Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * xlog.c
4 : * PostgreSQL write-ahead log manager
5 : *
6 : * The Write-Ahead Log (WAL) functionality is split into several source
7 : * files, in addition to this one:
8 : *
9 : * xloginsert.c - Functions for constructing WAL records
10 : * xlogrecovery.c - WAL recovery and standby code
11 : * xlogreader.c - Facility for reading WAL files and parsing WAL records
12 : * xlogutils.c - Helper functions for WAL redo routines
13 : *
14 : * This file contains functions for coordinating database startup and
15 : * checkpointing, and managing the write-ahead log buffers when the
16 : * system is running.
17 : *
18 : * StartupXLOG() is the main entry point of the startup process. It
19 : * coordinates database startup, performing WAL recovery, and the
20 : * transition from WAL recovery into normal operations.
21 : *
22 : * XLogInsertRecord() inserts a WAL record into the WAL buffers. Most
23 : * callers should not call this directly, but use the functions in
24 : * xloginsert.c to construct the WAL record. XLogFlush() can be used
25 : * to force the WAL to disk.
26 : *
27 : * In addition to those, there are many other functions for interrogating
28 : * the current system state, and for starting/stopping backups.
29 : *
30 : *
31 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
32 : * Portions Copyright (c) 1994, Regents of the University of California
33 : *
34 : * src/backend/access/transam/xlog.c
35 : *
36 : *-------------------------------------------------------------------------
37 : */
38 :
39 : #include "postgres.h"
40 :
41 : #include <ctype.h>
42 : #include <math.h>
43 : #include <time.h>
44 : #include <fcntl.h>
45 : #include <sys/stat.h>
46 : #include <sys/time.h>
47 : #include <unistd.h>
48 :
49 : #include "access/clog.h"
50 : #include "access/commit_ts.h"
51 : #include "access/heaptoast.h"
52 : #include "access/multixact.h"
53 : #include "access/rewriteheap.h"
54 : #include "access/subtrans.h"
55 : #include "access/timeline.h"
56 : #include "access/transam.h"
57 : #include "access/twophase.h"
58 : #include "access/xact.h"
59 : #include "access/xlog_internal.h"
60 : #include "access/xlogarchive.h"
61 : #include "access/xloginsert.h"
62 : #include "access/xlogreader.h"
63 : #include "access/xlogrecovery.h"
64 : #include "access/xlogutils.h"
65 : #include "access/xlogwait.h"
66 : #include "backup/basebackup.h"
67 : #include "catalog/catversion.h"
68 : #include "catalog/pg_control.h"
69 : #include "catalog/pg_database.h"
70 : #include "common/controldata_utils.h"
71 : #include "common/file_utils.h"
72 : #include "executor/instrument.h"
73 : #include "miscadmin.h"
74 : #include "pg_trace.h"
75 : #include "pgstat.h"
76 : #include "port/atomics.h"
77 : #include "postmaster/bgwriter.h"
78 : #include "postmaster/startup.h"
79 : #include "postmaster/walsummarizer.h"
80 : #include "postmaster/walwriter.h"
81 : #include "replication/origin.h"
82 : #include "replication/slot.h"
83 : #include "replication/slotsync.h"
84 : #include "replication/snapbuild.h"
85 : #include "replication/walreceiver.h"
86 : #include "replication/walsender.h"
87 : #include "storage/bufmgr.h"
88 : #include "storage/fd.h"
89 : #include "storage/ipc.h"
90 : #include "storage/large_object.h"
91 : #include "storage/latch.h"
92 : #include "storage/predicate.h"
93 : #include "storage/proc.h"
94 : #include "storage/procarray.h"
95 : #include "storage/reinit.h"
96 : #include "storage/spin.h"
97 : #include "storage/sync.h"
98 : #include "utils/guc_hooks.h"
99 : #include "utils/guc_tables.h"
100 : #include "utils/injection_point.h"
101 : #include "utils/pgstat_internal.h"
102 : #include "utils/ps_status.h"
103 : #include "utils/relmapper.h"
104 : #include "utils/snapmgr.h"
105 : #include "utils/timeout.h"
106 : #include "utils/timestamp.h"
107 : #include "utils/varlena.h"
108 :
109 : #ifdef WAL_DEBUG
110 : #include "utils/memutils.h"
111 : #endif
112 :
113 : /* timeline ID to be used when bootstrapping */
114 : #define BootstrapTimeLineID 1
115 :
116 : /* User-settable parameters */
117 : int max_wal_size_mb = 1024; /* 1 GB */
118 : int min_wal_size_mb = 80; /* 80 MB */
119 : int wal_keep_size_mb = 0;
120 : int XLOGbuffers = -1;
121 : int XLogArchiveTimeout = 0;
122 : int XLogArchiveMode = ARCHIVE_MODE_OFF;
123 : char *XLogArchiveCommand = NULL;
124 : bool EnableHotStandby = false;
125 : bool fullPageWrites = true;
126 : bool wal_log_hints = false;
127 : int wal_compression = WAL_COMPRESSION_NONE;
128 : char *wal_consistency_checking_string = NULL;
129 : bool *wal_consistency_checking = NULL;
130 : bool wal_init_zero = true;
131 : bool wal_recycle = true;
132 : bool log_checkpoints = true;
133 : int wal_sync_method = DEFAULT_WAL_SYNC_METHOD;
134 : int wal_level = WAL_LEVEL_REPLICA;
135 : int CommitDelay = 0; /* precommit delay in microseconds */
136 : int CommitSiblings = 5; /* # concurrent xacts needed to sleep */
137 : int wal_retrieve_retry_interval = 5000;
138 : int max_slot_wal_keep_size_mb = -1;
139 : int wal_decode_buffer_size = 512 * 1024;
140 : bool track_wal_io_timing = false;
141 :
142 : #ifdef WAL_DEBUG
143 : bool XLOG_DEBUG = false;
144 : #endif
145 :
146 : int wal_segment_size = DEFAULT_XLOG_SEG_SIZE;
147 :
148 : /*
149 : * Number of WAL insertion locks to use. A higher value allows more insertions
150 : * to happen concurrently, but adds some CPU overhead to flushing the WAL,
151 : * which needs to iterate all the locks.
152 : */
153 : #define NUM_XLOGINSERT_LOCKS 8
154 :
155 : /*
156 : * Max distance from last checkpoint, before triggering a new xlog-based
157 : * checkpoint.
158 : */
159 : int CheckPointSegments;
160 :
161 : /* Estimated distance between checkpoints, in bytes */
162 : static double CheckPointDistanceEstimate = 0;
163 : static double PrevCheckPointDistance = 0;
164 :
165 : /*
166 : * Track whether there were any deferred checks for custom resource managers
167 : * specified in wal_consistency_checking.
168 : */
169 : static bool check_wal_consistency_checking_deferred = false;
170 :
171 : /*
172 : * GUC support
173 : */
174 : const struct config_enum_entry wal_sync_method_options[] = {
175 : {"fsync", WAL_SYNC_METHOD_FSYNC, false},
176 : #ifdef HAVE_FSYNC_WRITETHROUGH
177 : {"fsync_writethrough", WAL_SYNC_METHOD_FSYNC_WRITETHROUGH, false},
178 : #endif
179 : {"fdatasync", WAL_SYNC_METHOD_FDATASYNC, false},
180 : #ifdef O_SYNC
181 : {"open_sync", WAL_SYNC_METHOD_OPEN, false},
182 : #endif
183 : #ifdef O_DSYNC
184 : {"open_datasync", WAL_SYNC_METHOD_OPEN_DSYNC, false},
185 : #endif
186 : {NULL, 0, false}
187 : };
188 :
189 :
190 : /*
191 : * Although only "on", "off", and "always" are documented,
192 : * we accept all the likely variants of "on" and "off".
193 : */
194 : const struct config_enum_entry archive_mode_options[] = {
195 : {"always", ARCHIVE_MODE_ALWAYS, false},
196 : {"on", ARCHIVE_MODE_ON, false},
197 : {"off", ARCHIVE_MODE_OFF, false},
198 : {"true", ARCHIVE_MODE_ON, true},
199 : {"false", ARCHIVE_MODE_OFF, true},
200 : {"yes", ARCHIVE_MODE_ON, true},
201 : {"no", ARCHIVE_MODE_OFF, true},
202 : {"1", ARCHIVE_MODE_ON, true},
203 : {"0", ARCHIVE_MODE_OFF, true},
204 : {NULL, 0, false}
205 : };
206 :
207 : /*
208 : * Statistics for current checkpoint are collected in this global struct.
209 : * Because only the checkpointer or a stand-alone backend can perform
210 : * checkpoints, this will be unused in normal backends.
211 : */
212 : CheckpointStatsData CheckpointStats;
213 :
214 : /*
215 : * During recovery, lastFullPageWrites keeps track of full_page_writes that
216 : * the replayed WAL records indicate. It's initialized with full_page_writes
217 : * that the recovery starting checkpoint record indicates, and then updated
218 : * each time XLOG_FPW_CHANGE record is replayed.
219 : */
220 : static bool lastFullPageWrites;
221 :
222 : /*
223 : * Local copy of the state tracked by SharedRecoveryState in shared memory,
224 : * It is false if SharedRecoveryState is RECOVERY_STATE_DONE. True actually
225 : * means "not known, need to check the shared state".
226 : */
227 : static bool LocalRecoveryInProgress = true;
228 :
229 : /*
230 : * Local state for XLogInsertAllowed():
231 : * 1: unconditionally allowed to insert XLOG
232 : * 0: unconditionally not allowed to insert XLOG
233 : * -1: must check RecoveryInProgress(); disallow until it is false
234 : * Most processes start with -1 and transition to 1 after seeing that recovery
235 : * is not in progress. But we can also force the value for special cases.
236 : * The coding in XLogInsertAllowed() depends on the first two of these states
237 : * being numerically the same as bool true and false.
238 : */
239 : static int LocalXLogInsertAllowed = -1;
240 :
241 : /*
242 : * ProcLastRecPtr points to the start of the last XLOG record inserted by the
243 : * current backend. It is updated for all inserts. XactLastRecEnd points to
244 : * end+1 of the last record, and is reset when we end a top-level transaction,
245 : * or start a new one; so it can be used to tell if the current transaction has
246 : * created any XLOG records.
247 : *
248 : * While in parallel mode, this may not be fully up to date. When committing,
249 : * a transaction can assume this covers all xlog records written either by the
250 : * user backend or by any parallel worker which was present at any point during
251 : * the transaction. But when aborting, or when still in parallel mode, other
252 : * parallel backends may have written WAL records at later LSNs than the value
253 : * stored here. The parallel leader advances its own copy, when necessary,
254 : * in WaitForParallelWorkersToFinish.
255 : */
256 : XLogRecPtr ProcLastRecPtr = InvalidXLogRecPtr;
257 : XLogRecPtr XactLastRecEnd = InvalidXLogRecPtr;
258 : XLogRecPtr XactLastCommitEnd = InvalidXLogRecPtr;
259 :
260 : /*
261 : * RedoRecPtr is this backend's local copy of the REDO record pointer
262 : * (which is almost but not quite the same as a pointer to the most recent
263 : * CHECKPOINT record). We update this from the shared-memory copy,
264 : * XLogCtl->Insert.RedoRecPtr, whenever we can safely do so (ie, when we
265 : * hold an insertion lock). See XLogInsertRecord for details. We are also
266 : * allowed to update from XLogCtl->RedoRecPtr if we hold the info_lck;
267 : * see GetRedoRecPtr.
268 : *
269 : * NB: Code that uses this variable must be prepared not only for the
270 : * possibility that it may be arbitrarily out of date, but also for the
271 : * possibility that it might be set to InvalidXLogRecPtr. We used to
272 : * initialize it as a side effect of the first call to RecoveryInProgress(),
273 : * which meant that most code that might use it could assume that it had a
274 : * real if perhaps stale value. That's no longer the case.
275 : */
276 : static XLogRecPtr RedoRecPtr;
277 :
278 : /*
279 : * doPageWrites is this backend's local copy of (fullPageWrites ||
280 : * runningBackups > 0). It is used together with RedoRecPtr to decide whether
281 : * a full-page image of a page need to be taken.
282 : *
283 : * NB: Initially this is false, and there's no guarantee that it will be
284 : * initialized to any other value before it is first used. Any code that
285 : * makes use of it must recheck the value after obtaining a WALInsertLock,
286 : * and respond appropriately if it turns out that the previous value wasn't
287 : * accurate.
288 : */
289 : static bool doPageWrites;
290 :
291 : /*----------
292 : * Shared-memory data structures for XLOG control
293 : *
294 : * LogwrtRqst indicates a byte position that we need to write and/or fsync
295 : * the log up to (all records before that point must be written or fsynced).
296 : * The positions already written/fsynced are maintained in logWriteResult
297 : * and logFlushResult using atomic access.
298 : * In addition to the shared variable, each backend has a private copy of
299 : * both in LogwrtResult, which is updated when convenient.
300 : *
301 : * The request bookkeeping is simpler: there is a shared XLogCtl->LogwrtRqst
302 : * (protected by info_lck), but we don't need to cache any copies of it.
303 : *
304 : * info_lck is only held long enough to read/update the protected variables,
305 : * so it's a plain spinlock. The other locks are held longer (potentially
306 : * over I/O operations), so we use LWLocks for them. These locks are:
307 : *
308 : * WALBufMappingLock: must be held to replace a page in the WAL buffer cache.
309 : * It is only held while initializing and changing the mapping. If the
310 : * contents of the buffer being replaced haven't been written yet, the mapping
311 : * lock is released while the write is done, and reacquired afterwards.
312 : *
313 : * WALWriteLock: must be held to write WAL buffers to disk (XLogWrite or
314 : * XLogFlush).
315 : *
316 : * ControlFileLock: must be held to read/update control file or create
317 : * new log file.
318 : *
319 : *----------
320 : */
321 :
322 : typedef struct XLogwrtRqst
323 : {
324 : XLogRecPtr Write; /* last byte + 1 to write out */
325 : XLogRecPtr Flush; /* last byte + 1 to flush */
326 : } XLogwrtRqst;
327 :
328 : typedef struct XLogwrtResult
329 : {
330 : XLogRecPtr Write; /* last byte + 1 written out */
331 : XLogRecPtr Flush; /* last byte + 1 flushed */
332 : } XLogwrtResult;
333 :
334 : /*
335 : * Inserting to WAL is protected by a small fixed number of WAL insertion
336 : * locks. To insert to the WAL, you must hold one of the locks - it doesn't
337 : * matter which one. To lock out other concurrent insertions, you must hold
338 : * of them. Each WAL insertion lock consists of a lightweight lock, plus an
339 : * indicator of how far the insertion has progressed (insertingAt).
340 : *
341 : * The insertingAt values are read when a process wants to flush WAL from
342 : * the in-memory buffers to disk, to check that all the insertions to the
343 : * region the process is about to write out have finished. You could simply
344 : * wait for all currently in-progress insertions to finish, but the
345 : * insertingAt indicator allows you to ignore insertions to later in the WAL,
346 : * so that you only wait for the insertions that are modifying the buffers
347 : * you're about to write out.
348 : *
349 : * This isn't just an optimization. If all the WAL buffers are dirty, an
350 : * inserter that's holding a WAL insert lock might need to evict an old WAL
351 : * buffer, which requires flushing the WAL. If it's possible for an inserter
352 : * to block on another inserter unnecessarily, deadlock can arise when two
353 : * inserters holding a WAL insert lock wait for each other to finish their
354 : * insertion.
355 : *
356 : * Small WAL records that don't cross a page boundary never update the value,
357 : * the WAL record is just copied to the page and the lock is released. But
358 : * to avoid the deadlock-scenario explained above, the indicator is always
359 : * updated before sleeping while holding an insertion lock.
360 : *
361 : * lastImportantAt contains the LSN of the last important WAL record inserted
362 : * using a given lock. This value is used to detect if there has been
363 : * important WAL activity since the last time some action, like a checkpoint,
364 : * was performed - allowing to not repeat the action if not. The LSN is
365 : * updated for all insertions, unless the XLOG_MARK_UNIMPORTANT flag was
366 : * set. lastImportantAt is never cleared, only overwritten by the LSN of newer
367 : * records. Tracking the WAL activity directly in WALInsertLock has the
368 : * advantage of not needing any additional locks to update the value.
369 : */
370 : typedef struct
371 : {
372 : LWLock lock;
373 : pg_atomic_uint64 insertingAt;
374 : XLogRecPtr lastImportantAt;
375 : } WALInsertLock;
376 :
377 : /*
378 : * All the WAL insertion locks are allocated as an array in shared memory. We
379 : * force the array stride to be a power of 2, which saves a few cycles in
380 : * indexing, but more importantly also ensures that individual slots don't
381 : * cross cache line boundaries. (Of course, we have to also ensure that the
382 : * array start address is suitably aligned.)
383 : */
384 : typedef union WALInsertLockPadded
385 : {
386 : WALInsertLock l;
387 : char pad[PG_CACHE_LINE_SIZE];
388 : } WALInsertLockPadded;
389 :
390 : /*
391 : * Session status of running backup, used for sanity checks in SQL-callable
392 : * functions to start and stop backups.
393 : */
394 : static SessionBackupState sessionBackupState = SESSION_BACKUP_NONE;
395 :
396 : /*
397 : * Shared state data for WAL insertion.
398 : */
399 : typedef struct XLogCtlInsert
400 : {
401 : slock_t insertpos_lck; /* protects CurrBytePos and PrevBytePos */
402 :
403 : /*
404 : * CurrBytePos is the end of reserved WAL. The next record will be
405 : * inserted at that position. PrevBytePos is the start position of the
406 : * previously inserted (or rather, reserved) record - it is copied to the
407 : * prev-link of the next record. These are stored as "usable byte
408 : * positions" rather than XLogRecPtrs (see XLogBytePosToRecPtr()).
409 : */
410 : uint64 CurrBytePos;
411 : uint64 PrevBytePos;
412 :
413 : /*
414 : * Make sure the above heavily-contended spinlock and byte positions are
415 : * on their own cache line. In particular, the RedoRecPtr and full page
416 : * write variables below should be on a different cache line. They are
417 : * read on every WAL insertion, but updated rarely, and we don't want
418 : * those reads to steal the cache line containing Curr/PrevBytePos.
419 : */
420 : char pad[PG_CACHE_LINE_SIZE];
421 :
422 : /*
423 : * fullPageWrites is the authoritative value used by all backends to
424 : * determine whether to write full-page image to WAL. This shared value,
425 : * instead of the process-local fullPageWrites, is required because, when
426 : * full_page_writes is changed by SIGHUP, we must WAL-log it before it
427 : * actually affects WAL-logging by backends. Checkpointer sets at startup
428 : * or after SIGHUP.
429 : *
430 : * To read these fields, you must hold an insertion lock. To modify them,
431 : * you must hold ALL the locks.
432 : */
433 : XLogRecPtr RedoRecPtr; /* current redo point for insertions */
434 : bool fullPageWrites;
435 :
436 : /*
437 : * runningBackups is a counter indicating the number of backups currently
438 : * in progress. lastBackupStart is the latest checkpoint redo location
439 : * used as a starting point for an online backup.
440 : */
441 : int runningBackups;
442 : XLogRecPtr lastBackupStart;
443 :
444 : /*
445 : * WAL insertion locks.
446 : */
447 : WALInsertLockPadded *WALInsertLocks;
448 : } XLogCtlInsert;
449 :
450 : /*
451 : * Total shared-memory state for XLOG.
452 : */
453 : typedef struct XLogCtlData
454 : {
455 : XLogCtlInsert Insert;
456 :
457 : /* Protected by info_lck: */
458 : XLogwrtRqst LogwrtRqst;
459 : XLogRecPtr RedoRecPtr; /* a recent copy of Insert->RedoRecPtr */
460 : XLogRecPtr asyncXactLSN; /* LSN of newest async commit/abort */
461 : XLogRecPtr replicationSlotMinLSN; /* oldest LSN needed by any slot */
462 :
463 : XLogSegNo lastRemovedSegNo; /* latest removed/recycled XLOG segment */
464 :
465 : /* Fake LSN counter, for unlogged relations. */
466 : pg_atomic_uint64 unloggedLSN;
467 :
468 : /* Time and LSN of last xlog segment switch. Protected by WALWriteLock. */
469 : pg_time_t lastSegSwitchTime;
470 : XLogRecPtr lastSegSwitchLSN;
471 :
472 : /* These are accessed using atomics -- info_lck not needed */
473 : pg_atomic_uint64 logInsertResult; /* last byte + 1 inserted to buffers */
474 : pg_atomic_uint64 logWriteResult; /* last byte + 1 written out */
475 : pg_atomic_uint64 logFlushResult; /* last byte + 1 flushed */
476 :
477 : /*
478 : * Latest initialized page in the cache (last byte position + 1).
479 : *
480 : * To change the identity of a buffer (and InitializedUpTo), you need to
481 : * hold WALBufMappingLock. To change the identity of a buffer that's
482 : * still dirty, the old page needs to be written out first, and for that
483 : * you need WALWriteLock, and you need to ensure that there are no
484 : * in-progress insertions to the page by calling
485 : * WaitXLogInsertionsToFinish().
486 : */
487 : XLogRecPtr InitializedUpTo;
488 :
489 : /*
490 : * These values do not change after startup, although the pointed-to pages
491 : * and xlblocks values certainly do. xlblocks values are protected by
492 : * WALBufMappingLock.
493 : */
494 : char *pages; /* buffers for unwritten XLOG pages */
495 : pg_atomic_uint64 *xlblocks; /* 1st byte ptr-s + XLOG_BLCKSZ */
496 : int XLogCacheBlck; /* highest allocated xlog buffer index */
497 :
498 : /*
499 : * InsertTimeLineID is the timeline into which new WAL is being inserted
500 : * and flushed. It is zero during recovery, and does not change once set.
501 : *
502 : * If we create a new timeline when the system was started up,
503 : * PrevTimeLineID is the old timeline's ID that we forked off from.
504 : * Otherwise it's equal to InsertTimeLineID.
505 : *
506 : * We set these fields while holding info_lck. Most that reads these
507 : * values knows that recovery is no longer in progress and so can safely
508 : * read the value without a lock, but code that could be run either during
509 : * or after recovery can take info_lck while reading these values.
510 : */
511 : TimeLineID InsertTimeLineID;
512 : TimeLineID PrevTimeLineID;
513 :
514 : /*
515 : * SharedRecoveryState indicates if we're still in crash or archive
516 : * recovery. Protected by info_lck.
517 : */
518 : RecoveryState SharedRecoveryState;
519 :
520 : /*
521 : * InstallXLogFileSegmentActive indicates whether the checkpointer should
522 : * arrange for future segments by recycling and/or PreallocXlogFiles().
523 : * Protected by ControlFileLock. Only the startup process changes it. If
524 : * true, anyone can use InstallXLogFileSegment(). If false, the startup
525 : * process owns the exclusive right to install segments, by reading from
526 : * the archive and possibly replacing existing files.
527 : */
528 : bool InstallXLogFileSegmentActive;
529 :
530 : /*
531 : * WalWriterSleeping indicates whether the WAL writer is currently in
532 : * low-power mode (and hence should be nudged if an async commit occurs).
533 : * Protected by info_lck.
534 : */
535 : bool WalWriterSleeping;
536 :
537 : /*
538 : * During recovery, we keep a copy of the latest checkpoint record here.
539 : * lastCheckPointRecPtr points to start of checkpoint record and
540 : * lastCheckPointEndPtr points to end+1 of checkpoint record. Used by the
541 : * checkpointer when it wants to create a restartpoint.
542 : *
543 : * Protected by info_lck.
544 : */
545 : XLogRecPtr lastCheckPointRecPtr;
546 : XLogRecPtr lastCheckPointEndPtr;
547 : CheckPoint lastCheckPoint;
548 :
549 : /*
550 : * lastFpwDisableRecPtr points to the start of the last replayed
551 : * XLOG_FPW_CHANGE record that instructs full_page_writes is disabled.
552 : */
553 : XLogRecPtr lastFpwDisableRecPtr;
554 :
555 : slock_t info_lck; /* locks shared variables shown above */
556 : } XLogCtlData;
557 :
558 : /*
559 : * Classification of XLogInsertRecord operations.
560 : */
561 : typedef enum
562 : {
563 : WALINSERT_NORMAL,
564 : WALINSERT_SPECIAL_SWITCH,
565 : WALINSERT_SPECIAL_CHECKPOINT
566 : } WalInsertClass;
567 :
568 : static XLogCtlData *XLogCtl = NULL;
569 :
570 : /* a private copy of XLogCtl->Insert.WALInsertLocks, for convenience */
571 : static WALInsertLockPadded *WALInsertLocks = NULL;
572 :
573 : /*
574 : * We maintain an image of pg_control in shared memory.
575 : */
576 : static ControlFileData *ControlFile = NULL;
577 :
578 : /*
579 : * Calculate the amount of space left on the page after 'endptr'. Beware
580 : * multiple evaluation!
581 : */
582 : #define INSERT_FREESPACE(endptr) \
583 : (((endptr) % XLOG_BLCKSZ == 0) ? 0 : (XLOG_BLCKSZ - (endptr) % XLOG_BLCKSZ))
584 :
585 : /* Macro to advance to next buffer index. */
586 : #define NextBufIdx(idx) \
587 : (((idx) == XLogCtl->XLogCacheBlck) ? 0 : ((idx) + 1))
588 :
589 : /*
590 : * XLogRecPtrToBufIdx returns the index of the WAL buffer that holds, or
591 : * would hold if it was in cache, the page containing 'recptr'.
592 : */
593 : #define XLogRecPtrToBufIdx(recptr) \
594 : (((recptr) / XLOG_BLCKSZ) % (XLogCtl->XLogCacheBlck + 1))
595 :
596 : /*
597 : * These are the number of bytes in a WAL page usable for WAL data.
598 : */
599 : #define UsableBytesInPage (XLOG_BLCKSZ - SizeOfXLogShortPHD)
600 :
601 : /*
602 : * Convert values of GUCs measured in megabytes to equiv. segment count.
603 : * Rounds down.
604 : */
605 : #define ConvertToXSegs(x, segsize) XLogMBVarToSegs((x), (segsize))
606 :
607 : /* The number of bytes in a WAL segment usable for WAL data. */
608 : static int UsableBytesInSegment;
609 :
610 : /*
611 : * Private, possibly out-of-date copy of shared LogwrtResult.
612 : * See discussion above.
613 : */
614 : static XLogwrtResult LogwrtResult = {0, 0};
615 :
616 : /*
617 : * Update local copy of shared XLogCtl->log{Write,Flush}Result
618 : *
619 : * It's critical that Flush always trails Write, so the order of the reads is
620 : * important, as is the barrier. See also XLogWrite.
621 : */
622 : #define RefreshXLogWriteResult(_target) \
623 : do { \
624 : _target.Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult); \
625 : pg_read_barrier(); \
626 : _target.Write = pg_atomic_read_u64(&XLogCtl->logWriteResult); \
627 : } while (0)
628 :
629 : /*
630 : * openLogFile is -1 or a kernel FD for an open log file segment.
631 : * openLogSegNo identifies the segment, and openLogTLI the corresponding TLI.
632 : * These variables are only used to write the XLOG, and so will normally refer
633 : * to the active segment.
634 : *
635 : * Note: call Reserve/ReleaseExternalFD to track consumption of this FD.
636 : */
637 : static int openLogFile = -1;
638 : static XLogSegNo openLogSegNo = 0;
639 : static TimeLineID openLogTLI = 0;
640 :
641 : /*
642 : * Local copies of equivalent fields in the control file. When running
643 : * crash recovery, LocalMinRecoveryPoint is set to InvalidXLogRecPtr as we
644 : * expect to replay all the WAL available, and updateMinRecoveryPoint is
645 : * switched to false to prevent any updates while replaying records.
646 : * Those values are kept consistent as long as crash recovery runs.
647 : */
648 : static XLogRecPtr LocalMinRecoveryPoint;
649 : static TimeLineID LocalMinRecoveryPointTLI;
650 : static bool updateMinRecoveryPoint = true;
651 :
652 : /* For WALInsertLockAcquire/Release functions */
653 : static int MyLockNo = 0;
654 : static bool holdingAllLocks = false;
655 :
656 : #ifdef WAL_DEBUG
657 : static MemoryContext walDebugCxt = NULL;
658 : #endif
659 :
660 : static void CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI,
661 : XLogRecPtr EndOfLog,
662 : TimeLineID newTLI);
663 : static void CheckRequiredParameterValues(void);
664 : static void XLogReportParameters(void);
665 : static int LocalSetXLogInsertAllowed(void);
666 : static void CreateEndOfRecoveryRecord(void);
667 : static XLogRecPtr CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn,
668 : XLogRecPtr pagePtr,
669 : TimeLineID newTLI);
670 : static void CheckPointGuts(XLogRecPtr checkPointRedo, int flags);
671 : static void KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo);
672 :
673 : static void AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli,
674 : bool opportunistic);
675 : static void XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible);
676 : static bool InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
677 : bool find_free, XLogSegNo max_segno,
678 : TimeLineID tli);
679 : static void XLogFileClose(void);
680 : static void PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli);
681 : static void RemoveTempXlogFiles(void);
682 : static void RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr,
683 : XLogRecPtr endptr, TimeLineID insertTLI);
684 : static void RemoveXlogFile(const struct dirent *segment_de,
685 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
686 : TimeLineID insertTLI);
687 : static void UpdateLastRemovedPtr(char *filename);
688 : static void ValidateXLOGDirectoryStructure(void);
689 : static void CleanupBackupHistory(void);
690 : static void UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force);
691 : static bool PerformRecoveryXLogAction(void);
692 : static void InitControlFile(uint64 sysidentifier, uint32 data_checksum_version);
693 : static void WriteControlFile(void);
694 : static void ReadControlFile(void);
695 : static void UpdateControlFile(void);
696 : static char *str_time(pg_time_t tnow, char *buf, size_t bufsize);
697 :
698 : static int get_sync_bit(int method);
699 :
700 : static void CopyXLogRecordToWAL(int write_len, bool isLogSwitch,
701 : XLogRecData *rdata,
702 : XLogRecPtr StartPos, XLogRecPtr EndPos,
703 : TimeLineID tli);
704 : static void ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos,
705 : XLogRecPtr *EndPos, XLogRecPtr *PrevPtr);
706 : static bool ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos,
707 : XLogRecPtr *PrevPtr);
708 : static XLogRecPtr WaitXLogInsertionsToFinish(XLogRecPtr upto);
709 : static char *GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli);
710 : static XLogRecPtr XLogBytePosToRecPtr(uint64 bytepos);
711 : static XLogRecPtr XLogBytePosToEndRecPtr(uint64 bytepos);
712 : static uint64 XLogRecPtrToBytePos(XLogRecPtr ptr);
713 :
714 : static void WALInsertLockAcquire(void);
715 : static void WALInsertLockAcquireExclusive(void);
716 : static void WALInsertLockRelease(void);
717 : static void WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt);
718 :
719 : /*
720 : * Insert an XLOG record represented by an already-constructed chain of data
721 : * chunks. This is a low-level routine; to construct the WAL record header
722 : * and data, use the higher-level routines in xloginsert.c.
723 : *
724 : * If 'fpw_lsn' is valid, it is the oldest LSN among the pages that this
725 : * WAL record applies to, that were not included in the record as full page
726 : * images. If fpw_lsn <= RedoRecPtr, the function does not perform the
727 : * insertion and returns InvalidXLogRecPtr. The caller can then recalculate
728 : * which pages need a full-page image, and retry. If fpw_lsn is invalid, the
729 : * record is always inserted.
730 : *
731 : * 'flags' gives more in-depth control on the record being inserted. See
732 : * XLogSetRecordFlags() for details.
733 : *
734 : * 'topxid_included' tells whether the top-transaction id is logged along with
735 : * current subtransaction. See XLogRecordAssemble().
736 : *
737 : * The first XLogRecData in the chain must be for the record header, and its
738 : * data must be MAXALIGNed. XLogInsertRecord fills in the xl_prev and
739 : * xl_crc fields in the header, the rest of the header must already be filled
740 : * by the caller.
741 : *
742 : * Returns XLOG pointer to end of record (beginning of next record).
743 : * This can be used as LSN for data pages affected by the logged action.
744 : * (LSN is the XLOG point up to which the XLOG must be flushed to disk
745 : * before the data page can be written out. This implements the basic
746 : * WAL rule "write the log before the data".)
747 : */
748 : XLogRecPtr
749 30780054 : XLogInsertRecord(XLogRecData *rdata,
750 : XLogRecPtr fpw_lsn,
751 : uint8 flags,
752 : int num_fpi,
753 : uint64 fpi_bytes,
754 : bool topxid_included)
755 : {
756 30780054 : XLogCtlInsert *Insert = &XLogCtl->Insert;
757 : pg_crc32c rdata_crc;
758 : bool inserted;
759 30780054 : XLogRecord *rechdr = (XLogRecord *) rdata->data;
760 30780054 : uint8 info = rechdr->xl_info & ~XLR_INFO_MASK;
761 30780054 : WalInsertClass class = WALINSERT_NORMAL;
762 : XLogRecPtr StartPos;
763 : XLogRecPtr EndPos;
764 30780054 : bool prevDoPageWrites = doPageWrites;
765 : TimeLineID insertTLI;
766 :
767 : /* Does this record type require special handling? */
768 30780054 : if (unlikely(rechdr->xl_rmid == RM_XLOG_ID))
769 : {
770 465954 : if (info == XLOG_SWITCH)
771 1604 : class = WALINSERT_SPECIAL_SWITCH;
772 464350 : else if (info == XLOG_CHECKPOINT_REDO)
773 1864 : class = WALINSERT_SPECIAL_CHECKPOINT;
774 : }
775 :
776 : /* we assume that all of the record header is in the first chunk */
777 : Assert(rdata->len >= SizeOfXLogRecord);
778 :
779 : /* cross-check on whether we should be here or not */
780 30780054 : if (!XLogInsertAllowed())
781 0 : elog(ERROR, "cannot make new WAL entries during recovery");
782 :
783 : /*
784 : * Given that we're not in recovery, InsertTimeLineID is set and can't
785 : * change, so we can read it without a lock.
786 : */
787 30780054 : insertTLI = XLogCtl->InsertTimeLineID;
788 :
789 : /*----------
790 : *
791 : * We have now done all the preparatory work we can without holding a
792 : * lock or modifying shared state. From here on, inserting the new WAL
793 : * record to the shared WAL buffer cache is a two-step process:
794 : *
795 : * 1. Reserve the right amount of space from the WAL. The current head of
796 : * reserved space is kept in Insert->CurrBytePos, and is protected by
797 : * insertpos_lck.
798 : *
799 : * 2. Copy the record to the reserved WAL space. This involves finding the
800 : * correct WAL buffer containing the reserved space, and copying the
801 : * record in place. This can be done concurrently in multiple processes.
802 : *
803 : * To keep track of which insertions are still in-progress, each concurrent
804 : * inserter acquires an insertion lock. In addition to just indicating that
805 : * an insertion is in progress, the lock tells others how far the inserter
806 : * has progressed. There is a small fixed number of insertion locks,
807 : * determined by NUM_XLOGINSERT_LOCKS. When an inserter crosses a page
808 : * boundary, it updates the value stored in the lock to the how far it has
809 : * inserted, to allow the previous buffer to be flushed.
810 : *
811 : * Holding onto an insertion lock also protects RedoRecPtr and
812 : * fullPageWrites from changing until the insertion is finished.
813 : *
814 : * Step 2 can usually be done completely in parallel. If the required WAL
815 : * page is not initialized yet, you have to grab WALBufMappingLock to
816 : * initialize it, but the WAL writer tries to do that ahead of insertions
817 : * to avoid that from happening in the critical path.
818 : *
819 : *----------
820 : */
821 30780054 : START_CRIT_SECTION();
822 :
823 30780054 : if (likely(class == WALINSERT_NORMAL))
824 : {
825 30776586 : WALInsertLockAcquire();
826 :
827 : /*
828 : * Check to see if my copy of RedoRecPtr is out of date. If so, may
829 : * have to go back and have the caller recompute everything. This can
830 : * only happen just after a checkpoint, so it's better to be slow in
831 : * this case and fast otherwise.
832 : *
833 : * Also check to see if fullPageWrites was just turned on or there's a
834 : * running backup (which forces full-page writes); if we weren't
835 : * already doing full-page writes then go back and recompute.
836 : *
837 : * If we aren't doing full-page writes then RedoRecPtr doesn't
838 : * actually affect the contents of the XLOG record, so we'll update
839 : * our local copy but not force a recomputation. (If doPageWrites was
840 : * just turned off, we could recompute the record without full pages,
841 : * but we choose not to bother.)
842 : */
843 30776586 : if (RedoRecPtr != Insert->RedoRecPtr)
844 : {
845 : Assert(RedoRecPtr < Insert->RedoRecPtr);
846 14588 : RedoRecPtr = Insert->RedoRecPtr;
847 : }
848 30776586 : doPageWrites = (Insert->fullPageWrites || Insert->runningBackups > 0);
849 :
850 30776586 : if (doPageWrites &&
851 30279070 : (!prevDoPageWrites ||
852 27792142 : (XLogRecPtrIsValid(fpw_lsn) && fpw_lsn <= RedoRecPtr)))
853 : {
854 : /*
855 : * Oops, some buffer now needs to be backed up that the caller
856 : * didn't back up. Start over.
857 : */
858 16000 : WALInsertLockRelease();
859 16000 : END_CRIT_SECTION();
860 16000 : return InvalidXLogRecPtr;
861 : }
862 :
863 : /*
864 : * Reserve space for the record in the WAL. This also sets the xl_prev
865 : * pointer.
866 : */
867 30760586 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
868 : &rechdr->xl_prev);
869 :
870 : /* Normal records are always inserted. */
871 30760586 : inserted = true;
872 : }
873 3468 : else if (class == WALINSERT_SPECIAL_SWITCH)
874 : {
875 : /*
876 : * In order to insert an XLOG_SWITCH record, we need to hold all of
877 : * the WAL insertion locks, not just one, so that no one else can
878 : * begin inserting a record until we've figured out how much space
879 : * remains in the current WAL segment and claimed all of it.
880 : *
881 : * Nonetheless, this case is simpler than the normal cases handled
882 : * below, which must check for changes in doPageWrites and RedoRecPtr.
883 : * Those checks are only needed for records that can contain buffer
884 : * references, and an XLOG_SWITCH record never does.
885 : */
886 : Assert(!XLogRecPtrIsValid(fpw_lsn));
887 1604 : WALInsertLockAcquireExclusive();
888 1604 : inserted = ReserveXLogSwitch(&StartPos, &EndPos, &rechdr->xl_prev);
889 : }
890 : else
891 : {
892 : Assert(class == WALINSERT_SPECIAL_CHECKPOINT);
893 :
894 : /*
895 : * We need to update both the local and shared copies of RedoRecPtr,
896 : * which means that we need to hold all the WAL insertion locks.
897 : * However, there can't be any buffer references, so as above, we need
898 : * not check RedoRecPtr before inserting the record; we just need to
899 : * update it afterwards.
900 : */
901 : Assert(!XLogRecPtrIsValid(fpw_lsn));
902 1864 : WALInsertLockAcquireExclusive();
903 1864 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
904 : &rechdr->xl_prev);
905 1864 : RedoRecPtr = Insert->RedoRecPtr = StartPos;
906 1864 : inserted = true;
907 : }
908 :
909 30764054 : if (inserted)
910 : {
911 : /*
912 : * Now that xl_prev has been filled in, calculate CRC of the record
913 : * header.
914 : */
915 30763936 : rdata_crc = rechdr->xl_crc;
916 30763936 : COMP_CRC32C(rdata_crc, rechdr, offsetof(XLogRecord, xl_crc));
917 30763936 : FIN_CRC32C(rdata_crc);
918 30763936 : rechdr->xl_crc = rdata_crc;
919 :
920 : /*
921 : * All the record data, including the header, is now ready to be
922 : * inserted. Copy the record in the space reserved.
923 : */
924 30763936 : CopyXLogRecordToWAL(rechdr->xl_tot_len,
925 : class == WALINSERT_SPECIAL_SWITCH, rdata,
926 : StartPos, EndPos, insertTLI);
927 :
928 : /*
929 : * Unless record is flagged as not important, update LSN of last
930 : * important record in the current slot. When holding all locks, just
931 : * update the first one.
932 : */
933 30763936 : if ((flags & XLOG_MARK_UNIMPORTANT) == 0)
934 : {
935 30521808 : int lockno = holdingAllLocks ? 0 : MyLockNo;
936 :
937 30521808 : WALInsertLocks[lockno].l.lastImportantAt = StartPos;
938 : }
939 : }
940 : else
941 : {
942 : /*
943 : * This was an xlog-switch record, but the current insert location was
944 : * already exactly at the beginning of a segment, so there was no need
945 : * to do anything.
946 : */
947 : }
948 :
949 : /*
950 : * Done! Let others know that we're finished.
951 : */
952 30764054 : WALInsertLockRelease();
953 :
954 30764054 : END_CRIT_SECTION();
955 :
956 30764054 : MarkCurrentTransactionIdLoggedIfAny();
957 :
958 : /*
959 : * Mark top transaction id is logged (if needed) so that we should not try
960 : * to log it again with the next WAL record in the current subtransaction.
961 : */
962 30764054 : if (topxid_included)
963 438 : MarkSubxactTopXidLogged();
964 :
965 : /*
966 : * Update shared LogwrtRqst.Write, if we crossed page boundary.
967 : */
968 30764054 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
969 : {
970 3404326 : SpinLockAcquire(&XLogCtl->info_lck);
971 : /* advance global request to include new block(s) */
972 3404326 : if (XLogCtl->LogwrtRqst.Write < EndPos)
973 3272560 : XLogCtl->LogwrtRqst.Write = EndPos;
974 3404326 : SpinLockRelease(&XLogCtl->info_lck);
975 3404326 : RefreshXLogWriteResult(LogwrtResult);
976 : }
977 :
978 : /*
979 : * If this was an XLOG_SWITCH record, flush the record and the empty
980 : * padding space that fills the rest of the segment, and perform
981 : * end-of-segment actions (eg, notifying archiver).
982 : */
983 30764054 : if (class == WALINSERT_SPECIAL_SWITCH)
984 : {
985 : TRACE_POSTGRESQL_WAL_SWITCH();
986 1604 : XLogFlush(EndPos);
987 :
988 : /*
989 : * Even though we reserved the rest of the segment for us, which is
990 : * reflected in EndPos, we return a pointer to just the end of the
991 : * xlog-switch record.
992 : */
993 1604 : if (inserted)
994 : {
995 1486 : EndPos = StartPos + SizeOfXLogRecord;
996 1486 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
997 : {
998 0 : uint64 offset = XLogSegmentOffset(EndPos, wal_segment_size);
999 :
1000 0 : if (offset == EndPos % XLOG_BLCKSZ)
1001 0 : EndPos += SizeOfXLogLongPHD;
1002 : else
1003 0 : EndPos += SizeOfXLogShortPHD;
1004 : }
1005 : }
1006 : }
1007 :
1008 : #ifdef WAL_DEBUG
1009 : if (XLOG_DEBUG)
1010 : {
1011 : static XLogReaderState *debug_reader = NULL;
1012 : XLogRecord *record;
1013 : DecodedXLogRecord *decoded;
1014 : StringInfoData buf;
1015 : StringInfoData recordBuf;
1016 : char *errormsg = NULL;
1017 : MemoryContext oldCxt;
1018 :
1019 : oldCxt = MemoryContextSwitchTo(walDebugCxt);
1020 :
1021 : initStringInfo(&buf);
1022 : appendStringInfo(&buf, "INSERT @ %X/%08X: ", LSN_FORMAT_ARGS(EndPos));
1023 :
1024 : /*
1025 : * We have to piece together the WAL record data from the XLogRecData
1026 : * entries, so that we can pass it to the rm_desc function as one
1027 : * contiguous chunk.
1028 : */
1029 : initStringInfo(&recordBuf);
1030 : for (; rdata != NULL; rdata = rdata->next)
1031 : appendBinaryStringInfo(&recordBuf, rdata->data, rdata->len);
1032 :
1033 : /* We also need temporary space to decode the record. */
1034 : record = (XLogRecord *) recordBuf.data;
1035 : decoded = (DecodedXLogRecord *)
1036 : palloc(DecodeXLogRecordRequiredSpace(record->xl_tot_len));
1037 :
1038 : if (!debug_reader)
1039 : debug_reader = XLogReaderAllocate(wal_segment_size, NULL,
1040 : XL_ROUTINE(.page_read = NULL,
1041 : .segment_open = NULL,
1042 : .segment_close = NULL),
1043 : NULL);
1044 : if (!debug_reader)
1045 : {
1046 : appendStringInfoString(&buf, "error decoding record: out of memory while allocating a WAL reading processor");
1047 : }
1048 : else if (!DecodeXLogRecord(debug_reader,
1049 : decoded,
1050 : record,
1051 : EndPos,
1052 : &errormsg))
1053 : {
1054 : appendStringInfo(&buf, "error decoding record: %s",
1055 : errormsg ? errormsg : "no error message");
1056 : }
1057 : else
1058 : {
1059 : appendStringInfoString(&buf, " - ");
1060 :
1061 : debug_reader->record = decoded;
1062 : xlog_outdesc(&buf, debug_reader);
1063 : debug_reader->record = NULL;
1064 : }
1065 : elog(LOG, "%s", buf.data);
1066 :
1067 : pfree(decoded);
1068 : pfree(buf.data);
1069 : pfree(recordBuf.data);
1070 : MemoryContextSwitchTo(oldCxt);
1071 : }
1072 : #endif
1073 :
1074 : /*
1075 : * Update our global variables
1076 : */
1077 30764054 : ProcLastRecPtr = StartPos;
1078 30764054 : XactLastRecEnd = EndPos;
1079 :
1080 : /* Report WAL traffic to the instrumentation. */
1081 30764054 : if (inserted)
1082 : {
1083 30763936 : pgWalUsage.wal_bytes += rechdr->xl_tot_len;
1084 30763936 : pgWalUsage.wal_records++;
1085 30763936 : pgWalUsage.wal_fpi += num_fpi;
1086 30763936 : pgWalUsage.wal_fpi_bytes += fpi_bytes;
1087 :
1088 : /* Required for the flush of pending stats WAL data */
1089 30763936 : pgstat_report_fixed = true;
1090 : }
1091 :
1092 30764054 : return EndPos;
1093 : }
1094 :
1095 : /*
1096 : * Reserves the right amount of space for a record of given size from the WAL.
1097 : * *StartPos is set to the beginning of the reserved section, *EndPos to
1098 : * its end+1. *PrevPtr is set to the beginning of the previous record; it is
1099 : * used to set the xl_prev of this record.
1100 : *
1101 : * This is the performance critical part of XLogInsert that must be serialized
1102 : * across backends. The rest can happen mostly in parallel. Try to keep this
1103 : * section as short as possible, insertpos_lck can be heavily contended on a
1104 : * busy system.
1105 : *
1106 : * NB: The space calculation here must match the code in CopyXLogRecordToWAL,
1107 : * where we actually copy the record to the reserved space.
1108 : *
1109 : * NB: Testing shows that XLogInsertRecord runs faster if this code is inlined;
1110 : * however, because there are two call sites, the compiler is reluctant to
1111 : * inline. We use pg_attribute_always_inline here to try to convince it.
1112 : */
1113 : static pg_attribute_always_inline void
1114 30762450 : ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1115 : XLogRecPtr *PrevPtr)
1116 : {
1117 30762450 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1118 : uint64 startbytepos;
1119 : uint64 endbytepos;
1120 : uint64 prevbytepos;
1121 :
1122 30762450 : size = MAXALIGN(size);
1123 :
1124 : /* All (non xlog-switch) records should contain data. */
1125 : Assert(size > SizeOfXLogRecord);
1126 :
1127 : /*
1128 : * The duration the spinlock needs to be held is minimized by minimizing
1129 : * the calculations that have to be done while holding the lock. The
1130 : * current tip of reserved WAL is kept in CurrBytePos, as a byte position
1131 : * that only counts "usable" bytes in WAL, that is, it excludes all WAL
1132 : * page headers. The mapping between "usable" byte positions and physical
1133 : * positions (XLogRecPtrs) can be done outside the locked region, and
1134 : * because the usable byte position doesn't include any headers, reserving
1135 : * X bytes from WAL is almost as simple as "CurrBytePos += X".
1136 : */
1137 30762450 : SpinLockAcquire(&Insert->insertpos_lck);
1138 :
1139 30762450 : startbytepos = Insert->CurrBytePos;
1140 30762450 : endbytepos = startbytepos + size;
1141 30762450 : prevbytepos = Insert->PrevBytePos;
1142 30762450 : Insert->CurrBytePos = endbytepos;
1143 30762450 : Insert->PrevBytePos = startbytepos;
1144 :
1145 30762450 : SpinLockRelease(&Insert->insertpos_lck);
1146 :
1147 30762450 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1148 30762450 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1149 30762450 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1150 :
1151 : /*
1152 : * Check that the conversions between "usable byte positions" and
1153 : * XLogRecPtrs work consistently in both directions.
1154 : */
1155 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1156 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1157 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1158 30762450 : }
1159 :
1160 : /*
1161 : * Like ReserveXLogInsertLocation(), but for an xlog-switch record.
