1 /*-------------------------------------------------------------------------
4 * PostgreSQL logical replay/reorder buffer management
7 * Copyright (c) 2012-2014, 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/relcache.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 probaly 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 a 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 a 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 if (change->data.tp.newtuple)
406 ReorderBufferReturnTupleBuf(rb, change->data.tp.newtuple);
407 change->data.tp.newtuple = NULL;
410 if (change->data.tp.oldtuple)
412 ReorderBufferReturnTupleBuf(rb, change->data.tp.oldtuple);
413 change->data.tp.oldtuple = NULL;
416 case REORDER_BUFFER_CHANGE_INTERNAL_SNAPSHOT:
417 if (change->data.snapshot)
419 ReorderBufferFreeSnap(rb, change->data.snapshot);
420 change->data.snapshot = NULL;
423 case REORDER_BUFFER_CHANGE_INTERNAL_COMMAND_ID:
425 case REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID:
429 /* check whether to put into the slab cache */
430 if (rb->nr_cached_changes < max_cached_changes)
432 rb->nr_cached_changes++;
433 dlist_push_head(&rb->cached_changes, &change->node);
434 VALGRIND_MAKE_MEM_UNDEFINED(change, sizeof(ReorderBufferChange));
435 VALGRIND_MAKE_MEM_DEFINED(&change->node, sizeof(change->node));
445 * Get a unused, possibly preallocated, ReorderBufferTupleBuf
447 ReorderBufferTupleBuf *
448 ReorderBufferGetTupleBuf(ReorderBuffer *rb)
450 ReorderBufferTupleBuf *tuple;
452 /* check the slab cache */
453 if (rb->nr_cached_tuplebufs)
455 rb->nr_cached_tuplebufs--;
456 tuple = slist_container(ReorderBufferTupleBuf, node,
457 slist_pop_head_node(&rb->cached_tuplebufs));
458 #ifdef USE_ASSERT_CHECKING
459 memset(tuple, 0xa9, sizeof(ReorderBufferTupleBuf));
464 tuple = (ReorderBufferTupleBuf *)
465 MemoryContextAlloc(rb->context, sizeof(ReorderBufferTupleBuf));
472 * Free an ReorderBufferTupleBuf.
474 * Deallocation might be delayed for efficiency purposes, for details check
475 * the comments above max_cached_changes's definition.
478 ReorderBufferReturnTupleBuf(ReorderBuffer *rb, ReorderBufferTupleBuf *tuple)
480 /* check whether to put into the slab cache */
481 if (rb->nr_cached_tuplebufs < max_cached_tuplebufs)
483 rb->nr_cached_tuplebufs++;
484 slist_push_head(&rb->cached_tuplebufs, &tuple->node);
485 VALGRIND_MAKE_MEM_UNDEFINED(tuple, sizeof(ReorderBufferTupleBuf));
486 VALGRIND_MAKE_MEM_DEFINED(&tuple->node, sizeof(tuple->node));
495 * Return the ReorderBufferTXN from the given buffer, specified by Xid.
496 * If create is true, and a transaction doesn't already exist, create it
497 * (with the given LSN, and as top transaction if that's specified);
498 * when this happens, is_new is set to true.
500 static ReorderBufferTXN *
501 ReorderBufferTXNByXid(ReorderBuffer *rb, TransactionId xid, bool create,
502 bool *is_new, XLogRecPtr lsn, bool create_as_top)
504 ReorderBufferTXN *txn;
505 ReorderBufferTXNByIdEnt *ent;
508 Assert(TransactionIdIsValid(xid));
509 Assert(!create || lsn != InvalidXLogRecPtr);
512 * Check the one-entry lookup cache first
514 if (TransactionIdIsValid(rb->by_txn_last_xid) &&
515 rb->by_txn_last_xid == xid)
517 txn = rb->by_txn_last_txn;
521 /* found it, and it's valid */
528 * cached as non-existant, and asked not to create? Then nothing else
533 /* otherwise fall through to create it */
537 * If the cache wasn't hit or it yielded an "does-not-exist" and we want
538 * to create an entry.
541 /* search the lookup table */
542 ent = (ReorderBufferTXNByIdEnt *)
543 hash_search(rb->by_txn,
545 create ? HASH_ENTER : HASH_FIND,
551 /* initialize the new entry, if creation was requested */
554 ent->txn = ReorderBufferGetTXN(rb);
557 txn->first_lsn = lsn;
558 txn->restart_decoding_lsn = rb->current_restart_decoding_lsn;
562 dlist_push_tail(&rb->toplevel_by_lsn, &txn->node);
563 AssertTXNLsnOrder(rb);
567 txn = NULL; /* not found and not asked to create */
570 rb->by_txn_last_xid = xid;
571 rb->by_txn_last_txn = txn;
576 Assert(!create || !!txn);
581 * Queue a change into a transaction so it can be replayed upon commit.
584 ReorderBufferQueueChange(ReorderBuffer *rb, TransactionId xid, XLogRecPtr lsn,
585 ReorderBufferChange *change)
587 ReorderBufferTXN *txn;
589 txn = ReorderBufferTXNByXid(rb, xid, true, NULL, lsn, true);
592 Assert(InvalidXLogRecPtr != lsn);
593 dlist_push_tail(&txn->changes, &change->node);
597 ReorderBufferCheckSerializeTXN(rb, txn);
601 AssertTXNLsnOrder(ReorderBuffer *rb)
603 #ifdef USE_ASSERT_CHECKING
605 XLogRecPtr prev_first_lsn = InvalidXLogRecPtr;
607 dlist_foreach(iter, &rb->toplevel_by_lsn)
609 ReorderBufferTXN *cur_txn;
611 cur_txn = dlist_container(ReorderBufferTXN, node, iter.cur);
612 Assert(cur_txn->first_lsn != InvalidXLogRecPtr);
614 if (cur_txn->end_lsn != InvalidXLogRecPtr)
615 Assert(cur_txn->first_lsn <= cur_txn->end_lsn);
617 if (prev_first_lsn != InvalidXLogRecPtr)
618 Assert(prev_first_lsn < cur_txn->first_lsn);
620 Assert(!cur_txn->is_known_as_subxact);
621 prev_first_lsn = cur_txn->first_lsn;
627 ReorderBufferGetOldestTXN(ReorderBuffer *rb)
629 ReorderBufferTXN *txn;
631 if (dlist_is_empty(&rb->toplevel_by_lsn))
634 AssertTXNLsnOrder(rb);
636 txn = dlist_head_element(ReorderBufferTXN, node, &rb->toplevel_by_lsn);
638 Assert(!txn->is_known_as_subxact);
639 Assert(txn->first_lsn != InvalidXLogRecPtr);
644 ReorderBufferSetRestartPoint(ReorderBuffer *rb, XLogRecPtr ptr)
646 rb->current_restart_decoding_lsn = ptr;
650 ReorderBufferAssignChild(ReorderBuffer *rb, TransactionId xid,
651 TransactionId subxid, XLogRecPtr lsn)
653 ReorderBufferTXN *txn;
654 ReorderBufferTXN *subtxn;
658 txn = ReorderBufferTXNByXid(rb, xid, true, &new_top, lsn, true);
659 subtxn = ReorderBufferTXNByXid(rb, subxid, true, &new_sub, lsn, false);
664 * we assign subtransactions to top level transaction even if we don't
665 * have data for it yet, assignment records frequently reference xids
666 * that have not yet produced any records. Knowing those aren't top
667 * level xids allows us to make processing cheaper in some places.
669 dlist_push_tail(&txn->subtxns, &subtxn->node);
672 else if (!subtxn->is_known_as_subxact)
674 subtxn->is_known_as_subxact = true;
675 Assert(subtxn->nsubtxns == 0);
677 /* remove from lsn order list of top-level transactions */
678 dlist_delete(&subtxn->node);
680 /* add to toplevel transaction */
681 dlist_push_tail(&txn->subtxns, &subtxn->node);
686 elog(ERROR, "existing subxact assigned to unknown toplevel xact");
691 * Associate a subtransaction with its toplevel transaction at commit
692 * time. There may be no further changes added after this.
695 ReorderBufferCommitChild(ReorderBuffer *rb, TransactionId xid,
696 TransactionId subxid, XLogRecPtr commit_lsn,
699 ReorderBufferTXN *txn;
700 ReorderBufferTXN *subtxn;
702 subtxn = ReorderBufferTXNByXid(rb, subxid, false, NULL,
703 InvalidXLogRecPtr, false);
706 * No need to do anything if that subtxn didn't contain any changes
711 txn = ReorderBufferTXNByXid(rb, xid, false, NULL, commit_lsn, true);
714 elog(ERROR, "subxact logged without previous toplevel record");
717 * Pass the our base snapshot to the parent transaction if it doesn't have
718 * one, or ours is older. That can happen if there are no changes in the
719 * toplevel transaction but in one of the child transactions. This allows
720 * the parent to simply use it's base snapshot initially.
722 if (txn->base_snapshot == NULL ||
723 txn->base_snapshot_lsn > subtxn->base_snapshot_lsn)
725 txn->base_snapshot = subtxn->base_snapshot;
726 txn->base_snapshot_lsn = subtxn->base_snapshot_lsn;
727 subtxn->base_snapshot = NULL;
728 subtxn->base_snapshot_lsn = InvalidXLogRecPtr;
731 subtxn->final_lsn = commit_lsn;
732 subtxn->end_lsn = end_lsn;
734 if (!subtxn->is_known_as_subxact)
736 subtxn->is_known_as_subxact = true;
737 Assert(subtxn->nsubtxns == 0);
739 /* remove from lsn order list of top-level transactions */
740 dlist_delete(&subtxn->node);
742 /* add to subtransaction list */
743 dlist_push_tail(&txn->subtxns, &subtxn->node);
750 * Support for efficiently iterating over a transaction's and its
751 * subtransactions' changes.
