Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * xlog.c
4 : * PostgreSQL write-ahead log manager
5 : *
6 : * The Write-Ahead Log (WAL) functionality is split into several source
7 : * files, in addition to this one:
8 : *
9 : * xloginsert.c - Functions for constructing WAL records
10 : * xlogrecovery.c - WAL recovery and standby code
11 : * xlogreader.c - Facility for reading WAL files and parsing WAL records
12 : * xlogutils.c - Helper functions for WAL redo routines
13 : *
14 : * This file contains functions for coordinating database startup and
15 : * checkpointing, and managing the write-ahead log buffers when the
16 : * system is running.
17 : *
18 : * StartupXLOG() is the main entry point of the startup process. It
19 : * coordinates database startup, performing WAL recovery, and the
20 : * transition from WAL recovery into normal operations.
21 : *
22 : * XLogInsertRecord() inserts a WAL record into the WAL buffers. Most
23 : * callers should not call this directly, but use the functions in
24 : * xloginsert.c to construct the WAL record. XLogFlush() can be used
25 : * to force the WAL to disk.
26 : *
27 : * In addition to those, there are many other functions for interrogating
28 : * the current system state, and for starting/stopping backups.
29 : *
30 : *
31 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
32 : * Portions Copyright (c) 1994, Regents of the University of California
33 : *
34 : * src/backend/access/transam/xlog.c
35 : *
36 : *-------------------------------------------------------------------------
37 : */
38 :
39 : #include "postgres.h"
40 :
41 : #include <ctype.h>
42 : #include <math.h>
43 : #include <time.h>
44 : #include <fcntl.h>
45 : #include <sys/stat.h>
46 : #include <sys/time.h>
47 : #include <unistd.h>
48 :
49 : #include "access/clog.h"
50 : #include "access/commit_ts.h"
51 : #include "access/heaptoast.h"
52 : #include "access/multixact.h"
53 : #include "access/rewriteheap.h"
54 : #include "access/subtrans.h"
55 : #include "access/timeline.h"
56 : #include "access/transam.h"
57 : #include "access/twophase.h"
58 : #include "access/xact.h"
59 : #include "access/xlog_internal.h"
60 : #include "access/xlogarchive.h"
61 : #include "access/xloginsert.h"
62 : #include "access/xlogreader.h"
63 : #include "access/xlogrecovery.h"
64 : #include "access/xlogutils.h"
65 : #include "backup/basebackup.h"
66 : #include "catalog/catversion.h"
67 : #include "catalog/pg_control.h"
68 : #include "catalog/pg_database.h"
69 : #include "common/controldata_utils.h"
70 : #include "common/file_utils.h"
71 : #include "executor/instrument.h"
72 : #include "miscadmin.h"
73 : #include "pg_trace.h"
74 : #include "pgstat.h"
75 : #include "port/atomics.h"
76 : #include "postmaster/bgwriter.h"
77 : #include "postmaster/startup.h"
78 : #include "postmaster/walsummarizer.h"
79 : #include "postmaster/walwriter.h"
80 : #include "replication/origin.h"
81 : #include "replication/slot.h"
82 : #include "replication/snapbuild.h"
83 : #include "replication/walreceiver.h"
84 : #include "replication/walsender.h"
85 : #include "storage/bufmgr.h"
86 : #include "storage/fd.h"
87 : #include "storage/ipc.h"
88 : #include "storage/large_object.h"
89 : #include "storage/latch.h"
90 : #include "storage/predicate.h"
91 : #include "storage/proc.h"
92 : #include "storage/procarray.h"
93 : #include "storage/reinit.h"
94 : #include "storage/spin.h"
95 : #include "storage/sync.h"
96 : #include "utils/guc_hooks.h"
97 : #include "utils/guc_tables.h"
98 : #include "utils/injection_point.h"
99 : #include "utils/ps_status.h"
100 : #include "utils/relmapper.h"
101 : #include "utils/snapmgr.h"
102 : #include "utils/timeout.h"
103 : #include "utils/timestamp.h"
104 : #include "utils/varlena.h"
105 :
106 : #ifdef WAL_DEBUG
107 : #include "utils/memutils.h"
108 : #endif
109 :
110 : /* timeline ID to be used when bootstrapping */
111 : #define BootstrapTimeLineID 1
112 :
113 : /* User-settable parameters */
114 : int max_wal_size_mb = 1024; /* 1 GB */
115 : int min_wal_size_mb = 80; /* 80 MB */
116 : int wal_keep_size_mb = 0;
117 : int XLOGbuffers = -1;
118 : int XLogArchiveTimeout = 0;
119 : int XLogArchiveMode = ARCHIVE_MODE_OFF;
120 : char *XLogArchiveCommand = NULL;
121 : bool EnableHotStandby = false;
122 : bool fullPageWrites = true;
123 : bool wal_log_hints = false;
124 : int wal_compression = WAL_COMPRESSION_NONE;
125 : char *wal_consistency_checking_string = NULL;
126 : bool *wal_consistency_checking = NULL;
127 : bool wal_init_zero = true;
128 : bool wal_recycle = true;
129 : bool log_checkpoints = true;
130 : int wal_sync_method = DEFAULT_WAL_SYNC_METHOD;
131 : int wal_level = WAL_LEVEL_REPLICA;
132 : int CommitDelay = 0; /* precommit delay in microseconds */
133 : int CommitSiblings = 5; /* # concurrent xacts needed to sleep */
134 : int wal_retrieve_retry_interval = 5000;
135 : int max_slot_wal_keep_size_mb = -1;
136 : int wal_decode_buffer_size = 512 * 1024;
137 : bool track_wal_io_timing = false;
138 :
139 : #ifdef WAL_DEBUG
140 : bool XLOG_DEBUG = false;
141 : #endif
142 :
143 : int wal_segment_size = DEFAULT_XLOG_SEG_SIZE;
144 :
145 : /*
146 : * Number of WAL insertion locks to use. A higher value allows more insertions
147 : * to happen concurrently, but adds some CPU overhead to flushing the WAL,
148 : * which needs to iterate all the locks.
149 : */
150 : #define NUM_XLOGINSERT_LOCKS 8
151 :
152 : /*
153 : * Max distance from last checkpoint, before triggering a new xlog-based
154 : * checkpoint.
155 : */
156 : int CheckPointSegments;
157 :
158 : /* Estimated distance between checkpoints, in bytes */
159 : static double CheckPointDistanceEstimate = 0;
160 : static double PrevCheckPointDistance = 0;
161 :
162 : /*
163 : * Track whether there were any deferred checks for custom resource managers
164 : * specified in wal_consistency_checking.
165 : */
166 : static bool check_wal_consistency_checking_deferred = false;
167 :
168 : /*
169 : * GUC support
170 : */
171 : const struct config_enum_entry wal_sync_method_options[] = {
172 : {"fsync", WAL_SYNC_METHOD_FSYNC, false},
173 : #ifdef HAVE_FSYNC_WRITETHROUGH
174 : {"fsync_writethrough", WAL_SYNC_METHOD_FSYNC_WRITETHROUGH, false},
175 : #endif
176 : {"fdatasync", WAL_SYNC_METHOD_FDATASYNC, false},
177 : #ifdef O_SYNC
178 : {"open_sync", WAL_SYNC_METHOD_OPEN, false},
179 : #endif
180 : #ifdef O_DSYNC
181 : {"open_datasync", WAL_SYNC_METHOD_OPEN_DSYNC, false},
182 : #endif
183 : {NULL, 0, false}
184 : };
185 :
186 :
187 : /*
188 : * Although only "on", "off", and "always" are documented,
189 : * we accept all the likely variants of "on" and "off".
190 : */
191 : const struct config_enum_entry archive_mode_options[] = {
192 : {"always", ARCHIVE_MODE_ALWAYS, false},
193 : {"on", ARCHIVE_MODE_ON, false},
194 : {"off", ARCHIVE_MODE_OFF, false},
195 : {"true", ARCHIVE_MODE_ON, true},
196 : {"false", ARCHIVE_MODE_OFF, true},
197 : {"yes", ARCHIVE_MODE_ON, true},
198 : {"no", ARCHIVE_MODE_OFF, true},
199 : {"1", ARCHIVE_MODE_ON, true},
200 : {"0", ARCHIVE_MODE_OFF, true},
201 : {NULL, 0, false}
202 : };
203 :
204 : /*
205 : * Statistics for current checkpoint are collected in this global struct.
206 : * Because only the checkpointer or a stand-alone backend can perform
207 : * checkpoints, this will be unused in normal backends.
208 : */
209 : CheckpointStatsData CheckpointStats;
210 :
211 : /*
212 : * During recovery, lastFullPageWrites keeps track of full_page_writes that
213 : * the replayed WAL records indicate. It's initialized with full_page_writes
214 : * that the recovery starting checkpoint record indicates, and then updated
215 : * each time XLOG_FPW_CHANGE record is replayed.
216 : */
217 : static bool lastFullPageWrites;
218 :
219 : /*
220 : * Local copy of the state tracked by SharedRecoveryState in shared memory,
221 : * It is false if SharedRecoveryState is RECOVERY_STATE_DONE. True actually
222 : * means "not known, need to check the shared state".
223 : */
224 : static bool LocalRecoveryInProgress = true;
225 :
226 : /*
227 : * Local state for XLogInsertAllowed():
228 : * 1: unconditionally allowed to insert XLOG
229 : * 0: unconditionally not allowed to insert XLOG
230 : * -1: must check RecoveryInProgress(); disallow until it is false
231 : * Most processes start with -1 and transition to 1 after seeing that recovery
232 : * is not in progress. But we can also force the value for special cases.
233 : * The coding in XLogInsertAllowed() depends on the first two of these states
234 : * being numerically the same as bool true and false.
235 : */
236 : static int LocalXLogInsertAllowed = -1;
237 :
238 : /*
239 : * ProcLastRecPtr points to the start of the last XLOG record inserted by the
240 : * current backend. It is updated for all inserts. XactLastRecEnd points to
241 : * end+1 of the last record, and is reset when we end a top-level transaction,
242 : * or start a new one; so it can be used to tell if the current transaction has
243 : * created any XLOG records.
244 : *
245 : * While in parallel mode, this may not be fully up to date. When committing,
246 : * a transaction can assume this covers all xlog records written either by the
247 : * user backend or by any parallel worker which was present at any point during
248 : * the transaction. But when aborting, or when still in parallel mode, other
249 : * parallel backends may have written WAL records at later LSNs than the value
250 : * stored here. The parallel leader advances its own copy, when necessary,
251 : * in WaitForParallelWorkersToFinish.
252 : */
253 : XLogRecPtr ProcLastRecPtr = InvalidXLogRecPtr;
254 : XLogRecPtr XactLastRecEnd = InvalidXLogRecPtr;
255 : XLogRecPtr XactLastCommitEnd = InvalidXLogRecPtr;
256 :
257 : /*
258 : * RedoRecPtr is this backend's local copy of the REDO record pointer
259 : * (which is almost but not quite the same as a pointer to the most recent
260 : * CHECKPOINT record). We update this from the shared-memory copy,
261 : * XLogCtl->Insert.RedoRecPtr, whenever we can safely do so (ie, when we
262 : * hold an insertion lock). See XLogInsertRecord for details. We are also
263 : * allowed to update from XLogCtl->RedoRecPtr if we hold the info_lck;
264 : * see GetRedoRecPtr.
265 : *
266 : * NB: Code that uses this variable must be prepared not only for the
267 : * possibility that it may be arbitrarily out of date, but also for the
268 : * possibility that it might be set to InvalidXLogRecPtr. We used to
269 : * initialize it as a side effect of the first call to RecoveryInProgress(),
270 : * which meant that most code that might use it could assume that it had a
271 : * real if perhaps stale value. That's no longer the case.
272 : */
273 : static XLogRecPtr RedoRecPtr;
274 :
275 : /*
276 : * doPageWrites is this backend's local copy of (fullPageWrites ||
277 : * runningBackups > 0). It is used together with RedoRecPtr to decide whether
278 : * a full-page image of a page need to be taken.
279 : *
280 : * NB: Initially this is false, and there's no guarantee that it will be
281 : * initialized to any other value before it is first used. Any code that
282 : * makes use of it must recheck the value after obtaining a WALInsertLock,
283 : * and respond appropriately if it turns out that the previous value wasn't
284 : * accurate.
285 : */
286 : static bool doPageWrites;
287 :
288 : /*----------
289 : * Shared-memory data structures for XLOG control
290 : *
291 : * LogwrtRqst indicates a byte position that we need to write and/or fsync
292 : * the log up to (all records before that point must be written or fsynced).
293 : * The positions already written/fsynced are maintained in logWriteResult
294 : * and logFlushResult using atomic access.
295 : * In addition to the shared variable, each backend has a private copy of
296 : * both in LogwrtResult, which is updated when convenient.
297 : *
298 : * The request bookkeeping is simpler: there is a shared XLogCtl->LogwrtRqst
299 : * (protected by info_lck), but we don't need to cache any copies of it.
300 : *
301 : * info_lck is only held long enough to read/update the protected variables,
302 : * so it's a plain spinlock. The other locks are held longer (potentially
303 : * over I/O operations), so we use LWLocks for them. These locks are:
304 : *
305 : * WALBufMappingLock: must be held to replace a page in the WAL buffer cache.
306 : * It is only held while initializing and changing the mapping. If the
307 : * contents of the buffer being replaced haven't been written yet, the mapping
308 : * lock is released while the write is done, and reacquired afterwards.
309 : *
310 : * WALWriteLock: must be held to write WAL buffers to disk (XLogWrite or
311 : * XLogFlush).
312 : *
313 : * ControlFileLock: must be held to read/update control file or create
314 : * new log file.
315 : *
316 : *----------
317 : */
318 :
319 : typedef struct XLogwrtRqst
320 : {
321 : XLogRecPtr Write; /* last byte + 1 to write out */
322 : XLogRecPtr Flush; /* last byte + 1 to flush */
323 : } XLogwrtRqst;
324 :
325 : typedef struct XLogwrtResult
326 : {
327 : XLogRecPtr Write; /* last byte + 1 written out */
328 : XLogRecPtr Flush; /* last byte + 1 flushed */
329 : } XLogwrtResult;
330 :
331 : /*
332 : * Inserting to WAL is protected by a small fixed number of WAL insertion
333 : * locks. To insert to the WAL, you must hold one of the locks - it doesn't
334 : * matter which one. To lock out other concurrent insertions, you must hold
335 : * of them. Each WAL insertion lock consists of a lightweight lock, plus an
336 : * indicator of how far the insertion has progressed (insertingAt).
337 : *
338 : * The insertingAt values are read when a process wants to flush WAL from
339 : * the in-memory buffers to disk, to check that all the insertions to the
340 : * region the process is about to write out have finished. You could simply
341 : * wait for all currently in-progress insertions to finish, but the
342 : * insertingAt indicator allows you to ignore insertions to later in the WAL,
343 : * so that you only wait for the insertions that are modifying the buffers
344 : * you're about to write out.
345 : *
346 : * This isn't just an optimization. If all the WAL buffers are dirty, an
347 : * inserter that's holding a WAL insert lock might need to evict an old WAL
348 : * buffer, which requires flushing the WAL. If it's possible for an inserter
349 : * to block on another inserter unnecessarily, deadlock can arise when two
350 : * inserters holding a WAL insert lock wait for each other to finish their
351 : * insertion.
352 : *
353 : * Small WAL records that don't cross a page boundary never update the value,
354 : * the WAL record is just copied to the page and the lock is released. But
355 : * to avoid the deadlock-scenario explained above, the indicator is always
356 : * updated before sleeping while holding an insertion lock.
357 : *
358 : * lastImportantAt contains the LSN of the last important WAL record inserted
359 : * using a given lock. This value is used to detect if there has been
360 : * important WAL activity since the last time some action, like a checkpoint,
361 : * was performed - allowing to not repeat the action if not. The LSN is
362 : * updated for all insertions, unless the XLOG_MARK_UNIMPORTANT flag was
363 : * set. lastImportantAt is never cleared, only overwritten by the LSN of newer
364 : * records. Tracking the WAL activity directly in WALInsertLock has the
365 : * advantage of not needing any additional locks to update the value.
366 : */
367 : typedef struct
368 : {
369 : LWLock lock;
370 : pg_atomic_uint64 insertingAt;
371 : XLogRecPtr lastImportantAt;
372 : } WALInsertLock;
373 :
374 : /*
375 : * All the WAL insertion locks are allocated as an array in shared memory. We
376 : * force the array stride to be a power of 2, which saves a few cycles in
377 : * indexing, but more importantly also ensures that individual slots don't
378 : * cross cache line boundaries. (Of course, we have to also ensure that the
379 : * array start address is suitably aligned.)
380 : */
381 : typedef union WALInsertLockPadded
382 : {
383 : WALInsertLock l;
384 : char pad[PG_CACHE_LINE_SIZE];
385 : } WALInsertLockPadded;
386 :
387 : /*
388 : * Session status of running backup, used for sanity checks in SQL-callable
389 : * functions to start and stop backups.
390 : */
391 : static SessionBackupState sessionBackupState = SESSION_BACKUP_NONE;
392 :
393 : /*
394 : * Shared state data for WAL insertion.
395 : */
396 : typedef struct XLogCtlInsert
397 : {
398 : slock_t insertpos_lck; /* protects CurrBytePos and PrevBytePos */
399 :
400 : /*
401 : * CurrBytePos is the end of reserved WAL. The next record will be
402 : * inserted at that position. PrevBytePos is the start position of the
403 : * previously inserted (or rather, reserved) record - it is copied to the
404 : * prev-link of the next record. These are stored as "usable byte
405 : * positions" rather than XLogRecPtrs (see XLogBytePosToRecPtr()).
406 : */
407 : uint64 CurrBytePos;
408 : uint64 PrevBytePos;
409 :
410 : /*
411 : * Make sure the above heavily-contended spinlock and byte positions are
412 : * on their own cache line. In particular, the RedoRecPtr and full page
413 : * write variables below should be on a different cache line. They are
414 : * read on every WAL insertion, but updated rarely, and we don't want
415 : * those reads to steal the cache line containing Curr/PrevBytePos.
416 : */
417 : char pad[PG_CACHE_LINE_SIZE];
418 :
419 : /*
420 : * fullPageWrites is the authoritative value used by all backends to
421 : * determine whether to write full-page image to WAL. This shared value,
422 : * instead of the process-local fullPageWrites, is required because, when
423 : * full_page_writes is changed by SIGHUP, we must WAL-log it before it
424 : * actually affects WAL-logging by backends. Checkpointer sets at startup
425 : * or after SIGHUP.
426 : *
427 : * To read these fields, you must hold an insertion lock. To modify them,
428 : * you must hold ALL the locks.
429 : */
430 : XLogRecPtr RedoRecPtr; /* current redo point for insertions */
431 : bool fullPageWrites;
432 :
433 : /*
434 : * runningBackups is a counter indicating the number of backups currently
435 : * in progress. lastBackupStart is the latest checkpoint redo location
436 : * used as a starting point for an online backup.
437 : */
438 : int runningBackups;
439 : XLogRecPtr lastBackupStart;
440 :
441 : /*
442 : * WAL insertion locks.
443 : */
444 : WALInsertLockPadded *WALInsertLocks;
445 : } XLogCtlInsert;
446 :
447 : /*
448 : * Total shared-memory state for XLOG.
449 : */
450 : typedef struct XLogCtlData
451 : {
452 : XLogCtlInsert Insert;
453 :
454 : /* Protected by info_lck: */
455 : XLogwrtRqst LogwrtRqst;
456 : XLogRecPtr RedoRecPtr; /* a recent copy of Insert->RedoRecPtr */
457 : FullTransactionId ckptFullXid; /* nextXid of latest checkpoint */
458 : XLogRecPtr asyncXactLSN; /* LSN of newest async commit/abort */
459 : XLogRecPtr replicationSlotMinLSN; /* oldest LSN needed by any slot */
460 :
461 : XLogSegNo lastRemovedSegNo; /* latest removed/recycled XLOG segment */
462 :
463 : /* Fake LSN counter, for unlogged relations. */
464 : pg_atomic_uint64 unloggedLSN;
465 :
466 : /* Time and LSN of last xlog segment switch. Protected by WALWriteLock. */
467 : pg_time_t lastSegSwitchTime;
468 : XLogRecPtr lastSegSwitchLSN;
469 :
470 : /* These are accessed using atomics -- info_lck not needed */
471 : pg_atomic_uint64 logInsertResult; /* last byte + 1 inserted to buffers */
472 : pg_atomic_uint64 logWriteResult; /* last byte + 1 written out */
473 : pg_atomic_uint64 logFlushResult; /* last byte + 1 flushed */
474 :
475 : /*
476 : * Latest initialized page in the cache (last byte position + 1).
477 : *
478 : * To change the identity of a buffer (and InitializedUpTo), you need to
479 : * hold WALBufMappingLock. To change the identity of a buffer that's
480 : * still dirty, the old page needs to be written out first, and for that
481 : * you need WALWriteLock, and you need to ensure that there are no
482 : * in-progress insertions to the page by calling
483 : * WaitXLogInsertionsToFinish().
484 : */
485 : XLogRecPtr InitializedUpTo;
486 :
487 : /*
488 : * These values do not change after startup, although the pointed-to pages
489 : * and xlblocks values certainly do. xlblocks values are protected by
490 : * WALBufMappingLock.
491 : */
492 : char *pages; /* buffers for unwritten XLOG pages */
493 : pg_atomic_uint64 *xlblocks; /* 1st byte ptr-s + XLOG_BLCKSZ */
494 : int XLogCacheBlck; /* highest allocated xlog buffer index */
495 :
496 : /*
497 : * InsertTimeLineID is the timeline into which new WAL is being inserted
498 : * and flushed. It is zero during recovery, and does not change once set.
499 : *
500 : * If we create a new timeline when the system was started up,
501 : * PrevTimeLineID is the old timeline's ID that we forked off from.
502 : * Otherwise it's equal to InsertTimeLineID.
503 : *
504 : * We set these fields while holding info_lck. Most that reads these
505 : * values knows that recovery is no longer in progress and so can safely
506 : * read the value without a lock, but code that could be run either during
507 : * or after recovery can take info_lck while reading these values.
508 : */
509 : TimeLineID InsertTimeLineID;
510 : TimeLineID PrevTimeLineID;
511 :
512 : /*
513 : * SharedRecoveryState indicates if we're still in crash or archive
514 : * recovery. Protected by info_lck.
515 : */
516 : RecoveryState SharedRecoveryState;
517 :
518 : /*
519 : * InstallXLogFileSegmentActive indicates whether the checkpointer should
520 : * arrange for future segments by recycling and/or PreallocXlogFiles().
521 : * Protected by ControlFileLock. Only the startup process changes it. If
522 : * true, anyone can use InstallXLogFileSegment(). If false, the startup
523 : * process owns the exclusive right to install segments, by reading from
524 : * the archive and possibly replacing existing files.
525 : */
526 : bool InstallXLogFileSegmentActive;
527 :
528 : /*
529 : * WalWriterSleeping indicates whether the WAL writer is currently in
530 : * low-power mode (and hence should be nudged if an async commit occurs).
531 : * Protected by info_lck.
532 : */
533 : bool WalWriterSleeping;
534 :
535 : /*
536 : * During recovery, we keep a copy of the latest checkpoint record here.
537 : * lastCheckPointRecPtr points to start of checkpoint record and
538 : * lastCheckPointEndPtr points to end+1 of checkpoint record. Used by the
539 : * checkpointer when it wants to create a restartpoint.
540 : *
541 : * Protected by info_lck.
542 : */
543 : XLogRecPtr lastCheckPointRecPtr;
544 : XLogRecPtr lastCheckPointEndPtr;
545 : CheckPoint lastCheckPoint;
546 :
547 : /*
548 : * lastFpwDisableRecPtr points to the start of the last replayed
549 : * XLOG_FPW_CHANGE record that instructs full_page_writes is disabled.
550 : */
551 : XLogRecPtr lastFpwDisableRecPtr;
552 :
553 : slock_t info_lck; /* locks shared variables shown above */
554 : } XLogCtlData;
555 :
556 : /*
557 : * Classification of XLogRecordInsert operations.
558 : */
559 : typedef enum
560 : {
561 : WALINSERT_NORMAL,
562 : WALINSERT_SPECIAL_SWITCH,
563 : WALINSERT_SPECIAL_CHECKPOINT
564 : } WalInsertClass;
565 :
566 : static XLogCtlData *XLogCtl = NULL;
567 :
568 : /* a private copy of XLogCtl->Insert.WALInsertLocks, for convenience */
569 : static WALInsertLockPadded *WALInsertLocks = NULL;
570 :
571 : /*
572 : * We maintain an image of pg_control in shared memory.
573 : */
574 : static ControlFileData *ControlFile = NULL;
575 :
576 : /*
577 : * Calculate the amount of space left on the page after 'endptr'. Beware
578 : * multiple evaluation!
579 : */
580 : #define INSERT_FREESPACE(endptr) \
581 : (((endptr) % XLOG_BLCKSZ == 0) ? 0 : (XLOG_BLCKSZ - (endptr) % XLOG_BLCKSZ))
582 :
583 : /* Macro to advance to next buffer index. */
584 : #define NextBufIdx(idx) \
585 : (((idx) == XLogCtl->XLogCacheBlck) ? 0 : ((idx) + 1))
586 :
587 : /*
588 : * XLogRecPtrToBufIdx returns the index of the WAL buffer that holds, or
589 : * would hold if it was in cache, the page containing 'recptr'.
590 : */
591 : #define XLogRecPtrToBufIdx(recptr) \
592 : (((recptr) / XLOG_BLCKSZ) % (XLogCtl->XLogCacheBlck + 1))
593 :
594 : /*
595 : * These are the number of bytes in a WAL page usable for WAL data.
596 : */
597 : #define UsableBytesInPage (XLOG_BLCKSZ - SizeOfXLogShortPHD)
598 :
599 : /*
600 : * Convert values of GUCs measured in megabytes to equiv. segment count.
601 : * Rounds down.
602 : */
603 : #define ConvertToXSegs(x, segsize) XLogMBVarToSegs((x), (segsize))
604 :
605 : /* The number of bytes in a WAL segment usable for WAL data. */
606 : static int UsableBytesInSegment;
607 :
608 : /*
609 : * Private, possibly out-of-date copy of shared LogwrtResult.
610 : * See discussion above.
611 : */
612 : static XLogwrtResult LogwrtResult = {0, 0};
613 :
614 : /*
615 : * Update local copy of shared XLogCtl->log{Write,Flush}Result
616 : *
617 : * It's critical that Flush always trails Write, so the order of the reads is
618 : * important, as is the barrier. See also XLogWrite.
619 : */
620 : #define RefreshXLogWriteResult(_target) \
621 : do { \
622 : _target.Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult); \
623 : pg_read_barrier(); \
624 : _target.Write = pg_atomic_read_u64(&XLogCtl->logWriteResult); \
625 : } while (0)
626 :
627 : /*
628 : * openLogFile is -1 or a kernel FD for an open log file segment.
629 : * openLogSegNo identifies the segment, and openLogTLI the corresponding TLI.
630 : * These variables are only used to write the XLOG, and so will normally refer
631 : * to the active segment.
632 : *
633 : * Note: call Reserve/ReleaseExternalFD to track consumption of this FD.
634 : */
635 : static int openLogFile = -1;
636 : static XLogSegNo openLogSegNo = 0;
637 : static TimeLineID openLogTLI = 0;
638 :
639 : /*
640 : * Local copies of equivalent fields in the control file. When running
641 : * crash recovery, LocalMinRecoveryPoint is set to InvalidXLogRecPtr as we
642 : * expect to replay all the WAL available, and updateMinRecoveryPoint is
643 : * switched to false to prevent any updates while replaying records.
644 : * Those values are kept consistent as long as crash recovery runs.
645 : */
646 : static XLogRecPtr LocalMinRecoveryPoint;
647 : static TimeLineID LocalMinRecoveryPointTLI;
648 : static bool updateMinRecoveryPoint = true;
649 :
650 : /* For WALInsertLockAcquire/Release functions */
651 : static int MyLockNo = 0;
652 : static bool holdingAllLocks = false;
653 :
654 : #ifdef WAL_DEBUG
655 : static MemoryContext walDebugCxt = NULL;
656 : #endif
657 :
658 : static void CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI,
659 : XLogRecPtr EndOfLog,
660 : TimeLineID newTLI);
661 : static void CheckRequiredParameterValues(void);
662 : static void XLogReportParameters(void);
663 : static int LocalSetXLogInsertAllowed(void);
664 : static void CreateEndOfRecoveryRecord(void);
665 : static XLogRecPtr CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn,
666 : XLogRecPtr pagePtr,
667 : TimeLineID newTLI);
668 : static void CheckPointGuts(XLogRecPtr checkPointRedo, int flags);
669 : static void KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo);
670 : static XLogRecPtr XLogGetReplicationSlotMinimumLSN(void);
671 :
672 : static void AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli,
673 : bool opportunistic);
674 : static void XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible);
675 : static bool InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
676 : bool find_free, XLogSegNo max_segno,
677 : TimeLineID tli);
678 : static void XLogFileClose(void);
679 : static void PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli);
680 : static void RemoveTempXlogFiles(void);
681 : static void RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr,
682 : XLogRecPtr endptr, TimeLineID insertTLI);
683 : static void RemoveXlogFile(const struct dirent *segment_de,
684 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
685 : TimeLineID insertTLI);
686 : static void UpdateLastRemovedPtr(char *filename);
687 : static void ValidateXLOGDirectoryStructure(void);
688 : static void CleanupBackupHistory(void);
689 : static void UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force);
690 : static bool PerformRecoveryXLogAction(void);
691 : static void InitControlFile(uint64 sysidentifier, uint32 data_checksum_version);
692 : static void WriteControlFile(void);
693 : static void ReadControlFile(void);
694 : static void UpdateControlFile(void);
695 : static char *str_time(pg_time_t tnow);
696 :
697 : static int get_sync_bit(int method);
698 :
699 : static void CopyXLogRecordToWAL(int write_len, bool isLogSwitch,
700 : XLogRecData *rdata,
701 : XLogRecPtr StartPos, XLogRecPtr EndPos,
702 : TimeLineID tli);
703 : static void ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos,
704 : XLogRecPtr *EndPos, XLogRecPtr *PrevPtr);
705 : static bool ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos,
706 : XLogRecPtr *PrevPtr);
707 : static XLogRecPtr WaitXLogInsertionsToFinish(XLogRecPtr upto);
708 : static char *GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli);
709 : static XLogRecPtr XLogBytePosToRecPtr(uint64 bytepos);
710 : static XLogRecPtr XLogBytePosToEndRecPtr(uint64 bytepos);
711 : static uint64 XLogRecPtrToBytePos(XLogRecPtr ptr);
712 :
713 : static void WALInsertLockAcquire(void);
714 : static void WALInsertLockAcquireExclusive(void);
715 : static void WALInsertLockRelease(void);
716 : static void WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt);
717 :
718 : /*
719 : * Insert an XLOG record represented by an already-constructed chain of data
720 : * chunks. This is a low-level routine; to construct the WAL record header
721 : * and data, use the higher-level routines in xloginsert.c.
722 : *
723 : * If 'fpw_lsn' is valid, it is the oldest LSN among the pages that this
724 : * WAL record applies to, that were not included in the record as full page
725 : * images. If fpw_lsn <= RedoRecPtr, the function does not perform the
726 : * insertion and returns InvalidXLogRecPtr. The caller can then recalculate
727 : * which pages need a full-page image, and retry. If fpw_lsn is invalid, the
728 : * record is always inserted.
729 : *
730 : * 'flags' gives more in-depth control on the record being inserted. See
731 : * XLogSetRecordFlags() for details.
732 : *
733 : * 'topxid_included' tells whether the top-transaction id is logged along with
734 : * current subtransaction. See XLogRecordAssemble().
735 : *
736 : * The first XLogRecData in the chain must be for the record header, and its
737 : * data must be MAXALIGNed. XLogInsertRecord fills in the xl_prev and
738 : * xl_crc fields in the header, the rest of the header must already be filled
739 : * by the caller.
740 : *
741 : * Returns XLOG pointer to end of record (beginning of next record).
742 : * This can be used as LSN for data pages affected by the logged action.
743 : * (LSN is the XLOG point up to which the XLOG must be flushed to disk
744 : * before the data page can be written out. This implements the basic
745 : * WAL rule "write the log before the data".)
746 : */
747 : XLogRecPtr
748 27699216 : XLogInsertRecord(XLogRecData *rdata,
749 : XLogRecPtr fpw_lsn,
750 : uint8 flags,
751 : int num_fpi,
752 : bool topxid_included)
753 : {
754 27699216 : XLogCtlInsert *Insert = &XLogCtl->Insert;
755 : pg_crc32c rdata_crc;
756 : bool inserted;
757 27699216 : XLogRecord *rechdr = (XLogRecord *) rdata->data;
758 27699216 : uint8 info = rechdr->xl_info & ~XLR_INFO_MASK;
759 27699216 : WalInsertClass class = WALINSERT_NORMAL;
760 : XLogRecPtr StartPos;
761 : XLogRecPtr EndPos;
762 27699216 : bool prevDoPageWrites = doPageWrites;
763 : TimeLineID insertTLI;
764 :
765 : /* Does this record type require special handling? */
766 27699216 : if (unlikely(rechdr->xl_rmid == RM_XLOG_ID))
767 : {
768 400384 : if (info == XLOG_SWITCH)
769 736 : class = WALINSERT_SPECIAL_SWITCH;
770 399648 : else if (info == XLOG_CHECKPOINT_REDO)
771 1056 : class = WALINSERT_SPECIAL_CHECKPOINT;
772 : }
773 :
774 : /* we assume that all of the record header is in the first chunk */
775 : Assert(rdata->len >= SizeOfXLogRecord);
776 :
777 : /* cross-check on whether we should be here or not */
778 27699216 : if (!XLogInsertAllowed())
779 0 : elog(ERROR, "cannot make new WAL entries during recovery");
780 :
781 : /*
782 : * Given that we're not in recovery, InsertTimeLineID is set and can't
783 : * change, so we can read it without a lock.
784 : */
785 27699216 : insertTLI = XLogCtl->InsertTimeLineID;
786 :
787 : /*----------
788 : *
789 : * We have now done all the preparatory work we can without holding a
790 : * lock or modifying shared state. From here on, inserting the new WAL
791 : * record to the shared WAL buffer cache is a two-step process:
792 : *
793 : * 1. Reserve the right amount of space from the WAL. The current head of
794 : * reserved space is kept in Insert->CurrBytePos, and is protected by
795 : * insertpos_lck.
796 : *
797 : * 2. Copy the record to the reserved WAL space. This involves finding the
798 : * correct WAL buffer containing the reserved space, and copying the
799 : * record in place. This can be done concurrently in multiple processes.
800 : *
801 : * To keep track of which insertions are still in-progress, each concurrent
802 : * inserter acquires an insertion lock. In addition to just indicating that
803 : * an insertion is in progress, the lock tells others how far the inserter
804 : * has progressed. There is a small fixed number of insertion locks,
805 : * determined by NUM_XLOGINSERT_LOCKS. When an inserter crosses a page
806 : * boundary, it updates the value stored in the lock to the how far it has
807 : * inserted, to allow the previous buffer to be flushed.
808 : *
809 : * Holding onto an insertion lock also protects RedoRecPtr and
810 : * fullPageWrites from changing until the insertion is finished.
811 : *
812 : * Step 2 can usually be done completely in parallel. If the required WAL
813 : * page is not initialized yet, you have to grab WALBufMappingLock to
814 : * initialize it, but the WAL writer tries to do that ahead of insertions
815 : * to avoid that from happening in the critical path.
816 : *
817 : *----------
818 : */
819 27699216 : START_CRIT_SECTION();
820 :
821 27699216 : if (likely(class == WALINSERT_NORMAL))
822 : {
823 27697424 : WALInsertLockAcquire();
824 :
825 : /*
826 : * Check to see if my copy of RedoRecPtr is out of date. If so, may
827 : * have to go back and have the caller recompute everything. This can
828 : * only happen just after a checkpoint, so it's better to be slow in
829 : * this case and fast otherwise.
830 : *
831 : * Also check to see if fullPageWrites was just turned on or there's a
832 : * running backup (which forces full-page writes); if we weren't
833 : * already doing full-page writes then go back and recompute.
834 : *
835 : * If we aren't doing full-page writes then RedoRecPtr doesn't
836 : * actually affect the contents of the XLOG record, so we'll update
837 : * our local copy but not force a recomputation. (If doPageWrites was
838 : * just turned off, we could recompute the record without full pages,
839 : * but we choose not to bother.)
840 : */
841 27697424 : if (RedoRecPtr != Insert->RedoRecPtr)
842 : {
843 : Assert(RedoRecPtr < Insert->RedoRecPtr);
844 12272 : RedoRecPtr = Insert->RedoRecPtr;
845 : }
846 27697424 : doPageWrites = (Insert->fullPageWrites || Insert->runningBackups > 0);
847 :
848 27697424 : if (doPageWrites &&
849 27289762 : (!prevDoPageWrites ||
850 25065278 : (fpw_lsn != InvalidXLogRecPtr && fpw_lsn <= RedoRecPtr)))
851 : {
852 : /*
853 : * Oops, some buffer now needs to be backed up that the caller
854 : * didn't back up. Start over.
855 : */
856 13710 : WALInsertLockRelease();
857 13710 : END_CRIT_SECTION();
858 13710 : return InvalidXLogRecPtr;
859 : }
860 :
861 : /*
862 : * Reserve space for the record in the WAL. This also sets the xl_prev
863 : * pointer.
864 : */
865 27683714 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
866 : &rechdr->xl_prev);
867 :
868 : /* Normal records are always inserted. */
869 27683714 : inserted = true;
870 : }
871 1792 : else if (class == WALINSERT_SPECIAL_SWITCH)
872 : {
873 : /*
874 : * In order to insert an XLOG_SWITCH record, we need to hold all of
875 : * the WAL insertion locks, not just one, so that no one else can
876 : * begin inserting a record until we've figured out how much space
877 : * remains in the current WAL segment and claimed all of it.
878 : *
879 : * Nonetheless, this case is simpler than the normal cases handled
880 : * below, which must check for changes in doPageWrites and RedoRecPtr.
881 : * Those checks are only needed for records that can contain buffer
882 : * references, and an XLOG_SWITCH record never does.
883 : */
884 : Assert(fpw_lsn == InvalidXLogRecPtr);
885 736 : WALInsertLockAcquireExclusive();
886 736 : inserted = ReserveXLogSwitch(&StartPos, &EndPos, &rechdr->xl_prev);
887 : }
888 : else
889 : {
890 : Assert(class == WALINSERT_SPECIAL_CHECKPOINT);
891 :
892 : /*
893 : * We need to update both the local and shared copies of RedoRecPtr,
894 : * which means that we need to hold all the WAL insertion locks.
895 : * However, there can't be any buffer references, so as above, we need
896 : * not check RedoRecPtr before inserting the record; we just need to
897 : * update it afterwards.
898 : */
899 : Assert(fpw_lsn == InvalidXLogRecPtr);
900 1056 : WALInsertLockAcquireExclusive();
901 1056 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
902 : &rechdr->xl_prev);
903 1056 : RedoRecPtr = Insert->RedoRecPtr = StartPos;
904 1056 : inserted = true;
905 : }
906 :
907 27685506 : if (inserted)
908 : {
909 : /*
910 : * Now that xl_prev has been filled in, calculate CRC of the record
911 : * header.
912 : */
913 27685392 : rdata_crc = rechdr->xl_crc;
914 27685392 : COMP_CRC32C(rdata_crc, rechdr, offsetof(XLogRecord, xl_crc));
915 27685392 : FIN_CRC32C(rdata_crc);
916 27685392 : rechdr->xl_crc = rdata_crc;
917 :
918 : /*
919 : * All the record data, including the header, is now ready to be
920 : * inserted. Copy the record in the space reserved.
921 : */
922 27685392 : CopyXLogRecordToWAL(rechdr->xl_tot_len,
923 : class == WALINSERT_SPECIAL_SWITCH, rdata,
924 : StartPos, EndPos, insertTLI);
925 :
926 : /*
927 : * Unless record is flagged as not important, update LSN of last
928 : * important record in the current slot. When holding all locks, just
929 : * update the first one.
930 : */
931 27685392 : if ((flags & XLOG_MARK_UNIMPORTANT) == 0)
932 : {
933 27517178 : int lockno = holdingAllLocks ? 0 : MyLockNo;
934 :
935 27517178 : WALInsertLocks[lockno].l.lastImportantAt = StartPos;
936 : }
937 : }
938 : else
939 : {
940 : /*
941 : * This was an xlog-switch record, but the current insert location was
942 : * already exactly at the beginning of a segment, so there was no need
943 : * to do anything.
944 : */
945 : }
946 :
947 : /*
948 : * Done! Let others know that we're finished.
949 : */
950 27685506 : WALInsertLockRelease();
951 :
952 27685506 : END_CRIT_SECTION();
953 :
954 27685506 : MarkCurrentTransactionIdLoggedIfAny();
955 :
956 : /*
957 : * Mark top transaction id is logged (if needed) so that we should not try
958 : * to log it again with the next WAL record in the current subtransaction.
959 : */
960 27685506 : if (topxid_included)
961 438 : MarkSubxactTopXidLogged();
962 :
963 : /*
964 : * Update shared LogwrtRqst.Write, if we crossed page boundary.
965 : */
966 27685506 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
967 : {
968 3158366 : SpinLockAcquire(&XLogCtl->info_lck);
969 : /* advance global request to include new block(s) */
970 3158366 : if (XLogCtl->LogwrtRqst.Write < EndPos)
971 3011004 : XLogCtl->LogwrtRqst.Write = EndPos;
972 3158366 : SpinLockRelease(&XLogCtl->info_lck);
973 3158366 : RefreshXLogWriteResult(LogwrtResult);
974 : }
975 :
976 : /*
977 : * If this was an XLOG_SWITCH record, flush the record and the empty
978 : * padding space that fills the rest of the segment, and perform
979 : * end-of-segment actions (eg, notifying archiver).
980 : */
981 27685506 : if (class == WALINSERT_SPECIAL_SWITCH)
982 : {
983 : TRACE_POSTGRESQL_WAL_SWITCH();
984 736 : XLogFlush(EndPos);
985 :
986 : /*
987 : * Even though we reserved the rest of the segment for us, which is
988 : * reflected in EndPos, we return a pointer to just the end of the
989 : * xlog-switch record.
990 : */
991 736 : if (inserted)
992 : {
993 622 : EndPos = StartPos + SizeOfXLogRecord;
994 622 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
995 : {
996 4 : uint64 offset = XLogSegmentOffset(EndPos, wal_segment_size);
997 :
998 4 : if (offset == EndPos % XLOG_BLCKSZ)
999 0 : EndPos += SizeOfXLogLongPHD;
1000 : else
1001 4 : EndPos += SizeOfXLogShortPHD;
1002 : }
1003 : }
1004 : }
1005 :
1006 : #ifdef WAL_DEBUG
1007 : if (XLOG_DEBUG)
1008 : {
1009 : static XLogReaderState *debug_reader = NULL;
1010 : XLogRecord *record;
1011 : DecodedXLogRecord *decoded;
1012 : StringInfoData buf;
1013 : StringInfoData recordBuf;
1014 : char *errormsg = NULL;
1015 : MemoryContext oldCxt;
1016 :
1017 : oldCxt = MemoryContextSwitchTo(walDebugCxt);
1018 :
1019 : initStringInfo(&buf);
1020 : appendStringInfo(&buf, "INSERT @ %X/%X: ", LSN_FORMAT_ARGS(EndPos));
1021 :
1022 : /*
1023 : * We have to piece together the WAL record data from the XLogRecData
1024 : * entries, so that we can pass it to the rm_desc function as one
1025 : * contiguous chunk.
1026 : */
1027 : initStringInfo(&recordBuf);
1028 : for (; rdata != NULL; rdata = rdata->next)
1029 : appendBinaryStringInfo(&recordBuf, rdata->data, rdata->len);
1030 :
1031 : /* We also need temporary space to decode the record. */
1032 : record = (XLogRecord *) recordBuf.data;
1033 : decoded = (DecodedXLogRecord *)
1034 : palloc(DecodeXLogRecordRequiredSpace(record->xl_tot_len));
1035 :
1036 : if (!debug_reader)
1037 : debug_reader = XLogReaderAllocate(wal_segment_size, NULL,
1038 : XL_ROUTINE(.page_read = NULL,
1039 : .segment_open = NULL,
1040 : .segment_close = NULL),
1041 : NULL);
1042 : if (!debug_reader)
1043 : {
1044 : appendStringInfoString(&buf, "error decoding record: out of memory while allocating a WAL reading processor");
1045 : }
1046 : else if (!DecodeXLogRecord(debug_reader,
1047 : decoded,
1048 : record,
1049 : EndPos,
1050 : &errormsg))
1051 : {
1052 : appendStringInfo(&buf, "error decoding record: %s",
1053 : errormsg ? errormsg : "no error message");
1054 : }
1055 : else
1056 : {
1057 : appendStringInfoString(&buf, " - ");
1058 :
1059 : debug_reader->record = decoded;
1060 : xlog_outdesc(&buf, debug_reader);
1061 : debug_reader->record = NULL;
1062 : }
1063 : elog(LOG, "%s", buf.data);
1064 :
1065 : pfree(decoded);
1066 : pfree(buf.data);
1067 : pfree(recordBuf.data);
1068 : MemoryContextSwitchTo(oldCxt);
1069 : }
1070 : #endif
1071 :
1072 : /*
1073 : * Update our global variables
1074 : */
1075 27685506 : ProcLastRecPtr = StartPos;
1076 27685506 : XactLastRecEnd = EndPos;
1077 :
1078 : /* Report WAL traffic to the instrumentation. */
1079 27685506 : if (inserted)
1080 : {
1081 27685392 : pgWalUsage.wal_bytes += rechdr->xl_tot_len;
1082 27685392 : pgWalUsage.wal_records++;
1083 27685392 : pgWalUsage.wal_fpi += num_fpi;
1084 : }
1085 :
1086 27685506 : return EndPos;
1087 : }
1088 :
1089 : /*
1090 : * Reserves the right amount of space for a record of given size from the WAL.
