1 /*-------------------------------------------------------------------------
4 * PostgreSQL write-ahead log manager
6 * The Write-Ahead Log (WAL) functionality is split into several source
7 * files, in addition to this one:
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
14 * This file contains functions for coordinating database startup and
15 * checkpointing, and managing the write-ahead log buffers when the
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.
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.
27 * In addition to those, there are many other functions for interrogating
28 * the current system state, and for starting/stopping backups.
31 * Portions Copyright (c) 1996-2023, PostgreSQL Global Development Group
32 * Portions Copyright (c) 1994, Regents of the University of California
34 * src/backend/access/transam/xlog.c
36 *-------------------------------------------------------------------------
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/xlogprefetcher.h"
63 #include "access/xlogreader.h"
64 #include "access/xlogrecovery.h"
65 #include "access/xlogutils.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"
76 #include "port/atomics.h"
77 #include "port/pg_iovec.h"
78 #include "postmaster/bgwriter.h"
79 #include "postmaster/startup.h"
80 #include "postmaster/walwriter.h"
81 #include "replication/logical.h"
82 #include "replication/origin.h"
83 #include "replication/slot.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/pmsignal.h"
93 #include "storage/predicate.h"
94 #include "storage/proc.h"
95 #include "storage/procarray.h"
96 #include "storage/reinit.h"
97 #include "storage/smgr.h"
98 #include "storage/spin.h"
99 #include "storage/sync.h"
100 #include "utils/guc_hooks.h"
101 #include "utils/guc_tables.h"
102 #include "utils/memutils.h"
103 #include "utils/ps_status.h"
104 #include "utils/relmapper.h"
105 #include "utils/pg_rusage.h"
106 #include "utils/snapmgr.h"
107 #include "utils/timeout.h"
108 #include "utils/timestamp.h"
109 #include "utils/varlena.h"
111 extern uint32 bootstrap_data_checksum_version
;
113 /* timeline ID to be used when bootstrapping */
114 #define BootstrapTimeLineID 1
116 /* User-settable parameters */
117 int max_wal_size_mb
= 1024; /* 1 GB */
118 int min_wal_size_mb
= 80; /* 80 MB */
119 int wal_keep_size_mb
= 0;
120 int XLOGbuffers
= -1;
121 int XLogArchiveTimeout
= 0;
122 int XLogArchiveMode
= ARCHIVE_MODE_OFF
;
123 char *XLogArchiveCommand
= NULL
;
124 bool EnableHotStandby
= false;
125 bool fullPageWrites
= true;
126 bool wal_log_hints
= false;
127 int wal_compression
= WAL_COMPRESSION_NONE
;
128 char *wal_consistency_checking_string
= NULL
;
129 bool *wal_consistency_checking
= NULL
;
130 bool wal_init_zero
= true;
131 bool wal_recycle
= true;
132 bool log_checkpoints
= true;
133 int sync_method
= DEFAULT_SYNC_METHOD
;
134 int wal_level
= WAL_LEVEL_REPLICA
;
135 int CommitDelay
= 0; /* precommit delay in microseconds */
136 int CommitSiblings
= 5; /* # concurrent xacts needed to sleep */
137 int wal_retrieve_retry_interval
= 5000;
138 int max_slot_wal_keep_size_mb
= -1;
139 int wal_decode_buffer_size
= 512 * 1024;
140 bool track_wal_io_timing
= false;
143 bool XLOG_DEBUG
= false;
146 int wal_segment_size
= DEFAULT_XLOG_SEG_SIZE
;
149 * Number of WAL insertion locks to use. A higher value allows more insertions
150 * to happen concurrently, but adds some CPU overhead to flushing the WAL,
151 * which needs to iterate all the locks.
153 #define NUM_XLOGINSERT_LOCKS 8
156 * Max distance from last checkpoint, before triggering a new xlog-based
159 int CheckPointSegments
;
161 /* Estimated distance between checkpoints, in bytes */
162 static double CheckPointDistanceEstimate
= 0;
163 static double PrevCheckPointDistance
= 0;
166 * Track whether there were any deferred checks for custom resource managers
167 * specified in wal_consistency_checking.
169 static bool check_wal_consistency_checking_deferred
= false;
174 const struct config_enum_entry sync_method_options
[] = {
175 {"fsync", SYNC_METHOD_FSYNC
, false},
176 #ifdef HAVE_FSYNC_WRITETHROUGH
177 {"fsync_writethrough", SYNC_METHOD_FSYNC_WRITETHROUGH
, false},
179 {"fdatasync", SYNC_METHOD_FDATASYNC
, false},
181 {"open_sync", SYNC_METHOD_OPEN
, false},
184 {"open_datasync", SYNC_METHOD_OPEN_DSYNC
, false},
191 * Although only "on", "off", and "always" are documented,
192 * we accept all the likely variants of "on" and "off".
194 const struct config_enum_entry archive_mode_options
[] = {
195 {"always", ARCHIVE_MODE_ALWAYS
, false},
196 {"on", ARCHIVE_MODE_ON
, false},
197 {"off", ARCHIVE_MODE_OFF
, false},
198 {"true", ARCHIVE_MODE_ON
, true},
199 {"false", ARCHIVE_MODE_OFF
, true},
200 {"yes", ARCHIVE_MODE_ON
, true},
201 {"no", ARCHIVE_MODE_OFF
, true},
202 {"1", ARCHIVE_MODE_ON
, true},
203 {"0", ARCHIVE_MODE_OFF
, true},
208 * Statistics for current checkpoint are collected in this global struct.
209 * Because only the checkpointer or a stand-alone backend can perform
210 * checkpoints, this will be unused in normal backends.
212 CheckpointStatsData CheckpointStats
;
215 * During recovery, lastFullPageWrites keeps track of full_page_writes that
216 * the replayed WAL records indicate. It's initialized with full_page_writes
217 * that the recovery starting checkpoint record indicates, and then updated
218 * each time XLOG_FPW_CHANGE record is replayed.
220 static bool lastFullPageWrites
;
223 * Local copy of the state tracked by SharedRecoveryState in shared memory,
224 * It is false if SharedRecoveryState is RECOVERY_STATE_DONE. True actually
225 * means "not known, need to check the shared state".
227 static bool LocalRecoveryInProgress
= true;
230 * Local state for XLogInsertAllowed():
231 * 1: unconditionally allowed to insert XLOG
232 * 0: unconditionally not allowed to insert XLOG
233 * -1: must check RecoveryInProgress(); disallow until it is false
234 * Most processes start with -1 and transition to 1 after seeing that recovery
235 * is not in progress. But we can also force the value for special cases.
236 * The coding in XLogInsertAllowed() depends on the first two of these states
237 * being numerically the same as bool true and false.
239 static int LocalXLogInsertAllowed
= -1;
242 * ProcLastRecPtr points to the start of the last XLOG record inserted by the
243 * current backend. It is updated for all inserts. XactLastRecEnd points to
244 * end+1 of the last record, and is reset when we end a top-level transaction,
245 * or start a new one; so it can be used to tell if the current transaction has
246 * created any XLOG records.
248 * While in parallel mode, this may not be fully up to date. When committing,
249 * a transaction can assume this covers all xlog records written either by the
250 * user backend or by any parallel worker which was present at any point during
251 * the transaction. But when aborting, or when still in parallel mode, other
252 * parallel backends may have written WAL records at later LSNs than the value
253 * stored here. The parallel leader advances its own copy, when necessary,
254 * in WaitForParallelWorkersToFinish.
256 XLogRecPtr ProcLastRecPtr
= InvalidXLogRecPtr
;
257 XLogRecPtr XactLastRecEnd
= InvalidXLogRecPtr
;
258 XLogRecPtr XactLastCommitEnd
= InvalidXLogRecPtr
;
261 * RedoRecPtr is this backend's local copy of the REDO record pointer
262 * (which is almost but not quite the same as a pointer to the most recent
263 * CHECKPOINT record). We update this from the shared-memory copy,
264 * XLogCtl->Insert.RedoRecPtr, whenever we can safely do so (ie, when we
265 * hold an insertion lock). See XLogInsertRecord for details. We are also
266 * allowed to update from XLogCtl->RedoRecPtr if we hold the info_lck;
269 * NB: Code that uses this variable must be prepared not only for the
270 * possibility that it may be arbitrarily out of date, but also for the
271 * possibility that it might be set to InvalidXLogRecPtr. We used to
272 * initialize it as a side effect of the first call to RecoveryInProgress(),
273 * which meant that most code that might use it could assume that it had a
274 * real if perhaps stale value. That's no longer the case.
276 static XLogRecPtr RedoRecPtr
;
279 * doPageWrites is this backend's local copy of (fullPageWrites ||
280 * runningBackups > 0). It is used together with RedoRecPtr to decide whether
281 * a full-page image of a page need to be taken.
283 * NB: Initially this is false, and there's no guarantee that it will be
284 * initialized to any other value before it is first used. Any code that
285 * makes use of it must recheck the value after obtaining a WALInsertLock,
286 * and respond appropriately if it turns out that the previous value wasn't
289 static bool doPageWrites
;
292 * Shared-memory data structures for XLOG control
294 * LogwrtRqst indicates a byte position that we need to write and/or fsync
295 * the log up to (all records before that point must be written or fsynced).
296 * LogwrtResult indicates the byte positions we have already written/fsynced.
297 * These structs are identical but are declared separately to indicate their
298 * slightly different functions.
300 * To read XLogCtl->LogwrtResult, you must hold either info_lck or
301 * WALWriteLock. To update it, you need to hold both locks. The point of
302 * this arrangement is that the value can be examined by code that already
303 * holds WALWriteLock without needing to grab info_lck as well. In addition
304 * to the shared variable, each backend has a private copy of LogwrtResult,
305 * which is updated when convenient.
307 * The request bookkeeping is simpler: there is a shared XLogCtl->LogwrtRqst
308 * (protected by info_lck), but we don't need to cache any copies of it.
310 * info_lck is only held long enough to read/update the protected variables,
311 * so it's a plain spinlock. The other locks are held longer (potentially
312 * over I/O operations), so we use LWLocks for them. These locks are:
314 * WALBufMappingLock: must be held to replace a page in the WAL buffer cache.
315 * It is only held while initializing and changing the mapping. If the
316 * contents of the buffer being replaced haven't been written yet, the mapping
317 * lock is released while the write is done, and reacquired afterwards.
319 * WALWriteLock: must be held to write WAL buffers to disk (XLogWrite or
322 * ControlFileLock: must be held to read/update control file or create
328 typedef struct XLogwrtRqst
330 XLogRecPtr Write
; /* last byte + 1 to write out */
331 XLogRecPtr Flush
; /* last byte + 1 to flush */
334 typedef struct XLogwrtResult
336 XLogRecPtr Write
; /* last byte + 1 written out */
337 XLogRecPtr Flush
; /* last byte + 1 flushed */
341 * Inserting to WAL is protected by a small fixed number of WAL insertion
342 * locks. To insert to the WAL, you must hold one of the locks - it doesn't
343 * matter which one. To lock out other concurrent insertions, you must hold
344 * of them. Each WAL insertion lock consists of a lightweight lock, plus an
345 * indicator of how far the insertion has progressed (insertingAt).
347 * The insertingAt values are read when a process wants to flush WAL from
348 * the in-memory buffers to disk, to check that all the insertions to the
349 * region the process is about to write out have finished. You could simply
350 * wait for all currently in-progress insertions to finish, but the
351 * insertingAt indicator allows you to ignore insertions to later in the WAL,
352 * so that you only wait for the insertions that are modifying the buffers
353 * you're about to write out.
355 * This isn't just an optimization. If all the WAL buffers are dirty, an
356 * inserter that's holding a WAL insert lock might need to evict an old WAL
357 * buffer, which requires flushing the WAL. If it's possible for an inserter
358 * to block on another inserter unnecessarily, deadlock can arise when two
359 * inserters holding a WAL insert lock wait for each other to finish their
362 * Small WAL records that don't cross a page boundary never update the value,
363 * the WAL record is just copied to the page and the lock is released. But
364 * to avoid the deadlock-scenario explained above, the indicator is always
365 * updated before sleeping while holding an insertion lock.
367 * lastImportantAt contains the LSN of the last important WAL record inserted
368 * using a given lock. This value is used to detect if there has been
369 * important WAL activity since the last time some action, like a checkpoint,
370 * was performed - allowing to not repeat the action if not. The LSN is
371 * updated for all insertions, unless the XLOG_MARK_UNIMPORTANT flag was
372 * set. lastImportantAt is never cleared, only overwritten by the LSN of newer
373 * records. Tracking the WAL activity directly in WALInsertLock has the
374 * advantage of not needing any additional locks to update the value.
379 XLogRecPtr insertingAt
;
380 XLogRecPtr lastImportantAt
;
384 * All the WAL insertion locks are allocated as an array in shared memory. We
385 * force the array stride to be a power of 2, which saves a few cycles in
386 * indexing, but more importantly also ensures that individual slots don't
387 * cross cache line boundaries. (Of course, we have to also ensure that the
388 * array start address is suitably aligned.)
390 typedef union WALInsertLockPadded
393 char pad
[PG_CACHE_LINE_SIZE
];
394 } WALInsertLockPadded
;
397 * Session status of running backup, used for sanity checks in SQL-callable
398 * functions to start and stop backups.
400 static SessionBackupState sessionBackupState
= SESSION_BACKUP_NONE
;
403 * Shared state data for WAL insertion.
405 typedef struct XLogCtlInsert
407 slock_t insertpos_lck
; /* protects CurrBytePos and PrevBytePos */
410 * CurrBytePos is the end of reserved WAL. The next record will be
411 * inserted at that position. PrevBytePos is the start position of the
412 * previously inserted (or rather, reserved) record - it is copied to the
413 * prev-link of the next record. These are stored as "usable byte
414 * positions" rather than XLogRecPtrs (see XLogBytePosToRecPtr()).
420 * Make sure the above heavily-contended spinlock and byte positions are
421 * on their own cache line. In particular, the RedoRecPtr and full page
422 * write variables below should be on a different cache line. They are
423 * read on every WAL insertion, but updated rarely, and we don't want
424 * those reads to steal the cache line containing Curr/PrevBytePos.
426 char pad
[PG_CACHE_LINE_SIZE
];
429 * fullPageWrites is the authoritative value used by all backends to
430 * determine whether to write full-page image to WAL. This shared value,
431 * instead of the process-local fullPageWrites, is required because, when
432 * full_page_writes is changed by SIGHUP, we must WAL-log it before it
433 * actually affects WAL-logging by backends. Checkpointer sets at startup
436 * To read these fields, you must hold an insertion lock. To modify them,
437 * you must hold ALL the locks.
439 XLogRecPtr RedoRecPtr
; /* current redo point for insertions */
443 * runningBackups is a counter indicating the number of backups currently
444 * in progress. lastBackupStart is the latest checkpoint redo location
445 * used as a starting point for an online backup.
448 XLogRecPtr lastBackupStart
;
451 * WAL insertion locks.
453 WALInsertLockPadded
*WALInsertLocks
;
457 * Total shared-memory state for XLOG.
459 typedef struct XLogCtlData
461 XLogCtlInsert Insert
;
463 /* Protected by info_lck: */
464 XLogwrtRqst LogwrtRqst
;
465 XLogRecPtr RedoRecPtr
; /* a recent copy of Insert->RedoRecPtr */
466 FullTransactionId ckptFullXid
; /* nextXid of latest checkpoint */
467 XLogRecPtr asyncXactLSN
; /* LSN of newest async commit/abort */
468 XLogRecPtr replicationSlotMinLSN
; /* oldest LSN needed by any slot */
470 XLogSegNo lastRemovedSegNo
; /* latest removed/recycled XLOG segment */
472 /* Fake LSN counter, for unlogged relations. Protected by ulsn_lck. */
473 XLogRecPtr unloggedLSN
;
476 /* Time and LSN of last xlog segment switch. Protected by WALWriteLock. */
477 pg_time_t lastSegSwitchTime
;
478 XLogRecPtr lastSegSwitchLSN
;
481 * Protected by info_lck and WALWriteLock (you must hold either lock to
482 * read it, but both to update)
484 XLogwrtResult LogwrtResult
;
487 * Latest initialized page in the cache (last byte position + 1).
489 * To change the identity of a buffer (and InitializedUpTo), you need to
490 * hold WALBufMappingLock. To change the identity of a buffer that's
491 * still dirty, the old page needs to be written out first, and for that
492 * you need WALWriteLock, and you need to ensure that there are no
493 * in-progress insertions to the page by calling
494 * WaitXLogInsertionsToFinish().
496 XLogRecPtr InitializedUpTo
;
499 * These values do not change after startup, although the pointed-to pages
500 * and xlblocks values certainly do. xlblocks values are protected by
503 char *pages
; /* buffers for unwritten XLOG pages */
504 XLogRecPtr
*xlblocks
; /* 1st byte ptr-s + XLOG_BLCKSZ */
505 int XLogCacheBlck
; /* highest allocated xlog buffer index */
508 * InsertTimeLineID is the timeline into which new WAL is being inserted
509 * and flushed. It is zero during recovery, and does not change once set.
511 * If we create a new timeline when the system was started up,
512 * PrevTimeLineID is the old timeline's ID that we forked off from.
513 * Otherwise it's equal to InsertTimeLineID.
515 TimeLineID InsertTimeLineID
;
516 TimeLineID PrevTimeLineID
;
519 * SharedRecoveryState indicates if we're still in crash or archive
520 * recovery. Protected by info_lck.
522 RecoveryState SharedRecoveryState
;
525 * InstallXLogFileSegmentActive indicates whether the checkpointer should
526 * arrange for future segments by recycling and/or PreallocXlogFiles().
527 * Protected by ControlFileLock. Only the startup process changes it. If
528 * true, anyone can use InstallXLogFileSegment(). If false, the startup
529 * process owns the exclusive right to install segments, by reading from
530 * the archive and possibly replacing existing files.
532 bool InstallXLogFileSegmentActive
;
535 * WalWriterSleeping indicates whether the WAL writer is currently in
536 * low-power mode (and hence should be nudged if an async commit occurs).
537 * Protected by info_lck.
539 bool WalWriterSleeping
;
542 * During recovery, we keep a copy of the latest checkpoint record here.
543 * lastCheckPointRecPtr points to start of checkpoint record and
544 * lastCheckPointEndPtr points to end+1 of checkpoint record. Used by the
545 * checkpointer when it wants to create a restartpoint.
547 * Protected by info_lck.
549 XLogRecPtr lastCheckPointRecPtr
;
550 XLogRecPtr lastCheckPointEndPtr
;
551 CheckPoint lastCheckPoint
;
554 * lastFpwDisableRecPtr points to the start of the last replayed
555 * XLOG_FPW_CHANGE record that instructs full_page_writes is disabled.
557 XLogRecPtr lastFpwDisableRecPtr
;
559 slock_t info_lck
; /* locks shared variables shown above */
562 static XLogCtlData
*XLogCtl
= NULL
;
564 /* a private copy of XLogCtl->Insert.WALInsertLocks, for convenience */
565 static WALInsertLockPadded
*WALInsertLocks
= NULL
;
568 * We maintain an image of pg_control in shared memory.
570 static ControlFileData
*ControlFile
= NULL
;
573 * Calculate the amount of space left on the page after 'endptr'. Beware
574 * multiple evaluation!
576 #define INSERT_FREESPACE(endptr) \
577 (((endptr) % XLOG_BLCKSZ == 0) ? 0 : (XLOG_BLCKSZ - (endptr) % XLOG_BLCKSZ))
579 /* Macro to advance to next buffer index. */
580 #define NextBufIdx(idx) \
581 (((idx) == XLogCtl->XLogCacheBlck) ? 0 : ((idx) + 1))
584 * XLogRecPtrToBufIdx returns the index of the WAL buffer that holds, or
585 * would hold if it was in cache, the page containing 'recptr'.
587 #define XLogRecPtrToBufIdx(recptr) \
588 (((recptr) / XLOG_BLCKSZ) % (XLogCtl->XLogCacheBlck + 1))
591 * These are the number of bytes in a WAL page usable for WAL data.
593 #define UsableBytesInPage (XLOG_BLCKSZ - SizeOfXLogShortPHD)
596 * Convert values of GUCs measured in megabytes to equiv. segment count.
599 #define ConvertToXSegs(x, segsize) XLogMBVarToSegs((x), (segsize))
601 /* The number of bytes in a WAL segment usable for WAL data. */
602 static int UsableBytesInSegment
;
605 * Private, possibly out-of-date copy of shared LogwrtResult.
606 * See discussion above.
608 static XLogwrtResult LogwrtResult
= {0, 0};
611 * openLogFile is -1 or a kernel FD for an open log file segment.
612 * openLogSegNo identifies the segment, and openLogTLI the corresponding TLI.
613 * These variables are only used to write the XLOG, and so will normally refer
614 * to the active segment.
616 * Note: call Reserve/ReleaseExternalFD to track consumption of this FD.
618 static int openLogFile
= -1;
619 static XLogSegNo openLogSegNo
= 0;
620 static TimeLineID openLogTLI
= 0;
623 * Local copies of equivalent fields in the control file. When running
624 * crash recovery, LocalMinRecoveryPoint is set to InvalidXLogRecPtr as we
625 * expect to replay all the WAL available, and updateMinRecoveryPoint is
626 * switched to false to prevent any updates while replaying records.
627 * Those values are kept consistent as long as crash recovery runs.
629 static XLogRecPtr LocalMinRecoveryPoint
;
630 static TimeLineID LocalMinRecoveryPointTLI
;
631 static bool updateMinRecoveryPoint
= true;
633 /* For WALInsertLockAcquire/Release functions */
634 static int MyLockNo
= 0;
635 static bool holdingAllLocks
= false;
638 static MemoryContext walDebugCxt
= NULL
;
641 static void CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI
,
644 static void CheckRequiredParameterValues(void);
645 static void XLogReportParameters(void);
646 static int LocalSetXLogInsertAllowed(void);
647 static void CreateEndOfRecoveryRecord(void);
648 static XLogRecPtr
CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn
,
651 static void CheckPointGuts(XLogRecPtr checkPointRedo
, int flags
);
652 static void KeepLogSeg(XLogRecPtr recptr
, XLogSegNo
*logSegNo
);
653 static XLogRecPtr
XLogGetReplicationSlotMinimumLSN(void);
655 static void AdvanceXLInsertBuffer(XLogRecPtr upto
, TimeLineID tli
,
657 static void XLogWrite(XLogwrtRqst WriteRqst
, TimeLineID tli
, bool flexible
);
658 static bool InstallXLogFileSegment(XLogSegNo
*segno
, char *tmppath
,
659 bool find_free
, XLogSegNo max_segno
,
661 static void XLogFileClose(void);
662 static void PreallocXlogFiles(XLogRecPtr endptr
, TimeLineID tli
);
663 static void RemoveTempXlogFiles(void);
664 static void RemoveOldXlogFiles(XLogSegNo segno
, XLogRecPtr lastredoptr
,
665 XLogRecPtr endptr
, TimeLineID insertTLI
);
666 static void RemoveXlogFile(const struct dirent
*segment_de
,
667 XLogSegNo recycleSegNo
, XLogSegNo
*endlogSegNo
,
668 TimeLineID insertTLI
);
669 static void UpdateLastRemovedPtr(char *filename
);
670 static void ValidateXLOGDirectoryStructure(void);
671 static void CleanupBackupHistory(void);
672 static void UpdateMinRecoveryPoint(XLogRecPtr lsn
, bool force
);
673 static bool PerformRecoveryXLogAction(void);
674 static void InitControlFile(uint64 sysidentifier
);
675 static void WriteControlFile(void);
676 static void ReadControlFile(void);
677 static void UpdateControlFile(void);
678 static char *str_time(pg_time_t tnow
);
680 static int get_sync_bit(int method
);
682 static void CopyXLogRecordToWAL(int write_len
, bool isLogSwitch
,
684 XLogRecPtr StartPos
, XLogRecPtr EndPos
,
686 static void ReserveXLogInsertLocation(int size
, XLogRecPtr
*StartPos
,
687 XLogRecPtr
*EndPos
, XLogRecPtr
*PrevPtr
);
688 static bool ReserveXLogSwitch(XLogRecPtr
*StartPos
, XLogRecPtr
*EndPos
,
689 XLogRecPtr
*PrevPtr
);
690 static XLogRecPtr
WaitXLogInsertionsToFinish(XLogRecPtr upto
);
691 static char *GetXLogBuffer(XLogRecPtr ptr
, TimeLineID tli
);
692 static XLogRecPtr
XLogBytePosToRecPtr(uint64 bytepos
);
693 static XLogRecPtr
XLogBytePosToEndRecPtr(uint64 bytepos
);
694 static uint64
XLogRecPtrToBytePos(XLogRecPtr ptr
);
695 static void GetOldestRestartPointFileName(char *fname
);
697 static void WALInsertLockAcquire(void);
698 static void WALInsertLockAcquireExclusive(void);
699 static void WALInsertLockRelease(void);
700 static void WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt
);
703 * Insert an XLOG record represented by an already-constructed chain of data
704 * chunks. This is a low-level routine; to construct the WAL record header
705 * and data, use the higher-level routines in xloginsert.c.
707 * If 'fpw_lsn' is valid, it is the oldest LSN among the pages that this
708 * WAL record applies to, that were not included in the record as full page
709 * images. If fpw_lsn <= RedoRecPtr, the function does not perform the
710 * insertion and returns InvalidXLogRecPtr. The caller can then recalculate
711 * which pages need a full-page image, and retry. If fpw_lsn is invalid, the
712 * record is always inserted.
714 * 'flags' gives more in-depth control on the record being inserted. See
715 * XLogSetRecordFlags() for details.
717 * 'topxid_included' tells whether the top-transaction id is logged along with
718 * current subtransaction. See XLogRecordAssemble().
720 * The first XLogRecData in the chain must be for the record header, and its
721 * data must be MAXALIGNed. XLogInsertRecord fills in the xl_prev and
722 * xl_crc fields in the header, the rest of the header must already be filled
725 * Returns XLOG pointer to end of record (beginning of next record).
726 * This can be used as LSN for data pages affected by the logged action.
727 * (LSN is the XLOG point up to which the XLOG must be flushed to disk
728 * before the data page can be written out. This implements the basic
729 * WAL rule "write the log before the data".)
732 XLogInsertRecord(XLogRecData
*rdata
,
736 bool topxid_included
)
738 XLogCtlInsert
*Insert
= &XLogCtl
->Insert
;
741 XLogRecord
*rechdr
= (XLogRecord
*) rdata
->data
;
742 uint8 info
= rechdr
->xl_info
& ~XLR_INFO_MASK
;
743 bool isLogSwitch
= (rechdr
->xl_rmid
== RM_XLOG_ID
&&
744 info
== XLOG_SWITCH
);
747 bool prevDoPageWrites
= doPageWrites
;
748 TimeLineID insertTLI
;
750 /* we assume that all of the record header is in the first chunk */
751 Assert(rdata
->len
>= SizeOfXLogRecord
);
753 /* cross-check on whether we should be here or not */
754 if (!XLogInsertAllowed())
755 elog(ERROR
, "cannot make new WAL entries during recovery");
758 * Given that we're not in recovery, InsertTimeLineID is set and can't
759 * change, so we can read it without a lock.
761 insertTLI
= XLogCtl
->InsertTimeLineID
;
765 * We have now done all the preparatory work we can without holding a
766 * lock or modifying shared state. From here on, inserting the new WAL
767 * record to the shared WAL buffer cache is a two-step process:
769 * 1. Reserve the right amount of space from the WAL. The current head of
770 * reserved space is kept in Insert->CurrBytePos, and is protected by
773 * 2. Copy the record to the reserved WAL space. This involves finding the
774 * correct WAL buffer containing the reserved space, and copying the
775 * record in place. This can be done concurrently in multiple processes.
777 * To keep track of which insertions are still in-progress, each concurrent
778 * inserter acquires an insertion lock. In addition to just indicating that
779 * an insertion is in progress, the lock tells others how far the inserter
780 * has progressed. There is a small fixed number of insertion locks,
781 * determined by NUM_XLOGINSERT_LOCKS. When an inserter crosses a page
782 * boundary, it updates the value stored in the lock to the how far it has
783 * inserted, to allow the previous buffer to be flushed.
785 * Holding onto an insertion lock also protects RedoRecPtr and
786 * fullPageWrites from changing until the insertion is finished.
788 * Step 2 can usually be done completely in parallel. If the required WAL
789 * page is not initialized yet, you have to grab WALBufMappingLock to
790 * initialize it, but the WAL writer tries to do that ahead of insertions
791 * to avoid that from happening in the critical path.
795 START_CRIT_SECTION();
797 WALInsertLockAcquireExclusive();
799 WALInsertLockAcquire();
802 * Check to see if my copy of RedoRecPtr is out of date. If so, may have
803 * to go back and have the caller recompute everything. This can only
804 * happen just after a checkpoint, so it's better to be slow in this case
805 * and fast otherwise.
807 * Also check to see if fullPageWrites was just turned on or there's a
808 * running backup (which forces full-page writes); if we weren't already
809 * doing full-page writes then go back and recompute.
811 * If we aren't doing full-page writes then RedoRecPtr doesn't actually
812 * affect the contents of the XLOG record, so we'll update our local copy
813 * but not force a recomputation. (If doPageWrites was just turned off,
814 * we could recompute the record without full pages, but we choose not to
817 if (RedoRecPtr
!= Insert
->RedoRecPtr
)
819 Assert(RedoRecPtr
< Insert
->RedoRecPtr
);
820 RedoRecPtr
= Insert
->RedoRecPtr
;
822 doPageWrites
= (Insert
->fullPageWrites
|| Insert
->runningBackups
> 0);
825 (!prevDoPageWrites
||
826 (fpw_lsn
!= InvalidXLogRecPtr
&& fpw_lsn
<= RedoRecPtr
)))
829 * Oops, some buffer now needs to be backed up that the caller didn't
830 * back up. Start over.
832 WALInsertLockRelease();
834 return InvalidXLogRecPtr
;
838 * Reserve space for the record in the WAL. This also sets the xl_prev
842 inserted
= ReserveXLogSwitch(&StartPos
, &EndPos
, &rechdr
->xl_prev
);
845 ReserveXLogInsertLocation(rechdr
->xl_tot_len
, &StartPos
, &EndPos
,
853 * Now that xl_prev has been filled in, calculate CRC of the record
856 rdata_crc
= rechdr
->xl_crc
;
857 COMP_CRC32C(rdata_crc
, rechdr
, offsetof(XLogRecord
, xl_crc
));
858 FIN_CRC32C(rdata_crc
);
859 rechdr
->xl_crc
= rdata_crc
;
862 * All the record data, including the header, is now ready to be
863 * inserted. Copy the record in the space reserved.
865 CopyXLogRecordToWAL(rechdr
->xl_tot_len
, isLogSwitch
, rdata
,
866 StartPos
, EndPos
, insertTLI
);
869 * Unless record is flagged as not important, update LSN of last
870 * important record in the current slot. When holding all locks, just
871 * update the first one.
873 if ((flags
& XLOG_MARK_UNIMPORTANT
) == 0)
875 int lockno
= holdingAllLocks
? 0 : MyLockNo
;
877 WALInsertLocks
[lockno
].l
.lastImportantAt
= StartPos
;
883 * This was an xlog-switch record, but the current insert location was
884 * already exactly at the beginning of a segment, so there was no need
890 * Done! Let others know that we're finished.
892 WALInsertLockRelease();
896 MarkCurrentTransactionIdLoggedIfAny();
899 * Mark top transaction id is logged (if needed) so that we should not try
900 * to log it again with the next WAL record in the current subtransaction.
903 MarkSubxactTopXidLogged();
906 * Update shared LogwrtRqst.Write, if we crossed page boundary.
908 if (StartPos
/ XLOG_BLCKSZ
!= EndPos
/ XLOG_BLCKSZ
)
910 SpinLockAcquire(&XLogCtl
->info_lck
);
911 /* advance global request to include new block(s) */
912 if (XLogCtl
->LogwrtRqst
.Write
< EndPos
)
913 XLogCtl
->LogwrtRqst
.Write
= EndPos
;
914 /* update local result copy while I have the chance */
915 LogwrtResult
= XLogCtl
->LogwrtResult
;
916 SpinLockRelease(&XLogCtl
->info_lck
);
920 * If this was an XLOG_SWITCH record, flush the record and the empty
921 * padding space that fills the rest of the segment, and perform
922 * end-of-segment actions (eg, notifying archiver).
926 TRACE_POSTGRESQL_WAL_SWITCH();
930 * Even though we reserved the rest of the segment for us, which is
931 * reflected in EndPos, we return a pointer to just the end of the
932 * xlog-switch record.
936 EndPos
= StartPos
+ SizeOfXLogRecord
;
937 if (StartPos
/ XLOG_BLCKSZ
!= EndPos
/ XLOG_BLCKSZ
)
939 uint64 offset
= XLogSegmentOffset(EndPos
, wal_segment_size
);
941 if (offset
== EndPos
% XLOG_BLCKSZ
)
942 EndPos
+= SizeOfXLogLongPHD
;
944 EndPos
+= SizeOfXLogShortPHD
;
952 static XLogReaderState
*debug_reader
= NULL
;
954 DecodedXLogRecord
*decoded
;
956 StringInfoData recordBuf
;
957 char *errormsg
= NULL
;
958 MemoryContext oldCxt
;
960 oldCxt
= MemoryContextSwitchTo(walDebugCxt
);
962 initStringInfo(&buf
);
963 appendStringInfo(&buf
, "INSERT @ %X/%X: ", LSN_FORMAT_ARGS(EndPos
));
966 * We have to piece together the WAL record data from the XLogRecData
967 * entries, so that we can pass it to the rm_desc function as one
970 initStringInfo(&recordBuf
);
971 for (; rdata
!= NULL
; rdata
= rdata
->next
)
972 appendBinaryStringInfo(&recordBuf
, rdata
->data
, rdata
->len
);
974 /* We also need temporary space to decode the record. */
975 record
= (XLogRecord
*) recordBuf
.data
;
976 decoded
= (DecodedXLogRecord
*)
977 palloc(DecodeXLogRecordRequiredSpace(record
->xl_tot_len
));
980 debug_reader
= XLogReaderAllocate(wal_segment_size
, NULL
,
985 appendStringInfoString(&buf
, "error decoding record: out of memory while allocating a WAL reading processor");
987 else if (!DecodeXLogRecord(debug_reader
,
993 appendStringInfo(&buf
, "error decoding record: %s",
994 errormsg
? errormsg
: "no error message");
998 appendStringInfoString(&buf
, " - ");
1000 debug_reader
->record
= decoded
;
1001 xlog_outdesc(&buf
, debug_reader
);
1002 debug_reader
->record
= NULL
;
1004 elog(LOG
, "%s", buf
.data
);
1008 pfree(recordBuf
.data
);
1009 MemoryContextSwitchTo(oldCxt
);
1014 * Update our global variables
1016 ProcLastRecPtr
= StartPos
;
1017 XactLastRecEnd
= EndPos
;
1019 /* Report WAL traffic to the instrumentation. */
1022 pgWalUsage
.wal_bytes
+= rechdr
->xl_tot_len
;
1023 pgWalUsage
.wal_records
++;
1024 pgWalUsage
.wal_fpi
+= num_fpi
;
1031 * Reserves the right amount of space for a record of given size from the WAL.
1032 * *StartPos is set to the beginning of the reserved section, *EndPos to
1033 * its end+1. *PrevPtr is set to the beginning of the previous record; it is
1034 * used to set the xl_prev of this record.
1036 * This is the performance critical part of XLogInsert that must be serialized
1037 * across backends. The rest can happen mostly in parallel. Try to keep this
1038 * section as short as possible, insertpos_lck can be heavily contended on a
1041 * NB: The space calculation here must match the code in CopyXLogRecordToWAL,
1042 * where we actually copy the record to the reserved space.
1045 ReserveXLogInsertLocation(int size
, XLogRecPtr
*StartPos
, XLogRecPtr
*EndPos
,
1046 XLogRecPtr
*PrevPtr
)
1048 XLogCtlInsert
*Insert
= &XLogCtl
->Insert
;
1049 uint64 startbytepos
;
1053 size
= MAXALIGN(size
);
1055 /* All (non xlog-switch) records should contain data. */
1056 Assert(size
> SizeOfXLogRecord
);
1059 * The duration the spinlock needs to be held is minimized by minimizing
1060 * the calculations that have to be done while holding the lock. The
1061 * current tip of reserved WAL is kept in CurrBytePos, as a byte position
1062 * that only counts "usable" bytes in WAL, that is, it excludes all WAL
1063 * page headers. The mapping between "usable" byte positions and physical
1064 * positions (XLogRecPtrs) can be done outside the locked region, and
1065 * because the usable byte position doesn't include any headers, reserving
1066 * X bytes from WAL is almost as simple as "CurrBytePos += X".
1068 SpinLockAcquire(&Insert
->insertpos_lck
);
1070 startbytepos
= Insert
->CurrBytePos
;
1071 endbytepos
= startbytepos
+ size
;
1072 prevbytepos
= Insert
->PrevBytePos
;
1073 Insert
->CurrBytePos
= endbytepos
;
1074 Insert
->PrevBytePos
= startbytepos
;
1076 SpinLockRelease(&Insert
->insertpos_lck
);
1078 *StartPos
= XLogBytePosToRecPtr(startbytepos
);
1079 *EndPos
= XLogBytePosToEndRecPtr(endbytepos
);
1080 *PrevPtr
= XLogBytePosToRecPtr(prevbytepos
);
1083 * Check that the conversions between "usable byte positions" and
1084 * XLogRecPtrs work consistently in both directions.
1086 Assert(XLogRecPtrToBytePos(*StartPos
) == startbytepos
);
1087 Assert(XLogRecPtrToBytePos(*EndPos
) == endbytepos
);
1088 Assert(XLogRecPtrToBytePos(*PrevPtr
) == prevbytepos
);
1092 * Like ReserveXLogInsertLocation(), but for an xlog-switch record.
1094 * A log-switch record is handled slightly differently. The rest of the
1095 * segment will be reserved for this insertion, as indicated by the returned
1096 * *EndPos value. However, if we are already at the beginning of the current
1097 * segment, *StartPos and *EndPos are set to the current location without
1098 * reserving any space, and the function returns false.
1101 ReserveXLogSwitch(XLogRecPtr
*StartPos
, XLogRecPtr
*EndPos
, XLogRecPtr
*PrevPtr
)
1103 XLogCtlInsert
*Insert
= &XLogCtl
->Insert
;
1104 uint64 startbytepos
;
1107 uint32 size
= MAXALIGN(SizeOfXLogRecord
);
1112 * These calculations are a bit heavy-weight to be done while holding a
1113 * spinlock, but since we're holding all the WAL insertion locks, there
1114 * are no other inserters competing for it. GetXLogInsertRecPtr() does
1115 * compete for it, but that's not called very frequently.
