Branch data 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) 2012-2026, 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 : : #include "utils/wait_event.h"
148 : :
149 : :
150 : : /*
151 : : * Starting a transaction -- which we need to do while exporting a snapshot --
152 : : * removes knowledge about the previously used resowner, so we save it here.
153 : : */
154 : : static ResourceOwner SavedResourceOwnerDuringExport = NULL;
155 : : static bool ExportInProgress = false;
156 : :
157 : : /* ->committed and ->catchange manipulation */
158 : : static void SnapBuildPurgeOlderTxn(SnapBuild *builder);
159 : :
160 : : /* snapshot building/manipulation/distribution functions */
161 : : static Snapshot SnapBuildBuildSnapshot(SnapBuild *builder);
162 : :
163 : : static void SnapBuildFreeSnapshot(Snapshot snap);
164 : :
165 : : static void SnapBuildSnapIncRefcount(Snapshot snap);
166 : :
167 : : static void SnapBuildDistributeSnapshotAndInval(SnapBuild *builder, XLogRecPtr lsn, TransactionId xid);
168 : :
169 : : static inline bool SnapBuildXidHasCatalogChanges(SnapBuild *builder, TransactionId xid,
170 : : uint32 xinfo);
171 : :
172 : : /* xlog reading helper functions for SnapBuildProcessRunningXacts */
173 : : static bool SnapBuildFindSnapshot(SnapBuild *builder, XLogRecPtr lsn,
174 : : xl_running_xacts *running);
175 : : static void SnapBuildWaitSnapshot(xl_running_xacts *running, TransactionId cutoff);
176 : :
177 : : /* serialization functions */
178 : : static void SnapBuildSerialize(SnapBuild *builder, XLogRecPtr lsn);
179 : : static bool SnapBuildRestore(SnapBuild *builder, XLogRecPtr lsn);
180 : : static void SnapBuildRestoreContents(int fd, void *dest, Size size, const char *path);
181 : :
182 : : /*
183 : : * Allocate a new snapshot builder.
184 : : *
185 : : * xmin_horizon is the xid >= which we can be sure no catalog rows have been
186 : : * removed, start_lsn is the LSN >= we want to replay commits.
187 : : */
188 : : SnapBuild *
189 : 1230 : AllocateSnapshotBuilder(ReorderBuffer *reorder,
190 : : TransactionId xmin_horizon,
191 : : XLogRecPtr start_lsn,
192 : : bool need_full_snapshot,
193 : : bool in_slot_creation,
194 : : XLogRecPtr two_phase_at)
195 : : {
196 : : MemoryContext context;
197 : : MemoryContext oldcontext;
198 : : SnapBuild *builder;
199 : :
200 : : /* allocate memory in own context, to have better accountability */
201 : 1230 : context = AllocSetContextCreate(CurrentMemoryContext,
202 : : "snapshot builder context",
203 : : ALLOCSET_DEFAULT_SIZES);
204 : 1230 : oldcontext = MemoryContextSwitchTo(context);
205 : :
206 : 1230 : builder = palloc0_object(SnapBuild);
207 : :
208 : 1230 : builder->state = SNAPBUILD_START;
209 : 1230 : builder->context = context;
210 : 1230 : builder->reorder = reorder;
211 : : /* Other struct members initialized by zeroing via palloc0 above */
212 : :
213 : 1230 : builder->committed.xcnt = 0;
214 : 1230 : builder->committed.xcnt_space = 128; /* arbitrary number */
215 : 1230 : builder->committed.xip =
216 : 1230 : palloc0_array(TransactionId, builder->committed.xcnt_space);
217 : 1230 : builder->committed.includes_all_transactions = true;
218 : :
219 : 1230 : builder->catchange.xcnt = 0;
220 : 1230 : builder->catchange.xip = NULL;
221 : :
222 : 1230 : builder->initial_xmin_horizon = xmin_horizon;
223 : 1230 : builder->start_decoding_at = start_lsn;
224 : 1230 : builder->in_slot_creation = in_slot_creation;
225 : 1230 : builder->building_full_snapshot = need_full_snapshot;
226 : 1230 : builder->two_phase_at = two_phase_at;
227 : :
228 : 1230 : MemoryContextSwitchTo(oldcontext);
229 : :
230 : 1230 : return builder;
231 : : }
232 : :
233 : : /*
234 : : * Free a snapshot builder.
235 : : */
236 : : void
237 : 956 : FreeSnapshotBuilder(SnapBuild *builder)
238 : : {
239 : 956 : MemoryContext context = builder->context;
240 : :
241 : : /* free snapshot explicitly, that contains some error checking */
242 [ + + ]: 956 : if (builder->snapshot != NULL)
243 : : {
244 : 236 : SnapBuildSnapDecRefcount(builder->snapshot);
245 : 236 : builder->snapshot = NULL;
246 : : }
247 : :
248 : : /* other resources are deallocated via memory context reset */
249 : 956 : MemoryContextDelete(context);
250 : 956 : }
251 : :
252 : : /*
253 : : * Free an unreferenced snapshot that has previously been built by us.
254 : : */
255 : : static void
256 : 1861 : SnapBuildFreeSnapshot(Snapshot snap)
257 : : {
258 : : /* make sure we don't get passed an external snapshot */
259 : : Assert(snap->snapshot_type == SNAPSHOT_HISTORIC_MVCC);
260 : :
261 : : /* make sure nobody modified our snapshot */
262 : : Assert(snap->curcid == FirstCommandId);
263 : : Assert(!snap->suboverflowed);
264 : : Assert(!snap->takenDuringRecovery);
265 : : Assert(snap->regd_count == 0);
266 : :
267 : : /* slightly more likely, so it's checked even without c-asserts */
268 [ - + ]: 1861 : if (snap->copied)
269 [ # # ]: 0 : elog(ERROR, "cannot free a copied snapshot");
270 : :
271 [ - + ]: 1861 : if (snap->active_count)
272 [ # # ]: 0 : elog(ERROR, "cannot free an active snapshot");
273 : :
274 : 1861 : pfree(snap);
275 : 1861 : }
276 : :
277 : : /*
278 : : * In which state of snapshot building are we?
279 : : */
280 : : SnapBuildState
281 : 2380415 : SnapBuildCurrentState(SnapBuild *builder)
282 : : {
283 : 2380415 : return builder->state;
284 : : }
285 : :
286 : : /*
287 : : * Return the LSN at which the two-phase decoding was first enabled.
288 : : */
289 : : XLogRecPtr
290 : 33 : SnapBuildGetTwoPhaseAt(SnapBuild *builder)
291 : : {
292 : 33 : return builder->two_phase_at;
293 : : }
294 : :
295 : : /*
296 : : * Set the LSN at which two-phase decoding is enabled.
297 : : */
298 : : void
299 : 8 : SnapBuildSetTwoPhaseAt(SnapBuild *builder, XLogRecPtr ptr)
300 : : {
301 : 8 : builder->two_phase_at = ptr;
302 : 8 : }
303 : :
304 : : /*
305 : : * Should the contents of transaction ending at 'ptr' be decoded?
306 : : */
307 : : bool
308 : 557773 : SnapBuildXactNeedsSkip(SnapBuild *builder, XLogRecPtr ptr)
309 : : {
310 : 557773 : return ptr < builder->start_decoding_at;
311 : : }
312 : :
313 : : /*
314 : : * Increase refcount of a snapshot.
315 : : *
316 : : * This is used when handing out a snapshot to some external resource or when
317 : : * adding a Snapshot as builder->snapshot.
318 : : */
319 : : static void
320 : 8015 : SnapBuildSnapIncRefcount(Snapshot snap)
321 : : {
322 : 8015 : snap->active_count++;
323 : 8015 : }
324 : :
325 : : /*
326 : : * Decrease refcount of a snapshot and free if the refcount reaches zero.
327 : : *
328 : : * Externally visible, so that external resources that have been handed an
329 : : * IncRef'ed Snapshot can adjust its refcount easily.
330 : : */
331 : : void
332 : 7701 : SnapBuildSnapDecRefcount(Snapshot snap)
333 : : {
334 : : /* make sure we don't get passed an external snapshot */
335 : : Assert(snap->snapshot_type == SNAPSHOT_HISTORIC_MVCC);
336 : :
337 : : /* make sure nobody modified our snapshot */
338 : : Assert(snap->curcid == FirstCommandId);
339 : : Assert(!snap->suboverflowed);
340 : : Assert(!snap->takenDuringRecovery);
341 : :
342 : : Assert(snap->regd_count == 0);
343 : :
344 : : Assert(snap->active_count > 0);
345 : :
346 : : /* slightly more likely, so it's checked even without casserts */
347 [ - + ]: 7701 : if (snap->copied)
348 [ # # ]: 0 : elog(ERROR, "cannot free a copied snapshot");
349 : :
350 : 7701 : snap->active_count--;
351 [ + + ]: 7701 : if (snap->active_count == 0)
352 : 1861 : SnapBuildFreeSnapshot(snap);
353 : 7701 : }
354 : :
355 : : /*
356 : : * Build a new snapshot, based on currently committed catalog-modifying
357 : : * transactions.
358 : : *
359 : : * In-progress transactions with catalog access are *not* allowed to modify
360 : : * these snapshots; they have to copy them and fill in appropriate ->curcid
361 : : * and ->subxip/subxcnt values.
362 : : */
363 : : static Snapshot
364 : 2369 : SnapBuildBuildSnapshot(SnapBuild *builder)
365 : : {
366 : : Snapshot snapshot;
367 : : Size ssize;
368 : :
369 : : Assert(builder->state >= SNAPBUILD_FULL_SNAPSHOT);
370 : :
371 : 2369 : ssize = sizeof(SnapshotData)
372 : 2369 : + sizeof(TransactionId) * builder->committed.xcnt
373 : 2369 : + sizeof(TransactionId) * 1 /* toplevel xid */ ;
374 : :
375 : 2369 : snapshot = MemoryContextAllocZero(builder->context, ssize);
376 : :
377 : 2369 : snapshot->snapshot_type = SNAPSHOT_HISTORIC_MVCC;
378 : :
379 : : /*
380 : : * We misuse the original meaning of SnapshotData's xip and subxip fields
381 : : * to make the more fitting for our needs.
382 : : *
383 : : * In the 'xip' array we store transactions that have to be treated as
384 : : * committed. Since we will only ever look at tuples from transactions
385 : : * that have modified the catalog it's more efficient to store those few
386 : : * that exist between xmin and xmax (frequently there are none).
387 : : *
388 : : * Snapshots that are used in transactions that have modified the catalog
389 : : * also use the 'subxip' array to store their toplevel xid and all the
390 : : * subtransaction xids so we can recognize when we need to treat rows as
391 : : * visible that are not in xip but still need to be visible. Subxip only
392 : : * gets filled when the transaction is copied into the context of a
393 : : * catalog modifying transaction since we otherwise share a snapshot
394 : : * between transactions. As long as a txn hasn't modified the catalog it
395 : : * doesn't need to treat any uncommitted rows as visible, so there is no
396 : : * need for those xids.
397 : : *
398 : : * Both arrays are qsort'ed so that we can use bsearch() on them.
