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) 1996-2025, PostgreSQL Global Development Group
32 : * Portions Copyright (c) 1994, Regents of the University of California
33 : *
34 : * src/backend/access/transam/xlog.c
35 : *
36 : *-------------------------------------------------------------------------
37 : */
38 :
39 : #include "postgres.h"
40 :
41 : #include <ctype.h>
42 : #include <math.h>
43 : #include <time.h>
44 : #include <fcntl.h>
45 : #include <sys/stat.h>
46 : #include <sys/time.h>
47 : #include <unistd.h>
48 :
49 : #include "access/clog.h"
50 : #include "access/commit_ts.h"
51 : #include "access/heaptoast.h"
52 : #include "access/multixact.h"
53 : #include "access/rewriteheap.h"
54 : #include "access/subtrans.h"
55 : #include "access/timeline.h"
56 : #include "access/transam.h"
57 : #include "access/twophase.h"
58 : #include "access/xact.h"
59 : #include "access/xlog_internal.h"
60 : #include "access/xlogarchive.h"
61 : #include "access/xloginsert.h"
62 : #include "access/xlogreader.h"
63 : #include "access/xlogrecovery.h"
64 : #include "access/xlogutils.h"
65 : #include "backup/basebackup.h"
66 : #include "catalog/catversion.h"
67 : #include "catalog/pg_control.h"
68 : #include "catalog/pg_database.h"
69 : #include "common/controldata_utils.h"
70 : #include "common/file_utils.h"
71 : #include "executor/instrument.h"
72 : #include "miscadmin.h"
73 : #include "pg_trace.h"
74 : #include "pgstat.h"
75 : #include "port/atomics.h"
76 : #include "postmaster/bgwriter.h"
77 : #include "postmaster/startup.h"
78 : #include "postmaster/walsummarizer.h"
79 : #include "postmaster/walwriter.h"
80 : #include "replication/origin.h"
81 : #include "replication/slot.h"
82 : #include "replication/snapbuild.h"
83 : #include "replication/walreceiver.h"
84 : #include "replication/walsender.h"
85 : #include "storage/bufmgr.h"
86 : #include "storage/fd.h"
87 : #include "storage/ipc.h"
88 : #include "storage/large_object.h"
89 : #include "storage/latch.h"
90 : #include "storage/predicate.h"
91 : #include "storage/proc.h"
92 : #include "storage/procarray.h"
93 : #include "storage/reinit.h"
94 : #include "storage/spin.h"
95 : #include "storage/sync.h"
96 : #include "utils/guc_hooks.h"
97 : #include "utils/guc_tables.h"
98 : #include "utils/injection_point.h"
99 : #include "utils/pgstat_internal.h"
100 : #include "utils/ps_status.h"
101 : #include "utils/relmapper.h"
102 : #include "utils/snapmgr.h"
103 : #include "utils/timeout.h"
104 : #include "utils/timestamp.h"
105 : #include "utils/varlena.h"
106 :
107 : #ifdef WAL_DEBUG
108 : #include "utils/memutils.h"
109 : #endif
110 :
111 : /* timeline ID to be used when bootstrapping */
112 : #define BootstrapTimeLineID 1
113 :
114 : /* User-settable parameters */
115 : int max_wal_size_mb = 1024; /* 1 GB */
116 : int min_wal_size_mb = 80; /* 80 MB */
117 : int wal_keep_size_mb = 0;
118 : int XLOGbuffers = -1;
119 : int XLogArchiveTimeout = 0;
120 : int XLogArchiveMode = ARCHIVE_MODE_OFF;
121 : char *XLogArchiveCommand = NULL;
122 : bool EnableHotStandby = false;
123 : bool fullPageWrites = true;
124 : bool wal_log_hints = false;
125 : int wal_compression = WAL_COMPRESSION_NONE;
126 : char *wal_consistency_checking_string = NULL;
127 : bool *wal_consistency_checking = NULL;
128 : bool wal_init_zero = true;
129 : bool wal_recycle = true;
130 : bool log_checkpoints = true;
131 : int wal_sync_method = DEFAULT_WAL_SYNC_METHOD;
132 : int wal_level = WAL_LEVEL_REPLICA;
133 : int CommitDelay = 0; /* precommit delay in microseconds */
134 : int CommitSiblings = 5; /* # concurrent xacts needed to sleep */
135 : int wal_retrieve_retry_interval = 5000;
136 : int max_slot_wal_keep_size_mb = -1;
137 : int wal_decode_buffer_size = 512 * 1024;
138 : bool track_wal_io_timing = false;
139 :
140 : #ifdef WAL_DEBUG
141 : bool XLOG_DEBUG = false;
142 : #endif
143 :
144 : int wal_segment_size = DEFAULT_XLOG_SEG_SIZE;
145 :
146 : /*
147 : * Number of WAL insertion locks to use. A higher value allows more insertions
148 : * to happen concurrently, but adds some CPU overhead to flushing the WAL,
149 : * which needs to iterate all the locks.
150 : */
151 : #define NUM_XLOGINSERT_LOCKS 8
152 :
153 : /*
154 : * Max distance from last checkpoint, before triggering a new xlog-based
155 : * checkpoint.
156 : */
157 : int CheckPointSegments;
158 :
159 : /* Estimated distance between checkpoints, in bytes */
160 : static double CheckPointDistanceEstimate = 0;
161 : static double PrevCheckPointDistance = 0;
162 :
163 : /*
164 : * Track whether there were any deferred checks for custom resource managers
165 : * specified in wal_consistency_checking.
166 : */
167 : static bool check_wal_consistency_checking_deferred = false;
168 :
169 : /*
170 : * GUC support
171 : */
172 : const struct config_enum_entry wal_sync_method_options[] = {
173 : {"fsync", WAL_SYNC_METHOD_FSYNC, false},
174 : #ifdef HAVE_FSYNC_WRITETHROUGH
175 : {"fsync_writethrough", WAL_SYNC_METHOD_FSYNC_WRITETHROUGH, false},
176 : #endif
177 : {"fdatasync", WAL_SYNC_METHOD_FDATASYNC, false},
178 : #ifdef O_SYNC
179 : {"open_sync", WAL_SYNC_METHOD_OPEN, false},
180 : #endif
181 : #ifdef O_DSYNC
182 : {"open_datasync", WAL_SYNC_METHOD_OPEN_DSYNC, false},
183 : #endif
184 : {NULL, 0, false}
185 : };
186 :
187 :
188 : /*
189 : * Although only "on", "off", and "always" are documented,
190 : * we accept all the likely variants of "on" and "off".
191 : */
192 : const struct config_enum_entry archive_mode_options[] = {
193 : {"always", ARCHIVE_MODE_ALWAYS, false},
194 : {"on", ARCHIVE_MODE_ON, false},
195 : {"off", ARCHIVE_MODE_OFF, false},
196 : {"true", ARCHIVE_MODE_ON, true},
197 : {"false", ARCHIVE_MODE_OFF, true},
198 : {"yes", ARCHIVE_MODE_ON, true},
199 : {"no", ARCHIVE_MODE_OFF, true},
200 : {"1", ARCHIVE_MODE_ON, true},
201 : {"0", ARCHIVE_MODE_OFF, true},
202 : {NULL, 0, false}
203 : };
204 :
205 : /*
206 : * Statistics for current checkpoint are collected in this global struct.
207 : * Because only the checkpointer or a stand-alone backend can perform
208 : * checkpoints, this will be unused in normal backends.
209 : */
210 : CheckpointStatsData CheckpointStats;
211 :
212 : /*
213 : * During recovery, lastFullPageWrites keeps track of full_page_writes that
214 : * the replayed WAL records indicate. It's initialized with full_page_writes
215 : * that the recovery starting checkpoint record indicates, and then updated
216 : * each time XLOG_FPW_CHANGE record is replayed.
217 : */
218 : static bool lastFullPageWrites;
219 :
220 : /*
221 : * Local copy of the state tracked by SharedRecoveryState in shared memory,
222 : * It is false if SharedRecoveryState is RECOVERY_STATE_DONE. True actually
223 : * means "not known, need to check the shared state".
224 : */
225 : static bool LocalRecoveryInProgress = true;
226 :
227 : /*
228 : * Local state for XLogInsertAllowed():
229 : * 1: unconditionally allowed to insert XLOG
230 : * 0: unconditionally not allowed to insert XLOG
231 : * -1: must check RecoveryInProgress(); disallow until it is false
232 : * Most processes start with -1 and transition to 1 after seeing that recovery
233 : * is not in progress. But we can also force the value for special cases.
234 : * The coding in XLogInsertAllowed() depends on the first two of these states
235 : * being numerically the same as bool true and false.
236 : */
237 : static int LocalXLogInsertAllowed = -1;
238 :
239 : /*
240 : * ProcLastRecPtr points to the start of the last XLOG record inserted by the
241 : * current backend. It is updated for all inserts. XactLastRecEnd points to
242 : * end+1 of the last record, and is reset when we end a top-level transaction,
243 : * or start a new one; so it can be used to tell if the current transaction has
244 : * created any XLOG records.
245 : *
246 : * While in parallel mode, this may not be fully up to date. When committing,
247 : * a transaction can assume this covers all xlog records written either by the
248 : * user backend or by any parallel worker which was present at any point during
249 : * the transaction. But when aborting, or when still in parallel mode, other
250 : * parallel backends may have written WAL records at later LSNs than the value
251 : * stored here. The parallel leader advances its own copy, when necessary,
252 : * in WaitForParallelWorkersToFinish.
253 : */
254 : XLogRecPtr ProcLastRecPtr = InvalidXLogRecPtr;
255 : XLogRecPtr XactLastRecEnd = InvalidXLogRecPtr;
256 : XLogRecPtr XactLastCommitEnd = InvalidXLogRecPtr;
257 :
258 : /*
259 : * RedoRecPtr is this backend's local copy of the REDO record pointer
260 : * (which is almost but not quite the same as a pointer to the most recent
261 : * CHECKPOINT record). We update this from the shared-memory copy,
262 : * XLogCtl->Insert.RedoRecPtr, whenever we can safely do so (ie, when we
263 : * hold an insertion lock). See XLogInsertRecord for details. We are also
264 : * allowed to update from XLogCtl->RedoRecPtr if we hold the info_lck;
265 : * see GetRedoRecPtr.
266 : *
267 : * NB: Code that uses this variable must be prepared not only for the
268 : * possibility that it may be arbitrarily out of date, but also for the
269 : * possibility that it might be set to InvalidXLogRecPtr. We used to
270 : * initialize it as a side effect of the first call to RecoveryInProgress(),
271 : * which meant that most code that might use it could assume that it had a
272 : * real if perhaps stale value. That's no longer the case.
273 : */
274 : static XLogRecPtr RedoRecPtr;
275 :
276 : /*
277 : * doPageWrites is this backend's local copy of (fullPageWrites ||
278 : * runningBackups > 0). It is used together with RedoRecPtr to decide whether
279 : * a full-page image of a page need to be taken.
280 : *
281 : * NB: Initially this is false, and there's no guarantee that it will be
282 : * initialized to any other value before it is first used. Any code that
283 : * makes use of it must recheck the value after obtaining a WALInsertLock,
284 : * and respond appropriately if it turns out that the previous value wasn't
285 : * accurate.
286 : */
287 : static bool doPageWrites;
288 :
289 : /*----------
290 : * Shared-memory data structures for XLOG control
291 : *
292 : * LogwrtRqst indicates a byte position that we need to write and/or fsync
293 : * the log up to (all records before that point must be written or fsynced).
294 : * The positions already written/fsynced are maintained in logWriteResult
295 : * and logFlushResult using atomic access.
296 : * In addition to the shared variable, each backend has a private copy of
297 : * both in LogwrtResult, which is updated when convenient.
298 : *
299 : * The request bookkeeping is simpler: there is a shared XLogCtl->LogwrtRqst
300 : * (protected by info_lck), but we don't need to cache any copies of it.
301 : *
302 : * info_lck is only held long enough to read/update the protected variables,
303 : * so it's a plain spinlock. The other locks are held longer (potentially
304 : * over I/O operations), so we use LWLocks for them. These locks are:
305 : *
306 : * WALBufMappingLock: must be held to replace a page in the WAL buffer cache.
307 : * It is only held while initializing and changing the mapping. If the
308 : * contents of the buffer being replaced haven't been written yet, the mapping
309 : * lock is released while the write is done, and reacquired afterwards.
310 : *
311 : * WALWriteLock: must be held to write WAL buffers to disk (XLogWrite or
312 : * XLogFlush).
313 : *
314 : * ControlFileLock: must be held to read/update control file or create
315 : * new log file.
316 : *
317 : *----------
318 : */
319 :
320 : typedef struct XLogwrtRqst
321 : {
322 : XLogRecPtr Write; /* last byte + 1 to write out */
323 : XLogRecPtr Flush; /* last byte + 1 to flush */
324 : } XLogwrtRqst;
325 :
326 : typedef struct XLogwrtResult
327 : {
328 : XLogRecPtr Write; /* last byte + 1 written out */
329 : XLogRecPtr Flush; /* last byte + 1 flushed */
330 : } XLogwrtResult;
331 :
332 : /*
333 : * Inserting to WAL is protected by a small fixed number of WAL insertion
334 : * locks. To insert to the WAL, you must hold one of the locks - it doesn't
335 : * matter which one. To lock out other concurrent insertions, you must hold
336 : * of them. Each WAL insertion lock consists of a lightweight lock, plus an
337 : * indicator of how far the insertion has progressed (insertingAt).
338 : *
339 : * The insertingAt values are read when a process wants to flush WAL from
340 : * the in-memory buffers to disk, to check that all the insertions to the
341 : * region the process is about to write out have finished. You could simply
342 : * wait for all currently in-progress insertions to finish, but the
343 : * insertingAt indicator allows you to ignore insertions to later in the WAL,
344 : * so that you only wait for the insertions that are modifying the buffers
345 : * you're about to write out.
346 : *
347 : * This isn't just an optimization. If all the WAL buffers are dirty, an
348 : * inserter that's holding a WAL insert lock might need to evict an old WAL
349 : * buffer, which requires flushing the WAL. If it's possible for an inserter
350 : * to block on another inserter unnecessarily, deadlock can arise when two
351 : * inserters holding a WAL insert lock wait for each other to finish their
352 : * insertion.
353 : *
354 : * Small WAL records that don't cross a page boundary never update the value,
355 : * the WAL record is just copied to the page and the lock is released. But
356 : * to avoid the deadlock-scenario explained above, the indicator is always
357 : * updated before sleeping while holding an insertion lock.
358 : *
359 : * lastImportantAt contains the LSN of the last important WAL record inserted
360 : * using a given lock. This value is used to detect if there has been
361 : * important WAL activity since the last time some action, like a checkpoint,
362 : * was performed - allowing to not repeat the action if not. The LSN is
363 : * updated for all insertions, unless the XLOG_MARK_UNIMPORTANT flag was
364 : * set. lastImportantAt is never cleared, only overwritten by the LSN of newer
365 : * records. Tracking the WAL activity directly in WALInsertLock has the
366 : * advantage of not needing any additional locks to update the value.
367 : */
368 : typedef struct
369 : {
370 : LWLock lock;
371 : pg_atomic_uint64 insertingAt;
372 : XLogRecPtr lastImportantAt;
373 : } WALInsertLock;
374 :
375 : /*
376 : * All the WAL insertion locks are allocated as an array in shared memory. We
377 : * force the array stride to be a power of 2, which saves a few cycles in
378 : * indexing, but more importantly also ensures that individual slots don't
379 : * cross cache line boundaries. (Of course, we have to also ensure that the
380 : * array start address is suitably aligned.)
381 : */
382 : typedef union WALInsertLockPadded
383 : {
384 : WALInsertLock l;
385 : char pad[PG_CACHE_LINE_SIZE];
386 : } WALInsertLockPadded;
387 :
388 : /*
389 : * Session status of running backup, used for sanity checks in SQL-callable
390 : * functions to start and stop backups.
391 : */
392 : static SessionBackupState sessionBackupState = SESSION_BACKUP_NONE;
393 :
394 : /*
395 : * Shared state data for WAL insertion.
396 : */
397 : typedef struct XLogCtlInsert
398 : {
399 : slock_t insertpos_lck; /* protects CurrBytePos and PrevBytePos */
400 :
401 : /*
402 : * CurrBytePos is the end of reserved WAL. The next record will be
403 : * inserted at that position. PrevBytePos is the start position of the
404 : * previously inserted (or rather, reserved) record - it is copied to the
405 : * prev-link of the next record. These are stored as "usable byte
406 : * positions" rather than XLogRecPtrs (see XLogBytePosToRecPtr()).
407 : */
408 : uint64 CurrBytePos;
409 : uint64 PrevBytePos;
410 :
411 : /*
412 : * Make sure the above heavily-contended spinlock and byte positions are
413 : * on their own cache line. In particular, the RedoRecPtr and full page
414 : * write variables below should be on a different cache line. They are
415 : * read on every WAL insertion, but updated rarely, and we don't want
416 : * those reads to steal the cache line containing Curr/PrevBytePos.
417 : */
418 : char pad[PG_CACHE_LINE_SIZE];
419 :
420 : /*
421 : * fullPageWrites is the authoritative value used by all backends to
422 : * determine whether to write full-page image to WAL. This shared value,
423 : * instead of the process-local fullPageWrites, is required because, when
424 : * full_page_writes is changed by SIGHUP, we must WAL-log it before it
425 : * actually affects WAL-logging by backends. Checkpointer sets at startup
426 : * or after SIGHUP.
427 : *
428 : * To read these fields, you must hold an insertion lock. To modify them,
429 : * you must hold ALL the locks.
430 : */
431 : XLogRecPtr RedoRecPtr; /* current redo point for insertions */
432 : bool fullPageWrites;
433 :
434 : /*
435 : * runningBackups is a counter indicating the number of backups currently
436 : * in progress. lastBackupStart is the latest checkpoint redo location
437 : * used as a starting point for an online backup.
438 : */
439 : int runningBackups;
440 : XLogRecPtr lastBackupStart;
441 :
442 : /*
443 : * WAL insertion locks.
444 : */
445 : WALInsertLockPadded *WALInsertLocks;
446 : } XLogCtlInsert;
447 :
448 : /*
449 : * Total shared-memory state for XLOG.
450 : */
451 : typedef struct XLogCtlData
452 : {
453 : XLogCtlInsert Insert;
454 :
455 : /* Protected by info_lck: */
456 : XLogwrtRqst LogwrtRqst;
457 : XLogRecPtr RedoRecPtr; /* a recent copy of Insert->RedoRecPtr */
458 : XLogRecPtr asyncXactLSN; /* LSN of newest async commit/abort */
459 : XLogRecPtr replicationSlotMinLSN; /* oldest LSN needed by any slot */
460 :
461 : XLogSegNo lastRemovedSegNo; /* latest removed/recycled XLOG segment */
462 :
463 : /* Fake LSN counter, for unlogged relations. */
464 : pg_atomic_uint64 unloggedLSN;
465 :
466 : /* Time and LSN of last xlog segment switch. Protected by WALWriteLock. */
467 : pg_time_t lastSegSwitchTime;
468 : XLogRecPtr lastSegSwitchLSN;
469 :
470 : /* These are accessed using atomics -- info_lck not needed */
471 : pg_atomic_uint64 logInsertResult; /* last byte + 1 inserted to buffers */
472 : pg_atomic_uint64 logWriteResult; /* last byte + 1 written out */
473 : pg_atomic_uint64 logFlushResult; /* last byte + 1 flushed */
474 :
475 : /*
476 : * Latest initialized page in the cache (last byte position + 1).
477 : *
478 : * To change the identity of a buffer (and InitializedUpTo), you need to
479 : * hold WALBufMappingLock. To change the identity of a buffer that's
480 : * still dirty, the old page needs to be written out first, and for that
481 : * you need WALWriteLock, and you need to ensure that there are no
482 : * in-progress insertions to the page by calling
483 : * WaitXLogInsertionsToFinish().
484 : */
485 : XLogRecPtr InitializedUpTo;
486 :
487 : /*
488 : * These values do not change after startup, although the pointed-to pages
489 : * and xlblocks values certainly do. xlblocks values are protected by
490 : * WALBufMappingLock.
491 : */
492 : char *pages; /* buffers for unwritten XLOG pages */
493 : pg_atomic_uint64 *xlblocks; /* 1st byte ptr-s + XLOG_BLCKSZ */
494 : int XLogCacheBlck; /* highest allocated xlog buffer index */
495 :
496 : /*
497 : * InsertTimeLineID is the timeline into which new WAL is being inserted
498 : * and flushed. It is zero during recovery, and does not change once set.
499 : *
500 : * If we create a new timeline when the system was started up,
501 : * PrevTimeLineID is the old timeline's ID that we forked off from.
502 : * Otherwise it's equal to InsertTimeLineID.
503 : *
504 : * We set these fields while holding info_lck. Most that reads these
505 : * values knows that recovery is no longer in progress and so can safely
506 : * read the value without a lock, but code that could be run either during
507 : * or after recovery can take info_lck while reading these values.
508 : */
509 : TimeLineID InsertTimeLineID;
510 : TimeLineID PrevTimeLineID;
511 :
512 : /*
513 : * SharedRecoveryState indicates if we're still in crash or archive
514 : * recovery. Protected by info_lck.
515 : */
516 : RecoveryState SharedRecoveryState;
517 :
518 : /*
519 : * InstallXLogFileSegmentActive indicates whether the checkpointer should
520 : * arrange for future segments by recycling and/or PreallocXlogFiles().
521 : * Protected by ControlFileLock. Only the startup process changes it. If
522 : * true, anyone can use InstallXLogFileSegment(). If false, the startup
523 : * process owns the exclusive right to install segments, by reading from
524 : * the archive and possibly replacing existing files.
525 : */
526 : bool InstallXLogFileSegmentActive;
527 :
528 : /*
529 : * WalWriterSleeping indicates whether the WAL writer is currently in
530 : * low-power mode (and hence should be nudged if an async commit occurs).
531 : * Protected by info_lck.
532 : */
533 : bool WalWriterSleeping;
534 :
535 : /*
536 : * During recovery, we keep a copy of the latest checkpoint record here.
537 : * lastCheckPointRecPtr points to start of checkpoint record and
538 : * lastCheckPointEndPtr points to end+1 of checkpoint record. Used by the
539 : * checkpointer when it wants to create a restartpoint.
540 : *
541 : * Protected by info_lck.
542 : */
543 : XLogRecPtr lastCheckPointRecPtr;
544 : XLogRecPtr lastCheckPointEndPtr;
545 : CheckPoint lastCheckPoint;
546 :
547 : /*
548 : * lastFpwDisableRecPtr points to the start of the last replayed
549 : * XLOG_FPW_CHANGE record that instructs full_page_writes is disabled.
550 : */
551 : XLogRecPtr lastFpwDisableRecPtr;
552 :
553 : slock_t info_lck; /* locks shared variables shown above */
554 : } XLogCtlData;
555 :
556 : /*
557 : * Classification of XLogInsertRecord operations.
558 : */
559 : typedef enum
560 : {
561 : WALINSERT_NORMAL,
562 : WALINSERT_SPECIAL_SWITCH,
563 : WALINSERT_SPECIAL_CHECKPOINT
564 : } WalInsertClass;
565 :
566 : static XLogCtlData *XLogCtl = NULL;
567 :
568 : /* a private copy of XLogCtl->Insert.WALInsertLocks, for convenience */
569 : static WALInsertLockPadded *WALInsertLocks = NULL;
570 :
571 : /*
572 : * We maintain an image of pg_control in shared memory.
573 : */
574 : static ControlFileData *ControlFile = NULL;
575 :
576 : /*
577 : * Calculate the amount of space left on the page after 'endptr'. Beware
578 : * multiple evaluation!