753 * We do by doing a k-way merge between transactions/subtransactions. For that
754 * we model the current heads of the different transactions as a binary heap
755 * so we easily know which (sub-)transaction has the change with the smallest
758 * We assume the changes in individual transactions are already sorted by LSN.
762 * Binary heap comparison function.
765 ReorderBufferIterCompare(Datum a, Datum b, void *arg)
767 ReorderBufferIterTXNState *state = (ReorderBufferIterTXNState *) arg;
768 XLogRecPtr pos_a = state->entries[DatumGetInt32(a)].lsn;
769 XLogRecPtr pos_b = state->entries[DatumGetInt32(b)].lsn;
773 else if (pos_a == pos_b)
779 * Allocate & initialize an iterator which iterates in lsn order over a
780 * transaction and all its subtransactions.
782 static ReorderBufferIterTXNState *
783 ReorderBufferIterTXNInit(ReorderBuffer *rb, ReorderBufferTXN *txn)
786 ReorderBufferIterTXNState *state;
787 dlist_iter cur_txn_i;
791 * Calculate the size of our heap: one element for every transaction that
792 * contains changes. (Besides the transactions already in the reorder
793 * buffer, we count the one we were directly passed.)
795 if (txn->nentries > 0)
798 dlist_foreach(cur_txn_i, &txn->subtxns)
800 ReorderBufferTXN *cur_txn;
802 cur_txn = dlist_container(ReorderBufferTXN, node, cur_txn_i.cur);
804 if (cur_txn->nentries > 0)
809 * TODO: Consider adding fastpath for the rather common nr_txns=1 case, no
810 * need to allocate/build a heap then.
813 /* allocate iteration state */
814 state = (ReorderBufferIterTXNState *)
815 MemoryContextAllocZero(rb->context,
816 sizeof(ReorderBufferIterTXNState) +
817 sizeof(ReorderBufferIterTXNEntry) * nr_txns);
819 state->nr_txns = nr_txns;
820 dlist_init(&state->old_change);
822 for (off = 0; off < state->nr_txns; off++)
824 state->entries[off].fd = -1;
825 state->entries[off].segno = 0;
829 state->heap = binaryheap_allocate(state->nr_txns,
830 ReorderBufferIterCompare,
834 * Now insert items into the binary heap, in an unordered fashion. (We
835 * will run a heap assembly step at the end; this is more efficient.)
840 /* add toplevel transaction if it contains changes */
841 if (txn->nentries > 0)
843 ReorderBufferChange *cur_change;
845 if (txn->nentries != txn->nentries_mem)
846 ReorderBufferRestoreChanges(rb, txn, &state->entries[off].fd,
847 &state->entries[off].segno);
849 cur_change = dlist_head_element(ReorderBufferChange, node,
852 state->entries[off].lsn = cur_change->lsn;
853 state->entries[off].change = cur_change;
854 state->entries[off].txn = txn;
856 binaryheap_add_unordered(state->heap, Int32GetDatum(off++));
859 /* add subtransactions if they contain changes */
860 dlist_foreach(cur_txn_i, &txn->subtxns)
862 ReorderBufferTXN *cur_txn;
864 cur_txn = dlist_container(ReorderBufferTXN, node, cur_txn_i.cur);
866 if (cur_txn->nentries > 0)
868 ReorderBufferChange *cur_change;
870 if (txn->nentries != txn->nentries_mem)
871 ReorderBufferRestoreChanges(rb, cur_txn,
872 &state->entries[off].fd,
873 &state->entries[off].segno);
875 cur_change = dlist_head_element(ReorderBufferChange, node,
878 state->entries[off].lsn = cur_change->lsn;
879 state->entries[off].change = cur_change;
880 state->entries[off].txn = cur_txn;
882 binaryheap_add_unordered(state->heap, Int32GetDatum(off++));
886 /* assemble a valid binary heap */
887 binaryheap_build(state->heap);
893 * Return the next change when iterating over a transaction and its
896 * Returns NULL when no further changes exist.
898 static ReorderBufferChange *
899 ReorderBufferIterTXNNext(ReorderBuffer *rb, ReorderBufferIterTXNState *state)
901 ReorderBufferChange *change;
902 ReorderBufferIterTXNEntry *entry;
905 /* nothing there anymore */
906 if (state->heap->bh_size == 0)
909 off = DatumGetInt32(binaryheap_first(state->heap));
910 entry = &state->entries[off];
912 /* free memory we might have "leaked" in the previous *Next call */
913 if (!dlist_is_empty(&state->old_change))
915 change = dlist_container(ReorderBufferChange, node,
916 dlist_pop_head_node(&state->old_change));
917 ReorderBufferReturnChange(rb, change);
918 Assert(dlist_is_empty(&state->old_change));
921 change = entry->change;
924 * update heap with information about which transaction has the next
925 * relevant change in LSN order
928 /* there are in-memory changes */
929 if (dlist_has_next(&entry->txn->changes, &entry->change->node))
931 dlist_node *next = dlist_next_node(&entry->txn->changes, &change->node);
932 ReorderBufferChange *next_change =
933 dlist_container(ReorderBufferChange, node, next);
935 /* txn stays the same */
936 state->entries[off].lsn = next_change->lsn;
937 state->entries[off].change = next_change;
939 binaryheap_replace_first(state->heap, Int32GetDatum(off));
943 /* try to load changes from disk */
944 if (entry->txn->nentries != entry->txn->nentries_mem)
947 * Ugly: restoring changes will reuse *Change records, thus delete the
948 * current one from the per-tx list and only free in the next call.
950 dlist_delete(&change->node);
951 dlist_push_tail(&state->old_change, &change->node);
953 if (ReorderBufferRestoreChanges(rb, entry->txn, &entry->fd,
954 &state->entries[off].segno))
956 /* successfully restored changes from disk */
957 ReorderBufferChange *next_change =
958 dlist_head_element(ReorderBufferChange, node,
959 &entry->txn->changes);
961 elog(DEBUG2, "restored %u/%u changes from disk",
962 (uint32) entry->txn->nentries_mem,
963 (uint32) entry->txn->nentries);
965 Assert(entry->txn->nentries_mem);
966 /* txn stays the same */
967 state->entries[off].lsn = next_change->lsn;
968 state->entries[off].change = next_change;
969 binaryheap_replace_first(state->heap, Int32GetDatum(off));
975 /* ok, no changes there anymore, remove */
976 binaryheap_remove_first(state->heap);
982 * Deallocate the iterator
985 ReorderBufferIterTXNFinish(ReorderBuffer *rb,
986 ReorderBufferIterTXNState *state)
990 for (off = 0; off < state->nr_txns; off++)
992 if (state->entries[off].fd != -1)
993 CloseTransientFile(state->entries[off].fd);
996 /* free memory we might have "leaked" in the last *Next call */
997 if (!dlist_is_empty(&state->old_change))
999 ReorderBufferChange *change;
1001 change = dlist_container(ReorderBufferChange, node,
1002 dlist_pop_head_node(&state->old_change));
1003 ReorderBufferReturnChange(rb, change);
1004 Assert(dlist_is_empty(&state->old_change));
1007 binaryheap_free(state->heap);
1012 * Cleanup the contents of a transaction, usually after the transaction
1013 * committed or aborted.
1016 ReorderBufferCleanupTXN(ReorderBuffer *rb, ReorderBufferTXN *txn)
1019 dlist_mutable_iter iter;
1021 /* cleanup subtransactions & their changes */
1022 dlist_foreach_modify(iter, &txn->subtxns)
1024 ReorderBufferTXN *subtxn;
1026 subtxn = dlist_container(ReorderBufferTXN, node, iter.cur);
1029 * Subtransactions are always associated to the toplevel TXN, even if
1030 * they originally were happening inside another subtxn, so we won't
1031 * ever recurse more than one level deep here.
1033 Assert(subtxn->is_known_as_subxact);
1034 Assert(subtxn->nsubtxns == 0);
1036 ReorderBufferCleanupTXN(rb, subtxn);
1039 /* cleanup changes in the toplevel txn */
1040 dlist_foreach_modify(iter, &txn->changes)
1042 ReorderBufferChange *change;
1044 change = dlist_container(ReorderBufferChange, node, iter.cur);
1046 ReorderBufferReturnChange(rb, change);
1050 * Cleanup the tuplecids we stored for decoding catalog snapshot access.
1051 * They are always stored in the toplevel transaction.
1053 dlist_foreach_modify(iter, &txn->tuplecids)
1055 ReorderBufferChange *change;
1057 change = dlist_container(ReorderBufferChange, node, iter.cur);
1058 Assert(change->action == REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID);
1059 ReorderBufferReturnChange(rb, change);
1062 if (txn->base_snapshot != NULL)
1064 SnapBuildSnapDecRefcount(txn->base_snapshot);
1065 txn->base_snapshot = NULL;
1066 txn->base_snapshot_lsn = InvalidXLogRecPtr;
1069 /* delete from list of known subxacts */
1070 if (txn->is_known_as_subxact)
1072 /* NB: nsubxacts count of parent will be too high now */
1073 dlist_delete(&txn->node);
1075 /* delete from LSN ordered list of toplevel TXNs */
1078 dlist_delete(&txn->node);
1081 /* now remove reference from buffer */
1082 hash_search(rb->by_txn,
1088 /* remove entries spilled to disk */
1089 if (txn->nentries != txn->nentries_mem)
1090 ReorderBufferRestoreCleanup(rb, txn);
1093 ReorderBufferReturnTXN(rb, txn);
1097 * Build a hash with a (relfilenode, ctid) -> (cmin, cmax) mapping for use by
1098 * tqual.c's HeapTupleSatisfiesHistoricMVCC.
