Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * xlog.c
4 : * PostgreSQL write-ahead log manager
5 : *
6 : * The Write-Ahead Log (WAL) functionality is split into several source
7 : * files, in addition to this one:
8 : *
9 : * xloginsert.c - Functions for constructing WAL records
10 : * xlogrecovery.c - WAL recovery and standby code
11 : * xlogreader.c - Facility for reading WAL files and parsing WAL records
12 : * xlogutils.c - Helper functions for WAL redo routines
13 : *
14 : * This file contains functions for coordinating database startup and
15 : * checkpointing, and managing the write-ahead log buffers when the
16 : * system is running.
17 : *
18 : * StartupXLOG() is the main entry point of the startup process. It
19 : * coordinates database startup, performing WAL recovery, and the
20 : * transition from WAL recovery into normal operations.
21 : *
22 : * XLogInsertRecord() inserts a WAL record into the WAL buffers. Most
23 : * callers should not call this directly, but use the functions in
24 : * xloginsert.c to construct the WAL record. XLogFlush() can be used
25 : * to force the WAL to disk.
26 : *
27 : * In addition to those, there are many other functions for interrogating
28 : * the current system state, and for starting/stopping backups.
29 : *
30 : *
31 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
32 : * Portions Copyright (c) 1994, Regents of the University of California
33 : *
34 : * src/backend/access/transam/xlog.c
35 : *
36 : *-------------------------------------------------------------------------
37 : */
38 :
39 : #include "postgres.h"
40 :
41 : #include <ctype.h>
42 : #include <math.h>
43 : #include <time.h>
44 : #include <fcntl.h>
45 : #include <sys/stat.h>
46 : #include <sys/time.h>
47 : #include <unistd.h>
48 :
49 : #include "access/clog.h"
50 : #include "access/commit_ts.h"
51 : #include "access/heaptoast.h"
52 : #include "access/multixact.h"
53 : #include "access/rewriteheap.h"
54 : #include "access/subtrans.h"
55 : #include "access/timeline.h"
56 : #include "access/transam.h"
57 : #include "access/twophase.h"
58 : #include "access/xact.h"
59 : #include "access/xlog_internal.h"
60 : #include "access/xlogarchive.h"
61 : #include "access/xloginsert.h"
62 : #include "access/xlogreader.h"
63 : #include "access/xlogrecovery.h"
64 : #include "access/xlogutils.h"
65 : #include "access/xlogwait.h"
66 : #include "backup/basebackup.h"
67 : #include "catalog/catversion.h"
68 : #include "catalog/pg_control.h"
69 : #include "catalog/pg_database.h"
70 : #include "common/controldata_utils.h"
71 : #include "common/file_utils.h"
72 : #include "executor/instrument.h"
73 : #include "miscadmin.h"
74 : #include "pg_trace.h"
75 : #include "pgstat.h"
76 : #include "port/atomics.h"
77 : #include "postmaster/bgwriter.h"
78 : #include "postmaster/startup.h"
79 : #include "postmaster/walsummarizer.h"
80 : #include "postmaster/walwriter.h"
81 : #include "replication/origin.h"
82 : #include "replication/slot.h"
83 : #include "replication/slotsync.h"
84 : #include "replication/snapbuild.h"
85 : #include "replication/walreceiver.h"
86 : #include "replication/walsender.h"
87 : #include "storage/bufmgr.h"
88 : #include "storage/fd.h"
89 : #include "storage/ipc.h"
90 : #include "storage/large_object.h"
91 : #include "storage/latch.h"
92 : #include "storage/predicate.h"
93 : #include "storage/proc.h"
94 : #include "storage/procarray.h"
95 : #include "storage/reinit.h"
96 : #include "storage/spin.h"
97 : #include "storage/sync.h"
98 : #include "utils/guc_hooks.h"
99 : #include "utils/guc_tables.h"
100 : #include "utils/injection_point.h"
101 : #include "utils/pgstat_internal.h"
102 : #include "utils/ps_status.h"
103 : #include "utils/relmapper.h"
104 : #include "utils/snapmgr.h"
105 : #include "utils/timeout.h"
106 : #include "utils/timestamp.h"
107 : #include "utils/varlena.h"
108 : #include "utils/wait_event.h"
109 :
110 : #ifdef WAL_DEBUG
111 : #include "utils/memutils.h"
112 : #endif
113 :
114 : /* timeline ID to be used when bootstrapping */
115 : #define BootstrapTimeLineID 1
116 :
117 : /* User-settable parameters */
118 : int max_wal_size_mb = 1024; /* 1 GB */
119 : int min_wal_size_mb = 80; /* 80 MB */
120 : int wal_keep_size_mb = 0;
121 : int XLOGbuffers = -1;
122 : int XLogArchiveTimeout = 0;
123 : int XLogArchiveMode = ARCHIVE_MODE_OFF;
124 : char *XLogArchiveCommand = NULL;
125 : bool EnableHotStandby = false;
126 : bool fullPageWrites = true;
127 : bool wal_log_hints = false;
128 : int wal_compression = WAL_COMPRESSION_NONE;
129 : char *wal_consistency_checking_string = NULL;
130 : bool *wal_consistency_checking = NULL;
131 : bool wal_init_zero = true;
132 : bool wal_recycle = true;
133 : bool log_checkpoints = true;
134 : int wal_sync_method = DEFAULT_WAL_SYNC_METHOD;
135 : int wal_level = WAL_LEVEL_REPLICA;
136 : int CommitDelay = 0; /* precommit delay in microseconds */
137 : int CommitSiblings = 5; /* # concurrent xacts needed to sleep */
138 : int wal_retrieve_retry_interval = 5000;
139 : int max_slot_wal_keep_size_mb = -1;
140 : int wal_decode_buffer_size = 512 * 1024;
141 : bool track_wal_io_timing = false;
142 :
143 : #ifdef WAL_DEBUG
144 : bool XLOG_DEBUG = false;
145 : #endif
146 :
147 : int wal_segment_size = DEFAULT_XLOG_SEG_SIZE;
148 :
149 : /*
150 : * Number of WAL insertion locks to use. A higher value allows more insertions
151 : * to happen concurrently, but adds some CPU overhead to flushing the WAL,
152 : * which needs to iterate all the locks.
153 : */
154 : #define NUM_XLOGINSERT_LOCKS 8
155 :
156 : /*
157 : * Max distance from last checkpoint, before triggering a new xlog-based
158 : * checkpoint.
159 : */
160 : int CheckPointSegments;
161 :
162 : /* Estimated distance between checkpoints, in bytes */
163 : static double CheckPointDistanceEstimate = 0;
164 : static double PrevCheckPointDistance = 0;
165 :
166 : /*
167 : * Track whether there were any deferred checks for custom resource managers
168 : * specified in wal_consistency_checking.
169 : */
170 : static bool check_wal_consistency_checking_deferred = false;
171 :
172 : /*
173 : * GUC support
174 : */
175 : const struct config_enum_entry wal_sync_method_options[] = {
176 : {"fsync", WAL_SYNC_METHOD_FSYNC, false},
177 : #ifdef HAVE_FSYNC_WRITETHROUGH
178 : {"fsync_writethrough", WAL_SYNC_METHOD_FSYNC_WRITETHROUGH, false},
179 : #endif
180 : {"fdatasync", WAL_SYNC_METHOD_FDATASYNC, false},
181 : #ifdef O_SYNC
182 : {"open_sync", WAL_SYNC_METHOD_OPEN, false},
183 : #endif
184 : #ifdef O_DSYNC
185 : {"open_datasync", WAL_SYNC_METHOD_OPEN_DSYNC, false},
186 : #endif
187 : {NULL, 0, false}
188 : };
189 :
190 :
191 : /*
192 : * Although only "on", "off", and "always" are documented,
193 : * we accept all the likely variants of "on" and "off".
194 : */
195 : const struct config_enum_entry archive_mode_options[] = {
196 : {"always", ARCHIVE_MODE_ALWAYS, false},
197 : {"on", ARCHIVE_MODE_ON, false},
198 : {"off", ARCHIVE_MODE_OFF, false},
199 : {"true", ARCHIVE_MODE_ON, true},
200 : {"false", ARCHIVE_MODE_OFF, true},
201 : {"yes", ARCHIVE_MODE_ON, true},
202 : {"no", ARCHIVE_MODE_OFF, true},
203 : {"1", ARCHIVE_MODE_ON, true},
204 : {"0", ARCHIVE_MODE_OFF, true},
205 : {NULL, 0, false}
206 : };
207 :
208 : /*
209 : * Statistics for current checkpoint are collected in this global struct.
210 : * Because only the checkpointer or a stand-alone backend can perform
211 : * checkpoints, this will be unused in normal backends.
212 : */
213 : CheckpointStatsData CheckpointStats;
214 :
215 : /*
216 : * During recovery, lastFullPageWrites keeps track of full_page_writes that
217 : * the replayed WAL records indicate. It's initialized with full_page_writes
218 : * that the recovery starting checkpoint record indicates, and then updated
219 : * each time XLOG_FPW_CHANGE record is replayed.
220 : */
221 : static bool lastFullPageWrites;
222 :
223 : /*
224 : * Local copy of the state tracked by SharedRecoveryState in shared memory,
225 : * It is false if SharedRecoveryState is RECOVERY_STATE_DONE. True actually
226 : * means "not known, need to check the shared state".
227 : */
228 : static bool LocalRecoveryInProgress = true;
229 :
230 : /*
231 : * Local state for XLogInsertAllowed():
232 : * 1: unconditionally allowed to insert XLOG
233 : * 0: unconditionally not allowed to insert XLOG
234 : * -1: must check RecoveryInProgress(); disallow until it is false
235 : * Most processes start with -1 and transition to 1 after seeing that recovery
236 : * is not in progress. But we can also force the value for special cases.
237 : * The coding in XLogInsertAllowed() depends on the first two of these states
238 : * being numerically the same as bool true and false.
239 : */
240 : static int LocalXLogInsertAllowed = -1;
241 :
242 : /*
243 : * ProcLastRecPtr points to the start of the last XLOG record inserted by the
244 : * current backend. It is updated for all inserts. XactLastRecEnd points to
245 : * end+1 of the last record, and is reset when we end a top-level transaction,
246 : * or start a new one; so it can be used to tell if the current transaction has
247 : * created any XLOG records.
248 : *
249 : * While in parallel mode, this may not be fully up to date. When committing,
250 : * a transaction can assume this covers all xlog records written either by the
251 : * user backend or by any parallel worker which was present at any point during
252 : * the transaction. But when aborting, or when still in parallel mode, other
253 : * parallel backends may have written WAL records at later LSNs than the value
254 : * stored here. The parallel leader advances its own copy, when necessary,
255 : * in WaitForParallelWorkersToFinish.
256 : */
257 : XLogRecPtr ProcLastRecPtr = InvalidXLogRecPtr;
258 : XLogRecPtr XactLastRecEnd = InvalidXLogRecPtr;
259 : XLogRecPtr XactLastCommitEnd = InvalidXLogRecPtr;
260 :
261 : /*
262 : * RedoRecPtr is this backend's local copy of the REDO record pointer
263 : * (which is almost but not quite the same as a pointer to the most recent
264 : * CHECKPOINT record). We update this from the shared-memory copy,
265 : * XLogCtl->Insert.RedoRecPtr, whenever we can safely do so (ie, when we
266 : * hold an insertion lock). See XLogInsertRecord for details. We are also
267 : * allowed to update from XLogCtl->RedoRecPtr if we hold the info_lck;
268 : * see GetRedoRecPtr.
269 : *
270 : * NB: Code that uses this variable must be prepared not only for the
271 : * possibility that it may be arbitrarily out of date, but also for the
272 : * possibility that it might be set to InvalidXLogRecPtr. We used to
273 : * initialize it as a side effect of the first call to RecoveryInProgress(),
274 : * which meant that most code that might use it could assume that it had a
275 : * real if perhaps stale value. That's no longer the case.
276 : */
277 : static XLogRecPtr RedoRecPtr;
278 :
279 : /*
280 : * doPageWrites is this backend's local copy of (fullPageWrites ||
281 : * runningBackups > 0). It is used together with RedoRecPtr to decide whether
282 : * a full-page image of a page need to be taken.
283 : *
284 : * NB: Initially this is false, and there's no guarantee that it will be
285 : * initialized to any other value before it is first used. Any code that
286 : * makes use of it must recheck the value after obtaining a WALInsertLock,
287 : * and respond appropriately if it turns out that the previous value wasn't
288 : * accurate.
289 : */
290 : static bool doPageWrites;
291 :
292 : /*----------
293 : * Shared-memory data structures for XLOG control
294 : *
295 : * LogwrtRqst indicates a byte position that we need to write and/or fsync
296 : * the log up to (all records before that point must be written or fsynced).
297 : * The positions already written/fsynced are maintained in logWriteResult
298 : * and logFlushResult using atomic access.
299 : * In addition to the shared variable, each backend has a private copy of
300 : * both in LogwrtResult, which is updated when convenient.
301 : *
302 : * The request bookkeeping is simpler: there is a shared XLogCtl->LogwrtRqst
303 : * (protected by info_lck), but we don't need to cache any copies of it.
304 : *
305 : * info_lck is only held long enough to read/update the protected variables,
306 : * so it's a plain spinlock. The other locks are held longer (potentially
307 : * over I/O operations), so we use LWLocks for them. These locks are:
308 : *
309 : * WALBufMappingLock: must be held to replace a page in the WAL buffer cache.
310 : * It is only held while initializing and changing the mapping. If the
311 : * contents of the buffer being replaced haven't been written yet, the mapping
312 : * lock is released while the write is done, and reacquired afterwards.
313 : *
314 : * WALWriteLock: must be held to write WAL buffers to disk (XLogWrite or
315 : * XLogFlush).
316 : *
317 : * ControlFileLock: must be held to read/update control file or create
318 : * new log file.
319 : *
320 : *----------
321 : */
322 :
323 : typedef struct XLogwrtRqst
324 : {
325 : XLogRecPtr Write; /* last byte + 1 to write out */
326 : XLogRecPtr Flush; /* last byte + 1 to flush */
327 : } XLogwrtRqst;
328 :
329 : typedef struct XLogwrtResult
330 : {
331 : XLogRecPtr Write; /* last byte + 1 written out */
332 : XLogRecPtr Flush; /* last byte + 1 flushed */
333 : } XLogwrtResult;
334 :
335 : /*
336 : * Inserting to WAL is protected by a small fixed number of WAL insertion
337 : * locks. To insert to the WAL, you must hold one of the locks - it doesn't
338 : * matter which one. To lock out other concurrent insertions, you must hold
339 : * of them. Each WAL insertion lock consists of a lightweight lock, plus an
340 : * indicator of how far the insertion has progressed (insertingAt).
341 : *
342 : * The insertingAt values are read when a process wants to flush WAL from
343 : * the in-memory buffers to disk, to check that all the insertions to the
344 : * region the process is about to write out have finished. You could simply
345 : * wait for all currently in-progress insertions to finish, but the
346 : * insertingAt indicator allows you to ignore insertions to later in the WAL,
347 : * so that you only wait for the insertions that are modifying the buffers
348 : * you're about to write out.
349 : *
350 : * This isn't just an optimization. If all the WAL buffers are dirty, an
351 : * inserter that's holding a WAL insert lock might need to evict an old WAL
352 : * buffer, which requires flushing the WAL. If it's possible for an inserter
353 : * to block on another inserter unnecessarily, deadlock can arise when two
354 : * inserters holding a WAL insert lock wait for each other to finish their
355 : * insertion.
356 : *
357 : * Small WAL records that don't cross a page boundary never update the value,
358 : * the WAL record is just copied to the page and the lock is released. But
359 : * to avoid the deadlock-scenario explained above, the indicator is always
360 : * updated before sleeping while holding an insertion lock.
361 : *
362 : * lastImportantAt contains the LSN of the last important WAL record inserted
363 : * using a given lock. This value is used to detect if there has been
364 : * important WAL activity since the last time some action, like a checkpoint,
365 : * was performed - allowing to not repeat the action if not. The LSN is
366 : * updated for all insertions, unless the XLOG_MARK_UNIMPORTANT flag was
367 : * set. lastImportantAt is never cleared, only overwritten by the LSN of newer
368 : * records. Tracking the WAL activity directly in WALInsertLock has the
369 : * advantage of not needing any additional locks to update the value.
370 : */
371 : typedef struct
372 : {
373 : LWLock lock;
374 : pg_atomic_uint64 insertingAt;
375 : XLogRecPtr lastImportantAt;
376 : } WALInsertLock;
377 :
378 : /*
379 : * All the WAL insertion locks are allocated as an array in shared memory. We
380 : * force the array stride to be a power of 2, which saves a few cycles in
381 : * indexing, but more importantly also ensures that individual slots don't
382 : * cross cache line boundaries. (Of course, we have to also ensure that the
383 : * array start address is suitably aligned.)
384 : */
385 : typedef union WALInsertLockPadded
386 : {
387 : WALInsertLock l;
388 : char pad[PG_CACHE_LINE_SIZE];
389 : } WALInsertLockPadded;
390 :
391 : /*
392 : * Session status of running backup, used for sanity checks in SQL-callable
393 : * functions to start and stop backups.
394 : */
395 : static SessionBackupState sessionBackupState = SESSION_BACKUP_NONE;
396 :
397 : /*
398 : * Shared state data for WAL insertion.
399 : */
400 : typedef struct XLogCtlInsert
401 : {
402 : slock_t insertpos_lck; /* protects CurrBytePos and PrevBytePos */
403 :
404 : /*
405 : * CurrBytePos is the end of reserved WAL. The next record will be
406 : * inserted at that position. PrevBytePos is the start position of the
407 : * previously inserted (or rather, reserved) record - it is copied to the
408 : * prev-link of the next record. These are stored as "usable byte
409 : * positions" rather than XLogRecPtrs (see XLogBytePosToRecPtr()).
410 : */
411 : uint64 CurrBytePos;
412 : uint64 PrevBytePos;
413 :
414 : /*
415 : * Make sure the above heavily-contended spinlock and byte positions are
416 : * on their own cache line. In particular, the RedoRecPtr and full page
417 : * write variables below should be on a different cache line. They are
418 : * read on every WAL insertion, but updated rarely, and we don't want
419 : * those reads to steal the cache line containing Curr/PrevBytePos.
420 : */
421 : char pad[PG_CACHE_LINE_SIZE];
422 :
423 : /*
424 : * fullPageWrites is the authoritative value used by all backends to
425 : * determine whether to write full-page image to WAL. This shared value,
426 : * instead of the process-local fullPageWrites, is required because, when
427 : * full_page_writes is changed by SIGHUP, we must WAL-log it before it
428 : * actually affects WAL-logging by backends. Checkpointer sets at startup
429 : * or after SIGHUP.
430 : *
431 : * To read these fields, you must hold an insertion lock. To modify them,
432 : * you must hold ALL the locks.
433 : */
434 : XLogRecPtr RedoRecPtr; /* current redo point for insertions */
435 : bool fullPageWrites;
436 :
437 : /*
438 : * runningBackups is a counter indicating the number of backups currently
439 : * in progress. lastBackupStart is the latest checkpoint redo location
440 : * used as a starting point for an online backup.
441 : */
442 : int runningBackups;
443 : XLogRecPtr lastBackupStart;
444 :
445 : /*
446 : * WAL insertion locks.
447 : */
448 : WALInsertLockPadded *WALInsertLocks;
449 : } XLogCtlInsert;
450 :
451 : /*
452 : * Total shared-memory state for XLOG.
453 : */
454 : typedef struct XLogCtlData
455 : {
456 : XLogCtlInsert Insert;
457 :
458 : /* Protected by info_lck: */
459 : XLogwrtRqst LogwrtRqst;
460 : XLogRecPtr RedoRecPtr; /* a recent copy of Insert->RedoRecPtr */
461 : XLogRecPtr asyncXactLSN; /* LSN of newest async commit/abort */
462 : XLogRecPtr replicationSlotMinLSN; /* oldest LSN needed by any slot */
463 :
464 : XLogSegNo lastRemovedSegNo; /* latest removed/recycled XLOG segment */
465 :
466 : /* Fake LSN counter, for unlogged relations. */
467 : pg_atomic_uint64 unloggedLSN;
468 :
469 : /* Time and LSN of last xlog segment switch. Protected by WALWriteLock. */
470 : pg_time_t lastSegSwitchTime;
471 : XLogRecPtr lastSegSwitchLSN;
472 :
473 : /* These are accessed using atomics -- info_lck not needed */
474 : pg_atomic_uint64 logInsertResult; /* last byte + 1 inserted to buffers */
475 : pg_atomic_uint64 logWriteResult; /* last byte + 1 written out */
476 : pg_atomic_uint64 logFlushResult; /* last byte + 1 flushed */
477 :
478 : /*
479 : * Latest initialized page in the cache (last byte position + 1).
480 : *
481 : * To change the identity of a buffer (and InitializedUpTo), you need to
482 : * hold WALBufMappingLock. To change the identity of a buffer that's
483 : * still dirty, the old page needs to be written out first, and for that
484 : * you need WALWriteLock, and you need to ensure that there are no
485 : * in-progress insertions to the page by calling
486 : * WaitXLogInsertionsToFinish().
487 : */
488 : XLogRecPtr InitializedUpTo;
489 :
490 : /*
491 : * These values do not change after startup, although the pointed-to pages
492 : * and xlblocks values certainly do. xlblocks values are protected by
493 : * WALBufMappingLock.
494 : */
495 : char *pages; /* buffers for unwritten XLOG pages */
496 : pg_atomic_uint64 *xlblocks; /* 1st byte ptr-s + XLOG_BLCKSZ */
497 : int XLogCacheBlck; /* highest allocated xlog buffer index */
498 :
499 : /*
500 : * InsertTimeLineID is the timeline into which new WAL is being inserted
501 : * and flushed. It is zero during recovery, and does not change once set.
502 : *
503 : * If we create a new timeline when the system was started up,
504 : * PrevTimeLineID is the old timeline's ID that we forked off from.
505 : * Otherwise it's equal to InsertTimeLineID.
506 : *
507 : * We set these fields while holding info_lck. Most that reads these
508 : * values knows that recovery is no longer in progress and so can safely
509 : * read the value without a lock, but code that could be run either during
510 : * or after recovery can take info_lck while reading these values.
511 : */
512 : TimeLineID InsertTimeLineID;
513 : TimeLineID PrevTimeLineID;
514 :
515 : /*
516 : * SharedRecoveryState indicates if we're still in crash or archive
517 : * recovery. Protected by info_lck.
518 : */
519 : RecoveryState SharedRecoveryState;
520 :
521 : /*
522 : * InstallXLogFileSegmentActive indicates whether the checkpointer should
523 : * arrange for future segments by recycling and/or PreallocXlogFiles().
524 : * Protected by ControlFileLock. Only the startup process changes it. If
525 : * true, anyone can use InstallXLogFileSegment(). If false, the startup
526 : * process owns the exclusive right to install segments, by reading from
527 : * the archive and possibly replacing existing files.
528 : */
529 : bool InstallXLogFileSegmentActive;
530 :
531 : /*
532 : * WalWriterSleeping indicates whether the WAL writer is currently in
533 : * low-power mode (and hence should be nudged if an async commit occurs).
534 : * Protected by info_lck.
535 : */
536 : bool WalWriterSleeping;
537 :
538 : /*
539 : * During recovery, we keep a copy of the latest checkpoint record here.
540 : * lastCheckPointRecPtr points to start of checkpoint record and
541 : * lastCheckPointEndPtr points to end+1 of checkpoint record. Used by the
542 : * checkpointer when it wants to create a restartpoint.
543 : *
544 : * Protected by info_lck.
545 : */
546 : XLogRecPtr lastCheckPointRecPtr;
547 : XLogRecPtr lastCheckPointEndPtr;
548 : CheckPoint lastCheckPoint;
549 :
550 : /*
551 : * lastFpwDisableRecPtr points to the start of the last replayed
552 : * XLOG_FPW_CHANGE record that instructs full_page_writes is disabled.
553 : */
554 : XLogRecPtr lastFpwDisableRecPtr;
555 :
556 : slock_t info_lck; /* locks shared variables shown above */
557 : } XLogCtlData;
558 :
559 : /*
560 : * Classification of XLogInsertRecord operations.
561 : */
562 : typedef enum
563 : {
564 : WALINSERT_NORMAL,
565 : WALINSERT_SPECIAL_SWITCH,
566 : WALINSERT_SPECIAL_CHECKPOINT
567 : } WalInsertClass;
568 :
569 : static XLogCtlData *XLogCtl = NULL;
570 :
571 : /* a private copy of XLogCtl->Insert.WALInsertLocks, for convenience */
572 : static WALInsertLockPadded *WALInsertLocks = NULL;
573 :
574 : /*
575 : * We maintain an image of pg_control in shared memory.
576 : */
577 : static ControlFileData *ControlFile = NULL;
578 :
579 : /*
580 : * Calculate the amount of space left on the page after 'endptr'. Beware
581 : * multiple evaluation!
582 : */
583 : #define INSERT_FREESPACE(endptr) \
584 : (((endptr) % XLOG_BLCKSZ == 0) ? 0 : (XLOG_BLCKSZ - (endptr) % XLOG_BLCKSZ))
585 :
586 : /* Macro to advance to next buffer index. */
587 : #define NextBufIdx(idx) \
588 : (((idx) == XLogCtl->XLogCacheBlck) ? 0 : ((idx) + 1))
589 :
590 : /*
591 : * XLogRecPtrToBufIdx returns the index of the WAL buffer that holds, or
592 : * would hold if it was in cache, the page containing 'recptr'.
593 : */
594 : #define XLogRecPtrToBufIdx(recptr) \
595 : (((recptr) / XLOG_BLCKSZ) % (XLogCtl->XLogCacheBlck + 1))
596 :
597 : /*
598 : * These are the number of bytes in a WAL page usable for WAL data.
599 : */
600 : #define UsableBytesInPage (XLOG_BLCKSZ - SizeOfXLogShortPHD)
601 :
602 : /*
603 : * Convert values of GUCs measured in megabytes to equiv. segment count.
604 : * Rounds down.
605 : */
606 : #define ConvertToXSegs(x, segsize) XLogMBVarToSegs((x), (segsize))
607 :
608 : /* The number of bytes in a WAL segment usable for WAL data. */
609 : static int UsableBytesInSegment;
610 :
611 : /*
612 : * Private, possibly out-of-date copy of shared LogwrtResult.
613 : * See discussion above.
614 : */
615 : static XLogwrtResult LogwrtResult = {0, 0};
616 :
617 : /*
618 : * Update local copy of shared XLogCtl->log{Write,Flush}Result
619 : *
620 : * It's critical that Flush always trails Write, so the order of the reads is
621 : * important, as is the barrier. See also XLogWrite.
622 : */
623 : #define RefreshXLogWriteResult(_target) \
624 : do { \
625 : _target.Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult); \
626 : pg_read_barrier(); \
627 : _target.Write = pg_atomic_read_u64(&XLogCtl->logWriteResult); \
628 : } while (0)
629 :
630 : /*
631 : * openLogFile is -1 or a kernel FD for an open log file segment.
632 : * openLogSegNo identifies the segment, and openLogTLI the corresponding TLI.
633 : * These variables are only used to write the XLOG, and so will normally refer
634 : * to the active segment.
635 : *
636 : * Note: call Reserve/ReleaseExternalFD to track consumption of this FD.
637 : */
638 : static int openLogFile = -1;
639 : static XLogSegNo openLogSegNo = 0;
640 : static TimeLineID openLogTLI = 0;
641 :
642 : /*
643 : * Local copies of equivalent fields in the control file. When running
644 : * crash recovery, LocalMinRecoveryPoint is set to InvalidXLogRecPtr as we
645 : * expect to replay all the WAL available, and updateMinRecoveryPoint is
646 : * switched to false to prevent any updates while replaying records.
647 : * Those values are kept consistent as long as crash recovery runs.
648 : */
649 : static XLogRecPtr LocalMinRecoveryPoint;
650 : static TimeLineID LocalMinRecoveryPointTLI;
651 : static bool updateMinRecoveryPoint = true;
652 :
653 : /* For WALInsertLockAcquire/Release functions */
654 : static int MyLockNo = 0;
655 : static bool holdingAllLocks = false;
656 :
657 : #ifdef WAL_DEBUG
658 : static MemoryContext walDebugCxt = NULL;
659 : #endif
660 :
661 : static void CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI,
662 : XLogRecPtr EndOfLog,
663 : TimeLineID newTLI);
664 : static void CheckRequiredParameterValues(void);
665 : static void XLogReportParameters(void);
666 : static int LocalSetXLogInsertAllowed(void);
667 : static void CreateEndOfRecoveryRecord(void);
668 : static XLogRecPtr CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn,
669 : XLogRecPtr pagePtr,
670 : TimeLineID newTLI);
671 : static void CheckPointGuts(XLogRecPtr checkPointRedo, int flags);
672 : static void KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo);
673 :
674 : static void AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli,
675 : bool opportunistic);
676 : static void XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible);
677 : static bool InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
678 : bool find_free, XLogSegNo max_segno,
679 : TimeLineID tli);
680 : static void XLogFileClose(void);
681 : static void PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli);
682 : static void RemoveTempXlogFiles(void);
683 : static void RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr,
684 : XLogRecPtr endptr, TimeLineID insertTLI);
685 : static void RemoveXlogFile(const struct dirent *segment_de,
686 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
687 : TimeLineID insertTLI);
688 : static void UpdateLastRemovedPtr(char *filename);
689 : static void ValidateXLOGDirectoryStructure(void);
690 : static void CleanupBackupHistory(void);
691 : static void UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force);
692 : static bool PerformRecoveryXLogAction(void);
693 : static void InitControlFile(uint64 sysidentifier, uint32 data_checksum_version);
694 : static void WriteControlFile(void);
695 : static void ReadControlFile(void);
696 : static void UpdateControlFile(void);
697 : static char *str_time(pg_time_t tnow, char *buf, size_t bufsize);
698 :
699 : static int get_sync_bit(int method);
700 :
701 : static void CopyXLogRecordToWAL(int write_len, bool isLogSwitch,
702 : XLogRecData *rdata,
703 : XLogRecPtr StartPos, XLogRecPtr EndPos,
704 : TimeLineID tli);
705 : static void ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos,
706 : XLogRecPtr *EndPos, XLogRecPtr *PrevPtr);
707 : static bool ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos,
708 : XLogRecPtr *PrevPtr);
709 : static XLogRecPtr WaitXLogInsertionsToFinish(XLogRecPtr upto);
710 : static char *GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli);
711 : static XLogRecPtr XLogBytePosToRecPtr(uint64 bytepos);
712 : static XLogRecPtr XLogBytePosToEndRecPtr(uint64 bytepos);
713 : static uint64 XLogRecPtrToBytePos(XLogRecPtr ptr);
714 :
715 : static void WALInsertLockAcquire(void);
716 : static void WALInsertLockAcquireExclusive(void);
717 : static void WALInsertLockRelease(void);
718 : static void WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt);
719 :
720 : /*
721 : * Insert an XLOG record represented by an already-constructed chain of data
722 : * chunks. This is a low-level routine; to construct the WAL record header
723 : * and data, use the higher-level routines in xloginsert.c.
724 : *
725 : * If 'fpw_lsn' is valid, it is the oldest LSN among the pages that this
726 : * WAL record applies to, that were not included in the record as full page
727 : * images. If fpw_lsn <= RedoRecPtr, the function does not perform the
728 : * insertion and returns InvalidXLogRecPtr. The caller can then recalculate
729 : * which pages need a full-page image, and retry. If fpw_lsn is invalid, the
730 : * record is always inserted.
731 : *
732 : * 'flags' gives more in-depth control on the record being inserted. See
733 : * XLogSetRecordFlags() for details.
734 : *
735 : * 'topxid_included' tells whether the top-transaction id is logged along with
736 : * current subtransaction. See XLogRecordAssemble().
737 : *
738 : * The first XLogRecData in the chain must be for the record header, and its
739 : * data must be MAXALIGNed. XLogInsertRecord fills in the xl_prev and
740 : * xl_crc fields in the header, the rest of the header must already be filled
741 : * by the caller.
742 : *
743 : * Returns XLOG pointer to end of record (beginning of next record).
744 : * This can be used as LSN for data pages affected by the logged action.
745 : * (LSN is the XLOG point up to which the XLOG must be flushed to disk
746 : * before the data page can be written out. This implements the basic
747 : * WAL rule "write the log before the data".)
748 : */
749 : XLogRecPtr
750 18264115 : XLogInsertRecord(XLogRecData *rdata,
751 : XLogRecPtr fpw_lsn,
752 : uint8 flags,
753 : int num_fpi,
754 : uint64 fpi_bytes,
755 : bool topxid_included)
756 : {
757 18264115 : XLogCtlInsert *Insert = &XLogCtl->Insert;
758 : pg_crc32c rdata_crc;
759 : bool inserted;
760 18264115 : XLogRecord *rechdr = (XLogRecord *) rdata->data;
761 18264115 : uint8 info = rechdr->xl_info & ~XLR_INFO_MASK;
762 18264115 : WalInsertClass class = WALINSERT_NORMAL;
763 : XLogRecPtr StartPos;
764 : XLogRecPtr EndPos;
765 18264115 : bool prevDoPageWrites = doPageWrites;
766 : TimeLineID insertTLI;
767 :
768 : /* Does this record type require special handling? */
769 18264115 : if (unlikely(rechdr->xl_rmid == RM_XLOG_ID))
770 : {
771 347067 : if (info == XLOG_SWITCH)
772 812 : class = WALINSERT_SPECIAL_SWITCH;
773 346255 : else if (info == XLOG_CHECKPOINT_REDO)
774 937 : class = WALINSERT_SPECIAL_CHECKPOINT;
775 : }
776 :
777 : /* we assume that all of the record header is in the first chunk */
778 : Assert(rdata->len >= SizeOfXLogRecord);
779 :
780 : /* cross-check on whether we should be here or not */
781 18264115 : if (!XLogInsertAllowed())
782 0 : elog(ERROR, "cannot make new WAL entries during recovery");
783 :
784 : /*
785 : * Given that we're not in recovery, InsertTimeLineID is set and can't
786 : * change, so we can read it without a lock.
787 : */
788 18264115 : insertTLI = XLogCtl->InsertTimeLineID;
789 :
790 : /*----------
791 : *
792 : * We have now done all the preparatory work we can without holding a
793 : * lock or modifying shared state. From here on, inserting the new WAL
794 : * record to the shared WAL buffer cache is a two-step process:
795 : *
796 : * 1. Reserve the right amount of space from the WAL. The current head of
797 : * reserved space is kept in Insert->CurrBytePos, and is protected by
798 : * insertpos_lck.
799 : *
800 : * 2. Copy the record to the reserved WAL space. This involves finding the
801 : * correct WAL buffer containing the reserved space, and copying the
802 : * record in place. This can be done concurrently in multiple processes.
803 : *
804 : * To keep track of which insertions are still in-progress, each concurrent
805 : * inserter acquires an insertion lock. In addition to just indicating that
806 : * an insertion is in progress, the lock tells others how far the inserter
807 : * has progressed. There is a small fixed number of insertion locks,
808 : * determined by NUM_XLOGINSERT_LOCKS. When an inserter crosses a page
809 : * boundary, it updates the value stored in the lock to the how far it has
810 : * inserted, to allow the previous buffer to be flushed.
811 : *
812 : * Holding onto an insertion lock also protects RedoRecPtr and
813 : * fullPageWrites from changing until the insertion is finished.
814 : *
815 : * Step 2 can usually be done completely in parallel. If the required WAL
816 : * page is not initialized yet, you have to grab WALBufMappingLock to
817 : * initialize it, but the WAL writer tries to do that ahead of insertions
818 : * to avoid that from happening in the critical path.
819 : *
820 : *----------
821 : */
822 18264115 : START_CRIT_SECTION();
823 :
824 18264115 : if (likely(class == WALINSERT_NORMAL))
825 : {
826 18262366 : WALInsertLockAcquire();
827 :
828 : /*
829 : * Check to see if my copy of RedoRecPtr is out of date. If so, may
830 : * have to go back and have the caller recompute everything. This can
831 : * only happen just after a checkpoint, so it's better to be slow in
832 : * this case and fast otherwise.
833 : *
834 : * Also check to see if fullPageWrites was just turned on or there's a
835 : * running backup (which forces full-page writes); if we weren't
836 : * already doing full-page writes then go back and recompute.
837 : *
838 : * If we aren't doing full-page writes then RedoRecPtr doesn't
839 : * actually affect the contents of the XLOG record, so we'll update
840 : * our local copy but not force a recomputation. (If doPageWrites was
841 : * just turned off, we could recompute the record without full pages,
842 : * but we choose not to bother.)
843 : */
844 18262366 : if (RedoRecPtr != Insert->RedoRecPtr)
845 : {
846 : Assert(RedoRecPtr < Insert->RedoRecPtr);
847 7906 : RedoRecPtr = Insert->RedoRecPtr;
848 : }
849 18262366 : doPageWrites = (Insert->fullPageWrites || Insert->runningBackups > 0);
850 :
851 18262366 : if (doPageWrites &&
852 17994854 : (!prevDoPageWrites ||
853 16567656 : (XLogRecPtrIsValid(fpw_lsn) && fpw_lsn <= RedoRecPtr)))
854 : {
855 : /*
856 : * Oops, some buffer now needs to be backed up that the caller
857 : * didn't back up. Start over.
858 : */
859 8568 : WALInsertLockRelease();
860 8568 : END_CRIT_SECTION();
861 8568 : return InvalidXLogRecPtr;
862 : }
863 :
864 : /*
865 : * Reserve space for the record in the WAL. This also sets the xl_prev
866 : * pointer.
867 : */
868 18253798 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
869 : &rechdr->xl_prev);
870 :
871 : /* Normal records are always inserted. */
872 18253798 : inserted = true;
873 : }
874 1749 : else if (class == WALINSERT_SPECIAL_SWITCH)
875 : {
876 : /*
877 : * In order to insert an XLOG_SWITCH record, we need to hold all of
878 : * the WAL insertion locks, not just one, so that no one else can
879 : * begin inserting a record until we've figured out how much space
880 : * remains in the current WAL segment and claimed all of it.
881 : *
882 : * Nonetheless, this case is simpler than the normal cases handled
883 : * below, which must check for changes in doPageWrites and RedoRecPtr.
884 : * Those checks are only needed for records that can contain buffer
885 : * references, and an XLOG_SWITCH record never does.
886 : */
887 : Assert(!XLogRecPtrIsValid(fpw_lsn));
888 812 : WALInsertLockAcquireExclusive();
889 812 : inserted = ReserveXLogSwitch(&StartPos, &EndPos, &rechdr->xl_prev);
890 : }
891 : else
892 : {
893 : Assert(class == WALINSERT_SPECIAL_CHECKPOINT);
894 :
895 : /*
896 : * We need to update both the local and shared copies of RedoRecPtr,
897 : * which means that we need to hold all the WAL insertion locks.
898 : * However, there can't be any buffer references, so as above, we need
899 : * not check RedoRecPtr before inserting the record; we just need to
900 : * update it afterwards.
901 : */
902 : Assert(!XLogRecPtrIsValid(fpw_lsn));
903 937 : WALInsertLockAcquireExclusive();
904 937 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
905 : &rechdr->xl_prev);
906 937 : RedoRecPtr = Insert->RedoRecPtr = StartPos;
907 937 : inserted = true;
908 : }
909 :
910 18255547 : if (inserted)
911 : {
912 : /*
913 : * Now that xl_prev has been filled in, calculate CRC of the record
914 : * header.
915 : */
916 18255485 : rdata_crc = rechdr->xl_crc;
917 18255485 : COMP_CRC32C(rdata_crc, rechdr, offsetof(XLogRecord, xl_crc));
918 18255485 : FIN_CRC32C(rdata_crc);
919 18255485 : rechdr->xl_crc = rdata_crc;
920 :
921 : /*
922 : * All the record data, including the header, is now ready to be
923 : * inserted. Copy the record in the space reserved.
924 : */
925 18255485 : CopyXLogRecordToWAL(rechdr->xl_tot_len,
926 : class == WALINSERT_SPECIAL_SWITCH, rdata,
927 : StartPos, EndPos, insertTLI);
928 :
929 : /*
930 : * Unless record is flagged as not important, update LSN of last
931 : * important record in the current slot. When holding all locks, just
932 : * update the first one.
933 : */
934 18255485 : if ((flags & XLOG_MARK_UNIMPORTANT) == 0)
935 : {
936 18036206 : int lockno = holdingAllLocks ? 0 : MyLockNo;
937 :
938 18036206 : WALInsertLocks[lockno].l.lastImportantAt = StartPos;
939 : }
940 : }
941 : else
942 : {
943 : /*
944 : * This was an xlog-switch record, but the current insert location was
945 : * already exactly at the beginning of a segment, so there was no need
946 : * to do anything.
947 : */
948 : }
949 :
950 : /*
951 : * Done! Let others know that we're finished.
952 : */
953 18255547 : WALInsertLockRelease();
954 :
955 18255547 : END_CRIT_SECTION();
956 :
957 18255547 : MarkCurrentTransactionIdLoggedIfAny();
958 :
959 : /*
960 : * Mark top transaction id is logged (if needed) so that we should not try
961 : * to log it again with the next WAL record in the current subtransaction.
962 : */
963 18255547 : if (topxid_included)
964 219 : MarkSubxactTopXidLogged();
965 :
966 : /*
967 : * Update shared LogwrtRqst.Write, if we crossed page boundary.
968 : */
969 18255547 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
970 : {
971 1769865 : SpinLockAcquire(&XLogCtl->info_lck);
972 : /* advance global request to include new block(s) */
973 1769865 : if (XLogCtl->LogwrtRqst.Write < EndPos)
974 1710319 : XLogCtl->LogwrtRqst.Write = EndPos;
975 1769865 : SpinLockRelease(&XLogCtl->info_lck);
976 1769865 : RefreshXLogWriteResult(LogwrtResult);
977 : }
978 :
979 : /*
980 : * If this was an XLOG_SWITCH record, flush the record and the empty
981 : * padding space that fills the rest of the segment, and perform
982 : * end-of-segment actions (eg, notifying archiver).
983 : */
984 18255547 : if (class == WALINSERT_SPECIAL_SWITCH)
985 : {
986 : TRACE_POSTGRESQL_WAL_SWITCH();
987 812 : XLogFlush(EndPos);
988 :
989 : /*
990 : * Even though we reserved the rest of the segment for us, which is
991 : * reflected in EndPos, we return a pointer to just the end of the
992 : * xlog-switch record.
993 : */
994 812 : if (inserted)
995 : {
996 750 : EndPos = StartPos + SizeOfXLogRecord;
997 750 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
998 : {
999 0 : uint64 offset = XLogSegmentOffset(EndPos, wal_segment_size);
1000 :
1001 0 : if (offset == EndPos % XLOG_BLCKSZ)
1002 0 : EndPos += SizeOfXLogLongPHD;
1003 : else
1004 0 : EndPos += SizeOfXLogShortPHD;
1005 : }
1006 : }
1007 : }
1008 :
1009 : #ifdef WAL_DEBUG
1010 : if (XLOG_DEBUG)
1011 : {
1012 : static XLogReaderState *debug_reader = NULL;
1013 : XLogRecord *record;
1014 : DecodedXLogRecord *decoded;
1015 : StringInfoData buf;
1016 : StringInfoData recordBuf;
1017 : char *errormsg = NULL;
1018 : MemoryContext oldCxt;
1019 :
1020 : oldCxt = MemoryContextSwitchTo(walDebugCxt);
1021 :
1022 : initStringInfo(&buf);
1023 : appendStringInfo(&buf, "INSERT @ %X/%08X: ", LSN_FORMAT_ARGS(EndPos));
1024 :
1025 : /*
1026 : * We have to piece together the WAL record data from the XLogRecData
1027 : * entries, so that we can pass it to the rm_desc function as one
1028 : * contiguous chunk.
1029 : */
1030 : initStringInfo(&recordBuf);
1031 : for (; rdata != NULL; rdata = rdata->next)
1032 : appendBinaryStringInfo(&recordBuf, rdata->data, rdata->len);
1033 :
1034 : /* We also need temporary space to decode the record. */
1035 : record = (XLogRecord *) recordBuf.data;
1036 : decoded = (DecodedXLogRecord *)
1037 : palloc(DecodeXLogRecordRequiredSpace(record->xl_tot_len));
1038 :
1039 : if (!debug_reader)
1040 : debug_reader = XLogReaderAllocate(wal_segment_size, NULL,
1041 : XL_ROUTINE(.page_read = NULL,
1042 : .segment_open = NULL,
1043 : .segment_close = NULL),
1044 : NULL);
1045 : if (!debug_reader)
1046 : {
1047 : appendStringInfoString(&buf, "error decoding record: out of memory while allocating a WAL reading processor");
1048 : }
1049 : else if (!DecodeXLogRecord(debug_reader,
1050 : decoded,
1051 : record,
1052 : EndPos,
1053 : &errormsg))
1054 : {
1055 : appendStringInfo(&buf, "error decoding record: %s",
1056 : errormsg ? errormsg : "no error message");
1057 : }
1058 : else
1059 : {
1060 : appendStringInfoString(&buf, " - ");
1061 :
1062 : debug_reader->record = decoded;
1063 : xlog_outdesc(&buf, debug_reader);
1064 : debug_reader->record = NULL;
1065 : }
1066 : elog(LOG, "%s", buf.data);
1067 :
1068 : pfree(decoded);
1069 : pfree(buf.data);
1070 : pfree(recordBuf.data);
1071 : MemoryContextSwitchTo(oldCxt);
1072 : }
1073 : #endif
1074 :
1075 : /*
1076 : * Update our global variables
1077 : */
1078 18255547 : ProcLastRecPtr = StartPos;
1079 18255547 : XactLastRecEnd = EndPos;
1080 :
1081 : /* Report WAL traffic to the instrumentation. */
1082 18255547 : if (inserted)
1083 : {
1084 18255485 : pgWalUsage.wal_bytes += rechdr->xl_tot_len;
1085 18255485 : pgWalUsage.wal_records++;
1086 18255485 : pgWalUsage.wal_fpi += num_fpi;
1087 18255485 : pgWalUsage.wal_fpi_bytes += fpi_bytes;
1088 :
1089 : /* Required for the flush of pending stats WAL data */
1090 18255485 : pgstat_report_fixed = true;
1091 : }
1092 :
1093 18255547 : return EndPos;
1094 : }
1095 :
1096 : /*
1097 : * Reserves the right amount of space for a record of given size from the WAL.
1098 : * *StartPos is set to the beginning of the reserved section, *EndPos to
1099 : * its end+1. *PrevPtr is set to the beginning of the previous record; it is
1100 : * used to set the xl_prev of this record.
1101 : *
1102 : * This is the performance critical part of XLogInsert that must be serialized
1103 : * across backends. The rest can happen mostly in parallel. Try to keep this
1104 : * section as short as possible, insertpos_lck can be heavily contended on a
1105 : * busy system.
1106 : *
1107 : * NB: The space calculation here must match the code in CopyXLogRecordToWAL,
1108 : * where we actually copy the record to the reserved space.
1109 : *
1110 : * NB: Testing shows that XLogInsertRecord runs faster if this code is inlined;
1111 : * however, because there are two call sites, the compiler is reluctant to
1112 : * inline. We use pg_attribute_always_inline here to try to convince it.
1113 : */
1114 : static pg_attribute_always_inline void
1115 18254735 : ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1116 : XLogRecPtr *PrevPtr)
1117 : {
1118 18254735 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1119 : uint64 startbytepos;
1120 : uint64 endbytepos;
1121 : uint64 prevbytepos;
1122 :
1123 18254735 : size = MAXALIGN(size);
1124 :
1125 : /* All (non xlog-switch) records should contain data. */
1126 : Assert(size > SizeOfXLogRecord);
1127 :
1128 : /*
1129 : * The duration the spinlock needs to be held is minimized by minimizing
1130 : * the calculations that have to be done while holding the lock. The
1131 : * current tip of reserved WAL is kept in CurrBytePos, as a byte position
1132 : * that only counts "usable" bytes in WAL, that is, it excludes all WAL
1133 : * page headers. The mapping between "usable" byte positions and physical
1134 : * positions (XLogRecPtrs) can be done outside the locked region, and
1135 : * because the usable byte position doesn't include any headers, reserving
1136 : * X bytes from WAL is almost as simple as "CurrBytePos += X".
1137 : */
1138 18254735 : SpinLockAcquire(&Insert->insertpos_lck);
1139 :
1140 18254735 : startbytepos = Insert->CurrBytePos;
1141 18254735 : endbytepos = startbytepos + size;
1142 18254735 : prevbytepos = Insert->PrevBytePos;
1143 18254735 : Insert->CurrBytePos = endbytepos;
1144 18254735 : Insert->PrevBytePos = startbytepos;
1145 :
1146 18254735 : SpinLockRelease(&Insert->insertpos_lck);
1147 :
1148 18254735 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1149 18254735 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1150 18254735 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1151 :
1152 : /*
1153 : * Check that the conversions between "usable byte positions" and
1154 : * XLogRecPtrs work consistently in both directions.
1155 : */
1156 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1157 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1158 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1159 18254735 : }
1160 :
1161 : /*
1162 : * Like ReserveXLogInsertLocation(), but for an xlog-switch record.
