1 /*-------------------------------------------------------------------------
4 * PostgreSQL transaction log manager
7 * Portions Copyright (c) 1996-2013, PostgreSQL Global Development Group
8 * Portions Copyright (c) 1994, Regents of the University of California
10 * src/backend/access/transam/xlog.c
12 *-------------------------------------------------------------------------
24 #include "access/clog.h"
25 #include "access/multixact.h"
26 #include "access/subtrans.h"
27 #include "access/timeline.h"
28 #include "access/transam.h"
29 #include "access/tuptoaster.h"
30 #include "access/twophase.h"
31 #include "access/xact.h"
32 #include "access/xlog_internal.h"
33 #include "access/xlogreader.h"
34 #include "access/xlogutils.h"
35 #include "catalog/catversion.h"
36 #include "catalog/pg_control.h"
37 #include "catalog/pg_database.h"
38 #include "miscadmin.h"
40 #include "postmaster/bgwriter.h"
41 #include "postmaster/startup.h"
42 #include "replication/walreceiver.h"
43 #include "replication/walsender.h"
44 #include "storage/barrier.h"
45 #include "storage/bufmgr.h"
46 #include "storage/fd.h"
47 #include "storage/ipc.h"
48 #include "storage/latch.h"
49 #include "storage/pmsignal.h"
50 #include "storage/predicate.h"
51 #include "storage/proc.h"
52 #include "storage/procarray.h"
53 #include "storage/reinit.h"
54 #include "storage/smgr.h"
55 #include "storage/spin.h"
56 #include "utils/builtins.h"
57 #include "utils/guc.h"
58 #include "utils/ps_status.h"
59 #include "utils/relmapper.h"
60 #include "utils/snapmgr.h"
61 #include "utils/timestamp.h"
64 extern uint32 bootstrap_data_checksum_version;
66 /* File path names (all relative to $PGDATA) */
67 #define RECOVERY_COMMAND_FILE "recovery.conf"
68 #define RECOVERY_COMMAND_DONE "recovery.done"
69 #define PROMOTE_SIGNAL_FILE "promote"
70 #define FALLBACK_PROMOTE_SIGNAL_FILE "fallback_promote"
73 /* User-settable parameters */
74 int CheckPointSegments = 3;
75 int wal_keep_segments = 0;
77 int XLogArchiveTimeout = 0;
78 bool XLogArchiveMode = false;
79 char *XLogArchiveCommand = NULL;
80 bool EnableHotStandby = false;
81 bool fullPageWrites = true;
82 bool log_checkpoints = false;
83 int sync_method = DEFAULT_SYNC_METHOD;
84 int wal_level = WAL_LEVEL_MINIMAL;
85 int CommitDelay = 0; /* precommit delay in microseconds */
86 int CommitSiblings = 5; /* # concurrent xacts needed to sleep */
87 int num_xloginsert_slots = 8;
90 bool XLOG_DEBUG = false;
94 * XLOGfileslop is the maximum number of preallocated future XLOG segments.
95 * When we are done with an old XLOG segment file, we will recycle it as a
96 * future XLOG segment as long as there aren't already XLOGfileslop future
97 * segments; else we'll delete it. This could be made a separate GUC
98 * variable, but at present I think it's sufficient to hardwire it as
99 * 2*CheckPointSegments+1. Under normal conditions, a checkpoint will free
100 * no more than 2*CheckPointSegments log segments, and we want to recycle all
101 * of them; the +1 allows boundary cases to happen without wasting a
102 * delete/create-segment cycle.
104 #define XLOGfileslop (2*CheckPointSegments + 1)
110 const struct config_enum_entry sync_method_options[] = {
111 {"fsync", SYNC_METHOD_FSYNC, false},
112 #ifdef HAVE_FSYNC_WRITETHROUGH
113 {"fsync_writethrough", SYNC_METHOD_FSYNC_WRITETHROUGH, false},
115 #ifdef HAVE_FDATASYNC
116 {"fdatasync", SYNC_METHOD_FDATASYNC, false},
118 #ifdef OPEN_SYNC_FLAG
119 {"open_sync", SYNC_METHOD_OPEN, false},
121 #ifdef OPEN_DATASYNC_FLAG
122 {"open_datasync", SYNC_METHOD_OPEN_DSYNC, false},
128 * Statistics for current checkpoint are collected in this global struct.
129 * Because only the background writer or a stand-alone backend can perform
130 * checkpoints, this will be unused in normal backends.
132 CheckpointStatsData CheckpointStats;
135 * ThisTimeLineID will be same in all backends --- it identifies current
136 * WAL timeline for the database system.
138 TimeLineID ThisTimeLineID = 0;
141 * Are we doing recovery from XLOG?
143 * This is only ever true in the startup process; it should be read as meaning
144 * "this process is replaying WAL records", rather than "the system is in
145 * recovery mode". It should be examined primarily by functions that need
146 * to act differently when called from a WAL redo function (e.g., to skip WAL
147 * logging). To check whether the system is in recovery regardless of which
148 * process you're running in, use RecoveryInProgress() but only after shared
149 * memory startup and lock initialization.
151 bool InRecovery = false;
153 /* Are we in Hot Standby mode? Only valid in startup process, see xlog.h */
154 HotStandbyState standbyState = STANDBY_DISABLED;
156 static XLogRecPtr LastRec;
158 /* Local copy of WalRcv->receivedUpto */
159 static XLogRecPtr receivedUpto = 0;
160 static TimeLineID receiveTLI = 0;
163 * During recovery, lastFullPageWrites keeps track of full_page_writes that
164 * the replayed WAL records indicate. It's initialized with full_page_writes
165 * that the recovery starting checkpoint record indicates, and then updated
166 * each time XLOG_FPW_CHANGE record is replayed.
168 static bool lastFullPageWrites;
171 * Local copy of SharedRecoveryInProgress variable. True actually means "not
172 * known, need to check the shared state".
174 static bool LocalRecoveryInProgress = true;
177 * Local copy of SharedHotStandbyActive variable. False actually means "not
178 * known, need to check the shared state".
180 static bool LocalHotStandbyActive = false;
183 * Local state for XLogInsertAllowed():
184 * 1: unconditionally allowed to insert XLOG
185 * 0: unconditionally not allowed to insert XLOG
186 * -1: must check RecoveryInProgress(); disallow until it is false
187 * Most processes start with -1 and transition to 1 after seeing that recovery
188 * is not in progress. But we can also force the value for special cases.
189 * The coding in XLogInsertAllowed() depends on the first two of these states
190 * being numerically the same as bool true and false.
192 static int LocalXLogInsertAllowed = -1;
195 * When ArchiveRecoveryRequested is set, archive recovery was requested,
196 * ie. recovery.conf file was present. When InArchiveRecovery is set, we are
197 * currently recovering using offline XLOG archives. These variables are only
198 * valid in the startup process.
200 * When ArchiveRecoveryRequested is true, but InArchiveRecovery is false, we're
201 * currently performing crash recovery using only XLOG files in pg_xlog, but
202 * will switch to using offline XLOG archives as soon as we reach the end of
205 bool ArchiveRecoveryRequested = false;
206 bool InArchiveRecovery = false;
208 /* Was the last xlog file restored from archive, or local? */
209 static bool restoredFromArchive = false;
211 /* options taken from recovery.conf for archive recovery */
212 char *recoveryRestoreCommand = NULL;
213 static char *recoveryEndCommand = NULL;
214 static char *archiveCleanupCommand = NULL;
215 static RecoveryTargetType recoveryTarget = RECOVERY_TARGET_UNSET;
216 static bool recoveryTargetInclusive = true;
217 static bool recoveryPauseAtTarget = true;
218 static TransactionId recoveryTargetXid;
219 static TimestampTz recoveryTargetTime;
220 static char *recoveryTargetName;
222 /* options taken from recovery.conf for XLOG streaming */
223 static bool StandbyModeRequested = false;
224 static char *PrimaryConnInfo = NULL;
225 static char *TriggerFile = NULL;
227 /* are we currently in standby mode? */
228 bool StandbyMode = false;
230 /* whether request for fast promotion has been made yet */
231 static bool fast_promote = false;
233 /* if recoveryStopsHere returns true, it saves actual stop xid/time/name here */
234 static TransactionId recoveryStopXid;
235 static TimestampTz recoveryStopTime;
236 static char recoveryStopName[MAXFNAMELEN];
237 static bool recoveryStopAfter;
240 * During normal operation, the only timeline we care about is ThisTimeLineID.
241 * During recovery, however, things are more complicated. To simplify life
242 * for rmgr code, we keep ThisTimeLineID set to the "current" timeline as we
243 * scan through the WAL history (that is, it is the line that was active when
244 * the currently-scanned WAL record was generated). We also need these
247 * recoveryTargetTLI: the desired timeline that we want to end in.
249 * recoveryTargetIsLatest: was the requested target timeline 'latest'?
251 * expectedTLEs: a list of TimeLineHistoryEntries for recoveryTargetTLI and the timelines of
252 * its known parents, newest first (so recoveryTargetTLI is always the
253 * first list member). Only these TLIs are expected to be seen in the WAL
254 * segments we read, and indeed only these TLIs will be considered as
255 * candidate WAL files to open at all.
257 * curFileTLI: the TLI appearing in the name of the current input WAL file.
258 * (This is not necessarily the same as ThisTimeLineID, because we could
259 * be scanning data that was copied from an ancestor timeline when the current
260 * file was created.) During a sequential scan we do not allow this value
263 static TimeLineID recoveryTargetTLI;
264 static bool recoveryTargetIsLatest = false;
265 static List *expectedTLEs;
266 static TimeLineID curFileTLI;
269 * ProcLastRecPtr points to the start of the last XLOG record inserted by the
270 * current backend. It is updated for all inserts. XactLastRecEnd points to
271 * end+1 of the last record, and is reset when we end a top-level transaction,
272 * or start a new one; so it can be used to tell if the current transaction has
273 * created any XLOG records.
275 static XLogRecPtr ProcLastRecPtr = InvalidXLogRecPtr;
277 XLogRecPtr XactLastRecEnd = InvalidXLogRecPtr;
280 * RedoRecPtr is this backend's local copy of the REDO record pointer
281 * (which is almost but not quite the same as a pointer to the most recent
282 * CHECKPOINT record). We update this from the shared-memory copy,
283 * XLogCtl->Insert.RedoRecPtr, whenever we can safely do so (ie, when we
284 * hold an insertion slot). See XLogInsert for details. We are also allowed
285 * to update from XLogCtl->RedoRecPtr if we hold the info_lck;
286 * see GetRedoRecPtr. A freshly spawned backend obtains the value during
289 static XLogRecPtr RedoRecPtr;
292 * RedoStartLSN points to the checkpoint's REDO location which is specified
293 * in a backup label file, backup history file or control file. In standby
294 * mode, XLOG streaming usually starts from the position where an invalid
295 * record was found. But if we fail to read even the initial checkpoint
296 * record, we use the REDO location instead of the checkpoint location as
297 * the start position of XLOG streaming. Otherwise we would have to jump
298 * backwards to the REDO location after reading the checkpoint record,
299 * because the REDO record can precede the checkpoint record.
301 static XLogRecPtr RedoStartLSN = InvalidXLogRecPtr;
304 * Shared-memory data structures for XLOG control
306 * LogwrtRqst indicates a byte position that we need to write and/or fsync
307 * the log up to (all records before that point must be written or fsynced).
308 * LogwrtResult indicates the byte positions we have already written/fsynced.
309 * These structs are identical but are declared separately to indicate their
310 * slightly different functions.
312 * To read XLogCtl->LogwrtResult, you must hold either info_lck or
313 * WALWriteLock. To update it, you need to hold both locks. The point of
314 * this arrangement is that the value can be examined by code that already
315 * holds WALWriteLock without needing to grab info_lck as well. In addition
316 * to the shared variable, each backend has a private copy of LogwrtResult,
317 * which is updated when convenient.
319 * The request bookkeeping is simpler: there is a shared XLogCtl->LogwrtRqst
320 * (protected by info_lck), but we don't need to cache any copies of it.
322 * info_lck is only held long enough to read/update the protected variables,
323 * so it's a plain spinlock. The other locks are held longer (potentially
324 * over I/O operations), so we use LWLocks for them. These locks are:
326 * WALBufMappingLock: must be held to replace a page in the WAL buffer cache.
327 * It is only held while initializing and changing the mapping. If the
328 * contents of the buffer being replaced haven't been written yet, the mapping
329 * lock is released while the write is done, and reacquired afterwards.
331 * WALWriteLock: must be held to write WAL buffers to disk (XLogWrite or
334 * ControlFileLock: must be held to read/update control file or create
337 * CheckpointLock: must be held to do a checkpoint or restartpoint (ensures
338 * only one checkpointer at a time; currently, with all checkpoints done by
339 * the checkpointer, this is just pro forma).
344 typedef struct XLogwrtRqst
346 XLogRecPtr Write; /* last byte + 1 to write out */
347 XLogRecPtr Flush; /* last byte + 1 to flush */
350 typedef struct XLogwrtResult
352 XLogRecPtr Write; /* last byte + 1 written out */
353 XLogRecPtr Flush; /* last byte + 1 flushed */
358 * A slot for inserting to the WAL. This is similar to an LWLock, the main
359 * difference is that there is an extra xlogInsertingAt field that is protected
360 * by the same mutex. Unlike an LWLock, a slot can only be acquired in
363 * The xlogInsertingAt field is used to advertise to other processes how far
364 * the slot owner has progressed in inserting the record. When a backend
365 * acquires a slot, it initializes xlogInsertingAt to 1, because it doesn't
366 * yet know where it's going to insert the record. That's conservative
367 * but correct; the new insertion is certainly going to go to a byte position
368 * greater than 1. If another backend needs to flush the WAL, it will have to
369 * wait for the new insertion. xlogInsertingAt is updated after finishing the
370 * insert or when crossing a page boundary, which will wake up anyone waiting
371 * for it, whether the wait was necessary in the first place or not.
373 * A process can wait on a slot in two modes: LW_EXCLUSIVE or
374 * LW_WAIT_UNTIL_FREE. LW_EXCLUSIVE works like in an lwlock; when the slot is
375 * released, the first LW_EXCLUSIVE waiter in the queue is woken up. Processes
376 * waiting in LW_WAIT_UNTIL_FREE mode are woken up whenever the slot is
377 * released, or xlogInsertingAt is updated. In other words, a process in
378 * LW_WAIT_UNTIL_FREE mode is woken up whenever the inserter makes any progress
379 * copying the record in place. LW_WAIT_UNTIL_FREE waiters are always added to
380 * the front of the queue, while LW_EXCLUSIVE waiters are appended to the end.
382 * To join the wait queue, a process must set MyProc->lwWaitMode to the mode
383 * it wants to wait in, MyProc->lwWaiting to true, and link MyProc to the head
384 * or tail of the wait queue. The same mechanism is used to wait on an LWLock,
385 * see lwlock.c for details.
389 slock_t mutex; /* protects the below fields */
390 XLogRecPtr xlogInsertingAt; /* insert has completed up to this point */
392 PGPROC *owner; /* for debugging purposes */
394 bool releaseOK; /* T if ok to release waiters */
395 char exclusive; /* # of exclusive holders (0 or 1) */
396 PGPROC *head; /* head of list of waiting PGPROCs */
397 PGPROC *tail; /* tail of list of waiting PGPROCs */
398 /* tail is undefined when head is NULL */
402 * All the slots are allocated as an array in shared memory. We force the
403 * array stride to be a power of 2, which saves a few cycles in indexing, but
404 * more importantly also ensures that individual slots don't cross cache line
405 * boundaries. (Of course, we have to also ensure that the array start
406 * address is suitably aligned.)
408 typedef union XLogInsertSlotPadded
411 char pad[CACHE_LINE_SIZE];
412 } XLogInsertSlotPadded;
415 * Shared state data for XLogInsert.
417 typedef struct XLogCtlInsert
419 slock_t insertpos_lck; /* protects CurrBytePos and PrevBytePos */
422 * CurrBytePos is the end of reserved WAL. The next record will be inserted
423 * at that position. PrevBytePos is the start position of the previously
424 * inserted (or rather, reserved) record - it is copied to the the prev-
425 * link of the next record. These are stored as "usable byte positions"
426 * rather than XLogRecPtrs (see XLogBytePosToRecPtr()).
432 * Make sure the above heavily-contended spinlock and byte positions are
433 * on their own cache line. In particular, the RedoRecPtr and full page
434 * write variables below should be on a different cache line. They are
435 * read on every WAL insertion, but updated rarely, and we don't want
436 * those reads to steal the cache line containing Curr/PrevBytePos.
438 char pad[CACHE_LINE_SIZE];
441 * fullPageWrites is the master copy used by all backends to determine
442 * whether to write full-page to WAL, instead of using process-local one.
443 * This is required because, when full_page_writes is changed by SIGHUP,
444 * we must WAL-log it before it actually affects WAL-logging by backends.
445 * Checkpointer sets at startup or after SIGHUP.
447 * To read these fields, you must hold an insertion slot. To modify them,
448 * you must hold ALL the slots.
450 XLogRecPtr RedoRecPtr; /* current redo point for insertions */
451 bool forcePageWrites; /* forcing full-page writes for PITR? */
455 * exclusiveBackup is true if a backup started with pg_start_backup() is
456 * in progress, and nonExclusiveBackups is a counter indicating the number
457 * of streaming base backups currently in progress. forcePageWrites is set
458 * to true when either of these is non-zero. lastBackupStart is the latest
459 * checkpoint redo location used as a starting point for an online backup.
461 bool exclusiveBackup;
462 int nonExclusiveBackups;
463 XLogRecPtr lastBackupStart;
465 /* insertion slots, see XLogInsertSlot struct above for details */
466 XLogInsertSlotPadded *insertSlots;
470 * Total shared-memory state for XLOG.
472 typedef struct XLogCtlData
474 XLogCtlInsert Insert;
476 /* Protected by info_lck: */
477 XLogwrtRqst LogwrtRqst;
478 XLogRecPtr RedoRecPtr; /* a recent copy of Insert->RedoRecPtr */
479 uint32 ckptXidEpoch; /* nextXID & epoch of latest checkpoint */
480 TransactionId ckptXid;
481 XLogRecPtr asyncXactLSN; /* LSN of newest async commit/abort */
482 XLogSegNo lastRemovedSegNo; /* latest removed/recycled XLOG
485 /* Fake LSN counter, for unlogged relations. Protected by ulsn_lck. */
486 XLogRecPtr unloggedLSN;
489 /* Time of last xlog segment switch. Protected by WALWriteLock. */
490 pg_time_t lastSegSwitchTime;
493 * Protected by info_lck and WALWriteLock (you must hold either lock to
494 * read it, but both to update)
496 XLogwrtResult LogwrtResult;
499 * Latest initialized page in the cache (last byte position + 1).
501 * To change the identity of a buffer (and InitializedUpTo), you need to
502 * hold WALBufMappingLock. To change the identity of a buffer that's still
503 * dirty, the old page needs to be written out first, and for that you
504 * need WALWriteLock, and you need to ensure that there are no in-progress
505 * insertions to the page by calling WaitXLogInsertionsToFinish().
507 XLogRecPtr InitializedUpTo;
510 * These values do not change after startup, although the pointed-to pages
511 * and xlblocks values certainly do. xlblock values are protected by
514 char *pages; /* buffers for unwritten XLOG pages */
515 XLogRecPtr *xlblocks; /* 1st byte ptr-s + XLOG_BLCKSZ */
516 int XLogCacheBlck; /* highest allocated xlog buffer index */
519 * Shared copy of ThisTimeLineID. Does not change after end-of-recovery.
520 * If we created a new timeline when the system was started up,
521 * PrevTimeLineID is the old timeline's ID that we forked off from.
522 * Otherwise it's equal to ThisTimeLineID.
524 TimeLineID ThisTimeLineID;
525 TimeLineID PrevTimeLineID;
528 * archiveCleanupCommand is read from recovery.conf but needs to be in
529 * shared memory so that the checkpointer process can access it.
531 char archiveCleanupCommand[MAXPGPATH];
534 * SharedRecoveryInProgress indicates if we're still in crash or archive
535 * recovery. Protected by info_lck.
537 bool SharedRecoveryInProgress;
540 * SharedHotStandbyActive indicates if we're still in crash or archive
541 * recovery. Protected by info_lck.
543 bool SharedHotStandbyActive;
546 * WalWriterSleeping indicates whether the WAL writer is currently in
547 * low-power mode (and hence should be nudged if an async commit occurs).
548 * Protected by info_lck.
550 bool WalWriterSleeping;
553 * recoveryWakeupLatch is used to wake up the startup process to continue
554 * WAL replay, if it is waiting for WAL to arrive or failover trigger file
557 Latch recoveryWakeupLatch;
560 * During recovery, we keep a copy of the latest checkpoint record here.
561 * Used by the background writer when it wants to create a restartpoint.
563 * Protected by info_lck.
565 XLogRecPtr lastCheckPointRecPtr;
566 CheckPoint lastCheckPoint;
569 * lastReplayedEndRecPtr points to end+1 of the last record successfully
570 * replayed. When we're currently replaying a record, ie. in a redo
571 * function, replayEndRecPtr points to the end+1 of the record being
572 * replayed, otherwise it's equal to lastReplayedEndRecPtr.
574 XLogRecPtr lastReplayedEndRecPtr;
575 TimeLineID lastReplayedTLI;
576 XLogRecPtr replayEndRecPtr;
577 TimeLineID replayEndTLI;
578 /* timestamp of last COMMIT/ABORT record replayed (or being replayed) */
579 TimestampTz recoveryLastXTime;
580 /* current effective recovery target timeline */
581 TimeLineID RecoveryTargetTLI;
584 * timestamp of when we started replaying the current chunk of WAL data,
585 * only relevant for replication or archive recovery
587 TimestampTz currentChunkStartTime;
588 /* Are we requested to pause recovery? */
592 * lastFpwDisableRecPtr points to the start of the last replayed
593 * XLOG_FPW_CHANGE record that instructs full_page_writes is disabled.
595 XLogRecPtr lastFpwDisableRecPtr;
597 slock_t info_lck; /* locks shared variables shown above */
600 static XLogCtlData *XLogCtl = NULL;
603 * We maintain an image of pg_control in shared memory.
605 static ControlFileData *ControlFile = NULL;
608 * Calculate the amount of space left on the page after 'endptr'. Beware
609 * multiple evaluation!
611 #define INSERT_FREESPACE(endptr) \
612 (((endptr) % XLOG_BLCKSZ == 0) ? 0 : (XLOG_BLCKSZ - (endptr) % XLOG_BLCKSZ))
614 /* Macro to advance to next buffer index. */
615 #define NextBufIdx(idx) \
616 (((idx) == XLogCtl->XLogCacheBlck) ? 0 : ((idx) + 1))
619 * XLogRecPtrToBufIdx returns the index of the WAL buffer that holds, or
620 * would hold if it was in cache, the page containing 'recptr'.
622 #define XLogRecPtrToBufIdx(recptr) \
623 (((recptr) / XLOG_BLCKSZ) % (XLogCtl->XLogCacheBlck + 1))
626 * These are the number of bytes in a WAL page and segment usable for WAL data.
628 #define UsableBytesInPage (XLOG_BLCKSZ - SizeOfXLogShortPHD)
629 #define UsableBytesInSegment ((XLOG_SEG_SIZE / XLOG_BLCKSZ) * UsableBytesInPage - (SizeOfXLogLongPHD - SizeOfXLogShortPHD))
632 * Private, possibly out-of-date copy of shared LogwrtResult.
633 * See discussion above.
635 static XLogwrtResult LogwrtResult = {0, 0};
638 * Codes indicating where we got a WAL file from during recovery, or where
639 * to attempt to get one.
643 XLOG_FROM_ANY = 0, /* request to read WAL from any source */
644 XLOG_FROM_ARCHIVE, /* restored using restore_command */
645 XLOG_FROM_PG_XLOG, /* existing file in pg_xlog */
646 XLOG_FROM_STREAM, /* streamed from master */
649 /* human-readable names for XLogSources, for debugging output */
650 static const char *xlogSourceNames[] = {"any", "archive", "pg_xlog", "stream"};
653 * openLogFile is -1 or a kernel FD for an open log file segment.
654 * When it's open, openLogOff is the current seek offset in the file.
655 * openLogSegNo identifies the segment. These variables are only
656 * used to write the XLOG, and so will normally refer to the active segment.
658 static int openLogFile = -1;
659 static XLogSegNo openLogSegNo = 0;
660 static uint32 openLogOff = 0;
663 * These variables are used similarly to the ones above, but for reading
664 * the XLOG. Note, however, that readOff generally represents the offset
665 * of the page just read, not the seek position of the FD itself, which
666 * will be just past that page. readLen indicates how much of the current
667 * page has been read into readBuf, and readSource indicates where we got
668 * the currently open file from.
670 static int readFile = -1;
671 static XLogSegNo readSegNo = 0;
672 static uint32 readOff = 0;
673 static uint32 readLen = 0;
674 static XLogSource readSource = 0; /* XLOG_FROM_* code */
677 * Keeps track of which source we're currently reading from. This is
678 * different from readSource in that this is always set, even when we don't
679 * currently have a WAL file open. If lastSourceFailed is set, our last
680 * attempt to read from currentSource failed, and we should try another source
683 static XLogSource currentSource = 0; /* XLOG_FROM_* code */
684 static bool lastSourceFailed = false;
686 typedef struct XLogPageReadPrivate
689 bool fetching_ckpt; /* are we fetching a checkpoint record? */
691 } XLogPageReadPrivate;
694 * These variables track when we last obtained some WAL data to process,
695 * and where we got it from. (XLogReceiptSource is initially the same as
696 * readSource, but readSource gets reset to zero when we don't have data
697 * to process right now. It is also different from currentSource, which
698 * also changes when we try to read from a source and fail, while
699 * XLogReceiptSource tracks where we last successfully read some WAL.)
701 static TimestampTz XLogReceiptTime = 0;
702 static XLogSource XLogReceiptSource = 0; /* XLOG_FROM_* code */
704 /* State information for XLOG reading */
705 static XLogRecPtr ReadRecPtr; /* start of last record read */
706 static XLogRecPtr EndRecPtr; /* end+1 of last record read */
708 static XLogRecPtr minRecoveryPoint; /* local copy of
709 * ControlFile->minRecoveryPoint */
710 static TimeLineID minRecoveryPointTLI;
711 static bool updateMinRecoveryPoint = true;
714 * Have we reached a consistent database state? In crash recovery, we have
715 * to replay all the WAL, so reachedConsistency is never set. During archive
716 * recovery, the database is consistent once minRecoveryPoint is reached.
718 bool reachedConsistency = false;
720 static bool InRedo = false;
722 /* Have we launched bgwriter during recovery? */
723 static bool bgwriterLaunched = false;
725 /* For WALInsertSlotAcquire/Release functions */
726 static int MySlotNo = 0;
727 static bool holdingAllSlots = false;
729 static void readRecoveryCommandFile(void);
730 static void exitArchiveRecovery(TimeLineID endTLI, XLogSegNo endLogSegNo);
731 static bool recoveryStopsHere(XLogRecord *record, bool *includeThis);
732 static void recoveryPausesHere(void);
733 static void SetLatestXTime(TimestampTz xtime);
734 static void SetCurrentChunkStartTime(TimestampTz xtime);
735 static void CheckRequiredParameterValues(void);
736 static void XLogReportParameters(void);
737 static void checkTimeLineSwitch(XLogRecPtr lsn, TimeLineID newTLI,
739 static void LocalSetXLogInsertAllowed(void);
740 static void CreateEndOfRecoveryRecord(void);
741 static void CheckPointGuts(XLogRecPtr checkPointRedo, int flags);
742 static void KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo);
744 static bool XLogCheckBuffer(XLogRecData *rdata, bool holdsExclusiveLock,
745 XLogRecPtr *lsn, BkpBlock *bkpb);
746 static Buffer RestoreBackupBlockContents(XLogRecPtr lsn, BkpBlock bkpb,
747 char *blk, bool get_cleanup_lock, bool keep_buffer);
748 static void AdvanceXLInsertBuffer(XLogRecPtr upto, bool opportunistic);
749 static bool XLogCheckpointNeeded(XLogSegNo new_segno);
750 static void XLogWrite(XLogwrtRqst WriteRqst, bool flexible);
751 static bool InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
752 bool find_free, int *max_advance,
754 static int XLogFileRead(XLogSegNo segno, int emode, TimeLineID tli,
755 int source, bool notexistOk);
756 static int XLogFileReadAnyTLI(XLogSegNo segno, int emode, int source);
757 static int XLogPageRead(XLogReaderState *xlogreader, XLogRecPtr targetPagePtr,
758 int reqLen, XLogRecPtr targetRecPtr, char *readBuf,
759 TimeLineID *readTLI);
760 static bool WaitForWALToBecomeAvailable(XLogRecPtr RecPtr, bool randAccess,
761 bool fetching_ckpt, XLogRecPtr tliRecPtr);
762 static int emode_for_corrupt_record(int emode, XLogRecPtr RecPtr);
763 static void XLogFileClose(void);
764 static void PreallocXlogFiles(XLogRecPtr endptr);
765 static void RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr endptr);
766 static void UpdateLastRemovedPtr(char *filename);
767 static void ValidateXLOGDirectoryStructure(void);
768 static void CleanupBackupHistory(void);
769 static void UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force);
770 static XLogRecord *ReadRecord(XLogReaderState *xlogreader, XLogRecPtr RecPtr,
771 int emode, bool fetching_ckpt);
772 static void CheckRecoveryConsistency(void);
773 static XLogRecord *ReadCheckpointRecord(XLogReaderState *xlogreader,
774 XLogRecPtr RecPtr, int whichChkpti, bool report);
775 static bool rescanLatestTimeLine(void);
776 static void WriteControlFile(void);
777 static void ReadControlFile(void);
778 static char *str_time(pg_time_t tnow);
779 static bool CheckForStandbyTrigger(void);
782 static void xlog_outrec(StringInfo buf, XLogRecord *record);
784 static void pg_start_backup_callback(int code, Datum arg);
785 static bool read_backup_label(XLogRecPtr *checkPointLoc,
786 bool *backupEndRequired, bool *backupFromStandby);
787 static void rm_redo_error_callback(void *arg);
788 static int get_sync_bit(int method);
790 static void CopyXLogRecordToWAL(int write_len, bool isLogSwitch,
792 XLogRecPtr StartPos, XLogRecPtr EndPos);
793 static void ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos,
794 XLogRecPtr *EndPos, XLogRecPtr *PrevPtr);
795 static bool ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos,
796 XLogRecPtr *PrevPtr);
797 static XLogRecPtr WaitXLogInsertionsToFinish(XLogRecPtr upto);
798 static void WakeupWaiters(XLogRecPtr EndPos);
799 static char *GetXLogBuffer(XLogRecPtr ptr);
800 static XLogRecPtr XLogBytePosToRecPtr(uint64 bytepos);
801 static XLogRecPtr XLogBytePosToEndRecPtr(uint64 bytepos);
802 static uint64 XLogRecPtrToBytePos(XLogRecPtr ptr);
804 static void WALInsertSlotAcquire(bool exclusive);
805 static void WALInsertSlotAcquireOne(int slotno);
806 static void WALInsertSlotRelease(void);
807 static void WALInsertSlotReleaseOne(int slotno);
810 * Insert an XLOG record having the specified RMID and info bytes,
811 * with the body of the record being the data chunk(s) described by
812 * the rdata chain (see xlog.h for notes about rdata).
814 * Returns XLOG pointer to end of record (beginning of next record).
815 * This can be used as LSN for data pages affected by the logged action.
816 * (LSN is the XLOG point up to which the XLOG must be flushed to disk
817 * before the data page can be written out. This implements the basic
818 * WAL rule "write the log before the data".)
820 * NB: this routine feels free to scribble on the XLogRecData structs,
821 * though not on the data they reference. This is OK since the XLogRecData
822 * structs are always just temporaries in the calling code.
825 XLogInsert(RmgrId rmid, uint8 info, XLogRecData *rdata)
827 XLogCtlInsert *Insert = &XLogCtl->Insert;
829 XLogRecData *rdt_lastnormal;
830 Buffer dtbuf[XLR_MAX_BKP_BLOCKS];
831 bool dtbuf_bkp[XLR_MAX_BKP_BLOCKS];
832 BkpBlock dtbuf_xlg[XLR_MAX_BKP_BLOCKS];
833 XLogRecPtr dtbuf_lsn[XLR_MAX_BKP_BLOCKS];
834 XLogRecData dtbuf_rdt1[XLR_MAX_BKP_BLOCKS];
835 XLogRecData dtbuf_rdt2[XLR_MAX_BKP_BLOCKS];
836 XLogRecData dtbuf_rdt3[XLR_MAX_BKP_BLOCKS];
843 bool isLogSwitch = (rmid == RM_XLOG_ID && info == XLOG_SWITCH);
845 uint8 info_orig = info;
846 static XLogRecord *rechdr;
852 rechdr = malloc(SizeOfXLogRecord);
854 elog(ERROR, "out of memory");
855 MemSet(rechdr, 0, SizeOfXLogRecord);
858 /* cross-check on whether we should be here or not */
859 if (!XLogInsertAllowed())
860 elog(ERROR, "cannot make new WAL entries during recovery");
862 /* info's high bits are reserved for use by me */
863 if (info & XLR_INFO_MASK)
864 elog(PANIC, "invalid xlog info mask %02X", info);
866 TRACE_POSTGRESQL_XLOG_INSERT(rmid, info);
869 * In bootstrap mode, we don't actually log anything but XLOG resources;
870 * return a phony record pointer.
872 if (IsBootstrapProcessingMode() && rmid != RM_XLOG_ID)
874 EndPos = SizeOfXLogLongPHD; /* start of 1st chkpt record */
879 * Here we scan the rdata chain, to determine which buffers must be backed
882 * We may have to loop back to here if a race condition is detected below.
883 * We could prevent the race by doing all this work while holding an
884 * insertion slot, but it seems better to avoid doing CRC calculations
887 * We add entries for backup blocks to the chain, so that they don't need
888 * any special treatment in the critical section where the chunks are
889 * copied into the WAL buffers. Those entries have to be unlinked from the
890 * chain if we have to loop back here.
893 for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
895 dtbuf[i] = InvalidBuffer;
896 dtbuf_bkp[i] = false;
900 * Decide if we need to do full-page writes in this XLOG record: true if
901 * full_page_writes is on or we have a PITR request for it. Since we
902 * don't yet have an insertion slot, fullPageWrites and forcePageWrites
903 * could change under us, but we'll recheck them once we have a slot.
905 doPageWrites = Insert->fullPageWrites || Insert->forcePageWrites;
910 if (rdt->buffer == InvalidBuffer)
912 /* Simple data, just include it */
917 /* Find info for buffer */
918 for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
920 if (rdt->buffer == dtbuf[i])
922 /* Buffer already referenced by earlier chain item */
932 if (dtbuf[i] == InvalidBuffer)
934 /* OK, put it in this slot */
935 dtbuf[i] = rdt->buffer;
936 if (doPageWrites && XLogCheckBuffer(rdt, true,
937 &(dtbuf_lsn[i]), &(dtbuf_xlg[i])))
948 if (i >= XLR_MAX_BKP_BLOCKS)
949 elog(PANIC, "can backup at most %d blocks per xlog record",
952 /* Break out of loop when rdt points to last chain item */
953 if (rdt->next == NULL)
959 * NOTE: We disallow len == 0 because it provides a useful bit of extra
960 * error checking in ReadRecord. This means that all callers of
961 * XLogInsert must supply at least some not-in-a-buffer data. However, we
962 * make an exception for XLOG SWITCH records because we don't want them to
963 * ever cross a segment boundary.
965 if (len == 0 && !isLogSwitch)
966 elog(PANIC, "invalid xlog record length %u", len);
969 * Make additional rdata chain entries for the backup blocks, so that we
970 * don't need to special-case them in the write loop. This modifies the
971 * original rdata chain, but we keep a pointer to the last regular entry,
972 * rdt_lastnormal, so that we can undo this if we have to loop back to the
975 * At the exit of this loop, write_len includes the backup block data.
977 * Also set the appropriate info bits to show which buffers were backed
978 * up. The XLR_BKP_BLOCK(N) bit corresponds to the N'th distinct buffer
979 * value (ignoring InvalidBuffer) appearing in the rdata chain.
981 rdt_lastnormal = rdt;
983 for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
991 info |= XLR_BKP_BLOCK(i);
993 bkpb = &(dtbuf_xlg[i]);
994 page = (char *) BufferGetBlock(dtbuf[i]);
996 rdt->next = &(dtbuf_rdt1[i]);
999 rdt->data = (char *) bkpb;
1000 rdt->len = sizeof(BkpBlock);
1001 write_len += sizeof(BkpBlock);
1003 rdt->next = &(dtbuf_rdt2[i]);
1006 if (bkpb->hole_length == 0)
1010 write_len += BLCKSZ;
1015 /* must skip the hole */
1017 rdt->len = bkpb->hole_offset;
1018 write_len += bkpb->hole_offset;
1020 rdt->next = &(dtbuf_rdt3[i]);
1023 rdt->data = page + (bkpb->hole_offset + bkpb->hole_length);
1024 rdt->len = BLCKSZ - (bkpb->hole_offset + bkpb->hole_length);
1025 write_len += rdt->len;
1031 * Calculate CRC of the data, including all the backup blocks
1033 * Note that the record header isn't added into the CRC initially since we
1034 * don't know the prev-link yet. Thus, the CRC will represent the CRC of
1035 * the whole record in the order: rdata, then backup blocks, then record
1038 INIT_CRC32(rdata_crc);
1039 for (rdt = rdata; rdt != NULL; rdt = rdt->next)
1040 COMP_CRC32(rdata_crc, rdt->data, rdt->len);
1043 * Construct record header (prev-link is filled in later, after reserving
1044 * the space for the record), and make that the first chunk in the chain.
1046 * The CRC calculated for the header here doesn't include prev-link,
1047 * because we don't know it yet. It will be added later.
1049 rechdr->xl_xid = GetCurrentTransactionIdIfAny();
1050 rechdr->xl_tot_len = SizeOfXLogRecord + write_len;
1051 rechdr->xl_len = len; /* doesn't include backup blocks */
1052 rechdr->xl_info = info;
1053 rechdr->xl_rmid = rmid;
1054 rechdr->xl_prev = InvalidXLogRecPtr;
1055 COMP_CRC32(rdata_crc, ((char *) rechdr), offsetof(XLogRecord, xl_prev));
1057 hdr_rdt.next = rdata;
1058 hdr_rdt.data = (char *) rechdr;
1059 hdr_rdt.len = SizeOfXLogRecord;
1060 write_len += SizeOfXLogRecord;
1064 * We have now done all the preparatory work we can without holding a
1065 * lock or modifying shared state. From here on, inserting the new WAL
1066 * record to the shared WAL buffer cache is a two-step process:
1068 * 1. Reserve the right amount of space from the WAL. The current head of
1069 * reserved space is kept in Insert->CurrBytePos, and is protected by
1072 * 2. Copy the record to the reserved WAL space. This involves finding the
1073 * correct WAL buffer containing the reserved space, and copying the
1074 * record in place. This can be done concurrently in multiple processes.
1076 * To keep track of which insertions are still in-progress, each concurrent
1077 * inserter allocates an "insertion slot", which tells others how far the
1078 * inserter has progressed. There is a small fixed number of insertion
1079 * slots, determined by the num_xloginsert_slots GUC. When an inserter
1080 * finishes, it updates the xlogInsertingAt of its slot to the end of the
1081 * record it inserted, to let others know that it's done. xlogInsertingAt
1082 * is also updated when crossing over to a new WAL buffer, to allow the
1083 * the previous buffer to be flushed.
1085 * Holding onto a slot also protects RedoRecPtr and fullPageWrites from
1086 * changing until the insertion is finished.
1088 * Step 2 can usually be done completely in parallel. If the required WAL
1089 * page is not initialized yet, you have to grab WALBufMappingLock to
1090 * initialize it, but the WAL writer tries to do that ahead of insertions
1091 * to avoid that from happening in the critical path.
1095 START_CRIT_SECTION();
1096 WALInsertSlotAcquire(isLogSwitch);
1099 * Check to see if my RedoRecPtr is out of date. If so, may have to go
1100 * back and recompute everything. This can only happen just after a
1101 * checkpoint, so it's better to be slow in this case and fast otherwise.
1103 * If we aren't doing full-page writes then RedoRecPtr doesn't actually
1104 * affect the contents of the XLOG record, so we'll update our local copy
1105 * but not force a recomputation.
1107 if (RedoRecPtr != Insert->RedoRecPtr)
1109 Assert(RedoRecPtr < Insert->RedoRecPtr);
1110 RedoRecPtr = Insert->RedoRecPtr;
1114 for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
1116 if (dtbuf[i] == InvalidBuffer)
1118 if (dtbuf_bkp[i] == false &&
1119 dtbuf_lsn[i] <= RedoRecPtr)
1122 * Oops, this buffer now needs to be backed up, but we
1123 * didn't think so above. Start over.
1125 WALInsertSlotRelease();
1127 rdt_lastnormal->next = NULL;
1136 * Also check to see if fullPageWrites or forcePageWrites was just turned
1137 * on; if we weren't already doing full-page writes then go back and
1138 * recompute. (If it was just turned off, we could recompute the record
1139 * without full pages, but we choose not to bother.)
1141 if ((Insert->fullPageWrites || Insert->forcePageWrites) && !doPageWrites)
1143 /* Oops, must redo it with full-page data. */
1144 WALInsertSlotRelease();
1146 rdt_lastnormal->next = NULL;
1152 * Reserve space for the record in the WAL. This also sets the xl_prev
1156 inserted = ReserveXLogSwitch(&StartPos, &EndPos, &rechdr->xl_prev);
1159 ReserveXLogInsertLocation(write_len, &StartPos, &EndPos,
1167 * Now that xl_prev has been filled in, finish CRC calculation of the
1170 COMP_CRC32(rdata_crc, ((char *) &rechdr->xl_prev), sizeof(XLogRecPtr));
1171 FIN_CRC32(rdata_crc);
1172 rechdr->xl_crc = rdata_crc;
1175 * All the record data, including the header, is now ready to be
1176 * inserted. Copy the record in the space reserved.
1178 CopyXLogRecordToWAL(write_len, isLogSwitch, &hdr_rdt, StartPos, EndPos);
1183 * This was an xlog-switch record, but the current insert location was
1184 * already exactly at the beginning of a segment, so there was no need
1190 * Done! Let others know that we're finished.
1192 WALInsertSlotRelease();
1197 * Update shared LogwrtRqst.Write, if we crossed page boundary.
