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 : static XLogRecPtr XLogGetReplicationSlotMinimumLSN(void);
673 :
674 : static void AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli,
675 : bool opportunistic);
676 : static void XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible);
677 : static bool InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
678 : bool find_free, XLogSegNo max_segno,
679 : TimeLineID tli);
680 : static void XLogFileClose(void);
681 : static void PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli);
682 : static void RemoveTempXlogFiles(void);
683 : static void RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr,
684 : XLogRecPtr endptr, TimeLineID insertTLI);
685 : static void RemoveXlogFile(const struct dirent *segment_de,
686 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
687 : TimeLineID insertTLI);
688 : static void UpdateLastRemovedPtr(char *filename);
689 : static void ValidateXLOGDirectoryStructure(void);
690 : static void CleanupBackupHistory(void);
691 : static void UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force);
692 : static bool PerformRecoveryXLogAction(void);
693 : static void InitControlFile(uint64 sysidentifier, uint32 data_checksum_version);
694 : static void WriteControlFile(void);
695 : static void ReadControlFile(void);
696 : static void UpdateControlFile(void);
697 : static char *str_time(pg_time_t tnow, char *buf, size_t bufsize);
698 :
699 : static int get_sync_bit(int method);
700 :
701 : static void CopyXLogRecordToWAL(int write_len, bool isLogSwitch,
702 : XLogRecData *rdata,
703 : XLogRecPtr StartPos, XLogRecPtr EndPos,
704 : TimeLineID tli);
705 : static void ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos,
706 : XLogRecPtr *EndPos, XLogRecPtr *PrevPtr);
707 : static bool ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos,
708 : XLogRecPtr *PrevPtr);
709 : static XLogRecPtr WaitXLogInsertionsToFinish(XLogRecPtr upto);
710 : static char *GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli);
711 : static XLogRecPtr XLogBytePosToRecPtr(uint64 bytepos);
712 : static XLogRecPtr XLogBytePosToEndRecPtr(uint64 bytepos);
713 : static uint64 XLogRecPtrToBytePos(XLogRecPtr ptr);
714 :
715 : static void WALInsertLockAcquire(void);
716 : static void WALInsertLockAcquireExclusive(void);
717 : static void WALInsertLockRelease(void);
718 : static void WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt);
719 :
720 : /*
721 : * Insert an XLOG record represented by an already-constructed chain of data
722 : * chunks. This is a low-level routine; to construct the WAL record header
723 : * and data, use the higher-level routines in xloginsert.c.
724 : *
725 : * If 'fpw_lsn' is valid, it is the oldest LSN among the pages that this
726 : * WAL record applies to, that were not included in the record as full page
727 : * images. If fpw_lsn <= RedoRecPtr, the function does not perform the
728 : * insertion and returns InvalidXLogRecPtr. The caller can then recalculate
729 : * which pages need a full-page image, and retry. If fpw_lsn is invalid, the
730 : * record is always inserted.
731 : *
732 : * 'flags' gives more in-depth control on the record being inserted. See
733 : * XLogSetRecordFlags() for details.
734 : *
735 : * 'topxid_included' tells whether the top-transaction id is logged along with
736 : * current subtransaction. See XLogRecordAssemble().
737 : *
738 : * The first XLogRecData in the chain must be for the record header, and its
739 : * data must be MAXALIGNed. XLogInsertRecord fills in the xl_prev and
740 : * xl_crc fields in the header, the rest of the header must already be filled
741 : * by the caller.
742 : *
743 : * Returns XLOG pointer to end of record (beginning of next record).
744 : * This can be used as LSN for data pages affected by the logged action.
745 : * (LSN is the XLOG point up to which the XLOG must be flushed to disk
746 : * before the data page can be written out. This implements the basic
747 : * WAL rule "write the log before the data".)
748 : */
749 : XLogRecPtr
750 30770350 : XLogInsertRecord(XLogRecData *rdata,
751 : XLogRecPtr fpw_lsn,
752 : uint8 flags,
753 : int num_fpi,
754 : uint64 fpi_bytes,
755 : bool topxid_included)
756 : {
757 30770350 : XLogCtlInsert *Insert = &XLogCtl->Insert;
758 : pg_crc32c rdata_crc;
759 : bool inserted;
760 30770350 : XLogRecord *rechdr = (XLogRecord *) rdata->data;
761 30770350 : uint8 info = rechdr->xl_info & ~XLR_INFO_MASK;
762 30770350 : WalInsertClass class = WALINSERT_NORMAL;
763 : XLogRecPtr StartPos;
764 : XLogRecPtr EndPos;
765 30770350 : bool prevDoPageWrites = doPageWrites;
766 : TimeLineID insertTLI;
767 :
768 : /* Does this record type require special handling? */
769 30770350 : if (unlikely(rechdr->xl_rmid == RM_XLOG_ID))
770 : {
771 462000 : if (info == XLOG_SWITCH)
772 1576 : class = WALINSERT_SPECIAL_SWITCH;
773 460424 : else if (info == XLOG_CHECKPOINT_REDO)
774 1852 : class = WALINSERT_SPECIAL_CHECKPOINT;
775 : }
776 :
777 : /* we assume that all of the record header is in the first chunk */
778 : Assert(rdata->len >= SizeOfXLogRecord);
779 :
780 : /* cross-check on whether we should be here or not */
781 30770350 : if (!XLogInsertAllowed())
782 0 : elog(ERROR, "cannot make new WAL entries during recovery");
783 :
784 : /*
785 : * Given that we're not in recovery, InsertTimeLineID is set and can't
786 : * change, so we can read it without a lock.
787 : */
788 30770350 : insertTLI = XLogCtl->InsertTimeLineID;
789 :
790 : /*----------
791 : *
792 : * We have now done all the preparatory work we can without holding a
793 : * lock or modifying shared state. From here on, inserting the new WAL
794 : * record to the shared WAL buffer cache is a two-step process:
795 : *
796 : * 1. Reserve the right amount of space from the WAL. The current head of
797 : * reserved space is kept in Insert->CurrBytePos, and is protected by
798 : * insertpos_lck.
799 : *
800 : * 2. Copy the record to the reserved WAL space. This involves finding the
801 : * correct WAL buffer containing the reserved space, and copying the
802 : * record in place. This can be done concurrently in multiple processes.
803 : *
804 : * To keep track of which insertions are still in-progress, each concurrent
805 : * inserter acquires an insertion lock. In addition to just indicating that
806 : * an insertion is in progress, the lock tells others how far the inserter
807 : * has progressed. There is a small fixed number of insertion locks,
808 : * determined by NUM_XLOGINSERT_LOCKS. When an inserter crosses a page
809 : * boundary, it updates the value stored in the lock to the how far it has
810 : * inserted, to allow the previous buffer to be flushed.
811 : *
812 : * Holding onto an insertion lock also protects RedoRecPtr and
813 : * fullPageWrites from changing until the insertion is finished.
814 : *
815 : * Step 2 can usually be done completely in parallel. If the required WAL
816 : * page is not initialized yet, you have to grab WALBufMappingLock to
817 : * initialize it, but the WAL writer tries to do that ahead of insertions
818 : * to avoid that from happening in the critical path.
819 : *
820 : *----------
821 : */
822 30770350 : START_CRIT_SECTION();
823 :
824 30770350 : if (likely(class == WALINSERT_NORMAL))
825 : {
826 30766922 : WALInsertLockAcquire();
827 :
828 : /*
829 : * Check to see if my copy of RedoRecPtr is out of date. If so, may
830 : * have to go back and have the caller recompute everything. This can
831 : * only happen just after a checkpoint, so it's better to be slow in
832 : * this case and fast otherwise.
833 : *
834 : * Also check to see if fullPageWrites was just turned on or there's a
835 : * running backup (which forces full-page writes); if we weren't
836 : * already doing full-page writes then go back and recompute.
837 : *
838 : * If we aren't doing full-page writes then RedoRecPtr doesn't
839 : * actually affect the contents of the XLOG record, so we'll update
840 : * our local copy but not force a recomputation. (If doPageWrites was
841 : * just turned off, we could recompute the record without full pages,
842 : * but we choose not to bother.)
843 : */
844 30766922 : if (RedoRecPtr != Insert->RedoRecPtr)
845 : {
846 : Assert(RedoRecPtr < Insert->RedoRecPtr);
847 14532 : RedoRecPtr = Insert->RedoRecPtr;
848 : }
849 30766922 : doPageWrites = (Insert->fullPageWrites || Insert->runningBackups > 0);
850 :
851 30766922 : if (doPageWrites &&
852 30269794 : (!prevDoPageWrites ||
853 27791954 : (XLogRecPtrIsValid(fpw_lsn) && fpw_lsn <= RedoRecPtr)))
854 : {
855 : /*
856 : * Oops, some buffer now needs to be backed up that the caller
857 : * didn't back up. Start over.
858 : */
859 15974 : WALInsertLockRelease();
860 15974 : END_CRIT_SECTION();
861 15974 : return InvalidXLogRecPtr;
862 : }
863 :
864 : /*
865 : * Reserve space for the record in the WAL. This also sets the xl_prev
866 : * pointer.
867 : */
868 30750948 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
869 : &rechdr->xl_prev);
870 :
871 : /* Normal records are always inserted. */
872 30750948 : inserted = true;
873 : }
874 3428 : else if (class == WALINSERT_SPECIAL_SWITCH)
875 : {
876 : /*
877 : * In order to insert an XLOG_SWITCH record, we need to hold all of
878 : * the WAL insertion locks, not just one, so that no one else can
879 : * begin inserting a record until we've figured out how much space
880 : * remains in the current WAL segment and claimed all of it.
881 : *
882 : * Nonetheless, this case is simpler than the normal cases handled
883 : * below, which must check for changes in doPageWrites and RedoRecPtr.
884 : * Those checks are only needed for records that can contain buffer
885 : * references, and an XLOG_SWITCH record never does.
886 : */
887 : Assert(!XLogRecPtrIsValid(fpw_lsn));
888 1576 : WALInsertLockAcquireExclusive();
889 1576 : inserted = ReserveXLogSwitch(&StartPos, &EndPos, &rechdr->xl_prev);
890 : }
891 : else
892 : {
893 : Assert(class == WALINSERT_SPECIAL_CHECKPOINT);
894 :
895 : /*
896 : * We need to update both the local and shared copies of RedoRecPtr,
897 : * which means that we need to hold all the WAL insertion locks.
898 : * However, there can't be any buffer references, so as above, we need
899 : * not check RedoRecPtr before inserting the record; we just need to
900 : * update it afterwards.
901 : */
902 : Assert(!XLogRecPtrIsValid(fpw_lsn));
903 1852 : WALInsertLockAcquireExclusive();
904 1852 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
905 : &rechdr->xl_prev);
906 1852 : RedoRecPtr = Insert->RedoRecPtr = StartPos;
907 1852 : inserted = true;
908 : }
909 :
910 30754376 : if (inserted)
911 : {
912 : /*
913 : * Now that xl_prev has been filled in, calculate CRC of the record
914 : * header.
915 : */
916 30754264 : rdata_crc = rechdr->xl_crc;
917 30754264 : COMP_CRC32C(rdata_crc, rechdr, offsetof(XLogRecord, xl_crc));
918 30754264 : FIN_CRC32C(rdata_crc);
919 30754264 : rechdr->xl_crc = rdata_crc;
920 :
921 : /*
922 : * All the record data, including the header, is now ready to be
923 : * inserted. Copy the record in the space reserved.
924 : */
925 30754264 : CopyXLogRecordToWAL(rechdr->xl_tot_len,
926 : class == WALINSERT_SPECIAL_SWITCH, rdata,
927 : StartPos, EndPos, insertTLI);
928 :
929 : /*
930 : * Unless record is flagged as not important, update LSN of last
931 : * important record in the current slot. When holding all locks, just
932 : * update the first one.
933 : */
934 30754264 : if ((flags & XLOG_MARK_UNIMPORTANT) == 0)
935 : {
936 30512640 : int lockno = holdingAllLocks ? 0 : MyLockNo;
937 :
938 30512640 : WALInsertLocks[lockno].l.lastImportantAt = StartPos;
939 : }
940 : }
941 : else
942 : {
943 : /*
944 : * This was an xlog-switch record, but the current insert location was
945 : * already exactly at the beginning of a segment, so there was no need
946 : * to do anything.
947 : */
948 : }
949 :
950 : /*
951 : * Done! Let others know that we're finished.
952 : */
953 30754376 : WALInsertLockRelease();
954 :
955 30754376 : END_CRIT_SECTION();
956 :
957 30754376 : MarkCurrentTransactionIdLoggedIfAny();
958 :
959 : /*
960 : * Mark top transaction id is logged (if needed) so that we should not try
961 : * to log it again with the next WAL record in the current subtransaction.
962 : */
963 30754376 : if (topxid_included)
964 438 : MarkSubxactTopXidLogged();
965 :
966 : /*
967 : * Update shared LogwrtRqst.Write, if we crossed page boundary.
968 : */
969 30754376 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
970 : {
971 3390266 : SpinLockAcquire(&XLogCtl->info_lck);
972 : /* advance global request to include new block(s) */
973 3390266 : if (XLogCtl->LogwrtRqst.Write < EndPos)
974 3220534 : XLogCtl->LogwrtRqst.Write = EndPos;
975 3390266 : SpinLockRelease(&XLogCtl->info_lck);
976 3390266 : RefreshXLogWriteResult(LogwrtResult);
977 : }
978 :
979 : /*
980 : * If this was an XLOG_SWITCH record, flush the record and the empty
981 : * padding space that fills the rest of the segment, and perform
982 : * end-of-segment actions (eg, notifying archiver).
983 : */
984 30754376 : if (class == WALINSERT_SPECIAL_SWITCH)
985 : {
986 : TRACE_POSTGRESQL_WAL_SWITCH();
987 1576 : XLogFlush(EndPos);
988 :
989 : /*
990 : * Even though we reserved the rest of the segment for us, which is
991 : * reflected in EndPos, we return a pointer to just the end of the
992 : * xlog-switch record.
993 : */
994 1576 : if (inserted)
995 : {
996 1464 : EndPos = StartPos + SizeOfXLogRecord;
997 1464 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
998 : {
999 2 : uint64 offset = XLogSegmentOffset(EndPos, wal_segment_size);
1000 :
1001 2 : if (offset == EndPos % XLOG_BLCKSZ)
1002 0 : EndPos += SizeOfXLogLongPHD;
1003 : else
1004 2 : EndPos += SizeOfXLogShortPHD;
1005 : }
1006 : }
1007 : }
1008 :
1009 : #ifdef WAL_DEBUG
1010 : if (XLOG_DEBUG)
1011 : {
1012 : static XLogReaderState *debug_reader = NULL;
1013 : XLogRecord *record;
1014 : DecodedXLogRecord *decoded;
1015 : StringInfoData buf;
1016 : StringInfoData recordBuf;
1017 : char *errormsg = NULL;
1018 : MemoryContext oldCxt;
1019 :
1020 : oldCxt = MemoryContextSwitchTo(walDebugCxt);
1021 :
1022 : initStringInfo(&buf);
1023 : appendStringInfo(&buf, "INSERT @ %X/%08X: ", LSN_FORMAT_ARGS(EndPos));
1024 :
1025 : /*
1026 : * We have to piece together the WAL record data from the XLogRecData
1027 : * entries, so that we can pass it to the rm_desc function as one
1028 : * contiguous chunk.
1029 : */
1030 : initStringInfo(&recordBuf);
1031 : for (; rdata != NULL; rdata = rdata->next)
1032 : appendBinaryStringInfo(&recordBuf, rdata->data, rdata->len);
1033 :
1034 : /* We also need temporary space to decode the record. */
1035 : record = (XLogRecord *) recordBuf.data;
1036 : decoded = (DecodedXLogRecord *)
1037 : palloc(DecodeXLogRecordRequiredSpace(record->xl_tot_len));
1038 :
1039 : if (!debug_reader)
1040 : debug_reader = XLogReaderAllocate(wal_segment_size, NULL,
1041 : XL_ROUTINE(.page_read = NULL,
1042 : .segment_open = NULL,
1043 : .segment_close = NULL),
1044 : NULL);
1045 : if (!debug_reader)
1046 : {
1047 : appendStringInfoString(&buf, "error decoding record: out of memory while allocating a WAL reading processor");
1048 : }
1049 : else if (!DecodeXLogRecord(debug_reader,
1050 : decoded,
1051 : record,
1052 : EndPos,
1053 : &errormsg))
1054 : {
1055 : appendStringInfo(&buf, "error decoding record: %s",
1056 : errormsg ? errormsg : "no error message");
1057 : }
1058 : else
1059 : {
1060 : appendStringInfoString(&buf, " - ");
1061 :
1062 : debug_reader->record = decoded;
1063 : xlog_outdesc(&buf, debug_reader);
1064 : debug_reader->record = NULL;
1065 : }
1066 : elog(LOG, "%s", buf.data);
1067 :
1068 : pfree(decoded);
1069 : pfree(buf.data);
1070 : pfree(recordBuf.data);
1071 : MemoryContextSwitchTo(oldCxt);
1072 : }
1073 : #endif
1074 :
1075 : /*
1076 : * Update our global variables
1077 : */
1078 30754376 : ProcLastRecPtr = StartPos;
1079 30754376 : XactLastRecEnd = EndPos;
1080 :
1081 : /* Report WAL traffic to the instrumentation. */
1082 30754376 : if (inserted)
1083 : {
1084 30754264 : pgWalUsage.wal_bytes += rechdr->xl_tot_len;
1085 30754264 : pgWalUsage.wal_records++;
1086 30754264 : pgWalUsage.wal_fpi += num_fpi;
1087 30754264 : pgWalUsage.wal_fpi_bytes += fpi_bytes;
1088 :
1089 : /* Required for the flush of pending stats WAL data */
1090 30754264 : pgstat_report_fixed = true;
1091 : }
1092 :
1093 30754376 : return EndPos;
1094 : }
1095 :
1096 : /*
1097 : * Reserves the right amount of space for a record of given size from the WAL.
1098 : * *StartPos is set to the beginning of the reserved section, *EndPos to
1099 : * its end+1. *PrevPtr is set to the beginning of the previous record; it is
1100 : * used to set the xl_prev of this record.
1101 : *
1102 : * This is the performance critical part of XLogInsert that must be serialized
1103 : * across backends. The rest can happen mostly in parallel. Try to keep this
1104 : * section as short as possible, insertpos_lck can be heavily contended on a
1105 : * busy system.
1106 : *
1107 : * NB: The space calculation here must match the code in CopyXLogRecordToWAL,
1108 : * where we actually copy the record to the reserved space.
1109 : *
1110 : * NB: Testing shows that XLogInsertRecord runs faster if this code is inlined;
1111 : * however, because there are two call sites, the compiler is reluctant to
1112 : * inline. We use pg_attribute_always_inline here to try to convince it.
1113 : */
1114 : static pg_attribute_always_inline void
1115 30752800 : ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1116 : XLogRecPtr *PrevPtr)
1117 : {
1118 30752800 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1119 : uint64 startbytepos;
1120 : uint64 endbytepos;
1121 : uint64 prevbytepos;
1122 :
1123 30752800 : size = MAXALIGN(size);
1124 :
1125 : /* All (non xlog-switch) records should contain data. */
1126 : Assert(size > SizeOfXLogRecord);
1127 :
1128 : /*
1129 : * The duration the spinlock needs to be held is minimized by minimizing
1130 : * the calculations that have to be done while holding the lock. The
1131 : * current tip of reserved WAL is kept in CurrBytePos, as a byte position
1132 : * that only counts "usable" bytes in WAL, that is, it excludes all WAL
1133 : * page headers. The mapping between "usable" byte positions and physical
1134 : * positions (XLogRecPtrs) can be done outside the locked region, and
1135 : * because the usable byte position doesn't include any headers, reserving
1136 : * X bytes from WAL is almost as simple as "CurrBytePos += X".
1137 : */
1138 30752800 : SpinLockAcquire(&Insert->insertpos_lck);
1139 :
1140 30752800 : startbytepos = Insert->CurrBytePos;
1141 30752800 : endbytepos = startbytepos + size;
1142 30752800 : prevbytepos = Insert->PrevBytePos;
1143 30752800 : Insert->CurrBytePos = endbytepos;
1144 30752800 : Insert->PrevBytePos = startbytepos;
1145 :
1146 30752800 : SpinLockRelease(&Insert->insertpos_lck);
1147 :
1148 30752800 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1149 30752800 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1150 30752800 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1151 :
1152 : /*
1153 : * Check that the conversions between "usable byte positions" and
1154 : * XLogRecPtrs work consistently in both directions.
1155 : */
1156 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1157 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1158 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1159 30752800 : }
1160 :
1161 : /*
1162 : * Like ReserveXLogInsertLocation(), but for an xlog-switch record.
1163 : *
1164 : * A log-switch record is handled slightly differently. The rest of the
1165 : * segment will be reserved for this insertion, as indicated by the returned
1166 : * *EndPos value. However, if we are already at the beginning of the current
1167 : * segment, *StartPos and *EndPos are set to the current location without
1168 : * reserving any space, and the function returns false.
1169 : */
1170 : static bool
1171 1576 : ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1172 : {
1173 1576 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1174 : uint64 startbytepos;
1175 : uint64 endbytepos;
1176 : uint64 prevbytepos;
1177 1576 : uint32 size = MAXALIGN(SizeOfXLogRecord);
1178 : XLogRecPtr ptr;
1179 : uint32 segleft;
1180 :
1181 : /*
1182 : * These calculations are a bit heavy-weight to be done while holding a
1183 : * spinlock, but since we're holding all the WAL insertion locks, there
1184 : * are no other inserters competing for it. GetXLogInsertRecPtr() does
1185 : * compete for it, but that's not called very frequently.
1186 : */
1187 1576 : SpinLockAcquire(&Insert->insertpos_lck);
1188 :
1189 1576 : startbytepos = Insert->CurrBytePos;
1190 :
1191 1576 : ptr = XLogBytePosToEndRecPtr(startbytepos);
1192 1576 : if (XLogSegmentOffset(ptr, wal_segment_size) == 0)
1193 : {
1194 112 : SpinLockRelease(&Insert->insertpos_lck);
1195 112 : *EndPos = *StartPos = ptr;
1196 112 : return false;
1197 : }
1198 :
1199 1464 : endbytepos = startbytepos + size;
1200 1464 : prevbytepos = Insert->PrevBytePos;
1201 :
1202 1464 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1203 1464 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1204 :
1205 1464 : segleft = wal_segment_size - XLogSegmentOffset(*EndPos, wal_segment_size);
1206 1464 : if (segleft != wal_segment_size)
1207 : {
1208 : /* consume the rest of the segment */
1209 1464 : *EndPos += segleft;
1210 1464 : endbytepos = XLogRecPtrToBytePos(*EndPos);
1211 : }
1212 1464 : Insert->CurrBytePos = endbytepos;
1213 1464 : Insert->PrevBytePos = startbytepos;
1214 :
1215 1464 : SpinLockRelease(&Insert->insertpos_lck);
1216 :
1217 1464 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1218 :
1219 : Assert(XLogSegmentOffset(*EndPos, wal_segment_size) == 0);
1220 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1221 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1222 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1223 :
1224 1464 : return true;
1225 : }
1226 :
1227 : /*
1228 : * Subroutine of XLogInsertRecord. Copies a WAL record to an already-reserved
1229 : * area in the WAL.
1230 : */
1231 : static void
1232 30754264 : CopyXLogRecordToWAL(int write_len, bool isLogSwitch, XLogRecData *rdata,
1233 : XLogRecPtr StartPos, XLogRecPtr EndPos, TimeLineID tli)
1234 : {
1235 : char *currpos;
1236 : int freespace;
1237 : int written;
1238 : XLogRecPtr CurrPos;
1239 : XLogPageHeader pagehdr;
1240 :
1241 : /*
1242 : * Get a pointer to the right place in the right WAL buffer to start
1243 : * inserting to.
1244 : */
1245 30754264 : CurrPos = StartPos;
1246 30754264 : currpos = GetXLogBuffer(CurrPos, tli);
1247 30754264 : freespace = INSERT_FREESPACE(CurrPos);
1248 :
1249 : /*
1250 : * there should be enough space for at least the first field (xl_tot_len)
1251 : * on this page.
1252 : */
1253 : Assert(freespace >= sizeof(uint32));
1254 :
1255 : /* Copy record data */
1256 30754264 : written = 0;
1257 145109586 : while (rdata != NULL)
1258 : {
1259 114355322 : const char *rdata_data = rdata->data;
1260 114355322 : int rdata_len = rdata->len;
1261 :
1262 117964058 : while (rdata_len > freespace)
1263 : {
1264 : /*
1265 : * Write what fits on this page, and continue on the next page.
1266 : */
1267 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || freespace == 0);
1268 3608736 : memcpy(currpos, rdata_data, freespace);
1269 3608736 : rdata_data += freespace;
1270 3608736 : rdata_len -= freespace;
1271 3608736 : written += freespace;
1272 3608736 : CurrPos += freespace;
1273 :
1274 : /*
1275 : * Get pointer to beginning of next page, and set the xlp_rem_len
1276 : * in the page header. Set XLP_FIRST_IS_CONTRECORD.
1277 : *
1278 : * It's safe to set the contrecord flag and xlp_rem_len without a
1279 : * lock on the page. All the other flags were already set when the
1280 : * page was initialized, in AdvanceXLInsertBuffer, and we're the
1281 : * only backend that needs to set the contrecord flag.
1282 : */
1283 3608736 : currpos = GetXLogBuffer(CurrPos, tli);
1284 3608736 : pagehdr = (XLogPageHeader) currpos;
1285 3608736 : pagehdr->xlp_rem_len = write_len - written;
1286 3608736 : pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1287 :
1288 : /* skip over the page header */
1289 3608736 : if (XLogSegmentOffset(CurrPos, wal_segment_size) == 0)
1290 : {
1291 2330 : CurrPos += SizeOfXLogLongPHD;
1292 2330 : currpos += SizeOfXLogLongPHD;
1293 : }
1294 : else
1295 : {
1296 3606406 : CurrPos += SizeOfXLogShortPHD;
1297 3606406 : currpos += SizeOfXLogShortPHD;
1298 : }
1299 3608736 : freespace = INSERT_FREESPACE(CurrPos);
1300 : }
1301 :
1302 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1303 114355322 : memcpy(currpos, rdata_data, rdata_len);
1304 114355322 : currpos += rdata_len;
1305 114355322 : CurrPos += rdata_len;
1306 114355322 : freespace -= rdata_len;
1307 114355322 : written += rdata_len;
1308 :
1309 114355322 : rdata = rdata->next;
1310 : }
1311 : Assert(written == write_len);
1312 :
1313 : /*
1314 : * If this was an xlog-switch, it's not enough to write the switch record,
1315 : * we also have to consume all the remaining space in the WAL segment. We
1316 : * have already reserved that space, but we need to actually fill it.
1317 : */
1318 30754264 : if (isLogSwitch && XLogSegmentOffset(CurrPos, wal_segment_size) != 0)
1319 : {
1320 : /* An xlog-switch record doesn't contain any data besides the header */
1321 : Assert(write_len == SizeOfXLogRecord);
1322 :
1323 : /* Assert that we did reserve the right amount of space */
1324 : Assert(XLogSegmentOffset(EndPos, wal_segment_size) == 0);
1325 :
1326 : /* Use up all the remaining space on the current page */
1327 1464 : CurrPos += freespace;
1328 :
1329 : /*
1330 : * Cause all remaining pages in the segment to be flushed, leaving the
1331 : * XLog position where it should be, at the start of the next segment.
1332 : * We do this one page at a time, to make sure we don't deadlock
1333 : * against ourselves if wal_buffers < wal_segment_size.
1334 : */
1335 1422228 : while (CurrPos < EndPos)
1336 : {
1337 : /*
1338 : * The minimal action to flush the page would be to call
1339 : * WALInsertLockUpdateInsertingAt(CurrPos) followed by
1340 : * AdvanceXLInsertBuffer(...). The page would be left initialized
1341 : * mostly to zeros, except for the page header (always the short
1342 : * variant, as this is never a segment's first page).
1343 : *
1344 : * The large vistas of zeros are good for compressibility, but the
1345 : * headers interrupting them every XLOG_BLCKSZ (with values that
1346 : * differ from page to page) are not. The effect varies with
1347 : * compression tool, but bzip2 for instance compresses about an
1348 : * order of magnitude worse if those headers are left in place.
1349 : *
1350 : * Rather than complicating AdvanceXLInsertBuffer itself (which is
1351 : * called in heavily-loaded circumstances as well as this lightly-
1352 : * loaded one) with variant behavior, we just use GetXLogBuffer
1353 : * (which itself calls the two methods we need) to get the pointer
1354 : * and zero most of the page. Then we just zero the page header.
1355 : */
1356 1420764 : currpos = GetXLogBuffer(CurrPos, tli);
1357 5683056 : MemSet(currpos, 0, SizeOfXLogShortPHD);
1358 :
1359 1420764 : CurrPos += XLOG_BLCKSZ;
1360 : }
1361 : }
1362 : else
1363 : {
1364 : /* Align the end position, so that the next record starts aligned */
1365 30752800 : CurrPos = MAXALIGN64(CurrPos);
1366 : }
1367 :
1368 30754264 : if (CurrPos != EndPos)
1369 0 : ereport(PANIC,
1370 : errcode(ERRCODE_DATA_CORRUPTED),
1371 : errmsg_internal("space reserved for WAL record does not match what was written"));
1372 30754264 : }
1373 :
1374 : /*
1375 : * Acquire a WAL insertion lock, for inserting to WAL.
1376 : */
1377 : static void
1378 30766944 : WALInsertLockAcquire(void)
1379 : {
1380 : bool immed;
1381 :
1382 : /*
1383 : * It doesn't matter which of the WAL insertion locks we acquire, so try
1384 : * the one we used last time. If the system isn't particularly busy, it's
1385 : * a good bet that it's still available, and it's good to have some
1386 : * affinity to a particular lock so that you don't unnecessarily bounce
1387 : * cache lines between processes when there's no contention.
1388 : *
1389 : * If this is the first time through in this backend, pick a lock
1390 : * (semi-)randomly. This allows the locks to be used evenly if you have a
1391 : * lot of very short connections.
1392 : */
1393 : static int lockToTry = -1;
1394 :
1395 30766944 : if (lockToTry == -1)
1396 16798 : lockToTry = MyProcNumber % NUM_XLOGINSERT_LOCKS;
1397 30766944 : MyLockNo = lockToTry;
1398 :
1399 : /*
1400 : * The insertingAt value is initially set to 0, as we don't know our
1401 : * insert location yet.
1402 : */
1403 30766944 : immed = LWLockAcquire(&WALInsertLocks[MyLockNo].l.lock, LW_EXCLUSIVE);
1404 30766944 : if (!immed)
1405 : {
1406 : /*
1407 : * If we couldn't get the lock immediately, try another lock next
1408 : * time. On a system with more insertion locks than concurrent
1409 : * inserters, this causes all the inserters to eventually migrate to a
1410 : * lock that no-one else is using. On a system with more inserters
1411 : * than locks, it still helps to distribute the inserters evenly
1412 : * across the locks.
1413 : */
1414 54172 : lockToTry = (lockToTry + 1) % NUM_XLOGINSERT_LOCKS;
1415 : }
1416 30766944 : }
1417 :
1418 : /*
1419 : * Acquire all WAL insertion locks, to prevent other backends from inserting
1420 : * to WAL.
1421 : */
1422 : static void
1423 8862 : WALInsertLockAcquireExclusive(void)
1424 : {
1425 : int i;
1426 :
1427 : /*
1428 : * When holding all the locks, all but the last lock's insertingAt
1429 : * indicator is set to 0xFFFFFFFFFFFFFFFF, which is higher than any real
1430 : * XLogRecPtr value, to make sure that no-one blocks waiting on those.
1431 : */
1432 70896 : for (i = 0; i < NUM_XLOGINSERT_LOCKS - 1; i++)
1433 : {
1434 62034 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1435 62034 : LWLockUpdateVar(&WALInsertLocks[i].l.lock,
1436 62034 : &WALInsertLocks[i].l.insertingAt,
1437 : PG_UINT64_MAX);
1438 : }
1439 : /* Variable value reset to 0 at release */
1440 8862 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1441 :
1442 8862 : holdingAllLocks = true;
1443 8862 : }
1444 :
1445 : /*
1446 : * Release our insertion lock (or locks, if we're holding them all).
1447 : *
1448 : * NB: Reset all variables to 0, so they cause LWLockWaitForVar to block the
1449 : * next time the lock is acquired.
1450 : */
1451 : static void
1452 30775806 : WALInsertLockRelease(void)
1453 : {
1454 30775806 : if (holdingAllLocks)
1455 : {
1456 : int i;
1457 :
1458 79758 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1459 70896 : LWLockReleaseClearVar(&WALInsertLocks[i].l.lock,
1460 70896 : &WALInsertLocks[i].l.insertingAt,
1461 : 0);
1462 :
1463 8862 : holdingAllLocks = false;
1464 : }
1465 : else
1466 : {
1467 30766944 : LWLockReleaseClearVar(&WALInsertLocks[MyLockNo].l.lock,
1468 30766944 : &WALInsertLocks[MyLockNo].l.insertingAt,
1469 : 0);
1470 : }
1471 30775806 : }
1472 :
1473 : /*
1474 : * Update our insertingAt value, to let others know that we've finished
1475 : * inserting up to that point.
1476 : */
1477 : static void
1478 5330104 : WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt)
1479 : {
1480 5330104 : if (holdingAllLocks)
1481 : {
1482 : /*
1483 : * We use the last lock to mark our actual position, see comments in
1484 : * WALInsertLockAcquireExclusive.
1485 : */
1486 1415354 : LWLockUpdateVar(&WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.lock,
1487 1415354 : &WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.insertingAt,
1488 : insertingAt);
1489 : }
1490 : else
1491 3914750 : LWLockUpdateVar(&WALInsertLocks[MyLockNo].l.lock,
1492 3914750 : &WALInsertLocks[MyLockNo].l.insertingAt,
1493 : insertingAt);
1494 5330104 : }
1495 :
1496 : /*
1497 : * Wait for any WAL insertions < upto to finish.
1498 : *
1499 : * Returns the location of the oldest insertion that is still in-progress.
1500 : * Any WAL prior to that point has been fully copied into WAL buffers, and
1501 : * can be flushed out to disk. Because this waits for any insertions older
1502 : * than 'upto' to finish, the return value is always >= 'upto'.
1503 : *
1504 : * Note: When you are about to write out WAL, you must call this function
1505 : * *before* acquiring WALWriteLock, to avoid deadlocks. This function might
1506 : * need to wait for an insertion to finish (or at least advance to next
1507 : * uninitialized page), and the inserter might need to evict an old WAL buffer
1508 : * to make room for a new one, which in turn requires WALWriteLock.
1509 : */
1510 : static XLogRecPtr
1511 4880086 : WaitXLogInsertionsToFinish(XLogRecPtr upto)
1512 : {
1513 : uint64 bytepos;
1514 : XLogRecPtr inserted;
1515 : XLogRecPtr reservedUpto;
1516 : XLogRecPtr finishedUpto;
1517 4880086 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1518 : int i;
1519 :
1520 4880086 : if (MyProc == NULL)
1521 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1522 :
1523 : /*
1524 : * Check if there's any work to do. Use a barrier to ensure we get the
1525 : * freshest value.
1526 : */
1527 4880086 : inserted = pg_atomic_read_membarrier_u64(&XLogCtl->logInsertResult);
1528 4880086 : if (upto <= inserted)
1529 3995254 : return inserted;
1530 :
1531 : /* Read the current insert position */
1532 884832 : SpinLockAcquire(&Insert->insertpos_lck);
1533 884832 : bytepos = Insert->CurrBytePos;
1534 884832 : SpinLockRelease(&Insert->insertpos_lck);
1535 884832 : reservedUpto = XLogBytePosToEndRecPtr(bytepos);
1536 :
1537 : /*
1538 : * No-one should request to flush a piece of WAL that hasn't even been
1539 : * reserved yet. However, it can happen if there is a block with a bogus
1540 : * LSN on disk, for example. XLogFlush checks for that situation and
1541 : * complains, but only after the flush. Here we just assume that to mean
1542 : * that all WAL that has been reserved needs to be finished. In this
1543 : * corner-case, the return value can be smaller than 'upto' argument.
1544 : */
1545 884832 : if (upto > reservedUpto)
1546 : {
1547 0 : ereport(LOG,
1548 : errmsg("request to flush past end of generated WAL; request %X/%08X, current position %X/%08X",
1549 : LSN_FORMAT_ARGS(upto), LSN_FORMAT_ARGS(reservedUpto)));
1550 0 : upto = reservedUpto;
1551 : }
1552 :
1553 : /*
1554 : * Loop through all the locks, sleeping on any in-progress insert older
1555 : * than 'upto'.
1556 : *
1557 : * finishedUpto is our return value, indicating the point upto which all
1558 : * the WAL insertions have been finished. Initialize it to the head of
1559 : * reserved WAL, and as we iterate through the insertion locks, back it
1560 : * out for any insertion that's still in progress.
1561 : */
1562 884832 : finishedUpto = reservedUpto;
1563 7963488 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1564 : {
1565 7078656 : XLogRecPtr insertingat = InvalidXLogRecPtr;
1566 :
1567 : do
1568 : {
1569 : /*
1570 : * See if this insertion is in progress. LWLockWaitForVar will
1571 : * wait for the lock to be released, or for the 'value' to be set
1572 : * by a LWLockUpdateVar call. When a lock is initially acquired,
1573 : * its value is 0 (InvalidXLogRecPtr), which means that we don't
1574 : * know where it's inserting yet. We will have to wait for it. If
1575 : * it's a small insertion, the record will most likely fit on the
1576 : * same page and the inserter will release the lock without ever
1577 : * calling LWLockUpdateVar. But if it has to sleep, it will
1578 : * advertise the insertion point with LWLockUpdateVar before
1579 : * sleeping.
1580 : *
1581 : * In this loop we are only waiting for insertions that started
1582 : * before WaitXLogInsertionsToFinish was called. The lack of
1583 : * memory barriers in the loop means that we might see locks as
1584 : * "unused" that have since become used. This is fine because
1585 : * they only can be used for later insertions that we would not
1586 : * want to wait on anyway. Not taking a lock to acquire the
1587 : * current insertingAt value means that we might see older
1588 : * insertingAt values. This is also fine, because if we read a
1589 : * value too old, we will add ourselves to the wait queue, which
1590 : * contains atomic operations.
1591 : */
1592 7228686 : if (LWLockWaitForVar(&WALInsertLocks[i].l.lock,
1593 7228686 : &WALInsertLocks[i].l.insertingAt,
1594 : insertingat, &insertingat))
1595 : {
1596 : /* the lock was free, so no insertion in progress */
1597 5160134 : insertingat = InvalidXLogRecPtr;
1598 5160134 : break;
1599 : }
1600 :
1601 : /*
1602 : * This insertion is still in progress. Have to wait, unless the
1603 : * inserter has proceeded past 'upto'.
1604 : */
1605 2068552 : } while (insertingat < upto);
1606 :
1607 7078656 : if (XLogRecPtrIsValid(insertingat) && insertingat < finishedUpto)
1608 749922 : finishedUpto = insertingat;
1609 : }
1610 :
1611 : /*
1612 : * Advance the limit we know to have been inserted and return the freshest
1613 : * value we know of, which might be beyond what we requested if somebody
1614 : * is concurrently doing this with an 'upto' pointer ahead of us.
1615 : */
1616 884832 : finishedUpto = pg_atomic_monotonic_advance_u64(&XLogCtl->logInsertResult,
1617 : finishedUpto);
1618 :
1619 884832 : return finishedUpto;
1620 : }
1621 :
1622 : /*
1623 : * Get a pointer to the right location in the WAL buffer containing the
1624 : * given XLogRecPtr.
1625 : *
1626 : * If the page is not initialized yet, it is initialized. That might require
1627 : * evicting an old dirty buffer from the buffer cache, which means I/O.
1628 : *
1629 : * The caller must ensure that the page containing the requested location
1630 : * isn't evicted yet, and won't be evicted. The way to ensure that is to
1631 : * hold onto a WAL insertion lock with the insertingAt position set to
1632 : * something <= ptr. GetXLogBuffer() will update insertingAt if it needs
1633 : * to evict an old page from the buffer. (This means that once you call
1634 : * GetXLogBuffer() with a given 'ptr', you must not access anything before
1635 : * that point anymore, and must not call GetXLogBuffer() with an older 'ptr'
1636 : * later, because older buffers might be recycled already)
1637 : */
1638 : static char *
1639 35783786 : GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli)
1640 : {
1641 : int idx;
1642 : XLogRecPtr endptr;
1643 : static uint64 cachedPage = 0;
1644 : static char *cachedPos = NULL;
1645 : XLogRecPtr expectedEndPtr;
1646 :
1647 : /*
1648 : * Fast path for the common case that we need to access again the same
1649 : * page as last time.
1650 : */
1651 35783786 : if (ptr / XLOG_BLCKSZ == cachedPage)
1652 : {
1653 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1654 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1655 29404126 : return cachedPos + ptr % XLOG_BLCKSZ;
1656 : }
1657 :
1658 : /*
1659 : * The XLog buffer cache is organized so that a page is always loaded to a
1660 : * particular buffer. That way we can easily calculate the buffer a given
1661 : * page must be loaded into, from the XLogRecPtr alone.
1662 : */
1663 6379660 : idx = XLogRecPtrToBufIdx(ptr);
1664 :
1665 : /*
1666 : * See what page is loaded in the buffer at the moment. It could be the
1667 : * page we're looking for, or something older. It can't be anything newer
1668 : * - that would imply the page we're looking for has already been written
1669 : * out to disk and evicted, and the caller is responsible for making sure
1670 : * that doesn't happen.
1671 : *
1672 : * We don't hold a lock while we read the value. If someone is just about
1673 : * to initialize or has just initialized the page, it's possible that we
1674 : * get InvalidXLogRecPtr. That's ok, we'll grab the mapping lock (in
1675 : * AdvanceXLInsertBuffer) and retry if we see anything other than the page
1676 : * we're looking for.
1677 : */
1678 6379660 : expectedEndPtr = ptr;
1679 6379660 : expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
1680 :
1681 6379660 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1682 6379660 : if (expectedEndPtr != endptr)
1683 : {
1684 : XLogRecPtr initializedUpto;
1685 :
1686 : /*
1687 : * Before calling AdvanceXLInsertBuffer(), which can block, let others
1688 : * know how far we're finished with inserting the record.
1689 : *
1690 : * NB: If 'ptr' points to just after the page header, advertise a
1691 : * position at the beginning of the page rather than 'ptr' itself. If
1692 : * there are no other insertions running, someone might try to flush
1693 : * up to our advertised location. If we advertised a position after
1694 : * the page header, someone might try to flush the page header, even
1695 : * though page might actually not be initialized yet. As the first
1696 : * inserter on the page, we are effectively responsible for making
1697 : * sure that it's initialized, before we let insertingAt to move past
1698 : * the page header.
1699 : */
1700 5330104 : if (ptr % XLOG_BLCKSZ == SizeOfXLogShortPHD &&
1701 12698 : XLogSegmentOffset(ptr, wal_segment_size) > XLOG_BLCKSZ)
1702 12698 : initializedUpto = ptr - SizeOfXLogShortPHD;
1703 5317406 : else if (ptr % XLOG_BLCKSZ == SizeOfXLogLongPHD &&
1704 2544 : XLogSegmentOffset(ptr, wal_segment_size) < XLOG_BLCKSZ)
1705 1244 : initializedUpto = ptr - SizeOfXLogLongPHD;
1706 : else
1707 5316162 : initializedUpto = ptr;
1708 :
1709 5330104 : WALInsertLockUpdateInsertingAt(initializedUpto);
1710 :
1711 5330104 : AdvanceXLInsertBuffer(ptr, tli, false);
1712 5330104 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1713 :
1714 5330104 : if (expectedEndPtr != endptr)
1715 0 : elog(PANIC, "could not find WAL buffer for %X/%08X",
1716 : LSN_FORMAT_ARGS(ptr));
1717 : }
1718 : else
1719 : {
1720 : /*
1721 : * Make sure the initialization of the page is visible to us, and
1722 : * won't arrive later to overwrite the WAL data we write on the page.
1723 : */
1724 1049556 : pg_memory_barrier();
1725 : }
1726 :
1727 : /*
1728 : * Found the buffer holding this page. Return a pointer to the right
1729 : * offset within the page.
1730 : */
1731 6379660 : cachedPage = ptr / XLOG_BLCKSZ;
1732 6379660 : cachedPos = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1733 :
1734 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1735 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1736 :
1737 6379660 : return cachedPos + ptr % XLOG_BLCKSZ;
1738 : }
1739 :
1740 : /*
1741 : * Read WAL data directly from WAL buffers, if available. Returns the number
1742 : * of bytes read successfully.
1743 : *
1744 : * Fewer than 'count' bytes may be read if some of the requested WAL data has
1745 : * already been evicted.
1746 : *
1747 : * No locks are taken.
1748 : *
1749 : * Caller should ensure that it reads no further than LogwrtResult.Write
1750 : * (which should have been updated by the caller when determining how far to
1751 : * read). The 'tli' argument is only used as a convenient safety check so that
1752 : * callers do not read from WAL buffers on a historical timeline.
1753 : */
1754 : Size
1755 196616 : WALReadFromBuffers(char *dstbuf, XLogRecPtr startptr, Size count,
1756 : TimeLineID tli)
1757 : {
1758 196616 : char *pdst = dstbuf;
1759 196616 : XLogRecPtr recptr = startptr;
1760 : XLogRecPtr inserted;
1761 196616 : Size nbytes = count;
1762 :
1763 196616 : if (RecoveryInProgress() || tli != GetWALInsertionTimeLine())
1764 1778 : return 0;
1765 :
1766 : Assert(XLogRecPtrIsValid(startptr));
1767 :
1768 : /*
1769 : * Caller should ensure that the requested data has been inserted into WAL
1770 : * buffers before we try to read it.