1091 : * *StartPos is set to the beginning of the reserved section, *EndPos to
1092 : * its end+1. *PrevPtr is set to the beginning of the previous record; it is
1093 : * used to set the xl_prev of this record.
1094 : *
1095 : * This is the performance critical part of XLogInsert that must be serialized
1096 : * across backends. The rest can happen mostly in parallel. Try to keep this
1097 : * section as short as possible, insertpos_lck can be heavily contended on a
1098 : * busy system.
1099 : *
1100 : * NB: The space calculation here must match the code in CopyXLogRecordToWAL,
1101 : * where we actually copy the record to the reserved space.
1102 : *
1103 : * NB: Testing shows that XLogInsertRecord runs faster if this code is inlined;
1104 : * however, because there are two call sites, the compiler is reluctant to
1105 : * inline. We use pg_attribute_always_inline here to try to convince it.
1106 : */
1107 : static pg_attribute_always_inline void
1108 27684770 : ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1109 : XLogRecPtr *PrevPtr)
1110 : {
1111 27684770 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1112 : uint64 startbytepos;
1113 : uint64 endbytepos;
1114 : uint64 prevbytepos;
1115 :
1116 27684770 : size = MAXALIGN(size);
1117 :
1118 : /* All (non xlog-switch) records should contain data. */
1119 : Assert(size > SizeOfXLogRecord);
1120 :
1121 : /*
1122 : * The duration the spinlock needs to be held is minimized by minimizing
1123 : * the calculations that have to be done while holding the lock. The
1124 : * current tip of reserved WAL is kept in CurrBytePos, as a byte position
1125 : * that only counts "usable" bytes in WAL, that is, it excludes all WAL
1126 : * page headers. The mapping between "usable" byte positions and physical
1127 : * positions (XLogRecPtrs) can be done outside the locked region, and
1128 : * because the usable byte position doesn't include any headers, reserving
1129 : * X bytes from WAL is almost as simple as "CurrBytePos += X".
1130 : */
1131 27684770 : SpinLockAcquire(&Insert->insertpos_lck);
1132 :
1133 27684770 : startbytepos = Insert->CurrBytePos;
1134 27684770 : endbytepos = startbytepos + size;
1135 27684770 : prevbytepos = Insert->PrevBytePos;
1136 27684770 : Insert->CurrBytePos = endbytepos;
1137 27684770 : Insert->PrevBytePos = startbytepos;
1138 :
1139 27684770 : SpinLockRelease(&Insert->insertpos_lck);
1140 :
1141 27684770 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1142 27684770 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1143 27684770 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1144 :
1145 : /*
1146 : * Check that the conversions between "usable byte positions" and
1147 : * XLogRecPtrs work consistently in both directions.
1148 : */
1149 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1150 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1151 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1152 27684770 : }
1153 :
1154 : /*
1155 : * Like ReserveXLogInsertLocation(), but for an xlog-switch record.
1156 : *
1157 : * A log-switch record is handled slightly differently. The rest of the
1158 : * segment will be reserved for this insertion, as indicated by the returned
1159 : * *EndPos value. However, if we are already at the beginning of the current
1160 : * segment, *StartPos and *EndPos are set to the current location without
1161 : * reserving any space, and the function returns false.
1162 : */
1163 : static bool
1164 736 : ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1165 : {
1166 736 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1167 : uint64 startbytepos;
1168 : uint64 endbytepos;
1169 : uint64 prevbytepos;
1170 736 : uint32 size = MAXALIGN(SizeOfXLogRecord);
1171 : XLogRecPtr ptr;
1172 : uint32 segleft;
1173 :
1174 : /*
1175 : * These calculations are a bit heavy-weight to be done while holding a
1176 : * spinlock, but since we're holding all the WAL insertion locks, there
1177 : * are no other inserters competing for it. GetXLogInsertRecPtr() does
1178 : * compete for it, but that's not called very frequently.
1179 : */
1180 736 : SpinLockAcquire(&Insert->insertpos_lck);
1181 :
1182 736 : startbytepos = Insert->CurrBytePos;
1183 :
1184 736 : ptr = XLogBytePosToEndRecPtr(startbytepos);
1185 736 : if (XLogSegmentOffset(ptr, wal_segment_size) == 0)
1186 : {
1187 114 : SpinLockRelease(&Insert->insertpos_lck);
1188 114 : *EndPos = *StartPos = ptr;
1189 114 : return false;
1190 : }
1191 :
1192 622 : endbytepos = startbytepos + size;
1193 622 : prevbytepos = Insert->PrevBytePos;
1194 :
1195 622 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1196 622 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1197 :
1198 622 : segleft = wal_segment_size - XLogSegmentOffset(*EndPos, wal_segment_size);
1199 622 : if (segleft != wal_segment_size)
1200 : {
1201 : /* consume the rest of the segment */
1202 622 : *EndPos += segleft;
1203 622 : endbytepos = XLogRecPtrToBytePos(*EndPos);
1204 : }
1205 622 : Insert->CurrBytePos = endbytepos;
1206 622 : Insert->PrevBytePos = startbytepos;
1207 :
1208 622 : SpinLockRelease(&Insert->insertpos_lck);
1209 :
1210 622 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1211 :
1212 : Assert(XLogSegmentOffset(*EndPos, wal_segment_size) == 0);
1213 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1214 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1215 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1216 :
1217 622 : return true;
1218 : }
1219 :
1220 : /*
1221 : * Subroutine of XLogInsertRecord. Copies a WAL record to an already-reserved
1222 : * area in the WAL.
1223 : */
1224 : static void
1225 27685392 : CopyXLogRecordToWAL(int write_len, bool isLogSwitch, XLogRecData *rdata,
1226 : XLogRecPtr StartPos, XLogRecPtr EndPos, TimeLineID tli)
1227 : {
1228 : char *currpos;
1229 : int freespace;
1230 : int written;
1231 : XLogRecPtr CurrPos;
1232 : XLogPageHeader pagehdr;
1233 :
1234 : /*
1235 : * Get a pointer to the right place in the right WAL buffer to start
1236 : * inserting to.
1237 : */
1238 27685392 : CurrPos = StartPos;
1239 27685392 : currpos = GetXLogBuffer(CurrPos, tli);
1240 27685392 : freespace = INSERT_FREESPACE(CurrPos);
1241 :
1242 : /*
1243 : * there should be enough space for at least the first field (xl_tot_len)
1244 : * on this page.
1245 : */
1246 : Assert(freespace >= sizeof(uint32));
1247 :
1248 : /* Copy record data */
1249 27685392 : written = 0;
1250 130752096 : while (rdata != NULL)
1251 : {
1252 103066704 : const char *rdata_data = rdata->data;
1253 103066704 : int rdata_len = rdata->len;
1254 :
1255 106416900 : while (rdata_len > freespace)
1256 : {
1257 : /*
1258 : * Write what fits on this page, and continue on the next page.
1259 : */
1260 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || freespace == 0);
1261 3350196 : memcpy(currpos, rdata_data, freespace);
1262 3350196 : rdata_data += freespace;
1263 3350196 : rdata_len -= freespace;
1264 3350196 : written += freespace;
1265 3350196 : CurrPos += freespace;
1266 :
1267 : /*
1268 : * Get pointer to beginning of next page, and set the xlp_rem_len
1269 : * in the page header. Set XLP_FIRST_IS_CONTRECORD.
1270 : *
1271 : * It's safe to set the contrecord flag and xlp_rem_len without a
1272 : * lock on the page. All the other flags were already set when the
1273 : * page was initialized, in AdvanceXLInsertBuffer, and we're the
1274 : * only backend that needs to set the contrecord flag.
1275 : */
1276 3350196 : currpos = GetXLogBuffer(CurrPos, tli);
1277 3350196 : pagehdr = (XLogPageHeader) currpos;
1278 3350196 : pagehdr->xlp_rem_len = write_len - written;
1279 3350196 : pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1280 :
1281 : /* skip over the page header */
1282 3350196 : if (XLogSegmentOffset(CurrPos, wal_segment_size) == 0)
1283 : {
1284 2234 : CurrPos += SizeOfXLogLongPHD;
1285 2234 : currpos += SizeOfXLogLongPHD;
1286 : }
1287 : else
1288 : {
1289 3347962 : CurrPos += SizeOfXLogShortPHD;
1290 3347962 : currpos += SizeOfXLogShortPHD;
1291 : }
1292 3350196 : freespace = INSERT_FREESPACE(CurrPos);
1293 : }
1294 :
1295 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1296 103066704 : memcpy(currpos, rdata_data, rdata_len);
1297 103066704 : currpos += rdata_len;
1298 103066704 : CurrPos += rdata_len;
1299 103066704 : freespace -= rdata_len;
1300 103066704 : written += rdata_len;
1301 :
1302 103066704 : rdata = rdata->next;
1303 : }
1304 : Assert(written == write_len);
1305 :
1306 : /*
1307 : * If this was an xlog-switch, it's not enough to write the switch record,
1308 : * we also have to consume all the remaining space in the WAL segment. We
1309 : * have already reserved that space, but we need to actually fill it.
1310 : */
1311 27685392 : if (isLogSwitch && XLogSegmentOffset(CurrPos, wal_segment_size) != 0)
1312 : {
1313 : /* An xlog-switch record doesn't contain any data besides the header */
1314 : Assert(write_len == SizeOfXLogRecord);
1315 :
1316 : /* Assert that we did reserve the right amount of space */
1317 : Assert(XLogSegmentOffset(EndPos, wal_segment_size) == 0);
1318 :
1319 : /* Use up all the remaining space on the current page */
1320 622 : CurrPos += freespace;
1321 :
1322 : /*
1323 : * Cause all remaining pages in the segment to be flushed, leaving the
1324 : * XLog position where it should be, at the start of the next segment.
1325 : * We do this one page at a time, to make sure we don't deadlock
1326 : * against ourselves if wal_buffers < wal_segment_size.
1327 : */
1328 952576 : while (CurrPos < EndPos)
1329 : {
1330 : /*
1331 : * The minimal action to flush the page would be to call
1332 : * WALInsertLockUpdateInsertingAt(CurrPos) followed by
1333 : * AdvanceXLInsertBuffer(...). The page would be left initialized
1334 : * mostly to zeros, except for the page header (always the short
1335 : * variant, as this is never a segment's first page).
1336 : *
1337 : * The large vistas of zeros are good for compressibility, but the
1338 : * headers interrupting them every XLOG_BLCKSZ (with values that
1339 : * differ from page to page) are not. The effect varies with
1340 : * compression tool, but bzip2 for instance compresses about an
1341 : * order of magnitude worse if those headers are left in place.
1342 : *
1343 : * Rather than complicating AdvanceXLInsertBuffer itself (which is
1344 : * called in heavily-loaded circumstances as well as this lightly-
1345 : * loaded one) with variant behavior, we just use GetXLogBuffer
1346 : * (which itself calls the two methods we need) to get the pointer
1347 : * and zero most of the page. Then we just zero the page header.
1348 : */
1349 951954 : currpos = GetXLogBuffer(CurrPos, tli);
1350 3807816 : MemSet(currpos, 0, SizeOfXLogShortPHD);
1351 :
1352 951954 : CurrPos += XLOG_BLCKSZ;
1353 : }
1354 : }
1355 : else
1356 : {
1357 : /* Align the end position, so that the next record starts aligned */
1358 27684770 : CurrPos = MAXALIGN64(CurrPos);
1359 : }
1360 :
1361 27685392 : if (CurrPos != EndPos)
1362 0 : ereport(PANIC,
1363 : errcode(ERRCODE_DATA_CORRUPTED),
1364 : errmsg_internal("space reserved for WAL record does not match what was written"));
1365 27685392 : }
1366 :
1367 : /*
1368 : * Acquire a WAL insertion lock, for inserting to WAL.
1369 : */
1370 : static void
1371 27697446 : WALInsertLockAcquire(void)
1372 : {
1373 : bool immed;
1374 :
1375 : /*
1376 : * It doesn't matter which of the WAL insertion locks we acquire, so try
1377 : * the one we used last time. If the system isn't particularly busy, it's
1378 : * a good bet that it's still available, and it's good to have some
1379 : * affinity to a particular lock so that you don't unnecessarily bounce
1380 : * cache lines between processes when there's no contention.
1381 : *
1382 : * If this is the first time through in this backend, pick a lock
1383 : * (semi-)randomly. This allows the locks to be used evenly if you have a
1384 : * lot of very short connections.
1385 : */
1386 : static int lockToTry = -1;
1387 :
1388 27697446 : if (lockToTry == -1)
1389 14050 : lockToTry = MyProcNumber % NUM_XLOGINSERT_LOCKS;
1390 27697446 : MyLockNo = lockToTry;
1391 :
1392 : /*
1393 : * The insertingAt value is initially set to 0, as we don't know our
1394 : * insert location yet.
1395 : */
1396 27697446 : immed = LWLockAcquire(&WALInsertLocks[MyLockNo].l.lock, LW_EXCLUSIVE);
1397 27697446 : if (!immed)
1398 : {
1399 : /*
1400 : * If we couldn't get the lock immediately, try another lock next
1401 : * time. On a system with more insertion locks than concurrent
1402 : * inserters, this causes all the inserters to eventually migrate to a
1403 : * lock that no-one else is using. On a system with more inserters
1404 : * than locks, it still helps to distribute the inserters evenly
1405 : * across the locks.
1406 : */
1407 77176 : lockToTry = (lockToTry + 1) % NUM_XLOGINSERT_LOCKS;
1408 : }
1409 27697446 : }
1410 :
1411 : /*
1412 : * Acquire all WAL insertion locks, to prevent other backends from inserting
1413 : * to WAL.
1414 : */
1415 : static void
1416 6172 : WALInsertLockAcquireExclusive(void)
1417 : {
1418 : int i;
1419 :
1420 : /*
1421 : * When holding all the locks, all but the last lock's insertingAt
1422 : * indicator is set to 0xFFFFFFFFFFFFFFFF, which is higher than any real
1423 : * XLogRecPtr value, to make sure that no-one blocks waiting on those.
1424 : */
1425 49376 : for (i = 0; i < NUM_XLOGINSERT_LOCKS - 1; i++)
1426 : {
1427 43204 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1428 43204 : LWLockUpdateVar(&WALInsertLocks[i].l.lock,
1429 43204 : &WALInsertLocks[i].l.insertingAt,
1430 : PG_UINT64_MAX);
1431 : }
1432 : /* Variable value reset to 0 at release */
1433 6172 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1434 :
1435 6172 : holdingAllLocks = true;
1436 6172 : }
1437 :
1438 : /*
1439 : * Release our insertion lock (or locks, if we're holding them all).
1440 : *
1441 : * NB: Reset all variables to 0, so they cause LWLockWaitForVar to block the
1442 : * next time the lock is acquired.
1443 : */
1444 : static void
1445 27703618 : WALInsertLockRelease(void)
1446 : {
1447 27703618 : if (holdingAllLocks)
1448 : {
1449 : int i;
1450 :
1451 55548 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1452 49376 : LWLockReleaseClearVar(&WALInsertLocks[i].l.lock,
1453 49376 : &WALInsertLocks[i].l.insertingAt,
1454 : 0);
1455 :
1456 6172 : holdingAllLocks = false;
1457 : }
1458 : else
1459 : {
1460 27697446 : LWLockReleaseClearVar(&WALInsertLocks[MyLockNo].l.lock,
1461 27697446 : &WALInsertLocks[MyLockNo].l.insertingAt,
1462 : 0);
1463 : }
1464 27703618 : }
1465 :
1466 : /*
1467 : * Update our insertingAt value, to let others know that we've finished
1468 : * inserting up to that point.
1469 : */
1470 : static void
1471 4793618 : WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt)
1472 : {
1473 4793618 : if (holdingAllLocks)
1474 : {
1475 : /*
1476 : * We use the last lock to mark our actual position, see comments in
1477 : * WALInsertLockAcquireExclusive.
1478 : */
1479 947598 : LWLockUpdateVar(&WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.lock,
1480 947598 : &WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.insertingAt,
1481 : insertingAt);
1482 : }
1483 : else
1484 3846020 : LWLockUpdateVar(&WALInsertLocks[MyLockNo].l.lock,
1485 3846020 : &WALInsertLocks[MyLockNo].l.insertingAt,
1486 : insertingAt);
1487 4793618 : }
1488 :
1489 : /*
1490 : * Wait for any WAL insertions < upto to finish.
1491 : *
1492 : * Returns the location of the oldest insertion that is still in-progress.
1493 : * Any WAL prior to that point has been fully copied into WAL buffers, and
1494 : * can be flushed out to disk. Because this waits for any insertions older
1495 : * than 'upto' to finish, the return value is always >= 'upto'.
1496 : *
1497 : * Note: When you are about to write out WAL, you must call this function
1498 : * *before* acquiring WALWriteLock, to avoid deadlocks. This function might
1499 : * need to wait for an insertion to finish (or at least advance to next
1500 : * uninitialized page), and the inserter might need to evict an old WAL buffer
1501 : * to make room for a new one, which in turn requires WALWriteLock.
1502 : */
1503 : static XLogRecPtr
1504 4397024 : WaitXLogInsertionsToFinish(XLogRecPtr upto)
1505 : {
1506 : uint64 bytepos;
1507 : XLogRecPtr inserted;
1508 : XLogRecPtr reservedUpto;
1509 : XLogRecPtr finishedUpto;
1510 4397024 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1511 : int i;
1512 :
1513 4397024 : if (MyProc == NULL)
1514 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1515 :
1516 : /*
1517 : * Check if there's any work to do. Use a barrier to ensure we get the
1518 : * freshest value.
1519 : */
1520 4397024 : inserted = pg_atomic_read_membarrier_u64(&XLogCtl->logInsertResult);
1521 4397024 : if (upto <= inserted)
1522 3651554 : return inserted;
1523 :
1524 : /* Read the current insert position */
1525 745470 : SpinLockAcquire(&Insert->insertpos_lck);
1526 745470 : bytepos = Insert->CurrBytePos;
1527 745470 : SpinLockRelease(&Insert->insertpos_lck);
1528 745470 : reservedUpto = XLogBytePosToEndRecPtr(bytepos);
1529 :
1530 : /*
1531 : * No-one should request to flush a piece of WAL that hasn't even been
1532 : * reserved yet. However, it can happen if there is a block with a bogus
1533 : * LSN on disk, for example. XLogFlush checks for that situation and
1534 : * complains, but only after the flush. Here we just assume that to mean
1535 : * that all WAL that has been reserved needs to be finished. In this
1536 : * corner-case, the return value can be smaller than 'upto' argument.
1537 : */
1538 745470 : if (upto > reservedUpto)
1539 : {
1540 0 : ereport(LOG,
1541 : (errmsg("request to flush past end of generated WAL; request %X/%X, current position %X/%X",
1542 : LSN_FORMAT_ARGS(upto), LSN_FORMAT_ARGS(reservedUpto))));
1543 0 : upto = reservedUpto;
1544 : }
1545 :
1546 : /*
1547 : * Loop through all the locks, sleeping on any in-progress insert older
1548 : * than 'upto'.
1549 : *
1550 : * finishedUpto is our return value, indicating the point upto which all
1551 : * the WAL insertions have been finished. Initialize it to the head of
1552 : * reserved WAL, and as we iterate through the insertion locks, back it
1553 : * out for any insertion that's still in progress.
1554 : */
1555 745470 : finishedUpto = reservedUpto;
1556 6709230 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1557 : {
1558 5963760 : XLogRecPtr insertingat = InvalidXLogRecPtr;
1559 :
1560 : do
1561 : {
1562 : /*
1563 : * See if this insertion is in progress. LWLockWaitForVar will
1564 : * wait for the lock to be released, or for the 'value' to be set
1565 : * by a LWLockUpdateVar call. When a lock is initially acquired,
1566 : * its value is 0 (InvalidXLogRecPtr), which means that we don't
1567 : * know where it's inserting yet. We will have to wait for it. If
1568 : * it's a small insertion, the record will most likely fit on the
1569 : * same page and the inserter will release the lock without ever
1570 : * calling LWLockUpdateVar. But if it has to sleep, it will
1571 : * advertise the insertion point with LWLockUpdateVar before
1572 : * sleeping.
1573 : *
1574 : * In this loop we are only waiting for insertions that started
1575 : * before WaitXLogInsertionsToFinish was called. The lack of
1576 : * memory barriers in the loop means that we might see locks as
1577 : * "unused" that have since become used. This is fine because
1578 : * they only can be used for later insertions that we would not
1579 : * want to wait on anyway. Not taking a lock to acquire the
1580 : * current insertingAt value means that we might see older
1581 : * insertingAt values. This is also fine, because if we read a
1582 : * value too old, we will add ourselves to the wait queue, which
1583 : * contains atomic operations.
1584 : */
1585 6028180 : if (LWLockWaitForVar(&WALInsertLocks[i].l.lock,
1586 6028180 : &WALInsertLocks[i].l.insertingAt,
1587 : insertingat, &insertingat))
1588 : {
1589 : /* the lock was free, so no insertion in progress */
1590 4475130 : insertingat = InvalidXLogRecPtr;
1591 4475130 : break;
1592 : }
1593 :
1594 : /*
1595 : * This insertion is still in progress. Have to wait, unless the
1596 : * inserter has proceeded past 'upto'.
1597 : */
1598 1553050 : } while (insertingat < upto);
1599 :
1600 5963760 : if (insertingat != InvalidXLogRecPtr && insertingat < finishedUpto)
1601 609998 : finishedUpto = insertingat;
1602 : }
1603 :
1604 : /*
1605 : * Advance the limit we know to have been inserted and return the freshest
1606 : * value we know of, which might be beyond what we requested if somebody
1607 : * is concurrently doing this with an 'upto' pointer ahead of us.
1608 : */
1609 745470 : finishedUpto = pg_atomic_monotonic_advance_u64(&XLogCtl->logInsertResult,
1610 : finishedUpto);
1611 :
1612 745470 : return finishedUpto;
1613 : }
1614 :
1615 : /*
1616 : * Get a pointer to the right location in the WAL buffer containing the
1617 : * given XLogRecPtr.
1618 : *
1619 : * If the page is not initialized yet, it is initialized. That might require
1620 : * evicting an old dirty buffer from the buffer cache, which means I/O.
1621 : *
1622 : * The caller must ensure that the page containing the requested location
1623 : * isn't evicted yet, and won't be evicted. The way to ensure that is to
1624 : * hold onto a WAL insertion lock with the insertingAt position set to
1625 : * something <= ptr. GetXLogBuffer() will update insertingAt if it needs
1626 : * to evict an old page from the buffer. (This means that once you call
1627 : * GetXLogBuffer() with a given 'ptr', you must not access anything before
1628 : * that point anymore, and must not call GetXLogBuffer() with an older 'ptr'
1629 : * later, because older buffers might be recycled already)
1630 : */
1631 : static char *
1632 31987564 : GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli)
1633 : {
1634 : int idx;
1635 : XLogRecPtr endptr;
1636 : static uint64 cachedPage = 0;
1637 : static char *cachedPos = NULL;
1638 : XLogRecPtr expectedEndPtr;
1639 :
1640 : /*
1641 : * Fast path for the common case that we need to access again the same
1642 : * page as last time.
1643 : */
1644 31987564 : if (ptr / XLOG_BLCKSZ == cachedPage)
1645 : {
1646 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1647 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1648 26243522 : return cachedPos + ptr % XLOG_BLCKSZ;
1649 : }
1650 :
1651 : /*
1652 : * The XLog buffer cache is organized so that a page is always loaded to a
1653 : * particular buffer. That way we can easily calculate the buffer a given
1654 : * page must be loaded into, from the XLogRecPtr alone.
1655 : */
1656 5744042 : idx = XLogRecPtrToBufIdx(ptr);
1657 :
1658 : /*
1659 : * See what page is loaded in the buffer at the moment. It could be the
1660 : * page we're looking for, or something older. It can't be anything newer
1661 : * - that would imply the page we're looking for has already been written
1662 : * out to disk and evicted, and the caller is responsible for making sure
1663 : * that doesn't happen.
1664 : *
1665 : * We don't hold a lock while we read the value. If someone is just about
1666 : * to initialize or has just initialized the page, it's possible that we
1667 : * get InvalidXLogRecPtr. That's ok, we'll grab the mapping lock (in
1668 : * AdvanceXLInsertBuffer) and retry if we see anything other than the page
1669 : * we're looking for.
1670 : */
1671 5744042 : expectedEndPtr = ptr;
1672 5744042 : expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
1673 :
1674 5744042 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1675 5744042 : if (expectedEndPtr != endptr)
1676 : {
1677 : XLogRecPtr initializedUpto;
1678 :
1679 : /*
1680 : * Before calling AdvanceXLInsertBuffer(), which can block, let others
1681 : * know how far we're finished with inserting the record.
1682 : *
1683 : * NB: If 'ptr' points to just after the page header, advertise a
1684 : * position at the beginning of the page rather than 'ptr' itself. If
1685 : * there are no other insertions running, someone might try to flush
1686 : * up to our advertised location. If we advertised a position after
1687 : * the page header, someone might try to flush the page header, even
1688 : * though page might actually not be initialized yet. As the first
1689 : * inserter on the page, we are effectively responsible for making
1690 : * sure that it's initialized, before we let insertingAt to move past
1691 : * the page header.
1692 : */
1693 4793618 : if (ptr % XLOG_BLCKSZ == SizeOfXLogShortPHD &&
1694 11844 : XLogSegmentOffset(ptr, wal_segment_size) > XLOG_BLCKSZ)
1695 11844 : initializedUpto = ptr - SizeOfXLogShortPHD;
1696 4781774 : else if (ptr % XLOG_BLCKSZ == SizeOfXLogLongPHD &&
1697 1680 : XLogSegmentOffset(ptr, wal_segment_size) < XLOG_BLCKSZ)
1698 398 : initializedUpto = ptr - SizeOfXLogLongPHD;
1699 : else
1700 4781376 : initializedUpto = ptr;
1701 :
1702 4793618 : WALInsertLockUpdateInsertingAt(initializedUpto);
1703 :
1704 4793618 : AdvanceXLInsertBuffer(ptr, tli, false);
1705 4793618 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1706 :
1707 4793618 : if (expectedEndPtr != endptr)
1708 0 : elog(PANIC, "could not find WAL buffer for %X/%X",
1709 : LSN_FORMAT_ARGS(ptr));
1710 : }
1711 : else
1712 : {
1713 : /*
1714 : * Make sure the initialization of the page is visible to us, and
1715 : * won't arrive later to overwrite the WAL data we write on the page.
1716 : */
1717 950424 : pg_memory_barrier();
1718 : }
1719 :
1720 : /*
1721 : * Found the buffer holding this page. Return a pointer to the right
1722 : * offset within the page.
1723 : */
1724 5744042 : cachedPage = ptr / XLOG_BLCKSZ;
1725 5744042 : cachedPos = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1726 :
1727 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1728 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1729 :
1730 5744042 : return cachedPos + ptr % XLOG_BLCKSZ;
1731 : }
1732 :
1733 : /*
1734 : * Read WAL data directly from WAL buffers, if available. Returns the number
1735 : * of bytes read successfully.
1736 : *
1737 : * Fewer than 'count' bytes may be read if some of the requested WAL data has
1738 : * already been evicted.
1739 : *
1740 : * No locks are taken.
1741 : *
1742 : * Caller should ensure that it reads no further than LogwrtResult.Write
1743 : * (which should have been updated by the caller when determining how far to
1744 : * read). The 'tli' argument is only used as a convenient safety check so that
1745 : * callers do not read from WAL buffers on a historical timeline.
1746 : */
1747 : Size
1748 185052 : WALReadFromBuffers(char *dstbuf, XLogRecPtr startptr, Size count,
1749 : TimeLineID tli)
1750 : {
1751 185052 : char *pdst = dstbuf;
1752 185052 : XLogRecPtr recptr = startptr;
1753 : XLogRecPtr inserted;
1754 185052 : Size nbytes = count;
1755 :
1756 185052 : if (RecoveryInProgress() || tli != GetWALInsertionTimeLine())
1757 1824 : return 0;
1758 :
1759 : Assert(!XLogRecPtrIsInvalid(startptr));
1760 :
1761 : /*
1762 : * Caller should ensure that the requested data has been inserted into WAL
1763 : * buffers before we try to read it.
1764 : */
1765 183228 : inserted = pg_atomic_read_u64(&XLogCtl->logInsertResult);
1766 183228 : if (startptr + count > inserted)
1767 0 : ereport(ERROR,
1768 : errmsg("cannot read past end of generated WAL: requested %X/%X, current position %X/%X",
1769 : LSN_FORMAT_ARGS(startptr + count),
1770 : LSN_FORMAT_ARGS(inserted)));
1771 :
1772 : /*
1773 : * Loop through the buffers without a lock. For each buffer, atomically
1774 : * read and verify the end pointer, then copy the data out, and finally
1775 : * re-read and re-verify the end pointer.
1776 : *
1777 : * Once a page is evicted, it never returns to the WAL buffers, so if the
1778 : * end pointer matches the expected end pointer before and after we copy
1779 : * the data, then the right page must have been present during the data
1780 : * copy. Read barriers are necessary to ensure that the data copy actually
1781 : * happens between the two verification steps.
1782 : *
1783 : * If either verification fails, we simply terminate the loop and return
1784 : * with the data that had been already copied out successfully.
1785 : */
1786 197074 : while (nbytes > 0)
1787 : {
1788 193158 : uint32 offset = recptr % XLOG_BLCKSZ;
1789 193158 : int idx = XLogRecPtrToBufIdx(recptr);
1790 : XLogRecPtr expectedEndPtr;
1791 : XLogRecPtr endptr;
1792 : const char *page;
1793 : const char *psrc;
1794 : Size npagebytes;
1795 :
1796 : /*
1797 : * Calculate the end pointer we expect in the xlblocks array if the
1798 : * correct page is present.
1799 : */
1800 193158 : expectedEndPtr = recptr + (XLOG_BLCKSZ - offset);
1801 :
1802 : /*
1803 : * First verification step: check that the correct page is present in
1804 : * the WAL buffers.
1805 : */
1806 193158 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1807 193158 : if (expectedEndPtr != endptr)
1808 179308 : break;
1809 :
1810 : /*
1811 : * The correct page is present (or was at the time the endptr was
1812 : * read; must re-verify later). Calculate pointer to source data and
1813 : * determine how much data to read from this page.
1814 : */
1815 13850 : page = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1816 13850 : psrc = page + offset;
1817 13850 : npagebytes = Min(nbytes, XLOG_BLCKSZ - offset);
1818 :
1819 : /*
1820 : * Ensure that the data copy and the first verification step are not
1821 : * reordered.
1822 : */
1823 13850 : pg_read_barrier();
1824 :
1825 : /* data copy */
1826 13850 : memcpy(pdst, psrc, npagebytes);
1827 :
1828 : /*
1829 : * Ensure that the data copy and the second verification step are not
1830 : * reordered.
1831 : */
1832 13850 : pg_read_barrier();
1833 :
1834 : /*
1835 : * Second verification step: check that the page we read from wasn't
1836 : * evicted while we were copying the data.
1837 : */
1838 13850 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1839 13850 : if (expectedEndPtr != endptr)
1840 4 : break;
1841 :
1842 13846 : pdst += npagebytes;
1843 13846 : recptr += npagebytes;
1844 13846 : nbytes -= npagebytes;
1845 : }
1846 :
1847 : Assert(pdst - dstbuf <= count);
1848 :
1849 183228 : return pdst - dstbuf;
1850 : }
1851 :
1852 : /*
1853 : * Converts a "usable byte position" to XLogRecPtr. A usable byte position
1854 : * is the position starting from the beginning of WAL, excluding all WAL
1855 : * page headers.
1856 : */
1857 : static XLogRecPtr
1858 55375776 : XLogBytePosToRecPtr(uint64 bytepos)
1859 : {
1860 : uint64 fullsegs;
1861 : uint64 fullpages;
1862 : uint64 bytesleft;
1863 : uint32 seg_offset;
1864 : XLogRecPtr result;
1865 :
1866 55375776 : fullsegs = bytepos / UsableBytesInSegment;
1867 55375776 : bytesleft = bytepos % UsableBytesInSegment;
1868 :
1869 55375776 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1870 : {
1871 : /* fits on first page of segment */
1872 90326 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1873 : }
1874 : else
1875 : {
1876 : /* account for the first page on segment with long header */
1877 55285450 : seg_offset = XLOG_BLCKSZ;
1878 55285450 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1879 :
1880 55285450 : fullpages = bytesleft / UsableBytesInPage;
1881 55285450 : bytesleft = bytesleft % UsableBytesInPage;
1882 :
1883 55285450 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1884 : }
1885 :
1886 55375776 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1887 :
1888 55375776 : return result;
1889 : }
1890 :
1891 : /*
1892 : * Like XLogBytePosToRecPtr, but if the position is at a page boundary,
1893 : * returns a pointer to the beginning of the page (ie. before page header),
1894 : * not to where the first xlog record on that page would go to. This is used
1895 : * when converting a pointer to the end of a record.
1896 : */
1897 : static XLogRecPtr
1898 28431598 : XLogBytePosToEndRecPtr(uint64 bytepos)
1899 : {
1900 : uint64 fullsegs;
1901 : uint64 fullpages;
1902 : uint64 bytesleft;
1903 : uint32 seg_offset;
1904 : XLogRecPtr result;
1905 :
1906 28431598 : fullsegs = bytepos / UsableBytesInSegment;
1907 28431598 : bytesleft = bytepos % UsableBytesInSegment;
1908 :
1909 28431598 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1910 : {
1911 : /* fits on first page of segment */
1912 140668 : if (bytesleft == 0)
1913 94392 : seg_offset = 0;
1914 : else
1915 46276 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1916 : }
1917 : else
1918 : {
1919 : /* account for the first page on segment with long header */
1920 28290930 : seg_offset = XLOG_BLCKSZ;
1921 28290930 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1922 :
1923 28290930 : fullpages = bytesleft / UsableBytesInPage;
1924 28290930 : bytesleft = bytesleft % UsableBytesInPage;
1925 :
1926 28290930 : if (bytesleft == 0)
1927 27508 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
1928 : else
1929 28263422 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1930 : }
1931 :
1932 28431598 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1933 :
1934 28431598 : return result;
1935 : }
1936 :
1937 : /*
1938 : * Convert an XLogRecPtr to a "usable byte position".
1939 : */
1940 : static uint64
1941 3690 : XLogRecPtrToBytePos(XLogRecPtr ptr)
1942 : {
1943 : uint64 fullsegs;
1944 : uint32 fullpages;
1945 : uint32 offset;
1946 : uint64 result;
1947 :
1948 3690 : XLByteToSeg(ptr, fullsegs, wal_segment_size);
1949 :
1950 3690 : fullpages = (XLogSegmentOffset(ptr, wal_segment_size)) / XLOG_BLCKSZ;
1951 3690 : offset = ptr % XLOG_BLCKSZ;
1952 :
1953 3690 : if (fullpages == 0)
1954 : {
1955 1060 : result = fullsegs * UsableBytesInSegment;
1956 1060 : if (offset > 0)
1957 : {
1958 : Assert(offset >= SizeOfXLogLongPHD);
1959 404 : result += offset - SizeOfXLogLongPHD;
1960 : }
1961 : }
1962 : else
1963 : {
1964 2630 : result = fullsegs * UsableBytesInSegment +
1965 2630 : (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
1966 2630 : (fullpages - 1) * UsableBytesInPage; /* full pages */
1967 2630 : if (offset > 0)
1968 : {
1969 : Assert(offset >= SizeOfXLogShortPHD);
1970 2608 : result += offset - SizeOfXLogShortPHD;
1971 : }
1972 : }
1973 :
1974 3690 : return result;
1975 : }
1976 :
1977 : /*
1978 : * Initialize XLOG buffers, writing out old buffers if they still contain
1979 : * unwritten data, upto the page containing 'upto'. Or if 'opportunistic' is
1980 : * true, initialize as many pages as we can without having to write out
1981 : * unwritten data. Any new pages are initialized to zeros, with pages headers
1982 : * initialized properly.
1983 : */
1984 : static void
1985 4800562 : AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
1986 : {
1987 4800562 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1988 : int nextidx;
1989 : XLogRecPtr OldPageRqstPtr;
1990 : XLogwrtRqst WriteRqst;
1991 4800562 : XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
1992 : XLogRecPtr NewPageBeginPtr;
1993 : XLogPageHeader NewPage;
1994 4800562 : int npages pg_attribute_unused() = 0;
1995 :
1996 4800562 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
1997 :
1998 : /*
1999 : * Now that we have the lock, check if someone initialized the page
2000 : * already.
2001 : */
2002 13345090 : while (upto >= XLogCtl->InitializedUpTo || opportunistic)
2003 : {
2004 8551472 : nextidx = XLogRecPtrToBufIdx(XLogCtl->InitializedUpTo);
2005 :
2006 : /*
2007 : * Get ending-offset of the buffer page we need to replace (this may
2008 : * be zero if the buffer hasn't been used yet). Fall through if it's
2009 : * already written out.
2010 : */
2011 8551472 : OldPageRqstPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]);
2012 8551472 : if (LogwrtResult.Write < OldPageRqstPtr)
2013 : {
2014 : /*
2015 : * Nope, got work to do. If we just want to pre-initialize as much
2016 : * as we can without flushing, give up now.
2017 : */
2018 4326768 : if (opportunistic)
2019 6944 : break;
2020 :
2021 : /* Advance shared memory write request position */
2022 4319824 : SpinLockAcquire(&XLogCtl->info_lck);
2023 4319824 : if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
2024 899330 : XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
2025 4319824 : SpinLockRelease(&XLogCtl->info_lck);
2026 :
2027 : /*
2028 : * Acquire an up-to-date LogwrtResult value and see if we still
2029 : * need to write it or if someone else already did.
2030 : */
2031 4319824 : RefreshXLogWriteResult(LogwrtResult);
2032 4319824 : if (LogwrtResult.Write < OldPageRqstPtr)
2033 : {
2034 : /*
2035 : * Must acquire write lock. Release WALBufMappingLock first,
2036 : * to make sure that all insertions that we need to wait for
2037 : * can finish (up to this same position). Otherwise we risk
2038 : * deadlock.
2039 : */
2040 4144126 : LWLockRelease(WALBufMappingLock);
2041 :
2042 4144126 : WaitXLogInsertionsToFinish(OldPageRqstPtr);
2043 :
2044 4144126 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2045 :
2046 4144126 : RefreshXLogWriteResult(LogwrtResult);
2047 4144126 : if (LogwrtResult.Write >= OldPageRqstPtr)
2048 : {
2049 : /* OK, someone wrote it already */
2050 664218 : LWLockRelease(WALWriteLock);
2051 : }
2052 : else
2053 : {
2054 : /* Have to write it ourselves */
2055 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2056 3479908 : WriteRqst.Write = OldPageRqstPtr;
2057 3479908 : WriteRqst.Flush = 0;
2058 3479908 : XLogWrite(WriteRqst, tli, false);
2059 3479908 : LWLockRelease(WALWriteLock);
2060 3479908 : PendingWalStats.wal_buffers_full++;
2061 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_DONE();
2062 : }
2063 : /* Re-acquire WALBufMappingLock and retry */
2064 4144126 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2065 4144126 : continue;
2066 : }
2067 : }
2068 :
2069 : /*
2070 : * Now the next buffer slot is free and we can set it up to be the
2071 : * next output page.
2072 : */
2073 4400402 : NewPageBeginPtr = XLogCtl->InitializedUpTo;
2074 4400402 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2075 :
2076 : Assert(XLogRecPtrToBufIdx(NewPageBeginPtr) == nextidx);
2077 :
2078 4400402 : NewPage = (XLogPageHeader) (XLogCtl->pages + nextidx * (Size) XLOG_BLCKSZ);
2079 :
2080 : /*
2081 : * Mark the xlblock with InvalidXLogRecPtr and issue a write barrier
2082 : * before initializing. Otherwise, the old page may be partially
2083 : * zeroed but look valid.
2084 : */
2085 4400402 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], InvalidXLogRecPtr);
2086 4400402 : pg_write_barrier();
2087 :
2088 : /*
2089 : * Be sure to re-zero the buffer so that bytes beyond what we've
2090 : * written will look like zeroes and not valid XLOG records...
2091 : */
2092 4400402 : MemSet((char *) NewPage, 0, XLOG_BLCKSZ);
2093 :
2094 : /*
2095 : * Fill the new page's header
2096 : */
2097 4400402 : NewPage->xlp_magic = XLOG_PAGE_MAGIC;
2098 :
2099 : /* NewPage->xlp_info = 0; */ /* done by memset */
2100 4400402 : NewPage->xlp_tli = tli;
2101 4400402 : NewPage->xlp_pageaddr = NewPageBeginPtr;
2102 :
2103 : /* NewPage->xlp_rem_len = 0; */ /* done by memset */
2104 :
2105 : /*
2106 : * If online backup is not in progress, mark the header to indicate
2107 : * that WAL records beginning in this page have removable backup
2108 : * blocks. This allows the WAL archiver to know whether it is safe to
2109 : * compress archived WAL data by transforming full-block records into
2110 : * the non-full-block format. It is sufficient to record this at the
2111 : * page level because we force a page switch (in fact a segment
2112 : * switch) when starting a backup, so the flag will be off before any
2113 : * records can be written during the backup. At the end of a backup,
2114 : * the last page will be marked as all unsafe when perhaps only part
2115 : * is unsafe, but at worst the archiver would miss the opportunity to
2116 : * compress a few records.
2117 : */
2118 4400402 : if (Insert->runningBackups == 0)
2119 4147138 : NewPage->xlp_info |= XLP_BKP_REMOVABLE;
2120 :
2121 : /*
2122 : * If first page of an XLOG segment file, make it a long header.
2123 : */
2124 4400402 : if ((XLogSegmentOffset(NewPage->xlp_pageaddr, wal_segment_size)) == 0)
2125 : {
2126 2646 : XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2127 :
2128 2646 : NewLongPage->xlp_sysid = ControlFile->system_identifier;
2129 2646 : NewLongPage->xlp_seg_size = wal_segment_size;
2130 2646 : NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2131 2646 : NewPage->xlp_info |= XLP_LONG_HEADER;
2132 : }
2133 :
2134 : /*
2135 : * Make sure the initialization of the page becomes visible to others
2136 : * before the xlblocks update. GetXLogBuffer() reads xlblocks without
2137 : * holding a lock.
2138 : */
2139 4400402 : pg_write_barrier();
2140 :
2141 4400402 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], NewPageEndPtr);
2142 4400402 : XLogCtl->InitializedUpTo = NewPageEndPtr;
2143 :
2144 4400402 : npages++;
2145 : }
2146 4800562 : LWLockRelease(WALBufMappingLock);
2147 :
2148 : #ifdef WAL_DEBUG
2149 : if (XLOG_DEBUG && npages > 0)
2150 : {
2151 : elog(DEBUG1, "initialized %d pages, up to %X/%X",
2152 : npages, LSN_FORMAT_ARGS(NewPageEndPtr));
2153 : }
2154 : #endif
2155 4800562 : }
2156 :
2157 : /*
2158 : * Calculate CheckPointSegments based on max_wal_size_mb and
2159 : * checkpoint_completion_target.
2160 : */
2161 : static void
2162 13678 : CalculateCheckpointSegments(void)
2163 : {
2164 : double target;
2165 :
2166 : /*-------
2167 : * Calculate the distance at which to trigger a checkpoint, to avoid
2168 : * exceeding max_wal_size_mb. This is based on two assumptions:
2169 : *
2170 : * a) we keep WAL for only one checkpoint cycle (prior to PG11 we kept
2171 : * WAL for two checkpoint cycles to allow us to recover from the
2172 : * secondary checkpoint if the first checkpoint failed, though we
2173 : * only did this on the primary anyway, not on standby. Keeping just
2174 : * one checkpoint simplifies processing and reduces disk space in
2175 : * many smaller databases.)
2176 : * b) during checkpoint, we consume checkpoint_completion_target *
2177 : * number of segments consumed between checkpoints.
2178 : *-------
2179 : */
2180 13678 : target = (double) ConvertToXSegs(max_wal_size_mb, wal_segment_size) /
2181 13678 : (1.0 + CheckPointCompletionTarget);
2182 :
2183 : /* round down */
2184 13678 : CheckPointSegments = (int) target;
2185 :
2186 13678 : if (CheckPointSegments < 1)
2187 18 : CheckPointSegments = 1;
2188 13678 : }
2189 :
2190 : void
2191 9962 : assign_max_wal_size(int newval, void *extra)
2192 : {
2193 9962 : max_wal_size_mb = newval;
2194 9962 : CalculateCheckpointSegments();
2195 9962 : }
2196 :
2197 : void
2198 1972 : assign_checkpoint_completion_target(double newval, void *extra)
2199 : {
2200 1972 : CheckPointCompletionTarget = newval;
2201 1972 : CalculateCheckpointSegments();
2202 1972 : }
2203 :
2204 : bool
2205 3808 : check_wal_segment_size(int *newval, void **extra, GucSource source)
2206 : {
2207 3808 : if (!IsValidWalSegSize(*newval))
2208 : {
2209 0 : GUC_check_errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.");
2210 0 : return false;
2211 : }
2212 :
2213 3808 : return true;
2214 : }
2215 :
2216 : /*
2217 : * GUC check_hook for max_slot_wal_keep_size
2218 : *
2219 : * We don't allow the value of max_slot_wal_keep_size other than -1 during the
2220 : * binary upgrade. See start_postmaster() in pg_upgrade for more details.
