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) 1996-2024, 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 "postmaster/auxprocess.h"
37 : #include "postmaster/interrupt.h"
38 : #include "postmaster/walsummarizer.h"
39 : #include "replication/walreceiver.h"
40 : #include "storage/fd.h"
41 : #include "storage/ipc.h"
42 : #include "storage/latch.h"
43 : #include "storage/lwlock.h"
44 : #include "storage/proc.h"
45 : #include "storage/procsignal.h"
46 : #include "storage/shmem.h"
47 : #include "utils/guc.h"
48 : #include "utils/memutils.h"
49 : #include "utils/wait_event.h"
50 :
51 : /*
52 : * Data in shared memory related to WAL summarization.
53 : */
54 : typedef struct
55 : {
56 : /*
57 : * These fields are protected by WALSummarizerLock.
58 : *
59 : * Until we've discovered what summary files already exist on disk and
60 : * stored that information in shared memory, initialized is false and the
61 : * other fields here contain no meaningful information. After that has
62 : * been done, initialized is true.
63 : *
64 : * summarized_tli and summarized_lsn indicate the last LSN and TLI at
65 : * which the next summary file will start. Normally, these are the LSN and
66 : * TLI at which the last file ended; in such case, lsn_is_exact is true.
67 : * If, however, the LSN is just an approximation, then lsn_is_exact is
68 : * false. This can happen if, for example, there are no existing WAL
69 : * summary files at startup. In that case, we have to derive the position
70 : * at which to start summarizing from the WAL files that exist on disk,
71 : * and so the LSN might point to the start of the next file even though
72 : * that might happen to be in the middle of a WAL record.
73 : *
74 : * summarizer_pgprocno is the proc number of the summarizer process, if
75 : * one is running, or else INVALID_PROC_NUMBER.
76 : *
77 : * pending_lsn is used by the summarizer to advertise the ending LSN of a
78 : * record it has recently read. It shouldn't ever be less than
79 : * summarized_lsn, but might be greater, because the summarizer buffers
80 : * data for a range of LSNs in memory before writing out a new file.
81 : */
82 : bool initialized;
83 : TimeLineID summarized_tli;
84 : XLogRecPtr summarized_lsn;
85 : bool lsn_is_exact;
86 : ProcNumber summarizer_pgprocno;
87 : XLogRecPtr pending_lsn;
88 :
89 : /*
90 : * This field handles its own synchronization.
91 : */
92 : ConditionVariable summary_file_cv;
93 : } WalSummarizerData;
94 :
95 : /*
96 : * Private data for our xlogreader's page read callback.
97 : */
98 : typedef struct
99 : {
100 : TimeLineID tli;
101 : bool historic;
102 : XLogRecPtr read_upto;
103 : bool end_of_wal;
104 : } SummarizerReadLocalXLogPrivate;
105 :
106 : /* Pointer to shared memory state. */
107 : static WalSummarizerData *WalSummarizerCtl;
108 :
109 : /*
110 : * When we reach end of WAL and need to read more, we sleep for a number of
111 : * milliseconds that is an integer multiple of MS_PER_SLEEP_QUANTUM. This is
112 : * the multiplier. It should vary between 1 and MAX_SLEEP_QUANTA, depending
113 : * on system activity. See summarizer_wait_for_wal() for how we adjust this.
114 : */
115 : static long sleep_quanta = 1;
116 :
117 : /*
118 : * The sleep time will always be a multiple of 200ms and will not exceed
119 : * thirty seconds (150 * 200 = 30 * 1000). Note that the timeout here needs
120 : * to be substantially less than the maximum amount of time for which an
121 : * incremental backup will wait for this process to catch up. Otherwise, an
122 : * incremental backup might time out on an idle system just because we sleep
123 : * for too long.
124 : */
125 : #define MAX_SLEEP_QUANTA 150
126 : #define MS_PER_SLEEP_QUANTUM 200
127 :
128 : /*
129 : * This is a count of the number of pages of WAL that we've read since the
130 : * last time we waited for more WAL to appear.
131 : */
132 : static long pages_read_since_last_sleep = 0;
133 :
134 : /*
135 : * Most recent RedoRecPtr value observed by MaybeRemoveOldWalSummaries.
136 : */
137 : static XLogRecPtr redo_pointer_at_last_summary_removal = InvalidXLogRecPtr;
138 :
139 : /*
140 : * GUC parameters
141 : */
142 : bool summarize_wal = false;
143 : int wal_summary_keep_time = 10 * HOURS_PER_DAY * MINS_PER_HOUR;
144 :
145 : static void WalSummarizerShutdown(int code, Datum arg);
146 : static XLogRecPtr GetLatestLSN(TimeLineID *tli);
147 : static void HandleWalSummarizerInterrupts(void);
148 : static XLogRecPtr SummarizeWAL(TimeLineID tli, XLogRecPtr start_lsn,
149 : bool exact, XLogRecPtr switch_lsn,
150 : XLogRecPtr maximum_lsn);
151 : static void SummarizeDbaseRecord(XLogReaderState *xlogreader,
152 : BlockRefTable *brtab);
153 : static void SummarizeSmgrRecord(XLogReaderState *xlogreader,
154 : BlockRefTable *brtab);
155 : static void SummarizeXactRecord(XLogReaderState *xlogreader,
156 : BlockRefTable *brtab);
157 : static bool SummarizeXlogRecord(XLogReaderState *xlogreader);
158 : static int summarizer_read_local_xlog_page(XLogReaderState *state,
159 : XLogRecPtr targetPagePtr,
160 : int reqLen,
161 : XLogRecPtr targetRecPtr,
162 : char *cur_page);
163 : static void summarizer_wait_for_wal(void);
164 : static void MaybeRemoveOldWalSummaries(void);
165 :
166 : /*
167 : * Amount of shared memory required for this module.
168 : */
169 : Size
170 5066 : WalSummarizerShmemSize(void)
171 : {
172 5066 : return sizeof(WalSummarizerData);
173 : }
174 :
175 : /*
176 : * Create or attach to shared memory segment for this module.
177 : */
178 : void
179 1768 : WalSummarizerShmemInit(void)
180 : {
181 : bool found;
182 :
183 1768 : WalSummarizerCtl = (WalSummarizerData *)
184 1768 : ShmemInitStruct("Wal Summarizer Ctl", WalSummarizerShmemSize(),
185 : &found);
186 :
187 1768 : if (!found)
188 : {
189 : /*
190 : * First time through, so initialize.
191 : *
192 : * We're just filling in dummy values here -- the real initialization
193 : * will happen when GetOldestUnsummarizedLSN() is called for the first
194 : * time.
195 : */
196 1768 : WalSummarizerCtl->initialized = false;
197 1768 : WalSummarizerCtl->summarized_tli = 0;
198 1768 : WalSummarizerCtl->summarized_lsn = InvalidXLogRecPtr;
199 1768 : WalSummarizerCtl->lsn_is_exact = false;
200 1768 : WalSummarizerCtl->summarizer_pgprocno = INVALID_PROC_NUMBER;
201 1768 : WalSummarizerCtl->pending_lsn = InvalidXLogRecPtr;
202 1768 : ConditionVariableInit(&WalSummarizerCtl->summary_file_cv);
203 : }
204 1768 : }
205 :
206 : /*
207 : * Entry point for walsummarizer process.
208 : */
209 : void
210 0 : WalSummarizerMain(char *startup_data, size_t startup_data_len)
211 : {
212 : sigjmp_buf local_sigjmp_buf;
213 : MemoryContext context;
214 :
215 : /*
216 : * Within this function, 'current_lsn' and 'current_tli' refer to the
217 : * point from which the next WAL summary file should start. 'exact' is
218 : * true if 'current_lsn' is known to be the start of a WAL record or WAL
219 : * segment, and false if it might be in the middle of a record someplace.