1771 : */
1772 194838 : inserted = pg_atomic_read_u64(&XLogCtl->logInsertResult);
1773 194838 : if (startptr + count > inserted)
1774 0 : ereport(ERROR,
1775 : errmsg("cannot read past end of generated WAL: requested %X/%08X, current position %X/%08X",
1776 : LSN_FORMAT_ARGS(startptr + count),
1777 : LSN_FORMAT_ARGS(inserted)));
1778 :
1779 : /*
1780 : * Loop through the buffers without a lock. For each buffer, atomically
1781 : * read and verify the end pointer, then copy the data out, and finally
1782 : * re-read and re-verify the end pointer.
1783 : *
1784 : * Once a page is evicted, it never returns to the WAL buffers, so if the
1785 : * end pointer matches the expected end pointer before and after we copy
1786 : * the data, then the right page must have been present during the data
1787 : * copy. Read barriers are necessary to ensure that the data copy actually
1788 : * happens between the two verification steps.
1789 : *
1790 : * If either verification fails, we simply terminate the loop and return
1791 : * with the data that had been already copied out successfully.
1792 : */
1793 206560 : while (nbytes > 0)
1794 : {
1795 202874 : uint32 offset = recptr % XLOG_BLCKSZ;
1796 202874 : int idx = XLogRecPtrToBufIdx(recptr);
1797 : XLogRecPtr expectedEndPtr;
1798 : XLogRecPtr endptr;
1799 : const char *page;
1800 : const char *psrc;
1801 : Size npagebytes;
1802 :
1803 : /*
1804 : * Calculate the end pointer we expect in the xlblocks array if the
1805 : * correct page is present.
1806 : */
1807 202874 : expectedEndPtr = recptr + (XLOG_BLCKSZ - offset);
1808 :
1809 : /*
1810 : * First verification step: check that the correct page is present in
1811 : * the WAL buffers.
1812 : */
1813 202874 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1814 202874 : if (expectedEndPtr != endptr)
1815 191150 : break;
1816 :
1817 : /*
1818 : * The correct page is present (or was at the time the endptr was
1819 : * read; must re-verify later). Calculate pointer to source data and
1820 : * determine how much data to read from this page.
1821 : */
1822 11724 : page = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1823 11724 : psrc = page + offset;
1824 11724 : npagebytes = Min(nbytes, XLOG_BLCKSZ - offset);
1825 :
1826 : /*
1827 : * Ensure that the data copy and the first verification step are not
1828 : * reordered.
1829 : */
1830 11724 : pg_read_barrier();
1831 :
1832 : /* data copy */
1833 11724 : memcpy(pdst, psrc, npagebytes);
1834 :
1835 : /*
1836 : * Ensure that the data copy and the second verification step are not
1837 : * reordered.
1838 : */
1839 11724 : pg_read_barrier();
1840 :
1841 : /*
1842 : * Second verification step: check that the page we read from wasn't
1843 : * evicted while we were copying the data.
1844 : */
1845 11724 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1846 11724 : if (expectedEndPtr != endptr)
1847 2 : break;
1848 :
1849 11722 : pdst += npagebytes;
1850 11722 : recptr += npagebytes;
1851 11722 : nbytes -= npagebytes;
1852 : }
1853 :
1854 : Assert(pdst - dstbuf <= count);
1855 :
1856 194838 : return pdst - dstbuf;
1857 : }
1858 :
1859 : /*
1860 : * Converts a "usable byte position" to XLogRecPtr. A usable byte position
1861 : * is the position starting from the beginning of WAL, excluding all WAL
1862 : * page headers.
1863 : */
1864 : static XLogRecPtr
1865 61513868 : XLogBytePosToRecPtr(uint64 bytepos)
1866 : {
1867 : uint64 fullsegs;
1868 : uint64 fullpages;
1869 : uint64 bytesleft;
1870 : uint32 seg_offset;
1871 : XLogRecPtr result;
1872 :
1873 61513868 : fullsegs = bytepos / UsableBytesInSegment;
1874 61513868 : bytesleft = bytepos % UsableBytesInSegment;
1875 :
1876 61513868 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1877 : {
1878 : /* fits on first page of segment */
1879 116838 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1880 : }
1881 : else
1882 : {
1883 : /* account for the first page on segment with long header */
1884 61397030 : seg_offset = XLOG_BLCKSZ;
1885 61397030 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1886 :
1887 61397030 : fullpages = bytesleft / UsableBytesInPage;
1888 61397030 : bytesleft = bytesleft % UsableBytesInPage;
1889 :
1890 61397030 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1891 : }
1892 :
1893 61513868 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1894 :
1895 61513868 : return result;
1896 : }
1897 :
1898 : /*
1899 : * Like XLogBytePosToRecPtr, but if the position is at a page boundary,
1900 : * returns a pointer to the beginning of the page (ie. before page header),
1901 : * not to where the first xlog record on that page would go to. This is used
1902 : * when converting a pointer to the end of a record.
1903 : */
1904 : static XLogRecPtr
1905 31640672 : XLogBytePosToEndRecPtr(uint64 bytepos)
1906 : {
1907 : uint64 fullsegs;
1908 : uint64 fullpages;
1909 : uint64 bytesleft;
1910 : uint32 seg_offset;
1911 : XLogRecPtr result;
1912 :
1913 31640672 : fullsegs = bytepos / UsableBytesInSegment;
1914 31640672 : bytesleft = bytepos % UsableBytesInSegment;
1915 :
1916 31640672 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1917 : {
1918 : /* fits on first page of segment */
1919 174956 : if (bytesleft == 0)
1920 113814 : seg_offset = 0;
1921 : else
1922 61142 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1923 : }
1924 : else
1925 : {
1926 : /* account for the first page on segment with long header */
1927 31465716 : seg_offset = XLOG_BLCKSZ;
1928 31465716 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1929 :
1930 31465716 : fullpages = bytesleft / UsableBytesInPage;
1931 31465716 : bytesleft = bytesleft % UsableBytesInPage;
1932 :
1933 31465716 : if (bytesleft == 0)
1934 29830 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
1935 : else
1936 31435886 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1937 : }
1938 :
1939 31640672 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1940 :
1941 31640672 : return result;
1942 : }
1943 :
1944 : /*
1945 : * Convert an XLogRecPtr to a "usable byte position".
1946 : */
1947 : static uint64
1948 5176 : XLogRecPtrToBytePos(XLogRecPtr ptr)
1949 : {
1950 : uint64 fullsegs;
1951 : uint32 fullpages;
1952 : uint32 offset;
1953 : uint64 result;
1954 :
1955 5176 : XLByteToSeg(ptr, fullsegs, wal_segment_size);
1956 :
1957 5176 : fullpages = (XLogSegmentOffset(ptr, wal_segment_size)) / XLOG_BLCKSZ;
1958 5176 : offset = ptr % XLOG_BLCKSZ;
1959 :
1960 5176 : if (fullpages == 0)
1961 : {
1962 2038 : result = fullsegs * UsableBytesInSegment;
1963 2038 : if (offset > 0)
1964 : {
1965 : Assert(offset >= SizeOfXLogLongPHD);
1966 538 : result += offset - SizeOfXLogLongPHD;
1967 : }
1968 : }
1969 : else
1970 : {
1971 3138 : result = fullsegs * UsableBytesInSegment +
1972 3138 : (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
1973 3138 : (fullpages - 1) * UsableBytesInPage; /* full pages */
1974 3138 : if (offset > 0)
1975 : {
1976 : Assert(offset >= SizeOfXLogShortPHD);
1977 3118 : result += offset - SizeOfXLogShortPHD;
1978 : }
1979 : }
1980 :
1981 5176 : return result;
1982 : }
1983 :
1984 : /*
1985 : * Initialize XLOG buffers, writing out old buffers if they still contain
1986 : * unwritten data, upto the page containing 'upto'. Or if 'opportunistic' is
1987 : * true, initialize as many pages as we can without having to write out
1988 : * unwritten data. Any new pages are initialized to zeros, with pages headers
1989 : * initialized properly.
1990 : */
1991 : static void
1992 5339422 : AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
1993 : {
1994 5339422 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1995 : int nextidx;
1996 : XLogRecPtr OldPageRqstPtr;
1997 : XLogwrtRqst WriteRqst;
1998 5339422 : XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
1999 : XLogRecPtr NewPageBeginPtr;
2000 : XLogPageHeader NewPage;
2001 5339422 : int npages pg_attribute_unused() = 0;
2002 :
2003 5339422 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2004 :
2005 : /*
2006 : * Now that we have the lock, check if someone initialized the page
2007 : * already.
2008 : */
2009 15092256 : while (upto >= XLogCtl->InitializedUpTo || opportunistic)
2010 : {
2011 9762152 : nextidx = XLogRecPtrToBufIdx(XLogCtl->InitializedUpTo);
2012 :
2013 : /*
2014 : * Get ending-offset of the buffer page we need to replace (this may
2015 : * be zero if the buffer hasn't been used yet). Fall through if it's
2016 : * already written out.
2017 : */
2018 9762152 : OldPageRqstPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]);
2019 9762152 : if (LogwrtResult.Write < OldPageRqstPtr)
2020 : {
2021 : /*
2022 : * Nope, got work to do. If we just want to pre-initialize as much
2023 : * as we can without flushing, give up now.
2024 : */
2025 4718686 : if (opportunistic)
2026 9318 : break;
2027 :
2028 : /* Advance shared memory write request position */
2029 4709368 : SpinLockAcquire(&XLogCtl->info_lck);
2030 4709368 : if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
2031 1282456 : XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
2032 4709368 : SpinLockRelease(&XLogCtl->info_lck);
2033 :
2034 : /*
2035 : * Acquire an up-to-date LogwrtResult value and see if we still
2036 : * need to write it or if someone else already did.
2037 : */
2038 4709368 : RefreshXLogWriteResult(LogwrtResult);
2039 4709368 : if (LogwrtResult.Write < OldPageRqstPtr)
2040 : {
2041 : /*
2042 : * Must acquire write lock. Release WALBufMappingLock first,
2043 : * to make sure that all insertions that we need to wait for
2044 : * can finish (up to this same position). Otherwise we risk
2045 : * deadlock.
2046 : */
2047 4590880 : LWLockRelease(WALBufMappingLock);
2048 :
2049 4590880 : WaitXLogInsertionsToFinish(OldPageRqstPtr);
2050 :
2051 4590880 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2052 :
2053 4590880 : RefreshXLogWriteResult(LogwrtResult);
2054 4590880 : if (LogwrtResult.Write >= OldPageRqstPtr)
2055 : {
2056 : /* OK, someone wrote it already */
2057 565746 : LWLockRelease(WALWriteLock);
2058 : }
2059 : else
2060 : {
2061 : /* Have to write it ourselves */
2062 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2063 4025134 : WriteRqst.Write = OldPageRqstPtr;
2064 4025134 : WriteRqst.Flush = 0;
2065 4025134 : XLogWrite(WriteRqst, tli, false);
2066 4025134 : LWLockRelease(WALWriteLock);
2067 4025134 : pgWalUsage.wal_buffers_full++;
2068 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_DONE();
2069 :
2070 : /*
2071 : * Required for the flush of pending stats WAL data, per
2072 : * update of pgWalUsage.
2073 : */
2074 4025134 : pgstat_report_fixed = true;
2075 : }
2076 : /* Re-acquire WALBufMappingLock and retry */
2077 4590880 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2078 4590880 : continue;
2079 : }
2080 : }
2081 :
2082 : /*
2083 : * Now the next buffer slot is free and we can set it up to be the
2084 : * next output page.
2085 : */
2086 5161954 : NewPageBeginPtr = XLogCtl->InitializedUpTo;
2087 5161954 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2088 :
2089 : Assert(XLogRecPtrToBufIdx(NewPageBeginPtr) == nextidx);
2090 :
2091 5161954 : NewPage = (XLogPageHeader) (XLogCtl->pages + nextidx * (Size) XLOG_BLCKSZ);
2092 :
2093 : /*
2094 : * Mark the xlblock with InvalidXLogRecPtr and issue a write barrier
2095 : * before initializing. Otherwise, the old page may be partially
2096 : * zeroed but look valid.
2097 : */
2098 5161954 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], InvalidXLogRecPtr);
2099 5161954 : pg_write_barrier();
2100 :
2101 : /*
2102 : * Be sure to re-zero the buffer so that bytes beyond what we've
2103 : * written will look like zeroes and not valid XLOG records...
2104 : */
2105 5161954 : MemSet(NewPage, 0, XLOG_BLCKSZ);
2106 :
2107 : /*
2108 : * Fill the new page's header
2109 : */
2110 5161954 : NewPage->xlp_magic = XLOG_PAGE_MAGIC;
2111 :
2112 : /* NewPage->xlp_info = 0; */ /* done by memset */
2113 5161954 : NewPage->xlp_tli = tli;
2114 5161954 : NewPage->xlp_pageaddr = NewPageBeginPtr;
2115 :
2116 : /* NewPage->xlp_rem_len = 0; */ /* done by memset */
2117 :
2118 : /*
2119 : * If online backup is not in progress, mark the header to indicate
2120 : * that WAL records beginning in this page have removable backup
2121 : * blocks. This allows the WAL archiver to know whether it is safe to
2122 : * compress archived WAL data by transforming full-block records into
2123 : * the non-full-block format. It is sufficient to record this at the
2124 : * page level because we force a page switch (in fact a segment
2125 : * switch) when starting a backup, so the flag will be off before any
2126 : * records can be written during the backup. At the end of a backup,
2127 : * the last page will be marked as all unsafe when perhaps only part
2128 : * is unsafe, but at worst the archiver would miss the opportunity to
2129 : * compress a few records.
2130 : */
2131 5161954 : if (Insert->runningBackups == 0)
2132 4890738 : NewPage->xlp_info |= XLP_BKP_REMOVABLE;
2133 :
2134 : /*
2135 : * If first page of an XLOG segment file, make it a long header.
2136 : */
2137 5161954 : if ((XLogSegmentOffset(NewPage->xlp_pageaddr, wal_segment_size)) == 0)
2138 : {
2139 3630 : XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2140 :
2141 3630 : NewLongPage->xlp_sysid = ControlFile->system_identifier;
2142 3630 : NewLongPage->xlp_seg_size = wal_segment_size;
2143 3630 : NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2144 3630 : NewPage->xlp_info |= XLP_LONG_HEADER;
2145 : }
2146 :
2147 : /*
2148 : * Make sure the initialization of the page becomes visible to others
2149 : * before the xlblocks update. GetXLogBuffer() reads xlblocks without
2150 : * holding a lock.
2151 : */
2152 5161954 : pg_write_barrier();
2153 :
2154 5161954 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], NewPageEndPtr);
2155 5161954 : XLogCtl->InitializedUpTo = NewPageEndPtr;
2156 :
2157 5161954 : npages++;
2158 : }
2159 5339422 : LWLockRelease(WALBufMappingLock);
2160 :
2161 : #ifdef WAL_DEBUG
2162 : if (XLOG_DEBUG && npages > 0)
2163 : {
2164 : elog(DEBUG1, "initialized %d pages, up to %X/%08X",
2165 : npages, LSN_FORMAT_ARGS(NewPageEndPtr));
2166 : }
2167 : #endif
2168 5339422 : }
2169 :
2170 : /*
2171 : * Calculate CheckPointSegments based on max_wal_size_mb and
2172 : * checkpoint_completion_target.
2173 : */
2174 : static void
2175 16172 : CalculateCheckpointSegments(void)
2176 : {
2177 : double target;
2178 :
2179 : /*-------
2180 : * Calculate the distance at which to trigger a checkpoint, to avoid
2181 : * exceeding max_wal_size_mb. This is based on two assumptions:
2182 : *
2183 : * a) we keep WAL for only one checkpoint cycle (prior to PG11 we kept
2184 : * WAL for two checkpoint cycles to allow us to recover from the
2185 : * secondary checkpoint if the first checkpoint failed, though we
2186 : * only did this on the primary anyway, not on standby. Keeping just
2187 : * one checkpoint simplifies processing and reduces disk space in
2188 : * many smaller databases.)
2189 : * b) during checkpoint, we consume checkpoint_completion_target *
2190 : * number of segments consumed between checkpoints.
2191 : *-------
2192 : */
2193 16172 : target = (double) ConvertToXSegs(max_wal_size_mb, wal_segment_size) /
2194 16172 : (1.0 + CheckPointCompletionTarget);
2195 :
2196 : /* round down */
2197 16172 : CheckPointSegments = (int) target;
2198 :
2199 16172 : if (CheckPointSegments < 1)
2200 20 : CheckPointSegments = 1;
2201 16172 : }
2202 :
2203 : void
2204 11742 : assign_max_wal_size(int newval, void *extra)
2205 : {
2206 11742 : max_wal_size_mb = newval;
2207 11742 : CalculateCheckpointSegments();
2208 11742 : }
2209 :
2210 : void
2211 2346 : assign_checkpoint_completion_target(double newval, void *extra)
2212 : {
2213 2346 : CheckPointCompletionTarget = newval;
2214 2346 : CalculateCheckpointSegments();
2215 2346 : }
2216 :
2217 : bool
2218 4534 : check_wal_segment_size(int *newval, void **extra, GucSource source)
2219 : {
2220 4534 : if (!IsValidWalSegSize(*newval))
2221 : {
2222 0 : GUC_check_errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.");
2223 0 : return false;
2224 : }
2225 :
2226 4534 : return true;
2227 : }
2228 :
2229 : /*
2230 : * At a checkpoint, how many WAL segments to recycle as preallocated future
2231 : * XLOG segments? Returns the highest segment that should be preallocated.
2232 : */
2233 : static XLogSegNo
2234 3550 : XLOGfileslop(XLogRecPtr lastredoptr)
2235 : {
2236 : XLogSegNo minSegNo;
2237 : XLogSegNo maxSegNo;
2238 : double distance;
2239 : XLogSegNo recycleSegNo;
2240 :
2241 : /*
2242 : * Calculate the segment numbers that min_wal_size_mb and max_wal_size_mb
2243 : * correspond to. Always recycle enough segments to meet the minimum, and
2244 : * remove enough segments to stay below the maximum.
2245 : */
2246 3550 : minSegNo = lastredoptr / wal_segment_size +
2247 3550 : ConvertToXSegs(min_wal_size_mb, wal_segment_size) - 1;
2248 3550 : maxSegNo = lastredoptr / wal_segment_size +
2249 3550 : ConvertToXSegs(max_wal_size_mb, wal_segment_size) - 1;
2250 :
2251 : /*
2252 : * Between those limits, recycle enough segments to get us through to the
2253 : * estimated end of next checkpoint.
2254 : *
2255 : * To estimate where the next checkpoint will finish, assume that the
2256 : * system runs steadily consuming CheckPointDistanceEstimate bytes between
2257 : * every checkpoint.
2258 : */
2259 3550 : distance = (1.0 + CheckPointCompletionTarget) * CheckPointDistanceEstimate;
2260 : /* add 10% for good measure. */
2261 3550 : distance *= 1.10;
2262 :
2263 3550 : recycleSegNo = (XLogSegNo) ceil(((double) lastredoptr + distance) /
2264 : wal_segment_size);
2265 :
2266 3550 : if (recycleSegNo < minSegNo)
2267 2518 : recycleSegNo = minSegNo;
2268 3550 : if (recycleSegNo > maxSegNo)
2269 788 : recycleSegNo = maxSegNo;
2270 :
2271 3550 : return recycleSegNo;
2272 : }
2273 :
2274 : /*
2275 : * Check whether we've consumed enough xlog space that a checkpoint is needed.
2276 : *
2277 : * new_segno indicates a log file that has just been filled up (or read
2278 : * during recovery). We measure the distance from RedoRecPtr to new_segno
2279 : * and see if that exceeds CheckPointSegments.
2280 : *
2281 : * Note: it is caller's responsibility that RedoRecPtr is up-to-date.
2282 : */
2283 : bool
2284 10156 : XLogCheckpointNeeded(XLogSegNo new_segno)
2285 : {
2286 : XLogSegNo old_segno;
2287 :
2288 10156 : XLByteToSeg(RedoRecPtr, old_segno, wal_segment_size);
2289 :
2290 10156 : if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2291 6580 : return true;
2292 3576 : return false;
2293 : }
2294 :
2295 : /*
2296 : * Write and/or fsync the log at least as far as WriteRqst indicates.
2297 : *
2298 : * If flexible == true, we don't have to write as far as WriteRqst, but
2299 : * may stop at any convenient boundary (such as a cache or logfile boundary).
2300 : * This option allows us to avoid uselessly issuing multiple writes when a
2301 : * single one would do.
2302 : *
2303 : * Must be called with WALWriteLock held. WaitXLogInsertionsToFinish(WriteRqst)
2304 : * must be called before grabbing the lock, to make sure the data is ready to
2305 : * write.
2306 : */
2307 : static void
2308 4300196 : XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible)
2309 : {
2310 : bool ispartialpage;
2311 : bool last_iteration;
2312 : bool finishing_seg;
2313 : int curridx;
2314 : int npages;
2315 : int startidx;
2316 : uint32 startoffset;
2317 :
2318 : /* We should always be inside a critical section here */
2319 : Assert(CritSectionCount > 0);
2320 :
2321 : /*
2322 : * Update local LogwrtResult (caller probably did this already, but...)
2323 : */
2324 4300196 : RefreshXLogWriteResult(LogwrtResult);
2325 :
2326 : /*
2327 : * Since successive pages in the xlog cache are consecutively allocated,
2328 : * we can usually gather multiple pages together and issue just one
2329 : * write() call. npages is the number of pages we have determined can be
2330 : * written together; startidx is the cache block index of the first one,
2331 : * and startoffset is the file offset at which it should go. The latter
2332 : * two variables are only valid when npages > 0, but we must initialize
2333 : * all of them to keep the compiler quiet.
2334 : */
2335 4300196 : npages = 0;
2336 4300196 : startidx = 0;
2337 4300196 : startoffset = 0;
2338 :
2339 : /*
2340 : * Within the loop, curridx is the cache block index of the page to
2341 : * consider writing. Begin at the buffer containing the next unwritten
2342 : * page, or last partially written page.
2343 : */
2344 4300196 : curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2345 :
2346 9333506 : while (LogwrtResult.Write < WriteRqst.Write)
2347 : {
2348 : /*
2349 : * Make sure we're not ahead of the insert process. This could happen
2350 : * if we're passed a bogus WriteRqst.Write that is past the end of the
2351 : * last page that's been initialized by AdvanceXLInsertBuffer.
2352 : */
2353 5299064 : XLogRecPtr EndPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[curridx]);
2354 :
2355 5299064 : if (LogwrtResult.Write >= EndPtr)
2356 0 : elog(PANIC, "xlog write request %X/%08X is past end of log %X/%08X",
2357 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2358 : LSN_FORMAT_ARGS(EndPtr));
2359 :
2360 : /* Advance LogwrtResult.Write to end of current buffer page */
2361 5299064 : LogwrtResult.Write = EndPtr;
2362 5299064 : ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2363 :
2364 5299064 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2365 : wal_segment_size))
2366 : {
2367 : /*
2368 : * Switch to new logfile segment. We cannot have any pending
2369 : * pages here (since we dump what we have at segment end).
2370 : */
2371 : Assert(npages == 0);
2372 27278 : if (openLogFile >= 0)
2373 12180 : XLogFileClose();
2374 27278 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2375 : wal_segment_size);
2376 27278 : openLogTLI = tli;
2377 :
2378 : /* create/use new log file */
2379 27278 : openLogFile = XLogFileInit(openLogSegNo, tli);
2380 27278 : ReserveExternalFD();
2381 : }
2382 :
2383 : /* Make sure we have the current logfile open */
2384 5299064 : if (openLogFile < 0)
2385 : {
2386 0 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2387 : wal_segment_size);
2388 0 : openLogTLI = tli;
2389 0 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2390 0 : ReserveExternalFD();
2391 : }
2392 :
2393 : /* Add current page to the set of pending pages-to-dump */
2394 5299064 : if (npages == 0)
2395 : {
2396 : /* first of group */
2397 4331158 : startidx = curridx;
2398 4331158 : startoffset = XLogSegmentOffset(LogwrtResult.Write - XLOG_BLCKSZ,
2399 : wal_segment_size);
2400 : }
2401 5299064 : npages++;
2402 :
2403 : /*
2404 : * Dump the set if this will be the last loop iteration, or if we are
2405 : * at the last page of the cache area (since the next page won't be
2406 : * contiguous in memory), or if we are at the end of the logfile
2407 : * segment.
2408 : */
2409 5299064 : last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2410 :
2411 10339734 : finishing_seg = !ispartialpage &&
2412 5040670 : (startoffset + npages * XLOG_BLCKSZ) >= wal_segment_size;
2413 :
2414 5299064 : if (last_iteration ||
2415 1000706 : curridx == XLogCtl->XLogCacheBlck ||
2416 : finishing_seg)
2417 : {
2418 : char *from;
2419 : Size nbytes;
2420 : Size nleft;
2421 : ssize_t written;
2422 : instr_time start;
2423 :
2424 : /* OK to write the page(s) */
2425 4331158 : from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2426 4331158 : nbytes = npages * (Size) XLOG_BLCKSZ;
2427 4331158 : nleft = nbytes;
2428 : do
2429 : {
2430 4331158 : errno = 0;
2431 :
2432 : /*
2433 : * Measure I/O timing to write WAL data, for pg_stat_io.
2434 : */
2435 4331158 : start = pgstat_prepare_io_time(track_wal_io_timing);
2436 :
2437 4331158 : pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE);
2438 4331158 : written = pg_pwrite(openLogFile, from, nleft, startoffset);
2439 4331158 : pgstat_report_wait_end();
2440 :
2441 4331158 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL,
2442 : IOOP_WRITE, start, 1, written);
2443 :
2444 4331158 : if (written <= 0)
2445 : {
2446 : char xlogfname[MAXFNAMELEN];
2447 : int save_errno;
2448 :
2449 0 : if (errno == EINTR)
2450 0 : continue;
2451 :
2452 0 : save_errno = errno;
2453 0 : XLogFileName(xlogfname, tli, openLogSegNo,
2454 : wal_segment_size);
2455 0 : errno = save_errno;
2456 0 : ereport(PANIC,
2457 : (errcode_for_file_access(),
2458 : errmsg("could not write to log file \"%s\" at offset %u, length %zu: %m",
2459 : xlogfname, startoffset, nleft)));
2460 : }
2461 4331158 : nleft -= written;
2462 4331158 : from += written;
2463 4331158 : startoffset += written;
2464 4331158 : } while (nleft > 0);
2465 :
2466 4331158 : npages = 0;
2467 :
2468 : /*
2469 : * If we just wrote the whole last page of a logfile segment,
2470 : * fsync the segment immediately. This avoids having to go back
2471 : * and re-open prior segments when an fsync request comes along
2472 : * later. Doing it here ensures that one and only one backend will
2473 : * perform this fsync.
2474 : *
2475 : * This is also the right place to notify the Archiver that the
2476 : * segment is ready to copy to archival storage, and to update the
2477 : * timer for archive_timeout, and to signal for a checkpoint if
2478 : * too many logfile segments have been used since the last
2479 : * checkpoint.
2480 : */
2481 4331158 : if (finishing_seg)
2482 : {
2483 3836 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2484 :
2485 : /* signal that we need to wakeup walsenders later */
2486 3836 : WalSndWakeupRequest();
2487 :
2488 3836 : LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2489 :
2490 3836 : if (XLogArchivingActive())
2491 812 : XLogArchiveNotifySeg(openLogSegNo, tli);
2492 :
2493 3836 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2494 3836 : XLogCtl->lastSegSwitchLSN = LogwrtResult.Flush;
2495 :
2496 : /*
2497 : * Request a checkpoint if we've consumed too much xlog since
2498 : * the last one. For speed, we first check using the local
2499 : * copy of RedoRecPtr, which might be out of date; if it looks
2500 : * like a checkpoint is needed, forcibly update RedoRecPtr and
2501 : * recheck.
2502 : */
2503 3836 : if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2504 : {
2505 508 : (void) GetRedoRecPtr();
2506 508 : if (XLogCheckpointNeeded(openLogSegNo))
2507 406 : RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2508 : }
2509 : }
2510 : }
2511 :
2512 5299064 : if (ispartialpage)
2513 : {
2514 : /* Only asked to write a partial page */
2515 258394 : LogwrtResult.Write = WriteRqst.Write;
2516 258394 : break;
2517 : }
2518 5040670 : curridx = NextBufIdx(curridx);
2519 :
2520 : /* If flexible, break out of loop as soon as we wrote something */
2521 5040670 : if (flexible && npages == 0)
2522 7360 : break;
2523 : }
2524 :
2525 : Assert(npages == 0);
2526 :
2527 : /*
2528 : * If asked to flush, do so
2529 : */
2530 4300196 : if (LogwrtResult.Flush < WriteRqst.Flush &&
2531 273570 : LogwrtResult.Flush < LogwrtResult.Write)
2532 : {
2533 : /*
2534 : * Could get here without iterating above loop, in which case we might
2535 : * have no open file or the wrong one. However, we do not need to
2536 : * fsync more than one file.
2537 : */
2538 273428 : if (wal_sync_method != WAL_SYNC_METHOD_OPEN &&
2539 273428 : wal_sync_method != WAL_SYNC_METHOD_OPEN_DSYNC)
2540 : {
2541 273428 : if (openLogFile >= 0 &&
2542 273418 : !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2543 : wal_segment_size))
2544 136 : XLogFileClose();
2545 273428 : if (openLogFile < 0)
2546 : {
2547 146 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2548 : wal_segment_size);
2549 146 : openLogTLI = tli;
2550 146 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2551 146 : ReserveExternalFD();
2552 : }
2553 :
2554 273428 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2555 : }
2556 :
2557 : /* signal that we need to wakeup walsenders later */
2558 273428 : WalSndWakeupRequest();
2559 :
2560 273428 : LogwrtResult.Flush = LogwrtResult.Write;
2561 : }
2562 :
2563 : /*
2564 : * Update shared-memory status
2565 : *
2566 : * We make sure that the shared 'request' values do not fall behind the
2567 : * 'result' values. This is not absolutely essential, but it saves some
2568 : * code in a couple of places.
2569 : */
2570 4300196 : SpinLockAcquire(&XLogCtl->info_lck);
2571 4300196 : if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
2572 238928 : XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
2573 4300196 : if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
2574 276484 : XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
2575 4300196 : SpinLockRelease(&XLogCtl->info_lck);
2576 :
2577 : /*
2578 : * We write Write first, bar, then Flush. When reading, the opposite must
2579 : * be done (with a matching barrier in between), so that we always see a
2580 : * Flush value that trails behind the Write value seen.
2581 : */
2582 4300196 : pg_atomic_write_u64(&XLogCtl->logWriteResult, LogwrtResult.Write);
2583 4300196 : pg_write_barrier();
2584 4300196 : pg_atomic_write_u64(&XLogCtl->logFlushResult, LogwrtResult.Flush);
2585 :
2586 : #ifdef USE_ASSERT_CHECKING
2587 : {
2588 : XLogRecPtr Flush;
2589 : XLogRecPtr Write;
2590 : XLogRecPtr Insert;
2591 :
2592 : Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult);
2593 : pg_read_barrier();
2594 : Write = pg_atomic_read_u64(&XLogCtl->logWriteResult);
2595 : pg_read_barrier();
2596 : Insert = pg_atomic_read_u64(&XLogCtl->logInsertResult);
2597 :
2598 : /* WAL written to disk is always ahead of WAL flushed */
2599 : Assert(Write >= Flush);
2600 :
2601 : /* WAL inserted to buffers is always ahead of WAL written */
2602 : Assert(Insert >= Write);
2603 : }
2604 : #endif
2605 4300196 : }
2606 :
2607 : /*
2608 : * Record the LSN for an asynchronous transaction commit/abort
2609 : * and nudge the WALWriter if there is work for it to do.
2610 : * (This should not be called for synchronous commits.)
2611 : */
2612 : void
2613 102108 : XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2614 : {
2615 102108 : XLogRecPtr WriteRqstPtr = asyncXactLSN;
2616 : bool sleeping;
2617 102108 : bool wakeup = false;
2618 : XLogRecPtr prevAsyncXactLSN;
2619 :
2620 102108 : SpinLockAcquire(&XLogCtl->info_lck);
2621 102108 : sleeping = XLogCtl->WalWriterSleeping;
2622 102108 : prevAsyncXactLSN = XLogCtl->asyncXactLSN;
2623 102108 : if (XLogCtl->asyncXactLSN < asyncXactLSN)
2624 101118 : XLogCtl->asyncXactLSN = asyncXactLSN;
2625 102108 : SpinLockRelease(&XLogCtl->info_lck);
2626 :
2627 : /*
2628 : * If somebody else already called this function with a more aggressive
2629 : * LSN, they will have done what we needed (and perhaps more).
2630 : */
2631 102108 : if (asyncXactLSN <= prevAsyncXactLSN)
2632 990 : return;
2633 :
2634 : /*
2635 : * If the WALWriter is sleeping, kick it to make it come out of low-power
2636 : * mode, so that this async commit will reach disk within the expected
2637 : * amount of time. Otherwise, determine whether it has enough WAL
2638 : * available to flush, the same way that XLogBackgroundFlush() does.
2639 : */
2640 101118 : if (sleeping)
2641 40 : wakeup = true;
2642 : else
2643 : {
2644 : int flushblocks;
2645 :
2646 101078 : RefreshXLogWriteResult(LogwrtResult);
2647 :
2648 101078 : flushblocks =
2649 101078 : WriteRqstPtr / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
2650 :
2651 101078 : if (WalWriterFlushAfter == 0 || flushblocks >= WalWriterFlushAfter)
2652 7632 : wakeup = true;
2653 : }
2654 :
2655 101118 : if (wakeup)
2656 : {
2657 7672 : volatile PROC_HDR *procglobal = ProcGlobal;
2658 7672 : ProcNumber walwriterProc = procglobal->walwriterProc;
2659 :
2660 7672 : if (walwriterProc != INVALID_PROC_NUMBER)
2661 448 : SetLatch(&GetPGProcByNumber(walwriterProc)->procLatch);
2662 : }
2663 : }
2664 :
2665 : /*
2666 : * Record the LSN up to which we can remove WAL because it's not required by
2667 : * any replication slot.
2668 : */
2669 : void
2670 50072 : XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn)
2671 : {
2672 50072 : SpinLockAcquire(&XLogCtl->info_lck);
2673 50072 : XLogCtl->replicationSlotMinLSN = lsn;
2674 50072 : SpinLockRelease(&XLogCtl->info_lck);
2675 50072 : }
2676 :
2677 :
2678 : /*
2679 : * Return the oldest LSN we must retain to satisfy the needs of some
2680 : * replication slot.
2681 : */
2682 : static XLogRecPtr
2683 4570 : XLogGetReplicationSlotMinimumLSN(void)
2684 : {
2685 : XLogRecPtr retval;
2686 :
2687 4570 : SpinLockAcquire(&XLogCtl->info_lck);
2688 4570 : retval = XLogCtl->replicationSlotMinLSN;
2689 4570 : SpinLockRelease(&XLogCtl->info_lck);
2690 :
2691 4570 : return retval;
2692 : }
2693 :
2694 : /*
2695 : * Advance minRecoveryPoint in control file.
2696 : *
2697 : * If we crash during recovery, we must reach this point again before the
2698 : * database is consistent.
2699 : *
2700 : * If 'force' is true, 'lsn' argument is ignored. Otherwise, minRecoveryPoint
2701 : * is only updated if it's not already greater than or equal to 'lsn'.
2702 : */
2703 : static void
2704 213792 : UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2705 : {
2706 : /* Quick check using our local copy of the variable */
2707 213792 : if (!updateMinRecoveryPoint || (!force && lsn <= LocalMinRecoveryPoint))
2708 200368 : return;
2709 :
2710 : /*
2711 : * An invalid minRecoveryPoint means that we need to recover all the WAL,
2712 : * i.e., we're doing crash recovery. We never modify the control file's
2713 : * value in that case, so we can short-circuit future checks here too. The
2714 : * local values of minRecoveryPoint and minRecoveryPointTLI should not be
2715 : * updated until crash recovery finishes. We only do this for the startup
2716 : * process as it should not update its own reference of minRecoveryPoint
2717 : * until it has finished crash recovery to make sure that all WAL
2718 : * available is replayed in this case. This also saves from extra locks
2719 : * taken on the control file from the startup process.
2720 : */
2721 13424 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint) && InRecovery)
2722 : {
2723 60 : updateMinRecoveryPoint = false;
2724 60 : return;
2725 : }
2726 :
2727 13364 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2728 :
2729 : /* update local copy */
2730 13364 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
2731 13364 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2732 :
2733 13364 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
2734 4 : updateMinRecoveryPoint = false;
2735 13360 : else if (force || LocalMinRecoveryPoint < lsn)
2736 : {
2737 : XLogRecPtr newMinRecoveryPoint;
2738 : TimeLineID newMinRecoveryPointTLI;
2739 :
2740 : /*
2741 : * To avoid having to update the control file too often, we update it
2742 : * all the way to the last record being replayed, even though 'lsn'
2743 : * would suffice for correctness. This also allows the 'force' case
2744 : * to not need a valid 'lsn' value.
2745 : *
2746 : * Another important reason for doing it this way is that the passed
2747 : * 'lsn' value could be bogus, i.e., past the end of available WAL, if
2748 : * the caller got it from a corrupted heap page. Accepting such a
2749 : * value as the min recovery point would prevent us from coming up at
2750 : * all. Instead, we just log a warning and continue with recovery.
2751 : * (See also the comments about corrupt LSNs in XLogFlush.)
2752 : */
2753 10978 : newMinRecoveryPoint = GetCurrentReplayRecPtr(&newMinRecoveryPointTLI);
2754 10978 : if (!force && newMinRecoveryPoint < lsn)
2755 0 : elog(WARNING,
2756 : "xlog min recovery request %X/%08X is past current point %X/%08X",
2757 : LSN_FORMAT_ARGS(lsn), LSN_FORMAT_ARGS(newMinRecoveryPoint));
2758 :
2759 : /* update control file */
2760 10978 : if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2761 : {
2762 10278 : ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2763 10278 : ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2764 10278 : UpdateControlFile();
2765 10278 : LocalMinRecoveryPoint = newMinRecoveryPoint;
2766 10278 : LocalMinRecoveryPointTLI = newMinRecoveryPointTLI;
2767 :
2768 10278 : ereport(DEBUG2,
2769 : errmsg_internal("updated min recovery point to %X/%08X on timeline %u",
2770 : LSN_FORMAT_ARGS(newMinRecoveryPoint),
2771 : newMinRecoveryPointTLI));
2772 : }
2773 : }
2774 13364 : LWLockRelease(ControlFileLock);
2775 : }
2776 :
2777 : /*
2778 : * Ensure that all XLOG data through the given position is flushed to disk.
2779 : *
2780 : * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
2781 : * already held, and we try to avoid acquiring it if possible.
2782 : */
2783 : void
2784 1398466 : XLogFlush(XLogRecPtr record)
2785 : {
2786 : XLogRecPtr WriteRqstPtr;
2787 : XLogwrtRqst WriteRqst;
2788 1398466 : TimeLineID insertTLI = XLogCtl->InsertTimeLineID;
2789 :
2790 : /*
2791 : * During REDO, we are reading not writing WAL. Therefore, instead of
2792 : * trying to flush the WAL, we should update minRecoveryPoint instead. We
2793 : * test XLogInsertAllowed(), not InRecovery, because we need checkpointer
2794 : * to act this way too, and because when it tries to write the
2795 : * end-of-recovery checkpoint, it should indeed flush.
2796 : */
2797 1398466 : if (!XLogInsertAllowed())
2798 : {
2799 212914 : UpdateMinRecoveryPoint(record, false);
2800 1106458 : return;
2801 : }
2802 :
2803 : /* Quick exit if already known flushed */
2804 1185552 : if (record <= LogwrtResult.Flush)
2805 893544 : return;
2806 :
2807 : #ifdef WAL_DEBUG
2808 : if (XLOG_DEBUG)
2809 : elog(LOG, "xlog flush request %X/%08X; write %X/%08X; flush %X/%08X",
2810 : LSN_FORMAT_ARGS(record),
2811 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2812 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2813 : #endif
2814 :
2815 292008 : START_CRIT_SECTION();
2816 :
2817 : /*
2818 : * Since fsync is usually a horribly expensive operation, we try to
2819 : * piggyback as much data as we can on each fsync: if we see any more data
2820 : * entered into the xlog buffer, we'll write and fsync that too, so that
2821 : * the final value of LogwrtResult.Flush is as large as possible. This
2822 : * gives us some chance of avoiding another fsync immediately after.
2823 : */
2824 :
2825 : /* initialize to given target; may increase below */
2826 292008 : WriteRqstPtr = record;
2827 :
2828 : /*
2829 : * Now wait until we get the write lock, or someone else does the flush
2830 : * for us.
2831 : */
2832 : for (;;)
2833 9914 : {
2834 : XLogRecPtr insertpos;
2835 :
2836 : /* done already? */
2837 301922 : RefreshXLogWriteResult(LogwrtResult);
2838 301922 : if (record <= LogwrtResult.Flush)
2839 22034 : break;
2840 :
2841 : /*
2842 : * Before actually performing the write, wait for all in-flight
2843 : * insertions to the pages we're about to write to finish.
2844 : */
2845 279888 : SpinLockAcquire(&XLogCtl->info_lck);
2846 279888 : if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
2847 19568 : WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
2848 279888 : SpinLockRelease(&XLogCtl->info_lck);
2849 279888 : insertpos = WaitXLogInsertionsToFinish(WriteRqstPtr);
2850 :
2851 : /*
2852 : * Try to get the write lock. If we can't get it immediately, wait
2853 : * until it's released, and recheck if we still need to do the flush
2854 : * or if the backend that held the lock did it for us already. This
2855 : * helps to maintain a good rate of group committing when the system
2856 : * is bottlenecked by the speed of fsyncing.
2857 : */
2858 279888 : if (!LWLockAcquireOrWait(WALWriteLock, LW_EXCLUSIVE))
2859 : {
2860 : /*
2861 : * The lock is now free, but we didn't acquire it yet. Before we
2862 : * do, loop back to check if someone else flushed the record for
2863 : * us already.
2864 : */
2865 9914 : continue;
2866 : }
2867 :
2868 : /* Got the lock; recheck whether request is satisfied */
2869 269974 : RefreshXLogWriteResult(LogwrtResult);
2870 269974 : if (record <= LogwrtResult.Flush)
2871 : {
2872 3916 : LWLockRelease(WALWriteLock);
2873 3916 : break;
2874 : }
2875 :
2876 : /*
2877 : * Sleep before flush! By adding a delay here, we may give further
2878 : * backends the opportunity to join the backlog of group commit
2879 : * followers; this can significantly improve transaction throughput,
2880 : * at the risk of increasing transaction latency.
2881 : *
2882 : * We do not sleep if enableFsync is not turned on, nor if there are
2883 : * fewer than CommitSiblings other backends with active transactions.
2884 : */
2885 266058 : if (CommitDelay > 0 && enableFsync &&
2886 0 : MinimumActiveBackends(CommitSiblings))
2887 : {
2888 0 : pgstat_report_wait_start(WAIT_EVENT_COMMIT_DELAY);
2889 0 : pg_usleep(CommitDelay);
2890 0 : pgstat_report_wait_end();
2891 :
2892 : /*
2893 : * Re-check how far we can now flush the WAL. It's generally not
2894 : * safe to call WaitXLogInsertionsToFinish while holding
2895 : * WALWriteLock, because an in-progress insertion might need to
2896 : * also grab WALWriteLock to make progress. But we know that all
2897 : * the insertions up to insertpos have already finished, because
2898 : * that's what the earlier WaitXLogInsertionsToFinish() returned.
2899 : * We're only calling it again to allow insertpos to be moved
2900 : * further forward, not to actually wait for anyone.
2901 : */
2902 0 : insertpos = WaitXLogInsertionsToFinish(insertpos);
2903 : }
2904 :
2905 : /* try to write/flush later additions to XLOG as well */
2906 266058 : WriteRqst.Write = insertpos;
2907 266058 : WriteRqst.Flush = insertpos;
2908 :
2909 266058 : XLogWrite(WriteRqst, insertTLI, false);
2910 :
2911 266058 : LWLockRelease(WALWriteLock);
2912 : /* done */
2913 266058 : break;
2914 : }
2915 :
2916 292008 : END_CRIT_SECTION();
2917 :
2918 : /* wake up walsenders now that we've released heavily contended locks */
2919 292008 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
2920 :
2921 : /*
2922 : * If we flushed an LSN that someone was waiting for, notify the waiters.
2923 : */
2924 584016 : if (waitLSNState &&
2925 292008 : (LogwrtResult.Flush >=
2926 292008 : pg_atomic_read_u64(&waitLSNState->minWaitedLSN[WAIT_LSN_TYPE_PRIMARY_FLUSH])))
2927 0 : WaitLSNWakeup(WAIT_LSN_TYPE_PRIMARY_FLUSH, LogwrtResult.Flush);
2928 :
2929 : /*
2930 : * If we still haven't flushed to the request point then we have a
2931 : * problem; most likely, the requested flush point is past end of XLOG.
2932 : * This has been seen to occur when a disk page has a corrupted LSN.