1162 : *
1163 : * A log-switch record is handled slightly differently. The rest of the
1164 : * segment will be reserved for this insertion, as indicated by the returned
1165 : * *EndPos value. However, if we are already at the beginning of the current
1166 : * segment, *StartPos and *EndPos are set to the current location without
1167 : * reserving any space, and the function returns false.
1168 : */
1169 : static bool
1170 1604 : ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1171 : {
1172 1604 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1173 : uint64 startbytepos;
1174 : uint64 endbytepos;
1175 : uint64 prevbytepos;
1176 1604 : uint32 size = MAXALIGN(SizeOfXLogRecord);
1177 : XLogRecPtr ptr;
1178 : uint32 segleft;
1179 :
1180 : /*
1181 : * These calculations are a bit heavy-weight to be done while holding a
1182 : * spinlock, but since we're holding all the WAL insertion locks, there
1183 : * are no other inserters competing for it. GetXLogInsertRecPtr() does
1184 : * compete for it, but that's not called very frequently.
1185 : */
1186 1604 : SpinLockAcquire(&Insert->insertpos_lck);
1187 :
1188 1604 : startbytepos = Insert->CurrBytePos;
1189 :
1190 1604 : ptr = XLogBytePosToEndRecPtr(startbytepos);
1191 1604 : if (XLogSegmentOffset(ptr, wal_segment_size) == 0)
1192 : {
1193 118 : SpinLockRelease(&Insert->insertpos_lck);
1194 118 : *EndPos = *StartPos = ptr;
1195 118 : return false;
1196 : }
1197 :
1198 1486 : endbytepos = startbytepos + size;
1199 1486 : prevbytepos = Insert->PrevBytePos;
1200 :
1201 1486 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1202 1486 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1203 :
1204 1486 : segleft = wal_segment_size - XLogSegmentOffset(*EndPos, wal_segment_size);
1205 1486 : if (segleft != wal_segment_size)
1206 : {
1207 : /* consume the rest of the segment */
1208 1486 : *EndPos += segleft;
1209 1486 : endbytepos = XLogRecPtrToBytePos(*EndPos);
1210 : }
1211 1486 : Insert->CurrBytePos = endbytepos;
1212 1486 : Insert->PrevBytePos = startbytepos;
1213 :
1214 1486 : SpinLockRelease(&Insert->insertpos_lck);
1215 :
1216 1486 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1217 :
1218 : Assert(XLogSegmentOffset(*EndPos, wal_segment_size) == 0);
1219 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1220 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1221 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1222 :
1223 1486 : return true;
1224 : }
1225 :
1226 : /*
1227 : * Subroutine of XLogInsertRecord. Copies a WAL record to an already-reserved
1228 : * area in the WAL.
1229 : */
1230 : static void
1231 30763936 : CopyXLogRecordToWAL(int write_len, bool isLogSwitch, XLogRecData *rdata,
1232 : XLogRecPtr StartPos, XLogRecPtr EndPos, TimeLineID tli)
1233 : {
1234 : char *currpos;
1235 : int freespace;
1236 : int written;
1237 : XLogRecPtr CurrPos;
1238 : XLogPageHeader pagehdr;
1239 :
1240 : /*
1241 : * Get a pointer to the right place in the right WAL buffer to start
1242 : * inserting to.
1243 : */
1244 30763936 : CurrPos = StartPos;
1245 30763936 : currpos = GetXLogBuffer(CurrPos, tli);
1246 30763936 : freespace = INSERT_FREESPACE(CurrPos);
1247 :
1248 : /*
1249 : * there should be enough space for at least the first field (xl_tot_len)
1250 : * on this page.
1251 : */
1252 : Assert(freespace >= sizeof(uint32));
1253 :
1254 : /* Copy record data */
1255 30763936 : written = 0;
1256 145158760 : while (rdata != NULL)
1257 : {
1258 114394824 : const char *rdata_data = rdata->data;
1259 114394824 : int rdata_len = rdata->len;
1260 :
1261 118018094 : while (rdata_len > freespace)
1262 : {
1263 : /*
1264 : * Write what fits on this page, and continue on the next page.
1265 : */
1266 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || freespace == 0);
1267 3623270 : memcpy(currpos, rdata_data, freespace);
1268 3623270 : rdata_data += freespace;
1269 3623270 : rdata_len -= freespace;
1270 3623270 : written += freespace;
1271 3623270 : CurrPos += freespace;
1272 :
1273 : /*
1274 : * Get pointer to beginning of next page, and set the xlp_rem_len
1275 : * in the page header. Set XLP_FIRST_IS_CONTRECORD.
1276 : *
1277 : * It's safe to set the contrecord flag and xlp_rem_len without a
1278 : * lock on the page. All the other flags were already set when the
1279 : * page was initialized, in AdvanceXLInsertBuffer, and we're the
1280 : * only backend that needs to set the contrecord flag.
1281 : */
1282 3623270 : currpos = GetXLogBuffer(CurrPos, tli);
1283 3623270 : pagehdr = (XLogPageHeader) currpos;
1284 3623270 : pagehdr->xlp_rem_len = write_len - written;
1285 3623270 : pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1286 :
1287 : /* skip over the page header */
1288 3623270 : if (XLogSegmentOffset(CurrPos, wal_segment_size) == 0)
1289 : {
1290 2346 : CurrPos += SizeOfXLogLongPHD;
1291 2346 : currpos += SizeOfXLogLongPHD;
1292 : }
1293 : else
1294 : {
1295 3620924 : CurrPos += SizeOfXLogShortPHD;
1296 3620924 : currpos += SizeOfXLogShortPHD;
1297 : }
1298 3623270 : freespace = INSERT_FREESPACE(CurrPos);
1299 : }
1300 :
1301 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1302 114394824 : memcpy(currpos, rdata_data, rdata_len);
1303 114394824 : currpos += rdata_len;
1304 114394824 : CurrPos += rdata_len;
1305 114394824 : freespace -= rdata_len;
1306 114394824 : written += rdata_len;
1307 :
1308 114394824 : rdata = rdata->next;
1309 : }
1310 : Assert(written == write_len);
1311 :
1312 : /*
1313 : * If this was an xlog-switch, it's not enough to write the switch record,
1314 : * we also have to consume all the remaining space in the WAL segment. We
1315 : * have already reserved that space, but we need to actually fill it.
1316 : */
1317 30763936 : if (isLogSwitch && XLogSegmentOffset(CurrPos, wal_segment_size) != 0)
1318 : {
1319 : /* An xlog-switch record doesn't contain any data besides the header */
1320 : Assert(write_len == SizeOfXLogRecord);
1321 :
1322 : /* Assert that we did reserve the right amount of space */
1323 : Assert(XLogSegmentOffset(EndPos, wal_segment_size) == 0);
1324 :
1325 : /* Use up all the remaining space on the current page */
1326 1486 : CurrPos += freespace;
1327 :
1328 : /*
1329 : * Cause all remaining pages in the segment to be flushed, leaving the
1330 : * XLog position where it should be, at the start of the next segment.
1331 : * We do this one page at a time, to make sure we don't deadlock
1332 : * against ourselves if wal_buffers < wal_segment_size.
1333 : */
1334 1465254 : while (CurrPos < EndPos)
1335 : {
1336 : /*
1337 : * The minimal action to flush the page would be to call
1338 : * WALInsertLockUpdateInsertingAt(CurrPos) followed by
1339 : * AdvanceXLInsertBuffer(...). The page would be left initialized
1340 : * mostly to zeros, except for the page header (always the short
1341 : * variant, as this is never a segment's first page).
1342 : *
1343 : * The large vistas of zeros are good for compressibility, but the
1344 : * headers interrupting them every XLOG_BLCKSZ (with values that
1345 : * differ from page to page) are not. The effect varies with
1346 : * compression tool, but bzip2 for instance compresses about an
1347 : * order of magnitude worse if those headers are left in place.
1348 : *
1349 : * Rather than complicating AdvanceXLInsertBuffer itself (which is
1350 : * called in heavily-loaded circumstances as well as this lightly-
1351 : * loaded one) with variant behavior, we just use GetXLogBuffer
1352 : * (which itself calls the two methods we need) to get the pointer
1353 : * and zero most of the page. Then we just zero the page header.
1354 : */
1355 1463768 : currpos = GetXLogBuffer(CurrPos, tli);
1356 5855072 : MemSet(currpos, 0, SizeOfXLogShortPHD);
1357 :
1358 1463768 : CurrPos += XLOG_BLCKSZ;
1359 : }
1360 : }
1361 : else
1362 : {
1363 : /* Align the end position, so that the next record starts aligned */
1364 30762450 : CurrPos = MAXALIGN64(CurrPos);
1365 : }
1366 :
1367 30763936 : if (CurrPos != EndPos)
1368 0 : ereport(PANIC,
1369 : errcode(ERRCODE_DATA_CORRUPTED),
1370 : errmsg_internal("space reserved for WAL record does not match what was written"));
1371 30763936 : }
1372 :
1373 : /*
1374 : * Acquire a WAL insertion lock, for inserting to WAL.
1375 : */
1376 : static void
1377 30776608 : WALInsertLockAcquire(void)
1378 : {
1379 : bool immed;
1380 :
1381 : /*
1382 : * It doesn't matter which of the WAL insertion locks we acquire, so try
1383 : * the one we used last time. If the system isn't particularly busy, it's
1384 : * a good bet that it's still available, and it's good to have some
1385 : * affinity to a particular lock so that you don't unnecessarily bounce
1386 : * cache lines between processes when there's no contention.
1387 : *
1388 : * If this is the first time through in this backend, pick a lock
1389 : * (semi-)randomly. This allows the locks to be used evenly if you have a
1390 : * lot of very short connections.
1391 : */
1392 : static int lockToTry = -1;
1393 :
1394 30776608 : if (lockToTry == -1)
1395 16878 : lockToTry = MyProcNumber % NUM_XLOGINSERT_LOCKS;
1396 30776608 : MyLockNo = lockToTry;
1397 :
1398 : /*
1399 : * The insertingAt value is initially set to 0, as we don't know our
1400 : * insert location yet.
1401 : */
1402 30776608 : immed = LWLockAcquire(&WALInsertLocks[MyLockNo].l.lock, LW_EXCLUSIVE);
1403 30776608 : if (!immed)
1404 : {
1405 : /*
1406 : * If we couldn't get the lock immediately, try another lock next
1407 : * time. On a system with more insertion locks than concurrent
1408 : * inserters, this causes all the inserters to eventually migrate to a
1409 : * lock that no-one else is using. On a system with more inserters
1410 : * than locks, it still helps to distribute the inserters evenly
1411 : * across the locks.
1412 : */
1413 24068 : lockToTry = (lockToTry + 1) % NUM_XLOGINSERT_LOCKS;
1414 : }
1415 30776608 : }
1416 :
1417 : /*
1418 : * Acquire all WAL insertion locks, to prevent other backends from inserting
1419 : * to WAL.
1420 : */
1421 : static void
1422 8946 : WALInsertLockAcquireExclusive(void)
1423 : {
1424 : int i;
1425 :
1426 : /*
1427 : * When holding all the locks, all but the last lock's insertingAt
1428 : * indicator is set to 0xFFFFFFFFFFFFFFFF, which is higher than any real
1429 : * XLogRecPtr value, to make sure that no-one blocks waiting on those.
1430 : */
1431 71568 : for (i = 0; i < NUM_XLOGINSERT_LOCKS - 1; i++)
1432 : {
1433 62622 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1434 62622 : LWLockUpdateVar(&WALInsertLocks[i].l.lock,
1435 62622 : &WALInsertLocks[i].l.insertingAt,
1436 : PG_UINT64_MAX);
1437 : }
1438 : /* Variable value reset to 0 at release */
1439 8946 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1440 :
1441 8946 : holdingAllLocks = true;
1442 8946 : }
1443 :
1444 : /*
1445 : * Release our insertion lock (or locks, if we're holding them all).
1446 : *
1447 : * NB: Reset all variables to 0, so they cause LWLockWaitForVar to block the
1448 : * next time the lock is acquired.
1449 : */
1450 : static void
1451 30785554 : WALInsertLockRelease(void)
1452 : {
1453 30785554 : if (holdingAllLocks)
1454 : {
1455 : int i;
1456 :
1457 80514 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1458 71568 : LWLockReleaseClearVar(&WALInsertLocks[i].l.lock,
1459 71568 : &WALInsertLocks[i].l.insertingAt,
1460 : 0);
1461 :
1462 8946 : holdingAllLocks = false;
1463 : }
1464 : else
1465 : {
1466 30776608 : LWLockReleaseClearVar(&WALInsertLocks[MyLockNo].l.lock,
1467 30776608 : &WALInsertLocks[MyLockNo].l.insertingAt,
1468 : 0);
1469 : }
1470 30785554 : }
1471 :
1472 : /*
1473 : * Update our insertingAt value, to let others know that we've finished
1474 : * inserting up to that point.
1475 : */
1476 : static void
1477 4988078 : WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt)
1478 : {
1479 4988078 : if (holdingAllLocks)
1480 : {
1481 : /*
1482 : * We use the last lock to mark our actual position, see comments in
1483 : * WALInsertLockAcquireExclusive.
1484 : */
1485 1456588 : LWLockUpdateVar(&WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.lock,
1486 1456588 : &WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.insertingAt,
1487 : insertingAt);
1488 : }
1489 : else
1490 3531490 : LWLockUpdateVar(&WALInsertLocks[MyLockNo].l.lock,
1491 3531490 : &WALInsertLocks[MyLockNo].l.insertingAt,
1492 : insertingAt);
1493 4988078 : }
1494 :
1495 : /*
1496 : * Wait for any WAL insertions < upto to finish.
1497 : *
1498 : * Returns the location of the oldest insertion that is still in-progress.
1499 : * Any WAL prior to that point has been fully copied into WAL buffers, and
1500 : * can be flushed out to disk. Because this waits for any insertions older
1501 : * than 'upto' to finish, the return value is always >= 'upto'.
1502 : *
1503 : * Note: When you are about to write out WAL, you must call this function
1504 : * *before* acquiring WALWriteLock, to avoid deadlocks. This function might
1505 : * need to wait for an insertion to finish (or at least advance to next
1506 : * uninitialized page), and the inserter might need to evict an old WAL buffer
1507 : * to make room for a new one, which in turn requires WALWriteLock.
1508 : */
1509 : static XLogRecPtr
1510 4700098 : WaitXLogInsertionsToFinish(XLogRecPtr upto)
1511 : {
1512 : uint64 bytepos;
1513 : XLogRecPtr inserted;
1514 : XLogRecPtr reservedUpto;
1515 : XLogRecPtr finishedUpto;
1516 4700098 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1517 : int i;
1518 :
1519 4700098 : if (MyProc == NULL)
1520 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1521 :
1522 : /*
1523 : * Check if there's any work to do. Use a barrier to ensure we get the
1524 : * freshest value.
1525 : */
1526 4700098 : inserted = pg_atomic_read_membarrier_u64(&XLogCtl->logInsertResult);
1527 4700098 : if (upto <= inserted)
1528 3758914 : return inserted;
1529 :
1530 : /* Read the current insert position */
1531 941184 : SpinLockAcquire(&Insert->insertpos_lck);
1532 941184 : bytepos = Insert->CurrBytePos;
1533 941184 : SpinLockRelease(&Insert->insertpos_lck);
1534 941184 : reservedUpto = XLogBytePosToEndRecPtr(bytepos);
1535 :
1536 : /*
1537 : * No-one should request to flush a piece of WAL that hasn't even been
1538 : * reserved yet. However, it can happen if there is a block with a bogus
1539 : * LSN on disk, for example. XLogFlush checks for that situation and
1540 : * complains, but only after the flush. Here we just assume that to mean
1541 : * that all WAL that has been reserved needs to be finished. In this
1542 : * corner-case, the return value can be smaller than 'upto' argument.
1543 : */
1544 941184 : if (upto > reservedUpto)
1545 : {
1546 0 : ereport(LOG,
1547 : errmsg("request to flush past end of generated WAL; request %X/%08X, current position %X/%08X",
1548 : LSN_FORMAT_ARGS(upto), LSN_FORMAT_ARGS(reservedUpto)));
1549 0 : upto = reservedUpto;
1550 : }
1551 :
1552 : /*
1553 : * Loop through all the locks, sleeping on any in-progress insert older
1554 : * than 'upto'.
1555 : *
1556 : * finishedUpto is our return value, indicating the point upto which all
1557 : * the WAL insertions have been finished. Initialize it to the head of
1558 : * reserved WAL, and as we iterate through the insertion locks, back it
1559 : * out for any insertion that's still in progress.
1560 : */
1561 941184 : finishedUpto = reservedUpto;
1562 8470656 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1563 : {
1564 7529472 : XLogRecPtr insertingat = InvalidXLogRecPtr;
1565 :
1566 : do
1567 : {
1568 : /*
1569 : * See if this insertion is in progress. LWLockWaitForVar will
1570 : * wait for the lock to be released, or for the 'value' to be set
1571 : * by a LWLockUpdateVar call. When a lock is initially acquired,
1572 : * its value is 0 (InvalidXLogRecPtr), which means that we don't
1573 : * know where it's inserting yet. We will have to wait for it. If
1574 : * it's a small insertion, the record will most likely fit on the
1575 : * same page and the inserter will release the lock without ever
1576 : * calling LWLockUpdateVar. But if it has to sleep, it will
1577 : * advertise the insertion point with LWLockUpdateVar before
1578 : * sleeping.
1579 : *
1580 : * In this loop we are only waiting for insertions that started
1581 : * before WaitXLogInsertionsToFinish was called. The lack of
1582 : * memory barriers in the loop means that we might see locks as
1583 : * "unused" that have since become used. This is fine because
1584 : * they only can be used for later insertions that we would not
1585 : * want to wait on anyway. Not taking a lock to acquire the
1586 : * current insertingAt value means that we might see older
1587 : * insertingAt values. This is also fine, because if we read a
1588 : * value too old, we will add ourselves to the wait queue, which
1589 : * contains atomic operations.
1590 : */
1591 7738900 : if (LWLockWaitForVar(&WALInsertLocks[i].l.lock,
1592 7738900 : &WALInsertLocks[i].l.insertingAt,
1593 : insertingat, &insertingat))
1594 : {
1595 : /* the lock was free, so no insertion in progress */
1596 5313766 : insertingat = InvalidXLogRecPtr;
1597 5313766 : break;
1598 : }
1599 :
1600 : /*
1601 : * This insertion is still in progress. Have to wait, unless the
1602 : * inserter has proceeded past 'upto'.
1603 : */
1604 2425134 : } while (insertingat < upto);
1605 :
1606 7529472 : if (XLogRecPtrIsValid(insertingat) && insertingat < finishedUpto)
1607 796182 : finishedUpto = insertingat;
1608 : }
1609 :
1610 : /*
1611 : * Advance the limit we know to have been inserted and return the freshest
1612 : * value we know of, which might be beyond what we requested if somebody
1613 : * is concurrently doing this with an 'upto' pointer ahead of us.
1614 : */
1615 941184 : finishedUpto = pg_atomic_monotonic_advance_u64(&XLogCtl->logInsertResult,
1616 : finishedUpto);
1617 :
1618 941184 : return finishedUpto;
1619 : }
1620 :
1621 : /*
1622 : * Get a pointer to the right location in the WAL buffer containing the
1623 : * given XLogRecPtr.
1624 : *
1625 : * If the page is not initialized yet, it is initialized. That might require
1626 : * evicting an old dirty buffer from the buffer cache, which means I/O.
1627 : *
1628 : * The caller must ensure that the page containing the requested location
1629 : * isn't evicted yet, and won't be evicted. The way to ensure that is to
1630 : * hold onto a WAL insertion lock with the insertingAt position set to
1631 : * something <= ptr. GetXLogBuffer() will update insertingAt if it needs
1632 : * to evict an old page from the buffer. (This means that once you call
1633 : * GetXLogBuffer() with a given 'ptr', you must not access anything before
1634 : * that point anymore, and must not call GetXLogBuffer() with an older 'ptr'
1635 : * later, because older buffers might be recycled already)
1636 : */
1637 : static char *
1638 35850996 : GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli)
1639 : {
1640 : int idx;
1641 : XLogRecPtr endptr;
1642 : static uint64 cachedPage = 0;
1643 : static char *cachedPos = NULL;
1644 : XLogRecPtr expectedEndPtr;
1645 :
1646 : /*
1647 : * Fast path for the common case that we need to access again the same
1648 : * page as last time.
1649 : */
1650 35850996 : if (ptr / XLOG_BLCKSZ == cachedPage)
1651 : {
1652 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1653 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1654 30059008 : return cachedPos + ptr % XLOG_BLCKSZ;
1655 : }
1656 :
1657 : /*
1658 : * The XLog buffer cache is organized so that a page is always loaded to a
1659 : * particular buffer. That way we can easily calculate the buffer a given
1660 : * page must be loaded into, from the XLogRecPtr alone.
1661 : */
1662 5791988 : idx = XLogRecPtrToBufIdx(ptr);
1663 :
1664 : /*
1665 : * See what page is loaded in the buffer at the moment. It could be the
1666 : * page we're looking for, or something older. It can't be anything newer
1667 : * - that would imply the page we're looking for has already been written
1668 : * out to disk and evicted, and the caller is responsible for making sure
1669 : * that doesn't happen.
1670 : *
1671 : * We don't hold a lock while we read the value. If someone is just about
1672 : * to initialize or has just initialized the page, it's possible that we
1673 : * get InvalidXLogRecPtr. That's ok, we'll grab the mapping lock (in
1674 : * AdvanceXLInsertBuffer) and retry if we see anything other than the page
1675 : * we're looking for.
1676 : */
1677 5791988 : expectedEndPtr = ptr;
1678 5791988 : expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
1679 :
1680 5791988 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1681 5791988 : if (expectedEndPtr != endptr)
1682 : {
1683 : XLogRecPtr initializedUpto;
1684 :
1685 : /*
1686 : * Before calling AdvanceXLInsertBuffer(), which can block, let others
1687 : * know how far we're finished with inserting the record.
1688 : *
1689 : * NB: If 'ptr' points to just after the page header, advertise a
1690 : * position at the beginning of the page rather than 'ptr' itself. If
1691 : * there are no other insertions running, someone might try to flush
1692 : * up to our advertised location. If we advertised a position after
1693 : * the page header, someone might try to flush the page header, even
1694 : * though page might actually not be initialized yet. As the first
1695 : * inserter on the page, we are effectively responsible for making
1696 : * sure that it's initialized, before we let insertingAt to move past
1697 : * the page header.
1698 : */
1699 4988078 : if (ptr % XLOG_BLCKSZ == SizeOfXLogShortPHD &&
1700 12846 : XLogSegmentOffset(ptr, wal_segment_size) > XLOG_BLCKSZ)
1701 12846 : initializedUpto = ptr - SizeOfXLogShortPHD;
1702 4975232 : else if (ptr % XLOG_BLCKSZ == SizeOfXLogLongPHD &&
1703 2032 : XLogSegmentOffset(ptr, wal_segment_size) < XLOG_BLCKSZ)
1704 1234 : initializedUpto = ptr - SizeOfXLogLongPHD;
1705 : else
1706 4973998 : initializedUpto = ptr;
1707 :
1708 4988078 : WALInsertLockUpdateInsertingAt(initializedUpto);
1709 :
1710 4988078 : AdvanceXLInsertBuffer(ptr, tli, false);
1711 4988078 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1712 :
1713 4988078 : if (expectedEndPtr != endptr)
1714 0 : elog(PANIC, "could not find WAL buffer for %X/%08X",
1715 : LSN_FORMAT_ARGS(ptr));
1716 : }
1717 : else
1718 : {
1719 : /*
1720 : * Make sure the initialization of the page is visible to us, and
1721 : * won't arrive later to overwrite the WAL data we write on the page.
1722 : */
1723 803910 : pg_memory_barrier();
1724 : }
1725 :
1726 : /*
1727 : * Found the buffer holding this page. Return a pointer to the right
1728 : * offset within the page.
1729 : */
1730 5791988 : cachedPage = ptr / XLOG_BLCKSZ;
1731 5791988 : cachedPos = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1732 :
1733 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1734 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1735 :
1736 5791988 : return cachedPos + ptr % XLOG_BLCKSZ;
1737 : }
1738 :
1739 : /*
1740 : * Read WAL data directly from WAL buffers, if available. Returns the number
1741 : * of bytes read successfully.
1742 : *
1743 : * Fewer than 'count' bytes may be read if some of the requested WAL data has
1744 : * already been evicted.
1745 : *
1746 : * No locks are taken.
1747 : *
1748 : * Caller should ensure that it reads no further than LogwrtResult.Write
1749 : * (which should have been updated by the caller when determining how far to
1750 : * read). The 'tli' argument is only used as a convenient safety check so that
1751 : * callers do not read from WAL buffers on a historical timeline.
1752 : */
1753 : Size
1754 212594 : WALReadFromBuffers(char *dstbuf, XLogRecPtr startptr, Size count,
1755 : TimeLineID tli)
1756 : {
1757 212594 : char *pdst = dstbuf;
1758 212594 : XLogRecPtr recptr = startptr;
1759 : XLogRecPtr inserted;
1760 212594 : Size nbytes = count;
1761 :
1762 212594 : if (RecoveryInProgress() || tli != GetWALInsertionTimeLine())
1763 1810 : return 0;
1764 :
1765 : Assert(XLogRecPtrIsValid(startptr));
1766 :
1767 : /*
1768 : * Caller should ensure that the requested data has been inserted into WAL
1769 : * buffers before we try to read it.
1770 : */
1771 210784 : inserted = pg_atomic_read_u64(&XLogCtl->logInsertResult);
1772 210784 : if (startptr + count > inserted)
1773 0 : ereport(ERROR,
1774 : errmsg("cannot read past end of generated WAL: requested %X/%08X, current position %X/%08X",
1775 : LSN_FORMAT_ARGS(startptr + count),
1776 : LSN_FORMAT_ARGS(inserted)));
1777 :
1778 : /*
1779 : * Loop through the buffers without a lock. For each buffer, atomically
1780 : * read and verify the end pointer, then copy the data out, and finally
1781 : * re-read and re-verify the end pointer.
1782 : *
1783 : * Once a page is evicted, it never returns to the WAL buffers, so if the
1784 : * end pointer matches the expected end pointer before and after we copy
1785 : * the data, then the right page must have been present during the data
1786 : * copy. Read barriers are necessary to ensure that the data copy actually
1787 : * happens between the two verification steps.
1788 : *
1789 : * If either verification fails, we simply terminate the loop and return
1790 : * with the data that had been already copied out successfully.
1791 : */
1792 250538 : while (nbytes > 0)
1793 : {
1794 238930 : uint32 offset = recptr % XLOG_BLCKSZ;
1795 238930 : int idx = XLogRecPtrToBufIdx(recptr);
1796 : XLogRecPtr expectedEndPtr;
1797 : XLogRecPtr endptr;
1798 : const char *page;
1799 : const char *psrc;
1800 : Size npagebytes;
1801 :
1802 : /*
1803 : * Calculate the end pointer we expect in the xlblocks array if the
1804 : * correct page is present.
1805 : */
1806 238930 : expectedEndPtr = recptr + (XLOG_BLCKSZ - offset);
1807 :
1808 : /*
1809 : * First verification step: check that the correct page is present in
1810 : * the WAL buffers.
1811 : */
1812 238930 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1813 238930 : if (expectedEndPtr != endptr)
1814 199176 : break;
1815 :
1816 : /*
1817 : * The correct page is present (or was at the time the endptr was
1818 : * read; must re-verify later). Calculate pointer to source data and
1819 : * determine how much data to read from this page.
1820 : */
1821 39754 : page = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1822 39754 : psrc = page + offset;
1823 39754 : npagebytes = Min(nbytes, XLOG_BLCKSZ - offset);
1824 :
1825 : /*
1826 : * Ensure that the data copy and the first verification step are not
1827 : * reordered.
1828 : */
1829 39754 : pg_read_barrier();
1830 :
1831 : /* data copy */
1832 39754 : memcpy(pdst, psrc, npagebytes);
1833 :
1834 : /*
1835 : * Ensure that the data copy and the second verification step are not
1836 : * reordered.
1837 : */
1838 39754 : pg_read_barrier();
1839 :
1840 : /*
1841 : * Second verification step: check that the page we read from wasn't
1842 : * evicted while we were copying the data.
1843 : */
1844 39754 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1845 39754 : if (expectedEndPtr != endptr)
1846 0 : break;
1847 :
1848 39754 : pdst += npagebytes;
1849 39754 : recptr += npagebytes;
1850 39754 : nbytes -= npagebytes;
1851 : }
1852 :
1853 : Assert(pdst - dstbuf <= count);
1854 :
1855 210784 : return pdst - dstbuf;
1856 : }
1857 :
1858 : /*
1859 : * Converts a "usable byte position" to XLogRecPtr. A usable byte position
1860 : * is the position starting from the beginning of WAL, excluding all WAL
1861 : * page headers.
1862 : */
1863 : static XLogRecPtr
1864 61533240 : XLogBytePosToRecPtr(uint64 bytepos)
1865 : {
1866 : uint64 fullsegs;
1867 : uint64 fullpages;
1868 : uint64 bytesleft;
1869 : uint32 seg_offset;
1870 : XLogRecPtr result;
1871 :
1872 61533240 : fullsegs = bytepos / UsableBytesInSegment;
1873 61533240 : bytesleft = bytepos % UsableBytesInSegment;
1874 :
1875 61533240 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1876 : {
1877 : /* fits on first page of segment */
1878 117876 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1879 : }
1880 : else
1881 : {
1882 : /* account for the first page on segment with long header */
1883 61415364 : seg_offset = XLOG_BLCKSZ;
1884 61415364 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1885 :
1886 61415364 : fullpages = bytesleft / UsableBytesInPage;
1887 61415364 : bytesleft = bytesleft % UsableBytesInPage;
1888 :
1889 61415364 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1890 : }
1891 :
1892 61533240 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1893 :
1894 61533240 : return result;
1895 : }
1896 :
1897 : /*
1898 : * Like XLogBytePosToRecPtr, but if the position is at a page boundary,
1899 : * returns a pointer to the beginning of the page (ie. before page header),
1900 : * not to where the first xlog record on that page would go to. This is used
1901 : * when converting a pointer to the end of a record.
1902 : */
1903 : static XLogRecPtr
1904 31706724 : XLogBytePosToEndRecPtr(uint64 bytepos)
1905 : {
1906 : uint64 fullsegs;
1907 : uint64 fullpages;
1908 : uint64 bytesleft;
1909 : uint32 seg_offset;
1910 : XLogRecPtr result;
1911 :
1912 31706724 : fullsegs = bytepos / UsableBytesInSegment;
1913 31706724 : bytesleft = bytepos % UsableBytesInSegment;
1914 :
1915 31706724 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1916 : {
1917 : /* fits on first page of segment */
1918 200996 : if (bytesleft == 0)
1919 139380 : seg_offset = 0;
1920 : else
1921 61616 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1922 : }
1923 : else
1924 : {
1925 : /* account for the first page on segment with long header */
1926 31505728 : seg_offset = XLOG_BLCKSZ;
1927 31505728 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1928 :
1929 31505728 : fullpages = bytesleft / UsableBytesInPage;
1930 31505728 : bytesleft = bytesleft % UsableBytesInPage;
1931 :
1932 31505728 : if (bytesleft == 0)
1933 30444 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
1934 : else
1935 31475284 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1936 : }
1937 :
1938 31706724 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1939 :
1940 31706724 : return result;
1941 : }
1942 :
1943 : /*
1944 : * Convert an XLogRecPtr to a "usable byte position".
1945 : */
1946 : static uint64
1947 5210 : XLogRecPtrToBytePos(XLogRecPtr ptr)
1948 : {
1949 : uint64 fullsegs;
1950 : uint32 fullpages;
1951 : uint32 offset;
1952 : uint64 result;
1953 :
1954 5210 : XLByteToSeg(ptr, fullsegs, wal_segment_size);
1955 :
1956 5210 : fullpages = (XLogSegmentOffset(ptr, wal_segment_size)) / XLOG_BLCKSZ;
1957 5210 : offset = ptr % XLOG_BLCKSZ;
1958 :
1959 5210 : if (fullpages == 0)
1960 : {
1961 2046 : result = fullsegs * UsableBytesInSegment;
1962 2046 : if (offset > 0)
1963 : {
1964 : Assert(offset >= SizeOfXLogLongPHD);
1965 518 : result += offset - SizeOfXLogLongPHD;
1966 : }
1967 : }
1968 : else
1969 : {
1970 3164 : result = fullsegs * UsableBytesInSegment +
1971 3164 : (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
1972 3164 : (fullpages - 1) * UsableBytesInPage; /* full pages */
1973 3164 : if (offset > 0)
1974 : {
1975 : Assert(offset >= SizeOfXLogShortPHD);
1976 3142 : result += offset - SizeOfXLogShortPHD;
1977 : }
1978 : }
1979 :
1980 5210 : return result;
1981 : }
1982 :
1983 : /*
1984 : * Initialize XLOG buffers, writing out old buffers if they still contain
1985 : * unwritten data, upto the page containing 'upto'. Or if 'opportunistic' is
1986 : * true, initialize as many pages as we can without having to write out
1987 : * unwritten data. Any new pages are initialized to zeros, with pages headers
1988 : * initialized properly.
1989 : */
1990 : static void
1991 4997730 : AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
1992 : {
1993 4997730 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1994 : int nextidx;
1995 : XLogRecPtr OldPageRqstPtr;
1996 : XLogwrtRqst WriteRqst;
1997 4997730 : XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
1998 : XLogRecPtr NewPageBeginPtr;
1999 : XLogPageHeader NewPage;
2000 4997730 : int npages pg_attribute_unused() = 0;
2001 :
2002 4997730 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2003 :
2004 : /*
2005 : * Now that we have the lock, check if someone initialized the page
2006 : * already.
2007 : */
2008 14615946 : while (upto >= XLogCtl->InitializedUpTo || opportunistic)
2009 : {
2010 9627868 : nextidx = XLogRecPtrToBufIdx(XLogCtl->InitializedUpTo);
2011 :
2012 : /*
2013 : * Get ending-offset of the buffer page we need to replace (this may
2014 : * be zero if the buffer hasn't been used yet). Fall through if it's
2015 : * already written out.
2016 : */
2017 9627868 : OldPageRqstPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]);
2018 9627868 : if (LogwrtResult.Write < OldPageRqstPtr)
2019 : {
2020 : /*
2021 : * Nope, got work to do. If we just want to pre-initialize as much
2022 : * as we can without flushing, give up now.
2023 : */
2024 4450416 : if (opportunistic)
2025 9652 : break;
2026 :
2027 : /* Advance shared memory write request position */
2028 4440764 : SpinLockAcquire(&XLogCtl->info_lck);
2029 4440764 : if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
2030 1366748 : XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
2031 4440764 : SpinLockRelease(&XLogCtl->info_lck);
2032 :
2033 : /*
2034 : * Acquire an up-to-date LogwrtResult value and see if we still
2035 : * need to write it or if someone else already did.
2036 : */
2037 4440764 : RefreshXLogWriteResult(LogwrtResult);
2038 4440764 : if (LogwrtResult.Write < OldPageRqstPtr)
2039 : {
2040 : /*
2041 : * Must acquire write lock. Release WALBufMappingLock first,
2042 : * to make sure that all insertions that we need to wait for
2043 : * can finish (up to this same position). Otherwise we risk
2044 : * deadlock.
2045 : */
2046 4412716 : LWLockRelease(WALBufMappingLock);
2047 :
2048 4412716 : WaitXLogInsertionsToFinish(OldPageRqstPtr);
2049 :
2050 4412716 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2051 :
2052 4412716 : RefreshXLogWriteResult(LogwrtResult);
2053 4412716 : if (LogwrtResult.Write >= OldPageRqstPtr)
2054 : {
2055 : /* OK, someone wrote it already */
2056 264114 : LWLockRelease(WALWriteLock);
2057 : }
2058 : else
2059 : {
2060 : /* Have to write it ourselves */
2061 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2062 4148602 : WriteRqst.Write = OldPageRqstPtr;
2063 4148602 : WriteRqst.Flush = 0;
2064 4148602 : XLogWrite(WriteRqst, tli, false);
2065 4148602 : LWLockRelease(WALWriteLock);
2066 4148602 : pgWalUsage.wal_buffers_full++;
2067 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_DONE();
2068 :
2069 : /*
2070 : * Required for the flush of pending stats WAL data, per
2071 : * update of pgWalUsage.
2072 : */
2073 4148602 : pgstat_report_fixed = true;
2074 : }
2075 : /* Re-acquire WALBufMappingLock and retry */
2076 4412716 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2077 4412716 : continue;
2078 : }
2079 : }
2080 :
2081 : /*
2082 : * Now the next buffer slot is free and we can set it up to be the
2083 : * next output page.
2084 : */
2085 5205500 : NewPageBeginPtr = XLogCtl->InitializedUpTo;
2086 5205500 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2087 :
2088 : Assert(XLogRecPtrToBufIdx(NewPageBeginPtr) == nextidx);
2089 :
2090 5205500 : NewPage = (XLogPageHeader) (XLogCtl->pages + nextidx * (Size) XLOG_BLCKSZ);
2091 :
2092 : /*
2093 : * Mark the xlblock with InvalidXLogRecPtr and issue a write barrier
2094 : * before initializing. Otherwise, the old page may be partially
2095 : * zeroed but look valid.
2096 : */
2097 5205500 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], InvalidXLogRecPtr);
2098 5205500 : pg_write_barrier();
2099 :
2100 : /*
2101 : * Be sure to re-zero the buffer so that bytes beyond what we've
2102 : * written will look like zeroes and not valid XLOG records...
2103 : */
2104 5205500 : MemSet(NewPage, 0, XLOG_BLCKSZ);
2105 :
2106 : /*
2107 : * Fill the new page's header
2108 : */
2109 5205500 : NewPage->xlp_magic = XLOG_PAGE_MAGIC;
2110 :
2111 : /* NewPage->xlp_info = 0; */ /* done by memset */
2112 5205500 : NewPage->xlp_tli = tli;
2113 5205500 : NewPage->xlp_pageaddr = NewPageBeginPtr;
2114 :
2115 : /* NewPage->xlp_rem_len = 0; */ /* done by memset */
2116 :
2117 : /*
2118 : * If online backup is not in progress, mark the header to indicate
2119 : * that WAL records beginning in this page have removable backup
2120 : * blocks. This allows the WAL archiver to know whether it is safe to
2121 : * compress archived WAL data by transforming full-block records into
2122 : * the non-full-block format. It is sufficient to record this at the
2123 : * page level because we force a page switch (in fact a segment
2124 : * switch) when starting a backup, so the flag will be off before any
2125 : * records can be written during the backup. At the end of a backup,
2126 : * the last page will be marked as all unsafe when perhaps only part
2127 : * is unsafe, but at worst the archiver would miss the opportunity to
2128 : * compress a few records.
2129 : */
2130 5205500 : if (Insert->runningBackups == 0)
2131 4922844 : NewPage->xlp_info |= XLP_BKP_REMOVABLE;
2132 :
2133 : /*
2134 : * If first page of an XLOG segment file, make it a long header.
2135 : */
2136 5205500 : if ((XLogSegmentOffset(NewPage->xlp_pageaddr, wal_segment_size)) == 0)
2137 : {
2138 3622 : XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2139 :
2140 3622 : NewLongPage->xlp_sysid = ControlFile->system_identifier;
2141 3622 : NewLongPage->xlp_seg_size = wal_segment_size;
2142 3622 : NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2143 3622 : NewPage->xlp_info |= XLP_LONG_HEADER;
2144 : }
2145 :
2146 : /*
2147 : * Make sure the initialization of the page becomes visible to others
2148 : * before the xlblocks update. GetXLogBuffer() reads xlblocks without
2149 : * holding a lock.
2150 : */
2151 5205500 : pg_write_barrier();
2152 :
2153 5205500 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], NewPageEndPtr);
2154 5205500 : XLogCtl->InitializedUpTo = NewPageEndPtr;
2155 :
2156 5205500 : npages++;
2157 : }
2158 4997730 : LWLockRelease(WALBufMappingLock);
2159 :
2160 : #ifdef WAL_DEBUG
2161 : if (XLOG_DEBUG && npages > 0)
2162 : {
2163 : elog(DEBUG1, "initialized %d pages, up to %X/%08X",
2164 : npages, LSN_FORMAT_ARGS(NewPageEndPtr));
2165 : }
2166 : #endif
2167 4997730 : }
2168 :
2169 : /*
2170 : * Calculate CheckPointSegments based on max_wal_size_mb and
2171 : * checkpoint_completion_target.
2172 : */
2173 : static void
2174 16416 : CalculateCheckpointSegments(void)
2175 : {
2176 : double target;
2177 :
2178 : /*-------
2179 : * Calculate the distance at which to trigger a checkpoint, to avoid
2180 : * exceeding max_wal_size_mb. This is based on two assumptions:
2181 : *
2182 : * a) we keep WAL for only one checkpoint cycle (prior to PG11 we kept
2183 : * WAL for two checkpoint cycles to allow us to recover from the
2184 : * secondary checkpoint if the first checkpoint failed, though we
2185 : * only did this on the primary anyway, not on standby. Keeping just
2186 : * one checkpoint simplifies processing and reduces disk space in
2187 : * many smaller databases.)
2188 : * b) during checkpoint, we consume checkpoint_completion_target *
2189 : * number of segments consumed between checkpoints.
2190 : *-------
2191 : */
2192 16416 : target = (double) ConvertToXSegs(max_wal_size_mb, wal_segment_size) /
2193 16416 : (1.0 + CheckPointCompletionTarget);
2194 :
2195 : /* round down */
2196 16416 : CheckPointSegments = (int) target;
2197 :
2198 16416 : if (CheckPointSegments < 1)
2199 20 : CheckPointSegments = 1;
2200 16416 : }
2201 :
2202 : void
2203 11968 : assign_max_wal_size(int newval, void *extra)
2204 : {
2205 11968 : max_wal_size_mb = newval;
2206 11968 : CalculateCheckpointSegments();
2207 11968 : }
2208 :
2209 : void
2210 2356 : assign_checkpoint_completion_target(double newval, void *extra)
2211 : {
2212 2356 : CheckPointCompletionTarget = newval;
2213 2356 : CalculateCheckpointSegments();
2214 2356 : }
2215 :
2216 : bool
2217 4552 : check_wal_segment_size(int *newval, void **extra, GucSource source)
2218 : {
2219 4552 : if (!IsValidWalSegSize(*newval))
2220 : {
2221 0 : GUC_check_errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.");
2222 0 : return false;
2223 : }
2224 :
2225 4552 : return true;
2226 : }
2227 :
2228 : /*
2229 : * At a checkpoint, how many WAL segments to recycle as preallocated future
2230 : * XLOG segments? Returns the highest segment that should be preallocated.
2231 : */
2232 : static XLogSegNo
2233 3586 : XLOGfileslop(XLogRecPtr lastredoptr)
2234 : {
2235 : XLogSegNo minSegNo;
2236 : XLogSegNo maxSegNo;
2237 : double distance;
2238 : XLogSegNo recycleSegNo;
2239 :
2240 : /*
2241 : * Calculate the segment numbers that min_wal_size_mb and max_wal_size_mb
2242 : * correspond to. Always recycle enough segments to meet the minimum, and
2243 : * remove enough segments to stay below the maximum.
2244 : */
2245 3586 : minSegNo = lastredoptr / wal_segment_size +
2246 3586 : ConvertToXSegs(min_wal_size_mb, wal_segment_size) - 1;
2247 3586 : maxSegNo = lastredoptr / wal_segment_size +
2248 3586 : ConvertToXSegs(max_wal_size_mb, wal_segment_size) - 1;
2249 :
2250 : /*
2251 : * Between those limits, recycle enough segments to get us through to the
2252 : * estimated end of next checkpoint.
2253 : *
2254 : * To estimate where the next checkpoint will finish, assume that the
2255 : * system runs steadily consuming CheckPointDistanceEstimate bytes between
2256 : * every checkpoint.
2257 : */
2258 3586 : distance = (1.0 + CheckPointCompletionTarget) * CheckPointDistanceEstimate;
2259 : /* add 10% for good measure. */
2260 3586 : distance *= 1.10;
2261 :
2262 3586 : recycleSegNo = (XLogSegNo) ceil(((double) lastredoptr + distance) /
2263 : wal_segment_size);
2264 :
2265 3586 : if (recycleSegNo < minSegNo)
2266 2528 : recycleSegNo = minSegNo;
2267 3586 : if (recycleSegNo > maxSegNo)
2268 810 : recycleSegNo = maxSegNo;
2269 :
2270 3586 : return recycleSegNo;
2271 : }
2272 :
2273 : /*
2274 : * Check whether we've consumed enough xlog space that a checkpoint is needed.
2275 : *
2276 : * new_segno indicates a log file that has just been filled up (or read
2277 : * during recovery). We measure the distance from RedoRecPtr to new_segno
2278 : * and see if that exceeds CheckPointSegments.
2279 : *
2280 : * Note: it is caller's responsibility that RedoRecPtr is up-to-date.
2281 : */
2282 : bool
2283 9590 : XLogCheckpointNeeded(XLogSegNo new_segno)
2284 : {
2285 : XLogSegNo old_segno;
2286 :
2287 9590 : XLByteToSeg(RedoRecPtr, old_segno, wal_segment_size);
2288 :
2289 9590 : if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2290 5980 : return true;
2291 3610 : return false;
2292 : }
2293 :
2294 : /*
2295 : * Write and/or fsync the log at least as far as WriteRqst indicates.
2296 : *
2297 : * If flexible == true, we don't have to write as far as WriteRqst, but
2298 : * may stop at any convenient boundary (such as a cache or logfile boundary).
2299 : * This option allows us to avoid uselessly issuing multiple writes when a
2300 : * single one would do.
2301 : *
2302 : * Must be called with WALWriteLock held. WaitXLogInsertionsToFinish(WriteRqst)
2303 : * must be called before grabbing the lock, to make sure the data is ready to
2304 : * write.
2305 : */
2306 : static void
2307 4424106 : XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible)
2308 : {
2309 : bool ispartialpage;
2310 : bool last_iteration;
2311 : bool finishing_seg;
2312 : int curridx;
2313 : int npages;
2314 : int startidx;
2315 : uint32 startoffset;
2316 :
2317 : /* We should always be inside a critical section here */
2318 : Assert(CritSectionCount > 0);
2319 :
2320 : /*
2321 : * Update local LogwrtResult (caller probably did this already, but...)
2322 : */
2323 4424106 : RefreshXLogWriteResult(LogwrtResult);
2324 :
2325 : /*
2326 : * Since successive pages in the xlog cache are consecutively allocated,
2327 : * we can usually gather multiple pages together and issue just one
2328 : * write() call. npages is the number of pages we have determined can be
2329 : * written together; startidx is the cache block index of the first one,
2330 : * and startoffset is the file offset at which it should go. The latter
2331 : * two variables are only valid when npages > 0, but we must initialize
2332 : * all of them to keep the compiler quiet.
2333 : */
2334 4424106 : npages = 0;
2335 4424106 : startidx = 0;
2336 4424106 : startoffset = 0;
2337 :
2338 : /*
2339 : * Within the loop, curridx is the cache block index of the page to
2340 : * consider writing. Begin at the buffer containing the next unwritten
2341 : * page, or last partially written page.
2342 : */
2343 4424106 : curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2344 :
2345 9514930 : while (LogwrtResult.Write < WriteRqst.Write)
2346 : {
2347 : /*
2348 : * Make sure we're not ahead of the insert process. This could happen
2349 : * if we're passed a bogus WriteRqst.Write that is past the end of the
2350 : * last page that's been initialized by AdvanceXLInsertBuffer.
2351 : */
2352 5356816 : XLogRecPtr EndPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[curridx]);
2353 :
2354 5356816 : if (LogwrtResult.Write >= EndPtr)
2355 0 : elog(PANIC, "xlog write request %X/%08X is past end of log %X/%08X",
2356 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2357 : LSN_FORMAT_ARGS(EndPtr));
2358 :
2359 : /* Advance LogwrtResult.Write to end of current buffer page */
2360 5356816 : LogwrtResult.Write = EndPtr;
2361 5356816 : ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2362 :
2363 5356816 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2364 : wal_segment_size))
2365 : {
2366 : /*
2367 : * Switch to new logfile segment. We cannot have any pending
2368 : * pages here (since we dump what we have at segment end).
2369 : */
2370 : Assert(npages == 0);
2371 27512 : if (openLogFile >= 0)
2372 12334 : XLogFileClose();
2373 27512 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2374 : wal_segment_size);
2375 27512 : openLogTLI = tli;
2376 :
2377 : /* create/use new log file */
2378 27512 : openLogFile = XLogFileInit(openLogSegNo, tli);
2379 27512 : ReserveExternalFD();
2380 : }
2381 :
2382 : /* Make sure we have the current logfile open */
2383 5356816 : if (openLogFile < 0)
2384 : {
2385 0 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2386 : wal_segment_size);
2387 0 : openLogTLI = tli;
2388 0 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2389 0 : ReserveExternalFD();
2390 : }
2391 :
2392 : /* Add current page to the set of pending pages-to-dump */
2393 5356816 : if (npages == 0)
2394 : {
2395 : /* first of group */
2396 4450494 : startidx = curridx;
2397 4450494 : startoffset = XLogSegmentOffset(LogwrtResult.Write - XLOG_BLCKSZ,
2398 : wal_segment_size);
2399 : }
2400 5356816 : npages++;
2401 :
2402 : /*
2403 : * Dump the set if this will be the last loop iteration, or if we are
2404 : * at the last page of the cache area (since the next page won't be
2405 : * contiguous in memory), or if we are at the end of the logfile
2406 : * segment.
