Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * walsummarizer.c
4 : *
5 : * Background process to perform WAL summarization, if it is enabled.
6 : * It continuously scans the write-ahead log and periodically emits a
7 : * summary file which indicates which blocks in which relation forks
8 : * were modified by WAL records in the LSN range covered by the summary
9 : * file. See walsummary.c and blkreftable.c for more details on the
10 : * naming and contents of WAL summary files.
11 : *
12 : * If configured to do, this background process will also remove WAL
13 : * summary files when the file timestamp is older than a configurable
14 : * threshold (but only if the WAL has been removed first).
15 : *
16 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
17 : *
18 : * IDENTIFICATION
19 : * src/backend/postmaster/walsummarizer.c
20 : *
21 : *-------------------------------------------------------------------------
22 : */
23 : #include "postgres.h"
24 :
25 : #include "access/timeline.h"
26 : #include "access/xlog.h"
27 : #include "access/xlog_internal.h"
28 : #include "access/xlogrecovery.h"
29 : #include "access/xlogutils.h"
30 : #include "backup/walsummary.h"
31 : #include "catalog/storage_xlog.h"
32 : #include "commands/dbcommands_xlog.h"
33 : #include "common/blkreftable.h"
34 : #include "libpq/pqsignal.h"
35 : #include "miscadmin.h"
36 : #include "pgstat.h"
37 : #include "postmaster/auxprocess.h"
38 : #include "postmaster/interrupt.h"
39 : #include "postmaster/walsummarizer.h"
40 : #include "replication/walreceiver.h"
41 : #include "storage/aio_subsys.h"
42 : #include "storage/fd.h"
43 : #include "storage/ipc.h"
44 : #include "storage/latch.h"
45 : #include "storage/lwlock.h"
46 : #include "storage/proc.h"
47 : #include "storage/procsignal.h"
48 : #include "storage/shmem.h"
49 : #include "utils/guc.h"
50 : #include "utils/memutils.h"
51 : #include "utils/wait_event.h"
52 :
53 : /*
54 : * Data in shared memory related to WAL summarization.
55 : */
56 : typedef struct
57 : {
58 : /*
59 : * These fields are protected by WALSummarizerLock.
60 : *
61 : * Until we've discovered what summary files already exist on disk and
62 : * stored that information in shared memory, initialized is false and the
63 : * other fields here contain no meaningful information. After that has
64 : * been done, initialized is true.
65 : *
66 : * summarized_tli and summarized_lsn indicate the last LSN and TLI at
67 : * which the next summary file will start. Normally, these are the LSN and
68 : * TLI at which the last file ended; in such case, lsn_is_exact is true.
69 : * If, however, the LSN is just an approximation, then lsn_is_exact is
70 : * false. This can happen if, for example, there are no existing WAL
71 : * summary files at startup. In that case, we have to derive the position
72 : * at which to start summarizing from the WAL files that exist on disk,
73 : * and so the LSN might point to the start of the next file even though
74 : * that might happen to be in the middle of a WAL record.
75 : *
76 : * summarizer_pgprocno is the proc number of the summarizer process, if
77 : * one is running, or else INVALID_PROC_NUMBER.
78 : *
79 : * pending_lsn is used by the summarizer to advertise the ending LSN of a
80 : * record it has recently read. It shouldn't ever be less than
81 : * summarized_lsn, but might be greater, because the summarizer buffers
82 : * data for a range of LSNs in memory before writing out a new file.
83 : */
84 : bool initialized;
85 : TimeLineID summarized_tli;
86 : XLogRecPtr summarized_lsn;
87 : bool lsn_is_exact;
88 : ProcNumber summarizer_pgprocno;
89 : XLogRecPtr pending_lsn;
90 :
91 : /*
92 : * This field handles its own synchronization.
93 : */
94 : ConditionVariable summary_file_cv;
95 : } WalSummarizerData;
96 :
97 : /*
98 : * Private data for our xlogreader's page read callback.
99 : */
100 : typedef struct
101 : {
102 : TimeLineID tli;
103 : bool historic;
104 : XLogRecPtr read_upto;
105 : bool end_of_wal;
106 : } SummarizerReadLocalXLogPrivate;
107 :
108 : /* Pointer to shared memory state. */
109 : static WalSummarizerData *WalSummarizerCtl;
110 :
111 : /*
112 : * When we reach end of WAL and need to read more, we sleep for a number of
113 : * milliseconds that is an integer multiple of MS_PER_SLEEP_QUANTUM. This is
114 : * the multiplier. It should vary between 1 and MAX_SLEEP_QUANTA, depending
115 : * on system activity. See summarizer_wait_for_wal() for how we adjust this.
116 : */
117 : static long sleep_quanta = 1;
118 :
119 : /*
120 : * The sleep time will always be a multiple of 200ms and will not exceed
121 : * thirty seconds (150 * 200 = 30 * 1000). Note that the timeout here needs
122 : * to be substantially less than the maximum amount of time for which an
123 : * incremental backup will wait for this process to catch up. Otherwise, an
124 : * incremental backup might time out on an idle system just because we sleep
125 : * for too long.
126 : */
127 : #define MAX_SLEEP_QUANTA 150
128 : #define MS_PER_SLEEP_QUANTUM 200
129 :
130 : /*
131 : * This is a count of the number of pages of WAL that we've read since the
132 : * last time we waited for more WAL to appear.
133 : */
134 : static long pages_read_since_last_sleep = 0;
135 :
136 : /*
137 : * Most recent RedoRecPtr value observed by MaybeRemoveOldWalSummaries.
138 : */
139 : static XLogRecPtr redo_pointer_at_last_summary_removal = InvalidXLogRecPtr;
140 :
141 : /*
142 : * GUC parameters
143 : */
144 : bool summarize_wal = false;
145 : int wal_summary_keep_time = 10 * HOURS_PER_DAY * MINS_PER_HOUR;
146 :
147 : static void WalSummarizerShutdown(int code, Datum arg);
148 : static XLogRecPtr GetLatestLSN(TimeLineID *tli);
149 : static void ProcessWalSummarizerInterrupts(void);
150 : static XLogRecPtr SummarizeWAL(TimeLineID tli, XLogRecPtr start_lsn,
151 : bool exact, XLogRecPtr switch_lsn,
152 : XLogRecPtr maximum_lsn);
153 : static void SummarizeDbaseRecord(XLogReaderState *xlogreader,
154 : BlockRefTable *brtab);
155 : static void SummarizeSmgrRecord(XLogReaderState *xlogreader,
156 : BlockRefTable *brtab);
157 : static void SummarizeXactRecord(XLogReaderState *xlogreader,
158 : BlockRefTable *brtab);
159 : static bool SummarizeXlogRecord(XLogReaderState *xlogreader,
160 : bool *new_fast_forward);
161 : static int summarizer_read_local_xlog_page(XLogReaderState *state,
162 : XLogRecPtr targetPagePtr,
163 : int reqLen,
164 : XLogRecPtr targetRecPtr,
165 : char *cur_page);
166 : static void summarizer_wait_for_wal(void);
167 : static void MaybeRemoveOldWalSummaries(void);
168 :
169 : /*
170 : * Amount of shared memory required for this module.
171 : */
172 : Size
173 5826 : WalSummarizerShmemSize(void)
174 : {
175 5826 : return sizeof(WalSummarizerData);
176 : }
177 :
178 : /*
179 : * Create or attach to shared memory segment for this module.
180 : */
181 : void
182 2032 : WalSummarizerShmemInit(void)
183 : {
184 : bool found;
185 :
186 2032 : WalSummarizerCtl = (WalSummarizerData *)
187 2032 : ShmemInitStruct("Wal Summarizer Ctl", WalSummarizerShmemSize(),
188 : &found);
189 :
190 2032 : if (!found)
191 : {
192 : /*
193 : * First time through, so initialize.
194 : *
195 : * We're just filling in dummy values here -- the real initialization
196 : * will happen when GetOldestUnsummarizedLSN() is called for the first
197 : * time.
198 : */
199 2032 : WalSummarizerCtl->initialized = false;
200 2032 : WalSummarizerCtl->summarized_tli = 0;
201 2032 : WalSummarizerCtl->summarized_lsn = InvalidXLogRecPtr;
202 2032 : WalSummarizerCtl->lsn_is_exact = false;
203 2032 : WalSummarizerCtl->summarizer_pgprocno = INVALID_PROC_NUMBER;
204 2032 : WalSummarizerCtl->pending_lsn = InvalidXLogRecPtr;
205 2032 : ConditionVariableInit(&WalSummarizerCtl->summary_file_cv);
206 : }
207 2032 : }
208 :
209 : /*
210 : * Entry point for walsummarizer process.
211 : */
212 : void
213 2 : WalSummarizerMain(const void *startup_data, size_t startup_data_len)
214 : {
215 : sigjmp_buf local_sigjmp_buf;
216 : MemoryContext context;
217 :
218 : /*
219 : * Within this function, 'current_lsn' and 'current_tli' refer to the
220 : * point from which the next WAL summary file should start. 'exact' is
221 : * true if 'current_lsn' is known to be the start of a WAL record or WAL
222 : * segment, and false if it might be in the middle of a record someplace.