1163 : *
1164 : * A log-switch record is handled slightly differently. The rest of the
1165 : * segment will be reserved for this insertion, as indicated by the returned
1166 : * *EndPos value. However, if we are already at the beginning of the current
1167 : * segment, *StartPos and *EndPos are set to the current location without
1168 : * reserving any space, and the function returns false.
1169 : */
1170 : static bool
1171 812 : ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1172 : {
1173 812 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1174 : uint64 startbytepos;
1175 : uint64 endbytepos;
1176 : uint64 prevbytepos;
1177 812 : uint32 size = MAXALIGN(SizeOfXLogRecord);
1178 : XLogRecPtr ptr;
1179 : uint32 segleft;
1180 :
1181 : /*
1182 : * These calculations are a bit heavy-weight to be done while holding a
1183 : * spinlock, but since we're holding all the WAL insertion locks, there
1184 : * are no other inserters competing for it. GetXLogInsertRecPtr() does
1185 : * compete for it, but that's not called very frequently.
1186 : */
1187 812 : SpinLockAcquire(&Insert->insertpos_lck);
1188 :
1189 812 : startbytepos = Insert->CurrBytePos;
1190 :
1191 812 : ptr = XLogBytePosToEndRecPtr(startbytepos);
1192 812 : if (XLogSegmentOffset(ptr, wal_segment_size) == 0)
1193 : {
1194 62 : SpinLockRelease(&Insert->insertpos_lck);
1195 62 : *EndPos = *StartPos = ptr;
1196 62 : return false;
1197 : }
1198 :
1199 750 : endbytepos = startbytepos + size;
1200 750 : prevbytepos = Insert->PrevBytePos;
1201 :
1202 750 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1203 750 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1204 :
1205 750 : segleft = wal_segment_size - XLogSegmentOffset(*EndPos, wal_segment_size);
1206 750 : if (segleft != wal_segment_size)
1207 : {
1208 : /* consume the rest of the segment */
1209 750 : *EndPos += segleft;
1210 750 : endbytepos = XLogRecPtrToBytePos(*EndPos);
1211 : }
1212 750 : Insert->CurrBytePos = endbytepos;
1213 750 : Insert->PrevBytePos = startbytepos;
1214 :
1215 750 : SpinLockRelease(&Insert->insertpos_lck);
1216 :
1217 750 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1218 :
1219 : Assert(XLogSegmentOffset(*EndPos, wal_segment_size) == 0);
1220 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1221 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1222 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1223 :
1224 750 : return true;
1225 : }
1226 :
1227 : /*
1228 : * Subroutine of XLogInsertRecord. Copies a WAL record to an already-reserved
1229 : * area in the WAL.
1230 : */
1231 : static void
1232 18255485 : CopyXLogRecordToWAL(int write_len, bool isLogSwitch, XLogRecData *rdata,
1233 : XLogRecPtr StartPos, XLogRecPtr EndPos, TimeLineID tli)
1234 : {
1235 : char *currpos;
1236 : int freespace;
1237 : int written;
1238 : XLogRecPtr CurrPos;
1239 : XLogPageHeader pagehdr;
1240 :
1241 : /*
1242 : * Get a pointer to the right place in the right WAL buffer to start
1243 : * inserting to.
1244 : */
1245 18255485 : CurrPos = StartPos;
1246 18255485 : currpos = GetXLogBuffer(CurrPos, tli);
1247 18255485 : freespace = INSERT_FREESPACE(CurrPos);
1248 :
1249 : /*
1250 : * there should be enough space for at least the first field (xl_tot_len)
1251 : * on this page.
1252 : */
1253 : Assert(freespace >= sizeof(uint32));
1254 :
1255 : /* Copy record data */
1256 18255485 : written = 0;
1257 85050760 : while (rdata != NULL)
1258 : {
1259 66795275 : const char *rdata_data = rdata->data;
1260 66795275 : int rdata_len = rdata->len;
1261 :
1262 68688712 : while (rdata_len > freespace)
1263 : {
1264 : /*
1265 : * Write what fits on this page, and continue on the next page.
1266 : */
1267 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || freespace == 0);
1268 1893437 : memcpy(currpos, rdata_data, freespace);
1269 1893437 : rdata_data += freespace;
1270 1893437 : rdata_len -= freespace;
1271 1893437 : written += freespace;
1272 1893437 : CurrPos += freespace;
1273 :
1274 : /*
1275 : * Get pointer to beginning of next page, and set the xlp_rem_len
1276 : * in the page header. Set XLP_FIRST_IS_CONTRECORD.
1277 : *
1278 : * It's safe to set the contrecord flag and xlp_rem_len without a
1279 : * lock on the page. All the other flags were already set when the
1280 : * page was initialized, in AdvanceXLInsertBuffer, and we're the
1281 : * only backend that needs to set the contrecord flag.
1282 : */
1283 1893437 : currpos = GetXLogBuffer(CurrPos, tli);
1284 1893437 : pagehdr = (XLogPageHeader) currpos;
1285 1893437 : pagehdr->xlp_rem_len = write_len - written;
1286 1893437 : pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1287 :
1288 : /* skip over the page header */
1289 1893437 : if (XLogSegmentOffset(CurrPos, wal_segment_size) == 0)
1290 : {
1291 1223 : CurrPos += SizeOfXLogLongPHD;
1292 1223 : currpos += SizeOfXLogLongPHD;
1293 : }
1294 : else
1295 : {
1296 1892214 : CurrPos += SizeOfXLogShortPHD;
1297 1892214 : currpos += SizeOfXLogShortPHD;
1298 : }
1299 1893437 : freespace = INSERT_FREESPACE(CurrPos);
1300 : }
1301 :
1302 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1303 66795275 : memcpy(currpos, rdata_data, rdata_len);
1304 66795275 : currpos += rdata_len;
1305 66795275 : CurrPos += rdata_len;
1306 66795275 : freespace -= rdata_len;
1307 66795275 : written += rdata_len;
1308 :
1309 66795275 : rdata = rdata->next;
1310 : }
1311 : Assert(written == write_len);
1312 :
1313 : /*
1314 : * If this was an xlog-switch, it's not enough to write the switch record,
1315 : * we also have to consume all the remaining space in the WAL segment. We
1316 : * have already reserved that space, but we need to actually fill it.
1317 : */
1318 18255485 : if (isLogSwitch && XLogSegmentOffset(CurrPos, wal_segment_size) != 0)
1319 : {
1320 : /* An xlog-switch record doesn't contain any data besides the header */
1321 : Assert(write_len == SizeOfXLogRecord);
1322 :
1323 : /* Assert that we did reserve the right amount of space */
1324 : Assert(XLogSegmentOffset(EndPos, wal_segment_size) == 0);
1325 :
1326 : /* Use up all the remaining space on the current page */
1327 750 : CurrPos += freespace;
1328 :
1329 : /*
1330 : * Cause all remaining pages in the segment to be flushed, leaving the
1331 : * XLog position where it should be, at the start of the next segment.
1332 : * We do this one page at a time, to make sure we don't deadlock
1333 : * against ourselves if wal_buffers < wal_segment_size.
1334 : */
1335 747642 : while (CurrPos < EndPos)
1336 : {
1337 : /*
1338 : * The minimal action to flush the page would be to call
1339 : * WALInsertLockUpdateInsertingAt(CurrPos) followed by
1340 : * AdvanceXLInsertBuffer(...). The page would be left initialized
1341 : * mostly to zeros, except for the page header (always the short
1342 : * variant, as this is never a segment's first page).
1343 : *
1344 : * The large vistas of zeros are good for compressibility, but the
1345 : * headers interrupting them every XLOG_BLCKSZ (with values that
1346 : * differ from page to page) are not. The effect varies with
1347 : * compression tool, but bzip2 for instance compresses about an
1348 : * order of magnitude worse if those headers are left in place.
1349 : *
1350 : * Rather than complicating AdvanceXLInsertBuffer itself (which is
1351 : * called in heavily-loaded circumstances as well as this lightly-
1352 : * loaded one) with variant behavior, we just use GetXLogBuffer
1353 : * (which itself calls the two methods we need) to get the pointer
1354 : * and zero most of the page. Then we just zero the page header.
1355 : */
1356 746892 : currpos = GetXLogBuffer(CurrPos, tli);
1357 2987568 : MemSet(currpos, 0, SizeOfXLogShortPHD);
1358 :
1359 746892 : CurrPos += XLOG_BLCKSZ;
1360 : }
1361 : }
1362 : else
1363 : {
1364 : /* Align the end position, so that the next record starts aligned */
1365 18254735 : CurrPos = MAXALIGN64(CurrPos);
1366 : }
1367 :
1368 18255485 : if (CurrPos != EndPos)
1369 0 : ereport(PANIC,
1370 : errcode(ERRCODE_DATA_CORRUPTED),
1371 : errmsg_internal("space reserved for WAL record does not match what was written"));
1372 18255485 : }
1373 :
1374 : /*
1375 : * Acquire a WAL insertion lock, for inserting to WAL.
1376 : */
1377 : static void
1378 18262376 : WALInsertLockAcquire(void)
1379 : {
1380 : bool immed;
1381 :
1382 : /*
1383 : * It doesn't matter which of the WAL insertion locks we acquire, so try
1384 : * the one we used last time. If the system isn't particularly busy, it's
1385 : * a good bet that it's still available, and it's good to have some
1386 : * affinity to a particular lock so that you don't unnecessarily bounce
1387 : * cache lines between processes when there's no contention.
1388 : *
1389 : * If this is the first time through in this backend, pick a lock
1390 : * (semi-)randomly. This allows the locks to be used evenly if you have a
1391 : * lot of very short connections.
1392 : */
1393 : static int lockToTry = -1;
1394 :
1395 18262376 : if (lockToTry == -1)
1396 9005 : lockToTry = MyProcNumber % NUM_XLOGINSERT_LOCKS;
1397 18262376 : MyLockNo = lockToTry;
1398 :
1399 : /*
1400 : * The insertingAt value is initially set to 0, as we don't know our
1401 : * insert location yet.
1402 : */
1403 18262376 : immed = LWLockAcquire(&WALInsertLocks[MyLockNo].l.lock, LW_EXCLUSIVE);
1404 18262376 : if (!immed)
1405 : {
1406 : /*
1407 : * If we couldn't get the lock immediately, try another lock next
1408 : * time. On a system with more insertion locks than concurrent
1409 : * inserters, this causes all the inserters to eventually migrate to a
1410 : * lock that no-one else is using. On a system with more inserters
1411 : * than locks, it still helps to distribute the inserters evenly
1412 : * across the locks.
1413 : */
1414 18048 : lockToTry = (lockToTry + 1) % NUM_XLOGINSERT_LOCKS;
1415 : }
1416 18262376 : }
1417 :
1418 : /*
1419 : * Acquire all WAL insertion locks, to prevent other backends from inserting
1420 : * to WAL.
1421 : */
1422 : static void
1423 4603 : WALInsertLockAcquireExclusive(void)
1424 : {
1425 : int i;
1426 :
1427 : /*
1428 : * When holding all the locks, all but the last lock's insertingAt
1429 : * indicator is set to 0xFFFFFFFFFFFFFFFF, which is higher than any real
1430 : * XLogRecPtr value, to make sure that no-one blocks waiting on those.
1431 : */
1432 36824 : for (i = 0; i < NUM_XLOGINSERT_LOCKS - 1; i++)
1433 : {
1434 32221 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1435 32221 : LWLockUpdateVar(&WALInsertLocks[i].l.lock,
1436 32221 : &WALInsertLocks[i].l.insertingAt,
1437 : PG_UINT64_MAX);
1438 : }
1439 : /* Variable value reset to 0 at release */
1440 4603 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1441 :
1442 4603 : holdingAllLocks = true;
1443 4603 : }
1444 :
1445 : /*
1446 : * Release our insertion lock (or locks, if we're holding them all).
1447 : *
1448 : * NB: Reset all variables to 0, so they cause LWLockWaitForVar to block the
1449 : * next time the lock is acquired.
1450 : */
1451 : static void
1452 18266979 : WALInsertLockRelease(void)
1453 : {
1454 18266979 : if (holdingAllLocks)
1455 : {
1456 : int i;
1457 :
1458 41427 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1459 36824 : LWLockReleaseClearVar(&WALInsertLocks[i].l.lock,
1460 36824 : &WALInsertLocks[i].l.insertingAt,
1461 : 0);
1462 :
1463 4603 : holdingAllLocks = false;
1464 : }
1465 : else
1466 : {
1467 18262376 : LWLockReleaseClearVar(&WALInsertLocks[MyLockNo].l.lock,
1468 18262376 : &WALInsertLocks[MyLockNo].l.insertingAt,
1469 : 0);
1470 : }
1471 18266979 : }
1472 :
1473 : /*
1474 : * Update our insertingAt value, to let others know that we've finished
1475 : * inserting up to that point.
1476 : */
1477 : static void
1478 2592661 : WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt)
1479 : {
1480 2592661 : if (holdingAllLocks)
1481 : {
1482 : /*
1483 : * We use the last lock to mark our actual position, see comments in
1484 : * WALInsertLockAcquireExclusive.
1485 : */
1486 744146 : LWLockUpdateVar(&WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.lock,
1487 744146 : &WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.insertingAt,
1488 : insertingAt);
1489 : }
1490 : else
1491 1848515 : LWLockUpdateVar(&WALInsertLocks[MyLockNo].l.lock,
1492 1848515 : &WALInsertLocks[MyLockNo].l.insertingAt,
1493 : insertingAt);
1494 2592661 : }
1495 :
1496 : /*
1497 : * Wait for any WAL insertions < upto to finish.
1498 : *
1499 : * Returns the location of the oldest insertion that is still in-progress.
1500 : * Any WAL prior to that point has been fully copied into WAL buffers, and
1501 : * can be flushed out to disk. Because this waits for any insertions older
1502 : * than 'upto' to finish, the return value is always >= 'upto'.
1503 : *
1504 : * Note: When you are about to write out WAL, you must call this function
1505 : * *before* acquiring WALWriteLock, to avoid deadlocks. This function might
1506 : * need to wait for an insertion to finish (or at least advance to next
1507 : * uninitialized page), and the inserter might need to evict an old WAL buffer
1508 : * to make room for a new one, which in turn requires WALWriteLock.
1509 : */
1510 : static XLogRecPtr
1511 2416807 : WaitXLogInsertionsToFinish(XLogRecPtr upto)
1512 : {
1513 : uint64 bytepos;
1514 : XLogRecPtr inserted;
1515 : XLogRecPtr reservedUpto;
1516 : XLogRecPtr finishedUpto;
1517 2416807 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1518 : int i;
1519 :
1520 2416807 : if (MyProc == NULL)
1521 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1522 :
1523 : /*
1524 : * Check if there's any work to do. Use a barrier to ensure we get the
1525 : * freshest value.
1526 : */
1527 2416807 : inserted = pg_atomic_read_membarrier_u64(&XLogCtl->logInsertResult);
1528 2416807 : if (upto <= inserted)
1529 1931186 : return inserted;
1530 :
1531 : /* Read the current insert position */
1532 485621 : SpinLockAcquire(&Insert->insertpos_lck);
1533 485621 : bytepos = Insert->CurrBytePos;
1534 485621 : SpinLockRelease(&Insert->insertpos_lck);
1535 485621 : reservedUpto = XLogBytePosToEndRecPtr(bytepos);
1536 :
1537 : /*
1538 : * No-one should request to flush a piece of WAL that hasn't even been
1539 : * reserved yet. However, it can happen if there is a block with a bogus
1540 : * LSN on disk, for example. XLogFlush checks for that situation and
1541 : * complains, but only after the flush. Here we just assume that to mean
1542 : * that all WAL that has been reserved needs to be finished. In this
1543 : * corner-case, the return value can be smaller than 'upto' argument.
1544 : */
1545 485621 : if (upto > reservedUpto)
1546 : {
1547 0 : ereport(LOG,
1548 : errmsg("request to flush past end of generated WAL; request %X/%08X, current position %X/%08X",
1549 : LSN_FORMAT_ARGS(upto), LSN_FORMAT_ARGS(reservedUpto)));
1550 0 : upto = reservedUpto;
1551 : }
1552 :
1553 : /*
1554 : * Loop through all the locks, sleeping on any in-progress insert older
1555 : * than 'upto'.
1556 : *
1557 : * finishedUpto is our return value, indicating the point upto which all
1558 : * the WAL insertions have been finished. Initialize it to the head of
1559 : * reserved WAL, and as we iterate through the insertion locks, back it
1560 : * out for any insertion that's still in progress.
1561 : */
1562 485621 : finishedUpto = reservedUpto;
1563 4370589 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1564 : {
1565 3884968 : XLogRecPtr insertingat = InvalidXLogRecPtr;
1566 :
1567 : do
1568 : {
1569 : /*
1570 : * See if this insertion is in progress. LWLockWaitForVar will
1571 : * wait for the lock to be released, or for the 'value' to be set
1572 : * by a LWLockUpdateVar call. When a lock is initially acquired,
1573 : * its value is 0 (InvalidXLogRecPtr), which means that we don't
1574 : * know where it's inserting yet. We will have to wait for it. If
1575 : * it's a small insertion, the record will most likely fit on the
1576 : * same page and the inserter will release the lock without ever
1577 : * calling LWLockUpdateVar. But if it has to sleep, it will
1578 : * advertise the insertion point with LWLockUpdateVar before
1579 : * sleeping.
1580 : *
1581 : * In this loop we are only waiting for insertions that started
1582 : * before WaitXLogInsertionsToFinish was called. The lack of
1583 : * memory barriers in the loop means that we might see locks as
1584 : * "unused" that have since become used. This is fine because
1585 : * they only can be used for later insertions that we would not
1586 : * want to wait on anyway. Not taking a lock to acquire the
1587 : * current insertingAt value means that we might see older
1588 : * insertingAt values. This is also fine, because if we read a
1589 : * value too old, we will add ourselves to the wait queue, which
1590 : * contains atomic operations.
1591 : */
1592 3982705 : if (LWLockWaitForVar(&WALInsertLocks[i].l.lock,
1593 3982705 : &WALInsertLocks[i].l.insertingAt,
1594 : insertingat, &insertingat))
1595 : {
1596 : /* the lock was free, so no insertion in progress */
1597 2786971 : insertingat = InvalidXLogRecPtr;
1598 2786971 : break;
1599 : }
1600 :
1601 : /*
1602 : * This insertion is still in progress. Have to wait, unless the
1603 : * inserter has proceeded past 'upto'.
1604 : */
1605 1195734 : } while (insertingat < upto);
1606 :
1607 3884968 : if (XLogRecPtrIsValid(insertingat) && insertingat < finishedUpto)
1608 390909 : finishedUpto = insertingat;
1609 : }
1610 :
1611 : /*
1612 : * Advance the limit we know to have been inserted and return the freshest
1613 : * value we know of, which might be beyond what we requested if somebody
1614 : * is concurrently doing this with an 'upto' pointer ahead of us.
1615 : */
1616 485621 : finishedUpto = pg_atomic_monotonic_advance_u64(&XLogCtl->logInsertResult,
1617 : finishedUpto);
1618 :
1619 485621 : return finishedUpto;
1620 : }
1621 :
1622 : /*
1623 : * Get a pointer to the right location in the WAL buffer containing the
1624 : * given XLogRecPtr.
1625 : *
1626 : * If the page is not initialized yet, it is initialized. That might require
1627 : * evicting an old dirty buffer from the buffer cache, which means I/O.
1628 : *
1629 : * The caller must ensure that the page containing the requested location
1630 : * isn't evicted yet, and won't be evicted. The way to ensure that is to
1631 : * hold onto a WAL insertion lock with the insertingAt position set to
1632 : * something <= ptr. GetXLogBuffer() will update insertingAt if it needs
1633 : * to evict an old page from the buffer. (This means that once you call
1634 : * GetXLogBuffer() with a given 'ptr', you must not access anything before
1635 : * that point anymore, and must not call GetXLogBuffer() with an older 'ptr'
1636 : * later, because older buffers might be recycled already)
1637 : */
1638 : static char *
1639 20895824 : GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli)
1640 : {
1641 : int idx;
1642 : XLogRecPtr endptr;
1643 : static uint64 cachedPage = 0;
1644 : static char *cachedPos = NULL;
1645 : XLogRecPtr expectedEndPtr;
1646 :
1647 : /*
1648 : * Fast path for the common case that we need to access again the same
1649 : * page as last time.
1650 : */
1651 20895824 : if (ptr / XLOG_BLCKSZ == cachedPage)
1652 : {
1653 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1654 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1655 17827609 : return cachedPos + ptr % XLOG_BLCKSZ;
1656 : }
1657 :
1658 : /*
1659 : * The XLog buffer cache is organized so that a page is always loaded to a
1660 : * particular buffer. That way we can easily calculate the buffer a given
1661 : * page must be loaded into, from the XLogRecPtr alone.
1662 : */
1663 3068215 : idx = XLogRecPtrToBufIdx(ptr);
1664 :
1665 : /*
1666 : * See what page is loaded in the buffer at the moment. It could be the
1667 : * page we're looking for, or something older. It can't be anything newer
1668 : * - that would imply the page we're looking for has already been written
1669 : * out to disk and evicted, and the caller is responsible for making sure
1670 : * that doesn't happen.
1671 : *
1672 : * We don't hold a lock while we read the value. If someone is just about
1673 : * to initialize or has just initialized the page, it's possible that we
1674 : * get InvalidXLogRecPtr. That's ok, we'll grab the mapping lock (in
1675 : * AdvanceXLInsertBuffer) and retry if we see anything other than the page
1676 : * we're looking for.
1677 : */
1678 3068215 : expectedEndPtr = ptr;
1679 3068215 : expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
1680 :
1681 3068215 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1682 3068215 : if (expectedEndPtr != endptr)
1683 : {
1684 : XLogRecPtr initializedUpto;
1685 :
1686 : /*
1687 : * Before calling AdvanceXLInsertBuffer(), which can block, let others
1688 : * know how far we're finished with inserting the record.
1689 : *
1690 : * NB: If 'ptr' points to just after the page header, advertise a
1691 : * position at the beginning of the page rather than 'ptr' itself. If
1692 : * there are no other insertions running, someone might try to flush
1693 : * up to our advertised location. If we advertised a position after
1694 : * the page header, someone might try to flush the page header, even
1695 : * though page might actually not be initialized yet. As the first
1696 : * inserter on the page, we are effectively responsible for making
1697 : * sure that it's initialized, before we let insertingAt to move past
1698 : * the page header.
1699 : */
1700 2592661 : if (ptr % XLOG_BLCKSZ == SizeOfXLogShortPHD &&
1701 7326 : XLogSegmentOffset(ptr, wal_segment_size) > XLOG_BLCKSZ)
1702 7326 : initializedUpto = ptr - SizeOfXLogShortPHD;
1703 2585335 : else if (ptr % XLOG_BLCKSZ == SizeOfXLogLongPHD &&
1704 1018 : XLogSegmentOffset(ptr, wal_segment_size) < XLOG_BLCKSZ)
1705 623 : initializedUpto = ptr - SizeOfXLogLongPHD;
1706 : else
1707 2584712 : initializedUpto = ptr;
1708 :
1709 2592661 : WALInsertLockUpdateInsertingAt(initializedUpto);
1710 :
1711 2592661 : AdvanceXLInsertBuffer(ptr, tli, false);
1712 2592661 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1713 :
1714 2592661 : if (expectedEndPtr != endptr)
1715 0 : elog(PANIC, "could not find WAL buffer for %X/%08X",
1716 : LSN_FORMAT_ARGS(ptr));
1717 : }
1718 : else
1719 : {
1720 : /*
1721 : * Make sure the initialization of the page is visible to us, and
1722 : * won't arrive later to overwrite the WAL data we write on the page.
1723 : */
1724 475554 : pg_memory_barrier();
1725 : }
1726 :
1727 : /*
1728 : * Found the buffer holding this page. Return a pointer to the right
1729 : * offset within the page.
1730 : */
1731 3068215 : cachedPage = ptr / XLOG_BLCKSZ;
1732 3068215 : cachedPos = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1733 :
1734 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1735 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1736 :
1737 3068215 : return cachedPos + ptr % XLOG_BLCKSZ;
1738 : }
1739 :
1740 : /*
1741 : * Read WAL data directly from WAL buffers, if available. Returns the number
1742 : * of bytes read successfully.
1743 : *
1744 : * Fewer than 'count' bytes may be read if some of the requested WAL data has
1745 : * already been evicted.
1746 : *
1747 : * No locks are taken.
1748 : *
1749 : * Caller should ensure that it reads no further than LogwrtResult.Write
1750 : * (which should have been updated by the caller when determining how far to
1751 : * read). The 'tli' argument is only used as a convenient safety check so that
1752 : * callers do not read from WAL buffers on a historical timeline.
1753 : */
1754 : Size
1755 105391 : WALReadFromBuffers(char *dstbuf, XLogRecPtr startptr, Size count,
1756 : TimeLineID tli)
1757 : {
1758 105391 : char *pdst = dstbuf;
1759 105391 : XLogRecPtr recptr = startptr;
1760 : XLogRecPtr inserted;
1761 105391 : Size nbytes = count;
1762 :
1763 105391 : if (RecoveryInProgress() || tli != GetWALInsertionTimeLine())
1764 909 : return 0;
1765 :
1766 : Assert(XLogRecPtrIsValid(startptr));
1767 :
1768 : /*
1769 : * Caller should ensure that the requested data has been inserted into WAL
1770 : * buffers before we try to read it.
1771 : */
1772 104482 : inserted = pg_atomic_read_u64(&XLogCtl->logInsertResult);
1773 104482 : if (startptr + count > inserted)
1774 0 : ereport(ERROR,
1775 : errmsg("cannot read past end of generated WAL: requested %X/%08X, current position %X/%08X",
1776 : LSN_FORMAT_ARGS(startptr + count),
1777 : LSN_FORMAT_ARGS(inserted)));
1778 :
1779 : /*
1780 : * Loop through the buffers without a lock. For each buffer, atomically
1781 : * read and verify the end pointer, then copy the data out, and finally
1782 : * re-read and re-verify the end pointer.
1783 : *
1784 : * Once a page is evicted, it never returns to the WAL buffers, so if the
1785 : * end pointer matches the expected end pointer before and after we copy
1786 : * the data, then the right page must have been present during the data
1787 : * copy. Read barriers are necessary to ensure that the data copy actually
1788 : * happens between the two verification steps.
1789 : *
1790 : * If either verification fails, we simply terminate the loop and return
1791 : * with the data that had been already copied out successfully.
1792 : */
1793 134744 : while (nbytes > 0)
1794 : {
1795 125971 : uint32 offset = recptr % XLOG_BLCKSZ;
1796 125971 : int idx = XLogRecPtrToBufIdx(recptr);
1797 : XLogRecPtr expectedEndPtr;
1798 : XLogRecPtr endptr;
1799 : const char *page;
1800 : const char *psrc;
1801 : Size npagebytes;
1802 :
1803 : /*
1804 : * Calculate the end pointer we expect in the xlblocks array if the
1805 : * correct page is present.
1806 : */
1807 125971 : expectedEndPtr = recptr + (XLOG_BLCKSZ - offset);
1808 :
1809 : /*
1810 : * First verification step: check that the correct page is present in
1811 : * the WAL buffers.
1812 : */
1813 125971 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1814 125971 : if (expectedEndPtr != endptr)
1815 95705 : break;
1816 :
1817 : /*
1818 : * The correct page is present (or was at the time the endptr was
1819 : * read; must re-verify later). Calculate pointer to source data and
1820 : * determine how much data to read from this page.
1821 : */
1822 30266 : page = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1823 30266 : psrc = page + offset;
1824 30266 : npagebytes = Min(nbytes, XLOG_BLCKSZ - offset);
1825 :
1826 : /*
1827 : * Ensure that the data copy and the first verification step are not
1828 : * reordered.
1829 : */
1830 30266 : pg_read_barrier();
1831 :
1832 : /* data copy */
1833 30266 : memcpy(pdst, psrc, npagebytes);
1834 :
1835 : /*
1836 : * Ensure that the data copy and the second verification step are not
1837 : * reordered.
1838 : */
1839 30266 : pg_read_barrier();
1840 :
1841 : /*
1842 : * Second verification step: check that the page we read from wasn't
1843 : * evicted while we were copying the data.
1844 : */
1845 30266 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1846 30266 : if (expectedEndPtr != endptr)
1847 4 : break;
1848 :
1849 30262 : pdst += npagebytes;
1850 30262 : recptr += npagebytes;
1851 30262 : nbytes -= npagebytes;
1852 : }
1853 :
1854 : Assert(pdst - dstbuf <= count);
1855 :
1856 104482 : return pdst - dstbuf;
1857 : }
1858 :
1859 : /*
1860 : * Converts a "usable byte position" to XLogRecPtr. A usable byte position
1861 : * is the position starting from the beginning of WAL, excluding all WAL
1862 : * page headers.
1863 : */
1864 : static XLogRecPtr
1865 36513802 : XLogBytePosToRecPtr(uint64 bytepos)
1866 : {
1867 : uint64 fullsegs;
1868 : uint64 fullpages;
1869 : uint64 bytesleft;
1870 : uint32 seg_offset;
1871 : XLogRecPtr result;
1872 :
1873 36513802 : fullsegs = bytepos / UsableBytesInSegment;
1874 36513802 : bytesleft = bytepos % UsableBytesInSegment;
1875 :
1876 36513802 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1877 : {
1878 : /* fits on first page of segment */
1879 63565 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1880 : }
1881 : else
1882 : {
1883 : /* account for the first page on segment with long header */
1884 36450237 : seg_offset = XLOG_BLCKSZ;
1885 36450237 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1886 :
1887 36450237 : fullpages = bytesleft / UsableBytesInPage;
1888 36450237 : bytesleft = bytesleft % UsableBytesInPage;
1889 :
1890 36450237 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1891 : }
1892 :
1893 36513802 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1894 :
1895 36513802 : return result;
1896 : }
1897 :
1898 : /*
1899 : * Like XLogBytePosToRecPtr, but if the position is at a page boundary,
1900 : * returns a pointer to the beginning of the page (ie. before page header),
1901 : * not to where the first xlog record on that page would go to. This is used
1902 : * when converting a pointer to the end of a record.
1903 : */
1904 : static XLogRecPtr
1905 18842400 : XLogBytePosToEndRecPtr(uint64 bytepos)
1906 : {
1907 : uint64 fullsegs;
1908 : uint64 fullpages;
1909 : uint64 bytesleft;
1910 : uint32 seg_offset;
1911 : XLogRecPtr result;
1912 :
1913 18842400 : fullsegs = bytepos / UsableBytesInSegment;
1914 18842400 : bytesleft = bytepos % UsableBytesInSegment;
1915 :
1916 18842400 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1917 : {
1918 : /* fits on first page of segment */
1919 104444 : if (bytesleft == 0)
1920 71247 : seg_offset = 0;
1921 : else
1922 33197 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1923 : }
1924 : else
1925 : {
1926 : /* account for the first page on segment with long header */
1927 18737956 : seg_offset = XLOG_BLCKSZ;
1928 18737956 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1929 :
1930 18737956 : fullpages = bytesleft / UsableBytesInPage;
1931 18737956 : bytesleft = bytesleft % UsableBytesInPage;
1932 :
1933 18737956 : if (bytesleft == 0)
1934 18259 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
1935 : else
1936 18719697 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1937 : }
1938 :
1939 18842400 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1940 :
1941 18842400 : return result;
1942 : }
1943 :
1944 : /*
1945 : * Convert an XLogRecPtr to a "usable byte position".
1946 : */
1947 : static uint64
1948 2690 : XLogRecPtrToBytePos(XLogRecPtr ptr)
1949 : {
1950 : uint64 fullsegs;
1951 : uint32 fullpages;
1952 : uint32 offset;
1953 : uint64 result;
1954 :
1955 2690 : XLByteToSeg(ptr, fullsegs, wal_segment_size);
1956 :
1957 2690 : fullpages = (XLogSegmentOffset(ptr, wal_segment_size)) / XLOG_BLCKSZ;
1958 2690 : offset = ptr % XLOG_BLCKSZ;
1959 :
1960 2690 : if (fullpages == 0)
1961 : {
1962 1048 : result = fullsegs * UsableBytesInSegment;
1963 1048 : if (offset > 0)
1964 : {
1965 : Assert(offset >= SizeOfXLogLongPHD);
1966 275 : result += offset - SizeOfXLogLongPHD;
1967 : }
1968 : }
1969 : else
1970 : {
1971 1642 : result = fullsegs * UsableBytesInSegment +
1972 1642 : (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
1973 1642 : (fullpages - 1) * UsableBytesInPage; /* full pages */
1974 1642 : if (offset > 0)
1975 : {
1976 : Assert(offset >= SizeOfXLogShortPHD);
1977 1633 : result += offset - SizeOfXLogShortPHD;
1978 : }
1979 : }
1980 :
1981 2690 : return result;
1982 : }
1983 :
1984 : /*
1985 : * Initialize XLOG buffers, writing out old buffers if they still contain
1986 : * unwritten data, upto the page containing 'upto'. Or if 'opportunistic' is
1987 : * true, initialize as many pages as we can without having to write out
1988 : * unwritten data. Any new pages are initialized to zeros, with pages headers
1989 : * initialized properly.
1990 : */
1991 : static void
1992 2597136 : AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
1993 : {
1994 : int nextidx;
1995 : XLogRecPtr OldPageRqstPtr;
1996 : XLogwrtRqst WriteRqst;
1997 2597136 : XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
1998 : XLogRecPtr NewPageBeginPtr;
1999 : XLogPageHeader NewPage;
2000 2597136 : int npages pg_attribute_unused() = 0;
2001 :
2002 2597136 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2003 :
2004 : /*
2005 : * Now that we have the lock, check if someone initialized the page
2006 : * already.
2007 : */
2008 7526933 : while (upto >= XLogCtl->InitializedUpTo || opportunistic)
2009 : {
2010 4934272 : nextidx = XLogRecPtrToBufIdx(XLogCtl->InitializedUpTo);
2011 :
2012 : /*
2013 : * Get ending-offset of the buffer page we need to replace (this may
2014 : * be zero if the buffer hasn't been used yet). Fall through if it's
2015 : * already written out.
2016 : */
2017 4934272 : OldPageRqstPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]);
2018 4934272 : if (LogwrtResult.Write < OldPageRqstPtr)
2019 : {
2020 : /*
2021 : * Nope, got work to do. If we just want to pre-initialize as much
2022 : * as we can without flushing, give up now.
2023 : */
2024 2275509 : if (opportunistic)
2025 4475 : break;
2026 :
2027 : /* Advance shared memory write request position */
2028 2271034 : SpinLockAcquire(&XLogCtl->info_lck);
2029 2271034 : if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
2030 697478 : XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
2031 2271034 : SpinLockRelease(&XLogCtl->info_lck);
2032 :
2033 : /*
2034 : * Acquire an up-to-date LogwrtResult value and see if we still
2035 : * need to write it or if someone else already did.
2036 : */
2037 2271034 : RefreshXLogWriteResult(LogwrtResult);
2038 2271034 : if (LogwrtResult.Write < OldPageRqstPtr)
2039 : {
2040 : /*
2041 : * Must acquire write lock. Release WALBufMappingLock first,
2042 : * to make sure that all insertions that we need to wait for
2043 : * can finish (up to this same position). Otherwise we risk
2044 : * deadlock.
2045 : */
2046 2253473 : LWLockRelease(WALBufMappingLock);
2047 :
2048 2253473 : WaitXLogInsertionsToFinish(OldPageRqstPtr);
2049 :
2050 2253473 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2051 :
2052 2253473 : RefreshXLogWriteResult(LogwrtResult);
2053 2253473 : if (LogwrtResult.Write >= OldPageRqstPtr)
2054 : {
2055 : /* OK, someone wrote it already */
2056 141278 : LWLockRelease(WALWriteLock);
2057 : }
2058 : else
2059 : {
2060 : /* Have to write it ourselves */
2061 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2062 2112195 : WriteRqst.Write = OldPageRqstPtr;
2063 2112195 : WriteRqst.Flush = InvalidXLogRecPtr;
2064 2112195 : XLogWrite(WriteRqst, tli, false);
2065 2112195 : LWLockRelease(WALWriteLock);
2066 2112195 : pgWalUsage.wal_buffers_full++;
2067 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_DONE();
2068 :
2069 : /*
2070 : * Required for the flush of pending stats WAL data, per
2071 : * update of pgWalUsage.
2072 : */
2073 2112195 : pgstat_report_fixed = true;
2074 : }
2075 : /* Re-acquire WALBufMappingLock and retry */
2076 2253473 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2077 2253473 : continue;
2078 : }
2079 : }
2080 :
2081 : /*
2082 : * Now the next buffer slot is free and we can set it up to be the
2083 : * next output page.
2084 : */
2085 2676324 : NewPageBeginPtr = XLogCtl->InitializedUpTo;
2086 2676324 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2087 :
2088 : Assert(XLogRecPtrToBufIdx(NewPageBeginPtr) == nextidx);
2089 :
2090 2676324 : NewPage = (XLogPageHeader) (XLogCtl->pages + nextidx * (Size) XLOG_BLCKSZ);
2091 :
2092 : /*
2093 : * Mark the xlblock with InvalidXLogRecPtr and issue a write barrier
2094 : * before initializing. Otherwise, the old page may be partially
2095 : * zeroed but look valid.
2096 : */
2097 2676324 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], InvalidXLogRecPtr);
2098 2676324 : pg_write_barrier();
2099 :
2100 : /*
2101 : * Be sure to re-zero the buffer so that bytes beyond what we've
2102 : * written will look like zeroes and not valid XLOG records...
2103 : */
2104 2676324 : MemSet(NewPage, 0, XLOG_BLCKSZ);
2105 :
2106 : /*
2107 : * Fill the new page's header
2108 : */
2109 2676324 : NewPage->xlp_magic = XLOG_PAGE_MAGIC;
2110 :
2111 : /* NewPage->xlp_info = 0; */ /* done by memset */
2112 2676324 : NewPage->xlp_tli = tli;
2113 2676324 : NewPage->xlp_pageaddr = NewPageBeginPtr;
2114 :
2115 : /* NewPage->xlp_rem_len = 0; */ /* done by memset */
2116 :
2117 : /*
2118 : * If first page of an XLOG segment file, make it a long header.
2119 : */
2120 2676324 : if ((XLogSegmentOffset(NewPage->xlp_pageaddr, wal_segment_size)) == 0)
2121 : {
2122 1866 : XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2123 :
2124 1866 : NewLongPage->xlp_sysid = ControlFile->system_identifier;
2125 1866 : NewLongPage->xlp_seg_size = wal_segment_size;
2126 1866 : NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2127 1866 : NewPage->xlp_info |= XLP_LONG_HEADER;
2128 : }
2129 :
2130 : /*
2131 : * Make sure the initialization of the page becomes visible to others
2132 : * before the xlblocks update. GetXLogBuffer() reads xlblocks without
2133 : * holding a lock.
2134 : */
2135 2676324 : pg_write_barrier();
2136 :
2137 2676324 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], NewPageEndPtr);
2138 2676324 : XLogCtl->InitializedUpTo = NewPageEndPtr;
2139 :
2140 2676324 : npages++;
2141 : }
2142 2597136 : LWLockRelease(WALBufMappingLock);
2143 :
2144 : #ifdef WAL_DEBUG
2145 : if (XLOG_DEBUG && npages > 0)
2146 : {
2147 : elog(DEBUG1, "initialized %d pages, up to %X/%08X",
2148 : npages, LSN_FORMAT_ARGS(NewPageEndPtr));
2149 : }
2150 : #endif
2151 2597136 : }
2152 :
2153 : /*
2154 : * Calculate CheckPointSegments based on max_wal_size_mb and
2155 : * checkpoint_completion_target.
2156 : */
2157 : static void
2158 9507 : CalculateCheckpointSegments(void)
2159 : {
2160 : double target;
2161 :
2162 : /*-------
2163 : * Calculate the distance at which to trigger a checkpoint, to avoid
2164 : * exceeding max_wal_size_mb. This is based on two assumptions:
2165 : *
2166 : * a) we keep WAL for only one checkpoint cycle (prior to PG11 we kept
2167 : * WAL for two checkpoint cycles to allow us to recover from the
2168 : * secondary checkpoint if the first checkpoint failed, though we
2169 : * only did this on the primary anyway, not on standby. Keeping just
2170 : * one checkpoint simplifies processing and reduces disk space in
2171 : * many smaller databases.)
2172 : * b) during checkpoint, we consume checkpoint_completion_target *
2173 : * number of segments consumed between checkpoints.
2174 : *-------
2175 : */
2176 9507 : target = (double) ConvertToXSegs(max_wal_size_mb, wal_segment_size) /
2177 9507 : (1.0 + CheckPointCompletionTarget);
2178 :
2179 : /* round down */
2180 9507 : CheckPointSegments = (int) target;
2181 :
2182 9507 : if (CheckPointSegments < 1)
2183 10 : CheckPointSegments = 1;
2184 9507 : }
2185 :
2186 : void
2187 7191 : assign_max_wal_size(int newval, void *extra)
2188 : {
2189 7191 : max_wal_size_mb = newval;
2190 7191 : CalculateCheckpointSegments();
2191 7191 : }
2192 :
2193 : void
2194 1222 : assign_checkpoint_completion_target(double newval, void *extra)
2195 : {
2196 1222 : CheckPointCompletionTarget = newval;
2197 1222 : CalculateCheckpointSegments();
2198 1222 : }
2199 :
2200 : bool
2201 2368 : check_wal_segment_size(int *newval, void **extra, GucSource source)
2202 : {
2203 2368 : if (!IsValidWalSegSize(*newval))
2204 : {
2205 0 : GUC_check_errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.");
2206 0 : return false;
2207 : }
2208 :
2209 2368 : return true;
2210 : }
2211 :
2212 : /*
2213 : * At a checkpoint, how many WAL segments to recycle as preallocated future
2214 : * XLOG segments? Returns the highest segment that should be preallocated.
2215 : */
2216 : static XLogSegNo
2217 1835 : XLOGfileslop(XLogRecPtr lastredoptr)
2218 : {
2219 : XLogSegNo minSegNo;
2220 : XLogSegNo maxSegNo;
2221 : double distance;
2222 : XLogSegNo recycleSegNo;
2223 :
2224 : /*
2225 : * Calculate the segment numbers that min_wal_size_mb and max_wal_size_mb
2226 : * correspond to. Always recycle enough segments to meet the minimum, and
2227 : * remove enough segments to stay below the maximum.
2228 : */
2229 1835 : minSegNo = lastredoptr / wal_segment_size +
2230 1835 : ConvertToXSegs(min_wal_size_mb, wal_segment_size) - 1;
2231 1835 : maxSegNo = lastredoptr / wal_segment_size +
2232 1835 : ConvertToXSegs(max_wal_size_mb, wal_segment_size) - 1;
2233 :
2234 : /*
2235 : * Between those limits, recycle enough segments to get us through to the
2236 : * estimated end of next checkpoint.
2237 : *
2238 : * To estimate where the next checkpoint will finish, assume that the
2239 : * system runs steadily consuming CheckPointDistanceEstimate bytes between
2240 : * every checkpoint.
2241 : */
2242 1835 : distance = (1.0 + CheckPointCompletionTarget) * CheckPointDistanceEstimate;
2243 : /* add 10% for good measure. */
2244 1835 : distance *= 1.10;
2245 :
2246 1835 : recycleSegNo = (XLogSegNo) ceil(((double) lastredoptr + distance) /
2247 : wal_segment_size);
2248 :
2249 1835 : if (recycleSegNo < minSegNo)
2250 1301 : recycleSegNo = minSegNo;
2251 1835 : if (recycleSegNo > maxSegNo)
2252 407 : recycleSegNo = maxSegNo;
2253 :
2254 1835 : return recycleSegNo;
2255 : }
2256 :
2257 : /*
2258 : * Check whether we've consumed enough xlog space that a checkpoint is needed.
2259 : *
2260 : * new_segno indicates a log file that has just been filled up (or read
2261 : * during recovery). We measure the distance from RedoRecPtr to new_segno
2262 : * and see if that exceeds CheckPointSegments.
2263 : *
2264 : * Note: it is caller's responsibility that RedoRecPtr is up-to-date.
2265 : */
2266 : bool
2267 5002 : XLogCheckpointNeeded(XLogSegNo new_segno)
2268 : {
2269 : XLogSegNo old_segno;
2270 :
2271 5002 : XLByteToSeg(RedoRecPtr, old_segno, wal_segment_size);
2272 :
2273 5002 : if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2274 3135 : return true;
2275 1867 : return false;
2276 : }
2277 :
2278 : /*
2279 : * Write and/or fsync the log at least as far as WriteRqst indicates.
2280 : *
2281 : * If flexible == true, we don't have to write as far as WriteRqst, but
2282 : * may stop at any convenient boundary (such as a cache or logfile boundary).
2283 : * This option allows us to avoid uselessly issuing multiple writes when a
2284 : * single one would do.
2285 : *
2286 : * Must be called with WALWriteLock held. WaitXLogInsertionsToFinish(WriteRqst)
2287 : * must be called before grabbing the lock, to make sure the data is ready to
2288 : * write.
2289 : */
2290 : static void
2291 2269864 : XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible)
2292 : {
2293 : bool ispartialpage;
2294 : bool last_iteration;
2295 : bool finishing_seg;
2296 : int curridx;
2297 : int npages;
2298 : int startidx;
2299 : uint32 startoffset;
2300 :
2301 : /* We should always be inside a critical section here */
2302 : Assert(CritSectionCount > 0);
2303 :
2304 : /*
2305 : * Update local LogwrtResult (caller probably did this already, but...)
2306 : */
2307 2269864 : RefreshXLogWriteResult(LogwrtResult);
2308 :
2309 : /*
2310 : * Since successive pages in the xlog cache are consecutively allocated,
2311 : * we can usually gather multiple pages together and issue just one
2312 : * write() call. npages is the number of pages we have determined can be
2313 : * written together; startidx is the cache block index of the first one,
2314 : * and startoffset is the file offset at which it should go. The latter
2315 : * two variables are only valid when npages > 0, but we must initialize
2316 : * all of them to keep the compiler quiet.
2317 : */
2318 2269864 : npages = 0;
2319 2269864 : startidx = 0;
2320 2269864 : startoffset = 0;
2321 :
2322 : /*
2323 : * Within the loop, curridx is the cache block index of the page to
2324 : * consider writing. Begin at the buffer containing the next unwritten
2325 : * page, or last partially written page.
2326 : */
2327 2269864 : curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2328 :
2329 4915327 : while (LogwrtResult.Write < WriteRqst.Write)
2330 : {
2331 : /*
2332 : * Make sure we're not ahead of the insert process. This could happen
2333 : * if we're passed a bogus WriteRqst.Write that is past the end of the
2334 : * last page that's been initialized by AdvanceXLInsertBuffer.
2335 : */
2336 2798669 : XLogRecPtr EndPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[curridx]);
2337 :
2338 2798669 : if (LogwrtResult.Write >= EndPtr)
2339 0 : elog(PANIC, "xlog write request %X/%08X is past end of log %X/%08X",
2340 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2341 : LSN_FORMAT_ARGS(EndPtr));
2342 :
2343 : /* Advance LogwrtResult.Write to end of current buffer page */
2344 2798669 : LogwrtResult.Write = EndPtr;
2345 2798669 : ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2346 :
2347 2798669 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2348 : wal_segment_size))
2349 : {
2350 : /*
2351 : * Switch to new logfile segment. We cannot have any pending
2352 : * pages here (since we dump what we have at segment end).
2353 : */
2354 : Assert(npages == 0);
2355 14695 : if (openLogFile >= 0)
2356 6603 : XLogFileClose();
2357 14695 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2358 : wal_segment_size);
2359 14695 : openLogTLI = tli;
2360 :
2361 : /* create/use new log file */
2362 14695 : openLogFile = XLogFileInit(openLogSegNo, tli);
2363 14695 : ReserveExternalFD();
2364 : }
2365 :
2366 : /* Make sure we have the current logfile open */
2367 2798669 : if (openLogFile < 0)
2368 : {
2369 0 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2370 : wal_segment_size);
2371 0 : openLogTLI = tli;
2372 0 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2373 0 : ReserveExternalFD();
2374 : }
2375 :
2376 : /* Add current page to the set of pending pages-to-dump */
2377 2798669 : if (npages == 0)
2378 : {
2379 : /* first of group */
2380 2284928 : startidx = curridx;
2381 2284928 : startoffset = XLogSegmentOffset(LogwrtResult.Write - XLOG_BLCKSZ,
2382 : wal_segment_size);
2383 : }
2384 2798669 : npages++;
2385 :
2386 : /*
2387 : * Dump the set if this will be the last loop iteration, or if we are
2388 : * at the last page of the cache area (since the next page won't be
2389 : * contiguous in memory), or if we are at the end of the logfile
2390 : * segment.
2391 : */
2392 2798669 : last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2393 :
2394 5447647 : finishing_seg = !ispartialpage &&
2395 2648978 : (startoffset + npages * XLOG_BLCKSZ) >= wal_segment_size;
2396 :
2397 2798669 : if (last_iteration ||
2398 530141 : curridx == XLogCtl->XLogCacheBlck ||
2399 : finishing_seg)
2400 : {
2401 : char *from;
2402 : Size nbytes;
2403 : Size nleft;
2404 : ssize_t written;
2405 : instr_time start;
2406 :
2407 : /* OK to write the page(s) */
2408 2284928 : from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2409 2284928 : nbytes = npages * (Size) XLOG_BLCKSZ;
2410 2284928 : nleft = nbytes;
2411 : do
2412 : {
2413 2284928 : errno = 0;
2414 :
2415 : /*
2416 : * Measure I/O timing to write WAL data, for pg_stat_io.