1199 if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
1201 /* use volatile pointer to prevent code rearrangement */
1202 volatile XLogCtlData *xlogctl = XLogCtl;
1204 SpinLockAcquire(&xlogctl->info_lck);
1205 /* advance global request to include new block(s) */
1206 if (xlogctl->LogwrtRqst.Write < EndPos)
1207 xlogctl->LogwrtRqst.Write = EndPos;
1208 /* update local result copy while I have the chance */
1209 LogwrtResult = xlogctl->LogwrtResult;
1210 SpinLockRelease(&xlogctl->info_lck);
1214 * If this was an XLOG_SWITCH record, flush the record and the empty
1215 * padding space that fills the rest of the segment, and perform
1216 * end-of-segment actions (eg, notifying archiver).
1220 TRACE_POSTGRESQL_XLOG_SWITCH();
1223 * Even though we reserved the rest of the segment for us, which is
1224 * reflected in EndPos, we return a pointer to just the end of the
1225 * xlog-switch record.
1229 EndPos = StartPos + SizeOfXLogRecord;
1230 if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
1232 if (EndPos % XLOG_SEG_SIZE == EndPos % XLOG_BLCKSZ)
1233 EndPos += SizeOfXLogLongPHD;
1235 EndPos += SizeOfXLogShortPHD;
1245 initStringInfo(&buf);
1246 appendStringInfo(&buf, "INSERT @ %X/%X: ",
1247 (uint32) (EndPos >> 32), (uint32) EndPos);
1248 xlog_outrec(&buf, rechdr);
1249 if (rdata->data != NULL)
1251 appendStringInfo(&buf, " - ");
1252 RmgrTable[rechdr->xl_rmid].rm_desc(&buf, rechdr->xl_info, rdata->data);
1254 elog(LOG, "%s", buf.data);
1260 * Update our global variables
1262 ProcLastRecPtr = StartPos;
1263 XactLastRecEnd = EndPos;
1269 * Reserves the right amount of space for a record of given size from the WAL.
1270 * *StartPos is set to the beginning of the reserved section, *EndPos to
1271 * its end+1. *PrevPtr is set to the beginning of the previous record; it is
1272 * used to set the xl_prev of this record.
1274 * This is the performance critical part of XLogInsert that must be serialized
1275 * across backends. The rest can happen mostly in parallel. Try to keep this
1276 * section as short as possible, insertpos_lck can be heavily contended on a
1279 * NB: The space calculation here must match the code in CopyXLogRecordToWAL,
1280 * where we actually copy the record to the reserved space.
1283 ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1284 XLogRecPtr *PrevPtr)
1286 volatile XLogCtlInsert *Insert = &XLogCtl->Insert;
1287 uint64 startbytepos;
1291 size = MAXALIGN(size);
1293 /* All (non xlog-switch) records should contain data. */
1294 Assert(size > SizeOfXLogRecord);
1297 * The duration the spinlock needs to be held is minimized by minimizing
1298 * the calculations that have to be done while holding the lock. The
1299 * current tip of reserved WAL is kept in CurrBytePos, as a byte position
1300 * that only counts "usable" bytes in WAL, that is, it excludes all WAL
1301 * page headers. The mapping between "usable" byte positions and physical
1302 * positions (XLogRecPtrs) can be done outside the locked region, and
1303 * because the usable byte position doesn't include any headers, reserving
1304 * X bytes from WAL is almost as simple as "CurrBytePos += X".
1306 SpinLockAcquire(&Insert->insertpos_lck);
1308 startbytepos = Insert->CurrBytePos;
1309 endbytepos = startbytepos + size;
1310 prevbytepos = Insert->PrevBytePos;
1311 Insert->CurrBytePos = endbytepos;
1312 Insert->PrevBytePos = startbytepos;
1314 SpinLockRelease(&Insert->insertpos_lck);
1316 *StartPos = XLogBytePosToRecPtr(startbytepos);
1317 *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1318 *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1321 * Check that the conversions between "usable byte positions" and
1322 * XLogRecPtrs work consistently in both directions.
1324 Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1325 Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1326 Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1330 * Like ReserveXLogInsertLocation(), but for an xlog-switch record.
1332 * A log-switch record is handled slightly differently. The rest of the
1333 * segment will be reserved for this insertion, as indicated by the returned
1334 * *EndPos value. However, if we are already at the beginning of the current
1335 * segment, *StartPos and *EndPos are set to the current location without
1336 * reserving any space, and the function returns false.
1339 ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1341 volatile XLogCtlInsert *Insert = &XLogCtl->Insert;
1342 uint64 startbytepos;
1345 uint32 size = SizeOfXLogRecord;
1350 * These calculations are a bit heavy-weight to be done while holding a
1351 * spinlock, but since we're holding all the WAL insertion slots, there
1352 * are no other inserters competing for it. GetXLogInsertRecPtr() does
1353 * compete for it, but that's not called very frequently.
1355 SpinLockAcquire(&Insert->insertpos_lck);
1357 startbytepos = Insert->CurrBytePos;
1359 ptr = XLogBytePosToEndRecPtr(startbytepos);
1360 if (ptr % XLOG_SEG_SIZE == 0)
1362 SpinLockRelease(&Insert->insertpos_lck);
1363 *EndPos = *StartPos = ptr;
1367 endbytepos = startbytepos + size;
1368 prevbytepos = Insert->PrevBytePos;
1370 *StartPos = XLogBytePosToRecPtr(startbytepos);
1371 *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1373 segleft = XLOG_SEG_SIZE - ((*EndPos) % XLOG_SEG_SIZE);
1374 if (segleft != XLOG_SEG_SIZE)
1376 /* consume the rest of the segment */
1378 endbytepos = XLogRecPtrToBytePos(*EndPos);
1380 Insert->CurrBytePos = endbytepos;
1381 Insert->PrevBytePos = startbytepos;
1383 SpinLockRelease(&Insert->insertpos_lck);
1385 *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1387 Assert((*EndPos) % XLOG_SEG_SIZE == 0);
1388 Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1389 Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1390 Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1396 * Subroutine of XLogInsert. Copies a WAL record to an already-reserved
1400 CopyXLogRecordToWAL(int write_len, bool isLogSwitch, XLogRecData *rdata,
1401 XLogRecPtr StartPos, XLogRecPtr EndPos)
1407 XLogPageHeader pagehdr;
1409 /* The first chunk is the record header */
1410 Assert(rdata->len == SizeOfXLogRecord);
1413 * Get a pointer to the right place in the right WAL buffer to start
1417 currpos = GetXLogBuffer(CurrPos);
1418 freespace = INSERT_FREESPACE(CurrPos);
1421 * there should be enough space for at least the first field (xl_tot_len)
1424 Assert(freespace >= sizeof(uint32));
1426 /* Copy record data */
1428 while (rdata != NULL)
1430 char *rdata_data = rdata->data;
1431 int rdata_len = rdata->len;
1433 while (rdata_len > freespace)
1436 * Write what fits on this page, and continue on the next page.
1438 Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || freespace == 0);
1439 memcpy(currpos, rdata_data, freespace);
1440 rdata_data += freespace;
1441 rdata_len -= freespace;
1442 written += freespace;
1443 CurrPos += freespace;
1446 * Get pointer to beginning of next page, and set the xlp_rem_len
1447 * in the page header. Set XLP_FIRST_IS_CONTRECORD.
1449 * It's safe to set the contrecord flag and xlp_rem_len without a
1450 * lock on the page. All the other flags were already set when the
1451 * page was initialized, in AdvanceXLInsertBuffer, and we're the
1452 * only backend that needs to set the contrecord flag.
1454 currpos = GetXLogBuffer(CurrPos);
1455 pagehdr = (XLogPageHeader) currpos;
1456 pagehdr->xlp_rem_len = write_len - written;
1457 pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1459 /* skip over the page header */
1460 if (CurrPos % XLogSegSize == 0)
1462 CurrPos += SizeOfXLogLongPHD;
1463 currpos += SizeOfXLogLongPHD;
1467 CurrPos += SizeOfXLogShortPHD;
1468 currpos += SizeOfXLogShortPHD;
1470 freespace = INSERT_FREESPACE(CurrPos);
1473 Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1474 memcpy(currpos, rdata_data, rdata_len);
1475 currpos += rdata_len;
1476 CurrPos += rdata_len;
1477 freespace -= rdata_len;
1478 written += rdata_len;
1480 rdata = rdata->next;
1482 Assert(written == write_len);
1484 /* Align the end position, so that the next record starts aligned */
1485 CurrPos = MAXALIGN(CurrPos);
1488 * If this was an xlog-switch, it's not enough to write the switch record,
1489 * we also have to consume all the remaining space in the WAL segment.
1490 * We have already reserved it for us, but we still need to make sure it's
1491 * allocated and zeroed in the WAL buffers so that when the caller (or
1492 * someone else) does XLogWrite(), it can really write out all the zeros.
1494 if (isLogSwitch && CurrPos % XLOG_SEG_SIZE != 0)
1496 /* An xlog-switch record doesn't contain any data besides the header */
1497 Assert(write_len == SizeOfXLogRecord);
1500 * We do this one page at a time, to make sure we don't deadlock
1501 * against ourselves if wal_buffers < XLOG_SEG_SIZE.
1503 Assert(EndPos % XLogSegSize == 0);
1505 /* Use up all the remaining space on the first page */
1506 CurrPos += freespace;
1508 while (CurrPos < EndPos)
1510 /* initialize the next page (if not initialized already) */
1511 WakeupWaiters(CurrPos);
1512 AdvanceXLInsertBuffer(CurrPos, false);
1513 CurrPos += XLOG_BLCKSZ;
1517 if (CurrPos != EndPos)
1518 elog(PANIC, "space reserved for WAL record does not match what was written");
1522 * Allocate a slot for insertion.
1524 * In exclusive mode, all slots are reserved for the current process. That
1525 * blocks all concurrent insertions.
1528 WALInsertSlotAcquire(bool exclusive)
1534 for (i = 0; i < num_xloginsert_slots; i++)
1535 WALInsertSlotAcquireOne(i);
1536 holdingAllSlots = true;
1539 WALInsertSlotAcquireOne(-1);
1543 * Workhorse of WALInsertSlotAcquire. Acquires the given slot, or an arbitrary
1544 * one if slotno == -1. The index of the slot that was acquired is stored in
1547 * This is more or less equivalent to LWLockAcquire().
1550 WALInsertSlotAcquireOne(int slotno)
1552 volatile XLogInsertSlot *slot;
1553 PGPROC *proc = MyProc;
1556 static int slotToTry = -1;
1559 * Try to use the slot we used last time. If the system isn't particularly
1560 * busy, it's a good bet that it's available, and it's good to have some
1561 * affinity to a particular slot so that you don't unnecessarily bounce
1562 * cache lines between processes when there is no contention.
1564 * If this is the first time through in this backend, pick a slot
1565 * (semi-)randomly. This allows the slots to be used evenly if you have a
1566 * lot of very short connections.
1572 if (slotToTry == -1)
1573 slotToTry = MyProc->pgprocno % num_xloginsert_slots;
1574 MySlotNo = slotToTry;
1578 * We can't wait if we haven't got a PGPROC. This should only occur
1579 * during bootstrap or shared memory initialization. Put an Assert here
1580 * to catch unsafe coding practices.
1582 Assert(MyProc != NULL);
1585 * Lock out cancel/die interrupts until we exit the code section protected
1586 * by the slot. This ensures that interrupts will not interfere with
1587 * manipulations of data structures in shared memory. There is no cleanup
1588 * mechanism to release the slot if the backend dies while holding one,
1589 * so make this a critical section.
1591 START_CRIT_SECTION();
1594 * Loop here to try to acquire slot after each time we are signaled by
1595 * WALInsertSlotRelease.
1601 slot = &XLogCtl->Insert.insertSlots[MySlotNo].slot;
1603 /* Acquire mutex. Time spent holding mutex should be short! */
1604 SpinLockAcquire(&slot->mutex);
1606 /* If retrying, allow WALInsertSlotRelease to release waiters again */
1608 slot->releaseOK = true;
1610 /* If I can get the slot, do so quickly. */
1611 if (slot->exclusive == 0)
1620 break; /* got the lock */
1622 Assert(slot->owner != MyProc);
1625 * Add myself to wait queue.
1627 proc->lwWaiting = true;
1628 proc->lwWaitMode = LW_EXCLUSIVE;
1629 proc->lwWaitLink = NULL;
1630 if (slot->head == NULL)
1633 slot->tail->lwWaitLink = proc;
1636 /* Can release the mutex now */
1637 SpinLockRelease(&slot->mutex);
1640 * Wait until awakened.
1642 * Since we share the process wait semaphore with the regular lock
1643 * manager and ProcWaitForSignal, and we may need to acquire a slot
1644 * while one of those is pending, it is possible that we get awakened
1645 * for a reason other than being signaled by WALInsertSlotRelease. If
1646 * so, loop back and wait again. Once we've gotten the slot,
1647 * re-increment the sema by the number of additional signals received,
1648 * so that the lock manager or signal manager will see the received
1649 * signal when it next waits.
1653 /* "false" means cannot accept cancel/die interrupt here. */
1654 PGSemaphoreLock(&proc->sem, false);
1655 if (!proc->lwWaiting)
1660 /* Now loop back and try to acquire lock again. */
1667 * Normally, we initialize the xlogInsertingAt value of the slot to 1,
1668 * because we don't yet know where in the WAL we're going to insert. It's
1669 * not critical what it points to right now - leaving it to a too small
1670 * value just means that WaitXlogInsertionsToFinish() might wait on us
1671 * unnecessarily, until we update the value (when we finish the insert or
1672 * move to next page).
1674 * If we're grabbing all the slots, however, stamp all but the last one
1675 * with InvalidXLogRecPtr, meaning there is no insert in progress. The last
1676 * slot is the one that we will update as we proceed with the insert, the
1677 * rest are held just to keep off other inserters.
1679 if (slotno != -1 && slotno != num_xloginsert_slots - 1)
1680 slot->xlogInsertingAt = InvalidXLogRecPtr;
1682 slot->xlogInsertingAt = 1;
1684 /* We are done updating shared state of the slot itself. */
1685 SpinLockRelease(&slot->mutex);
1688 * Fix the process wait semaphore's count for any absorbed wakeups.
1690 while (extraWaits-- > 0)
1691 PGSemaphoreUnlock(&proc->sem);
1694 * If we couldn't get the slot immediately, try another slot next time.
1695 * On a system with more insertion slots than concurrent inserters, this
1696 * causes all the inserters to eventually migrate to a slot that no-one
1697 * else is using. On a system with more inserters than slots, it still
1698 * causes the inserters to be distributed quite evenly across the slots.
1700 if (slotno != -1 && retry)
1701 slotToTry = (slotToTry + 1) % num_xloginsert_slots;
1705 * Wait for the given slot to become free, or for its xlogInsertingAt location
1706 * to change to something else than 'waitptr'. In other words, wait for the
1707 * inserter using the given slot to finish its insertion, or to at least make
1711 WaitOnSlot(volatile XLogInsertSlot *slot, XLogRecPtr waitptr)
1713 PGPROC *proc = MyProc;
1717 * Lock out cancel/die interrupts while we sleep on the slot. There is
1718 * no cleanup mechanism to remove us from the wait queue if we got
1724 * Loop here to try to acquire lock after each time we are signaled.
1730 /* Acquire mutex. Time spent holding mutex should be short! */
1731 SpinLockAcquire(&slot->mutex);
1733 /* If I can get the lock, do so quickly. */
1734 if (slot->exclusive == 0 || slot->xlogInsertingAt != waitptr)
1740 break; /* the lock was free */
1742 Assert(slot->owner != MyProc);
1745 * Add myself to wait queue.
1747 proc->lwWaiting = true;
1748 proc->lwWaitMode = LW_WAIT_UNTIL_FREE;
1749 proc->lwWaitLink = NULL;
1751 /* waiters are added to the front of the queue */
1752 proc->lwWaitLink = slot->head;
1753 if (slot->head == NULL)
1757 /* Can release the mutex now */
1758 SpinLockRelease(&slot->mutex);
1761 * Wait until awakened.
1763 * Since we share the process wait semaphore with other things, like
1764 * the regular lock manager and ProcWaitForSignal, and we may need to
1765 * acquire an LWLock while one of those is pending, it is possible that
1766 * we get awakened for a reason other than being signaled by
1767 * LWLockRelease. If so, loop back and wait again. Once we've gotten
1768 * the LWLock, re-increment the sema by the number of additional
1769 * signals received, so that the lock manager or signal manager will
1770 * see the received signal when it next waits.
1774 /* "false" means cannot accept cancel/die interrupt here. */
1775 PGSemaphoreLock(&proc->sem, false);
1776 if (!proc->lwWaiting)
1781 /* Now loop back and try to acquire lock again. */
1784 /* We are done updating shared state of the lock itself. */
1785 SpinLockRelease(&slot->mutex);
1788 * Fix the process wait semaphore's count for any absorbed wakeups.
1790 while (extraWaits-- > 0)
1791 PGSemaphoreUnlock(&proc->sem);
1794 * Now okay to allow cancel/die interrupts.
1796 RESUME_INTERRUPTS();
1800 * Wake up all processes waiting for us with WaitOnSlot(). Sets our
1801 * xlogInsertingAt value to EndPos, without releasing the slot.
1804 WakeupWaiters(XLogRecPtr EndPos)
1806 volatile XLogInsertSlot *slot = &XLogCtl->Insert.insertSlots[MySlotNo].slot;
1812 * If we have already reported progress up to the same point, do nothing.
1813 * No other process can modify xlogInsertingAt, so we can check this before
1814 * grabbing the spinlock.
1816 if (slot->xlogInsertingAt == EndPos)
1818 /* xlogInsertingAt should not go backwards */
1819 Assert(slot->xlogInsertingAt < EndPos);
1821 /* Acquire mutex. Time spent holding mutex should be short! */
1822 SpinLockAcquire(&slot->mutex);
1824 /* we should own the slot */
1825 Assert(slot->exclusive == 1 && slot->owner == MyProc);
1827 slot->xlogInsertingAt = EndPos;
1830 * See if there are any waiters that need to be woken up.
1838 /* LW_WAIT_UNTIL_FREE waiters are always in the front of the queue */
1839 next = proc->lwWaitLink;
1840 while (next && next->lwWaitMode == LW_WAIT_UNTIL_FREE)
1843 next = next->lwWaitLink;
1846 /* proc is now the last PGPROC to be released */
1848 proc->lwWaitLink = NULL;
1851 /* We are done updating shared state of the lock itself. */
1852 SpinLockRelease(&slot->mutex);
1855 * Awaken any waiters I removed from the queue.
1857 while (head != NULL)
1860 head = proc->lwWaitLink;
1861 proc->lwWaitLink = NULL;
1862 proc->lwWaiting = false;
1863 PGSemaphoreUnlock(&proc->sem);
1868 * Release our insertion slot (or slots, if we're holding them all).
1871 WALInsertSlotRelease(void)
1875 if (holdingAllSlots)
1877 for (i = 0; i < num_xloginsert_slots; i++)
1878 WALInsertSlotReleaseOne(i);
1879 holdingAllSlots = false;
1882 WALInsertSlotReleaseOne(MySlotNo);
1886 WALInsertSlotReleaseOne(int slotno)
1888 volatile XLogInsertSlot *slot = &XLogCtl->Insert.insertSlots[slotno].slot;
1892 /* Acquire mutex. Time spent holding mutex should be short! */
1893 SpinLockAcquire(&slot->mutex);
1895 /* we must be holding it */
1896 Assert(slot->exclusive == 1 && slot->owner == MyProc);
1898 slot->xlogInsertingAt = InvalidXLogRecPtr;
1900 /* Release my hold on the slot */
1901 slot->exclusive = 0;
1905 * See if I need to awaken any waiters..
1910 if (slot->releaseOK)
1913 * Remove the to-be-awakened PGPROCs from the queue.
1915 bool releaseOK = true;
1920 * First wake up any backends that want to be woken up without
1921 * acquiring the lock. These are always in the front of the queue.
1923 while (proc->lwWaitMode == LW_WAIT_UNTIL_FREE && proc->lwWaitLink)
1924 proc = proc->lwWaitLink;
1927 * Awaken the first exclusive-waiter, if any.
1929 if (proc->lwWaitLink)
1931 Assert(proc->lwWaitLink->lwWaitMode == LW_EXCLUSIVE);
1932 proc = proc->lwWaitLink;
1935 /* proc is now the last PGPROC to be released */
1936 slot->head = proc->lwWaitLink;
1937 proc->lwWaitLink = NULL;
1939 slot->releaseOK = releaseOK;
1945 /* We are done updating shared state of the slot itself. */
1946 SpinLockRelease(&slot->mutex);
1949 * Awaken any waiters I removed from the queue.
1951 while (head != NULL)
1954 head = proc->lwWaitLink;
1955 proc->lwWaitLink = NULL;
1956 proc->lwWaiting = false;
1957 PGSemaphoreUnlock(&proc->sem);
1961 * Now okay to allow cancel/die interrupts.
1968 * Wait for any WAL insertions < upto to finish.
1970 * Returns the location of the oldest insertion that is still in-progress.
1971 * Any WAL prior to that point has been fully copied into WAL buffers, and
1972 * can be flushed out to disk. Because this waits for any insertions older
1973 * than 'upto' to finish, the return value is always >= 'upto'.
1975 * Note: When you are about to write out WAL, you must call this function
1976 * *before* acquiring WALWriteLock, to avoid deadlocks. This function might
1977 * need to wait for an insertion to finish (or at least advance to next
1978 * uninitialized page), and the inserter might need to evict an old WAL buffer
1979 * to make room for a new one, which in turn requires WALWriteLock.
1982 WaitXLogInsertionsToFinish(XLogRecPtr upto)
1985 XLogRecPtr reservedUpto;
1986 XLogRecPtr finishedUpto;
1987 volatile XLogCtlInsert *Insert = &XLogCtl->Insert;
1991 elog(PANIC, "cannot wait without a PGPROC structure");
1993 /* Read the current insert position */
1994 SpinLockAcquire(&Insert->insertpos_lck);
1995 bytepos = Insert->CurrBytePos;
1996 SpinLockRelease(&Insert->insertpos_lck);
1997 reservedUpto = XLogBytePosToEndRecPtr(bytepos);
2000 * No-one should request to flush a piece of WAL that hasn't even been
2001 * reserved yet. However, it can happen if there is a block with a bogus
2002 * LSN on disk, for example. XLogFlush checks for that situation and
2003 * complains, but only after the flush. Here we just assume that to mean
2004 * that all WAL that has been reserved needs to be finished. In this
2005 * corner-case, the return value can be smaller than 'upto' argument.
2007 if (upto > reservedUpto)
2009 elog(LOG, "request to flush past end of generated WAL; request %X/%X, currpos %X/%X",
2010 (uint32) (upto >> 32), (uint32) upto,
2011 (uint32) (reservedUpto >> 32), (uint32) reservedUpto);
2012 upto = reservedUpto;
2016 * finishedUpto is our return value, indicating the point upto which
2017 * all the WAL insertions have been finished. Initialize it to the head
2018 * of reserved WAL, and as we iterate through the insertion slots, back it
2019 * out for any insertion that's still in progress.
2021 finishedUpto = reservedUpto;
2024 * Loop through all the slots, sleeping on any in-progress insert older
2027 for (i = 0; i < num_xloginsert_slots; i++)
2029 volatile XLogInsertSlot *slot = &XLogCtl->Insert.insertSlots[i].slot;
2030 XLogRecPtr insertingat;
2034 * We can check if the slot is in use without grabbing the spinlock.
2035 * The spinlock acquisition of insertpos_lck before this loop acts
2036 * as a memory barrier. If someone acquires the slot after that, it
2037 * can't possibly be inserting to anything < reservedUpto. If it was
2038 * acquired before that, an unlocked test will return true.
2040 if (!slot->exclusive)
2043 SpinLockAcquire(&slot->mutex);
2044 /* re-check now that we have the lock */
2045 if (!slot->exclusive)
2047 SpinLockRelease(&slot->mutex);
2050 insertingat = slot->xlogInsertingAt;
2051 SpinLockRelease(&slot->mutex);
2053 if (insertingat == InvalidXLogRecPtr)
2056 * slot is reserved just to hold off other inserters, there is no
2057 * actual insert in progress.
2063 * This insertion is still in progress. Do we need to wait for it?
2065 * When an inserter acquires a slot, it doesn't reset 'insertingat', so
2066 * it will initially point to the old value of some already-finished
2067 * insertion. The inserter will update the value as soon as it finishes
2068 * the insertion, moves to the next page, or has to do I/O to flush an
2069 * old dirty buffer. That means that when we see a slot with
2070 * insertingat value < upto, we don't know if that insertion is still
2071 * truly in progress, or if the slot is reused by a new inserter that
2072 * hasn't updated the insertingat value yet. We have to assume it's the
2075 if (insertingat < upto)
2077 WaitOnSlot(slot, insertingat);
2083 * We don't need to wait for this insertion, but update the
2086 if (insertingat < finishedUpto)
2087 finishedUpto = insertingat;
2090 return finishedUpto;
2094 * Get a pointer to the right location in the WAL buffer containing the
2097 * If the page is not initialized yet, it is initialized. That might require
2098 * evicting an old dirty buffer from the buffer cache, which means I/O.
2100 * The caller must ensure that the page containing the requested location
2101 * isn't evicted yet, and won't be evicted. The way to ensure that is to
2102 * hold onto an XLogInsertSlot with the xlogInsertingAt position set to
2103 * something <= ptr. GetXLogBuffer() will update xlogInsertingAt if it needs
2104 * to evict an old page from the buffer. (This means that once you call
2105 * GetXLogBuffer() with a given 'ptr', you must not access anything before
2106 * that point anymore, and must not call GetXLogBuffer() with an older 'ptr'
2107 * later, because older buffers might be recycled already)
2110 GetXLogBuffer(XLogRecPtr ptr)
2114 static uint64 cachedPage = 0;
2115 static char *cachedPos = NULL;
2116 XLogRecPtr expectedEndPtr;
2119 * Fast path for the common case that we need to access again the same
2120 * page as last time.
2122 if (ptr / XLOG_BLCKSZ == cachedPage)
2124 Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
2125 Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
2126 return cachedPos + ptr % XLOG_BLCKSZ;
2130 * The XLog buffer cache is organized so that a page is always loaded
2131 * to a particular buffer. That way we can easily calculate the buffer
2132 * a given page must be loaded into, from the XLogRecPtr alone.
2134 idx = XLogRecPtrToBufIdx(ptr);
2137 * See what page is loaded in the buffer at the moment. It could be the
2138 * page we're looking for, or something older. It can't be anything newer
2139 * - that would imply the page we're looking for has already been written
2140 * out to disk and evicted, and the caller is responsible for making sure
2141 * that doesn't happen.
2143 * However, we don't hold a lock while we read the value. If someone has
2144 * just initialized the page, it's possible that we get a "torn read" of
2145 * the XLogRecPtr if 64-bit fetches are not atomic on this platform. In
2146 * that case we will see a bogus value. That's ok, we'll grab the mapping
2147 * lock (in AdvanceXLInsertBuffer) and retry if we see anything else than
2148 * the page we're looking for. But it means that when we do this unlocked
2149 * read, we might see a value that appears to be ahead of the page we're
2150 * looking for. Don't PANIC on that, until we've verified the value while
2153 expectedEndPtr = ptr;
2154 expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
2156 endptr = XLogCtl->xlblocks[idx];
2157 if (expectedEndPtr != endptr)
2160 * Let others know that we're finished inserting the record up
2161 * to the page boundary.
2163 WakeupWaiters(expectedEndPtr - XLOG_BLCKSZ);
2165 AdvanceXLInsertBuffer(ptr, false);
2166 endptr = XLogCtl->xlblocks[idx];
2168 if (expectedEndPtr != endptr)
2169 elog(PANIC, "could not find WAL buffer for %X/%X",
2170 (uint32) (ptr >> 32) , (uint32) ptr);
2175 * Make sure the initialization of the page is visible to us, and
2176 * won't arrive later to overwrite the WAL data we write on the page.
2178 pg_memory_barrier();
2182 * Found the buffer holding this page. Return a pointer to the right
2183 * offset within the page.
2185 cachedPage = ptr / XLOG_BLCKSZ;
2186 cachedPos = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
2188 Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
2189 Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
2191 return cachedPos + ptr % XLOG_BLCKSZ;
2195 * Converts a "usable byte position" to XLogRecPtr. A usable byte position
2196 * is the position starting from the beginning of WAL, excluding all WAL
2200 XLogBytePosToRecPtr(uint64 bytepos)
2208 fullsegs = bytepos / UsableBytesInSegment;
2209 bytesleft = bytepos % UsableBytesInSegment;
2211 if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
2213 /* fits on first page of segment */
2214 seg_offset = bytesleft + SizeOfXLogLongPHD;
2218 /* account for the first page on segment with long header */
2219 seg_offset = XLOG_BLCKSZ;
2220 bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
2222 fullpages = bytesleft / UsableBytesInPage;
2223 bytesleft = bytesleft % UsableBytesInPage;
2225 seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
2228 XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, result);
2234 * Like XLogBytePosToRecPtr, but if the position is at a page boundary,
2235 * returns a pointer to the beginning of the page (ie. before page header),
2236 * not to where the first xlog record on that page would go to. This is used
2237 * when converting a pointer to the end of a record.
2240 XLogBytePosToEndRecPtr(uint64 bytepos)
2248 fullsegs = bytepos / UsableBytesInSegment;
2249 bytesleft = bytepos % UsableBytesInSegment;
2251 if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
2253 /* fits on first page of segment */
2257 seg_offset = bytesleft + SizeOfXLogLongPHD;
2261 /* account for the first page on segment with long header */
2262 seg_offset = XLOG_BLCKSZ;
2263 bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
2265 fullpages = bytesleft / UsableBytesInPage;
2266 bytesleft = bytesleft % UsableBytesInPage;
2269 seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
2271 seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
2274 XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, result);
2280 * Convert an XLogRecPtr to a "usable byte position".
2283 XLogRecPtrToBytePos(XLogRecPtr ptr)
2290 XLByteToSeg(ptr, fullsegs);
2292 fullpages = (ptr % XLOG_SEG_SIZE) / XLOG_BLCKSZ;
2293 offset = ptr % XLOG_BLCKSZ;
2297 result = fullsegs * UsableBytesInSegment;
2300 Assert(offset >= SizeOfXLogLongPHD);
2301 result += offset - SizeOfXLogLongPHD;
2306 result = fullsegs * UsableBytesInSegment +
2307 (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
2308 (fullpages - 1) * UsableBytesInPage; /* full pages */
2311 Assert(offset >= SizeOfXLogShortPHD);
2312 result += offset - SizeOfXLogShortPHD;
2320 * Determine whether the buffer referenced by an XLogRecData item has to
2321 * be backed up, and if so fill a BkpBlock struct for it. In any case
2322 * save the buffer's LSN at *lsn.
2325 XLogCheckBuffer(XLogRecData *rdata, bool holdsExclusiveLock,
2326 XLogRecPtr *lsn, BkpBlock *bkpb)
2330 page = BufferGetPage(rdata->buffer);
2333 * We assume page LSN is first data on *every* page that can be passed to
2334 * XLogInsert, whether it has the standard page layout or not. We don't
2335 * need to take the buffer header lock for PageGetLSN if we hold an
2336 * exclusive lock on the page and/or the relation.
2338 if (holdsExclusiveLock)
2339 *lsn = PageGetLSN(page);
2341 *lsn = BufferGetLSNAtomic(rdata->buffer);
2343 if (*lsn <= RedoRecPtr)
2346 * The page needs to be backed up, so set up *bkpb
2348 BufferGetTag(rdata->buffer, &bkpb->node, &bkpb->fork, &bkpb->block);
2350 if (rdata->buffer_std)
2352 /* Assume we can omit data between pd_lower and pd_upper */
2353 uint16 lower = ((PageHeader) page)->pd_lower;
2354 uint16 upper = ((PageHeader) page)->pd_upper;
2356 if (lower >= SizeOfPageHeaderData &&
2360 bkpb->hole_offset = lower;
2361 bkpb->hole_length = upper - lower;
2365 /* No "hole" to compress out */
2366 bkpb->hole_offset = 0;
2367 bkpb->hole_length = 0;
2372 /* Not a standard page header, don't try to eliminate "hole" */
2373 bkpb->hole_offset = 0;
2374 bkpb->hole_length = 0;
2377 return true; /* buffer requires backup */
2380 return false; /* buffer does not need to be backed up */
2384 * Initialize XLOG buffers, writing out old buffers if they still contain
2385 * unwritten data, upto the page containing 'upto'. Or if 'opportunistic' is
2386 * true, initialize as many pages as we can without having to write out
2387 * unwritten data. Any new pages are initialized to zeros, with pages headers
2388 * initialized properly.
2391 AdvanceXLInsertBuffer(XLogRecPtr upto, bool opportunistic)
2393 XLogCtlInsert *Insert = &XLogCtl->Insert;
2395 XLogRecPtr OldPageRqstPtr;
2396 XLogwrtRqst WriteRqst;
2397 XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
2398 XLogRecPtr NewPageBeginPtr;
2399 XLogPageHeader NewPage;
2402 LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2405 * Now that we have the lock, check if someone initialized the page
2408 while (upto >= XLogCtl->InitializedUpTo || opportunistic)
2410 nextidx = XLogRecPtrToBufIdx(XLogCtl->InitializedUpTo);
2413 * Get ending-offset of the buffer page we need to replace (this may
2414 * be zero if the buffer hasn't been used yet). Fall through if it's
2415 * already written out.
2417 OldPageRqstPtr = XLogCtl->xlblocks[nextidx];
2418 if (LogwrtResult.Write < OldPageRqstPtr)
2421 * Nope, got work to do. If we just want to pre-initialize as much
2422 * as we can without flushing, give up now.
2427 /* Before waiting, get info_lck and update LogwrtResult */
2429 /* use volatile pointer to prevent code rearrangement */
2430 volatile XLogCtlData *xlogctl = XLogCtl;
2432 SpinLockAcquire(&xlogctl->info_lck);
2433 if (xlogctl->LogwrtRqst.Write < OldPageRqstPtr)
2434 xlogctl->LogwrtRqst.Write = OldPageRqstPtr;
2435 LogwrtResult = xlogctl->LogwrtResult;
2436 SpinLockRelease(&xlogctl->info_lck);
2440 * Now that we have an up-to-date LogwrtResult value, see if we
2441 * still need to write it or if someone else already did.
2443 if (LogwrtResult.Write < OldPageRqstPtr)
2446 * Must acquire write lock. Release WALBufMappingLock first,
2447 * to make sure that all insertions that we need to wait for
2448 * can finish (up to this same position). Otherwise we risk
2451 LWLockRelease(WALBufMappingLock);
2453 WaitXLogInsertionsToFinish(OldPageRqstPtr);
2455 LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2457 LogwrtResult = XLogCtl->LogwrtResult;
2458 if (LogwrtResult.Write >= OldPageRqstPtr)
2460 /* OK, someone wrote it already */
2461 LWLockRelease(WALWriteLock);
2465 /* Have to write it ourselves */
2466 TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2467 WriteRqst.Write = OldPageRqstPtr;
2468 WriteRqst.Flush = 0;
2469 XLogWrite(WriteRqst, false);
2470 LWLockRelease(WALWriteLock);
2471 TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_DONE();
2473 /* Re-acquire WALBufMappingLock and retry */
2474 LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2480 * Now the next buffer slot is free and we can set it up to be the next
2483 NewPageBeginPtr = XLogCtl->InitializedUpTo;
2484 NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2486 Assert(XLogRecPtrToBufIdx(NewPageBeginPtr) == nextidx);
2488 NewPage = (XLogPageHeader) (XLogCtl->pages + nextidx * (Size) XLOG_BLCKSZ);
2491 * Be sure to re-zero the buffer so that bytes beyond what we've
2492 * written will look like zeroes and not valid XLOG records...
2494 MemSet((char *) NewPage, 0, XLOG_BLCKSZ);
2497 * Fill the new page's header
2499 NewPage ->xlp_magic = XLOG_PAGE_MAGIC;
2501 /* NewPage->xlp_info = 0; */ /* done by memset */
2502 NewPage ->xlp_tli = ThisTimeLineID;
2503 NewPage ->xlp_pageaddr = NewPageBeginPtr;
2504 /* NewPage->xlp_rem_len = 0; */ /* done by memset */
2507 * If online backup is not in progress, mark the header to indicate
2508 * that* WAL records beginning in this page have removable backup
2509 * blocks. This allows the WAL archiver to know whether it is safe to
2510 * compress archived WAL data by transforming full-block records into
2511 * the non-full-block format. It is sufficient to record this at the
2512 * page level because we force a page switch (in fact a segment switch)
2513 * when starting a backup, so the flag will be off before any records
2514 * can be written during the backup. At the end of a backup, the last
2515 * page will be marked as all unsafe when perhaps only part is unsafe,
2516 * but at worst the archiver would miss the opportunity to compress a
2519 if (!Insert->forcePageWrites)
2520 NewPage ->xlp_info |= XLP_BKP_REMOVABLE;
2523 * If first page of an XLOG segment file, make it a long header.
2525 if ((NewPage->xlp_pageaddr % XLogSegSize) == 0)
2527 XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2529 NewLongPage->xlp_sysid = ControlFile->system_identifier;
2530 NewLongPage->xlp_seg_size = XLogSegSize;
2531 NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2532 NewPage ->xlp_info |= XLP_LONG_HEADER;
2536 * Make sure the initialization of the page becomes visible to others
2537 * before the xlblocks update. GetXLogBuffer() reads xlblocks without
2542 *((volatile XLogRecPtr *) &XLogCtl->xlblocks[nextidx]) = NewPageEndPtr;
2544 XLogCtl->InitializedUpTo = NewPageEndPtr;
2548 LWLockRelease(WALBufMappingLock);
2553 elog(DEBUG1, "initialized %d pages, upto %X/%X",
2554 npages, (uint32) (NewPageEndPtr >> 32), (uint32) NewPageEndPtr);
2560 * Check whether we've consumed enough xlog space that a checkpoint is needed.
2562 * new_segno indicates a log file that has just been filled up (or read
2563 * during recovery). We measure the distance from RedoRecPtr to new_segno
2564 * and see if that exceeds CheckPointSegments.
2566 * Note: it is caller's responsibility that RedoRecPtr is up-to-date.
2569 XLogCheckpointNeeded(XLogSegNo new_segno)
2571 XLogSegNo old_segno;
2573 XLByteToSeg(RedoRecPtr, old_segno);
2575 if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2581 * Write and/or fsync the log at least as far as WriteRqst indicates.
2583 * If flexible == TRUE, we don't have to write as far as WriteRqst, but
2584 * may stop at any convenient boundary (such as a cache or logfile boundary).
2585 * This option allows us to avoid uselessly issuing multiple writes when a
2586 * single one would do.
2588 * Must be called with WALWriteLock held. WaitXLogInsertionsToFinish(WriteRqst)
2589 * must be called before grabbing the lock, to make sure the data is ready to
2593 XLogWrite(XLogwrtRqst WriteRqst, bool flexible)
2596 bool last_iteration;
2604 /* We should always be inside a critical section here */
2605 Assert(CritSectionCount > 0);
2608 * Update local LogwrtResult (caller probably did this already, but...)
2610 LogwrtResult = XLogCtl->LogwrtResult;
2613 * Since successive pages in the xlog cache are consecutively allocated,
2614 * we can usually gather multiple pages together and issue just one
2615 * write() call. npages is the number of pages we have determined can be
2616 * written together; startidx is the cache block index of the first one,
2617 * and startoffset is the file offset at which it should go. The latter
2618 * two variables are only valid when npages > 0, but we must initialize
2619 * all of them to keep the compiler quiet.
2626 * Within the loop, curridx is the cache block index of the page to
2627 * consider writing. Begin at the buffer containing the next unwritten
2628 * page, or last partially written page.
2630 curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2632 while (LogwrtResult.Write < WriteRqst.Write)
2635 * Make sure we're not ahead of the insert process. This could happen
2636 * if we're passed a bogus WriteRqst.Write that is past the end of the
2637 * last page that's been initialized by AdvanceXLInsertBuffer.
2639 XLogRecPtr EndPtr = XLogCtl->xlblocks[curridx];
2640 if (LogwrtResult.Write >= EndPtr)
2641 elog(PANIC, "xlog write request %X/%X is past end of log %X/%X",
2642 (uint32) (LogwrtResult.Write >> 32),
2643 (uint32) LogwrtResult.Write,
2644 (uint32) (EndPtr >> 32), (uint32) EndPtr);
2646 /* Advance LogwrtResult.Write to end of current buffer page */
2647 LogwrtResult.Write = EndPtr;
2648 ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2650 if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo))
2653 * Switch to new logfile segment. We cannot have any pending
2654 * pages here (since we dump what we have at segment end).
2656 Assert(npages == 0);
2657 if (openLogFile >= 0)
2659 XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo);
2661 /* create/use new log file */
2662 use_existent = true;
2663 openLogFile = XLogFileInit(openLogSegNo, &use_existent, true);
2667 /* Make sure we have the current logfile open */
2668 if (openLogFile < 0)
2670 XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo);
2671 openLogFile = XLogFileOpen(openLogSegNo);
2675 /* Add current page to the set of pending pages-to-dump */
2678 /* first of group */
2680 startoffset = (LogwrtResult.Write - XLOG_BLCKSZ) % XLogSegSize;
2685 * Dump the set if this will be the last loop iteration, or if we are
2686 * at the last page of the cache area (since the next page won't be
2687 * contiguous in memory), or if we are at the end of the logfile
2690 last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2692 finishing_seg = !ispartialpage &&
2693 (startoffset + npages * XLOG_BLCKSZ) >= XLogSegSize;
2695 if (last_iteration ||
2696 curridx == XLogCtl->XLogCacheBlck ||
2704 /* Need to seek in the file? */
2705 if (openLogOff != startoffset)
2707 if (lseek(openLogFile, (off_t) startoffset, SEEK_SET) < 0)
2709 (errcode_for_file_access(),
2710 errmsg("could not seek in log file %s to offset %u: %m",
2711 XLogFileNameP(ThisTimeLineID, openLogSegNo),
2713 openLogOff = startoffset;
2716 /* OK to write the page(s) */
2717 from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2718 nbytes = npages * (Size) XLOG_BLCKSZ;
2723 written = write(openLogFile, from, nleft);
2729 (errcode_for_file_access(),
2730 errmsg("could not write to log file %s "
2731 "at offset %u, length %lu: %m",
2732 XLogFileNameP(ThisTimeLineID, openLogSegNo),
2733 openLogOff, (unsigned long) nbytes)));
2737 } while (nleft > 0);
2739 /* Update state for write */
2740 openLogOff += nbytes;
2744 * If we just wrote the whole last page of a logfile segment,
2745 * fsync the segment immediately. This avoids having to go back
2746 * and re-open prior segments when an fsync request comes along
2747 * later. Doing it here ensures that one and only one backend will
2748 * perform this fsync.