223 : *
224 : * 'switch_lsn' and 'switch_tli', if set, are the LSN at which we need to
225 : * switch to a new timeline and the timeline to which we need to switch.
226 : * If not set, we either haven't figured out the answers yet or we're
227 : * already on the latest timeline.
228 : */
229 : XLogRecPtr current_lsn;
230 : TimeLineID current_tli;
231 : bool exact;
232 2 : XLogRecPtr switch_lsn = InvalidXLogRecPtr;
233 2 : TimeLineID switch_tli = 0;
234 :
235 : Assert(startup_data_len == 0);
236 :
237 2 : MyBackendType = B_WAL_SUMMARIZER;
238 2 : AuxiliaryProcessMainCommon();
239 :
240 2 : ereport(DEBUG1,
241 : (errmsg_internal("WAL summarizer started")));
242 :
243 : /*
244 : * Properly accept or ignore signals the postmaster might send us
245 : *
246 : * We have no particular use for SIGINT at the moment, but seems
247 : * reasonable to treat like SIGTERM.
248 : */
249 2 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
250 2 : pqsignal(SIGINT, SignalHandlerForShutdownRequest);
251 2 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
252 : /* SIGQUIT handler was already set up by InitPostmasterChild */
253 2 : pqsignal(SIGALRM, SIG_IGN);
254 2 : pqsignal(SIGPIPE, SIG_IGN);
255 2 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
256 2 : pqsignal(SIGUSR2, SIG_IGN); /* not used */
257 :
258 : /* Advertise ourselves. */
259 2 : on_shmem_exit(WalSummarizerShutdown, (Datum) 0);
260 2 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
261 2 : WalSummarizerCtl->summarizer_pgprocno = MyProcNumber;
262 2 : LWLockRelease(WALSummarizerLock);
263 :
264 : /* Create and switch to a memory context that we can reset on error. */
265 2 : context = AllocSetContextCreate(TopMemoryContext,
266 : "Wal Summarizer",
267 : ALLOCSET_DEFAULT_SIZES);
268 2 : MemoryContextSwitchTo(context);
269 :
270 : /*
271 : * Reset some signals that are accepted by postmaster but not here
272 : */
273 2 : pqsignal(SIGCHLD, SIG_DFL);
274 :
275 : /*
276 : * If an exception is encountered, processing resumes here.
277 : */
278 2 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
279 : {
280 : /* Since not using PG_TRY, must reset error stack by hand */
281 0 : error_context_stack = NULL;
282 :
283 : /* Prevent interrupts while cleaning up */
284 0 : HOLD_INTERRUPTS();
285 :
286 : /* Report the error to the server log */
287 0 : EmitErrorReport();
288 :
289 : /* Release resources we might have acquired. */
290 0 : LWLockReleaseAll();
291 0 : ConditionVariableCancelSleep();
292 0 : pgstat_report_wait_end();
293 0 : pgaio_error_cleanup();
294 0 : ReleaseAuxProcessResources(false);
295 0 : AtEOXact_Files(false);
296 0 : AtEOXact_HashTables(false);
297 :
298 : /*
299 : * Now return to normal top-level context and clear ErrorContext for
300 : * next time.
301 : */
302 0 : MemoryContextSwitchTo(context);
303 0 : FlushErrorState();
304 :
305 : /* Flush any leaked data in the top-level context */
306 0 : MemoryContextReset(context);
307 :
308 : /* Now we can allow interrupts again */
309 0 : RESUME_INTERRUPTS();
310 :
311 : /*
312 : * Sleep for 10 seconds before attempting to resume operations in
313 : * order to avoid excessive logging.
314 : *
315 : * Many of the likely error conditions are things that will repeat
316 : * every time. For example, if the WAL can't be read or the summary
317 : * can't be written, only administrator action will cure the problem.
318 : * So a really fast retry time doesn't seem to be especially
319 : * beneficial, and it will clutter the logs.
320 : */
321 0 : (void) WaitLatch(NULL,
322 : WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
323 : 10000,
324 : WAIT_EVENT_WAL_SUMMARIZER_ERROR);
325 : }
326 :
327 : /* We can now handle ereport(ERROR) */
328 2 : PG_exception_stack = &local_sigjmp_buf;
329 :
330 : /*
331 : * Unblock signals (they were blocked when the postmaster forked us)
332 : */
333 2 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
334 :
335 : /*
336 : * Fetch information about previous progress from shared memory, and ask
337 : * GetOldestUnsummarizedLSN to reset pending_lsn to summarized_lsn. We
338 : * might be recovering from an error, and if so, pending_lsn might have
339 : * advanced past summarized_lsn, but any WAL we read previously has been
340 : * lost and will need to be reread.
341 : *
342 : * If we discover that WAL summarization is not enabled, just exit.
343 : */
344 2 : current_lsn = GetOldestUnsummarizedLSN(¤t_tli, &exact);
345 2 : if (XLogRecPtrIsInvalid(current_lsn))
346 0 : proc_exit(0);
347 :
348 : /*
349 : * Loop forever
350 : */
351 : for (;;)
352 16 : {
353 : XLogRecPtr latest_lsn;
354 : TimeLineID latest_tli;
355 : XLogRecPtr end_of_summary_lsn;
356 :
357 : /* Flush any leaked data in the top-level context */
358 18 : MemoryContextReset(context);
359 :
360 : /* Process any signals received recently. */
361 18 : ProcessWalSummarizerInterrupts();
362 :
363 : /* If it's time to remove any old WAL summaries, do that now. */
364 18 : MaybeRemoveOldWalSummaries();
365 :
366 : /* Find the LSN and TLI up to which we can safely summarize. */
367 18 : latest_lsn = GetLatestLSN(&latest_tli);
368 :
369 : /*
370 : * If we're summarizing a historic timeline and we haven't yet
371 : * computed the point at which to switch to the next timeline, do that
372 : * now.
373 : *
374 : * Note that if this is a standby, what was previously the current
375 : * timeline could become historic at any time.
376 : *
377 : * We could try to make this more efficient by caching the results of
378 : * readTimeLineHistory when latest_tli has not changed, but since we
379 : * only have to do this once per timeline switch, we probably wouldn't
380 : * save any significant amount of work in practice.
381 : */
382 18 : if (current_tli != latest_tli && XLogRecPtrIsInvalid(switch_lsn))
383 : {
384 0 : List *tles = readTimeLineHistory(latest_tli);
385 :
386 0 : switch_lsn = tliSwitchPoint(current_tli, tles, &switch_tli);
387 0 : ereport(DEBUG1,
388 : errmsg_internal("switch point from TLI %u to TLI %u is at %X/%X",
389 : current_tli, switch_tli, LSN_FORMAT_ARGS(switch_lsn)));
390 : }
391 :
392 : /*
393 : * If we've reached the switch LSN, we can't summarize anything else
394 : * on this timeline. Switch to the next timeline and go around again,
395 : * backing up to the exact switch point if we passed it.
396 : */
397 18 : if (!XLogRecPtrIsInvalid(switch_lsn) && current_lsn >= switch_lsn)
398 : {
399 : /* Restart summarization from switch point. */
400 0 : current_tli = switch_tli;
401 0 : current_lsn = switch_lsn;
402 :
403 : /* Next timeline and switch point, if any, not yet known. */
404 0 : switch_lsn = InvalidXLogRecPtr;
405 0 : switch_tli = 0;
406 :
407 : /* Update (really, rewind, if needed) state in shared memory. */
408 0 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
409 0 : WalSummarizerCtl->summarized_lsn = current_lsn;
410 0 : WalSummarizerCtl->summarized_tli = current_tli;
411 0 : WalSummarizerCtl->lsn_is_exact = true;
412 0 : WalSummarizerCtl->pending_lsn = current_lsn;
413 0 : LWLockRelease(WALSummarizerLock);
414 :
415 0 : continue;
416 : }
417 :
418 : /* Summarize WAL. */
419 18 : end_of_summary_lsn = SummarizeWAL(current_tli,
420 : current_lsn, exact,
421 : switch_lsn, latest_lsn);
422 : Assert(!XLogRecPtrIsInvalid(end_of_summary_lsn));
423 : Assert(end_of_summary_lsn >= current_lsn);
424 :
425 : /*
426 : * Update state for next loop iteration.
427 : *
428 : * Next summary file should start from exactly where this one ended.
429 : */
430 16 : current_lsn = end_of_summary_lsn;
431 16 : exact = true;
432 :
433 : /* Update state in shared memory. */
434 16 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
435 16 : WalSummarizerCtl->summarized_lsn = end_of_summary_lsn;
436 16 : WalSummarizerCtl->summarized_tli = current_tli;
437 16 : WalSummarizerCtl->lsn_is_exact = true;
438 16 : WalSummarizerCtl->pending_lsn = end_of_summary_lsn;
439 16 : LWLockRelease(WALSummarizerLock);
440 :
441 : /* Wake up anyone waiting for more summary files to be written. */
442 16 : ConditionVariableBroadcast(&WalSummarizerCtl->summary_file_cv);
443 : }
444 : }
445 :
446 : /*
447 : * Get information about the state of the WAL summarizer.