579 : */
580 : #define INSERT_FREESPACE(endptr) \
581 : (((endptr) % XLOG_BLCKSZ == 0) ? 0 : (XLOG_BLCKSZ - (endptr) % XLOG_BLCKSZ))
582 :
583 : /* Macro to advance to next buffer index. */
584 : #define NextBufIdx(idx) \
585 : (((idx) == XLogCtl->XLogCacheBlck) ? 0 : ((idx) + 1))
586 :
587 : /*
588 : * XLogRecPtrToBufIdx returns the index of the WAL buffer that holds, or
589 : * would hold if it was in cache, the page containing 'recptr'.
590 : */
591 : #define XLogRecPtrToBufIdx(recptr) \
592 : (((recptr) / XLOG_BLCKSZ) % (XLogCtl->XLogCacheBlck + 1))
593 :
594 : /*
595 : * These are the number of bytes in a WAL page usable for WAL data.
596 : */
597 : #define UsableBytesInPage (XLOG_BLCKSZ - SizeOfXLogShortPHD)
598 :
599 : /*
600 : * Convert values of GUCs measured in megabytes to equiv. segment count.
601 : * Rounds down.
602 : */
603 : #define ConvertToXSegs(x, segsize) XLogMBVarToSegs((x), (segsize))
604 :
605 : /* The number of bytes in a WAL segment usable for WAL data. */
606 : static int UsableBytesInSegment;
607 :
608 : /*
609 : * Private, possibly out-of-date copy of shared LogwrtResult.
610 : * See discussion above.
611 : */
612 : static XLogwrtResult LogwrtResult = {0, 0};
613 :
614 : /*
615 : * Update local copy of shared XLogCtl->log{Write,Flush}Result
616 : *
617 : * It's critical that Flush always trails Write, so the order of the reads is
618 : * important, as is the barrier. See also XLogWrite.
619 : */
620 : #define RefreshXLogWriteResult(_target) \
621 : do { \
622 : _target.Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult); \
623 : pg_read_barrier(); \
624 : _target.Write = pg_atomic_read_u64(&XLogCtl->logWriteResult); \
625 : } while (0)
626 :
627 : /*
628 : * openLogFile is -1 or a kernel FD for an open log file segment.
629 : * openLogSegNo identifies the segment, and openLogTLI the corresponding TLI.
630 : * These variables are only used to write the XLOG, and so will normally refer
631 : * to the active segment.
632 : *
633 : * Note: call Reserve/ReleaseExternalFD to track consumption of this FD.
634 : */
635 : static int openLogFile = -1;
636 : static XLogSegNo openLogSegNo = 0;
637 : static TimeLineID openLogTLI = 0;
638 :
639 : /*
640 : * Local copies of equivalent fields in the control file. When running
641 : * crash recovery, LocalMinRecoveryPoint is set to InvalidXLogRecPtr as we
642 : * expect to replay all the WAL available, and updateMinRecoveryPoint is
643 : * switched to false to prevent any updates while replaying records.
644 : * Those values are kept consistent as long as crash recovery runs.
645 : */
646 : static XLogRecPtr LocalMinRecoveryPoint;
647 : static TimeLineID LocalMinRecoveryPointTLI;
648 : static bool updateMinRecoveryPoint = true;
649 :
650 : /* For WALInsertLockAcquire/Release functions */
651 : static int MyLockNo = 0;
652 : static bool holdingAllLocks = false;
653 :
654 : #ifdef WAL_DEBUG
655 : static MemoryContext walDebugCxt = NULL;
656 : #endif
657 :
658 : static void CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI,
659 : XLogRecPtr EndOfLog,
660 : TimeLineID newTLI);
661 : static void CheckRequiredParameterValues(void);
662 : static void XLogReportParameters(void);
663 : static int LocalSetXLogInsertAllowed(void);
664 : static void CreateEndOfRecoveryRecord(void);
665 : static XLogRecPtr CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn,
666 : XLogRecPtr pagePtr,
667 : TimeLineID newTLI);
668 : static void CheckPointGuts(XLogRecPtr checkPointRedo, int flags);
669 : static void KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo);
670 : static XLogRecPtr XLogGetReplicationSlotMinimumLSN(void);
671 :
672 : static void AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli,
673 : bool opportunistic);
674 : static void XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible);
675 : static bool InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
676 : bool find_free, XLogSegNo max_segno,
677 : TimeLineID tli);
678 : static void XLogFileClose(void);
679 : static void PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli);
680 : static void RemoveTempXlogFiles(void);
681 : static void RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr,
682 : XLogRecPtr endptr, TimeLineID insertTLI);
683 : static void RemoveXlogFile(const struct dirent *segment_de,
684 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
685 : TimeLineID insertTLI);
686 : static void UpdateLastRemovedPtr(char *filename);
687 : static void ValidateXLOGDirectoryStructure(void);
688 : static void CleanupBackupHistory(void);
689 : static void UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force);
690 : static bool PerformRecoveryXLogAction(void);
691 : static void InitControlFile(uint64 sysidentifier, uint32 data_checksum_version);
692 : static void WriteControlFile(void);
693 : static void ReadControlFile(void);
694 : static void UpdateControlFile(void);
695 : static char *str_time(pg_time_t tnow, char *buf, size_t bufsize);
696 :
697 : static int get_sync_bit(int method);
698 :
699 : static void CopyXLogRecordToWAL(int write_len, bool isLogSwitch,
700 : XLogRecData *rdata,
701 : XLogRecPtr StartPos, XLogRecPtr EndPos,
702 : TimeLineID tli);
703 : static void ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos,
704 : XLogRecPtr *EndPos, XLogRecPtr *PrevPtr);
705 : static bool ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos,
706 : XLogRecPtr *PrevPtr);
707 : static XLogRecPtr WaitXLogInsertionsToFinish(XLogRecPtr upto);
708 : static char *GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli);
709 : static XLogRecPtr XLogBytePosToRecPtr(uint64 bytepos);
710 : static XLogRecPtr XLogBytePosToEndRecPtr(uint64 bytepos);
711 : static uint64 XLogRecPtrToBytePos(XLogRecPtr ptr);
712 :
713 : static void WALInsertLockAcquire(void);
714 : static void WALInsertLockAcquireExclusive(void);
715 : static void WALInsertLockRelease(void);
716 : static void WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt);
717 :
718 : /*
719 : * Insert an XLOG record represented by an already-constructed chain of data
720 : * chunks. This is a low-level routine; to construct the WAL record header
721 : * and data, use the higher-level routines in xloginsert.c.
722 : *
723 : * If 'fpw_lsn' is valid, it is the oldest LSN among the pages that this
724 : * WAL record applies to, that were not included in the record as full page
725 : * images. If fpw_lsn <= RedoRecPtr, the function does not perform the
726 : * insertion and returns InvalidXLogRecPtr. The caller can then recalculate
727 : * which pages need a full-page image, and retry. If fpw_lsn is invalid, the
728 : * record is always inserted.
729 : *
730 : * 'flags' gives more in-depth control on the record being inserted. See
731 : * XLogSetRecordFlags() for details.
732 : *
733 : * 'topxid_included' tells whether the top-transaction id is logged along with
734 : * current subtransaction. See XLogRecordAssemble().
735 : *
736 : * The first XLogRecData in the chain must be for the record header, and its
737 : * data must be MAXALIGNed. XLogInsertRecord fills in the xl_prev and
738 : * xl_crc fields in the header, the rest of the header must already be filled
739 : * by the caller.
740 : *
741 : * Returns XLOG pointer to end of record (beginning of next record).
742 : * This can be used as LSN for data pages affected by the logged action.
743 : * (LSN is the XLOG point up to which the XLOG must be flushed to disk
744 : * before the data page can be written out. This implements the basic
745 : * WAL rule "write the log before the data".)
746 : */
747 : XLogRecPtr
748 29139472 : XLogInsertRecord(XLogRecData *rdata,
749 : XLogRecPtr fpw_lsn,
750 : uint8 flags,
751 : int num_fpi,
752 : bool topxid_included)
753 : {
754 29139472 : XLogCtlInsert *Insert = &XLogCtl->Insert;
755 : pg_crc32c rdata_crc;
756 : bool inserted;
757 29139472 : XLogRecord *rechdr = (XLogRecord *) rdata->data;
758 29139472 : uint8 info = rechdr->xl_info & ~XLR_INFO_MASK;
759 29139472 : WalInsertClass class = WALINSERT_NORMAL;
760 : XLogRecPtr StartPos;
761 : XLogRecPtr EndPos;
762 29139472 : bool prevDoPageWrites = doPageWrites;
763 : TimeLineID insertTLI;
764 :
765 : /* Does this record type require special handling? */
766 29139472 : if (unlikely(rechdr->xl_rmid == RM_XLOG_ID))
767 : {
768 441144 : if (info == XLOG_SWITCH)
769 1550 : class = WALINSERT_SPECIAL_SWITCH;
770 439594 : else if (info == XLOG_CHECKPOINT_REDO)
771 1806 : class = WALINSERT_SPECIAL_CHECKPOINT;
772 : }
773 :
774 : /* we assume that all of the record header is in the first chunk */
775 : Assert(rdata->len >= SizeOfXLogRecord);
776 :
777 : /* cross-check on whether we should be here or not */
778 29139472 : if (!XLogInsertAllowed())
779 0 : elog(ERROR, "cannot make new WAL entries during recovery");
780 :
781 : /*
782 : * Given that we're not in recovery, InsertTimeLineID is set and can't
783 : * change, so we can read it without a lock.
784 : */
785 29139472 : insertTLI = XLogCtl->InsertTimeLineID;
786 :
787 : /*----------
788 : *
789 : * We have now done all the preparatory work we can without holding a
790 : * lock or modifying shared state. From here on, inserting the new WAL
791 : * record to the shared WAL buffer cache is a two-step process:
792 : *
793 : * 1. Reserve the right amount of space from the WAL. The current head of
794 : * reserved space is kept in Insert->CurrBytePos, and is protected by
795 : * insertpos_lck.
796 : *
797 : * 2. Copy the record to the reserved WAL space. This involves finding the
798 : * correct WAL buffer containing the reserved space, and copying the
799 : * record in place. This can be done concurrently in multiple processes.
800 : *
801 : * To keep track of which insertions are still in-progress, each concurrent
802 : * inserter acquires an insertion lock. In addition to just indicating that
803 : * an insertion is in progress, the lock tells others how far the inserter
804 : * has progressed. There is a small fixed number of insertion locks,
805 : * determined by NUM_XLOGINSERT_LOCKS. When an inserter crosses a page
806 : * boundary, it updates the value stored in the lock to the how far it has
807 : * inserted, to allow the previous buffer to be flushed.
808 : *
809 : * Holding onto an insertion lock also protects RedoRecPtr and
810 : * fullPageWrites from changing until the insertion is finished.
811 : *
812 : * Step 2 can usually be done completely in parallel. If the required WAL
813 : * page is not initialized yet, you have to grab WALBufMappingLock to
814 : * initialize it, but the WAL writer tries to do that ahead of insertions
815 : * to avoid that from happening in the critical path.
816 : *
817 : *----------
818 : */
819 29139472 : START_CRIT_SECTION();
820 :
821 29139472 : if (likely(class == WALINSERT_NORMAL))
822 : {
823 29136116 : WALInsertLockAcquire();
824 :
825 : /*
826 : * Check to see if my copy of RedoRecPtr is out of date. If so, may
827 : * have to go back and have the caller recompute everything. This can
828 : * only happen just after a checkpoint, so it's better to be slow in
829 : * this case and fast otherwise.
830 : *
831 : * Also check to see if fullPageWrites was just turned on or there's a
832 : * running backup (which forces full-page writes); if we weren't
833 : * already doing full-page writes then go back and recompute.
834 : *
835 : * If we aren't doing full-page writes then RedoRecPtr doesn't
836 : * actually affect the contents of the XLOG record, so we'll update
837 : * our local copy but not force a recomputation. (If doPageWrites was
838 : * just turned off, we could recompute the record without full pages,
839 : * but we choose not to bother.)
840 : */
841 29136116 : if (RedoRecPtr != Insert->RedoRecPtr)
842 : {
843 : Assert(RedoRecPtr < Insert->RedoRecPtr);
844 13784 : RedoRecPtr = Insert->RedoRecPtr;
845 : }
846 29136116 : doPageWrites = (Insert->fullPageWrites || Insert->runningBackups > 0);
847 :
848 29136116 : if (doPageWrites &&
849 28650790 : (!prevDoPageWrites ||
850 26298346 : (fpw_lsn != InvalidXLogRecPtr && fpw_lsn <= RedoRecPtr)))
851 : {
852 : /*
853 : * Oops, some buffer now needs to be backed up that the caller
854 : * didn't back up. Start over.
855 : */
856 15298 : WALInsertLockRelease();
857 15298 : END_CRIT_SECTION();
858 15298 : return InvalidXLogRecPtr;
859 : }
860 :
861 : /*
862 : * Reserve space for the record in the WAL. This also sets the xl_prev
863 : * pointer.
864 : */
865 29120818 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
866 : &rechdr->xl_prev);
867 :
868 : /* Normal records are always inserted. */
869 29120818 : inserted = true;
870 : }
871 3356 : else if (class == WALINSERT_SPECIAL_SWITCH)
872 : {
873 : /*
874 : * In order to insert an XLOG_SWITCH record, we need to hold all of
875 : * the WAL insertion locks, not just one, so that no one else can
876 : * begin inserting a record until we've figured out how much space
877 : * remains in the current WAL segment and claimed all of it.
878 : *
879 : * Nonetheless, this case is simpler than the normal cases handled
880 : * below, which must check for changes in doPageWrites and RedoRecPtr.
881 : * Those checks are only needed for records that can contain buffer
882 : * references, and an XLOG_SWITCH record never does.
883 : */
884 : Assert(fpw_lsn == InvalidXLogRecPtr);
885 1550 : WALInsertLockAcquireExclusive();
886 1550 : inserted = ReserveXLogSwitch(&StartPos, &EndPos, &rechdr->xl_prev);
887 : }
888 : else
889 : {
890 : Assert(class == WALINSERT_SPECIAL_CHECKPOINT);
891 :
892 : /*
893 : * We need to update both the local and shared copies of RedoRecPtr,
894 : * which means that we need to hold all the WAL insertion locks.
895 : * However, there can't be any buffer references, so as above, we need
896 : * not check RedoRecPtr before inserting the record; we just need to
897 : * update it afterwards.
898 : */
899 : Assert(fpw_lsn == InvalidXLogRecPtr);
900 1806 : WALInsertLockAcquireExclusive();
901 1806 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
902 : &rechdr->xl_prev);
903 1806 : RedoRecPtr = Insert->RedoRecPtr = StartPos;
904 1806 : inserted = true;
905 : }
906 :
907 29124174 : if (inserted)
908 : {
909 : /*
910 : * Now that xl_prev has been filled in, calculate CRC of the record
911 : * header.
912 : */
913 29124060 : rdata_crc = rechdr->xl_crc;
914 29124060 : COMP_CRC32C(rdata_crc, rechdr, offsetof(XLogRecord, xl_crc));
915 29124060 : FIN_CRC32C(rdata_crc);
916 29124060 : rechdr->xl_crc = rdata_crc;
917 :
918 : /*
919 : * All the record data, including the header, is now ready to be
920 : * inserted. Copy the record in the space reserved.
921 : */
922 29124060 : CopyXLogRecordToWAL(rechdr->xl_tot_len,
923 : class == WALINSERT_SPECIAL_SWITCH, rdata,
924 : StartPos, EndPos, insertTLI);
925 :
926 : /*
927 : * Unless record is flagged as not important, update LSN of last
928 : * important record in the current slot. When holding all locks, just
929 : * update the first one.
930 : */
931 29124060 : if ((flags & XLOG_MARK_UNIMPORTANT) == 0)
932 : {
933 28943150 : int lockno = holdingAllLocks ? 0 : MyLockNo;
934 :
935 28943150 : WALInsertLocks[lockno].l.lastImportantAt = StartPos;
936 : }
937 : }
938 : else
939 : {
940 : /*
941 : * This was an xlog-switch record, but the current insert location was
942 : * already exactly at the beginning of a segment, so there was no need
943 : * to do anything.
944 : */
945 : }
946 :
947 : /*
948 : * Done! Let others know that we're finished.
949 : */
950 29124174 : WALInsertLockRelease();
951 :
952 29124174 : END_CRIT_SECTION();
953 :
954 29124174 : MarkCurrentTransactionIdLoggedIfAny();
955 :
956 : /*
957 : * Mark top transaction id is logged (if needed) so that we should not try
958 : * to log it again with the next WAL record in the current subtransaction.
959 : */
960 29124174 : if (topxid_included)
961 438 : MarkSubxactTopXidLogged();
962 :
963 : /*
964 : * Update shared LogwrtRqst.Write, if we crossed page boundary.
965 : */
966 29124174 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
967 : {
968 3303590 : SpinLockAcquire(&XLogCtl->info_lck);
969 : /* advance global request to include new block(s) */
970 3303590 : if (XLogCtl->LogwrtRqst.Write < EndPos)
971 3104588 : XLogCtl->LogwrtRqst.Write = EndPos;
972 3303590 : SpinLockRelease(&XLogCtl->info_lck);
973 3303590 : RefreshXLogWriteResult(LogwrtResult);
974 : }
975 :
976 : /*
977 : * If this was an XLOG_SWITCH record, flush the record and the empty
978 : * padding space that fills the rest of the segment, and perform
979 : * end-of-segment actions (eg, notifying archiver).
980 : */
981 29124174 : if (class == WALINSERT_SPECIAL_SWITCH)
982 : {
983 : TRACE_POSTGRESQL_WAL_SWITCH();
984 1550 : XLogFlush(EndPos);
985 :
986 : /*
987 : * Even though we reserved the rest of the segment for us, which is
988 : * reflected in EndPos, we return a pointer to just the end of the
989 : * xlog-switch record.
990 : */
991 1550 : if (inserted)
992 : {
993 1436 : EndPos = StartPos + SizeOfXLogRecord;
994 1436 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
995 : {
996 0 : uint64 offset = XLogSegmentOffset(EndPos, wal_segment_size);
997 :
998 0 : if (offset == EndPos % XLOG_BLCKSZ)
999 0 : EndPos += SizeOfXLogLongPHD;
1000 : else
1001 0 : EndPos += SizeOfXLogShortPHD;
1002 : }
1003 : }
1004 : }
1005 :
1006 : #ifdef WAL_DEBUG
1007 : if (XLOG_DEBUG)
1008 : {
1009 : static XLogReaderState *debug_reader = NULL;
1010 : XLogRecord *record;
1011 : DecodedXLogRecord *decoded;
1012 : StringInfoData buf;
1013 : StringInfoData recordBuf;
1014 : char *errormsg = NULL;
1015 : MemoryContext oldCxt;
1016 :
1017 : oldCxt = MemoryContextSwitchTo(walDebugCxt);
1018 :
1019 : initStringInfo(&buf);
1020 : appendStringInfo(&buf, "INSERT @ %X/%08X: ", LSN_FORMAT_ARGS(EndPos));
1021 :
1022 : /*
1023 : * We have to piece together the WAL record data from the XLogRecData
1024 : * entries, so that we can pass it to the rm_desc function as one
1025 : * contiguous chunk.
1026 : */
1027 : initStringInfo(&recordBuf);
1028 : for (; rdata != NULL; rdata = rdata->next)
1029 : appendBinaryStringInfo(&recordBuf, rdata->data, rdata->len);
1030 :
1031 : /* We also need temporary space to decode the record. */
1032 : record = (XLogRecord *) recordBuf.data;
1033 : decoded = (DecodedXLogRecord *)
1034 : palloc(DecodeXLogRecordRequiredSpace(record->xl_tot_len));
1035 :
1036 : if (!debug_reader)
1037 : debug_reader = XLogReaderAllocate(wal_segment_size, NULL,
1038 : XL_ROUTINE(.page_read = NULL,
1039 : .segment_open = NULL,
1040 : .segment_close = NULL),
1041 : NULL);
1042 : if (!debug_reader)
1043 : {
1044 : appendStringInfoString(&buf, "error decoding record: out of memory while allocating a WAL reading processor");
1045 : }
1046 : else if (!DecodeXLogRecord(debug_reader,
1047 : decoded,
1048 : record,
1049 : EndPos,
1050 : &errormsg))
1051 : {
1052 : appendStringInfo(&buf, "error decoding record: %s",
1053 : errormsg ? errormsg : "no error message");
1054 : }
1055 : else
1056 : {
1057 : appendStringInfoString(&buf, " - ");
1058 :
1059 : debug_reader->record = decoded;
1060 : xlog_outdesc(&buf, debug_reader);
1061 : debug_reader->record = NULL;
1062 : }
1063 : elog(LOG, "%s", buf.data);
1064 :
1065 : pfree(decoded);
1066 : pfree(buf.data);
1067 : pfree(recordBuf.data);
1068 : MemoryContextSwitchTo(oldCxt);
1069 : }
1070 : #endif
1071 :
1072 : /*
1073 : * Update our global variables
1074 : */
1075 29124174 : ProcLastRecPtr = StartPos;
1076 29124174 : XactLastRecEnd = EndPos;
1077 :
1078 : /* Report WAL traffic to the instrumentation. */
1079 29124174 : if (inserted)
1080 : {
1081 29124060 : pgWalUsage.wal_bytes += rechdr->xl_tot_len;
1082 29124060 : pgWalUsage.wal_records++;
1083 29124060 : pgWalUsage.wal_fpi += num_fpi;
1084 :
1085 : /* Required for the flush of pending stats WAL data */
1086 29124060 : pgstat_report_fixed = true;
1087 : }
1088 :
1089 29124174 : return EndPos;
1090 : }
1091 :
1092 : /*
1093 : * Reserves the right amount of space for a record of given size from the WAL.
1094 : * *StartPos is set to the beginning of the reserved section, *EndPos to
1095 : * its end+1. *PrevPtr is set to the beginning of the previous record; it is
1096 : * used to set the xl_prev of this record.
1097 : *
1098 : * This is the performance critical part of XLogInsert that must be serialized
1099 : * across backends. The rest can happen mostly in parallel. Try to keep this
1100 : * section as short as possible, insertpos_lck can be heavily contended on a
1101 : * busy system.
1102 : *
1103 : * NB: The space calculation here must match the code in CopyXLogRecordToWAL,
1104 : * where we actually copy the record to the reserved space.
1105 : *
1106 : * NB: Testing shows that XLogInsertRecord runs faster if this code is inlined;
1107 : * however, because there are two call sites, the compiler is reluctant to
1108 : * inline. We use pg_attribute_always_inline here to try to convince it.
1109 : */
1110 : static pg_attribute_always_inline void
1111 29122624 : ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1112 : XLogRecPtr *PrevPtr)
1113 : {
1114 29122624 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1115 : uint64 startbytepos;
1116 : uint64 endbytepos;
1117 : uint64 prevbytepos;
1118 :
1119 29122624 : size = MAXALIGN(size);
1120 :
1121 : /* All (non xlog-switch) records should contain data. */
1122 : Assert(size > SizeOfXLogRecord);
1123 :
1124 : /*
1125 : * The duration the spinlock needs to be held is minimized by minimizing
1126 : * the calculations that have to be done while holding the lock. The
1127 : * current tip of reserved WAL is kept in CurrBytePos, as a byte position
1128 : * that only counts "usable" bytes in WAL, that is, it excludes all WAL
1129 : * page headers. The mapping between "usable" byte positions and physical
1130 : * positions (XLogRecPtrs) can be done outside the locked region, and
1131 : * because the usable byte position doesn't include any headers, reserving
1132 : * X bytes from WAL is almost as simple as "CurrBytePos += X".
1133 : */
1134 29122624 : SpinLockAcquire(&Insert->insertpos_lck);
1135 :
1136 29122624 : startbytepos = Insert->CurrBytePos;
1137 29122624 : endbytepos = startbytepos + size;
1138 29122624 : prevbytepos = Insert->PrevBytePos;
1139 29122624 : Insert->CurrBytePos = endbytepos;
1140 29122624 : Insert->PrevBytePos = startbytepos;
1141 :
1142 29122624 : SpinLockRelease(&Insert->insertpos_lck);
1143 :
1144 29122624 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1145 29122624 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1146 29122624 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1147 :
1148 : /*
1149 : * Check that the conversions between "usable byte positions" and
1150 : * XLogRecPtrs work consistently in both directions.
