Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * snapbuild.c
4 : *
5 : * Infrastructure for building historic catalog snapshots based on contents
6 : * of the WAL, for the purpose of decoding heapam.c style values in the
7 : * WAL.
8 : *
9 : * NOTES:
10 : *
11 : * We build snapshots which can *only* be used to read catalog contents and we
12 : * do so by reading and interpreting the WAL stream. The aim is to build a
13 : * snapshot that behaves the same as a freshly taken MVCC snapshot would have
14 : * at the time the XLogRecord was generated.
15 : *
16 : * To build the snapshots we reuse the infrastructure built for Hot
17 : * Standby. The in-memory snapshots we build look different than HS' because
18 : * we have different needs. To successfully decode data from the WAL we only
19 : * need to access catalog tables and (sys|rel|cat)cache, not the actual user
20 : * tables since the data we decode is wholly contained in the WAL
21 : * records. Also, our snapshots need to be different in comparison to normal
22 : * MVCC ones because in contrast to those we cannot fully rely on the clog and
23 : * pg_subtrans for information about committed transactions because they might
24 : * commit in the future from the POV of the WAL entry we're currently
25 : * decoding. This definition has the advantage that we only need to prevent
26 : * removal of catalog rows, while normal table's rows can still be
27 : * removed. This is achieved by using the replication slot mechanism.
28 : *
29 : * As the percentage of transactions modifying the catalog normally is fairly
30 : * small in comparisons to ones only manipulating user data, we keep track of
31 : * the committed catalog modifying ones inside [xmin, xmax) instead of keeping
32 : * track of all running transactions like it's done in a normal snapshot. Note
33 : * that we're generally only looking at transactions that have acquired an
34 : * xid. That is we keep a list of transactions between snapshot->(xmin, xmax)
35 : * that we consider committed, everything else is considered aborted/in
36 : * progress. That also allows us not to care about subtransactions before they
37 : * have committed which means this module, in contrast to HS, doesn't have to
38 : * care about suboverflowed subtransactions and similar.
39 : *
40 : * One complexity of doing this is that to e.g. handle mixed DDL/DML
41 : * transactions we need Snapshots that see intermediate versions of the
42 : * catalog in a transaction. During normal operation this is achieved by using
43 : * CommandIds/cmin/cmax. The problem with that however is that for space
44 : * efficiency reasons, the cmin and cmax are not included in WAL records. We
45 : * cannot read the cmin/cmax from the tuple itself, either, because it is
46 : * reset on crash recovery. Even if we could, we could not decode combocids
47 : * which are only tracked in the original backend's memory. To work around
48 : * that, heapam writes an extra WAL record (XLOG_HEAP2_NEW_CID) every time a
49 : * catalog row is modified, which includes the cmin and cmax of the
50 : * tuple. During decoding, we insert the ctid->(cmin,cmax) mappings into the
51 : * reorder buffer, and use them at visibility checks instead of the cmin/cmax
52 : * on the tuple itself. Check the reorderbuffer.c's comment above
53 : * ResolveCminCmaxDuringDecoding() for details.
54 : *
55 : * To facilitate all this we need our own visibility routine, as the normal
56 : * ones are optimized for different usecases.
57 : *
58 : * To replace the normal catalog snapshots with decoding ones use the
59 : * SetupHistoricSnapshot() and TeardownHistoricSnapshot() functions.
60 : *
61 : *
62 : *
63 : * The snapbuild machinery is starting up in several stages, as illustrated
64 : * by the following graph describing the SnapBuild->state transitions:
65 : *
66 : * +-------------------------+
67 : * +----| START |-------------+
68 : * | +-------------------------+ |
69 : * | | |
70 : * | | |
71 : * | running_xacts #1 |
72 : * | | |
73 : * | | |
74 : * | v |
75 : * | +-------------------------+ v
76 : * | | BUILDING_SNAPSHOT |------------>|
77 : * | +-------------------------+ |
78 : * | | |
79 : * | | |
80 : * | running_xacts #2, xacts from #1 finished |
81 : * | | |
82 : * | | |
83 : * | v |
84 : * | +-------------------------+ v
85 : * | | FULL_SNAPSHOT |------------>|
86 : * | +-------------------------+ |
87 : * | | |
88 : * running_xacts | saved snapshot
89 : * with zero xacts | at running_xacts's lsn
90 : * | | |
91 : * | running_xacts with xacts from #2 finished |
92 : * | | |
93 : * | v |
94 : * | +-------------------------+ |
95 : * +--->|SNAPBUILD_CONSISTENT |<------------+
96 : * +-------------------------+
97 : *
98 : * Initially the machinery is in the START stage. When an xl_running_xacts
99 : * record is read that is sufficiently new (above the safe xmin horizon),
100 : * there's a state transition. If there were no running xacts when the
101 : * xl_running_xacts record was generated, we'll directly go into CONSISTENT
102 : * state, otherwise we'll switch to the BUILDING_SNAPSHOT state. Having a full
103 : * snapshot means that all transactions that start henceforth can be decoded
104 : * in their entirety, but transactions that started previously can't. In
105 : * FULL_SNAPSHOT we'll switch into CONSISTENT once all those previously
106 : * running transactions have committed or aborted.
107 : *
108 : * Only transactions that commit after CONSISTENT state has been reached will
109 : * be replayed, even though they might have started while still in
110 : * FULL_SNAPSHOT. That ensures that we'll reach a point where no previous
111 : * changes has been exported, but all the following ones will be. That point
112 : * is a convenient point to initialize replication from, which is why we
113 : * export a snapshot at that point, which *can* be used to read normal data.
114 : *
115 : * Copyright (c) 2012-2024, PostgreSQL Global Development Group
116 : *
117 : * IDENTIFICATION
118 : * src/backend/replication/logical/snapbuild.c
119 : *
120 : *-------------------------------------------------------------------------
121 : */
122 :
123 : #include "postgres.h"
124 :
125 : #include <sys/stat.h>
126 : #include <unistd.h>
127 :
128 : #include "access/heapam_xlog.h"
129 : #include "access/transam.h"
130 : #include "access/xact.h"
131 : #include "common/file_utils.h"
132 : #include "miscadmin.h"
133 : #include "pgstat.h"
134 : #include "replication/logical.h"
135 : #include "replication/reorderbuffer.h"
136 : #include "replication/snapbuild.h"
137 : #include "replication/snapbuild_internal.h"
138 : #include "storage/fd.h"
139 : #include "storage/lmgr.h"
140 : #include "storage/proc.h"
141 : #include "storage/procarray.h"
142 : #include "storage/standby.h"
143 : #include "utils/builtins.h"
144 : #include "utils/memutils.h"
145 : #include "utils/snapmgr.h"
146 : #include "utils/snapshot.h"
147 : /*
148 : * Starting a transaction -- which we need to do while exporting a snapshot --
149 : * removes knowledge about the previously used resowner, so we save it here.
150 : */
151 : static ResourceOwner SavedResourceOwnerDuringExport = NULL;
152 : static bool ExportInProgress = false;
153 :
154 : /* ->committed and ->catchange manipulation */
155 : static void SnapBuildPurgeOlderTxn(SnapBuild *builder);
156 :
157 : /* snapshot building/manipulation/distribution functions */
158 : static Snapshot SnapBuildBuildSnapshot(SnapBuild *builder);
159 :
160 : static void SnapBuildFreeSnapshot(Snapshot snap);
161 :
162 : static void SnapBuildSnapIncRefcount(Snapshot snap);
163 :
164 : static void SnapBuildDistributeNewCatalogSnapshot(SnapBuild *builder, XLogRecPtr lsn);
165 :
166 : static inline bool SnapBuildXidHasCatalogChanges(SnapBuild *builder, TransactionId xid,
167 : uint32 xinfo);
168 :
169 : /* xlog reading helper functions for SnapBuildProcessRunningXacts */
170 : static bool SnapBuildFindSnapshot(SnapBuild *builder, XLogRecPtr lsn, xl_running_xacts *running);
171 : static void SnapBuildWaitSnapshot(xl_running_xacts *running, TransactionId cutoff);
172 :
173 : /* serialization functions */
174 : static void SnapBuildSerialize(SnapBuild *builder, XLogRecPtr lsn);
175 : static bool SnapBuildRestore(SnapBuild *builder, XLogRecPtr lsn);
176 : static void SnapBuildRestoreContents(int fd, char *dest, Size size, const char *path);
177 :
178 : /*
179 : * Allocate a new snapshot builder.
180 : *
181 : * xmin_horizon is the xid >= which we can be sure no catalog rows have been
182 : * removed, start_lsn is the LSN >= we want to replay commits.
183 : */
184 : SnapBuild *
185 2006 : AllocateSnapshotBuilder(ReorderBuffer *reorder,
186 : TransactionId xmin_horizon,
187 : XLogRecPtr start_lsn,
188 : bool need_full_snapshot,
189 : bool in_slot_creation,
190 : XLogRecPtr two_phase_at)
191 : {
192 : MemoryContext context;
193 : MemoryContext oldcontext;
194 : SnapBuild *builder;
195 :
196 : /* allocate memory in own context, to have better accountability */
197 2006 : context = AllocSetContextCreate(CurrentMemoryContext,
198 : "snapshot builder context",
199 : ALLOCSET_DEFAULT_SIZES);
200 2006 : oldcontext = MemoryContextSwitchTo(context);
201 :
202 2006 : builder = palloc0(sizeof(SnapBuild));
203 :
204 2006 : builder->state = SNAPBUILD_START;
205 2006 : builder->context = context;
206 2006 : builder->reorder = reorder;
207 : /* Other struct members initialized by zeroing via palloc0 above */
208 :
209 2006 : builder->committed.xcnt = 0;
210 2006 : builder->committed.xcnt_space = 128; /* arbitrary number */
211 2006 : builder->committed.xip =
212 2006 : palloc0(builder->committed.xcnt_space * sizeof(TransactionId));
213 2006 : builder->committed.includes_all_transactions = true;
214 :
215 2006 : builder->catchange.xcnt = 0;
216 2006 : builder->catchange.xip = NULL;
217 :
218 2006 : builder->initial_xmin_horizon = xmin_horizon;
219 2006 : builder->start_decoding_at = start_lsn;
220 2006 : builder->in_slot_creation = in_slot_creation;
221 2006 : builder->building_full_snapshot = need_full_snapshot;
222 2006 : builder->two_phase_at = two_phase_at;
223 :
224 2006 : MemoryContextSwitchTo(oldcontext);
225 :
226 2006 : return builder;
227 : }
228 :
229 : /*
230 : * Free a snapshot builder.
231 : */
232 : void
233 1646 : FreeSnapshotBuilder(SnapBuild *builder)
234 : {
235 1646 : MemoryContext context = builder->context;
236 :
237 : /* free snapshot explicitly, that contains some error checking */
238 1646 : if (builder->snapshot != NULL)
239 : {
240 408 : SnapBuildSnapDecRefcount(builder->snapshot);
241 408 : builder->snapshot = NULL;
242 : }
243 :
244 : /* other resources are deallocated via memory context reset */
245 1646 : MemoryContextDelete(context);
246 1646 : }
247 :
248 : /*
249 : * Free an unreferenced snapshot that has previously been built by us.