2750 * This is also the right place to notify the Archiver that the
2751 * segment is ready to copy to archival storage, and to update the
2752 * timer for archive_timeout, and to signal for a checkpoint if
2753 * too many logfile segments have been used since the last
2758 issue_xlog_fsync(openLogFile, openLogSegNo);
2760 /* signal that we need to wakeup walsenders later */
2761 WalSndWakeupRequest();
2763 LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2765 if (XLogArchivingActive())
2766 XLogArchiveNotifySeg(openLogSegNo);
2768 XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2771 * Request a checkpoint if we've consumed too much xlog since
2772 * the last one. For speed, we first check using the local
2773 * copy of RedoRecPtr, which might be out of date; if it looks
2774 * like a checkpoint is needed, forcibly update RedoRecPtr and
2777 if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2779 (void) GetRedoRecPtr();
2780 if (XLogCheckpointNeeded(openLogSegNo))
2781 RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2788 /* Only asked to write a partial page */
2789 LogwrtResult.Write = WriteRqst.Write;
2792 curridx = NextBufIdx(curridx);
2794 /* If flexible, break out of loop as soon as we wrote something */
2795 if (flexible && npages == 0)
2799 Assert(npages == 0);
2802 * If asked to flush, do so
2804 if (LogwrtResult.Flush < WriteRqst.Flush &&
2805 LogwrtResult.Flush < LogwrtResult.Write)
2809 * Could get here without iterating above loop, in which case we might
2810 * have no open file or the wrong one. However, we do not need to
2811 * fsync more than one file.
2813 if (sync_method != SYNC_METHOD_OPEN &&
2814 sync_method != SYNC_METHOD_OPEN_DSYNC)
2816 if (openLogFile >= 0 &&
2817 !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo))
2819 if (openLogFile < 0)
2821 XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo);
2822 openLogFile = XLogFileOpen(openLogSegNo);
2826 issue_xlog_fsync(openLogFile, openLogSegNo);
2829 /* signal that we need to wakeup walsenders later */
2830 WalSndWakeupRequest();
2832 LogwrtResult.Flush = LogwrtResult.Write;
2836 * Update shared-memory status
2838 * We make sure that the shared 'request' values do not fall behind the
2839 * 'result' values. This is not absolutely essential, but it saves some
2840 * code in a couple of places.
2843 /* use volatile pointer to prevent code rearrangement */
2844 volatile XLogCtlData *xlogctl = XLogCtl;
2846 SpinLockAcquire(&xlogctl->info_lck);
2847 xlogctl->LogwrtResult = LogwrtResult;
2848 if (xlogctl->LogwrtRqst.Write < LogwrtResult.Write)
2849 xlogctl->LogwrtRqst.Write = LogwrtResult.Write;
2850 if (xlogctl->LogwrtRqst.Flush < LogwrtResult.Flush)
2851 xlogctl->LogwrtRqst.Flush = LogwrtResult.Flush;
2852 SpinLockRelease(&xlogctl->info_lck);
2857 * Record the LSN for an asynchronous transaction commit/abort
2858 * and nudge the WALWriter if there is work for it to do.
2859 * (This should not be called for synchronous commits.)
2862 XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2864 XLogRecPtr WriteRqstPtr = asyncXactLSN;
2867 /* use volatile pointer to prevent code rearrangement */
2868 volatile XLogCtlData *xlogctl = XLogCtl;
2870 SpinLockAcquire(&xlogctl->info_lck);
2871 LogwrtResult = xlogctl->LogwrtResult;
2872 sleeping = xlogctl->WalWriterSleeping;
2873 if (xlogctl->asyncXactLSN < asyncXactLSN)
2874 xlogctl->asyncXactLSN = asyncXactLSN;
2875 SpinLockRelease(&xlogctl->info_lck);
2878 * If the WALWriter is sleeping, we should kick it to make it come out of
2879 * low-power mode. Otherwise, determine whether there's a full page of
2880 * WAL available to write.
2884 /* back off to last completed page boundary */
2885 WriteRqstPtr -= WriteRqstPtr % XLOG_BLCKSZ;
2887 /* if we have already flushed that far, we're done */
2888 if (WriteRqstPtr <= LogwrtResult.Flush)
2893 * Nudge the WALWriter: it has a full page of WAL to write, or we want it
2894 * to come out of low-power mode so that this async commit will reach disk
2895 * within the expected amount of time.
2897 if (ProcGlobal->walwriterLatch)
2898 SetLatch(ProcGlobal->walwriterLatch);
2902 * Advance minRecoveryPoint in control file.
2904 * If we crash during recovery, we must reach this point again before the
2905 * database is consistent.
2907 * If 'force' is true, 'lsn' argument is ignored. Otherwise, minRecoveryPoint
2908 * is only updated if it's not already greater than or equal to 'lsn'.
2911 UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2913 /* Quick check using our local copy of the variable */
2914 if (!updateMinRecoveryPoint || (!force && lsn <= minRecoveryPoint))
2917 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2919 /* update local copy */
2920 minRecoveryPoint = ControlFile->minRecoveryPoint;
2921 minRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2924 * An invalid minRecoveryPoint means that we need to recover all the WAL,
2925 * i.e., we're doing crash recovery. We never modify the control file's
2926 * value in that case, so we can short-circuit future checks here too.
2928 if (minRecoveryPoint == 0)
2929 updateMinRecoveryPoint = false;
2930 else if (force || minRecoveryPoint < lsn)
2932 /* use volatile pointer to prevent code rearrangement */
2933 volatile XLogCtlData *xlogctl = XLogCtl;
2934 XLogRecPtr newMinRecoveryPoint;
2935 TimeLineID newMinRecoveryPointTLI;
2938 * To avoid having to update the control file too often, we update it
2939 * all the way to the last record being replayed, even though 'lsn'
2940 * would suffice for correctness. This also allows the 'force' case
2941 * to not need a valid 'lsn' value.
2943 * Another important reason for doing it this way is that the passed
2944 * 'lsn' value could be bogus, i.e., past the end of available WAL, if
2945 * the caller got it from a corrupted heap page. Accepting such a
2946 * value as the min recovery point would prevent us from coming up at
2947 * all. Instead, we just log a warning and continue with recovery.
2948 * (See also the comments about corrupt LSNs in XLogFlush.)
2950 SpinLockAcquire(&xlogctl->info_lck);
2951 newMinRecoveryPoint = xlogctl->replayEndRecPtr;
2952 newMinRecoveryPointTLI = xlogctl->replayEndTLI;
2953 SpinLockRelease(&xlogctl->info_lck);
2955 if (!force && newMinRecoveryPoint < lsn)
2957 "xlog min recovery request %X/%X is past current point %X/%X",
2958 (uint32) (lsn >> 32), (uint32) lsn,
2959 (uint32) (newMinRecoveryPoint >> 32),
2960 (uint32) newMinRecoveryPoint);
2962 /* update control file */
2963 if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2965 ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2966 ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2967 UpdateControlFile();
2968 minRecoveryPoint = newMinRecoveryPoint;
2969 minRecoveryPointTLI = newMinRecoveryPointTLI;
2972 (errmsg("updated min recovery point to %X/%X on timeline %u",
2973 (uint32) (minRecoveryPoint >> 32),
2974 (uint32) minRecoveryPoint,
2975 newMinRecoveryPointTLI)));
2978 LWLockRelease(ControlFileLock);
2982 * Ensure that all XLOG data through the given position is flushed to disk.
2984 * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
2985 * already held, and we try to avoid acquiring it if possible.
2988 XLogFlush(XLogRecPtr record)
2990 XLogRecPtr WriteRqstPtr;
2991 XLogwrtRqst WriteRqst;
2994 * During REDO, we are reading not writing WAL. Therefore, instead of
2995 * trying to flush the WAL, we should update minRecoveryPoint instead. We
2996 * test XLogInsertAllowed(), not InRecovery, because we need checkpointer
2997 * to act this way too, and because when it tries to write the
2998 * end-of-recovery checkpoint, it should indeed flush.
3000 if (!XLogInsertAllowed())
3002 UpdateMinRecoveryPoint(record, false);
3006 /* Quick exit if already known flushed */
3007 if (record <= LogwrtResult.Flush)
3012 elog(LOG, "xlog flush request %X/%X; write %X/%X; flush %X/%X",
3013 (uint32) (record >> 32), (uint32) record,
3014 (uint32) (LogwrtResult.Write >> 32), (uint32) LogwrtResult.Write,
3015 (uint32) (LogwrtResult.Flush >> 32), (uint32) LogwrtResult.Flush);
3018 START_CRIT_SECTION();
3021 * Since fsync is usually a horribly expensive operation, we try to
3022 * piggyback as much data as we can on each fsync: if we see any more data
3023 * entered into the xlog buffer, we'll write and fsync that too, so that
3024 * the final value of LogwrtResult.Flush is as large as possible. This
3025 * gives us some chance of avoiding another fsync immediately after.
3028 /* initialize to given target; may increase below */
3029 WriteRqstPtr = record;
3032 * Now wait until we get the write lock, or someone else does the flush
3037 /* use volatile pointer to prevent code rearrangement */
3038 volatile XLogCtlData *xlogctl = XLogCtl;
3039 XLogRecPtr insertpos;
3041 /* read LogwrtResult and update local state */
3042 SpinLockAcquire(&xlogctl->info_lck);
3043 if (WriteRqstPtr < xlogctl->LogwrtRqst.Write)
3044 WriteRqstPtr = xlogctl->LogwrtRqst.Write;
3045 LogwrtResult = xlogctl->LogwrtResult;
3046 SpinLockRelease(&xlogctl->info_lck);
3049 if (record <= LogwrtResult.Flush)
3053 * Before actually performing the write, wait for all in-flight
3054 * insertions to the pages we're about to write to finish.
3056 insertpos = WaitXLogInsertionsToFinish(WriteRqstPtr);
3059 * Try to get the write lock. If we can't get it immediately, wait
3060 * until it's released, and recheck if we still need to do the flush
3061 * or if the backend that held the lock did it for us already. This
3062 * helps to maintain a good rate of group committing when the system
3063 * is bottlenecked by the speed of fsyncing.
3065 if (!LWLockAcquireOrWait(WALWriteLock, LW_EXCLUSIVE))
3068 * The lock is now free, but we didn't acquire it yet. Before we
3069 * do, loop back to check if someone else flushed the record for
3075 /* Got the lock; recheck whether request is satisfied */
3076 LogwrtResult = XLogCtl->LogwrtResult;
3077 if (record <= LogwrtResult.Flush)
3079 LWLockRelease(WALWriteLock);
3084 * Sleep before flush! By adding a delay here, we may give further
3085 * backends the opportunity to join the backlog of group commit
3086 * followers; this can significantly improve transaction throughput,
3087 * at the risk of increasing transaction latency.
3089 * We do not sleep if enableFsync is not turned on, nor if there are
3090 * fewer than CommitSiblings other backends with active transactions.
3092 if (CommitDelay > 0 && enableFsync &&
3093 MinimumActiveBackends(CommitSiblings))
3095 pg_usleep(CommitDelay);
3098 * Re-check how far we can now flush the WAL. It's generally not
3099 * safe to call WaitXLogInsetionsToFinish while holding
3100 * WALWriteLock, because an in-progress insertion might need to
3101 * also grab WALWriteLock to make progress. But we know that all
3102 * the insertions up to insertpos have already finished, because
3103 * that's what the earlier WaitXLogInsertionsToFinish() returned.
3104 * We're only calling it again to allow insertpos to be moved
3105 * further forward, not to actually wait for anyone.
3107 insertpos = WaitXLogInsertionsToFinish(insertpos);
3110 /* try to write/flush later additions to XLOG as well */
3111 WriteRqst.Write = insertpos;
3112 WriteRqst.Flush = insertpos;
3114 XLogWrite(WriteRqst, false);
3116 LWLockRelease(WALWriteLock);
3123 /* wake up walsenders now that we've released heavily contended locks */
3124 WalSndWakeupProcessRequests();
3127 * If we still haven't flushed to the request point then we have a
3128 * problem; most likely, the requested flush point is past end of XLOG.
3129 * This has been seen to occur when a disk page has a corrupted LSN.
3131 * Formerly we treated this as a PANIC condition, but that hurts the
3132 * system's robustness rather than helping it: we do not want to take down
3133 * the whole system due to corruption on one data page. In particular, if
3134 * the bad page is encountered again during recovery then we would be
3135 * unable to restart the database at all! (This scenario actually
3136 * happened in the field several times with 7.1 releases.) As of 8.4, bad
3137 * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
3138 * the only time we can reach here during recovery is while flushing the
3139 * end-of-recovery checkpoint record, and we don't expect that to have a
3142 * Note that for calls from xact.c, the ERROR will be promoted to PANIC
3143 * since xact.c calls this routine inside a critical section. However,
3144 * calls from bufmgr.c are not within critical sections and so we will not
3145 * force a restart for a bad LSN on a data page.
3147 if (LogwrtResult.Flush < record)
3149 "xlog flush request %X/%X is not satisfied --- flushed only to %X/%X",
3150 (uint32) (record >> 32), (uint32) record,
3151 (uint32) (LogwrtResult.Flush >> 32), (uint32) LogwrtResult.Flush);
3155 * Flush xlog, but without specifying exactly where to flush to.
3157 * We normally flush only completed blocks; but if there is nothing to do on
3158 * that basis, we check for unflushed async commits in the current incomplete
3159 * block, and flush through the latest one of those. Thus, if async commits
3160 * are not being used, we will flush complete blocks only. We can guarantee
3161 * that async commits reach disk after at most three cycles; normally only
3162 * one or two. (When flushing complete blocks, we allow XLogWrite to write
3163 * "flexibly", meaning it can stop at the end of the buffer ring; this makes a
3164 * difference only with very high load or long wal_writer_delay, but imposes
3165 * one extra cycle for the worst case for async commits.)
3167 * This routine is invoked periodically by the background walwriter process.
3169 * Returns TRUE if we flushed anything.
3172 XLogBackgroundFlush(void)
3174 XLogRecPtr WriteRqstPtr;
3175 bool flexible = true;
3176 bool wrote_something = false;
3178 /* XLOG doesn't need flushing during recovery */
3179 if (RecoveryInProgress())
3182 /* read LogwrtResult and update local state */
3184 /* use volatile pointer to prevent code rearrangement */
3185 volatile XLogCtlData *xlogctl = XLogCtl;
3187 SpinLockAcquire(&xlogctl->info_lck);
3188 LogwrtResult = xlogctl->LogwrtResult;
3189 WriteRqstPtr = xlogctl->LogwrtRqst.Write;
3190 SpinLockRelease(&xlogctl->info_lck);
3193 /* back off to last completed page boundary */
3194 WriteRqstPtr -= WriteRqstPtr % XLOG_BLCKSZ;
3196 /* if we have already flushed that far, consider async commit records */
3197 if (WriteRqstPtr <= LogwrtResult.Flush)
3199 /* use volatile pointer to prevent code rearrangement */
3200 volatile XLogCtlData *xlogctl = XLogCtl;
3202 SpinLockAcquire(&xlogctl->info_lck);
3203 WriteRqstPtr = xlogctl->asyncXactLSN;
3204 SpinLockRelease(&xlogctl->info_lck);
3205 flexible = false; /* ensure it all gets written */
3209 * If already known flushed, we're done. Just need to check if we are
3210 * holding an open file handle to a logfile that's no longer in use,
3211 * preventing the file from being deleted.
3213 if (WriteRqstPtr <= LogwrtResult.Flush)
3215 if (openLogFile >= 0)
3217 if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo))
3227 elog(LOG, "xlog bg flush request %X/%X; write %X/%X; flush %X/%X",
3228 (uint32) (WriteRqstPtr >> 32), (uint32) WriteRqstPtr,
3229 (uint32) (LogwrtResult.Write >> 32), (uint32) LogwrtResult.Write,
3230 (uint32) (LogwrtResult.Flush >> 32), (uint32) LogwrtResult.Flush);
3233 START_CRIT_SECTION();
3235 /* now wait for any in-progress insertions to finish and get write lock */
3236 WaitXLogInsertionsToFinish(WriteRqstPtr);
3237 LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3238 LogwrtResult = XLogCtl->LogwrtResult;
3239 if (WriteRqstPtr > LogwrtResult.Flush)
3241 XLogwrtRqst WriteRqst;
3243 WriteRqst.Write = WriteRqstPtr;
3244 WriteRqst.Flush = WriteRqstPtr;
3245 XLogWrite(WriteRqst, flexible);
3246 wrote_something = true;
3248 LWLockRelease(WALWriteLock);
3252 /* wake up walsenders now that we've released heavily contended locks */
3253 WalSndWakeupProcessRequests();
3256 * Great, done. To take some work off the critical path, try to initialize
3257 * as many of the no-longer-needed WAL buffers for future use as we can.
3259 AdvanceXLInsertBuffer(InvalidXLogRecPtr, true);
3261 return wrote_something;
3265 * Test whether XLOG data has been flushed up to (at least) the given position.
3267 * Returns true if a flush is still needed. (It may be that someone else
3268 * is already in process of flushing that far, however.)
3271 XLogNeedsFlush(XLogRecPtr record)
3274 * During recovery, we don't flush WAL but update minRecoveryPoint
3275 * instead. So "needs flush" is taken to mean whether minRecoveryPoint
3276 * would need to be updated.
3278 if (RecoveryInProgress())
3280 /* Quick exit if already known updated */
3281 if (record <= minRecoveryPoint || !updateMinRecoveryPoint)
3285 * Update local copy of minRecoveryPoint. But if the lock is busy,
3286 * just return a conservative guess.
3288 if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3290 minRecoveryPoint = ControlFile->minRecoveryPoint;
3291 minRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3292 LWLockRelease(ControlFileLock);
3295 * An invalid minRecoveryPoint means that we need to recover all the
3296 * WAL, i.e., we're doing crash recovery. We never modify the control
3297 * file's value in that case, so we can short-circuit future checks
3300 if (minRecoveryPoint == 0)
3301 updateMinRecoveryPoint = false;
3304 if (record <= minRecoveryPoint || !updateMinRecoveryPoint)
3310 /* Quick exit if already known flushed */
3311 if (record <= LogwrtResult.Flush)
3314 /* read LogwrtResult and update local state */
3316 /* use volatile pointer to prevent code rearrangement */
3317 volatile XLogCtlData *xlogctl = XLogCtl;
3319 SpinLockAcquire(&xlogctl->info_lck);
3320 LogwrtResult = xlogctl->LogwrtResult;
3321 SpinLockRelease(&xlogctl->info_lck);
3325 if (record <= LogwrtResult.Flush)
3332 * Create a new XLOG file segment, or open a pre-existing one.
3334 * log, seg: identify segment to be created/opened.
3336 * *use_existent: if TRUE, OK to use a pre-existing file (else, any
3337 * pre-existing file will be deleted). On return, TRUE if a pre-existing
3340 * use_lock: if TRUE, acquire ControlFileLock while moving file into
3341 * place. This should be TRUE except during bootstrap log creation. The
3342 * caller must *not* hold the lock at call.
3344 * Returns FD of opened file.
3346 * Note: errors here are ERROR not PANIC because we might or might not be
3347 * inside a critical section (eg, during checkpoint there is no reason to
3348 * take down the system on failure). They will promote to PANIC if we are
3349 * in a critical section.
3352 XLogFileInit(XLogSegNo logsegno, bool *use_existent, bool use_lock)
3354 char path[MAXPGPATH];
3355 char tmppath[MAXPGPATH];
3357 XLogSegNo installed_segno;
3362 XLogFilePath(path, ThisTimeLineID, logsegno);
3365 * Try to use existent file (checkpoint maker may have created it already)
3369 fd = BasicOpenFile(path, O_RDWR | PG_BINARY | get_sync_bit(sync_method),
3373 if (errno != ENOENT)
3375 (errcode_for_file_access(),
3376 errmsg("could not open file \"%s\": %m", path)));
3383 * Initialize an empty (all zeroes) segment. NOTE: it is possible that
3384 * another process is doing the same thing. If so, we will end up
3385 * pre-creating an extra log segment. That seems OK, and better than
3386 * holding the lock throughout this lengthy process.
3388 elog(DEBUG2, "creating and filling new WAL file");
3390 snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3395 * Allocate a buffer full of zeros. This is done before opening the file
3396 * so that we don't leak the file descriptor if palloc fails.
3398 * Note: palloc zbuffer, instead of just using a local char array, to
3399 * ensure it is reasonably well-aligned; this may save a few cycles
3400 * transferring data to the kernel.
3402 zbuffer = (char *) palloc0(XLOG_BLCKSZ);
3404 /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3405 fd = BasicOpenFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY,
3409 (errcode_for_file_access(),
3410 errmsg("could not create file \"%s\": %m", tmppath)));
3413 * Zero-fill the file. We have to do this the hard way to ensure that all
3414 * the file space has really been allocated --- on platforms that allow
3415 * "holes" in files, just seeking to the end doesn't allocate intermediate
3416 * space. This way, we know that we have all the space and (after the
3417 * fsync below) that all the indirect blocks are down on disk. Therefore,
3418 * fdatasync(2) or O_DSYNC will be sufficient to sync future writes to the
3421 for (nbytes = 0; nbytes < XLogSegSize; nbytes += XLOG_BLCKSZ)
3424 if ((int) write(fd, zbuffer, XLOG_BLCKSZ) != (int) XLOG_BLCKSZ)
3426 int save_errno = errno;
3429 * If we fail to make the file, delete it to release disk space
3435 /* if write didn't set errno, assume problem is no disk space */
3436 errno = save_errno ? save_errno : ENOSPC;
3439 (errcode_for_file_access(),
3440 errmsg("could not write to file \"%s\": %m", tmppath)));
3445 if (pg_fsync(fd) != 0)
3449 (errcode_for_file_access(),
3450 errmsg("could not fsync file \"%s\": %m", tmppath)));
3455 (errcode_for_file_access(),
3456 errmsg("could not close file \"%s\": %m", tmppath)));
3459 * Now move the segment into place with its final name.
3461 * If caller didn't want to use a pre-existing file, get rid of any
3462 * pre-existing file. Otherwise, cope with possibility that someone else
3463 * has created the file while we were filling ours: if so, use ours to
3464 * pre-create a future log segment.
3466 installed_segno = logsegno;
3467 max_advance = XLOGfileslop;
3468 if (!InstallXLogFileSegment(&installed_segno, tmppath,
3469 *use_existent, &max_advance,
3473 * No need for any more future segments, or InstallXLogFileSegment()
3474 * failed to rename the file into place. If the rename failed, opening
3475 * the file below will fail.
3480 /* Set flag to tell caller there was no existent file */
3481 *use_existent = false;
3483 /* Now open original target segment (might not be file I just made) */
3484 fd = BasicOpenFile(path, O_RDWR | PG_BINARY | get_sync_bit(sync_method),
3488 (errcode_for_file_access(),
3489 errmsg("could not open file \"%s\": %m", path)));
3491 elog(DEBUG2, "done creating and filling new WAL file");
3497 * Create a new XLOG file segment by copying a pre-existing one.
3499 * destsegno: identify segment to be created.
3501 * srcTLI, srclog, srcseg: identify segment to be copied (could be from
3502 * a different timeline)
3504 * Currently this is only used during recovery, and so there are no locking
3505 * considerations. But we should be just as tense as XLogFileInit to avoid
3506 * emplacing a bogus file.
3509 XLogFileCopy(XLogSegNo destsegno, TimeLineID srcTLI, XLogSegNo srcsegno)
3511 char path[MAXPGPATH];
3512 char tmppath[MAXPGPATH];
3513 char buffer[XLOG_BLCKSZ];
3519 * Open the source file
3521 XLogFilePath(path, srcTLI, srcsegno);
3522 srcfd = OpenTransientFile(path, O_RDONLY | PG_BINARY, 0);
3525 (errcode_for_file_access(),
3526 errmsg("could not open file \"%s\": %m", path)));
3529 * Copy into a temp file name.
3531 snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3535 /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3536 fd = OpenTransientFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY,
3540 (errcode_for_file_access(),
3541 errmsg("could not create file \"%s\": %m", tmppath)));
3544 * Do the data copying.
3546 for (nbytes = 0; nbytes < XLogSegSize; nbytes += sizeof(buffer))
3549 if ((int) read(srcfd, buffer, sizeof(buffer)) != (int) sizeof(buffer))
3553 (errcode_for_file_access(),
3554 errmsg("could not read file \"%s\": %m", path)));
3557 (errmsg("not enough data in file \"%s\"", path)));
3560 if ((int) write(fd, buffer, sizeof(buffer)) != (int) sizeof(buffer))
3562 int save_errno = errno;
3565 * If we fail to make the file, delete it to release disk space
3568 /* if write didn't set errno, assume problem is no disk space */
3569 errno = save_errno ? save_errno : ENOSPC;
3572 (errcode_for_file_access(),
3573 errmsg("could not write to file \"%s\": %m", tmppath)));
3577 if (pg_fsync(fd) != 0)
3579 (errcode_for_file_access(),
3580 errmsg("could not fsync file \"%s\": %m", tmppath)));
3582 if (CloseTransientFile(fd))
3584 (errcode_for_file_access(),
3585 errmsg("could not close file \"%s\": %m", tmppath)));
3587 CloseTransientFile(srcfd);
3590 * Now move the segment into place with its final name.
3592 if (!InstallXLogFileSegment(&destsegno, tmppath, false, NULL, false))
3593 elog(ERROR, "InstallXLogFileSegment should not have failed");
3597 * Install a new XLOG segment file as a current or future log segment.
3599 * This is used both to install a newly-created segment (which has a temp
3600 * filename while it's being created) and to recycle an old segment.
3602 * *segno: identify segment to install as (or first possible target).
3603 * When find_free is TRUE, this is modified on return to indicate the
3604 * actual installation location or last segment searched.
3606 * tmppath: initial name of file to install. It will be renamed into place.
3608 * find_free: if TRUE, install the new segment at the first empty segno
3609 * number at or after the passed numbers. If FALSE, install the new segment
3610 * exactly where specified, deleting any existing segment file there.
3612 * *max_advance: maximum number of segno slots to advance past the starting
3613 * point. Fail if no free slot is found in this range. On return, reduced
3614 * by the number of slots skipped over. (Irrelevant, and may be NULL,
3615 * when find_free is FALSE.)
3617 * use_lock: if TRUE, acquire ControlFileLock while moving file into
3618 * place. This should be TRUE except during bootstrap log creation. The
3619 * caller must *not* hold the lock at call.
3621 * Returns TRUE if the file was installed successfully. FALSE indicates that
3622 * max_advance limit was exceeded, or an error occurred while renaming the
3626 InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
3627 bool find_free, int *max_advance,
3630 char path[MAXPGPATH];
3631 struct stat stat_buf;
3633 XLogFilePath(path, ThisTimeLineID, *segno);
3636 * We want to be sure that only one process does this at a time.
3639 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3643 /* Force installation: get rid of any pre-existing segment file */
3648 /* Find a free slot to put it in */
3649 while (stat(path, &stat_buf) == 0)
3651 if (*max_advance <= 0)
3653 /* Failed to find a free slot within specified range */
3655 LWLockRelease(ControlFileLock);
3660 XLogFilePath(path, ThisTimeLineID, *segno);
3665 * Prefer link() to rename() here just to be really sure that we don't
3666 * overwrite an existing logfile. However, there shouldn't be one, so
3667 * rename() is an acceptable substitute except for the truly paranoid.
3669 #if HAVE_WORKING_LINK
3670 if (link(tmppath, path) < 0)
3673 LWLockRelease(ControlFileLock);
3675 (errcode_for_file_access(),
3676 errmsg("could not link file \"%s\" to \"%s\" (initialization of log file): %m",
3682 if (rename(tmppath, path) < 0)
3685 LWLockRelease(ControlFileLock);
3687 (errcode_for_file_access(),
3688 errmsg("could not rename file \"%s\" to \"%s\" (initialization of log file): %m",
3695 LWLockRelease(ControlFileLock);
3701 * Open a pre-existing logfile segment for writing.
3704 XLogFileOpen(XLogSegNo segno)
3706 char path[MAXPGPATH];
3709 XLogFilePath(path, ThisTimeLineID, segno);
3711 fd = BasicOpenFile(path, O_RDWR | PG_BINARY | get_sync_bit(sync_method),
3715 (errcode_for_file_access(),
3716 errmsg("could not open transaction log file \"%s\": %m", path)));
3722 * Open a logfile segment for reading (during recovery).
3724 * If source == XLOG_FROM_ARCHIVE, the segment is retrieved from archive.
3725 * Otherwise, it's assumed to be already available in pg_xlog.
3728 XLogFileRead(XLogSegNo segno, int emode, TimeLineID tli,
3729 int source, bool notfoundOk)
3731 char xlogfname[MAXFNAMELEN];
3732 char activitymsg[MAXFNAMELEN + 16];
3733 char path[MAXPGPATH];
3736 XLogFileName(xlogfname, tli, segno);
3740 case XLOG_FROM_ARCHIVE:
3741 /* Report recovery progress in PS display */
3742 snprintf(activitymsg, sizeof(activitymsg), "waiting for %s",
3744 set_ps_display(activitymsg, false);
3746 restoredFromArchive = RestoreArchivedFile(path, xlogfname,
3750 if (!restoredFromArchive)
3754 case XLOG_FROM_PG_XLOG:
3755 case XLOG_FROM_STREAM:
3756 XLogFilePath(path, tli, segno);
3757 restoredFromArchive = false;
3761 elog(ERROR, "invalid XLogFileRead source %d", source);
3765 * If the segment was fetched from archival storage, replace the existing
3766 * xlog segment (if any) with the archival version.
3768 if (source == XLOG_FROM_ARCHIVE)
3770 KeepFileRestoredFromArchive(path, xlogfname);
3773 * Set path to point at the new file in pg_xlog.
3775 snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlogfname);
3778 fd = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
3784 /* Report recovery progress in PS display */
3785 snprintf(activitymsg, sizeof(activitymsg), "recovering %s",
3787 set_ps_display(activitymsg, false);
3789 /* Track source of data in assorted state variables */
3790 readSource = source;
3791 XLogReceiptSource = source;
3792 /* In FROM_STREAM case, caller tracks receipt time, not me */
3793 if (source != XLOG_FROM_STREAM)
3794 XLogReceiptTime = GetCurrentTimestamp();
3798 if (errno != ENOENT || !notfoundOk) /* unexpected failure? */
3800 (errcode_for_file_access(),
3801 errmsg("could not open file \"%s\": %m", path)));
3806 * Open a logfile segment for reading (during recovery).
3808 * This version searches for the segment with any TLI listed in expectedTLEs.
3811 XLogFileReadAnyTLI(XLogSegNo segno, int emode, int source)
3813 char path[MAXPGPATH];
3819 * Loop looking for a suitable timeline ID: we might need to read any of
3820 * the timelines listed in expectedTLEs.
3822 * We expect curFileTLI on entry to be the TLI of the preceding file in
3823 * sequence, or 0 if there was no predecessor. We do not allow curFileTLI
3824 * to go backwards; this prevents us from picking up the wrong file when a
3825 * parent timeline extends to higher segment numbers than the child we
3828 * If we haven't read the timeline history file yet, read it now, so that
3829 * we know which TLIs to scan. We don't save the list in expectedTLEs,
3830 * however, unless we actually find a valid segment. That way if there is
3831 * neither a timeline history file nor a WAL segment in the archive, and
3832 * streaming replication is set up, we'll read the timeline history file
3833 * streamed from the master when we start streaming, instead of recovering
3834 * with a dummy history generated here.
3837 tles = expectedTLEs;
3839 tles = readTimeLineHistory(recoveryTargetTLI);
3843 TimeLineID tli = ((TimeLineHistoryEntry *) lfirst(cell))->tli;
3845 if (tli < curFileTLI)
3846 break; /* don't bother looking at too-old TLIs */
3848 if (source == XLOG_FROM_ANY || source == XLOG_FROM_ARCHIVE)
3850 fd = XLogFileRead(segno, emode, tli,
3851 XLOG_FROM_ARCHIVE, true);
3854 elog(DEBUG1, "got WAL segment from archive");
3856 expectedTLEs = tles;
3861 if (source == XLOG_FROM_ANY || source == XLOG_FROM_PG_XLOG)
3863 fd = XLogFileRead(segno, emode, tli,
3864 XLOG_FROM_PG_XLOG, true);
3868 expectedTLEs = tles;
3874 /* Couldn't find it. For simplicity, complain about front timeline */
3875 XLogFilePath(path, recoveryTargetTLI, segno);
3878 (errcode_for_file_access(),
3879 errmsg("could not open file \"%s\": %m", path)));
3884 * Close the current logfile segment for writing.
3889 Assert(openLogFile >= 0);
3892 * WAL segment files will not be re-read in normal operation, so we advise
3893 * the OS to release any cached pages. But do not do so if WAL archiving
3894 * or streaming is active, because archiver and walsender process could
3895 * use the cache to read the WAL segment.
3897 #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3898 if (!XLogIsNeeded())
3899 (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3902 if (close(openLogFile))
3904 (errcode_for_file_access(),
3905 errmsg("could not close log file %s: %m",
3906 XLogFileNameP(ThisTimeLineID, openLogSegNo))));
3911 * Preallocate log files beyond the specified log endpoint.
3913 * XXX this is currently extremely conservative, since it forces only one
3914 * future log segment to exist, and even that only if we are 75% done with
3915 * the current one. This is only appropriate for very low-WAL-volume systems.
3916 * High-volume systems will be OK once they've built up a sufficient set of
3917 * recycled log segments, but the startup transient is likely to include
3918 * a lot of segment creations by foreground processes, which is not so good.
3921 PreallocXlogFiles(XLogRecPtr endptr)
3923 XLogSegNo _logSegNo;
3927 XLByteToPrevSeg(endptr, _logSegNo);
3928 if ((endptr - 1) % XLogSegSize >= (uint32) (0.75 * XLogSegSize))
3931 use_existent = true;
3932 lf = XLogFileInit(_logSegNo, &use_existent, true);
3935 CheckpointStats.ckpt_segs_added++;
3940 * Throws an error if the given log segment has already been removed or
3941 * recycled. The caller should only pass a segment that it knows to have
3942 * existed while the server has been running, as this function always
3943 * succeeds if no WAL segments have been removed since startup.
3944 * 'tli' is only used in the error message.
3947 CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3949 /* use volatile pointer to prevent code rearrangement */
3950 volatile XLogCtlData *xlogctl = XLogCtl;
3951 XLogSegNo lastRemovedSegNo;
3953 SpinLockAcquire(&xlogctl->info_lck);
3954 lastRemovedSegNo = xlogctl->lastRemovedSegNo;
3955 SpinLockRelease(&xlogctl->info_lck);
3957 if (segno <= lastRemovedSegNo)
3959 char filename[MAXFNAMELEN];
3961 XLogFileName(filename, tli, segno);
3963 (errcode_for_file_access(),
3964 errmsg("requested WAL segment %s has already been removed",
3970 * Update the last removed segno pointer in shared memory, to reflect
3971 * that the given XLOG file has been removed.
3974 UpdateLastRemovedPtr(char *filename)
3976 /* use volatile pointer to prevent code rearrangement */
3977 volatile XLogCtlData *xlogctl = XLogCtl;
3981 XLogFromFileName(filename, &tli, &segno);
3983 SpinLockAcquire(&xlogctl->info_lck);
3984 if (segno > xlogctl->lastRemovedSegNo)
3985 xlogctl->lastRemovedSegNo = segno;
3986 SpinLockRelease(&xlogctl->info_lck);
3990 * Recycle or remove all log files older or equal to passed segno
3992 * endptr is current (or recent) end of xlog; this is used to determine
3993 * whether we want to recycle rather than delete no-longer-wanted log files.
3996 RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr endptr)
3998 XLogSegNo endlogSegNo;
4001 struct dirent *xlde;
4002 char lastoff[MAXFNAMELEN];
4003 char path[MAXPGPATH];
4006 char newpath[MAXPGPATH];
4008 struct stat statbuf;
4011 * Initialize info about where to try to recycle to. We allow recycling
4012 * segments up to XLOGfileslop segments beyond the current XLOG location.
4014 XLByteToPrevSeg(endptr, endlogSegNo);
4015 max_advance = XLOGfileslop;
4017 xldir = AllocateDir(XLOGDIR);
4020 (errcode_for_file_access(),
4021 errmsg("could not open transaction log directory \"%s\": %m",
4025 * Construct a filename of the last segment to be kept. The timeline ID
4026 * doesn't matter, we ignore that in the comparison. (During recovery,
4027 * ThisTimeLineID isn't set, so we can't use that.)
4029 XLogFileName(lastoff, 0, segno);
4031 elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
4034 while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4037 * We ignore the timeline part of the XLOG segment identifiers in
4038 * deciding whether a segment is still needed. This ensures that we
4039 * won't prematurely remove a segment from a parent timeline. We could
4040 * probably be a little more proactive about removing segments of
4041 * non-parent timelines, but that would be a whole lot more
4044 * We use the alphanumeric sorting property of the filenames to decide
4045 * which ones are earlier than the lastoff segment.
4047 if (strlen(xlde->d_name) == 24 &&
4048 strspn(xlde->d_name, "0123456789ABCDEF") == 24 &&
4049 strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
4051 if (XLogArchiveCheckDone(xlde->d_name))
4053 snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
4055 /* Update the last removed location in shared memory first */
4056 UpdateLastRemovedPtr(xlde->d_name);
4059 * Before deleting the file, see if it can be recycled as a
4060 * future log segment. Only recycle normal files, pg_standby
4061 * for example can create symbolic links pointing to a
4062 * separate archive directory.
4064 if (lstat(path, &statbuf) == 0 && S_ISREG(statbuf.st_mode) &&
4065 InstallXLogFileSegment(&endlogSegNo, path,
4066 true, &max_advance, true))
4069 (errmsg("recycled transaction log file \"%s\"",
4071 CheckpointStats.ckpt_segs_recycled++;
4072 /* Needn't recheck that slot on future iterations */
4073 if (max_advance > 0)
4081 /* No need for any more future segments... */
4085 (errmsg("removing transaction log file \"%s\"",
4091 * On Windows, if another process (e.g another backend)
4092 * holds the file open in FILE_SHARE_DELETE mode, unlink
4093 * will succeed, but the file will still show up in
4094 * directory listing until the last handle is closed. To
4095 * avoid confusing the lingering deleted file for a live
4096 * WAL file that needs to be archived, rename it before
4099 * If another process holds the file open without
4100 * FILE_SHARE_DELETE flag, rename will fail. We'll try
4101 * again at the next checkpoint.
4103 snprintf(newpath, MAXPGPATH, "%s.deleted", path);
4104 if (rename(path, newpath) != 0)
4107 (errcode_for_file_access(),
4108 errmsg("could not rename old transaction log file \"%s\": %m",
4112 rc = unlink(newpath);
4119 (errcode_for_file_access(),
4120 errmsg("could not remove old transaction log file \"%s\": %m",
4124 CheckpointStats.ckpt_segs_removed++;
4127 XLogArchiveCleanup(xlde->d_name);
4136 * Verify whether pg_xlog and pg_xlog/archive_status exist.
4137 * If the latter does not exist, recreate it.
4139 * It is not the goal of this function to verify the contents of these
4140 * directories, but to help in cases where someone has performed a cluster
4141 * copy for PITR purposes but omitted pg_xlog from the copy.
4143 * We could also recreate pg_xlog if it doesn't exist, but a deliberate
4144 * policy decision was made not to. It is fairly common for pg_xlog to be
4145 * a symlink, and if that was the DBA's intent then automatically making a
4146 * plain directory would result in degraded performance with no notice.
4149 ValidateXLOGDirectoryStructure(void)
4151 char path[MAXPGPATH];
4152 struct stat stat_buf;
4154 /* Check for pg_xlog; if it doesn't exist, error out */
4155 if (stat(XLOGDIR, &stat_buf) != 0 ||
4156 !S_ISDIR(stat_buf.st_mode))
4158 (errmsg("required WAL directory \"%s\" does not exist",
4161 /* Check for archive_status */
4162 snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4163 if (stat(path, &stat_buf) == 0)
4165 /* Check for weird cases where it exists but isn't a directory */
4166 if (!S_ISDIR(stat_buf.st_mode))
4168 (errmsg("required WAL directory \"%s\" does not exist",
4174 (errmsg("creating missing WAL directory \"%s\"", path)));
4175 if (mkdir(path, S_IRWXU) < 0)
4177 (errmsg("could not create missing directory \"%s\": %m",
4183 * Remove previous backup history files. This also retries creation of
4184 * .ready files for any backup history files for which XLogArchiveNotify
4188 CleanupBackupHistory(void)
4191 struct dirent *xlde;
4192 char path[MAXPGPATH];
4194 xldir = AllocateDir(XLOGDIR);
4197 (errcode_for_file_access(),
4198 errmsg("could not open transaction log directory \"%s\": %m",
4201 while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4203 if (strlen(xlde->d_name) > 24 &&
4204 strspn(xlde->d_name, "0123456789ABCDEF") == 24 &&
4205 strcmp(xlde->d_name + strlen(xlde->d_name) - strlen(".backup"),
4208 if (XLogArchiveCheckDone(xlde->d_name))
4211 (errmsg("removing transaction log backup history file \"%s\"",
4213 snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
4215 XLogArchiveCleanup(xlde->d_name);
4224 * Restore a full-page image from a backup block attached to an XLOG record.
4226 * lsn: LSN of the XLOG record being replayed
4227 * record: the complete XLOG record
4228 * block_index: which backup block to restore (0 .. XLR_MAX_BKP_BLOCKS - 1)
4229 * get_cleanup_lock: TRUE to get a cleanup rather than plain exclusive lock
4230 * keep_buffer: TRUE to return the buffer still locked and pinned
4232 * Returns the buffer number containing the page. Note this is not terribly
4233 * useful unless keep_buffer is specified as TRUE.
4235 * Note: when a backup block is available in XLOG, we restore it
4236 * unconditionally, even if the page in the database appears newer.
4237 * This is to protect ourselves against database pages that were partially
4238 * or incorrectly written during a crash. We assume that the XLOG data
4239 * must be good because it has passed a CRC check, while the database
4240 * page might not be. This will force us to replay all subsequent
4241 * modifications of the page that appear in XLOG, rather than possibly
4242 * ignoring them as already applied, but that's not a huge drawback.
4244 * If 'get_cleanup_lock' is true, a cleanup lock is obtained on the buffer,
4245 * else a normal exclusive lock is used. During crash recovery, that's just
4246 * pro forma because there can't be any regular backends in the system, but
4247 * in hot standby mode the distinction is important.
4249 * If 'keep_buffer' is true, return without releasing the buffer lock and pin;
4250 * then caller is responsible for doing UnlockReleaseBuffer() later. This
4251 * is needed in some cases when replaying XLOG records that touch multiple
4252 * pages, to prevent inconsistent states from being visible to other backends.
4253 * (Again, that's only important in hot standby mode.)