448 : */
449 : void
450 0 : GetWalSummarizerState(TimeLineID *summarized_tli, XLogRecPtr *summarized_lsn,
451 : XLogRecPtr *pending_lsn, int *summarizer_pid)
452 : {
453 0 : LWLockAcquire(WALSummarizerLock, LW_SHARED);
454 0 : if (!WalSummarizerCtl->initialized)
455 : {
456 : /*
457 : * If initialized is false, the rest of the structure contents are
458 : * undefined.
459 : */
460 0 : *summarized_tli = 0;
461 0 : *summarized_lsn = InvalidXLogRecPtr;
462 0 : *pending_lsn = InvalidXLogRecPtr;
463 0 : *summarizer_pid = -1;
464 : }
465 : else
466 : {
467 0 : int summarizer_pgprocno = WalSummarizerCtl->summarizer_pgprocno;
468 :
469 0 : *summarized_tli = WalSummarizerCtl->summarized_tli;
470 0 : *summarized_lsn = WalSummarizerCtl->summarized_lsn;
471 0 : if (summarizer_pgprocno == INVALID_PROC_NUMBER)
472 : {
473 : /*
474 : * If the summarizer has exited, the fact that it had processed
475 : * beyond summarized_lsn is irrelevant now.
476 : */
477 0 : *pending_lsn = WalSummarizerCtl->summarized_lsn;
478 0 : *summarizer_pid = -1;
479 : }
480 : else
481 : {
482 0 : *pending_lsn = WalSummarizerCtl->pending_lsn;
483 :
484 : /*
485 : * We're not fussed about inexact answers here, since they could
486 : * become stale instantly, so we don't bother taking the lock, but
487 : * make sure that invalid PID values are normalized to -1.
488 : */
489 0 : *summarizer_pid = GetPGProcByNumber(summarizer_pgprocno)->pid;
490 0 : if (*summarizer_pid <= 0)
491 0 : *summarizer_pid = -1;
492 : }
493 : }
494 0 : LWLockRelease(WALSummarizerLock);
495 0 : }
496 :
497 : /*
498 : * Get the oldest LSN in this server's timeline history that has not yet been
499 : * summarized, and update shared memory state as appropriate.
500 : *
501 : * If *tli != NULL, it will be set to the TLI for the LSN that is returned.
502 : *
503 : * If *lsn_is_exact != NULL, it will be set to true if the returned LSN is
504 : * necessarily the start of a WAL record and false if it's just the beginning
505 : * of a WAL segment.
506 : */
507 : XLogRecPtr
508 3458 : GetOldestUnsummarizedLSN(TimeLineID *tli, bool *lsn_is_exact)
509 : {
510 : TimeLineID latest_tli;
511 : int n;
512 : List *tles;
513 3458 : XLogRecPtr unsummarized_lsn = InvalidXLogRecPtr;
514 3458 : TimeLineID unsummarized_tli = 0;
515 3458 : bool should_make_exact = false;
516 : List *existing_summaries;
517 : ListCell *lc;
518 3458 : bool am_wal_summarizer = AmWalSummarizerProcess();
519 :
520 : /* If not summarizing WAL, do nothing. */
521 3458 : if (!summarize_wal)
522 3452 : return InvalidXLogRecPtr;
523 :
524 : /*
525 : * If we are not the WAL summarizer process, then we normally just want to
526 : * read the values from shared memory. However, as an exception, if shared
527 : * memory hasn't been initialized yet, then we need to do that so that we
528 : * can read legal values and not remove any WAL too early.
529 : */
530 6 : if (!am_wal_summarizer)
531 : {
532 4 : LWLockAcquire(WALSummarizerLock, LW_SHARED);
533 :
534 4 : if (WalSummarizerCtl->initialized)
535 : {
536 4 : unsummarized_lsn = WalSummarizerCtl->summarized_lsn;
537 4 : if (tli != NULL)
538 0 : *tli = WalSummarizerCtl->summarized_tli;
539 4 : if (lsn_is_exact != NULL)
540 0 : *lsn_is_exact = WalSummarizerCtl->lsn_is_exact;
541 4 : LWLockRelease(WALSummarizerLock);
542 4 : return unsummarized_lsn;
543 : }
544 :
545 0 : LWLockRelease(WALSummarizerLock);
546 : }
547 :
548 : /*
549 : * Find the oldest timeline on which WAL still exists, and the earliest
550 : * segment for which it exists.
551 : *
552 : * Note that we do this every time the WAL summarizer process restarts or
553 : * recovers from an error, in case the contents of pg_wal have changed
554 : * under us e.g. if some files were removed, either manually - which
555 : * shouldn't really happen, but might - or by postgres itself, if
556 : * summarize_wal was turned off and then back on again.
557 : */
558 2 : (void) GetLatestLSN(&latest_tli);
559 2 : tles = readTimeLineHistory(latest_tli);
560 2 : for (n = list_length(tles) - 1; n >= 0; --n)
561 : {
562 2 : TimeLineHistoryEntry *tle = list_nth(tles, n);
563 : XLogSegNo oldest_segno;
564 :
565 2 : oldest_segno = XLogGetOldestSegno(tle->tli);
566 2 : if (oldest_segno != 0)
567 : {
568 : /* Compute oldest LSN that still exists on disk. */
569 2 : XLogSegNoOffsetToRecPtr(oldest_segno, 0, wal_segment_size,
570 : unsummarized_lsn);
571 :
572 2 : unsummarized_tli = tle->tli;
573 2 : break;
574 : }
575 : }
576 :
577 : /*
578 : * Don't try to summarize anything older than the end LSN of the newest
579 : * summary file that exists for this timeline.
580 : */
581 : existing_summaries =
582 2 : GetWalSummaries(unsummarized_tli,
583 : InvalidXLogRecPtr, InvalidXLogRecPtr);
584 2 : foreach(lc, existing_summaries)
585 : {
586 0 : WalSummaryFile *ws = lfirst(lc);
587 :
588 0 : if (ws->end_lsn > unsummarized_lsn)
589 : {
590 0 : unsummarized_lsn = ws->end_lsn;
591 0 : should_make_exact = true;
592 : }
593 : }
594 :
595 : /* It really should not be possible for us to find no WAL. */
596 2 : if (unsummarized_tli == 0)
597 0 : ereport(ERROR,
598 : errcode(ERRCODE_INTERNAL_ERROR),
599 : errmsg_internal("no WAL found on timeline %u", latest_tli));
600 :
601 : /*
602 : * If we're the WAL summarizer, we always want to store the values we just
603 : * computed into shared memory, because those are the values we're going
604 : * to use to drive our operation, and so they are the authoritative
605 : * values. Otherwise, we only store values into shared memory if shared
606 : * memory is uninitialized. Our values are not canonical in such a case,
607 : * but it's better to have something than nothing, to guide WAL retention.
608 : */
609 2 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
610 2 : if (am_wal_summarizer || !WalSummarizerCtl->initialized)
611 : {
612 2 : WalSummarizerCtl->initialized = true;
613 2 : WalSummarizerCtl->summarized_lsn = unsummarized_lsn;
614 2 : WalSummarizerCtl->summarized_tli = unsummarized_tli;
615 2 : WalSummarizerCtl->lsn_is_exact = should_make_exact;
616 2 : WalSummarizerCtl->pending_lsn = unsummarized_lsn;
617 : }
618 : else
619 0 : unsummarized_lsn = WalSummarizerCtl->summarized_lsn;
620 :
621 : /* Also return the to the caller as required. */
622 2 : if (tli != NULL)
623 2 : *tli = WalSummarizerCtl->summarized_tli;
624 2 : if (lsn_is_exact != NULL)
625 2 : *lsn_is_exact = WalSummarizerCtl->lsn_is_exact;
626 2 : LWLockRelease(WALSummarizerLock);
627 :
628 2 : return unsummarized_lsn;
629 : }
630 :
631 : /*
632 : * Wake up the WAL summarizer process.
633 : *
634 : * This might not work, because there's no guarantee that the WAL summarizer
635 : * process was successfully started, and it also might have started but
636 : * subsequently terminated. So, under normal circumstances, this will get the
637 : * latch set, but there's no guarantee.
638 : */
639 : void
640 2228 : WakeupWalSummarizer(void)
641 : {
642 : ProcNumber pgprocno;
643 :
644 2228 : if (WalSummarizerCtl == NULL)
645 0 : return;
646 :
647 2228 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
648 2228 : pgprocno = WalSummarizerCtl->summarizer_pgprocno;
649 2228 : LWLockRelease(WALSummarizerLock);
650 :
651 2228 : if (pgprocno != INVALID_PROC_NUMBER)
652 2 : SetLatch(&ProcGlobal->allProcs[pgprocno].procLatch);
653 : }
654 :
655 : /*
656 : * Wait until WAL summarization reaches the given LSN, but time out with an
657 : * error if the summarizer seems to be stick.
658 : *
659 : * Returns immediately if summarize_wal is turned off while we wait. Caller
660 : * is expected to handle this case, if necessary.