1117 SpinLockAcquire(&Insert
->insertpos_lck
);
1119 startbytepos
= Insert
->CurrBytePos
;
1121 ptr
= XLogBytePosToEndRecPtr(startbytepos
);
1122 if (XLogSegmentOffset(ptr
, wal_segment_size
) == 0)
1124 SpinLockRelease(&Insert
->insertpos_lck
);
1125 *EndPos
= *StartPos
= ptr
;
1129 endbytepos
= startbytepos
+ size
;
1130 prevbytepos
= Insert
->PrevBytePos
;
1132 *StartPos
= XLogBytePosToRecPtr(startbytepos
);
1133 *EndPos
= XLogBytePosToEndRecPtr(endbytepos
);
1135 segleft
= wal_segment_size
- XLogSegmentOffset(*EndPos
, wal_segment_size
);
1136 if (segleft
!= wal_segment_size
)
1138 /* consume the rest of the segment */
1140 endbytepos
= XLogRecPtrToBytePos(*EndPos
);
1142 Insert
->CurrBytePos
= endbytepos
;
1143 Insert
->PrevBytePos
= startbytepos
;
1145 SpinLockRelease(&Insert
->insertpos_lck
);
1147 *PrevPtr
= XLogBytePosToRecPtr(prevbytepos
);
1149 Assert(XLogSegmentOffset(*EndPos
, wal_segment_size
) == 0);
1150 Assert(XLogRecPtrToBytePos(*EndPos
) == endbytepos
);
1151 Assert(XLogRecPtrToBytePos(*StartPos
) == startbytepos
);
1152 Assert(XLogRecPtrToBytePos(*PrevPtr
) == prevbytepos
);
1158 * Subroutine of XLogInsertRecord. Copies a WAL record to an already-reserved
1162 CopyXLogRecordToWAL(int write_len
, bool isLogSwitch
, XLogRecData
*rdata
,
1163 XLogRecPtr StartPos
, XLogRecPtr EndPos
, TimeLineID tli
)
1169 XLogPageHeader pagehdr
;
1172 * Get a pointer to the right place in the right WAL buffer to start
1176 currpos
= GetXLogBuffer(CurrPos
, tli
);
1177 freespace
= INSERT_FREESPACE(CurrPos
);
1180 * there should be enough space for at least the first field (xl_tot_len)
1183 Assert(freespace
>= sizeof(uint32
));
1185 /* Copy record data */
1187 while (rdata
!= NULL
)
1189 char *rdata_data
= rdata
->data
;
1190 int rdata_len
= rdata
->len
;
1192 while (rdata_len
> freespace
)
1195 * Write what fits on this page, and continue on the next page.
1197 Assert(CurrPos
% XLOG_BLCKSZ
>= SizeOfXLogShortPHD
|| freespace
== 0);
1198 memcpy(currpos
, rdata_data
, freespace
);
1199 rdata_data
+= freespace
;
1200 rdata_len
-= freespace
;
1201 written
+= freespace
;
1202 CurrPos
+= freespace
;
1205 * Get pointer to beginning of next page, and set the xlp_rem_len
1206 * in the page header. Set XLP_FIRST_IS_CONTRECORD.
1208 * It's safe to set the contrecord flag and xlp_rem_len without a
1209 * lock on the page. All the other flags were already set when the
1210 * page was initialized, in AdvanceXLInsertBuffer, and we're the
1211 * only backend that needs to set the contrecord flag.
1213 currpos
= GetXLogBuffer(CurrPos
, tli
);
1214 pagehdr
= (XLogPageHeader
) currpos
;
1215 pagehdr
->xlp_rem_len
= write_len
- written
;
1216 pagehdr
->xlp_info
|= XLP_FIRST_IS_CONTRECORD
;
1218 /* skip over the page header */
1219 if (XLogSegmentOffset(CurrPos
, wal_segment_size
) == 0)
1221 CurrPos
+= SizeOfXLogLongPHD
;
1222 currpos
+= SizeOfXLogLongPHD
;
1226 CurrPos
+= SizeOfXLogShortPHD
;
1227 currpos
+= SizeOfXLogShortPHD
;
1229 freespace
= INSERT_FREESPACE(CurrPos
);
1232 Assert(CurrPos
% XLOG_BLCKSZ
>= SizeOfXLogShortPHD
|| rdata_len
== 0);
1233 memcpy(currpos
, rdata_data
, rdata_len
);
1234 currpos
+= rdata_len
;
1235 CurrPos
+= rdata_len
;
1236 freespace
-= rdata_len
;
1237 written
+= rdata_len
;
1239 rdata
= rdata
->next
;
1241 Assert(written
== write_len
);
1244 * If this was an xlog-switch, it's not enough to write the switch record,
1245 * we also have to consume all the remaining space in the WAL segment. We
1246 * have already reserved that space, but we need to actually fill it.
1248 if (isLogSwitch
&& XLogSegmentOffset(CurrPos
, wal_segment_size
) != 0)
1250 /* An xlog-switch record doesn't contain any data besides the header */
1251 Assert(write_len
== SizeOfXLogRecord
);
1253 /* Assert that we did reserve the right amount of space */
1254 Assert(XLogSegmentOffset(EndPos
, wal_segment_size
) == 0);
1256 /* Use up all the remaining space on the current page */
1257 CurrPos
+= freespace
;
1260 * Cause all remaining pages in the segment to be flushed, leaving the
1261 * XLog position where it should be, at the start of the next segment.
1262 * We do this one page at a time, to make sure we don't deadlock
1263 * against ourselves if wal_buffers < wal_segment_size.
1265 while (CurrPos
< EndPos
)
1268 * The minimal action to flush the page would be to call
1269 * WALInsertLockUpdateInsertingAt(CurrPos) followed by
1270 * AdvanceXLInsertBuffer(...). The page would be left initialized
1271 * mostly to zeros, except for the page header (always the short
1272 * variant, as this is never a segment's first page).
1274 * The large vistas of zeros are good for compressibility, but the
1275 * headers interrupting them every XLOG_BLCKSZ (with values that
1276 * differ from page to page) are not. The effect varies with
1277 * compression tool, but bzip2 for instance compresses about an
1278 * order of magnitude worse if those headers are left in place.
1280 * Rather than complicating AdvanceXLInsertBuffer itself (which is
1281 * called in heavily-loaded circumstances as well as this lightly-
1282 * loaded one) with variant behavior, we just use GetXLogBuffer
1283 * (which itself calls the two methods we need) to get the pointer
1284 * and zero most of the page. Then we just zero the page header.
1286 currpos
= GetXLogBuffer(CurrPos
, tli
);
1287 MemSet(currpos
, 0, SizeOfXLogShortPHD
);
1289 CurrPos
+= XLOG_BLCKSZ
;
1294 /* Align the end position, so that the next record starts aligned */
1295 CurrPos
= MAXALIGN64(CurrPos
);
1298 if (CurrPos
!= EndPos
)
1299 elog(PANIC
, "space reserved for WAL record does not match what was written");
1303 * Acquire a WAL insertion lock, for inserting to WAL.
1306 WALInsertLockAcquire(void)
1311 * It doesn't matter which of the WAL insertion locks we acquire, so try
1312 * the one we used last time. If the system isn't particularly busy, it's
1313 * a good bet that it's still available, and it's good to have some
1314 * affinity to a particular lock so that you don't unnecessarily bounce
1315 * cache lines between processes when there's no contention.
1317 * If this is the first time through in this backend, pick a lock
1318 * (semi-)randomly. This allows the locks to be used evenly if you have a
1319 * lot of very short connections.
1321 static int lockToTry
= -1;
1323 if (lockToTry
== -1)
1324 lockToTry
= MyProc
->pgprocno
% NUM_XLOGINSERT_LOCKS
;
1325 MyLockNo
= lockToTry
;
1328 * The insertingAt value is initially set to 0, as we don't know our
1329 * insert location yet.
1331 immed
= LWLockAcquire(&WALInsertLocks
[MyLockNo
].l
.lock
, LW_EXCLUSIVE
);
1335 * If we couldn't get the lock immediately, try another lock next
1336 * time. On a system with more insertion locks than concurrent
1337 * inserters, this causes all the inserters to eventually migrate to a
1338 * lock that no-one else is using. On a system with more inserters
1339 * than locks, it still helps to distribute the inserters evenly
1342 lockToTry
= (lockToTry
+ 1) % NUM_XLOGINSERT_LOCKS
;
1347 * Acquire all WAL insertion locks, to prevent other backends from inserting
1351 WALInsertLockAcquireExclusive(void)
1356 * When holding all the locks, all but the last lock's insertingAt
1357 * indicator is set to 0xFFFFFFFFFFFFFFFF, which is higher than any real
1358 * XLogRecPtr value, to make sure that no-one blocks waiting on those.
1360 for (i
= 0; i
< NUM_XLOGINSERT_LOCKS
- 1; i
++)
1362 LWLockAcquire(&WALInsertLocks
[i
].l
.lock
, LW_EXCLUSIVE
);
1363 LWLockUpdateVar(&WALInsertLocks
[i
].l
.lock
,
1364 &WALInsertLocks
[i
].l
.insertingAt
,
1367 /* Variable value reset to 0 at release */
1368 LWLockAcquire(&WALInsertLocks
[i
].l
.lock
, LW_EXCLUSIVE
);
1370 holdingAllLocks
= true;
1374 * Release our insertion lock (or locks, if we're holding them all).
1376 * NB: Reset all variables to 0, so they cause LWLockWaitForVar to block the
1377 * next time the lock is acquired.
1380 WALInsertLockRelease(void)
1382 if (holdingAllLocks
)
1386 for (i
= 0; i
< NUM_XLOGINSERT_LOCKS
; i
++)
1387 LWLockReleaseClearVar(&WALInsertLocks
[i
].l
.lock
,
1388 &WALInsertLocks
[i
].l
.insertingAt
,
1391 holdingAllLocks
= false;
1395 LWLockReleaseClearVar(&WALInsertLocks
[MyLockNo
].l
.lock
,
1396 &WALInsertLocks
[MyLockNo
].l
.insertingAt
,
1402 * Update our insertingAt value, to let others know that we've finished
1403 * inserting up to that point.
1406 WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt
)
1408 if (holdingAllLocks
)
1411 * We use the last lock to mark our actual position, see comments in
1412 * WALInsertLockAcquireExclusive.
1414 LWLockUpdateVar(&WALInsertLocks
[NUM_XLOGINSERT_LOCKS
- 1].l
.lock
,
1415 &WALInsertLocks
[NUM_XLOGINSERT_LOCKS
- 1].l
.insertingAt
,
1419 LWLockUpdateVar(&WALInsertLocks
[MyLockNo
].l
.lock
,
1420 &WALInsertLocks
[MyLockNo
].l
.insertingAt
,
1425 * Wait for any WAL insertions < upto to finish.
1427 * Returns the location of the oldest insertion that is still in-progress.
1428 * Any WAL prior to that point has been fully copied into WAL buffers, and
1429 * can be flushed out to disk. Because this waits for any insertions older
1430 * than 'upto' to finish, the return value is always >= 'upto'.
1432 * Note: When you are about to write out WAL, you must call this function
1433 * *before* acquiring WALWriteLock, to avoid deadlocks. This function might
1434 * need to wait for an insertion to finish (or at least advance to next
1435 * uninitialized page), and the inserter might need to evict an old WAL buffer
1436 * to make room for a new one, which in turn requires WALWriteLock.
1439 WaitXLogInsertionsToFinish(XLogRecPtr upto
)
1442 XLogRecPtr reservedUpto
;
1443 XLogRecPtr finishedUpto
;
1444 XLogCtlInsert
*Insert
= &XLogCtl
->Insert
;
1448 elog(PANIC
, "cannot wait without a PGPROC structure");
1450 /* Read the current insert position */
1451 SpinLockAcquire(&Insert
->insertpos_lck
);
1452 bytepos
= Insert
->CurrBytePos
;
1453 SpinLockRelease(&Insert
->insertpos_lck
);
1454 reservedUpto
= XLogBytePosToEndRecPtr(bytepos
);
1457 * No-one should request to flush a piece of WAL that hasn't even been
1458 * reserved yet. However, it can happen if there is a block with a bogus
1459 * LSN on disk, for example. XLogFlush checks for that situation and
1460 * complains, but only after the flush. Here we just assume that to mean
1461 * that all WAL that has been reserved needs to be finished. In this
1462 * corner-case, the return value can be smaller than 'upto' argument.
1464 if (upto
> reservedUpto
)
1467 (errmsg("request to flush past end of generated WAL; request %X/%X, current position %X/%X",
1468 LSN_FORMAT_ARGS(upto
), LSN_FORMAT_ARGS(reservedUpto
))));
1469 upto
= reservedUpto
;
1473 * Loop through all the locks, sleeping on any in-progress insert older
1476 * finishedUpto is our return value, indicating the point upto which all
1477 * the WAL insertions have been finished. Initialize it to the head of
1478 * reserved WAL, and as we iterate through the insertion locks, back it
1479 * out for any insertion that's still in progress.
1481 finishedUpto
= reservedUpto
;
1482 for (i
= 0; i
< NUM_XLOGINSERT_LOCKS
; i
++)
1484 XLogRecPtr insertingat
= InvalidXLogRecPtr
;
1489 * See if this insertion is in progress. LWLockWaitForVar will
1490 * wait for the lock to be released, or for the 'value' to be set
1491 * by a LWLockUpdateVar call. When a lock is initially acquired,
1492 * its value is 0 (InvalidXLogRecPtr), which means that we don't
1493 * know where it's inserting yet. We will have to wait for it. If
1494 * it's a small insertion, the record will most likely fit on the
1495 * same page and the inserter will release the lock without ever
1496 * calling LWLockUpdateVar. But if it has to sleep, it will
1497 * advertise the insertion point with LWLockUpdateVar before
1500 if (LWLockWaitForVar(&WALInsertLocks
[i
].l
.lock
,
1501 &WALInsertLocks
[i
].l
.insertingAt
,
1502 insertingat
, &insertingat
))
1504 /* the lock was free, so no insertion in progress */
1505 insertingat
= InvalidXLogRecPtr
;
1510 * This insertion is still in progress. Have to wait, unless the
1511 * inserter has proceeded past 'upto'.
1513 } while (insertingat
< upto
);
1515 if (insertingat
!= InvalidXLogRecPtr
&& insertingat
< finishedUpto
)
1516 finishedUpto
= insertingat
;
1518 return finishedUpto
;
1522 * Get a pointer to the right location in the WAL buffer containing the
1525 * If the page is not initialized yet, it is initialized. That might require
1526 * evicting an old dirty buffer from the buffer cache, which means I/O.
1528 * The caller must ensure that the page containing the requested location
1529 * isn't evicted yet, and won't be evicted. The way to ensure that is to
1530 * hold onto a WAL insertion lock with the insertingAt position set to
1531 * something <= ptr. GetXLogBuffer() will update insertingAt if it needs
1532 * to evict an old page from the buffer. (This means that once you call
1533 * GetXLogBuffer() with a given 'ptr', you must not access anything before
1534 * that point anymore, and must not call GetXLogBuffer() with an older 'ptr'
1535 * later, because older buffers might be recycled already)
1538 GetXLogBuffer(XLogRecPtr ptr
, TimeLineID tli
)
1542 static uint64 cachedPage
= 0;
1543 static char *cachedPos
= NULL
;
1544 XLogRecPtr expectedEndPtr
;
1547 * Fast path for the common case that we need to access again the same
1548 * page as last time.
1550 if (ptr
/ XLOG_BLCKSZ
== cachedPage
)
1552 Assert(((XLogPageHeader
) cachedPos
)->xlp_magic
== XLOG_PAGE_MAGIC
);
1553 Assert(((XLogPageHeader
) cachedPos
)->xlp_pageaddr
== ptr
- (ptr
% XLOG_BLCKSZ
));
1554 return cachedPos
+ ptr
% XLOG_BLCKSZ
;
1558 * The XLog buffer cache is organized so that a page is always loaded to a
1559 * particular buffer. That way we can easily calculate the buffer a given
1560 * page must be loaded into, from the XLogRecPtr alone.
1562 idx
= XLogRecPtrToBufIdx(ptr
);
1565 * See what page is loaded in the buffer at the moment. It could be the
1566 * page we're looking for, or something older. It can't be anything newer
1567 * - that would imply the page we're looking for has already been written
1568 * out to disk and evicted, and the caller is responsible for making sure
1569 * that doesn't happen.
1571 * However, we don't hold a lock while we read the value. If someone has
1572 * just initialized the page, it's possible that we get a "torn read" of
1573 * the XLogRecPtr if 64-bit fetches are not atomic on this platform. In
1574 * that case we will see a bogus value. That's ok, we'll grab the mapping
1575 * lock (in AdvanceXLInsertBuffer) and retry if we see anything else than
1576 * the page we're looking for. But it means that when we do this unlocked
1577 * read, we might see a value that appears to be ahead of the page we're
1578 * looking for. Don't PANIC on that, until we've verified the value while
1581 expectedEndPtr
= ptr
;
1582 expectedEndPtr
+= XLOG_BLCKSZ
- ptr
% XLOG_BLCKSZ
;
1584 endptr
= XLogCtl
->xlblocks
[idx
];
1585 if (expectedEndPtr
!= endptr
)
1587 XLogRecPtr initializedUpto
;
1590 * Before calling AdvanceXLInsertBuffer(), which can block, let others
1591 * know how far we're finished with inserting the record.
1593 * NB: If 'ptr' points to just after the page header, advertise a
1594 * position at the beginning of the page rather than 'ptr' itself. If
1595 * there are no other insertions running, someone might try to flush
1596 * up to our advertised location. If we advertised a position after
1597 * the page header, someone might try to flush the page header, even
1598 * though page might actually not be initialized yet. As the first
1599 * inserter on the page, we are effectively responsible for making
1600 * sure that it's initialized, before we let insertingAt to move past
1603 if (ptr
% XLOG_BLCKSZ
== SizeOfXLogShortPHD
&&
1604 XLogSegmentOffset(ptr
, wal_segment_size
) > XLOG_BLCKSZ
)
1605 initializedUpto
= ptr
- SizeOfXLogShortPHD
;
1606 else if (ptr
% XLOG_BLCKSZ
== SizeOfXLogLongPHD
&&
1607 XLogSegmentOffset(ptr
, wal_segment_size
) < XLOG_BLCKSZ
)
1608 initializedUpto
= ptr
- SizeOfXLogLongPHD
;
1610 initializedUpto
= ptr
;
1612 WALInsertLockUpdateInsertingAt(initializedUpto
);
1614 AdvanceXLInsertBuffer(ptr
, tli
, false);
1615 endptr
= XLogCtl
->xlblocks
[idx
];
1617 if (expectedEndPtr
!= endptr
)
1618 elog(PANIC
, "could not find WAL buffer for %X/%X",
1619 LSN_FORMAT_ARGS(ptr
));
1624 * Make sure the initialization of the page is visible to us, and
1625 * won't arrive later to overwrite the WAL data we write on the page.
1627 pg_memory_barrier();
1631 * Found the buffer holding this page. Return a pointer to the right
1632 * offset within the page.
1634 cachedPage
= ptr
/ XLOG_BLCKSZ
;
1635 cachedPos
= XLogCtl
->pages
+ idx
* (Size
) XLOG_BLCKSZ
;
1637 Assert(((XLogPageHeader
) cachedPos
)->xlp_magic
== XLOG_PAGE_MAGIC
);
1638 Assert(((XLogPageHeader
) cachedPos
)->xlp_pageaddr
== ptr
- (ptr
% XLOG_BLCKSZ
));
1640 return cachedPos
+ ptr
% XLOG_BLCKSZ
;
1644 * Converts a "usable byte position" to XLogRecPtr. A usable byte position
1645 * is the position starting from the beginning of WAL, excluding all WAL
1649 XLogBytePosToRecPtr(uint64 bytepos
)
1657 fullsegs
= bytepos
/ UsableBytesInSegment
;
1658 bytesleft
= bytepos
% UsableBytesInSegment
;
1660 if (bytesleft
< XLOG_BLCKSZ
- SizeOfXLogLongPHD
)
1662 /* fits on first page of segment */
1663 seg_offset
= bytesleft
+ SizeOfXLogLongPHD
;
1667 /* account for the first page on segment with long header */
1668 seg_offset
= XLOG_BLCKSZ
;
1669 bytesleft
-= XLOG_BLCKSZ
- SizeOfXLogLongPHD
;
1671 fullpages
= bytesleft
/ UsableBytesInPage
;
1672 bytesleft
= bytesleft
% UsableBytesInPage
;
1674 seg_offset
+= fullpages
* XLOG_BLCKSZ
+ bytesleft
+ SizeOfXLogShortPHD
;
1677 XLogSegNoOffsetToRecPtr(fullsegs
, seg_offset
, wal_segment_size
, result
);
1683 * Like XLogBytePosToRecPtr, but if the position is at a page boundary,
1684 * returns a pointer to the beginning of the page (ie. before page header),
1685 * not to where the first xlog record on that page would go to. This is used
1686 * when converting a pointer to the end of a record.
1689 XLogBytePosToEndRecPtr(uint64 bytepos
)
1697 fullsegs
= bytepos
/ UsableBytesInSegment
;
1698 bytesleft
= bytepos
% UsableBytesInSegment
;
1700 if (bytesleft
< XLOG_BLCKSZ
- SizeOfXLogLongPHD
)
1702 /* fits on first page of segment */
1706 seg_offset
= bytesleft
+ SizeOfXLogLongPHD
;
1710 /* account for the first page on segment with long header */
1711 seg_offset
= XLOG_BLCKSZ
;
1712 bytesleft
-= XLOG_BLCKSZ
- SizeOfXLogLongPHD
;
1714 fullpages
= bytesleft
/ UsableBytesInPage
;
1715 bytesleft
= bytesleft
% UsableBytesInPage
;
1718 seg_offset
+= fullpages
* XLOG_BLCKSZ
+ bytesleft
;
1720 seg_offset
+= fullpages
* XLOG_BLCKSZ
+ bytesleft
+ SizeOfXLogShortPHD
;
1723 XLogSegNoOffsetToRecPtr(fullsegs
, seg_offset
, wal_segment_size
, result
);
1729 * Convert an XLogRecPtr to a "usable byte position".
1732 XLogRecPtrToBytePos(XLogRecPtr ptr
)
1739 XLByteToSeg(ptr
, fullsegs
, wal_segment_size
);
1741 fullpages
= (XLogSegmentOffset(ptr
, wal_segment_size
)) / XLOG_BLCKSZ
;
1742 offset
= ptr
% XLOG_BLCKSZ
;
1746 result
= fullsegs
* UsableBytesInSegment
;
1749 Assert(offset
>= SizeOfXLogLongPHD
);
1750 result
+= offset
- SizeOfXLogLongPHD
;
1755 result
= fullsegs
* UsableBytesInSegment
+
1756 (XLOG_BLCKSZ
- SizeOfXLogLongPHD
) + /* account for first page */
1757 (fullpages
- 1) * UsableBytesInPage
; /* full pages */
1760 Assert(offset
>= SizeOfXLogShortPHD
);
1761 result
+= offset
- SizeOfXLogShortPHD
;
1769 * Initialize XLOG buffers, writing out old buffers if they still contain
1770 * unwritten data, upto the page containing 'upto'. Or if 'opportunistic' is
1771 * true, initialize as many pages as we can without having to write out
1772 * unwritten data. Any new pages are initialized to zeros, with pages headers
1773 * initialized properly.
1776 AdvanceXLInsertBuffer(XLogRecPtr upto
, TimeLineID tli
, bool opportunistic
)
1778 XLogCtlInsert
*Insert
= &XLogCtl
->Insert
;
1780 XLogRecPtr OldPageRqstPtr
;
1781 XLogwrtRqst WriteRqst
;
1782 XLogRecPtr NewPageEndPtr
= InvalidXLogRecPtr
;
1783 XLogRecPtr NewPageBeginPtr
;
1784 XLogPageHeader NewPage
;
1785 int npages
pg_attribute_unused() = 0;
1787 LWLockAcquire(WALBufMappingLock
, LW_EXCLUSIVE
);
1790 * Now that we have the lock, check if someone initialized the page
1793 while (upto
>= XLogCtl
->InitializedUpTo
|| opportunistic
)
1795 nextidx
= XLogRecPtrToBufIdx(XLogCtl
->InitializedUpTo
);
1798 * Get ending-offset of the buffer page we need to replace (this may
1799 * be zero if the buffer hasn't been used yet). Fall through if it's
1800 * already written out.
1802 OldPageRqstPtr
= XLogCtl
->xlblocks
[nextidx
];
1803 if (LogwrtResult
.Write
< OldPageRqstPtr
)
1806 * Nope, got work to do. If we just want to pre-initialize as much
1807 * as we can without flushing, give up now.
1812 /* Before waiting, get info_lck and update LogwrtResult */
1813 SpinLockAcquire(&XLogCtl
->info_lck
);
1814 if (XLogCtl
->LogwrtRqst
.Write
< OldPageRqstPtr
)
1815 XLogCtl
->LogwrtRqst
.Write
= OldPageRqstPtr
;
1816 LogwrtResult
= XLogCtl
->LogwrtResult
;
1817 SpinLockRelease(&XLogCtl
->info_lck
);
1820 * Now that we have an up-to-date LogwrtResult value, see if we
1821 * still need to write it or if someone else already did.
1823 if (LogwrtResult
.Write
< OldPageRqstPtr
)
1826 * Must acquire write lock. Release WALBufMappingLock first,
1827 * to make sure that all insertions that we need to wait for
1828 * can finish (up to this same position). Otherwise we risk
1831 LWLockRelease(WALBufMappingLock
);
1833 WaitXLogInsertionsToFinish(OldPageRqstPtr
);
1835 LWLockAcquire(WALWriteLock
, LW_EXCLUSIVE
);
1837 LogwrtResult
= XLogCtl
->LogwrtResult
;
1838 if (LogwrtResult
.Write
>= OldPageRqstPtr
)
1840 /* OK, someone wrote it already */
1841 LWLockRelease(WALWriteLock
);
1845 /* Have to write it ourselves */
1846 TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
1847 WriteRqst
.Write
= OldPageRqstPtr
;
1848 WriteRqst
.Flush
= 0;
1849 XLogWrite(WriteRqst
, tli
, false);
1850 LWLockRelease(WALWriteLock
);
1851 PendingWalStats
.wal_buffers_full
++;
1852 TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_DONE();
1854 /* Re-acquire WALBufMappingLock and retry */
1855 LWLockAcquire(WALBufMappingLock
, LW_EXCLUSIVE
);
1861 * Now the next buffer slot is free and we can set it up to be the
1864 NewPageBeginPtr
= XLogCtl
->InitializedUpTo
;
1865 NewPageEndPtr
= NewPageBeginPtr
+ XLOG_BLCKSZ
;
1867 Assert(XLogRecPtrToBufIdx(NewPageBeginPtr
) == nextidx
);
1869 NewPage
= (XLogPageHeader
) (XLogCtl
->pages
+ nextidx
* (Size
) XLOG_BLCKSZ
);
1872 * Be sure to re-zero the buffer so that bytes beyond what we've
1873 * written will look like zeroes and not valid XLOG records...
1875 MemSet((char *) NewPage
, 0, XLOG_BLCKSZ
);
1878 * Fill the new page's header
1880 NewPage
->xlp_magic
= XLOG_PAGE_MAGIC
;
1882 /* NewPage->xlp_info = 0; */ /* done by memset */
1883 NewPage
->xlp_tli
= tli
;
1884 NewPage
->xlp_pageaddr
= NewPageBeginPtr
;
1886 /* NewPage->xlp_rem_len = 0; */ /* done by memset */
1889 * If online backup is not in progress, mark the header to indicate
1890 * that WAL records beginning in this page have removable backup
1891 * blocks. This allows the WAL archiver to know whether it is safe to
1892 * compress archived WAL data by transforming full-block records into
1893 * the non-full-block format. It is sufficient to record this at the
1894 * page level because we force a page switch (in fact a segment
1895 * switch) when starting a backup, so the flag will be off before any
1896 * records can be written during the backup. At the end of a backup,
1897 * the last page will be marked as all unsafe when perhaps only part
1898 * is unsafe, but at worst the archiver would miss the opportunity to
1899 * compress a few records.
1901 if (Insert
->runningBackups
== 0)
1902 NewPage
->xlp_info
|= XLP_BKP_REMOVABLE
;
1905 * If first page of an XLOG segment file, make it a long header.
1907 if ((XLogSegmentOffset(NewPage
->xlp_pageaddr
, wal_segment_size
)) == 0)
1909 XLogLongPageHeader NewLongPage
= (XLogLongPageHeader
) NewPage
;
1911 NewLongPage
->xlp_sysid
= ControlFile
->system_identifier
;
1912 NewLongPage
->xlp_seg_size
= wal_segment_size
;
1913 NewLongPage
->xlp_xlog_blcksz
= XLOG_BLCKSZ
;
1914 NewPage
->xlp_info
|= XLP_LONG_HEADER
;
1918 * Make sure the initialization of the page becomes visible to others
1919 * before the xlblocks update. GetXLogBuffer() reads xlblocks without
1924 *((volatile XLogRecPtr
*) &XLogCtl
->xlblocks
[nextidx
]) = NewPageEndPtr
;
1926 XLogCtl
->InitializedUpTo
= NewPageEndPtr
;
1930 LWLockRelease(WALBufMappingLock
);
1933 if (XLOG_DEBUG
&& npages
> 0)
1935 elog(DEBUG1
, "initialized %d pages, up to %X/%X",
1936 npages
, LSN_FORMAT_ARGS(NewPageEndPtr
));
1942 * Calculate CheckPointSegments based on max_wal_size_mb and
1943 * checkpoint_completion_target.
1946 CalculateCheckpointSegments(void)
1951 * Calculate the distance at which to trigger a checkpoint, to avoid
1952 * exceeding max_wal_size_mb. This is based on two assumptions:
1954 * a) we keep WAL for only one checkpoint cycle (prior to PG11 we kept
1955 * WAL for two checkpoint cycles to allow us to recover from the
1956 * secondary checkpoint if the first checkpoint failed, though we
1957 * only did this on the primary anyway, not on standby. Keeping just
1958 * one checkpoint simplifies processing and reduces disk space in
1959 * many smaller databases.)
1960 * b) during checkpoint, we consume checkpoint_completion_target *
1961 * number of segments consumed between checkpoints.
1964 target
= (double) ConvertToXSegs(max_wal_size_mb
, wal_segment_size
) /
1965 (1.0 + CheckPointCompletionTarget
);
1968 CheckPointSegments
= (int) target
;
1970 if (CheckPointSegments
< 1)
1971 CheckPointSegments
= 1;
1975 assign_max_wal_size(int newval
, void *extra
)
1977 max_wal_size_mb
= newval
;
1978 CalculateCheckpointSegments();
1982 assign_checkpoint_completion_target(double newval
, void *extra
)
1984 CheckPointCompletionTarget
= newval
;
1985 CalculateCheckpointSegments();
1989 * At a checkpoint, how many WAL segments to recycle as preallocated future
1990 * XLOG segments? Returns the highest segment that should be preallocated.
1993 XLOGfileslop(XLogRecPtr lastredoptr
)
1998 XLogSegNo recycleSegNo
;
2001 * Calculate the segment numbers that min_wal_size_mb and max_wal_size_mb
2002 * correspond to. Always recycle enough segments to meet the minimum, and
2003 * remove enough segments to stay below the maximum.
2005 minSegNo
= lastredoptr
/ wal_segment_size
+
2006 ConvertToXSegs(min_wal_size_mb
, wal_segment_size
) - 1;
2007 maxSegNo
= lastredoptr
/ wal_segment_size
+
2008 ConvertToXSegs(max_wal_size_mb
, wal_segment_size
) - 1;
2011 * Between those limits, recycle enough segments to get us through to the
2012 * estimated end of next checkpoint.
2014 * To estimate where the next checkpoint will finish, assume that the
2015 * system runs steadily consuming CheckPointDistanceEstimate bytes between
2018 distance
= (1.0 + CheckPointCompletionTarget
) * CheckPointDistanceEstimate
;
2019 /* add 10% for good measure. */
2022 recycleSegNo
= (XLogSegNo
) ceil(((double) lastredoptr
+ distance
) /
2025 if (recycleSegNo
< minSegNo
)
2026 recycleSegNo
= minSegNo
;
2027 if (recycleSegNo
> maxSegNo
)
2028 recycleSegNo
= maxSegNo
;
2030 return recycleSegNo
;
2034 * Check whether we've consumed enough xlog space that a checkpoint is needed.
2036 * new_segno indicates a log file that has just been filled up (or read
2037 * during recovery). We measure the distance from RedoRecPtr to new_segno
2038 * and see if that exceeds CheckPointSegments.
2040 * Note: it is caller's responsibility that RedoRecPtr is up-to-date.
2043 XLogCheckpointNeeded(XLogSegNo new_segno
)
2045 XLogSegNo old_segno
;
2047 XLByteToSeg(RedoRecPtr
, old_segno
, wal_segment_size
);
2049 if (new_segno
>= old_segno
+ (uint64
) (CheckPointSegments
- 1))
2055 * Write and/or fsync the log at least as far as WriteRqst indicates.
2057 * If flexible == true, we don't have to write as far as WriteRqst, but
2058 * may stop at any convenient boundary (such as a cache or logfile boundary).
2059 * This option allows us to avoid uselessly issuing multiple writes when a
2060 * single one would do.
2062 * Must be called with WALWriteLock held. WaitXLogInsertionsToFinish(WriteRqst)
2063 * must be called before grabbing the lock, to make sure the data is ready to
2067 XLogWrite(XLogwrtRqst WriteRqst
, TimeLineID tli
, bool flexible
)
2070 bool last_iteration
;
2077 /* We should always be inside a critical section here */
2078 Assert(CritSectionCount
> 0);
2081 * Update local LogwrtResult (caller probably did this already, but...)
2083 LogwrtResult
= XLogCtl
->LogwrtResult
;
2086 * Since successive pages in the xlog cache are consecutively allocated,
2087 * we can usually gather multiple pages together and issue just one
2088 * write() call. npages is the number of pages we have determined can be
2089 * written together; startidx is the cache block index of the first one,
2090 * and startoffset is the file offset at which it should go. The latter
2091 * two variables are only valid when npages > 0, but we must initialize
2092 * all of them to keep the compiler quiet.
2099 * Within the loop, curridx is the cache block index of the page to
2100 * consider writing. Begin at the buffer containing the next unwritten
2101 * page, or last partially written page.
2103 curridx
= XLogRecPtrToBufIdx(LogwrtResult
.Write
);
2105 while (LogwrtResult
.Write
< WriteRqst
.Write
)
2108 * Make sure we're not ahead of the insert process. This could happen
2109 * if we're passed a bogus WriteRqst.Write that is past the end of the
2110 * last page that's been initialized by AdvanceXLInsertBuffer.
2112 XLogRecPtr EndPtr
= XLogCtl
->xlblocks
[curridx
];
2114 if (LogwrtResult
.Write
>= EndPtr
)
2115 elog(PANIC
, "xlog write request %X/%X is past end of log %X/%X",
2116 LSN_FORMAT_ARGS(LogwrtResult
.Write
),
2117 LSN_FORMAT_ARGS(EndPtr
));
2119 /* Advance LogwrtResult.Write to end of current buffer page */
2120 LogwrtResult
.Write
= EndPtr
;
2121 ispartialpage
= WriteRqst
.Write
< LogwrtResult
.Write
;
2123 if (!XLByteInPrevSeg(LogwrtResult
.Write
, openLogSegNo
,
2127 * Switch to new logfile segment. We cannot have any pending
2128 * pages here (since we dump what we have at segment end).
2130 Assert(npages
== 0);
2131 if (openLogFile
>= 0)
2133 XLByteToPrevSeg(LogwrtResult
.Write
, openLogSegNo
,
2137 /* create/use new log file */
2138 openLogFile
= XLogFileInit(openLogSegNo
, tli
);
2139 ReserveExternalFD();
2142 /* Make sure we have the current logfile open */
2143 if (openLogFile
< 0)
2145 XLByteToPrevSeg(LogwrtResult
.Write
, openLogSegNo
,
2148 openLogFile
= XLogFileOpen(openLogSegNo
, tli
);
2149 ReserveExternalFD();
2152 /* Add current page to the set of pending pages-to-dump */
2155 /* first of group */
2157 startoffset
= XLogSegmentOffset(LogwrtResult
.Write
- XLOG_BLCKSZ
,
2163 * Dump the set if this will be the last loop iteration, or if we are
2164 * at the last page of the cache area (since the next page won't be
2165 * contiguous in memory), or if we are at the end of the logfile
2168 last_iteration
= WriteRqst
.Write
<= LogwrtResult
.Write
;
2170 finishing_seg
= !ispartialpage
&&
2171 (startoffset
+ npages
* XLOG_BLCKSZ
) >= wal_segment_size
;
2173 if (last_iteration
||
2174 curridx
== XLogCtl
->XLogCacheBlck
||
2183 /* OK to write the page(s) */
2184 from
= XLogCtl
->pages
+ startidx
* (Size
) XLOG_BLCKSZ
;
2185 nbytes
= npages
* (Size
) XLOG_BLCKSZ
;
2191 /* Measure I/O timing to write WAL data */
2192 if (track_wal_io_timing
)
2193 INSTR_TIME_SET_CURRENT(start
);
2195 pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE
);
2196 written
= pg_pwrite(openLogFile
, from
, nleft
, startoffset
);
2197 pgstat_report_wait_end();
2200 * Increment the I/O timing and the number of times WAL data
2201 * were written out to disk.
2203 if (track_wal_io_timing
)
2205 instr_time duration
;
2207 INSTR_TIME_SET_CURRENT(duration
);
2208 INSTR_TIME_SUBTRACT(duration
, start
);
2209 PendingWalStats
.wal_write_time
+= INSTR_TIME_GET_MICROSEC(duration
);
2212 PendingWalStats
.wal_write
++;
2216 char xlogfname
[MAXFNAMELEN
];
2223 XLogFileName(xlogfname
, tli
, openLogSegNo
,
2227 (errcode_for_file_access(),
2228 errmsg("could not write to log file %s "
2229 "at offset %u, length %zu: %m",
2230 xlogfname
, startoffset
, nleft
)));
2234 startoffset
+= written
;
2235 } while (nleft
> 0);
2240 * If we just wrote the whole last page of a logfile segment,
2241 * fsync the segment immediately. This avoids having to go back
2242 * and re-open prior segments when an fsync request comes along
2243 * later. Doing it here ensures that one and only one backend will
2244 * perform this fsync.