220 : *
221 : * 'switch_lsn' and 'switch_tli', if set, are the LSN at which we need to
222 : * switch to a new timeline and the timeline to which we need to switch.
223 : * If not set, we either haven't figured out the answers yet or we're
224 : * already on the latest timeline.
225 : */
226 : XLogRecPtr current_lsn;
227 : TimeLineID current_tli;
228 : bool exact;
229 0 : XLogRecPtr switch_lsn = InvalidXLogRecPtr;
230 0 : TimeLineID switch_tli = 0;
231 :
232 : Assert(startup_data_len == 0);
233 :
234 0 : MyBackendType = B_WAL_SUMMARIZER;
235 0 : AuxiliaryProcessMainCommon();
236 :
237 0 : ereport(DEBUG1,
238 : (errmsg_internal("WAL summarizer started")));
239 :
240 : /*
241 : * Properly accept or ignore signals the postmaster might send us
242 : *
243 : * We have no particular use for SIGINT at the moment, but seems
244 : * reasonable to treat like SIGTERM.
245 : */
246 0 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
247 0 : pqsignal(SIGINT, SignalHandlerForShutdownRequest);
248 0 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
249 : /* SIGQUIT handler was already set up by InitPostmasterChild */
250 0 : pqsignal(SIGALRM, SIG_IGN);
251 0 : pqsignal(SIGPIPE, SIG_IGN);
252 0 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
253 0 : pqsignal(SIGUSR2, SIG_IGN); /* not used */
254 :
255 : /* Advertise ourselves. */
256 0 : on_shmem_exit(WalSummarizerShutdown, (Datum) 0);
257 0 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
258 0 : WalSummarizerCtl->summarizer_pgprocno = MyProcNumber;
259 0 : LWLockRelease(WALSummarizerLock);
260 :
261 : /* Create and switch to a memory context that we can reset on error. */
262 0 : context = AllocSetContextCreate(TopMemoryContext,
263 : "Wal Summarizer",
264 : ALLOCSET_DEFAULT_SIZES);
265 0 : MemoryContextSwitchTo(context);
266 :
267 : /*
268 : * Reset some signals that are accepted by postmaster but not here
269 : */
270 0 : pqsignal(SIGCHLD, SIG_DFL);
271 :
272 : /*
273 : * If an exception is encountered, processing resumes here.
274 : */
275 0 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
276 : {
277 : /* Since not using PG_TRY, must reset error stack by hand */
278 0 : error_context_stack = NULL;
279 :
280 : /* Prevent interrupts while cleaning up */
281 0 : HOLD_INTERRUPTS();
282 :
283 : /* Report the error to the server log */
284 0 : EmitErrorReport();
285 :
286 : /* Release resources we might have acquired. */
287 0 : LWLockReleaseAll();
288 0 : ConditionVariableCancelSleep();
289 0 : pgstat_report_wait_end();
290 0 : ReleaseAuxProcessResources(false);
291 0 : AtEOXact_Files(false);
292 0 : AtEOXact_HashTables(false);
293 :
294 : /*
295 : * Now return to normal top-level context and clear ErrorContext for
296 : * next time.
297 : */
298 0 : MemoryContextSwitchTo(context);
299 0 : FlushErrorState();
300 :
301 : /* Flush any leaked data in the top-level context */
302 0 : MemoryContextReset(context);
303 :
304 : /* Now we can allow interrupts again */
305 0 : RESUME_INTERRUPTS();
306 :
307 : /*
308 : * Sleep for 10 seconds before attempting to resume operations in
309 : * order to avoid excessive logging.
310 : *
311 : * Many of the likely error conditions are things that will repeat
312 : * every time. For example, if the WAL can't be read or the summary
313 : * can't be written, only administrator action will cure the problem.
314 : * So a really fast retry time doesn't seem to be especially
315 : * beneficial, and it will clutter the logs.
316 : */
317 0 : (void) WaitLatch(MyLatch,
318 : WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
319 : 10000,
320 : WAIT_EVENT_WAL_SUMMARIZER_ERROR);
321 : }
322 :
323 : /* We can now handle ereport(ERROR) */
324 0 : PG_exception_stack = &local_sigjmp_buf;
325 :
326 : /*
327 : * Unblock signals (they were blocked when the postmaster forked us)
328 : */
329 0 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
330 :
331 : /*
332 : * Fetch information about previous progress from shared memory, and ask
333 : * GetOldestUnsummarizedLSN to reset pending_lsn to summarized_lsn. We
334 : * might be recovering from an error, and if so, pending_lsn might have
335 : * advanced past summarized_lsn, but any WAL we read previously has been
336 : * lost and will need to be reread.
337 : *
338 : * If we discover that WAL summarization is not enabled, just exit.
339 : */
340 0 : current_lsn = GetOldestUnsummarizedLSN(¤t_tli, &exact, true);
341 0 : if (XLogRecPtrIsInvalid(current_lsn))
342 0 : proc_exit(0);
343 :
344 : /*
345 : * Loop forever
346 : */
347 : for (;;)
348 0 : {
349 : XLogRecPtr latest_lsn;
350 : TimeLineID latest_tli;
351 : XLogRecPtr end_of_summary_lsn;
352 :
353 : /* Flush any leaked data in the top-level context */
354 0 : MemoryContextReset(context);
355 :
356 : /* Process any signals received recently. */
357 0 : HandleWalSummarizerInterrupts();
358 :
359 : /* If it's time to remove any old WAL summaries, do that now. */
360 0 : MaybeRemoveOldWalSummaries();
361 :
362 : /* Find the LSN and TLI up to which we can safely summarize. */
363 0 : latest_lsn = GetLatestLSN(&latest_tli);
364 :
365 : /*
366 : * If we're summarizing a historic timeline and we haven't yet
367 : * computed the point at which to switch to the next timeline, do that
368 : * now.
369 : *
370 : * Note that if this is a standby, what was previously the current
371 : * timeline could become historic at any time.
372 : *
373 : * We could try to make this more efficient by caching the results of
374 : * readTimeLineHistory when latest_tli has not changed, but since we
375 : * only have to do this once per timeline switch, we probably wouldn't
376 : * save any significant amount of work in practice.
377 : */
378 0 : if (current_tli != latest_tli && XLogRecPtrIsInvalid(switch_lsn))
379 : {
380 0 : List *tles = readTimeLineHistory(latest_tli);
381 :
382 0 : switch_lsn = tliSwitchPoint(current_tli, tles, &switch_tli);
383 0 : ereport(DEBUG1,
384 : errmsg("switch point from TLI %u to TLI %u is at %X/%X",
385 : current_tli, switch_tli, LSN_FORMAT_ARGS(switch_lsn)));
386 : }
387 :
388 : /*
389 : * If we've reached the switch LSN, we can't summarize anything else
390 : * on this timeline. Switch to the next timeline and go around again.
391 : */
392 0 : if (!XLogRecPtrIsInvalid(switch_lsn) && current_lsn >= switch_lsn)
393 : {
394 0 : current_tli = switch_tli;
395 0 : switch_lsn = InvalidXLogRecPtr;
396 0 : switch_tli = 0;
397 0 : continue;
398 : }
399 :
400 : /* Summarize WAL. */
401 0 : end_of_summary_lsn = SummarizeWAL(current_tli,
402 : current_lsn, exact,
403 : switch_lsn, latest_lsn);
404 : Assert(!XLogRecPtrIsInvalid(end_of_summary_lsn));
405 : Assert(end_of_summary_lsn >= current_lsn);
406 :
407 : /*
408 : * Update state for next loop iteration.
409 : *
410 : * Next summary file should start from exactly where this one ended.
