_lsn = commit_lsn;
1305 txn->end_lsn = end_lsn;
1306 txn->commit_time = commit_time;
1307 txn->origin_id = origin_id;
1308 txn->origin_lsn = origin_lsn;
1310 /* serialize the last bunch of changes if we need start earlier anyway */
1311 if (txn->nentries_mem != txn->nentries)
1312 ReorderBufferSerializeTXN(rb, txn);
1315 * If this transaction didn't have any real changes in our database, it's
1316 * OK not to have a snapshot. Note that ReorderBufferCommitChild will have
1317 * transferred its snapshot to this transaction if it had one and the
1318 * toplevel tx didn't.
1320 if (txn->base_snapshot == NULL)
1322 Assert(txn->ninvalidations == 0);
1323 ReorderBufferCleanupTXN(rb, txn);
1327 snapshot_now = txn->base_snapshot;
1329 /* build data to be able to lookup the CommandIds of catalog tuples */
1330 ReorderBufferBuildTupleCidHash(rb, txn);
1332 /* setup the initial snapshot */
1333 SetupHistoricSnapshot(snapshot_now, txn->tuplecid_hash);
1336 * Decoding needs access to syscaches et al., which in turn use
1337 * heavyweight locks and such. Thus we need to have enough state around to
1338 * keep track of those. The easiest way is to simply use a transaction
1339 * internally. That also allows us to easily enforce that nothing writes
1340 * to the database by checking for xid assignments.
1342 * When we're called via the SQL SRF there's already a transaction
1343 * started, so start an explicit subtransaction there.
1345 using_subtxn = IsTransactionOrTransactionBlock();
1349 ReorderBufferChange *change;
1350 ReorderBufferChange *specinsert = NULL;
1353 BeginInternalSubTransaction("replay");
1355 StartTransactionCommand();
1359 iterstate = ReorderBufferIterTXNInit(rb, txn);
1360 while ((change = ReorderBufferIterTXNNext(rb, iterstate)) != NULL)
1362 Relation relation = NULL;
1365 switch (change->action)
1367 case REORDER_BUFFER_CHANGE_INTERNAL_SPEC_CONFIRM:
1370 * Confirmation for speculative insertion arrived. Simply
1371 * use as a normal record. It'll be cleaned up at the end
1372 * of INSERT processing.
1374 Assert(specinsert->data.tp.oldtuple == NULL);
1375 change = specinsert;
1376 change->action = REORDER_BUFFER_CHANGE_INSERT;
1378 /* intentionally fall through */
1379 case REORDER_BUFFER_CHANGE_INSERT:
1380 case REORDER_BUFFER_CHANGE_UPDATE:
1381 case REORDER_BUFFER_CHANGE_DELETE:
1382 Assert(snapshot_now);
1384 reloid = RelidByRelfilenode(change->data.tp.relnode.spcNode,
1385 change->data.tp.relnode.relNode);
1388 * Catalog tuple without data, emitted while catalog was
1389 * in the process of being rewritten.
1391 if (reloid == InvalidOid &&
1392 change->data.tp.newtuple == NULL &&
1393 change->data.tp.oldtuple == NULL)
1395 else if (reloid == InvalidOid)
1396 elog(ERROR, "could not map filenode \"%s\" to relation OID",
1397 relpathperm(change->data.tp.relnode,
1400 relation = RelationIdGetRelation(reloid);
1402 if (relation == NULL)
1403 elog(ERROR, "could not open relation with OID %u (for filenode \"%s\")",
1405 relpathperm(change->data.tp.relnode,
1408 if (!RelationIsLogicallyLogged(relation))
1412 * For now ignore sequence changes entirely. Most of the
1413 * time they don't log changes using records we
1414 * understand, so it doesn't make sense to handle the few
1417 if (relation->rd_rel->relkind == RELKIND_SEQUENCE)
1420 /* user-triggered change */
1421 if (!IsToastRelation(relation))
1423 ReorderBufferToastReplace(rb, txn, relation, change);
1424 rb->apply_change(rb, txn, relation, change);
1427 * Only clear reassembled toast chunks if we're sure
1428 * they're not required anymore. The creator of the
1431 if (change->data.tp.clear_toast_afterwards)
1432 ReorderBufferToastReset(rb, txn);
1434 /* we're not interested in toast deletions */
1435 else if (change->action == REORDER_BUFFER_CHANGE_INSERT)
1438 * Need to reassemble the full toasted Datum in
1439 * memory, to ensure the chunks don't get reused till
1440 * we're done remove it from the list of this
1441 * transaction's changes. Otherwise it will get
1442 * freed/reused while restoring spooled data from
1445 dlist_delete(&change->node);
1446 ReorderBufferToastAppendChunk(rb, txn, relation,
1453 * Either speculative insertion was confirmed, or it was
1454 * unsuccessful and the record isn't needed anymore.
1456 if (specinsert != NULL)
1458 ReorderBufferReturnChange(rb, specinsert);
1462 if (relation != NULL)
1464 RelationClose(relation);
1469 case REORDER_BUFFER_CHANGE_INTERNAL_SPEC_INSERT:
1472 * Speculative insertions are dealt with by delaying the
1473 * processing of the insert until the confirmation record
1474 * arrives. For that we simply unlink the record from the
1475 * chain, so it does not get freed/reused while restoring
1476 * spooled data from disk.
1478 * This is safe in the face of concurrent catalog changes
1479 * because the relevant relation can't be changed between
1480 * speculative insertion and confirmation due to
1481 * CheckTableNotInUse() and locking.
1484 /* clear out a pending (and thus failed) speculation */
1485 if (specinsert != NULL)
1487 ReorderBufferReturnChange(rb, specinsert);
1491 /* and memorize the pending insertion */
1492 dlist_delete(&change->node);
1493 specinsert = change;
1496 case REORDER_BUFFER_CHANGE_INTERNAL_SNAPSHOT:
1497 /* get rid of the old */
1498 TeardownHistoricSnapshot(false);
1500 if (snapshot_now->copied)
1502 ReorderBufferFreeSnap(rb, snapshot_now);
1504 ReorderBufferCopySnap(rb, change->data.snapshot,
1509 * Restored from disk, need to be careful not to double
1510 * free. We could introduce refcounting for that, but for
1511 * now this seems infrequent enough not to care.