2246 * This is also the right place to notify the Archiver that the
2247 * segment is ready to copy to archival storage, and to update the
2248 * timer for archive_timeout, and to signal for a checkpoint if
2249 * too many logfile segments have been used since the last
2254 issue_xlog_fsync(openLogFile
, openLogSegNo
, tli
);
2256 /* signal that we need to wakeup walsenders later */
2257 WalSndWakeupRequest();
2259 LogwrtResult
.Flush
= LogwrtResult
.Write
; /* end of page */
2261 if (XLogArchivingActive())
2262 XLogArchiveNotifySeg(openLogSegNo
, tli
);
2264 XLogCtl
->lastSegSwitchTime
= (pg_time_t
) time(NULL
);
2265 XLogCtl
->lastSegSwitchLSN
= LogwrtResult
.Flush
;
2268 * Request a checkpoint if we've consumed too much xlog since
2269 * the last one. For speed, we first check using the local
2270 * copy of RedoRecPtr, which might be out of date; if it looks
2271 * like a checkpoint is needed, forcibly update RedoRecPtr and
2274 if (IsUnderPostmaster
&& XLogCheckpointNeeded(openLogSegNo
))
2276 (void) GetRedoRecPtr();
2277 if (XLogCheckpointNeeded(openLogSegNo
))
2278 RequestCheckpoint(CHECKPOINT_CAUSE_XLOG
);
2285 /* Only asked to write a partial page */
2286 LogwrtResult
.Write
= WriteRqst
.Write
;
2289 curridx
= NextBufIdx(curridx
);
2291 /* If flexible, break out of loop as soon as we wrote something */
2292 if (flexible
&& npages
== 0)
2296 Assert(npages
== 0);
2299 * If asked to flush, do so
2301 if (LogwrtResult
.Flush
< WriteRqst
.Flush
&&
2302 LogwrtResult
.Flush
< LogwrtResult
.Write
)
2305 * Could get here without iterating above loop, in which case we might
2306 * have no open file or the wrong one. However, we do not need to
2307 * fsync more than one file.
2309 if (sync_method
!= SYNC_METHOD_OPEN
&&
2310 sync_method
!= SYNC_METHOD_OPEN_DSYNC
)
2312 if (openLogFile
>= 0 &&
2313 !XLByteInPrevSeg(LogwrtResult
.Write
, openLogSegNo
,
2316 if (openLogFile
< 0)
2318 XLByteToPrevSeg(LogwrtResult
.Write
, openLogSegNo
,
2321 openLogFile
= XLogFileOpen(openLogSegNo
, tli
);
2322 ReserveExternalFD();
2325 issue_xlog_fsync(openLogFile
, openLogSegNo
, tli
);
2328 /* signal that we need to wakeup walsenders later */
2329 WalSndWakeupRequest();
2331 LogwrtResult
.Flush
= LogwrtResult
.Write
;
2335 * Update shared-memory status
2337 * We make sure that the shared 'request' values do not fall behind the
2338 * 'result' values. This is not absolutely essential, but it saves some
2339 * code in a couple of places.
2342 SpinLockAcquire(&XLogCtl
->info_lck
);
2343 XLogCtl
->LogwrtResult
= LogwrtResult
;
2344 if (XLogCtl
->LogwrtRqst
.Write
< LogwrtResult
.Write
)
2345 XLogCtl
->LogwrtRqst
.Write
= LogwrtResult
.Write
;
2346 if (XLogCtl
->LogwrtRqst
.Flush
< LogwrtResult
.Flush
)
2347 XLogCtl
->LogwrtRqst
.Flush
= LogwrtResult
.Flush
;
2348 SpinLockRelease(&XLogCtl
->info_lck
);
2353 * Record the LSN for an asynchronous transaction commit/abort
2354 * and nudge the WALWriter if there is work for it to do.
2355 * (This should not be called for synchronous commits.)
2358 XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN
)
2360 XLogRecPtr WriteRqstPtr
= asyncXactLSN
;
2363 SpinLockAcquire(&XLogCtl
->info_lck
);
2364 LogwrtResult
= XLogCtl
->LogwrtResult
;
2365 sleeping
= XLogCtl
->WalWriterSleeping
;
2366 if (XLogCtl
->asyncXactLSN
< asyncXactLSN
)
2367 XLogCtl
->asyncXactLSN
= asyncXactLSN
;
2368 SpinLockRelease(&XLogCtl
->info_lck
);
2371 * If the WALWriter is sleeping, we should kick it to make it come out of
2372 * low-power mode. Otherwise, determine whether there's a full page of
2373 * WAL available to write.
2377 /* back off to last completed page boundary */
2378 WriteRqstPtr
-= WriteRqstPtr
% XLOG_BLCKSZ
;
2380 /* if we have already flushed that far, we're done */
2381 if (WriteRqstPtr
<= LogwrtResult
.Flush
)
2386 * Nudge the WALWriter: it has a full page of WAL to write, or we want it
2387 * to come out of low-power mode so that this async commit will reach disk
2388 * within the expected amount of time.
2390 if (ProcGlobal
->walwriterLatch
)
2391 SetLatch(ProcGlobal
->walwriterLatch
);
2395 * Record the LSN up to which we can remove WAL because it's not required by
2396 * any replication slot.
2399 XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn
)
2401 SpinLockAcquire(&XLogCtl
->info_lck
);
2402 XLogCtl
->replicationSlotMinLSN
= lsn
;
2403 SpinLockRelease(&XLogCtl
->info_lck
);
2408 * Return the oldest LSN we must retain to satisfy the needs of some
2412 XLogGetReplicationSlotMinimumLSN(void)
2416 SpinLockAcquire(&XLogCtl
->info_lck
);
2417 retval
= XLogCtl
->replicationSlotMinLSN
;
2418 SpinLockRelease(&XLogCtl
->info_lck
);
2424 * Advance minRecoveryPoint in control file.
2426 * If we crash during recovery, we must reach this point again before the
2427 * database is consistent.
2429 * If 'force' is true, 'lsn' argument is ignored. Otherwise, minRecoveryPoint
2430 * is only updated if it's not already greater than or equal to 'lsn'.
2433 UpdateMinRecoveryPoint(XLogRecPtr lsn
, bool force
)
2435 /* Quick check using our local copy of the variable */
2436 if (!updateMinRecoveryPoint
|| (!force
&& lsn
<= LocalMinRecoveryPoint
))
2440 * An invalid minRecoveryPoint means that we need to recover all the WAL,
2441 * i.e., we're doing crash recovery. We never modify the control file's
2442 * value in that case, so we can short-circuit future checks here too. The
2443 * local values of minRecoveryPoint and minRecoveryPointTLI should not be
2444 * updated until crash recovery finishes. We only do this for the startup
2445 * process as it should not update its own reference of minRecoveryPoint
2446 * until it has finished crash recovery to make sure that all WAL
2447 * available is replayed in this case. This also saves from extra locks
2448 * taken on the control file from the startup process.
2450 if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint
) && InRecovery
)
2452 updateMinRecoveryPoint
= false;
2456 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
2458 /* update local copy */
2459 LocalMinRecoveryPoint
= ControlFile
->minRecoveryPoint
;
2460 LocalMinRecoveryPointTLI
= ControlFile
->minRecoveryPointTLI
;
2462 if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint
))
2463 updateMinRecoveryPoint
= false;
2464 else if (force
|| LocalMinRecoveryPoint
< lsn
)
2466 XLogRecPtr newMinRecoveryPoint
;
2467 TimeLineID newMinRecoveryPointTLI
;
2470 * To avoid having to update the control file too often, we update it
2471 * all the way to the last record being replayed, even though 'lsn'
2472 * would suffice for correctness. This also allows the 'force' case
2473 * to not need a valid 'lsn' value.
2475 * Another important reason for doing it this way is that the passed
2476 * 'lsn' value could be bogus, i.e., past the end of available WAL, if
2477 * the caller got it from a corrupted heap page. Accepting such a
2478 * value as the min recovery point would prevent us from coming up at
2479 * all. Instead, we just log a warning and continue with recovery.
2480 * (See also the comments about corrupt LSNs in XLogFlush.)
2482 newMinRecoveryPoint
= GetCurrentReplayRecPtr(&newMinRecoveryPointTLI
);
2483 if (!force
&& newMinRecoveryPoint
< lsn
)
2485 "xlog min recovery request %X/%X is past current point %X/%X",
2486 LSN_FORMAT_ARGS(lsn
), LSN_FORMAT_ARGS(newMinRecoveryPoint
));
2488 /* update control file */
2489 if (ControlFile
->minRecoveryPoint
< newMinRecoveryPoint
)
2491 ControlFile
->minRecoveryPoint
= newMinRecoveryPoint
;
2492 ControlFile
->minRecoveryPointTLI
= newMinRecoveryPointTLI
;
2493 UpdateControlFile();
2494 LocalMinRecoveryPoint
= newMinRecoveryPoint
;
2495 LocalMinRecoveryPointTLI
= newMinRecoveryPointTLI
;
2498 (errmsg_internal("updated min recovery point to %X/%X on timeline %u",
2499 LSN_FORMAT_ARGS(newMinRecoveryPoint
),
2500 newMinRecoveryPointTLI
)));
2503 LWLockRelease(ControlFileLock
);
2507 * Ensure that all XLOG data through the given position is flushed to disk.
2509 * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
2510 * already held, and we try to avoid acquiring it if possible.
2513 XLogFlush(XLogRecPtr record
)
2515 XLogRecPtr WriteRqstPtr
;
2516 XLogwrtRqst WriteRqst
;
2517 TimeLineID insertTLI
= XLogCtl
->InsertTimeLineID
;
2520 * During REDO, we are reading not writing WAL. Therefore, instead of
2521 * trying to flush the WAL, we should update minRecoveryPoint instead. We
2522 * test XLogInsertAllowed(), not InRecovery, because we need checkpointer
2523 * to act this way too, and because when it tries to write the
2524 * end-of-recovery checkpoint, it should indeed flush.
2526 if (!XLogInsertAllowed())
2528 UpdateMinRecoveryPoint(record
, false);
2532 /* Quick exit if already known flushed */
2533 if (record
<= LogwrtResult
.Flush
)
2538 elog(LOG
, "xlog flush request %X/%X; write %X/%X; flush %X/%X",
2539 LSN_FORMAT_ARGS(record
),
2540 LSN_FORMAT_ARGS(LogwrtResult
.Write
),
2541 LSN_FORMAT_ARGS(LogwrtResult
.Flush
));
2544 START_CRIT_SECTION();
2547 * Since fsync is usually a horribly expensive operation, we try to
2548 * piggyback as much data as we can on each fsync: if we see any more data
2549 * entered into the xlog buffer, we'll write and fsync that too, so that
2550 * the final value of LogwrtResult.Flush is as large as possible. This
2551 * gives us some chance of avoiding another fsync immediately after.
2554 /* initialize to given target; may increase below */
2555 WriteRqstPtr
= record
;
2558 * Now wait until we get the write lock, or someone else does the flush
2563 XLogRecPtr insertpos
;
2565 /* read LogwrtResult and update local state */
2566 SpinLockAcquire(&XLogCtl
->info_lck
);
2567 if (WriteRqstPtr
< XLogCtl
->LogwrtRqst
.Write
)
2568 WriteRqstPtr
= XLogCtl
->LogwrtRqst
.Write
;
2569 LogwrtResult
= XLogCtl
->LogwrtResult
;
2570 SpinLockRelease(&XLogCtl
->info_lck
);
2573 if (record
<= LogwrtResult
.Flush
)
2577 * Before actually performing the write, wait for all in-flight
2578 * insertions to the pages we're about to write to finish.
2580 insertpos
= WaitXLogInsertionsToFinish(WriteRqstPtr
);
2583 * Try to get the write lock. If we can't get it immediately, wait
2584 * until it's released, and recheck if we still need to do the flush
2585 * or if the backend that held the lock did it for us already. This
2586 * helps to maintain a good rate of group committing when the system
2587 * is bottlenecked by the speed of fsyncing.
2589 if (!LWLockAcquireOrWait(WALWriteLock
, LW_EXCLUSIVE
))
2592 * The lock is now free, but we didn't acquire it yet. Before we
2593 * do, loop back to check if someone else flushed the record for
2599 /* Got the lock; recheck whether request is satisfied */
2600 LogwrtResult
= XLogCtl
->LogwrtResult
;
2601 if (record
<= LogwrtResult
.Flush
)
2603 LWLockRelease(WALWriteLock
);
2608 * Sleep before flush! By adding a delay here, we may give further
2609 * backends the opportunity to join the backlog of group commit
2610 * followers; this can significantly improve transaction throughput,
2611 * at the risk of increasing transaction latency.
2613 * We do not sleep if enableFsync is not turned on, nor if there are
2614 * fewer than CommitSiblings other backends with active transactions.
2616 if (CommitDelay
> 0 && enableFsync
&&
2617 MinimumActiveBackends(CommitSiblings
))
2619 pg_usleep(CommitDelay
);
2622 * Re-check how far we can now flush the WAL. It's generally not
2623 * safe to call WaitXLogInsertionsToFinish while holding
2624 * WALWriteLock, because an in-progress insertion might need to
2625 * also grab WALWriteLock to make progress. But we know that all
2626 * the insertions up to insertpos have already finished, because
2627 * that's what the earlier WaitXLogInsertionsToFinish() returned.
2628 * We're only calling it again to allow insertpos to be moved
2629 * further forward, not to actually wait for anyone.
2631 insertpos
= WaitXLogInsertionsToFinish(insertpos
);
2634 /* try to write/flush later additions to XLOG as well */
2635 WriteRqst
.Write
= insertpos
;
2636 WriteRqst
.Flush
= insertpos
;
2638 XLogWrite(WriteRqst
, insertTLI
, false);
2640 LWLockRelease(WALWriteLock
);
2647 /* wake up walsenders now that we've released heavily contended locks */
2648 WalSndWakeupProcessRequests();
2651 * If we still haven't flushed to the request point then we have a
2652 * problem; most likely, the requested flush point is past end of XLOG.
2653 * This has been seen to occur when a disk page has a corrupted LSN.
2655 * Formerly we treated this as a PANIC condition, but that hurts the
2656 * system's robustness rather than helping it: we do not want to take down
2657 * the whole system due to corruption on one data page. In particular, if
2658 * the bad page is encountered again during recovery then we would be
2659 * unable to restart the database at all! (This scenario actually
2660 * happened in the field several times with 7.1 releases.) As of 8.4, bad
2661 * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
2662 * the only time we can reach here during recovery is while flushing the
2663 * end-of-recovery checkpoint record, and we don't expect that to have a
2666 * Note that for calls from xact.c, the ERROR will be promoted to PANIC
2667 * since xact.c calls this routine inside a critical section. However,
2668 * calls from bufmgr.c are not within critical sections and so we will not
2669 * force a restart for a bad LSN on a data page.
2671 if (LogwrtResult
.Flush
< record
)
2673 "xlog flush request %X/%X is not satisfied --- flushed only to %X/%X",
2674 LSN_FORMAT_ARGS(record
),
2675 LSN_FORMAT_ARGS(LogwrtResult
.Flush
));
2679 * Write & flush xlog, but without specifying exactly where to.
2681 * We normally write only completed blocks; but if there is nothing to do on
2682 * that basis, we check for unwritten async commits in the current incomplete
2683 * block, and write through the latest one of those. Thus, if async commits
2684 * are not being used, we will write complete blocks only.
2686 * If, based on the above, there's anything to write we do so immediately. But
2687 * to avoid calling fsync, fdatasync et. al. at a rate that'd impact
2688 * concurrent IO, we only flush WAL every wal_writer_delay ms, or if there's
2689 * more than wal_writer_flush_after unflushed blocks.
2691 * We can guarantee that async commits reach disk after at most three
2692 * wal_writer_delay cycles. (When flushing complete blocks, we allow XLogWrite
2693 * to write "flexibly", meaning it can stop at the end of the buffer ring;
2694 * this makes a difference only with very high load or long wal_writer_delay,
2695 * but imposes one extra cycle for the worst case for async commits.)
2697 * This routine is invoked periodically by the background walwriter process.
2699 * Returns true if there was any work to do, even if we skipped flushing due
2700 * to wal_writer_delay/wal_writer_flush_after.
2703 XLogBackgroundFlush(void)
2705 XLogwrtRqst WriteRqst
;
2706 bool flexible
= true;
2707 static TimestampTz lastflush
;
2710 TimeLineID insertTLI
;
2712 /* XLOG doesn't need flushing during recovery */
2713 if (RecoveryInProgress())
2717 * Since we're not in recovery, InsertTimeLineID is set and can't change,
2718 * so we can read it without a lock.
2720 insertTLI
= XLogCtl
->InsertTimeLineID
;
2722 /* read LogwrtResult and update local state */
2723 SpinLockAcquire(&XLogCtl
->info_lck
);
2724 LogwrtResult
= XLogCtl
->LogwrtResult
;
2725 WriteRqst
= XLogCtl
->LogwrtRqst
;
2726 SpinLockRelease(&XLogCtl
->info_lck
);
2728 /* back off to last completed page boundary */
2729 WriteRqst
.Write
-= WriteRqst
.Write
% XLOG_BLCKSZ
;
2731 /* if we have already flushed that far, consider async commit records */
2732 if (WriteRqst
.Write
<= LogwrtResult
.Flush
)
2734 SpinLockAcquire(&XLogCtl
->info_lck
);
2735 WriteRqst
.Write
= XLogCtl
->asyncXactLSN
;
2736 SpinLockRelease(&XLogCtl
->info_lck
);
2737 flexible
= false; /* ensure it all gets written */
2741 * If already known flushed, we're done. Just need to check if we are
2742 * holding an open file handle to a logfile that's no longer in use,
2743 * preventing the file from being deleted.
2745 if (WriteRqst
.Write
<= LogwrtResult
.Flush
)
2747 if (openLogFile
>= 0)
2749 if (!XLByteInPrevSeg(LogwrtResult
.Write
, openLogSegNo
,
2759 * Determine how far to flush WAL, based on the wal_writer_delay and
2760 * wal_writer_flush_after GUCs.
2762 now
= GetCurrentTimestamp();
2764 WriteRqst
.Write
/ XLOG_BLCKSZ
- LogwrtResult
.Flush
/ XLOG_BLCKSZ
;
2766 if (WalWriterFlushAfter
== 0 || lastflush
== 0)
2768 /* first call, or block based limits disabled */
2769 WriteRqst
.Flush
= WriteRqst
.Write
;
2772 else if (TimestampDifferenceExceeds(lastflush
, now
, WalWriterDelay
))
2775 * Flush the writes at least every WalWriterDelay ms. This is
2776 * important to bound the amount of time it takes for an asynchronous
2777 * commit to hit disk.
2779 WriteRqst
.Flush
= WriteRqst
.Write
;
2782 else if (flushbytes
>= WalWriterFlushAfter
)
2784 /* exceeded wal_writer_flush_after blocks, flush */
2785 WriteRqst
.Flush
= WriteRqst
.Write
;
2790 /* no flushing, this time round */
2791 WriteRqst
.Flush
= 0;
2796 elog(LOG
, "xlog bg flush request write %X/%X; flush: %X/%X, current is write %X/%X; flush %X/%X",
2797 LSN_FORMAT_ARGS(WriteRqst
.Write
),
2798 LSN_FORMAT_ARGS(WriteRqst
.Flush
),
2799 LSN_FORMAT_ARGS(LogwrtResult
.Write
),
2800 LSN_FORMAT_ARGS(LogwrtResult
.Flush
));
2803 START_CRIT_SECTION();
2805 /* now wait for any in-progress insertions to finish and get write lock */
2806 WaitXLogInsertionsToFinish(WriteRqst
.Write
);
2807 LWLockAcquire(WALWriteLock
, LW_EXCLUSIVE
);
2808 LogwrtResult
= XLogCtl
->LogwrtResult
;
2809 if (WriteRqst
.Write
> LogwrtResult
.Write
||
2810 WriteRqst
.Flush
> LogwrtResult
.Flush
)
2812 XLogWrite(WriteRqst
, insertTLI
, flexible
);
2814 LWLockRelease(WALWriteLock
);
2818 /* wake up walsenders now that we've released heavily contended locks */
2819 WalSndWakeupProcessRequests();
2822 * Great, done. To take some work off the critical path, try to initialize
2823 * as many of the no-longer-needed WAL buffers for future use as we can.
2825 AdvanceXLInsertBuffer(InvalidXLogRecPtr
, insertTLI
, true);
2828 * If we determined that we need to write data, but somebody else
2829 * wrote/flushed already, it should be considered as being active, to
2830 * avoid hibernating too early.
2836 * Test whether XLOG data has been flushed up to (at least) the given position.
2838 * Returns true if a flush is still needed. (It may be that someone else
2839 * is already in process of flushing that far, however.)
2842 XLogNeedsFlush(XLogRecPtr record
)
2845 * During recovery, we don't flush WAL but update minRecoveryPoint
2846 * instead. So "needs flush" is taken to mean whether minRecoveryPoint
2847 * would need to be updated.
2849 if (RecoveryInProgress())
2852 * An invalid minRecoveryPoint means that we need to recover all the
2853 * WAL, i.e., we're doing crash recovery. We never modify the control
2854 * file's value in that case, so we can short-circuit future checks
2855 * here too. This triggers a quick exit path for the startup process,
2856 * which cannot update its local copy of minRecoveryPoint as long as
2857 * it has not replayed all WAL available when doing crash recovery.
2859 if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint
) && InRecovery
)
2860 updateMinRecoveryPoint
= false;
2862 /* Quick exit if already known to be updated or cannot be updated */
2863 if (record
<= LocalMinRecoveryPoint
|| !updateMinRecoveryPoint
)
2867 * Update local copy of minRecoveryPoint. But if the lock is busy,
2868 * just return a conservative guess.
2870 if (!LWLockConditionalAcquire(ControlFileLock
, LW_SHARED
))
2872 LocalMinRecoveryPoint
= ControlFile
->minRecoveryPoint
;
2873 LocalMinRecoveryPointTLI
= ControlFile
->minRecoveryPointTLI
;
2874 LWLockRelease(ControlFileLock
);
2877 * Check minRecoveryPoint for any other process than the startup
2878 * process doing crash recovery, which should not update the control
2879 * file value if crash recovery is still running.
2881 if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint
))
2882 updateMinRecoveryPoint
= false;
2885 if (record
<= LocalMinRecoveryPoint
|| !updateMinRecoveryPoint
)
2891 /* Quick exit if already known flushed */
2892 if (record
<= LogwrtResult
.Flush
)
2895 /* read LogwrtResult and update local state */
2896 SpinLockAcquire(&XLogCtl
->info_lck
);
2897 LogwrtResult
= XLogCtl
->LogwrtResult
;
2898 SpinLockRelease(&XLogCtl
->info_lck
);
2901 if (record
<= LogwrtResult
.Flush
)
2908 * Try to make a given XLOG file segment exist.
2910 * logsegno: identify segment.
2912 * *added: on return, true if this call raised the number of extant segments.
2914 * path: on return, this char[MAXPGPATH] has the path to the logsegno file.
2916 * Returns -1 or FD of opened file. A -1 here is not an error; a caller
2917 * wanting an open segment should attempt to open "path", which usually will
2918 * succeed. (This is weird, but it's efficient for the callers.)
2921 XLogFileInitInternal(XLogSegNo logsegno
, TimeLineID logtli
,
2922 bool *added
, char *path
)
2924 char tmppath
[MAXPGPATH
];
2925 XLogSegNo installed_segno
;
2926 XLogSegNo max_segno
;
2930 Assert(logtli
!= 0);
2932 XLogFilePath(path
, logtli
, logsegno
, wal_segment_size
);
2935 * Try to use existent file (checkpoint maker may have created it already)
2938 fd
= BasicOpenFile(path
, O_RDWR
| PG_BINARY
| get_sync_bit(sync_method
));
2941 if (errno
!= ENOENT
)
2943 (errcode_for_file_access(),
2944 errmsg("could not open file \"%s\": %m", path
)));
2950 * Initialize an empty (all zeroes) segment. NOTE: it is possible that
2951 * another process is doing the same thing. If so, we will end up
2952 * pre-creating an extra log segment. That seems OK, and better than
2953 * holding the lock throughout this lengthy process.
2955 elog(DEBUG2
, "creating and filling new WAL file");
2957 snprintf(tmppath
, MAXPGPATH
, XLOGDIR
"/xlogtemp.%d", (int) getpid());
2961 /* do not use get_sync_bit() here --- want to fsync only at end of fill */
2962 fd
= BasicOpenFile(tmppath
, O_RDWR
| O_CREAT
| O_EXCL
| PG_BINARY
);
2965 (errcode_for_file_access(),
2966 errmsg("could not create file \"%s\": %m", tmppath
)));
2968 pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE
);
2975 * Zero-fill the file. With this setting, we do this the hard way to
2976 * ensure that all the file space has really been allocated. On
2977 * platforms that allow "holes" in files, just seeking to the end
2978 * doesn't allocate intermediate space. This way, we know that we
2979 * have all the space and (after the fsync below) that all the
2980 * indirect blocks are down on disk. Therefore, fdatasync(2) or
2981 * O_DSYNC will be sufficient to sync future writes to the log file.
2983 rc
= pg_pwrite_zeros(fd
, wal_segment_size
);
2991 * Otherwise, seeking to the end and writing a solitary byte is
2995 if (pg_pwrite(fd
, "\0", 1, wal_segment_size
- 1) != 1)
2997 /* if write didn't set errno, assume no disk space */
2998 save_errno
= errno
? errno
: ENOSPC
;
3001 pgstat_report_wait_end();
3006 * If we fail to make the file, delete it to release disk space
3015 (errcode_for_file_access(),
3016 errmsg("could not write to file \"%s\": %m", tmppath
)));
3019 pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC
);
3020 if (pg_fsync(fd
) != 0)
3026 (errcode_for_file_access(),
3027 errmsg("could not fsync file \"%s\": %m", tmppath
)));
3029 pgstat_report_wait_end();
3033 (errcode_for_file_access(),
3034 errmsg("could not close file \"%s\": %m", tmppath
)));
3037 * Now move the segment into place with its final name. Cope with
3038 * possibility that someone else has created the file while we were
3039 * filling ours: if so, use ours to pre-create a future log segment.
3041 installed_segno
= logsegno
;
3044 * XXX: What should we use as max_segno? We used to use XLOGfileslop when
3045 * that was a constant, but that was always a bit dubious: normally, at a
3046 * checkpoint, XLOGfileslop was the offset from the checkpoint record, but
3047 * here, it was the offset from the insert location. We can't do the
3048 * normal XLOGfileslop calculation here because we don't have access to
3049 * the prior checkpoint's redo location. So somewhat arbitrarily, just use
3050 * CheckPointSegments.
3052 max_segno
= logsegno
+ CheckPointSegments
;
3053 if (InstallXLogFileSegment(&installed_segno
, tmppath
, true, max_segno
,
3057 elog(DEBUG2
, "done creating and filling new WAL file");
3062 * No need for any more future segments, or InstallXLogFileSegment()
3063 * failed to rename the file into place. If the rename failed, a
3064 * caller opening the file may fail.
3067 elog(DEBUG2
, "abandoned new WAL file");
3074 * Create a new XLOG file segment, or open a pre-existing one.
3076 * logsegno: identify segment to be created/opened.
3078 * Returns FD of opened file.
3080 * Note: errors here are ERROR not PANIC because we might or might not be
3081 * inside a critical section (eg, during checkpoint there is no reason to
3082 * take down the system on failure). They will promote to PANIC if we are
3083 * in a critical section.
3086 XLogFileInit(XLogSegNo logsegno
, TimeLineID logtli
)
3089 char path
[MAXPGPATH
];
3092 Assert(logtli
!= 0);
3094 fd
= XLogFileInitInternal(logsegno
, logtli
, &ignore_added
, path
);
3098 /* Now open original target segment (might not be file I just made) */
3099 fd
= BasicOpenFile(path
, O_RDWR
| PG_BINARY
| get_sync_bit(sync_method
));
3102 (errcode_for_file_access(),
3103 errmsg("could not open file \"%s\": %m", path
)));
3108 * Create a new XLOG file segment by copying a pre-existing one.
3110 * destsegno: identify segment to be created.
3112 * srcTLI, srcsegno: identify segment to be copied (could be from
3113 * a different timeline)
3115 * upto: how much of the source file to copy (the rest is filled with
3118 * Currently this is only used during recovery, and so there are no locking
3119 * considerations. But we should be just as tense as XLogFileInit to avoid
3120 * emplacing a bogus file.
3123 XLogFileCopy(TimeLineID destTLI
, XLogSegNo destsegno
,
3124 TimeLineID srcTLI
, XLogSegNo srcsegno
,
3127 char path
[MAXPGPATH
];
3128 char tmppath
[MAXPGPATH
];
3129 PGAlignedXLogBlock buffer
;
3135 * Open the source file
3137 XLogFilePath(path
, srcTLI
, srcsegno
, wal_segment_size
);
3138 srcfd
= OpenTransientFile(path
, O_RDONLY
| PG_BINARY
);
3141 (errcode_for_file_access(),
3142 errmsg("could not open file \"%s\": %m", path
)));
3145 * Copy into a temp file name.
3147 snprintf(tmppath
, MAXPGPATH
, XLOGDIR
"/xlogtemp.%d", (int) getpid());
3151 /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3152 fd
= OpenTransientFile(tmppath
, O_RDWR
| O_CREAT
| O_EXCL
| PG_BINARY
);
3155 (errcode_for_file_access(),
3156 errmsg("could not create file \"%s\": %m", tmppath
)));
3159 * Do the data copying.
3161 for (nbytes
= 0; nbytes
< wal_segment_size
; nbytes
+= sizeof(buffer
))
3165 nread
= upto
- nbytes
;
3168 * The part that is not read from the source file is filled with
3171 if (nread
< sizeof(buffer
))
3172 memset(buffer
.data
, 0, sizeof(buffer
));
3178 if (nread
> sizeof(buffer
))
3179 nread
= sizeof(buffer
);
3180 pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ
);
3181 r
= read(srcfd
, buffer
.data
, nread
);
3186 (errcode_for_file_access(),
3187 errmsg("could not read file \"%s\": %m",
3191 (errcode(ERRCODE_DATA_CORRUPTED
),
3192 errmsg("could not read file \"%s\": read %d of %zu",
3193 path
, r
, (Size
) nread
)));
3195 pgstat_report_wait_end();
3198 pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_WRITE
);
3199 if ((int) write(fd
, buffer
.data
, sizeof(buffer
)) != (int) sizeof(buffer
))
3201 int save_errno
= errno
;
3204 * If we fail to make the file, delete it to release disk space
3207 /* if write didn't set errno, assume problem is no disk space */
3208 errno
= save_errno
? save_errno
: ENOSPC
;
3211 (errcode_for_file_access(),
3212 errmsg("could not write to file \"%s\": %m", tmppath
)));
3214 pgstat_report_wait_end();
3217 pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_SYNC
);
3218 if (pg_fsync(fd
) != 0)
3219 ereport(data_sync_elevel(ERROR
),
3220 (errcode_for_file_access(),
3221 errmsg("could not fsync file \"%s\": %m", tmppath
)));
3222 pgstat_report_wait_end();
3224 if (CloseTransientFile(fd
) != 0)
3226 (errcode_for_file_access(),
3227 errmsg("could not close file \"%s\": %m", tmppath
)));
3229 if (CloseTransientFile(srcfd
) != 0)
3231 (errcode_for_file_access(),
3232 errmsg("could not close file \"%s\": %m", path
)));
3235 * Now move the segment into place with its final name.
3237 if (!InstallXLogFileSegment(&destsegno
, tmppath
, false, 0, destTLI
))
3238 elog(ERROR
, "InstallXLogFileSegment should not have failed");
3242 * Install a new XLOG segment file as a current or future log segment.
3244 * This is used both to install a newly-created segment (which has a temp
3245 * filename while it's being created) and to recycle an old segment.
3247 * *segno: identify segment to install as (or first possible target).
3248 * When find_free is true, this is modified on return to indicate the
3249 * actual installation location or last segment searched.
3251 * tmppath: initial name of file to install. It will be renamed into place.
3253 * find_free: if true, install the new segment at the first empty segno
3254 * number at or after the passed numbers. If false, install the new segment
3255 * exactly where specified, deleting any existing segment file there.
3257 * max_segno: maximum segment number to install the new file as. Fail if no
3258 * free slot is found between *segno and max_segno. (Ignored when find_free
3261 * tli: The timeline on which the new segment should be installed.
3263 * Returns true if the file was installed successfully. false indicates that
3264 * max_segno limit was exceeded, the startup process has disabled this
3265 * function for now, or an error occurred while renaming the file into place.
3268 InstallXLogFileSegment(XLogSegNo
*segno
, char *tmppath
,
3269 bool find_free
, XLogSegNo max_segno
, TimeLineID tli
)
3271 char path
[MAXPGPATH
];
3272 struct stat stat_buf
;
3276 XLogFilePath(path
, tli
, *segno
, wal_segment_size
);
3278 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
3279 if (!XLogCtl
->InstallXLogFileSegmentActive
)
3281 LWLockRelease(ControlFileLock
);
3287 /* Force installation: get rid of any pre-existing segment file */
3288 durable_unlink(path
, DEBUG1
);
3292 /* Find a free slot to put it in */
3293 while (stat(path
, &stat_buf
) == 0)
3295 if ((*segno
) >= max_segno
)
3297 /* Failed to find a free slot within specified range */
3298 LWLockRelease(ControlFileLock
);
3302 XLogFilePath(path
, tli
, *segno
, wal_segment_size
);
3306 Assert(access(path
, F_OK
) != 0 && errno
== ENOENT
);
3307 if (durable_rename(tmppath
, path
, LOG
) != 0)
3309 LWLockRelease(ControlFileLock
);
3310 /* durable_rename already emitted log message */
3314 LWLockRelease(ControlFileLock
);
3320 * Open a pre-existing logfile segment for writing.
3323 XLogFileOpen(XLogSegNo segno
, TimeLineID tli
)
3325 char path
[MAXPGPATH
];
3328 XLogFilePath(path
, tli
, segno
, wal_segment_size
);
3330 fd
= BasicOpenFile(path
, O_RDWR
| PG_BINARY
| get_sync_bit(sync_method
));
3333 (errcode_for_file_access(),
3334 errmsg("could not open file \"%s\": %m", path
)));
3340 * Close the current logfile segment for writing.
3345 Assert(openLogFile
>= 0);
3348 * WAL segment files will not be re-read in normal operation, so we advise
3349 * the OS to release any cached pages. But do not do so if WAL archiving
3350 * or streaming is active, because archiver and walsender process could
3351 * use the cache to read the WAL segment.
3353 #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3354 if (!XLogIsNeeded())
3355 (void) posix_fadvise(openLogFile
, 0, 0, POSIX_FADV_DONTNEED
);
3358 if (close(openLogFile
) != 0)
3360 char xlogfname
[MAXFNAMELEN
];
3361 int save_errno
= errno
;
3363 XLogFileName(xlogfname
, openLogTLI
, openLogSegNo
, wal_segment_size
);
3366 (errcode_for_file_access(),
3367 errmsg("could not close file \"%s\": %m", xlogfname
)));
3371 ReleaseExternalFD();
3375 * Preallocate log files beyond the specified log endpoint.
3377 * XXX this is currently extremely conservative, since it forces only one
3378 * future log segment to exist, and even that only if we are 75% done with
3379 * the current one. This is only appropriate for very low-WAL-volume systems.
3380 * High-volume systems will be OK once they've built up a sufficient set of
3381 * recycled log segments, but the startup transient is likely to include
3382 * a lot of segment creations by foreground processes, which is not so good.
3384 * XLogFileInitInternal() can ereport(ERROR). All known causes indicate big
3385 * trouble; for example, a full filesystem is one cause. The checkpoint WAL
3386 * and/or ControlFile updates already completed. If a RequestCheckpoint()
3387 * initiated the present checkpoint and an ERROR ends this function, the
3388 * command that called RequestCheckpoint() fails. That's not ideal, but it's
3389 * not worth contorting more functions to use caller-specified elevel values.
3390 * (With or without RequestCheckpoint(), an ERROR forestalls some inessential
3391 * reporting and resource reclamation.)
3394 PreallocXlogFiles(XLogRecPtr endptr
, TimeLineID tli
)
3396 XLogSegNo _logSegNo
;
3399 char path
[MAXPGPATH
];
3402 if (!XLogCtl
->InstallXLogFileSegmentActive
)
3403 return; /* unlocked check says no */
3405 XLByteToPrevSeg(endptr
, _logSegNo
, wal_segment_size
);
3406 offset
= XLogSegmentOffset(endptr
- 1, wal_segment_size
);
3407 if (offset
>= (uint32
) (0.75 * wal_segment_size
))
3410 lf
= XLogFileInitInternal(_logSegNo
, tli
, &added
, path
);
3414 CheckpointStats
.ckpt_segs_added
++;
3419 * Throws an error if the given log segment has already been removed or
3420 * recycled. The caller should only pass a segment that it knows to have
3421 * existed while the server has been running, as this function always
3422 * succeeds if no WAL segments have been removed since startup.
3423 * 'tli' is only used in the error message.
3425 * Note: this function guarantees to keep errno unchanged on return.
3426 * This supports callers that use this to possibly deliver a better
3427 * error message about a missing file, while still being able to throw
3428 * a normal file-access error afterwards, if this does return.
3431 CheckXLogRemoved(XLogSegNo segno
, TimeLineID tli
)
3433 int save_errno
= errno
;
3434 XLogSegNo lastRemovedSegNo
;
3436 SpinLockAcquire(&XLogCtl
->info_lck
);
3437 lastRemovedSegNo
= XLogCtl
->lastRemovedSegNo
;
3438 SpinLockRelease(&XLogCtl
->info_lck
);
3440 if (segno
<= lastRemovedSegNo
)
3442 char filename
[MAXFNAMELEN
];
3444 XLogFileName(filename
, tli
, segno
, wal_segment_size
);
3447 (errcode_for_file_access(),
3448 errmsg("requested WAL segment %s has already been removed",
3455 * Return the last WAL segment removed, or 0 if no segment has been removed
3458 * NB: the result can be out of date arbitrarily fast, the caller has to deal
3462 XLogGetLastRemovedSegno(void)
3464 XLogSegNo lastRemovedSegNo
;
3466 SpinLockAcquire(&XLogCtl
->info_lck
);
3467 lastRemovedSegNo
= XLogCtl
->lastRemovedSegNo
;
3468 SpinLockRelease(&XLogCtl
->info_lck
);
3470 return lastRemovedSegNo
;
3475 * Update the last removed segno pointer in shared memory, to reflect that the
3476 * given XLOG file has been removed.