411 : */
412 0 : current_lsn = end_of_summary_lsn;
413 0 : exact = true;
414 :
415 : /* Update state in shared memory. */
416 0 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
417 : Assert(WalSummarizerCtl->pending_lsn <= end_of_summary_lsn);
418 0 : WalSummarizerCtl->summarized_lsn = end_of_summary_lsn;
419 0 : WalSummarizerCtl->summarized_tli = current_tli;
420 0 : WalSummarizerCtl->lsn_is_exact = true;
421 0 : WalSummarizerCtl->pending_lsn = end_of_summary_lsn;
422 0 : LWLockRelease(WALSummarizerLock);
423 :
424 : /* Wake up anyone waiting for more summary files to be written. */
425 0 : ConditionVariableBroadcast(&WalSummarizerCtl->summary_file_cv);
426 : }
427 : }
428 :
429 : /*
430 : * Get information about the state of the WAL summarizer.
431 : */
432 : void
433 0 : GetWalSummarizerState(TimeLineID *summarized_tli, XLogRecPtr *summarized_lsn,
434 : XLogRecPtr *pending_lsn, int *summarizer_pid)
435 : {
436 0 : LWLockAcquire(WALSummarizerLock, LW_SHARED);
437 0 : if (!WalSummarizerCtl->initialized)
438 : {
439 : /*
440 : * If initialized is false, the rest of the structure contents are
441 : * undefined.
442 : */
443 0 : *summarized_tli = 0;
444 0 : *summarized_lsn = InvalidXLogRecPtr;
445 0 : *pending_lsn = InvalidXLogRecPtr;
446 0 : *summarizer_pid = -1;
447 : }
448 : else
449 : {
450 0 : int summarizer_pgprocno = WalSummarizerCtl->summarizer_pgprocno;
451 :
452 0 : *summarized_tli = WalSummarizerCtl->summarized_tli;
453 0 : *summarized_lsn = WalSummarizerCtl->summarized_lsn;
454 0 : if (summarizer_pgprocno == INVALID_PROC_NUMBER)
455 : {
456 : /*
457 : * If the summarizer has exited, the fact that it had processed
458 : * beyond summarized_lsn is irrelevant now.
459 : */
460 0 : *pending_lsn = WalSummarizerCtl->summarized_lsn;
461 0 : *summarizer_pid = -1;
462 : }
463 : else
464 : {
465 0 : *pending_lsn = WalSummarizerCtl->pending_lsn;
466 :
467 : /*
468 : * We're not fussed about inexact answers here, since they could
469 : * become stale instantly, so we don't bother taking the lock, but
470 : * make sure that invalid PID values are normalized to -1.
471 : */
472 0 : *summarizer_pid = GetPGProcByNumber(summarizer_pgprocno)->pid;
473 0 : if (*summarizer_pid <= 0)
474 0 : *summarizer_pid = -1;
475 : }
476 : }
477 0 : LWLockRelease(WALSummarizerLock);
478 0 : }
479 :
480 : /*
481 : * Get the oldest LSN in this server's timeline history that has not yet been
482 : * summarized.
483 : *
484 : * If *tli != NULL, it will be set to the TLI for the LSN that is returned.
485 : *
486 : * If *lsn_is_exact != NULL, it will be set to true if the returned LSN is
487 : * necessarily the start of a WAL record and false if it's just the beginning
488 : * of a WAL segment.
489 : *
490 : * If reset_pending_lsn is true, resets the pending_lsn in shared memory to
491 : * be equal to the summarized_lsn.
492 : */
493 : XLogRecPtr
494 2504 : GetOldestUnsummarizedLSN(TimeLineID *tli, bool *lsn_is_exact,
495 : bool reset_pending_lsn)
496 : {
497 : TimeLineID latest_tli;
498 2504 : LWLockMode mode = reset_pending_lsn ? LW_EXCLUSIVE : LW_SHARED;
499 : int n;
500 : List *tles;
501 2504 : XLogRecPtr unsummarized_lsn = InvalidXLogRecPtr;
502 2504 : TimeLineID unsummarized_tli = 0;
503 2504 : bool should_make_exact = false;
504 : List *existing_summaries;
505 : ListCell *lc;
506 :
507 : /* If not summarizing WAL, do nothing. */
508 2504 : if (!summarize_wal)
509 2504 : return InvalidXLogRecPtr;
510 :
511 : /*
512 : * Unless we need to reset the pending_lsn, we initially acquire the lock
513 : * in shared mode and try to fetch the required information. If we acquire
514 : * in shared mode and find that the data structure hasn't been
515 : * initialized, we reacquire the lock in exclusive mode so that we can
516 : * initialize it. However, if someone else does that first before we get
517 : * the lock, then we can just return the requested information after all.
518 : */
519 : while (1)
520 : {
521 0 : LWLockAcquire(WALSummarizerLock, mode);
522 :
523 0 : if (WalSummarizerCtl->initialized)
524 : {
525 0 : unsummarized_lsn = WalSummarizerCtl->summarized_lsn;
526 0 : if (tli != NULL)
527 0 : *tli = WalSummarizerCtl->summarized_tli;
528 0 : if (lsn_is_exact != NULL)
529 0 : *lsn_is_exact = WalSummarizerCtl->lsn_is_exact;
530 0 : if (reset_pending_lsn)
531 0 : WalSummarizerCtl->pending_lsn =
532 0 : WalSummarizerCtl->summarized_lsn;
533 0 : LWLockRelease(WALSummarizerLock);
534 0 : return unsummarized_lsn;
535 : }
536 :
537 0 : if (mode == LW_EXCLUSIVE)
538 0 : break;
539 :
540 0 : LWLockRelease(WALSummarizerLock);
541 0 : mode = LW_EXCLUSIVE;
542 : }
543 :
544 : /*
545 : * The data structure needs to be initialized, and we are the first to
546 : * obtain the lock in exclusive mode, so it's our job to do that
547 : * initialization.
548 : *
549 : * So, find the oldest timeline on which WAL still exists, and the
550 : * earliest segment for which it exists.
551 : */
552 0 : (void) GetLatestLSN(&latest_tli);
553 0 : tles = readTimeLineHistory(latest_tli);
554 0 : for (n = list_length(tles) - 1; n >= 0; --n)
555 : {
556 0 : TimeLineHistoryEntry *tle = list_nth(tles, n);
557 : XLogSegNo oldest_segno;
558 :
559 0 : oldest_segno = XLogGetOldestSegno(tle->tli);
560 0 : if (oldest_segno != 0)
561 : {
562 : /* Compute oldest LSN that still exists on disk. */
563 0 : XLogSegNoOffsetToRecPtr(oldest_segno, 0, wal_segment_size,
564 : unsummarized_lsn);
565 :
566 0 : unsummarized_tli = tle->tli;
567 0 : break;
568 : }
569 : }
570 :
571 : /* It really should not be possible for us to find no WAL. */
572 0 : if (unsummarized_tli == 0)
573 0 : ereport(ERROR,
574 : errcode(ERRCODE_INTERNAL_ERROR),
575 : errmsg_internal("no WAL found on timeline %u", latest_tli));
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 0 : GetWalSummaries(unsummarized_tli,
583 : InvalidXLogRecPtr, InvalidXLogRecPtr);
584 0 : 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 : /* Update shared memory with the discovered values. */
596 0 : WalSummarizerCtl->initialized = true;
597 0 : WalSummarizerCtl->summarized_lsn = unsummarized_lsn;
598 0 : WalSummarizerCtl->summarized_tli = unsummarized_tli;
599 0 : WalSummarizerCtl->lsn_is_exact = should_make_exact;
600 0 : WalSummarizerCtl->pending_lsn = unsummarized_lsn;
601 :
602 : /* Also return the to the caller as required. */
603 0 : if (tli != NULL)
604 0 : *tli = WalSummarizerCtl->summarized_tli;
605 0 : if (lsn_is_exact != NULL)
606 0 : *lsn_is_exact = WalSummarizerCtl->lsn_is_exact;
607 0 : LWLockRelease(WALSummarizerLock);
608 :
609 0 : return unsummarized_lsn;
610 : }
611 :
612 : /*
613 : * Attempt to set the WAL summarizer's latch.