399 : : */
400 : : Assert(TransactionIdIsNormal(builder->xmin));
401 : : Assert(TransactionIdIsNormal(builder->xmax));
402 : :
403 : 2369 : snapshot->xmin = builder->xmin;
404 : 2369 : snapshot->xmax = builder->xmax;
405 : :
406 : : /* store all transactions to be treated as committed by this snapshot */
407 : 2369 : snapshot->xip =
408 : 2369 : (TransactionId *) ((char *) snapshot + sizeof(SnapshotData));
409 : 2369 : snapshot->xcnt = builder->committed.xcnt;
410 : 2369 : memcpy(snapshot->xip,
411 : 2369 : builder->committed.xip,
412 : 2369 : builder->committed.xcnt * sizeof(TransactionId));
413 : :
414 : : /* sort so we can bsearch() */
415 : 2369 : qsort(snapshot->xip, snapshot->xcnt, sizeof(TransactionId), xidComparator);
416 : :
417 : : /*
418 : : * Initially, subxip is empty, i.e. it's a snapshot to be used by
419 : : * transactions that don't modify the catalog. Will be filled by
420 : : * ReorderBufferCopySnap() if necessary.
421 : : */
422 : 2369 : snapshot->subxcnt = 0;
423 : 2369 : snapshot->subxip = NULL;
424 : :
425 : 2369 : snapshot->suboverflowed = false;
426 : 2369 : snapshot->takenDuringRecovery = false;
427 : 2369 : snapshot->copied = false;
428 : 2369 : snapshot->curcid = FirstCommandId;
429 : 2369 : snapshot->active_count = 0;
430 : 2369 : snapshot->regd_count = 0;
431 : 2369 : snapshot->snapXactCompletionCount = 0;
432 : :
433 : 2369 : return snapshot;
434 : : }
435 : :
436 : : /*
437 : : * Build the initial slot snapshot and convert it to a normal snapshot that
438 : : * is understood by HeapTupleSatisfiesMVCC.
439 : : *
440 : : * The snapshot will be usable directly in current transaction or exported
441 : : * for loading in different transaction.
442 : : */
443 : : Snapshot
444 : 222 : SnapBuildInitialSnapshot(SnapBuild *builder)
445 : : {
446 : : Snapshot snap;
447 : : TransactionId xid;
448 : : TransactionId safeXid;
449 : : TransactionId *newxip;
450 : 222 : int newxcnt = 0;
451 : :
452 : : Assert(XactIsoLevel == XACT_REPEATABLE_READ);
453 : : Assert(builder->building_full_snapshot);
454 : :
455 : : /* don't allow older snapshots */
456 : 222 : InvalidateCatalogSnapshot(); /* about to overwrite MyProc->xmin */
457 [ - + ]: 222 : if (HaveRegisteredOrActiveSnapshot())
458 [ # # ]: 0 : elog(ERROR, "cannot build an initial slot snapshot when snapshots exist");
459 : : Assert(!HistoricSnapshotActive());
460 : :
461 [ - + ]: 222 : if (builder->state != SNAPBUILD_CONSISTENT)
462 [ # # ]: 0 : elog(ERROR, "cannot build an initial slot snapshot before reaching a consistent state");
463 : :
464 [ - + ]: 222 : if (!builder->committed.includes_all_transactions)
465 [ # # ]: 0 : elog(ERROR, "cannot build an initial slot snapshot, not all transactions are monitored anymore");
466 : :
467 : : /* so we don't overwrite the existing value */
468 [ - + ]: 222 : if (TransactionIdIsValid(MyProc->xmin))
469 [ # # ]: 0 : elog(ERROR, "cannot build an initial slot snapshot when MyProc->xmin already is valid");
470 : :
471 : 222 : snap = SnapBuildBuildSnapshot(builder);
472 : :
473 : : /*
474 : : * We know that snap->xmin is alive, enforced by the logical xmin
475 : : * mechanism. Due to that we can do this without locks, we're only
476 : : * changing our own value.
477 : : *
478 : : * Building an initial snapshot is expensive and an unenforced xmin
479 : : * horizon would have bad consequences, therefore always double-check that
480 : : * the horizon is enforced.
481 : : */
482 : 222 : LWLockAcquire(ProcArrayLock, LW_SHARED);
483 : 222 : safeXid = GetOldestSafeDecodingTransactionId(false);
484 : 222 : LWLockRelease(ProcArrayLock);
485 : :
486 [ - + ]: 222 : if (TransactionIdFollows(safeXid, snap->xmin))
487 [ # # ]: 0 : elog(ERROR, "cannot build an initial slot snapshot as oldest safe xid %u follows snapshot's xmin %u",
488 : : safeXid, snap->xmin);
489 : :
490 : 222 : MyProc->xmin = snap->xmin;
491 : :
492 : : /* allocate in transaction context */
493 : 222 : newxip = palloc_array(TransactionId, GetMaxSnapshotXidCount());
494 : :
495 : : /*
496 : : * snapbuild.c builds transactions in an "inverted" manner, which means it
497 : : * stores committed transactions in ->xip, not ones in progress. Build a
498 : : * classical snapshot by marking all non-committed transactions as
499 : : * in-progress. This can be expensive.
500 : : */
501 [ - + ]: 222 : for (xid = snap->xmin; NormalTransactionIdPrecedes(xid, snap->xmax);)
502 : : {
503 : : void *test;
504 : :
505 : : /*
506 : : * Check whether transaction committed using the decoding snapshot
507 : : * meaning of ->xip.
508 : : */
509 : 0 : test = bsearch(&xid, snap->xip, snap->xcnt,
510 : : sizeof(TransactionId), xidComparator);
511 : :
512 [ # # ]: 0 : if (test == NULL)
513 : : {
514 [ # # ]: 0 : if (newxcnt >= GetMaxSnapshotXidCount())
515 [ # # ]: 0 : ereport(ERROR,
516 : : (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
517 : : errmsg("initial slot snapshot too large")));
518 : :
519 : 0 : newxip[newxcnt++] = xid;
520 : : }
521 : :
522 [ # # ]: 0 : TransactionIdAdvance(xid);
523 : : }
524 : :
525 : : /* adjust remaining snapshot fields as needed */
526 : 222 : snap->snapshot_type = SNAPSHOT_MVCC;
527 : 222 : snap->xcnt = newxcnt;
528 : 222 : snap->xip = newxip;
529 : :
530 : 222 : return snap;
531 : : }
532 : :
533 : : /*
534 : : * Export a snapshot so it can be set in another session with SET TRANSACTION
535 : : * SNAPSHOT.
536 : : *
537 : : * For that we need to start a transaction in the current backend as the
538 : : * importing side checks whether the source transaction is still open to make
539 : : * sure the xmin horizon hasn't advanced since then.
540 : : */
541 : : const char *
542 : 1 : SnapBuildExportSnapshot(SnapBuild *builder)
543 : : {
544 : : Snapshot snap;
545 : : char *snapname;
546 : :
547 [ - + ]: 1 : if (IsTransactionOrTransactionBlock())
548 [ # # ]: 0 : elog(ERROR, "cannot export a snapshot from within a transaction");
549 : :
550 [ - + ]: 1 : if (SavedResourceOwnerDuringExport)
551 [ # # ]: 0 : elog(ERROR, "can only export one snapshot at a time");
552 : :
553 : 1 : SavedResourceOwnerDuringExport = CurrentResourceOwner;
554 : 1 : ExportInProgress = true;
555 : :
556 : 1 : StartTransactionCommand();
557 : :
558 : : /* There doesn't seem to a nice API to set these */
559 : 1 : XactIsoLevel = XACT_REPEATABLE_READ;
560 : 1 : XactReadOnly = true;
561 : :
562 : 1 : snap = SnapBuildInitialSnapshot(builder);
563 : :
564 : : /*
565 : : * now that we've built a plain snapshot, make it active and use the
566 : : * normal mechanisms for exporting it
567 : : */
568 : 1 : snapname = ExportSnapshot(snap);
569 : :
570 [ + - ]: 1 : ereport(LOG,
571 : : (errmsg_plural("exported logical decoding snapshot: \"%s\" with %u transaction ID",
572 : : "exported logical decoding snapshot: \"%s\" with %u transaction IDs",
573 : : snap->xcnt,
574 : : snapname, snap->xcnt)));
575 : 1 : return snapname;
576 : : }
577 : :
578 : : /*
579 : : * Ensure there is a snapshot and if not build one for current transaction.
580 : : */
581 : : Snapshot
582 : 9 : SnapBuildGetOrBuildSnapshot(SnapBuild *builder)
583 : : {
584 : : Assert(builder->state == SNAPBUILD_CONSISTENT);
585 : :
586 : : /* only build a new snapshot if we don't have a prebuilt one */
587 [ + + ]: 9 : if (builder->snapshot == NULL)
588 : : {
589 : 2 : builder->snapshot = SnapBuildBuildSnapshot(builder);
590 : : /* increase refcount for the snapshot builder */
591 : 2 : SnapBuildSnapIncRefcount(builder->snapshot);
592 : : }
593 : :
594 : 9 : return builder->snapshot;
595 : : }
596 : :
597 : : /*
598 : : * Reset a previously SnapBuildExportSnapshot()'ed snapshot if there is
599 : : * any. Aborts the previously started transaction and resets the resource
600 : : * owner back to its original value.
601 : : */
602 : : void
603 : 5978 : SnapBuildClearExportedSnapshot(void)
604 : : {
605 : : ResourceOwner tmpResOwner;
606 : :
607 : : /* nothing exported, that is the usual case */
608 [ + + ]: 5978 : if (!ExportInProgress)
609 : 5977 : return;
610 : :
611 [ - + ]: 1 : if (!IsTransactionState())
612 [ # # ]: 0 : elog(ERROR, "clearing exported snapshot in wrong transaction state");
613 : :
614 : : /*
615 : : * AbortCurrentTransaction() takes care of resetting the snapshot state,
616 : : * so remember SavedResourceOwnerDuringExport.
617 : : */
618 : 1 : tmpResOwner = SavedResourceOwnerDuringExport;
619 : :
620 : : /* make sure nothing could have ever happened */
621 : 1 : AbortCurrentTransaction();
622 : :
623 : 1 : CurrentResourceOwner = tmpResOwner;
624 : : }
625 : :
626 : : /*
627 : : * Clear snapshot export state during transaction abort.
628 : : */
629 : : void
630 : 35946 : SnapBuildResetExportedSnapshotState(void)
631 : : {
632 : 35946 : SavedResourceOwnerDuringExport = NULL;
633 : 35946 : ExportInProgress = false;
634 : 35946 : }
635 : :
636 : : /*
637 : : * Handle the effects of a single heap change, appropriate to the current state
638 : : * of the snapshot builder and returns whether changes made at (xid, lsn) can
639 : : * be decoded.
640 : : */
641 : : bool
642 : 1589187 : SnapBuildProcessChange(SnapBuild *builder, TransactionId xid, XLogRecPtr lsn)
643 : : {
644 : : /*
645 : : * We can't handle data in transactions if we haven't built a snapshot
646 : : * yet, so don't store them.
647 : : */
648 [ - + ]: 1589187 : if (builder->state < SNAPBUILD_FULL_SNAPSHOT)
649 : 0 : return false;
650 : :
651 : : /*
652 : : * No point in keeping track of changes in transactions that we don't have
653 : : * enough information about to decode. This means that they started before
654 : : * we got into the SNAPBUILD_FULL_SNAPSHOT state.
655 : : */
656 [ + + - + ]: 1589190 : if (builder->state < SNAPBUILD_CONSISTENT &&
657 : 3 : TransactionIdPrecedes(xid, builder->next_phase_at))
658 : 0 : return false;
659 : :
660 : : /*
661 : : * If the reorderbuffer doesn't yet have a snapshot, add one now, it will
662 : : * be needed to decode the change we're currently processing.