3479 UpdateLastRemovedPtr(char *filename
)
3484 XLogFromFileName(filename
, &tli
, &segno
, wal_segment_size
);
3486 SpinLockAcquire(&XLogCtl
->info_lck
);
3487 if (segno
> XLogCtl
->lastRemovedSegNo
)
3488 XLogCtl
->lastRemovedSegNo
= segno
;
3489 SpinLockRelease(&XLogCtl
->info_lck
);
3493 * Remove all temporary log files in pg_wal
3495 * This is called at the beginning of recovery after a previous crash,
3496 * at a point where no other processes write fresh WAL data.
3499 RemoveTempXlogFiles(void)
3502 struct dirent
*xlde
;
3504 elog(DEBUG2
, "removing all temporary WAL segments");
3506 xldir
= AllocateDir(XLOGDIR
);
3507 while ((xlde
= ReadDir(xldir
, XLOGDIR
)) != NULL
)
3509 char path
[MAXPGPATH
];
3511 if (strncmp(xlde
->d_name
, "xlogtemp.", 9) != 0)
3514 snprintf(path
, MAXPGPATH
, XLOGDIR
"/%s", xlde
->d_name
);
3516 elog(DEBUG2
, "removed temporary WAL segment \"%s\"", path
);
3522 * Recycle or remove all log files older or equal to passed segno.
3524 * endptr is current (or recent) end of xlog, and lastredoptr is the
3525 * redo pointer of the last checkpoint. These are used to determine
3526 * whether we want to recycle rather than delete no-longer-wanted log files.
3528 * insertTLI is the current timeline for XLOG insertion. Any recycled
3529 * segments should be reused for this timeline.
3532 RemoveOldXlogFiles(XLogSegNo segno
, XLogRecPtr lastredoptr
, XLogRecPtr endptr
,
3533 TimeLineID insertTLI
)
3536 struct dirent
*xlde
;
3537 char lastoff
[MAXFNAMELEN
];
3538 XLogSegNo endlogSegNo
;
3539 XLogSegNo recycleSegNo
;
3541 /* Initialize info about where to try to recycle to */
3542 XLByteToSeg(endptr
, endlogSegNo
, wal_segment_size
);
3543 recycleSegNo
= XLOGfileslop(lastredoptr
);
3546 * Construct a filename of the last segment to be kept. The timeline ID
3547 * doesn't matter, we ignore that in the comparison. (During recovery,
3548 * InsertTimeLineID isn't set, so we can't use that.)
3550 XLogFileName(lastoff
, 0, segno
, wal_segment_size
);
3552 elog(DEBUG2
, "attempting to remove WAL segments older than log file %s",
3555 xldir
= AllocateDir(XLOGDIR
);
3557 while ((xlde
= ReadDir(xldir
, XLOGDIR
)) != NULL
)
3559 /* Ignore files that are not XLOG segments */
3560 if (!IsXLogFileName(xlde
->d_name
) &&
3561 !IsPartialXLogFileName(xlde
->d_name
))
3565 * We ignore the timeline part of the XLOG segment identifiers in
3566 * deciding whether a segment is still needed. This ensures that we
3567 * won't prematurely remove a segment from a parent timeline. We could
3568 * probably be a little more proactive about removing segments of
3569 * non-parent timelines, but that would be a whole lot more
3572 * We use the alphanumeric sorting property of the filenames to decide
3573 * which ones are earlier than the lastoff segment.
3575 if (strcmp(xlde
->d_name
+ 8, lastoff
+ 8) <= 0)
3577 if (XLogArchiveCheckDone(xlde
->d_name
))
3579 /* Update the last removed location in shared memory first */
3580 UpdateLastRemovedPtr(xlde
->d_name
);
3582 RemoveXlogFile(xlde
, recycleSegNo
, &endlogSegNo
, insertTLI
);
3591 * Remove WAL files that are not part of the given timeline's history.
3593 * This is called during recovery, whenever we switch to follow a new
3594 * timeline, and at the end of recovery when we create a new timeline. We
3595 * wouldn't otherwise care about extra WAL files lying in pg_wal, but they
3596 * might be leftover pre-allocated or recycled WAL segments on the old timeline
3597 * that we haven't used yet, and contain garbage. If we just leave them in
3598 * pg_wal, they will eventually be archived, and we can't let that happen.
3599 * Files that belong to our timeline history are valid, because we have
3600 * successfully replayed them, but from others we can't be sure.
3602 * 'switchpoint' is the current point in WAL where we switch to new timeline,
3603 * and 'newTLI' is the new timeline we switch to.
3606 RemoveNonParentXlogFiles(XLogRecPtr switchpoint
, TimeLineID newTLI
)
3609 struct dirent
*xlde
;
3610 char switchseg
[MAXFNAMELEN
];
3611 XLogSegNo endLogSegNo
;
3612 XLogSegNo switchLogSegNo
;
3613 XLogSegNo recycleSegNo
;
3616 * Initialize info about where to begin the work. This will recycle,
3617 * somewhat arbitrarily, 10 future segments.
3619 XLByteToPrevSeg(switchpoint
, switchLogSegNo
, wal_segment_size
);
3620 XLByteToSeg(switchpoint
, endLogSegNo
, wal_segment_size
);
3621 recycleSegNo
= endLogSegNo
+ 10;
3624 * Construct a filename of the last segment to be kept.
3626 XLogFileName(switchseg
, newTLI
, switchLogSegNo
, wal_segment_size
);
3628 elog(DEBUG2
, "attempting to remove WAL segments newer than log file %s",
3631 xldir
= AllocateDir(XLOGDIR
);
3633 while ((xlde
= ReadDir(xldir
, XLOGDIR
)) != NULL
)
3635 /* Ignore files that are not XLOG segments */
3636 if (!IsXLogFileName(xlde
->d_name
))
3640 * Remove files that are on a timeline older than the new one we're
3641 * switching to, but with a segment number >= the first segment on the
3644 if (strncmp(xlde
->d_name
, switchseg
, 8) < 0 &&
3645 strcmp(xlde
->d_name
+ 8, switchseg
+ 8) > 0)
3648 * If the file has already been marked as .ready, however, don't
3649 * remove it yet. It should be OK to remove it - files that are
3650 * not part of our timeline history are not required for recovery
3651 * - but seems safer to let them be archived and removed later.
3653 if (!XLogArchiveIsReady(xlde
->d_name
))
3654 RemoveXlogFile(xlde
, recycleSegNo
, &endLogSegNo
, newTLI
);
3662 * Recycle or remove a log file that's no longer needed.
3664 * segment_de is the dirent structure of the segment to recycle or remove.
3665 * recycleSegNo is the segment number to recycle up to. endlogSegNo is
3666 * the segment number of the current (or recent) end of WAL.
3668 * endlogSegNo gets incremented if the segment is recycled so as it is not
3669 * checked again with future callers of this function.
3671 * insertTLI is the current timeline for XLOG insertion. Any recycled segments
3672 * should be used for this timeline.
3675 RemoveXlogFile(const struct dirent
*segment_de
,
3676 XLogSegNo recycleSegNo
, XLogSegNo
*endlogSegNo
,
3677 TimeLineID insertTLI
)
3679 char path
[MAXPGPATH
];
3681 char newpath
[MAXPGPATH
];
3683 const char *segname
= segment_de
->d_name
;
3685 snprintf(path
, MAXPGPATH
, XLOGDIR
"/%s", segname
);
3688 * Before deleting the file, see if it can be recycled as a future log
3689 * segment. Only recycle normal files, because we don't want to recycle
3690 * symbolic links pointing to a separate archive directory.
3693 *endlogSegNo
<= recycleSegNo
&&
3694 XLogCtl
->InstallXLogFileSegmentActive
&& /* callee rechecks this */
3695 get_dirent_type(path
, segment_de
, false, DEBUG2
) == PGFILETYPE_REG
&&
3696 InstallXLogFileSegment(endlogSegNo
, path
,
3697 true, recycleSegNo
, insertTLI
))
3700 (errmsg_internal("recycled write-ahead log file \"%s\"",
3702 CheckpointStats
.ckpt_segs_recycled
++;
3703 /* Needn't recheck that slot on future iterations */
3708 /* No need for any more future segments, or recycling failed ... */
3712 (errmsg_internal("removing write-ahead log file \"%s\"",
3718 * On Windows, if another process (e.g another backend) holds the file
3719 * open in FILE_SHARE_DELETE mode, unlink will succeed, but the file
3720 * will still show up in directory listing until the last handle is
3721 * closed. To avoid confusing the lingering deleted file for a live
3722 * WAL file that needs to be archived, rename it before deleting it.
3724 * If another process holds the file open without FILE_SHARE_DELETE
3725 * flag, rename will fail. We'll try again at the next checkpoint.
3727 snprintf(newpath
, MAXPGPATH
, "%s.deleted", path
);
3728 if (rename(path
, newpath
) != 0)
3731 (errcode_for_file_access(),
3732 errmsg("could not rename file \"%s\": %m",
3736 rc
= durable_unlink(newpath
, LOG
);
3738 rc
= durable_unlink(path
, LOG
);
3742 /* Message already logged by durable_unlink() */
3745 CheckpointStats
.ckpt_segs_removed
++;
3748 XLogArchiveCleanup(segname
);
3752 * Verify whether pg_wal and pg_wal/archive_status exist.
3753 * If the latter does not exist, recreate it.
3755 * It is not the goal of this function to verify the contents of these
3756 * directories, but to help in cases where someone has performed a cluster
3757 * copy for PITR purposes but omitted pg_wal from the copy.
3759 * We could also recreate pg_wal if it doesn't exist, but a deliberate
3760 * policy decision was made not to. It is fairly common for pg_wal to be
3761 * a symlink, and if that was the DBA's intent then automatically making a
3762 * plain directory would result in degraded performance with no notice.
3765 ValidateXLOGDirectoryStructure(void)
3767 char path
[MAXPGPATH
];
3768 struct stat stat_buf
;
3770 /* Check for pg_wal; if it doesn't exist, error out */
3771 if (stat(XLOGDIR
, &stat_buf
) != 0 ||
3772 !S_ISDIR(stat_buf
.st_mode
))
3774 (errmsg("required WAL directory \"%s\" does not exist",
3777 /* Check for archive_status */
3778 snprintf(path
, MAXPGPATH
, XLOGDIR
"/archive_status");
3779 if (stat(path
, &stat_buf
) == 0)
3781 /* Check for weird cases where it exists but isn't a directory */
3782 if (!S_ISDIR(stat_buf
.st_mode
))
3784 (errmsg("required WAL directory \"%s\" does not exist",
3790 (errmsg("creating missing WAL directory \"%s\"", path
)));
3791 if (MakePGDirectory(path
) < 0)
3793 (errmsg("could not create missing directory \"%s\": %m",
3799 * Remove previous backup history files. This also retries creation of
3800 * .ready files for any backup history files for which XLogArchiveNotify
3804 CleanupBackupHistory(void)
3807 struct dirent
*xlde
;
3808 char path
[MAXPGPATH
+ sizeof(XLOGDIR
)];
3810 xldir
= AllocateDir(XLOGDIR
);
3812 while ((xlde
= ReadDir(xldir
, XLOGDIR
)) != NULL
)
3814 if (IsBackupHistoryFileName(xlde
->d_name
))
3816 if (XLogArchiveCheckDone(xlde
->d_name
))
3818 elog(DEBUG2
, "removing WAL backup history file \"%s\"",
3820 snprintf(path
, sizeof(path
), XLOGDIR
"/%s", xlde
->d_name
);
3822 XLogArchiveCleanup(xlde
->d_name
);
3831 * I/O routines for pg_control
3833 * *ControlFile is a buffer in shared memory that holds an image of the
3834 * contents of pg_control. WriteControlFile() initializes pg_control
3835 * given a preloaded buffer, ReadControlFile() loads the buffer from
3836 * the pg_control file (during postmaster or standalone-backend startup),
3837 * and UpdateControlFile() rewrites pg_control after we modify xlog state.
3838 * InitControlFile() fills the buffer with initial values.
3840 * For simplicity, WriteControlFile() initializes the fields of pg_control
3841 * that are related to checking backend/database compatibility, and
3842 * ReadControlFile() verifies they are correct. We could split out the
3843 * I/O and compatibility-check functions, but there seems no need currently.
3847 InitControlFile(uint64 sysidentifier
)
3849 char mock_auth_nonce
[MOCK_AUTH_NONCE_LEN
];
3852 * Generate a random nonce. This is used for authentication requests that
3853 * will fail because the user does not exist. The nonce is used to create
3854 * a genuine-looking password challenge for the non-existent user, in lieu
3855 * of an actual stored password.
3857 if (!pg_strong_random(mock_auth_nonce
, MOCK_AUTH_NONCE_LEN
))
3859 (errcode(ERRCODE_INTERNAL_ERROR
),
3860 errmsg("could not generate secret authorization token")));
3862 memset(ControlFile
, 0, sizeof(ControlFileData
));
3863 /* Initialize pg_control status fields */
3864 ControlFile
->system_identifier
= sysidentifier
;
3865 memcpy(ControlFile
->mock_authentication_nonce
, mock_auth_nonce
, MOCK_AUTH_NONCE_LEN
);
3866 ControlFile
->state
= DB_SHUTDOWNED
;
3867 ControlFile
->unloggedLSN
= FirstNormalUnloggedLSN
;
3869 /* Set important parameter values for use when replaying WAL */
3870 ControlFile
->MaxConnections
= MaxConnections
;
3871 ControlFile
->max_worker_processes
= max_worker_processes
;
3872 ControlFile
->max_wal_senders
= max_wal_senders
;
3873 ControlFile
->max_prepared_xacts
= max_prepared_xacts
;
3874 ControlFile
->max_locks_per_xact
= max_locks_per_xact
;
3875 ControlFile
->wal_level
= wal_level
;
3876 ControlFile
->wal_log_hints
= wal_log_hints
;
3877 ControlFile
->track_commit_timestamp
= track_commit_timestamp
;
3878 ControlFile
->data_checksum_version
= bootstrap_data_checksum_version
;
3882 WriteControlFile(void)
3885 char buffer
[PG_CONTROL_FILE_SIZE
]; /* need not be aligned */
3888 * Initialize version and compatibility-check fields
3890 ControlFile
->pg_control_version
= PG_CONTROL_VERSION
;
3891 ControlFile
->catalog_version_no
= CATALOG_VERSION_NO
;
3893 ControlFile
->maxAlign
= MAXIMUM_ALIGNOF
;
3894 ControlFile
->floatFormat
= FLOATFORMAT_VALUE
;
3896 ControlFile
->blcksz
= BLCKSZ
;
3897 ControlFile
->relseg_size
= RELSEG_SIZE
;
3898 ControlFile
->xlog_blcksz
= XLOG_BLCKSZ
;
3899 ControlFile
->xlog_seg_size
= wal_segment_size
;
3901 ControlFile
->nameDataLen
= NAMEDATALEN
;
3902 ControlFile
->indexMaxKeys
= INDEX_MAX_KEYS
;
3904 ControlFile
->toast_max_chunk_size
= TOAST_MAX_CHUNK_SIZE
;
3905 ControlFile
->loblksize
= LOBLKSIZE
;
3907 ControlFile
->float8ByVal
= FLOAT8PASSBYVAL
;
3909 /* Contents are protected with a CRC */
3910 INIT_CRC32C(ControlFile
->crc
);
3911 COMP_CRC32C(ControlFile
->crc
,
3912 (char *) ControlFile
,
3913 offsetof(ControlFileData
, crc
));
3914 FIN_CRC32C(ControlFile
->crc
);
3917 * We write out PG_CONTROL_FILE_SIZE bytes into pg_control, zero-padding
3918 * the excess over sizeof(ControlFileData). This reduces the odds of
3919 * premature-EOF errors when reading pg_control. We'll still fail when we
3920 * check the contents of the file, but hopefully with a more specific
3921 * error than "couldn't read pg_control".
3923 memset(buffer
, 0, PG_CONTROL_FILE_SIZE
);
3924 memcpy(buffer
, ControlFile
, sizeof(ControlFileData
));
3926 fd
= BasicOpenFile(XLOG_CONTROL_FILE
,
3927 O_RDWR
| O_CREAT
| O_EXCL
| PG_BINARY
);
3930 (errcode_for_file_access(),
3931 errmsg("could not create file \"%s\": %m",
3932 XLOG_CONTROL_FILE
)));
3935 pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_WRITE
);
3936 if (write(fd
, buffer
, PG_CONTROL_FILE_SIZE
) != PG_CONTROL_FILE_SIZE
)
3938 /* if write didn't set errno, assume problem is no disk space */
3942 (errcode_for_file_access(),
3943 errmsg("could not write to file \"%s\": %m",
3944 XLOG_CONTROL_FILE
)));
3946 pgstat_report_wait_end();
3948 pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_SYNC
);
3949 if (pg_fsync(fd
) != 0)
3951 (errcode_for_file_access(),
3952 errmsg("could not fsync file \"%s\": %m",
3953 XLOG_CONTROL_FILE
)));
3954 pgstat_report_wait_end();
3958 (errcode_for_file_access(),
3959 errmsg("could not close file \"%s\": %m",
3960 XLOG_CONTROL_FILE
)));
3964 ReadControlFile(void)
3968 static char wal_segsz_str
[20];
3974 fd
= BasicOpenFile(XLOG_CONTROL_FILE
,
3975 O_RDWR
| PG_BINARY
);
3978 (errcode_for_file_access(),
3979 errmsg("could not open file \"%s\": %m",
3980 XLOG_CONTROL_FILE
)));
3982 pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ
);
3983 r
= read(fd
, ControlFile
, sizeof(ControlFileData
));
3984 if (r
!= sizeof(ControlFileData
))
3988 (errcode_for_file_access(),
3989 errmsg("could not read file \"%s\": %m",
3990 XLOG_CONTROL_FILE
)));
3993 (errcode(ERRCODE_DATA_CORRUPTED
),
3994 errmsg("could not read file \"%s\": read %d of %zu",
3995 XLOG_CONTROL_FILE
, r
, sizeof(ControlFileData
))));
3997 pgstat_report_wait_end();
4002 * Check for expected pg_control format version. If this is wrong, the
4003 * CRC check will likely fail because we'll be checking the wrong number
4004 * of bytes. Complaining about wrong version will probably be more
4005 * enlightening than complaining about wrong CRC.
4008 if (ControlFile
->pg_control_version
!= PG_CONTROL_VERSION
&& ControlFile
->pg_control_version
% 65536 == 0 && ControlFile
->pg_control_version
/ 65536 != 0)
4010 (errmsg("database files are incompatible with server"),
4011 errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4012 " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4013 ControlFile
->pg_control_version
, ControlFile
->pg_control_version
,
4014 PG_CONTROL_VERSION
, PG_CONTROL_VERSION
),
4015 errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4017 if (ControlFile
->pg_control_version
!= PG_CONTROL_VERSION
)
4019 (errmsg("database files are incompatible with server"),
4020 errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4021 " but the server was compiled with PG_CONTROL_VERSION %d.",
4022 ControlFile
->pg_control_version
, PG_CONTROL_VERSION
),
4023 errhint("It looks like you need to initdb.")));
4025 /* Now check the CRC. */
4028 (char *) ControlFile
,
4029 offsetof(ControlFileData
, crc
));
4032 if (!EQ_CRC32C(crc
, ControlFile
->crc
))
4034 (errmsg("incorrect checksum in control file")));
4037 * Do compatibility checking immediately. If the database isn't
4038 * compatible with the backend executable, we want to abort before we can
4039 * possibly do any damage.
4041 if (ControlFile
->catalog_version_no
!= CATALOG_VERSION_NO
)
4043 (errmsg("database files are incompatible with server"),
4044 errdetail("The database cluster was initialized with CATALOG_VERSION_NO %d,"
4045 " but the server was compiled with CATALOG_VERSION_NO %d.",
4046 ControlFile
->catalog_version_no
, CATALOG_VERSION_NO
),
4047 errhint("It looks like you need to initdb.")));
4048 if (ControlFile
->maxAlign
!= MAXIMUM_ALIGNOF
)
4050 (errmsg("database files are incompatible with server"),
4051 errdetail("The database cluster was initialized with MAXALIGN %d,"
4052 " but the server was compiled with MAXALIGN %d.",
4053 ControlFile
->maxAlign
, MAXIMUM_ALIGNOF
),
4054 errhint("It looks like you need to initdb.")));
4055 if (ControlFile
->floatFormat
!= FLOATFORMAT_VALUE
)
4057 (errmsg("database files are incompatible with server"),
4058 errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4059 errhint("It looks like you need to initdb.")));
4060 if (ControlFile
->blcksz
!= BLCKSZ
)
4062 (errmsg("database files are incompatible with server"),
4063 errdetail("The database cluster was initialized with BLCKSZ %d,"
4064 " but the server was compiled with BLCKSZ %d.",
4065 ControlFile
->blcksz
, BLCKSZ
),
4066 errhint("It looks like you need to recompile or initdb.")));
4067 if (ControlFile
->relseg_size
!= RELSEG_SIZE
)
4069 (errmsg("database files are incompatible with server"),
4070 errdetail("The database cluster was initialized with RELSEG_SIZE %d,"
4071 " but the server was compiled with RELSEG_SIZE %d.",
4072 ControlFile
->relseg_size
, RELSEG_SIZE
),
4073 errhint("It looks like you need to recompile or initdb.")));
4074 if (ControlFile
->xlog_blcksz
!= XLOG_BLCKSZ
)
4076 (errmsg("database files are incompatible with server"),
4077 errdetail("The database cluster was initialized with XLOG_BLCKSZ %d,"
4078 " but the server was compiled with XLOG_BLCKSZ %d.",
4079 ControlFile
->xlog_blcksz
, XLOG_BLCKSZ
),
4080 errhint("It looks like you need to recompile or initdb.")));
4081 if (ControlFile
->nameDataLen
!= NAMEDATALEN
)
4083 (errmsg("database files are incompatible with server"),
4084 errdetail("The database cluster was initialized with NAMEDATALEN %d,"
4085 " but the server was compiled with NAMEDATALEN %d.",
4086 ControlFile
->nameDataLen
, NAMEDATALEN
),
4087 errhint("It looks like you need to recompile or initdb.")));
4088 if (ControlFile
->indexMaxKeys
!= INDEX_MAX_KEYS
)
4090 (errmsg("database files are incompatible with server"),
4091 errdetail("The database cluster was initialized with INDEX_MAX_KEYS %d,"
4092 " but the server was compiled with INDEX_MAX_KEYS %d.",
4093 ControlFile
->indexMaxKeys
, INDEX_MAX_KEYS
),
4094 errhint("It looks like you need to recompile or initdb.")));
4095 if (ControlFile
->toast_max_chunk_size
!= TOAST_MAX_CHUNK_SIZE
)
4097 (errmsg("database files are incompatible with server"),
4098 errdetail("The database cluster was initialized with TOAST_MAX_CHUNK_SIZE %d,"
4099 " but the server was compiled with TOAST_MAX_CHUNK_SIZE %d.",
4100 ControlFile
->toast_max_chunk_size
, (int) TOAST_MAX_CHUNK_SIZE
),
4101 errhint("It looks like you need to recompile or initdb.")));
4102 if (ControlFile
->loblksize
!= LOBLKSIZE
)
4104 (errmsg("database files are incompatible with server"),
4105 errdetail("The database cluster was initialized with LOBLKSIZE %d,"
4106 " but the server was compiled with LOBLKSIZE %d.",
4107 ControlFile
->loblksize
, (int) LOBLKSIZE
),
4108 errhint("It looks like you need to recompile or initdb.")));
4110 #ifdef USE_FLOAT8_BYVAL
4111 if (ControlFile
->float8ByVal
!= true)
4113 (errmsg("database files are incompatible with server"),
4114 errdetail("The database cluster was initialized without USE_FLOAT8_BYVAL"
4115 " but the server was compiled with USE_FLOAT8_BYVAL."),
4116 errhint("It looks like you need to recompile or initdb.")));
4118 if (ControlFile
->float8ByVal
!= false)
4120 (errmsg("database files are incompatible with server"),
4121 errdetail("The database cluster was initialized with USE_FLOAT8_BYVAL"
4122 " but the server was compiled without USE_FLOAT8_BYVAL."),
4123 errhint("It looks like you need to recompile or initdb.")));
4126 wal_segment_size
= ControlFile
->xlog_seg_size
;
4128 if (!IsValidWalSegSize(wal_segment_size
))
4129 ereport(ERROR
, (errcode(ERRCODE_INVALID_PARAMETER_VALUE
),
4130 errmsg_plural("WAL segment size must be a power of two between 1 MB and 1 GB, but the control file specifies %d byte",
4131 "WAL segment size must be a power of two between 1 MB and 1 GB, but the control file specifies %d bytes",
4133 wal_segment_size
)));
4135 snprintf(wal_segsz_str
, sizeof(wal_segsz_str
), "%d", wal_segment_size
);
4136 SetConfigOption("wal_segment_size", wal_segsz_str
, PGC_INTERNAL
,
4137 PGC_S_DYNAMIC_DEFAULT
);
4139 /* check and update variables dependent on wal_segment_size */
4140 if (ConvertToXSegs(min_wal_size_mb
, wal_segment_size
) < 2)
4141 ereport(ERROR
, (errcode(ERRCODE_INVALID_PARAMETER_VALUE
),
4142 errmsg("\"min_wal_size\" must be at least twice \"wal_segment_size\"")));
4144 if (ConvertToXSegs(max_wal_size_mb
, wal_segment_size
) < 2)
4145 ereport(ERROR
, (errcode(ERRCODE_INVALID_PARAMETER_VALUE
),
4146 errmsg("\"max_wal_size\" must be at least twice \"wal_segment_size\"")));
4148 UsableBytesInSegment
=
4149 (wal_segment_size
/ XLOG_BLCKSZ
* UsableBytesInPage
) -
4150 (SizeOfXLogLongPHD
- SizeOfXLogShortPHD
);
4152 CalculateCheckpointSegments();
4154 /* Make the initdb settings visible as GUC variables, too */
4155 SetConfigOption("data_checksums", DataChecksumsEnabled() ? "yes" : "no",
4156 PGC_INTERNAL
, PGC_S_DYNAMIC_DEFAULT
);
4160 * Utility wrapper to update the control file. Note that the control
4161 * file gets flushed.
4164 UpdateControlFile(void)
4166 update_controlfile(DataDir
, ControlFile
, true);
4170 * Returns the unique system identifier from control file.
4173 GetSystemIdentifier(void)
4175 Assert(ControlFile
!= NULL
);
4176 return ControlFile
->system_identifier
;
4180 * Returns the random nonce from control file.
4183 GetMockAuthenticationNonce(void)
4185 Assert(ControlFile
!= NULL
);
4186 return ControlFile
->mock_authentication_nonce
;
4190 * Are checksums enabled for data pages?
4193 DataChecksumsEnabled(void)
4195 Assert(ControlFile
!= NULL
);
4196 return (ControlFile
->data_checksum_version
> 0);
4200 * Returns a fake LSN for unlogged relations.
4202 * Each call generates an LSN that is greater than any previous value
4203 * returned. The current counter value is saved and restored across clean
4204 * shutdowns, but like unlogged relations, does not survive a crash. This can
4205 * be used in lieu of real LSN values returned by XLogInsert, if you need an
4206 * LSN-like increasing sequence of numbers without writing any WAL.
4209 GetFakeLSNForUnloggedRel(void)
4211 XLogRecPtr nextUnloggedLSN
;
4213 /* increment the unloggedLSN counter, need SpinLock */
4214 SpinLockAcquire(&XLogCtl
->ulsn_lck
);
4215 nextUnloggedLSN
= XLogCtl
->unloggedLSN
++;
4216 SpinLockRelease(&XLogCtl
->ulsn_lck
);
4218 return nextUnloggedLSN
;
4222 * Auto-tune the number of XLOG buffers.
4224 * The preferred setting for wal_buffers is about 3% of shared_buffers, with
4225 * a maximum of one XLOG segment (there is little reason to think that more
4226 * is helpful, at least so long as we force an fsync when switching log files)
4227 * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
4228 * 9.1, when auto-tuning was added).
4230 * This should not be called until NBuffers has received its final value.
4233 XLOGChooseNumBuffers(void)
4237 xbuffers
= NBuffers
/ 32;
4238 if (xbuffers
> (wal_segment_size
/ XLOG_BLCKSZ
))
4239 xbuffers
= (wal_segment_size
/ XLOG_BLCKSZ
);
4246 * GUC check_hook for wal_buffers
4249 check_wal_buffers(int *newval
, void **extra
, GucSource source
)
4252 * -1 indicates a request for auto-tune.
4257 * If we haven't yet changed the boot_val default of -1, just let it
4258 * be. We'll fix it when XLOGShmemSize is called.
4260 if (XLOGbuffers
== -1)
4263 /* Otherwise, substitute the auto-tune value */
4264 *newval
= XLOGChooseNumBuffers();
4268 * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
4269 * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
4270 * the case, we just silently treat such values as a request for the
4271 * minimum. (We could throw an error instead, but that doesn't seem very
4281 * GUC check_hook for wal_consistency_checking
4284 check_wal_consistency_checking(char **newval
, void **extra
, GucSource source
)
4289 bool newwalconsistency
[RM_MAX_ID
+ 1];
4291 /* Initialize the array */
4292 MemSet(newwalconsistency
, 0, (RM_MAX_ID
+ 1) * sizeof(bool));
4294 /* Need a modifiable copy of string */
4295 rawstring
= pstrdup(*newval
);
4297 /* Parse string into list of identifiers */
4298 if (!SplitIdentifierString(rawstring
, ',', &elemlist
))
4300 /* syntax error in list */
4301 GUC_check_errdetail("List syntax is invalid.");
4303 list_free(elemlist
);
4307 foreach(l
, elemlist
)
4309 char *tok
= (char *) lfirst(l
);
4312 /* Check for 'all'. */
4313 if (pg_strcasecmp(tok
, "all") == 0)
4315 for (rmid
= 0; rmid
<= RM_MAX_ID
; rmid
++)
4316 if (RmgrIdExists(rmid
) && GetRmgr(rmid
).rm_mask
!= NULL
)
4317 newwalconsistency
[rmid
] = true;
4321 /* Check if the token matches any known resource manager. */
4324 for (rmid
= 0; rmid
<= RM_MAX_ID
; rmid
++)
4326 if (RmgrIdExists(rmid
) && GetRmgr(rmid
).rm_mask
!= NULL
&&
4327 pg_strcasecmp(tok
, GetRmgr(rmid
).rm_name
) == 0)
4329 newwalconsistency
[rmid
] = true;
4337 * During startup, it might be a not-yet-loaded custom
4338 * resource manager. Defer checking until
4339 * InitializeWalConsistencyChecking().
4341 if (!process_shared_preload_libraries_done
)
4343 check_wal_consistency_checking_deferred
= true;
4347 GUC_check_errdetail("Unrecognized key word: \"%s\".", tok
);
4349 list_free(elemlist
);
4357 list_free(elemlist
);
4359 /* assign new value */
4360 *extra
= guc_malloc(ERROR
, (RM_MAX_ID
+ 1) * sizeof(bool));
4361 memcpy(*extra
, newwalconsistency
, (RM_MAX_ID
+ 1) * sizeof(bool));
4366 * GUC assign_hook for wal_consistency_checking
4369 assign_wal_consistency_checking(const char *newval
, void *extra
)
4372 * If some checks were deferred, it's possible that the checks will fail
4373 * later during InitializeWalConsistencyChecking(). But in that case, the
4374 * postmaster will exit anyway, so it's safe to proceed with the
4377 * Any built-in resource managers specified are assigned immediately,
4378 * which affects WAL created before shared_preload_libraries are
4379 * processed. Any custom resource managers specified won't be assigned
4380 * until after shared_preload_libraries are processed, but that's OK
4381 * because WAL for a custom resource manager can't be written before the
4382 * module is loaded anyway.
4384 wal_consistency_checking
= extra
;
4388 * InitializeWalConsistencyChecking: run after loading custom resource managers
4390 * If any unknown resource managers were specified in the
4391 * wal_consistency_checking GUC, processing was deferred. Now that
4392 * shared_preload_libraries have been loaded, process wal_consistency_checking
4396 InitializeWalConsistencyChecking(void)
4398 Assert(process_shared_preload_libraries_done
);
4400 if (check_wal_consistency_checking_deferred
)
4402 struct config_generic
*guc
;
4404 guc
= find_option("wal_consistency_checking", false, false, ERROR
);
4406 check_wal_consistency_checking_deferred
= false;
4408 set_config_option_ext("wal_consistency_checking",
4409 wal_consistency_checking_string
,
4410 guc
->scontext
, guc
->source
, guc
->srole
,
4411 GUC_ACTION_SET
, true, ERROR
, false);
4413 /* checking should not be deferred again */
4414 Assert(!check_wal_consistency_checking_deferred
);
4419 * GUC show_hook for archive_command
4422 show_archive_command(void)
4424 if (XLogArchivingActive())
4425 return XLogArchiveCommand
;
4427 return "(disabled)";
4431 * GUC show_hook for in_hot_standby
4434 show_in_hot_standby(void)
4437 * We display the actual state based on shared memory, so that this GUC
4438 * reports up-to-date state if examined intra-query. The underlying
4439 * variable (in_hot_standby_guc) changes only when we transmit a new value
4442 return RecoveryInProgress() ? "on" : "off";
4447 * Read the control file, set respective GUCs.
4449 * This is to be called during startup, including a crash recovery cycle,
4450 * unless in bootstrap mode, where no control file yet exists. As there's no
4451 * usable shared memory yet (its sizing can depend on the contents of the
4452 * control file!), first store the contents in local memory. XLOGShmemInit()
4453 * will then copy it to shared memory later.
4455 * reset just controls whether previous contents are to be expected (in the
4456 * reset case, there's a dangling pointer into old shared memory), or not.
4459 LocalProcessControlFile(bool reset
)
4461 Assert(reset
|| ControlFile
== NULL
);
4462 ControlFile
= palloc(sizeof(ControlFileData
));
4467 * Initialization of shared memory for XLOG
4475 * If the value of wal_buffers is -1, use the preferred auto-tune value.
4476 * This isn't an amazingly clean place to do this, but we must wait till
4477 * NBuffers has received its final value, and must do it before using the
4478 * value of XLOGbuffers to do anything important.
4480 * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
4481 * However, if the DBA explicitly set wal_buffers = -1 in the config file,
4482 * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
4483 * the matter with PGC_S_OVERRIDE.
4485 if (XLOGbuffers
== -1)
4489 snprintf(buf
, sizeof(buf
), "%d", XLOGChooseNumBuffers());
4490 SetConfigOption("wal_buffers", buf
, PGC_POSTMASTER
,
4491 PGC_S_DYNAMIC_DEFAULT
);
4492 if (XLOGbuffers
== -1) /* failed to apply it? */
4493 SetConfigOption("wal_buffers", buf
, PGC_POSTMASTER
,
4496 Assert(XLOGbuffers
> 0);
4499 size
= sizeof(XLogCtlData
);
4501 /* WAL insertion locks, plus alignment */
4502 size
= add_size(size
, mul_size(sizeof(WALInsertLockPadded
), NUM_XLOGINSERT_LOCKS
+ 1));
4503 /* xlblocks array */
4504 size
= add_size(size
, mul_size(sizeof(XLogRecPtr
), XLOGbuffers
));
4505 /* extra alignment padding for XLOG I/O buffers */
4506 size
= add_size(size
, XLOG_BLCKSZ
);
4507 /* and the buffers themselves */
4508 size
= add_size(size
, mul_size(XLOG_BLCKSZ
, XLOGbuffers
));
4511 * Note: we don't count ControlFileData, it comes out of the "slop factor"
4512 * added by CreateSharedMemoryAndSemaphores. This lets us use this
4513 * routine again below to compute the actual allocation size.
4526 ControlFileData
*localControlFile
;
4531 * Create a memory context for WAL debugging that's exempt from the normal
4532 * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
4533 * an allocation fails, but wal_debug is not for production use anyway.
4535 if (walDebugCxt
== NULL
)
4537 walDebugCxt
= AllocSetContextCreate(TopMemoryContext
,
4539 ALLOCSET_DEFAULT_SIZES
);
4540 MemoryContextAllowInCriticalSection(walDebugCxt
, true);
4545 XLogCtl
= (XLogCtlData
*)
4546 ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog
);
4548 localControlFile
= ControlFile
;
4549 ControlFile
= (ControlFileData
*)
4550 ShmemInitStruct("Control File", sizeof(ControlFileData
), &foundCFile
);
4552 if (foundCFile
|| foundXLog
)
4554 /* both should be present or neither */
4555 Assert(foundCFile
&& foundXLog
);
4557 /* Initialize local copy of WALInsertLocks */
4558 WALInsertLocks
= XLogCtl
->Insert
.WALInsertLocks
;
4560 if (localControlFile
)
4561 pfree(localControlFile
);
4564 memset(XLogCtl
, 0, sizeof(XLogCtlData
));
4567 * Already have read control file locally, unless in bootstrap mode. Move
4568 * contents into shared memory.
4570 if (localControlFile
)
4572 memcpy(ControlFile
, localControlFile
, sizeof(ControlFileData
));
4573 pfree(localControlFile
);
4577 * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
4578 * multiple of the alignment for same, so no extra alignment padding is
4581 allocptr
= ((char *) XLogCtl
) + sizeof(XLogCtlData
);
4582 XLogCtl
->xlblocks
= (XLogRecPtr
*) allocptr
;
4583 memset(XLogCtl
->xlblocks
, 0, sizeof(XLogRecPtr
) * XLOGbuffers
);
4584 allocptr
+= sizeof(XLogRecPtr
) * XLOGbuffers
;
4587 /* WAL insertion locks. Ensure they're aligned to the full padded size */
4588 allocptr
+= sizeof(WALInsertLockPadded
) -
4589 ((uintptr_t) allocptr
) % sizeof(WALInsertLockPadded
);
4590 WALInsertLocks
= XLogCtl
->Insert
.WALInsertLocks
=
4591 (WALInsertLockPadded
*) allocptr
;
4592 allocptr
+= sizeof(WALInsertLockPadded
) * NUM_XLOGINSERT_LOCKS
;
4594 for (i
= 0; i
< NUM_XLOGINSERT_LOCKS
; i
++)
4596 LWLockInitialize(&WALInsertLocks
[i
].l
.lock
, LWTRANCHE_WAL_INSERT
);
4597 WALInsertLocks
[i
].l
.insertingAt
= InvalidXLogRecPtr
;
4598 WALInsertLocks
[i
].l
.lastImportantAt
= InvalidXLogRecPtr
;
4602 * Align the start of the page buffers to a full xlog block size boundary.
4603 * This simplifies some calculations in XLOG insertion. It is also
4604 * required for O_DIRECT.