2933 : *
2934 : * Formerly we treated this as a PANIC condition, but that hurts the
2935 : * system's robustness rather than helping it: we do not want to take down
2936 : * the whole system due to corruption on one data page. In particular, if
2937 : * the bad page is encountered again during recovery then we would be
2938 : * unable to restart the database at all! (This scenario actually
2939 : * happened in the field several times with 7.1 releases.) As of 8.4, bad
2940 : * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
2941 : * the only time we can reach here during recovery is while flushing the
2942 : * end-of-recovery checkpoint record, and we don't expect that to have a
2943 : * bad LSN.
2944 : *
2945 : * Note that for calls from xact.c, the ERROR will be promoted to PANIC
2946 : * since xact.c calls this routine inside a critical section. However,
2947 : * calls from bufmgr.c are not within critical sections and so we will not
2948 : * force a restart for a bad LSN on a data page.
2949 : */
2950 292008 : if (LogwrtResult.Flush < record)
2951 0 : elog(ERROR,
2952 : "xlog flush request %X/%08X is not satisfied --- flushed only to %X/%08X",
2953 : LSN_FORMAT_ARGS(record),
2954 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2955 :
2956 : /*
2957 : * Cross-check XLogNeedsFlush(). Some of the checks of XLogFlush() and
2958 : * XLogNeedsFlush() are duplicated, and this assertion ensures that these
2959 : * remain consistent.
2960 : */
2961 : Assert(!XLogNeedsFlush(record));
2962 : }
2963 :
2964 : /*
2965 : * Write & flush xlog, but without specifying exactly where to.
2966 : *
2967 : * We normally write only completed blocks; but if there is nothing to do on
2968 : * that basis, we check for unwritten async commits in the current incomplete
2969 : * block, and write through the latest one of those. Thus, if async commits
2970 : * are not being used, we will write complete blocks only.
2971 : *
2972 : * If, based on the above, there's anything to write we do so immediately. But
2973 : * to avoid calling fsync, fdatasync et. al. at a rate that'd impact
2974 : * concurrent IO, we only flush WAL every wal_writer_delay ms, or if there's
2975 : * more than wal_writer_flush_after unflushed blocks.
2976 : *
2977 : * We can guarantee that async commits reach disk after at most three
2978 : * wal_writer_delay cycles. (When flushing complete blocks, we allow XLogWrite
2979 : * to write "flexibly", meaning it can stop at the end of the buffer ring;
2980 : * this makes a difference only with very high load or long wal_writer_delay,
2981 : * but imposes one extra cycle for the worst case for async commits.)
2982 : *
2983 : * This routine is invoked periodically by the background walwriter process.
2984 : *
2985 : * Returns true if there was any work to do, even if we skipped flushing due
2986 : * to wal_writer_delay/wal_writer_flush_after.
2987 : */
2988 : bool
2989 29602 : XLogBackgroundFlush(void)
2990 : {
2991 : XLogwrtRqst WriteRqst;
2992 29602 : bool flexible = true;
2993 : static TimestampTz lastflush;
2994 : TimestampTz now;
2995 : int flushblocks;
2996 : TimeLineID insertTLI;
2997 :
2998 : /* XLOG doesn't need flushing during recovery */
2999 29602 : if (RecoveryInProgress())
3000 12 : return false;
3001 :
3002 : /*
3003 : * Since we're not in recovery, InsertTimeLineID is set and can't change,
3004 : * so we can read it without a lock.
3005 : */
3006 29590 : insertTLI = XLogCtl->InsertTimeLineID;
3007 :
3008 : /* read updated LogwrtRqst */
3009 29590 : SpinLockAcquire(&XLogCtl->info_lck);
3010 29590 : WriteRqst = XLogCtl->LogwrtRqst;
3011 29590 : SpinLockRelease(&XLogCtl->info_lck);
3012 :
3013 : /* back off to last completed page boundary */
3014 29590 : WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
3015 :
3016 : /* if we have already flushed that far, consider async commit records */
3017 29590 : RefreshXLogWriteResult(LogwrtResult);
3018 29590 : if (WriteRqst.Write <= LogwrtResult.Flush)
3019 : {
3020 21840 : SpinLockAcquire(&XLogCtl->info_lck);
3021 21840 : WriteRqst.Write = XLogCtl->asyncXactLSN;
3022 21840 : SpinLockRelease(&XLogCtl->info_lck);
3023 21840 : flexible = false; /* ensure it all gets written */
3024 : }
3025 :
3026 : /*
3027 : * If already known flushed, we're done. Just need to check if we are
3028 : * holding an open file handle to a logfile that's no longer in use,
3029 : * preventing the file from being deleted.
3030 : */
3031 29590 : if (WriteRqst.Write <= LogwrtResult.Flush)
3032 : {
3033 20272 : if (openLogFile >= 0)
3034 : {
3035 12658 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
3036 : wal_segment_size))
3037 : {
3038 276 : XLogFileClose();
3039 : }
3040 : }
3041 20272 : return false;
3042 : }
3043 :
3044 : /*
3045 : * Determine how far to flush WAL, based on the wal_writer_delay and
3046 : * wal_writer_flush_after GUCs.
3047 : *
3048 : * Note that XLogSetAsyncXactLSN() performs similar calculation based on
3049 : * wal_writer_flush_after, to decide when to wake us up. Make sure the
3050 : * logic is the same in both places if you change this.
3051 : */
3052 9318 : now = GetCurrentTimestamp();
3053 9318 : flushblocks =
3054 9318 : WriteRqst.Write / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
3055 :
3056 9318 : if (WalWriterFlushAfter == 0 || lastflush == 0)
3057 : {
3058 : /* first call, or block based limits disabled */
3059 534 : WriteRqst.Flush = WriteRqst.Write;
3060 534 : lastflush = now;
3061 : }
3062 8784 : else if (TimestampDifferenceExceeds(lastflush, now, WalWriterDelay))
3063 : {
3064 : /*
3065 : * Flush the writes at least every WalWriterDelay ms. This is
3066 : * important to bound the amount of time it takes for an asynchronous
3067 : * commit to hit disk.
3068 : */
3069 8476 : WriteRqst.Flush = WriteRqst.Write;
3070 8476 : lastflush = now;
3071 : }
3072 308 : else if (flushblocks >= WalWriterFlushAfter)
3073 : {
3074 : /* exceeded wal_writer_flush_after blocks, flush */
3075 282 : WriteRqst.Flush = WriteRqst.Write;
3076 282 : lastflush = now;
3077 : }
3078 : else
3079 : {
3080 : /* no flushing, this time round */
3081 26 : WriteRqst.Flush = 0;
3082 : }
3083 :
3084 : #ifdef WAL_DEBUG
3085 : if (XLOG_DEBUG)
3086 : elog(LOG, "xlog bg flush request write %X/%08X; flush: %X/%08X, current is write %X/%08X; flush %X/%08X",
3087 : LSN_FORMAT_ARGS(WriteRqst.Write),
3088 : LSN_FORMAT_ARGS(WriteRqst.Flush),
3089 : LSN_FORMAT_ARGS(LogwrtResult.Write),
3090 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
3091 : #endif
3092 :
3093 9318 : START_CRIT_SECTION();
3094 :
3095 : /* now wait for any in-progress insertions to finish and get write lock */
3096 9318 : WaitXLogInsertionsToFinish(WriteRqst.Write);
3097 9318 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3098 9318 : RefreshXLogWriteResult(LogwrtResult);
3099 9318 : if (WriteRqst.Write > LogwrtResult.Write ||
3100 444 : WriteRqst.Flush > LogwrtResult.Flush)
3101 : {
3102 9004 : XLogWrite(WriteRqst, insertTLI, flexible);
3103 : }
3104 9318 : LWLockRelease(WALWriteLock);
3105 :
3106 9318 : END_CRIT_SECTION();
3107 :
3108 : /* wake up walsenders now that we've released heavily contended locks */
3109 9318 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3110 :
3111 : /*
3112 : * If we flushed an LSN that someone was waiting for, notify the waiters.
3113 : */
3114 18636 : if (waitLSNState &&
3115 9318 : (LogwrtResult.Flush >=
3116 9318 : pg_atomic_read_u64(&waitLSNState->minWaitedLSN[WAIT_LSN_TYPE_PRIMARY_FLUSH])))
3117 0 : WaitLSNWakeup(WAIT_LSN_TYPE_PRIMARY_FLUSH, LogwrtResult.Flush);
3118 :
3119 : /*
3120 : * Great, done. To take some work off the critical path, try to initialize
3121 : * as many of the no-longer-needed WAL buffers for future use as we can.
3122 : */
3123 9318 : AdvanceXLInsertBuffer(InvalidXLogRecPtr, insertTLI, true);
3124 :
3125 : /*
3126 : * If we determined that we need to write data, but somebody else
3127 : * wrote/flushed already, it should be considered as being active, to
3128 : * avoid hibernating too early.
3129 : */
3130 9318 : return true;
3131 : }
3132 :
3133 : /*
3134 : * Test whether XLOG data has been flushed up to (at least) the given
3135 : * position, or whether the minimum recovery point has been updated past
3136 : * the given position.
3137 : *
3138 : * Returns true if a flush is still needed, or if the minimum recovery point
3139 : * must be updated.
3140 : *
3141 : * It is possible that someone else is already in the process of flushing
3142 : * that far, or has updated the minimum recovery point up to the given
3143 : * position.
3144 : */
3145 : bool
3146 17735378 : XLogNeedsFlush(XLogRecPtr record)
3147 : {
3148 : /*
3149 : * During recovery, we don't flush WAL but update minRecoveryPoint
3150 : * instead. So "needs flush" is taken to mean whether minRecoveryPoint
3151 : * would need to be updated.
3152 : *
3153 : * Using XLogInsertAllowed() rather than RecoveryInProgress() matters for
3154 : * the case of an end-of-recovery checkpoint, where WAL data is flushed.
3155 : * This check should be consistent with the one in XLogFlush().
3156 : */
3157 17735378 : if (!XLogInsertAllowed())
3158 : {
3159 : /* Quick exit if already known to be updated or cannot be updated */
3160 1268958 : if (!updateMinRecoveryPoint || record <= LocalMinRecoveryPoint)
3161 1246204 : return false;
3162 :
3163 : /*
3164 : * An invalid minRecoveryPoint means that we need to recover all the
3165 : * WAL, i.e., we're doing crash recovery. We never modify the control
3166 : * file's value in that case, so we can short-circuit future checks
3167 : * here too. This triggers a quick exit path for the startup process,
3168 : * which cannot update its local copy of minRecoveryPoint as long as
3169 : * it has not replayed all WAL available when doing crash recovery.
3170 : */
3171 22754 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint) && InRecovery)
3172 : {
3173 0 : updateMinRecoveryPoint = false;
3174 0 : return false;
3175 : }
3176 :
3177 : /*
3178 : * Update local copy of minRecoveryPoint. But if the lock is busy,
3179 : * just return a conservative guess.
3180 : */
3181 22754 : if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3182 0 : return true;
3183 22754 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
3184 22754 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3185 22754 : LWLockRelease(ControlFileLock);
3186 :
3187 : /*
3188 : * Check minRecoveryPoint for any other process than the startup
3189 : * process doing crash recovery, which should not update the control
3190 : * file value if crash recovery is still running.
3191 : */
3192 22754 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
3193 0 : updateMinRecoveryPoint = false;
3194 :
3195 : /* check again */
3196 22754 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3197 150 : return false;
3198 : else
3199 22604 : return true;
3200 : }
3201 :
3202 : /* Quick exit if already known flushed */
3203 16466420 : if (record <= LogwrtResult.Flush)
3204 16046744 : return false;
3205 :
3206 : /* read LogwrtResult and update local state */
3207 419676 : RefreshXLogWriteResult(LogwrtResult);
3208 :
3209 : /* check again */
3210 419676 : if (record <= LogwrtResult.Flush)
3211 6906 : return false;
3212 :
3213 412770 : return true;
3214 : }
3215 :
3216 : /*
3217 : * Try to make a given XLOG file segment exist.
3218 : *
3219 : * logsegno: identify segment.
3220 : *
3221 : * *added: on return, true if this call raised the number of extant segments.
3222 : *
3223 : * path: on return, this char[MAXPGPATH] has the path to the logsegno file.
3224 : *
3225 : * Returns -1 or FD of opened file. A -1 here is not an error; a caller
3226 : * wanting an open segment should attempt to open "path", which usually will
3227 : * succeed. (This is weird, but it's efficient for the callers.)
3228 : */
3229 : static int
3230 29634 : XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,
3231 : bool *added, char *path)
3232 : {
3233 : char tmppath[MAXPGPATH];
3234 : XLogSegNo installed_segno;
3235 : XLogSegNo max_segno;
3236 : int fd;
3237 : int save_errno;
3238 29634 : int open_flags = O_RDWR | O_CREAT | O_EXCL | PG_BINARY;
3239 : instr_time io_start;
3240 :
3241 : Assert(logtli != 0);
3242 :
3243 29634 : XLogFilePath(path, logtli, logsegno, wal_segment_size);
3244 :
3245 : /*
3246 : * Try to use existent file (checkpoint maker may have created it already)
3247 : */
3248 29634 : *added = false;
3249 29634 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3250 29634 : get_sync_bit(wal_sync_method));
3251 29634 : if (fd < 0)
3252 : {
3253 2968 : if (errno != ENOENT)
3254 0 : ereport(ERROR,
3255 : (errcode_for_file_access(),
3256 : errmsg("could not open file \"%s\": %m", path)));
3257 : }
3258 : else
3259 26666 : return fd;
3260 :
3261 : /*
3262 : * Initialize an empty (all zeroes) segment. NOTE: it is possible that
3263 : * another process is doing the same thing. If so, we will end up
3264 : * pre-creating an extra log segment. That seems OK, and better than
3265 : * holding the lock throughout this lengthy process.
3266 : */
3267 2968 : elog(DEBUG2, "creating and filling new WAL file");
3268 :
3269 2968 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3270 :
3271 2968 : unlink(tmppath);
3272 :
3273 2968 : if (io_direct_flags & IO_DIRECT_WAL_INIT)
3274 0 : open_flags |= PG_O_DIRECT;
3275 :
3276 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3277 2968 : fd = BasicOpenFile(tmppath, open_flags);
3278 2968 : if (fd < 0)
3279 0 : ereport(ERROR,
3280 : (errcode_for_file_access(),
3281 : errmsg("could not create file \"%s\": %m", tmppath)));
3282 :
3283 : /* Measure I/O timing when initializing segment */
3284 2968 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3285 :
3286 2968 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE);
3287 2968 : save_errno = 0;
3288 2968 : if (wal_init_zero)
3289 : {
3290 : ssize_t rc;
3291 :
3292 : /*
3293 : * Zero-fill the file. With this setting, we do this the hard way to
3294 : * ensure that all the file space has really been allocated. On
3295 : * platforms that allow "holes" in files, just seeking to the end
3296 : * doesn't allocate intermediate space. This way, we know that we
3297 : * have all the space and (after the fsync below) that all the
3298 : * indirect blocks are down on disk. Therefore, fdatasync(2) or
3299 : * O_DSYNC will be sufficient to sync future writes to the log file.
3300 : */
3301 2968 : rc = pg_pwrite_zeros(fd, wal_segment_size, 0);
3302 :
3303 2968 : if (rc < 0)
3304 0 : save_errno = errno;
3305 : }
3306 : else
3307 : {
3308 : /*
3309 : * Otherwise, seeking to the end and writing a solitary byte is
3310 : * enough.
3311 : */
3312 0 : errno = 0;
3313 0 : if (pg_pwrite(fd, "\0", 1, wal_segment_size - 1) != 1)
3314 : {
3315 : /* if write didn't set errno, assume no disk space */
3316 0 : save_errno = errno ? errno : ENOSPC;
3317 : }
3318 : }
3319 2968 : pgstat_report_wait_end();
3320 :
3321 : /*
3322 : * A full segment worth of data is written when using wal_init_zero. One
3323 : * byte is written when not using it.
3324 : */
3325 2968 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT, IOOP_WRITE,
3326 : io_start, 1,
3327 2968 : wal_init_zero ? wal_segment_size : 1);
3328 :
3329 2968 : if (save_errno)
3330 : {
3331 : /*
3332 : * If we fail to make the file, delete it to release disk space
3333 : */
3334 0 : unlink(tmppath);
3335 :
3336 0 : close(fd);
3337 :
3338 0 : errno = save_errno;
3339 :
3340 0 : ereport(ERROR,
3341 : (errcode_for_file_access(),
3342 : errmsg("could not write to file \"%s\": %m", tmppath)));
3343 : }
3344 :
3345 : /* Measure I/O timing when flushing segment */
3346 2968 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3347 :
3348 2968 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC);
3349 2968 : if (pg_fsync(fd) != 0)
3350 : {
3351 0 : save_errno = errno;
3352 0 : close(fd);
3353 0 : errno = save_errno;
3354 0 : ereport(ERROR,
3355 : (errcode_for_file_access(),
3356 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3357 : }
3358 2968 : pgstat_report_wait_end();
3359 :
3360 2968 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT,
3361 : IOOP_FSYNC, io_start, 1, 0);
3362 :
3363 2968 : if (close(fd) != 0)
3364 0 : ereport(ERROR,
3365 : (errcode_for_file_access(),
3366 : errmsg("could not close file \"%s\": %m", tmppath)));
3367 :
3368 : /*
3369 : * Now move the segment into place with its final name. Cope with
3370 : * possibility that someone else has created the file while we were
3371 : * filling ours: if so, use ours to pre-create a future log segment.
3372 : */
3373 2968 : installed_segno = logsegno;
3374 :
3375 : /*
3376 : * XXX: What should we use as max_segno? We used to use XLOGfileslop when
3377 : * that was a constant, but that was always a bit dubious: normally, at a
3378 : * checkpoint, XLOGfileslop was the offset from the checkpoint record, but
3379 : * here, it was the offset from the insert location. We can't do the
3380 : * normal XLOGfileslop calculation here because we don't have access to
3381 : * the prior checkpoint's redo location. So somewhat arbitrarily, just use
3382 : * CheckPointSegments.
3383 : */
3384 2968 : max_segno = logsegno + CheckPointSegments;
3385 2968 : if (InstallXLogFileSegment(&installed_segno, tmppath, true, max_segno,
3386 : logtli))
3387 : {
3388 2968 : *added = true;
3389 2968 : elog(DEBUG2, "done creating and filling new WAL file");
3390 : }
3391 : else
3392 : {
3393 : /*
3394 : * No need for any more future segments, or InstallXLogFileSegment()
3395 : * failed to rename the file into place. If the rename failed, a
3396 : * caller opening the file may fail.
3397 : */
3398 0 : unlink(tmppath);
3399 0 : elog(DEBUG2, "abandoned new WAL file");
3400 : }
3401 :
3402 2968 : return -1;
3403 : }
3404 :
3405 : /*
3406 : * Create a new XLOG file segment, or open a pre-existing one.
3407 : *
3408 : * logsegno: identify segment to be created/opened.
3409 : *
3410 : * Returns FD of opened file.
3411 : *
3412 : * Note: errors here are ERROR not PANIC because we might or might not be
3413 : * inside a critical section (eg, during checkpoint there is no reason to
3414 : * take down the system on failure). They will promote to PANIC if we are
3415 : * in a critical section.
3416 : */
3417 : int
3418 29110 : XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)
3419 : {
3420 : bool ignore_added;
3421 : char path[MAXPGPATH];
3422 : int fd;
3423 :
3424 : Assert(logtli != 0);
3425 :
3426 29110 : fd = XLogFileInitInternal(logsegno, logtli, &ignore_added, path);
3427 29110 : if (fd >= 0)
3428 26430 : return fd;
3429 :
3430 : /* Now open original target segment (might not be file I just made) */
3431 2680 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3432 2680 : get_sync_bit(wal_sync_method));
3433 2680 : if (fd < 0)
3434 0 : ereport(ERROR,
3435 : (errcode_for_file_access(),
3436 : errmsg("could not open file \"%s\": %m", path)));
3437 2680 : return fd;
3438 : }
3439 :
3440 : /*
3441 : * Create a new XLOG file segment by copying a pre-existing one.
3442 : *
3443 : * destsegno: identify segment to be created.
3444 : *
3445 : * srcTLI, srcsegno: identify segment to be copied (could be from
3446 : * a different timeline)
3447 : *
3448 : * upto: how much of the source file to copy (the rest is filled with
3449 : * zeros)
3450 : *
3451 : * Currently this is only used during recovery, and so there are no locking
3452 : * considerations. But we should be just as tense as XLogFileInit to avoid
3453 : * emplacing a bogus file.
3454 : */
3455 : static void
3456 90 : XLogFileCopy(TimeLineID destTLI, XLogSegNo destsegno,
3457 : TimeLineID srcTLI, XLogSegNo srcsegno,
3458 : int upto)
3459 : {
3460 : char path[MAXPGPATH];
3461 : char tmppath[MAXPGPATH];
3462 : PGAlignedXLogBlock buffer;
3463 : int srcfd;
3464 : int fd;
3465 : int nbytes;
3466 :
3467 : /*
3468 : * Open the source file
3469 : */
3470 90 : XLogFilePath(path, srcTLI, srcsegno, wal_segment_size);
3471 90 : srcfd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
3472 90 : if (srcfd < 0)
3473 0 : ereport(ERROR,
3474 : (errcode_for_file_access(),
3475 : errmsg("could not open file \"%s\": %m", path)));
3476 :
3477 : /*
3478 : * Copy into a temp file name.
3479 : */
3480 90 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3481 :
3482 90 : unlink(tmppath);
3483 :
3484 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3485 90 : fd = OpenTransientFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
3486 90 : if (fd < 0)
3487 0 : ereport(ERROR,
3488 : (errcode_for_file_access(),
3489 : errmsg("could not create file \"%s\": %m", tmppath)));
3490 :
3491 : /*
3492 : * Do the data copying.
3493 : */
3494 184410 : for (nbytes = 0; nbytes < wal_segment_size; nbytes += sizeof(buffer))
3495 : {
3496 : int nread;
3497 :
3498 184320 : nread = upto - nbytes;
3499 :
3500 : /*
3501 : * The part that is not read from the source file is filled with
3502 : * zeros.
3503 : */
3504 184320 : if (nread < sizeof(buffer))
3505 90 : memset(buffer.data, 0, sizeof(buffer));
3506 :
3507 184320 : if (nread > 0)
3508 : {
3509 : int r;
3510 :
3511 5452 : if (nread > sizeof(buffer))
3512 5362 : nread = sizeof(buffer);
3513 5452 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ);
3514 5452 : r = read(srcfd, buffer.data, nread);
3515 5452 : if (r != nread)
3516 : {
3517 0 : if (r < 0)
3518 0 : ereport(ERROR,
3519 : (errcode_for_file_access(),
3520 : errmsg("could not read file \"%s\": %m",
3521 : path)));
3522 : else
3523 0 : ereport(ERROR,
3524 : (errcode(ERRCODE_DATA_CORRUPTED),
3525 : errmsg("could not read file \"%s\": read %d of %zu",
3526 : path, r, (Size) nread)));
3527 : }
3528 5452 : pgstat_report_wait_end();
3529 : }
3530 184320 : errno = 0;
3531 184320 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_WRITE);
3532 184320 : if ((int) write(fd, buffer.data, sizeof(buffer)) != (int) sizeof(buffer))
3533 : {
3534 0 : int save_errno = errno;
3535 :
3536 : /*
3537 : * If we fail to make the file, delete it to release disk space
3538 : */
3539 0 : unlink(tmppath);
3540 : /* if write didn't set errno, assume problem is no disk space */
3541 0 : errno = save_errno ? save_errno : ENOSPC;
3542 :
3543 0 : ereport(ERROR,
3544 : (errcode_for_file_access(),
3545 : errmsg("could not write to file \"%s\": %m", tmppath)));
3546 : }
3547 184320 : pgstat_report_wait_end();
3548 : }
3549 :
3550 90 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_SYNC);
3551 90 : if (pg_fsync(fd) != 0)
3552 0 : ereport(data_sync_elevel(ERROR),
3553 : (errcode_for_file_access(),
3554 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3555 90 : pgstat_report_wait_end();
3556 :
3557 90 : if (CloseTransientFile(fd) != 0)
3558 0 : ereport(ERROR,
3559 : (errcode_for_file_access(),
3560 : errmsg("could not close file \"%s\": %m", tmppath)));
3561 :
3562 90 : if (CloseTransientFile(srcfd) != 0)
3563 0 : ereport(ERROR,
3564 : (errcode_for_file_access(),
3565 : errmsg("could not close file \"%s\": %m", path)));
3566 :
3567 : /*
3568 : * Now move the segment into place with its final name.
3569 : */
3570 90 : if (!InstallXLogFileSegment(&destsegno, tmppath, false, 0, destTLI))
3571 0 : elog(ERROR, "InstallXLogFileSegment should not have failed");
3572 90 : }
3573 :
3574 : /*
3575 : * Install a new XLOG segment file as a current or future log segment.
3576 : *
3577 : * This is used both to install a newly-created segment (which has a temp
3578 : * filename while it's being created) and to recycle an old segment.
3579 : *
3580 : * *segno: identify segment to install as (or first possible target).
3581 : * When find_free is true, this is modified on return to indicate the
3582 : * actual installation location or last segment searched.
3583 : *
3584 : * tmppath: initial name of file to install. It will be renamed into place.
3585 : *
3586 : * find_free: if true, install the new segment at the first empty segno
3587 : * number at or after the passed numbers. If false, install the new segment
3588 : * exactly where specified, deleting any existing segment file there.
3589 : *
3590 : * max_segno: maximum segment number to install the new file as. Fail if no
3591 : * free slot is found between *segno and max_segno. (Ignored when find_free
3592 : * is false.)
3593 : *
3594 : * tli: The timeline on which the new segment should be installed.
3595 : *
3596 : * Returns true if the file was installed successfully. false indicates that
3597 : * max_segno limit was exceeded, the startup process has disabled this
3598 : * function for now, or an error occurred while renaming the file into place.
3599 : */
3600 : static bool
3601 5814 : InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
3602 : bool find_free, XLogSegNo max_segno, TimeLineID tli)
3603 : {
3604 : char path[MAXPGPATH];
3605 : struct stat stat_buf;
3606 :
3607 : Assert(tli != 0);
3608 :
3609 5814 : XLogFilePath(path, tli, *segno, wal_segment_size);
3610 :
3611 5814 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3612 5814 : if (!XLogCtl->InstallXLogFileSegmentActive)
3613 : {
3614 0 : LWLockRelease(ControlFileLock);
3615 0 : return false;
3616 : }
3617 :
3618 5814 : if (!find_free)
3619 : {
3620 : /* Force installation: get rid of any pre-existing segment file */
3621 90 : durable_unlink(path, DEBUG1);
3622 : }
3623 : else
3624 : {
3625 : /* Find a free slot to put it in */
3626 8806 : while (stat(path, &stat_buf) == 0)
3627 : {
3628 3128 : if ((*segno) >= max_segno)
3629 : {
3630 : /* Failed to find a free slot within specified range */
3631 46 : LWLockRelease(ControlFileLock);
3632 46 : return false;
3633 : }
3634 3082 : (*segno)++;
3635 3082 : XLogFilePath(path, tli, *segno, wal_segment_size);
3636 : }
3637 : }
3638 :
3639 : Assert(access(path, F_OK) != 0 && errno == ENOENT);
3640 5768 : if (durable_rename(tmppath, path, LOG) != 0)
3641 : {
3642 0 : LWLockRelease(ControlFileLock);
3643 : /* durable_rename already emitted log message */
3644 0 : return false;
3645 : }
3646 :
3647 5768 : LWLockRelease(ControlFileLock);
3648 :
3649 5768 : return true;
3650 : }
3651 :
3652 : /*
3653 : * Open a pre-existing logfile segment for writing.
3654 : */
3655 : int
3656 146 : XLogFileOpen(XLogSegNo segno, TimeLineID tli)
3657 : {
3658 : char path[MAXPGPATH];
3659 : int fd;
3660 :
3661 146 : XLogFilePath(path, tli, segno, wal_segment_size);
3662 :
3663 146 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3664 146 : get_sync_bit(wal_sync_method));
3665 146 : if (fd < 0)
3666 0 : ereport(PANIC,
3667 : (errcode_for_file_access(),
3668 : errmsg("could not open file \"%s\": %m", path)));
3669 :
3670 146 : return fd;
3671 : }
3672 :
3673 : /*
3674 : * Close the current logfile segment for writing.
3675 : */
3676 : static void
3677 12592 : XLogFileClose(void)
3678 : {
3679 : Assert(openLogFile >= 0);
3680 :
3681 : /*
3682 : * WAL segment files will not be re-read in normal operation, so we advise
3683 : * the OS to release any cached pages. But do not do so if WAL archiving
3684 : * or streaming is active, because archiver and walsender process could
3685 : * use the cache to read the WAL segment.
3686 : */
3687 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3688 12592 : if (!XLogIsNeeded() && (io_direct_flags & IO_DIRECT_WAL) == 0)
3689 224 : (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3690 : #endif
3691 :
3692 12592 : if (close(openLogFile) != 0)
3693 : {
3694 : char xlogfname[MAXFNAMELEN];
3695 0 : int save_errno = errno;
3696 :
3697 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo, wal_segment_size);
3698 0 : errno = save_errno;
3699 0 : ereport(PANIC,
3700 : (errcode_for_file_access(),
3701 : errmsg("could not close file \"%s\": %m", xlogfname)));
3702 : }
3703 :
3704 12592 : openLogFile = -1;
3705 12592 : ReleaseExternalFD();
3706 12592 : }
3707 :
3708 : /*
3709 : * Preallocate log files beyond the specified log endpoint.
3710 : *
3711 : * XXX this is currently extremely conservative, since it forces only one
3712 : * future log segment to exist, and even that only if we are 75% done with
3713 : * the current one. This is only appropriate for very low-WAL-volume systems.
3714 : * High-volume systems will be OK once they've built up a sufficient set of
3715 : * recycled log segments, but the startup transient is likely to include
3716 : * a lot of segment creations by foreground processes, which is not so good.
3717 : *
3718 : * XLogFileInitInternal() can ereport(ERROR). All known causes indicate big
3719 : * trouble; for example, a full filesystem is one cause. The checkpoint WAL
3720 : * and/or ControlFile updates already completed. If a RequestCheckpoint()
3721 : * initiated the present checkpoint and an ERROR ends this function, the
3722 : * command that called RequestCheckpoint() fails. That's not ideal, but it's
3723 : * not worth contorting more functions to use caller-specified elevel values.
3724 : * (With or without RequestCheckpoint(), an ERROR forestalls some inessential
3725 : * reporting and resource reclamation.)
3726 : */
3727 : static void
3728 4100 : PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli)
3729 : {
3730 : XLogSegNo _logSegNo;
3731 : int lf;
3732 : bool added;
3733 : char path[MAXPGPATH];
3734 : uint64 offset;
3735 :
3736 4100 : if (!XLogCtl->InstallXLogFileSegmentActive)
3737 20 : return; /* unlocked check says no */
3738 :
3739 4080 : XLByteToPrevSeg(endptr, _logSegNo, wal_segment_size);
3740 4080 : offset = XLogSegmentOffset(endptr - 1, wal_segment_size);
3741 4080 : if (offset >= (uint32) (0.75 * wal_segment_size))
3742 : {
3743 524 : _logSegNo++;
3744 524 : lf = XLogFileInitInternal(_logSegNo, tli, &added, path);
3745 524 : if (lf >= 0)
3746 236 : close(lf);
3747 524 : if (added)
3748 288 : CheckpointStats.ckpt_segs_added++;
3749 : }
3750 : }
3751 :
3752 : /*
3753 : * Throws an error if the given log segment has already been removed or
3754 : * recycled. The caller should only pass a segment that it knows to have
3755 : * existed while the server has been running, as this function always
3756 : * succeeds if no WAL segments have been removed since startup.
3757 : * 'tli' is only used in the error message.
3758 : *
3759 : * Note: this function guarantees to keep errno unchanged on return.
3760 : * This supports callers that use this to possibly deliver a better
3761 : * error message about a missing file, while still being able to throw
3762 : * a normal file-access error afterwards, if this does return.
3763 : */
3764 : void
3765 229990 : CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3766 : {
3767 229990 : int save_errno = errno;
3768 : XLogSegNo lastRemovedSegNo;
3769 :
3770 229990 : SpinLockAcquire(&XLogCtl->info_lck);
3771 229990 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3772 229990 : SpinLockRelease(&XLogCtl->info_lck);
3773 :
3774 229990 : if (segno <= lastRemovedSegNo)
3775 : {
3776 : char filename[MAXFNAMELEN];
3777 :
3778 0 : XLogFileName(filename, tli, segno, wal_segment_size);
3779 0 : errno = save_errno;
3780 0 : ereport(ERROR,
3781 : (errcode_for_file_access(),
3782 : errmsg("requested WAL segment %s has already been removed",
3783 : filename)));
3784 : }
3785 229990 : errno = save_errno;
3786 229990 : }
3787 :
3788 : /*
3789 : * Return the last WAL segment removed, or 0 if no segment has been removed
3790 : * since startup.
3791 : *
3792 : * NB: the result can be out of date arbitrarily fast, the caller has to deal
3793 : * with that.
3794 : */
3795 : XLogSegNo
3796 2260 : XLogGetLastRemovedSegno(void)
3797 : {
3798 : XLogSegNo lastRemovedSegNo;
3799 :
3800 2260 : SpinLockAcquire(&XLogCtl->info_lck);
3801 2260 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3802 2260 : SpinLockRelease(&XLogCtl->info_lck);
3803 :
3804 2260 : return lastRemovedSegNo;
3805 : }
3806 :
3807 : /*
3808 : * Return the oldest WAL segment on the given TLI that still exists in
3809 : * XLOGDIR, or 0 if none.
3810 : */
3811 : XLogSegNo
3812 12 : XLogGetOldestSegno(TimeLineID tli)
3813 : {
3814 : DIR *xldir;
3815 : struct dirent *xlde;
3816 12 : XLogSegNo oldest_segno = 0;
3817 :
3818 12 : xldir = AllocateDir(XLOGDIR);
3819 82 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3820 : {
3821 : TimeLineID file_tli;
3822 : XLogSegNo file_segno;
3823 :
3824 : /* Ignore files that are not XLOG segments. */
3825 70 : if (!IsXLogFileName(xlde->d_name))
3826 48 : continue;
3827 :
3828 : /* Parse filename to get TLI and segno. */
3829 22 : XLogFromFileName(xlde->d_name, &file_tli, &file_segno,
3830 : wal_segment_size);
3831 :
3832 : /* Ignore anything that's not from the TLI of interest. */
3833 22 : if (tli != file_tli)
3834 0 : continue;
3835 :
3836 : /* If it's the oldest so far, update oldest_segno. */
3837 22 : if (oldest_segno == 0 || file_segno < oldest_segno)
3838 20 : oldest_segno = file_segno;
3839 : }
3840 :
3841 12 : FreeDir(xldir);
3842 12 : return oldest_segno;
3843 : }
3844 :
3845 : /*
3846 : * Update the last removed segno pointer in shared memory, to reflect that the
3847 : * given XLOG file has been removed.
3848 : */
3849 : static void
3850 5232 : UpdateLastRemovedPtr(char *filename)
3851 : {
3852 : uint32 tli;
3853 : XLogSegNo segno;
3854 :
3855 5232 : XLogFromFileName(filename, &tli, &segno, wal_segment_size);
3856 :
3857 5232 : SpinLockAcquire(&XLogCtl->info_lck);
3858 5232 : if (segno > XLogCtl->lastRemovedSegNo)
3859 2308 : XLogCtl->lastRemovedSegNo = segno;
3860 5232 : SpinLockRelease(&XLogCtl->info_lck);
3861 5232 : }
3862 :
3863 : /*
3864 : * Remove all temporary log files in pg_wal
3865 : *
3866 : * This is called at the beginning of recovery after a previous crash,
3867 : * at a point where no other processes write fresh WAL data.
3868 : */
3869 : static void
3870 368 : RemoveTempXlogFiles(void)
3871 : {
3872 : DIR *xldir;
3873 : struct dirent *xlde;
3874 :
3875 368 : elog(DEBUG2, "removing all temporary WAL segments");
3876 :
3877 368 : xldir = AllocateDir(XLOGDIR);
3878 2492 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3879 : {
3880 : char path[MAXPGPATH];
3881 :
3882 2124 : if (strncmp(xlde->d_name, "xlogtemp.", 9) != 0)
3883 2124 : continue;
3884 :
3885 0 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
3886 0 : unlink(path);
3887 0 : elog(DEBUG2, "removed temporary WAL segment \"%s\"", path);
3888 : }
3889 368 : FreeDir(xldir);
3890 368 : }
3891 :
3892 : /*
3893 : * Recycle or remove all log files older or equal to passed segno.
3894 : *
3895 : * endptr is current (or recent) end of xlog, and lastredoptr is the
3896 : * redo pointer of the last checkpoint. These are used to determine
3897 : * whether we want to recycle rather than delete no-longer-wanted log files.
3898 : *
3899 : * insertTLI is the current timeline for XLOG insertion. Any recycled
3900 : * segments should be reused for this timeline.
3901 : */
3902 : static void
3903 3550 : RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr, XLogRecPtr endptr,
3904 : TimeLineID insertTLI)
3905 : {
3906 : DIR *xldir;
3907 : struct dirent *xlde;
3908 : char lastoff[MAXFNAMELEN];
3909 : XLogSegNo endlogSegNo;
3910 : XLogSegNo recycleSegNo;
3911 :
3912 : /* Initialize info about where to try to recycle to */
3913 3550 : XLByteToSeg(endptr, endlogSegNo, wal_segment_size);
3914 3550 : recycleSegNo = XLOGfileslop(lastredoptr);
3915 :
3916 : /*
3917 : * Construct a filename of the last segment to be kept. The timeline ID
3918 : * doesn't matter, we ignore that in the comparison. (During recovery,
3919 : * InsertTimeLineID isn't set, so we can't use that.)
3920 : */
3921 3550 : XLogFileName(lastoff, 0, segno, wal_segment_size);
3922 :
3923 3550 : elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
3924 : lastoff);
3925 :
3926 3550 : xldir = AllocateDir(XLOGDIR);
3927 :
3928 57428 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3929 : {
3930 : /* Ignore files that are not XLOG segments */
3931 53878 : if (!IsXLogFileName(xlde->d_name) &&
3932 15318 : !IsPartialXLogFileName(xlde->d_name))
3933 15310 : continue;
3934 :
3935 : /*
3936 : * We ignore the timeline part of the XLOG segment identifiers in
3937 : * deciding whether a segment is still needed. This ensures that we
3938 : * won't prematurely remove a segment from a parent timeline. We could
3939 : * probably be a little more proactive about removing segments of
3940 : * non-parent timelines, but that would be a whole lot more
3941 : * complicated.
3942 : *
3943 : * We use the alphanumeric sorting property of the filenames to decide
3944 : * which ones are earlier than the lastoff segment.
3945 : */
3946 38568 : if (strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
3947 : {
3948 10350 : if (XLogArchiveCheckDone(xlde->d_name))
3949 : {
3950 : /* Update the last removed location in shared memory first */
3951 5232 : UpdateLastRemovedPtr(xlde->d_name);
3952 :
3953 5232 : RemoveXlogFile(xlde, recycleSegNo, &endlogSegNo, insertTLI);
3954 : }
3955 : }
3956 : }
3957 :
3958 3550 : FreeDir(xldir);
3959 3550 : }
3960 :
3961 : /*
3962 : * Recycle or remove WAL files that are not part of the given timeline's
3963 : * history.
3964 : *
3965 : * This is called during recovery, whenever we switch to follow a new
3966 : * timeline, and at the end of recovery when we create a new timeline. We
3967 : * wouldn't otherwise care about extra WAL files lying in pg_wal, but they
3968 : * might be leftover pre-allocated or recycled WAL segments on the old timeline
3969 : * that we haven't used yet, and contain garbage. If we just leave them in
3970 : * pg_wal, they will eventually be archived, and we can't let that happen.
3971 : * Files that belong to our timeline history are valid, because we have
3972 : * successfully replayed them, but from others we can't be sure.
3973 : *
3974 : * 'switchpoint' is the current point in WAL where we switch to new timeline,
3975 : * and 'newTLI' is the new timeline we switch to.
3976 : */
3977 : void
3978 134 : RemoveNonParentXlogFiles(XLogRecPtr switchpoint, TimeLineID newTLI)
3979 : {
3980 : DIR *xldir;
3981 : struct dirent *xlde;
3982 : char switchseg[MAXFNAMELEN];
3983 : XLogSegNo endLogSegNo;
3984 : XLogSegNo switchLogSegNo;
3985 : XLogSegNo recycleSegNo;
3986 :
3987 : /*
3988 : * Initialize info about where to begin the work. This will recycle,
3989 : * somewhat arbitrarily, 10 future segments.
3990 : */
3991 134 : XLByteToPrevSeg(switchpoint, switchLogSegNo, wal_segment_size);
3992 134 : XLByteToSeg(switchpoint, endLogSegNo, wal_segment_size);
3993 134 : recycleSegNo = endLogSegNo + 10;
3994 :
3995 : /*
3996 : * Construct a filename of the last segment to be kept.
3997 : */
3998 134 : XLogFileName(switchseg, newTLI, switchLogSegNo, wal_segment_size);
3999 :
4000 134 : elog(DEBUG2, "attempting to remove WAL segments newer than log file %s",
4001 : switchseg);
4002 :
4003 134 : xldir = AllocateDir(XLOGDIR);
4004 :
4005 1276 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4006 : {
4007 : /* Ignore files that are not XLOG segments */
4008 1142 : if (!IsXLogFileName(xlde->d_name))
4009 706 : continue;
4010 :
4011 : /*
4012 : * Remove files that are on a timeline older than the new one we're
4013 : * switching to, but with a segment number >= the first segment on the
4014 : * new timeline.
4015 : */
4016 436 : if (strncmp(xlde->d_name, switchseg, 8) < 0 &&
4017 284 : strcmp(xlde->d_name + 8, switchseg + 8) > 0)
4018 : {
4019 : /*
4020 : * If the file has already been marked as .ready, however, don't
4021 : * remove it yet. It should be OK to remove it - files that are
4022 : * not part of our timeline history are not required for recovery
4023 : * - but seems safer to let them be archived and removed later.
4024 : */
4025 34 : if (!XLogArchiveIsReady(xlde->d_name))
4026 34 : RemoveXlogFile(xlde, recycleSegNo, &endLogSegNo, newTLI);
4027 : }
4028 : }
4029 :
4030 134 : FreeDir(xldir);
4031 134 : }
4032 :
4033 : /*
4034 : * Recycle or remove a log file that's no longer needed.
4035 : *
4036 : * segment_de is the dirent structure of the segment to recycle or remove.
4037 : * recycleSegNo is the segment number to recycle up to. endlogSegNo is
4038 : * the segment number of the current (or recent) end of WAL.
4039 : *
4040 : * endlogSegNo gets incremented if the segment is recycled so as it is not
4041 : * checked again with future callers of this function.
4042 : *
4043 : * insertTLI is the current timeline for XLOG insertion. Any recycled segments
4044 : * should be used for this timeline.
4045 : */
4046 : static void
4047 5266 : RemoveXlogFile(const struct dirent *segment_de,
4048 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
4049 : TimeLineID insertTLI)
4050 : {
4051 : char path[MAXPGPATH];
4052 : #ifdef WIN32
4053 : char newpath[MAXPGPATH];
4054 : #endif
4055 5266 : const char *segname = segment_de->d_name;
4056 :
4057 5266 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", segname);
4058 :
4059 : /*
4060 : * Before deleting the file, see if it can be recycled as a future log
4061 : * segment. Only recycle normal files, because we don't want to recycle
4062 : * symbolic links pointing to a separate archive directory.
4063 : */
4064 5266 : if (wal_recycle &&
4065 5266 : *endlogSegNo <= recycleSegNo &&
4066 6168 : XLogCtl->InstallXLogFileSegmentActive && /* callee rechecks this */
4067 5512 : get_dirent_type(path, segment_de, false, DEBUG2) == PGFILETYPE_REG &&
4068 2756 : InstallXLogFileSegment(endlogSegNo, path,
4069 : true, recycleSegNo, insertTLI))
4070 : {
4071 2710 : ereport(DEBUG2,
4072 : (errmsg_internal("recycled write-ahead log file \"%s\"",
4073 : segname)));
4074 2710 : CheckpointStats.ckpt_segs_recycled++;
4075 : /* Needn't recheck that slot on future iterations */
4076 2710 : (*endlogSegNo)++;
4077 : }
4078 : else
4079 : {
4080 : /* No need for any more future segments, or recycling failed ... */
4081 : int rc;
4082 :
4083 2556 : ereport(DEBUG2,
4084 : (errmsg_internal("removing write-ahead log file \"%s\"",
4085 : segname)));
4086 :
4087 : #ifdef WIN32
4088 :
4089 : /*
4090 : * On Windows, if another process (e.g another backend) holds the file
4091 : * open in FILE_SHARE_DELETE mode, unlink will succeed, but the file
4092 : * will still show up in directory listing until the last handle is
4093 : * closed. To avoid confusing the lingering deleted file for a live
4094 : * WAL file that needs to be archived, rename it before deleting it.
4095 : *
4096 : * If another process holds the file open without FILE_SHARE_DELETE
4097 : * flag, rename will fail. We'll try again at the next checkpoint.