663 : : */
664 [ + + ]: 1589187 : if (!ReorderBufferXidHasBaseSnapshot(builder->reorder, xid))
665 : : {
666 : : /* only build a new snapshot if we don't have a prebuilt one */
667 [ + + ]: 4185 : if (builder->snapshot == NULL)
668 : : {
669 : 466 : builder->snapshot = SnapBuildBuildSnapshot(builder);
670 : : /* increase refcount for the snapshot builder */
671 : 466 : SnapBuildSnapIncRefcount(builder->snapshot);
672 : : }
673 : :
674 : : /*
675 : : * Increase refcount for the transaction we're handing the snapshot
676 : : * out to.
677 : : */
678 : 4185 : SnapBuildSnapIncRefcount(builder->snapshot);
679 : 4185 : ReorderBufferSetBaseSnapshot(builder->reorder, xid, lsn,
680 : : builder->snapshot);
681 : : }
682 : :
683 : 1589187 : return true;
684 : : }
685 : :
686 : : /*
687 : : * Do CommandId/combo CID handling after reading an xl_heap_new_cid record.
688 : : * This implies that a transaction has done some form of write to system
689 : : * catalogs.
690 : : */
691 : : void
692 : 28117 : SnapBuildProcessNewCid(SnapBuild *builder, TransactionId xid,
693 : : XLogRecPtr lsn, xl_heap_new_cid *xlrec)
694 : : {
695 : : CommandId cid;
696 : :
697 : : /*
698 : : * we only log new_cid's if a catalog tuple was modified, so mark the
699 : : * transaction as containing catalog modifications
700 : : */
701 : 28117 : ReorderBufferXidSetCatalogChanges(builder->reorder, xid, lsn);
702 : :
703 : 28117 : ReorderBufferAddNewTupleCids(builder->reorder, xlrec->top_xid, lsn,
704 : : xlrec->target_locator, xlrec->target_tid,
705 : : xlrec->cmin, xlrec->cmax,
706 : : xlrec->combocid);
707 : :
708 : : /* figure out new command id */
709 [ + + ]: 28117 : if (xlrec->cmin != InvalidCommandId &&
710 [ + + ]: 23623 : xlrec->cmax != InvalidCommandId)
711 : 3320 : cid = Max(xlrec->cmin, xlrec->cmax);
712 [ + + ]: 24797 : else if (xlrec->cmax != InvalidCommandId)
713 : 4494 : cid = xlrec->cmax;
714 [ + - ]: 20303 : else if (xlrec->cmin != InvalidCommandId)
715 : 20303 : cid = xlrec->cmin;
716 : : else
717 : : {
718 : 0 : cid = InvalidCommandId; /* silence compiler */
719 [ # # ]: 0 : elog(ERROR, "xl_heap_new_cid record without a valid CommandId");
720 : : }
721 : :
722 : 28117 : ReorderBufferAddNewCommandId(builder->reorder, xid, lsn, cid + 1);
723 : 28117 : }
724 : :
725 : : /*
726 : : * Add a new Snapshot and invalidation messages to all transactions we're
727 : : * decoding that currently are in-progress so they can see new catalog contents
728 : : * made by the transaction that just committed. This is necessary because those
729 : : * in-progress transactions will use the new catalog's contents from here on
730 : : * (at the very least everything they do needs to be compatible with newer
731 : : * catalog contents).
732 : : */
733 : : static void
734 : 1671 : SnapBuildDistributeSnapshotAndInval(SnapBuild *builder, XLogRecPtr lsn, TransactionId xid)
735 : : {
736 : : dlist_iter txn_i;
737 : : ReorderBufferTXN *txn;
738 : :
739 : : /*
740 : : * Iterate through all toplevel transactions. This can include
741 : : * subtransactions which we just don't yet know to be that, but that's
742 : : * fine, they will just get an unnecessary snapshot and invalidations
743 : : * queued.
744 : : */
745 [ + - + + ]: 3376 : dlist_foreach(txn_i, &builder->reorder->toplevel_by_lsn)
746 : : {
747 : 1705 : txn = dlist_container(ReorderBufferTXN, node, txn_i.cur);
748 : :
749 : : Assert(TransactionIdIsValid(txn->xid));
750 : :
751 : : /*
752 : : * If we don't have a base snapshot yet, there are no changes in this
753 : : * transaction which in turn implies we don't yet need a snapshot at
754 : : * all. We'll add a snapshot when the first change gets queued.
755 : : *
756 : : * Similarly, we don't need to add invalidations to a transaction
757 : : * whose base snapshot is not yet set. Once a base snapshot is built,
758 : : * it will include the xids of committed transactions that have
759 : : * modified the catalog, thus reflecting the new catalog contents. The
760 : : * existing catalog cache will have already been invalidated after
761 : : * processing the invalidations in the transaction that modified
762 : : * catalogs, ensuring that a fresh cache is constructed during
763 : : * decoding.
764 : : *
765 : : * NB: This works correctly even for subtransactions because
766 : : * ReorderBufferAssignChild() takes care to transfer the base snapshot
767 : : * to the top-level transaction, and while iterating the changequeue
768 : : * we'll get the change from the subtxn.
769 : : */
770 [ + + ]: 1705 : if (!ReorderBufferXidHasBaseSnapshot(builder->reorder, txn->xid))
771 : 2 : continue;
772 : :
773 : : /*
774 : : * We don't need to add snapshot or invalidations to prepared
775 : : * transactions as they should not see the new catalog contents.
776 : : */
777 [ + + ]: 1703 : if (rbtxn_is_prepared(txn))
778 : 28 : continue;
779 : :
780 [ + + ]: 1675 : elog(DEBUG2, "adding a new snapshot and invalidations to %u at %X/%08X",
781 : : txn->xid, LSN_FORMAT_ARGS(lsn));
782 : :
783 : : /*
784 : : * increase the snapshot's refcount for the transaction we are handing
785 : : * it out to
786 : : */
787 : 1675 : SnapBuildSnapIncRefcount(builder->snapshot);
788 : 1675 : ReorderBufferAddSnapshot(builder->reorder, txn->xid, lsn,
789 : : builder->snapshot);
790 : :
791 : : /*
792 : : * Add invalidation messages to the reorder buffer of in-progress
793 : : * transactions except the current committed transaction, for which we
794 : : * will execute invalidations at the end.
795 : : *
796 : : * It is required, otherwise, we will end up using the stale catcache
797 : : * contents built by the current transaction even after its decoding,
798 : : * which should have been invalidated due to concurrent catalog
799 : : * changing transaction.
800 : : *
801 : : * Distribute only the invalidation messages generated by the current
802 : : * committed transaction. Invalidation messages received from other
803 : : * transactions would have already been propagated to the relevant
804 : : * in-progress transactions. This transaction would have processed
805 : : * those invalidations, ensuring that subsequent transactions observe
806 : : * a consistent cache state.
807 : : */
808 [ + + ]: 1675 : if (txn->xid != xid)
809 : : {
810 : : uint32 ninvalidations;
811 : 32 : SharedInvalidationMessage *msgs = NULL;
812 : :
813 : 32 : ninvalidations = ReorderBufferGetInvalidations(builder->reorder,
814 : : xid, &msgs);
815 : :
816 [ + + ]: 32 : if (ninvalidations > 0)
817 : : {
818 : : Assert(msgs != NULL);
819 : :
820 : 28 : ReorderBufferAddDistributedInvalidations(builder->reorder,
821 : : txn->xid, lsn,
822 : : ninvalidations, msgs);
823 : : }
824 : : }
825 : : }
826 : 1671 : }
827 : :
828 : : /*
829 : : * Keep track of a new catalog changing transaction that has committed.
830 : : */
831 : : static void
832 : 1680 : SnapBuildAddCommittedTxn(SnapBuild *builder, TransactionId xid)
833 : : {
834 : : Assert(TransactionIdIsValid(xid));
835 : :
836 [ - + ]: 1680 : if (builder->committed.xcnt == builder->committed.xcnt_space)
837 : : {
838 : 0 : builder->committed.xcnt_space = builder->committed.xcnt_space * 2 + 1;
839 : :
840 [ # # ]: 0 : elog(DEBUG1, "increasing space for committed transactions to %u",
841 : : (uint32) builder->committed.xcnt_space);
842 : :
843 : 0 : builder->committed.xip = repalloc_array(builder->committed.xip,
844 : : TransactionId,
845 : : builder->committed.xcnt_space);
846 : : }
847 : :
848 : : /*
849 : : * TODO: It might make sense to keep the array sorted here instead of
850 : : * doing it every time we build a new snapshot. On the other hand this
851 : : * gets called repeatedly when a transaction with subtransactions commits.
852 : : */
853 : 1680 : builder->committed.xip[builder->committed.xcnt++] = xid;
854 : 1680 : }
855 : :
856 : : /*
857 : : * Remove knowledge about transactions we treat as committed or containing catalog
858 : : * changes that are smaller than ->xmin. Those won't ever get checked via
859 : : * the ->committed or ->catchange array, respectively. The committed xids will
860 : : * get checked via the clog machinery.
861 : : *
862 : : * We can ideally remove the transaction from catchange array once it is
863 : : * finished (committed/aborted) but that could be costly as we need to maintain
864 : : * the xids order in the array.
865 : : */
866 : : static void
867 : 537 : SnapBuildPurgeOlderTxn(SnapBuild *builder)
868 : : {
869 : : TransactionId *workspace;
870 : 537 : int surviving_xids = 0;
871 : :
872 : : /* not ready yet */
873 [ - + ]: 537 : if (!TransactionIdIsNormal(builder->xmin))
874 : 0 : return;
875 : :
876 : : /* TODO: Neater algorithm than just copying and iterating? */
877 : : workspace =
878 : 537 : MemoryContextAlloc(builder->context,
879 : 537 : builder->committed.xcnt * sizeof(TransactionId));
880 : :
881 : : /* copy xids that still are interesting to workspace */
882 [ + + ]: 923 : for (size_t off = 0; off < builder->committed.xcnt; off++)
883 : : {
884 [ + + ]: 386 : if (NormalTransactionIdPrecedes(builder->committed.xip[off],
885 : : builder->xmin))
886 : : ; /* remove */
887 : : else
888 : 1 : workspace[surviving_xids++] = builder->committed.xip[off];
889 : : }
890 : :
891 : : /* copy workspace back to persistent state */
892 : 537 : memcpy(builder->committed.xip, workspace,
893 : : surviving_xids * sizeof(TransactionId));
894 : :
895 [ - + ]: 537 : elog(DEBUG3, "purged committed transactions from %u to %u, xmin: %u, xmax: %u",
896 : : (uint32) builder->committed.xcnt, (uint32) surviving_xids,
897 : : builder->xmin, builder->xmax);
898 : 537 : builder->committed.xcnt = surviving_xids;
899 : :
900 : 537 : pfree(workspace);
901 : :
902 : : /*
903 : : * Purge xids in ->catchange as well. The purged array must also be sorted
904 : : * in xidComparator order.
905 : : */
906 [ + + ]: 537 : if (builder->catchange.xcnt > 0)
907 : : {
908 : : size_t off;
909 : :
910 : : /*
911 : : * Since catchange.xip is sorted, we find the lower bound of xids that
912 : : * are still interesting.