4606 allocptr
= (char *) TYPEALIGN(XLOG_BLCKSZ
, allocptr
);
4607 XLogCtl
->pages
= allocptr
;
4608 memset(XLogCtl
->pages
, 0, (Size
) XLOG_BLCKSZ
* XLOGbuffers
);
4611 * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
4612 * in additional info.)
4614 XLogCtl
->XLogCacheBlck
= XLOGbuffers
- 1;
4615 XLogCtl
->SharedRecoveryState
= RECOVERY_STATE_CRASH
;
4616 XLogCtl
->InstallXLogFileSegmentActive
= false;
4617 XLogCtl
->WalWriterSleeping
= false;
4619 SpinLockInit(&XLogCtl
->Insert
.insertpos_lck
);
4620 SpinLockInit(&XLogCtl
->info_lck
);
4621 SpinLockInit(&XLogCtl
->ulsn_lck
);
4625 * This func must be called ONCE on system install. It creates pg_control
4626 * and the initial XLOG segment.
4631 CheckPoint checkPoint
;
4633 XLogPageHeader page
;
4634 XLogLongPageHeader longpage
;
4637 uint64 sysidentifier
;
4641 /* allow ordinary WAL segment creation, like StartupXLOG() would */
4642 SetInstallXLogFileSegmentActive();
4645 * Select a hopefully-unique system identifier code for this installation.
4646 * We use the result of gettimeofday(), including the fractional seconds
4647 * field, as being about as unique as we can easily get. (Think not to
4648 * use random(), since it hasn't been seeded and there's no portable way
4649 * to seed it other than the system clock value...) The upper half of the
4650 * uint64 value is just the tv_sec part, while the lower half contains the
4651 * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
4652 * PID for a little extra uniqueness. A person knowing this encoding can
4653 * determine the initialization time of the installation, which could
4654 * perhaps be useful sometimes.
4656 gettimeofday(&tv
, NULL
);
4657 sysidentifier
= ((uint64
) tv
.tv_sec
) << 32;
4658 sysidentifier
|= ((uint64
) tv
.tv_usec
) << 12;
4659 sysidentifier
|= getpid() & 0xFFF;
4661 /* page buffer must be aligned suitably for O_DIRECT */
4662 buffer
= (char *) palloc(XLOG_BLCKSZ
+ XLOG_BLCKSZ
);
4663 page
= (XLogPageHeader
) TYPEALIGN(XLOG_BLCKSZ
, buffer
);
4664 memset(page
, 0, XLOG_BLCKSZ
);
4667 * Set up information for the initial checkpoint record
4669 * The initial checkpoint record is written to the beginning of the WAL
4670 * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
4671 * used, so that we can use 0/0 to mean "before any valid WAL segment".
4673 checkPoint
.redo
= wal_segment_size
+ SizeOfXLogLongPHD
;
4674 checkPoint
.ThisTimeLineID
= BootstrapTimeLineID
;
4675 checkPoint
.PrevTimeLineID
= BootstrapTimeLineID
;
4676 checkPoint
.fullPageWrites
= fullPageWrites
;
4677 checkPoint
.nextXid
=
4678 FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId
);
4679 checkPoint
.nextOid
= FirstGenbkiObjectId
;
4680 checkPoint
.nextMulti
= FirstMultiXactId
;
4681 checkPoint
.nextMultiOffset
= 0;
4682 checkPoint
.oldestXid
= FirstNormalTransactionId
;
4683 checkPoint
.oldestXidDB
= Template1DbOid
;
4684 checkPoint
.oldestMulti
= FirstMultiXactId
;
4685 checkPoint
.oldestMultiDB
= Template1DbOid
;
4686 checkPoint
.oldestCommitTsXid
= InvalidTransactionId
;
4687 checkPoint
.newestCommitTsXid
= InvalidTransactionId
;
4688 checkPoint
.time
= (pg_time_t
) time(NULL
);
4689 checkPoint
.oldestActiveXid
= InvalidTransactionId
;
4691 ShmemVariableCache
->nextXid
= checkPoint
.nextXid
;
4692 ShmemVariableCache
->nextOid
= checkPoint
.nextOid
;
4693 ShmemVariableCache
->oidCount
= 0;
4694 MultiXactSetNextMXact(checkPoint
.nextMulti
, checkPoint
.nextMultiOffset
);
4695 AdvanceOldestClogXid(checkPoint
.oldestXid
);
4696 SetTransactionIdLimit(checkPoint
.oldestXid
, checkPoint
.oldestXidDB
);
4697 SetMultiXactIdLimit(checkPoint
.oldestMulti
, checkPoint
.oldestMultiDB
, true);
4698 SetCommitTsLimit(InvalidTransactionId
, InvalidTransactionId
);
4700 /* Set up the XLOG page header */
4701 page
->xlp_magic
= XLOG_PAGE_MAGIC
;
4702 page
->xlp_info
= XLP_LONG_HEADER
;
4703 page
->xlp_tli
= BootstrapTimeLineID
;
4704 page
->xlp_pageaddr
= wal_segment_size
;
4705 longpage
= (XLogLongPageHeader
) page
;
4706 longpage
->xlp_sysid
= sysidentifier
;
4707 longpage
->xlp_seg_size
= wal_segment_size
;
4708 longpage
->xlp_xlog_blcksz
= XLOG_BLCKSZ
;
4710 /* Insert the initial checkpoint record */
4711 recptr
= ((char *) page
+ SizeOfXLogLongPHD
);
4712 record
= (XLogRecord
*) recptr
;
4713 record
->xl_prev
= 0;
4714 record
->xl_xid
= InvalidTransactionId
;
4715 record
->xl_tot_len
= SizeOfXLogRecord
+ SizeOfXLogRecordDataHeaderShort
+ sizeof(checkPoint
);
4716 record
->xl_info
= XLOG_CHECKPOINT_SHUTDOWN
;
4717 record
->xl_rmid
= RM_XLOG_ID
;
4718 recptr
+= SizeOfXLogRecord
;
4719 /* fill the XLogRecordDataHeaderShort struct */
4720 *(recptr
++) = (char) XLR_BLOCK_ID_DATA_SHORT
;
4721 *(recptr
++) = sizeof(checkPoint
);
4722 memcpy(recptr
, &checkPoint
, sizeof(checkPoint
));
4723 recptr
+= sizeof(checkPoint
);
4724 Assert(recptr
- (char *) record
== record
->xl_tot_len
);
4727 COMP_CRC32C(crc
, ((char *) record
) + SizeOfXLogRecord
, record
->xl_tot_len
- SizeOfXLogRecord
);
4728 COMP_CRC32C(crc
, (char *) record
, offsetof(XLogRecord
, xl_crc
));
4730 record
->xl_crc
= crc
;
4732 /* Create first XLOG segment file */
4733 openLogTLI
= BootstrapTimeLineID
;
4734 openLogFile
= XLogFileInit(1, BootstrapTimeLineID
);
4737 * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
4738 * close the file again in a moment.
4741 /* Write the first page with the initial record */
4743 pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE
);
4744 if (write(openLogFile
, page
, XLOG_BLCKSZ
) != XLOG_BLCKSZ
)
4746 /* if write didn't set errno, assume problem is no disk space */
4750 (errcode_for_file_access(),
4751 errmsg("could not write bootstrap write-ahead log file: %m")));
4753 pgstat_report_wait_end();
4755 pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC
);
4756 if (pg_fsync(openLogFile
) != 0)
4758 (errcode_for_file_access(),
4759 errmsg("could not fsync bootstrap write-ahead log file: %m")));
4760 pgstat_report_wait_end();
4762 if (close(openLogFile
) != 0)
4764 (errcode_for_file_access(),
4765 errmsg("could not close bootstrap write-ahead log file: %m")));
4769 /* Now create pg_control */
4770 InitControlFile(sysidentifier
);
4771 ControlFile
->time
= checkPoint
.time
;
4772 ControlFile
->checkPoint
= checkPoint
.redo
;
4773 ControlFile
->checkPointCopy
= checkPoint
;
4775 /* some additional ControlFile fields are set in WriteControlFile() */
4778 /* Bootstrap the commit log, too */
4780 BootStrapCommitTs();
4781 BootStrapSUBTRANS();
4782 BootStrapMultiXact();
4787 * Force control file to be read - in contrast to normal processing we'd
4788 * otherwise never run the checks and GUC related initializations therein.
4794 str_time(pg_time_t tnow
)
4796 static char buf
[128];
4798 pg_strftime(buf
, sizeof(buf
),
4799 "%Y-%m-%d %H:%M:%S %Z",
4800 pg_localtime(&tnow
, log_timezone
));
4806 * Initialize the first WAL segment on new timeline.
4809 XLogInitNewTimeline(TimeLineID endTLI
, XLogRecPtr endOfLog
, TimeLineID newTLI
)
4811 char xlogfname
[MAXFNAMELEN
];
4812 XLogSegNo endLogSegNo
;
4813 XLogSegNo startLogSegNo
;
4815 /* we always switch to a new timeline after archive recovery */
4816 Assert(endTLI
!= newTLI
);
4819 * Update min recovery point one last time.
4821 UpdateMinRecoveryPoint(InvalidXLogRecPtr
, true);
4824 * Calculate the last segment on the old timeline, and the first segment
4825 * on the new timeline. If the switch happens in the middle of a segment,
4826 * they are the same, but if the switch happens exactly at a segment
4827 * boundary, startLogSegNo will be endLogSegNo + 1.
4829 XLByteToPrevSeg(endOfLog
, endLogSegNo
, wal_segment_size
);
4830 XLByteToSeg(endOfLog
, startLogSegNo
, wal_segment_size
);
4833 * Initialize the starting WAL segment for the new timeline. If the switch
4834 * happens in the middle of a segment, copy data from the last WAL segment
4835 * of the old timeline up to the switch point, to the starting WAL segment
4836 * on the new timeline.
4838 if (endLogSegNo
== startLogSegNo
)
4841 * Make a copy of the file on the new timeline.
4843 * Writing WAL isn't allowed yet, so there are no locking
4844 * considerations. But we should be just as tense as XLogFileInit to
4845 * avoid emplacing a bogus file.
4847 XLogFileCopy(newTLI
, endLogSegNo
, endTLI
, endLogSegNo
,
4848 XLogSegmentOffset(endOfLog
, wal_segment_size
));
4853 * The switch happened at a segment boundary, so just create the next
4854 * segment on the new timeline.
4858 fd
= XLogFileInit(startLogSegNo
, newTLI
);
4862 int save_errno
= errno
;
4864 XLogFileName(xlogfname
, newTLI
, startLogSegNo
, wal_segment_size
);
4867 (errcode_for_file_access(),
4868 errmsg("could not close file \"%s\": %m", xlogfname
)));
4873 * Let's just make real sure there are not .ready or .done flags posted
4874 * for the new segment.
4876 XLogFileName(xlogfname
, newTLI
, startLogSegNo
, wal_segment_size
);
4877 XLogArchiveCleanup(xlogfname
);
4881 * Perform cleanup actions at the conclusion of archive recovery.
4884 CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI
, XLogRecPtr EndOfLog
,
4888 * Execute the recovery_end_command, if any.
4890 if (recoveryEndCommand
&& strcmp(recoveryEndCommand
, "") != 0)
4892 char lastRestartPointFname
[MAXFNAMELEN
];
4894 GetOldestRestartPointFileName(lastRestartPointFname
);
4895 shell_recovery_end(lastRestartPointFname
);
4899 * We switched to a new timeline. Clean up segments on the old timeline.
4901 * If there are any higher-numbered segments on the old timeline, remove
4902 * them. They might contain valid WAL, but they might also be
4903 * pre-allocated files containing garbage. In any case, they are not part
4904 * of the new timeline's history so we don't need them.
4906 RemoveNonParentXlogFiles(EndOfLog
, newTLI
);
4909 * If the switch happened in the middle of a segment, what to do with the
4910 * last, partial segment on the old timeline? If we don't archive it, and
4911 * the server that created the WAL never archives it either (e.g. because
4912 * it was hit by a meteor), it will never make it to the archive. That's
4913 * OK from our point of view, because the new segment that we created with
4914 * the new TLI contains all the WAL from the old timeline up to the switch
4915 * point. But if you later try to do PITR to the "missing" WAL on the old
4916 * timeline, recovery won't find it in the archive. It's physically
4917 * present in the new file with new TLI, but recovery won't look there
4918 * when it's recovering to the older timeline. On the other hand, if we
4919 * archive the partial segment, and the original server on that timeline
4920 * is still running and archives the completed version of the same segment
4921 * later, it will fail. (We used to do that in 9.4 and below, and it
4922 * caused such problems).
4924 * As a compromise, we rename the last segment with the .partial suffix,
4925 * and archive it. Archive recovery will never try to read .partial
4926 * segments, so they will normally go unused. But in the odd PITR case,
4927 * the administrator can copy them manually to the pg_wal directory
4928 * (removing the suffix). They can be useful in debugging, too.
4930 * If a .done or .ready file already exists for the old timeline, however,
4931 * we had already determined that the segment is complete, so we can let
4932 * it be archived normally. (In particular, if it was restored from the
4933 * archive to begin with, it's expected to have a .done file).
4935 if (XLogSegmentOffset(EndOfLog
, wal_segment_size
) != 0 &&
4936 XLogArchivingActive())
4938 char origfname
[MAXFNAMELEN
];
4939 XLogSegNo endLogSegNo
;
4941 XLByteToPrevSeg(EndOfLog
, endLogSegNo
, wal_segment_size
);
4942 XLogFileName(origfname
, EndOfLogTLI
, endLogSegNo
, wal_segment_size
);
4944 if (!XLogArchiveIsReadyOrDone(origfname
))
4946 char origpath
[MAXPGPATH
];
4947 char partialfname
[MAXFNAMELEN
];
4948 char partialpath
[MAXPGPATH
];
4950 XLogFilePath(origpath
, EndOfLogTLI
, endLogSegNo
, wal_segment_size
);
4951 snprintf(partialfname
, MAXFNAMELEN
, "%s.partial", origfname
);
4952 snprintf(partialpath
, MAXPGPATH
, "%s.partial", origpath
);
4955 * Make sure there's no .done or .ready file for the .partial
4958 XLogArchiveCleanup(partialfname
);
4960 durable_rename(origpath
, partialpath
, ERROR
);
4961 XLogArchiveNotify(partialfname
);
4967 * Check to see if required parameters are set high enough on this server
4968 * for various aspects of recovery operation.
4970 * Note that all the parameters which this function tests need to be
4971 * listed in Administrator's Overview section in high-availability.sgml.
4972 * If you change them, don't forget to update the list.
4975 CheckRequiredParameterValues(void)
4978 * For archive recovery, the WAL must be generated with at least 'replica'
4981 if (ArchiveRecoveryRequested
&& ControlFile
->wal_level
== WAL_LEVEL_MINIMAL
)
4984 (errmsg("WAL was generated with wal_level=minimal, cannot continue recovering"),
4985 errdetail("This happens if you temporarily set wal_level=minimal on the server."),
4986 errhint("Use a backup taken after setting wal_level to higher than minimal.")));
4990 * For Hot Standby, the WAL must be generated with 'replica' mode, and we
4991 * must have at least as many backend slots as the primary.
4993 if (ArchiveRecoveryRequested
&& EnableHotStandby
)
4995 /* We ignore autovacuum_max_workers when we make this test. */
4996 RecoveryRequiresIntParameter("max_connections",
4998 ControlFile
->MaxConnections
);
4999 RecoveryRequiresIntParameter("max_worker_processes",
5000 max_worker_processes
,
5001 ControlFile
->max_worker_processes
);
5002 RecoveryRequiresIntParameter("max_wal_senders",
5004 ControlFile
->max_wal_senders
);
5005 RecoveryRequiresIntParameter("max_prepared_transactions",
5007 ControlFile
->max_prepared_xacts
);
5008 RecoveryRequiresIntParameter("max_locks_per_transaction",
5010 ControlFile
->max_locks_per_xact
);
5015 * This must be called ONCE during postmaster or standalone-backend startup
5020 XLogCtlInsert
*Insert
;
5021 CheckPoint checkPoint
;
5025 bool haveBackupLabel
;
5026 XLogRecPtr EndOfLog
;
5027 TimeLineID EndOfLogTLI
;
5029 bool performedWalRecovery
;
5030 EndOfWalRecoveryInfo
*endOfRecoveryInfo
;
5031 XLogRecPtr abortedRecPtr
;
5032 XLogRecPtr missingContrecPtr
;
5033 TransactionId oldestActiveXID
;
5034 bool promoted
= false;
5037 * We should have an aux process resource owner to use, and we should not
5038 * be in a transaction that's installed some other resowner.
5040 Assert(AuxProcessResourceOwner
!= NULL
);
5041 Assert(CurrentResourceOwner
== NULL
||
5042 CurrentResourceOwner
== AuxProcessResourceOwner
);
5043 CurrentResourceOwner
= AuxProcessResourceOwner
;
5046 * Check that contents look valid.
5048 if (!XRecOffIsValid(ControlFile
->checkPoint
))
5050 (errmsg("control file contains invalid checkpoint location")));
5052 switch (ControlFile
->state
)
5057 * This is the expected case, so don't be chatty in standalone
5060 ereport(IsPostmasterEnvironment
? LOG
: NOTICE
,
5061 (errmsg("database system was shut down at %s",
5062 str_time(ControlFile
->time
))));
5065 case DB_SHUTDOWNED_IN_RECOVERY
:
5067 (errmsg("database system was shut down in recovery at %s",
5068 str_time(ControlFile
->time
))));
5071 case DB_SHUTDOWNING
:
5073 (errmsg("database system shutdown was interrupted; last known up at %s",
5074 str_time(ControlFile
->time
))));
5077 case DB_IN_CRASH_RECOVERY
:
5079 (errmsg("database system was interrupted while in recovery at %s",
5080 str_time(ControlFile
->time
)),
5081 errhint("This probably means that some data is corrupted and"
5082 " you will have to use the last backup for recovery.")));
5085 case DB_IN_ARCHIVE_RECOVERY
:
5087 (errmsg("database system was interrupted while in recovery at log time %s",
5088 str_time(ControlFile
->checkPointCopy
.time
)),
5089 errhint("If this has occurred more than once some data might be corrupted"
5090 " and you might need to choose an earlier recovery target.")));
5093 case DB_IN_PRODUCTION
:
5095 (errmsg("database system was interrupted; last known up at %s",
5096 str_time(ControlFile
->time
))));
5101 (errmsg("control file contains invalid database cluster state")));
5104 /* This is just to allow attaching to startup process with a debugger */
5105 #ifdef XLOG_REPLAY_DELAY
5106 if (ControlFile
->state
!= DB_SHUTDOWNED
)
5107 pg_usleep(60000000L);
5111 * Verify that pg_wal and pg_wal/archive_status exist. In cases where
5112 * someone has performed a copy for PITR, these directories may have been
5113 * excluded and need to be re-created.
5115 ValidateXLOGDirectoryStructure();
5117 /* Set up timeout handler needed to report startup progress. */
5118 if (!IsBootstrapProcessingMode())
5119 RegisterTimeout(STARTUP_PROGRESS_TIMEOUT
,
5120 startup_progress_timeout_handler
);
5123 * If we previously crashed, perform a couple of actions:
5125 * - The pg_wal directory may still include some temporary WAL segments
5126 * used when creating a new segment, so perform some clean up to not
5127 * bloat this path. This is done first as there is no point to sync
5128 * this temporary data.
5130 * - There might be data which we had written, intending to fsync it, but
5131 * which we had not actually fsync'd yet. Therefore, a power failure in
5132 * the near future might cause earlier unflushed writes to be lost, even
5133 * though more recent data written to disk from here on would be
5134 * persisted. To avoid that, fsync the entire data directory.
5136 if (ControlFile
->state
!= DB_SHUTDOWNED
&&
5137 ControlFile
->state
!= DB_SHUTDOWNED_IN_RECOVERY
)
5139 RemoveTempXlogFiles();
5140 SyncDataDirectory();
5147 * Prepare for WAL recovery if needed.
5149 * InitWalRecovery analyzes the control file and the backup label file, if
5150 * any. It updates the in-memory ControlFile buffer according to the
5151 * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5152 * It also applies the tablespace map file, if any.
5154 InitWalRecovery(ControlFile
, &wasShutdown
,
5155 &haveBackupLabel
, &haveTblspcMap
);
5156 checkPoint
= ControlFile
->checkPointCopy
;
5158 /* initialize shared memory variables from the checkpoint record */
5159 ShmemVariableCache
->nextXid
= checkPoint
.nextXid
;
5160 ShmemVariableCache
->nextOid
= checkPoint
.nextOid
;
5161 ShmemVariableCache
->oidCount
= 0;
5162 MultiXactSetNextMXact(checkPoint
.nextMulti
, checkPoint
.nextMultiOffset
);
5163 AdvanceOldestClogXid(checkPoint
.oldestXid
);
5164 SetTransactionIdLimit(checkPoint
.oldestXid
, checkPoint
.oldestXidDB
);
5165 SetMultiXactIdLimit(checkPoint
.oldestMulti
, checkPoint
.oldestMultiDB
, true);
5166 SetCommitTsLimit(checkPoint
.oldestCommitTsXid
,
5167 checkPoint
.newestCommitTsXid
);
5168 XLogCtl
->ckptFullXid
= checkPoint
.nextXid
;
5171 * Clear out any old relcache cache files. This is *necessary* if we do
5172 * any WAL replay, since that would probably result in the cache files
5173 * being out of sync with database reality. In theory we could leave them
5174 * in place if the database had been cleanly shut down, but it seems
5175 * safest to just remove them always and let them be rebuilt during the
5176 * first backend startup. These files needs to be removed from all
5177 * directories including pg_tblspc, however the symlinks are created only
5178 * after reading tablespace_map file in case of archive recovery from
5179 * backup, so needs to clear old relcache files here after creating
5182 RelationCacheInitFileRemove();
5185 * Initialize replication slots, before there's a chance to remove
5186 * required resources.
5188 StartupReplicationSlots();
5191 * Startup logical state, needs to be setup now so we have proper data
5192 * during crash recovery.
5194 StartupReorderBuffer();
5197 * Startup CLOG. This must be done after ShmemVariableCache->nextXid has
5198 * been initialized and before we accept connections or begin WAL replay.
5203 * Startup MultiXact. We need to do this early to be able to replay
5209 * Ditto for commit timestamps. Activate the facility if the setting is
5210 * enabled in the control file, as there should be no tracking of commit
5211 * timestamps done when the setting was disabled. This facility can be
5212 * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
5214 if (ControlFile
->track_commit_timestamp
)
5218 * Recover knowledge about replay progress of known replication partners.
5220 StartupReplicationOrigin();
5223 * Initialize unlogged LSN. On a clean shutdown, it's restored from the
5224 * control file. On recovery, all unlogged relations are blown away, so
5225 * the unlogged LSN counter can be reset too.
5227 if (ControlFile
->state
== DB_SHUTDOWNED
)
5228 XLogCtl
->unloggedLSN
= ControlFile
->unloggedLSN
;
5230 XLogCtl
->unloggedLSN
= FirstNormalUnloggedLSN
;
5233 * Copy any missing timeline history files between 'now' and the recovery
5234 * target timeline from archive to pg_wal. While we don't need those files
5235 * ourselves - the history file of the recovery target timeline covers all
5236 * the previous timelines in the history too - a cascading standby server
5237 * might be interested in them. Or, if you archive the WAL from this
5238 * server to a different archive than the primary, it'd be good for all
5239 * the history files to get archived there after failover, so that you can
5240 * use one of the old timelines as a PITR target. Timeline history files
5241 * are small, so it's better to copy them unnecessarily than not copy them
5244 restoreTimeLineHistoryFiles(checkPoint
.ThisTimeLineID
, recoveryTargetTLI
);
5247 * Before running in recovery, scan pg_twophase and fill in its status to
5248 * be able to work on entries generated by redo. Doing a scan before
5249 * taking any recovery action has the merit to discard any 2PC files that
5250 * are newer than the first record to replay, saving from any conflicts at
5251 * replay. This avoids as well any subsequent scans when doing recovery
5252 * of the on-disk two-phase data.
5254 restoreTwoPhaseData();
5257 * When starting with crash recovery, reset pgstat data - it might not be
5258 * valid. Otherwise restore pgstat data. It's safe to do this here,
5259 * because postmaster will not yet have started any other processes.
5261 * NB: Restoring replication slot stats relies on slot state to have
5262 * already been restored from disk.
5264 * TODO: With a bit of extra work we could just start with a pgstat file
5265 * associated with the checkpoint redo location we're starting from.
5268 pgstat_discard_stats();
5270 pgstat_restore_stats();
5272 lastFullPageWrites
= checkPoint
.fullPageWrites
;
5274 RedoRecPtr
= XLogCtl
->RedoRecPtr
= XLogCtl
->Insert
.RedoRecPtr
= checkPoint
.redo
;
5275 doPageWrites
= lastFullPageWrites
;
5280 /* Initialize state for RecoveryInProgress() */
5281 SpinLockAcquire(&XLogCtl
->info_lck
);
5282 if (InArchiveRecovery
)
5283 XLogCtl
->SharedRecoveryState
= RECOVERY_STATE_ARCHIVE
;
5285 XLogCtl
->SharedRecoveryState
= RECOVERY_STATE_CRASH
;
5286 SpinLockRelease(&XLogCtl
->info_lck
);
5289 * Update pg_control to show that we are recovering and to show the
5290 * selected checkpoint as the place we are starting from. We also mark
5291 * pg_control with any minimum recovery stop point obtained from a
5292 * backup history file.
5294 * No need to hold ControlFileLock yet, we aren't up far enough.
5296 UpdateControlFile();
5299 * If there was a backup label file, it's done its job and the info
5300 * has now been propagated into pg_control. We must get rid of the
5301 * label file so that if we crash during recovery, we'll pick up at
5302 * the latest recovery restartpoint instead of going all the way back
5303 * to the backup start point. It seems prudent though to just rename
5304 * the file out of the way rather than delete it completely.
5306 if (haveBackupLabel
)
5308 unlink(BACKUP_LABEL_OLD
);
5309 durable_rename(BACKUP_LABEL_FILE
, BACKUP_LABEL_OLD
, FATAL
);
5313 * If there was a tablespace_map file, it's done its job and the
5314 * symlinks have been created. We must get rid of the map file so
5315 * that if we crash during recovery, we don't create symlinks again.
5316 * It seems prudent though to just rename the file out of the way
5317 * rather than delete it completely.
5321 unlink(TABLESPACE_MAP_OLD
);
5322 durable_rename(TABLESPACE_MAP
, TABLESPACE_MAP_OLD
, FATAL
);
5326 * Initialize our local copy of minRecoveryPoint. When doing crash
5327 * recovery we want to replay up to the end of WAL. Particularly, in
5328 * the case of a promoted standby minRecoveryPoint value in the
5329 * control file is only updated after the first checkpoint. However,
5330 * if the instance crashes before the first post-recovery checkpoint
5331 * is completed then recovery will use a stale location causing the
5332 * startup process to think that there are still invalid page
5333 * references when checking for data consistency.
5335 if (InArchiveRecovery
)
5337 LocalMinRecoveryPoint
= ControlFile
->minRecoveryPoint
;
5338 LocalMinRecoveryPointTLI
= ControlFile
->minRecoveryPointTLI
;
5342 LocalMinRecoveryPoint
= InvalidXLogRecPtr
;
5343 LocalMinRecoveryPointTLI
= 0;
5346 /* Check that the GUCs used to generate the WAL allow recovery */
5347 CheckRequiredParameterValues();
5350 * We're in recovery, so unlogged relations may be trashed and must be
5351 * reset. This should be done BEFORE allowing Hot Standby
5352 * connections, so that read-only backends don't try to read whatever
5353 * garbage is left over from before.
5355 ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP
);
5358 * Likewise, delete any saved transaction snapshot files that got left
5359 * behind by crashed backends.
5361 DeleteAllExportedSnapshotFiles();
5364 * Initialize for Hot Standby, if enabled. We won't let backends in
5365 * yet, not until we've reached the min recovery point specified in
5366 * control file and we've established a recovery snapshot from a
5367 * running-xacts WAL record.
5369 if (ArchiveRecoveryRequested
&& EnableHotStandby
)
5371 TransactionId
*xids
;
5375 (errmsg_internal("initializing for hot standby")));
5377 InitRecoveryTransactionEnvironment();
5380 oldestActiveXID
= PrescanPreparedTransactions(&xids
, &nxids
);
5382 oldestActiveXID
= checkPoint
.oldestActiveXid
;
5383 Assert(TransactionIdIsValid(oldestActiveXID
));
5385 /* Tell procarray about the range of xids it has to deal with */
5386 ProcArrayInitRecovery(XidFromFullTransactionId(ShmemVariableCache
->nextXid
));
5389 * Startup subtrans only. CLOG, MultiXact and commit timestamp
5390 * have already been started up and other SLRUs are not maintained
5391 * during recovery and need not be started yet.
5393 StartupSUBTRANS(oldestActiveXID
);
5396 * If we're beginning at a shutdown checkpoint, we know that
5397 * nothing was running on the primary at this point. So fake-up an
5398 * empty running-xacts record and use that here and now. Recover
5399 * additional standby state for prepared transactions.
5403 RunningTransactionsData running
;
5404 TransactionId latestCompletedXid
;
5407 * Construct a RunningTransactions snapshot representing a
5408 * shut down server, with only prepared transactions still
5409 * alive. We're never overflowed at this point because all
5410 * subxids are listed with their parent prepared transactions.
5412 running
.xcnt
= nxids
;
5413 running
.subxcnt
= 0;
5414 running
.subxid_overflow
= false;
5415 running
.nextXid
= XidFromFullTransactionId(checkPoint
.nextXid
);
5416 running
.oldestRunningXid
= oldestActiveXID
;
5417 latestCompletedXid
= XidFromFullTransactionId(checkPoint
.nextXid
);
5418 TransactionIdRetreat(latestCompletedXid
);
5419 Assert(TransactionIdIsNormal(latestCompletedXid
));
5420 running
.latestCompletedXid
= latestCompletedXid
;
5421 running
.xids
= xids
;
5423 ProcArrayApplyRecoveryInfo(&running
);
5425 StandbyRecoverPreparedTransactions();
5430 * We're all set for replaying the WAL now. Do it.
5432 PerformWalRecovery();
5433 performedWalRecovery
= true;
5436 performedWalRecovery
= false;
5439 * Finish WAL recovery.
5441 endOfRecoveryInfo
= FinishWalRecovery();
5442 EndOfLog
= endOfRecoveryInfo
->endOfLog
;
5443 EndOfLogTLI
= endOfRecoveryInfo
->endOfLogTLI
;
5444 abortedRecPtr
= endOfRecoveryInfo
->abortedRecPtr
;
5445 missingContrecPtr
= endOfRecoveryInfo
->missingContrecPtr
;
5448 * Reset ps status display, so as no information related to recovery
5454 * When recovering from a backup (we are in recovery, and archive recovery
5455 * was requested), complain if we did not roll forward far enough to reach
5456 * the point where the database is consistent. For regular online
5457 * backup-from-primary, that means reaching the end-of-backup WAL record
5458 * (at which point we reset backupStartPoint to be Invalid), for
5459 * backup-from-replica (which can't inject records into the WAL stream),
5460 * that point is when we reach the minRecoveryPoint in pg_control (which
5461 * we purposefully copy last when backing up from a replica). For
5462 * pg_rewind (which creates a backup_label with a method of "pg_rewind")
5463 * or snapshot-style backups (which don't), backupEndRequired will be set
5466 * Note: it is indeed okay to look at the local variable
5467 * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
5468 * might be further ahead --- ControlFile->minRecoveryPoint cannot have
5469 * been advanced beyond the WAL we processed.
5472 (EndOfLog
< LocalMinRecoveryPoint
||
5473 !XLogRecPtrIsInvalid(ControlFile
->backupStartPoint
)))
5476 * Ran off end of WAL before reaching end-of-backup WAL record, or
5477 * minRecoveryPoint. That's a bad sign, indicating that you tried to
5478 * recover from an online backup but never called pg_backup_stop(), or
5479 * you didn't archive all the WAL needed.
5481 if (ArchiveRecoveryRequested
|| ControlFile
->backupEndRequired
)
5483 if (!XLogRecPtrIsInvalid(ControlFile
->backupStartPoint
) || ControlFile
->backupEndRequired
)
5485 (errmsg("WAL ends before end of online backup"),
5486 errhint("All WAL generated while online backup was taken must be available at recovery.")));
5489 (errmsg("WAL ends before consistent recovery point")));
5494 * Reset unlogged relations to the contents of their INIT fork. This is
5495 * done AFTER recovery is complete so as to include any unlogged relations
5496 * created during recovery, but BEFORE recovery is marked as having
5497 * completed successfully. Otherwise we'd not retry if any of the post
5498 * end-of-recovery steps fail.
5501 ResetUnloggedRelations(UNLOGGED_RELATION_INIT
);
5504 * Pre-scan prepared transactions to find out the range of XIDs present.
5505 * This information is not quite needed yet, but it is positioned here so
5506 * as potential problems are detected before any on-disk change is done.
5508 oldestActiveXID
= PrescanPreparedTransactions(NULL
, NULL
);
5511 * Allow ordinary WAL segment creation before possibly switching to a new
5512 * timeline, which creates a new segment, and after the last ReadRecord().
5514 SetInstallXLogFileSegmentActive();
5517 * Consider whether we need to assign a new timeline ID.
5519 * If we did archive recovery, we always assign a new ID. This handles a
5520 * couple of issues. If we stopped short of the end of WAL during
5521 * recovery, then we are clearly generating a new timeline and must assign
5522 * it a unique new ID. Even if we ran to the end, modifying the current
5523 * last segment is problematic because it may result in trying to
5524 * overwrite an already-archived copy of that segment, and we encourage
5525 * DBAs to make their archive_commands reject that. We can dodge the
5526 * problem by making the new active segment have a new timeline ID.
5528 * In a normal crash recovery, we can just extend the timeline we were in.
5530 newTLI
= endOfRecoveryInfo
->lastRecTLI
;
5531 if (ArchiveRecoveryRequested
)
5533 newTLI
= findNewestTimeLine(recoveryTargetTLI
) + 1;
5535 (errmsg("selected new timeline ID: %u", newTLI
)));
5538 * Make a writable copy of the last WAL segment. (Note that we also
5539 * have a copy of the last block of the old WAL in
5540 * endOfRecovery->lastPage; we will use that below.)
5542 XLogInitNewTimeline(EndOfLogTLI
, EndOfLog
, newTLI
);
5545 * Remove the signal files out of the way, so that we don't
5546 * accidentally re-enter archive recovery mode in a subsequent crash.
5548 if (endOfRecoveryInfo
->standby_signal_file_found
)
5549 durable_unlink(STANDBY_SIGNAL_FILE
, FATAL
);
5551 if (endOfRecoveryInfo
->recovery_signal_file_found
)
5552 durable_unlink(RECOVERY_SIGNAL_FILE
, FATAL
);
5555 * Write the timeline history file, and have it archived. After this
5556 * point (or rather, as soon as the file is archived), the timeline
5557 * will appear as "taken" in the WAL archive and to any standby
5558 * servers. If we crash before actually switching to the new
5559 * timeline, standby servers will nevertheless think that we switched
5560 * to the new timeline, and will try to connect to the new timeline.
5561 * To minimize the window for that, try to do as little as possible
5562 * between here and writing the end-of-recovery record.
5564 writeTimeLineHistory(newTLI
, recoveryTargetTLI
,
5565 EndOfLog
, endOfRecoveryInfo
->recoveryStopReason
);
5568 (errmsg("archive recovery complete")));
5571 /* Save the selected TimeLineID in shared memory, too */
5572 XLogCtl
->InsertTimeLineID
= newTLI
;
5573 XLogCtl
->PrevTimeLineID
= endOfRecoveryInfo
->lastRecTLI
;
5576 * Actually, if WAL ended in an incomplete record, skip the parts that
5577 * made it through and start writing after the portion that persisted.
5578 * (It's critical to first write an OVERWRITE_CONTRECORD message, which
5579 * we'll do as soon as we're open for writing new WAL.)
5581 if (!XLogRecPtrIsInvalid(missingContrecPtr
))
5584 * We should only have a missingContrecPtr if we're not switching to
5585 * a new timeline. When a timeline switch occurs, WAL is copied from
5586 * the old timeline to the new only up to the end of the last complete
5587 * record, so there can't be an incomplete WAL record that we need to
5590 Assert(newTLI
== endOfRecoveryInfo
->lastRecTLI
);
5591 Assert(!XLogRecPtrIsInvalid(abortedRecPtr
));
5592 EndOfLog
= missingContrecPtr
;
5596 * Prepare to write WAL starting at EndOfLog location, and init xlog
5597 * buffer cache using the block containing the last record from the
5598 * previous incarnation.
5600 Insert
= &XLogCtl
->Insert
;
5601 Insert
->PrevBytePos
= XLogRecPtrToBytePos(endOfRecoveryInfo
->lastRec
);
5602 Insert
->CurrBytePos
= XLogRecPtrToBytePos(EndOfLog
);
5605 * Tricky point here: lastPage contains the *last* block that the LastRec
5606 * record spans, not the one it starts in. The last block is indeed the
5607 * one we want to use.
5609 if (EndOfLog
% XLOG_BLCKSZ
!= 0)
5615 firstIdx
= XLogRecPtrToBufIdx(EndOfLog
);
5616 len
= EndOfLog
- endOfRecoveryInfo
->lastPageBeginPtr
;
5617 Assert(len
< XLOG_BLCKSZ
);
5619 /* Copy the valid part of the last block, and zero the rest */
5620 page
= &XLogCtl
->pages
[firstIdx
* XLOG_BLCKSZ
];
5621 memcpy(page
, endOfRecoveryInfo
->lastPage
, len
);
5622 memset(page
+ len
, 0, XLOG_BLCKSZ
- len
);
5624 XLogCtl
->xlblocks
[firstIdx
] = endOfRecoveryInfo
->lastPageBeginPtr
+ XLOG_BLCKSZ
;
5625 XLogCtl
->InitializedUpTo
= endOfRecoveryInfo
->lastPageBeginPtr
+ XLOG_BLCKSZ
;
5630 * There is no partial block to copy. Just set InitializedUpTo, and
5631 * let the first attempt to insert a log record to initialize the next
5634 XLogCtl
->InitializedUpTo
= EndOfLog
;
5637 LogwrtResult
.Write
= LogwrtResult
.Flush
= EndOfLog
;
5639 XLogCtl
->LogwrtResult
= LogwrtResult
;
5641 XLogCtl
->LogwrtRqst
.Write
= EndOfLog
;
5642 XLogCtl
->LogwrtRqst
.Flush
= EndOfLog
;
5645 * Preallocate additional log files, if wanted.
5647 PreallocXlogFiles(EndOfLog
, newTLI
);
5650 * Okay, we're officially UP.