4098 : */
4099 : snprintf(newpath, MAXPGPATH, "%s.deleted", path);
4100 : if (rename(path, newpath) != 0)
4101 : {
4102 : ereport(LOG,
4103 : (errcode_for_file_access(),
4104 : errmsg("could not rename file \"%s\": %m",
4105 : path)));
4106 : return;
4107 : }
4108 : rc = durable_unlink(newpath, LOG);
4109 : #else
4110 2556 : rc = durable_unlink(path, LOG);
4111 : #endif
4112 2556 : if (rc != 0)
4113 : {
4114 : /* Message already logged by durable_unlink() */
4115 0 : return;
4116 : }
4117 2556 : CheckpointStats.ckpt_segs_removed++;
4118 : }
4119 :
4120 5266 : XLogArchiveCleanup(segname);
4121 : }
4122 :
4123 : /*
4124 : * Verify whether pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
4125 : * If the latter do not exist, recreate them.
4126 : *
4127 : * It is not the goal of this function to verify the contents of these
4128 : * directories, but to help in cases where someone has performed a cluster
4129 : * copy for PITR purposes but omitted pg_wal from the copy.
4130 : *
4131 : * We could also recreate pg_wal if it doesn't exist, but a deliberate
4132 : * policy decision was made not to. It is fairly common for pg_wal to be
4133 : * a symlink, and if that was the DBA's intent then automatically making a
4134 : * plain directory would result in degraded performance with no notice.
4135 : */
4136 : static void
4137 1982 : ValidateXLOGDirectoryStructure(void)
4138 : {
4139 : char path[MAXPGPATH];
4140 : struct stat stat_buf;
4141 :
4142 : /* Check for pg_wal; if it doesn't exist, error out */
4143 1982 : if (stat(XLOGDIR, &stat_buf) != 0 ||
4144 1982 : !S_ISDIR(stat_buf.st_mode))
4145 0 : ereport(FATAL,
4146 : (errcode_for_file_access(),
4147 : errmsg("required WAL directory \"%s\" does not exist",
4148 : XLOGDIR)));
4149 :
4150 : /* Check for archive_status */
4151 1982 : snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4152 1982 : if (stat(path, &stat_buf) == 0)
4153 : {
4154 : /* Check for weird cases where it exists but isn't a directory */
4155 1980 : if (!S_ISDIR(stat_buf.st_mode))
4156 0 : ereport(FATAL,
4157 : (errcode_for_file_access(),
4158 : errmsg("required WAL directory \"%s\" does not exist",
4159 : path)));
4160 : }
4161 : else
4162 : {
4163 2 : ereport(LOG,
4164 : (errmsg("creating missing WAL directory \"%s\"", path)));
4165 2 : if (MakePGDirectory(path) < 0)
4166 0 : ereport(FATAL,
4167 : (errcode_for_file_access(),
4168 : errmsg("could not create missing directory \"%s\": %m",
4169 : path)));
4170 : }
4171 :
4172 : /* Check for summaries */
4173 1982 : snprintf(path, MAXPGPATH, XLOGDIR "/summaries");
4174 1982 : if (stat(path, &stat_buf) == 0)
4175 : {
4176 : /* Check for weird cases where it exists but isn't a directory */
4177 1980 : if (!S_ISDIR(stat_buf.st_mode))
4178 0 : ereport(FATAL,
4179 : (errmsg("required WAL directory \"%s\" does not exist",
4180 : path)));
4181 : }
4182 : else
4183 : {
4184 2 : ereport(LOG,
4185 : (errmsg("creating missing WAL directory \"%s\"", path)));
4186 2 : if (MakePGDirectory(path) < 0)
4187 0 : ereport(FATAL,
4188 : (errmsg("could not create missing directory \"%s\": %m",
4189 : path)));
4190 : }
4191 1982 : }
4192 :
4193 : /*
4194 : * Remove previous backup history files. This also retries creation of
4195 : * .ready files for any backup history files for which XLogArchiveNotify
4196 : * failed earlier.
4197 : */
4198 : static void
4199 302 : CleanupBackupHistory(void)
4200 : {
4201 : DIR *xldir;
4202 : struct dirent *xlde;
4203 : char path[MAXPGPATH + sizeof(XLOGDIR)];
4204 :
4205 302 : xldir = AllocateDir(XLOGDIR);
4206 :
4207 3038 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4208 : {
4209 2434 : if (IsBackupHistoryFileName(xlde->d_name))
4210 : {
4211 318 : if (XLogArchiveCheckDone(xlde->d_name))
4212 : {
4213 252 : elog(DEBUG2, "removing WAL backup history file \"%s\"",
4214 : xlde->d_name);
4215 252 : snprintf(path, sizeof(path), XLOGDIR "/%s", xlde->d_name);
4216 252 : unlink(path);
4217 252 : XLogArchiveCleanup(xlde->d_name);
4218 : }
4219 : }
4220 : }
4221 :
4222 302 : FreeDir(xldir);
4223 302 : }
4224 :
4225 : /*
4226 : * I/O routines for pg_control
4227 : *
4228 : * *ControlFile is a buffer in shared memory that holds an image of the
4229 : * contents of pg_control. WriteControlFile() initializes pg_control
4230 : * given a preloaded buffer, ReadControlFile() loads the buffer from
4231 : * the pg_control file (during postmaster or standalone-backend startup),
4232 : * and UpdateControlFile() rewrites pg_control after we modify xlog state.
4233 : * InitControlFile() fills the buffer with initial values.
4234 : *
4235 : * For simplicity, WriteControlFile() initializes the fields of pg_control
4236 : * that are related to checking backend/database compatibility, and
4237 : * ReadControlFile() verifies they are correct. We could split out the
4238 : * I/O and compatibility-check functions, but there seems no need currently.
4239 : */
4240 :
4241 : static void
4242 102 : InitControlFile(uint64 sysidentifier, uint32 data_checksum_version)
4243 : {
4244 : char mock_auth_nonce[MOCK_AUTH_NONCE_LEN];
4245 :
4246 : /*
4247 : * Generate a random nonce. This is used for authentication requests that
4248 : * will fail because the user does not exist. The nonce is used to create
4249 : * a genuine-looking password challenge for the non-existent user, in lieu
4250 : * of an actual stored password.
4251 : */
4252 102 : if (!pg_strong_random(mock_auth_nonce, MOCK_AUTH_NONCE_LEN))
4253 0 : ereport(PANIC,
4254 : (errcode(ERRCODE_INTERNAL_ERROR),
4255 : errmsg("could not generate secret authorization token")));
4256 :
4257 102 : memset(ControlFile, 0, sizeof(ControlFileData));
4258 : /* Initialize pg_control status fields */
4259 102 : ControlFile->system_identifier = sysidentifier;
4260 102 : memcpy(ControlFile->mock_authentication_nonce, mock_auth_nonce, MOCK_AUTH_NONCE_LEN);
4261 102 : ControlFile->state = DB_SHUTDOWNED;
4262 102 : ControlFile->unloggedLSN = FirstNormalUnloggedLSN;
4263 :
4264 : /* Set important parameter values for use when replaying WAL */
4265 102 : ControlFile->MaxConnections = MaxConnections;
4266 102 : ControlFile->max_worker_processes = max_worker_processes;
4267 102 : ControlFile->max_wal_senders = max_wal_senders;
4268 102 : ControlFile->max_prepared_xacts = max_prepared_xacts;
4269 102 : ControlFile->max_locks_per_xact = max_locks_per_xact;
4270 102 : ControlFile->wal_level = wal_level;
4271 102 : ControlFile->wal_log_hints = wal_log_hints;
4272 102 : ControlFile->track_commit_timestamp = track_commit_timestamp;
4273 102 : ControlFile->data_checksum_version = data_checksum_version;
4274 102 : }
4275 :
4276 : static void
4277 102 : WriteControlFile(void)
4278 : {
4279 : int fd;
4280 : char buffer[PG_CONTROL_FILE_SIZE]; /* need not be aligned */
4281 :
4282 : /*
4283 : * Initialize version and compatibility-check fields
4284 : */
4285 102 : ControlFile->pg_control_version = PG_CONTROL_VERSION;
4286 102 : ControlFile->catalog_version_no = CATALOG_VERSION_NO;
4287 :
4288 102 : ControlFile->maxAlign = MAXIMUM_ALIGNOF;
4289 102 : ControlFile->floatFormat = FLOATFORMAT_VALUE;
4290 :
4291 102 : ControlFile->blcksz = BLCKSZ;
4292 102 : ControlFile->relseg_size = RELSEG_SIZE;
4293 102 : ControlFile->slru_pages_per_segment = SLRU_PAGES_PER_SEGMENT;
4294 102 : ControlFile->xlog_blcksz = XLOG_BLCKSZ;
4295 102 : ControlFile->xlog_seg_size = wal_segment_size;
4296 :
4297 102 : ControlFile->nameDataLen = NAMEDATALEN;
4298 102 : ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
4299 :
4300 102 : ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
4301 102 : ControlFile->loblksize = LOBLKSIZE;
4302 :
4303 102 : ControlFile->float8ByVal = true; /* vestigial */
4304 :
4305 : /*
4306 : * Initialize the default 'char' signedness.
4307 : *
4308 : * The signedness of the char type is implementation-defined. For instance
4309 : * on x86 architecture CPUs, the char data type is typically treated as
4310 : * signed by default, whereas on aarch architecture CPUs, it is typically
4311 : * treated as unsigned by default. In v17 or earlier, we accidentally let
4312 : * C implementation signedness affect persistent data. This led to
4313 : * inconsistent results when comparing char data across different
4314 : * platforms.
4315 : *
4316 : * This flag can be used as a hint to ensure consistent behavior for
4317 : * pre-v18 data files that store data sorted by the 'char' type on disk,
4318 : * especially in cross-platform replication scenarios.
4319 : *
4320 : * Newly created database clusters unconditionally set the default char
4321 : * signedness to true. pg_upgrade changes this flag for clusters that were
4322 : * initialized on signedness=false platforms. As a result,
4323 : * signedness=false setting will become rare over time. If we had known
4324 : * about this problem during the last development cycle that forced initdb
4325 : * (v8.3), we would have made all clusters signed or all clusters
4326 : * unsigned. Making pg_upgrade the only source of signedness=false will
4327 : * cause the population of database clusters to converge toward that
4328 : * retrospective ideal.
4329 : */
4330 102 : ControlFile->default_char_signedness = true;
4331 :
4332 : /* Contents are protected with a CRC */
4333 102 : INIT_CRC32C(ControlFile->crc);
4334 102 : COMP_CRC32C(ControlFile->crc,
4335 : ControlFile,
4336 : offsetof(ControlFileData, crc));
4337 102 : FIN_CRC32C(ControlFile->crc);
4338 :
4339 : /*
4340 : * We write out PG_CONTROL_FILE_SIZE bytes into pg_control, zero-padding
4341 : * the excess over sizeof(ControlFileData). This reduces the odds of
4342 : * premature-EOF errors when reading pg_control. We'll still fail when we
4343 : * check the contents of the file, but hopefully with a more specific
4344 : * error than "couldn't read pg_control".
4345 : */
4346 102 : memset(buffer, 0, PG_CONTROL_FILE_SIZE);
4347 102 : memcpy(buffer, ControlFile, sizeof(ControlFileData));
4348 :
4349 102 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4350 : O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
4351 102 : if (fd < 0)
4352 0 : ereport(PANIC,
4353 : (errcode_for_file_access(),
4354 : errmsg("could not create file \"%s\": %m",
4355 : XLOG_CONTROL_FILE)));
4356 :
4357 102 : errno = 0;
4358 102 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_WRITE);
4359 102 : if (write(fd, buffer, PG_CONTROL_FILE_SIZE) != PG_CONTROL_FILE_SIZE)
4360 : {
4361 : /* if write didn't set errno, assume problem is no disk space */
4362 0 : if (errno == 0)
4363 0 : errno = ENOSPC;
4364 0 : ereport(PANIC,
4365 : (errcode_for_file_access(),
4366 : errmsg("could not write to file \"%s\": %m",
4367 : XLOG_CONTROL_FILE)));
4368 : }
4369 102 : pgstat_report_wait_end();
4370 :
4371 102 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_SYNC);
4372 102 : if (pg_fsync(fd) != 0)
4373 0 : ereport(PANIC,
4374 : (errcode_for_file_access(),
4375 : errmsg("could not fsync file \"%s\": %m",
4376 : XLOG_CONTROL_FILE)));
4377 102 : pgstat_report_wait_end();
4378 :
4379 102 : if (close(fd) != 0)
4380 0 : ereport(PANIC,
4381 : (errcode_for_file_access(),
4382 : errmsg("could not close file \"%s\": %m",
4383 : XLOG_CONTROL_FILE)));
4384 102 : }
4385 :
4386 : static void
4387 2084 : ReadControlFile(void)
4388 : {
4389 : pg_crc32c crc;
4390 : int fd;
4391 : char wal_segsz_str[20];
4392 : int r;
4393 :
4394 : /*
4395 : * Read data...
4396 : */
4397 2084 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4398 : O_RDWR | PG_BINARY);
4399 2084 : if (fd < 0)
4400 0 : ereport(PANIC,
4401 : (errcode_for_file_access(),
4402 : errmsg("could not open file \"%s\": %m",
4403 : XLOG_CONTROL_FILE)));
4404 :
4405 2084 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ);
4406 2084 : r = read(fd, ControlFile, sizeof(ControlFileData));
4407 2084 : if (r != sizeof(ControlFileData))
4408 : {
4409 0 : if (r < 0)
4410 0 : ereport(PANIC,
4411 : (errcode_for_file_access(),
4412 : errmsg("could not read file \"%s\": %m",
4413 : XLOG_CONTROL_FILE)));
4414 : else
4415 0 : ereport(PANIC,
4416 : (errcode(ERRCODE_DATA_CORRUPTED),
4417 : errmsg("could not read file \"%s\": read %d of %zu",
4418 : XLOG_CONTROL_FILE, r, sizeof(ControlFileData))));
4419 : }
4420 2084 : pgstat_report_wait_end();
4421 :
4422 2084 : close(fd);
4423 :
4424 : /*
4425 : * Check for expected pg_control format version. If this is wrong, the
4426 : * CRC check will likely fail because we'll be checking the wrong number
4427 : * of bytes. Complaining about wrong version will probably be more
4428 : * enlightening than complaining about wrong CRC.
4429 : */
4430 :
4431 2084 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
4432 0 : ereport(FATAL,
4433 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4434 : errmsg("database files are incompatible with server"),
4435 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4436 : " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4437 : ControlFile->pg_control_version, ControlFile->pg_control_version,
4438 : PG_CONTROL_VERSION, PG_CONTROL_VERSION),
4439 : errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4440 :
4441 2084 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
4442 0 : ereport(FATAL,
4443 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4444 : errmsg("database files are incompatible with server"),
4445 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4446 : " but the server was compiled with PG_CONTROL_VERSION %d.",
4447 : ControlFile->pg_control_version, PG_CONTROL_VERSION),
4448 : errhint("It looks like you need to initdb.")));
4449 :
4450 : /* Now check the CRC. */
4451 2084 : INIT_CRC32C(crc);
4452 2084 : COMP_CRC32C(crc,
4453 : ControlFile,
4454 : offsetof(ControlFileData, crc));
4455 2084 : FIN_CRC32C(crc);
4456 :
4457 2084 : if (!EQ_CRC32C(crc, ControlFile->crc))
4458 0 : ereport(FATAL,
4459 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4460 : errmsg("incorrect checksum in control file")));
4461 :
4462 : /*
4463 : * Do compatibility checking immediately. If the database isn't
4464 : * compatible with the backend executable, we want to abort before we can
4465 : * possibly do any damage.
4466 : */
4467 2084 : if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
4468 0 : ereport(FATAL,
4469 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4470 : errmsg("database files are incompatible with server"),
4471 : /* translator: %s is a variable name and %d is its value */
4472 : errdetail("The database cluster was initialized with %s %d,"
4473 : " but the server was compiled with %s %d.",
4474 : "CATALOG_VERSION_NO", ControlFile->catalog_version_no,
4475 : "CATALOG_VERSION_NO", CATALOG_VERSION_NO),
4476 : errhint("It looks like you need to initdb.")));
4477 2084 : if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
4478 0 : ereport(FATAL,
4479 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4480 : errmsg("database files are incompatible with server"),
4481 : /* translator: %s is a variable name and %d is its value */
4482 : errdetail("The database cluster was initialized with %s %d,"
4483 : " but the server was compiled with %s %d.",
4484 : "MAXALIGN", ControlFile->maxAlign,
4485 : "MAXALIGN", MAXIMUM_ALIGNOF),
4486 : errhint("It looks like you need to initdb.")));
4487 2084 : if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
4488 0 : ereport(FATAL,
4489 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4490 : errmsg("database files are incompatible with server"),
4491 : errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4492 : errhint("It looks like you need to initdb.")));
4493 2084 : if (ControlFile->blcksz != BLCKSZ)
4494 0 : ereport(FATAL,
4495 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4496 : errmsg("database files are incompatible with server"),
4497 : /* translator: %s is a variable name and %d is its value */
4498 : errdetail("The database cluster was initialized with %s %d,"
4499 : " but the server was compiled with %s %d.",
4500 : "BLCKSZ", ControlFile->blcksz,
4501 : "BLCKSZ", BLCKSZ),
4502 : errhint("It looks like you need to recompile or initdb.")));
4503 2084 : if (ControlFile->relseg_size != RELSEG_SIZE)
4504 0 : ereport(FATAL,
4505 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4506 : errmsg("database files are incompatible with server"),
4507 : /* translator: %s is a variable name and %d is its value */
4508 : errdetail("The database cluster was initialized with %s %d,"
4509 : " but the server was compiled with %s %d.",
4510 : "RELSEG_SIZE", ControlFile->relseg_size,
4511 : "RELSEG_SIZE", RELSEG_SIZE),
4512 : errhint("It looks like you need to recompile or initdb.")));
4513 2084 : if (ControlFile->slru_pages_per_segment != SLRU_PAGES_PER_SEGMENT)
4514 0 : ereport(FATAL,
4515 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4516 : errmsg("database files are incompatible with server"),
4517 : /* translator: %s is a variable name and %d is its value */
4518 : errdetail("The database cluster was initialized with %s %d,"
4519 : " but the server was compiled with %s %d.",
4520 : "SLRU_PAGES_PER_SEGMENT", ControlFile->slru_pages_per_segment,
4521 : "SLRU_PAGES_PER_SEGMENT", SLRU_PAGES_PER_SEGMENT),
4522 : errhint("It looks like you need to recompile or initdb.")));
4523 2084 : if (ControlFile->xlog_blcksz != XLOG_BLCKSZ)
4524 0 : ereport(FATAL,
4525 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4526 : errmsg("database files are incompatible with server"),
4527 : /* translator: %s is a variable name and %d is its value */
4528 : errdetail("The database cluster was initialized with %s %d,"
4529 : " but the server was compiled with %s %d.",
4530 : "XLOG_BLCKSZ", ControlFile->xlog_blcksz,
4531 : "XLOG_BLCKSZ", XLOG_BLCKSZ),
4532 : errhint("It looks like you need to recompile or initdb.")));
4533 2084 : if (ControlFile->nameDataLen != NAMEDATALEN)
4534 0 : ereport(FATAL,
4535 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4536 : errmsg("database files are incompatible with server"),
4537 : /* translator: %s is a variable name and %d is its value */
4538 : errdetail("The database cluster was initialized with %s %d,"
4539 : " but the server was compiled with %s %d.",
4540 : "NAMEDATALEN", ControlFile->nameDataLen,
4541 : "NAMEDATALEN", NAMEDATALEN),
4542 : errhint("It looks like you need to recompile or initdb.")));
4543 2084 : if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
4544 0 : ereport(FATAL,
4545 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4546 : errmsg("database files are incompatible with server"),
4547 : /* translator: %s is a variable name and %d is its value */
4548 : errdetail("The database cluster was initialized with %s %d,"
4549 : " but the server was compiled with %s %d.",
4550 : "INDEX_MAX_KEYS", ControlFile->indexMaxKeys,
4551 : "INDEX_MAX_KEYS", INDEX_MAX_KEYS),
4552 : errhint("It looks like you need to recompile or initdb.")));
4553 2084 : if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
4554 0 : ereport(FATAL,
4555 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4556 : errmsg("database files are incompatible with server"),
4557 : /* translator: %s is a variable name and %d is its value */
4558 : errdetail("The database cluster was initialized with %s %d,"
4559 : " but the server was compiled with %s %d.",
4560 : "TOAST_MAX_CHUNK_SIZE", ControlFile->toast_max_chunk_size,
4561 : "TOAST_MAX_CHUNK_SIZE", (int) TOAST_MAX_CHUNK_SIZE),
4562 : errhint("It looks like you need to recompile or initdb.")));
4563 2084 : if (ControlFile->loblksize != LOBLKSIZE)
4564 0 : ereport(FATAL,
4565 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4566 : errmsg("database files are incompatible with server"),
4567 : /* translator: %s is a variable name and %d is its value */
4568 : errdetail("The database cluster was initialized with %s %d,"
4569 : " but the server was compiled with %s %d.",
4570 : "LOBLKSIZE", ControlFile->loblksize,
4571 : "LOBLKSIZE", (int) LOBLKSIZE),
4572 : errhint("It looks like you need to recompile or initdb.")));
4573 :
4574 : Assert(ControlFile->float8ByVal); /* vestigial, not worth an error msg */
4575 :
4576 2084 : wal_segment_size = ControlFile->xlog_seg_size;
4577 :
4578 2084 : if (!IsValidWalSegSize(wal_segment_size))
4579 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4580 : errmsg_plural("invalid WAL segment size in control file (%d byte)",
4581 : "invalid WAL segment size in control file (%d bytes)",
4582 : wal_segment_size,
4583 : wal_segment_size),
4584 : errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.")));
4585 :
4586 2084 : snprintf(wal_segsz_str, sizeof(wal_segsz_str), "%d", wal_segment_size);
4587 2084 : SetConfigOption("wal_segment_size", wal_segsz_str, PGC_INTERNAL,
4588 : PGC_S_DYNAMIC_DEFAULT);
4589 :
4590 : /* check and update variables dependent on wal_segment_size */
4591 2084 : if (ConvertToXSegs(min_wal_size_mb, wal_segment_size) < 2)
4592 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4593 : /* translator: both %s are GUC names */
4594 : errmsg("\"%s\" must be at least twice \"%s\"",
4595 : "min_wal_size", "wal_segment_size")));
4596 :
4597 2084 : if (ConvertToXSegs(max_wal_size_mb, wal_segment_size) < 2)
4598 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4599 : /* translator: both %s are GUC names */
4600 : errmsg("\"%s\" must be at least twice \"%s\"",
4601 : "max_wal_size", "wal_segment_size")));
4602 :
4603 2084 : UsableBytesInSegment =
4604 2084 : (wal_segment_size / XLOG_BLCKSZ * UsableBytesInPage) -
4605 : (SizeOfXLogLongPHD - SizeOfXLogShortPHD);
4606 :
4607 2084 : CalculateCheckpointSegments();
4608 :
4609 : /* Make the initdb settings visible as GUC variables, too */
4610 2084 : SetConfigOption("data_checksums", DataChecksumsEnabled() ? "yes" : "no",
4611 : PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
4612 2084 : }
4613 :
4614 : /*
4615 : * Utility wrapper to update the control file. Note that the control
4616 : * file gets flushed.
4617 : */
4618 : static void
4619 18772 : UpdateControlFile(void)
4620 : {
4621 18772 : update_controlfile(DataDir, ControlFile, true);
4622 18772 : }
4623 :
4624 : /*
4625 : * Returns the unique system identifier from control file.
4626 : */
4627 : uint64
4628 2894 : GetSystemIdentifier(void)
4629 : {
4630 : Assert(ControlFile != NULL);
4631 2894 : return ControlFile->system_identifier;
4632 : }
4633 :
4634 : /*
4635 : * Returns the random nonce from control file.
4636 : */
4637 : char *
4638 2 : GetMockAuthenticationNonce(void)
4639 : {
4640 : Assert(ControlFile != NULL);
4641 2 : return ControlFile->mock_authentication_nonce;
4642 : }
4643 :
4644 : /*
4645 : * Are checksums enabled for data pages?
4646 : */
4647 : bool
4648 19950770 : DataChecksumsEnabled(void)
4649 : {
4650 : Assert(ControlFile != NULL);
4651 19950770 : return (ControlFile->data_checksum_version > 0);
4652 : }
4653 :
4654 : /*
4655 : * Return true if the cluster was initialized on a platform where the
4656 : * default signedness of char is "signed". This function exists for code
4657 : * that deals with pre-v18 data files that store data sorted by the 'char'
4658 : * type on disk (e.g., GIN and GiST indexes). See the comments in
4659 : * WriteControlFile() for details.
4660 : */
4661 : bool
4662 6 : GetDefaultCharSignedness(void)
4663 : {
4664 6 : return ControlFile->default_char_signedness;
4665 : }
4666 :
4667 : /*
4668 : * Returns a fake LSN for unlogged relations.
4669 : *
4670 : * Each call generates an LSN that is greater than any previous value
4671 : * returned. The current counter value is saved and restored across clean
4672 : * shutdowns, but like unlogged relations, does not survive a crash. This can
4673 : * be used in lieu of real LSN values returned by XLogInsert, if you need an
4674 : * LSN-like increasing sequence of numbers without writing any WAL.
4675 : */
4676 : XLogRecPtr
4677 66 : GetFakeLSNForUnloggedRel(void)
4678 : {
4679 66 : return pg_atomic_fetch_add_u64(&XLogCtl->unloggedLSN, 1);
4680 : }
4681 :
4682 : /*
4683 : * Auto-tune the number of XLOG buffers.
4684 : *
4685 : * The preferred setting for wal_buffers is about 3% of shared_buffers, with
4686 : * a maximum of one XLOG segment (there is little reason to think that more
4687 : * is helpful, at least so long as we force an fsync when switching log files)
4688 : * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
4689 : * 9.1, when auto-tuning was added).
4690 : *
4691 : * This should not be called until NBuffers has received its final value.
4692 : */
4693 : static int
4694 2268 : XLOGChooseNumBuffers(void)
4695 : {
4696 : int xbuffers;
4697 :
4698 2268 : xbuffers = NBuffers / 32;
4699 2268 : if (xbuffers > (wal_segment_size / XLOG_BLCKSZ))
4700 48 : xbuffers = (wal_segment_size / XLOG_BLCKSZ);
4701 2268 : if (xbuffers < 8)
4702 882 : xbuffers = 8;
4703 2268 : return xbuffers;
4704 : }
4705 :
4706 : /*
4707 : * GUC check_hook for wal_buffers
4708 : */
4709 : bool
4710 4614 : check_wal_buffers(int *newval, void **extra, GucSource source)
4711 : {
4712 : /*
4713 : * -1 indicates a request for auto-tune.
4714 : */
4715 4614 : if (*newval == -1)
4716 : {
4717 : /*
4718 : * If we haven't yet changed the boot_val default of -1, just let it
4719 : * be. We'll fix it when XLOGShmemSize is called.
4720 : */
4721 2346 : if (XLOGbuffers == -1)
4722 2346 : return true;
4723 :
4724 : /* Otherwise, substitute the auto-tune value */
4725 0 : *newval = XLOGChooseNumBuffers();
4726 : }
4727 :
4728 : /*
4729 : * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
4730 : * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
4731 : * the case, we just silently treat such values as a request for the
4732 : * minimum. (We could throw an error instead, but that doesn't seem very
4733 : * helpful.)
4734 : */
4735 2268 : if (*newval < 4)
4736 0 : *newval = 4;
4737 :
4738 2268 : return true;
4739 : }
4740 :
4741 : /*
4742 : * GUC check_hook for wal_consistency_checking
4743 : */
4744 : bool
4745 4230 : check_wal_consistency_checking(char **newval, void **extra, GucSource source)
4746 : {
4747 : char *rawstring;
4748 : List *elemlist;
4749 : ListCell *l;
4750 : bool newwalconsistency[RM_MAX_ID + 1];
4751 :
4752 : /* Initialize the array */
4753 139590 : MemSet(newwalconsistency, 0, (RM_MAX_ID + 1) * sizeof(bool));
4754 :
4755 : /* Need a modifiable copy of string */
4756 4230 : rawstring = pstrdup(*newval);
4757 :
4758 : /* Parse string into list of identifiers */
4759 4230 : if (!SplitIdentifierString(rawstring, ',', &elemlist))
4760 : {
4761 : /* syntax error in list */
4762 0 : GUC_check_errdetail("List syntax is invalid.");
4763 0 : pfree(rawstring);
4764 0 : list_free(elemlist);
4765 0 : return false;
4766 : }
4767 :
4768 5178 : foreach(l, elemlist)
4769 : {
4770 948 : char *tok = (char *) lfirst(l);
4771 : int rmid;
4772 :
4773 : /* Check for 'all'. */
4774 948 : if (pg_strcasecmp(tok, "all") == 0)
4775 : {
4776 242608 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4777 241664 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL)
4778 9440 : newwalconsistency[rmid] = true;
4779 : }
4780 : else
4781 : {
4782 : /* Check if the token matches any known resource manager. */
4783 4 : bool found = false;
4784 :
4785 72 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4786 : {
4787 108 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL &&
4788 36 : pg_strcasecmp(tok, GetRmgr(rmid).rm_name) == 0)
4789 : {
4790 4 : newwalconsistency[rmid] = true;
4791 4 : found = true;
4792 4 : break;
4793 : }
4794 : }
4795 4 : if (!found)
4796 : {
4797 : /*
4798 : * During startup, it might be a not-yet-loaded custom
4799 : * resource manager. Defer checking until
4800 : * InitializeWalConsistencyChecking().
4801 : */
4802 0 : if (!process_shared_preload_libraries_done)
4803 : {
4804 0 : check_wal_consistency_checking_deferred = true;
4805 : }
4806 : else
4807 : {
4808 0 : GUC_check_errdetail("Unrecognized key word: \"%s\".", tok);
4809 0 : pfree(rawstring);
4810 0 : list_free(elemlist);
4811 0 : return false;
4812 : }
4813 : }
4814 : }
4815 : }
4816 :
4817 4230 : pfree(rawstring);
4818 4230 : list_free(elemlist);
4819 :
4820 : /* assign new value */
4821 4230 : *extra = guc_malloc(LOG, (RM_MAX_ID + 1) * sizeof(bool));
4822 4230 : if (!*extra)
4823 0 : return false;
4824 4230 : memcpy(*extra, newwalconsistency, (RM_MAX_ID + 1) * sizeof(bool));
4825 4230 : return true;
4826 : }
4827 :
4828 : /*
4829 : * GUC assign_hook for wal_consistency_checking
4830 : */
4831 : void
4832 4228 : assign_wal_consistency_checking(const char *newval, void *extra)
4833 : {
4834 : /*
4835 : * If some checks were deferred, it's possible that the checks will fail
4836 : * later during InitializeWalConsistencyChecking(). But in that case, the
4837 : * postmaster will exit anyway, so it's safe to proceed with the
4838 : * assignment.
4839 : *
4840 : * Any built-in resource managers specified are assigned immediately,
4841 : * which affects WAL created before shared_preload_libraries are
4842 : * processed. Any custom resource managers specified won't be assigned
4843 : * until after shared_preload_libraries are processed, but that's OK
4844 : * because WAL for a custom resource manager can't be written before the
4845 : * module is loaded anyway.
4846 : */
4847 4228 : wal_consistency_checking = extra;
4848 4228 : }
4849 :
4850 : /*
4851 : * InitializeWalConsistencyChecking: run after loading custom resource managers
4852 : *
4853 : * If any unknown resource managers were specified in the
4854 : * wal_consistency_checking GUC, processing was deferred. Now that
4855 : * shared_preload_libraries have been loaded, process wal_consistency_checking
4856 : * again.
4857 : */
4858 : void
4859 1962 : InitializeWalConsistencyChecking(void)
4860 : {
4861 : Assert(process_shared_preload_libraries_done);
4862 :
4863 1962 : if (check_wal_consistency_checking_deferred)
4864 : {
4865 : struct config_generic *guc;
4866 :
4867 0 : guc = find_option("wal_consistency_checking", false, false, ERROR);
4868 :
4869 0 : check_wal_consistency_checking_deferred = false;
4870 :
4871 0 : set_config_option_ext("wal_consistency_checking",
4872 : wal_consistency_checking_string,
4873 : guc->scontext, guc->source, guc->srole,
4874 : GUC_ACTION_SET, true, ERROR, false);
4875 :
4876 : /* checking should not be deferred again */
4877 : Assert(!check_wal_consistency_checking_deferred);
4878 : }
4879 1962 : }
4880 :
4881 : /*
4882 : * GUC show_hook for archive_command
4883 : */
4884 : const char *
4885 3452 : show_archive_command(void)
4886 : {
4887 3452 : if (XLogArchivingActive())
4888 4 : return XLogArchiveCommand;
4889 : else
4890 3448 : return "(disabled)";
4891 : }
4892 :
4893 : /*
4894 : * GUC show_hook for in_hot_standby
4895 : */
4896 : const char *
4897 31720 : show_in_hot_standby(void)
4898 : {
4899 : /*
4900 : * We display the actual state based on shared memory, so that this GUC
4901 : * reports up-to-date state if examined intra-query. The underlying
4902 : * variable (in_hot_standby_guc) changes only when we transmit a new value
4903 : * to the client.
4904 : */
4905 31720 : return RecoveryInProgress() ? "on" : "off";
4906 : }
4907 :
4908 : /*
4909 : * GUC show_hook for effective_wal_level
4910 : */
4911 : const char *
4912 3516 : show_effective_wal_level(void)
4913 : {
4914 3516 : if (wal_level == WAL_LEVEL_MINIMAL)
4915 392 : return "minimal";
4916 :
4917 : /*
4918 : * During recovery, effective_wal_level reflects the primary's
4919 : * configuration rather than the local wal_level value.
4920 : */
4921 3124 : if (RecoveryInProgress())
4922 44 : return IsXLogLogicalInfoEnabled() ? "logical" : "replica";
4923 :
4924 3080 : return XLogLogicalInfoActive() ? "logical" : "replica";
4925 : }
4926 :
4927 : /*
4928 : * Read the control file, set respective GUCs.
4929 : *
4930 : * This is to be called during startup, including a crash recovery cycle,
4931 : * unless in bootstrap mode, where no control file yet exists. As there's no
4932 : * usable shared memory yet (its sizing can depend on the contents of the
4933 : * control file!), first store the contents in local memory. XLOGShmemInit()
4934 : * will then copy it to shared memory later.
4935 : *
4936 : * reset just controls whether previous contents are to be expected (in the
4937 : * reset case, there's a dangling pointer into old shared memory), or not.
4938 : */
4939 : void
4940 1982 : LocalProcessControlFile(bool reset)
4941 : {
4942 : Assert(reset || ControlFile == NULL);
4943 1982 : ControlFile = palloc_object(ControlFileData);
4944 1982 : ReadControlFile();
4945 1982 : }
4946 :
4947 : /*
4948 : * Get the wal_level from the control file. For a standby, this value should be
4949 : * considered as its active wal_level, because it may be different from what
4950 : * was originally configured on standby.
4951 : */
4952 : WalLevel
4953 0 : GetActiveWalLevelOnStandby(void)
4954 : {
4955 0 : return ControlFile->wal_level;
4956 : }
4957 :
4958 : /*
4959 : * Initialization of shared memory for XLOG
4960 : */
4961 : Size
4962 6510 : XLOGShmemSize(void)
4963 : {
4964 : Size size;
4965 :
4966 : /*
4967 : * If the value of wal_buffers is -1, use the preferred auto-tune value.
4968 : * This isn't an amazingly clean place to do this, but we must wait till
4969 : * NBuffers has received its final value, and must do it before using the
4970 : * value of XLOGbuffers to do anything important.
4971 : *
4972 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
4973 : * However, if the DBA explicitly set wal_buffers = -1 in the config file,
4974 : * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
4975 : * the matter with PGC_S_OVERRIDE.
4976 : */
4977 6510 : if (XLOGbuffers == -1)
4978 : {
4979 : char buf[32];
4980 :
4981 2268 : snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
4982 2268 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4983 : PGC_S_DYNAMIC_DEFAULT);
4984 2268 : if (XLOGbuffers == -1) /* failed to apply it? */
4985 0 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4986 : PGC_S_OVERRIDE);
4987 : }
4988 : Assert(XLOGbuffers > 0);
4989 :
4990 : /* XLogCtl */
4991 6510 : size = sizeof(XLogCtlData);
4992 :
4993 : /* WAL insertion locks, plus alignment */
4994 6510 : size = add_size(size, mul_size(sizeof(WALInsertLockPadded), NUM_XLOGINSERT_LOCKS + 1));
4995 : /* xlblocks array */
4996 6510 : size = add_size(size, mul_size(sizeof(pg_atomic_uint64), XLOGbuffers));
4997 : /* extra alignment padding for XLOG I/O buffers */
4998 6510 : size = add_size(size, Max(XLOG_BLCKSZ, PG_IO_ALIGN_SIZE));
4999 : /* and the buffers themselves */
5000 6510 : size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
5001 :
5002 : /*
5003 : * Note: we don't count ControlFileData, it comes out of the "slop factor"
5004 : * added by CreateSharedMemoryAndSemaphores. This lets us use this
5005 : * routine again below to compute the actual allocation size.
5006 : */
5007 :
5008 6510 : return size;
5009 : }
5010 :
5011 : void
5012 2272 : XLOGShmemInit(void)
5013 : {
5014 : bool foundCFile,
5015 : foundXLog;
5016 : char *allocptr;
5017 : int i;
5018 : ControlFileData *localControlFile;
5019 :
5020 : #ifdef WAL_DEBUG
5021 :
5022 : /*
5023 : * Create a memory context for WAL debugging that's exempt from the normal
5024 : * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
5025 : * an allocation fails, but wal_debug is not for production use anyway.
5026 : */
5027 : if (walDebugCxt == NULL)
5028 : {
5029 : walDebugCxt = AllocSetContextCreate(TopMemoryContext,
5030 : "WAL Debug",
5031 : ALLOCSET_DEFAULT_SIZES);
5032 : MemoryContextAllowInCriticalSection(walDebugCxt, true);
5033 : }
5034 : #endif
5035 :
5036 :
5037 2272 : XLogCtl = (XLogCtlData *)
5038 2272 : ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog);
5039 :
5040 2272 : localControlFile = ControlFile;
5041 2272 : ControlFile = (ControlFileData *)
5042 2272 : ShmemInitStruct("Control File", sizeof(ControlFileData), &foundCFile);
5043 :
5044 2272 : if (foundCFile || foundXLog)
5045 : {
5046 : /* both should be present or neither */
5047 : Assert(foundCFile && foundXLog);
5048 :
5049 : /* Initialize local copy of WALInsertLocks */
5050 0 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks;
5051 :
5052 0 : if (localControlFile)
5053 0 : pfree(localControlFile);
5054 0 : return;
5055 : }
5056 2272 : memset(XLogCtl, 0, sizeof(XLogCtlData));
5057 :
5058 : /*
5059 : * Already have read control file locally, unless in bootstrap mode. Move
5060 : * contents into shared memory.
5061 : */
5062 2272 : if (localControlFile)
5063 : {
5064 1966 : memcpy(ControlFile, localControlFile, sizeof(ControlFileData));
5065 1966 : pfree(localControlFile);
5066 : }
5067 :
5068 : /*
5069 : * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
5070 : * multiple of the alignment for same, so no extra alignment padding is
5071 : * needed here.
5072 : */
5073 2272 : allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
5074 2272 : XLogCtl->xlblocks = (pg_atomic_uint64 *) allocptr;
5075 2272 : allocptr += sizeof(pg_atomic_uint64) * XLOGbuffers;
5076 :
5077 650490 : for (i = 0; i < XLOGbuffers; i++)
5078 : {
5079 648218 : pg_atomic_init_u64(&XLogCtl->xlblocks[i], InvalidXLogRecPtr);
5080 : }
5081 :
5082 : /* WAL insertion locks. Ensure they're aligned to the full padded size */
5083 2272 : allocptr += sizeof(WALInsertLockPadded) -
5084 2272 : ((uintptr_t) allocptr) % sizeof(WALInsertLockPadded);
5085 2272 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks =
5086 : (WALInsertLockPadded *) allocptr;
5087 2272 : allocptr += sizeof(WALInsertLockPadded) * NUM_XLOGINSERT_LOCKS;
5088 :
5089 20448 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
5090 : {
5091 18176 : LWLockInitialize(&WALInsertLocks[i].l.lock, LWTRANCHE_WAL_INSERT);
5092 18176 : pg_atomic_init_u64(&WALInsertLocks[i].l.insertingAt, InvalidXLogRecPtr);
5093 18176 : WALInsertLocks[i].l.lastImportantAt = InvalidXLogRecPtr;
5094 : }
5095 :
5096 : /*
5097 : * Align the start of the page buffers to a full xlog block size boundary.
5098 : * This simplifies some calculations in XLOG insertion. It is also
5099 : * required for O_DIRECT.
5100 : */
5101 2272 : allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5102 2272 : XLogCtl->pages = allocptr;
5103 2272 : memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5104 :
5105 : /*
5106 : * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5107 : * in additional info.)
5108 : */
5109 2272 : XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5110 2272 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5111 2272 : XLogCtl->InstallXLogFileSegmentActive = false;
5112 2272 : XLogCtl->WalWriterSleeping = false;
5113 :
5114 2272 : SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5115 2272 : SpinLockInit(&XLogCtl->info_lck);
5116 2272 : pg_atomic_init_u64(&XLogCtl->logInsertResult, InvalidXLogRecPtr);
5117 2272 : pg_atomic_init_u64(&XLogCtl->logWriteResult, InvalidXLogRecPtr);
5118 2272 : pg_atomic_init_u64(&XLogCtl->logFlushResult, InvalidXLogRecPtr);
5119 2272 : pg_atomic_init_u64(&XLogCtl->unloggedLSN, InvalidXLogRecPtr);
5120 : }
5121 :
5122 : /*
5123 : * This func must be called ONCE on system install. It creates pg_control
5124 : * and the initial XLOG segment.
5125 : */
5126 : void
5127 102 : BootStrapXLOG(uint32 data_checksum_version)
5128 : {
5129 : CheckPoint checkPoint;
5130 : PGAlignedXLogBlock buffer;
5131 : XLogPageHeader page;
5132 : XLogLongPageHeader longpage;
5133 : XLogRecord *record;
5134 : char *recptr;
5135 : uint64 sysidentifier;
5136 : struct timeval tv;
5137 : pg_crc32c crc;
5138 :
5139 : /* allow ordinary WAL segment creation, like StartupXLOG() would */
5140 102 : SetInstallXLogFileSegmentActive();
5141 :
5142 : /*
5143 : * Select a hopefully-unique system identifier code for this installation.
5144 : * We use the result of gettimeofday(), including the fractional seconds
5145 : * field, as being about as unique as we can easily get. (Think not to
5146 : * use random(), since it hasn't been seeded and there's no portable way
5147 : * to seed it other than the system clock value...) The upper half of the
5148 : * uint64 value is just the tv_sec part, while the lower half contains the
5149 : * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
5150 : * PID for a little extra uniqueness. A person knowing this encoding can
5151 : * determine the initialization time of the installation, which could
5152 : * perhaps be useful sometimes.
5153 : */
5154 102 : gettimeofday(&tv, NULL);
5155 102 : sysidentifier = ((uint64) tv.tv_sec) << 32;
5156 102 : sysidentifier |= ((uint64) tv.tv_usec) << 12;
5157 102 : sysidentifier |= getpid() & 0xFFF;
5158 :
5159 102 : memset(&buffer, 0, sizeof buffer);
5160 102 : page = (XLogPageHeader) &buffer;
5161 :
5162 : /*
5163 : * Set up information for the initial checkpoint record
5164 : *
5165 : * The initial checkpoint record is written to the beginning of the WAL
5166 : * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5167 : * used, so that we can use 0/0 to mean "before any valid WAL segment".
5168 : */
5169 102 : checkPoint.redo = wal_segment_size + SizeOfXLogLongPHD;
5170 102 : checkPoint.ThisTimeLineID = BootstrapTimeLineID;
5171 102 : checkPoint.PrevTimeLineID = BootstrapTimeLineID;
5172 102 : checkPoint.fullPageWrites = fullPageWrites;
5173 102 : checkPoint.logicalDecodingEnabled = (wal_level == WAL_LEVEL_LOGICAL);
5174 102 : checkPoint.wal_level = wal_level;
5175 : checkPoint.nextXid =
5176 102 : FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId);
5177 102 : checkPoint.nextOid = FirstGenbkiObjectId;
5178 102 : checkPoint.nextMulti = FirstMultiXactId;
5179 102 : checkPoint.nextMultiOffset = 1;
5180 102 : checkPoint.oldestXid = FirstNormalTransactionId;
5181 102 : checkPoint.oldestXidDB = Template1DbOid;
5182 102 : checkPoint.oldestMulti = FirstMultiXactId;
5183 102 : checkPoint.oldestMultiDB = Template1DbOid;
5184 102 : checkPoint.oldestCommitTsXid = InvalidTransactionId;
5185 102 : checkPoint.newestCommitTsXid = InvalidTransactionId;
5186 102 : checkPoint.time = (pg_time_t) time(NULL);
5187 102 : checkPoint.oldestActiveXid = InvalidTransactionId;
5188 :
5189 102 : TransamVariables->nextXid = checkPoint.nextXid;
5190 102 : TransamVariables->nextOid = checkPoint.nextOid;
5191 102 : TransamVariables->oidCount = 0;
5192 102 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5193 102 : AdvanceOldestClogXid(checkPoint.oldestXid);
5194 102 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5195 102 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
5196 102 : SetCommitTsLimit(InvalidTransactionId, InvalidTransactionId);
5197 :
5198 : /* Set up the XLOG page header */
5199 102 : page->xlp_magic = XLOG_PAGE_MAGIC;
5200 102 : page->xlp_info = XLP_LONG_HEADER;
5201 102 : page->xlp_tli = BootstrapTimeLineID;
5202 102 : page->xlp_pageaddr = wal_segment_size;
5203 102 : longpage = (XLogLongPageHeader) page;
5204 102 : longpage->xlp_sysid = sysidentifier;
5205 102 : longpage->xlp_seg_size = wal_segment_size;
5206 102 : longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5207 :
5208 : /* Insert the initial checkpoint record */
5209 102 : recptr = ((char *) page + SizeOfXLogLongPHD);
5210 102 : record = (XLogRecord *) recptr;
5211 102 : record->xl_prev = 0;
5212 102 : record->xl_xid = InvalidTransactionId;
5213 102 : record->xl_tot_len = SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(checkPoint);
5214 102 : record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5215 102 : record->xl_rmid = RM_XLOG_ID;
5216 102 : recptr += SizeOfXLogRecord;
5217 : /* fill the XLogRecordDataHeaderShort struct */
5218 102 : *(recptr++) = (char) XLR_BLOCK_ID_DATA_SHORT;
5219 102 : *(recptr++) = sizeof(checkPoint);
5220 102 : memcpy(recptr, &checkPoint, sizeof(checkPoint));
5221 102 : recptr += sizeof(checkPoint);
5222 : Assert(recptr - (char *) record == record->xl_tot_len);
5223 :
5224 102 : INIT_CRC32C(crc);
5225 102 : COMP_CRC32C(crc, ((char *) record) + SizeOfXLogRecord, record->xl_tot_len - SizeOfXLogRecord);
5226 102 : COMP_CRC32C(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5227 102 : FIN_CRC32C(crc);
5228 102 : record->xl_crc = crc;
5229 :
5230 : /* Create first XLOG segment file */
5231 102 : openLogTLI = BootstrapTimeLineID;
5232 102 : openLogFile = XLogFileInit(1, BootstrapTimeLineID);
5233 :
5234 : /*
5235 : * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
5236 : * close the file again in a moment.