913 : : */
914 [ + + ]: 7 : for (off = 0; off < builder->catchange.xcnt; off++)
915 : : {
916 [ + + ]: 5 : if (TransactionIdFollowsOrEquals(builder->catchange.xip[off],
917 : : builder->xmin))
918 : 1 : break;
919 : : }
920 : :
921 : 3 : surviving_xids = builder->catchange.xcnt - off;
922 : :
923 [ + + ]: 3 : if (surviving_xids > 0)
924 : : {
925 : 1 : memmove(builder->catchange.xip, &(builder->catchange.xip[off]),
926 : : surviving_xids * sizeof(TransactionId));
927 : : }
928 : : else
929 : : {
930 : 2 : pfree(builder->catchange.xip);
931 : 2 : builder->catchange.xip = NULL;
932 : : }
933 : :
934 [ - + ]: 3 : elog(DEBUG3, "purged catalog modifying transactions from %u to %u, xmin: %u, xmax: %u",
935 : : (uint32) builder->catchange.xcnt, (uint32) surviving_xids,
936 : : builder->xmin, builder->xmax);
937 : 3 : builder->catchange.xcnt = surviving_xids;
938 : : }
939 : : }
940 : :
941 : : /*
942 : : * Handle everything that needs to be done when a transaction commits
943 : : */
944 : : void
945 : 4027 : SnapBuildCommitTxn(SnapBuild *builder, XLogRecPtr lsn, TransactionId xid,
946 : : int nsubxacts, TransactionId *subxacts, uint32 xinfo)
947 : : {
948 : : int nxact;
949 : :
950 : 4027 : bool needs_snapshot = false;
951 : 4027 : bool needs_timetravel = false;
952 : 4027 : bool sub_needs_timetravel = false;
953 : :
954 : 4027 : TransactionId xmax = xid;
955 : :
956 : : /*
957 : : * Transactions preceding BUILDING_SNAPSHOT will neither be decoded, nor
958 : : * will they be part of a snapshot. So we don't need to record anything.
959 : : */
960 [ + - ]: 4027 : if (builder->state == SNAPBUILD_START ||
961 [ - + - - ]: 4027 : (builder->state == SNAPBUILD_BUILDING_SNAPSHOT &&
962 : 0 : TransactionIdPrecedes(xid, builder->next_phase_at)))
963 : : {
964 : : /* ensure that only commits after this are getting replayed */
965 [ # # ]: 0 : if (builder->start_decoding_at <= lsn)
966 : 0 : builder->start_decoding_at = lsn + 1;
967 : 0 : return;
968 : : }
969 : :
970 [ + + ]: 4027 : if (builder->state < SNAPBUILD_CONSISTENT)
971 : : {
972 : : /* ensure that only commits after this are getting replayed */
973 [ + + ]: 5 : if (builder->start_decoding_at <= lsn)
974 : 2 : builder->start_decoding_at = lsn + 1;
975 : :
976 : : /*
977 : : * If building an exportable snapshot, force xid to be tracked, even
978 : : * if the transaction didn't modify the catalog.
979 : : */
980 [ - + ]: 5 : if (builder->building_full_snapshot)
981 : : {
982 : 0 : needs_timetravel = true;
983 : : }
984 : : }
985 : :
986 [ + + ]: 5251 : for (nxact = 0; nxact < nsubxacts; nxact++)
987 : : {
988 : 1224 : TransactionId subxid = subxacts[nxact];
989 : :
990 : : /*
991 : : * Add subtransaction to base snapshot if catalog modifying, we don't
992 : : * distinguish to toplevel transactions there.
993 : : */
994 [ + + ]: 1224 : if (SnapBuildXidHasCatalogChanges(builder, subxid, xinfo))
995 : : {
996 : 9 : sub_needs_timetravel = true;
997 : 9 : needs_snapshot = true;
998 : :
999 [ - + ]: 9 : elog(DEBUG1, "found subtransaction %u:%u with catalog changes",
1000 : : xid, subxid);
1001 : :
1002 : 9 : SnapBuildAddCommittedTxn(builder, subxid);
1003 : :
1004 [ + - ]: 9 : if (NormalTransactionIdFollows(subxid, xmax))
1005 : 9 : xmax = subxid;
1006 : : }
1007 : :
1008 : : /*
1009 : : * If we're forcing timetravel we also need visibility information
1010 : : * about subtransaction, so keep track of subtransaction's state, even
1011 : : * if not catalog modifying. Don't need to distribute a snapshot in
1012 : : * that case.
1013 : : */
1014 [ - + ]: 1215 : else if (needs_timetravel)
1015 : : {
1016 : 0 : SnapBuildAddCommittedTxn(builder, subxid);
1017 [ # # ]: 0 : if (NormalTransactionIdFollows(subxid, xmax))
1018 : 0 : xmax = subxid;
1019 : : }
1020 : : }
1021 : :
1022 : : /* if top-level modified catalog, it'll need a snapshot */
1023 [ + + ]: 4027 : if (SnapBuildXidHasCatalogChanges(builder, xid, xinfo))
1024 : : {
1025 [ + + ]: 1670 : elog(DEBUG2, "found top level transaction %u, with catalog changes",
1026 : : xid);
1027 : 1670 : needs_snapshot = true;
1028 : 1670 : needs_timetravel = true;
1029 : 1670 : SnapBuildAddCommittedTxn(builder, xid);
1030 : : }
1031 [ + + ]: 2357 : else if (sub_needs_timetravel)
1032 : : {
1033 : : /* track toplevel txn as well, subxact alone isn't meaningful */
1034 [ - + ]: 1 : elog(DEBUG2, "forced transaction %u to do timetravel due to one of its subtransactions",
1035 : : xid);
1036 : 1 : needs_timetravel = true;
1037 : 1 : SnapBuildAddCommittedTxn(builder, xid);
1038 : : }
1039 [ - + ]: 2356 : else if (needs_timetravel)
1040 : : {
1041 [ # # ]: 0 : elog(DEBUG2, "forced transaction %u to do timetravel", xid);
1042 : :
1043 : 0 : SnapBuildAddCommittedTxn(builder, xid);
1044 : : }
1045 : :
1046 [ + + ]: 4027 : if (!needs_timetravel)
1047 : : {
1048 : : /* record that we cannot export a general snapshot anymore */
1049 : 2356 : builder->committed.includes_all_transactions = false;
1050 : : }
1051 : :
1052 : : Assert(!needs_snapshot || needs_timetravel);
1053 : :
1054 : : /*
1055 : : * Adjust xmax of the snapshot builder, we only do that for committed,
1056 : : * catalog modifying, transactions, everything else isn't interesting for
1057 : : * us since we'll never look at the respective rows.
1058 : : */
1059 [ + + ]: 4027 : if (needs_timetravel &&
1060 [ + - + - ]: 3342 : (!TransactionIdIsValid(builder->xmax) ||
1061 : 1671 : TransactionIdFollowsOrEquals(xmax, builder->xmax)))
1062 : : {
1063 : 1671 : builder->xmax = xmax;
1064 [ - + ]: 1671 : TransactionIdAdvance(builder->xmax);
1065 : : }
1066 : :
1067 : : /* if there's any reason to build a historic snapshot, do so now */
1068 [ + + ]: 4027 : if (needs_snapshot)
1069 : : {
1070 : : /*
1071 : : * If we haven't built a complete snapshot yet there's no need to hand
1072 : : * it out, it wouldn't (and couldn't) be used anyway.
1073 : : */
1074 [ - + ]: 1671 : if (builder->state < SNAPBUILD_FULL_SNAPSHOT)
1075 : 0 : return;
1076 : :
1077 : : /*
1078 : : * Decrease the snapshot builder's refcount of the old snapshot, note
1079 : : * that it still will be used if it has been handed out to the
1080 : : * reorderbuffer earlier.
1081 : : */
1082 [ + - ]: 1671 : if (builder->snapshot)
1083 : 1671 : SnapBuildSnapDecRefcount(builder->snapshot);
1084 : :
1085 : 1671 : builder->snapshot = SnapBuildBuildSnapshot(builder);
1086 : :
1087 : : /* we might need to execute invalidations, add snapshot */
1088 [ + + ]: 1671 : if (!ReorderBufferXidHasBaseSnapshot(builder->reorder, xid))
1089 : : {
1090 : 8 : SnapBuildSnapIncRefcount(builder->snapshot);
1091 : 8 : ReorderBufferSetBaseSnapshot(builder->reorder, xid, lsn,
1092 : : builder->snapshot);
1093 : : }
1094 : :
1095 : : /* refcount of the snapshot builder for the new snapshot */
1096 : 1671 : SnapBuildSnapIncRefcount(builder->snapshot);
1097 : :
1098 : : /*
1099 : : * Add a new catalog snapshot and invalidations messages to all
1100 : : * currently running transactions.
1101 : : */
1102 : 1671 : SnapBuildDistributeSnapshotAndInval(builder, lsn, xid);
1103 : : }
1104 : : }
1105 : :
1106 : : /*
1107 : : * Check the reorder buffer and the snapshot to see if the given transaction has
1108 : : * modified catalogs.
1109 : : */
1110 : : static inline bool
1111 : 5251 : SnapBuildXidHasCatalogChanges(SnapBuild *builder, TransactionId xid,
1112 : : uint32 xinfo)
1113 : : {
1114 [ + + ]: 5251 : if (ReorderBufferXidHasCatalogChanges(builder->reorder, xid))
1115 : 1675 : return true;
1116 : :
1117 : : /*
1118 : : * The transactions that have changed catalogs must have invalidation
1119 : : * info.
1120 : : */
1121 [ + + ]: 3576 : if (!(xinfo & XACT_XINFO_HAS_INVALS))
1122 : 3568 : return false;
1123 : :
1124 : : /* Check the catchange XID array */
1125 [ + + + - ]: 12 : return ((builder->catchange.xcnt > 0) &&
1126 : 4 : (bsearch(&xid, builder->catchange.xip, builder->catchange.xcnt,
1127 : : sizeof(TransactionId), xidComparator) != NULL));
1128 : : }
1129 : :
1130 : : /* -----------------------------------
1131 : : * Snapshot building functions dealing with xlog records
1132 : : * -----------------------------------
1133 : : */
1134 : :
1135 : : /*
1136 : : * Process a running xacts record, and use its information to first build a
1137 : : * historic snapshot and later to release resources that aren't needed
1138 : : * anymore.
1139 : : */
1140 : : void
1141 : 1715 : SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xacts *running)
1142 : : {
1143 : : ReorderBufferTXN *txn;
1144 : : TransactionId xmin;
1145 : :
1146 : : /*
1147 : : * If we're not consistent yet, inspect the record to see whether it
1148 : : * allows to get closer to being consistent. If we are consistent, dump
1149 : : * our snapshot so others or we, after a restart, can use it.
1150 : : */
1151 [ + + ]: 1715 : if (builder->state < SNAPBUILD_CONSISTENT)
1152 : : {
1153 : : /* returns false if there's no point in performing cleanup just yet */
1154 [ + + ]: 1201 : if (!SnapBuildFindSnapshot(builder, lsn, running))
1155 : 1176 : return;
1156 : : }
1157 : : else
1158 : 514 : SnapBuildSerialize(builder, lsn);
1159 : :
1160 : : /*
1161 : : * Update range of interesting xids based on the running xacts
1162 : : * information. We don't increase ->xmax using it, because once we are in
1163 : : * a consistent state we can do that ourselves and much more efficiently
1164 : : * so, because we only need to do it for catalog transactions since we
1165 : : * only ever look at those.