614 : *
615 : * This might not work, because there's no guarantee that the WAL summarizer
616 : * process was successfully started, and it also might have started but
617 : * subsequently terminated. So, under normal circumstances, this will get the
618 : * latch set, but there's no guarantee.
619 : */
620 : void
621 1628 : SetWalSummarizerLatch(void)
622 : {
623 : ProcNumber pgprocno;
624 :
625 1628 : if (WalSummarizerCtl == NULL)
626 0 : return;
627 :
628 1628 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
629 1628 : pgprocno = WalSummarizerCtl->summarizer_pgprocno;
630 1628 : LWLockRelease(WALSummarizerLock);
631 :
632 1628 : if (pgprocno != INVALID_PROC_NUMBER)
633 0 : SetLatch(&ProcGlobal->allProcs[pgprocno].procLatch);
634 : }
635 :
636 : /*
637 : * Wait until WAL summarization reaches the given LSN, but not longer than
638 : * the given timeout.
639 : *
640 : * The return value is the first still-unsummarized LSN. If it's greater than
641 : * or equal to the passed LSN, then that LSN was reached. If not, we timed out.
642 : *
643 : * Either way, *pending_lsn is set to the value taken from WalSummarizerCtl.
644 : */
645 : XLogRecPtr
646 14 : WaitForWalSummarization(XLogRecPtr lsn, long timeout, XLogRecPtr *pending_lsn)
647 : {
648 14 : TimestampTz start_time = GetCurrentTimestamp();
649 14 : TimestampTz deadline = TimestampTzPlusMilliseconds(start_time, timeout);
650 : XLogRecPtr summarized_lsn;
651 :
652 : Assert(!XLogRecPtrIsInvalid(lsn));
653 : Assert(timeout > 0);
654 :
655 : while (1)
656 8 : {
657 : TimestampTz now;
658 : long remaining_timeout;
659 :
660 : /*
661 : * If the LSN summarized on disk has reached the target value, stop.
662 : */
663 22 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
664 22 : summarized_lsn = WalSummarizerCtl->summarized_lsn;
665 22 : *pending_lsn = WalSummarizerCtl->pending_lsn;
666 22 : LWLockRelease(WALSummarizerLock);
667 22 : if (summarized_lsn >= lsn)
668 14 : break;
669 :
670 : /* Timeout reached? If yes, stop. */
671 8 : now = GetCurrentTimestamp();
672 8 : remaining_timeout = TimestampDifferenceMilliseconds(now, deadline);
673 8 : if (remaining_timeout <= 0)
674 0 : break;
675 :
676 : /* Wait and see. */
677 8 : ConditionVariableTimedSleep(&WalSummarizerCtl->summary_file_cv,
678 : remaining_timeout,
679 : WAIT_EVENT_WAL_SUMMARY_READY);
680 : }
681 :
682 14 : return summarized_lsn;
683 : }
684 :
685 : /*
686 : * On exit, update shared memory to make it clear that we're no longer
687 : * running.
688 : */
689 : static void
690 0 : WalSummarizerShutdown(int code, Datum arg)
691 : {
692 0 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
693 0 : WalSummarizerCtl->summarizer_pgprocno = INVALID_PROC_NUMBER;
694 0 : LWLockRelease(WALSummarizerLock);
695 0 : }
696 :
697 : /*
698 : * Get the latest LSN that is eligible to be summarized, and set *tli to the
699 : * corresponding timeline.
700 : */
701 : static XLogRecPtr
702 0 : GetLatestLSN(TimeLineID *tli)
703 : {
704 0 : if (!RecoveryInProgress())
705 : {
706 : /* Don't summarize WAL before it's flushed. */
707 0 : return GetFlushRecPtr(tli);
708 : }
709 : else
710 : {
711 : XLogRecPtr flush_lsn;
712 : TimeLineID flush_tli;
713 : XLogRecPtr replay_lsn;
714 : TimeLineID replay_tli;
715 :
716 : /*
717 : * What we really want to know is how much WAL has been flushed to
718 : * disk, but the only flush position available is the one provided by
719 : * the walreceiver, which may not be running, because this could be
720 : * crash recovery or recovery via restore_command. So use either the
721 : * WAL receiver's flush position or the replay position, whichever is
722 : * further ahead, on the theory that if the WAL has been replayed then
723 : * it must also have been flushed to disk.
724 : */
725 0 : flush_lsn = GetWalRcvFlushRecPtr(NULL, &flush_tli);
726 0 : replay_lsn = GetXLogReplayRecPtr(&replay_tli);
727 0 : if (flush_lsn > replay_lsn)
728 : {
729 0 : *tli = flush_tli;
730 0 : return flush_lsn;
731 : }
732 : else
733 : {
734 0 : *tli = replay_tli;
735 0 : return replay_lsn;
736 : }
737 : }
738 : }
739 :
740 : /*
741 : * Interrupt handler for main loop of WAL summarizer process.
742 : */
743 : static void
744 0 : HandleWalSummarizerInterrupts(void)
745 : {
746 0 : if (ProcSignalBarrierPending)
747 0 : ProcessProcSignalBarrier();
748 :
749 0 : if (ConfigReloadPending)
750 : {
751 0 : ConfigReloadPending = false;
752 0 : ProcessConfigFile(PGC_SIGHUP);
753 : }
754 :
755 0 : if (ShutdownRequestPending || !summarize_wal)
756 : {
757 0 : ereport(DEBUG1,
758 : errmsg_internal("WAL summarizer shutting down"));
759 0 : proc_exit(0);
760 : }
761 :
762 : /* Perform logging of memory contexts of this process */
763 0 : if (LogMemoryContextPending)
764 0 : ProcessLogMemoryContextInterrupt();
765 0 : }
766 :
767 : /*
768 : * Summarize a range of WAL records on a single timeline.
769 : *
770 : * 'tli' is the timeline to be summarized.
771 : *
772 : * 'start_lsn' is the point at which we should start summarizing. If this
773 : * value comes from the end LSN of the previous record as returned by the
774 : * xlogreader machinery, 'exact' should be true; otherwise, 'exact' should
775 : * be false, and this function will search forward for the start of a valid
776 : * WAL record.
777 : *
778 : * 'switch_lsn' is the point at which we should switch to a later timeline,
779 : * if we're summarizing a historic timeline.
780 : *
781 : * 'maximum_lsn' identifies the point beyond which we can't count on being
782 : * able to read any more WAL. It should be the switch point when reading a
783 : * historic timeline, or the most-recently-measured end of WAL when reading
784 : * the current timeline.
785 : *
786 : * The return value is the LSN at which the WAL summary actually ends. Most
787 : * often, a summary file ends because we notice that a checkpoint has
788 : * occurred and reach the redo pointer of that checkpoint, but sometimes
789 : * we stop for other reasons, such as a timeline switch.
790 : */
791 : static XLogRecPtr
792 0 : SummarizeWAL(TimeLineID tli, XLogRecPtr start_lsn, bool exact,
793 : XLogRecPtr switch_lsn, XLogRecPtr maximum_lsn)
794 : {
795 : SummarizerReadLocalXLogPrivate *private_data;
796 : XLogReaderState *xlogreader;
797 : XLogRecPtr summary_start_lsn;
798 0 : XLogRecPtr summary_end_lsn = switch_lsn;
799 : char temp_path[MAXPGPATH];
800 : char final_path[MAXPGPATH];
801 : WalSummaryIO io;
802 0 : BlockRefTable *brtab = CreateEmptyBlockRefTable();
803 :
804 : /* Initialize private data for xlogreader. */
805 : private_data = (SummarizerReadLocalXLogPrivate *)
806 0 : palloc0(sizeof(SummarizerReadLocalXLogPrivate));
807 0 : private_data->tli = tli;
808 0 : private_data->historic = !XLogRecPtrIsInvalid(switch_lsn);
809 0 : private_data->read_upto = maximum_lsn;
810 :
811 : /* Create xlogreader. */
812 0 : xlogreader = XLogReaderAllocate(wal_segment_size, NULL,
813 0 : XL_ROUTINE(.page_read = &summarizer_read_local_xlog_page,
814 : .segment_open = &wal_segment_open,
815 : .segment_close = &wal_segment_close),
816 : private_data);
817 0 : if (xlogreader == NULL)
818 0 : ereport(ERROR,
819 : (errcode(ERRCODE_OUT_OF_MEMORY),
820 : errmsg("out of memory"),
821 : errdetail("Failed while allocating a WAL reading processor.")));
822 :
823 : /*
824 : * When exact = false, we're starting from an arbitrary point in the WAL
825 : * and must search forward for the start of the next record.