1101 ReorderBufferBuildTupleCidHash(ReorderBuffer *rb, ReorderBufferTXN *txn)
1106 if (!txn->has_catalog_changes || dlist_is_empty(&txn->tuplecids))
1109 memset(&hash_ctl, 0, sizeof(hash_ctl));
1111 hash_ctl.keysize = sizeof(ReorderBufferTupleCidKey);
1112 hash_ctl.entrysize = sizeof(ReorderBufferTupleCidEnt);
1113 hash_ctl.hcxt = rb->context;
1116 * create the hash with the exact number of to-be-stored tuplecids from
1119 txn->tuplecid_hash =
1120 hash_create("ReorderBufferTupleCid", txn->ntuplecids, &hash_ctl,
1121 HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1123 dlist_foreach(iter, &txn->tuplecids)
1125 ReorderBufferTupleCidKey key;
1126 ReorderBufferTupleCidEnt *ent;
1128 ReorderBufferChange *change;
1130 change = dlist_container(ReorderBufferChange, node, iter.cur);
1132 Assert(change->action == REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID);
1134 /* be careful about padding */
1135 memset(&key, 0, sizeof(ReorderBufferTupleCidKey));
1137 key.relnode = change->data.tuplecid.node;
1139 ItemPointerCopy(&change->data.tuplecid.tid,
1142 ent = (ReorderBufferTupleCidEnt *)
1143 hash_search(txn->tuplecid_hash,
1145 HASH_ENTER | HASH_FIND,
1149 ent->cmin = change->data.tuplecid.cmin;
1150 ent->cmax = change->data.tuplecid.cmax;
1151 ent->combocid = change->data.tuplecid.combocid;
1155 Assert(ent->cmin == change->data.tuplecid.cmin);
1156 Assert(ent->cmax == InvalidCommandId ||
1157 ent->cmax == change->data.tuplecid.cmax);
1160 * if the tuple got valid in this transaction and now got deleted
1161 * we already have a valid cmin stored. The cmax will be
1162 * InvalidCommandId though.
1164 ent->cmax = change->data.tuplecid.cmax;
1170 * Copy a provided snapshot so we can modify it privately. This is needed so
1171 * that catalog modifying transactions can look into intermediate catalog
1175 ReorderBufferCopySnap(ReorderBuffer *rb, Snapshot orig_snap,
1176 ReorderBufferTXN *txn, CommandId cid)
1183 size = sizeof(SnapshotData) +
1184 sizeof(TransactionId) * orig_snap->xcnt +
1185 sizeof(TransactionId) * (txn->nsubtxns + 1);
1187 snap = MemoryContextAllocZero(rb->context, size);
1188 memcpy(snap, orig_snap, sizeof(SnapshotData));
1190 snap->copied = true;
1191 snap->active_count = 0;
1192 snap->regd_count = 1;
1193 snap->xip = (TransactionId *) (snap + 1);
1195 memcpy(snap->xip, orig_snap->xip, sizeof(TransactionId) * snap->xcnt);
1198 * snap->subxip contains all txids that belong to our transaction which we
1199 * need to check via cmin/cmax. Thats why we store the toplevel
1200 * transaction in there as well.
1202 snap->subxip = snap->xip + snap->xcnt;
1203 snap->subxip[i++] = txn->xid;
1206 * nsubxcnt isn't decreased when subtransactions abort, so count manually.
1207 * Since it's an upper boundary it is safe to use it for the allocation
1212 dlist_foreach(iter, &txn->subtxns)
1214 ReorderBufferTXN *sub_txn;
1216 sub_txn = dlist_container(ReorderBufferTXN, node, iter.cur);
1217 snap->subxip[i++] = sub_txn->xid;
1221 /* sort so we can bsearch() later */
1222 qsort(snap->subxip, snap->subxcnt, sizeof(TransactionId), xidComparator);
1224 /* store the specified current CommandId */
1231 * Free a previously ReorderBufferCopySnap'ed snapshot
1234 ReorderBufferFreeSnap(ReorderBuffer *rb, Snapshot snap)
1239 SnapBuildSnapDecRefcount(snap);
1243 * Perform the replay of a transaction and it's non-aborted subtransactions.
1245 * Subtransactions previously have to be processed by
1246 * ReorderBufferCommitChild(), even if previously assigned to the toplevel
1247 * transaction with ReorderBufferAssignChild.
1249 * We currently can only decode a transaction's contents in when their commit
1250 * record is read because that's currently the only place where we know about
1251 * cache invalidations. Thus, once a toplevel commit is read, we iterate over
1252 * the top and subtransactions (using a k-way merge) and replay the changes in
1256 ReorderBufferCommit(ReorderBuffer *rb, TransactionId xid,
1257 XLogRecPtr commit_lsn, XLogRecPtr end_lsn,
1258 TimestampTz commit_time)
1260 ReorderBufferTXN *txn;
1261 ReorderBufferIterTXNState *iterstate = NULL;
1262 ReorderBufferChange *change;
1264 volatile CommandId command_id = FirstCommandId;
1265 volatile Snapshot snapshot_now = NULL;
1266 volatile bool using_subtxn = false;
1268 txn = ReorderBufferTXNByXid(rb, xid, false, NULL, InvalidXLogRecPtr,
1271 /* unknown transaction, nothing to replay */
1275 txn->final_lsn = commit_lsn;
1276 txn->end_lsn = end_lsn;
1277 txn->commit_time = commit_time;
1279 /* serialize the last bunch of changes if we need start earlier anyway */
1280 if (txn->nentries_mem != txn->nentries)
1281 ReorderBufferSerializeTXN(rb, txn);
1284 * If this transaction didn't have any real changes in our database, it's
1285 * OK not to have a snapshot. Note that ReorderBufferCommitChild will have
1286 * transferred its snapshot to this transaction if it had one and the
1287 * toplevel tx didn't.
1289 if (txn->base_snapshot == NULL)
1291 Assert(txn->ninvalidations == 0);
1292 ReorderBufferCleanupTXN(rb, txn);
1296 snapshot_now = txn->base_snapshot;
1298 /* build data to be able to lookup the CommandIds of catalog tuples */
1299 ReorderBufferBuildTupleCidHash(rb, txn);
1301 /* setup the initial snapshot */
1302 SetupHistoricSnapshot(snapshot_now, txn->tuplecid_hash);
1308 * Decoding needs access to syscaches et al., which in turn use
1309 * heavyweight locks and such. Thus we need to have enough state
1310 * around to keep track of those. The easiest way is to simply use a
1311 * transaction internally. That also allows us to easily enforce that
1312 * nothing writes to the database by checking for xid assignments.
1314 * When we're called via the SQL SRF there's already a transaction
1315 * started, so start an explicit subtransaction there.
1317 using_subtxn = IsTransactionOrTransactionBlock();
1320 BeginInternalSubTransaction("replay");
1322 StartTransactionCommand();
1326 iterstate = ReorderBufferIterTXNInit(rb, txn);
1327 while ((change = ReorderBufferIterTXNNext(rb, iterstate)))
1329 Relation relation = NULL;
1332 switch (change->action)
1334 case REORDER_BUFFER_CHANGE_INSERT:
1335 case REORDER_BUFFER_CHANGE_UPDATE:
1336 case REORDER_BUFFER_CHANGE_DELETE:
1337 Assert(snapshot_now);
1339 reloid = RelidByRelfilenode(change->data.tp.relnode.spcNode,
1340 change->data.tp.relnode.relNode);
1343 * Catalog tuple without data, emitted while catalog was
1344 * in the process of being rewritten.
1346 if (reloid == InvalidOid &&
1347 change->data.tp.newtuple == NULL &&
1348 change->data.tp.oldtuple == NULL)
1350 else if (reloid == InvalidOid)
1351 elog(ERROR, "could not map filenode \"%s\" to relation OID",
1352 relpathperm(change->data.tp.relnode,
1355 relation = RelationIdGetRelation(reloid);
1357 if (relation == NULL)
1358 elog(ERROR, "could not open relation with OID %u (for filenode \"%s\")",
1360 relpathperm(change->data.tp.relnode,
1363 if (RelationIsLogicallyLogged(relation))
1366 * For now ignore sequence changes entirely. Most of
1367 * the time they don't log changes using records we
1368 * understand, so it doesn't make sense to handle the
1371 if (relation->rd_rel->relkind == RELKIND_SEQUENCE)
1374 /* user-triggered change */
1375 else if (!IsToastRelation(relation))
1377 ReorderBufferToastReplace(rb, txn, relation, change);
1378 rb->apply_change(rb, txn, relation, change);
1381 * Only clear reassembled toast chunks if we're
1382 * sure they're not required anymore. The creator
1383 * of the tuple tells us.
1385 if (change->data.tp.clear_toast_afterwards)
1386 ReorderBufferToastReset(rb, txn);
1388 /* we're not interested in toast deletions */
1389 else if (change->action == REORDER_BUFFER_CHANGE_INSERT)
1392 * Need to reassemble the full toasted Datum in
1393 * memory, to ensure the chunks don't get reused
1394 * till we're done remove it from the list of this
1395 * transaction's changes. Otherwise it will get
1396 * freed/reused while restoring spooled data from
1399 dlist_delete(&change->node);
1400 ReorderBufferToastAppendChunk(rb, txn, relation,
1405 RelationClose(relation);
1407 case REORDER_BUFFER_CHANGE_INTERNAL_SNAPSHOT:
1408 /* get rid of the old */
1409 TeardownHistoricSnapshot(false);
1411 if (snapshot_now->copied)
1413 ReorderBufferFreeSnap(rb, snapshot_now);
1415 ReorderBufferCopySnap(rb, change->data.snapshot,
1420 * Restored from disk, need to be careful not to double
1421 * free. We could introduce refcounting for that, but for
1422 * now this seems infrequent enough not to care.