661 : */
662 : void
663 24 : WaitForWalSummarization(XLogRecPtr lsn)
664 : {
665 : TimestampTz initial_time,
666 : cycle_time,
667 : current_time;
668 24 : XLogRecPtr prior_pending_lsn = InvalidXLogRecPtr;
669 24 : int deadcycles = 0;
670 :
671 24 : initial_time = cycle_time = GetCurrentTimestamp();
672 :
673 : while (1)
674 16 : {
675 40 : long timeout_in_ms = 10000;
676 : XLogRecPtr summarized_lsn;
677 : XLogRecPtr pending_lsn;
678 :
679 40 : CHECK_FOR_INTERRUPTS();
680 :
681 : /* If WAL summarization is disabled while we're waiting, give up. */
682 40 : if (!summarize_wal)
683 0 : return;
684 :
685 : /*
686 : * If the LSN summarized on disk has reached the target value, stop.
687 : */
688 40 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
689 40 : summarized_lsn = WalSummarizerCtl->summarized_lsn;
690 40 : pending_lsn = WalSummarizerCtl->pending_lsn;
691 40 : LWLockRelease(WALSummarizerLock);
692 :
693 : /* If WAL summarization has progressed sufficiently, stop waiting. */
694 40 : if (summarized_lsn >= lsn)
695 24 : break;
696 :
697 : /* Recheck current time. */
698 16 : current_time = GetCurrentTimestamp();
699 :
700 : /* Have we finished the current cycle of waiting? */
701 16 : if (TimestampDifferenceMilliseconds(cycle_time,
702 : current_time) >= timeout_in_ms)
703 : {
704 : long elapsed_seconds;
705 :
706 : /* Begin new wait cycle. */
707 0 : cycle_time = TimestampTzPlusMilliseconds(cycle_time,
708 : timeout_in_ms);
709 :
710 : /*
711 : * Keep track of the number of cycles during which there has been
712 : * no progression of pending_lsn. If pending_lsn is not advancing,
713 : * that means that not only are no new files appearing on disk,
714 : * but we're not even incorporating new records into the in-memory
715 : * state.
716 : */
717 0 : if (pending_lsn > prior_pending_lsn)
718 : {
719 0 : prior_pending_lsn = pending_lsn;
720 0 : deadcycles = 0;
721 : }
722 : else
723 0 : ++deadcycles;
724 :
725 : /*
726 : * If we've managed to wait for an entire minute without the WAL
727 : * summarizer absorbing a single WAL record, error out; probably
728 : * something is wrong.
729 : *
730 : * We could consider also erroring out if the summarizer is taking
731 : * too long to catch up, but it's not clear what rate of progress
732 : * would be acceptable and what would be too slow. So instead, we
733 : * just try to error out in the case where there's no progress at
734 : * all. That seems likely to catch a reasonable number of the
735 : * things that can go wrong in practice (e.g. the summarizer
736 : * process is completely hung, say because somebody hooked up a
737 : * debugger to it or something) without giving up too quickly when
738 : * the system is just slow.
739 : */
740 0 : if (deadcycles >= 6)
741 0 : ereport(ERROR,
742 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
743 : errmsg("WAL summarization is not progressing"),
744 : errdetail("Summarization is needed through %X/%X, but is stuck at %X/%X on disk and %X/%X in memory.",
745 : LSN_FORMAT_ARGS(lsn),
746 : LSN_FORMAT_ARGS(summarized_lsn),
747 : LSN_FORMAT_ARGS(pending_lsn))));
748 :
749 :
750 : /*
751 : * Otherwise, just let the user know what's happening.
752 : */
753 0 : elapsed_seconds =
754 0 : TimestampDifferenceMilliseconds(initial_time,
755 : current_time) / 1000;
756 0 : ereport(WARNING,
757 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
758 : errmsg_plural("still waiting for WAL summarization through %X/%X after %ld second",
759 : "still waiting for WAL summarization through %X/%X after %ld seconds",
760 : elapsed_seconds,
761 : LSN_FORMAT_ARGS(lsn),
762 : elapsed_seconds),
763 : errdetail("Summarization has reached %X/%X on disk and %X/%X in memory.",
764 : LSN_FORMAT_ARGS(summarized_lsn),
765 : LSN_FORMAT_ARGS(pending_lsn))));
766 : }
767 :
768 : /*
769 : * Align the wait time to prevent drift. This doesn't really matter,
770 : * but we'd like the warnings about how long we've been waiting to say
771 : * 10 seconds, 20 seconds, 30 seconds, 40 seconds ... without ever
772 : * drifting to something that is not a multiple of ten.
773 : */
774 16 : timeout_in_ms -=
775 16 : TimestampDifferenceMilliseconds(cycle_time, current_time);
776 :
777 : /* Wait and see. */
778 16 : ConditionVariableTimedSleep(&WalSummarizerCtl->summary_file_cv,
779 : timeout_in_ms,
780 : WAIT_EVENT_WAL_SUMMARY_READY);
781 : }
782 :
783 24 : ConditionVariableCancelSleep();
784 : }
785 :
786 : /*
787 : * On exit, update shared memory to make it clear that we're no longer
788 : * running.
789 : */
790 : static void
791 2 : WalSummarizerShutdown(int code, Datum arg)
792 : {
793 2 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
794 2 : WalSummarizerCtl->summarizer_pgprocno = INVALID_PROC_NUMBER;
795 2 : LWLockRelease(WALSummarizerLock);
796 2 : }
797 :
798 : /*
799 : * Get the latest LSN that is eligible to be summarized, and set *tli to the
800 : * corresponding timeline.
801 : */
802 : static XLogRecPtr
803 32 : GetLatestLSN(TimeLineID *tli)
804 : {
805 32 : if (!RecoveryInProgress())
806 : {
807 : /* Don't summarize WAL before it's flushed. */
808 32 : return GetFlushRecPtr(tli);
809 : }
810 : else
811 : {
812 : XLogRecPtr flush_lsn;
813 : TimeLineID flush_tli;
814 : XLogRecPtr replay_lsn;
815 : TimeLineID replay_tli;
816 : TimeLineID insert_tli;
817 :
818 : /*
819 : * After the insert TLI has been set and before the control file has
820 : * been updated to show the DB in production, RecoveryInProgress()
821 : * will return true, because it's not yet safe for all backends to
822 : * begin writing WAL. However, replay has already ceased, so from our
823 : * point of view, recovery is already over. We should summarize up to
824 : * where replay stopped and then prepare to resume at the start of the
825 : * insert timeline.
826 : */
827 0 : if ((insert_tli = GetWALInsertionTimeLineIfSet()) != 0)
828 : {
829 0 : *tli = insert_tli;
830 0 : return GetXLogReplayRecPtr(NULL);
831 : }
832 :
833 : /*
834 : * What we really want to know is how much WAL has been flushed to
835 : * disk, but the only flush position available is the one provided by
836 : * the walreceiver, which may not be running, because this could be
837 : * crash recovery or recovery via restore_command. So use either the
838 : * WAL receiver's flush position or the replay position, whichever is
839 : * further ahead, on the theory that if the WAL has been replayed then
840 : * it must also have been flushed to disk.
841 : */
842 0 : flush_lsn = GetWalRcvFlushRecPtr(NULL, &flush_tli);
843 0 : replay_lsn = GetXLogReplayRecPtr(&replay_tli);
844 0 : if (flush_lsn > replay_lsn)
845 : {
846 0 : *tli = flush_tli;
847 0 : return flush_lsn;
848 : }
849 : else
850 : {
851 0 : *tli = replay_tli;
852 0 : return replay_lsn;
853 : }
854 : }
855 : }
856 :
857 : /*
858 : * Interrupt handler for main loop of WAL summarizer process.
859 : */
860 : static void
861 50472 : ProcessWalSummarizerInterrupts(void)
862 : {
863 50472 : if (ProcSignalBarrierPending)
864 0 : ProcessProcSignalBarrier();
865 :
866 50472 : if (ConfigReloadPending)
867 : {
868 0 : ConfigReloadPending = false;
869 0 : ProcessConfigFile(PGC_SIGHUP);
870 : }
871 :
872 50472 : if (ShutdownRequestPending || !summarize_wal)
873 : {
874 2 : ereport(DEBUG1,
875 : errmsg_internal("WAL summarizer shutting down"));
876 2 : proc_exit(0);
877 : }
878 :
879 : /* Perform logging of memory contexts of this process */
880 50470 : if (LogMemoryContextPending)
881 0 : ProcessLogMemoryContextInterrupt();
882 50470 : }
883 :
884 : /*
885 : * Summarize a range of WAL records on a single timeline.
886 : *
887 : * 'tli' is the timeline to be summarized.
888 : *
889 : * 'start_lsn' is the point at which we should start summarizing. If this
890 : * value comes from the end LSN of the previous record as returned by the
891 : * xlogreader machinery, 'exact' should be true; otherwise, 'exact' should
892 : * be false, and this function will search forward for the start of a valid
893 : * WAL record.
894 : *
895 : * 'switch_lsn' is the point at which we should switch to a later timeline,
896 : * if we're summarizing a historic timeline.
897 : *
898 : * 'maximum_lsn' identifies the point beyond which we can't count on being
899 : * able to read any more WAL. It should be the switch point when reading a
900 : * historic timeline, or the most-recently-measured end of WAL when reading
901 : * the current timeline.