2417 : */
2418 2284928 : start = pgstat_prepare_io_time(track_wal_io_timing);
2419 :
2420 2284928 : pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE);
2421 2284928 : written = pg_pwrite(openLogFile, from, nleft, startoffset);
2422 2284928 : pgstat_report_wait_end();
2423 :
2424 2284928 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL,
2425 : IOOP_WRITE, start, 1, written);
2426 :
2427 2284928 : if (written <= 0)
2428 : {
2429 : char xlogfname[MAXFNAMELEN];
2430 : int save_errno;
2431 :
2432 0 : if (errno == EINTR)
2433 0 : continue;
2434 :
2435 0 : save_errno = errno;
2436 0 : XLogFileName(xlogfname, tli, openLogSegNo,
2437 : wal_segment_size);
2438 0 : errno = save_errno;
2439 0 : ereport(PANIC,
2440 : (errcode_for_file_access(),
2441 : errmsg("could not write to log file \"%s\" at offset %u, length %zu: %m",
2442 : xlogfname, startoffset, nleft)));
2443 : }
2444 2284928 : nleft -= written;
2445 2284928 : from += written;
2446 2284928 : startoffset += written;
2447 2284928 : } while (nleft > 0);
2448 :
2449 2284928 : npages = 0;
2450 :
2451 : /*
2452 : * If we just wrote the whole last page of a logfile segment,
2453 : * fsync the segment immediately. This avoids having to go back
2454 : * and re-open prior segments when an fsync request comes along
2455 : * later. Doing it here ensures that one and only one backend will
2456 : * perform this fsync.
2457 : *
2458 : * This is also the right place to notify the Archiver that the
2459 : * segment is ready to copy to archival storage, and to update the
2460 : * timer for archive_timeout, and to signal for a checkpoint if
2461 : * too many logfile segments have been used since the last
2462 : * checkpoint.
2463 : */
2464 2284928 : if (finishing_seg)
2465 : {
2466 1993 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2467 :
2468 : /* signal that we need to wakeup walsenders later */
2469 1993 : WalSndWakeupRequest();
2470 :
2471 1993 : LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2472 :
2473 1993 : if (XLogArchivingActive())
2474 412 : XLogArchiveNotifySeg(openLogSegNo, tli);
2475 :
2476 1993 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2477 1993 : XLogCtl->lastSegSwitchLSN = LogwrtResult.Flush;
2478 :
2479 : /*
2480 : * Request a checkpoint if we've consumed too much xlog since
2481 : * the last one. For speed, we first check using the local
2482 : * copy of RedoRecPtr, which might be out of date; if it looks
2483 : * like a checkpoint is needed, forcibly update RedoRecPtr and
2484 : * recheck.
2485 : */
2486 1993 : if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2487 : {
2488 269 : (void) GetRedoRecPtr();
2489 269 : if (XLogCheckpointNeeded(openLogSegNo))
2490 215 : RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2491 : }
2492 : }
2493 : }
2494 :
2495 2798669 : if (ispartialpage)
2496 : {
2497 : /* Only asked to write a partial page */
2498 149691 : LogwrtResult.Write = WriteRqst.Write;
2499 149691 : break;
2500 : }
2501 2648978 : curridx = NextBufIdx(curridx);
2502 :
2503 : /* If flexible, break out of loop as soon as we wrote something */
2504 2648978 : if (flexible && npages == 0)
2505 3515 : break;
2506 : }
2507 :
2508 : Assert(npages == 0);
2509 :
2510 : /*
2511 : * If asked to flush, do so
2512 : */
2513 2269864 : if (LogwrtResult.Flush < WriteRqst.Flush &&
2514 156887 : LogwrtResult.Flush < LogwrtResult.Write)
2515 : {
2516 : /*
2517 : * Could get here without iterating above loop, in which case we might
2518 : * have no open file or the wrong one. However, we do not need to
2519 : * fsync more than one file.
2520 : */
2521 156811 : if (wal_sync_method != WAL_SYNC_METHOD_OPEN &&
2522 156811 : wal_sync_method != WAL_SYNC_METHOD_OPEN_DSYNC)
2523 : {
2524 156811 : if (openLogFile >= 0 &&
2525 156800 : !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2526 : wal_segment_size))
2527 86 : XLogFileClose();
2528 156811 : if (openLogFile < 0)
2529 : {
2530 97 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2531 : wal_segment_size);
2532 97 : openLogTLI = tli;
2533 97 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2534 97 : ReserveExternalFD();
2535 : }
2536 :
2537 156811 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2538 : }
2539 :
2540 : /* signal that we need to wakeup walsenders later */
2541 156811 : WalSndWakeupRequest();
2542 :
2543 156811 : LogwrtResult.Flush = LogwrtResult.Write;
2544 : }
2545 :
2546 : /*
2547 : * Update shared-memory status
2548 : *
2549 : * We make sure that the shared 'request' values do not fall behind the
2550 : * 'result' values. This is not absolutely essential, but it saves some
2551 : * code in a couple of places.
2552 : */
2553 2269864 : SpinLockAcquire(&XLogCtl->info_lck);
2554 2269864 : if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
2555 140043 : XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
2556 2269864 : if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
2557 158386 : XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
2558 2269864 : SpinLockRelease(&XLogCtl->info_lck);
2559 :
2560 : /*
2561 : * We write Write first, bar, then Flush. When reading, the opposite must
2562 : * be done (with a matching barrier in between), so that we always see a
2563 : * Flush value that trails behind the Write value seen.
2564 : */
2565 2269864 : pg_atomic_write_u64(&XLogCtl->logWriteResult, LogwrtResult.Write);
2566 2269864 : pg_write_barrier();
2567 2269864 : pg_atomic_write_u64(&XLogCtl->logFlushResult, LogwrtResult.Flush);
2568 :
2569 : #ifdef USE_ASSERT_CHECKING
2570 : {
2571 : XLogRecPtr Flush;
2572 : XLogRecPtr Write;
2573 : XLogRecPtr Insert;
2574 :
2575 : Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult);
2576 : pg_read_barrier();
2577 : Write = pg_atomic_read_u64(&XLogCtl->logWriteResult);
2578 : pg_read_barrier();
2579 : Insert = pg_atomic_read_u64(&XLogCtl->logInsertResult);
2580 :
2581 : /* WAL written to disk is always ahead of WAL flushed */
2582 : Assert(Write >= Flush);
2583 :
2584 : /* WAL inserted to buffers is always ahead of WAL written */
2585 : Assert(Insert >= Write);
2586 : }
2587 : #endif
2588 2269864 : }
2589 :
2590 : /*
2591 : * Record the LSN for an asynchronous transaction commit/abort
2592 : * and nudge the WALWriter if there is work for it to do.
2593 : * (This should not be called for synchronous commits.)
2594 : */
2595 : void
2596 55858 : XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2597 : {
2598 55858 : XLogRecPtr WriteRqstPtr = asyncXactLSN;
2599 : bool sleeping;
2600 55858 : bool wakeup = false;
2601 : XLogRecPtr prevAsyncXactLSN;
2602 :
2603 55858 : SpinLockAcquire(&XLogCtl->info_lck);
2604 55858 : sleeping = XLogCtl->WalWriterSleeping;
2605 55858 : prevAsyncXactLSN = XLogCtl->asyncXactLSN;
2606 55858 : if (XLogCtl->asyncXactLSN < asyncXactLSN)
2607 55339 : XLogCtl->asyncXactLSN = asyncXactLSN;
2608 55858 : SpinLockRelease(&XLogCtl->info_lck);
2609 :
2610 : /*
2611 : * If somebody else already called this function with a more aggressive
2612 : * LSN, they will have done what we needed (and perhaps more).
2613 : */
2614 55858 : if (asyncXactLSN <= prevAsyncXactLSN)
2615 519 : return;
2616 :
2617 : /*
2618 : * If the WALWriter is sleeping, kick it to make it come out of low-power
2619 : * mode, so that this async commit will reach disk within the expected
2620 : * amount of time. Otherwise, determine whether it has enough WAL
2621 : * available to flush, the same way that XLogBackgroundFlush() does.
2622 : */
2623 55339 : if (sleeping)
2624 58 : wakeup = true;
2625 : else
2626 : {
2627 : int flushblocks;
2628 :
2629 55281 : RefreshXLogWriteResult(LogwrtResult);
2630 :
2631 55281 : flushblocks =
2632 55281 : WriteRqstPtr / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
2633 :
2634 55281 : if (WalWriterFlushAfter == 0 || flushblocks >= WalWriterFlushAfter)
2635 4615 : wakeup = true;
2636 : }
2637 :
2638 55339 : if (wakeup)
2639 : {
2640 4673 : volatile PROC_HDR *procglobal = ProcGlobal;
2641 4673 : ProcNumber walwriterProc = procglobal->walwriterProc;
2642 :
2643 4673 : if (walwriterProc != INVALID_PROC_NUMBER)
2644 330 : SetLatch(&GetPGProcByNumber(walwriterProc)->procLatch);
2645 : }
2646 : }
2647 :
2648 : /*
2649 : * Record the LSN up to which we can remove WAL because it's not required by
2650 : * any replication slot.
2651 : */
2652 : void
2653 43558 : XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn)
2654 : {
2655 43558 : SpinLockAcquire(&XLogCtl->info_lck);
2656 43558 : XLogCtl->replicationSlotMinLSN = lsn;
2657 43558 : SpinLockRelease(&XLogCtl->info_lck);
2658 43558 : }
2659 :
2660 :
2661 : /*
2662 : * Return the oldest LSN we must retain to satisfy the needs of some
2663 : * replication slot.
2664 : */
2665 : XLogRecPtr
2666 2460 : XLogGetReplicationSlotMinimumLSN(void)
2667 : {
2668 : XLogRecPtr retval;
2669 :
2670 2460 : SpinLockAcquire(&XLogCtl->info_lck);
2671 2460 : retval = XLogCtl->replicationSlotMinLSN;
2672 2460 : SpinLockRelease(&XLogCtl->info_lck);
2673 :
2674 2460 : return retval;
2675 : }
2676 :
2677 : /*
2678 : * Advance minRecoveryPoint in control file.
2679 : *
2680 : * If we crash during recovery, we must reach this point again before the
2681 : * database is consistent.
2682 : *
2683 : * If 'force' is true, 'lsn' argument is ignored. Otherwise, minRecoveryPoint
2684 : * is only updated if it's not already greater than or equal to 'lsn'.
2685 : */
2686 : static void
2687 112979 : UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2688 : {
2689 : /* Quick check using our local copy of the variable */
2690 112979 : if (!updateMinRecoveryPoint || (!force && lsn <= LocalMinRecoveryPoint))
2691 105815 : return;
2692 :
2693 : /*
2694 : * An invalid minRecoveryPoint means that we need to recover all the WAL,
2695 : * i.e., we're doing crash recovery. We never modify the control file's
2696 : * value in that case, so we can short-circuit future checks here too. The
2697 : * local values of minRecoveryPoint and minRecoveryPointTLI should not be
2698 : * updated until crash recovery finishes. We only do this for the startup
2699 : * process as it should not update its own reference of minRecoveryPoint
2700 : * until it has finished crash recovery to make sure that all WAL
2701 : * available is replayed in this case. This also saves from extra locks
2702 : * taken on the control file from the startup process.
2703 : */
2704 7164 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint) && InRecovery)
2705 : {
2706 31 : updateMinRecoveryPoint = false;
2707 31 : return;
2708 : }
2709 :
2710 7133 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2711 :
2712 : /* update local copy */
2713 7133 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
2714 7133 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2715 :
2716 7133 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
2717 3 : updateMinRecoveryPoint = false;
2718 7130 : else if (force || LocalMinRecoveryPoint < lsn)
2719 : {
2720 : XLogRecPtr newMinRecoveryPoint;
2721 : TimeLineID newMinRecoveryPointTLI;
2722 :
2723 : /*
2724 : * To avoid having to update the control file too often, we update it
2725 : * all the way to the last record being replayed, even though 'lsn'
2726 : * would suffice for correctness. This also allows the 'force' case
2727 : * to not need a valid 'lsn' value.
2728 : *
2729 : * Another important reason for doing it this way is that the passed
2730 : * 'lsn' value could be bogus, i.e., past the end of available WAL, if
2731 : * the caller got it from a corrupted heap page. Accepting such a
2732 : * value as the min recovery point would prevent us from coming up at
2733 : * all. Instead, we just log a warning and continue with recovery.
2734 : * (See also the comments about corrupt LSNs in XLogFlush.)
2735 : */
2736 5722 : newMinRecoveryPoint = GetCurrentReplayRecPtr(&newMinRecoveryPointTLI);
2737 5722 : if (!force && newMinRecoveryPoint < lsn)
2738 0 : elog(WARNING,
2739 : "xlog min recovery request %X/%08X is past current point %X/%08X",
2740 : LSN_FORMAT_ARGS(lsn), LSN_FORMAT_ARGS(newMinRecoveryPoint));
2741 :
2742 : /* update control file */
2743 5722 : if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2744 : {
2745 5369 : ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2746 5369 : ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2747 5369 : UpdateControlFile();
2748 5369 : LocalMinRecoveryPoint = newMinRecoveryPoint;
2749 5369 : LocalMinRecoveryPointTLI = newMinRecoveryPointTLI;
2750 :
2751 5369 : ereport(DEBUG2,
2752 : errmsg_internal("updated min recovery point to %X/%08X on timeline %u",
2753 : LSN_FORMAT_ARGS(newMinRecoveryPoint),
2754 : newMinRecoveryPointTLI));
2755 : }
2756 : }
2757 7133 : LWLockRelease(ControlFileLock);
2758 : }
2759 :
2760 : /*
2761 : * Ensure that all XLOG data through the given position is flushed to disk.
2762 : *
2763 : * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
2764 : * already held, and we try to avoid acquiring it if possible.
2765 : */
2766 : void
2767 762321 : XLogFlush(XLogRecPtr record)
2768 : {
2769 : XLogRecPtr WriteRqstPtr;
2770 : XLogwrtRqst WriteRqst;
2771 762321 : TimeLineID insertTLI = XLogCtl->InsertTimeLineID;
2772 :
2773 : /*
2774 : * During REDO, we are reading not writing WAL. Therefore, instead of
2775 : * trying to flush the WAL, we should update minRecoveryPoint instead. We
2776 : * test XLogInsertAllowed(), not InRecovery, because we need checkpointer
2777 : * to act this way too, and because when it tries to write the
2778 : * end-of-recovery checkpoint, it should indeed flush.
2779 : */
2780 762321 : if (!XLogInsertAllowed())
2781 : {
2782 112514 : UpdateMinRecoveryPoint(record, false);
2783 593977 : return;
2784 : }
2785 :
2786 : /* Quick exit if already known flushed */
2787 649807 : if (record <= LogwrtResult.Flush)
2788 481463 : return;
2789 :
2790 : #ifdef WAL_DEBUG
2791 : if (XLOG_DEBUG)
2792 : elog(LOG, "xlog flush request %X/%08X; write %X/%08X; flush %X/%08X",
2793 : LSN_FORMAT_ARGS(record),
2794 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2795 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2796 : #endif
2797 :
2798 168344 : START_CRIT_SECTION();
2799 :
2800 : /*
2801 : * Since fsync is usually a horribly expensive operation, we try to
2802 : * piggyback as much data as we can on each fsync: if we see any more data
2803 : * entered into the xlog buffer, we'll write and fsync that too, so that
2804 : * the final value of LogwrtResult.Flush is as large as possible. This
2805 : * gives us some chance of avoiding another fsync immediately after.
2806 : */
2807 :
2808 : /* initialize to given target; may increase below */
2809 168344 : WriteRqstPtr = record;
2810 :
2811 : /*
2812 : * Now wait until we get the write lock, or someone else does the flush
2813 : * for us.
2814 : */
2815 : for (;;)
2816 3324 : {
2817 : XLogRecPtr insertpos;
2818 :
2819 : /* done already? */
2820 171668 : RefreshXLogWriteResult(LogwrtResult);
2821 171668 : if (record <= LogwrtResult.Flush)
2822 12809 : break;
2823 :
2824 : /*
2825 : * Before actually performing the write, wait for all in-flight
2826 : * insertions to the pages we're about to write to finish.
2827 : */
2828 158859 : SpinLockAcquire(&XLogCtl->info_lck);
2829 158859 : if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
2830 8923 : WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
2831 158859 : SpinLockRelease(&XLogCtl->info_lck);
2832 158859 : insertpos = WaitXLogInsertionsToFinish(WriteRqstPtr);
2833 :
2834 : /*
2835 : * Try to get the write lock. If we can't get it immediately, wait
2836 : * until it's released, and recheck if we still need to do the flush
2837 : * or if the backend that held the lock did it for us already. This
2838 : * helps to maintain a good rate of group committing when the system
2839 : * is bottlenecked by the speed of fsyncing.
2840 : */
2841 158859 : if (!LWLockAcquireOrWait(WALWriteLock, LW_EXCLUSIVE))
2842 : {
2843 : /*
2844 : * The lock is now free, but we didn't acquire it yet. Before we
2845 : * do, loop back to check if someone else flushed the record for
2846 : * us already.
2847 : */
2848 3324 : continue;
2849 : }
2850 :
2851 : /* Got the lock; recheck whether request is satisfied */
2852 155535 : RefreshXLogWriteResult(LogwrtResult);
2853 155535 : if (record <= LogwrtResult.Flush)
2854 : {
2855 2259 : LWLockRelease(WALWriteLock);
2856 2259 : break;
2857 : }
2858 :
2859 : /*
2860 : * Sleep before flush! By adding a delay here, we may give further
2861 : * backends the opportunity to join the backlog of group commit
2862 : * followers; this can significantly improve transaction throughput,
2863 : * at the risk of increasing transaction latency.
2864 : *
2865 : * We do not sleep if enableFsync is not turned on, nor if there are
2866 : * fewer than CommitSiblings other backends with active transactions.
2867 : */
2868 153276 : if (CommitDelay > 0 && enableFsync &&
2869 0 : MinimumActiveBackends(CommitSiblings))
2870 : {
2871 0 : pgstat_report_wait_start(WAIT_EVENT_COMMIT_DELAY);
2872 0 : pg_usleep(CommitDelay);
2873 0 : pgstat_report_wait_end();
2874 :
2875 : /*
2876 : * Re-check how far we can now flush the WAL. It's generally not
2877 : * safe to call WaitXLogInsertionsToFinish while holding
2878 : * WALWriteLock, because an in-progress insertion might need to
2879 : * also grab WALWriteLock to make progress. But we know that all
2880 : * the insertions up to insertpos have already finished, because
2881 : * that's what the earlier WaitXLogInsertionsToFinish() returned.
2882 : * We're only calling it again to allow insertpos to be moved
2883 : * further forward, not to actually wait for anyone.
2884 : */
2885 0 : insertpos = WaitXLogInsertionsToFinish(insertpos);
2886 : }
2887 :
2888 : /* try to write/flush later additions to XLOG as well */
2889 153276 : WriteRqst.Write = insertpos;
2890 153276 : WriteRqst.Flush = insertpos;
2891 :
2892 153276 : XLogWrite(WriteRqst, insertTLI, false);
2893 :
2894 153276 : LWLockRelease(WALWriteLock);
2895 : /* done */
2896 153276 : break;
2897 : }
2898 :
2899 168344 : END_CRIT_SECTION();
2900 :
2901 : /* wake up walsenders now that we've released heavily contended locks */
2902 168344 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
2903 :
2904 : /*
2905 : * If we flushed an LSN that someone was waiting for, notify the waiters.
2906 : */
2907 336688 : if (waitLSNState &&
2908 168344 : (LogwrtResult.Flush >=
2909 168344 : pg_atomic_read_u64(&waitLSNState->minWaitedLSN[WAIT_LSN_TYPE_PRIMARY_FLUSH])))
2910 0 : WaitLSNWakeup(WAIT_LSN_TYPE_PRIMARY_FLUSH, LogwrtResult.Flush);
2911 :
2912 : /*
2913 : * If we still haven't flushed to the request point then we have a
2914 : * problem; most likely, the requested flush point is past end of XLOG.
2915 : * This has been seen to occur when a disk page has a corrupted LSN.
2916 : *
2917 : * Formerly we treated this as a PANIC condition, but that hurts the
2918 : * system's robustness rather than helping it: we do not want to take down
2919 : * the whole system due to corruption on one data page. In particular, if
2920 : * the bad page is encountered again during recovery then we would be
2921 : * unable to restart the database at all! (This scenario actually
2922 : * happened in the field several times with 7.1 releases.) As of 8.4, bad
2923 : * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
2924 : * the only time we can reach here during recovery is while flushing the
2925 : * end-of-recovery checkpoint record, and we don't expect that to have a
2926 : * bad LSN.
2927 : *
2928 : * Note that for calls from xact.c, the ERROR will be promoted to PANIC
2929 : * since xact.c calls this routine inside a critical section. However,
2930 : * calls from bufmgr.c are not within critical sections and so we will not
2931 : * force a restart for a bad LSN on a data page.
2932 : */
2933 168344 : if (LogwrtResult.Flush < record)
2934 0 : elog(ERROR,
2935 : "xlog flush request %X/%08X is not satisfied --- flushed only to %X/%08X",
2936 : LSN_FORMAT_ARGS(record),
2937 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2938 :
2939 : /*
2940 : * Cross-check XLogNeedsFlush(). Some of the checks of XLogFlush() and
2941 : * XLogNeedsFlush() are duplicated, and this assertion ensures that these
2942 : * remain consistent.
2943 : */
2944 : Assert(!XLogNeedsFlush(record));
2945 : }
2946 :
2947 : /*
2948 : * Write & flush xlog, but without specifying exactly where to.
2949 : *
2950 : * We normally write only completed blocks; but if there is nothing to do on
2951 : * that basis, we check for unwritten async commits in the current incomplete
2952 : * block, and write through the latest one of those. Thus, if async commits
2953 : * are not being used, we will write complete blocks only.
2954 : *
2955 : * If, based on the above, there's anything to write we do so immediately. But
2956 : * to avoid calling fsync, fdatasync et. al. at a rate that'd impact
2957 : * concurrent IO, we only flush WAL every wal_writer_delay ms, or if there's
2958 : * more than wal_writer_flush_after unflushed blocks.
2959 : *
2960 : * We can guarantee that async commits reach disk after at most three
2961 : * wal_writer_delay cycles. (When flushing complete blocks, we allow XLogWrite
2962 : * to write "flexibly", meaning it can stop at the end of the buffer ring;
2963 : * this makes a difference only with very high load or long wal_writer_delay,
2964 : * but imposes one extra cycle for the worst case for async commits.)
2965 : *
2966 : * This routine is invoked periodically by the background walwriter process.
2967 : *
2968 : * Returns true if there was any work to do, even if we skipped flushing due
2969 : * to wal_writer_delay/wal_writer_flush_after.
2970 : */
2971 : bool
2972 15410 : XLogBackgroundFlush(void)
2973 : {
2974 : XLogwrtRqst WriteRqst;
2975 15410 : bool flexible = true;
2976 : static TimestampTz lastflush;
2977 : TimestampTz now;
2978 : int flushblocks;
2979 : TimeLineID insertTLI;
2980 :
2981 : /* XLOG doesn't need flushing during recovery */
2982 15410 : if (RecoveryInProgress())
2983 0 : return false;
2984 :
2985 : /*
2986 : * Since we're not in recovery, InsertTimeLineID is set and can't change,
2987 : * so we can read it without a lock.
2988 : */
2989 15410 : insertTLI = XLogCtl->InsertTimeLineID;
2990 :
2991 : /* read updated LogwrtRqst */
2992 15410 : SpinLockAcquire(&XLogCtl->info_lck);
2993 15410 : WriteRqst = XLogCtl->LogwrtRqst;
2994 15410 : SpinLockRelease(&XLogCtl->info_lck);
2995 :
2996 : /* back off to last completed page boundary */
2997 15410 : WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
2998 :
2999 : /* if we have already flushed that far, consider async commit records */
3000 15410 : RefreshXLogWriteResult(LogwrtResult);
3001 15410 : if (WriteRqst.Write <= LogwrtResult.Flush)
3002 : {
3003 11777 : SpinLockAcquire(&XLogCtl->info_lck);
3004 11777 : WriteRqst.Write = XLogCtl->asyncXactLSN;
3005 11777 : SpinLockRelease(&XLogCtl->info_lck);
3006 11777 : flexible = false; /* ensure it all gets written */
3007 : }
3008 :
3009 : /*
3010 : * If already known flushed, we're done. Just need to check if we are
3011 : * holding an open file handle to a logfile that's no longer in use,
3012 : * preventing the file from being deleted.
3013 : */
3014 15410 : if (WriteRqst.Write <= LogwrtResult.Flush)
3015 : {
3016 10935 : if (openLogFile >= 0)
3017 : {
3018 6767 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
3019 : wal_segment_size))
3020 : {
3021 186 : XLogFileClose();
3022 : }
3023 : }
3024 10935 : return false;
3025 : }
3026 :
3027 : /*
3028 : * Determine how far to flush WAL, based on the wal_writer_delay and
3029 : * wal_writer_flush_after GUCs.
3030 : *
3031 : * Note that XLogSetAsyncXactLSN() performs similar calculation based on
3032 : * wal_writer_flush_after, to decide when to wake us up. Make sure the
3033 : * logic is the same in both places if you change this.
3034 : */
3035 4475 : now = GetCurrentTimestamp();
3036 4475 : flushblocks =
3037 4475 : WriteRqst.Write / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
3038 :
3039 4475 : if (WalWriterFlushAfter == 0 || lastflush == 0)
3040 : {
3041 : /* first call, or block based limits disabled */
3042 257 : WriteRqst.Flush = WriteRqst.Write;
3043 257 : lastflush = now;
3044 : }
3045 4218 : else if (TimestampDifferenceExceeds(lastflush, now, WalWriterDelay))
3046 : {
3047 : /*
3048 : * Flush the writes at least every WalWriterDelay ms. This is
3049 : * important to bound the amount of time it takes for an asynchronous
3050 : * commit to hit disk.
3051 : */
3052 4006 : WriteRqst.Flush = WriteRqst.Write;
3053 4006 : lastflush = now;
3054 : }
3055 212 : else if (flushblocks >= WalWriterFlushAfter)
3056 : {
3057 : /* exceeded wal_writer_flush_after blocks, flush */
3058 181 : WriteRqst.Flush = WriteRqst.Write;
3059 181 : lastflush = now;
3060 : }
3061 : else
3062 : {
3063 : /* no flushing, this time round */
3064 31 : WriteRqst.Flush = InvalidXLogRecPtr;
3065 : }
3066 :
3067 : #ifdef WAL_DEBUG
3068 : if (XLOG_DEBUG)
3069 : elog(LOG, "xlog bg flush request write %X/%08X; flush: %X/%08X, current is write %X/%08X; flush %X/%08X",
3070 : LSN_FORMAT_ARGS(WriteRqst.Write),
3071 : LSN_FORMAT_ARGS(WriteRqst.Flush),
3072 : LSN_FORMAT_ARGS(LogwrtResult.Write),
3073 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
3074 : #endif
3075 :
3076 4475 : START_CRIT_SECTION();
3077 :
3078 : /* now wait for any in-progress insertions to finish and get write lock */
3079 4475 : WaitXLogInsertionsToFinish(WriteRqst.Write);
3080 4475 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3081 4475 : RefreshXLogWriteResult(LogwrtResult);
3082 4475 : if (WriteRqst.Write > LogwrtResult.Write ||
3083 127 : WriteRqst.Flush > LogwrtResult.Flush)
3084 : {
3085 4393 : XLogWrite(WriteRqst, insertTLI, flexible);
3086 : }
3087 4475 : LWLockRelease(WALWriteLock);
3088 :
3089 4475 : END_CRIT_SECTION();
3090 :
3091 : /* wake up walsenders now that we've released heavily contended locks */
3092 4475 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3093 :
3094 : /*
3095 : * If we flushed an LSN that someone was waiting for, notify the waiters.
3096 : */
3097 8950 : if (waitLSNState &&
3098 4475 : (LogwrtResult.Flush >=
3099 4475 : pg_atomic_read_u64(&waitLSNState->minWaitedLSN[WAIT_LSN_TYPE_PRIMARY_FLUSH])))
3100 0 : WaitLSNWakeup(WAIT_LSN_TYPE_PRIMARY_FLUSH, LogwrtResult.Flush);
3101 :
3102 : /*
3103 : * Great, done. To take some work off the critical path, try to initialize
3104 : * as many of the no-longer-needed WAL buffers for future use as we can.
3105 : */
3106 4475 : AdvanceXLInsertBuffer(InvalidXLogRecPtr, insertTLI, true);
3107 :
3108 : /*
3109 : * If we determined that we need to write data, but somebody else
3110 : * wrote/flushed already, it should be considered as being active, to
3111 : * avoid hibernating too early.
3112 : */
3113 4475 : return true;
3114 : }
3115 :
3116 : /*
3117 : * Test whether XLOG data has been flushed up to (at least) the given
3118 : * position, or whether the minimum recovery point has been updated past
3119 : * the given position.
3120 : *
3121 : * Returns true if a flush is still needed, or if the minimum recovery point
3122 : * must be updated.
3123 : *
3124 : * It is possible that someone else is already in the process of flushing
3125 : * that far, or has updated the minimum recovery point up to the given
3126 : * position.
3127 : */
3128 : bool
3129 10629000 : XLogNeedsFlush(XLogRecPtr record)
3130 : {
3131 : /*
3132 : * During recovery, we don't flush WAL but update minRecoveryPoint
3133 : * instead. So "needs flush" is taken to mean whether minRecoveryPoint
3134 : * would need to be updated.
3135 : *
3136 : * Using XLogInsertAllowed() rather than RecoveryInProgress() matters for
3137 : * the case of an end-of-recovery checkpoint, where WAL data is flushed.
3138 : * This check should be consistent with the one in XLogFlush().
3139 : */
3140 10629000 : if (!XLogInsertAllowed())
3141 : {
3142 : /* Quick exit if already known to be updated or cannot be updated */
3143 496585 : if (!updateMinRecoveryPoint || record <= LocalMinRecoveryPoint)
3144 489136 : return false;
3145 :
3146 : /*
3147 : * An invalid minRecoveryPoint means that we need to recover all the
3148 : * WAL, i.e., we're doing crash recovery. We never modify the control
3149 : * file's value in that case, so we can short-circuit future checks
3150 : * here too. This triggers a quick exit path for the startup process,
3151 : * which cannot update its local copy of minRecoveryPoint as long as
3152 : * it has not replayed all WAL available when doing crash recovery.
3153 : */
3154 7449 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint) && InRecovery)
3155 : {
3156 0 : updateMinRecoveryPoint = false;
3157 0 : return false;
3158 : }
3159 :
3160 : /*
3161 : * Update local copy of minRecoveryPoint. But if the lock is busy,
3162 : * just return a conservative guess.
3163 : */
3164 7449 : if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3165 0 : return true;
3166 7449 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
3167 7449 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3168 7449 : LWLockRelease(ControlFileLock);
3169 :
3170 : /*
3171 : * Check minRecoveryPoint for any other process than the startup
3172 : * process doing crash recovery, which should not update the control
3173 : * file value if crash recovery is still running.
3174 : */
3175 7449 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
3176 0 : updateMinRecoveryPoint = false;
3177 :
3178 : /* check again */
3179 7449 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3180 88 : return false;
3181 : else
3182 7361 : return true;
3183 : }
3184 :
3185 : /* Quick exit if already known flushed */
3186 10132415 : if (record <= LogwrtResult.Flush)
3187 9911513 : return false;
3188 :
3189 : /* read LogwrtResult and update local state */
3190 220902 : RefreshXLogWriteResult(LogwrtResult);
3191 :
3192 : /* check again */
3193 220902 : if (record <= LogwrtResult.Flush)
3194 2843 : return false;
3195 :
3196 218059 : return true;
3197 : }
3198 :
3199 : /*
3200 : * Try to make a given XLOG file segment exist.
3201 : *
3202 : * logsegno: identify segment.
3203 : *
3204 : * *added: on return, true if this call raised the number of extant segments.
3205 : *
3206 : * path: on return, this char[MAXPGPATH] has the path to the logsegno file.
3207 : *
3208 : * Returns -1 or FD of opened file. A -1 here is not an error; a caller
3209 : * wanting an open segment should attempt to open "path", which usually will
3210 : * succeed. (This is weird, but it's efficient for the callers.)
3211 : */
3212 : static int
3213 15853 : XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,
3214 : bool *added, char *path)
3215 : {
3216 : char tmppath[MAXPGPATH];
3217 : XLogSegNo installed_segno;
3218 : XLogSegNo max_segno;
3219 : int fd;
3220 : int save_errno;
3221 15853 : int open_flags = O_RDWR | O_CREAT | O_EXCL | PG_BINARY;
3222 : instr_time io_start;
3223 :
3224 : Assert(logtli != 0);
3225 :
3226 15853 : XLogFilePath(path, logtli, logsegno, wal_segment_size);
3227 :
3228 : /*
3229 : * Try to use existent file (checkpoint maker may have created it already)
3230 : */
3231 15853 : *added = false;
3232 15853 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3233 15853 : get_sync_bit(wal_sync_method));
3234 15853 : if (fd < 0)
3235 : {
3236 1379 : if (errno != ENOENT)
3237 0 : ereport(ERROR,
3238 : (errcode_for_file_access(),
3239 : errmsg("could not open file \"%s\": %m", path)));
3240 : }
3241 : else
3242 14474 : return fd;
3243 :
3244 : /*
3245 : * Initialize an empty (all zeroes) segment. NOTE: it is possible that
3246 : * another process is doing the same thing. If so, we will end up
3247 : * pre-creating an extra log segment. That seems OK, and better than
3248 : * holding the lock throughout this lengthy process.
3249 : */
3250 1379 : elog(DEBUG2, "creating and filling new WAL file");
3251 :
3252 1379 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3253 :
3254 1379 : unlink(tmppath);
3255 :
3256 1379 : if (io_direct_flags & IO_DIRECT_WAL_INIT)
3257 0 : open_flags |= PG_O_DIRECT;
3258 :
3259 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3260 1379 : fd = BasicOpenFile(tmppath, open_flags);
3261 1379 : if (fd < 0)
3262 0 : ereport(ERROR,
3263 : (errcode_for_file_access(),
3264 : errmsg("could not create file \"%s\": %m", tmppath)));
3265 :
3266 : /* Measure I/O timing when initializing segment */
3267 1379 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3268 :
3269 1379 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE);
3270 1379 : save_errno = 0;
3271 1379 : if (wal_init_zero)
3272 : {
3273 : ssize_t rc;
3274 :
3275 : /*
3276 : * Zero-fill the file. With this setting, we do this the hard way to
3277 : * ensure that all the file space has really been allocated. On
3278 : * platforms that allow "holes" in files, just seeking to the end
3279 : * doesn't allocate intermediate space. This way, we know that we
3280 : * have all the space and (after the fsync below) that all the
3281 : * indirect blocks are down on disk. Therefore, fdatasync(2) or
3282 : * O_DSYNC will be sufficient to sync future writes to the log file.
3283 : */
3284 1379 : rc = pg_pwrite_zeros(fd, wal_segment_size, 0);
3285 :
3286 1379 : if (rc < 0)
3287 0 : save_errno = errno;
3288 : }
3289 : else
3290 : {
3291 : /*
3292 : * Otherwise, seeking to the end and writing a solitary byte is
3293 : * enough.
3294 : */
3295 0 : errno = 0;
3296 0 : if (pg_pwrite(fd, "\0", 1, wal_segment_size - 1) != 1)
3297 : {
3298 : /* if write didn't set errno, assume no disk space */
3299 0 : save_errno = errno ? errno : ENOSPC;
3300 : }
3301 : }
3302 1379 : pgstat_report_wait_end();
3303 :
3304 : /*
3305 : * A full segment worth of data is written when using wal_init_zero. One
3306 : * byte is written when not using it.
3307 : */
3308 1379 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT, IOOP_WRITE,
3309 : io_start, 1,
3310 1379 : wal_init_zero ? wal_segment_size : 1);
3311 :
3312 1379 : if (save_errno)
3313 : {
3314 : /*
3315 : * If we fail to make the file, delete it to release disk space
3316 : */
3317 0 : unlink(tmppath);
3318 :
3319 0 : close(fd);
3320 :
3321 0 : errno = save_errno;
3322 :
3323 0 : ereport(ERROR,
3324 : (errcode_for_file_access(),
3325 : errmsg("could not write to file \"%s\": %m", tmppath)));
3326 : }
3327 :
3328 : /* Measure I/O timing when flushing segment */
3329 1379 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3330 :
3331 1379 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC);
3332 1379 : if (pg_fsync(fd) != 0)
3333 : {
3334 0 : save_errno = errno;
3335 0 : close(fd);
3336 0 : errno = save_errno;
3337 0 : ereport(ERROR,
3338 : (errcode_for_file_access(),
3339 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3340 : }
3341 1379 : pgstat_report_wait_end();
3342 :
3343 1379 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT,
3344 : IOOP_FSYNC, io_start, 1, 0);
3345 :
3346 1379 : if (close(fd) != 0)
3347 0 : ereport(ERROR,
3348 : (errcode_for_file_access(),
3349 : errmsg("could not close file \"%s\": %m", tmppath)));
3350 :
3351 : /*
3352 : * Now move the segment into place with its final name. Cope with
3353 : * possibility that someone else has created the file while we were
3354 : * filling ours: if so, use ours to pre-create a future log segment.
3355 : */
3356 1379 : installed_segno = logsegno;
3357 :
3358 : /*
3359 : * XXX: What should we use as max_segno? We used to use XLOGfileslop when
3360 : * that was a constant, but that was always a bit dubious: normally, at a
3361 : * checkpoint, XLOGfileslop was the offset from the checkpoint record, but
3362 : * here, it was the offset from the insert location. We can't do the
3363 : * normal XLOGfileslop calculation here because we don't have access to
3364 : * the prior checkpoint's redo location. So somewhat arbitrarily, just use
3365 : * CheckPointSegments.
3366 : */
3367 1379 : max_segno = logsegno + CheckPointSegments;
3368 1379 : if (InstallXLogFileSegment(&installed_segno, tmppath, true, max_segno,
3369 : logtli))
3370 : {
3371 1379 : *added = true;
3372 1379 : elog(DEBUG2, "done creating and filling new WAL file");
3373 : }
3374 : else
3375 : {
3376 : /*
3377 : * No need for any more future segments, or InstallXLogFileSegment()
3378 : * failed to rename the file into place. If the rename failed, a
3379 : * caller opening the file may fail.
3380 : */
3381 0 : unlink(tmppath);
3382 0 : elog(DEBUG2, "abandoned new WAL file");
3383 : }
3384 :
3385 1379 : return -1;
3386 : }
3387 :
3388 : /*
3389 : * Create a new XLOG file segment, or open a pre-existing one.
3390 : *
3391 : * logsegno: identify segment to be created/opened.
3392 : *
3393 : * Returns FD of opened file.
3394 : *
3395 : * Note: errors here are ERROR not PANIC because we might or might not be
3396 : * inside a critical section (eg, during checkpoint there is no reason to
3397 : * take down the system on failure). They will promote to PANIC if we are
3398 : * in a critical section.
3399 : */
3400 : int
3401 15629 : XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)
3402 : {
3403 : bool ignore_added;
3404 : char path[MAXPGPATH];
3405 : int fd;
3406 :
3407 : Assert(logtli != 0);
3408 :
3409 15629 : fd = XLogFileInitInternal(logsegno, logtli, &ignore_added, path);
3410 15629 : if (fd >= 0)
3411 14314 : return fd;
3412 :
3413 : /* Now open original target segment (might not be file I just made) */
3414 1315 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3415 1315 : get_sync_bit(wal_sync_method));
3416 1315 : if (fd < 0)
3417 0 : ereport(ERROR,
3418 : (errcode_for_file_access(),
3419 : errmsg("could not open file \"%s\": %m", path)));
3420 1315 : return fd;
3421 : }
3422 :
3423 : /*
3424 : * Create a new XLOG file segment by copying a pre-existing one.
3425 : *
3426 : * destsegno: identify segment to be created.
3427 : *
3428 : * srcTLI, srcsegno: identify segment to be copied (could be from
3429 : * a different timeline)
3430 : *
3431 : * upto: how much of the source file to copy (the rest is filled with
3432 : * zeros)
3433 : *
3434 : * Currently this is only used during recovery, and so there are no locking
3435 : * considerations. But we should be just as tense as XLogFileInit to avoid
3436 : * emplacing a bogus file.
3437 : */
3438 : static void
3439 43 : XLogFileCopy(TimeLineID destTLI, XLogSegNo destsegno,
3440 : TimeLineID srcTLI, XLogSegNo srcsegno,
3441 : int upto)
3442 : {
3443 : char path[MAXPGPATH];
3444 : char tmppath[MAXPGPATH];
3445 : PGAlignedXLogBlock buffer;
3446 : int srcfd;
3447 : int fd;
3448 : int nbytes;
3449 :
3450 : /*
3451 : * Open the source file
3452 : */
3453 43 : XLogFilePath(path, srcTLI, srcsegno, wal_segment_size);
3454 43 : srcfd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
3455 43 : if (srcfd < 0)
3456 0 : ereport(ERROR,
3457 : (errcode_for_file_access(),
3458 : errmsg("could not open file \"%s\": %m", path)));
3459 :
3460 : /*
3461 : * Copy into a temp file name.
3462 : */
3463 43 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3464 :
3465 43 : unlink(tmppath);
3466 :
3467 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3468 43 : fd = OpenTransientFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
3469 43 : if (fd < 0)
3470 0 : ereport(ERROR,
3471 : (errcode_for_file_access(),
3472 : errmsg("could not create file \"%s\": %m", tmppath)));
3473 :
3474 : /*
3475 : * Do the data copying.
3476 : */
3477 88107 : for (nbytes = 0; nbytes < wal_segment_size; nbytes += sizeof(buffer))
3478 : {
3479 : int nread;
3480 :
3481 88064 : nread = upto - nbytes;
3482 :
3483 : /*
3484 : * The part that is not read from the source file is filled with
3485 : * zeros.
3486 : */
3487 88064 : if (nread < sizeof(buffer))
3488 43 : memset(buffer.data, 0, sizeof(buffer));
3489 :
3490 88064 : if (nread > 0)
3491 : {
3492 : int r;
3493 :
3494 2780 : if (nread > sizeof(buffer))
3495 2737 : nread = sizeof(buffer);
3496 2780 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ);
3497 2780 : r = read(srcfd, buffer.data, nread);
3498 2780 : if (r != nread)
3499 : {
3500 0 : if (r < 0)
3501 0 : ereport(ERROR,
3502 : (errcode_for_file_access(),
3503 : errmsg("could not read file \"%s\": %m",
3504 : path)));
3505 : else
3506 0 : ereport(ERROR,
3507 : (errcode(ERRCODE_DATA_CORRUPTED),
3508 : errmsg("could not read file \"%s\": read %d of %zu",
3509 : path, r, (Size) nread)));
3510 : }
3511 2780 : pgstat_report_wait_end();
3512 : }
3513 88064 : errno = 0;
3514 88064 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_WRITE);
3515 88064 : if ((int) write(fd, buffer.data, sizeof(buffer)) != (int) sizeof(buffer))
3516 : {
3517 0 : int save_errno = errno;
3518 :
3519 : /*
3520 : * If we fail to make the file, delete it to release disk space
3521 : */
3522 0 : unlink(tmppath);
3523 : /* if write didn't set errno, assume problem is no disk space */
3524 0 : errno = save_errno ? save_errno : ENOSPC;
3525 :
3526 0 : ereport(ERROR,
3527 : (errcode_for_file_access(),
3528 : errmsg("could not write to file \"%s\": %m", tmppath)));
3529 : }
3530 88064 : pgstat_report_wait_end();
3531 : }
3532 :
3533 43 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_SYNC);
3534 43 : if (pg_fsync(fd) != 0)
3535 0 : ereport(data_sync_elevel(ERROR),
3536 : (errcode_for_file_access(),
3537 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3538 43 : pgstat_report_wait_end();
3539 :
3540 43 : if (CloseTransientFile(fd) != 0)
3541 0 : ereport(ERROR,
3542 : (errcode_for_file_access(),
3543 : errmsg("could not close file \"%s\": %m", tmppath)));
3544 :
3545 43 : if (CloseTransientFile(srcfd) != 0)
3546 0 : ereport(ERROR,
3547 : (errcode_for_file_access(),
3548 : errmsg("could not close file \"%s\": %m", path)));
3549 :
3550 : /*
3551 : * Now move the segment into place with its final name.
3552 : */
3553 43 : if (!InstallXLogFileSegment(&destsegno, tmppath, false, 0, destTLI))
3554 0 : elog(ERROR, "InstallXLogFileSegment should not have failed");
3555 43 : }
3556 :
3557 : /*
3558 : * Install a new XLOG segment file as a current or future log segment.
3559 : *
3560 : * This is used both to install a newly-created segment (which has a temp
3561 : * filename while it's being created) and to recycle an old segment.
3562 : *
3563 : * *segno: identify segment to install as (or first possible target).
3564 : * When find_free is true, this is modified on return to indicate the
3565 : * actual installation location or last segment searched.
3566 : *
3567 : * tmppath: initial name of file to install. It will be renamed into place.
3568 : *
3569 : * find_free: if true, install the new segment at the first empty segno
3570 : * number at or after the passed numbers. If false, install the new segment
3571 : * exactly where specified, deleting any existing segment file there.
3572 : *
3573 : * max_segno: maximum segment number to install the new file as. Fail if no
3574 : * free slot is found between *segno and max_segno. (Ignored when find_free
3575 : * is false.)
3576 : *
3577 : * tli: The timeline on which the new segment should be installed.
3578 : *
3579 : * Returns true if the file was installed successfully. false indicates that
3580 : * max_segno limit was exceeded, the startup process has disabled this
3581 : * function for now, or an error occurred while renaming the file into place.
3582 : */
3583 : static bool
3584 3186 : InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
3585 : bool find_free, XLogSegNo max_segno, TimeLineID tli)
3586 : {
3587 : char path[MAXPGPATH];
3588 : struct stat stat_buf;
3589 :
3590 : Assert(tli != 0);
3591 :
3592 3186 : XLogFilePath(path, tli, *segno, wal_segment_size);
3593 :
3594 3186 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3595 3186 : if (!XLogCtl->InstallXLogFileSegmentActive)
3596 : {
3597 0 : LWLockRelease(ControlFileLock);
3598 0 : return false;
3599 : }
3600 :
3601 3186 : if (!find_free)
3602 : {
3603 : /* Force installation: get rid of any pre-existing segment file */
3604 43 : durable_unlink(path, DEBUG1);
3605 : }
3606 : else
3607 : {
3608 : /* Find a free slot to put it in */
3609 4553 : while (stat(path, &stat_buf) == 0)
3610 : {
3611 1629 : if ((*segno) >= max_segno)
3612 : {
3613 : /* Failed to find a free slot within specified range */
3614 219 : LWLockRelease(ControlFileLock);
3615 219 : return false;
3616 : }
3617 1410 : (*segno)++;
3618 1410 : XLogFilePath(path, tli, *segno, wal_segment_size);
3619 : }
3620 : }
3621 :
3622 : Assert(access(path, F_OK) != 0 && errno == ENOENT);
3623 2967 : if (durable_rename(tmppath, path, LOG) != 0)
3624 : {
3625 0 : LWLockRelease(ControlFileLock);
3626 : /* durable_rename already emitted log message */
3627 0 : return false;
3628 : }
3629 :
3630 2967 : LWLockRelease(ControlFileLock);
3631 :
3632 2967 : return true;
3633 : }
3634 :
3635 : /*
3636 : * Open a pre-existing logfile segment for writing.
3637 : */
3638 : int
3639 97 : XLogFileOpen(XLogSegNo segno, TimeLineID tli)
3640 : {
3641 : char path[MAXPGPATH];
3642 : int fd;
3643 :
3644 97 : XLogFilePath(path, tli, segno, wal_segment_size);
3645 :
3646 97 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3647 97 : get_sync_bit(wal_sync_method));
3648 97 : if (fd < 0)
3649 0 : ereport(PANIC,
3650 : (errcode_for_file_access(),
3651 : errmsg("could not open file \"%s\": %m", path)));
3652 :
3653 97 : return fd;
3654 : }
3655 :
3656 : /*
3657 : * Close the current logfile segment for writing.
3658 : */
3659 : static void
3660 6875 : XLogFileClose(void)
3661 : {
3662 : Assert(openLogFile >= 0);
3663 :
3664 : /*
3665 : * WAL segment files will not be re-read in normal operation, so we advise
3666 : * the OS to release any cached pages. But do not do so if WAL archiving
3667 : * or streaming is active, because archiver and walsender process could
3668 : * use the cache to read the WAL segment.