1424 else if (change->data.snapshot->copied)
1427 ReorderBufferCopySnap(rb, change->data.snapshot,
1432 snapshot_now = change->data.snapshot;
1436 /* and continue with the new one */
1437 SetupHistoricSnapshot(snapshot_now, txn->tuplecid_hash);
1440 case REORDER_BUFFER_CHANGE_INTERNAL_COMMAND_ID:
1441 Assert(change->data.command_id != InvalidCommandId);
1443 if (command_id < change->data.command_id)
1445 command_id = change->data.command_id;
1447 if (!snapshot_now->copied)
1449 /* we don't use the global one anymore */
1450 snapshot_now = ReorderBufferCopySnap(rb, snapshot_now,
1454 snapshot_now->curcid = command_id;
1456 TeardownHistoricSnapshot(false);
1457 SetupHistoricSnapshot(snapshot_now, txn->tuplecid_hash);
1460 * Every time the CommandId is incremented, we could
1461 * see new catalog contents, so execute all
1464 ReorderBufferExecuteInvalidations(rb, txn);
1469 case REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID:
1470 elog(ERROR, "tuplecid value in changequeue");
1475 ReorderBufferIterTXNFinish(rb, iterstate);
1477 /* call commit callback */
1478 rb->commit(rb, txn, commit_lsn);
1480 /* this is just a sanity check against bad output plugin behaviour */
1481 if (GetCurrentTransactionIdIfAny() != InvalidTransactionId)
1482 elog(ERROR, "output plugin used XID %u",
1483 GetCurrentTransactionId());
1486 TeardownHistoricSnapshot(false);
1489 * Aborting the current (sub-)transaction as a whole has the right
1490 * semantics. We want all locks acquired in here to be released, not
1491 * reassigned to the parent and we do not want any database access
1492 * have persistent effects.
1494 AbortCurrentTransaction();
1496 /* make sure there's no cache pollution */
1497 ReorderBufferExecuteInvalidations(rb, txn);
1500 RollbackAndReleaseCurrentSubTransaction();
1502 if (snapshot_now->copied)
1503 ReorderBufferFreeSnap(rb, snapshot_now);
1505 /* remove potential on-disk data, and deallocate */
1506 ReorderBufferCleanupTXN(rb, txn);
1510 /* TODO: Encapsulate cleanup from the PG_TRY and PG_CATCH blocks */
1512 ReorderBufferIterTXNFinish(rb, iterstate);
1514 TeardownHistoricSnapshot(true);
1517 * Force cache invalidation to happen outside of a valid transaction
1518 * to prevent catalog access as we just caught an error.
1520 AbortCurrentTransaction();
1522 /* make sure there's no cache pollution */
1523 ReorderBufferExecuteInvalidations(rb, txn);
1526 RollbackAndReleaseCurrentSubTransaction();
1528 if (snapshot_now->copied)
1529 ReorderBufferFreeSnap(rb, snapshot_now);
1531 /* remove potential on-disk data, and deallocate */
1532 ReorderBufferCleanupTXN(rb, txn);
1540 * Abort a transaction that possibly has previous changes. Needs to be first
1541 * called for subtransactions and then for the toplevel xid.
1543 * NB: Transactions handled here have to have actively aborted (i.e. have
1544 * produced an abort record). Implicitly aborted transactions are handled via
1545 * ReorderBufferAbortOld(); transactions we're just not interesteded in, but
1546 * which have committed are handled in ReorderBufferForget().
1548 * This function purges this transaction and its contents from memory and
1552 ReorderBufferAbort(ReorderBuffer *rb, TransactionId xid, XLogRecPtr lsn)
1554 ReorderBufferTXN *txn;
1556 txn = ReorderBufferTXNByXid(rb, xid, false, NULL, InvalidXLogRecPtr,
1559 /* unknown, nothing to remove */
1564 txn->final_lsn = lsn;
1566 /* remove potential on-disk data, and deallocate */
1567 ReorderBufferCleanupTXN(rb, txn);
1571 * Abort all transactions that aren't actually running anymore because the
1574 * NB: These really have to be transactions that have aborted due to a server
1575 * crash/immediate restart, as we don't deal with invalidations here.
1578 ReorderBufferAbortOld(ReorderBuffer *rb, TransactionId oldestRunningXid)
1580 dlist_mutable_iter it;
1583 * Iterate through all (potential) toplevel TXNs and abort all that are
1584 * older than what possibly can be running. Once we've found the first
1585 * that is alive we stop, there might be some that acquired an xid earlier
1586 * but started writing later, but it's unlikely and they will cleaned up
1587 * in a later call to ReorderBufferAbortOld().
1589 dlist_foreach_modify(it, &rb->toplevel_by_lsn)
1591 ReorderBufferTXN *txn;
1593 txn = dlist_container(ReorderBufferTXN, node, it.cur);
1595 if (TransactionIdPrecedes(txn->xid, oldestRunningXid))
1597 elog(DEBUG1, "aborting old transaction %u", txn->xid);
1599 /* remove potential on-disk data, and deallocate this tx */
1600 ReorderBufferCleanupTXN(rb, txn);
1608 * Forget the contents of a transaction if we aren't interested in it's
1609 * contents. Needs to be first called for subtransactions and then for the
1612 * This is significantly different to ReorderBufferAbort() because
1613 * transactions that have committed need to be treated differenly from aborted
1614 * ones since they may have modified the catalog.
1616 * Note that this is only allowed to be called in the moment a transaction
1617 * commit has just been read, not earlier; otherwise later records referring
1618 * to this xid might re-create the transaction incompletely.
1621 ReorderBufferForget(ReorderBuffer *rb, TransactionId xid, XLogRecPtr lsn)
1623 ReorderBufferTXN *txn;
1625 txn = ReorderBufferTXNByXid(rb, xid, false, NULL, InvalidXLogRecPtr,
1628 /* unknown, nothing to forget */
1633 txn->final_lsn = lsn;
1636 * Proccess cache invalidation messages if there are any. Even if we're
1637 * not interested in the transaction's contents, it could have manipulated
1638 * the catalog and we need to update the caches according to that.
1640 if (txn->base_snapshot != NULL && txn->ninvalidations > 0)
1642 bool use_subtxn = IsTransactionOrTransactionBlock();
1645 BeginInternalSubTransaction("replay");
1648 * Force invalidations to happen outside of a valid transaction - that
1649 * way entries will just be marked as invalid without accessing the
1650 * catalog. That's advantageous because we don't need to setup the
1651 * full state necessary for catalog access.
1654 AbortCurrentTransaction();
1656 ReorderBufferExecuteInvalidations(rb, txn);
1659 RollbackAndReleaseCurrentSubTransaction();
1662 Assert(txn->ninvalidations == 0);
1664 /* remove potential on-disk data, and deallocate */
1665 ReorderBufferCleanupTXN(rb, txn);
1670 * Check whether a transaction is already known in this module.xs
1673 ReorderBufferIsXidKnown(ReorderBuffer *rb, TransactionId xid)
1675 ReorderBufferTXN *txn;
1677 txn = ReorderBufferTXNByXid(rb, xid, false, NULL, InvalidXLogRecPtr,
1683 * Add a new snapshot to this transaction that may only used after lsn 'lsn'
1684 * because the previous snapshot doesn't describe the catalog correctly for
1688 ReorderBufferAddSnapshot(ReorderBuffer *rb, TransactionId xid,
1689 XLogRecPtr lsn, Snapshot snap)
1691 ReorderBufferChange *change = ReorderBufferGetChange(rb);
1693 change->data.snapshot = snap;
1694 change->action = REORDER_BUFFER_CHANGE_INTERNAL_SNAPSHOT;
1696 ReorderBufferQueueChange(rb, xid, lsn, change);
1700 * Setup the base snapshot of a transaction. The base snapshot is the snapshot
1701 * that is used to decode all changes until either this transaction modifies
1702 * the catalog or another catalog modifying transaction commits.
1704 * Needs to be called before any changes are added with
1705 * ReorderBufferQueueChange().
1708 ReorderBufferSetBaseSnapshot(ReorderBuffer *rb, TransactionId xid,
1709 XLogRecPtr lsn, Snapshot snap)
1711 ReorderBufferTXN *txn;
1714 txn = ReorderBufferTXNByXid(rb, xid, true, &is_new, lsn, true);
1715 Assert(txn->base_snapshot == NULL);
1716 Assert(snap != NULL);
1718 txn->base_snapshot = snap;
1719 txn->base_snapshot_lsn = lsn;
1723 * Access the catalog with this CommandId at this point in the changestream.
1725 * May only be called for command ids > 1
1728 ReorderBufferAddNewCommandId(ReorderBuffer *rb, TransactionId xid,
1729 XLogRecPtr lsn, CommandId cid)
1731 ReorderBufferChange *change = ReorderBufferGetChange(rb);
1733 change->data.command_id = cid;
1734 change->action = REORDER_BUFFER_CHANGE_INTERNAL_COMMAND_ID;
1736 ReorderBufferQueueChange(rb, xid, lsn, change);
1741 * Add new (relfilenode, tid) -> (cmin, cmax) mappings.