902 : *
903 : * The return value is the LSN at which the WAL summary actually ends. Most
904 : * often, a summary file ends because we notice that a checkpoint has
905 : * occurred and reach the redo pointer of that checkpoint, but sometimes
906 : * we stop for other reasons, such as a timeline switch.
907 : */
908 : static XLogRecPtr
909 18 : SummarizeWAL(TimeLineID tli, XLogRecPtr start_lsn, bool exact,
910 : XLogRecPtr switch_lsn, XLogRecPtr maximum_lsn)
911 : {
912 : SummarizerReadLocalXLogPrivate *private_data;
913 : XLogReaderState *xlogreader;
914 : XLogRecPtr summary_start_lsn;
915 18 : XLogRecPtr summary_end_lsn = switch_lsn;
916 : char temp_path[MAXPGPATH];
917 : char final_path[MAXPGPATH];
918 : WalSummaryIO io;
919 18 : BlockRefTable *brtab = CreateEmptyBlockRefTable();
920 18 : bool fast_forward = true;
921 :
922 : /* Initialize private data for xlogreader. */
923 : private_data = (SummarizerReadLocalXLogPrivate *)
924 18 : palloc0(sizeof(SummarizerReadLocalXLogPrivate));
925 18 : private_data->tli = tli;
926 18 : private_data->historic = !XLogRecPtrIsInvalid(switch_lsn);
927 18 : private_data->read_upto = maximum_lsn;
928 :
929 : /* Create xlogreader. */
930 18 : xlogreader = XLogReaderAllocate(wal_segment_size, NULL,
931 18 : XL_ROUTINE(.page_read = &summarizer_read_local_xlog_page,
932 : .segment_open = &wal_segment_open,
933 : .segment_close = &wal_segment_close),
934 : private_data);
935 18 : if (xlogreader == NULL)
936 0 : ereport(ERROR,
937 : (errcode(ERRCODE_OUT_OF_MEMORY),
938 : errmsg("out of memory"),
939 : errdetail("Failed while allocating a WAL reading processor.")));
940 :
941 : /*
942 : * When exact = false, we're starting from an arbitrary point in the WAL
943 : * and must search forward for the start of the next record.
944 : *
945 : * When exact = true, start_lsn should be either the LSN where a record
946 : * begins, or the LSN of a page where the page header is immediately
947 : * followed by the start of a new record. XLogBeginRead should tolerate
948 : * either case.
949 : *
950 : * We need to allow for both cases because the behavior of xlogreader
951 : * varies. When a record spans two or more xlog pages, the ending LSN
952 : * reported by xlogreader will be the starting LSN of the following
953 : * record, but when an xlog page boundary falls between two records, the
954 : * end LSN for the first will be reported as the first byte of the
955 : * following page. We can't know until we read that page how large the
956 : * header will be, but we'll have to skip over it to find the next record.
957 : */
958 18 : if (exact)
959 : {
960 : /*
961 : * Even if start_lsn is the beginning of a page rather than the
962 : * beginning of the first record on that page, we should still use it
963 : * as the start LSN for the summary file. That's because we detect
964 : * missing summary files by looking for cases where the end LSN of one
965 : * file is less than the start LSN of the next file. When only a page
966 : * header is skipped, nothing has been missed.
967 : */
968 16 : XLogBeginRead(xlogreader, start_lsn);
969 16 : summary_start_lsn = start_lsn;
970 : }
971 : else
972 : {
973 2 : summary_start_lsn = XLogFindNextRecord(xlogreader, start_lsn);
974 2 : if (XLogRecPtrIsInvalid(summary_start_lsn))
975 : {
976 : /*
977 : * If we hit end-of-WAL while trying to find the next valid
978 : * record, we must be on a historic timeline that has no valid
979 : * records that begin after start_lsn and before end of WAL.
980 : */
981 0 : if (private_data->end_of_wal)
982 : {
983 0 : ereport(DEBUG1,
984 : errmsg_internal("could not read WAL from timeline %u at %X/%X: end of WAL at %X/%X",
985 : tli,
986 : LSN_FORMAT_ARGS(start_lsn),
987 : LSN_FORMAT_ARGS(private_data->read_upto)));
988 :
989 : /*
990 : * The timeline ends at or after start_lsn, without containing
991 : * any records. Thus, we must make sure the main loop does not
992 : * iterate. If start_lsn is the end of the timeline, then we
993 : * won't actually emit an empty summary file, but otherwise,
994 : * we must, to capture the fact that the LSN range in question
995 : * contains no interesting WAL records.
996 : */
997 0 : summary_start_lsn = start_lsn;
998 0 : summary_end_lsn = private_data->read_upto;
999 0 : switch_lsn = xlogreader->EndRecPtr;
1000 : }
1001 : else
1002 0 : ereport(ERROR,
1003 : (errmsg("could not find a valid record after %X/%X",
1004 : LSN_FORMAT_ARGS(start_lsn))));
1005 : }
1006 :
1007 : /* We shouldn't go backward. */
1008 : Assert(summary_start_lsn >= start_lsn);
1009 : }
1010 :
1011 : /*
1012 : * Main loop: read xlog records one by one.
1013 : */
1014 : while (1)
1015 48470 : {
1016 : int block_id;
1017 : char *errormsg;
1018 : XLogRecord *record;
1019 : uint8 rmid;
1020 :
1021 48488 : ProcessWalSummarizerInterrupts();
1022 :
1023 : /* We shouldn't go backward. */
1024 : Assert(summary_start_lsn <= xlogreader->EndRecPtr);
1025 :
1026 : /* Now read the next record. */
1027 48488 : record = XLogReadRecord(xlogreader, &errormsg);
1028 48486 : if (record == NULL)
1029 : {
1030 0 : if (private_data->end_of_wal)
1031 : {
1032 : /*
1033 : * This timeline must be historic and must end before we were
1034 : * able to read a complete record.
1035 : */
1036 0 : ereport(DEBUG1,
1037 : errmsg_internal("could not read WAL from timeline %u at %X/%X: end of WAL at %X/%X",
1038 : tli,
1039 : LSN_FORMAT_ARGS(xlogreader->EndRecPtr),
1040 : LSN_FORMAT_ARGS(private_data->read_upto)));
1041 : /* Summary ends at end of WAL. */
1042 0 : summary_end_lsn = private_data->read_upto;
1043 0 : break;
1044 : }
1045 0 : if (errormsg)
1046 0 : ereport(ERROR,
1047 : (errcode_for_file_access(),
1048 : errmsg("could not read WAL from timeline %u at %X/%X: %s",
1049 : tli, LSN_FORMAT_ARGS(xlogreader->EndRecPtr),
1050 : errormsg)));
1051 : else
1052 0 : ereport(ERROR,
1053 : (errcode_for_file_access(),
1054 : errmsg("could not read WAL from timeline %u at %X/%X",
1055 : tli, LSN_FORMAT_ARGS(xlogreader->EndRecPtr))));
1056 : }
1057 :
1058 : /* We shouldn't go backward. */
1059 : Assert(summary_start_lsn <= xlogreader->EndRecPtr);
1060 :
1061 48486 : if (!XLogRecPtrIsInvalid(switch_lsn) &&
1062 0 : xlogreader->ReadRecPtr >= switch_lsn)
1063 : {
1064 : /*
1065 : * Whoops! We've read a record that *starts* after the switch LSN,
1066 : * contrary to our goal of reading only until we hit the first
1067 : * record that ends at or after the switch LSN. Pretend we didn't
1068 : * read it after all by bailing out of this loop right here,
1069 : * before we do anything with this record.
1070 : *
1071 : * This can happen because the last record before the switch LSN
1072 : * might be continued across multiple pages, and then we might
1073 : * come to a page with XLP_FIRST_IS_OVERWRITE_CONTRECORD set. In
1074 : * that case, the record that was continued across multiple pages
1075 : * is incomplete and will be disregarded, and the read will
1076 : * restart from the beginning of the page that is flagged
1077 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD.
1078 : *
1079 : * If this case occurs, we can fairly say that the current summary
1080 : * file ends at the switch LSN exactly. The first record on the
1081 : * page marked XLP_FIRST_IS_OVERWRITE_CONTRECORD will be
1082 : * discovered when generating the next summary file.
1083 : */
1084 0 : summary_end_lsn = switch_lsn;
1085 0 : break;
1086 : }
1087 :
1088 : /*
1089 : * Certain types of records require special handling. Redo points and
1090 : * shutdown checkpoints trigger creation of new summary files and can
1091 : * also cause us to enter or exit "fast forward" mode. Other types of
1092 : * records can require special updates to the block reference table.
1093 : */
1094 48486 : rmid = XLogRecGetRmid(xlogreader);
1095 48486 : if (rmid == RM_XLOG_ID)
1096 : {
1097 : bool new_fast_forward;
1098 :
1099 : /*
1100 : * If we've already processed some WAL records when we hit a redo
1101 : * point or shutdown checkpoint, then we stop summarization before
1102 : * including this record in the current file, so that it will be
1103 : * the first record in the next file.
1104 : *
1105 : * When we hit one of those record types as the first record in a
1106 : * file, we adjust our notion of whether we're fast-forwarding.
1107 : * Any WAL generated with wal_level=minimal must be skipped
1108 : * without actually generating any summary file, because an
1109 : * incremental backup that crosses such WAL would be unsafe.