1513 else if (change->data.snapshot->copied)
1516 ReorderBufferCopySnap(rb, change->data.snapshot,
1521 snapshot_now = change->data.snapshot;
1525 /* and continue with the new one */
1526 SetupHistoricSnapshot(snapshot_now, txn->tuplecid_hash);
1529 case REORDER_BUFFER_CHANGE_INTERNAL_COMMAND_ID:
1530 Assert(change->data.command_id != InvalidCommandId);
1532 if (command_id < change->data.command_id)
1534 command_id = change->data.command_id;
1536 if (!snapshot_now->copied)
1538 /* we don't use the global one anymore */
1539 snapshot_now = ReorderBufferCopySnap(rb, snapshot_now,
1543 snapshot_now->curcid = command_id;
1545 TeardownHistoricSnapshot(false);
1546 SetupHistoricSnapshot(snapshot_now, txn->tuplecid_hash);
1549 * Every time the CommandId is incremented, we could
1550 * see new catalog contents, so execute all
1553 ReorderBufferExecuteInvalidations(rb, txn);
1558 case REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID:
1559 elog(ERROR, "tuplecid value in changequeue");
1565 * There's a speculative insertion remaining, just clean in up, it
1566 * can't have been successful, otherwise we'd gotten a confirmation
1571 ReorderBufferReturnChange(rb, specinsert);
1575 /* clean up the iterator */
1576 ReorderBufferIterTXNFinish(rb, iterstate);
1579 /* call commit callback */
1580 rb->commit(rb, txn, commit_lsn);
1582 /* this is just a sanity check against bad output plugin behaviour */
1583 if (GetCurrentTransactionIdIfAny() != InvalidTransactionId)
1584 elog(ERROR, "output plugin used XID %u",
1585 GetCurrentTransactionId());
1588 TeardownHistoricSnapshot(false);
1591 * Aborting the current (sub-)transaction as a whole has the right
1592 * semantics. We want all locks acquired in here to be released, not
1593 * reassigned to the parent and we do not want any database access
1594 * have persistent effects.
1596 AbortCurrentTransaction();
1598 /* make sure there's no cache pollution */
1599 ReorderBufferExecuteInvalidations(rb, txn);
1602 RollbackAndReleaseCurrentSubTransaction();
1604 if (snapshot_now->copied)
1605 ReorderBufferFreeSnap(rb, snapshot_now);
1607 /* remove potential on-disk data, and deallocate */
1608 ReorderBufferCleanupTXN(rb, txn);
1612 /* TODO: Encapsulate cleanup from the PG_TRY and PG_CATCH blocks */
1614 ReorderBufferIterTXNFinish(rb, iterstate);
1616 TeardownHistoricSnapshot(true);
1619 * Force cache invalidation to happen outside of a valid transaction
1620 * to prevent catalog access as we just caught an error.
1622 AbortCurrentTransaction();
1624 /* make sure there's no cache pollution */
1625 ReorderBufferExecuteInvalidations(rb, txn);
1628 RollbackAndReleaseCurrentSubTransaction();
1630 if (snapshot_now->copied)
1631 ReorderBufferFreeSnap(rb, snapshot_now);
1633 /* remove potential on-disk data, and deallocate */
1634 ReorderBufferCleanupTXN(rb, txn);
1642 * Abort a transaction that possibly has previous changes. Needs to be first
1643 * called for subtransactions and then for the toplevel xid.
1645 * NB: Transactions handled here have to have actively aborted (i.e. have
1646 * produced an abort record). Implicitly aborted transactions are handled via
1647 * ReorderBufferAbortOld(); transactions we're just not interesteded in, but
1648 * which have committed are handled in ReorderBufferForget().
1650 * This function purges this transaction and its contents from memory and
1654 ReorderBufferAbort(ReorderBuffer *rb, TransactionId xid, XLogRecPtr lsn)
1656 ReorderBufferTXN *txn;
1658 txn = ReorderBufferTXNByXid(rb, xid, false, NULL, InvalidXLogRecPtr,
1661 /* unknown, nothing to remove */
1666 txn->final_lsn = lsn;
1668 /* remove potential on-disk data, and deallocate */
1669 ReorderBufferCleanupTXN(rb, txn);
1673 * Abort all transactions that aren't actually running anymore because the
1676 * NB: These really have to be transactions that have aborted due to a server
1677 * crash/immediate restart, as we don't deal with invalidations here.
1680 ReorderBufferAbortOld(ReorderBuffer *rb, TransactionId oldestRunningXid)
1682 dlist_mutable_iter it;
1685 * Iterate through all (potential) toplevel TXNs and abort all that are
1686 * older than what possibly can be running. Once we've found the first
1687 * that is alive we stop, there might be some that acquired an xid earlier
1688 * but started writing later, but it's unlikely and they will cleaned up
1689 * in a later call to ReorderBufferAbortOld().
1691 dlist_foreach_modify(it, &rb->toplevel_by_lsn)
1693 ReorderBufferTXN *txn;
1695 txn = dlist_container(ReorderBufferTXN, node, it.cur);
1697 if (TransactionIdPrecedes(txn->xid, oldestRunningXid))
1699 elog(DEBUG1, "aborting old transaction %u", txn->xid);
1701 /* remove potential on-disk data, and deallocate this tx */
1702 ReorderBufferCleanupTXN(rb, txn);
1710 * Forget the contents of a transaction if we aren't interested in it's
1711 * contents. Needs to be first called for subtransactions and then for the
1714 * This is significantly different to ReorderBufferAbort() because
1715 * transactions that have committed need to be treated differenly from aborted
1716 * ones since they may have modified the catalog.
1718 * Note that this is only allowed to be called in the moment a transaction
1719 * commit has just been read, not earlier; otherwise later records referring
1720 * to this xid might re-create the transaction incompletely.
1723 ReorderBufferForget(ReorderBuffer *rb, TransactionId xid, XLogRecPtr lsn)
1725 ReorderBufferTXN *txn;
1727 txn = ReorderBufferTXNByXid(rb, xid, false, NULL, InvalidXLogRecPtr,
1730 /* unknown, nothing to forget */
1735 txn->final_lsn = lsn;
1738 * Process cache invalidation messages if there are any. Even if we're not
1739 * interested in the transaction's contents, it could have manipulated the
1740 * catalog and we need to update the caches according to that.
1742 if (txn->base_snapshot != NULL && txn->ninvalidations > 0)
1744 bool use_subtxn = IsTransactionOrTransactionBlock();
1747 BeginInternalSubTransaction("replay");
1750 * Force invalidations to happen outside of a valid transaction - that
1751 * way entries will just be marked as invalid without accessing the
1752 * catalog. That's advantageous because we don't need to setup the
1753 * full state necessary for catalog access.
1756 AbortCurrentTransaction();
1758 ReorderBufferExecuteInvalidations(rb, txn);
1761 RollbackAndReleaseCurrentSubTransaction();
1764 Assert(txn->ninvalidations == 0);
1766 /* remove potential on-disk data, and deallocate */
1767 ReorderBufferCleanupTXN(rb, txn);
1772 * Tell reorderbuffer about an xid seen in the WAL stream. Has to be called at
1773 * least once for every xid in XLogRecord->xl_xid (other places in records
1774 * may, but do not have to be passed through here).
1776 * Reorderbuffer keeps some datastructures about transactions in LSN order,
1777 * for efficiency. To do that it has to know about when transactions are seen
1778 * first in the WAL. As many types of records are not actually interesting for
1779 * logical decoding, they do not necessarily pass though here.