1166 : : *
1167 : : * NB: We only increase xmax when a catalog modifying transaction commits
1168 : : * (see SnapBuildCommitTxn). Because of this, xmax can be lower than
1169 : : * xmin, which looks odd but is correct and actually more efficient, since
1170 : : * we hit fast paths in heapam_visibility.c.
1171 : : */
1172 : 537 : builder->xmin = running->oldestRunningXid;
1173 : :
1174 : : /* Remove transactions we don't need to keep track off anymore */
1175 : 537 : SnapBuildPurgeOlderTxn(builder);
1176 : :
1177 : : /*
1178 : : * Advance the xmin limit for the current replication slot, to allow
1179 : : * vacuum to clean up the tuples this slot has been protecting.
1180 : : *
1181 : : * The reorderbuffer might have an xmin among the currently running
1182 : : * snapshots; use it if so. If not, we need only consider the snapshots
1183 : : * we'll produce later, which can't be less than the oldest running xid in
1184 : : * the record we're reading now.
1185 : : */
1186 : 537 : xmin = ReorderBufferGetOldestXmin(builder->reorder);
1187 [ + + ]: 537 : if (xmin == InvalidTransactionId)
1188 : 488 : xmin = running->oldestRunningXid;
1189 [ - + ]: 537 : elog(DEBUG3, "xmin: %u, xmax: %u, oldest running: %u, oldest xmin: %u",
1190 : : builder->xmin, builder->xmax, running->oldestRunningXid, xmin);
1191 : 537 : LogicalIncreaseXminForSlot(lsn, xmin);
1192 : :
1193 : : /*
1194 : : * Also tell the slot where we can restart decoding from. We don't want to
1195 : : * do that after every commit because changing that implies an fsync of
1196 : : * the logical slot's state file, so we only do it every time we see a
1197 : : * running xacts record.
1198 : : *
1199 : : * Do so by looking for the oldest in progress transaction (determined by
1200 : : * the first LSN of any of its relevant records). Every transaction
1201 : : * remembers the last location we stored the snapshot to disk before its
1202 : : * beginning. That point is where we can restart from.
1203 : : */
1204 : :
1205 : : /*
1206 : : * Can't know about a serialized snapshot's location if we're not
1207 : : * consistent.
1208 : : */
1209 [ + + ]: 537 : if (builder->state < SNAPBUILD_CONSISTENT)
1210 : 18 : return;
1211 : :
1212 : 519 : txn = ReorderBufferGetOldestTXN(builder->reorder);
1213 : :
1214 : : /*
1215 : : * oldest ongoing txn might have started when we didn't yet serialize
1216 : : * anything because we hadn't reached a consistent state yet.
1217 : : */
1218 [ + + + + ]: 519 : if (txn != NULL && XLogRecPtrIsValid(txn->restart_decoding_lsn))
1219 : 24 : LogicalIncreaseRestartDecodingForSlot(lsn, txn->restart_decoding_lsn);
1220 : :
1221 : : /*
1222 : : * No in-progress transaction, can reuse the last serialized snapshot if
1223 : : * we have one.
1224 : : */
1225 [ + + ]: 495 : else if (txn == NULL &&
1226 [ + + ]: 461 : XLogRecPtrIsValid(builder->reorder->current_restart_decoding_lsn) &&
1227 [ + - ]: 459 : XLogRecPtrIsValid(builder->last_serialized_snapshot))
1228 : 459 : LogicalIncreaseRestartDecodingForSlot(lsn,
1229 : : builder->last_serialized_snapshot);
1230 : : }
1231 : :
1232 : :
1233 : : /*
1234 : : * Build the start of a snapshot that's capable of decoding the catalog.
1235 : : *
1236 : : * Helper function for SnapBuildProcessRunningXacts() while we're not yet
1237 : : * consistent.
1238 : : *
1239 : : * Returns true if there is a point in performing internal maintenance/cleanup
1240 : : * using the xl_running_xacts record.
1241 : : */
1242 : : static bool
1243 : 1201 : SnapBuildFindSnapshot(SnapBuild *builder, XLogRecPtr lsn, xl_running_xacts *running)
1244 : : {
1245 : : /* ---
1246 : : * Build catalog decoding snapshot incrementally using information about
1247 : : * the currently running transactions. There are several ways to do that:
1248 : : *
1249 : : * a) There were no running transactions when the xl_running_xacts record
1250 : : * was inserted, jump to CONSISTENT immediately. We might find such a
1251 : : * state while waiting on c)'s sub-states.
1252 : : *
1253 : : * b) This (in a previous run) or another decoding slot serialized a
1254 : : * snapshot to disk that we can use. Can't use this method while finding
1255 : : * the start point for decoding changes as the restart LSN would be an
1256 : : * arbitrary LSN but we need to find the start point to extract changes
1257 : : * where we won't see the data for partial transactions. Also, we cannot
1258 : : * use this method when a slot needs a full snapshot for export or direct
1259 : : * use, as that snapshot will only contain catalog modifying transactions.
1260 : : *
1261 : : * c) First incrementally build a snapshot for catalog tuples
1262 : : * (BUILDING_SNAPSHOT), that requires all, already in-progress,
1263 : : * transactions to finish. Every transaction starting after that
1264 : : * (FULL_SNAPSHOT state), has enough information to be decoded. But
1265 : : * for older running transactions no viable snapshot exists yet, so
1266 : : * CONSISTENT will only be reached once all of those have finished.
1267 : : * ---
1268 : : */
1269 : :
1270 : : /*
1271 : : * xl_running_xacts record is older than what we can use, we might not
1272 : : * have all necessary catalog rows anymore.
1273 : : */
1274 [ + + ]: 1201 : if (TransactionIdIsNormal(builder->initial_xmin_horizon) &&
1275 [ + + ]: 512 : NormalTransactionIdPrecedes(running->oldestRunningXid,
1276 : : builder->initial_xmin_horizon))
1277 : : {
1278 [ - + ]: 1 : ereport(DEBUG1,
1279 : : errmsg_internal("skipping snapshot at %X/%08X while building logical decoding snapshot, xmin horizon too low",
1280 : : LSN_FORMAT_ARGS(lsn)),
1281 : : errdetail_internal("initial xmin horizon of %u vs the snapshot's %u",
1282 : : builder->initial_xmin_horizon, running->oldestRunningXid));
1283 : :
1284 : :
1285 : 1 : SnapBuildWaitSnapshot(running, builder->initial_xmin_horizon);
1286 : :
1287 : 1 : return true;
1288 : : }
1289 : :
1290 : : /*
1291 : : * a) No transaction were running, we can jump to consistent.
1292 : : *
1293 : : * This is not affected by races around xl_running_xacts, because we can
1294 : : * miss transaction commits, but currently not transactions starting.
1295 : : *
1296 : : * NB: We might have already started to incrementally assemble a snapshot,
1297 : : * so we need to be careful to deal with that.
1298 : : */
1299 [ + + ]: 1200 : if (running->oldestRunningXid == running->nextXid)
1300 : : {
1301 [ + + ]: 1168 : if (!XLogRecPtrIsValid(builder->start_decoding_at) ||
1302 [ + + ]: 663 : builder->start_decoding_at <= lsn)
1303 : : /* can decode everything after this */
1304 : 506 : builder->start_decoding_at = lsn + 1;
1305 : :
1306 : : /* As no transactions were running xmin/xmax can be trivially set. */
1307 : 1168 : builder->xmin = running->nextXid; /* < are finished */
1308 : 1168 : builder->xmax = running->nextXid; /* >= are running */
1309 : :
1310 : : /* so we can safely use the faster comparisons */
1311 : : Assert(TransactionIdIsNormal(builder->xmin));
1312 : : Assert(TransactionIdIsNormal(builder->xmax));
1313 : :
1314 : 1168 : builder->state = SNAPBUILD_CONSISTENT;
1315 : 1168 : builder->next_phase_at = InvalidTransactionId;
1316 : :
1317 [ + + + + ]: 1168 : ereport(LogicalDecodingLogLevel(),
1318 : : errmsg("logical decoding found consistent point at %X/%08X",
1319 : : LSN_FORMAT_ARGS(lsn)),
1320 : : errdetail("There are no running transactions."));
1321 : :
1322 : 1168 : return false;
1323 : : }
1324 : :
1325 : : /*
1326 : : * b) valid on disk state and while neither building full snapshot nor
1327 : : * creating a slot.
1328 : : */
1329 [ + + ]: 32 : else if (!builder->building_full_snapshot &&
1330 [ + + + + ]: 50 : !builder->in_slot_creation &&
1331 : 19 : SnapBuildRestore(builder, lsn))
1332 : : {
1333 : : /* there won't be any state to cleanup */
1334 : 8 : return false;
1335 : : }
1336 : :
1337 : : /*
1338 : : * c) transition from START to BUILDING_SNAPSHOT.
1339 : : *
1340 : : * In START state, and a xl_running_xacts record with running xacts is
1341 : : * encountered. In that case, switch to BUILDING_SNAPSHOT state, and
1342 : : * record xl_running_xacts->nextXid. Once all running xacts have finished
1343 : : * (i.e. they're all >= nextXid), we have a complete catalog snapshot. It
1344 : : * might look that we could use xl_running_xacts's ->xids information to
1345 : : * get there quicker, but that is problematic because transactions marked
1346 : : * as running, might already have inserted their commit record - it's
1347 : : * infeasible to change that with locking.
1348 : : */
1349 [ + + ]: 24 : else if (builder->state == SNAPBUILD_START)
1350 : : {
1351 : 13 : builder->state = SNAPBUILD_BUILDING_SNAPSHOT;
1352 : 13 : builder->next_phase_at = running->nextXid;
1353 : :
1354 : : /*
1355 : : * Start with an xmin/xmax that's correct for future, when all the
1356 : : * currently running transactions have finished. We'll update both
1357 : : * while waiting for the pending transactions to finish.
1358 : : */
1359 : 13 : builder->xmin = running->nextXid; /* < are finished */
1360 : 13 : builder->xmax = running->nextXid; /* >= are running */
1361 : :
1362 : : /* so we can safely use the faster comparisons */
1363 : : Assert(TransactionIdIsNormal(builder->xmin));
1364 : : Assert(TransactionIdIsNormal(builder->xmax));
1365 : :
1366 [ + - ]: 13 : ereport(LOG,
1367 : : errmsg("logical decoding found initial starting point at %X/%08X",
1368 : : LSN_FORMAT_ARGS(lsn)),
1369 : : errdetail("Waiting for transactions (approximately %d) older than %u to end.",
1370 : : running->xcnt, running->nextXid));
1371 : :
1372 : 13 : SnapBuildWaitSnapshot(running, running->nextXid);
1373 : : }
1374 : :
1375 : : /*
1376 : : * c) transition from BUILDING_SNAPSHOT to FULL_SNAPSHOT.
1377 : : *
1378 : : * In BUILDING_SNAPSHOT state, and this xl_running_xacts' oldestRunningXid
1379 : : * is >= than nextXid from when we switched to BUILDING_SNAPSHOT. This
1380 : : * means all transactions starting afterwards have enough information to
1381 : : * be decoded. Switch to FULL_SNAPSHOT.