1744 ReorderBufferAddNewTupleCids(ReorderBuffer *rb, TransactionId xid,
1745 XLogRecPtr lsn, RelFileNode node,
1746 ItemPointerData tid, CommandId cmin,
1747 CommandId cmax, CommandId combocid)
1749 ReorderBufferChange *change = ReorderBufferGetChange(rb);
1750 ReorderBufferTXN *txn;
1752 txn = ReorderBufferTXNByXid(rb, xid, true, NULL, lsn, true);
1754 change->data.tuplecid.node = node;
1755 change->data.tuplecid.tid = tid;
1756 change->data.tuplecid.cmin = cmin;
1757 change->data.tuplecid.cmax = cmax;
1758 change->data.tuplecid.combocid = combocid;
1760 change->action = REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID;
1762 dlist_push_tail(&txn->tuplecids, &change->node);
1767 * Setup the invalidation of the toplevel transaction.
1769 * This needs to be done before ReorderBufferCommit is called!
1772 ReorderBufferAddInvalidations(ReorderBuffer *rb, TransactionId xid,
1773 XLogRecPtr lsn, Size nmsgs,
1774 SharedInvalidationMessage *msgs)
1776 ReorderBufferTXN *txn;
1778 txn = ReorderBufferTXNByXid(rb, xid, true, NULL, lsn, true);
1780 if (txn->ninvalidations != 0)
1781 elog(ERROR, "only ever add one set of invalidations");
1785 txn->ninvalidations = nmsgs;
1786 txn->invalidations = (SharedInvalidationMessage *)
1787 MemoryContextAlloc(rb->context,
1788 sizeof(SharedInvalidationMessage) * nmsgs);
1789 memcpy(txn->invalidations, msgs,
1790 sizeof(SharedInvalidationMessage) * nmsgs);
1794 * Apply all invalidations we know. Possibly we only need parts at this point
1795 * in the changestream but we don't know which those are.
1798 ReorderBufferExecuteInvalidations(ReorderBuffer *rb, ReorderBufferTXN *txn)
1802 for (i = 0; i < txn->ninvalidations; i++)
1803 LocalExecuteInvalidationMessage(&txn->invalidations[i]);
1807 * Mark a transaction as containing catalog changes
1810 ReorderBufferXidSetCatalogChanges(ReorderBuffer *rb, TransactionId xid,
1813 ReorderBufferTXN *txn;
1815 txn = ReorderBufferTXNByXid(rb, xid, true, NULL, lsn, true);
1817 txn->has_catalog_changes = true;
1821 * Query whether a transaction is already *known* to contain catalog
1822 * changes. This can be wrong until directly before the commit!
1825 ReorderBufferXidHasCatalogChanges(ReorderBuffer *rb, TransactionId xid)
1827 ReorderBufferTXN *txn;
1829 txn = ReorderBufferTXNByXid(rb, xid, false, NULL, InvalidXLogRecPtr,
1834 return txn->has_catalog_changes;
1838 * Have we already added the first snapshot?
1841 ReorderBufferXidHasBaseSnapshot(ReorderBuffer *rb, TransactionId xid)
1843 ReorderBufferTXN *txn;
1845 txn = ReorderBufferTXNByXid(rb, xid, false, NULL, InvalidXLogRecPtr,
1848 /* transaction isn't known yet, ergo no snapshot */
1853 * TODO: It would be a nice improvement if we would check the toplevel
1854 * transaction in subtransactions, but we'd need to keep track of a bit
1857 return txn->base_snapshot != NULL;
1862 * ---------------------------------------
1863 * Disk serialization support
1864 * ---------------------------------------
1868 * Ensure the IO buffer is >= sz.
1871 ReorderBufferSerializeReserve(ReorderBuffer *rb, Size sz)
1873 if (!rb->outbufsize)
1875 rb->outbuf = MemoryContextAlloc(rb->context, sz);
1876 rb->outbufsize = sz;
1878 else if (rb->outbufsize < sz)
1880 rb->outbuf = repalloc(rb->outbuf, sz);
1881 rb->outbufsize = sz;
1886 * Check whether the transaction tx should spill its data to disk.
1889 ReorderBufferCheckSerializeTXN(ReorderBuffer *rb, ReorderBufferTXN *txn)
1892 * TODO: improve accounting so we cheaply can take subtransactions into
1895 if (txn->nentries_mem >= max_changes_in_memory)
1897 ReorderBufferSerializeTXN(rb, txn);
1898 Assert(txn->nentries_mem == 0);
1903 * Spill data of a large transaction (and its subtransactions) to disk.
1906 ReorderBufferSerializeTXN(ReorderBuffer *rb, ReorderBufferTXN *txn)
1908 dlist_iter subtxn_i;
1909 dlist_mutable_iter change_i;
1911 XLogSegNo curOpenSegNo = 0;
1913 char path[MAXPGPATH];
1915 elog(DEBUG2, "spill %u changes in XID %u to disk",
1916 (uint32) txn->nentries_mem, txn->xid);
1918 /* do the same to all child TXs */
1919 dlist_foreach(subtxn_i, &txn->subtxns)
1921 ReorderBufferTXN *subtxn;
1923 subtxn = dlist_container(ReorderBufferTXN, node, subtxn_i.cur);
1924 ReorderBufferSerializeTXN(rb, subtxn);
1927 /* serialize changestream */
1928 dlist_foreach_modify(change_i, &txn->changes)
1930 ReorderBufferChange *change;
1932 change = dlist_container(ReorderBufferChange, node, change_i.cur);
1935 * store in segment in which it belongs by start lsn, don't split over
1936 * multiple segments tho
1938 if (fd == -1 || XLByteInSeg(change->lsn, curOpenSegNo))
1943 CloseTransientFile(fd);
1945 XLByteToSeg(change->lsn, curOpenSegNo);
1946 XLogSegNoOffsetToRecPtr(curOpenSegNo, 0, recptr);
1949 * No need to care about TLIs here, only used during a single run,
1950 * so each LSN only maps to a specific WAL record.
1952 sprintf(path, "pg_replslot/%s/xid-%u-lsn-%X-%X.snap",
1953 NameStr(MyReplicationSlot->data.name), txn->xid,
1954 (uint32) (recptr >> 32), (uint32) recptr);
1956 /* open segment, create it if necessary */
1957 fd = OpenTransientFile(path,
1958 O_CREAT | O_WRONLY | O_APPEND | PG_BINARY,
1963 (errcode_for_file_access(),
1964 errmsg("could not open file \"%s\": %m",
1968 ReorderBufferSerializeChange(rb, txn, fd, change);
1969 dlist_delete(&change->node);
1970 ReorderBufferReturnChange(rb, change);
1975 Assert(spilled == txn->nentries_mem);
1976 Assert(dlist_is_empty(&txn->changes));
1977 txn->nentries_mem = 0;
1980 CloseTransientFile(fd);
1984 * Serialize individual change to disk.
1987 ReorderBufferSerializeChange(ReorderBuffer *rb, ReorderBufferTXN *txn,
1988 int fd, ReorderBufferChange *change)
1990 ReorderBufferDiskChange *ondisk;
1991 Size sz = sizeof(ReorderBufferDiskChange);
1993 ReorderBufferSerializeReserve(rb, sz);
1995 ondisk = (ReorderBufferDiskChange *) rb->outbuf;
1996 memcpy(&ondisk->change, change, sizeof(ReorderBufferChange));
1998 switch (change->action)
2000 case REORDER_BUFFER_CHANGE_INSERT:
2002 case REORDER_BUFFER_CHANGE_UPDATE:
2004 case REORDER_BUFFER_CHANGE_DELETE:
2007 ReorderBufferTupleBuf *oldtup,
2012 oldtup = change->data.tp.oldtuple;
2013 newtup = change->data.tp.newtuple;
2016 oldlen = offsetof(ReorderBufferTupleBuf, data)
2017 +oldtup->tuple.t_len
2018 - offsetof(HeapTupleHeaderData, t_bits);
2021 newlen = offsetof(ReorderBufferTupleBuf, data)
2022 +newtup->tuple.t_len
2023 - offsetof(HeapTupleHeaderData, t_bits);
2028 /* make sure we have enough space */
2029 ReorderBufferSerializeReserve(rb, sz);
2031 data = ((char *) rb->outbuf) + sizeof(ReorderBufferDiskChange);
2032 /* might have been reallocated above */
2033 ondisk = (ReorderBufferDiskChange *) rb->outbuf;
2037 memcpy(data, oldtup, oldlen);
2043 memcpy(data, newtup, newlen);
2048 case REORDER_BUFFER_CHANGE_INTERNAL_SNAPSHOT:
2053 snap = change->data.snapshot;
2055 sz += sizeof(SnapshotData) +
2056 sizeof(TransactionId) * snap->xcnt +
2057 sizeof(TransactionId) * snap->subxcnt
2060 /* make sure we have enough space */
2061 ReorderBufferSerializeReserve(rb, sz);
2062 data = ((char *) rb->outbuf) + sizeof(ReorderBufferDiskChange);
2063 /* might have been reallocated above */
2064 ondisk = (ReorderBufferDiskChange *) rb->outbuf;
2066 memcpy(data, snap, sizeof(SnapshotData));
2067 data += sizeof(SnapshotData);
2071 memcpy(data, snap->xip,
2072 sizeof(TransactionId) * snap->xcnt);
2073 data += sizeof(TransactionId) * snap->xcnt;
2078 memcpy(data, snap->subxip,
2079 sizeof(TransactionId) * snap->subxcnt);
2080 data += sizeof(TransactionId) * snap->subxcnt;
2084 case REORDER_BUFFER_CHANGE_INTERNAL_COMMAND_ID:
2085 /* ReorderBufferChange contains everything important */
2087 case REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID:
2088 /* ReorderBufferChange contains everything important */
2094 if (write(fd, rb->outbuf, ondisk->size) != ondisk->size)
2096 CloseTransientFile(fd);
2098 (errcode_for_file_access(),
2099 errmsg("could not write to data file for XID %u: %m",
2103 Assert(ondisk->change.action == change->action);
2107 * Restore a number of changes spilled to disk back into memory.