1782 ReorderBufferProcessXid(ReorderBuffer *rb, TransactionId xid, XLogRecPtr lsn)
1784 /* many records won't have an xid assigned, centralize check here */
1785 if (xid != InvalidTransactionId)
1786 ReorderBufferTXNByXid(rb, xid, true, NULL, lsn, true);
1790 * Add a new snapshot to this transaction that may only used after lsn 'lsn'
1791 * because the previous snapshot doesn't describe the catalog correctly for
1795 ReorderBufferAddSnapshot(ReorderBuffer *rb, TransactionId xid,
1796 XLogRecPtr lsn, Snapshot snap)
1798 ReorderBufferChange *change = ReorderBufferGetChange(rb);
1800 change->data.snapshot = snap;
1801 change->action = REORDER_BUFFER_CHANGE_INTERNAL_SNAPSHOT;
1803 ReorderBufferQueueChange(rb, xid, lsn, change);
1807 * Setup the base snapshot of a transaction. The base snapshot is the snapshot
1808 * that is used to decode all changes until either this transaction modifies
1809 * the catalog or another catalog modifying transaction commits.
1811 * Needs to be called before any changes are added with
1812 * ReorderBufferQueueChange().
1815 ReorderBufferSetBaseSnapshot(ReorderBuffer *rb, TransactionId xid,
1816 XLogRecPtr lsn, Snapshot snap)
1818 ReorderBufferTXN *txn;
1821 txn = ReorderBufferTXNByXid(rb, xid, true, &is_new, lsn, true);
1822 Assert(txn->base_snapshot == NULL);
1823 Assert(snap != NULL);
1825 txn->base_snapshot = snap;
1826 txn->base_snapshot_lsn = lsn;
1830 * Access the catalog with this CommandId at this point in the changestream.
1832 * May only be called for command ids > 1
1835 ReorderBufferAddNewCommandId(ReorderBuffer *rb, TransactionId xid,
1836 XLogRecPtr lsn, CommandId cid)
1838 ReorderBufferChange *change = ReorderBufferGetChange(rb);
1840 change->data.command_id = cid;
1841 change->action = REORDER_BUFFER_CHANGE_INTERNAL_COMMAND_ID;
1843 ReorderBufferQueueChange(rb, xid, lsn, change);
1848 * Add new (relfilenode, tid) -> (cmin, cmax) mappings.
1851 ReorderBufferAddNewTupleCids(ReorderBuffer *rb, TransactionId xid,
1852 XLogRecPtr lsn, RelFileNode node,
1853 ItemPointerData tid, CommandId cmin,
1854 CommandId cmax, CommandId combocid)
1856 ReorderBufferChange *change = ReorderBufferGetChange(rb);
1857 ReorderBufferTXN *txn;
1859 txn = ReorderBufferTXNByXid(rb, xid, true, NULL, lsn, true);
1861 change->data.tuplecid.node = node;
1862 change->data.tuplecid.tid = tid;
1863 change->data.tuplecid.cmin = cmin;
1864 change->data.tuplecid.cmax = cmax;
1865 change->data.tuplecid.combocid = combocid;
1867 change->action = REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID;
1869 dlist_push_tail(&txn->tuplecids, &change->node);
1874 * Setup the invalidation of the toplevel transaction.
1876 * This needs to be done before ReorderBufferCommit is called!
1879 ReorderBufferAddInvalidations(ReorderBuffer *rb, TransactionId xid,
1880 XLogRecPtr lsn, Size nmsgs,
1881 SharedInvalidationMessage *msgs)
1883 ReorderBufferTXN *txn;
1885 txn = ReorderBufferTXNByXid(rb, xid, true, NULL, lsn, true);
1887 if (txn->ninvalidations != 0)
1888 elog(ERROR, "only ever add one set of invalidations");
1892 txn->ninvalidations = nmsgs;
1893 txn->invalidations = (SharedInvalidationMessage *)
1894 MemoryContextAlloc(rb->context,
1895 sizeof(SharedInvalidationMessage) * nmsgs);
1896 memcpy(txn->invalidations, msgs,
1897 sizeof(SharedInvalidationMessage) * nmsgs);
1901 * Apply all invalidations we know. Possibly we only need parts at this point
1902 * in the changestream but we don't know which those are.
1905 ReorderBufferExecuteInvalidations(ReorderBuffer *rb, ReorderBufferTXN *txn)
1909 for (i = 0; i < txn->ninvalidations; i++)
1910 LocalExecuteInvalidationMessage(&txn->invalidations[i]);
1914 * Mark a transaction as containing catalog changes
1917 ReorderBufferXidSetCatalogChanges(ReorderBuffer *rb, TransactionId xid,
1920 ReorderBufferTXN *txn;
1922 txn = ReorderBufferTXNByXid(rb, xid, true, NULL, lsn, true);
1924 txn->has_catalog_changes = true;
1928 * Query whether a transaction is already *known* to contain catalog
1929 * changes. This can be wrong until directly before the commit!
1932 ReorderBufferXidHasCatalogChanges(ReorderBuffer *rb, TransactionId xid)
1934 ReorderBufferTXN *txn;
1936 txn = ReorderBufferTXNByXid(rb, xid, false, NULL, InvalidXLogRecPtr,
1941 return txn->has_catalog_changes;
1945 * Have we already added the first snapshot?
1948 ReorderBufferXidHasBaseSnapshot(ReorderBuffer *rb, TransactionId xid)
1950 ReorderBufferTXN *txn;
1952 txn = ReorderBufferTXNByXid(rb, xid, false, NULL, InvalidXLogRecPtr,
1955 /* transaction isn't known yet, ergo no snapshot */
1960 * TODO: It would be a nice improvement if we would check the toplevel
1961 * transaction in subtransactions, but we'd need to keep track of a bit
1964 return txn->base_snapshot != NULL;
1969 * ---------------------------------------
1970 * Disk serialization support
1971 * ---------------------------------------
1975 * Ensure the IO buffer is >= sz.
1978 ReorderBufferSerializeReserve(ReorderBuffer *rb, Size sz)
1980 if (!rb->outbufsize)
1982 rb->outbuf = MemoryContextAlloc(rb->context, sz);
1983 rb->outbufsize = sz;
1985 else if (rb->outbufsize < sz)
1987 rb->outbuf = repalloc(rb->outbuf, sz);
1988 rb->outbufsize = sz;
1993 * Check whether the transaction tx should spill its data to disk.
1996 ReorderBufferCheckSerializeTXN(ReorderBuffer *rb, ReorderBufferTXN *txn)
1999 * TODO: improve accounting so we cheaply can take subtransactions into
2002 if (txn->nentries_mem >= max_changes_in_memory)
2004 ReorderBufferSerializeTXN(rb, txn);
2005 Assert(txn->nentries_mem == 0);
2010 * Spill data of a large transaction (and its subtransactions) to disk.