2221 : */
2222 : bool
2223 2180 : check_max_slot_wal_keep_size(int *newval, void **extra, GucSource source)
2224 : {
2225 2180 : if (IsBinaryUpgrade && *newval != -1)
2226 : {
2227 0 : GUC_check_errdetail("\"%s\" must be set to -1 during binary upgrade mode.",
2228 : "max_slot_wal_keep_size");
2229 0 : return false;
2230 : }
2231 :
2232 2180 : return true;
2233 : }
2234 :
2235 : /*
2236 : * At a checkpoint, how many WAL segments to recycle as preallocated future
2237 : * XLOG segments? Returns the highest segment that should be preallocated.
2238 : */
2239 : static XLogSegNo
2240 2468 : XLOGfileslop(XLogRecPtr lastredoptr)
2241 : {
2242 : XLogSegNo minSegNo;
2243 : XLogSegNo maxSegNo;
2244 : double distance;
2245 : XLogSegNo recycleSegNo;
2246 :
2247 : /*
2248 : * Calculate the segment numbers that min_wal_size_mb and max_wal_size_mb
2249 : * correspond to. Always recycle enough segments to meet the minimum, and
2250 : * remove enough segments to stay below the maximum.
2251 : */
2252 2468 : minSegNo = lastredoptr / wal_segment_size +
2253 2468 : ConvertToXSegs(min_wal_size_mb, wal_segment_size) - 1;
2254 2468 : maxSegNo = lastredoptr / wal_segment_size +
2255 2468 : ConvertToXSegs(max_wal_size_mb, wal_segment_size) - 1;
2256 :
2257 : /*
2258 : * Between those limits, recycle enough segments to get us through to the
2259 : * estimated end of next checkpoint.
2260 : *
2261 : * To estimate where the next checkpoint will finish, assume that the
2262 : * system runs steadily consuming CheckPointDistanceEstimate bytes between
2263 : * every checkpoint.
2264 : */
2265 2468 : distance = (1.0 + CheckPointCompletionTarget) * CheckPointDistanceEstimate;
2266 : /* add 10% for good measure. */
2267 2468 : distance *= 1.10;
2268 :
2269 2468 : recycleSegNo = (XLogSegNo) ceil(((double) lastredoptr + distance) /
2270 : wal_segment_size);
2271 :
2272 2468 : if (recycleSegNo < minSegNo)
2273 1532 : recycleSegNo = minSegNo;
2274 2468 : if (recycleSegNo > maxSegNo)
2275 730 : recycleSegNo = maxSegNo;
2276 :
2277 2468 : return recycleSegNo;
2278 : }
2279 :
2280 : /*
2281 : * Check whether we've consumed enough xlog space that a checkpoint is needed.
2282 : *
2283 : * new_segno indicates a log file that has just been filled up (or read
2284 : * during recovery). We measure the distance from RedoRecPtr to new_segno
2285 : * and see if that exceeds CheckPointSegments.
2286 : *
2287 : * Note: it is caller's responsibility that RedoRecPtr is up-to-date.
2288 : */
2289 : bool
2290 7870 : XLogCheckpointNeeded(XLogSegNo new_segno)
2291 : {
2292 : XLogSegNo old_segno;
2293 :
2294 7870 : XLByteToSeg(RedoRecPtr, old_segno, wal_segment_size);
2295 :
2296 7870 : if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2297 5232 : return true;
2298 2638 : return false;
2299 : }
2300 :
2301 : /*
2302 : * Write and/or fsync the log at least as far as WriteRqst indicates.
2303 : *
2304 : * If flexible == true, we don't have to write as far as WriteRqst, but
2305 : * may stop at any convenient boundary (such as a cache or logfile boundary).
2306 : * This option allows us to avoid uselessly issuing multiple writes when a
2307 : * single one would do.
2308 : *
2309 : * Must be called with WALWriteLock held. WaitXLogInsertionsToFinish(WriteRqst)
2310 : * must be called before grabbing the lock, to make sure the data is ready to
2311 : * write.
2312 : */
2313 : static void
2314 3720022 : XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible)
2315 : {
2316 : bool ispartialpage;
2317 : bool last_iteration;
2318 : bool finishing_seg;
2319 : int curridx;
2320 : int npages;
2321 : int startidx;
2322 : uint32 startoffset;
2323 :
2324 : /* We should always be inside a critical section here */
2325 : Assert(CritSectionCount > 0);
2326 :
2327 : /*
2328 : * Update local LogwrtResult (caller probably did this already, but...)
2329 : */
2330 3720022 : RefreshXLogWriteResult(LogwrtResult);
2331 :
2332 : /*
2333 : * Since successive pages in the xlog cache are consecutively allocated,
2334 : * we can usually gather multiple pages together and issue just one
2335 : * write() call. npages is the number of pages we have determined can be
2336 : * written together; startidx is the cache block index of the first one,
2337 : * and startoffset is the file offset at which it should go. The latter
2338 : * two variables are only valid when npages > 0, but we must initialize
2339 : * all of them to keep the compiler quiet.
2340 : */
2341 3720022 : npages = 0;
2342 3720022 : startidx = 0;
2343 3720022 : startoffset = 0;
2344 :
2345 : /*
2346 : * Within the loop, curridx is the cache block index of the page to
2347 : * consider writing. Begin at the buffer containing the next unwritten
2348 : * page, or last partially written page.
2349 : */
2350 3720022 : curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2351 :
2352 8027428 : while (LogwrtResult.Write < WriteRqst.Write)
2353 : {
2354 : /*
2355 : * Make sure we're not ahead of the insert process. This could happen
2356 : * if we're passed a bogus WriteRqst.Write that is past the end of the
2357 : * last page that's been initialized by AdvanceXLInsertBuffer.
2358 : */
2359 4541756 : XLogRecPtr EndPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[curridx]);
2360 :
2361 4541756 : if (LogwrtResult.Write >= EndPtr)
2362 0 : elog(PANIC, "xlog write request %X/%X is past end of log %X/%X",
2363 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2364 : LSN_FORMAT_ARGS(EndPtr));
2365 :
2366 : /* Advance LogwrtResult.Write to end of current buffer page */
2367 4541756 : LogwrtResult.Write = EndPtr;
2368 4541756 : ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2369 :
2370 4541756 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2371 : wal_segment_size))
2372 : {
2373 : /*
2374 : * Switch to new logfile segment. We cannot have any pending
2375 : * pages here (since we dump what we have at segment end).
2376 : */
2377 : Assert(npages == 0);
2378 22062 : if (openLogFile >= 0)
2379 9062 : XLogFileClose();
2380 22062 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2381 : wal_segment_size);
2382 22062 : openLogTLI = tli;
2383 :
2384 : /* create/use new log file */
2385 22062 : openLogFile = XLogFileInit(openLogSegNo, tli);
2386 22062 : ReserveExternalFD();
2387 : }
2388 :
2389 : /* Make sure we have the current logfile open */
2390 4541756 : if (openLogFile < 0)
2391 : {
2392 0 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2393 : wal_segment_size);
2394 0 : openLogTLI = tli;
2395 0 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2396 0 : ReserveExternalFD();
2397 : }
2398 :
2399 : /* Add current page to the set of pending pages-to-dump */
2400 4541756 : if (npages == 0)
2401 : {
2402 : /* first of group */
2403 3749354 : startidx = curridx;
2404 3749354 : startoffset = XLogSegmentOffset(LogwrtResult.Write - XLOG_BLCKSZ,
2405 : wal_segment_size);
2406 : }
2407 4541756 : npages++;
2408 :
2409 : /*
2410 : * Dump the set if this will be the last loop iteration, or if we are
2411 : * at the last page of the cache area (since the next page won't be
2412 : * contiguous in memory), or if we are at the end of the logfile
2413 : * segment.
2414 : */
2415 4541756 : last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2416 :
2417 8854978 : finishing_seg = !ispartialpage &&
2418 4313222 : (startoffset + npages * XLOG_BLCKSZ) >= wal_segment_size;
2419 :
2420 4541756 : if (last_iteration ||
2421 822742 : curridx == XLogCtl->XLogCacheBlck ||
2422 : finishing_seg)
2423 : {
2424 : char *from;
2425 : Size nbytes;
2426 : Size nleft;
2427 : ssize_t written;
2428 : instr_time start;
2429 :
2430 : /* OK to write the page(s) */
2431 3749354 : from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2432 3749354 : nbytes = npages * (Size) XLOG_BLCKSZ;
2433 3749354 : nleft = nbytes;
2434 : do
2435 : {
2436 3749354 : errno = 0;
2437 :
2438 : /* Measure I/O timing to write WAL data */
2439 3749354 : if (track_wal_io_timing)
2440 0 : INSTR_TIME_SET_CURRENT(start);
2441 : else
2442 3749354 : INSTR_TIME_SET_ZERO(start);
2443 :
2444 3749354 : pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE);
2445 3749354 : written = pg_pwrite(openLogFile, from, nleft, startoffset);
2446 3749354 : pgstat_report_wait_end();
2447 :
2448 : /*
2449 : * Increment the I/O timing and the number of times WAL data
2450 : * were written out to disk.
2451 : */
2452 3749354 : if (track_wal_io_timing)
2453 : {
2454 : instr_time end;
2455 :
2456 0 : INSTR_TIME_SET_CURRENT(end);
2457 0 : INSTR_TIME_ACCUM_DIFF(PendingWalStats.wal_write_time, end, start);
2458 : }
2459 :
2460 3749354 : PendingWalStats.wal_write++;
2461 :
2462 3749354 : if (written <= 0)
2463 : {
2464 : char xlogfname[MAXFNAMELEN];
2465 : int save_errno;
2466 :
2467 0 : if (errno == EINTR)
2468 0 : continue;
2469 :
2470 0 : save_errno = errno;
2471 0 : XLogFileName(xlogfname, tli, openLogSegNo,
2472 : wal_segment_size);
2473 0 : errno = save_errno;
2474 0 : ereport(PANIC,
2475 : (errcode_for_file_access(),
2476 : errmsg("could not write to log file \"%s\" at offset %u, length %zu: %m",
2477 : xlogfname, startoffset, nleft)));
2478 : }
2479 3749354 : nleft -= written;
2480 3749354 : from += written;
2481 3749354 : startoffset += written;
2482 3749354 : } while (nleft > 0);
2483 :
2484 3749354 : npages = 0;
2485 :
2486 : /*
2487 : * If we just wrote the whole last page of a logfile segment,
2488 : * fsync the segment immediately. This avoids having to go back
2489 : * and re-open prior segments when an fsync request comes along
2490 : * later. Doing it here ensures that one and only one backend will
2491 : * perform this fsync.
2492 : *
2493 : * This is also the right place to notify the Archiver that the
2494 : * segment is ready to copy to archival storage, and to update the
2495 : * timer for archive_timeout, and to signal for a checkpoint if
2496 : * too many logfile segments have been used since the last
2497 : * checkpoint.
2498 : */
2499 3749354 : if (finishing_seg)
2500 : {
2501 2872 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2502 :
2503 : /* signal that we need to wakeup walsenders later */
2504 2872 : WalSndWakeupRequest();
2505 :
2506 2872 : LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2507 :
2508 2872 : if (XLogArchivingActive())
2509 156 : XLogArchiveNotifySeg(openLogSegNo, tli);
2510 :
2511 2872 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2512 2872 : XLogCtl->lastSegSwitchLSN = LogwrtResult.Flush;
2513 :
2514 : /*
2515 : * Request a checkpoint if we've consumed too much xlog since
2516 : * the last one. For speed, we first check using the local
2517 : * copy of RedoRecPtr, which might be out of date; if it looks
2518 : * like a checkpoint is needed, forcibly update RedoRecPtr and
2519 : * recheck.
2520 : */
2521 2872 : if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2522 : {
2523 466 : (void) GetRedoRecPtr();
2524 466 : if (XLogCheckpointNeeded(openLogSegNo))
2525 390 : RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2526 : }
2527 : }
2528 : }
2529 :
2530 4541756 : if (ispartialpage)
2531 : {
2532 : /* Only asked to write a partial page */
2533 228534 : LogwrtResult.Write = WriteRqst.Write;
2534 228534 : break;
2535 : }
2536 4313222 : curridx = NextBufIdx(curridx);
2537 :
2538 : /* If flexible, break out of loop as soon as we wrote something */
2539 4313222 : if (flexible && npages == 0)
2540 5816 : break;
2541 : }
2542 :
2543 : Assert(npages == 0);
2544 :
2545 : /*
2546 : * If asked to flush, do so
2547 : */
2548 3720022 : if (LogwrtResult.Flush < WriteRqst.Flush &&
2549 239450 : LogwrtResult.Flush < LogwrtResult.Write)
2550 : {
2551 : /*
2552 : * Could get here without iterating above loop, in which case we might
2553 : * have no open file or the wrong one. However, we do not need to
2554 : * fsync more than one file.
2555 : */
2556 239352 : if (wal_sync_method != WAL_SYNC_METHOD_OPEN &&
2557 239352 : wal_sync_method != WAL_SYNC_METHOD_OPEN_DSYNC)
2558 : {
2559 239352 : if (openLogFile >= 0 &&
2560 239332 : !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2561 : wal_segment_size))
2562 42 : XLogFileClose();
2563 239352 : if (openLogFile < 0)
2564 : {
2565 62 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2566 : wal_segment_size);
2567 62 : openLogTLI = tli;
2568 62 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2569 62 : ReserveExternalFD();
2570 : }
2571 :
2572 239352 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2573 : }
2574 :
2575 : /* signal that we need to wakeup walsenders later */
2576 239352 : WalSndWakeupRequest();
2577 :
2578 239352 : LogwrtResult.Flush = LogwrtResult.Write;
2579 : }
2580 :
2581 : /*
2582 : * Update shared-memory status
2583 : *
2584 : * We make sure that the shared 'request' values do not fall behind the
2585 : * 'result' values. This is not absolutely essential, but it saves some
2586 : * code in a couple of places.
2587 : */
2588 3720022 : SpinLockAcquire(&XLogCtl->info_lck);
2589 3720022 : if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
2590 212368 : XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
2591 3720022 : if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
2592 241508 : XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
2593 3720022 : SpinLockRelease(&XLogCtl->info_lck);
2594 :
2595 : /*
2596 : * We write Write first, bar, then Flush. When reading, the opposite must
2597 : * be done (with a matching barrier in between), so that we always see a
2598 : * Flush value that trails behind the Write value seen.
2599 : */
2600 3720022 : pg_atomic_write_u64(&XLogCtl->logWriteResult, LogwrtResult.Write);
2601 3720022 : pg_write_barrier();
2602 3720022 : pg_atomic_write_u64(&XLogCtl->logFlushResult, LogwrtResult.Flush);
2603 :
2604 : #ifdef USE_ASSERT_CHECKING
2605 : {
2606 : XLogRecPtr Flush;
2607 : XLogRecPtr Write;
2608 : XLogRecPtr Insert;
2609 :
2610 : Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult);
2611 : pg_read_barrier();
2612 : Write = pg_atomic_read_u64(&XLogCtl->logWriteResult);
2613 : pg_read_barrier();
2614 : Insert = pg_atomic_read_u64(&XLogCtl->logInsertResult);
2615 :
2616 : /* WAL written to disk is always ahead of WAL flushed */
2617 : Assert(Write >= Flush);
2618 :
2619 : /* WAL inserted to buffers is always ahead of WAL written */
2620 : Assert(Insert >= Write);
2621 : }
2622 : #endif
2623 3720022 : }
2624 :
2625 : /*
2626 : * Record the LSN for an asynchronous transaction commit/abort
2627 : * and nudge the WALWriter if there is work for it to do.
2628 : * (This should not be called for synchronous commits.)
2629 : */
2630 : void
2631 83322 : XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2632 : {
2633 83322 : XLogRecPtr WriteRqstPtr = asyncXactLSN;
2634 : bool sleeping;
2635 83322 : bool wakeup = false;
2636 : XLogRecPtr prevAsyncXactLSN;
2637 :
2638 83322 : SpinLockAcquire(&XLogCtl->info_lck);
2639 83322 : sleeping = XLogCtl->WalWriterSleeping;
2640 83322 : prevAsyncXactLSN = XLogCtl->asyncXactLSN;
2641 83322 : if (XLogCtl->asyncXactLSN < asyncXactLSN)
2642 82550 : XLogCtl->asyncXactLSN = asyncXactLSN;
2643 83322 : SpinLockRelease(&XLogCtl->info_lck);
2644 :
2645 : /*
2646 : * If somebody else already called this function with a more aggressive
2647 : * LSN, they will have done what we needed (and perhaps more).
2648 : */
2649 83322 : if (asyncXactLSN <= prevAsyncXactLSN)
2650 772 : return;
2651 :
2652 : /*
2653 : * If the WALWriter is sleeping, kick it to make it come out of low-power
2654 : * mode, so that this async commit will reach disk within the expected
2655 : * amount of time. Otherwise, determine whether it has enough WAL
2656 : * available to flush, the same way that XLogBackgroundFlush() does.
2657 : */
2658 82550 : if (sleeping)
2659 22 : wakeup = true;
2660 : else
2661 : {
2662 : int flushblocks;
2663 :
2664 82528 : RefreshXLogWriteResult(LogwrtResult);
2665 :
2666 82528 : flushblocks =
2667 82528 : WriteRqstPtr / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
2668 :
2669 82528 : if (WalWriterFlushAfter == 0 || flushblocks >= WalWriterFlushAfter)
2670 6774 : wakeup = true;
2671 : }
2672 :
2673 82550 : if (wakeup)
2674 : {
2675 6796 : volatile PROC_HDR *procglobal = ProcGlobal;
2676 6796 : ProcNumber walwriterProc = procglobal->walwriterProc;
2677 :
2678 6796 : if (walwriterProc != INVALID_PROC_NUMBER)
2679 386 : SetLatch(&GetPGProcByNumber(walwriterProc)->procLatch);
2680 : }
2681 : }
2682 :
2683 : /*
2684 : * Record the LSN up to which we can remove WAL because it's not required by
2685 : * any replication slot.
2686 : */
2687 : void
2688 17024 : XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn)
2689 : {
2690 17024 : SpinLockAcquire(&XLogCtl->info_lck);
2691 17024 : XLogCtl->replicationSlotMinLSN = lsn;
2692 17024 : SpinLockRelease(&XLogCtl->info_lck);
2693 17024 : }
2694 :
2695 :
2696 : /*
2697 : * Return the oldest LSN we must retain to satisfy the needs of some
2698 : * replication slot.
2699 : */
2700 : static XLogRecPtr
2701 3298 : XLogGetReplicationSlotMinimumLSN(void)
2702 : {
2703 : XLogRecPtr retval;
2704 :
2705 3298 : SpinLockAcquire(&XLogCtl->info_lck);
2706 3298 : retval = XLogCtl->replicationSlotMinLSN;
2707 3298 : SpinLockRelease(&XLogCtl->info_lck);
2708 :
2709 3298 : return retval;
2710 : }
2711 :
2712 : /*
2713 : * Advance minRecoveryPoint in control file.
2714 : *
2715 : * If we crash during recovery, we must reach this point again before the
2716 : * database is consistent.
2717 : *
2718 : * If 'force' is true, 'lsn' argument is ignored. Otherwise, minRecoveryPoint
2719 : * is only updated if it's not already greater than or equal to 'lsn'.
2720 : */
2721 : static void
2722 182636 : UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2723 : {
2724 : /* Quick check using our local copy of the variable */
2725 182636 : if (!updateMinRecoveryPoint || (!force && lsn <= LocalMinRecoveryPoint))
2726 171262 : return;
2727 :
2728 : /*
2729 : * An invalid minRecoveryPoint means that we need to recover all the WAL,
2730 : * i.e., we're doing crash recovery. We never modify the control file's
2731 : * value in that case, so we can short-circuit future checks here too. The
2732 : * local values of minRecoveryPoint and minRecoveryPointTLI should not be
2733 : * updated until crash recovery finishes. We only do this for the startup
2734 : * process as it should not update its own reference of minRecoveryPoint
2735 : * until it has finished crash recovery to make sure that all WAL
2736 : * available is replayed in this case. This also saves from extra locks
2737 : * taken on the control file from the startup process.
2738 : */
2739 11374 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint) && InRecovery)
2740 : {
2741 62 : updateMinRecoveryPoint = false;
2742 62 : return;
2743 : }
2744 :
2745 11312 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2746 :
2747 : /* update local copy */
2748 11312 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
2749 11312 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2750 :
2751 11312 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint))
2752 4 : updateMinRecoveryPoint = false;
2753 11308 : else if (force || LocalMinRecoveryPoint < lsn)
2754 : {
2755 : XLogRecPtr newMinRecoveryPoint;
2756 : TimeLineID newMinRecoveryPointTLI;
2757 :
2758 : /*
2759 : * To avoid having to update the control file too often, we update it
2760 : * all the way to the last record being replayed, even though 'lsn'
2761 : * would suffice for correctness. This also allows the 'force' case
2762 : * to not need a valid 'lsn' value.
2763 : *
2764 : * Another important reason for doing it this way is that the passed
2765 : * 'lsn' value could be bogus, i.e., past the end of available WAL, if
2766 : * the caller got it from a corrupted heap page. Accepting such a
2767 : * value as the min recovery point would prevent us from coming up at
2768 : * all. Instead, we just log a warning and continue with recovery.
2769 : * (See also the comments about corrupt LSNs in XLogFlush.)
2770 : */
2771 9302 : newMinRecoveryPoint = GetCurrentReplayRecPtr(&newMinRecoveryPointTLI);
2772 9302 : if (!force && newMinRecoveryPoint < lsn)
2773 0 : elog(WARNING,
2774 : "xlog min recovery request %X/%X is past current point %X/%X",
2775 : LSN_FORMAT_ARGS(lsn), LSN_FORMAT_ARGS(newMinRecoveryPoint));
2776 :
2777 : /* update control file */
2778 9302 : if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2779 : {
2780 9250 : ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2781 9250 : ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2782 9250 : UpdateControlFile();
2783 9250 : LocalMinRecoveryPoint = newMinRecoveryPoint;
2784 9250 : LocalMinRecoveryPointTLI = newMinRecoveryPointTLI;
2785 :
2786 9250 : ereport(DEBUG2,
2787 : (errmsg_internal("updated min recovery point to %X/%X on timeline %u",
2788 : LSN_FORMAT_ARGS(newMinRecoveryPoint),
2789 : newMinRecoveryPointTLI)));
2790 : }
2791 : }
2792 11312 : LWLockRelease(ControlFileLock);
2793 : }
2794 :
2795 : /*
2796 : * Ensure that all XLOG data through the given position is flushed to disk.
2797 : *
2798 : * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
2799 : * already held, and we try to avoid acquiring it if possible.
2800 : */
2801 : void
2802 1185964 : XLogFlush(XLogRecPtr record)
2803 : {
2804 : XLogRecPtr WriteRqstPtr;
2805 : XLogwrtRqst WriteRqst;
2806 1185964 : TimeLineID insertTLI = XLogCtl->InsertTimeLineID;
2807 :
2808 : /*
2809 : * During REDO, we are reading not writing WAL. Therefore, instead of
2810 : * trying to flush the WAL, we should update minRecoveryPoint instead. We
2811 : * test XLogInsertAllowed(), not InRecovery, because we need checkpointer
2812 : * to act this way too, and because when it tries to write the
2813 : * end-of-recovery checkpoint, it should indeed flush.
2814 : */
2815 1185964 : if (!XLogInsertAllowed())
2816 : {
2817 182384 : UpdateMinRecoveryPoint(record, false);
2818 933100 : return;
2819 : }
2820 :
2821 : /* Quick exit if already known flushed */
2822 1003580 : if (record <= LogwrtResult.Flush)
2823 750716 : return;
2824 :
2825 : #ifdef WAL_DEBUG
2826 : if (XLOG_DEBUG)
2827 : elog(LOG, "xlog flush request %X/%X; write %X/%X; flush %X/%X",
2828 : LSN_FORMAT_ARGS(record),
2829 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2830 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2831 : #endif
2832 :
2833 252864 : START_CRIT_SECTION();
2834 :
2835 : /*
2836 : * Since fsync is usually a horribly expensive operation, we try to
2837 : * piggyback as much data as we can on each fsync: if we see any more data
2838 : * entered into the xlog buffer, we'll write and fsync that too, so that
2839 : * the final value of LogwrtResult.Flush is as large as possible. This
2840 : * gives us some chance of avoiding another fsync immediately after.
2841 : */
2842 :
2843 : /* initialize to given target; may increase below */
2844 252864 : WriteRqstPtr = record;
2845 :
2846 : /*
2847 : * Now wait until we get the write lock, or someone else does the flush
2848 : * for us.
2849 : */
2850 : for (;;)
2851 10994 : {
2852 : XLogRecPtr insertpos;
2853 :
2854 : /* done already? */
2855 263858 : RefreshXLogWriteResult(LogwrtResult);
2856 263858 : if (record <= LogwrtResult.Flush)
2857 17904 : break;
2858 :
2859 : /*
2860 : * Before actually performing the write, wait for all in-flight
2861 : * insertions to the pages we're about to write to finish.
2862 : */
2863 245954 : SpinLockAcquire(&XLogCtl->info_lck);
2864 245954 : if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
2865 14962 : WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
2866 245954 : SpinLockRelease(&XLogCtl->info_lck);
2867 245954 : insertpos = WaitXLogInsertionsToFinish(WriteRqstPtr);
2868 :
2869 : /*
2870 : * Try to get the write lock. If we can't get it immediately, wait
2871 : * until it's released, and recheck if we still need to do the flush
2872 : * or if the backend that held the lock did it for us already. This
2873 : * helps to maintain a good rate of group committing when the system
2874 : * is bottlenecked by the speed of fsyncing.
2875 : */
2876 245954 : if (!LWLockAcquireOrWait(WALWriteLock, LW_EXCLUSIVE))
2877 : {
2878 : /*
2879 : * The lock is now free, but we didn't acquire it yet. Before we
2880 : * do, loop back to check if someone else flushed the record for
2881 : * us already.
2882 : */
2883 10994 : continue;
2884 : }
2885 :
2886 : /* Got the lock; recheck whether request is satisfied */
2887 234960 : RefreshXLogWriteResult(LogwrtResult);
2888 234960 : if (record <= LogwrtResult.Flush)
2889 : {
2890 1570 : LWLockRelease(WALWriteLock);
2891 1570 : break;
2892 : }
2893 :
2894 : /*
2895 : * Sleep before flush! By adding a delay here, we may give further
2896 : * backends the opportunity to join the backlog of group commit
2897 : * followers; this can significantly improve transaction throughput,
2898 : * at the risk of increasing transaction latency.
2899 : *
2900 : * We do not sleep if enableFsync is not turned on, nor if there are
2901 : * fewer than CommitSiblings other backends with active transactions.
2902 : */
2903 233390 : if (CommitDelay > 0 && enableFsync &&
2904 0 : MinimumActiveBackends(CommitSiblings))
2905 : {
2906 0 : pg_usleep(CommitDelay);
2907 :
2908 : /*
2909 : * Re-check how far we can now flush the WAL. It's generally not
2910 : * safe to call WaitXLogInsertionsToFinish while holding
2911 : * WALWriteLock, because an in-progress insertion might need to
2912 : * also grab WALWriteLock to make progress. But we know that all
2913 : * the insertions up to insertpos have already finished, because
2914 : * that's what the earlier WaitXLogInsertionsToFinish() returned.
2915 : * We're only calling it again to allow insertpos to be moved
2916 : * further forward, not to actually wait for anyone.
2917 : */
2918 0 : insertpos = WaitXLogInsertionsToFinish(insertpos);
2919 : }
2920 :
2921 : /* try to write/flush later additions to XLOG as well */
2922 233390 : WriteRqst.Write = insertpos;
2923 233390 : WriteRqst.Flush = insertpos;
2924 :
2925 233390 : XLogWrite(WriteRqst, insertTLI, false);
2926 :
2927 233390 : LWLockRelease(WALWriteLock);
2928 : /* done */
2929 233390 : break;
2930 : }
2931 :
2932 252864 : END_CRIT_SECTION();
2933 :
2934 : /* wake up walsenders now that we've released heavily contended locks */
2935 252864 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
2936 :
2937 : /*
2938 : * If we still haven't flushed to the request point then we have a
2939 : * problem; most likely, the requested flush point is past end of XLOG.
2940 : * This has been seen to occur when a disk page has a corrupted LSN.
2941 : *
2942 : * Formerly we treated this as a PANIC condition, but that hurts the
2943 : * system's robustness rather than helping it: we do not want to take down
2944 : * the whole system due to corruption on one data page. In particular, if
2945 : * the bad page is encountered again during recovery then we would be
2946 : * unable to restart the database at all! (This scenario actually
2947 : * happened in the field several times with 7.1 releases.) As of 8.4, bad
2948 : * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
2949 : * the only time we can reach here during recovery is while flushing the
2950 : * end-of-recovery checkpoint record, and we don't expect that to have a
2951 : * bad LSN.
2952 : *
2953 : * Note that for calls from xact.c, the ERROR will be promoted to PANIC
2954 : * since xact.c calls this routine inside a critical section. However,
2955 : * calls from bufmgr.c are not within critical sections and so we will not
2956 : * force a restart for a bad LSN on a data page.
2957 : */
2958 252864 : if (LogwrtResult.Flush < record)
2959 0 : elog(ERROR,
2960 : "xlog flush request %X/%X is not satisfied --- flushed only to %X/%X",
2961 : LSN_FORMAT_ARGS(record),
2962 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2963 : }
2964 :
2965 : /*
2966 : * Write & flush xlog, but without specifying exactly where to.
2967 : *
2968 : * We normally write only completed blocks; but if there is nothing to do on
2969 : * that basis, we check for unwritten async commits in the current incomplete
2970 : * block, and write through the latest one of those. Thus, if async commits
2971 : * are not being used, we will write complete blocks only.
2972 : *
2973 : * If, based on the above, there's anything to write we do so immediately. But
2974 : * to avoid calling fsync, fdatasync et. al. at a rate that'd impact
2975 : * concurrent IO, we only flush WAL every wal_writer_delay ms, or if there's
2976 : * more than wal_writer_flush_after unflushed blocks.
2977 : *
2978 : * We can guarantee that async commits reach disk after at most three
2979 : * wal_writer_delay cycles. (When flushing complete blocks, we allow XLogWrite
2980 : * to write "flexibly", meaning it can stop at the end of the buffer ring;
2981 : * this makes a difference only with very high load or long wal_writer_delay,
2982 : * but imposes one extra cycle for the worst case for async commits.)
2983 : *
2984 : * This routine is invoked periodically by the background walwriter process.
2985 : *
2986 : * Returns true if there was any work to do, even if we skipped flushing due
2987 : * to wal_writer_delay/wal_writer_flush_after.
2988 : */
2989 : bool
2990 38730 : XLogBackgroundFlush(void)
2991 : {
2992 : XLogwrtRqst WriteRqst;
2993 38730 : bool flexible = true;
2994 : static TimestampTz lastflush;
2995 : TimestampTz now;
2996 : int flushblocks;
2997 : TimeLineID insertTLI;
2998 :
2999 : /* XLOG doesn't need flushing during recovery */
3000 38730 : if (RecoveryInProgress())
3001 16 : return false;
3002 :
3003 : /*
3004 : * Since we're not in recovery, InsertTimeLineID is set and can't change,
3005 : * so we can read it without a lock.
3006 : */
3007 38714 : insertTLI = XLogCtl->InsertTimeLineID;
3008 :
3009 : /* read updated LogwrtRqst */
3010 38714 : SpinLockAcquire(&XLogCtl->info_lck);
3011 38714 : WriteRqst = XLogCtl->LogwrtRqst;
3012 38714 : SpinLockRelease(&XLogCtl->info_lck);
3013 :
3014 : /* back off to last completed page boundary */
3015 38714 : WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
3016 :
3017 : /* if we have already flushed that far, consider async commit records */
3018 38714 : RefreshXLogWriteResult(LogwrtResult);
3019 38714 : if (WriteRqst.Write <= LogwrtResult.Flush)
3020 : {
3021 32644 : SpinLockAcquire(&XLogCtl->info_lck);
3022 32644 : WriteRqst.Write = XLogCtl->asyncXactLSN;
3023 32644 : SpinLockRelease(&XLogCtl->info_lck);
3024 32644 : flexible = false; /* ensure it all gets written */
3025 : }
3026 :
3027 : /*
3028 : * If already known flushed, we're done. Just need to check if we are
3029 : * holding an open file handle to a logfile that's no longer in use,
3030 : * preventing the file from being deleted.
3031 : */
3032 38714 : if (WriteRqst.Write <= LogwrtResult.Flush)
3033 : {
3034 31770 : if (openLogFile >= 0)
3035 : {
3036 6532 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
3037 : wal_segment_size))
3038 : {
3039 244 : XLogFileClose();
3040 : }
3041 : }
3042 31770 : return false;
3043 : }
3044 :
3045 : /*
3046 : * Determine how far to flush WAL, based on the wal_writer_delay and
3047 : * wal_writer_flush_after GUCs.
3048 : *
3049 : * Note that XLogSetAsyncXactLSN() performs similar calculation based on
3050 : * wal_writer_flush_after, to decide when to wake us up. Make sure the
3051 : * logic is the same in both places if you change this.
3052 : */
3053 6944 : now = GetCurrentTimestamp();
3054 6944 : flushblocks =
3055 6944 : WriteRqst.Write / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
3056 :
3057 6944 : if (WalWriterFlushAfter == 0 || lastflush == 0)
3058 : {
3059 : /* first call, or block based limits disabled */
3060 368 : WriteRqst.Flush = WriteRqst.Write;
3061 368 : lastflush = now;
3062 : }
3063 6576 : else if (TimestampDifferenceExceeds(lastflush, now, WalWriterDelay))
3064 : {
3065 : /*
3066 : * Flush the writes at least every WalWriterDelay ms. This is
3067 : * important to bound the amount of time it takes for an asynchronous
3068 : * commit to hit disk.
3069 : */
3070 6248 : WriteRqst.Flush = WriteRqst.Write;
3071 6248 : lastflush = now;
3072 : }
3073 328 : else if (flushblocks >= WalWriterFlushAfter)
3074 : {
3075 : /* exceeded wal_writer_flush_after blocks, flush */
3076 306 : WriteRqst.Flush = WriteRqst.Write;
3077 306 : lastflush = now;
3078 : }
3079 : else
3080 : {
3081 : /* no flushing, this time round */
3082 22 : WriteRqst.Flush = 0;
3083 : }
3084 :
3085 : #ifdef WAL_DEBUG
3086 : if (XLOG_DEBUG)
3087 : elog(LOG, "xlog bg flush request write %X/%X; flush: %X/%X, current is write %X/%X; flush %X/%X",
3088 : LSN_FORMAT_ARGS(WriteRqst.Write),
3089 : LSN_FORMAT_ARGS(WriteRqst.Flush),
3090 : LSN_FORMAT_ARGS(LogwrtResult.Write),
3091 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
3092 : #endif
3093 :
3094 6944 : START_CRIT_SECTION();
3095 :
3096 : /* now wait for any in-progress insertions to finish and get write lock */
3097 6944 : WaitXLogInsertionsToFinish(WriteRqst.Write);
3098 6944 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3099 6944 : RefreshXLogWriteResult(LogwrtResult);
3100 6944 : if (WriteRqst.Write > LogwrtResult.Write ||
3101 272 : WriteRqst.Flush > LogwrtResult.Flush)
3102 : {
3103 6724 : XLogWrite(WriteRqst, insertTLI, flexible);
3104 : }
3105 6944 : LWLockRelease(WALWriteLock);
3106 :
3107 6944 : END_CRIT_SECTION();
3108 :
3109 : /* wake up walsenders now that we've released heavily contended locks */
3110 6944 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3111 :
3112 : /*
3113 : * Great, done. To take some work off the critical path, try to initialize
3114 : * as many of the no-longer-needed WAL buffers for future use as we can.
3115 : */
3116 6944 : AdvanceXLInsertBuffer(InvalidXLogRecPtr, insertTLI, true);
3117 :
3118 : /*
3119 : * If we determined that we need to write data, but somebody else
3120 : * wrote/flushed already, it should be considered as being active, to
3121 : * avoid hibernating too early.
3122 : */
3123 6944 : return true;
3124 : }
3125 :
3126 : /*
3127 : * Test whether XLOG data has been flushed up to (at least) the given position.
3128 : *
3129 : * Returns true if a flush is still needed. (It may be that someone else
3130 : * is already in process of flushing that far, however.)
3131 : */
3132 : bool
3133 17076810 : XLogNeedsFlush(XLogRecPtr record)
3134 : {
3135 : /*
3136 : * During recovery, we don't flush WAL but update minRecoveryPoint
3137 : * instead. So "needs flush" is taken to mean whether minRecoveryPoint
3138 : * would need to be updated.
3139 : */
3140 17076810 : if (RecoveryInProgress())
3141 : {
3142 : /*
3143 : * An invalid minRecoveryPoint means that we need to recover all the
3144 : * WAL, i.e., we're doing crash recovery. We never modify the control
3145 : * file's value in that case, so we can short-circuit future checks
3146 : * here too. This triggers a quick exit path for the startup process,
3147 : * which cannot update its local copy of minRecoveryPoint as long as
3148 : * it has not replayed all WAL available when doing crash recovery.
3149 : */
3150 1638464 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint) && InRecovery)
3151 0 : updateMinRecoveryPoint = false;
3152 :
3153 : /* Quick exit if already known to be updated or cannot be updated */
3154 1638464 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3155 1615738 : return false;
3156 :
3157 : /*
3158 : * Update local copy of minRecoveryPoint. But if the lock is busy,
3159 : * just return a conservative guess.
3160 : */
3161 22726 : if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3162 0 : return true;
3163 22726 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
3164 22726 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3165 22726 : LWLockRelease(ControlFileLock);
3166 :
3167 : /*
3168 : * Check minRecoveryPoint for any other process than the startup
3169 : * process doing crash recovery, which should not update the control
3170 : * file value if crash recovery is still running.
3171 : */
3172 22726 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint))
3173 0 : updateMinRecoveryPoint = false;
3174 :
3175 : /* check again */
3176 22726 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3177 152 : return false;
3178 : else
3179 22574 : return true;
3180 : }
3181 :
3182 : /* Quick exit if already known flushed */
3183 15438346 : if (record <= LogwrtResult.Flush)
3184 15002192 : return false;
3185 :
3186 : /* read LogwrtResult and update local state */
3187 436154 : RefreshXLogWriteResult(LogwrtResult);
3188 :
3189 : /* check again */
3190 436154 : if (record <= LogwrtResult.Flush)
3191 4104 : return false;
3192 :
3193 432050 : return true;
3194 : }
3195 :
3196 : /*
3197 : * Try to make a given XLOG file segment exist.
3198 : *
3199 : * logsegno: identify segment.
3200 : *
3201 : * *added: on return, true if this call raised the number of extant segments.
3202 : *
3203 : * path: on return, this char[MAXPGPATH] has the path to the logsegno file.
3204 : *
3205 : * Returns -1 or FD of opened file. A -1 here is not an error; a caller
3206 : * wanting an open segment should attempt to open "path", which usually will
3207 : * succeed. (This is weird, but it's efficient for the callers.)
3208 : */
3209 : static int
3210 24260 : XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,
3211 : bool *added, char *path)
3212 : {
3213 : char tmppath[MAXPGPATH];
3214 : XLogSegNo installed_segno;
3215 : XLogSegNo max_segno;
3216 : int fd;
3217 : int save_errno;
3218 24260 : int open_flags = O_RDWR | O_CREAT | O_EXCL | PG_BINARY;
3219 :
3220 : Assert(logtli != 0);
3221 :
3222 24260 : XLogFilePath(path, logtli, logsegno, wal_segment_size);
3223 :
3224 : /*
3225 : * Try to use existent file (checkpoint maker may have created it already)
3226 : */
3227 24260 : *added = false;
3228 24260 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3229 24260 : get_sync_bit(wal_sync_method));
3230 24260 : if (fd < 0)
3231 : {
3232 2570 : if (errno != ENOENT)
3233 0 : ereport(ERROR,
3234 : (errcode_for_file_access(),
3235 : errmsg("could not open file \"%s\": %m", path)));
3236 : }
3237 : else
3238 21690 : return fd;
3239 :
3240 : /*
3241 : * Initialize an empty (all zeroes) segment. NOTE: it is possible that
3242 : * another process is doing the same thing. If so, we will end up
3243 : * pre-creating an extra log segment. That seems OK, and better than
3244 : * holding the lock throughout this lengthy process.
3245 : */
3246 2570 : elog(DEBUG2, "creating and filling new WAL file");
3247 :
3248 2570 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3249 :
3250 2570 : unlink(tmppath);
3251 :
3252 2570 : if (io_direct_flags & IO_DIRECT_WAL_INIT)
3253 0 : open_flags |= PG_O_DIRECT;
3254 :
3255 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3256 2570 : fd = BasicOpenFile(tmppath, open_flags);
3257 2570 : if (fd < 0)
3258 0 : ereport(ERROR,
3259 : (errcode_for_file_access(),
3260 : errmsg("could not create file \"%s\": %m", tmppath)));
3261 :
3262 2570 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE);
3263 2570 : save_errno = 0;
3264 2570 : if (wal_init_zero)
3265 : {
3266 : ssize_t rc;
3267 :
3268 : /*
3269 : * Zero-fill the file. With this setting, we do this the hard way to
3270 : * ensure that all the file space has really been allocated. On
3271 : * platforms that allow "holes" in files, just seeking to the end
3272 : * doesn't allocate intermediate space. This way, we know that we
3273 : * have all the space and (after the fsync below) that all the
3274 : * indirect blocks are down on disk. Therefore, fdatasync(2) or
3275 : * O_DSYNC will be sufficient to sync future writes to the log file.
3276 : */
3277 2570 : rc = pg_pwrite_zeros(fd, wal_segment_size, 0);
3278 :
3279 2570 : if (rc < 0)
3280 0 : save_errno = errno;
3281 : }
3282 : else
3283 : {
3284 : /*
3285 : * Otherwise, seeking to the end and writing a solitary byte is
3286 : * enough.
3287 : */
3288 0 : errno = 0;
3289 0 : if (pg_pwrite(fd, "\0", 1, wal_segment_size - 1) != 1)
3290 : {
3291 : /* if write didn't set errno, assume no disk space */
3292 0 : save_errno = errno ? errno : ENOSPC;
3293 : }
3294 : }
3295 2570 : pgstat_report_wait_end();
3296 :
3297 2570 : if (save_errno)
3298 : {
3299 : /*
3300 : * If we fail to make the file, delete it to release disk space
3301 : */
3302 0 : unlink(tmppath);
3303 :
3304 0 : close(fd);
3305 :
3306 0 : errno = save_errno;
3307 :
3308 0 : ereport(ERROR,
3309 : (errcode_for_file_access(),
3310 : errmsg("could not write to file \"%s\": %m", tmppath)));
3311 : }
3312 :
3313 2570 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC);
3314 2570 : if (pg_fsync(fd) != 0)
3315 : {
3316 0 : save_errno = errno;
3317 0 : close(fd);
3318 0 : errno = save_errno;
3319 0 : ereport(ERROR,
3320 : (errcode_for_file_access(),
3321 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3322 : }
3323 2570 : pgstat_report_wait_end();
3324 :
3325 2570 : if (close(fd) != 0)
3326 0 : ereport(ERROR,
3327 : (errcode_for_file_access(),
3328 : errmsg("could not close file \"%s\": %m", tmppath)));
3329 :
3330 : /*
3331 : * Now move the segment into place with its final name. Cope with
3332 : * possibility that someone else has created the file while we were
3333 : * filling ours: if so, use ours to pre-create a future log segment.
3334 : */
3335 2570 : installed_segno = logsegno;
3336 :
3337 : /*
3338 : * XXX: What should we use as max_segno? We used to use XLOGfileslop when
3339 : * that was a constant, but that was always a bit dubious: normally, at a
3340 : * checkpoint, XLOGfileslop was the offset from the checkpoint record, but
3341 : * here, it was the offset from the insert location. We can't do the
3342 : * normal XLOGfileslop calculation here because we don't have access to
3343 : * the prior checkpoint's redo location. So somewhat arbitrarily, just use
3344 : * CheckPointSegments.
3345 : */
3346 2570 : max_segno = logsegno + CheckPointSegments;
3347 2570 : if (InstallXLogFileSegment(&installed_segno, tmppath, true, max_segno,
3348 : logtli))
3349 : {
3350 2568 : *added = true;
3351 2568 : elog(DEBUG2, "done creating and filling new WAL file");
3352 : }
3353 : else
3354 : {
3355 : /*
3356 : * No need for any more future segments, or InstallXLogFileSegment()
3357 : * failed to rename the file into place. If the rename failed, a
3358 : * caller opening the file may fail.
3359 : */
3360 2 : unlink(tmppath);
3361 2 : elog(DEBUG2, "abandoned new WAL file");
3362 : }
3363 :
3364 2570 : return -1;
3365 : }
3366 :
3367 : /*
3368 : * Create a new XLOG file segment, or open a pre-existing one.
3369 : *
3370 : * logsegno: identify segment to be created/opened.
3371 : *
3372 : * Returns FD of opened file.
3373 : *
3374 : * Note: errors here are ERROR not PANIC because we might or might not be
3375 : * inside a critical section (eg, during checkpoint there is no reason to
3376 : * take down the system on failure). They will promote to PANIC if we are
3377 : * in a critical section.
3378 : */
3379 : int
3380 23764 : XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)
3381 : {
3382 : bool ignore_added;
3383 : char path[MAXPGPATH];
3384 : int fd;
3385 :
3386 : Assert(logtli != 0);
3387 :
3388 23764 : fd = XLogFileInitInternal(logsegno, logtli, &ignore_added, path);
3389 23764 : if (fd >= 0)
3390 21494 : return fd;
3391 :
3392 : /* Now open original target segment (might not be file I just made) */
3393 2270 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3394 2270 : get_sync_bit(wal_sync_method));
3395 2270 : if (fd < 0)
3396 0 : ereport(ERROR,
3397 : (errcode_for_file_access(),
3398 : errmsg("could not open file \"%s\": %m", path)));
3399 2270 : return fd;
3400 : }
3401 :
3402 : /*
3403 : * Create a new XLOG file segment by copying a pre-existing one.