2407 : */
2408 5356816 : last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2409 :
2410 10455290 : finishing_seg = !ispartialpage &&
2411 5098474 : (startoffset + npages * XLOG_BLCKSZ) >= wal_segment_size;
2412 :
2413 5356816 : if (last_iteration ||
2414 936472 : curridx == XLogCtl->XLogCacheBlck ||
2415 : finishing_seg)
2416 : {
2417 : char *from;
2418 : Size nbytes;
2419 : Size nleft;
2420 : ssize_t written;
2421 : instr_time start;
2422 :
2423 : /* OK to write the page(s) */
2424 4450494 : from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2425 4450494 : nbytes = npages * (Size) XLOG_BLCKSZ;
2426 4450494 : nleft = nbytes;
2427 : do
2428 : {
2429 4450494 : errno = 0;
2430 :
2431 : /*
2432 : * Measure I/O timing to write WAL data, for pg_stat_io.
2433 : */
2434 4450494 : start = pgstat_prepare_io_time(track_wal_io_timing);
2435 :
2436 4450494 : pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE);
2437 4450494 : written = pg_pwrite(openLogFile, from, nleft, startoffset);
2438 4450494 : pgstat_report_wait_end();
2439 :
2440 4450494 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL,
2441 : IOOP_WRITE, start, 1, written);
2442 :
2443 4450494 : if (written <= 0)
2444 : {
2445 : char xlogfname[MAXFNAMELEN];
2446 : int save_errno;
2447 :
2448 0 : if (errno == EINTR)
2449 0 : continue;
2450 :
2451 0 : save_errno = errno;
2452 0 : XLogFileName(xlogfname, tli, openLogSegNo,
2453 : wal_segment_size);
2454 0 : errno = save_errno;
2455 0 : ereport(PANIC,
2456 : (errcode_for_file_access(),
2457 : errmsg("could not write to log file \"%s\" at offset %u, length %zu: %m",
2458 : xlogfname, startoffset, nleft)));
2459 : }
2460 4450494 : nleft -= written;
2461 4450494 : from += written;
2462 4450494 : startoffset += written;
2463 4450494 : } while (nleft > 0);
2464 :
2465 4450494 : npages = 0;
2466 :
2467 : /*
2468 : * If we just wrote the whole last page of a logfile segment,
2469 : * fsync the segment immediately. This avoids having to go back
2470 : * and re-open prior segments when an fsync request comes along
2471 : * later. Doing it here ensures that one and only one backend will
2472 : * perform this fsync.
2473 : *
2474 : * This is also the right place to notify the Archiver that the
2475 : * segment is ready to copy to archival storage, and to update the
2476 : * timer for archive_timeout, and to signal for a checkpoint if
2477 : * too many logfile segments have been used since the last
2478 : * checkpoint.
2479 : */
2480 4450494 : if (finishing_seg)
2481 : {
2482 3854 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2483 :
2484 : /* signal that we need to wakeup walsenders later */
2485 3854 : WalSndWakeupRequest();
2486 :
2487 3854 : LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2488 :
2489 3854 : if (XLogArchivingActive())
2490 822 : XLogArchiveNotifySeg(openLogSegNo, tli);
2491 :
2492 3854 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2493 3854 : XLogCtl->lastSegSwitchLSN = LogwrtResult.Flush;
2494 :
2495 : /*
2496 : * Request a checkpoint if we've consumed too much xlog since
2497 : * the last one. For speed, we first check using the local
2498 : * copy of RedoRecPtr, which might be out of date; if it looks
2499 : * like a checkpoint is needed, forcibly update RedoRecPtr and
2500 : * recheck.
2501 : */
2502 3854 : if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2503 : {
2504 516 : (void) GetRedoRecPtr();
2505 516 : if (XLogCheckpointNeeded(openLogSegNo))
2506 420 : RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2507 : }
2508 : }
2509 : }
2510 :
2511 5356816 : if (ispartialpage)
2512 : {
2513 : /* Only asked to write a partial page */
2514 258342 : LogwrtResult.Write = WriteRqst.Write;
2515 258342 : break;
2516 : }
2517 5098474 : curridx = NextBufIdx(curridx);
2518 :
2519 : /* If flexible, break out of loop as soon as we wrote something */
2520 5098474 : if (flexible && npages == 0)
2521 7650 : break;
2522 : }
2523 :
2524 : Assert(npages == 0);
2525 :
2526 : /*
2527 : * If asked to flush, do so
2528 : */
2529 4424106 : if (LogwrtResult.Flush < WriteRqst.Flush &&
2530 273992 : LogwrtResult.Flush < LogwrtResult.Write)
2531 : {
2532 : /*
2533 : * Could get here without iterating above loop, in which case we might
2534 : * have no open file or the wrong one. However, we do not need to
2535 : * fsync more than one file.
2536 : */
2537 273874 : if (wal_sync_method != WAL_SYNC_METHOD_OPEN &&
2538 273874 : wal_sync_method != WAL_SYNC_METHOD_OPEN_DSYNC)
2539 : {
2540 273874 : if (openLogFile >= 0 &&
2541 273842 : !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2542 : wal_segment_size))
2543 260 : XLogFileClose();
2544 273874 : if (openLogFile < 0)
2545 : {
2546 292 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2547 : wal_segment_size);
2548 292 : openLogTLI = tli;
2549 292 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2550 292 : ReserveExternalFD();
2551 : }
2552 :
2553 273874 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2554 : }
2555 :
2556 : /* signal that we need to wakeup walsenders later */
2557 273874 : WalSndWakeupRequest();
2558 :
2559 273874 : LogwrtResult.Flush = LogwrtResult.Write;
2560 : }
2561 :
2562 : /*
2563 : * Update shared-memory status
2564 : *
2565 : * We make sure that the shared 'request' values do not fall behind the
2566 : * 'result' values. This is not absolutely essential, but it saves some
2567 : * code in a couple of places.
2568 : */
2569 4424106 : SpinLockAcquire(&XLogCtl->info_lck);
2570 4424106 : if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
2571 239686 : XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
2572 4424106 : if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
2573 276960 : XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
2574 4424106 : SpinLockRelease(&XLogCtl->info_lck);
2575 :
2576 : /*
2577 : * We write Write first, bar, then Flush. When reading, the opposite must
2578 : * be done (with a matching barrier in between), so that we always see a
2579 : * Flush value that trails behind the Write value seen.
2580 : */
2581 4424106 : pg_atomic_write_u64(&XLogCtl->logWriteResult, LogwrtResult.Write);
2582 4424106 : pg_write_barrier();
2583 4424106 : pg_atomic_write_u64(&XLogCtl->logFlushResult, LogwrtResult.Flush);
2584 :
2585 : #ifdef USE_ASSERT_CHECKING
2586 : {
2587 : XLogRecPtr Flush;
2588 : XLogRecPtr Write;
2589 : XLogRecPtr Insert;
2590 :
2591 : Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult);
2592 : pg_read_barrier();
2593 : Write = pg_atomic_read_u64(&XLogCtl->logWriteResult);
2594 : pg_read_barrier();
2595 : Insert = pg_atomic_read_u64(&XLogCtl->logInsertResult);
2596 :
2597 : /* WAL written to disk is always ahead of WAL flushed */
2598 : Assert(Write >= Flush);
2599 :
2600 : /* WAL inserted to buffers is always ahead of WAL written */
2601 : Assert(Insert >= Write);
2602 : }
2603 : #endif
2604 4424106 : }
2605 :
2606 : /*
2607 : * Record the LSN for an asynchronous transaction commit/abort
2608 : * and nudge the WALWriter if there is work for it to do.
2609 : * (This should not be called for synchronous commits.)
2610 : */
2611 : void
2612 99552 : XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2613 : {
2614 99552 : XLogRecPtr WriteRqstPtr = asyncXactLSN;
2615 : bool sleeping;
2616 99552 : bool wakeup = false;
2617 : XLogRecPtr prevAsyncXactLSN;
2618 :
2619 99552 : SpinLockAcquire(&XLogCtl->info_lck);
2620 99552 : sleeping = XLogCtl->WalWriterSleeping;
2621 99552 : prevAsyncXactLSN = XLogCtl->asyncXactLSN;
2622 99552 : if (XLogCtl->asyncXactLSN < asyncXactLSN)
2623 98528 : XLogCtl->asyncXactLSN = asyncXactLSN;
2624 99552 : SpinLockRelease(&XLogCtl->info_lck);
2625 :
2626 : /*
2627 : * If somebody else already called this function with a more aggressive
2628 : * LSN, they will have done what we needed (and perhaps more).
2629 : */
2630 99552 : if (asyncXactLSN <= prevAsyncXactLSN)
2631 1024 : return;
2632 :
2633 : /*
2634 : * If the WALWriter is sleeping, kick it to make it come out of low-power
2635 : * mode, so that this async commit will reach disk within the expected
2636 : * amount of time. Otherwise, determine whether it has enough WAL
2637 : * available to flush, the same way that XLogBackgroundFlush() does.
2638 : */
2639 98528 : if (sleeping)
2640 44 : wakeup = true;
2641 : else
2642 : {
2643 : int flushblocks;
2644 :
2645 98484 : RefreshXLogWriteResult(LogwrtResult);
2646 :
2647 98484 : flushblocks =
2648 98484 : WriteRqstPtr / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
2649 :
2650 98484 : if (WalWriterFlushAfter == 0 || flushblocks >= WalWriterFlushAfter)
2651 7566 : wakeup = true;
2652 : }
2653 :
2654 98528 : if (wakeup)
2655 : {
2656 7610 : volatile PROC_HDR *procglobal = ProcGlobal;
2657 7610 : ProcNumber walwriterProc = procglobal->walwriterProc;
2658 :
2659 7610 : if (walwriterProc != INVALID_PROC_NUMBER)
2660 398 : SetLatch(&GetPGProcByNumber(walwriterProc)->procLatch);
2661 : }
2662 : }
2663 :
2664 : /*
2665 : * Record the LSN up to which we can remove WAL because it's not required by
2666 : * any replication slot.
2667 : */
2668 : void
2669 65322 : XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn)
2670 : {
2671 65322 : SpinLockAcquire(&XLogCtl->info_lck);
2672 65322 : XLogCtl->replicationSlotMinLSN = lsn;
2673 65322 : SpinLockRelease(&XLogCtl->info_lck);
2674 65322 : }
2675 :
2676 :
2677 : /*
2678 : * Return the oldest LSN we must retain to satisfy the needs of some
2679 : * replication slot.
2680 : */
2681 : XLogRecPtr
2682 4850 : XLogGetReplicationSlotMinimumLSN(void)
2683 : {
2684 : XLogRecPtr retval;
2685 :
2686 4850 : SpinLockAcquire(&XLogCtl->info_lck);
2687 4850 : retval = XLogCtl->replicationSlotMinLSN;
2688 4850 : SpinLockRelease(&XLogCtl->info_lck);
2689 :
2690 4850 : return retval;
2691 : }
2692 :
2693 : /*
2694 : * Advance minRecoveryPoint in control file.
2695 : *
2696 : * If we crash during recovery, we must reach this point again before the
2697 : * database is consistent.
2698 : *
2699 : * If 'force' is true, 'lsn' argument is ignored. Otherwise, minRecoveryPoint
2700 : * is only updated if it's not already greater than or equal to 'lsn'.
2701 : */
2702 : static void
2703 224028 : UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2704 : {
2705 : /* Quick check using our local copy of the variable */
2706 224028 : if (!updateMinRecoveryPoint || (!force && lsn <= LocalMinRecoveryPoint))
2707 209868 : return;
2708 :
2709 : /*
2710 : * An invalid minRecoveryPoint means that we need to recover all the WAL,
2711 : * i.e., we're doing crash recovery. We never modify the control file's
2712 : * value in that case, so we can short-circuit future checks here too. The
2713 : * local values of minRecoveryPoint and minRecoveryPointTLI should not be
2714 : * updated until crash recovery finishes. We only do this for the startup
2715 : * process as it should not update its own reference of minRecoveryPoint
2716 : * until it has finished crash recovery to make sure that all WAL
2717 : * available is replayed in this case. This also saves from extra locks
2718 : * taken on the control file from the startup process.
2719 : */
2720 14160 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint) && InRecovery)
2721 : {
2722 62 : updateMinRecoveryPoint = false;
2723 62 : return;
2724 : }
2725 :
2726 14098 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2727 :
2728 : /* update local copy */
2729 14098 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
2730 14098 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2731 :
2732 14098 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
2733 6 : updateMinRecoveryPoint = false;
2734 14092 : else if (force || LocalMinRecoveryPoint < lsn)
2735 : {
2736 : XLogRecPtr newMinRecoveryPoint;
2737 : TimeLineID newMinRecoveryPointTLI;
2738 :
2739 : /*
2740 : * To avoid having to update the control file too often, we update it
2741 : * all the way to the last record being replayed, even though 'lsn'
2742 : * would suffice for correctness. This also allows the 'force' case
2743 : * to not need a valid 'lsn' value.
2744 : *
2745 : * Another important reason for doing it this way is that the passed
2746 : * 'lsn' value could be bogus, i.e., past the end of available WAL, if
2747 : * the caller got it from a corrupted heap page. Accepting such a
2748 : * value as the min recovery point would prevent us from coming up at
2749 : * all. Instead, we just log a warning and continue with recovery.
2750 : * (See also the comments about corrupt LSNs in XLogFlush.)
2751 : */
2752 11230 : newMinRecoveryPoint = GetCurrentReplayRecPtr(&newMinRecoveryPointTLI);
2753 11230 : if (!force && newMinRecoveryPoint < lsn)
2754 0 : elog(WARNING,
2755 : "xlog min recovery request %X/%08X is past current point %X/%08X",
2756 : LSN_FORMAT_ARGS(lsn), LSN_FORMAT_ARGS(newMinRecoveryPoint));
2757 :
2758 : /* update control file */
2759 11230 : if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2760 : {
2761 10532 : ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2762 10532 : ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2763 10532 : UpdateControlFile();
2764 10532 : LocalMinRecoveryPoint = newMinRecoveryPoint;
2765 10532 : LocalMinRecoveryPointTLI = newMinRecoveryPointTLI;
2766 :
2767 10532 : ereport(DEBUG2,
2768 : errmsg_internal("updated min recovery point to %X/%08X on timeline %u",
2769 : LSN_FORMAT_ARGS(newMinRecoveryPoint),
2770 : newMinRecoveryPointTLI));
2771 : }
2772 : }
2773 14098 : LWLockRelease(ControlFileLock);
2774 : }
2775 :
2776 : /*
2777 : * Ensure that all XLOG data through the given position is flushed to disk.
2778 : *
2779 : * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
2780 : * already held, and we try to avoid acquiring it if possible.
2781 : */
2782 : void
2783 1416870 : XLogFlush(XLogRecPtr record)
2784 : {
2785 : XLogRecPtr WriteRqstPtr;
2786 : XLogwrtRqst WriteRqst;
2787 1416870 : TimeLineID insertTLI = XLogCtl->InsertTimeLineID;
2788 :
2789 : /*
2790 : * During REDO, we are reading not writing WAL. Therefore, instead of
2791 : * trying to flush the WAL, we should update minRecoveryPoint instead. We
2792 : * test XLogInsertAllowed(), not InRecovery, because we need checkpointer
2793 : * to act this way too, and because when it tries to write the
2794 : * end-of-recovery checkpoint, it should indeed flush.
2795 : */
2796 1416870 : if (!XLogInsertAllowed())
2797 : {
2798 223082 : UpdateMinRecoveryPoint(record, false);
2799 1124740 : return;
2800 : }
2801 :
2802 : /* Quick exit if already known flushed */
2803 1193788 : if (record <= LogwrtResult.Flush)
2804 901658 : return;
2805 :
2806 : #ifdef WAL_DEBUG
2807 : if (XLOG_DEBUG)
2808 : elog(LOG, "xlog flush request %X/%08X; write %X/%08X; flush %X/%08X",
2809 : LSN_FORMAT_ARGS(record),
2810 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2811 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2812 : #endif
2813 :
2814 292130 : START_CRIT_SECTION();
2815 :
2816 : /*
2817 : * Since fsync is usually a horribly expensive operation, we try to
2818 : * piggyback as much data as we can on each fsync: if we see any more data
2819 : * entered into the xlog buffer, we'll write and fsync that too, so that
2820 : * the final value of LogwrtResult.Flush is as large as possible. This
2821 : * gives us some chance of avoiding another fsync immediately after.
2822 : */
2823 :
2824 : /* initialize to given target; may increase below */
2825 292130 : WriteRqstPtr = record;
2826 :
2827 : /*
2828 : * Now wait until we get the write lock, or someone else does the flush
2829 : * for us.
2830 : */
2831 : for (;;)
2832 6316 : {
2833 : XLogRecPtr insertpos;
2834 :
2835 : /* done already? */
2836 298446 : RefreshXLogWriteResult(LogwrtResult);
2837 298446 : if (record <= LogwrtResult.Flush)
2838 20716 : break;
2839 :
2840 : /*
2841 : * Before actually performing the write, wait for all in-flight
2842 : * insertions to the pages we're about to write to finish.
2843 : */
2844 277730 : SpinLockAcquire(&XLogCtl->info_lck);
2845 277730 : if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
2846 17776 : WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
2847 277730 : SpinLockRelease(&XLogCtl->info_lck);
2848 277730 : insertpos = WaitXLogInsertionsToFinish(WriteRqstPtr);
2849 :
2850 : /*
2851 : * Try to get the write lock. If we can't get it immediately, wait
2852 : * until it's released, and recheck if we still need to do the flush
2853 : * or if the backend that held the lock did it for us already. This
2854 : * helps to maintain a good rate of group committing when the system
2855 : * is bottlenecked by the speed of fsyncing.
2856 : */
2857 277730 : if (!LWLockAcquireOrWait(WALWriteLock, LW_EXCLUSIVE))
2858 : {
2859 : /*
2860 : * The lock is now free, but we didn't acquire it yet. Before we
2861 : * do, loop back to check if someone else flushed the record for
2862 : * us already.
2863 : */
2864 6316 : continue;
2865 : }
2866 :
2867 : /* Got the lock; recheck whether request is satisfied */
2868 271414 : RefreshXLogWriteResult(LogwrtResult);
2869 271414 : if (record <= LogwrtResult.Flush)
2870 : {
2871 5324 : LWLockRelease(WALWriteLock);
2872 5324 : break;
2873 : }
2874 :
2875 : /*
2876 : * Sleep before flush! By adding a delay here, we may give further
2877 : * backends the opportunity to join the backlog of group commit
2878 : * followers; this can significantly improve transaction throughput,
2879 : * at the risk of increasing transaction latency.
2880 : *
2881 : * We do not sleep if enableFsync is not turned on, nor if there are
2882 : * fewer than CommitSiblings other backends with active transactions.
2883 : */
2884 266090 : if (CommitDelay > 0 && enableFsync &&
2885 0 : MinimumActiveBackends(CommitSiblings))
2886 : {
2887 0 : pgstat_report_wait_start(WAIT_EVENT_COMMIT_DELAY);
2888 0 : pg_usleep(CommitDelay);
2889 0 : pgstat_report_wait_end();
2890 :
2891 : /*
2892 : * Re-check how far we can now flush the WAL. It's generally not
2893 : * safe to call WaitXLogInsertionsToFinish while holding
2894 : * WALWriteLock, because an in-progress insertion might need to
2895 : * also grab WALWriteLock to make progress. But we know that all
2896 : * the insertions up to insertpos have already finished, because
2897 : * that's what the earlier WaitXLogInsertionsToFinish() returned.
2898 : * We're only calling it again to allow insertpos to be moved
2899 : * further forward, not to actually wait for anyone.
2900 : */
2901 0 : insertpos = WaitXLogInsertionsToFinish(insertpos);
2902 : }
2903 :
2904 : /* try to write/flush later additions to XLOG as well */
2905 266090 : WriteRqst.Write = insertpos;
2906 266090 : WriteRqst.Flush = insertpos;
2907 :
2908 266090 : XLogWrite(WriteRqst, insertTLI, false);
2909 :
2910 266090 : LWLockRelease(WALWriteLock);
2911 : /* done */
2912 266090 : break;
2913 : }
2914 :
2915 292130 : END_CRIT_SECTION();
2916 :
2917 : /* wake up walsenders now that we've released heavily contended locks */
2918 292130 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
2919 :
2920 : /*
2921 : * If we flushed an LSN that someone was waiting for, notify the waiters.
2922 : */
2923 584260 : if (waitLSNState &&
2924 292130 : (LogwrtResult.Flush >=
2925 292130 : pg_atomic_read_u64(&waitLSNState->minWaitedLSN[WAIT_LSN_TYPE_PRIMARY_FLUSH])))
2926 0 : WaitLSNWakeup(WAIT_LSN_TYPE_PRIMARY_FLUSH, LogwrtResult.Flush);
2927 :
2928 : /*
2929 : * If we still haven't flushed to the request point then we have a
2930 : * problem; most likely, the requested flush point is past end of XLOG.
2931 : * This has been seen to occur when a disk page has a corrupted LSN.
2932 : *
2933 : * Formerly we treated this as a PANIC condition, but that hurts the
2934 : * system's robustness rather than helping it: we do not want to take down
2935 : * the whole system due to corruption on one data page. In particular, if
2936 : * the bad page is encountered again during recovery then we would be
2937 : * unable to restart the database at all! (This scenario actually
2938 : * happened in the field several times with 7.1 releases.) As of 8.4, bad
2939 : * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
2940 : * the only time we can reach here during recovery is while flushing the
2941 : * end-of-recovery checkpoint record, and we don't expect that to have a
2942 : * bad LSN.
2943 : *
2944 : * Note that for calls from xact.c, the ERROR will be promoted to PANIC
2945 : * since xact.c calls this routine inside a critical section. However,
2946 : * calls from bufmgr.c are not within critical sections and so we will not
2947 : * force a restart for a bad LSN on a data page.
2948 : */
2949 292130 : if (LogwrtResult.Flush < record)
2950 0 : elog(ERROR,
2951 : "xlog flush request %X/%08X is not satisfied --- flushed only to %X/%08X",
2952 : LSN_FORMAT_ARGS(record),
2953 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2954 :
2955 : /*
2956 : * Cross-check XLogNeedsFlush(). Some of the checks of XLogFlush() and
2957 : * XLogNeedsFlush() are duplicated, and this assertion ensures that these
2958 : * remain consistent.
2959 : */
2960 : Assert(!XLogNeedsFlush(record));
2961 : }
2962 :
2963 : /*
2964 : * Write & flush xlog, but without specifying exactly where to.
2965 : *
2966 : * We normally write only completed blocks; but if there is nothing to do on
2967 : * that basis, we check for unwritten async commits in the current incomplete
2968 : * block, and write through the latest one of those. Thus, if async commits
2969 : * are not being used, we will write complete blocks only.
2970 : *
2971 : * If, based on the above, there's anything to write we do so immediately. But
2972 : * to avoid calling fsync, fdatasync et. al. at a rate that'd impact
2973 : * concurrent IO, we only flush WAL every wal_writer_delay ms, or if there's
2974 : * more than wal_writer_flush_after unflushed blocks.
2975 : *
2976 : * We can guarantee that async commits reach disk after at most three
2977 : * wal_writer_delay cycles. (When flushing complete blocks, we allow XLogWrite
2978 : * to write "flexibly", meaning it can stop at the end of the buffer ring;
2979 : * this makes a difference only with very high load or long wal_writer_delay,
2980 : * but imposes one extra cycle for the worst case for async commits.)
2981 : *
2982 : * This routine is invoked periodically by the background walwriter process.
2983 : *
2984 : * Returns true if there was any work to do, even if we skipped flushing due
2985 : * to wal_writer_delay/wal_writer_flush_after.
2986 : */
2987 : bool
2988 37278 : XLogBackgroundFlush(void)
2989 : {
2990 : XLogwrtRqst WriteRqst;
2991 37278 : bool flexible = true;
2992 : static TimestampTz lastflush;
2993 : TimestampTz now;
2994 : int flushblocks;
2995 : TimeLineID insertTLI;
2996 :
2997 : /* XLOG doesn't need flushing during recovery */
2998 37278 : if (RecoveryInProgress())
2999 1780 : return false;
3000 :
3001 : /*
3002 : * Since we're not in recovery, InsertTimeLineID is set and can't change,
3003 : * so we can read it without a lock.
3004 : */
3005 35498 : insertTLI = XLogCtl->InsertTimeLineID;
3006 :
3007 : /* read updated LogwrtRqst */
3008 35498 : SpinLockAcquire(&XLogCtl->info_lck);
3009 35498 : WriteRqst = XLogCtl->LogwrtRqst;
3010 35498 : SpinLockRelease(&XLogCtl->info_lck);
3011 :
3012 : /* back off to last completed page boundary */
3013 35498 : WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
3014 :
3015 : /* if we have already flushed that far, consider async commit records */
3016 35498 : RefreshXLogWriteResult(LogwrtResult);
3017 35498 : if (WriteRqst.Write <= LogwrtResult.Flush)
3018 : {
3019 27490 : SpinLockAcquire(&XLogCtl->info_lck);
3020 27490 : WriteRqst.Write = XLogCtl->asyncXactLSN;
3021 27490 : SpinLockRelease(&XLogCtl->info_lck);
3022 27490 : flexible = false; /* ensure it all gets written */
3023 : }
3024 :
3025 : /*
3026 : * If already known flushed, we're done. Just need to check if we are
3027 : * holding an open file handle to a logfile that's no longer in use,
3028 : * preventing the file from being deleted.
3029 : */
3030 35498 : if (WriteRqst.Write <= LogwrtResult.Flush)
3031 : {
3032 25846 : if (openLogFile >= 0)
3033 : {
3034 19642 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
3035 : wal_segment_size))
3036 : {
3037 324 : XLogFileClose();
3038 : }
3039 : }
3040 25846 : return false;
3041 : }
3042 :
3043 : /*
3044 : * Determine how far to flush WAL, based on the wal_writer_delay and
3045 : * wal_writer_flush_after GUCs.
3046 : *
3047 : * Note that XLogSetAsyncXactLSN() performs similar calculation based on
3048 : * wal_writer_flush_after, to decide when to wake us up. Make sure the
3049 : * logic is the same in both places if you change this.
3050 : */
3051 9652 : now = GetCurrentTimestamp();
3052 9652 : flushblocks =
3053 9652 : WriteRqst.Write / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
3054 :
3055 9652 : if (WalWriterFlushAfter == 0 || lastflush == 0)
3056 : {
3057 : /* first call, or block based limits disabled */
3058 542 : WriteRqst.Flush = WriteRqst.Write;
3059 542 : lastflush = now;
3060 : }
3061 9110 : else if (TimestampDifferenceExceeds(lastflush, now, WalWriterDelay))
3062 : {
3063 : /*
3064 : * Flush the writes at least every WalWriterDelay ms. This is
3065 : * important to bound the amount of time it takes for an asynchronous
3066 : * commit to hit disk.
3067 : */
3068 8828 : WriteRqst.Flush = WriteRqst.Write;
3069 8828 : lastflush = now;
3070 : }
3071 282 : else if (flushblocks >= WalWriterFlushAfter)
3072 : {
3073 : /* exceeded wal_writer_flush_after blocks, flush */
3074 260 : WriteRqst.Flush = WriteRqst.Write;
3075 260 : lastflush = now;
3076 : }
3077 : else
3078 : {
3079 : /* no flushing, this time round */
3080 22 : WriteRqst.Flush = 0;
3081 : }
3082 :
3083 : #ifdef WAL_DEBUG
3084 : if (XLOG_DEBUG)
3085 : elog(LOG, "xlog bg flush request write %X/%08X; flush: %X/%08X, current is write %X/%08X; flush %X/%08X",
3086 : LSN_FORMAT_ARGS(WriteRqst.Write),
3087 : LSN_FORMAT_ARGS(WriteRqst.Flush),
3088 : LSN_FORMAT_ARGS(LogwrtResult.Write),
3089 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
3090 : #endif
3091 :
3092 9652 : START_CRIT_SECTION();
3093 :
3094 : /* now wait for any in-progress insertions to finish and get write lock */
3095 9652 : WaitXLogInsertionsToFinish(WriteRqst.Write);
3096 9652 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3097 9652 : RefreshXLogWriteResult(LogwrtResult);
3098 9652 : if (WriteRqst.Write > LogwrtResult.Write ||
3099 394 : WriteRqst.Flush > LogwrtResult.Flush)
3100 : {
3101 9414 : XLogWrite(WriteRqst, insertTLI, flexible);
3102 : }
3103 9652 : LWLockRelease(WALWriteLock);
3104 :
3105 9652 : END_CRIT_SECTION();
3106 :
3107 : /* wake up walsenders now that we've released heavily contended locks */
3108 9652 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3109 :
3110 : /*
3111 : * If we flushed an LSN that someone was waiting for, notify the waiters.
3112 : */
3113 19304 : if (waitLSNState &&
3114 9652 : (LogwrtResult.Flush >=
3115 9652 : pg_atomic_read_u64(&waitLSNState->minWaitedLSN[WAIT_LSN_TYPE_PRIMARY_FLUSH])))
3116 0 : WaitLSNWakeup(WAIT_LSN_TYPE_PRIMARY_FLUSH, LogwrtResult.Flush);
3117 :
3118 : /*
3119 : * Great, done. To take some work off the critical path, try to initialize
3120 : * as many of the no-longer-needed WAL buffers for future use as we can.
3121 : */
3122 9652 : AdvanceXLInsertBuffer(InvalidXLogRecPtr, insertTLI, true);
3123 :
3124 : /*
3125 : * If we determined that we need to write data, but somebody else
3126 : * wrote/flushed already, it should be considered as being active, to
3127 : * avoid hibernating too early.
3128 : */
3129 9652 : return true;
3130 : }
3131 :
3132 : /*
3133 : * Test whether XLOG data has been flushed up to (at least) the given
3134 : * position, or whether the minimum recovery point has been updated past
3135 : * the given position.
3136 : *
3137 : * Returns true if a flush is still needed, or if the minimum recovery point
3138 : * must be updated.
3139 : *
3140 : * It is possible that someone else is already in the process of flushing
3141 : * that far, or has updated the minimum recovery point up to the given
3142 : * position.
3143 : */
3144 : bool
3145 17730872 : XLogNeedsFlush(XLogRecPtr record)
3146 : {
3147 : /*
3148 : * During recovery, we don't flush WAL but update minRecoveryPoint
3149 : * instead. So "needs flush" is taken to mean whether minRecoveryPoint
3150 : * would need to be updated.
3151 : *
3152 : * Using XLogInsertAllowed() rather than RecoveryInProgress() matters for
3153 : * the case of an end-of-recovery checkpoint, where WAL data is flushed.
3154 : * This check should be consistent with the one in XLogFlush().
3155 : */
3156 17730872 : if (!XLogInsertAllowed())
3157 : {
3158 : /* Quick exit if already known to be updated or cannot be updated */
3159 1212172 : if (!updateMinRecoveryPoint || record <= LocalMinRecoveryPoint)
3160 1190204 : return false;
3161 :
3162 : /*
3163 : * An invalid minRecoveryPoint means that we need to recover all the
3164 : * WAL, i.e., we're doing crash recovery. We never modify the control
3165 : * file's value in that case, so we can short-circuit future checks
3166 : * here too. This triggers a quick exit path for the startup process,
3167 : * which cannot update its local copy of minRecoveryPoint as long as
3168 : * it has not replayed all WAL available when doing crash recovery.
3169 : */
3170 21968 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint) && InRecovery)
3171 : {
3172 0 : updateMinRecoveryPoint = false;
3173 0 : return false;
3174 : }
3175 :
3176 : /*
3177 : * Update local copy of minRecoveryPoint. But if the lock is busy,
3178 : * just return a conservative guess.
3179 : */
3180 21968 : if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3181 0 : return true;
3182 21968 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
3183 21968 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3184 21968 : LWLockRelease(ControlFileLock);
3185 :
3186 : /*
3187 : * Check minRecoveryPoint for any other process than the startup
3188 : * process doing crash recovery, which should not update the control
3189 : * file value if crash recovery is still running.
3190 : */
3191 21968 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
3192 0 : updateMinRecoveryPoint = false;
3193 :
3194 : /* check again */
3195 21968 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3196 150 : return false;
3197 : else
3198 21818 : return true;
3199 : }
3200 :
3201 : /* Quick exit if already known flushed */
3202 16518700 : if (record <= LogwrtResult.Flush)
3203 16024926 : return false;
3204 :
3205 : /* read LogwrtResult and update local state */
3206 493774 : RefreshXLogWriteResult(LogwrtResult);
3207 :
3208 : /* check again */
3209 493774 : if (record <= LogwrtResult.Flush)
3210 6374 : return false;
3211 :
3212 487400 : return true;
3213 : }
3214 :
3215 : /*
3216 : * Try to make a given XLOG file segment exist.
3217 : *
3218 : * logsegno: identify segment.
3219 : *
3220 : * *added: on return, true if this call raised the number of extant segments.
3221 : *
3222 : * path: on return, this char[MAXPGPATH] has the path to the logsegno file.
3223 : *
3224 : * Returns -1 or FD of opened file. A -1 here is not an error; a caller
3225 : * wanting an open segment should attempt to open "path", which usually will
3226 : * succeed. (This is weird, but it's efficient for the callers.)
3227 : */
3228 : static int
3229 29836 : XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,
3230 : bool *added, char *path)
3231 : {
3232 : char tmppath[MAXPGPATH];
3233 : XLogSegNo installed_segno;
3234 : XLogSegNo max_segno;
3235 : int fd;
3236 : int save_errno;
3237 29836 : int open_flags = O_RDWR | O_CREAT | O_EXCL | PG_BINARY;
3238 : instr_time io_start;
3239 :
3240 : Assert(logtli != 0);
3241 :
3242 29836 : XLogFilePath(path, logtli, logsegno, wal_segment_size);
3243 :
3244 : /*
3245 : * Try to use existent file (checkpoint maker may have created it already)
3246 : */
3247 29836 : *added = false;
3248 29836 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3249 29836 : get_sync_bit(wal_sync_method));
3250 29836 : if (fd < 0)
3251 : {
3252 3168 : if (errno != ENOENT)
3253 0 : ereport(ERROR,
3254 : (errcode_for_file_access(),
3255 : errmsg("could not open file \"%s\": %m", path)));
3256 : }
3257 : else
3258 26668 : return fd;
3259 :
3260 : /*
3261 : * Initialize an empty (all zeroes) segment. NOTE: it is possible that
3262 : * another process is doing the same thing. If so, we will end up
3263 : * pre-creating an extra log segment. That seems OK, and better than
3264 : * holding the lock throughout this lengthy process.
3265 : */
3266 3168 : elog(DEBUG2, "creating and filling new WAL file");
3267 :
3268 3168 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3269 :
3270 3168 : unlink(tmppath);
3271 :
3272 3168 : if (io_direct_flags & IO_DIRECT_WAL_INIT)
3273 0 : open_flags |= PG_O_DIRECT;
3274 :
3275 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3276 3168 : fd = BasicOpenFile(tmppath, open_flags);
3277 3168 : if (fd < 0)
3278 0 : ereport(ERROR,
3279 : (errcode_for_file_access(),
3280 : errmsg("could not create file \"%s\": %m", tmppath)));
3281 :
3282 : /* Measure I/O timing when initializing segment */
3283 3168 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3284 :
3285 3168 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE);
3286 3168 : save_errno = 0;
3287 3168 : if (wal_init_zero)
3288 : {
3289 : ssize_t rc;
3290 :
3291 : /*
3292 : * Zero-fill the file. With this setting, we do this the hard way to
3293 : * ensure that all the file space has really been allocated. On
3294 : * platforms that allow "holes" in files, just seeking to the end
3295 : * doesn't allocate intermediate space. This way, we know that we
3296 : * have all the space and (after the fsync below) that all the
3297 : * indirect blocks are down on disk. Therefore, fdatasync(2) or
3298 : * O_DSYNC will be sufficient to sync future writes to the log file.
3299 : */
3300 3168 : rc = pg_pwrite_zeros(fd, wal_segment_size, 0);
3301 :
3302 3168 : if (rc < 0)
3303 0 : save_errno = errno;
3304 : }
3305 : else
3306 : {
3307 : /*
3308 : * Otherwise, seeking to the end and writing a solitary byte is
3309 : * enough.
3310 : */
3311 0 : errno = 0;
3312 0 : if (pg_pwrite(fd, "\0", 1, wal_segment_size - 1) != 1)
3313 : {
3314 : /* if write didn't set errno, assume no disk space */
3315 0 : save_errno = errno ? errno : ENOSPC;
3316 : }
3317 : }
3318 3168 : pgstat_report_wait_end();
3319 :
3320 : /*
3321 : * A full segment worth of data is written when using wal_init_zero. One
3322 : * byte is written when not using it.
3323 : */
3324 3168 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT, IOOP_WRITE,
3325 : io_start, 1,
3326 3168 : wal_init_zero ? wal_segment_size : 1);
3327 :
3328 3168 : if (save_errno)
3329 : {
3330 : /*
3331 : * If we fail to make the file, delete it to release disk space
3332 : */
3333 0 : unlink(tmppath);
3334 :
3335 0 : close(fd);
3336 :
3337 0 : errno = save_errno;
3338 :
3339 0 : ereport(ERROR,
3340 : (errcode_for_file_access(),
3341 : errmsg("could not write to file \"%s\": %m", tmppath)));
3342 : }
3343 :
3344 : /* Measure I/O timing when flushing segment */
3345 3168 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3346 :
3347 3168 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC);
3348 3168 : if (pg_fsync(fd) != 0)
3349 : {
3350 0 : save_errno = errno;
3351 0 : close(fd);
3352 0 : errno = save_errno;
3353 0 : ereport(ERROR,
3354 : (errcode_for_file_access(),
3355 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3356 : }
3357 3168 : pgstat_report_wait_end();
3358 :
3359 3168 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT,
3360 : IOOP_FSYNC, io_start, 1, 0);
3361 :
3362 3168 : if (close(fd) != 0)
3363 0 : ereport(ERROR,
3364 : (errcode_for_file_access(),
3365 : errmsg("could not close file \"%s\": %m", tmppath)));
3366 :
3367 : /*
3368 : * Now move the segment into place with its final name. Cope with
3369 : * possibility that someone else has created the file while we were
3370 : * filling ours: if so, use ours to pre-create a future log segment.
3371 : */
3372 3168 : installed_segno = logsegno;
3373 :
3374 : /*
3375 : * XXX: What should we use as max_segno? We used to use XLOGfileslop when
3376 : * that was a constant, but that was always a bit dubious: normally, at a
3377 : * checkpoint, XLOGfileslop was the offset from the checkpoint record, but
3378 : * here, it was the offset from the insert location. We can't do the
3379 : * normal XLOGfileslop calculation here because we don't have access to
3380 : * the prior checkpoint's redo location. So somewhat arbitrarily, just use
3381 : * CheckPointSegments.
3382 : */
3383 3168 : max_segno = logsegno + CheckPointSegments;
3384 3168 : if (InstallXLogFileSegment(&installed_segno, tmppath, true, max_segno,
3385 : logtli))
3386 : {
3387 3168 : *added = true;
3388 3168 : elog(DEBUG2, "done creating and filling new WAL file");
3389 : }
3390 : else
3391 : {
3392 : /*
3393 : * No need for any more future segments, or InstallXLogFileSegment()
3394 : * failed to rename the file into place. If the rename failed, a
3395 : * caller opening the file may fail.
3396 : */
3397 0 : unlink(tmppath);
3398 0 : elog(DEBUG2, "abandoned new WAL file");
3399 : }
3400 :
3401 3168 : return -1;
3402 : }
3403 :
3404 : /*
3405 : * Create a new XLOG file segment, or open a pre-existing one.
3406 : *
3407 : * logsegno: identify segment to be created/opened.
3408 : *
3409 : * Returns FD of opened file.
3410 : *
3411 : * Note: errors here are ERROR not PANIC because we might or might not be
3412 : * inside a critical section (eg, during checkpoint there is no reason to
3413 : * take down the system on failure). They will promote to PANIC if we are
3414 : * in a critical section.
3415 : */
3416 : int
3417 29354 : XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)
3418 : {
3419 : bool ignore_added;
3420 : char path[MAXPGPATH];
3421 : int fd;
3422 :
3423 : Assert(logtli != 0);
3424 :
3425 29354 : fd = XLogFileInitInternal(logsegno, logtli, &ignore_added, path);
3426 29354 : if (fd >= 0)
3427 26356 : return fd;
3428 :
3429 : /* Now open original target segment (might not be file I just made) */
3430 2998 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3431 2998 : get_sync_bit(wal_sync_method));
3432 2998 : if (fd < 0)
3433 0 : ereport(ERROR,
3434 : (errcode_for_file_access(),
3435 : errmsg("could not open file \"%s\": %m", path)));
3436 2998 : return fd;
3437 : }
3438 :
3439 : /*
3440 : * Create a new XLOG file segment by copying a pre-existing one.
3441 : *
3442 : * destsegno: identify segment to be created.
3443 : *
3444 : * srcTLI, srcsegno: identify segment to be copied (could be from
3445 : * a different timeline)
3446 : *
3447 : * upto: how much of the source file to copy (the rest is filled with
3448 : * zeros)
3449 : *
3450 : * Currently this is only used during recovery, and so there are no locking
3451 : * considerations. But we should be just as tense as XLogFileInit to avoid
3452 : * emplacing a bogus file.
3453 : */
3454 : static void
3455 86 : XLogFileCopy(TimeLineID destTLI, XLogSegNo destsegno,
3456 : TimeLineID srcTLI, XLogSegNo srcsegno,
3457 : int upto)
3458 : {
3459 : char path[MAXPGPATH];
3460 : char tmppath[MAXPGPATH];
3461 : PGAlignedXLogBlock buffer;
3462 : int srcfd;
3463 : int fd;
3464 : int nbytes;
3465 :
3466 : /*
3467 : * Open the source file
3468 : */
3469 86 : XLogFilePath(path, srcTLI, srcsegno, wal_segment_size);
3470 86 : srcfd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
3471 86 : if (srcfd < 0)
3472 0 : ereport(ERROR,
3473 : (errcode_for_file_access(),
3474 : errmsg("could not open file \"%s\": %m", path)));
3475 :
3476 : /*
3477 : * Copy into a temp file name.
3478 : */
3479 86 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3480 :
3481 86 : unlink(tmppath);
3482 :
3483 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3484 86 : fd = OpenTransientFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
3485 86 : if (fd < 0)
3486 0 : ereport(ERROR,
3487 : (errcode_for_file_access(),
3488 : errmsg("could not create file \"%s\": %m", tmppath)));
3489 :
3490 : /*
3491 : * Do the data copying.
3492 : */
3493 176214 : for (nbytes = 0; nbytes < wal_segment_size; nbytes += sizeof(buffer))
3494 : {
3495 : int nread;
3496 :
3497 176128 : nread = upto - nbytes;
3498 :
3499 : /*
3500 : * The part that is not read from the source file is filled with
3501 : * zeros.
3502 : */
3503 176128 : if (nread < sizeof(buffer))
3504 86 : memset(buffer.data, 0, sizeof(buffer));
3505 :
3506 176128 : if (nread > 0)
3507 : {
3508 : int r;
3509 :
3510 5482 : if (nread > sizeof(buffer))
3511 5396 : nread = sizeof(buffer);
3512 5482 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ);
3513 5482 : r = read(srcfd, buffer.data, nread);
3514 5482 : if (r != nread)
3515 : {
3516 0 : if (r < 0)
3517 0 : ereport(ERROR,
3518 : (errcode_for_file_access(),
3519 : errmsg("could not read file \"%s\": %m",
3520 : path)));
3521 : else
3522 0 : ereport(ERROR,
3523 : (errcode(ERRCODE_DATA_CORRUPTED),
3524 : errmsg("could not read file \"%s\": read %d of %zu",
3525 : path, r, (Size) nread)));
3526 : }
3527 5482 : pgstat_report_wait_end();
3528 : }
3529 176128 : errno = 0;
3530 176128 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_WRITE);
3531 176128 : if ((int) write(fd, buffer.data, sizeof(buffer)) != (int) sizeof(buffer))
3532 : {
3533 0 : int save_errno = errno;
3534 :
3535 : /*
3536 : * If we fail to make the file, delete it to release disk space
3537 : */
3538 0 : unlink(tmppath);
3539 : /* if write didn't set errno, assume problem is no disk space */
3540 0 : errno = save_errno ? save_errno : ENOSPC;
3541 :
3542 0 : ereport(ERROR,
3543 : (errcode_for_file_access(),
3544 : errmsg("could not write to file \"%s\": %m", tmppath)));
3545 : }
3546 176128 : pgstat_report_wait_end();
3547 : }
3548 :
3549 86 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_SYNC);
3550 86 : if (pg_fsync(fd) != 0)
3551 0 : ereport(data_sync_elevel(ERROR),
3552 : (errcode_for_file_access(),
3553 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3554 86 : pgstat_report_wait_end();
3555 :
3556 86 : if (CloseTransientFile(fd) != 0)
3557 0 : ereport(ERROR,
3558 : (errcode_for_file_access(),
3559 : errmsg("could not close file \"%s\": %m", tmppath)));
3560 :
3561 86 : if (CloseTransientFile(srcfd) != 0)
3562 0 : ereport(ERROR,
3563 : (errcode_for_file_access(),
3564 : errmsg("could not close file \"%s\": %m", path)));
3565 :
3566 : /*
3567 : * Now move the segment into place with its final name.
3568 : */
3569 86 : if (!InstallXLogFileSegment(&destsegno, tmppath, false, 0, destTLI))
3570 0 : elog(ERROR, "InstallXLogFileSegment should not have failed");
3571 86 : }
3572 :
3573 : /*
3574 : * Install a new XLOG segment file as a current or future log segment.
3575 : *
3576 : * This is used both to install a newly-created segment (which has a temp
3577 : * filename while it's being created) and to recycle an old segment.
3578 : *
3579 : * *segno: identify segment to install as (or first possible target).
3580 : * When find_free is true, this is modified on return to indicate the
3581 : * actual installation location or last segment searched.
3582 : *
3583 : * tmppath: initial name of file to install. It will be renamed into place.
3584 : *
3585 : * find_free: if true, install the new segment at the first empty segno
3586 : * number at or after the passed numbers. If false, install the new segment
3587 : * exactly where specified, deleting any existing segment file there.
3588 : *
3589 : * max_segno: maximum segment number to install the new file as. Fail if no
3590 : * free slot is found between *segno and max_segno. (Ignored when find_free
3591 : * is false.)
3592 : *
3593 : * tli: The timeline on which the new segment should be installed.
3594 : *
3595 : * Returns true if the file was installed successfully. false indicates that
3596 : * max_segno limit was exceeded, the startup process has disabled this
3597 : * function for now, or an error occurred while renaming the file into place.
3598 : */
3599 : static bool
3600 6006 : InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
3601 : bool find_free, XLogSegNo max_segno, TimeLineID tli)
3602 : {
3603 : char path[MAXPGPATH];
3604 : struct stat stat_buf;
3605 :
3606 : Assert(tli != 0);
3607 :
3608 6006 : XLogFilePath(path, tli, *segno, wal_segment_size);
3609 :
3610 6006 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3611 6006 : if (!XLogCtl->InstallXLogFileSegmentActive)
3612 : {
3613 0 : LWLockRelease(ControlFileLock);
3614 0 : return false;
3615 : }
3616 :
3617 6006 : if (!find_free)
3618 : {
3619 : /* Force installation: get rid of any pre-existing segment file */
3620 86 : durable_unlink(path, DEBUG1);
3621 : }
3622 : else
3623 : {
3624 : /* Find a free slot to put it in */
3625 8136 : while (stat(path, &stat_buf) == 0)
3626 : {
3627 2436 : if ((*segno) >= max_segno)
3628 : {
3629 : /* Failed to find a free slot within specified range */
3630 220 : LWLockRelease(ControlFileLock);
3631 220 : return false;
3632 : }
3633 2216 : (*segno)++;
3634 2216 : XLogFilePath(path, tli, *segno, wal_segment_size);
3635 : }
3636 : }
3637 :
3638 : Assert(access(path, F_OK) != 0 && errno == ENOENT);
3639 5786 : if (durable_rename(tmppath, path, LOG) != 0)
3640 : {
3641 0 : LWLockRelease(ControlFileLock);
3642 : /* durable_rename already emitted log message */
3643 0 : return false;
3644 : }
3645 :
3646 5786 : LWLockRelease(ControlFileLock);
3647 :
3648 5786 : return true;
3649 : }
3650 :
3651 : /*
3652 : * Open a pre-existing logfile segment for writing.