3669 : */
3670 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3671 6875 : if (!XLogIsNeeded() && (io_direct_flags & IO_DIRECT_WAL) == 0)
3672 239 : (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3673 : #endif
3674 :
3675 6875 : if (close(openLogFile) != 0)
3676 : {
3677 : char xlogfname[MAXFNAMELEN];
3678 0 : int save_errno = errno;
3679 :
3680 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo, wal_segment_size);
3681 0 : errno = save_errno;
3682 0 : ereport(PANIC,
3683 : (errcode_for_file_access(),
3684 : errmsg("could not close file \"%s\": %m", xlogfname)));
3685 : }
3686 :
3687 6875 : openLogFile = -1;
3688 6875 : ReleaseExternalFD();
3689 6875 : }
3690 :
3691 : /*
3692 : * Preallocate log files beyond the specified log endpoint.
3693 : *
3694 : * XXX this is currently extremely conservative, since it forces only one
3695 : * future log segment to exist, and even that only if we are 75% done with
3696 : * the current one. This is only appropriate for very low-WAL-volume systems.
3697 : * High-volume systems will be OK once they've built up a sufficient set of
3698 : * recycled log segments, but the startup transient is likely to include
3699 : * a lot of segment creations by foreground processes, which is not so good.
3700 : *
3701 : * XLogFileInitInternal() can ereport(ERROR). All known causes indicate big
3702 : * trouble; for example, a full filesystem is one cause. The checkpoint WAL
3703 : * and/or ControlFile updates already completed. If a RequestCheckpoint()
3704 : * initiated the present checkpoint and an ERROR ends this function, the
3705 : * command that called RequestCheckpoint() fails. That's not ideal, but it's
3706 : * not worth contorting more functions to use caller-specified elevel values.
3707 : * (With or without RequestCheckpoint(), an ERROR forestalls some inessential
3708 : * reporting and resource reclamation.)
3709 : */
3710 : static void
3711 2112 : PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli)
3712 : {
3713 : XLogSegNo _logSegNo;
3714 : int lf;
3715 : bool added;
3716 : char path[MAXPGPATH];
3717 : uint64 offset;
3718 :
3719 2112 : if (!XLogCtl->InstallXLogFileSegmentActive)
3720 12 : return; /* unlocked check says no */
3721 :
3722 2100 : XLByteToPrevSeg(endptr, _logSegNo, wal_segment_size);
3723 2100 : offset = XLogSegmentOffset(endptr - 1, wal_segment_size);
3724 2100 : if (offset >= (uint32) (0.75 * wal_segment_size))
3725 : {
3726 224 : _logSegNo++;
3727 224 : lf = XLogFileInitInternal(_logSegNo, tli, &added, path);
3728 224 : if (lf >= 0)
3729 160 : close(lf);
3730 224 : if (added)
3731 64 : CheckpointStats.ckpt_segs_added++;
3732 : }
3733 : }
3734 :
3735 : /*
3736 : * Throws an error if the given log segment has already been removed or
3737 : * recycled. The caller should only pass a segment that it knows to have
3738 : * existed while the server has been running, as this function always
3739 : * succeeds if no WAL segments have been removed since startup.
3740 : * 'tli' is only used in the error message.
3741 : *
3742 : * Note: this function guarantees to keep errno unchanged on return.
3743 : * This supports callers that use this to possibly deliver a better
3744 : * error message about a missing file, while still being able to throw
3745 : * a normal file-access error afterwards, if this does return.
3746 : */
3747 : void
3748 125269 : CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3749 : {
3750 125269 : int save_errno = errno;
3751 : XLogSegNo lastRemovedSegNo;
3752 :
3753 125269 : SpinLockAcquire(&XLogCtl->info_lck);
3754 125269 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3755 125269 : SpinLockRelease(&XLogCtl->info_lck);
3756 :
3757 125269 : if (segno <= lastRemovedSegNo)
3758 : {
3759 : char filename[MAXFNAMELEN];
3760 :
3761 0 : XLogFileName(filename, tli, segno, wal_segment_size);
3762 0 : errno = save_errno;
3763 0 : ereport(ERROR,
3764 : (errcode_for_file_access(),
3765 : errmsg("requested WAL segment %s has already been removed",
3766 : filename)));
3767 : }
3768 125269 : errno = save_errno;
3769 125269 : }
3770 :
3771 : /*
3772 : * Return the last WAL segment removed, or 0 if no segment has been removed
3773 : * since startup.
3774 : *
3775 : * NB: the result can be out of date arbitrarily fast, the caller has to deal
3776 : * with that.
3777 : */
3778 : XLogSegNo
3779 1271 : XLogGetLastRemovedSegno(void)
3780 : {
3781 : XLogSegNo lastRemovedSegNo;
3782 :
3783 1271 : SpinLockAcquire(&XLogCtl->info_lck);
3784 1271 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3785 1271 : SpinLockRelease(&XLogCtl->info_lck);
3786 :
3787 1271 : return lastRemovedSegNo;
3788 : }
3789 :
3790 : /*
3791 : * Return the oldest WAL segment on the given TLI that still exists in
3792 : * XLOGDIR, or 0 if none.
3793 : */
3794 : XLogSegNo
3795 7 : XLogGetOldestSegno(TimeLineID tli)
3796 : {
3797 : DIR *xldir;
3798 : struct dirent *xlde;
3799 7 : XLogSegNo oldest_segno = 0;
3800 :
3801 7 : xldir = AllocateDir(XLOGDIR);
3802 50 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3803 : {
3804 : TimeLineID file_tli;
3805 : XLogSegNo file_segno;
3806 :
3807 : /* Ignore files that are not XLOG segments. */
3808 43 : if (!IsXLogFileName(xlde->d_name))
3809 29 : continue;
3810 :
3811 : /* Parse filename to get TLI and segno. */
3812 14 : XLogFromFileName(xlde->d_name, &file_tli, &file_segno,
3813 : wal_segment_size);
3814 :
3815 : /* Ignore anything that's not from the TLI of interest. */
3816 14 : if (tli != file_tli)
3817 0 : continue;
3818 :
3819 : /* If it's the oldest so far, update oldest_segno. */
3820 14 : if (oldest_segno == 0 || file_segno < oldest_segno)
3821 9 : oldest_segno = file_segno;
3822 : }
3823 :
3824 7 : FreeDir(xldir);
3825 7 : return oldest_segno;
3826 : }
3827 :
3828 : /*
3829 : * Update the last removed segno pointer in shared memory, to reflect that the
3830 : * given XLOG file has been removed.
3831 : */
3832 : static void
3833 2708 : UpdateLastRemovedPtr(char *filename)
3834 : {
3835 : uint32 tli;
3836 : XLogSegNo segno;
3837 :
3838 2708 : XLogFromFileName(filename, &tli, &segno, wal_segment_size);
3839 :
3840 2708 : SpinLockAcquire(&XLogCtl->info_lck);
3841 2708 : if (segno > XLogCtl->lastRemovedSegNo)
3842 1271 : XLogCtl->lastRemovedSegNo = segno;
3843 2708 : SpinLockRelease(&XLogCtl->info_lck);
3844 2708 : }
3845 :
3846 : /*
3847 : * Remove all temporary log files in pg_wal
3848 : *
3849 : * This is called at the beginning of recovery after a previous crash,
3850 : * at a point where no other processes write fresh WAL data.
3851 : */
3852 : static void
3853 186 : RemoveTempXlogFiles(void)
3854 : {
3855 : DIR *xldir;
3856 : struct dirent *xlde;
3857 :
3858 186 : elog(DEBUG2, "removing all temporary WAL segments");
3859 :
3860 186 : xldir = AllocateDir(XLOGDIR);
3861 1259 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3862 : {
3863 : char path[MAXPGPATH];
3864 :
3865 1073 : if (strncmp(xlde->d_name, "xlogtemp.", 9) != 0)
3866 1073 : continue;
3867 :
3868 0 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
3869 0 : unlink(path);
3870 0 : elog(DEBUG2, "removed temporary WAL segment \"%s\"", path);
3871 : }
3872 186 : FreeDir(xldir);
3873 186 : }
3874 :
3875 : /*
3876 : * Recycle or remove all log files older or equal to passed segno.
3877 : *
3878 : * endptr is current (or recent) end of xlog, and lastredoptr is the
3879 : * redo pointer of the last checkpoint. These are used to determine
3880 : * whether we want to recycle rather than delete no-longer-wanted log files.
3881 : *
3882 : * insertTLI is the current timeline for XLOG insertion. Any recycled
3883 : * segments should be reused for this timeline.
3884 : */
3885 : static void
3886 1835 : RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr, XLogRecPtr endptr,
3887 : TimeLineID insertTLI)
3888 : {
3889 : DIR *xldir;
3890 : struct dirent *xlde;
3891 : char lastoff[MAXFNAMELEN];
3892 : XLogSegNo endlogSegNo;
3893 : XLogSegNo recycleSegNo;
3894 :
3895 : /* Initialize info about where to try to recycle to */
3896 1835 : XLByteToSeg(endptr, endlogSegNo, wal_segment_size);
3897 1835 : recycleSegNo = XLOGfileslop(lastredoptr);
3898 :
3899 : /*
3900 : * Construct a filename of the last segment to be kept. The timeline ID
3901 : * doesn't matter, we ignore that in the comparison. (During recovery,
3902 : * InsertTimeLineID isn't set, so we can't use that.)
3903 : */
3904 1835 : XLogFileName(lastoff, 0, segno, wal_segment_size);
3905 :
3906 1835 : elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
3907 : lastoff);
3908 :
3909 1835 : xldir = AllocateDir(XLOGDIR);
3910 :
3911 20799 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3912 : {
3913 : /* Ignore files that are not XLOG segments */
3914 18964 : if (!IsXLogFileName(xlde->d_name) &&
3915 7794 : !IsPartialXLogFileName(xlde->d_name))
3916 7792 : continue;
3917 :
3918 : /*
3919 : * We ignore the timeline part of the XLOG segment identifiers in
3920 : * deciding whether a segment is still needed. This ensures that we
3921 : * won't prematurely remove a segment from a parent timeline. We could
3922 : * probably be a little more proactive about removing segments of
3923 : * non-parent timelines, but that would be a whole lot more
3924 : * complicated.
3925 : *
3926 : * We use the alphanumeric sorting property of the filenames to decide
3927 : * which ones are earlier than the lastoff segment.
3928 : */
3929 11172 : if (strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
3930 : {
3931 4536 : if (XLogArchiveCheckDone(xlde->d_name))
3932 : {
3933 : /* Update the last removed location in shared memory first */
3934 2708 : UpdateLastRemovedPtr(xlde->d_name);
3935 :
3936 2708 : RemoveXlogFile(xlde, recycleSegNo, &endlogSegNo, insertTLI);
3937 : }
3938 : }
3939 : }
3940 :
3941 1835 : FreeDir(xldir);
3942 1835 : }
3943 :
3944 : /*
3945 : * Recycle or remove WAL files that are not part of the given timeline's
3946 : * history.
3947 : *
3948 : * This is called during recovery, whenever we switch to follow a new
3949 : * timeline, and at the end of recovery when we create a new timeline. We
3950 : * wouldn't otherwise care about extra WAL files lying in pg_wal, but they
3951 : * might be leftover pre-allocated or recycled WAL segments on the old timeline
3952 : * that we haven't used yet, and contain garbage. If we just leave them in
3953 : * pg_wal, they will eventually be archived, and we can't let that happen.
3954 : * Files that belong to our timeline history are valid, because we have
3955 : * successfully replayed them, but from others we can't be sure.
3956 : *
3957 : * 'switchpoint' is the current point in WAL where we switch to new timeline,
3958 : * and 'newTLI' is the new timeline we switch to.
3959 : */
3960 : void
3961 66 : RemoveNonParentXlogFiles(XLogRecPtr switchpoint, TimeLineID newTLI)
3962 : {
3963 : DIR *xldir;
3964 : struct dirent *xlde;
3965 : char switchseg[MAXFNAMELEN];
3966 : XLogSegNo endLogSegNo;
3967 : XLogSegNo switchLogSegNo;
3968 : XLogSegNo recycleSegNo;
3969 :
3970 : /*
3971 : * Initialize info about where to begin the work. This will recycle,
3972 : * somewhat arbitrarily, 10 future segments.
3973 : */
3974 66 : XLByteToPrevSeg(switchpoint, switchLogSegNo, wal_segment_size);
3975 66 : XLByteToSeg(switchpoint, endLogSegNo, wal_segment_size);
3976 66 : recycleSegNo = endLogSegNo + 10;
3977 :
3978 : /*
3979 : * Construct a filename of the last segment to be kept.
3980 : */
3981 66 : XLogFileName(switchseg, newTLI, switchLogSegNo, wal_segment_size);
3982 :
3983 66 : elog(DEBUG2, "attempting to remove WAL segments newer than log file %s",
3984 : switchseg);
3985 :
3986 66 : xldir = AllocateDir(XLOGDIR);
3987 :
3988 625 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3989 : {
3990 : /* Ignore files that are not XLOG segments */
3991 559 : if (!IsXLogFileName(xlde->d_name))
3992 348 : continue;
3993 :
3994 : /*
3995 : * Remove files that are on a timeline older than the new one we're
3996 : * switching to, but with a segment number >= the first segment on the
3997 : * new timeline.
3998 : */
3999 211 : if (strncmp(xlde->d_name, switchseg, 8) < 0 &&
4000 137 : strcmp(xlde->d_name + 8, switchseg + 8) > 0)
4001 : {
4002 : /*
4003 : * If the file has already been marked as .ready, however, don't
4004 : * remove it yet. It should be OK to remove it - files that are
4005 : * not part of our timeline history are not required for recovery
4006 : * - but seems safer to let them be archived and removed later.
4007 : */
4008 16 : if (!XLogArchiveIsReady(xlde->d_name))
4009 16 : RemoveXlogFile(xlde, recycleSegNo, &endLogSegNo, newTLI);
4010 : }
4011 : }
4012 :
4013 66 : FreeDir(xldir);
4014 66 : }
4015 :
4016 : /*
4017 : * Recycle or remove a log file that's no longer needed.
4018 : *
4019 : * segment_de is the dirent structure of the segment to recycle or remove.
4020 : * recycleSegNo is the segment number to recycle up to. endlogSegNo is
4021 : * the segment number of the current (or recent) end of WAL.
4022 : *
4023 : * endlogSegNo gets incremented if the segment is recycled so as it is not
4024 : * checked again with future callers of this function.
4025 : *
4026 : * insertTLI is the current timeline for XLOG insertion. Any recycled segments
4027 : * should be used for this timeline.
4028 : */
4029 : static void
4030 2724 : RemoveXlogFile(const struct dirent *segment_de,
4031 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
4032 : TimeLineID insertTLI)
4033 : {
4034 : char path[MAXPGPATH];
4035 : #ifdef WIN32
4036 : char newpath[MAXPGPATH];
4037 : #endif
4038 2724 : const char *segname = segment_de->d_name;
4039 :
4040 2724 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", segname);
4041 :
4042 : /*
4043 : * Before deleting the file, see if it can be recycled as a future log
4044 : * segment. Only recycle normal files, because we don't want to recycle
4045 : * symbolic links pointing to a separate archive directory.
4046 : */
4047 2724 : if (wal_recycle &&
4048 2724 : *endlogSegNo <= recycleSegNo &&
4049 3856 : XLogCtl->InstallXLogFileSegmentActive && /* callee rechecks this */
4050 3528 : get_dirent_type(path, segment_de, false, DEBUG2) == PGFILETYPE_REG &&
4051 1764 : InstallXLogFileSegment(endlogSegNo, path,
4052 : true, recycleSegNo, insertTLI))
4053 : {
4054 1545 : ereport(DEBUG2,
4055 : (errmsg_internal("recycled write-ahead log file \"%s\"",
4056 : segname)));
4057 1545 : CheckpointStats.ckpt_segs_recycled++;
4058 : /* Needn't recheck that slot on future iterations */
4059 1545 : (*endlogSegNo)++;
4060 : }
4061 : else
4062 : {
4063 : /* No need for any more future segments, or recycling failed ... */
4064 : int rc;
4065 :
4066 1179 : ereport(DEBUG2,
4067 : (errmsg_internal("removing write-ahead log file \"%s\"",
4068 : segname)));
4069 :
4070 : #ifdef WIN32
4071 :
4072 : /*
4073 : * On Windows, if another process (e.g another backend) holds the file
4074 : * open in FILE_SHARE_DELETE mode, unlink will succeed, but the file
4075 : * will still show up in directory listing until the last handle is
4076 : * closed. To avoid confusing the lingering deleted file for a live
4077 : * WAL file that needs to be archived, rename it before deleting it.
4078 : *
4079 : * If another process holds the file open without FILE_SHARE_DELETE
4080 : * flag, rename will fail. We'll try again at the next checkpoint.
4081 : */
4082 : snprintf(newpath, MAXPGPATH, "%s.deleted", path);
4083 : if (rename(path, newpath) != 0)
4084 : {
4085 : ereport(LOG,
4086 : (errcode_for_file_access(),
4087 : errmsg("could not rename file \"%s\": %m",
4088 : path)));
4089 : return;
4090 : }
4091 : rc = durable_unlink(newpath, LOG);
4092 : #else
4093 1179 : rc = durable_unlink(path, LOG);
4094 : #endif
4095 1179 : if (rc != 0)
4096 : {
4097 : /* Message already logged by durable_unlink() */
4098 0 : return;
4099 : }
4100 1179 : CheckpointStats.ckpt_segs_removed++;
4101 : }
4102 :
4103 2724 : XLogArchiveCleanup(segname);
4104 : }
4105 :
4106 : /*
4107 : * Verify whether pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
4108 : * If the latter do not exist, recreate them.
4109 : *
4110 : * It is not the goal of this function to verify the contents of these
4111 : * directories, but to help in cases where someone has performed a cluster
4112 : * copy for PITR purposes but omitted pg_wal from the copy.
4113 : *
4114 : * We could also recreate pg_wal if it doesn't exist, but a deliberate
4115 : * policy decision was made not to. It is fairly common for pg_wal to be
4116 : * a symlink, and if that was the DBA's intent then automatically making a
4117 : * plain directory would result in degraded performance with no notice.
4118 : */
4119 : static void
4120 1035 : ValidateXLOGDirectoryStructure(void)
4121 : {
4122 : char path[MAXPGPATH];
4123 : struct stat stat_buf;
4124 :
4125 : /* Check for pg_wal; if it doesn't exist, error out */
4126 1035 : if (stat(XLOGDIR, &stat_buf) != 0 ||
4127 1035 : !S_ISDIR(stat_buf.st_mode))
4128 0 : ereport(FATAL,
4129 : (errcode_for_file_access(),
4130 : errmsg("required WAL directory \"%s\" does not exist",
4131 : XLOGDIR)));
4132 :
4133 : /* Check for archive_status */
4134 1035 : snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4135 1035 : if (stat(path, &stat_buf) == 0)
4136 : {
4137 : /* Check for weird cases where it exists but isn't a directory */
4138 1034 : if (!S_ISDIR(stat_buf.st_mode))
4139 0 : ereport(FATAL,
4140 : (errcode_for_file_access(),
4141 : errmsg("required WAL directory \"%s\" does not exist",
4142 : path)));
4143 : }
4144 : else
4145 : {
4146 1 : ereport(LOG,
4147 : (errmsg("creating missing WAL directory \"%s\"", path)));
4148 1 : if (MakePGDirectory(path) < 0)
4149 0 : ereport(FATAL,
4150 : (errcode_for_file_access(),
4151 : errmsg("could not create missing directory \"%s\": %m",
4152 : path)));
4153 : }
4154 :
4155 : /* Check for summaries */
4156 1035 : snprintf(path, MAXPGPATH, XLOGDIR "/summaries");
4157 1035 : if (stat(path, &stat_buf) == 0)
4158 : {
4159 : /* Check for weird cases where it exists but isn't a directory */
4160 1034 : if (!S_ISDIR(stat_buf.st_mode))
4161 0 : ereport(FATAL,
4162 : (errmsg("required WAL directory \"%s\" does not exist",
4163 : path)));
4164 : }
4165 : else
4166 : {
4167 1 : ereport(LOG,
4168 : (errmsg("creating missing WAL directory \"%s\"", path)));
4169 1 : if (MakePGDirectory(path) < 0)
4170 0 : ereport(FATAL,
4171 : (errmsg("could not create missing directory \"%s\": %m",
4172 : path)));
4173 : }
4174 1035 : }
4175 :
4176 : /*
4177 : * Remove previous backup history files. This also retries creation of
4178 : * .ready files for any backup history files for which XLogArchiveNotify
4179 : * failed earlier.
4180 : */
4181 : static void
4182 161 : CleanupBackupHistory(void)
4183 : {
4184 : DIR *xldir;
4185 : struct dirent *xlde;
4186 : char path[MAXPGPATH + sizeof(XLOGDIR)];
4187 :
4188 161 : xldir = AllocateDir(XLOGDIR);
4189 :
4190 1664 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4191 : {
4192 1342 : if (IsBackupHistoryFileName(xlde->d_name))
4193 : {
4194 171 : if (XLogArchiveCheckDone(xlde->d_name))
4195 : {
4196 135 : elog(DEBUG2, "removing WAL backup history file \"%s\"",
4197 : xlde->d_name);
4198 135 : snprintf(path, sizeof(path), XLOGDIR "/%s", xlde->d_name);
4199 135 : unlink(path);
4200 135 : XLogArchiveCleanup(xlde->d_name);
4201 : }
4202 : }
4203 : }
4204 :
4205 161 : FreeDir(xldir);
4206 161 : }
4207 :
4208 : /*
4209 : * I/O routines for pg_control
4210 : *
4211 : * *ControlFile is a buffer in shared memory that holds an image of the
4212 : * contents of pg_control. WriteControlFile() initializes pg_control
4213 : * given a preloaded buffer, ReadControlFile() loads the buffer from
4214 : * the pg_control file (during postmaster or standalone-backend startup),
4215 : * and UpdateControlFile() rewrites pg_control after we modify xlog state.
4216 : * InitControlFile() fills the buffer with initial values.
4217 : *
4218 : * For simplicity, WriteControlFile() initializes the fields of pg_control
4219 : * that are related to checking backend/database compatibility, and
4220 : * ReadControlFile() verifies they are correct. We could split out the
4221 : * I/O and compatibility-check functions, but there seems no need currently.
4222 : */
4223 :
4224 : static void
4225 51 : InitControlFile(uint64 sysidentifier, uint32 data_checksum_version)
4226 : {
4227 : char mock_auth_nonce[MOCK_AUTH_NONCE_LEN];
4228 :
4229 : /*
4230 : * Generate a random nonce. This is used for authentication requests that
4231 : * will fail because the user does not exist. The nonce is used to create
4232 : * a genuine-looking password challenge for the non-existent user, in lieu
4233 : * of an actual stored password.
4234 : */
4235 51 : if (!pg_strong_random(mock_auth_nonce, MOCK_AUTH_NONCE_LEN))
4236 0 : ereport(PANIC,
4237 : (errcode(ERRCODE_INTERNAL_ERROR),
4238 : errmsg("could not generate secret authorization token")));
4239 :
4240 51 : memset(ControlFile, 0, sizeof(ControlFileData));
4241 : /* Initialize pg_control status fields */
4242 51 : ControlFile->system_identifier = sysidentifier;
4243 51 : memcpy(ControlFile->mock_authentication_nonce, mock_auth_nonce, MOCK_AUTH_NONCE_LEN);
4244 51 : ControlFile->state = DB_SHUTDOWNED;
4245 51 : ControlFile->unloggedLSN = FirstNormalUnloggedLSN;
4246 :
4247 : /* Set important parameter values for use when replaying WAL */
4248 51 : ControlFile->MaxConnections = MaxConnections;
4249 51 : ControlFile->max_worker_processes = max_worker_processes;
4250 51 : ControlFile->max_wal_senders = max_wal_senders;
4251 51 : ControlFile->max_prepared_xacts = max_prepared_xacts;
4252 51 : ControlFile->max_locks_per_xact = max_locks_per_xact;
4253 51 : ControlFile->wal_level = wal_level;
4254 51 : ControlFile->wal_log_hints = wal_log_hints;
4255 51 : ControlFile->track_commit_timestamp = track_commit_timestamp;
4256 51 : ControlFile->data_checksum_version = data_checksum_version;
4257 51 : }
4258 :
4259 : static void
4260 51 : WriteControlFile(void)
4261 : {
4262 : int fd;
4263 : char buffer[PG_CONTROL_FILE_SIZE]; /* need not be aligned */
4264 :
4265 : /*
4266 : * Initialize version and compatibility-check fields
4267 : */
4268 51 : ControlFile->pg_control_version = PG_CONTROL_VERSION;
4269 51 : ControlFile->catalog_version_no = CATALOG_VERSION_NO;
4270 :
4271 51 : ControlFile->maxAlign = MAXIMUM_ALIGNOF;
4272 51 : ControlFile->floatFormat = FLOATFORMAT_VALUE;
4273 :
4274 51 : ControlFile->blcksz = BLCKSZ;
4275 51 : ControlFile->relseg_size = RELSEG_SIZE;
4276 51 : ControlFile->slru_pages_per_segment = SLRU_PAGES_PER_SEGMENT;
4277 51 : ControlFile->xlog_blcksz = XLOG_BLCKSZ;
4278 51 : ControlFile->xlog_seg_size = wal_segment_size;
4279 :
4280 51 : ControlFile->nameDataLen = NAMEDATALEN;
4281 51 : ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
4282 :
4283 51 : ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
4284 51 : ControlFile->loblksize = LOBLKSIZE;
4285 :
4286 51 : ControlFile->float8ByVal = true; /* vestigial */
4287 :
4288 : /*
4289 : * Initialize the default 'char' signedness.
4290 : *
4291 : * The signedness of the char type is implementation-defined. For instance
4292 : * on x86 architecture CPUs, the char data type is typically treated as
4293 : * signed by default, whereas on aarch architecture CPUs, it is typically
4294 : * treated as unsigned by default. In v17 or earlier, we accidentally let
4295 : * C implementation signedness affect persistent data. This led to
4296 : * inconsistent results when comparing char data across different
4297 : * platforms.
4298 : *
4299 : * This flag can be used as a hint to ensure consistent behavior for
4300 : * pre-v18 data files that store data sorted by the 'char' type on disk,
4301 : * especially in cross-platform replication scenarios.
4302 : *
4303 : * Newly created database clusters unconditionally set the default char
4304 : * signedness to true. pg_upgrade changes this flag for clusters that were
4305 : * initialized on signedness=false platforms. As a result,
4306 : * signedness=false setting will become rare over time. If we had known
4307 : * about this problem during the last development cycle that forced initdb
4308 : * (v8.3), we would have made all clusters signed or all clusters
4309 : * unsigned. Making pg_upgrade the only source of signedness=false will
4310 : * cause the population of database clusters to converge toward that
4311 : * retrospective ideal.
4312 : */
4313 51 : ControlFile->default_char_signedness = true;
4314 :
4315 : /* Contents are protected with a CRC */
4316 51 : INIT_CRC32C(ControlFile->crc);
4317 51 : COMP_CRC32C(ControlFile->crc,
4318 : ControlFile,
4319 : offsetof(ControlFileData, crc));
4320 51 : FIN_CRC32C(ControlFile->crc);
4321 :
4322 : /*
4323 : * We write out PG_CONTROL_FILE_SIZE bytes into pg_control, zero-padding
4324 : * the excess over sizeof(ControlFileData). This reduces the odds of
4325 : * premature-EOF errors when reading pg_control. We'll still fail when we
4326 : * check the contents of the file, but hopefully with a more specific
4327 : * error than "couldn't read pg_control".
4328 : */
4329 51 : memset(buffer, 0, PG_CONTROL_FILE_SIZE);
4330 51 : memcpy(buffer, ControlFile, sizeof(ControlFileData));
4331 :
4332 51 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4333 : O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
4334 51 : if (fd < 0)
4335 0 : ereport(PANIC,
4336 : (errcode_for_file_access(),
4337 : errmsg("could not create file \"%s\": %m",
4338 : XLOG_CONTROL_FILE)));
4339 :
4340 51 : errno = 0;
4341 51 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_WRITE);
4342 51 : if (write(fd, buffer, PG_CONTROL_FILE_SIZE) != PG_CONTROL_FILE_SIZE)
4343 : {
4344 : /* if write didn't set errno, assume problem is no disk space */
4345 0 : if (errno == 0)
4346 0 : errno = ENOSPC;
4347 0 : ereport(PANIC,
4348 : (errcode_for_file_access(),
4349 : errmsg("could not write to file \"%s\": %m",
4350 : XLOG_CONTROL_FILE)));
4351 : }
4352 51 : pgstat_report_wait_end();
4353 :
4354 51 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_SYNC);
4355 51 : if (pg_fsync(fd) != 0)
4356 0 : ereport(PANIC,
4357 : (errcode_for_file_access(),
4358 : errmsg("could not fsync file \"%s\": %m",
4359 : XLOG_CONTROL_FILE)));
4360 51 : pgstat_report_wait_end();
4361 :
4362 51 : if (close(fd) != 0)
4363 0 : ereport(PANIC,
4364 : (errcode_for_file_access(),
4365 : errmsg("could not close file \"%s\": %m",
4366 : XLOG_CONTROL_FILE)));
4367 51 : }
4368 :
4369 : static void
4370 1094 : ReadControlFile(void)
4371 : {
4372 : pg_crc32c crc;
4373 : int fd;
4374 : char wal_segsz_str[20];
4375 : int r;
4376 :
4377 : /*
4378 : * Read data...
4379 : */
4380 1094 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4381 : O_RDWR | PG_BINARY);
4382 1094 : if (fd < 0)
4383 0 : ereport(PANIC,
4384 : (errcode_for_file_access(),
4385 : errmsg("could not open file \"%s\": %m",
4386 : XLOG_CONTROL_FILE)));
4387 :
4388 1094 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ);
4389 1094 : r = read(fd, ControlFile, sizeof(ControlFileData));
4390 1094 : if (r != sizeof(ControlFileData))
4391 : {
4392 0 : if (r < 0)
4393 0 : ereport(PANIC,
4394 : (errcode_for_file_access(),
4395 : errmsg("could not read file \"%s\": %m",
4396 : XLOG_CONTROL_FILE)));
4397 : else
4398 0 : ereport(PANIC,
4399 : (errcode(ERRCODE_DATA_CORRUPTED),
4400 : errmsg("could not read file \"%s\": read %d of %zu",
4401 : XLOG_CONTROL_FILE, r, sizeof(ControlFileData))));
4402 : }
4403 1094 : pgstat_report_wait_end();
4404 :
4405 1094 : close(fd);
4406 :
4407 : /*
4408 : * Check for expected pg_control format version. If this is wrong, the
4409 : * CRC check will likely fail because we'll be checking the wrong number
4410 : * of bytes. Complaining about wrong version will probably be more
4411 : * enlightening than complaining about wrong CRC.
4412 : */
4413 :
4414 1094 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
4415 0 : ereport(FATAL,
4416 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4417 : errmsg("database files are incompatible with server"),
4418 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4419 : " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4420 : ControlFile->pg_control_version, ControlFile->pg_control_version,
4421 : PG_CONTROL_VERSION, PG_CONTROL_VERSION),
4422 : errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4423 :
4424 1094 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
4425 0 : ereport(FATAL,
4426 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4427 : errmsg("database files are incompatible with server"),
4428 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4429 : " but the server was compiled with PG_CONTROL_VERSION %d.",
4430 : ControlFile->pg_control_version, PG_CONTROL_VERSION),
4431 : errhint("It looks like you need to initdb.")));
4432 :
4433 : /* Now check the CRC. */
4434 1094 : INIT_CRC32C(crc);
4435 1094 : COMP_CRC32C(crc,
4436 : ControlFile,
4437 : offsetof(ControlFileData, crc));
4438 1094 : FIN_CRC32C(crc);
4439 :
4440 1094 : if (!EQ_CRC32C(crc, ControlFile->crc))
4441 0 : ereport(FATAL,
4442 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4443 : errmsg("incorrect checksum in control file")));
4444 :
4445 : /*
4446 : * Do compatibility checking immediately. If the database isn't
4447 : * compatible with the backend executable, we want to abort before we can
4448 : * possibly do any damage.
4449 : */
4450 1094 : if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
4451 0 : ereport(FATAL,
4452 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4453 : errmsg("database files are incompatible with server"),
4454 : /* translator: %s is a variable name and %d is its value */
4455 : errdetail("The database cluster was initialized with %s %d,"
4456 : " but the server was compiled with %s %d.",
4457 : "CATALOG_VERSION_NO", ControlFile->catalog_version_no,
4458 : "CATALOG_VERSION_NO", CATALOG_VERSION_NO),
4459 : errhint("It looks like you need to initdb.")));
4460 1094 : if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
4461 0 : ereport(FATAL,
4462 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4463 : errmsg("database files are incompatible with server"),
4464 : /* translator: %s is a variable name and %d is its value */
4465 : errdetail("The database cluster was initialized with %s %d,"
4466 : " but the server was compiled with %s %d.",
4467 : "MAXALIGN", ControlFile->maxAlign,
4468 : "MAXALIGN", MAXIMUM_ALIGNOF),
4469 : errhint("It looks like you need to initdb.")));
4470 1094 : if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
4471 0 : ereport(FATAL,
4472 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4473 : errmsg("database files are incompatible with server"),
4474 : errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4475 : errhint("It looks like you need to initdb.")));
4476 1094 : if (ControlFile->blcksz != BLCKSZ)
4477 0 : ereport(FATAL,
4478 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4479 : errmsg("database files are incompatible with server"),
4480 : /* translator: %s is a variable name and %d is its value */
4481 : errdetail("The database cluster was initialized with %s %d,"
4482 : " but the server was compiled with %s %d.",
4483 : "BLCKSZ", ControlFile->blcksz,
4484 : "BLCKSZ", BLCKSZ),
4485 : errhint("It looks like you need to recompile or initdb.")));
4486 1094 : if (ControlFile->relseg_size != RELSEG_SIZE)
4487 0 : ereport(FATAL,
4488 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4489 : errmsg("database files are incompatible with server"),
4490 : /* translator: %s is a variable name and %d is its value */
4491 : errdetail("The database cluster was initialized with %s %d,"
4492 : " but the server was compiled with %s %d.",
4493 : "RELSEG_SIZE", ControlFile->relseg_size,
4494 : "RELSEG_SIZE", RELSEG_SIZE),
4495 : errhint("It looks like you need to recompile or initdb.")));
4496 1094 : if (ControlFile->slru_pages_per_segment != SLRU_PAGES_PER_SEGMENT)
4497 0 : ereport(FATAL,
4498 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4499 : errmsg("database files are incompatible with server"),
4500 : /* translator: %s is a variable name and %d is its value */
4501 : errdetail("The database cluster was initialized with %s %d,"
4502 : " but the server was compiled with %s %d.",
4503 : "SLRU_PAGES_PER_SEGMENT", ControlFile->slru_pages_per_segment,
4504 : "SLRU_PAGES_PER_SEGMENT", SLRU_PAGES_PER_SEGMENT),
4505 : errhint("It looks like you need to recompile or initdb.")));
4506 1094 : if (ControlFile->xlog_blcksz != XLOG_BLCKSZ)
4507 0 : ereport(FATAL,
4508 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4509 : errmsg("database files are incompatible with server"),
4510 : /* translator: %s is a variable name and %d is its value */
4511 : errdetail("The database cluster was initialized with %s %d,"
4512 : " but the server was compiled with %s %d.",
4513 : "XLOG_BLCKSZ", ControlFile->xlog_blcksz,
4514 : "XLOG_BLCKSZ", XLOG_BLCKSZ),
4515 : errhint("It looks like you need to recompile or initdb.")));
4516 1094 : if (ControlFile->nameDataLen != NAMEDATALEN)
4517 0 : ereport(FATAL,
4518 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4519 : errmsg("database files are incompatible with server"),
4520 : /* translator: %s is a variable name and %d is its value */
4521 : errdetail("The database cluster was initialized with %s %d,"
4522 : " but the server was compiled with %s %d.",
4523 : "NAMEDATALEN", ControlFile->nameDataLen,
4524 : "NAMEDATALEN", NAMEDATALEN),
4525 : errhint("It looks like you need to recompile or initdb.")));
4526 1094 : if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
4527 0 : ereport(FATAL,
4528 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4529 : errmsg("database files are incompatible with server"),
4530 : /* translator: %s is a variable name and %d is its value */
4531 : errdetail("The database cluster was initialized with %s %d,"
4532 : " but the server was compiled with %s %d.",
4533 : "INDEX_MAX_KEYS", ControlFile->indexMaxKeys,
4534 : "INDEX_MAX_KEYS", INDEX_MAX_KEYS),
4535 : errhint("It looks like you need to recompile or initdb.")));
4536 1094 : if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
4537 0 : ereport(FATAL,
4538 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4539 : errmsg("database files are incompatible with server"),
4540 : /* translator: %s is a variable name and %d is its value */
4541 : errdetail("The database cluster was initialized with %s %d,"
4542 : " but the server was compiled with %s %d.",
4543 : "TOAST_MAX_CHUNK_SIZE", ControlFile->toast_max_chunk_size,
4544 : "TOAST_MAX_CHUNK_SIZE", (int) TOAST_MAX_CHUNK_SIZE),
4545 : errhint("It looks like you need to recompile or initdb.")));
4546 1094 : if (ControlFile->loblksize != LOBLKSIZE)
4547 0 : ereport(FATAL,
4548 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4549 : errmsg("database files are incompatible with server"),
4550 : /* translator: %s is a variable name and %d is its value */
4551 : errdetail("The database cluster was initialized with %s %d,"
4552 : " but the server was compiled with %s %d.",
4553 : "LOBLKSIZE", ControlFile->loblksize,
4554 : "LOBLKSIZE", (int) LOBLKSIZE),
4555 : errhint("It looks like you need to recompile or initdb.")));
4556 :
4557 : Assert(ControlFile->float8ByVal); /* vestigial, not worth an error msg */
4558 :
4559 1094 : wal_segment_size = ControlFile->xlog_seg_size;
4560 :
4561 1094 : if (!IsValidWalSegSize(wal_segment_size))
4562 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4563 : errmsg_plural("invalid WAL segment size in control file (%d byte)",
4564 : "invalid WAL segment size in control file (%d bytes)",
4565 : wal_segment_size,
4566 : wal_segment_size),
4567 : errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.")));
4568 :
4569 1094 : snprintf(wal_segsz_str, sizeof(wal_segsz_str), "%d", wal_segment_size);
4570 1094 : SetConfigOption("wal_segment_size", wal_segsz_str, PGC_INTERNAL,
4571 : PGC_S_DYNAMIC_DEFAULT);
4572 :
4573 : /* check and update variables dependent on wal_segment_size */
4574 1094 : if (ConvertToXSegs(min_wal_size_mb, wal_segment_size) < 2)
4575 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4576 : /* translator: both %s are GUC names */
4577 : errmsg("\"%s\" must be at least twice \"%s\"",
4578 : "min_wal_size", "wal_segment_size")));
4579 :
4580 1094 : if (ConvertToXSegs(max_wal_size_mb, wal_segment_size) < 2)
4581 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4582 : /* translator: both %s are GUC names */
4583 : errmsg("\"%s\" must be at least twice \"%s\"",
4584 : "max_wal_size", "wal_segment_size")));
4585 :
4586 1094 : UsableBytesInSegment =
4587 1094 : (wal_segment_size / XLOG_BLCKSZ * UsableBytesInPage) -
4588 : (SizeOfXLogLongPHD - SizeOfXLogShortPHD);
4589 :
4590 1094 : CalculateCheckpointSegments();
4591 :
4592 : /* Make the initdb settings visible as GUC variables, too */
4593 1094 : SetConfigOption("data_checksums", DataChecksumsEnabled() ? "yes" : "no",
4594 : PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
4595 1094 : }
4596 :
4597 : /*
4598 : * Utility wrapper to update the control file. Note that the control
4599 : * file gets flushed.
4600 : */
4601 : static void
4602 9789 : UpdateControlFile(void)
4603 : {
4604 9789 : update_controlfile(DataDir, ControlFile, true);
4605 9789 : }
4606 :
4607 : /*
4608 : * Returns the unique system identifier from control file.
4609 : */
4610 : uint64
4611 1542 : GetSystemIdentifier(void)
4612 : {
4613 : Assert(ControlFile != NULL);
4614 1542 : return ControlFile->system_identifier;
4615 : }
4616 :
4617 : /*
4618 : * Returns the random nonce from control file.
4619 : */
4620 : char *
4621 2 : GetMockAuthenticationNonce(void)
4622 : {
4623 : Assert(ControlFile != NULL);
4624 2 : return ControlFile->mock_authentication_nonce;
4625 : }
4626 :
4627 : /*
4628 : * Are checksums enabled for data pages?
4629 : */
4630 : bool
4631 2960466 : DataChecksumsEnabled(void)
4632 : {
4633 : Assert(ControlFile != NULL);
4634 2960466 : return (ControlFile->data_checksum_version > 0);
4635 : }
4636 :
4637 : /*
4638 : * Return true if the cluster was initialized on a platform where the
4639 : * default signedness of char is "signed". This function exists for code
4640 : * that deals with pre-v18 data files that store data sorted by the 'char'
4641 : * type on disk (e.g., GIN and GiST indexes). See the comments in
4642 : * WriteControlFile() for details.
4643 : */
4644 : bool
4645 4 : GetDefaultCharSignedness(void)
4646 : {
4647 4 : return ControlFile->default_char_signedness;
4648 : }
4649 :
4650 : /*
4651 : * Returns a fake LSN for unlogged relations.
4652 : *
4653 : * Each call generates an LSN that is greater than any previous value
4654 : * returned. The current counter value is saved and restored across clean
4655 : * shutdowns, but like unlogged relations, does not survive a crash. This can
4656 : * be used in lieu of real LSN values returned by XLogInsert, if you need an
4657 : * LSN-like increasing sequence of numbers without writing any WAL.
4658 : */
4659 : XLogRecPtr
4660 201460 : GetFakeLSNForUnloggedRel(void)
4661 : {
4662 201460 : return pg_atomic_fetch_add_u64(&XLogCtl->unloggedLSN, 1);
4663 : }
4664 :
4665 : /*
4666 : * Auto-tune the number of XLOG buffers.
4667 : *
4668 : * The preferred setting for wal_buffers is about 3% of shared_buffers, with
4669 : * a maximum of one XLOG segment (there is little reason to think that more
4670 : * is helpful, at least so long as we force an fsync when switching log files)
4671 : * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
4672 : * 9.1, when auto-tuning was added).
4673 : *
4674 : * This should not be called until NBuffers has received its final value.
4675 : */
4676 : static int
4677 1178 : XLOGChooseNumBuffers(void)
4678 : {
4679 : int xbuffers;
4680 :
4681 1178 : xbuffers = NBuffers / 32;
4682 1178 : if (xbuffers > (wal_segment_size / XLOG_BLCKSZ))
4683 24 : xbuffers = (wal_segment_size / XLOG_BLCKSZ);
4684 1178 : if (xbuffers < 8)
4685 448 : xbuffers = 8;
4686 1178 : return xbuffers;
4687 : }
4688 :
4689 : /*
4690 : * GUC check_hook for wal_buffers
4691 : */
4692 : bool
4693 2400 : check_wal_buffers(int *newval, void **extra, GucSource source)
4694 : {
4695 : /*
4696 : * -1 indicates a request for auto-tune.
4697 : */
4698 2400 : if (*newval == -1)
4699 : {
4700 : /*
4701 : * If we haven't yet changed the boot_val default of -1, just let it
4702 : * be. We'll fix it when XLOGShmemSize is called.
4703 : */
4704 1222 : if (XLOGbuffers == -1)
4705 1222 : return true;
4706 :
4707 : /* Otherwise, substitute the auto-tune value */
4708 0 : *newval = XLOGChooseNumBuffers();
4709 : }
4710 :
4711 : /*
4712 : * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
4713 : * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
4714 : * the case, we just silently treat such values as a request for the
4715 : * minimum. (We could throw an error instead, but that doesn't seem very
4716 : * helpful.)
4717 : */
4718 1178 : if (*newval < 4)
4719 0 : *newval = 4;
4720 :
4721 1178 : return true;
4722 : }
4723 :
4724 : /*
4725 : * GUC check_hook for wal_consistency_checking
4726 : */
4727 : bool
4728 2204 : check_wal_consistency_checking(char **newval, void **extra, GucSource source)
4729 : {
4730 : char *rawstring;
4731 : List *elemlist;
4732 : ListCell *l;
4733 : bool newwalconsistency[RM_MAX_ID + 1];
4734 :
4735 : /* Initialize the array */
4736 72732 : MemSet(newwalconsistency, 0, (RM_MAX_ID + 1) * sizeof(bool));
4737 :
4738 : /* Need a modifiable copy of string */
4739 2204 : rawstring = pstrdup(*newval);
4740 :
4741 : /* Parse string into list of identifiers */
4742 2204 : if (!SplitIdentifierString(rawstring, ',', &elemlist))
4743 : {
4744 : /* syntax error in list */
4745 0 : GUC_check_errdetail("List syntax is invalid.");
4746 0 : pfree(rawstring);
4747 0 : list_free(elemlist);
4748 0 : return false;
4749 : }
4750 :
4751 2698 : foreach(l, elemlist)
4752 : {
4753 494 : char *tok = (char *) lfirst(l);
4754 : int rmid;
4755 :
4756 : /* Check for 'all'. */
4757 494 : if (pg_strcasecmp(tok, "all") == 0)
4758 : {
4759 126444 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4760 125952 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL)
4761 4920 : newwalconsistency[rmid] = true;
4762 : }
4763 : else
4764 : {
4765 : /* Check if the token matches any known resource manager. */
4766 2 : bool found = false;
4767 :
4768 36 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4769 : {
4770 54 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL &&
4771 18 : pg_strcasecmp(tok, GetRmgr(rmid).rm_name) == 0)
4772 : {
4773 2 : newwalconsistency[rmid] = true;
4774 2 : found = true;
4775 2 : break;
4776 : }
4777 : }
4778 2 : if (!found)
4779 : {
4780 : /*
4781 : * During startup, it might be a not-yet-loaded custom
4782 : * resource manager. Defer checking until
4783 : * InitializeWalConsistencyChecking().
4784 : */
4785 0 : if (!process_shared_preload_libraries_done)
4786 : {
4787 0 : check_wal_consistency_checking_deferred = true;
4788 : }
4789 : else
4790 : {
4791 0 : GUC_check_errdetail("Unrecognized key word: \"%s\".", tok);
4792 0 : pfree(rawstring);
4793 0 : list_free(elemlist);
4794 0 : return false;
4795 : }
4796 : }
4797 : }
4798 : }
4799 :
4800 2204 : pfree(rawstring);
4801 2204 : list_free(elemlist);
4802 :
4803 : /* assign new value */
4804 2204 : *extra = guc_malloc(LOG, (RM_MAX_ID + 1) * sizeof(bool));
4805 2204 : if (!*extra)
4806 0 : return false;
4807 2204 : memcpy(*extra, newwalconsistency, (RM_MAX_ID + 1) * sizeof(bool));
4808 2204 : return true;
4809 : }
4810 :
4811 : /*
4812 : * GUC assign_hook for wal_consistency_checking
4813 : */
4814 : void
4815 2203 : assign_wal_consistency_checking(const char *newval, void *extra)
4816 : {
4817 : /*
4818 : * If some checks were deferred, it's possible that the checks will fail
4819 : * later during InitializeWalConsistencyChecking(). But in that case, the
4820 : * postmaster will exit anyway, so it's safe to proceed with the
4821 : * assignment.
4822 : *
4823 : * Any built-in resource managers specified are assigned immediately,
4824 : * which affects WAL created before shared_preload_libraries are
4825 : * processed. Any custom resource managers specified won't be assigned
4826 : * until after shared_preload_libraries are processed, but that's OK
4827 : * because WAL for a custom resource manager can't be written before the
4828 : * module is loaded anyway.
4829 : */
4830 2203 : wal_consistency_checking = extra;
4831 2203 : }
4832 :
4833 : /*
4834 : * InitializeWalConsistencyChecking: run after loading custom resource managers
4835 : *
4836 : * If any unknown resource managers were specified in the
4837 : * wal_consistency_checking GUC, processing was deferred. Now that
4838 : * shared_preload_libraries have been loaded, process wal_consistency_checking
4839 : * again.
4840 : */
4841 : void
4842 1025 : InitializeWalConsistencyChecking(void)
4843 : {
4844 : Assert(process_shared_preload_libraries_done);
4845 :
4846 1025 : if (check_wal_consistency_checking_deferred)
4847 : {
4848 : struct config_generic *guc;
4849 :
4850 0 : guc = find_option("wal_consistency_checking", false, false, ERROR);
4851 :
4852 0 : check_wal_consistency_checking_deferred = false;
4853 :
4854 0 : set_config_option_ext("wal_consistency_checking",
4855 : wal_consistency_checking_string,
4856 : guc->scontext, guc->source, guc->srole,
4857 : GUC_ACTION_SET, true, ERROR, false);
4858 :
4859 : /* checking should not be deferred again */
4860 : Assert(!check_wal_consistency_checking_deferred);
4861 : }
4862 1025 : }
4863 :
4864 : /*
4865 : * GUC show_hook for archive_command
4866 : */
4867 : const char *
4868 1900 : show_archive_command(void)
4869 : {
4870 1900 : if (XLogArchivingActive())
4871 106 : return XLogArchiveCommand;
4872 : else
4873 1794 : return "(disabled)";
4874 : }
4875 :
4876 : /*
4877 : * GUC show_hook for in_hot_standby
4878 : */
4879 : const char *
4880 16590 : show_in_hot_standby(void)
4881 : {
4882 : /*
4883 : * We display the actual state based on shared memory, so that this GUC
4884 : * reports up-to-date state if examined intra-query. The underlying
4885 : * variable (in_hot_standby_guc) changes only when we transmit a new value
4886 : * to the client.