1382 : : */
1383 [ + + + + ]: 17 : else if (builder->state == SNAPBUILD_BUILDING_SNAPSHOT &&
1384 : 6 : TransactionIdPrecedesOrEquals(builder->next_phase_at,
1385 : : running->oldestRunningXid))
1386 : : {
1387 : 5 : builder->state = SNAPBUILD_FULL_SNAPSHOT;
1388 : 5 : builder->next_phase_at = running->nextXid;
1389 : :
1390 [ + - ]: 5 : ereport(LOG,
1391 : : errmsg("logical decoding found initial consistent point at %X/%08X",
1392 : : LSN_FORMAT_ARGS(lsn)),
1393 : : errdetail("Waiting for transactions (approximately %d) older than %u to end.",
1394 : : running->xcnt, running->nextXid));
1395 : :
1396 : 5 : SnapBuildWaitSnapshot(running, running->nextXid);
1397 : : }
1398 : :
1399 : : /*
1400 : : * c) transition from FULL_SNAPSHOT to CONSISTENT.
1401 : : *
1402 : : * In FULL_SNAPSHOT state, and this xl_running_xacts' oldestRunningXid is
1403 : : * >= than nextXid from when we switched to FULL_SNAPSHOT. This means all
1404 : : * transactions that are currently in progress have a catalog snapshot,
1405 : : * and all their changes have been collected. Switch to CONSISTENT.
1406 : : */
1407 [ + + + - ]: 11 : else if (builder->state == SNAPBUILD_FULL_SNAPSHOT &&
1408 : 5 : TransactionIdPrecedesOrEquals(builder->next_phase_at,
1409 : : running->oldestRunningXid))
1410 : : {
1411 : 5 : builder->state = SNAPBUILD_CONSISTENT;
1412 : 5 : builder->next_phase_at = InvalidTransactionId;
1413 : :
1414 [ + - - + ]: 5 : ereport(LogicalDecodingLogLevel(),
1415 : : errmsg("logical decoding found consistent point at %X/%08X",
1416 : : LSN_FORMAT_ARGS(lsn)),
1417 : : errdetail("There are no old transactions anymore."));
1418 : : }
1419 : :
1420 : : /*
1421 : : * We already started to track running xacts and need to wait for all
1422 : : * in-progress ones to finish. We fall through to the normal processing of
1423 : : * records so incremental cleanup can be performed.
1424 : : */
1425 : 22 : return true;
1426 : : }
1427 : :
1428 : : /* ---
1429 : : * Iterate through xids in record, wait for all older than the cutoff to
1430 : : * finish. Then, if possible, log a new xl_running_xacts record.
1431 : : *
1432 : : * This isn't required for the correctness of decoding, but to:
1433 : : * a) allow isolationtester to notice that we're currently waiting for
1434 : : * something.
1435 : : * b) log a new xl_running_xacts record where it'd be helpful, without having
1436 : : * to wait for bgwriter or checkpointer.
1437 : : * ---
1438 : : */
1439 : : static void
1440 : 19 : SnapBuildWaitSnapshot(xl_running_xacts *running, TransactionId cutoff)
1441 : : {
1442 : : int off;
1443 : :
1444 [ + + ]: 35 : for (off = 0; off < running->xcnt; off++)
1445 : : {
1446 : 18 : TransactionId xid = running->xids[off];
1447 : :
1448 : : /*
1449 : : * Upper layers should prevent that we ever need to wait on ourselves.
1450 : : * Check anyway, since failing to do so would either result in an
1451 : : * endless wait or an Assert() failure.
1452 : : */
1453 [ - + ]: 18 : if (TransactionIdIsCurrentTransactionId(xid))
1454 [ # # ]: 0 : elog(ERROR, "waiting for ourselves");
1455 : :
1456 [ - + ]: 18 : if (TransactionIdFollows(xid, cutoff))
1457 : 0 : continue;
1458 : :
1459 : 18 : XactLockTableWait(xid, NULL, NULL, XLTW_None);
1460 : : }
1461 : :
1462 : : /*
1463 : : * All transactions we needed to finish finished - try to ensure there is
1464 : : * another xl_running_xacts record in a timely manner, without having to
1465 : : * wait for bgwriter or checkpointer to log one. During recovery we can't
1466 : : * enforce that, so we'll have to wait.
1467 : : */
1468 [ + - ]: 17 : if (!RecoveryInProgress())
1469 : : {
1470 : 17 : LogStandbySnapshot();
1471 : : }
1472 : 17 : }
1473 : :
1474 : : #define SnapBuildOnDiskConstantSize \
1475 : : offsetof(SnapBuildOnDisk, builder)
1476 : : #define SnapBuildOnDiskNotChecksummedSize \
1477 : : offsetof(SnapBuildOnDisk, version)
1478 : :
1479 : : #define SNAPBUILD_MAGIC 0x51A1E001
1480 : : #define SNAPBUILD_VERSION 6
1481 : :
1482 : : /*
1483 : : * Store/Load a snapshot from disk, depending on the snapshot builder's state.
1484 : : *
1485 : : * Supposed to be used by external (i.e. not snapbuild.c) code that just read
1486 : : * a record that's a potential location for a serialized snapshot.
1487 : : */
1488 : : void
1489 : 91 : SnapBuildSerializationPoint(SnapBuild *builder, XLogRecPtr lsn)
1490 : : {
1491 [ - + ]: 91 : if (builder->state < SNAPBUILD_CONSISTENT)
1492 : 0 : SnapBuildRestore(builder, lsn);
1493 : : else
1494 : 91 : SnapBuildSerialize(builder, lsn);
1495 : 91 : }
1496 : :
1497 : : /*
1498 : : * Serialize the snapshot 'builder' at the location 'lsn' if it hasn't already
1499 : : * been done by another decoding process.
1500 : : */
1501 : : static void
1502 : 605 : SnapBuildSerialize(SnapBuild *builder, XLogRecPtr lsn)
1503 : : {
1504 : : Size needed_length;
1505 : 605 : SnapBuildOnDisk *ondisk = NULL;
1506 : 605 : TransactionId *catchange_xip = NULL;
1507 : : MemoryContext old_ctx;
1508 : : size_t catchange_xcnt;
1509 : : char *ondisk_c;
1510 : : int fd;
1511 : : char tmppath[MAXPGPATH];
1512 : : char path[MAXPGPATH];
1513 : : int ret;
1514 : : struct stat stat_buf;
1515 : : Size sz;
1516 : :
1517 : : Assert(XLogRecPtrIsValid(lsn));
1518 : : Assert(!XLogRecPtrIsValid(builder->last_serialized_snapshot) ||
1519 : : builder->last_serialized_snapshot <= lsn);
1520 : :
1521 : : /*
1522 : : * no point in serializing if we cannot continue to work immediately after
1523 : : * restoring the snapshot
1524 : : */
1525 [ - + ]: 605 : if (builder->state < SNAPBUILD_CONSISTENT)
1526 : 0 : return;
1527 : :
1528 : : /* consistent snapshots have no next phase */
1529 : : Assert(builder->next_phase_at == InvalidTransactionId);
1530 : :
1531 : : /*
1532 : : * We identify snapshots by the LSN they are valid for. We don't need to
1533 : : * include timelines in the name as each LSN maps to exactly one timeline
1534 : : * unless the user used pg_resetwal or similar. If a user did so, there's
1535 : : * no hope continuing to decode anyway.
1536 : : */
1537 : 605 : sprintf(path, "%s/%X-%X.snap",
1538 : : PG_LOGICAL_SNAPSHOTS_DIR,
1539 : 605 : LSN_FORMAT_ARGS(lsn));
1540 : :
1541 : : /*
1542 : : * first check whether some other backend already has written the snapshot
1543 : : * for this LSN. It's perfectly fine if there's none, so we accept ENOENT
1544 : : * as a valid state. Everything else is an unexpected error.
1545 : : */
1546 : 605 : ret = stat(path, &stat_buf);
1547 : :
1548 [ + + - + ]: 605 : if (ret != 0 && errno != ENOENT)
1549 [ # # ]: 0 : ereport(ERROR,
1550 : : (errcode_for_file_access(),
1551 : : errmsg("could not stat file \"%s\": %m", path)));
1552 : :
1553 [ + + ]: 605 : else if (ret == 0)
1554 : : {
1555 : : /*
1556 : : * somebody else has already serialized to this point, don't overwrite
1557 : : * but remember location, so we don't need to read old data again.
1558 : : *
1559 : : * To be sure it has been synced to disk after the rename() from the
1560 : : * tempfile filename to the real filename, we just repeat the fsync.
1561 : : * That ought to be cheap because in most scenarios it should already
1562 : : * be safely on disk.
1563 : : */
1564 : 269 : fsync_fname(path, false);
1565 : 269 : fsync_fname(PG_LOGICAL_SNAPSHOTS_DIR, true);
1566 : :
1567 : 269 : builder->last_serialized_snapshot = lsn;
1568 : 269 : goto out;
1569 : : }
1570 : :
1571 : : /*
1572 : : * there is an obvious race condition here between the time we stat(2) the
1573 : : * file and us writing the file. But we rename the file into place
1574 : : * atomically and all files created need to contain the same data anyway,
1575 : : * so this is perfectly fine, although a bit of a resource waste. Locking
1576 : : * seems like pointless complication.
1577 : : */
1578 [ + + ]: 336 : elog(DEBUG1, "serializing snapshot to %s", path);
1579 : :
1580 : : /* to make sure only we will write to this tempfile, include pid */
1581 : 336 : sprintf(tmppath, "%s/%X-%X.snap.%d.tmp",
1582 : : PG_LOGICAL_SNAPSHOTS_DIR,
1583 : 336 : LSN_FORMAT_ARGS(lsn), MyProcPid);
1584 : :
1585 : : /*
1586 : : * Unlink temporary file if it already exists, needs to have been before a
1587 : : * crash/error since we won't enter this function twice from within a
1588 : : * single decoding slot/backend and the temporary file contains the pid of
1589 : : * the current process.