826 : *
827 : * When exact = true, start_lsn should be either the LSN where a record
828 : * begins, or the LSN of a page where the page header is immediately
829 : * followed by the start of a new record. XLogBeginRead should tolerate
830 : * either case.
831 : *
832 : * We need to allow for both cases because the behavior of xlogreader
833 : * varies. When a record spans two or more xlog pages, the ending LSN
834 : * reported by xlogreader will be the starting LSN of the following
835 : * record, but when an xlog page boundary falls between two records, the
836 : * end LSN for the first will be reported as the first byte of the
837 : * following page. We can't know until we read that page how large the
838 : * header will be, but we'll have to skip over it to find the next record.
839 : */
840 0 : if (exact)
841 : {
842 : /*
843 : * Even if start_lsn is the beginning of a page rather than the
844 : * beginning of the first record on that page, we should still use it
845 : * as the start LSN for the summary file. That's because we detect
846 : * missing summary files by looking for cases where the end LSN of one
847 : * file is less than the start LSN of the next file. When only a page
848 : * header is skipped, nothing has been missed.
849 : */
850 0 : XLogBeginRead(xlogreader, start_lsn);
851 0 : summary_start_lsn = start_lsn;
852 : }
853 : else
854 : {
855 0 : summary_start_lsn = XLogFindNextRecord(xlogreader, start_lsn);
856 0 : if (XLogRecPtrIsInvalid(summary_start_lsn))
857 : {
858 : /*
859 : * If we hit end-of-WAL while trying to find the next valid
860 : * record, we must be on a historic timeline that has no valid
861 : * records that begin after start_lsn and before end of WAL.
862 : */
863 0 : if (private_data->end_of_wal)
864 : {
865 0 : ereport(DEBUG1,
866 : errmsg_internal("could not read WAL from timeline %u at %X/%X: end of WAL at %X/%X",
867 : tli,
868 : LSN_FORMAT_ARGS(start_lsn),
869 : LSN_FORMAT_ARGS(private_data->read_upto)));
870 :
871 : /*
872 : * The timeline ends at or after start_lsn, without containing
873 : * any records. Thus, we must make sure the main loop does not
874 : * iterate. If start_lsn is the end of the timeline, then we
875 : * won't actually emit an empty summary file, but otherwise,
876 : * we must, to capture the fact that the LSN range in question
877 : * contains no interesting WAL records.
878 : */
879 0 : summary_start_lsn = start_lsn;
880 0 : summary_end_lsn = private_data->read_upto;
881 0 : switch_lsn = xlogreader->EndRecPtr;
882 : }
883 : else
884 0 : ereport(ERROR,
885 : (errmsg("could not find a valid record after %X/%X",
886 : LSN_FORMAT_ARGS(start_lsn))));
887 : }
888 :
889 : /* We shouldn't go backward. */
890 : Assert(summary_start_lsn >= start_lsn);
891 : }
892 :
893 : /*
894 : * Main loop: read xlog records one by one.
895 : */
896 : while (1)
897 0 : {
898 : int block_id;
899 : char *errormsg;
900 : XLogRecord *record;
901 0 : bool stop_requested = false;
902 :
903 0 : HandleWalSummarizerInterrupts();
904 :
905 : /* We shouldn't go backward. */
906 : Assert(summary_start_lsn <= xlogreader->EndRecPtr);
907 :
908 : /* Now read the next record. */
909 0 : record = XLogReadRecord(xlogreader, &errormsg);
910 0 : if (record == NULL)
911 : {
912 0 : if (private_data->end_of_wal)
913 : {
914 : /*
915 : * This timeline must be historic and must end before we were
916 : * able to read a complete record.
917 : */
918 0 : ereport(DEBUG1,
919 : errmsg_internal("could not read WAL from timeline %u at %X/%X: end of WAL at %X/%X",
920 : tli,
921 : LSN_FORMAT_ARGS(xlogreader->EndRecPtr),
922 : LSN_FORMAT_ARGS(private_data->read_upto)));
923 : /* Summary ends at end of WAL. */
924 0 : summary_end_lsn = private_data->read_upto;
925 0 : break;
926 : }
927 0 : if (errormsg)
928 0 : ereport(ERROR,
929 : (errcode_for_file_access(),
930 : errmsg("could not read WAL from timeline %u at %X/%X: %s",
931 : tli, LSN_FORMAT_ARGS(xlogreader->EndRecPtr),
932 : errormsg)));
933 : else
934 0 : ereport(ERROR,
935 : (errcode_for_file_access(),
936 : errmsg("could not read WAL from timeline %u at %X/%X",
937 : tli, LSN_FORMAT_ARGS(xlogreader->EndRecPtr))));
938 : }
939 :
940 : /* We shouldn't go backward. */
941 : Assert(summary_start_lsn <= xlogreader->EndRecPtr);
942 :
943 0 : if (!XLogRecPtrIsInvalid(switch_lsn) &&
944 0 : xlogreader->ReadRecPtr >= switch_lsn)
945 : {
946 : /*
947 : * Whoops! We've read a record that *starts* after the switch LSN,
948 : * contrary to our goal of reading only until we hit the first
949 : * record that ends at or after the switch LSN. Pretend we didn't
950 : * read it after all by bailing out of this loop right here,
951 : * before we do anything with this record.
952 : *
953 : * This can happen because the last record before the switch LSN
954 : * might be continued across multiple pages, and then we might
955 : * come to a page with XLP_FIRST_IS_OVERWRITE_CONTRECORD set. In
956 : * that case, the record that was continued across multiple pages
957 : * is incomplete and will be disregarded, and the read will
958 : * restart from the beginning of the page that is flagged
959 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD.
960 : *
961 : * If this case occurs, we can fairly say that the current summary
962 : * file ends at the switch LSN exactly. The first record on the
963 : * page marked XLP_FIRST_IS_OVERWRITE_CONTRECORD will be
964 : * discovered when generating the next summary file.
965 : */
966 0 : summary_end_lsn = switch_lsn;
967 0 : break;
968 : }
969 :
970 : /* Special handling for particular types of WAL records. */
971 0 : switch (XLogRecGetRmid(xlogreader))
972 : {
973 0 : case RM_DBASE_ID:
974 0 : SummarizeDbaseRecord(xlogreader, brtab);
975 0 : break;
976 0 : case RM_SMGR_ID:
977 0 : SummarizeSmgrRecord(xlogreader, brtab);
978 0 : break;
979 0 : case RM_XACT_ID:
980 0 : SummarizeXactRecord(xlogreader, brtab);
981 0 : break;
982 0 : case RM_XLOG_ID:
983 0 : stop_requested = SummarizeXlogRecord(xlogreader);
984 0 : break;
985 0 : default:
986 0 : break;
987 : }
988 :
989 : /*
990 : * If we've been told that it's time to end this WAL summary file, do
991 : * so. As an exception, if there's nothing included in this WAL
992 : * summary file yet, then stopping doesn't make any sense, and we
993 : * should wait until the next stop point instead.