5237 : */
5238 :
5239 : /* Write the first page with the initial record */
5240 102 : errno = 0;
5241 102 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE);
5242 102 : if (write(openLogFile, &buffer, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5243 : {
5244 : /* if write didn't set errno, assume problem is no disk space */
5245 0 : if (errno == 0)
5246 0 : errno = ENOSPC;
5247 0 : ereport(PANIC,
5248 : (errcode_for_file_access(),
5249 : errmsg("could not write bootstrap write-ahead log file: %m")));
5250 : }
5251 102 : pgstat_report_wait_end();
5252 :
5253 102 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC);
5254 102 : if (pg_fsync(openLogFile) != 0)
5255 0 : ereport(PANIC,
5256 : (errcode_for_file_access(),
5257 : errmsg("could not fsync bootstrap write-ahead log file: %m")));
5258 102 : pgstat_report_wait_end();
5259 :
5260 102 : if (close(openLogFile) != 0)
5261 0 : ereport(PANIC,
5262 : (errcode_for_file_access(),
5263 : errmsg("could not close bootstrap write-ahead log file: %m")));
5264 :
5265 102 : openLogFile = -1;
5266 :
5267 : /* Now create pg_control */
5268 102 : InitControlFile(sysidentifier, data_checksum_version);
5269 102 : ControlFile->time = checkPoint.time;
5270 102 : ControlFile->checkPoint = checkPoint.redo;
5271 102 : ControlFile->checkPointCopy = checkPoint;
5272 :
5273 : /* some additional ControlFile fields are set in WriteControlFile() */
5274 102 : WriteControlFile();
5275 :
5276 : /* Bootstrap the commit log, too */
5277 102 : BootStrapCLOG();
5278 102 : BootStrapCommitTs();
5279 102 : BootStrapSUBTRANS();
5280 102 : BootStrapMultiXact();
5281 :
5282 : /*
5283 : * Force control file to be read - in contrast to normal processing we'd
5284 : * otherwise never run the checks and GUC related initializations therein.
5285 : */
5286 102 : ReadControlFile();
5287 102 : }
5288 :
5289 : static char *
5290 1758 : str_time(pg_time_t tnow, char *buf, size_t bufsize)
5291 : {
5292 1758 : pg_strftime(buf, bufsize,
5293 : "%Y-%m-%d %H:%M:%S %Z",
5294 1758 : pg_localtime(&tnow, log_timezone));
5295 :
5296 1758 : return buf;
5297 : }
5298 :
5299 : /*
5300 : * Initialize the first WAL segment on new timeline.
5301 : */
5302 : static void
5303 110 : XLogInitNewTimeline(TimeLineID endTLI, XLogRecPtr endOfLog, TimeLineID newTLI)
5304 : {
5305 : char xlogfname[MAXFNAMELEN];
5306 : XLogSegNo endLogSegNo;
5307 : XLogSegNo startLogSegNo;
5308 :
5309 : /* we always switch to a new timeline after archive recovery */
5310 : Assert(endTLI != newTLI);
5311 :
5312 : /*
5313 : * Update min recovery point one last time.
5314 : */
5315 110 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5316 :
5317 : /*
5318 : * Calculate the last segment on the old timeline, and the first segment
5319 : * on the new timeline. If the switch happens in the middle of a segment,
5320 : * they are the same, but if the switch happens exactly at a segment
5321 : * boundary, startLogSegNo will be endLogSegNo + 1.
5322 : */
5323 110 : XLByteToPrevSeg(endOfLog, endLogSegNo, wal_segment_size);
5324 110 : XLByteToSeg(endOfLog, startLogSegNo, wal_segment_size);
5325 :
5326 : /*
5327 : * Initialize the starting WAL segment for the new timeline. If the switch
5328 : * happens in the middle of a segment, copy data from the last WAL segment
5329 : * of the old timeline up to the switch point, to the starting WAL segment
5330 : * on the new timeline.
5331 : */
5332 110 : if (endLogSegNo == startLogSegNo)
5333 : {
5334 : /*
5335 : * Make a copy of the file on the new timeline.
5336 : *
5337 : * Writing WAL isn't allowed yet, so there are no locking
5338 : * considerations. But we should be just as tense as XLogFileInit to
5339 : * avoid emplacing a bogus file.
5340 : */
5341 90 : XLogFileCopy(newTLI, endLogSegNo, endTLI, endLogSegNo,
5342 90 : XLogSegmentOffset(endOfLog, wal_segment_size));
5343 : }
5344 : else
5345 : {
5346 : /*
5347 : * The switch happened at a segment boundary, so just create the next
5348 : * segment on the new timeline.
5349 : */
5350 : int fd;
5351 :
5352 20 : fd = XLogFileInit(startLogSegNo, newTLI);
5353 :
5354 20 : if (close(fd) != 0)
5355 : {
5356 0 : int save_errno = errno;
5357 :
5358 0 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5359 0 : errno = save_errno;
5360 0 : ereport(ERROR,
5361 : (errcode_for_file_access(),
5362 : errmsg("could not close file \"%s\": %m", xlogfname)));
5363 : }
5364 : }
5365 :
5366 : /*
5367 : * Let's just make real sure there are not .ready or .done flags posted
5368 : * for the new segment.
5369 : */
5370 110 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5371 110 : XLogArchiveCleanup(xlogfname);
5372 110 : }
5373 :
5374 : /*
5375 : * Perform cleanup actions at the conclusion of archive recovery.
5376 : */
5377 : static void
5378 110 : CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI, XLogRecPtr EndOfLog,
5379 : TimeLineID newTLI)
5380 : {
5381 : /*
5382 : * Execute the recovery_end_command, if any.
5383 : */
5384 110 : if (recoveryEndCommand && strcmp(recoveryEndCommand, "") != 0)
5385 4 : ExecuteRecoveryCommand(recoveryEndCommand,
5386 : "recovery_end_command",
5387 : true,
5388 : WAIT_EVENT_RECOVERY_END_COMMAND);
5389 :
5390 : /*
5391 : * We switched to a new timeline. Clean up segments on the old timeline.
5392 : *
5393 : * If there are any higher-numbered segments on the old timeline, remove
5394 : * them. They might contain valid WAL, but they might also be
5395 : * pre-allocated files containing garbage. In any case, they are not part
5396 : * of the new timeline's history so we don't need them.
5397 : */
5398 110 : RemoveNonParentXlogFiles(EndOfLog, newTLI);
5399 :
5400 : /*
5401 : * If the switch happened in the middle of a segment, what to do with the
5402 : * last, partial segment on the old timeline? If we don't archive it, and
5403 : * the server that created the WAL never archives it either (e.g. because
5404 : * it was hit by a meteor), it will never make it to the archive. That's
5405 : * OK from our point of view, because the new segment that we created with
5406 : * the new TLI contains all the WAL from the old timeline up to the switch
5407 : * point. But if you later try to do PITR to the "missing" WAL on the old
5408 : * timeline, recovery won't find it in the archive. It's physically
5409 : * present in the new file with new TLI, but recovery won't look there
5410 : * when it's recovering to the older timeline. On the other hand, if we
5411 : * archive the partial segment, and the original server on that timeline
5412 : * is still running and archives the completed version of the same segment
5413 : * later, it will fail. (We used to do that in 9.4 and below, and it
5414 : * caused such problems).
5415 : *
5416 : * As a compromise, we rename the last segment with the .partial suffix,
5417 : * and archive it. Archive recovery will never try to read .partial
5418 : * segments, so they will normally go unused. But in the odd PITR case,
5419 : * the administrator can copy them manually to the pg_wal directory
5420 : * (removing the suffix). They can be useful in debugging, too.
5421 : *
5422 : * If a .done or .ready file already exists for the old timeline, however,
5423 : * we had already determined that the segment is complete, so we can let
5424 : * it be archived normally. (In particular, if it was restored from the
5425 : * archive to begin with, it's expected to have a .done file).
5426 : */
5427 110 : if (XLogSegmentOffset(EndOfLog, wal_segment_size) != 0 &&
5428 : XLogArchivingActive())
5429 : {
5430 : char origfname[MAXFNAMELEN];
5431 : XLogSegNo endLogSegNo;
5432 :
5433 20 : XLByteToPrevSeg(EndOfLog, endLogSegNo, wal_segment_size);
5434 20 : XLogFileName(origfname, EndOfLogTLI, endLogSegNo, wal_segment_size);
5435 :
5436 20 : if (!XLogArchiveIsReadyOrDone(origfname))
5437 : {
5438 : char origpath[MAXPGPATH];
5439 : char partialfname[MAXFNAMELEN];
5440 : char partialpath[MAXPGPATH];
5441 :
5442 : /*
5443 : * If we're summarizing WAL, we can't rename the partial file
5444 : * until the summarizer finishes with it, else it will fail.
5445 : */
5446 12 : if (summarize_wal)
5447 2 : WaitForWalSummarization(EndOfLog);
5448 :
5449 12 : XLogFilePath(origpath, EndOfLogTLI, endLogSegNo, wal_segment_size);
5450 12 : snprintf(partialfname, MAXFNAMELEN, "%s.partial", origfname);
5451 12 : snprintf(partialpath, MAXPGPATH, "%s.partial", origpath);
5452 :
5453 : /*
5454 : * Make sure there's no .done or .ready file for the .partial
5455 : * file.
5456 : */
5457 12 : XLogArchiveCleanup(partialfname);
5458 :
5459 12 : durable_rename(origpath, partialpath, ERROR);
5460 12 : XLogArchiveNotify(partialfname);
5461 : }
5462 : }
5463 110 : }
5464 :
5465 : /*
5466 : * Check to see if required parameters are set high enough on this server
5467 : * for various aspects of recovery operation.
5468 : *
5469 : * Note that all the parameters which this function tests need to be
5470 : * listed in Administrator's Overview section in high-availability.sgml.
5471 : * If you change them, don't forget to update the list.
5472 : */
5473 : static void
5474 520 : CheckRequiredParameterValues(void)
5475 : {
5476 : /*
5477 : * For archive recovery, the WAL must be generated with at least 'replica'
5478 : * wal_level.
5479 : */
5480 520 : if (ArchiveRecoveryRequested && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5481 : {
5482 4 : ereport(FATAL,
5483 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5484 : errmsg("WAL was generated with \"wal_level=minimal\", cannot continue recovering"),
5485 : errdetail("This happens if you temporarily set \"wal_level=minimal\" on the server."),
5486 : errhint("Use a backup taken after setting \"wal_level\" to higher than \"minimal\".")));
5487 : }
5488 :
5489 : /*
5490 : * For Hot Standby, the WAL must be generated with 'replica' mode, and we
5491 : * must have at least as many backend slots as the primary.
5492 : */
5493 516 : if (ArchiveRecoveryRequested && EnableHotStandby)
5494 : {
5495 : /* We ignore autovacuum_worker_slots when we make this test. */
5496 270 : RecoveryRequiresIntParameter("max_connections",
5497 : MaxConnections,
5498 270 : ControlFile->MaxConnections);
5499 270 : RecoveryRequiresIntParameter("max_worker_processes",
5500 : max_worker_processes,
5501 270 : ControlFile->max_worker_processes);
5502 270 : RecoveryRequiresIntParameter("max_wal_senders",
5503 : max_wal_senders,
5504 270 : ControlFile->max_wal_senders);
5505 270 : RecoveryRequiresIntParameter("max_prepared_transactions",
5506 : max_prepared_xacts,
5507 270 : ControlFile->max_prepared_xacts);
5508 270 : RecoveryRequiresIntParameter("max_locks_per_transaction",
5509 : max_locks_per_xact,
5510 270 : ControlFile->max_locks_per_xact);
5511 : }
5512 516 : }
5513 :
5514 : /*
5515 : * This must be called ONCE during postmaster or standalone-backend startup
5516 : */
5517 : void
5518 1982 : StartupXLOG(void)
5519 : {
5520 : XLogCtlInsert *Insert;
5521 : CheckPoint checkPoint;
5522 : bool wasShutdown;
5523 : bool didCrash;
5524 : bool haveTblspcMap;
5525 : bool haveBackupLabel;
5526 : XLogRecPtr EndOfLog;
5527 : TimeLineID EndOfLogTLI;
5528 : TimeLineID newTLI;
5529 : bool performedWalRecovery;
5530 : EndOfWalRecoveryInfo *endOfRecoveryInfo;
5531 : XLogRecPtr abortedRecPtr;
5532 : XLogRecPtr missingContrecPtr;
5533 : TransactionId oldestActiveXID;
5534 1982 : bool promoted = false;
5535 : char timebuf[128];
5536 :
5537 : /*
5538 : * We should have an aux process resource owner to use, and we should not
5539 : * be in a transaction that's installed some other resowner.
5540 : */
5541 : Assert(AuxProcessResourceOwner != NULL);
5542 : Assert(CurrentResourceOwner == NULL ||
5543 : CurrentResourceOwner == AuxProcessResourceOwner);
5544 1982 : CurrentResourceOwner = AuxProcessResourceOwner;
5545 :
5546 : /*
5547 : * Check that contents look valid.
5548 : */
5549 1982 : if (!XRecOffIsValid(ControlFile->checkPoint))
5550 0 : ereport(FATAL,
5551 : (errcode(ERRCODE_DATA_CORRUPTED),
5552 : errmsg("control file contains invalid checkpoint location")));
5553 :
5554 1982 : switch (ControlFile->state)
5555 : {
5556 1548 : case DB_SHUTDOWNED:
5557 :
5558 : /*
5559 : * This is the expected case, so don't be chatty in standalone
5560 : * mode
5561 : */
5562 1548 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
5563 : (errmsg("database system was shut down at %s",
5564 : str_time(ControlFile->time,
5565 : timebuf, sizeof(timebuf)))));
5566 1548 : break;
5567 :
5568 66 : case DB_SHUTDOWNED_IN_RECOVERY:
5569 66 : ereport(LOG,
5570 : (errmsg("database system was shut down in recovery at %s",
5571 : str_time(ControlFile->time,
5572 : timebuf, sizeof(timebuf)))));
5573 66 : break;
5574 :
5575 0 : case DB_SHUTDOWNING:
5576 0 : ereport(LOG,
5577 : (errmsg("database system shutdown was interrupted; last known up at %s",
5578 : str_time(ControlFile->time,
5579 : timebuf, sizeof(timebuf)))));
5580 0 : break;
5581 :
5582 0 : case DB_IN_CRASH_RECOVERY:
5583 0 : ereport(LOG,
5584 : (errmsg("database system was interrupted while in recovery at %s",
5585 : str_time(ControlFile->time,
5586 : timebuf, sizeof(timebuf))),
5587 : errhint("This probably means that some data is corrupted and"
5588 : " you will have to use the last backup for recovery.")));
5589 0 : break;
5590 :
5591 14 : case DB_IN_ARCHIVE_RECOVERY:
5592 14 : ereport(LOG,
5593 : (errmsg("database system was interrupted while in recovery at log time %s",
5594 : str_time(ControlFile->checkPointCopy.time,
5595 : timebuf, sizeof(timebuf))),
5596 : errhint("If this has occurred more than once some data might be corrupted"
5597 : " and you might need to choose an earlier recovery target.")));
5598 14 : break;
5599 :
5600 354 : case DB_IN_PRODUCTION:
5601 354 : ereport(LOG,
5602 : (errmsg("database system was interrupted; last known up at %s",
5603 : str_time(ControlFile->time,
5604 : timebuf, sizeof(timebuf)))));
5605 354 : break;
5606 :
5607 0 : default:
5608 0 : ereport(FATAL,
5609 : (errcode(ERRCODE_DATA_CORRUPTED),
5610 : errmsg("control file contains invalid database cluster state")));
5611 : }
5612 :
5613 : /* This is just to allow attaching to startup process with a debugger */
5614 : #ifdef XLOG_REPLAY_DELAY
5615 : if (ControlFile->state != DB_SHUTDOWNED)
5616 : pg_usleep(60000000L);
5617 : #endif
5618 :
5619 : /*
5620 : * Verify that pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
5621 : * In cases where someone has performed a copy for PITR, these directories
5622 : * may have been excluded and need to be re-created.
5623 : */
5624 1982 : ValidateXLOGDirectoryStructure();
5625 :
5626 : /* Set up timeout handler needed to report startup progress. */
5627 1982 : if (!IsBootstrapProcessingMode())
5628 1880 : RegisterTimeout(STARTUP_PROGRESS_TIMEOUT,
5629 : startup_progress_timeout_handler);
5630 :
5631 : /*----------
5632 : * If we previously crashed, perform a couple of actions:
5633 : *
5634 : * - The pg_wal directory may still include some temporary WAL segments
5635 : * used when creating a new segment, so perform some clean up to not
5636 : * bloat this path. This is done first as there is no point to sync
5637 : * this temporary data.
5638 : *
5639 : * - There might be data which we had written, intending to fsync it, but
5640 : * which we had not actually fsync'd yet. Therefore, a power failure in
5641 : * the near future might cause earlier unflushed writes to be lost, even
5642 : * though more recent data written to disk from here on would be
5643 : * persisted. To avoid that, fsync the entire data directory.
5644 : */
5645 1982 : if (ControlFile->state != DB_SHUTDOWNED &&
5646 434 : ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
5647 : {
5648 368 : RemoveTempXlogFiles();
5649 368 : SyncDataDirectory();
5650 368 : didCrash = true;
5651 : }
5652 : else
5653 1614 : didCrash = false;
5654 :
5655 : /*
5656 : * Prepare for WAL recovery if needed.
5657 : *
5658 : * InitWalRecovery analyzes the control file and the backup label file, if
5659 : * any. It updates the in-memory ControlFile buffer according to the
5660 : * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5661 : * It also applies the tablespace map file, if any.
5662 : */
5663 1982 : InitWalRecovery(ControlFile, &wasShutdown,
5664 : &haveBackupLabel, &haveTblspcMap);
5665 1980 : checkPoint = ControlFile->checkPointCopy;
5666 :
5667 : /* initialize shared memory variables from the checkpoint record */
5668 1980 : TransamVariables->nextXid = checkPoint.nextXid;
5669 1980 : TransamVariables->nextOid = checkPoint.nextOid;
5670 1980 : TransamVariables->oidCount = 0;
5671 1980 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5672 1980 : AdvanceOldestClogXid(checkPoint.oldestXid);
5673 1980 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5674 1980 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
5675 1980 : SetCommitTsLimit(checkPoint.oldestCommitTsXid,
5676 : checkPoint.newestCommitTsXid);
5677 :
5678 : /*
5679 : * Clear out any old relcache cache files. This is *necessary* if we do
5680 : * any WAL replay, since that would probably result in the cache files
5681 : * being out of sync with database reality. In theory we could leave them
5682 : * in place if the database had been cleanly shut down, but it seems
5683 : * safest to just remove them always and let them be rebuilt during the
5684 : * first backend startup. These files needs to be removed from all
5685 : * directories including pg_tblspc, however the symlinks are created only
5686 : * after reading tablespace_map file in case of archive recovery from
5687 : * backup, so needs to clear old relcache files here after creating
5688 : * symlinks.
5689 : */
5690 1980 : RelationCacheInitFileRemove();
5691 :
5692 : /*
5693 : * Initialize replication slots, before there's a chance to remove
5694 : * required resources.
5695 : */
5696 1980 : StartupReplicationSlots();
5697 :
5698 : /*
5699 : * Startup the logical decoding status with the last status stored in the
5700 : * checkpoint record.
5701 : */
5702 1976 : StartupLogicalDecodingStatus(checkPoint.logicalDecodingEnabled);
5703 :
5704 : /*
5705 : * Startup logical state, needs to be setup now so we have proper data
5706 : * during crash recovery.
5707 : */
5708 1976 : StartupReorderBuffer();
5709 :
5710 : /*
5711 : * Startup CLOG. This must be done after TransamVariables->nextXid has
5712 : * been initialized and before we accept connections or begin WAL replay.
5713 : */
5714 1976 : StartupCLOG();
5715 :
5716 : /*
5717 : * Startup MultiXact. We need to do this early to be able to replay
5718 : * truncations.
5719 : */
5720 1976 : StartupMultiXact();
5721 :
5722 : /*
5723 : * Ditto for commit timestamps. Activate the facility if the setting is
5724 : * enabled in the control file, as there should be no tracking of commit
5725 : * timestamps done when the setting was disabled. This facility can be
5726 : * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
5727 : */
5728 1976 : if (ControlFile->track_commit_timestamp)
5729 28 : StartupCommitTs();
5730 :
5731 : /*
5732 : * Recover knowledge about replay progress of known replication partners.
5733 : */
5734 1976 : StartupReplicationOrigin();
5735 :
5736 : /*
5737 : * Initialize unlogged LSN. On a clean shutdown, it's restored from the
5738 : * control file. On recovery, all unlogged relations are blown away, so
5739 : * the unlogged LSN counter can be reset too.
5740 : */
5741 1976 : if (ControlFile->state == DB_SHUTDOWNED)
5742 1532 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5743 1532 : ControlFile->unloggedLSN);
5744 : else
5745 444 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5746 : FirstNormalUnloggedLSN);
5747 :
5748 : /*
5749 : * Copy any missing timeline history files between 'now' and the recovery
5750 : * target timeline from archive to pg_wal. While we don't need those files
5751 : * ourselves - the history file of the recovery target timeline covers all
5752 : * the previous timelines in the history too - a cascading standby server
5753 : * might be interested in them. Or, if you archive the WAL from this
5754 : * server to a different archive than the primary, it'd be good for all
5755 : * the history files to get archived there after failover, so that you can
5756 : * use one of the old timelines as a PITR target. Timeline history files
5757 : * are small, so it's better to copy them unnecessarily than not copy them
5758 : * and regret later.
5759 : */
5760 1976 : restoreTimeLineHistoryFiles(checkPoint.ThisTimeLineID, recoveryTargetTLI);
5761 :
5762 : /*
5763 : * Before running in recovery, scan pg_twophase and fill in its status to
5764 : * be able to work on entries generated by redo. Doing a scan before
5765 : * taking any recovery action has the merit to discard any 2PC files that
5766 : * are newer than the first record to replay, saving from any conflicts at
5767 : * replay. This avoids as well any subsequent scans when doing recovery
5768 : * of the on-disk two-phase data.
5769 : */
5770 1976 : restoreTwoPhaseData();
5771 :
5772 : /*
5773 : * When starting with crash recovery, reset pgstat data - it might not be
5774 : * valid. Otherwise restore pgstat data. It's safe to do this here,
5775 : * because postmaster will not yet have started any other processes.
5776 : *
5777 : * NB: Restoring replication slot stats relies on slot state to have
5778 : * already been restored from disk.
5779 : *
5780 : * TODO: With a bit of extra work we could just start with a pgstat file
5781 : * associated with the checkpoint redo location we're starting from.
5782 : */
5783 1976 : if (didCrash)
5784 366 : pgstat_discard_stats();
5785 : else
5786 1610 : pgstat_restore_stats();
5787 :
5788 1976 : lastFullPageWrites = checkPoint.fullPageWrites;
5789 :
5790 1976 : RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
5791 1976 : doPageWrites = lastFullPageWrites;
5792 :
5793 : /* REDO */
5794 1976 : if (InRecovery)
5795 : {
5796 : /* Initialize state for RecoveryInProgress() */
5797 444 : SpinLockAcquire(&XLogCtl->info_lck);
5798 444 : if (InArchiveRecovery)
5799 238 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
5800 : else
5801 206 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5802 444 : SpinLockRelease(&XLogCtl->info_lck);
5803 :
5804 : /*
5805 : * Update pg_control to show that we are recovering and to show the
5806 : * selected checkpoint as the place we are starting from. We also mark
5807 : * pg_control with any minimum recovery stop point obtained from a
5808 : * backup history file.
5809 : *
5810 : * No need to hold ControlFileLock yet, we aren't up far enough.
5811 : */
5812 444 : UpdateControlFile();
5813 :
5814 : /*
5815 : * If there was a backup label file, it's done its job and the info
5816 : * has now been propagated into pg_control. We must get rid of the
5817 : * label file so that if we crash during recovery, we'll pick up at
5818 : * the latest recovery restartpoint instead of going all the way back
5819 : * to the backup start point. It seems prudent though to just rename
5820 : * the file out of the way rather than delete it completely.
5821 : */
5822 444 : if (haveBackupLabel)
5823 : {
5824 156 : unlink(BACKUP_LABEL_OLD);
5825 156 : durable_rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD, FATAL);
5826 : }
5827 :
5828 : /*
5829 : * If there was a tablespace_map file, it's done its job and the
5830 : * symlinks have been created. We must get rid of the map file so
5831 : * that if we crash during recovery, we don't create symlinks again.
5832 : * It seems prudent though to just rename the file out of the way
5833 : * rather than delete it completely.
5834 : */
5835 444 : if (haveTblspcMap)
5836 : {
5837 4 : unlink(TABLESPACE_MAP_OLD);
5838 4 : durable_rename(TABLESPACE_MAP, TABLESPACE_MAP_OLD, FATAL);
5839 : }
5840 :
5841 : /*
5842 : * Initialize our local copy of minRecoveryPoint. When doing crash
5843 : * recovery we want to replay up to the end of WAL. Particularly, in
5844 : * the case of a promoted standby minRecoveryPoint value in the
5845 : * control file is only updated after the first checkpoint. However,
5846 : * if the instance crashes before the first post-recovery checkpoint
5847 : * is completed then recovery will use a stale location causing the
5848 : * startup process to think that there are still invalid page
5849 : * references when checking for data consistency.
5850 : */
5851 444 : if (InArchiveRecovery)
5852 : {
5853 238 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
5854 238 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
5855 : }
5856 : else
5857 : {
5858 206 : LocalMinRecoveryPoint = InvalidXLogRecPtr;
5859 206 : LocalMinRecoveryPointTLI = 0;
5860 : }
5861 :
5862 : /* Check that the GUCs used to generate the WAL allow recovery */
5863 444 : CheckRequiredParameterValues();
5864 :
5865 : /*
5866 : * We're in recovery, so unlogged relations may be trashed and must be
5867 : * reset. This should be done BEFORE allowing Hot Standby
5868 : * connections, so that read-only backends don't try to read whatever
5869 : * garbage is left over from before.
5870 : */
5871 444 : ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
5872 :
5873 : /*
5874 : * Likewise, delete any saved transaction snapshot files that got left
5875 : * behind by crashed backends.
5876 : */
5877 444 : DeleteAllExportedSnapshotFiles();
5878 :
5879 : /*
5880 : * Initialize for Hot Standby, if enabled. We won't let backends in
5881 : * yet, not until we've reached the min recovery point specified in
5882 : * control file and we've established a recovery snapshot from a
5883 : * running-xacts WAL record.
5884 : */
5885 444 : if (ArchiveRecoveryRequested && EnableHotStandby)
5886 : {
5887 : TransactionId *xids;
5888 : int nxids;
5889 :
5890 226 : ereport(DEBUG1,
5891 : (errmsg_internal("initializing for hot standby")));
5892 :
5893 226 : InitRecoveryTransactionEnvironment();
5894 :
5895 226 : if (wasShutdown)
5896 52 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
5897 : else
5898 174 : oldestActiveXID = checkPoint.oldestActiveXid;
5899 : Assert(TransactionIdIsValid(oldestActiveXID));
5900 :
5901 : /* Tell procarray about the range of xids it has to deal with */
5902 226 : ProcArrayInitRecovery(XidFromFullTransactionId(TransamVariables->nextXid));
5903 :
5904 : /*
5905 : * Startup subtrans only. CLOG, MultiXact and commit timestamp
5906 : * have already been started up and other SLRUs are not maintained
5907 : * during recovery and need not be started yet.
5908 : */
5909 226 : StartupSUBTRANS(oldestActiveXID);
5910 :
5911 : /*
5912 : * If we're beginning at a shutdown checkpoint, we know that
5913 : * nothing was running on the primary at this point. So fake-up an
5914 : * empty running-xacts record and use that here and now. Recover
5915 : * additional standby state for prepared transactions.
5916 : */
5917 226 : if (wasShutdown)
5918 : {
5919 : RunningTransactionsData running;
5920 : TransactionId latestCompletedXid;
5921 :
5922 : /* Update pg_subtrans entries for any prepared transactions */
5923 52 : StandbyRecoverPreparedTransactions();
5924 :
5925 : /*
5926 : * Construct a RunningTransactions snapshot representing a
5927 : * shut down server, with only prepared transactions still
5928 : * alive. We're never overflowed at this point because all
5929 : * subxids are listed with their parent prepared transactions.
5930 : */
5931 52 : running.xcnt = nxids;
5932 52 : running.subxcnt = 0;
5933 52 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
5934 52 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
5935 52 : running.oldestRunningXid = oldestActiveXID;
5936 52 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
5937 52 : TransactionIdRetreat(latestCompletedXid);
5938 : Assert(TransactionIdIsNormal(latestCompletedXid));
5939 52 : running.latestCompletedXid = latestCompletedXid;
5940 52 : running.xids = xids;
5941 :
5942 52 : ProcArrayApplyRecoveryInfo(&running);
5943 : }
5944 : }
5945 :
5946 : /*
5947 : * We're all set for replaying the WAL now. Do it.
5948 : */
5949 444 : PerformWalRecovery();
5950 324 : performedWalRecovery = true;
5951 : }
5952 : else
5953 1532 : performedWalRecovery = false;
5954 :
5955 : /*
5956 : * Finish WAL recovery.
5957 : */
5958 1856 : endOfRecoveryInfo = FinishWalRecovery();
5959 1856 : EndOfLog = endOfRecoveryInfo->endOfLog;
5960 1856 : EndOfLogTLI = endOfRecoveryInfo->endOfLogTLI;
5961 1856 : abortedRecPtr = endOfRecoveryInfo->abortedRecPtr;
5962 1856 : missingContrecPtr = endOfRecoveryInfo->missingContrecPtr;
5963 :
5964 : /*
5965 : * Reset ps status display, so as no information related to recovery shows
5966 : * up.
5967 : */
5968 1856 : set_ps_display("");
5969 :
5970 : /*
5971 : * When recovering from a backup (we are in recovery, and archive recovery
5972 : * was requested), complain if we did not roll forward far enough to reach
5973 : * the point where the database is consistent. For regular online
5974 : * backup-from-primary, that means reaching the end-of-backup WAL record
5975 : * (at which point we reset backupStartPoint to be Invalid), for
5976 : * backup-from-replica (which can't inject records into the WAL stream),
5977 : * that point is when we reach the minRecoveryPoint in pg_control (which
5978 : * we purposefully copy last when backing up from a replica). For
5979 : * pg_rewind (which creates a backup_label with a method of "pg_rewind")
5980 : * or snapshot-style backups (which don't), backupEndRequired will be set
5981 : * to false.
5982 : *
5983 : * Note: it is indeed okay to look at the local variable
5984 : * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
5985 : * might be further ahead --- ControlFile->minRecoveryPoint cannot have
5986 : * been advanced beyond the WAL we processed.
5987 : */
5988 1856 : if (InRecovery &&
5989 324 : (EndOfLog < LocalMinRecoveryPoint ||
5990 324 : XLogRecPtrIsValid(ControlFile->backupStartPoint)))
5991 : {
5992 : /*
5993 : * Ran off end of WAL before reaching end-of-backup WAL record, or
5994 : * minRecoveryPoint. That's a bad sign, indicating that you tried to
5995 : * recover from an online backup but never called pg_backup_stop(), or
5996 : * you didn't archive all the WAL needed.
5997 : */
5998 0 : if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
5999 : {
6000 0 : if (XLogRecPtrIsValid(ControlFile->backupStartPoint) || ControlFile->backupEndRequired)
6001 0 : ereport(FATAL,
6002 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6003 : errmsg("WAL ends before end of online backup"),
6004 : errhint("All WAL generated while online backup was taken must be available at recovery.")));
6005 : else
6006 0 : ereport(FATAL,
6007 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6008 : errmsg("WAL ends before consistent recovery point")));
6009 : }
6010 : }
6011 :
6012 : /*
6013 : * Reset unlogged relations to the contents of their INIT fork. This is
6014 : * done AFTER recovery is complete so as to include any unlogged relations
6015 : * created during recovery, but BEFORE recovery is marked as having
6016 : * completed successfully. Otherwise we'd not retry if any of the post
6017 : * end-of-recovery steps fail.
6018 : */
6019 1856 : if (InRecovery)
6020 324 : ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
6021 :
6022 : /*
6023 : * Pre-scan prepared transactions to find out the range of XIDs present.
6024 : * This information is not quite needed yet, but it is positioned here so
6025 : * as potential problems are detected before any on-disk change is done.
6026 : */
6027 1856 : oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
6028 :
6029 : /*
6030 : * Allow ordinary WAL segment creation before possibly switching to a new
6031 : * timeline, which creates a new segment, and after the last ReadRecord().
6032 : */
6033 1856 : SetInstallXLogFileSegmentActive();
6034 :
6035 : /*
6036 : * Consider whether we need to assign a new timeline ID.
6037 : *
6038 : * If we did archive recovery, we always assign a new ID. This handles a
6039 : * couple of issues. If we stopped short of the end of WAL during
6040 : * recovery, then we are clearly generating a new timeline and must assign
6041 : * it a unique new ID. Even if we ran to the end, modifying the current
6042 : * last segment is problematic because it may result in trying to
6043 : * overwrite an already-archived copy of that segment, and we encourage
6044 : * DBAs to make their archive_commands reject that. We can dodge the
6045 : * problem by making the new active segment have a new timeline ID.
6046 : *
6047 : * In a normal crash recovery, we can just extend the timeline we were in.
6048 : */
6049 1856 : newTLI = endOfRecoveryInfo->lastRecTLI;
6050 1856 : if (ArchiveRecoveryRequested)
6051 : {
6052 110 : newTLI = findNewestTimeLine(recoveryTargetTLI) + 1;
6053 110 : ereport(LOG,
6054 : (errmsg("selected new timeline ID: %u", newTLI)));
6055 :
6056 : /*
6057 : * Make a writable copy of the last WAL segment. (Note that we also
6058 : * have a copy of the last block of the old WAL in
6059 : * endOfRecovery->lastPage; we will use that below.)
6060 : */
6061 110 : XLogInitNewTimeline(EndOfLogTLI, EndOfLog, newTLI);
6062 :
6063 : /*
6064 : * Remove the signal files out of the way, so that we don't
6065 : * accidentally re-enter archive recovery mode in a subsequent crash.
6066 : */
6067 110 : if (endOfRecoveryInfo->standby_signal_file_found)
6068 104 : durable_unlink(STANDBY_SIGNAL_FILE, FATAL);
6069 :
6070 110 : if (endOfRecoveryInfo->recovery_signal_file_found)
6071 6 : durable_unlink(RECOVERY_SIGNAL_FILE, FATAL);
6072 :
6073 : /*
6074 : * Write the timeline history file, and have it archived. After this
6075 : * point (or rather, as soon as the file is archived), the timeline
6076 : * will appear as "taken" in the WAL archive and to any standby
6077 : * servers. If we crash before actually switching to the new
6078 : * timeline, standby servers will nevertheless think that we switched
6079 : * to the new timeline, and will try to connect to the new timeline.
6080 : * To minimize the window for that, try to do as little as possible
6081 : * between here and writing the end-of-recovery record.
6082 : */
6083 110 : writeTimeLineHistory(newTLI, recoveryTargetTLI,
6084 : EndOfLog, endOfRecoveryInfo->recoveryStopReason);
6085 :
6086 110 : ereport(LOG,
6087 : (errmsg("archive recovery complete")));
6088 : }
6089 :
6090 : /* Save the selected TimeLineID in shared memory, too */
6091 1856 : SpinLockAcquire(&XLogCtl->info_lck);
6092 1856 : XLogCtl->InsertTimeLineID = newTLI;
6093 1856 : XLogCtl->PrevTimeLineID = endOfRecoveryInfo->lastRecTLI;
6094 1856 : SpinLockRelease(&XLogCtl->info_lck);
6095 :
6096 : /*
6097 : * Actually, if WAL ended in an incomplete record, skip the parts that
6098 : * made it through and start writing after the portion that persisted.
6099 : * (It's critical to first write an OVERWRITE_CONTRECORD message, which
6100 : * we'll do as soon as we're open for writing new WAL.)
6101 : */
6102 1856 : if (XLogRecPtrIsValid(missingContrecPtr))
6103 : {
6104 : /*
6105 : * We should only have a missingContrecPtr if we're not switching to a
6106 : * new timeline. When a timeline switch occurs, WAL is copied from the
6107 : * old timeline to the new only up to the end of the last complete
6108 : * record, so there can't be an incomplete WAL record that we need to
6109 : * disregard.
6110 : */
6111 : Assert(newTLI == endOfRecoveryInfo->lastRecTLI);
6112 : Assert(XLogRecPtrIsValid(abortedRecPtr));
6113 22 : EndOfLog = missingContrecPtr;
6114 : }
6115 :
6116 : /*
6117 : * Prepare to write WAL starting at EndOfLog location, and init xlog
6118 : * buffer cache using the block containing the last record from the
6119 : * previous incarnation.
6120 : */
6121 1856 : Insert = &XLogCtl->Insert;
6122 1856 : Insert->PrevBytePos = XLogRecPtrToBytePos(endOfRecoveryInfo->lastRec);
6123 1856 : Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
6124 :
6125 : /*
6126 : * Tricky point here: lastPage contains the *last* block that the LastRec
6127 : * record spans, not the one it starts in. The last block is indeed the
6128 : * one we want to use.
6129 : */
6130 1856 : if (EndOfLog % XLOG_BLCKSZ != 0)
6131 : {
6132 : char *page;
6133 : int len;
6134 : int firstIdx;
6135 :
6136 1800 : firstIdx = XLogRecPtrToBufIdx(EndOfLog);
6137 1800 : len = EndOfLog - endOfRecoveryInfo->lastPageBeginPtr;
6138 : Assert(len < XLOG_BLCKSZ);
6139 :
6140 : /* Copy the valid part of the last block, and zero the rest */
6141 1800 : page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
6142 1800 : memcpy(page, endOfRecoveryInfo->lastPage, len);
6143 1800 : memset(page + len, 0, XLOG_BLCKSZ - len);
6144 :
6145 1800 : pg_atomic_write_u64(&XLogCtl->xlblocks[firstIdx], endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6146 1800 : XLogCtl->InitializedUpTo = endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ;
6147 : }
6148 : else
6149 : {
6150 : /*
6151 : * There is no partial block to copy. Just set InitializedUpTo, and
6152 : * let the first attempt to insert a log record to initialize the next
6153 : * buffer.
6154 : */
6155 56 : XLogCtl->InitializedUpTo = EndOfLog;
6156 : }
6157 :
6158 : /*
6159 : * Update local and shared status. This is OK to do without any locks
6160 : * because no other process can be reading or writing WAL yet.
6161 : */
6162 1856 : LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
6163 1856 : pg_atomic_write_u64(&XLogCtl->logInsertResult, EndOfLog);
6164 1856 : pg_atomic_write_u64(&XLogCtl->logWriteResult, EndOfLog);
6165 1856 : pg_atomic_write_u64(&XLogCtl->logFlushResult, EndOfLog);
6166 1856 : XLogCtl->LogwrtRqst.Write = EndOfLog;
6167 1856 : XLogCtl->LogwrtRqst.Flush = EndOfLog;
6168 :
6169 : /*
6170 : * Preallocate additional log files, if wanted.
6171 : */
6172 1856 : PreallocXlogFiles(EndOfLog, newTLI);
6173 :
6174 : /*
6175 : * Okay, we're officially UP.
6176 : */
6177 1856 : InRecovery = false;
6178 :
6179 : /* start the archive_timeout timer and LSN running */
6180 1856 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
6181 1856 : XLogCtl->lastSegSwitchLSN = EndOfLog;
6182 :
6183 : /* also initialize latestCompletedXid, to nextXid - 1 */
6184 1856 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
6185 1856 : TransamVariables->latestCompletedXid = TransamVariables->nextXid;
6186 1856 : FullTransactionIdRetreat(&TransamVariables->latestCompletedXid);
6187 1856 : LWLockRelease(ProcArrayLock);
6188 :
6189 : /*
6190 : * Start up subtrans, if not already done for hot standby. (commit
6191 : * timestamps are started below, if necessary.)
6192 : */
6193 1856 : if (standbyState == STANDBY_DISABLED)
6194 1746 : StartupSUBTRANS(oldestActiveXID);
6195 :
6196 : /*
6197 : * Perform end of recovery actions for any SLRUs that need it.
6198 : */
6199 1856 : TrimCLOG();
6200 1856 : TrimMultiXact();
6201 :
6202 : /*
6203 : * Reload shared-memory state for prepared transactions. This needs to
6204 : * happen before renaming the last partial segment of the old timeline as
6205 : * it may be possible that we have to recover some transactions from it.
6206 : */
6207 1856 : RecoverPreparedTransactions();
6208 :
6209 : /* Shut down xlogreader */
6210 1856 : ShutdownWalRecovery();
6211 :
6212 : /* Enable WAL writes for this backend only. */
6213 1856 : LocalSetXLogInsertAllowed();
6214 :
6215 : /* If necessary, write overwrite-contrecord before doing anything else */
6216 1856 : if (XLogRecPtrIsValid(abortedRecPtr))
6217 : {
6218 : Assert(XLogRecPtrIsValid(missingContrecPtr));
6219 22 : CreateOverwriteContrecordRecord(abortedRecPtr, missingContrecPtr, newTLI);
6220 : }
6221 :
6222 : /*
6223 : * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
6224 : * record before resource manager writes cleanup WAL records or checkpoint
6225 : * record is written.
6226 : */
6227 1856 : Insert->fullPageWrites = lastFullPageWrites;
6228 1856 : UpdateFullPageWrites();
6229 :
6230 : /*
6231 : * Emit checkpoint or end-of-recovery record in XLOG, if required.
6232 : */
6233 1856 : if (performedWalRecovery)
6234 324 : promoted = PerformRecoveryXLogAction();
6235 :
6236 : /*
6237 : * If any of the critical GUCs have changed, log them before we allow
6238 : * backends to write WAL.
6239 : */
6240 1856 : XLogReportParameters();
6241 :
6242 : /* If this is archive recovery, perform post-recovery cleanup actions. */
6243 1856 : if (ArchiveRecoveryRequested)
6244 110 : CleanupAfterArchiveRecovery(EndOfLogTLI, EndOfLog, newTLI);
6245 :
6246 : /*
6247 : * Local WAL inserts enabled, so it's time to finish initialization of
6248 : * commit timestamp.
6249 : */
6250 1856 : CompleteCommitTsInitialization();
6251 :
6252 : /*
6253 : * Update logical decoding status in shared memory and write an
6254 : * XLOG_LOGICAL_DECODING_STATUS_CHANGE, if necessary.
6255 : */
6256 1856 : UpdateLogicalDecodingStatusEndOfRecovery();
6257 :
6258 : /* Clean up EndOfWalRecoveryInfo data to appease Valgrind leak checking */
6259 1856 : if (endOfRecoveryInfo->lastPage)
6260 1822 : pfree(endOfRecoveryInfo->lastPage);
6261 1856 : pfree(endOfRecoveryInfo->recoveryStopReason);
6262 1856 : pfree(endOfRecoveryInfo);
6263 :
6264 : /*
6265 : * All done with end-of-recovery actions.
6266 : *
6267 : * Now allow backends to write WAL and update the control file status in
6268 : * consequence. SharedRecoveryState, that controls if backends can write
6269 : * WAL, is updated while holding ControlFileLock to prevent other backends
6270 : * to look at an inconsistent state of the control file in shared memory.
6271 : * There is still a small window during which backends can write WAL and
6272 : * the control file is still referring to a system not in DB_IN_PRODUCTION
6273 : * state while looking at the on-disk control file.
6274 : *
6275 : * Also, we use info_lck to update SharedRecoveryState to ensure that
6276 : * there are no race conditions concerning visibility of other recent
6277 : * updates to shared memory.