2013 ReorderBufferSerializeTXN(ReorderBuffer *rb, ReorderBufferTXN *txn)
2015 dlist_iter subtxn_i;
2016 dlist_mutable_iter change_i;
2018 XLogSegNo curOpenSegNo = 0;
2020 char path[MAXPGPATH];
2022 elog(DEBUG2, "spill %u changes in XID %u to disk",
2023 (uint32) txn->nentries_mem, txn->xid);
2025 /* do the same to all child TXs */
2026 dlist_foreach(subtxn_i, &txn->subtxns)
2028 ReorderBufferTXN *subtxn;
2030 subtxn = dlist_container(ReorderBufferTXN, node, subtxn_i.cur);
2031 ReorderBufferSerializeTXN(rb, subtxn);
2034 /* serialize changestream */
2035 dlist_foreach_modify(change_i, &txn->changes)
2037 ReorderBufferChange *change;
2039 change = dlist_container(ReorderBufferChange, node, change_i.cur);
2042 * store in segment in which it belongs by start lsn, don't split over
2043 * multiple segments tho
2045 if (fd == -1 || !XLByteInSeg(change->lsn, curOpenSegNo))
2050 CloseTransientFile(fd);
2052 XLByteToSeg(change->lsn, curOpenSegNo);
2053 XLogSegNoOffsetToRecPtr(curOpenSegNo, 0, recptr);
2056 * No need to care about TLIs here, only used during a single run,
2057 * so each LSN only maps to a specific WAL record.
2059 sprintf(path, "pg_replslot/%s/xid-%u-lsn-%X-%X.snap",
2060 NameStr(MyReplicationSlot->data.name), txn->xid,
2061 (uint32) (recptr >> 32), (uint32) recptr);
2063 /* open segment, create it if necessary */
2064 fd = OpenTransientFile(path,
2065 O_CREAT | O_WRONLY | O_APPEND | PG_BINARY,
2070 (errcode_for_file_access(),
2071 errmsg("could not open file \"%s\": %m",
2075 ReorderBufferSerializeChange(rb, txn, fd, change);
2076 dlist_delete(&change->node);
2077 ReorderBufferReturnChange(rb, change);
2082 Assert(spilled == txn->nentries_mem);
2083 Assert(dlist_is_empty(&txn->changes));
2084 txn->nentries_mem = 0;
2087 CloseTransientFile(fd);
2091 * Serialize individual change to disk.
2094 ReorderBufferSerializeChange(ReorderBuffer *rb, ReorderBufferTXN *txn,
2095 int fd, ReorderBufferChange *change)
2097 ReorderBufferDiskChange *ondisk;
2098 Size sz = sizeof(ReorderBufferDiskChange);
2100 ReorderBufferSerializeReserve(rb, sz);
2102 ondisk = (ReorderBufferDiskChange *) rb->outbuf;
2103 memcpy(&ondisk->change, change, sizeof(ReorderBufferChange));
2105 switch (change->action)
2107 /* fall through these, they're all similar enough */
2108 case REORDER_BUFFER_CHANGE_INSERT:
2109 case REORDER_BUFFER_CHANGE_UPDATE:
2110 case REORDER_BUFFER_CHANGE_DELETE:
2111 case REORDER_BUFFER_CHANGE_INTERNAL_SPEC_INSERT:
2114 ReorderBufferTupleBuf *oldtup,
2119 oldtup = change->data.tp.oldtuple;
2120 newtup = change->data.tp.newtuple;
2124 sz += sizeof(HeapTupleData);
2125 oldlen = oldtup->tuple.t_len;
2131 sz += sizeof(HeapTupleData);
2132 newlen = newtup->tuple.t_len;
2136 /* make sure we have enough space */
2137 ReorderBufferSerializeReserve(rb, sz);
2139 data = ((char *) rb->outbuf) + sizeof(ReorderBufferDiskChange);
2140 /* might have been reallocated above */
2141 ondisk = (ReorderBufferDiskChange *) rb->outbuf;
2145 memcpy(data, &oldtup->tuple, sizeof(HeapTupleData));
2146 data += sizeof(HeapTupleData);
2148 memcpy(data, oldtup->tuple.t_data, oldlen);
2154 memcpy(data, &newtup->tuple, sizeof(HeapTupleData));
2155 data += sizeof(HeapTupleData);
2157 memcpy(data, newtup->tuple.t_data, newlen);
2162 case REORDER_BUFFER_CHANGE_INTERNAL_SNAPSHOT:
2167 snap = change->data.snapshot;
2169 sz += sizeof(SnapshotData) +
2170 sizeof(TransactionId) * snap->xcnt +
2171 sizeof(TransactionId) * snap->subxcnt
2174 /* make sure we have enough space */
2175 ReorderBufferSerializeReserve(rb, sz);
2176 data = ((char *) rb->outbuf) + sizeof(ReorderBufferDiskChange);
2177 /* might have been reallocated above */
2178 ondisk = (ReorderBufferDiskChange *) rb->outbuf;
2180 memcpy(data, snap, sizeof(SnapshotData));
2181 data += sizeof(SnapshotData);
2185 memcpy(data, snap->xip,
2186 sizeof(TransactionId) * snap->xcnt);
2187 data += sizeof(TransactionId) * snap->xcnt;
2192 memcpy(data, snap->subxip,
2193 sizeof(TransactionId) * snap->subxcnt);
2194 data += sizeof(TransactionId) * snap->subxcnt;
2198 case REORDER_BUFFER_CHANGE_INTERNAL_SPEC_CONFIRM:
2199 case REORDER_BUFFER_CHANGE_INTERNAL_COMMAND_ID:
2200 case REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID:
2201 /* ReorderBufferChange contains everything important */
2207 if (write(fd, rb->outbuf, ondisk->size) != ondisk->size)
2209 CloseTransientFile(fd);
2211 (errcode_for_file_access(),
2212 errmsg("could not write to data file for XID %u: %m",
2216 Assert(ondisk->change.action == change->action);
2220 * Restore a number of changes spilled to disk back into memory.
2223 ReorderBufferRestoreChanges(ReorderBuffer *rb, ReorderBufferTXN *txn,
2224 int *fd, XLogSegNo *segno)
2227 XLogSegNo last_segno;
2228 dlist_mutable_iter cleanup_iter;
2230 Assert(txn->first_lsn != InvalidXLogRecPtr);
2231 Assert(txn->final_lsn != InvalidXLogRecPtr);
2233 /* free current entries, so we have memory for more */
2234 dlist_foreach_modify(cleanup_iter, &txn->changes)
2236 ReorderBufferChange *cleanup =
2237 dlist_container(ReorderBufferChange, node, cleanup_iter.cur);
2239 dlist_delete(&cleanup->node);
2240 ReorderBufferReturnChange(rb, cleanup);
2242 txn->nentries_mem = 0;
2243 Assert(dlist_is_empty(&txn->changes));
2245 XLByteToSeg(txn->final_lsn, last_segno);
2247 while (restored < max_changes_in_memory && *segno <= last_segno)
2250 ReorderBufferDiskChange *ondisk;
2255 char path[MAXPGPATH];
2260 XLByteToSeg(txn->first_lsn, *segno);
2263 Assert(*segno != 0 || dlist_is_empty(&txn->changes));
2264 XLogSegNoOffsetToRecPtr(*segno, 0, recptr);
2267 * No need to care about TLIs here, only used during a single run,
2268 * so each LSN only maps to a specific WAL record.