3404 : *
3405 : * destsegno: identify segment to be created.
3406 : *
3407 : * srcTLI, srcsegno: identify segment to be copied (could be from
3408 : * a different timeline)
3409 : *
3410 : * upto: how much of the source file to copy (the rest is filled with
3411 : * zeros)
3412 : *
3413 : * Currently this is only used during recovery, and so there are no locking
3414 : * considerations. But we should be just as tense as XLogFileInit to avoid
3415 : * emplacing a bogus file.
3416 : */
3417 : static void
3418 78 : XLogFileCopy(TimeLineID destTLI, XLogSegNo destsegno,
3419 : TimeLineID srcTLI, XLogSegNo srcsegno,
3420 : int upto)
3421 : {
3422 : char path[MAXPGPATH];
3423 : char tmppath[MAXPGPATH];
3424 : PGAlignedXLogBlock buffer;
3425 : int srcfd;
3426 : int fd;
3427 : int nbytes;
3428 :
3429 : /*
3430 : * Open the source file
3431 : */
3432 78 : XLogFilePath(path, srcTLI, srcsegno, wal_segment_size);
3433 78 : srcfd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
3434 78 : if (srcfd < 0)
3435 0 : ereport(ERROR,
3436 : (errcode_for_file_access(),
3437 : errmsg("could not open file \"%s\": %m", path)));
3438 :
3439 : /*
3440 : * Copy into a temp file name.
3441 : */
3442 78 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3443 :
3444 78 : unlink(tmppath);
3445 :
3446 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3447 78 : fd = OpenTransientFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
3448 78 : if (fd < 0)
3449 0 : ereport(ERROR,
3450 : (errcode_for_file_access(),
3451 : errmsg("could not create file \"%s\": %m", tmppath)));
3452 :
3453 : /*
3454 : * Do the data copying.
3455 : */
3456 159822 : for (nbytes = 0; nbytes < wal_segment_size; nbytes += sizeof(buffer))
3457 : {
3458 : int nread;
3459 :
3460 159744 : nread = upto - nbytes;
3461 :
3462 : /*
3463 : * The part that is not read from the source file is filled with
3464 : * zeros.
3465 : */
3466 159744 : if (nread < sizeof(buffer))
3467 78 : memset(buffer.data, 0, sizeof(buffer));
3468 :
3469 159744 : if (nread > 0)
3470 : {
3471 : int r;
3472 :
3473 5376 : if (nread > sizeof(buffer))
3474 5298 : nread = sizeof(buffer);
3475 5376 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ);
3476 5376 : r = read(srcfd, buffer.data, nread);
3477 5376 : if (r != nread)
3478 : {
3479 0 : if (r < 0)
3480 0 : ereport(ERROR,
3481 : (errcode_for_file_access(),
3482 : errmsg("could not read file \"%s\": %m",
3483 : path)));
3484 : else
3485 0 : ereport(ERROR,
3486 : (errcode(ERRCODE_DATA_CORRUPTED),
3487 : errmsg("could not read file \"%s\": read %d of %zu",
3488 : path, r, (Size) nread)));
3489 : }
3490 5376 : pgstat_report_wait_end();
3491 : }
3492 159744 : errno = 0;
3493 159744 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_WRITE);
3494 159744 : if ((int) write(fd, buffer.data, sizeof(buffer)) != (int) sizeof(buffer))
3495 : {
3496 0 : int save_errno = errno;
3497 :
3498 : /*
3499 : * If we fail to make the file, delete it to release disk space
3500 : */
3501 0 : unlink(tmppath);
3502 : /* if write didn't set errno, assume problem is no disk space */
3503 0 : errno = save_errno ? save_errno : ENOSPC;
3504 :
3505 0 : ereport(ERROR,
3506 : (errcode_for_file_access(),
3507 : errmsg("could not write to file \"%s\": %m", tmppath)));
3508 : }
3509 159744 : pgstat_report_wait_end();
3510 : }
3511 :
3512 78 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_SYNC);
3513 78 : if (pg_fsync(fd) != 0)
3514 0 : ereport(data_sync_elevel(ERROR),
3515 : (errcode_for_file_access(),
3516 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3517 78 : pgstat_report_wait_end();
3518 :
3519 78 : if (CloseTransientFile(fd) != 0)
3520 0 : ereport(ERROR,
3521 : (errcode_for_file_access(),
3522 : errmsg("could not close file \"%s\": %m", tmppath)));
3523 :
3524 78 : if (CloseTransientFile(srcfd) != 0)
3525 0 : ereport(ERROR,
3526 : (errcode_for_file_access(),
3527 : errmsg("could not close file \"%s\": %m", path)));
3528 :
3529 : /*
3530 : * Now move the segment into place with its final name.
3531 : */
3532 78 : if (!InstallXLogFileSegment(&destsegno, tmppath, false, 0, destTLI))
3533 0 : elog(ERROR, "InstallXLogFileSegment should not have failed");
3534 78 : }
3535 :
3536 : /*
3537 : * Install a new XLOG segment file as a current or future log segment.
3538 : *
3539 : * This is used both to install a newly-created segment (which has a temp
3540 : * filename while it's being created) and to recycle an old segment.
3541 : *
3542 : * *segno: identify segment to install as (or first possible target).
3543 : * When find_free is true, this is modified on return to indicate the
3544 : * actual installation location or last segment searched.
3545 : *
3546 : * tmppath: initial name of file to install. It will be renamed into place.
3547 : *
3548 : * find_free: if true, install the new segment at the first empty segno
3549 : * number at or after the passed numbers. If false, install the new segment
3550 : * exactly where specified, deleting any existing segment file there.
3551 : *
3552 : * max_segno: maximum segment number to install the new file as. Fail if no
3553 : * free slot is found between *segno and max_segno. (Ignored when find_free
3554 : * is false.)
3555 : *
3556 : * tli: The timeline on which the new segment should be installed.
3557 : *
3558 : * Returns true if the file was installed successfully. false indicates that
3559 : * max_segno limit was exceeded, the startup process has disabled this
3560 : * function for now, or an error occurred while renaming the file into place.
3561 : */
3562 : static bool
3563 4694 : InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
3564 : bool find_free, XLogSegNo max_segno, TimeLineID tli)
3565 : {
3566 : char path[MAXPGPATH];
3567 : struct stat stat_buf;
3568 :
3569 : Assert(tli != 0);
3570 :
3571 4694 : XLogFilePath(path, tli, *segno, wal_segment_size);
3572 :
3573 4694 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3574 4694 : if (!XLogCtl->InstallXLogFileSegmentActive)
3575 : {
3576 0 : LWLockRelease(ControlFileLock);
3577 0 : return false;
3578 : }
3579 :
3580 4694 : if (!find_free)
3581 : {
3582 : /* Force installation: get rid of any pre-existing segment file */
3583 78 : durable_unlink(path, DEBUG1);
3584 : }
3585 : else
3586 : {
3587 : /* Find a free slot to put it in */
3588 6596 : while (stat(path, &stat_buf) == 0)
3589 : {
3590 2002 : if ((*segno) >= max_segno)
3591 : {
3592 : /* Failed to find a free slot within specified range */
3593 22 : LWLockRelease(ControlFileLock);
3594 22 : return false;
3595 : }
3596 1980 : (*segno)++;
3597 1980 : XLogFilePath(path, tli, *segno, wal_segment_size);
3598 : }
3599 : }
3600 :
3601 : Assert(access(path, F_OK) != 0 && errno == ENOENT);
3602 4672 : if (durable_rename(tmppath, path, LOG) != 0)
3603 : {
3604 0 : LWLockRelease(ControlFileLock);
3605 : /* durable_rename already emitted log message */
3606 0 : return false;
3607 : }
3608 :
3609 4672 : LWLockRelease(ControlFileLock);
3610 :
3611 4672 : return true;
3612 : }
3613 :
3614 : /*
3615 : * Open a pre-existing logfile segment for writing.
3616 : */
3617 : int
3618 62 : XLogFileOpen(XLogSegNo segno, TimeLineID tli)
3619 : {
3620 : char path[MAXPGPATH];
3621 : int fd;
3622 :
3623 62 : XLogFilePath(path, tli, segno, wal_segment_size);
3624 :
3625 62 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3626 62 : get_sync_bit(wal_sync_method));
3627 62 : if (fd < 0)
3628 0 : ereport(PANIC,
3629 : (errcode_for_file_access(),
3630 : errmsg("could not open file \"%s\": %m", path)));
3631 :
3632 62 : return fd;
3633 : }
3634 :
3635 : /*
3636 : * Close the current logfile segment for writing.
3637 : */
3638 : static void
3639 9348 : XLogFileClose(void)
3640 : {
3641 : Assert(openLogFile >= 0);
3642 :
3643 : /*
3644 : * WAL segment files will not be re-read in normal operation, so we advise
3645 : * the OS to release any cached pages. But do not do so if WAL archiving
3646 : * or streaming is active, because archiver and walsender process could
3647 : * use the cache to read the WAL segment.
3648 : */
3649 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3650 9348 : if (!XLogIsNeeded() && (io_direct_flags & IO_DIRECT_WAL) == 0)
3651 2284 : (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3652 : #endif
3653 :
3654 9348 : if (close(openLogFile) != 0)
3655 : {
3656 : char xlogfname[MAXFNAMELEN];
3657 0 : int save_errno = errno;
3658 :
3659 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo, wal_segment_size);
3660 0 : errno = save_errno;
3661 0 : ereport(PANIC,
3662 : (errcode_for_file_access(),
3663 : errmsg("could not close file \"%s\": %m", xlogfname)));
3664 : }
3665 :
3666 9348 : openLogFile = -1;
3667 9348 : ReleaseExternalFD();
3668 9348 : }
3669 :
3670 : /*
3671 : * Preallocate log files beyond the specified log endpoint.
3672 : *
3673 : * XXX this is currently extremely conservative, since it forces only one
3674 : * future log segment to exist, and even that only if we are 75% done with
3675 : * the current one. This is only appropriate for very low-WAL-volume systems.
3676 : * High-volume systems will be OK once they've built up a sufficient set of
3677 : * recycled log segments, but the startup transient is likely to include
3678 : * a lot of segment creations by foreground processes, which is not so good.
3679 : *
3680 : * XLogFileInitInternal() can ereport(ERROR). All known causes indicate big
3681 : * trouble; for example, a full filesystem is one cause. The checkpoint WAL
3682 : * and/or ControlFile updates already completed. If a RequestCheckpoint()
3683 : * initiated the present checkpoint and an ERROR ends this function, the
3684 : * command that called RequestCheckpoint() fails. That's not ideal, but it's
3685 : * not worth contorting more functions to use caller-specified elevel values.
3686 : * (With or without RequestCheckpoint(), an ERROR forestalls some inessential
3687 : * reporting and resource reclamation.)
3688 : */
3689 : static void
3690 2956 : PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli)
3691 : {
3692 : XLogSegNo _logSegNo;
3693 : int lf;
3694 : bool added;
3695 : char path[MAXPGPATH];
3696 : uint64 offset;
3697 :
3698 2956 : if (!XLogCtl->InstallXLogFileSegmentActive)
3699 6 : return; /* unlocked check says no */
3700 :
3701 2950 : XLByteToPrevSeg(endptr, _logSegNo, wal_segment_size);
3702 2950 : offset = XLogSegmentOffset(endptr - 1, wal_segment_size);
3703 2950 : if (offset >= (uint32) (0.75 * wal_segment_size))
3704 : {
3705 496 : _logSegNo++;
3706 496 : lf = XLogFileInitInternal(_logSegNo, tli, &added, path);
3707 496 : if (lf >= 0)
3708 196 : close(lf);
3709 496 : if (added)
3710 298 : CheckpointStats.ckpt_segs_added++;
3711 : }
3712 : }
3713 :
3714 : /*
3715 : * Throws an error if the given log segment has already been removed or
3716 : * recycled. The caller should only pass a segment that it knows to have
3717 : * existed while the server has been running, as this function always
3718 : * succeeds if no WAL segments have been removed since startup.
3719 : * 'tli' is only used in the error message.
3720 : *
3721 : * Note: this function guarantees to keep errno unchanged on return.
3722 : * This supports callers that use this to possibly deliver a better
3723 : * error message about a missing file, while still being able to throw
3724 : * a normal file-access error afterwards, if this does return.
3725 : */
3726 : void
3727 242054 : CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3728 : {
3729 242054 : int save_errno = errno;
3730 : XLogSegNo lastRemovedSegNo;
3731 :
3732 242054 : SpinLockAcquire(&XLogCtl->info_lck);
3733 242054 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3734 242054 : SpinLockRelease(&XLogCtl->info_lck);
3735 :
3736 242054 : if (segno <= lastRemovedSegNo)
3737 : {
3738 : char filename[MAXFNAMELEN];
3739 :
3740 0 : XLogFileName(filename, tli, segno, wal_segment_size);
3741 0 : errno = save_errno;
3742 0 : ereport(ERROR,
3743 : (errcode_for_file_access(),
3744 : errmsg("requested WAL segment %s has already been removed",
3745 : filename)));
3746 : }
3747 242054 : errno = save_errno;
3748 242054 : }
3749 :
3750 : /*
3751 : * Return the last WAL segment removed, or 0 if no segment has been removed
3752 : * since startup.
3753 : *
3754 : * NB: the result can be out of date arbitrarily fast, the caller has to deal
3755 : * with that.
3756 : */
3757 : XLogSegNo
3758 1958 : XLogGetLastRemovedSegno(void)
3759 : {
3760 : XLogSegNo lastRemovedSegNo;
3761 :
3762 1958 : SpinLockAcquire(&XLogCtl->info_lck);
3763 1958 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3764 1958 : SpinLockRelease(&XLogCtl->info_lck);
3765 :
3766 1958 : return lastRemovedSegNo;
3767 : }
3768 :
3769 : /*
3770 : * Return the oldest WAL segment on the given TLI that still exists in
3771 : * XLOGDIR, or 0 if none.
3772 : */
3773 : XLogSegNo
3774 10 : XLogGetOldestSegno(TimeLineID tli)
3775 : {
3776 : DIR *xldir;
3777 : struct dirent *xlde;
3778 10 : XLogSegNo oldest_segno = 0;
3779 :
3780 10 : xldir = AllocateDir(XLOGDIR);
3781 66 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3782 : {
3783 : TimeLineID file_tli;
3784 : XLogSegNo file_segno;
3785 :
3786 : /* Ignore files that are not XLOG segments. */
3787 56 : if (!IsXLogFileName(xlde->d_name))
3788 40 : continue;
3789 :
3790 : /* Parse filename to get TLI and segno. */
3791 16 : XLogFromFileName(xlde->d_name, &file_tli, &file_segno,
3792 : wal_segment_size);
3793 :
3794 : /* Ignore anything that's not from the TLI of interest. */
3795 16 : if (tli != file_tli)
3796 0 : continue;
3797 :
3798 : /* If it's the oldest so far, update oldest_segno. */
3799 16 : if (oldest_segno == 0 || file_segno < oldest_segno)
3800 14 : oldest_segno = file_segno;
3801 : }
3802 :
3803 10 : FreeDir(xldir);
3804 10 : return oldest_segno;
3805 : }
3806 :
3807 : /*
3808 : * Update the last removed segno pointer in shared memory, to reflect that the
3809 : * given XLOG file has been removed.
3810 : */
3811 : static void
3812 3788 : UpdateLastRemovedPtr(char *filename)
3813 : {
3814 : uint32 tli;
3815 : XLogSegNo segno;
3816 :
3817 3788 : XLogFromFileName(filename, &tli, &segno, wal_segment_size);
3818 :
3819 3788 : SpinLockAcquire(&XLogCtl->info_lck);
3820 3788 : if (segno > XLogCtl->lastRemovedSegNo)
3821 1666 : XLogCtl->lastRemovedSegNo = segno;
3822 3788 : SpinLockRelease(&XLogCtl->info_lck);
3823 3788 : }
3824 :
3825 : /*
3826 : * Remove all temporary log files in pg_wal
3827 : *
3828 : * This is called at the beginning of recovery after a previous crash,
3829 : * at a point where no other processes write fresh WAL data.
3830 : */
3831 : static void
3832 340 : RemoveTempXlogFiles(void)
3833 : {
3834 : DIR *xldir;
3835 : struct dirent *xlde;
3836 :
3837 340 : elog(DEBUG2, "removing all temporary WAL segments");
3838 :
3839 340 : xldir = AllocateDir(XLOGDIR);
3840 2202 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3841 : {
3842 : char path[MAXPGPATH];
3843 :
3844 1862 : if (strncmp(xlde->d_name, "xlogtemp.", 9) != 0)
3845 1862 : continue;
3846 :
3847 0 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
3848 0 : unlink(path);
3849 0 : elog(DEBUG2, "removed temporary WAL segment \"%s\"", path);
3850 : }
3851 340 : FreeDir(xldir);
3852 340 : }
3853 :
3854 : /*
3855 : * Recycle or remove all log files older or equal to passed segno.
3856 : *
3857 : * endptr is current (or recent) end of xlog, and lastredoptr is the
3858 : * redo pointer of the last checkpoint. These are used to determine
3859 : * whether we want to recycle rather than delete no-longer-wanted log files.
3860 : *
3861 : * insertTLI is the current timeline for XLOG insertion. Any recycled
3862 : * segments should be reused for this timeline.
3863 : */
3864 : static void
3865 2468 : RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr, XLogRecPtr endptr,
3866 : TimeLineID insertTLI)
3867 : {
3868 : DIR *xldir;
3869 : struct dirent *xlde;
3870 : char lastoff[MAXFNAMELEN];
3871 : XLogSegNo endlogSegNo;
3872 : XLogSegNo recycleSegNo;
3873 :
3874 : /* Initialize info about where to try to recycle to */
3875 2468 : XLByteToSeg(endptr, endlogSegNo, wal_segment_size);
3876 2468 : recycleSegNo = XLOGfileslop(lastredoptr);
3877 :
3878 : /*
3879 : * Construct a filename of the last segment to be kept. The timeline ID
3880 : * doesn't matter, we ignore that in the comparison. (During recovery,
3881 : * InsertTimeLineID isn't set, so we can't use that.)
3882 : */
3883 2468 : XLogFileName(lastoff, 0, segno, wal_segment_size);
3884 :
3885 2468 : elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
3886 : lastoff);
3887 :
3888 2468 : xldir = AllocateDir(XLOGDIR);
3889 :
3890 43226 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3891 : {
3892 : /* Ignore files that are not XLOG segments */
3893 40758 : if (!IsXLogFileName(xlde->d_name) &&
3894 10312 : !IsPartialXLogFileName(xlde->d_name))
3895 10308 : continue;
3896 :
3897 : /*
3898 : * We ignore the timeline part of the XLOG segment identifiers in
3899 : * deciding whether a segment is still needed. This ensures that we
3900 : * won't prematurely remove a segment from a parent timeline. We could
3901 : * probably be a little more proactive about removing segments of
3902 : * non-parent timelines, but that would be a whole lot more
3903 : * complicated.
3904 : *
3905 : * We use the alphanumeric sorting property of the filenames to decide
3906 : * which ones are earlier than the lastoff segment.
3907 : */
3908 30450 : if (strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
3909 : {
3910 3806 : if (XLogArchiveCheckDone(xlde->d_name))
3911 : {
3912 : /* Update the last removed location in shared memory first */
3913 3788 : UpdateLastRemovedPtr(xlde->d_name);
3914 :
3915 3788 : RemoveXlogFile(xlde, recycleSegNo, &endlogSegNo, insertTLI);
3916 : }
3917 : }
3918 : }
3919 :
3920 2468 : FreeDir(xldir);
3921 2468 : }
3922 :
3923 : /*
3924 : * Recycle or remove WAL files that are not part of the given timeline's
3925 : * history.
3926 : *
3927 : * This is called during recovery, whenever we switch to follow a new
3928 : * timeline, and at the end of recovery when we create a new timeline. We
3929 : * wouldn't otherwise care about extra WAL files lying in pg_wal, but they
3930 : * might be leftover pre-allocated or recycled WAL segments on the old timeline
3931 : * that we haven't used yet, and contain garbage. If we just leave them in
3932 : * pg_wal, they will eventually be archived, and we can't let that happen.
3933 : * Files that belong to our timeline history are valid, because we have
3934 : * successfully replayed them, but from others we can't be sure.
3935 : *
3936 : * 'switchpoint' is the current point in WAL where we switch to new timeline,
3937 : * and 'newTLI' is the new timeline we switch to.
3938 : */
3939 : void
3940 118 : RemoveNonParentXlogFiles(XLogRecPtr switchpoint, TimeLineID newTLI)
3941 : {
3942 : DIR *xldir;
3943 : struct dirent *xlde;
3944 : char switchseg[MAXFNAMELEN];
3945 : XLogSegNo endLogSegNo;
3946 : XLogSegNo switchLogSegNo;
3947 : XLogSegNo recycleSegNo;
3948 :
3949 : /*
3950 : * Initialize info about where to begin the work. This will recycle,
3951 : * somewhat arbitrarily, 10 future segments.
3952 : */
3953 118 : XLByteToPrevSeg(switchpoint, switchLogSegNo, wal_segment_size);
3954 118 : XLByteToSeg(switchpoint, endLogSegNo, wal_segment_size);
3955 118 : recycleSegNo = endLogSegNo + 10;
3956 :
3957 : /*
3958 : * Construct a filename of the last segment to be kept.
3959 : */
3960 118 : XLogFileName(switchseg, newTLI, switchLogSegNo, wal_segment_size);
3961 :
3962 118 : elog(DEBUG2, "attempting to remove WAL segments newer than log file %s",
3963 : switchseg);
3964 :
3965 118 : xldir = AllocateDir(XLOGDIR);
3966 :
3967 1126 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3968 : {
3969 : /* Ignore files that are not XLOG segments */
3970 1008 : if (!IsXLogFileName(xlde->d_name))
3971 626 : continue;
3972 :
3973 : /*
3974 : * Remove files that are on a timeline older than the new one we're
3975 : * switching to, but with a segment number >= the first segment on the
3976 : * new timeline.
3977 : */
3978 382 : if (strncmp(xlde->d_name, switchseg, 8) < 0 &&
3979 248 : strcmp(xlde->d_name + 8, switchseg + 8) > 0)
3980 : {
3981 : /*
3982 : * If the file has already been marked as .ready, however, don't
3983 : * remove it yet. It should be OK to remove it - files that are
3984 : * not part of our timeline history are not required for recovery
3985 : * - but seems safer to let them be archived and removed later.
3986 : */
3987 32 : if (!XLogArchiveIsReady(xlde->d_name))
3988 32 : RemoveXlogFile(xlde, recycleSegNo, &endLogSegNo, newTLI);
3989 : }
3990 : }
3991 :
3992 118 : FreeDir(xldir);
3993 118 : }
3994 :
3995 : /*
3996 : * Recycle or remove a log file that's no longer needed.
3997 : *
3998 : * segment_de is the dirent structure of the segment to recycle or remove.
3999 : * recycleSegNo is the segment number to recycle up to. endlogSegNo is
4000 : * the segment number of the current (or recent) end of WAL.
4001 : *
4002 : * endlogSegNo gets incremented if the segment is recycled so as it is not
4003 : * checked again with future callers of this function.
4004 : *
4005 : * insertTLI is the current timeline for XLOG insertion. Any recycled segments
4006 : * should be used for this timeline.
4007 : */
4008 : static void
4009 3820 : RemoveXlogFile(const struct dirent *segment_de,
4010 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
4011 : TimeLineID insertTLI)
4012 : {
4013 : char path[MAXPGPATH];
4014 : #ifdef WIN32
4015 : char newpath[MAXPGPATH];
4016 : #endif
4017 3820 : const char *segname = segment_de->d_name;
4018 :
4019 3820 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", segname);
4020 :
4021 : /*
4022 : * Before deleting the file, see if it can be recycled as a future log
4023 : * segment. Only recycle normal files, because we don't want to recycle
4024 : * symbolic links pointing to a separate archive directory.
4025 : */
4026 3820 : if (wal_recycle &&
4027 3820 : *endlogSegNo <= recycleSegNo &&
4028 4096 : XLogCtl->InstallXLogFileSegmentActive && /* callee rechecks this */
4029 4092 : get_dirent_type(path, segment_de, false, DEBUG2) == PGFILETYPE_REG &&
4030 2046 : InstallXLogFileSegment(endlogSegNo, path,
4031 : true, recycleSegNo, insertTLI))
4032 : {
4033 2026 : ereport(DEBUG2,
4034 : (errmsg_internal("recycled write-ahead log file \"%s\"",
4035 : segname)));
4036 2026 : CheckpointStats.ckpt_segs_recycled++;
4037 : /* Needn't recheck that slot on future iterations */
4038 2026 : (*endlogSegNo)++;
4039 : }
4040 : else
4041 : {
4042 : /* No need for any more future segments, or recycling failed ... */
4043 : int rc;
4044 :
4045 1794 : ereport(DEBUG2,
4046 : (errmsg_internal("removing write-ahead log file \"%s\"",
4047 : segname)));
4048 :
4049 : #ifdef WIN32
4050 :
4051 : /*
4052 : * On Windows, if another process (e.g another backend) holds the file
4053 : * open in FILE_SHARE_DELETE mode, unlink will succeed, but the file
4054 : * will still show up in directory listing until the last handle is
4055 : * closed. To avoid confusing the lingering deleted file for a live
4056 : * WAL file that needs to be archived, rename it before deleting it.
4057 : *
4058 : * If another process holds the file open without FILE_SHARE_DELETE
4059 : * flag, rename will fail. We'll try again at the next checkpoint.
4060 : */
4061 : snprintf(newpath, MAXPGPATH, "%s.deleted", path);
4062 : if (rename(path, newpath) != 0)
4063 : {
4064 : ereport(LOG,
4065 : (errcode_for_file_access(),
4066 : errmsg("could not rename file \"%s\": %m",
4067 : path)));
4068 : return;
4069 : }
4070 : rc = durable_unlink(newpath, LOG);
4071 : #else
4072 1794 : rc = durable_unlink(path, LOG);
4073 : #endif
4074 1794 : if (rc != 0)
4075 : {
4076 : /* Message already logged by durable_unlink() */
4077 0 : return;
4078 : }
4079 1794 : CheckpointStats.ckpt_segs_removed++;
4080 : }
4081 :
4082 3820 : XLogArchiveCleanup(segname);
4083 : }
4084 :
4085 : /*
4086 : * Verify whether pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
4087 : * If the latter do not exist, recreate them.
4088 : *
4089 : * It is not the goal of this function to verify the contents of these
4090 : * directories, but to help in cases where someone has performed a cluster
4091 : * copy for PITR purposes but omitted pg_wal from the copy.
4092 : *
4093 : * We could also recreate pg_wal if it doesn't exist, but a deliberate
4094 : * policy decision was made not to. It is fairly common for pg_wal to be
4095 : * a symlink, and if that was the DBA's intent then automatically making a
4096 : * plain directory would result in degraded performance with no notice.
4097 : */
4098 : static void
4099 1640 : ValidateXLOGDirectoryStructure(void)
4100 : {
4101 : char path[MAXPGPATH];
4102 : struct stat stat_buf;
4103 :
4104 : /* Check for pg_wal; if it doesn't exist, error out */
4105 1640 : if (stat(XLOGDIR, &stat_buf) != 0 ||
4106 1640 : !S_ISDIR(stat_buf.st_mode))
4107 0 : ereport(FATAL,
4108 : (errcode_for_file_access(),
4109 : errmsg("required WAL directory \"%s\" does not exist",
4110 : XLOGDIR)));
4111 :
4112 : /* Check for archive_status */
4113 1640 : snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4114 1640 : if (stat(path, &stat_buf) == 0)
4115 : {
4116 : /* Check for weird cases where it exists but isn't a directory */
4117 1638 : if (!S_ISDIR(stat_buf.st_mode))
4118 0 : ereport(FATAL,
4119 : (errcode_for_file_access(),
4120 : errmsg("required WAL directory \"%s\" does not exist",
4121 : path)));
4122 : }
4123 : else
4124 : {
4125 2 : ereport(LOG,
4126 : (errmsg("creating missing WAL directory \"%s\"", path)));
4127 2 : if (MakePGDirectory(path) < 0)
4128 0 : ereport(FATAL,
4129 : (errcode_for_file_access(),
4130 : errmsg("could not create missing directory \"%s\": %m",
4131 : path)));
4132 : }
4133 :
4134 : /* Check for summaries */
4135 1640 : snprintf(path, MAXPGPATH, XLOGDIR "/summaries");
4136 1640 : if (stat(path, &stat_buf) == 0)
4137 : {
4138 : /* Check for weird cases where it exists but isn't a directory */
4139 1638 : if (!S_ISDIR(stat_buf.st_mode))
4140 0 : ereport(FATAL,
4141 : (errmsg("required WAL directory \"%s\" does not exist",
4142 : path)));
4143 : }
4144 : else
4145 : {
4146 2 : ereport(LOG,
4147 : (errmsg("creating missing WAL directory \"%s\"", path)));
4148 2 : if (MakePGDirectory(path) < 0)
4149 0 : ereport(FATAL,
4150 : (errmsg("could not create missing directory \"%s\": %m",
4151 : path)));
4152 : }
4153 1640 : }
4154 :
4155 : /*
4156 : * Remove previous backup history files. This also retries creation of
4157 : * .ready files for any backup history files for which XLogArchiveNotify
4158 : * failed earlier.
4159 : */
4160 : static void
4161 288 : CleanupBackupHistory(void)
4162 : {
4163 : DIR *xldir;
4164 : struct dirent *xlde;
4165 : char path[MAXPGPATH + sizeof(XLOGDIR)];
4166 :
4167 288 : xldir = AllocateDir(XLOGDIR);
4168 :
4169 2602 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4170 : {
4171 2314 : if (IsBackupHistoryFileName(xlde->d_name))
4172 : {
4173 304 : if (XLogArchiveCheckDone(xlde->d_name))
4174 : {
4175 244 : elog(DEBUG2, "removing WAL backup history file \"%s\"",
4176 : xlde->d_name);
4177 244 : snprintf(path, sizeof(path), XLOGDIR "/%s", xlde->d_name);
4178 244 : unlink(path);
4179 244 : XLogArchiveCleanup(xlde->d_name);
4180 : }
4181 : }
4182 : }
4183 :
4184 288 : FreeDir(xldir);
4185 288 : }
4186 :
4187 : /*
4188 : * I/O routines for pg_control
4189 : *
4190 : * *ControlFile is a buffer in shared memory that holds an image of the
4191 : * contents of pg_control. WriteControlFile() initializes pg_control
4192 : * given a preloaded buffer, ReadControlFile() loads the buffer from
4193 : * the pg_control file (during postmaster or standalone-backend startup),
4194 : * and UpdateControlFile() rewrites pg_control after we modify xlog state.
4195 : * InitControlFile() fills the buffer with initial values.
4196 : *
4197 : * For simplicity, WriteControlFile() initializes the fields of pg_control
4198 : * that are related to checking backend/database compatibility, and
4199 : * ReadControlFile() verifies they are correct. We could split out the
4200 : * I/O and compatibility-check functions, but there seems no need currently.
4201 : */
4202 :
4203 : static void
4204 90 : InitControlFile(uint64 sysidentifier, uint32 data_checksum_version)
4205 : {
4206 : char mock_auth_nonce[MOCK_AUTH_NONCE_LEN];
4207 :
4208 : /*
4209 : * Generate a random nonce. This is used for authentication requests that
4210 : * will fail because the user does not exist. The nonce is used to create
4211 : * a genuine-looking password challenge for the non-existent user, in lieu
4212 : * of an actual stored password.
4213 : */
4214 90 : if (!pg_strong_random(mock_auth_nonce, MOCK_AUTH_NONCE_LEN))
4215 0 : ereport(PANIC,
4216 : (errcode(ERRCODE_INTERNAL_ERROR),
4217 : errmsg("could not generate secret authorization token")));
4218 :
4219 90 : memset(ControlFile, 0, sizeof(ControlFileData));
4220 : /* Initialize pg_control status fields */
4221 90 : ControlFile->system_identifier = sysidentifier;
4222 90 : memcpy(ControlFile->mock_authentication_nonce, mock_auth_nonce, MOCK_AUTH_NONCE_LEN);
4223 90 : ControlFile->state = DB_SHUTDOWNED;
4224 90 : ControlFile->unloggedLSN = FirstNormalUnloggedLSN;
4225 :
4226 : /* Set important parameter values for use when replaying WAL */
4227 90 : ControlFile->MaxConnections = MaxConnections;
4228 90 : ControlFile->max_worker_processes = max_worker_processes;
4229 90 : ControlFile->max_wal_senders = max_wal_senders;
4230 90 : ControlFile->max_prepared_xacts = max_prepared_xacts;
4231 90 : ControlFile->max_locks_per_xact = max_locks_per_xact;
4232 90 : ControlFile->wal_level = wal_level;
4233 90 : ControlFile->wal_log_hints = wal_log_hints;
4234 90 : ControlFile->track_commit_timestamp = track_commit_timestamp;
4235 90 : ControlFile->data_checksum_version = data_checksum_version;
4236 90 : }
4237 :
4238 : static void
4239 90 : WriteControlFile(void)
4240 : {
4241 : int fd;
4242 : char buffer[PG_CONTROL_FILE_SIZE]; /* need not be aligned */
4243 :
4244 : /*
4245 : * Initialize version and compatibility-check fields
4246 : */
4247 90 : ControlFile->pg_control_version = PG_CONTROL_VERSION;
4248 90 : ControlFile->catalog_version_no = CATALOG_VERSION_NO;
4249 :
4250 90 : ControlFile->maxAlign = MAXIMUM_ALIGNOF;
4251 90 : ControlFile->floatFormat = FLOATFORMAT_VALUE;
4252 :
4253 90 : ControlFile->blcksz = BLCKSZ;
4254 90 : ControlFile->relseg_size = RELSEG_SIZE;
4255 90 : ControlFile->xlog_blcksz = XLOG_BLCKSZ;
4256 90 : ControlFile->xlog_seg_size = wal_segment_size;
4257 :
4258 90 : ControlFile->nameDataLen = NAMEDATALEN;
4259 90 : ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
4260 :
4261 90 : ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
4262 90 : ControlFile->loblksize = LOBLKSIZE;
4263 :
4264 90 : ControlFile->float8ByVal = FLOAT8PASSBYVAL;
4265 :
4266 : /* Contents are protected with a CRC */
4267 90 : INIT_CRC32C(ControlFile->crc);
4268 90 : COMP_CRC32C(ControlFile->crc,
4269 : (char *) ControlFile,
4270 : offsetof(ControlFileData, crc));
4271 90 : FIN_CRC32C(ControlFile->crc);
4272 :
4273 : /*
4274 : * We write out PG_CONTROL_FILE_SIZE bytes into pg_control, zero-padding
4275 : * the excess over sizeof(ControlFileData). This reduces the odds of
4276 : * premature-EOF errors when reading pg_control. We'll still fail when we
4277 : * check the contents of the file, but hopefully with a more specific
4278 : * error than "couldn't read pg_control".
4279 : */
4280 90 : memset(buffer, 0, PG_CONTROL_FILE_SIZE);
4281 90 : memcpy(buffer, ControlFile, sizeof(ControlFileData));
4282 :
4283 90 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4284 : O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
4285 90 : if (fd < 0)
4286 0 : ereport(PANIC,
4287 : (errcode_for_file_access(),
4288 : errmsg("could not create file \"%s\": %m",
4289 : XLOG_CONTROL_FILE)));
4290 :
4291 90 : errno = 0;
4292 90 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_WRITE);
4293 90 : if (write(fd, buffer, PG_CONTROL_FILE_SIZE) != PG_CONTROL_FILE_SIZE)
4294 : {
4295 : /* if write didn't set errno, assume problem is no disk space */
4296 0 : if (errno == 0)
4297 0 : errno = ENOSPC;
4298 0 : ereport(PANIC,
4299 : (errcode_for_file_access(),
4300 : errmsg("could not write to file \"%s\": %m",
4301 : XLOG_CONTROL_FILE)));
4302 : }
4303 90 : pgstat_report_wait_end();
4304 :
4305 90 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_SYNC);
4306 90 : if (pg_fsync(fd) != 0)
4307 0 : ereport(PANIC,
4308 : (errcode_for_file_access(),
4309 : errmsg("could not fsync file \"%s\": %m",
4310 : XLOG_CONTROL_FILE)));
4311 90 : pgstat_report_wait_end();
4312 :
4313 90 : if (close(fd) != 0)
4314 0 : ereport(PANIC,
4315 : (errcode_for_file_access(),
4316 : errmsg("could not close file \"%s\": %m",
4317 : XLOG_CONTROL_FILE)));
4318 90 : }
4319 :
4320 : static void
4321 1744 : ReadControlFile(void)
4322 : {
4323 : pg_crc32c crc;
4324 : int fd;
4325 : char wal_segsz_str[20];
4326 : int r;
4327 :
4328 : /*
4329 : * Read data...
4330 : */
4331 1744 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4332 : O_RDWR | PG_BINARY);
4333 1744 : if (fd < 0)
4334 0 : ereport(PANIC,
4335 : (errcode_for_file_access(),
4336 : errmsg("could not open file \"%s\": %m",
4337 : XLOG_CONTROL_FILE)));
4338 :
4339 1744 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ);
4340 1744 : r = read(fd, ControlFile, sizeof(ControlFileData));
4341 1744 : if (r != sizeof(ControlFileData))
4342 : {
4343 0 : if (r < 0)
4344 0 : ereport(PANIC,
4345 : (errcode_for_file_access(),
4346 : errmsg("could not read file \"%s\": %m",
4347 : XLOG_CONTROL_FILE)));
4348 : else
4349 0 : ereport(PANIC,
4350 : (errcode(ERRCODE_DATA_CORRUPTED),
4351 : errmsg("could not read file \"%s\": read %d of %zu",
4352 : XLOG_CONTROL_FILE, r, sizeof(ControlFileData))));
4353 : }
4354 1744 : pgstat_report_wait_end();
4355 :
4356 1744 : close(fd);
4357 :
4358 : /*
4359 : * Check for expected pg_control format version. If this is wrong, the
4360 : * CRC check will likely fail because we'll be checking the wrong number
4361 : * of bytes. Complaining about wrong version will probably be more
4362 : * enlightening than complaining about wrong CRC.
4363 : */
4364 :
4365 1744 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
4366 0 : ereport(FATAL,
4367 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4368 : errmsg("database files are incompatible with server"),
4369 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4370 : " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4371 : ControlFile->pg_control_version, ControlFile->pg_control_version,
4372 : PG_CONTROL_VERSION, PG_CONTROL_VERSION),
4373 : errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4374 :
4375 1744 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
4376 0 : ereport(FATAL,
4377 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4378 : errmsg("database files are incompatible with server"),
4379 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4380 : " but the server was compiled with PG_CONTROL_VERSION %d.",
4381 : ControlFile->pg_control_version, PG_CONTROL_VERSION),
4382 : errhint("It looks like you need to initdb.")));
4383 :
4384 : /* Now check the CRC. */
4385 1744 : INIT_CRC32C(crc);
4386 1744 : COMP_CRC32C(crc,
4387 : (char *) ControlFile,
4388 : offsetof(ControlFileData, crc));
4389 1744 : FIN_CRC32C(crc);
4390 :
4391 1744 : if (!EQ_CRC32C(crc, ControlFile->crc))
4392 0 : ereport(FATAL,
4393 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4394 : errmsg("incorrect checksum in control file")));
4395 :
4396 : /*
4397 : * Do compatibility checking immediately. If the database isn't
4398 : * compatible with the backend executable, we want to abort before we can
4399 : * possibly do any damage.
4400 : */
4401 1744 : if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
4402 0 : ereport(FATAL,
4403 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4404 : errmsg("database files are incompatible with server"),
4405 : /* translator: %s is a variable name and %d is its value */
4406 : errdetail("The database cluster was initialized with %s %d,"
4407 : " but the server was compiled with %s %d.",
4408 : "CATALOG_VERSION_NO", ControlFile->catalog_version_no,
4409 : "CATALOG_VERSION_NO", CATALOG_VERSION_NO),
4410 : errhint("It looks like you need to initdb.")));
4411 1744 : if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
4412 0 : ereport(FATAL,
4413 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4414 : errmsg("database files are incompatible with server"),
4415 : /* translator: %s is a variable name and %d is its value */
4416 : errdetail("The database cluster was initialized with %s %d,"
4417 : " but the server was compiled with %s %d.",
4418 : "MAXALIGN", ControlFile->maxAlign,
4419 : "MAXALIGN", MAXIMUM_ALIGNOF),
4420 : errhint("It looks like you need to initdb.")));
4421 1744 : if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
4422 0 : ereport(FATAL,
4423 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4424 : errmsg("database files are incompatible with server"),
4425 : errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4426 : errhint("It looks like you need to initdb.")));
4427 1744 : if (ControlFile->blcksz != BLCKSZ)
4428 0 : ereport(FATAL,
4429 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4430 : errmsg("database files are incompatible with server"),
4431 : /* translator: %s is a variable name and %d is its value */
4432 : errdetail("The database cluster was initialized with %s %d,"
4433 : " but the server was compiled with %s %d.",
4434 : "BLCKSZ", ControlFile->blcksz,
4435 : "BLCKSZ", BLCKSZ),
4436 : errhint("It looks like you need to recompile or initdb.")));
4437 1744 : if (ControlFile->relseg_size != RELSEG_SIZE)
4438 0 : ereport(FATAL,
4439 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4440 : errmsg("database files are incompatible with server"),
4441 : /* translator: %s is a variable name and %d is its value */
4442 : errdetail("The database cluster was initialized with %s %d,"
4443 : " but the server was compiled with %s %d.",
4444 : "RELSEG_SIZE", ControlFile->relseg_size,
4445 : "RELSEG_SIZE", RELSEG_SIZE),
4446 : errhint("It looks like you need to recompile or initdb.")));
4447 1744 : if (ControlFile->xlog_blcksz != XLOG_BLCKSZ)
4448 0 : ereport(FATAL,
4449 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4450 : errmsg("database files are incompatible with server"),
4451 : /* translator: %s is a variable name and %d is its value */
4452 : errdetail("The database cluster was initialized with %s %d,"
4453 : " but the server was compiled with %s %d.",
4454 : "XLOG_BLCKSZ", ControlFile->xlog_blcksz,
4455 : "XLOG_BLCKSZ", XLOG_BLCKSZ),
4456 : errhint("It looks like you need to recompile or initdb.")));
4457 1744 : if (ControlFile->nameDataLen != NAMEDATALEN)
4458 0 : ereport(FATAL,
4459 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4460 : errmsg("database files are incompatible with server"),
4461 : /* translator: %s is a variable name and %d is its value */
4462 : errdetail("The database cluster was initialized with %s %d,"
4463 : " but the server was compiled with %s %d.",
4464 : "NAMEDATALEN", ControlFile->nameDataLen,
4465 : "NAMEDATALEN", NAMEDATALEN),
4466 : errhint("It looks like you need to recompile or initdb.")));
4467 1744 : if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
4468 0 : ereport(FATAL,
4469 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4470 : errmsg("database files are incompatible with server"),
4471 : /* translator: %s is a variable name and %d is its value */
4472 : errdetail("The database cluster was initialized with %s %d,"
4473 : " but the server was compiled with %s %d.",
4474 : "INDEX_MAX_KEYS", ControlFile->indexMaxKeys,
4475 : "INDEX_MAX_KEYS", INDEX_MAX_KEYS),
4476 : errhint("It looks like you need to recompile or initdb.")));
4477 1744 : if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
4478 0 : ereport(FATAL,
4479 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4480 : errmsg("database files are incompatible with server"),
4481 : /* translator: %s is a variable name and %d is its value */
4482 : errdetail("The database cluster was initialized with %s %d,"
4483 : " but the server was compiled with %s %d.",
4484 : "TOAST_MAX_CHUNK_SIZE", ControlFile->toast_max_chunk_size,
4485 : "TOAST_MAX_CHUNK_SIZE", (int) TOAST_MAX_CHUNK_SIZE),
4486 : errhint("It looks like you need to recompile or initdb.")));
4487 1744 : if (ControlFile->loblksize != LOBLKSIZE)
4488 0 : ereport(FATAL,
4489 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4490 : errmsg("database files are incompatible with server"),
4491 : /* translator: %s is a variable name and %d is its value */
4492 : errdetail("The database cluster was initialized with %s %d,"
4493 : " but the server was compiled with %s %d.",
4494 : "LOBLKSIZE", ControlFile->loblksize,
4495 : "LOBLKSIZE", (int) LOBLKSIZE),
4496 : errhint("It looks like you need to recompile or initdb.")));
4497 :
4498 : #ifdef USE_FLOAT8_BYVAL
4499 1744 : if (ControlFile->float8ByVal != true)
4500 0 : ereport(FATAL,
4501 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4502 : errmsg("database files are incompatible with server"),
4503 : errdetail("The database cluster was initialized without USE_FLOAT8_BYVAL"
4504 : " but the server was compiled with USE_FLOAT8_BYVAL."),
4505 : errhint("It looks like you need to recompile or initdb.")));
4506 : #else
4507 : if (ControlFile->float8ByVal != false)
4508 : ereport(FATAL,
4509 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4510 : errmsg("database files are incompatible with server"),
4511 : errdetail("The database cluster was initialized with USE_FLOAT8_BYVAL"
4512 : " but the server was compiled without USE_FLOAT8_BYVAL."),
4513 : errhint("It looks like you need to recompile or initdb.")));
4514 : #endif
4515 :
4516 1744 : wal_segment_size = ControlFile->xlog_seg_size;
4517 :
4518 1744 : if (!IsValidWalSegSize(wal_segment_size))
4519 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4520 : errmsg_plural("invalid WAL segment size in control file (%d byte)",
4521 : "invalid WAL segment size in control file (%d bytes)",
4522 : wal_segment_size,
4523 : wal_segment_size),
4524 : errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.")));
4525 :
4526 1744 : snprintf(wal_segsz_str, sizeof(wal_segsz_str), "%d", wal_segment_size);
4527 1744 : SetConfigOption("wal_segment_size", wal_segsz_str, PGC_INTERNAL,
4528 : PGC_S_DYNAMIC_DEFAULT);
4529 :
4530 : /* check and update variables dependent on wal_segment_size */
4531 1744 : if (ConvertToXSegs(min_wal_size_mb, wal_segment_size) < 2)
4532 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4533 : /* translator: both %s are GUC names */
4534 : errmsg("\"%s\" must be at least twice \"%s\"",
4535 : "min_wal_size", "wal_segment_size")));
4536 :
4537 1744 : if (ConvertToXSegs(max_wal_size_mb, wal_segment_size) < 2)
4538 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4539 : /* translator: both %s are GUC names */
4540 : errmsg("\"%s\" must be at least twice \"%s\"",
4541 : "max_wal_size", "wal_segment_size")));
4542 :
4543 1744 : UsableBytesInSegment =
4544 1744 : (wal_segment_size / XLOG_BLCKSZ * UsableBytesInPage) -
4545 : (SizeOfXLogLongPHD - SizeOfXLogShortPHD);
4546 :
4547 1744 : CalculateCheckpointSegments();
4548 :
4549 : /* Make the initdb settings visible as GUC variables, too */
4550 1744 : SetConfigOption("data_checksums", DataChecksumsEnabled() ? "yes" : "no",
4551 : PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
4552 1744 : }
4553 :
4554 : /*
4555 : * Utility wrapper to update the control file. Note that the control
4556 : * file gets flushed.