4256 RestoreBackupBlock(XLogRecPtr lsn, XLogRecord *record, int block_index,
4257 bool get_cleanup_lock, bool keep_buffer)
4263 /* Locate requested BkpBlock in the record */
4264 blk = (char *) XLogRecGetData(record) + record->xl_len;
4265 for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
4267 if (!(record->xl_info & XLR_BKP_BLOCK(i)))
4270 memcpy(&bkpb, blk, sizeof(BkpBlock));
4271 blk += sizeof(BkpBlock);
4273 if (i == block_index)
4275 /* Found it, apply the update */
4276 return RestoreBackupBlockContents(lsn, bkpb, blk, get_cleanup_lock,
4280 blk += BLCKSZ - bkpb.hole_length;
4283 /* Caller specified a bogus block_index */
4284 elog(ERROR, "failed to restore block_index %d", block_index);
4285 return InvalidBuffer; /* keep compiler quiet */
4289 * Workhorse for RestoreBackupBlock usable without an xlog record
4291 * Restores a full-page image from BkpBlock and a data pointer.
4294 RestoreBackupBlockContents(XLogRecPtr lsn, BkpBlock bkpb, char *blk,
4295 bool get_cleanup_lock, bool keep_buffer)
4300 buffer = XLogReadBufferExtended(bkpb.node, bkpb.fork, bkpb.block,
4302 Assert(BufferIsValid(buffer));
4303 if (get_cleanup_lock)
4304 LockBufferForCleanup(buffer);
4306 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
4308 page = (Page) BufferGetPage(buffer);
4310 if (bkpb.hole_length == 0)
4312 memcpy((char *) page, blk, BLCKSZ);
4316 memcpy((char *) page, blk, bkpb.hole_offset);
4317 /* must zero-fill the hole */
4318 MemSet((char *) page + bkpb.hole_offset, 0, bkpb.hole_length);
4319 memcpy((char *) page + (bkpb.hole_offset + bkpb.hole_length),
4320 blk + bkpb.hole_offset,
4321 BLCKSZ - (bkpb.hole_offset + bkpb.hole_length));
4325 * The checksum value on this page is currently invalid. We don't need to
4326 * reset it here since it will be set before being written.
4329 PageSetLSN(page, lsn);
4330 MarkBufferDirty(buffer);
4333 UnlockReleaseBuffer(buffer);
4339 * Attempt to read an XLOG record.
4341 * If RecPtr is not NULL, try to read a record at that position. Otherwise
4342 * try to read a record just after the last one previously read.
4344 * If no valid record is available, returns NULL, or fails if emode is PANIC.
4345 * (emode must be either PANIC, LOG). In standby mode, retries until a valid
4346 * record is available.
4348 * The record is copied into readRecordBuf, so that on successful return,
4349 * the returned record pointer always points there.
4352 ReadRecord(XLogReaderState *xlogreader, XLogRecPtr RecPtr, int emode,
4356 XLogPageReadPrivate *private = (XLogPageReadPrivate *) xlogreader->private_data;
4358 /* Pass through parameters to XLogPageRead */
4359 private->fetching_ckpt = fetching_ckpt;
4360 private->emode = emode;
4361 private->randAccess = (RecPtr != InvalidXLogRecPtr);
4363 /* This is the first attempt to read this page. */
4364 lastSourceFailed = false;
4370 record = XLogReadRecord(xlogreader, RecPtr, &errormsg);
4371 ReadRecPtr = xlogreader->ReadRecPtr;
4372 EndRecPtr = xlogreader->EndRecPtr;
4382 * We only end up here without a message when XLogPageRead()
4383 * failed - in that case we already logged something. In
4384 * StandbyMode that only happens if we have been triggered, so we
4385 * shouldn't loop anymore in that case.
4388 ereport(emode_for_corrupt_record(emode,
4389 RecPtr ? RecPtr : EndRecPtr),
4390 (errmsg_internal("%s", errormsg) /* already translated */ ));
4394 * Check page TLI is one of the expected values.
4396 else if (!tliInHistory(xlogreader->latestPageTLI, expectedTLEs))
4398 char fname[MAXFNAMELEN];
4402 XLByteToSeg(xlogreader->latestPagePtr, segno);
4403 offset = xlogreader->latestPagePtr % XLogSegSize;
4404 XLogFileName(fname, xlogreader->readPageTLI, segno);
4405 ereport(emode_for_corrupt_record(emode,
4406 RecPtr ? RecPtr : EndRecPtr),
4407 (errmsg("unexpected timeline ID %u in log segment %s, offset %u",
4408 xlogreader->latestPageTLI,
4416 /* Great, got a record */
4421 /* No valid record available from this source */
4422 lastSourceFailed = true;
4425 * If archive recovery was requested, but we were still doing
4426 * crash recovery, switch to archive recovery and retry using the
4427 * offline archive. We have now replayed all the valid WAL in
4428 * pg_xlog, so we are presumably now consistent.
4430 * We require that there's at least some valid WAL present in
4431 * pg_xlog, however (!fetch_ckpt). We could recover using the WAL
4432 * from the archive, even if pg_xlog is completely empty, but we'd
4433 * have no idea how far we'd have to replay to reach consistency.
4434 * So err on the safe side and give up.
4436 if (!InArchiveRecovery && ArchiveRecoveryRequested &&
4440 (errmsg_internal("reached end of WAL in pg_xlog, entering archive recovery")));
4441 InArchiveRecovery = true;
4442 if (StandbyModeRequested)
4445 /* initialize minRecoveryPoint to this record */
4446 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
4447 ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
4448 if (ControlFile->minRecoveryPoint < EndRecPtr)
4450 ControlFile->minRecoveryPoint = EndRecPtr;
4451 ControlFile->minRecoveryPointTLI = ThisTimeLineID;
4453 /* update local copy */
4454 minRecoveryPoint = ControlFile->minRecoveryPoint;
4455 minRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
4457 UpdateControlFile();
4458 LWLockRelease(ControlFileLock);
4460 CheckRecoveryConsistency();
4463 * Before we retry, reset lastSourceFailed and currentSource
4464 * so that we will check the archive next.
4466 lastSourceFailed = false;
4472 /* In standby mode, loop back to retry. Otherwise, give up. */
4473 if (StandbyMode && !CheckForStandbyTrigger())
4482 * Scan for new timelines that might have appeared in the archive since we
4485 * If there are any, the function changes recovery target TLI to the latest
4486 * one and returns 'true'.
4489 rescanLatestTimeLine(void)
4491 List *newExpectedTLEs;
4494 TimeLineID newtarget;
4495 TimeLineID oldtarget = recoveryTargetTLI;
4496 TimeLineHistoryEntry *currentTle = NULL;
4498 newtarget = findNewestTimeLine(recoveryTargetTLI);
4499 if (newtarget == recoveryTargetTLI)
4501 /* No new timelines found */
4506 * Determine the list of expected TLIs for the new TLI
4509 newExpectedTLEs = readTimeLineHistory(newtarget);
4512 * If the current timeline is not part of the history of the new timeline,
4513 * we cannot proceed to it.
4516 foreach(cell, newExpectedTLEs)
4518 currentTle = (TimeLineHistoryEntry *) lfirst(cell);
4520 if (currentTle->tli == recoveryTargetTLI)
4529 (errmsg("new timeline %u is not a child of database system timeline %u",
4536 * The current timeline was found in the history file, but check that the
4537 * next timeline was forked off from it *after* the current recovery
4540 if (currentTle->end < EndRecPtr)
4543 (errmsg("new timeline %u forked off current database system timeline %u before current recovery point %X/%X",
4546 (uint32) (EndRecPtr >> 32), (uint32) EndRecPtr)));
4550 /* The new timeline history seems valid. Switch target */
4551 recoveryTargetTLI = newtarget;
4552 list_free_deep(expectedTLEs);
4553 expectedTLEs = newExpectedTLEs;
4556 * As in StartupXLOG(), try to ensure we have all the history files
4557 * between the old target and new target in pg_xlog.
4559 restoreTimeLineHistoryFiles(oldtarget + 1, newtarget);
4562 (errmsg("new target timeline is %u",
4563 recoveryTargetTLI)));
4569 * I/O routines for pg_control
4571 * *ControlFile is a buffer in shared memory that holds an image of the
4572 * contents of pg_control. WriteControlFile() initializes pg_control
4573 * given a preloaded buffer, ReadControlFile() loads the buffer from
4574 * the pg_control file (during postmaster or standalone-backend startup),
4575 * and UpdateControlFile() rewrites pg_control after we modify xlog state.
4577 * For simplicity, WriteControlFile() initializes the fields of pg_control
4578 * that are related to checking backend/database compatibility, and
4579 * ReadControlFile() verifies they are correct. We could split out the
4580 * I/O and compatibility-check functions, but there seems no need currently.
4583 WriteControlFile(void)
4586 char buffer[PG_CONTROL_SIZE]; /* need not be aligned */
4589 * Initialize version and compatibility-check fields
4591 ControlFile->pg_control_version = PG_CONTROL_VERSION;
4592 ControlFile->catalog_version_no = CATALOG_VERSION_NO;
4594 ControlFile->maxAlign = MAXIMUM_ALIGNOF;
4595 ControlFile->floatFormat = FLOATFORMAT_VALUE;
4597 ControlFile->blcksz = BLCKSZ;
4598 ControlFile->relseg_size = RELSEG_SIZE;
4599 ControlFile->xlog_blcksz = XLOG_BLCKSZ;
4600 ControlFile->xlog_seg_size = XLOG_SEG_SIZE;
4602 ControlFile->nameDataLen = NAMEDATALEN;
4603 ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
4605 ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
4607 #ifdef HAVE_INT64_TIMESTAMP
4608 ControlFile->enableIntTimes = true;
4610 ControlFile->enableIntTimes = false;
4612 ControlFile->float4ByVal = FLOAT4PASSBYVAL;
4613 ControlFile->float8ByVal = FLOAT8PASSBYVAL;
4615 /* Contents are protected with a CRC */
4616 INIT_CRC32(ControlFile->crc);
4617 COMP_CRC32(ControlFile->crc,
4618 (char *) ControlFile,
4619 offsetof(ControlFileData, crc));
4620 FIN_CRC32(ControlFile->crc);
4623 * We write out PG_CONTROL_SIZE bytes into pg_control, zero-padding the
4624 * excess over sizeof(ControlFileData). This reduces the odds of
4625 * premature-EOF errors when reading pg_control. We'll still fail when we
4626 * check the contents of the file, but hopefully with a more specific
4627 * error than "couldn't read pg_control".
4629 if (sizeof(ControlFileData) > PG_CONTROL_SIZE)
4630 elog(PANIC, "sizeof(ControlFileData) is larger than PG_CONTROL_SIZE; fix either one");
4632 memset(buffer, 0, PG_CONTROL_SIZE);
4633 memcpy(buffer, ControlFile, sizeof(ControlFileData));
4635 fd = BasicOpenFile(XLOG_CONTROL_FILE,
4636 O_RDWR | O_CREAT | O_EXCL | PG_BINARY,
4640 (errcode_for_file_access(),
4641 errmsg("could not create control file \"%s\": %m",
4642 XLOG_CONTROL_FILE)));
4645 if (write(fd, buffer, PG_CONTROL_SIZE) != PG_CONTROL_SIZE)
4647 /* if write didn't set errno, assume problem is no disk space */
4651 (errcode_for_file_access(),
4652 errmsg("could not write to control file: %m")));
4655 if (pg_fsync(fd) != 0)
4657 (errcode_for_file_access(),
4658 errmsg("could not fsync control file: %m")));
4662 (errcode_for_file_access(),
4663 errmsg("could not close control file: %m")));
4667 ReadControlFile(void)
4675 fd = BasicOpenFile(XLOG_CONTROL_FILE,
4680 (errcode_for_file_access(),
4681 errmsg("could not open control file \"%s\": %m",
4682 XLOG_CONTROL_FILE)));
4684 if (read(fd, ControlFile, sizeof(ControlFileData)) != sizeof(ControlFileData))
4686 (errcode_for_file_access(),
4687 errmsg("could not read from control file: %m")));
4692 * Check for expected pg_control format version. If this is wrong, the
4693 * CRC check will likely fail because we'll be checking the wrong number
4694 * of bytes. Complaining about wrong version will probably be more
4695 * enlightening than complaining about wrong CRC.
4698 if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
4700 (errmsg("database files are incompatible with server"),
4701 errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4702 " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4703 ControlFile->pg_control_version, ControlFile->pg_control_version,
4704 PG_CONTROL_VERSION, PG_CONTROL_VERSION),
4705 errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4707 if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
4709 (errmsg("database files are incompatible with server"),
4710 errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4711 " but the server was compiled with PG_CONTROL_VERSION %d.",
4712 ControlFile->pg_control_version, PG_CONTROL_VERSION),
4713 errhint("It looks like you need to initdb.")));
4715 /* Now check the CRC. */
4718 (char *) ControlFile,
4719 offsetof(ControlFileData, crc));
4722 if (!EQ_CRC32(crc, ControlFile->crc))
4724 (errmsg("incorrect checksum in control file")));
4727 * Do compatibility checking immediately. If the database isn't
4728 * compatible with the backend executable, we want to abort before we can
4729 * possibly do any damage.
4731 if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
4733 (errmsg("database files are incompatible with server"),
4734 errdetail("The database cluster was initialized with CATALOG_VERSION_NO %d,"
4735 " but the server was compiled with CATALOG_VERSION_NO %d.",
4736 ControlFile->catalog_version_no, CATALOG_VERSION_NO),
4737 errhint("It looks like you need to initdb.")));
4738 if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
4740 (errmsg("database files are incompatible with server"),
4741 errdetail("The database cluster was initialized with MAXALIGN %d,"
4742 " but the server was compiled with MAXALIGN %d.",
4743 ControlFile->maxAlign, MAXIMUM_ALIGNOF),
4744 errhint("It looks like you need to initdb.")));
4745 if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
4747 (errmsg("database files are incompatible with server"),
4748 errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4749 errhint("It looks like you need to initdb.")));
4750 if (ControlFile->blcksz != BLCKSZ)
4752 (errmsg("database files are incompatible with server"),
4753 errdetail("The database cluster was initialized with BLCKSZ %d,"
4754 " but the server was compiled with BLCKSZ %d.",
4755 ControlFile->blcksz, BLCKSZ),
4756 errhint("It looks like you need to recompile or initdb.")));
4757 if (ControlFile->relseg_size != RELSEG_SIZE)
4759 (errmsg("database files are incompatible with server"),
4760 errdetail("The database cluster was initialized with RELSEG_SIZE %d,"
4761 " but the server was compiled with RELSEG_SIZE %d.",
4762 ControlFile->relseg_size, RELSEG_SIZE),
4763 errhint("It looks like you need to recompile or initdb.")));
4764 if (ControlFile->xlog_blcksz != XLOG_BLCKSZ)
4766 (errmsg("database files are incompatible with server"),
4767 errdetail("The database cluster was initialized with XLOG_BLCKSZ %d,"
4768 " but the server was compiled with XLOG_BLCKSZ %d.",
4769 ControlFile->xlog_blcksz, XLOG_BLCKSZ),
4770 errhint("It looks like you need to recompile or initdb.")));
4771 if (ControlFile->xlog_seg_size != XLOG_SEG_SIZE)
4773 (errmsg("database files are incompatible with server"),
4774 errdetail("The database cluster was initialized with XLOG_SEG_SIZE %d,"
4775 " but the server was compiled with XLOG_SEG_SIZE %d.",
4776 ControlFile->xlog_seg_size, XLOG_SEG_SIZE),
4777 errhint("It looks like you need to recompile or initdb.")));
4778 if (ControlFile->nameDataLen != NAMEDATALEN)
4780 (errmsg("database files are incompatible with server"),
4781 errdetail("The database cluster was initialized with NAMEDATALEN %d,"
4782 " but the server was compiled with NAMEDATALEN %d.",
4783 ControlFile->nameDataLen, NAMEDATALEN),
4784 errhint("It looks like you need to recompile or initdb.")));
4785 if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
4787 (errmsg("database files are incompatible with server"),
4788 errdetail("The database cluster was initialized with INDEX_MAX_KEYS %d,"
4789 " but the server was compiled with INDEX_MAX_KEYS %d.",
4790 ControlFile->indexMaxKeys, INDEX_MAX_KEYS),
4791 errhint("It looks like you need to recompile or initdb.")));
4792 if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
4794 (errmsg("database files are incompatible with server"),
4795 errdetail("The database cluster was initialized with TOAST_MAX_CHUNK_SIZE %d,"
4796 " but the server was compiled with TOAST_MAX_CHUNK_SIZE %d.",
4797 ControlFile->toast_max_chunk_size, (int) TOAST_MAX_CHUNK_SIZE),
4798 errhint("It looks like you need to recompile or initdb.")));
4800 #ifdef HAVE_INT64_TIMESTAMP
4801 if (ControlFile->enableIntTimes != true)
4803 (errmsg("database files are incompatible with server"),
4804 errdetail("The database cluster was initialized without HAVE_INT64_TIMESTAMP"
4805 " but the server was compiled with HAVE_INT64_TIMESTAMP."),
4806 errhint("It looks like you need to recompile or initdb.")));
4808 if (ControlFile->enableIntTimes != false)
4810 (errmsg("database files are incompatible with server"),
4811 errdetail("The database cluster was initialized with HAVE_INT64_TIMESTAMP"
4812 " but the server was compiled without HAVE_INT64_TIMESTAMP."),
4813 errhint("It looks like you need to recompile or initdb.")));
4816 #ifdef USE_FLOAT4_BYVAL
4817 if (ControlFile->float4ByVal != true)
4819 (errmsg("database files are incompatible with server"),
4820 errdetail("The database cluster was initialized without USE_FLOAT4_BYVAL"
4821 " but the server was compiled with USE_FLOAT4_BYVAL."),
4822 errhint("It looks like you need to recompile or initdb.")));
4824 if (ControlFile->float4ByVal != false)
4826 (errmsg("database files are incompatible with server"),
4827 errdetail("The database cluster was initialized with USE_FLOAT4_BYVAL"
4828 " but the server was compiled without USE_FLOAT4_BYVAL."),
4829 errhint("It looks like you need to recompile or initdb.")));
4832 #ifdef USE_FLOAT8_BYVAL
4833 if (ControlFile->float8ByVal != true)
4835 (errmsg("database files are incompatible with server"),
4836 errdetail("The database cluster was initialized without USE_FLOAT8_BYVAL"
4837 " but the server was compiled with USE_FLOAT8_BYVAL."),
4838 errhint("It looks like you need to recompile or initdb.")));
4840 if (ControlFile->float8ByVal != false)
4842 (errmsg("database files are incompatible with server"),
4843 errdetail("The database cluster was initialized with USE_FLOAT8_BYVAL"
4844 " but the server was compiled without USE_FLOAT8_BYVAL."),
4845 errhint("It looks like you need to recompile or initdb.")));
4848 /* Make the fixed settings visible as GUC variables, too */
4849 SetConfigOption("data_checksums", DataChecksumsEnabled() ? "yes" : "no",
4850 PGC_INTERNAL, PGC_S_OVERRIDE);
4854 UpdateControlFile(void)
4858 INIT_CRC32(ControlFile->crc);
4859 COMP_CRC32(ControlFile->crc,
4860 (char *) ControlFile,
4861 offsetof(ControlFileData, crc));
4862 FIN_CRC32(ControlFile->crc);
4864 fd = BasicOpenFile(XLOG_CONTROL_FILE,
4869 (errcode_for_file_access(),
4870 errmsg("could not open control file \"%s\": %m",
4871 XLOG_CONTROL_FILE)));
4874 if (write(fd, ControlFile, sizeof(ControlFileData)) != sizeof(ControlFileData))
4876 /* if write didn't set errno, assume problem is no disk space */
4880 (errcode_for_file_access(),
4881 errmsg("could not write to control file: %m")));
4884 if (pg_fsync(fd) != 0)
4886 (errcode_for_file_access(),
4887 errmsg("could not fsync control file: %m")));
4891 (errcode_for_file_access(),
4892 errmsg("could not close control file: %m")));
4896 * Returns the unique system identifier from control file.
4899 GetSystemIdentifier(void)
4901 Assert(ControlFile != NULL);
4902 return ControlFile->system_identifier;
4906 * Are checksums enabled for data pages?
4909 DataChecksumsEnabled(void)
4911 Assert(ControlFile != NULL);
4912 return (ControlFile->data_checksum_version > 0);
4916 * Returns a fake LSN for unlogged relations.
4918 * Each call generates an LSN that is greater than any previous value
4919 * returned. The current counter value is saved and restored across clean
4920 * shutdowns, but like unlogged relations, does not survive a crash. This can
4921 * be used in lieu of real LSN values returned by XLogInsert, if you need an
4922 * LSN-like increasing sequence of numbers without writing any WAL.
4925 GetFakeLSNForUnloggedRel(void)
4927 XLogRecPtr nextUnloggedLSN;
4929 /* use volatile pointer to prevent code rearrangement */
4930 volatile XLogCtlData *xlogctl = XLogCtl;
4932 /* increment the unloggedLSN counter, need SpinLock */
4933 SpinLockAcquire(&xlogctl->ulsn_lck);
4934 nextUnloggedLSN = xlogctl->unloggedLSN++;
4935 SpinLockRelease(&xlogctl->ulsn_lck);
4937 return nextUnloggedLSN;
4941 * Auto-tune the number of XLOG buffers.
4943 * The preferred setting for wal_buffers is about 3% of shared_buffers, with
4944 * a maximum of one XLOG segment (there is little reason to think that more
4945 * is helpful, at least so long as we force an fsync when switching log files)
4946 * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
4947 * 9.1, when auto-tuning was added).
4949 * This should not be called until NBuffers has received its final value.
4952 XLOGChooseNumBuffers(void)
4956 xbuffers = NBuffers / 32;
4957 if (xbuffers > XLOG_SEG_SIZE / XLOG_BLCKSZ)
4958 xbuffers = XLOG_SEG_SIZE / XLOG_BLCKSZ;
4965 * GUC check_hook for wal_buffers
4968 check_wal_buffers(int *newval, void **extra, GucSource source)
4971 * -1 indicates a request for auto-tune.
4976 * If we haven't yet changed the boot_val default of -1, just let it
4977 * be. We'll fix it when XLOGShmemSize is called.
4979 if (XLOGbuffers == -1)
4982 /* Otherwise, substitute the auto-tune value */
4983 *newval = XLOGChooseNumBuffers();
4987 * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
4988 * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
4989 * the case, we just silently treat such values as a request for the
4990 * minimum. (We could throw an error instead, but that doesn't seem very
5000 * Initialization of shared memory for XLOG
5008 * If the value of wal_buffers is -1, use the preferred auto-tune value.
5009 * This isn't an amazingly clean place to do this, but we must wait till
5010 * NBuffers has received its final value, and must do it before using the
5011 * value of XLOGbuffers to do anything important.
5013 if (XLOGbuffers == -1)
5017 snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
5018 SetConfigOption("wal_buffers", buf, PGC_POSTMASTER, PGC_S_OVERRIDE);
5020 Assert(XLOGbuffers > 0);
5023 size = sizeof(XLogCtlData);
5025 /* xlog insertion slots, plus alignment */
5026 size = add_size(size, mul_size(sizeof(XLogInsertSlotPadded), num_xloginsert_slots + 1));
5027 /* xlblocks array */
5028 size = add_size(size, mul_size(sizeof(XLogRecPtr), XLOGbuffers));
5029 /* extra alignment padding for XLOG I/O buffers */
5030 size = add_size(size, XLOG_BLCKSZ);
5031 /* and the buffers themselves */
5032 size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
5035 * Note: we don't count ControlFileData, it comes out of the "slop factor"
5036 * added by CreateSharedMemoryAndSemaphores. This lets us use this
5037 * routine again below to compute the actual allocation size.
5051 ControlFile = (ControlFileData *)
5052 ShmemInitStruct("Control File", sizeof(ControlFileData), &foundCFile);
5053 XLogCtl = (XLogCtlData *)
5054 ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog);
5056 if (foundCFile || foundXLog)
5058 /* both should be present or neither */
5059 Assert(foundCFile && foundXLog);
5062 memset(XLogCtl, 0, sizeof(XLogCtlData));
5065 * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
5066 * multiple of the alignment for same, so no extra alignment padding is
5069 allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
5070 XLogCtl->xlblocks = (XLogRecPtr *) allocptr;
5071 memset(XLogCtl->xlblocks, 0, sizeof(XLogRecPtr) * XLOGbuffers);
5072 allocptr += sizeof(XLogRecPtr) * XLOGbuffers;
5074 /* Xlog insertion slots. Ensure they're aligned to the full padded size */
5075 allocptr += sizeof(XLogInsertSlotPadded) -
5076 ((uintptr_t) allocptr) % sizeof(XLogInsertSlotPadded);
5077 XLogCtl->Insert.insertSlots = (XLogInsertSlotPadded *) allocptr;
5078 allocptr += sizeof(XLogInsertSlotPadded) * num_xloginsert_slots;
5081 * Align the start of the page buffers to a full xlog block size boundary.
5082 * This simplifies some calculations in XLOG insertion. It is also required
5085 allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5086 XLogCtl->pages = allocptr;
5087 memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5090 * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5091 * in additional info.)
5093 XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5094 XLogCtl->SharedRecoveryInProgress = true;
5095 XLogCtl->SharedHotStandbyActive = false;
5096 XLogCtl->WalWriterSleeping = false;
5098 for (i = 0; i < num_xloginsert_slots; i++)
5100 XLogInsertSlot *slot = &XLogCtl->Insert.insertSlots[i].slot;
5101 SpinLockInit(&slot->mutex);
5102 slot->xlogInsertingAt = InvalidXLogRecPtr;
5105 slot->releaseOK = true;
5106 slot->exclusive = 0;
5111 SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5112 SpinLockInit(&XLogCtl->info_lck);
5113 SpinLockInit(&XLogCtl->ulsn_lck);
5114 InitSharedLatch(&XLogCtl->recoveryWakeupLatch);
5117 * If we are not in bootstrap mode, pg_control should already exist. Read
5118 * and validate it immediately (see comments in ReadControlFile() for the
5121 if (!IsBootstrapProcessingMode())
5126 * This func must be called ONCE on system install. It creates pg_control
5127 * and the initial XLOG segment.
5132 CheckPoint checkPoint;
5134 XLogPageHeader page;
5135 XLogLongPageHeader longpage;
5138 uint64 sysidentifier;
5143 * Select a hopefully-unique system identifier code for this installation.
5144 * We use the result of gettimeofday(), including the fractional seconds
5145 * field, as being about as unique as we can easily get. (Think not to
5146 * use random(), since it hasn't been seeded and there's no portable way
5147 * to seed it other than the system clock value...) The upper half of the
5148 * uint64 value is just the tv_sec part, while the lower half is the XOR
5149 * of tv_sec and tv_usec. This is to ensure that we don't lose uniqueness
5150 * unnecessarily if "uint64" is really only 32 bits wide. A person
5151 * knowing this encoding can determine the initialization time of the
5152 * installation, which could perhaps be useful sometimes.
5154 gettimeofday(&tv, NULL);
5155 sysidentifier = ((uint64) tv.tv_sec) << 32;
5156 sysidentifier |= (uint32) (tv.tv_sec | tv.tv_usec);
5158 /* First timeline ID is always 1 */
5161 /* page buffer must be aligned suitably for O_DIRECT */
5162 buffer = (char *) palloc(XLOG_BLCKSZ + XLOG_BLCKSZ);
5163 page = (XLogPageHeader) TYPEALIGN(XLOG_BLCKSZ, buffer);
5164 memset(page, 0, XLOG_BLCKSZ);
5167 * Set up information for the initial checkpoint record
5169 * The initial checkpoint record is written to the beginning of the WAL
5170 * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5171 * used, so that we can use 0/0 to mean "before any valid WAL segment".
5173 checkPoint.redo = XLogSegSize + SizeOfXLogLongPHD;
5174 checkPoint.ThisTimeLineID = ThisTimeLineID;
5175 checkPoint.PrevTimeLineID = ThisTimeLineID;
5176 checkPoint.fullPageWrites = fullPageWrites;
5177 checkPoint.nextXidEpoch = 0;
5178 checkPoint.nextXid = FirstNormalTransactionId;
5179 checkPoint.nextOid = FirstBootstrapObjectId;
5180 checkPoint.nextMulti = FirstMultiXactId;
5181 checkPoint.nextMultiOffset = 0;
5182 checkPoint.oldestXid = FirstNormalTransactionId;
5183 checkPoint.oldestXidDB = TemplateDbOid;
5184 checkPoint.oldestMulti = FirstMultiXactId;
5185 checkPoint.oldestMultiDB = TemplateDbOid;
5186 checkPoint.time = (pg_time_t) time(NULL);
5187 checkPoint.oldestActiveXid = InvalidTransactionId;
5189 ShmemVariableCache->nextXid = checkPoint.nextXid;
5190 ShmemVariableCache->nextOid = checkPoint.nextOid;
5191 ShmemVariableCache->oidCount = 0;
5192 MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5193 SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5194 SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
5196 /* Set up the XLOG page header */
5197 page->xlp_magic = XLOG_PAGE_MAGIC;
5198 page->xlp_info = XLP_LONG_HEADER;
5199 page->xlp_tli = ThisTimeLineID;
5200 page->xlp_pageaddr = XLogSegSize;
5201 longpage = (XLogLongPageHeader) page;
5202 longpage->xlp_sysid = sysidentifier;
5203 longpage->xlp_seg_size = XLogSegSize;
5204 longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5206 /* Insert the initial checkpoint record */
5207 record = (XLogRecord *) ((char *) page + SizeOfXLogLongPHD);
5208 record->xl_prev = 0;
5209 record->xl_xid = InvalidTransactionId;
5210 record->xl_tot_len = SizeOfXLogRecord + sizeof(checkPoint);
5211 record->xl_len = sizeof(checkPoint);
5212 record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5213 record->xl_rmid = RM_XLOG_ID;
5214 memcpy(XLogRecGetData(record), &checkPoint, sizeof(checkPoint));
5217 COMP_CRC32(crc, &checkPoint, sizeof(checkPoint));
5218 COMP_CRC32(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5220 record->xl_crc = crc;
5222 /* Create first XLOG segment file */
5223 use_existent = false;
5224 openLogFile = XLogFileInit(1, &use_existent, false);
5226 /* Write the first page with the initial record */
5228 if (write(openLogFile, page, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5230 /* if write didn't set errno, assume problem is no disk space */
5234 (errcode_for_file_access(),
5235 errmsg("could not write bootstrap transaction log file: %m")));
5238 if (pg_fsync(openLogFile) != 0)
5240 (errcode_for_file_access(),
5241 errmsg("could not fsync bootstrap transaction log file: %m")));
5243 if (close(openLogFile))
5245 (errcode_for_file_access(),
5246 errmsg("could not close bootstrap transaction log file: %m")));
5250 /* Now create pg_control */
5252 memset(ControlFile, 0, sizeof(ControlFileData));
5253 /* Initialize pg_control status fields */
5254 ControlFile->system_identifier = sysidentifier;
5255 ControlFile->state = DB_SHUTDOWNED;
5256 ControlFile->time = checkPoint.time;
5257 ControlFile->checkPoint = checkPoint.redo;
5258 ControlFile->checkPointCopy = checkPoint;
5259 ControlFile->unloggedLSN = 1;
5261 /* Set important parameter values for use when replaying WAL */
5262 ControlFile->MaxConnections = MaxConnections;
5263 ControlFile->max_worker_processes = max_worker_processes;
5264 ControlFile->max_prepared_xacts = max_prepared_xacts;
5265 ControlFile->max_locks_per_xact = max_locks_per_xact;
5266 ControlFile->wal_level = wal_level;
5267 ControlFile->data_checksum_version = bootstrap_data_checksum_version;
5269 /* some additional ControlFile fields are set in WriteControlFile() */
5273 /* Bootstrap the commit log, too */
5275 BootStrapSUBTRANS();
5276 BootStrapMultiXact();
5282 str_time(pg_time_t tnow)
5284 static char buf[128];
5286 pg_strftime(buf, sizeof(buf),
5287 "%Y-%m-%d %H:%M:%S %Z",
5288 pg_localtime(&tnow, log_timezone));
5294 * See if there is a recovery command file (recovery.conf), and if so
5295 * read in parameters for archive recovery and XLOG streaming.
5297 * The file is parsed using the main configuration parser.
5300 readRecoveryCommandFile(void)
5303 TimeLineID rtli = 0;
5304 bool rtliGiven = false;
5305 ConfigVariable *item,
5309 fd = AllocateFile(RECOVERY_COMMAND_FILE, "r");
5312 if (errno == ENOENT)
5313 return; /* not there, so no archive recovery */
5315 (errcode_for_file_access(),
5316 errmsg("could not open recovery command file \"%s\": %m",
5317 RECOVERY_COMMAND_FILE)));
5321 * Since we're asking ParseConfigFp() to report errors as FATAL, there's
5322 * no need to check the return value.
5324 (void) ParseConfigFp(fd, RECOVERY_COMMAND_FILE, 0, FATAL, &head, &tail);
5328 for (item = head; item; item = item->next)
5330 if (strcmp(item->name, "restore_command") == 0)
5332 recoveryRestoreCommand = pstrdup(item->value);
5334 (errmsg_internal("restore_command = '%s'",
5335 recoveryRestoreCommand)));
5337 else if (strcmp(item->name, "recovery_end_command") == 0)
5339 recoveryEndCommand = pstrdup(item->value);
5341 (errmsg_internal("recovery_end_command = '%s'",
5342 recoveryEndCommand)));
5344 else if (strcmp(item->name, "archive_cleanup_command") == 0)
5346 archiveCleanupCommand = pstrdup(item->value);
5348 (errmsg_internal("archive_cleanup_command = '%s'",
5349 archiveCleanupCommand)));
5351 else if (strcmp(item->name, "pause_at_recovery_target") == 0)
5353 if (!parse_bool(item->value, &recoveryPauseAtTarget))
5355 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
5356 errmsg("parameter \"%s\" requires a Boolean value", "pause_at_recovery_target")));
5358 (errmsg_internal("pause_at_recovery_target = '%s'",
5361 else if (strcmp(item->name, "recovery_target_timeline") == 0)
5364 if (strcmp(item->value, "latest") == 0)
5369 rtli = (TimeLineID) strtoul(item->value, NULL, 0);
5370 if (errno == EINVAL || errno == ERANGE)
5372 (errmsg("recovery_target_timeline is not a valid number: \"%s\"",
5377 (errmsg_internal("recovery_target_timeline = %u", rtli)));
5380 (errmsg_internal("recovery_target_timeline = latest")));
5382 else if (strcmp(item->name, "recovery_target_xid") == 0)
5385 recoveryTargetXid = (TransactionId) strtoul(item->value, NULL, 0);
5386 if (errno == EINVAL || errno == ERANGE)
5388 (errmsg("recovery_target_xid is not a valid number: \"%s\"",
5391 (errmsg_internal("recovery_target_xid = %u",
5392 recoveryTargetXid)));
5393 recoveryTarget = RECOVERY_TARGET_XID;
5395 else if (strcmp(item->name, "recovery_target_time") == 0)
5398 * if recovery_target_xid or recovery_target_name specified, then
5399 * this overrides recovery_target_time
5401 if (recoveryTarget == RECOVERY_TARGET_XID ||
5402 recoveryTarget == RECOVERY_TARGET_NAME)
5404 recoveryTarget = RECOVERY_TARGET_TIME;
5407 * Convert the time string given by the user to TimestampTz form.
5409 recoveryTargetTime =
5410 DatumGetTimestampTz(DirectFunctionCall3(timestamptz_in,
5411 CStringGetDatum(item->value),
5412 ObjectIdGetDatum(InvalidOid),
5413 Int32GetDatum(-1)));
5415 (errmsg_internal("recovery_target_time = '%s'",
5416 timestamptz_to_str(recoveryTargetTime))));
5418 else if (strcmp(item->name, "recovery_target_name") == 0)
5421 * if recovery_target_xid specified, then this overrides
5422 * recovery_target_name
5424 if (recoveryTarget == RECOVERY_TARGET_XID)
5426 recoveryTarget = RECOVERY_TARGET_NAME;
5428 recoveryTargetName = pstrdup(item->value);
5429 if (strlen(recoveryTargetName) >= MAXFNAMELEN)
5431 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
5432 errmsg("recovery_target_name is too long (maximum %d characters)",
5436 (errmsg_internal("recovery_target_name = '%s'",
5437 recoveryTargetName)));
5439 else if (strcmp(item->name, "recovery_target_inclusive") == 0)
5442 * does nothing if a recovery_target is not also set
5444 if (!parse_bool(item->value, &recoveryTargetInclusive))
5446 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
5447 errmsg("parameter \"%s\" requires a Boolean value",
5448 "recovery_target_inclusive")));
5450 (errmsg_internal("recovery_target_inclusive = %s",
5453 else if (strcmp(item->name, "standby_mode") == 0)
5455 if (!parse_bool(item->value, &StandbyModeRequested))
5457 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
5458 errmsg("parameter \"%s\" requires a Boolean value",
5461 (errmsg_internal("standby_mode = '%s'", item->value)));
5463 else if (strcmp(item->name, "primary_conninfo") == 0)
5465 PrimaryConnInfo = pstrdup(item->value);
5467 (errmsg_internal("primary_conninfo = '%s'",
5470 else if (strcmp(item->name, "trigger_file") == 0)
5472 TriggerFile = pstrdup(item->value);
5474 (errmsg_internal("trigger_file = '%s'",
5479 (errmsg("unrecognized recovery parameter \"%s\"",
5484 * Check for compulsory parameters
5486 if (StandbyModeRequested)
5488 if (PrimaryConnInfo == NULL && recoveryRestoreCommand == NULL)
5490 (errmsg("recovery command file \"%s\" specified neither primary_conninfo nor restore_command",
5491 RECOVERY_COMMAND_FILE),
5492 errhint("The database server will regularly poll the pg_xlog subdirectory to check for files placed there.")));
5496 if (recoveryRestoreCommand == NULL)
5498 (errmsg("recovery command file \"%s\" must specify restore_command when standby mode is not enabled",
5499 RECOVERY_COMMAND_FILE)));
5502 /* Enable fetching from archive recovery area */
5503 ArchiveRecoveryRequested = true;
5506 * If user specified recovery_target_timeline, validate it or compute the
5507 * "latest" value. We can't do this until after we've gotten the restore
5508 * command and set InArchiveRecovery, because we need to fetch timeline
5509 * history files from the archive.
5515 /* Timeline 1 does not have a history file, all else should */
5516 if (rtli != 1 && !existsTimeLineHistory(rtli))
5518 (errmsg("recovery target timeline %u does not exist",
5520 recoveryTargetTLI = rtli;
5521 recoveryTargetIsLatest = false;
5525 /* We start the "latest" search from pg_control's timeline */
5526 recoveryTargetTLI = findNewestTimeLine(recoveryTargetTLI);
5527 recoveryTargetIsLatest = true;
5531 FreeConfigVariables(head);
5535 * Exit archive-recovery state
5538 exitArchiveRecovery(TimeLineID endTLI, XLogSegNo endLogSegNo)
5540 char recoveryPath[MAXPGPATH];
5541 char xlogpath[MAXPGPATH];
5544 * We are no longer in archive recovery state.
5546 InArchiveRecovery = false;
5549 * Update min recovery point one last time.
5551 UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5554 * If the ending log segment is still open, close it (to avoid problems on
5555 * Windows with trying to rename or delete an open file).
5564 * If we are establishing a new timeline, we have to copy data from the
5565 * last WAL segment of the old timeline to create a starting WAL segment
5566 * for the new timeline.
5568 * Notify the archiver that the last WAL segment of the old timeline is
5569 * ready to copy to archival storage. Otherwise, it is not archived for a
5572 if (endTLI != ThisTimeLineID)
5574 XLogFileCopy(endLogSegNo, endTLI, endLogSegNo);
5576 if (XLogArchivingActive())
5578 XLogFileName(xlogpath, endTLI, endLogSegNo);
5579 XLogArchiveNotify(xlogpath);
5584 * Let's just make real sure there are not .ready or .done flags posted
5585 * for the new segment.
5587 XLogFileName(xlogpath, ThisTimeLineID, endLogSegNo);
5588 XLogArchiveCleanup(xlogpath);
5591 * Since there might be a partial WAL segment named RECOVERYXLOG, get rid
5594 snprintf(recoveryPath, MAXPGPATH, XLOGDIR "/RECOVERYXLOG");
5595 unlink(recoveryPath); /* ignore any error */
5597 /* Get rid of any remaining recovered timeline-history file, too */
5598 snprintf(recoveryPath, MAXPGPATH, XLOGDIR "/RECOVERYHISTORY");
5599 unlink(recoveryPath); /* ignore any error */
5602 * Rename the config file out of the way, so that we don't accidentally
5603 * re-enter archive recovery mode in a subsequent crash.
5605 unlink(RECOVERY_COMMAND_DONE);
5606 if (rename(RECOVERY_COMMAND_FILE, RECOVERY_COMMAND_DONE) != 0)
5608 (errcode_for_file_access(),
5609 errmsg("could not rename file \"%s\" to \"%s\": %m",
5610 RECOVERY_COMMAND_FILE, RECOVERY_COMMAND_DONE)));
5613 (errmsg("archive recovery complete")));
5617 * For point-in-time recovery, this function decides whether we want to
5618 * stop applying the XLOG at or after the current record.
5620 * Returns TRUE if we are stopping, FALSE otherwise. On TRUE return,
5621 * *includeThis is set TRUE if we should apply this record before stopping.
5623 * We also track the timestamp of the latest applied COMMIT/ABORT
5624 * record in XLogCtl->recoveryLastXTime, for logging purposes.
5625 * Also, some information is saved in recoveryStopXid et al for use in
5626 * annotating the new timeline's history file.