2270 sprintf(path, "pg_replslot/%s/xid-%u-lsn-%X-%X.snap",
2271 NameStr(MyReplicationSlot->data.name), txn->xid,
2272 (uint32) (recptr >> 32), (uint32) recptr);
2274 *fd = OpenTransientFile(path, O_RDONLY | PG_BINARY, 0);
2275 if (*fd < 0 && errno == ENOENT)
2283 (errcode_for_file_access(),
2284 errmsg("could not open file \"%s\": %m",
2290 * Read the statically sized part of a change which has information
2291 * about the total size. If we couldn't read a record, we're at the
2294 ReorderBufferSerializeReserve(rb, sizeof(ReorderBufferDiskChange));
2295 readBytes = read(*fd, rb->outbuf, sizeof(ReorderBufferDiskChange));
2300 CloseTransientFile(*fd);
2305 else if (readBytes < 0)
2307 (errcode_for_file_access(),
2308 errmsg("could not read from reorderbuffer spill file: %m")));
2309 else if (readBytes != sizeof(ReorderBufferDiskChange))
2311 (errcode_for_file_access(),
2312 errmsg("could not read from reorderbuffer spill file: read %d instead of %u bytes",
2314 (uint32) sizeof(ReorderBufferDiskChange))));
2316 ondisk = (ReorderBufferDiskChange *) rb->outbuf;
2318 ReorderBufferSerializeReserve(rb,
2319 sizeof(ReorderBufferDiskChange) + ondisk->size);
2320 ondisk = (ReorderBufferDiskChange *) rb->outbuf;
2322 readBytes = read(*fd, rb->outbuf + sizeof(ReorderBufferDiskChange),
2323 ondisk->size - sizeof(ReorderBufferDiskChange));
2327 (errcode_for_file_access(),
2328 errmsg("could not read from reorderbuffer spill file: %m")));
2329 else if (readBytes != ondisk->size - sizeof(ReorderBufferDiskChange))
2331 (errcode_for_file_access(),
2332 errmsg("could not read from reorderbuffer spill file: read %d instead of %u bytes",
2334 (uint32) (ondisk->size - sizeof(ReorderBufferDiskChange)))));
2337 * ok, read a full change from disk, now restore it into proper
2340 ReorderBufferRestoreChange(rb, txn, rb->outbuf);
2348 * Convert change from its on-disk format to in-memory format and queue it onto
2349 * the TXN's ->changes list.
2352 ReorderBufferRestoreChange(ReorderBuffer *rb, ReorderBufferTXN *txn,
2355 ReorderBufferDiskChange *ondisk;
2356 ReorderBufferChange *change;
2358 ondisk = (ReorderBufferDiskChange *) data;
2360 change = ReorderBufferGetChange(rb);
2362 /* copy static part */
2363 memcpy(change, &ondisk->change, sizeof(ReorderBufferChange));
2365 data += sizeof(ReorderBufferDiskChange);
2367 /* restore individual stuff */
2368 switch (change->action)
2370 /* fall through these, they're all similar enough */
2371 case REORDER_BUFFER_CHANGE_INSERT:
2372 case REORDER_BUFFER_CHANGE_UPDATE:
2373 case REORDER_BUFFER_CHANGE_DELETE:
2374 case REORDER_BUFFER_CHANGE_INTERNAL_SPEC_INSERT:
2375 if (change->data.tp.oldtuple)
2377 Size tuplelen = ((HeapTuple) data)->t_len;
2379 change->data.tp.oldtuple =
2380 ReorderBufferGetTupleBuf(rb, tuplelen - SizeofHeapTupleHeader);
2382 /* restore ->tuple */
2383 memcpy(&change->data.tp.oldtuple->tuple, data,
2384 sizeof(HeapTupleData));
2385 data += sizeof(HeapTupleData);
2387 /* reset t_data pointer into the new tuplebuf */
2388 change->data.tp.oldtuple->tuple.t_data =
2389 ReorderBufferTupleBufData(change->data.tp.oldtuple);
2391 /* restore tuple data itself */
2392 memcpy(change->data.tp.oldtuple->tuple.t_data, data, tuplelen);
2396 if (change->data.tp.newtuple)
2398 Size tuplelen = ((HeapTuple) data)->t_len;
2400 change->data.tp.newtuple =
2401 ReorderBufferGetTupleBuf(rb, tuplelen - SizeofHeapTupleHeader);
2403 /* restore ->tuple */
2404 memcpy(&change->data.tp.newtuple->tuple, data,
2405 sizeof(HeapTupleData));
2406 data += sizeof(HeapTupleData);
2408 /* reset t_data pointer into the new tuplebuf */
2409 change->data.tp.newtuple->tuple.t_data =
2410 ReorderBufferTupleBufData(change->data.tp.newtuple);
2412 /* restore tuple data itself */
2413 memcpy(change->data.tp.newtuple->tuple.t_data, data, tuplelen);
2418 case REORDER_BUFFER_CHANGE_INTERNAL_SNAPSHOT:
2424 oldsnap = (Snapshot) data;
2426 size = sizeof(SnapshotData) +
2427 sizeof(TransactionId) * oldsnap->xcnt +
2428 sizeof(TransactionId) * (oldsnap->subxcnt + 0);
2430 change->data.snapshot = MemoryContextAllocZero(rb->context, size);
2432 newsnap = change->data.snapshot;
2434 memcpy(newsnap, data, size);
2435 newsnap->xip = (TransactionId *)
2436 (((char *) newsnap) + sizeof(SnapshotData));
2437 newsnap->subxip = newsnap->xip + newsnap->xcnt;
2438 newsnap->copied = true;
2441 /* the base struct contains all the data, easy peasy */
2442 case REORDER_BUFFER_CHANGE_INTERNAL_SPEC_CONFIRM:
2443 case REORDER_BUFFER_CHANGE_INTERNAL_COMMAND_ID:
2444 case REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID:
2448 dlist_push_tail(&txn->changes, &change->node);
2449 txn->nentries_mem++;
2453 * Remove all on-disk stored for the passed in transaction.
2456 ReorderBufferRestoreCleanup(ReorderBuffer *rb, ReorderBufferTXN *txn)
2462 Assert(txn->first_lsn != InvalidXLogRecPtr);
2463 Assert(txn->final_lsn != InvalidXLogRecPtr);
2465 XLByteToSeg(txn->first_lsn, first);
2466 XLByteToSeg(txn->final_lsn, last);
2468 /* iterate over all possible filenames, and delete them */
2469 for (cur = first; cur <= last; cur++)
2471 char path[MAXPGPATH];
2474 XLogSegNoOffsetToRecPtr(cur, 0, recptr);
2476 sprintf(path, "pg_replslot/%s/xid-%u-lsn-%X-%X.snap",
2477 NameStr(MyReplicationSlot->data.name), txn->xid,
2478 (uint32) (recptr >> 32), (uint32) recptr);
2479 if (unlink(path) != 0 && errno != ENOENT)
2481 (errcode_for_file_access(),
2482 errmsg("could not remove file \"%s\": %m", path)));
2487 * Delete all data spilled to disk after we've restarted/crashed. It will be
2488 * recreated when the respective slots are reused.