4557 : */
4558 : static void
4559 15824 : UpdateControlFile(void)
4560 : {
4561 15824 : update_controlfile(DataDir, ControlFile, true);
4562 15824 : }
4563 :
4564 : /*
4565 : * Returns the unique system identifier from control file.
4566 : */
4567 : uint64
4568 2620 : GetSystemIdentifier(void)
4569 : {
4570 : Assert(ControlFile != NULL);
4571 2620 : return ControlFile->system_identifier;
4572 : }
4573 :
4574 : /*
4575 : * Returns the random nonce from control file.
4576 : */
4577 : char *
4578 2 : GetMockAuthenticationNonce(void)
4579 : {
4580 : Assert(ControlFile != NULL);
4581 2 : return ControlFile->mock_authentication_nonce;
4582 : }
4583 :
4584 : /*
4585 : * Are checksums enabled for data pages?
4586 : */
4587 : bool
4588 21622724 : DataChecksumsEnabled(void)
4589 : {
4590 : Assert(ControlFile != NULL);
4591 21622724 : return (ControlFile->data_checksum_version > 0);
4592 : }
4593 :
4594 : /*
4595 : * Returns a fake LSN for unlogged relations.
4596 : *
4597 : * Each call generates an LSN that is greater than any previous value
4598 : * returned. The current counter value is saved and restored across clean
4599 : * shutdowns, but like unlogged relations, does not survive a crash. This can
4600 : * be used in lieu of real LSN values returned by XLogInsert, if you need an
4601 : * LSN-like increasing sequence of numbers without writing any WAL.
4602 : */
4603 : XLogRecPtr
4604 66 : GetFakeLSNForUnloggedRel(void)
4605 : {
4606 66 : return pg_atomic_fetch_add_u64(&XLogCtl->unloggedLSN, 1);
4607 : }
4608 :
4609 : /*
4610 : * Auto-tune the number of XLOG buffers.
4611 : *
4612 : * The preferred setting for wal_buffers is about 3% of shared_buffers, with
4613 : * a maximum of one XLOG segment (there is little reason to think that more
4614 : * is helpful, at least so long as we force an fsync when switching log files)
4615 : * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
4616 : * 9.1, when auto-tuning was added).
4617 : *
4618 : * This should not be called until NBuffers has received its final value.
4619 : */
4620 : static int
4621 1904 : XLOGChooseNumBuffers(void)
4622 : {
4623 : int xbuffers;
4624 :
4625 1904 : xbuffers = NBuffers / 32;
4626 1904 : if (xbuffers > (wal_segment_size / XLOG_BLCKSZ))
4627 42 : xbuffers = (wal_segment_size / XLOG_BLCKSZ);
4628 1904 : if (xbuffers < 8)
4629 748 : xbuffers = 8;
4630 1904 : return xbuffers;
4631 : }
4632 :
4633 : /*
4634 : * GUC check_hook for wal_buffers
4635 : */
4636 : bool
4637 3876 : check_wal_buffers(int *newval, void **extra, GucSource source)
4638 : {
4639 : /*
4640 : * -1 indicates a request for auto-tune.
4641 : */
4642 3876 : if (*newval == -1)
4643 : {
4644 : /*
4645 : * If we haven't yet changed the boot_val default of -1, just let it
4646 : * be. We'll fix it when XLOGShmemSize is called.
4647 : */
4648 1972 : if (XLOGbuffers == -1)
4649 1972 : return true;
4650 :
4651 : /* Otherwise, substitute the auto-tune value */
4652 0 : *newval = XLOGChooseNumBuffers();
4653 : }
4654 :
4655 : /*
4656 : * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
4657 : * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
4658 : * the case, we just silently treat such values as a request for the
4659 : * minimum. (We could throw an error instead, but that doesn't seem very
4660 : * helpful.)
4661 : */
4662 1904 : if (*newval < 4)
4663 0 : *newval = 4;
4664 :
4665 1904 : return true;
4666 : }
4667 :
4668 : /*
4669 : * GUC check_hook for wal_consistency_checking
4670 : */
4671 : bool
4672 3752 : check_wal_consistency_checking(char **newval, void **extra, GucSource source)
4673 : {
4674 : char *rawstring;
4675 : List *elemlist;
4676 : ListCell *l;
4677 : bool newwalconsistency[RM_MAX_ID + 1];
4678 :
4679 : /* Initialize the array */
4680 123816 : MemSet(newwalconsistency, 0, (RM_MAX_ID + 1) * sizeof(bool));
4681 :
4682 : /* Need a modifiable copy of string */
4683 3752 : rawstring = pstrdup(*newval);
4684 :
4685 : /* Parse string into list of identifiers */
4686 3752 : if (!SplitIdentifierString(rawstring, ',', &elemlist))
4687 : {
4688 : /* syntax error in list */
4689 0 : GUC_check_errdetail("List syntax is invalid.");
4690 0 : pfree(rawstring);
4691 0 : list_free(elemlist);
4692 0 : return false;
4693 : }
4694 :
4695 4648 : foreach(l, elemlist)
4696 : {
4697 896 : char *tok = (char *) lfirst(l);
4698 : int rmid;
4699 :
4700 : /* Check for 'all'. */
4701 896 : if (pg_strcasecmp(tok, "all") == 0)
4702 : {
4703 229244 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4704 228352 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL)
4705 8920 : newwalconsistency[rmid] = true;
4706 : }
4707 : else
4708 : {
4709 : /* Check if the token matches any known resource manager. */
4710 4 : bool found = false;
4711 :
4712 72 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4713 : {
4714 108 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL &&
4715 36 : pg_strcasecmp(tok, GetRmgr(rmid).rm_name) == 0)
4716 : {
4717 4 : newwalconsistency[rmid] = true;
4718 4 : found = true;
4719 4 : break;
4720 : }
4721 : }
4722 4 : if (!found)
4723 : {
4724 : /*
4725 : * During startup, it might be a not-yet-loaded custom
4726 : * resource manager. Defer checking until
4727 : * InitializeWalConsistencyChecking().
4728 : */
4729 0 : if (!process_shared_preload_libraries_done)
4730 : {
4731 0 : check_wal_consistency_checking_deferred = true;
4732 : }
4733 : else
4734 : {
4735 0 : GUC_check_errdetail("Unrecognized key word: \"%s\".", tok);
4736 0 : pfree(rawstring);
4737 0 : list_free(elemlist);
4738 0 : return false;
4739 : }
4740 : }
4741 : }
4742 : }
4743 :
4744 3752 : pfree(rawstring);
4745 3752 : list_free(elemlist);
4746 :
4747 : /* assign new value */
4748 3752 : *extra = guc_malloc(ERROR, (RM_MAX_ID + 1) * sizeof(bool));
4749 3752 : memcpy(*extra, newwalconsistency, (RM_MAX_ID + 1) * sizeof(bool));
4750 3752 : return true;
4751 : }
4752 :
4753 : /*
4754 : * GUC assign_hook for wal_consistency_checking
4755 : */
4756 : void
4757 3750 : assign_wal_consistency_checking(const char *newval, void *extra)
4758 : {
4759 : /*
4760 : * If some checks were deferred, it's possible that the checks will fail
4761 : * later during InitializeWalConsistencyChecking(). But in that case, the
4762 : * postmaster will exit anyway, so it's safe to proceed with the
4763 : * assignment.
4764 : *
4765 : * Any built-in resource managers specified are assigned immediately,
4766 : * which affects WAL created before shared_preload_libraries are
4767 : * processed. Any custom resource managers specified won't be assigned
4768 : * until after shared_preload_libraries are processed, but that's OK
4769 : * because WAL for a custom resource manager can't be written before the
4770 : * module is loaded anyway.
4771 : */
4772 3750 : wal_consistency_checking = extra;
4773 3750 : }
4774 :
4775 : /*
4776 : * InitializeWalConsistencyChecking: run after loading custom resource managers
4777 : *
4778 : * If any unknown resource managers were specified in the
4779 : * wal_consistency_checking GUC, processing was deferred. Now that
4780 : * shared_preload_libraries have been loaded, process wal_consistency_checking
4781 : * again.
4782 : */
4783 : void
4784 1634 : InitializeWalConsistencyChecking(void)
4785 : {
4786 : Assert(process_shared_preload_libraries_done);
4787 :
4788 1634 : if (check_wal_consistency_checking_deferred)
4789 : {
4790 : struct config_generic *guc;
4791 :
4792 0 : guc = find_option("wal_consistency_checking", false, false, ERROR);
4793 :
4794 0 : check_wal_consistency_checking_deferred = false;
4795 :
4796 0 : set_config_option_ext("wal_consistency_checking",
4797 : wal_consistency_checking_string,
4798 : guc->scontext, guc->source, guc->srole,
4799 : GUC_ACTION_SET, true, ERROR, false);
4800 :
4801 : /* checking should not be deferred again */
4802 : Assert(!check_wal_consistency_checking_deferred);
4803 : }
4804 1634 : }
4805 :
4806 : /*
4807 : * GUC show_hook for archive_command
4808 : */
4809 : const char *
4810 3404 : show_archive_command(void)
4811 : {
4812 3404 : if (XLogArchivingActive())
4813 0 : return XLogArchiveCommand;
4814 : else
4815 3404 : return "(disabled)";
4816 : }
4817 :
4818 : /*
4819 : * GUC show_hook for in_hot_standby
4820 : */
4821 : const char *
4822 26866 : show_in_hot_standby(void)
4823 : {
4824 : /*
4825 : * We display the actual state based on shared memory, so that this GUC
4826 : * reports up-to-date state if examined intra-query. The underlying
4827 : * variable (in_hot_standby_guc) changes only when we transmit a new value
4828 : * to the client.
4829 : */
4830 26866 : return RecoveryInProgress() ? "on" : "off";
4831 : }
4832 :
4833 : /*
4834 : * Read the control file, set respective GUCs.
4835 : *
4836 : * This is to be called during startup, including a crash recovery cycle,
4837 : * unless in bootstrap mode, where no control file yet exists. As there's no
4838 : * usable shared memory yet (its sizing can depend on the contents of the
4839 : * control file!), first store the contents in local memory. XLOGShmemInit()
4840 : * will then copy it to shared memory later.
4841 : *
4842 : * reset just controls whether previous contents are to be expected (in the
4843 : * reset case, there's a dangling pointer into old shared memory), or not.
4844 : */
4845 : void
4846 1654 : LocalProcessControlFile(bool reset)
4847 : {
4848 : Assert(reset || ControlFile == NULL);
4849 1654 : ControlFile = palloc(sizeof(ControlFileData));
4850 1654 : ReadControlFile();
4851 1654 : }
4852 :
4853 : /*
4854 : * Get the wal_level from the control file. For a standby, this value should be
4855 : * considered as its active wal_level, because it may be different from what
4856 : * was originally configured on standby.
4857 : */
4858 : WalLevel
4859 132 : GetActiveWalLevelOnStandby(void)
4860 : {
4861 132 : return ControlFile->wal_level;
4862 : }
4863 :
4864 : /*
4865 : * Initialization of shared memory for XLOG
4866 : */
4867 : Size
4868 5454 : XLOGShmemSize(void)
4869 : {
4870 : Size size;
4871 :
4872 : /*
4873 : * If the value of wal_buffers is -1, use the preferred auto-tune value.
4874 : * This isn't an amazingly clean place to do this, but we must wait till
4875 : * NBuffers has received its final value, and must do it before using the
4876 : * value of XLOGbuffers to do anything important.
4877 : *
4878 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
4879 : * However, if the DBA explicitly set wal_buffers = -1 in the config file,
4880 : * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
4881 : * the matter with PGC_S_OVERRIDE.
4882 : */
4883 5454 : if (XLOGbuffers == -1)
4884 : {
4885 : char buf[32];
4886 :
4887 1904 : snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
4888 1904 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4889 : PGC_S_DYNAMIC_DEFAULT);
4890 1904 : if (XLOGbuffers == -1) /* failed to apply it? */
4891 0 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4892 : PGC_S_OVERRIDE);
4893 : }
4894 : Assert(XLOGbuffers > 0);
4895 :
4896 : /* XLogCtl */
4897 5454 : size = sizeof(XLogCtlData);
4898 :
4899 : /* WAL insertion locks, plus alignment */
4900 5454 : size = add_size(size, mul_size(sizeof(WALInsertLockPadded), NUM_XLOGINSERT_LOCKS + 1));
4901 : /* xlblocks array */
4902 5454 : size = add_size(size, mul_size(sizeof(pg_atomic_uint64), XLOGbuffers));
4903 : /* extra alignment padding for XLOG I/O buffers */
4904 5454 : size = add_size(size, Max(XLOG_BLCKSZ, PG_IO_ALIGN_SIZE));
4905 : /* and the buffers themselves */
4906 5454 : size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
4907 :
4908 : /*
4909 : * Note: we don't count ControlFileData, it comes out of the "slop factor"
4910 : * added by CreateSharedMemoryAndSemaphores. This lets us use this
4911 : * routine again below to compute the actual allocation size.
4912 : */
4913 :
4914 5454 : return size;
4915 : }
4916 :
4917 : void
4918 1908 : XLOGShmemInit(void)
4919 : {
4920 : bool foundCFile,
4921 : foundXLog;
4922 : char *allocptr;
4923 : int i;
4924 : ControlFileData *localControlFile;
4925 :
4926 : #ifdef WAL_DEBUG
4927 :
4928 : /*
4929 : * Create a memory context for WAL debugging that's exempt from the normal
4930 : * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
4931 : * an allocation fails, but wal_debug is not for production use anyway.
4932 : */
4933 : if (walDebugCxt == NULL)
4934 : {
4935 : walDebugCxt = AllocSetContextCreate(TopMemoryContext,
4936 : "WAL Debug",
4937 : ALLOCSET_DEFAULT_SIZES);
4938 : MemoryContextAllowInCriticalSection(walDebugCxt, true);
4939 : }
4940 : #endif
4941 :
4942 :
4943 1908 : XLogCtl = (XLogCtlData *)
4944 1908 : ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog);
4945 :
4946 1908 : localControlFile = ControlFile;
4947 1908 : ControlFile = (ControlFileData *)
4948 1908 : ShmemInitStruct("Control File", sizeof(ControlFileData), &foundCFile);
4949 :
4950 1908 : if (foundCFile || foundXLog)
4951 : {
4952 : /* both should be present or neither */
4953 : Assert(foundCFile && foundXLog);
4954 :
4955 : /* Initialize local copy of WALInsertLocks */
4956 0 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks;
4957 :
4958 0 : if (localControlFile)
4959 0 : pfree(localControlFile);
4960 0 : return;
4961 : }
4962 1908 : memset(XLogCtl, 0, sizeof(XLogCtlData));
4963 :
4964 : /*
4965 : * Already have read control file locally, unless in bootstrap mode. Move
4966 : * contents into shared memory.
4967 : */
4968 1908 : if (localControlFile)
4969 : {
4970 1638 : memcpy(ControlFile, localControlFile, sizeof(ControlFileData));
4971 1638 : pfree(localControlFile);
4972 : }
4973 :
4974 : /*
4975 : * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
4976 : * multiple of the alignment for same, so no extra alignment padding is
4977 : * needed here.
4978 : */
4979 1908 : allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
4980 1908 : XLogCtl->xlblocks = (pg_atomic_uint64 *) allocptr;
4981 1908 : allocptr += sizeof(pg_atomic_uint64) * XLOGbuffers;
4982 :
4983 539370 : for (i = 0; i < XLOGbuffers; i++)
4984 : {
4985 537462 : pg_atomic_init_u64(&XLogCtl->xlblocks[i], InvalidXLogRecPtr);
4986 : }
4987 :
4988 : /* WAL insertion locks. Ensure they're aligned to the full padded size */
4989 1908 : allocptr += sizeof(WALInsertLockPadded) -
4990 1908 : ((uintptr_t) allocptr) % sizeof(WALInsertLockPadded);
4991 1908 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks =
4992 : (WALInsertLockPadded *) allocptr;
4993 1908 : allocptr += sizeof(WALInsertLockPadded) * NUM_XLOGINSERT_LOCKS;
4994 :
4995 17172 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
4996 : {
4997 15264 : LWLockInitialize(&WALInsertLocks[i].l.lock, LWTRANCHE_WAL_INSERT);
4998 15264 : pg_atomic_init_u64(&WALInsertLocks[i].l.insertingAt, InvalidXLogRecPtr);
4999 15264 : WALInsertLocks[i].l.lastImportantAt = InvalidXLogRecPtr;
5000 : }
5001 :
5002 : /*
5003 : * Align the start of the page buffers to a full xlog block size boundary.
5004 : * This simplifies some calculations in XLOG insertion. It is also
5005 : * required for O_DIRECT.
5006 : */
5007 1908 : allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5008 1908 : XLogCtl->pages = allocptr;
5009 1908 : memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5010 :
5011 : /*
5012 : * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5013 : * in additional info.)
5014 : */
5015 1908 : XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5016 1908 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5017 1908 : XLogCtl->InstallXLogFileSegmentActive = false;
5018 1908 : XLogCtl->WalWriterSleeping = false;
5019 :
5020 1908 : SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5021 1908 : SpinLockInit(&XLogCtl->info_lck);
5022 1908 : pg_atomic_init_u64(&XLogCtl->logInsertResult, InvalidXLogRecPtr);
5023 1908 : pg_atomic_init_u64(&XLogCtl->logWriteResult, InvalidXLogRecPtr);
5024 1908 : pg_atomic_init_u64(&XLogCtl->logFlushResult, InvalidXLogRecPtr);
5025 1908 : pg_atomic_init_u64(&XLogCtl->unloggedLSN, InvalidXLogRecPtr);
5026 : }
5027 :
5028 : /*
5029 : * This func must be called ONCE on system install. It creates pg_control
5030 : * and the initial XLOG segment.
5031 : */
5032 : void
5033 90 : BootStrapXLOG(uint32 data_checksum_version)
5034 : {
5035 : CheckPoint checkPoint;
5036 : char *buffer;
5037 : XLogPageHeader page;
5038 : XLogLongPageHeader longpage;
5039 : XLogRecord *record;
5040 : char *recptr;
5041 : uint64 sysidentifier;
5042 : struct timeval tv;
5043 : pg_crc32c crc;
5044 :
5045 : /* allow ordinary WAL segment creation, like StartupXLOG() would */
5046 90 : SetInstallXLogFileSegmentActive();
5047 :
5048 : /*
5049 : * Select a hopefully-unique system identifier code for this installation.
5050 : * We use the result of gettimeofday(), including the fractional seconds
5051 : * field, as being about as unique as we can easily get. (Think not to
5052 : * use random(), since it hasn't been seeded and there's no portable way
5053 : * to seed it other than the system clock value...) The upper half of the
5054 : * uint64 value is just the tv_sec part, while the lower half contains the
5055 : * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
5056 : * PID for a little extra uniqueness. A person knowing this encoding can
5057 : * determine the initialization time of the installation, which could
5058 : * perhaps be useful sometimes.
5059 : */
5060 90 : gettimeofday(&tv, NULL);
5061 90 : sysidentifier = ((uint64) tv.tv_sec) << 32;
5062 90 : sysidentifier |= ((uint64) tv.tv_usec) << 12;
5063 90 : sysidentifier |= getpid() & 0xFFF;
5064 :
5065 : /* page buffer must be aligned suitably for O_DIRECT */
5066 90 : buffer = (char *) palloc(XLOG_BLCKSZ + XLOG_BLCKSZ);
5067 90 : page = (XLogPageHeader) TYPEALIGN(XLOG_BLCKSZ, buffer);
5068 90 : memset(page, 0, XLOG_BLCKSZ);
5069 :
5070 : /*
5071 : * Set up information for the initial checkpoint record
5072 : *
5073 : * The initial checkpoint record is written to the beginning of the WAL
5074 : * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5075 : * used, so that we can use 0/0 to mean "before any valid WAL segment".
5076 : */
5077 90 : checkPoint.redo = wal_segment_size + SizeOfXLogLongPHD;
5078 90 : checkPoint.ThisTimeLineID = BootstrapTimeLineID;
5079 90 : checkPoint.PrevTimeLineID = BootstrapTimeLineID;
5080 90 : checkPoint.fullPageWrites = fullPageWrites;
5081 90 : checkPoint.wal_level = wal_level;
5082 : checkPoint.nextXid =
5083 90 : FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId);
5084 90 : checkPoint.nextOid = FirstGenbkiObjectId;
5085 90 : checkPoint.nextMulti = FirstMultiXactId;
5086 90 : checkPoint.nextMultiOffset = 0;
5087 90 : checkPoint.oldestXid = FirstNormalTransactionId;
5088 90 : checkPoint.oldestXidDB = Template1DbOid;
5089 90 : checkPoint.oldestMulti = FirstMultiXactId;
5090 90 : checkPoint.oldestMultiDB = Template1DbOid;
5091 90 : checkPoint.oldestCommitTsXid = InvalidTransactionId;
5092 90 : checkPoint.newestCommitTsXid = InvalidTransactionId;
5093 90 : checkPoint.time = (pg_time_t) time(NULL);
5094 90 : checkPoint.oldestActiveXid = InvalidTransactionId;
5095 :
5096 90 : TransamVariables->nextXid = checkPoint.nextXid;
5097 90 : TransamVariables->nextOid = checkPoint.nextOid;
5098 90 : TransamVariables->oidCount = 0;
5099 90 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5100 90 : AdvanceOldestClogXid(checkPoint.oldestXid);
5101 90 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5102 90 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5103 90 : SetCommitTsLimit(InvalidTransactionId, InvalidTransactionId);
5104 :
5105 : /* Set up the XLOG page header */
5106 90 : page->xlp_magic = XLOG_PAGE_MAGIC;
5107 90 : page->xlp_info = XLP_LONG_HEADER;
5108 90 : page->xlp_tli = BootstrapTimeLineID;
5109 90 : page->xlp_pageaddr = wal_segment_size;
5110 90 : longpage = (XLogLongPageHeader) page;
5111 90 : longpage->xlp_sysid = sysidentifier;
5112 90 : longpage->xlp_seg_size = wal_segment_size;
5113 90 : longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5114 :
5115 : /* Insert the initial checkpoint record */
5116 90 : recptr = ((char *) page + SizeOfXLogLongPHD);
5117 90 : record = (XLogRecord *) recptr;
5118 90 : record->xl_prev = 0;
5119 90 : record->xl_xid = InvalidTransactionId;
5120 90 : record->xl_tot_len = SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(checkPoint);
5121 90 : record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5122 90 : record->xl_rmid = RM_XLOG_ID;
5123 90 : recptr += SizeOfXLogRecord;
5124 : /* fill the XLogRecordDataHeaderShort struct */
5125 90 : *(recptr++) = (char) XLR_BLOCK_ID_DATA_SHORT;
5126 90 : *(recptr++) = sizeof(checkPoint);
5127 90 : memcpy(recptr, &checkPoint, sizeof(checkPoint));
5128 90 : recptr += sizeof(checkPoint);
5129 : Assert(recptr - (char *) record == record->xl_tot_len);
5130 :
5131 90 : INIT_CRC32C(crc);
5132 90 : COMP_CRC32C(crc, ((char *) record) + SizeOfXLogRecord, record->xl_tot_len - SizeOfXLogRecord);
5133 90 : COMP_CRC32C(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5134 90 : FIN_CRC32C(crc);
5135 90 : record->xl_crc = crc;
5136 :
5137 : /* Create first XLOG segment file */
5138 90 : openLogTLI = BootstrapTimeLineID;
5139 90 : openLogFile = XLogFileInit(1, BootstrapTimeLineID);
5140 :
5141 : /*
5142 : * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
5143 : * close the file again in a moment.
5144 : */
5145 :
5146 : /* Write the first page with the initial record */
5147 90 : errno = 0;
5148 90 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE);
5149 90 : if (write(openLogFile, page, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5150 : {
5151 : /* if write didn't set errno, assume problem is no disk space */
5152 0 : if (errno == 0)
5153 0 : errno = ENOSPC;
5154 0 : ereport(PANIC,
5155 : (errcode_for_file_access(),
5156 : errmsg("could not write bootstrap write-ahead log file: %m")));
5157 : }
5158 90 : pgstat_report_wait_end();
5159 :
5160 90 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC);
5161 90 : if (pg_fsync(openLogFile) != 0)
5162 0 : ereport(PANIC,
5163 : (errcode_for_file_access(),
5164 : errmsg("could not fsync bootstrap write-ahead log file: %m")));
5165 90 : pgstat_report_wait_end();
5166 :
5167 90 : if (close(openLogFile) != 0)
5168 0 : ereport(PANIC,
5169 : (errcode_for_file_access(),
5170 : errmsg("could not close bootstrap write-ahead log file: %m")));
5171 :
5172 90 : openLogFile = -1;
5173 :
5174 : /* Now create pg_control */
5175 90 : InitControlFile(sysidentifier, data_checksum_version);
5176 90 : ControlFile->time = checkPoint.time;
5177 90 : ControlFile->checkPoint = checkPoint.redo;
5178 90 : ControlFile->checkPointCopy = checkPoint;
5179 :
5180 : /* some additional ControlFile fields are set in WriteControlFile() */
5181 90 : WriteControlFile();
5182 :
5183 : /* Bootstrap the commit log, too */
5184 90 : BootStrapCLOG();
5185 90 : BootStrapCommitTs();
5186 90 : BootStrapSUBTRANS();
5187 90 : BootStrapMultiXact();
5188 :
5189 90 : pfree(buffer);
5190 :
5191 : /*
5192 : * Force control file to be read - in contrast to normal processing we'd
5193 : * otherwise never run the checks and GUC related initializations therein.
5194 : */
5195 90 : ReadControlFile();
5196 90 : }
5197 :
5198 : static char *
5199 1460 : str_time(pg_time_t tnow)
5200 : {
5201 1460 : char *buf = palloc(128);
5202 :
5203 1460 : pg_strftime(buf, 128,
5204 : "%Y-%m-%d %H:%M:%S %Z",
5205 1460 : pg_localtime(&tnow, log_timezone));
5206 :
5207 1460 : return buf;
5208 : }
5209 :
5210 : /*
5211 : * Initialize the first WAL segment on new timeline.
5212 : */
5213 : static void
5214 96 : XLogInitNewTimeline(TimeLineID endTLI, XLogRecPtr endOfLog, TimeLineID newTLI)
5215 : {
5216 : char xlogfname[MAXFNAMELEN];
5217 : XLogSegNo endLogSegNo;
5218 : XLogSegNo startLogSegNo;
5219 :
5220 : /* we always switch to a new timeline after archive recovery */
5221 : Assert(endTLI != newTLI);
5222 :
5223 : /*
5224 : * Update min recovery point one last time.
5225 : */
5226 96 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5227 :
5228 : /*
5229 : * Calculate the last segment on the old timeline, and the first segment
5230 : * on the new timeline. If the switch happens in the middle of a segment,
5231 : * they are the same, but if the switch happens exactly at a segment
5232 : * boundary, startLogSegNo will be endLogSegNo + 1.
5233 : */
5234 96 : XLByteToPrevSeg(endOfLog, endLogSegNo, wal_segment_size);
5235 96 : XLByteToSeg(endOfLog, startLogSegNo, wal_segment_size);
5236 :
5237 : /*
5238 : * Initialize the starting WAL segment for the new timeline. If the switch
5239 : * happens in the middle of a segment, copy data from the last WAL segment
5240 : * of the old timeline up to the switch point, to the starting WAL segment
5241 : * on the new timeline.
5242 : */
5243 96 : if (endLogSegNo == startLogSegNo)
5244 : {
5245 : /*
5246 : * Make a copy of the file on the new timeline.
5247 : *
5248 : * Writing WAL isn't allowed yet, so there are no locking
5249 : * considerations. But we should be just as tense as XLogFileInit to
5250 : * avoid emplacing a bogus file.
5251 : */
5252 78 : XLogFileCopy(newTLI, endLogSegNo, endTLI, endLogSegNo,
5253 78 : XLogSegmentOffset(endOfLog, wal_segment_size));
5254 : }
5255 : else
5256 : {
5257 : /*
5258 : * The switch happened at a segment boundary, so just create the next
5259 : * segment on the new timeline.
5260 : */
5261 : int fd;
5262 :
5263 18 : fd = XLogFileInit(startLogSegNo, newTLI);
5264 :
5265 18 : if (close(fd) != 0)
5266 : {
5267 0 : int save_errno = errno;
5268 :
5269 0 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5270 0 : errno = save_errno;
5271 0 : ereport(ERROR,
5272 : (errcode_for_file_access(),
5273 : errmsg("could not close file \"%s\": %m", xlogfname)));
5274 : }
5275 : }
5276 :
5277 : /*
5278 : * Let's just make real sure there are not .ready or .done flags posted
5279 : * for the new segment.
5280 : */
5281 96 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5282 96 : XLogArchiveCleanup(xlogfname);
5283 96 : }
5284 :
5285 : /*
5286 : * Perform cleanup actions at the conclusion of archive recovery.
5287 : */
5288 : static void
5289 96 : CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI, XLogRecPtr EndOfLog,
5290 : TimeLineID newTLI)
5291 : {
5292 : /*
5293 : * Execute the recovery_end_command, if any.
5294 : */
5295 96 : if (recoveryEndCommand && strcmp(recoveryEndCommand, "") != 0)
5296 4 : ExecuteRecoveryCommand(recoveryEndCommand,
5297 : "recovery_end_command",
5298 : true,
5299 : WAIT_EVENT_RECOVERY_END_COMMAND);
5300 :
5301 : /*
5302 : * We switched to a new timeline. Clean up segments on the old timeline.
5303 : *
5304 : * If there are any higher-numbered segments on the old timeline, remove
5305 : * them. They might contain valid WAL, but they might also be
5306 : * pre-allocated files containing garbage. In any case, they are not part
5307 : * of the new timeline's history so we don't need them.
5308 : */
5309 96 : RemoveNonParentXlogFiles(EndOfLog, newTLI);
5310 :
5311 : /*
5312 : * If the switch happened in the middle of a segment, what to do with the
5313 : * last, partial segment on the old timeline? If we don't archive it, and
5314 : * the server that created the WAL never archives it either (e.g. because
5315 : * it was hit by a meteor), it will never make it to the archive. That's
5316 : * OK from our point of view, because the new segment that we created with
5317 : * the new TLI contains all the WAL from the old timeline up to the switch
5318 : * point. But if you later try to do PITR to the "missing" WAL on the old
5319 : * timeline, recovery won't find it in the archive. It's physically
5320 : * present in the new file with new TLI, but recovery won't look there
5321 : * when it's recovering to the older timeline. On the other hand, if we
5322 : * archive the partial segment, and the original server on that timeline
5323 : * is still running and archives the completed version of the same segment
5324 : * later, it will fail. (We used to do that in 9.4 and below, and it
5325 : * caused such problems).
5326 : *
5327 : * As a compromise, we rename the last segment with the .partial suffix,
5328 : * and archive it. Archive recovery will never try to read .partial
5329 : * segments, so they will normally go unused. But in the odd PITR case,
5330 : * the administrator can copy them manually to the pg_wal directory
5331 : * (removing the suffix). They can be useful in debugging, too.
5332 : *
5333 : * If a .done or .ready file already exists for the old timeline, however,
5334 : * we had already determined that the segment is complete, so we can let
5335 : * it be archived normally. (In particular, if it was restored from the
5336 : * archive to begin with, it's expected to have a .done file).
5337 : */
5338 96 : if (XLogSegmentOffset(EndOfLog, wal_segment_size) != 0 &&
5339 : XLogArchivingActive())
5340 : {
5341 : char origfname[MAXFNAMELEN];
5342 : XLogSegNo endLogSegNo;
5343 :
5344 16 : XLByteToPrevSeg(EndOfLog, endLogSegNo, wal_segment_size);
5345 16 : XLogFileName(origfname, EndOfLogTLI, endLogSegNo, wal_segment_size);
5346 :
5347 16 : if (!XLogArchiveIsReadyOrDone(origfname))
5348 : {
5349 : char origpath[MAXPGPATH];
5350 : char partialfname[MAXFNAMELEN];
5351 : char partialpath[MAXPGPATH];
5352 :
5353 : /*
5354 : * If we're summarizing WAL, we can't rename the partial file
5355 : * until the summarizer finishes with it, else it will fail.
5356 : */
5357 10 : if (summarize_wal)
5358 2 : WaitForWalSummarization(EndOfLog);
5359 :
5360 10 : XLogFilePath(origpath, EndOfLogTLI, endLogSegNo, wal_segment_size);
5361 10 : snprintf(partialfname, MAXFNAMELEN, "%s.partial", origfname);
5362 10 : snprintf(partialpath, MAXPGPATH, "%s.partial", origpath);
5363 :
5364 : /*
5365 : * Make sure there's no .done or .ready file for the .partial
5366 : * file.
5367 : */
5368 10 : XLogArchiveCleanup(partialfname);
5369 :
5370 10 : durable_rename(origpath, partialpath, ERROR);
5371 10 : XLogArchiveNotify(partialfname);
5372 : }
5373 : }
5374 96 : }
5375 :
5376 : /*
5377 : * Check to see if required parameters are set high enough on this server
5378 : * for various aspects of recovery operation.
5379 : *
5380 : * Note that all the parameters which this function tests need to be
5381 : * listed in Administrator's Overview section in high-availability.sgml.
5382 : * If you change them, don't forget to update the list.
5383 : */
5384 : static void
5385 470 : CheckRequiredParameterValues(void)
5386 : {
5387 : /*
5388 : * For archive recovery, the WAL must be generated with at least 'replica'
5389 : * wal_level.
5390 : */
5391 470 : if (ArchiveRecoveryRequested && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5392 : {
5393 4 : ereport(FATAL,
5394 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5395 : errmsg("WAL was generated with \"wal_level=minimal\", cannot continue recovering"),
5396 : errdetail("This happens if you temporarily set \"wal_level=minimal\" on the server."),
5397 : errhint("Use a backup taken after setting \"wal_level\" to higher than \"minimal\".")));
5398 : }
5399 :
5400 : /*
5401 : * For Hot Standby, the WAL must be generated with 'replica' mode, and we
5402 : * must have at least as many backend slots as the primary.
5403 : */
5404 466 : if (ArchiveRecoveryRequested && EnableHotStandby)
5405 : {
5406 : /* We ignore autovacuum_max_workers when we make this test. */
5407 228 : RecoveryRequiresIntParameter("max_connections",
5408 : MaxConnections,
5409 228 : ControlFile->MaxConnections);
5410 228 : RecoveryRequiresIntParameter("max_worker_processes",
5411 : max_worker_processes,
5412 228 : ControlFile->max_worker_processes);
5413 228 : RecoveryRequiresIntParameter("max_wal_senders",
5414 : max_wal_senders,
5415 228 : ControlFile->max_wal_senders);
5416 228 : RecoveryRequiresIntParameter("max_prepared_transactions",
5417 : max_prepared_xacts,
5418 228 : ControlFile->max_prepared_xacts);
5419 228 : RecoveryRequiresIntParameter("max_locks_per_transaction",
5420 : max_locks_per_xact,
5421 228 : ControlFile->max_locks_per_xact);
5422 : }
5423 466 : }
5424 :
5425 : /*
5426 : * This must be called ONCE during postmaster or standalone-backend startup
5427 : */
5428 : void
5429 1640 : StartupXLOG(void)
5430 : {
5431 : XLogCtlInsert *Insert;
5432 : CheckPoint checkPoint;
5433 : bool wasShutdown;
5434 : bool didCrash;
5435 : bool haveTblspcMap;
5436 : bool haveBackupLabel;
5437 : XLogRecPtr EndOfLog;
5438 : TimeLineID EndOfLogTLI;
5439 : TimeLineID newTLI;
5440 : bool performedWalRecovery;
5441 : EndOfWalRecoveryInfo *endOfRecoveryInfo;
5442 : XLogRecPtr abortedRecPtr;
5443 : XLogRecPtr missingContrecPtr;
5444 : TransactionId oldestActiveXID;
5445 1640 : bool promoted = false;
5446 :
5447 : /*
5448 : * We should have an aux process resource owner to use, and we should not
5449 : * be in a transaction that's installed some other resowner.
5450 : */
5451 : Assert(AuxProcessResourceOwner != NULL);
5452 : Assert(CurrentResourceOwner == NULL ||
5453 : CurrentResourceOwner == AuxProcessResourceOwner);
5454 1640 : CurrentResourceOwner = AuxProcessResourceOwner;
5455 :
5456 : /*
5457 : * Check that contents look valid.
5458 : */
5459 1640 : if (!XRecOffIsValid(ControlFile->checkPoint))
5460 0 : ereport(FATAL,
5461 : (errcode(ERRCODE_DATA_CORRUPTED),
5462 : errmsg("control file contains invalid checkpoint location")));
5463 :
5464 1640 : switch (ControlFile->state)
5465 : {
5466 1244 : case DB_SHUTDOWNED:
5467 :
5468 : /*
5469 : * This is the expected case, so don't be chatty in standalone
5470 : * mode
5471 : */
5472 1244 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
5473 : (errmsg("database system was shut down at %s",
5474 : str_time(ControlFile->time))));
5475 1244 : break;
5476 :
5477 56 : case DB_SHUTDOWNED_IN_RECOVERY:
5478 56 : ereport(LOG,
5479 : (errmsg("database system was shut down in recovery at %s",
5480 : str_time(ControlFile->time))));
5481 56 : break;
5482 :
5483 0 : case DB_SHUTDOWNING:
5484 0 : ereport(LOG,
5485 : (errmsg("database system shutdown was interrupted; last known up at %s",
5486 : str_time(ControlFile->time))));
5487 0 : break;
5488 :
5489 0 : case DB_IN_CRASH_RECOVERY:
5490 0 : ereport(LOG,
5491 : (errmsg("database system was interrupted while in recovery at %s",
5492 : str_time(ControlFile->time)),
5493 : errhint("This probably means that some data is corrupted and"
5494 : " you will have to use the last backup for recovery.")));
5495 0 : break;
5496 :
5497 12 : case DB_IN_ARCHIVE_RECOVERY:
5498 12 : ereport(LOG,
5499 : (errmsg("database system was interrupted while in recovery at log time %s",
5500 : str_time(ControlFile->checkPointCopy.time)),
5501 : errhint("If this has occurred more than once some data might be corrupted"
5502 : " and you might need to choose an earlier recovery target.")));
5503 12 : break;
5504 :
5505 328 : case DB_IN_PRODUCTION:
5506 328 : ereport(LOG,
5507 : (errmsg("database system was interrupted; last known up at %s",
5508 : str_time(ControlFile->time))));
5509 328 : break;
5510 :
5511 0 : default:
5512 0 : ereport(FATAL,
5513 : (errcode(ERRCODE_DATA_CORRUPTED),
5514 : errmsg("control file contains invalid database cluster state")));
5515 : }
5516 :
5517 : /* This is just to allow attaching to startup process with a debugger */
5518 : #ifdef XLOG_REPLAY_DELAY
5519 : if (ControlFile->state != DB_SHUTDOWNED)
5520 : pg_usleep(60000000L);
5521 : #endif
5522 :
5523 : /*
5524 : * Verify that pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
5525 : * In cases where someone has performed a copy for PITR, these directories
5526 : * may have been excluded and need to be re-created.
5527 : */
5528 1640 : ValidateXLOGDirectoryStructure();
5529 :
5530 : /* Set up timeout handler needed to report startup progress. */
5531 1640 : if (!IsBootstrapProcessingMode())
5532 1550 : RegisterTimeout(STARTUP_PROGRESS_TIMEOUT,
5533 : startup_progress_timeout_handler);
5534 :
5535 : /*----------
5536 : * If we previously crashed, perform a couple of actions:
5537 : *
5538 : * - The pg_wal directory may still include some temporary WAL segments
5539 : * used when creating a new segment, so perform some clean up to not
5540 : * bloat this path. This is done first as there is no point to sync
5541 : * this temporary data.
5542 : *
5543 : * - There might be data which we had written, intending to fsync it, but
5544 : * which we had not actually fsync'd yet. Therefore, a power failure in
5545 : * the near future might cause earlier unflushed writes to be lost, even
5546 : * though more recent data written to disk from here on would be
5547 : * persisted. To avoid that, fsync the entire data directory.
5548 : */
5549 1640 : if (ControlFile->state != DB_SHUTDOWNED &&
5550 396 : ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
5551 : {
5552 340 : RemoveTempXlogFiles();
5553 340 : SyncDataDirectory();
5554 340 : didCrash = true;
5555 : }
5556 : else
5557 1300 : didCrash = false;
5558 :
5559 : /*
5560 : * Prepare for WAL recovery if needed.
5561 : *
5562 : * InitWalRecovery analyzes the control file and the backup label file, if
5563 : * any. It updates the in-memory ControlFile buffer according to the
5564 : * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5565 : * It also applies the tablespace map file, if any.
5566 : */
5567 1640 : InitWalRecovery(ControlFile, &wasShutdown,
5568 : &haveBackupLabel, &haveTblspcMap);
5569 1640 : checkPoint = ControlFile->checkPointCopy;
5570 :
5571 : /* initialize shared memory variables from the checkpoint record */
5572 1640 : TransamVariables->nextXid = checkPoint.nextXid;
5573 1640 : TransamVariables->nextOid = checkPoint.nextOid;
5574 1640 : TransamVariables->oidCount = 0;
5575 1640 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5576 1640 : AdvanceOldestClogXid(checkPoint.oldestXid);
5577 1640 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5578 1640 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5579 1640 : SetCommitTsLimit(checkPoint.oldestCommitTsXid,
5580 : checkPoint.newestCommitTsXid);
5581 1640 : XLogCtl->ckptFullXid = checkPoint.nextXid;
5582 :
5583 : /*
5584 : * Clear out any old relcache cache files. This is *necessary* if we do
5585 : * any WAL replay, since that would probably result in the cache files
5586 : * being out of sync with database reality. In theory we could leave them
5587 : * in place if the database had been cleanly shut down, but it seems
5588 : * safest to just remove them always and let them be rebuilt during the
5589 : * first backend startup. These files needs to be removed from all
5590 : * directories including pg_tblspc, however the symlinks are created only
5591 : * after reading tablespace_map file in case of archive recovery from
5592 : * backup, so needs to clear old relcache files here after creating
5593 : * symlinks.
5594 : */
5595 1640 : RelationCacheInitFileRemove();
5596 :
5597 : /*
5598 : * Initialize replication slots, before there's a chance to remove
5599 : * required resources.
5600 : */
5601 1640 : StartupReplicationSlots();
5602 :
5603 : /*
5604 : * Startup logical state, needs to be setup now so we have proper data
5605 : * during crash recovery.
5606 : */
5607 1640 : StartupReorderBuffer();
5608 :
5609 : /*
5610 : * Startup CLOG. This must be done after TransamVariables->nextXid has
5611 : * been initialized and before we accept connections or begin WAL replay.
5612 : */
5613 1640 : StartupCLOG();
5614 :
5615 : /*
5616 : * Startup MultiXact. We need to do this early to be able to replay
5617 : * truncations.
5618 : */
5619 1640 : StartupMultiXact();
5620 :
5621 : /*
5622 : * Ditto for commit timestamps. Activate the facility if the setting is
5623 : * enabled in the control file, as there should be no tracking of commit
5624 : * timestamps done when the setting was disabled. This facility can be
5625 : * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
5626 : */
5627 1640 : if (ControlFile->track_commit_timestamp)
5628 20 : StartupCommitTs();
5629 :
5630 : /*
5631 : * Recover knowledge about replay progress of known replication partners.
5632 : */
5633 1640 : StartupReplicationOrigin();
5634 :
5635 : /*
5636 : * Initialize unlogged LSN. On a clean shutdown, it's restored from the
5637 : * control file. On recovery, all unlogged relations are blown away, so
5638 : * the unlogged LSN counter can be reset too.