994 : */
995 0 : if (stop_requested && xlogreader->ReadRecPtr > summary_start_lsn)
996 : {
997 0 : summary_end_lsn = xlogreader->ReadRecPtr;
998 0 : break;
999 : }
1000 :
1001 : /* Feed block references from xlog record to block reference table. */
1002 0 : for (block_id = 0; block_id <= XLogRecMaxBlockId(xlogreader);
1003 0 : block_id++)
1004 : {
1005 : RelFileLocator rlocator;
1006 : ForkNumber forknum;
1007 : BlockNumber blocknum;
1008 :
1009 0 : if (!XLogRecGetBlockTagExtended(xlogreader, block_id, &rlocator,
1010 : &forknum, &blocknum, NULL))
1011 0 : continue;
1012 :
1013 : /*
1014 : * As we do elsewhere, ignore the FSM fork, because it's not fully
1015 : * WAL-logged.
1016 : */
1017 0 : if (forknum != FSM_FORKNUM)
1018 0 : BlockRefTableMarkBlockModified(brtab, &rlocator, forknum,
1019 : blocknum);
1020 : }
1021 :
1022 : /* Update our notion of where this summary file ends. */
1023 0 : summary_end_lsn = xlogreader->EndRecPtr;
1024 :
1025 : /* Also update shared memory. */
1026 0 : LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
1027 : Assert(summary_end_lsn >= WalSummarizerCtl->pending_lsn);
1028 : Assert(summary_end_lsn >= WalSummarizerCtl->summarized_lsn);
1029 0 : WalSummarizerCtl->pending_lsn = summary_end_lsn;
1030 0 : LWLockRelease(WALSummarizerLock);
1031 :
1032 : /*
1033 : * If we have a switch LSN and have reached it, stop before reading
1034 : * the next record.
1035 : */
1036 0 : if (!XLogRecPtrIsInvalid(switch_lsn) &&
1037 0 : xlogreader->EndRecPtr >= switch_lsn)
1038 0 : break;
1039 : }
1040 :
1041 : /* Destroy xlogreader. */
1042 0 : pfree(xlogreader->private_data);
1043 0 : XLogReaderFree(xlogreader);
1044 :
1045 : /*
1046 : * If a timeline switch occurs, we may fail to make any progress at all
1047 : * before exiting the loop above. If that happens, we don't write a WAL
1048 : * summary file at all.
1049 : */
1050 0 : if (summary_end_lsn > summary_start_lsn)
1051 : {
1052 : /* Generate temporary and final path name. */
1053 0 : snprintf(temp_path, MAXPGPATH,
1054 : XLOGDIR "/summaries/temp.summary");
1055 0 : snprintf(final_path, MAXPGPATH,
1056 : XLOGDIR "/summaries/%08X%08X%08X%08X%08X.summary",
1057 : tli,
1058 0 : LSN_FORMAT_ARGS(summary_start_lsn),
1059 0 : LSN_FORMAT_ARGS(summary_end_lsn));
1060 :
1061 : /* Open the temporary file for writing. */
1062 0 : io.filepos = 0;
1063 0 : io.file = PathNameOpenFile(temp_path, O_WRONLY | O_CREAT | O_TRUNC);
1064 0 : if (io.file < 0)
1065 0 : ereport(ERROR,
1066 : (errcode_for_file_access(),
1067 : errmsg("could not create file \"%s\": %m", temp_path)));
1068 :
1069 : /* Write the data. */
1070 0 : WriteBlockRefTable(brtab, WriteWalSummary, &io);
1071 :
1072 : /* Close temporary file and shut down xlogreader. */
1073 0 : FileClose(io.file);
1074 :
1075 : /* Tell the user what we did. */
1076 0 : ereport(DEBUG1,
1077 : errmsg("summarized WAL on TLI %u from %X/%X to %X/%X",
1078 : tli,
1079 : LSN_FORMAT_ARGS(summary_start_lsn),
1080 : LSN_FORMAT_ARGS(summary_end_lsn)));
1081 :
1082 : /* Durably rename the new summary into place. */
1083 0 : durable_rename(temp_path, final_path, ERROR);
1084 : }
1085 :
1086 0 : return summary_end_lsn;
1087 : }
1088 :
1089 : /*
1090 : * Special handling for WAL records with RM_DBASE_ID.
1091 : */
1092 : static void
1093 0 : SummarizeDbaseRecord(XLogReaderState *xlogreader, BlockRefTable *brtab)
1094 : {
1095 0 : uint8 info = XLogRecGetInfo(xlogreader) & ~XLR_INFO_MASK;
1096 :
1097 : /*
1098 : * We use relfilenode zero for a given database OID and tablespace OID to
1099 : * indicate that all relations with that pair of IDs have been recreated
1100 : * if they exist at all. Effectively, we're setting a limit block of 0 for
1101 : * all such relfilenodes.
1102 : *
1103 : * Technically, this special handling is only needed in the case of
1104 : * XLOG_DBASE_CREATE_FILE_COPY, because that can create a whole bunch of
1105 : * relation files in a directory without logging anything specific to each
1106 : * one. If we didn't mark the whole DB OID/TS OID combination in some way,
1107 : * then a tablespace that was dropped after the reference backup and
1108 : * recreated using the FILE_COPY method prior to the incremental backup
1109 : * would look just like one that was never touched at all, which would be
1110 : * catastrophic.
1111 : *
1112 : * But it seems best to adopt this treatment for all records that drop or
1113 : * create a DB OID/TS OID combination. That's similar to how we treat the
1114 : * limit block for individual relations, and it's an extra layer of safety
1115 : * here. We can never lose data by marking more stuff as needing to be
1116 : * backed up in full.
1117 : */
1118 0 : if (info == XLOG_DBASE_CREATE_FILE_COPY)
1119 : {
1120 : xl_dbase_create_file_copy_rec *xlrec;
1121 : RelFileLocator rlocator;
1122 :
1123 0 : xlrec =
1124 0 : (xl_dbase_create_file_copy_rec *) XLogRecGetData(xlogreader);
1125 0 : rlocator.spcOid = xlrec->tablespace_id;
1126 0 : rlocator.dbOid = xlrec->db_id;
1127 0 : rlocator.relNumber = 0;
1128 0 : BlockRefTableSetLimitBlock(brtab, &rlocator, MAIN_FORKNUM, 0);
1129 : }
1130 0 : else if (info == XLOG_DBASE_CREATE_WAL_LOG)
1131 : {
1132 : xl_dbase_create_wal_log_rec *xlrec;
1133 : RelFileLocator rlocator;
1134 :
1135 0 : xlrec = (xl_dbase_create_wal_log_rec *) XLogRecGetData(xlogreader);
1136 0 : rlocator.spcOid = xlrec->tablespace_id;
1137 0 : rlocator.dbOid = xlrec->db_id;
1138 0 : rlocator.relNumber = 0;
1139 0 : BlockRefTableSetLimitBlock(brtab, &rlocator, MAIN_FORKNUM, 0);
1140 : }
1141 0 : else if (info == XLOG_DBASE_DROP)
1142 : {
1143 : xl_dbase_drop_rec *xlrec;
1144 : RelFileLocator rlocator;
1145 : int i;
1146 :
1147 0 : xlrec = (xl_dbase_drop_rec *) XLogRecGetData(xlogreader);
1148 0 : rlocator.dbOid = xlrec->db_id;
1149 0 : rlocator.relNumber = 0;
1150 0 : for (i = 0; i < xlrec->ntablespaces; ++i)
1151 : {
1152 0 : rlocator.spcOid = xlrec->tablespace_ids[i];
1153 0 : BlockRefTableSetLimitBlock(brtab, &rlocator, MAIN_FORKNUM, 0);
1154 : }
1155 : }
1156 0 : }
1157 :
1158 : /*
1159 : * Special handling for WAL records with RM_SMGR_ID.