1151 : */
1152 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1153 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1154 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1155 29122624 : }
1156 :
1157 : /*
1158 : * Like ReserveXLogInsertLocation(), but for an xlog-switch record.
1159 : *
1160 : * A log-switch record is handled slightly differently. The rest of the
1161 : * segment will be reserved for this insertion, as indicated by the returned
1162 : * *EndPos value. However, if we are already at the beginning of the current
1163 : * segment, *StartPos and *EndPos are set to the current location without
1164 : * reserving any space, and the function returns false.
1165 : */
1166 : static bool
1167 1550 : ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1168 : {
1169 1550 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1170 : uint64 startbytepos;
1171 : uint64 endbytepos;
1172 : uint64 prevbytepos;
1173 1550 : uint32 size = MAXALIGN(SizeOfXLogRecord);
1174 : XLogRecPtr ptr;
1175 : uint32 segleft;
1176 :
1177 : /*
1178 : * These calculations are a bit heavy-weight to be done while holding a
1179 : * spinlock, but since we're holding all the WAL insertion locks, there
1180 : * are no other inserters competing for it. GetXLogInsertRecPtr() does
1181 : * compete for it, but that's not called very frequently.
1182 : */
1183 1550 : SpinLockAcquire(&Insert->insertpos_lck);
1184 :
1185 1550 : startbytepos = Insert->CurrBytePos;
1186 :
1187 1550 : ptr = XLogBytePosToEndRecPtr(startbytepos);
1188 1550 : if (XLogSegmentOffset(ptr, wal_segment_size) == 0)
1189 : {
1190 114 : SpinLockRelease(&Insert->insertpos_lck);
1191 114 : *EndPos = *StartPos = ptr;
1192 114 : return false;
1193 : }
1194 :
1195 1436 : endbytepos = startbytepos + size;
1196 1436 : prevbytepos = Insert->PrevBytePos;
1197 :
1198 1436 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1199 1436 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1200 :
1201 1436 : segleft = wal_segment_size - XLogSegmentOffset(*EndPos, wal_segment_size);
1202 1436 : if (segleft != wal_segment_size)
1203 : {
1204 : /* consume the rest of the segment */
1205 1436 : *EndPos += segleft;
1206 1436 : endbytepos = XLogRecPtrToBytePos(*EndPos);
1207 : }
1208 1436 : Insert->CurrBytePos = endbytepos;
1209 1436 : Insert->PrevBytePos = startbytepos;
1210 :
1211 1436 : SpinLockRelease(&Insert->insertpos_lck);
1212 :
1213 1436 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1214 :
1215 : Assert(XLogSegmentOffset(*EndPos, wal_segment_size) == 0);
1216 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1217 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1218 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1219 :
1220 1436 : return true;
1221 : }
1222 :
1223 : /*
1224 : * Subroutine of XLogInsertRecord. Copies a WAL record to an already-reserved
1225 : * area in the WAL.
1226 : */
1227 : static void
1228 29124060 : CopyXLogRecordToWAL(int write_len, bool isLogSwitch, XLogRecData *rdata,
1229 : XLogRecPtr StartPos, XLogRecPtr EndPos, TimeLineID tli)
1230 : {
1231 : char *currpos;
1232 : int freespace;
1233 : int written;
1234 : XLogRecPtr CurrPos;
1235 : XLogPageHeader pagehdr;
1236 :
1237 : /*
1238 : * Get a pointer to the right place in the right WAL buffer to start
1239 : * inserting to.
1240 : */
1241 29124060 : CurrPos = StartPos;
1242 29124060 : currpos = GetXLogBuffer(CurrPos, tli);
1243 29124060 : freespace = INSERT_FREESPACE(CurrPos);
1244 :
1245 : /*
1246 : * there should be enough space for at least the first field (xl_tot_len)
1247 : * on this page.
1248 : */
1249 : Assert(freespace >= sizeof(uint32));
1250 :
1251 : /* Copy record data */
1252 29124060 : written = 0;
1253 137545706 : while (rdata != NULL)
1254 : {
1255 108421646 : const char *rdata_data = rdata->data;
1256 108421646 : int rdata_len = rdata->len;
1257 :
1258 111954822 : while (rdata_len > freespace)
1259 : {
1260 : /*
1261 : * Write what fits on this page, and continue on the next page.
1262 : */
1263 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || freespace == 0);
1264 3533176 : memcpy(currpos, rdata_data, freespace);
1265 3533176 : rdata_data += freespace;
1266 3533176 : rdata_len -= freespace;
1267 3533176 : written += freespace;
1268 3533176 : CurrPos += freespace;
1269 :
1270 : /*
1271 : * Get pointer to beginning of next page, and set the xlp_rem_len
1272 : * in the page header. Set XLP_FIRST_IS_CONTRECORD.
1273 : *
1274 : * It's safe to set the contrecord flag and xlp_rem_len without a
1275 : * lock on the page. All the other flags were already set when the
1276 : * page was initialized, in AdvanceXLInsertBuffer, and we're the
1277 : * only backend that needs to set the contrecord flag.
1278 : */
1279 3533176 : currpos = GetXLogBuffer(CurrPos, tli);
1280 3533176 : pagehdr = (XLogPageHeader) currpos;
1281 3533176 : pagehdr->xlp_rem_len = write_len - written;
1282 3533176 : pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1283 :
1284 : /* skip over the page header */
1285 3533176 : if (XLogSegmentOffset(CurrPos, wal_segment_size) == 0)
1286 : {
1287 2366 : CurrPos += SizeOfXLogLongPHD;
1288 2366 : currpos += SizeOfXLogLongPHD;
1289 : }
1290 : else
1291 : {
1292 3530810 : CurrPos += SizeOfXLogShortPHD;
1293 3530810 : currpos += SizeOfXLogShortPHD;
1294 : }
1295 3533176 : freespace = INSERT_FREESPACE(CurrPos);
1296 : }
1297 :
1298 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1299 108421646 : memcpy(currpos, rdata_data, rdata_len);
1300 108421646 : currpos += rdata_len;
1301 108421646 : CurrPos += rdata_len;
1302 108421646 : freespace -= rdata_len;
1303 108421646 : written += rdata_len;
1304 :
1305 108421646 : rdata = rdata->next;
1306 : }
1307 : Assert(written == write_len);
1308 :
1309 : /*
1310 : * If this was an xlog-switch, it's not enough to write the switch record,
1311 : * we also have to consume all the remaining space in the WAL segment. We
1312 : * have already reserved that space, but we need to actually fill it.
1313 : */
1314 29124060 : if (isLogSwitch && XLogSegmentOffset(CurrPos, wal_segment_size) != 0)
1315 : {
1316 : /* An xlog-switch record doesn't contain any data besides the header */
1317 : Assert(write_len == SizeOfXLogRecord);
1318 :
1319 : /* Assert that we did reserve the right amount of space */
1320 : Assert(XLogSegmentOffset(EndPos, wal_segment_size) == 0);
1321 :
1322 : /* Use up all the remaining space on the current page */
1323 1436 : CurrPos += freespace;
1324 :
1325 : /*
1326 : * Cause all remaining pages in the segment to be flushed, leaving the
1327 : * XLog position where it should be, at the start of the next segment.
1328 : * We do this one page at a time, to make sure we don't deadlock
1329 : * against ourselves if wal_buffers < wal_segment_size.
1330 : */
1331 1373460 : while (CurrPos < EndPos)
1332 : {
1333 : /*
1334 : * The minimal action to flush the page would be to call
1335 : * WALInsertLockUpdateInsertingAt(CurrPos) followed by
1336 : * AdvanceXLInsertBuffer(...). The page would be left initialized
1337 : * mostly to zeros, except for the page header (always the short
1338 : * variant, as this is never a segment's first page).
1339 : *
1340 : * The large vistas of zeros are good for compressibility, but the
1341 : * headers interrupting them every XLOG_BLCKSZ (with values that
1342 : * differ from page to page) are not. The effect varies with
1343 : * compression tool, but bzip2 for instance compresses about an
1344 : * order of magnitude worse if those headers are left in place.
1345 : *
1346 : * Rather than complicating AdvanceXLInsertBuffer itself (which is
1347 : * called in heavily-loaded circumstances as well as this lightly-
1348 : * loaded one) with variant behavior, we just use GetXLogBuffer
1349 : * (which itself calls the two methods we need) to get the pointer
1350 : * and zero most of the page. Then we just zero the page header.
1351 : */
1352 1372024 : currpos = GetXLogBuffer(CurrPos, tli);
1353 5488096 : MemSet(currpos, 0, SizeOfXLogShortPHD);
1354 :
1355 1372024 : CurrPos += XLOG_BLCKSZ;
1356 : }
1357 : }
1358 : else
1359 : {
1360 : /* Align the end position, so that the next record starts aligned */
1361 29122624 : CurrPos = MAXALIGN64(CurrPos);
1362 : }
1363 :
1364 29124060 : if (CurrPos != EndPos)
1365 0 : ereport(PANIC,
1366 : errcode(ERRCODE_DATA_CORRUPTED),
1367 : errmsg_internal("space reserved for WAL record does not match what was written"));
1368 29124060 : }
1369 :
1370 : /*
1371 : * Acquire a WAL insertion lock, for inserting to WAL.
1372 : */
1373 : static void
1374 29136138 : WALInsertLockAcquire(void)
1375 : {
1376 : bool immed;
1377 :
1378 : /*
1379 : * It doesn't matter which of the WAL insertion locks we acquire, so try
1380 : * the one we used last time. If the system isn't particularly busy, it's
1381 : * a good bet that it's still available, and it's good to have some
1382 : * affinity to a particular lock so that you don't unnecessarily bounce
1383 : * cache lines between processes when there's no contention.
1384 : *
1385 : * If this is the first time through in this backend, pick a lock
1386 : * (semi-)randomly. This allows the locks to be used evenly if you have a
1387 : * lot of very short connections.
1388 : */
1389 : static int lockToTry = -1;
1390 :
1391 29136138 : if (lockToTry == -1)
1392 16004 : lockToTry = MyProcNumber % NUM_XLOGINSERT_LOCKS;
1393 29136138 : MyLockNo = lockToTry;
1394 :
1395 : /*
1396 : * The insertingAt value is initially set to 0, as we don't know our
1397 : * insert location yet.
1398 : */
1399 29136138 : immed = LWLockAcquire(&WALInsertLocks[MyLockNo].l.lock, LW_EXCLUSIVE);
1400 29136138 : if (!immed)
1401 : {
1402 : /*
1403 : * If we couldn't get the lock immediately, try another lock next
1404 : * time. On a system with more insertion locks than concurrent
1405 : * inserters, this causes all the inserters to eventually migrate to a
1406 : * lock that no-one else is using. On a system with more inserters
1407 : * than locks, it still helps to distribute the inserters evenly
1408 : * across the locks.
1409 : */
1410 70388 : lockToTry = (lockToTry + 1) % NUM_XLOGINSERT_LOCKS;
1411 : }
1412 29136138 : }
1413 :
1414 : /*
1415 : * Acquire all WAL insertion locks, to prevent other backends from inserting
1416 : * to WAL.
1417 : */
1418 : static void
1419 8554 : WALInsertLockAcquireExclusive(void)
1420 : {
1421 : int i;
1422 :
1423 : /*
1424 : * When holding all the locks, all but the last lock's insertingAt
1425 : * indicator is set to 0xFFFFFFFFFFFFFFFF, which is higher than any real
1426 : * XLogRecPtr value, to make sure that no-one blocks waiting on those.
1427 : */
1428 68432 : for (i = 0; i < NUM_XLOGINSERT_LOCKS - 1; i++)
1429 : {
1430 59878 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1431 59878 : LWLockUpdateVar(&WALInsertLocks[i].l.lock,
1432 59878 : &WALInsertLocks[i].l.insertingAt,
1433 : PG_UINT64_MAX);
1434 : }
1435 : /* Variable value reset to 0 at release */
1436 8554 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1437 :
1438 8554 : holdingAllLocks = true;
1439 8554 : }
1440 :
1441 : /*
1442 : * Release our insertion lock (or locks, if we're holding them all).
1443 : *
1444 : * NB: Reset all variables to 0, so they cause LWLockWaitForVar to block the
1445 : * next time the lock is acquired.
1446 : */
1447 : static void
1448 29144692 : WALInsertLockRelease(void)
1449 : {
1450 29144692 : if (holdingAllLocks)
1451 : {
1452 : int i;
1453 :
1454 76986 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1455 68432 : LWLockReleaseClearVar(&WALInsertLocks[i].l.lock,
1456 68432 : &WALInsertLocks[i].l.insertingAt,
1457 : 0);
1458 :
1459 8554 : holdingAllLocks = false;
1460 : }
1461 : else
1462 : {
1463 29136138 : LWLockReleaseClearVar(&WALInsertLocks[MyLockNo].l.lock,
1464 29136138 : &WALInsertLocks[MyLockNo].l.insertingAt,
1465 : 0);
1466 : }
1467 29144692 : }
1468 :
1469 : /*
1470 : * Update our insertingAt value, to let others know that we've finished
1471 : * inserting up to that point.
1472 : */
1473 : static void
1474 5216884 : WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt)
1475 : {
1476 5216884 : if (holdingAllLocks)
1477 : {
1478 : /*
1479 : * We use the last lock to mark our actual position, see comments in
1480 : * WALInsertLockAcquireExclusive.
1481 : */
1482 1363838 : LWLockUpdateVar(&WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.lock,
1483 1363838 : &WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.insertingAt,
1484 : insertingAt);
1485 : }
1486 : else
1487 3853046 : LWLockUpdateVar(&WALInsertLocks[MyLockNo].l.lock,
1488 3853046 : &WALInsertLocks[MyLockNo].l.insertingAt,
1489 : insertingAt);
1490 5216884 : }
1491 :
1492 : /*
1493 : * Wait for any WAL insertions < upto to finish.
1494 : *
1495 : * Returns the location of the oldest insertion that is still in-progress.
1496 : * Any WAL prior to that point has been fully copied into WAL buffers, and
1497 : * can be flushed out to disk. Because this waits for any insertions older
1498 : * than 'upto' to finish, the return value is always >= 'upto'.
1499 : *
1500 : * Note: When you are about to write out WAL, you must call this function
1501 : * *before* acquiring WALWriteLock, to avoid deadlocks. This function might
1502 : * need to wait for an insertion to finish (or at least advance to next
1503 : * uninitialized page), and the inserter might need to evict an old WAL buffer
1504 : * to make room for a new one, which in turn requires WALWriteLock.
1505 : */
1506 : static XLogRecPtr
1507 4784050 : WaitXLogInsertionsToFinish(XLogRecPtr upto)
1508 : {
1509 : uint64 bytepos;
1510 : XLogRecPtr inserted;
1511 : XLogRecPtr reservedUpto;
1512 : XLogRecPtr finishedUpto;
1513 4784050 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1514 : int i;
1515 :
1516 4784050 : if (MyProc == NULL)
1517 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1518 :
1519 : /*
1520 : * Check if there's any work to do. Use a barrier to ensure we get the
1521 : * freshest value.
1522 : */
1523 4784050 : inserted = pg_atomic_read_membarrier_u64(&XLogCtl->logInsertResult);
1524 4784050 : if (upto <= inserted)
1525 3913960 : return inserted;
1526 :
1527 : /* Read the current insert position */
1528 870090 : SpinLockAcquire(&Insert->insertpos_lck);
1529 870090 : bytepos = Insert->CurrBytePos;
1530 870090 : SpinLockRelease(&Insert->insertpos_lck);
1531 870090 : reservedUpto = XLogBytePosToEndRecPtr(bytepos);
1532 :
1533 : /*
1534 : * No-one should request to flush a piece of WAL that hasn't even been
1535 : * reserved yet. However, it can happen if there is a block with a bogus
1536 : * LSN on disk, for example. XLogFlush checks for that situation and
1537 : * complains, but only after the flush. Here we just assume that to mean
1538 : * that all WAL that has been reserved needs to be finished. In this
1539 : * corner-case, the return value can be smaller than 'upto' argument.
1540 : */
1541 870090 : if (upto > reservedUpto)
1542 : {
1543 0 : ereport(LOG,
1544 : errmsg("request to flush past end of generated WAL; request %X/%08X, current position %X/%08X",
1545 : LSN_FORMAT_ARGS(upto), LSN_FORMAT_ARGS(reservedUpto)));
1546 0 : upto = reservedUpto;
1547 : }
1548 :
1549 : /*
1550 : * Loop through all the locks, sleeping on any in-progress insert older
1551 : * than 'upto'.
1552 : *
1553 : * finishedUpto is our return value, indicating the point upto which all
1554 : * the WAL insertions have been finished. Initialize it to the head of
1555 : * reserved WAL, and as we iterate through the insertion locks, back it
1556 : * out for any insertion that's still in progress.
1557 : */
1558 870090 : finishedUpto = reservedUpto;
1559 7830810 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1560 : {
1561 6960720 : XLogRecPtr insertingat = InvalidXLogRecPtr;
1562 :
1563 : do
1564 : {
1565 : /*
1566 : * See if this insertion is in progress. LWLockWaitForVar will
1567 : * wait for the lock to be released, or for the 'value' to be set
1568 : * by a LWLockUpdateVar call. When a lock is initially acquired,
1569 : * its value is 0 (InvalidXLogRecPtr), which means that we don't
1570 : * know where it's inserting yet. We will have to wait for it. If
1571 : * it's a small insertion, the record will most likely fit on the
1572 : * same page and the inserter will release the lock without ever
1573 : * calling LWLockUpdateVar. But if it has to sleep, it will
1574 : * advertise the insertion point with LWLockUpdateVar before
1575 : * sleeping.
1576 : *
1577 : * In this loop we are only waiting for insertions that started
1578 : * before WaitXLogInsertionsToFinish was called. The lack of
1579 : * memory barriers in the loop means that we might see locks as
1580 : * "unused" that have since become used. This is fine because
1581 : * they only can be used for later insertions that we would not
1582 : * want to wait on anyway. Not taking a lock to acquire the
1583 : * current insertingAt value means that we might see older
1584 : * insertingAt values. This is also fine, because if we read a
1585 : * value too old, we will add ourselves to the wait queue, which
1586 : * contains atomic operations.
1587 : */
1588 7117704 : if (LWLockWaitForVar(&WALInsertLocks[i].l.lock,
1589 7117704 : &WALInsertLocks[i].l.insertingAt,
1590 : insertingat, &insertingat))
1591 : {
1592 : /* the lock was free, so no insertion in progress */
1593 5030012 : insertingat = InvalidXLogRecPtr;
1594 5030012 : break;
1595 : }
1596 :
1597 : /*
1598 : * This insertion is still in progress. Have to wait, unless the
1599 : * inserter has proceeded past 'upto'.
1600 : */
1601 2087692 : } while (insertingat < upto);
1602 :
1603 6960720 : if (insertingat != InvalidXLogRecPtr && insertingat < finishedUpto)
1604 767308 : finishedUpto = insertingat;
1605 : }
1606 :
1607 : /*
1608 : * Advance the limit we know to have been inserted and return the freshest
1609 : * value we know of, which might be beyond what we requested if somebody
1610 : * is concurrently doing this with an 'upto' pointer ahead of us.
1611 : */
1612 870090 : finishedUpto = pg_atomic_monotonic_advance_u64(&XLogCtl->logInsertResult,
1613 : finishedUpto);
1614 :
1615 870090 : return finishedUpto;
1616 : }
1617 :
1618 : /*
1619 : * Get a pointer to the right location in the WAL buffer containing the
1620 : * given XLogRecPtr.
1621 : *
1622 : * If the page is not initialized yet, it is initialized. That might require
1623 : * evicting an old dirty buffer from the buffer cache, which means I/O.
1624 : *
1625 : * The caller must ensure that the page containing the requested location
1626 : * isn't evicted yet, and won't be evicted. The way to ensure that is to
1627 : * hold onto a WAL insertion lock with the insertingAt position set to
1628 : * something <= ptr. GetXLogBuffer() will update insertingAt if it needs
1629 : * to evict an old page from the buffer. (This means that once you call
1630 : * GetXLogBuffer() with a given 'ptr', you must not access anything before
1631 : * that point anymore, and must not call GetXLogBuffer() with an older 'ptr'
1632 : * later, because older buffers might be recycled already)
1633 : */
1634 : static char *
1635 34029282 : GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli)
1636 : {
1637 : int idx;
1638 : XLogRecPtr endptr;
1639 : static uint64 cachedPage = 0;
1640 : static char *cachedPos = NULL;
1641 : XLogRecPtr expectedEndPtr;
1642 :
1643 : /*
1644 : * Fast path for the common case that we need to access again the same
1645 : * page as last time.
1646 : */
1647 34029282 : if (ptr / XLOG_BLCKSZ == cachedPage)
1648 : {
1649 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1650 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1651 27783130 : return cachedPos + ptr % XLOG_BLCKSZ;
1652 : }
1653 :
1654 : /*
1655 : * The XLog buffer cache is organized so that a page is always loaded to a
1656 : * particular buffer. That way we can easily calculate the buffer a given
1657 : * page must be loaded into, from the XLogRecPtr alone.
1658 : */
1659 6246152 : idx = XLogRecPtrToBufIdx(ptr);
1660 :
1661 : /*
1662 : * See what page is loaded in the buffer at the moment. It could be the
1663 : * page we're looking for, or something older. It can't be anything newer
1664 : * - that would imply the page we're looking for has already been written
1665 : * out to disk and evicted, and the caller is responsible for making sure
1666 : * that doesn't happen.
1667 : *
1668 : * We don't hold a lock while we read the value. If someone is just about
1669 : * to initialize or has just initialized the page, it's possible that we
1670 : * get InvalidXLogRecPtr. That's ok, we'll grab the mapping lock (in
1671 : * AdvanceXLInsertBuffer) and retry if we see anything other than the page
1672 : * we're looking for.
1673 : */
1674 6246152 : expectedEndPtr = ptr;
1675 6246152 : expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
1676 :
1677 6246152 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1678 6246152 : if (expectedEndPtr != endptr)
1679 : {
1680 : XLogRecPtr initializedUpto;
1681 :
1682 : /*
1683 : * Before calling AdvanceXLInsertBuffer(), which can block, let others
1684 : * know how far we're finished with inserting the record.
1685 : *
1686 : * NB: If 'ptr' points to just after the page header, advertise a
1687 : * position at the beginning of the page rather than 'ptr' itself. If
1688 : * there are no other insertions running, someone might try to flush
1689 : * up to our advertised location. If we advertised a position after
1690 : * the page header, someone might try to flush the page header, even
1691 : * though page might actually not be initialized yet. As the first
1692 : * inserter on the page, we are effectively responsible for making
1693 : * sure that it's initialized, before we let insertingAt to move past
1694 : * the page header.
1695 : */
1696 5216884 : if (ptr % XLOG_BLCKSZ == SizeOfXLogShortPHD &&
1697 12242 : XLogSegmentOffset(ptr, wal_segment_size) > XLOG_BLCKSZ)
1698 12242 : initializedUpto = ptr - SizeOfXLogShortPHD;
1699 5204642 : else if (ptr % XLOG_BLCKSZ == SizeOfXLogLongPHD &&
1700 2470 : XLogSegmentOffset(ptr, wal_segment_size) < XLOG_BLCKSZ)
1701 1196 : initializedUpto = ptr - SizeOfXLogLongPHD;
1702 : else
1703 5203446 : initializedUpto = ptr;
1704 :
1705 5216884 : WALInsertLockUpdateInsertingAt(initializedUpto);
1706 :
1707 5216884 : AdvanceXLInsertBuffer(ptr, tli, false);
1708 5216884 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1709 :
1710 5216884 : if (expectedEndPtr != endptr)
1711 0 : elog(PANIC, "could not find WAL buffer for %X/%08X",
1712 : LSN_FORMAT_ARGS(ptr));
1713 : }
1714 : else
1715 : {
1716 : /*
1717 : * Make sure the initialization of the page is visible to us, and
1718 : * won't arrive later to overwrite the WAL data we write on the page.