5639 : */
5640 1640 : if (ControlFile->state == DB_SHUTDOWNED)
5641 1230 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5642 1230 : ControlFile->unloggedLSN);
5643 : else
5644 410 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5645 : FirstNormalUnloggedLSN);
5646 :
5647 : /*
5648 : * Copy any missing timeline history files between 'now' and the recovery
5649 : * target timeline from archive to pg_wal. While we don't need those files
5650 : * ourselves - the history file of the recovery target timeline covers all
5651 : * the previous timelines in the history too - a cascading standby server
5652 : * might be interested in them. Or, if you archive the WAL from this
5653 : * server to a different archive than the primary, it'd be good for all
5654 : * the history files to get archived there after failover, so that you can
5655 : * use one of the old timelines as a PITR target. Timeline history files
5656 : * are small, so it's better to copy them unnecessarily than not copy them
5657 : * and regret later.
5658 : */
5659 1640 : restoreTimeLineHistoryFiles(checkPoint.ThisTimeLineID, recoveryTargetTLI);
5660 :
5661 : /*
5662 : * Before running in recovery, scan pg_twophase and fill in its status to
5663 : * be able to work on entries generated by redo. Doing a scan before
5664 : * taking any recovery action has the merit to discard any 2PC files that
5665 : * are newer than the first record to replay, saving from any conflicts at
5666 : * replay. This avoids as well any subsequent scans when doing recovery
5667 : * of the on-disk two-phase data.
5668 : */
5669 1640 : restoreTwoPhaseData();
5670 :
5671 : /*
5672 : * When starting with crash recovery, reset pgstat data - it might not be
5673 : * valid. Otherwise restore pgstat data. It's safe to do this here,
5674 : * because postmaster will not yet have started any other processes.
5675 : *
5676 : * NB: Restoring replication slot stats relies on slot state to have
5677 : * already been restored from disk.
5678 : *
5679 : * TODO: With a bit of extra work we could just start with a pgstat file
5680 : * associated with the checkpoint redo location we're starting from.
5681 : */
5682 1640 : if (didCrash)
5683 340 : pgstat_discard_stats();
5684 : else
5685 1300 : pgstat_restore_stats(checkPoint.redo);
5686 :
5687 1640 : lastFullPageWrites = checkPoint.fullPageWrites;
5688 :
5689 1640 : RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
5690 1640 : doPageWrites = lastFullPageWrites;
5691 :
5692 : /* REDO */
5693 1640 : if (InRecovery)
5694 : {
5695 : /* Initialize state for RecoveryInProgress() */
5696 410 : SpinLockAcquire(&XLogCtl->info_lck);
5697 410 : if (InArchiveRecovery)
5698 210 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
5699 : else
5700 200 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5701 410 : SpinLockRelease(&XLogCtl->info_lck);
5702 :
5703 : /*
5704 : * Update pg_control to show that we are recovering and to show the
5705 : * selected checkpoint as the place we are starting from. We also mark
5706 : * pg_control with any minimum recovery stop point obtained from a
5707 : * backup history file.
5708 : *
5709 : * No need to hold ControlFileLock yet, we aren't up far enough.
5710 : */
5711 410 : UpdateControlFile();
5712 :
5713 : /*
5714 : * If there was a backup label file, it's done its job and the info
5715 : * has now been propagated into pg_control. We must get rid of the
5716 : * label file so that if we crash during recovery, we'll pick up at
5717 : * the latest recovery restartpoint instead of going all the way back
5718 : * to the backup start point. It seems prudent though to just rename
5719 : * the file out of the way rather than delete it completely.
5720 : */
5721 410 : if (haveBackupLabel)
5722 : {
5723 136 : unlink(BACKUP_LABEL_OLD);
5724 136 : durable_rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD, FATAL);
5725 : }
5726 :
5727 : /*
5728 : * If there was a tablespace_map file, it's done its job and the
5729 : * symlinks have been created. We must get rid of the map file so
5730 : * that if we crash during recovery, we don't create symlinks again.
5731 : * It seems prudent though to just rename the file out of the way
5732 : * rather than delete it completely.
5733 : */
5734 410 : if (haveTblspcMap)
5735 : {
5736 4 : unlink(TABLESPACE_MAP_OLD);
5737 4 : durable_rename(TABLESPACE_MAP, TABLESPACE_MAP_OLD, FATAL);
5738 : }
5739 :
5740 : /*
5741 : * Initialize our local copy of minRecoveryPoint. When doing crash
5742 : * recovery we want to replay up to the end of WAL. Particularly, in
5743 : * the case of a promoted standby minRecoveryPoint value in the
5744 : * control file is only updated after the first checkpoint. However,
5745 : * if the instance crashes before the first post-recovery checkpoint
5746 : * is completed then recovery will use a stale location causing the
5747 : * startup process to think that there are still invalid page
5748 : * references when checking for data consistency.
5749 : */
5750 410 : if (InArchiveRecovery)
5751 : {
5752 210 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
5753 210 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
5754 : }
5755 : else
5756 : {
5757 200 : LocalMinRecoveryPoint = InvalidXLogRecPtr;
5758 200 : LocalMinRecoveryPointTLI = 0;
5759 : }
5760 :
5761 : /* Check that the GUCs used to generate the WAL allow recovery */
5762 410 : CheckRequiredParameterValues();
5763 :
5764 : /*
5765 : * We're in recovery, so unlogged relations may be trashed and must be
5766 : * reset. This should be done BEFORE allowing Hot Standby
5767 : * connections, so that read-only backends don't try to read whatever
5768 : * garbage is left over from before.
5769 : */
5770 410 : ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
5771 :
5772 : /*
5773 : * Likewise, delete any saved transaction snapshot files that got left
5774 : * behind by crashed backends.
5775 : */
5776 410 : DeleteAllExportedSnapshotFiles();
5777 :
5778 : /*
5779 : * Initialize for Hot Standby, if enabled. We won't let backends in
5780 : * yet, not until we've reached the min recovery point specified in
5781 : * control file and we've established a recovery snapshot from a
5782 : * running-xacts WAL record.
5783 : */
5784 410 : if (ArchiveRecoveryRequested && EnableHotStandby)
5785 : {
5786 : TransactionId *xids;
5787 : int nxids;
5788 :
5789 198 : ereport(DEBUG1,
5790 : (errmsg_internal("initializing for hot standby")));
5791 :
5792 198 : InitRecoveryTransactionEnvironment();
5793 :
5794 198 : if (wasShutdown)
5795 48 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
5796 : else
5797 150 : oldestActiveXID = checkPoint.oldestActiveXid;
5798 : Assert(TransactionIdIsValid(oldestActiveXID));
5799 :
5800 : /* Tell procarray about the range of xids it has to deal with */
5801 198 : ProcArrayInitRecovery(XidFromFullTransactionId(TransamVariables->nextXid));
5802 :
5803 : /*
5804 : * Startup subtrans only. CLOG, MultiXact and commit timestamp
5805 : * have already been started up and other SLRUs are not maintained
5806 : * during recovery and need not be started yet.
5807 : */
5808 198 : StartupSUBTRANS(oldestActiveXID);
5809 :
5810 : /*
5811 : * If we're beginning at a shutdown checkpoint, we know that
5812 : * nothing was running on the primary at this point. So fake-up an
5813 : * empty running-xacts record and use that here and now. Recover
5814 : * additional standby state for prepared transactions.
5815 : */
5816 198 : if (wasShutdown)
5817 : {
5818 : RunningTransactionsData running;
5819 : TransactionId latestCompletedXid;
5820 :
5821 : /* Update pg_subtrans entries for any prepared transactions */
5822 48 : StandbyRecoverPreparedTransactions();
5823 :
5824 : /*
5825 : * Construct a RunningTransactions snapshot representing a
5826 : * shut down server, with only prepared transactions still
5827 : * alive. We're never overflowed at this point because all
5828 : * subxids are listed with their parent prepared transactions.
5829 : */
5830 48 : running.xcnt = nxids;
5831 48 : running.subxcnt = 0;
5832 48 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
5833 48 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
5834 48 : running.oldestRunningXid = oldestActiveXID;
5835 48 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
5836 48 : TransactionIdRetreat(latestCompletedXid);
5837 : Assert(TransactionIdIsNormal(latestCompletedXid));
5838 48 : running.latestCompletedXid = latestCompletedXid;
5839 48 : running.xids = xids;
5840 :
5841 48 : ProcArrayApplyRecoveryInfo(&running);
5842 : }
5843 : }
5844 :
5845 : /*
5846 : * We're all set for replaying the WAL now. Do it.
5847 : */
5848 410 : PerformWalRecovery();
5849 304 : performedWalRecovery = true;
5850 : }
5851 : else
5852 1230 : performedWalRecovery = false;
5853 :
5854 : /*
5855 : * Finish WAL recovery.
5856 : */
5857 1534 : endOfRecoveryInfo = FinishWalRecovery();
5858 1534 : EndOfLog = endOfRecoveryInfo->endOfLog;
5859 1534 : EndOfLogTLI = endOfRecoveryInfo->endOfLogTLI;
5860 1534 : abortedRecPtr = endOfRecoveryInfo->abortedRecPtr;
5861 1534 : missingContrecPtr = endOfRecoveryInfo->missingContrecPtr;
5862 :
5863 : /*
5864 : * Reset ps status display, so as no information related to recovery shows
5865 : * up.
5866 : */
5867 1534 : set_ps_display("");
5868 :
5869 : /*
5870 : * When recovering from a backup (we are in recovery, and archive recovery
5871 : * was requested), complain if we did not roll forward far enough to reach
5872 : * the point where the database is consistent. For regular online
5873 : * backup-from-primary, that means reaching the end-of-backup WAL record
5874 : * (at which point we reset backupStartPoint to be Invalid), for
5875 : * backup-from-replica (which can't inject records into the WAL stream),
5876 : * that point is when we reach the minRecoveryPoint in pg_control (which
5877 : * we purposefully copy last when backing up from a replica). For
5878 : * pg_rewind (which creates a backup_label with a method of "pg_rewind")
5879 : * or snapshot-style backups (which don't), backupEndRequired will be set
5880 : * to false.
5881 : *
5882 : * Note: it is indeed okay to look at the local variable
5883 : * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
5884 : * might be further ahead --- ControlFile->minRecoveryPoint cannot have
5885 : * been advanced beyond the WAL we processed.
5886 : */
5887 1534 : if (InRecovery &&
5888 304 : (EndOfLog < LocalMinRecoveryPoint ||
5889 304 : !XLogRecPtrIsInvalid(ControlFile->backupStartPoint)))
5890 : {
5891 : /*
5892 : * Ran off end of WAL before reaching end-of-backup WAL record, or
5893 : * minRecoveryPoint. That's a bad sign, indicating that you tried to
5894 : * recover from an online backup but never called pg_backup_stop(), or
5895 : * you didn't archive all the WAL needed.
5896 : */
5897 0 : if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
5898 : {
5899 0 : if (!XLogRecPtrIsInvalid(ControlFile->backupStartPoint) || ControlFile->backupEndRequired)
5900 0 : ereport(FATAL,
5901 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5902 : errmsg("WAL ends before end of online backup"),
5903 : errhint("All WAL generated while online backup was taken must be available at recovery.")));
5904 : else
5905 0 : ereport(FATAL,
5906 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5907 : errmsg("WAL ends before consistent recovery point")));
5908 : }
5909 : }
5910 :
5911 : /*
5912 : * Reset unlogged relations to the contents of their INIT fork. This is
5913 : * done AFTER recovery is complete so as to include any unlogged relations
5914 : * created during recovery, but BEFORE recovery is marked as having
5915 : * completed successfully. Otherwise we'd not retry if any of the post
5916 : * end-of-recovery steps fail.
5917 : */
5918 1534 : if (InRecovery)
5919 304 : ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
5920 :
5921 : /*
5922 : * Pre-scan prepared transactions to find out the range of XIDs present.
5923 : * This information is not quite needed yet, but it is positioned here so
5924 : * as potential problems are detected before any on-disk change is done.
5925 : */
5926 1534 : oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
5927 :
5928 : /*
5929 : * Allow ordinary WAL segment creation before possibly switching to a new
5930 : * timeline, which creates a new segment, and after the last ReadRecord().
5931 : */
5932 1534 : SetInstallXLogFileSegmentActive();
5933 :
5934 : /*
5935 : * Consider whether we need to assign a new timeline ID.
5936 : *
5937 : * If we did archive recovery, we always assign a new ID. This handles a
5938 : * couple of issues. If we stopped short of the end of WAL during
5939 : * recovery, then we are clearly generating a new timeline and must assign
5940 : * it a unique new ID. Even if we ran to the end, modifying the current
5941 : * last segment is problematic because it may result in trying to
5942 : * overwrite an already-archived copy of that segment, and we encourage
5943 : * DBAs to make their archive_commands reject that. We can dodge the
5944 : * problem by making the new active segment have a new timeline ID.
5945 : *
5946 : * In a normal crash recovery, we can just extend the timeline we were in.
5947 : */
5948 1534 : newTLI = endOfRecoveryInfo->lastRecTLI;
5949 1534 : if (ArchiveRecoveryRequested)
5950 : {
5951 96 : newTLI = findNewestTimeLine(recoveryTargetTLI) + 1;
5952 96 : ereport(LOG,
5953 : (errmsg("selected new timeline ID: %u", newTLI)));
5954 :
5955 : /*
5956 : * Make a writable copy of the last WAL segment. (Note that we also
5957 : * have a copy of the last block of the old WAL in
5958 : * endOfRecovery->lastPage; we will use that below.)
5959 : */
5960 96 : XLogInitNewTimeline(EndOfLogTLI, EndOfLog, newTLI);
5961 :
5962 : /*
5963 : * Remove the signal files out of the way, so that we don't
5964 : * accidentally re-enter archive recovery mode in a subsequent crash.
5965 : */
5966 96 : if (endOfRecoveryInfo->standby_signal_file_found)
5967 90 : durable_unlink(STANDBY_SIGNAL_FILE, FATAL);
5968 :
5969 96 : if (endOfRecoveryInfo->recovery_signal_file_found)
5970 6 : durable_unlink(RECOVERY_SIGNAL_FILE, FATAL);
5971 :
5972 : /*
5973 : * Write the timeline history file, and have it archived. After this
5974 : * point (or rather, as soon as the file is archived), the timeline
5975 : * will appear as "taken" in the WAL archive and to any standby
5976 : * servers. If we crash before actually switching to the new
5977 : * timeline, standby servers will nevertheless think that we switched
5978 : * to the new timeline, and will try to connect to the new timeline.
5979 : * To minimize the window for that, try to do as little as possible
5980 : * between here and writing the end-of-recovery record.
5981 : */
5982 96 : writeTimeLineHistory(newTLI, recoveryTargetTLI,
5983 : EndOfLog, endOfRecoveryInfo->recoveryStopReason);
5984 :
5985 96 : ereport(LOG,
5986 : (errmsg("archive recovery complete")));
5987 : }
5988 :
5989 : /* Save the selected TimeLineID in shared memory, too */
5990 1534 : SpinLockAcquire(&XLogCtl->info_lck);
5991 1534 : XLogCtl->InsertTimeLineID = newTLI;
5992 1534 : XLogCtl->PrevTimeLineID = endOfRecoveryInfo->lastRecTLI;
5993 1534 : SpinLockRelease(&XLogCtl->info_lck);
5994 :
5995 : /*
5996 : * Actually, if WAL ended in an incomplete record, skip the parts that
5997 : * made it through and start writing after the portion that persisted.
5998 : * (It's critical to first write an OVERWRITE_CONTRECORD message, which
5999 : * we'll do as soon as we're open for writing new WAL.)
6000 : */
6001 1534 : if (!XLogRecPtrIsInvalid(missingContrecPtr))
6002 : {
6003 : /*
6004 : * We should only have a missingContrecPtr if we're not switching to a
6005 : * new timeline. When a timeline switch occurs, WAL is copied from the
6006 : * old timeline to the new only up to the end of the last complete
6007 : * record, so there can't be an incomplete WAL record that we need to
6008 : * disregard.
6009 : */
6010 : Assert(newTLI == endOfRecoveryInfo->lastRecTLI);
6011 : Assert(!XLogRecPtrIsInvalid(abortedRecPtr));
6012 22 : EndOfLog = missingContrecPtr;
6013 : }
6014 :
6015 : /*
6016 : * Prepare to write WAL starting at EndOfLog location, and init xlog
6017 : * buffer cache using the block containing the last record from the
6018 : * previous incarnation.
6019 : */
6020 1534 : Insert = &XLogCtl->Insert;
6021 1534 : Insert->PrevBytePos = XLogRecPtrToBytePos(endOfRecoveryInfo->lastRec);
6022 1534 : Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
6023 :
6024 : /*
6025 : * Tricky point here: lastPage contains the *last* block that the LastRec
6026 : * record spans, not the one it starts in. The last block is indeed the
6027 : * one we want to use.
6028 : */
6029 1534 : if (EndOfLog % XLOG_BLCKSZ != 0)
6030 : {
6031 : char *page;
6032 : int len;
6033 : int firstIdx;
6034 :
6035 1478 : firstIdx = XLogRecPtrToBufIdx(EndOfLog);
6036 1478 : len = EndOfLog - endOfRecoveryInfo->lastPageBeginPtr;
6037 : Assert(len < XLOG_BLCKSZ);
6038 :
6039 : /* Copy the valid part of the last block, and zero the rest */
6040 1478 : page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
6041 1478 : memcpy(page, endOfRecoveryInfo->lastPage, len);
6042 1478 : memset(page + len, 0, XLOG_BLCKSZ - len);
6043 :
6044 1478 : pg_atomic_write_u64(&XLogCtl->xlblocks[firstIdx], endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6045 1478 : XLogCtl->InitializedUpTo = endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ;
6046 : }
6047 : else
6048 : {
6049 : /*
6050 : * There is no partial block to copy. Just set InitializedUpTo, and
6051 : * let the first attempt to insert a log record to initialize the next
6052 : * buffer.
6053 : */
6054 56 : XLogCtl->InitializedUpTo = EndOfLog;
6055 : }
6056 :
6057 : /*
6058 : * Update local and shared status. This is OK to do without any locks
6059 : * because no other process can be reading or writing WAL yet.
6060 : */
6061 1534 : LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
6062 1534 : pg_atomic_write_u64(&XLogCtl->logInsertResult, EndOfLog);
6063 1534 : pg_atomic_write_u64(&XLogCtl->logWriteResult, EndOfLog);
6064 1534 : pg_atomic_write_u64(&XLogCtl->logFlushResult, EndOfLog);
6065 1534 : XLogCtl->LogwrtRqst.Write = EndOfLog;
6066 1534 : XLogCtl->LogwrtRqst.Flush = EndOfLog;
6067 :
6068 : /*
6069 : * Preallocate additional log files, if wanted.
6070 : */
6071 1534 : PreallocXlogFiles(EndOfLog, newTLI);
6072 :
6073 : /*
6074 : * Okay, we're officially UP.
6075 : */
6076 1534 : InRecovery = false;
6077 :
6078 : /* start the archive_timeout timer and LSN running */
6079 1534 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
6080 1534 : XLogCtl->lastSegSwitchLSN = EndOfLog;
6081 :
6082 : /* also initialize latestCompletedXid, to nextXid - 1 */
6083 1534 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
6084 1534 : TransamVariables->latestCompletedXid = TransamVariables->nextXid;
6085 1534 : FullTransactionIdRetreat(&TransamVariables->latestCompletedXid);
6086 1534 : LWLockRelease(ProcArrayLock);
6087 :
6088 : /*
6089 : * Start up subtrans, if not already done for hot standby. (commit
6090 : * timestamps are started below, if necessary.)
6091 : */
6092 1534 : if (standbyState == STANDBY_DISABLED)
6093 1438 : StartupSUBTRANS(oldestActiveXID);
6094 :
6095 : /*
6096 : * Perform end of recovery actions for any SLRUs that need it.
6097 : */
6098 1534 : TrimCLOG();
6099 1534 : TrimMultiXact();
6100 :
6101 : /*
6102 : * Reload shared-memory state for prepared transactions. This needs to
6103 : * happen before renaming the last partial segment of the old timeline as
6104 : * it may be possible that we have to recovery some transactions from it.
6105 : */
6106 1534 : RecoverPreparedTransactions();
6107 :
6108 : /* Shut down xlogreader */
6109 1534 : ShutdownWalRecovery();
6110 :
6111 : /* Enable WAL writes for this backend only. */
6112 1534 : LocalSetXLogInsertAllowed();
6113 :
6114 : /* If necessary, write overwrite-contrecord before doing anything else */
6115 1534 : if (!XLogRecPtrIsInvalid(abortedRecPtr))
6116 : {
6117 : Assert(!XLogRecPtrIsInvalid(missingContrecPtr));
6118 22 : CreateOverwriteContrecordRecord(abortedRecPtr, missingContrecPtr, newTLI);
6119 : }
6120 :
6121 : /*
6122 : * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
6123 : * record before resource manager writes cleanup WAL records or checkpoint
6124 : * record is written.
6125 : */
6126 1534 : Insert->fullPageWrites = lastFullPageWrites;
6127 1534 : UpdateFullPageWrites();
6128 :
6129 : /*
6130 : * Emit checkpoint or end-of-recovery record in XLOG, if required.
6131 : */
6132 1534 : if (performedWalRecovery)
6133 304 : promoted = PerformRecoveryXLogAction();
6134 :
6135 : /*
6136 : * If any of the critical GUCs have changed, log them before we allow
6137 : * backends to write WAL.
6138 : */
6139 1534 : XLogReportParameters();
6140 :
6141 : /* If this is archive recovery, perform post-recovery cleanup actions. */
6142 1534 : if (ArchiveRecoveryRequested)
6143 96 : CleanupAfterArchiveRecovery(EndOfLogTLI, EndOfLog, newTLI);
6144 :
6145 : /*
6146 : * Local WAL inserts enabled, so it's time to finish initialization of
6147 : * commit timestamp.
6148 : */
6149 1534 : CompleteCommitTsInitialization();
6150 :
6151 : /*
6152 : * All done with end-of-recovery actions.
6153 : *
6154 : * Now allow backends to write WAL and update the control file status in
6155 : * consequence. SharedRecoveryState, that controls if backends can write
6156 : * WAL, is updated while holding ControlFileLock to prevent other backends
6157 : * to look at an inconsistent state of the control file in shared memory.
6158 : * There is still a small window during which backends can write WAL and
6159 : * the control file is still referring to a system not in DB_IN_PRODUCTION
6160 : * state while looking at the on-disk control file.
6161 : *
6162 : * Also, we use info_lck to update SharedRecoveryState to ensure that
6163 : * there are no race conditions concerning visibility of other recent
6164 : * updates to shared memory.
6165 : */
6166 1534 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6167 1534 : ControlFile->state = DB_IN_PRODUCTION;
6168 :
6169 1534 : SpinLockAcquire(&XLogCtl->info_lck);
6170 1534 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_DONE;
6171 1534 : SpinLockRelease(&XLogCtl->info_lck);
6172 :
6173 1534 : UpdateControlFile();
6174 1534 : LWLockRelease(ControlFileLock);
6175 :
6176 : /*
6177 : * Shutdown the recovery environment. This must occur after
6178 : * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
6179 : * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
6180 : * any session building a snapshot will not rely on KnownAssignedXids as
6181 : * RecoveryInProgress() would return false at this stage. This is
6182 : * particularly critical for prepared 2PC transactions, that would still
6183 : * need to be included in snapshots once recovery has ended.
6184 : */
6185 1534 : if (standbyState != STANDBY_DISABLED)
6186 96 : ShutdownRecoveryTransactionEnvironment();
6187 :
6188 : /*
6189 : * If there were cascading standby servers connected to us, nudge any wal
6190 : * sender processes to notice that we've been promoted.
6191 : */
6192 1534 : WalSndWakeup(true, true);
6193 :
6194 : /*
6195 : * If this was a promotion, request an (online) checkpoint now. This isn't
6196 : * required for consistency, but the last restartpoint might be far back,
6197 : * and in case of a crash, recovering from it might take a longer than is
6198 : * appropriate now that we're not in standby mode anymore.
6199 : */
6200 1534 : if (promoted)
6201 82 : RequestCheckpoint(CHECKPOINT_FORCE);
6202 1534 : }
6203 :
6204 : /*
6205 : * Callback from PerformWalRecovery(), called when we switch from crash
6206 : * recovery to archive recovery mode. Updates the control file accordingly.
6207 : */
6208 : void
6209 4 : SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr, TimeLineID replayTLI)
6210 : {
6211 : /* initialize minRecoveryPoint to this record */
6212 4 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6213 4 : ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6214 4 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6215 : {
6216 4 : ControlFile->minRecoveryPoint = EndRecPtr;
6217 4 : ControlFile->minRecoveryPointTLI = replayTLI;
6218 : }
6219 : /* update local copy */
6220 4 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6221 4 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6222 :
6223 : /*
6224 : * The startup process can update its local copy of minRecoveryPoint from
6225 : * this point.
6226 : */
6227 4 : updateMinRecoveryPoint = true;
6228 :
6229 4 : UpdateControlFile();
6230 :
6231 : /*
6232 : * We update SharedRecoveryState while holding the lock on ControlFileLock
6233 : * so both states are consistent in shared memory.
6234 : */
6235 4 : SpinLockAcquire(&XLogCtl->info_lck);
6236 4 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6237 4 : SpinLockRelease(&XLogCtl->info_lck);
6238 :
6239 4 : LWLockRelease(ControlFileLock);
6240 4 : }
6241 :
6242 : /*
6243 : * Callback from PerformWalRecovery(), called when we reach the end of backup.
6244 : * Updates the control file accordingly.
6245 : */
6246 : void
6247 136 : ReachedEndOfBackup(XLogRecPtr EndRecPtr, TimeLineID tli)
6248 : {
6249 : /*
6250 : * We have reached the end of base backup, as indicated by pg_control. The
6251 : * data on disk is now consistent (unless minRecoveryPoint is further
6252 : * ahead, which can happen if we crashed during previous recovery). Reset
6253 : * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
6254 : * make sure we don't allow starting up at an earlier point even if
6255 : * recovery is stopped and restarted soon after this.
6256 : */
6257 136 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6258 :
6259 136 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6260 : {
6261 128 : ControlFile->minRecoveryPoint = EndRecPtr;
6262 128 : ControlFile->minRecoveryPointTLI = tli;
6263 : }
6264 :
6265 136 : ControlFile->backupStartPoint = InvalidXLogRecPtr;
6266 136 : ControlFile->backupEndPoint = InvalidXLogRecPtr;
6267 136 : ControlFile->backupEndRequired = false;
6268 136 : UpdateControlFile();
6269 :
6270 136 : LWLockRelease(ControlFileLock);
6271 136 : }
6272 :
6273 : /*
6274 : * Perform whatever XLOG actions are necessary at end of REDO.
6275 : *
6276 : * The goal here is to make sure that we'll be able to recover properly if
6277 : * we crash again. If we choose to write a checkpoint, we'll write a shutdown
6278 : * checkpoint rather than an on-line one. This is not particularly critical,
6279 : * but since we may be assigning a new TLI, using a shutdown checkpoint allows
6280 : * us to have the rule that TLI only changes in shutdown checkpoints, which
6281 : * allows some extra error checking in xlog_redo.
6282 : */
6283 : static bool
6284 304 : PerformRecoveryXLogAction(void)
6285 : {
6286 304 : bool promoted = false;
6287 :
6288 : /*
6289 : * Perform a checkpoint to update all our recovery activity to disk.
6290 : *
6291 : * Note that we write a shutdown checkpoint rather than an on-line one.
6292 : * This is not particularly critical, but since we may be assigning a new
6293 : * TLI, using a shutdown checkpoint allows us to have the rule that TLI
6294 : * only changes in shutdown checkpoints, which allows some extra error
6295 : * checking in xlog_redo.
6296 : *
6297 : * In promotion, only create a lightweight end-of-recovery record instead
6298 : * of a full checkpoint. A checkpoint is requested later, after we're
6299 : * fully out of recovery mode and already accepting queries.
6300 : */
6301 400 : if (ArchiveRecoveryRequested && IsUnderPostmaster &&
6302 96 : PromoteIsTriggered())
6303 : {
6304 82 : promoted = true;
6305 :
6306 : /*
6307 : * Insert a special WAL record to mark the end of recovery, since we
6308 : * aren't doing a checkpoint. That means that the checkpointer process
6309 : * may likely be in the middle of a time-smoothed restartpoint and
6310 : * could continue to be for minutes after this. That sounds strange,
6311 : * but the effect is roughly the same and it would be stranger to try
6312 : * to come out of the restartpoint and then checkpoint. We request a
6313 : * checkpoint later anyway, just for safety.
6314 : */
6315 82 : CreateEndOfRecoveryRecord();
6316 : }
6317 : else
6318 : {
6319 222 : RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
6320 : CHECKPOINT_IMMEDIATE |
6321 : CHECKPOINT_WAIT);
6322 : }
6323 :
6324 304 : return promoted;
6325 : }
6326 :
6327 : /*
6328 : * Is the system still in recovery?
6329 : *
6330 : * Unlike testing InRecovery, this works in any process that's connected to
6331 : * shared memory.
6332 : */
6333 : bool
6334 163777884 : RecoveryInProgress(void)
6335 : {
6336 : /*
6337 : * We check shared state each time only until we leave recovery mode. We
6338 : * can't re-enter recovery, so there's no need to keep checking after the
6339 : * shared variable has once been seen false.
6340 : */
6341 163777884 : if (!LocalRecoveryInProgress)
6342 158697854 : return false;
6343 : else
6344 : {
6345 : /*
6346 : * use volatile pointer to make sure we make a fresh read of the
6347 : * shared variable.
6348 : */
6349 5080030 : volatile XLogCtlData *xlogctl = XLogCtl;
6350 :
6351 5080030 : LocalRecoveryInProgress = (xlogctl->SharedRecoveryState != RECOVERY_STATE_DONE);
6352 :
6353 : /*
6354 : * Note: We don't need a memory barrier when we're still in recovery.
6355 : * We might exit recovery immediately after return, so the caller
6356 : * can't rely on 'true' meaning that we're still in recovery anyway.
6357 : */
6358 :
6359 5080030 : return LocalRecoveryInProgress;
6360 : }
6361 : }
6362 :
6363 : /*
6364 : * Returns current recovery state from shared memory.
6365 : *
6366 : * This returned state is kept consistent with the contents of the control
6367 : * file. See details about the possible values of RecoveryState in xlog.h.
6368 : */
6369 : RecoveryState
6370 114 : GetRecoveryState(void)
6371 : {
6372 : RecoveryState retval;
6373 :
6374 114 : SpinLockAcquire(&XLogCtl->info_lck);
6375 114 : retval = XLogCtl->SharedRecoveryState;
6376 114 : SpinLockRelease(&XLogCtl->info_lck);
6377 :
6378 114 : return retval;
6379 : }
6380 :
6381 : /*
6382 : * Is this process allowed to insert new WAL records?
6383 : *
6384 : * Ordinarily this is essentially equivalent to !RecoveryInProgress().
6385 : * But we also have provisions for forcing the result "true" or "false"
6386 : * within specific processes regardless of the global state.
6387 : */
6388 : bool
6389 57688486 : XLogInsertAllowed(void)
6390 : {
6391 : /*
6392 : * If value is "unconditionally true" or "unconditionally false", just
6393 : * return it. This provides the normal fast path once recovery is known
6394 : * done.
6395 : */
6396 57688486 : if (LocalXLogInsertAllowed >= 0)
6397 57491406 : return (bool) LocalXLogInsertAllowed;
6398 :
6399 : /*
6400 : * Else, must check to see if we're still in recovery.
6401 : */
6402 197080 : if (RecoveryInProgress())
6403 182384 : return false;
6404 :
6405 : /*
6406 : * On exit from recovery, reset to "unconditionally true", since there is
6407 : * no need to keep checking.
6408 : */
6409 14696 : LocalXLogInsertAllowed = 1;
6410 14696 : return true;
6411 : }
6412 :
6413 : /*
6414 : * Make XLogInsertAllowed() return true in the current process only.
6415 : *
6416 : * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
6417 : * and even call LocalSetXLogInsertAllowed() again after that.
6418 : *
6419 : * Returns the previous value of LocalXLogInsertAllowed.
6420 : */
6421 : static int
6422 1592 : LocalSetXLogInsertAllowed(void)
6423 : {
6424 1592 : int oldXLogAllowed = LocalXLogInsertAllowed;
6425 :
6426 1592 : LocalXLogInsertAllowed = 1;
6427 :
6428 1592 : return oldXLogAllowed;
6429 : }
6430 :
6431 : /*
6432 : * Return the current Redo pointer from shared memory.
6433 : *
6434 : * As a side-effect, the local RedoRecPtr copy is updated.
6435 : */
6436 : XLogRecPtr
6437 567830 : GetRedoRecPtr(void)
6438 : {
6439 : XLogRecPtr ptr;
6440 :
6441 : /*
6442 : * The possibly not up-to-date copy in XlogCtl is enough. Even if we
6443 : * grabbed a WAL insertion lock to read the authoritative value in
6444 : * Insert->RedoRecPtr, someone might update it just after we've released
6445 : * the lock.
6446 : */
6447 567830 : SpinLockAcquire(&XLogCtl->info_lck);
6448 567830 : ptr = XLogCtl->RedoRecPtr;
6449 567830 : SpinLockRelease(&XLogCtl->info_lck);
6450 :
6451 567830 : if (RedoRecPtr < ptr)
6452 2552 : RedoRecPtr = ptr;
6453 :
6454 567830 : return RedoRecPtr;
6455 : }
6456 :
6457 : /*
6458 : * Return information needed to decide whether a modified block needs a
6459 : * full-page image to be included in the WAL record.
6460 : *
6461 : * The returned values are cached copies from backend-private memory, and
6462 : * possibly out-of-date or, indeed, uninitialized, in which case they will
6463 : * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6464 : * re-check them against up-to-date values, while holding the WAL insert lock.
6465 : */
6466 : void
6467 27965538 : GetFullPageWriteInfo(XLogRecPtr *RedoRecPtr_p, bool *doPageWrites_p)
6468 : {
6469 27965538 : *RedoRecPtr_p = RedoRecPtr;
6470 27965538 : *doPageWrites_p = doPageWrites;
6471 27965538 : }
6472 :
6473 : /*
6474 : * GetInsertRecPtr -- Returns the current insert position.
6475 : *
6476 : * NOTE: The value *actually* returned is the position of the last full
6477 : * xlog page. It lags behind the real insert position by at most 1 page.
6478 : * For that, we don't need to scan through WAL insertion locks, and an
6479 : * approximation is enough for the current usage of this function.
6480 : */
6481 : XLogRecPtr
6482 11426 : GetInsertRecPtr(void)
6483 : {
6484 : XLogRecPtr recptr;
6485 :
6486 11426 : SpinLockAcquire(&XLogCtl->info_lck);
6487 11426 : recptr = XLogCtl->LogwrtRqst.Write;
6488 11426 : SpinLockRelease(&XLogCtl->info_lck);
6489 :
6490 11426 : return recptr;
6491 : }
6492 :
6493 : /*
6494 : * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6495 : * position known to be fsync'd to disk. This should only be used on a
6496 : * system that is known not to be in recovery.
6497 : */
6498 : XLogRecPtr
6499 381118 : GetFlushRecPtr(TimeLineID *insertTLI)
6500 : {
6501 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6502 :
6503 381118 : RefreshXLogWriteResult(LogwrtResult);
6504 :
6505 : /*
6506 : * If we're writing and flushing WAL, the time line can't be changing, so
6507 : * no lock is required.
6508 : */
6509 381118 : if (insertTLI)
6510 49114 : *insertTLI = XLogCtl->InsertTimeLineID;
6511 :
6512 381118 : return LogwrtResult.Flush;
6513 : }
6514 :
6515 : /*
6516 : * GetWALInsertionTimeLine -- Returns the current timeline of a system that
6517 : * is not in recovery.
6518 : */
6519 : TimeLineID
6520 227074 : GetWALInsertionTimeLine(void)
6521 : {
6522 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6523 :
6524 : /* Since the value can't be changing, no lock is required. */
6525 227074 : return XLogCtl->InsertTimeLineID;
6526 : }
6527 :
6528 : /*
6529 : * GetWALInsertionTimeLineIfSet -- If the system is not in recovery, returns
6530 : * the WAL insertion timeline; else, returns 0. Wherever possible, use
6531 : * GetWALInsertionTimeLine() instead, since it's cheaper. Note that this
6532 : * function decides recovery has ended as soon as the insert TLI is set, which
6533 : * happens before we set XLogCtl->SharedRecoveryState to RECOVERY_STATE_DONE.
6534 : */
6535 : TimeLineID
6536 0 : GetWALInsertionTimeLineIfSet(void)
6537 : {
6538 : TimeLineID insertTLI;
6539 :
6540 0 : SpinLockAcquire(&XLogCtl->info_lck);
6541 0 : insertTLI = XLogCtl->InsertTimeLineID;
6542 0 : SpinLockRelease(&XLogCtl->info_lck);
6543 :
6544 0 : return insertTLI;
6545 : }
6546 :
6547 : /*
6548 : * GetLastImportantRecPtr -- Returns the LSN of the last important record
6549 : * inserted. All records not explicitly marked as unimportant are considered
6550 : * important.
6551 : *
6552 : * The LSN is determined by computing the maximum of
6553 : * WALInsertLocks[i].lastImportantAt.
6554 : */
6555 : XLogRecPtr
6556 2156 : GetLastImportantRecPtr(void)
6557 : {
6558 2156 : XLogRecPtr res = InvalidXLogRecPtr;
6559 : int i;
6560 :
6561 19404 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
6562 : {
6563 : XLogRecPtr last_important;
6564 :
6565 : /*
6566 : * Need to take a lock to prevent torn reads of the LSN, which are
6567 : * possible on some of the supported platforms. WAL insert locks only
6568 : * support exclusive mode, so we have to use that.
6569 : */
6570 17248 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
6571 17248 : last_important = WALInsertLocks[i].l.lastImportantAt;
6572 17248 : LWLockRelease(&WALInsertLocks[i].l.lock);
6573 :
6574 17248 : if (res < last_important)
6575 3916 : res = last_important;
6576 : }
6577 :
6578 2156 : return res;
6579 : }
6580 :
6581 : /*
6582 : * Get the time and LSN of the last xlog segment switch
6583 : */
6584 : pg_time_t
6585 0 : GetLastSegSwitchData(XLogRecPtr *lastSwitchLSN)
6586 : {
6587 : pg_time_t result;
6588 :
6589 : /* Need WALWriteLock, but shared lock is sufficient */
6590 0 : LWLockAcquire(WALWriteLock, LW_SHARED);
6591 0 : result = XLogCtl->lastSegSwitchTime;
6592 0 : *lastSwitchLSN = XLogCtl->lastSegSwitchLSN;
6593 0 : LWLockRelease(WALWriteLock);
6594 :
6595 0 : return result;
6596 : }
6597 :
6598 : /*
6599 : * This must be called ONCE during postmaster or standalone-backend shutdown
6600 : */
6601 : void
6602 1088 : ShutdownXLOG(int code, Datum arg)
6603 : {
6604 : /*
6605 : * We should have an aux process resource owner to use, and we should not
6606 : * be in a transaction that's installed some other resowner.
6607 : */
6608 : Assert(AuxProcessResourceOwner != NULL);
6609 : Assert(CurrentResourceOwner == NULL ||
6610 : CurrentResourceOwner == AuxProcessResourceOwner);
6611 1088 : CurrentResourceOwner = AuxProcessResourceOwner;
6612 :
6613 : /* Don't be chatty in standalone mode */
6614 1088 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
6615 : (errmsg("shutting down")));
6616 :
6617 : /*
6618 : * Signal walsenders to move to stopping state.
6619 : */
6620 1088 : WalSndInitStopping();
6621 :
6622 : /*
6623 : * Wait for WAL senders to be in stopping state. This prevents commands
6624 : * from writing new WAL.
6625 : */
6626 1088 : WalSndWaitStopping();
6627 :
6628 1088 : if (RecoveryInProgress())
6629 100 : CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_IMMEDIATE);
6630 : else
6631 : {
6632 : /*
6633 : * If archiving is enabled, rotate the last XLOG file so that all the
6634 : * remaining records are archived (postmaster wakes up the archiver
6635 : * process one more time at the end of shutdown). The checkpoint
6636 : * record will go to the next XLOG file and won't be archived (yet).
6637 : */
6638 988 : if (XLogArchivingActive())
6639 26 : RequestXLogSwitch(false);
6640 :
6641 988 : CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_IMMEDIATE);
6642 : }
6643 1088 : }
6644 :
6645 : /*
6646 : * Log start of a checkpoint.
6647 : */
6648 : static void
6649 1944 : LogCheckpointStart(int flags, bool restartpoint)
6650 : {
6651 1944 : if (restartpoint)
6652 366 : ereport(LOG,
6653 : /* translator: the placeholders show checkpoint options */
6654 : (errmsg("restartpoint starting:%s%s%s%s%s%s%s%s",
6655 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6656 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6657 : (flags & CHECKPOINT_IMMEDIATE) ? " immediate" : "",
6658 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6659 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6660 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6661 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6662 : (flags & CHECKPOINT_FLUSH_ALL) ? " flush-all" : "")));
6663 : else
6664 1578 : ereport(LOG,
6665 : /* translator: the placeholders show checkpoint options */
6666 : (errmsg("checkpoint starting:%s%s%s%s%s%s%s%s",
6667 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6668 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6669 : (flags & CHECKPOINT_IMMEDIATE) ? " immediate" : "",
6670 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6671 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6672 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6673 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6674 : (flags & CHECKPOINT_FLUSH_ALL) ? " flush-all" : "")));
6675 1944 : }
6676 :
6677 : /*
6678 : * Log end of a checkpoint.
6679 : */
6680 : static void
6681 2468 : LogCheckpointEnd(bool restartpoint)
6682 : {
6683 : long write_msecs,
6684 : sync_msecs,
6685 : total_msecs,
6686 : longest_msecs,
6687 : average_msecs;
6688 : uint64 average_sync_time;
6689 :
6690 2468 : CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
6691 :
6692 2468 : write_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_write_t,
6693 : CheckpointStats.ckpt_sync_t);
6694 :
6695 2468 : sync_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_sync_t,
6696 : CheckpointStats.ckpt_sync_end_t);
6697 :
6698 : /* Accumulate checkpoint timing summary data, in milliseconds. */
6699 2468 : PendingCheckpointerStats.write_time += write_msecs;
6700 2468 : PendingCheckpointerStats.sync_time += sync_msecs;
6701 :
6702 : /*
6703 : * All of the published timing statistics are accounted for. Only
6704 : * continue if a log message is to be written.
6705 : */
6706 2468 : if (!log_checkpoints)
6707 524 : return;
6708 :
6709 1944 : total_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_start_t,
6710 : CheckpointStats.ckpt_end_t);
6711 :
6712 : /*
6713 : * Timing values returned from CheckpointStats are in microseconds.
6714 : * Convert to milliseconds for consistent printing.
6715 : */
6716 1944 : longest_msecs = (long) ((CheckpointStats.ckpt_longest_sync + 999) / 1000);
6717 :
6718 1944 : average_sync_time = 0;
6719 1944 : if (CheckpointStats.ckpt_sync_rels > 0)
6720 0 : average_sync_time = CheckpointStats.ckpt_agg_sync_time /
6721 0 : CheckpointStats.ckpt_sync_rels;
6722 1944 : average_msecs = (long) ((average_sync_time + 999) / 1000);
6723 :
6724 : /*
6725 : * ControlFileLock is not required to see ControlFile->checkPoint and
6726 : * ->checkPointCopy here as we are the only updator of those variables at
6727 : * this moment.
6728 : */
6729 1944 : if (restartpoint)
6730 366 : ereport(LOG,
6731 : (errmsg("restartpoint complete: wrote %d buffers (%.1f%%), "
6732 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6733 : "%d removed, %d recycled; write=%ld.%03d s, "
6734 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6735 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6736 : "estimate=%d kB; lsn=%X/%X, redo lsn=%X/%X",
6737 : CheckpointStats.ckpt_bufs_written,
6738 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6739 : CheckpointStats.ckpt_slru_written,
6740 : CheckpointStats.ckpt_segs_added,
6741 : CheckpointStats.ckpt_segs_removed,
6742 : CheckpointStats.ckpt_segs_recycled,
6743 : write_msecs / 1000, (int) (write_msecs % 1000),
6744 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6745 : total_msecs / 1000, (int) (total_msecs % 1000),
6746 : CheckpointStats.ckpt_sync_rels,
6747 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6748 : average_msecs / 1000, (int) (average_msecs % 1000),
6749 : (int) (PrevCheckPointDistance / 1024.0),
6750 : (int) (CheckPointDistanceEstimate / 1024.0),
6751 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6752 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6753 : else
6754 1578 : ereport(LOG,
6755 : (errmsg("checkpoint complete: wrote %d buffers (%.1f%%), "
6756 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6757 : "%d removed, %d recycled; write=%ld.%03d s, "
6758 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6759 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6760 : "estimate=%d kB; lsn=%X/%X, redo lsn=%X/%X",
6761 : CheckpointStats.ckpt_bufs_written,
6762 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6763 : CheckpointStats.ckpt_slru_written,
6764 : CheckpointStats.ckpt_segs_added,
6765 : CheckpointStats.ckpt_segs_removed,
6766 : CheckpointStats.ckpt_segs_recycled,
6767 : write_msecs / 1000, (int) (write_msecs % 1000),
6768 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6769 : total_msecs / 1000, (int) (total_msecs % 1000),
6770 : CheckpointStats.ckpt_sync_rels,
6771 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6772 : average_msecs / 1000, (int) (average_msecs % 1000),
6773 : (int) (PrevCheckPointDistance / 1024.0),
6774 : (int) (CheckPointDistanceEstimate / 1024.0),
6775 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6776 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6777 : }
6778 :
6779 : /*
6780 : * Update the estimate of distance between checkpoints.
6781 : *
6782 : * The estimate is used to calculate the number of WAL segments to keep
6783 : * preallocated, see XLOGfileslop().