5629 recoveryStopsHere(XLogRecord *record, bool *includeThis)
5633 TimestampTz recordXtime;
5634 char recordRPName[MAXFNAMELEN];
5636 /* We only consider stopping at COMMIT, ABORT or RESTORE POINT records */
5637 if (record->xl_rmid != RM_XACT_ID && record->xl_rmid != RM_XLOG_ID)
5639 record_info = record->xl_info & ~XLR_INFO_MASK;
5640 if (record->xl_rmid == RM_XACT_ID && record_info == XLOG_XACT_COMMIT_COMPACT)
5642 xl_xact_commit_compact *recordXactCommitData;
5644 recordXactCommitData = (xl_xact_commit_compact *) XLogRecGetData(record);
5645 recordXtime = recordXactCommitData->xact_time;
5647 else if (record->xl_rmid == RM_XACT_ID && record_info == XLOG_XACT_COMMIT)
5649 xl_xact_commit *recordXactCommitData;
5651 recordXactCommitData = (xl_xact_commit *) XLogRecGetData(record);
5652 recordXtime = recordXactCommitData->xact_time;
5654 else if (record->xl_rmid == RM_XACT_ID && record_info == XLOG_XACT_ABORT)
5656 xl_xact_abort *recordXactAbortData;
5658 recordXactAbortData = (xl_xact_abort *) XLogRecGetData(record);
5659 recordXtime = recordXactAbortData->xact_time;
5661 else if (record->xl_rmid == RM_XLOG_ID && record_info == XLOG_RESTORE_POINT)
5663 xl_restore_point *recordRestorePointData;
5665 recordRestorePointData = (xl_restore_point *) XLogRecGetData(record);
5666 recordXtime = recordRestorePointData->rp_time;
5667 strncpy(recordRPName, recordRestorePointData->rp_name, MAXFNAMELEN);
5672 /* Do we have a PITR target at all? */
5673 if (recoveryTarget == RECOVERY_TARGET_UNSET)
5676 * Save timestamp of latest transaction commit/abort if this is a
5677 * transaction record
5679 if (record->xl_rmid == RM_XACT_ID)
5680 SetLatestXTime(recordXtime);
5684 if (recoveryTarget == RECOVERY_TARGET_XID)
5687 * There can be only one transaction end record with this exact
5690 * when testing for an xid, we MUST test for equality only, since
5691 * transactions are numbered in the order they start, not the order
5692 * they complete. A higher numbered xid will complete before you about
5693 * 50% of the time...
5695 stopsHere = (record->xl_xid == recoveryTargetXid);
5697 *includeThis = recoveryTargetInclusive;
5699 else if (recoveryTarget == RECOVERY_TARGET_NAME)
5702 * There can be many restore points that share the same name, so we
5703 * stop at the first one
5705 stopsHere = (strcmp(recordRPName, recoveryTargetName) == 0);
5708 * Ignore recoveryTargetInclusive because this is not a transaction
5711 *includeThis = false;
5716 * There can be many transactions that share the same commit time, so
5717 * we stop after the last one, if we are inclusive, or stop at the
5718 * first one if we are exclusive
5720 if (recoveryTargetInclusive)
5721 stopsHere = (recordXtime > recoveryTargetTime);
5723 stopsHere = (recordXtime >= recoveryTargetTime);
5725 *includeThis = false;
5730 recoveryStopXid = record->xl_xid;
5731 recoveryStopTime = recordXtime;
5732 recoveryStopAfter = *includeThis;
5734 if (record_info == XLOG_XACT_COMMIT_COMPACT || record_info == XLOG_XACT_COMMIT)
5736 if (recoveryStopAfter)
5738 (errmsg("recovery stopping after commit of transaction %u, time %s",
5740 timestamptz_to_str(recoveryStopTime))));
5743 (errmsg("recovery stopping before commit of transaction %u, time %s",
5745 timestamptz_to_str(recoveryStopTime))));
5747 else if (record_info == XLOG_XACT_ABORT)
5749 if (recoveryStopAfter)
5751 (errmsg("recovery stopping after abort of transaction %u, time %s",
5753 timestamptz_to_str(recoveryStopTime))));
5756 (errmsg("recovery stopping before abort of transaction %u, time %s",
5758 timestamptz_to_str(recoveryStopTime))));
5762 strncpy(recoveryStopName, recordRPName, MAXFNAMELEN);
5765 (errmsg("recovery stopping at restore point \"%s\", time %s",
5767 timestamptz_to_str(recoveryStopTime))));
5771 * Note that if we use a RECOVERY_TARGET_TIME then we can stop at a
5772 * restore point since they are timestamped, though the latest
5773 * transaction time is not updated.
5775 if (record->xl_rmid == RM_XACT_ID && recoveryStopAfter)
5776 SetLatestXTime(recordXtime);
5778 else if (record->xl_rmid == RM_XACT_ID)
5779 SetLatestXTime(recordXtime);
5785 * Wait until shared recoveryPause flag is cleared.
5787 * XXX Could also be done with shared latch, avoiding the pg_usleep loop.
5788 * Probably not worth the trouble though. This state shouldn't be one that
5789 * anyone cares about server power consumption in.
5792 recoveryPausesHere(void)
5794 /* Don't pause unless users can connect! */
5795 if (!LocalHotStandbyActive)
5799 (errmsg("recovery has paused"),
5800 errhint("Execute pg_xlog_replay_resume() to continue.")));
5802 while (RecoveryIsPaused())
5804 pg_usleep(1000000L); /* 1000 ms */
5805 HandleStartupProcInterrupts();
5810 RecoveryIsPaused(void)
5812 /* use volatile pointer to prevent code rearrangement */
5813 volatile XLogCtlData *xlogctl = XLogCtl;
5816 SpinLockAcquire(&xlogctl->info_lck);
5817 recoveryPause = xlogctl->recoveryPause;
5818 SpinLockRelease(&xlogctl->info_lck);
5820 return recoveryPause;
5824 SetRecoveryPause(bool recoveryPause)
5826 /* use volatile pointer to prevent code rearrangement */
5827 volatile XLogCtlData *xlogctl = XLogCtl;
5829 SpinLockAcquire(&xlogctl->info_lck);
5830 xlogctl->recoveryPause = recoveryPause;
5831 SpinLockRelease(&xlogctl->info_lck);
5835 * Save timestamp of latest processed commit/abort record.
5837 * We keep this in XLogCtl, not a simple static variable, so that it can be
5838 * seen by processes other than the startup process. Note in particular
5839 * that CreateRestartPoint is executed in the checkpointer.
5842 SetLatestXTime(TimestampTz xtime)
5844 /* use volatile pointer to prevent code rearrangement */
5845 volatile XLogCtlData *xlogctl = XLogCtl;
5847 SpinLockAcquire(&xlogctl->info_lck);
5848 xlogctl->recoveryLastXTime = xtime;
5849 SpinLockRelease(&xlogctl->info_lck);
5853 * Fetch timestamp of latest processed commit/abort record.
5856 GetLatestXTime(void)
5858 /* use volatile pointer to prevent code rearrangement */
5859 volatile XLogCtlData *xlogctl = XLogCtl;
5862 SpinLockAcquire(&xlogctl->info_lck);
5863 xtime = xlogctl->recoveryLastXTime;
5864 SpinLockRelease(&xlogctl->info_lck);
5870 * Save timestamp of the next chunk of WAL records to apply.
5872 * We keep this in XLogCtl, not a simple static variable, so that it can be
5873 * seen by all backends.
5876 SetCurrentChunkStartTime(TimestampTz xtime)
5878 /* use volatile pointer to prevent code rearrangement */
5879 volatile XLogCtlData *xlogctl = XLogCtl;
5881 SpinLockAcquire(&xlogctl->info_lck);
5882 xlogctl->currentChunkStartTime = xtime;
5883 SpinLockRelease(&xlogctl->info_lck);
5887 * Fetch timestamp of latest processed commit/abort record.
5888 * Startup process maintains an accurate local copy in XLogReceiptTime
5891 GetCurrentChunkReplayStartTime(void)
5893 /* use volatile pointer to prevent code rearrangement */
5894 volatile XLogCtlData *xlogctl = XLogCtl;
5897 SpinLockAcquire(&xlogctl->info_lck);
5898 xtime = xlogctl->currentChunkStartTime;
5899 SpinLockRelease(&xlogctl->info_lck);
5905 * Returns time of receipt of current chunk of XLOG data, as well as
5906 * whether it was received from streaming replication or from archives.
5909 GetXLogReceiptTime(TimestampTz *rtime, bool *fromStream)
5912 * This must be executed in the startup process, since we don't export the
5913 * relevant state to shared memory.
5917 *rtime = XLogReceiptTime;
5918 *fromStream = (XLogReceiptSource == XLOG_FROM_STREAM);
5922 * Note that text field supplied is a parameter name and does not require
5925 #define RecoveryRequiresIntParameter(param_name, currValue, minValue) \
5927 if ((currValue) < (minValue)) \
5929 (errcode(ERRCODE_INVALID_PARAMETER_VALUE), \
5930 errmsg("hot standby is not possible because " \
5931 "%s = %d is a lower setting than on the master server " \
5932 "(its value was %d)", \
5939 * Check to see if required parameters are set high enough on this server
5940 * for various aspects of recovery operation.
5943 CheckRequiredParameterValues(void)
5946 * For archive recovery, the WAL must be generated with at least 'archive'
5949 if (InArchiveRecovery && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5952 (errmsg("WAL was generated with wal_level=minimal, data may be missing"),
5953 errhint("This happens if you temporarily set wal_level=minimal without taking a new base backup.")));
5957 * For Hot Standby, the WAL must be generated with 'hot_standby' mode, and
5958 * we must have at least as many backend slots as the primary.
5960 if (InArchiveRecovery && EnableHotStandby)
5962 if (ControlFile->wal_level < WAL_LEVEL_HOT_STANDBY)
5964 (errmsg("hot standby is not possible because wal_level was not set to \"hot_standby\" on the master server"),
5965 errhint("Either set wal_level to \"hot_standby\" on the master, or turn off hot_standby here.")));
5967 /* We ignore autovacuum_max_workers when we make this test. */
5968 RecoveryRequiresIntParameter("max_connections",
5970 ControlFile->MaxConnections);
5971 RecoveryRequiresIntParameter("max_worker_processes",
5972 max_worker_processes,
5973 ControlFile->max_worker_processes);
5974 RecoveryRequiresIntParameter("max_prepared_transactions",
5976 ControlFile->max_prepared_xacts);
5977 RecoveryRequiresIntParameter("max_locks_per_transaction",
5979 ControlFile->max_locks_per_xact);
5984 * This must be called ONCE during postmaster or standalone-backend startup
5989 XLogCtlInsert *Insert;
5990 CheckPoint checkPoint;
5992 bool reachedStopPoint = false;
5993 bool haveBackupLabel = false;
5997 XLogSegNo endLogSegNo;
5998 TimeLineID PrevTimeLineID;
6000 TransactionId oldestActiveXID;
6001 bool backupEndRequired = false;
6002 bool backupFromStandby = false;
6003 DBState dbstate_at_startup;
6004 XLogReaderState *xlogreader;
6005 XLogPageReadPrivate private;
6006 bool fast_promoted = false;
6009 * Read control file and check XLOG status looks valid.
6011 * Note: in most control paths, *ControlFile is already valid and we need
6012 * not do ReadControlFile() here, but might as well do it to be sure.
6016 if (ControlFile->state < DB_SHUTDOWNED ||
6017 ControlFile->state > DB_IN_PRODUCTION ||
6018 !XRecOffIsValid(ControlFile->checkPoint))
6020 (errmsg("control file contains invalid data")));
6022 if (ControlFile->state == DB_SHUTDOWNED)
6024 /* This is the expected case, so don't be chatty in standalone mode */
6025 ereport(IsPostmasterEnvironment ? LOG : NOTICE,
6026 (errmsg("database system was shut down at %s",
6027 str_time(ControlFile->time))));
6029 else if (ControlFile->state == DB_SHUTDOWNED_IN_RECOVERY)
6031 (errmsg("database system was shut down in recovery at %s",
6032 str_time(ControlFile->time))));
6033 else if (ControlFile->state == DB_SHUTDOWNING)
6035 (errmsg("database system shutdown was interrupted; last known up at %s",
6036 str_time(ControlFile->time))));
6037 else if (ControlFile->state == DB_IN_CRASH_RECOVERY)
6039 (errmsg("database system was interrupted while in recovery at %s",
6040 str_time(ControlFile->time)),
6041 errhint("This probably means that some data is corrupted and"
6042 " you will have to use the last backup for recovery.")));
6043 else if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
6045 (errmsg("database system was interrupted while in recovery at log time %s",
6046 str_time(ControlFile->checkPointCopy.time)),
6047 errhint("If this has occurred more than once some data might be corrupted"
6048 " and you might need to choose an earlier recovery target.")));
6049 else if (ControlFile->state == DB_IN_PRODUCTION)
6051 (errmsg("database system was interrupted; last known up at %s",
6052 str_time(ControlFile->time))));
6054 /* This is just to allow attaching to startup process with a debugger */
6055 #ifdef XLOG_REPLAY_DELAY
6056 if (ControlFile->state != DB_SHUTDOWNED)
6057 pg_usleep(60000000L);
6061 * Verify that pg_xlog and pg_xlog/archive_status exist. In cases where
6062 * someone has performed a copy for PITR, these directories may have been
6063 * excluded and need to be re-created.
6065 ValidateXLOGDirectoryStructure();
6068 * Clear out any old relcache cache files. This is *necessary* if we do
6069 * any WAL replay, since that would probably result in the cache files
6070 * being out of sync with database reality. In theory we could leave them
6071 * in place if the database had been cleanly shut down, but it seems
6072 * safest to just remove them always and let them be rebuilt during the
6073 * first backend startup.
6075 RelationCacheInitFileRemove();
6078 * Initialize on the assumption we want to recover to the latest timeline
6079 * that's active according to pg_control.
6081 if (ControlFile->minRecoveryPointTLI >
6082 ControlFile->checkPointCopy.ThisTimeLineID)
6083 recoveryTargetTLI = ControlFile->minRecoveryPointTLI;
6085 recoveryTargetTLI = ControlFile->checkPointCopy.ThisTimeLineID;
6088 * Check for recovery control file, and if so set up state for offline
6091 readRecoveryCommandFile();
6094 * Save archive_cleanup_command in shared memory so that other processes
6097 strncpy(XLogCtl->archiveCleanupCommand,
6098 archiveCleanupCommand ? archiveCleanupCommand : "",
6099 sizeof(XLogCtl->archiveCleanupCommand));
6101 if (ArchiveRecoveryRequested)
6103 if (StandbyModeRequested)
6105 (errmsg("entering standby mode")));
6106 else if (recoveryTarget == RECOVERY_TARGET_XID)
6108 (errmsg("starting point-in-time recovery to XID %u",
6109 recoveryTargetXid)));
6110 else if (recoveryTarget == RECOVERY_TARGET_TIME)
6112 (errmsg("starting point-in-time recovery to %s",
6113 timestamptz_to_str(recoveryTargetTime))));
6114 else if (recoveryTarget == RECOVERY_TARGET_NAME)
6116 (errmsg("starting point-in-time recovery to \"%s\"",
6117 recoveryTargetName)));
6120 (errmsg("starting archive recovery")));
6124 * Take ownership of the wakeup latch if we're going to sleep during
6127 if (StandbyModeRequested)
6128 OwnLatch(&XLogCtl->recoveryWakeupLatch);
6130 /* Set up XLOG reader facility */
6131 MemSet(&private, 0, sizeof(XLogPageReadPrivate));
6132 xlogreader = XLogReaderAllocate(&XLogPageRead, &private);
6135 (errcode(ERRCODE_OUT_OF_MEMORY),
6136 errmsg("out of memory"),
6137 errdetail("Failed while allocating an XLog reading processor.")));
6138 xlogreader->system_identifier = ControlFile->system_identifier;
6140 if (read_backup_label(&checkPointLoc, &backupEndRequired,
6141 &backupFromStandby))
6144 * Archive recovery was requested, and thanks to the backup label
6145 * file, we know how far we need to replay to reach consistency. Enter
6146 * archive recovery directly.
6148 InArchiveRecovery = true;
6149 if (StandbyModeRequested)
6153 * When a backup_label file is present, we want to roll forward from
6154 * the checkpoint it identifies, rather than using pg_control.
6156 record = ReadCheckpointRecord(xlogreader, checkPointLoc, 0, true);
6159 memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
6160 wasShutdown = (record->xl_info == XLOG_CHECKPOINT_SHUTDOWN);
6162 (errmsg("checkpoint record is at %X/%X",
6163 (uint32) (checkPointLoc >> 32), (uint32) checkPointLoc)));
6164 InRecovery = true; /* force recovery even if SHUTDOWNED */
6167 * Make sure that REDO location exists. This may not be the case
6168 * if there was a crash during an online backup, which left a
6169 * backup_label around that references a WAL segment that's
6170 * already been archived.
6172 if (checkPoint.redo < checkPointLoc)
6174 if (!ReadRecord(xlogreader, checkPoint.redo, LOG, false))
6176 (errmsg("could not find redo location referenced by checkpoint record"),
6177 errhint("If you are not restoring from a backup, try removing the file \"%s/backup_label\".", DataDir)));
6183 (errmsg("could not locate required checkpoint record"),
6184 errhint("If you are not restoring from a backup, try removing the file \"%s/backup_label\".", DataDir)));
6185 wasShutdown = false; /* keep compiler quiet */
6187 /* set flag to delete it later */
6188 haveBackupLabel = true;
6193 * It's possible that archive recovery was requested, but we don't
6194 * know how far we need to replay the WAL before we reach consistency.
6195 * This can happen for example if a base backup is taken from a
6196 * running server using an atomic filesystem snapshot, without calling
6197 * pg_start/stop_backup. Or if you just kill a running master server
6198 * and put it into archive recovery by creating a recovery.conf file.
6200 * Our strategy in that case is to perform crash recovery first,
6201 * replaying all the WAL present in pg_xlog, and only enter archive
6202 * recovery after that.
6204 * But usually we already know how far we need to replay the WAL (up
6205 * to minRecoveryPoint, up to backupEndPoint, or until we see an
6206 * end-of-backup record), and we can enter archive recovery directly.
6208 if (ArchiveRecoveryRequested &&
6209 (ControlFile->minRecoveryPoint != InvalidXLogRecPtr ||
6210 ControlFile->backupEndRequired ||
6211 ControlFile->backupEndPoint != InvalidXLogRecPtr ||
6212 ControlFile->state == DB_SHUTDOWNED))
6214 InArchiveRecovery = true;
6215 if (StandbyModeRequested)
6220 * Get the last valid checkpoint record. If the latest one according
6221 * to pg_control is broken, try the next-to-last one.
6223 checkPointLoc = ControlFile->checkPoint;
6224 RedoStartLSN = ControlFile->checkPointCopy.redo;
6225 record = ReadCheckpointRecord(xlogreader, checkPointLoc, 1, true);
6229 (errmsg("checkpoint record is at %X/%X",
6230 (uint32) (checkPointLoc >> 32), (uint32) checkPointLoc)));
6232 else if (StandbyMode)
6235 * The last valid checkpoint record required for a streaming
6236 * recovery exists in neither standby nor the primary.
6239 (errmsg("could not locate a valid checkpoint record")));
6243 checkPointLoc = ControlFile->prevCheckPoint;
6244 record = ReadCheckpointRecord(xlogreader, checkPointLoc, 2, true);
6248 (errmsg("using previous checkpoint record at %X/%X",
6249 (uint32) (checkPointLoc >> 32), (uint32) checkPointLoc)));
6250 InRecovery = true; /* force recovery even if SHUTDOWNED */
6254 (errmsg("could not locate a valid checkpoint record")));
6256 memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
6257 wasShutdown = (record->xl_info == XLOG_CHECKPOINT_SHUTDOWN);
6261 * If the location of the checkpoint record is not on the expected
6262 * timeline in the history of the requested timeline, we cannot proceed:
6263 * the backup is not part of the history of the requested timeline.
6265 Assert(expectedTLEs); /* was initialized by reading checkpoint
6267 if (tliOfPointInHistory(checkPointLoc, expectedTLEs) !=
6268 checkPoint.ThisTimeLineID)
6270 XLogRecPtr switchpoint;
6273 * tliSwitchPoint will throw an error if the checkpoint's timeline is
6274 * not in expectedTLEs at all.
6276 switchpoint = tliSwitchPoint(ControlFile->checkPointCopy.ThisTimeLineID, expectedTLEs, NULL);
6278 (errmsg("requested timeline %u is not a child of this server's history",
6280 errdetail("Latest checkpoint is at %X/%X on timeline %u, but in the history of the requested timeline, the server forked off from that timeline at %X/%X.",
6281 (uint32) (ControlFile->checkPoint >> 32),
6282 (uint32) ControlFile->checkPoint,
6283 ControlFile->checkPointCopy.ThisTimeLineID,
6284 (uint32) (switchpoint >> 32),
6285 (uint32) switchpoint)));
6289 * The min recovery point should be part of the requested timeline's
6292 if (!XLogRecPtrIsInvalid(ControlFile->minRecoveryPoint) &&
6293 tliOfPointInHistory(ControlFile->minRecoveryPoint - 1, expectedTLEs) !=
6294 ControlFile->minRecoveryPointTLI)
6296 (errmsg("requested timeline %u does not contain minimum recovery point %X/%X on timeline %u",
6298 (uint32) (ControlFile->minRecoveryPoint >> 32),
6299 (uint32) ControlFile->minRecoveryPoint,
6300 ControlFile->minRecoveryPointTLI)));
6302 LastRec = RecPtr = checkPointLoc;
6305 (errmsg("redo record is at %X/%X; shutdown %s",
6306 (uint32) (checkPoint.redo >> 32), (uint32) checkPoint.redo,
6307 wasShutdown ? "TRUE" : "FALSE")));
6309 (errmsg("next transaction ID: %u/%u; next OID: %u",
6310 checkPoint.nextXidEpoch, checkPoint.nextXid,
6311 checkPoint.nextOid)));
6313 (errmsg("next MultiXactId: %u; next MultiXactOffset: %u",
6314 checkPoint.nextMulti, checkPoint.nextMultiOffset)));
6316 (errmsg("oldest unfrozen transaction ID: %u, in database %u",
6317 checkPoint.oldestXid, checkPoint.oldestXidDB)));
6319 (errmsg("oldest MultiXactId: %u, in database %u",
6320 checkPoint.oldestMulti, checkPoint.oldestMultiDB)));
6321 if (!TransactionIdIsNormal(checkPoint.nextXid))
6323 (errmsg("invalid next transaction ID")));
6325 /* initialize shared memory variables from the checkpoint record */
6326 ShmemVariableCache->nextXid = checkPoint.nextXid;
6327 ShmemVariableCache->nextOid = checkPoint.nextOid;
6328 ShmemVariableCache->oidCount = 0;
6329 MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
6330 SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
6331 SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
6332 XLogCtl->ckptXidEpoch = checkPoint.nextXidEpoch;
6333 XLogCtl->ckptXid = checkPoint.nextXid;
6336 * Initialize unlogged LSN. On a clean shutdown, it's restored from the
6337 * control file. On recovery, all unlogged relations are blown away, so
6338 * the unlogged LSN counter can be reset too.
6340 if (ControlFile->state == DB_SHUTDOWNED)
6341 XLogCtl->unloggedLSN = ControlFile->unloggedLSN;
6343 XLogCtl->unloggedLSN = 1;
6346 * We must replay WAL entries using the same TimeLineID they were created
6347 * under, so temporarily adopt the TLI indicated by the checkpoint (see
6348 * also xlog_redo()).
6350 ThisTimeLineID = checkPoint.ThisTimeLineID;
6353 * Copy any missing timeline history files between 'now' and the recovery
6354 * target timeline from archive to pg_xlog. While we don't need those
6355 * files ourselves - the history file of the recovery target timeline
6356 * covers all the previous timelines in the history too - a cascading
6357 * standby server might be interested in them. Or, if you archive the WAL
6358 * from this server to a different archive than the master, it'd be good
6359 * for all the history files to get archived there after failover, so that
6360 * you can use one of the old timelines as a PITR target. Timeline history
6361 * files are small, so it's better to copy them unnecessarily than not
6362 * copy them and regret later.
6364 restoreTimeLineHistoryFiles(ThisTimeLineID, recoveryTargetTLI);
6366 lastFullPageWrites = checkPoint.fullPageWrites;
6368 RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
6370 if (RecPtr < checkPoint.redo)
6372 (errmsg("invalid redo in checkpoint record")));
6375 * Check whether we need to force recovery from WAL. If it appears to
6376 * have been a clean shutdown and we did not have a recovery.conf file,
6377 * then assume no recovery needed.
6379 if (checkPoint.redo < RecPtr)
6383 (errmsg("invalid redo record in shutdown checkpoint")));
6386 else if (ControlFile->state != DB_SHUTDOWNED)
6388 else if (ArchiveRecoveryRequested)
6390 /* force recovery due to presence of recovery.conf */
6399 /* use volatile pointer to prevent code rearrangement */
6400 volatile XLogCtlData *xlogctl = XLogCtl;
6403 * Update pg_control to show that we are recovering and to show the
6404 * selected checkpoint as the place we are starting from. We also mark
6405 * pg_control with any minimum recovery stop point obtained from a
6406 * backup history file.
6408 dbstate_at_startup = ControlFile->state;
6409 if (InArchiveRecovery)
6410 ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6414 (errmsg("database system was not properly shut down; "
6415 "automatic recovery in progress")));
6416 if (recoveryTargetTLI > ControlFile->checkPointCopy.ThisTimeLineID)
6418 (errmsg("crash recovery starts in timeline %u "
6419 "and has target timeline %u",
6420 ControlFile->checkPointCopy.ThisTimeLineID,
6421 recoveryTargetTLI)));
6422 ControlFile->state = DB_IN_CRASH_RECOVERY;
6424 ControlFile->prevCheckPoint = ControlFile->checkPoint;
6425 ControlFile->checkPoint = checkPointLoc;
6426 ControlFile->checkPointCopy = checkPoint;
6427 if (InArchiveRecovery)
6429 /* initialize minRecoveryPoint if not set yet */
6430 if (ControlFile->minRecoveryPoint < checkPoint.redo)
6432 ControlFile->minRecoveryPoint = checkPoint.redo;
6433 ControlFile->minRecoveryPointTLI = checkPoint.ThisTimeLineID;
6438 * Set backupStartPoint if we're starting recovery from a base backup.
6440 * Set backupEndPoint and use minRecoveryPoint as the backup end
6441 * location if we're starting recovery from a base backup which was
6442 * taken from the standby. In this case, the database system status in
6443 * pg_control must indicate DB_IN_ARCHIVE_RECOVERY. If not, which
6444 * means that backup is corrupted, so we cancel recovery.
6446 if (haveBackupLabel)
6448 ControlFile->backupStartPoint = checkPoint.redo;
6449 ControlFile->backupEndRequired = backupEndRequired;
6451 if (backupFromStandby)
6453 if (dbstate_at_startup != DB_IN_ARCHIVE_RECOVERY)
6455 (errmsg("backup_label contains data inconsistent with control file"),
6456 errhint("This means that the backup is corrupted and you will "
6457 "have to use another backup for recovery.")));
6458 ControlFile->backupEndPoint = ControlFile->minRecoveryPoint;
6461 ControlFile->time = (pg_time_t) time(NULL);
6462 /* No need to hold ControlFileLock yet, we aren't up far enough */
6463 UpdateControlFile();
6465 /* initialize our local copy of minRecoveryPoint */
6466 minRecoveryPoint = ControlFile->minRecoveryPoint;
6467 minRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6470 * Reset pgstat data, because it may be invalid after recovery.
6475 * If there was a backup label file, it's done its job and the info
6476 * has now been propagated into pg_control. We must get rid of the
6477 * label file so that if we crash during recovery, we'll pick up at
6478 * the latest recovery restartpoint instead of going all the way back
6479 * to the backup start point. It seems prudent though to just rename
6480 * the file out of the way rather than delete it completely.
6482 if (haveBackupLabel)
6484 unlink(BACKUP_LABEL_OLD);
6485 if (rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD) != 0)
6487 (errcode_for_file_access(),
6488 errmsg("could not rename file \"%s\" to \"%s\": %m",
6489 BACKUP_LABEL_FILE, BACKUP_LABEL_OLD)));
6492 /* Check that the GUCs used to generate the WAL allow recovery */
6493 CheckRequiredParameterValues();
6496 * We're in recovery, so unlogged relations may be trashed and must be
6497 * reset. This should be done BEFORE allowing Hot Standby
6498 * connections, so that read-only backends don't try to read whatever
6499 * garbage is left over from before.
6501 ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
6504 * Likewise, delete any saved transaction snapshot files that got left
6505 * behind by crashed backends.
6507 DeleteAllExportedSnapshotFiles();
6510 * Initialize for Hot Standby, if enabled. We won't let backends in
6511 * yet, not until we've reached the min recovery point specified in
6512 * control file and we've established a recovery snapshot from a
6513 * running-xacts WAL record.
6515 if (ArchiveRecoveryRequested && EnableHotStandby)
6517 TransactionId *xids;
6521 (errmsg("initializing for hot standby")));
6523 InitRecoveryTransactionEnvironment();
6526 oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
6528 oldestActiveXID = checkPoint.oldestActiveXid;
6529 Assert(TransactionIdIsValid(oldestActiveXID));
6531 /* Tell procarray about the range of xids it has to deal with */
6532 ProcArrayInitRecovery(ShmemVariableCache->nextXid);
6535 * Startup commit log and subtrans only. Other SLRUs are not
6536 * maintained during recovery and need not be started yet.
6539 StartupSUBTRANS(oldestActiveXID);
6542 * If we're beginning at a shutdown checkpoint, we know that
6543 * nothing was running on the master at this point. So fake-up an
6544 * empty running-xacts record and use that here and now. Recover
6545 * additional standby state for prepared transactions.
6549 RunningTransactionsData running;
6550 TransactionId latestCompletedXid;
6553 * Construct a RunningTransactions snapshot representing a
6554 * shut down server, with only prepared transactions still
6555 * alive. We're never overflowed at this point because all
6556 * subxids are listed with their parent prepared transactions.
6558 running.xcnt = nxids;
6559 running.subxcnt = 0;
6560 running.subxid_overflow = false;
6561 running.nextXid = checkPoint.nextXid;
6562 running.oldestRunningXid = oldestActiveXID;
6563 latestCompletedXid = checkPoint.nextXid;
6564 TransactionIdRetreat(latestCompletedXid);
6565 Assert(TransactionIdIsNormal(latestCompletedXid));
6566 running.latestCompletedXid = latestCompletedXid;
6567 running.xids = xids;
6569 ProcArrayApplyRecoveryInfo(&running);
6571 StandbyRecoverPreparedTransactions(false);
6575 /* Initialize resource managers */
6576 for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
6578 if (RmgrTable[rmid].rm_startup != NULL)
6579 RmgrTable[rmid].rm_startup();
6583 * Initialize shared replayEndRecPtr, lastReplayedEndRecPtr, and
6584 * recoveryLastXTime.
6586 * This is slightly confusing if we're starting from an online
6587 * checkpoint; we've just read and replayed the checkpoint record, but
6588 * we're going to start replay from its redo pointer, which precedes
6589 * the location of the checkpoint record itself. So even though the
6590 * last record we've replayed is indeed ReadRecPtr, we haven't
6591 * replayed all the preceding records yet. That's OK for the current
6592 * use of these variables.
6594 SpinLockAcquire(&xlogctl->info_lck);
6595 xlogctl->replayEndRecPtr = ReadRecPtr;
6596 xlogctl->replayEndTLI = ThisTimeLineID;
6597 xlogctl->lastReplayedEndRecPtr = EndRecPtr;
6598 xlogctl->lastReplayedTLI = ThisTimeLineID;
6599 xlogctl->recoveryLastXTime = 0;
6600 xlogctl->currentChunkStartTime = 0;
6601 xlogctl->recoveryPause = false;
6602 SpinLockRelease(&xlogctl->info_lck);
6604 /* Also ensure XLogReceiptTime has a sane value */
6605 XLogReceiptTime = GetCurrentTimestamp();
6608 * Let postmaster know we've started redo now, so that it can launch
6609 * checkpointer to perform restartpoints. We don't bother during
6610 * crash recovery as restartpoints can only be performed during
6611 * archive recovery. And we'd like to keep crash recovery simple, to
6612 * avoid introducing bugs that could affect you when recovering after
6615 * After this point, we can no longer assume that we're the only
6616 * process in addition to postmaster! Also, fsync requests are
6617 * subsequently to be handled by the checkpointer, not locally.
6619 if (ArchiveRecoveryRequested && IsUnderPostmaster)
6621 PublishStartupProcessInformation();
6622 SetForwardFsyncRequests();
6623 SendPostmasterSignal(PMSIGNAL_RECOVERY_STARTED);
6624 bgwriterLaunched = true;
6628 * Allow read-only connections immediately if we're consistent
6631 CheckRecoveryConsistency();
6634 * Find the first record that logically follows the checkpoint --- it
6635 * might physically precede it, though.
6637 if (checkPoint.redo < RecPtr)
6639 /* back up to find the record */
6640 record = ReadRecord(xlogreader, checkPoint.redo, PANIC, false);
6644 /* just have to read next record after CheckPoint */
6645 record = ReadRecord(xlogreader, InvalidXLogRecPtr, LOG, false);
6650 bool recoveryContinue = true;
6651 bool recoveryApply = true;
6652 ErrorContextCallback errcallback;
6658 (errmsg("redo starts at %X/%X",
6659 (uint32) (ReadRecPtr >> 32), (uint32) ReadRecPtr)));
6662 * main redo apply loop
6666 bool switchedTLI = false;
6670 (rmid == RM_XACT_ID && trace_recovery_messages <= DEBUG2) ||
6671 (rmid != RM_XACT_ID && trace_recovery_messages <= DEBUG3))
6675 initStringInfo(&buf);
6676 appendStringInfo(&buf, "REDO @ %X/%X; LSN %X/%X: ",
6677 (uint32) (ReadRecPtr >> 32), (uint32) ReadRecPtr,
6678 (uint32) (EndRecPtr >> 32), (uint32) EndRecPtr);
6679 xlog_outrec(&buf, record);
6680 appendStringInfo(&buf, " - ");
6681 RmgrTable[record->xl_rmid].rm_desc(&buf,
6683 XLogRecGetData(record));
6684 elog(LOG, "%s", buf.data);
6689 /* Handle interrupt signals of startup process */
6690 HandleStartupProcInterrupts();
6693 * Pause WAL replay, if requested by a hot-standby session via
6694 * SetRecoveryPause().
6696 * Note that we intentionally don't take the info_lck spinlock
6697 * here. We might therefore read a slightly stale value of
6698 * the recoveryPause flag, but it can't be very stale (no
6699 * worse than the last spinlock we did acquire). Since a
6700 * pause request is a pretty asynchronous thing anyway,
6701 * possibly responding to it one WAL record later than we
6702 * otherwise would is a minor issue, so it doesn't seem worth
6703 * adding another spinlock cycle to prevent that.
6705 if (xlogctl->recoveryPause)
6706 recoveryPausesHere();
6709 * Have we reached our recovery target?
6711 if (recoveryStopsHere(record, &recoveryApply))
6713 if (recoveryPauseAtTarget)
6715 SetRecoveryPause(true);
6716 recoveryPausesHere();
6718 reachedStopPoint = true; /* see below */
6719 recoveryContinue = false;
6721 /* Exit loop if we reached non-inclusive recovery target */
6726 /* Setup error traceback support for ereport() */
6727 errcallback.callback = rm_redo_error_callback;
6728 errcallback.arg = (void *) record;
6729 errcallback.previous = error_context_stack;
6730 error_context_stack = &errcallback;
6733 * ShmemVariableCache->nextXid must be beyond record's xid.
6735 * We don't expect anyone else to modify nextXid, hence we
6736 * don't need to hold a lock while examining it. We still
6737 * acquire the lock to modify it, though.
6739 if (TransactionIdFollowsOrEquals(record->xl_xid,
6740 ShmemVariableCache->nextXid))
6742 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
6743 ShmemVariableCache->nextXid = record->xl_xid;
6744 TransactionIdAdvance(ShmemVariableCache->nextXid);
6745 LWLockRelease(XidGenLock);
6749 * Before replaying this record, check if this record causes
6750 * the current timeline to change. The record is already
6751 * considered to be part of the new timeline, so we update
6752 * ThisTimeLineID before replaying it. That's important so
6753 * that replayEndTLI, which is recorded as the minimum
6754 * recovery point's TLI if recovery stops after this record,
6757 if (record->xl_rmid == RM_XLOG_ID)
6759 TimeLineID newTLI = ThisTimeLineID;
6760 TimeLineID prevTLI = ThisTimeLineID;
6761 uint8 info = record->xl_info & ~XLR_INFO_MASK;
6763 if (info == XLOG_CHECKPOINT_SHUTDOWN)
6765 CheckPoint checkPoint;
6767 memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
6768 newTLI = checkPoint.ThisTimeLineID;
6769 prevTLI = checkPoint.PrevTimeLineID;
6771 else if (info == XLOG_END_OF_RECOVERY)
6773 xl_end_of_recovery xlrec;
6775 memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
6776 newTLI = xlrec.ThisTimeLineID;
6777 prevTLI = xlrec.PrevTimeLineID;
6780 if (newTLI != ThisTimeLineID)
6782 /* Check that it's OK to switch to this TLI */
6783 checkTimeLineSwitch(EndRecPtr, newTLI, prevTLI);
6785 /* Following WAL records should be run with new TLI */
6786 ThisTimeLineID = newTLI;
6792 * Update shared replayEndRecPtr before replaying this record,
6793 * so that XLogFlush will update minRecoveryPoint correctly.
6795 SpinLockAcquire(&xlogctl->info_lck);
6796 xlogctl->replayEndRecPtr = EndRecPtr;
6797 xlogctl->replayEndTLI = ThisTimeLineID;
6798 SpinLockRelease(&xlogctl->info_lck);
6801 * If we are attempting to enter Hot Standby mode, process
6804 if (standbyState >= STANDBY_INITIALIZED &&
6805 TransactionIdIsValid(record->xl_xid))
6806 RecordKnownAssignedTransactionIds(record->xl_xid);
6808 /* Now apply the WAL record itself */
6809 RmgrTable[record->xl_rmid].rm_redo(EndRecPtr, record);
6811 /* Pop the error context stack */
6812 error_context_stack = errcallback.previous;
6815 * Update lastReplayedEndRecPtr after this record has been
6816 * successfully replayed.
6818 SpinLockAcquire(&xlogctl->info_lck);
6819 xlogctl->lastReplayedEndRecPtr = EndRecPtr;
6820 xlogctl->lastReplayedTLI = ThisTimeLineID;
6821 SpinLockRelease(&xlogctl->info_lck);
6823 /* Remember this record as the last-applied one */
6824 LastRec = ReadRecPtr;
6826 /* Allow read-only connections if we're consistent now */
6827 CheckRecoveryConsistency();
6830 * If this record was a timeline switch, wake up any
6831 * walsenders to notice that we are on a new timeline.
6833 if (switchedTLI && AllowCascadeReplication())
6836 /* Exit loop if we reached inclusive recovery target */
6837 if (!recoveryContinue)
6840 /* Else, try to fetch the next WAL record */
6841 record = ReadRecord(xlogreader, InvalidXLogRecPtr, LOG, false);
6842 } while (record != NULL);
6845 * end of main redo apply loop
6849 (errmsg("redo done at %X/%X",
6850 (uint32) (ReadRecPtr >> 32), (uint32) ReadRecPtr)));
6851 xtime = GetLatestXTime();
6854 (errmsg("last completed transaction was at log time %s",
6855 timestamptz_to_str(xtime))));
6860 /* there are no WAL records following the checkpoint */
6862 (errmsg("redo is not required")));
6867 * Kill WAL receiver, if it's still running, before we continue to write
6868 * the startup checkpoint record. It will trump over the checkpoint and
6869 * subsequent records if it's still alive when we start writing WAL.
6874 * We don't need the latch anymore. It's not strictly necessary to disown
6875 * it, but let's do it for the sake of tidiness.
6877 if (StandbyModeRequested)
6878 DisownLatch(&XLogCtl->recoveryWakeupLatch);
6881 * We are now done reading the xlog from stream. Turn off streaming
6882 * recovery to force fetching the files (which would be required at end of
6883 * recovery, e.g., timeline history file) from archive or pg_xlog.
6885 StandbyMode = false;
6888 * Re-fetch the last valid or last applied record, so we can identify the
6889 * exact endpoint of what we consider the valid portion of WAL.
6891 record = ReadRecord(xlogreader, LastRec, PANIC, false);
6892 EndOfLog = EndRecPtr;
6893 XLByteToPrevSeg(EndOfLog, endLogSegNo);
6896 * Complain if we did not roll forward far enough to render the backup
6897 * dump consistent. Note: it is indeed okay to look at the local variable
6898 * minRecoveryPoint here, even though ControlFile->minRecoveryPoint might
6899 * be further ahead --- ControlFile->minRecoveryPoint cannot have been
6900 * advanced beyond the WAL we processed.
6903 (EndOfLog < minRecoveryPoint ||
6904 !XLogRecPtrIsInvalid(ControlFile->backupStartPoint)))
6906 if (reachedStopPoint)
6908 /* stopped because of stop request */
6910 (errmsg("requested recovery stop point is before consistent recovery point")));
6914 * Ran off end of WAL before reaching end-of-backup WAL record, or
6915 * minRecoveryPoint. That's usually a bad sign, indicating that you
6916 * tried to recover from an online backup but never called
6917 * pg_stop_backup(), or you didn't archive all the WAL up to that
6918 * point. However, this also happens in crash recovery, if the system
6919 * crashes while an online backup is in progress. We must not treat
6920 * that as an error, or the database will refuse to start up.
6922 if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
6924 if (ControlFile->backupEndRequired)
6926 (errmsg("WAL ends before end of online backup"),
6927 errhint("All WAL generated while online backup was taken must be available at recovery.")));
6928 else if (!XLogRecPtrIsInvalid(ControlFile->backupStartPoint))
6930 (errmsg("WAL ends before end of online backup"),
6931 errhint("Online backup started with pg_start_backup() must be ended with pg_stop_backup(), and all WAL up to that point must be available at recovery.")));
6934 (errmsg("WAL ends before consistent recovery point")));
6939 * Consider whether we need to assign a new timeline ID.
6941 * If we are doing an archive recovery, we always assign a new ID. This
6942 * handles a couple of issues. If we stopped short of the end of WAL
6943 * during recovery, then we are clearly generating a new timeline and must
6944 * assign it a unique new ID. Even if we ran to the end, modifying the
6945 * current last segment is problematic because it may result in trying to
6946 * overwrite an already-archived copy of that segment, and we encourage
6947 * DBAs to make their archive_commands reject that. We can dodge the
6948 * problem by making the new active segment have a new timeline ID.
6950 * In a normal crash recovery, we can just extend the timeline we were in.
6952 PrevTimeLineID = ThisTimeLineID;
6953 if (ArchiveRecoveryRequested)
6957 Assert(InArchiveRecovery);
6959 ThisTimeLineID = findNewestTimeLine(recoveryTargetTLI) + 1;
6961 (errmsg("selected new timeline ID: %u", ThisTimeLineID)));
6964 * Create a comment for the history file to explain why and where
6967 if (recoveryTarget == RECOVERY_TARGET_XID)
6968 snprintf(reason, sizeof(reason),
6969 "%s transaction %u",
6970 recoveryStopAfter ? "after" : "before",
6972 else if (recoveryTarget == RECOVERY_TARGET_TIME)
6973 snprintf(reason, sizeof(reason),
6975 recoveryStopAfter ? "after" : "before",
6976 timestamptz_to_str(recoveryStopTime));
6977 else if (recoveryTarget == RECOVERY_TARGET_NAME)
6978 snprintf(reason, sizeof(reason),
6979 "at restore point \"%s\"",
6982 snprintf(reason, sizeof(reason), "no recovery target specified");
6984 writeTimeLineHistory(ThisTimeLineID, recoveryTargetTLI,
6988 /* Save the selected TimeLineID in shared memory, too */
6989 XLogCtl->ThisTimeLineID = ThisTimeLineID;
6990 XLogCtl->PrevTimeLineID = PrevTimeLineID;
6993 * We are now done reading the old WAL. Turn off archive fetching if it
6994 * was active, and make a writable copy of the last WAL segment. (Note
6995 * that we also have a copy of the last block of the old WAL in readBuf;
6996 * we will use that below.)