1719 : */
1720 1029268 : pg_memory_barrier();
1721 : }
1722 :
1723 : /*
1724 : * Found the buffer holding this page. Return a pointer to the right
1725 : * offset within the page.
1726 : */
1727 6246152 : cachedPage = ptr / XLOG_BLCKSZ;
1728 6246152 : cachedPos = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1729 :
1730 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1731 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1732 :
1733 6246152 : return cachedPos + ptr % XLOG_BLCKSZ;
1734 : }
1735 :
1736 : /*
1737 : * Read WAL data directly from WAL buffers, if available. Returns the number
1738 : * of bytes read successfully.
1739 : *
1740 : * Fewer than 'count' bytes may be read if some of the requested WAL data has
1741 : * already been evicted.
1742 : *
1743 : * No locks are taken.
1744 : *
1745 : * Caller should ensure that it reads no further than LogwrtResult.Write
1746 : * (which should have been updated by the caller when determining how far to
1747 : * read). The 'tli' argument is only used as a convenient safety check so that
1748 : * callers do not read from WAL buffers on a historical timeline.
1749 : */
1750 : Size
1751 196900 : WALReadFromBuffers(char *dstbuf, XLogRecPtr startptr, Size count,
1752 : TimeLineID tli)
1753 : {
1754 196900 : char *pdst = dstbuf;
1755 196900 : XLogRecPtr recptr = startptr;
1756 : XLogRecPtr inserted;
1757 196900 : Size nbytes = count;
1758 :
1759 196900 : if (RecoveryInProgress() || tli != GetWALInsertionTimeLine())
1760 1828 : return 0;
1761 :
1762 : Assert(!XLogRecPtrIsInvalid(startptr));
1763 :
1764 : /*
1765 : * Caller should ensure that the requested data has been inserted into WAL
1766 : * buffers before we try to read it.
1767 : */
1768 195072 : inserted = pg_atomic_read_u64(&XLogCtl->logInsertResult);
1769 195072 : if (startptr + count > inserted)
1770 0 : ereport(ERROR,
1771 : errmsg("cannot read past end of generated WAL: requested %X/%08X, current position %X/%08X",
1772 : LSN_FORMAT_ARGS(startptr + count),
1773 : LSN_FORMAT_ARGS(inserted)));
1774 :
1775 : /*
1776 : * Loop through the buffers without a lock. For each buffer, atomically
1777 : * read and verify the end pointer, then copy the data out, and finally
1778 : * re-read and re-verify the end pointer.
1779 : *
1780 : * Once a page is evicted, it never returns to the WAL buffers, so if the
1781 : * end pointer matches the expected end pointer before and after we copy
1782 : * the data, then the right page must have been present during the data
1783 : * copy. Read barriers are necessary to ensure that the data copy actually
1784 : * happens between the two verification steps.
1785 : *
1786 : * If either verification fails, we simply terminate the loop and return
1787 : * with the data that had been already copied out successfully.
1788 : */
1789 208402 : while (nbytes > 0)
1790 : {
1791 204372 : uint32 offset = recptr % XLOG_BLCKSZ;
1792 204372 : int idx = XLogRecPtrToBufIdx(recptr);
1793 : XLogRecPtr expectedEndPtr;
1794 : XLogRecPtr endptr;
1795 : const char *page;
1796 : const char *psrc;
1797 : Size npagebytes;
1798 :
1799 : /*
1800 : * Calculate the end pointer we expect in the xlblocks array if the
1801 : * correct page is present.
1802 : */
1803 204372 : expectedEndPtr = recptr + (XLOG_BLCKSZ - offset);
1804 :
1805 : /*
1806 : * First verification step: check that the correct page is present in
1807 : * the WAL buffers.
1808 : */
1809 204372 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1810 204372 : if (expectedEndPtr != endptr)
1811 191042 : break;
1812 :
1813 : /*
1814 : * The correct page is present (or was at the time the endptr was
1815 : * read; must re-verify later). Calculate pointer to source data and
1816 : * determine how much data to read from this page.
1817 : */
1818 13330 : page = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1819 13330 : psrc = page + offset;
1820 13330 : npagebytes = Min(nbytes, XLOG_BLCKSZ - offset);
1821 :
1822 : /*
1823 : * Ensure that the data copy and the first verification step are not
1824 : * reordered.
1825 : */
1826 13330 : pg_read_barrier();
1827 :
1828 : /* data copy */
1829 13330 : memcpy(pdst, psrc, npagebytes);
1830 :
1831 : /*
1832 : * Ensure that the data copy and the second verification step are not
1833 : * reordered.
1834 : */
1835 13330 : pg_read_barrier();
1836 :
1837 : /*
1838 : * Second verification step: check that the page we read from wasn't
1839 : * evicted while we were copying the data.
1840 : */
1841 13330 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1842 13330 : if (expectedEndPtr != endptr)
1843 0 : break;
1844 :
1845 13330 : pdst += npagebytes;
1846 13330 : recptr += npagebytes;
1847 13330 : nbytes -= npagebytes;
1848 : }
1849 :
1850 : Assert(pdst - dstbuf <= count);
1851 :
1852 195072 : return pdst - dstbuf;
1853 : }
1854 :
1855 : /*
1856 : * Converts a "usable byte position" to XLogRecPtr. A usable byte position
1857 : * is the position starting from the beginning of WAL, excluding all WAL
1858 : * page headers.
1859 : */
1860 : static XLogRecPtr
1861 58253314 : XLogBytePosToRecPtr(uint64 bytepos)
1862 : {
1863 : uint64 fullsegs;
1864 : uint64 fullpages;
1865 : uint64 bytesleft;
1866 : uint32 seg_offset;
1867 : XLogRecPtr result;
1868 :
1869 58253314 : fullsegs = bytepos / UsableBytesInSegment;
1870 58253314 : bytesleft = bytepos % UsableBytesInSegment;
1871 :
1872 58253314 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1873 : {
1874 : /* fits on first page of segment */
1875 100238 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1876 : }
1877 : else
1878 : {
1879 : /* account for the first page on segment with long header */
1880 58153076 : seg_offset = XLOG_BLCKSZ;
1881 58153076 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1882 :
1883 58153076 : fullpages = bytesleft / UsableBytesInPage;
1884 58153076 : bytesleft = bytesleft % UsableBytesInPage;
1885 :
1886 58153076 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1887 : }
1888 :
1889 58253314 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1890 :
1891 58253314 : return result;
1892 : }
1893 :
1894 : /*
1895 : * Like XLogBytePosToRecPtr, but if the position is at a page boundary,
1896 : * returns a pointer to the beginning of the page (ie. before page header),
1897 : * not to where the first xlog record on that page would go to. This is used
1898 : * when converting a pointer to the end of a record.
1899 : */
1900 : static XLogRecPtr
1901 29995700 : XLogBytePosToEndRecPtr(uint64 bytepos)
1902 : {
1903 : uint64 fullsegs;
1904 : uint64 fullpages;
1905 : uint64 bytesleft;
1906 : uint32 seg_offset;
1907 : XLogRecPtr result;
1908 :
1909 29995700 : fullsegs = bytepos / UsableBytesInSegment;
1910 29995700 : bytesleft = bytepos % UsableBytesInSegment;
1911 :
1912 29995700 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1913 : {
1914 : /* fits on first page of segment */
1915 160708 : if (bytesleft == 0)
1916 107986 : seg_offset = 0;
1917 : else
1918 52722 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1919 : }
1920 : else
1921 : {
1922 : /* account for the first page on segment with long header */
1923 29834992 : seg_offset = XLOG_BLCKSZ;
1924 29834992 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1925 :
1926 29834992 : fullpages = bytesleft / UsableBytesInPage;
1927 29834992 : bytesleft = bytesleft % UsableBytesInPage;
1928 :
1929 29834992 : if (bytesleft == 0)
1930 28784 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
1931 : else
1932 29806208 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1933 : }
1934 :
1935 29995700 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1936 :
1937 29995700 : return result;
1938 : }
1939 :
1940 : /*
1941 : * Convert an XLogRecPtr to a "usable byte position".
1942 : */
1943 : static uint64
1944 4980 : XLogRecPtrToBytePos(XLogRecPtr ptr)
1945 : {
1946 : uint64 fullsegs;
1947 : uint32 fullpages;
1948 : uint32 offset;
1949 : uint64 result;
1950 :
1951 4980 : XLByteToSeg(ptr, fullsegs, wal_segment_size);
1952 :
1953 4980 : fullpages = (XLogSegmentOffset(ptr, wal_segment_size)) / XLOG_BLCKSZ;
1954 4980 : offset = ptr % XLOG_BLCKSZ;
1955 :
1956 4980 : if (fullpages == 0)
1957 : {
1958 1950 : result = fullsegs * UsableBytesInSegment;
1959 1950 : if (offset > 0)
1960 : {
1961 : Assert(offset >= SizeOfXLogLongPHD);
1962 478 : result += offset - SizeOfXLogLongPHD;
1963 : }
1964 : }
1965 : else
1966 : {
1967 3030 : result = fullsegs * UsableBytesInSegment +
1968 3030 : (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
1969 3030 : (fullpages - 1) * UsableBytesInPage; /* full pages */
1970 3030 : if (offset > 0)
1971 : {
1972 : Assert(offset >= SizeOfXLogShortPHD);
1973 3010 : result += offset - SizeOfXLogShortPHD;
1974 : }
1975 : }
1976 :
1977 4980 : return result;
1978 : }
1979 :
1980 : /*
1981 : * Initialize XLOG buffers, writing out old buffers if they still contain
1982 : * unwritten data, upto the page containing 'upto'. Or if 'opportunistic' is
1983 : * true, initialize as many pages as we can without having to write out
1984 : * unwritten data. Any new pages are initialized to zeros, with pages headers
1985 : * initialized properly.
1986 : */
1987 : static void
1988 5226912 : AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
1989 : {
1990 5226912 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1991 : int nextidx;
1992 : XLogRecPtr OldPageRqstPtr;
1993 : XLogwrtRqst WriteRqst;
1994 5226912 : XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
1995 : XLogRecPtr NewPageBeginPtr;
1996 : XLogPageHeader NewPage;
1997 5226912 : int npages pg_attribute_unused() = 0;
1998 :
1999 5226912 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2000 :
2001 : /*
2002 : * Now that we have the lock, check if someone initialized the page
2003 : * already.
2004 : */
2005 14747672 : while (upto >= XLogCtl->InitializedUpTo || opportunistic)
2006 : {
2007 9530788 : nextidx = XLogRecPtrToBufIdx(XLogCtl->InitializedUpTo);
2008 :
2009 : /*
2010 : * Get ending-offset of the buffer page we need to replace (this may
2011 : * be zero if the buffer hasn't been used yet). Fall through if it's
2012 : * already written out.
2013 : */
2014 9530788 : OldPageRqstPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]);
2015 9530788 : if (LogwrtResult.Write < OldPageRqstPtr)
2016 : {
2017 : /*
2018 : * Nope, got work to do. If we just want to pre-initialize as much
2019 : * as we can without flushing, give up now.
2020 : */
2021 4659492 : if (opportunistic)
2022 10028 : break;
2023 :
2024 : /* Advance shared memory write request position */
2025 4649464 : SpinLockAcquire(&XLogCtl->info_lck);
2026 4649464 : if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
2027 1234896 : XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
2028 4649464 : SpinLockRelease(&XLogCtl->info_lck);
2029 :
2030 : /*
2031 : * Acquire an up-to-date LogwrtResult value and see if we still
2032 : * need to write it or if someone else already did.
2033 : */
2034 4649464 : RefreshXLogWriteResult(LogwrtResult);
2035 4649464 : if (LogwrtResult.Write < OldPageRqstPtr)
2036 : {
2037 : /*
2038 : * Must acquire write lock. Release WALBufMappingLock first,
2039 : * to make sure that all insertions that we need to wait for
2040 : * can finish (up to this same position). Otherwise we risk
2041 : * deadlock.
2042 : */
2043 4508382 : LWLockRelease(WALBufMappingLock);
2044 :
2045 4508382 : WaitXLogInsertionsToFinish(OldPageRqstPtr);
2046 :
2047 4508382 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2048 :
2049 4508382 : RefreshXLogWriteResult(LogwrtResult);
2050 4508382 : if (LogwrtResult.Write >= OldPageRqstPtr)
2051 : {
2052 : /* OK, someone wrote it already */
2053 579322 : LWLockRelease(WALWriteLock);
2054 : }
2055 : else
2056 : {
2057 : /* Have to write it ourselves */
2058 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2059 3929060 : WriteRqst.Write = OldPageRqstPtr;
2060 3929060 : WriteRqst.Flush = 0;
2061 3929060 : XLogWrite(WriteRqst, tli, false);
2062 3929060 : LWLockRelease(WALWriteLock);
2063 3929060 : pgWalUsage.wal_buffers_full++;
2064 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_DONE();
2065 :
2066 : /*
2067 : * Required for the flush of pending stats WAL data, per
2068 : * update of pgWalUsage.
2069 : */
2070 3929060 : pgstat_report_fixed = true;
2071 : }
2072 : /* Re-acquire WALBufMappingLock and retry */
2073 4508382 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2074 4508382 : continue;
2075 : }
2076 : }
2077 :
2078 : /*
2079 : * Now the next buffer slot is free and we can set it up to be the
2080 : * next output page.
2081 : */
2082 5012378 : NewPageBeginPtr = XLogCtl->InitializedUpTo;
2083 5012378 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2084 :
2085 : Assert(XLogRecPtrToBufIdx(NewPageBeginPtr) == nextidx);
2086 :
2087 5012378 : NewPage = (XLogPageHeader) (XLogCtl->pages + nextidx * (Size) XLOG_BLCKSZ);
2088 :
2089 : /*
2090 : * Mark the xlblock with InvalidXLogRecPtr and issue a write barrier
2091 : * before initializing. Otherwise, the old page may be partially
2092 : * zeroed but look valid.
2093 : */
2094 5012378 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], InvalidXLogRecPtr);
2095 5012378 : pg_write_barrier();
2096 :
2097 : /*
2098 : * Be sure to re-zero the buffer so that bytes beyond what we've
2099 : * written will look like zeroes and not valid XLOG records...
2100 : */
2101 5012378 : MemSet(NewPage, 0, XLOG_BLCKSZ);
2102 :
2103 : /*
2104 : * Fill the new page's header
2105 : */
2106 5012378 : NewPage->xlp_magic = XLOG_PAGE_MAGIC;
2107 :
2108 : /* NewPage->xlp_info = 0; */ /* done by memset */
2109 5012378 : NewPage->xlp_tli = tli;
2110 5012378 : NewPage->xlp_pageaddr = NewPageBeginPtr;
2111 :
2112 : /* NewPage->xlp_rem_len = 0; */ /* done by memset */
2113 :
2114 : /*
2115 : * If online backup is not in progress, mark the header to indicate
2116 : * that WAL records beginning in this page have removable backup
2117 : * blocks. This allows the WAL archiver to know whether it is safe to
2118 : * compress archived WAL data by transforming full-block records into
2119 : * the non-full-block format. It is sufficient to record this at the
2120 : * page level because we force a page switch (in fact a segment
2121 : * switch) when starting a backup, so the flag will be off before any
2122 : * records can be written during the backup. At the end of a backup,
2123 : * the last page will be marked as all unsafe when perhaps only part
2124 : * is unsafe, but at worst the archiver would miss the opportunity to
2125 : * compress a few records.
2126 : */
2127 5012378 : if (Insert->runningBackups == 0)
2128 4756020 : NewPage->xlp_info |= XLP_BKP_REMOVABLE;
2129 :
2130 : /*
2131 : * If first page of an XLOG segment file, make it a long header.
2132 : */
2133 5012378 : if ((XLogSegmentOffset(NewPage->xlp_pageaddr, wal_segment_size)) == 0)
2134 : {
2135 3594 : XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2136 :
2137 3594 : NewLongPage->xlp_sysid = ControlFile->system_identifier;
2138 3594 : NewLongPage->xlp_seg_size = wal_segment_size;
2139 3594 : NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2140 3594 : NewPage->xlp_info |= XLP_LONG_HEADER;
2141 : }
2142 :
2143 : /*
2144 : * Make sure the initialization of the page becomes visible to others
2145 : * before the xlblocks update. GetXLogBuffer() reads xlblocks without
2146 : * holding a lock.
2147 : */
2148 5012378 : pg_write_barrier();
2149 :
2150 5012378 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], NewPageEndPtr);
2151 5012378 : XLogCtl->InitializedUpTo = NewPageEndPtr;
2152 :
2153 5012378 : npages++;
2154 : }
2155 5226912 : LWLockRelease(WALBufMappingLock);
2156 :
2157 : #ifdef WAL_DEBUG
2158 : if (XLOG_DEBUG && npages > 0)
2159 : {
2160 : elog(DEBUG1, "initialized %d pages, up to %X/%08X",
2161 : npages, LSN_FORMAT_ARGS(NewPageEndPtr));
2162 : }
2163 : #endif
2164 5226912 : }
2165 :
2166 : /*
2167 : * Calculate CheckPointSegments based on max_wal_size_mb and
2168 : * checkpoint_completion_target.
2169 : */
2170 : static void
2171 15344 : CalculateCheckpointSegments(void)
2172 : {
2173 : double target;
2174 :
2175 : /*-------
2176 : * Calculate the distance at which to trigger a checkpoint, to avoid
2177 : * exceeding max_wal_size_mb. This is based on two assumptions:
2178 : *
2179 : * a) we keep WAL for only one checkpoint cycle (prior to PG11 we kept
2180 : * WAL for two checkpoint cycles to allow us to recover from the
2181 : * secondary checkpoint if the first checkpoint failed, though we
2182 : * only did this on the primary anyway, not on standby. Keeping just
2183 : * one checkpoint simplifies processing and reduces disk space in
2184 : * many smaller databases.)
2185 : * b) during checkpoint, we consume checkpoint_completion_target *
2186 : * number of segments consumed between checkpoints.
2187 : *-------
2188 : */
2189 15344 : target = (double) ConvertToXSegs(max_wal_size_mb, wal_segment_size) /
2190 15344 : (1.0 + CheckPointCompletionTarget);
2191 :
2192 : /* round down */
2193 15344 : CheckPointSegments = (int) target;
2194 :
2195 15344 : if (CheckPointSegments < 1)
2196 20 : CheckPointSegments = 1;
2197 15344 : }
2198 :
2199 : void
2200 11104 : assign_max_wal_size(int newval, void *extra)
2201 : {
2202 11104 : max_wal_size_mb = newval;
2203 11104 : CalculateCheckpointSegments();
2204 11104 : }
2205 :
2206 : void
2207 2250 : assign_checkpoint_completion_target(double newval, void *extra)
2208 : {
2209 2250 : CheckPointCompletionTarget = newval;
2210 2250 : CalculateCheckpointSegments();
2211 2250 : }
2212 :
2213 : bool
2214 4342 : check_wal_segment_size(int *newval, void **extra, GucSource source)
2215 : {
2216 4342 : if (!IsValidWalSegSize(*newval))
2217 : {
2218 0 : GUC_check_errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.");
2219 0 : return false;
2220 : }
2221 :
2222 4342 : return true;
2223 : }
2224 :
2225 : /*
2226 : * At a checkpoint, how many WAL segments to recycle as preallocated future
2227 : * XLOG segments? Returns the highest segment that should be preallocated.
2228 : */
2229 : static XLogSegNo
2230 3428 : XLOGfileslop(XLogRecPtr lastredoptr)
2231 : {
2232 : XLogSegNo minSegNo;
2233 : XLogSegNo maxSegNo;
2234 : double distance;
2235 : XLogSegNo recycleSegNo;
2236 :
2237 : /*
2238 : * Calculate the segment numbers that min_wal_size_mb and max_wal_size_mb
2239 : * correspond to. Always recycle enough segments to meet the minimum, and
2240 : * remove enough segments to stay below the maximum.
2241 : */
2242 3428 : minSegNo = lastredoptr / wal_segment_size +
2243 3428 : ConvertToXSegs(min_wal_size_mb, wal_segment_size) - 1;
2244 3428 : maxSegNo = lastredoptr / wal_segment_size +
2245 3428 : ConvertToXSegs(max_wal_size_mb, wal_segment_size) - 1;
2246 :
2247 : /*
2248 : * Between those limits, recycle enough segments to get us through to the
2249 : * estimated end of next checkpoint.
2250 : *
2251 : * To estimate where the next checkpoint will finish, assume that the
2252 : * system runs steadily consuming CheckPointDistanceEstimate bytes between
2253 : * every checkpoint.
2254 : */
2255 3428 : distance = (1.0 + CheckPointCompletionTarget) * CheckPointDistanceEstimate;
2256 : /* add 10% for good measure. */
2257 3428 : distance *= 1.10;
2258 :
2259 3428 : recycleSegNo = (XLogSegNo) ceil(((double) lastredoptr + distance) /
2260 : wal_segment_size);
2261 :
2262 3428 : if (recycleSegNo < minSegNo)
2263 2434 : recycleSegNo = minSegNo;
2264 3428 : if (recycleSegNo > maxSegNo)
2265 766 : recycleSegNo = maxSegNo;
2266 :
2267 3428 : return recycleSegNo;
2268 : }
2269 :
2270 : /*
2271 : * Check whether we've consumed enough xlog space that a checkpoint is needed.
2272 : *
2273 : * new_segno indicates a log file that has just been filled up (or read
2274 : * during recovery). We measure the distance from RedoRecPtr to new_segno
2275 : * and see if that exceeds CheckPointSegments.
2276 : *
2277 : * Note: it is caller's responsibility that RedoRecPtr is up-to-date.
2278 : */
2279 : bool
2280 10298 : XLogCheckpointNeeded(XLogSegNo new_segno)
2281 : {
2282 : XLogSegNo old_segno;
2283 :
2284 10298 : XLByteToSeg(RedoRecPtr, old_segno, wal_segment_size);
2285 :
2286 10298 : if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2287 6714 : return true;
2288 3584 : return false;
2289 : }
2290 :
2291 : /*
2292 : * Write and/or fsync the log at least as far as WriteRqst indicates.
2293 : *
2294 : * If flexible == true, we don't have to write as far as WriteRqst, but
2295 : * may stop at any convenient boundary (such as a cache or logfile boundary).
2296 : * This option allows us to avoid uselessly issuing multiple writes when a
2297 : * single one would do.
2298 : *
2299 : * Must be called with WALWriteLock held. WaitXLogInsertionsToFinish(WriteRqst)
2300 : * must be called before grabbing the lock, to make sure the data is ready to
2301 : * write.
2302 : */
2303 : static void
2304 4189466 : XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible)
2305 : {
2306 : bool ispartialpage;
2307 : bool last_iteration;
2308 : bool finishing_seg;
2309 : int curridx;
2310 : int npages;
2311 : int startidx;
2312 : uint32 startoffset;
2313 :
2314 : /* We should always be inside a critical section here */
2315 : Assert(CritSectionCount > 0);
2316 :
2317 : /*
2318 : * Update local LogwrtResult (caller probably did this already, but...)
2319 : */
2320 4189466 : RefreshXLogWriteResult(LogwrtResult);
2321 :
2322 : /*
2323 : * Since successive pages in the xlog cache are consecutively allocated,
2324 : * we can usually gather multiple pages together and issue just one
2325 : * write() call. npages is the number of pages we have determined can be
2326 : * written together; startidx is the cache block index of the first one,
2327 : * and startoffset is the file offset at which it should go. The latter
2328 : * two variables are only valid when npages > 0, but we must initialize
2329 : * all of them to keep the compiler quiet.