5654 /* start the archive_timeout timer and LSN running */
5655 XLogCtl
->lastSegSwitchTime
= (pg_time_t
) time(NULL
);
5656 XLogCtl
->lastSegSwitchLSN
= EndOfLog
;
5658 /* also initialize latestCompletedXid, to nextXid - 1 */
5659 LWLockAcquire(ProcArrayLock
, LW_EXCLUSIVE
);
5660 ShmemVariableCache
->latestCompletedXid
= ShmemVariableCache
->nextXid
;
5661 FullTransactionIdRetreat(&ShmemVariableCache
->latestCompletedXid
);
5662 LWLockRelease(ProcArrayLock
);
5665 * Start up subtrans, if not already done for hot standby. (commit
5666 * timestamps are started below, if necessary.)
5668 if (standbyState
== STANDBY_DISABLED
)
5669 StartupSUBTRANS(oldestActiveXID
);
5672 * Perform end of recovery actions for any SLRUs that need it.
5677 /* Reload shared-memory state for prepared transactions */
5678 RecoverPreparedTransactions();
5680 /* Shut down xlogreader */
5681 ShutdownWalRecovery();
5683 /* Enable WAL writes for this backend only. */
5684 LocalSetXLogInsertAllowed();
5686 /* If necessary, write overwrite-contrecord before doing anything else */
5687 if (!XLogRecPtrIsInvalid(abortedRecPtr
))
5689 Assert(!XLogRecPtrIsInvalid(missingContrecPtr
));
5690 CreateOverwriteContrecordRecord(abortedRecPtr
, missingContrecPtr
, newTLI
);
5694 * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
5695 * record before resource manager writes cleanup WAL records or checkpoint
5696 * record is written.
5698 Insert
->fullPageWrites
= lastFullPageWrites
;
5699 UpdateFullPageWrites();
5702 * Emit checkpoint or end-of-recovery record in XLOG, if required.
5704 if (performedWalRecovery
)
5705 promoted
= PerformRecoveryXLogAction();
5708 * If any of the critical GUCs have changed, log them before we allow
5709 * backends to write WAL.
5711 XLogReportParameters();
5713 /* If this is archive recovery, perform post-recovery cleanup actions. */
5714 if (ArchiveRecoveryRequested
)
5715 CleanupAfterArchiveRecovery(EndOfLogTLI
, EndOfLog
, newTLI
);
5718 * Local WAL inserts enabled, so it's time to finish initialization of
5721 CompleteCommitTsInitialization();
5724 * All done with end-of-recovery actions.
5726 * Now allow backends to write WAL and update the control file status in
5727 * consequence. SharedRecoveryState, that controls if backends can write
5728 * WAL, is updated while holding ControlFileLock to prevent other backends
5729 * to look at an inconsistent state of the control file in shared memory.
5730 * There is still a small window during which backends can write WAL and
5731 * the control file is still referring to a system not in DB_IN_PRODUCTION
5732 * state while looking at the on-disk control file.
5734 * Also, we use info_lck to update SharedRecoveryState to ensure that
5735 * there are no race conditions concerning visibility of other recent
5736 * updates to shared memory.
5738 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
5739 ControlFile
->state
= DB_IN_PRODUCTION
;
5741 SpinLockAcquire(&XLogCtl
->info_lck
);
5742 XLogCtl
->SharedRecoveryState
= RECOVERY_STATE_DONE
;
5743 SpinLockRelease(&XLogCtl
->info_lck
);
5745 UpdateControlFile();
5746 LWLockRelease(ControlFileLock
);
5749 * Shutdown the recovery environment. This must occur after
5750 * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
5751 * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
5752 * any session building a snapshot will not rely on KnownAssignedXids as
5753 * RecoveryInProgress() would return false at this stage. This is
5754 * particularly critical for prepared 2PC transactions, that would still
5755 * need to be included in snapshots once recovery has ended.
5757 if (standbyState
!= STANDBY_DISABLED
)
5758 ShutdownRecoveryTransactionEnvironment();
5761 * If there were cascading standby servers connected to us, nudge any wal
5762 * sender processes to notice that we've been promoted.
5767 * If this was a promotion, request an (online) checkpoint now. This isn't
5768 * required for consistency, but the last restartpoint might be far back,
5769 * and in case of a crash, recovering from it might take a longer than is
5770 * appropriate now that we're not in standby mode anymore.
5773 RequestCheckpoint(CHECKPOINT_FORCE
);
5777 * Callback from PerformWalRecovery(), called when we switch from crash
5778 * recovery to archive recovery mode. Updates the control file accordingly.
5781 SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr
, TimeLineID replayTLI
)
5783 /* initialize minRecoveryPoint to this record */
5784 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
5785 ControlFile
->state
= DB_IN_ARCHIVE_RECOVERY
;
5786 if (ControlFile
->minRecoveryPoint
< EndRecPtr
)
5788 ControlFile
->minRecoveryPoint
= EndRecPtr
;
5789 ControlFile
->minRecoveryPointTLI
= replayTLI
;
5791 /* update local copy */
5792 LocalMinRecoveryPoint
= ControlFile
->minRecoveryPoint
;
5793 LocalMinRecoveryPointTLI
= ControlFile
->minRecoveryPointTLI
;
5796 * The startup process can update its local copy of minRecoveryPoint from
5799 updateMinRecoveryPoint
= true;
5801 UpdateControlFile();
5804 * We update SharedRecoveryState while holding the lock on ControlFileLock
5805 * so both states are consistent in shared memory.
5807 SpinLockAcquire(&XLogCtl
->info_lck
);
5808 XLogCtl
->SharedRecoveryState
= RECOVERY_STATE_ARCHIVE
;
5809 SpinLockRelease(&XLogCtl
->info_lck
);
5811 LWLockRelease(ControlFileLock
);
5815 * Callback from PerformWalRecovery(), called when we reach the end of backup.
5816 * Updates the control file accordingly.
5819 ReachedEndOfBackup(XLogRecPtr EndRecPtr
, TimeLineID tli
)
5822 * We have reached the end of base backup, as indicated by pg_control. The
5823 * data on disk is now consistent (unless minRecovery point is further
5824 * ahead, which can happen if we crashed during previous recovery). Reset
5825 * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
5826 * make sure we don't allow starting up at an earlier point even if
5827 * recovery is stopped and restarted soon after this.
5829 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
5831 if (ControlFile
->minRecoveryPoint
< EndRecPtr
)
5833 ControlFile
->minRecoveryPoint
= EndRecPtr
;
5834 ControlFile
->minRecoveryPointTLI
= tli
;
5837 ControlFile
->backupStartPoint
= InvalidXLogRecPtr
;
5838 ControlFile
->backupEndPoint
= InvalidXLogRecPtr
;
5839 ControlFile
->backupEndRequired
= false;
5840 UpdateControlFile();
5842 LWLockRelease(ControlFileLock
);
5846 * Perform whatever XLOG actions are necessary at end of REDO.
5848 * The goal here is to make sure that we'll be able to recover properly if
5849 * we crash again. If we choose to write a checkpoint, we'll write a shutdown
5850 * checkpoint rather than an on-line one. This is not particularly critical,
5851 * but since we may be assigning a new TLI, using a shutdown checkpoint allows
5852 * us to have the rule that TLI only changes in shutdown checkpoints, which
5853 * allows some extra error checking in xlog_redo.
5856 PerformRecoveryXLogAction(void)
5858 bool promoted
= false;
5861 * Perform a checkpoint to update all our recovery activity to disk.
5863 * Note that we write a shutdown checkpoint rather than an on-line one.
5864 * This is not particularly critical, but since we may be assigning a new
5865 * TLI, using a shutdown checkpoint allows us to have the rule that TLI
5866 * only changes in shutdown checkpoints, which allows some extra error
5867 * checking in xlog_redo.
5869 * In promotion, only create a lightweight end-of-recovery record instead
5870 * of a full checkpoint. A checkpoint is requested later, after we're
5871 * fully out of recovery mode and already accepting queries.
5873 if (ArchiveRecoveryRequested
&& IsUnderPostmaster
&&
5874 PromoteIsTriggered())
5879 * Insert a special WAL record to mark the end of recovery, since we
5880 * aren't doing a checkpoint. That means that the checkpointer process
5881 * may likely be in the middle of a time-smoothed restartpoint and
5882 * could continue to be for minutes after this. That sounds strange,
5883 * but the effect is roughly the same and it would be stranger to try
5884 * to come out of the restartpoint and then checkpoint. We request a
5885 * checkpoint later anyway, just for safety.
5887 CreateEndOfRecoveryRecord();
5891 RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY
|
5892 CHECKPOINT_IMMEDIATE
|
5900 * Is the system still in recovery?
5902 * Unlike testing InRecovery, this works in any process that's connected to
5906 RecoveryInProgress(void)
5909 * We check shared state each time only until we leave recovery mode. We
5910 * can't re-enter recovery, so there's no need to keep checking after the
5911 * shared variable has once been seen false.
5913 if (!LocalRecoveryInProgress
)
5918 * use volatile pointer to make sure we make a fresh read of the
5921 volatile XLogCtlData
*xlogctl
= XLogCtl
;
5923 LocalRecoveryInProgress
= (xlogctl
->SharedRecoveryState
!= RECOVERY_STATE_DONE
);
5926 * Note: We don't need a memory barrier when we're still in recovery.
5927 * We might exit recovery immediately after return, so the caller
5928 * can't rely on 'true' meaning that we're still in recovery anyway.
5931 return LocalRecoveryInProgress
;
5936 * Returns current recovery state from shared memory.
5938 * This returned state is kept consistent with the contents of the control
5939 * file. See details about the possible values of RecoveryState in xlog.h.
5942 GetRecoveryState(void)
5944 RecoveryState retval
;
5946 SpinLockAcquire(&XLogCtl
->info_lck
);
5947 retval
= XLogCtl
->SharedRecoveryState
;
5948 SpinLockRelease(&XLogCtl
->info_lck
);
5954 * Is this process allowed to insert new WAL records?
5956 * Ordinarily this is essentially equivalent to !RecoveryInProgress().
5957 * But we also have provisions for forcing the result "true" or "false"
5958 * within specific processes regardless of the global state.
5961 XLogInsertAllowed(void)
5964 * If value is "unconditionally true" or "unconditionally false", just
5965 * return it. This provides the normal fast path once recovery is known
5968 if (LocalXLogInsertAllowed
>= 0)
5969 return (bool) LocalXLogInsertAllowed
;
5972 * Else, must check to see if we're still in recovery.
5974 if (RecoveryInProgress())
5978 * On exit from recovery, reset to "unconditionally true", since there is
5979 * no need to keep checking.
5981 LocalXLogInsertAllowed
= 1;
5986 * Make XLogInsertAllowed() return true in the current process only.
5988 * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
5989 * and even call LocalSetXLogInsertAllowed() again after that.
5991 * Returns the previous value of LocalXLogInsertAllowed.
5994 LocalSetXLogInsertAllowed(void)
5996 int oldXLogAllowed
= LocalXLogInsertAllowed
;
5998 LocalXLogInsertAllowed
= 1;
6000 return oldXLogAllowed
;
6004 * Return the current Redo pointer from shared memory.
6006 * As a side-effect, the local RedoRecPtr copy is updated.
6014 * The possibly not up-to-date copy in XlogCtl is enough. Even if we
6015 * grabbed a WAL insertion lock to read the authoritative value in
6016 * Insert->RedoRecPtr, someone might update it just after we've released
6019 SpinLockAcquire(&XLogCtl
->info_lck
);
6020 ptr
= XLogCtl
->RedoRecPtr
;
6021 SpinLockRelease(&XLogCtl
->info_lck
);
6023 if (RedoRecPtr
< ptr
)
6030 * Return information needed to decide whether a modified block needs a
6031 * full-page image to be included in the WAL record.
6033 * The returned values are cached copies from backend-private memory, and
6034 * possibly out-of-date or, indeed, uninitialized, in which case they will
6035 * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6036 * re-check them against up-to-date values, while holding the WAL insert lock.
6039 GetFullPageWriteInfo(XLogRecPtr
*RedoRecPtr_p
, bool *doPageWrites_p
)
6041 *RedoRecPtr_p
= RedoRecPtr
;
6042 *doPageWrites_p
= doPageWrites
;
6046 * GetInsertRecPtr -- Returns the current insert position.
6048 * NOTE: The value *actually* returned is the position of the last full
6049 * xlog page. It lags behind the real insert position by at most 1 page.
6050 * For that, we don't need to scan through WAL insertion locks, and an
6051 * approximation is enough for the current usage of this function.
6054 GetInsertRecPtr(void)
6058 SpinLockAcquire(&XLogCtl
->info_lck
);
6059 recptr
= XLogCtl
->LogwrtRqst
.Write
;
6060 SpinLockRelease(&XLogCtl
->info_lck
);
6066 * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6067 * position known to be fsync'd to disk. This should only be used on a
6068 * system that is known not to be in recovery.
6071 GetFlushRecPtr(TimeLineID
*insertTLI
)
6073 Assert(XLogCtl
->SharedRecoveryState
== RECOVERY_STATE_DONE
);
6075 SpinLockAcquire(&XLogCtl
->info_lck
);
6076 LogwrtResult
= XLogCtl
->LogwrtResult
;
6077 SpinLockRelease(&XLogCtl
->info_lck
);
6080 * If we're writing and flushing WAL, the time line can't be changing, so
6081 * no lock is required.
6084 *insertTLI
= XLogCtl
->InsertTimeLineID
;
6086 return LogwrtResult
.Flush
;
6090 * GetWALInsertionTimeLine -- Returns the current timeline of a system that
6091 * is not in recovery.
6094 GetWALInsertionTimeLine(void)
6096 Assert(XLogCtl
->SharedRecoveryState
== RECOVERY_STATE_DONE
);
6098 /* Since the value can't be changing, no lock is required. */
6099 return XLogCtl
->InsertTimeLineID
;
6103 * GetLastImportantRecPtr -- Returns the LSN of the last important record
6104 * inserted. All records not explicitly marked as unimportant are considered
6107 * The LSN is determined by computing the maximum of
6108 * WALInsertLocks[i].lastImportantAt.
6111 GetLastImportantRecPtr(void)
6113 XLogRecPtr res
= InvalidXLogRecPtr
;
6116 for (i
= 0; i
< NUM_XLOGINSERT_LOCKS
; i
++)
6118 XLogRecPtr last_important
;
6121 * Need to take a lock to prevent torn reads of the LSN, which are
6122 * possible on some of the supported platforms. WAL insert locks only
6123 * support exclusive mode, so we have to use that.
6125 LWLockAcquire(&WALInsertLocks
[i
].l
.lock
, LW_EXCLUSIVE
);
6126 last_important
= WALInsertLocks
[i
].l
.lastImportantAt
;
6127 LWLockRelease(&WALInsertLocks
[i
].l
.lock
);
6129 if (res
< last_important
)
6130 res
= last_important
;
6137 * Get the time and LSN of the last xlog segment switch
6140 GetLastSegSwitchData(XLogRecPtr
*lastSwitchLSN
)
6144 /* Need WALWriteLock, but shared lock is sufficient */
6145 LWLockAcquire(WALWriteLock
, LW_SHARED
);
6146 result
= XLogCtl
->lastSegSwitchTime
;
6147 *lastSwitchLSN
= XLogCtl
->lastSegSwitchLSN
;
6148 LWLockRelease(WALWriteLock
);
6154 * This must be called ONCE during postmaster or standalone-backend shutdown
6157 ShutdownXLOG(int code
, Datum arg
)
6160 * We should have an aux process resource owner to use, and we should not
6161 * be in a transaction that's installed some other resowner.
6163 Assert(AuxProcessResourceOwner
!= NULL
);
6164 Assert(CurrentResourceOwner
== NULL
||
6165 CurrentResourceOwner
== AuxProcessResourceOwner
);
6166 CurrentResourceOwner
= AuxProcessResourceOwner
;
6168 /* Don't be chatty in standalone mode */
6169 ereport(IsPostmasterEnvironment
? LOG
: NOTICE
,
6170 (errmsg("shutting down")));
6173 * Signal walsenders to move to stopping state.
6175 WalSndInitStopping();
6178 * Wait for WAL senders to be in stopping state. This prevents commands
6179 * from writing new WAL.
6181 WalSndWaitStopping();
6183 if (RecoveryInProgress())
6184 CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN
| CHECKPOINT_IMMEDIATE
);
6188 * If archiving is enabled, rotate the last XLOG file so that all the
6189 * remaining records are archived (postmaster wakes up the archiver
6190 * process one more time at the end of shutdown). The checkpoint
6191 * record will go to the next XLOG file and won't be archived (yet).
6193 if (XLogArchivingActive())
6194 RequestXLogSwitch(false);
6196 CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN
| CHECKPOINT_IMMEDIATE
);
6201 * Log start of a checkpoint.
6204 LogCheckpointStart(int flags
, bool restartpoint
)
6208 /* translator: the placeholders show checkpoint options */
6209 (errmsg("restartpoint starting:%s%s%s%s%s%s%s%s",
6210 (flags
& CHECKPOINT_IS_SHUTDOWN
) ? " shutdown" : "",
6211 (flags
& CHECKPOINT_END_OF_RECOVERY
) ? " end-of-recovery" : "",
6212 (flags
& CHECKPOINT_IMMEDIATE
) ? " immediate" : "",
6213 (flags
& CHECKPOINT_FORCE
) ? " force" : "",
6214 (flags
& CHECKPOINT_WAIT
) ? " wait" : "",
6215 (flags
& CHECKPOINT_CAUSE_XLOG
) ? " wal" : "",
6216 (flags
& CHECKPOINT_CAUSE_TIME
) ? " time" : "",
6217 (flags
& CHECKPOINT_FLUSH_ALL
) ? " flush-all" : "")));
6220 /* translator: the placeholders show checkpoint options */
6221 (errmsg("checkpoint starting:%s%s%s%s%s%s%s%s",
6222 (flags
& CHECKPOINT_IS_SHUTDOWN
) ? " shutdown" : "",
6223 (flags
& CHECKPOINT_END_OF_RECOVERY
) ? " end-of-recovery" : "",
6224 (flags
& CHECKPOINT_IMMEDIATE
) ? " immediate" : "",
6225 (flags
& CHECKPOINT_FORCE
) ? " force" : "",
6226 (flags
& CHECKPOINT_WAIT
) ? " wait" : "",
6227 (flags
& CHECKPOINT_CAUSE_XLOG
) ? " wal" : "",
6228 (flags
& CHECKPOINT_CAUSE_TIME
) ? " time" : "",
6229 (flags
& CHECKPOINT_FLUSH_ALL
) ? " flush-all" : "")));
6233 * Log end of a checkpoint.
6236 LogCheckpointEnd(bool restartpoint
)
6243 uint64 average_sync_time
;
6245 CheckpointStats
.ckpt_end_t
= GetCurrentTimestamp();
6247 write_msecs
= TimestampDifferenceMilliseconds(CheckpointStats
.ckpt_write_t
,
6248 CheckpointStats
.ckpt_sync_t
);
6250 sync_msecs
= TimestampDifferenceMilliseconds(CheckpointStats
.ckpt_sync_t
,
6251 CheckpointStats
.ckpt_sync_end_t
);
6253 /* Accumulate checkpoint timing summary data, in milliseconds. */
6254 PendingCheckpointerStats
.checkpoint_write_time
+= write_msecs
;
6255 PendingCheckpointerStats
.checkpoint_sync_time
+= sync_msecs
;
6258 * All of the published timing statistics are accounted for. Only
6259 * continue if a log message is to be written.
6261 if (!log_checkpoints
)
6264 total_msecs
= TimestampDifferenceMilliseconds(CheckpointStats
.ckpt_start_t
,
6265 CheckpointStats
.ckpt_end_t
);
6268 * Timing values returned from CheckpointStats are in microseconds.
6269 * Convert to milliseconds for consistent printing.
6271 longest_msecs
= (long) ((CheckpointStats
.ckpt_longest_sync
+ 999) / 1000);
6273 average_sync_time
= 0;
6274 if (CheckpointStats
.ckpt_sync_rels
> 0)
6275 average_sync_time
= CheckpointStats
.ckpt_agg_sync_time
/
6276 CheckpointStats
.ckpt_sync_rels
;
6277 average_msecs
= (long) ((average_sync_time
+ 999) / 1000);
6280 * ControlFileLock is not required to see ControlFile->checkPoint and
6281 * ->checkPointCopy here as we are the only updator of those variables at
6286 (errmsg("restartpoint complete: wrote %d buffers (%.1f%%); "
6287 "%d WAL file(s) added, %d removed, %d recycled; "
6288 "write=%ld.%03d s, sync=%ld.%03d s, total=%ld.%03d s; "
6289 "sync files=%d, longest=%ld.%03d s, average=%ld.%03d s; "
6290 "distance=%d kB, estimate=%d kB; "
6291 "lsn=%X/%X, redo lsn=%X/%X",
6292 CheckpointStats
.ckpt_bufs_written
,
6293 (double) CheckpointStats
.ckpt_bufs_written
* 100 / NBuffers
,
6294 CheckpointStats
.ckpt_segs_added
,
6295 CheckpointStats
.ckpt_segs_removed
,
6296 CheckpointStats
.ckpt_segs_recycled
,
6297 write_msecs
/ 1000, (int) (write_msecs
% 1000),
6298 sync_msecs
/ 1000, (int) (sync_msecs
% 1000),
6299 total_msecs
/ 1000, (int) (total_msecs
% 1000),
6300 CheckpointStats
.ckpt_sync_rels
,
6301 longest_msecs
/ 1000, (int) (longest_msecs
% 1000),
6302 average_msecs
/ 1000, (int) (average_msecs
% 1000),
6303 (int) (PrevCheckPointDistance
/ 1024.0),
6304 (int) (CheckPointDistanceEstimate
/ 1024.0),
6305 LSN_FORMAT_ARGS(ControlFile
->checkPoint
),
6306 LSN_FORMAT_ARGS(ControlFile
->checkPointCopy
.redo
))));
6309 (errmsg("checkpoint complete: wrote %d buffers (%.1f%%); "
6310 "%d WAL file(s) added, %d removed, %d recycled; "
6311 "write=%ld.%03d s, sync=%ld.%03d s, total=%ld.%03d s; "
6312 "sync files=%d, longest=%ld.%03d s, average=%ld.%03d s; "
6313 "distance=%d kB, estimate=%d kB; "
6314 "lsn=%X/%X, redo lsn=%X/%X",
6315 CheckpointStats
.ckpt_bufs_written
,
6316 (double) CheckpointStats
.ckpt_bufs_written
* 100 / NBuffers
,
6317 CheckpointStats
.ckpt_segs_added
,
6318 CheckpointStats
.ckpt_segs_removed
,
6319 CheckpointStats
.ckpt_segs_recycled
,
6320 write_msecs
/ 1000, (int) (write_msecs
% 1000),
6321 sync_msecs
/ 1000, (int) (sync_msecs
% 1000),
6322 total_msecs
/ 1000, (int) (total_msecs
% 1000),
6323 CheckpointStats
.ckpt_sync_rels
,
6324 longest_msecs
/ 1000, (int) (longest_msecs
% 1000),
6325 average_msecs
/ 1000, (int) (average_msecs
% 1000),
6326 (int) (PrevCheckPointDistance
/ 1024.0),
6327 (int) (CheckPointDistanceEstimate
/ 1024.0),
6328 LSN_FORMAT_ARGS(ControlFile
->checkPoint
),
6329 LSN_FORMAT_ARGS(ControlFile
->checkPointCopy
.redo
))));
6333 * Update the estimate of distance between checkpoints.
6335 * The estimate is used to calculate the number of WAL segments to keep
6336 * preallocated, see XLOGfileslop().
6339 UpdateCheckPointDistanceEstimate(uint64 nbytes
)
6342 * To estimate the number of segments consumed between checkpoints, keep a
6343 * moving average of the amount of WAL generated in previous checkpoint
6344 * cycles. However, if the load is bursty, with quiet periods and busy
6345 * periods, we want to cater for the peak load. So instead of a plain
6346 * moving average, let the average decline slowly if the previous cycle
6347 * used less WAL than estimated, but bump it up immediately if it used
6350 * When checkpoints are triggered by max_wal_size, this should converge to
6351 * CheckpointSegments * wal_segment_size,
6353 * Note: This doesn't pay any attention to what caused the checkpoint.
6354 * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
6355 * starting a base backup, are counted the same as those created
6356 * automatically. The slow-decline will largely mask them out, if they are
6357 * not frequent. If they are frequent, it seems reasonable to count them
6358 * in as any others; if you issue a manual checkpoint every 5 minutes and
6359 * never let a timed checkpoint happen, it makes sense to base the
6360 * preallocation on that 5 minute interval rather than whatever
6361 * checkpoint_timeout is set to.
6363 PrevCheckPointDistance
= nbytes
;
6364 if (CheckPointDistanceEstimate
< nbytes
)
6365 CheckPointDistanceEstimate
= nbytes
;
6367 CheckPointDistanceEstimate
=
6368 (0.90 * CheckPointDistanceEstimate
+ 0.10 * (double) nbytes
);
6372 * Update the ps display for a process running a checkpoint. Note that
6373 * this routine should not do any allocations so as it can be called
6374 * from a critical section.
6377 update_checkpoint_display(int flags
, bool restartpoint
, bool reset
)
6380 * The status is reported only for end-of-recovery and shutdown
6381 * checkpoints or shutdown restartpoints. Updating the ps display is
6382 * useful in those situations as it may not be possible to rely on
6383 * pg_stat_activity to see the status of the checkpointer or the startup
6386 if ((flags
& (CHECKPOINT_END_OF_RECOVERY
| CHECKPOINT_IS_SHUTDOWN
)) == 0)
6393 char activitymsg
[128];
6395 snprintf(activitymsg
, sizeof(activitymsg
), "performing %s%s%s",
6396 (flags
& CHECKPOINT_END_OF_RECOVERY
) ? "end-of-recovery " : "",
6397 (flags
& CHECKPOINT_IS_SHUTDOWN
) ? "shutdown " : "",
6398 restartpoint
? "restartpoint" : "checkpoint");
6399 set_ps_display(activitymsg
);
6405 * Perform a checkpoint --- either during shutdown, or on-the-fly
6407 * flags is a bitwise OR of the following:
6408 * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
6409 * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
6410 * CHECKPOINT_IMMEDIATE: finish the checkpoint ASAP,
6411 * ignoring checkpoint_completion_target parameter.
6412 * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
6413 * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
6414 * CHECKPOINT_END_OF_RECOVERY).
6415 * CHECKPOINT_FLUSH_ALL: also flush buffers of unlogged tables.
6417 * Note: flags contains other bits, of interest here only for logging purposes.
6418 * In particular note that this routine is synchronous and does not pay
6419 * attention to CHECKPOINT_WAIT.
6421 * If !shutdown then we are writing an online checkpoint. This is a very special
6422 * kind of operation and WAL record because the checkpoint action occurs over
6423 * a period of time yet logically occurs at just a single LSN. The logical
6424 * position of the WAL record (redo ptr) is the same or earlier than the
6425 * physical position. When we replay WAL we locate the checkpoint via its
6426 * physical position then read the redo ptr and actually start replay at the
6427 * earlier logical position. Note that we don't write *anything* to WAL at
6428 * the logical position, so that location could be any other kind of WAL record.
6429 * All of this mechanism allows us to continue working while we checkpoint.
6430 * As a result, timing of actions is critical here and be careful to note that
6431 * this function will likely take minutes to execute on a busy system.
6434 CreateCheckPoint(int flags
)
6437 CheckPoint checkPoint
;
6439 XLogSegNo _logSegNo
;
6440 XLogCtlInsert
*Insert
= &XLogCtl
->Insert
;
6442 XLogRecPtr PriorRedoPtr
;
6443 XLogRecPtr curInsert
;
6444 XLogRecPtr last_important_lsn
;
6445 VirtualTransactionId
*vxids
;
6447 int oldXLogAllowed
= 0;
6450 * An end-of-recovery checkpoint is really a shutdown checkpoint, just
6451 * issued at a different time.
6453 if (flags
& (CHECKPOINT_IS_SHUTDOWN
| CHECKPOINT_END_OF_RECOVERY
))
6459 if (RecoveryInProgress() && (flags
& CHECKPOINT_END_OF_RECOVERY
) == 0)
6460 elog(ERROR
, "can't create a checkpoint during recovery");
6463 * Prepare to accumulate statistics.
6465 * Note: because it is possible for log_checkpoints to change while a
6466 * checkpoint proceeds, we always accumulate stats, even if
6467 * log_checkpoints is currently off.
6469 MemSet(&CheckpointStats
, 0, sizeof(CheckpointStats
));
6470 CheckpointStats
.ckpt_start_t
= GetCurrentTimestamp();
6473 * Let smgr prepare for checkpoint; this has to happen outside the
6474 * critical section and before we determine the REDO pointer. Note that
6475 * smgr must not do anything that'd have to be undone if we decide no
6476 * checkpoint is needed.
6478 SyncPreCheckpoint();
6481 * Use a critical section to force system panic if we have trouble.
6483 START_CRIT_SECTION();
6487 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
6488 ControlFile
->state
= DB_SHUTDOWNING
;
6489 UpdateControlFile();
6490 LWLockRelease(ControlFileLock
);
6493 /* Begin filling in the checkpoint WAL record */
6494 MemSet(&checkPoint
, 0, sizeof(checkPoint
));
6495 checkPoint
.time
= (pg_time_t
) time(NULL
);
6498 * For Hot Standby, derive the oldestActiveXid before we fix the redo
6499 * pointer. This allows us to begin accumulating changes to assemble our
6500 * starting snapshot of locks and transactions.
6502 if (!shutdown
&& XLogStandbyInfoActive())
6503 checkPoint
.oldestActiveXid
= GetOldestActiveTransactionId();
6505 checkPoint
.oldestActiveXid
= InvalidTransactionId
;
6508 * Get location of last important record before acquiring insert locks (as
6509 * GetLastImportantRecPtr() also locks WAL locks).
6511 last_important_lsn
= GetLastImportantRecPtr();
6514 * We must block concurrent insertions while examining insert state to
6515 * determine the checkpoint REDO pointer.
6517 WALInsertLockAcquireExclusive();
6518 curInsert
= XLogBytePosToRecPtr(Insert
->CurrBytePos
);
6521 * If this isn't a shutdown or forced checkpoint, and if there has been no
6522 * WAL activity requiring a checkpoint, skip it. The idea here is to
6523 * avoid inserting duplicate checkpoints when the system is idle.
6525 if ((flags
& (CHECKPOINT_IS_SHUTDOWN
| CHECKPOINT_END_OF_RECOVERY
|
6526 CHECKPOINT_FORCE
)) == 0)
6528 if (last_important_lsn
== ControlFile
->checkPoint
)
6530 WALInsertLockRelease();
6533 (errmsg_internal("checkpoint skipped because system is idle")));
6539 * An end-of-recovery checkpoint is created before anyone is allowed to
6540 * write WAL. To allow us to write the checkpoint record, temporarily
6541 * enable XLogInsertAllowed.
6543 if (flags
& CHECKPOINT_END_OF_RECOVERY
)
6544 oldXLogAllowed
= LocalSetXLogInsertAllowed();
6546 checkPoint
.ThisTimeLineID
= XLogCtl
->InsertTimeLineID
;
6547 if (flags
& CHECKPOINT_END_OF_RECOVERY
)
6548 checkPoint
.PrevTimeLineID
= XLogCtl
->PrevTimeLineID
;
6550 checkPoint
.PrevTimeLineID
= checkPoint
.ThisTimeLineID
;
6552 checkPoint
.fullPageWrites
= Insert
->fullPageWrites
;
6555 * Compute new REDO record ptr = location of next XLOG record.
6557 * NB: this is NOT necessarily where the checkpoint record itself will be,
6558 * since other backends may insert more XLOG records while we're off doing
6559 * the buffer flush work. Those XLOG records are logically after the
6560 * checkpoint, even though physically before it. Got that?
6562 freespace
= INSERT_FREESPACE(curInsert
);
6565 if (XLogSegmentOffset(curInsert
, wal_segment_size
) == 0)
6566 curInsert
+= SizeOfXLogLongPHD
;
6568 curInsert
+= SizeOfXLogShortPHD
;
6570 checkPoint
.redo
= curInsert
;
6573 * Here we update the shared RedoRecPtr for future XLogInsert calls; this
6574 * must be done while holding all the insertion locks.
6576 * Note: if we fail to complete the checkpoint, RedoRecPtr will be left
6577 * pointing past where it really needs to point. This is okay; the only
6578 * consequence is that XLogInsert might back up whole buffers that it
6579 * didn't really need to. We can't postpone advancing RedoRecPtr because
6580 * XLogInserts that happen while we are dumping buffers must assume that
6581 * their buffer changes are not included in the checkpoint.
6583 RedoRecPtr
= XLogCtl
->Insert
.RedoRecPtr
= checkPoint
.redo
;
6586 * Now we can release the WAL insertion locks, allowing other xacts to
6587 * proceed while we are flushing disk buffers.
6589 WALInsertLockRelease();
6591 /* Update the info_lck-protected copy of RedoRecPtr as well */
6592 SpinLockAcquire(&XLogCtl
->info_lck
);
6593 XLogCtl
->RedoRecPtr
= checkPoint
.redo
;
6594 SpinLockRelease(&XLogCtl
->info_lck
);
6597 * If enabled, log checkpoint start. We postpone this until now so as not
6598 * to log anything if we decided to skip the checkpoint.
6600 if (log_checkpoints
)
6601 LogCheckpointStart(flags
, false);
6603 /* Update the process title */
6604 update_checkpoint_display(flags
, false, false);
6606 TRACE_POSTGRESQL_CHECKPOINT_START(flags
);
6609 * Get the other info we need for the checkpoint record.
6611 * We don't need to save oldestClogXid in the checkpoint, it only matters
6612 * for the short period in which clog is being truncated, and if we crash
6613 * during that we'll redo the clog truncation and fix up oldestClogXid
6616 LWLockAcquire(XidGenLock
, LW_SHARED
);
6617 checkPoint
.nextXid
= ShmemVariableCache
->nextXid
;
6618 checkPoint
.oldestXid
= ShmemVariableCache
->oldestXid
;
6619 checkPoint
.oldestXidDB
= ShmemVariableCache
->oldestXidDB
;
6620 LWLockRelease(XidGenLock
);
6622 LWLockAcquire(CommitTsLock
, LW_SHARED
);
6623 checkPoint
.oldestCommitTsXid
= ShmemVariableCache
->oldestCommitTsXid
;
6624 checkPoint
.newestCommitTsXid
= ShmemVariableCache
->newestCommitTsXid
;
6625 LWLockRelease(CommitTsLock
);
6627 LWLockAcquire(OidGenLock
, LW_SHARED
);
6628 checkPoint
.nextOid
= ShmemVariableCache
->nextOid
;
6630 checkPoint
.nextOid
+= ShmemVariableCache
->oidCount
;
6631 LWLockRelease(OidGenLock
);
6633 MultiXactGetCheckptMulti(shutdown
,
6634 &checkPoint
.nextMulti
,
6635 &checkPoint
.nextMultiOffset
,
6636 &checkPoint
.oldestMulti
,
6637 &checkPoint
.oldestMultiDB
);
6640 * Having constructed the checkpoint record, ensure all shmem disk buffers
6641 * and commit-log buffers are flushed to disk.
6643 * This I/O could fail for various reasons. If so, we will fail to
6644 * complete the checkpoint, but there is no reason to force a system
6645 * panic. Accordingly, exit critical section while doing it.
6650 * In some cases there are groups of actions that must all occur on one
6651 * side or the other of a checkpoint record. Before flushing the
6652 * checkpoint record we must explicitly wait for any backend currently
6653 * performing those groups of actions.
6655 * One example is end of transaction, so we must wait for any transactions
6656 * that are currently in commit critical sections. If an xact inserted
6657 * its commit record into XLOG just before the REDO point, then a crash
6658 * restart from the REDO point would not replay that record, which means
6659 * that our flushing had better include the xact's update of pg_xact. So
6660 * we wait till he's out of his commit critical section before proceeding.
6661 * See notes in RecordTransactionCommit().
6663 * Because we've already released the insertion locks, this test is a bit
6664 * fuzzy: it is possible that we will wait for xacts we didn't really need
6665 * to wait for. But the delay should be short and it seems better to make
6666 * checkpoint take a bit longer than to hold off insertions longer than
6667 * necessary. (In fact, the whole reason we have this issue is that xact.c
6668 * does commit record XLOG insertion and clog update as two separate steps
6669 * protected by different locks, but again that seems best on grounds of
6670 * minimizing lock contention.)
6672 * A transaction that has not yet set delayChkptFlags when we look cannot
6673 * be at risk, since it has not inserted its commit record yet; and one
6674 * that's already cleared it is not at risk either, since it's done fixing
6675 * clog and we will correctly flush the update below. So we cannot miss
6676 * any xacts we need to wait for.
6678 vxids
= GetVirtualXIDsDelayingChkpt(&nvxids
, DELAY_CHKPT_START
);
6683 pg_usleep(10000L); /* wait for 10 msec */
6684 } while (HaveVirtualXIDsDelayingChkpt(vxids
, nvxids
,
6685 DELAY_CHKPT_START
));
6689 CheckPointGuts(checkPoint
.redo
, flags
);
6691 vxids
= GetVirtualXIDsDelayingChkpt(&nvxids
, DELAY_CHKPT_COMPLETE
);
6696 pg_usleep(10000L); /* wait for 10 msec */
6697 } while (HaveVirtualXIDsDelayingChkpt(vxids
, nvxids
,
6698 DELAY_CHKPT_COMPLETE
));
6703 * Take a snapshot of running transactions and write this to WAL. This
6704 * allows us to reconstruct the state of running transactions during
6705 * archive recovery, if required. Skip, if this info disabled.
6707 * If we are shutting down, or Startup process is completing crash
6708 * recovery we don't need to write running xact data.
6710 if (!shutdown
&& XLogStandbyInfoActive())
6711 LogStandbySnapshot();
6713 START_CRIT_SECTION();
6716 * Now insert the checkpoint record into XLOG.
6719 XLogRegisterData((char *) (&checkPoint
), sizeof(checkPoint
));
6720 recptr
= XLogInsert(RM_XLOG_ID
,
6721 shutdown
? XLOG_CHECKPOINT_SHUTDOWN
:
6722 XLOG_CHECKPOINT_ONLINE
);
6727 * We mustn't write any new WAL after a shutdown checkpoint, or it will be
6728 * overwritten at next startup. No-one should even try, this just allows
6729 * sanity-checking. In the case of an end-of-recovery checkpoint, we want
6730 * to just temporarily disable writing until the system has exited
6735 if (flags
& CHECKPOINT_END_OF_RECOVERY
)
6736 LocalXLogInsertAllowed
= oldXLogAllowed
;
6738 LocalXLogInsertAllowed
= 0; /* never again write WAL */
6742 * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
6743 * = end of actual checkpoint record.