1110 : */
1111 676 : if (SummarizeXlogRecord(xlogreader, &new_fast_forward))
1112 : {
1113 34 : if (xlogreader->ReadRecPtr > summary_start_lsn)
1114 : {
1115 16 : summary_end_lsn = xlogreader->ReadRecPtr;
1116 16 : break;
1117 : }
1118 : else
1119 18 : fast_forward = new_fast_forward;
1120 : }
1121 : }
1122 47810 : else if (!fast_forward)
1123 : {
1124 : /*
1125 : * This switch handles record types that require extra updates to
1126 : * the contents of the block reference table.
1127 : */
1128 47810 : switch (rmid)
1129 : {
1130 4 : case RM_DBASE_ID:
1131 4 : SummarizeDbaseRecord(xlogreader, brtab);
1132 4 : break;
1133 30 : case RM_SMGR_ID:
1134 30 : SummarizeSmgrRecord(xlogreader, brtab);
1135 30 : break;
1136 1496 : case RM_XACT_ID:
1137 1496 : SummarizeXactRecord(xlogreader, brtab);
1138 1496 : break;
1139 : }
1140 0 : }
1141 :
1142 : /*
1143 : * If we're in fast-forward mode, we don't really need to do anything.
1144 : * Otherwise, feed block references from xlog record to block
1145 : * reference table.
1146 : */
1147 48470 : if (!fast_forward)
1148 : {
1149 96532 : for (block_id = 0; block_id <= XLogRecMaxBlockId(xlogreader);
1150 48062 : block_id++)
1151 : {
1152 : RelFileLocator rlocator;
1153 : ForkNumber forknum;
1154 : BlockNumber blocknum;
1155 :
1156 48062 : if (!XLogRecGetBlockTagExtended(xlogreader, block_id, &rlocator,
1157 : &forknum, &blocknum, NULL))
1158 40 : continue;
1159 :
1160 : /*
1161 : * As we do elsewhere, ignore the FSM fork, because it's not
1162 : * fully WAL-logged.
1163 : */
1164 48022 : if (forknum != FSM_FORKNUM)
1165 47740 : BlockRefTableMarkBlockModified(brtab, &rlocator, forknum,
1166 : blocknum);
1167 : }
1168 : }
1169 :
1170 : /* Update our notion of where this summary file ends. */
1171 48470 : summary_end_lsn = xlogreader->EndRecPtr;
1172 :
1173 : /* Also update shared memory. */
1174 48470 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
1175 : Assert(summary_end_lsn >= WalSummarizerCtl->summarized_lsn);
1176 48470 : WalSummarizerCtl->pending_lsn = summary_end_lsn;
1177 48470 : LWLockRelease(WALSummarizerLock);
1178 :
1179 : /*
1180 : * If we have a switch LSN and have reached it, stop before reading
1181 : * the next record.
1182 : */
1183 48470 : if (!XLogRecPtrIsInvalid(switch_lsn) &&
1184 0 : xlogreader->EndRecPtr >= switch_lsn)
1185 0 : break;
1186 : }
1187 :
1188 : /* Destroy xlogreader. */
1189 16 : pfree(xlogreader->private_data);
1190 16 : XLogReaderFree(xlogreader);
1191 :
1192 : /*
1193 : * If a timeline switch occurs, we may fail to make any progress at all
1194 : * before exiting the loop above. If that happens, we don't write a WAL
1195 : * summary file at all. We can also skip writing a file if we're in
1196 : * fast-forward mode.
1197 : */
1198 16 : if (summary_end_lsn > summary_start_lsn && !fast_forward)
1199 : {
1200 : /* Generate temporary and final path name. */
1201 16 : snprintf(temp_path, MAXPGPATH,
1202 : XLOGDIR "/summaries/temp.summary");
1203 16 : snprintf(final_path, MAXPGPATH,
1204 : XLOGDIR "/summaries/%08X%08X%08X%08X%08X.summary",
1205 : tli,
1206 16 : LSN_FORMAT_ARGS(summary_start_lsn),
1207 16 : LSN_FORMAT_ARGS(summary_end_lsn));
1208 :
1209 : /* Open the temporary file for writing. */
1210 16 : io.filepos = 0;
1211 16 : io.file = PathNameOpenFile(temp_path, O_WRONLY | O_CREAT | O_TRUNC);
1212 16 : if (io.file < 0)
1213 0 : ereport(ERROR,
1214 : (errcode_for_file_access(),
1215 : errmsg("could not create file \"%s\": %m", temp_path)));
1216 :
1217 : /* Write the data. */
1218 16 : WriteBlockRefTable(brtab, WriteWalSummary, &io);
1219 :
1220 : /* Close temporary file and shut down xlogreader. */
1221 16 : FileClose(io.file);
1222 :
1223 : /* Tell the user what we did. */
1224 16 : ereport(DEBUG1,
1225 : errmsg_internal("summarized WAL on TLI %u from %X/%X to %X/%X",
1226 : tli,
1227 : LSN_FORMAT_ARGS(summary_start_lsn),
1228 : LSN_FORMAT_ARGS(summary_end_lsn)));
1229 :
1230 : /* Durably rename the new summary into place. */
1231 16 : durable_rename(temp_path, final_path, ERROR);
1232 : }
1233 :
1234 : /* If we skipped a non-zero amount of WAL, log a debug message. */
1235 16 : if (summary_end_lsn > summary_start_lsn && fast_forward)
1236 0 : ereport(DEBUG1,
1237 : errmsg_internal("skipped summarizing WAL on TLI %u from %X/%X to %X/%X",
1238 : tli,
1239 : LSN_FORMAT_ARGS(summary_start_lsn),
1240 : LSN_FORMAT_ARGS(summary_end_lsn)));
1241 :
1242 16 : return summary_end_lsn;
1243 : }
1244 :
1245 : /*
1246 : * Special handling for WAL records with RM_DBASE_ID.
1247 : */
1248 : static void
1249 4 : SummarizeDbaseRecord(XLogReaderState *xlogreader, BlockRefTable *brtab)
1250 : {
1251 4 : uint8 info = XLogRecGetInfo(xlogreader) & ~XLR_INFO_MASK;
1252 :
1253 : /*
1254 : * We use relfilenode zero for a given database OID and tablespace OID to
1255 : * indicate that all relations with that pair of IDs have been recreated
1256 : * if they exist at all. Effectively, we're setting a limit block of 0 for
1257 : * all such relfilenodes.
1258 : *
1259 : * Technically, this special handling is only needed in the case of
1260 : * XLOG_DBASE_CREATE_FILE_COPY, because that can create a whole bunch of
1261 : * relation files in a directory without logging anything specific to each
1262 : * one. If we didn't mark the whole DB OID/TS OID combination in some way,
1263 : * then a tablespace that was dropped after the reference backup and
1264 : * recreated using the FILE_COPY method prior to the incremental backup
1265 : * would look just like one that was never touched at all, which would be
1266 : * catastrophic.
1267 : *
1268 : * But it seems best to adopt this treatment for all records that drop or
1269 : * create a DB OID/TS OID combination. That's similar to how we treat the
1270 : * limit block for individual relations, and it's an extra layer of safety
1271 : * here. We can never lose data by marking more stuff as needing to be
1272 : * backed up in full.
1273 : */
1274 4 : if (info == XLOG_DBASE_CREATE_FILE_COPY)
1275 : {
1276 : xl_dbase_create_file_copy_rec *xlrec;
1277 : RelFileLocator rlocator;
1278 :
1279 4 : xlrec =
1280 4 : (xl_dbase_create_file_copy_rec *) XLogRecGetData(xlogreader);
1281 4 : rlocator.spcOid = xlrec->tablespace_id;
1282 4 : rlocator.dbOid = xlrec->db_id;
1283 4 : rlocator.relNumber = 0;
1284 4 : BlockRefTableSetLimitBlock(brtab, &rlocator, MAIN_FORKNUM, 0);
1285 : }
1286 0 : else if (info == XLOG_DBASE_CREATE_WAL_LOG)
1287 : {
1288 : xl_dbase_create_wal_log_rec *xlrec;
1289 : RelFileLocator rlocator;
1290 :
1291 0 : xlrec = (xl_dbase_create_wal_log_rec *) XLogRecGetData(xlogreader);
1292 0 : rlocator.spcOid = xlrec->tablespace_id;
1293 0 : rlocator.dbOid = xlrec->db_id;
1294 0 : rlocator.relNumber = 0;
1295 0 : BlockRefTableSetLimitBlock(brtab, &rlocator, MAIN_FORKNUM, 0);
1296 : }
1297 0 : else if (info == XLOG_DBASE_DROP)
1298 : {
1299 : xl_dbase_drop_rec *xlrec;
1300 : RelFileLocator rlocator;
1301 : int i;
1302 :
1303 0 : xlrec = (xl_dbase_drop_rec *) XLogRecGetData(xlogreader);
1304 0 : rlocator.dbOid = xlrec->db_id;
1305 0 : rlocator.relNumber = 0;
1306 0 : for (i = 0; i < xlrec->ntablespaces; ++i)
1307 : {
1308 0 : rlocator.spcOid = xlrec->tablespace_ids[i];
1309 0 : BlockRefTableSetLimitBlock(brtab, &rlocator, MAIN_FORKNUM, 0);
1310 : }
1311 : }
1312 4 : }
1313 :
1314 : /*
1315 : * Special handling for WAL records with RM_SMGR_ID.