250 : */
251 : static void
252 2368 : SnapBuildFreeSnapshot(Snapshot snap)
253 : {
254 : /* make sure we don't get passed an external snapshot */
255 : Assert(snap->snapshot_type == SNAPSHOT_HISTORIC_MVCC);
256 :
257 : /* make sure nobody modified our snapshot */
258 : Assert(snap->curcid == FirstCommandId);
259 : Assert(!snap->suboverflowed);
260 : Assert(!snap->takenDuringRecovery);
261 : Assert(snap->regd_count == 0);
262 :
263 : /* slightly more likely, so it's checked even without c-asserts */
264 2368 : if (snap->copied)
265 0 : elog(ERROR, "cannot free a copied snapshot");
266 :
267 2368 : if (snap->active_count)
268 0 : elog(ERROR, "cannot free an active snapshot");
269 :
270 2368 : pfree(snap);
271 2368 : }
272 :
273 : /*
274 : * In which state of snapshot building are we?
275 : */
276 : SnapBuildState
277 4681174 : SnapBuildCurrentState(SnapBuild *builder)
278 : {
279 4681174 : return builder->state;
280 : }
281 :
282 : /*
283 : * Return the LSN at which the two-phase decoding was first enabled.
284 : */
285 : XLogRecPtr
286 64 : SnapBuildGetTwoPhaseAt(SnapBuild *builder)
287 : {
288 64 : return builder->two_phase_at;
289 : }
290 :
291 : /*
292 : * Set the LSN at which two-phase decoding is enabled.
293 : */
294 : void
295 14 : SnapBuildSetTwoPhaseAt(SnapBuild *builder, XLogRecPtr ptr)
296 : {
297 14 : builder->two_phase_at = ptr;
298 14 : }
299 :
300 : /*
301 : * Should the contents of transaction ending at 'ptr' be decoded?
302 : */
303 : bool
304 1011528 : SnapBuildXactNeedsSkip(SnapBuild *builder, XLogRecPtr ptr)
305 : {
306 1011528 : return ptr < builder->start_decoding_at;
307 : }
308 :
309 : /*
310 : * Increase refcount of a snapshot.
311 : *
312 : * This is used when handing out a snapshot to some external resource or when
313 : * adding a Snapshot as builder->snapshot.
314 : */
315 : static void
316 10326 : SnapBuildSnapIncRefcount(Snapshot snap)
317 : {
318 10326 : snap->active_count++;
319 10326 : }
320 :
321 : /*
322 : * Decrease refcount of a snapshot and free if the refcount reaches zero.
323 : *
324 : * Externally visible, so that external resources that have been handed an
325 : * IncRef'ed Snapshot can adjust its refcount easily.
326 : */
327 : void
328 9932 : SnapBuildSnapDecRefcount(Snapshot snap)
329 : {
330 : /* make sure we don't get passed an external snapshot */
331 : Assert(snap->snapshot_type == SNAPSHOT_HISTORIC_MVCC);
332 :
333 : /* make sure nobody modified our snapshot */
334 : Assert(snap->curcid == FirstCommandId);
335 : Assert(!snap->suboverflowed);
336 : Assert(!snap->takenDuringRecovery);
337 :
338 : Assert(snap->regd_count == 0);
339 :
340 : Assert(snap->active_count > 0);
341 :
342 : /* slightly more likely, so it's checked even without casserts */
343 9932 : if (snap->copied)
344 0 : elog(ERROR, "cannot free a copied snapshot");
345 :
346 9932 : snap->active_count--;
347 9932 : if (snap->active_count == 0)
348 2368 : SnapBuildFreeSnapshot(snap);
349 9932 : }
350 :
351 : /*
352 : * Build a new snapshot, based on currently committed catalog-modifying
353 : * transactions.
354 : *
355 : * In-progress transactions with catalog access are *not* allowed to modify
356 : * these snapshots; they have to copy them and fill in appropriate ->curcid
357 : * and ->subxip/subxcnt values.
358 : */
359 : static Snapshot
360 3108 : SnapBuildBuildSnapshot(SnapBuild *builder)
361 : {
362 : Snapshot snapshot;
363 : Size ssize;
364 :
365 : Assert(builder->state >= SNAPBUILD_FULL_SNAPSHOT);
366 :
367 3108 : ssize = sizeof(SnapshotData)
368 3108 : + sizeof(TransactionId) * builder->committed.xcnt
369 3108 : + sizeof(TransactionId) * 1 /* toplevel xid */ ;
370 :
371 3108 : snapshot = MemoryContextAllocZero(builder->context, ssize);
372 :
373 3108 : snapshot->snapshot_type = SNAPSHOT_HISTORIC_MVCC;
374 :
375 : /*
376 : * We misuse the original meaning of SnapshotData's xip and subxip fields
377 : * to make the more fitting for our needs.
378 : *
379 : * In the 'xip' array we store transactions that have to be treated as
380 : * committed. Since we will only ever look at tuples from transactions
381 : * that have modified the catalog it's more efficient to store those few
382 : * that exist between xmin and xmax (frequently there are none).
383 : *
384 : * Snapshots that are used in transactions that have modified the catalog
385 : * also use the 'subxip' array to store their toplevel xid and all the
386 : * subtransaction xids so we can recognize when we need to treat rows as
387 : * visible that are not in xip but still need to be visible. Subxip only
388 : * gets filled when the transaction is copied into the context of a
389 : * catalog modifying transaction since we otherwise share a snapshot
390 : * between transactions. As long as a txn hasn't modified the catalog it
391 : * doesn't need to treat any uncommitted rows as visible, so there is no
392 : * need for those xids.
393 : *
394 : * Both arrays are qsort'ed so that we can use bsearch() on them.
395 : */
396 : Assert(TransactionIdIsNormal(builder->xmin));
397 : Assert(TransactionIdIsNormal(builder->xmax));
398 :
399 3108 : snapshot->xmin = builder->xmin;
400 3108 : snapshot->xmax = builder->xmax;
401 :
402 : /* store all transactions to be treated as committed by this snapshot */
403 3108 : snapshot->xip =
404 3108 : (TransactionId *) ((char *) snapshot + sizeof(SnapshotData));
405 3108 : snapshot->xcnt = builder->committed.xcnt;
406 3108 : memcpy(snapshot->xip,
407 3108 : builder->committed.xip,
408 3108 : builder->committed.xcnt * sizeof(TransactionId));
409 :
410 : /* sort so we can bsearch() */
411 3108 : qsort(snapshot->xip, snapshot->xcnt, sizeof(TransactionId), xidComparator);
412 :
413 : /*
414 : * Initially, subxip is empty, i.e. it's a snapshot to be used by
415 : * transactions that don't modify the catalog. Will be filled by
416 : * ReorderBufferCopySnap() if necessary.
417 : */
418 3108 : snapshot->subxcnt = 0;
419 3108 : snapshot->subxip = NULL;
420 :
421 3108 : snapshot->suboverflowed = false;
422 3108 : snapshot->takenDuringRecovery = false;
423 3108 : snapshot->copied = false;
424 3108 : snapshot->curcid = FirstCommandId;
425 3108 : snapshot->active_count = 0;
426 3108 : snapshot->regd_count = 0;
427 3108 : snapshot->snapXactCompletionCount = 0;
428 :
429 3108 : return snapshot;
430 : }
431 :
432 : /*
433 : * Build the initial slot snapshot and convert it to a normal snapshot that
434 : * is understood by HeapTupleSatisfiesMVCC.
435 : *
436 : * The snapshot will be usable directly in current transaction or exported
437 : * for loading in different transaction.
438 : */
439 : Snapshot
440 368 : SnapBuildInitialSnapshot(SnapBuild *builder)
441 : {
442 : Snapshot snap;
443 : TransactionId xid;
444 : TransactionId safeXid;
445 : TransactionId *newxip;
446 368 : int newxcnt = 0;
447 :
448 : Assert(XactIsoLevel == XACT_REPEATABLE_READ);
449 : Assert(builder->building_full_snapshot);
450 :
451 : /* don't allow older snapshots */
452 368 : InvalidateCatalogSnapshot(); /* about to overwrite MyProc->xmin */
453 368 : if (HaveRegisteredOrActiveSnapshot())
454 0 : elog(ERROR, "cannot build an initial slot snapshot when snapshots exist");
455 : Assert(!HistoricSnapshotActive());
456 :
457 368 : if (builder->state != SNAPBUILD_CONSISTENT)
458 0 : elog(ERROR, "cannot build an initial slot snapshot before reaching a consistent state");
459 :
460 368 : if (!builder->committed.includes_all_transactions)
461 0 : elog(ERROR, "cannot build an initial slot snapshot, not all transactions are monitored anymore");
462 :
463 : /* so we don't overwrite the existing value */
464 368 : if (TransactionIdIsValid(MyProc->xmin))
465 0 : elog(ERROR, "cannot build an initial slot snapshot when MyProc->xmin already is valid");
466 :
467 368 : snap = SnapBuildBuildSnapshot(builder);
468 :
469 : /*
470 : * We know that snap->xmin is alive, enforced by the logical xmin
471 : * mechanism. Due to that we can do this without locks, we're only
472 : * changing our own value.
473 : *
474 : * Building an initial snapshot is expensive and an unenforced xmin
475 : * horizon would have bad consequences, therefore always double-check that
476 : * the horizon is enforced.
477 : */
478 368 : LWLockAcquire(ProcArrayLock, LW_SHARED);
479 368 : safeXid = GetOldestSafeDecodingTransactionId(false);
480 368 : LWLockRelease(ProcArrayLock);
481 :
482 368 : if (TransactionIdFollows(safeXid, snap->xmin))
483 0 : elog(ERROR, "cannot build an initial slot snapshot as oldest safe xid %u follows snapshot's xmin %u",
484 : safeXid, snap->xmin);
485 :
486 368 : MyProc->xmin = snap->xmin;
487 :
488 : /* allocate in transaction context */
489 : newxip = (TransactionId *)
490 368 : palloc(sizeof(TransactionId) * GetMaxSnapshotXidCount());
491 :
492 : /*
493 : * snapbuild.c builds transactions in an "inverted" manner, which means it
494 : * stores committed transactions in ->xip, not ones in progress. Build a
495 : * classical snapshot by marking all non-committed transactions as
496 : * in-progress. This can be expensive.
497 : */
498 368 : for (xid = snap->xmin; NormalTransactionIdPrecedes(xid, snap->xmax);)
499 : {
500 : void *test;
501 :
502 : /*
503 : * Check whether transaction committed using the decoding snapshot
504 : * meaning of ->xip.