6278 : */
6279 1856 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6280 1856 : ControlFile->state = DB_IN_PRODUCTION;
6281 :
6282 1856 : SpinLockAcquire(&XLogCtl->info_lck);
6283 1856 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_DONE;
6284 1856 : SpinLockRelease(&XLogCtl->info_lck);
6285 :
6286 1856 : UpdateControlFile();
6287 1856 : LWLockRelease(ControlFileLock);
6288 :
6289 : /*
6290 : * Wake up the checkpointer process as there might be a request to disable
6291 : * logical decoding by concurrent slot drop.
6292 : */
6293 1856 : WakeupCheckpointer();
6294 :
6295 : /*
6296 : * Wake up all waiters. They need to report an error that recovery was
6297 : * ended before reaching the target LSN.
6298 : */
6299 1856 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_REPLAY, InvalidXLogRecPtr);
6300 1856 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_WRITE, InvalidXLogRecPtr);
6301 1856 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_FLUSH, InvalidXLogRecPtr);
6302 :
6303 : /*
6304 : * Shutdown the recovery environment. This must occur after
6305 : * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
6306 : * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
6307 : * any session building a snapshot will not rely on KnownAssignedXids as
6308 : * RecoveryInProgress() would return false at this stage. This is
6309 : * particularly critical for prepared 2PC transactions, that would still
6310 : * need to be included in snapshots once recovery has ended.
6311 : */
6312 1856 : if (standbyState != STANDBY_DISABLED)
6313 110 : ShutdownRecoveryTransactionEnvironment();
6314 :
6315 : /*
6316 : * If there were cascading standby servers connected to us, nudge any wal
6317 : * sender processes to notice that we've been promoted.
6318 : */
6319 1856 : WalSndWakeup(true, true);
6320 :
6321 : /*
6322 : * If this was a promotion, request an (online) checkpoint now. This isn't
6323 : * required for consistency, but the last restartpoint might be far back,
6324 : * and in case of a crash, recovering from it might take a longer than is
6325 : * appropriate now that we're not in standby mode anymore.
6326 : */
6327 1856 : if (promoted)
6328 96 : RequestCheckpoint(CHECKPOINT_FORCE);
6329 1856 : }
6330 :
6331 : /*
6332 : * Callback from PerformWalRecovery(), called when we switch from crash
6333 : * recovery to archive recovery mode. Updates the control file accordingly.
6334 : */
6335 : void
6336 4 : SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr, TimeLineID replayTLI)
6337 : {
6338 : /* initialize minRecoveryPoint to this record */
6339 4 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6340 4 : ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6341 4 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6342 : {
6343 4 : ControlFile->minRecoveryPoint = EndRecPtr;
6344 4 : ControlFile->minRecoveryPointTLI = replayTLI;
6345 : }
6346 : /* update local copy */
6347 4 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6348 4 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6349 :
6350 : /*
6351 : * The startup process can update its local copy of minRecoveryPoint from
6352 : * this point.
6353 : */
6354 4 : updateMinRecoveryPoint = true;
6355 :
6356 4 : UpdateControlFile();
6357 :
6358 : /*
6359 : * We update SharedRecoveryState while holding the lock on ControlFileLock
6360 : * so both states are consistent in shared memory.
6361 : */
6362 4 : SpinLockAcquire(&XLogCtl->info_lck);
6363 4 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6364 4 : SpinLockRelease(&XLogCtl->info_lck);
6365 :
6366 4 : LWLockRelease(ControlFileLock);
6367 4 : }
6368 :
6369 : /*
6370 : * Callback from PerformWalRecovery(), called when we reach the end of backup.
6371 : * Updates the control file accordingly.
6372 : */
6373 : void
6374 156 : ReachedEndOfBackup(XLogRecPtr EndRecPtr, TimeLineID tli)
6375 : {
6376 : /*
6377 : * We have reached the end of base backup, as indicated by pg_control. The
6378 : * data on disk is now consistent (unless minRecoveryPoint is further
6379 : * ahead, which can happen if we crashed during previous recovery). Reset
6380 : * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
6381 : * make sure we don't allow starting up at an earlier point even if
6382 : * recovery is stopped and restarted soon after this.
6383 : */
6384 156 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6385 :
6386 156 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6387 : {
6388 146 : ControlFile->minRecoveryPoint = EndRecPtr;
6389 146 : ControlFile->minRecoveryPointTLI = tli;
6390 : }
6391 :
6392 156 : ControlFile->backupStartPoint = InvalidXLogRecPtr;
6393 156 : ControlFile->backupEndPoint = InvalidXLogRecPtr;
6394 156 : ControlFile->backupEndRequired = false;
6395 156 : UpdateControlFile();
6396 :
6397 156 : LWLockRelease(ControlFileLock);
6398 156 : }
6399 :
6400 : /*
6401 : * Perform whatever XLOG actions are necessary at end of REDO.
6402 : *
6403 : * The goal here is to make sure that we'll be able to recover properly if
6404 : * we crash again. If we choose to write a checkpoint, we'll write a shutdown
6405 : * checkpoint rather than an on-line one. This is not particularly critical,
6406 : * but since we may be assigning a new TLI, using a shutdown checkpoint allows
6407 : * us to have the rule that TLI only changes in shutdown checkpoints, which
6408 : * allows some extra error checking in xlog_redo.
6409 : */
6410 : static bool
6411 324 : PerformRecoveryXLogAction(void)
6412 : {
6413 324 : bool promoted = false;
6414 :
6415 : /*
6416 : * Perform a checkpoint to update all our recovery activity to disk.
6417 : *
6418 : * Note that we write a shutdown checkpoint rather than an on-line one.
6419 : * This is not particularly critical, but since we may be assigning a new
6420 : * TLI, using a shutdown checkpoint allows us to have the rule that TLI
6421 : * only changes in shutdown checkpoints, which allows some extra error
6422 : * checking in xlog_redo.
6423 : *
6424 : * In promotion, only create a lightweight end-of-recovery record instead
6425 : * of a full checkpoint. A checkpoint is requested later, after we're
6426 : * fully out of recovery mode and already accepting queries.
6427 : */
6428 434 : if (ArchiveRecoveryRequested && IsUnderPostmaster &&
6429 110 : PromoteIsTriggered())
6430 : {
6431 96 : promoted = true;
6432 :
6433 : /*
6434 : * Insert a special WAL record to mark the end of recovery, since we
6435 : * aren't doing a checkpoint. That means that the checkpointer process
6436 : * may likely be in the middle of a time-smoothed restartpoint and
6437 : * could continue to be for minutes after this. That sounds strange,
6438 : * but the effect is roughly the same and it would be stranger to try
6439 : * to come out of the restartpoint and then checkpoint. We request a
6440 : * checkpoint later anyway, just for safety.
6441 : */
6442 96 : CreateEndOfRecoveryRecord();
6443 : }
6444 : else
6445 : {
6446 228 : RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
6447 : CHECKPOINT_FAST |
6448 : CHECKPOINT_WAIT);
6449 : }
6450 :
6451 324 : return promoted;
6452 : }
6453 :
6454 : /*
6455 : * Is the system still in recovery?
6456 : *
6457 : * Unlike testing InRecovery, this works in any process that's connected to
6458 : * shared memory.
6459 : */
6460 : bool
6461 161018146 : RecoveryInProgress(void)
6462 : {
6463 : /*
6464 : * We check shared state each time only until we leave recovery mode. We
6465 : * can't re-enter recovery, so there's no need to keep checking after the
6466 : * shared variable has once been seen false.
6467 : */
6468 161018146 : if (!LocalRecoveryInProgress)
6469 156441786 : return false;
6470 : else
6471 : {
6472 : /*
6473 : * use volatile pointer to make sure we make a fresh read of the
6474 : * shared variable.
6475 : */
6476 4576360 : volatile XLogCtlData *xlogctl = XLogCtl;
6477 :
6478 4576360 : LocalRecoveryInProgress = (xlogctl->SharedRecoveryState != RECOVERY_STATE_DONE);
6479 :
6480 : /*
6481 : * Note: We don't need a memory barrier when we're still in recovery.
6482 : * We might exit recovery immediately after return, so the caller
6483 : * can't rely on 'true' meaning that we're still in recovery anyway.
6484 : */
6485 :
6486 4576360 : return LocalRecoveryInProgress;
6487 : }
6488 : }
6489 :
6490 : /*
6491 : * Returns current recovery state from shared memory.
6492 : *
6493 : * This returned state is kept consistent with the contents of the control
6494 : * file. See details about the possible values of RecoveryState in xlog.h.
6495 : */
6496 : RecoveryState
6497 6426 : GetRecoveryState(void)
6498 : {
6499 : RecoveryState retval;
6500 :
6501 6426 : SpinLockAcquire(&XLogCtl->info_lck);
6502 6426 : retval = XLogCtl->SharedRecoveryState;
6503 6426 : SpinLockRelease(&XLogCtl->info_lck);
6504 :
6505 6426 : return retval;
6506 : }
6507 :
6508 : /*
6509 : * Is this process allowed to insert new WAL records?
6510 : *
6511 : * Ordinarily this is essentially equivalent to !RecoveryInProgress().
6512 : * But we also have provisions for forcing the result "true" or "false"
6513 : * within specific processes regardless of the global state.
6514 : */
6515 : bool
6516 81945318 : XLogInsertAllowed(void)
6517 : {
6518 : /*
6519 : * If value is "unconditionally true" or "unconditionally false", just
6520 : * return it. This provides the normal fast path once recovery is known
6521 : * done.
6522 : */
6523 81945318 : if (LocalXLogInsertAllowed >= 0)
6524 80443884 : return (bool) LocalXLogInsertAllowed;
6525 :
6526 : /*
6527 : * Else, must check to see if we're still in recovery.
6528 : */
6529 1501434 : if (RecoveryInProgress())
6530 1481872 : return false;
6531 :
6532 : /*
6533 : * On exit from recovery, reset to "unconditionally true", since there is
6534 : * no need to keep checking.
6535 : */
6536 19562 : LocalXLogInsertAllowed = 1;
6537 19562 : return true;
6538 : }
6539 :
6540 : /*
6541 : * Make XLogInsertAllowed() return true in the current process only.
6542 : *
6543 : * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
6544 : * and even call LocalSetXLogInsertAllowed() again after that.
6545 : *
6546 : * Returns the previous value of LocalXLogInsertAllowed.
6547 : */
6548 : static int
6549 1912 : LocalSetXLogInsertAllowed(void)
6550 : {
6551 1912 : int oldXLogAllowed = LocalXLogInsertAllowed;
6552 :
6553 1912 : LocalXLogInsertAllowed = 1;
6554 :
6555 1912 : return oldXLogAllowed;
6556 : }
6557 :
6558 : /*
6559 : * Return the current Redo pointer from shared memory.
6560 : *
6561 : * As a side-effect, the local RedoRecPtr copy is updated.
6562 : */
6563 : XLogRecPtr
6564 602270 : GetRedoRecPtr(void)
6565 : {
6566 : XLogRecPtr ptr;
6567 :
6568 : /*
6569 : * The possibly not up-to-date copy in XlogCtl is enough. Even if we
6570 : * grabbed a WAL insertion lock to read the authoritative value in
6571 : * Insert->RedoRecPtr, someone might update it just after we've released
6572 : * the lock.
6573 : */
6574 602270 : SpinLockAcquire(&XLogCtl->info_lck);
6575 602270 : ptr = XLogCtl->RedoRecPtr;
6576 602270 : SpinLockRelease(&XLogCtl->info_lck);
6577 :
6578 602270 : if (RedoRecPtr < ptr)
6579 3194 : RedoRecPtr = ptr;
6580 :
6581 602270 : return RedoRecPtr;
6582 : }
6583 :
6584 : /*
6585 : * Return information needed to decide whether a modified block needs a
6586 : * full-page image to be included in the WAL record.
6587 : *
6588 : * The returned values are cached copies from backend-private memory, and
6589 : * possibly out-of-date or, indeed, uninitialized, in which case they will
6590 : * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6591 : * re-check them against up-to-date values, while holding the WAL insert lock.
6592 : */
6593 : void
6594 31068090 : GetFullPageWriteInfo(XLogRecPtr *RedoRecPtr_p, bool *doPageWrites_p)
6595 : {
6596 31068090 : *RedoRecPtr_p = RedoRecPtr;
6597 31068090 : *doPageWrites_p = doPageWrites;
6598 31068090 : }
6599 :
6600 : /*
6601 : * GetInsertRecPtr -- Returns the current insert position.
6602 : *
6603 : * NOTE: The value *actually* returned is the position of the last full
6604 : * xlog page. It lags behind the real insert position by at most 1 page.
6605 : * For that, we don't need to scan through WAL insertion locks, and an
6606 : * approximation is enough for the current usage of this function.
6607 : */
6608 : XLogRecPtr
6609 13984 : GetInsertRecPtr(void)
6610 : {
6611 : XLogRecPtr recptr;
6612 :
6613 13984 : SpinLockAcquire(&XLogCtl->info_lck);
6614 13984 : recptr = XLogCtl->LogwrtRqst.Write;
6615 13984 : SpinLockRelease(&XLogCtl->info_lck);
6616 :
6617 13984 : return recptr;
6618 : }
6619 :
6620 : /*
6621 : * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6622 : * position known to be fsync'd to disk. This should only be used on a
6623 : * system that is known not to be in recovery.
6624 : */
6625 : XLogRecPtr
6626 441762 : GetFlushRecPtr(TimeLineID *insertTLI)
6627 : {
6628 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6629 :
6630 441762 : RefreshXLogWriteResult(LogwrtResult);
6631 :
6632 : /*
6633 : * If we're writing and flushing WAL, the time line can't be changing, so
6634 : * no lock is required.
6635 : */
6636 441762 : if (insertTLI)
6637 51404 : *insertTLI = XLogCtl->InsertTimeLineID;
6638 :
6639 441762 : return LogwrtResult.Flush;
6640 : }
6641 :
6642 : /*
6643 : * GetWALInsertionTimeLine -- Returns the current timeline of a system that
6644 : * is not in recovery.
6645 : */
6646 : TimeLineID
6647 214058 : GetWALInsertionTimeLine(void)
6648 : {
6649 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6650 :
6651 : /* Since the value can't be changing, no lock is required. */
6652 214058 : return XLogCtl->InsertTimeLineID;
6653 : }
6654 :
6655 : /*
6656 : * GetWALInsertionTimeLineIfSet -- If the system is not in recovery, returns
6657 : * the WAL insertion timeline; else, returns 0. Wherever possible, use
6658 : * GetWALInsertionTimeLine() instead, since it's cheaper. Note that this
6659 : * function decides recovery has ended as soon as the insert TLI is set, which
6660 : * happens before we set XLogCtl->SharedRecoveryState to RECOVERY_STATE_DONE.
6661 : */
6662 : TimeLineID
6663 0 : GetWALInsertionTimeLineIfSet(void)
6664 : {
6665 : TimeLineID insertTLI;
6666 :
6667 0 : SpinLockAcquire(&XLogCtl->info_lck);
6668 0 : insertTLI = XLogCtl->InsertTimeLineID;
6669 0 : SpinLockRelease(&XLogCtl->info_lck);
6670 :
6671 0 : return insertTLI;
6672 : }
6673 :
6674 : /*
6675 : * GetLastImportantRecPtr -- Returns the LSN of the last important record
6676 : * inserted. All records not explicitly marked as unimportant are considered
6677 : * important.
6678 : *
6679 : * The LSN is determined by computing the maximum of
6680 : * WALInsertLocks[i].lastImportantAt.
6681 : */
6682 : XLogRecPtr
6683 3242 : GetLastImportantRecPtr(void)
6684 : {
6685 3242 : XLogRecPtr res = InvalidXLogRecPtr;
6686 : int i;
6687 :
6688 29178 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
6689 : {
6690 : XLogRecPtr last_important;
6691 :
6692 : /*
6693 : * Need to take a lock to prevent torn reads of the LSN, which are
6694 : * possible on some of the supported platforms. WAL insert locks only
6695 : * support exclusive mode, so we have to use that.
6696 : */
6697 25936 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
6698 25936 : last_important = WALInsertLocks[i].l.lastImportantAt;
6699 25936 : LWLockRelease(&WALInsertLocks[i].l.lock);
6700 :
6701 25936 : if (res < last_important)
6702 5604 : res = last_important;
6703 : }
6704 :
6705 3242 : return res;
6706 : }
6707 :
6708 : /*
6709 : * Get the time and LSN of the last xlog segment switch
6710 : */
6711 : pg_time_t
6712 0 : GetLastSegSwitchData(XLogRecPtr *lastSwitchLSN)
6713 : {
6714 : pg_time_t result;
6715 :
6716 : /* Need WALWriteLock, but shared lock is sufficient */
6717 0 : LWLockAcquire(WALWriteLock, LW_SHARED);
6718 0 : result = XLogCtl->lastSegSwitchTime;
6719 0 : *lastSwitchLSN = XLogCtl->lastSegSwitchLSN;
6720 0 : LWLockRelease(WALWriteLock);
6721 :
6722 0 : return result;
6723 : }
6724 :
6725 : /*
6726 : * This must be called ONCE during postmaster or standalone-backend shutdown
6727 : */
6728 : void
6729 1364 : ShutdownXLOG(int code, Datum arg)
6730 : {
6731 : /*
6732 : * We should have an aux process resource owner to use, and we should not
6733 : * be in a transaction that's installed some other resowner.
6734 : */
6735 : Assert(AuxProcessResourceOwner != NULL);
6736 : Assert(CurrentResourceOwner == NULL ||
6737 : CurrentResourceOwner == AuxProcessResourceOwner);
6738 1364 : CurrentResourceOwner = AuxProcessResourceOwner;
6739 :
6740 : /* Don't be chatty in standalone mode */
6741 1364 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
6742 : (errmsg("shutting down")));
6743 :
6744 : /*
6745 : * Signal walsenders to move to stopping state.
6746 : */
6747 1364 : WalSndInitStopping();
6748 :
6749 : /*
6750 : * Wait for WAL senders to be in stopping state. This prevents commands
6751 : * from writing new WAL.
6752 : */
6753 1364 : WalSndWaitStopping();
6754 :
6755 1364 : if (RecoveryInProgress())
6756 114 : CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
6757 : else
6758 : {
6759 : /*
6760 : * If archiving is enabled, rotate the last XLOG file so that all the
6761 : * remaining records are archived (postmaster wakes up the archiver
6762 : * process one more time at the end of shutdown). The checkpoint
6763 : * record will go to the next XLOG file and won't be archived (yet).
6764 : */
6765 1250 : if (XLogArchivingActive())
6766 28 : RequestXLogSwitch(false);
6767 :
6768 1250 : CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
6769 : }
6770 1364 : }
6771 :
6772 : /*
6773 : * Log start of a checkpoint.
6774 : */
6775 : static void
6776 2954 : LogCheckpointStart(int flags, bool restartpoint)
6777 : {
6778 2954 : if (restartpoint)
6779 392 : ereport(LOG,
6780 : /* translator: the placeholders show checkpoint options */
6781 : (errmsg("restartpoint starting:%s%s%s%s%s%s%s%s",
6782 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6783 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6784 : (flags & CHECKPOINT_FAST) ? " fast" : "",
6785 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6786 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6787 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6788 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6789 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "")));
6790 : else
6791 2562 : ereport(LOG,
6792 : /* translator: the placeholders show checkpoint options */
6793 : (errmsg("checkpoint starting:%s%s%s%s%s%s%s%s",
6794 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6795 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6796 : (flags & CHECKPOINT_FAST) ? " fast" : "",
6797 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6798 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6799 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6800 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6801 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "")));
6802 2954 : }
6803 :
6804 : /*
6805 : * Log end of a checkpoint.
6806 : */
6807 : static void
6808 3550 : LogCheckpointEnd(bool restartpoint)
6809 : {
6810 : long write_msecs,
6811 : sync_msecs,
6812 : total_msecs,
6813 : longest_msecs,
6814 : average_msecs;
6815 : uint64 average_sync_time;
6816 :
6817 3550 : CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
6818 :
6819 3550 : write_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_write_t,
6820 : CheckpointStats.ckpt_sync_t);
6821 :
6822 3550 : sync_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_sync_t,
6823 : CheckpointStats.ckpt_sync_end_t);
6824 :
6825 : /* Accumulate checkpoint timing summary data, in milliseconds. */
6826 3550 : PendingCheckpointerStats.write_time += write_msecs;
6827 3550 : PendingCheckpointerStats.sync_time += sync_msecs;
6828 :
6829 : /*
6830 : * All of the published timing statistics are accounted for. Only
6831 : * continue if a log message is to be written.
6832 : */
6833 3550 : if (!log_checkpoints)
6834 596 : return;
6835 :
6836 2954 : total_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_start_t,
6837 : CheckpointStats.ckpt_end_t);
6838 :
6839 : /*
6840 : * Timing values returned from CheckpointStats are in microseconds.
6841 : * Convert to milliseconds for consistent printing.
6842 : */
6843 2954 : longest_msecs = (long) ((CheckpointStats.ckpt_longest_sync + 999) / 1000);
6844 :
6845 2954 : average_sync_time = 0;
6846 2954 : if (CheckpointStats.ckpt_sync_rels > 0)
6847 0 : average_sync_time = CheckpointStats.ckpt_agg_sync_time /
6848 0 : CheckpointStats.ckpt_sync_rels;
6849 2954 : average_msecs = (long) ((average_sync_time + 999) / 1000);
6850 :
6851 : /*
6852 : * ControlFileLock is not required to see ControlFile->checkPoint and
6853 : * ->checkPointCopy here as we are the only updator of those variables at
6854 : * this moment.
6855 : */
6856 2954 : if (restartpoint)
6857 392 : ereport(LOG,
6858 : (errmsg("restartpoint complete: wrote %d buffers (%.1f%%), "
6859 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6860 : "%d removed, %d recycled; write=%ld.%03d s, "
6861 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6862 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6863 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6864 : CheckpointStats.ckpt_bufs_written,
6865 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6866 : CheckpointStats.ckpt_slru_written,
6867 : CheckpointStats.ckpt_segs_added,
6868 : CheckpointStats.ckpt_segs_removed,
6869 : CheckpointStats.ckpt_segs_recycled,
6870 : write_msecs / 1000, (int) (write_msecs % 1000),
6871 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6872 : total_msecs / 1000, (int) (total_msecs % 1000),
6873 : CheckpointStats.ckpt_sync_rels,
6874 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6875 : average_msecs / 1000, (int) (average_msecs % 1000),
6876 : (int) (PrevCheckPointDistance / 1024.0),
6877 : (int) (CheckPointDistanceEstimate / 1024.0),
6878 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6879 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6880 : else
6881 2562 : ereport(LOG,
6882 : (errmsg("checkpoint complete: wrote %d buffers (%.1f%%), "
6883 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6884 : "%d removed, %d recycled; write=%ld.%03d s, "
6885 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6886 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6887 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6888 : CheckpointStats.ckpt_bufs_written,
6889 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6890 : CheckpointStats.ckpt_slru_written,
6891 : CheckpointStats.ckpt_segs_added,
6892 : CheckpointStats.ckpt_segs_removed,
6893 : CheckpointStats.ckpt_segs_recycled,
6894 : write_msecs / 1000, (int) (write_msecs % 1000),
6895 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6896 : total_msecs / 1000, (int) (total_msecs % 1000),
6897 : CheckpointStats.ckpt_sync_rels,
6898 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6899 : average_msecs / 1000, (int) (average_msecs % 1000),
6900 : (int) (PrevCheckPointDistance / 1024.0),
6901 : (int) (CheckPointDistanceEstimate / 1024.0),
6902 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6903 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6904 : }
6905 :
6906 : /*
6907 : * Update the estimate of distance between checkpoints.
6908 : *
6909 : * The estimate is used to calculate the number of WAL segments to keep
6910 : * preallocated, see XLOGfileslop().
6911 : */
6912 : static void
6913 3550 : UpdateCheckPointDistanceEstimate(uint64 nbytes)
6914 : {
6915 : /*
6916 : * To estimate the number of segments consumed between checkpoints, keep a
6917 : * moving average of the amount of WAL generated in previous checkpoint
6918 : * cycles. However, if the load is bursty, with quiet periods and busy
6919 : * periods, we want to cater for the peak load. So instead of a plain
6920 : * moving average, let the average decline slowly if the previous cycle
6921 : * used less WAL than estimated, but bump it up immediately if it used
6922 : * more.
6923 : *
6924 : * When checkpoints are triggered by max_wal_size, this should converge to
6925 : * CheckpointSegments * wal_segment_size,
6926 : *
6927 : * Note: This doesn't pay any attention to what caused the checkpoint.
6928 : * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
6929 : * starting a base backup, are counted the same as those created
6930 : * automatically. The slow-decline will largely mask them out, if they are
6931 : * not frequent. If they are frequent, it seems reasonable to count them
6932 : * in as any others; if you issue a manual checkpoint every 5 minutes and
6933 : * never let a timed checkpoint happen, it makes sense to base the
6934 : * preallocation on that 5 minute interval rather than whatever
6935 : * checkpoint_timeout is set to.
6936 : */
6937 3550 : PrevCheckPointDistance = nbytes;
6938 3550 : if (CheckPointDistanceEstimate < nbytes)
6939 1526 : CheckPointDistanceEstimate = nbytes;
6940 : else
6941 2024 : CheckPointDistanceEstimate =
6942 2024 : (0.90 * CheckPointDistanceEstimate + 0.10 * (double) nbytes);
6943 3550 : }
6944 :
6945 : /*
6946 : * Update the ps display for a process running a checkpoint. Note that
6947 : * this routine should not do any allocations so as it can be called
6948 : * from a critical section.
6949 : */
6950 : static void
6951 7100 : update_checkpoint_display(int flags, bool restartpoint, bool reset)
6952 : {
6953 : /*
6954 : * The status is reported only for end-of-recovery and shutdown
6955 : * checkpoints or shutdown restartpoints. Updating the ps display is
6956 : * useful in those situations as it may not be possible to rely on
6957 : * pg_stat_activity to see the status of the checkpointer or the startup
6958 : * process.
6959 : */
6960 7100 : if ((flags & (CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IS_SHUTDOWN)) == 0)
6961 4400 : return;
6962 :
6963 2700 : if (reset)
6964 1350 : set_ps_display("");
6965 : else
6966 : {
6967 : char activitymsg[128];
6968 :
6969 4050 : snprintf(activitymsg, sizeof(activitymsg), "performing %s%s%s",
6970 1350 : (flags & CHECKPOINT_END_OF_RECOVERY) ? "end-of-recovery " : "",
6971 1350 : (flags & CHECKPOINT_IS_SHUTDOWN) ? "shutdown " : "",
6972 : restartpoint ? "restartpoint" : "checkpoint");
6973 1350 : set_ps_display(activitymsg);
6974 : }
6975 : }
6976 :
6977 :
6978 : /*
6979 : * Perform a checkpoint --- either during shutdown, or on-the-fly
6980 : *
6981 : * flags is a bitwise OR of the following:
6982 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
6983 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
6984 : * CHECKPOINT_FAST: finish the checkpoint ASAP, ignoring
6985 : * checkpoint_completion_target parameter.
6986 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
6987 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
6988 : * CHECKPOINT_END_OF_RECOVERY).
6989 : * CHECKPOINT_FLUSH_UNLOGGED: also flush buffers of unlogged tables.
6990 : *
6991 : * Note: flags contains other bits, of interest here only for logging purposes.
6992 : * In particular note that this routine is synchronous and does not pay
6993 : * attention to CHECKPOINT_WAIT.
6994 : *
6995 : * If !shutdown then we are writing an online checkpoint. An XLOG_CHECKPOINT_REDO
6996 : * record is inserted into WAL at the logical location of the checkpoint, before
6997 : * flushing anything to disk, and when the checkpoint is eventually completed,
6998 : * and it is from this point that WAL replay will begin in the case of a recovery
6999 : * from this checkpoint. Once everything is written to disk, an
7000 : * XLOG_CHECKPOINT_ONLINE record is written to complete the checkpoint, and
7001 : * points back to the earlier XLOG_CHECKPOINT_REDO record. This mechanism allows
7002 : * other write-ahead log records to be written while the checkpoint is in
7003 : * progress, but we must be very careful about order of operations. This function
7004 : * may take many minutes to execute on a busy system.
7005 : *
7006 : * On the other hand, when shutdown is true, concurrent insertion into the
7007 : * write-ahead log is impossible, so there is no need for two separate records.
7008 : * In this case, we only insert an XLOG_CHECKPOINT_SHUTDOWN record, and it's
7009 : * both the record marking the completion of the checkpoint and the location
7010 : * from which WAL replay would begin if needed.
7011 : *
7012 : * Returns true if a new checkpoint was performed, or false if it was skipped
7013 : * because the system was idle.
7014 : */
7015 : bool
7016 3160 : CreateCheckPoint(int flags)
7017 : {
7018 : bool shutdown;
7019 : CheckPoint checkPoint;
7020 : XLogRecPtr recptr;
7021 : XLogSegNo _logSegNo;
7022 3160 : XLogCtlInsert *Insert = &XLogCtl->Insert;
7023 : uint32 freespace;
7024 : XLogRecPtr PriorRedoPtr;
7025 : XLogRecPtr last_important_lsn;
7026 : VirtualTransactionId *vxids;
7027 : int nvxids;
7028 3160 : int oldXLogAllowed = 0;
7029 :
7030 : /*
7031 : * An end-of-recovery checkpoint is really a shutdown checkpoint, just
7032 : * issued at a different time.
7033 : */
7034 3160 : if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
7035 1306 : shutdown = true;
7036 : else
7037 1854 : shutdown = false;
7038 :
7039 : /* sanity check */
7040 3160 : if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
7041 0 : elog(ERROR, "can't create a checkpoint during recovery");
7042 :
7043 : /*
7044 : * Prepare to accumulate statistics.
7045 : *
7046 : * Note: because it is possible for log_checkpoints to change while a
7047 : * checkpoint proceeds, we always accumulate stats, even if
7048 : * log_checkpoints is currently off.
7049 : */
7050 34760 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7051 3160 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7052 :
7053 : /*
7054 : * Let smgr prepare for checkpoint; this has to happen outside the
7055 : * critical section and before we determine the REDO pointer. Note that
7056 : * smgr must not do anything that'd have to be undone if we decide no
7057 : * checkpoint is needed.
7058 : */
7059 3160 : SyncPreCheckpoint();
7060 :
7061 : /* Run these points outside the critical section. */
7062 3160 : INJECTION_POINT("create-checkpoint-initial", NULL);
7063 3160 : INJECTION_POINT_LOAD("create-checkpoint-run");
7064 :
7065 : /*
7066 : * Use a critical section to force system panic if we have trouble.
7067 : */
7068 3160 : START_CRIT_SECTION();
7069 :
7070 3160 : if (shutdown)
7071 : {
7072 1306 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7073 1306 : ControlFile->state = DB_SHUTDOWNING;
7074 1306 : UpdateControlFile();
7075 1306 : LWLockRelease(ControlFileLock);
7076 : }
7077 :
7078 : /* Begin filling in the checkpoint WAL record */
7079 41080 : MemSet(&checkPoint, 0, sizeof(checkPoint));
7080 3160 : checkPoint.time = (pg_time_t) time(NULL);
7081 :
7082 : /*
7083 : * For Hot Standby, derive the oldestActiveXid before we fix the redo
7084 : * pointer. This allows us to begin accumulating changes to assemble our
7085 : * starting snapshot of locks and transactions.
7086 : */
7087 3160 : if (!shutdown && XLogStandbyInfoActive())
7088 1798 : checkPoint.oldestActiveXid = GetOldestActiveTransactionId(false, true);
7089 : else
7090 1362 : checkPoint.oldestActiveXid = InvalidTransactionId;
7091 :
7092 : /*
7093 : * Get location of last important record before acquiring insert locks (as
7094 : * GetLastImportantRecPtr() also locks WAL locks).
7095 : */
7096 3160 : last_important_lsn = GetLastImportantRecPtr();
7097 :
7098 : /*
7099 : * If this isn't a shutdown or forced checkpoint, and if there has been no
7100 : * WAL activity requiring a checkpoint, skip it. The idea here is to
7101 : * avoid inserting duplicate checkpoints when the system is idle.
7102 : */
7103 3160 : if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
7104 : CHECKPOINT_FORCE)) == 0)
7105 : {
7106 386 : if (last_important_lsn == ControlFile->checkPoint)
7107 : {
7108 2 : END_CRIT_SECTION();
7109 2 : ereport(DEBUG1,
7110 : (errmsg_internal("checkpoint skipped because system is idle")));
7111 2 : return false;
7112 : }
7113 : }
7114 :
7115 : /*
7116 : * An end-of-recovery checkpoint is created before anyone is allowed to
7117 : * write WAL. To allow us to write the checkpoint record, temporarily
7118 : * enable XLogInsertAllowed.
7119 : */
7120 3158 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7121 56 : oldXLogAllowed = LocalSetXLogInsertAllowed();
7122 :
7123 3158 : checkPoint.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7124 3158 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7125 56 : checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7126 : else
7127 3102 : checkPoint.PrevTimeLineID = checkPoint.ThisTimeLineID;
7128 :
7129 : /*
7130 : * We must block concurrent insertions while examining insert state.
7131 : */
7132 3158 : WALInsertLockAcquireExclusive();
7133 :
7134 3158 : checkPoint.fullPageWrites = Insert->fullPageWrites;
7135 3158 : checkPoint.wal_level = wal_level;
7136 :
7137 3158 : if (shutdown)
7138 : {
7139 1306 : XLogRecPtr curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
7140 :
7141 : /*
7142 : * Compute new REDO record ptr = location of next XLOG record.
7143 : *
7144 : * Since this is a shutdown checkpoint, there can't be any concurrent
7145 : * WAL insertion.
7146 : */
7147 1306 : freespace = INSERT_FREESPACE(curInsert);
7148 1306 : if (freespace == 0)
7149 : {
7150 0 : if (XLogSegmentOffset(curInsert, wal_segment_size) == 0)
7151 0 : curInsert += SizeOfXLogLongPHD;
7152 : else
7153 0 : curInsert += SizeOfXLogShortPHD;
7154 : }
7155 1306 : checkPoint.redo = curInsert;
7156 :
7157 : /*
7158 : * Here we update the shared RedoRecPtr for future XLogInsert calls;
7159 : * this must be done while holding all the insertion locks.
7160 : *
7161 : * Note: if we fail to complete the checkpoint, RedoRecPtr will be
7162 : * left pointing past where it really needs to point. This is okay;
7163 : * the only consequence is that XLogInsert might back up whole buffers
7164 : * that it didn't really need to. We can't postpone advancing
7165 : * RedoRecPtr because XLogInserts that happen while we are dumping
7166 : * buffers must assume that their buffer changes are not included in
7167 : * the checkpoint.
7168 : */
7169 1306 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
7170 : }
7171 :
7172 : /*
7173 : * Now we can release the WAL insertion locks, allowing other xacts to
7174 : * proceed while we are flushing disk buffers.
7175 : */
7176 3158 : WALInsertLockRelease();
7177 :
7178 : /*
7179 : * If this is an online checkpoint, we have not yet determined the redo
7180 : * point. We do so now by inserting the special XLOG_CHECKPOINT_REDO
7181 : * record; the LSN at which it starts becomes the new redo pointer. We
7182 : * don't do this for a shutdown checkpoint, because in that case no WAL
7183 : * can be written between the redo point and the insertion of the
7184 : * checkpoint record itself, so the checkpoint record itself serves to
7185 : * mark the redo point.
7186 : */
7187 3158 : if (!shutdown)
7188 : {
7189 : /* Include WAL level in record for WAL summarizer's benefit. */
7190 1852 : XLogBeginInsert();
7191 1852 : XLogRegisterData(&wal_level, sizeof(wal_level));
7192 1852 : (void) XLogInsert(RM_XLOG_ID, XLOG_CHECKPOINT_REDO);
7193 :
7194 : /*
7195 : * XLogInsertRecord will have updated XLogCtl->Insert.RedoRecPtr in
7196 : * shared memory and RedoRecPtr in backend-local memory, but we need
7197 : * to copy that into the record that will be inserted when the
7198 : * checkpoint is complete.
7199 : */
7200 1852 : checkPoint.redo = RedoRecPtr;
7201 : }
7202 :
7203 : /* Update the info_lck-protected copy of RedoRecPtr as well */
7204 3158 : SpinLockAcquire(&XLogCtl->info_lck);
7205 3158 : XLogCtl->RedoRecPtr = checkPoint.redo;
7206 3158 : SpinLockRelease(&XLogCtl->info_lck);
7207 :
7208 : /*
7209 : * If enabled, log checkpoint start. We postpone this until now so as not
7210 : * to log anything if we decided to skip the checkpoint.
7211 : */
7212 3158 : if (log_checkpoints)
7213 2562 : LogCheckpointStart(flags, false);
7214 :
7215 3158 : INJECTION_POINT_CACHED("create-checkpoint-run", NULL);
7216 :
7217 : /* Update the process title */
7218 3158 : update_checkpoint_display(flags, false, false);
7219 :
7220 : TRACE_POSTGRESQL_CHECKPOINT_START(flags);
7221 :
7222 : /*
7223 : * Get the other info we need for the checkpoint record.
7224 : *
7225 : * We don't need to save oldestClogXid in the checkpoint, it only matters
7226 : * for the short period in which clog is being truncated, and if we crash
7227 : * during that we'll redo the clog truncation and fix up oldestClogXid
7228 : * there.
7229 : */
7230 3158 : LWLockAcquire(XidGenLock, LW_SHARED);
7231 3158 : checkPoint.nextXid = TransamVariables->nextXid;
7232 3158 : checkPoint.oldestXid = TransamVariables->oldestXid;
7233 3158 : checkPoint.oldestXidDB = TransamVariables->oldestXidDB;
7234 3158 : LWLockRelease(XidGenLock);
7235 :
7236 3158 : LWLockAcquire(CommitTsLock, LW_SHARED);
7237 3158 : checkPoint.oldestCommitTsXid = TransamVariables->oldestCommitTsXid;
7238 3158 : checkPoint.newestCommitTsXid = TransamVariables->newestCommitTsXid;
7239 3158 : LWLockRelease(CommitTsLock);
7240 :
7241 3158 : LWLockAcquire(OidGenLock, LW_SHARED);
7242 3158 : checkPoint.nextOid = TransamVariables->nextOid;
7243 3158 : if (!shutdown)
7244 1852 : checkPoint.nextOid += TransamVariables->oidCount;
7245 3158 : LWLockRelease(OidGenLock);
7246 :
7247 3158 : checkPoint.logicalDecodingEnabled = IsLogicalDecodingEnabled();
7248 :
7249 3158 : MultiXactGetCheckptMulti(shutdown,
7250 : &checkPoint.nextMulti,
7251 : &checkPoint.nextMultiOffset,
7252 : &checkPoint.oldestMulti,
7253 : &checkPoint.oldestMultiDB);
7254 :
7255 : /*
7256 : * Having constructed the checkpoint record, ensure all shmem disk buffers
7257 : * and commit-log buffers are flushed to disk.
7258 : *
7259 : * This I/O could fail for various reasons. If so, we will fail to
7260 : * complete the checkpoint, but there is no reason to force a system
7261 : * panic. Accordingly, exit critical section while doing it.
7262 : */
7263 3158 : END_CRIT_SECTION();
7264 :
7265 : /*
7266 : * In some cases there are groups of actions that must all occur on one
7267 : * side or the other of a checkpoint record. Before flushing the
7268 : * checkpoint record we must explicitly wait for any backend currently
7269 : * performing those groups of actions.
7270 : *
7271 : * One example is end of transaction, so we must wait for any transactions
7272 : * that are currently in commit critical sections. If an xact inserted
7273 : * its commit record into XLOG just before the REDO point, then a crash
7274 : * restart from the REDO point would not replay that record, which means
7275 : * that our flushing had better include the xact's update of pg_xact. So
7276 : * we wait till he's out of his commit critical section before proceeding.
7277 : * See notes in RecordTransactionCommit().
7278 : *
7279 : * Because we've already released the insertion locks, this test is a bit
7280 : * fuzzy: it is possible that we will wait for xacts we didn't really need
7281 : * to wait for. But the delay should be short and it seems better to make
7282 : * checkpoint take a bit longer than to hold off insertions longer than
7283 : * necessary. (In fact, the whole reason we have this issue is that xact.c
7284 : * does commit record XLOG insertion and clog update as two separate steps
7285 : * protected by different locks, but again that seems best on grounds of
7286 : * minimizing lock contention.)
7287 : *
7288 : * A transaction that has not yet set delayChkptFlags when we look cannot
7289 : * be at risk, since it has not inserted its commit record yet; and one
7290 : * that's already cleared it is not at risk either, since it's done fixing
7291 : * clog and we will correctly flush the update below. So we cannot miss
7292 : * any xacts we need to wait for.
7293 : */
7294 3158 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
7295 3158 : if (nvxids > 0)
7296 : {
7297 : do
7298 : {
7299 : /*
7300 : * Keep absorbing fsync requests while we wait. There could even
7301 : * be a deadlock if we don't, if the process that prevents the
7302 : * checkpoint is trying to add a request to the queue.
7303 : */
7304 58 : AbsorbSyncRequests();
7305 :
7306 58 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
7307 58 : pg_usleep(10000L); /* wait for 10 msec */
7308 58 : pgstat_report_wait_end();
7309 58 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7310 : DELAY_CHKPT_START));
7311 : }
7312 3158 : pfree(vxids);
7313 :
7314 3158 : CheckPointGuts(checkPoint.redo, flags);
7315 :
7316 3158 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
7317 3158 : if (nvxids > 0)
7318 : {
7319 : do
7320 : {
7321 0 : AbsorbSyncRequests();
7322 :
7323 0 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_COMPLETE);
7324 0 : pg_usleep(10000L); /* wait for 10 msec */
7325 0 : pgstat_report_wait_end();
7326 0 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7327 : DELAY_CHKPT_COMPLETE));
7328 : }
7329 3158 : pfree(vxids);
7330 :
7331 : /*
7332 : * Take a snapshot of running transactions and write this to WAL. This
7333 : * allows us to reconstruct the state of running transactions during
7334 : * archive recovery, if required. Skip, if this info disabled.
7335 : *
7336 : * If we are shutting down, or Startup process is completing crash
7337 : * recovery we don't need to write running xact data.
7338 : */
7339 3158 : if (!shutdown && XLogStandbyInfoActive())
7340 1796 : LogStandbySnapshot();
7341 :
7342 3158 : START_CRIT_SECTION();
7343 :
7344 : /*
7345 : * Now insert the checkpoint record into XLOG.
7346 : */
7347 3158 : XLogBeginInsert();
7348 3158 : XLogRegisterData(&checkPoint, sizeof(checkPoint));
7349 3158 : recptr = XLogInsert(RM_XLOG_ID,
7350 : shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
7351 : XLOG_CHECKPOINT_ONLINE);
7352 :
7353 3158 : XLogFlush(recptr);
7354 :
7355 : /*
7356 : * We mustn't write any new WAL after a shutdown checkpoint, or it will be
7357 : * overwritten at next startup. No-one should even try, this just allows
7358 : * sanity-checking. In the case of an end-of-recovery checkpoint, we want
7359 : * to just temporarily disable writing until the system has exited
7360 : * recovery.
7361 : */
7362 3158 : if (shutdown)
7363 : {
7364 1306 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7365 56 : LocalXLogInsertAllowed = oldXLogAllowed;
7366 : else
7367 1250 : LocalXLogInsertAllowed = 0; /* never again write WAL */
7368 : }
7369 :
7370 : /*
7371 : * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
7372 : * = end of actual checkpoint record.
7373 : */
7374 3158 : if (shutdown && checkPoint.redo != ProcLastRecPtr)
7375 0 : ereport(PANIC,
7376 : (errmsg("concurrent write-ahead log activity while database system is shutting down")));
7377 :
7378 : /*
7379 : * Remember the prior checkpoint's redo ptr for
7380 : * UpdateCheckPointDistanceEstimate()
7381 : */
7382 3158 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7383 :
7384 : /*
7385 : * Update the control file.
7386 : */
7387 3158 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7388 3158 : if (shutdown)
7389 1306 : ControlFile->state = DB_SHUTDOWNED;
7390 3158 : ControlFile->checkPoint = ProcLastRecPtr;
7391 3158 : ControlFile->checkPointCopy = checkPoint;
7392 : /* crash recovery should always recover to the end of WAL */
7393 3158 : ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
7394 3158 : ControlFile->minRecoveryPointTLI = 0;
7395 :
7396 : /*
7397 : * Persist unloggedLSN value. It's reset on crash recovery, so this goes
7398 : * unused on non-shutdown checkpoints, but seems useful to store it always
7399 : * for debugging purposes.
7400 : */
7401 3158 : ControlFile->unloggedLSN = pg_atomic_read_membarrier_u64(&XLogCtl->unloggedLSN);
7402 :
7403 3158 : UpdateControlFile();
7404 3158 : LWLockRelease(ControlFileLock);
7405 :
7406 : /*
7407 : * We are now done with critical updates; no need for system panic if we
7408 : * have trouble while fooling with old log segments.
7409 : */
7410 3158 : END_CRIT_SECTION();
7411 :
7412 : /*
7413 : * WAL summaries end when the next XLOG_CHECKPOINT_REDO or
7414 : * XLOG_CHECKPOINT_SHUTDOWN record is reached. This is the first point
7415 : * where (a) we're not inside of a critical section and (b) we can be
7416 : * certain that the relevant record has been flushed to disk, which must
7417 : * happen before it can be summarized.
7418 : *
7419 : * If this is a shutdown checkpoint, then this happens reasonably
7420 : * promptly: we've only just inserted and flushed the
7421 : * XLOG_CHECKPOINT_SHUTDOWN record. If this is not a shutdown checkpoint,
7422 : * then this might not be very prompt at all: the XLOG_CHECKPOINT_REDO
7423 : * record was written before we began flushing data to disk, and that
7424 : * could be many minutes ago at this point. However, we don't XLogFlush()
7425 : * after inserting that record, so we're not guaranteed that it's on disk
7426 : * until after the above call that flushes the XLOG_CHECKPOINT_ONLINE
7427 : * record.