3653 : */
3654 : int
3655 292 : XLogFileOpen(XLogSegNo segno, TimeLineID tli)
3656 : {
3657 : char path[MAXPGPATH];
3658 : int fd;
3659 :
3660 292 : XLogFilePath(path, tli, segno, wal_segment_size);
3661 :
3662 292 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3663 292 : get_sync_bit(wal_sync_method));
3664 292 : if (fd < 0)
3665 0 : ereport(PANIC,
3666 : (errcode_for_file_access(),
3667 : errmsg("could not open file \"%s\": %m", path)));
3668 :
3669 292 : return fd;
3670 : }
3671 :
3672 : /*
3673 : * Close the current logfile segment for writing.
3674 : */
3675 : static void
3676 12918 : XLogFileClose(void)
3677 : {
3678 : Assert(openLogFile >= 0);
3679 :
3680 : /*
3681 : * WAL segment files will not be re-read in normal operation, so we advise
3682 : * the OS to release any cached pages. But do not do so if WAL archiving
3683 : * or streaming is active, because archiver and walsender process could
3684 : * use the cache to read the WAL segment.
3685 : */
3686 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3687 12918 : if (!XLogIsNeeded() && (io_direct_flags & IO_DIRECT_WAL) == 0)
3688 226 : (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3689 : #endif
3690 :
3691 12918 : if (close(openLogFile) != 0)
3692 : {
3693 : char xlogfname[MAXFNAMELEN];
3694 0 : int save_errno = errno;
3695 :
3696 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo, wal_segment_size);
3697 0 : errno = save_errno;
3698 0 : ereport(PANIC,
3699 : (errcode_for_file_access(),
3700 : errmsg("could not close file \"%s\": %m", xlogfname)));
3701 : }
3702 :
3703 12918 : openLogFile = -1;
3704 12918 : ReleaseExternalFD();
3705 12918 : }
3706 :
3707 : /*
3708 : * Preallocate log files beyond the specified log endpoint.
3709 : *
3710 : * XXX this is currently extremely conservative, since it forces only one
3711 : * future log segment to exist, and even that only if we are 75% done with
3712 : * the current one. This is only appropriate for very low-WAL-volume systems.
3713 : * High-volume systems will be OK once they've built up a sufficient set of
3714 : * recycled log segments, but the startup transient is likely to include
3715 : * a lot of segment creations by foreground processes, which is not so good.
3716 : *
3717 : * XLogFileInitInternal() can ereport(ERROR). All known causes indicate big
3718 : * trouble; for example, a full filesystem is one cause. The checkpoint WAL
3719 : * and/or ControlFile updates already completed. If a RequestCheckpoint()
3720 : * initiated the present checkpoint and an ERROR ends this function, the
3721 : * command that called RequestCheckpoint() fails. That's not ideal, but it's
3722 : * not worth contorting more functions to use caller-specified elevel values.
3723 : * (With or without RequestCheckpoint(), an ERROR forestalls some inessential
3724 : * reporting and resource reclamation.)
3725 : */
3726 : static void
3727 4134 : PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli)
3728 : {
3729 : XLogSegNo _logSegNo;
3730 : int lf;
3731 : bool added;
3732 : char path[MAXPGPATH];
3733 : uint64 offset;
3734 :
3735 4134 : if (!XLogCtl->InstallXLogFileSegmentActive)
3736 16 : return; /* unlocked check says no */
3737 :
3738 4118 : XLByteToPrevSeg(endptr, _logSegNo, wal_segment_size);
3739 4118 : offset = XLogSegmentOffset(endptr - 1, wal_segment_size);
3740 4118 : if (offset >= (uint32) (0.75 * wal_segment_size))
3741 : {
3742 482 : _logSegNo++;
3743 482 : lf = XLogFileInitInternal(_logSegNo, tli, &added, path);
3744 482 : if (lf >= 0)
3745 312 : close(lf);
3746 482 : if (added)
3747 170 : CheckpointStats.ckpt_segs_added++;
3748 : }
3749 : }
3750 :
3751 : /*
3752 : * Throws an error if the given log segment has already been removed or
3753 : * recycled. The caller should only pass a segment that it knows to have
3754 : * existed while the server has been running, as this function always
3755 : * succeeds if no WAL segments have been removed since startup.
3756 : * 'tli' is only used in the error message.
3757 : *
3758 : * Note: this function guarantees to keep errno unchanged on return.
3759 : * This supports callers that use this to possibly deliver a better
3760 : * error message about a missing file, while still being able to throw
3761 : * a normal file-access error afterwards, if this does return.
3762 : */
3763 : void
3764 252292 : CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3765 : {
3766 252292 : int save_errno = errno;
3767 : XLogSegNo lastRemovedSegNo;
3768 :
3769 252292 : SpinLockAcquire(&XLogCtl->info_lck);
3770 252292 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3771 252292 : SpinLockRelease(&XLogCtl->info_lck);
3772 :
3773 252292 : if (segno <= lastRemovedSegNo)
3774 : {
3775 : char filename[MAXFNAMELEN];
3776 :
3777 0 : XLogFileName(filename, tli, segno, wal_segment_size);
3778 0 : errno = save_errno;
3779 0 : ereport(ERROR,
3780 : (errcode_for_file_access(),
3781 : errmsg("requested WAL segment %s has already been removed",
3782 : filename)));
3783 : }
3784 252292 : errno = save_errno;
3785 252292 : }
3786 :
3787 : /*
3788 : * Return the last WAL segment removed, or 0 if no segment has been removed
3789 : * since startup.
3790 : *
3791 : * NB: the result can be out of date arbitrarily fast, the caller has to deal
3792 : * with that.
3793 : */
3794 : XLogSegNo
3795 2502 : XLogGetLastRemovedSegno(void)
3796 : {
3797 : XLogSegNo lastRemovedSegNo;
3798 :
3799 2502 : SpinLockAcquire(&XLogCtl->info_lck);
3800 2502 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3801 2502 : SpinLockRelease(&XLogCtl->info_lck);
3802 :
3803 2502 : return lastRemovedSegNo;
3804 : }
3805 :
3806 : /*
3807 : * Return the oldest WAL segment on the given TLI that still exists in
3808 : * XLOGDIR, or 0 if none.
3809 : */
3810 : XLogSegNo
3811 14 : XLogGetOldestSegno(TimeLineID tli)
3812 : {
3813 : DIR *xldir;
3814 : struct dirent *xlde;
3815 14 : XLogSegNo oldest_segno = 0;
3816 :
3817 14 : xldir = AllocateDir(XLOGDIR);
3818 100 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3819 : {
3820 : TimeLineID file_tli;
3821 : XLogSegNo file_segno;
3822 :
3823 : /* Ignore files that are not XLOG segments. */
3824 86 : if (!IsXLogFileName(xlde->d_name))
3825 58 : continue;
3826 :
3827 : /* Parse filename to get TLI and segno. */
3828 28 : XLogFromFileName(xlde->d_name, &file_tli, &file_segno,
3829 : wal_segment_size);
3830 :
3831 : /* Ignore anything that's not from the TLI of interest. */
3832 28 : if (tli != file_tli)
3833 0 : continue;
3834 :
3835 : /* If it's the oldest so far, update oldest_segno. */
3836 28 : if (oldest_segno == 0 || file_segno < oldest_segno)
3837 16 : oldest_segno = file_segno;
3838 : }
3839 :
3840 14 : FreeDir(xldir);
3841 14 : return oldest_segno;
3842 : }
3843 :
3844 : /*
3845 : * Update the last removed segno pointer in shared memory, to reflect that the
3846 : * given XLOG file has been removed.
3847 : */
3848 : static void
3849 4872 : UpdateLastRemovedPtr(char *filename)
3850 : {
3851 : uint32 tli;
3852 : XLogSegNo segno;
3853 :
3854 4872 : XLogFromFileName(filename, &tli, &segno, wal_segment_size);
3855 :
3856 4872 : SpinLockAcquire(&XLogCtl->info_lck);
3857 4872 : if (segno > XLogCtl->lastRemovedSegNo)
3858 2072 : XLogCtl->lastRemovedSegNo = segno;
3859 4872 : SpinLockRelease(&XLogCtl->info_lck);
3860 4872 : }
3861 :
3862 : /*
3863 : * Remove all temporary log files in pg_wal
3864 : *
3865 : * This is called at the beginning of recovery after a previous crash,
3866 : * at a point where no other processes write fresh WAL data.
3867 : */
3868 : static void
3869 374 : RemoveTempXlogFiles(void)
3870 : {
3871 : DIR *xldir;
3872 : struct dirent *xlde;
3873 :
3874 374 : elog(DEBUG2, "removing all temporary WAL segments");
3875 :
3876 374 : xldir = AllocateDir(XLOGDIR);
3877 2532 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3878 : {
3879 : char path[MAXPGPATH];
3880 :
3881 2158 : if (strncmp(xlde->d_name, "xlogtemp.", 9) != 0)
3882 2158 : continue;
3883 :
3884 0 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
3885 0 : unlink(path);
3886 0 : elog(DEBUG2, "removed temporary WAL segment \"%s\"", path);
3887 : }
3888 374 : FreeDir(xldir);
3889 374 : }
3890 :
3891 : /*
3892 : * Recycle or remove all log files older or equal to passed segno.
3893 : *
3894 : * endptr is current (or recent) end of xlog, and lastredoptr is the
3895 : * redo pointer of the last checkpoint. These are used to determine
3896 : * whether we want to recycle rather than delete no-longer-wanted log files.
3897 : *
3898 : * insertTLI is the current timeline for XLOG insertion. Any recycled
3899 : * segments should be reused for this timeline.
3900 : */
3901 : static void
3902 3586 : RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr, XLogRecPtr endptr,
3903 : TimeLineID insertTLI)
3904 : {
3905 : DIR *xldir;
3906 : struct dirent *xlde;
3907 : char lastoff[MAXFNAMELEN];
3908 : XLogSegNo endlogSegNo;
3909 : XLogSegNo recycleSegNo;
3910 :
3911 : /* Initialize info about where to try to recycle to */
3912 3586 : XLByteToSeg(endptr, endlogSegNo, wal_segment_size);
3913 3586 : recycleSegNo = XLOGfileslop(lastredoptr);
3914 :
3915 : /*
3916 : * Construct a filename of the last segment to be kept. The timeline ID
3917 : * doesn't matter, we ignore that in the comparison. (During recovery,
3918 : * InsertTimeLineID isn't set, so we can't use that.)
3919 : */
3920 3586 : XLogFileName(lastoff, 0, segno, wal_segment_size);
3921 :
3922 3586 : elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
3923 : lastoff);
3924 :
3925 3586 : xldir = AllocateDir(XLOGDIR);
3926 :
3927 118488 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3928 : {
3929 : /* Ignore files that are not XLOG segments */
3930 114902 : if (!IsXLogFileName(xlde->d_name) &&
3931 15314 : !IsPartialXLogFileName(xlde->d_name))
3932 15310 : continue;
3933 :
3934 : /*
3935 : * We ignore the timeline part of the XLOG segment identifiers in
3936 : * deciding whether a segment is still needed. This ensures that we
3937 : * won't prematurely remove a segment from a parent timeline. We could
3938 : * probably be a little more proactive about removing segments of
3939 : * non-parent timelines, but that would be a whole lot more
3940 : * complicated.
3941 : *
3942 : * We use the alphanumeric sorting property of the filenames to decide
3943 : * which ones are earlier than the lastoff segment.
3944 : */
3945 99592 : if (strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
3946 : {
3947 83880 : if (XLogArchiveCheckDone(xlde->d_name))
3948 : {
3949 : /* Update the last removed location in shared memory first */
3950 4872 : UpdateLastRemovedPtr(xlde->d_name);
3951 :
3952 4872 : RemoveXlogFile(xlde, recycleSegNo, &endlogSegNo, insertTLI);
3953 : }
3954 : }
3955 : }
3956 :
3957 3586 : FreeDir(xldir);
3958 3586 : }
3959 :
3960 : /*
3961 : * Recycle or remove WAL files that are not part of the given timeline's
3962 : * history.
3963 : *
3964 : * This is called during recovery, whenever we switch to follow a new
3965 : * timeline, and at the end of recovery when we create a new timeline. We
3966 : * wouldn't otherwise care about extra WAL files lying in pg_wal, but they
3967 : * might be leftover pre-allocated or recycled WAL segments on the old timeline
3968 : * that we haven't used yet, and contain garbage. If we just leave them in
3969 : * pg_wal, they will eventually be archived, and we can't let that happen.
3970 : * Files that belong to our timeline history are valid, because we have
3971 : * successfully replayed them, but from others we can't be sure.
3972 : *
3973 : * 'switchpoint' is the current point in WAL where we switch to new timeline,
3974 : * and 'newTLI' is the new timeline we switch to.
3975 : */
3976 : void
3977 134 : RemoveNonParentXlogFiles(XLogRecPtr switchpoint, TimeLineID newTLI)
3978 : {
3979 : DIR *xldir;
3980 : struct dirent *xlde;
3981 : char switchseg[MAXFNAMELEN];
3982 : XLogSegNo endLogSegNo;
3983 : XLogSegNo switchLogSegNo;
3984 : XLogSegNo recycleSegNo;
3985 :
3986 : /*
3987 : * Initialize info about where to begin the work. This will recycle,
3988 : * somewhat arbitrarily, 10 future segments.
3989 : */
3990 134 : XLByteToPrevSeg(switchpoint, switchLogSegNo, wal_segment_size);
3991 134 : XLByteToSeg(switchpoint, endLogSegNo, wal_segment_size);
3992 134 : recycleSegNo = endLogSegNo + 10;
3993 :
3994 : /*
3995 : * Construct a filename of the last segment to be kept.
3996 : */
3997 134 : XLogFileName(switchseg, newTLI, switchLogSegNo, wal_segment_size);
3998 :
3999 134 : elog(DEBUG2, "attempting to remove WAL segments newer than log file %s",
4000 : switchseg);
4001 :
4002 134 : xldir = AllocateDir(XLOGDIR);
4003 :
4004 1270 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4005 : {
4006 : /* Ignore files that are not XLOG segments */
4007 1136 : if (!IsXLogFileName(xlde->d_name))
4008 706 : continue;
4009 :
4010 : /*
4011 : * Remove files that are on a timeline older than the new one we're
4012 : * switching to, but with a segment number >= the first segment on the
4013 : * new timeline.
4014 : */
4015 430 : if (strncmp(xlde->d_name, switchseg, 8) < 0 &&
4016 278 : strcmp(xlde->d_name + 8, switchseg + 8) > 0)
4017 : {
4018 : /*
4019 : * If the file has already been marked as .ready, however, don't
4020 : * remove it yet. It should be OK to remove it - files that are
4021 : * not part of our timeline history are not required for recovery
4022 : * - but seems safer to let them be archived and removed later.
4023 : */
4024 32 : if (!XLogArchiveIsReady(xlde->d_name))
4025 32 : RemoveXlogFile(xlde, recycleSegNo, &endLogSegNo, newTLI);
4026 : }
4027 : }
4028 :
4029 134 : FreeDir(xldir);
4030 134 : }
4031 :
4032 : /*
4033 : * Recycle or remove a log file that's no longer needed.
4034 : *
4035 : * segment_de is the dirent structure of the segment to recycle or remove.
4036 : * recycleSegNo is the segment number to recycle up to. endlogSegNo is
4037 : * the segment number of the current (or recent) end of WAL.
4038 : *
4039 : * endlogSegNo gets incremented if the segment is recycled so as it is not
4040 : * checked again with future callers of this function.
4041 : *
4042 : * insertTLI is the current timeline for XLOG insertion. Any recycled segments
4043 : * should be used for this timeline.
4044 : */
4045 : static void
4046 4904 : RemoveXlogFile(const struct dirent *segment_de,
4047 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
4048 : TimeLineID insertTLI)
4049 : {
4050 : char path[MAXPGPATH];
4051 : #ifdef WIN32
4052 : char newpath[MAXPGPATH];
4053 : #endif
4054 4904 : const char *segname = segment_de->d_name;
4055 :
4056 4904 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", segname);
4057 :
4058 : /*
4059 : * Before deleting the file, see if it can be recycled as a future log
4060 : * segment. Only recycle normal files, because we don't want to recycle
4061 : * symbolic links pointing to a separate archive directory.
4062 : */
4063 4904 : if (wal_recycle &&
4064 4904 : *endlogSegNo <= recycleSegNo &&
4065 6152 : XLogCtl->InstallXLogFileSegmentActive && /* callee rechecks this */
4066 5504 : get_dirent_type(path, segment_de, false, DEBUG2) == PGFILETYPE_REG &&
4067 2752 : InstallXLogFileSegment(endlogSegNo, path,
4068 : true, recycleSegNo, insertTLI))
4069 : {
4070 2532 : ereport(DEBUG2,
4071 : (errmsg_internal("recycled write-ahead log file \"%s\"",
4072 : segname)));
4073 2532 : CheckpointStats.ckpt_segs_recycled++;
4074 : /* Needn't recheck that slot on future iterations */
4075 2532 : (*endlogSegNo)++;
4076 : }
4077 : else
4078 : {
4079 : /* No need for any more future segments, or recycling failed ... */
4080 : int rc;
4081 :
4082 2372 : ereport(DEBUG2,
4083 : (errmsg_internal("removing write-ahead log file \"%s\"",
4084 : segname)));
4085 :
4086 : #ifdef WIN32
4087 :
4088 : /*
4089 : * On Windows, if another process (e.g another backend) holds the file
4090 : * open in FILE_SHARE_DELETE mode, unlink will succeed, but the file
4091 : * will still show up in directory listing until the last handle is
4092 : * closed. To avoid confusing the lingering deleted file for a live
4093 : * WAL file that needs to be archived, rename it before deleting it.
4094 : *
4095 : * If another process holds the file open without FILE_SHARE_DELETE
4096 : * flag, rename will fail. We'll try again at the next checkpoint.
4097 : */
4098 : snprintf(newpath, MAXPGPATH, "%s.deleted", path);
4099 : if (rename(path, newpath) != 0)
4100 : {
4101 : ereport(LOG,
4102 : (errcode_for_file_access(),
4103 : errmsg("could not rename file \"%s\": %m",
4104 : path)));
4105 : return;
4106 : }
4107 : rc = durable_unlink(newpath, LOG);
4108 : #else
4109 2372 : rc = durable_unlink(path, LOG);
4110 : #endif
4111 2372 : if (rc != 0)
4112 : {
4113 : /* Message already logged by durable_unlink() */
4114 0 : return;
4115 : }
4116 2372 : CheckpointStats.ckpt_segs_removed++;
4117 : }
4118 :
4119 4904 : XLogArchiveCleanup(segname);
4120 : }
4121 :
4122 : /*
4123 : * Verify whether pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
4124 : * If the latter do not exist, recreate them.
4125 : *
4126 : * It is not the goal of this function to verify the contents of these
4127 : * directories, but to help in cases where someone has performed a cluster
4128 : * copy for PITR purposes but omitted pg_wal from the copy.
4129 : *
4130 : * We could also recreate pg_wal if it doesn't exist, but a deliberate
4131 : * policy decision was made not to. It is fairly common for pg_wal to be
4132 : * a symlink, and if that was the DBA's intent then automatically making a
4133 : * plain directory would result in degraded performance with no notice.
4134 : */
4135 : static void
4136 1992 : ValidateXLOGDirectoryStructure(void)
4137 : {
4138 : char path[MAXPGPATH];
4139 : struct stat stat_buf;
4140 :
4141 : /* Check for pg_wal; if it doesn't exist, error out */
4142 1992 : if (stat(XLOGDIR, &stat_buf) != 0 ||
4143 1992 : !S_ISDIR(stat_buf.st_mode))
4144 0 : ereport(FATAL,
4145 : (errcode_for_file_access(),
4146 : errmsg("required WAL directory \"%s\" does not exist",
4147 : XLOGDIR)));
4148 :
4149 : /* Check for archive_status */
4150 1992 : snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4151 1992 : if (stat(path, &stat_buf) == 0)
4152 : {
4153 : /* Check for weird cases where it exists but isn't a directory */
4154 1990 : if (!S_ISDIR(stat_buf.st_mode))
4155 0 : ereport(FATAL,
4156 : (errcode_for_file_access(),
4157 : errmsg("required WAL directory \"%s\" does not exist",
4158 : path)));
4159 : }
4160 : else
4161 : {
4162 2 : ereport(LOG,
4163 : (errmsg("creating missing WAL directory \"%s\"", path)));
4164 2 : if (MakePGDirectory(path) < 0)
4165 0 : ereport(FATAL,
4166 : (errcode_for_file_access(),
4167 : errmsg("could not create missing directory \"%s\": %m",
4168 : path)));
4169 : }
4170 :
4171 : /* Check for summaries */
4172 1992 : snprintf(path, MAXPGPATH, XLOGDIR "/summaries");
4173 1992 : if (stat(path, &stat_buf) == 0)
4174 : {
4175 : /* Check for weird cases where it exists but isn't a directory */
4176 1990 : if (!S_ISDIR(stat_buf.st_mode))
4177 0 : ereport(FATAL,
4178 : (errmsg("required WAL directory \"%s\" does not exist",
4179 : path)));
4180 : }
4181 : else
4182 : {
4183 2 : ereport(LOG,
4184 : (errmsg("creating missing WAL directory \"%s\"", path)));
4185 2 : if (MakePGDirectory(path) < 0)
4186 0 : ereport(FATAL,
4187 : (errmsg("could not create missing directory \"%s\": %m",
4188 : path)));
4189 : }
4190 1992 : }
4191 :
4192 : /*
4193 : * Remove previous backup history files. This also retries creation of
4194 : * .ready files for any backup history files for which XLogArchiveNotify
4195 : * failed earlier.
4196 : */
4197 : static void
4198 314 : CleanupBackupHistory(void)
4199 : {
4200 : DIR *xldir;
4201 : struct dirent *xlde;
4202 : char path[MAXPGPATH + sizeof(XLOGDIR)];
4203 :
4204 314 : xldir = AllocateDir(XLOGDIR);
4205 :
4206 3234 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4207 : {
4208 2606 : if (IsBackupHistoryFileName(xlde->d_name))
4209 : {
4210 332 : if (XLogArchiveCheckDone(xlde->d_name))
4211 : {
4212 262 : elog(DEBUG2, "removing WAL backup history file \"%s\"",
4213 : xlde->d_name);
4214 262 : snprintf(path, sizeof(path), XLOGDIR "/%s", xlde->d_name);
4215 262 : unlink(path);
4216 262 : XLogArchiveCleanup(xlde->d_name);
4217 : }
4218 : }
4219 : }
4220 :
4221 314 : FreeDir(xldir);
4222 314 : }
4223 :
4224 : /*
4225 : * I/O routines for pg_control
4226 : *
4227 : * *ControlFile is a buffer in shared memory that holds an image of the
4228 : * contents of pg_control. WriteControlFile() initializes pg_control
4229 : * given a preloaded buffer, ReadControlFile() loads the buffer from
4230 : * the pg_control file (during postmaster or standalone-backend startup),
4231 : * and UpdateControlFile() rewrites pg_control after we modify xlog state.
4232 : * InitControlFile() fills the buffer with initial values.
4233 : *
4234 : * For simplicity, WriteControlFile() initializes the fields of pg_control
4235 : * that are related to checking backend/database compatibility, and
4236 : * ReadControlFile() verifies they are correct. We could split out the
4237 : * I/O and compatibility-check functions, but there seems no need currently.
4238 : */
4239 :
4240 : static void
4241 102 : InitControlFile(uint64 sysidentifier, uint32 data_checksum_version)
4242 : {
4243 : char mock_auth_nonce[MOCK_AUTH_NONCE_LEN];
4244 :
4245 : /*
4246 : * Generate a random nonce. This is used for authentication requests that
4247 : * will fail because the user does not exist. The nonce is used to create
4248 : * a genuine-looking password challenge for the non-existent user, in lieu
4249 : * of an actual stored password.
4250 : */
4251 102 : if (!pg_strong_random(mock_auth_nonce, MOCK_AUTH_NONCE_LEN))
4252 0 : ereport(PANIC,
4253 : (errcode(ERRCODE_INTERNAL_ERROR),
4254 : errmsg("could not generate secret authorization token")));
4255 :
4256 102 : memset(ControlFile, 0, sizeof(ControlFileData));
4257 : /* Initialize pg_control status fields */
4258 102 : ControlFile->system_identifier = sysidentifier;
4259 102 : memcpy(ControlFile->mock_authentication_nonce, mock_auth_nonce, MOCK_AUTH_NONCE_LEN);
4260 102 : ControlFile->state = DB_SHUTDOWNED;
4261 102 : ControlFile->unloggedLSN = FirstNormalUnloggedLSN;
4262 :
4263 : /* Set important parameter values for use when replaying WAL */
4264 102 : ControlFile->MaxConnections = MaxConnections;
4265 102 : ControlFile->max_worker_processes = max_worker_processes;
4266 102 : ControlFile->max_wal_senders = max_wal_senders;
4267 102 : ControlFile->max_prepared_xacts = max_prepared_xacts;
4268 102 : ControlFile->max_locks_per_xact = max_locks_per_xact;
4269 102 : ControlFile->wal_level = wal_level;
4270 102 : ControlFile->wal_log_hints = wal_log_hints;
4271 102 : ControlFile->track_commit_timestamp = track_commit_timestamp;
4272 102 : ControlFile->data_checksum_version = data_checksum_version;
4273 102 : }
4274 :
4275 : static void
4276 102 : WriteControlFile(void)
4277 : {
4278 : int fd;
4279 : char buffer[PG_CONTROL_FILE_SIZE]; /* need not be aligned */
4280 :
4281 : /*
4282 : * Initialize version and compatibility-check fields
4283 : */
4284 102 : ControlFile->pg_control_version = PG_CONTROL_VERSION;
4285 102 : ControlFile->catalog_version_no = CATALOG_VERSION_NO;
4286 :
4287 102 : ControlFile->maxAlign = MAXIMUM_ALIGNOF;
4288 102 : ControlFile->floatFormat = FLOATFORMAT_VALUE;
4289 :
4290 102 : ControlFile->blcksz = BLCKSZ;
4291 102 : ControlFile->relseg_size = RELSEG_SIZE;
4292 102 : ControlFile->slru_pages_per_segment = SLRU_PAGES_PER_SEGMENT;
4293 102 : ControlFile->xlog_blcksz = XLOG_BLCKSZ;
4294 102 : ControlFile->xlog_seg_size = wal_segment_size;
4295 :
4296 102 : ControlFile->nameDataLen = NAMEDATALEN;
4297 102 : ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
4298 :
4299 102 : ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
4300 102 : ControlFile->loblksize = LOBLKSIZE;
4301 :
4302 102 : ControlFile->float8ByVal = true; /* vestigial */
4303 :
4304 : /*
4305 : * Initialize the default 'char' signedness.
4306 : *
4307 : * The signedness of the char type is implementation-defined. For instance
4308 : * on x86 architecture CPUs, the char data type is typically treated as
4309 : * signed by default, whereas on aarch architecture CPUs, it is typically
4310 : * treated as unsigned by default. In v17 or earlier, we accidentally let
4311 : * C implementation signedness affect persistent data. This led to
4312 : * inconsistent results when comparing char data across different
4313 : * platforms.
4314 : *
4315 : * This flag can be used as a hint to ensure consistent behavior for
4316 : * pre-v18 data files that store data sorted by the 'char' type on disk,
4317 : * especially in cross-platform replication scenarios.
4318 : *
4319 : * Newly created database clusters unconditionally set the default char
4320 : * signedness to true. pg_upgrade changes this flag for clusters that were
4321 : * initialized on signedness=false platforms. As a result,
4322 : * signedness=false setting will become rare over time. If we had known
4323 : * about this problem during the last development cycle that forced initdb
4324 : * (v8.3), we would have made all clusters signed or all clusters
4325 : * unsigned. Making pg_upgrade the only source of signedness=false will
4326 : * cause the population of database clusters to converge toward that
4327 : * retrospective ideal.
4328 : */
4329 102 : ControlFile->default_char_signedness = true;
4330 :
4331 : /* Contents are protected with a CRC */
4332 102 : INIT_CRC32C(ControlFile->crc);
4333 102 : COMP_CRC32C(ControlFile->crc,
4334 : ControlFile,
4335 : offsetof(ControlFileData, crc));
4336 102 : FIN_CRC32C(ControlFile->crc);
4337 :
4338 : /*
4339 : * We write out PG_CONTROL_FILE_SIZE bytes into pg_control, zero-padding
4340 : * the excess over sizeof(ControlFileData). This reduces the odds of
4341 : * premature-EOF errors when reading pg_control. We'll still fail when we
4342 : * check the contents of the file, but hopefully with a more specific
4343 : * error than "couldn't read pg_control".
4344 : */
4345 102 : memset(buffer, 0, PG_CONTROL_FILE_SIZE);
4346 102 : memcpy(buffer, ControlFile, sizeof(ControlFileData));
4347 :
4348 102 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4349 : O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
4350 102 : if (fd < 0)
4351 0 : ereport(PANIC,
4352 : (errcode_for_file_access(),
4353 : errmsg("could not create file \"%s\": %m",
4354 : XLOG_CONTROL_FILE)));
4355 :
4356 102 : errno = 0;
4357 102 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_WRITE);
4358 102 : if (write(fd, buffer, PG_CONTROL_FILE_SIZE) != PG_CONTROL_FILE_SIZE)
4359 : {
4360 : /* if write didn't set errno, assume problem is no disk space */
4361 0 : if (errno == 0)
4362 0 : errno = ENOSPC;
4363 0 : ereport(PANIC,
4364 : (errcode_for_file_access(),
4365 : errmsg("could not write to file \"%s\": %m",
4366 : XLOG_CONTROL_FILE)));
4367 : }
4368 102 : pgstat_report_wait_end();
4369 :
4370 102 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_SYNC);
4371 102 : if (pg_fsync(fd) != 0)
4372 0 : ereport(PANIC,
4373 : (errcode_for_file_access(),
4374 : errmsg("could not fsync file \"%s\": %m",
4375 : XLOG_CONTROL_FILE)));
4376 102 : pgstat_report_wait_end();
4377 :
4378 102 : if (close(fd) != 0)
4379 0 : ereport(PANIC,
4380 : (errcode_for_file_access(),
4381 : errmsg("could not close file \"%s\": %m",
4382 : XLOG_CONTROL_FILE)));
4383 102 : }
4384 :
4385 : static void
4386 2092 : ReadControlFile(void)
4387 : {
4388 : pg_crc32c crc;
4389 : int fd;
4390 : char wal_segsz_str[20];
4391 : int r;
4392 :
4393 : /*
4394 : * Read data...
4395 : */
4396 2092 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4397 : O_RDWR | PG_BINARY);
4398 2092 : if (fd < 0)
4399 0 : ereport(PANIC,
4400 : (errcode_for_file_access(),
4401 : errmsg("could not open file \"%s\": %m",
4402 : XLOG_CONTROL_FILE)));
4403 :
4404 2092 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ);
4405 2092 : r = read(fd, ControlFile, sizeof(ControlFileData));
4406 2092 : if (r != sizeof(ControlFileData))
4407 : {
4408 0 : if (r < 0)
4409 0 : ereport(PANIC,
4410 : (errcode_for_file_access(),
4411 : errmsg("could not read file \"%s\": %m",
4412 : XLOG_CONTROL_FILE)));
4413 : else
4414 0 : ereport(PANIC,
4415 : (errcode(ERRCODE_DATA_CORRUPTED),
4416 : errmsg("could not read file \"%s\": read %d of %zu",
4417 : XLOG_CONTROL_FILE, r, sizeof(ControlFileData))));
4418 : }
4419 2092 : pgstat_report_wait_end();
4420 :
4421 2092 : close(fd);
4422 :
4423 : /*
4424 : * Check for expected pg_control format version. If this is wrong, the
4425 : * CRC check will likely fail because we'll be checking the wrong number
4426 : * of bytes. Complaining about wrong version will probably be more
4427 : * enlightening than complaining about wrong CRC.
4428 : */
4429 :
4430 2092 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
4431 0 : ereport(FATAL,
4432 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4433 : errmsg("database files are incompatible with server"),
4434 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4435 : " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4436 : ControlFile->pg_control_version, ControlFile->pg_control_version,
4437 : PG_CONTROL_VERSION, PG_CONTROL_VERSION),
4438 : errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4439 :
4440 2092 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
4441 0 : ereport(FATAL,
4442 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4443 : errmsg("database files are incompatible with server"),
4444 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4445 : " but the server was compiled with PG_CONTROL_VERSION %d.",
4446 : ControlFile->pg_control_version, PG_CONTROL_VERSION),
4447 : errhint("It looks like you need to initdb.")));
4448 :
4449 : /* Now check the CRC. */
4450 2092 : INIT_CRC32C(crc);
4451 2092 : COMP_CRC32C(crc,
4452 : ControlFile,
4453 : offsetof(ControlFileData, crc));
4454 2092 : FIN_CRC32C(crc);
4455 :
4456 2092 : if (!EQ_CRC32C(crc, ControlFile->crc))
4457 0 : ereport(FATAL,
4458 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4459 : errmsg("incorrect checksum in control file")));
4460 :
4461 : /*
4462 : * Do compatibility checking immediately. If the database isn't
4463 : * compatible with the backend executable, we want to abort before we can
4464 : * possibly do any damage.
4465 : */
4466 2092 : if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
4467 0 : ereport(FATAL,
4468 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4469 : errmsg("database files are incompatible with server"),
4470 : /* translator: %s is a variable name and %d is its value */
4471 : errdetail("The database cluster was initialized with %s %d,"
4472 : " but the server was compiled with %s %d.",
4473 : "CATALOG_VERSION_NO", ControlFile->catalog_version_no,
4474 : "CATALOG_VERSION_NO", CATALOG_VERSION_NO),
4475 : errhint("It looks like you need to initdb.")));
4476 2092 : if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
4477 0 : ereport(FATAL,
4478 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4479 : errmsg("database files are incompatible with server"),
4480 : /* translator: %s is a variable name and %d is its value */
4481 : errdetail("The database cluster was initialized with %s %d,"
4482 : " but the server was compiled with %s %d.",
4483 : "MAXALIGN", ControlFile->maxAlign,
4484 : "MAXALIGN", MAXIMUM_ALIGNOF),
4485 : errhint("It looks like you need to initdb.")));
4486 2092 : if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
4487 0 : ereport(FATAL,
4488 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4489 : errmsg("database files are incompatible with server"),
4490 : errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4491 : errhint("It looks like you need to initdb.")));
4492 2092 : if (ControlFile->blcksz != BLCKSZ)
4493 0 : ereport(FATAL,
4494 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4495 : errmsg("database files are incompatible with server"),
4496 : /* translator: %s is a variable name and %d is its value */
4497 : errdetail("The database cluster was initialized with %s %d,"
4498 : " but the server was compiled with %s %d.",
4499 : "BLCKSZ", ControlFile->blcksz,
4500 : "BLCKSZ", BLCKSZ),
4501 : errhint("It looks like you need to recompile or initdb.")));
4502 2092 : if (ControlFile->relseg_size != RELSEG_SIZE)
4503 0 : ereport(FATAL,
4504 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4505 : errmsg("database files are incompatible with server"),
4506 : /* translator: %s is a variable name and %d is its value */
4507 : errdetail("The database cluster was initialized with %s %d,"
4508 : " but the server was compiled with %s %d.",
4509 : "RELSEG_SIZE", ControlFile->relseg_size,
4510 : "RELSEG_SIZE", RELSEG_SIZE),
4511 : errhint("It looks like you need to recompile or initdb.")));
4512 2092 : if (ControlFile->slru_pages_per_segment != SLRU_PAGES_PER_SEGMENT)
4513 0 : ereport(FATAL,
4514 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4515 : errmsg("database files are incompatible with server"),
4516 : /* translator: %s is a variable name and %d is its value */
4517 : errdetail("The database cluster was initialized with %s %d,"
4518 : " but the server was compiled with %s %d.",
4519 : "SLRU_PAGES_PER_SEGMENT", ControlFile->slru_pages_per_segment,
4520 : "SLRU_PAGES_PER_SEGMENT", SLRU_PAGES_PER_SEGMENT),
4521 : errhint("It looks like you need to recompile or initdb.")));
4522 2092 : if (ControlFile->xlog_blcksz != XLOG_BLCKSZ)
4523 0 : ereport(FATAL,
4524 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4525 : errmsg("database files are incompatible with server"),
4526 : /* translator: %s is a variable name and %d is its value */
4527 : errdetail("The database cluster was initialized with %s %d,"
4528 : " but the server was compiled with %s %d.",
4529 : "XLOG_BLCKSZ", ControlFile->xlog_blcksz,
4530 : "XLOG_BLCKSZ", XLOG_BLCKSZ),
4531 : errhint("It looks like you need to recompile or initdb.")));
4532 2092 : if (ControlFile->nameDataLen != NAMEDATALEN)
4533 0 : ereport(FATAL,
4534 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4535 : errmsg("database files are incompatible with server"),
4536 : /* translator: %s is a variable name and %d is its value */
4537 : errdetail("The database cluster was initialized with %s %d,"
4538 : " but the server was compiled with %s %d.",
4539 : "NAMEDATALEN", ControlFile->nameDataLen,
4540 : "NAMEDATALEN", NAMEDATALEN),
4541 : errhint("It looks like you need to recompile or initdb.")));
4542 2092 : if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
4543 0 : ereport(FATAL,
4544 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4545 : errmsg("database files are incompatible with server"),
4546 : /* translator: %s is a variable name and %d is its value */
4547 : errdetail("The database cluster was initialized with %s %d,"
4548 : " but the server was compiled with %s %d.",
4549 : "INDEX_MAX_KEYS", ControlFile->indexMaxKeys,
4550 : "INDEX_MAX_KEYS", INDEX_MAX_KEYS),
4551 : errhint("It looks like you need to recompile or initdb.")));
4552 2092 : if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
4553 0 : ereport(FATAL,
4554 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4555 : errmsg("database files are incompatible with server"),
4556 : /* translator: %s is a variable name and %d is its value */
4557 : errdetail("The database cluster was initialized with %s %d,"
4558 : " but the server was compiled with %s %d.",
4559 : "TOAST_MAX_CHUNK_SIZE", ControlFile->toast_max_chunk_size,
4560 : "TOAST_MAX_CHUNK_SIZE", (int) TOAST_MAX_CHUNK_SIZE),
4561 : errhint("It looks like you need to recompile or initdb.")));
4562 2092 : if (ControlFile->loblksize != LOBLKSIZE)
4563 0 : ereport(FATAL,
4564 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4565 : errmsg("database files are incompatible with server"),
4566 : /* translator: %s is a variable name and %d is its value */
4567 : errdetail("The database cluster was initialized with %s %d,"
4568 : " but the server was compiled with %s %d.",
4569 : "LOBLKSIZE", ControlFile->loblksize,
4570 : "LOBLKSIZE", (int) LOBLKSIZE),
4571 : errhint("It looks like you need to recompile or initdb.")));
4572 :
4573 : Assert(ControlFile->float8ByVal); /* vestigial, not worth an error msg */
4574 :
4575 2092 : wal_segment_size = ControlFile->xlog_seg_size;
4576 :
4577 2092 : if (!IsValidWalSegSize(wal_segment_size))
4578 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4579 : errmsg_plural("invalid WAL segment size in control file (%d byte)",
4580 : "invalid WAL segment size in control file (%d bytes)",
4581 : wal_segment_size,
4582 : wal_segment_size),
4583 : errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.")));
4584 :
4585 2092 : snprintf(wal_segsz_str, sizeof(wal_segsz_str), "%d", wal_segment_size);
4586 2092 : SetConfigOption("wal_segment_size", wal_segsz_str, PGC_INTERNAL,
4587 : PGC_S_DYNAMIC_DEFAULT);
4588 :
4589 : /* check and update variables dependent on wal_segment_size */
4590 2092 : if (ConvertToXSegs(min_wal_size_mb, wal_segment_size) < 2)
4591 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4592 : /* translator: both %s are GUC names */
4593 : errmsg("\"%s\" must be at least twice \"%s\"",
4594 : "min_wal_size", "wal_segment_size")));
4595 :
4596 2092 : if (ConvertToXSegs(max_wal_size_mb, wal_segment_size) < 2)
4597 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4598 : /* translator: both %s are GUC names */
4599 : errmsg("\"%s\" must be at least twice \"%s\"",
4600 : "max_wal_size", "wal_segment_size")));
4601 :
4602 2092 : UsableBytesInSegment =
4603 2092 : (wal_segment_size / XLOG_BLCKSZ * UsableBytesInPage) -
4604 : (SizeOfXLogLongPHD - SizeOfXLogShortPHD);
4605 :
4606 2092 : CalculateCheckpointSegments();
4607 :
4608 : /* Make the initdb settings visible as GUC variables, too */
4609 2092 : SetConfigOption("data_checksums", DataChecksumsEnabled() ? "yes" : "no",
4610 : PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
4611 2092 : }
4612 :
4613 : /*
4614 : * Utility wrapper to update the control file. Note that the control
4615 : * file gets flushed.
4616 : */
4617 : static void
4618 19096 : UpdateControlFile(void)
4619 : {
4620 19096 : update_controlfile(DataDir, ControlFile, true);
4621 19096 : }
4622 :
4623 : /*
4624 : * Returns the unique system identifier from control file.
4625 : */
4626 : uint64
4627 2944 : GetSystemIdentifier(void)
4628 : {
4629 : Assert(ControlFile != NULL);
4630 2944 : return ControlFile->system_identifier;
4631 : }
4632 :
4633 : /*
4634 : * Returns the random nonce from control file.
4635 : */
4636 : char *
4637 2 : GetMockAuthenticationNonce(void)
4638 : {
4639 : Assert(ControlFile != NULL);
4640 2 : return ControlFile->mock_authentication_nonce;
4641 : }
4642 :
4643 : /*
4644 : * Are checksums enabled for data pages?
4645 : */
4646 : bool
4647 20454998 : DataChecksumsEnabled(void)
4648 : {
4649 : Assert(ControlFile != NULL);
4650 20454998 : return (ControlFile->data_checksum_version > 0);
4651 : }
4652 :
4653 : /*
4654 : * Return true if the cluster was initialized on a platform where the
4655 : * default signedness of char is "signed". This function exists for code
4656 : * that deals with pre-v18 data files that store data sorted by the 'char'
4657 : * type on disk (e.g., GIN and GiST indexes). See the comments in
4658 : * WriteControlFile() for details.
4659 : */
4660 : bool
4661 6 : GetDefaultCharSignedness(void)
4662 : {
4663 6 : return ControlFile->default_char_signedness;
4664 : }
4665 :
4666 : /*
4667 : * Returns a fake LSN for unlogged relations.
4668 : *
4669 : * Each call generates an LSN that is greater than any previous value
4670 : * returned. The current counter value is saved and restored across clean
4671 : * shutdowns, but like unlogged relations, does not survive a crash. This can
4672 : * be used in lieu of real LSN values returned by XLogInsert, if you need an
4673 : * LSN-like increasing sequence of numbers without writing any WAL.
4674 : */
4675 : XLogRecPtr
4676 66 : GetFakeLSNForUnloggedRel(void)
4677 : {
4678 66 : return pg_atomic_fetch_add_u64(&XLogCtl->unloggedLSN, 1);
4679 : }
4680 :
4681 : /*
4682 : * Auto-tune the number of XLOG buffers.
4683 : *
4684 : * The preferred setting for wal_buffers is about 3% of shared_buffers, with
4685 : * a maximum of one XLOG segment (there is little reason to think that more
4686 : * is helpful, at least so long as we force an fsync when switching log files)
4687 : * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
4688 : * 9.1, when auto-tuning was added).
4689 : *
4690 : * This should not be called until NBuffers has received its final value.
4691 : */
4692 : static int
4693 2276 : XLOGChooseNumBuffers(void)
4694 : {
4695 : int xbuffers;
4696 :
4697 2276 : xbuffers = NBuffers / 32;
4698 2276 : if (xbuffers > (wal_segment_size / XLOG_BLCKSZ))
4699 48 : xbuffers = (wal_segment_size / XLOG_BLCKSZ);
4700 2276 : if (xbuffers < 8)
4701 892 : xbuffers = 8;
4702 2276 : return xbuffers;
4703 : }
4704 :
4705 : /*
4706 : * GUC check_hook for wal_buffers
4707 : */
4708 : bool
4709 4632 : check_wal_buffers(int *newval, void **extra, GucSource source)
4710 : {
4711 : /*
4712 : * -1 indicates a request for auto-tune.
4713 : */
4714 4632 : if (*newval == -1)
4715 : {
4716 : /*
4717 : * If we haven't yet changed the boot_val default of -1, just let it
4718 : * be. We'll fix it when XLOGShmemSize is called.
4719 : */
4720 2356 : if (XLOGbuffers == -1)
4721 2356 : return true;
4722 :
4723 : /* Otherwise, substitute the auto-tune value */
4724 0 : *newval = XLOGChooseNumBuffers();
4725 : }
4726 :
4727 : /*
4728 : * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
4729 : * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
4730 : * the case, we just silently treat such values as a request for the
4731 : * minimum. (We could throw an error instead, but that doesn't seem very
4732 : * helpful.)
4733 : */
4734 2276 : if (*newval < 4)
4735 0 : *newval = 4;
4736 :
4737 2276 : return true;
4738 : }
4739 :
4740 : /*
4741 : * GUC check_hook for wal_consistency_checking
4742 : */
4743 : bool
4744 4312 : check_wal_consistency_checking(char **newval, void **extra, GucSource source)
4745 : {
4746 : char *rawstring;
4747 : List *elemlist;
4748 : ListCell *l;
4749 : bool newwalconsistency[RM_MAX_ID + 1];
4750 :
4751 : /* Initialize the array */
4752 142296 : MemSet(newwalconsistency, 0, (RM_MAX_ID + 1) * sizeof(bool));
4753 :
4754 : /* Need a modifiable copy of string */
4755 4312 : rawstring = pstrdup(*newval);
4756 :
4757 : /* Parse string into list of identifiers */
4758 4312 : if (!SplitIdentifierString(rawstring, ',', &elemlist))
4759 : {
4760 : /* syntax error in list */
4761 0 : GUC_check_errdetail("List syntax is invalid.");
4762 0 : pfree(rawstring);
4763 0 : list_free(elemlist);
4764 0 : return false;
4765 : }
4766 :
4767 5296 : foreach(l, elemlist)
4768 : {
4769 984 : char *tok = (char *) lfirst(l);
4770 : int rmid;
4771 :
4772 : /* Check for 'all'. */
4773 984 : if (pg_strcasecmp(tok, "all") == 0)
4774 : {
4775 251860 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4776 250880 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL)
4777 9800 : newwalconsistency[rmid] = true;
4778 : }
4779 : else
4780 : {
4781 : /* Check if the token matches any known resource manager. */
4782 4 : bool found = false;
4783 :
4784 72 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4785 : {
4786 108 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL &&
4787 36 : pg_strcasecmp(tok, GetRmgr(rmid).rm_name) == 0)
4788 : {
4789 4 : newwalconsistency[rmid] = true;
4790 4 : found = true;
4791 4 : break;
4792 : }
4793 : }
4794 4 : if (!found)
4795 : {
4796 : /*
4797 : * During startup, it might be a not-yet-loaded custom
4798 : * resource manager. Defer checking until
4799 : * InitializeWalConsistencyChecking().
4800 : */
4801 0 : if (!process_shared_preload_libraries_done)
4802 : {
4803 0 : check_wal_consistency_checking_deferred = true;
4804 : }
4805 : else
4806 : {
4807 0 : GUC_check_errdetail("Unrecognized key word: \"%s\".", tok);
4808 0 : pfree(rawstring);
4809 0 : list_free(elemlist);
4810 0 : return false;
4811 : }
4812 : }
4813 : }
4814 : }
4815 :
4816 4312 : pfree(rawstring);
4817 4312 : list_free(elemlist);
4818 :
4819 : /* assign new value */
4820 4312 : *extra = guc_malloc(LOG, (RM_MAX_ID + 1) * sizeof(bool));
4821 4312 : if (!*extra)
4822 0 : return false;
4823 4312 : memcpy(*extra, newwalconsistency, (RM_MAX_ID + 1) * sizeof(bool));
4824 4312 : return true;
4825 : }
4826 :
4827 : /*
4828 : * GUC assign_hook for wal_consistency_checking
4829 : */
4830 : void
4831 4310 : assign_wal_consistency_checking(const char *newval, void *extra)
4832 : {
4833 : /*
4834 : * If some checks were deferred, it's possible that the checks will fail
4835 : * later during InitializeWalConsistencyChecking(). But in that case, the
4836 : * postmaster will exit anyway, so it's safe to proceed with the
4837 : * assignment.
4838 : *
4839 : * Any built-in resource managers specified are assigned immediately,
4840 : * which affects WAL created before shared_preload_libraries are
4841 : * processed. Any custom resource managers specified won't be assigned
4842 : * until after shared_preload_libraries are processed, but that's OK
4843 : * because WAL for a custom resource manager can't be written before the
4844 : * module is loaded anyway.
4845 : */
4846 4310 : wal_consistency_checking = extra;
4847 4310 : }
4848 :
4849 : /*
4850 : * InitializeWalConsistencyChecking: run after loading custom resource managers
4851 : *
4852 : * If any unknown resource managers were specified in the
4853 : * wal_consistency_checking GUC, processing was deferred. Now that
4854 : * shared_preload_libraries have been loaded, process wal_consistency_checking
4855 : * again.