4887 : */
4888 16590 : return RecoveryInProgress() ? "on" : "off";
4889 : }
4890 :
4891 : /*
4892 : * GUC show_hook for effective_wal_level
4893 : */
4894 : const char *
4895 1933 : show_effective_wal_level(void)
4896 : {
4897 1933 : if (wal_level == WAL_LEVEL_MINIMAL)
4898 276 : return "minimal";
4899 :
4900 : /*
4901 : * During recovery, effective_wal_level reflects the primary's
4902 : * configuration rather than the local wal_level value.
4903 : */
4904 1657 : if (RecoveryInProgress())
4905 22 : return IsXLogLogicalInfoEnabled() ? "logical" : "replica";
4906 :
4907 1635 : return XLogLogicalInfoActive() ? "logical" : "replica";
4908 : }
4909 :
4910 : /*
4911 : * Read the control file, set respective GUCs.
4912 : *
4913 : * This is to be called during startup, including a crash recovery cycle,
4914 : * unless in bootstrap mode, where no control file yet exists. As there's no
4915 : * usable shared memory yet (its sizing can depend on the contents of the
4916 : * control file!), first store the contents in local memory. XLOGShmemInit()
4917 : * will then copy it to shared memory later.
4918 : *
4919 : * reset just controls whether previous contents are to be expected (in the
4920 : * reset case, there's a dangling pointer into old shared memory), or not.
4921 : */
4922 : void
4923 1043 : LocalProcessControlFile(bool reset)
4924 : {
4925 : Assert(reset || ControlFile == NULL);
4926 1043 : ControlFile = palloc_object(ControlFileData);
4927 1043 : ReadControlFile();
4928 1043 : }
4929 :
4930 : /*
4931 : * Get the wal_level from the control file. For a standby, this value should be
4932 : * considered as its active wal_level, because it may be different from what
4933 : * was originally configured on standby.
4934 : */
4935 : WalLevel
4936 0 : GetActiveWalLevelOnStandby(void)
4937 : {
4938 0 : return ControlFile->wal_level;
4939 : }
4940 :
4941 : /*
4942 : * Initialization of shared memory for XLOG
4943 : */
4944 : Size
4945 3387 : XLOGShmemSize(void)
4946 : {
4947 : Size size;
4948 :
4949 : /*
4950 : * If the value of wal_buffers is -1, use the preferred auto-tune value.
4951 : * This isn't an amazingly clean place to do this, but we must wait till
4952 : * NBuffers has received its final value, and must do it before using the
4953 : * value of XLOGbuffers to do anything important.
4954 : *
4955 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
4956 : * However, if the DBA explicitly set wal_buffers = -1 in the config file,
4957 : * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
4958 : * the matter with PGC_S_OVERRIDE.
4959 : */
4960 3387 : if (XLOGbuffers == -1)
4961 : {
4962 : char buf[32];
4963 :
4964 1178 : snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
4965 1178 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4966 : PGC_S_DYNAMIC_DEFAULT);
4967 1178 : if (XLOGbuffers == -1) /* failed to apply it? */
4968 0 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4969 : PGC_S_OVERRIDE);
4970 : }
4971 : Assert(XLOGbuffers > 0);
4972 :
4973 : /* XLogCtl */
4974 3387 : size = sizeof(XLogCtlData);
4975 :
4976 : /* WAL insertion locks, plus alignment */
4977 3387 : size = add_size(size, mul_size(sizeof(WALInsertLockPadded), NUM_XLOGINSERT_LOCKS + 1));
4978 : /* xlblocks array */
4979 3387 : size = add_size(size, mul_size(sizeof(pg_atomic_uint64), XLOGbuffers));
4980 : /* extra alignment padding for XLOG I/O buffers */
4981 3387 : size = add_size(size, Max(XLOG_BLCKSZ, PG_IO_ALIGN_SIZE));
4982 : /* and the buffers themselves */
4983 3387 : size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
4984 :
4985 : /*
4986 : * Note: we don't count ControlFileData, it comes out of the "slop factor"
4987 : * added by CreateSharedMemoryAndSemaphores. This lets us use this
4988 : * routine again below to compute the actual allocation size.
4989 : */
4990 :
4991 3387 : return size;
4992 : }
4993 :
4994 : void
4995 1180 : XLOGShmemInit(void)
4996 : {
4997 : bool foundCFile,
4998 : foundXLog;
4999 : char *allocptr;
5000 : int i;
5001 : ControlFileData *localControlFile;
5002 :
5003 : #ifdef WAL_DEBUG
5004 :
5005 : /*
5006 : * Create a memory context for WAL debugging that's exempt from the normal
5007 : * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
5008 : * an allocation fails, but wal_debug is not for production use anyway.
5009 : */
5010 : if (walDebugCxt == NULL)
5011 : {
5012 : walDebugCxt = AllocSetContextCreate(TopMemoryContext,
5013 : "WAL Debug",
5014 : ALLOCSET_DEFAULT_SIZES);
5015 : MemoryContextAllowInCriticalSection(walDebugCxt, true);
5016 : }
5017 : #endif
5018 :
5019 :
5020 1180 : XLogCtl = (XLogCtlData *)
5021 1180 : ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog);
5022 :
5023 1180 : localControlFile = ControlFile;
5024 1180 : ControlFile = (ControlFileData *)
5025 1180 : ShmemInitStruct("Control File", sizeof(ControlFileData), &foundCFile);
5026 :
5027 1180 : if (foundCFile || foundXLog)
5028 : {
5029 : /* both should be present or neither */
5030 : Assert(foundCFile && foundXLog);
5031 :
5032 : /* Initialize local copy of WALInsertLocks */
5033 0 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks;
5034 :
5035 0 : if (localControlFile)
5036 0 : pfree(localControlFile);
5037 0 : return;
5038 : }
5039 1180 : memset(XLogCtl, 0, sizeof(XLogCtlData));
5040 :
5041 : /*
5042 : * Already have read control file locally, unless in bootstrap mode. Move
5043 : * contents into shared memory.
5044 : */
5045 1180 : if (localControlFile)
5046 : {
5047 1027 : memcpy(ControlFile, localControlFile, sizeof(ControlFileData));
5048 1027 : pfree(localControlFile);
5049 : }
5050 :
5051 : /*
5052 : * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
5053 : * multiple of the alignment for same, so no extra alignment padding is
5054 : * needed here.
5055 : */
5056 1180 : allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
5057 1180 : XLogCtl->xlblocks = (pg_atomic_uint64 *) allocptr;
5058 1180 : allocptr += sizeof(pg_atomic_uint64) * XLOGbuffers;
5059 :
5060 344289 : for (i = 0; i < XLOGbuffers; i++)
5061 : {
5062 343109 : pg_atomic_init_u64(&XLogCtl->xlblocks[i], InvalidXLogRecPtr);
5063 : }
5064 :
5065 : /* WAL insertion locks. Ensure they're aligned to the full padded size */
5066 1180 : allocptr += sizeof(WALInsertLockPadded) -
5067 1180 : ((uintptr_t) allocptr) % sizeof(WALInsertLockPadded);
5068 1180 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks =
5069 : (WALInsertLockPadded *) allocptr;
5070 1180 : allocptr += sizeof(WALInsertLockPadded) * NUM_XLOGINSERT_LOCKS;
5071 :
5072 10620 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
5073 : {
5074 9440 : LWLockInitialize(&WALInsertLocks[i].l.lock, LWTRANCHE_WAL_INSERT);
5075 9440 : pg_atomic_init_u64(&WALInsertLocks[i].l.insertingAt, InvalidXLogRecPtr);
5076 9440 : WALInsertLocks[i].l.lastImportantAt = InvalidXLogRecPtr;
5077 : }
5078 :
5079 : /*
5080 : * Align the start of the page buffers to a full xlog block size boundary.
5081 : * This simplifies some calculations in XLOG insertion. It is also
5082 : * required for O_DIRECT.
5083 : */
5084 1180 : allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5085 1180 : XLogCtl->pages = allocptr;
5086 1180 : memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5087 :
5088 : /*
5089 : * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5090 : * in additional info.)
5091 : */
5092 1180 : XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5093 1180 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5094 1180 : XLogCtl->InstallXLogFileSegmentActive = false;
5095 1180 : XLogCtl->WalWriterSleeping = false;
5096 :
5097 1180 : SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5098 1180 : SpinLockInit(&XLogCtl->info_lck);
5099 1180 : pg_atomic_init_u64(&XLogCtl->logInsertResult, InvalidXLogRecPtr);
5100 1180 : pg_atomic_init_u64(&XLogCtl->logWriteResult, InvalidXLogRecPtr);
5101 1180 : pg_atomic_init_u64(&XLogCtl->logFlushResult, InvalidXLogRecPtr);
5102 1180 : pg_atomic_init_u64(&XLogCtl->unloggedLSN, InvalidXLogRecPtr);
5103 : }
5104 :
5105 : /*
5106 : * This func must be called ONCE on system install. It creates pg_control
5107 : * and the initial XLOG segment.
5108 : */
5109 : void
5110 51 : BootStrapXLOG(uint32 data_checksum_version)
5111 : {
5112 : CheckPoint checkPoint;
5113 : PGAlignedXLogBlock buffer;
5114 : XLogPageHeader page;
5115 : XLogLongPageHeader longpage;
5116 : XLogRecord *record;
5117 : char *recptr;
5118 : uint64 sysidentifier;
5119 : struct timeval tv;
5120 : pg_crc32c crc;
5121 :
5122 : /* allow ordinary WAL segment creation, like StartupXLOG() would */
5123 51 : SetInstallXLogFileSegmentActive();
5124 :
5125 : /*
5126 : * Select a hopefully-unique system identifier code for this installation.
5127 : * We use the result of gettimeofday(), including the fractional seconds
5128 : * field, as being about as unique as we can easily get. (Think not to
5129 : * use random(), since it hasn't been seeded and there's no portable way
5130 : * to seed it other than the system clock value...) The upper half of the
5131 : * uint64 value is just the tv_sec part, while the lower half contains the
5132 : * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
5133 : * PID for a little extra uniqueness. A person knowing this encoding can
5134 : * determine the initialization time of the installation, which could
5135 : * perhaps be useful sometimes.
5136 : */
5137 51 : gettimeofday(&tv, NULL);
5138 51 : sysidentifier = ((uint64) tv.tv_sec) << 32;
5139 51 : sysidentifier |= ((uint64) tv.tv_usec) << 12;
5140 51 : sysidentifier |= getpid() & 0xFFF;
5141 :
5142 51 : memset(&buffer, 0, sizeof buffer);
5143 51 : page = (XLogPageHeader) &buffer;
5144 :
5145 : /*
5146 : * Set up information for the initial checkpoint record
5147 : *
5148 : * The initial checkpoint record is written to the beginning of the WAL
5149 : * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5150 : * used, so that we can use 0/0 to mean "before any valid WAL segment".
5151 : */
5152 51 : checkPoint.redo = wal_segment_size + SizeOfXLogLongPHD;
5153 51 : checkPoint.ThisTimeLineID = BootstrapTimeLineID;
5154 51 : checkPoint.PrevTimeLineID = BootstrapTimeLineID;
5155 51 : checkPoint.fullPageWrites = fullPageWrites;
5156 51 : checkPoint.logicalDecodingEnabled = (wal_level == WAL_LEVEL_LOGICAL);
5157 51 : checkPoint.wal_level = wal_level;
5158 : checkPoint.nextXid =
5159 51 : FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId);
5160 51 : checkPoint.nextOid = FirstGenbkiObjectId;
5161 51 : checkPoint.nextMulti = FirstMultiXactId;
5162 51 : checkPoint.nextMultiOffset = 1;
5163 51 : checkPoint.oldestXid = FirstNormalTransactionId;
5164 51 : checkPoint.oldestXidDB = Template1DbOid;
5165 51 : checkPoint.oldestMulti = FirstMultiXactId;
5166 51 : checkPoint.oldestMultiDB = Template1DbOid;
5167 51 : checkPoint.oldestCommitTsXid = InvalidTransactionId;
5168 51 : checkPoint.newestCommitTsXid = InvalidTransactionId;
5169 51 : checkPoint.time = (pg_time_t) time(NULL);
5170 51 : checkPoint.oldestActiveXid = InvalidTransactionId;
5171 :
5172 51 : TransamVariables->nextXid = checkPoint.nextXid;
5173 51 : TransamVariables->nextOid = checkPoint.nextOid;
5174 51 : TransamVariables->oidCount = 0;
5175 51 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5176 51 : AdvanceOldestClogXid(checkPoint.oldestXid);
5177 51 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5178 51 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
5179 51 : SetCommitTsLimit(InvalidTransactionId, InvalidTransactionId);
5180 :
5181 : /* Set up the XLOG page header */
5182 51 : page->xlp_magic = XLOG_PAGE_MAGIC;
5183 51 : page->xlp_info = XLP_LONG_HEADER;
5184 51 : page->xlp_tli = BootstrapTimeLineID;
5185 51 : page->xlp_pageaddr = wal_segment_size;
5186 51 : longpage = (XLogLongPageHeader) page;
5187 51 : longpage->xlp_sysid = sysidentifier;
5188 51 : longpage->xlp_seg_size = wal_segment_size;
5189 51 : longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5190 :
5191 : /* Insert the initial checkpoint record */
5192 51 : recptr = ((char *) page + SizeOfXLogLongPHD);
5193 51 : record = (XLogRecord *) recptr;
5194 51 : record->xl_prev = InvalidXLogRecPtr;
5195 51 : record->xl_xid = InvalidTransactionId;
5196 51 : record->xl_tot_len = SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(checkPoint);
5197 51 : record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5198 51 : record->xl_rmid = RM_XLOG_ID;
5199 51 : recptr += SizeOfXLogRecord;
5200 : /* fill the XLogRecordDataHeaderShort struct */
5201 51 : *(recptr++) = (char) XLR_BLOCK_ID_DATA_SHORT;
5202 51 : *(recptr++) = sizeof(checkPoint);
5203 51 : memcpy(recptr, &checkPoint, sizeof(checkPoint));
5204 51 : recptr += sizeof(checkPoint);
5205 : Assert(recptr - (char *) record == record->xl_tot_len);
5206 :
5207 51 : INIT_CRC32C(crc);
5208 51 : COMP_CRC32C(crc, ((char *) record) + SizeOfXLogRecord, record->xl_tot_len - SizeOfXLogRecord);
5209 51 : COMP_CRC32C(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5210 51 : FIN_CRC32C(crc);
5211 51 : record->xl_crc = crc;
5212 :
5213 : /* Create first XLOG segment file */
5214 51 : openLogTLI = BootstrapTimeLineID;
5215 51 : openLogFile = XLogFileInit(1, BootstrapTimeLineID);
5216 :
5217 : /*
5218 : * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
5219 : * close the file again in a moment.
5220 : */
5221 :
5222 : /* Write the first page with the initial record */
5223 51 : errno = 0;
5224 51 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE);
5225 51 : if (write(openLogFile, &buffer, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5226 : {
5227 : /* if write didn't set errno, assume problem is no disk space */
5228 0 : if (errno == 0)
5229 0 : errno = ENOSPC;
5230 0 : ereport(PANIC,
5231 : (errcode_for_file_access(),
5232 : errmsg("could not write bootstrap write-ahead log file: %m")));
5233 : }
5234 51 : pgstat_report_wait_end();
5235 :
5236 51 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC);
5237 51 : if (pg_fsync(openLogFile) != 0)
5238 0 : ereport(PANIC,
5239 : (errcode_for_file_access(),
5240 : errmsg("could not fsync bootstrap write-ahead log file: %m")));
5241 51 : pgstat_report_wait_end();
5242 :
5243 51 : if (close(openLogFile) != 0)
5244 0 : ereport(PANIC,
5245 : (errcode_for_file_access(),
5246 : errmsg("could not close bootstrap write-ahead log file: %m")));
5247 :
5248 51 : openLogFile = -1;
5249 :
5250 : /* Now create pg_control */
5251 51 : InitControlFile(sysidentifier, data_checksum_version);
5252 51 : ControlFile->time = checkPoint.time;
5253 51 : ControlFile->checkPoint = checkPoint.redo;
5254 51 : ControlFile->checkPointCopy = checkPoint;
5255 :
5256 : /* some additional ControlFile fields are set in WriteControlFile() */
5257 51 : WriteControlFile();
5258 :
5259 : /* Bootstrap the commit log, too */
5260 51 : BootStrapCLOG();
5261 51 : BootStrapCommitTs();
5262 51 : BootStrapSUBTRANS();
5263 51 : BootStrapMultiXact();
5264 :
5265 : /*
5266 : * Force control file to be read - in contrast to normal processing we'd
5267 : * otherwise never run the checks and GUC related initializations therein.
5268 : */
5269 51 : ReadControlFile();
5270 51 : }
5271 :
5272 : static char *
5273 923 : str_time(pg_time_t tnow, char *buf, size_t bufsize)
5274 : {
5275 923 : pg_strftime(buf, bufsize,
5276 : "%Y-%m-%d %H:%M:%S %Z",
5277 923 : pg_localtime(&tnow, log_timezone));
5278 :
5279 923 : return buf;
5280 : }
5281 :
5282 : /*
5283 : * Initialize the first WAL segment on new timeline.
5284 : */
5285 : static void
5286 55 : XLogInitNewTimeline(TimeLineID endTLI, XLogRecPtr endOfLog, TimeLineID newTLI)
5287 : {
5288 : char xlogfname[MAXFNAMELEN];
5289 : XLogSegNo endLogSegNo;
5290 : XLogSegNo startLogSegNo;
5291 :
5292 : /* we always switch to a new timeline after archive recovery */
5293 : Assert(endTLI != newTLI);
5294 :
5295 : /*
5296 : * Update min recovery point one last time.
5297 : */
5298 55 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5299 :
5300 : /*
5301 : * Calculate the last segment on the old timeline, and the first segment
5302 : * on the new timeline. If the switch happens in the middle of a segment,
5303 : * they are the same, but if the switch happens exactly at a segment
5304 : * boundary, startLogSegNo will be endLogSegNo + 1.
5305 : */
5306 55 : XLByteToPrevSeg(endOfLog, endLogSegNo, wal_segment_size);
5307 55 : XLByteToSeg(endOfLog, startLogSegNo, wal_segment_size);
5308 :
5309 : /*
5310 : * Initialize the starting WAL segment for the new timeline. If the switch
5311 : * happens in the middle of a segment, copy data from the last WAL segment
5312 : * of the old timeline up to the switch point, to the starting WAL segment
5313 : * on the new timeline.
5314 : */
5315 55 : if (endLogSegNo == startLogSegNo)
5316 : {
5317 : /*
5318 : * Make a copy of the file on the new timeline.
5319 : *
5320 : * Writing WAL isn't allowed yet, so there are no locking
5321 : * considerations. But we should be just as tense as XLogFileInit to
5322 : * avoid emplacing a bogus file.
5323 : */
5324 43 : XLogFileCopy(newTLI, endLogSegNo, endTLI, endLogSegNo,
5325 43 : XLogSegmentOffset(endOfLog, wal_segment_size));
5326 : }
5327 : else
5328 : {
5329 : /*
5330 : * The switch happened at a segment boundary, so just create the next
5331 : * segment on the new timeline.
5332 : */
5333 : int fd;
5334 :
5335 12 : fd = XLogFileInit(startLogSegNo, newTLI);
5336 :
5337 12 : if (close(fd) != 0)
5338 : {
5339 0 : int save_errno = errno;
5340 :
5341 0 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5342 0 : errno = save_errno;
5343 0 : ereport(ERROR,
5344 : (errcode_for_file_access(),
5345 : errmsg("could not close file \"%s\": %m", xlogfname)));
5346 : }
5347 : }
5348 :
5349 : /*
5350 : * Let's just make real sure there are not .ready or .done flags posted
5351 : * for the new segment.
5352 : */
5353 55 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5354 55 : XLogArchiveCleanup(xlogfname);
5355 55 : }
5356 :
5357 : /*
5358 : * Perform cleanup actions at the conclusion of archive recovery.
5359 : */
5360 : static void
5361 55 : CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI, XLogRecPtr EndOfLog,
5362 : TimeLineID newTLI)
5363 : {
5364 : /*
5365 : * Execute the recovery_end_command, if any.
5366 : */
5367 55 : if (recoveryEndCommand && strcmp(recoveryEndCommand, "") != 0)
5368 2 : ExecuteRecoveryCommand(recoveryEndCommand,
5369 : "recovery_end_command",
5370 : true,
5371 : WAIT_EVENT_RECOVERY_END_COMMAND);
5372 :
5373 : /*
5374 : * We switched to a new timeline. Clean up segments on the old timeline.
5375 : *
5376 : * If there are any higher-numbered segments on the old timeline, remove
5377 : * them. They might contain valid WAL, but they might also be
5378 : * pre-allocated files containing garbage. In any case, they are not part
5379 : * of the new timeline's history so we don't need them.
5380 : */
5381 55 : RemoveNonParentXlogFiles(EndOfLog, newTLI);
5382 :
5383 : /*
5384 : * If the switch happened in the middle of a segment, what to do with the
5385 : * last, partial segment on the old timeline? If we don't archive it, and
5386 : * the server that created the WAL never archives it either (e.g. because
5387 : * it was hit by a meteor), it will never make it to the archive. That's
5388 : * OK from our point of view, because the new segment that we created with
5389 : * the new TLI contains all the WAL from the old timeline up to the switch
5390 : * point. But if you later try to do PITR to the "missing" WAL on the old
5391 : * timeline, recovery won't find it in the archive. It's physically
5392 : * present in the new file with new TLI, but recovery won't look there
5393 : * when it's recovering to the older timeline. On the other hand, if we
5394 : * archive the partial segment, and the original server on that timeline
5395 : * is still running and archives the completed version of the same segment
5396 : * later, it will fail. (We used to do that in 9.4 and below, and it
5397 : * caused such problems).
5398 : *
5399 : * As a compromise, we rename the last segment with the .partial suffix,
5400 : * and archive it. Archive recovery will never try to read .partial
5401 : * segments, so they will normally go unused. But in the odd PITR case,
5402 : * the administrator can copy them manually to the pg_wal directory
5403 : * (removing the suffix). They can be useful in debugging, too.
5404 : *
5405 : * If a .done or .ready file already exists for the old timeline, however,
5406 : * we had already determined that the segment is complete, so we can let
5407 : * it be archived normally. (In particular, if it was restored from the
5408 : * archive to begin with, it's expected to have a .done file).
5409 : */
5410 55 : if (XLogSegmentOffset(EndOfLog, wal_segment_size) != 0 &&
5411 : XLogArchivingActive())
5412 : {
5413 : char origfname[MAXFNAMELEN];
5414 : XLogSegNo endLogSegNo;
5415 :
5416 9 : XLByteToPrevSeg(EndOfLog, endLogSegNo, wal_segment_size);
5417 9 : XLogFileName(origfname, EndOfLogTLI, endLogSegNo, wal_segment_size);
5418 :
5419 9 : if (!XLogArchiveIsReadyOrDone(origfname))
5420 : {
5421 : char origpath[MAXPGPATH];
5422 : char partialfname[MAXFNAMELEN];
5423 : char partialpath[MAXPGPATH];
5424 :
5425 : /*
5426 : * If we're summarizing WAL, we can't rename the partial file
5427 : * until the summarizer finishes with it, else it will fail.
5428 : */
5429 5 : if (summarize_wal)
5430 1 : WaitForWalSummarization(EndOfLog);
5431 :
5432 5 : XLogFilePath(origpath, EndOfLogTLI, endLogSegNo, wal_segment_size);
5433 5 : snprintf(partialfname, MAXFNAMELEN, "%s.partial", origfname);
5434 5 : snprintf(partialpath, MAXPGPATH, "%s.partial", origpath);
5435 :
5436 : /*
5437 : * Make sure there's no .done or .ready file for the .partial
5438 : * file.
5439 : */
5440 5 : XLogArchiveCleanup(partialfname);
5441 :
5442 5 : durable_rename(origpath, partialpath, ERROR);
5443 5 : XLogArchiveNotify(partialfname);
5444 : }
5445 : }
5446 55 : }
5447 :
5448 : /*
5449 : * Check to see if required parameters are set high enough on this server
5450 : * for various aspects of recovery operation.
5451 : *
5452 : * Note that all the parameters which this function tests need to be
5453 : * listed in Administrator's Overview section in high-availability.sgml.
5454 : * If you change them, don't forget to update the list.
5455 : */
5456 : static void
5457 262 : CheckRequiredParameterValues(void)
5458 : {
5459 : /*
5460 : * For archive recovery, the WAL must be generated with at least 'replica'
5461 : * wal_level.
5462 : */
5463 262 : if (ArchiveRecoveryRequested && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5464 : {
5465 2 : ereport(FATAL,
5466 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5467 : errmsg("WAL was generated with \"wal_level=minimal\", cannot continue recovering"),
5468 : errdetail("This happens if you temporarily set \"wal_level=minimal\" on the server."),
5469 : errhint("Use a backup taken after setting \"wal_level\" to higher than \"minimal\".")));
5470 : }
5471 :
5472 : /*
5473 : * For Hot Standby, the WAL must be generated with 'replica' mode, and we
5474 : * must have at least as many backend slots as the primary.
5475 : */
5476 260 : if (ArchiveRecoveryRequested && EnableHotStandby)
5477 : {
5478 : /* We ignore autovacuum_worker_slots when we make this test. */
5479 136 : RecoveryRequiresIntParameter("max_connections",
5480 : MaxConnections,
5481 136 : ControlFile->MaxConnections);
5482 136 : RecoveryRequiresIntParameter("max_worker_processes",
5483 : max_worker_processes,
5484 136 : ControlFile->max_worker_processes);
5485 136 : RecoveryRequiresIntParameter("max_wal_senders",
5486 : max_wal_senders,
5487 136 : ControlFile->max_wal_senders);
5488 136 : RecoveryRequiresIntParameter("max_prepared_transactions",
5489 : max_prepared_xacts,
5490 136 : ControlFile->max_prepared_xacts);
5491 136 : RecoveryRequiresIntParameter("max_locks_per_transaction",
5492 : max_locks_per_xact,
5493 136 : ControlFile->max_locks_per_xact);
5494 : }
5495 260 : }
5496 :
5497 : /*
5498 : * This must be called ONCE during postmaster or standalone-backend startup
5499 : */
5500 : void
5501 1035 : StartupXLOG(void)
5502 : {
5503 : XLogCtlInsert *Insert;
5504 : CheckPoint checkPoint;
5505 : bool wasShutdown;
5506 : bool didCrash;
5507 : bool haveTblspcMap;
5508 : bool haveBackupLabel;
5509 : XLogRecPtr EndOfLog;
5510 : TimeLineID EndOfLogTLI;
5511 : TimeLineID newTLI;
5512 : bool performedWalRecovery;
5513 : EndOfWalRecoveryInfo *endOfRecoveryInfo;
5514 : XLogRecPtr abortedRecPtr;
5515 : XLogRecPtr missingContrecPtr;
5516 : TransactionId oldestActiveXID;
5517 1035 : bool promoted = false;
5518 : char timebuf[128];
5519 :
5520 : /*
5521 : * We should have an aux process resource owner to use, and we should not
5522 : * be in a transaction that's installed some other resowner.
5523 : */
5524 : Assert(AuxProcessResourceOwner != NULL);
5525 : Assert(CurrentResourceOwner == NULL ||
5526 : CurrentResourceOwner == AuxProcessResourceOwner);
5527 1035 : CurrentResourceOwner = AuxProcessResourceOwner;
5528 :
5529 : /*
5530 : * Check that contents look valid.
5531 : */
5532 1035 : if (!XRecOffIsValid(ControlFile->checkPoint))
5533 0 : ereport(FATAL,
5534 : (errcode(ERRCODE_DATA_CORRUPTED),
5535 : errmsg("control file contains invalid checkpoint location")));
5536 :
5537 1035 : switch (ControlFile->state)
5538 : {
5539 816 : case DB_SHUTDOWNED:
5540 :
5541 : /*
5542 : * This is the expected case, so don't be chatty in standalone
5543 : * mode
5544 : */
5545 816 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
5546 : (errmsg("database system was shut down at %s",
5547 : str_time(ControlFile->time,
5548 : timebuf, sizeof(timebuf)))));
5549 816 : break;
5550 :
5551 33 : case DB_SHUTDOWNED_IN_RECOVERY:
5552 33 : ereport(LOG,
5553 : (errmsg("database system was shut down in recovery at %s",
5554 : str_time(ControlFile->time,
5555 : timebuf, sizeof(timebuf)))));
5556 33 : break;
5557 :
5558 0 : case DB_SHUTDOWNING:
5559 0 : ereport(LOG,
5560 : (errmsg("database system shutdown was interrupted; last known up at %s",
5561 : str_time(ControlFile->time,
5562 : timebuf, sizeof(timebuf)))));
5563 0 : break;
5564 :
5565 0 : case DB_IN_CRASH_RECOVERY:
5566 0 : ereport(LOG,
5567 : (errmsg("database system was interrupted while in recovery at %s",
5568 : str_time(ControlFile->time,
5569 : timebuf, sizeof(timebuf))),
5570 : errhint("This probably means that some data is corrupted and"
5571 : " you will have to use the last backup for recovery.")));
5572 0 : break;
5573 :
5574 7 : case DB_IN_ARCHIVE_RECOVERY:
5575 7 : ereport(LOG,
5576 : (errmsg("database system was interrupted while in recovery at log time %s",
5577 : str_time(ControlFile->checkPointCopy.time,
5578 : timebuf, sizeof(timebuf))),
5579 : errhint("If this has occurred more than once some data might be corrupted"
5580 : " and you might need to choose an earlier recovery target.")));
5581 7 : break;
5582 :
5583 179 : case DB_IN_PRODUCTION:
5584 179 : ereport(LOG,
5585 : (errmsg("database system was interrupted; last known up at %s",
5586 : str_time(ControlFile->time,
5587 : timebuf, sizeof(timebuf)))));
5588 179 : break;
5589 :
5590 0 : default:
5591 0 : ereport(FATAL,
5592 : (errcode(ERRCODE_DATA_CORRUPTED),
5593 : errmsg("control file contains invalid database cluster state")));
5594 : }
5595 :
5596 : /* This is just to allow attaching to startup process with a debugger */
5597 : #ifdef XLOG_REPLAY_DELAY
5598 : if (ControlFile->state != DB_SHUTDOWNED)
5599 : pg_usleep(60000000L);
5600 : #endif
5601 :
5602 : /*
5603 : * Verify that pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
5604 : * In cases where someone has performed a copy for PITR, these directories
5605 : * may have been excluded and need to be re-created.
5606 : */
5607 1035 : ValidateXLOGDirectoryStructure();
5608 :
5609 : /* Set up timeout handler needed to report startup progress. */
5610 1035 : if (!IsBootstrapProcessingMode())
5611 984 : RegisterTimeout(STARTUP_PROGRESS_TIMEOUT,
5612 : startup_progress_timeout_handler);
5613 :
5614 : /*----------
5615 : * If we previously crashed, perform a couple of actions:
5616 : *
5617 : * - The pg_wal directory may still include some temporary WAL segments
5618 : * used when creating a new segment, so perform some clean up to not
5619 : * bloat this path. This is done first as there is no point to sync
5620 : * this temporary data.
5621 : *
5622 : * - There might be data which we had written, intending to fsync it, but
5623 : * which we had not actually fsync'd yet. Therefore, a power failure in
5624 : * the near future might cause earlier unflushed writes to be lost, even
5625 : * though more recent data written to disk from here on would be
5626 : * persisted. To avoid that, fsync the entire data directory.
5627 : */
5628 1035 : if (ControlFile->state != DB_SHUTDOWNED &&
5629 219 : ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
5630 : {
5631 186 : RemoveTempXlogFiles();
5632 186 : SyncDataDirectory();
5633 186 : didCrash = true;
5634 : }
5635 : else
5636 849 : didCrash = false;
5637 :
5638 : /*
5639 : * Prepare for WAL recovery if needed.
5640 : *
5641 : * InitWalRecovery analyzes the control file and the backup label file, if
5642 : * any. It updates the in-memory ControlFile buffer according to the
5643 : * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5644 : * It also applies the tablespace map file, if any.
5645 : */
5646 1035 : InitWalRecovery(ControlFile, &wasShutdown,
5647 : &haveBackupLabel, &haveTblspcMap);
5648 1033 : checkPoint = ControlFile->checkPointCopy;
5649 :
5650 : /* initialize shared memory variables from the checkpoint record */
5651 1033 : TransamVariables->nextXid = checkPoint.nextXid;
5652 1033 : TransamVariables->nextOid = checkPoint.nextOid;
5653 1033 : TransamVariables->oidCount = 0;
5654 1033 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5655 1033 : AdvanceOldestClogXid(checkPoint.oldestXid);
5656 1033 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5657 1033 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
5658 1033 : SetCommitTsLimit(checkPoint.oldestCommitTsXid,
5659 : checkPoint.newestCommitTsXid);
5660 :
5661 : /*
5662 : * Clear out any old relcache cache files. This is *necessary* if we do
5663 : * any WAL replay, since that would probably result in the cache files
5664 : * being out of sync with database reality. In theory we could leave them
5665 : * in place if the database had been cleanly shut down, but it seems
5666 : * safest to just remove them always and let them be rebuilt during the
5667 : * first backend startup. These files needs to be removed from all
5668 : * directories including pg_tblspc, however the symlinks are created only
5669 : * after reading tablespace_map file in case of archive recovery from
5670 : * backup, so needs to clear old relcache files here after creating
5671 : * symlinks.
5672 : */
5673 1033 : RelationCacheInitFileRemove();
5674 :
5675 : /*
5676 : * Initialize replication slots, before there's a chance to remove
5677 : * required resources.
5678 : */
5679 1033 : StartupReplicationSlots();
5680 :
5681 : /*
5682 : * Startup the logical decoding status with the last status stored in the
5683 : * checkpoint record.
5684 : */
5685 1031 : StartupLogicalDecodingStatus(checkPoint.logicalDecodingEnabled);
5686 :
5687 : /*
5688 : * Startup logical state, needs to be setup now so we have proper data
5689 : * during crash recovery.
5690 : */
5691 1031 : StartupReorderBuffer();
5692 :
5693 : /*
5694 : * Startup CLOG. This must be done after TransamVariables->nextXid has
5695 : * been initialized and before we accept connections or begin WAL replay.
5696 : */
5697 1031 : StartupCLOG();
5698 :
5699 : /*
5700 : * Startup MultiXact. We need to do this early to be able to replay
5701 : * truncations.
5702 : */
5703 1031 : StartupMultiXact();
5704 :
5705 : /*
5706 : * Ditto for commit timestamps. Activate the facility if the setting is
5707 : * enabled in the control file, as there should be no tracking of commit
5708 : * timestamps done when the setting was disabled. This facility can be
5709 : * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
5710 : */
5711 1031 : if (ControlFile->track_commit_timestamp)
5712 14 : StartupCommitTs();
5713 :
5714 : /*
5715 : * Recover knowledge about replay progress of known replication partners.
5716 : */
5717 1031 : StartupReplicationOrigin();
5718 :
5719 : /*
5720 : * Initialize unlogged LSN. On a clean shutdown, it's restored from the
5721 : * control file. On recovery, all unlogged relations are blown away, so
5722 : * the unlogged LSN counter can be reset too.
5723 : */
5724 1031 : if (ControlFile->state == DB_SHUTDOWNED)
5725 807 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5726 807 : ControlFile->unloggedLSN);
5727 : else
5728 224 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5729 : FirstNormalUnloggedLSN);
5730 :
5731 : /*
5732 : * Copy any missing timeline history files between 'now' and the recovery
5733 : * target timeline from archive to pg_wal. While we don't need those files
5734 : * ourselves - the history file of the recovery target timeline covers all
5735 : * the previous timelines in the history too - a cascading standby server
5736 : * might be interested in them. Or, if you archive the WAL from this
5737 : * server to a different archive than the primary, it'd be good for all
5738 : * the history files to get archived there after failover, so that you can
5739 : * use one of the old timelines as a PITR target. Timeline history files
5740 : * are small, so it's better to copy them unnecessarily than not copy them
5741 : * and regret later.
5742 : */
5743 1031 : restoreTimeLineHistoryFiles(checkPoint.ThisTimeLineID, recoveryTargetTLI);
5744 :
5745 : /*
5746 : * Before running in recovery, scan pg_twophase and fill in its status to
5747 : * be able to work on entries generated by redo. Doing a scan before
5748 : * taking any recovery action has the merit to discard any 2PC files that
5749 : * are newer than the first record to replay, saving from any conflicts at
5750 : * replay. This avoids as well any subsequent scans when doing recovery
5751 : * of the on-disk two-phase data.
5752 : */
5753 1031 : restoreTwoPhaseData();
5754 :
5755 : /*
5756 : * When starting with crash recovery, reset pgstat data - it might not be
5757 : * valid. Otherwise restore pgstat data. It's safe to do this here,
5758 : * because postmaster will not yet have started any other processes.
5759 : *
5760 : * NB: Restoring replication slot stats relies on slot state to have
5761 : * already been restored from disk.
5762 : *
5763 : * TODO: With a bit of extra work we could just start with a pgstat file
5764 : * associated with the checkpoint redo location we're starting from.
5765 : */
5766 1031 : if (didCrash)
5767 184 : pgstat_discard_stats();
5768 : else
5769 847 : pgstat_restore_stats();
5770 :
5771 1031 : lastFullPageWrites = checkPoint.fullPageWrites;
5772 :
5773 1031 : RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
5774 1031 : doPageWrites = lastFullPageWrites;
5775 :
5776 : /* REDO */
5777 1031 : if (InRecovery)
5778 : {
5779 : /* Initialize state for RecoveryInProgress() */
5780 224 : SpinLockAcquire(&XLogCtl->info_lck);
5781 224 : if (InArchiveRecovery)
5782 122 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
5783 : else
5784 102 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5785 224 : SpinLockRelease(&XLogCtl->info_lck);
5786 :
5787 : /*
5788 : * Update pg_control to show that we are recovering and to show the
5789 : * selected checkpoint as the place we are starting from. We also mark
5790 : * pg_control with any minimum recovery stop point obtained from a
5791 : * backup history file.
5792 : *
5793 : * No need to hold ControlFileLock yet, we aren't up far enough.
5794 : */
5795 224 : UpdateControlFile();
5796 :
5797 : /*
5798 : * If there was a backup label file, it's done its job and the info
5799 : * has now been propagated into pg_control. We must get rid of the
5800 : * label file so that if we crash during recovery, we'll pick up at
5801 : * the latest recovery restartpoint instead of going all the way back
5802 : * to the backup start point. It seems prudent though to just rename
5803 : * the file out of the way rather than delete it completely.
5804 : */
5805 224 : if (haveBackupLabel)
5806 : {
5807 80 : unlink(BACKUP_LABEL_OLD);
5808 80 : durable_rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD, FATAL);
5809 : }
5810 :
5811 : /*
5812 : * If there was a tablespace_map file, it's done its job and the
5813 : * symlinks have been created. We must get rid of the map file so
5814 : * that if we crash during recovery, we don't create symlinks again.
5815 : * It seems prudent though to just rename the file out of the way
5816 : * rather than delete it completely.
5817 : */
5818 224 : if (haveTblspcMap)
5819 : {
5820 2 : unlink(TABLESPACE_MAP_OLD);
5821 2 : durable_rename(TABLESPACE_MAP, TABLESPACE_MAP_OLD, FATAL);
5822 : }
5823 :
5824 : /*
5825 : * Initialize our local copy of minRecoveryPoint. When doing crash
5826 : * recovery we want to replay up to the end of WAL. Particularly, in
5827 : * the case of a promoted standby minRecoveryPoint value in the
5828 : * control file is only updated after the first checkpoint. However,
5829 : * if the instance crashes before the first post-recovery checkpoint
5830 : * is completed then recovery will use a stale location causing the
5831 : * startup process to think that there are still invalid page
5832 : * references when checking for data consistency.
5833 : */
5834 224 : if (InArchiveRecovery)
5835 : {
5836 122 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
5837 122 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
5838 : }
5839 : else
5840 : {
5841 102 : LocalMinRecoveryPoint = InvalidXLogRecPtr;
5842 102 : LocalMinRecoveryPointTLI = 0;
5843 : }
5844 :
5845 : /* Check that the GUCs used to generate the WAL allow recovery */
5846 224 : CheckRequiredParameterValues();
5847 :
5848 : /*
5849 : * We're in recovery, so unlogged relations may be trashed and must be
5850 : * reset. This should be done BEFORE allowing Hot Standby
5851 : * connections, so that read-only backends don't try to read whatever
5852 : * garbage is left over from before.
5853 : */
5854 224 : ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
5855 :
5856 : /*
5857 : * Likewise, delete any saved transaction snapshot files that got left
5858 : * behind by crashed backends.
5859 : */
5860 224 : DeleteAllExportedSnapshotFiles();
5861 :
5862 : /*
5863 : * Initialize for Hot Standby, if enabled. We won't let backends in
5864 : * yet, not until we've reached the min recovery point specified in
5865 : * control file and we've established a recovery snapshot from a
5866 : * running-xacts WAL record.
5867 : */
5868 224 : if (ArchiveRecoveryRequested && EnableHotStandby)
5869 : {
5870 : TransactionId *xids;
5871 : int nxids;
5872 :
5873 114 : ereport(DEBUG1,
5874 : (errmsg_internal("initializing for hot standby")));
5875 :
5876 114 : InitRecoveryTransactionEnvironment();
5877 :
5878 114 : if (wasShutdown)
5879 26 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
5880 : else
5881 88 : oldestActiveXID = checkPoint.oldestActiveXid;
5882 : Assert(TransactionIdIsValid(oldestActiveXID));
5883 :
5884 : /* Tell procarray about the range of xids it has to deal with */
5885 114 : ProcArrayInitRecovery(XidFromFullTransactionId(TransamVariables->nextXid));
5886 :
5887 : /*
5888 : * Startup subtrans only. CLOG, MultiXact and commit timestamp
5889 : * have already been started up and other SLRUs are not maintained
5890 : * during recovery and need not be started yet.
5891 : */
5892 114 : StartupSUBTRANS(oldestActiveXID);
5893 :
5894 : /*
5895 : * If we're beginning at a shutdown checkpoint, we know that
5896 : * nothing was running on the primary at this point. So fake-up an
5897 : * empty running-xacts record and use that here and now. Recover
5898 : * additional standby state for prepared transactions.
5899 : */
5900 114 : if (wasShutdown)
5901 : {
5902 : RunningTransactionsData running;
5903 : TransactionId latestCompletedXid;
5904 :
5905 : /* Update pg_subtrans entries for any prepared transactions */
5906 26 : StandbyRecoverPreparedTransactions();
5907 :
5908 : /*
5909 : * Construct a RunningTransactions snapshot representing a
5910 : * shut down server, with only prepared transactions still
5911 : * alive. We're never overflowed at this point because all
5912 : * subxids are listed with their parent prepared transactions.
5913 : */
5914 26 : running.xcnt = nxids;
5915 26 : running.subxcnt = 0;
5916 26 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
5917 26 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
5918 26 : running.oldestRunningXid = oldestActiveXID;
5919 26 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
5920 26 : TransactionIdRetreat(latestCompletedXid);
5921 : Assert(TransactionIdIsNormal(latestCompletedXid));
5922 26 : running.latestCompletedXid = latestCompletedXid;
5923 26 : running.xids = xids;
5924 :
5925 26 : ProcArrayApplyRecoveryInfo(&running);
5926 : }
5927 : }
5928 :
5929 : /*
5930 : * We're all set for replaying the WAL now. Do it.
5931 : */
5932 224 : PerformWalRecovery();
5933 163 : performedWalRecovery = true;
5934 : }
5935 : else
5936 807 : performedWalRecovery = false;
5937 :
5938 : /*
5939 : * Finish WAL recovery.
5940 : */
5941 970 : endOfRecoveryInfo = FinishWalRecovery();
5942 970 : EndOfLog = endOfRecoveryInfo->endOfLog;
5943 970 : EndOfLogTLI = endOfRecoveryInfo->endOfLogTLI;
5944 970 : abortedRecPtr = endOfRecoveryInfo->abortedRecPtr;
5945 970 : missingContrecPtr = endOfRecoveryInfo->missingContrecPtr;
5946 :
5947 : /*
5948 : * Reset ps status display, so as no information related to recovery shows
5949 : * up.
5950 : */
5951 970 : set_ps_display("");
5952 :
5953 : /*
5954 : * When recovering from a backup (we are in recovery, and archive recovery
5955 : * was requested), complain if we did not roll forward far enough to reach
5956 : * the point where the database is consistent. For regular online
5957 : * backup-from-primary, that means reaching the end-of-backup WAL record
5958 : * (at which point we reset backupStartPoint to be Invalid), for
5959 : * backup-from-replica (which can't inject records into the WAL stream),
5960 : * that point is when we reach the minRecoveryPoint in pg_control (which
5961 : * we purposefully copy last when backing up from a replica). For
5962 : * pg_rewind (which creates a backup_label with a method of "pg_rewind")
5963 : * or snapshot-style backups (which don't), backupEndRequired will be set
5964 : * to false.
5965 : *
5966 : * Note: it is indeed okay to look at the local variable
5967 : * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
5968 : * might be further ahead --- ControlFile->minRecoveryPoint cannot have
5969 : * been advanced beyond the WAL we processed.
5970 : */
5971 970 : if (InRecovery &&
5972 163 : (EndOfLog < LocalMinRecoveryPoint ||
5973 163 : XLogRecPtrIsValid(ControlFile->backupStartPoint)))
5974 : {
5975 : /*
5976 : * Ran off end of WAL before reaching end-of-backup WAL record, or
5977 : * minRecoveryPoint. That's a bad sign, indicating that you tried to
5978 : * recover from an online backup but never called pg_backup_stop(), or
5979 : * you didn't archive all the WAL needed.
5980 : */
5981 0 : if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
5982 : {
5983 0 : if (XLogRecPtrIsValid(ControlFile->backupStartPoint) || ControlFile->backupEndRequired)
5984 0 : ereport(FATAL,
5985 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5986 : errmsg("WAL ends before end of online backup"),
5987 : errhint("All WAL generated while online backup was taken must be available at recovery.")));
5988 : else
5989 0 : ereport(FATAL,
5990 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5991 : errmsg("WAL ends before consistent recovery point")));
5992 : }
5993 : }
5994 :
5995 : /*
5996 : * Reset unlogged relations to the contents of their INIT fork. This is
5997 : * done AFTER recovery is complete so as to include any unlogged relations
5998 : * created during recovery, but BEFORE recovery is marked as having
5999 : * completed successfully. Otherwise we'd not retry if any of the post
6000 : * end-of-recovery steps fail.
6001 : */
6002 970 : if (InRecovery)
6003 163 : ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
6004 :
6005 : /*
6006 : * Pre-scan prepared transactions to find out the range of XIDs present.
6007 : * This information is not quite needed yet, but it is positioned here so
6008 : * as potential problems are detected before any on-disk change is done.
6009 : */
6010 970 : oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
6011 :
6012 : /*
6013 : * Allow ordinary WAL segment creation before possibly switching to a new
6014 : * timeline, which creates a new segment, and after the last ReadRecord().
6015 : */
6016 970 : SetInstallXLogFileSegmentActive();
6017 :
6018 : /*
6019 : * Consider whether we need to assign a new timeline ID.
6020 : *
6021 : * If we did archive recovery, we always assign a new ID. This handles a
6022 : * couple of issues. If we stopped short of the end of WAL during
6023 : * recovery, then we are clearly generating a new timeline and must assign
6024 : * it a unique new ID. Even if we ran to the end, modifying the current
6025 : * last segment is problematic because it may result in trying to
6026 : * overwrite an already-archived copy of that segment, and we encourage
6027 : * DBAs to make their archive_commands reject that. We can dodge the
6028 : * problem by making the new active segment have a new timeline ID.
6029 : *
6030 : * In a normal crash recovery, we can just extend the timeline we were in.
6031 : */
6032 970 : newTLI = endOfRecoveryInfo->lastRecTLI;
6033 970 : if (ArchiveRecoveryRequested)
6034 : {
6035 55 : newTLI = findNewestTimeLine(recoveryTargetTLI) + 1;
6036 55 : ereport(LOG,
6037 : (errmsg("selected new timeline ID: %u", newTLI)));
6038 :
6039 : /*
6040 : * Make a writable copy of the last WAL segment. (Note that we also
6041 : * have a copy of the last block of the old WAL in
6042 : * endOfRecovery->lastPage; we will use that below.)
6043 : */
6044 55 : XLogInitNewTimeline(EndOfLogTLI, EndOfLog, newTLI);
6045 :
6046 : /*
6047 : * Remove the signal files out of the way, so that we don't
6048 : * accidentally re-enter archive recovery mode in a subsequent crash.