505 : */
506 0 : test = bsearch(&xid, snap->xip, snap->xcnt,
507 : sizeof(TransactionId), xidComparator);
508 :
509 0 : if (test == NULL)
510 : {
511 0 : if (newxcnt >= GetMaxSnapshotXidCount())
512 0 : ereport(ERROR,
513 : (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
514 : errmsg("initial slot snapshot too large")));
515 :
516 0 : newxip[newxcnt++] = xid;
517 : }
518 :
519 0 : TransactionIdAdvance(xid);
520 : }
521 :
522 : /* adjust remaining snapshot fields as needed */
523 368 : snap->snapshot_type = SNAPSHOT_MVCC;
524 368 : snap->xcnt = newxcnt;
525 368 : snap->xip = newxip;
526 :
527 368 : return snap;
528 : }
529 :
530 : /*
531 : * Export a snapshot so it can be set in another session with SET TRANSACTION
532 : * SNAPSHOT.
533 : *
534 : * For that we need to start a transaction in the current backend as the
535 : * importing side checks whether the source transaction is still open to make
536 : * sure the xmin horizon hasn't advanced since then.
537 : */
538 : const char *
539 0 : SnapBuildExportSnapshot(SnapBuild *builder)
540 : {
541 : Snapshot snap;
542 : char *snapname;
543 :
544 0 : if (IsTransactionOrTransactionBlock())
545 0 : elog(ERROR, "cannot export a snapshot from within a transaction");
546 :
547 0 : if (SavedResourceOwnerDuringExport)
548 0 : elog(ERROR, "can only export one snapshot at a time");
549 :
550 0 : SavedResourceOwnerDuringExport = CurrentResourceOwner;
551 0 : ExportInProgress = true;
552 :
553 0 : StartTransactionCommand();
554 :
555 : /* There doesn't seem to a nice API to set these */
556 0 : XactIsoLevel = XACT_REPEATABLE_READ;
557 0 : XactReadOnly = true;
558 :
559 0 : snap = SnapBuildInitialSnapshot(builder);
560 :
561 : /*
562 : * now that we've built a plain snapshot, make it active and use the
563 : * normal mechanisms for exporting it
564 : */
565 0 : snapname = ExportSnapshot(snap);
566 :
567 0 : ereport(LOG,
568 : (errmsg_plural("exported logical decoding snapshot: \"%s\" with %u transaction ID",
569 : "exported logical decoding snapshot: \"%s\" with %u transaction IDs",
570 : snap->xcnt,
571 : snapname, snap->xcnt)));
572 0 : return snapname;
573 : }
574 :
575 : /*
576 : * Ensure there is a snapshot and if not build one for current transaction.
577 : */
578 : Snapshot
579 16 : SnapBuildGetOrBuildSnapshot(SnapBuild *builder)
580 : {
581 : Assert(builder->state == SNAPBUILD_CONSISTENT);
582 :
583 : /* only build a new snapshot if we don't have a prebuilt one */
584 16 : if (builder->snapshot == NULL)
585 : {
586 2 : builder->snapshot = SnapBuildBuildSnapshot(builder);
587 : /* increase refcount for the snapshot builder */
588 2 : SnapBuildSnapIncRefcount(builder->snapshot);
589 : }
590 :
591 16 : return builder->snapshot;
592 : }
593 :
594 : /*
595 : * Reset a previously SnapBuildExportSnapshot()'ed snapshot if there is
596 : * any. Aborts the previously started transaction and resets the resource
597 : * owner back to its original value.
598 : */
599 : void
600 9676 : SnapBuildClearExportedSnapshot(void)
601 : {
602 : ResourceOwner tmpResOwner;
603 :
604 : /* nothing exported, that is the usual case */
605 9676 : if (!ExportInProgress)
606 9676 : return;
607 :
608 0 : if (!IsTransactionState())
609 0 : elog(ERROR, "clearing exported snapshot in wrong transaction state");
610 :
611 : /*
612 : * AbortCurrentTransaction() takes care of resetting the snapshot state,
613 : * so remember SavedResourceOwnerDuringExport.
614 : */
615 0 : tmpResOwner = SavedResourceOwnerDuringExport;
616 :
617 : /* make sure nothing could have ever happened */
618 0 : AbortCurrentTransaction();
619 :
620 0 : CurrentResourceOwner = tmpResOwner;
621 : }
622 :
623 : /*
624 : * Clear snapshot export state during transaction abort.
625 : */
626 : void
627 46892 : SnapBuildResetExportedSnapshotState(void)
628 : {
629 46892 : SavedResourceOwnerDuringExport = NULL;
630 46892 : ExportInProgress = false;
631 46892 : }
632 :
633 : /*
634 : * Handle the effects of a single heap change, appropriate to the current state
635 : * of the snapshot builder and returns whether changes made at (xid, lsn) can
636 : * be decoded.
637 : */
638 : bool
639 3218798 : SnapBuildProcessChange(SnapBuild *builder, TransactionId xid, XLogRecPtr lsn)
640 : {
641 : /*
642 : * We can't handle data in transactions if we haven't built a snapshot
643 : * yet, so don't store them.
644 : */
645 3218798 : if (builder->state < SNAPBUILD_FULL_SNAPSHOT)
646 0 : return false;
647 :
648 : /*
649 : * No point in keeping track of changes in transactions that we don't have
650 : * enough information about to decode. This means that they started before
651 : * we got into the SNAPBUILD_FULL_SNAPSHOT state.
652 : */
653 3218804 : if (builder->state < SNAPBUILD_CONSISTENT &&
654 6 : TransactionIdPrecedes(xid, builder->next_phase_at))
655 0 : return false;
656 :
657 : /*
658 : * If the reorderbuffer doesn't yet have a snapshot, add one now, it will
659 : * be needed to decode the change we're currently processing.
660 : */
661 3218798 : if (!ReorderBufferXidHasBaseSnapshot(builder->reorder, xid))
662 : {
663 : /* only build a new snapshot if we don't have a prebuilt one */
664 5500 : if (builder->snapshot == NULL)
665 : {
666 678 : builder->snapshot = SnapBuildBuildSnapshot(builder);
667 : /* increase refcount for the snapshot builder */
668 678 : SnapBuildSnapIncRefcount(builder->snapshot);
669 : }
670 :
671 : /*
672 : * Increase refcount for the transaction we're handing the snapshot
673 : * out to.
674 : */
675 5500 : SnapBuildSnapIncRefcount(builder->snapshot);
676 5500 : ReorderBufferSetBaseSnapshot(builder->reorder, xid, lsn,
677 : builder->snapshot);
678 : }
679 :
680 3218798 : return true;
681 : }
682 :
683 : /*
684 : * Do CommandId/combo CID handling after reading an xl_heap_new_cid record.
685 : * This implies that a transaction has done some form of write to system
686 : * catalogs.
687 : */
688 : void
689 46596 : SnapBuildProcessNewCid(SnapBuild *builder, TransactionId xid,
690 : XLogRecPtr lsn, xl_heap_new_cid *xlrec)
691 : {
692 : CommandId cid;
693 :
694 : /*
695 : * we only log new_cid's if a catalog tuple was modified, so mark the
696 : * transaction as containing catalog modifications
697 : */
698 46596 : ReorderBufferXidSetCatalogChanges(builder->reorder, xid, lsn);
699 :
700 46596 : ReorderBufferAddNewTupleCids(builder->reorder, xlrec->top_xid, lsn,
701 : xlrec->target_locator, xlrec->target_tid,
702 : xlrec->cmin, xlrec->cmax,
703 : xlrec->combocid);
704 :
705 : /* figure out new command id */
706 46596 : if (xlrec->cmin != InvalidCommandId &&
707 39308 : xlrec->cmax != InvalidCommandId)
708 6386 : cid = Max(xlrec->cmin, xlrec->cmax);
709 40210 : else if (xlrec->cmax != InvalidCommandId)
710 7288 : cid = xlrec->cmax;
711 32922 : else if (xlrec->cmin != InvalidCommandId)
712 32922 : cid = xlrec->cmin;
713 : else
714 : {
715 0 : cid = InvalidCommandId; /* silence compiler */
716 0 : elog(ERROR, "xl_heap_new_cid record without a valid CommandId");
717 : }
718 :
719 46596 : ReorderBufferAddNewCommandId(builder->reorder, xid, lsn, cid + 1);
720 46596 : }
721 :
722 : /*
723 : * Add a new Snapshot to all transactions we're decoding that currently are
724 : * in-progress so they can see new catalog contents made by the transaction
725 : * that just committed. This is necessary because those in-progress
726 : * transactions will use the new catalog's contents from here on (at the very
727 : * least everything they do needs to be compatible with newer catalog
728 : * contents).
729 : */
730 : static void
731 2042 : SnapBuildDistributeNewCatalogSnapshot(SnapBuild *builder, XLogRecPtr lsn)
732 : {
733 : dlist_iter txn_i;
734 : ReorderBufferTXN *txn;
735 :
736 : /*
737 : * Iterate through all toplevel transactions. This can include
738 : * subtransactions which we just don't yet know to be that, but that's
739 : * fine, they will just get an unnecessary snapshot queued.
740 : */
741 4146 : dlist_foreach(txn_i, &builder->reorder->toplevel_by_lsn)
742 : {
743 2104 : txn = dlist_container(ReorderBufferTXN, node, txn_i.cur);
744 :
745 : Assert(TransactionIdIsValid(txn->xid));
746 :
747 : /*
748 : * If we don't have a base snapshot yet, there are no changes in this
749 : * transaction which in turn implies we don't yet need a snapshot at
750 : * all. We'll add a snapshot when the first change gets queued.
751 : *
752 : * NB: This works correctly even for subtransactions because
753 : * ReorderBufferAssignChild() takes care to transfer the base snapshot
754 : * to the top-level transaction, and while iterating the changequeue
755 : * we'll get the change from the subtxn.
756 : */
757 2104 : if (!ReorderBufferXidHasBaseSnapshot(builder->reorder, txn->xid))
758 4 : continue;
759 :
760 : /*
761 : * We don't need to add snapshot to prepared transactions as they
762 : * should not see the new catalog contents.
763 : */
764 2100 : if (rbtxn_prepared(txn) || rbtxn_skip_prepared(txn))
765 56 : continue;
766 :
767 2044 : elog(DEBUG2, "adding a new snapshot to %u at %X/%X",
768 : txn->xid, LSN_FORMAT_ARGS(lsn));
769 :
770 : /*
771 : * increase the snapshot's refcount for the transaction we are handing
772 : * it out to
773 : */
774 2044 : SnapBuildSnapIncRefcount(builder->snapshot);
775 2044 : ReorderBufferAddSnapshot(builder->reorder, txn->xid, lsn,
776 : builder->snapshot);
777 : }
778 2042 : }
779 :
780 : /*
781 : * Keep track of a new catalog changing transaction that has committed.