7428 : */
7429 3158 : WakeupWalSummarizer();
7430 :
7431 : /*
7432 : * Let smgr do post-checkpoint cleanup (eg, deleting old files).
7433 : */
7434 3158 : SyncPostCheckpoint();
7435 :
7436 : /*
7437 : * Update the average distance between checkpoints if the prior checkpoint
7438 : * exists.
7439 : */
7440 3158 : if (XLogRecPtrIsValid(PriorRedoPtr))
7441 3158 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7442 :
7443 3158 : INJECTION_POINT("checkpoint-before-old-wal-removal", NULL);
7444 :
7445 : /*
7446 : * Delete old log files, those no longer needed for last checkpoint to
7447 : * prevent the disk holding the xlog from growing full.
7448 : */
7449 3158 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7450 3158 : KeepLogSeg(recptr, &_logSegNo);
7451 3158 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7452 : _logSegNo, InvalidOid,
7453 : InvalidTransactionId))
7454 : {
7455 : /*
7456 : * Some slots have been invalidated; recalculate the old-segment
7457 : * horizon, starting again from RedoRecPtr.
7458 : */
7459 8 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7460 8 : KeepLogSeg(recptr, &_logSegNo);
7461 : }
7462 3158 : _logSegNo--;
7463 3158 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
7464 : checkPoint.ThisTimeLineID);
7465 :
7466 : /*
7467 : * Make more log segments if needed. (Do this after recycling old log
7468 : * segments, since that may supply some of the needed files.)
7469 : */
7470 3158 : if (!shutdown)
7471 1852 : PreallocXlogFiles(recptr, checkPoint.ThisTimeLineID);
7472 :
7473 : /*
7474 : * Truncate pg_subtrans if possible. We can throw away all data before
7475 : * the oldest XMIN of any running transaction. No future transaction will
7476 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7477 : * in subtrans.c). During recovery, though, we mustn't do this because
7478 : * StartupSUBTRANS hasn't been called yet.
7479 : */
7480 3158 : if (!RecoveryInProgress())
7481 3102 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7482 :
7483 : /* Real work is done; log and update stats. */
7484 3158 : LogCheckpointEnd(false);
7485 :
7486 : /* Reset the process title */
7487 3158 : update_checkpoint_display(flags, false, true);
7488 :
7489 : TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
7490 : NBuffers,
7491 : CheckpointStats.ckpt_segs_added,
7492 : CheckpointStats.ckpt_segs_removed,
7493 : CheckpointStats.ckpt_segs_recycled);
7494 :
7495 3158 : return true;
7496 : }
7497 :
7498 : /*
7499 : * Mark the end of recovery in WAL though without running a full checkpoint.
7500 : * We can expect that a restartpoint is likely to be in progress as we
7501 : * do this, though we are unwilling to wait for it to complete.
7502 : *
7503 : * CreateRestartPoint() allows for the case where recovery may end before
7504 : * the restartpoint completes so there is no concern of concurrent behaviour.
7505 : */
7506 : static void
7507 96 : CreateEndOfRecoveryRecord(void)
7508 : {
7509 : xl_end_of_recovery xlrec;
7510 : XLogRecPtr recptr;
7511 :
7512 : /* sanity check */
7513 96 : if (!RecoveryInProgress())
7514 0 : elog(ERROR, "can only be used to end recovery");
7515 :
7516 96 : xlrec.end_time = GetCurrentTimestamp();
7517 96 : xlrec.wal_level = wal_level;
7518 :
7519 96 : WALInsertLockAcquireExclusive();
7520 96 : xlrec.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7521 96 : xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7522 96 : WALInsertLockRelease();
7523 :
7524 96 : START_CRIT_SECTION();
7525 :
7526 96 : XLogBeginInsert();
7527 96 : XLogRegisterData(&xlrec, sizeof(xl_end_of_recovery));
7528 96 : recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY);
7529 :
7530 96 : XLogFlush(recptr);
7531 :
7532 : /*
7533 : * Update the control file so that crash recovery can follow the timeline
7534 : * changes to this point.
7535 : */
7536 96 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7537 96 : ControlFile->minRecoveryPoint = recptr;
7538 96 : ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
7539 96 : UpdateControlFile();
7540 96 : LWLockRelease(ControlFileLock);
7541 :
7542 96 : END_CRIT_SECTION();
7543 96 : }
7544 :
7545 : /*
7546 : * Write an OVERWRITE_CONTRECORD message.
7547 : *
7548 : * When on WAL replay we expect a continuation record at the start of a page
7549 : * that is not there, recovery ends and WAL writing resumes at that point.
7550 : * But it's wrong to resume writing new WAL back at the start of the record
7551 : * that was broken, because downstream consumers of that WAL (physical
7552 : * replicas) are not prepared to "rewind". So the first action after
7553 : * finishing replay of all valid WAL must be to write a record of this type
7554 : * at the point where the contrecord was missing; to support xlogreader
7555 : * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
7556 : * to the page header where the record occurs. xlogreader has an ad-hoc
7557 : * mechanism to report metadata about the broken record, which is what we
7558 : * use here.
7559 : *
7560 : * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
7561 : * skip the record it was reading, and pass back the LSN of the skipped
7562 : * record, so that its caller can verify (on "replay" of that record) that the
7563 : * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
7564 : *
7565 : * 'aborted_lsn' is the beginning position of the record that was incomplete.
7566 : * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
7567 : * beginning of the XLOG page where the record is to be inserted. They must
7568 : * match the current WAL insert position, they're passed here just so that we
7569 : * can verify that.
7570 : */
7571 : static XLogRecPtr
7572 22 : CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn, XLogRecPtr pagePtr,
7573 : TimeLineID newTLI)
7574 : {
7575 : xl_overwrite_contrecord xlrec;
7576 : XLogRecPtr recptr;
7577 : XLogPageHeader pagehdr;
7578 : XLogRecPtr startPos;
7579 :
7580 : /* sanity checks */
7581 22 : if (!RecoveryInProgress())
7582 0 : elog(ERROR, "can only be used at end of recovery");
7583 22 : if (pagePtr % XLOG_BLCKSZ != 0)
7584 0 : elog(ERROR, "invalid position for missing continuation record %X/%08X",
7585 : LSN_FORMAT_ARGS(pagePtr));
7586 :
7587 : /* The current WAL insert position should be right after the page header */
7588 22 : startPos = pagePtr;
7589 22 : if (XLogSegmentOffset(startPos, wal_segment_size) == 0)
7590 2 : startPos += SizeOfXLogLongPHD;
7591 : else
7592 20 : startPos += SizeOfXLogShortPHD;
7593 22 : recptr = GetXLogInsertRecPtr();
7594 22 : if (recptr != startPos)
7595 0 : elog(ERROR, "invalid WAL insert position %X/%08X for OVERWRITE_CONTRECORD",
7596 : LSN_FORMAT_ARGS(recptr));
7597 :
7598 22 : START_CRIT_SECTION();
7599 :
7600 : /*
7601 : * Initialize the XLOG page header (by GetXLogBuffer), and set the
7602 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
7603 : *
7604 : * No other backend is allowed to write WAL yet, so acquiring the WAL
7605 : * insertion lock is just pro forma.
7606 : */
7607 22 : WALInsertLockAcquire();
7608 22 : pagehdr = (XLogPageHeader) GetXLogBuffer(pagePtr, newTLI);
7609 22 : pagehdr->xlp_info |= XLP_FIRST_IS_OVERWRITE_CONTRECORD;
7610 22 : WALInsertLockRelease();
7611 :
7612 : /*
7613 : * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
7614 : * page. We know it becomes the first record, because no other backend is
7615 : * allowed to write WAL yet.
7616 : */
7617 22 : XLogBeginInsert();
7618 22 : xlrec.overwritten_lsn = aborted_lsn;
7619 22 : xlrec.overwrite_time = GetCurrentTimestamp();
7620 22 : XLogRegisterData(&xlrec, sizeof(xl_overwrite_contrecord));
7621 22 : recptr = XLogInsert(RM_XLOG_ID, XLOG_OVERWRITE_CONTRECORD);
7622 :
7623 : /* check that the record was inserted to the right place */
7624 22 : if (ProcLastRecPtr != startPos)
7625 0 : elog(ERROR, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%08X",
7626 : LSN_FORMAT_ARGS(ProcLastRecPtr));
7627 :
7628 22 : XLogFlush(recptr);
7629 :
7630 22 : END_CRIT_SECTION();
7631 :
7632 22 : return recptr;
7633 : }
7634 :
7635 : /*
7636 : * Flush all data in shared memory to disk, and fsync
7637 : *
7638 : * This is the common code shared between regular checkpoints and
7639 : * recovery restartpoints.
7640 : */
7641 : static void
7642 3550 : CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
7643 : {
7644 3550 : CheckPointRelationMap();
7645 3550 : CheckPointReplicationSlots(flags & CHECKPOINT_IS_SHUTDOWN);
7646 3550 : CheckPointSnapBuild();
7647 3550 : CheckPointLogicalRewriteHeap();
7648 3550 : CheckPointReplicationOrigin();
7649 :
7650 : /* Write out all dirty data in SLRUs and the main buffer pool */
7651 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags);
7652 3550 : CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
7653 3550 : CheckPointCLOG();
7654 3550 : CheckPointCommitTs();
7655 3550 : CheckPointSUBTRANS();
7656 3550 : CheckPointMultiXact();
7657 3550 : CheckPointPredicate();
7658 3550 : CheckPointBuffers(flags);
7659 :
7660 : /* Perform all queued up fsyncs */
7661 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
7662 3550 : CheckpointStats.ckpt_sync_t = GetCurrentTimestamp();
7663 3550 : ProcessSyncRequests();
7664 3550 : CheckpointStats.ckpt_sync_end_t = GetCurrentTimestamp();
7665 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
7666 :
7667 : /* We deliberately delay 2PC checkpointing as long as possible */
7668 3550 : CheckPointTwoPhase(checkPointRedo);
7669 3550 : }
7670 :
7671 : /*
7672 : * Save a checkpoint for recovery restart if appropriate
7673 : *
7674 : * This function is called each time a checkpoint record is read from XLOG.
7675 : * It must determine whether the checkpoint represents a safe restartpoint or
7676 : * not. If so, the checkpoint record is stashed in shared memory so that
7677 : * CreateRestartPoint can consult it. (Note that the latter function is
7678 : * executed by the checkpointer, while this one will be executed by the
7679 : * startup process.)
7680 : */
7681 : static void
7682 1454 : RecoveryRestartPoint(const CheckPoint *checkPoint, XLogReaderState *record)
7683 : {
7684 : /*
7685 : * Also refrain from creating a restartpoint if we have seen any
7686 : * references to non-existent pages. Restarting recovery from the
7687 : * restartpoint would not see the references, so we would lose the
7688 : * cross-check that the pages belonged to a relation that was dropped
7689 : * later.
7690 : */
7691 1454 : if (XLogHaveInvalidPages())
7692 : {
7693 0 : elog(DEBUG2,
7694 : "could not record restart point at %X/%08X because there are unresolved references to invalid pages",
7695 : LSN_FORMAT_ARGS(checkPoint->redo));
7696 0 : return;
7697 : }
7698 :
7699 : /*
7700 : * Copy the checkpoint record to shared memory, so that checkpointer can
7701 : * work out the next time it wants to perform a restartpoint.
7702 : */
7703 1454 : SpinLockAcquire(&XLogCtl->info_lck);
7704 1454 : XLogCtl->lastCheckPointRecPtr = record->ReadRecPtr;
7705 1454 : XLogCtl->lastCheckPointEndPtr = record->EndRecPtr;
7706 1454 : XLogCtl->lastCheckPoint = *checkPoint;
7707 1454 : SpinLockRelease(&XLogCtl->info_lck);
7708 : }
7709 :
7710 : /*
7711 : * Establish a restartpoint if possible.
7712 : *
7713 : * This is similar to CreateCheckPoint, but is used during WAL recovery
7714 : * to establish a point from which recovery can roll forward without
7715 : * replaying the entire recovery log.
7716 : *
7717 : * Returns true if a new restartpoint was established. We can only establish
7718 : * a restartpoint if we have replayed a safe checkpoint record since last
7719 : * restartpoint.
7720 : */
7721 : bool
7722 1160 : CreateRestartPoint(int flags)
7723 : {
7724 : XLogRecPtr lastCheckPointRecPtr;
7725 : XLogRecPtr lastCheckPointEndPtr;
7726 : CheckPoint lastCheckPoint;
7727 : XLogRecPtr PriorRedoPtr;
7728 : XLogRecPtr receivePtr;
7729 : XLogRecPtr replayPtr;
7730 : TimeLineID replayTLI;
7731 : XLogRecPtr endptr;
7732 : XLogSegNo _logSegNo;
7733 : TimestampTz xtime;
7734 :
7735 : /* Concurrent checkpoint/restartpoint cannot happen */
7736 : Assert(!IsUnderPostmaster || MyBackendType == B_CHECKPOINTER);
7737 :
7738 : /* Get a local copy of the last safe checkpoint record. */
7739 1160 : SpinLockAcquire(&XLogCtl->info_lck);
7740 1160 : lastCheckPointRecPtr = XLogCtl->lastCheckPointRecPtr;
7741 1160 : lastCheckPointEndPtr = XLogCtl->lastCheckPointEndPtr;
7742 1160 : lastCheckPoint = XLogCtl->lastCheckPoint;
7743 1160 : SpinLockRelease(&XLogCtl->info_lck);
7744 :
7745 : /*
7746 : * Check that we're still in recovery mode. It's ok if we exit recovery
7747 : * mode after this check, the restart point is valid anyway.
7748 : */
7749 1160 : if (!RecoveryInProgress())
7750 : {
7751 0 : ereport(DEBUG2,
7752 : (errmsg_internal("skipping restartpoint, recovery has already ended")));
7753 0 : return false;
7754 : }
7755 :
7756 : /*
7757 : * If the last checkpoint record we've replayed is already our last
7758 : * restartpoint, we can't perform a new restart point. We still update
7759 : * minRecoveryPoint in that case, so that if this is a shutdown restart
7760 : * point, we won't start up earlier than before. That's not strictly
7761 : * necessary, but when hot standby is enabled, it would be rather weird if
7762 : * the database opened up for read-only connections at a point-in-time
7763 : * before the last shutdown. Such time travel is still possible in case of
7764 : * immediate shutdown, though.
7765 : *
7766 : * We don't explicitly advance minRecoveryPoint when we do create a
7767 : * restartpoint. It's assumed that flushing the buffers will do that as a
7768 : * side-effect.
7769 : */
7770 1160 : if (!XLogRecPtrIsValid(lastCheckPointRecPtr) ||
7771 494 : lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
7772 : {
7773 768 : ereport(DEBUG2,
7774 : errmsg_internal("skipping restartpoint, already performed at %X/%08X",
7775 : LSN_FORMAT_ARGS(lastCheckPoint.redo)));
7776 :
7777 768 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
7778 768 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7779 : {
7780 70 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7781 70 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7782 70 : UpdateControlFile();
7783 70 : LWLockRelease(ControlFileLock);
7784 : }
7785 768 : return false;
7786 : }
7787 :
7788 : /*
7789 : * Update the shared RedoRecPtr so that the startup process can calculate
7790 : * the number of segments replayed since last restartpoint, and request a
7791 : * restartpoint if it exceeds CheckPointSegments.
7792 : *
7793 : * Like in CreateCheckPoint(), hold off insertions to update it, although
7794 : * during recovery this is just pro forma, because no WAL insertions are
7795 : * happening.
7796 : */
7797 392 : WALInsertLockAcquireExclusive();
7798 392 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = lastCheckPoint.redo;
7799 392 : WALInsertLockRelease();
7800 :
7801 : /* Also update the info_lck-protected copy */
7802 392 : SpinLockAcquire(&XLogCtl->info_lck);
7803 392 : XLogCtl->RedoRecPtr = lastCheckPoint.redo;
7804 392 : SpinLockRelease(&XLogCtl->info_lck);
7805 :
7806 : /*
7807 : * Prepare to accumulate statistics.
7808 : *
7809 : * Note: because it is possible for log_checkpoints to change while a
7810 : * checkpoint proceeds, we always accumulate stats, even if
7811 : * log_checkpoints is currently off.
7812 : */
7813 4312 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7814 392 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7815 :
7816 392 : if (log_checkpoints)
7817 392 : LogCheckpointStart(flags, true);
7818 :
7819 : /* Update the process title */
7820 392 : update_checkpoint_display(flags, true, false);
7821 :
7822 392 : CheckPointGuts(lastCheckPoint.redo, flags);
7823 :
7824 : /*
7825 : * This location needs to be after CheckPointGuts() to ensure that some
7826 : * work has already happened during this checkpoint.
7827 : */
7828 392 : INJECTION_POINT("create-restart-point", NULL);
7829 :
7830 : /*
7831 : * Remember the prior checkpoint's redo ptr for
7832 : * UpdateCheckPointDistanceEstimate()
7833 : */
7834 392 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7835 :
7836 : /*
7837 : * Update pg_control, using current time. Check that it still shows an
7838 : * older checkpoint, else do nothing; this is a quick hack to make sure
7839 : * nothing really bad happens if somehow we get here after the
7840 : * end-of-recovery checkpoint.
7841 : */
7842 392 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7843 392 : if (ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
7844 : {
7845 : /*
7846 : * Update the checkpoint information. We do this even if the cluster
7847 : * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
7848 : * segments recycled below.
7849 : */
7850 392 : ControlFile->checkPoint = lastCheckPointRecPtr;
7851 392 : ControlFile->checkPointCopy = lastCheckPoint;
7852 :
7853 : /*
7854 : * Ensure minRecoveryPoint is past the checkpoint record and update it
7855 : * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
7856 : * this will have happened already while writing out dirty buffers,
7857 : * but not necessarily - e.g. because no buffers were dirtied. We do
7858 : * this because a backup performed in recovery uses minRecoveryPoint
7859 : * to determine which WAL files must be included in the backup, and
7860 : * the file (or files) containing the checkpoint record must be
7861 : * included, at a minimum. Note that for an ordinary restart of
7862 : * recovery there's no value in having the minimum recovery point any
7863 : * earlier than this anyway, because redo will begin just after the
7864 : * checkpoint record.
7865 : */
7866 392 : if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
7867 : {
7868 392 : if (ControlFile->minRecoveryPoint < lastCheckPointEndPtr)
7869 : {
7870 40 : ControlFile->minRecoveryPoint = lastCheckPointEndPtr;
7871 40 : ControlFile->minRecoveryPointTLI = lastCheckPoint.ThisTimeLineID;
7872 :
7873 : /* update local copy */
7874 40 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
7875 40 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
7876 : }
7877 392 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7878 44 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7879 : }
7880 392 : UpdateControlFile();
7881 : }
7882 392 : LWLockRelease(ControlFileLock);
7883 :
7884 : /*
7885 : * Update the average distance between checkpoints/restartpoints if the
7886 : * prior checkpoint exists.
7887 : */
7888 392 : if (XLogRecPtrIsValid(PriorRedoPtr))
7889 392 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7890 :
7891 : /*
7892 : * Delete old log files, those no longer needed for last restartpoint to
7893 : * prevent the disk holding the xlog from growing full.
7894 : */
7895 392 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7896 :
7897 : /*
7898 : * Retreat _logSegNo using the current end of xlog replayed or received,
7899 : * whichever is later.
7900 : */
7901 392 : receivePtr = GetWalRcvFlushRecPtr(NULL, NULL);
7902 392 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
7903 392 : endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
7904 392 : KeepLogSeg(endptr, &_logSegNo);
7905 392 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7906 : _logSegNo, InvalidOid,
7907 : InvalidTransactionId))
7908 : {
7909 : /*
7910 : * Some slots have been invalidated; recalculate the old-segment
7911 : * horizon, starting again from RedoRecPtr.
7912 : */
7913 2 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7914 2 : KeepLogSeg(endptr, &_logSegNo);
7915 : }
7916 392 : _logSegNo--;
7917 :
7918 : /*
7919 : * Try to recycle segments on a useful timeline. If we've been promoted
7920 : * since the beginning of this restartpoint, use the new timeline chosen
7921 : * at end of recovery. If we're still in recovery, use the timeline we're
7922 : * currently replaying.
7923 : *
7924 : * There is no guarantee that the WAL segments will be useful on the
7925 : * current timeline; if recovery proceeds to a new timeline right after
7926 : * this, the pre-allocated WAL segments on this timeline will not be used,
7927 : * and will go wasted until recycled on the next restartpoint. We'll live
7928 : * with that.
7929 : */
7930 392 : if (!RecoveryInProgress())
7931 0 : replayTLI = XLogCtl->InsertTimeLineID;
7932 :
7933 392 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, endptr, replayTLI);
7934 :
7935 : /*
7936 : * Make more log segments if needed. (Do this after recycling old log
7937 : * segments, since that may supply some of the needed files.)
7938 : */
7939 392 : PreallocXlogFiles(endptr, replayTLI);
7940 :
7941 : /*
7942 : * Truncate pg_subtrans if possible. We can throw away all data before
7943 : * the oldest XMIN of any running transaction. No future transaction will
7944 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7945 : * in subtrans.c). When hot standby is disabled, though, we mustn't do
7946 : * this because StartupSUBTRANS hasn't been called yet.
7947 : */
7948 392 : if (EnableHotStandby)
7949 392 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7950 :
7951 : /* Real work is done; log and update stats. */
7952 392 : LogCheckpointEnd(true);
7953 :
7954 : /* Reset the process title */
7955 392 : update_checkpoint_display(flags, true, true);
7956 :
7957 392 : xtime = GetLatestXTime();
7958 392 : ereport((log_checkpoints ? LOG : DEBUG2),
7959 : errmsg("recovery restart point at %X/%08X",
7960 : LSN_FORMAT_ARGS(lastCheckPoint.redo)),
7961 : xtime ? errdetail("Last completed transaction was at log time %s.",
7962 : timestamptz_to_str(xtime)) : 0);
7963 :
7964 : /*
7965 : * Finally, execute archive_cleanup_command, if any.
7966 : */
7967 392 : if (archiveCleanupCommand && strcmp(archiveCleanupCommand, "") != 0)
7968 0 : ExecuteRecoveryCommand(archiveCleanupCommand,
7969 : "archive_cleanup_command",
7970 : false,
7971 : WAIT_EVENT_ARCHIVE_CLEANUP_COMMAND);
7972 :
7973 392 : return true;
7974 : }
7975 :
7976 : /*
7977 : * Report availability of WAL for the given target LSN
7978 : * (typically a slot's restart_lsn)
7979 : *
7980 : * Returns one of the following enum values:
7981 : *
7982 : * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
7983 : * max_wal_size.
7984 : *
7985 : * * WALAVAIL_EXTENDED means it is still available by preserving extra
7986 : * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
7987 : * than max_wal_size, this state is not returned.
7988 : *
7989 : * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
7990 : * remove reserved segments. The walsender using this slot may return to the
7991 : * above.
7992 : *
7993 : * * WALAVAIL_REMOVED means it has been removed. A replication stream on
7994 : * a slot with this LSN cannot continue. (Any associated walsender
7995 : * processes should have been terminated already.)
7996 : *
7997 : * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
7998 : */
7999 : WALAvailability
8000 1088 : GetWALAvailability(XLogRecPtr targetLSN)
8001 : {
8002 : XLogRecPtr currpos; /* current write LSN */
8003 : XLogSegNo currSeg; /* segid of currpos */
8004 : XLogSegNo targetSeg; /* segid of targetLSN */
8005 : XLogSegNo oldestSeg; /* actual oldest segid */
8006 : XLogSegNo oldestSegMaxWalSize; /* oldest segid kept by max_wal_size */
8007 : XLogSegNo oldestSlotSeg; /* oldest segid kept by slot */
8008 : uint64 keepSegs;
8009 :
8010 : /*
8011 : * slot does not reserve WAL. Either deactivated, or has never been active
8012 : */
8013 1088 : if (!XLogRecPtrIsValid(targetLSN))
8014 78 : return WALAVAIL_INVALID_LSN;
8015 :
8016 : /*
8017 : * Calculate the oldest segment currently reserved by all slots,
8018 : * considering wal_keep_size and max_slot_wal_keep_size. Initialize
8019 : * oldestSlotSeg to the current segment.
8020 : */
8021 1010 : currpos = GetXLogWriteRecPtr();
8022 1010 : XLByteToSeg(currpos, oldestSlotSeg, wal_segment_size);
8023 1010 : KeepLogSeg(currpos, &oldestSlotSeg);
8024 :
8025 : /*
8026 : * Find the oldest extant segment file. We get 1 until checkpoint removes
8027 : * the first WAL segment file since startup, which causes the status being
8028 : * wrong under certain abnormal conditions but that doesn't actually harm.
8029 : */
8030 1010 : oldestSeg = XLogGetLastRemovedSegno() + 1;
8031 :
8032 : /* calculate oldest segment by max_wal_size */
8033 1010 : XLByteToSeg(currpos, currSeg, wal_segment_size);
8034 1010 : keepSegs = ConvertToXSegs(max_wal_size_mb, wal_segment_size) + 1;
8035 :
8036 1010 : if (currSeg > keepSegs)
8037 16 : oldestSegMaxWalSize = currSeg - keepSegs;
8038 : else
8039 994 : oldestSegMaxWalSize = 1;
8040 :
8041 : /* the segment we care about */
8042 1010 : XLByteToSeg(targetLSN, targetSeg, wal_segment_size);
8043 :
8044 : /*
8045 : * No point in returning reserved or extended status values if the
8046 : * targetSeg is known to be lost.
8047 : */
8048 1010 : if (targetSeg >= oldestSlotSeg)
8049 : {
8050 : /* show "reserved" when targetSeg is within max_wal_size */
8051 1008 : if (targetSeg >= oldestSegMaxWalSize)
8052 1004 : return WALAVAIL_RESERVED;
8053 :
8054 : /* being retained by slots exceeding max_wal_size */
8055 4 : return WALAVAIL_EXTENDED;
8056 : }
8057 :
8058 : /* WAL segments are no longer retained but haven't been removed yet */
8059 2 : if (targetSeg >= oldestSeg)
8060 2 : return WALAVAIL_UNRESERVED;
8061 :
8062 : /* Definitely lost */
8063 0 : return WALAVAIL_REMOVED;
8064 : }
8065 :
8066 :
8067 : /*
8068 : * Retreat *logSegNo to the last segment that we need to retain because of
8069 : * either wal_keep_size or replication slots.
8070 : *
8071 : * This is calculated by subtracting wal_keep_size from the given xlog
8072 : * location, recptr and by making sure that that result is below the
8073 : * requirement of replication slots. For the latter criterion we do consider
8074 : * the effects of max_slot_wal_keep_size: reserve at most that much space back
8075 : * from recptr.
8076 : *
8077 : * Note about replication slots: if this function calculates a value
8078 : * that's further ahead than what slots need reserved, then affected
8079 : * slots need to be invalidated and this function invoked again.
8080 : * XXX it might be a good idea to rewrite this function so that
8081 : * invalidation is optionally done here, instead.
8082 : */
8083 : static void
8084 4570 : KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
8085 : {
8086 : XLogSegNo currSegNo;
8087 : XLogSegNo segno;
8088 : XLogRecPtr keep;
8089 :
8090 4570 : XLByteToSeg(recptr, currSegNo, wal_segment_size);
8091 4570 : segno = currSegNo;
8092 :
8093 : /* Calculate how many segments are kept by slots. */
8094 4570 : keep = XLogGetReplicationSlotMinimumLSN();
8095 4570 : if (XLogRecPtrIsValid(keep) && keep < recptr)
8096 : {
8097 1324 : XLByteToSeg(keep, segno, wal_segment_size);
8098 :
8099 : /*
8100 : * Account for max_slot_wal_keep_size to avoid keeping more than
8101 : * configured. However, don't do that during a binary upgrade: if
8102 : * slots were to be invalidated because of this, it would not be
8103 : * possible to preserve logical ones during the upgrade.
8104 : */
8105 1324 : if (max_slot_wal_keep_size_mb >= 0 && !IsBinaryUpgrade)
8106 : {
8107 : uint64 slot_keep_segs;
8108 :
8109 46 : slot_keep_segs =
8110 46 : ConvertToXSegs(max_slot_wal_keep_size_mb, wal_segment_size);
8111 :
8112 46 : if (currSegNo - segno > slot_keep_segs)
8113 12 : segno = currSegNo - slot_keep_segs;
8114 : }
8115 : }
8116 :
8117 : /*
8118 : * If WAL summarization is in use, don't remove WAL that has yet to be
8119 : * summarized.
8120 : */
8121 4570 : keep = GetOldestUnsummarizedLSN(NULL, NULL);
8122 4570 : if (XLogRecPtrIsValid(keep))
8123 : {
8124 : XLogSegNo unsummarized_segno;
8125 :
8126 4 : XLByteToSeg(keep, unsummarized_segno, wal_segment_size);
8127 4 : if (unsummarized_segno < segno)
8128 2 : segno = unsummarized_segno;
8129 : }
8130 :
8131 : /* but, keep at least wal_keep_size if that's set */
8132 4570 : if (wal_keep_size_mb > 0)
8133 : {
8134 : uint64 keep_segs;
8135 :
8136 148 : keep_segs = ConvertToXSegs(wal_keep_size_mb, wal_segment_size);
8137 148 : if (currSegNo - segno < keep_segs)
8138 : {
8139 : /* avoid underflow, don't go below 1 */
8140 148 : if (currSegNo <= keep_segs)
8141 140 : segno = 1;
8142 : else
8143 8 : segno = currSegNo - keep_segs;
8144 : }
8145 : }
8146 :
8147 : /* don't delete WAL segments newer than the calculated segment */
8148 4570 : if (segno < *logSegNo)
8149 762 : *logSegNo = segno;
8150 4570 : }
8151 :
8152 : /*
8153 : * Write a NEXTOID log record
8154 : */
8155 : void
8156 1262 : XLogPutNextOid(Oid nextOid)
8157 : {
8158 1262 : XLogBeginInsert();
8159 1262 : XLogRegisterData(&nextOid, sizeof(Oid));
8160 1262 : (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID);
8161 :
8162 : /*
8163 : * We need not flush the NEXTOID record immediately, because any of the
8164 : * just-allocated OIDs could only reach disk as part of a tuple insert or
8165 : * update that would have its own XLOG record that must follow the NEXTOID
8166 : * record. Therefore, the standard buffer LSN interlock applied to those
8167 : * records will ensure no such OID reaches disk before the NEXTOID record
8168 : * does.
8169 : *
8170 : * Note, however, that the above statement only covers state "within" the
8171 : * database. When we use a generated OID as a file or directory name, we
8172 : * are in a sense violating the basic WAL rule, because that filesystem
8173 : * change may reach disk before the NEXTOID WAL record does. The impact
8174 : * of this is that if a database crash occurs immediately afterward, we
8175 : * might after restart re-generate the same OID and find that it conflicts
8176 : * with the leftover file or directory. But since for safety's sake we
8177 : * always loop until finding a nonconflicting filename, this poses no real
8178 : * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8179 : */
8180 1262 : }
8181 :
8182 : /*
8183 : * Write an XLOG SWITCH record.
8184 : *
8185 : * Here we just blindly issue an XLogInsert request for the record.
8186 : * All the magic happens inside XLogInsert.
8187 : *
8188 : * The return value is either the end+1 address of the switch record,
8189 : * or the end+1 address of the prior segment if we did not need to
8190 : * write a switch record because we are already at segment start.
8191 : */
8192 : XLogRecPtr
8193 1576 : RequestXLogSwitch(bool mark_unimportant)
8194 : {
8195 : XLogRecPtr RecPtr;
8196 :
8197 : /* XLOG SWITCH has no data */
8198 1576 : XLogBeginInsert();
8199 :
8200 1576 : if (mark_unimportant)
8201 0 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8202 1576 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH);
8203 :
8204 1576 : return RecPtr;
8205 : }
8206 :
8207 : /*
8208 : * Write a RESTORE POINT record
8209 : */
8210 : XLogRecPtr
8211 6 : XLogRestorePoint(const char *rpName)
8212 : {
8213 : XLogRecPtr RecPtr;
8214 : xl_restore_point xlrec;
8215 :
8216 6 : xlrec.rp_time = GetCurrentTimestamp();
8217 6 : strlcpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8218 :
8219 6 : XLogBeginInsert();
8220 6 : XLogRegisterData(&xlrec, sizeof(xl_restore_point));
8221 :
8222 6 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT);
8223 :
8224 6 : ereport(LOG,
8225 : errmsg("restore point \"%s\" created at %X/%08X",
8226 : rpName, LSN_FORMAT_ARGS(RecPtr)));
8227 :
8228 6 : return RecPtr;
8229 : }
8230 :
8231 : /*
8232 : * Check if any of the GUC parameters that are critical for hot standby
8233 : * have changed, and update the value in pg_control file if necessary.
8234 : */
8235 : static void
8236 1856 : XLogReportParameters(void)
8237 : {
8238 1856 : if (wal_level != ControlFile->wal_level ||
8239 1372 : wal_log_hints != ControlFile->wal_log_hints ||
8240 1200 : MaxConnections != ControlFile->MaxConnections ||
8241 1198 : max_worker_processes != ControlFile->max_worker_processes ||
8242 1192 : max_wal_senders != ControlFile->max_wal_senders ||
8243 1142 : max_prepared_xacts != ControlFile->max_prepared_xacts ||
8244 944 : max_locks_per_xact != ControlFile->max_locks_per_xact ||
8245 944 : track_commit_timestamp != ControlFile->track_commit_timestamp)
8246 : {
8247 : /*
8248 : * The change in number of backend slots doesn't need to be WAL-logged
8249 : * if archiving is not enabled, as you can't start archive recovery
8250 : * with wal_level=minimal anyway. We don't really care about the
8251 : * values in pg_control either if wal_level=minimal, but seems better
8252 : * to keep them up-to-date to avoid confusion.
8253 : */
8254 936 : if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8255 : {
8256 : xl_parameter_change xlrec;
8257 : XLogRecPtr recptr;
8258 :
8259 886 : xlrec.MaxConnections = MaxConnections;
8260 886 : xlrec.max_worker_processes = max_worker_processes;
8261 886 : xlrec.max_wal_senders = max_wal_senders;
8262 886 : xlrec.max_prepared_xacts = max_prepared_xacts;
8263 886 : xlrec.max_locks_per_xact = max_locks_per_xact;
8264 886 : xlrec.wal_level = wal_level;
8265 886 : xlrec.wal_log_hints = wal_log_hints;
8266 886 : xlrec.track_commit_timestamp = track_commit_timestamp;
8267 :
8268 886 : XLogBeginInsert();
8269 886 : XLogRegisterData(&xlrec, sizeof(xlrec));
8270 :
8271 886 : recptr = XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE);
8272 886 : XLogFlush(recptr);
8273 : }
8274 :
8275 936 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8276 :
8277 936 : ControlFile->MaxConnections = MaxConnections;
8278 936 : ControlFile->max_worker_processes = max_worker_processes;
8279 936 : ControlFile->max_wal_senders = max_wal_senders;
8280 936 : ControlFile->max_prepared_xacts = max_prepared_xacts;
8281 936 : ControlFile->max_locks_per_xact = max_locks_per_xact;
8282 936 : ControlFile->wal_level = wal_level;
8283 936 : ControlFile->wal_log_hints = wal_log_hints;
8284 936 : ControlFile->track_commit_timestamp = track_commit_timestamp;
8285 936 : UpdateControlFile();
8286 :
8287 936 : LWLockRelease(ControlFileLock);
8288 : }
8289 1856 : }
8290 :
8291 : /*
8292 : * Update full_page_writes in shared memory, and write an
8293 : * XLOG_FPW_CHANGE record if necessary.
8294 : *
8295 : * Note: this function assumes there is no other process running
8296 : * concurrently that could update it.
8297 : */
8298 : void
8299 3124 : UpdateFullPageWrites(void)
8300 : {
8301 3124 : XLogCtlInsert *Insert = &XLogCtl->Insert;
8302 : bool recoveryInProgress;
8303 :
8304 : /*
8305 : * Do nothing if full_page_writes has not been changed.
8306 : *
8307 : * It's safe to check the shared full_page_writes without the lock,
8308 : * because we assume that there is no concurrently running process which
8309 : * can update it.
8310 : */
8311 3124 : if (fullPageWrites == Insert->fullPageWrites)
8312 2326 : return;
8313 :
8314 : /*
8315 : * Perform this outside critical section so that the WAL insert
8316 : * initialization done by RecoveryInProgress() doesn't trigger an
8317 : * assertion failure.
8318 : */
8319 798 : recoveryInProgress = RecoveryInProgress();
8320 :
8321 798 : START_CRIT_SECTION();
8322 :
8323 : /*
8324 : * It's always safe to take full page images, even when not strictly
8325 : * required, but not the other round. So if we're setting full_page_writes
8326 : * to true, first set it true and then write the WAL record. If we're
8327 : * setting it to false, first write the WAL record and then set the global
8328 : * flag.
8329 : */
8330 798 : if (fullPageWrites)
8331 : {
8332 776 : WALInsertLockAcquireExclusive();
8333 776 : Insert->fullPageWrites = true;
8334 776 : WALInsertLockRelease();
8335 : }
8336 :
8337 : /*
8338 : * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8339 : * full_page_writes during archive recovery, if required.
8340 : */
8341 798 : if (XLogStandbyInfoActive() && !recoveryInProgress)
8342 : {
8343 0 : XLogBeginInsert();
8344 0 : XLogRegisterData(&fullPageWrites, sizeof(bool));
8345 :
8346 0 : XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
8347 : }
8348 :
8349 798 : if (!fullPageWrites)
8350 : {
8351 22 : WALInsertLockAcquireExclusive();
8352 22 : Insert->fullPageWrites = false;
8353 22 : WALInsertLockRelease();
8354 : }
8355 798 : END_CRIT_SECTION();
8356 : }
8357 :
8358 : /*
8359 : * XLOG resource manager's routines
8360 : *
8361 : * Definitions of info values are in include/catalog/pg_control.h, though
8362 : * not all record types are related to control file updates.
8363 : *
8364 : * NOTE: Some XLOG record types that are directly related to WAL recovery
8365 : * are handled in xlogrecovery_redo().
8366 : */
8367 : void
8368 88048 : xlog_redo(XLogReaderState *record)
8369 : {
8370 88048 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8371 88048 : XLogRecPtr lsn = record->EndRecPtr;
8372 :
8373 : /*
8374 : * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
8375 : * XLOG_FPI_FOR_HINT records.
8376 : */
8377 : Assert(info == XLOG_FPI || info == XLOG_FPI_FOR_HINT ||
8378 : !XLogRecHasAnyBlockRefs(record));
8379 :
8380 88048 : if (info == XLOG_NEXTOID)
8381 : {
8382 : Oid nextOid;
8383 :
8384 : /*
8385 : * We used to try to take the maximum of TransamVariables->nextOid and
8386 : * the recorded nextOid, but that fails if the OID counter wraps
8387 : * around. Since no OID allocation should be happening during replay
8388 : * anyway, better to just believe the record exactly. We still take
8389 : * OidGenLock while setting the variable, just in case.
8390 : */
8391 192 : memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
8392 192 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8393 192 : TransamVariables->nextOid = nextOid;
8394 192 : TransamVariables->oidCount = 0;
8395 192 : LWLockRelease(OidGenLock);
8396 : }
8397 87856 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
8398 : {
8399 : CheckPoint checkPoint;
8400 : TimeLineID replayTLI;
8401 :
8402 78 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8403 : /* In a SHUTDOWN checkpoint, believe the counters exactly */
8404 78 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8405 78 : TransamVariables->nextXid = checkPoint.nextXid;
8406 78 : LWLockRelease(XidGenLock);
8407 78 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8408 78 : TransamVariables->nextOid = checkPoint.nextOid;
8409 78 : TransamVariables->oidCount = 0;
8410 78 : LWLockRelease(OidGenLock);
8411 78 : MultiXactSetNextMXact(checkPoint.nextMulti,
8412 : checkPoint.nextMultiOffset);
8413 :
8414 78 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8415 : checkPoint.oldestMultiDB);
8416 :
8417 : /*
8418 : * No need to set oldestClogXid here as well; it'll be set when we
8419 : * redo an xl_clog_truncate if it changed since initialization.
8420 : */
8421 78 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
8422 :
8423 : /*
8424 : * If we see a shutdown checkpoint while waiting for an end-of-backup
8425 : * record, the backup was canceled and the end-of-backup record will
8426 : * never arrive.
8427 : */
8428 78 : if (ArchiveRecoveryRequested &&
8429 76 : XLogRecPtrIsValid(ControlFile->backupStartPoint) &&
8430 0 : !XLogRecPtrIsValid(ControlFile->backupEndPoint))
8431 0 : ereport(PANIC,
8432 : (errmsg("online backup was canceled, recovery cannot continue")));
8433 :
8434 : /*
8435 : * If we see a shutdown checkpoint, we know that nothing was running
8436 : * on the primary at this point. So fake-up an empty running-xacts
8437 : * record and use that here and now. Recover additional standby state
8438 : * for prepared transactions.
8439 : */
8440 78 : if (standbyState >= STANDBY_INITIALIZED)
8441 : {
8442 : TransactionId *xids;
8443 : int nxids;
8444 : TransactionId oldestActiveXID;
8445 : TransactionId latestCompletedXid;
8446 : RunningTransactionsData running;
8447 :
8448 72 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
8449 :
8450 : /* Update pg_subtrans entries for any prepared transactions */
8451 72 : StandbyRecoverPreparedTransactions();
8452 :
8453 : /*
8454 : * Construct a RunningTransactions snapshot representing a shut
8455 : * down server, with only prepared transactions still alive. We're
8456 : * never overflowed at this point because all subxids are listed
8457 : * with their parent prepared transactions.
8458 : */
8459 72 : running.xcnt = nxids;
8460 72 : running.subxcnt = 0;
8461 72 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
8462 72 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
8463 72 : running.oldestRunningXid = oldestActiveXID;
8464 72 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
8465 72 : TransactionIdRetreat(latestCompletedXid);
8466 : Assert(TransactionIdIsNormal(latestCompletedXid));
8467 72 : running.latestCompletedXid = latestCompletedXid;
8468 72 : running.xids = xids;
8469 :
8470 72 : ProcArrayApplyRecoveryInfo(&running);
8471 : }
8472 :
8473 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8474 78 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8475 78 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8476 78 : LWLockRelease(ControlFileLock);
8477 :
8478 : /*
8479 : * We should've already switched to the new TLI before replaying this
8480 : * record.
8481 : */
8482 78 : (void) GetCurrentReplayRecPtr(&replayTLI);
8483 78 : if (checkPoint.ThisTimeLineID != replayTLI)
8484 0 : ereport(PANIC,
8485 : (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
8486 : checkPoint.ThisTimeLineID, replayTLI)));
8487 :
8488 78 : RecoveryRestartPoint(&checkPoint, record);
8489 :
8490 : /*
8491 : * After replaying a checkpoint record, free all smgr objects.
8492 : * Otherwise we would never do so for dropped relations, as the
8493 : * startup does not process shared invalidation messages or call
8494 : * AtEOXact_SMgr().
8495 : */
8496 78 : smgrdestroyall();
8497 : }
8498 87778 : else if (info == XLOG_CHECKPOINT_ONLINE)
8499 : {
8500 : CheckPoint checkPoint;
8501 : TimeLineID replayTLI;
8502 :
8503 1376 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8504 : /* In an ONLINE checkpoint, treat the XID counter as a minimum */
8505 1376 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8506 1376 : if (FullTransactionIdPrecedes(TransamVariables->nextXid,
8507 : checkPoint.nextXid))
8508 0 : TransamVariables->nextXid = checkPoint.nextXid;
8509 1376 : LWLockRelease(XidGenLock);
8510 :
8511 : /*
8512 : * We ignore the nextOid counter in an ONLINE checkpoint, preferring
8513 : * to track OID assignment through XLOG_NEXTOID records. The nextOid
8514 : * counter is from the start of the checkpoint and might well be stale
8515 : * compared to later XLOG_NEXTOID records. We could try to take the
8516 : * maximum of the nextOid counter and our latest value, but since
8517 : * there's no particular guarantee about the speed with which the OID
8518 : * counter wraps around, that's a risky thing to do. In any case,
8519 : * users of the nextOid counter are required to avoid assignment of
8520 : * duplicates, so that a somewhat out-of-date value should be safe.
8521 : */
8522 :
8523 : /* Handle multixact */
8524 1376 : MultiXactAdvanceNextMXact(checkPoint.nextMulti,
8525 : checkPoint.nextMultiOffset);
8526 :
8527 : /*
8528 : * NB: This may perform multixact truncation when replaying WAL
8529 : * generated by an older primary.
8530 : */
8531 1376 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8532 : checkPoint.oldestMultiDB);
8533 1376 : if (TransactionIdPrecedes(TransamVariables->oldestXid,
8534 : checkPoint.oldestXid))
8535 0 : SetTransactionIdLimit(checkPoint.oldestXid,
8536 : checkPoint.oldestXidDB);
8537 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8538 1376 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8539 1376 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8540 1376 : LWLockRelease(ControlFileLock);
8541 :
8542 : /* TLI should not change in an on-line checkpoint */
8543 1376 : (void) GetCurrentReplayRecPtr(&replayTLI);
8544 1376 : if (checkPoint.ThisTimeLineID != replayTLI)
8545 0 : ereport(PANIC,
8546 : (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
8547 : checkPoint.ThisTimeLineID, replayTLI)));
8548 :
8549 1376 : RecoveryRestartPoint(&checkPoint, record);
8550 :
8551 : /*
8552 : * After replaying a checkpoint record, free all smgr objects.