1160 : */
1161 : static void
1162 0 : SummarizeSmgrRecord(XLogReaderState *xlogreader, BlockRefTable *brtab)
1163 : {
1164 0 : uint8 info = XLogRecGetInfo(xlogreader) & ~XLR_INFO_MASK;
1165 :
1166 0 : if (info == XLOG_SMGR_CREATE)
1167 : {
1168 : xl_smgr_create *xlrec;
1169 :
1170 : /*
1171 : * If a new relation fork is created on disk, there is no point
1172 : * tracking anything about which blocks have been modified, because
1173 : * the whole thing will be new. Hence, set the limit block for this
1174 : * fork to 0.
1175 : *
1176 : * Ignore the FSM fork, which is not fully WAL-logged.
1177 : */
1178 0 : xlrec = (xl_smgr_create *) XLogRecGetData(xlogreader);
1179 :
1180 0 : if (xlrec->forkNum != FSM_FORKNUM)
1181 0 : BlockRefTableSetLimitBlock(brtab, &xlrec->rlocator,
1182 : xlrec->forkNum, 0);
1183 : }
1184 0 : else if (info == XLOG_SMGR_TRUNCATE)
1185 : {
1186 : xl_smgr_truncate *xlrec;
1187 :
1188 0 : xlrec = (xl_smgr_truncate *) XLogRecGetData(xlogreader);
1189 :
1190 : /*
1191 : * If a relation fork is truncated on disk, there is no point in
1192 : * tracking anything about block modifications beyond the truncation
1193 : * point.
1194 : *
1195 : * We ignore SMGR_TRUNCATE_FSM here because the FSM isn't fully
1196 : * WAL-logged and thus we can't track modified blocks for it anyway.
1197 : */
1198 0 : if ((xlrec->flags & SMGR_TRUNCATE_HEAP) != 0)
1199 0 : BlockRefTableSetLimitBlock(brtab, &xlrec->rlocator,
1200 : MAIN_FORKNUM, xlrec->blkno);
1201 0 : if ((xlrec->flags & SMGR_TRUNCATE_VM) != 0)
1202 0 : BlockRefTableSetLimitBlock(brtab, &xlrec->rlocator,
1203 : VISIBILITYMAP_FORKNUM, xlrec->blkno);
1204 : }
1205 0 : }
1206 :
1207 : /*
1208 : * Special handling for WAL records with RM_XACT_ID.
1209 : */
1210 : static void
1211 0 : SummarizeXactRecord(XLogReaderState *xlogreader, BlockRefTable *brtab)
1212 : {
1213 0 : uint8 info = XLogRecGetInfo(xlogreader) & ~XLR_INFO_MASK;
1214 0 : uint8 xact_info = info & XLOG_XACT_OPMASK;
1215 :
1216 0 : if (xact_info == XLOG_XACT_COMMIT ||
1217 : xact_info == XLOG_XACT_COMMIT_PREPARED)
1218 0 : {
1219 0 : xl_xact_commit *xlrec = (xl_xact_commit *) XLogRecGetData(xlogreader);
1220 : xl_xact_parsed_commit parsed;
1221 : int i;
1222 :
1223 : /*
1224 : * Don't track modified blocks for any relations that were removed on
1225 : * commit.
1226 : */
1227 0 : ParseCommitRecord(XLogRecGetInfo(xlogreader), xlrec, &parsed);
1228 0 : for (i = 0; i < parsed.nrels; ++i)
1229 : {
1230 : ForkNumber forknum;
1231 :
1232 0 : for (forknum = 0; forknum <= MAX_FORKNUM; ++forknum)
1233 0 : if (forknum != FSM_FORKNUM)
1234 0 : BlockRefTableSetLimitBlock(brtab, &parsed.xlocators[i],
1235 : forknum, 0);
1236 : }
1237 : }
1238 0 : else if (xact_info == XLOG_XACT_ABORT ||
1239 : xact_info == XLOG_XACT_ABORT_PREPARED)
1240 : {
1241 0 : xl_xact_abort *xlrec = (xl_xact_abort *) XLogRecGetData(xlogreader);
1242 : xl_xact_parsed_abort parsed;
1243 : int i;
1244 :
1245 : /*
1246 : * Don't track modified blocks for any relations that were removed on
1247 : * abort.
1248 : */
1249 0 : ParseAbortRecord(XLogRecGetInfo(xlogreader), xlrec, &parsed);
1250 0 : for (i = 0; i < parsed.nrels; ++i)
1251 : {
1252 : ForkNumber forknum;
1253 :
1254 0 : for (forknum = 0; forknum <= MAX_FORKNUM; ++forknum)
1255 0 : if (forknum != FSM_FORKNUM)
1256 0 : BlockRefTableSetLimitBlock(brtab, &parsed.xlocators[i],
1257 : forknum, 0);
1258 : }
1259 : }
1260 0 : }
1261 :
1262 : /*
1263 : * Special handling for WAL records with RM_XLOG_ID.
1264 : */
1265 : static bool
1266 0 : SummarizeXlogRecord(XLogReaderState *xlogreader)
1267 : {
1268 0 : uint8 info = XLogRecGetInfo(xlogreader) & ~XLR_INFO_MASK;
1269 :
1270 0 : if (info == XLOG_CHECKPOINT_REDO || info == XLOG_CHECKPOINT_SHUTDOWN)
1271 : {
1272 : /*
1273 : * This is an LSN at which redo might begin, so we'd like
1274 : * summarization to stop just before this WAL record.
1275 : */
1276 0 : return true;
1277 : }
1278 :
1279 0 : return false;
1280 : }
1281 :
1282 : /*
1283 : * Similar to read_local_xlog_page, but limited to read from one particular
1284 : * timeline. If the end of WAL is reached, it will wait for more if reading
1285 : * from the current timeline, or give up if reading from a historic timeline.
1286 : * In the latter case, it will also set private_data->end_of_wal = true.
1287 : *
1288 : * Caller must set private_data->tli to the TLI of interest,
1289 : * private_data->read_upto to the lowest LSN that is not known to be safe
1290 : * to read on that timeline, and private_data->historic to true if and only
1291 : * if the timeline is not the current timeline. This function will update
1292 : * private_data->read_upto and private_data->historic if more WAL appears
1293 : * on the current timeline or if the current timeline becomes historic.
1294 : */
1295 : static int
1296 0 : summarizer_read_local_xlog_page(XLogReaderState *state,
1297 : XLogRecPtr targetPagePtr, int reqLen,
1298 : XLogRecPtr targetRecPtr, char *cur_page)
1299 : {
1300 : int count;
1301 : WALReadError errinfo;
1302 : SummarizerReadLocalXLogPrivate *private_data;
1303 :
1304 0 : HandleWalSummarizerInterrupts();
1305 :
1306 0 : private_data = (SummarizerReadLocalXLogPrivate *)
1307 : state->private_data;
1308 :
1309 : while (1)
1310 : {
1311 0 : if (targetPagePtr + XLOG_BLCKSZ <= private_data->read_upto)
1312 : {
1313 : /*
1314 : * more than one block available; read only that block, have
1315 : * caller come back if they need more.
1316 : */
1317 0 : count = XLOG_BLCKSZ;
1318 0 : break;
1319 : }
1320 0 : else if (targetPagePtr + reqLen > private_data->read_upto)
1321 : {
1322 : /* We don't seem to have enough data. */
1323 0 : if (private_data->historic)
1324 : {
1325 : /*
1326 : * This is a historic timeline, so there will never be any
1327 : * more data than we have currently.
1328 : */
1329 0 : private_data->end_of_wal = true;
1330 0 : return -1;
1331 : }
1332 : else
1333 : {
1334 : XLogRecPtr latest_lsn;
1335 : TimeLineID latest_tli;
1336 :
1337 : /*
1338 : * This is - or at least was up until very recently - the
1339 : * current timeline, so more data might show up. Delay here
1340 : * so we don't tight-loop.