2330 : */
2331 4189466 : npages = 0;
2332 4189466 : startidx = 0;
2333 4189466 : startoffset = 0;
2334 :
2335 : /*
2336 : * Within the loop, curridx is the cache block index of the page to
2337 : * consider writing. Begin at the buffer containing the next unwritten
2338 : * page, or last partially written page.
2339 : */
2340 4189466 : curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2341 :
2342 9099454 : while (LogwrtResult.Write < WriteRqst.Write)
2343 : {
2344 : /*
2345 : * Make sure we're not ahead of the insert process. This could happen
2346 : * if we're passed a bogus WriteRqst.Write that is past the end of the
2347 : * last page that's been initialized by AdvanceXLInsertBuffer.
2348 : */
2349 5160862 : XLogRecPtr EndPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[curridx]);
2350 :
2351 5160862 : if (LogwrtResult.Write >= EndPtr)
2352 0 : elog(PANIC, "xlog write request %X/%08X is past end of log %X/%08X",
2353 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2354 : LSN_FORMAT_ARGS(EndPtr));
2355 :
2356 : /* Advance LogwrtResult.Write to end of current buffer page */
2357 5160862 : LogwrtResult.Write = EndPtr;
2358 5160862 : ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2359 :
2360 5160862 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2361 : wal_segment_size))
2362 : {
2363 : /*
2364 : * Switch to new logfile segment. We cannot have any pending
2365 : * pages here (since we dump what we have at segment end).
2366 : */
2367 : Assert(npages == 0);
2368 26252 : if (openLogFile >= 0)
2369 11782 : XLogFileClose();
2370 26252 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2371 : wal_segment_size);
2372 26252 : openLogTLI = tli;
2373 :
2374 : /* create/use new log file */
2375 26252 : openLogFile = XLogFileInit(openLogSegNo, tli);
2376 26252 : ReserveExternalFD();
2377 : }
2378 :
2379 : /* Make sure we have the current logfile open */
2380 5160862 : if (openLogFile < 0)
2381 : {
2382 0 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2383 : wal_segment_size);
2384 0 : openLogTLI = tli;
2385 0 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2386 0 : ReserveExternalFD();
2387 : }
2388 :
2389 : /* Add current page to the set of pending pages-to-dump */
2390 5160862 : if (npages == 0)
2391 : {
2392 : /* first of group */
2393 4217904 : startidx = curridx;
2394 4217904 : startoffset = XLogSegmentOffset(LogwrtResult.Write - XLOG_BLCKSZ,
2395 : wal_segment_size);
2396 : }
2397 5160862 : npages++;
2398 :
2399 : /*
2400 : * Dump the set if this will be the last loop iteration, or if we are
2401 : * at the last page of the cache area (since the next page won't be
2402 : * contiguous in memory), or if we are at the end of the logfile
2403 : * segment.
2404 : */
2405 5160862 : last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2406 :
2407 10078700 : finishing_seg = !ispartialpage &&
2408 4917838 : (startoffset + npages * XLOG_BLCKSZ) >= wal_segment_size;
2409 :
2410 5160862 : if (last_iteration ||
2411 974144 : curridx == XLogCtl->XLogCacheBlck ||
2412 : finishing_seg)
2413 : {
2414 : char *from;
2415 : Size nbytes;
2416 : Size nleft;
2417 : ssize_t written;
2418 : instr_time start;
2419 :
2420 : /* OK to write the page(s) */
2421 4217904 : from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2422 4217904 : nbytes = npages * (Size) XLOG_BLCKSZ;
2423 4217904 : nleft = nbytes;
2424 : do
2425 : {
2426 4217904 : errno = 0;
2427 :
2428 : /*
2429 : * Measure I/O timing to write WAL data, for pg_stat_io.
2430 : */
2431 4217904 : start = pgstat_prepare_io_time(track_wal_io_timing);
2432 :
2433 4217904 : pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE);
2434 4217904 : written = pg_pwrite(openLogFile, from, nleft, startoffset);
2435 4217904 : pgstat_report_wait_end();
2436 :
2437 4217904 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL,
2438 : IOOP_WRITE, start, 1, written);
2439 :
2440 4217904 : if (written <= 0)
2441 : {
2442 : char xlogfname[MAXFNAMELEN];
2443 : int save_errno;
2444 :
2445 0 : if (errno == EINTR)
2446 0 : continue;
2447 :
2448 0 : save_errno = errno;
2449 0 : XLogFileName(xlogfname, tli, openLogSegNo,
2450 : wal_segment_size);
2451 0 : errno = save_errno;
2452 0 : ereport(PANIC,
2453 : (errcode_for_file_access(),
2454 : errmsg("could not write to log file \"%s\" at offset %u, length %zu: %m",
2455 : xlogfname, startoffset, nleft)));
2456 : }
2457 4217904 : nleft -= written;
2458 4217904 : from += written;
2459 4217904 : startoffset += written;
2460 4217904 : } while (nleft > 0);
2461 :
2462 4217904 : npages = 0;
2463 :
2464 : /*
2465 : * If we just wrote the whole last page of a logfile segment,
2466 : * fsync the segment immediately. This avoids having to go back
2467 : * and re-open prior segments when an fsync request comes along
2468 : * later. Doing it here ensures that one and only one backend will
2469 : * perform this fsync.
2470 : *
2471 : * This is also the right place to notify the Archiver that the
2472 : * segment is ready to copy to archival storage, and to update the
2473 : * timer for archive_timeout, and to signal for a checkpoint if
2474 : * too many logfile segments have been used since the last
2475 : * checkpoint.
2476 : */
2477 4217904 : if (finishing_seg)
2478 : {
2479 3818 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2480 :
2481 : /* signal that we need to wakeup walsenders later */
2482 3818 : WalSndWakeupRequest();
2483 :
2484 3818 : LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2485 :
2486 3818 : if (XLogArchivingActive())
2487 810 : XLogArchiveNotifySeg(openLogSegNo, tli);
2488 :
2489 3818 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2490 3818 : XLogCtl->lastSegSwitchLSN = LogwrtResult.Flush;
2491 :
2492 : /*
2493 : * Request a checkpoint if we've consumed too much xlog since
2494 : * the last one. For speed, we first check using the local
2495 : * copy of RedoRecPtr, which might be out of date; if it looks
2496 : * like a checkpoint is needed, forcibly update RedoRecPtr and
2497 : * recheck.
2498 : */
2499 3818 : if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2500 : {
2501 476 : (void) GetRedoRecPtr();
2502 476 : if (XLogCheckpointNeeded(openLogSegNo))
2503 390 : RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2504 : }
2505 : }
2506 : }
2507 :
2508 5160862 : if (ispartialpage)
2509 : {
2510 : /* Only asked to write a partial page */
2511 243024 : LogwrtResult.Write = WriteRqst.Write;
2512 243024 : break;
2513 : }
2514 4917838 : curridx = NextBufIdx(curridx);
2515 :
2516 : /* If flexible, break out of loop as soon as we wrote something */
2517 4917838 : if (flexible && npages == 0)
2518 7850 : break;
2519 : }
2520 :
2521 : Assert(npages == 0);
2522 :
2523 : /*
2524 : * If asked to flush, do so
2525 : */
2526 4189466 : if (LogwrtResult.Flush < WriteRqst.Flush &&
2527 258944 : LogwrtResult.Flush < LogwrtResult.Write)
2528 : {
2529 : /*
2530 : * Could get here without iterating above loop, in which case we might
2531 : * have no open file or the wrong one. However, we do not need to
2532 : * fsync more than one file.
2533 : */
2534 258822 : if (wal_sync_method != WAL_SYNC_METHOD_OPEN &&
2535 258822 : wal_sync_method != WAL_SYNC_METHOD_OPEN_DSYNC)
2536 : {
2537 258822 : if (openLogFile >= 0 &&
2538 258784 : !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2539 : wal_segment_size))
2540 286 : XLogFileClose();
2541 258822 : if (openLogFile < 0)
2542 : {
2543 324 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2544 : wal_segment_size);
2545 324 : openLogTLI = tli;
2546 324 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2547 324 : ReserveExternalFD();
2548 : }
2549 :
2550 258822 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2551 : }
2552 :
2553 : /* signal that we need to wakeup walsenders later */
2554 258822 : WalSndWakeupRequest();
2555 :
2556 258822 : LogwrtResult.Flush = LogwrtResult.Write;
2557 : }
2558 :
2559 : /*
2560 : * Update shared-memory status
2561 : *
2562 : * We make sure that the shared 'request' values do not fall behind the
2563 : * 'result' values. This is not absolutely essential, but it saves some
2564 : * code in a couple of places.
2565 : */
2566 4189466 : SpinLockAcquire(&XLogCtl->info_lck);
2567 4189466 : if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
2568 224158 : XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
2569 4189466 : if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
2570 261840 : XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
2571 4189466 : SpinLockRelease(&XLogCtl->info_lck);
2572 :
2573 : /*
2574 : * We write Write first, bar, then Flush. When reading, the opposite must
2575 : * be done (with a matching barrier in between), so that we always see a
2576 : * Flush value that trails behind the Write value seen.
2577 : */
2578 4189466 : pg_atomic_write_u64(&XLogCtl->logWriteResult, LogwrtResult.Write);
2579 4189466 : pg_write_barrier();
2580 4189466 : pg_atomic_write_u64(&XLogCtl->logFlushResult, LogwrtResult.Flush);
2581 :
2582 : #ifdef USE_ASSERT_CHECKING
2583 : {
2584 : XLogRecPtr Flush;
2585 : XLogRecPtr Write;
2586 : XLogRecPtr Insert;
2587 :
2588 : Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult);
2589 : pg_read_barrier();
2590 : Write = pg_atomic_read_u64(&XLogCtl->logWriteResult);
2591 : pg_read_barrier();
2592 : Insert = pg_atomic_read_u64(&XLogCtl->logInsertResult);
2593 :
2594 : /* WAL written to disk is always ahead of WAL flushed */
2595 : Assert(Write >= Flush);
2596 :
2597 : /* WAL inserted to buffers is always ahead of WAL written */
2598 : Assert(Insert >= Write);
2599 : }
2600 : #endif
2601 4189466 : }
2602 :
2603 : /*
2604 : * Record the LSN for an asynchronous transaction commit/abort
2605 : * and nudge the WALWriter if there is work for it to do.
2606 : * (This should not be called for synchronous commits.)
2607 : */
2608 : void
2609 94504 : XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2610 : {
2611 94504 : XLogRecPtr WriteRqstPtr = asyncXactLSN;
2612 : bool sleeping;
2613 94504 : bool wakeup = false;
2614 : XLogRecPtr prevAsyncXactLSN;
2615 :
2616 94504 : SpinLockAcquire(&XLogCtl->info_lck);
2617 94504 : sleeping = XLogCtl->WalWriterSleeping;
2618 94504 : prevAsyncXactLSN = XLogCtl->asyncXactLSN;
2619 94504 : if (XLogCtl->asyncXactLSN < asyncXactLSN)
2620 93564 : XLogCtl->asyncXactLSN = asyncXactLSN;
2621 94504 : SpinLockRelease(&XLogCtl->info_lck);
2622 :
2623 : /*
2624 : * If somebody else already called this function with a more aggressive
2625 : * LSN, they will have done what we needed (and perhaps more).
2626 : */
2627 94504 : if (asyncXactLSN <= prevAsyncXactLSN)
2628 940 : return;
2629 :
2630 : /*
2631 : * If the WALWriter is sleeping, kick it to make it come out of low-power
2632 : * mode, so that this async commit will reach disk within the expected
2633 : * amount of time. Otherwise, determine whether it has enough WAL
2634 : * available to flush, the same way that XLogBackgroundFlush() does.
2635 : */
2636 93564 : if (sleeping)
2637 56 : wakeup = true;
2638 : else
2639 : {
2640 : int flushblocks;
2641 :
2642 93508 : RefreshXLogWriteResult(LogwrtResult);
2643 :
2644 93508 : flushblocks =
2645 93508 : WriteRqstPtr / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
2646 :
2647 93508 : if (WalWriterFlushAfter == 0 || flushblocks >= WalWriterFlushAfter)
2648 7418 : wakeup = true;
2649 : }
2650 :
2651 93564 : if (wakeup)
2652 : {
2653 7474 : volatile PROC_HDR *procglobal = ProcGlobal;
2654 7474 : ProcNumber walwriterProc = procglobal->walwriterProc;
2655 :
2656 7474 : if (walwriterProc != INVALID_PROC_NUMBER)
2657 418 : SetLatch(&GetPGProcByNumber(walwriterProc)->procLatch);
2658 : }
2659 : }
2660 :
2661 : /*
2662 : * Record the LSN up to which we can remove WAL because it's not required by
2663 : * any replication slot.
2664 : */
2665 : void
2666 51882 : XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn)
2667 : {
2668 51882 : SpinLockAcquire(&XLogCtl->info_lck);
2669 51882 : XLogCtl->replicationSlotMinLSN = lsn;
2670 51882 : SpinLockRelease(&XLogCtl->info_lck);
2671 51882 : }
2672 :
2673 :
2674 : /*
2675 : * Return the oldest LSN we must retain to satisfy the needs of some
2676 : * replication slot.
2677 : */
2678 : static XLogRecPtr
2679 4394 : XLogGetReplicationSlotMinimumLSN(void)
2680 : {
2681 : XLogRecPtr retval;
2682 :
2683 4394 : SpinLockAcquire(&XLogCtl->info_lck);
2684 4394 : retval = XLogCtl->replicationSlotMinLSN;
2685 4394 : SpinLockRelease(&XLogCtl->info_lck);
2686 :
2687 4394 : return retval;
2688 : }
2689 :
2690 : /*
2691 : * Advance minRecoveryPoint in control file.
2692 : *
2693 : * If we crash during recovery, we must reach this point again before the
2694 : * database is consistent.
2695 : *
2696 : * If 'force' is true, 'lsn' argument is ignored. Otherwise, minRecoveryPoint
2697 : * is only updated if it's not already greater than or equal to 'lsn'.
2698 : */
2699 : static void
2700 213352 : UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2701 : {
2702 : /* Quick check using our local copy of the variable */
2703 213352 : if (!updateMinRecoveryPoint || (!force && lsn <= LocalMinRecoveryPoint))
2704 199932 : return;
2705 :
2706 : /*
2707 : * An invalid minRecoveryPoint means that we need to recover all the WAL,
2708 : * i.e., we're doing crash recovery. We never modify the control file's
2709 : * value in that case, so we can short-circuit future checks here too. The
2710 : * local values of minRecoveryPoint and minRecoveryPointTLI should not be
2711 : * updated until crash recovery finishes. We only do this for the startup
2712 : * process as it should not update its own reference of minRecoveryPoint
2713 : * until it has finished crash recovery to make sure that all WAL
2714 : * available is replayed in this case. This also saves from extra locks
2715 : * taken on the control file from the startup process.
2716 : */
2717 13420 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint) && InRecovery)
2718 : {
2719 60 : updateMinRecoveryPoint = false;
2720 60 : return;
2721 : }
2722 :
2723 13360 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2724 :
2725 : /* update local copy */
2726 13360 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
2727 13360 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2728 :
2729 13360 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint))
2730 4 : updateMinRecoveryPoint = false;
2731 13356 : else if (force || LocalMinRecoveryPoint < lsn)
2732 : {
2733 : XLogRecPtr newMinRecoveryPoint;
2734 : TimeLineID newMinRecoveryPointTLI;
2735 :
2736 : /*
2737 : * To avoid having to update the control file too often, we update it
2738 : * all the way to the last record being replayed, even though 'lsn'
2739 : * would suffice for correctness. This also allows the 'force' case
2740 : * to not need a valid 'lsn' value.
2741 : *
2742 : * Another important reason for doing it this way is that the passed
2743 : * 'lsn' value could be bogus, i.e., past the end of available WAL, if
2744 : * the caller got it from a corrupted heap page. Accepting such a
2745 : * value as the min recovery point would prevent us from coming up at
2746 : * all. Instead, we just log a warning and continue with recovery.
2747 : * (See also the comments about corrupt LSNs in XLogFlush.)
2748 : */
2749 10672 : newMinRecoveryPoint = GetCurrentReplayRecPtr(&newMinRecoveryPointTLI);
2750 10672 : if (!force && newMinRecoveryPoint < lsn)
2751 0 : elog(WARNING,
2752 : "xlog min recovery request %X/%08X is past current point %X/%08X",
2753 : LSN_FORMAT_ARGS(lsn), LSN_FORMAT_ARGS(newMinRecoveryPoint));
2754 :
2755 : /* update control file */
2756 10672 : if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2757 : {
2758 9968 : ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2759 9968 : ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2760 9968 : UpdateControlFile();
2761 9968 : LocalMinRecoveryPoint = newMinRecoveryPoint;
2762 9968 : LocalMinRecoveryPointTLI = newMinRecoveryPointTLI;
2763 :
2764 9968 : ereport(DEBUG2,
2765 : errmsg_internal("updated min recovery point to %X/%08X on timeline %u",
2766 : LSN_FORMAT_ARGS(newMinRecoveryPoint),
2767 : newMinRecoveryPointTLI));
2768 : }
2769 : }
2770 13360 : LWLockRelease(ControlFileLock);
2771 : }
2772 :
2773 : /*
2774 : * Ensure that all XLOG data through the given position is flushed to disk.
2775 : *
2776 : * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
2777 : * already held, and we try to avoid acquiring it if possible.
2778 : */
2779 : void
2780 1368724 : XLogFlush(XLogRecPtr record)
2781 : {
2782 : XLogRecPtr WriteRqstPtr;
2783 : XLogwrtRqst WriteRqst;
2784 1368724 : TimeLineID insertTLI = XLogCtl->InsertTimeLineID;
2785 :
2786 : /*
2787 : * During REDO, we are reading not writing WAL. Therefore, instead of
2788 : * trying to flush the WAL, we should update minRecoveryPoint instead. We
2789 : * test XLogInsertAllowed(), not InRecovery, because we need checkpointer
2790 : * to act this way too, and because when it tries to write the
2791 : * end-of-recovery checkpoint, it should indeed flush.
2792 : */
2793 1368724 : if (!XLogInsertAllowed())
2794 : {
2795 212460 : UpdateMinRecoveryPoint(record, false);
2796 1093428 : return;
2797 : }
2798 :
2799 : /* Quick exit if already known flushed */
2800 1156264 : if (record <= LogwrtResult.Flush)
2801 880968 : return;
2802 :
2803 : #ifdef WAL_DEBUG
2804 : if (XLOG_DEBUG)
2805 : elog(LOG, "xlog flush request %X/%08X; write %X/%08X; flush %X/%08X",
2806 : LSN_FORMAT_ARGS(record),
2807 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2808 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2809 : #endif
2810 :
2811 275296 : START_CRIT_SECTION();
2812 :
2813 : /*
2814 : * Since fsync is usually a horribly expensive operation, we try to
2815 : * piggyback as much data as we can on each fsync: if we see any more data
2816 : * entered into the xlog buffer, we'll write and fsync that too, so that
2817 : * the final value of LogwrtResult.Flush is as large as possible. This
2818 : * gives us some chance of avoiding another fsync immediately after.
2819 : */
2820 :
2821 : /* initialize to given target; may increase below */
2822 275296 : WriteRqstPtr = record;
2823 :
2824 : /*
2825 : * Now wait until we get the write lock, or someone else does the flush
2826 : * for us.
2827 : */
2828 : for (;;)
2829 11098 : {
2830 : XLogRecPtr insertpos;
2831 :
2832 : /* done already? */
2833 286394 : RefreshXLogWriteResult(LogwrtResult);
2834 286394 : if (record <= LogwrtResult.Flush)
2835 20754 : break;
2836 :
2837 : /*
2838 : * Before actually performing the write, wait for all in-flight
2839 : * insertions to the pages we're about to write to finish.
2840 : */
2841 265640 : SpinLockAcquire(&XLogCtl->info_lck);
2842 265640 : if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
2843 19854 : WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
2844 265640 : SpinLockRelease(&XLogCtl->info_lck);
2845 265640 : insertpos = WaitXLogInsertionsToFinish(WriteRqstPtr);
2846 :
2847 : /*
2848 : * Try to get the write lock. If we can't get it immediately, wait
2849 : * until it's released, and recheck if we still need to do the flush
2850 : * or if the backend that held the lock did it for us already. This
2851 : * helps to maintain a good rate of group committing when the system
2852 : * is bottlenecked by the speed of fsyncing.
2853 : */
2854 265640 : if (!LWLockAcquireOrWait(WALWriteLock, LW_EXCLUSIVE))
2855 : {
2856 : /*
2857 : * The lock is now free, but we didn't acquire it yet. Before we
2858 : * do, loop back to check if someone else flushed the record for
2859 : * us already.
2860 : */
2861 11098 : continue;
2862 : }
2863 :
2864 : /* Got the lock; recheck whether request is satisfied */
2865 254542 : RefreshXLogWriteResult(LogwrtResult);
2866 254542 : if (record <= LogwrtResult.Flush)
2867 : {
2868 3800 : LWLockRelease(WALWriteLock);
2869 3800 : break;
2870 : }
2871 :
2872 : /*
2873 : * Sleep before flush! By adding a delay here, we may give further
2874 : * backends the opportunity to join the backlog of group commit
2875 : * followers; this can significantly improve transaction throughput,
2876 : * at the risk of increasing transaction latency.
2877 : *
2878 : * We do not sleep if enableFsync is not turned on, nor if there are
2879 : * fewer than CommitSiblings other backends with active transactions.
2880 : */
2881 250742 : if (CommitDelay > 0 && enableFsync &&
2882 0 : MinimumActiveBackends(CommitSiblings))
2883 : {
2884 0 : pg_usleep(CommitDelay);
2885 :
2886 : /*
2887 : * Re-check how far we can now flush the WAL. It's generally not
2888 : * safe to call WaitXLogInsertionsToFinish while holding
2889 : * WALWriteLock, because an in-progress insertion might need to
2890 : * also grab WALWriteLock to make progress. But we know that all
2891 : * the insertions up to insertpos have already finished, because
2892 : * that's what the earlier WaitXLogInsertionsToFinish() returned.
2893 : * We're only calling it again to allow insertpos to be moved
2894 : * further forward, not to actually wait for anyone.
2895 : */
2896 0 : insertpos = WaitXLogInsertionsToFinish(insertpos);
2897 : }
2898 :
2899 : /* try to write/flush later additions to XLOG as well */
2900 250742 : WriteRqst.Write = insertpos;
2901 250742 : WriteRqst.Flush = insertpos;
2902 :
2903 250742 : XLogWrite(WriteRqst, insertTLI, false);
2904 :
2905 250742 : LWLockRelease(WALWriteLock);
2906 : /* done */
2907 250742 : break;
2908 : }
2909 :
2910 275296 : END_CRIT_SECTION();
2911 :
2912 : /* wake up walsenders now that we've released heavily contended locks */
2913 275296 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
2914 :
2915 : /*
2916 : * If we still haven't flushed to the request point then we have a
2917 : * problem; most likely, the requested flush point is past end of XLOG.
2918 : * This has been seen to occur when a disk page has a corrupted LSN.
2919 : *
2920 : * Formerly we treated this as a PANIC condition, but that hurts the
2921 : * system's robustness rather than helping it: we do not want to take down
2922 : * the whole system due to corruption on one data page. In particular, if
2923 : * the bad page is encountered again during recovery then we would be
2924 : * unable to restart the database at all! (This scenario actually
2925 : * happened in the field several times with 7.1 releases.) As of 8.4, bad
2926 : * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
2927 : * the only time we can reach here during recovery is while flushing the
2928 : * end-of-recovery checkpoint record, and we don't expect that to have a
2929 : * bad LSN.
2930 : *
2931 : * Note that for calls from xact.c, the ERROR will be promoted to PANIC
2932 : * since xact.c calls this routine inside a critical section. However,
2933 : * calls from bufmgr.c are not within critical sections and so we will not
2934 : * force a restart for a bad LSN on a data page.