6784 : */
6785 : static void
6786 2468 : UpdateCheckPointDistanceEstimate(uint64 nbytes)
6787 : {
6788 : /*
6789 : * To estimate the number of segments consumed between checkpoints, keep a
6790 : * moving average of the amount of WAL generated in previous checkpoint
6791 : * cycles. However, if the load is bursty, with quiet periods and busy
6792 : * periods, we want to cater for the peak load. So instead of a plain
6793 : * moving average, let the average decline slowly if the previous cycle
6794 : * used less WAL than estimated, but bump it up immediately if it used
6795 : * more.
6796 : *
6797 : * When checkpoints are triggered by max_wal_size, this should converge to
6798 : * CheckpointSegments * wal_segment_size,
6799 : *
6800 : * Note: This doesn't pay any attention to what caused the checkpoint.
6801 : * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
6802 : * starting a base backup, are counted the same as those created
6803 : * automatically. The slow-decline will largely mask them out, if they are
6804 : * not frequent. If they are frequent, it seems reasonable to count them
6805 : * in as any others; if you issue a manual checkpoint every 5 minutes and
6806 : * never let a timed checkpoint happen, it makes sense to base the
6807 : * preallocation on that 5 minute interval rather than whatever
6808 : * checkpoint_timeout is set to.
6809 : */
6810 2468 : PrevCheckPointDistance = nbytes;
6811 2468 : if (CheckPointDistanceEstimate < nbytes)
6812 1250 : CheckPointDistanceEstimate = nbytes;
6813 : else
6814 1218 : CheckPointDistanceEstimate =
6815 1218 : (0.90 * CheckPointDistanceEstimate + 0.10 * (double) nbytes);
6816 2468 : }
6817 :
6818 : /*
6819 : * Update the ps display for a process running a checkpoint. Note that
6820 : * this routine should not do any allocations so as it can be called
6821 : * from a critical section.
6822 : */
6823 : static void
6824 4936 : update_checkpoint_display(int flags, bool restartpoint, bool reset)
6825 : {
6826 : /*
6827 : * The status is reported only for end-of-recovery and shutdown
6828 : * checkpoints or shutdown restartpoints. Updating the ps display is
6829 : * useful in those situations as it may not be possible to rely on
6830 : * pg_stat_activity to see the status of the checkpointer or the startup
6831 : * process.
6832 : */
6833 4936 : if ((flags & (CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IS_SHUTDOWN)) == 0)
6834 2764 : return;
6835 :
6836 2172 : if (reset)
6837 1086 : set_ps_display("");
6838 : else
6839 : {
6840 : char activitymsg[128];
6841 :
6842 3258 : snprintf(activitymsg, sizeof(activitymsg), "performing %s%s%s",
6843 1086 : (flags & CHECKPOINT_END_OF_RECOVERY) ? "end-of-recovery " : "",
6844 1086 : (flags & CHECKPOINT_IS_SHUTDOWN) ? "shutdown " : "",
6845 : restartpoint ? "restartpoint" : "checkpoint");
6846 1086 : set_ps_display(activitymsg);
6847 : }
6848 : }
6849 :
6850 :
6851 : /*
6852 : * Perform a checkpoint --- either during shutdown, or on-the-fly
6853 : *
6854 : * flags is a bitwise OR of the following:
6855 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
6856 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
6857 : * CHECKPOINT_IMMEDIATE: finish the checkpoint ASAP,
6858 : * ignoring checkpoint_completion_target parameter.
6859 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
6860 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
6861 : * CHECKPOINT_END_OF_RECOVERY).
6862 : * CHECKPOINT_FLUSH_ALL: also flush buffers of unlogged tables.
6863 : *
6864 : * Note: flags contains other bits, of interest here only for logging purposes.
6865 : * In particular note that this routine is synchronous and does not pay
6866 : * attention to CHECKPOINT_WAIT.
6867 : *
6868 : * If !shutdown then we are writing an online checkpoint. An XLOG_CHECKPOINT_REDO
6869 : * record is inserted into WAL at the logical location of the checkpoint, before
6870 : * flushing anything to disk, and when the checkpoint is eventually completed,
6871 : * and it is from this point that WAL replay will begin in the case of a recovery
6872 : * from this checkpoint. Once everything is written to disk, an
6873 : * XLOG_CHECKPOINT_ONLINE record is written to complete the checkpoint, and
6874 : * points back to the earlier XLOG_CHECKPOINT_REDO record. This mechanism allows
6875 : * other write-ahead log records to be written while the checkpoint is in
6876 : * progress, but we must be very careful about order of operations. This function
6877 : * may take many minutes to execute on a busy system.
6878 : *
6879 : * On the other hand, when shutdown is true, concurrent insertion into the
6880 : * write-ahead log is impossible, so there is no need for two separate records.
6881 : * In this case, we only insert an XLOG_CHECKPOINT_SHUTDOWN record, and it's
6882 : * both the record marking the completion of the checkpoint and the location
6883 : * from which WAL replay would begin if needed.
6884 : *
6885 : * Returns true if a new checkpoint was performed, or false if it was skipped
6886 : * because the system was idle.
6887 : */
6888 : bool
6889 2102 : CreateCheckPoint(int flags)
6890 : {
6891 : bool shutdown;
6892 : CheckPoint checkPoint;
6893 : XLogRecPtr recptr;
6894 : XLogSegNo _logSegNo;
6895 2102 : XLogCtlInsert *Insert = &XLogCtl->Insert;
6896 : uint32 freespace;
6897 : XLogRecPtr PriorRedoPtr;
6898 : XLogRecPtr last_important_lsn;
6899 : VirtualTransactionId *vxids;
6900 : int nvxids;
6901 2102 : int oldXLogAllowed = 0;
6902 :
6903 : /*
6904 : * An end-of-recovery checkpoint is really a shutdown checkpoint, just
6905 : * issued at a different time.
6906 : */
6907 2102 : if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
6908 1046 : shutdown = true;
6909 : else
6910 1056 : shutdown = false;
6911 :
6912 : /* sanity check */
6913 2102 : if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
6914 0 : elog(ERROR, "can't create a checkpoint during recovery");
6915 :
6916 : /*
6917 : * Prepare to accumulate statistics.
6918 : *
6919 : * Note: because it is possible for log_checkpoints to change while a
6920 : * checkpoint proceeds, we always accumulate stats, even if
6921 : * log_checkpoints is currently off.
6922 : */
6923 23122 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
6924 2102 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
6925 :
6926 : /*
6927 : * Let smgr prepare for checkpoint; this has to happen outside the
6928 : * critical section and before we determine the REDO pointer. Note that
6929 : * smgr must not do anything that'd have to be undone if we decide no
6930 : * checkpoint is needed.
6931 : */
6932 2102 : SyncPreCheckpoint();
6933 :
6934 : /*
6935 : * Use a critical section to force system panic if we have trouble.
6936 : */
6937 2102 : START_CRIT_SECTION();
6938 :
6939 2102 : if (shutdown)
6940 : {
6941 1046 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6942 1046 : ControlFile->state = DB_SHUTDOWNING;
6943 1046 : UpdateControlFile();
6944 1046 : LWLockRelease(ControlFileLock);
6945 : }
6946 :
6947 : /* Begin filling in the checkpoint WAL record */
6948 25224 : MemSet(&checkPoint, 0, sizeof(checkPoint));
6949 2102 : checkPoint.time = (pg_time_t) time(NULL);
6950 :
6951 : /*
6952 : * For Hot Standby, derive the oldestActiveXid before we fix the redo
6953 : * pointer. This allows us to begin accumulating changes to assemble our
6954 : * starting snapshot of locks and transactions.
6955 : */
6956 2102 : if (!shutdown && XLogStandbyInfoActive())
6957 994 : checkPoint.oldestActiveXid = GetOldestActiveTransactionId();
6958 : else
6959 1108 : checkPoint.oldestActiveXid = InvalidTransactionId;
6960 :
6961 : /*
6962 : * Get location of last important record before acquiring insert locks (as
6963 : * GetLastImportantRecPtr() also locks WAL locks).
6964 : */
6965 2102 : last_important_lsn = GetLastImportantRecPtr();
6966 :
6967 : /*
6968 : * If this isn't a shutdown or forced checkpoint, and if there has been no
6969 : * WAL activity requiring a checkpoint, skip it. The idea here is to
6970 : * avoid inserting duplicate checkpoints when the system is idle.
6971 : */
6972 2102 : if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
6973 : CHECKPOINT_FORCE)) == 0)
6974 : {
6975 354 : if (last_important_lsn == ControlFile->checkPoint)
6976 : {
6977 0 : END_CRIT_SECTION();
6978 0 : ereport(DEBUG1,
6979 : (errmsg_internal("checkpoint skipped because system is idle")));
6980 0 : return false;
6981 : }
6982 : }
6983 :
6984 : /*
6985 : * An end-of-recovery checkpoint is created before anyone is allowed to
6986 : * write WAL. To allow us to write the checkpoint record, temporarily
6987 : * enable XLogInsertAllowed.
6988 : */
6989 2102 : if (flags & CHECKPOINT_END_OF_RECOVERY)
6990 58 : oldXLogAllowed = LocalSetXLogInsertAllowed();
6991 :
6992 2102 : checkPoint.ThisTimeLineID = XLogCtl->InsertTimeLineID;
6993 2102 : if (flags & CHECKPOINT_END_OF_RECOVERY)
6994 58 : checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
6995 : else
6996 2044 : checkPoint.PrevTimeLineID = checkPoint.ThisTimeLineID;
6997 :
6998 : /*
6999 : * We must block concurrent insertions while examining insert state.
7000 : */
7001 2102 : WALInsertLockAcquireExclusive();
7002 :
7003 2102 : checkPoint.fullPageWrites = Insert->fullPageWrites;
7004 2102 : checkPoint.wal_level = wal_level;
7005 :
7006 2102 : if (shutdown)
7007 : {
7008 1046 : XLogRecPtr curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
7009 :
7010 : /*
7011 : * Compute new REDO record ptr = location of next XLOG record.
7012 : *
7013 : * Since this is a shutdown checkpoint, there can't be any concurrent
7014 : * WAL insertion.
7015 : */
7016 1046 : freespace = INSERT_FREESPACE(curInsert);
7017 1046 : if (freespace == 0)
7018 : {
7019 0 : if (XLogSegmentOffset(curInsert, wal_segment_size) == 0)
7020 0 : curInsert += SizeOfXLogLongPHD;
7021 : else
7022 0 : curInsert += SizeOfXLogShortPHD;
7023 : }
7024 1046 : checkPoint.redo = curInsert;
7025 :
7026 : /*
7027 : * Here we update the shared RedoRecPtr for future XLogInsert calls;
7028 : * this must be done while holding all the insertion locks.
7029 : *
7030 : * Note: if we fail to complete the checkpoint, RedoRecPtr will be
7031 : * left pointing past where it really needs to point. This is okay;
7032 : * the only consequence is that XLogInsert might back up whole buffers
7033 : * that it didn't really need to. We can't postpone advancing
7034 : * RedoRecPtr because XLogInserts that happen while we are dumping
7035 : * buffers must assume that their buffer changes are not included in
7036 : * the checkpoint.
7037 : */
7038 1046 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
7039 : }
7040 :
7041 : /*
7042 : * Now we can release the WAL insertion locks, allowing other xacts to
7043 : * proceed while we are flushing disk buffers.
7044 : */
7045 2102 : WALInsertLockRelease();
7046 :
7047 : /*
7048 : * If this is an online checkpoint, we have not yet determined the redo
7049 : * point. We do so now by inserting the special XLOG_CHECKPOINT_REDO
7050 : * record; the LSN at which it starts becomes the new redo pointer. We
7051 : * don't do this for a shutdown checkpoint, because in that case no WAL
7052 : * can be written between the redo point and the insertion of the
7053 : * checkpoint record itself, so the checkpoint record itself serves to
7054 : * mark the redo point.
7055 : */
7056 2102 : if (!shutdown)
7057 : {
7058 : /* Include WAL level in record for WAL summarizer's benefit. */
7059 1056 : XLogBeginInsert();
7060 1056 : XLogRegisterData((char *) &wal_level, sizeof(wal_level));
7061 1056 : (void) XLogInsert(RM_XLOG_ID, XLOG_CHECKPOINT_REDO);
7062 :
7063 : /*
7064 : * XLogInsertRecord will have updated XLogCtl->Insert.RedoRecPtr in
7065 : * shared memory and RedoRecPtr in backend-local memory, but we need
7066 : * to copy that into the record that will be inserted when the
7067 : * checkpoint is complete.
7068 : */
7069 1056 : checkPoint.redo = RedoRecPtr;
7070 : }
7071 :
7072 : /* Update the info_lck-protected copy of RedoRecPtr as well */
7073 2102 : SpinLockAcquire(&XLogCtl->info_lck);
7074 2102 : XLogCtl->RedoRecPtr = checkPoint.redo;
7075 2102 : SpinLockRelease(&XLogCtl->info_lck);
7076 :
7077 : /*
7078 : * If enabled, log checkpoint start. We postpone this until now so as not
7079 : * to log anything if we decided to skip the checkpoint.
7080 : */
7081 2102 : if (log_checkpoints)
7082 1578 : LogCheckpointStart(flags, false);
7083 :
7084 : /* Update the process title */
7085 2102 : update_checkpoint_display(flags, false, false);
7086 :
7087 : TRACE_POSTGRESQL_CHECKPOINT_START(flags);
7088 :
7089 : /*
7090 : * Get the other info we need for the checkpoint record.
7091 : *
7092 : * We don't need to save oldestClogXid in the checkpoint, it only matters
7093 : * for the short period in which clog is being truncated, and if we crash
7094 : * during that we'll redo the clog truncation and fix up oldestClogXid
7095 : * there.
7096 : */
7097 2102 : LWLockAcquire(XidGenLock, LW_SHARED);
7098 2102 : checkPoint.nextXid = TransamVariables->nextXid;
7099 2102 : checkPoint.oldestXid = TransamVariables->oldestXid;
7100 2102 : checkPoint.oldestXidDB = TransamVariables->oldestXidDB;
7101 2102 : LWLockRelease(XidGenLock);
7102 :
7103 2102 : LWLockAcquire(CommitTsLock, LW_SHARED);
7104 2102 : checkPoint.oldestCommitTsXid = TransamVariables->oldestCommitTsXid;
7105 2102 : checkPoint.newestCommitTsXid = TransamVariables->newestCommitTsXid;
7106 2102 : LWLockRelease(CommitTsLock);
7107 :
7108 2102 : LWLockAcquire(OidGenLock, LW_SHARED);
7109 2102 : checkPoint.nextOid = TransamVariables->nextOid;
7110 2102 : if (!shutdown)
7111 1056 : checkPoint.nextOid += TransamVariables->oidCount;
7112 2102 : LWLockRelease(OidGenLock);
7113 :
7114 2102 : MultiXactGetCheckptMulti(shutdown,
7115 : &checkPoint.nextMulti,
7116 : &checkPoint.nextMultiOffset,
7117 : &checkPoint.oldestMulti,
7118 : &checkPoint.oldestMultiDB);
7119 :
7120 : /*
7121 : * Having constructed the checkpoint record, ensure all shmem disk buffers
7122 : * and commit-log buffers are flushed to disk.
7123 : *
7124 : * This I/O could fail for various reasons. If so, we will fail to
7125 : * complete the checkpoint, but there is no reason to force a system
7126 : * panic. Accordingly, exit critical section while doing it.
7127 : */
7128 2102 : END_CRIT_SECTION();
7129 :
7130 : /*
7131 : * In some cases there are groups of actions that must all occur on one
7132 : * side or the other of a checkpoint record. Before flushing the
7133 : * checkpoint record we must explicitly wait for any backend currently
7134 : * performing those groups of actions.
7135 : *
7136 : * One example is end of transaction, so we must wait for any transactions
7137 : * that are currently in commit critical sections. If an xact inserted
7138 : * its commit record into XLOG just before the REDO point, then a crash
7139 : * restart from the REDO point would not replay that record, which means
7140 : * that our flushing had better include the xact's update of pg_xact. So
7141 : * we wait till he's out of his commit critical section before proceeding.
7142 : * See notes in RecordTransactionCommit().
7143 : *
7144 : * Because we've already released the insertion locks, this test is a bit
7145 : * fuzzy: it is possible that we will wait for xacts we didn't really need
7146 : * to wait for. But the delay should be short and it seems better to make
7147 : * checkpoint take a bit longer than to hold off insertions longer than
7148 : * necessary. (In fact, the whole reason we have this issue is that xact.c
7149 : * does commit record XLOG insertion and clog update as two separate steps
7150 : * protected by different locks, but again that seems best on grounds of
7151 : * minimizing lock contention.)
7152 : *
7153 : * A transaction that has not yet set delayChkptFlags when we look cannot
7154 : * be at risk, since it has not inserted its commit record yet; and one
7155 : * that's already cleared it is not at risk either, since it's done fixing
7156 : * clog and we will correctly flush the update below. So we cannot miss
7157 : * any xacts we need to wait for.
7158 : */
7159 2102 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
7160 2102 : if (nvxids > 0)
7161 : {
7162 : do
7163 : {
7164 : /*
7165 : * Keep absorbing fsync requests while we wait. There could even
7166 : * be a deadlock if we don't, if the process that prevents the
7167 : * checkpoint is trying to add a request to the queue.
7168 : */
7169 58 : AbsorbSyncRequests();
7170 :
7171 58 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
7172 58 : pg_usleep(10000L); /* wait for 10 msec */
7173 58 : pgstat_report_wait_end();
7174 58 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7175 58 : DELAY_CHKPT_START));
7176 : }
7177 2102 : pfree(vxids);
7178 :
7179 2102 : CheckPointGuts(checkPoint.redo, flags);
7180 :
7181 2102 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
7182 2102 : if (nvxids > 0)
7183 : {
7184 : do
7185 : {
7186 0 : AbsorbSyncRequests();
7187 :
7188 0 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_COMPLETE);
7189 0 : pg_usleep(10000L); /* wait for 10 msec */
7190 0 : pgstat_report_wait_end();
7191 0 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7192 0 : DELAY_CHKPT_COMPLETE));
7193 : }
7194 2102 : pfree(vxids);
7195 :
7196 : /*
7197 : * Take a snapshot of running transactions and write this to WAL. This
7198 : * allows us to reconstruct the state of running transactions during
7199 : * archive recovery, if required. Skip, if this info disabled.
7200 : *
7201 : * If we are shutting down, or Startup process is completing crash
7202 : * recovery we don't need to write running xact data.
7203 : */
7204 2102 : if (!shutdown && XLogStandbyInfoActive())
7205 994 : LogStandbySnapshot();
7206 :
7207 2102 : START_CRIT_SECTION();
7208 :
7209 : /*
7210 : * Now insert the checkpoint record into XLOG.
7211 : */
7212 2102 : XLogBeginInsert();
7213 2102 : XLogRegisterData((char *) (&checkPoint), sizeof(checkPoint));
7214 2102 : recptr = XLogInsert(RM_XLOG_ID,
7215 : shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
7216 : XLOG_CHECKPOINT_ONLINE);
7217 :
7218 2102 : XLogFlush(recptr);
7219 :
7220 : /*
7221 : * We mustn't write any new WAL after a shutdown checkpoint, or it will be
7222 : * overwritten at next startup. No-one should even try, this just allows
7223 : * sanity-checking. In the case of an end-of-recovery checkpoint, we want
7224 : * to just temporarily disable writing until the system has exited
7225 : * recovery.
7226 : */
7227 2102 : if (shutdown)
7228 : {
7229 1046 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7230 58 : LocalXLogInsertAllowed = oldXLogAllowed;
7231 : else
7232 988 : LocalXLogInsertAllowed = 0; /* never again write WAL */
7233 : }
7234 :
7235 : /*
7236 : * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
7237 : * = end of actual checkpoint record.
7238 : */
7239 2102 : if (shutdown && checkPoint.redo != ProcLastRecPtr)
7240 0 : ereport(PANIC,
7241 : (errmsg("concurrent write-ahead log activity while database system is shutting down")));
7242 :
7243 : /*
7244 : * Remember the prior checkpoint's redo ptr for
7245 : * UpdateCheckPointDistanceEstimate()
7246 : */
7247 2102 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7248 :
7249 : /*
7250 : * Update the control file.
7251 : */
7252 2102 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7253 2102 : if (shutdown)
7254 1046 : ControlFile->state = DB_SHUTDOWNED;
7255 2102 : ControlFile->checkPoint = ProcLastRecPtr;
7256 2102 : ControlFile->checkPointCopy = checkPoint;
7257 : /* crash recovery should always recover to the end of WAL */
7258 2102 : ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
7259 2102 : ControlFile->minRecoveryPointTLI = 0;
7260 :
7261 : /*
7262 : * Persist unloggedLSN value. It's reset on crash recovery, so this goes
7263 : * unused on non-shutdown checkpoints, but seems useful to store it always
7264 : * for debugging purposes.
7265 : */
7266 2102 : ControlFile->unloggedLSN = pg_atomic_read_membarrier_u64(&XLogCtl->unloggedLSN);
7267 :
7268 2102 : UpdateControlFile();
7269 2102 : LWLockRelease(ControlFileLock);
7270 :
7271 : /* Update shared-memory copy of checkpoint XID/epoch */
7272 2102 : SpinLockAcquire(&XLogCtl->info_lck);
7273 2102 : XLogCtl->ckptFullXid = checkPoint.nextXid;
7274 2102 : SpinLockRelease(&XLogCtl->info_lck);
7275 :
7276 : /*
7277 : * We are now done with critical updates; no need for system panic if we
7278 : * have trouble while fooling with old log segments.
7279 : */
7280 2102 : END_CRIT_SECTION();
7281 :
7282 : /*
7283 : * WAL summaries end when the next XLOG_CHECKPOINT_REDO or
7284 : * XLOG_CHECKPOINT_SHUTDOWN record is reached. This is the first point
7285 : * where (a) we're not inside of a critical section and (b) we can be
7286 : * certain that the relevant record has been flushed to disk, which must
7287 : * happen before it can be summarized.
7288 : *
7289 : * If this is a shutdown checkpoint, then this happens reasonably
7290 : * promptly: we've only just inserted and flushed the
7291 : * XLOG_CHECKPOINT_SHUTDOWN record. If this is not a shutdown checkpoint,
7292 : * then this might not be very prompt at all: the XLOG_CHECKPOINT_REDO
7293 : * record was written before we began flushing data to disk, and that
7294 : * could be many minutes ago at this point. However, we don't XLogFlush()
7295 : * after inserting that record, so we're not guaranteed that it's on disk
7296 : * until after the above call that flushes the XLOG_CHECKPOINT_ONLINE
7297 : * record.
7298 : */
7299 2102 : WakeupWalSummarizer();
7300 :
7301 : /*
7302 : * Let smgr do post-checkpoint cleanup (eg, deleting old files).
7303 : */
7304 2102 : SyncPostCheckpoint();
7305 :
7306 : /*
7307 : * Update the average distance between checkpoints if the prior checkpoint
7308 : * exists.
7309 : */
7310 2102 : if (PriorRedoPtr != InvalidXLogRecPtr)
7311 2102 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7312 :
7313 : /*
7314 : * Delete old log files, those no longer needed for last checkpoint to
7315 : * prevent the disk holding the xlog from growing full.
7316 : */
7317 2102 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7318 2102 : KeepLogSeg(recptr, &_logSegNo);
7319 2102 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED,
7320 : _logSegNo, InvalidOid,
7321 : InvalidTransactionId))
7322 : {
7323 : /*
7324 : * Some slots have been invalidated; recalculate the old-segment
7325 : * horizon, starting again from RedoRecPtr.
7326 : */
7327 6 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7328 6 : KeepLogSeg(recptr, &_logSegNo);
7329 : }
7330 2102 : _logSegNo--;
7331 2102 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
7332 : checkPoint.ThisTimeLineID);
7333 :
7334 : /*
7335 : * Make more log segments if needed. (Do this after recycling old log
7336 : * segments, since that may supply some of the needed files.)
7337 : */
7338 2102 : if (!shutdown)
7339 1056 : PreallocXlogFiles(recptr, checkPoint.ThisTimeLineID);
7340 :
7341 : /*
7342 : * Truncate pg_subtrans if possible. We can throw away all data before
7343 : * the oldest XMIN of any running transaction. No future transaction will
7344 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7345 : * in subtrans.c). During recovery, though, we mustn't do this because
7346 : * StartupSUBTRANS hasn't been called yet.
7347 : */
7348 2102 : if (!RecoveryInProgress())
7349 2044 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7350 :
7351 : /* Real work is done; log and update stats. */
7352 2102 : LogCheckpointEnd(false);
7353 :
7354 : /* Reset the process title */
7355 2102 : update_checkpoint_display(flags, false, true);
7356 :
7357 : TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
7358 : NBuffers,
7359 : CheckpointStats.ckpt_segs_added,
7360 : CheckpointStats.ckpt_segs_removed,
7361 : CheckpointStats.ckpt_segs_recycled);
7362 :
7363 2102 : return true;
7364 : }
7365 :
7366 : /*
7367 : * Mark the end of recovery in WAL though without running a full checkpoint.
7368 : * We can expect that a restartpoint is likely to be in progress as we
7369 : * do this, though we are unwilling to wait for it to complete.
7370 : *
7371 : * CreateRestartPoint() allows for the case where recovery may end before
7372 : * the restartpoint completes so there is no concern of concurrent behaviour.
7373 : */
7374 : static void
7375 82 : CreateEndOfRecoveryRecord(void)
7376 : {
7377 : xl_end_of_recovery xlrec;
7378 : XLogRecPtr recptr;
7379 :
7380 : /* sanity check */
7381 82 : if (!RecoveryInProgress())
7382 0 : elog(ERROR, "can only be used to end recovery");
7383 :
7384 82 : xlrec.end_time = GetCurrentTimestamp();
7385 82 : xlrec.wal_level = wal_level;
7386 :
7387 82 : WALInsertLockAcquireExclusive();
7388 82 : xlrec.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7389 82 : xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7390 82 : WALInsertLockRelease();
7391 :
7392 82 : START_CRIT_SECTION();
7393 :
7394 82 : XLogBeginInsert();
7395 82 : XLogRegisterData((char *) &xlrec, sizeof(xl_end_of_recovery));
7396 82 : recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY);
7397 :
7398 82 : XLogFlush(recptr);
7399 :
7400 : /*
7401 : * Update the control file so that crash recovery can follow the timeline
7402 : * changes to this point.
7403 : */
7404 82 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7405 82 : ControlFile->minRecoveryPoint = recptr;
7406 82 : ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
7407 82 : UpdateControlFile();
7408 82 : LWLockRelease(ControlFileLock);
7409 :
7410 82 : END_CRIT_SECTION();
7411 82 : }
7412 :
7413 : /*
7414 : * Write an OVERWRITE_CONTRECORD message.
7415 : *
7416 : * When on WAL replay we expect a continuation record at the start of a page
7417 : * that is not there, recovery ends and WAL writing resumes at that point.
7418 : * But it's wrong to resume writing new WAL back at the start of the record
7419 : * that was broken, because downstream consumers of that WAL (physical
7420 : * replicas) are not prepared to "rewind". So the first action after
7421 : * finishing replay of all valid WAL must be to write a record of this type
7422 : * at the point where the contrecord was missing; to support xlogreader
7423 : * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
7424 : * to the page header where the record occurs. xlogreader has an ad-hoc
7425 : * mechanism to report metadata about the broken record, which is what we
7426 : * use here.
7427 : *
7428 : * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
7429 : * skip the record it was reading, and pass back the LSN of the skipped
7430 : * record, so that its caller can verify (on "replay" of that record) that the
7431 : * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
7432 : *
7433 : * 'aborted_lsn' is the beginning position of the record that was incomplete.
7434 : * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
7435 : * beginning of the XLOG page where the record is to be inserted. They must
7436 : * match the current WAL insert position, they're passed here just so that we
7437 : * can verify that.
7438 : */
7439 : static XLogRecPtr
7440 22 : CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn, XLogRecPtr pagePtr,
7441 : TimeLineID newTLI)
7442 : {
7443 : xl_overwrite_contrecord xlrec;
7444 : XLogRecPtr recptr;
7445 : XLogPageHeader pagehdr;
7446 : XLogRecPtr startPos;
7447 :
7448 : /* sanity checks */
7449 22 : if (!RecoveryInProgress())
7450 0 : elog(ERROR, "can only be used at end of recovery");
7451 22 : if (pagePtr % XLOG_BLCKSZ != 0)
7452 0 : elog(ERROR, "invalid position for missing continuation record %X/%X",
7453 : LSN_FORMAT_ARGS(pagePtr));
7454 :
7455 : /* The current WAL insert position should be right after the page header */
7456 22 : startPos = pagePtr;
7457 22 : if (XLogSegmentOffset(startPos, wal_segment_size) == 0)
7458 2 : startPos += SizeOfXLogLongPHD;
7459 : else
7460 20 : startPos += SizeOfXLogShortPHD;
7461 22 : recptr = GetXLogInsertRecPtr();
7462 22 : if (recptr != startPos)
7463 0 : elog(ERROR, "invalid WAL insert position %X/%X for OVERWRITE_CONTRECORD",
7464 : LSN_FORMAT_ARGS(recptr));
7465 :
7466 22 : START_CRIT_SECTION();
7467 :
7468 : /*
7469 : * Initialize the XLOG page header (by GetXLogBuffer), and set the
7470 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
7471 : *
7472 : * No other backend is allowed to write WAL yet, so acquiring the WAL
7473 : * insertion lock is just pro forma.
7474 : */
7475 22 : WALInsertLockAcquire();
7476 22 : pagehdr = (XLogPageHeader) GetXLogBuffer(pagePtr, newTLI);
7477 22 : pagehdr->xlp_info |= XLP_FIRST_IS_OVERWRITE_CONTRECORD;
7478 22 : WALInsertLockRelease();
7479 :
7480 : /*
7481 : * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
7482 : * page. We know it becomes the first record, because no other backend is
7483 : * allowed to write WAL yet.
7484 : */
7485 22 : XLogBeginInsert();
7486 22 : xlrec.overwritten_lsn = aborted_lsn;
7487 22 : xlrec.overwrite_time = GetCurrentTimestamp();
7488 22 : XLogRegisterData((char *) &xlrec, sizeof(xl_overwrite_contrecord));
7489 22 : recptr = XLogInsert(RM_XLOG_ID, XLOG_OVERWRITE_CONTRECORD);
7490 :
7491 : /* check that the record was inserted to the right place */
7492 22 : if (ProcLastRecPtr != startPos)
7493 0 : elog(ERROR, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%X",
7494 : LSN_FORMAT_ARGS(ProcLastRecPtr));
7495 :
7496 22 : XLogFlush(recptr);
7497 :
7498 22 : END_CRIT_SECTION();
7499 :
7500 22 : return recptr;
7501 : }
7502 :
7503 : /*
7504 : * Flush all data in shared memory to disk, and fsync
7505 : *
7506 : * This is the common code shared between regular checkpoints and
7507 : * recovery restartpoints.
7508 : */
7509 : static void
7510 2468 : CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
7511 : {
7512 2468 : CheckPointRelationMap();
7513 2468 : CheckPointReplicationSlots(flags & CHECKPOINT_IS_SHUTDOWN);
7514 2468 : CheckPointSnapBuild();
7515 2468 : CheckPointLogicalRewriteHeap();
7516 2468 : CheckPointReplicationOrigin();
7517 :
7518 : /* Write out all dirty data in SLRUs and the main buffer pool */
7519 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags);
7520 2468 : CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
7521 2468 : CheckPointCLOG();
7522 2468 : CheckPointCommitTs();
7523 2468 : CheckPointSUBTRANS();
7524 2468 : CheckPointMultiXact();
7525 2468 : CheckPointPredicate();
7526 2468 : CheckPointBuffers(flags);
7527 :
7528 : /* Perform all queued up fsyncs */
7529 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
7530 2468 : CheckpointStats.ckpt_sync_t = GetCurrentTimestamp();
7531 2468 : ProcessSyncRequests();
7532 2468 : CheckpointStats.ckpt_sync_end_t = GetCurrentTimestamp();
7533 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
7534 :
7535 : /* We deliberately delay 2PC checkpointing as long as possible */
7536 2468 : CheckPointTwoPhase(checkPointRedo);
7537 2468 : }
7538 :
7539 : /*
7540 : * Save a checkpoint for recovery restart if appropriate
7541 : *
7542 : * This function is called each time a checkpoint record is read from XLOG.
7543 : * It must determine whether the checkpoint represents a safe restartpoint or
7544 : * not. If so, the checkpoint record is stashed in shared memory so that
7545 : * CreateRestartPoint can consult it. (Note that the latter function is
7546 : * executed by the checkpointer, while this one will be executed by the
7547 : * startup process.)
7548 : */
7549 : static void
7550 744 : RecoveryRestartPoint(const CheckPoint *checkPoint, XLogReaderState *record)
7551 : {
7552 : /*
7553 : * Also refrain from creating a restartpoint if we have seen any
7554 : * references to non-existent pages. Restarting recovery from the
7555 : * restartpoint would not see the references, so we would lose the
7556 : * cross-check that the pages belonged to a relation that was dropped
7557 : * later.
7558 : */
7559 744 : if (XLogHaveInvalidPages())
7560 : {
7561 0 : elog(DEBUG2,
7562 : "could not record restart point at %X/%X because there "
7563 : "are unresolved references to invalid pages",
7564 : LSN_FORMAT_ARGS(checkPoint->redo));
7565 0 : return;
7566 : }
7567 :
7568 : /*
7569 : * Copy the checkpoint record to shared memory, so that checkpointer can
7570 : * work out the next time it wants to perform a restartpoint.
7571 : */
7572 744 : SpinLockAcquire(&XLogCtl->info_lck);
7573 744 : XLogCtl->lastCheckPointRecPtr = record->ReadRecPtr;
7574 744 : XLogCtl->lastCheckPointEndPtr = record->EndRecPtr;
7575 744 : XLogCtl->lastCheckPoint = *checkPoint;
7576 744 : SpinLockRelease(&XLogCtl->info_lck);
7577 : }
7578 :
7579 : /*
7580 : * Establish a restartpoint if possible.
7581 : *
7582 : * This is similar to CreateCheckPoint, but is used during WAL recovery
7583 : * to establish a point from which recovery can roll forward without
7584 : * replaying the entire recovery log.
7585 : *
7586 : * Returns true if a new restartpoint was established. We can only establish
7587 : * a restartpoint if we have replayed a safe checkpoint record since last
7588 : * restartpoint.
7589 : */
7590 : bool
7591 522 : CreateRestartPoint(int flags)
7592 : {
7593 : XLogRecPtr lastCheckPointRecPtr;
7594 : XLogRecPtr lastCheckPointEndPtr;
7595 : CheckPoint lastCheckPoint;
7596 : XLogRecPtr PriorRedoPtr;
7597 : XLogRecPtr receivePtr;
7598 : XLogRecPtr replayPtr;
7599 : TimeLineID replayTLI;
7600 : XLogRecPtr endptr;
7601 : XLogSegNo _logSegNo;
7602 : TimestampTz xtime;
7603 :
7604 : /* Concurrent checkpoint/restartpoint cannot happen */
7605 : Assert(!IsUnderPostmaster || MyBackendType == B_CHECKPOINTER);
7606 :
7607 : /* Get a local copy of the last safe checkpoint record. */
7608 522 : SpinLockAcquire(&XLogCtl->info_lck);
7609 522 : lastCheckPointRecPtr = XLogCtl->lastCheckPointRecPtr;
7610 522 : lastCheckPointEndPtr = XLogCtl->lastCheckPointEndPtr;
7611 522 : lastCheckPoint = XLogCtl->lastCheckPoint;
7612 522 : SpinLockRelease(&XLogCtl->info_lck);
7613 :
7614 : /*
7615 : * Check that we're still in recovery mode. It's ok if we exit recovery
7616 : * mode after this check, the restart point is valid anyway.
7617 : */
7618 522 : if (!RecoveryInProgress())
7619 : {
7620 0 : ereport(DEBUG2,
7621 : (errmsg_internal("skipping restartpoint, recovery has already ended")));
7622 0 : return false;
7623 : }
7624 :
7625 : /*
7626 : * If the last checkpoint record we've replayed is already our last
7627 : * restartpoint, we can't perform a new restart point. We still update
7628 : * minRecoveryPoint in that case, so that if this is a shutdown restart
7629 : * point, we won't start up earlier than before. That's not strictly
7630 : * necessary, but when hot standby is enabled, it would be rather weird if
7631 : * the database opened up for read-only connections at a point-in-time
7632 : * before the last shutdown. Such time travel is still possible in case of
7633 : * immediate shutdown, though.
7634 : *
7635 : * We don't explicitly advance minRecoveryPoint when we do create a
7636 : * restartpoint. It's assumed that flushing the buffers will do that as a
7637 : * side-effect.
7638 : */
7639 522 : if (XLogRecPtrIsInvalid(lastCheckPointRecPtr) ||
7640 506 : lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
7641 : {
7642 156 : ereport(DEBUG2,
7643 : (errmsg_internal("skipping restartpoint, already performed at %X/%X",
7644 : LSN_FORMAT_ARGS(lastCheckPoint.redo))));
7645 :
7646 156 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
7647 156 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7648 : {
7649 60 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7650 60 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7651 60 : UpdateControlFile();
7652 60 : LWLockRelease(ControlFileLock);
7653 : }
7654 156 : return false;
7655 : }
7656 :
7657 : /*
7658 : * Update the shared RedoRecPtr so that the startup process can calculate
7659 : * the number of segments replayed since last restartpoint, and request a
7660 : * restartpoint if it exceeds CheckPointSegments.
7661 : *
7662 : * Like in CreateCheckPoint(), hold off insertions to update it, although
7663 : * during recovery this is just pro forma, because no WAL insertions are
7664 : * happening.
7665 : */
7666 366 : WALInsertLockAcquireExclusive();
7667 366 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = lastCheckPoint.redo;
7668 366 : WALInsertLockRelease();
7669 :
7670 : /* Also update the info_lck-protected copy */
7671 366 : SpinLockAcquire(&XLogCtl->info_lck);
7672 366 : XLogCtl->RedoRecPtr = lastCheckPoint.redo;
7673 366 : SpinLockRelease(&XLogCtl->info_lck);
7674 :
7675 : /*
7676 : * Prepare to accumulate statistics.
7677 : *
7678 : * Note: because it is possible for log_checkpoints to change while a
7679 : * checkpoint proceeds, we always accumulate stats, even if
7680 : * log_checkpoints is currently off.
7681 : */
7682 4026 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7683 366 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7684 :
7685 366 : if (log_checkpoints)
7686 366 : LogCheckpointStart(flags, true);
7687 :
7688 : /* Update the process title */
7689 366 : update_checkpoint_display(flags, true, false);
7690 :
7691 366 : CheckPointGuts(lastCheckPoint.redo, flags);
7692 :
7693 : /*
7694 : * This location needs to be after CheckPointGuts() to ensure that some
7695 : * work has already happened during this checkpoint.
7696 : */
7697 366 : INJECTION_POINT("create-restart-point");
7698 :
7699 : /*
7700 : * Remember the prior checkpoint's redo ptr for
7701 : * UpdateCheckPointDistanceEstimate()
7702 : */
7703 366 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7704 :
7705 : /*
7706 : * Update pg_control, using current time. Check that it still shows an
7707 : * older checkpoint, else do nothing; this is a quick hack to make sure
7708 : * nothing really bad happens if somehow we get here after the
7709 : * end-of-recovery checkpoint.
7710 : */
7711 366 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7712 366 : if (ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
7713 : {
7714 : /*
7715 : * Update the checkpoint information. We do this even if the cluster
7716 : * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
7717 : * segments recycled below.
7718 : */
7719 366 : ControlFile->checkPoint = lastCheckPointRecPtr;
7720 366 : ControlFile->checkPointCopy = lastCheckPoint;
7721 :
7722 : /*
7723 : * Ensure minRecoveryPoint is past the checkpoint record and update it
7724 : * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
7725 : * this will have happened already while writing out dirty buffers,
7726 : * but not necessarily - e.g. because no buffers were dirtied. We do
7727 : * this because a backup performed in recovery uses minRecoveryPoint
7728 : * to determine which WAL files must be included in the backup, and
7729 : * the file (or files) containing the checkpoint record must be
7730 : * included, at a minimum. Note that for an ordinary restart of
7731 : * recovery there's no value in having the minimum recovery point any
7732 : * earlier than this anyway, because redo will begin just after the
7733 : * checkpoint record.
7734 : */
7735 366 : if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
7736 : {
7737 366 : if (ControlFile->minRecoveryPoint < lastCheckPointEndPtr)
7738 : {
7739 38 : ControlFile->minRecoveryPoint = lastCheckPointEndPtr;
7740 38 : ControlFile->minRecoveryPointTLI = lastCheckPoint.ThisTimeLineID;
7741 :
7742 : /* update local copy */
7743 38 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
7744 38 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
7745 : }
7746 366 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7747 40 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7748 : }
7749 366 : UpdateControlFile();
7750 : }
7751 366 : LWLockRelease(ControlFileLock);
7752 :
7753 : /*
7754 : * Update the average distance between checkpoints/restartpoints if the
7755 : * prior checkpoint exists.
7756 : */
7757 366 : if (PriorRedoPtr != InvalidXLogRecPtr)
7758 366 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7759 :
7760 : /*
7761 : * Delete old log files, those no longer needed for last restartpoint to
7762 : * prevent the disk holding the xlog from growing full.
7763 : */
7764 366 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7765 :
7766 : /*
7767 : * Retreat _logSegNo using the current end of xlog replayed or received,
7768 : * whichever is later.
7769 : */
7770 366 : receivePtr = GetWalRcvFlushRecPtr(NULL, NULL);
7771 366 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
7772 366 : endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
7773 366 : KeepLogSeg(endptr, &_logSegNo);
7774 366 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED,
7775 : _logSegNo, InvalidOid,
7776 : InvalidTransactionId))
7777 : {
7778 : /*
7779 : * Some slots have been invalidated; recalculate the old-segment
7780 : * horizon, starting again from RedoRecPtr.
7781 : */
7782 2 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7783 2 : KeepLogSeg(endptr, &_logSegNo);
7784 : }
7785 366 : _logSegNo--;
7786 :
7787 : /*
7788 : * Try to recycle segments on a useful timeline. If we've been promoted
7789 : * since the beginning of this restartpoint, use the new timeline chosen
7790 : * at end of recovery. If we're still in recovery, use the timeline we're
7791 : * currently replaying.
7792 : *
7793 : * There is no guarantee that the WAL segments will be useful on the
7794 : * current timeline; if recovery proceeds to a new timeline right after
7795 : * this, the pre-allocated WAL segments on this timeline will not be used,
7796 : * and will go wasted until recycled on the next restartpoint. We'll live
7797 : * with that.
7798 : */
7799 366 : if (!RecoveryInProgress())
7800 0 : replayTLI = XLogCtl->InsertTimeLineID;
7801 :
7802 366 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, endptr, replayTLI);
7803 :
7804 : /*
7805 : * Make more log segments if needed. (Do this after recycling old log
7806 : * segments, since that may supply some of the needed files.)
7807 : */
7808 366 : PreallocXlogFiles(endptr, replayTLI);
7809 :
7810 : /*
7811 : * Truncate pg_subtrans if possible. We can throw away all data before
7812 : * the oldest XMIN of any running transaction. No future transaction will
7813 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7814 : * in subtrans.c). When hot standby is disabled, though, we mustn't do
7815 : * this because StartupSUBTRANS hasn't been called yet.
7816 : */
7817 366 : if (EnableHotStandby)
7818 366 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7819 :
7820 : /* Real work is done; log and update stats. */
7821 366 : LogCheckpointEnd(true);
7822 :
7823 : /* Reset the process title */
7824 366 : update_checkpoint_display(flags, true, true);
7825 :
7826 366 : xtime = GetLatestXTime();
7827 366 : ereport((log_checkpoints ? LOG : DEBUG2),
7828 : (errmsg("recovery restart point at %X/%X",
7829 : LSN_FORMAT_ARGS(lastCheckPoint.redo)),
7830 : xtime ? errdetail("Last completed transaction was at log time %s.",
7831 : timestamptz_to_str(xtime)) : 0));
7832 :
7833 : /*
7834 : * Finally, execute archive_cleanup_command, if any.
7835 : */
7836 366 : if (archiveCleanupCommand && strcmp(archiveCleanupCommand, "") != 0)
7837 0 : ExecuteRecoveryCommand(archiveCleanupCommand,
7838 : "archive_cleanup_command",
7839 : false,
7840 : WAIT_EVENT_ARCHIVE_CLEANUP_COMMAND);
7841 :
7842 366 : return true;
7843 : }
7844 :
7845 : /*
7846 : * Report availability of WAL for the given target LSN
7847 : * (typically a slot's restart_lsn)
7848 : *
7849 : * Returns one of the following enum values:
7850 : *
7851 : * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
7852 : * max_wal_size.
7853 : *
7854 : * * WALAVAIL_EXTENDED means it is still available by preserving extra
7855 : * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
7856 : * than max_wal_size, this state is not returned.
7857 : *
7858 : * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
7859 : * remove reserved segments. The walsender using this slot may return to the
7860 : * above.
7861 : *
7862 : * * WALAVAIL_REMOVED means it has been removed. A replication stream on
7863 : * a slot with this LSN cannot continue. (Any associated walsender
7864 : * processes should have been terminated already.)