782 : */
783 : static void
784 2064 : SnapBuildAddCommittedTxn(SnapBuild *builder, TransactionId xid)
785 : {
786 : Assert(TransactionIdIsValid(xid));
787 :
788 2064 : if (builder->committed.xcnt == builder->committed.xcnt_space)
789 : {
790 0 : builder->committed.xcnt_space = builder->committed.xcnt_space * 2 + 1;
791 :
792 0 : elog(DEBUG1, "increasing space for committed transactions to %u",
793 : (uint32) builder->committed.xcnt_space);
794 :
795 0 : builder->committed.xip = repalloc(builder->committed.xip,
796 0 : builder->committed.xcnt_space * sizeof(TransactionId));
797 : }
798 :
799 : /*
800 : * TODO: It might make sense to keep the array sorted here instead of
801 : * doing it every time we build a new snapshot. On the other hand this
802 : * gets called repeatedly when a transaction with subtransactions commits.
803 : */
804 2064 : builder->committed.xip[builder->committed.xcnt++] = xid;
805 2064 : }
806 :
807 : /*
808 : * Remove knowledge about transactions we treat as committed or containing catalog
809 : * changes that are smaller than ->xmin. Those won't ever get checked via
810 : * the ->committed or ->catchange array, respectively. The committed xids will
811 : * get checked via the clog machinery.
812 : *
813 : * We can ideally remove the transaction from catchange array once it is
814 : * finished (committed/aborted) but that could be costly as we need to maintain
815 : * the xids order in the array.
816 : */
817 : static void
818 706 : SnapBuildPurgeOlderTxn(SnapBuild *builder)
819 : {
820 : int off;
821 : TransactionId *workspace;
822 706 : int surviving_xids = 0;
823 :
824 : /* not ready yet */
825 706 : if (!TransactionIdIsNormal(builder->xmin))
826 0 : return;
827 :
828 : /* TODO: Neater algorithm than just copying and iterating? */
829 : workspace =
830 706 : MemoryContextAlloc(builder->context,
831 706 : builder->committed.xcnt * sizeof(TransactionId));
832 :
833 : /* copy xids that still are interesting to workspace */
834 1124 : for (off = 0; off < builder->committed.xcnt; off++)
835 : {
836 418 : if (NormalTransactionIdPrecedes(builder->committed.xip[off],
837 : builder->xmin))
838 : ; /* remove */
839 : else
840 0 : workspace[surviving_xids++] = builder->committed.xip[off];
841 : }
842 :
843 : /* copy workspace back to persistent state */
844 706 : memcpy(builder->committed.xip, workspace,
845 : surviving_xids * sizeof(TransactionId));
846 :
847 706 : elog(DEBUG3, "purged committed transactions from %u to %u, xmin: %u, xmax: %u",
848 : (uint32) builder->committed.xcnt, (uint32) surviving_xids,
849 : builder->xmin, builder->xmax);
850 706 : builder->committed.xcnt = surviving_xids;
851 :
852 706 : pfree(workspace);
853 :
854 : /*
855 : * Purge xids in ->catchange as well. The purged array must also be sorted
856 : * in xidComparator order.
857 : */
858 706 : if (builder->catchange.xcnt > 0)
859 : {
860 : /*
861 : * Since catchange.xip is sorted, we find the lower bound of xids that
862 : * are still interesting.
863 : */
864 20 : for (off = 0; off < builder->catchange.xcnt; off++)
865 : {
866 14 : if (TransactionIdFollowsOrEquals(builder->catchange.xip[off],
867 : builder->xmin))
868 2 : break;
869 : }
870 :
871 8 : surviving_xids = builder->catchange.xcnt - off;
872 :
873 8 : if (surviving_xids > 0)
874 : {
875 2 : memmove(builder->catchange.xip, &(builder->catchange.xip[off]),
876 : surviving_xids * sizeof(TransactionId));
877 : }
878 : else
879 : {
880 6 : pfree(builder->catchange.xip);
881 6 : builder->catchange.xip = NULL;
882 : }
883 :
884 8 : elog(DEBUG3, "purged catalog modifying transactions from %u to %u, xmin: %u, xmax: %u",
885 : (uint32) builder->catchange.xcnt, (uint32) surviving_xids,
886 : builder->xmin, builder->xmax);
887 8 : builder->catchange.xcnt = surviving_xids;
888 : }
889 : }
890 :
891 : /*
892 : * Handle everything that needs to be done when a transaction commits
893 : */
894 : void
895 5416 : SnapBuildCommitTxn(SnapBuild *builder, XLogRecPtr lsn, TransactionId xid,
896 : int nsubxacts, TransactionId *subxacts, uint32 xinfo)
897 : {
898 : int nxact;
899 :
900 5416 : bool needs_snapshot = false;
901 5416 : bool needs_timetravel = false;
902 5416 : bool sub_needs_timetravel = false;
903 :
904 5416 : TransactionId xmax = xid;
905 :
906 : /*
907 : * Transactions preceding BUILDING_SNAPSHOT will neither be decoded, nor
908 : * will they be part of a snapshot. So we don't need to record anything.
909 : */
910 5416 : if (builder->state == SNAPBUILD_START ||
911 5416 : (builder->state == SNAPBUILD_BUILDING_SNAPSHOT &&
912 0 : TransactionIdPrecedes(xid, builder->next_phase_at)))
913 : {
914 : /* ensure that only commits after this are getting replayed */
915 0 : if (builder->start_decoding_at <= lsn)
916 0 : builder->start_decoding_at = lsn + 1;
917 0 : return;
918 : }
919 :
920 5416 : if (builder->state < SNAPBUILD_CONSISTENT)
921 : {
922 : /* ensure that only commits after this are getting replayed */
923 10 : if (builder->start_decoding_at <= lsn)
924 4 : builder->start_decoding_at = lsn + 1;
925 :
926 : /*
927 : * If building an exportable snapshot, force xid to be tracked, even
928 : * if the transaction didn't modify the catalog.
929 : */
930 10 : if (builder->building_full_snapshot)
931 : {
932 0 : needs_timetravel = true;
933 : }
934 : }
935 :
936 7886 : for (nxact = 0; nxact < nsubxacts; nxact++)
937 : {
938 2470 : TransactionId subxid = subxacts[nxact];
939 :
940 : /*
941 : * Add subtransaction to base snapshot if catalog modifying, we don't
942 : * distinguish to toplevel transactions there.
943 : */
944 2470 : if (SnapBuildXidHasCatalogChanges(builder, subxid, xinfo))
945 : {
946 22 : sub_needs_timetravel = true;
947 22 : needs_snapshot = true;
948 :
949 22 : elog(DEBUG1, "found subtransaction %u:%u with catalog changes",
950 : xid, subxid);
951 :
952 22 : SnapBuildAddCommittedTxn(builder, subxid);
953 :
954 22 : if (NormalTransactionIdFollows(subxid, xmax))
955 22 : xmax = subxid;
956 : }
957 :
958 : /*
959 : * If we're forcing timetravel we also need visibility information
960 : * about subtransaction, so keep track of subtransaction's state, even
961 : * if not catalog modifying. Don't need to distribute a snapshot in
962 : * that case.
963 : */
964 2448 : else if (needs_timetravel)
965 : {
966 0 : SnapBuildAddCommittedTxn(builder, subxid);
967 0 : if (NormalTransactionIdFollows(subxid, xmax))
968 0 : xmax = subxid;
969 : }
970 : }
971 :
972 : /* if top-level modified catalog, it'll need a snapshot */
973 5416 : if (SnapBuildXidHasCatalogChanges(builder, xid, xinfo))
974 : {
975 2040 : elog(DEBUG2, "found top level transaction %u, with catalog changes",
976 : xid);
977 2040 : needs_snapshot = true;
978 2040 : needs_timetravel = true;
979 2040 : SnapBuildAddCommittedTxn(builder, xid);
980 : }
981 3376 : else if (sub_needs_timetravel)
982 : {
983 : /* track toplevel txn as well, subxact alone isn't meaningful */
984 2 : elog(DEBUG2, "forced transaction %u to do timetravel due to one of its subtransactions",
985 : xid);
986 2 : needs_timetravel = true;
987 2 : SnapBuildAddCommittedTxn(builder, xid);
988 : }
989 3374 : else if (needs_timetravel)
990 : {
991 0 : elog(DEBUG2, "forced transaction %u to do timetravel", xid);
992 :
993 0 : SnapBuildAddCommittedTxn(builder, xid);
994 : }
995 :
996 5416 : if (!needs_timetravel)
997 : {
998 : /* record that we cannot export a general snapshot anymore */
999 3374 : builder->committed.includes_all_transactions = false;
1000 : }
1001 :
1002 : Assert(!needs_snapshot || needs_timetravel);
1003 :
1004 : /*
1005 : * Adjust xmax of the snapshot builder, we only do that for committed,
1006 : * catalog modifying, transactions, everything else isn't interesting for
1007 : * us since we'll never look at the respective rows.
1008 : */
1009 5416 : if (needs_timetravel &&
1010 4084 : (!TransactionIdIsValid(builder->xmax) ||
1011 2042 : TransactionIdFollowsOrEquals(xmax, builder->xmax)))
1012 : {
1013 2038 : builder->xmax = xmax;
1014 2038 : TransactionIdAdvance(builder->xmax);
1015 : }
1016 :
1017 : /* if there's any reason to build a historic snapshot, do so now */
1018 5416 : if (needs_snapshot)
1019 : {
1020 : /*
1021 : * If we haven't built a complete snapshot yet there's no need to hand
1022 : * it out, it wouldn't (and couldn't) be used anyway.
1023 : */
1024 2042 : if (builder->state < SNAPBUILD_FULL_SNAPSHOT)
1025 0 : return;
1026 :
1027 : /*
1028 : * Decrease the snapshot builder's refcount of the old snapshot, note
1029 : * that it still will be used if it has been handed out to the
1030 : * reorderbuffer earlier.
1031 : */
1032 2042 : if (builder->snapshot)
1033 2032 : SnapBuildSnapDecRefcount(builder->snapshot);
1034 :
1035 2042 : builder->snapshot = SnapBuildBuildSnapshot(builder);
1036 :
1037 : /* we might need to execute invalidations, add snapshot */
1038 2042 : if (!ReorderBufferXidHasBaseSnapshot(builder->reorder, xid))
1039 : {
1040 42 : SnapBuildSnapIncRefcount(builder->snapshot);
1041 42 : ReorderBufferSetBaseSnapshot(builder->reorder, xid, lsn,
1042 : builder->snapshot);
1043 : }
1044 :
1045 : /* refcount of the snapshot builder for the new snapshot */
1046 2042 : SnapBuildSnapIncRefcount(builder->snapshot);
1047 :
1048 : /* add a new catalog snapshot to all currently running transactions */
1049 2042 : SnapBuildDistributeNewCatalogSnapshot(builder, lsn);
1050 : }
1051 : }
1052 :
1053 : /*
1054 : * Check the reorder buffer and the snapshot to see if the given transaction has
1055 : * modified catalogs.