2491 StartupReorderBuffer(void)
2494 struct dirent *logical_de;
2497 struct dirent *spill_de;
2499 logical_dir = AllocateDir("pg_replslot");
2500 while ((logical_de = ReadDir(logical_dir, "pg_replslot")) != NULL)
2502 struct stat statbuf;
2503 char path[MAXPGPATH];
2505 if (strcmp(logical_de->d_name, ".") == 0 ||
2506 strcmp(logical_de->d_name, "..") == 0)
2509 /* if it cannot be a slot, skip the directory */
2510 if (!ReplicationSlotValidateName(logical_de->d_name, DEBUG2))
2514 * ok, has to be a surviving logical slot, iterate and delete
2515 * everythign starting with xid-*
2517 sprintf(path, "pg_replslot/%s", logical_de->d_name);
2519 /* we're only creating directories here, skip if it's not our's */
2520 if (lstat(path, &statbuf) == 0 && !S_ISDIR(statbuf.st_mode))
2523 spill_dir = AllocateDir(path);
2524 while ((spill_de = ReadDir(spill_dir, path)) != NULL)
2526 if (strcmp(spill_de->d_name, ".") == 0 ||
2527 strcmp(spill_de->d_name, "..") == 0)
2530 /* only look at names that can be ours */
2531 if (strncmp(spill_de->d_name, "xid", 3) == 0)
2533 sprintf(path, "pg_replslot/%s/%s", logical_de->d_name,
2536 if (unlink(path) != 0)
2538 (errcode_for_file_access(),
2539 errmsg("could not remove file \"%s\": %m",
2545 FreeDir(logical_dir);
2548 /* ---------------------------------------
2549 * toast reassembly support
2550 * ---------------------------------------
2554 * Initialize per tuple toast reconstruction support.
2557 ReorderBufferToastInitHash(ReorderBuffer *rb, ReorderBufferTXN *txn)
2561 Assert(txn->toast_hash == NULL);
2563 memset(&hash_ctl, 0, sizeof(hash_ctl));
2564 hash_ctl.keysize = sizeof(Oid);
2565 hash_ctl.entrysize = sizeof(ReorderBufferToastEnt);
2566 hash_ctl.hcxt = rb->context;
2567 txn->toast_hash = hash_create("ReorderBufferToastHash", 5, &hash_ctl,
2568 HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
2572 * Per toast-chunk handling for toast reconstruction
2574 * Appends a toast chunk so we can reconstruct it when the tuple "owning" the
2575 * toasted Datum comes along.
2578 ReorderBufferToastAppendChunk(ReorderBuffer *rb, ReorderBufferTXN *txn,
2579 Relation relation, ReorderBufferChange *change)
2581 ReorderBufferToastEnt *ent;
2582 ReorderBufferTupleBuf *newtup;
2587 TupleDesc desc = RelationGetDescr(relation);
2591 if (txn->toast_hash == NULL)
2592 ReorderBufferToastInitHash(rb, txn);
2594 Assert(IsToastRelation(relation));
2596 newtup = change->data.tp.newtuple;
2597 chunk_id = DatumGetObjectId(fastgetattr(&newtup->tuple, 1, desc, &isnull));
2599 chunk_seq = DatumGetInt32(fastgetattr(&newtup->tuple, 2, desc, &isnull));
2602 ent = (ReorderBufferToastEnt *)
2603 hash_search(txn->toast_hash,
2610 Assert(ent->chunk_id == chunk_id);
2611 ent->num_chunks = 0;
2612 ent->last_chunk_seq = 0;
2614 ent->reconstructed = NULL;
2615 dlist_init(&ent->chunks);
2618 elog(ERROR, "got sequence entry %d for toast chunk %u instead of seq 0",
2619 chunk_seq, chunk_id);
2621 else if (found && chunk_seq != ent->last_chunk_seq + 1)
2622 elog(ERROR, "got sequence entry %d for toast chunk %u instead of seq %d",
2623 chunk_seq, chunk_id, ent->last_chunk_seq + 1);
2625 chunk = DatumGetPointer(fastgetattr(&newtup->tuple, 3, desc, &isnull));
2628 /* calculate size so we can allocate the right size at once later */
2629 if (!VARATT_IS_EXTENDED(chunk))
2630 chunksize = VARSIZE(chunk) - VARHDRSZ;
2631 else if (VARATT_IS_SHORT(chunk))
2632 /* could happen due to heap_form_tuple doing its thing */
2633 chunksize = VARSIZE_SHORT(chunk) - VARHDRSZ_SHORT;
2635 elog(ERROR, "unexpected type of toast chunk");
2637 ent->size += chunksize;
2638 ent->last_chunk_seq = chunk_seq;
2640 dlist_push_tail(&ent->chunks, &change->node);
2644 * Rejigger change->newtuple to point to in-memory toast tuples instead to
2645 * on-disk toast tuples that may not longer exist (think DROP TABLE or VACUUM).
2647 * We cannot replace unchanged toast tuples though, so those will still point
2648 * to on-disk toast data.
2651 ReorderBufferToastReplace(ReorderBuffer *rb, ReorderBufferTXN *txn,
2652 Relation relation, ReorderBufferChange *change)
2661 TupleDesc toast_desc;
2662 MemoryContext oldcontext;
2663 ReorderBufferTupleBuf *newtup;
2665 /* no toast tuples changed */
2666 if (txn->toast_hash == NULL)
2669 oldcontext = MemoryContextSwitchTo(rb->context);
2671 /* we should only have toast tuples in an INSERT or UPDATE */
2672 Assert(change->data.tp.newtuple);
2674 desc = RelationGetDescr(relation);
2676 toast_rel = RelationIdGetRelation(relation->rd_rel->reltoastrelid);
2677 toast_desc = RelationGetDescr(toast_rel);
2679 /* should we allocate from stack instead? */
2680 attrs = palloc0(sizeof(Datum) * desc->natts);
2681 isnull = palloc0(sizeof(bool) * desc->natts);
2682 free = palloc0(sizeof(bool) * desc->natts);
2684 newtup = change->data.tp.newtuple;
2686 heap_deform_tuple(&newtup->tuple, desc, attrs, isnull);
2688 for (natt = 0; natt < desc->natts; natt++)
2690 Form_pg_attribute attr = desc->attrs[natt];
2691 ReorderBufferToastEnt *ent;
2692 struct varlena *varlena;
2694 /* va_rawsize is the size of the original datum -- including header */
2695 struct varatt_external toast_pointer;
2696 struct varatt_indirect redirect_pointer;
2697 struct varlena *new_datum = NULL;
2698 struct varlena *reconstructed;
2702 /* system columns aren't toasted */
2703 if (attr->attnum < 0)
2706 if (attr->attisdropped)
2709 /* not a varlena datatype */
2710 if (attr->attlen != -1)
2717 /* ok, we know we have a toast datum */
2718 varlena = (struct varlena *) DatumGetPointer(attrs[natt]);
2720 /* no need to do anything if the tuple isn't external */
2721 if (!VARATT_IS_EXTERNAL(varlena))
2724 VARATT_EXTERNAL_GET_POINTER(toast_pointer, varlena);
2727 * Check whether the toast tuple changed, replace if so.