6049 : */
6050 55 : if (endOfRecoveryInfo->standby_signal_file_found)
6051 52 : durable_unlink(STANDBY_SIGNAL_FILE, FATAL);
6052 :
6053 55 : if (endOfRecoveryInfo->recovery_signal_file_found)
6054 4 : durable_unlink(RECOVERY_SIGNAL_FILE, FATAL);
6055 :
6056 : /*
6057 : * Write the timeline history file, and have it archived. After this
6058 : * point (or rather, as soon as the file is archived), the timeline
6059 : * will appear as "taken" in the WAL archive and to any standby
6060 : * servers. If we crash before actually switching to the new
6061 : * timeline, standby servers will nevertheless think that we switched
6062 : * to the new timeline, and will try to connect to the new timeline.
6063 : * To minimize the window for that, try to do as little as possible
6064 : * between here and writing the end-of-recovery record.
6065 : */
6066 55 : writeTimeLineHistory(newTLI, recoveryTargetTLI,
6067 : EndOfLog, endOfRecoveryInfo->recoveryStopReason);
6068 :
6069 55 : ereport(LOG,
6070 : (errmsg("archive recovery complete")));
6071 : }
6072 :
6073 : /* Save the selected TimeLineID in shared memory, too */
6074 970 : SpinLockAcquire(&XLogCtl->info_lck);
6075 970 : XLogCtl->InsertTimeLineID = newTLI;
6076 970 : XLogCtl->PrevTimeLineID = endOfRecoveryInfo->lastRecTLI;
6077 970 : SpinLockRelease(&XLogCtl->info_lck);
6078 :
6079 : /*
6080 : * Actually, if WAL ended in an incomplete record, skip the parts that
6081 : * made it through and start writing after the portion that persisted.
6082 : * (It's critical to first write an OVERWRITE_CONTRECORD message, which
6083 : * we'll do as soon as we're open for writing new WAL.)
6084 : */
6085 970 : if (XLogRecPtrIsValid(missingContrecPtr))
6086 : {
6087 : /*
6088 : * We should only have a missingContrecPtr if we're not switching to a
6089 : * new timeline. When a timeline switch occurs, WAL is copied from the
6090 : * old timeline to the new only up to the end of the last complete
6091 : * record, so there can't be an incomplete WAL record that we need to
6092 : * disregard.
6093 : */
6094 : Assert(newTLI == endOfRecoveryInfo->lastRecTLI);
6095 : Assert(XLogRecPtrIsValid(abortedRecPtr));
6096 10 : EndOfLog = missingContrecPtr;
6097 : }
6098 :
6099 : /*
6100 : * Prepare to write WAL starting at EndOfLog location, and init xlog
6101 : * buffer cache using the block containing the last record from the
6102 : * previous incarnation.
6103 : */
6104 970 : Insert = &XLogCtl->Insert;
6105 970 : Insert->PrevBytePos = XLogRecPtrToBytePos(endOfRecoveryInfo->lastRec);
6106 970 : Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
6107 :
6108 : /*
6109 : * Tricky point here: lastPage contains the *last* block that the LastRec
6110 : * record spans, not the one it starts in. The last block is indeed the
6111 : * one we want to use.
6112 : */
6113 970 : if (EndOfLog % XLOG_BLCKSZ != 0)
6114 : {
6115 : char *page;
6116 : int len;
6117 : int firstIdx;
6118 :
6119 938 : firstIdx = XLogRecPtrToBufIdx(EndOfLog);
6120 938 : len = EndOfLog - endOfRecoveryInfo->lastPageBeginPtr;
6121 : Assert(len < XLOG_BLCKSZ);
6122 :
6123 : /* Copy the valid part of the last block, and zero the rest */
6124 938 : page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
6125 938 : memcpy(page, endOfRecoveryInfo->lastPage, len);
6126 938 : memset(page + len, 0, XLOG_BLCKSZ - len);
6127 :
6128 938 : pg_atomic_write_u64(&XLogCtl->xlblocks[firstIdx], endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6129 938 : XLogCtl->InitializedUpTo = endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ;
6130 : }
6131 : else
6132 : {
6133 : /*
6134 : * There is no partial block to copy. Just set InitializedUpTo, and
6135 : * let the first attempt to insert a log record to initialize the next
6136 : * buffer.
6137 : */
6138 32 : XLogCtl->InitializedUpTo = EndOfLog;
6139 : }
6140 :
6141 : /*
6142 : * Update local and shared status. This is OK to do without any locks
6143 : * because no other process can be reading or writing WAL yet.
6144 : */
6145 970 : LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
6146 970 : pg_atomic_write_u64(&XLogCtl->logInsertResult, EndOfLog);
6147 970 : pg_atomic_write_u64(&XLogCtl->logWriteResult, EndOfLog);
6148 970 : pg_atomic_write_u64(&XLogCtl->logFlushResult, EndOfLog);
6149 970 : XLogCtl->LogwrtRqst.Write = EndOfLog;
6150 970 : XLogCtl->LogwrtRqst.Flush = EndOfLog;
6151 :
6152 : /*
6153 : * Preallocate additional log files, if wanted.
6154 : */
6155 970 : PreallocXlogFiles(EndOfLog, newTLI);
6156 :
6157 : /*
6158 : * Okay, we're officially UP.
6159 : */
6160 970 : InRecovery = false;
6161 :
6162 : /* start the archive_timeout timer and LSN running */
6163 970 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
6164 970 : XLogCtl->lastSegSwitchLSN = EndOfLog;
6165 :
6166 : /* also initialize latestCompletedXid, to nextXid - 1 */
6167 970 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
6168 970 : TransamVariables->latestCompletedXid = TransamVariables->nextXid;
6169 970 : FullTransactionIdRetreat(&TransamVariables->latestCompletedXid);
6170 970 : LWLockRelease(ProcArrayLock);
6171 :
6172 : /*
6173 : * Start up subtrans, if not already done for hot standby. (commit
6174 : * timestamps are started below, if necessary.)
6175 : */
6176 970 : if (standbyState == STANDBY_DISABLED)
6177 915 : StartupSUBTRANS(oldestActiveXID);
6178 :
6179 : /*
6180 : * Perform end of recovery actions for any SLRUs that need it.
6181 : */
6182 970 : TrimCLOG();
6183 970 : TrimMultiXact();
6184 :
6185 : /*
6186 : * Reload shared-memory state for prepared transactions. This needs to
6187 : * happen before renaming the last partial segment of the old timeline as
6188 : * it may be possible that we have to recover some transactions from it.
6189 : */
6190 970 : RecoverPreparedTransactions();
6191 :
6192 : /* Shut down xlogreader */
6193 970 : ShutdownWalRecovery();
6194 :
6195 : /* Enable WAL writes for this backend only. */
6196 970 : LocalSetXLogInsertAllowed();
6197 :
6198 : /* If necessary, write overwrite-contrecord before doing anything else */
6199 970 : if (XLogRecPtrIsValid(abortedRecPtr))
6200 : {
6201 : Assert(XLogRecPtrIsValid(missingContrecPtr));
6202 10 : CreateOverwriteContrecordRecord(abortedRecPtr, missingContrecPtr, newTLI);
6203 : }
6204 :
6205 : /*
6206 : * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
6207 : * record before resource manager writes cleanup WAL records or checkpoint
6208 : * record is written.
6209 : */
6210 970 : Insert->fullPageWrites = lastFullPageWrites;
6211 970 : UpdateFullPageWrites();
6212 :
6213 : /*
6214 : * Emit checkpoint or end-of-recovery record in XLOG, if required.
6215 : */
6216 970 : if (performedWalRecovery)
6217 163 : promoted = PerformRecoveryXLogAction();
6218 :
6219 : /*
6220 : * If any of the critical GUCs have changed, log them before we allow
6221 : * backends to write WAL.
6222 : */
6223 970 : XLogReportParameters();
6224 :
6225 : /* If this is archive recovery, perform post-recovery cleanup actions. */
6226 970 : if (ArchiveRecoveryRequested)
6227 55 : CleanupAfterArchiveRecovery(EndOfLogTLI, EndOfLog, newTLI);
6228 :
6229 : /*
6230 : * Local WAL inserts enabled, so it's time to finish initialization of
6231 : * commit timestamp.
6232 : */
6233 970 : CompleteCommitTsInitialization();
6234 :
6235 : /*
6236 : * Update logical decoding status in shared memory and write an
6237 : * XLOG_LOGICAL_DECODING_STATUS_CHANGE, if necessary.
6238 : */
6239 970 : UpdateLogicalDecodingStatusEndOfRecovery();
6240 :
6241 : /* Clean up EndOfWalRecoveryInfo data to appease Valgrind leak checking */
6242 970 : if (endOfRecoveryInfo->lastPage)
6243 948 : pfree(endOfRecoveryInfo->lastPage);
6244 970 : pfree(endOfRecoveryInfo->recoveryStopReason);
6245 970 : pfree(endOfRecoveryInfo);
6246 :
6247 : /*
6248 : * All done with end-of-recovery actions.
6249 : *
6250 : * Now allow backends to write WAL and update the control file status in
6251 : * consequence. SharedRecoveryState, that controls if backends can write
6252 : * WAL, is updated while holding ControlFileLock to prevent other backends
6253 : * to look at an inconsistent state of the control file in shared memory.
6254 : * There is still a small window during which backends can write WAL and
6255 : * the control file is still referring to a system not in DB_IN_PRODUCTION
6256 : * state while looking at the on-disk control file.
6257 : *
6258 : * Also, we use info_lck to update SharedRecoveryState to ensure that
6259 : * there are no race conditions concerning visibility of other recent
6260 : * updates to shared memory.
6261 : */
6262 970 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6263 970 : ControlFile->state = DB_IN_PRODUCTION;
6264 :
6265 970 : SpinLockAcquire(&XLogCtl->info_lck);
6266 970 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_DONE;
6267 970 : SpinLockRelease(&XLogCtl->info_lck);
6268 :
6269 970 : UpdateControlFile();
6270 970 : LWLockRelease(ControlFileLock);
6271 :
6272 : /*
6273 : * Wake up the checkpointer process as there might be a request to disable
6274 : * logical decoding by concurrent slot drop.
6275 : */
6276 970 : WakeupCheckpointer();
6277 :
6278 : /*
6279 : * Wake up all waiters. They need to report an error that recovery was
6280 : * ended before reaching the target LSN.
6281 : */
6282 970 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_REPLAY, InvalidXLogRecPtr);
6283 970 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_WRITE, InvalidXLogRecPtr);
6284 970 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_FLUSH, InvalidXLogRecPtr);
6285 :
6286 : /*
6287 : * Shutdown the recovery environment. This must occur after
6288 : * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
6289 : * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
6290 : * any session building a snapshot will not rely on KnownAssignedXids as
6291 : * RecoveryInProgress() would return false at this stage. This is
6292 : * particularly critical for prepared 2PC transactions, that would still
6293 : * need to be included in snapshots once recovery has ended.
6294 : */
6295 970 : if (standbyState != STANDBY_DISABLED)
6296 55 : ShutdownRecoveryTransactionEnvironment();
6297 :
6298 : /*
6299 : * If there were cascading standby servers connected to us, nudge any wal
6300 : * sender processes to notice that we've been promoted.
6301 : */
6302 970 : WalSndWakeup(true, true);
6303 :
6304 : /*
6305 : * If this was a promotion, request an (online) checkpoint now. This isn't
6306 : * required for consistency, but the last restartpoint might be far back,
6307 : * and in case of a crash, recovering from it might take a longer than is
6308 : * appropriate now that we're not in standby mode anymore.
6309 : */
6310 970 : if (promoted)
6311 48 : RequestCheckpoint(CHECKPOINT_FORCE);
6312 970 : }
6313 :
6314 : /*
6315 : * Callback from PerformWalRecovery(), called when we switch from crash
6316 : * recovery to archive recovery mode. Updates the control file accordingly.
6317 : */
6318 : void
6319 1 : SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr, TimeLineID replayTLI)
6320 : {
6321 : /* initialize minRecoveryPoint to this record */
6322 1 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6323 1 : ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6324 1 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6325 : {
6326 1 : ControlFile->minRecoveryPoint = EndRecPtr;
6327 1 : ControlFile->minRecoveryPointTLI = replayTLI;
6328 : }
6329 : /* update local copy */
6330 1 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6331 1 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6332 :
6333 : /*
6334 : * The startup process can update its local copy of minRecoveryPoint from
6335 : * this point.
6336 : */
6337 1 : updateMinRecoveryPoint = true;
6338 :
6339 1 : UpdateControlFile();
6340 :
6341 : /*
6342 : * We update SharedRecoveryState while holding the lock on ControlFileLock
6343 : * so both states are consistent in shared memory.
6344 : */
6345 1 : SpinLockAcquire(&XLogCtl->info_lck);
6346 1 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6347 1 : SpinLockRelease(&XLogCtl->info_lck);
6348 :
6349 1 : LWLockRelease(ControlFileLock);
6350 1 : }
6351 :
6352 : /*
6353 : * Callback from PerformWalRecovery(), called when we reach the end of backup.
6354 : * Updates the control file accordingly.
6355 : */
6356 : void
6357 80 : ReachedEndOfBackup(XLogRecPtr EndRecPtr, TimeLineID tli)
6358 : {
6359 : /*
6360 : * We have reached the end of base backup, as indicated by pg_control. The
6361 : * data on disk is now consistent (unless minRecoveryPoint is further
6362 : * ahead, which can happen if we crashed during previous recovery). Reset
6363 : * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
6364 : * make sure we don't allow starting up at an earlier point even if
6365 : * recovery is stopped and restarted soon after this.
6366 : */
6367 80 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6368 :
6369 80 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6370 : {
6371 75 : ControlFile->minRecoveryPoint = EndRecPtr;
6372 75 : ControlFile->minRecoveryPointTLI = tli;
6373 : }
6374 :
6375 80 : ControlFile->backupStartPoint = InvalidXLogRecPtr;
6376 80 : ControlFile->backupEndPoint = InvalidXLogRecPtr;
6377 80 : ControlFile->backupEndRequired = false;
6378 80 : UpdateControlFile();
6379 :
6380 80 : LWLockRelease(ControlFileLock);
6381 80 : }
6382 :
6383 : /*
6384 : * Perform whatever XLOG actions are necessary at end of REDO.
6385 : *
6386 : * The goal here is to make sure that we'll be able to recover properly if
6387 : * we crash again. If we choose to write a checkpoint, we'll write a shutdown
6388 : * checkpoint rather than an on-line one. This is not particularly critical,
6389 : * but since we may be assigning a new TLI, using a shutdown checkpoint allows
6390 : * us to have the rule that TLI only changes in shutdown checkpoints, which
6391 : * allows some extra error checking in xlog_redo.
6392 : */
6393 : static bool
6394 163 : PerformRecoveryXLogAction(void)
6395 : {
6396 163 : bool promoted = false;
6397 :
6398 : /*
6399 : * Perform a checkpoint to update all our recovery activity to disk.
6400 : *
6401 : * Note that we write a shutdown checkpoint rather than an on-line one.
6402 : * This is not particularly critical, but since we may be assigning a new
6403 : * TLI, using a shutdown checkpoint allows us to have the rule that TLI
6404 : * only changes in shutdown checkpoints, which allows some extra error
6405 : * checking in xlog_redo.
6406 : *
6407 : * In promotion, only create a lightweight end-of-recovery record instead
6408 : * of a full checkpoint. A checkpoint is requested later, after we're
6409 : * fully out of recovery mode and already accepting queries.
6410 : */
6411 218 : if (ArchiveRecoveryRequested && IsUnderPostmaster &&
6412 55 : PromoteIsTriggered())
6413 : {
6414 48 : promoted = true;
6415 :
6416 : /*
6417 : * Insert a special WAL record to mark the end of recovery, since we
6418 : * aren't doing a checkpoint. That means that the checkpointer process
6419 : * may likely be in the middle of a time-smoothed restartpoint and
6420 : * could continue to be for minutes after this. That sounds strange,
6421 : * but the effect is roughly the same and it would be stranger to try
6422 : * to come out of the restartpoint and then checkpoint. We request a
6423 : * checkpoint later anyway, just for safety.
6424 : */
6425 48 : CreateEndOfRecoveryRecord();
6426 : }
6427 : else
6428 : {
6429 115 : RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
6430 : CHECKPOINT_FAST |
6431 : CHECKPOINT_WAIT);
6432 : }
6433 :
6434 163 : return promoted;
6435 : }
6436 :
6437 : /*
6438 : * Is the system still in recovery?
6439 : *
6440 : * Unlike testing InRecovery, this works in any process that's connected to
6441 : * shared memory.
6442 : */
6443 : bool
6444 96475915 : RecoveryInProgress(void)
6445 : {
6446 : /*
6447 : * We check shared state each time only until we leave recovery mode. We
6448 : * can't re-enter recovery, so there's no need to keep checking after the
6449 : * shared variable has once been seen false.
6450 : */
6451 96475915 : if (!LocalRecoveryInProgress)
6452 94548825 : return false;
6453 : else
6454 : {
6455 : /*
6456 : * use volatile pointer to make sure we make a fresh read of the
6457 : * shared variable.
6458 : */
6459 1927090 : volatile XLogCtlData *xlogctl = XLogCtl;
6460 :
6461 1927090 : LocalRecoveryInProgress = (xlogctl->SharedRecoveryState != RECOVERY_STATE_DONE);
6462 :
6463 : /*
6464 : * Note: We don't need a memory barrier when we're still in recovery.
6465 : * We might exit recovery immediately after return, so the caller
6466 : * can't rely on 'true' meaning that we're still in recovery anyway.
6467 : */
6468 :
6469 1927090 : return LocalRecoveryInProgress;
6470 : }
6471 : }
6472 :
6473 : /*
6474 : * Returns current recovery state from shared memory.
6475 : *
6476 : * This returned state is kept consistent with the contents of the control
6477 : * file. See details about the possible values of RecoveryState in xlog.h.
6478 : */
6479 : RecoveryState
6480 2530 : GetRecoveryState(void)
6481 : {
6482 : RecoveryState retval;
6483 :
6484 2530 : SpinLockAcquire(&XLogCtl->info_lck);
6485 2530 : retval = XLogCtl->SharedRecoveryState;
6486 2530 : SpinLockRelease(&XLogCtl->info_lck);
6487 :
6488 2530 : return retval;
6489 : }
6490 :
6491 : /*
6492 : * Is this process allowed to insert new WAL records?
6493 : *
6494 : * Ordinarily this is essentially equivalent to !RecoveryInProgress().
6495 : * But we also have provisions for forcing the result "true" or "false"
6496 : * within specific processes regardless of the global state.
6497 : */
6498 : bool
6499 48562100 : XLogInsertAllowed(void)
6500 : {
6501 : /*
6502 : * If value is "unconditionally true" or "unconditionally false", just
6503 : * return it. This provides the normal fast path once recovery is known
6504 : * done.
6505 : */
6506 48562100 : if (LocalXLogInsertAllowed >= 0)
6507 47942671 : return (bool) LocalXLogInsertAllowed;
6508 :
6509 : /*
6510 : * Else, must check to see if we're still in recovery.
6511 : */
6512 619429 : if (RecoveryInProgress())
6513 609099 : return false;
6514 :
6515 : /*
6516 : * On exit from recovery, reset to "unconditionally true", since there is
6517 : * no need to keep checking.
6518 : */
6519 10330 : LocalXLogInsertAllowed = 1;
6520 10330 : return true;
6521 : }
6522 :
6523 : /*
6524 : * Make XLogInsertAllowed() return true in the current process only.
6525 : *
6526 : * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
6527 : * and even call LocalSetXLogInsertAllowed() again after that.
6528 : *
6529 : * Returns the previous value of LocalXLogInsertAllowed.
6530 : */
6531 : static int
6532 1000 : LocalSetXLogInsertAllowed(void)
6533 : {
6534 1000 : int oldXLogAllowed = LocalXLogInsertAllowed;
6535 :
6536 1000 : LocalXLogInsertAllowed = 1;
6537 :
6538 1000 : return oldXLogAllowed;
6539 : }
6540 :
6541 : /*
6542 : * Return the current Redo pointer from shared memory.
6543 : *
6544 : * As a side-effect, the local RedoRecPtr copy is updated.
6545 : */
6546 : XLogRecPtr
6547 360808 : GetRedoRecPtr(void)
6548 : {
6549 : XLogRecPtr ptr;
6550 :
6551 : /*
6552 : * The possibly not up-to-date copy in XlogCtl is enough. Even if we
6553 : * grabbed a WAL insertion lock to read the authoritative value in
6554 : * Insert->RedoRecPtr, someone might update it just after we've released
6555 : * the lock.
6556 : */
6557 360808 : SpinLockAcquire(&XLogCtl->info_lck);
6558 360808 : ptr = XLogCtl->RedoRecPtr;
6559 360808 : SpinLockRelease(&XLogCtl->info_lck);
6560 :
6561 360808 : if (RedoRecPtr < ptr)
6562 1721 : RedoRecPtr = ptr;
6563 :
6564 360808 : return RedoRecPtr;
6565 : }
6566 :
6567 : /*
6568 : * Return information needed to decide whether a modified block needs a
6569 : * full-page image to be included in the WAL record.
6570 : *
6571 : * The returned values are cached copies from backend-private memory, and
6572 : * possibly out-of-date or, indeed, uninitialized, in which case they will
6573 : * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6574 : * re-check them against up-to-date values, while holding the WAL insert lock.
6575 : */
6576 : void
6577 18432917 : GetFullPageWriteInfo(XLogRecPtr *RedoRecPtr_p, bool *doPageWrites_p)
6578 : {
6579 18432917 : *RedoRecPtr_p = RedoRecPtr;
6580 18432917 : *doPageWrites_p = doPageWrites;
6581 18432917 : }
6582 :
6583 : /*
6584 : * GetInsertRecPtr -- Returns the current insert position.
6585 : *
6586 : * NOTE: The value *actually* returned is the position of the last full
6587 : * xlog page. It lags behind the real insert position by at most 1 page.
6588 : * For that, we don't need to scan through WAL insertion locks, and an
6589 : * approximation is enough for the current usage of this function.
6590 : */
6591 : XLogRecPtr
6592 6893 : GetInsertRecPtr(void)
6593 : {
6594 : XLogRecPtr recptr;
6595 :
6596 6893 : SpinLockAcquire(&XLogCtl->info_lck);
6597 6893 : recptr = XLogCtl->LogwrtRqst.Write;
6598 6893 : SpinLockRelease(&XLogCtl->info_lck);
6599 :
6600 6893 : return recptr;
6601 : }
6602 :
6603 : /*
6604 : * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6605 : * position known to be fsync'd to disk. This should only be used on a
6606 : * system that is known not to be in recovery.
6607 : */
6608 : XLogRecPtr
6609 214873 : GetFlushRecPtr(TimeLineID *insertTLI)
6610 : {
6611 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6612 :
6613 214873 : RefreshXLogWriteResult(LogwrtResult);
6614 :
6615 : /*
6616 : * If we're writing and flushing WAL, the time line can't be changing, so
6617 : * no lock is required.
6618 : */
6619 214873 : if (insertTLI)
6620 24283 : *insertTLI = XLogCtl->InsertTimeLineID;
6621 :
6622 214873 : return LogwrtResult.Flush;
6623 : }
6624 :
6625 : /*
6626 : * GetWALInsertionTimeLine -- Returns the current timeline of a system that
6627 : * is not in recovery.
6628 : */
6629 : TimeLineID
6630 116710 : GetWALInsertionTimeLine(void)
6631 : {
6632 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6633 :
6634 : /* Since the value can't be changing, no lock is required. */
6635 116710 : return XLogCtl->InsertTimeLineID;
6636 : }
6637 :
6638 : /*
6639 : * GetWALInsertionTimeLineIfSet -- If the system is not in recovery, returns
6640 : * the WAL insertion timeline; else, returns 0. Wherever possible, use
6641 : * GetWALInsertionTimeLine() instead, since it's cheaper. Note that this
6642 : * function decides recovery has ended as soon as the insert TLI is set, which
6643 : * happens before we set XLogCtl->SharedRecoveryState to RECOVERY_STATE_DONE.
6644 : */
6645 : TimeLineID
6646 1 : GetWALInsertionTimeLineIfSet(void)
6647 : {
6648 : TimeLineID insertTLI;
6649 :
6650 1 : SpinLockAcquire(&XLogCtl->info_lck);
6651 1 : insertTLI = XLogCtl->InsertTimeLineID;
6652 1 : SpinLockRelease(&XLogCtl->info_lck);
6653 :
6654 1 : return insertTLI;
6655 : }
6656 :
6657 : /*
6658 : * GetLastImportantRecPtr -- Returns the LSN of the last important record
6659 : * inserted. All records not explicitly marked as unimportant are considered
6660 : * important.
6661 : *
6662 : * The LSN is determined by computing the maximum of
6663 : * WALInsertLocks[i].lastImportantAt.
6664 : */
6665 : XLogRecPtr
6666 1684 : GetLastImportantRecPtr(void)
6667 : {
6668 1684 : XLogRecPtr res = InvalidXLogRecPtr;
6669 : int i;
6670 :
6671 15156 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
6672 : {
6673 : XLogRecPtr last_important;
6674 :
6675 : /*
6676 : * Need to take a lock to prevent torn reads of the LSN, which are
6677 : * possible on some of the supported platforms. WAL insert locks only
6678 : * support exclusive mode, so we have to use that.
6679 : */
6680 13472 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
6681 13472 : last_important = WALInsertLocks[i].l.lastImportantAt;
6682 13472 : LWLockRelease(&WALInsertLocks[i].l.lock);
6683 :
6684 13472 : if (res < last_important)
6685 2910 : res = last_important;
6686 : }
6687 :
6688 1684 : return res;
6689 : }
6690 :
6691 : /*
6692 : * Get the time and LSN of the last xlog segment switch
6693 : */
6694 : pg_time_t
6695 0 : GetLastSegSwitchData(XLogRecPtr *lastSwitchLSN)
6696 : {
6697 : pg_time_t result;
6698 :
6699 : /* Need WALWriteLock, but shared lock is sufficient */
6700 0 : LWLockAcquire(WALWriteLock, LW_SHARED);
6701 0 : result = XLogCtl->lastSegSwitchTime;
6702 0 : *lastSwitchLSN = XLogCtl->lastSegSwitchLSN;
6703 0 : LWLockRelease(WALWriteLock);
6704 :
6705 0 : return result;
6706 : }
6707 :
6708 : /*
6709 : * This must be called ONCE during postmaster or standalone-backend shutdown
6710 : */
6711 : void
6712 721 : ShutdownXLOG(int code, Datum arg)
6713 : {
6714 : /*
6715 : * We should have an aux process resource owner to use, and we should not
6716 : * be in a transaction that's installed some other resowner.
6717 : */
6718 : Assert(AuxProcessResourceOwner != NULL);
6719 : Assert(CurrentResourceOwner == NULL ||
6720 : CurrentResourceOwner == AuxProcessResourceOwner);
6721 721 : CurrentResourceOwner = AuxProcessResourceOwner;
6722 :
6723 : /* Don't be chatty in standalone mode */
6724 721 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
6725 : (errmsg("shutting down")));
6726 :
6727 : /*
6728 : * Signal walsenders to move to stopping state.
6729 : */
6730 721 : WalSndInitStopping();
6731 :
6732 : /*
6733 : * Wait for WAL senders to be in stopping state. This prevents commands
6734 : * from writing new WAL.
6735 : */
6736 721 : WalSndWaitStopping();
6737 :
6738 721 : if (RecoveryInProgress())
6739 58 : CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
6740 : else
6741 : {
6742 : /*
6743 : * If archiving is enabled, rotate the last XLOG file so that all the
6744 : * remaining records are archived (postmaster wakes up the archiver
6745 : * process one more time at the end of shutdown). The checkpoint
6746 : * record will go to the next XLOG file and won't be archived (yet).
6747 : */
6748 663 : if (XLogArchivingActive())
6749 16 : RequestXLogSwitch(false);
6750 :
6751 663 : CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
6752 : }
6753 721 : }
6754 :
6755 : /*
6756 : * Format checkpoint request flags as a space-separated string for
6757 : * log messages.
6758 : */
6759 : static const char *
6760 3074 : CheckpointFlagsString(int flags)
6761 : {
6762 : static char buf[128];
6763 :
6764 24592 : snprintf(buf, sizeof(buf), "%s%s%s%s%s%s%s%s",
6765 3074 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6766 3074 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6767 3074 : (flags & CHECKPOINT_FAST) ? " fast" : "",
6768 3074 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6769 3074 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6770 3074 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6771 3074 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6772 3074 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "");
6773 :
6774 3074 : return buf;
6775 : }
6776 :
6777 : /*
6778 : * Log start of a checkpoint.
6779 : */
6780 : static void
6781 1537 : LogCheckpointStart(int flags, bool restartpoint)
6782 : {
6783 1537 : if (restartpoint)
6784 205 : ereport(LOG,
6785 : /* translator: the placeholder shows checkpoint options */
6786 : (errmsg("restartpoint starting:%s",
6787 : CheckpointFlagsString(flags))));
6788 : else
6789 1332 : ereport(LOG,
6790 : /* translator: the placeholder shows checkpoint options */
6791 : (errmsg("checkpoint starting:%s",
6792 : CheckpointFlagsString(flags))));
6793 1537 : }
6794 :
6795 : /*
6796 : * Log end of a checkpoint.
6797 : */
6798 : static void
6799 1835 : LogCheckpointEnd(bool restartpoint, int flags)
6800 : {
6801 : long write_msecs,
6802 : sync_msecs,
6803 : total_msecs,
6804 : longest_msecs,
6805 : average_msecs;
6806 : uint64 average_sync_time;
6807 :
6808 1835 : CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
6809 :
6810 1835 : write_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_write_t,
6811 : CheckpointStats.ckpt_sync_t);
6812 :
6813 1835 : sync_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_sync_t,
6814 : CheckpointStats.ckpt_sync_end_t);
6815 :
6816 : /* Accumulate checkpoint timing summary data, in milliseconds. */
6817 1835 : PendingCheckpointerStats.write_time += write_msecs;
6818 1835 : PendingCheckpointerStats.sync_time += sync_msecs;
6819 :
6820 : /*
6821 : * All of the published timing statistics are accounted for. Only
6822 : * continue if a log message is to be written.
6823 : */
6824 1835 : if (!log_checkpoints)
6825 298 : return;
6826 :
6827 1537 : total_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_start_t,
6828 : CheckpointStats.ckpt_end_t);
6829 :
6830 : /*
6831 : * Timing values returned from CheckpointStats are in microseconds.
6832 : * Convert to milliseconds for consistent printing.
6833 : */
6834 1537 : longest_msecs = (long) ((CheckpointStats.ckpt_longest_sync + 999) / 1000);
6835 :
6836 1537 : average_sync_time = 0;
6837 1537 : if (CheckpointStats.ckpt_sync_rels > 0)
6838 0 : average_sync_time = CheckpointStats.ckpt_agg_sync_time /
6839 0 : CheckpointStats.ckpt_sync_rels;
6840 1537 : average_msecs = (long) ((average_sync_time + 999) / 1000);
6841 :
6842 : /*
6843 : * ControlFileLock is not required to see ControlFile->checkPoint and
6844 : * ->checkPointCopy here as we are the only updator of those variables at
6845 : * this moment.
6846 : */
6847 1537 : if (restartpoint)
6848 205 : ereport(LOG,
6849 : (errmsg("restartpoint complete:%s: wrote %d buffers (%.1f%%), "
6850 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6851 : "%d removed, %d recycled; write=%ld.%03d s, "
6852 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6853 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6854 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6855 : CheckpointFlagsString(flags),
6856 : CheckpointStats.ckpt_bufs_written,
6857 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6858 : CheckpointStats.ckpt_slru_written,
6859 : CheckpointStats.ckpt_segs_added,
6860 : CheckpointStats.ckpt_segs_removed,
6861 : CheckpointStats.ckpt_segs_recycled,
6862 : write_msecs / 1000, (int) (write_msecs % 1000),
6863 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6864 : total_msecs / 1000, (int) (total_msecs % 1000),
6865 : CheckpointStats.ckpt_sync_rels,
6866 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6867 : average_msecs / 1000, (int) (average_msecs % 1000),
6868 : (int) (PrevCheckPointDistance / 1024.0),
6869 : (int) (CheckPointDistanceEstimate / 1024.0),
6870 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6871 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6872 : else
6873 1332 : ereport(LOG,
6874 : (errmsg("checkpoint complete:%s: wrote %d buffers (%.1f%%), "
6875 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6876 : "%d removed, %d recycled; write=%ld.%03d s, "
6877 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6878 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6879 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6880 : CheckpointFlagsString(flags),
6881 : CheckpointStats.ckpt_bufs_written,
6882 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6883 : CheckpointStats.ckpt_slru_written,
6884 : CheckpointStats.ckpt_segs_added,
6885 : CheckpointStats.ckpt_segs_removed,
6886 : CheckpointStats.ckpt_segs_recycled,
6887 : write_msecs / 1000, (int) (write_msecs % 1000),
6888 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6889 : total_msecs / 1000, (int) (total_msecs % 1000),
6890 : CheckpointStats.ckpt_sync_rels,
6891 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6892 : average_msecs / 1000, (int) (average_msecs % 1000),
6893 : (int) (PrevCheckPointDistance / 1024.0),
6894 : (int) (CheckPointDistanceEstimate / 1024.0),
6895 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6896 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6897 : }
6898 :
6899 : /*
6900 : * Update the estimate of distance between checkpoints.
6901 : *
6902 : * The estimate is used to calculate the number of WAL segments to keep
6903 : * preallocated, see XLOGfileslop().
6904 : */
6905 : static void
6906 1835 : UpdateCheckPointDistanceEstimate(uint64 nbytes)
6907 : {
6908 : /*
6909 : * To estimate the number of segments consumed between checkpoints, keep a
6910 : * moving average of the amount of WAL generated in previous checkpoint
6911 : * cycles. However, if the load is bursty, with quiet periods and busy
6912 : * periods, we want to cater for the peak load. So instead of a plain
6913 : * moving average, let the average decline slowly if the previous cycle
6914 : * used less WAL than estimated, but bump it up immediately if it used
6915 : * more.
6916 : *
6917 : * When checkpoints are triggered by max_wal_size, this should converge to
6918 : * CheckpointSegments * wal_segment_size,
6919 : *
6920 : * Note: This doesn't pay any attention to what caused the checkpoint.
6921 : * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
6922 : * starting a base backup, are counted the same as those created
6923 : * automatically. The slow-decline will largely mask them out, if they are
6924 : * not frequent. If they are frequent, it seems reasonable to count them
6925 : * in as any others; if you issue a manual checkpoint every 5 minutes and
6926 : * never let a timed checkpoint happen, it makes sense to base the
6927 : * preallocation on that 5 minute interval rather than whatever
6928 : * checkpoint_timeout is set to.
6929 : */
6930 1835 : PrevCheckPointDistance = nbytes;
6931 1835 : if (CheckPointDistanceEstimate < nbytes)
6932 809 : CheckPointDistanceEstimate = nbytes;
6933 : else
6934 1026 : CheckPointDistanceEstimate =
6935 1026 : (0.90 * CheckPointDistanceEstimate + 0.10 * (double) nbytes);
6936 1835 : }
6937 :
6938 : /*
6939 : * Update the ps display for a process running a checkpoint. Note that
6940 : * this routine should not do any allocations so as it can be called
6941 : * from a critical section.
6942 : */
6943 : static void
6944 3670 : update_checkpoint_display(int flags, bool restartpoint, bool reset)
6945 : {
6946 : /*
6947 : * The status is reported only for end-of-recovery and shutdown
6948 : * checkpoints or shutdown restartpoints. Updating the ps display is
6949 : * useful in those situations as it may not be possible to rely on
6950 : * pg_stat_activity to see the status of the checkpointer or the startup
6951 : * process.
6952 : */
6953 3670 : if ((flags & (CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IS_SHUTDOWN)) == 0)
6954 2240 : return;
6955 :
6956 1430 : if (reset)
6957 715 : set_ps_display("");
6958 : else
6959 : {
6960 : char activitymsg[128];
6961 :
6962 2145 : snprintf(activitymsg, sizeof(activitymsg), "performing %s%s%s",
6963 715 : (flags & CHECKPOINT_END_OF_RECOVERY) ? "end-of-recovery " : "",
6964 715 : (flags & CHECKPOINT_IS_SHUTDOWN) ? "shutdown " : "",
6965 : restartpoint ? "restartpoint" : "checkpoint");
6966 715 : set_ps_display(activitymsg);
6967 : }
6968 : }
6969 :
6970 :
6971 : /*
6972 : * Perform a checkpoint --- either during shutdown, or on-the-fly
6973 : *
6974 : * flags is a bitwise OR of the following:
6975 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
6976 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
6977 : * CHECKPOINT_FAST: finish the checkpoint ASAP, ignoring
6978 : * checkpoint_completion_target parameter.
6979 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
6980 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
6981 : * CHECKPOINT_END_OF_RECOVERY).
6982 : * CHECKPOINT_FLUSH_UNLOGGED: also flush buffers of unlogged tables.
6983 : *
6984 : * Note: flags contains other bits, of interest here only for logging purposes.
6985 : * In particular note that this routine is synchronous and does not pay
6986 : * attention to CHECKPOINT_WAIT.
6987 : *
6988 : * If !shutdown then we are writing an online checkpoint. An XLOG_CHECKPOINT_REDO
6989 : * record is inserted into WAL at the logical location of the checkpoint, before
6990 : * flushing anything to disk, and when the checkpoint is eventually completed,
6991 : * and it is from this point that WAL replay will begin in the case of a recovery
6992 : * from this checkpoint. Once everything is written to disk, an
6993 : * XLOG_CHECKPOINT_ONLINE record is written to complete the checkpoint, and
6994 : * points back to the earlier XLOG_CHECKPOINT_REDO record. This mechanism allows
6995 : * other write-ahead log records to be written while the checkpoint is in
6996 : * progress, but we must be very careful about order of operations. This function
6997 : * may take many minutes to execute on a busy system.
6998 : *
6999 : * On the other hand, when shutdown is true, concurrent insertion into the
7000 : * write-ahead log is impossible, so there is no need for two separate records.
7001 : * In this case, we only insert an XLOG_CHECKPOINT_SHUTDOWN record, and it's
7002 : * both the record marking the completion of the checkpoint and the location
7003 : * from which WAL replay would begin if needed.
7004 : *
7005 : * Returns true if a new checkpoint was performed, or false if it was skipped
7006 : * because the system was idle.
7007 : */
7008 : bool
7009 1631 : CreateCheckPoint(int flags)
7010 : {
7011 : bool shutdown;
7012 : CheckPoint checkPoint;
7013 : XLogRecPtr recptr;
7014 : XLogSegNo _logSegNo;
7015 1631 : XLogCtlInsert *Insert = &XLogCtl->Insert;
7016 : uint32 freespace;
7017 : XLogRecPtr PriorRedoPtr;
7018 : XLogRecPtr last_important_lsn;
7019 : VirtualTransactionId *vxids;
7020 : int nvxids;
7021 1631 : int oldXLogAllowed = 0;
7022 :
7023 : /*
7024 : * An end-of-recovery checkpoint is really a shutdown checkpoint, just
7025 : * issued at a different time.
7026 : */
7027 1631 : if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
7028 693 : shutdown = true;
7029 : else
7030 938 : shutdown = false;
7031 :
7032 : /* sanity check */
7033 1631 : if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
7034 0 : elog(ERROR, "can't create a checkpoint during recovery");
7035 :
7036 : /*
7037 : * Prepare to accumulate statistics.
7038 : *
7039 : * Note: because it is possible for log_checkpoints to change while a
7040 : * checkpoint proceeds, we always accumulate stats, even if
7041 : * log_checkpoints is currently off.
7042 : */
7043 17941 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7044 1631 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7045 :
7046 : /*
7047 : * Let smgr prepare for checkpoint; this has to happen outside the
7048 : * critical section and before we determine the REDO pointer. Note that
7049 : * smgr must not do anything that'd have to be undone if we decide no
7050 : * checkpoint is needed.
7051 : */
7052 1631 : SyncPreCheckpoint();
7053 :
7054 : /* Run these points outside the critical section. */
7055 1631 : INJECTION_POINT("create-checkpoint-initial", NULL);
7056 1631 : INJECTION_POINT_LOAD("create-checkpoint-run");
7057 :
7058 : /*
7059 : * Use a critical section to force system panic if we have trouble.
7060 : */
7061 1631 : START_CRIT_SECTION();
7062 :
7063 1631 : if (shutdown)
7064 : {
7065 693 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7066 693 : ControlFile->state = DB_SHUTDOWNING;
7067 693 : UpdateControlFile();
7068 693 : LWLockRelease(ControlFileLock);
7069 : }
7070 :
7071 : /* Begin filling in the checkpoint WAL record */
7072 21203 : MemSet(&checkPoint, 0, sizeof(checkPoint));
7073 1631 : checkPoint.time = (pg_time_t) time(NULL);
7074 :
7075 : /*
7076 : * For Hot Standby, derive the oldestActiveXid before we fix the redo
7077 : * pointer. This allows us to begin accumulating changes to assemble our
7078 : * starting snapshot of locks and transactions.
7079 : */
7080 1631 : if (!shutdown && XLogStandbyInfoActive())
7081 907 : checkPoint.oldestActiveXid = GetOldestActiveTransactionId(false, true);
7082 : else
7083 724 : checkPoint.oldestActiveXid = InvalidTransactionId;
7084 :
7085 : /*
7086 : * Get location of last important record before acquiring insert locks (as
7087 : * GetLastImportantRecPtr() also locks WAL locks).
7088 : */
7089 1631 : last_important_lsn = GetLastImportantRecPtr();
7090 :
7091 : /*
7092 : * If this isn't a shutdown or forced checkpoint, and if there has been no
7093 : * WAL activity requiring a checkpoint, skip it. The idea here is to
7094 : * avoid inserting duplicate checkpoints when the system is idle.
7095 : */
7096 1631 : if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
7097 : CHECKPOINT_FORCE)) == 0)
7098 : {
7099 197 : if (last_important_lsn == ControlFile->checkPoint)
7100 : {
7101 1 : END_CRIT_SECTION();
7102 1 : ereport(DEBUG1,
7103 : (errmsg_internal("checkpoint skipped because system is idle")));
7104 1 : return false;
7105 : }
7106 : }
7107 :
7108 : /*
7109 : * An end-of-recovery checkpoint is created before anyone is allowed to
7110 : * write WAL. To allow us to write the checkpoint record, temporarily
7111 : * enable XLogInsertAllowed.
7112 : */
7113 1630 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7114 30 : oldXLogAllowed = LocalSetXLogInsertAllowed();
7115 :
7116 1630 : checkPoint.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7117 1630 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7118 30 : checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7119 : else
7120 1600 : checkPoint.PrevTimeLineID = checkPoint.ThisTimeLineID;
7121 :
7122 : /*
7123 : * We must block concurrent insertions while examining insert state.
7124 : */
7125 1630 : WALInsertLockAcquireExclusive();
7126 :
7127 1630 : checkPoint.fullPageWrites = Insert->fullPageWrites;
7128 1630 : checkPoint.wal_level = wal_level;
7129 :
7130 1630 : if (shutdown)
7131 : {
7132 693 : XLogRecPtr curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
7133 :
7134 : /*
7135 : * Compute new REDO record ptr = location of next XLOG record.
7136 : *
7137 : * Since this is a shutdown checkpoint, there can't be any concurrent
7138 : * WAL insertion.
7139 : */
7140 693 : freespace = INSERT_FREESPACE(curInsert);
7141 693 : if (freespace == 0)
7142 : {
7143 0 : if (XLogSegmentOffset(curInsert, wal_segment_size) == 0)
7144 0 : curInsert += SizeOfXLogLongPHD;
7145 : else
7146 0 : curInsert += SizeOfXLogShortPHD;
7147 : }
7148 693 : checkPoint.redo = curInsert;
7149 :
7150 : /*
7151 : * Here we update the shared RedoRecPtr for future XLogInsert calls;
7152 : * this must be done while holding all the insertion locks.
7153 : *
7154 : * Note: if we fail to complete the checkpoint, RedoRecPtr will be
7155 : * left pointing past where it really needs to point. This is okay;
7156 : * the only consequence is that XLogInsert might back up whole buffers
7157 : * that it didn't really need to. We can't postpone advancing
7158 : * RedoRecPtr because XLogInserts that happen while we are dumping
7159 : * buffers must assume that their buffer changes are not included in
7160 : * the checkpoint.
7161 : */
7162 693 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
7163 : }
7164 :
7165 : /*
7166 : * Now we can release the WAL insertion locks, allowing other xacts to
7167 : * proceed while we are flushing disk buffers.
7168 : */
7169 1630 : WALInsertLockRelease();
7170 :
7171 : /*
7172 : * If this is an online checkpoint, we have not yet determined the redo
7173 : * point. We do so now by inserting the special XLOG_CHECKPOINT_REDO
7174 : * record; the LSN at which it starts becomes the new redo pointer. We
7175 : * don't do this for a shutdown checkpoint, because in that case no WAL
7176 : * can be written between the redo point and the insertion of the
7177 : * checkpoint record itself, so the checkpoint record itself serves to
7178 : * mark the redo point.
7179 : */
7180 1630 : if (!shutdown)
7181 : {
7182 : /* Include WAL level in record for WAL summarizer's benefit. */
7183 937 : XLogBeginInsert();
7184 937 : XLogRegisterData(&wal_level, sizeof(wal_level));
7185 937 : (void) XLogInsert(RM_XLOG_ID, XLOG_CHECKPOINT_REDO);
7186 :
7187 : /*
7188 : * XLogInsertRecord will have updated XLogCtl->Insert.RedoRecPtr in
7189 : * shared memory and RedoRecPtr in backend-local memory, but we need
7190 : * to copy that into the record that will be inserted when the
7191 : * checkpoint is complete.
7192 : */
7193 937 : checkPoint.redo = RedoRecPtr;
7194 : }
7195 :
7196 : /* Update the info_lck-protected copy of RedoRecPtr as well */
7197 1630 : SpinLockAcquire(&XLogCtl->info_lck);
7198 1630 : XLogCtl->RedoRecPtr = checkPoint.redo;
7199 1630 : SpinLockRelease(&XLogCtl->info_lck);
7200 :
7201 : /*
7202 : * If enabled, log checkpoint start. We postpone this until now so as not
7203 : * to log anything if we decided to skip the checkpoint.
7204 : */
7205 1630 : if (log_checkpoints)
7206 1332 : LogCheckpointStart(flags, false);
7207 :
7208 1630 : INJECTION_POINT_CACHED("create-checkpoint-run", NULL);
7209 :
7210 : /* Update the process title */
7211 1630 : update_checkpoint_display(flags, false, false);
7212 :
7213 : TRACE_POSTGRESQL_CHECKPOINT_START(flags);
7214 :
7215 : /*
7216 : * Get the other info we need for the checkpoint record.
7217 : *
7218 : * We don't need to save oldestClogXid in the checkpoint, it only matters
7219 : * for the short period in which clog is being truncated, and if we crash
7220 : * during that we'll redo the clog truncation and fix up oldestClogXid
7221 : * there.
7222 : */
7223 1630 : LWLockAcquire(XidGenLock, LW_SHARED);
7224 1630 : checkPoint.nextXid = TransamVariables->nextXid;
7225 1630 : checkPoint.oldestXid = TransamVariables->oldestXid;
7226 1630 : checkPoint.oldestXidDB = TransamVariables->oldestXidDB;
7227 1630 : LWLockRelease(XidGenLock);
7228 :
7229 1630 : LWLockAcquire(CommitTsLock, LW_SHARED);
7230 1630 : checkPoint.oldestCommitTsXid = TransamVariables->oldestCommitTsXid;
7231 1630 : checkPoint.newestCommitTsXid = TransamVariables->newestCommitTsXid;
7232 1630 : LWLockRelease(CommitTsLock);
7233 :
7234 1630 : LWLockAcquire(OidGenLock, LW_SHARED);
7235 1630 : checkPoint.nextOid = TransamVariables->nextOid;
7236 1630 : if (!shutdown)
7237 937 : checkPoint.nextOid += TransamVariables->oidCount;
7238 1630 : LWLockRelease(OidGenLock);
7239 :
7240 1630 : checkPoint.logicalDecodingEnabled = IsLogicalDecodingEnabled();
7241 :
7242 1630 : MultiXactGetCheckptMulti(shutdown,
7243 : &checkPoint.nextMulti,
7244 : &checkPoint.nextMultiOffset,
7245 : &checkPoint.oldestMulti,
7246 : &checkPoint.oldestMultiDB);
7247 :
7248 : /*
7249 : * Having constructed the checkpoint record, ensure all shmem disk buffers
7250 : * and commit-log buffers are flushed to disk.
7251 : *
7252 : * This I/O could fail for various reasons. If so, we will fail to
7253 : * complete the checkpoint, but there is no reason to force a system
7254 : * panic. Accordingly, exit critical section while doing it.
7255 : */
7256 1630 : END_CRIT_SECTION();
7257 :
7258 : /*
7259 : * In some cases there are groups of actions that must all occur on one
7260 : * side or the other of a checkpoint record. Before flushing the
7261 : * checkpoint record we must explicitly wait for any backend currently
7262 : * performing those groups of actions.
7263 : *
7264 : * One example is end of transaction, so we must wait for any transactions
7265 : * that are currently in commit critical sections. If an xact inserted
7266 : * its commit record into XLOG just before the REDO point, then a crash
7267 : * restart from the REDO point would not replay that record, which means
7268 : * that our flushing had better include the xact's update of pg_xact. So
7269 : * we wait till he's out of his commit critical section before proceeding.
7270 : * See notes in RecordTransactionCommit().