6998 if (ArchiveRecoveryRequested)
6999 exitArchiveRecovery(xlogreader->readPageTLI, endLogSegNo);
7002 * Prepare to write WAL starting at EndOfLog position, and init xlog
7003 * buffer cache using the block containing the last record from the
7004 * previous incarnation.
7006 openLogSegNo = endLogSegNo;
7007 openLogFile = XLogFileOpen(openLogSegNo);
7009 Insert = &XLogCtl->Insert;
7010 Insert->PrevBytePos = XLogRecPtrToBytePos(LastRec);
7011 Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
7014 * Tricky point here: readBuf contains the *last* block that the LastRec
7015 * record spans, not the one it starts in. The last block is indeed the
7016 * one we want to use.
7018 if (EndOfLog % XLOG_BLCKSZ != 0)
7023 XLogRecPtr pageBeginPtr;
7025 pageBeginPtr = EndOfLog - (EndOfLog % XLOG_BLCKSZ);
7026 Assert(readOff == pageBeginPtr % XLogSegSize);
7028 firstIdx = XLogRecPtrToBufIdx(EndOfLog);
7030 /* Copy the valid part of the last block, and zero the rest */
7031 page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
7032 len = EndOfLog % XLOG_BLCKSZ;
7033 memcpy(page, xlogreader->readBuf, len);
7034 memset(page + len, 0, XLOG_BLCKSZ - len);
7036 XLogCtl->xlblocks[firstIdx] = pageBeginPtr + XLOG_BLCKSZ;
7037 XLogCtl->InitializedUpTo = pageBeginPtr + XLOG_BLCKSZ;
7042 * There is no partial block to copy. Just set InitializedUpTo,
7043 * and let the first attempt to insert a log record to initialize
7046 XLogCtl->InitializedUpTo = EndOfLog;
7049 LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
7051 XLogCtl->LogwrtResult = LogwrtResult;
7053 XLogCtl->LogwrtRqst.Write = EndOfLog;
7054 XLogCtl->LogwrtRqst.Flush = EndOfLog;
7056 /* Pre-scan prepared transactions to find out the range of XIDs present */
7057 oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
7060 * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
7061 * record before resource manager writes cleanup WAL records or checkpoint
7062 * record is written.
7064 Insert->fullPageWrites = lastFullPageWrites;
7065 LocalSetXLogInsertAllowed();
7066 UpdateFullPageWrites();
7067 LocalXLogInsertAllowed = -1;
7074 * Resource managers might need to write WAL records, eg, to record
7075 * index cleanup actions. So temporarily enable XLogInsertAllowed in
7076 * this process only.
7078 LocalSetXLogInsertAllowed();
7081 * Allow resource managers to do any required cleanup.
7083 for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
7085 if (RmgrTable[rmid].rm_cleanup != NULL)
7086 RmgrTable[rmid].rm_cleanup();
7089 /* Disallow XLogInsert again */
7090 LocalXLogInsertAllowed = -1;
7093 * Perform a checkpoint to update all our recovery activity to disk.
7095 * Note that we write a shutdown checkpoint rather than an on-line
7096 * one. This is not particularly critical, but since we may be
7097 * assigning a new TLI, using a shutdown checkpoint allows us to have
7098 * the rule that TLI only changes in shutdown checkpoints, which
7099 * allows some extra error checking in xlog_redo.
7101 * In fast promotion, only create a lightweight end-of-recovery record
7102 * instead of a full checkpoint. A checkpoint is requested later,
7103 * after we're fully out of recovery mode and already accepting
7106 if (bgwriterLaunched)
7110 checkPointLoc = ControlFile->prevCheckPoint;
7113 * Confirm the last checkpoint is available for us to recover
7114 * from if we fail. Note that we don't check for the secondary
7115 * checkpoint since that isn't available in most base backups.
7117 record = ReadCheckpointRecord(xlogreader, checkPointLoc, 1, false);
7120 fast_promoted = true;
7123 * Insert a special WAL record to mark the end of
7124 * recovery, since we aren't doing a checkpoint. That
7125 * means that the checkpointer process may likely be in
7126 * the middle of a time-smoothed restartpoint and could
7127 * continue to be for minutes after this. That sounds
7128 * strange, but the effect is roughly the same and it
7129 * would be stranger to try to come out of the
7130 * restartpoint and then checkpoint. We request a
7131 * checkpoint later anyway, just for safety.
7133 CreateEndOfRecoveryRecord();
7138 RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
7139 CHECKPOINT_IMMEDIATE |
7143 CreateCheckPoint(CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IMMEDIATE);
7146 * And finally, execute the recovery_end_command, if any.
7148 if (recoveryEndCommand)
7149 ExecuteRecoveryCommand(recoveryEndCommand,
7150 "recovery_end_command",
7155 * Preallocate additional log files, if wanted.
7157 PreallocXlogFiles(EndOfLog);
7160 * Reset initial contents of unlogged relations. This has to be done
7161 * AFTER recovery is complete so that any unlogged relations created
7162 * during recovery also get picked up.
7165 ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
7168 * Okay, we're officially UP.
7172 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7173 ControlFile->state = DB_IN_PRODUCTION;
7174 ControlFile->time = (pg_time_t) time(NULL);
7175 UpdateControlFile();
7176 LWLockRelease(ControlFileLock);
7178 /* start the archive_timeout timer running */
7179 XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
7181 /* also initialize latestCompletedXid, to nextXid - 1 */
7182 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
7183 ShmemVariableCache->latestCompletedXid = ShmemVariableCache->nextXid;
7184 TransactionIdRetreat(ShmemVariableCache->latestCompletedXid);
7185 LWLockRelease(ProcArrayLock);
7188 * Start up the commit log and subtrans, if not already done for hot
7191 if (standbyState == STANDBY_DISABLED)
7194 StartupSUBTRANS(oldestActiveXID);
7198 * Perform end of recovery actions for any SLRUs that need it.
7203 /* Reload shared-memory state for prepared transactions */
7204 RecoverPreparedTransactions();
7207 * Shutdown the recovery environment. This must occur after
7208 * RecoverPreparedTransactions(), see notes for lock_twophase_recover()
7210 if (standbyState != STANDBY_DISABLED)
7211 ShutdownRecoveryTransactionEnvironment();
7213 /* Shut down xlogreader */
7219 XLogReaderFree(xlogreader);
7222 * If any of the critical GUCs have changed, log them before we allow
7223 * backends to write WAL.
7225 LocalSetXLogInsertAllowed();
7226 XLogReportParameters();
7229 * All done. Allow backends to write WAL. (Although the bool flag is
7230 * probably atomic in itself, we use the info_lck here to ensure that
7231 * there are no race conditions concerning visibility of other recent
7232 * updates to shared memory.)
7235 /* use volatile pointer to prevent code rearrangement */
7236 volatile XLogCtlData *xlogctl = XLogCtl;
7238 SpinLockAcquire(&xlogctl->info_lck);
7239 xlogctl->SharedRecoveryInProgress = false;
7240 SpinLockRelease(&xlogctl->info_lck);
7244 * If there were cascading standby servers connected to us, nudge any wal
7245 * sender processes to notice that we've been promoted.
7250 * If this was a fast promotion, request an (online) checkpoint now. This
7251 * isn't required for consistency, but the last restartpoint might be far
7252 * back, and in case of a crash, recovering from it might take a longer
7253 * than is appropriate now that we're not in standby mode anymore.
7256 RequestCheckpoint(CHECKPOINT_FORCE);
7260 * Checks if recovery has reached a consistent state. When consistency is
7261 * reached and we have a valid starting standby snapshot, tell postmaster
7262 * that it can start accepting read-only connections.
7265 CheckRecoveryConsistency(void)
7268 * During crash recovery, we don't reach a consistent state until we've
7269 * replayed all the WAL.
7271 if (XLogRecPtrIsInvalid(minRecoveryPoint))
7275 * Have we reached the point where our base backup was completed?
7277 if (!XLogRecPtrIsInvalid(ControlFile->backupEndPoint) &&
7278 ControlFile->backupEndPoint <= EndRecPtr)
7281 * We have reached the end of base backup, as indicated by pg_control.
7282 * The data on disk is now consistent. Reset backupStartPoint and
7283 * backupEndPoint, and update minRecoveryPoint to make sure we don't
7284 * allow starting up at an earlier point even if recovery is stopped
7285 * and restarted soon after this.
7287 elog(DEBUG1, "end of backup reached");
7289 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7291 if (ControlFile->minRecoveryPoint < EndRecPtr)
7292 ControlFile->minRecoveryPoint = EndRecPtr;
7294 ControlFile->backupStartPoint = InvalidXLogRecPtr;
7295 ControlFile->backupEndPoint = InvalidXLogRecPtr;
7296 ControlFile->backupEndRequired = false;
7297 UpdateControlFile();
7299 LWLockRelease(ControlFileLock);
7303 * Have we passed our safe starting point? Note that minRecoveryPoint is
7304 * known to be incorrectly set if ControlFile->backupEndRequired, until
7305 * the XLOG_BACKUP_RECORD arrives to advise us of the correct
7306 * minRecoveryPoint. All we know prior to that is that we're not
7309 if (!reachedConsistency && !ControlFile->backupEndRequired &&
7310 minRecoveryPoint <= XLogCtl->lastReplayedEndRecPtr &&
7311 XLogRecPtrIsInvalid(ControlFile->backupStartPoint))
7314 * Check to see if the XLOG sequence contained any unresolved
7315 * references to uninitialized pages.
7317 XLogCheckInvalidPages();
7319 reachedConsistency = true;
7321 (errmsg("consistent recovery state reached at %X/%X",
7322 (uint32) (XLogCtl->lastReplayedEndRecPtr >> 32),
7323 (uint32) XLogCtl->lastReplayedEndRecPtr)));
7327 * Have we got a valid starting snapshot that will allow queries to be
7328 * run? If so, we can tell postmaster that the database is consistent now,
7329 * enabling connections.
7331 if (standbyState == STANDBY_SNAPSHOT_READY &&
7332 !LocalHotStandbyActive &&
7333 reachedConsistency &&
7336 /* use volatile pointer to prevent code rearrangement */
7337 volatile XLogCtlData *xlogctl = XLogCtl;
7339 SpinLockAcquire(&xlogctl->info_lck);
7340 xlogctl->SharedHotStandbyActive = true;
7341 SpinLockRelease(&xlogctl->info_lck);
7343 LocalHotStandbyActive = true;
7345 SendPostmasterSignal(PMSIGNAL_BEGIN_HOT_STANDBY);
7350 * Is the system still in recovery?
7352 * Unlike testing InRecovery, this works in any process that's connected to
7355 * As a side-effect, we initialize the local TimeLineID and RedoRecPtr
7356 * variables the first time we see that recovery is finished.
7359 RecoveryInProgress(void)
7362 * We check shared state each time only until we leave recovery mode. We
7363 * can't re-enter recovery, so there's no need to keep checking after the
7364 * shared variable has once been seen false.
7366 if (!LocalRecoveryInProgress)
7370 /* use volatile pointer to prevent code rearrangement */
7371 volatile XLogCtlData *xlogctl = XLogCtl;
7373 /* spinlock is essential on machines with weak memory ordering! */
7374 SpinLockAcquire(&xlogctl->info_lck);
7375 LocalRecoveryInProgress = xlogctl->SharedRecoveryInProgress;
7376 SpinLockRelease(&xlogctl->info_lck);
7379 * Initialize TimeLineID and RedoRecPtr when we discover that recovery
7380 * is finished. InitPostgres() relies upon this behaviour to ensure
7381 * that InitXLOGAccess() is called at backend startup. (If you change
7382 * this, see also LocalSetXLogInsertAllowed.)
7384 if (!LocalRecoveryInProgress)
7387 return LocalRecoveryInProgress;
7392 * Is HotStandby active yet? This is only important in special backends
7393 * since normal backends won't ever be able to connect until this returns
7394 * true. Postmaster knows this by way of signal, not via shared memory.
7396 * Unlike testing standbyState, this works in any process that's connected to
7400 HotStandbyActive(void)
7403 * We check shared state each time only until Hot Standby is active. We
7404 * can't de-activate Hot Standby, so there's no need to keep checking
7405 * after the shared variable has once been seen true.
7407 if (LocalHotStandbyActive)
7411 /* use volatile pointer to prevent code rearrangement */
7412 volatile XLogCtlData *xlogctl = XLogCtl;
7414 /* spinlock is essential on machines with weak memory ordering! */
7415 SpinLockAcquire(&xlogctl->info_lck);
7416 LocalHotStandbyActive = xlogctl->SharedHotStandbyActive;
7417 SpinLockRelease(&xlogctl->info_lck);
7419 return LocalHotStandbyActive;
7424 * Is this process allowed to insert new WAL records?
7426 * Ordinarily this is essentially equivalent to !RecoveryInProgress().
7427 * But we also have provisions for forcing the result "true" or "false"
7428 * within specific processes regardless of the global state.
7431 XLogInsertAllowed(void)
7434 * If value is "unconditionally true" or "unconditionally false", just
7435 * return it. This provides the normal fast path once recovery is known
7438 if (LocalXLogInsertAllowed >= 0)
7439 return (bool) LocalXLogInsertAllowed;
7442 * Else, must check to see if we're still in recovery.
7444 if (RecoveryInProgress())
7448 * On exit from recovery, reset to "unconditionally true", since there is
7449 * no need to keep checking.
7451 LocalXLogInsertAllowed = 1;
7456 * Make XLogInsertAllowed() return true in the current process only.
7458 * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
7459 * and even call LocalSetXLogInsertAllowed() again after that.
7462 LocalSetXLogInsertAllowed(void)
7464 Assert(LocalXLogInsertAllowed == -1);
7465 LocalXLogInsertAllowed = 1;
7467 /* Initialize as RecoveryInProgress() would do when switching state */
7472 * Subroutine to try to fetch and validate a prior checkpoint record.
7474 * whichChkpt identifies the checkpoint (merely for reporting purposes).
7475 * 1 for "primary", 2 for "secondary", 0 for "other" (backup_label)
7478 ReadCheckpointRecord(XLogReaderState *xlogreader, XLogRecPtr RecPtr,
7479 int whichChkpt, bool report)
7483 if (!XRecOffIsValid(RecPtr))
7492 (errmsg("invalid primary checkpoint link in control file")));
7496 (errmsg("invalid secondary checkpoint link in control file")));
7500 (errmsg("invalid checkpoint link in backup_label file")));
7506 record = ReadRecord(xlogreader, RecPtr, LOG, true);
7517 (errmsg("invalid primary checkpoint record")));
7521 (errmsg("invalid secondary checkpoint record")));
7525 (errmsg("invalid checkpoint record")));
7530 if (record->xl_rmid != RM_XLOG_ID)
7536 (errmsg("invalid resource manager ID in primary checkpoint record")));
7540 (errmsg("invalid resource manager ID in secondary checkpoint record")));
7544 (errmsg("invalid resource manager ID in checkpoint record")));
7549 if (record->xl_info != XLOG_CHECKPOINT_SHUTDOWN &&
7550 record->xl_info != XLOG_CHECKPOINT_ONLINE)
7556 (errmsg("invalid xl_info in primary checkpoint record")));
7560 (errmsg("invalid xl_info in secondary checkpoint record")));
7564 (errmsg("invalid xl_info in checkpoint record")));
7569 if (record->xl_len != sizeof(CheckPoint) ||
7570 record->xl_tot_len != SizeOfXLogRecord + sizeof(CheckPoint))
7576 (errmsg("invalid length of primary checkpoint record")));
7580 (errmsg("invalid length of secondary checkpoint record")));
7584 (errmsg("invalid length of checkpoint record")));
7593 * This must be called during startup of a backend process, except that
7594 * it need not be called in a standalone backend (which does StartupXLOG
7595 * instead). We need to initialize the local copies of ThisTimeLineID and
7598 * Note: before Postgres 8.0, we went to some effort to keep the postmaster
7599 * process's copies of ThisTimeLineID and RedoRecPtr valid too. This was
7600 * unnecessary however, since the postmaster itself never touches XLOG anyway.
7603 InitXLOGAccess(void)
7605 /* ThisTimeLineID doesn't change so we need no lock to copy it */
7606 ThisTimeLineID = XLogCtl->ThisTimeLineID;
7607 Assert(ThisTimeLineID != 0 || IsBootstrapProcessingMode());
7609 /* Use GetRedoRecPtr to copy the RedoRecPtr safely */
7610 (void) GetRedoRecPtr();
7614 * Return the current Redo pointer from shared memory.
7616 * As a side-effect, the local RedoRecPtr copy is updated.
7621 /* use volatile pointer to prevent code rearrangement */
7622 volatile XLogCtlData *xlogctl = XLogCtl;
7626 * The possibly not up-to-date copy in XlogCtl is enough. Even if we
7627 * grabbed a WAL insertion slot to read the master copy, someone might
7628 * update it just after we've released the lock.
7630 SpinLockAcquire(&xlogctl->info_lck);
7631 ptr = xlogctl->RedoRecPtr;
7632 SpinLockRelease(&xlogctl->info_lck);
7634 if (RedoRecPtr < ptr)
7641 * GetInsertRecPtr -- Returns the current insert position.
7643 * NOTE: The value *actually* returned is the position of the last full
7644 * xlog page. It lags behind the real insert position by at most 1 page.
7645 * For that, we don't need to scan through WAL insertion slots, and an
7646 * approximation is enough for the current usage of this function.
7649 GetInsertRecPtr(void)
7651 /* use volatile pointer to prevent code rearrangement */
7652 volatile XLogCtlData *xlogctl = XLogCtl;
7655 SpinLockAcquire(&xlogctl->info_lck);
7656 recptr = xlogctl->LogwrtRqst.Write;
7657 SpinLockRelease(&xlogctl->info_lck);
7663 * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
7664 * position known to be fsync'd to disk.
7667 GetFlushRecPtr(void)
7669 /* use volatile pointer to prevent code rearrangement */
7670 volatile XLogCtlData *xlogctl = XLogCtl;
7673 SpinLockAcquire(&xlogctl->info_lck);
7674 recptr = xlogctl->LogwrtResult.Flush;
7675 SpinLockRelease(&xlogctl->info_lck);
7681 * Get the time of the last xlog segment switch
7684 GetLastSegSwitchTime(void)
7688 /* Need WALWriteLock, but shared lock is sufficient */
7689 LWLockAcquire(WALWriteLock, LW_SHARED);
7690 result = XLogCtl->lastSegSwitchTime;
7691 LWLockRelease(WALWriteLock);
7697 * GetNextXidAndEpoch - get the current nextXid value and associated epoch
7699 * This is exported for use by code that would like to have 64-bit XIDs.
7700 * We don't really support such things, but all XIDs within the system
7701 * can be presumed "close to" the result, and thus the epoch associated
7702 * with them can be determined.
7705 GetNextXidAndEpoch(TransactionId *xid, uint32 *epoch)
7707 uint32 ckptXidEpoch;
7708 TransactionId ckptXid;
7709 TransactionId nextXid;
7711 /* Must read checkpoint info first, else have race condition */
7713 /* use volatile pointer to prevent code rearrangement */
7714 volatile XLogCtlData *xlogctl = XLogCtl;
7716 SpinLockAcquire(&xlogctl->info_lck);
7717 ckptXidEpoch = xlogctl->ckptXidEpoch;
7718 ckptXid = xlogctl->ckptXid;
7719 SpinLockRelease(&xlogctl->info_lck);
7722 /* Now fetch current nextXid */
7723 nextXid = ReadNewTransactionId();
7726 * nextXid is certainly logically later than ckptXid. So if it's
7727 * numerically less, it must have wrapped into the next epoch.
7729 if (nextXid < ckptXid)
7733 *epoch = ckptXidEpoch;
7737 * This must be called ONCE during postmaster or standalone-backend shutdown
7740 ShutdownXLOG(int code, Datum arg)
7742 /* Don't be chatty in standalone mode */
7743 ereport(IsPostmasterEnvironment ? LOG : NOTICE,
7744 (errmsg("shutting down")));
7746 if (RecoveryInProgress())
7747 CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_IMMEDIATE);
7751 * If archiving is enabled, rotate the last XLOG file so that all the
7752 * remaining records are archived (postmaster wakes up the archiver
7753 * process one more time at the end of shutdown). The checkpoint
7754 * record will go to the next XLOG file and won't be archived (yet).
7756 if (XLogArchivingActive() && XLogArchiveCommandSet())
7757 RequestXLogSwitch();
7759 CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_IMMEDIATE);
7763 ShutdownMultiXact();
7765 /* Don't be chatty in standalone mode */
7766 ereport(IsPostmasterEnvironment ? LOG : NOTICE,
7767 (errmsg("database system is shut down")));
7771 * Log start of a checkpoint.
7774 LogCheckpointStart(int flags, bool restartpoint)
7779 * XXX: This is hopelessly untranslatable. We could call gettext_noop for
7780 * the main message, but what about all the flags?
7783 msg = "restartpoint starting:%s%s%s%s%s%s%s";
7785 msg = "checkpoint starting:%s%s%s%s%s%s%s";
7788 (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
7789 (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
7790 (flags & CHECKPOINT_IMMEDIATE) ? " immediate" : "",
7791 (flags & CHECKPOINT_FORCE) ? " force" : "",
7792 (flags & CHECKPOINT_WAIT) ? " wait" : "",
7793 (flags & CHECKPOINT_CAUSE_XLOG) ? " xlog" : "",
7794 (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "");
7798 * Log end of a checkpoint.
7801 LogCheckpointEnd(bool restartpoint)
7813 uint64 average_sync_time;
7815 CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
7817 TimestampDifference(CheckpointStats.ckpt_write_t,
7818 CheckpointStats.ckpt_sync_t,
7819 &write_secs, &write_usecs);
7821 TimestampDifference(CheckpointStats.ckpt_sync_t,
7822 CheckpointStats.ckpt_sync_end_t,
7823 &sync_secs, &sync_usecs);
7825 /* Accumulate checkpoint timing summary data, in milliseconds. */
7826 BgWriterStats.m_checkpoint_write_time +=
7827 write_secs * 1000 + write_usecs / 1000;
7828 BgWriterStats.m_checkpoint_sync_time +=
7829 sync_secs * 1000 + sync_usecs / 1000;
7832 * All of the published timing statistics are accounted for. Only
7833 * continue if a log message is to be written.
7835 if (!log_checkpoints)
7838 TimestampDifference(CheckpointStats.ckpt_start_t,
7839 CheckpointStats.ckpt_end_t,
7840 &total_secs, &total_usecs);
7843 * Timing values returned from CheckpointStats are in microseconds.
7844 * Convert to the second plus microsecond form that TimestampDifference
7845 * returns for homogeneous printing.
7847 longest_secs = (long) (CheckpointStats.ckpt_longest_sync / 1000000);
7848 longest_usecs = CheckpointStats.ckpt_longest_sync -
7849 (uint64) longest_secs *1000000;
7851 average_sync_time = 0;
7852 if (CheckpointStats.ckpt_sync_rels > 0)
7853 average_sync_time = CheckpointStats.ckpt_agg_sync_time /
7854 CheckpointStats.ckpt_sync_rels;
7855 average_secs = (long) (average_sync_time / 1000000);
7856 average_usecs = average_sync_time - (uint64) average_secs *1000000;
7859 elog(LOG, "restartpoint complete: wrote %d buffers (%.1f%%); "
7860 "%d transaction log file(s) added, %d removed, %d recycled; "
7861 "write=%ld.%03d s, sync=%ld.%03d s, total=%ld.%03d s; "
7862 "sync files=%d, longest=%ld.%03d s, average=%ld.%03d s",
7863 CheckpointStats.ckpt_bufs_written,
7864 (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
7865 CheckpointStats.ckpt_segs_added,
7866 CheckpointStats.ckpt_segs_removed,
7867 CheckpointStats.ckpt_segs_recycled,
7868 write_secs, write_usecs / 1000,
7869 sync_secs, sync_usecs / 1000,
7870 total_secs, total_usecs / 1000,
7871 CheckpointStats.ckpt_sync_rels,
7872 longest_secs, longest_usecs / 1000,
7873 average_secs, average_usecs / 1000);
7875 elog(LOG, "checkpoint complete: wrote %d buffers (%.1f%%); "
7876 "%d transaction log file(s) added, %d removed, %d recycled; "
7877 "write=%ld.%03d s, sync=%ld.%03d s, total=%ld.%03d s; "
7878 "sync files=%d, longest=%ld.%03d s, average=%ld.%03d s",
7879 CheckpointStats.ckpt_bufs_written,
7880 (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
7881 CheckpointStats.ckpt_segs_added,
7882 CheckpointStats.ckpt_segs_removed,
7883 CheckpointStats.ckpt_segs_recycled,
7884 write_secs, write_usecs / 1000,
7885 sync_secs, sync_usecs / 1000,
7886 total_secs, total_usecs / 1000,
7887 CheckpointStats.ckpt_sync_rels,
7888 longest_secs, longest_usecs / 1000,
7889 average_secs, average_usecs / 1000);
7893 * Perform a checkpoint --- either during shutdown, or on-the-fly
7895 * flags is a bitwise OR of the following:
7896 * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
7897 * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
7898 * CHECKPOINT_IMMEDIATE: finish the checkpoint ASAP,
7899 * ignoring checkpoint_completion_target parameter.
7900 * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
7901 * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
7902 * CHECKPOINT_END_OF_RECOVERY).
7904 * Note: flags contains other bits, of interest here only for logging purposes.
7905 * In particular note that this routine is synchronous and does not pay
7906 * attention to CHECKPOINT_WAIT.
7908 * If !shutdown then we are writing an online checkpoint. This is a very special
7909 * kind of operation and WAL record because the checkpoint action occurs over
7910 * a period of time yet logically occurs at just a single LSN. The logical
7911 * position of the WAL record (redo ptr) is the same or earlier than the
7912 * physical position. When we replay WAL we locate the checkpoint via its
7913 * physical position then read the redo ptr and actually start replay at the
7914 * earlier logical position. Note that we don't write *anything* to WAL at
7915 * the logical position, so that location could be any other kind of WAL record.
7916 * All of this mechanism allows us to continue working while we checkpoint.
7917 * As a result, timing of actions is critical here and be careful to note that
7918 * this function will likely take minutes to execute on a busy system.
7921 CreateCheckPoint(int flags)
7923 /* use volatile pointer to prevent code rearrangement */
7924 volatile XLogCtlData *xlogctl = XLogCtl;
7926 CheckPoint checkPoint;
7928 XLogCtlInsert *Insert = &XLogCtl->Insert;
7931 XLogSegNo _logSegNo;
7932 XLogRecPtr curInsert;
7933 VirtualTransactionId *vxids;
7937 * An end-of-recovery checkpoint is really a shutdown checkpoint, just
7938 * issued at a different time.
7940 if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
7946 if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
7947 elog(ERROR, "can't create a checkpoint during recovery");
7950 * Acquire CheckpointLock to ensure only one checkpoint happens at a time.
7951 * (This is just pro forma, since in the present system structure there is
7952 * only one process that is allowed to issue checkpoints at any given
7955 LWLockAcquire(CheckpointLock, LW_EXCLUSIVE);
7958 * Prepare to accumulate statistics.
7960 * Note: because it is possible for log_checkpoints to change while a
7961 * checkpoint proceeds, we always accumulate stats, even if
7962 * log_checkpoints is currently off.
7964 MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7965 CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7968 * Use a critical section to force system panic if we have trouble.
7970 START_CRIT_SECTION();
7974 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7975 ControlFile->state = DB_SHUTDOWNING;
7976 ControlFile->time = (pg_time_t) time(NULL);
7977 UpdateControlFile();
7978 LWLockRelease(ControlFileLock);
7982 * Let smgr prepare for checkpoint; this has to happen before we determine
7983 * the REDO pointer. Note that smgr must not do anything that'd have to
7984 * be undone if we decide no checkpoint is needed.
7988 /* Begin filling in the checkpoint WAL record */
7989 MemSet(&checkPoint, 0, sizeof(checkPoint));
7990 checkPoint.time = (pg_time_t) time(NULL);
7993 * For Hot Standby, derive the oldestActiveXid before we fix the redo
7994 * pointer. This allows us to begin accumulating changes to assemble our
7995 * starting snapshot of locks and transactions.
7997 if (!shutdown && XLogStandbyInfoActive())
7998 checkPoint.oldestActiveXid = GetOldestActiveTransactionId();
8000 checkPoint.oldestActiveXid = InvalidTransactionId;
8003 * We must block concurrent insertions while examining insert state to
8004 * determine the checkpoint REDO pointer.
8006 WALInsertSlotAcquire(true);
8007 curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
8010 * If this isn't a shutdown or forced checkpoint, and we have not inserted
8011 * any XLOG records since the start of the last checkpoint, skip the
8012 * checkpoint. The idea here is to avoid inserting duplicate checkpoints
8013 * when the system is idle. That wastes log space, and more importantly it
8014 * exposes us to possible loss of both current and previous checkpoint
8015 * records if the machine crashes just as we're writing the update.
8016 * (Perhaps it'd make even more sense to checkpoint only when the previous
8017 * checkpoint record is in a different xlog page?)
8019 * We have to make two tests to determine that nothing has happened since
8020 * the start of the last checkpoint: current insertion point must match
8021 * the end of the last checkpoint record, and its redo pointer must point
8024 if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
8025 CHECKPOINT_FORCE)) == 0)
8027 if (curInsert == ControlFile->checkPoint +
8028 MAXALIGN(SizeOfXLogRecord + sizeof(CheckPoint)) &&
8029 ControlFile->checkPoint == ControlFile->checkPointCopy.redo)
8031 WALInsertSlotRelease();
8032 LWLockRelease(CheckpointLock);
8039 * An end-of-recovery checkpoint is created before anyone is allowed to
8040 * write WAL. To allow us to write the checkpoint record, temporarily
8041 * enable XLogInsertAllowed. (This also ensures ThisTimeLineID is
8042 * initialized, which we need here and in AdvanceXLInsertBuffer.)
8044 if (flags & CHECKPOINT_END_OF_RECOVERY)
8045 LocalSetXLogInsertAllowed();
8047 checkPoint.ThisTimeLineID = ThisTimeLineID;
8048 if (flags & CHECKPOINT_END_OF_RECOVERY)
8049 checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
8051 checkPoint.PrevTimeLineID = ThisTimeLineID;
8053 checkPoint.fullPageWrites = Insert->fullPageWrites;
8056 * Compute new REDO record ptr = location of next XLOG record.
8058 * NB: this is NOT necessarily where the checkpoint record itself will be,
8059 * since other backends may insert more XLOG records while we're off doing
8060 * the buffer flush work. Those XLOG records are logically after the
8061 * checkpoint, even though physically before it. Got that?
8063 freespace = INSERT_FREESPACE(curInsert);
8066 if (curInsert % XLogSegSize == 0)
8067 curInsert += SizeOfXLogLongPHD;
8069 curInsert += SizeOfXLogShortPHD;
8071 checkPoint.redo = curInsert;
8074 * Here we update the shared RedoRecPtr for future XLogInsert calls; this
8075 * must be done while holding the insertion slots.
8077 * Note: if we fail to complete the checkpoint, RedoRecPtr will be left
8078 * pointing past where it really needs to point. This is okay; the only
8079 * consequence is that XLogInsert might back up whole buffers that it
8080 * didn't really need to. We can't postpone advancing RedoRecPtr because
8081 * XLogInserts that happen while we are dumping buffers must assume that
8082 * their buffer changes are not included in the checkpoint.
8084 RedoRecPtr = xlogctl->Insert.RedoRecPtr = checkPoint.redo;
8087 * Now we can release the WAL insertion slots, allowing other xacts to
8088 * proceed while we are flushing disk buffers.
8090 WALInsertSlotRelease();
8092 /* Update the info_lck-protected copy of RedoRecPtr as well */
8093 SpinLockAcquire(&xlogctl->info_lck);
8094 xlogctl->RedoRecPtr = checkPoint.redo;
8095 SpinLockRelease(&xlogctl->info_lck);
8098 * If enabled, log checkpoint start. We postpone this until now so as not
8099 * to log anything if we decided to skip the checkpoint.
8101 if (log_checkpoints)
8102 LogCheckpointStart(flags, false);
8104 TRACE_POSTGRESQL_CHECKPOINT_START(flags);
8107 * In some cases there are groups of actions that must all occur on one
8108 * side or the other of a checkpoint record. Before flushing the
8109 * checkpoint record we must explicitly wait for any backend currently
8110 * performing those groups of actions.
8112 * One example is end of transaction, so we must wait for any transactions
8113 * that are currently in commit critical sections. If an xact inserted
8114 * its commit record into XLOG just before the REDO point, then a crash
8115 * restart from the REDO point would not replay that record, which means
8116 * that our flushing had better include the xact's update of pg_clog. So
8117 * we wait till he's out of his commit critical section before proceeding.
8118 * See notes in RecordTransactionCommit().
8120 * Because we've already released the insertion slots, this test is a bit
8121 * fuzzy: it is possible that we will wait for xacts we didn't really need
8122 * to wait for. But the delay should be short and it seems better to make
8123 * checkpoint take a bit longer than to hold off insertions longer than
8125 * (In fact, the whole reason we have this issue is that xact.c does
8126 * commit record XLOG insertion and clog update as two separate steps
8127 * protected by different locks, but again that seems best on grounds of
8128 * minimizing lock contention.)
8130 * A transaction that has not yet set delayChkpt when we look cannot be at
8131 * risk, since he's not inserted his commit record yet; and one that's
8132 * already cleared it is not at risk either, since he's done fixing clog
8133 * and we will correctly flush the update below. So we cannot miss any
8134 * xacts we need to wait for.
8136 vxids = GetVirtualXIDsDelayingChkpt(&nvxids);
8141 pg_usleep(10000L); /* wait for 10 msec */
8142 } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids));
8147 * Get the other info we need for the checkpoint record.
8149 LWLockAcquire(XidGenLock, LW_SHARED);
8150 checkPoint.nextXid = ShmemVariableCache->nextXid;
8151 checkPoint.oldestXid = ShmemVariableCache->oldestXid;
8152 checkPoint.oldestXidDB = ShmemVariableCache->oldestXidDB;
8153 LWLockRelease(XidGenLock);
8155 /* Increase XID epoch if we've wrapped around since last checkpoint */
8156 checkPoint.nextXidEpoch = ControlFile->checkPointCopy.nextXidEpoch;
8157 if (checkPoint.nextXid < ControlFile->checkPointCopy.nextXid)
8158 checkPoint.nextXidEpoch++;
8160 LWLockAcquire(OidGenLock, LW_SHARED);
8161 checkPoint.nextOid = ShmemVariableCache->nextOid;
8163 checkPoint.nextOid += ShmemVariableCache->oidCount;
8164 LWLockRelease(OidGenLock);
8166 MultiXactGetCheckptMulti(shutdown,
8167 &checkPoint.nextMulti,
8168 &checkPoint.nextMultiOffset,
8169 &checkPoint.oldestMulti,
8170 &checkPoint.oldestMultiDB);
8173 * Having constructed the checkpoint record, ensure all shmem disk buffers
8174 * and commit-log buffers are flushed to disk.
8176 * This I/O could fail for various reasons. If so, we will fail to
8177 * complete the checkpoint, but there is no reason to force a system
8178 * panic. Accordingly, exit critical section while doing it.
8182 CheckPointGuts(checkPoint.redo, flags);
8185 * Take a snapshot of running transactions and write this to WAL. This
8186 * allows us to reconstruct the state of running transactions during
8187 * archive recovery, if required. Skip, if this info disabled.
8189 * If we are shutting down, or Startup process is completing crash
8190 * recovery we don't need to write running xact data.
8192 if (!shutdown && XLogStandbyInfoActive())
8193 LogStandbySnapshot();
8195 START_CRIT_SECTION();
8198 * Now insert the checkpoint record into XLOG.
8200 rdata.data = (char *) (&checkPoint);
8201 rdata.len = sizeof(checkPoint);
8202 rdata.buffer = InvalidBuffer;
8205 recptr = XLogInsert(RM_XLOG_ID,
8206 shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
8207 XLOG_CHECKPOINT_ONLINE,
8213 * We mustn't write any new WAL after a shutdown checkpoint, or it will be
8214 * overwritten at next startup. No-one should even try, this just allows
8215 * sanity-checking. In the case of an end-of-recovery checkpoint, we want
8216 * to just temporarily disable writing until the system has exited
8221 if (flags & CHECKPOINT_END_OF_RECOVERY)
8222 LocalXLogInsertAllowed = -1; /* return to "check" state */
8224 LocalXLogInsertAllowed = 0; /* never again write WAL */
8228 * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
8229 * = end of actual checkpoint record.
8231 if (shutdown && checkPoint.redo != ProcLastRecPtr)
8233 (errmsg("concurrent transaction log activity while database system is shutting down")));
8236 * Select point at which we can truncate the log, which we base on the
8237 * prior checkpoint's earliest info.
8239 XLByteToSeg(ControlFile->checkPointCopy.redo, _logSegNo);
8242 * Update the control file.
8244 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8246 ControlFile->state = DB_SHUTDOWNED;
8247 ControlFile->prevCheckPoint = ControlFile->checkPoint;
8248 ControlFile->checkPoint = ProcLastRecPtr;
8249 ControlFile->checkPointCopy = checkPoint;
8250 ControlFile->time = (pg_time_t) time(NULL);
8251 /* crash recovery should always recover to the end of WAL */
8252 ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
8253 ControlFile->minRecoveryPointTLI = 0;
8256 * Persist unloggedLSN value. It's reset on crash recovery, so this goes
8257 * unused on non-shutdown checkpoints, but seems useful to store it always
8258 * for debugging purposes.
8260 SpinLockAcquire(&XLogCtl->ulsn_lck);
8261 ControlFile->unloggedLSN = XLogCtl->unloggedLSN;
8262 SpinLockRelease(&XLogCtl->ulsn_lck);
8264 UpdateControlFile();
8265 LWLockRelease(ControlFileLock);
8267 /* Update shared-memory copy of checkpoint XID/epoch */
8269 /* use volatile pointer to prevent code rearrangement */
8270 volatile XLogCtlData *xlogctl = XLogCtl;
8272 SpinLockAcquire(&xlogctl->info_lck);
8273 xlogctl->ckptXidEpoch = checkPoint.nextXidEpoch;
8274 xlogctl->ckptXid = checkPoint.nextXid;
8275 SpinLockRelease(&xlogctl->info_lck);
8279 * We are now done with critical updates; no need for system panic if we
8280 * have trouble while fooling with old log segments.
8285 * Let smgr do post-checkpoint cleanup (eg, deleting old files).
8290 * Delete old log files (those no longer needed even for previous
8291 * checkpoint or the standbys in XLOG streaming).
8295 KeepLogSeg(recptr, &_logSegNo);
8297 RemoveOldXlogFiles(_logSegNo, recptr);
8301 * Make more log segments if needed. (Do this after recycling old log
8302 * segments, since that may supply some of the needed files.)
8305 PreallocXlogFiles(recptr);
8308 * Truncate pg_subtrans if possible. We can throw away all data before
8309 * the oldest XMIN of any running transaction. No future transaction will
8310 * attempt to reference any pg_subtrans entry older than that (see Asserts
8311 * in subtrans.c). During recovery, though, we mustn't do this because
8312 * StartupSUBTRANS hasn't been called yet.
8314 if (!RecoveryInProgress())
8315 TruncateSUBTRANS(GetOldestXmin(true, false));
8317 /* Real work is done, but log and update stats before releasing lock. */
8318 LogCheckpointEnd(false);
8320 TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
8322 CheckpointStats.ckpt_segs_added,
8323 CheckpointStats.ckpt_segs_removed,
8324 CheckpointStats.ckpt_segs_recycled);
8326 LWLockRelease(CheckpointLock);
8330 * Mark the end of recovery in WAL though without running a full checkpoint.
8331 * We can expect that a restartpoint is likely to be in progress as we
8332 * do this, though we are unwilling to wait for it to complete. So be
8333 * careful to avoid taking the CheckpointLock anywhere here.
8335 * CreateRestartPoint() allows for the case where recovery may end before
8336 * the restartpoint completes so there is no concern of concurrent behaviour.
8339 CreateEndOfRecoveryRecord(void)
8341 xl_end_of_recovery xlrec;
8346 if (!RecoveryInProgress())
8347 elog(ERROR, "can only be used to end recovery");
8349 xlrec.end_time = time(NULL);
8351 WALInsertSlotAcquire(true);
8352 xlrec.ThisTimeLineID = ThisTimeLineID;
8353 xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
8354 WALInsertSlotRelease();
8356 LocalSetXLogInsertAllowed();
8358 START_CRIT_SECTION();
8360 rdata.data = (char *) &xlrec;
8361 rdata.len = sizeof(xl_end_of_recovery);
8362 rdata.buffer = InvalidBuffer;
8365 recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY, &rdata);
8370 * Update the control file so that crash recovery can follow the timeline
8371 * changes to this point.
8373 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8374 ControlFile->time = (pg_time_t) xlrec.end_time;
8375 ControlFile->minRecoveryPoint = recptr;
8376 ControlFile->minRecoveryPointTLI = ThisTimeLineID;
8377 UpdateControlFile();
8378 LWLockRelease(ControlFileLock);
8382 LocalXLogInsertAllowed = -1; /* return to "check" state */
8386 * Flush all data in shared memory to disk, and fsync
8388 * This is the common code shared between regular checkpoints and
8389 * recovery restartpoints.
8392 CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
8395 CheckPointSUBTRANS();
8396 CheckPointMultiXact();
8397 CheckPointPredicate();
8398 CheckPointRelationMap();
8399 CheckPointBuffers(flags); /* performs all required fsyncs */
8400 /* We deliberately delay 2PC checkpointing as long as possible */
8401 CheckPointTwoPhase(checkPointRedo);
8405 * Save a checkpoint for recovery restart if appropriate
8407 * This function is called each time a checkpoint record is read from XLOG.