1316 : */
1317 : static void
1318 30 : SummarizeSmgrRecord(XLogReaderState *xlogreader, BlockRefTable *brtab)
1319 : {
1320 30 : uint8 info = XLogRecGetInfo(xlogreader) & ~XLR_INFO_MASK;
1321 :
1322 30 : if (info == XLOG_SMGR_CREATE)
1323 : {
1324 : xl_smgr_create *xlrec;
1325 :
1326 : /*
1327 : * If a new relation fork is created on disk, there is no point
1328 : * tracking anything about which blocks have been modified, because
1329 : * the whole thing will be new. Hence, set the limit block for this
1330 : * fork to 0.
1331 : *
1332 : * Ignore the FSM fork, which is not fully WAL-logged.
1333 : */
1334 30 : xlrec = (xl_smgr_create *) XLogRecGetData(xlogreader);
1335 :
1336 30 : if (xlrec->forkNum != FSM_FORKNUM)
1337 30 : BlockRefTableSetLimitBlock(brtab, &xlrec->rlocator,
1338 : xlrec->forkNum, 0);
1339 : }
1340 0 : else if (info == XLOG_SMGR_TRUNCATE)
1341 : {
1342 : xl_smgr_truncate *xlrec;
1343 :
1344 0 : xlrec = (xl_smgr_truncate *) XLogRecGetData(xlogreader);
1345 :
1346 : /*
1347 : * If a relation fork is truncated on disk, there is no point in
1348 : * tracking anything about block modifications beyond the truncation
1349 : * point.
1350 : *
1351 : * We ignore SMGR_TRUNCATE_FSM here because the FSM isn't fully
1352 : * WAL-logged and thus we can't track modified blocks for it anyway.
1353 : */
1354 0 : if ((xlrec->flags & SMGR_TRUNCATE_HEAP) != 0)
1355 0 : BlockRefTableSetLimitBlock(brtab, &xlrec->rlocator,
1356 : MAIN_FORKNUM, xlrec->blkno);
1357 0 : if ((xlrec->flags & SMGR_TRUNCATE_VM) != 0)
1358 0 : BlockRefTableSetLimitBlock(brtab, &xlrec->rlocator,
1359 : VISIBILITYMAP_FORKNUM, xlrec->blkno);
1360 : }
1361 30 : }
1362 :
1363 : /*
1364 : * Special handling for WAL records with RM_XACT_ID.
1365 : */
1366 : static void
1367 1496 : SummarizeXactRecord(XLogReaderState *xlogreader, BlockRefTable *brtab)
1368 : {
1369 1496 : uint8 info = XLogRecGetInfo(xlogreader) & ~XLR_INFO_MASK;
1370 1496 : uint8 xact_info = info & XLOG_XACT_OPMASK;
1371 :
1372 1496 : if (xact_info == XLOG_XACT_COMMIT ||
1373 : xact_info == XLOG_XACT_COMMIT_PREPARED)
1374 1496 : {
1375 1496 : xl_xact_commit *xlrec = (xl_xact_commit *) XLogRecGetData(xlogreader);
1376 : xl_xact_parsed_commit parsed;
1377 : int i;
1378 :
1379 : /*
1380 : * Don't track modified blocks for any relations that were removed on
1381 : * commit.
1382 : */
1383 1496 : ParseCommitRecord(XLogRecGetInfo(xlogreader), xlrec, &parsed);
1384 1496 : for (i = 0; i < parsed.nrels; ++i)
1385 : {
1386 : ForkNumber forknum;
1387 :
1388 0 : for (forknum = 0; forknum <= MAX_FORKNUM; ++forknum)
1389 0 : if (forknum != FSM_FORKNUM)
1390 0 : BlockRefTableSetLimitBlock(brtab, &parsed.xlocators[i],
1391 : forknum, 0);
1392 : }
1393 : }
1394 0 : else if (xact_info == XLOG_XACT_ABORT ||
1395 : xact_info == XLOG_XACT_ABORT_PREPARED)
1396 : {
1397 0 : xl_xact_abort *xlrec = (xl_xact_abort *) XLogRecGetData(xlogreader);
1398 : xl_xact_parsed_abort parsed;
1399 : int i;
1400 :
1401 : /*
1402 : * Don't track modified blocks for any relations that were removed on
1403 : * abort.
1404 : */
1405 0 : ParseAbortRecord(XLogRecGetInfo(xlogreader), xlrec, &parsed);
1406 0 : for (i = 0; i < parsed.nrels; ++i)
1407 : {
1408 : ForkNumber forknum;
1409 :
1410 0 : for (forknum = 0; forknum <= MAX_FORKNUM; ++forknum)
1411 0 : if (forknum != FSM_FORKNUM)
1412 0 : BlockRefTableSetLimitBlock(brtab, &parsed.xlocators[i],
1413 : forknum, 0);
1414 : }
1415 : }
1416 1496 : }
1417 :
1418 : /*
1419 : * Special handling for WAL records with RM_XLOG_ID.
1420 : *
1421 : * The return value is true if WAL summarization should stop before this
1422 : * record and false otherwise. When the return value is true,
1423 : * *new_fast_forward indicates whether future processing should be done
1424 : * in fast forward mode (i.e. read WAL without emitting summaries) or not.
1425 : */
1426 : static bool
1427 676 : SummarizeXlogRecord(XLogReaderState *xlogreader, bool *new_fast_forward)
1428 : {
1429 676 : uint8 info = XLogRecGetInfo(xlogreader) & ~XLR_INFO_MASK;
1430 : int record_wal_level;
1431 :
1432 676 : if (info == XLOG_CHECKPOINT_REDO)
1433 : {
1434 : /* Payload is wal_level at the time record was written. */
1435 20 : memcpy(&record_wal_level, XLogRecGetData(xlogreader), sizeof(int));
1436 : }
1437 656 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
1438 : {
1439 : CheckPoint rec_ckpt;
1440 :
1441 : /* Extract wal_level at time record was written from payload. */
1442 10 : memcpy(&rec_ckpt, XLogRecGetData(xlogreader), sizeof(CheckPoint));
1443 10 : record_wal_level = rec_ckpt.wal_level;
1444 : }
1445 646 : else if (info == XLOG_PARAMETER_CHANGE)
1446 : {
1447 : xl_parameter_change xlrec;
1448 :
1449 : /* Extract wal_level at time record was written from payload. */
1450 4 : memcpy(&xlrec, XLogRecGetData(xlogreader),
1451 : sizeof(xl_parameter_change));
1452 4 : record_wal_level = xlrec.wal_level;
1453 : }
1454 642 : else if (info == XLOG_END_OF_RECOVERY)
1455 : {
1456 : xl_end_of_recovery xlrec;
1457 :
1458 : /* Extract wal_level at time record was written from payload. */
1459 0 : memcpy(&xlrec, XLogRecGetData(xlogreader), sizeof(xl_end_of_recovery));
1460 0 : record_wal_level = xlrec.wal_level;
1461 : }
1462 : else
1463 : {
1464 : /* No special handling required. Return false. */
1465 642 : return false;
1466 : }
1467 :
1468 : /*
1469 : * Redo can only begin at an XLOG_CHECKPOINT_REDO or
1470 : * XLOG_CHECKPOINT_SHUTDOWN record, so we want WAL summarization to begin
1471 : * at those points. Hence, when those records are encountered, return
1472 : * true, so that we stop just before summarizing either of those records.
1473 : *
1474 : * We also reach here if we just saw XLOG_END_OF_RECOVERY or
1475 : * XLOG_PARAMETER_CHANGE. These are not places where recovery can start,
1476 : * but they're still relevant here. A new timeline can begin with
1477 : * XLOG_END_OF_RECOVERY, so we need to confirm the WAL level at that
1478 : * point; and a restart can provoke XLOG_PARAMETER_CHANGE after an
1479 : * intervening change to postgresql.conf, which might force us to stop
1480 : * summarizing.
1481 : */
1482 34 : *new_fast_forward = (record_wal_level == WAL_LEVEL_MINIMAL);
1483 34 : return true;
1484 : }
1485 :
1486 : /*
1487 : * Similar to read_local_xlog_page, but limited to read from one particular
1488 : * timeline. If the end of WAL is reached, it will wait for more if reading
1489 : * from the current timeline, or give up if reading from a historic timeline.
1490 : * In the latter case, it will also set private_data->end_of_wal = true.
1491 : *
1492 : * Caller must set private_data->tli to the TLI of interest,
1493 : * private_data->read_upto to the lowest LSN that is not known to be safe
1494 : * to read on that timeline, and private_data->historic to true if and only
1495 : * if the timeline is not the current timeline. This function will update
1496 : * private_data->read_upto and private_data->historic if more WAL appears
1497 : * on the current timeline or if the current timeline becomes historic.