6745 if (shutdown
&& checkPoint
.redo
!= ProcLastRecPtr
)
6747 (errmsg("concurrent write-ahead log activity while database system is shutting down")));
6750 * Remember the prior checkpoint's redo ptr for
6751 * UpdateCheckPointDistanceEstimate()
6753 PriorRedoPtr
= ControlFile
->checkPointCopy
.redo
;
6756 * Update the control file.
6758 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
6760 ControlFile
->state
= DB_SHUTDOWNED
;
6761 ControlFile
->checkPoint
= ProcLastRecPtr
;
6762 ControlFile
->checkPointCopy
= checkPoint
;
6763 /* crash recovery should always recover to the end of WAL */
6764 ControlFile
->minRecoveryPoint
= InvalidXLogRecPtr
;
6765 ControlFile
->minRecoveryPointTLI
= 0;
6768 * Persist unloggedLSN value. It's reset on crash recovery, so this goes
6769 * unused on non-shutdown checkpoints, but seems useful to store it always
6770 * for debugging purposes.
6772 SpinLockAcquire(&XLogCtl
->ulsn_lck
);
6773 ControlFile
->unloggedLSN
= XLogCtl
->unloggedLSN
;
6774 SpinLockRelease(&XLogCtl
->ulsn_lck
);
6776 UpdateControlFile();
6777 LWLockRelease(ControlFileLock
);
6779 /* Update shared-memory copy of checkpoint XID/epoch */
6780 SpinLockAcquire(&XLogCtl
->info_lck
);
6781 XLogCtl
->ckptFullXid
= checkPoint
.nextXid
;
6782 SpinLockRelease(&XLogCtl
->info_lck
);
6785 * We are now done with critical updates; no need for system panic if we
6786 * have trouble while fooling with old log segments.
6791 * Let smgr do post-checkpoint cleanup (eg, deleting old files).
6793 SyncPostCheckpoint();
6796 * Update the average distance between checkpoints if the prior checkpoint
6799 if (PriorRedoPtr
!= InvalidXLogRecPtr
)
6800 UpdateCheckPointDistanceEstimate(RedoRecPtr
- PriorRedoPtr
);
6803 * Delete old log files, those no longer needed for last checkpoint to
6804 * prevent the disk holding the xlog from growing full.
6806 XLByteToSeg(RedoRecPtr
, _logSegNo
, wal_segment_size
);
6807 KeepLogSeg(recptr
, &_logSegNo
);
6808 if (InvalidateObsoleteReplicationSlots(_logSegNo
))
6811 * Some slots have been invalidated; recalculate the old-segment
6812 * horizon, starting again from RedoRecPtr.
6814 XLByteToSeg(RedoRecPtr
, _logSegNo
, wal_segment_size
);
6815 KeepLogSeg(recptr
, &_logSegNo
);
6818 RemoveOldXlogFiles(_logSegNo
, RedoRecPtr
, recptr
,
6819 checkPoint
.ThisTimeLineID
);
6822 * Make more log segments if needed. (Do this after recycling old log
6823 * segments, since that may supply some of the needed files.)
6826 PreallocXlogFiles(recptr
, checkPoint
.ThisTimeLineID
);
6829 * Truncate pg_subtrans if possible. We can throw away all data before
6830 * the oldest XMIN of any running transaction. No future transaction will
6831 * attempt to reference any pg_subtrans entry older than that (see Asserts
6832 * in subtrans.c). During recovery, though, we mustn't do this because
6833 * StartupSUBTRANS hasn't been called yet.
6835 if (!RecoveryInProgress())
6836 TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
6838 /* Real work is done; log and update stats. */
6839 LogCheckpointEnd(false);
6841 /* Reset the process title */
6842 update_checkpoint_display(flags
, false, true);
6844 TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats
.ckpt_bufs_written
,
6846 CheckpointStats
.ckpt_segs_added
,
6847 CheckpointStats
.ckpt_segs_removed
,
6848 CheckpointStats
.ckpt_segs_recycled
);
6852 * Mark the end of recovery in WAL though without running a full checkpoint.
6853 * We can expect that a restartpoint is likely to be in progress as we
6854 * do this, though we are unwilling to wait for it to complete.
6856 * CreateRestartPoint() allows for the case where recovery may end before
6857 * the restartpoint completes so there is no concern of concurrent behaviour.
6860 CreateEndOfRecoveryRecord(void)
6862 xl_end_of_recovery xlrec
;
6866 if (!RecoveryInProgress())
6867 elog(ERROR
, "can only be used to end recovery");
6869 xlrec
.end_time
= GetCurrentTimestamp();
6871 WALInsertLockAcquireExclusive();
6872 xlrec
.ThisTimeLineID
= XLogCtl
->InsertTimeLineID
;
6873 xlrec
.PrevTimeLineID
= XLogCtl
->PrevTimeLineID
;
6874 WALInsertLockRelease();
6876 START_CRIT_SECTION();
6879 XLogRegisterData((char *) &xlrec
, sizeof(xl_end_of_recovery
));
6880 recptr
= XLogInsert(RM_XLOG_ID
, XLOG_END_OF_RECOVERY
);
6885 * Update the control file so that crash recovery can follow the timeline
6886 * changes to this point.
6888 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
6889 ControlFile
->minRecoveryPoint
= recptr
;
6890 ControlFile
->minRecoveryPointTLI
= xlrec
.ThisTimeLineID
;
6891 UpdateControlFile();
6892 LWLockRelease(ControlFileLock
);
6898 * Write an OVERWRITE_CONTRECORD message.
6900 * When on WAL replay we expect a continuation record at the start of a page
6901 * that is not there, recovery ends and WAL writing resumes at that point.
6902 * But it's wrong to resume writing new WAL back at the start of the record
6903 * that was broken, because downstream consumers of that WAL (physical
6904 * replicas) are not prepared to "rewind". So the first action after
6905 * finishing replay of all valid WAL must be to write a record of this type
6906 * at the point where the contrecord was missing; to support xlogreader
6907 * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
6908 * to the page header where the record occurs. xlogreader has an ad-hoc
6909 * mechanism to report metadata about the broken record, which is what we
6912 * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
6913 * skip the record it was reading, and pass back the LSN of the skipped
6914 * record, so that its caller can verify (on "replay" of that record) that the
6915 * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
6917 * 'aborted_lsn' is the beginning position of the record that was incomplete.
6918 * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
6919 * beginning of the XLOG page where the record is to be inserted. They must
6920 * match the current WAL insert position, they're passed here just so that we
6924 CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn
, XLogRecPtr pagePtr
,
6927 xl_overwrite_contrecord xlrec
;
6929 XLogPageHeader pagehdr
;
6930 XLogRecPtr startPos
;
6933 if (!RecoveryInProgress())
6934 elog(ERROR
, "can only be used at end of recovery");
6935 if (pagePtr
% XLOG_BLCKSZ
!= 0)
6936 elog(ERROR
, "invalid position for missing continuation record %X/%X",
6937 LSN_FORMAT_ARGS(pagePtr
));
6939 /* The current WAL insert position should be right after the page header */
6941 if (XLogSegmentOffset(startPos
, wal_segment_size
) == 0)
6942 startPos
+= SizeOfXLogLongPHD
;
6944 startPos
+= SizeOfXLogShortPHD
;
6945 recptr
= GetXLogInsertRecPtr();
6946 if (recptr
!= startPos
)
6947 elog(ERROR
, "invalid WAL insert position %X/%X for OVERWRITE_CONTRECORD",
6948 LSN_FORMAT_ARGS(recptr
));
6950 START_CRIT_SECTION();
6953 * Initialize the XLOG page header (by GetXLogBuffer), and set the
6954 * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
6956 * No other backend is allowed to write WAL yet, so acquiring the WAL
6957 * insertion lock is just pro forma.
6959 WALInsertLockAcquire();
6960 pagehdr
= (XLogPageHeader
) GetXLogBuffer(pagePtr
, newTLI
);
6961 pagehdr
->xlp_info
|= XLP_FIRST_IS_OVERWRITE_CONTRECORD
;
6962 WALInsertLockRelease();
6965 * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
6966 * page. We know it becomes the first record, because no other backend is
6967 * allowed to write WAL yet.
6970 xlrec
.overwritten_lsn
= aborted_lsn
;
6971 xlrec
.overwrite_time
= GetCurrentTimestamp();
6972 XLogRegisterData((char *) &xlrec
, sizeof(xl_overwrite_contrecord
));
6973 recptr
= XLogInsert(RM_XLOG_ID
, XLOG_OVERWRITE_CONTRECORD
);
6975 /* check that the record was inserted to the right place */
6976 if (ProcLastRecPtr
!= startPos
)
6977 elog(ERROR
, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%X",
6978 LSN_FORMAT_ARGS(ProcLastRecPtr
));
6988 * Flush all data in shared memory to disk, and fsync
6990 * This is the common code shared between regular checkpoints and
6991 * recovery restartpoints.
6994 CheckPointGuts(XLogRecPtr checkPointRedo
, int flags
)
6996 CheckPointRelationMap();
6997 CheckPointReplicationSlots();
6998 CheckPointSnapBuild();
6999 CheckPointLogicalRewriteHeap();
7000 CheckPointReplicationOrigin();
7002 /* Write out all dirty data in SLRUs and the main buffer pool */
7003 TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags
);
7004 CheckpointStats
.ckpt_write_t
= GetCurrentTimestamp();
7006 CheckPointCommitTs();
7007 CheckPointSUBTRANS();
7008 CheckPointMultiXact();
7009 CheckPointPredicate();
7010 CheckPointBuffers(flags
);
7012 /* Perform all queued up fsyncs */
7013 TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
7014 CheckpointStats
.ckpt_sync_t
= GetCurrentTimestamp();
7015 ProcessSyncRequests();
7016 CheckpointStats
.ckpt_sync_end_t
= GetCurrentTimestamp();
7017 TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
7019 /* We deliberately delay 2PC checkpointing as long as possible */
7020 CheckPointTwoPhase(checkPointRedo
);
7024 * Save a checkpoint for recovery restart if appropriate
7026 * This function is called each time a checkpoint record is read from XLOG.
7027 * It must determine whether the checkpoint represents a safe restartpoint or
7028 * not. If so, the checkpoint record is stashed in shared memory so that
7029 * CreateRestartPoint can consult it. (Note that the latter function is
7030 * executed by the checkpointer, while this one will be executed by the
7034 RecoveryRestartPoint(const CheckPoint
*checkPoint
, XLogReaderState
*record
)
7037 * Also refrain from creating a restartpoint if we have seen any
7038 * references to non-existent pages. Restarting recovery from the
7039 * restartpoint would not see the references, so we would lose the
7040 * cross-check that the pages belonged to a relation that was dropped
7043 if (XLogHaveInvalidPages())
7045 elog(trace_recovery(DEBUG2
),
7046 "could not record restart point at %X/%X because there "
7047 "are unresolved references to invalid pages",
7048 LSN_FORMAT_ARGS(checkPoint
->redo
));
7053 * Copy the checkpoint record to shared memory, so that checkpointer can
7054 * work out the next time it wants to perform a restartpoint.
7056 SpinLockAcquire(&XLogCtl
->info_lck
);
7057 XLogCtl
->lastCheckPointRecPtr
= record
->ReadRecPtr
;
7058 XLogCtl
->lastCheckPointEndPtr
= record
->EndRecPtr
;
7059 XLogCtl
->lastCheckPoint
= *checkPoint
;
7060 SpinLockRelease(&XLogCtl
->info_lck
);
7064 * Establish a restartpoint if possible.
7066 * This is similar to CreateCheckPoint, but is used during WAL recovery
7067 * to establish a point from which recovery can roll forward without
7068 * replaying the entire recovery log.
7070 * Returns true if a new restartpoint was established. We can only establish
7071 * a restartpoint if we have replayed a safe checkpoint record since last
7075 CreateRestartPoint(int flags
)
7077 XLogRecPtr lastCheckPointRecPtr
;
7078 XLogRecPtr lastCheckPointEndPtr
;
7079 CheckPoint lastCheckPoint
;
7080 XLogRecPtr PriorRedoPtr
;
7081 XLogRecPtr receivePtr
;
7082 XLogRecPtr replayPtr
;
7083 TimeLineID replayTLI
;
7085 XLogSegNo _logSegNo
;
7088 /* Concurrent checkpoint/restartpoint cannot happen */
7089 Assert(!IsUnderPostmaster
|| MyBackendType
== B_CHECKPOINTER
);
7091 /* Get a local copy of the last safe checkpoint record. */
7092 SpinLockAcquire(&XLogCtl
->info_lck
);
7093 lastCheckPointRecPtr
= XLogCtl
->lastCheckPointRecPtr
;
7094 lastCheckPointEndPtr
= XLogCtl
->lastCheckPointEndPtr
;
7095 lastCheckPoint
= XLogCtl
->lastCheckPoint
;
7096 SpinLockRelease(&XLogCtl
->info_lck
);
7099 * Check that we're still in recovery mode. It's ok if we exit recovery
7100 * mode after this check, the restart point is valid anyway.
7102 if (!RecoveryInProgress())
7105 (errmsg_internal("skipping restartpoint, recovery has already ended")));
7110 * If the last checkpoint record we've replayed is already our last
7111 * restartpoint, we can't perform a new restart point. We still update
7112 * minRecoveryPoint in that case, so that if this is a shutdown restart
7113 * point, we won't start up earlier than before. That's not strictly
7114 * necessary, but when hot standby is enabled, it would be rather weird if
7115 * the database opened up for read-only connections at a point-in-time
7116 * before the last shutdown. Such time travel is still possible in case of
7117 * immediate shutdown, though.
7119 * We don't explicitly advance minRecoveryPoint when we do create a
7120 * restartpoint. It's assumed that flushing the buffers will do that as a
7123 if (XLogRecPtrIsInvalid(lastCheckPointRecPtr
) ||
7124 lastCheckPoint
.redo
<= ControlFile
->checkPointCopy
.redo
)
7127 (errmsg_internal("skipping restartpoint, already performed at %X/%X",
7128 LSN_FORMAT_ARGS(lastCheckPoint
.redo
))));
7130 UpdateMinRecoveryPoint(InvalidXLogRecPtr
, true);
7131 if (flags
& CHECKPOINT_IS_SHUTDOWN
)
7133 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
7134 ControlFile
->state
= DB_SHUTDOWNED_IN_RECOVERY
;
7135 UpdateControlFile();
7136 LWLockRelease(ControlFileLock
);
7142 * Update the shared RedoRecPtr so that the startup process can calculate
7143 * the number of segments replayed since last restartpoint, and request a
7144 * restartpoint if it exceeds CheckPointSegments.
7146 * Like in CreateCheckPoint(), hold off insertions to update it, although
7147 * during recovery this is just pro forma, because no WAL insertions are
7150 WALInsertLockAcquireExclusive();
7151 RedoRecPtr
= XLogCtl
->Insert
.RedoRecPtr
= lastCheckPoint
.redo
;
7152 WALInsertLockRelease();
7154 /* Also update the info_lck-protected copy */
7155 SpinLockAcquire(&XLogCtl
->info_lck
);
7156 XLogCtl
->RedoRecPtr
= lastCheckPoint
.redo
;
7157 SpinLockRelease(&XLogCtl
->info_lck
);
7160 * Prepare to accumulate statistics.
7162 * Note: because it is possible for log_checkpoints to change while a
7163 * checkpoint proceeds, we always accumulate stats, even if
7164 * log_checkpoints is currently off.
7166 MemSet(&CheckpointStats
, 0, sizeof(CheckpointStats
));
7167 CheckpointStats
.ckpt_start_t
= GetCurrentTimestamp();
7169 if (log_checkpoints
)
7170 LogCheckpointStart(flags
, true);
7172 /* Update the process title */
7173 update_checkpoint_display(flags
, true, false);
7175 CheckPointGuts(lastCheckPoint
.redo
, flags
);
7178 * Remember the prior checkpoint's redo ptr for
7179 * UpdateCheckPointDistanceEstimate()
7181 PriorRedoPtr
= ControlFile
->checkPointCopy
.redo
;
7184 * Update pg_control, using current time. Check that it still shows an
7185 * older checkpoint, else do nothing; this is a quick hack to make sure
7186 * nothing really bad happens if somehow we get here after the
7187 * end-of-recovery checkpoint.
7189 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
7190 if (ControlFile
->checkPointCopy
.redo
< lastCheckPoint
.redo
)
7193 * Update the checkpoint information. We do this even if the cluster
7194 * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
7195 * segments recycled below.
7197 ControlFile
->checkPoint
= lastCheckPointRecPtr
;
7198 ControlFile
->checkPointCopy
= lastCheckPoint
;
7201 * Ensure minRecoveryPoint is past the checkpoint record and update it
7202 * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
7203 * this will have happened already while writing out dirty buffers,
7204 * but not necessarily - e.g. because no buffers were dirtied. We do
7205 * this because a backup performed in recovery uses minRecoveryPoint
7206 * to determine which WAL files must be included in the backup, and
7207 * the file (or files) containing the checkpoint record must be
7208 * included, at a minimum. Note that for an ordinary restart of
7209 * recovery there's no value in having the minimum recovery point any
7210 * earlier than this anyway, because redo will begin just after the
7211 * checkpoint record.
7213 if (ControlFile
->state
== DB_IN_ARCHIVE_RECOVERY
)
7215 if (ControlFile
->minRecoveryPoint
< lastCheckPointEndPtr
)
7217 ControlFile
->minRecoveryPoint
= lastCheckPointEndPtr
;
7218 ControlFile
->minRecoveryPointTLI
= lastCheckPoint
.ThisTimeLineID
;
7220 /* update local copy */
7221 LocalMinRecoveryPoint
= ControlFile
->minRecoveryPoint
;
7222 LocalMinRecoveryPointTLI
= ControlFile
->minRecoveryPointTLI
;
7224 if (flags
& CHECKPOINT_IS_SHUTDOWN
)
7225 ControlFile
->state
= DB_SHUTDOWNED_IN_RECOVERY
;
7227 UpdateControlFile();
7229 LWLockRelease(ControlFileLock
);
7232 * Update the average distance between checkpoints/restartpoints if the
7233 * prior checkpoint exists.
7235 if (PriorRedoPtr
!= InvalidXLogRecPtr
)
7236 UpdateCheckPointDistanceEstimate(RedoRecPtr
- PriorRedoPtr
);
7239 * Delete old log files, those no longer needed for last restartpoint to
7240 * prevent the disk holding the xlog from growing full.
7242 XLByteToSeg(RedoRecPtr
, _logSegNo
, wal_segment_size
);
7245 * Retreat _logSegNo using the current end of xlog replayed or received,
7246 * whichever is later.
7248 receivePtr
= GetWalRcvFlushRecPtr(NULL
, NULL
);
7249 replayPtr
= GetXLogReplayRecPtr(&replayTLI
);
7250 endptr
= (receivePtr
< replayPtr
) ? replayPtr
: receivePtr
;
7251 KeepLogSeg(endptr
, &_logSegNo
);
7252 if (InvalidateObsoleteReplicationSlots(_logSegNo
))
7255 * Some slots have been invalidated; recalculate the old-segment
7256 * horizon, starting again from RedoRecPtr.
7258 XLByteToSeg(RedoRecPtr
, _logSegNo
, wal_segment_size
);
7259 KeepLogSeg(endptr
, &_logSegNo
);
7264 * Try to recycle segments on a useful timeline. If we've been promoted
7265 * since the beginning of this restartpoint, use the new timeline chosen
7266 * at end of recovery. If we're still in recovery, use the timeline we're
7267 * currently replaying.
7269 * There is no guarantee that the WAL segments will be useful on the
7270 * current timeline; if recovery proceeds to a new timeline right after
7271 * this, the pre-allocated WAL segments on this timeline will not be used,
7272 * and will go wasted until recycled on the next restartpoint. We'll live
7275 if (!RecoveryInProgress())
7276 replayTLI
= XLogCtl
->InsertTimeLineID
;
7278 RemoveOldXlogFiles(_logSegNo
, RedoRecPtr
, endptr
, replayTLI
);
7281 * Make more log segments if needed. (Do this after recycling old log
7282 * segments, since that may supply some of the needed files.)
7284 PreallocXlogFiles(endptr
, replayTLI
);
7287 * Truncate pg_subtrans if possible. We can throw away all data before
7288 * the oldest XMIN of any running transaction. No future transaction will
7289 * attempt to reference any pg_subtrans entry older than that (see Asserts
7290 * in subtrans.c). When hot standby is disabled, though, we mustn't do
7291 * this because StartupSUBTRANS hasn't been called yet.
7293 if (EnableHotStandby
)
7294 TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7296 /* Real work is done; log and update stats. */
7297 LogCheckpointEnd(true);
7299 /* Reset the process title */
7300 update_checkpoint_display(flags
, true, true);
7302 xtime
= GetLatestXTime();
7303 ereport((log_checkpoints
? LOG
: DEBUG2
),
7304 (errmsg("recovery restart point at %X/%X",
7305 LSN_FORMAT_ARGS(lastCheckPoint
.redo
)),
7306 xtime
? errdetail("Last completed transaction was at log time %s.",
7307 timestamptz_to_str(xtime
)) : 0));
7310 * Finally, execute archive_cleanup_command, if any.
7312 if (archiveCleanupCommand
&& strcmp(archiveCleanupCommand
, "") != 0)
7314 char lastRestartPointFname
[MAXFNAMELEN
];
7316 GetOldestRestartPointFileName(lastRestartPointFname
);
7317 shell_archive_cleanup(lastRestartPointFname
);
7324 * Report availability of WAL for the given target LSN
7325 * (typically a slot's restart_lsn)
7327 * Returns one of the following enum values:
7329 * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
7332 * * WALAVAIL_EXTENDED means it is still available by preserving extra
7333 * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
7334 * than max_wal_size, this state is not returned.
7336 * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
7337 * remove reserved segments. The walsender using this slot may return to the
7340 * * WALAVAIL_REMOVED means it has been removed. A replication stream on
7341 * a slot with this LSN cannot continue. (Any associated walsender
7342 * processes should have been terminated already.)
7344 * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
7347 GetWALAvailability(XLogRecPtr targetLSN
)
7349 XLogRecPtr currpos
; /* current write LSN */
7350 XLogSegNo currSeg
; /* segid of currpos */
7351 XLogSegNo targetSeg
; /* segid of targetLSN */
7352 XLogSegNo oldestSeg
; /* actual oldest segid */
7353 XLogSegNo oldestSegMaxWalSize
; /* oldest segid kept by max_wal_size */
7354 XLogSegNo oldestSlotSeg
; /* oldest segid kept by slot */
7358 * slot does not reserve WAL. Either deactivated, or has never been active
7360 if (XLogRecPtrIsInvalid(targetLSN
))
7361 return WALAVAIL_INVALID_LSN
;
7364 * Calculate the oldest segment currently reserved by all slots,
7365 * considering wal_keep_size and max_slot_wal_keep_size. Initialize
7366 * oldestSlotSeg to the current segment.
7368 currpos
= GetXLogWriteRecPtr();
7369 XLByteToSeg(currpos
, oldestSlotSeg
, wal_segment_size
);
7370 KeepLogSeg(currpos
, &oldestSlotSeg
);
7373 * Find the oldest extant segment file. We get 1 until checkpoint removes
7374 * the first WAL segment file since startup, which causes the status being
7375 * wrong under certain abnormal conditions but that doesn't actually harm.
7377 oldestSeg
= XLogGetLastRemovedSegno() + 1;
7379 /* calculate oldest segment by max_wal_size */
7380 XLByteToSeg(currpos
, currSeg
, wal_segment_size
);
7381 keepSegs
= ConvertToXSegs(max_wal_size_mb
, wal_segment_size
) + 1;
7383 if (currSeg
> keepSegs
)
7384 oldestSegMaxWalSize
= currSeg
- keepSegs
;
7386 oldestSegMaxWalSize
= 1;
7388 /* the segment we care about */
7389 XLByteToSeg(targetLSN
, targetSeg
, wal_segment_size
);
7392 * No point in returning reserved or extended status values if the
7393 * targetSeg is known to be lost.
7395 if (targetSeg
>= oldestSlotSeg
)
7397 /* show "reserved" when targetSeg is within max_wal_size */
7398 if (targetSeg
>= oldestSegMaxWalSize
)
7399 return WALAVAIL_RESERVED
;
7401 /* being retained by slots exceeding max_wal_size */
7402 return WALAVAIL_EXTENDED
;
7405 /* WAL segments are no longer retained but haven't been removed yet */
7406 if (targetSeg
>= oldestSeg
)
7407 return WALAVAIL_UNRESERVED
;
7409 /* Definitely lost */
7410 return WALAVAIL_REMOVED
;
7415 * Retreat *logSegNo to the last segment that we need to retain because of
7416 * either wal_keep_size or replication slots.
7418 * This is calculated by subtracting wal_keep_size from the given xlog
7419 * location, recptr and by making sure that that result is below the
7420 * requirement of replication slots. For the latter criterion we do consider
7421 * the effects of max_slot_wal_keep_size: reserve at most that much space back
7424 * Note about replication slots: if this function calculates a value
7425 * that's further ahead than what slots need reserved, then affected
7426 * slots need to be invalidated and this function invoked again.
7427 * XXX it might be a good idea to rewrite this function so that
7428 * invalidation is optionally done here, instead.
7431 KeepLogSeg(XLogRecPtr recptr
, XLogSegNo
*logSegNo
)
7433 XLogSegNo currSegNo
;
7437 XLByteToSeg(recptr
, currSegNo
, wal_segment_size
);
7441 * Calculate how many segments are kept by slots first, adjusting for
7442 * max_slot_wal_keep_size.
7444 keep
= XLogGetReplicationSlotMinimumLSN();
7445 if (keep
!= InvalidXLogRecPtr
)
7447 XLByteToSeg(keep
, segno
, wal_segment_size
);
7449 /* Cap by max_slot_wal_keep_size ... */
7450 if (max_slot_wal_keep_size_mb
>= 0)
7452 uint64 slot_keep_segs
;
7455 ConvertToXSegs(max_slot_wal_keep_size_mb
, wal_segment_size
);
7457 if (currSegNo
- segno
> slot_keep_segs
)
7458 segno
= currSegNo
- slot_keep_segs
;
7462 /* but, keep at least wal_keep_size if that's set */
7463 if (wal_keep_size_mb
> 0)
7467 keep_segs
= ConvertToXSegs(wal_keep_size_mb
, wal_segment_size
);
7468 if (currSegNo
- segno
< keep_segs
)
7470 /* avoid underflow, don't go below 1 */
7471 if (currSegNo
<= keep_segs
)
7474 segno
= currSegNo
- keep_segs
;
7478 /* don't delete WAL segments newer than the calculated segment */
7479 if (segno
< *logSegNo
)
7484 * Write a NEXTOID log record
7487 XLogPutNextOid(Oid nextOid
)
7490 XLogRegisterData((char *) (&nextOid
), sizeof(Oid
));
7491 (void) XLogInsert(RM_XLOG_ID
, XLOG_NEXTOID
);
7494 * We need not flush the NEXTOID record immediately, because any of the
7495 * just-allocated OIDs could only reach disk as part of a tuple insert or
7496 * update that would have its own XLOG record that must follow the NEXTOID
7497 * record. Therefore, the standard buffer LSN interlock applied to those
7498 * records will ensure no such OID reaches disk before the NEXTOID record
7501 * Note, however, that the above statement only covers state "within" the
7502 * database. When we use a generated OID as a file or directory name, we
7503 * are in a sense violating the basic WAL rule, because that filesystem
7504 * change may reach disk before the NEXTOID WAL record does. The impact
7505 * of this is that if a database crash occurs immediately afterward, we
7506 * might after restart re-generate the same OID and find that it conflicts
7507 * with the leftover file or directory. But since for safety's sake we
7508 * always loop until finding a nonconflicting filename, this poses no real
7509 * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
7514 * Write an XLOG SWITCH record.
7516 * Here we just blindly issue an XLogInsert request for the record.
7517 * All the magic happens inside XLogInsert.
7519 * The return value is either the end+1 address of the switch record,
7520 * or the end+1 address of the prior segment if we did not need to
7521 * write a switch record because we are already at segment start.
7524 RequestXLogSwitch(bool mark_unimportant
)
7528 /* XLOG SWITCH has no data */
7531 if (mark_unimportant
)
7532 XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT
);
7533 RecPtr
= XLogInsert(RM_XLOG_ID
, XLOG_SWITCH
);
7539 * Write a RESTORE POINT record
7542 XLogRestorePoint(const char *rpName
)
7545 xl_restore_point xlrec
;
7547 xlrec
.rp_time
= GetCurrentTimestamp();
7548 strlcpy(xlrec
.rp_name
, rpName
, MAXFNAMELEN
);
7551 XLogRegisterData((char *) &xlrec
, sizeof(xl_restore_point
));
7553 RecPtr
= XLogInsert(RM_XLOG_ID
, XLOG_RESTORE_POINT
);
7556 (errmsg("restore point \"%s\" created at %X/%X",
7557 rpName
, LSN_FORMAT_ARGS(RecPtr
))));
7563 * Check if any of the GUC parameters that are critical for hot standby
7564 * have changed, and update the value in pg_control file if necessary.
7567 XLogReportParameters(void)
7569 if (wal_level
!= ControlFile
->wal_level
||
7570 wal_log_hints
!= ControlFile
->wal_log_hints
||
7571 MaxConnections
!= ControlFile
->MaxConnections
||
7572 max_worker_processes
!= ControlFile
->max_worker_processes
||
7573 max_wal_senders
!= ControlFile
->max_wal_senders
||
7574 max_prepared_xacts
!= ControlFile
->max_prepared_xacts
||
7575 max_locks_per_xact
!= ControlFile
->max_locks_per_xact
||
7576 track_commit_timestamp
!= ControlFile
->track_commit_timestamp
)
7579 * The change in number of backend slots doesn't need to be WAL-logged
7580 * if archiving is not enabled, as you can't start archive recovery
7581 * with wal_level=minimal anyway. We don't really care about the
7582 * values in pg_control either if wal_level=minimal, but seems better
7583 * to keep them up-to-date to avoid confusion.
7585 if (wal_level
!= ControlFile
->wal_level
|| XLogIsNeeded())
7587 xl_parameter_change xlrec
;
7590 xlrec
.MaxConnections
= MaxConnections
;
7591 xlrec
.max_worker_processes
= max_worker_processes
;
7592 xlrec
.max_wal_senders
= max_wal_senders
;
7593 xlrec
.max_prepared_xacts
= max_prepared_xacts
;
7594 xlrec
.max_locks_per_xact
= max_locks_per_xact
;
7595 xlrec
.wal_level
= wal_level
;
7596 xlrec
.wal_log_hints
= wal_log_hints
;
7597 xlrec
.track_commit_timestamp
= track_commit_timestamp
;
7600 XLogRegisterData((char *) &xlrec
, sizeof(xlrec
));
7602 recptr
= XLogInsert(RM_XLOG_ID
, XLOG_PARAMETER_CHANGE
);
7606 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
7608 ControlFile
->MaxConnections
= MaxConnections
;
7609 ControlFile
->max_worker_processes
= max_worker_processes
;
7610 ControlFile
->max_wal_senders
= max_wal_senders
;
7611 ControlFile
->max_prepared_xacts
= max_prepared_xacts
;
7612 ControlFile
->max_locks_per_xact
= max_locks_per_xact
;
7613 ControlFile
->wal_level
= wal_level
;
7614 ControlFile
->wal_log_hints
= wal_log_hints
;
7615 ControlFile
->track_commit_timestamp
= track_commit_timestamp
;
7616 UpdateControlFile();
7618 LWLockRelease(ControlFileLock
);
7623 * Update full_page_writes in shared memory, and write an
7624 * XLOG_FPW_CHANGE record if necessary.
7626 * Note: this function assumes there is no other process running
7627 * concurrently that could update it.
7630 UpdateFullPageWrites(void)
7632 XLogCtlInsert
*Insert
= &XLogCtl
->Insert
;
7633 bool recoveryInProgress
;
7636 * Do nothing if full_page_writes has not been changed.
7638 * It's safe to check the shared full_page_writes without the lock,
7639 * because we assume that there is no concurrently running process which
7642 if (fullPageWrites
== Insert
->fullPageWrites
)
7646 * Perform this outside critical section so that the WAL insert
7647 * initialization done by RecoveryInProgress() doesn't trigger an
7648 * assertion failure.
7650 recoveryInProgress
= RecoveryInProgress();
7652 START_CRIT_SECTION();
7655 * It's always safe to take full page images, even when not strictly
7656 * required, but not the other round. So if we're setting full_page_writes
7657 * to true, first set it true and then write the WAL record. If we're
7658 * setting it to false, first write the WAL record and then set the global
7663 WALInsertLockAcquireExclusive();
7664 Insert
->fullPageWrites
= true;
7665 WALInsertLockRelease();
7669 * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
7670 * full_page_writes during archive recovery, if required.
7672 if (XLogStandbyInfoActive() && !recoveryInProgress
)
7675 XLogRegisterData((char *) (&fullPageWrites
), sizeof(bool));
7677 XLogInsert(RM_XLOG_ID
, XLOG_FPW_CHANGE
);
7680 if (!fullPageWrites
)
7682 WALInsertLockAcquireExclusive();
7683 Insert
->fullPageWrites
= false;
7684 WALInsertLockRelease();
7690 * XLOG resource manager's routines
7692 * Definitions of info values are in include/catalog/pg_control.h, though
7693 * not all record types are related to control file updates.
7695 * NOTE: Some XLOG record types that are directly related to WAL recovery
7696 * are handled in xlogrecovery_redo().
7699 xlog_redo(XLogReaderState
*record
)
7701 uint8 info
= XLogRecGetInfo(record
) & ~XLR_INFO_MASK
;
7702 XLogRecPtr lsn
= record
->EndRecPtr
;
7705 * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
7706 * XLOG_FPI_FOR_HINT records.
7708 Assert(info
== XLOG_FPI
|| info
== XLOG_FPI_FOR_HINT
||
7709 !XLogRecHasAnyBlockRefs(record
));
7711 if (info
== XLOG_NEXTOID
)
7716 * We used to try to take the maximum of ShmemVariableCache->nextOid
7717 * and the recorded nextOid, but that fails if the OID counter wraps
7718 * around. Since no OID allocation should be happening during replay
7719 * anyway, better to just believe the record exactly. We still take
7720 * OidGenLock while setting the variable, just in case.
7722 memcpy(&nextOid
, XLogRecGetData(record
), sizeof(Oid
));
7723 LWLockAcquire(OidGenLock
, LW_EXCLUSIVE
);
7724 ShmemVariableCache
->nextOid
= nextOid
;
7725 ShmemVariableCache
->oidCount
= 0;
7726 LWLockRelease(OidGenLock
);
7728 else if (info
== XLOG_CHECKPOINT_SHUTDOWN
)
7730 CheckPoint checkPoint
;
7731 TimeLineID replayTLI
;
7733 memcpy(&checkPoint
, XLogRecGetData(record
), sizeof(CheckPoint
));
7734 /* In a SHUTDOWN checkpoint, believe the counters exactly */
7735 LWLockAcquire(XidGenLock
, LW_EXCLUSIVE
);
7736 ShmemVariableCache
->nextXid
= checkPoint
.nextXid
;
7737 LWLockRelease(XidGenLock
);
7738 LWLockAcquire(OidGenLock
, LW_EXCLUSIVE
);
7739 ShmemVariableCache
->nextOid
= checkPoint
.nextOid
;
7740 ShmemVariableCache
->oidCount
= 0;
7741 LWLockRelease(OidGenLock
);
7742 MultiXactSetNextMXact(checkPoint
.nextMulti
,
7743 checkPoint
.nextMultiOffset
);
7745 MultiXactAdvanceOldest(checkPoint
.oldestMulti
,
7746 checkPoint
.oldestMultiDB
);
7749 * No need to set oldestClogXid here as well; it'll be set when we
7750 * redo an xl_clog_truncate if it changed since initialization.
7752 SetTransactionIdLimit(checkPoint
.oldestXid
, checkPoint
.oldestXidDB
);
7755 * If we see a shutdown checkpoint while waiting for an end-of-backup
7756 * record, the backup was canceled and the end-of-backup record will
7759 if (ArchiveRecoveryRequested
&&
7760 !XLogRecPtrIsInvalid(ControlFile
->backupStartPoint
) &&
7761 XLogRecPtrIsInvalid(ControlFile
->backupEndPoint
))
7763 (errmsg("online backup was canceled, recovery cannot continue")));
7766 * If we see a shutdown checkpoint, we know that nothing was running
7767 * on the primary at this point. So fake-up an empty running-xacts
7768 * record and use that here and now. Recover additional standby state
7769 * for prepared transactions.
7771 if (standbyState
>= STANDBY_INITIALIZED
)
7773 TransactionId
*xids
;
7775 TransactionId oldestActiveXID
;
7776 TransactionId latestCompletedXid
;
7777 RunningTransactionsData running
;
7779 oldestActiveXID
= PrescanPreparedTransactions(&xids
, &nxids
);
7782 * Construct a RunningTransactions snapshot representing a shut
7783 * down server, with only prepared transactions still alive. We're
7784 * never overflowed at this point because all subxids are listed
7785 * with their parent prepared transactions.