8408 * It must determine whether the checkpoint represents a safe restartpoint or
8409 * not. If so, the checkpoint record is stashed in shared memory so that
8410 * CreateRestartPoint can consult it. (Note that the latter function is
8411 * executed by the checkpointer, while this one will be executed by the
8415 RecoveryRestartPoint(const CheckPoint *checkPoint)
8419 /* use volatile pointer to prevent code rearrangement */
8420 volatile XLogCtlData *xlogctl = XLogCtl;
8423 * Is it safe to restartpoint? We must ask each of the resource managers
8424 * whether they have any partial state information that might prevent a
8425 * correct restart from this point. If so, we skip this opportunity, but
8426 * return at the next checkpoint record for another try.
8428 for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
8430 if (RmgrTable[rmid].rm_safe_restartpoint != NULL)
8431 if (!(RmgrTable[rmid].rm_safe_restartpoint()))
8433 elog(trace_recovery(DEBUG2),
8434 "RM %d not safe to record restart point at %X/%X",
8436 (uint32) (checkPoint->redo >> 32),
8437 (uint32) checkPoint->redo);
8443 * Also refrain from creating a restartpoint if we have seen any
8444 * references to non-existent pages. Restarting recovery from the
8445 * restartpoint would not see the references, so we would lose the
8446 * cross-check that the pages belonged to a relation that was dropped
8449 if (XLogHaveInvalidPages())
8451 elog(trace_recovery(DEBUG2),
8452 "could not record restart point at %X/%X because there "
8453 "are unresolved references to invalid pages",
8454 (uint32) (checkPoint->redo >> 32),
8455 (uint32) checkPoint->redo);
8460 * Copy the checkpoint record to shared memory, so that checkpointer can
8461 * work out the next time it wants to perform a restartpoint.
8463 SpinLockAcquire(&xlogctl->info_lck);
8464 xlogctl->lastCheckPointRecPtr = ReadRecPtr;
8465 xlogctl->lastCheckPoint = *checkPoint;
8466 SpinLockRelease(&xlogctl->info_lck);
8470 * Establish a restartpoint if possible.
8472 * This is similar to CreateCheckPoint, but is used during WAL recovery
8473 * to establish a point from which recovery can roll forward without
8474 * replaying the entire recovery log.
8476 * Returns true if a new restartpoint was established. We can only establish
8477 * a restartpoint if we have replayed a safe checkpoint record since last
8481 CreateRestartPoint(int flags)
8483 XLogRecPtr lastCheckPointRecPtr;
8484 CheckPoint lastCheckPoint;
8485 XLogSegNo _logSegNo;
8488 /* use volatile pointer to prevent code rearrangement */
8489 volatile XLogCtlData *xlogctl = XLogCtl;
8492 * Acquire CheckpointLock to ensure only one restartpoint or checkpoint
8493 * happens at a time.
8495 LWLockAcquire(CheckpointLock, LW_EXCLUSIVE);
8497 /* Get a local copy of the last safe checkpoint record. */
8498 SpinLockAcquire(&xlogctl->info_lck);
8499 lastCheckPointRecPtr = xlogctl->lastCheckPointRecPtr;
8500 lastCheckPoint = xlogctl->lastCheckPoint;
8501 SpinLockRelease(&xlogctl->info_lck);
8504 * Check that we're still in recovery mode. It's ok if we exit recovery
8505 * mode after this check, the restart point is valid anyway.
8507 if (!RecoveryInProgress())
8510 (errmsg("skipping restartpoint, recovery has already ended")));
8511 LWLockRelease(CheckpointLock);
8516 * If the last checkpoint record we've replayed is already our last
8517 * restartpoint, we can't perform a new restart point. We still update
8518 * minRecoveryPoint in that case, so that if this is a shutdown restart
8519 * point, we won't start up earlier than before. That's not strictly
8520 * necessary, but when hot standby is enabled, it would be rather weird if
8521 * the database opened up for read-only connections at a point-in-time
8522 * before the last shutdown. Such time travel is still possible in case of
8523 * immediate shutdown, though.
8525 * We don't explicitly advance minRecoveryPoint when we do create a
8526 * restartpoint. It's assumed that flushing the buffers will do that as a
8529 if (XLogRecPtrIsInvalid(lastCheckPointRecPtr) ||
8530 lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
8533 (errmsg("skipping restartpoint, already performed at %X/%X",
8534 (uint32) (lastCheckPoint.redo >> 32),
8535 (uint32) lastCheckPoint.redo)));
8537 UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
8538 if (flags & CHECKPOINT_IS_SHUTDOWN)
8540 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8541 ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
8542 ControlFile->time = (pg_time_t) time(NULL);
8543 UpdateControlFile();
8544 LWLockRelease(ControlFileLock);
8546 LWLockRelease(CheckpointLock);
8551 * Update the shared RedoRecPtr so that the startup process can calculate
8552 * the number of segments replayed since last restartpoint, and request a
8553 * restartpoint if it exceeds checkpoint_segments.
8555 * Like in CreateCheckPoint(), hold off insertions to update it, although
8556 * during recovery this is just pro forma, because no WAL insertions are
8559 WALInsertSlotAcquire(true);
8560 xlogctl->Insert.RedoRecPtr = lastCheckPoint.redo;
8561 WALInsertSlotRelease();
8563 /* Also update the info_lck-protected copy */
8564 SpinLockAcquire(&xlogctl->info_lck);
8565 xlogctl->RedoRecPtr = lastCheckPoint.redo;
8566 SpinLockRelease(&xlogctl->info_lck);
8569 * Prepare to accumulate statistics.
8571 * Note: because it is possible for log_checkpoints to change while a
8572 * checkpoint proceeds, we always accumulate stats, even if
8573 * log_checkpoints is currently off.
8575 MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
8576 CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
8578 if (log_checkpoints)
8579 LogCheckpointStart(flags, true);
8581 CheckPointGuts(lastCheckPoint.redo, flags);
8584 * Select point at which we can truncate the xlog, which we base on the
8585 * prior checkpoint's earliest info.
8587 XLByteToSeg(ControlFile->checkPointCopy.redo, _logSegNo);
8590 * Update pg_control, using current time. Check that it still shows
8591 * IN_ARCHIVE_RECOVERY state and an older checkpoint, else do nothing;
8592 * this is a quick hack to make sure nothing really bad happens if somehow
8593 * we get here after the end-of-recovery checkpoint.
8595 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8596 if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY &&
8597 ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
8599 ControlFile->prevCheckPoint = ControlFile->checkPoint;
8600 ControlFile->checkPoint = lastCheckPointRecPtr;
8601 ControlFile->checkPointCopy = lastCheckPoint;
8602 ControlFile->time = (pg_time_t) time(NULL);
8603 if (flags & CHECKPOINT_IS_SHUTDOWN)
8604 ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
8605 UpdateControlFile();
8607 LWLockRelease(ControlFileLock);
8610 * Delete old log files (those no longer needed even for previous
8611 * checkpoint/restartpoint) to prevent the disk holding the xlog from
8616 XLogRecPtr receivePtr;
8617 XLogRecPtr replayPtr;
8618 TimeLineID replayTLI;
8622 * Get the current end of xlog replayed or received, whichever is
8625 receivePtr = GetWalRcvWriteRecPtr(NULL, NULL);
8626 replayPtr = GetXLogReplayRecPtr(&replayTLI);
8627 endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
8629 KeepLogSeg(endptr, &_logSegNo);
8633 * Try to recycle segments on a useful timeline. If we've been promoted
8634 * since the beginning of this restartpoint, use the new timeline
8635 * chosen at end of recovery (RecoveryInProgress() sets ThisTimeLineID
8636 * in that case). If we're still in recovery, use the timeline we're
8637 * currently replaying.
8639 * There is no guarantee that the WAL segments will be useful on the
8640 * current timeline; if recovery proceeds to a new timeline right
8641 * after this, the pre-allocated WAL segments on this timeline will
8642 * not be used, and will go wasted until recycled on the next
8643 * restartpoint. We'll live with that.
8645 if (RecoveryInProgress())
8646 ThisTimeLineID = replayTLI;
8648 RemoveOldXlogFiles(_logSegNo, endptr);
8651 * Make more log segments if needed. (Do this after recycling old log
8652 * segments, since that may supply some of the needed files.)
8654 PreallocXlogFiles(endptr);
8657 * ThisTimeLineID is normally not set when we're still in recovery.
8658 * However, recycling/preallocating segments above needed
8659 * ThisTimeLineID to determine which timeline to install the segments
8660 * on. Reset it now, to restore the normal state of affairs for
8661 * debugging purposes.
8663 if (RecoveryInProgress())
8668 * Truncate pg_subtrans if possible. We can throw away all data before
8669 * the oldest XMIN of any running transaction. No future transaction will
8670 * attempt to reference any pg_subtrans entry older than that (see Asserts
8671 * in subtrans.c). When hot standby is disabled, though, we mustn't do
8672 * this because StartupSUBTRANS hasn't been called yet.
8674 if (EnableHotStandby)
8675 TruncateSUBTRANS(GetOldestXmin(true, false));
8677 /* Real work is done, but log and update before releasing lock. */
8678 LogCheckpointEnd(true);
8680 xtime = GetLatestXTime();
8681 ereport((log_checkpoints ? LOG : DEBUG2),
8682 (errmsg("recovery restart point at %X/%X",
8683 (uint32) (lastCheckPoint.redo >> 32), (uint32) lastCheckPoint.redo),
8684 xtime ? errdetail("last completed transaction was at log time %s",
8685 timestamptz_to_str(xtime)) : 0));
8687 LWLockRelease(CheckpointLock);
8690 * Finally, execute archive_cleanup_command, if any.
8692 if (XLogCtl->archiveCleanupCommand[0])
8693 ExecuteRecoveryCommand(XLogCtl->archiveCleanupCommand,
8694 "archive_cleanup_command",
8701 * Retreat *logSegNo to the last segment that we need to retain because of
8702 * wal_keep_segments. This is calculated by subtracting wal_keep_segments
8703 * from the given xlog location, recptr.
8706 KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
8710 if (wal_keep_segments == 0)
8713 XLByteToSeg(recptr, segno);
8715 /* avoid underflow, don't go below 1 */
8716 if (segno <= wal_keep_segments)
8719 segno = segno - wal_keep_segments;
8721 /* don't delete WAL segments newer than the calculated segment */
8722 if (segno < *logSegNo)
8727 * Write a NEXTOID log record
8730 XLogPutNextOid(Oid nextOid)
8734 rdata.data = (char *) (&nextOid);
8735 rdata.len = sizeof(Oid);
8736 rdata.buffer = InvalidBuffer;
8738 (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID, &rdata);
8741 * We need not flush the NEXTOID record immediately, because any of the
8742 * just-allocated OIDs could only reach disk as part of a tuple insert or
8743 * update that would have its own XLOG record that must follow the NEXTOID
8744 * record. Therefore, the standard buffer LSN interlock applied to those
8745 * records will ensure no such OID reaches disk before the NEXTOID record
8748 * Note, however, that the above statement only covers state "within" the
8749 * database. When we use a generated OID as a file or directory name, we
8750 * are in a sense violating the basic WAL rule, because that filesystem
8751 * change may reach disk before the NEXTOID WAL record does. The impact
8752 * of this is that if a database crash occurs immediately afterward, we
8753 * might after restart re-generate the same OID and find that it conflicts
8754 * with the leftover file or directory. But since for safety's sake we
8755 * always loop until finding a nonconflicting filename, this poses no real
8756 * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8761 * Write an XLOG SWITCH record.
8763 * Here we just blindly issue an XLogInsert request for the record.
8764 * All the magic happens inside XLogInsert.
8766 * The return value is either the end+1 address of the switch record,
8767 * or the end+1 address of the prior segment if we did not need to
8768 * write a switch record because we are already at segment start.
8771 RequestXLogSwitch(void)
8776 /* XLOG SWITCH, alone among xlog record types, has no data */
8777 rdata.buffer = InvalidBuffer;
8782 RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH, &rdata);
8788 * Write a RESTORE POINT record
8791 XLogRestorePoint(const char *rpName)
8795 xl_restore_point xlrec;
8797 xlrec.rp_time = GetCurrentTimestamp();
8798 strncpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8800 rdata.buffer = InvalidBuffer;
8801 rdata.data = (char *) &xlrec;
8802 rdata.len = sizeof(xl_restore_point);
8805 RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT, &rdata);
8808 (errmsg("restore point \"%s\" created at %X/%X",
8809 rpName, (uint32) (RecPtr >> 32), (uint32) RecPtr)));
8815 * Write a backup block if needed when we are setting a hint. Note that
8816 * this may be called for a variety of page types, not just heaps.
8818 * Callable while holding just share lock on the buffer content.
8820 * We can't use the plain backup block mechanism since that relies on the
8821 * Buffer being exclusively locked. Since some modifications (setting LSN, hint
8822 * bits) are allowed in a sharelocked buffer that can lead to wal checksum
8823 * failures. So instead we copy the page and insert the copied data as normal
8826 * We only need to do something if page has not yet been full page written in
8827 * this checkpoint round. The LSN of the inserted wal record is returned if we
8828 * had to write, InvalidXLogRecPtr otherwise.
8830 * It is possible that multiple concurrent backends could attempt to write WAL
8831 * records. In that case, multiple copies of the same block would be recorded
8832 * in separate WAL records by different backends, though that is still OK from
8833 * a correctness perspective.
8836 XLogSaveBufferForHint(Buffer buffer, bool buffer_std)
8838 XLogRecPtr recptr = InvalidXLogRecPtr;
8840 XLogRecData rdata[2];
8844 * Ensure no checkpoint can change our view of RedoRecPtr.
8846 Assert(MyPgXact->delayChkpt);
8849 * Update RedoRecPtr so XLogCheckBuffer can make the right decision
8854 * Setup phony rdata element for use within XLogCheckBuffer only. We reuse
8855 * and reset rdata for any actual WAL record insert.
8857 rdata[0].buffer = buffer;
8858 rdata[0].buffer_std = buffer_std;
8861 * Check buffer while not holding an exclusive lock.
8863 if (XLogCheckBuffer(rdata, false, &lsn, &bkpb))
8865 char copied_buffer[BLCKSZ];
8866 char *origdata = (char *) BufferGetBlock(buffer);
8869 * Copy buffer so we don't have to worry about concurrent hint bit or
8870 * lsn updates. We assume pd_lower/upper cannot be changed without an
8871 * exclusive lock, so the contents bkp are not racy.
8873 * With buffer_std set to false, XLogCheckBuffer() sets hole_length and
8874 * hole_offset to 0; so the following code is safe for either case.
8876 memcpy(copied_buffer, origdata, bkpb.hole_offset);
8877 memcpy(copied_buffer + bkpb.hole_offset,
8878 origdata + bkpb.hole_offset + bkpb.hole_length,
8879 BLCKSZ - bkpb.hole_offset - bkpb.hole_length);
8882 * Header for backup block.
8884 rdata[0].data = (char *) &bkpb;
8885 rdata[0].len = sizeof(BkpBlock);
8886 rdata[0].buffer = InvalidBuffer;
8887 rdata[0].next = &(rdata[1]);
8890 * Save copy of the buffer.
8892 rdata[1].data = copied_buffer;
8893 rdata[1].len = BLCKSZ - bkpb.hole_length;
8894 rdata[1].buffer = InvalidBuffer;
8895 rdata[1].next = NULL;
8897 recptr = XLogInsert(RM_XLOG_ID, XLOG_FPI, rdata);
8904 * Check if any of the GUC parameters that are critical for hot standby
8905 * have changed, and update the value in pg_control file if necessary.
8908 XLogReportParameters(void)
8910 if (wal_level != ControlFile->wal_level ||
8911 MaxConnections != ControlFile->MaxConnections ||
8912 max_worker_processes != ControlFile->max_worker_processes ||
8913 max_prepared_xacts != ControlFile->max_prepared_xacts ||
8914 max_locks_per_xact != ControlFile->max_locks_per_xact)
8917 * The change in number of backend slots doesn't need to be WAL-logged
8918 * if archiving is not enabled, as you can't start archive recovery
8919 * with wal_level=minimal anyway. We don't really care about the
8920 * values in pg_control either if wal_level=minimal, but seems better
8921 * to keep them up-to-date to avoid confusion.
8923 if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8926 xl_parameter_change xlrec;
8928 xlrec.MaxConnections = MaxConnections;
8929 xlrec.max_worker_processes = max_worker_processes;
8930 xlrec.max_prepared_xacts = max_prepared_xacts;
8931 xlrec.max_locks_per_xact = max_locks_per_xact;
8932 xlrec.wal_level = wal_level;
8934 rdata.buffer = InvalidBuffer;
8935 rdata.data = (char *) &xlrec;
8936 rdata.len = sizeof(xlrec);
8939 XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE, &rdata);
8942 ControlFile->MaxConnections = MaxConnections;
8943 ControlFile->max_worker_processes = max_worker_processes;
8944 ControlFile->max_prepared_xacts = max_prepared_xacts;
8945 ControlFile->max_locks_per_xact = max_locks_per_xact;
8946 ControlFile->wal_level = wal_level;
8947 UpdateControlFile();
8952 * Update full_page_writes in shared memory, and write an
8953 * XLOG_FPW_CHANGE record if necessary.
8955 * Note: this function assumes there is no other process running
8956 * concurrently that could update it.
8959 UpdateFullPageWrites(void)
8961 XLogCtlInsert *Insert = &XLogCtl->Insert;
8964 * Do nothing if full_page_writes has not been changed.
8966 * It's safe to check the shared full_page_writes without the lock,
8967 * because we assume that there is no concurrently running process which
8970 if (fullPageWrites == Insert->fullPageWrites)
8973 START_CRIT_SECTION();
8976 * It's always safe to take full page images, even when not strictly
8977 * required, but not the other round. So if we're setting full_page_writes
8978 * to true, first set it true and then write the WAL record. If we're
8979 * setting it to false, first write the WAL record and then set the global
8984 WALInsertSlotAcquire(true);
8985 Insert->fullPageWrites = true;
8986 WALInsertSlotRelease();
8990 * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8991 * full_page_writes during archive recovery, if required.
8993 if (XLogStandbyInfoActive() && !RecoveryInProgress())
8997 rdata.data = (char *) (&fullPageWrites);
8998 rdata.len = sizeof(bool);
8999 rdata.buffer = InvalidBuffer;
9002 XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE, &rdata);
9005 if (!fullPageWrites)
9007 WALInsertSlotAcquire(true);
9008 Insert->fullPageWrites = false;
9009 WALInsertSlotRelease();
9015 * Check that it's OK to switch to new timeline during recovery.
9017 * 'lsn' is the address of the shutdown checkpoint record we're about to
9018 * replay. (Currently, timeline can only change at a shutdown checkpoint).
9021 checkTimeLineSwitch(XLogRecPtr lsn, TimeLineID newTLI, TimeLineID prevTLI)
9023 /* Check that the record agrees on what the current (old) timeline is */
9024 if (prevTLI != ThisTimeLineID)
9026 (errmsg("unexpected previous timeline ID %u (current timeline ID %u) in checkpoint record",
9027 prevTLI, ThisTimeLineID)));
9030 * The new timeline better be in the list of timelines we expect to see,
9031 * according to the timeline history. It should also not decrease.
9033 if (newTLI < ThisTimeLineID || !tliInHistory(newTLI, expectedTLEs))
9035 (errmsg("unexpected timeline ID %u (after %u) in checkpoint record",
9036 newTLI, ThisTimeLineID)));
9039 * If we have not yet reached min recovery point, and we're about to
9040 * switch to a timeline greater than the timeline of the min recovery
9041 * point: trouble. After switching to the new timeline, we could not
9042 * possibly visit the min recovery point on the correct timeline anymore.
9043 * This can happen if there is a newer timeline in the archive that
9044 * branched before the timeline the min recovery point is on, and you
9045 * attempt to do PITR to the new timeline.
9047 if (!XLogRecPtrIsInvalid(minRecoveryPoint) &&
9048 lsn < minRecoveryPoint &&
9049 newTLI > minRecoveryPointTLI)
9051 (errmsg("unexpected timeline ID %u in checkpoint record, before reaching minimum recovery point %X/%X on timeline %u",
9053 (uint32) (minRecoveryPoint >> 32),
9054 (uint32) minRecoveryPoint,
9055 minRecoveryPointTLI)));
9061 * XLOG resource manager's routines
9063 * Definitions of info values are in include/catalog/pg_control.h, though
9064 * not all record types are related to control file updates.
9067 xlog_redo(XLogRecPtr lsn, XLogRecord *record)
9069 uint8 info = record->xl_info & ~XLR_INFO_MASK;
9071 /* Backup blocks are not used by XLOG rmgr */
9072 Assert(!(record->xl_info & XLR_BKP_BLOCK_MASK));
9074 if (info == XLOG_NEXTOID)
9079 * We used to try to take the maximum of ShmemVariableCache->nextOid
9080 * and the recorded nextOid, but that fails if the OID counter wraps
9081 * around. Since no OID allocation should be happening during replay
9082 * anyway, better to just believe the record exactly. We still take
9083 * OidGenLock while setting the variable, just in case.
9085 memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
9086 LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
9087 ShmemVariableCache->nextOid = nextOid;
9088 ShmemVariableCache->oidCount = 0;
9089 LWLockRelease(OidGenLock);
9091 else if (info == XLOG_CHECKPOINT_SHUTDOWN)
9093 CheckPoint checkPoint;
9095 memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
9096 /* In a SHUTDOWN checkpoint, believe the counters exactly */
9097 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
9098 ShmemVariableCache->nextXid = checkPoint.nextXid;
9099 LWLockRelease(XidGenLock);
9100 LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
9101 ShmemVariableCache->nextOid = checkPoint.nextOid;
9102 ShmemVariableCache->oidCount = 0;
9103 LWLockRelease(OidGenLock);
9104 MultiXactSetNextMXact(checkPoint.nextMulti,
9105 checkPoint.nextMultiOffset);
9106 SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
9107 SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
9110 * If we see a shutdown checkpoint while waiting for an end-of-backup
9111 * record, the backup was canceled and the end-of-backup record will
9114 if (ArchiveRecoveryRequested &&
9115 !XLogRecPtrIsInvalid(ControlFile->backupStartPoint) &&
9116 XLogRecPtrIsInvalid(ControlFile->backupEndPoint))
9118 (errmsg("online backup was canceled, recovery cannot continue")));
9121 * If we see a shutdown checkpoint, we know that nothing was running
9122 * on the master at this point. So fake-up an empty running-xacts
9123 * record and use that here and now. Recover additional standby state
9124 * for prepared transactions.
9126 if (standbyState >= STANDBY_INITIALIZED)
9128 TransactionId *xids;
9130 TransactionId oldestActiveXID;
9131 TransactionId latestCompletedXid;
9132 RunningTransactionsData running;
9134 oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
9137 * Construct a RunningTransactions snapshot representing a shut
9138 * down server, with only prepared transactions still alive. We're
9139 * never overflowed at this point because all subxids are listed
9140 * with their parent prepared transactions.
9142 running.xcnt = nxids;
9143 running.subxcnt = 0;
9144 running.subxid_overflow = false;
9145 running.nextXid = checkPoint.nextXid;
9146 running.oldestRunningXid = oldestActiveXID;
9147 latestCompletedXid = checkPoint.nextXid;
9148 TransactionIdRetreat(latestCompletedXid);
9149 Assert(TransactionIdIsNormal(latestCompletedXid));
9150 running.latestCompletedXid = latestCompletedXid;
9151 running.xids = xids;
9153 ProcArrayApplyRecoveryInfo(&running);
9155 StandbyRecoverPreparedTransactions(true);
9158 /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
9159 ControlFile->checkPointCopy.nextXidEpoch = checkPoint.nextXidEpoch;
9160 ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
9162 /* Update shared-memory copy of checkpoint XID/epoch */
9164 /* use volatile pointer to prevent code rearrangement */
9165 volatile XLogCtlData *xlogctl = XLogCtl;
9167 SpinLockAcquire(&xlogctl->info_lck);
9168 xlogctl->ckptXidEpoch = checkPoint.nextXidEpoch;
9169 xlogctl->ckptXid = checkPoint.nextXid;
9170 SpinLockRelease(&xlogctl->info_lck);
9174 * We should've already switched to the new TLI before replaying this
9177 if (checkPoint.ThisTimeLineID != ThisTimeLineID)
9179 (errmsg("unexpected timeline ID %u (should be %u) in checkpoint record",
9180 checkPoint.ThisTimeLineID, ThisTimeLineID)));
9182 RecoveryRestartPoint(&checkPoint);
9184 else if (info == XLOG_CHECKPOINT_ONLINE)
9186 CheckPoint checkPoint;
9188 memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
9189 /* In an ONLINE checkpoint, treat the XID counter as a minimum */
9190 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
9191 if (TransactionIdPrecedes(ShmemVariableCache->nextXid,
9192 checkPoint.nextXid))
9193 ShmemVariableCache->nextXid = checkPoint.nextXid;
9194 LWLockRelease(XidGenLock);
9195 /* ... but still treat OID counter as exact */
9196 LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
9197 ShmemVariableCache->nextOid = checkPoint.nextOid;
9198 ShmemVariableCache->oidCount = 0;
9199 LWLockRelease(OidGenLock);
9200 MultiXactAdvanceNextMXact(checkPoint.nextMulti,
9201 checkPoint.nextMultiOffset);
9202 if (TransactionIdPrecedes(ShmemVariableCache->oldestXid,
9203 checkPoint.oldestXid))
9204 SetTransactionIdLimit(checkPoint.oldestXid,
9205 checkPoint.oldestXidDB);
9206 MultiXactAdvanceOldest(checkPoint.oldestMulti,
9207 checkPoint.oldestMultiDB);
9209 /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
9210 ControlFile->checkPointCopy.nextXidEpoch = checkPoint.nextXidEpoch;
9211 ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
9213 /* Update shared-memory copy of checkpoint XID/epoch */
9215 /* use volatile pointer to prevent code rearrangement */
9216 volatile XLogCtlData *xlogctl = XLogCtl;
9218 SpinLockAcquire(&xlogctl->info_lck);
9219 xlogctl->ckptXidEpoch = checkPoint.nextXidEpoch;
9220 xlogctl->ckptXid = checkPoint.nextXid;
9221 SpinLockRelease(&xlogctl->info_lck);
9224 /* TLI should not change in an on-line checkpoint */
9225 if (checkPoint.ThisTimeLineID != ThisTimeLineID)
9227 (errmsg("unexpected timeline ID %u (should be %u) in checkpoint record",
9228 checkPoint.ThisTimeLineID, ThisTimeLineID)));
9230 RecoveryRestartPoint(&checkPoint);
9232 else if (info == XLOG_END_OF_RECOVERY)
9234 xl_end_of_recovery xlrec;
9236 memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
9239 * For Hot Standby, we could treat this like a Shutdown Checkpoint,
9240 * but this case is rarer and harder to test, so the benefit doesn't
9241 * outweigh the potential extra cost of maintenance.
9245 * We should've already switched to the new TLI before replaying this
9248 if (xlrec.ThisTimeLineID != ThisTimeLineID)
9250 (errmsg("unexpected timeline ID %u (should be %u) in checkpoint record",
9251 xlrec.ThisTimeLineID, ThisTimeLineID)));
9253 else if (info == XLOG_NOOP)
9255 /* nothing to do here */
9257 else if (info == XLOG_SWITCH)
9259 /* nothing to do here */
9261 else if (info == XLOG_RESTORE_POINT)
9263 /* nothing to do here */
9265 else if (info == XLOG_FPI)
9271 * Full-page image (FPI) records contain a backup block stored "inline"
9272 * in the normal data since the locking when writing hint records isn't
9273 * sufficient to use the normal backup block mechanism, which assumes
9274 * exclusive lock on the buffer supplied.
9276 * Since the only change in these backup block are hint bits, there
9277 * are no recovery conflicts generated.
9279 * This also means there is no corresponding API call for this, so an
9280 * smgr implementation has no need to implement anything. Which means
9281 * nothing is needed in md.c etc
9283 data = XLogRecGetData(record);
9284 memcpy(&bkpb, data, sizeof(BkpBlock));
9285 data += sizeof(BkpBlock);
9287 RestoreBackupBlockContents(lsn, bkpb, data, false, false);
9289 else if (info == XLOG_BACKUP_END)
9291 XLogRecPtr startpoint;
9293 memcpy(&startpoint, XLogRecGetData(record), sizeof(startpoint));
9295 if (ControlFile->backupStartPoint == startpoint)
9298 * We have reached the end of base backup, the point where
9299 * pg_stop_backup() was done. The data on disk is now consistent.
9300 * Reset backupStartPoint, and update minRecoveryPoint to make
9301 * sure we don't allow starting up at an earlier point even if
9302 * recovery is stopped and restarted soon after this.
9304 elog(DEBUG1, "end of backup reached");
9306 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9308 if (ControlFile->minRecoveryPoint < lsn)
9310 ControlFile->minRecoveryPoint = lsn;
9311 ControlFile->minRecoveryPointTLI = ThisTimeLineID;
9313 ControlFile->backupStartPoint = InvalidXLogRecPtr;
9314 ControlFile->backupEndRequired = false;
9315 UpdateControlFile();
9317 LWLockRelease(ControlFileLock);
9320 else if (info == XLOG_PARAMETER_CHANGE)
9322 xl_parameter_change xlrec;
9324 /* Update our copy of the parameters in pg_control */
9325 memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
9327 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9328 ControlFile->MaxConnections = xlrec.MaxConnections;
9329 ControlFile->max_worker_processes = xlrec.max_worker_processes;
9330 ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
9331 ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
9332 ControlFile->wal_level = xlrec.wal_level;
9335 * Update minRecoveryPoint to ensure that if recovery is aborted, we
9336 * recover back up to this point before allowing hot standby again.
9337 * This is particularly important if wal_level was set to 'archive'
9338 * before, and is now 'hot_standby', to ensure you don't run queries
9339 * against the WAL preceding the wal_level change. Same applies to
9340 * decreasing max_* settings.
9342 minRecoveryPoint = ControlFile->minRecoveryPoint;
9343 minRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
9344 if (minRecoveryPoint != 0 && minRecoveryPoint < lsn)
9346 ControlFile->minRecoveryPoint = lsn;
9347 ControlFile->minRecoveryPointTLI = ThisTimeLineID;
9350 UpdateControlFile();
9351 LWLockRelease(ControlFileLock);
9353 /* Check to see if any changes to max_connections give problems */
9354 CheckRequiredParameterValues();
9356 else if (info == XLOG_FPW_CHANGE)
9358 /* use volatile pointer to prevent code rearrangement */
9359 volatile XLogCtlData *xlogctl = XLogCtl;
9362 memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
9365 * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
9366 * do_pg_start_backup() and do_pg_stop_backup() can check whether
9367 * full_page_writes has been disabled during online backup.
9371 SpinLockAcquire(&xlogctl->info_lck);
9372 if (xlogctl->lastFpwDisableRecPtr < ReadRecPtr)
9373 xlogctl->lastFpwDisableRecPtr = ReadRecPtr;
9374 SpinLockRelease(&xlogctl->info_lck);
9377 /* Keep track of full_page_writes */
9378 lastFullPageWrites = fpw;
9385 xlog_outrec(StringInfo buf, XLogRecord *record)
9389 appendStringInfo(buf, "prev %X/%X; xid %u",
9390 (uint32) (record->xl_prev >> 32),
9391 (uint32) record->xl_prev,
9394 appendStringInfo(buf, "; len %u",
9397 for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
9399 if (record->xl_info & XLR_BKP_BLOCK(i))
9400 appendStringInfo(buf, "; bkpb%d", i);
9403 appendStringInfo(buf, ": %s", RmgrTable[record->xl_rmid].rm_name);
9405 #endif /* WAL_DEBUG */
9409 * Return the (possible) sync flag used for opening a file, depending on the
9410 * value of the GUC wal_sync_method.
9413 get_sync_bit(int method)
9415 int o_direct_flag = 0;
9417 /* If fsync is disabled, never open in sync mode */
9422 * Optimize writes by bypassing kernel cache with O_DIRECT when using
9423 * O_SYNC/O_FSYNC and O_DSYNC. But only if archiving and streaming are
9424 * disabled, otherwise the archive command or walsender process will read
9425 * the WAL soon after writing it, which is guaranteed to cause a physical
9426 * read if we bypassed the kernel cache. We also skip the
9427 * posix_fadvise(POSIX_FADV_DONTNEED) call in XLogFileClose() for the same
9430 * Never use O_DIRECT in walreceiver process for similar reasons; the WAL
9431 * written by walreceiver is normally read by the startup process soon
9432 * after its written. Also, walreceiver performs unaligned writes, which
9433 * don't work with O_DIRECT, so it is required for correctness too.
9435 if (!XLogIsNeeded() && !AmWalReceiverProcess())
9436 o_direct_flag = PG_O_DIRECT;
9441 * enum values for all sync options are defined even if they are
9442 * not supported on the current platform. But if not, they are
9443 * not included in the enum option array, and therefore will never
9446 case SYNC_METHOD_FSYNC:
9447 case SYNC_METHOD_FSYNC_WRITETHROUGH:
9448 case SYNC_METHOD_FDATASYNC:
9450 #ifdef OPEN_SYNC_FLAG
9451 case SYNC_METHOD_OPEN:
9452 return OPEN_SYNC_FLAG | o_direct_flag;
9454 #ifdef OPEN_DATASYNC_FLAG
9455 case SYNC_METHOD_OPEN_DSYNC:
9456 return OPEN_DATASYNC_FLAG | o_direct_flag;
9459 /* can't happen (unless we are out of sync with option array) */
9460 elog(ERROR, "unrecognized wal_sync_method: %d", method);
9461 return 0; /* silence warning */
9469 assign_xlog_sync_method(int new_sync_method, void *extra)
9471 if (sync_method != new_sync_method)
9474 * To ensure that no blocks escape unsynced, force an fsync on the
9475 * currently open log segment (if any). Also, if the open flag is
9476 * changing, close the log file so it will be reopened (with new flag
9479 if (openLogFile >= 0)
9481 if (pg_fsync(openLogFile) != 0)
9483 (errcode_for_file_access(),
9484 errmsg("could not fsync log segment %s: %m",
9485 XLogFileNameP(ThisTimeLineID, openLogSegNo))));
9486 if (get_sync_bit(sync_method) != get_sync_bit(new_sync_method))
9494 * Issue appropriate kind of fsync (if any) for an XLOG output file.
9496 * 'fd' is a file descriptor for the XLOG file to be fsync'd.
9497 * 'log' and 'seg' are for error reporting purposes.
9500 issue_xlog_fsync(int fd, XLogSegNo segno)
9502 switch (sync_method)
9504 case SYNC_METHOD_FSYNC:
9505 if (pg_fsync_no_writethrough(fd) != 0)
9507 (errcode_for_file_access(),
9508 errmsg("could not fsync log file %s: %m",
9509 XLogFileNameP(ThisTimeLineID, segno))));
9511 #ifdef HAVE_FSYNC_WRITETHROUGH
9512 case SYNC_METHOD_FSYNC_WRITETHROUGH:
9513 if (pg_fsync_writethrough(fd) != 0)
9515 (errcode_for_file_access(),
9516 errmsg("could not fsync write-through log file %s: %m",
9517 XLogFileNameP(ThisTimeLineID, segno))));
9520 #ifdef HAVE_FDATASYNC
9521 case SYNC_METHOD_FDATASYNC:
9522 if (pg_fdatasync(fd) != 0)
9524 (errcode_for_file_access(),
9525 errmsg("could not fdatasync log file %s: %m",
9526 XLogFileNameP(ThisTimeLineID, segno))));
9529 case SYNC_METHOD_OPEN:
9530 case SYNC_METHOD_OPEN_DSYNC:
9531 /* write synced it already */
9534 elog(PANIC, "unrecognized wal_sync_method: %d", sync_method);
9540 * Return the filename of given log segment, as a palloc'd string.
9543 XLogFileNameP(TimeLineID tli, XLogSegNo segno)
9545 char *result = palloc(MAXFNAMELEN);
9547 XLogFileName(result, tli, segno);
9552 * do_pg_start_backup is the workhorse of the user-visible pg_start_backup()
9553 * function. It creates the necessary starting checkpoint and constructs the
9554 * backup label file.
9556 * There are two kind of backups: exclusive and non-exclusive. An exclusive
9557 * backup is started with pg_start_backup(), and there can be only one active
9558 * at a time. The backup label file of an exclusive backup is written to
9559 * $PGDATA/backup_label, and it is removed by pg_stop_backup().
9561 * A non-exclusive backup is used for the streaming base backups (see
9562 * src/backend/replication/basebackup.c). The difference to exclusive backups
9563 * is that the backup label file is not written to disk. Instead, its would-be
9564 * contents are returned in *labelfile, and the caller is responsible for
9565 * including it in the backup archive as 'backup_label'. There can be many
9566 * non-exclusive backups active at the same time, and they don't conflict
9567 * with an exclusive backup either.
9569 * Returns the minimum WAL position that must be present to restore from this
9570 * backup, and the corresponding timeline ID in *starttli_p.
9572 * Every successfully started non-exclusive backup must be stopped by calling
9573 * do_pg_stop_backup() or do_pg_abort_backup().
9576 do_pg_start_backup(const char *backupidstr, bool fast, TimeLineID *starttli_p,
9579 bool exclusive = (labelfile == NULL);
9580 bool backup_started_in_recovery = false;
9581 XLogRecPtr checkpointloc;
9582 XLogRecPtr startpoint;
9583 TimeLineID starttli;
9584 pg_time_t stamp_time;
9586 char xlogfilename[MAXFNAMELEN];
9587 XLogSegNo _logSegNo;
9588 struct stat stat_buf;
9590 StringInfoData labelfbuf;
9592 backup_started_in_recovery = RecoveryInProgress();
9594 if (!superuser() && !has_rolreplication(GetUserId()))
9596 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
9597 errmsg("must be superuser or replication role to run a backup")));
9600 * Currently only non-exclusive backup can be taken during recovery.
9602 if (backup_started_in_recovery && exclusive)
9604 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9605 errmsg("recovery is in progress"),
9606 errhint("WAL control functions cannot be executed during recovery.")));
9609 * During recovery, we don't need to check WAL level. Because, if WAL
9610 * level is not sufficient, it's impossible to get here during recovery.
9612 if (!backup_started_in_recovery && !XLogIsNeeded())
9614 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9615 errmsg("WAL level not sufficient for making an online backup"),
9616 errhint("wal_level must be set to \"archive\" or \"hot_standby\" at server start.")));
9618 if (strlen(backupidstr) > MAXPGPATH)
9620 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
9621 errmsg("backup label too long (max %d bytes)",
9625 * Mark backup active in shared memory. We must do full-page WAL writes
9626 * during an on-line backup even if not doing so at other times, because
9627 * it's quite possible for the backup dump to obtain a "torn" (partially
9628 * written) copy of a database page if it reads the page concurrently with
9629 * our write to the same page. This can be fixed as long as the first
9630 * write to the page in the WAL sequence is a full-page write. Hence, we
9631 * turn on forcePageWrites and then force a CHECKPOINT, to ensure there
9632 * are no dirty pages in shared memory that might get dumped while the
9633 * backup is in progress without having a corresponding WAL record. (Once
9634 * the backup is complete, we need not force full-page writes anymore,
9635 * since we expect that any pages not modified during the backup interval
9636 * must have been correctly captured by the backup.)
9638 * Note that forcePageWrites has no effect during an online backup from
9641 * We must hold all the insertion slots to change the value of
9642 * forcePageWrites, to ensure adequate interlocking against XLogInsert().
9644 WALInsertSlotAcquire(true);
9647 if (XLogCtl->Insert.exclusiveBackup)
9649 WALInsertSlotRelease();
9651 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9652 errmsg("a backup is already in progress"),
9653 errhint("Run pg_stop_backup() and try again.")));
9655 XLogCtl->Insert.exclusiveBackup = true;
9658 XLogCtl->Insert.nonExclusiveBackups++;
9659 XLogCtl->Insert.forcePageWrites = true;
9660 WALInsertSlotRelease();
9662 /* Ensure we release forcePageWrites if fail below */
9663 PG_ENSURE_ERROR_CLEANUP(pg_start_backup_callback, (Datum) BoolGetDatum(exclusive));
9665 bool gotUniqueStartpoint = false;
9668 * Force an XLOG file switch before the checkpoint, to ensure that the
9669 * WAL segment the checkpoint is written to doesn't contain pages with
9670 * old timeline IDs. That would otherwise happen if you called
9671 * pg_start_backup() right after restoring from a PITR archive: the
9672 * first WAL segment containing the startup checkpoint has pages in
9673 * the beginning with the old timeline ID. That can cause trouble at
9674 * recovery: we won't have a history file covering the old timeline if
9675 * pg_xlog directory was not included in the base backup and the WAL
9676 * archive was cleared too before starting the backup.
9678 * This also ensures that we have emitted a WAL page header that has
9679 * XLP_BKP_REMOVABLE off before we emit the checkpoint record.
9680 * Therefore, if a WAL archiver (such as pglesslog) is trying to
9681 * compress out removable backup blocks, it won't remove any that
9682 * occur after this point.
9684 * During recovery, we skip forcing XLOG file switch, which means that
9685 * the backup taken during recovery is not available for the special
9686 * recovery case described above.
9688 if (!backup_started_in_recovery)
9689 RequestXLogSwitch();
9696 * Force a CHECKPOINT. Aside from being necessary to prevent torn
9697 * page problems, this guarantees that two successive backup runs
9698 * will have different checkpoint positions and hence different
9699 * history file names, even if nothing happened in between.
9701 * During recovery, establish a restartpoint if possible. We use
9702 * the last restartpoint as the backup starting checkpoint. This
9703 * means that two successive backup runs can have same checkpoint
9706 * Since the fact that we are executing do_pg_start_backup()
9707 * during recovery means that checkpointer is running, we can use
9708 * RequestCheckpoint() to establish a restartpoint.
9710 * We use CHECKPOINT_IMMEDIATE only if requested by user (via
9711 * passing fast = true). Otherwise this can take awhile.
9713 RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
9714 (fast ? CHECKPOINT_IMMEDIATE : 0));
9717 * Now we need to fetch the checkpoint record location, and also
9718 * its REDO pointer. The oldest point in WAL that would be needed
9719 * to restore starting from the checkpoint is precisely the REDO
9722 LWLockAcquire(ControlFileLock, LW_SHARED);
9723 checkpointloc = ControlFile->checkPoint;
9724 startpoint = ControlFile->checkPointCopy.redo;
9725 starttli = ControlFile->checkPointCopy.ThisTimeLineID;
9726 checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
9727 LWLockRelease(ControlFileLock);
9729 if (backup_started_in_recovery)
9731 /* use volatile pointer to prevent code rearrangement */
9732 volatile XLogCtlData *xlogctl = XLogCtl;
9736 * Check to see if all WAL replayed during online backup
9737 * (i.e., since last restartpoint used as backup starting
9738 * checkpoint) contain full-page writes.