8553 : * Otherwise we would never do so for dropped relations, as the
8554 : * startup does not process shared invalidation messages or call
8555 : * AtEOXact_SMgr().
8556 : */
8557 1376 : smgrdestroyall();
8558 : }
8559 86402 : else if (info == XLOG_OVERWRITE_CONTRECORD)
8560 : {
8561 : /* nothing to do here, handled in xlogrecovery_redo() */
8562 : }
8563 86400 : else if (info == XLOG_END_OF_RECOVERY)
8564 : {
8565 : xl_end_of_recovery xlrec;
8566 : TimeLineID replayTLI;
8567 :
8568 22 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
8569 :
8570 : /*
8571 : * For Hot Standby, we could treat this like a Shutdown Checkpoint,
8572 : * but this case is rarer and harder to test, so the benefit doesn't
8573 : * outweigh the potential extra cost of maintenance.
8574 : */
8575 :
8576 : /*
8577 : * We should've already switched to the new TLI before replaying this
8578 : * record.
8579 : */
8580 22 : (void) GetCurrentReplayRecPtr(&replayTLI);
8581 22 : if (xlrec.ThisTimeLineID != replayTLI)
8582 0 : ereport(PANIC,
8583 : (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
8584 : xlrec.ThisTimeLineID, replayTLI)));
8585 : }
8586 86378 : else if (info == XLOG_NOOP)
8587 : {
8588 : /* nothing to do here */
8589 : }
8590 86378 : else if (info == XLOG_SWITCH)
8591 : {
8592 : /* nothing to do here */
8593 : }
8594 85474 : else if (info == XLOG_RESTORE_POINT)
8595 : {
8596 : /* nothing to do here, handled in xlogrecovery.c */
8597 : }
8598 85464 : else if (info == XLOG_FPI || info == XLOG_FPI_FOR_HINT)
8599 : {
8600 : /*
8601 : * XLOG_FPI records contain nothing else but one or more block
8602 : * references. Every block reference must include a full-page image
8603 : * even if full_page_writes was disabled when the record was generated
8604 : * - otherwise there would be no point in this record.
8605 : *
8606 : * XLOG_FPI_FOR_HINT records are generated when a page needs to be
8607 : * WAL-logged because of a hint bit update. They are only generated
8608 : * when checksums and/or wal_log_hints are enabled. They may include
8609 : * no full-page images if full_page_writes was disabled when they were
8610 : * generated. In this case there is nothing to do here.
8611 : *
8612 : * No recovery conflicts are generated by these generic records - if a
8613 : * resource manager needs to generate conflicts, it has to define a
8614 : * separate WAL record type and redo routine.
8615 : */
8616 177104 : for (uint8 block_id = 0; block_id <= XLogRecMaxBlockId(record); block_id++)
8617 : {
8618 : Buffer buffer;
8619 :
8620 93316 : if (!XLogRecHasBlockImage(record, block_id))
8621 : {
8622 132 : if (info == XLOG_FPI)
8623 0 : elog(ERROR, "XLOG_FPI record did not contain a full-page image");
8624 132 : continue;
8625 : }
8626 :
8627 93184 : if (XLogReadBufferForRedo(record, block_id, &buffer) != BLK_RESTORED)
8628 0 : elog(ERROR, "unexpected XLogReadBufferForRedo result when restoring backup block");
8629 93184 : UnlockReleaseBuffer(buffer);
8630 : }
8631 : }
8632 1676 : else if (info == XLOG_BACKUP_END)
8633 : {
8634 : /* nothing to do here, handled in xlogrecovery_redo() */
8635 : }
8636 1490 : else if (info == XLOG_PARAMETER_CHANGE)
8637 : {
8638 : xl_parameter_change xlrec;
8639 :
8640 : /* Update our copy of the parameters in pg_control */
8641 76 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
8642 :
8643 76 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8644 76 : ControlFile->MaxConnections = xlrec.MaxConnections;
8645 76 : ControlFile->max_worker_processes = xlrec.max_worker_processes;
8646 76 : ControlFile->max_wal_senders = xlrec.max_wal_senders;
8647 76 : ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
8648 76 : ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
8649 76 : ControlFile->wal_level = xlrec.wal_level;
8650 76 : ControlFile->wal_log_hints = xlrec.wal_log_hints;
8651 :
8652 : /*
8653 : * Update minRecoveryPoint to ensure that if recovery is aborted, we
8654 : * recover back up to this point before allowing hot standby again.
8655 : * This is important if the max_* settings are decreased, to ensure
8656 : * you don't run queries against the WAL preceding the change. The
8657 : * local copies cannot be updated as long as crash recovery is
8658 : * happening and we expect all the WAL to be replayed.
8659 : */
8660 76 : if (InArchiveRecovery)
8661 : {
8662 46 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
8663 46 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
8664 : }
8665 76 : if (XLogRecPtrIsValid(LocalMinRecoveryPoint) && LocalMinRecoveryPoint < lsn)
8666 : {
8667 : TimeLineID replayTLI;
8668 :
8669 24 : (void) GetCurrentReplayRecPtr(&replayTLI);
8670 24 : ControlFile->minRecoveryPoint = lsn;
8671 24 : ControlFile->minRecoveryPointTLI = replayTLI;
8672 : }
8673 :
8674 76 : CommitTsParameterChange(xlrec.track_commit_timestamp,
8675 76 : ControlFile->track_commit_timestamp);
8676 76 : ControlFile->track_commit_timestamp = xlrec.track_commit_timestamp;
8677 :
8678 76 : UpdateControlFile();
8679 76 : LWLockRelease(ControlFileLock);
8680 :
8681 : /* Check to see if any parameter change gives a problem on recovery */
8682 76 : CheckRequiredParameterValues();
8683 : }
8684 1414 : else if (info == XLOG_FPW_CHANGE)
8685 : {
8686 : bool fpw;
8687 :
8688 0 : memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
8689 :
8690 : /*
8691 : * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
8692 : * do_pg_backup_start() and do_pg_backup_stop() can check whether
8693 : * full_page_writes has been disabled during online backup.
8694 : */
8695 0 : if (!fpw)
8696 : {
8697 0 : SpinLockAcquire(&XLogCtl->info_lck);
8698 0 : if (XLogCtl->lastFpwDisableRecPtr < record->ReadRecPtr)
8699 0 : XLogCtl->lastFpwDisableRecPtr = record->ReadRecPtr;
8700 0 : SpinLockRelease(&XLogCtl->info_lck);
8701 : }
8702 :
8703 : /* Keep track of full_page_writes */
8704 0 : lastFullPageWrites = fpw;
8705 : }
8706 1414 : else if (info == XLOG_CHECKPOINT_REDO)
8707 : {
8708 : /* nothing to do here, just for informational purposes */
8709 : }
8710 36 : else if (info == XLOG_LOGICAL_DECODING_STATUS_CHANGE)
8711 : {
8712 : bool status;
8713 :
8714 36 : memcpy(&status, XLogRecGetData(record), sizeof(bool));
8715 :
8716 : /*
8717 : * We need to toggle the logical decoding status and update the
8718 : * XLogLogicalInfo cache of processes synchronously because
8719 : * XLogLogicalInfoActive() is used even during read-only queries
8720 : * (e.g., via RelationIsAccessibleInLogicalDecoding()). In the
8721 : * 'disable' case, it is safe to invalidate existing slots after
8722 : * disabling logical decoding because logical decoding cannot process
8723 : * subsequent WAL records, which may not contain logical information.
8724 : */
8725 36 : if (status)
8726 18 : EnableLogicalDecoding();
8727 : else
8728 18 : DisableLogicalDecoding();
8729 :
8730 36 : elog(DEBUG1, "update logical decoding status to %d during recovery",
8731 : status);
8732 :
8733 36 : if (InRecovery && InHotStandby)
8734 : {
8735 32 : if (!status)
8736 : {
8737 : /*
8738 : * Invalidate logical slots if we are in hot standby and the
8739 : * primary disabled logical decoding.
8740 : */
8741 18 : InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_LEVEL,
8742 : 0, InvalidOid,
8743 : InvalidTransactionId);
8744 : }
8745 14 : else if (sync_replication_slots)
8746 : {
8747 : /*
8748 : * Signal the postmaster to launch the slotsync worker.
8749 : *
8750 : * XXX: For simplicity, we keep the slotsync worker running
8751 : * even after logical decoding is disabled. A future
8752 : * improvement can consider starting and stopping the worker
8753 : * based on logical decoding status change.
8754 : */
8755 0 : kill(PostmasterPid, SIGUSR1);
8756 : }
8757 : }
8758 : }
8759 88044 : }
8760 :
8761 : /*
8762 : * Return the extra open flags used for opening a file, depending on the
8763 : * value of the GUCs wal_sync_method, fsync and debug_io_direct.
8764 : */
8765 : static int
8766 32462 : get_sync_bit(int method)
8767 : {
8768 32462 : int o_direct_flag = 0;
8769 :
8770 : /*
8771 : * Use O_DIRECT if requested, except in walreceiver process. The WAL
8772 : * written by walreceiver is normally read by the startup process soon
8773 : * after it's written. Also, walreceiver performs unaligned writes, which
8774 : * don't work with O_DIRECT, so it is required for correctness too.
8775 : */
8776 32462 : if ((io_direct_flags & IO_DIRECT_WAL) && !AmWalReceiverProcess())
8777 18 : o_direct_flag = PG_O_DIRECT;
8778 :
8779 : /* If fsync is disabled, never open in sync mode */
8780 32462 : if (!enableFsync)
8781 32462 : return o_direct_flag;
8782 :
8783 0 : switch (method)
8784 : {
8785 : /*
8786 : * enum values for all sync options are defined even if they are
8787 : * not supported on the current platform. But if not, they are
8788 : * not included in the enum option array, and therefore will never
8789 : * be seen here.
8790 : */
8791 0 : case WAL_SYNC_METHOD_FSYNC:
8792 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8793 : case WAL_SYNC_METHOD_FDATASYNC:
8794 0 : return o_direct_flag;
8795 : #ifdef O_SYNC
8796 0 : case WAL_SYNC_METHOD_OPEN:
8797 0 : return O_SYNC | o_direct_flag;
8798 : #endif
8799 : #ifdef O_DSYNC
8800 0 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8801 0 : return O_DSYNC | o_direct_flag;
8802 : #endif
8803 0 : default:
8804 : /* can't happen (unless we are out of sync with option array) */
8805 0 : elog(ERROR, "unrecognized \"wal_sync_method\": %d", method);
8806 : return 0; /* silence warning */
8807 : }
8808 : }
8809 :
8810 : /*
8811 : * GUC support
8812 : */
8813 : void
8814 2346 : assign_wal_sync_method(int new_wal_sync_method, void *extra)
8815 : {
8816 2346 : if (wal_sync_method != new_wal_sync_method)
8817 : {
8818 : /*
8819 : * To ensure that no blocks escape unsynced, force an fsync on the
8820 : * currently open log segment (if any). Also, if the open flag is
8821 : * changing, close the log file so it will be reopened (with new flag
8822 : * bit) at next use.
8823 : */
8824 0 : if (openLogFile >= 0)
8825 : {
8826 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN);
8827 0 : if (pg_fsync(openLogFile) != 0)
8828 : {
8829 : char xlogfname[MAXFNAMELEN];
8830 : int save_errno;
8831 :
8832 0 : save_errno = errno;
8833 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo,
8834 : wal_segment_size);
8835 0 : errno = save_errno;
8836 0 : ereport(PANIC,
8837 : (errcode_for_file_access(),
8838 : errmsg("could not fsync file \"%s\": %m", xlogfname)));
8839 : }
8840 :
8841 0 : pgstat_report_wait_end();
8842 0 : if (get_sync_bit(wal_sync_method) != get_sync_bit(new_wal_sync_method))
8843 0 : XLogFileClose();
8844 : }
8845 : }
8846 2346 : }
8847 :
8848 :
8849 : /*
8850 : * Issue appropriate kind of fsync (if any) for an XLOG output file.
8851 : *
8852 : * 'fd' is a file descriptor for the XLOG file to be fsync'd.
8853 : * 'segno' is for error reporting purposes.
8854 : */
8855 : void
8856 324854 : issue_xlog_fsync(int fd, XLogSegNo segno, TimeLineID tli)
8857 : {
8858 324854 : char *msg = NULL;
8859 : instr_time start;
8860 :
8861 : Assert(tli != 0);
8862 :
8863 : /*
8864 : * Quick exit if fsync is disabled or write() has already synced the WAL
8865 : * file.
8866 : */
8867 324854 : if (!enableFsync ||
8868 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN ||
8869 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN_DSYNC)
8870 324854 : return;
8871 :
8872 : /*
8873 : * Measure I/O timing to sync the WAL file for pg_stat_io.
8874 : */
8875 0 : start = pgstat_prepare_io_time(track_wal_io_timing);
8876 :
8877 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC);
8878 0 : switch (wal_sync_method)
8879 : {
8880 0 : case WAL_SYNC_METHOD_FSYNC:
8881 0 : if (pg_fsync_no_writethrough(fd) != 0)
8882 0 : msg = _("could not fsync file \"%s\": %m");
8883 0 : break;
8884 : #ifdef HAVE_FSYNC_WRITETHROUGH
8885 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8886 : if (pg_fsync_writethrough(fd) != 0)
8887 : msg = _("could not fsync write-through file \"%s\": %m");
8888 : break;
8889 : #endif
8890 0 : case WAL_SYNC_METHOD_FDATASYNC:
8891 0 : if (pg_fdatasync(fd) != 0)
8892 0 : msg = _("could not fdatasync file \"%s\": %m");
8893 0 : break;
8894 0 : case WAL_SYNC_METHOD_OPEN:
8895 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8896 : /* not reachable */
8897 : Assert(false);
8898 0 : break;
8899 0 : default:
8900 0 : ereport(PANIC,
8901 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8902 : errmsg_internal("unrecognized \"wal_sync_method\": %d", wal_sync_method));
8903 : break;
8904 : }
8905 :
8906 : /* PANIC if failed to fsync */
8907 0 : if (msg)
8908 : {
8909 : char xlogfname[MAXFNAMELEN];
8910 0 : int save_errno = errno;
8911 :
8912 0 : XLogFileName(xlogfname, tli, segno, wal_segment_size);
8913 0 : errno = save_errno;
8914 0 : ereport(PANIC,
8915 : (errcode_for_file_access(),
8916 : errmsg(msg, xlogfname)));
8917 : }
8918 :
8919 0 : pgstat_report_wait_end();
8920 :
8921 0 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL, IOOP_FSYNC,
8922 : start, 1, 0);
8923 : }
8924 :
8925 : /*
8926 : * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
8927 : * function. It creates the necessary starting checkpoint and constructs the
8928 : * backup state and tablespace map.
8929 : *
8930 : * Input parameters are "state" (the backup state), "fast" (if true, we do
8931 : * the checkpoint in fast mode), and "tablespaces" (if non-NULL, indicates a
8932 : * list of tablespaceinfo structs describing the cluster's tablespaces.).
8933 : *
8934 : * The tablespace map contents are appended to passed-in parameter
8935 : * tablespace_map and the caller is responsible for including it in the backup
8936 : * archive as 'tablespace_map'. The tablespace_map file is required mainly for
8937 : * tar format in windows as native windows utilities are not able to create
8938 : * symlinks while extracting files from tar. However for consistency and
8939 : * platform-independence, we do it the same way everywhere.
8940 : *
8941 : * It fills in "state" with the information required for the backup, such
8942 : * as the minimum WAL location that must be present to restore from this
8943 : * backup (starttli) and the corresponding timeline ID (starttli).
8944 : *
8945 : * Every successfully started backup must be stopped by calling
8946 : * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
8947 : * backups active at the same time.
8948 : *
8949 : * It is the responsibility of the caller of this function to verify the
8950 : * permissions of the calling user!
8951 : */
8952 : void
8953 332 : do_pg_backup_start(const char *backupidstr, bool fast, List **tablespaces,
8954 : BackupState *state, StringInfo tblspcmapfile)
8955 : {
8956 : bool backup_started_in_recovery;
8957 :
8958 : Assert(state != NULL);
8959 332 : backup_started_in_recovery = RecoveryInProgress();
8960 :
8961 : /*
8962 : * During recovery, we don't need to check WAL level. Because, if WAL
8963 : * level is not sufficient, it's impossible to get here during recovery.
8964 : */
8965 332 : if (!backup_started_in_recovery && !XLogIsNeeded())
8966 0 : ereport(ERROR,
8967 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8968 : errmsg("WAL level not sufficient for making an online backup"),
8969 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
8970 :
8971 332 : if (strlen(backupidstr) > MAXPGPATH)
8972 2 : ereport(ERROR,
8973 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8974 : errmsg("backup label too long (max %d bytes)",
8975 : MAXPGPATH)));
8976 :
8977 330 : strlcpy(state->name, backupidstr, sizeof(state->name));
8978 :
8979 : /*
8980 : * Mark backup active in shared memory. We must do full-page WAL writes
8981 : * during an on-line backup even if not doing so at other times, because
8982 : * it's quite possible for the backup dump to obtain a "torn" (partially
8983 : * written) copy of a database page if it reads the page concurrently with
8984 : * our write to the same page. This can be fixed as long as the first
8985 : * write to the page in the WAL sequence is a full-page write. Hence, we
8986 : * increment runningBackups then force a CHECKPOINT, to ensure there are
8987 : * no dirty pages in shared memory that might get dumped while the backup
8988 : * is in progress without having a corresponding WAL record. (Once the
8989 : * backup is complete, we need not force full-page writes anymore, since
8990 : * we expect that any pages not modified during the backup interval must
8991 : * have been correctly captured by the backup.)
8992 : *
8993 : * Note that forcing full-page writes has no effect during an online
8994 : * backup from the standby.
8995 : *
8996 : * We must hold all the insertion locks to change the value of
8997 : * runningBackups, to ensure adequate interlocking against
8998 : * XLogInsertRecord().
8999 : */
9000 330 : WALInsertLockAcquireExclusive();
9001 330 : XLogCtl->Insert.runningBackups++;
9002 330 : WALInsertLockRelease();
9003 :
9004 : /*
9005 : * Ensure we decrement runningBackups if we fail below. NB -- for this to
9006 : * work correctly, it is critical that sessionBackupState is only updated
9007 : * after this block is over.
9008 : */
9009 330 : PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9010 : {
9011 330 : bool gotUniqueStartpoint = false;
9012 : DIR *tblspcdir;
9013 : struct dirent *de;
9014 : tablespaceinfo *ti;
9015 : int datadirpathlen;
9016 :
9017 : /*
9018 : * Force an XLOG file switch before the checkpoint, to ensure that the
9019 : * WAL segment the checkpoint is written to doesn't contain pages with
9020 : * old timeline IDs. That would otherwise happen if you called
9021 : * pg_backup_start() right after restoring from a PITR archive: the
9022 : * first WAL segment containing the startup checkpoint has pages in
9023 : * the beginning with the old timeline ID. That can cause trouble at
9024 : * recovery: we won't have a history file covering the old timeline if
9025 : * pg_wal directory was not included in the base backup and the WAL
9026 : * archive was cleared too before starting the backup.
9027 : *
9028 : * This also ensures that we have emitted a WAL page header that has
9029 : * XLP_BKP_REMOVABLE off before we emit the checkpoint record.
9030 : * Therefore, if a WAL archiver (such as pglesslog) is trying to
9031 : * compress out removable backup blocks, it won't remove any that
9032 : * occur after this point.
9033 : *
9034 : * During recovery, we skip forcing XLOG file switch, which means that
9035 : * the backup taken during recovery is not available for the special
9036 : * recovery case described above.
9037 : */
9038 330 : if (!backup_started_in_recovery)
9039 314 : RequestXLogSwitch(false);
9040 :
9041 : do
9042 : {
9043 : bool checkpointfpw;
9044 :
9045 : /*
9046 : * Force a CHECKPOINT. Aside from being necessary to prevent torn
9047 : * page problems, this guarantees that two successive backup runs
9048 : * will have different checkpoint positions and hence different
9049 : * history file names, even if nothing happened in between.
9050 : *
9051 : * During recovery, establish a restartpoint if possible. We use
9052 : * the last restartpoint as the backup starting checkpoint. This
9053 : * means that two successive backup runs can have same checkpoint
9054 : * positions.
9055 : *
9056 : * Since the fact that we are executing do_pg_backup_start()
9057 : * during recovery means that checkpointer is running, we can use
9058 : * RequestCheckpoint() to establish a restartpoint.
9059 : *
9060 : * We use CHECKPOINT_FAST only if requested by user (via passing
9061 : * fast = true). Otherwise this can take awhile.
9062 : */
9063 330 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
9064 : (fast ? CHECKPOINT_FAST : 0));
9065 :
9066 : /*
9067 : * Now we need to fetch the checkpoint record location, and also
9068 : * its REDO pointer. The oldest point in WAL that would be needed
9069 : * to restore starting from the checkpoint is precisely the REDO
9070 : * pointer.
9071 : */
9072 330 : LWLockAcquire(ControlFileLock, LW_SHARED);
9073 330 : state->checkpointloc = ControlFile->checkPoint;
9074 330 : state->startpoint = ControlFile->checkPointCopy.redo;
9075 330 : state->starttli = ControlFile->checkPointCopy.ThisTimeLineID;
9076 330 : checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
9077 330 : LWLockRelease(ControlFileLock);
9078 :
9079 330 : if (backup_started_in_recovery)
9080 : {
9081 : XLogRecPtr recptr;
9082 :
9083 : /*
9084 : * Check to see if all WAL replayed during online backup
9085 : * (i.e., since last restartpoint used as backup starting
9086 : * checkpoint) contain full-page writes.
9087 : */
9088 16 : SpinLockAcquire(&XLogCtl->info_lck);
9089 16 : recptr = XLogCtl->lastFpwDisableRecPtr;
9090 16 : SpinLockRelease(&XLogCtl->info_lck);
9091 :
9092 16 : if (!checkpointfpw || state->startpoint <= recptr)
9093 0 : ereport(ERROR,
9094 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9095 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9096 : "since last restartpoint"),
9097 : errhint("This means that the backup being taken on the standby "
9098 : "is corrupt and should not be used. "
9099 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9100 : "and then try an online backup again.")));
9101 :
9102 : /*
9103 : * During recovery, since we don't use the end-of-backup WAL
9104 : * record and don't write the backup history file, the
9105 : * starting WAL location doesn't need to be unique. This means
9106 : * that two base backups started at the same time might use
9107 : * the same checkpoint as starting locations.
9108 : */
9109 16 : gotUniqueStartpoint = true;
9110 : }
9111 :
9112 : /*
9113 : * If two base backups are started at the same time (in WAL sender
9114 : * processes), we need to make sure that they use different
9115 : * checkpoints as starting locations, because we use the starting
9116 : * WAL location as a unique identifier for the base backup in the
9117 : * end-of-backup WAL record and when we write the backup history
9118 : * file. Perhaps it would be better generate a separate unique ID
9119 : * for each backup instead of forcing another checkpoint, but
9120 : * taking a checkpoint right after another is not that expensive
9121 : * either because only few buffers have been dirtied yet.
9122 : */
9123 330 : WALInsertLockAcquireExclusive();
9124 330 : if (XLogCtl->Insert.lastBackupStart < state->startpoint)
9125 : {
9126 330 : XLogCtl->Insert.lastBackupStart = state->startpoint;
9127 330 : gotUniqueStartpoint = true;
9128 : }
9129 330 : WALInsertLockRelease();
9130 330 : } while (!gotUniqueStartpoint);
9131 :
9132 : /*
9133 : * Construct tablespace_map file.
9134 : */
9135 330 : datadirpathlen = strlen(DataDir);
9136 :
9137 : /* Collect information about all tablespaces */
9138 330 : tblspcdir = AllocateDir(PG_TBLSPC_DIR);
9139 1064 : while ((de = ReadDir(tblspcdir, PG_TBLSPC_DIR)) != NULL)
9140 : {
9141 : char fullpath[MAXPGPATH + sizeof(PG_TBLSPC_DIR)];
9142 : char linkpath[MAXPGPATH];
9143 734 : char *relpath = NULL;
9144 : char *s;
9145 : PGFileType de_type;
9146 : char *badp;
9147 : Oid tsoid;
9148 :
9149 : /*
9150 : * Try to parse the directory name as an unsigned integer.
9151 : *
9152 : * Tablespace directories should be positive integers that can be
9153 : * represented in 32 bits, with no leading zeroes or trailing
9154 : * garbage. If we come across a name that doesn't meet those
9155 : * criteria, skip it.
9156 : */
9157 734 : if (de->d_name[0] < '1' || de->d_name[1] > '9')
9158 660 : continue;
9159 74 : errno = 0;
9160 74 : tsoid = strtoul(de->d_name, &badp, 10);
9161 74 : if (*badp != '\0' || errno == EINVAL || errno == ERANGE)
9162 0 : continue;
9163 :
9164 74 : snprintf(fullpath, sizeof(fullpath), "%s/%s", PG_TBLSPC_DIR, de->d_name);
9165 :
9166 74 : de_type = get_dirent_type(fullpath, de, false, ERROR);
9167 :
9168 74 : if (de_type == PGFILETYPE_LNK)
9169 : {
9170 : StringInfoData escapedpath;
9171 : int rllen;
9172 :
9173 46 : rllen = readlink(fullpath, linkpath, sizeof(linkpath));
9174 46 : if (rllen < 0)
9175 : {
9176 0 : ereport(WARNING,
9177 : (errmsg("could not read symbolic link \"%s\": %m",
9178 : fullpath)));
9179 0 : continue;
9180 : }
9181 46 : else if (rllen >= sizeof(linkpath))
9182 : {
9183 0 : ereport(WARNING,
9184 : (errmsg("symbolic link \"%s\" target is too long",
9185 : fullpath)));
9186 0 : continue;
9187 : }
9188 46 : linkpath[rllen] = '\0';
9189 :
9190 : /*
9191 : * Relpath holds the relative path of the tablespace directory
9192 : * when it's located within PGDATA, or NULL if it's located
9193 : * elsewhere.
9194 : */
9195 46 : if (rllen > datadirpathlen &&
9196 2 : strncmp(linkpath, DataDir, datadirpathlen) == 0 &&
9197 0 : IS_DIR_SEP(linkpath[datadirpathlen]))
9198 0 : relpath = pstrdup(linkpath + datadirpathlen + 1);
9199 :
9200 : /*
9201 : * Add a backslash-escaped version of the link path to the
9202 : * tablespace map file.
9203 : */
9204 46 : initStringInfo(&escapedpath);
9205 1124 : for (s = linkpath; *s; s++)
9206 : {
9207 1078 : if (*s == '\n' || *s == '\r' || *s == '\\')
9208 0 : appendStringInfoChar(&escapedpath, '\\');
9209 1078 : appendStringInfoChar(&escapedpath, *s);
9210 : }
9211 46 : appendStringInfo(tblspcmapfile, "%s %s\n",
9212 46 : de->d_name, escapedpath.data);
9213 46 : pfree(escapedpath.data);
9214 : }
9215 28 : else if (de_type == PGFILETYPE_DIR)
9216 : {
9217 : /*
9218 : * It's possible to use allow_in_place_tablespaces to create
9219 : * directories directly under pg_tblspc, for testing purposes
9220 : * only.
9221 : *
9222 : * In this case, we store a relative path rather than an
9223 : * absolute path into the tablespaceinfo.
9224 : */
9225 28 : snprintf(linkpath, sizeof(linkpath), "%s/%s",
9226 28 : PG_TBLSPC_DIR, de->d_name);
9227 28 : relpath = pstrdup(linkpath);
9228 : }
9229 : else
9230 : {
9231 : /* Skip any other file type that appears here. */
9232 0 : continue;
9233 : }
9234 :
9235 74 : ti = palloc_object(tablespaceinfo);
9236 74 : ti->oid = tsoid;
9237 74 : ti->path = pstrdup(linkpath);
9238 74 : ti->rpath = relpath;
9239 74 : ti->size = -1;
9240 :
9241 74 : if (tablespaces)
9242 74 : *tablespaces = lappend(*tablespaces, ti);
9243 : }
9244 330 : FreeDir(tblspcdir);
9245 :
9246 330 : state->starttime = (pg_time_t) time(NULL);
9247 : }
9248 330 : PG_END_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9249 :
9250 330 : state->started_in_recovery = backup_started_in_recovery;
9251 :
9252 : /*
9253 : * Mark that the start phase has correctly finished for the backup.
9254 : */
9255 330 : sessionBackupState = SESSION_BACKUP_RUNNING;
9256 330 : }
9257 :
9258 : /*
9259 : * Utility routine to fetch the session-level status of a backup running.
9260 : */
9261 : SessionBackupState
9262 372 : get_backup_status(void)
9263 : {
9264 372 : return sessionBackupState;
9265 : }
9266 :
9267 : /*
9268 : * do_pg_backup_stop
9269 : *
9270 : * Utility function called at the end of an online backup. It creates history
9271 : * file (if required), resets sessionBackupState and so on. It can optionally
9272 : * wait for WAL segments to be archived.
9273 : *
9274 : * "state" is filled with the information necessary to restore from this
9275 : * backup with its stop LSN (stoppoint), its timeline ID (stoptli), etc.
9276 : *
9277 : * It is the responsibility of the caller of this function to verify the
9278 : * permissions of the calling user!
9279 : */
9280 : void
9281 318 : do_pg_backup_stop(BackupState *state, bool waitforarchive)
9282 : {
9283 318 : bool backup_stopped_in_recovery = false;
9284 : char histfilepath[MAXPGPATH];
9285 : char lastxlogfilename[MAXFNAMELEN];
9286 : char histfilename[MAXFNAMELEN];
9287 : XLogSegNo _logSegNo;
9288 : FILE *fp;
9289 : int seconds_before_warning;
9290 318 : int waits = 0;
9291 318 : bool reported_waiting = false;
9292 :
9293 : Assert(state != NULL);
9294 :
9295 318 : backup_stopped_in_recovery = RecoveryInProgress();
9296 :
9297 : /*
9298 : * During recovery, we don't need to check WAL level. Because, if WAL
9299 : * level is not sufficient, it's impossible to get here during recovery.
9300 : */
9301 318 : if (!backup_stopped_in_recovery && !XLogIsNeeded())
9302 0 : ereport(ERROR,
9303 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9304 : errmsg("WAL level not sufficient for making an online backup"),
9305 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9306 :
9307 : /*
9308 : * OK to update backup counter and session-level lock.
9309 : *
9310 : * Note that CHECK_FOR_INTERRUPTS() must not occur while updating them,
9311 : * otherwise they can be updated inconsistently, which might cause
9312 : * do_pg_abort_backup() to fail.
9313 : */
9314 318 : WALInsertLockAcquireExclusive();
9315 :
9316 : /*
9317 : * It is expected that each do_pg_backup_start() call is matched by
9318 : * exactly one do_pg_backup_stop() call.
9319 : */
9320 : Assert(XLogCtl->Insert.runningBackups > 0);
9321 318 : XLogCtl->Insert.runningBackups--;
9322 :
9323 : /*
9324 : * Clean up session-level lock.
9325 : *
9326 : * You might think that WALInsertLockRelease() can be called before
9327 : * cleaning up session-level lock because session-level lock doesn't need
9328 : * to be protected with WAL insertion lock. But since
9329 : * CHECK_FOR_INTERRUPTS() can occur in it, session-level lock must be
9330 : * cleaned up before it.
9331 : */
9332 318 : sessionBackupState = SESSION_BACKUP_NONE;
9333 :
9334 318 : WALInsertLockRelease();
9335 :
9336 : /*
9337 : * If we are taking an online backup from the standby, we confirm that the
9338 : * standby has not been promoted during the backup.
9339 : */
9340 318 : if (state->started_in_recovery && !backup_stopped_in_recovery)
9341 0 : ereport(ERROR,
9342 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9343 : errmsg("the standby was promoted during online backup"),
9344 : errhint("This means that the backup being taken is corrupt "
9345 : "and should not be used. "
9346 : "Try taking another online backup.")));
9347 :
9348 : /*
9349 : * During recovery, we don't write an end-of-backup record. We assume that
9350 : * pg_control was backed up last and its minimum recovery point can be
9351 : * available as the backup end location. Since we don't have an
9352 : * end-of-backup record, we use the pg_control value to check whether
9353 : * we've reached the end of backup when starting recovery from this
9354 : * backup. We have no way of checking if pg_control wasn't backed up last
9355 : * however.
9356 : *
9357 : * We don't force a switch to new WAL file but it is still possible to
9358 : * wait for all the required files to be archived if waitforarchive is
9359 : * true. This is okay if we use the backup to start a standby and fetch
9360 : * the missing WAL using streaming replication. But in the case of an
9361 : * archive recovery, a user should set waitforarchive to true and wait for
9362 : * them to be archived to ensure that all the required files are
9363 : * available.
9364 : *
9365 : * We return the current minimum recovery point as the backup end
9366 : * location. Note that it can be greater than the exact backup end
9367 : * location if the minimum recovery point is updated after the backup of
9368 : * pg_control. This is harmless for current uses.
9369 : *
9370 : * XXX currently a backup history file is for informational and debug
9371 : * purposes only. It's not essential for an online backup. Furthermore,
9372 : * even if it's created, it will not be archived during recovery because
9373 : * an archiver is not invoked. So it doesn't seem worthwhile to write a
9374 : * backup history file during recovery.
9375 : */
9376 318 : if (backup_stopped_in_recovery)
9377 : {
9378 : XLogRecPtr recptr;
9379 :
9380 : /*
9381 : * Check to see if all WAL replayed during online backup contain
9382 : * full-page writes.
9383 : */
9384 16 : SpinLockAcquire(&XLogCtl->info_lck);
9385 16 : recptr = XLogCtl->lastFpwDisableRecPtr;
9386 16 : SpinLockRelease(&XLogCtl->info_lck);
9387 :
9388 16 : if (state->startpoint <= recptr)
9389 0 : ereport(ERROR,
9390 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9391 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9392 : "during online backup"),
9393 : errhint("This means that the backup being taken on the standby "
9394 : "is corrupt and should not be used. "
9395 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9396 : "and then try an online backup again.")));
9397 :
9398 :
9399 16 : LWLockAcquire(ControlFileLock, LW_SHARED);
9400 16 : state->stoppoint = ControlFile->minRecoveryPoint;
9401 16 : state->stoptli = ControlFile->minRecoveryPointTLI;
9402 16 : LWLockRelease(ControlFileLock);
9403 : }
9404 : else
9405 : {
9406 : char *history_file;
9407 :
9408 : /*
9409 : * Write the backup-end xlog record
9410 : */
9411 302 : XLogBeginInsert();
9412 302 : XLogRegisterData(&state->startpoint,
9413 : sizeof(state->startpoint));
9414 302 : state->stoppoint = XLogInsert(RM_XLOG_ID, XLOG_BACKUP_END);
9415 :
9416 : /*
9417 : * Given that we're not in recovery, InsertTimeLineID is set and can't
9418 : * change, so we can read it without a lock.
9419 : */
9420 302 : state->stoptli = XLogCtl->InsertTimeLineID;
9421 :
9422 : /*
9423 : * Force a switch to a new xlog segment file, so that the backup is
9424 : * valid as soon as archiver moves out the current segment file.
9425 : */
9426 302 : RequestXLogSwitch(false);
9427 :
9428 302 : state->stoptime = (pg_time_t) time(NULL);
9429 :
9430 : /*
9431 : * Write the backup history file
9432 : */
9433 302 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9434 302 : BackupHistoryFilePath(histfilepath, state->stoptli, _logSegNo,
9435 : state->startpoint, wal_segment_size);
9436 302 : fp = AllocateFile(histfilepath, "w");
9437 302 : if (!fp)
9438 0 : ereport(ERROR,
9439 : (errcode_for_file_access(),
9440 : errmsg("could not create file \"%s\": %m",
9441 : histfilepath)));
9442 :
9443 : /* Build and save the contents of the backup history file */
9444 302 : history_file = build_backup_content(state, true);
9445 302 : fprintf(fp, "%s", history_file);
9446 302 : pfree(history_file);
9447 :
9448 302 : if (fflush(fp) || ferror(fp) || FreeFile(fp))
9449 0 : ereport(ERROR,
9450 : (errcode_for_file_access(),
9451 : errmsg("could not write file \"%s\": %m",
9452 : histfilepath)));
9453 :
9454 : /*
9455 : * Clean out any no-longer-needed history files. As a side effect,
9456 : * this will post a .ready file for the newly created history file,
9457 : * notifying the archiver that history file may be archived
9458 : * immediately.
9459 : */
9460 302 : CleanupBackupHistory();
9461 : }
9462 :
9463 : /*
9464 : * If archiving is enabled, wait for all the required WAL files to be
9465 : * archived before returning. If archiving isn't enabled, the required WAL
9466 : * needs to be transported via streaming replication (hopefully with
9467 : * wal_keep_size set high enough), or some more exotic mechanism like
9468 : * polling and copying files from pg_wal with script. We have no knowledge
9469 : * of those mechanisms, so it's up to the user to ensure that he gets all
9470 : * the required WAL.
9471 : *
9472 : * We wait until both the last WAL file filled during backup and the
9473 : * history file have been archived, and assume that the alphabetic sorting
9474 : * property of the WAL files ensures any earlier WAL files are safely
9475 : * archived as well.
9476 : *
9477 : * We wait forever, since archive_command is supposed to work and we
9478 : * assume the admin wanted his backup to work completely. If you don't
9479 : * wish to wait, then either waitforarchive should be passed in as false,
9480 : * or you can set statement_timeout. Also, some notices are issued to
9481 : * clue in anyone who might be doing this interactively.
9482 : */
9483 :
9484 318 : if (waitforarchive &&
9485 20 : ((!backup_stopped_in_recovery && XLogArchivingActive()) ||
9486 2 : (backup_stopped_in_recovery && XLogArchivingAlways())))
9487 : {
9488 8 : XLByteToPrevSeg(state->stoppoint, _logSegNo, wal_segment_size);
9489 8 : XLogFileName(lastxlogfilename, state->stoptli, _logSegNo,
9490 : wal_segment_size);
9491 :
9492 8 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9493 8 : BackupHistoryFileName(histfilename, state->stoptli, _logSegNo,
9494 : state->startpoint, wal_segment_size);
9495 :
9496 8 : seconds_before_warning = 60;
9497 8 : waits = 0;
9498 :
9499 26 : while (XLogArchiveIsBusy(lastxlogfilename) ||
9500 8 : XLogArchiveIsBusy(histfilename))
9501 : {
9502 10 : CHECK_FOR_INTERRUPTS();
9503 :
9504 10 : if (!reported_waiting && waits > 5)
9505 : {
9506 0 : ereport(NOTICE,
9507 : (errmsg("base backup done, waiting for required WAL segments to be archived")));
9508 0 : reported_waiting = true;
9509 : }
9510 :
9511 10 : (void) WaitLatch(MyLatch,
9512 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
9513 : 1000L,
9514 : WAIT_EVENT_BACKUP_WAIT_WAL_ARCHIVE);
9515 10 : ResetLatch(MyLatch);
9516 :
9517 10 : if (++waits >= seconds_before_warning)
9518 : {
9519 0 : seconds_before_warning *= 2; /* This wraps in >10 years... */
9520 0 : ereport(WARNING,
9521 : (errmsg("still waiting for all required WAL segments to be archived (%d seconds elapsed)",
9522 : waits),
9523 : errhint("Check that your \"archive_command\" is executing properly. "
9524 : "You can safely cancel this backup, "
9525 : "but the database backup will not be usable without all the WAL segments.")));
9526 : }
9527 : }
9528 :
9529 8 : ereport(NOTICE,
9530 : (errmsg("all required WAL segments have been archived")));
9531 : }
9532 310 : else if (waitforarchive)
9533 12 : ereport(NOTICE,
9534 : (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
9535 318 : }
9536 :
9537 :
9538 : /*
9539 : * do_pg_abort_backup: abort a running backup
9540 : *
9541 : * This does just the most basic steps of do_pg_backup_stop(), by taking the
9542 : * system out of backup mode, thus making it a lot more safe to call from
9543 : * an error handler.
9544 : *
9545 : * 'arg' indicates that it's being called during backup setup; so
9546 : * sessionBackupState has not been modified yet, but runningBackups has
9547 : * already been incremented. When it's false, then it's invoked as a
9548 : * before_shmem_exit handler, and therefore we must not change state
9549 : * unless sessionBackupState indicates that a backup is actually running.
9550 : *
9551 : * NB: This gets used as a PG_ENSURE_ERROR_CLEANUP callback and
9552 : * before_shmem_exit handler, hence the odd-looking signature.
9553 : */
9554 : void
9555 16 : do_pg_abort_backup(int code, Datum arg)
9556 : {
9557 16 : bool during_backup_start = DatumGetBool(arg);
9558 :
9559 : /* If called during backup start, there shouldn't be one already running */
9560 : Assert(!during_backup_start || sessionBackupState == SESSION_BACKUP_NONE);
9561 :
9562 16 : if (during_backup_start || sessionBackupState != SESSION_BACKUP_NONE)
9563 : {
9564 12 : WALInsertLockAcquireExclusive();
9565 : Assert(XLogCtl->Insert.runningBackups > 0);
9566 12 : XLogCtl->Insert.runningBackups--;
9567 :
9568 12 : sessionBackupState = SESSION_BACKUP_NONE;
9569 12 : WALInsertLockRelease();
9570 :
9571 12 : if (!during_backup_start)
9572 12 : ereport(WARNING,
9573 : errmsg("aborting backup due to backend exiting before pg_backup_stop was called"));
9574 : }
9575 16 : }
9576 :
9577 : /*
9578 : * Register a handler that will warn about unterminated backups at end of
9579 : * session, unless this has already been done.
9580 : */
9581 : void
9582 8 : register_persistent_abort_backup_handler(void)
9583 : {
9584 : static bool already_done = false;
9585 :
9586 8 : if (already_done)
9587 2 : return;
9588 6 : before_shmem_exit(do_pg_abort_backup, BoolGetDatum(false));
9589 6 : already_done = true;
9590 : }
9591 :
9592 : /*
9593 : * Get latest WAL insert pointer
9594 : */
9595 : XLogRecPtr
9596 4034 : GetXLogInsertRecPtr(void)
9597 : {
9598 4034 : XLogCtlInsert *Insert = &XLogCtl->Insert;
9599 : uint64 current_bytepos;
9600 :
9601 4034 : SpinLockAcquire(&Insert->insertpos_lck);
9602 4034 : current_bytepos = Insert->CurrBytePos;
9603 4034 : SpinLockRelease(&Insert->insertpos_lck);
9604 :
9605 4034 : return XLogBytePosToRecPtr(current_bytepos);
9606 : }
9607 :
9608 : /*
9609 : * Get latest WAL write pointer
9610 : */
9611 : XLogRecPtr
9612 11854 : GetXLogWriteRecPtr(void)
9613 : {
9614 11854 : RefreshXLogWriteResult(LogwrtResult);
9615 :
9616 11854 : return LogwrtResult.Write;
9617 : }
9618 :
9619 : /*
9620 : * Returns the redo pointer of the last checkpoint or restartpoint. This is
9621 : * the oldest point in WAL that we still need, if we have to restart recovery.
9622 : */
9623 : void
9624 790 : GetOldestRestartPoint(XLogRecPtr *oldrecptr, TimeLineID *oldtli)
9625 : {
9626 790 : LWLockAcquire(ControlFileLock, LW_SHARED);
9627 790 : *oldrecptr = ControlFile->checkPointCopy.redo;
9628 790 : *oldtli = ControlFile->checkPointCopy.ThisTimeLineID;
9629 790 : LWLockRelease(ControlFileLock);
9630 790 : }
9631 :
9632 : /* Thin wrapper around ShutdownWalRcv(). */
9633 : void
9634 1978 : XLogShutdownWalRcv(void)
9635 : {
9636 : Assert(AmStartupProcess() || !IsUnderPostmaster);
9637 :
9638 1978 : ShutdownWalRcv();
9639 1978 : ResetInstallXLogFileSegmentActive();
9640 1978 : }
9641 :
9642 : /* Enable WAL file recycling and preallocation. */
9643 : void
9644 2368 : SetInstallXLogFileSegmentActive(void)
9645 : {
9646 2368 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9647 2368 : XLogCtl->InstallXLogFileSegmentActive = true;
9648 2368 : LWLockRelease(ControlFileLock);
9649 2368 : }
9650 :
9651 : /* Disable WAL file recycling and preallocation. */
9652 : void
9653 2290 : ResetInstallXLogFileSegmentActive(void)
9654 : {
9655 2290 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9656 2290 : XLogCtl->InstallXLogFileSegmentActive = false;
9657 2290 : LWLockRelease(ControlFileLock);
9658 2290 : }
9659 :
9660 : bool
9661 0 : IsInstallXLogFileSegmentActive(void)
9662 : {
9663 : bool result;
9664 :
9665 0 : LWLockAcquire(ControlFileLock, LW_SHARED);
9666 0 : result = XLogCtl->InstallXLogFileSegmentActive;
9667 0 : LWLockRelease(ControlFileLock);
9668 :
9669 0 : return result;
9670 : }
9671 :
9672 : /*
9673 : * Update the WalWriterSleeping flag.
9674 : */
9675 : void
9676 1040 : SetWalWriterSleeping(bool sleeping)
9677 : {
9678 1040 : SpinLockAcquire(&XLogCtl->info_lck);
9679 1040 : XLogCtl->WalWriterSleeping = sleeping;
9680 1040 : SpinLockRelease(&XLogCtl->info_lck);
9681 1040 : }
|