7787 running
.xcnt
= nxids
;
7788 running
.subxcnt
= 0;
7789 running
.subxid_overflow
= false;
7790 running
.nextXid
= XidFromFullTransactionId(checkPoint
.nextXid
);
7791 running
.oldestRunningXid
= oldestActiveXID
;
7792 latestCompletedXid
= XidFromFullTransactionId(checkPoint
.nextXid
);
7793 TransactionIdRetreat(latestCompletedXid
);
7794 Assert(TransactionIdIsNormal(latestCompletedXid
));
7795 running
.latestCompletedXid
= latestCompletedXid
;
7796 running
.xids
= xids
;
7798 ProcArrayApplyRecoveryInfo(&running
);
7800 StandbyRecoverPreparedTransactions();
7803 /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
7804 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
7805 ControlFile
->checkPointCopy
.nextXid
= checkPoint
.nextXid
;
7806 LWLockRelease(ControlFileLock
);
7808 /* Update shared-memory copy of checkpoint XID/epoch */
7809 SpinLockAcquire(&XLogCtl
->info_lck
);
7810 XLogCtl
->ckptFullXid
= checkPoint
.nextXid
;
7811 SpinLockRelease(&XLogCtl
->info_lck
);
7814 * We should've already switched to the new TLI before replaying this
7817 (void) GetCurrentReplayRecPtr(&replayTLI
);
7818 if (checkPoint
.ThisTimeLineID
!= replayTLI
)
7820 (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
7821 checkPoint
.ThisTimeLineID
, replayTLI
)));
7823 RecoveryRestartPoint(&checkPoint
, record
);
7825 else if (info
== XLOG_CHECKPOINT_ONLINE
)
7827 CheckPoint checkPoint
;
7828 TimeLineID replayTLI
;
7830 memcpy(&checkPoint
, XLogRecGetData(record
), sizeof(CheckPoint
));
7831 /* In an ONLINE checkpoint, treat the XID counter as a minimum */
7832 LWLockAcquire(XidGenLock
, LW_EXCLUSIVE
);
7833 if (FullTransactionIdPrecedes(ShmemVariableCache
->nextXid
,
7834 checkPoint
.nextXid
))
7835 ShmemVariableCache
->nextXid
= checkPoint
.nextXid
;
7836 LWLockRelease(XidGenLock
);
7839 * We ignore the nextOid counter in an ONLINE checkpoint, preferring
7840 * to track OID assignment through XLOG_NEXTOID records. The nextOid
7841 * counter is from the start of the checkpoint and might well be stale
7842 * compared to later XLOG_NEXTOID records. We could try to take the
7843 * maximum of the nextOid counter and our latest value, but since
7844 * there's no particular guarantee about the speed with which the OID
7845 * counter wraps around, that's a risky thing to do. In any case,
7846 * users of the nextOid counter are required to avoid assignment of
7847 * duplicates, so that a somewhat out-of-date value should be safe.
7850 /* Handle multixact */
7851 MultiXactAdvanceNextMXact(checkPoint
.nextMulti
,
7852 checkPoint
.nextMultiOffset
);
7855 * NB: This may perform multixact truncation when replaying WAL
7856 * generated by an older primary.
7858 MultiXactAdvanceOldest(checkPoint
.oldestMulti
,
7859 checkPoint
.oldestMultiDB
);
7860 if (TransactionIdPrecedes(ShmemVariableCache
->oldestXid
,
7861 checkPoint
.oldestXid
))
7862 SetTransactionIdLimit(checkPoint
.oldestXid
,
7863 checkPoint
.oldestXidDB
);
7864 /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
7865 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
7866 ControlFile
->checkPointCopy
.nextXid
= checkPoint
.nextXid
;
7867 LWLockRelease(ControlFileLock
);
7869 /* Update shared-memory copy of checkpoint XID/epoch */
7870 SpinLockAcquire(&XLogCtl
->info_lck
);
7871 XLogCtl
->ckptFullXid
= checkPoint
.nextXid
;
7872 SpinLockRelease(&XLogCtl
->info_lck
);
7874 /* TLI should not change in an on-line checkpoint */
7875 (void) GetCurrentReplayRecPtr(&replayTLI
);
7876 if (checkPoint
.ThisTimeLineID
!= replayTLI
)
7878 (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
7879 checkPoint
.ThisTimeLineID
, replayTLI
)));
7881 RecoveryRestartPoint(&checkPoint
, record
);
7883 else if (info
== XLOG_OVERWRITE_CONTRECORD
)
7885 /* nothing to do here, handled in xlogrecovery_redo() */
7887 else if (info
== XLOG_END_OF_RECOVERY
)
7889 xl_end_of_recovery xlrec
;
7890 TimeLineID replayTLI
;
7892 memcpy(&xlrec
, XLogRecGetData(record
), sizeof(xl_end_of_recovery
));
7895 * For Hot Standby, we could treat this like a Shutdown Checkpoint,
7896 * but this case is rarer and harder to test, so the benefit doesn't
7897 * outweigh the potential extra cost of maintenance.
7901 * We should've already switched to the new TLI before replaying this
7904 (void) GetCurrentReplayRecPtr(&replayTLI
);
7905 if (xlrec
.ThisTimeLineID
!= replayTLI
)
7907 (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
7908 xlrec
.ThisTimeLineID
, replayTLI
)));
7910 else if (info
== XLOG_NOOP
)
7912 /* nothing to do here */
7914 else if (info
== XLOG_SWITCH
)
7916 /* nothing to do here */
7918 else if (info
== XLOG_RESTORE_POINT
)
7920 /* nothing to do here, handled in xlogrecovery.c */
7922 else if (info
== XLOG_FPI
|| info
== XLOG_FPI_FOR_HINT
)
7925 * XLOG_FPI records contain nothing else but one or more block
7926 * references. Every block reference must include a full-page image
7927 * even if full_page_writes was disabled when the record was generated
7928 * - otherwise there would be no point in this record.
7930 * XLOG_FPI_FOR_HINT records are generated when a page needs to be
7931 * WAL-logged because of a hint bit update. They are only generated
7932 * when checksums and/or wal_log_hints are enabled. They may include
7933 * no full-page images if full_page_writes was disabled when they were
7934 * generated. In this case there is nothing to do here.
7936 * No recovery conflicts are generated by these generic records - if a
7937 * resource manager needs to generate conflicts, it has to define a
7938 * separate WAL record type and redo routine.
7940 for (uint8 block_id
= 0; block_id
<= XLogRecMaxBlockId(record
); block_id
++)
7944 if (!XLogRecHasBlockImage(record
, block_id
))
7946 if (info
== XLOG_FPI
)
7947 elog(ERROR
, "XLOG_FPI record did not contain a full-page image");
7951 if (XLogReadBufferForRedo(record
, block_id
, &buffer
) != BLK_RESTORED
)
7952 elog(ERROR
, "unexpected XLogReadBufferForRedo result when restoring backup block");
7953 UnlockReleaseBuffer(buffer
);
7956 else if (info
== XLOG_BACKUP_END
)
7958 /* nothing to do here, handled in xlogrecovery_redo() */
7960 else if (info
== XLOG_PARAMETER_CHANGE
)
7962 xl_parameter_change xlrec
;
7964 /* Update our copy of the parameters in pg_control */
7965 memcpy(&xlrec
, XLogRecGetData(record
), sizeof(xl_parameter_change
));
7967 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
7968 ControlFile
->MaxConnections
= xlrec
.MaxConnections
;
7969 ControlFile
->max_worker_processes
= xlrec
.max_worker_processes
;
7970 ControlFile
->max_wal_senders
= xlrec
.max_wal_senders
;
7971 ControlFile
->max_prepared_xacts
= xlrec
.max_prepared_xacts
;
7972 ControlFile
->max_locks_per_xact
= xlrec
.max_locks_per_xact
;
7973 ControlFile
->wal_level
= xlrec
.wal_level
;
7974 ControlFile
->wal_log_hints
= xlrec
.wal_log_hints
;
7977 * Update minRecoveryPoint to ensure that if recovery is aborted, we
7978 * recover back up to this point before allowing hot standby again.
7979 * This is important if the max_* settings are decreased, to ensure
7980 * you don't run queries against the WAL preceding the change. The
7981 * local copies cannot be updated as long as crash recovery is
7982 * happening and we expect all the WAL to be replayed.
7984 if (InArchiveRecovery
)
7986 LocalMinRecoveryPoint
= ControlFile
->minRecoveryPoint
;
7987 LocalMinRecoveryPointTLI
= ControlFile
->minRecoveryPointTLI
;
7989 if (LocalMinRecoveryPoint
!= InvalidXLogRecPtr
&& LocalMinRecoveryPoint
< lsn
)
7991 TimeLineID replayTLI
;
7993 (void) GetCurrentReplayRecPtr(&replayTLI
);
7994 ControlFile
->minRecoveryPoint
= lsn
;
7995 ControlFile
->minRecoveryPointTLI
= replayTLI
;
7998 CommitTsParameterChange(xlrec
.track_commit_timestamp
,
7999 ControlFile
->track_commit_timestamp
);
8000 ControlFile
->track_commit_timestamp
= xlrec
.track_commit_timestamp
;
8002 UpdateControlFile();
8003 LWLockRelease(ControlFileLock
);
8005 /* Check to see if any parameter change gives a problem on recovery */
8006 CheckRequiredParameterValues();
8008 else if (info
== XLOG_FPW_CHANGE
)
8012 memcpy(&fpw
, XLogRecGetData(record
), sizeof(bool));
8015 * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
8016 * do_pg_backup_start() and do_pg_backup_stop() can check whether
8017 * full_page_writes has been disabled during online backup.
8021 SpinLockAcquire(&XLogCtl
->info_lck
);
8022 if (XLogCtl
->lastFpwDisableRecPtr
< record
->ReadRecPtr
)
8023 XLogCtl
->lastFpwDisableRecPtr
= record
->ReadRecPtr
;
8024 SpinLockRelease(&XLogCtl
->info_lck
);
8027 /* Keep track of full_page_writes */
8028 lastFullPageWrites
= fpw
;
8033 * Return the (possible) sync flag used for opening a file, depending on the
8034 * value of the GUC wal_sync_method.
8037 get_sync_bit(int method
)
8039 int o_direct_flag
= 0;
8041 /* If fsync is disabled, never open in sync mode */
8046 * Optimize writes by bypassing kernel cache with O_DIRECT when using
8047 * O_SYNC and O_DSYNC. But only if archiving and streaming are disabled,
8048 * otherwise the archive command or walsender process will read the WAL
8049 * soon after writing it, which is guaranteed to cause a physical read if
8050 * we bypassed the kernel cache. We also skip the
8051 * posix_fadvise(POSIX_FADV_DONTNEED) call in XLogFileClose() for the same
8054 * Never use O_DIRECT in walreceiver process for similar reasons; the WAL
8055 * written by walreceiver is normally read by the startup process soon
8056 * after it's written. Also, walreceiver performs unaligned writes, which
8057 * don't work with O_DIRECT, so it is required for correctness too.
8059 if (!XLogIsNeeded() && !AmWalReceiverProcess())
8060 o_direct_flag
= PG_O_DIRECT
;
8065 * enum values for all sync options are defined even if they are
8066 * not supported on the current platform. But if not, they are
8067 * not included in the enum option array, and therefore will never
8070 case SYNC_METHOD_FSYNC
:
8071 case SYNC_METHOD_FSYNC_WRITETHROUGH
:
8072 case SYNC_METHOD_FDATASYNC
:
8075 case SYNC_METHOD_OPEN
:
8076 return O_SYNC
| o_direct_flag
;
8079 case SYNC_METHOD_OPEN_DSYNC
:
8080 return O_DSYNC
| o_direct_flag
;
8083 /* can't happen (unless we are out of sync with option array) */
8084 elog(ERROR
, "unrecognized wal_sync_method: %d", method
);
8085 return 0; /* silence warning */
8093 assign_xlog_sync_method(int new_sync_method
, void *extra
)
8095 if (sync_method
!= new_sync_method
)
8098 * To ensure that no blocks escape unsynced, force an fsync on the
8099 * currently open log segment (if any). Also, if the open flag is
8100 * changing, close the log file so it will be reopened (with new flag
8103 if (openLogFile
>= 0)
8105 pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN
);
8106 if (pg_fsync(openLogFile
) != 0)
8108 char xlogfname
[MAXFNAMELEN
];
8112 XLogFileName(xlogfname
, openLogTLI
, openLogSegNo
,
8116 (errcode_for_file_access(),
8117 errmsg("could not fsync file \"%s\": %m", xlogfname
)));
8120 pgstat_report_wait_end();
8121 if (get_sync_bit(sync_method
) != get_sync_bit(new_sync_method
))
8129 * Issue appropriate kind of fsync (if any) for an XLOG output file.
8131 * 'fd' is a file descriptor for the XLOG file to be fsync'd.
8132 * 'segno' is for error reporting purposes.
8135 issue_xlog_fsync(int fd
, XLogSegNo segno
, TimeLineID tli
)
8143 * Quick exit if fsync is disabled or write() has already synced the WAL
8147 sync_method
== SYNC_METHOD_OPEN
||
8148 sync_method
== SYNC_METHOD_OPEN_DSYNC
)
8151 /* Measure I/O timing to sync the WAL file */
8152 if (track_wal_io_timing
)
8153 INSTR_TIME_SET_CURRENT(start
);
8155 pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC
);
8156 switch (sync_method
)
8158 case SYNC_METHOD_FSYNC
:
8159 if (pg_fsync_no_writethrough(fd
) != 0)
8160 msg
= _("could not fsync file \"%s\": %m");
8162 #ifdef HAVE_FSYNC_WRITETHROUGH
8163 case SYNC_METHOD_FSYNC_WRITETHROUGH
:
8164 if (pg_fsync_writethrough(fd
) != 0)
8165 msg
= _("could not fsync write-through file \"%s\": %m");
8168 case SYNC_METHOD_FDATASYNC
:
8169 if (pg_fdatasync(fd
) != 0)
8170 msg
= _("could not fdatasync file \"%s\": %m");
8172 case SYNC_METHOD_OPEN
:
8173 case SYNC_METHOD_OPEN_DSYNC
:
8178 elog(PANIC
, "unrecognized wal_sync_method: %d", sync_method
);
8182 /* PANIC if failed to fsync */
8185 char xlogfname
[MAXFNAMELEN
];
8186 int save_errno
= errno
;
8188 XLogFileName(xlogfname
, tli
, segno
, wal_segment_size
);
8191 (errcode_for_file_access(),
8192 errmsg(msg
, xlogfname
)));
8195 pgstat_report_wait_end();
8198 * Increment the I/O timing and the number of times WAL files were synced.
8200 if (track_wal_io_timing
)
8202 instr_time duration
;
8204 INSTR_TIME_SET_CURRENT(duration
);
8205 INSTR_TIME_SUBTRACT(duration
, start
);
8206 PendingWalStats
.wal_sync_time
+= INSTR_TIME_GET_MICROSEC(duration
);
8209 PendingWalStats
.wal_sync
++;
8213 * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
8214 * function. It creates the necessary starting checkpoint and constructs the
8215 * backup state and tablespace map.
8217 * Input parameters are "state" (the backup state), "fast" (if true, we do
8218 * the checkpoint in immediate mode to make it faster), and "tablespaces"
8219 * (if non-NULL, indicates a list of tablespaceinfo structs describing the
8220 * cluster's tablespaces.).
8222 * The tablespace map contents are appended to passed-in parameter
8223 * tablespace_map and the caller is responsible for including it in the backup
8224 * archive as 'tablespace_map'. The tablespace_map file is required mainly for
8225 * tar format in windows as native windows utilities are not able to create
8226 * symlinks while extracting files from tar. However for consistency and
8227 * platform-independence, we do it the same way everywhere.
8229 * It fills in "state" with the information required for the backup, such
8230 * as the minimum WAL location that must be present to restore from this
8231 * backup (starttli) and the corresponding timeline ID (starttli).
8233 * Every successfully started backup must be stopped by calling
8234 * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
8235 * backups active at the same time.
8237 * It is the responsibility of the caller of this function to verify the
8238 * permissions of the calling user!
8241 do_pg_backup_start(const char *backupidstr
, bool fast
, List
**tablespaces
,
8242 BackupState
*state
, StringInfo tblspcmapfile
)
8244 bool backup_started_in_recovery
;
8246 Assert(state
!= NULL
);
8247 backup_started_in_recovery
= RecoveryInProgress();
8250 * During recovery, we don't need to check WAL level. Because, if WAL
8251 * level is not sufficient, it's impossible to get here during recovery.
8253 if (!backup_started_in_recovery
&& !XLogIsNeeded())
8255 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE
),
8256 errmsg("WAL level not sufficient for making an online backup"),
8257 errhint("wal_level must be set to \"replica\" or \"logical\" at server start.")));
8259 if (strlen(backupidstr
) > MAXPGPATH
)
8261 (errcode(ERRCODE_INVALID_PARAMETER_VALUE
),
8262 errmsg("backup label too long (max %d bytes)",
8265 memcpy(state
->name
, backupidstr
, strlen(backupidstr
));
8268 * Mark backup active in shared memory. We must do full-page WAL writes
8269 * during an on-line backup even if not doing so at other times, because
8270 * it's quite possible for the backup dump to obtain a "torn" (partially
8271 * written) copy of a database page if it reads the page concurrently with
8272 * our write to the same page. This can be fixed as long as the first
8273 * write to the page in the WAL sequence is a full-page write. Hence, we
8274 * increment runningBackups then force a CHECKPOINT, to ensure there are
8275 * no dirty pages in shared memory that might get dumped while the backup
8276 * is in progress without having a corresponding WAL record. (Once the
8277 * backup is complete, we need not force full-page writes anymore, since
8278 * we expect that any pages not modified during the backup interval must
8279 * have been correctly captured by the backup.)
8281 * Note that forcing full-page writes has no effect during an online
8282 * backup from the standby.
8284 * We must hold all the insertion locks to change the value of
8285 * runningBackups, to ensure adequate interlocking against
8286 * XLogInsertRecord().
8288 WALInsertLockAcquireExclusive();
8289 XLogCtl
->Insert
.runningBackups
++;
8290 WALInsertLockRelease();
8293 * Ensure we decrement runningBackups if we fail below. NB -- for this to
8294 * work correctly, it is critical that sessionBackupState is only updated
8295 * after this block is over.
8297 PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup
, DatumGetBool(true));
8299 bool gotUniqueStartpoint
= false;
8306 * Force an XLOG file switch before the checkpoint, to ensure that the
8307 * WAL segment the checkpoint is written to doesn't contain pages with
8308 * old timeline IDs. That would otherwise happen if you called
8309 * pg_backup_start() right after restoring from a PITR archive: the
8310 * first WAL segment containing the startup checkpoint has pages in
8311 * the beginning with the old timeline ID. That can cause trouble at
8312 * recovery: we won't have a history file covering the old timeline if
8313 * pg_wal directory was not included in the base backup and the WAL
8314 * archive was cleared too before starting the backup.
8316 * This also ensures that we have emitted a WAL page header that has
8317 * XLP_BKP_REMOVABLE off before we emit the checkpoint record.
8318 * Therefore, if a WAL archiver (such as pglesslog) is trying to
8319 * compress out removable backup blocks, it won't remove any that
8320 * occur after this point.
8322 * During recovery, we skip forcing XLOG file switch, which means that
8323 * the backup taken during recovery is not available for the special
8324 * recovery case described above.
8326 if (!backup_started_in_recovery
)
8327 RequestXLogSwitch(false);
8334 * Force a CHECKPOINT. Aside from being necessary to prevent torn
8335 * page problems, this guarantees that two successive backup runs
8336 * will have different checkpoint positions and hence different
8337 * history file names, even if nothing happened in between.
8339 * During recovery, establish a restartpoint if possible. We use
8340 * the last restartpoint as the backup starting checkpoint. This
8341 * means that two successive backup runs can have same checkpoint
8344 * Since the fact that we are executing do_pg_backup_start()
8345 * during recovery means that checkpointer is running, we can use
8346 * RequestCheckpoint() to establish a restartpoint.
8348 * We use CHECKPOINT_IMMEDIATE only if requested by user (via
8349 * passing fast = true). Otherwise this can take awhile.
8351 RequestCheckpoint(CHECKPOINT_FORCE
| CHECKPOINT_WAIT
|
8352 (fast
? CHECKPOINT_IMMEDIATE
: 0));
8355 * Now we need to fetch the checkpoint record location, and also
8356 * its REDO pointer. The oldest point in WAL that would be needed
8357 * to restore starting from the checkpoint is precisely the REDO
8360 LWLockAcquire(ControlFileLock
, LW_SHARED
);
8361 state
->checkpointloc
= ControlFile
->checkPoint
;
8362 state
->startpoint
= ControlFile
->checkPointCopy
.redo
;
8363 state
->starttli
= ControlFile
->checkPointCopy
.ThisTimeLineID
;
8364 checkpointfpw
= ControlFile
->checkPointCopy
.fullPageWrites
;
8365 LWLockRelease(ControlFileLock
);
8367 if (backup_started_in_recovery
)
8372 * Check to see if all WAL replayed during online backup
8373 * (i.e., since last restartpoint used as backup starting
8374 * checkpoint) contain full-page writes.
8376 SpinLockAcquire(&XLogCtl
->info_lck
);
8377 recptr
= XLogCtl
->lastFpwDisableRecPtr
;
8378 SpinLockRelease(&XLogCtl
->info_lck
);
8380 if (!checkpointfpw
|| state
->startpoint
<= recptr
)
8382 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE
),
8383 errmsg("WAL generated with full_page_writes=off was replayed "
8384 "since last restartpoint"),
8385 errhint("This means that the backup being taken on the standby "
8386 "is corrupt and should not be used. "
8387 "Enable full_page_writes and run CHECKPOINT on the primary, "
8388 "and then try an online backup again.")));
8391 * During recovery, since we don't use the end-of-backup WAL
8392 * record and don't write the backup history file, the
8393 * starting WAL location doesn't need to be unique. This means
8394 * that two base backups started at the same time might use
8395 * the same checkpoint as starting locations.
8397 gotUniqueStartpoint
= true;
8401 * If two base backups are started at the same time (in WAL sender
8402 * processes), we need to make sure that they use different
8403 * checkpoints as starting locations, because we use the starting
8404 * WAL location as a unique identifier for the base backup in the
8405 * end-of-backup WAL record and when we write the backup history
8406 * file. Perhaps it would be better generate a separate unique ID
8407 * for each backup instead of forcing another checkpoint, but
8408 * taking a checkpoint right after another is not that expensive
8409 * either because only few buffers have been dirtied yet.
8411 WALInsertLockAcquireExclusive();
8412 if (XLogCtl
->Insert
.lastBackupStart
< state
->startpoint
)
8414 XLogCtl
->Insert
.lastBackupStart
= state
->startpoint
;
8415 gotUniqueStartpoint
= true;
8417 WALInsertLockRelease();
8418 } while (!gotUniqueStartpoint
);
8421 * Construct tablespace_map file.
8423 datadirpathlen
= strlen(DataDir
);
8425 /* Collect information about all tablespaces */
8426 tblspcdir
= AllocateDir("pg_tblspc");
8427 while ((de
= ReadDir(tblspcdir
, "pg_tblspc")) != NULL
)
8429 char fullpath
[MAXPGPATH
+ 10];
8430 char linkpath
[MAXPGPATH
];
8431 char *relpath
= NULL
;
8433 StringInfoData escapedpath
;
8436 /* Skip anything that doesn't look like a tablespace */
8437 if (strspn(de
->d_name
, "0123456789") != strlen(de
->d_name
))
8440 snprintf(fullpath
, sizeof(fullpath
), "pg_tblspc/%s", de
->d_name
);
8443 * Skip anything that isn't a symlink/junction. For testing only,
8444 * we sometimes use allow_in_place_tablespaces to create
8445 * directories directly under pg_tblspc, which would fail below.
8447 if (get_dirent_type(fullpath
, de
, false, ERROR
) != PGFILETYPE_LNK
)
8450 rllen
= readlink(fullpath
, linkpath
, sizeof(linkpath
));
8454 (errmsg("could not read symbolic link \"%s\": %m",
8458 else if (rllen
>= sizeof(linkpath
))
8461 (errmsg("symbolic link \"%s\" target is too long",
8465 linkpath
[rllen
] = '\0';
8468 * Build a backslash-escaped version of the link path to include
8469 * in the tablespace map file.
8471 initStringInfo(&escapedpath
);
8472 for (s
= linkpath
; *s
; s
++)
8474 if (*s
== '\n' || *s
== '\r' || *s
== '\\')
8475 appendStringInfoChar(&escapedpath
, '\\');
8476 appendStringInfoChar(&escapedpath
, *s
);
8480 * Relpath holds the relative path of the tablespace directory
8481 * when it's located within PGDATA, or NULL if it's located
8484 if (rllen
> datadirpathlen
&&
8485 strncmp(linkpath
, DataDir
, datadirpathlen
) == 0 &&
8486 IS_DIR_SEP(linkpath
[datadirpathlen
]))
8487 relpath
= linkpath
+ datadirpathlen
+ 1;
8489 ti
= palloc(sizeof(tablespaceinfo
));
8490 ti
->oid
= pstrdup(de
->d_name
);
8491 ti
->path
= pstrdup(linkpath
);
8492 ti
->rpath
= relpath
? pstrdup(relpath
) : NULL
;
8496 *tablespaces
= lappend(*tablespaces
, ti
);
8498 appendStringInfo(tblspcmapfile
, "%s %s\n",
8499 ti
->oid
, escapedpath
.data
);
8501 pfree(escapedpath
.data
);
8505 state
->starttime
= (pg_time_t
) time(NULL
);
8507 PG_END_ENSURE_ERROR_CLEANUP(do_pg_abort_backup
, DatumGetBool(true));
8509 state
->started_in_recovery
= backup_started_in_recovery
;
8512 * Mark that the start phase has correctly finished for the backup.
8514 sessionBackupState
= SESSION_BACKUP_RUNNING
;
8518 * Utility routine to fetch the session-level status of a backup running.
8521 get_backup_status(void)
8523 return sessionBackupState
;
8529 * Utility function called at the end of an online backup. It creates history
8530 * file (if required), resets sessionBackupState and so on. It can optionally
8531 * wait for WAL segments to be archived.
8533 * "state" is filled with the information necessary to restore from this
8534 * backup with its stop LSN (stoppoint), its timeline ID (stoptli), etc.
8536 * It is the responsibility of the caller of this function to verify the
8537 * permissions of the calling user!
8540 do_pg_backup_stop(BackupState
*state
, bool waitforarchive
)
8542 bool backup_stopped_in_recovery
= false;
8543 char histfilepath
[MAXPGPATH
];
8544 char lastxlogfilename
[MAXFNAMELEN
];
8545 char histfilename
[MAXFNAMELEN
];
8546 XLogSegNo _logSegNo
;
8548 int seconds_before_warning
;
8550 bool reported_waiting
= false;
8552 Assert(state
!= NULL
);
8554 backup_stopped_in_recovery
= RecoveryInProgress();
8557 * During recovery, we don't need to check WAL level. Because, if WAL
8558 * level is not sufficient, it's impossible to get here during recovery.
8560 if (!backup_stopped_in_recovery
&& !XLogIsNeeded())
8562 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE
),
8563 errmsg("WAL level not sufficient for making an online backup"),
8564 errhint("wal_level must be set to \"replica\" or \"logical\" at server start.")));
8567 * OK to update backup counter and session-level lock.
8569 * Note that CHECK_FOR_INTERRUPTS() must not occur while updating them,
8570 * otherwise they can be updated inconsistently, which might cause
8571 * do_pg_abort_backup() to fail.
8573 WALInsertLockAcquireExclusive();
8576 * It is expected that each do_pg_backup_start() call is matched by
8577 * exactly one do_pg_backup_stop() call.
8579 Assert(XLogCtl
->Insert
.runningBackups
> 0);
8580 XLogCtl
->Insert
.runningBackups
--;
8583 * Clean up session-level lock.
8585 * You might think that WALInsertLockRelease() can be called before
8586 * cleaning up session-level lock because session-level lock doesn't need
8587 * to be protected with WAL insertion lock. But since
8588 * CHECK_FOR_INTERRUPTS() can occur in it, session-level lock must be
8589 * cleaned up before it.
8591 sessionBackupState
= SESSION_BACKUP_NONE
;
8593 WALInsertLockRelease();
8596 * If we are taking an online backup from the standby, we confirm that the
8597 * standby has not been promoted during the backup.
8599 if (state
->started_in_recovery
&& !backup_stopped_in_recovery
)
8601 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE
),
8602 errmsg("the standby was promoted during online backup"),
8603 errhint("This means that the backup being taken is corrupt "
8604 "and should not be used. "
8605 "Try taking another online backup.")));
8608 * During recovery, we don't write an end-of-backup record. We assume that
8609 * pg_control was backed up last and its minimum recovery point can be
8610 * available as the backup end location. Since we don't have an
8611 * end-of-backup record, we use the pg_control value to check whether
8612 * we've reached the end of backup when starting recovery from this
8613 * backup. We have no way of checking if pg_control wasn't backed up last
8616 * We don't force a switch to new WAL file but it is still possible to
8617 * wait for all the required files to be archived if waitforarchive is
8618 * true. This is okay if we use the backup to start a standby and fetch
8619 * the missing WAL using streaming replication. But in the case of an
8620 * archive recovery, a user should set waitforarchive to true and wait for
8621 * them to be archived to ensure that all the required files are
8624 * We return the current minimum recovery point as the backup end
8625 * location. Note that it can be greater than the exact backup end
8626 * location if the minimum recovery point is updated after the backup of
8627 * pg_control. This is harmless for current uses.
8629 * XXX currently a backup history file is for informational and debug
8630 * purposes only. It's not essential for an online backup. Furthermore,
8631 * even if it's created, it will not be archived during recovery because
8632 * an archiver is not invoked. So it doesn't seem worthwhile to write a
8633 * backup history file during recovery.
8635 if (backup_stopped_in_recovery
)
8640 * Check to see if all WAL replayed during online backup contain
8643 SpinLockAcquire(&XLogCtl
->info_lck
);
8644 recptr
= XLogCtl
->lastFpwDisableRecPtr
;
8645 SpinLockRelease(&XLogCtl
->info_lck
);
8647 if (state
->startpoint
<= recptr
)
8649 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE
),
8650 errmsg("WAL generated with full_page_writes=off was replayed "
8651 "during online backup"),
8652 errhint("This means that the backup being taken on the standby "
8653 "is corrupt and should not be used. "
8654 "Enable full_page_writes and run CHECKPOINT on the primary, "
8655 "and then try an online backup again.")));
8658 LWLockAcquire(ControlFileLock
, LW_SHARED
);
8659 state
->stoppoint
= ControlFile
->minRecoveryPoint
;
8660 state
->stoptli
= ControlFile
->minRecoveryPointTLI
;
8661 LWLockRelease(ControlFileLock
);
8668 * Write the backup-end xlog record
8671 XLogRegisterData((char *) (&state
->startpoint
),
8672 sizeof(state
->startpoint
));
8673 state
->stoppoint
= XLogInsert(RM_XLOG_ID
, XLOG_BACKUP_END
);
8676 * Given that we're not in recovery, InsertTimeLineID is set and can't
8677 * change, so we can read it without a lock.
8679 state
->stoptli
= XLogCtl
->InsertTimeLineID
;
8682 * Force a switch to a new xlog segment file, so that the backup is
8683 * valid as soon as archiver moves out the current segment file.
8685 RequestXLogSwitch(false);
8687 state
->stoptime
= (pg_time_t
) time(NULL
);
8690 * Write the backup history file
8692 XLByteToSeg(state
->startpoint
, _logSegNo
, wal_segment_size
);
8693 BackupHistoryFilePath(histfilepath
, state
->stoptli
, _logSegNo
,
8694 state
->startpoint
, wal_segment_size
);
8695 fp
= AllocateFile(histfilepath
, "w");
8698 (errcode_for_file_access(),
8699 errmsg("could not create file \"%s\": %m",
8702 /* Build and save the contents of the backup history file */
8703 history_file
= build_backup_content(state
, true);
8704 fprintf(fp
, "%s", history_file
);
8705 pfree(history_file
);
8707 if (fflush(fp
) || ferror(fp
) || FreeFile(fp
))
8709 (errcode_for_file_access(),
8710 errmsg("could not write file \"%s\": %m",
8714 * Clean out any no-longer-needed history files. As a side effect,
8715 * this will post a .ready file for the newly created history file,
8716 * notifying the archiver that history file may be archived
8719 CleanupBackupHistory();
8723 * If archiving is enabled, wait for all the required WAL files to be
8724 * archived before returning. If archiving isn't enabled, the required WAL
8725 * needs to be transported via streaming replication (hopefully with
8726 * wal_keep_size set high enough), or some more exotic mechanism like
8727 * polling and copying files from pg_wal with script. We have no knowledge
8728 * of those mechanisms, so it's up to the user to ensure that he gets all
8731 * We wait until both the last WAL file filled during backup and the
8732 * history file have been archived, and assume that the alphabetic sorting
8733 * property of the WAL files ensures any earlier WAL files are safely
8736 * We wait forever, since archive_command is supposed to work and we
8737 * assume the admin wanted his backup to work completely. If you don't
8738 * wish to wait, then either waitforarchive should be passed in as false,
8739 * or you can set statement_timeout. Also, some notices are issued to
8740 * clue in anyone who might be doing this interactively.
8743 if (waitforarchive
&&
8744 ((!backup_stopped_in_recovery
&& XLogArchivingActive()) ||
8745 (backup_stopped_in_recovery
&& XLogArchivingAlways())))
8747 XLByteToPrevSeg(state
->stoppoint
, _logSegNo
, wal_segment_size
);
8748 XLogFileName(lastxlogfilename
, state
->stoptli
, _logSegNo
,
8751 XLByteToSeg(state
->startpoint
, _logSegNo
, wal_segment_size
);
8752 BackupHistoryFileName(histfilename
, state
->stoptli
, _logSegNo
,
8753 state
->startpoint
, wal_segment_size
);
8755 seconds_before_warning
= 60;
8758 while (XLogArchiveIsBusy(lastxlogfilename
) ||
8759 XLogArchiveIsBusy(histfilename
))
8761 CHECK_FOR_INTERRUPTS();
8763 if (!reported_waiting
&& waits
> 5)
8766 (errmsg("base backup done, waiting for required WAL segments to be archived")));
8767 reported_waiting
= true;
8770 (void) WaitLatch(MyLatch
,
8771 WL_LATCH_SET
| WL_TIMEOUT
| WL_EXIT_ON_PM_DEATH
,
8773 WAIT_EVENT_BACKUP_WAIT_WAL_ARCHIVE
);
8774 ResetLatch(MyLatch
);
8776 if (++waits
>= seconds_before_warning
)
8778 seconds_before_warning
*= 2; /* This wraps in >10 years... */
8780 (errmsg("still waiting for all required WAL segments to be archived (%d seconds elapsed)",
8782 errhint("Check that your archive_command is executing properly. "
8783 "You can safely cancel this backup, "
8784 "but the database backup will not be usable without all the WAL segments.")));
8789 (errmsg("all required WAL segments have been archived")));
8791 else if (waitforarchive
)
8793 (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
8798 * do_pg_abort_backup: abort a running backup
8800 * This does just the most basic steps of do_pg_backup_stop(), by taking the
8801 * system out of backup mode, thus making it a lot more safe to call from
8804 * 'arg' indicates that it's being called during backup setup; so
8805 * sessionBackupState has not been modified yet, but runningBackups has
8806 * already been incremented. When it's false, then it's invoked as a
8807 * before_shmem_exit handler, and therefore we must not change state
8808 * unless sessionBackupState indicates that a backup is actually running.
8810 * NB: This gets used as a PG_ENSURE_ERROR_CLEANUP callback and
8811 * before_shmem_exit handler, hence the odd-looking signature.
8814 do_pg_abort_backup(int code
, Datum arg
)
8816 bool during_backup_start
= DatumGetBool(arg
);
8818 /* If called during backup start, there shouldn't be one already running */
8819 Assert(!during_backup_start
|| sessionBackupState
== SESSION_BACKUP_NONE
);
8821 if (during_backup_start
|| sessionBackupState
!= SESSION_BACKUP_NONE
)
8823 WALInsertLockAcquireExclusive();
8824 Assert(XLogCtl
->Insert
.runningBackups
> 0);
8825 XLogCtl
->Insert
.runningBackups
--;
8827 sessionBackupState
= SESSION_BACKUP_NONE
;
8828 WALInsertLockRelease();
8830 if (!during_backup_start
)
8832 errmsg("aborting backup due to backend exiting before pg_backup_stop was called"));
8837 * Register a handler that will warn about unterminated backups at end of
8838 * session, unless this has already been done.
8841 register_persistent_abort_backup_handler(void)
8843 static bool already_done
= false;
8847 before_shmem_exit(do_pg_abort_backup
, DatumGetBool(false));
8848 already_done
= true;
8852 * Get latest WAL insert pointer
8855 GetXLogInsertRecPtr(void)
8857 XLogCtlInsert
*Insert
= &XLogCtl
->Insert
;
8858 uint64 current_bytepos
;
8860 SpinLockAcquire(&Insert
->insertpos_lck
);
8861 current_bytepos
= Insert
->CurrBytePos
;
8862 SpinLockRelease(&Insert
->insertpos_lck
);
8864 return XLogBytePosToRecPtr(current_bytepos
);
8868 * Get latest WAL write pointer
8871 GetXLogWriteRecPtr(void)
8873 SpinLockAcquire(&XLogCtl
->info_lck
);
8874 LogwrtResult
= XLogCtl
->LogwrtResult
;
8875 SpinLockRelease(&XLogCtl
->info_lck
);
8877 return LogwrtResult
.Write
;
8881 * Returns the redo pointer of the last checkpoint or restartpoint. This is
8882 * the oldest point in WAL that we still need, if we have to restart recovery.
8885 GetOldestRestartPoint(XLogRecPtr
*oldrecptr
, TimeLineID
*oldtli
)
8887 LWLockAcquire(ControlFileLock
, LW_SHARED
);
8888 *oldrecptr
= ControlFile
->checkPointCopy
.redo
;
8889 *oldtli
= ControlFile
->checkPointCopy
.ThisTimeLineID
;
8890 LWLockRelease(ControlFileLock
);
8894 * Returns the WAL file name for the last checkpoint or restartpoint. This is
8895 * the oldest WAL file that we still need if we have to restart recovery.
8898 GetOldestRestartPointFileName(char *fname
)
8900 XLogRecPtr restartRedoPtr
;
8901 TimeLineID restartTli
;
8902 XLogSegNo restartSegNo
;
8904 GetOldestRestartPoint(&restartRedoPtr
, &restartTli
);
8905 XLByteToSeg(restartRedoPtr
, restartSegNo
, wal_segment_size
);
8906 XLogFileName(fname
, restartTli
, restartSegNo
, wal_segment_size
);
8909 /* Thin wrapper around ShutdownWalRcv(). */
8911 XLogShutdownWalRcv(void)
8915 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
8916 XLogCtl
->InstallXLogFileSegmentActive
= false;
8917 LWLockRelease(ControlFileLock
);
8920 /* Enable WAL file recycling and preallocation. */
8922 SetInstallXLogFileSegmentActive(void)
8924 LWLockAcquire(ControlFileLock
, LW_EXCLUSIVE
);
8925 XLogCtl
->InstallXLogFileSegmentActive
= true;
8926 LWLockRelease(ControlFileLock
);
8930 IsInstallXLogFileSegmentActive(void)
8934 LWLockAcquire(ControlFileLock
, LW_SHARED
);
8935 result
= XLogCtl
->InstallXLogFileSegmentActive
;
8936 LWLockRelease(ControlFileLock
);
8942 * Update the WalWriterSleeping flag.
8945 SetWalWriterSleeping(bool sleeping
)
8947 SpinLockAcquire(&XLogCtl
->info_lck
);
8948 XLogCtl
->WalWriterSleeping
= sleeping
;
8949 SpinLockRelease(&XLogCtl
->info_lck
);