2729 ent = (ReorderBufferToastEnt *)
2730 hash_search(txn->toast_hash,
2731 (void *) &toast_pointer.va_valueid,
2738 (struct varlena *) palloc0(INDIRECT_POINTER_SIZE);
2742 reconstructed = palloc0(toast_pointer.va_rawsize);
2744 ent->reconstructed = reconstructed;
2746 /* stitch toast tuple back together from its parts */
2747 dlist_foreach(it, &ent->chunks)
2750 ReorderBufferChange *cchange;
2751 ReorderBufferTupleBuf *ctup;
2754 cchange = dlist_container(ReorderBufferChange, node, it.cur);
2755 ctup = cchange->data.tp.newtuple;
2756 chunk = DatumGetPointer(
2757 fastgetattr(&ctup->tuple, 3, toast_desc, &isnull));
2760 Assert(!VARATT_IS_EXTERNAL(chunk));
2761 Assert(!VARATT_IS_SHORT(chunk));
2763 memcpy(VARDATA(reconstructed) + data_done,
2765 VARSIZE(chunk) - VARHDRSZ);
2766 data_done += VARSIZE(chunk) - VARHDRSZ;
2768 Assert(data_done == toast_pointer.va_extsize);
2770 /* make sure its marked as compressed or not */
2771 if (VARATT_EXTERNAL_IS_COMPRESSED(toast_pointer))
2772 SET_VARSIZE_COMPRESSED(reconstructed, data_done + VARHDRSZ);
2774 SET_VARSIZE(reconstructed, data_done + VARHDRSZ);
2776 memset(&redirect_pointer, 0, sizeof(redirect_pointer));
2777 redirect_pointer.pointer = reconstructed;
2779 SET_VARTAG_EXTERNAL(new_datum, VARTAG_INDIRECT);
2780 memcpy(VARDATA_EXTERNAL(new_datum), &redirect_pointer,
2781 sizeof(redirect_pointer));
2783 attrs[natt] = PointerGetDatum(new_datum);
2787 * Build tuple in separate memory & copy tuple back into the tuplebuf
2788 * passed to the output plugin. We can't directly heap_fill_tuple() into
2789 * the tuplebuf because attrs[] will point back into the current content.
2791 tmphtup = heap_form_tuple(desc, attrs, isnull);
2792 Assert(newtup->tuple.t_len <= MaxHeapTupleSize);
2793 Assert(ReorderBufferTupleBufData(newtup) == newtup->tuple.t_data);
2795 memcpy(newtup->tuple.t_data, tmphtup->t_data, tmphtup->t_len);
2796 newtup->tuple.t_len = tmphtup->t_len;
2799 * free resources we won't further need, more persistent stuff will be
2800 * free'd in ReorderBufferToastReset().
2802 RelationClose(toast_rel);
2804 for (natt = 0; natt < desc->natts; natt++)
2807 pfree(DatumGetPointer(attrs[natt]));
2813 MemoryContextSwitchTo(oldcontext);
2817 * Free all resources allocated for toast reconstruction.
2820 ReorderBufferToastReset(ReorderBuffer *rb, ReorderBufferTXN *txn)
2822 HASH_SEQ_STATUS hstat;
2823 ReorderBufferToastEnt *ent;
2825 if (txn->toast_hash == NULL)
2828 /* sequentially walk over the hash and free everything */
2829 hash_seq_init(&hstat, txn->toast_hash);
2830 while ((ent = (ReorderBufferToastEnt *) hash_seq_search(&hstat)) != NULL)
2832 dlist_mutable_iter it;
2834 if (ent->reconstructed != NULL)
2835 pfree(ent->reconstructed);
2837 dlist_foreach_modify(it, &ent->chunks)
2839 ReorderBufferChange *change =
2840 dlist_container(ReorderBufferChange, node, it.cur);
2842 dlist_delete(&change->node);
2843 ReorderBufferReturnChange(rb, change);
2847 hash_destroy(txn->toast_hash);
2848 txn->toast_hash = NULL;
2852 /* ---------------------------------------
2853 * Visibility support for logical decoding
2856 * Lookup actual cmin/cmax values when using decoding snapshot. We can't
2857 * always rely on stored cmin/cmax values because of two scenarios:
2859 * * A tuple got changed multiple times during a single transaction and thus
2860 * has got a combocid. Combocid's are only valid for the duration of a
2861 * single transaction.
2862 * * A tuple with a cmin but no cmax (and thus no combocid) got
2863 * deleted/updated in another transaction than the one which created it
2864 * which we are looking at right now. As only one of cmin, cmax or combocid
2865 * is actually stored in the heap we don't have access to the value we
2868 * To resolve those problems we have a per-transaction hash of (cmin,
2869 * cmax) tuples keyed by (relfilenode, ctid) which contains the actual
2870 * (cmin, cmax) values. That also takes care of combocids by simply
2871 * not caring about them at all. As we have the real cmin/cmax values
2872 * combocids aren't interesting.
2874 * As we only care about catalog tuples here the overhead of this
2875 * hashtable should be acceptable.
2877 * Heap rewrites complicate this a bit, check rewriteheap.c for
2879 * -------------------------------------------------------------------------
2882 /* struct for qsort()ing mapping files by lsn somewhat efficiently */
2883 typedef struct RewriteMappingFile
2886 char fname[MAXPGPATH];
2887 } RewriteMappingFile;
2891 DisplayMapping(HTAB *tuplecid_data)
2893 HASH_SEQ_STATUS hstat;
2894 ReorderBufferTupleCidEnt *ent;
2896 hash_seq_init(&hstat, tuplecid_data);
2897 while ((ent = (ReorderBufferTupleCidEnt *) hash_seq_search(&hstat)) != NULL)
2899 elog(DEBUG3, "mapping: node: %u/%u/%u tid: %u/%u cmin: %u, cmax: %u",
2900 ent->key.relnode.dbNode,
2901 ent->key.relnode.spcNode,
2902 ent->key.relnode.relNode,
2903 BlockIdGetBlockNumber(&ent->key.tid.ip_blkid),
2904 ent->key.tid.ip_posid,
2913 * Apply a single mapping file to tuplecid_data.
2915 * The mapping file has to have been verified to be a) committed b) for our
2916 * transaction c) applied in LSN order.