4856 : */
4857 : void
4858 1970 : InitializeWalConsistencyChecking(void)
4859 : {
4860 : Assert(process_shared_preload_libraries_done);
4861 :
4862 1970 : if (check_wal_consistency_checking_deferred)
4863 : {
4864 : struct config_generic *guc;
4865 :
4866 0 : guc = find_option("wal_consistency_checking", false, false, ERROR);
4867 :
4868 0 : check_wal_consistency_checking_deferred = false;
4869 :
4870 0 : set_config_option_ext("wal_consistency_checking",
4871 : wal_consistency_checking_string,
4872 : guc->scontext, guc->source, guc->srole,
4873 : GUC_ACTION_SET, true, ERROR, false);
4874 :
4875 : /* checking should not be deferred again */
4876 : Assert(!check_wal_consistency_checking_deferred);
4877 : }
4878 1970 : }
4879 :
4880 : /*
4881 : * GUC show_hook for archive_command
4882 : */
4883 : const char *
4884 3466 : show_archive_command(void)
4885 : {
4886 3466 : if (XLogArchivingActive())
4887 4 : return XLogArchiveCommand;
4888 : else
4889 3462 : return "(disabled)";
4890 : }
4891 :
4892 : /*
4893 : * GUC show_hook for in_hot_standby
4894 : */
4895 : const char *
4896 31738 : show_in_hot_standby(void)
4897 : {
4898 : /*
4899 : * We display the actual state based on shared memory, so that this GUC
4900 : * reports up-to-date state if examined intra-query. The underlying
4901 : * variable (in_hot_standby_guc) changes only when we transmit a new value
4902 : * to the client.
4903 : */
4904 31738 : return RecoveryInProgress() ? "on" : "off";
4905 : }
4906 :
4907 : /*
4908 : * GUC show_hook for effective_wal_level
4909 : */
4910 : const char *
4911 3530 : show_effective_wal_level(void)
4912 : {
4913 3530 : if (wal_level == WAL_LEVEL_MINIMAL)
4914 388 : return "minimal";
4915 :
4916 : /*
4917 : * During recovery, effective_wal_level reflects the primary's
4918 : * configuration rather than the local wal_level value.
4919 : */
4920 3142 : if (RecoveryInProgress())
4921 44 : return IsXLogLogicalInfoEnabled() ? "logical" : "replica";
4922 :
4923 3098 : return XLogLogicalInfoActive() ? "logical" : "replica";
4924 : }
4925 :
4926 : /*
4927 : * Read the control file, set respective GUCs.
4928 : *
4929 : * This is to be called during startup, including a crash recovery cycle,
4930 : * unless in bootstrap mode, where no control file yet exists. As there's no
4931 : * usable shared memory yet (its sizing can depend on the contents of the
4932 : * control file!), first store the contents in local memory. XLOGShmemInit()
4933 : * will then copy it to shared memory later.
4934 : *
4935 : * reset just controls whether previous contents are to be expected (in the
4936 : * reset case, there's a dangling pointer into old shared memory), or not.
4937 : */
4938 : void
4939 1990 : LocalProcessControlFile(bool reset)
4940 : {
4941 : Assert(reset || ControlFile == NULL);
4942 1990 : ControlFile = palloc_object(ControlFileData);
4943 1990 : ReadControlFile();
4944 1990 : }
4945 :
4946 : /*
4947 : * Get the wal_level from the control file. For a standby, this value should be
4948 : * considered as its active wal_level, because it may be different from what
4949 : * was originally configured on standby.
4950 : */
4951 : WalLevel
4952 0 : GetActiveWalLevelOnStandby(void)
4953 : {
4954 0 : return ControlFile->wal_level;
4955 : }
4956 :
4957 : /*
4958 : * Initialization of shared memory for XLOG
4959 : */
4960 : Size
4961 6534 : XLOGShmemSize(void)
4962 : {
4963 : Size size;
4964 :
4965 : /*
4966 : * If the value of wal_buffers is -1, use the preferred auto-tune value.
4967 : * This isn't an amazingly clean place to do this, but we must wait till
4968 : * NBuffers has received its final value, and must do it before using the
4969 : * value of XLOGbuffers to do anything important.
4970 : *
4971 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
4972 : * However, if the DBA explicitly set wal_buffers = -1 in the config file,
4973 : * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
4974 : * the matter with PGC_S_OVERRIDE.
4975 : */
4976 6534 : if (XLOGbuffers == -1)
4977 : {
4978 : char buf[32];
4979 :
4980 2276 : snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
4981 2276 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4982 : PGC_S_DYNAMIC_DEFAULT);
4983 2276 : if (XLOGbuffers == -1) /* failed to apply it? */
4984 0 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4985 : PGC_S_OVERRIDE);
4986 : }
4987 : Assert(XLOGbuffers > 0);
4988 :
4989 : /* XLogCtl */
4990 6534 : size = sizeof(XLogCtlData);
4991 :
4992 : /* WAL insertion locks, plus alignment */
4993 6534 : size = add_size(size, mul_size(sizeof(WALInsertLockPadded), NUM_XLOGINSERT_LOCKS + 1));
4994 : /* xlblocks array */
4995 6534 : size = add_size(size, mul_size(sizeof(pg_atomic_uint64), XLOGbuffers));
4996 : /* extra alignment padding for XLOG I/O buffers */
4997 6534 : size = add_size(size, Max(XLOG_BLCKSZ, PG_IO_ALIGN_SIZE));
4998 : /* and the buffers themselves */
4999 6534 : size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
5000 :
5001 : /*
5002 : * Note: we don't count ControlFileData, it comes out of the "slop factor"
5003 : * added by CreateSharedMemoryAndSemaphores. This lets us use this
5004 : * routine again below to compute the actual allocation size.
5005 : */
5006 :
5007 6534 : return size;
5008 : }
5009 :
5010 : void
5011 2280 : XLOGShmemInit(void)
5012 : {
5013 : bool foundCFile,
5014 : foundXLog;
5015 : char *allocptr;
5016 : int i;
5017 : ControlFileData *localControlFile;
5018 :
5019 : #ifdef WAL_DEBUG
5020 :
5021 : /*
5022 : * Create a memory context for WAL debugging that's exempt from the normal
5023 : * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
5024 : * an allocation fails, but wal_debug is not for production use anyway.
5025 : */
5026 : if (walDebugCxt == NULL)
5027 : {
5028 : walDebugCxt = AllocSetContextCreate(TopMemoryContext,
5029 : "WAL Debug",
5030 : ALLOCSET_DEFAULT_SIZES);
5031 : MemoryContextAllowInCriticalSection(walDebugCxt, true);
5032 : }
5033 : #endif
5034 :
5035 :
5036 2280 : XLogCtl = (XLogCtlData *)
5037 2280 : ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog);
5038 :
5039 2280 : localControlFile = ControlFile;
5040 2280 : ControlFile = (ControlFileData *)
5041 2280 : ShmemInitStruct("Control File", sizeof(ControlFileData), &foundCFile);
5042 :
5043 2280 : if (foundCFile || foundXLog)
5044 : {
5045 : /* both should be present or neither */
5046 : Assert(foundCFile && foundXLog);
5047 :
5048 : /* Initialize local copy of WALInsertLocks */
5049 0 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks;
5050 :
5051 0 : if (localControlFile)
5052 0 : pfree(localControlFile);
5053 0 : return;
5054 : }
5055 2280 : memset(XLogCtl, 0, sizeof(XLogCtlData));
5056 :
5057 : /*
5058 : * Already have read control file locally, unless in bootstrap mode. Move
5059 : * contents into shared memory.
5060 : */
5061 2280 : if (localControlFile)
5062 : {
5063 1974 : memcpy(ControlFile, localControlFile, sizeof(ControlFileData));
5064 1974 : pfree(localControlFile);
5065 : }
5066 :
5067 : /*
5068 : * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
5069 : * multiple of the alignment for same, so no extra alignment padding is
5070 : * needed here.
5071 : */
5072 2280 : allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
5073 2280 : XLogCtl->xlblocks = (pg_atomic_uint64 *) allocptr;
5074 2280 : allocptr += sizeof(pg_atomic_uint64) * XLOGbuffers;
5075 :
5076 649554 : for (i = 0; i < XLOGbuffers; i++)
5077 : {
5078 647274 : pg_atomic_init_u64(&XLogCtl->xlblocks[i], InvalidXLogRecPtr);
5079 : }
5080 :
5081 : /* WAL insertion locks. Ensure they're aligned to the full padded size */
5082 2280 : allocptr += sizeof(WALInsertLockPadded) -
5083 2280 : ((uintptr_t) allocptr) % sizeof(WALInsertLockPadded);
5084 2280 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks =
5085 : (WALInsertLockPadded *) allocptr;
5086 2280 : allocptr += sizeof(WALInsertLockPadded) * NUM_XLOGINSERT_LOCKS;
5087 :
5088 20520 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
5089 : {
5090 18240 : LWLockInitialize(&WALInsertLocks[i].l.lock, LWTRANCHE_WAL_INSERT);
5091 18240 : pg_atomic_init_u64(&WALInsertLocks[i].l.insertingAt, InvalidXLogRecPtr);
5092 18240 : WALInsertLocks[i].l.lastImportantAt = InvalidXLogRecPtr;
5093 : }
5094 :
5095 : /*
5096 : * Align the start of the page buffers to a full xlog block size boundary.
5097 : * This simplifies some calculations in XLOG insertion. It is also
5098 : * required for O_DIRECT.
5099 : */
5100 2280 : allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5101 2280 : XLogCtl->pages = allocptr;
5102 2280 : memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5103 :
5104 : /*
5105 : * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5106 : * in additional info.)
5107 : */
5108 2280 : XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5109 2280 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5110 2280 : XLogCtl->InstallXLogFileSegmentActive = false;
5111 2280 : XLogCtl->WalWriterSleeping = false;
5112 :
5113 2280 : SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5114 2280 : SpinLockInit(&XLogCtl->info_lck);
5115 2280 : pg_atomic_init_u64(&XLogCtl->logInsertResult, InvalidXLogRecPtr);
5116 2280 : pg_atomic_init_u64(&XLogCtl->logWriteResult, InvalidXLogRecPtr);
5117 2280 : pg_atomic_init_u64(&XLogCtl->logFlushResult, InvalidXLogRecPtr);
5118 2280 : pg_atomic_init_u64(&XLogCtl->unloggedLSN, InvalidXLogRecPtr);
5119 : }
5120 :
5121 : /*
5122 : * This func must be called ONCE on system install. It creates pg_control
5123 : * and the initial XLOG segment.
5124 : */
5125 : void
5126 102 : BootStrapXLOG(uint32 data_checksum_version)
5127 : {
5128 : CheckPoint checkPoint;
5129 : PGAlignedXLogBlock buffer;
5130 : XLogPageHeader page;
5131 : XLogLongPageHeader longpage;
5132 : XLogRecord *record;
5133 : char *recptr;
5134 : uint64 sysidentifier;
5135 : struct timeval tv;
5136 : pg_crc32c crc;
5137 :
5138 : /* allow ordinary WAL segment creation, like StartupXLOG() would */
5139 102 : SetInstallXLogFileSegmentActive();
5140 :
5141 : /*
5142 : * Select a hopefully-unique system identifier code for this installation.
5143 : * We use the result of gettimeofday(), including the fractional seconds
5144 : * field, as being about as unique as we can easily get. (Think not to
5145 : * use random(), since it hasn't been seeded and there's no portable way
5146 : * to seed it other than the system clock value...) The upper half of the
5147 : * uint64 value is just the tv_sec part, while the lower half contains the
5148 : * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
5149 : * PID for a little extra uniqueness. A person knowing this encoding can
5150 : * determine the initialization time of the installation, which could
5151 : * perhaps be useful sometimes.
5152 : */
5153 102 : gettimeofday(&tv, NULL);
5154 102 : sysidentifier = ((uint64) tv.tv_sec) << 32;
5155 102 : sysidentifier |= ((uint64) tv.tv_usec) << 12;
5156 102 : sysidentifier |= getpid() & 0xFFF;
5157 :
5158 102 : memset(&buffer, 0, sizeof buffer);
5159 102 : page = (XLogPageHeader) &buffer;
5160 :
5161 : /*
5162 : * Set up information for the initial checkpoint record
5163 : *
5164 : * The initial checkpoint record is written to the beginning of the WAL
5165 : * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5166 : * used, so that we can use 0/0 to mean "before any valid WAL segment".
5167 : */
5168 102 : checkPoint.redo = wal_segment_size + SizeOfXLogLongPHD;
5169 102 : checkPoint.ThisTimeLineID = BootstrapTimeLineID;
5170 102 : checkPoint.PrevTimeLineID = BootstrapTimeLineID;
5171 102 : checkPoint.fullPageWrites = fullPageWrites;
5172 102 : checkPoint.logicalDecodingEnabled = (wal_level == WAL_LEVEL_LOGICAL);
5173 102 : checkPoint.wal_level = wal_level;
5174 : checkPoint.nextXid =
5175 102 : FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId);
5176 102 : checkPoint.nextOid = FirstGenbkiObjectId;
5177 102 : checkPoint.nextMulti = FirstMultiXactId;
5178 102 : checkPoint.nextMultiOffset = 1;
5179 102 : checkPoint.oldestXid = FirstNormalTransactionId;
5180 102 : checkPoint.oldestXidDB = Template1DbOid;
5181 102 : checkPoint.oldestMulti = FirstMultiXactId;
5182 102 : checkPoint.oldestMultiDB = Template1DbOid;
5183 102 : checkPoint.oldestCommitTsXid = InvalidTransactionId;
5184 102 : checkPoint.newestCommitTsXid = InvalidTransactionId;
5185 102 : checkPoint.time = (pg_time_t) time(NULL);
5186 102 : checkPoint.oldestActiveXid = InvalidTransactionId;
5187 :
5188 102 : TransamVariables->nextXid = checkPoint.nextXid;
5189 102 : TransamVariables->nextOid = checkPoint.nextOid;
5190 102 : TransamVariables->oidCount = 0;
5191 102 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5192 102 : AdvanceOldestClogXid(checkPoint.oldestXid);
5193 102 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5194 102 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
5195 102 : SetCommitTsLimit(InvalidTransactionId, InvalidTransactionId);
5196 :
5197 : /* Set up the XLOG page header */
5198 102 : page->xlp_magic = XLOG_PAGE_MAGIC;
5199 102 : page->xlp_info = XLP_LONG_HEADER;
5200 102 : page->xlp_tli = BootstrapTimeLineID;
5201 102 : page->xlp_pageaddr = wal_segment_size;
5202 102 : longpage = (XLogLongPageHeader) page;
5203 102 : longpage->xlp_sysid = sysidentifier;
5204 102 : longpage->xlp_seg_size = wal_segment_size;
5205 102 : longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5206 :
5207 : /* Insert the initial checkpoint record */
5208 102 : recptr = ((char *) page + SizeOfXLogLongPHD);
5209 102 : record = (XLogRecord *) recptr;
5210 102 : record->xl_prev = 0;
5211 102 : record->xl_xid = InvalidTransactionId;
5212 102 : record->xl_tot_len = SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(checkPoint);
5213 102 : record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5214 102 : record->xl_rmid = RM_XLOG_ID;
5215 102 : recptr += SizeOfXLogRecord;
5216 : /* fill the XLogRecordDataHeaderShort struct */
5217 102 : *(recptr++) = (char) XLR_BLOCK_ID_DATA_SHORT;
5218 102 : *(recptr++) = sizeof(checkPoint);
5219 102 : memcpy(recptr, &checkPoint, sizeof(checkPoint));
5220 102 : recptr += sizeof(checkPoint);
5221 : Assert(recptr - (char *) record == record->xl_tot_len);
5222 :
5223 102 : INIT_CRC32C(crc);
5224 102 : COMP_CRC32C(crc, ((char *) record) + SizeOfXLogRecord, record->xl_tot_len - SizeOfXLogRecord);
5225 102 : COMP_CRC32C(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5226 102 : FIN_CRC32C(crc);
5227 102 : record->xl_crc = crc;
5228 :
5229 : /* Create first XLOG segment file */
5230 102 : openLogTLI = BootstrapTimeLineID;
5231 102 : openLogFile = XLogFileInit(1, BootstrapTimeLineID);
5232 :
5233 : /*
5234 : * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
5235 : * close the file again in a moment.
5236 : */
5237 :
5238 : /* Write the first page with the initial record */
5239 102 : errno = 0;
5240 102 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE);
5241 102 : if (write(openLogFile, &buffer, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5242 : {
5243 : /* if write didn't set errno, assume problem is no disk space */
5244 0 : if (errno == 0)
5245 0 : errno = ENOSPC;
5246 0 : ereport(PANIC,
5247 : (errcode_for_file_access(),
5248 : errmsg("could not write bootstrap write-ahead log file: %m")));
5249 : }
5250 102 : pgstat_report_wait_end();
5251 :
5252 102 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC);
5253 102 : if (pg_fsync(openLogFile) != 0)
5254 0 : ereport(PANIC,
5255 : (errcode_for_file_access(),
5256 : errmsg("could not fsync bootstrap write-ahead log file: %m")));
5257 102 : pgstat_report_wait_end();
5258 :
5259 102 : if (close(openLogFile) != 0)
5260 0 : ereport(PANIC,
5261 : (errcode_for_file_access(),
5262 : errmsg("could not close bootstrap write-ahead log file: %m")));
5263 :
5264 102 : openLogFile = -1;
5265 :
5266 : /* Now create pg_control */
5267 102 : InitControlFile(sysidentifier, data_checksum_version);
5268 102 : ControlFile->time = checkPoint.time;
5269 102 : ControlFile->checkPoint = checkPoint.redo;
5270 102 : ControlFile->checkPointCopy = checkPoint;
5271 :
5272 : /* some additional ControlFile fields are set in WriteControlFile() */
5273 102 : WriteControlFile();
5274 :
5275 : /* Bootstrap the commit log, too */
5276 102 : BootStrapCLOG();
5277 102 : BootStrapCommitTs();
5278 102 : BootStrapSUBTRANS();
5279 102 : BootStrapMultiXact();
5280 :
5281 : /*
5282 : * Force control file to be read - in contrast to normal processing we'd
5283 : * otherwise never run the checks and GUC related initializations therein.
5284 : */
5285 102 : ReadControlFile();
5286 102 : }
5287 :
5288 : static char *
5289 1768 : str_time(pg_time_t tnow, char *buf, size_t bufsize)
5290 : {
5291 1768 : pg_strftime(buf, bufsize,
5292 : "%Y-%m-%d %H:%M:%S %Z",
5293 1768 : pg_localtime(&tnow, log_timezone));
5294 :
5295 1768 : return buf;
5296 : }
5297 :
5298 : /*
5299 : * Initialize the first WAL segment on new timeline.
5300 : */
5301 : static void
5302 110 : XLogInitNewTimeline(TimeLineID endTLI, XLogRecPtr endOfLog, TimeLineID newTLI)
5303 : {
5304 : char xlogfname[MAXFNAMELEN];
5305 : XLogSegNo endLogSegNo;
5306 : XLogSegNo startLogSegNo;
5307 :
5308 : /* we always switch to a new timeline after archive recovery */
5309 : Assert(endTLI != newTLI);
5310 :
5311 : /*
5312 : * Update min recovery point one last time.
5313 : */
5314 110 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5315 :
5316 : /*
5317 : * Calculate the last segment on the old timeline, and the first segment
5318 : * on the new timeline. If the switch happens in the middle of a segment,
5319 : * they are the same, but if the switch happens exactly at a segment
5320 : * boundary, startLogSegNo will be endLogSegNo + 1.
5321 : */
5322 110 : XLByteToPrevSeg(endOfLog, endLogSegNo, wal_segment_size);
5323 110 : XLByteToSeg(endOfLog, startLogSegNo, wal_segment_size);
5324 :
5325 : /*
5326 : * Initialize the starting WAL segment for the new timeline. If the switch
5327 : * happens in the middle of a segment, copy data from the last WAL segment
5328 : * of the old timeline up to the switch point, to the starting WAL segment
5329 : * on the new timeline.
5330 : */
5331 110 : if (endLogSegNo == startLogSegNo)
5332 : {
5333 : /*
5334 : * Make a copy of the file on the new timeline.
5335 : *
5336 : * Writing WAL isn't allowed yet, so there are no locking
5337 : * considerations. But we should be just as tense as XLogFileInit to
5338 : * avoid emplacing a bogus file.
5339 : */
5340 86 : XLogFileCopy(newTLI, endLogSegNo, endTLI, endLogSegNo,
5341 86 : XLogSegmentOffset(endOfLog, wal_segment_size));
5342 : }
5343 : else
5344 : {
5345 : /*
5346 : * The switch happened at a segment boundary, so just create the next
5347 : * segment on the new timeline.
5348 : */
5349 : int fd;
5350 :
5351 24 : fd = XLogFileInit(startLogSegNo, newTLI);
5352 :
5353 24 : if (close(fd) != 0)
5354 : {
5355 0 : int save_errno = errno;
5356 :
5357 0 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5358 0 : errno = save_errno;
5359 0 : ereport(ERROR,
5360 : (errcode_for_file_access(),
5361 : errmsg("could not close file \"%s\": %m", xlogfname)));
5362 : }
5363 : }
5364 :
5365 : /*
5366 : * Let's just make real sure there are not .ready or .done flags posted
5367 : * for the new segment.
5368 : */
5369 110 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5370 110 : XLogArchiveCleanup(xlogfname);
5371 110 : }
5372 :
5373 : /*
5374 : * Perform cleanup actions at the conclusion of archive recovery.
5375 : */
5376 : static void
5377 110 : CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI, XLogRecPtr EndOfLog,
5378 : TimeLineID newTLI)
5379 : {
5380 : /*
5381 : * Execute the recovery_end_command, if any.
5382 : */
5383 110 : if (recoveryEndCommand && strcmp(recoveryEndCommand, "") != 0)
5384 4 : ExecuteRecoveryCommand(recoveryEndCommand,
5385 : "recovery_end_command",
5386 : true,
5387 : WAIT_EVENT_RECOVERY_END_COMMAND);
5388 :
5389 : /*
5390 : * We switched to a new timeline. Clean up segments on the old timeline.
5391 : *
5392 : * If there are any higher-numbered segments on the old timeline, remove
5393 : * them. They might contain valid WAL, but they might also be
5394 : * pre-allocated files containing garbage. In any case, they are not part
5395 : * of the new timeline's history so we don't need them.
5396 : */
5397 110 : RemoveNonParentXlogFiles(EndOfLog, newTLI);
5398 :
5399 : /*
5400 : * If the switch happened in the middle of a segment, what to do with the
5401 : * last, partial segment on the old timeline? If we don't archive it, and
5402 : * the server that created the WAL never archives it either (e.g. because
5403 : * it was hit by a meteor), it will never make it to the archive. That's
5404 : * OK from our point of view, because the new segment that we created with
5405 : * the new TLI contains all the WAL from the old timeline up to the switch
5406 : * point. But if you later try to do PITR to the "missing" WAL on the old
5407 : * timeline, recovery won't find it in the archive. It's physically
5408 : * present in the new file with new TLI, but recovery won't look there
5409 : * when it's recovering to the older timeline. On the other hand, if we
5410 : * archive the partial segment, and the original server on that timeline
5411 : * is still running and archives the completed version of the same segment
5412 : * later, it will fail. (We used to do that in 9.4 and below, and it
5413 : * caused such problems).
5414 : *
5415 : * As a compromise, we rename the last segment with the .partial suffix,
5416 : * and archive it. Archive recovery will never try to read .partial
5417 : * segments, so they will normally go unused. But in the odd PITR case,
5418 : * the administrator can copy them manually to the pg_wal directory
5419 : * (removing the suffix). They can be useful in debugging, too.
5420 : *
5421 : * If a .done or .ready file already exists for the old timeline, however,
5422 : * we had already determined that the segment is complete, so we can let
5423 : * it be archived normally. (In particular, if it was restored from the
5424 : * archive to begin with, it's expected to have a .done file).
5425 : */
5426 110 : if (XLogSegmentOffset(EndOfLog, wal_segment_size) != 0 &&
5427 : XLogArchivingActive())
5428 : {
5429 : char origfname[MAXFNAMELEN];
5430 : XLogSegNo endLogSegNo;
5431 :
5432 18 : XLByteToPrevSeg(EndOfLog, endLogSegNo, wal_segment_size);
5433 18 : XLogFileName(origfname, EndOfLogTLI, endLogSegNo, wal_segment_size);
5434 :
5435 18 : if (!XLogArchiveIsReadyOrDone(origfname))
5436 : {
5437 : char origpath[MAXPGPATH];
5438 : char partialfname[MAXFNAMELEN];
5439 : char partialpath[MAXPGPATH];
5440 :
5441 : /*
5442 : * If we're summarizing WAL, we can't rename the partial file
5443 : * until the summarizer finishes with it, else it will fail.
5444 : */
5445 10 : if (summarize_wal)
5446 2 : WaitForWalSummarization(EndOfLog);
5447 :
5448 10 : XLogFilePath(origpath, EndOfLogTLI, endLogSegNo, wal_segment_size);
5449 10 : snprintf(partialfname, MAXFNAMELEN, "%s.partial", origfname);
5450 10 : snprintf(partialpath, MAXPGPATH, "%s.partial", origpath);
5451 :
5452 : /*
5453 : * Make sure there's no .done or .ready file for the .partial
5454 : * file.
5455 : */
5456 10 : XLogArchiveCleanup(partialfname);
5457 :
5458 10 : durable_rename(origpath, partialpath, ERROR);
5459 10 : XLogArchiveNotify(partialfname);
5460 : }
5461 : }
5462 110 : }
5463 :
5464 : /*
5465 : * Check to see if required parameters are set high enough on this server
5466 : * for various aspects of recovery operation.
5467 : *
5468 : * Note that all the parameters which this function tests need to be
5469 : * listed in Administrator's Overview section in high-availability.sgml.
5470 : * If you change them, don't forget to update the list.
5471 : */
5472 : static void
5473 526 : CheckRequiredParameterValues(void)
5474 : {
5475 : /*
5476 : * For archive recovery, the WAL must be generated with at least 'replica'
5477 : * wal_level.
5478 : */
5479 526 : if (ArchiveRecoveryRequested && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5480 : {
5481 4 : ereport(FATAL,
5482 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5483 : errmsg("WAL was generated with \"wal_level=minimal\", cannot continue recovering"),
5484 : errdetail("This happens if you temporarily set \"wal_level=minimal\" on the server."),
5485 : errhint("Use a backup taken after setting \"wal_level\" to higher than \"minimal\".")));
5486 : }
5487 :
5488 : /*
5489 : * For Hot Standby, the WAL must be generated with 'replica' mode, and we
5490 : * must have at least as many backend slots as the primary.
5491 : */
5492 522 : if (ArchiveRecoveryRequested && EnableHotStandby)
5493 : {
5494 : /* We ignore autovacuum_worker_slots when we make this test. */
5495 274 : RecoveryRequiresIntParameter("max_connections",
5496 : MaxConnections,
5497 274 : ControlFile->MaxConnections);
5498 274 : RecoveryRequiresIntParameter("max_worker_processes",
5499 : max_worker_processes,
5500 274 : ControlFile->max_worker_processes);
5501 274 : RecoveryRequiresIntParameter("max_wal_senders",
5502 : max_wal_senders,
5503 274 : ControlFile->max_wal_senders);
5504 274 : RecoveryRequiresIntParameter("max_prepared_transactions",
5505 : max_prepared_xacts,
5506 274 : ControlFile->max_prepared_xacts);
5507 274 : RecoveryRequiresIntParameter("max_locks_per_transaction",
5508 : max_locks_per_xact,
5509 274 : ControlFile->max_locks_per_xact);
5510 : }
5511 522 : }
5512 :
5513 : /*
5514 : * This must be called ONCE during postmaster or standalone-backend startup
5515 : */
5516 : void
5517 1992 : StartupXLOG(void)
5518 : {
5519 : XLogCtlInsert *Insert;
5520 : CheckPoint checkPoint;
5521 : bool wasShutdown;
5522 : bool didCrash;
5523 : bool haveTblspcMap;
5524 : bool haveBackupLabel;
5525 : XLogRecPtr EndOfLog;
5526 : TimeLineID EndOfLogTLI;
5527 : TimeLineID newTLI;
5528 : bool performedWalRecovery;
5529 : EndOfWalRecoveryInfo *endOfRecoveryInfo;
5530 : XLogRecPtr abortedRecPtr;
5531 : XLogRecPtr missingContrecPtr;
5532 : TransactionId oldestActiveXID;
5533 1992 : bool promoted = false;
5534 : char timebuf[128];
5535 :
5536 : /*
5537 : * We should have an aux process resource owner to use, and we should not
5538 : * be in a transaction that's installed some other resowner.
5539 : */
5540 : Assert(AuxProcessResourceOwner != NULL);
5541 : Assert(CurrentResourceOwner == NULL ||
5542 : CurrentResourceOwner == AuxProcessResourceOwner);
5543 1992 : CurrentResourceOwner = AuxProcessResourceOwner;
5544 :
5545 : /*
5546 : * Check that contents look valid.
5547 : */
5548 1992 : if (!XRecOffIsValid(ControlFile->checkPoint))
5549 0 : ereport(FATAL,
5550 : (errcode(ERRCODE_DATA_CORRUPTED),
5551 : errmsg("control file contains invalid checkpoint location")));
5552 :
5553 1992 : switch (ControlFile->state)
5554 : {
5555 1552 : case DB_SHUTDOWNED:
5556 :
5557 : /*
5558 : * This is the expected case, so don't be chatty in standalone
5559 : * mode
5560 : */
5561 1552 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
5562 : (errmsg("database system was shut down at %s",
5563 : str_time(ControlFile->time,
5564 : timebuf, sizeof(timebuf)))));
5565 1552 : break;
5566 :
5567 66 : case DB_SHUTDOWNED_IN_RECOVERY:
5568 66 : ereport(LOG,
5569 : (errmsg("database system was shut down in recovery at %s",
5570 : str_time(ControlFile->time,
5571 : timebuf, sizeof(timebuf)))));
5572 66 : break;
5573 :
5574 0 : case DB_SHUTDOWNING:
5575 0 : ereport(LOG,
5576 : (errmsg("database system shutdown was interrupted; last known up at %s",
5577 : str_time(ControlFile->time,
5578 : timebuf, sizeof(timebuf)))));
5579 0 : break;
5580 :
5581 0 : case DB_IN_CRASH_RECOVERY:
5582 0 : ereport(LOG,
5583 : (errmsg("database system was interrupted while in recovery at %s",
5584 : str_time(ControlFile->time,
5585 : timebuf, sizeof(timebuf))),
5586 : errhint("This probably means that some data is corrupted and"
5587 : " you will have to use the last backup for recovery.")));
5588 0 : break;
5589 :
5590 14 : case DB_IN_ARCHIVE_RECOVERY:
5591 14 : ereport(LOG,
5592 : (errmsg("database system was interrupted while in recovery at log time %s",
5593 : str_time(ControlFile->checkPointCopy.time,
5594 : timebuf, sizeof(timebuf))),
5595 : errhint("If this has occurred more than once some data might be corrupted"
5596 : " and you might need to choose an earlier recovery target.")));
5597 14 : break;
5598 :
5599 360 : case DB_IN_PRODUCTION:
5600 360 : ereport(LOG,
5601 : (errmsg("database system was interrupted; last known up at %s",
5602 : str_time(ControlFile->time,
5603 : timebuf, sizeof(timebuf)))));
5604 360 : break;
5605 :
5606 0 : default:
5607 0 : ereport(FATAL,
5608 : (errcode(ERRCODE_DATA_CORRUPTED),
5609 : errmsg("control file contains invalid database cluster state")));
5610 : }
5611 :
5612 : /* This is just to allow attaching to startup process with a debugger */
5613 : #ifdef XLOG_REPLAY_DELAY
5614 : if (ControlFile->state != DB_SHUTDOWNED)
5615 : pg_usleep(60000000L);
5616 : #endif
5617 :
5618 : /*
5619 : * Verify that pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
5620 : * In cases where someone has performed a copy for PITR, these directories
5621 : * may have been excluded and need to be re-created.
5622 : */
5623 1992 : ValidateXLOGDirectoryStructure();
5624 :
5625 : /* Set up timeout handler needed to report startup progress. */
5626 1992 : if (!IsBootstrapProcessingMode())
5627 1890 : RegisterTimeout(STARTUP_PROGRESS_TIMEOUT,
5628 : startup_progress_timeout_handler);
5629 :
5630 : /*----------
5631 : * If we previously crashed, perform a couple of actions:
5632 : *
5633 : * - The pg_wal directory may still include some temporary WAL segments
5634 : * used when creating a new segment, so perform some clean up to not
5635 : * bloat this path. This is done first as there is no point to sync
5636 : * this temporary data.
5637 : *
5638 : * - There might be data which we had written, intending to fsync it, but
5639 : * which we had not actually fsync'd yet. Therefore, a power failure in
5640 : * the near future might cause earlier unflushed writes to be lost, even
5641 : * though more recent data written to disk from here on would be
5642 : * persisted. To avoid that, fsync the entire data directory.
5643 : */
5644 1992 : if (ControlFile->state != DB_SHUTDOWNED &&
5645 440 : ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
5646 : {
5647 374 : RemoveTempXlogFiles();
5648 374 : SyncDataDirectory();
5649 374 : didCrash = true;
5650 : }
5651 : else
5652 1618 : didCrash = false;
5653 :
5654 : /*
5655 : * Prepare for WAL recovery if needed.
5656 : *
5657 : * InitWalRecovery analyzes the control file and the backup label file, if
5658 : * any. It updates the in-memory ControlFile buffer according to the
5659 : * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5660 : * It also applies the tablespace map file, if any.
5661 : */
5662 1992 : InitWalRecovery(ControlFile, &wasShutdown,
5663 : &haveBackupLabel, &haveTblspcMap);
5664 1990 : checkPoint = ControlFile->checkPointCopy;
5665 :
5666 : /* initialize shared memory variables from the checkpoint record */
5667 1990 : TransamVariables->nextXid = checkPoint.nextXid;
5668 1990 : TransamVariables->nextOid = checkPoint.nextOid;
5669 1990 : TransamVariables->oidCount = 0;
5670 1990 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5671 1990 : AdvanceOldestClogXid(checkPoint.oldestXid);
5672 1990 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5673 1990 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
5674 1990 : SetCommitTsLimit(checkPoint.oldestCommitTsXid,
5675 : checkPoint.newestCommitTsXid);
5676 :
5677 : /*
5678 : * Clear out any old relcache cache files. This is *necessary* if we do
5679 : * any WAL replay, since that would probably result in the cache files
5680 : * being out of sync with database reality. In theory we could leave them
5681 : * in place if the database had been cleanly shut down, but it seems
5682 : * safest to just remove them always and let them be rebuilt during the
5683 : * first backend startup. These files needs to be removed from all
5684 : * directories including pg_tblspc, however the symlinks are created only
5685 : * after reading tablespace_map file in case of archive recovery from
5686 : * backup, so needs to clear old relcache files here after creating
5687 : * symlinks.
5688 : */
5689 1990 : RelationCacheInitFileRemove();
5690 :
5691 : /*
5692 : * Initialize replication slots, before there's a chance to remove
5693 : * required resources.
5694 : */
5695 1990 : StartupReplicationSlots();
5696 :
5697 : /*
5698 : * Startup the logical decoding status with the last status stored in the
5699 : * checkpoint record.
5700 : */
5701 1986 : StartupLogicalDecodingStatus(checkPoint.logicalDecodingEnabled);
5702 :
5703 : /*
5704 : * Startup logical state, needs to be setup now so we have proper data
5705 : * during crash recovery.
5706 : */
5707 1986 : StartupReorderBuffer();
5708 :
5709 : /*
5710 : * Startup CLOG. This must be done after TransamVariables->nextXid has
5711 : * been initialized and before we accept connections or begin WAL replay.
5712 : */
5713 1986 : StartupCLOG();
5714 :
5715 : /*
5716 : * Startup MultiXact. We need to do this early to be able to replay
5717 : * truncations.
5718 : */
5719 1986 : StartupMultiXact();
5720 :
5721 : /*
5722 : * Ditto for commit timestamps. Activate the facility if the setting is
5723 : * enabled in the control file, as there should be no tracking of commit
5724 : * timestamps done when the setting was disabled. This facility can be
5725 : * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
5726 : */
5727 1986 : if (ControlFile->track_commit_timestamp)
5728 28 : StartupCommitTs();
5729 :
5730 : /*
5731 : * Recover knowledge about replay progress of known replication partners.
5732 : */
5733 1986 : StartupReplicationOrigin();
5734 :
5735 : /*
5736 : * Initialize unlogged LSN. On a clean shutdown, it's restored from the
5737 : * control file. On recovery, all unlogged relations are blown away, so
5738 : * the unlogged LSN counter can be reset too.
5739 : */
5740 1986 : if (ControlFile->state == DB_SHUTDOWNED)
5741 1536 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5742 1536 : ControlFile->unloggedLSN);
5743 : else
5744 450 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5745 : FirstNormalUnloggedLSN);
5746 :
5747 : /*
5748 : * Copy any missing timeline history files between 'now' and the recovery
5749 : * target timeline from archive to pg_wal. While we don't need those files
5750 : * ourselves - the history file of the recovery target timeline covers all
5751 : * the previous timelines in the history too - a cascading standby server
5752 : * might be interested in them. Or, if you archive the WAL from this
5753 : * server to a different archive than the primary, it'd be good for all
5754 : * the history files to get archived there after failover, so that you can
5755 : * use one of the old timelines as a PITR target. Timeline history files
5756 : * are small, so it's better to copy them unnecessarily than not copy them
5757 : * and regret later.
5758 : */
5759 1986 : restoreTimeLineHistoryFiles(checkPoint.ThisTimeLineID, recoveryTargetTLI);
5760 :
5761 : /*
5762 : * Before running in recovery, scan pg_twophase and fill in its status to
5763 : * be able to work on entries generated by redo. Doing a scan before
5764 : * taking any recovery action has the merit to discard any 2PC files that
5765 : * are newer than the first record to replay, saving from any conflicts at
5766 : * replay. This avoids as well any subsequent scans when doing recovery
5767 : * of the on-disk two-phase data.
5768 : */
5769 1986 : restoreTwoPhaseData();
5770 :
5771 : /*
5772 : * When starting with crash recovery, reset pgstat data - it might not be
5773 : * valid. Otherwise restore pgstat data. It's safe to do this here,
5774 : * because postmaster will not yet have started any other processes.
5775 : *
5776 : * NB: Restoring replication slot stats relies on slot state to have
5777 : * already been restored from disk.
5778 : *
5779 : * TODO: With a bit of extra work we could just start with a pgstat file
5780 : * associated with the checkpoint redo location we're starting from.
5781 : */
5782 1986 : if (didCrash)
5783 372 : pgstat_discard_stats();
5784 : else
5785 1614 : pgstat_restore_stats();
5786 :
5787 1986 : lastFullPageWrites = checkPoint.fullPageWrites;
5788 :
5789 1986 : RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
5790 1986 : doPageWrites = lastFullPageWrites;
5791 :
5792 : /* REDO */
5793 1986 : if (InRecovery)
5794 : {
5795 : /* Initialize state for RecoveryInProgress() */
5796 450 : SpinLockAcquire(&XLogCtl->info_lck);
5797 450 : if (InArchiveRecovery)
5798 244 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
5799 : else
5800 206 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5801 450 : SpinLockRelease(&XLogCtl->info_lck);
5802 :
5803 : /*
5804 : * Update pg_control to show that we are recovering and to show the
5805 : * selected checkpoint as the place we are starting from. We also mark
5806 : * pg_control with any minimum recovery stop point obtained from a
5807 : * backup history file.
5808 : *
5809 : * No need to hold ControlFileLock yet, we aren't up far enough.
5810 : */
5811 450 : UpdateControlFile();
5812 :
5813 : /*
5814 : * If there was a backup label file, it's done its job and the info
5815 : * has now been propagated into pg_control. We must get rid of the
5816 : * label file so that if we crash during recovery, we'll pick up at
5817 : * the latest recovery restartpoint instead of going all the way back
5818 : * to the backup start point. It seems prudent though to just rename
5819 : * the file out of the way rather than delete it completely.
5820 : */
5821 450 : if (haveBackupLabel)
5822 : {
5823 162 : unlink(BACKUP_LABEL_OLD);
5824 162 : durable_rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD, FATAL);
5825 : }
5826 :
5827 : /*
5828 : * If there was a tablespace_map file, it's done its job and the
5829 : * symlinks have been created. We must get rid of the map file so
5830 : * that if we crash during recovery, we don't create symlinks again.
5831 : * It seems prudent though to just rename the file out of the way
5832 : * rather than delete it completely.
5833 : */
5834 450 : if (haveTblspcMap)
5835 : {
5836 4 : unlink(TABLESPACE_MAP_OLD);
5837 4 : durable_rename(TABLESPACE_MAP, TABLESPACE_MAP_OLD, FATAL);
5838 : }
5839 :
5840 : /*
5841 : * Initialize our local copy of minRecoveryPoint. When doing crash
5842 : * recovery we want to replay up to the end of WAL. Particularly, in
5843 : * the case of a promoted standby minRecoveryPoint value in the
5844 : * control file is only updated after the first checkpoint. However,
5845 : * if the instance crashes before the first post-recovery checkpoint
5846 : * is completed then recovery will use a stale location causing the
5847 : * startup process to think that there are still invalid page
5848 : * references when checking for data consistency.
5849 : */
5850 450 : if (InArchiveRecovery)
5851 : {
5852 244 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
5853 244 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
5854 : }
5855 : else
5856 : {
5857 206 : LocalMinRecoveryPoint = InvalidXLogRecPtr;
5858 206 : LocalMinRecoveryPointTLI = 0;
5859 : }
5860 :
5861 : /* Check that the GUCs used to generate the WAL allow recovery */
5862 450 : CheckRequiredParameterValues();
5863 :
5864 : /*
5865 : * We're in recovery, so unlogged relations may be trashed and must be
5866 : * reset. This should be done BEFORE allowing Hot Standby
5867 : * connections, so that read-only backends don't try to read whatever
5868 : * garbage is left over from before.
5869 : */
5870 450 : ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
5871 :
5872 : /*
5873 : * Likewise, delete any saved transaction snapshot files that got left
5874 : * behind by crashed backends.
5875 : */
5876 450 : DeleteAllExportedSnapshotFiles();
5877 :
5878 : /*
5879 : * Initialize for Hot Standby, if enabled. We won't let backends in
5880 : * yet, not until we've reached the min recovery point specified in
5881 : * control file and we've established a recovery snapshot from a
5882 : * running-xacts WAL record.
5883 : */
5884 450 : if (ArchiveRecoveryRequested && EnableHotStandby)
5885 : {
5886 : TransactionId *xids;
5887 : int nxids;
5888 :
5889 230 : ereport(DEBUG1,
5890 : (errmsg_internal("initializing for hot standby")));
5891 :
5892 230 : InitRecoveryTransactionEnvironment();
5893 :
5894 230 : if (wasShutdown)
5895 52 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
5896 : else
5897 178 : oldestActiveXID = checkPoint.oldestActiveXid;
5898 : Assert(TransactionIdIsValid(oldestActiveXID));
5899 :
5900 : /* Tell procarray about the range of xids it has to deal with */
5901 230 : ProcArrayInitRecovery(XidFromFullTransactionId(TransamVariables->nextXid));
5902 :
5903 : /*
5904 : * Startup subtrans only. CLOG, MultiXact and commit timestamp
5905 : * have already been started up and other SLRUs are not maintained
5906 : * during recovery and need not be started yet.
5907 : */
5908 230 : StartupSUBTRANS(oldestActiveXID);
5909 :
5910 : /*
5911 : * If we're beginning at a shutdown checkpoint, we know that
5912 : * nothing was running on the primary at this point. So fake-up an
5913 : * empty running-xacts record and use that here and now. Recover
5914 : * additional standby state for prepared transactions.
5915 : */
5916 230 : if (wasShutdown)
5917 : {
5918 : RunningTransactionsData running;
5919 : TransactionId latestCompletedXid;
5920 :
5921 : /* Update pg_subtrans entries for any prepared transactions */
5922 52 : StandbyRecoverPreparedTransactions();
5923 :
5924 : /*
5925 : * Construct a RunningTransactions snapshot representing a
5926 : * shut down server, with only prepared transactions still
5927 : * alive. We're never overflowed at this point because all
5928 : * subxids are listed with their parent prepared transactions.
5929 : */
5930 52 : running.xcnt = nxids;
5931 52 : running.subxcnt = 0;
5932 52 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
5933 52 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
5934 52 : running.oldestRunningXid = oldestActiveXID;
5935 52 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
5936 52 : TransactionIdRetreat(latestCompletedXid);
5937 : Assert(TransactionIdIsNormal(latestCompletedXid));
5938 52 : running.latestCompletedXid = latestCompletedXid;
5939 52 : running.xids = xids;
5940 :
5941 52 : ProcArrayApplyRecoveryInfo(&running);
5942 : }
5943 : }
5944 :
5945 : /*
5946 : * We're all set for replaying the WAL now. Do it.
5947 : */
5948 450 : PerformWalRecovery();
5949 326 : performedWalRecovery = true;
5950 : }
5951 : else
5952 1536 : performedWalRecovery = false;
5953 :
5954 : /*
5955 : * Finish WAL recovery.
5956 : */
5957 1862 : endOfRecoveryInfo = FinishWalRecovery();
5958 1862 : EndOfLog = endOfRecoveryInfo->endOfLog;
5959 1862 : EndOfLogTLI = endOfRecoveryInfo->endOfLogTLI;
5960 1862 : abortedRecPtr = endOfRecoveryInfo->abortedRecPtr;
5961 1862 : missingContrecPtr = endOfRecoveryInfo->missingContrecPtr;
5962 :
5963 : /*
5964 : * Reset ps status display, so as no information related to recovery shows
5965 : * up.
5966 : */
5967 1862 : set_ps_display("");
5968 :
5969 : /*
5970 : * When recovering from a backup (we are in recovery, and archive recovery
5971 : * was requested), complain if we did not roll forward far enough to reach
5972 : * the point where the database is consistent. For regular online
5973 : * backup-from-primary, that means reaching the end-of-backup WAL record
5974 : * (at which point we reset backupStartPoint to be Invalid), for
5975 : * backup-from-replica (which can't inject records into the WAL stream),
5976 : * that point is when we reach the minRecoveryPoint in pg_control (which
5977 : * we purposefully copy last when backing up from a replica). For
5978 : * pg_rewind (which creates a backup_label with a method of "pg_rewind")
5979 : * or snapshot-style backups (which don't), backupEndRequired will be set
5980 : * to false.
5981 : *
5982 : * Note: it is indeed okay to look at the local variable
5983 : * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
5984 : * might be further ahead --- ControlFile->minRecoveryPoint cannot have
5985 : * been advanced beyond the WAL we processed.
5986 : */
5987 1862 : if (InRecovery &&
5988 326 : (EndOfLog < LocalMinRecoveryPoint ||
5989 326 : XLogRecPtrIsValid(ControlFile->backupStartPoint)))
5990 : {
5991 : /*
5992 : * Ran off end of WAL before reaching end-of-backup WAL record, or
5993 : * minRecoveryPoint. That's a bad sign, indicating that you tried to
5994 : * recover from an online backup but never called pg_backup_stop(), or
5995 : * you didn't archive all the WAL needed.
5996 : */
5997 0 : if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
5998 : {
5999 0 : if (XLogRecPtrIsValid(ControlFile->backupStartPoint) || ControlFile->backupEndRequired)
6000 0 : ereport(FATAL,
6001 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6002 : errmsg("WAL ends before end of online backup"),
6003 : errhint("All WAL generated while online backup was taken must be available at recovery.")));
6004 : else
6005 0 : ereport(FATAL,
6006 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6007 : errmsg("WAL ends before consistent recovery point")));
6008 : }
6009 : }
6010 :
6011 : /*
6012 : * Reset unlogged relations to the contents of their INIT fork. This is
6013 : * done AFTER recovery is complete so as to include any unlogged relations
6014 : * created during recovery, but BEFORE recovery is marked as having
6015 : * completed successfully. Otherwise we'd not retry if any of the post
6016 : * end-of-recovery steps fail.
6017 : */
6018 1862 : if (InRecovery)
6019 326 : ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
6020 :
6021 : /*
6022 : * Pre-scan prepared transactions to find out the range of XIDs present.
6023 : * This information is not quite needed yet, but it is positioned here so
6024 : * as potential problems are detected before any on-disk change is done.
6025 : */
6026 1862 : oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
6027 :
6028 : /*
6029 : * Allow ordinary WAL segment creation before possibly switching to a new
6030 : * timeline, which creates a new segment, and after the last ReadRecord().
6031 : */
6032 1862 : SetInstallXLogFileSegmentActive();
6033 :
6034 : /*
6035 : * Consider whether we need to assign a new timeline ID.
6036 : *
6037 : * If we did archive recovery, we always assign a new ID. This handles a
6038 : * couple of issues. If we stopped short of the end of WAL during
6039 : * recovery, then we are clearly generating a new timeline and must assign
6040 : * it a unique new ID. Even if we ran to the end, modifying the current
6041 : * last segment is problematic because it may result in trying to
6042 : * overwrite an already-archived copy of that segment, and we encourage
6043 : * DBAs to make their archive_commands reject that. We can dodge the
6044 : * problem by making the new active segment have a new timeline ID.