7271 : *
7272 : * Because we've already released the insertion locks, this test is a bit
7273 : * fuzzy: it is possible that we will wait for xacts we didn't really need
7274 : * to wait for. But the delay should be short and it seems better to make
7275 : * checkpoint take a bit longer than to hold off insertions longer than
7276 : * necessary. (In fact, the whole reason we have this issue is that xact.c
7277 : * does commit record XLOG insertion and clog update as two separate steps
7278 : * protected by different locks, but again that seems best on grounds of
7279 : * minimizing lock contention.)
7280 : *
7281 : * A transaction that has not yet set delayChkptFlags when we look cannot
7282 : * be at risk, since it has not inserted its commit record yet; and one
7283 : * that's already cleared it is not at risk either, since it's done fixing
7284 : * clog and we will correctly flush the update below. So we cannot miss
7285 : * any xacts we need to wait for.
7286 : */
7287 1630 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
7288 1630 : if (nvxids > 0)
7289 : {
7290 : do
7291 : {
7292 : /*
7293 : * Keep absorbing fsync requests while we wait. There could even
7294 : * be a deadlock if we don't, if the process that prevents the
7295 : * checkpoint is trying to add a request to the queue.
7296 : */
7297 29 : AbsorbSyncRequests();
7298 :
7299 29 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
7300 29 : pg_usleep(10000L); /* wait for 10 msec */
7301 29 : pgstat_report_wait_end();
7302 29 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7303 : DELAY_CHKPT_START));
7304 : }
7305 1630 : pfree(vxids);
7306 :
7307 1630 : CheckPointGuts(checkPoint.redo, flags);
7308 :
7309 1630 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
7310 1630 : if (nvxids > 0)
7311 : {
7312 : do
7313 : {
7314 0 : AbsorbSyncRequests();
7315 :
7316 0 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_COMPLETE);
7317 0 : pg_usleep(10000L); /* wait for 10 msec */
7318 0 : pgstat_report_wait_end();
7319 0 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7320 : DELAY_CHKPT_COMPLETE));
7321 : }
7322 1630 : pfree(vxids);
7323 :
7324 : /*
7325 : * Take a snapshot of running transactions and write this to WAL. This
7326 : * allows us to reconstruct the state of running transactions during
7327 : * archive recovery, if required. Skip, if this info disabled.
7328 : *
7329 : * If we are shutting down, or Startup process is completing crash
7330 : * recovery we don't need to write running xact data.
7331 : */
7332 1630 : if (!shutdown && XLogStandbyInfoActive())
7333 906 : LogStandbySnapshot();
7334 :
7335 1630 : START_CRIT_SECTION();
7336 :
7337 : /*
7338 : * Now insert the checkpoint record into XLOG.
7339 : */
7340 1630 : XLogBeginInsert();
7341 1630 : XLogRegisterData(&checkPoint, sizeof(checkPoint));
7342 1630 : recptr = XLogInsert(RM_XLOG_ID,
7343 : shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
7344 : XLOG_CHECKPOINT_ONLINE);
7345 :
7346 1630 : XLogFlush(recptr);
7347 :
7348 : /*
7349 : * We mustn't write any new WAL after a shutdown checkpoint, or it will be
7350 : * overwritten at next startup. No-one should even try, this just allows
7351 : * sanity-checking. In the case of an end-of-recovery checkpoint, we want
7352 : * to just temporarily disable writing until the system has exited
7353 : * recovery.
7354 : */
7355 1630 : if (shutdown)
7356 : {
7357 693 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7358 30 : LocalXLogInsertAllowed = oldXLogAllowed;
7359 : else
7360 663 : LocalXLogInsertAllowed = 0; /* never again write WAL */
7361 : }
7362 :
7363 : /*
7364 : * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
7365 : * = end of actual checkpoint record.
7366 : */
7367 1630 : if (shutdown && checkPoint.redo != ProcLastRecPtr)
7368 0 : ereport(PANIC,
7369 : (errmsg("concurrent write-ahead log activity while database system is shutting down")));
7370 :
7371 : /*
7372 : * Remember the prior checkpoint's redo ptr for
7373 : * UpdateCheckPointDistanceEstimate()
7374 : */
7375 1630 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7376 :
7377 : /*
7378 : * Update the control file.
7379 : */
7380 1630 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7381 1630 : if (shutdown)
7382 693 : ControlFile->state = DB_SHUTDOWNED;
7383 1630 : ControlFile->checkPoint = ProcLastRecPtr;
7384 1630 : ControlFile->checkPointCopy = checkPoint;
7385 : /* crash recovery should always recover to the end of WAL */
7386 1630 : ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
7387 1630 : ControlFile->minRecoveryPointTLI = 0;
7388 :
7389 : /*
7390 : * Persist unloggedLSN value. It's reset on crash recovery, so this goes
7391 : * unused on non-shutdown checkpoints, but seems useful to store it always
7392 : * for debugging purposes.
7393 : */
7394 1630 : ControlFile->unloggedLSN = pg_atomic_read_membarrier_u64(&XLogCtl->unloggedLSN);
7395 :
7396 1630 : UpdateControlFile();
7397 1630 : LWLockRelease(ControlFileLock);
7398 :
7399 : /*
7400 : * We are now done with critical updates; no need for system panic if we
7401 : * have trouble while fooling with old log segments.
7402 : */
7403 1630 : END_CRIT_SECTION();
7404 :
7405 : /*
7406 : * WAL summaries end when the next XLOG_CHECKPOINT_REDO or
7407 : * XLOG_CHECKPOINT_SHUTDOWN record is reached. This is the first point
7408 : * where (a) we're not inside of a critical section and (b) we can be
7409 : * certain that the relevant record has been flushed to disk, which must
7410 : * happen before it can be summarized.
7411 : *
7412 : * If this is a shutdown checkpoint, then this happens reasonably
7413 : * promptly: we've only just inserted and flushed the
7414 : * XLOG_CHECKPOINT_SHUTDOWN record. If this is not a shutdown checkpoint,
7415 : * then this might not be very prompt at all: the XLOG_CHECKPOINT_REDO
7416 : * record was written before we began flushing data to disk, and that
7417 : * could be many minutes ago at this point. However, we don't XLogFlush()
7418 : * after inserting that record, so we're not guaranteed that it's on disk
7419 : * until after the above call that flushes the XLOG_CHECKPOINT_ONLINE
7420 : * record.
7421 : */
7422 1630 : WakeupWalSummarizer();
7423 :
7424 : /*
7425 : * Let smgr do post-checkpoint cleanup (eg, deleting old files).
7426 : */
7427 1630 : SyncPostCheckpoint();
7428 :
7429 : /*
7430 : * Update the average distance between checkpoints if the prior checkpoint
7431 : * exists.
7432 : */
7433 1630 : if (XLogRecPtrIsValid(PriorRedoPtr))
7434 1630 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7435 :
7436 1630 : INJECTION_POINT("checkpoint-before-old-wal-removal", NULL);
7437 :
7438 : /*
7439 : * Delete old log files, those no longer needed for last checkpoint to
7440 : * prevent the disk holding the xlog from growing full.
7441 : */
7442 1630 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7443 1630 : KeepLogSeg(recptr, &_logSegNo);
7444 1630 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7445 : _logSegNo, InvalidOid,
7446 : InvalidTransactionId))
7447 : {
7448 : /*
7449 : * Some slots have been invalidated; recalculate the old-segment
7450 : * horizon, starting again from RedoRecPtr.
7451 : */
7452 4 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7453 4 : KeepLogSeg(recptr, &_logSegNo);
7454 : }
7455 1630 : _logSegNo--;
7456 1630 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
7457 : checkPoint.ThisTimeLineID);
7458 :
7459 : /*
7460 : * Make more log segments if needed. (Do this after recycling old log
7461 : * segments, since that may supply some of the needed files.)
7462 : */
7463 1630 : if (!shutdown)
7464 937 : PreallocXlogFiles(recptr, checkPoint.ThisTimeLineID);
7465 :
7466 : /*
7467 : * Truncate pg_subtrans if possible. We can throw away all data before
7468 : * the oldest XMIN of any running transaction. No future transaction will
7469 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7470 : * in subtrans.c). During recovery, though, we mustn't do this because
7471 : * StartupSUBTRANS hasn't been called yet.
7472 : */
7473 1630 : if (!RecoveryInProgress())
7474 1600 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7475 :
7476 : /* Real work is done; log and update stats. */
7477 1630 : LogCheckpointEnd(false, flags);
7478 :
7479 : /* Reset the process title */
7480 1630 : update_checkpoint_display(flags, false, true);
7481 :
7482 : TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
7483 : NBuffers,
7484 : CheckpointStats.ckpt_segs_added,
7485 : CheckpointStats.ckpt_segs_removed,
7486 : CheckpointStats.ckpt_segs_recycled);
7487 :
7488 1630 : return true;
7489 : }
7490 :
7491 : /*
7492 : * Mark the end of recovery in WAL though without running a full checkpoint.
7493 : * We can expect that a restartpoint is likely to be in progress as we
7494 : * do this, though we are unwilling to wait for it to complete.
7495 : *
7496 : * CreateRestartPoint() allows for the case where recovery may end before
7497 : * the restartpoint completes so there is no concern of concurrent behaviour.
7498 : */
7499 : static void
7500 48 : CreateEndOfRecoveryRecord(void)
7501 : {
7502 : xl_end_of_recovery xlrec;
7503 : XLogRecPtr recptr;
7504 :
7505 : /* sanity check */
7506 48 : if (!RecoveryInProgress())
7507 0 : elog(ERROR, "can only be used to end recovery");
7508 :
7509 48 : xlrec.end_time = GetCurrentTimestamp();
7510 48 : xlrec.wal_level = wal_level;
7511 :
7512 48 : WALInsertLockAcquireExclusive();
7513 48 : xlrec.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7514 48 : xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7515 48 : WALInsertLockRelease();
7516 :
7517 48 : START_CRIT_SECTION();
7518 :
7519 48 : XLogBeginInsert();
7520 48 : XLogRegisterData(&xlrec, sizeof(xl_end_of_recovery));
7521 48 : recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY);
7522 :
7523 48 : XLogFlush(recptr);
7524 :
7525 : /*
7526 : * Update the control file so that crash recovery can follow the timeline
7527 : * changes to this point.
7528 : */
7529 48 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7530 48 : ControlFile->minRecoveryPoint = recptr;
7531 48 : ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
7532 48 : UpdateControlFile();
7533 48 : LWLockRelease(ControlFileLock);
7534 :
7535 48 : END_CRIT_SECTION();
7536 48 : }
7537 :
7538 : /*
7539 : * Write an OVERWRITE_CONTRECORD message.
7540 : *
7541 : * When on WAL replay we expect a continuation record at the start of a page
7542 : * that is not there, recovery ends and WAL writing resumes at that point.
7543 : * But it's wrong to resume writing new WAL back at the start of the record
7544 : * that was broken, because downstream consumers of that WAL (physical
7545 : * replicas) are not prepared to "rewind". So the first action after
7546 : * finishing replay of all valid WAL must be to write a record of this type
7547 : * at the point where the contrecord was missing; to support xlogreader
7548 : * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
7549 : * to the page header where the record occurs. xlogreader has an ad-hoc
7550 : * mechanism to report metadata about the broken record, which is what we
7551 : * use here.
7552 : *
7553 : * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
7554 : * skip the record it was reading, and pass back the LSN of the skipped
7555 : * record, so that its caller can verify (on "replay" of that record) that the
7556 : * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
7557 : *
7558 : * 'aborted_lsn' is the beginning position of the record that was incomplete.
7559 : * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
7560 : * beginning of the XLOG page where the record is to be inserted. They must
7561 : * match the current WAL insert position, they're passed here just so that we
7562 : * can verify that.
7563 : */
7564 : static XLogRecPtr
7565 10 : CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn, XLogRecPtr pagePtr,
7566 : TimeLineID newTLI)
7567 : {
7568 : xl_overwrite_contrecord xlrec;
7569 : XLogRecPtr recptr;
7570 : XLogPageHeader pagehdr;
7571 : XLogRecPtr startPos;
7572 :
7573 : /* sanity checks */
7574 10 : if (!RecoveryInProgress())
7575 0 : elog(ERROR, "can only be used at end of recovery");
7576 10 : if (pagePtr % XLOG_BLCKSZ != 0)
7577 0 : elog(ERROR, "invalid position for missing continuation record %X/%08X",
7578 : LSN_FORMAT_ARGS(pagePtr));
7579 :
7580 : /* The current WAL insert position should be right after the page header */
7581 10 : startPos = pagePtr;
7582 10 : if (XLogSegmentOffset(startPos, wal_segment_size) == 0)
7583 1 : startPos += SizeOfXLogLongPHD;
7584 : else
7585 9 : startPos += SizeOfXLogShortPHD;
7586 10 : recptr = GetXLogInsertRecPtr();
7587 10 : if (recptr != startPos)
7588 0 : elog(ERROR, "invalid WAL insert position %X/%08X for OVERWRITE_CONTRECORD",
7589 : LSN_FORMAT_ARGS(recptr));
7590 :
7591 10 : START_CRIT_SECTION();
7592 :
7593 : /*
7594 : * Initialize the XLOG page header (by GetXLogBuffer), and set the
7595 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
7596 : *
7597 : * No other backend is allowed to write WAL yet, so acquiring the WAL
7598 : * insertion lock is just pro forma.
7599 : */
7600 10 : WALInsertLockAcquire();
7601 10 : pagehdr = (XLogPageHeader) GetXLogBuffer(pagePtr, newTLI);
7602 10 : pagehdr->xlp_info |= XLP_FIRST_IS_OVERWRITE_CONTRECORD;
7603 10 : WALInsertLockRelease();
7604 :
7605 : /*
7606 : * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
7607 : * page. We know it becomes the first record, because no other backend is
7608 : * allowed to write WAL yet.
7609 : */
7610 10 : XLogBeginInsert();
7611 10 : xlrec.overwritten_lsn = aborted_lsn;
7612 10 : xlrec.overwrite_time = GetCurrentTimestamp();
7613 10 : XLogRegisterData(&xlrec, sizeof(xl_overwrite_contrecord));
7614 10 : recptr = XLogInsert(RM_XLOG_ID, XLOG_OVERWRITE_CONTRECORD);
7615 :
7616 : /* check that the record was inserted to the right place */
7617 10 : if (ProcLastRecPtr != startPos)
7618 0 : elog(ERROR, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%08X",
7619 : LSN_FORMAT_ARGS(ProcLastRecPtr));
7620 :
7621 10 : XLogFlush(recptr);
7622 :
7623 10 : END_CRIT_SECTION();
7624 :
7625 10 : return recptr;
7626 : }
7627 :
7628 : /*
7629 : * Flush all data in shared memory to disk, and fsync
7630 : *
7631 : * This is the common code shared between regular checkpoints and
7632 : * recovery restartpoints.
7633 : */
7634 : static void
7635 1835 : CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
7636 : {
7637 1835 : CheckPointRelationMap();
7638 1835 : CheckPointReplicationSlots(flags & CHECKPOINT_IS_SHUTDOWN);
7639 1835 : CheckPointSnapBuild();
7640 1835 : CheckPointLogicalRewriteHeap();
7641 1835 : CheckPointReplicationOrigin();
7642 :
7643 : /* Write out all dirty data in SLRUs and the main buffer pool */
7644 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags);
7645 1835 : CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
7646 1835 : CheckPointCLOG();
7647 1835 : CheckPointCommitTs();
7648 1835 : CheckPointSUBTRANS();
7649 1835 : CheckPointMultiXact();
7650 1835 : CheckPointPredicate();
7651 1835 : CheckPointBuffers(flags);
7652 :
7653 : /* Perform all queued up fsyncs */
7654 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
7655 1835 : CheckpointStats.ckpt_sync_t = GetCurrentTimestamp();
7656 1835 : ProcessSyncRequests();
7657 1835 : CheckpointStats.ckpt_sync_end_t = GetCurrentTimestamp();
7658 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
7659 :
7660 : /* We deliberately delay 2PC checkpointing as long as possible */
7661 1835 : CheckPointTwoPhase(checkPointRedo);
7662 1835 : }
7663 :
7664 : /*
7665 : * Save a checkpoint for recovery restart if appropriate
7666 : *
7667 : * This function is called each time a checkpoint record is read from XLOG.
7668 : * It must determine whether the checkpoint represents a safe restartpoint or
7669 : * not. If so, the checkpoint record is stashed in shared memory so that
7670 : * CreateRestartPoint can consult it. (Note that the latter function is
7671 : * executed by the checkpointer, while this one will be executed by the
7672 : * startup process.)
7673 : */
7674 : static void
7675 735 : RecoveryRestartPoint(const CheckPoint *checkPoint, XLogReaderState *record)
7676 : {
7677 : /*
7678 : * Also refrain from creating a restartpoint if we have seen any
7679 : * references to non-existent pages. Restarting recovery from the
7680 : * restartpoint would not see the references, so we would lose the
7681 : * cross-check that the pages belonged to a relation that was dropped
7682 : * later.
7683 : */
7684 735 : if (XLogHaveInvalidPages())
7685 : {
7686 0 : elog(DEBUG2,
7687 : "could not record restart point at %X/%08X because there are unresolved references to invalid pages",
7688 : LSN_FORMAT_ARGS(checkPoint->redo));
7689 0 : return;
7690 : }
7691 :
7692 : /*
7693 : * Copy the checkpoint record to shared memory, so that checkpointer can
7694 : * work out the next time it wants to perform a restartpoint.
7695 : */
7696 735 : SpinLockAcquire(&XLogCtl->info_lck);
7697 735 : XLogCtl->lastCheckPointRecPtr = record->ReadRecPtr;
7698 735 : XLogCtl->lastCheckPointEndPtr = record->EndRecPtr;
7699 735 : XLogCtl->lastCheckPoint = *checkPoint;
7700 735 : SpinLockRelease(&XLogCtl->info_lck);
7701 : }
7702 :
7703 : /*
7704 : * Establish a restartpoint if possible.
7705 : *
7706 : * This is similar to CreateCheckPoint, but is used during WAL recovery
7707 : * to establish a point from which recovery can roll forward without
7708 : * replaying the entire recovery log.
7709 : *
7710 : * Returns true if a new restartpoint was established. We can only establish
7711 : * a restartpoint if we have replayed a safe checkpoint record since last
7712 : * restartpoint.
7713 : */
7714 : bool
7715 615 : CreateRestartPoint(int flags)
7716 : {
7717 : XLogRecPtr lastCheckPointRecPtr;
7718 : XLogRecPtr lastCheckPointEndPtr;
7719 : CheckPoint lastCheckPoint;
7720 : XLogRecPtr PriorRedoPtr;
7721 : XLogRecPtr receivePtr;
7722 : XLogRecPtr replayPtr;
7723 : TimeLineID replayTLI;
7724 : XLogRecPtr endptr;
7725 : XLogSegNo _logSegNo;
7726 : TimestampTz xtime;
7727 :
7728 : /* Concurrent checkpoint/restartpoint cannot happen */
7729 : Assert(!IsUnderPostmaster || MyBackendType == B_CHECKPOINTER);
7730 :
7731 : /* Get a local copy of the last safe checkpoint record. */
7732 615 : SpinLockAcquire(&XLogCtl->info_lck);
7733 615 : lastCheckPointRecPtr = XLogCtl->lastCheckPointRecPtr;
7734 615 : lastCheckPointEndPtr = XLogCtl->lastCheckPointEndPtr;
7735 615 : lastCheckPoint = XLogCtl->lastCheckPoint;
7736 615 : SpinLockRelease(&XLogCtl->info_lck);
7737 :
7738 : /*
7739 : * Check that we're still in recovery mode. It's ok if we exit recovery
7740 : * mode after this check, the restart point is valid anyway.
7741 : */
7742 615 : if (!RecoveryInProgress())
7743 : {
7744 0 : ereport(DEBUG2,
7745 : (errmsg_internal("skipping restartpoint, recovery has already ended")));
7746 0 : return false;
7747 : }
7748 :
7749 : /*
7750 : * If the last checkpoint record we've replayed is already our last
7751 : * restartpoint, we can't perform a new restart point. We still update
7752 : * minRecoveryPoint in that case, so that if this is a shutdown restart
7753 : * point, we won't start up earlier than before. That's not strictly
7754 : * necessary, but when hot standby is enabled, it would be rather weird if
7755 : * the database opened up for read-only connections at a point-in-time
7756 : * before the last shutdown. Such time travel is still possible in case of
7757 : * immediate shutdown, though.
7758 : *
7759 : * We don't explicitly advance minRecoveryPoint when we do create a
7760 : * restartpoint. It's assumed that flushing the buffers will do that as a
7761 : * side-effect.
7762 : */
7763 615 : if (!XLogRecPtrIsValid(lastCheckPointRecPtr) ||
7764 284 : lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
7765 : {
7766 410 : ereport(DEBUG2,
7767 : errmsg_internal("skipping restartpoint, already performed at %X/%08X",
7768 : LSN_FORMAT_ARGS(lastCheckPoint.redo)));
7769 :
7770 410 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
7771 410 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7772 : {
7773 36 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7774 36 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7775 36 : UpdateControlFile();
7776 36 : LWLockRelease(ControlFileLock);
7777 : }
7778 410 : return false;
7779 : }
7780 :
7781 : /*
7782 : * Update the shared RedoRecPtr so that the startup process can calculate
7783 : * the number of segments replayed since last restartpoint, and request a
7784 : * restartpoint if it exceeds CheckPointSegments.
7785 : *
7786 : * Like in CreateCheckPoint(), hold off insertions to update it, although
7787 : * during recovery this is just pro forma, because no WAL insertions are
7788 : * happening.
7789 : */
7790 205 : WALInsertLockAcquireExclusive();
7791 205 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = lastCheckPoint.redo;
7792 205 : WALInsertLockRelease();
7793 :
7794 : /* Also update the info_lck-protected copy */
7795 205 : SpinLockAcquire(&XLogCtl->info_lck);
7796 205 : XLogCtl->RedoRecPtr = lastCheckPoint.redo;
7797 205 : SpinLockRelease(&XLogCtl->info_lck);
7798 :
7799 : /*
7800 : * Prepare to accumulate statistics.
7801 : *
7802 : * Note: because it is possible for log_checkpoints to change while a
7803 : * checkpoint proceeds, we always accumulate stats, even if
7804 : * log_checkpoints is currently off.
7805 : */
7806 2255 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7807 205 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7808 :
7809 205 : if (log_checkpoints)
7810 205 : LogCheckpointStart(flags, true);
7811 :
7812 : /* Update the process title */
7813 205 : update_checkpoint_display(flags, true, false);
7814 :
7815 205 : CheckPointGuts(lastCheckPoint.redo, flags);
7816 :
7817 : /*
7818 : * This location needs to be after CheckPointGuts() to ensure that some
7819 : * work has already happened during this checkpoint.
7820 : */
7821 205 : INJECTION_POINT("create-restart-point", NULL);
7822 :
7823 : /*
7824 : * Remember the prior checkpoint's redo ptr for
7825 : * UpdateCheckPointDistanceEstimate()
7826 : */
7827 205 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7828 :
7829 : /*
7830 : * Update pg_control, using current time. Check that it still shows an
7831 : * older checkpoint, else do nothing; this is a quick hack to make sure
7832 : * nothing really bad happens if somehow we get here after the
7833 : * end-of-recovery checkpoint.
7834 : */
7835 205 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7836 205 : if (ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
7837 : {
7838 : /*
7839 : * Update the checkpoint information. We do this even if the cluster
7840 : * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
7841 : * segments recycled below.
7842 : */
7843 205 : ControlFile->checkPoint = lastCheckPointRecPtr;
7844 205 : ControlFile->checkPointCopy = lastCheckPoint;
7845 :
7846 : /*
7847 : * Ensure minRecoveryPoint is past the checkpoint record and update it
7848 : * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
7849 : * this will have happened already while writing out dirty buffers,
7850 : * but not necessarily - e.g. because no buffers were dirtied. We do
7851 : * this because a backup performed in recovery uses minRecoveryPoint
7852 : * to determine which WAL files must be included in the backup, and
7853 : * the file (or files) containing the checkpoint record must be
7854 : * included, at a minimum. Note that for an ordinary restart of
7855 : * recovery there's no value in having the minimum recovery point any
7856 : * earlier than this anyway, because redo will begin just after the
7857 : * checkpoint record.
7858 : */
7859 205 : if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
7860 : {
7861 205 : if (ControlFile->minRecoveryPoint < lastCheckPointEndPtr)
7862 : {
7863 19 : ControlFile->minRecoveryPoint = lastCheckPointEndPtr;
7864 19 : ControlFile->minRecoveryPointTLI = lastCheckPoint.ThisTimeLineID;
7865 :
7866 : /* update local copy */
7867 19 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
7868 19 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
7869 : }
7870 205 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7871 22 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7872 : }
7873 205 : UpdateControlFile();
7874 : }
7875 205 : LWLockRelease(ControlFileLock);
7876 :
7877 : /*
7878 : * Update the average distance between checkpoints/restartpoints if the
7879 : * prior checkpoint exists.
7880 : */
7881 205 : if (XLogRecPtrIsValid(PriorRedoPtr))
7882 205 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7883 :
7884 : /*
7885 : * Delete old log files, those no longer needed for last restartpoint to
7886 : * prevent the disk holding the xlog from growing full.
7887 : */
7888 205 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7889 :
7890 : /*
7891 : * Retreat _logSegNo using the current end of xlog replayed or received,
7892 : * whichever is later.
7893 : */
7894 205 : receivePtr = GetWalRcvFlushRecPtr(NULL, NULL);
7895 205 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
7896 205 : endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
7897 205 : KeepLogSeg(endptr, &_logSegNo);
7898 :
7899 205 : INJECTION_POINT("restartpoint-before-slot-invalidation", NULL);
7900 :
7901 205 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7902 : _logSegNo, InvalidOid,
7903 : InvalidTransactionId))
7904 : {
7905 : /*
7906 : * Some slots have been invalidated; recalculate the old-segment
7907 : * horizon, starting again from RedoRecPtr.
7908 : */
7909 1 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7910 1 : KeepLogSeg(endptr, &_logSegNo);
7911 : }
7912 205 : _logSegNo--;
7913 :
7914 : /*
7915 : * Try to recycle segments on a useful timeline. If we've been promoted
7916 : * since the beginning of this restartpoint, use the new timeline chosen
7917 : * at end of recovery. If we're still in recovery, use the timeline we're
7918 : * currently replaying.
7919 : *
7920 : * There is no guarantee that the WAL segments will be useful on the
7921 : * current timeline; if recovery proceeds to a new timeline right after
7922 : * this, the pre-allocated WAL segments on this timeline will not be used,
7923 : * and will go wasted until recycled on the next restartpoint. We'll live
7924 : * with that.
7925 : */
7926 205 : if (!RecoveryInProgress())
7927 0 : replayTLI = XLogCtl->InsertTimeLineID;
7928 :
7929 205 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, endptr, replayTLI);
7930 :
7931 : /*
7932 : * Make more log segments if needed. (Do this after recycling old log
7933 : * segments, since that may supply some of the needed files.)
7934 : */
7935 205 : PreallocXlogFiles(endptr, replayTLI);
7936 :
7937 : /*
7938 : * Truncate pg_subtrans if possible. We can throw away all data before
7939 : * the oldest XMIN of any running transaction. No future transaction will
7940 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7941 : * in subtrans.c). When hot standby is disabled, though, we mustn't do
7942 : * this because StartupSUBTRANS hasn't been called yet.
7943 : */
7944 205 : if (EnableHotStandby)
7945 205 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7946 :
7947 : /* Real work is done; log and update stats. */
7948 205 : LogCheckpointEnd(true, flags);
7949 :
7950 : /* Reset the process title */
7951 205 : update_checkpoint_display(flags, true, true);
7952 :
7953 205 : xtime = GetLatestXTime();
7954 205 : ereport((log_checkpoints ? LOG : DEBUG2),
7955 : errmsg("recovery restart point at %X/%08X",
7956 : LSN_FORMAT_ARGS(lastCheckPoint.redo)),
7957 : xtime ? errdetail("Last completed transaction was at log time %s.",
7958 : timestamptz_to_str(xtime)) : 0);
7959 :
7960 : /*
7961 : * Finally, execute archive_cleanup_command, if any.
7962 : */
7963 205 : if (archiveCleanupCommand && strcmp(archiveCleanupCommand, "") != 0)
7964 0 : ExecuteRecoveryCommand(archiveCleanupCommand,
7965 : "archive_cleanup_command",
7966 : false,
7967 : WAIT_EVENT_ARCHIVE_CLEANUP_COMMAND);
7968 :
7969 205 : return true;
7970 : }
7971 :
7972 : /*
7973 : * Report availability of WAL for the given target LSN
7974 : * (typically a slot's restart_lsn)
7975 : *
7976 : * Returns one of the following enum values:
7977 : *
7978 : * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
7979 : * max_wal_size.
7980 : *
7981 : * * WALAVAIL_EXTENDED means it is still available by preserving extra
7982 : * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
7983 : * than max_wal_size, this state is not returned.
7984 : *
7985 : * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
7986 : * remove reserved segments. The walsender using this slot may return to the
7987 : * above.
7988 : *
7989 : * * WALAVAIL_REMOVED means it has been removed. A replication stream on
7990 : * a slot with this LSN cannot continue. (Any associated walsender
7991 : * processes should have been terminated already.)
7992 : *
7993 : * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
7994 : */
7995 : WALAvailability
7996 640 : GetWALAvailability(XLogRecPtr targetLSN)
7997 : {
7998 : XLogRecPtr currpos; /* current write LSN */
7999 : XLogSegNo currSeg; /* segid of currpos */
8000 : XLogSegNo targetSeg; /* segid of targetLSN */
8001 : XLogSegNo oldestSeg; /* actual oldest segid */
8002 : XLogSegNo oldestSegMaxWalSize; /* oldest segid kept by max_wal_size */
8003 : XLogSegNo oldestSlotSeg; /* oldest segid kept by slot */
8004 : uint64 keepSegs;
8005 :
8006 : /*
8007 : * slot does not reserve WAL. Either deactivated, or has never been active
8008 : */
8009 640 : if (!XLogRecPtrIsValid(targetLSN))
8010 27 : return WALAVAIL_INVALID_LSN;
8011 :
8012 : /*
8013 : * Calculate the oldest segment currently reserved by all slots,
8014 : * considering wal_keep_size and max_slot_wal_keep_size. Initialize
8015 : * oldestSlotSeg to the current segment.
8016 : */
8017 613 : currpos = GetXLogWriteRecPtr();
8018 613 : XLByteToSeg(currpos, oldestSlotSeg, wal_segment_size);
8019 613 : KeepLogSeg(currpos, &oldestSlotSeg);
8020 :
8021 : /*
8022 : * Find the oldest extant segment file. We get 1 until checkpoint removes
8023 : * the first WAL segment file since startup, which causes the status being
8024 : * wrong under certain abnormal conditions but that doesn't actually harm.
8025 : */
8026 613 : oldestSeg = XLogGetLastRemovedSegno() + 1;
8027 :
8028 : /* calculate oldest segment by max_wal_size */
8029 613 : XLByteToSeg(currpos, currSeg, wal_segment_size);
8030 613 : keepSegs = ConvertToXSegs(max_wal_size_mb, wal_segment_size) + 1;
8031 :
8032 613 : if (currSeg > keepSegs)
8033 8 : oldestSegMaxWalSize = currSeg - keepSegs;
8034 : else
8035 605 : oldestSegMaxWalSize = 1;
8036 :
8037 : /* the segment we care about */
8038 613 : XLByteToSeg(targetLSN, targetSeg, wal_segment_size);
8039 :
8040 : /*
8041 : * No point in returning reserved or extended status values if the
8042 : * targetSeg is known to be lost.
8043 : */
8044 613 : if (targetSeg >= oldestSlotSeg)
8045 : {
8046 : /* show "reserved" when targetSeg is within max_wal_size */
8047 612 : if (targetSeg >= oldestSegMaxWalSize)
8048 610 : return WALAVAIL_RESERVED;
8049 :
8050 : /* being retained by slots exceeding max_wal_size */
8051 2 : return WALAVAIL_EXTENDED;
8052 : }
8053 :
8054 : /* WAL segments are no longer retained but haven't been removed yet */
8055 1 : if (targetSeg >= oldestSeg)
8056 1 : return WALAVAIL_UNRESERVED;
8057 :
8058 : /* Definitely lost */
8059 0 : return WALAVAIL_REMOVED;
8060 : }
8061 :
8062 :
8063 : /*
8064 : * Retreat *logSegNo to the last segment that we need to retain because of
8065 : * either wal_keep_size or replication slots.
8066 : *
8067 : * This is calculated by subtracting wal_keep_size from the given xlog
8068 : * location, recptr and by making sure that that result is below the
8069 : * requirement of replication slots. For the latter criterion we do consider
8070 : * the effects of max_slot_wal_keep_size: reserve at most that much space back
8071 : * from recptr.
8072 : *
8073 : * Note about replication slots: if this function calculates a value
8074 : * that's further ahead than what slots need reserved, then affected
8075 : * slots need to be invalidated and this function invoked again.
8076 : * XXX it might be a good idea to rewrite this function so that
8077 : * invalidation is optionally done here, instead.
8078 : */
8079 : static void
8080 2453 : KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
8081 : {
8082 : XLogSegNo currSegNo;
8083 : XLogSegNo segno;
8084 : XLogRecPtr keep;
8085 :
8086 2453 : XLByteToSeg(recptr, currSegNo, wal_segment_size);
8087 2453 : segno = currSegNo;
8088 :
8089 : /* Calculate how many segments are kept by slots. */
8090 2453 : keep = XLogGetReplicationSlotMinimumLSN();
8091 2453 : if (XLogRecPtrIsValid(keep) && keep < recptr)
8092 : {
8093 772 : XLByteToSeg(keep, segno, wal_segment_size);
8094 :
8095 : /*
8096 : * Account for max_slot_wal_keep_size to avoid keeping more than
8097 : * configured. However, don't do that during a binary upgrade: if
8098 : * slots were to be invalidated because of this, it would not be
8099 : * possible to preserve logical ones during the upgrade.
8100 : */
8101 772 : if (max_slot_wal_keep_size_mb >= 0 && !IsBinaryUpgrade)
8102 : {
8103 : uint64 slot_keep_segs;
8104 :
8105 23 : slot_keep_segs =
8106 23 : ConvertToXSegs(max_slot_wal_keep_size_mb, wal_segment_size);
8107 :
8108 23 : if (currSegNo - segno > slot_keep_segs)
8109 6 : segno = currSegNo - slot_keep_segs;
8110 : }
8111 : }
8112 :
8113 : /*
8114 : * If WAL summarization is in use, don't remove WAL that has yet to be
8115 : * summarized.
8116 : */
8117 2453 : keep = GetOldestUnsummarizedLSN(NULL, NULL);
8118 2453 : if (XLogRecPtrIsValid(keep))
8119 : {
8120 : XLogSegNo unsummarized_segno;
8121 :
8122 7 : XLByteToSeg(keep, unsummarized_segno, wal_segment_size);
8123 7 : if (unsummarized_segno < segno)
8124 7 : segno = unsummarized_segno;
8125 : }
8126 :
8127 : /* but, keep at least wal_keep_size if that's set */
8128 2453 : if (wal_keep_size_mb > 0)
8129 : {
8130 : uint64 keep_segs;
8131 :
8132 74 : keep_segs = ConvertToXSegs(wal_keep_size_mb, wal_segment_size);
8133 74 : if (currSegNo - segno < keep_segs)
8134 : {
8135 : /* avoid underflow, don't go below 1 */
8136 74 : if (currSegNo <= keep_segs)
8137 70 : segno = 1;
8138 : else
8139 4 : segno = currSegNo - keep_segs;
8140 : }
8141 : }
8142 :
8143 : /* don't delete WAL segments newer than the calculated segment */
8144 2453 : if (segno < *logSegNo)
8145 364 : *logSegNo = segno;
8146 2453 : }
8147 :
8148 : /*
8149 : * Write a NEXTOID log record
8150 : */
8151 : void
8152 654 : XLogPutNextOid(Oid nextOid)
8153 : {
8154 654 : XLogBeginInsert();
8155 654 : XLogRegisterData(&nextOid, sizeof(Oid));
8156 654 : (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID);
8157 :
8158 : /*
8159 : * We need not flush the NEXTOID record immediately, because any of the
8160 : * just-allocated OIDs could only reach disk as part of a tuple insert or
8161 : * update that would have its own XLOG record that must follow the NEXTOID
8162 : * record. Therefore, the standard buffer LSN interlock applied to those
8163 : * records will ensure no such OID reaches disk before the NEXTOID record
8164 : * does.
8165 : *
8166 : * Note, however, that the above statement only covers state "within" the
8167 : * database. When we use a generated OID as a file or directory name, we
8168 : * are in a sense violating the basic WAL rule, because that filesystem
8169 : * change may reach disk before the NEXTOID WAL record does. The impact
8170 : * of this is that if a database crash occurs immediately afterward, we
8171 : * might after restart re-generate the same OID and find that it conflicts
8172 : * with the leftover file or directory. But since for safety's sake we
8173 : * always loop until finding a nonconflicting filename, this poses no real
8174 : * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8175 : */
8176 654 : }
8177 :
8178 : /*
8179 : * Write an XLOG SWITCH record.
8180 : *
8181 : * Here we just blindly issue an XLogInsert request for the record.
8182 : * All the magic happens inside XLogInsert.
8183 : *
8184 : * The return value is either the end+1 address of the switch record,
8185 : * or the end+1 address of the prior segment if we did not need to
8186 : * write a switch record because we are already at segment start.
8187 : */
8188 : XLogRecPtr
8189 812 : RequestXLogSwitch(bool mark_unimportant)
8190 : {
8191 : XLogRecPtr RecPtr;
8192 :
8193 : /* XLOG SWITCH has no data */
8194 812 : XLogBeginInsert();
8195 :
8196 812 : if (mark_unimportant)
8197 0 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8198 812 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH);
8199 :
8200 812 : return RecPtr;
8201 : }
8202 :
8203 : /*
8204 : * Write a RESTORE POINT record
8205 : */
8206 : XLogRecPtr
8207 3 : XLogRestorePoint(const char *rpName)
8208 : {
8209 : XLogRecPtr RecPtr;
8210 : xl_restore_point xlrec;
8211 :
8212 3 : xlrec.rp_time = GetCurrentTimestamp();
8213 3 : strlcpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8214 :
8215 3 : XLogBeginInsert();
8216 3 : XLogRegisterData(&xlrec, sizeof(xl_restore_point));
8217 :
8218 3 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT);
8219 :
8220 3 : ereport(LOG,
8221 : errmsg("restore point \"%s\" created at %X/%08X",
8222 : rpName, LSN_FORMAT_ARGS(RecPtr)));
8223 :
8224 3 : return RecPtr;
8225 : }
8226 :
8227 : /*
8228 : * Write an empty XLOG record to assign a distinct LSN.
8229 : *
8230 : * This is used by some index AMs when building indexes on permanent relations
8231 : * with wal_level=minimal. In that scenario, WAL-logging will start after
8232 : * commit, but the index AM needs distinct LSNs to detect concurrent page
8233 : * modifications. When the current WAL insert position hasn't advanced since
8234 : * the last call, we emit a dummy record to ensure we get a new, distinct LSN.
8235 : */
8236 : XLogRecPtr
8237 94833 : XLogAssignLSN(void)
8238 : {
8239 94833 : int dummy = 0;
8240 :
8241 : /*
8242 : * Records other than XLOG_SWITCH must have content. We use an integer 0
8243 : * to satisfy this restriction.
8244 : */
8245 94833 : XLogBeginInsert();
8246 94833 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8247 94833 : XLogRegisterData(&dummy, sizeof(dummy));
8248 94833 : return XLogInsert(RM_XLOG_ID, XLOG_ASSIGN_LSN);
8249 : }
8250 :
8251 : /*
8252 : * Check if any of the GUC parameters that are critical for hot standby
8253 : * have changed, and update the value in pg_control file if necessary.
8254 : */
8255 : static void
8256 970 : XLogReportParameters(void)
8257 : {
8258 970 : if (wal_level != ControlFile->wal_level ||
8259 708 : wal_log_hints != ControlFile->wal_log_hints ||
8260 621 : MaxConnections != ControlFile->MaxConnections ||
8261 620 : max_worker_processes != ControlFile->max_worker_processes ||
8262 617 : max_wal_senders != ControlFile->max_wal_senders ||
8263 590 : max_prepared_xacts != ControlFile->max_prepared_xacts ||
8264 487 : max_locks_per_xact != ControlFile->max_locks_per_xact ||
8265 487 : track_commit_timestamp != ControlFile->track_commit_timestamp)
8266 : {
8267 : /*
8268 : * The change in number of backend slots doesn't need to be WAL-logged
8269 : * if archiving is not enabled, as you can't start archive recovery
8270 : * with wal_level=minimal anyway. We don't really care about the
8271 : * values in pg_control either if wal_level=minimal, but seems better
8272 : * to keep them up-to-date to avoid confusion.
8273 : */
8274 495 : if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8275 : {
8276 : xl_parameter_change xlrec;
8277 : XLogRecPtr recptr;
8278 :
8279 468 : xlrec.MaxConnections = MaxConnections;
8280 468 : xlrec.max_worker_processes = max_worker_processes;
8281 468 : xlrec.max_wal_senders = max_wal_senders;
8282 468 : xlrec.max_prepared_xacts = max_prepared_xacts;
8283 468 : xlrec.max_locks_per_xact = max_locks_per_xact;
8284 468 : xlrec.wal_level = wal_level;
8285 468 : xlrec.wal_log_hints = wal_log_hints;
8286 468 : xlrec.track_commit_timestamp = track_commit_timestamp;
8287 :
8288 468 : XLogBeginInsert();
8289 468 : XLogRegisterData(&xlrec, sizeof(xlrec));
8290 :
8291 468 : recptr = XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE);
8292 468 : XLogFlush(recptr);
8293 : }
8294 :
8295 495 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8296 :
8297 495 : ControlFile->MaxConnections = MaxConnections;
8298 495 : ControlFile->max_worker_processes = max_worker_processes;
8299 495 : ControlFile->max_wal_senders = max_wal_senders;
8300 495 : ControlFile->max_prepared_xacts = max_prepared_xacts;
8301 495 : ControlFile->max_locks_per_xact = max_locks_per_xact;
8302 495 : ControlFile->wal_level = wal_level;
8303 495 : ControlFile->wal_log_hints = wal_log_hints;
8304 495 : ControlFile->track_commit_timestamp = track_commit_timestamp;
8305 495 : UpdateControlFile();
8306 :
8307 495 : LWLockRelease(ControlFileLock);
8308 : }
8309 970 : }
8310 :
8311 : /*
8312 : * Update full_page_writes in shared memory, and write an
8313 : * XLOG_FPW_CHANGE record if necessary.
8314 : *
8315 : * Note: this function assumes there is no other process running
8316 : * concurrently that could update it.
8317 : */
8318 : void
8319 1655 : UpdateFullPageWrites(void)
8320 : {
8321 1655 : XLogCtlInsert *Insert = &XLogCtl->Insert;
8322 : bool recoveryInProgress;
8323 :
8324 : /*
8325 : * Do nothing if full_page_writes has not been changed.
8326 : *
8327 : * It's safe to check the shared full_page_writes without the lock,
8328 : * because we assume that there is no concurrently running process which
8329 : * can update it.
8330 : */
8331 1655 : if (fullPageWrites == Insert->fullPageWrites)
8332 1212 : return;
8333 :
8334 : /*
8335 : * Perform this outside critical section so that the WAL insert
8336 : * initialization done by RecoveryInProgress() doesn't trigger an
8337 : * assertion failure.
8338 : */
8339 443 : recoveryInProgress = RecoveryInProgress();
8340 :
8341 443 : START_CRIT_SECTION();
8342 :
8343 : /*
8344 : * It's always safe to take full page images, even when not strictly
8345 : * required, but not the other round. So if we're setting full_page_writes
8346 : * to true, first set it true and then write the WAL record. If we're
8347 : * setting it to false, first write the WAL record and then set the global
8348 : * flag.
8349 : */
8350 443 : if (fullPageWrites)
8351 : {
8352 431 : WALInsertLockAcquireExclusive();
8353 431 : Insert->fullPageWrites = true;
8354 431 : WALInsertLockRelease();
8355 : }
8356 :
8357 : /*
8358 : * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8359 : * full_page_writes during archive recovery, if required.
8360 : */
8361 443 : if (XLogStandbyInfoActive() && !recoveryInProgress)
8362 : {
8363 0 : XLogBeginInsert();
8364 0 : XLogRegisterData(&fullPageWrites, sizeof(bool));
8365 :
8366 0 : XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
8367 : }
8368 :
8369 443 : if (!fullPageWrites)
8370 : {
8371 12 : WALInsertLockAcquireExclusive();
8372 12 : Insert->fullPageWrites = false;
8373 12 : WALInsertLockRelease();
8374 : }
8375 443 : END_CRIT_SECTION();
8376 : }
8377 :
8378 : /*
8379 : * XLOG resource manager's routines
8380 : *
8381 : * Definitions of info values are in include/catalog/pg_control.h, though
8382 : * not all record types are related to control file updates.
8383 : *
8384 : * NOTE: Some XLOG record types that are directly related to WAL recovery
8385 : * are handled in xlogrecovery_redo().
8386 : */
8387 : void
8388 106897 : xlog_redo(XLogReaderState *record)
8389 : {
8390 106897 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8391 106897 : XLogRecPtr lsn = record->EndRecPtr;
8392 :
8393 : /*
8394 : * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
8395 : * XLOG_FPI_FOR_HINT records.
8396 : */
8397 : Assert(info == XLOG_FPI || info == XLOG_FPI_FOR_HINT ||
8398 : !XLogRecHasAnyBlockRefs(record));
8399 :
8400 106897 : if (info == XLOG_NEXTOID)
8401 : {
8402 : Oid nextOid;
8403 :
8404 : /*
8405 : * We used to try to take the maximum of TransamVariables->nextOid and
8406 : * the recorded nextOid, but that fails if the OID counter wraps
8407 : * around. Since no OID allocation should be happening during replay
8408 : * anyway, better to just believe the record exactly. We still take
8409 : * OidGenLock while setting the variable, just in case.
8410 : */
8411 95 : memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
8412 95 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8413 95 : TransamVariables->nextOid = nextOid;
8414 95 : TransamVariables->oidCount = 0;
8415 95 : LWLockRelease(OidGenLock);
8416 : }
8417 106802 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
8418 : {
8419 : CheckPoint checkPoint;
8420 : TimeLineID replayTLI;
8421 :
8422 41 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8423 : /* In a SHUTDOWN checkpoint, believe the counters exactly */
8424 41 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8425 41 : TransamVariables->nextXid = checkPoint.nextXid;
8426 41 : LWLockRelease(XidGenLock);
8427 41 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8428 41 : TransamVariables->nextOid = checkPoint.nextOid;
8429 41 : TransamVariables->oidCount = 0;
8430 41 : LWLockRelease(OidGenLock);
8431 41 : MultiXactSetNextMXact(checkPoint.nextMulti,
8432 : checkPoint.nextMultiOffset);
8433 :
8434 41 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8435 : checkPoint.oldestMultiDB);
8436 :
8437 : /*
8438 : * No need to set oldestClogXid here as well; it'll be set when we
8439 : * redo an xl_clog_truncate if it changed since initialization.
8440 : */
8441 41 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
8442 :
8443 : /*
8444 : * If we see a shutdown checkpoint while waiting for an end-of-backup
8445 : * record, the backup was canceled and the end-of-backup record will
8446 : * never arrive.
8447 : */
8448 41 : if (ArchiveRecoveryRequested &&
8449 41 : XLogRecPtrIsValid(ControlFile->backupStartPoint) &&
8450 0 : !XLogRecPtrIsValid(ControlFile->backupEndPoint))
8451 0 : ereport(PANIC,
8452 : (errmsg("online backup was canceled, recovery cannot continue")));
8453 :
8454 : /*
8455 : * If we see a shutdown checkpoint, we know that nothing was running
8456 : * on the primary at this point. So fake-up an empty running-xacts
8457 : * record and use that here and now. Recover additional standby state
8458 : * for prepared transactions.
8459 : */
8460 41 : if (standbyState >= STANDBY_INITIALIZED)
8461 : {
8462 : TransactionId *xids;
8463 : int nxids;
8464 : TransactionId oldestActiveXID;
8465 : TransactionId latestCompletedXid;
8466 : RunningTransactionsData running;
8467 :
8468 39 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
8469 :
8470 : /* Update pg_subtrans entries for any prepared transactions */
8471 39 : StandbyRecoverPreparedTransactions();
8472 :
8473 : /*
8474 : * Construct a RunningTransactions snapshot representing a shut
8475 : * down server, with only prepared transactions still alive. We're
8476 : * never overflowed at this point because all subxids are listed
8477 : * with their parent prepared transactions.
8478 : */
8479 39 : running.xcnt = nxids;
8480 39 : running.subxcnt = 0;
8481 39 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
8482 39 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
8483 39 : running.oldestRunningXid = oldestActiveXID;
8484 39 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
8485 39 : TransactionIdRetreat(latestCompletedXid);
8486 : Assert(TransactionIdIsNormal(latestCompletedXid));
8487 39 : running.latestCompletedXid = latestCompletedXid;
8488 39 : running.xids = xids;
8489 :
8490 39 : ProcArrayApplyRecoveryInfo(&running);
8491 : }
8492 :
8493 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8494 41 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8495 41 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8496 41 : LWLockRelease(ControlFileLock);
8497 :
8498 : /*
8499 : * We should've already switched to the new TLI before replaying this
8500 : * record.