2919 ApplyLogicalMappingFile(HTAB *tuplecid_data, Oid relid, const char *fname)
2921 char path[MAXPGPATH];
2924 LogicalRewriteMappingData map;
2926 sprintf(path, "pg_logical/mappings/%s", fname);
2927 fd = OpenTransientFile(path, O_RDONLY | PG_BINARY, 0);
2930 (errmsg("could not open file \"%s\": %m", path)));
2934 ReorderBufferTupleCidKey key;
2935 ReorderBufferTupleCidEnt *ent;
2936 ReorderBufferTupleCidEnt *new_ent;
2939 /* be careful about padding */
2940 memset(&key, 0, sizeof(ReorderBufferTupleCidKey));
2942 /* read all mappings till the end of the file */
2943 readBytes = read(fd, &map, sizeof(LogicalRewriteMappingData));
2947 (errcode_for_file_access(),
2948 errmsg("could not read file \"%s\": %m",
2950 else if (readBytes == 0) /* EOF */
2952 else if (readBytes != sizeof(LogicalRewriteMappingData))
2954 (errcode_for_file_access(),
2955 errmsg("could not read from file \"%s\": read %d instead of %d bytes",
2957 (int32) sizeof(LogicalRewriteMappingData))));
2959 key.relnode = map.old_node;
2960 ItemPointerCopy(&map.old_tid,
2964 ent = (ReorderBufferTupleCidEnt *)
2965 hash_search(tuplecid_data,
2970 /* no existing mapping, no need to update */
2974 key.relnode = map.new_node;
2975 ItemPointerCopy(&map.new_tid,
2978 new_ent = (ReorderBufferTupleCidEnt *)
2979 hash_search(tuplecid_data,
2987 * Make sure the existing mapping makes sense. We sometime update
2988 * old records that did not yet have a cmax (e.g. pg_class' own
2989 * entry while rewriting it) during rewrites, so allow that.
2991 Assert(ent->cmin == InvalidCommandId || ent->cmin == new_ent->cmin);
2992 Assert(ent->cmax == InvalidCommandId || ent->cmax == new_ent->cmax);
2996 /* update mapping */
2997 new_ent->cmin = ent->cmin;
2998 new_ent->cmax = ent->cmax;
2999 new_ent->combocid = ent->combocid;
3006 * Check whether the TransactionOId 'xid' is in the pre-sorted array 'xip'.
3009 TransactionIdInArray(TransactionId xid, TransactionId *xip, Size num)
3011 return bsearch(&xid, xip, num,
3012 sizeof(TransactionId), xidComparator) != NULL;
3016 * qsort() comparator for sorting RewriteMappingFiles in LSN order.
3019 file_sort_by_lsn(const void *a_p, const void *b_p)
3021 RewriteMappingFile *a = *(RewriteMappingFile **) a_p;
3022 RewriteMappingFile *b = *(RewriteMappingFile **) b_p;
3024 if (a->lsn < b->lsn)
3026 else if (a->lsn > b->lsn)
3032 * Apply any existing logical remapping files if there are any targeted at our
3033 * transaction for relid.
3036 UpdateLogicalMappings(HTAB *tuplecid_data, Oid relid, Snapshot snapshot)
3039 struct dirent *mapping_de;
3042 RewriteMappingFile **files_a;
3044 Oid dboid = IsSharedRelation(relid) ? InvalidOid : MyDatabaseId;
3046 mapping_dir = AllocateDir("pg_logical/mappings");
3047 while ((mapping_de = ReadDir(mapping_dir, "pg_logical/mappings")) != NULL)
3051 TransactionId f_mapped_xid;
3052 TransactionId f_create_xid;
3056 RewriteMappingFile *f;
3058 if (strcmp(mapping_de->d_name, ".") == 0 ||
3059 strcmp(mapping_de->d_name, "..") == 0)
3062 /* Ignore files that aren't ours */
3063 if (strncmp(mapping_de->d_name, "map-", 4) != 0)
3066 if (sscanf(mapping_de->d_name, LOGICAL_REWRITE_FORMAT,
3067 &f_dboid, &f_relid, &f_hi, &f_lo,
3068 &f_mapped_xid, &f_create_xid) != 6)
3069 elog(ERROR, "could not parse filename \"%s\"", mapping_de->d_name);
3071 f_lsn = ((uint64) f_hi) << 32 | f_lo;
3073 /* mapping for another database */
3074 if (f_dboid != dboid)
3077 /* mapping for another relation */
3078 if (f_relid != relid)
3081 /* did the creating transaction abort? */
3082 if (!TransactionIdDidCommit(f_create_xid))
3085 /* not for our transaction */
3086 if (!TransactionIdInArray(f_mapped_xid, snapshot->subxip, snapshot->subxcnt))
3089 /* ok, relevant, queue for apply */
3090 f = palloc(sizeof(RewriteMappingFile));
3092 strcpy(f->fname, mapping_de->d_name);
3093 files = lappend(files, f);
3095 FreeDir(mapping_dir);
3097 /* build array we can easily sort */
3098 files_a = palloc(list_length(files) * sizeof(RewriteMappingFile *));
3100 foreach(file, files)
3102 files_a[off++] = lfirst(file);
3105 /* sort files so we apply them in LSN order */
3106 qsort(files_a, list_length(files), sizeof(RewriteMappingFile *),
3109 for (off = 0; off < list_length(files); off++)
3111 RewriteMappingFile *f = files_a[off];
3113 elog(DEBUG1, "applying mapping: \"%s\" in %u", f->fname,
3114 snapshot->subxip[0]);
3115 ApplyLogicalMappingFile(tuplecid_data, relid, f->fname);
3121 * Lookup cmin/cmax of a tuple, during logical decoding where we can't rely on
3125 ResolveCminCmaxDuringDecoding(HTAB *tuplecid_data,
3127 HeapTuple htup, Buffer buffer,
3128 CommandId *cmin, CommandId *cmax)
3130 ReorderBufferTupleCidKey key;
3131 ReorderBufferTupleCidEnt *ent;
3133 BlockNumber blockno;
3134 bool updated_mapping = false;
3136 /* be careful about padding */
3137 memset(&key, 0, sizeof(key));
3139 Assert(!BufferIsLocal(buffer));
3142 * get relfilenode from the buffer, no convenient way to access it other
3145 BufferGetTag(buffer, &key.relnode, &forkno, &blockno);
3147 /* tuples can only be in the main fork */
3148 Assert(forkno == MAIN_FORKNUM);
3149 Assert(blockno == ItemPointerGetBlockNumber(&htup->t_self));
3151 ItemPointerCopy(&htup->t_self,
3155 ent = (ReorderBufferTupleCidEnt *)
3156 hash_search(tuplecid_data,
3162 * failed to find a mapping, check whether the table was rewritten and
3163 * apply mapping if so, but only do that once - there can be no new
3164 * mappings while we are in here since we have to hold a lock on the
3167 if (ent == NULL && !updated_mapping)
3169 UpdateLogicalMappings(tuplecid_data, htup->t_tableOid, snapshot);
3170 /* now check but don't update for a mapping again */
3171 updated_mapping = true;
3174 else if (ent == NULL)