1056 : */
1057 : static inline bool
1058 7886 : SnapBuildXidHasCatalogChanges(SnapBuild *builder, TransactionId xid,
1059 : uint32 xinfo)
1060 : {
1061 7886 : if (ReorderBufferXidHasCatalogChanges(builder->reorder, xid))
1062 2050 : return true;
1063 :
1064 : /*
1065 : * The transactions that have changed catalogs must have invalidation
1066 : * info.
1067 : */
1068 5836 : if (!(xinfo & XACT_XINFO_HAS_INVALS))
1069 5816 : return false;
1070 :
1071 : /* Check the catchange XID array */
1072 32 : return ((builder->catchange.xcnt > 0) &&
1073 12 : (bsearch(&xid, builder->catchange.xip, builder->catchange.xcnt,
1074 : sizeof(TransactionId), xidComparator) != NULL));
1075 : }
1076 :
1077 : /* -----------------------------------
1078 : * Snapshot building functions dealing with xlog records
1079 : * -----------------------------------
1080 : */
1081 :
1082 : /*
1083 : * Process a running xacts record, and use its information to first build a
1084 : * historic snapshot and later to release resources that aren't needed
1085 : * anymore.
1086 : */
1087 : void
1088 2654 : SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xacts *running)
1089 : {
1090 : ReorderBufferTXN *txn;
1091 : TransactionId xmin;
1092 :
1093 : /*
1094 : * If we're not consistent yet, inspect the record to see whether it
1095 : * allows to get closer to being consistent. If we are consistent, dump
1096 : * our snapshot so others or we, after a restart, can use it.
1097 : */
1098 2654 : if (builder->state < SNAPBUILD_CONSISTENT)
1099 : {
1100 : /* returns false if there's no point in performing cleanup just yet */
1101 1992 : if (!SnapBuildFindSnapshot(builder, lsn, running))
1102 1944 : return;
1103 : }
1104 : else
1105 662 : SnapBuildSerialize(builder, lsn);
1106 :
1107 : /*
1108 : * Update range of interesting xids based on the running xacts
1109 : * information. We don't increase ->xmax using it, because once we are in
1110 : * a consistent state we can do that ourselves and much more efficiently
1111 : * so, because we only need to do it for catalog transactions since we
1112 : * only ever look at those.
1113 : *
1114 : * NB: We only increase xmax when a catalog modifying transaction commits
1115 : * (see SnapBuildCommitTxn). Because of this, xmax can be lower than
1116 : * xmin, which looks odd but is correct and actually more efficient, since
1117 : * we hit fast paths in heapam_visibility.c.
1118 : */
1119 706 : builder->xmin = running->oldestRunningXid;
1120 :
1121 : /* Remove transactions we don't need to keep track off anymore */
1122 706 : SnapBuildPurgeOlderTxn(builder);
1123 :
1124 : /*
1125 : * Advance the xmin limit for the current replication slot, to allow
1126 : * vacuum to clean up the tuples this slot has been protecting.
1127 : *
1128 : * The reorderbuffer might have an xmin among the currently running
1129 : * snapshots; use it if so. If not, we need only consider the snapshots
1130 : * we'll produce later, which can't be less than the oldest running xid in
1131 : * the record we're reading now.
1132 : */
1133 706 : xmin = ReorderBufferGetOldestXmin(builder->reorder);
1134 706 : if (xmin == InvalidTransactionId)
1135 618 : xmin = running->oldestRunningXid;
1136 706 : elog(DEBUG3, "xmin: %u, xmax: %u, oldest running: %u, oldest xmin: %u",
1137 : builder->xmin, builder->xmax, running->oldestRunningXid, xmin);
1138 706 : LogicalIncreaseXminForSlot(lsn, xmin);
1139 :
1140 : /*
1141 : * Also tell the slot where we can restart decoding from. We don't want to
1142 : * do that after every commit because changing that implies an fsync of
1143 : * the logical slot's state file, so we only do it every time we see a
1144 : * running xacts record.
1145 : *
1146 : * Do so by looking for the oldest in progress transaction (determined by
1147 : * the first LSN of any of its relevant records). Every transaction
1148 : * remembers the last location we stored the snapshot to disk before its
1149 : * beginning. That point is where we can restart from.
1150 : */
1151 :
1152 : /*
1153 : * Can't know about a serialized snapshot's location if we're not
1154 : * consistent.
1155 : */
1156 706 : if (builder->state < SNAPBUILD_CONSISTENT)
1157 34 : return;
1158 :
1159 672 : txn = ReorderBufferGetOldestTXN(builder->reorder);
1160 :
1161 : /*
1162 : * oldest ongoing txn might have started when we didn't yet serialize
1163 : * anything because we hadn't reached a consistent state yet.
1164 : */
1165 672 : if (txn != NULL && txn->restart_decoding_lsn != InvalidXLogRecPtr)
1166 50 : LogicalIncreaseRestartDecodingForSlot(lsn, txn->restart_decoding_lsn);
1167 :
1168 : /*
1169 : * No in-progress transaction, can reuse the last serialized snapshot if
1170 : * we have one.
1171 : */
1172 622 : else if (txn == NULL &&
1173 564 : builder->reorder->current_restart_decoding_lsn != InvalidXLogRecPtr &&
1174 560 : builder->last_serialized_snapshot != InvalidXLogRecPtr)
1175 560 : LogicalIncreaseRestartDecodingForSlot(lsn,
1176 : builder->last_serialized_snapshot);
1177 : }
1178 :
1179 :
1180 : /*
1181 : * Build the start of a snapshot that's capable of decoding the catalog.
1182 : *
1183 : * Helper function for SnapBuildProcessRunningXacts() while we're not yet
1184 : * consistent.
1185 : *
1186 : * Returns true if there is a point in performing internal maintenance/cleanup
1187 : * using the xl_running_xacts record.
1188 : */
1189 : static bool
1190 1992 : SnapBuildFindSnapshot(SnapBuild *builder, XLogRecPtr lsn, xl_running_xacts *running)
1191 : {
1192 : /* ---
1193 : * Build catalog decoding snapshot incrementally using information about
1194 : * the currently running transactions. There are several ways to do that:
1195 : *
1196 : * a) There were no running transactions when the xl_running_xacts record
1197 : * was inserted, jump to CONSISTENT immediately. We might find such a
1198 : * state while waiting on c)'s sub-states.
1199 : *
1200 : * b) This (in a previous run) or another decoding slot serialized a
1201 : * snapshot to disk that we can use. Can't use this method while finding
1202 : * the start point for decoding changes as the restart LSN would be an
1203 : * arbitrary LSN but we need to find the start point to extract changes
1204 : * where we won't see the data for partial transactions. Also, we cannot
1205 : * use this method when a slot needs a full snapshot for export or direct
1206 : * use, as that snapshot will only contain catalog modifying transactions.
1207 : *
1208 : * c) First incrementally build a snapshot for catalog tuples
1209 : * (BUILDING_SNAPSHOT), that requires all, already in-progress,
1210 : * transactions to finish. Every transaction starting after that
1211 : * (FULL_SNAPSHOT state), has enough information to be decoded. But
1212 : * for older running transactions no viable snapshot exists yet, so
1213 : * CONSISTENT will only be reached once all of those have finished.
1214 : * ---
1215 : */
1216 :
1217 : /*
1218 : * xl_running_xacts record is older than what we can use, we might not
1219 : * have all necessary catalog rows anymore.
1220 : */
1221 1992 : if (TransactionIdIsNormal(builder->initial_xmin_horizon) &&
1222 856 : NormalTransactionIdPrecedes(running->oldestRunningXid,
1223 : builder->initial_xmin_horizon))
1224 : {
1225 0 : ereport(DEBUG1,
1226 : (errmsg_internal("skipping snapshot at %X/%X while building logical decoding snapshot, xmin horizon too low",
1227 : LSN_FORMAT_ARGS(lsn)),
1228 : errdetail_internal("initial xmin horizon of %u vs the snapshot's %u",
1229 : builder->initial_xmin_horizon, running->oldestRunningXid)));
1230 :
1231 :
1232 0 : SnapBuildWaitSnapshot(running, builder->initial_xmin_horizon);
1233 :
1234 0 : return true;
1235 : }
1236 :
1237 : /*
1238 : * a) No transaction were running, we can jump to consistent.
1239 : *
1240 : * This is not affected by races around xl_running_xacts, because we can
1241 : * miss transaction commits, but currently not transactions starting.
1242 : *
1243 : * NB: We might have already started to incrementally assemble a snapshot,
1244 : * so we need to be careful to deal with that.
1245 : */
1246 1992 : if (running->oldestRunningXid == running->nextXid)
1247 : {
1248 1926 : if (builder->start_decoding_at == InvalidXLogRecPtr ||
1249 1086 : builder->start_decoding_at <= lsn)
1250 : /* can decode everything after this */
1251 842 : builder->start_decoding_at = lsn + 1;
1252 :
1253 : /* As no transactions were running xmin/xmax can be trivially set. */
1254 1926 : builder->xmin = running->nextXid; /* < are finished */
1255 1926 : builder->xmax = running->nextXid; /* >= are running */
1256 :
1257 : /* so we can safely use the faster comparisons */
1258 : Assert(TransactionIdIsNormal(builder->xmin));
1259 : Assert(TransactionIdIsNormal(builder->xmax));
1260 :
1261 1926 : builder->state = SNAPBUILD_CONSISTENT;
1262 1926 : builder->next_phase_at = InvalidTransactionId;
1263 :
1264 1926 : ereport(LOG,
1265 : (errmsg("logical decoding found consistent point at %X/%X",
1266 : LSN_FORMAT_ARGS(lsn)),
1267 : errdetail("There are no running transactions.")));
1268 :
1269 1926 : return false;
1270 : }
1271 :
1272 : /*
1273 : * b) valid on disk state and while neither building full snapshot nor
1274 : * creating a slot.
1275 : */
1276 66 : else if (!builder->building_full_snapshot &&
1277 104 : !builder->in_slot_creation &&
1278 40 : SnapBuildRestore(builder, lsn))
1279 : {
1280 : /* there won't be any state to cleanup */
1281 18 : return false;
1282 : }
1283 :
1284 : /*
1285 : * c) transition from START to BUILDING_SNAPSHOT.
1286 : *
1287 : * In START state, and a xl_running_xacts record with running xacts is
1288 : * encountered. In that case, switch to BUILDING_SNAPSHOT state, and
1289 : * record xl_running_xacts->nextXid. Once all running xacts have finished
1290 : * (i.e. they're all >= nextXid), we have a complete catalog snapshot. It
1291 : * might look that we could use xl_running_xacts's ->xids information to
1292 : * get there quicker, but that is problematic because transactions marked
1293 : * as running, might already have inserted their commit record - it's
1294 : * infeasible to change that with locking.
1295 : */
1296 48 : else if (builder->state == SNAPBUILD_START)
1297 : {
1298 26 : builder->state = SNAPBUILD_BUILDING_SNAPSHOT;
1299 26 : builder->next_phase_at = running->nextXid;
1300 :
1301 : /*
1302 : * Start with an xmin/xmax that's correct for future, when all the
1303 : * currently running transactions have finished. We'll update both
1304 : * while waiting for the pending transactions to finish.