1590 : : */
1591 [ + - - + ]: 336 : if (unlink(tmppath) != 0 && errno != ENOENT)
1592 [ # # ]: 0 : ereport(ERROR,
1593 : : (errcode_for_file_access(),
1594 : : errmsg("could not remove file \"%s\": %m", tmppath)));
1595 : :
1596 : 336 : old_ctx = MemoryContextSwitchTo(builder->context);
1597 : :
1598 : : /* Get the catalog modifying transactions that are yet not committed */
1599 : 336 : catchange_xip = ReorderBufferGetCatalogChangesXacts(builder->reorder);
1600 : 336 : catchange_xcnt = dclist_count(&builder->reorder->catchange_txns);
1601 : :
1602 : 336 : needed_length = sizeof(SnapBuildOnDisk) +
1603 : 336 : sizeof(TransactionId) * (builder->committed.xcnt + catchange_xcnt);
1604 : :
1605 : 336 : ondisk_c = palloc0(needed_length);
1606 : 336 : ondisk = (SnapBuildOnDisk *) ondisk_c;
1607 : 336 : ondisk->magic = SNAPBUILD_MAGIC;
1608 : 336 : ondisk->version = SNAPBUILD_VERSION;
1609 : 336 : ondisk->length = needed_length;
1610 : 336 : INIT_CRC32C(ondisk->checksum);
1611 : 336 : COMP_CRC32C(ondisk->checksum,
1612 : : ((char *) ondisk) + SnapBuildOnDiskNotChecksummedSize,
1613 : : SnapBuildOnDiskConstantSize - SnapBuildOnDiskNotChecksummedSize);
1614 : 336 : ondisk_c += sizeof(SnapBuildOnDisk);
1615 : :
1616 : 336 : memcpy(&ondisk->builder, builder, sizeof(SnapBuild));
1617 : : /* NULL-ify memory-only data */
1618 : 336 : ondisk->builder.context = NULL;
1619 : 336 : ondisk->builder.snapshot = NULL;
1620 : 336 : ondisk->builder.reorder = NULL;
1621 : 336 : ondisk->builder.committed.xip = NULL;
1622 : 336 : ondisk->builder.catchange.xip = NULL;
1623 : : /* update catchange only on disk data */
1624 : 336 : ondisk->builder.catchange.xcnt = catchange_xcnt;
1625 : :
1626 : 336 : COMP_CRC32C(ondisk->checksum,
1627 : : &ondisk->builder,
1628 : : sizeof(SnapBuild));
1629 : :
1630 : : /* copy committed xacts */
1631 [ + + ]: 336 : if (builder->committed.xcnt > 0)
1632 : : {
1633 : 61 : sz = sizeof(TransactionId) * builder->committed.xcnt;
1634 : 61 : memcpy(ondisk_c, builder->committed.xip, sz);
1635 : 61 : COMP_CRC32C(ondisk->checksum, ondisk_c, sz);
1636 : 61 : ondisk_c += sz;
1637 : : }
1638 : :
1639 : : /* copy catalog modifying xacts */
1640 [ + + ]: 336 : if (catchange_xcnt > 0)
1641 : : {
1642 : 8 : sz = sizeof(TransactionId) * catchange_xcnt;
1643 : 8 : memcpy(ondisk_c, catchange_xip, sz);
1644 : 8 : COMP_CRC32C(ondisk->checksum, ondisk_c, sz);
1645 : 8 : ondisk_c += sz;
1646 : : }
1647 : :
1648 : 336 : FIN_CRC32C(ondisk->checksum);
1649 : :
1650 : : /* we have valid data now, open tempfile and write it there */
1651 : 336 : fd = OpenTransientFile(tmppath,
1652 : : O_CREAT | O_EXCL | O_WRONLY | PG_BINARY);
1653 [ - + ]: 336 : if (fd < 0)
1654 [ # # ]: 0 : ereport(ERROR,
1655 : : (errcode_for_file_access(),
1656 : : errmsg("could not open file \"%s\": %m", tmppath)));
1657 : :
1658 : 336 : errno = 0;
1659 : 336 : pgstat_report_wait_start(WAIT_EVENT_SNAPBUILD_WRITE);
1660 [ - + ]: 336 : if ((write(fd, ondisk, needed_length)) != needed_length)
1661 : : {
1662 : 0 : int save_errno = errno;
1663 : :
1664 : 0 : CloseTransientFile(fd);
1665 : :
1666 : : /* if write didn't set errno, assume problem is no disk space */
1667 [ # # ]: 0 : errno = save_errno ? save_errno : ENOSPC;
1668 [ # # ]: 0 : ereport(ERROR,
1669 : : (errcode_for_file_access(),
1670 : : errmsg("could not write to file \"%s\": %m", tmppath)));
1671 : : }
1672 : 336 : pgstat_report_wait_end();
1673 : :
1674 : : /*
1675 : : * fsync the file before renaming so that even if we crash after this we
1676 : : * have either a fully valid file or nothing.
1677 : : *
1678 : : * It's safe to just ERROR on fsync() here because we'll retry the whole
1679 : : * operation including the writes.
1680 : : *
1681 : : * TODO: Do the fsync() via checkpoints/restartpoints, doing it here has
1682 : : * some noticeable overhead since it's performed synchronously during
1683 : : * decoding?
1684 : : */
1685 : 336 : pgstat_report_wait_start(WAIT_EVENT_SNAPBUILD_SYNC);
1686 [ - + ]: 336 : if (pg_fsync(fd) != 0)
1687 : : {
1688 : 0 : int save_errno = errno;
1689 : :
1690 : 0 : CloseTransientFile(fd);
1691 : 0 : errno = save_errno;
1692 [ # # ]: 0 : ereport(ERROR,
1693 : : (errcode_for_file_access(),
1694 : : errmsg("could not fsync file \"%s\": %m", tmppath)));
1695 : : }
1696 : 336 : pgstat_report_wait_end();
1697 : :
1698 [ - + ]: 336 : if (CloseTransientFile(fd) != 0)
1699 [ # # ]: 0 : ereport(ERROR,
1700 : : (errcode_for_file_access(),
1701 : : errmsg("could not close file \"%s\": %m", tmppath)));
1702 : :
1703 : 336 : fsync_fname(PG_LOGICAL_SNAPSHOTS_DIR, true);
1704 : :
1705 : : /*
1706 : : * We may overwrite the work from some other backend, but that's ok, our
1707 : : * snapshot is valid as well, we'll just have done some superfluous work.
1708 : : */
1709 [ - + ]: 336 : if (rename(tmppath, path) != 0)
1710 : : {
1711 [ # # ]: 0 : ereport(ERROR,
1712 : : (errcode_for_file_access(),
1713 : : errmsg("could not rename file \"%s\" to \"%s\": %m",
1714 : : tmppath, path)));
1715 : : }
1716 : :
1717 : : /* make sure we persist */
1718 : 336 : fsync_fname(path, false);
1719 : 336 : fsync_fname(PG_LOGICAL_SNAPSHOTS_DIR, true);
1720 : :
1721 : : /*
1722 : : * Now there's no way we can lose the dumped state anymore, remember this
1723 : : * as a serialization point.
1724 : : */
1725 : 336 : builder->last_serialized_snapshot = lsn;
1726 : :
1727 : 336 : MemoryContextSwitchTo(old_ctx);
1728 : :
1729 : 605 : out:
1730 : 605 : ReorderBufferSetRestartPoint(builder->reorder,
1731 : : builder->last_serialized_snapshot);
1732 : : /* be tidy */
1733 [ + + ]: 605 : if (ondisk)
1734 : 336 : pfree(ondisk);
1735 [ + + ]: 605 : if (catchange_xip)
1736 : 8 : pfree(catchange_xip);
1737 : : }
1738 : :
1739 : : /*
1740 : : * Restore the logical snapshot file contents to 'ondisk'.
1741 : : *
1742 : : * 'context' is the memory context where the catalog modifying/committed xid
1743 : : * will live.
1744 : : * If 'missing_ok' is true, will not throw an error if the file is not found.
1745 : : */
1746 : : bool
1747 : 21 : SnapBuildRestoreSnapshot(SnapBuildOnDisk *ondisk, XLogRecPtr lsn,
1748 : : MemoryContext context, bool missing_ok)
1749 : : {
1750 : : int fd;
1751 : : pg_crc32c checksum;
1752 : : Size sz;
1753 : : char path[MAXPGPATH];
1754 : :
1755 : 21 : sprintf(path, "%s/%X-%X.snap",
1756 : : PG_LOGICAL_SNAPSHOTS_DIR,
1757 : 21 : LSN_FORMAT_ARGS(lsn));
1758 : :
1759 : 21 : fd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
1760 : :
1761 [ + + ]: 21 : if (fd < 0)
1762 : : {
1763 [ + - + - ]: 11 : if (missing_ok && errno == ENOENT)
1764 : 11 : return false;
1765 : :
1766 [ # # ]: 0 : ereport(ERROR,
1767 : : (errcode_for_file_access(),
1768 : : errmsg("could not open file \"%s\": %m", path)));
1769 : : }
1770 : :
1771 : : /* ----
1772 : : * Make sure the snapshot had been stored safely to disk, that's normally
1773 : : * cheap.
1774 : : * Note that we do not need PANIC here, nobody will be able to use the
1775 : : * slot without fsyncing, and saving it won't succeed without an fsync()
1776 : : * either...
1777 : : * ----
1778 : : */
1779 : 10 : fsync_fname(path, false);
1780 : 10 : fsync_fname(PG_LOGICAL_SNAPSHOTS_DIR, true);
1781 : :
1782 : : /* read statically sized portion of snapshot */
1783 : 10 : SnapBuildRestoreContents(fd, ondisk, SnapBuildOnDiskConstantSize, path);
1784 : :
1785 [ - + ]: 10 : if (ondisk->magic != SNAPBUILD_MAGIC)
1786 [ # # ]: 0 : ereport(ERROR,
1787 : : (errcode(ERRCODE_DATA_CORRUPTED),
1788 : : errmsg("snapbuild state file \"%s\" has wrong magic number: %u instead of %u",
1789 : : path, ondisk->magic, SNAPBUILD_MAGIC)));
1790 : :
1791 [ - + ]: 10 : if (ondisk->version != SNAPBUILD_VERSION)
1792 [ # # ]: 0 : ereport(ERROR,
1793 : : (errcode(ERRCODE_DATA_CORRUPTED),
1794 : : errmsg("snapbuild state file \"%s\" has unsupported version: %u instead of %u",
1795 : : path, ondisk->version, SNAPBUILD_VERSION)));
1796 : :
1797 : 10 : INIT_CRC32C(checksum);
1798 : 10 : COMP_CRC32C(checksum,
1799 : : ((char *) ondisk) + SnapBuildOnDiskNotChecksummedSize,
1800 : : SnapBuildOnDiskConstantSize - SnapBuildOnDiskNotChecksummedSize);
1801 : :
1802 : : /* read SnapBuild */
1803 : 10 : SnapBuildRestoreContents(fd, &ondisk->builder, sizeof(SnapBuild), path);
1804 : 10 : COMP_CRC32C(checksum, &ondisk->builder, sizeof(SnapBuild));
1805 : :
1806 : : /* restore committed xacts information */
1807 [ + + ]: 10 : if (ondisk->builder.committed.xcnt > 0)
1808 : : {
1809 : 5 : sz = sizeof(TransactionId) * ondisk->builder.committed.xcnt;
1810 : 5 : ondisk->builder.committed.xip = MemoryContextAllocZero(context, sz);
1811 : 5 : SnapBuildRestoreContents(fd, ondisk->builder.committed.xip, sz, path);
1812 : 5 : COMP_CRC32C(checksum, ondisk->builder.committed.xip, sz);
1813 : : }
1814 : :
1815 : : /* restore catalog modifying xacts information */
1816 [ + + ]: 10 : if (ondisk->builder.catchange.xcnt > 0)
1817 : : {
1818 : 4 : sz = sizeof(TransactionId) * ondisk->builder.catchange.xcnt;
1819 : 4 : ondisk->builder.catchange.xip = MemoryContextAllocZero(context, sz);
1820 : 4 : SnapBuildRestoreContents(fd, ondisk->builder.catchange.xip, sz, path);
1821 : 4 : COMP_CRC32C(checksum, ondisk->builder.catchange.xip, sz);
1822 : : }
1823 : :
1824 [ - + ]: 10 : if (CloseTransientFile(fd) != 0)
1825 [ # # ]: 0 : ereport(ERROR,
1826 : : (errcode_for_file_access(),
1827 : : errmsg("could not close file \"%s\": %m", path)));
1828 : :
1829 : 10 : FIN_CRC32C(checksum);
1830 : :
1831 : : /* verify checksum of what we've read */
1832 [ - + ]: 10 : if (!EQ_CRC32C(checksum, ondisk->checksum))
1833 [ # # ]: 0 : ereport(ERROR,
1834 : : (errcode(ERRCODE_DATA_CORRUPTED),
1835 : : errmsg("checksum mismatch for snapbuild state file \"%s\": is %u, should be %u",
1836 : : path, checksum, ondisk->checksum)));
1837 : :
1838 : 10 : return true;
1839 : : }
1840 : :
1841 : : /*
1842 : : * Restore a snapshot into 'builder' if previously one has been stored at the
1843 : : * location indicated by 'lsn'. Returns true if successful, false otherwise.