1498 : */
1499 : static int
1500 1934 : summarizer_read_local_xlog_page(XLogReaderState *state,
1501 : XLogRecPtr targetPagePtr, int reqLen,
1502 : XLogRecPtr targetRecPtr, char *cur_page)
1503 : {
1504 : int count;
1505 : WALReadError errinfo;
1506 : SummarizerReadLocalXLogPrivate *private_data;
1507 :
1508 1934 : ProcessWalSummarizerInterrupts();
1509 :
1510 1934 : private_data = (SummarizerReadLocalXLogPrivate *)
1511 : state->private_data;
1512 :
1513 : while (1)
1514 : {
1515 1946 : if (targetPagePtr + XLOG_BLCKSZ <= private_data->read_upto)
1516 : {
1517 : /*
1518 : * more than one block available; read only that block, have
1519 : * caller come back if they need more.
1520 : */
1521 1918 : count = XLOG_BLCKSZ;
1522 1918 : break;
1523 : }
1524 28 : else if (targetPagePtr + reqLen > private_data->read_upto)
1525 : {
1526 : /* We don't seem to have enough data. */
1527 14 : if (private_data->historic)
1528 : {
1529 : /*
1530 : * This is a historic timeline, so there will never be any
1531 : * more data than we have currently.
1532 : */
1533 0 : private_data->end_of_wal = true;
1534 0 : return -1;
1535 : }
1536 : else
1537 : {
1538 : XLogRecPtr latest_lsn;
1539 : TimeLineID latest_tli;
1540 :
1541 : /*
1542 : * This is - or at least was up until very recently - the
1543 : * current timeline, so more data might show up. Delay here
1544 : * so we don't tight-loop.
1545 : */
1546 14 : ProcessWalSummarizerInterrupts();
1547 12 : summarizer_wait_for_wal();
1548 :
1549 : /* Recheck end-of-WAL. */
1550 12 : latest_lsn = GetLatestLSN(&latest_tli);
1551 12 : if (private_data->tli == latest_tli)
1552 : {
1553 : /* Still the current timeline, update max LSN. */
1554 : Assert(latest_lsn >= private_data->read_upto);
1555 12 : private_data->read_upto = latest_lsn;
1556 : }
1557 : else
1558 : {
1559 0 : List *tles = readTimeLineHistory(latest_tli);
1560 : XLogRecPtr switchpoint;
1561 :
1562 : /*
1563 : * The timeline we're scanning is no longer the latest
1564 : * one. Figure out when it ended.
1565 : */
1566 0 : private_data->historic = true;
1567 0 : switchpoint = tliSwitchPoint(private_data->tli, tles,
1568 : NULL);
1569 :
1570 : /*
1571 : * Allow reads up to exactly the switch point.
1572 : *
1573 : * It's possible that this will cause read_upto to move
1574 : * backwards, because we might have been promoted before
1575 : * reaching the end of the previous timeline. In that
1576 : * case, the next loop iteration will likely conclude that
1577 : * we've reached end of WAL.
1578 : */
1579 0 : private_data->read_upto = switchpoint;
1580 :
1581 : /* Debugging output. */
1582 0 : ereport(DEBUG1,
1583 : errmsg_internal("timeline %u became historic, can read up to %X/%X",
1584 : private_data->tli, LSN_FORMAT_ARGS(private_data->read_upto)));
1585 : }
1586 :
1587 : /* Go around and try again. */
1588 : }
1589 : }
1590 : else
1591 : {
1592 : /* enough bytes available to satisfy the request */
1593 14 : count = private_data->read_upto - targetPagePtr;
1594 14 : break;
1595 : }
1596 : }
1597 :
1598 1932 : if (!WALRead(state, cur_page, targetPagePtr, count,
1599 : private_data->tli, &errinfo))
1600 0 : WALReadRaiseError(&errinfo);
1601 :
1602 : /* Track that we read a page, for sleep time calculation. */
1603 1932 : ++pages_read_since_last_sleep;
1604 :
1605 : /* number of valid bytes in the buffer */
1606 1932 : return count;
1607 : }
1608 :
1609 : /*
1610 : * Sleep for long enough that we believe it's likely that more WAL will
1611 : * be available afterwards.
1612 : */
1613 : static void
1614 12 : summarizer_wait_for_wal(void)
1615 : {
1616 12 : if (pages_read_since_last_sleep == 0)
1617 : {
1618 : /*
1619 : * No pages were read since the last sleep, so double the sleep time,
1620 : * but not beyond the maximum allowable value.
1621 : */
1622 4 : sleep_quanta = Min(sleep_quanta * 2, MAX_SLEEP_QUANTA);
1623 : }
1624 8 : else if (pages_read_since_last_sleep > 1)
1625 : {
1626 : /*
1627 : * Multiple pages were read since the last sleep, so reduce the sleep
1628 : * time.
1629 : *
1630 : * A large burst of activity should be able to quickly reduce the
1631 : * sleep time to the minimum, but we don't want a handful of extra WAL
1632 : * records to provoke a strong reaction. We choose to reduce the sleep
1633 : * time by 1 quantum for each page read beyond the first, which is a
1634 : * fairly arbitrary way of trying to be reactive without overreacting.
1635 : */
1636 6 : if (pages_read_since_last_sleep > sleep_quanta - 1)
1637 6 : sleep_quanta = 1;
1638 : else
1639 0 : sleep_quanta -= pages_read_since_last_sleep;
1640 : }
1641 :
1642 : /* Report pending statistics to the cumulative stats system. */
1643 12 : pgstat_report_wal(false);
1644 :
1645 : /* OK, now sleep. */
1646 12 : (void) WaitLatch(MyLatch,
1647 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
1648 : sleep_quanta * MS_PER_SLEEP_QUANTUM,
1649 : WAIT_EVENT_WAL_SUMMARIZER_WAL);
1650 12 : ResetLatch(MyLatch);
1651 :
1652 : /* Reset count of pages read. */
1653 12 : pages_read_since_last_sleep = 0;
1654 12 : }
1655 :
1656 : /*
1657 : * Remove WAL summaries whose mtimes are older than wal_summary_keep_time.
1658 : */
1659 : static void
1660 18 : MaybeRemoveOldWalSummaries(void)
1661 : {
1662 18 : XLogRecPtr redo_pointer = GetRedoRecPtr();
1663 : List *wslist;
1664 : time_t cutoff_time;
1665 :
1666 : /* If WAL summary removal is disabled, don't do anything. */
1667 18 : if (wal_summary_keep_time == 0)
1668 0 : return;
1669 :
1670 : /*
1671 : * If the redo pointer has not advanced, don't do anything.
1672 : *
1673 : * This has the effect that we only try to remove old WAL summary files
1674 : * once per checkpoint cycle.
1675 : */
1676 18 : if (redo_pointer == redo_pointer_at_last_summary_removal)
1677 14 : return;
1678 4 : redo_pointer_at_last_summary_removal = redo_pointer;
1679 :
1680 : /*
1681 : * Files should only be removed if the last modification time precedes the
1682 : * cutoff time we compute here.
1683 : */
1684 4 : cutoff_time = time(NULL) - wal_summary_keep_time * SECS_PER_MINUTE;
1685 :
1686 : /* Get all the summaries that currently exist. */
1687 4 : wslist = GetWalSummaries(0, InvalidXLogRecPtr, InvalidXLogRecPtr);
1688 :
1689 : /* Loop until all summaries have been considered for removal. */
1690 6 : while (wslist != NIL)
1691 : {
1692 : ListCell *lc;
1693 : XLogSegNo oldest_segno;
1694 2 : XLogRecPtr oldest_lsn = InvalidXLogRecPtr;
1695 : TimeLineID selected_tli;
1696 :
1697 2 : ProcessWalSummarizerInterrupts();
1698 :
1699 : /*
1700 : * Pick a timeline for which some summary files still exist on disk,
1701 : * and find the oldest LSN that still exists on disk for that
1702 : * timeline.
1703 : */
1704 2 : selected_tli = ((WalSummaryFile *) linitial(wslist))->tli;
1705 2 : oldest_segno = XLogGetOldestSegno(selected_tli);
1706 2 : if (oldest_segno != 0)
1707 2 : XLogSegNoOffsetToRecPtr(oldest_segno, 0, wal_segment_size,
1708 : oldest_lsn);
1709 :
1710 :
1711 : /* Consider each WAL file on the selected timeline in turn. */
1712 18 : foreach(lc, wslist)
1713 : {
1714 16 : WalSummaryFile *ws = lfirst(lc);
1715 :
1716 16 : ProcessWalSummarizerInterrupts();
1717 :
1718 : /* If it's not on this timeline, it's not time to consider it. */
1719 16 : if (selected_tli != ws->tli)
1720 0 : continue;
1721 :
1722 : /*
1723 : * If the WAL doesn't exist any more, we can remove it if the file
1724 : * modification time is old enough.
1725 : */
1726 16 : if (XLogRecPtrIsInvalid(oldest_lsn) || ws->end_lsn <= oldest_lsn)
1727 0 : RemoveWalSummaryIfOlderThan(ws, cutoff_time);
1728 :
1729 : /*
1730 : * Whether we removed the file or not, we need not consider it
1731 : * again.
1732 : */
1733 16 : wslist = foreach_delete_current(wslist, lc);
1734 16 : pfree(ws);
1735 : }
1736 : }
1737 : }
|