2110 ReorderBufferRestoreChanges(ReorderBuffer *rb, ReorderBufferTXN *txn,
2111 int *fd, XLogSegNo *segno)
2114 XLogSegNo last_segno;
2115 dlist_mutable_iter cleanup_iter;
2117 Assert(txn->first_lsn != InvalidXLogRecPtr);
2118 Assert(txn->final_lsn != InvalidXLogRecPtr);
2120 /* free current entries, so we have memory for more */
2121 dlist_foreach_modify(cleanup_iter, &txn->changes)
2123 ReorderBufferChange *cleanup =
2124 dlist_container(ReorderBufferChange, node, cleanup_iter.cur);
2126 dlist_delete(&cleanup->node);
2127 ReorderBufferReturnChange(rb, cleanup);
2129 txn->nentries_mem = 0;
2130 Assert(dlist_is_empty(&txn->changes));
2132 XLByteToSeg(txn->final_lsn, last_segno);
2134 while (restored < max_changes_in_memory && *segno <= last_segno)
2137 ReorderBufferDiskChange *ondisk;
2142 char path[MAXPGPATH];
2147 XLByteToSeg(txn->first_lsn, *segno);
2150 Assert(*segno != 0 || dlist_is_empty(&txn->changes));
2151 XLogSegNoOffsetToRecPtr(*segno, 0, recptr);
2154 * No need to care about TLIs here, only used during a single run,
2155 * so each LSN only maps to a specific WAL record.
2157 sprintf(path, "pg_replslot/%s/xid-%u-lsn-%X-%X.snap",
2158 NameStr(MyReplicationSlot->data.name), txn->xid,
2159 (uint32) (recptr >> 32), (uint32) recptr);
2161 *fd = OpenTransientFile(path, O_RDONLY | PG_BINARY, 0);
2162 if (*fd < 0 && errno == ENOENT)
2170 (errcode_for_file_access(),
2171 errmsg("could not open file \"%s\": %m",
2177 * Read the statically sized part of a change which has information
2178 * about the total size. If we couldn't read a record, we're at the
2181 ReorderBufferSerializeReserve(rb, sizeof(ReorderBufferDiskChange));
2182 readBytes = read(*fd, rb->outbuf, sizeof(ReorderBufferDiskChange));
2187 CloseTransientFile(*fd);
2192 else if (readBytes < 0)
2194 (errcode_for_file_access(),
2195 errmsg("could not read from reorderbuffer spill file: %m")));
2196 else if (readBytes != sizeof(ReorderBufferDiskChange))
2198 (errcode_for_file_access(),
2199 errmsg("could not read from reorderbuffer spill file: read %d instead of %u bytes",
2201 (uint32) sizeof(ReorderBufferDiskChange))));
2203 ondisk = (ReorderBufferDiskChange *) rb->outbuf;
2205 ReorderBufferSerializeReserve(rb,
2206 sizeof(ReorderBufferDiskChange) + ondisk->size);
2207 ondisk = (ReorderBufferDiskChange *) rb->outbuf;
2209 readBytes = read(*fd, rb->outbuf + sizeof(ReorderBufferDiskChange),
2210 ondisk->size - sizeof(ReorderBufferDiskChange));
2214 (errcode_for_file_access(),
2215 errmsg("could not read from reorderbuffer spill file: %m")));
2216 else if (readBytes != ondisk->size - sizeof(ReorderBufferDiskChange))
2218 (errcode_for_file_access(),
2219 errmsg("could not read from reorderbuffer spill file: read %d instead of %u bytes",
2221 (uint32) (ondisk->size - sizeof(ReorderBufferDiskChange)))));
2224 * ok, read a full change from disk, now restore it into proper
2227 ReorderBufferRestoreChange(rb, txn, rb->outbuf);
2235 * Convert change from its on-disk format to in-memory format and queue it onto
2236 * the TXN's ->changes list.
2239 ReorderBufferRestoreChange(ReorderBuffer *rb, ReorderBufferTXN *txn,
2242 ReorderBufferDiskChange *ondisk;
2243 ReorderBufferChange *change;
2245 ondisk = (ReorderBufferDiskChange *) data;
2247 change = ReorderBufferGetChange(rb);
2249 /* copy static part */
2250 memcpy(change, &ondisk->change, sizeof(ReorderBufferChange));
2252 data += sizeof(ReorderBufferDiskChange);
2254 /* restore individual stuff */
2255 switch (change->action)
2257 case REORDER_BUFFER_CHANGE_INSERT:
2259 case REORDER_BUFFER_CHANGE_UPDATE:
2261 case REORDER_BUFFER_CHANGE_DELETE:
2262 if (change->data.tp.newtuple)
2264 Size len = offsetof(ReorderBufferTupleBuf, data)
2265 +((ReorderBufferTupleBuf *) data)->tuple.t_len
2266 - offsetof(HeapTupleHeaderData, t_bits);
2268 change->data.tp.newtuple = ReorderBufferGetTupleBuf(rb);
2269 memcpy(change->data.tp.newtuple, data, len);
2270 change->data.tp.newtuple->tuple.t_data =
2271 &change->data.tp.newtuple->header;
2275 if (change->data.tp.oldtuple)
2277 Size len = offsetof(ReorderBufferTupleBuf, data)
2278 +((ReorderBufferTupleBuf *) data)->tuple.t_len
2279 - offsetof(HeapTupleHeaderData, t_bits);
2281 change->data.tp.oldtuple = ReorderBufferGetTupleBuf(rb);
2282 memcpy(change->data.tp.oldtuple, data, len);
2283 change->data.tp.oldtuple->tuple.t_data =
2284 &change->data.tp.oldtuple->header;
2288 case REORDER_BUFFER_CHANGE_INTERNAL_SNAPSHOT:
2294 oldsnap = (Snapshot) data;
2296 size = sizeof(SnapshotData) +
2297 sizeof(TransactionId) * oldsnap->xcnt +
2298 sizeof(TransactionId) * (oldsnap->subxcnt + 0);
2300 change->data.snapshot = MemoryContextAllocZero(rb->context, size);
2302 newsnap = change->data.snapshot;
2304 memcpy(newsnap, data, size);
2305 newsnap->xip = (TransactionId *)
2306 (((char *) newsnap) + sizeof(SnapshotData));
2307 newsnap->subxip = newsnap->xip + newsnap->xcnt;
2308 newsnap->copied = true;
2311 /* the base struct contains all the data, easy peasy */
2312 case REORDER_BUFFER_CHANGE_INTERNAL_COMMAND_ID:
2313 case REORDER_BUFFER_CHANGE_INTERNAL_TUPLECID:
2317 dlist_push_tail(&txn->changes, &change->node);
2318 txn->nentries_mem++;
2322 * Remove all on-disk stored for the passed in transaction.
2325 ReorderBufferRestoreCleanup(ReorderBuffer *rb, ReorderBufferTXN *txn)
2331 Assert(txn->first_lsn != InvalidXLogRecPtr);
2332 Assert(txn->final_lsn != InvalidXLogRecPtr);
2334 XLByteToSeg(txn->first_lsn, first);
2335 XLByteToSeg(txn->final_lsn, last);
2337 /* iterate over all possible filenames, and delete them */
2338 for (cur = first; cur <= last; cur++)
2340 char path[MAXPGPATH];
2343 XLogSegNoOffsetToRecPtr(cur, 0, recptr);
2345 sprintf(path, "pg_replslot/%s/xid-%u-lsn-%X-%X.snap",
2346 NameStr(MyReplicationSlot->data.name), txn->xid,
2347 (uint32) (recptr >> 32), (uint32) recptr);
2348 if (unlink(path) != 0 && errno != ENOENT)
2350 (errcode_for_file_access(),
2351 errmsg("could not remove file \"%s\": %m", path)));
2356 * Delete all data spilled to disk after we've restarted/crashed. It will be
2357 * recreated when the respective slots are reused.
2360 StartupReorderBuffer(void)
2363 struct dirent *logical_de;
2366 struct dirent *spill_de;
2368 logical_dir = AllocateDir("pg_replslot");
2369 while ((logical_de = ReadDir(logical_dir, "pg_replslot")) != NULL)
2371 struct stat statbuf;
2372 char path[MAXPGPATH];
2374 if (strcmp(logical_de->d_name, ".") == 0 ||
2375 strcmp(logical_de->d_name, "..") == 0)
2378 /* if it cannot be a slot, skip the directory */
2379 if (!ReplicationSlotValidateName(logical_de->d_name, DEBUG2))
2383 * ok, has to be a surviving logical slot, iterate and delete
2384 * everythign starting with xid-*
2386 sprintf(path, "pg_replslot/%s", logical_de->d_name);
2388 /* we're only creating directories here, skip if it's not our's */
2389 if (lstat(path, &statbuf) == 0 && !S_ISDIR(statbuf.st_mode))
2392 spill_dir = AllocateDir(path);
2393 while ((spill_de = ReadDir(spill_dir, path)) != NULL)
2395 if (strcmp(spill_de->d_name, ".") == 0 ||
2396 strcmp(spill_de->d_name, "..") == 0)
2399 /* only look at names that can be ours */
2400 if (strncmp(spill_de->d_name, "xid", 3) == 0)
2402 sprintf(path, "pg_replslot/%s/%s", logical_de->d_name,
2405 if (unlink(path) != 0)
2407 (errcode_for_file_access(),
2408 errmsg("could not remove file \"%s\": %m",
2414 FreeDir(logical_dir);
2417 /* ---------------------------------------
2418 * toast reassembly support
2419 * ---------------------------------------
2423 * Initialize per tuple toast reconstruction support.