1305 : */
1306 26 : builder->xmin = running->nextXid; /* < are finished */
1307 26 : builder->xmax = running->nextXid; /* >= are running */
1308 :
1309 : /* so we can safely use the faster comparisons */
1310 : Assert(TransactionIdIsNormal(builder->xmin));
1311 : Assert(TransactionIdIsNormal(builder->xmax));
1312 :
1313 26 : ereport(LOG,
1314 : (errmsg("logical decoding found initial starting point at %X/%X",
1315 : LSN_FORMAT_ARGS(lsn)),
1316 : errdetail("Waiting for transactions (approximately %d) older than %u to end.",
1317 : running->xcnt, running->nextXid)));
1318 :
1319 26 : SnapBuildWaitSnapshot(running, running->nextXid);
1320 : }
1321 :
1322 : /*
1323 : * c) transition from BUILDING_SNAPSHOT to FULL_SNAPSHOT.
1324 : *
1325 : * In BUILDING_SNAPSHOT state, and this xl_running_xacts' oldestRunningXid
1326 : * is >= than nextXid from when we switched to BUILDING_SNAPSHOT. This
1327 : * means all transactions starting afterwards have enough information to
1328 : * be decoded. Switch to FULL_SNAPSHOT.
1329 : */
1330 34 : else if (builder->state == SNAPBUILD_BUILDING_SNAPSHOT &&
1331 12 : TransactionIdPrecedesOrEquals(builder->next_phase_at,
1332 : running->oldestRunningXid))
1333 : {
1334 10 : builder->state = SNAPBUILD_FULL_SNAPSHOT;
1335 10 : builder->next_phase_at = running->nextXid;
1336 :
1337 10 : ereport(LOG,
1338 : (errmsg("logical decoding found initial consistent point at %X/%X",
1339 : LSN_FORMAT_ARGS(lsn)),
1340 : errdetail("Waiting for transactions (approximately %d) older than %u to end.",
1341 : running->xcnt, running->nextXid)));
1342 :
1343 10 : SnapBuildWaitSnapshot(running, running->nextXid);
1344 : }
1345 :
1346 : /*
1347 : * c) transition from FULL_SNAPSHOT to CONSISTENT.
1348 : *
1349 : * In FULL_SNAPSHOT state, and this xl_running_xacts' oldestRunningXid is
1350 : * >= than nextXid from when we switched to FULL_SNAPSHOT. This means all
1351 : * transactions that are currently in progress have a catalog snapshot,
1352 : * and all their changes have been collected. Switch to CONSISTENT.
1353 : */
1354 22 : else if (builder->state == SNAPBUILD_FULL_SNAPSHOT &&
1355 10 : TransactionIdPrecedesOrEquals(builder->next_phase_at,
1356 : running->oldestRunningXid))
1357 : {
1358 10 : builder->state = SNAPBUILD_CONSISTENT;
1359 10 : builder->next_phase_at = InvalidTransactionId;
1360 :
1361 10 : ereport(LOG,
1362 : (errmsg("logical decoding found consistent point at %X/%X",
1363 : LSN_FORMAT_ARGS(lsn)),
1364 : errdetail("There are no old transactions anymore.")));
1365 : }
1366 :
1367 : /*
1368 : * We already started to track running xacts and need to wait for all
1369 : * in-progress ones to finish. We fall through to the normal processing of
1370 : * records so incremental cleanup can be performed.
1371 : */
1372 44 : return true;
1373 : }
1374 :
1375 : /* ---
1376 : * Iterate through xids in record, wait for all older than the cutoff to
1377 : * finish. Then, if possible, log a new xl_running_xacts record.
1378 : *
1379 : * This isn't required for the correctness of decoding, but to:
1380 : * a) allow isolationtester to notice that we're currently waiting for
1381 : * something.
1382 : * b) log a new xl_running_xacts record where it'd be helpful, without having
1383 : * to wait for bgwriter or checkpointer.
1384 : * ---
1385 : */
1386 : static void
1387 36 : SnapBuildWaitSnapshot(xl_running_xacts *running, TransactionId cutoff)
1388 : {
1389 : int off;
1390 :
1391 68 : for (off = 0; off < running->xcnt; off++)
1392 : {
1393 36 : TransactionId xid = running->xids[off];
1394 :
1395 : /*
1396 : * Upper layers should prevent that we ever need to wait on ourselves.
1397 : * Check anyway, since failing to do so would either result in an
1398 : * endless wait or an Assert() failure.
1399 : */
1400 36 : if (TransactionIdIsCurrentTransactionId(xid))
1401 0 : elog(ERROR, "waiting for ourselves");
1402 :
1403 36 : if (TransactionIdFollows(xid, cutoff))
1404 0 : continue;
1405 :
1406 36 : XactLockTableWait(xid, NULL, NULL, XLTW_None);
1407 : }
1408 :
1409 : /*
1410 : * All transactions we needed to finish finished - try to ensure there is
1411 : * another xl_running_xacts record in a timely manner, without having to
1412 : * wait for bgwriter or checkpointer to log one. During recovery we can't
1413 : * enforce that, so we'll have to wait.
1414 : */
1415 32 : if (!RecoveryInProgress())
1416 : {
1417 32 : LogStandbySnapshot();
1418 : }
1419 32 : }
1420 :
1421 : #define SnapBuildOnDiskConstantSize \
1422 : offsetof(SnapBuildOnDisk, builder)
1423 : #define SnapBuildOnDiskNotChecksummedSize \
1424 : offsetof(SnapBuildOnDisk, version)
1425 :
1426 : #define SNAPBUILD_MAGIC 0x51A1E001
1427 : #define SNAPBUILD_VERSION 6
1428 :
1429 : /*
1430 : * Store/Load a snapshot from disk, depending on the snapshot builder's state.
1431 : *
1432 : * Supposed to be used by external (i.e. not snapbuild.c) code that just read
1433 : * a record that's a potential location for a serialized snapshot.
1434 : */
1435 : void
1436 76 : SnapBuildSerializationPoint(SnapBuild *builder, XLogRecPtr lsn)
1437 : {
1438 76 : if (builder->state < SNAPBUILD_CONSISTENT)
1439 0 : SnapBuildRestore(builder, lsn);
1440 : else
1441 76 : SnapBuildSerialize(builder, lsn);
1442 76 : }
1443 :
1444 : /*
1445 : * Serialize the snapshot 'builder' at the location 'lsn' if it hasn't already
1446 : * been done by another decoding process.
1447 : */
1448 : static void
1449 738 : SnapBuildSerialize(SnapBuild *builder, XLogRecPtr lsn)
1450 : {
1451 : Size needed_length;
1452 738 : SnapBuildOnDisk *ondisk = NULL;
1453 738 : TransactionId *catchange_xip = NULL;
1454 : MemoryContext old_ctx;
1455 : size_t catchange_xcnt;
1456 : char *ondisk_c;
1457 : int fd;
1458 : char tmppath[MAXPGPATH];
1459 : char path[MAXPGPATH];
1460 : int ret;
1461 : struct stat stat_buf;
1462 : Size sz;
1463 :
1464 : Assert(lsn != InvalidXLogRecPtr);
1465 : Assert(builder->last_serialized_snapshot == InvalidXLogRecPtr ||
1466 : builder->last_serialized_snapshot <= lsn);
1467 :
1468 : /*
1469 : * no point in serializing if we cannot continue to work immediately after
1470 : * restoring the snapshot
1471 : */
1472 738 : if (builder->state < SNAPBUILD_CONSISTENT)
1473 0 : return;
1474 :
1475 : /* consistent snapshots have no next phase */
1476 : Assert(builder->next_phase_at == InvalidTransactionId);
1477 :
1478 : /*
1479 : * We identify snapshots by the LSN they are valid for. We don't need to
1480 : * include timelines in the name as each LSN maps to exactly one timeline
1481 : * unless the user used pg_resetwal or similar. If a user did so, there's
1482 : * no hope continuing to decode anyway.
1483 : */
1484 738 : sprintf(path, "%s/%X-%X.snap",
1485 : PG_LOGICAL_SNAPSHOTS_DIR,
1486 738 : LSN_FORMAT_ARGS(lsn));
1487 :
1488 : /*
1489 : * first check whether some other backend already has written the snapshot
1490 : * for this LSN. It's perfectly fine if there's none, so we accept ENOENT
1491 : * as a valid state. Everything else is an unexpected error.
1492 : */
1493 738 : ret = stat(path, &stat_buf);
1494 :
1495 738 : if (ret != 0 && errno != ENOENT)
1496 0 : ereport(ERROR,
1497 : (errcode_for_file_access(),
1498 : errmsg("could not stat file \"%s\": %m", path)));
1499 :
1500 738 : else if (ret == 0)
1501 : {
1502 : /*
1503 : * somebody else has already serialized to this point, don't overwrite
1504 : * but remember location, so we don't need to read old data again.
1505 : *
1506 : * To be sure it has been synced to disk after the rename() from the
1507 : * tempfile filename to the real filename, we just repeat the fsync.
1508 : * That ought to be cheap because in most scenarios it should already
1509 : * be safely on disk.
1510 : */
1511 186 : fsync_fname(path, false);
1512 186 : fsync_fname(PG_LOGICAL_SNAPSHOTS_DIR, true);
1513 :
1514 186 : builder->last_serialized_snapshot = lsn;
1515 186 : goto out;
1516 : }
1517 :
1518 : /*
1519 : * there is an obvious race condition here between the time we stat(2) the
1520 : * file and us writing the file. But we rename the file into place
1521 : * atomically and all files created need to contain the same data anyway,
1522 : * so this is perfectly fine, although a bit of a resource waste. Locking
1523 : * seems like pointless complication.
1524 : */
1525 552 : elog(DEBUG1, "serializing snapshot to %s", path);
1526 :
1527 : /* to make sure only we will write to this tempfile, include pid */
1528 552 : sprintf(tmppath, "%s/%X-%X.snap.%d.tmp",
1529 : PG_LOGICAL_SNAPSHOTS_DIR,
1530 552 : LSN_FORMAT_ARGS(lsn), MyProcPid);
1531 :
1532 : /*
1533 : * Unlink temporary file if it already exists, needs to have been before a
1534 : * crash/error since we won't enter this function twice from within a
1535 : * single decoding slot/backend and the temporary file contains the pid of
1536 : * the current process.