9740 SpinLockAcquire(&xlogctl->info_lck);
9741 recptr = xlogctl->lastFpwDisableRecPtr;
9742 SpinLockRelease(&xlogctl->info_lck);
9744 if (!checkpointfpw || startpoint <= recptr)
9746 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9747 errmsg("WAL generated with full_page_writes=off was replayed "
9748 "since last restartpoint"),
9749 errhint("This means that the backup being taken on the standby "
9750 "is corrupt and should not be used. "
9751 "Enable full_page_writes and run CHECKPOINT on the master, "
9752 "and then try an online backup again.")));
9755 * During recovery, since we don't use the end-of-backup WAL
9756 * record and don't write the backup history file, the
9757 * starting WAL location doesn't need to be unique. This means
9758 * that two base backups started at the same time might use
9759 * the same checkpoint as starting locations.
9761 gotUniqueStartpoint = true;
9765 * If two base backups are started at the same time (in WAL sender
9766 * processes), we need to make sure that they use different
9767 * checkpoints as starting locations, because we use the starting
9768 * WAL location as a unique identifier for the base backup in the
9769 * end-of-backup WAL record and when we write the backup history
9770 * file. Perhaps it would be better generate a separate unique ID
9771 * for each backup instead of forcing another checkpoint, but
9772 * taking a checkpoint right after another is not that expensive
9773 * either because only few buffers have been dirtied yet.
9775 WALInsertSlotAcquire(true);
9776 if (XLogCtl->Insert.lastBackupStart < startpoint)
9778 XLogCtl->Insert.lastBackupStart = startpoint;
9779 gotUniqueStartpoint = true;
9781 WALInsertSlotRelease();
9782 } while (!gotUniqueStartpoint);
9784 XLByteToSeg(startpoint, _logSegNo);
9785 XLogFileName(xlogfilename, ThisTimeLineID, _logSegNo);
9788 * Construct backup label file
9790 initStringInfo(&labelfbuf);
9792 /* Use the log timezone here, not the session timezone */
9793 stamp_time = (pg_time_t) time(NULL);
9794 pg_strftime(strfbuf, sizeof(strfbuf),
9795 "%Y-%m-%d %H:%M:%S %Z",
9796 pg_localtime(&stamp_time, log_timezone));
9797 appendStringInfo(&labelfbuf, "START WAL LOCATION: %X/%X (file %s)\n",
9798 (uint32) (startpoint >> 32), (uint32) startpoint, xlogfilename);
9799 appendStringInfo(&labelfbuf, "CHECKPOINT LOCATION: %X/%X\n",
9800 (uint32) (checkpointloc >> 32), (uint32) checkpointloc);
9801 appendStringInfo(&labelfbuf, "BACKUP METHOD: %s\n",
9802 exclusive ? "pg_start_backup" : "streamed");
9803 appendStringInfo(&labelfbuf, "BACKUP FROM: %s\n",
9804 backup_started_in_recovery ? "standby" : "master");
9805 appendStringInfo(&labelfbuf, "START TIME: %s\n", strfbuf);
9806 appendStringInfo(&labelfbuf, "LABEL: %s\n", backupidstr);
9809 * Okay, write the file, or return its contents to caller.
9814 * Check for existing backup label --- implies a backup is already
9815 * running. (XXX given that we checked exclusiveBackup above,
9816 * maybe it would be OK to just unlink any such label file?)
9818 if (stat(BACKUP_LABEL_FILE, &stat_buf) != 0)
9820 if (errno != ENOENT)
9822 (errcode_for_file_access(),
9823 errmsg("could not stat file \"%s\": %m",
9824 BACKUP_LABEL_FILE)));
9828 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9829 errmsg("a backup is already in progress"),
9830 errhint("If you're sure there is no backup in progress, remove file \"%s\" and try again.",
9831 BACKUP_LABEL_FILE)));
9833 fp = AllocateFile(BACKUP_LABEL_FILE, "w");
9837 (errcode_for_file_access(),
9838 errmsg("could not create file \"%s\": %m",
9839 BACKUP_LABEL_FILE)));
9840 if (fwrite(labelfbuf.data, labelfbuf.len, 1, fp) != 1 ||
9842 pg_fsync(fileno(fp)) != 0 ||
9846 (errcode_for_file_access(),
9847 errmsg("could not write file \"%s\": %m",
9848 BACKUP_LABEL_FILE)));
9849 pfree(labelfbuf.data);
9852 *labelfile = labelfbuf.data;
9854 PG_END_ENSURE_ERROR_CLEANUP(pg_start_backup_callback, (Datum) BoolGetDatum(exclusive));
9857 * We're done. As a convenience, return the starting WAL location.
9860 *starttli_p = starttli;
9864 /* Error cleanup callback for pg_start_backup */
9866 pg_start_backup_callback(int code, Datum arg)
9868 bool exclusive = DatumGetBool(arg);
9870 /* Update backup counters and forcePageWrites on failure */
9871 WALInsertSlotAcquire(true);
9874 Assert(XLogCtl->Insert.exclusiveBackup);
9875 XLogCtl->Insert.exclusiveBackup = false;
9879 Assert(XLogCtl->Insert.nonExclusiveBackups > 0);
9880 XLogCtl->Insert.nonExclusiveBackups--;
9883 if (!XLogCtl->Insert.exclusiveBackup &&
9884 XLogCtl->Insert.nonExclusiveBackups == 0)
9886 XLogCtl->Insert.forcePageWrites = false;
9888 WALInsertSlotRelease();
9892 * do_pg_stop_backup is the workhorse of the user-visible pg_stop_backup()
9895 * If labelfile is NULL, this stops an exclusive backup. Otherwise this stops
9896 * the non-exclusive backup specified by 'labelfile'.
9898 * Returns the last WAL position that must be present to restore from this
9899 * backup, and the corresponding timeline ID in *stoptli_p.
9902 do_pg_stop_backup(char *labelfile, bool waitforarchive, TimeLineID *stoptli_p)
9904 bool exclusive = (labelfile == NULL);
9905 bool backup_started_in_recovery = false;
9906 XLogRecPtr startpoint;
9907 XLogRecPtr stoppoint;
9910 pg_time_t stamp_time;
9912 char histfilepath[MAXPGPATH];
9913 char startxlogfilename[MAXFNAMELEN];
9914 char stopxlogfilename[MAXFNAMELEN];
9915 char lastxlogfilename[MAXFNAMELEN];
9916 char histfilename[MAXFNAMELEN];
9917 char backupfrom[20];
9918 XLogSegNo _logSegNo;
9922 int seconds_before_warning;
9924 bool reported_waiting = false;
9930 backup_started_in_recovery = RecoveryInProgress();
9932 if (!superuser() && !has_rolreplication(GetUserId()))
9934 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
9935 (errmsg("must be superuser or replication role to run a backup"))));
9938 * Currently only non-exclusive backup can be taken during recovery.
9940 if (backup_started_in_recovery && exclusive)
9942 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9943 errmsg("recovery is in progress"),
9944 errhint("WAL control functions cannot be executed during recovery.")));
9947 * During recovery, we don't need to check WAL level. Because, if WAL
9948 * level is not sufficient, it's impossible to get here during recovery.
9950 if (!backup_started_in_recovery && !XLogIsNeeded())
9952 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9953 errmsg("WAL level not sufficient for making an online backup"),
9954 errhint("wal_level must be set to \"archive\" or \"hot_standby\" at server start.")));
9957 * OK to update backup counters and forcePageWrites
9959 WALInsertSlotAcquire(true);
9961 XLogCtl->Insert.exclusiveBackup = false;
9965 * The user-visible pg_start/stop_backup() functions that operate on
9966 * exclusive backups can be called at any time, but for non-exclusive
9967 * backups, it is expected that each do_pg_start_backup() call is
9968 * matched by exactly one do_pg_stop_backup() call.
9970 Assert(XLogCtl->Insert.nonExclusiveBackups > 0);
9971 XLogCtl->Insert.nonExclusiveBackups--;
9974 if (!XLogCtl->Insert.exclusiveBackup &&
9975 XLogCtl->Insert.nonExclusiveBackups == 0)
9977 XLogCtl->Insert.forcePageWrites = false;
9979 WALInsertSlotRelease();
9984 * Read the existing label file into memory.
9986 struct stat statbuf;
9989 if (stat(BACKUP_LABEL_FILE, &statbuf))
9991 if (errno != ENOENT)
9993 (errcode_for_file_access(),
9994 errmsg("could not stat file \"%s\": %m",
9995 BACKUP_LABEL_FILE)));
9997 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9998 errmsg("a backup is not in progress")));
10001 lfp = AllocateFile(BACKUP_LABEL_FILE, "r");
10005 (errcode_for_file_access(),
10006 errmsg("could not read file \"%s\": %m",
10007 BACKUP_LABEL_FILE)));
10009 labelfile = palloc(statbuf.st_size + 1);
10010 r = fread(labelfile, statbuf.st_size, 1, lfp);
10011 labelfile[statbuf.st_size] = '\0';
10014 * Close and remove the backup label file
10016 if (r != 1 || ferror(lfp) || FreeFile(lfp))
10018 (errcode_for_file_access(),
10019 errmsg("could not read file \"%s\": %m",
10020 BACKUP_LABEL_FILE)));
10021 if (unlink(BACKUP_LABEL_FILE) != 0)
10023 (errcode_for_file_access(),
10024 errmsg("could not remove file \"%s\": %m",
10025 BACKUP_LABEL_FILE)));
10029 * Read and parse the START WAL LOCATION line (this code is pretty crude,
10030 * but we are not expecting any variability in the file format).
10032 if (sscanf(labelfile, "START WAL LOCATION: %X/%X (file %24s)%c",
10033 &hi, &lo, startxlogfilename,
10034 &ch) != 4 || ch != '\n')
10036 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
10037 errmsg("invalid data in file \"%s\"", BACKUP_LABEL_FILE)));
10038 startpoint = ((uint64) hi) << 32 | lo;
10039 remaining = strchr(labelfile, '\n') + 1; /* %n is not portable enough */
10042 * Parse the BACKUP FROM line. If we are taking an online backup from the
10043 * standby, we confirm that the standby has not been promoted during the
10046 ptr = strstr(remaining, "BACKUP FROM:");
10047 if (!ptr || sscanf(ptr, "BACKUP FROM: %19s\n", backupfrom) != 1)
10049 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
10050 errmsg("invalid data in file \"%s\"", BACKUP_LABEL_FILE)));
10051 if (strcmp(backupfrom, "standby") == 0 && !backup_started_in_recovery)
10053 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
10054 errmsg("the standby was promoted during online backup"),
10055 errhint("This means that the backup being taken is corrupt "
10056 "and should not be used. "
10057 "Try taking another online backup.")));
10060 * During recovery, we don't write an end-of-backup record. We assume that
10061 * pg_control was backed up last and its minimum recovery point can be
10062 * available as the backup end location. Since we don't have an
10063 * end-of-backup record, we use the pg_control value to check whether
10064 * we've reached the end of backup when starting recovery from this
10065 * backup. We have no way of checking if pg_control wasn't backed up last
10068 * We don't force a switch to new WAL file and wait for all the required
10069 * files to be archived. This is okay if we use the backup to start the
10070 * standby. But, if it's for an archive recovery, to ensure all the
10071 * required files are available, a user should wait for them to be
10072 * archived, or include them into the backup.
10074 * We return the current minimum recovery point as the backup end
10075 * location. Note that it can be greater than the exact backup end
10076 * location if the minimum recovery point is updated after the backup of
10077 * pg_control. This is harmless for current uses.
10079 * XXX currently a backup history file is for informational and debug
10080 * purposes only. It's not essential for an online backup. Furthermore,
10081 * even if it's created, it will not be archived during recovery because
10082 * an archiver is not invoked. So it doesn't seem worthwhile to write a
10083 * backup history file during recovery.
10085 if (backup_started_in_recovery)
10087 /* use volatile pointer to prevent code rearrangement */
10088 volatile XLogCtlData *xlogctl = XLogCtl;
10092 * Check to see if all WAL replayed during online backup contain
10093 * full-page writes.
10095 SpinLockAcquire(&xlogctl->info_lck);
10096 recptr = xlogctl->lastFpwDisableRecPtr;
10097 SpinLockRelease(&xlogctl->info_lck);
10099 if (startpoint <= recptr)
10101 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
10102 errmsg("WAL generated with full_page_writes=off was replayed "
10103 "during online backup"),
10104 errhint("This means that the backup being taken on the standby "
10105 "is corrupt and should not be used. "
10106 "Enable full_page_writes and run CHECKPOINT on the master, "
10107 "and then try an online backup again.")));
10110 LWLockAcquire(ControlFileLock, LW_SHARED);
10111 stoppoint = ControlFile->minRecoveryPoint;
10112 stoptli = ControlFile->minRecoveryPointTLI;
10113 LWLockRelease(ControlFileLock);
10116 *stoptli_p = stoptli;
10121 * Write the backup-end xlog record
10123 rdata.data = (char *) (&startpoint);
10124 rdata.len = sizeof(startpoint);
10125 rdata.buffer = InvalidBuffer;
10127 stoppoint = XLogInsert(RM_XLOG_ID, XLOG_BACKUP_END, &rdata);
10128 stoptli = ThisTimeLineID;
10131 * Force a switch to a new xlog segment file, so that the backup is valid
10132 * as soon as archiver moves out the current segment file.
10134 RequestXLogSwitch();
10136 XLByteToPrevSeg(stoppoint, _logSegNo);
10137 XLogFileName(stopxlogfilename, ThisTimeLineID, _logSegNo);
10139 /* Use the log timezone here, not the session timezone */
10140 stamp_time = (pg_time_t) time(NULL);
10141 pg_strftime(strfbuf, sizeof(strfbuf),
10142 "%Y-%m-%d %H:%M:%S %Z",
10143 pg_localtime(&stamp_time, log_timezone));
10146 * Write the backup history file
10148 XLByteToSeg(startpoint, _logSegNo);
10149 BackupHistoryFilePath(histfilepath, ThisTimeLineID, _logSegNo,
10150 (uint32) (startpoint % XLogSegSize));
10151 fp = AllocateFile(histfilepath, "w");
10154 (errcode_for_file_access(),
10155 errmsg("could not create file \"%s\": %m",
10157 fprintf(fp, "START WAL LOCATION: %X/%X (file %s)\n",
10158 (uint32) (startpoint >> 32), (uint32) startpoint, startxlogfilename);
10159 fprintf(fp, "STOP WAL LOCATION: %X/%X (file %s)\n",
10160 (uint32) (stoppoint >> 32), (uint32) stoppoint, stopxlogfilename);
10161 /* transfer remaining lines from label to history file */
10162 fprintf(fp, "%s", remaining);
10163 fprintf(fp, "STOP TIME: %s\n", strfbuf);
10164 if (fflush(fp) || ferror(fp) || FreeFile(fp))
10166 (errcode_for_file_access(),
10167 errmsg("could not write file \"%s\": %m",
10171 * Clean out any no-longer-needed history files. As a side effect, this
10172 * will post a .ready file for the newly created history file, notifying
10173 * the archiver that history file may be archived immediately.
10175 CleanupBackupHistory();
10178 * If archiving is enabled, wait for all the required WAL files to be
10179 * archived before returning. If archiving isn't enabled, the required WAL
10180 * needs to be transported via streaming replication (hopefully with
10181 * wal_keep_segments set high enough), or some more exotic mechanism like
10182 * polling and copying files from pg_xlog with script. We have no
10183 * knowledge of those mechanisms, so it's up to the user to ensure that he
10184 * gets all the required WAL.
10186 * We wait until both the last WAL file filled during backup and the
10187 * history file have been archived, and assume that the alphabetic sorting
10188 * property of the WAL files ensures any earlier WAL files are safely
10189 * archived as well.
10191 * We wait forever, since archive_command is supposed to work and we
10192 * assume the admin wanted his backup to work completely. If you don't
10193 * wish to wait, you can set statement_timeout. Also, some notices are
10194 * issued to clue in anyone who might be doing this interactively.
10196 if (waitforarchive && XLogArchivingActive())
10198 XLByteToPrevSeg(stoppoint, _logSegNo);
10199 XLogFileName(lastxlogfilename, ThisTimeLineID, _logSegNo);
10201 XLByteToSeg(startpoint, _logSegNo);
10202 BackupHistoryFileName(histfilename, ThisTimeLineID, _logSegNo,
10203 (uint32) (startpoint % XLogSegSize));
10205 seconds_before_warning = 60;
10208 while (XLogArchiveIsBusy(lastxlogfilename) ||
10209 XLogArchiveIsBusy(histfilename))
10211 CHECK_FOR_INTERRUPTS();
10213 if (!reported_waiting && waits > 5)
10216 (errmsg("pg_stop_backup cleanup done, waiting for required WAL segments to be archived")));
10217 reported_waiting = true;
10220 pg_usleep(1000000L);
10222 if (++waits >= seconds_before_warning)
10224 seconds_before_warning *= 2; /* This wraps in >10 years... */
10226 (errmsg("pg_stop_backup still waiting for all required WAL segments to be archived (%d seconds elapsed)",
10228 errhint("Check that your archive_command is executing properly. "
10229 "pg_stop_backup can be canceled safely, "
10230 "but the database backup will not be usable without all the WAL segments.")));
10235 (errmsg("pg_stop_backup complete, all required WAL segments have been archived")));
10237 else if (waitforarchive)
10239 (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
10242 * We're done. As a convenience, return the ending WAL location.
10245 *stoptli_p = stoptli;
10251 * do_pg_abort_backup: abort a running backup
10253 * This does just the most basic steps of do_pg_stop_backup(), by taking the
10254 * system out of backup mode, thus making it a lot more safe to call from
10255 * an error handler.
10257 * NB: This is only for aborting a non-exclusive backup that doesn't write
10258 * backup_label. A backup started with pg_stop_backup() needs to be finished
10259 * with pg_stop_backup().
10262 do_pg_abort_backup(void)
10264 WALInsertSlotAcquire(true);
10265 Assert(XLogCtl->Insert.nonExclusiveBackups > 0);
10266 XLogCtl->Insert.nonExclusiveBackups--;
10268 if (!XLogCtl->Insert.exclusiveBackup &&
10269 XLogCtl->Insert.nonExclusiveBackups == 0)
10271 XLogCtl->Insert.forcePageWrites = false;
10273 WALInsertSlotRelease();
10277 * Get latest redo apply position.
10279 * Exported to allow WALReceiver to read the pointer directly.
10282 GetXLogReplayRecPtr(TimeLineID *replayTLI)
10284 /* use volatile pointer to prevent code rearrangement */
10285 volatile XLogCtlData *xlogctl = XLogCtl;
10289 SpinLockAcquire(&xlogctl->info_lck);
10290 recptr = xlogctl->lastReplayedEndRecPtr;
10291 tli = xlogctl->lastReplayedTLI;
10292 SpinLockRelease(&xlogctl->info_lck);
10300 * Get latest WAL insert pointer
10303 GetXLogInsertRecPtr(void)
10305 volatile XLogCtlInsert *Insert = &XLogCtl->Insert;
10306 uint64 current_bytepos;
10308 SpinLockAcquire(&Insert->insertpos_lck);
10309 current_bytepos = Insert->CurrBytePos;
10310 SpinLockRelease(&Insert->insertpos_lck);
10312 return XLogBytePosToRecPtr(current_bytepos);
10316 * Get latest WAL write pointer
10319 GetXLogWriteRecPtr(void)
10322 /* use volatile pointer to prevent code rearrangement */
10323 volatile XLogCtlData *xlogctl = XLogCtl;
10325 SpinLockAcquire(&xlogctl->info_lck);
10326 LogwrtResult = xlogctl->LogwrtResult;
10327 SpinLockRelease(&xlogctl->info_lck);
10330 return LogwrtResult.Write;
10334 * Returns the redo pointer of the last checkpoint or restartpoint. This is
10335 * the oldest point in WAL that we still need, if we have to restart recovery.
10338 GetOldestRestartPoint(XLogRecPtr *oldrecptr, TimeLineID *oldtli)
10340 LWLockAcquire(ControlFileLock, LW_SHARED);
10341 *oldrecptr = ControlFile->checkPointCopy.redo;
10342 *oldtli = ControlFile->checkPointCopy.ThisTimeLineID;
10343 LWLockRelease(ControlFileLock);
10347 * read_backup_label: check to see if a backup_label file is present
10349 * If we see a backup_label during recovery, we assume that we are recovering
10350 * from a backup dump file, and we therefore roll forward from the checkpoint
10351 * identified by the label file, NOT what pg_control says. This avoids the
10352 * problem that pg_control might have been archived one or more checkpoints
10353 * later than the start of the dump, and so if we rely on it as the start
10354 * point, we will fail to restore a consistent database state.
10356 * Returns TRUE if a backup_label was found (and fills the checkpoint
10357 * location and its REDO location into *checkPointLoc and RedoStartLSN,
10358 * respectively); returns FALSE if not. If this backup_label came from a
10359 * streamed backup, *backupEndRequired is set to TRUE. If this backup_label
10360 * was created during recovery, *backupFromStandby is set to TRUE.
10363 read_backup_label(XLogRecPtr *checkPointLoc, bool *backupEndRequired,
10364 bool *backupFromStandby)
10366 char startxlogfilename[MAXFNAMELEN];
10370 char backuptype[20];
10371 char backupfrom[20];
10375 *backupEndRequired = false;
10376 *backupFromStandby = false;
10379 * See if label file is present
10381 lfp = AllocateFile(BACKUP_LABEL_FILE, "r");
10384 if (errno != ENOENT)
10386 (errcode_for_file_access(),
10387 errmsg("could not read file \"%s\": %m",
10388 BACKUP_LABEL_FILE)));
10389 return false; /* it's not there, all is fine */
10393 * Read and parse the START WAL LOCATION and CHECKPOINT lines (this code
10394 * is pretty crude, but we are not expecting any variability in the file
10397 if (fscanf(lfp, "START WAL LOCATION: %X/%X (file %08X%16s)%c",
10398 &hi, &lo, &tli, startxlogfilename, &ch) != 5 || ch != '\n')
10400 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
10401 errmsg("invalid data in file \"%s\"", BACKUP_LABEL_FILE)));
10402 RedoStartLSN = ((uint64) hi) << 32 | lo;
10403 if (fscanf(lfp, "CHECKPOINT LOCATION: %X/%X%c",
10404 &hi, &lo, &ch) != 3 || ch != '\n')
10406 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
10407 errmsg("invalid data in file \"%s\"", BACKUP_LABEL_FILE)));
10408 *checkPointLoc = ((uint64) hi) << 32 | lo;
10411 * BACKUP METHOD and BACKUP FROM lines are new in 9.2. We can't restore
10412 * from an older backup anyway, but since the information on it is not
10413 * strictly required, don't error out if it's missing for some reason.
10415 if (fscanf(lfp, "BACKUP METHOD: %19s\n", backuptype) == 1)
10417 if (strcmp(backuptype, "streamed") == 0)
10418 *backupEndRequired = true;
10421 if (fscanf(lfp, "BACKUP FROM: %19s\n", backupfrom) == 1)
10423 if (strcmp(backupfrom, "standby") == 0)
10424 *backupFromStandby = true;
10427 if (ferror(lfp) || FreeFile(lfp))
10429 (errcode_for_file_access(),
10430 errmsg("could not read file \"%s\": %m",
10431 BACKUP_LABEL_FILE)));
10437 * Error context callback for errors occurring during rm_redo().
10440 rm_redo_error_callback(void *arg)
10442 XLogRecord *record = (XLogRecord *) arg;
10443 StringInfoData buf;
10445 initStringInfo(&buf);
10446 RmgrTable[record->xl_rmid].rm_desc(&buf,
10448 XLogRecGetData(record));
10450 /* don't bother emitting empty description */
10452 errcontext("xlog redo %s", buf.data);
10458 * BackupInProgress: check if online backup mode is active
10460 * This is done by checking for existence of the "backup_label" file.
10463 BackupInProgress(void)
10465 struct stat stat_buf;
10467 return (stat(BACKUP_LABEL_FILE, &stat_buf) == 0);
10471 * CancelBackup: rename the "backup_label" file to cancel backup mode
10473 * If the "backup_label" file exists, it will be renamed to "backup_label.old".
10474 * Note that this will render an online backup in progress useless.
10475 * To correctly finish an online backup, pg_stop_backup must be called.
10480 struct stat stat_buf;
10482 /* if the file is not there, return */
10483 if (stat(BACKUP_LABEL_FILE, &stat_buf) < 0)
10486 /* remove leftover file from previously canceled backup if it exists */
10487 unlink(BACKUP_LABEL_OLD);
10489 if (rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD) == 0)
10492 (errmsg("online backup mode canceled"),
10493 errdetail("\"%s\" was renamed to \"%s\".",
10494 BACKUP_LABEL_FILE, BACKUP_LABEL_OLD)));
10499 (errcode_for_file_access(),
10500 errmsg("online backup mode was not canceled"),
10501 errdetail("Could not rename \"%s\" to \"%s\": %m.",
10502 BACKUP_LABEL_FILE, BACKUP_LABEL_OLD)));
10507 * Read the XLOG page containing RecPtr into readBuf (if not read already).
10508 * Returns number of bytes read, if the page is read successfully, or -1
10509 * in case of errors. When errors occur, they are ereport'ed, but only
10510 * if they have not been previously reported.
10512 * This is responsible for restoring files from archive as needed, as well
10513 * as for waiting for the requested WAL record to arrive in standby mode.
10515 * 'emode' specifies the log level used for reporting "file not found" or
10516 * "end of WAL" situations in archive recovery, or in standby mode when a
10517 * trigger file is found. If set to WARNING or below, XLogPageRead() returns
10518 * false in those situations, on higher log levels the ereport() won't
10521 * In standby mode, if after a successful return of XLogPageRead() the
10522 * caller finds the record it's interested in to be broken, it should
10523 * ereport the error with the level determined by
10524 * emode_for_corrupt_record(), and then set lastSourceFailed
10525 * and call XLogPageRead() again with the same arguments. This lets
10526 * XLogPageRead() to try fetching the record from another source, or to
10530 XLogPageRead(XLogReaderState *xlogreader, XLogRecPtr targetPagePtr, int reqLen,
10531 XLogRecPtr targetRecPtr, char *readBuf, TimeLineID *readTLI)
10533 XLogPageReadPrivate *private =
10534 (XLogPageReadPrivate *) xlogreader->private_data;
10535 int emode = private->emode;
10536 uint32 targetPageOff;
10537 XLogSegNo targetSegNo PG_USED_FOR_ASSERTS_ONLY;
10539 XLByteToSeg(targetPagePtr, targetSegNo);
10540 targetPageOff = targetPagePtr % XLogSegSize;
10543 * See if we need to switch to a new segment because the requested record
10544 * is not in the currently open one.
10546 if (readFile >= 0 && !XLByteInSeg(targetPagePtr, readSegNo))
10549 * Request a restartpoint if we've replayed too much xlog since the
10552 if (StandbyModeRequested && bgwriterLaunched)
10554 if (XLogCheckpointNeeded(readSegNo))
10556 (void) GetRedoRecPtr();
10557 if (XLogCheckpointNeeded(readSegNo))
10558 RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
10567 XLByteToSeg(targetPagePtr, readSegNo);
10570 /* See if we need to retrieve more data */
10571 if (readFile < 0 ||
10572 (readSource == XLOG_FROM_STREAM &&
10573 receivedUpto < targetPagePtr + reqLen))
10575 if (!WaitForWALToBecomeAvailable(targetPagePtr + reqLen,
10576 private->randAccess,
10577 private->fetching_ckpt,
10591 * At this point, we have the right segment open and if we're streaming we
10592 * know the requested record is in it.
10594 Assert(readFile != -1);
10597 * If the current segment is being streamed from master, calculate how
10598 * much of the current page we have received already. We know the
10599 * requested record has been received, but this is for the benefit of
10600 * future calls, to allow quick exit at the top of this function.
10602 if (readSource == XLOG_FROM_STREAM)
10604 if (((targetPagePtr) / XLOG_BLCKSZ) != (receivedUpto / XLOG_BLCKSZ))
10605 readLen = XLOG_BLCKSZ;
10607 readLen = receivedUpto % XLogSegSize - targetPageOff;
10610 readLen = XLOG_BLCKSZ;
10612 /* Read the requested page */
10613 readOff = targetPageOff;
10614 if (lseek(readFile, (off_t) readOff, SEEK_SET) < 0)
10616 char fname[MAXFNAMELEN];
10618 XLogFileName(fname, curFileTLI, readSegNo);
10619 ereport(emode_for_corrupt_record(emode, targetPagePtr + reqLen),
10620 (errcode_for_file_access(),
10621 errmsg("could not seek in log segment %s to offset %u: %m",
10623 goto next_record_is_invalid;
10626 if (read(readFile, readBuf, XLOG_BLCKSZ) != XLOG_BLCKSZ)
10628 char fname[MAXFNAMELEN];
10630 XLogFileName(fname, curFileTLI, readSegNo);
10631 ereport(emode_for_corrupt_record(emode, targetPagePtr + reqLen),
10632 (errcode_for_file_access(),
10633 errmsg("could not read from log segment %s, offset %u: %m",
10635 goto next_record_is_invalid;
10638 Assert(targetSegNo == readSegNo);
10639 Assert(targetPageOff == readOff);
10640 Assert(reqLen <= readLen);
10642 *readTLI = curFileTLI;
10645 next_record_is_invalid:
10646 lastSourceFailed = true;
10654 /* In standby-mode, keep trying */
10662 * Open the WAL segment containing WAL position 'RecPtr'.
10664 * The segment can be fetched via restore_command, or via walreceiver having
10665 * streamed the record, or it can already be present in pg_xlog. Checking
10666 * pg_xlog is mainly for crash recovery, but it will be polled in standby mode
10667 * too, in case someone copies a new segment directly to pg_xlog. That is not
10668 * documented or recommended, though.
10670 * If 'fetching_ckpt' is true, we're fetching a checkpoint record, and should
10671 * prepare to read WAL starting from RedoStartLSN after this.
10673 * 'RecPtr' might not point to the beginning of the record we're interested
10674 * in, it might also point to the page or segment header. In that case,
10675 * 'tliRecPtr' is the position of the WAL record we're interested in. It is
10676 * used to decide which timeline to stream the requested WAL from.
10678 * If the the record is not immediately available, the function returns false
10679 * if we're not in standby mode. In standby mode, waits for it to become
10682 * When the requested record becomes available, the function opens the file
10683 * containing it (if not open already), and returns true. When end of standby
10684 * mode is triggered by the user, and there is no more WAL available, returns
10688 WaitForWALToBecomeAvailable(XLogRecPtr RecPtr, bool randAccess,
10689 bool fetching_ckpt, XLogRecPtr tliRecPtr)
10691 static pg_time_t last_fail_time = 0;
10695 * Standby mode is implemented by a state machine:
10697 * 1. Read from archive (XLOG_FROM_ARCHIVE)
10698 * 2. Read from pg_xlog (XLOG_FROM_PG_XLOG)
10699 * 3. Check trigger file
10700 * 4. Read from primary server via walreceiver (XLOG_FROM_STREAM)
10701 * 5. Rescan timelines
10702 * 6. Sleep 5 seconds, and loop back to 1.
10704 * Failure to read from the current source advances the state machine to
10705 * the next state. In addition, successfully reading a file from pg_xlog
10706 * moves the state machine from state 2 back to state 1 (we always prefer
10707 * files in the archive over files in pg_xlog).
10709 * 'currentSource' indicates the current state. There are no currentSource
10710 * values for "check trigger", "rescan timelines", and "sleep" states,
10711 * those actions are taken when reading from the previous source fails, as
10712 * part of advancing to the next state.
10715 if (!InArchiveRecovery)
10716 currentSource = XLOG_FROM_PG_XLOG;
10717 else if (currentSource == 0)
10718 currentSource = XLOG_FROM_ARCHIVE;
10722 int oldSource = currentSource;
10725 * First check if we failed to read from the current source, and
10726 * advance the state machine if so. The failure to read might've
10727 * happened outside this function, e.g when a CRC check fails on a
10728 * record, or within this loop.
10730 if (lastSourceFailed)
10732 switch (currentSource)
10734 case XLOG_FROM_ARCHIVE:
10735 currentSource = XLOG_FROM_PG_XLOG;
10738 case XLOG_FROM_PG_XLOG:
10741 * Check to see if the trigger file exists. Note that we
10742 * do this only after failure, so when you create the
10743 * trigger file, we still finish replaying as much as we
10744 * can from archive and pg_xlog before failover.
10746 if (StandbyMode && CheckForStandbyTrigger())
10753 * Not in standby mode, and we've now tried the archive
10760 * If primary_conninfo is set, launch walreceiver to try
10761 * to stream the missing WAL.
10763 * If fetching_ckpt is TRUE, RecPtr points to the initial
10764 * checkpoint location. In that case, we use RedoStartLSN
10765 * as the streaming start position instead of RecPtr, so
10766 * that when we later jump backwards to start redo at
10767 * RedoStartLSN, we will have the logs streamed already.
10769 if (PrimaryConnInfo)
10776 ptr = RedoStartLSN;
10777 tli = ControlFile->checkPointCopy.ThisTimeLineID;
10782 tli = tliOfPointInHistory(tliRecPtr, expectedTLEs);
10784 if (curFileTLI > 0 && tli < curFileTLI)
10785 elog(ERROR, "according to history file, WAL location %X/%X belongs to timeline %u, but previous recovered WAL file came from timeline %u",
10786 (uint32) (ptr >> 32), (uint32) ptr,
10790 RequestXLogStreaming(tli, ptr, PrimaryConnInfo);
10795 * Move to XLOG_FROM_STREAM state in either case. We'll
10796 * get immediate failure if we didn't launch walreceiver,
10797 * and move on to the next state.
10799 currentSource = XLOG_FROM_STREAM;
10802 case XLOG_FROM_STREAM:
10805 * Failure while streaming. Most likely, we got here
10806 * because streaming replication was terminated, or
10807 * promotion was triggered. But we also get here if we
10808 * find an invalid record in the WAL streamed from master,
10809 * in which case something is seriously wrong. There's
10810 * little chance that the problem will just go away, but
10811 * PANIC is not good for availability either, especially
10812 * in hot standby mode. So, we treat that the same as
10813 * disconnection, and retry from archive/pg_xlog again.
10814 * The WAL in the archive should be identical to what was
10815 * streamed, so it's unlikely that it helps, but one can
10820 * Before we leave XLOG_FROM_STREAM state, make sure that
10821 * walreceiver is not active, so that it won't overwrite
10822 * WAL that we restore from archive.
10824 if (WalRcvStreaming())
10828 * Before we sleep, re-scan for possible new timelines if
10829 * we were requested to recover to the latest timeline.
10831 if (recoveryTargetIsLatest)
10833 if (rescanLatestTimeLine())
10835 currentSource = XLOG_FROM_ARCHIVE;
10841 * XLOG_FROM_STREAM is the last state in our state
10842 * machine, so we've exhausted all the options for
10843 * obtaining the requested WAL. We're going to loop back
10844 * and retry from the archive, but if it hasn't been long
10845 * since last attempt, sleep 5 seconds to avoid
10848 now = (pg_time_t) time(NULL);
10849 if ((now - last_fail_time) < 5)
10851 pg_usleep(1000000L * (5 - (now - last_fail_time)));
10852 now = (pg_time_t) time(NULL);
10854 last_fail_time = now;
10855 currentSource = XLOG_FROM_ARCHIVE;
10859 elog(ERROR, "unexpected WAL source %d", currentSource);
10862 else if (currentSource == XLOG_FROM_PG_XLOG)
10865 * We just successfully read a file in pg_xlog. We prefer files in
10866 * the archive over ones in pg_xlog, so try the next file again
10867 * from the archive first.
10869 if (InArchiveRecovery)
10870 currentSource = XLOG_FROM_ARCHIVE;
10873 if (currentSource != oldSource)
10874 elog(DEBUG2, "switched WAL source from %s to %s after %s",
10875 xlogSourceNames[oldSource], xlogSourceNames[currentSource],
10876 lastSourceFailed ? "failure" : "success");
10879 * We've now handled possible failure. Try to read from the chosen
10882 lastSourceFailed = false;
10884 switch (currentSource)
10886 case XLOG_FROM_ARCHIVE:
10887 case XLOG_FROM_PG_XLOG:
10888 /* Close any old file we might have open. */
10894 /* Reset curFileTLI if random fetch. */
10899 * Try to restore the file from archive, or read an existing
10900 * file from pg_xlog.
10902 readFile = XLogFileReadAnyTLI(readSegNo, DEBUG2, currentSource);
10904 return true; /* success! */
10907 * Nope, not found in archive or pg_xlog.
10909 lastSourceFailed = true;
10912 case XLOG_FROM_STREAM:
10917 * Check if WAL receiver is still active.
10919 if (!WalRcvStreaming())
10921 lastSourceFailed = true;
10926 * Walreceiver is active, so see if new data has arrived.
10928 * We only advance XLogReceiptTime when we obtain fresh
10929 * WAL from walreceiver and observe that we had already
10930 * processed everything before the most recent "chunk"
10931 * that it flushed to disk. In steady state where we are
10932 * keeping up with the incoming data, XLogReceiptTime will
10933 * be updated on each cycle. When we are behind,
10934 * XLogReceiptTime will not advance, so the grace time
10935 * allotted to conflicting queries will decrease.
10937 if (RecPtr < receivedUpto)
10941 XLogRecPtr latestChunkStart;
10943 receivedUpto = GetWalRcvWriteRecPtr(&latestChunkStart, &receiveTLI);
10944 if (RecPtr < receivedUpto && receiveTLI == curFileTLI)
10947 if (latestChunkStart <= RecPtr)
10949 XLogReceiptTime = GetCurrentTimestamp();
10950 SetCurrentChunkStartTime(XLogReceiptTime);
10959 * Great, streamed far enough. Open the file if it's
10960 * not open already. Also read the timeline history
10961 * file if we haven't initialized timeline history
10962 * yet; it should be streamed over and present in
10963 * pg_xlog by now. Use XLOG_FROM_STREAM so that
10964 * source info is set correctly and XLogReceiptTime
10970 expectedTLEs = readTimeLineHistory(receiveTLI);
10971 readFile = XLogFileRead(readSegNo, PANIC,
10973 XLOG_FROM_STREAM, false);
10974 Assert(readFile >= 0);
10978 /* just make sure source info is correct... */
10979 readSource = XLOG_FROM_STREAM;
10980 XLogReceiptSource = XLOG_FROM_STREAM;
10987 * Data not here yet. Check for trigger, then wait for
10988 * walreceiver to wake us up when new WAL arrives.
10990 if (CheckForStandbyTrigger())
10993 * Note that we don't "return false" immediately here.
10994 * After being triggered, we still want to replay all
10995 * the WAL that was already streamed. It's in pg_xlog
10996 * now, so we just treat this as a failure, and the
10997 * state machine will move on to replay the streamed
10998 * WAL from pg_xlog, and then recheck the trigger and
11001 lastSourceFailed = true;
11006 * Wait for more WAL to arrive. Time out after 5 seconds,
11007 * like when polling the archive, to react to a trigger
11010 WaitLatch(&XLogCtl->recoveryWakeupLatch,
11011 WL_LATCH_SET | WL_TIMEOUT,
11013 ResetLatch(&XLogCtl->recoveryWakeupLatch);
11018 elog(ERROR, "unexpected WAL source %d", currentSource);
11022 * This possibly-long loop needs to handle interrupts of startup
11025 HandleStartupProcInterrupts();
11026 } while (StandbyMode);
11032 * Determine what log level should be used to report a corrupt WAL record
11033 * in the current WAL page, previously read by XLogPageRead().
11035 * 'emode' is the error mode that would be used to report a file-not-found
11036 * or legitimate end-of-WAL situation. Generally, we use it as-is, but if
11037 * we're retrying the exact same record that we've tried previously, only
11038 * complain the first time to keep the noise down. However, we only do when
11039 * reading from pg_xlog, because we don't expect any invalid records in archive
11040 * or in records streamed from master. Files in the archive should be complete,
11041 * and we should never hit the end of WAL because we stop and wait for more WAL
11042 * to arrive before replaying it.
11044 * NOTE: This function remembers the RecPtr value it was last called with,
11045 * to suppress repeated messages about the same record. Only call this when
11046 * you are about to ereport(), or you might cause a later message to be
11047 * erroneously suppressed.
11050 emode_for_corrupt_record(int emode, XLogRecPtr RecPtr)
11052 static XLogRecPtr lastComplaint = 0;
11054 if (readSource == XLOG_FROM_PG_XLOG && emode == LOG)
11056 if (RecPtr == lastComplaint)
11059 lastComplaint = RecPtr;
11065 * Check to see whether the user-specified trigger file exists and whether a
11066 * promote request has arrived. If either condition holds, return true.
11069 CheckForStandbyTrigger(void)
11071 struct stat stat_buf;
11072 static bool triggered = false;
11077 if (IsPromoteTriggered())
11080 * In 9.1 and 9.2 the postmaster unlinked the promote file inside the
11081 * signal handler. It now leaves the file in place and lets the
11082 * Startup process do the unlink. This allows Startup to know whether
11083 * it should create a full checkpoint before starting up (fallback
11084 * mode). Fast promotion takes precedence.
11086 if (stat(PROMOTE_SIGNAL_FILE, &stat_buf) == 0)
11088 unlink(PROMOTE_SIGNAL_FILE);
11089 unlink(FALLBACK_PROMOTE_SIGNAL_FILE);
11090 fast_promote = true;
11092 else if (stat(FALLBACK_PROMOTE_SIGNAL_FILE, &stat_buf) == 0)
11094 unlink(FALLBACK_PROMOTE_SIGNAL_FILE);
11095 fast_promote = false;
11098 ereport(LOG, (errmsg("received promote request")));
11100 ResetPromoteTriggered();
11105 if (TriggerFile == NULL)
11108 if (stat(TriggerFile, &stat_buf) == 0)
11111 (errmsg("trigger file found: %s", TriggerFile)));
11112 unlink(TriggerFile);
11114 fast_promote = true;
11121 * Check to see if a promote request has arrived. Should be
11122 * called by postmaster after receiving SIGUSR1.
11125 CheckPromoteSignal(void)
11127 struct stat stat_buf;
11129 if (stat(PROMOTE_SIGNAL_FILE, &stat_buf) == 0 ||
11130 stat(FALLBACK_PROMOTE_SIGNAL_FILE, &stat_buf) == 0)
11137 * Wake up startup process to replay newly arrived WAL, or to notice that
11138 * failover has been requested.
11141 WakeupRecovery(void)
11143 SetLatch(&XLogCtl->recoveryWakeupLatch);
11147 * Update the WalWriterSleeping flag.
11150 SetWalWriterSleeping(bool sleeping)
11152 /* use volatile pointer to prevent code rearrangement */
11153 volatile XLogCtlData *xlogctl = XLogCtl;
11155 SpinLockAcquire(&xlogctl->info_lck);
11156 xlogctl->WalWriterSleeping = sleeping;
11157 SpinLockRelease(&xlogctl->info_lck);