2935 : */
2936 275296 : if (LogwrtResult.Flush < record)
2937 0 : elog(ERROR,
2938 : "xlog flush request %X/%08X is not satisfied --- flushed only to %X/%08X",
2939 : LSN_FORMAT_ARGS(record),
2940 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2941 : }
2942 :
2943 : /*
2944 : * Write & flush xlog, but without specifying exactly where to.
2945 : *
2946 : * We normally write only completed blocks; but if there is nothing to do on
2947 : * that basis, we check for unwritten async commits in the current incomplete
2948 : * block, and write through the latest one of those. Thus, if async commits
2949 : * are not being used, we will write complete blocks only.
2950 : *
2951 : * If, based on the above, there's anything to write we do so immediately. But
2952 : * to avoid calling fsync, fdatasync et. al. at a rate that'd impact
2953 : * concurrent IO, we only flush WAL every wal_writer_delay ms, or if there's
2954 : * more than wal_writer_flush_after unflushed blocks.
2955 : *
2956 : * We can guarantee that async commits reach disk after at most three
2957 : * wal_writer_delay cycles. (When flushing complete blocks, we allow XLogWrite
2958 : * to write "flexibly", meaning it can stop at the end of the buffer ring;
2959 : * this makes a difference only with very high load or long wal_writer_delay,
2960 : * but imposes one extra cycle for the worst case for async commits.)
2961 : *
2962 : * This routine is invoked periodically by the background walwriter process.
2963 : *
2964 : * Returns true if there was any work to do, even if we skipped flushing due
2965 : * to wal_writer_delay/wal_writer_flush_after.
2966 : */
2967 : bool
2968 34884 : XLogBackgroundFlush(void)
2969 : {
2970 : XLogwrtRqst WriteRqst;
2971 34884 : bool flexible = true;
2972 : static TimestampTz lastflush;
2973 : TimestampTz now;
2974 : int flushblocks;
2975 : TimeLineID insertTLI;
2976 :
2977 : /* XLOG doesn't need flushing during recovery */
2978 34884 : if (RecoveryInProgress())
2979 1844 : return false;
2980 :
2981 : /*
2982 : * Since we're not in recovery, InsertTimeLineID is set and can't change,
2983 : * so we can read it without a lock.
2984 : */
2985 33040 : insertTLI = XLogCtl->InsertTimeLineID;
2986 :
2987 : /* read updated LogwrtRqst */
2988 33040 : SpinLockAcquire(&XLogCtl->info_lck);
2989 33040 : WriteRqst = XLogCtl->LogwrtRqst;
2990 33040 : SpinLockRelease(&XLogCtl->info_lck);
2991 :
2992 : /* back off to last completed page boundary */
2993 33040 : WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
2994 :
2995 : /* if we have already flushed that far, consider async commit records */
2996 33040 : RefreshXLogWriteResult(LogwrtResult);
2997 33040 : if (WriteRqst.Write <= LogwrtResult.Flush)
2998 : {
2999 24524 : SpinLockAcquire(&XLogCtl->info_lck);
3000 24524 : WriteRqst.Write = XLogCtl->asyncXactLSN;
3001 24524 : SpinLockRelease(&XLogCtl->info_lck);
3002 24524 : flexible = false; /* ensure it all gets written */
3003 : }
3004 :
3005 : /*
3006 : * If already known flushed, we're done. Just need to check if we are
3007 : * holding an open file handle to a logfile that's no longer in use,
3008 : * preventing the file from being deleted.
3009 : */
3010 33040 : if (WriteRqst.Write <= LogwrtResult.Flush)
3011 : {
3012 23012 : if (openLogFile >= 0)
3013 : {
3014 14584 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
3015 : wal_segment_size))
3016 : {
3017 290 : XLogFileClose();
3018 : }
3019 : }
3020 23012 : return false;
3021 : }
3022 :
3023 : /*
3024 : * Determine how far to flush WAL, based on the wal_writer_delay and
3025 : * wal_writer_flush_after GUCs.
3026 : *
3027 : * Note that XLogSetAsyncXactLSN() performs similar calculation based on
3028 : * wal_writer_flush_after, to decide when to wake us up. Make sure the
3029 : * logic is the same in both places if you change this.
3030 : */
3031 10028 : now = GetCurrentTimestamp();
3032 10028 : flushblocks =
3033 10028 : WriteRqst.Write / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
3034 :
3035 10028 : if (WalWriterFlushAfter == 0 || lastflush == 0)
3036 : {
3037 : /* first call, or block based limits disabled */
3038 500 : WriteRqst.Flush = WriteRqst.Write;
3039 500 : lastflush = now;
3040 : }
3041 9528 : else if (TimestampDifferenceExceeds(lastflush, now, WalWriterDelay))
3042 : {
3043 : /*
3044 : * Flush the writes at least every WalWriterDelay ms. This is
3045 : * important to bound the amount of time it takes for an asynchronous
3046 : * commit to hit disk.
3047 : */
3048 9276 : WriteRqst.Flush = WriteRqst.Write;
3049 9276 : lastflush = now;
3050 : }
3051 252 : else if (flushblocks >= WalWriterFlushAfter)
3052 : {
3053 : /* exceeded wal_writer_flush_after blocks, flush */
3054 232 : WriteRqst.Flush = WriteRqst.Write;
3055 232 : lastflush = now;
3056 : }
3057 : else
3058 : {
3059 : /* no flushing, this time round */
3060 20 : WriteRqst.Flush = 0;
3061 : }
3062 :
3063 : #ifdef WAL_DEBUG
3064 : if (XLOG_DEBUG)
3065 : elog(LOG, "xlog bg flush request write %X/%08X; flush: %X/%08X, current is write %X/%08X; flush %X/%08X",
3066 : LSN_FORMAT_ARGS(WriteRqst.Write),
3067 : LSN_FORMAT_ARGS(WriteRqst.Flush),
3068 : LSN_FORMAT_ARGS(LogwrtResult.Write),
3069 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
3070 : #endif
3071 :
3072 10028 : START_CRIT_SECTION();
3073 :
3074 : /* now wait for any in-progress insertions to finish and get write lock */
3075 10028 : WaitXLogInsertionsToFinish(WriteRqst.Write);
3076 10028 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3077 10028 : RefreshXLogWriteResult(LogwrtResult);
3078 10028 : if (WriteRqst.Write > LogwrtResult.Write ||
3079 726 : WriteRqst.Flush > LogwrtResult.Flush)
3080 : {
3081 9664 : XLogWrite(WriteRqst, insertTLI, flexible);
3082 : }
3083 10028 : LWLockRelease(WALWriteLock);
3084 :
3085 10028 : END_CRIT_SECTION();
3086 :
3087 : /* wake up walsenders now that we've released heavily contended locks */
3088 10028 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3089 :
3090 : /*
3091 : * Great, done. To take some work off the critical path, try to initialize
3092 : * as many of the no-longer-needed WAL buffers for future use as we can.
3093 : */
3094 10028 : AdvanceXLInsertBuffer(InvalidXLogRecPtr, insertTLI, true);
3095 :
3096 : /*
3097 : * If we determined that we need to write data, but somebody else
3098 : * wrote/flushed already, it should be considered as being active, to
3099 : * avoid hibernating too early.
3100 : */
3101 10028 : return true;
3102 : }
3103 :
3104 : /*
3105 : * Test whether XLOG data has been flushed up to (at least) the given position.
3106 : *
3107 : * Returns true if a flush is still needed. (It may be that someone else
3108 : * is already in process of flushing that far, however.)
3109 : */
3110 : bool
3111 17050268 : XLogNeedsFlush(XLogRecPtr record)
3112 : {
3113 : /*
3114 : * During recovery, we don't flush WAL but update minRecoveryPoint
3115 : * instead. So "needs flush" is taken to mean whether minRecoveryPoint
3116 : * would need to be updated.
3117 : */
3118 17050268 : if (RecoveryInProgress())
3119 : {
3120 : /*
3121 : * An invalid minRecoveryPoint means that we need to recover all the
3122 : * WAL, i.e., we're doing crash recovery. We never modify the control
3123 : * file's value in that case, so we can short-circuit future checks
3124 : * here too. This triggers a quick exit path for the startup process,
3125 : * which cannot update its local copy of minRecoveryPoint as long as
3126 : * it has not replayed all WAL available when doing crash recovery.
3127 : */
3128 1011082 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint) && InRecovery)
3129 0 : updateMinRecoveryPoint = false;
3130 :
3131 : /* Quick exit if already known to be updated or cannot be updated */
3132 1011082 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3133 993414 : return false;
3134 :
3135 : /*
3136 : * Update local copy of minRecoveryPoint. But if the lock is busy,
3137 : * just return a conservative guess.
3138 : */
3139 17668 : if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3140 0 : return true;
3141 17668 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
3142 17668 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3143 17668 : LWLockRelease(ControlFileLock);
3144 :
3145 : /*
3146 : * Check minRecoveryPoint for any other process than the startup
3147 : * process doing crash recovery, which should not update the control
3148 : * file value if crash recovery is still running.
3149 : */
3150 17668 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint))
3151 0 : updateMinRecoveryPoint = false;
3152 :
3153 : /* check again */
3154 17668 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3155 152 : return false;
3156 : else
3157 17516 : return true;
3158 : }
3159 :
3160 : /* Quick exit if already known flushed */
3161 16039186 : if (record <= LogwrtResult.Flush)
3162 15673918 : return false;
3163 :
3164 : /* read LogwrtResult and update local state */
3165 365268 : RefreshXLogWriteResult(LogwrtResult);
3166 :
3167 : /* check again */
3168 365268 : if (record <= LogwrtResult.Flush)
3169 6142 : return false;
3170 :
3171 359126 : return true;
3172 : }
3173 :
3174 : /*
3175 : * Try to make a given XLOG file segment exist.
3176 : *
3177 : * logsegno: identify segment.
3178 : *
3179 : * *added: on return, true if this call raised the number of extant segments.
3180 : *
3181 : * path: on return, this char[MAXPGPATH] has the path to the logsegno file.
3182 : *
3183 : * Returns -1 or FD of opened file. A -1 here is not an error; a caller
3184 : * wanting an open segment should attempt to open "path", which usually will
3185 : * succeed. (This is weird, but it's efficient for the callers.)
3186 : */
3187 : static int
3188 28528 : XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,
3189 : bool *added, char *path)
3190 : {
3191 : char tmppath[MAXPGPATH];
3192 : XLogSegNo installed_segno;
3193 : XLogSegNo max_segno;
3194 : int fd;
3195 : int save_errno;
3196 28528 : int open_flags = O_RDWR | O_CREAT | O_EXCL | PG_BINARY;
3197 : instr_time io_start;
3198 :
3199 : Assert(logtli != 0);
3200 :
3201 28528 : XLogFilePath(path, logtli, logsegno, wal_segment_size);
3202 :
3203 : /*
3204 : * Try to use existent file (checkpoint maker may have created it already)
3205 : */
3206 28528 : *added = false;
3207 28528 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3208 28528 : get_sync_bit(wal_sync_method));
3209 28528 : if (fd < 0)
3210 : {
3211 2874 : if (errno != ENOENT)
3212 0 : ereport(ERROR,
3213 : (errcode_for_file_access(),
3214 : errmsg("could not open file \"%s\": %m", path)));
3215 : }
3216 : else
3217 25654 : return fd;
3218 :
3219 : /*
3220 : * Initialize an empty (all zeroes) segment. NOTE: it is possible that
3221 : * another process is doing the same thing. If so, we will end up
3222 : * pre-creating an extra log segment. That seems OK, and better than
3223 : * holding the lock throughout this lengthy process.
3224 : */
3225 2874 : elog(DEBUG2, "creating and filling new WAL file");
3226 :
3227 2874 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3228 :
3229 2874 : unlink(tmppath);
3230 :
3231 2874 : if (io_direct_flags & IO_DIRECT_WAL_INIT)
3232 0 : open_flags |= PG_O_DIRECT;
3233 :
3234 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3235 2874 : fd = BasicOpenFile(tmppath, open_flags);
3236 2874 : if (fd < 0)
3237 0 : ereport(ERROR,
3238 : (errcode_for_file_access(),
3239 : errmsg("could not create file \"%s\": %m", tmppath)));
3240 :
3241 : /* Measure I/O timing when initializing segment */
3242 2874 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3243 :
3244 2874 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE);
3245 2874 : save_errno = 0;
3246 2874 : if (wal_init_zero)
3247 : {
3248 : ssize_t rc;
3249 :
3250 : /*
3251 : * Zero-fill the file. With this setting, we do this the hard way to
3252 : * ensure that all the file space has really been allocated. On
3253 : * platforms that allow "holes" in files, just seeking to the end
3254 : * doesn't allocate intermediate space. This way, we know that we
3255 : * have all the space and (after the fsync below) that all the
3256 : * indirect blocks are down on disk. Therefore, fdatasync(2) or
3257 : * O_DSYNC will be sufficient to sync future writes to the log file.
3258 : */
3259 2874 : rc = pg_pwrite_zeros(fd, wal_segment_size, 0);
3260 :
3261 2874 : if (rc < 0)
3262 0 : save_errno = errno;
3263 : }
3264 : else
3265 : {
3266 : /*
3267 : * Otherwise, seeking to the end and writing a solitary byte is
3268 : * enough.
3269 : */
3270 0 : errno = 0;
3271 0 : if (pg_pwrite(fd, "\0", 1, wal_segment_size - 1) != 1)
3272 : {
3273 : /* if write didn't set errno, assume no disk space */
3274 0 : save_errno = errno ? errno : ENOSPC;
3275 : }
3276 : }
3277 2874 : pgstat_report_wait_end();
3278 :
3279 : /*
3280 : * A full segment worth of data is written when using wal_init_zero. One
3281 : * byte is written when not using it.
3282 : */
3283 2874 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT, IOOP_WRITE,
3284 : io_start, 1,
3285 2874 : wal_init_zero ? wal_segment_size : 1);
3286 :
3287 2874 : if (save_errno)
3288 : {
3289 : /*
3290 : * If we fail to make the file, delete it to release disk space
3291 : */
3292 0 : unlink(tmppath);
3293 :
3294 0 : close(fd);
3295 :
3296 0 : errno = save_errno;
3297 :
3298 0 : ereport(ERROR,
3299 : (errcode_for_file_access(),
3300 : errmsg("could not write to file \"%s\": %m", tmppath)));
3301 : }
3302 :
3303 : /* Measure I/O timing when flushing segment */
3304 2874 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3305 :
3306 2874 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC);
3307 2874 : if (pg_fsync(fd) != 0)
3308 : {
3309 0 : save_errno = errno;
3310 0 : close(fd);
3311 0 : errno = save_errno;
3312 0 : ereport(ERROR,
3313 : (errcode_for_file_access(),
3314 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3315 : }
3316 2874 : pgstat_report_wait_end();
3317 :
3318 2874 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT,
3319 : IOOP_FSYNC, io_start, 1, 0);
3320 :
3321 2874 : if (close(fd) != 0)
3322 0 : ereport(ERROR,
3323 : (errcode_for_file_access(),
3324 : errmsg("could not close file \"%s\": %m", tmppath)));
3325 :
3326 : /*
3327 : * Now move the segment into place with its final name. Cope with
3328 : * possibility that someone else has created the file while we were
3329 : * filling ours: if so, use ours to pre-create a future log segment.
3330 : */
3331 2874 : installed_segno = logsegno;
3332 :
3333 : /*
3334 : * XXX: What should we use as max_segno? We used to use XLOGfileslop when
3335 : * that was a constant, but that was always a bit dubious: normally, at a
3336 : * checkpoint, XLOGfileslop was the offset from the checkpoint record, but
3337 : * here, it was the offset from the insert location. We can't do the
3338 : * normal XLOGfileslop calculation here because we don't have access to
3339 : * the prior checkpoint's redo location. So somewhat arbitrarily, just use
3340 : * CheckPointSegments.
3341 : */
3342 2874 : max_segno = logsegno + CheckPointSegments;
3343 2874 : if (InstallXLogFileSegment(&installed_segno, tmppath, true, max_segno,
3344 : logtli))
3345 : {
3346 2874 : *added = true;
3347 2874 : elog(DEBUG2, "done creating and filling new WAL file");
3348 : }
3349 : else
3350 : {
3351 : /*
3352 : * No need for any more future segments, or InstallXLogFileSegment()
3353 : * failed to rename the file into place. If the rename failed, a
3354 : * caller opening the file may fail.
3355 : */
3356 0 : unlink(tmppath);
3357 0 : elog(DEBUG2, "abandoned new WAL file");
3358 : }
3359 :
3360 2874 : return -1;
3361 : }
3362 :
3363 : /*
3364 : * Create a new XLOG file segment, or open a pre-existing one.
3365 : *
3366 : * logsegno: identify segment to be created/opened.
3367 : *
3368 : * Returns FD of opened file.
3369 : *
3370 : * Note: errors here are ERROR not PANIC because we might or might not be
3371 : * inside a critical section (eg, during checkpoint there is no reason to
3372 : * take down the system on failure). They will promote to PANIC if we are
3373 : * in a critical section.
3374 : */
3375 : int
3376 28020 : XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)
3377 : {
3378 : bool ignore_added;
3379 : char path[MAXPGPATH];
3380 : int fd;
3381 :
3382 : Assert(logtli != 0);
3383 :
3384 28020 : fd = XLogFileInitInternal(logsegno, logtli, &ignore_added, path);
3385 28020 : if (fd >= 0)
3386 25436 : return fd;
3387 :
3388 : /* Now open original target segment (might not be file I just made) */
3389 2584 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3390 2584 : get_sync_bit(wal_sync_method));
3391 2584 : if (fd < 0)
3392 0 : ereport(ERROR,
3393 : (errcode_for_file_access(),
3394 : errmsg("could not open file \"%s\": %m", path)));
3395 2584 : return fd;
3396 : }
3397 :
3398 : /*
3399 : * Create a new XLOG file segment by copying a pre-existing one.
3400 : *
3401 : * destsegno: identify segment to be created.
3402 : *
3403 : * srcTLI, srcsegno: identify segment to be copied (could be from
3404 : * a different timeline)
3405 : *
3406 : * upto: how much of the source file to copy (the rest is filled with
3407 : * zeros)
3408 : *
3409 : * Currently this is only used during recovery, and so there are no locking
3410 : * considerations. But we should be just as tense as XLogFileInit to avoid
3411 : * emplacing a bogus file.
3412 : */
3413 : static void
3414 78 : XLogFileCopy(TimeLineID destTLI, XLogSegNo destsegno,
3415 : TimeLineID srcTLI, XLogSegNo srcsegno,
3416 : int upto)
3417 : {
3418 : char path[MAXPGPATH];
3419 : char tmppath[MAXPGPATH];
3420 : PGAlignedXLogBlock buffer;
3421 : int srcfd;
3422 : int fd;
3423 : int nbytes;
3424 :
3425 : /*
3426 : * Open the source file
3427 : */
3428 78 : XLogFilePath(path, srcTLI, srcsegno, wal_segment_size);
3429 78 : srcfd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
3430 78 : if (srcfd < 0)
3431 0 : ereport(ERROR,
3432 : (errcode_for_file_access(),
3433 : errmsg("could not open file \"%s\": %m", path)));
3434 :
3435 : /*
3436 : * Copy into a temp file name.
3437 : */
3438 78 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3439 :
3440 78 : unlink(tmppath);
3441 :
3442 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3443 78 : fd = OpenTransientFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
3444 78 : if (fd < 0)
3445 0 : ereport(ERROR,
3446 : (errcode_for_file_access(),
3447 : errmsg("could not create file \"%s\": %m", tmppath)));
3448 :
3449 : /*
3450 : * Do the data copying.
3451 : */
3452 159822 : for (nbytes = 0; nbytes < wal_segment_size; nbytes += sizeof(buffer))
3453 : {
3454 : int nread;
3455 :
3456 159744 : nread = upto - nbytes;
3457 :
3458 : /*
3459 : * The part that is not read from the source file is filled with
3460 : * zeros.
3461 : */
3462 159744 : if (nread < sizeof(buffer))
3463 78 : memset(buffer.data, 0, sizeof(buffer));
3464 :
3465 159744 : if (nread > 0)
3466 : {
3467 : int r;
3468 :
3469 5362 : if (nread > sizeof(buffer))
3470 5284 : nread = sizeof(buffer);
3471 5362 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ);
3472 5362 : r = read(srcfd, buffer.data, nread);
3473 5362 : if (r != nread)
3474 : {
3475 0 : if (r < 0)
3476 0 : ereport(ERROR,
3477 : (errcode_for_file_access(),
3478 : errmsg("could not read file \"%s\": %m",
3479 : path)));
3480 : else
3481 0 : ereport(ERROR,
3482 : (errcode(ERRCODE_DATA_CORRUPTED),
3483 : errmsg("could not read file \"%s\": read %d of %zu",
3484 : path, r, (Size) nread)));
3485 : }
3486 5362 : pgstat_report_wait_end();
3487 : }
3488 159744 : errno = 0;
3489 159744 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_WRITE);
3490 159744 : if ((int) write(fd, buffer.data, sizeof(buffer)) != (int) sizeof(buffer))
3491 : {
3492 0 : int save_errno = errno;
3493 :
3494 : /*
3495 : * If we fail to make the file, delete it to release disk space
3496 : */
3497 0 : unlink(tmppath);
3498 : /* if write didn't set errno, assume problem is no disk space */
3499 0 : errno = save_errno ? save_errno : ENOSPC;
3500 :
3501 0 : ereport(ERROR,
3502 : (errcode_for_file_access(),
3503 : errmsg("could not write to file \"%s\": %m", tmppath)));
3504 : }
3505 159744 : pgstat_report_wait_end();
3506 : }
3507 :
3508 78 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_SYNC);
3509 78 : if (pg_fsync(fd) != 0)
3510 0 : ereport(data_sync_elevel(ERROR),
3511 : (errcode_for_file_access(),
3512 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3513 78 : pgstat_report_wait_end();
3514 :
3515 78 : if (CloseTransientFile(fd) != 0)
3516 0 : ereport(ERROR,
3517 : (errcode_for_file_access(),
3518 : errmsg("could not close file \"%s\": %m", tmppath)));
3519 :
3520 78 : if (CloseTransientFile(srcfd) != 0)
3521 0 : ereport(ERROR,
3522 : (errcode_for_file_access(),
3523 : errmsg("could not close file \"%s\": %m", path)));
3524 :
3525 : /*
3526 : * Now move the segment into place with its final name.
3527 : */
3528 78 : if (!InstallXLogFileSegment(&destsegno, tmppath, false, 0, destTLI))
3529 0 : elog(ERROR, "InstallXLogFileSegment should not have failed");
3530 78 : }
3531 :
3532 : /*
3533 : * Install a new XLOG segment file as a current or future log segment.
3534 : *
3535 : * This is used both to install a newly-created segment (which has a temp
3536 : * filename while it's being created) and to recycle an old segment.
3537 : *
3538 : * *segno: identify segment to install as (or first possible target).
3539 : * When find_free is true, this is modified on return to indicate the
3540 : * actual installation location or last segment searched.
3541 : *
3542 : * tmppath: initial name of file to install. It will be renamed into place.
3543 : *
3544 : * find_free: if true, install the new segment at the first empty segno
3545 : * number at or after the passed numbers. If false, install the new segment
3546 : * exactly where specified, deleting any existing segment file there.
3547 : *
3548 : * max_segno: maximum segment number to install the new file as. Fail if no
3549 : * free slot is found between *segno and max_segno. (Ignored when find_free
3550 : * is false.)
3551 : *
3552 : * tli: The timeline on which the new segment should be installed.
3553 : *
3554 : * Returns true if the file was installed successfully. false indicates that
3555 : * max_segno limit was exceeded, the startup process has disabled this
3556 : * function for now, or an error occurred while renaming the file into place.