6045 : *
6046 : * In a normal crash recovery, we can just extend the timeline we were in.
6047 : */
6048 1862 : newTLI = endOfRecoveryInfo->lastRecTLI;
6049 1862 : if (ArchiveRecoveryRequested)
6050 : {
6051 110 : newTLI = findNewestTimeLine(recoveryTargetTLI) + 1;
6052 110 : ereport(LOG,
6053 : (errmsg("selected new timeline ID: %u", newTLI)));
6054 :
6055 : /*
6056 : * Make a writable copy of the last WAL segment. (Note that we also
6057 : * have a copy of the last block of the old WAL in
6058 : * endOfRecovery->lastPage; we will use that below.)
6059 : */
6060 110 : XLogInitNewTimeline(EndOfLogTLI, EndOfLog, newTLI);
6061 :
6062 : /*
6063 : * Remove the signal files out of the way, so that we don't
6064 : * accidentally re-enter archive recovery mode in a subsequent crash.
6065 : */
6066 110 : if (endOfRecoveryInfo->standby_signal_file_found)
6067 104 : durable_unlink(STANDBY_SIGNAL_FILE, FATAL);
6068 :
6069 110 : if (endOfRecoveryInfo->recovery_signal_file_found)
6070 6 : durable_unlink(RECOVERY_SIGNAL_FILE, FATAL);
6071 :
6072 : /*
6073 : * Write the timeline history file, and have it archived. After this
6074 : * point (or rather, as soon as the file is archived), the timeline
6075 : * will appear as "taken" in the WAL archive and to any standby
6076 : * servers. If we crash before actually switching to the new
6077 : * timeline, standby servers will nevertheless think that we switched
6078 : * to the new timeline, and will try to connect to the new timeline.
6079 : * To minimize the window for that, try to do as little as possible
6080 : * between here and writing the end-of-recovery record.
6081 : */
6082 110 : writeTimeLineHistory(newTLI, recoveryTargetTLI,
6083 : EndOfLog, endOfRecoveryInfo->recoveryStopReason);
6084 :
6085 110 : ereport(LOG,
6086 : (errmsg("archive recovery complete")));
6087 : }
6088 :
6089 : /* Save the selected TimeLineID in shared memory, too */
6090 1862 : SpinLockAcquire(&XLogCtl->info_lck);
6091 1862 : XLogCtl->InsertTimeLineID = newTLI;
6092 1862 : XLogCtl->PrevTimeLineID = endOfRecoveryInfo->lastRecTLI;
6093 1862 : SpinLockRelease(&XLogCtl->info_lck);
6094 :
6095 : /*
6096 : * Actually, if WAL ended in an incomplete record, skip the parts that
6097 : * made it through and start writing after the portion that persisted.
6098 : * (It's critical to first write an OVERWRITE_CONTRECORD message, which
6099 : * we'll do as soon as we're open for writing new WAL.)
6100 : */
6101 1862 : if (XLogRecPtrIsValid(missingContrecPtr))
6102 : {
6103 : /*
6104 : * We should only have a missingContrecPtr if we're not switching to a
6105 : * new timeline. When a timeline switch occurs, WAL is copied from the
6106 : * old timeline to the new only up to the end of the last complete
6107 : * record, so there can't be an incomplete WAL record that we need to
6108 : * disregard.
6109 : */
6110 : Assert(newTLI == endOfRecoveryInfo->lastRecTLI);
6111 : Assert(XLogRecPtrIsValid(abortedRecPtr));
6112 22 : EndOfLog = missingContrecPtr;
6113 : }
6114 :
6115 : /*
6116 : * Prepare to write WAL starting at EndOfLog location, and init xlog
6117 : * buffer cache using the block containing the last record from the
6118 : * previous incarnation.
6119 : */
6120 1862 : Insert = &XLogCtl->Insert;
6121 1862 : Insert->PrevBytePos = XLogRecPtrToBytePos(endOfRecoveryInfo->lastRec);
6122 1862 : Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
6123 :
6124 : /*
6125 : * Tricky point here: lastPage contains the *last* block that the LastRec
6126 : * record spans, not the one it starts in. The last block is indeed the
6127 : * one we want to use.
6128 : */
6129 1862 : if (EndOfLog % XLOG_BLCKSZ != 0)
6130 : {
6131 : char *page;
6132 : int len;
6133 : int firstIdx;
6134 :
6135 1798 : firstIdx = XLogRecPtrToBufIdx(EndOfLog);
6136 1798 : len = EndOfLog - endOfRecoveryInfo->lastPageBeginPtr;
6137 : Assert(len < XLOG_BLCKSZ);
6138 :
6139 : /* Copy the valid part of the last block, and zero the rest */
6140 1798 : page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
6141 1798 : memcpy(page, endOfRecoveryInfo->lastPage, len);
6142 1798 : memset(page + len, 0, XLOG_BLCKSZ - len);
6143 :
6144 1798 : pg_atomic_write_u64(&XLogCtl->xlblocks[firstIdx], endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6145 1798 : XLogCtl->InitializedUpTo = endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ;
6146 : }
6147 : else
6148 : {
6149 : /*
6150 : * There is no partial block to copy. Just set InitializedUpTo, and
6151 : * let the first attempt to insert a log record to initialize the next
6152 : * buffer.
6153 : */
6154 64 : XLogCtl->InitializedUpTo = EndOfLog;
6155 : }
6156 :
6157 : /*
6158 : * Update local and shared status. This is OK to do without any locks
6159 : * because no other process can be reading or writing WAL yet.
6160 : */
6161 1862 : LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
6162 1862 : pg_atomic_write_u64(&XLogCtl->logInsertResult, EndOfLog);
6163 1862 : pg_atomic_write_u64(&XLogCtl->logWriteResult, EndOfLog);
6164 1862 : pg_atomic_write_u64(&XLogCtl->logFlushResult, EndOfLog);
6165 1862 : XLogCtl->LogwrtRqst.Write = EndOfLog;
6166 1862 : XLogCtl->LogwrtRqst.Flush = EndOfLog;
6167 :
6168 : /*
6169 : * Preallocate additional log files, if wanted.
6170 : */
6171 1862 : PreallocXlogFiles(EndOfLog, newTLI);
6172 :
6173 : /*
6174 : * Okay, we're officially UP.
6175 : */
6176 1862 : InRecovery = false;
6177 :
6178 : /* start the archive_timeout timer and LSN running */
6179 1862 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
6180 1862 : XLogCtl->lastSegSwitchLSN = EndOfLog;
6181 :
6182 : /* also initialize latestCompletedXid, to nextXid - 1 */
6183 1862 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
6184 1862 : TransamVariables->latestCompletedXid = TransamVariables->nextXid;
6185 1862 : FullTransactionIdRetreat(&TransamVariables->latestCompletedXid);
6186 1862 : LWLockRelease(ProcArrayLock);
6187 :
6188 : /*
6189 : * Start up subtrans, if not already done for hot standby. (commit
6190 : * timestamps are started below, if necessary.)
6191 : */
6192 1862 : if (standbyState == STANDBY_DISABLED)
6193 1752 : StartupSUBTRANS(oldestActiveXID);
6194 :
6195 : /*
6196 : * Perform end of recovery actions for any SLRUs that need it.
6197 : */
6198 1862 : TrimCLOG();
6199 1862 : TrimMultiXact();
6200 :
6201 : /*
6202 : * Reload shared-memory state for prepared transactions. This needs to
6203 : * happen before renaming the last partial segment of the old timeline as
6204 : * it may be possible that we have to recover some transactions from it.
6205 : */
6206 1862 : RecoverPreparedTransactions();
6207 :
6208 : /* Shut down xlogreader */
6209 1862 : ShutdownWalRecovery();
6210 :
6211 : /* Enable WAL writes for this backend only. */
6212 1862 : LocalSetXLogInsertAllowed();
6213 :
6214 : /* If necessary, write overwrite-contrecord before doing anything else */
6215 1862 : if (XLogRecPtrIsValid(abortedRecPtr))
6216 : {
6217 : Assert(XLogRecPtrIsValid(missingContrecPtr));
6218 22 : CreateOverwriteContrecordRecord(abortedRecPtr, missingContrecPtr, newTLI);
6219 : }
6220 :
6221 : /*
6222 : * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
6223 : * record before resource manager writes cleanup WAL records or checkpoint
6224 : * record is written.
6225 : */
6226 1862 : Insert->fullPageWrites = lastFullPageWrites;
6227 1862 : UpdateFullPageWrites();
6228 :
6229 : /*
6230 : * Emit checkpoint or end-of-recovery record in XLOG, if required.
6231 : */
6232 1862 : if (performedWalRecovery)
6233 326 : promoted = PerformRecoveryXLogAction();
6234 :
6235 : /*
6236 : * If any of the critical GUCs have changed, log them before we allow
6237 : * backends to write WAL.
6238 : */
6239 1862 : XLogReportParameters();
6240 :
6241 : /* If this is archive recovery, perform post-recovery cleanup actions. */
6242 1862 : if (ArchiveRecoveryRequested)
6243 110 : CleanupAfterArchiveRecovery(EndOfLogTLI, EndOfLog, newTLI);
6244 :
6245 : /*
6246 : * Local WAL inserts enabled, so it's time to finish initialization of
6247 : * commit timestamp.
6248 : */
6249 1862 : CompleteCommitTsInitialization();
6250 :
6251 : /*
6252 : * Update logical decoding status in shared memory and write an
6253 : * XLOG_LOGICAL_DECODING_STATUS_CHANGE, if necessary.
6254 : */
6255 1862 : UpdateLogicalDecodingStatusEndOfRecovery();
6256 :
6257 : /* Clean up EndOfWalRecoveryInfo data to appease Valgrind leak checking */
6258 1862 : if (endOfRecoveryInfo->lastPage)
6259 1820 : pfree(endOfRecoveryInfo->lastPage);
6260 1862 : pfree(endOfRecoveryInfo->recoveryStopReason);
6261 1862 : pfree(endOfRecoveryInfo);
6262 :
6263 : /*
6264 : * All done with end-of-recovery actions.
6265 : *
6266 : * Now allow backends to write WAL and update the control file status in
6267 : * consequence. SharedRecoveryState, that controls if backends can write
6268 : * WAL, is updated while holding ControlFileLock to prevent other backends
6269 : * to look at an inconsistent state of the control file in shared memory.
6270 : * There is still a small window during which backends can write WAL and
6271 : * the control file is still referring to a system not in DB_IN_PRODUCTION
6272 : * state while looking at the on-disk control file.
6273 : *
6274 : * Also, we use info_lck to update SharedRecoveryState to ensure that
6275 : * there are no race conditions concerning visibility of other recent
6276 : * updates to shared memory.
6277 : */
6278 1862 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6279 1862 : ControlFile->state = DB_IN_PRODUCTION;
6280 :
6281 1862 : SpinLockAcquire(&XLogCtl->info_lck);
6282 1862 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_DONE;
6283 1862 : SpinLockRelease(&XLogCtl->info_lck);
6284 :
6285 1862 : UpdateControlFile();
6286 1862 : LWLockRelease(ControlFileLock);
6287 :
6288 : /*
6289 : * Wake up the checkpointer process as there might be a request to disable
6290 : * logical decoding by concurrent slot drop.
6291 : */
6292 1862 : WakeupCheckpointer();
6293 :
6294 : /*
6295 : * Wake up all waiters. They need to report an error that recovery was
6296 : * ended before reaching the target LSN.
6297 : */
6298 1862 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_REPLAY, InvalidXLogRecPtr);
6299 1862 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_WRITE, InvalidXLogRecPtr);
6300 1862 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_FLUSH, InvalidXLogRecPtr);
6301 :
6302 : /*
6303 : * Shutdown the recovery environment. This must occur after
6304 : * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
6305 : * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
6306 : * any session building a snapshot will not rely on KnownAssignedXids as
6307 : * RecoveryInProgress() would return false at this stage. This is
6308 : * particularly critical for prepared 2PC transactions, that would still
6309 : * need to be included in snapshots once recovery has ended.
6310 : */
6311 1862 : if (standbyState != STANDBY_DISABLED)
6312 110 : ShutdownRecoveryTransactionEnvironment();
6313 :
6314 : /*
6315 : * If there were cascading standby servers connected to us, nudge any wal
6316 : * sender processes to notice that we've been promoted.
6317 : */
6318 1862 : WalSndWakeup(true, true);
6319 :
6320 : /*
6321 : * If this was a promotion, request an (online) checkpoint now. This isn't
6322 : * required for consistency, but the last restartpoint might be far back,
6323 : * and in case of a crash, recovering from it might take a longer than is
6324 : * appropriate now that we're not in standby mode anymore.
6325 : */
6326 1862 : if (promoted)
6327 96 : RequestCheckpoint(CHECKPOINT_FORCE);
6328 1862 : }
6329 :
6330 : /*
6331 : * Callback from PerformWalRecovery(), called when we switch from crash
6332 : * recovery to archive recovery mode. Updates the control file accordingly.
6333 : */
6334 : void
6335 4 : SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr, TimeLineID replayTLI)
6336 : {
6337 : /* initialize minRecoveryPoint to this record */
6338 4 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6339 4 : ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6340 4 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6341 : {
6342 4 : ControlFile->minRecoveryPoint = EndRecPtr;
6343 4 : ControlFile->minRecoveryPointTLI = replayTLI;
6344 : }
6345 : /* update local copy */
6346 4 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6347 4 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6348 :
6349 : /*
6350 : * The startup process can update its local copy of minRecoveryPoint from
6351 : * this point.
6352 : */
6353 4 : updateMinRecoveryPoint = true;
6354 :
6355 4 : UpdateControlFile();
6356 :
6357 : /*
6358 : * We update SharedRecoveryState while holding the lock on ControlFileLock
6359 : * so both states are consistent in shared memory.
6360 : */
6361 4 : SpinLockAcquire(&XLogCtl->info_lck);
6362 4 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6363 4 : SpinLockRelease(&XLogCtl->info_lck);
6364 :
6365 4 : LWLockRelease(ControlFileLock);
6366 4 : }
6367 :
6368 : /*
6369 : * Callback from PerformWalRecovery(), called when we reach the end of backup.
6370 : * Updates the control file accordingly.
6371 : */
6372 : void
6373 162 : ReachedEndOfBackup(XLogRecPtr EndRecPtr, TimeLineID tli)
6374 : {
6375 : /*
6376 : * We have reached the end of base backup, as indicated by pg_control. The
6377 : * data on disk is now consistent (unless minRecoveryPoint is further
6378 : * ahead, which can happen if we crashed during previous recovery). Reset
6379 : * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
6380 : * make sure we don't allow starting up at an earlier point even if
6381 : * recovery is stopped and restarted soon after this.
6382 : */
6383 162 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6384 :
6385 162 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6386 : {
6387 152 : ControlFile->minRecoveryPoint = EndRecPtr;
6388 152 : ControlFile->minRecoveryPointTLI = tli;
6389 : }
6390 :
6391 162 : ControlFile->backupStartPoint = InvalidXLogRecPtr;
6392 162 : ControlFile->backupEndPoint = InvalidXLogRecPtr;
6393 162 : ControlFile->backupEndRequired = false;
6394 162 : UpdateControlFile();
6395 :
6396 162 : LWLockRelease(ControlFileLock);
6397 162 : }
6398 :
6399 : /*
6400 : * Perform whatever XLOG actions are necessary at end of REDO.
6401 : *
6402 : * The goal here is to make sure that we'll be able to recover properly if
6403 : * we crash again. If we choose to write a checkpoint, we'll write a shutdown
6404 : * checkpoint rather than an on-line one. This is not particularly critical,
6405 : * but since we may be assigning a new TLI, using a shutdown checkpoint allows
6406 : * us to have the rule that TLI only changes in shutdown checkpoints, which
6407 : * allows some extra error checking in xlog_redo.
6408 : */
6409 : static bool
6410 326 : PerformRecoveryXLogAction(void)
6411 : {
6412 326 : bool promoted = false;
6413 :
6414 : /*
6415 : * Perform a checkpoint to update all our recovery activity to disk.
6416 : *
6417 : * Note that we write a shutdown checkpoint rather than an on-line one.
6418 : * This is not particularly critical, but since we may be assigning a new
6419 : * TLI, using a shutdown checkpoint allows us to have the rule that TLI
6420 : * only changes in shutdown checkpoints, which allows some extra error
6421 : * checking in xlog_redo.
6422 : *
6423 : * In promotion, only create a lightweight end-of-recovery record instead
6424 : * of a full checkpoint. A checkpoint is requested later, after we're
6425 : * fully out of recovery mode and already accepting queries.
6426 : */
6427 436 : if (ArchiveRecoveryRequested && IsUnderPostmaster &&
6428 110 : PromoteIsTriggered())
6429 : {
6430 96 : promoted = true;
6431 :
6432 : /*
6433 : * Insert a special WAL record to mark the end of recovery, since we
6434 : * aren't doing a checkpoint. That means that the checkpointer process
6435 : * may likely be in the middle of a time-smoothed restartpoint and
6436 : * could continue to be for minutes after this. That sounds strange,
6437 : * but the effect is roughly the same and it would be stranger to try
6438 : * to come out of the restartpoint and then checkpoint. We request a
6439 : * checkpoint later anyway, just for safety.
6440 : */
6441 96 : CreateEndOfRecoveryRecord();
6442 : }
6443 : else
6444 : {
6445 230 : RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
6446 : CHECKPOINT_FAST |
6447 : CHECKPOINT_WAIT);
6448 : }
6449 :
6450 326 : return promoted;
6451 : }
6452 :
6453 : /*
6454 : * Is the system still in recovery?
6455 : *
6456 : * Unlike testing InRecovery, this works in any process that's connected to
6457 : * shared memory.
6458 : */
6459 : bool
6460 162992080 : RecoveryInProgress(void)
6461 : {
6462 : /*
6463 : * We check shared state each time only until we leave recovery mode. We
6464 : * can't re-enter recovery, so there's no need to keep checking after the
6465 : * shared variable has once been seen false.
6466 : */
6467 162992080 : if (!LocalRecoveryInProgress)
6468 158484326 : return false;
6469 : else
6470 : {
6471 : /*
6472 : * use volatile pointer to make sure we make a fresh read of the
6473 : * shared variable.
6474 : */
6475 4507754 : volatile XLogCtlData *xlogctl = XLogCtl;
6476 :
6477 4507754 : LocalRecoveryInProgress = (xlogctl->SharedRecoveryState != RECOVERY_STATE_DONE);
6478 :
6479 : /*
6480 : * Note: We don't need a memory barrier when we're still in recovery.
6481 : * We might exit recovery immediately after return, so the caller
6482 : * can't rely on 'true' meaning that we're still in recovery anyway.
6483 : */
6484 :
6485 4507754 : return LocalRecoveryInProgress;
6486 : }
6487 : }
6488 :
6489 : /*
6490 : * Returns current recovery state from shared memory.
6491 : *
6492 : * This returned state is kept consistent with the contents of the control
6493 : * file. See details about the possible values of RecoveryState in xlog.h.
6494 : */
6495 : RecoveryState
6496 79954 : GetRecoveryState(void)
6497 : {
6498 : RecoveryState retval;
6499 :
6500 79954 : SpinLockAcquire(&XLogCtl->info_lck);
6501 79954 : retval = XLogCtl->SharedRecoveryState;
6502 79954 : SpinLockRelease(&XLogCtl->info_lck);
6503 :
6504 79954 : return retval;
6505 : }
6506 :
6507 : /*
6508 : * Is this process allowed to insert new WAL records?
6509 : *
6510 : * Ordinarily this is essentially equivalent to !RecoveryInProgress().
6511 : * But we also have provisions for forcing the result "true" or "false"
6512 : * within specific processes regardless of the global state.
6513 : */
6514 : bool
6515 81979488 : XLogInsertAllowed(void)
6516 : {
6517 : /*
6518 : * If value is "unconditionally true" or "unconditionally false", just
6519 : * return it. This provides the normal fast path once recovery is known
6520 : * done.
6521 : */
6522 81979488 : if (LocalXLogInsertAllowed >= 0)
6523 80524646 : return (bool) LocalXLogInsertAllowed;
6524 :
6525 : /*
6526 : * Else, must check to see if we're still in recovery.
6527 : */
6528 1454842 : if (RecoveryInProgress())
6529 1435338 : return false;
6530 :
6531 : /*
6532 : * On exit from recovery, reset to "unconditionally true", since there is
6533 : * no need to keep checking.
6534 : */
6535 19504 : LocalXLogInsertAllowed = 1;
6536 19504 : return true;
6537 : }
6538 :
6539 : /*
6540 : * Make XLogInsertAllowed() return true in the current process only.
6541 : *
6542 : * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
6543 : * and even call LocalSetXLogInsertAllowed() again after that.
6544 : *
6545 : * Returns the previous value of LocalXLogInsertAllowed.
6546 : */
6547 : static int
6548 1920 : LocalSetXLogInsertAllowed(void)
6549 : {
6550 1920 : int oldXLogAllowed = LocalXLogInsertAllowed;
6551 :
6552 1920 : LocalXLogInsertAllowed = 1;
6553 :
6554 1920 : return oldXLogAllowed;
6555 : }
6556 :
6557 : /*
6558 : * Return the current Redo pointer from shared memory.
6559 : *
6560 : * As a side-effect, the local RedoRecPtr copy is updated.
6561 : */
6562 : XLogRecPtr
6563 606008 : GetRedoRecPtr(void)
6564 : {
6565 : XLogRecPtr ptr;
6566 :
6567 : /*
6568 : * The possibly not up-to-date copy in XlogCtl is enough. Even if we
6569 : * grabbed a WAL insertion lock to read the authoritative value in
6570 : * Insert->RedoRecPtr, someone might update it just after we've released
6571 : * the lock.
6572 : */
6573 606008 : SpinLockAcquire(&XLogCtl->info_lck);
6574 606008 : ptr = XLogCtl->RedoRecPtr;
6575 606008 : SpinLockRelease(&XLogCtl->info_lck);
6576 :
6577 606008 : if (RedoRecPtr < ptr)
6578 3338 : RedoRecPtr = ptr;
6579 :
6580 606008 : return RedoRecPtr;
6581 : }
6582 :
6583 : /*
6584 : * Return information needed to decide whether a modified block needs a
6585 : * full-page image to be included in the WAL record.
6586 : *
6587 : * The returned values are cached copies from backend-private memory, and
6588 : * possibly out-of-date or, indeed, uninitialized, in which case they will
6589 : * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6590 : * re-check them against up-to-date values, while holding the WAL insert lock.
6591 : */
6592 : void
6593 31078276 : GetFullPageWriteInfo(XLogRecPtr *RedoRecPtr_p, bool *doPageWrites_p)
6594 : {
6595 31078276 : *RedoRecPtr_p = RedoRecPtr;
6596 31078276 : *doPageWrites_p = doPageWrites;
6597 31078276 : }
6598 :
6599 : /*
6600 : * GetInsertRecPtr -- Returns the current insert position.
6601 : *
6602 : * NOTE: The value *actually* returned is the position of the last full
6603 : * xlog page. It lags behind the real insert position by at most 1 page.
6604 : * For that, we don't need to scan through WAL insertion locks, and an
6605 : * approximation is enough for the current usage of this function.
6606 : */
6607 : XLogRecPtr
6608 13966 : GetInsertRecPtr(void)
6609 : {
6610 : XLogRecPtr recptr;
6611 :
6612 13966 : SpinLockAcquire(&XLogCtl->info_lck);
6613 13966 : recptr = XLogCtl->LogwrtRqst.Write;
6614 13966 : SpinLockRelease(&XLogCtl->info_lck);
6615 :
6616 13966 : return recptr;
6617 : }
6618 :
6619 : /*
6620 : * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6621 : * position known to be fsync'd to disk. This should only be used on a
6622 : * system that is known not to be in recovery.
6623 : */
6624 : XLogRecPtr
6625 505120 : GetFlushRecPtr(TimeLineID *insertTLI)
6626 : {
6627 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6628 :
6629 505120 : RefreshXLogWriteResult(LogwrtResult);
6630 :
6631 : /*
6632 : * If we're writing and flushing WAL, the time line can't be changing, so
6633 : * no lock is required.
6634 : */
6635 505120 : if (insertTLI)
6636 48108 : *insertTLI = XLogCtl->InsertTimeLineID;
6637 :
6638 505120 : return LogwrtResult.Flush;
6639 : }
6640 :
6641 : /*
6642 : * GetWALInsertionTimeLine -- Returns the current timeline of a system that
6643 : * is not in recovery.
6644 : */
6645 : TimeLineID
6646 234528 : GetWALInsertionTimeLine(void)
6647 : {
6648 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6649 :
6650 : /* Since the value can't be changing, no lock is required. */
6651 234528 : return XLogCtl->InsertTimeLineID;
6652 : }
6653 :
6654 : /*
6655 : * GetWALInsertionTimeLineIfSet -- If the system is not in recovery, returns
6656 : * the WAL insertion timeline; else, returns 0. Wherever possible, use
6657 : * GetWALInsertionTimeLine() instead, since it's cheaper. Note that this
6658 : * function decides recovery has ended as soon as the insert TLI is set, which
6659 : * happens before we set XLogCtl->SharedRecoveryState to RECOVERY_STATE_DONE.
6660 : */
6661 : TimeLineID
6662 2 : GetWALInsertionTimeLineIfSet(void)
6663 : {
6664 : TimeLineID insertTLI;
6665 :
6666 2 : SpinLockAcquire(&XLogCtl->info_lck);
6667 2 : insertTLI = XLogCtl->InsertTimeLineID;
6668 2 : SpinLockRelease(&XLogCtl->info_lck);
6669 :
6670 2 : return insertTLI;
6671 : }
6672 :
6673 : /*
6674 : * GetLastImportantRecPtr -- Returns the LSN of the last important record
6675 : * inserted. All records not explicitly marked as unimportant are considered
6676 : * important.
6677 : *
6678 : * The LSN is determined by computing the maximum of
6679 : * WALInsertLocks[i].lastImportantAt.
6680 : */
6681 : XLogRecPtr
6682 3296 : GetLastImportantRecPtr(void)
6683 : {
6684 3296 : XLogRecPtr res = InvalidXLogRecPtr;
6685 : int i;
6686 :
6687 29664 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
6688 : {
6689 : XLogRecPtr last_important;
6690 :
6691 : /*
6692 : * Need to take a lock to prevent torn reads of the LSN, which are
6693 : * possible on some of the supported platforms. WAL insert locks only
6694 : * support exclusive mode, so we have to use that.
6695 : */
6696 26368 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
6697 26368 : last_important = WALInsertLocks[i].l.lastImportantAt;
6698 26368 : LWLockRelease(&WALInsertLocks[i].l.lock);
6699 :
6700 26368 : if (res < last_important)
6701 5716 : res = last_important;
6702 : }
6703 :
6704 3296 : return res;
6705 : }
6706 :
6707 : /*
6708 : * Get the time and LSN of the last xlog segment switch
6709 : */
6710 : pg_time_t
6711 0 : GetLastSegSwitchData(XLogRecPtr *lastSwitchLSN)
6712 : {
6713 : pg_time_t result;
6714 :
6715 : /* Need WALWriteLock, but shared lock is sufficient */
6716 0 : LWLockAcquire(WALWriteLock, LW_SHARED);
6717 0 : result = XLogCtl->lastSegSwitchTime;
6718 0 : *lastSwitchLSN = XLogCtl->lastSegSwitchLSN;
6719 0 : LWLockRelease(WALWriteLock);
6720 :
6721 0 : return result;
6722 : }
6723 :
6724 : /*
6725 : * This must be called ONCE during postmaster or standalone-backend shutdown
6726 : */
6727 : void
6728 1372 : ShutdownXLOG(int code, Datum arg)
6729 : {
6730 : /*
6731 : * We should have an aux process resource owner to use, and we should not
6732 : * be in a transaction that's installed some other resowner.
6733 : */
6734 : Assert(AuxProcessResourceOwner != NULL);
6735 : Assert(CurrentResourceOwner == NULL ||
6736 : CurrentResourceOwner == AuxProcessResourceOwner);
6737 1372 : CurrentResourceOwner = AuxProcessResourceOwner;
6738 :
6739 : /* Don't be chatty in standalone mode */
6740 1372 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
6741 : (errmsg("shutting down")));
6742 :
6743 : /*
6744 : * Signal walsenders to move to stopping state.
6745 : */
6746 1372 : WalSndInitStopping();
6747 :
6748 : /*
6749 : * Wait for WAL senders to be in stopping state. This prevents commands
6750 : * from writing new WAL.
6751 : */
6752 1372 : WalSndWaitStopping();
6753 :
6754 1372 : if (RecoveryInProgress())
6755 116 : CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
6756 : else
6757 : {
6758 : /*
6759 : * If archiving is enabled, rotate the last XLOG file so that all the
6760 : * remaining records are archived (postmaster wakes up the archiver
6761 : * process one more time at the end of shutdown). The checkpoint
6762 : * record will go to the next XLOG file and won't be archived (yet).
6763 : */
6764 1256 : if (XLogArchivingActive())
6765 32 : RequestXLogSwitch(false);
6766 :
6767 1256 : CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
6768 : }
6769 1372 : }
6770 :
6771 : /*
6772 : * Log start of a checkpoint.
6773 : */
6774 : static void
6775 2990 : LogCheckpointStart(int flags, bool restartpoint)
6776 : {
6777 2990 : if (restartpoint)
6778 408 : ereport(LOG,
6779 : /* translator: the placeholders show checkpoint options */
6780 : (errmsg("restartpoint starting:%s%s%s%s%s%s%s%s",
6781 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6782 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6783 : (flags & CHECKPOINT_FAST) ? " fast" : "",
6784 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6785 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6786 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6787 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6788 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "")));
6789 : else
6790 2582 : ereport(LOG,
6791 : /* translator: the placeholders show checkpoint options */
6792 : (errmsg("checkpoint starting:%s%s%s%s%s%s%s%s",
6793 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6794 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6795 : (flags & CHECKPOINT_FAST) ? " fast" : "",
6796 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6797 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6798 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6799 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6800 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "")));
6801 2990 : }
6802 :
6803 : /*
6804 : * Log end of a checkpoint.
6805 : */
6806 : static void
6807 3586 : LogCheckpointEnd(bool restartpoint)
6808 : {
6809 : long write_msecs,
6810 : sync_msecs,
6811 : total_msecs,
6812 : longest_msecs,
6813 : average_msecs;
6814 : uint64 average_sync_time;
6815 :
6816 3586 : CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
6817 :
6818 3586 : write_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_write_t,
6819 : CheckpointStats.ckpt_sync_t);
6820 :
6821 3586 : sync_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_sync_t,
6822 : CheckpointStats.ckpt_sync_end_t);
6823 :
6824 : /* Accumulate checkpoint timing summary data, in milliseconds. */
6825 3586 : PendingCheckpointerStats.write_time += write_msecs;
6826 3586 : PendingCheckpointerStats.sync_time += sync_msecs;
6827 :
6828 : /*
6829 : * All of the published timing statistics are accounted for. Only
6830 : * continue if a log message is to be written.
6831 : */
6832 3586 : if (!log_checkpoints)
6833 596 : return;
6834 :
6835 2990 : total_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_start_t,
6836 : CheckpointStats.ckpt_end_t);
6837 :
6838 : /*
6839 : * Timing values returned from CheckpointStats are in microseconds.
6840 : * Convert to milliseconds for consistent printing.
6841 : */
6842 2990 : longest_msecs = (long) ((CheckpointStats.ckpt_longest_sync + 999) / 1000);
6843 :
6844 2990 : average_sync_time = 0;
6845 2990 : if (CheckpointStats.ckpt_sync_rels > 0)
6846 0 : average_sync_time = CheckpointStats.ckpt_agg_sync_time /
6847 0 : CheckpointStats.ckpt_sync_rels;
6848 2990 : average_msecs = (long) ((average_sync_time + 999) / 1000);
6849 :
6850 : /*
6851 : * ControlFileLock is not required to see ControlFile->checkPoint and
6852 : * ->checkPointCopy here as we are the only updator of those variables at
6853 : * this moment.
6854 : */
6855 2990 : if (restartpoint)
6856 408 : ereport(LOG,
6857 : (errmsg("restartpoint complete: wrote %d buffers (%.1f%%), "
6858 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6859 : "%d removed, %d recycled; write=%ld.%03d s, "
6860 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6861 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6862 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6863 : CheckpointStats.ckpt_bufs_written,
6864 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6865 : CheckpointStats.ckpt_slru_written,
6866 : CheckpointStats.ckpt_segs_added,
6867 : CheckpointStats.ckpt_segs_removed,
6868 : CheckpointStats.ckpt_segs_recycled,
6869 : write_msecs / 1000, (int) (write_msecs % 1000),
6870 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6871 : total_msecs / 1000, (int) (total_msecs % 1000),
6872 : CheckpointStats.ckpt_sync_rels,
6873 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6874 : average_msecs / 1000, (int) (average_msecs % 1000),
6875 : (int) (PrevCheckPointDistance / 1024.0),
6876 : (int) (CheckPointDistanceEstimate / 1024.0),
6877 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6878 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6879 : else
6880 2582 : ereport(LOG,
6881 : (errmsg("checkpoint complete: wrote %d buffers (%.1f%%), "
6882 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6883 : "%d removed, %d recycled; write=%ld.%03d s, "
6884 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6885 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6886 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6887 : CheckpointStats.ckpt_bufs_written,
6888 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6889 : CheckpointStats.ckpt_slru_written,
6890 : CheckpointStats.ckpt_segs_added,
6891 : CheckpointStats.ckpt_segs_removed,
6892 : CheckpointStats.ckpt_segs_recycled,
6893 : write_msecs / 1000, (int) (write_msecs % 1000),
6894 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6895 : total_msecs / 1000, (int) (total_msecs % 1000),
6896 : CheckpointStats.ckpt_sync_rels,
6897 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6898 : average_msecs / 1000, (int) (average_msecs % 1000),
6899 : (int) (PrevCheckPointDistance / 1024.0),
6900 : (int) (CheckPointDistanceEstimate / 1024.0),
6901 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6902 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6903 : }
6904 :
6905 : /*
6906 : * Update the estimate of distance between checkpoints.
6907 : *
6908 : * The estimate is used to calculate the number of WAL segments to keep
6909 : * preallocated, see XLOGfileslop().
6910 : */
6911 : static void
6912 3586 : UpdateCheckPointDistanceEstimate(uint64 nbytes)
6913 : {
6914 : /*
6915 : * To estimate the number of segments consumed between checkpoints, keep a
6916 : * moving average of the amount of WAL generated in previous checkpoint
6917 : * cycles. However, if the load is bursty, with quiet periods and busy
6918 : * periods, we want to cater for the peak load. So instead of a plain
6919 : * moving average, let the average decline slowly if the previous cycle
6920 : * used less WAL than estimated, but bump it up immediately if it used
6921 : * more.
6922 : *
6923 : * When checkpoints are triggered by max_wal_size, this should converge to
6924 : * CheckpointSegments * wal_segment_size,
6925 : *
6926 : * Note: This doesn't pay any attention to what caused the checkpoint.
6927 : * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
6928 : * starting a base backup, are counted the same as those created
6929 : * automatically. The slow-decline will largely mask them out, if they are
6930 : * not frequent. If they are frequent, it seems reasonable to count them
6931 : * in as any others; if you issue a manual checkpoint every 5 minutes and
6932 : * never let a timed checkpoint happen, it makes sense to base the
6933 : * preallocation on that 5 minute interval rather than whatever
6934 : * checkpoint_timeout is set to.
6935 : */
6936 3586 : PrevCheckPointDistance = nbytes;
6937 3586 : if (CheckPointDistanceEstimate < nbytes)
6938 1550 : CheckPointDistanceEstimate = nbytes;
6939 : else
6940 2036 : CheckPointDistanceEstimate =
6941 2036 : (0.90 * CheckPointDistanceEstimate + 0.10 * (double) nbytes);
6942 3586 : }
6943 :
6944 : /*
6945 : * Update the ps display for a process running a checkpoint. Note that
6946 : * this routine should not do any allocations so as it can be called
6947 : * from a critical section.
6948 : */
6949 : static void
6950 7172 : update_checkpoint_display(int flags, bool restartpoint, bool reset)
6951 : {
6952 : /*
6953 : * The status is reported only for end-of-recovery and shutdown
6954 : * checkpoints or shutdown restartpoints. Updating the ps display is
6955 : * useful in those situations as it may not be possible to rely on
6956 : * pg_stat_activity to see the status of the checkpointer or the startup
6957 : * process.
6958 : */
6959 7172 : if ((flags & (CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IS_SHUTDOWN)) == 0)
6960 4460 : return;
6961 :
6962 2712 : if (reset)
6963 1356 : set_ps_display("");
6964 : else
6965 : {
6966 : char activitymsg[128];
6967 :
6968 4068 : snprintf(activitymsg, sizeof(activitymsg), "performing %s%s%s",
6969 1356 : (flags & CHECKPOINT_END_OF_RECOVERY) ? "end-of-recovery " : "",
6970 1356 : (flags & CHECKPOINT_IS_SHUTDOWN) ? "shutdown " : "",
6971 : restartpoint ? "restartpoint" : "checkpoint");
6972 1356 : set_ps_display(activitymsg);
6973 : }
6974 : }
6975 :
6976 :
6977 : /*
6978 : * Perform a checkpoint --- either during shutdown, or on-the-fly
6979 : *
6980 : * flags is a bitwise OR of the following:
6981 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
6982 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
6983 : * CHECKPOINT_FAST: finish the checkpoint ASAP, ignoring
6984 : * checkpoint_completion_target parameter.
6985 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
6986 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
6987 : * CHECKPOINT_END_OF_RECOVERY).
6988 : * CHECKPOINT_FLUSH_UNLOGGED: also flush buffers of unlogged tables.
6989 : *
6990 : * Note: flags contains other bits, of interest here only for logging purposes.
6991 : * In particular note that this routine is synchronous and does not pay
6992 : * attention to CHECKPOINT_WAIT.
6993 : *
6994 : * If !shutdown then we are writing an online checkpoint. An XLOG_CHECKPOINT_REDO
6995 : * record is inserted into WAL at the logical location of the checkpoint, before
6996 : * flushing anything to disk, and when the checkpoint is eventually completed,
6997 : * and it is from this point that WAL replay will begin in the case of a recovery
6998 : * from this checkpoint. Once everything is written to disk, an
6999 : * XLOG_CHECKPOINT_ONLINE record is written to complete the checkpoint, and
7000 : * points back to the earlier XLOG_CHECKPOINT_REDO record. This mechanism allows
7001 : * other write-ahead log records to be written while the checkpoint is in
7002 : * progress, but we must be very careful about order of operations. This function
7003 : * may take many minutes to execute on a busy system.
7004 : *
7005 : * On the other hand, when shutdown is true, concurrent insertion into the
7006 : * write-ahead log is impossible, so there is no need for two separate records.
7007 : * In this case, we only insert an XLOG_CHECKPOINT_SHUTDOWN record, and it's
7008 : * both the record marking the completion of the checkpoint and the location
7009 : * from which WAL replay would begin if needed.
7010 : *
7011 : * Returns true if a new checkpoint was performed, or false if it was skipped
7012 : * because the system was idle.
7013 : */
7014 : bool
7015 3182 : CreateCheckPoint(int flags)
7016 : {
7017 : bool shutdown;
7018 : CheckPoint checkPoint;
7019 : XLogRecPtr recptr;
7020 : XLogSegNo _logSegNo;
7021 3182 : XLogCtlInsert *Insert = &XLogCtl->Insert;
7022 : uint32 freespace;
7023 : XLogRecPtr PriorRedoPtr;
7024 : XLogRecPtr last_important_lsn;
7025 : VirtualTransactionId *vxids;
7026 : int nvxids;
7027 3182 : int oldXLogAllowed = 0;
7028 :
7029 : /*
7030 : * An end-of-recovery checkpoint is really a shutdown checkpoint, just
7031 : * issued at a different time.
7032 : */
7033 3182 : if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
7034 1314 : shutdown = true;
7035 : else
7036 1868 : shutdown = false;
7037 :
7038 : /* sanity check */
7039 3182 : if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
7040 0 : elog(ERROR, "can't create a checkpoint during recovery");
7041 :
7042 : /*
7043 : * Prepare to accumulate statistics.
7044 : *
7045 : * Note: because it is possible for log_checkpoints to change while a
7046 : * checkpoint proceeds, we always accumulate stats, even if
7047 : * log_checkpoints is currently off.
7048 : */
7049 35002 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7050 3182 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7051 :
7052 : /*
7053 : * Let smgr prepare for checkpoint; this has to happen outside the
7054 : * critical section and before we determine the REDO pointer. Note that
7055 : * smgr must not do anything that'd have to be undone if we decide no
7056 : * checkpoint is needed.
7057 : */
7058 3182 : SyncPreCheckpoint();
7059 :
7060 : /* Run these points outside the critical section. */
7061 3182 : INJECTION_POINT("create-checkpoint-initial", NULL);
7062 3182 : INJECTION_POINT_LOAD("create-checkpoint-run");
7063 :
7064 : /*
7065 : * Use a critical section to force system panic if we have trouble.
7066 : */
7067 3182 : START_CRIT_SECTION();
7068 :
7069 3182 : if (shutdown)
7070 : {
7071 1314 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7072 1314 : ControlFile->state = DB_SHUTDOWNING;
7073 1314 : UpdateControlFile();
7074 1314 : LWLockRelease(ControlFileLock);
7075 : }
7076 :
7077 : /* Begin filling in the checkpoint WAL record */
7078 41366 : MemSet(&checkPoint, 0, sizeof(checkPoint));
7079 3182 : checkPoint.time = (pg_time_t) time(NULL);
7080 :
7081 : /*
7082 : * For Hot Standby, derive the oldestActiveXid before we fix the redo
7083 : * pointer. This allows us to begin accumulating changes to assemble our
7084 : * starting snapshot of locks and transactions.
7085 : */
7086 3182 : if (!shutdown && XLogStandbyInfoActive())
7087 1812 : checkPoint.oldestActiveXid = GetOldestActiveTransactionId(false, true);
7088 : else
7089 1370 : checkPoint.oldestActiveXid = InvalidTransactionId;
7090 :
7091 : /*
7092 : * Get location of last important record before acquiring insert locks (as
7093 : * GetLastImportantRecPtr() also locks WAL locks).
7094 : */
7095 3182 : last_important_lsn = GetLastImportantRecPtr();
7096 :
7097 : /*
7098 : * If this isn't a shutdown or forced checkpoint, and if there has been no
7099 : * WAL activity requiring a checkpoint, skip it. The idea here is to
7100 : * avoid inserting duplicate checkpoints when the system is idle.
7101 : */
7102 3182 : if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
7103 : CHECKPOINT_FORCE)) == 0)
7104 : {
7105 394 : if (last_important_lsn == ControlFile->checkPoint)
7106 : {
7107 4 : END_CRIT_SECTION();
7108 4 : ereport(DEBUG1,
7109 : (errmsg_internal("checkpoint skipped because system is idle")));
7110 4 : return false;
7111 : }
7112 : }
7113 :
7114 : /*
7115 : * An end-of-recovery checkpoint is created before anyone is allowed to
7116 : * write WAL. To allow us to write the checkpoint record, temporarily
7117 : * enable XLogInsertAllowed.
7118 : */
7119 3178 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7120 58 : oldXLogAllowed = LocalSetXLogInsertAllowed();
7121 :
7122 3178 : checkPoint.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7123 3178 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7124 58 : checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7125 : else
7126 3120 : checkPoint.PrevTimeLineID = checkPoint.ThisTimeLineID;
7127 :
7128 : /*
7129 : * We must block concurrent insertions while examining insert state.
7130 : */
7131 3178 : WALInsertLockAcquireExclusive();
7132 :
7133 3178 : checkPoint.fullPageWrites = Insert->fullPageWrites;
7134 3178 : checkPoint.wal_level = wal_level;
7135 :
7136 3178 : if (shutdown)
7137 : {
7138 1314 : XLogRecPtr curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
7139 :
7140 : /*
7141 : * Compute new REDO record ptr = location of next XLOG record.
7142 : *
7143 : * Since this is a shutdown checkpoint, there can't be any concurrent
7144 : * WAL insertion.
7145 : */
7146 1314 : freespace = INSERT_FREESPACE(curInsert);
7147 1314 : if (freespace == 0)
7148 : {
7149 0 : if (XLogSegmentOffset(curInsert, wal_segment_size) == 0)
7150 0 : curInsert += SizeOfXLogLongPHD;
7151 : else
7152 0 : curInsert += SizeOfXLogShortPHD;
7153 : }
7154 1314 : checkPoint.redo = curInsert;
7155 :
7156 : /*
7157 : * Here we update the shared RedoRecPtr for future XLogInsert calls;
7158 : * this must be done while holding all the insertion locks.
7159 : *
7160 : * Note: if we fail to complete the checkpoint, RedoRecPtr will be
7161 : * left pointing past where it really needs to point. This is okay;
7162 : * the only consequence is that XLogInsert might back up whole buffers
7163 : * that it didn't really need to. We can't postpone advancing
7164 : * RedoRecPtr because XLogInserts that happen while we are dumping
7165 : * buffers must assume that their buffer changes are not included in
7166 : * the checkpoint.
7167 : */
7168 1314 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
7169 : }
7170 :
7171 : /*
7172 : * Now we can release the WAL insertion locks, allowing other xacts to
7173 : * proceed while we are flushing disk buffers.
7174 : */
7175 3178 : WALInsertLockRelease();
7176 :
7177 : /*
7178 : * If this is an online checkpoint, we have not yet determined the redo
7179 : * point. We do so now by inserting the special XLOG_CHECKPOINT_REDO
7180 : * record; the LSN at which it starts becomes the new redo pointer. We
7181 : * don't do this for a shutdown checkpoint, because in that case no WAL
7182 : * can be written between the redo point and the insertion of the
7183 : * checkpoint record itself, so the checkpoint record itself serves to
7184 : * mark the redo point.
7185 : */
7186 3178 : if (!shutdown)
7187 : {
7188 : /* Include WAL level in record for WAL summarizer's benefit. */
7189 1864 : XLogBeginInsert();
7190 1864 : XLogRegisterData(&wal_level, sizeof(wal_level));
7191 1864 : (void) XLogInsert(RM_XLOG_ID, XLOG_CHECKPOINT_REDO);
7192 :
7193 : /*
7194 : * XLogInsertRecord will have updated XLogCtl->Insert.RedoRecPtr in
7195 : * shared memory and RedoRecPtr in backend-local memory, but we need
7196 : * to copy that into the record that will be inserted when the
7197 : * checkpoint is complete.
7198 : */
7199 1864 : checkPoint.redo = RedoRecPtr;
7200 : }
7201 :
7202 : /* Update the info_lck-protected copy of RedoRecPtr as well */
7203 3178 : SpinLockAcquire(&XLogCtl->info_lck);
7204 3178 : XLogCtl->RedoRecPtr = checkPoint.redo;
7205 3178 : SpinLockRelease(&XLogCtl->info_lck);
7206 :
7207 : /*
7208 : * If enabled, log checkpoint start. We postpone this until now so as not
7209 : * to log anything if we decided to skip the checkpoint.
7210 : */
7211 3178 : if (log_checkpoints)
7212 2582 : LogCheckpointStart(flags, false);
7213 :
7214 3178 : INJECTION_POINT_CACHED("create-checkpoint-run", NULL);
7215 :
7216 : /* Update the process title */
7217 3178 : update_checkpoint_display(flags, false, false);
7218 :
7219 : TRACE_POSTGRESQL_CHECKPOINT_START(flags);
7220 :
7221 : /*
7222 : * Get the other info we need for the checkpoint record.
7223 : *
7224 : * We don't need to save oldestClogXid in the checkpoint, it only matters
7225 : * for the short period in which clog is being truncated, and if we crash
7226 : * during that we'll redo the clog truncation and fix up oldestClogXid
7227 : * there.
7228 : */
7229 3178 : LWLockAcquire(XidGenLock, LW_SHARED);
7230 3178 : checkPoint.nextXid = TransamVariables->nextXid;
7231 3178 : checkPoint.oldestXid = TransamVariables->oldestXid;
7232 3178 : checkPoint.oldestXidDB = TransamVariables->oldestXidDB;
7233 3178 : LWLockRelease(XidGenLock);
7234 :
7235 3178 : LWLockAcquire(CommitTsLock, LW_SHARED);
7236 3178 : checkPoint.oldestCommitTsXid = TransamVariables->oldestCommitTsXid;
7237 3178 : checkPoint.newestCommitTsXid = TransamVariables->newestCommitTsXid;
7238 3178 : LWLockRelease(CommitTsLock);
7239 :
7240 3178 : LWLockAcquire(OidGenLock, LW_SHARED);
7241 3178 : checkPoint.nextOid = TransamVariables->nextOid;
7242 3178 : if (!shutdown)
7243 1864 : checkPoint.nextOid += TransamVariables->oidCount;
7244 3178 : LWLockRelease(OidGenLock);
7245 :
7246 3178 : checkPoint.logicalDecodingEnabled = IsLogicalDecodingEnabled();
7247 :
7248 3178 : MultiXactGetCheckptMulti(shutdown,
7249 : &checkPoint.nextMulti,
7250 : &checkPoint.nextMultiOffset,
7251 : &checkPoint.oldestMulti,
7252 : &checkPoint.oldestMultiDB);
7253 :
7254 : /*
7255 : * Having constructed the checkpoint record, ensure all shmem disk buffers
7256 : * and commit-log buffers are flushed to disk.