1537 : */
1538 552 : if (unlink(tmppath) != 0 && errno != ENOENT)
1539 0 : ereport(ERROR,
1540 : (errcode_for_file_access(),
1541 : errmsg("could not remove file \"%s\": %m", tmppath)));
1542 :
1543 552 : old_ctx = MemoryContextSwitchTo(builder->context);
1544 :
1545 : /* Get the catalog modifying transactions that are yet not committed */
1546 552 : catchange_xip = ReorderBufferGetCatalogChangesXacts(builder->reorder);
1547 552 : catchange_xcnt = dclist_count(&builder->reorder->catchange_txns);
1548 :
1549 552 : needed_length = sizeof(SnapBuildOnDisk) +
1550 552 : sizeof(TransactionId) * (builder->committed.xcnt + catchange_xcnt);
1551 :
1552 552 : ondisk_c = palloc0(needed_length);
1553 552 : ondisk = (SnapBuildOnDisk *) ondisk_c;
1554 552 : ondisk->magic = SNAPBUILD_MAGIC;
1555 552 : ondisk->version = SNAPBUILD_VERSION;
1556 552 : ondisk->length = needed_length;
1557 552 : INIT_CRC32C(ondisk->checksum);
1558 552 : COMP_CRC32C(ondisk->checksum,
1559 : ((char *) ondisk) + SnapBuildOnDiskNotChecksummedSize,
1560 : SnapBuildOnDiskConstantSize - SnapBuildOnDiskNotChecksummedSize);
1561 552 : ondisk_c += sizeof(SnapBuildOnDisk);
1562 :
1563 552 : memcpy(&ondisk->builder, builder, sizeof(SnapBuild));
1564 : /* NULL-ify memory-only data */
1565 552 : ondisk->builder.context = NULL;
1566 552 : ondisk->builder.snapshot = NULL;
1567 552 : ondisk->builder.reorder = NULL;
1568 552 : ondisk->builder.committed.xip = NULL;
1569 552 : ondisk->builder.catchange.xip = NULL;
1570 : /* update catchange only on disk data */
1571 552 : ondisk->builder.catchange.xcnt = catchange_xcnt;
1572 :
1573 552 : COMP_CRC32C(ondisk->checksum,
1574 : &ondisk->builder,
1575 : sizeof(SnapBuild));
1576 :
1577 : /* copy committed xacts */
1578 552 : if (builder->committed.xcnt > 0)
1579 : {
1580 96 : sz = sizeof(TransactionId) * builder->committed.xcnt;
1581 96 : memcpy(ondisk_c, builder->committed.xip, sz);
1582 96 : COMP_CRC32C(ondisk->checksum, ondisk_c, sz);
1583 96 : ondisk_c += sz;
1584 : }
1585 :
1586 : /* copy catalog modifying xacts */
1587 552 : if (catchange_xcnt > 0)
1588 : {
1589 18 : sz = sizeof(TransactionId) * catchange_xcnt;
1590 18 : memcpy(ondisk_c, catchange_xip, sz);
1591 18 : COMP_CRC32C(ondisk->checksum, ondisk_c, sz);
1592 18 : ondisk_c += sz;
1593 : }
1594 :
1595 552 : FIN_CRC32C(ondisk->checksum);
1596 :
1597 : /* we have valid data now, open tempfile and write it there */
1598 552 : fd = OpenTransientFile(tmppath,
1599 : O_CREAT | O_EXCL | O_WRONLY | PG_BINARY);
1600 552 : if (fd < 0)
1601 0 : ereport(ERROR,
1602 : (errcode_for_file_access(),
1603 : errmsg("could not open file \"%s\": %m", tmppath)));
1604 :
1605 552 : errno = 0;
1606 552 : pgstat_report_wait_start(WAIT_EVENT_SNAPBUILD_WRITE);
1607 552 : if ((write(fd, ondisk, needed_length)) != needed_length)
1608 : {
1609 0 : int save_errno = errno;
1610 :
1611 0 : CloseTransientFile(fd);
1612 :
1613 : /* if write didn't set errno, assume problem is no disk space */
1614 0 : errno = save_errno ? save_errno : ENOSPC;
1615 0 : ereport(ERROR,
1616 : (errcode_for_file_access(),
1617 : errmsg("could not write to file \"%s\": %m", tmppath)));
1618 : }
1619 552 : pgstat_report_wait_end();
1620 :
1621 : /*
1622 : * fsync the file before renaming so that even if we crash after this we
1623 : * have either a fully valid file or nothing.
1624 : *
1625 : * It's safe to just ERROR on fsync() here because we'll retry the whole
1626 : * operation including the writes.
1627 : *
1628 : * TODO: Do the fsync() via checkpoints/restartpoints, doing it here has
1629 : * some noticeable overhead since it's performed synchronously during
1630 : * decoding?
1631 : */
1632 552 : pgstat_report_wait_start(WAIT_EVENT_SNAPBUILD_SYNC);
1633 552 : if (pg_fsync(fd) != 0)
1634 : {
1635 0 : int save_errno = errno;
1636 :
1637 0 : CloseTransientFile(fd);
1638 0 : errno = save_errno;
1639 0 : ereport(ERROR,
1640 : (errcode_for_file_access(),
1641 : errmsg("could not fsync file \"%s\": %m", tmppath)));
1642 : }
1643 552 : pgstat_report_wait_end();
1644 :
1645 552 : if (CloseTransientFile(fd) != 0)
1646 0 : ereport(ERROR,
1647 : (errcode_for_file_access(),
1648 : errmsg("could not close file \"%s\": %m", tmppath)));
1649 :
1650 552 : fsync_fname(PG_LOGICAL_SNAPSHOTS_DIR, true);
1651 :
1652 : /*
1653 : * We may overwrite the work from some other backend, but that's ok, our
1654 : * snapshot is valid as well, we'll just have done some superfluous work.
1655 : */
1656 552 : if (rename(tmppath, path) != 0)
1657 : {
1658 0 : ereport(ERROR,
1659 : (errcode_for_file_access(),
1660 : errmsg("could not rename file \"%s\" to \"%s\": %m",
1661 : tmppath, path)));
1662 : }
1663 :
1664 : /* make sure we persist */
1665 552 : fsync_fname(path, false);
1666 552 : fsync_fname(PG_LOGICAL_SNAPSHOTS_DIR, true);
1667 :
1668 : /*
1669 : * Now there's no way we can lose the dumped state anymore, remember this
1670 : * as a serialization point.
1671 : */
1672 552 : builder->last_serialized_snapshot = lsn;
1673 :
1674 552 : MemoryContextSwitchTo(old_ctx);
1675 :
1676 738 : out:
1677 738 : ReorderBufferSetRestartPoint(builder->reorder,
1678 : builder->last_serialized_snapshot);
1679 : /* be tidy */
1680 738 : if (ondisk)
1681 552 : pfree(ondisk);
1682 738 : if (catchange_xip)
1683 18 : pfree(catchange_xip);
1684 : }
1685 :
1686 : /*
1687 : * Restore the logical snapshot file contents to 'ondisk'.
1688 : *
1689 : * 'context' is the memory context where the catalog modifying/committed xid
1690 : * will live.
1691 : * If 'missing_ok' is true, will not throw an error if the file is not found.
1692 : */
1693 : bool
1694 48 : SnapBuildRestoreSnapshot(SnapBuildOnDisk *ondisk, const char *path,
1695 : MemoryContext context, bool missing_ok)
1696 : {
1697 : int fd;
1698 : pg_crc32c checksum;
1699 : Size sz;
1700 :
1701 48 : fd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
1702 :
1703 48 : if (fd < 0)
1704 : {
1705 22 : if (missing_ok && errno == ENOENT)
1706 22 : return false;
1707 :
1708 0 : ereport(ERROR,
1709 : (errcode_for_file_access(),
1710 : errmsg("could not open file \"%s\": %m", path)));
1711 : }
1712 :
1713 : /* ----
1714 : * Make sure the snapshot had been stored safely to disk, that's normally
1715 : * cheap.
1716 : * Note that we do not need PANIC here, nobody will be able to use the
1717 : * slot without fsyncing, and saving it won't succeed without an fsync()
1718 : * either...
1719 : * ----
1720 : */
1721 26 : fsync_fname(path, false);
1722 26 : fsync_fname(PG_LOGICAL_SNAPSHOTS_DIR, true);
1723 :
1724 : /* read statically sized portion of snapshot */
1725 26 : SnapBuildRestoreContents(fd, (char *) ondisk, SnapBuildOnDiskConstantSize, path);
1726 :
1727 26 : if (ondisk->magic != SNAPBUILD_MAGIC)
1728 0 : ereport(ERROR,
1729 : (errcode(ERRCODE_DATA_CORRUPTED),
1730 : errmsg("snapbuild state file \"%s\" has wrong magic number: %u instead of %u",
1731 : path, ondisk->magic, SNAPBUILD_MAGIC)));
1732 :
1733 26 : if (ondisk->version != SNAPBUILD_VERSION)
1734 0 : ereport(ERROR,
1735 : (errcode(ERRCODE_DATA_CORRUPTED),
1736 : errmsg("snapbuild state file \"%s\" has unsupported version: %u instead of %u",
1737 : path, ondisk->version, SNAPBUILD_VERSION)));
1738 :
1739 26 : INIT_CRC32C(checksum);
1740 26 : COMP_CRC32C(checksum,
1741 : ((char *) ondisk) + SnapBuildOnDiskNotChecksummedSize,
1742 : SnapBuildOnDiskConstantSize - SnapBuildOnDiskNotChecksummedSize);
1743 :
1744 : /* read SnapBuild */
1745 26 : SnapBuildRestoreContents(fd, (char *) &ondisk->builder, sizeof(SnapBuild), path);
1746 26 : COMP_CRC32C(checksum, &ondisk->builder, sizeof(SnapBuild));
1747 :
1748 : /* restore committed xacts information */
1749 26 : if (ondisk->builder.committed.xcnt > 0)
1750 : {
1751 8 : sz = sizeof(TransactionId) * ondisk->builder.committed.xcnt;
1752 8 : ondisk->builder.committed.xip = MemoryContextAllocZero(context, sz);
1753 8 : SnapBuildRestoreContents(fd, (char *) ondisk->builder.committed.xip, sz, path);
1754 8 : COMP_CRC32C(checksum, ondisk->builder.committed.xip, sz);
1755 : }
1756 :
1757 : /* restore catalog modifying xacts information */
1758 26 : if (ondisk->builder.catchange.xcnt > 0)
1759 : {
1760 12 : sz = sizeof(TransactionId) * ondisk->builder.catchange.xcnt;
1761 12 : ondisk->builder.catchange.xip = MemoryContextAllocZero(context, sz);
1762 12 : SnapBuildRestoreContents(fd, (char *) ondisk->builder.catchange.xip, sz, path);
1763 12 : COMP_CRC32C(checksum, ondisk->builder.catchange.xip, sz);
1764 : }
1765 :
1766 26 : if (CloseTransientFile(fd) != 0)
1767 0 : ereport(ERROR,
1768 : (errcode_for_file_access(),
1769 : errmsg("could not close file \"%s\": %m", path)));
1770 :
1771 26 : FIN_CRC32C(checksum);
1772 :
1773 : /* verify checksum of what we've read */
1774 26 : if (!EQ_CRC32C(checksum, ondisk->checksum))
1775 0 : ereport(ERROR,
1776 : (errcode(ERRCODE_DATA_CORRUPTED),
1777 : errmsg("checksum mismatch for snapbuild state file \"%s\": is %u, should be %u",
1778 : path, checksum, ondisk->checksum)));
1779 :
1780 26 : return true;
1781 : }
1782 :
1783 : /*
1784 : * Restore a snapshot into 'builder' if previously one has been stored at the
1785 : * location indicated by 'lsn'. Returns true if successful, false otherwise.