3557 : */
3558 : static bool
3559 5728 : InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
3560 : bool find_free, XLogSegNo max_segno, TimeLineID tli)
3561 : {
3562 : char path[MAXPGPATH];
3563 : struct stat stat_buf;
3564 :
3565 : Assert(tli != 0);
3566 :
3567 5728 : XLogFilePath(path, tli, *segno, wal_segment_size);
3568 :
3569 5728 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3570 5728 : if (!XLogCtl->InstallXLogFileSegmentActive)
3571 : {
3572 0 : LWLockRelease(ControlFileLock);
3573 0 : return false;
3574 : }
3575 :
3576 5728 : if (!find_free)
3577 : {
3578 : /* Force installation: get rid of any pre-existing segment file */
3579 78 : durable_unlink(path, DEBUG1);
3580 : }
3581 : else
3582 : {
3583 : /* Find a free slot to put it in */
3584 8108 : while (stat(path, &stat_buf) == 0)
3585 : {
3586 2530 : if ((*segno) >= max_segno)
3587 : {
3588 : /* Failed to find a free slot within specified range */
3589 72 : LWLockRelease(ControlFileLock);
3590 72 : return false;
3591 : }
3592 2458 : (*segno)++;
3593 2458 : XLogFilePath(path, tli, *segno, wal_segment_size);
3594 : }
3595 : }
3596 :
3597 : Assert(access(path, F_OK) != 0 && errno == ENOENT);
3598 5656 : if (durable_rename(tmppath, path, LOG) != 0)
3599 : {
3600 0 : LWLockRelease(ControlFileLock);
3601 : /* durable_rename already emitted log message */
3602 0 : return false;
3603 : }
3604 :
3605 5656 : LWLockRelease(ControlFileLock);
3606 :
3607 5656 : return true;
3608 : }
3609 :
3610 : /*
3611 : * Open a pre-existing logfile segment for writing.
3612 : */
3613 : int
3614 324 : XLogFileOpen(XLogSegNo segno, TimeLineID tli)
3615 : {
3616 : char path[MAXPGPATH];
3617 : int fd;
3618 :
3619 324 : XLogFilePath(path, tli, segno, wal_segment_size);
3620 :
3621 324 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3622 324 : get_sync_bit(wal_sync_method));
3623 324 : if (fd < 0)
3624 0 : ereport(PANIC,
3625 : (errcode_for_file_access(),
3626 : errmsg("could not open file \"%s\": %m", path)));
3627 :
3628 324 : return fd;
3629 : }
3630 :
3631 : /*
3632 : * Close the current logfile segment for writing.
3633 : */
3634 : static void
3635 12358 : XLogFileClose(void)
3636 : {
3637 : Assert(openLogFile >= 0);
3638 :
3639 : /*
3640 : * WAL segment files will not be re-read in normal operation, so we advise
3641 : * the OS to release any cached pages. But do not do so if WAL archiving
3642 : * or streaming is active, because archiver and walsender process could
3643 : * use the cache to read the WAL segment.
3644 : */
3645 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3646 12358 : if (!XLogIsNeeded() && (io_direct_flags & IO_DIRECT_WAL) == 0)
3647 2770 : (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3648 : #endif
3649 :
3650 12358 : if (close(openLogFile) != 0)
3651 : {
3652 : char xlogfname[MAXFNAMELEN];
3653 0 : int save_errno = errno;
3654 :
3655 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo, wal_segment_size);
3656 0 : errno = save_errno;
3657 0 : ereport(PANIC,
3658 : (errcode_for_file_access(),
3659 : errmsg("could not close file \"%s\": %m", xlogfname)));
3660 : }
3661 :
3662 12358 : openLogFile = -1;
3663 12358 : ReleaseExternalFD();
3664 12358 : }
3665 :
3666 : /*
3667 : * Preallocate log files beyond the specified log endpoint.
3668 : *
3669 : * XXX this is currently extremely conservative, since it forces only one
3670 : * future log segment to exist, and even that only if we are 75% done with
3671 : * the current one. This is only appropriate for very low-WAL-volume systems.
3672 : * High-volume systems will be OK once they've built up a sufficient set of
3673 : * recycled log segments, but the startup transient is likely to include
3674 : * a lot of segment creations by foreground processes, which is not so good.
3675 : *
3676 : * XLogFileInitInternal() can ereport(ERROR). All known causes indicate big
3677 : * trouble; for example, a full filesystem is one cause. The checkpoint WAL
3678 : * and/or ControlFile updates already completed. If a RequestCheckpoint()
3679 : * initiated the present checkpoint and an ERROR ends this function, the
3680 : * command that called RequestCheckpoint() fails. That's not ideal, but it's
3681 : * not worth contorting more functions to use caller-specified elevel values.
3682 : * (With or without RequestCheckpoint(), an ERROR forestalls some inessential
3683 : * reporting and resource reclamation.)
3684 : */
3685 : static void
3686 3958 : PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli)
3687 : {
3688 : XLogSegNo _logSegNo;
3689 : int lf;
3690 : bool added;
3691 : char path[MAXPGPATH];
3692 : uint64 offset;
3693 :
3694 3958 : if (!XLogCtl->InstallXLogFileSegmentActive)
3695 14 : return; /* unlocked check says no */
3696 :
3697 3944 : XLByteToPrevSeg(endptr, _logSegNo, wal_segment_size);
3698 3944 : offset = XLogSegmentOffset(endptr - 1, wal_segment_size);
3699 3944 : if (offset >= (uint32) (0.75 * wal_segment_size))
3700 : {
3701 508 : _logSegNo++;
3702 508 : lf = XLogFileInitInternal(_logSegNo, tli, &added, path);
3703 508 : if (lf >= 0)
3704 218 : close(lf);
3705 508 : if (added)
3706 290 : CheckpointStats.ckpt_segs_added++;
3707 : }
3708 : }
3709 :
3710 : /*
3711 : * Throws an error if the given log segment has already been removed or
3712 : * recycled. The caller should only pass a segment that it knows to have
3713 : * existed while the server has been running, as this function always
3714 : * succeeds if no WAL segments have been removed since startup.
3715 : * 'tli' is only used in the error message.
3716 : *
3717 : * Note: this function guarantees to keep errno unchanged on return.
3718 : * This supports callers that use this to possibly deliver a better
3719 : * error message about a missing file, while still being able to throw
3720 : * a normal file-access error afterwards, if this does return.
3721 : */
3722 : void
3723 230550 : CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3724 : {
3725 230550 : int save_errno = errno;
3726 : XLogSegNo lastRemovedSegNo;
3727 :
3728 230550 : SpinLockAcquire(&XLogCtl->info_lck);
3729 230550 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3730 230550 : SpinLockRelease(&XLogCtl->info_lck);
3731 :
3732 230550 : if (segno <= lastRemovedSegNo)
3733 : {
3734 : char filename[MAXFNAMELEN];
3735 :
3736 0 : XLogFileName(filename, tli, segno, wal_segment_size);
3737 0 : errno = save_errno;
3738 0 : ereport(ERROR,
3739 : (errcode_for_file_access(),
3740 : errmsg("requested WAL segment %s has already been removed",
3741 : filename)));
3742 : }
3743 230550 : errno = save_errno;
3744 230550 : }
3745 :
3746 : /*
3747 : * Return the last WAL segment removed, or 0 if no segment has been removed
3748 : * since startup.
3749 : *
3750 : * NB: the result can be out of date arbitrarily fast, the caller has to deal
3751 : * with that.
3752 : */
3753 : XLogSegNo
3754 2164 : XLogGetLastRemovedSegno(void)
3755 : {
3756 : XLogSegNo lastRemovedSegNo;
3757 :
3758 2164 : SpinLockAcquire(&XLogCtl->info_lck);
3759 2164 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3760 2164 : SpinLockRelease(&XLogCtl->info_lck);
3761 :
3762 2164 : return lastRemovedSegNo;
3763 : }
3764 :
3765 : /*
3766 : * Return the oldest WAL segment on the given TLI that still exists in
3767 : * XLOGDIR, or 0 if none.
3768 : */
3769 : XLogSegNo
3770 10 : XLogGetOldestSegno(TimeLineID tli)
3771 : {
3772 : DIR *xldir;
3773 : struct dirent *xlde;
3774 10 : XLogSegNo oldest_segno = 0;
3775 :
3776 10 : xldir = AllocateDir(XLOGDIR);
3777 68 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3778 : {
3779 : TimeLineID file_tli;
3780 : XLogSegNo file_segno;
3781 :
3782 : /* Ignore files that are not XLOG segments. */
3783 58 : if (!IsXLogFileName(xlde->d_name))
3784 40 : continue;
3785 :
3786 : /* Parse filename to get TLI and segno. */
3787 18 : XLogFromFileName(xlde->d_name, &file_tli, &file_segno,
3788 : wal_segment_size);
3789 :
3790 : /* Ignore anything that's not from the TLI of interest. */
3791 18 : if (tli != file_tli)
3792 0 : continue;
3793 :
3794 : /* If it's the oldest so far, update oldest_segno. */
3795 18 : if (oldest_segno == 0 || file_segno < oldest_segno)
3796 16 : oldest_segno = file_segno;
3797 : }
3798 :
3799 10 : FreeDir(xldir);
3800 10 : return oldest_segno;
3801 : }
3802 :
3803 : /*
3804 : * Update the last removed segno pointer in shared memory, to reflect that the
3805 : * given XLOG file has been removed.
3806 : */
3807 : static void
3808 5034 : UpdateLastRemovedPtr(char *filename)
3809 : {
3810 : uint32 tli;
3811 : XLogSegNo segno;
3812 :
3813 5034 : XLogFromFileName(filename, &tli, &segno, wal_segment_size);
3814 :
3815 5034 : SpinLockAcquire(&XLogCtl->info_lck);
3816 5034 : if (segno > XLogCtl->lastRemovedSegNo)
3817 2040 : XLogCtl->lastRemovedSegNo = segno;
3818 5034 : SpinLockRelease(&XLogCtl->info_lck);
3819 5034 : }
3820 :
3821 : /*
3822 : * Remove all temporary log files in pg_wal
3823 : *
3824 : * This is called at the beginning of recovery after a previous crash,
3825 : * at a point where no other processes write fresh WAL data.
3826 : */
3827 : static void
3828 350 : RemoveTempXlogFiles(void)
3829 : {
3830 : DIR *xldir;
3831 : struct dirent *xlde;
3832 :
3833 350 : elog(DEBUG2, "removing all temporary WAL segments");
3834 :
3835 350 : xldir = AllocateDir(XLOGDIR);
3836 2384 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3837 : {
3838 : char path[MAXPGPATH];
3839 :
3840 2034 : if (strncmp(xlde->d_name, "xlogtemp.", 9) != 0)
3841 2034 : continue;
3842 :
3843 0 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
3844 0 : unlink(path);
3845 0 : elog(DEBUG2, "removed temporary WAL segment \"%s\"", path);
3846 : }
3847 350 : FreeDir(xldir);
3848 350 : }
3849 :
3850 : /*
3851 : * Recycle or remove all log files older or equal to passed segno.
3852 : *
3853 : * endptr is current (or recent) end of xlog, and lastredoptr is the
3854 : * redo pointer of the last checkpoint. These are used to determine
3855 : * whether we want to recycle rather than delete no-longer-wanted log files.
3856 : *
3857 : * insertTLI is the current timeline for XLOG insertion. Any recycled
3858 : * segments should be reused for this timeline.
3859 : */
3860 : static void
3861 3428 : RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr, XLogRecPtr endptr,
3862 : TimeLineID insertTLI)
3863 : {
3864 : DIR *xldir;
3865 : struct dirent *xlde;
3866 : char lastoff[MAXFNAMELEN];
3867 : XLogSegNo endlogSegNo;
3868 : XLogSegNo recycleSegNo;
3869 :
3870 : /* Initialize info about where to try to recycle to */
3871 3428 : XLByteToSeg(endptr, endlogSegNo, wal_segment_size);
3872 3428 : recycleSegNo = XLOGfileslop(lastredoptr);
3873 :
3874 : /*
3875 : * Construct a filename of the last segment to be kept. The timeline ID
3876 : * doesn't matter, we ignore that in the comparison. (During recovery,
3877 : * InsertTimeLineID isn't set, so we can't use that.)
3878 : */
3879 3428 : XLogFileName(lastoff, 0, segno, wal_segment_size);
3880 :
3881 3428 : elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
3882 : lastoff);
3883 :
3884 3428 : xldir = AllocateDir(XLOGDIR);
3885 :
3886 84618 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3887 : {
3888 : /* Ignore files that are not XLOG segments */
3889 81190 : if (!IsXLogFileName(xlde->d_name) &&
3890 14794 : !IsPartialXLogFileName(xlde->d_name))
3891 14790 : continue;
3892 :
3893 : /*
3894 : * We ignore the timeline part of the XLOG segment identifiers in
3895 : * deciding whether a segment is still needed. This ensures that we
3896 : * won't prematurely remove a segment from a parent timeline. We could
3897 : * probably be a little more proactive about removing segments of
3898 : * non-parent timelines, but that would be a whole lot more
3899 : * complicated.
3900 : *
3901 : * We use the alphanumeric sorting property of the filenames to decide
3902 : * which ones are earlier than the lastoff segment.
3903 : */
3904 66400 : if (strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
3905 : {
3906 40230 : if (XLogArchiveCheckDone(xlde->d_name))
3907 : {
3908 : /* Update the last removed location in shared memory first */
3909 5034 : UpdateLastRemovedPtr(xlde->d_name);
3910 :
3911 5034 : RemoveXlogFile(xlde, recycleSegNo, &endlogSegNo, insertTLI);
3912 : }
3913 : }
3914 : }
3915 :
3916 3428 : FreeDir(xldir);
3917 3428 : }
3918 :
3919 : /*
3920 : * Recycle or remove WAL files that are not part of the given timeline's
3921 : * history.
3922 : *
3923 : * This is called during recovery, whenever we switch to follow a new
3924 : * timeline, and at the end of recovery when we create a new timeline. We
3925 : * wouldn't otherwise care about extra WAL files lying in pg_wal, but they
3926 : * might be leftover pre-allocated or recycled WAL segments on the old timeline
3927 : * that we haven't used yet, and contain garbage. If we just leave them in
3928 : * pg_wal, they will eventually be archived, and we can't let that happen.
3929 : * Files that belong to our timeline history are valid, because we have
3930 : * successfully replayed them, but from others we can't be sure.
3931 : *
3932 : * 'switchpoint' is the current point in WAL where we switch to new timeline,
3933 : * and 'newTLI' is the new timeline we switch to.
3934 : */
3935 : void
3936 120 : RemoveNonParentXlogFiles(XLogRecPtr switchpoint, TimeLineID newTLI)
3937 : {
3938 : DIR *xldir;
3939 : struct dirent *xlde;
3940 : char switchseg[MAXFNAMELEN];
3941 : XLogSegNo endLogSegNo;
3942 : XLogSegNo switchLogSegNo;
3943 : XLogSegNo recycleSegNo;
3944 :
3945 : /*
3946 : * Initialize info about where to begin the work. This will recycle,
3947 : * somewhat arbitrarily, 10 future segments.
3948 : */
3949 120 : XLByteToPrevSeg(switchpoint, switchLogSegNo, wal_segment_size);
3950 120 : XLByteToSeg(switchpoint, endLogSegNo, wal_segment_size);
3951 120 : recycleSegNo = endLogSegNo + 10;
3952 :
3953 : /*
3954 : * Construct a filename of the last segment to be kept.
3955 : */
3956 120 : XLogFileName(switchseg, newTLI, switchLogSegNo, wal_segment_size);
3957 :
3958 120 : elog(DEBUG2, "attempting to remove WAL segments newer than log file %s",
3959 : switchseg);
3960 :
3961 120 : xldir = AllocateDir(XLOGDIR);
3962 :
3963 1144 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3964 : {
3965 : /* Ignore files that are not XLOG segments */
3966 1024 : if (!IsXLogFileName(xlde->d_name))
3967 636 : continue;
3968 :
3969 : /*
3970 : * Remove files that are on a timeline older than the new one we're
3971 : * switching to, but with a segment number >= the first segment on the
3972 : * new timeline.
3973 : */
3974 388 : if (strncmp(xlde->d_name, switchseg, 8) < 0 &&
3975 250 : strcmp(xlde->d_name + 8, switchseg + 8) > 0)
3976 : {
3977 : /*
3978 : * If the file has already been marked as .ready, however, don't
3979 : * remove it yet. It should be OK to remove it - files that are
3980 : * not part of our timeline history are not required for recovery
3981 : * - but seems safer to let them be archived and removed later.
3982 : */
3983 34 : if (!XLogArchiveIsReady(xlde->d_name))
3984 34 : RemoveXlogFile(xlde, recycleSegNo, &endLogSegNo, newTLI);
3985 : }
3986 : }
3987 :
3988 120 : FreeDir(xldir);
3989 120 : }
3990 :
3991 : /*
3992 : * Recycle or remove a log file that's no longer needed.
3993 : *
3994 : * segment_de is the dirent structure of the segment to recycle or remove.
3995 : * recycleSegNo is the segment number to recycle up to. endlogSegNo is
3996 : * the segment number of the current (or recent) end of WAL.
3997 : *
3998 : * endlogSegNo gets incremented if the segment is recycled so as it is not
3999 : * checked again with future callers of this function.
4000 : *
4001 : * insertTLI is the current timeline for XLOG insertion. Any recycled segments
4002 : * should be used for this timeline.
4003 : */
4004 : static void
4005 5068 : RemoveXlogFile(const struct dirent *segment_de,
4006 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
4007 : TimeLineID insertTLI)
4008 : {
4009 : char path[MAXPGPATH];
4010 : #ifdef WIN32
4011 : char newpath[MAXPGPATH];
4012 : #endif
4013 5068 : const char *segname = segment_de->d_name;
4014 :
4015 5068 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", segname);
4016 :
4017 : /*
4018 : * Before deleting the file, see if it can be recycled as a future log
4019 : * segment. Only recycle normal files, because we don't want to recycle
4020 : * symbolic links pointing to a separate archive directory.
4021 : */
4022 5068 : if (wal_recycle &&
4023 5068 : *endlogSegNo <= recycleSegNo &&
4024 6204 : XLogCtl->InstallXLogFileSegmentActive && /* callee rechecks this */
4025 5552 : get_dirent_type(path, segment_de, false, DEBUG2) == PGFILETYPE_REG &&
4026 2776 : InstallXLogFileSegment(endlogSegNo, path,
4027 : true, recycleSegNo, insertTLI))
4028 : {
4029 2704 : ereport(DEBUG2,
4030 : (errmsg_internal("recycled write-ahead log file \"%s\"",
4031 : segname)));
4032 2704 : CheckpointStats.ckpt_segs_recycled++;
4033 : /* Needn't recheck that slot on future iterations */
4034 2704 : (*endlogSegNo)++;
4035 : }
4036 : else
4037 : {
4038 : /* No need for any more future segments, or recycling failed ... */
4039 : int rc;
4040 :
4041 2364 : ereport(DEBUG2,
4042 : (errmsg_internal("removing write-ahead log file \"%s\"",
4043 : segname)));
4044 :
4045 : #ifdef WIN32
4046 :
4047 : /*
4048 : * On Windows, if another process (e.g another backend) holds the file
4049 : * open in FILE_SHARE_DELETE mode, unlink will succeed, but the file
4050 : * will still show up in directory listing until the last handle is
4051 : * closed. To avoid confusing the lingering deleted file for a live
4052 : * WAL file that needs to be archived, rename it before deleting it.
4053 : *
4054 : * If another process holds the file open without FILE_SHARE_DELETE
4055 : * flag, rename will fail. We'll try again at the next checkpoint.
4056 : */
4057 : snprintf(newpath, MAXPGPATH, "%s.deleted", path);
4058 : if (rename(path, newpath) != 0)
4059 : {
4060 : ereport(LOG,
4061 : (errcode_for_file_access(),
4062 : errmsg("could not rename file \"%s\": %m",
4063 : path)));
4064 : return;
4065 : }
4066 : rc = durable_unlink(newpath, LOG);
4067 : #else
4068 2364 : rc = durable_unlink(path, LOG);
4069 : #endif
4070 2364 : if (rc != 0)
4071 : {
4072 : /* Message already logged by durable_unlink() */
4073 0 : return;
4074 : }
4075 2364 : CheckpointStats.ckpt_segs_removed++;
4076 : }
4077 :
4078 5068 : XLogArchiveCleanup(segname);
4079 : }
4080 :
4081 : /*
4082 : * Verify whether pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
4083 : * If the latter do not exist, recreate them.
4084 : *
4085 : * It is not the goal of this function to verify the contents of these
4086 : * directories, but to help in cases where someone has performed a cluster
4087 : * copy for PITR purposes but omitted pg_wal from the copy.
4088 : *
4089 : * We could also recreate pg_wal if it doesn't exist, but a deliberate
4090 : * policy decision was made not to. It is fairly common for pg_wal to be
4091 : * a symlink, and if that was the DBA's intent then automatically making a
4092 : * plain directory would result in degraded performance with no notice.
4093 : */
4094 : static void
4095 1890 : ValidateXLOGDirectoryStructure(void)
4096 : {
4097 : char path[MAXPGPATH];
4098 : struct stat stat_buf;
4099 :
4100 : /* Check for pg_wal; if it doesn't exist, error out */
4101 1890 : if (stat(XLOGDIR, &stat_buf) != 0 ||
4102 1890 : !S_ISDIR(stat_buf.st_mode))
4103 0 : ereport(FATAL,
4104 : (errcode_for_file_access(),
4105 : errmsg("required WAL directory \"%s\" does not exist",
4106 : XLOGDIR)));
4107 :
4108 : /* Check for archive_status */
4109 1890 : snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4110 1890 : if (stat(path, &stat_buf) == 0)
4111 : {
4112 : /* Check for weird cases where it exists but isn't a directory */
4113 1888 : if (!S_ISDIR(stat_buf.st_mode))
4114 0 : ereport(FATAL,
4115 : (errcode_for_file_access(),
4116 : errmsg("required WAL directory \"%s\" does not exist",
4117 : path)));
4118 : }
4119 : else
4120 : {
4121 2 : ereport(LOG,
4122 : (errmsg("creating missing WAL directory \"%s\"", path)));
4123 2 : if (MakePGDirectory(path) < 0)
4124 0 : ereport(FATAL,
4125 : (errcode_for_file_access(),
4126 : errmsg("could not create missing directory \"%s\": %m",
4127 : path)));
4128 : }
4129 :
4130 : /* Check for summaries */
4131 1890 : snprintf(path, MAXPGPATH, XLOGDIR "/summaries");
4132 1890 : if (stat(path, &stat_buf) == 0)
4133 : {
4134 : /* Check for weird cases where it exists but isn't a directory */
4135 1888 : if (!S_ISDIR(stat_buf.st_mode))
4136 0 : ereport(FATAL,
4137 : (errmsg("required WAL directory \"%s\" does not exist",
4138 : path)));
4139 : }
4140 : else
4141 : {
4142 2 : ereport(LOG,
4143 : (errmsg("creating missing WAL directory \"%s\"", path)));
4144 2 : if (MakePGDirectory(path) < 0)
4145 0 : ereport(FATAL,
4146 : (errmsg("could not create missing directory \"%s\": %m",
4147 : path)));
4148 : }
4149 1890 : }
4150 :
4151 : /*
4152 : * Remove previous backup history files. This also retries creation of
4153 : * .ready files for any backup history files for which XLogArchiveNotify
4154 : * failed earlier.