7257 : *
7258 : * This I/O could fail for various reasons. If so, we will fail to
7259 : * complete the checkpoint, but there is no reason to force a system
7260 : * panic. Accordingly, exit critical section while doing it.
7261 : */
7262 3178 : END_CRIT_SECTION();
7263 :
7264 : /*
7265 : * In some cases there are groups of actions that must all occur on one
7266 : * side or the other of a checkpoint record. Before flushing the
7267 : * checkpoint record we must explicitly wait for any backend currently
7268 : * performing those groups of actions.
7269 : *
7270 : * One example is end of transaction, so we must wait for any transactions
7271 : * that are currently in commit critical sections. If an xact inserted
7272 : * its commit record into XLOG just before the REDO point, then a crash
7273 : * restart from the REDO point would not replay that record, which means
7274 : * that our flushing had better include the xact's update of pg_xact. So
7275 : * we wait till he's out of his commit critical section before proceeding.
7276 : * See notes in RecordTransactionCommit().
7277 : *
7278 : * Because we've already released the insertion locks, this test is a bit
7279 : * fuzzy: it is possible that we will wait for xacts we didn't really need
7280 : * to wait for. But the delay should be short and it seems better to make
7281 : * checkpoint take a bit longer than to hold off insertions longer than
7282 : * necessary. (In fact, the whole reason we have this issue is that xact.c
7283 : * does commit record XLOG insertion and clog update as two separate steps
7284 : * protected by different locks, but again that seems best on grounds of
7285 : * minimizing lock contention.)
7286 : *
7287 : * A transaction that has not yet set delayChkptFlags when we look cannot
7288 : * be at risk, since it has not inserted its commit record yet; and one
7289 : * that's already cleared it is not at risk either, since it's done fixing
7290 : * clog and we will correctly flush the update below. So we cannot miss
7291 : * any xacts we need to wait for.
7292 : */
7293 3178 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
7294 3178 : if (nvxids > 0)
7295 : {
7296 : do
7297 : {
7298 : /*
7299 : * Keep absorbing fsync requests while we wait. There could even
7300 : * be a deadlock if we don't, if the process that prevents the
7301 : * checkpoint is trying to add a request to the queue.
7302 : */
7303 116 : AbsorbSyncRequests();
7304 :
7305 116 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
7306 116 : pg_usleep(10000L); /* wait for 10 msec */
7307 116 : pgstat_report_wait_end();
7308 116 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7309 : DELAY_CHKPT_START));
7310 : }
7311 3178 : pfree(vxids);
7312 :
7313 3178 : CheckPointGuts(checkPoint.redo, flags);
7314 :
7315 3178 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
7316 3178 : if (nvxids > 0)
7317 : {
7318 : do
7319 : {
7320 0 : AbsorbSyncRequests();
7321 :
7322 0 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_COMPLETE);
7323 0 : pg_usleep(10000L); /* wait for 10 msec */
7324 0 : pgstat_report_wait_end();
7325 0 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7326 : DELAY_CHKPT_COMPLETE));
7327 : }
7328 3178 : pfree(vxids);
7329 :
7330 : /*
7331 : * Take a snapshot of running transactions and write this to WAL. This
7332 : * allows us to reconstruct the state of running transactions during
7333 : * archive recovery, if required. Skip, if this info disabled.
7334 : *
7335 : * If we are shutting down, or Startup process is completing crash
7336 : * recovery we don't need to write running xact data.
7337 : */
7338 3178 : if (!shutdown && XLogStandbyInfoActive())
7339 1808 : LogStandbySnapshot();
7340 :
7341 3178 : START_CRIT_SECTION();
7342 :
7343 : /*
7344 : * Now insert the checkpoint record into XLOG.
7345 : */
7346 3178 : XLogBeginInsert();
7347 3178 : XLogRegisterData(&checkPoint, sizeof(checkPoint));
7348 3178 : recptr = XLogInsert(RM_XLOG_ID,
7349 : shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
7350 : XLOG_CHECKPOINT_ONLINE);
7351 :
7352 3178 : XLogFlush(recptr);
7353 :
7354 : /*
7355 : * We mustn't write any new WAL after a shutdown checkpoint, or it will be
7356 : * overwritten at next startup. No-one should even try, this just allows
7357 : * sanity-checking. In the case of an end-of-recovery checkpoint, we want
7358 : * to just temporarily disable writing until the system has exited
7359 : * recovery.
7360 : */
7361 3178 : if (shutdown)
7362 : {
7363 1314 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7364 58 : LocalXLogInsertAllowed = oldXLogAllowed;
7365 : else
7366 1256 : LocalXLogInsertAllowed = 0; /* never again write WAL */
7367 : }
7368 :
7369 : /*
7370 : * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
7371 : * = end of actual checkpoint record.
7372 : */
7373 3178 : if (shutdown && checkPoint.redo != ProcLastRecPtr)
7374 0 : ereport(PANIC,
7375 : (errmsg("concurrent write-ahead log activity while database system is shutting down")));
7376 :
7377 : /*
7378 : * Remember the prior checkpoint's redo ptr for
7379 : * UpdateCheckPointDistanceEstimate()
7380 : */
7381 3178 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7382 :
7383 : /*
7384 : * Update the control file.
7385 : */
7386 3178 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7387 3178 : if (shutdown)
7388 1314 : ControlFile->state = DB_SHUTDOWNED;
7389 3178 : ControlFile->checkPoint = ProcLastRecPtr;
7390 3178 : ControlFile->checkPointCopy = checkPoint;
7391 : /* crash recovery should always recover to the end of WAL */
7392 3178 : ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
7393 3178 : ControlFile->minRecoveryPointTLI = 0;
7394 :
7395 : /*
7396 : * Persist unloggedLSN value. It's reset on crash recovery, so this goes
7397 : * unused on non-shutdown checkpoints, but seems useful to store it always
7398 : * for debugging purposes.
7399 : */
7400 3178 : ControlFile->unloggedLSN = pg_atomic_read_membarrier_u64(&XLogCtl->unloggedLSN);
7401 :
7402 3178 : UpdateControlFile();
7403 3178 : LWLockRelease(ControlFileLock);
7404 :
7405 : /*
7406 : * We are now done with critical updates; no need for system panic if we
7407 : * have trouble while fooling with old log segments.
7408 : */
7409 3178 : END_CRIT_SECTION();
7410 :
7411 : /*
7412 : * WAL summaries end when the next XLOG_CHECKPOINT_REDO or
7413 : * XLOG_CHECKPOINT_SHUTDOWN record is reached. This is the first point
7414 : * where (a) we're not inside of a critical section and (b) we can be
7415 : * certain that the relevant record has been flushed to disk, which must
7416 : * happen before it can be summarized.
7417 : *
7418 : * If this is a shutdown checkpoint, then this happens reasonably
7419 : * promptly: we've only just inserted and flushed the
7420 : * XLOG_CHECKPOINT_SHUTDOWN record. If this is not a shutdown checkpoint,
7421 : * then this might not be very prompt at all: the XLOG_CHECKPOINT_REDO
7422 : * record was written before we began flushing data to disk, and that
7423 : * could be many minutes ago at this point. However, we don't XLogFlush()
7424 : * after inserting that record, so we're not guaranteed that it's on disk
7425 : * until after the above call that flushes the XLOG_CHECKPOINT_ONLINE
7426 : * record.
7427 : */
7428 3178 : WakeupWalSummarizer();
7429 :
7430 : /*
7431 : * Let smgr do post-checkpoint cleanup (eg, deleting old files).
7432 : */
7433 3178 : SyncPostCheckpoint();
7434 :
7435 : /*
7436 : * Update the average distance between checkpoints if the prior checkpoint
7437 : * exists.
7438 : */
7439 3178 : if (XLogRecPtrIsValid(PriorRedoPtr))
7440 3178 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7441 :
7442 3178 : INJECTION_POINT("checkpoint-before-old-wal-removal", NULL);
7443 :
7444 : /*
7445 : * Delete old log files, those no longer needed for last checkpoint to
7446 : * prevent the disk holding the xlog from growing full.
7447 : */
7448 3178 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7449 3178 : KeepLogSeg(recptr, &_logSegNo);
7450 3178 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7451 : _logSegNo, InvalidOid,
7452 : InvalidTransactionId))
7453 : {
7454 : /*
7455 : * Some slots have been invalidated; recalculate the old-segment
7456 : * horizon, starting again from RedoRecPtr.
7457 : */
7458 8 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7459 8 : KeepLogSeg(recptr, &_logSegNo);
7460 : }
7461 3178 : _logSegNo--;
7462 3178 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
7463 : checkPoint.ThisTimeLineID);
7464 :
7465 : /*
7466 : * Make more log segments if needed. (Do this after recycling old log
7467 : * segments, since that may supply some of the needed files.)
7468 : */
7469 3178 : if (!shutdown)
7470 1864 : PreallocXlogFiles(recptr, checkPoint.ThisTimeLineID);
7471 :
7472 : /*
7473 : * Truncate pg_subtrans if possible. We can throw away all data before
7474 : * the oldest XMIN of any running transaction. No future transaction will
7475 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7476 : * in subtrans.c). During recovery, though, we mustn't do this because
7477 : * StartupSUBTRANS hasn't been called yet.
7478 : */
7479 3178 : if (!RecoveryInProgress())
7480 3120 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7481 :
7482 : /* Real work is done; log and update stats. */
7483 3178 : LogCheckpointEnd(false);
7484 :
7485 : /* Reset the process title */
7486 3178 : update_checkpoint_display(flags, false, true);
7487 :
7488 : TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
7489 : NBuffers,
7490 : CheckpointStats.ckpt_segs_added,
7491 : CheckpointStats.ckpt_segs_removed,
7492 : CheckpointStats.ckpt_segs_recycled);
7493 :
7494 3178 : return true;
7495 : }
7496 :
7497 : /*
7498 : * Mark the end of recovery in WAL though without running a full checkpoint.
7499 : * We can expect that a restartpoint is likely to be in progress as we
7500 : * do this, though we are unwilling to wait for it to complete.
7501 : *
7502 : * CreateRestartPoint() allows for the case where recovery may end before
7503 : * the restartpoint completes so there is no concern of concurrent behaviour.
7504 : */
7505 : static void
7506 96 : CreateEndOfRecoveryRecord(void)
7507 : {
7508 : xl_end_of_recovery xlrec;
7509 : XLogRecPtr recptr;
7510 :
7511 : /* sanity check */
7512 96 : if (!RecoveryInProgress())
7513 0 : elog(ERROR, "can only be used to end recovery");
7514 :
7515 96 : xlrec.end_time = GetCurrentTimestamp();
7516 96 : xlrec.wal_level = wal_level;
7517 :
7518 96 : WALInsertLockAcquireExclusive();
7519 96 : xlrec.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7520 96 : xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7521 96 : WALInsertLockRelease();
7522 :
7523 96 : START_CRIT_SECTION();
7524 :
7525 96 : XLogBeginInsert();
7526 96 : XLogRegisterData(&xlrec, sizeof(xl_end_of_recovery));
7527 96 : recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY);
7528 :
7529 96 : XLogFlush(recptr);
7530 :
7531 : /*
7532 : * Update the control file so that crash recovery can follow the timeline
7533 : * changes to this point.
7534 : */
7535 96 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7536 96 : ControlFile->minRecoveryPoint = recptr;
7537 96 : ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
7538 96 : UpdateControlFile();
7539 96 : LWLockRelease(ControlFileLock);
7540 :
7541 96 : END_CRIT_SECTION();
7542 96 : }
7543 :
7544 : /*
7545 : * Write an OVERWRITE_CONTRECORD message.
7546 : *
7547 : * When on WAL replay we expect a continuation record at the start of a page
7548 : * that is not there, recovery ends and WAL writing resumes at that point.
7549 : * But it's wrong to resume writing new WAL back at the start of the record
7550 : * that was broken, because downstream consumers of that WAL (physical
7551 : * replicas) are not prepared to "rewind". So the first action after
7552 : * finishing replay of all valid WAL must be to write a record of this type
7553 : * at the point where the contrecord was missing; to support xlogreader
7554 : * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
7555 : * to the page header where the record occurs. xlogreader has an ad-hoc
7556 : * mechanism to report metadata about the broken record, which is what we
7557 : * use here.
7558 : *
7559 : * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
7560 : * skip the record it was reading, and pass back the LSN of the skipped
7561 : * record, so that its caller can verify (on "replay" of that record) that the
7562 : * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
7563 : *
7564 : * 'aborted_lsn' is the beginning position of the record that was incomplete.
7565 : * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
7566 : * beginning of the XLOG page where the record is to be inserted. They must
7567 : * match the current WAL insert position, they're passed here just so that we
7568 : * can verify that.
7569 : */
7570 : static XLogRecPtr
7571 22 : CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn, XLogRecPtr pagePtr,
7572 : TimeLineID newTLI)
7573 : {
7574 : xl_overwrite_contrecord xlrec;
7575 : XLogRecPtr recptr;
7576 : XLogPageHeader pagehdr;
7577 : XLogRecPtr startPos;
7578 :
7579 : /* sanity checks */
7580 22 : if (!RecoveryInProgress())
7581 0 : elog(ERROR, "can only be used at end of recovery");
7582 22 : if (pagePtr % XLOG_BLCKSZ != 0)
7583 0 : elog(ERROR, "invalid position for missing continuation record %X/%08X",
7584 : LSN_FORMAT_ARGS(pagePtr));
7585 :
7586 : /* The current WAL insert position should be right after the page header */
7587 22 : startPos = pagePtr;
7588 22 : if (XLogSegmentOffset(startPos, wal_segment_size) == 0)
7589 2 : startPos += SizeOfXLogLongPHD;
7590 : else
7591 20 : startPos += SizeOfXLogShortPHD;
7592 22 : recptr = GetXLogInsertRecPtr();
7593 22 : if (recptr != startPos)
7594 0 : elog(ERROR, "invalid WAL insert position %X/%08X for OVERWRITE_CONTRECORD",
7595 : LSN_FORMAT_ARGS(recptr));
7596 :
7597 22 : START_CRIT_SECTION();
7598 :
7599 : /*
7600 : * Initialize the XLOG page header (by GetXLogBuffer), and set the
7601 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
7602 : *
7603 : * No other backend is allowed to write WAL yet, so acquiring the WAL
7604 : * insertion lock is just pro forma.
7605 : */
7606 22 : WALInsertLockAcquire();
7607 22 : pagehdr = (XLogPageHeader) GetXLogBuffer(pagePtr, newTLI);
7608 22 : pagehdr->xlp_info |= XLP_FIRST_IS_OVERWRITE_CONTRECORD;
7609 22 : WALInsertLockRelease();
7610 :
7611 : /*
7612 : * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
7613 : * page. We know it becomes the first record, because no other backend is
7614 : * allowed to write WAL yet.
7615 : */
7616 22 : XLogBeginInsert();
7617 22 : xlrec.overwritten_lsn = aborted_lsn;
7618 22 : xlrec.overwrite_time = GetCurrentTimestamp();
7619 22 : XLogRegisterData(&xlrec, sizeof(xl_overwrite_contrecord));
7620 22 : recptr = XLogInsert(RM_XLOG_ID, XLOG_OVERWRITE_CONTRECORD);
7621 :
7622 : /* check that the record was inserted to the right place */
7623 22 : if (ProcLastRecPtr != startPos)
7624 0 : elog(ERROR, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%08X",
7625 : LSN_FORMAT_ARGS(ProcLastRecPtr));
7626 :
7627 22 : XLogFlush(recptr);
7628 :
7629 22 : END_CRIT_SECTION();
7630 :
7631 22 : return recptr;
7632 : }
7633 :
7634 : /*
7635 : * Flush all data in shared memory to disk, and fsync
7636 : *
7637 : * This is the common code shared between regular checkpoints and
7638 : * recovery restartpoints.
7639 : */
7640 : static void
7641 3586 : CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
7642 : {
7643 3586 : CheckPointRelationMap();
7644 3586 : CheckPointReplicationSlots(flags & CHECKPOINT_IS_SHUTDOWN);
7645 3586 : CheckPointSnapBuild();
7646 3586 : CheckPointLogicalRewriteHeap();
7647 3586 : CheckPointReplicationOrigin();
7648 :
7649 : /* Write out all dirty data in SLRUs and the main buffer pool */
7650 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags);
7651 3586 : CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
7652 3586 : CheckPointCLOG();
7653 3586 : CheckPointCommitTs();
7654 3586 : CheckPointSUBTRANS();
7655 3586 : CheckPointMultiXact();
7656 3586 : CheckPointPredicate();
7657 3586 : CheckPointBuffers(flags);
7658 :
7659 : /* Perform all queued up fsyncs */
7660 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
7661 3586 : CheckpointStats.ckpt_sync_t = GetCurrentTimestamp();
7662 3586 : ProcessSyncRequests();
7663 3586 : CheckpointStats.ckpt_sync_end_t = GetCurrentTimestamp();
7664 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
7665 :
7666 : /* We deliberately delay 2PC checkpointing as long as possible */
7667 3586 : CheckPointTwoPhase(checkPointRedo);
7668 3586 : }
7669 :
7670 : /*
7671 : * Save a checkpoint for recovery restart if appropriate
7672 : *
7673 : * This function is called each time a checkpoint record is read from XLOG.
7674 : * It must determine whether the checkpoint represents a safe restartpoint or
7675 : * not. If so, the checkpoint record is stashed in shared memory so that
7676 : * CreateRestartPoint can consult it. (Note that the latter function is
7677 : * executed by the checkpointer, while this one will be executed by the
7678 : * startup process.)
7679 : */
7680 : static void
7681 1468 : RecoveryRestartPoint(const CheckPoint *checkPoint, XLogReaderState *record)
7682 : {
7683 : /*
7684 : * Also refrain from creating a restartpoint if we have seen any
7685 : * references to non-existent pages. Restarting recovery from the
7686 : * restartpoint would not see the references, so we would lose the
7687 : * cross-check that the pages belonged to a relation that was dropped
7688 : * later.
7689 : */
7690 1468 : if (XLogHaveInvalidPages())
7691 : {
7692 0 : elog(DEBUG2,
7693 : "could not record restart point at %X/%08X because there are unresolved references to invalid pages",
7694 : LSN_FORMAT_ARGS(checkPoint->redo));
7695 0 : return;
7696 : }
7697 :
7698 : /*
7699 : * Copy the checkpoint record to shared memory, so that checkpointer can
7700 : * work out the next time it wants to perform a restartpoint.
7701 : */
7702 1468 : SpinLockAcquire(&XLogCtl->info_lck);
7703 1468 : XLogCtl->lastCheckPointRecPtr = record->ReadRecPtr;
7704 1468 : XLogCtl->lastCheckPointEndPtr = record->EndRecPtr;
7705 1468 : XLogCtl->lastCheckPoint = *checkPoint;
7706 1468 : SpinLockRelease(&XLogCtl->info_lck);
7707 : }
7708 :
7709 : /*
7710 : * Establish a restartpoint if possible.
7711 : *
7712 : * This is similar to CreateCheckPoint, but is used during WAL recovery
7713 : * to establish a point from which recovery can roll forward without
7714 : * replaying the entire recovery log.
7715 : *
7716 : * Returns true if a new restartpoint was established. We can only establish
7717 : * a restartpoint if we have replayed a safe checkpoint record since last
7718 : * restartpoint.
7719 : */
7720 : bool
7721 1244 : CreateRestartPoint(int flags)
7722 : {
7723 : XLogRecPtr lastCheckPointRecPtr;
7724 : XLogRecPtr lastCheckPointEndPtr;
7725 : CheckPoint lastCheckPoint;
7726 : XLogRecPtr PriorRedoPtr;
7727 : XLogRecPtr receivePtr;
7728 : XLogRecPtr replayPtr;
7729 : TimeLineID replayTLI;
7730 : XLogRecPtr endptr;
7731 : XLogSegNo _logSegNo;
7732 : TimestampTz xtime;
7733 :
7734 : /* Concurrent checkpoint/restartpoint cannot happen */
7735 : Assert(!IsUnderPostmaster || MyBackendType == B_CHECKPOINTER);
7736 :
7737 : /* Get a local copy of the last safe checkpoint record. */
7738 1244 : SpinLockAcquire(&XLogCtl->info_lck);
7739 1244 : lastCheckPointRecPtr = XLogCtl->lastCheckPointRecPtr;
7740 1244 : lastCheckPointEndPtr = XLogCtl->lastCheckPointEndPtr;
7741 1244 : lastCheckPoint = XLogCtl->lastCheckPoint;
7742 1244 : SpinLockRelease(&XLogCtl->info_lck);
7743 :
7744 : /*
7745 : * Check that we're still in recovery mode. It's ok if we exit recovery
7746 : * mode after this check, the restart point is valid anyway.
7747 : */
7748 1244 : if (!RecoveryInProgress())
7749 : {
7750 0 : ereport(DEBUG2,
7751 : (errmsg_internal("skipping restartpoint, recovery has already ended")));
7752 0 : return false;
7753 : }
7754 :
7755 : /*
7756 : * If the last checkpoint record we've replayed is already our last
7757 : * restartpoint, we can't perform a new restart point. We still update
7758 : * minRecoveryPoint in that case, so that if this is a shutdown restart
7759 : * point, we won't start up earlier than before. That's not strictly
7760 : * necessary, but when hot standby is enabled, it would be rather weird if
7761 : * the database opened up for read-only connections at a point-in-time
7762 : * before the last shutdown. Such time travel is still possible in case of
7763 : * immediate shutdown, though.
7764 : *
7765 : * We don't explicitly advance minRecoveryPoint when we do create a
7766 : * restartpoint. It's assumed that flushing the buffers will do that as a
7767 : * side-effect.
7768 : */
7769 1244 : if (!XLogRecPtrIsValid(lastCheckPointRecPtr) ||
7770 580 : lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
7771 : {
7772 836 : ereport(DEBUG2,
7773 : errmsg_internal("skipping restartpoint, already performed at %X/%08X",
7774 : LSN_FORMAT_ARGS(lastCheckPoint.redo)));
7775 :
7776 836 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
7777 836 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7778 : {
7779 74 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7780 74 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7781 74 : UpdateControlFile();
7782 74 : LWLockRelease(ControlFileLock);
7783 : }
7784 836 : return false;
7785 : }
7786 :
7787 : /*
7788 : * Update the shared RedoRecPtr so that the startup process can calculate
7789 : * the number of segments replayed since last restartpoint, and request a
7790 : * restartpoint if it exceeds CheckPointSegments.
7791 : *
7792 : * Like in CreateCheckPoint(), hold off insertions to update it, although
7793 : * during recovery this is just pro forma, because no WAL insertions are
7794 : * happening.
7795 : */
7796 408 : WALInsertLockAcquireExclusive();
7797 408 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = lastCheckPoint.redo;
7798 408 : WALInsertLockRelease();
7799 :
7800 : /* Also update the info_lck-protected copy */
7801 408 : SpinLockAcquire(&XLogCtl->info_lck);
7802 408 : XLogCtl->RedoRecPtr = lastCheckPoint.redo;
7803 408 : SpinLockRelease(&XLogCtl->info_lck);
7804 :
7805 : /*
7806 : * Prepare to accumulate statistics.
7807 : *
7808 : * Note: because it is possible for log_checkpoints to change while a
7809 : * checkpoint proceeds, we always accumulate stats, even if
7810 : * log_checkpoints is currently off.
7811 : */
7812 4488 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7813 408 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7814 :
7815 408 : if (log_checkpoints)
7816 408 : LogCheckpointStart(flags, true);
7817 :
7818 : /* Update the process title */
7819 408 : update_checkpoint_display(flags, true, false);
7820 :
7821 408 : CheckPointGuts(lastCheckPoint.redo, flags);
7822 :
7823 : /*
7824 : * This location needs to be after CheckPointGuts() to ensure that some
7825 : * work has already happened during this checkpoint.
7826 : */
7827 408 : INJECTION_POINT("create-restart-point", NULL);
7828 :
7829 : /*
7830 : * Remember the prior checkpoint's redo ptr for
7831 : * UpdateCheckPointDistanceEstimate()
7832 : */
7833 408 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7834 :
7835 : /*
7836 : * Update pg_control, using current time. Check that it still shows an
7837 : * older checkpoint, else do nothing; this is a quick hack to make sure
7838 : * nothing really bad happens if somehow we get here after the
7839 : * end-of-recovery checkpoint.
7840 : */
7841 408 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7842 408 : if (ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
7843 : {
7844 : /*
7845 : * Update the checkpoint information. We do this even if the cluster
7846 : * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
7847 : * segments recycled below.
7848 : */
7849 408 : ControlFile->checkPoint = lastCheckPointRecPtr;
7850 408 : ControlFile->checkPointCopy = lastCheckPoint;
7851 :
7852 : /*
7853 : * Ensure minRecoveryPoint is past the checkpoint record and update it
7854 : * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
7855 : * this will have happened already while writing out dirty buffers,
7856 : * but not necessarily - e.g. because no buffers were dirtied. We do
7857 : * this because a backup performed in recovery uses minRecoveryPoint
7858 : * to determine which WAL files must be included in the backup, and
7859 : * the file (or files) containing the checkpoint record must be
7860 : * included, at a minimum. Note that for an ordinary restart of
7861 : * recovery there's no value in having the minimum recovery point any
7862 : * earlier than this anyway, because redo will begin just after the
7863 : * checkpoint record.
7864 : */
7865 408 : if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
7866 : {
7867 408 : if (ControlFile->minRecoveryPoint < lastCheckPointEndPtr)
7868 : {
7869 38 : ControlFile->minRecoveryPoint = lastCheckPointEndPtr;
7870 38 : ControlFile->minRecoveryPointTLI = lastCheckPoint.ThisTimeLineID;
7871 :
7872 : /* update local copy */
7873 38 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
7874 38 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
7875 : }
7876 408 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7877 42 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7878 : }
7879 408 : UpdateControlFile();
7880 : }
7881 408 : LWLockRelease(ControlFileLock);
7882 :
7883 : /*
7884 : * Update the average distance between checkpoints/restartpoints if the
7885 : * prior checkpoint exists.
7886 : */
7887 408 : if (XLogRecPtrIsValid(PriorRedoPtr))
7888 408 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7889 :
7890 : /*
7891 : * Delete old log files, those no longer needed for last restartpoint to
7892 : * prevent the disk holding the xlog from growing full.
7893 : */
7894 408 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7895 :
7896 : /*
7897 : * Retreat _logSegNo using the current end of xlog replayed or received,
7898 : * whichever is later.
7899 : */
7900 408 : receivePtr = GetWalRcvFlushRecPtr(NULL, NULL);
7901 408 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
7902 408 : endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
7903 408 : KeepLogSeg(endptr, &_logSegNo);
7904 :
7905 408 : INJECTION_POINT("restartpoint-before-slot-invalidation", NULL);
7906 :
7907 408 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7908 : _logSegNo, InvalidOid,
7909 : InvalidTransactionId))
7910 : {
7911 : /*
7912 : * Some slots have been invalidated; recalculate the old-segment
7913 : * horizon, starting again from RedoRecPtr.
7914 : */
7915 2 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7916 2 : KeepLogSeg(endptr, &_logSegNo);
7917 : }
7918 408 : _logSegNo--;
7919 :
7920 : /*
7921 : * Try to recycle segments on a useful timeline. If we've been promoted
7922 : * since the beginning of this restartpoint, use the new timeline chosen
7923 : * at end of recovery. If we're still in recovery, use the timeline we're
7924 : * currently replaying.
7925 : *
7926 : * There is no guarantee that the WAL segments will be useful on the
7927 : * current timeline; if recovery proceeds to a new timeline right after
7928 : * this, the pre-allocated WAL segments on this timeline will not be used,
7929 : * and will go wasted until recycled on the next restartpoint. We'll live
7930 : * with that.
7931 : */
7932 408 : if (!RecoveryInProgress())
7933 0 : replayTLI = XLogCtl->InsertTimeLineID;
7934 :
7935 408 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, endptr, replayTLI);
7936 :
7937 : /*
7938 : * Make more log segments if needed. (Do this after recycling old log
7939 : * segments, since that may supply some of the needed files.)
7940 : */
7941 408 : PreallocXlogFiles(endptr, replayTLI);
7942 :
7943 : /*
7944 : * Truncate pg_subtrans if possible. We can throw away all data before
7945 : * the oldest XMIN of any running transaction. No future transaction will
7946 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7947 : * in subtrans.c). When hot standby is disabled, though, we mustn't do
7948 : * this because StartupSUBTRANS hasn't been called yet.
7949 : */
7950 408 : if (EnableHotStandby)
7951 408 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7952 :
7953 : /* Real work is done; log and update stats. */
7954 408 : LogCheckpointEnd(true);
7955 :
7956 : /* Reset the process title */
7957 408 : update_checkpoint_display(flags, true, true);
7958 :
7959 408 : xtime = GetLatestXTime();
7960 408 : ereport((log_checkpoints ? LOG : DEBUG2),
7961 : errmsg("recovery restart point at %X/%08X",
7962 : LSN_FORMAT_ARGS(lastCheckPoint.redo)),
7963 : xtime ? errdetail("Last completed transaction was at log time %s.",
7964 : timestamptz_to_str(xtime)) : 0);
7965 :
7966 : /*
7967 : * Finally, execute archive_cleanup_command, if any.
7968 : */
7969 408 : if (archiveCleanupCommand && strcmp(archiveCleanupCommand, "") != 0)
7970 0 : ExecuteRecoveryCommand(archiveCleanupCommand,
7971 : "archive_cleanup_command",
7972 : false,
7973 : WAIT_EVENT_ARCHIVE_CLEANUP_COMMAND);
7974 :
7975 408 : return true;
7976 : }
7977 :
7978 : /*
7979 : * Report availability of WAL for the given target LSN
7980 : * (typically a slot's restart_lsn)
7981 : *
7982 : * Returns one of the following enum values:
7983 : *
7984 : * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
7985 : * max_wal_size.
7986 : *
7987 : * * WALAVAIL_EXTENDED means it is still available by preserving extra
7988 : * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
7989 : * than max_wal_size, this state is not returned.
7990 : *
7991 : * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
7992 : * remove reserved segments. The walsender using this slot may return to the
7993 : * above.
7994 : *
7995 : * * WALAVAIL_REMOVED means it has been removed. A replication stream on
7996 : * a slot with this LSN cannot continue. (Any associated walsender
7997 : * processes should have been terminated already.)
7998 : *
7999 : * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
8000 : */
8001 : WALAvailability
8002 1312 : GetWALAvailability(XLogRecPtr targetLSN)
8003 : {
8004 : XLogRecPtr currpos; /* current write LSN */
8005 : XLogSegNo currSeg; /* segid of currpos */
8006 : XLogSegNo targetSeg; /* segid of targetLSN */
8007 : XLogSegNo oldestSeg; /* actual oldest segid */
8008 : XLogSegNo oldestSegMaxWalSize; /* oldest segid kept by max_wal_size */
8009 : XLogSegNo oldestSlotSeg; /* oldest segid kept by slot */
8010 : uint64 keepSegs;
8011 :
8012 : /*
8013 : * slot does not reserve WAL. Either deactivated, or has never been active
8014 : */
8015 1312 : if (!XLogRecPtrIsValid(targetLSN))
8016 74 : return WALAVAIL_INVALID_LSN;
8017 :
8018 : /*
8019 : * Calculate the oldest segment currently reserved by all slots,
8020 : * considering wal_keep_size and max_slot_wal_keep_size. Initialize
8021 : * oldestSlotSeg to the current segment.
8022 : */
8023 1238 : currpos = GetXLogWriteRecPtr();
8024 1238 : XLByteToSeg(currpos, oldestSlotSeg, wal_segment_size);
8025 1238 : KeepLogSeg(currpos, &oldestSlotSeg);
8026 :
8027 : /*
8028 : * Find the oldest extant segment file. We get 1 until checkpoint removes
8029 : * the first WAL segment file since startup, which causes the status being
8030 : * wrong under certain abnormal conditions but that doesn't actually harm.
8031 : */
8032 1238 : oldestSeg = XLogGetLastRemovedSegno() + 1;
8033 :
8034 : /* calculate oldest segment by max_wal_size */
8035 1238 : XLByteToSeg(currpos, currSeg, wal_segment_size);
8036 1238 : keepSegs = ConvertToXSegs(max_wal_size_mb, wal_segment_size) + 1;
8037 :
8038 1238 : if (currSeg > keepSegs)
8039 16 : oldestSegMaxWalSize = currSeg - keepSegs;
8040 : else
8041 1222 : oldestSegMaxWalSize = 1;
8042 :
8043 : /* the segment we care about */
8044 1238 : XLByteToSeg(targetLSN, targetSeg, wal_segment_size);
8045 :
8046 : /*
8047 : * No point in returning reserved or extended status values if the
8048 : * targetSeg is known to be lost.
8049 : */
8050 1238 : if (targetSeg >= oldestSlotSeg)
8051 : {
8052 : /* show "reserved" when targetSeg is within max_wal_size */
8053 1236 : if (targetSeg >= oldestSegMaxWalSize)
8054 1232 : return WALAVAIL_RESERVED;
8055 :
8056 : /* being retained by slots exceeding max_wal_size */
8057 4 : return WALAVAIL_EXTENDED;
8058 : }
8059 :
8060 : /* WAL segments are no longer retained but haven't been removed yet */
8061 2 : if (targetSeg >= oldestSeg)
8062 2 : return WALAVAIL_UNRESERVED;
8063 :
8064 : /* Definitely lost */
8065 0 : return WALAVAIL_REMOVED;
8066 : }
8067 :
8068 :
8069 : /*
8070 : * Retreat *logSegNo to the last segment that we need to retain because of
8071 : * either wal_keep_size or replication slots.
8072 : *
8073 : * This is calculated by subtracting wal_keep_size from the given xlog
8074 : * location, recptr and by making sure that that result is below the
8075 : * requirement of replication slots. For the latter criterion we do consider
8076 : * the effects of max_slot_wal_keep_size: reserve at most that much space back
8077 : * from recptr.
8078 : *
8079 : * Note about replication slots: if this function calculates a value
8080 : * that's further ahead than what slots need reserved, then affected
8081 : * slots need to be invalidated and this function invoked again.
8082 : * XXX it might be a good idea to rewrite this function so that
8083 : * invalidation is optionally done here, instead.
8084 : */
8085 : static void
8086 4834 : KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
8087 : {
8088 : XLogSegNo currSegNo;
8089 : XLogSegNo segno;
8090 : XLogRecPtr keep;
8091 :
8092 4834 : XLByteToSeg(recptr, currSegNo, wal_segment_size);
8093 4834 : segno = currSegNo;
8094 :
8095 : /* Calculate how many segments are kept by slots. */
8096 4834 : keep = XLogGetReplicationSlotMinimumLSN();
8097 4834 : if (XLogRecPtrIsValid(keep) && keep < recptr)
8098 : {
8099 1536 : XLByteToSeg(keep, segno, wal_segment_size);
8100 :
8101 : /*
8102 : * Account for max_slot_wal_keep_size to avoid keeping more than
8103 : * configured. However, don't do that during a binary upgrade: if
8104 : * slots were to be invalidated because of this, it would not be
8105 : * possible to preserve logical ones during the upgrade.
8106 : */
8107 1536 : if (max_slot_wal_keep_size_mb >= 0 && !IsBinaryUpgrade)
8108 : {
8109 : uint64 slot_keep_segs;
8110 :
8111 46 : slot_keep_segs =
8112 46 : ConvertToXSegs(max_slot_wal_keep_size_mb, wal_segment_size);
8113 :
8114 46 : if (currSegNo - segno > slot_keep_segs)
8115 12 : segno = currSegNo - slot_keep_segs;
8116 : }
8117 : }
8118 :
8119 : /*
8120 : * If WAL summarization is in use, don't remove WAL that has yet to be
8121 : * summarized.
8122 : */
8123 4834 : keep = GetOldestUnsummarizedLSN(NULL, NULL);
8124 4834 : if (XLogRecPtrIsValid(keep))
8125 : {
8126 : XLogSegNo unsummarized_segno;
8127 :
8128 14 : XLByteToSeg(keep, unsummarized_segno, wal_segment_size);
8129 14 : if (unsummarized_segno < segno)
8130 10 : segno = unsummarized_segno;
8131 : }
8132 :
8133 : /* but, keep at least wal_keep_size if that's set */
8134 4834 : if (wal_keep_size_mb > 0)
8135 : {
8136 : uint64 keep_segs;
8137 :
8138 148 : keep_segs = ConvertToXSegs(wal_keep_size_mb, wal_segment_size);
8139 148 : if (currSegNo - segno < keep_segs)
8140 : {
8141 : /* avoid underflow, don't go below 1 */
8142 148 : if (currSegNo <= keep_segs)
8143 140 : segno = 1;
8144 : else
8145 8 : segno = currSegNo - keep_segs;
8146 : }
8147 : }
8148 :
8149 : /* don't delete WAL segments newer than the calculated segment */
8150 4834 : if (segno < *logSegNo)
8151 742 : *logSegNo = segno;
8152 4834 : }
8153 :
8154 : /*
8155 : * Write a NEXTOID log record
8156 : */
8157 : void
8158 1260 : XLogPutNextOid(Oid nextOid)
8159 : {
8160 1260 : XLogBeginInsert();
8161 1260 : XLogRegisterData(&nextOid, sizeof(Oid));
8162 1260 : (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID);
8163 :
8164 : /*
8165 : * We need not flush the NEXTOID record immediately, because any of the
8166 : * just-allocated OIDs could only reach disk as part of a tuple insert or
8167 : * update that would have its own XLOG record that must follow the NEXTOID
8168 : * record. Therefore, the standard buffer LSN interlock applied to those
8169 : * records will ensure no such OID reaches disk before the NEXTOID record
8170 : * does.
8171 : *
8172 : * Note, however, that the above statement only covers state "within" the
8173 : * database. When we use a generated OID as a file or directory name, we
8174 : * are in a sense violating the basic WAL rule, because that filesystem
8175 : * change may reach disk before the NEXTOID WAL record does. The impact
8176 : * of this is that if a database crash occurs immediately afterward, we
8177 : * might after restart re-generate the same OID and find that it conflicts
8178 : * with the leftover file or directory. But since for safety's sake we
8179 : * always loop until finding a nonconflicting filename, this poses no real
8180 : * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8181 : */
8182 1260 : }
8183 :
8184 : /*
8185 : * Write an XLOG SWITCH record.
8186 : *
8187 : * Here we just blindly issue an XLogInsert request for the record.
8188 : * All the magic happens inside XLogInsert.
8189 : *
8190 : * The return value is either the end+1 address of the switch record,
8191 : * or the end+1 address of the prior segment if we did not need to
8192 : * write a switch record because we are already at segment start.
8193 : */
8194 : XLogRecPtr
8195 1604 : RequestXLogSwitch(bool mark_unimportant)
8196 : {
8197 : XLogRecPtr RecPtr;
8198 :
8199 : /* XLOG SWITCH has no data */
8200 1604 : XLogBeginInsert();
8201 :
8202 1604 : if (mark_unimportant)
8203 0 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8204 1604 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH);
8205 :
8206 1604 : return RecPtr;
8207 : }
8208 :
8209 : /*
8210 : * Write a RESTORE POINT record
8211 : */
8212 : XLogRecPtr
8213 6 : XLogRestorePoint(const char *rpName)
8214 : {
8215 : XLogRecPtr RecPtr;
8216 : xl_restore_point xlrec;
8217 :
8218 6 : xlrec.rp_time = GetCurrentTimestamp();
8219 6 : strlcpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8220 :
8221 6 : XLogBeginInsert();
8222 6 : XLogRegisterData(&xlrec, sizeof(xl_restore_point));
8223 :
8224 6 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT);
8225 :
8226 6 : ereport(LOG,
8227 : errmsg("restore point \"%s\" created at %X/%08X",
8228 : rpName, LSN_FORMAT_ARGS(RecPtr)));
8229 :
8230 6 : return RecPtr;
8231 : }
8232 :
8233 : /*
8234 : * Check if any of the GUC parameters that are critical for hot standby
8235 : * have changed, and update the value in pg_control file if necessary.
8236 : */
8237 : static void
8238 1862 : XLogReportParameters(void)
8239 : {
8240 1862 : if (wal_level != ControlFile->wal_level ||
8241 1378 : wal_log_hints != ControlFile->wal_log_hints ||
8242 1204 : MaxConnections != ControlFile->MaxConnections ||
8243 1202 : max_worker_processes != ControlFile->max_worker_processes ||
8244 1196 : max_wal_senders != ControlFile->max_wal_senders ||
8245 1146 : max_prepared_xacts != ControlFile->max_prepared_xacts ||
8246 946 : max_locks_per_xact != ControlFile->max_locks_per_xact ||
8247 946 : track_commit_timestamp != ControlFile->track_commit_timestamp)
8248 : {
8249 : /*
8250 : * The change in number of backend slots doesn't need to be WAL-logged
8251 : * if archiving is not enabled, as you can't start archive recovery
8252 : * with wal_level=minimal anyway. We don't really care about the
8253 : * values in pg_control either if wal_level=minimal, but seems better
8254 : * to keep them up-to-date to avoid confusion.
8255 : */
8256 940 : if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8257 : {
8258 : xl_parameter_change xlrec;
8259 : XLogRecPtr recptr;
8260 :
8261 890 : xlrec.MaxConnections = MaxConnections;
8262 890 : xlrec.max_worker_processes = max_worker_processes;
8263 890 : xlrec.max_wal_senders = max_wal_senders;
8264 890 : xlrec.max_prepared_xacts = max_prepared_xacts;
8265 890 : xlrec.max_locks_per_xact = max_locks_per_xact;
8266 890 : xlrec.wal_level = wal_level;
8267 890 : xlrec.wal_log_hints = wal_log_hints;
8268 890 : xlrec.track_commit_timestamp = track_commit_timestamp;
8269 :
8270 890 : XLogBeginInsert();
8271 890 : XLogRegisterData(&xlrec, sizeof(xlrec));
8272 :
8273 890 : recptr = XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE);
8274 890 : XLogFlush(recptr);
8275 : }
8276 :
8277 940 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8278 :
8279 940 : ControlFile->MaxConnections = MaxConnections;
8280 940 : ControlFile->max_worker_processes = max_worker_processes;
8281 940 : ControlFile->max_wal_senders = max_wal_senders;
8282 940 : ControlFile->max_prepared_xacts = max_prepared_xacts;
8283 940 : ControlFile->max_locks_per_xact = max_locks_per_xact;
8284 940 : ControlFile->wal_level = wal_level;
8285 940 : ControlFile->wal_log_hints = wal_log_hints;
8286 940 : ControlFile->track_commit_timestamp = track_commit_timestamp;
8287 940 : UpdateControlFile();
8288 :
8289 940 : LWLockRelease(ControlFileLock);
8290 : }
8291 1862 : }
8292 :
8293 : /*
8294 : * Update full_page_writes in shared memory, and write an
8295 : * XLOG_FPW_CHANGE record if necessary.
8296 : *
8297 : * Note: this function assumes there is no other process running
8298 : * concurrently that could update it.
8299 : */
8300 : void
8301 3138 : UpdateFullPageWrites(void)
8302 : {
8303 3138 : XLogCtlInsert *Insert = &XLogCtl->Insert;
8304 : bool recoveryInProgress;
8305 :
8306 : /*
8307 : * Do nothing if full_page_writes has not been changed.
8308 : *
8309 : * It's safe to check the shared full_page_writes without the lock,
8310 : * because we assume that there is no concurrently running process which
8311 : * can update it.
8312 : */
8313 3138 : if (fullPageWrites == Insert->fullPageWrites)
8314 2368 : return;
8315 :
8316 : /*
8317 : * Perform this outside critical section so that the WAL insert
8318 : * initialization done by RecoveryInProgress() doesn't trigger an
8319 : * assertion failure.
8320 : */
8321 770 : recoveryInProgress = RecoveryInProgress();
8322 :
8323 770 : START_CRIT_SECTION();
8324 :
8325 : /*
8326 : * It's always safe to take full page images, even when not strictly
8327 : * required, but not the other round. So if we're setting full_page_writes
8328 : * to true, first set it true and then write the WAL record. If we're
8329 : * setting it to false, first write the WAL record and then set the global
8330 : * flag.
8331 : */
8332 770 : if (fullPageWrites)
8333 : {
8334 748 : WALInsertLockAcquireExclusive();
8335 748 : Insert->fullPageWrites = true;
8336 748 : WALInsertLockRelease();
8337 : }
8338 :
8339 : /*
8340 : * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8341 : * full_page_writes during archive recovery, if required.
8342 : */
8343 770 : if (XLogStandbyInfoActive() && !recoveryInProgress)
8344 : {
8345 0 : XLogBeginInsert();
8346 0 : XLogRegisterData(&fullPageWrites, sizeof(bool));
8347 :
8348 0 : XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
8349 : }
8350 :
8351 770 : if (!fullPageWrites)
8352 : {
8353 22 : WALInsertLockAcquireExclusive();
8354 22 : Insert->fullPageWrites = false;
8355 22 : WALInsertLockRelease();
8356 : }
8357 770 : END_CRIT_SECTION();
8358 : }
8359 :
8360 : /*
8361 : * XLOG resource manager's routines
8362 : *
8363 : * Definitions of info values are in include/catalog/pg_control.h, though
8364 : * not all record types are related to control file updates.
8365 : *
8366 : * NOTE: Some XLOG record types that are directly related to WAL recovery
8367 : * are handled in xlogrecovery_redo().
8368 : */
8369 : void
8370 89570 : xlog_redo(XLogReaderState *record)
8371 : {
8372 89570 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8373 89570 : XLogRecPtr lsn = record->EndRecPtr;
8374 :
8375 : /*
8376 : * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
8377 : * XLOG_FPI_FOR_HINT records.
8378 : */
8379 : Assert(info == XLOG_FPI || info == XLOG_FPI_FOR_HINT ||
8380 : !XLogRecHasAnyBlockRefs(record));
8381 :
8382 89570 : if (info == XLOG_NEXTOID)
8383 : {
8384 : Oid nextOid;
8385 :
8386 : /*
8387 : * We used to try to take the maximum of TransamVariables->nextOid and
8388 : * the recorded nextOid, but that fails if the OID counter wraps
8389 : * around. Since no OID allocation should be happening during replay
8390 : * anyway, better to just believe the record exactly. We still take
8391 : * OidGenLock while setting the variable, just in case.
8392 : */
8393 192 : memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
8394 192 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8395 192 : TransamVariables->nextOid = nextOid;
8396 192 : TransamVariables->oidCount = 0;
8397 192 : LWLockRelease(OidGenLock);
8398 : }
8399 89378 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
8400 : {
8401 : CheckPoint checkPoint;
8402 : TimeLineID replayTLI;
8403 :
8404 78 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8405 : /* In a SHUTDOWN checkpoint, believe the counters exactly */
8406 78 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8407 78 : TransamVariables->nextXid = checkPoint.nextXid;
8408 78 : LWLockRelease(XidGenLock);
8409 78 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8410 78 : TransamVariables->nextOid = checkPoint.nextOid;
8411 78 : TransamVariables->oidCount = 0;
8412 78 : LWLockRelease(OidGenLock);
8413 78 : MultiXactSetNextMXact(checkPoint.nextMulti,
8414 : checkPoint.nextMultiOffset);
8415 :
8416 78 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8417 : checkPoint.oldestMultiDB);
8418 :
8419 : /*
8420 : * No need to set oldestClogXid here as well; it'll be set when we
8421 : * redo an xl_clog_truncate if it changed since initialization.
8422 : */
8423 78 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
8424 :
8425 : /*
8426 : * If we see a shutdown checkpoint while waiting for an end-of-backup
8427 : * record, the backup was canceled and the end-of-backup record will
8428 : * never arrive.
8429 : */
8430 78 : if (ArchiveRecoveryRequested &&
8431 76 : XLogRecPtrIsValid(ControlFile->backupStartPoint) &&
8432 0 : !XLogRecPtrIsValid(ControlFile->backupEndPoint))
8433 0 : ereport(PANIC,
8434 : (errmsg("online backup was canceled, recovery cannot continue")));
8435 :
8436 : /*
8437 : * If we see a shutdown checkpoint, we know that nothing was running
8438 : * on the primary at this point. So fake-up an empty running-xacts
8439 : * record and use that here and now. Recover additional standby state
8440 : * for prepared transactions.
8441 : */
8442 78 : if (standbyState >= STANDBY_INITIALIZED)
8443 : {
8444 : TransactionId *xids;
8445 : int nxids;
8446 : TransactionId oldestActiveXID;
8447 : TransactionId latestCompletedXid;
8448 : RunningTransactionsData running;
8449 :
8450 72 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
8451 :
8452 : /* Update pg_subtrans entries for any prepared transactions */
8453 72 : StandbyRecoverPreparedTransactions();
8454 :
8455 : /*
8456 : * Construct a RunningTransactions snapshot representing a shut
8457 : * down server, with only prepared transactions still alive. We're
8458 : * never overflowed at this point because all subxids are listed
8459 : * with their parent prepared transactions.