2426 ReorderBufferToastInitHash(ReorderBuffer *rb, ReorderBufferTXN *txn)
2430 Assert(txn->toast_hash == NULL);
2432 memset(&hash_ctl, 0, sizeof(hash_ctl));
2433 hash_ctl.keysize = sizeof(Oid);
2434 hash_ctl.entrysize = sizeof(ReorderBufferToastEnt);
2435 hash_ctl.hcxt = rb->context;
2436 txn->toast_hash = hash_create("ReorderBufferToastHash", 5, &hash_ctl,
2437 HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
2441 * Per toast-chunk handling for toast reconstruction
2443 * Appends a toast chunk so we can reconstruct it when the tuple "owning" the
2444 * toasted Datum comes along.
2447 ReorderBufferToastAppendChunk(ReorderBuffer *rb, ReorderBufferTXN *txn,
2448 Relation relation, ReorderBufferChange *change)
2450 ReorderBufferToastEnt *ent;
2451 ReorderBufferTupleBuf *newtup;
2456 TupleDesc desc = RelationGetDescr(relation);
2460 if (txn->toast_hash == NULL)
2461 ReorderBufferToastInitHash(rb, txn);
2463 Assert(IsToastRelation(relation));
2465 newtup = change->data.tp.newtuple;
2466 chunk_id = DatumGetObjectId(fastgetattr(&newtup->tuple, 1, desc, &isnull));
2468 chunk_seq = DatumGetInt32(fastgetattr(&newtup->tuple, 2, desc, &isnull));
2471 ent = (ReorderBufferToastEnt *)
2472 hash_search(txn->toast_hash,
2479 Assert(ent->chunk_id == chunk_id);
2480 ent->num_chunks = 0;
2481 ent->last_chunk_seq = 0;
2483 ent->reconstructed = NULL;
2484 dlist_init(&ent->chunks);
2487 elog(ERROR, "got sequence entry %d for toast chunk %u instead of seq 0",
2488 chunk_seq, chunk_id);
2490 else if (found && chunk_seq != ent->last_chunk_seq + 1)
2491 elog(ERROR, "got sequence entry %d for toast chunk %u instead of seq %d",
2492 chunk_seq, chunk_id, ent->last_chunk_seq + 1);
2494 chunk = DatumGetPointer(fastgetattr(&newtup->tuple, 3, desc, &isnull));
2497 /* calculate size so we can allocate the right size at once later */
2498 if (!VARATT_IS_EXTENDED(chunk))
2499 chunksize = VARSIZE(chunk) - VARHDRSZ;
2500 else if (VARATT_IS_SHORT(chunk))
2501 /* could happen due to heap_form_tuple doing its thing */
2502 chunksize = VARSIZE_SHORT(chunk) - VARHDRSZ_SHORT;
2504 elog(ERROR, "unexpected type of toast chunk");
2506 ent->size += chunksize;
2507 ent->last_chunk_seq = chunk_seq;
2509 dlist_push_tail(&ent->chunks, &change->node);
2513 * Rejigger change->newtuple to point to in-memory toast tuples instead to
2514 * on-disk toast tuples that may not longer exist (think DROP TABLE or VACUUM).
2516 * We cannot replace unchanged toast tuples though, so those will still point
2517 * to on-disk toast data.
2520 ReorderBufferToastReplace(ReorderBuffer *rb, ReorderBufferTXN *txn,
2521 Relation relation, ReorderBufferChange *change)
2530 TupleDesc toast_desc;
2531 MemoryContext oldcontext;
2532 ReorderBufferTupleBuf *newtup;
2534 /* no toast tuples changed */
2535 if (txn->toast_hash == NULL)
2538 oldcontext = MemoryContextSwitchTo(rb->context);
2540 /* we should only have toast tuples in an INSERT or UPDATE */
2541 Assert(change->data.tp.newtuple);
2543 desc = RelationGetDescr(relation);
2545 toast_rel = RelationIdGetRelation(relation->rd_rel->reltoastrelid);
2546 toast_desc = RelationGetDescr(toast_rel);
2548 /* should we allocate from stack instead? */
2549 attrs = palloc0(sizeof(Datum) * desc->natts);
2550 isnull = palloc0(sizeof(bool) * desc->natts);
2551 free = palloc0(sizeof(bool) * desc->natts);
2553 newtup = change->data.tp.newtuple;
2555 heap_deform_tuple(&newtup->tuple, desc, attrs, isnull);
2557 for (natt = 0; natt < desc->natts; natt++)
2559 Form_pg_attribute attr = desc->attrs[natt];
2560 ReorderBufferToastEnt *ent;
2561 struct varlena *varlena;
2563 /* va_rawsize is the size of the original datum -- including header */
2564 struct varatt_external toast_pointer;
2565 struct varatt_indirect redirect_pointer;
2566 struct varlena *new_datum = NULL;
2567 struct varlena *reconstructed;
2571 /* system columns aren't toasted */
2572 if (attr->attnum < 0)
2575 if (attr->attisdropped)
2578 /* not a varlena datatype */
2579 if (attr->attlen != -1)
2586 /* ok, we know we have a toast datum */
2587 varlena = (struct varlena *) DatumGetPointer(attrs[natt]);
2589 /* no need to do anything if the tuple isn't external */
2590 if (!VARATT_IS_EXTERNAL(varlena))
2593 VARATT_EXTERNAL_GET_POINTER(toast_pointer, varlena);
2596 * Check whether the toast tuple changed, replace if so.
2598 ent = (ReorderBufferToastEnt *)
2599 hash_search(txn->toast_hash,
2600 (void *) &toast_pointer.va_valueid,
2607 (struct varlena *) palloc0(INDIRECT_POINTER_SIZE);
2611 reconstructed = palloc0(toast_pointer.va_rawsize);
2613 ent->reconstructed = reconstructed;
2615 /* stitch toast tuple back together from its parts */
2616 dlist_foreach(it, &ent->chunks)
2619 ReorderBufferChange *cchange;
2620 ReorderBufferTupleBuf *ctup;
2623 cchange = dlist_container(ReorderBufferChange, node, it.cur);
2624 ctup = cchange->data.tp.newtuple;
2625 chunk = DatumGetPointer(
2626 fastgetattr(&ctup->tuple, 3, toast_desc, &isnull));
2629 Assert(!VARATT_IS_EXTERNAL(chunk));
2630 Assert(!VARATT_IS_SHORT(chunk));
2632 memcpy(VARDATA(reconstructed) + data_done,
2634 VARSIZE(chunk) - VARHDRSZ);
2635 data_done += VARSIZE(chunk) - VARHDRSZ;
2637 Assert(data_done == toast_pointer.va_extsize);
2639 /* make sure its marked as compressed or not */
2640 if (VARATT_EXTERNAL_IS_COMPRESSED(toast_pointer))
2641 SET_VARSIZE_COMPRESSED(reconstructed, data_done + VARHDRSZ);
2643 SET_VARSIZE(reconstructed, data_done + VARHDRSZ);
2645 memset(&redirect_pointer, 0, sizeof(redirect_pointer));
2646 redirect_pointer.pointer = reconstructed;
2648 SET_VARTAG_EXTERNAL(new_datum, VARTAG_INDIRECT);
2649 memcpy(VARDATA_EXTERNAL(new_datum), &redirect_pointer,
2650 sizeof(redirect_pointer));
2652 attrs[natt] = PointerGetDatum(new_datum);
2656 * Build tuple in separate memory & copy tuple back into the tuplebuf
2657 * passed to the output plugin. We can't directly heap_fill_tuple() into
2658 * the tuplebuf because attrs[] will point back into the current content.
2660 tmphtup = heap_form_tuple(desc, attrs, isnull);
2661 Assert(newtup->tuple.t_len <= MaxHeapTupleSize);
2662 Assert(&newtup->header == newtup->tuple.t_data);
2664 memcpy(newtup->tuple.t_data, tmphtup->t_data, tmphtup->t_len);
2665 newtup->tuple.t_len = tmphtup->t_len;
2668 * free resources we won't further need, more persistent stuff will be
2669 * free'd in ReorderBufferToastReset().
2671 RelationClose(toast_rel);
2673 for (natt = 0; natt < desc->natts; natt++)
2676 pfree(DatumGetPointer(attrs[natt]));
2682 MemoryContextSwitchTo(oldcontext);
2686 * Free all resources allocated for toast reconstruction.
2689 ReorderBufferToastReset(ReorderBuffer *rb, ReorderBufferTXN *txn)
2691 HASH_SEQ_STATUS hstat;
2692 ReorderBufferToastEnt *ent;
2694 if (txn->toast_hash == NULL)
2697 /* sequentially walk over the hash and free everything */
2698 hash_seq_init(&hstat, txn->toast_hash);
2699 while ((ent = (ReorderBufferToastEnt *) hash_seq_search(&hstat)) != NULL)
2701 dlist_mutable_iter it;
2703 if (ent->reconstructed != NULL)
2704 pfree(ent->reconstructed);
2706 dlist_foreach_modify(it, &ent->chunks)
2708 ReorderBufferChange *change =
2709 dlist_container(ReorderBufferChange, node, it.cur);
2711 dlist_delete(&change->node);
2712 ReorderBufferReturnChange(rb, change);
2716 hash_destroy(txn->toast_hash);
2717 txn->toast_hash = NULL;
2721 /* ---------------------------------------
2722 * Visibility support for logical decoding
2725 * Lookup actual cmin/cmax values when using decoding snapshot. We can't
2726 * always rely on stored cmin/cmax values because of two scenarios:
2728 * * A tuple got changed multiple times during a single transaction and thus
2729 * has got a combocid. Combocid's are only valid for the duration of a
2730 * single transaction.