7865 : *
7866 : * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
7867 : */
7868 : WALAvailability
7869 850 : GetWALAvailability(XLogRecPtr targetLSN)
7870 : {
7871 : XLogRecPtr currpos; /* current write LSN */
7872 : XLogSegNo currSeg; /* segid of currpos */
7873 : XLogSegNo targetSeg; /* segid of targetLSN */
7874 : XLogSegNo oldestSeg; /* actual oldest segid */
7875 : XLogSegNo oldestSegMaxWalSize; /* oldest segid kept by max_wal_size */
7876 : XLogSegNo oldestSlotSeg; /* oldest segid kept by slot */
7877 : uint64 keepSegs;
7878 :
7879 : /*
7880 : * slot does not reserve WAL. Either deactivated, or has never been active
7881 : */
7882 850 : if (XLogRecPtrIsInvalid(targetLSN))
7883 28 : return WALAVAIL_INVALID_LSN;
7884 :
7885 : /*
7886 : * Calculate the oldest segment currently reserved by all slots,
7887 : * considering wal_keep_size and max_slot_wal_keep_size. Initialize
7888 : * oldestSlotSeg to the current segment.
7889 : */
7890 822 : currpos = GetXLogWriteRecPtr();
7891 822 : XLByteToSeg(currpos, oldestSlotSeg, wal_segment_size);
7892 822 : KeepLogSeg(currpos, &oldestSlotSeg);
7893 :
7894 : /*
7895 : * Find the oldest extant segment file. We get 1 until checkpoint removes
7896 : * the first WAL segment file since startup, which causes the status being
7897 : * wrong under certain abnormal conditions but that doesn't actually harm.
7898 : */
7899 822 : oldestSeg = XLogGetLastRemovedSegno() + 1;
7900 :
7901 : /* calculate oldest segment by max_wal_size */
7902 822 : XLByteToSeg(currpos, currSeg, wal_segment_size);
7903 822 : keepSegs = ConvertToXSegs(max_wal_size_mb, wal_segment_size) + 1;
7904 :
7905 822 : if (currSeg > keepSegs)
7906 16 : oldestSegMaxWalSize = currSeg - keepSegs;
7907 : else
7908 806 : oldestSegMaxWalSize = 1;
7909 :
7910 : /* the segment we care about */
7911 822 : XLByteToSeg(targetLSN, targetSeg, wal_segment_size);
7912 :
7913 : /*
7914 : * No point in returning reserved or extended status values if the
7915 : * targetSeg is known to be lost.
7916 : */
7917 822 : if (targetSeg >= oldestSlotSeg)
7918 : {
7919 : /* show "reserved" when targetSeg is within max_wal_size */
7920 820 : if (targetSeg >= oldestSegMaxWalSize)
7921 816 : return WALAVAIL_RESERVED;
7922 :
7923 : /* being retained by slots exceeding max_wal_size */
7924 4 : return WALAVAIL_EXTENDED;
7925 : }
7926 :
7927 : /* WAL segments are no longer retained but haven't been removed yet */
7928 2 : if (targetSeg >= oldestSeg)
7929 2 : return WALAVAIL_UNRESERVED;
7930 :
7931 : /* Definitely lost */
7932 0 : return WALAVAIL_REMOVED;
7933 : }
7934 :
7935 :
7936 : /*
7937 : * Retreat *logSegNo to the last segment that we need to retain because of
7938 : * either wal_keep_size or replication slots.
7939 : *
7940 : * This is calculated by subtracting wal_keep_size from the given xlog
7941 : * location, recptr and by making sure that that result is below the
7942 : * requirement of replication slots. For the latter criterion we do consider
7943 : * the effects of max_slot_wal_keep_size: reserve at most that much space back
7944 : * from recptr.
7945 : *
7946 : * Note about replication slots: if this function calculates a value
7947 : * that's further ahead than what slots need reserved, then affected
7948 : * slots need to be invalidated and this function invoked again.
7949 : * XXX it might be a good idea to rewrite this function so that
7950 : * invalidation is optionally done here, instead.
7951 : */
7952 : static void
7953 3298 : KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
7954 : {
7955 : XLogSegNo currSegNo;
7956 : XLogSegNo segno;
7957 : XLogRecPtr keep;
7958 :
7959 3298 : XLByteToSeg(recptr, currSegNo, wal_segment_size);
7960 3298 : segno = currSegNo;
7961 :
7962 : /*
7963 : * Calculate how many segments are kept by slots first, adjusting for
7964 : * max_slot_wal_keep_size.
7965 : */
7966 3298 : keep = XLogGetReplicationSlotMinimumLSN();
7967 3298 : if (keep != InvalidXLogRecPtr && keep < recptr)
7968 : {
7969 1044 : XLByteToSeg(keep, segno, wal_segment_size);
7970 :
7971 : /* Cap by max_slot_wal_keep_size ... */
7972 1044 : if (max_slot_wal_keep_size_mb >= 0)
7973 : {
7974 : uint64 slot_keep_segs;
7975 :
7976 38 : slot_keep_segs =
7977 38 : ConvertToXSegs(max_slot_wal_keep_size_mb, wal_segment_size);
7978 :
7979 38 : if (currSegNo - segno > slot_keep_segs)
7980 10 : segno = currSegNo - slot_keep_segs;
7981 : }
7982 : }
7983 :
7984 : /*
7985 : * If WAL summarization is in use, don't remove WAL that has yet to be
7986 : * summarized.
7987 : */
7988 3298 : keep = GetOldestUnsummarizedLSN(NULL, NULL);
7989 3298 : if (keep != InvalidXLogRecPtr)
7990 : {
7991 : XLogSegNo unsummarized_segno;
7992 :
7993 4 : XLByteToSeg(keep, unsummarized_segno, wal_segment_size);
7994 4 : if (unsummarized_segno < segno)
7995 4 : segno = unsummarized_segno;
7996 : }
7997 :
7998 : /* but, keep at least wal_keep_size if that's set */
7999 3298 : if (wal_keep_size_mb > 0)
8000 : {
8001 : uint64 keep_segs;
8002 :
8003 138 : keep_segs = ConvertToXSegs(wal_keep_size_mb, wal_segment_size);
8004 138 : if (currSegNo - segno < keep_segs)
8005 : {
8006 : /* avoid underflow, don't go below 1 */
8007 138 : if (currSegNo <= keep_segs)
8008 130 : segno = 1;
8009 : else
8010 8 : segno = currSegNo - keep_segs;
8011 : }
8012 : }
8013 :
8014 : /* don't delete WAL segments newer than the calculated segment */
8015 3298 : if (segno < *logSegNo)
8016 606 : *logSegNo = segno;
8017 3298 : }
8018 :
8019 : /*
8020 : * Write a NEXTOID log record
8021 : */
8022 : void
8023 1074 : XLogPutNextOid(Oid nextOid)
8024 : {
8025 1074 : XLogBeginInsert();
8026 1074 : XLogRegisterData((char *) (&nextOid), sizeof(Oid));
8027 1074 : (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID);
8028 :
8029 : /*
8030 : * We need not flush the NEXTOID record immediately, because any of the
8031 : * just-allocated OIDs could only reach disk as part of a tuple insert or
8032 : * update that would have its own XLOG record that must follow the NEXTOID
8033 : * record. Therefore, the standard buffer LSN interlock applied to those
8034 : * records will ensure no such OID reaches disk before the NEXTOID record
8035 : * does.
8036 : *
8037 : * Note, however, that the above statement only covers state "within" the
8038 : * database. When we use a generated OID as a file or directory name, we
8039 : * are in a sense violating the basic WAL rule, because that filesystem
8040 : * change may reach disk before the NEXTOID WAL record does. The impact
8041 : * of this is that if a database crash occurs immediately afterward, we
8042 : * might after restart re-generate the same OID and find that it conflicts
8043 : * with the leftover file or directory. But since for safety's sake we
8044 : * always loop until finding a nonconflicting filename, this poses no real
8045 : * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8046 : */
8047 1074 : }
8048 :
8049 : /*
8050 : * Write an XLOG SWITCH record.
8051 : *
8052 : * Here we just blindly issue an XLogInsert request for the record.
8053 : * All the magic happens inside XLogInsert.
8054 : *
8055 : * The return value is either the end+1 address of the switch record,
8056 : * or the end+1 address of the prior segment if we did not need to
8057 : * write a switch record because we are already at segment start.
8058 : */
8059 : XLogRecPtr
8060 736 : RequestXLogSwitch(bool mark_unimportant)
8061 : {
8062 : XLogRecPtr RecPtr;
8063 :
8064 : /* XLOG SWITCH has no data */
8065 736 : XLogBeginInsert();
8066 :
8067 736 : if (mark_unimportant)
8068 0 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8069 736 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH);
8070 :
8071 736 : return RecPtr;
8072 : }
8073 :
8074 : /*
8075 : * Write a RESTORE POINT record
8076 : */
8077 : XLogRecPtr
8078 6 : XLogRestorePoint(const char *rpName)
8079 : {
8080 : XLogRecPtr RecPtr;
8081 : xl_restore_point xlrec;
8082 :
8083 6 : xlrec.rp_time = GetCurrentTimestamp();
8084 6 : strlcpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8085 :
8086 6 : XLogBeginInsert();
8087 6 : XLogRegisterData((char *) &xlrec, sizeof(xl_restore_point));
8088 :
8089 6 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT);
8090 :
8091 6 : ereport(LOG,
8092 : (errmsg("restore point \"%s\" created at %X/%X",
8093 : rpName, LSN_FORMAT_ARGS(RecPtr))));
8094 :
8095 6 : return RecPtr;
8096 : }
8097 :
8098 : /*
8099 : * Check if any of the GUC parameters that are critical for hot standby
8100 : * have changed, and update the value in pg_control file if necessary.
8101 : */
8102 : static void
8103 1534 : XLogReportParameters(void)
8104 : {
8105 1534 : if (wal_level != ControlFile->wal_level ||
8106 1132 : wal_log_hints != ControlFile->wal_log_hints ||
8107 978 : MaxConnections != ControlFile->MaxConnections ||
8108 976 : max_worker_processes != ControlFile->max_worker_processes ||
8109 974 : max_wal_senders != ControlFile->max_wal_senders ||
8110 952 : max_prepared_xacts != ControlFile->max_prepared_xacts ||
8111 780 : max_locks_per_xact != ControlFile->max_locks_per_xact ||
8112 780 : track_commit_timestamp != ControlFile->track_commit_timestamp)
8113 : {
8114 : /*
8115 : * The change in number of backend slots doesn't need to be WAL-logged
8116 : * if archiving is not enabled, as you can't start archive recovery
8117 : * with wal_level=minimal anyway. We don't really care about the
8118 : * values in pg_control either if wal_level=minimal, but seems better
8119 : * to keep them up-to-date to avoid confusion.
8120 : */
8121 774 : if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8122 : {
8123 : xl_parameter_change xlrec;
8124 : XLogRecPtr recptr;
8125 :
8126 752 : xlrec.MaxConnections = MaxConnections;
8127 752 : xlrec.max_worker_processes = max_worker_processes;
8128 752 : xlrec.max_wal_senders = max_wal_senders;
8129 752 : xlrec.max_prepared_xacts = max_prepared_xacts;
8130 752 : xlrec.max_locks_per_xact = max_locks_per_xact;
8131 752 : xlrec.wal_level = wal_level;
8132 752 : xlrec.wal_log_hints = wal_log_hints;
8133 752 : xlrec.track_commit_timestamp = track_commit_timestamp;
8134 :
8135 752 : XLogBeginInsert();
8136 752 : XLogRegisterData((char *) &xlrec, sizeof(xlrec));
8137 :
8138 752 : recptr = XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE);
8139 752 : XLogFlush(recptr);
8140 : }
8141 :
8142 774 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8143 :
8144 774 : ControlFile->MaxConnections = MaxConnections;
8145 774 : ControlFile->max_worker_processes = max_worker_processes;
8146 774 : ControlFile->max_wal_senders = max_wal_senders;
8147 774 : ControlFile->max_prepared_xacts = max_prepared_xacts;
8148 774 : ControlFile->max_locks_per_xact = max_locks_per_xact;
8149 774 : ControlFile->wal_level = wal_level;
8150 774 : ControlFile->wal_log_hints = wal_log_hints;
8151 774 : ControlFile->track_commit_timestamp = track_commit_timestamp;
8152 774 : UpdateControlFile();
8153 :
8154 774 : LWLockRelease(ControlFileLock);
8155 : }
8156 1534 : }
8157 :
8158 : /*
8159 : * Update full_page_writes in shared memory, and write an
8160 : * XLOG_FPW_CHANGE record if necessary.
8161 : *
8162 : * Note: this function assumes there is no other process running
8163 : * concurrently that could update it.
8164 : */
8165 : void
8166 2532 : UpdateFullPageWrites(void)
8167 : {
8168 2532 : XLogCtlInsert *Insert = &XLogCtl->Insert;
8169 : bool recoveryInProgress;
8170 :
8171 : /*
8172 : * Do nothing if full_page_writes has not been changed.
8173 : *
8174 : * It's safe to check the shared full_page_writes without the lock,
8175 : * because we assume that there is no concurrently running process which
8176 : * can update it.
8177 : */
8178 2532 : if (fullPageWrites == Insert->fullPageWrites)
8179 1656 : return;
8180 :
8181 : /*
8182 : * Perform this outside critical section so that the WAL insert
8183 : * initialization done by RecoveryInProgress() doesn't trigger an
8184 : * assertion failure.
8185 : */
8186 876 : recoveryInProgress = RecoveryInProgress();
8187 :
8188 876 : START_CRIT_SECTION();
8189 :
8190 : /*
8191 : * It's always safe to take full page images, even when not strictly
8192 : * required, but not the other round. So if we're setting full_page_writes
8193 : * to true, first set it true and then write the WAL record. If we're
8194 : * setting it to false, first write the WAL record and then set the global
8195 : * flag.
8196 : */
8197 876 : if (fullPageWrites)
8198 : {
8199 868 : WALInsertLockAcquireExclusive();
8200 868 : Insert->fullPageWrites = true;
8201 868 : WALInsertLockRelease();
8202 : }
8203 :
8204 : /*
8205 : * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8206 : * full_page_writes during archive recovery, if required.
8207 : */
8208 876 : if (XLogStandbyInfoActive() && !recoveryInProgress)
8209 : {
8210 0 : XLogBeginInsert();
8211 0 : XLogRegisterData((char *) (&fullPageWrites), sizeof(bool));
8212 :
8213 0 : XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
8214 : }
8215 :
8216 876 : if (!fullPageWrites)
8217 : {
8218 8 : WALInsertLockAcquireExclusive();
8219 8 : Insert->fullPageWrites = false;
8220 8 : WALInsertLockRelease();
8221 : }
8222 876 : END_CRIT_SECTION();
8223 : }
8224 :
8225 : /*
8226 : * XLOG resource manager's routines
8227 : *
8228 : * Definitions of info values are in include/catalog/pg_control.h, though
8229 : * not all record types are related to control file updates.
8230 : *
8231 : * NOTE: Some XLOG record types that are directly related to WAL recovery
8232 : * are handled in xlogrecovery_redo().
8233 : */
8234 : void
8235 79122 : xlog_redo(XLogReaderState *record)
8236 : {
8237 79122 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8238 79122 : XLogRecPtr lsn = record->EndRecPtr;
8239 :
8240 : /*
8241 : * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
8242 : * XLOG_FPI_FOR_HINT records.
8243 : */
8244 : Assert(info == XLOG_FPI || info == XLOG_FPI_FOR_HINT ||
8245 : !XLogRecHasAnyBlockRefs(record));
8246 :
8247 79122 : if (info == XLOG_NEXTOID)
8248 : {
8249 : Oid nextOid;
8250 :
8251 : /*
8252 : * We used to try to take the maximum of TransamVariables->nextOid and
8253 : * the recorded nextOid, but that fails if the OID counter wraps
8254 : * around. Since no OID allocation should be happening during replay
8255 : * anyway, better to just believe the record exactly. We still take
8256 : * OidGenLock while setting the variable, just in case.
8257 : */
8258 178 : memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
8259 178 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8260 178 : TransamVariables->nextOid = nextOid;
8261 178 : TransamVariables->oidCount = 0;
8262 178 : LWLockRelease(OidGenLock);
8263 : }
8264 78944 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
8265 : {
8266 : CheckPoint checkPoint;
8267 : TimeLineID replayTLI;
8268 :
8269 68 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8270 : /* In a SHUTDOWN checkpoint, believe the counters exactly */
8271 68 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8272 68 : TransamVariables->nextXid = checkPoint.nextXid;
8273 68 : LWLockRelease(XidGenLock);
8274 68 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8275 68 : TransamVariables->nextOid = checkPoint.nextOid;
8276 68 : TransamVariables->oidCount = 0;
8277 68 : LWLockRelease(OidGenLock);
8278 68 : MultiXactSetNextMXact(checkPoint.nextMulti,
8279 : checkPoint.nextMultiOffset);
8280 :
8281 68 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8282 : checkPoint.oldestMultiDB);
8283 :
8284 : /*
8285 : * No need to set oldestClogXid here as well; it'll be set when we
8286 : * redo an xl_clog_truncate if it changed since initialization.
8287 : */
8288 68 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
8289 :
8290 : /*
8291 : * If we see a shutdown checkpoint while waiting for an end-of-backup
8292 : * record, the backup was canceled and the end-of-backup record will
8293 : * never arrive.
8294 : */
8295 68 : if (ArchiveRecoveryRequested &&
8296 66 : !XLogRecPtrIsInvalid(ControlFile->backupStartPoint) &&
8297 0 : XLogRecPtrIsInvalid(ControlFile->backupEndPoint))
8298 0 : ereport(PANIC,
8299 : (errmsg("online backup was canceled, recovery cannot continue")));
8300 :
8301 : /*
8302 : * If we see a shutdown checkpoint, we know that nothing was running
8303 : * on the primary at this point. So fake-up an empty running-xacts
8304 : * record and use that here and now. Recover additional standby state
8305 : * for prepared transactions.
8306 : */
8307 68 : if (standbyState >= STANDBY_INITIALIZED)
8308 : {
8309 : TransactionId *xids;
8310 : int nxids;
8311 : TransactionId oldestActiveXID;
8312 : TransactionId latestCompletedXid;
8313 : RunningTransactionsData running;
8314 :
8315 62 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
8316 :
8317 : /* Update pg_subtrans entries for any prepared transactions */
8318 62 : StandbyRecoverPreparedTransactions();
8319 :
8320 : /*
8321 : * Construct a RunningTransactions snapshot representing a shut
8322 : * down server, with only prepared transactions still alive. We're
8323 : * never overflowed at this point because all subxids are listed
8324 : * with their parent prepared transactions.
8325 : */
8326 62 : running.xcnt = nxids;
8327 62 : running.subxcnt = 0;
8328 62 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
8329 62 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
8330 62 : running.oldestRunningXid = oldestActiveXID;
8331 62 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
8332 62 : TransactionIdRetreat(latestCompletedXid);
8333 : Assert(TransactionIdIsNormal(latestCompletedXid));
8334 62 : running.latestCompletedXid = latestCompletedXid;
8335 62 : running.xids = xids;
8336 :
8337 62 : ProcArrayApplyRecoveryInfo(&running);
8338 : }
8339 :
8340 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8341 68 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8342 68 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8343 68 : LWLockRelease(ControlFileLock);
8344 :
8345 : /* Update shared-memory copy of checkpoint XID/epoch */
8346 68 : SpinLockAcquire(&XLogCtl->info_lck);
8347 68 : XLogCtl->ckptFullXid = checkPoint.nextXid;
8348 68 : SpinLockRelease(&XLogCtl->info_lck);
8349 :
8350 : /*
8351 : * We should've already switched to the new TLI before replaying this
8352 : * record.
8353 : */
8354 68 : (void) GetCurrentReplayRecPtr(&replayTLI);
8355 68 : if (checkPoint.ThisTimeLineID != replayTLI)
8356 0 : ereport(PANIC,
8357 : (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
8358 : checkPoint.ThisTimeLineID, replayTLI)));
8359 :
8360 68 : RecoveryRestartPoint(&checkPoint, record);
8361 : }
8362 78876 : else if (info == XLOG_CHECKPOINT_ONLINE)
8363 : {
8364 : CheckPoint checkPoint;
8365 : TimeLineID replayTLI;
8366 :
8367 676 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8368 : /* In an ONLINE checkpoint, treat the XID counter as a minimum */
8369 676 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8370 676 : if (FullTransactionIdPrecedes(TransamVariables->nextXid,
8371 : checkPoint.nextXid))
8372 0 : TransamVariables->nextXid = checkPoint.nextXid;
8373 676 : LWLockRelease(XidGenLock);
8374 :
8375 : /*
8376 : * We ignore the nextOid counter in an ONLINE checkpoint, preferring
8377 : * to track OID assignment through XLOG_NEXTOID records. The nextOid
8378 : * counter is from the start of the checkpoint and might well be stale
8379 : * compared to later XLOG_NEXTOID records. We could try to take the
8380 : * maximum of the nextOid counter and our latest value, but since
8381 : * there's no particular guarantee about the speed with which the OID
8382 : * counter wraps around, that's a risky thing to do. In any case,
8383 : * users of the nextOid counter are required to avoid assignment of
8384 : * duplicates, so that a somewhat out-of-date value should be safe.
8385 : */
8386 :
8387 : /* Handle multixact */
8388 676 : MultiXactAdvanceNextMXact(checkPoint.nextMulti,
8389 : checkPoint.nextMultiOffset);
8390 :
8391 : /*
8392 : * NB: This may perform multixact truncation when replaying WAL
8393 : * generated by an older primary.
8394 : */
8395 676 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8396 : checkPoint.oldestMultiDB);
8397 676 : if (TransactionIdPrecedes(TransamVariables->oldestXid,
8398 : checkPoint.oldestXid))
8399 0 : SetTransactionIdLimit(checkPoint.oldestXid,
8400 : checkPoint.oldestXidDB);
8401 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8402 676 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8403 676 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8404 676 : LWLockRelease(ControlFileLock);
8405 :
8406 : /* Update shared-memory copy of checkpoint XID/epoch */
8407 676 : SpinLockAcquire(&XLogCtl->info_lck);
8408 676 : XLogCtl->ckptFullXid = checkPoint.nextXid;
8409 676 : SpinLockRelease(&XLogCtl->info_lck);
8410 :
8411 : /* TLI should not change in an on-line checkpoint */
8412 676 : (void) GetCurrentReplayRecPtr(&replayTLI);
8413 676 : if (checkPoint.ThisTimeLineID != replayTLI)
8414 0 : ereport(PANIC,
8415 : (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
8416 : checkPoint.ThisTimeLineID, replayTLI)));
8417 :
8418 676 : RecoveryRestartPoint(&checkPoint, record);
8419 : }
8420 78200 : else if (info == XLOG_OVERWRITE_CONTRECORD)
8421 : {
8422 : /* nothing to do here, handled in xlogrecovery_redo() */
8423 : }
8424 78198 : else if (info == XLOG_END_OF_RECOVERY)
8425 : {
8426 : xl_end_of_recovery xlrec;
8427 : TimeLineID replayTLI;
8428 :
8429 20 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
8430 :
8431 : /*
8432 : * For Hot Standby, we could treat this like a Shutdown Checkpoint,
8433 : * but this case is rarer and harder to test, so the benefit doesn't
8434 : * outweigh the potential extra cost of maintenance.
8435 : */
8436 :
8437 : /*
8438 : * We should've already switched to the new TLI before replaying this
8439 : * record.
8440 : */
8441 20 : (void) GetCurrentReplayRecPtr(&replayTLI);
8442 20 : if (xlrec.ThisTimeLineID != replayTLI)
8443 0 : ereport(PANIC,
8444 : (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
8445 : xlrec.ThisTimeLineID, replayTLI)));
8446 : }
8447 78178 : else if (info == XLOG_NOOP)
8448 : {
8449 : /* nothing to do here */
8450 : }
8451 78178 : else if (info == XLOG_SWITCH)
8452 : {
8453 : /* nothing to do here */
8454 : }
8455 77942 : else if (info == XLOG_RESTORE_POINT)
8456 : {
8457 : /* nothing to do here, handled in xlogrecovery.c */
8458 : }
8459 77932 : else if (info == XLOG_FPI || info == XLOG_FPI_FOR_HINT)
8460 : {
8461 : /*
8462 : * XLOG_FPI records contain nothing else but one or more block
8463 : * references. Every block reference must include a full-page image
8464 : * even if full_page_writes was disabled when the record was generated
8465 : * - otherwise there would be no point in this record.
8466 : *
8467 : * XLOG_FPI_FOR_HINT records are generated when a page needs to be
8468 : * WAL-logged because of a hint bit update. They are only generated
8469 : * when checksums and/or wal_log_hints are enabled. They may include
8470 : * no full-page images if full_page_writes was disabled when they were
8471 : * generated. In this case there is nothing to do here.
8472 : *
8473 : * No recovery conflicts are generated by these generic records - if a
8474 : * resource manager needs to generate conflicts, it has to define a
8475 : * separate WAL record type and redo routine.
8476 : */
8477 163636 : for (uint8 block_id = 0; block_id <= XLogRecMaxBlockId(record); block_id++)
8478 : {
8479 : Buffer buffer;
8480 :
8481 86604 : if (!XLogRecHasBlockImage(record, block_id))
8482 : {
8483 134 : if (info == XLOG_FPI)
8484 0 : elog(ERROR, "XLOG_FPI record did not contain a full-page image");
8485 134 : continue;
8486 : }
8487 :
8488 86470 : if (XLogReadBufferForRedo(record, block_id, &buffer) != BLK_RESTORED)
8489 0 : elog(ERROR, "unexpected XLogReadBufferForRedo result when restoring backup block");
8490 86470 : UnlockReleaseBuffer(buffer);
8491 : }
8492 : }
8493 900 : else if (info == XLOG_BACKUP_END)
8494 : {
8495 : /* nothing to do here, handled in xlogrecovery_redo() */
8496 : }
8497 738 : else if (info == XLOG_PARAMETER_CHANGE)
8498 : {
8499 : xl_parameter_change xlrec;
8500 :
8501 : /* Update our copy of the parameters in pg_control */
8502 60 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
8503 :
8504 : /*
8505 : * Invalidate logical slots if we are in hot standby and the primary
8506 : * does not have a WAL level sufficient for logical decoding. No need
8507 : * to search for potentially conflicting logically slots if standby is
8508 : * running with wal_level lower than logical, because in that case, we
8509 : * would have either disallowed creation of logical slots or
8510 : * invalidated existing ones.
8511 : */
8512 60 : if (InRecovery && InHotStandby &&
8513 30 : xlrec.wal_level < WAL_LEVEL_LOGICAL &&
8514 12 : wal_level >= WAL_LEVEL_LOGICAL)
8515 8 : InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_LEVEL,
8516 : 0, InvalidOid,
8517 : InvalidTransactionId);
8518 :
8519 60 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8520 60 : ControlFile->MaxConnections = xlrec.MaxConnections;
8521 60 : ControlFile->max_worker_processes = xlrec.max_worker_processes;
8522 60 : ControlFile->max_wal_senders = xlrec.max_wal_senders;
8523 60 : ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
8524 60 : ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
8525 60 : ControlFile->wal_level = xlrec.wal_level;
8526 60 : ControlFile->wal_log_hints = xlrec.wal_log_hints;
8527 :
8528 : /*
8529 : * Update minRecoveryPoint to ensure that if recovery is aborted, we
8530 : * recover back up to this point before allowing hot standby again.
8531 : * This is important if the max_* settings are decreased, to ensure
8532 : * you don't run queries against the WAL preceding the change. The
8533 : * local copies cannot be updated as long as crash recovery is
8534 : * happening and we expect all the WAL to be replayed.
8535 : */
8536 60 : if (InArchiveRecovery)
8537 : {
8538 32 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
8539 32 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
8540 : }
8541 60 : if (LocalMinRecoveryPoint != InvalidXLogRecPtr && LocalMinRecoveryPoint < lsn)
8542 : {
8543 : TimeLineID replayTLI;
8544 :
8545 14 : (void) GetCurrentReplayRecPtr(&replayTLI);
8546 14 : ControlFile->minRecoveryPoint = lsn;
8547 14 : ControlFile->minRecoveryPointTLI = replayTLI;
8548 : }
8549 :
8550 60 : CommitTsParameterChange(xlrec.track_commit_timestamp,
8551 60 : ControlFile->track_commit_timestamp);
8552 60 : ControlFile->track_commit_timestamp = xlrec.track_commit_timestamp;
8553 :
8554 60 : UpdateControlFile();
8555 60 : LWLockRelease(ControlFileLock);
8556 :
8557 : /* Check to see if any parameter change gives a problem on recovery */
8558 60 : CheckRequiredParameterValues();
8559 : }
8560 678 : else if (info == XLOG_FPW_CHANGE)
8561 : {
8562 : bool fpw;
8563 :
8564 0 : memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
8565 :
8566 : /*
8567 : * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
8568 : * do_pg_backup_start() and do_pg_backup_stop() can check whether
8569 : * full_page_writes has been disabled during online backup.
8570 : */
8571 0 : if (!fpw)
8572 : {
8573 0 : SpinLockAcquire(&XLogCtl->info_lck);
8574 0 : if (XLogCtl->lastFpwDisableRecPtr < record->ReadRecPtr)
8575 0 : XLogCtl->lastFpwDisableRecPtr = record->ReadRecPtr;
8576 0 : SpinLockRelease(&XLogCtl->info_lck);
8577 : }
8578 :
8579 : /* Keep track of full_page_writes */
8580 0 : lastFullPageWrites = fpw;
8581 : }
8582 : else if (info == XLOG_CHECKPOINT_REDO)
8583 : {
8584 : /* nothing to do here, just for informational purposes */
8585 : }
8586 79118 : }
8587 :
8588 : /*
8589 : * Return the extra open flags used for opening a file, depending on the
8590 : * value of the GUCs wal_sync_method, fsync and debug_io_direct.
8591 : */
8592 : static int
8593 26592 : get_sync_bit(int method)
8594 : {
8595 26592 : int o_direct_flag = 0;
8596 :
8597 : /*
8598 : * Use O_DIRECT if requested, except in walreceiver process. The WAL
8599 : * written by walreceiver is normally read by the startup process soon
8600 : * after it's written. Also, walreceiver performs unaligned writes, which
8601 : * don't work with O_DIRECT, so it is required for correctness too.
8602 : */
8603 26592 : if ((io_direct_flags & IO_DIRECT_WAL) && !AmWalReceiverProcess())
8604 16 : o_direct_flag = PG_O_DIRECT;
8605 :
8606 : /* If fsync is disabled, never open in sync mode */
8607 26592 : if (!enableFsync)
8608 26592 : return o_direct_flag;
8609 :
8610 0 : switch (method)
8611 : {
8612 : /*
8613 : * enum values for all sync options are defined even if they are
8614 : * not supported on the current platform. But if not, they are
8615 : * not included in the enum option array, and therefore will never
8616 : * be seen here.
8617 : */
8618 0 : case WAL_SYNC_METHOD_FSYNC:
8619 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8620 : case WAL_SYNC_METHOD_FDATASYNC:
8621 0 : return o_direct_flag;
8622 : #ifdef O_SYNC
8623 0 : case WAL_SYNC_METHOD_OPEN:
8624 0 : return O_SYNC | o_direct_flag;
8625 : #endif
8626 : #ifdef O_DSYNC
8627 0 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8628 0 : return O_DSYNC | o_direct_flag;
8629 : #endif
8630 0 : default:
8631 : /* can't happen (unless we are out of sync with option array) */
8632 0 : elog(ERROR, "unrecognized \"wal_sync_method\": %d", method);
8633 : return 0; /* silence warning */
8634 : }
8635 : }
8636 :
8637 : /*
8638 : * GUC support
8639 : */
8640 : void
8641 1972 : assign_wal_sync_method(int new_wal_sync_method, void *extra)
8642 : {
8643 1972 : if (wal_sync_method != new_wal_sync_method)
8644 : {
8645 : /*
8646 : * To ensure that no blocks escape unsynced, force an fsync on the
8647 : * currently open log segment (if any). Also, if the open flag is
8648 : * changing, close the log file so it will be reopened (with new flag
8649 : * bit) at next use.
8650 : */
8651 0 : if (openLogFile >= 0)
8652 : {
8653 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN);
8654 0 : if (pg_fsync(openLogFile) != 0)
8655 : {
8656 : char xlogfname[MAXFNAMELEN];
8657 : int save_errno;
8658 :
8659 0 : save_errno = errno;
8660 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo,
8661 : wal_segment_size);
8662 0 : errno = save_errno;
8663 0 : ereport(PANIC,
8664 : (errcode_for_file_access(),
8665 : errmsg("could not fsync file \"%s\": %m", xlogfname)));
8666 : }
8667 :
8668 0 : pgstat_report_wait_end();
8669 0 : if (get_sync_bit(wal_sync_method) != get_sync_bit(new_wal_sync_method))
8670 0 : XLogFileClose();
8671 : }
8672 : }
8673 1972 : }
8674 :
8675 :
8676 : /*
8677 : * Issue appropriate kind of fsync (if any) for an XLOG output file.
8678 : *
8679 : * 'fd' is a file descriptor for the XLOG file to be fsync'd.
8680 : * 'segno' is for error reporting purposes.
8681 : */
8682 : void
8683 257070 : issue_xlog_fsync(int fd, XLogSegNo segno, TimeLineID tli)
8684 : {
8685 257070 : char *msg = NULL;
8686 : instr_time start;
8687 :
8688 : Assert(tli != 0);
8689 :
8690 : /*
8691 : * Quick exit if fsync is disabled or write() has already synced the WAL
8692 : * file.
8693 : */
8694 257070 : if (!enableFsync ||
8695 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN ||
8696 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN_DSYNC)
8697 257070 : return;
8698 :
8699 : /* Measure I/O timing to sync the WAL file */
8700 0 : if (track_wal_io_timing)
8701 0 : INSTR_TIME_SET_CURRENT(start);
8702 : else
8703 0 : INSTR_TIME_SET_ZERO(start);
8704 :
8705 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC);
8706 0 : switch (wal_sync_method)
8707 : {
8708 0 : case WAL_SYNC_METHOD_FSYNC:
8709 0 : if (pg_fsync_no_writethrough(fd) != 0)
8710 0 : msg = _("could not fsync file \"%s\": %m");
8711 0 : break;
8712 : #ifdef HAVE_FSYNC_WRITETHROUGH
8713 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8714 : if (pg_fsync_writethrough(fd) != 0)
8715 : msg = _("could not fsync write-through file \"%s\": %m");
8716 : break;
8717 : #endif
8718 0 : case WAL_SYNC_METHOD_FDATASYNC:
8719 0 : if (pg_fdatasync(fd) != 0)
8720 0 : msg = _("could not fdatasync file \"%s\": %m");
8721 0 : break;
8722 0 : case WAL_SYNC_METHOD_OPEN:
8723 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8724 : /* not reachable */
8725 : Assert(false);
8726 0 : break;
8727 0 : default:
8728 0 : ereport(PANIC,
8729 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8730 : errmsg_internal("unrecognized \"wal_sync_method\": %d", wal_sync_method));
8731 : break;
8732 : }
8733 :
8734 : /* PANIC if failed to fsync */
8735 0 : if (msg)
8736 : {
8737 : char xlogfname[MAXFNAMELEN];
8738 0 : int save_errno = errno;
8739 :
8740 0 : XLogFileName(xlogfname, tli, segno, wal_segment_size);
8741 0 : errno = save_errno;
8742 0 : ereport(PANIC,
8743 : (errcode_for_file_access(),
8744 : errmsg(msg, xlogfname)));
8745 : }
8746 :
8747 0 : pgstat_report_wait_end();
8748 :
8749 : /*
8750 : * Increment the I/O timing and the number of times WAL files were synced.
8751 : */
8752 0 : if (track_wal_io_timing)
8753 : {
8754 : instr_time end;
8755 :
8756 0 : INSTR_TIME_SET_CURRENT(end);
8757 0 : INSTR_TIME_ACCUM_DIFF(PendingWalStats.wal_sync_time, end, start);
8758 : }
8759 :
8760 0 : PendingWalStats.wal_sync++;
8761 : }
8762 :
8763 : /*
8764 : * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
8765 : * function. It creates the necessary starting checkpoint and constructs the
8766 : * backup state and tablespace map.
8767 : *
8768 : * Input parameters are "state" (the backup state), "fast" (if true, we do
8769 : * the checkpoint in immediate mode to make it faster), and "tablespaces"
8770 : * (if non-NULL, indicates a list of tablespaceinfo structs describing the
8771 : * cluster's tablespaces.).
8772 : *
8773 : * The tablespace map contents are appended to passed-in parameter
8774 : * tablespace_map and the caller is responsible for including it in the backup
8775 : * archive as 'tablespace_map'. The tablespace_map file is required mainly for
8776 : * tar format in windows as native windows utilities are not able to create
8777 : * symlinks while extracting files from tar. However for consistency and
8778 : * platform-independence, we do it the same way everywhere.
8779 : *
8780 : * It fills in "state" with the information required for the backup, such
8781 : * as the minimum WAL location that must be present to restore from this
8782 : * backup (starttli) and the corresponding timeline ID (starttli).
8783 : *
8784 : * Every successfully started backup must be stopped by calling
8785 : * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
8786 : * backups active at the same time.
8787 : *
8788 : * It is the responsibility of the caller of this function to verify the
8789 : * permissions of the calling user!
8790 : */
8791 : void
8792 320 : do_pg_backup_start(const char *backupidstr, bool fast, List **tablespaces,
8793 : BackupState *state, StringInfo tblspcmapfile)
8794 : {
8795 : bool backup_started_in_recovery;
8796 :
8797 : Assert(state != NULL);
8798 320 : backup_started_in_recovery = RecoveryInProgress();
8799 :
8800 : /*
8801 : * During recovery, we don't need to check WAL level. Because, if WAL
8802 : * level is not sufficient, it's impossible to get here during recovery.
8803 : */
8804 320 : if (!backup_started_in_recovery && !XLogIsNeeded())
8805 0 : ereport(ERROR,
8806 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8807 : errmsg("WAL level not sufficient for making an online backup"),
8808 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
8809 :
8810 320 : if (strlen(backupidstr) > MAXPGPATH)
8811 2 : ereport(ERROR,
8812 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8813 : errmsg("backup label too long (max %d bytes)",
8814 : MAXPGPATH)));
8815 :
8816 318 : strlcpy(state->name, backupidstr, sizeof(state->name));
8817 :
8818 : /*
8819 : * Mark backup active in shared memory. We must do full-page WAL writes
8820 : * during an on-line backup even if not doing so at other times, because
8821 : * it's quite possible for the backup dump to obtain a "torn" (partially
8822 : * written) copy of a database page if it reads the page concurrently with
8823 : * our write to the same page. This can be fixed as long as the first
8824 : * write to the page in the WAL sequence is a full-page write. Hence, we
8825 : * increment runningBackups then force a CHECKPOINT, to ensure there are
8826 : * no dirty pages in shared memory that might get dumped while the backup
8827 : * is in progress without having a corresponding WAL record. (Once the
8828 : * backup is complete, we need not force full-page writes anymore, since
8829 : * we expect that any pages not modified during the backup interval must
8830 : * have been correctly captured by the backup.)
8831 : *
8832 : * Note that forcing full-page writes has no effect during an online
8833 : * backup from the standby.
8834 : *
8835 : * We must hold all the insertion locks to change the value of
8836 : * runningBackups, to ensure adequate interlocking against
8837 : * XLogInsertRecord().
8838 : */
8839 318 : WALInsertLockAcquireExclusive();
8840 318 : XLogCtl->Insert.runningBackups++;
8841 318 : WALInsertLockRelease();
8842 :
8843 : /*
8844 : * Ensure we decrement runningBackups if we fail below. NB -- for this to
8845 : * work correctly, it is critical that sessionBackupState is only updated
8846 : * after this block is over.
8847 : */
8848 318 : PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, DatumGetBool(true));
8849 : {
8850 318 : bool gotUniqueStartpoint = false;
8851 : DIR *tblspcdir;
8852 : struct dirent *de;
8853 : tablespaceinfo *ti;
8854 : int datadirpathlen;
8855 :
8856 : /*
8857 : * Force an XLOG file switch before the checkpoint, to ensure that the
8858 : * WAL segment the checkpoint is written to doesn't contain pages with
8859 : * old timeline IDs. That would otherwise happen if you called
8860 : * pg_backup_start() right after restoring from a PITR archive: the
8861 : * first WAL segment containing the startup checkpoint has pages in
8862 : * the beginning with the old timeline ID. That can cause trouble at
8863 : * recovery: we won't have a history file covering the old timeline if
8864 : * pg_wal directory was not included in the base backup and the WAL
8865 : * archive was cleared too before starting the backup.
8866 : *
8867 : * This also ensures that we have emitted a WAL page header that has
8868 : * XLP_BKP_REMOVABLE off before we emit the checkpoint record.
8869 : * Therefore, if a WAL archiver (such as pglesslog) is trying to
8870 : * compress out removable backup blocks, it won't remove any that
8871 : * occur after this point.
8872 : *
8873 : * During recovery, we skip forcing XLOG file switch, which means that
8874 : * the backup taken during recovery is not available for the special
8875 : * recovery case described above.
8876 : */
8877 318 : if (!backup_started_in_recovery)
8878 304 : RequestXLogSwitch(false);
8879 :
8880 : do
8881 : {
8882 : bool checkpointfpw;
8883 :
8884 : /*
8885 : * Force a CHECKPOINT. Aside from being necessary to prevent torn
8886 : * page problems, this guarantees that two successive backup runs
8887 : * will have different checkpoint positions and hence different
8888 : * history file names, even if nothing happened in between.
8889 : *
8890 : * During recovery, establish a restartpoint if possible. We use
8891 : * the last restartpoint as the backup starting checkpoint. This
8892 : * means that two successive backup runs can have same checkpoint
8893 : * positions.
8894 : *
8895 : * Since the fact that we are executing do_pg_backup_start()
8896 : * during recovery means that checkpointer is running, we can use
8897 : * RequestCheckpoint() to establish a restartpoint.
8898 : *
8899 : * We use CHECKPOINT_IMMEDIATE only if requested by user (via
8900 : * passing fast = true). Otherwise this can take awhile.
8901 : */
8902 318 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
8903 : (fast ? CHECKPOINT_IMMEDIATE : 0));
8904 :
8905 : /*
8906 : * Now we need to fetch the checkpoint record location, and also
8907 : * its REDO pointer. The oldest point in WAL that would be needed
8908 : * to restore starting from the checkpoint is precisely the REDO
8909 : * pointer.
8910 : */
8911 318 : LWLockAcquire(ControlFileLock, LW_SHARED);
8912 318 : state->checkpointloc = ControlFile->checkPoint;
8913 318 : state->startpoint = ControlFile->checkPointCopy.redo;
8914 318 : state->starttli = ControlFile->checkPointCopy.ThisTimeLineID;
8915 318 : checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
8916 318 : LWLockRelease(ControlFileLock);
8917 :
8918 318 : if (backup_started_in_recovery)
8919 : {
8920 : XLogRecPtr recptr;
8921 :
8922 : /*
8923 : * Check to see if all WAL replayed during online backup
8924 : * (i.e., since last restartpoint used as backup starting
8925 : * checkpoint) contain full-page writes.
8926 : */
8927 14 : SpinLockAcquire(&XLogCtl->info_lck);
8928 14 : recptr = XLogCtl->lastFpwDisableRecPtr;
8929 14 : SpinLockRelease(&XLogCtl->info_lck);
8930 :
8931 14 : if (!checkpointfpw || state->startpoint <= recptr)
8932 0 : ereport(ERROR,
8933 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8934 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
8935 : "since last restartpoint"),
8936 : errhint("This means that the backup being taken on the standby "
8937 : "is corrupt and should not be used. "
8938 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
8939 : "and then try an online backup again.")));
8940 :
8941 : /*
8942 : * During recovery, since we don't use the end-of-backup WAL
8943 : * record and don't write the backup history file, the
8944 : * starting WAL location doesn't need to be unique. This means
8945 : * that two base backups started at the same time might use
8946 : * the same checkpoint as starting locations.
8947 : */
8948 14 : gotUniqueStartpoint = true;
8949 : }
8950 :
8951 : /*
8952 : * If two base backups are started at the same time (in WAL sender
8953 : * processes), we need to make sure that they use different
8954 : * checkpoints as starting locations, because we use the starting
8955 : * WAL location as a unique identifier for the base backup in the
8956 : * end-of-backup WAL record and when we write the backup history
8957 : * file. Perhaps it would be better generate a separate unique ID
8958 : * for each backup instead of forcing another checkpoint, but
8959 : * taking a checkpoint right after another is not that expensive
8960 : * either because only few buffers have been dirtied yet.
8961 : */
8962 318 : WALInsertLockAcquireExclusive();
8963 318 : if (XLogCtl->Insert.lastBackupStart < state->startpoint)
8964 : {
8965 318 : XLogCtl->Insert.lastBackupStart = state->startpoint;
8966 318 : gotUniqueStartpoint = true;
8967 : }
8968 318 : WALInsertLockRelease();
8969 318 : } while (!gotUniqueStartpoint);
8970 :
8971 : /*
8972 : * Construct tablespace_map file.
8973 : */
8974 318 : datadirpathlen = strlen(DataDir);
8975 :
8976 : /* Collect information about all tablespaces */
8977 318 : tblspcdir = AllocateDir(PG_TBLSPC_DIR);
8978 1026 : while ((de = ReadDir(tblspcdir, PG_TBLSPC_DIR)) != NULL)
8979 : {
8980 : char fullpath[MAXPGPATH + sizeof(PG_TBLSPC_DIR)];
8981 : char linkpath[MAXPGPATH];
8982 708 : char *relpath = NULL;
8983 : char *s;
8984 : PGFileType de_type;
8985 : char *badp;
8986 : Oid tsoid;
8987 :
8988 : /*
8989 : * Try to parse the directory name as an unsigned integer.
8990 : *
8991 : * Tablespace directories should be positive integers that can be
8992 : * represented in 32 bits, with no leading zeroes or trailing
8993 : * garbage. If we come across a name that doesn't meet those
8994 : * criteria, skip it.
8995 : */
8996 708 : if (de->d_name[0] < '1' || de->d_name[1] > '9')
8997 |