1786 : */
1787 : static bool
1788 40 : SnapBuildRestore(SnapBuild *builder, XLogRecPtr lsn)
1789 : {
1790 : SnapBuildOnDisk ondisk;
1791 : char path[MAXPGPATH];
1792 :
1793 : /* no point in loading a snapshot if we're already there */
1794 40 : if (builder->state == SNAPBUILD_CONSISTENT)
1795 0 : return false;
1796 :
1797 40 : sprintf(path, "%s/%X-%X.snap",
1798 : PG_LOGICAL_SNAPSHOTS_DIR,
1799 40 : LSN_FORMAT_ARGS(lsn));
1800 :
1801 : /* validate and restore the snapshot to 'ondisk' */
1802 40 : if (!SnapBuildRestoreSnapshot(&ondisk, path, builder->context, true))
1803 22 : return false;
1804 :
1805 : /*
1806 : * ok, we now have a sensible snapshot here, figure out if it has more
1807 : * information than we have.
1808 : */
1809 :
1810 : /*
1811 : * We are only interested in consistent snapshots for now, comparing
1812 : * whether one incomplete snapshot is more "advanced" seems to be
1813 : * unnecessarily complex.
1814 : */
1815 18 : if (ondisk.builder.state < SNAPBUILD_CONSISTENT)
1816 0 : goto snapshot_not_interesting;
1817 :
1818 : /*
1819 : * Don't use a snapshot that requires an xmin that we cannot guarantee to
1820 : * be available.
1821 : */
1822 18 : if (TransactionIdPrecedes(ondisk.builder.xmin, builder->initial_xmin_horizon))
1823 0 : goto snapshot_not_interesting;
1824 :
1825 : /*
1826 : * Consistent snapshots have no next phase. Reset next_phase_at as it is
1827 : * possible that an old value may remain.
1828 : */
1829 : Assert(ondisk.builder.next_phase_at == InvalidTransactionId);
1830 18 : builder->next_phase_at = InvalidTransactionId;
1831 :
1832 : /* ok, we think the snapshot is sensible, copy over everything important */
1833 18 : builder->xmin = ondisk.builder.xmin;
1834 18 : builder->xmax = ondisk.builder.xmax;
1835 18 : builder->state = ondisk.builder.state;
1836 :
1837 18 : builder->committed.xcnt = ondisk.builder.committed.xcnt;
1838 : /* We only allocated/stored xcnt, not xcnt_space xids ! */
1839 : /* don't overwrite preallocated xip, if we don't have anything here */
1840 18 : if (builder->committed.xcnt > 0)
1841 : {
1842 4 : pfree(builder->committed.xip);
1843 4 : builder->committed.xcnt_space = ondisk.builder.committed.xcnt;
1844 4 : builder->committed.xip = ondisk.builder.committed.xip;
1845 : }
1846 18 : ondisk.builder.committed.xip = NULL;
1847 :
1848 : /* set catalog modifying transactions */
1849 18 : if (builder->catchange.xip)
1850 0 : pfree(builder->catchange.xip);
1851 18 : builder->catchange.xcnt = ondisk.builder.catchange.xcnt;
1852 18 : builder->catchange.xip = ondisk.builder.catchange.xip;
1853 18 : ondisk.builder.catchange.xip = NULL;
1854 :
1855 : /* our snapshot is not interesting anymore, build a new one */
1856 18 : if (builder->snapshot != NULL)
1857 : {
1858 0 : SnapBuildSnapDecRefcount(builder->snapshot);
1859 : }
1860 18 : builder->snapshot = SnapBuildBuildSnapshot(builder);
1861 18 : SnapBuildSnapIncRefcount(builder->snapshot);
1862 :
1863 18 : ReorderBufferSetRestartPoint(builder->reorder, lsn);
1864 :
1865 : Assert(builder->state == SNAPBUILD_CONSISTENT);
1866 :
1867 18 : ereport(LOG,
1868 : (errmsg("logical decoding found consistent point at %X/%X",
1869 : LSN_FORMAT_ARGS(lsn)),
1870 : errdetail("Logical decoding will begin using saved snapshot.")));
1871 18 : return true;
1872 :
1873 0 : snapshot_not_interesting:
1874 0 : if (ondisk.builder.committed.xip != NULL)
1875 0 : pfree(ondisk.builder.committed.xip);
1876 0 : if (ondisk.builder.catchange.xip != NULL)
1877 0 : pfree(ondisk.builder.catchange.xip);
1878 0 : return false;
1879 : }
1880 :
1881 : /*
1882 : * Read the contents of the serialized snapshot to 'dest'.
1883 : */
1884 : static void
1885 72 : SnapBuildRestoreContents(int fd, char *dest, Size size, const char *path)
1886 : {
1887 : int readBytes;
1888 :
1889 72 : pgstat_report_wait_start(WAIT_EVENT_SNAPBUILD_READ);
1890 72 : readBytes = read(fd, dest, size);
1891 72 : pgstat_report_wait_end();
1892 72 : if (readBytes != size)
1893 : {
1894 0 : int save_errno = errno;
1895 :
1896 0 : CloseTransientFile(fd);
1897 :
1898 0 : if (readBytes < 0)
1899 : {
1900 0 : errno = save_errno;
1901 0 : ereport(ERROR,
1902 : (errcode_for_file_access(),
1903 : errmsg("could not read file \"%s\": %m", path)));
1904 : }
1905 : else
1906 0 : ereport(ERROR,
1907 : (errcode(ERRCODE_DATA_CORRUPTED),
1908 : errmsg("could not read file \"%s\": read %d of %zu",
1909 : path, readBytes, size)));
1910 : }
1911 72 : }
1912 :
1913 : /*
1914 : * Remove all serialized snapshots that are not required anymore because no
1915 : * slot can need them. This doesn't actually have to run during a checkpoint,
1916 : * but it's a convenient point to schedule this.
1917 : *
1918 : * NB: We run this during checkpoints even if logical decoding is disabled so
1919 : * we cleanup old slots at some point after it got disabled.
1920 : */
1921 : void
1922 2470 : CheckPointSnapBuild(void)
1923 : {
1924 : XLogRecPtr cutoff;
1925 : XLogRecPtr redo;
1926 : DIR *snap_dir;
1927 : struct dirent *snap_de;
1928 : char path[MAXPGPATH + sizeof(PG_LOGICAL_SNAPSHOTS_DIR)];
1929 :
1930 : /*
1931 : * We start off with a minimum of the last redo pointer. No new
1932 : * replication slot will start before that, so that's a safe upper bound
1933 : * for removal.
1934 : */
1935 2470 : redo = GetRedoRecPtr();
1936 :
1937 : /* now check for the restart ptrs from existing slots */
1938 2470 : cutoff = ReplicationSlotsComputeLogicalRestartLSN();
1939 :
1940 : /* don't start earlier than the restart lsn */
1941 2470 : if (redo < cutoff)
1942 2 : cutoff = redo;
1943 :
1944 2470 : snap_dir = AllocateDir(PG_LOGICAL_SNAPSHOTS_DIR);
1945 7854 : while ((snap_de = ReadDir(snap_dir, PG_LOGICAL_SNAPSHOTS_DIR)) != NULL)
1946 : {
1947 : uint32 hi;
1948 : uint32 lo;
1949 : XLogRecPtr lsn;
1950 : PGFileType de_type;
1951 :
1952 5384 : if (strcmp(snap_de->d_name, ".") == 0 ||
1953 2914 : strcmp(snap_de->d_name, "..") == 0)
1954 4940 : continue;
1955 :
1956 444 : snprintf(path, sizeof(path), "%s/%s", PG_LOGICAL_SNAPSHOTS_DIR, snap_de->d_name);
1957 444 : de_type = get_dirent_type(path, snap_de, false, DEBUG1);
1958 :
1959 444 : if (de_type != PGFILETYPE_ERROR && de_type != PGFILETYPE_REG)
1960 : {
1961 0 : elog(DEBUG1, "only regular files expected: %s", path);
1962 0 : continue;
1963 : }
1964 :
1965 : /*
1966 : * temporary filenames from SnapBuildSerialize() include the LSN and
1967 : * everything but are postfixed by .$pid.tmp. We can just remove them
1968 : * the same as other files because there can be none that are
1969 : * currently being written that are older than cutoff.
1970 : *
1971 : * We just log a message if a file doesn't fit the pattern, it's
1972 : * probably some editors lock/state file or similar...
1973 : */
1974 444 : if (sscanf(snap_de->d_name, "%X-%X.snap", &hi, &lo) != 2)
1975 : {
1976 0 : ereport(LOG,
1977 : (errmsg("could not parse file name \"%s\"", path)));
1978 0 : continue;
1979 : }
1980 :
1981 444 : lsn = ((uint64) hi) << 32 | lo;
1982 :
1983 : /* check whether we still need it */
1984 444 : if (lsn < cutoff || cutoff == InvalidXLogRecPtr)
1985 : {
1986 320 : elog(DEBUG1, "removing snapbuild snapshot %s", path);
1987 :
1988 : /*
1989 : * It's not particularly harmful, though strange, if we can't
1990 : * remove the file here. Don't prevent the checkpoint from
1991 : * completing, that'd be a cure worse than the disease.
1992 : */
1993 320 : if (unlink(path) < 0)
1994 : {
1995 0 : ereport(LOG,
1996 : (errcode_for_file_access(),
1997 : errmsg("could not remove file \"%s\": %m",
1998 : path)));
1999 0 : continue;
2000 : }
2001 : }
2002 : }
2003 2470 : FreeDir(snap_dir);
2004 2470 : }
2005 :
2006 : /*
2007 : * Check if a logical snapshot at the specified point has been serialized.
2008 : */
2009 : bool
2010 20 : SnapBuildSnapshotExists(XLogRecPtr lsn)
2011 : {
2012 : char path[MAXPGPATH];
2013 : int ret;
2014 : struct stat stat_buf;
2015 :
2016 20 : sprintf(path, "%s/%X-%X.snap",
2017 : PG_LOGICAL_SNAPSHOTS_DIR,
2018 20 : LSN_FORMAT_ARGS(lsn));
2019 :
2020 20 : ret = stat(path, &stat_buf);
2021 :
2022 20 : if (ret != 0 && errno != ENOENT)
2023 0 : ereport(ERROR,
2024 : (errcode_for_file_access(),
2025 : errmsg("could not stat file \"%s\": %m", path)));
2026 :
2027 20 : return ret == 0;
2028 : }
|