2731 * * A tuple with a cmin but no cmax (and thus no combocid) got
2732 * deleted/updated in another transaction than the one which created it
2733 * which we are looking at right now. As only one of cmin, cmax or combocid
2734 * is actually stored in the heap we don't have access to the value we
2737 * To resolve those problems we have a per-transaction hash of (cmin,
2738 * cmax) tuples keyed by (relfilenode, ctid) which contains the actual
2739 * (cmin, cmax) values. That also takes care of combocids by simply
2740 * not caring about them at all. As we have the real cmin/cmax values
2741 * combocids aren't interesting.
2743 * As we only care about catalog tuples here the overhead of this
2744 * hashtable should be acceptable.
2746 * Heap rewrites complicate this a bit, check rewriteheap.c for
2748 * -------------------------------------------------------------------------
2751 /* struct for qsort()ing mapping files by lsn somewhat efficiently */
2752 typedef struct RewriteMappingFile
2755 char fname[MAXPGPATH];
2756 } RewriteMappingFile;
2760 DisplayMapping(HTAB *tuplecid_data)
2762 HASH_SEQ_STATUS hstat;
2763 ReorderBufferTupleCidEnt *ent;
2765 hash_seq_init(&hstat, tuplecid_data);
2766 while ((ent = (ReorderBufferTupleCidEnt *) hash_seq_search(&hstat)) != NULL)
2768 elog(DEBUG3, "mapping: node: %u/%u/%u tid: %u/%u cmin: %u, cmax: %u",
2769 ent->key.relnode.dbNode,
2770 ent->key.relnode.spcNode,
2771 ent->key.relnode.relNode,
2772 BlockIdGetBlockNumber(&ent->key.tid.ip_blkid),
2773 ent->key.tid.ip_posid,
2782 * Apply a single mapping file to tuplecid_data.
2784 * The mapping file has to have been verified to be a) committed b) for our
2785 * transaction c) applied in LSN order.
2788 ApplyLogicalMappingFile(HTAB *tuplecid_data, Oid relid, const char *fname)
2790 char path[MAXPGPATH];
2793 LogicalRewriteMappingData map;
2795 sprintf(path, "pg_logical/mappings/%s", fname);
2796 fd = OpenTransientFile(path, O_RDONLY | PG_BINARY, 0);
2799 (errmsg("could not open file \"%s\": %m", path)));
2803 ReorderBufferTupleCidKey key;
2804 ReorderBufferTupleCidEnt *ent;
2805 ReorderBufferTupleCidEnt *new_ent;
2808 /* be careful about padding */
2809 memset(&key, 0, sizeof(ReorderBufferTupleCidKey));
2811 /* read all mappings till the end of the file */
2812 readBytes = read(fd, &map, sizeof(LogicalRewriteMappingData));
2816 (errcode_for_file_access(),
2817 errmsg("could not read file \"%s\": %m",
2819 else if (readBytes == 0) /* EOF */
2821 else if (readBytes != sizeof(LogicalRewriteMappingData))
2823 (errcode_for_file_access(),
2824 errmsg("could not read from file \"%s\": read %d instead of %d bytes",
2826 (int32) sizeof(LogicalRewriteMappingData))));
2828 key.relnode = map.old_node;
2829 ItemPointerCopy(&map.old_tid,
2833 ent = (ReorderBufferTupleCidEnt *)
2834 hash_search(tuplecid_data,
2839 /* no existing mapping, no need to update */
2843 key.relnode = map.new_node;
2844 ItemPointerCopy(&map.new_tid,
2847 new_ent = (ReorderBufferTupleCidEnt *)
2848 hash_search(tuplecid_data,
2856 * Make sure the existing mapping makes sense. We sometime update
2857 * old records that did not yet have a cmax (e.g. pg_class' own
2858 * entry while rewriting it) during rewrites, so allow that.
2860 Assert(ent->cmin == InvalidCommandId || ent->cmin == new_ent->cmin);
2861 Assert(ent->cmax == InvalidCommandId || ent->cmax == new_ent->cmax);
2865 /* update mapping */
2866 new_ent->cmin = ent->cmin;
2867 new_ent->cmax = ent->cmax;
2868 new_ent->combocid = ent->combocid;
2875 * Check whether the TransactionOId 'xid' is in the pre-sorted array 'xip'.
2878 TransactionIdInArray(TransactionId xid, TransactionId *xip, Size num)
2880 return bsearch(&xid, xip, num,
2881 sizeof(TransactionId), xidComparator) != NULL;
2885 * qsort() comparator for sorting RewriteMappingFiles in LSN order.
2888 file_sort_by_lsn(const void *a_p, const void *b_p)
2890 RewriteMappingFile *a = *(RewriteMappingFile **) a_p;
2891 RewriteMappingFile *b = *(RewriteMappingFile **) b_p;
2893 if (a->lsn < b->lsn)
2895 else if (a->lsn > b->lsn)
2901 * Apply any existing logical remapping files if there are any targeted at our
2902 * transaction for relid.
2905 UpdateLogicalMappings(HTAB *tuplecid_data, Oid relid, Snapshot snapshot)
2908 struct dirent *mapping_de;
2911 RewriteMappingFile **files_a;
2913 Oid dboid = IsSharedRelation(relid) ? InvalidOid : MyDatabaseId;
2915 mapping_dir = AllocateDir("pg_logical/mappings");
2916 while ((mapping_de = ReadDir(mapping_dir, "pg_logical/mappings")) != NULL)
2920 TransactionId f_mapped_xid;
2921 TransactionId f_create_xid;
2925 RewriteMappingFile *f;
2927 if (strcmp(mapping_de->d_name, ".") == 0 ||
2928 strcmp(mapping_de->d_name, "..") == 0)
2931 /* Ignore files that aren't ours */
2932 if (strncmp(mapping_de->d_name, "map-", 4) != 0)
2935 if (sscanf(mapping_de->d_name, LOGICAL_REWRITE_FORMAT,
2936 &f_dboid, &f_relid, &f_hi, &f_lo,
2937 &f_mapped_xid, &f_create_xid) != 6)
2938 elog(ERROR, "could not parse filename \"%s\"", mapping_de->d_name);
2940 f_lsn = ((uint64) f_hi) << 32 | f_lo;
2942 /* mapping for another database */
2943 if (f_dboid != dboid)
2946 /* mapping for another relation */
2947 if (f_relid != relid)
2950 /* did the creating transaction abort? */
2951 if (!TransactionIdDidCommit(f_create_xid))
2954 /* not for our transaction */
2955 if (!TransactionIdInArray(f_mapped_xid, snapshot->subxip, snapshot->subxcnt))
2958 /* ok, relevant, queue for apply */
2959 f = palloc(sizeof(RewriteMappingFile));
2961 strcpy(f->fname, mapping_de->d_name);
2962 files = lappend(files, f);
2964 FreeDir(mapping_dir);
2966 /* build array we can easily sort */
2967 files_a = palloc(list_length(files) * sizeof(RewriteMappingFile *));
2969 foreach(file, files)
2971 files_a[off++] = lfirst(file);
2974 /* sort files so we apply them in LSN order */
2975 qsort(files_a, list_length(files), sizeof(RewriteMappingFile *),
2978 for (off = 0; off < list_length(files); off++)
2980 RewriteMappingFile *f = files_a[off];
2982 elog(DEBUG1, "applying mapping: \"%s\" in %u", f->fname,
2983 snapshot->subxip[0]);
2984 ApplyLogicalMappingFile(tuplecid_data, relid, f->fname);
2990 * Lookup cmin/cmax of a tuple, during logical decoding where we can't rely on
2994 ResolveCminCmaxDuringDecoding(HTAB *tuplecid_data,
2996 HeapTuple htup, Buffer buffer,
2997 CommandId *cmin, CommandId *cmax)
2999 ReorderBufferTupleCidKey key;
3000 ReorderBufferTupleCidEnt *ent;
3002 BlockNumber blockno;
3003 bool updated_mapping = false;
3005 /* be careful about padding */
3006 memset(&key, 0, sizeof(key));
3008 Assert(!BufferIsLocal(buffer));
3011 * get relfilenode from the buffer, no convenient way to access it other
3014 BufferGetTag(buffer, &key.relnode, &forkno, &blockno);
3016 /* tuples can only be in the main fork */
3017 Assert(forkno == MAIN_FORKNUM);
3018 Assert(blockno == ItemPointerGetBlockNumber(&htup->t_self));
3020 ItemPointerCopy(&htup->t_self,
3024 ent = (ReorderBufferTupleCidEnt *)
3025 hash_search(tuplecid_data,
3031 * failed to find a mapping, check whether the table was rewritten and
3032 * apply mapping if so, but only do that once - there can be no new
3033 * mappings while we are in here since we have to hold a lock on the
3036 if (ent == NULL && !updated_mapping)
3038 UpdateLogicalMappings(tuplecid_data, htup->t_tableOid, snapshot);
3039 /* now check but don't update for a mapping again */
3040 updated_mapping = true;
3043 else if (ent == NULL)