8501 : */
8502 41 : (void) GetCurrentReplayRecPtr(&replayTLI);
8503 41 : if (checkPoint.ThisTimeLineID != replayTLI)
8504 0 : ereport(PANIC,
8505 : (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
8506 : checkPoint.ThisTimeLineID, replayTLI)));
8507 :
8508 41 : RecoveryRestartPoint(&checkPoint, record);
8509 :
8510 : /*
8511 : * After replaying a checkpoint record, free all smgr objects.
8512 : * Otherwise we would never do so for dropped relations, as the
8513 : * startup does not process shared invalidation messages or call
8514 : * AtEOXact_SMgr().
8515 : */
8516 41 : smgrdestroyall();
8517 : }
8518 106761 : else if (info == XLOG_CHECKPOINT_ONLINE)
8519 : {
8520 : CheckPoint checkPoint;
8521 : TimeLineID replayTLI;
8522 :
8523 694 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8524 : /* In an ONLINE checkpoint, treat the XID counter as a minimum */
8525 694 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8526 694 : if (FullTransactionIdPrecedes(TransamVariables->nextXid,
8527 : checkPoint.nextXid))
8528 0 : TransamVariables->nextXid = checkPoint.nextXid;
8529 694 : LWLockRelease(XidGenLock);
8530 :
8531 : /*
8532 : * We ignore the nextOid counter in an ONLINE checkpoint, preferring
8533 : * to track OID assignment through XLOG_NEXTOID records. The nextOid
8534 : * counter is from the start of the checkpoint and might well be stale
8535 : * compared to later XLOG_NEXTOID records. We could try to take the
8536 : * maximum of the nextOid counter and our latest value, but since
8537 : * there's no particular guarantee about the speed with which the OID
8538 : * counter wraps around, that's a risky thing to do. In any case,
8539 : * users of the nextOid counter are required to avoid assignment of
8540 : * duplicates, so that a somewhat out-of-date value should be safe.
8541 : */
8542 :
8543 : /* Handle multixact */
8544 694 : MultiXactAdvanceNextMXact(checkPoint.nextMulti,
8545 : checkPoint.nextMultiOffset);
8546 :
8547 : /*
8548 : * NB: This may perform multixact truncation when replaying WAL
8549 : * generated by an older primary.
8550 : */
8551 694 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8552 : checkPoint.oldestMultiDB);
8553 694 : if (TransactionIdPrecedes(TransamVariables->oldestXid,
8554 : checkPoint.oldestXid))
8555 0 : SetTransactionIdLimit(checkPoint.oldestXid,
8556 : checkPoint.oldestXidDB);
8557 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8558 694 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8559 694 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8560 694 : LWLockRelease(ControlFileLock);
8561 :
8562 : /* TLI should not change in an on-line checkpoint */
8563 694 : (void) GetCurrentReplayRecPtr(&replayTLI);
8564 694 : if (checkPoint.ThisTimeLineID != replayTLI)
8565 0 : ereport(PANIC,
8566 : (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
8567 : checkPoint.ThisTimeLineID, replayTLI)));
8568 :
8569 694 : RecoveryRestartPoint(&checkPoint, record);
8570 :
8571 : /*
8572 : * After replaying a checkpoint record, free all smgr objects.
8573 : * Otherwise we would never do so for dropped relations, as the
8574 : * startup does not process shared invalidation messages or call
8575 : * AtEOXact_SMgr().
8576 : */
8577 694 : smgrdestroyall();
8578 : }
8579 106067 : else if (info == XLOG_OVERWRITE_CONTRECORD)
8580 : {
8581 : /* nothing to do here, handled in xlogrecovery_redo() */
8582 : }
8583 106066 : else if (info == XLOG_END_OF_RECOVERY)
8584 : {
8585 : xl_end_of_recovery xlrec;
8586 : TimeLineID replayTLI;
8587 :
8588 10 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
8589 :
8590 : /*
8591 : * For Hot Standby, we could treat this like a Shutdown Checkpoint,
8592 : * but this case is rarer and harder to test, so the benefit doesn't
8593 : * outweigh the potential extra cost of maintenance.
8594 : */
8595 :
8596 : /*
8597 : * We should've already switched to the new TLI before replaying this
8598 : * record.
8599 : */
8600 10 : (void) GetCurrentReplayRecPtr(&replayTLI);
8601 10 : if (xlrec.ThisTimeLineID != replayTLI)
8602 0 : ereport(PANIC,
8603 : (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
8604 : xlrec.ThisTimeLineID, replayTLI)));
8605 : }
8606 106056 : else if (info == XLOG_NOOP)
8607 : {
8608 : /* nothing to do here */
8609 : }
8610 106056 : else if (info == XLOG_SWITCH)
8611 : {
8612 : /* nothing to do here */
8613 : }
8614 105602 : else if (info == XLOG_RESTORE_POINT)
8615 : {
8616 : /* nothing to do here, handled in xlogrecovery.c */
8617 : }
8618 105597 : else if (info == XLOG_ASSIGN_LSN)
8619 : {
8620 : /* nothing to do here, see XLogGetFakeLSN() */
8621 : }
8622 43271 : else if (info == XLOG_FPI || info == XLOG_FPI_FOR_HINT)
8623 : {
8624 : /*
8625 : * XLOG_FPI records contain nothing else but one or more block
8626 : * references. Every block reference must include a full-page image
8627 : * even if full_page_writes was disabled when the record was generated
8628 : * - otherwise there would be no point in this record.
8629 : *
8630 : * XLOG_FPI_FOR_HINT records are generated when a page needs to be
8631 : * WAL-logged because of a hint bit update. They are only generated
8632 : * when checksums and/or wal_log_hints are enabled. They may include
8633 : * no full-page images if full_page_writes was disabled when they were
8634 : * generated. In this case there is nothing to do here.
8635 : *
8636 : * No recovery conflicts are generated by these generic records - if a
8637 : * resource manager needs to generate conflicts, it has to define a
8638 : * separate WAL record type and redo routine.
8639 : */
8640 89731 : for (uint8 block_id = 0; block_id <= XLogRecMaxBlockId(record); block_id++)
8641 : {
8642 : Buffer buffer;
8643 :
8644 47306 : if (!XLogRecHasBlockImage(record, block_id))
8645 : {
8646 66 : if (info == XLOG_FPI)
8647 0 : elog(ERROR, "XLOG_FPI record did not contain a full-page image");
8648 66 : continue;
8649 : }
8650 :
8651 47240 : if (XLogReadBufferForRedo(record, block_id, &buffer) != BLK_RESTORED)
8652 0 : elog(ERROR, "unexpected XLogReadBufferForRedo result when restoring backup block");
8653 47240 : UnlockReleaseBuffer(buffer);
8654 : }
8655 : }
8656 846 : else if (info == XLOG_BACKUP_END)
8657 : {
8658 : /* nothing to do here, handled in xlogrecovery_redo() */
8659 : }
8660 751 : else if (info == XLOG_PARAMETER_CHANGE)
8661 : {
8662 : xl_parameter_change xlrec;
8663 :
8664 : /* Update our copy of the parameters in pg_control */
8665 38 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
8666 :
8667 38 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8668 38 : ControlFile->MaxConnections = xlrec.MaxConnections;
8669 38 : ControlFile->max_worker_processes = xlrec.max_worker_processes;
8670 38 : ControlFile->max_wal_senders = xlrec.max_wal_senders;
8671 38 : ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
8672 38 : ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
8673 38 : ControlFile->wal_level = xlrec.wal_level;
8674 38 : ControlFile->wal_log_hints = xlrec.wal_log_hints;
8675 :
8676 : /*
8677 : * Update minRecoveryPoint to ensure that if recovery is aborted, we
8678 : * recover back up to this point before allowing hot standby again.
8679 : * This is important if the max_* settings are decreased, to ensure
8680 : * you don't run queries against the WAL preceding the change. The
8681 : * local copies cannot be updated as long as crash recovery is
8682 : * happening and we expect all the WAL to be replayed.
8683 : */
8684 38 : if (InArchiveRecovery)
8685 : {
8686 23 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
8687 23 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
8688 : }
8689 38 : if (XLogRecPtrIsValid(LocalMinRecoveryPoint) && LocalMinRecoveryPoint < lsn)
8690 : {
8691 : TimeLineID replayTLI;
8692 :
8693 12 : (void) GetCurrentReplayRecPtr(&replayTLI);
8694 12 : ControlFile->minRecoveryPoint = lsn;
8695 12 : ControlFile->minRecoveryPointTLI = replayTLI;
8696 : }
8697 :
8698 38 : CommitTsParameterChange(xlrec.track_commit_timestamp,
8699 38 : ControlFile->track_commit_timestamp);
8700 38 : ControlFile->track_commit_timestamp = xlrec.track_commit_timestamp;
8701 :
8702 38 : UpdateControlFile();
8703 38 : LWLockRelease(ControlFileLock);
8704 :
8705 : /* Check to see if any parameter change gives a problem on recovery */
8706 38 : CheckRequiredParameterValues();
8707 : }
8708 713 : else if (info == XLOG_FPW_CHANGE)
8709 : {
8710 : bool fpw;
8711 :
8712 0 : memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
8713 :
8714 : /*
8715 : * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
8716 : * do_pg_backup_start() and do_pg_backup_stop() can check whether
8717 : * full_page_writes has been disabled during online backup.
8718 : */
8719 0 : if (!fpw)
8720 : {
8721 0 : SpinLockAcquire(&XLogCtl->info_lck);
8722 0 : if (XLogCtl->lastFpwDisableRecPtr < record->ReadRecPtr)
8723 0 : XLogCtl->lastFpwDisableRecPtr = record->ReadRecPtr;
8724 0 : SpinLockRelease(&XLogCtl->info_lck);
8725 : }
8726 :
8727 : /* Keep track of full_page_writes */
8728 0 : lastFullPageWrites = fpw;
8729 : }
8730 713 : else if (info == XLOG_CHECKPOINT_REDO)
8731 : {
8732 : /* nothing to do here, just for informational purposes */
8733 : }
8734 18 : else if (info == XLOG_LOGICAL_DECODING_STATUS_CHANGE)
8735 : {
8736 : bool status;
8737 :
8738 18 : memcpy(&status, XLogRecGetData(record), sizeof(bool));
8739 :
8740 : /*
8741 : * We need to toggle the logical decoding status and update the
8742 : * XLogLogicalInfo cache of processes synchronously because
8743 : * XLogLogicalInfoActive() is used even during read-only queries
8744 : * (e.g., via RelationIsAccessibleInLogicalDecoding()). In the
8745 : * 'disable' case, it is safe to invalidate existing slots after
8746 : * disabling logical decoding because logical decoding cannot process
8747 : * subsequent WAL records, which may not contain logical information.
8748 : */
8749 18 : if (status)
8750 9 : EnableLogicalDecoding();
8751 : else
8752 9 : DisableLogicalDecoding();
8753 :
8754 18 : elog(DEBUG1, "update logical decoding status to %d during recovery",
8755 : status);
8756 :
8757 18 : if (InRecovery && InHotStandby)
8758 : {
8759 16 : if (!status)
8760 : {
8761 : /*
8762 : * Invalidate logical slots if we are in hot standby and the
8763 : * primary disabled logical decoding.
8764 : */
8765 9 : InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_LEVEL,
8766 : 0, InvalidOid,
8767 : InvalidTransactionId);
8768 : }
8769 7 : else if (sync_replication_slots)
8770 : {
8771 : /*
8772 : * Signal the postmaster to launch the slotsync worker.
8773 : *
8774 : * XXX: For simplicity, we keep the slotsync worker running
8775 : * even after logical decoding is disabled. A future
8776 : * improvement can consider starting and stopping the worker
8777 : * based on logical decoding status change.
8778 : */
8779 0 : kill(PostmasterPid, SIGUSR1);
8780 : }
8781 : }
8782 : }
8783 106895 : }
8784 :
8785 : /*
8786 : * Return the extra open flags used for opening a file, depending on the
8787 : * value of the GUCs wal_sync_method, fsync and debug_io_direct.
8788 : */
8789 : static int
8790 17265 : get_sync_bit(int method)
8791 : {
8792 17265 : int o_direct_flag = 0;
8793 :
8794 : /*
8795 : * Use O_DIRECT if requested, except in walreceiver process. The WAL
8796 : * written by walreceiver is normally read by the startup process soon
8797 : * after it's written. Also, walreceiver performs unaligned writes, which
8798 : * don't work with O_DIRECT, so it is required for correctness too.
8799 : */
8800 17265 : if ((io_direct_flags & IO_DIRECT_WAL) && !AmWalReceiverProcess())
8801 10 : o_direct_flag = PG_O_DIRECT;
8802 :
8803 : /* If fsync is disabled, never open in sync mode */
8804 17265 : if (!enableFsync)
8805 17265 : return o_direct_flag;
8806 :
8807 0 : switch (method)
8808 : {
8809 : /*
8810 : * enum values for all sync options are defined even if they are
8811 : * not supported on the current platform. But if not, they are
8812 : * not included in the enum option array, and therefore will never
8813 : * be seen here.
8814 : */
8815 0 : case WAL_SYNC_METHOD_FSYNC:
8816 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8817 : case WAL_SYNC_METHOD_FDATASYNC:
8818 0 : return o_direct_flag;
8819 : #ifdef O_SYNC
8820 0 : case WAL_SYNC_METHOD_OPEN:
8821 0 : return O_SYNC | o_direct_flag;
8822 : #endif
8823 : #ifdef O_DSYNC
8824 0 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8825 0 : return O_DSYNC | o_direct_flag;
8826 : #endif
8827 0 : default:
8828 : /* can't happen (unless we are out of sync with option array) */
8829 0 : elog(ERROR, "unrecognized \"wal_sync_method\": %d", method);
8830 : return 0; /* silence warning */
8831 : }
8832 : }
8833 :
8834 : /*
8835 : * GUC support
8836 : */
8837 : void
8838 1222 : assign_wal_sync_method(int new_wal_sync_method, void *extra)
8839 : {
8840 1222 : if (wal_sync_method != new_wal_sync_method)
8841 : {
8842 : /*
8843 : * To ensure that no blocks escape unsynced, force an fsync on the
8844 : * currently open log segment (if any). Also, if the open flag is
8845 : * changing, close the log file so it will be reopened (with new flag
8846 : * bit) at next use.
8847 : */
8848 0 : if (openLogFile >= 0)
8849 : {
8850 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN);
8851 0 : if (pg_fsync(openLogFile) != 0)
8852 : {
8853 : char xlogfname[MAXFNAMELEN];
8854 : int save_errno;
8855 :
8856 0 : save_errno = errno;
8857 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo,
8858 : wal_segment_size);
8859 0 : errno = save_errno;
8860 0 : ereport(PANIC,
8861 : (errcode_for_file_access(),
8862 : errmsg("could not fsync file \"%s\": %m", xlogfname)));
8863 : }
8864 :
8865 0 : pgstat_report_wait_end();
8866 0 : if (get_sync_bit(wal_sync_method) != get_sync_bit(new_wal_sync_method))
8867 0 : XLogFileClose();
8868 : }
8869 : }
8870 1222 : }
8871 :
8872 :
8873 : /*
8874 : * Issue appropriate kind of fsync (if any) for an XLOG output file.
8875 : *
8876 : * 'fd' is a file descriptor for the XLOG file to be fsync'd.
8877 : * 'segno' is for error reporting purposes.
8878 : */
8879 : void
8880 201047 : issue_xlog_fsync(int fd, XLogSegNo segno, TimeLineID tli)
8881 : {
8882 201047 : char *msg = NULL;
8883 : instr_time start;
8884 :
8885 : Assert(tli != 0);
8886 :
8887 : /*
8888 : * Quick exit if fsync is disabled or write() has already synced the WAL
8889 : * file.
8890 : */
8891 201047 : if (!enableFsync ||
8892 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN ||
8893 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN_DSYNC)
8894 201047 : return;
8895 :
8896 : /*
8897 : * Measure I/O timing to sync the WAL file for pg_stat_io.
8898 : */
8899 0 : start = pgstat_prepare_io_time(track_wal_io_timing);
8900 :
8901 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC);
8902 0 : switch (wal_sync_method)
8903 : {
8904 0 : case WAL_SYNC_METHOD_FSYNC:
8905 0 : if (pg_fsync_no_writethrough(fd) != 0)
8906 0 : msg = _("could not fsync file \"%s\": %m");
8907 0 : break;
8908 : #ifdef HAVE_FSYNC_WRITETHROUGH
8909 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8910 : if (pg_fsync_writethrough(fd) != 0)
8911 : msg = _("could not fsync write-through file \"%s\": %m");
8912 : break;
8913 : #endif
8914 0 : case WAL_SYNC_METHOD_FDATASYNC:
8915 0 : if (pg_fdatasync(fd) != 0)
8916 0 : msg = _("could not fdatasync file \"%s\": %m");
8917 0 : break;
8918 0 : case WAL_SYNC_METHOD_OPEN:
8919 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8920 : /* not reachable */
8921 : Assert(false);
8922 0 : break;
8923 0 : default:
8924 0 : ereport(PANIC,
8925 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8926 : errmsg_internal("unrecognized \"wal_sync_method\": %d", wal_sync_method));
8927 : break;
8928 : }
8929 :
8930 : /* PANIC if failed to fsync */
8931 0 : if (msg)
8932 : {
8933 : char xlogfname[MAXFNAMELEN];
8934 0 : int save_errno = errno;
8935 :
8936 0 : XLogFileName(xlogfname, tli, segno, wal_segment_size);
8937 0 : errno = save_errno;
8938 0 : ereport(PANIC,
8939 : (errcode_for_file_access(),
8940 : errmsg(msg, xlogfname)));
8941 : }
8942 :
8943 0 : pgstat_report_wait_end();
8944 :
8945 0 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL, IOOP_FSYNC,
8946 : start, 1, 0);
8947 : }
8948 :
8949 : /*
8950 : * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
8951 : * function. It creates the necessary starting checkpoint and constructs the
8952 : * backup state and tablespace map.
8953 : *
8954 : * Input parameters are "state" (the backup state), "fast" (if true, we do
8955 : * the checkpoint in fast mode), and "tablespaces" (if non-NULL, indicates a
8956 : * list of tablespaceinfo structs describing the cluster's tablespaces.).
8957 : *
8958 : * The tablespace map contents are appended to passed-in parameter
8959 : * tablespace_map and the caller is responsible for including it in the backup
8960 : * archive as 'tablespace_map'. The tablespace_map file is required mainly for
8961 : * tar format in windows as native windows utilities are not able to create
8962 : * symlinks while extracting files from tar. However for consistency and
8963 : * platform-independence, we do it the same way everywhere.
8964 : *
8965 : * It fills in "state" with the information required for the backup, such
8966 : * as the minimum WAL location that must be present to restore from this
8967 : * backup (starttli) and the corresponding timeline ID (starttli).
8968 : *
8969 : * Every successfully started backup must be stopped by calling
8970 : * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
8971 : * backups active at the same time.
8972 : *
8973 : * It is the responsibility of the caller of this function to verify the
8974 : * permissions of the calling user!
8975 : */
8976 : void
8977 177 : do_pg_backup_start(const char *backupidstr, bool fast, List **tablespaces,
8978 : BackupState *state, StringInfo tblspcmapfile)
8979 : {
8980 : bool backup_started_in_recovery;
8981 :
8982 : Assert(state != NULL);
8983 177 : backup_started_in_recovery = RecoveryInProgress();
8984 :
8985 : /*
8986 : * During recovery, we don't need to check WAL level. Because, if WAL
8987 : * level is not sufficient, it's impossible to get here during recovery.
8988 : */
8989 177 : if (!backup_started_in_recovery && !XLogIsNeeded())
8990 0 : ereport(ERROR,
8991 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8992 : errmsg("WAL level not sufficient for making an online backup"),
8993 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
8994 :
8995 177 : if (strlen(backupidstr) > MAXPGPATH)
8996 1 : ereport(ERROR,
8997 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8998 : errmsg("backup label too long (max %d bytes)",
8999 : MAXPGPATH)));
9000 :
9001 176 : strlcpy(state->name, backupidstr, sizeof(state->name));
9002 :
9003 : /*
9004 : * Mark backup active in shared memory. We must do full-page WAL writes
9005 : * during an on-line backup even if not doing so at other times, because
9006 : * it's quite possible for the backup dump to obtain a "torn" (partially
9007 : * written) copy of a database page if it reads the page concurrently with
9008 : * our write to the same page. This can be fixed as long as the first
9009 : * write to the page in the WAL sequence is a full-page write. Hence, we
9010 : * increment runningBackups then force a CHECKPOINT, to ensure there are
9011 : * no dirty pages in shared memory that might get dumped while the backup
9012 : * is in progress without having a corresponding WAL record. (Once the
9013 : * backup is complete, we need not force full-page writes anymore, since
9014 : * we expect that any pages not modified during the backup interval must
9015 : * have been correctly captured by the backup.)
9016 : *
9017 : * Note that forcing full-page writes has no effect during an online
9018 : * backup from the standby.
9019 : *
9020 : * We must hold all the insertion locks to change the value of
9021 : * runningBackups, to ensure adequate interlocking against
9022 : * XLogInsertRecord().
9023 : */
9024 176 : WALInsertLockAcquireExclusive();
9025 176 : XLogCtl->Insert.runningBackups++;
9026 176 : WALInsertLockRelease();
9027 :
9028 : /*
9029 : * Ensure we decrement runningBackups if we fail below. NB -- for this to
9030 : * work correctly, it is critical that sessionBackupState is only updated
9031 : * after this block is over.
9032 : */
9033 176 : PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9034 : {
9035 176 : bool gotUniqueStartpoint = false;
9036 : DIR *tblspcdir;
9037 : struct dirent *de;
9038 : tablespaceinfo *ti;
9039 : int datadirpathlen;
9040 :
9041 : /*
9042 : * Force an XLOG file switch before the checkpoint, to ensure that the
9043 : * WAL segment the checkpoint is written to doesn't contain pages with
9044 : * old timeline IDs. That would otherwise happen if you called
9045 : * pg_backup_start() right after restoring from a PITR archive: the
9046 : * first WAL segment containing the startup checkpoint has pages in
9047 : * the beginning with the old timeline ID. That can cause trouble at
9048 : * recovery: we won't have a history file covering the old timeline if
9049 : * pg_wal directory was not included in the base backup and the WAL
9050 : * archive was cleared too before starting the backup.
9051 : *
9052 : * During recovery, we skip forcing XLOG file switch, which means that
9053 : * the backup taken during recovery is not available for the special
9054 : * recovery case described above.
9055 : */
9056 176 : if (!backup_started_in_recovery)
9057 168 : RequestXLogSwitch(false);
9058 :
9059 : do
9060 : {
9061 : bool checkpointfpw;
9062 :
9063 : /*
9064 : * Force a CHECKPOINT. Aside from being necessary to prevent torn
9065 : * page problems, this guarantees that two successive backup runs
9066 : * will have different checkpoint positions and hence different
9067 : * history file names, even if nothing happened in between.
9068 : *
9069 : * During recovery, establish a restartpoint if possible. We use
9070 : * the last restartpoint as the backup starting checkpoint. This
9071 : * means that two successive backup runs can have same checkpoint
9072 : * positions.
9073 : *
9074 : * Since the fact that we are executing do_pg_backup_start()
9075 : * during recovery means that checkpointer is running, we can use
9076 : * RequestCheckpoint() to establish a restartpoint.
9077 : *
9078 : * We use CHECKPOINT_FAST only if requested by user (via passing
9079 : * fast = true). Otherwise this can take awhile.
9080 : */
9081 176 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
9082 : (fast ? CHECKPOINT_FAST : 0));
9083 :
9084 : /*
9085 : * Now we need to fetch the checkpoint record location, and also
9086 : * its REDO pointer. The oldest point in WAL that would be needed
9087 : * to restore starting from the checkpoint is precisely the REDO
9088 : * pointer.
9089 : */
9090 176 : LWLockAcquire(ControlFileLock, LW_SHARED);
9091 176 : state->checkpointloc = ControlFile->checkPoint;
9092 176 : state->startpoint = ControlFile->checkPointCopy.redo;
9093 176 : state->starttli = ControlFile->checkPointCopy.ThisTimeLineID;
9094 176 : checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
9095 176 : LWLockRelease(ControlFileLock);
9096 :
9097 176 : if (backup_started_in_recovery)
9098 : {
9099 : XLogRecPtr recptr;
9100 :
9101 : /*
9102 : * Check to see if all WAL replayed during online backup
9103 : * (i.e., since last restartpoint used as backup starting
9104 : * checkpoint) contain full-page writes.
9105 : */
9106 8 : SpinLockAcquire(&XLogCtl->info_lck);
9107 8 : recptr = XLogCtl->lastFpwDisableRecPtr;
9108 8 : SpinLockRelease(&XLogCtl->info_lck);
9109 :
9110 8 : if (!checkpointfpw || state->startpoint <= recptr)
9111 0 : ereport(ERROR,
9112 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9113 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9114 : "since last restartpoint"),
9115 : errhint("This means that the backup being taken on the standby "
9116 : "is corrupt and should not be used. "
9117 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9118 : "and then try an online backup again.")));
9119 :
9120 : /*
9121 : * During recovery, since we don't use the end-of-backup WAL
9122 : * record and don't write the backup history file, the
9123 : * starting WAL location doesn't need to be unique. This means
9124 : * that two base backups started at the same time might use
9125 : * the same checkpoint as starting locations.
9126 : */
9127 8 : gotUniqueStartpoint = true;
9128 : }
9129 :
9130 : /*
9131 : * If two base backups are started at the same time (in WAL sender
9132 : * processes), we need to make sure that they use different
9133 : * checkpoints as starting locations, because we use the starting
9134 : * WAL location as a unique identifier for the base backup in the
9135 : * end-of-backup WAL record and when we write the backup history
9136 : * file. Perhaps it would be better generate a separate unique ID
9137 : * for each backup instead of forcing another checkpoint, but
9138 : * taking a checkpoint right after another is not that expensive
9139 : * either because only few buffers have been dirtied yet.
9140 : */
9141 176 : WALInsertLockAcquireExclusive();
9142 176 : if (XLogCtl->Insert.lastBackupStart < state->startpoint)
9143 : {
9144 176 : XLogCtl->Insert.lastBackupStart = state->startpoint;
9145 176 : gotUniqueStartpoint = true;
9146 : }
9147 176 : WALInsertLockRelease();
9148 176 : } while (!gotUniqueStartpoint);
9149 :
9150 : /*
9151 : * Construct tablespace_map file.
9152 : */
9153 176 : datadirpathlen = strlen(DataDir);
9154 :
9155 : /* Collect information about all tablespaces */
9156 176 : tblspcdir = AllocateDir(PG_TBLSPC_DIR);
9157 566 : while ((de = ReadDir(tblspcdir, PG_TBLSPC_DIR)) != NULL)
9158 : {
9159 : char fullpath[MAXPGPATH + sizeof(PG_TBLSPC_DIR)];
9160 : char linkpath[MAXPGPATH];
9161 390 : char *relpath = NULL;
9162 : char *s;
9163 : PGFileType de_type;
9164 : char *badp;
9165 : Oid tsoid;
9166 :
9167 : /*
9168 : * Try to parse the directory name as an unsigned integer.
9169 : *
9170 : * Tablespace directories should be positive integers that can be
9171 : * represented in 32 bits, with no leading zeroes or trailing
9172 : * garbage. If we come across a name that doesn't meet those
9173 : * criteria, skip it.
9174 : */
9175 390 : if (de->d_name[0] < '1' || de->d_name[1] > '9')
9176 352 : continue;
9177 38 : errno = 0;
9178 38 : tsoid = strtoul(de->d_name, &badp, 10);
9179 38 : if (*badp != '\0' || errno == EINVAL || errno == ERANGE)
9180 0 : continue;
9181 :
9182 38 : snprintf(fullpath, sizeof(fullpath), "%s/%s", PG_TBLSPC_DIR, de->d_name);
9183 :
9184 38 : de_type = get_dirent_type(fullpath, de, false, ERROR);
9185 :
9186 38 : if (de_type == PGFILETYPE_LNK)
9187 : {
9188 : StringInfoData escapedpath;
9189 : int rllen;
9190 :
9191 23 : rllen = readlink(fullpath, linkpath, sizeof(linkpath));
9192 23 : if (rllen < 0)
9193 : {
9194 0 : ereport(WARNING,
9195 : (errmsg("could not read symbolic link \"%s\": %m",
9196 : fullpath)));
9197 0 : continue;
9198 : }
9199 23 : else if (rllen >= sizeof(linkpath))
9200 : {
9201 0 : ereport(WARNING,
9202 : (errmsg("symbolic link \"%s\" target is too long",
9203 : fullpath)));
9204 0 : continue;
9205 : }
9206 23 : linkpath[rllen] = '\0';
9207 :
9208 : /*
9209 : * Relpath holds the relative path of the tablespace directory
9210 : * when it's located within PGDATA, or NULL if it's located
9211 : * elsewhere.
9212 : */
9213 23 : if (rllen > datadirpathlen &&
9214 1 : strncmp(linkpath, DataDir, datadirpathlen) == 0 &&
9215 0 : IS_DIR_SEP(linkpath[datadirpathlen]))
9216 0 : relpath = pstrdup(linkpath + datadirpathlen + 1);
9217 :
9218 : /*
9219 : * Add a backslash-escaped version of the link path to the
9220 : * tablespace map file.
9221 : */
9222 23 : initStringInfo(&escapedpath);
9223 562 : for (s = linkpath; *s; s++)
9224 : {
9225 539 : if (*s == '\n' || *s == '\r' || *s == '\\')
9226 0 : appendStringInfoChar(&escapedpath, '\\');
9227 539 : appendStringInfoChar(&escapedpath, *s);
9228 : }
9229 23 : appendStringInfo(tblspcmapfile, "%s %s\n",
9230 23 : de->d_name, escapedpath.data);
9231 23 : pfree(escapedpath.data);
9232 : }
9233 15 : else if (de_type == PGFILETYPE_DIR)
9234 : {
9235 : /*
9236 : * It's possible to use allow_in_place_tablespaces to create
9237 : * directories directly under pg_tblspc, for testing purposes
9238 : * only.
9239 : *
9240 : * In this case, we store a relative path rather than an
9241 : * absolute path into the tablespaceinfo.
9242 : */
9243 15 : snprintf(linkpath, sizeof(linkpath), "%s/%s",
9244 15 : PG_TBLSPC_DIR, de->d_name);
9245 15 : relpath = pstrdup(linkpath);
9246 : }
9247 : else
9248 : {
9249 : /* Skip any other file type that appears here. */
9250 0 : continue;
9251 : }
9252 :
9253 38 : ti = palloc_object(tablespaceinfo);
9254 38 : ti->oid = tsoid;
9255 38 : ti->path = pstrdup(linkpath);
9256 38 : ti->rpath = relpath;
9257 38 : ti->size = -1;
9258 :
9259 38 : if (tablespaces)
9260 38 : *tablespaces = lappend(*tablespaces, ti);
9261 : }
9262 176 : FreeDir(tblspcdir);
9263 :
9264 176 : state->starttime = (pg_time_t) time(NULL);
9265 : }
9266 176 : PG_END_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9267 :
9268 176 : state->started_in_recovery = backup_started_in_recovery;
9269 :
9270 : /*
9271 : * Mark that the start phase has correctly finished for the backup.
9272 : */
9273 176 : sessionBackupState = SESSION_BACKUP_RUNNING;
9274 176 : }
9275 :
9276 : /*
9277 : * Utility routine to fetch the session-level status of a backup running.
9278 : */
9279 : SessionBackupState
9280 199 : get_backup_status(void)
9281 : {
9282 199 : return sessionBackupState;
9283 : }
9284 :
9285 : /*
9286 : * do_pg_backup_stop
9287 : *
9288 : * Utility function called at the end of an online backup. It creates history
9289 : * file (if required), resets sessionBackupState and so on. It can optionally
9290 : * wait for WAL segments to be archived.
9291 : *
9292 : * "state" is filled with the information necessary to restore from this
9293 : * backup with its stop LSN (stoppoint), its timeline ID (stoptli), etc.
9294 : *
9295 : * It is the responsibility of the caller of this function to verify the
9296 : * permissions of the calling user!
9297 : */
9298 : void
9299 169 : do_pg_backup_stop(BackupState *state, bool waitforarchive)
9300 : {
9301 169 : bool backup_stopped_in_recovery = false;
9302 : char histfilepath[MAXPGPATH];
9303 : char lastxlogfilename[MAXFNAMELEN];
9304 : char histfilename[MAXFNAMELEN];
9305 : XLogSegNo _logSegNo;
9306 : FILE *fp;
9307 : int seconds_before_warning;
9308 169 : int waits = 0;
9309 169 : bool reported_waiting = false;
9310 :
9311 : Assert(state != NULL);
9312 :
9313 169 : backup_stopped_in_recovery = RecoveryInProgress();
9314 :
9315 : /*
9316 : * During recovery, we don't need to check WAL level. Because, if WAL
9317 : * level is not sufficient, it's impossible to get here during recovery.
9318 : */
9319 169 : if (!backup_stopped_in_recovery && !XLogIsNeeded())
9320 0 : ereport(ERROR,
9321 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9322 : errmsg("WAL level not sufficient for making an online backup"),
9323 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9324 :
9325 : /*
9326 : * OK to update backup counter and session-level lock.
9327 : *
9328 : * Note that CHECK_FOR_INTERRUPTS() must not occur while updating them,
9329 : * otherwise they can be updated inconsistently, which might cause
9330 : * do_pg_abort_backup() to fail.
9331 : */
9332 169 : WALInsertLockAcquireExclusive();
9333 :
9334 : /*
9335 : * It is expected that each do_pg_backup_start() call is matched by
9336 : * exactly one do_pg_backup_stop() call.
9337 : */
9338 : Assert(XLogCtl->Insert.runningBackups > 0);
9339 169 : XLogCtl->Insert.runningBackups--;
9340 :
9341 : /*
9342 : * Clean up session-level lock.
9343 : *
9344 : * You might think that WALInsertLockRelease() can be called before
9345 : * cleaning up session-level lock because session-level lock doesn't need
9346 : * to be protected with WAL insertion lock. But since
9347 : * CHECK_FOR_INTERRUPTS() can occur in it, session-level lock must be
9348 : * cleaned up before it.
9349 : */
9350 169 : sessionBackupState = SESSION_BACKUP_NONE;
9351 :
9352 169 : WALInsertLockRelease();
9353 :
9354 : /*
9355 : * If we are taking an online backup from the standby, we confirm that the
9356 : * standby has not been promoted during the backup.
9357 : */
9358 169 : if (state->started_in_recovery && !backup_stopped_in_recovery)
9359 0 : ereport(ERROR,
9360 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9361 : errmsg("the standby was promoted during online backup"),
9362 : errhint("This means that the backup being taken is corrupt "
9363 : "and should not be used. "
9364 : "Try taking another online backup.")));
9365 :
9366 : /*
9367 : * During recovery, we don't write an end-of-backup record. We assume that
9368 : * pg_control was backed up last and its minimum recovery point can be
9369 : * available as the backup end location. Since we don't have an
9370 : * end-of-backup record, we use the pg_control value to check whether
9371 : * we've reached the end of backup when starting recovery from this
9372 : * backup. We have no way of checking if pg_control wasn't backed up last
9373 : * however.
9374 : *
9375 : * We don't force a switch to new WAL file but it is still possible to
9376 : * wait for all the required files to be archived if waitforarchive is
9377 : * true. This is okay if we use the backup to start a standby and fetch
9378 : * the missing WAL using streaming replication. But in the case of an
9379 : * archive recovery, a user should set waitforarchive to true and wait for
9380 : * them to be archived to ensure that all the required files are
9381 : * available.
9382 : *
9383 : * We return the current minimum recovery point as the backup end
9384 : * location. Note that it can be greater than the exact backup end
9385 : * location if the minimum recovery point is updated after the backup of
9386 : * pg_control. This is harmless for current uses.
9387 : *
9388 : * XXX currently a backup history file is for informational and debug
9389 : * purposes only. It's not essential for an online backup. Furthermore,
9390 : * even if it's created, it will not be archived during recovery because
9391 : * an archiver is not invoked. So it doesn't seem worthwhile to write a
9392 : * backup history file during recovery.
9393 : */
9394 169 : if (backup_stopped_in_recovery)
9395 : {
9396 : XLogRecPtr recptr;
9397 :
9398 : /*
9399 : * Check to see if all WAL replayed during online backup contain
9400 : * full-page writes.
9401 : */
9402 8 : SpinLockAcquire(&XLogCtl->info_lck);
9403 8 : recptr = XLogCtl->lastFpwDisableRecPtr;
9404 8 : SpinLockRelease(&XLogCtl->info_lck);
9405 :
9406 8 : if (state->startpoint <= recptr)
9407 0 : ereport(ERROR,
9408 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9409 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9410 : "during online backup"),
9411 : errhint("This means that the backup being taken on the standby "
9412 : "is corrupt and should not be used. "
9413 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9414 : "and then try an online backup again.")));
9415 :
9416 :
9417 8 : LWLockAcquire(ControlFileLock, LW_SHARED);
9418 8 : state->stoppoint = ControlFile->minRecoveryPoint;
9419 8 : state->stoptli = ControlFile->minRecoveryPointTLI;
9420 8 : LWLockRelease(ControlFileLock);
9421 : }
9422 : else
9423 : {
9424 : char *history_file;
9425 :
9426 : /*
9427 : * Write the backup-end xlog record
9428 : */
9429 161 : XLogBeginInsert();
9430 161 : XLogRegisterData(&state->startpoint,
9431 : sizeof(state->startpoint));
9432 161 : state->stoppoint = XLogInsert(RM_XLOG_ID, XLOG_BACKUP_END);
9433 :
9434 : /*
9435 : * Given that we're not in recovery, InsertTimeLineID is set and can't
9436 : * change, so we can read it without a lock.
9437 : */
9438 161 : state->stoptli = XLogCtl->InsertTimeLineID;
9439 :
9440 : /*
9441 : * Force a switch to a new xlog segment file, so that the backup is
9442 : * valid as soon as archiver moves out the current segment file.
9443 : */
9444 161 : RequestXLogSwitch(false);
9445 :
9446 161 : state->stoptime = (pg_time_t) time(NULL);
9447 :
9448 : /*
9449 : * Write the backup history file
9450 : */
9451 161 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9452 161 : BackupHistoryFilePath(histfilepath, state->stoptli, _logSegNo,
9453 : state->startpoint, wal_segment_size);
9454 161 : fp = AllocateFile(histfilepath, "w");
9455 161 : if (!fp)
9456 0 : ereport(ERROR,
9457 : (errcode_for_file_access(),
9458 : errmsg("could not create file \"%s\": %m",
9459 : histfilepath)));
9460 :
9461 : /* Build and save the contents of the backup history file */
9462 161 : history_file = build_backup_content(state, true);
9463 161 : fprintf(fp, "%s", history_file);
9464 161 : pfree(history_file);
9465 :
9466 161 : if (fflush(fp) || ferror(fp) || FreeFile(fp))
9467 0 : ereport(ERROR,
9468 : (errcode_for_file_access(),
9469 : errmsg("could not write file \"%s\": %m",
9470 : histfilepath)));
9471 :
9472 : /*
9473 : * Clean out any no-longer-needed history files. As a side effect,
9474 : * this will post a .ready file for the newly created history file,
9475 : * notifying the archiver that history file may be archived
9476 : * immediately.
9477 : */
9478 161 : CleanupBackupHistory();
9479 : }
9480 :
9481 : /*
9482 : * If archiving is enabled, wait for all the required WAL files to be
9483 : * archived before returning. If archiving isn't enabled, the required WAL
9484 : * needs to be transported via streaming replication (hopefully with
9485 : * wal_keep_size set high enough), or some more exotic mechanism like
9486 : * polling and copying files from pg_wal with script. We have no knowledge
9487 : * of those mechanisms, so it's up to the user to ensure that he gets all
9488 : * the required WAL.
9489 : *
9490 : * We wait until both the last WAL file filled during backup and the
9491 : * history file have been archived, and assume that the alphabetic sorting
9492 : * property of the WAL files ensures any earlier WAL files are safely
9493 : * archived as well.
9494 : *
9495 : * We wait forever, since archive_command is supposed to work and we
9496 : * assume the admin wanted his backup to work completely. If you don't
9497 : * wish to wait, then either waitforarchive should be passed in as false,
9498 : * or you can set statement_timeout. Also, some notices are issued to
9499 : * clue in anyone who might be doing this interactively.
9500 : */
9501 :
9502 169 : if (waitforarchive &&
9503 12 : ((!backup_stopped_in_recovery && XLogArchivingActive()) ||
9504 1 : (backup_stopped_in_recovery && XLogArchivingAlways())))
9505 : {
9506 5 : XLByteToPrevSeg(state->stoppoint, _logSegNo, wal_segment_size);
9507 5 : XLogFileName(lastxlogfilename, state->stoptli, _logSegNo,
9508 : wal_segment_size);
9509 :
9510 5 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9511 5 : BackupHistoryFileName(histfilename, state->stoptli, _logSegNo,
9512 : state->startpoint, wal_segment_size);
9513 :
9514 5 : seconds_before_warning = 60;
9515 5 : waits = 0;
9516 :
9517 15 : while (XLogArchiveIsBusy(lastxlogfilename) ||
9518 5 : XLogArchiveIsBusy(histfilename))
9519 : {
9520 5 : CHECK_FOR_INTERRUPTS();
9521 :
9522 5 : if (!reported_waiting && waits > 5)
9523 : {
9524 0 : ereport(NOTICE,
9525 : (errmsg("base backup done, waiting for required WAL segments to be archived")));
9526 0 : reported_waiting = true;
9527 : }
9528 :
9529 5 : (void) WaitLatch(MyLatch,
9530 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
9531 : 1000L,
9532 : WAIT_EVENT_BACKUP_WAIT_WAL_ARCHIVE);
9533 5 : ResetLatch(MyLatch);
9534 :
9535 5 : if (++waits >= seconds_before_warning)
9536 : {
9537 0 : seconds_before_warning *= 2; /* This wraps in >10 years... */
9538 0 : ereport(WARNING,
9539 : (errmsg("still waiting for all required WAL segments to be archived (%d seconds elapsed)",
9540 : waits),
9541 : errhint("Check that your \"archive_command\" is executing properly. "
9542 : "You can safely cancel this backup, "
9543 : "but the database backup will not be usable without all the WAL segments.")));
9544 : }
9545 : }
9546 :
9547 5 : ereport(NOTICE,
9548 : (errmsg("all required WAL segments have been archived")));
9549 : }
9550 164 : else if (waitforarchive)
9551 7 : ereport(NOTICE,
9552 : (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
9553 169 : }
9554 :
9555 :
9556 : /*
9557 : * do_pg_abort_backup: abort a running backup
9558 : *
9559 : * This does just the most basic steps of do_pg_backup_stop(), by taking the
9560 : * system out of backup mode, thus making it a lot more safe to call from
9561 : * an error handler.
9562 : *
9563 : * 'arg' indicates that it's being called during backup setup; so
9564 : * sessionBackupState has not been modified yet, but runningBackups has
9565 : * already been incremented. When it's false, then it's invoked as a
9566 : * before_shmem_exit handler, and therefore we must not change state
9567 : * unless sessionBackupState indicates that a backup is actually running.
9568 : *
9569 : * NB: This gets used as a PG_ENSURE_ERROR_CLEANUP callback and
9570 : * before_shmem_exit handler, hence the odd-looking signature.
9571 : */
9572 : void
9573 10 : do_pg_abort_backup(int code, Datum arg)
9574 : {
9575 10 : bool during_backup_start = DatumGetBool(arg);
9576 :
9577 : /* If called during backup start, there shouldn't be one already running */
9578 : Assert(!during_backup_start || sessionBackupState == SESSION_BACKUP_NONE);
9579 :
9580 10 : if (during_backup_start || sessionBackupState != SESSION_BACKUP_NONE)
9581 : {
9582 7 : WALInsertLockAcquireExclusive();
9583 : Assert(XLogCtl->Insert.runningBackups > 0);
9584 7 : XLogCtl->Insert.runningBackups--;
9585 :
9586 7 : sessionBackupState = SESSION_BACKUP_NONE;
9587 7 : WALInsertLockRelease();
9588 :
9589 7 : if (!during_backup_start)
9590 7 : ereport(WARNING,
9591 : errmsg("aborting backup due to backend exiting before pg_backup_stop was called"));
9592 : }
9593 10 : }
9594 :
9595 : /*
9596 : * Register a handler that will warn about unterminated backups at end of
9597 : * session, unless this has already been done.
9598 : */
9599 : void
9600 5 : register_persistent_abort_backup_handler(void)
9601 : {
9602 : static bool already_done = false;
9603 :
9604 5 : if (already_done)
9605 1 : return;
9606 4 : before_shmem_exit(do_pg_abort_backup, BoolGetDatum(false));
9607 4 : already_done = true;
9608 : }
9609 :
9610 : /*
9611 : * Get latest WAL insert pointer
9612 : */
9613 : XLogRecPtr
9614 2139 : GetXLogInsertRecPtr(void)
9615 : {
9616 2139 : XLogCtlInsert *Insert = &XLogCtl->Insert;
9617 : uint64 current_bytepos;
9618 :
9619 2139 : SpinLockAcquire(&Insert->insertpos_lck);
9620 2139 : current_bytepos = Insert->CurrBytePos;
9621 2139 : SpinLockRelease(&Insert->insertpos_lck);
9622 :
9623 2139 : return XLogBytePosToRecPtr(current_bytepos);
9624 : }
9625 :
9626 : /*
9627 : * Get latest WAL record end pointer
9628 : */
9629 : XLogRecPtr
9630 100482 : GetXLogInsertEndRecPtr(void)
9631 : {
9632 100482 : XLogCtlInsert *Insert = &XLogCtl->Insert;
9633 : uint64 current_bytepos;
9634 :
9635 100482 : SpinLockAcquire(&Insert->insertpos_lck);
9636 100482 : current_bytepos = Insert->CurrBytePos;
9637 100482 : SpinLockRelease(&Insert->insertpos_lck);
9638 :
9639 100482 : return XLogBytePosToEndRecPtr(current_bytepos);
9640 : }
9641 :
9642 : /*
9643 : * Get latest WAL write pointer
9644 : */
9645 : XLogRecPtr
9646 13633 : GetXLogWriteRecPtr(void)
9647 : {
9648 13633 : RefreshXLogWriteResult(LogwrtResult);
9649 :
9650 13633 : return LogwrtResult.Write;
9651 : }
9652 :
9653 : /*
9654 : * Returns the redo pointer of the last checkpoint or restartpoint. This is
9655 : * the oldest point in WAL that we still need, if we have to restart recovery.
9656 : */
9657 : void
9658 392 : GetOldestRestartPoint(XLogRecPtr *oldrecptr, TimeLineID *oldtli)
9659 : {
9660 392 : LWLockAcquire(ControlFileLock, LW_SHARED);
9661 392 : *oldrecptr = ControlFile->checkPointCopy.redo;
9662 392 : *oldtli = ControlFile->checkPointCopy.ThisTimeLineID;
9663 392 : LWLockRelease(ControlFileLock);
9664 392 : }
9665 :
9666 : /* Thin wrapper around ShutdownWalRcv(). */
9667 : void
9668 1032 : XLogShutdownWalRcv(void)
9669 : {
9670 : Assert(AmStartupProcess() || !IsUnderPostmaster);
9671 :
9672 1032 : ShutdownWalRcv();
9673 1032 : ResetInstallXLogFileSegmentActive();
9674 1032 : }
9675 :
9676 : /* Enable WAL file recycling and preallocation. */
9677 : void
9678 1210 : SetInstallXLogFileSegmentActive(void)
9679 : {
9680 1210 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9681 1210 : XLogCtl->InstallXLogFileSegmentActive = true;
9682 1210 : LWLockRelease(ControlFileLock);
9683 1210 : }
9684 :
9685 : /* Disable WAL file recycling and preallocation. */
9686 : void
9687 1185 : ResetInstallXLogFileSegmentActive(void)
9688 : {
9689 1185 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9690 1185 : XLogCtl->InstallXLogFileSegmentActive = false;
9691 1185 : LWLockRelease(ControlFileLock);
9692 1185 : }
9693 :
9694 : bool
9695 0 : IsInstallXLogFileSegmentActive(void)
9696 : {
9697 : bool result;
9698 :
9699 0 : LWLockAcquire(ControlFileLock, LW_SHARED);
9700 0 : result = XLogCtl->InstallXLogFileSegmentActive;
9701 0 : LWLockRelease(ControlFileLock);
9702 :
9703 0 : return result;
9704 : }
9705 :
9706 : /*
9707 : * Update the WalWriterSleeping flag.
9708 : */
9709 : void
9710 595 : SetWalWriterSleeping(bool sleeping)
9711 : {
9712 595 : SpinLockAcquire(&XLogCtl->info_lck);
9713 595 : XLogCtl->WalWriterSleeping = sleeping;
9714 595 : SpinLockRelease(&XLogCtl->info_lck);
9715 595 : }
|