1844 : : */
1845 : : static bool
1846 : 19 : SnapBuildRestore(SnapBuild *builder, XLogRecPtr lsn)
1847 : : {
1848 : : SnapBuildOnDisk ondisk;
1849 : :
1850 : : /* no point in loading a snapshot if we're already there */
1851 [ - + ]: 19 : if (builder->state == SNAPBUILD_CONSISTENT)
1852 : 0 : return false;
1853 : :
1854 : : /* validate and restore the snapshot to 'ondisk' */
1855 [ + + ]: 19 : if (!SnapBuildRestoreSnapshot(&ondisk, lsn, builder->context, true))
1856 : 11 : return false;
1857 : :
1858 : : /*
1859 : : * ok, we now have a sensible snapshot here, figure out if it has more
1860 : : * information than we have.
1861 : : */
1862 : :
1863 : : /*
1864 : : * We are only interested in consistent snapshots for now, comparing
1865 : : * whether one incomplete snapshot is more "advanced" seems to be
1866 : : * unnecessarily complex.
1867 : : */
1868 [ - + ]: 8 : if (ondisk.builder.state < SNAPBUILD_CONSISTENT)
1869 : 0 : goto snapshot_not_interesting;
1870 : :
1871 : : /*
1872 : : * Don't use a snapshot that requires an xmin that we cannot guarantee to
1873 : : * be available.
1874 : : */
1875 [ - + ]: 8 : if (TransactionIdPrecedes(ondisk.builder.xmin, builder->initial_xmin_horizon))
1876 : 0 : goto snapshot_not_interesting;
1877 : :
1878 : : /*
1879 : : * Consistent snapshots have no next phase. Reset next_phase_at as it is
1880 : : * possible that an old value may remain.
1881 : : */
1882 : : Assert(ondisk.builder.next_phase_at == InvalidTransactionId);
1883 : 8 : builder->next_phase_at = InvalidTransactionId;
1884 : :
1885 : : /* ok, we think the snapshot is sensible, copy over everything important */
1886 : 8 : builder->xmin = ondisk.builder.xmin;
1887 : 8 : builder->xmax = ondisk.builder.xmax;
1888 : 8 : builder->state = ondisk.builder.state;
1889 : :
1890 : 8 : builder->committed.xcnt = ondisk.builder.committed.xcnt;
1891 : : /* We only allocated/stored xcnt, not xcnt_space xids ! */
1892 : : /* don't overwrite preallocated xip, if we don't have anything here */
1893 [ + + ]: 8 : if (builder->committed.xcnt > 0)
1894 : : {
1895 : 3 : pfree(builder->committed.xip);
1896 : 3 : builder->committed.xcnt_space = ondisk.builder.committed.xcnt;
1897 : 3 : builder->committed.xip = ondisk.builder.committed.xip;
1898 : : }
1899 : 8 : ondisk.builder.committed.xip = NULL;
1900 : :
1901 : : /* set catalog modifying transactions */
1902 [ - + ]: 8 : if (builder->catchange.xip)
1903 : 0 : pfree(builder->catchange.xip);
1904 : 8 : builder->catchange.xcnt = ondisk.builder.catchange.xcnt;
1905 : 8 : builder->catchange.xip = ondisk.builder.catchange.xip;
1906 : 8 : ondisk.builder.catchange.xip = NULL;
1907 : :
1908 : : /* our snapshot is not interesting anymore, build a new one */
1909 [ - + ]: 8 : if (builder->snapshot != NULL)
1910 : : {
1911 : 0 : SnapBuildSnapDecRefcount(builder->snapshot);
1912 : : }
1913 : 8 : builder->snapshot = SnapBuildBuildSnapshot(builder);
1914 : 8 : SnapBuildSnapIncRefcount(builder->snapshot);
1915 : :
1916 : 8 : ReorderBufferSetRestartPoint(builder->reorder, lsn);
1917 : :
1918 : : Assert(builder->state == SNAPBUILD_CONSISTENT);
1919 : :
1920 [ + + + + ]: 8 : ereport(LogicalDecodingLogLevel(),
1921 : : errmsg("logical decoding found consistent point at %X/%08X",
1922 : : LSN_FORMAT_ARGS(lsn)),
1923 : : errdetail("Logical decoding will begin using saved snapshot."));
1924 : 8 : return true;
1925 : :
1926 : 0 : snapshot_not_interesting:
1927 [ # # ]: 0 : if (ondisk.builder.committed.xip != NULL)
1928 : 0 : pfree(ondisk.builder.committed.xip);
1929 [ # # ]: 0 : if (ondisk.builder.catchange.xip != NULL)
1930 : 0 : pfree(ondisk.builder.catchange.xip);
1931 : 0 : return false;
1932 : : }
1933 : :
1934 : : /*
1935 : : * Read the contents of the serialized snapshot to 'dest'.
1936 : : */
1937 : : static void
1938 : 29 : SnapBuildRestoreContents(int fd, void *dest, Size size, const char *path)
1939 : : {
1940 : : ssize_t readBytes;
1941 : :
1942 : 29 : pgstat_report_wait_start(WAIT_EVENT_SNAPBUILD_READ);
1943 : 29 : readBytes = read(fd, dest, size);
1944 : 29 : pgstat_report_wait_end();
1945 [ - + ]: 29 : if (readBytes != size)
1946 : : {
1947 : 0 : int save_errno = errno;
1948 : :
1949 : 0 : CloseTransientFile(fd);
1950 : :
1951 [ # # ]: 0 : if (readBytes < 0)
1952 : : {
1953 : 0 : errno = save_errno;
1954 [ # # ]: 0 : ereport(ERROR,
1955 : : (errcode_for_file_access(),
1956 : : errmsg("could not read file \"%s\": %m", path)));
1957 : : }
1958 : : else
1959 [ # # ]: 0 : ereport(ERROR,
1960 : : (errcode(ERRCODE_DATA_CORRUPTED),
1961 : : errmsg("could not read file \"%s\": read %zd of %zu",
1962 : : path, readBytes, size)));
1963 : : }
1964 : 29 : }
1965 : :
1966 : : /*
1967 : : * Remove all serialized snapshots that are not required anymore because no
1968 : : * slot can need them. This doesn't actually have to run during a checkpoint,
1969 : : * but it's a convenient point to schedule this.
1970 : : *
1971 : : * NB: We run this during checkpoints even if logical decoding is disabled so
1972 : : * we cleanup old slots at some point after it got disabled.
1973 : : */
1974 : : void
1975 : 1950 : CheckPointSnapBuild(void)
1976 : : {
1977 : : XLogRecPtr cutoff;
1978 : : XLogRecPtr redo;
1979 : : DIR *snap_dir;
1980 : : struct dirent *snap_de;
1981 : : char path[MAXPGPATH + sizeof(PG_LOGICAL_SNAPSHOTS_DIR)];
1982 : :
1983 : : /*
1984 : : * We start off with a minimum of the last redo pointer. No new
1985 : : * replication slot will start before that, so that's a safe upper bound
1986 : : * for removal.
1987 : : */
1988 : 1950 : redo = GetRedoRecPtr();
1989 : :
1990 : : /* now check for the restart ptrs from existing slots */
1991 : 1950 : cutoff = ReplicationSlotsComputeLogicalRestartLSN();
1992 : :
1993 : : /* don't start earlier than the restart lsn */
1994 [ + + ]: 1950 : if (redo < cutoff)
1995 : 1 : cutoff = redo;
1996 : :
1997 : 1950 : snap_dir = AllocateDir(PG_LOGICAL_SNAPSHOTS_DIR);
1998 [ + + ]: 6153 : while ((snap_de = ReadDir(snap_dir, PG_LOGICAL_SNAPSHOTS_DIR)) != NULL)
1999 : : {
2000 : : uint32 hi;
2001 : : uint32 lo;
2002 : : XLogRecPtr lsn;
2003 : : PGFileType de_type;
2004 : :
2005 [ + + ]: 4203 : if (strcmp(snap_de->d_name, ".") == 0 ||
2006 [ + + ]: 2253 : strcmp(snap_de->d_name, "..") == 0)
2007 : 3900 : continue;
2008 : :
2009 : 303 : snprintf(path, sizeof(path), "%s/%s", PG_LOGICAL_SNAPSHOTS_DIR, snap_de->d_name);
2010 : 303 : de_type = get_dirent_type(path, snap_de, false, DEBUG1);
2011 : :
2012 [ + - - + ]: 303 : if (de_type != PGFILETYPE_ERROR && de_type != PGFILETYPE_REG)
2013 : : {
2014 [ # # ]: 0 : elog(DEBUG1, "only regular files expected: %s", path);
2015 : 0 : continue;
2016 : : }
2017 : :
2018 : : /*
2019 : : * temporary filenames from SnapBuildSerialize() include the LSN and
2020 : : * everything but are postfixed by .$pid.tmp. We can just remove them
2021 : : * the same as other files because there can be none that are
2022 : : * currently being written that are older than cutoff.
2023 : : *
2024 : : * We just log a message if a file doesn't fit the pattern, it's
2025 : : * probably some editors lock/state file or similar...
2026 : : */
2027 [ - + ]: 303 : if (sscanf(snap_de->d_name, "%X-%X.snap", &hi, &lo) != 2)
2028 : : {
2029 [ # # ]: 0 : ereport(LOG,
2030 : : (errmsg("could not parse file name \"%s\"", path)));
2031 : 0 : continue;
2032 : : }
2033 : :
2034 : 303 : lsn = ((uint64) hi) << 32 | lo;
2035 : :
2036 : : /* check whether we still need it */
2037 [ + + + + ]: 303 : if (lsn < cutoff || !XLogRecPtrIsValid(cutoff))
2038 : : {
2039 [ + + ]: 197 : elog(DEBUG1, "removing snapbuild snapshot %s", path);
2040 : :
2041 : : /*
2042 : : * It's not particularly harmful, though strange, if we can't
2043 : : * remove the file here. Don't prevent the checkpoint from
2044 : : * completing, that'd be a cure worse than the disease.
2045 : : */
2046 [ - + ]: 197 : if (unlink(path) < 0)
2047 : : {
2048 [ # # ]: 0 : ereport(LOG,
2049 : : (errcode_for_file_access(),
2050 : : errmsg("could not remove file \"%s\": %m",
2051 : : path)));
2052 : 0 : continue;
2053 : : }
2054 : : }
2055 : : }
2056 : 1950 : FreeDir(snap_dir);
2057 : 1950 : }
2058 : :
2059 : : /*
2060 : : * Check if a logical snapshot at the specified point has been serialized.
2061 : : */
2062 : : bool
2063 : 14 : SnapBuildSnapshotExists(XLogRecPtr lsn)
2064 : : {
2065 : : char path[MAXPGPATH];
2066 : : int ret;
2067 : : struct stat stat_buf;
2068 : :
2069 : 14 : sprintf(path, "%s/%X-%X.snap",
2070 : : PG_LOGICAL_SNAPSHOTS_DIR,
2071 : 14 : LSN_FORMAT_ARGS(lsn));
2072 : :
2073 : 14 : ret = stat(path, &stat_buf);
2074 : :
2075 [ + + - + ]: 14 : if (ret != 0 && errno != ENOENT)
2076 [ # # ]: 0 : ereport(ERROR,
2077 : : (errcode_for_file_access(),
2078 : : errmsg("could not stat file \"%s\": %m", path)));
2079 : :
2080 : 14 : return ret == 0;
2081 : : }
|