8460 : */
8461 72 : running.xcnt = nxids;
8462 72 : running.subxcnt = 0;
8463 72 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
8464 72 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
8465 72 : running.oldestRunningXid = oldestActiveXID;
8466 72 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
8467 72 : TransactionIdRetreat(latestCompletedXid);
8468 : Assert(TransactionIdIsNormal(latestCompletedXid));
8469 72 : running.latestCompletedXid = latestCompletedXid;
8470 72 : running.xids = xids;
8471 :
8472 72 : ProcArrayApplyRecoveryInfo(&running);
8473 : }
8474 :
8475 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8476 78 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8477 78 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8478 78 : LWLockRelease(ControlFileLock);
8479 :
8480 : /*
8481 : * We should've already switched to the new TLI before replaying this
8482 : * record.
8483 : */
8484 78 : (void) GetCurrentReplayRecPtr(&replayTLI);
8485 78 : if (checkPoint.ThisTimeLineID != replayTLI)
8486 0 : ereport(PANIC,
8487 : (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
8488 : checkPoint.ThisTimeLineID, replayTLI)));
8489 :
8490 78 : RecoveryRestartPoint(&checkPoint, record);
8491 :
8492 : /*
8493 : * After replaying a checkpoint record, free all smgr objects.
8494 : * Otherwise we would never do so for dropped relations, as the
8495 : * startup does not process shared invalidation messages or call
8496 : * AtEOXact_SMgr().
8497 : */
8498 78 : smgrdestroyall();
8499 : }
8500 89300 : else if (info == XLOG_CHECKPOINT_ONLINE)
8501 : {
8502 : CheckPoint checkPoint;
8503 : TimeLineID replayTLI;
8504 :
8505 1390 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8506 : /* In an ONLINE checkpoint, treat the XID counter as a minimum */
8507 1390 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8508 1390 : if (FullTransactionIdPrecedes(TransamVariables->nextXid,
8509 : checkPoint.nextXid))
8510 0 : TransamVariables->nextXid = checkPoint.nextXid;
8511 1390 : LWLockRelease(XidGenLock);
8512 :
8513 : /*
8514 : * We ignore the nextOid counter in an ONLINE checkpoint, preferring
8515 : * to track OID assignment through XLOG_NEXTOID records. The nextOid
8516 : * counter is from the start of the checkpoint and might well be stale
8517 : * compared to later XLOG_NEXTOID records. We could try to take the
8518 : * maximum of the nextOid counter and our latest value, but since
8519 : * there's no particular guarantee about the speed with which the OID
8520 : * counter wraps around, that's a risky thing to do. In any case,
8521 : * users of the nextOid counter are required to avoid assignment of
8522 : * duplicates, so that a somewhat out-of-date value should be safe.
8523 : */
8524 :
8525 : /* Handle multixact */
8526 1390 : MultiXactAdvanceNextMXact(checkPoint.nextMulti,
8527 : checkPoint.nextMultiOffset);
8528 :
8529 : /*
8530 : * NB: This may perform multixact truncation when replaying WAL
8531 : * generated by an older primary.
8532 : */
8533 1390 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8534 : checkPoint.oldestMultiDB);
8535 1390 : if (TransactionIdPrecedes(TransamVariables->oldestXid,
8536 : checkPoint.oldestXid))
8537 0 : SetTransactionIdLimit(checkPoint.oldestXid,
8538 : checkPoint.oldestXidDB);
8539 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8540 1390 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8541 1390 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8542 1390 : LWLockRelease(ControlFileLock);
8543 :
8544 : /* TLI should not change in an on-line checkpoint */
8545 1390 : (void) GetCurrentReplayRecPtr(&replayTLI);
8546 1390 : if (checkPoint.ThisTimeLineID != replayTLI)
8547 0 : ereport(PANIC,
8548 : (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
8549 : checkPoint.ThisTimeLineID, replayTLI)));
8550 :
8551 1390 : RecoveryRestartPoint(&checkPoint, record);
8552 :
8553 : /*
8554 : * After replaying a checkpoint record, free all smgr objects.
8555 : * Otherwise we would never do so for dropped relations, as the
8556 : * startup does not process shared invalidation messages or call
8557 : * AtEOXact_SMgr().
8558 : */
8559 1390 : smgrdestroyall();
8560 : }
8561 87910 : else if (info == XLOG_OVERWRITE_CONTRECORD)
8562 : {
8563 : /* nothing to do here, handled in xlogrecovery_redo() */
8564 : }
8565 87908 : else if (info == XLOG_END_OF_RECOVERY)
8566 : {
8567 : xl_end_of_recovery xlrec;
8568 : TimeLineID replayTLI;
8569 :
8570 22 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
8571 :
8572 : /*
8573 : * For Hot Standby, we could treat this like a Shutdown Checkpoint,
8574 : * but this case is rarer and harder to test, so the benefit doesn't
8575 : * outweigh the potential extra cost of maintenance.
8576 : */
8577 :
8578 : /*
8579 : * We should've already switched to the new TLI before replaying this
8580 : * record.
8581 : */
8582 22 : (void) GetCurrentReplayRecPtr(&replayTLI);
8583 22 : if (xlrec.ThisTimeLineID != replayTLI)
8584 0 : ereport(PANIC,
8585 : (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
8586 : xlrec.ThisTimeLineID, replayTLI)));
8587 : }
8588 87886 : else if (info == XLOG_NOOP)
8589 : {
8590 : /* nothing to do here */
8591 : }
8592 87886 : else if (info == XLOG_SWITCH)
8593 : {
8594 : /* nothing to do here */
8595 : }
8596 86972 : else if (info == XLOG_RESTORE_POINT)
8597 : {
8598 : /* nothing to do here, handled in xlogrecovery.c */
8599 : }
8600 86962 : else if (info == XLOG_FPI || info == XLOG_FPI_FOR_HINT)
8601 : {
8602 : /*
8603 : * XLOG_FPI records contain nothing else but one or more block
8604 : * references. Every block reference must include a full-page image
8605 : * even if full_page_writes was disabled when the record was generated
8606 : * - otherwise there would be no point in this record.
8607 : *
8608 : * XLOG_FPI_FOR_HINT records are generated when a page needs to be
8609 : * WAL-logged because of a hint bit update. They are only generated
8610 : * when checksums and/or wal_log_hints are enabled. They may include
8611 : * no full-page images if full_page_writes was disabled when they were
8612 : * generated. In this case there is nothing to do here.
8613 : *
8614 : * No recovery conflicts are generated by these generic records - if a
8615 : * resource manager needs to generate conflicts, it has to define a
8616 : * separate WAL record type and redo routine.
8617 : */
8618 180260 : for (uint8 block_id = 0; block_id <= XLogRecMaxBlockId(record); block_id++)
8619 : {
8620 : Buffer buffer;
8621 :
8622 94994 : if (!XLogRecHasBlockImage(record, block_id))
8623 : {
8624 132 : if (info == XLOG_FPI)
8625 0 : elog(ERROR, "XLOG_FPI record did not contain a full-page image");
8626 132 : continue;
8627 : }
8628 :
8629 94862 : if (XLogReadBufferForRedo(record, block_id, &buffer) != BLK_RESTORED)
8630 0 : elog(ERROR, "unexpected XLogReadBufferForRedo result when restoring backup block");
8631 94862 : UnlockReleaseBuffer(buffer);
8632 : }
8633 : }
8634 1696 : else if (info == XLOG_BACKUP_END)
8635 : {
8636 : /* nothing to do here, handled in xlogrecovery_redo() */
8637 : }
8638 1504 : else if (info == XLOG_PARAMETER_CHANGE)
8639 : {
8640 : xl_parameter_change xlrec;
8641 :
8642 : /* Update our copy of the parameters in pg_control */
8643 76 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
8644 :
8645 76 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8646 76 : ControlFile->MaxConnections = xlrec.MaxConnections;
8647 76 : ControlFile->max_worker_processes = xlrec.max_worker_processes;
8648 76 : ControlFile->max_wal_senders = xlrec.max_wal_senders;
8649 76 : ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
8650 76 : ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
8651 76 : ControlFile->wal_level = xlrec.wal_level;
8652 76 : ControlFile->wal_log_hints = xlrec.wal_log_hints;
8653 :
8654 : /*
8655 : * Update minRecoveryPoint to ensure that if recovery is aborted, we
8656 : * recover back up to this point before allowing hot standby again.
8657 : * This is important if the max_* settings are decreased, to ensure
8658 : * you don't run queries against the WAL preceding the change. The
8659 : * local copies cannot be updated as long as crash recovery is
8660 : * happening and we expect all the WAL to be replayed.
8661 : */
8662 76 : if (InArchiveRecovery)
8663 : {
8664 46 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
8665 46 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
8666 : }
8667 76 : if (XLogRecPtrIsValid(LocalMinRecoveryPoint) && LocalMinRecoveryPoint < lsn)
8668 : {
8669 : TimeLineID replayTLI;
8670 :
8671 24 : (void) GetCurrentReplayRecPtr(&replayTLI);
8672 24 : ControlFile->minRecoveryPoint = lsn;
8673 24 : ControlFile->minRecoveryPointTLI = replayTLI;
8674 : }
8675 :
8676 76 : CommitTsParameterChange(xlrec.track_commit_timestamp,
8677 76 : ControlFile->track_commit_timestamp);
8678 76 : ControlFile->track_commit_timestamp = xlrec.track_commit_timestamp;
8679 :
8680 76 : UpdateControlFile();
8681 76 : LWLockRelease(ControlFileLock);
8682 :
8683 : /* Check to see if any parameter change gives a problem on recovery */
8684 76 : CheckRequiredParameterValues();
8685 : }
8686 1428 : else if (info == XLOG_FPW_CHANGE)
8687 : {
8688 : bool fpw;
8689 :
8690 0 : memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
8691 :
8692 : /*
8693 : * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
8694 : * do_pg_backup_start() and do_pg_backup_stop() can check whether
8695 : * full_page_writes has been disabled during online backup.
8696 : */
8697 0 : if (!fpw)
8698 : {
8699 0 : SpinLockAcquire(&XLogCtl->info_lck);
8700 0 : if (XLogCtl->lastFpwDisableRecPtr < record->ReadRecPtr)
8701 0 : XLogCtl->lastFpwDisableRecPtr = record->ReadRecPtr;
8702 0 : SpinLockRelease(&XLogCtl->info_lck);
8703 : }
8704 :
8705 : /* Keep track of full_page_writes */
8706 0 : lastFullPageWrites = fpw;
8707 : }
8708 1428 : else if (info == XLOG_CHECKPOINT_REDO)
8709 : {
8710 : /* nothing to do here, just for informational purposes */
8711 : }
8712 36 : else if (info == XLOG_LOGICAL_DECODING_STATUS_CHANGE)
8713 : {
8714 : bool status;
8715 :
8716 36 : memcpy(&status, XLogRecGetData(record), sizeof(bool));
8717 :
8718 : /*
8719 : * We need to toggle the logical decoding status and update the
8720 : * XLogLogicalInfo cache of processes synchronously because
8721 : * XLogLogicalInfoActive() is used even during read-only queries
8722 : * (e.g., via RelationIsAccessibleInLogicalDecoding()). In the
8723 : * 'disable' case, it is safe to invalidate existing slots after
8724 : * disabling logical decoding because logical decoding cannot process
8725 : * subsequent WAL records, which may not contain logical information.
8726 : */
8727 36 : if (status)
8728 18 : EnableLogicalDecoding();
8729 : else
8730 18 : DisableLogicalDecoding();
8731 :
8732 36 : elog(DEBUG1, "update logical decoding status to %d during recovery",
8733 : status);
8734 :
8735 36 : if (InRecovery && InHotStandby)
8736 : {
8737 32 : if (!status)
8738 : {
8739 : /*
8740 : * Invalidate logical slots if we are in hot standby and the
8741 : * primary disabled logical decoding.
8742 : */
8743 18 : InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_LEVEL,
8744 : 0, InvalidOid,
8745 : InvalidTransactionId);
8746 : }
8747 14 : else if (sync_replication_slots)
8748 : {
8749 : /*
8750 : * Signal the postmaster to launch the slotsync worker.
8751 : *
8752 : * XXX: For simplicity, we keep the slotsync worker running
8753 : * even after logical decoding is disabled. A future
8754 : * improvement can consider starting and stopping the worker
8755 : * based on logical decoding status change.
8756 : */
8757 0 : kill(PostmasterPid, SIGUSR1);
8758 : }
8759 : }
8760 : }
8761 89566 : }
8762 :
8763 : /*
8764 : * Return the extra open flags used for opening a file, depending on the
8765 : * value of the GUCs wal_sync_method, fsync and debug_io_direct.
8766 : */
8767 : static int
8768 33126 : get_sync_bit(int method)
8769 : {
8770 33126 : int o_direct_flag = 0;
8771 :
8772 : /*
8773 : * Use O_DIRECT if requested, except in walreceiver process. The WAL
8774 : * written by walreceiver is normally read by the startup process soon
8775 : * after it's written. Also, walreceiver performs unaligned writes, which
8776 : * don't work with O_DIRECT, so it is required for correctness too.
8777 : */
8778 33126 : if ((io_direct_flags & IO_DIRECT_WAL) && !AmWalReceiverProcess())
8779 20 : o_direct_flag = PG_O_DIRECT;
8780 :
8781 : /* If fsync is disabled, never open in sync mode */
8782 33126 : if (!enableFsync)
8783 33126 : return o_direct_flag;
8784 :
8785 0 : switch (method)
8786 : {
8787 : /*
8788 : * enum values for all sync options are defined even if they are
8789 : * not supported on the current platform. But if not, they are
8790 : * not included in the enum option array, and therefore will never
8791 : * be seen here.
8792 : */
8793 0 : case WAL_SYNC_METHOD_FSYNC:
8794 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8795 : case WAL_SYNC_METHOD_FDATASYNC:
8796 0 : return o_direct_flag;
8797 : #ifdef O_SYNC
8798 0 : case WAL_SYNC_METHOD_OPEN:
8799 0 : return O_SYNC | o_direct_flag;
8800 : #endif
8801 : #ifdef O_DSYNC
8802 0 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8803 0 : return O_DSYNC | o_direct_flag;
8804 : #endif
8805 0 : default:
8806 : /* can't happen (unless we are out of sync with option array) */
8807 0 : elog(ERROR, "unrecognized \"wal_sync_method\": %d", method);
8808 : return 0; /* silence warning */
8809 : }
8810 : }
8811 :
8812 : /*
8813 : * GUC support
8814 : */
8815 : void
8816 2356 : assign_wal_sync_method(int new_wal_sync_method, void *extra)
8817 : {
8818 2356 : if (wal_sync_method != new_wal_sync_method)
8819 : {
8820 : /*
8821 : * To ensure that no blocks escape unsynced, force an fsync on the
8822 : * currently open log segment (if any). Also, if the open flag is
8823 : * changing, close the log file so it will be reopened (with new flag
8824 : * bit) at next use.
8825 : */
8826 0 : if (openLogFile >= 0)
8827 : {
8828 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN);
8829 0 : if (pg_fsync(openLogFile) != 0)
8830 : {
8831 : char xlogfname[MAXFNAMELEN];
8832 : int save_errno;
8833 :
8834 0 : save_errno = errno;
8835 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo,
8836 : wal_segment_size);
8837 0 : errno = save_errno;
8838 0 : ereport(PANIC,
8839 : (errcode_for_file_access(),
8840 : errmsg("could not fsync file \"%s\": %m", xlogfname)));
8841 : }
8842 :
8843 0 : pgstat_report_wait_end();
8844 0 : if (get_sync_bit(wal_sync_method) != get_sync_bit(new_wal_sync_method))
8845 0 : XLogFileClose();
8846 : }
8847 : }
8848 2356 : }
8849 :
8850 :
8851 : /*
8852 : * Issue appropriate kind of fsync (if any) for an XLOG output file.
8853 : *
8854 : * 'fd' is a file descriptor for the XLOG file to be fsync'd.
8855 : * 'segno' is for error reporting purposes.
8856 : */
8857 : void
8858 340620 : issue_xlog_fsync(int fd, XLogSegNo segno, TimeLineID tli)
8859 : {
8860 340620 : char *msg = NULL;
8861 : instr_time start;
8862 :
8863 : Assert(tli != 0);
8864 :
8865 : /*
8866 : * Quick exit if fsync is disabled or write() has already synced the WAL
8867 : * file.
8868 : */
8869 340620 : if (!enableFsync ||
8870 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN ||
8871 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN_DSYNC)
8872 340620 : return;
8873 :
8874 : /*
8875 : * Measure I/O timing to sync the WAL file for pg_stat_io.
8876 : */
8877 0 : start = pgstat_prepare_io_time(track_wal_io_timing);
8878 :
8879 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC);
8880 0 : switch (wal_sync_method)
8881 : {
8882 0 : case WAL_SYNC_METHOD_FSYNC:
8883 0 : if (pg_fsync_no_writethrough(fd) != 0)
8884 0 : msg = _("could not fsync file \"%s\": %m");
8885 0 : break;
8886 : #ifdef HAVE_FSYNC_WRITETHROUGH
8887 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8888 : if (pg_fsync_writethrough(fd) != 0)
8889 : msg = _("could not fsync write-through file \"%s\": %m");
8890 : break;
8891 : #endif
8892 0 : case WAL_SYNC_METHOD_FDATASYNC:
8893 0 : if (pg_fdatasync(fd) != 0)
8894 0 : msg = _("could not fdatasync file \"%s\": %m");
8895 0 : break;
8896 0 : case WAL_SYNC_METHOD_OPEN:
8897 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8898 : /* not reachable */
8899 : Assert(false);
8900 0 : break;
8901 0 : default:
8902 0 : ereport(PANIC,
8903 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8904 : errmsg_internal("unrecognized \"wal_sync_method\": %d", wal_sync_method));
8905 : break;
8906 : }
8907 :
8908 : /* PANIC if failed to fsync */
8909 0 : if (msg)
8910 : {
8911 : char xlogfname[MAXFNAMELEN];
8912 0 : int save_errno = errno;
8913 :
8914 0 : XLogFileName(xlogfname, tli, segno, wal_segment_size);
8915 0 : errno = save_errno;
8916 0 : ereport(PANIC,
8917 : (errcode_for_file_access(),
8918 : errmsg(msg, xlogfname)));
8919 : }
8920 :
8921 0 : pgstat_report_wait_end();
8922 :
8923 0 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL, IOOP_FSYNC,
8924 : start, 1, 0);
8925 : }
8926 :
8927 : /*
8928 : * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
8929 : * function. It creates the necessary starting checkpoint and constructs the
8930 : * backup state and tablespace map.
8931 : *
8932 : * Input parameters are "state" (the backup state), "fast" (if true, we do
8933 : * the checkpoint in fast mode), and "tablespaces" (if non-NULL, indicates a
8934 : * list of tablespaceinfo structs describing the cluster's tablespaces.).
8935 : *
8936 : * The tablespace map contents are appended to passed-in parameter
8937 : * tablespace_map and the caller is responsible for including it in the backup
8938 : * archive as 'tablespace_map'. The tablespace_map file is required mainly for
8939 : * tar format in windows as native windows utilities are not able to create
8940 : * symlinks while extracting files from tar. However for consistency and
8941 : * platform-independence, we do it the same way everywhere.
8942 : *
8943 : * It fills in "state" with the information required for the backup, such
8944 : * as the minimum WAL location that must be present to restore from this
8945 : * backup (starttli) and the corresponding timeline ID (starttli).
8946 : *
8947 : * Every successfully started backup must be stopped by calling
8948 : * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
8949 : * backups active at the same time.
8950 : *
8951 : * It is the responsibility of the caller of this function to verify the
8952 : * permissions of the calling user!
8953 : */
8954 : void
8955 344 : do_pg_backup_start(const char *backupidstr, bool fast, List **tablespaces,
8956 : BackupState *state, StringInfo tblspcmapfile)
8957 : {
8958 : bool backup_started_in_recovery;
8959 :
8960 : Assert(state != NULL);
8961 344 : backup_started_in_recovery = RecoveryInProgress();
8962 :
8963 : /*
8964 : * During recovery, we don't need to check WAL level. Because, if WAL
8965 : * level is not sufficient, it's impossible to get here during recovery.
8966 : */
8967 344 : if (!backup_started_in_recovery && !XLogIsNeeded())
8968 0 : ereport(ERROR,
8969 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8970 : errmsg("WAL level not sufficient for making an online backup"),
8971 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
8972 :
8973 344 : if (strlen(backupidstr) > MAXPGPATH)
8974 2 : ereport(ERROR,
8975 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8976 : errmsg("backup label too long (max %d bytes)",
8977 : MAXPGPATH)));
8978 :
8979 342 : strlcpy(state->name, backupidstr, sizeof(state->name));
8980 :
8981 : /*
8982 : * Mark backup active in shared memory. We must do full-page WAL writes
8983 : * during an on-line backup even if not doing so at other times, because
8984 : * it's quite possible for the backup dump to obtain a "torn" (partially
8985 : * written) copy of a database page if it reads the page concurrently with
8986 : * our write to the same page. This can be fixed as long as the first
8987 : * write to the page in the WAL sequence is a full-page write. Hence, we
8988 : * increment runningBackups then force a CHECKPOINT, to ensure there are
8989 : * no dirty pages in shared memory that might get dumped while the backup
8990 : * is in progress without having a corresponding WAL record. (Once the
8991 : * backup is complete, we need not force full-page writes anymore, since
8992 : * we expect that any pages not modified during the backup interval must
8993 : * have been correctly captured by the backup.)
8994 : *
8995 : * Note that forcing full-page writes has no effect during an online
8996 : * backup from the standby.
8997 : *
8998 : * We must hold all the insertion locks to change the value of
8999 : * runningBackups, to ensure adequate interlocking against
9000 : * XLogInsertRecord().
9001 : */
9002 342 : WALInsertLockAcquireExclusive();
9003 342 : XLogCtl->Insert.runningBackups++;
9004 342 : WALInsertLockRelease();
9005 :
9006 : /*
9007 : * Ensure we decrement runningBackups if we fail below. NB -- for this to
9008 : * work correctly, it is critical that sessionBackupState is only updated
9009 : * after this block is over.
9010 : */
9011 342 : PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9012 : {
9013 342 : bool gotUniqueStartpoint = false;
9014 : DIR *tblspcdir;
9015 : struct dirent *de;
9016 : tablespaceinfo *ti;
9017 : int datadirpathlen;
9018 :
9019 : /*
9020 : * Force an XLOG file switch before the checkpoint, to ensure that the
9021 : * WAL segment the checkpoint is written to doesn't contain pages with
9022 : * old timeline IDs. That would otherwise happen if you called
9023 : * pg_backup_start() right after restoring from a PITR archive: the
9024 : * first WAL segment containing the startup checkpoint has pages in
9025 : * the beginning with the old timeline ID. That can cause trouble at
9026 : * recovery: we won't have a history file covering the old timeline if
9027 : * pg_wal directory was not included in the base backup and the WAL
9028 : * archive was cleared too before starting the backup.
9029 : *
9030 : * This also ensures that we have emitted a WAL page header that has
9031 : * XLP_BKP_REMOVABLE off before we emit the checkpoint record.
9032 : * Therefore, if a WAL archiver (such as pglesslog) is trying to
9033 : * compress out removable backup blocks, it won't remove any that
9034 : * occur after this point.
9035 : *
9036 : * During recovery, we skip forcing XLOG file switch, which means that
9037 : * the backup taken during recovery is not available for the special
9038 : * recovery case described above.
9039 : */
9040 342 : if (!backup_started_in_recovery)
9041 326 : RequestXLogSwitch(false);
9042 :
9043 : do
9044 : {
9045 : bool checkpointfpw;
9046 :
9047 : /*
9048 : * Force a CHECKPOINT. Aside from being necessary to prevent torn
9049 : * page problems, this guarantees that two successive backup runs
9050 : * will have different checkpoint positions and hence different
9051 : * history file names, even if nothing happened in between.
9052 : *
9053 : * During recovery, establish a restartpoint if possible. We use
9054 : * the last restartpoint as the backup starting checkpoint. This
9055 : * means that two successive backup runs can have same checkpoint
9056 : * positions.
9057 : *
9058 : * Since the fact that we are executing do_pg_backup_start()
9059 : * during recovery means that checkpointer is running, we can use
9060 : * RequestCheckpoint() to establish a restartpoint.
9061 : *
9062 : * We use CHECKPOINT_FAST only if requested by user (via passing
9063 : * fast = true). Otherwise this can take awhile.
9064 : */
9065 342 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
9066 : (fast ? CHECKPOINT_FAST : 0));
9067 :
9068 : /*
9069 : * Now we need to fetch the checkpoint record location, and also
9070 : * its REDO pointer. The oldest point in WAL that would be needed
9071 : * to restore starting from the checkpoint is precisely the REDO
9072 : * pointer.
9073 : */
9074 342 : LWLockAcquire(ControlFileLock, LW_SHARED);
9075 342 : state->checkpointloc = ControlFile->checkPoint;
9076 342 : state->startpoint = ControlFile->checkPointCopy.redo;
9077 342 : state->starttli = ControlFile->checkPointCopy.ThisTimeLineID;
9078 342 : checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
9079 342 : LWLockRelease(ControlFileLock);
9080 :
9081 342 : if (backup_started_in_recovery)
9082 : {
9083 : XLogRecPtr recptr;
9084 :
9085 : /*
9086 : * Check to see if all WAL replayed during online backup
9087 : * (i.e., since last restartpoint used as backup starting
9088 : * checkpoint) contain full-page writes.
9089 : */
9090 16 : SpinLockAcquire(&XLogCtl->info_lck);
9091 16 : recptr = XLogCtl->lastFpwDisableRecPtr;
9092 16 : SpinLockRelease(&XLogCtl->info_lck);
9093 :
9094 16 : if (!checkpointfpw || state->startpoint <= recptr)
9095 0 : ereport(ERROR,
9096 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9097 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9098 : "since last restartpoint"),
9099 : errhint("This means that the backup being taken on the standby "
9100 : "is corrupt and should not be used. "
9101 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9102 : "and then try an online backup again.")));
9103 :
9104 : /*
9105 : * During recovery, since we don't use the end-of-backup WAL
9106 : * record and don't write the backup history file, the
9107 : * starting WAL location doesn't need to be unique. This means
9108 : * that two base backups started at the same time might use
9109 : * the same checkpoint as starting locations.
9110 : */
9111 16 : gotUniqueStartpoint = true;
9112 : }
9113 :
9114 : /*
9115 : * If two base backups are started at the same time (in WAL sender
9116 : * processes), we need to make sure that they use different
9117 : * checkpoints as starting locations, because we use the starting
9118 : * WAL location as a unique identifier for the base backup in the
9119 : * end-of-backup WAL record and when we write the backup history
9120 : * file. Perhaps it would be better generate a separate unique ID
9121 : * for each backup instead of forcing another checkpoint, but
9122 : * taking a checkpoint right after another is not that expensive
9123 : * either because only few buffers have been dirtied yet.
9124 : */
9125 342 : WALInsertLockAcquireExclusive();
9126 342 : if (XLogCtl->Insert.lastBackupStart < state->startpoint)
9127 : {
9128 342 : XLogCtl->Insert.lastBackupStart = state->startpoint;
9129 342 : gotUniqueStartpoint = true;
9130 : }
9131 342 : WALInsertLockRelease();
9132 342 : } while (!gotUniqueStartpoint);
9133 :
9134 : /*
9135 : * Construct tablespace_map file.
9136 : */
9137 342 : datadirpathlen = strlen(DataDir);
9138 :
9139 : /* Collect information about all tablespaces */
9140 342 : tblspcdir = AllocateDir(PG_TBLSPC_DIR);
9141 1098 : while ((de = ReadDir(tblspcdir, PG_TBLSPC_DIR)) != NULL)
9142 : {
9143 : char fullpath[MAXPGPATH + sizeof(PG_TBLSPC_DIR)];
9144 : char linkpath[MAXPGPATH];
9145 756 : char *relpath = NULL;
9146 : char *s;
9147 : PGFileType de_type;
9148 : char *badp;
9149 : Oid tsoid;
9150 :
9151 : /*
9152 : * Try to parse the directory name as an unsigned integer.
9153 : *
9154 : * Tablespace directories should be positive integers that can be
9155 : * represented in 32 bits, with no leading zeroes or trailing
9156 : * garbage. If we come across a name that doesn't meet those
9157 : * criteria, skip it.
9158 : */
9159 756 : if (de->d_name[0] < '1' || de->d_name[1] > '9')
9160 684 : continue;
9161 72 : errno = 0;
9162 72 : tsoid = strtoul(de->d_name, &badp, 10);
9163 72 : if (*badp != '\0' || errno == EINVAL || errno == ERANGE)
9164 0 : continue;
9165 :
9166 72 : snprintf(fullpath, sizeof(fullpath), "%s/%s", PG_TBLSPC_DIR, de->d_name);
9167 :
9168 72 : de_type = get_dirent_type(fullpath, de, false, ERROR);
9169 :
9170 72 : if (de_type == PGFILETYPE_LNK)
9171 : {
9172 : StringInfoData escapedpath;
9173 : int rllen;
9174 :
9175 44 : rllen = readlink(fullpath, linkpath, sizeof(linkpath));
9176 44 : if (rllen < 0)
9177 : {
9178 0 : ereport(WARNING,
9179 : (errmsg("could not read symbolic link \"%s\": %m",
9180 : fullpath)));
9181 0 : continue;
9182 : }
9183 44 : else if (rllen >= sizeof(linkpath))
9184 : {
9185 0 : ereport(WARNING,
9186 : (errmsg("symbolic link \"%s\" target is too long",
9187 : fullpath)));
9188 0 : continue;
9189 : }
9190 44 : linkpath[rllen] = '\0';
9191 :
9192 : /*
9193 : * Relpath holds the relative path of the tablespace directory
9194 : * when it's located within PGDATA, or NULL if it's located
9195 : * elsewhere.
9196 : */
9197 44 : if (rllen > datadirpathlen &&
9198 2 : strncmp(linkpath, DataDir, datadirpathlen) == 0 &&
9199 0 : IS_DIR_SEP(linkpath[datadirpathlen]))
9200 0 : relpath = pstrdup(linkpath + datadirpathlen + 1);
9201 :
9202 : /*
9203 : * Add a backslash-escaped version of the link path to the
9204 : * tablespace map file.
9205 : */
9206 44 : initStringInfo(&escapedpath);
9207 1092 : for (s = linkpath; *s; s++)
9208 : {
9209 1048 : if (*s == '\n' || *s == '\r' || *s == '\\')
9210 0 : appendStringInfoChar(&escapedpath, '\\');
9211 1048 : appendStringInfoChar(&escapedpath, *s);
9212 : }
9213 44 : appendStringInfo(tblspcmapfile, "%s %s\n",
9214 44 : de->d_name, escapedpath.data);
9215 44 : pfree(escapedpath.data);
9216 : }
9217 28 : else if (de_type == PGFILETYPE_DIR)
9218 : {
9219 : /*
9220 : * It's possible to use allow_in_place_tablespaces to create
9221 : * directories directly under pg_tblspc, for testing purposes
9222 : * only.
9223 : *
9224 : * In this case, we store a relative path rather than an
9225 : * absolute path into the tablespaceinfo.
9226 : */
9227 28 : snprintf(linkpath, sizeof(linkpath), "%s/%s",
9228 28 : PG_TBLSPC_DIR, de->d_name);
9229 28 : relpath = pstrdup(linkpath);
9230 : }
9231 : else
9232 : {
9233 : /* Skip any other file type that appears here. */
9234 0 : continue;
9235 : }
9236 :
9237 72 : ti = palloc_object(tablespaceinfo);
9238 72 : ti->oid = tsoid;
9239 72 : ti->path = pstrdup(linkpath);
9240 72 : ti->rpath = relpath;
9241 72 : ti->size = -1;
9242 :
9243 72 : if (tablespaces)
9244 72 : *tablespaces = lappend(*tablespaces, ti);
9245 : }
9246 342 : FreeDir(tblspcdir);
9247 :
9248 342 : state->starttime = (pg_time_t) time(NULL);
9249 : }
9250 342 : PG_END_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9251 :
9252 342 : state->started_in_recovery = backup_started_in_recovery;
9253 :
9254 : /*
9255 : * Mark that the start phase has correctly finished for the backup.
9256 : */
9257 342 : sessionBackupState = SESSION_BACKUP_RUNNING;
9258 342 : }
9259 :
9260 : /*
9261 : * Utility routine to fetch the session-level status of a backup running.
9262 : */
9263 : SessionBackupState
9264 386 : get_backup_status(void)
9265 : {
9266 386 : return sessionBackupState;
9267 : }
9268 :
9269 : /*
9270 : * do_pg_backup_stop
9271 : *
9272 : * Utility function called at the end of an online backup. It creates history
9273 : * file (if required), resets sessionBackupState and so on. It can optionally
9274 : * wait for WAL segments to be archived.
9275 : *
9276 : * "state" is filled with the information necessary to restore from this
9277 : * backup with its stop LSN (stoppoint), its timeline ID (stoptli), etc.
9278 : *
9279 : * It is the responsibility of the caller of this function to verify the
9280 : * permissions of the calling user!
9281 : */
9282 : void
9283 330 : do_pg_backup_stop(BackupState *state, bool waitforarchive)
9284 : {
9285 330 : bool backup_stopped_in_recovery = false;
9286 : char histfilepath[MAXPGPATH];
9287 : char lastxlogfilename[MAXFNAMELEN];
9288 : char histfilename[MAXFNAMELEN];
9289 : XLogSegNo _logSegNo;
9290 : FILE *fp;
9291 : int seconds_before_warning;
9292 330 : int waits = 0;
9293 330 : bool reported_waiting = false;
9294 :
9295 : Assert(state != NULL);
9296 :
9297 330 : backup_stopped_in_recovery = RecoveryInProgress();
9298 :
9299 : /*
9300 : * During recovery, we don't need to check WAL level. Because, if WAL
9301 : * level is not sufficient, it's impossible to get here during recovery.
9302 : */
9303 330 : if (!backup_stopped_in_recovery && !XLogIsNeeded())
9304 0 : ereport(ERROR,
9305 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9306 : errmsg("WAL level not sufficient for making an online backup"),
9307 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9308 :
9309 : /*
9310 : * OK to update backup counter and session-level lock.
9311 : *
9312 : * Note that CHECK_FOR_INTERRUPTS() must not occur while updating them,
9313 : * otherwise they can be updated inconsistently, which might cause
9314 : * do_pg_abort_backup() to fail.
9315 : */
9316 330 : WALInsertLockAcquireExclusive();
9317 :
9318 : /*
9319 : * It is expected that each do_pg_backup_start() call is matched by
9320 : * exactly one do_pg_backup_stop() call.
9321 : */
9322 : Assert(XLogCtl->Insert.runningBackups > 0);
9323 330 : XLogCtl->Insert.runningBackups--;
9324 :
9325 : /*
9326 : * Clean up session-level lock.
9327 : *
9328 : * You might think that WALInsertLockRelease() can be called before
9329 : * cleaning up session-level lock because session-level lock doesn't need
9330 : * to be protected with WAL insertion lock. But since
9331 : * CHECK_FOR_INTERRUPTS() can occur in it, session-level lock must be
9332 : * cleaned up before it.
9333 : */
9334 330 : sessionBackupState = SESSION_BACKUP_NONE;
9335 :
9336 330 : WALInsertLockRelease();
9337 :
9338 : /*
9339 : * If we are taking an online backup from the standby, we confirm that the
9340 : * standby has not been promoted during the backup.
9341 : */
9342 330 : if (state->started_in_recovery && !backup_stopped_in_recovery)
9343 0 : ereport(ERROR,
9344 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9345 : errmsg("the standby was promoted during online backup"),
9346 : errhint("This means that the backup being taken is corrupt "
9347 : "and should not be used. "
9348 : "Try taking another online backup.")));
9349 :
9350 : /*
9351 : * During recovery, we don't write an end-of-backup record. We assume that
9352 : * pg_control was backed up last and its minimum recovery point can be
9353 : * available as the backup end location. Since we don't have an
9354 : * end-of-backup record, we use the pg_control value to check whether
9355 : * we've reached the end of backup when starting recovery from this
9356 : * backup. We have no way of checking if pg_control wasn't backed up last
9357 : * however.
9358 : *
9359 : * We don't force a switch to new WAL file but it is still possible to
9360 : * wait for all the required files to be archived if waitforarchive is
9361 : * true. This is okay if we use the backup to start a standby and fetch
9362 : * the missing WAL using streaming replication. But in the case of an
9363 : * archive recovery, a user should set waitforarchive to true and wait for
9364 : * them to be archived to ensure that all the required files are
9365 : * available.
9366 : *
9367 : * We return the current minimum recovery point as the backup end
9368 : * location. Note that it can be greater than the exact backup end
9369 : * location if the minimum recovery point is updated after the backup of
9370 : * pg_control. This is harmless for current uses.
9371 : *
9372 : * XXX currently a backup history file is for informational and debug
9373 : * purposes only. It's not essential for an online backup. Furthermore,
9374 : * even if it's created, it will not be archived during recovery because
9375 : * an archiver is not invoked. So it doesn't seem worthwhile to write a
9376 : * backup history file during recovery.
9377 : */
9378 330 : if (backup_stopped_in_recovery)
9379 : {
9380 : XLogRecPtr recptr;
9381 :
9382 : /*
9383 : * Check to see if all WAL replayed during online backup contain
9384 : * full-page writes.
9385 : */
9386 16 : SpinLockAcquire(&XLogCtl->info_lck);
9387 16 : recptr = XLogCtl->lastFpwDisableRecPtr;
9388 16 : SpinLockRelease(&XLogCtl->info_lck);
9389 :
9390 16 : if (state->startpoint <= recptr)
9391 0 : ereport(ERROR,
9392 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9393 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9394 : "during online backup"),
9395 : errhint("This means that the backup being taken on the standby "
9396 : "is corrupt and should not be used. "
9397 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9398 : "and then try an online backup again.")));
9399 :
9400 :
9401 16 : LWLockAcquire(ControlFileLock, LW_SHARED);
9402 16 : state->stoppoint = ControlFile->minRecoveryPoint;
9403 16 : state->stoptli = ControlFile->minRecoveryPointTLI;
9404 16 : LWLockRelease(ControlFileLock);
9405 : }
9406 : else
9407 : {
9408 : char *history_file;
9409 :
9410 : /*
9411 : * Write the backup-end xlog record
9412 : */
9413 314 : XLogBeginInsert();
9414 314 : XLogRegisterData(&state->startpoint,
9415 : sizeof(state->startpoint));
9416 314 : state->stoppoint = XLogInsert(RM_XLOG_ID, XLOG_BACKUP_END);
9417 :
9418 : /*
9419 : * Given that we're not in recovery, InsertTimeLineID is set and can't
9420 : * change, so we can read it without a lock.
9421 : */
9422 314 : state->stoptli = XLogCtl->InsertTimeLineID;
9423 :
9424 : /*
9425 : * Force a switch to a new xlog segment file, so that the backup is
9426 : * valid as soon as archiver moves out the current segment file.
9427 : */
9428 314 : RequestXLogSwitch(false);
9429 :
9430 314 : state->stoptime = (pg_time_t) time(NULL);
9431 :
9432 : /*
9433 : * Write the backup history file
9434 : */
9435 314 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9436 314 : BackupHistoryFilePath(histfilepath, state->stoptli, _logSegNo,
9437 : state->startpoint, wal_segment_size);
9438 314 : fp = AllocateFile(histfilepath, "w");
9439 314 : if (!fp)
9440 0 : ereport(ERROR,
9441 : (errcode_for_file_access(),
9442 : errmsg("could not create file \"%s\": %m",
9443 : histfilepath)));
9444 :
9445 : /* Build and save the contents of the backup history file */
9446 314 : history_file = build_backup_content(state, true);
9447 314 : fprintf(fp, "%s", history_file);
9448 314 : pfree(history_file);
9449 :
9450 314 : if (fflush(fp) || ferror(fp) || FreeFile(fp))
9451 0 : ereport(ERROR,
9452 : (errcode_for_file_access(),
9453 : errmsg("could not write file \"%s\": %m",
9454 : histfilepath)));
9455 :
9456 : /*
9457 : * Clean out any no-longer-needed history files. As a side effect,
9458 : * this will post a .ready file for the newly created history file,
9459 : * notifying the archiver that history file may be archived
9460 : * immediately.
9461 : */
9462 314 : CleanupBackupHistory();
9463 : }
9464 :
9465 : /*
9466 : * If archiving is enabled, wait for all the required WAL files to be
9467 : * archived before returning. If archiving isn't enabled, the required WAL
9468 : * needs to be transported via streaming replication (hopefully with
9469 : * wal_keep_size set high enough), or some more exotic mechanism like
9470 : * polling and copying files from pg_wal with script. We have no knowledge
9471 : * of those mechanisms, so it's up to the user to ensure that he gets all
9472 : * the required WAL.
9473 : *
9474 : * We wait until both the last WAL file filled during backup and the
9475 : * history file have been archived, and assume that the alphabetic sorting
9476 : * property of the WAL files ensures any earlier WAL files are safely
9477 : * archived as well.
9478 : *
9479 : * We wait forever, since archive_command is supposed to work and we
9480 : * assume the admin wanted his backup to work completely. If you don't
9481 : * wish to wait, then either waitforarchive should be passed in as false,
9482 : * or you can set statement_timeout. Also, some notices are issued to
9483 : * clue in anyone who might be doing this interactively.
9484 : */
9485 :
9486 330 : if (waitforarchive &&
9487 22 : ((!backup_stopped_in_recovery && XLogArchivingActive()) ||
9488 2 : (backup_stopped_in_recovery && XLogArchivingAlways())))
9489 : {
9490 10 : XLByteToPrevSeg(state->stoppoint, _logSegNo, wal_segment_size);
9491 10 : XLogFileName(lastxlogfilename, state->stoptli, _logSegNo,
9492 : wal_segment_size);
9493 :
9494 10 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9495 10 : BackupHistoryFileName(histfilename, state->stoptli, _logSegNo,
9496 : state->startpoint, wal_segment_size);
9497 :
9498 10 : seconds_before_warning = 60;
9499 10 : waits = 0;
9500 :
9501 30 : while (XLogArchiveIsBusy(lastxlogfilename) ||
9502 10 : XLogArchiveIsBusy(histfilename))
9503 : {
9504 10 : CHECK_FOR_INTERRUPTS();
9505 :
9506 10 : if (!reported_waiting && waits > 5)
9507 : {
9508 0 : ereport(NOTICE,
9509 : (errmsg("base backup done, waiting for required WAL segments to be archived")));
9510 0 : reported_waiting = true;
9511 : }
9512 :
9513 10 : (void) WaitLatch(MyLatch,
9514 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
9515 : 1000L,
9516 : WAIT_EVENT_BACKUP_WAIT_WAL_ARCHIVE);
9517 10 : ResetLatch(MyLatch);
9518 :
9519 10 : if (++waits >= seconds_before_warning)
9520 : {
9521 0 : seconds_before_warning *= 2; /* This wraps in >10 years... */
9522 0 : ereport(WARNING,
9523 : (errmsg("still waiting for all required WAL segments to be archived (%d seconds elapsed)",
9524 : waits),
9525 : errhint("Check that your \"archive_command\" is executing properly. "
9526 : "You can safely cancel this backup, "
9527 : "but the database backup will not be usable without all the WAL segments.")));
9528 : }
9529 : }
9530 :
9531 10 : ereport(NOTICE,
9532 : (errmsg("all required WAL segments have been archived")));
9533 : }
9534 320 : else if (waitforarchive)
9535 12 : ereport(NOTICE,
9536 : (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
9537 330 : }
9538 :
9539 :
9540 : /*
9541 : * do_pg_abort_backup: abort a running backup
9542 : *
9543 : * This does just the most basic steps of do_pg_backup_stop(), by taking the
9544 : * system out of backup mode, thus making it a lot more safe to call from
9545 : * an error handler.
9546 : *
9547 : * 'arg' indicates that it's being called during backup setup; so
9548 : * sessionBackupState has not been modified yet, but runningBackups has
9549 : * already been incremented. When it's false, then it's invoked as a
9550 : * before_shmem_exit handler, and therefore we must not change state
9551 : * unless sessionBackupState indicates that a backup is actually running.
9552 : *
9553 : * NB: This gets used as a PG_ENSURE_ERROR_CLEANUP callback and
9554 : * before_shmem_exit handler, hence the odd-looking signature.
9555 : */
9556 : void
9557 18 : do_pg_abort_backup(int code, Datum arg)
9558 : {
9559 18 : bool during_backup_start = DatumGetBool(arg);
9560 :
9561 : /* If called during backup start, there shouldn't be one already running */
9562 : Assert(!during_backup_start || sessionBackupState == SESSION_BACKUP_NONE);
9563 :
9564 18 : if (during_backup_start || sessionBackupState != SESSION_BACKUP_NONE)
9565 : {
9566 12 : WALInsertLockAcquireExclusive();
9567 : Assert(XLogCtl->Insert.runningBackups > 0);
9568 12 : XLogCtl->Insert.runningBackups--;
9569 :
9570 12 : sessionBackupState = SESSION_BACKUP_NONE;
9571 12 : WALInsertLockRelease();
9572 :
9573 12 : if (!during_backup_start)
9574 12 : ereport(WARNING,
9575 : errmsg("aborting backup due to backend exiting before pg_backup_stop was called"));
9576 : }
9577 18 : }
9578 :
9579 : /*
9580 : * Register a handler that will warn about unterminated backups at end of
9581 : * session, unless this has already been done.
9582 : */
9583 : void
9584 10 : register_persistent_abort_backup_handler(void)
9585 : {
9586 : static bool already_done = false;
9587 :
9588 10 : if (already_done)
9589 2 : return;
9590 8 : before_shmem_exit(do_pg_abort_backup, BoolGetDatum(false));
9591 8 : already_done = true;
9592 : }
9593 :
9594 : /*
9595 : * Get latest WAL insert pointer
9596 : */
9597 : XLogRecPtr
9598 4054 : GetXLogInsertRecPtr(void)
9599 : {
9600 4054 : XLogCtlInsert *Insert = &XLogCtl->Insert;
9601 : uint64 current_bytepos;
9602 :
9603 4054 : SpinLockAcquire(&Insert->insertpos_lck);
9604 4054 : current_bytepos = Insert->CurrBytePos;
9605 4054 : SpinLockRelease(&Insert->insertpos_lck);
9606 :
9607 4054 : return XLogBytePosToRecPtr(current_bytepos);
9608 : }
9609 :
9610 : /*
9611 : * Get latest WAL write pointer
9612 : */
9613 : XLogRecPtr
9614 21272 : GetXLogWriteRecPtr(void)
9615 : {
9616 21272 : RefreshXLogWriteResult(LogwrtResult);
9617 :
9618 21272 : return LogwrtResult.Write;
9619 : }
9620 :
9621 : /*
9622 : * Returns the redo pointer of the last checkpoint or restartpoint. This is
9623 : * the oldest point in WAL that we still need, if we have to restart recovery.
9624 : */
9625 : void
9626 788 : GetOldestRestartPoint(XLogRecPtr *oldrecptr, TimeLineID *oldtli)
9627 : {
9628 788 : LWLockAcquire(ControlFileLock, LW_SHARED);
9629 788 : *oldrecptr = ControlFile->checkPointCopy.redo;
9630 788 : *oldtli = ControlFile->checkPointCopy.ThisTimeLineID;
9631 788 : LWLockRelease(ControlFileLock);
9632 788 : }
9633 :
9634 : /* Thin wrapper around ShutdownWalRcv(). */
9635 : void
9636 1990 : XLogShutdownWalRcv(void)
9637 : {
9638 : Assert(AmStartupProcess() || !IsUnderPostmaster);
9639 :
9640 1990 : ShutdownWalRcv();
9641 1990 : ResetInstallXLogFileSegmentActive();
9642 1990 : }
9643 :
9644 : /* Enable WAL file recycling and preallocation. */
9645 : void
9646 2360 : SetInstallXLogFileSegmentActive(void)
9647 : {
9648 2360 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9649 2360 : XLogCtl->InstallXLogFileSegmentActive = true;
9650 2360 : LWLockRelease(ControlFileLock);
9651 2360 : }
9652 :
9653 : /* Disable WAL file recycling and preallocation. */
9654 : void
9655 2286 : ResetInstallXLogFileSegmentActive(void)
9656 : {
9657 2286 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9658 2286 : XLogCtl->InstallXLogFileSegmentActive = false;
9659 2286 : LWLockRelease(ControlFileLock);
9660 2286 : }
9661 :
9662 : bool
9663 0 : IsInstallXLogFileSegmentActive(void)
9664 : {
9665 : bool result;
9666 :
9667 0 : LWLockAcquire(ControlFileLock, LW_SHARED);
9668 0 : result = XLogCtl->InstallXLogFileSegmentActive;
9669 0 : LWLockRelease(ControlFileLock);
9670 :
9671 0 : return result;
9672 : }
9673 :
9674 : /*
9675 : * Update the WalWriterSleeping flag.
9676 : */
9677 : void
9678 1060 : SetWalWriterSleeping(bool sleeping)
9679 : {
9680 1060 : SpinLockAcquire(&XLogCtl->info_lck);
9681 1060 : XLogCtl->WalWriterSleeping = sleeping;
9682 1060 : SpinLockRelease(&XLogCtl->info_lck);
9683 1060 : }
|