1341 : */
1342 0 : HandleWalSummarizerInterrupts();
1343 0 : summarizer_wait_for_wal();
1344 :
1345 : /* Recheck end-of-WAL. */
1346 0 : latest_lsn = GetLatestLSN(&latest_tli);
1347 0 : if (private_data->tli == latest_tli)
1348 : {
1349 : /* Still the current timeline, update max LSN. */
1350 : Assert(latest_lsn >= private_data->read_upto);
1351 0 : private_data->read_upto = latest_lsn;
1352 : }
1353 : else
1354 : {
1355 0 : List *tles = readTimeLineHistory(latest_tli);
1356 : XLogRecPtr switchpoint;
1357 :
1358 : /*
1359 : * The timeline we're scanning is no longer the latest
1360 : * one. Figure out when it ended.
1361 : */
1362 0 : private_data->historic = true;
1363 0 : switchpoint = tliSwitchPoint(private_data->tli, tles,
1364 : NULL);
1365 :
1366 : /*
1367 : * Allow reads up to exactly the switch point.
1368 : *
1369 : * It's possible that this will cause read_upto to move
1370 : * backwards, because walreceiver might have read a
1371 : * partial record and flushed it to disk, and we'd view
1372 : * that data as safe to read. However, the
1373 : * XLOG_END_OF_RECOVERY record will be written at the end
1374 : * of the last complete WAL record, not at the end of the
1375 : * WAL that we've flushed to disk.
1376 : *
1377 : * So switchpoint < private->read_upto is possible here,
1378 : * but switchpoint < state->EndRecPtr should not be.
1379 : */
1380 : Assert(switchpoint >= state->EndRecPtr);
1381 0 : private_data->read_upto = switchpoint;
1382 :
1383 : /* Debugging output. */
1384 0 : ereport(DEBUG1,
1385 : errmsg("timeline %u became historic, can read up to %X/%X",
1386 : private_data->tli, LSN_FORMAT_ARGS(private_data->read_upto)));
1387 : }
1388 :
1389 : /* Go around and try again. */
1390 : }
1391 : }
1392 : else
1393 : {
1394 : /* enough bytes available to satisfy the request */
1395 0 : count = private_data->read_upto - targetPagePtr;
1396 0 : break;
1397 : }
1398 : }
1399 :
1400 0 : if (!WALRead(state, cur_page, targetPagePtr, count,
1401 : private_data->tli, &errinfo))
1402 0 : WALReadRaiseError(&errinfo);
1403 :
1404 : /* Track that we read a page, for sleep time calculation. */
1405 0 : ++pages_read_since_last_sleep;
1406 :
1407 : /* number of valid bytes in the buffer */
1408 0 : return count;
1409 : }
1410 :
1411 : /*
1412 : * Sleep for long enough that we believe it's likely that more WAL will
1413 : * be available afterwards.
1414 : */
1415 : static void
1416 0 : summarizer_wait_for_wal(void)
1417 : {
1418 0 : if (pages_read_since_last_sleep == 0)
1419 : {
1420 : /*
1421 : * No pages were read since the last sleep, so double the sleep time,
1422 : * but not beyond the maximum allowable value.
1423 : */
1424 0 : sleep_quanta = Min(sleep_quanta * 2, MAX_SLEEP_QUANTA);
1425 : }
1426 0 : else if (pages_read_since_last_sleep > 1)
1427 : {
1428 : /*
1429 : * Multiple pages were read since the last sleep, so reduce the sleep
1430 : * time.
1431 : *
1432 : * A large burst of activity should be able to quickly reduce the
1433 : * sleep time to the minimum, but we don't want a handful of extra WAL
1434 : * records to provoke a strong reaction. We choose to reduce the sleep
1435 : * time by 1 quantum for each page read beyond the first, which is a
1436 : * fairly arbitrary way of trying to be reactive without overreacting.
1437 : */
1438 0 : if (pages_read_since_last_sleep > sleep_quanta - 1)
1439 0 : sleep_quanta = 1;
1440 : else
1441 0 : sleep_quanta -= pages_read_since_last_sleep;
1442 : }
1443 :
1444 : /* OK, now sleep. */
1445 0 : (void) WaitLatch(MyLatch,
1446 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
1447 : sleep_quanta * MS_PER_SLEEP_QUANTUM,
1448 : WAIT_EVENT_WAL_SUMMARIZER_WAL);
1449 0 : ResetLatch(MyLatch);
1450 :
1451 : /* Reset count of pages read. */
1452 0 : pages_read_since_last_sleep = 0;
1453 0 : }
1454 :
1455 : /*
1456 : * Remove WAL summaries whose mtimes are older than wal_summary_keep_time.
1457 : */
1458 : static void
1459 0 : MaybeRemoveOldWalSummaries(void)
1460 : {
1461 0 : XLogRecPtr redo_pointer = GetRedoRecPtr();
1462 : List *wslist;
1463 : time_t cutoff_time;
1464 :
1465 : /* If WAL summary removal is disabled, don't do anything. */
1466 0 : if (wal_summary_keep_time == 0)
1467 0 : return;
1468 :
1469 : /*
1470 : * If the redo pointer has not advanced, don't do anything.
1471 : *
1472 : * This has the effect that we only try to remove old WAL summary files
1473 : * once per checkpoint cycle.
1474 : */
1475 0 : if (redo_pointer == redo_pointer_at_last_summary_removal)
1476 0 : return;
1477 0 : redo_pointer_at_last_summary_removal = redo_pointer;
1478 :
1479 : /*
1480 : * Files should only be removed if the last modification time precedes the
1481 : * cutoff time we compute here.
1482 : */
1483 0 : cutoff_time = time(NULL) - wal_summary_keep_time * SECS_PER_MINUTE;
1484 :
1485 : /* Get all the summaries that currently exist. */
1486 0 : wslist = GetWalSummaries(0, InvalidXLogRecPtr, InvalidXLogRecPtr);
1487 :
1488 : /* Loop until all summaries have been considered for removal. */
1489 0 : while (wslist != NIL)
1490 : {
1491 : ListCell *lc;
1492 : XLogSegNo oldest_segno;
1493 0 : XLogRecPtr oldest_lsn = InvalidXLogRecPtr;
1494 : TimeLineID selected_tli;
1495 :
1496 0 : HandleWalSummarizerInterrupts();
1497 :
1498 : /*
1499 : * Pick a timeline for which some summary files still exist on disk,
1500 : * and find the oldest LSN that still exists on disk for that
1501 : * timeline.
1502 : */
1503 0 : selected_tli = ((WalSummaryFile *) linitial(wslist))->tli;
1504 0 : oldest_segno = XLogGetOldestSegno(selected_tli);
1505 0 : if (oldest_segno != 0)
1506 0 : XLogSegNoOffsetToRecPtr(oldest_segno, 0, wal_segment_size,
1507 : oldest_lsn);
1508 :
1509 :
1510 : /* Consider each WAL file on the selected timeline in turn. */
1511 0 : foreach(lc, wslist)
1512 : {
1513 0 : WalSummaryFile *ws = lfirst(lc);
1514 :
1515 0 : HandleWalSummarizerInterrupts();
1516 :
1517 : /* If it's not on this timeline, it's not time to consider it. */
1518 0 : if (selected_tli != ws->tli)
1519 0 : continue;
1520 :
1521 : /*
1522 : * If the WAL doesn't exist any more, we can remove it if the file
1523 : * modification time is old enough.
1524 : */
1525 0 : if (XLogRecPtrIsInvalid(oldest_lsn) || ws->end_lsn <= oldest_lsn)
1526 0 : RemoveWalSummaryIfOlderThan(ws, cutoff_time);
1527 :
1528 : /*
1529 : * Whether we removed the file or not, we need not consider it
1530 : * again.
1531 : */
1532 0 : wslist = foreach_delete_current(wslist, lc);
1533 0 : pfree(ws);
1534 : }
1535 : }
1536 : }
|