4155 : */
4156 : static void
4157 296 : CleanupBackupHistory(void)
4158 : {
4159 : DIR *xldir;
4160 : struct dirent *xlde;
4161 : char path[MAXPGPATH + sizeof(XLOGDIR)];
4162 :
4163 296 : xldir = AllocateDir(XLOGDIR);
4164 :
4165 2974 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4166 : {
4167 2382 : if (IsBackupHistoryFileName(xlde->d_name))
4168 : {
4169 312 : if (XLogArchiveCheckDone(xlde->d_name))
4170 : {
4171 246 : elog(DEBUG2, "removing WAL backup history file \"%s\"",
4172 : xlde->d_name);
4173 246 : snprintf(path, sizeof(path), XLOGDIR "/%s", xlde->d_name);
4174 246 : unlink(path);
4175 246 : XLogArchiveCleanup(xlde->d_name);
4176 : }
4177 : }
4178 : }
4179 :
4180 296 : FreeDir(xldir);
4181 296 : }
4182 :
4183 : /*
4184 : * I/O routines for pg_control
4185 : *
4186 : * *ControlFile is a buffer in shared memory that holds an image of the
4187 : * contents of pg_control. WriteControlFile() initializes pg_control
4188 : * given a preloaded buffer, ReadControlFile() loads the buffer from
4189 : * the pg_control file (during postmaster or standalone-backend startup),
4190 : * and UpdateControlFile() rewrites pg_control after we modify xlog state.
4191 : * InitControlFile() fills the buffer with initial values.
4192 : *
4193 : * For simplicity, WriteControlFile() initializes the fields of pg_control
4194 : * that are related to checking backend/database compatibility, and
4195 : * ReadControlFile() verifies they are correct. We could split out the
4196 : * I/O and compatibility-check functions, but there seems no need currently.
4197 : */
4198 :
4199 : static void
4200 100 : InitControlFile(uint64 sysidentifier, uint32 data_checksum_version)
4201 : {
4202 : char mock_auth_nonce[MOCK_AUTH_NONCE_LEN];
4203 :
4204 : /*
4205 : * Generate a random nonce. This is used for authentication requests that
4206 : * will fail because the user does not exist. The nonce is used to create
4207 : * a genuine-looking password challenge for the non-existent user, in lieu
4208 : * of an actual stored password.
4209 : */
4210 100 : if (!pg_strong_random(mock_auth_nonce, MOCK_AUTH_NONCE_LEN))
4211 0 : ereport(PANIC,
4212 : (errcode(ERRCODE_INTERNAL_ERROR),
4213 : errmsg("could not generate secret authorization token")));
4214 :
4215 100 : memset(ControlFile, 0, sizeof(ControlFileData));
4216 : /* Initialize pg_control status fields */
4217 100 : ControlFile->system_identifier = sysidentifier;
4218 100 : memcpy(ControlFile->mock_authentication_nonce, mock_auth_nonce, MOCK_AUTH_NONCE_LEN);
4219 100 : ControlFile->state = DB_SHUTDOWNED;
4220 100 : ControlFile->unloggedLSN = FirstNormalUnloggedLSN;
4221 :
4222 : /* Set important parameter values for use when replaying WAL */
4223 100 : ControlFile->MaxConnections = MaxConnections;
4224 100 : ControlFile->max_worker_processes = max_worker_processes;
4225 100 : ControlFile->max_wal_senders = max_wal_senders;
4226 100 : ControlFile->max_prepared_xacts = max_prepared_xacts;
4227 100 : ControlFile->max_locks_per_xact = max_locks_per_xact;
4228 100 : ControlFile->wal_level = wal_level;
4229 100 : ControlFile->wal_log_hints = wal_log_hints;
4230 100 : ControlFile->track_commit_timestamp = track_commit_timestamp;
4231 100 : ControlFile->data_checksum_version = data_checksum_version;
4232 100 : }
4233 :
4234 : static void
4235 100 : WriteControlFile(void)
4236 : {
4237 : int fd;
4238 : char buffer[PG_CONTROL_FILE_SIZE]; /* need not be aligned */
4239 :
4240 : /*
4241 : * Initialize version and compatibility-check fields
4242 : */
4243 100 : ControlFile->pg_control_version = PG_CONTROL_VERSION;
4244 100 : ControlFile->catalog_version_no = CATALOG_VERSION_NO;
4245 :
4246 100 : ControlFile->maxAlign = MAXIMUM_ALIGNOF;
4247 100 : ControlFile->floatFormat = FLOATFORMAT_VALUE;
4248 :
4249 100 : ControlFile->blcksz = BLCKSZ;
4250 100 : ControlFile->relseg_size = RELSEG_SIZE;
4251 100 : ControlFile->xlog_blcksz = XLOG_BLCKSZ;
4252 100 : ControlFile->xlog_seg_size = wal_segment_size;
4253 :
4254 100 : ControlFile->nameDataLen = NAMEDATALEN;
4255 100 : ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
4256 :
4257 100 : ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
4258 100 : ControlFile->loblksize = LOBLKSIZE;
4259 :
4260 100 : ControlFile->float8ByVal = true; /* vestigial */
4261 :
4262 : /*
4263 : * Initialize the default 'char' signedness.
4264 : *
4265 : * The signedness of the char type is implementation-defined. For instance
4266 : * on x86 architecture CPUs, the char data type is typically treated as
4267 : * signed by default, whereas on aarch architecture CPUs, it is typically
4268 : * treated as unsigned by default. In v17 or earlier, we accidentally let
4269 : * C implementation signedness affect persistent data. This led to
4270 : * inconsistent results when comparing char data across different
4271 : * platforms.
4272 : *
4273 : * This flag can be used as a hint to ensure consistent behavior for
4274 : * pre-v18 data files that store data sorted by the 'char' type on disk,
4275 : * especially in cross-platform replication scenarios.
4276 : *
4277 : * Newly created database clusters unconditionally set the default char
4278 : * signedness to true. pg_upgrade changes this flag for clusters that were
4279 : * initialized on signedness=false platforms. As a result,
4280 : * signedness=false setting will become rare over time. If we had known
4281 : * about this problem during the last development cycle that forced initdb
4282 : * (v8.3), we would have made all clusters signed or all clusters
4283 : * unsigned. Making pg_upgrade the only source of signedness=false will
4284 : * cause the population of database clusters to converge toward that
4285 : * retrospective ideal.
4286 : */
4287 100 : ControlFile->default_char_signedness = true;
4288 :
4289 : /* Contents are protected with a CRC */
4290 100 : INIT_CRC32C(ControlFile->crc);
4291 100 : COMP_CRC32C(ControlFile->crc,
4292 : ControlFile,
4293 : offsetof(ControlFileData, crc));
4294 100 : FIN_CRC32C(ControlFile->crc);
4295 :
4296 : /*
4297 : * We write out PG_CONTROL_FILE_SIZE bytes into pg_control, zero-padding
4298 : * the excess over sizeof(ControlFileData). This reduces the odds of
4299 : * premature-EOF errors when reading pg_control. We'll still fail when we
4300 : * check the contents of the file, but hopefully with a more specific
4301 : * error than "couldn't read pg_control".
4302 : */
4303 100 : memset(buffer, 0, PG_CONTROL_FILE_SIZE);
4304 100 : memcpy(buffer, ControlFile, sizeof(ControlFileData));
4305 :
4306 100 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4307 : O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
4308 100 : if (fd < 0)
4309 0 : ereport(PANIC,
4310 : (errcode_for_file_access(),
4311 : errmsg("could not create file \"%s\": %m",
4312 : XLOG_CONTROL_FILE)));
4313 :
4314 100 : errno = 0;
4315 100 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_WRITE);
4316 100 : if (write(fd, buffer, PG_CONTROL_FILE_SIZE) != PG_CONTROL_FILE_SIZE)
4317 : {
4318 : /* if write didn't set errno, assume problem is no disk space */
4319 0 : if (errno == 0)
4320 0 : errno = ENOSPC;
4321 0 : ereport(PANIC,
4322 : (errcode_for_file_access(),
4323 : errmsg("could not write to file \"%s\": %m",
4324 : XLOG_CONTROL_FILE)));
4325 : }
4326 100 : pgstat_report_wait_end();
4327 :
4328 100 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_SYNC);
4329 100 : if (pg_fsync(fd) != 0)
4330 0 : ereport(PANIC,
4331 : (errcode_for_file_access(),
4332 : errmsg("could not fsync file \"%s\": %m",
4333 : XLOG_CONTROL_FILE)));
4334 100 : pgstat_report_wait_end();
4335 :
4336 100 : if (close(fd) != 0)
4337 0 : ereport(PANIC,
4338 : (errcode_for_file_access(),
4339 : errmsg("could not close file \"%s\": %m",
4340 : XLOG_CONTROL_FILE)));
4341 100 : }
4342 :
4343 : static void
4344 1990 : ReadControlFile(void)
4345 : {
4346 : pg_crc32c crc;
4347 : int fd;
4348 : char wal_segsz_str[20];
4349 : int r;
4350 :
4351 : /*
4352 : * Read data...
4353 : */
4354 1990 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4355 : O_RDWR | PG_BINARY);
4356 1990 : if (fd < 0)
4357 0 : ereport(PANIC,
4358 : (errcode_for_file_access(),
4359 : errmsg("could not open file \"%s\": %m",
4360 : XLOG_CONTROL_FILE)));
4361 :
4362 1990 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ);
4363 1990 : r = read(fd, ControlFile, sizeof(ControlFileData));
4364 1990 : if (r != sizeof(ControlFileData))
4365 : {
4366 0 : if (r < 0)
4367 0 : ereport(PANIC,
4368 : (errcode_for_file_access(),
4369 : errmsg("could not read file \"%s\": %m",
4370 : XLOG_CONTROL_FILE)));
4371 : else
4372 0 : ereport(PANIC,
4373 : (errcode(ERRCODE_DATA_CORRUPTED),
4374 : errmsg("could not read file \"%s\": read %d of %zu",
4375 : XLOG_CONTROL_FILE, r, sizeof(ControlFileData))));
4376 : }
4377 1990 : pgstat_report_wait_end();
4378 :
4379 1990 : close(fd);
4380 :
4381 : /*
4382 : * Check for expected pg_control format version. If this is wrong, the
4383 : * CRC check will likely fail because we'll be checking the wrong number
4384 : * of bytes. Complaining about wrong version will probably be more
4385 : * enlightening than complaining about wrong CRC.
4386 : */
4387 :
4388 1990 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
4389 0 : ereport(FATAL,
4390 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4391 : errmsg("database files are incompatible with server"),
4392 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4393 : " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4394 : ControlFile->pg_control_version, ControlFile->pg_control_version,
4395 : PG_CONTROL_VERSION, PG_CONTROL_VERSION),
4396 : errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4397 :
4398 1990 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
4399 0 : ereport(FATAL,
4400 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4401 : errmsg("database files are incompatible with server"),
4402 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4403 : " but the server was compiled with PG_CONTROL_VERSION %d.",
4404 : ControlFile->pg_control_version, PG_CONTROL_VERSION),
4405 : errhint("It looks like you need to initdb.")));
4406 :
4407 : /* Now check the CRC. */
4408 1990 : INIT_CRC32C(crc);
4409 1990 : COMP_CRC32C(crc,
4410 : ControlFile,
4411 : offsetof(ControlFileData, crc));
4412 1990 : FIN_CRC32C(crc);
4413 :
4414 1990 : if (!EQ_CRC32C(crc, ControlFile->crc))
4415 0 : ereport(FATAL,
4416 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4417 : errmsg("incorrect checksum in control file")));
4418 :
4419 : /*
4420 : * Do compatibility checking immediately. If the database isn't
4421 : * compatible with the backend executable, we want to abort before we can
4422 : * possibly do any damage.
4423 : */
4424 1990 : if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
4425 0 : ereport(FATAL,
4426 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4427 : errmsg("database files are incompatible with server"),
4428 : /* translator: %s is a variable name and %d is its value */
4429 : errdetail("The database cluster was initialized with %s %d,"
4430 : " but the server was compiled with %s %d.",
4431 : "CATALOG_VERSION_NO", ControlFile->catalog_version_no,
4432 : "CATALOG_VERSION_NO", CATALOG_VERSION_NO),
4433 : errhint("It looks like you need to initdb.")));
4434 1990 : if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
4435 0 : ereport(FATAL,
4436 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4437 : errmsg("database files are incompatible with server"),
4438 : /* translator: %s is a variable name and %d is its value */
4439 : errdetail("The database cluster was initialized with %s %d,"
4440 : " but the server was compiled with %s %d.",
4441 : "MAXALIGN", ControlFile->maxAlign,
4442 : "MAXALIGN", MAXIMUM_ALIGNOF),
4443 : errhint("It looks like you need to initdb.")));
4444 1990 : if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
4445 0 : ereport(FATAL,
4446 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4447 : errmsg("database files are incompatible with server"),
4448 : errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4449 : errhint("It looks like you need to initdb.")));
4450 1990 : if (ControlFile->blcksz != BLCKSZ)
4451 0 : ereport(FATAL,
4452 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4453 : errmsg("database files are incompatible with server"),
4454 : /* translator: %s is a variable name and %d is its value */
4455 : errdetail("The database cluster was initialized with %s %d,"
4456 : " but the server was compiled with %s %d.",
4457 : "BLCKSZ", ControlFile->blcksz,
4458 : "BLCKSZ", BLCKSZ),
4459 : errhint("It looks like you need to recompile or initdb.")));
4460 1990 : if (ControlFile->relseg_size != RELSEG_SIZE)
4461 0 : ereport(FATAL,
4462 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4463 : errmsg("database files are incompatible with server"),
4464 : /* translator: %s is a variable name and %d is its value */
4465 : errdetail("The database cluster was initialized with %s %d,"
4466 : " but the server was compiled with %s %d.",
4467 : "RELSEG_SIZE", ControlFile->relseg_size,
4468 : "RELSEG_SIZE", RELSEG_SIZE),
4469 : errhint("It looks like you need to recompile or initdb.")));
4470 1990 : if (ControlFile->xlog_blcksz != XLOG_BLCKSZ)
4471 0 : ereport(FATAL,
4472 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4473 : errmsg("database files are incompatible with server"),
4474 : /* translator: %s is a variable name and %d is its value */
4475 : errdetail("The database cluster was initialized with %s %d,"
4476 : " but the server was compiled with %s %d.",
4477 : "XLOG_BLCKSZ", ControlFile->xlog_blcksz,
4478 : "XLOG_BLCKSZ", XLOG_BLCKSZ),
4479 : errhint("It looks like you need to recompile or initdb.")));
4480 1990 : if (ControlFile->nameDataLen != NAMEDATALEN)
4481 0 : ereport(FATAL,
4482 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4483 : errmsg("database files are incompatible with server"),
4484 : /* translator: %s is a variable name and %d is its value */
4485 : errdetail("The database cluster was initialized with %s %d,"
4486 : " but the server was compiled with %s %d.",
4487 : "NAMEDATALEN", ControlFile->nameDataLen,
4488 : "NAMEDATALEN", NAMEDATALEN),
4489 : errhint("It looks like you need to recompile or initdb.")));
4490 1990 : if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
4491 0 : ereport(FATAL,
4492 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4493 : errmsg("database files are incompatible with server"),
4494 : /* translator: %s is a variable name and %d is its value */
4495 : errdetail("The database cluster was initialized with %s %d,"
4496 : " but the server was compiled with %s %d.",
4497 : "INDEX_MAX_KEYS", ControlFile->indexMaxKeys,
4498 : "INDEX_MAX_KEYS", INDEX_MAX_KEYS),
4499 : errhint("It looks like you need to recompile or initdb.")));
4500 1990 : if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
4501 0 : ereport(FATAL,
4502 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4503 : errmsg("database files are incompatible with server"),
4504 : /* translator: %s is a variable name and %d is its value */
4505 : errdetail("The database cluster was initialized with %s %d,"
4506 : " but the server was compiled with %s %d.",
4507 : "TOAST_MAX_CHUNK_SIZE", ControlFile->toast_max_chunk_size,
4508 : "TOAST_MAX_CHUNK_SIZE", (int) TOAST_MAX_CHUNK_SIZE),
4509 : errhint("It looks like you need to recompile or initdb.")));
4510 1990 : if (ControlFile->loblksize != LOBLKSIZE)
4511 0 : ereport(FATAL,
4512 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4513 : errmsg("database files are incompatible with server"),
4514 : /* translator: %s is a variable name and %d is its value */
4515 : errdetail("The database cluster was initialized with %s %d,"
4516 : " but the server was compiled with %s %d.",
4517 : "LOBLKSIZE", ControlFile->loblksize,
4518 : "LOBLKSIZE", (int) LOBLKSIZE),
4519 : errhint("It looks like you need to recompile or initdb.")));
4520 :
4521 : Assert(ControlFile->float8ByVal); /* vestigial, not worth an error msg */
4522 :
4523 1990 : wal_segment_size = ControlFile->xlog_seg_size;
4524 :
4525 1990 : if (!IsValidWalSegSize(wal_segment_size))
4526 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4527 : errmsg_plural("invalid WAL segment size in control file (%d byte)",
4528 : "invalid WAL segment size in control file (%d bytes)",
4529 : wal_segment_size,
4530 : wal_segment_size),
4531 : errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.")));
4532 :
4533 1990 : snprintf(wal_segsz_str, sizeof(wal_segsz_str), "%d", wal_segment_size);
4534 1990 : SetConfigOption("wal_segment_size", wal_segsz_str, PGC_INTERNAL,
4535 : PGC_S_DYNAMIC_DEFAULT);
4536 :
4537 : /* check and update variables dependent on wal_segment_size */
4538 1990 : if (ConvertToXSegs(min_wal_size_mb, wal_segment_size) < 2)
4539 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4540 : /* translator: both %s are GUC names */
4541 : errmsg("\"%s\" must be at least twice \"%s\"",
4542 : "min_wal_size", "wal_segment_size")));
4543 :
4544 1990 : if (ConvertToXSegs(max_wal_size_mb, wal_segment_size) < 2)
4545 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4546 : /* translator: both %s are GUC names */
4547 : errmsg("\"%s\" must be at least twice \"%s\"",
4548 : "max_wal_size", "wal_segment_size")));
4549 :
4550 1990 : UsableBytesInSegment =
4551 1990 : (wal_segment_size / XLOG_BLCKSZ * UsableBytesInPage) -
4552 : (SizeOfXLogLongPHD - SizeOfXLogShortPHD);
4553 :
4554 1990 : CalculateCheckpointSegments();
4555 :
4556 : /* Make the initdb settings visible as GUC variables, too */
4557 1990 : SetConfigOption("data_checksums", DataChecksumsEnabled() ? "yes" : "no",
4558 : PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
4559 1990 : }
4560 :
4561 : /*
4562 : * Utility wrapper to update the control file. Note that the control
4563 : * file gets flushed.
4564 : */
4565 : static void
4566 18084 : UpdateControlFile(void)
4567 : {
4568 18084 : update_controlfile(DataDir, ControlFile, true);
4569 18084 : }
4570 :
4571 : /*
4572 : * Returns the unique system identifier from control file.
4573 : */
4574 : uint64
4575 2794 : GetSystemIdentifier(void)
4576 : {
4577 : Assert(ControlFile != NULL);
4578 2794 : return ControlFile->system_identifier;
4579 : }
4580 :
4581 : /*
4582 : * Returns the random nonce from control file.
4583 : */
4584 : char *
4585 2 : GetMockAuthenticationNonce(void)
4586 : {
4587 : Assert(ControlFile != NULL);
4588 2 : return ControlFile->mock_authentication_nonce;
4589 : }
4590 :
4591 : /*
4592 : * Are checksums enabled for data pages?
4593 : */
4594 : bool
4595 20141340 : DataChecksumsEnabled(void)
4596 : {
4597 : Assert(ControlFile != NULL);
4598 20141340 : return (ControlFile->data_checksum_version > 0);
4599 : }
4600 :
4601 : /*
4602 : * Return true if the cluster was initialized on a platform where the
4603 : * default signedness of char is "signed". This function exists for code
4604 : * that deals with pre-v18 data files that store data sorted by the 'char'
4605 : * type on disk (e.g., GIN and GiST indexes). See the comments in
4606 : * WriteControlFile() for details.
4607 : */
4608 : bool
4609 6 : GetDefaultCharSignedness(void)
4610 : {
4611 6 : return ControlFile->default_char_signedness;
4612 : }
4613 :
4614 : /*
4615 : * Returns a fake LSN for unlogged relations.
4616 : *
4617 : * Each call generates an LSN that is greater than any previous value
4618 : * returned. The current counter value is saved and restored across clean
4619 : * shutdowns, but like unlogged relations, does not survive a crash. This can
4620 : * be used in lieu of real LSN values returned by XLogInsert, if you need an
4621 : * LSN-like increasing sequence of numbers without writing any WAL.
4622 : */
4623 : XLogRecPtr
4624 66 : GetFakeLSNForUnloggedRel(void)
4625 : {
4626 66 : return pg_atomic_fetch_add_u64(&XLogCtl->unloggedLSN, 1);
4627 : }
4628 :
4629 : /*
4630 : * Auto-tune the number of XLOG buffers.
4631 : *
4632 : * The preferred setting for wal_buffers is about 3% of shared_buffers, with
4633 : * a maximum of one XLOG segment (there is little reason to think that more
4634 : * is helpful, at least so long as we force an fsync when switching log files)
4635 : * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
4636 : * 9.1, when auto-tuning was added).
4637 : *
4638 : * This should not be called until NBuffers has received its final value.
4639 : */
4640 : static int
4641 2170 : XLOGChooseNumBuffers(void)
4642 : {
4643 : int xbuffers;
4644 :
4645 2170 : xbuffers = NBuffers / 32;
4646 2170 : if (xbuffers > (wal_segment_size / XLOG_BLCKSZ))
4647 48 : xbuffers = (wal_segment_size / XLOG_BLCKSZ);
4648 2170 : if (xbuffers < 8)
4649 844 : xbuffers = 8;
4650 2170 : return xbuffers;
4651 : }
4652 :
4653 : /*
4654 : * GUC check_hook for wal_buffers
4655 : */
4656 : bool
4657 4420 : check_wal_buffers(int *newval, void **extra, GucSource source)
4658 : {
4659 : /*
4660 : * -1 indicates a request for auto-tune.
4661 : */
4662 4420 : if (*newval == -1)
4663 : {
4664 : /*
4665 : * If we haven't yet changed the boot_val default of -1, just let it
4666 : * be. We'll fix it when XLOGShmemSize is called.
4667 : */
4668 2250 : if (XLOGbuffers == -1)
4669 2250 : return true;
4670 :
4671 : /* Otherwise, substitute the auto-tune value */
4672 0 : *newval = XLOGChooseNumBuffers();
4673 : }
4674 :
4675 : /*
4676 : * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
4677 : * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
4678 : * the case, we just silently treat such values as a request for the
4679 : * minimum. (We could throw an error instead, but that doesn't seem very
4680 : * helpful.)
4681 : */
4682 2170 : if (*newval < 4)
4683 0 : *newval = 4;
4684 :
4685 2170 : return true;
4686 : }
4687 :
4688 : /*
4689 : * GUC check_hook for wal_consistency_checking
4690 : */
4691 : bool
4692 4042 : check_wal_consistency_checking(char **newval, void **extra, GucSource source)
4693 : {
4694 : char *rawstring;
4695 : List *elemlist;
4696 : ListCell *l;
4697 : bool newwalconsistency[RM_MAX_ID + 1];
4698 :
4699 : /* Initialize the array */
4700 133386 : MemSet(newwalconsistency, 0, (RM_MAX_ID + 1) * sizeof(bool));
4701 :
4702 : /* Need a modifiable copy of string */
4703 4042 : rawstring = pstrdup(*newval);
4704 :
4705 : /* Parse string into list of identifiers */
4706 4042 : if (!SplitIdentifierString(rawstring, ',', &elemlist))
4707 : {
4708 : /* syntax error in list */
4709 0 : GUC_check_errdetail("List syntax is invalid.");
4710 0 : pfree(rawstring);
4711 0 : list_free(elemlist);
4712 0 : return false;
4713 : }
4714 :
4715 4944 : foreach(l, elemlist)
4716 : {
4717 902 : char *tok = (char *) lfirst(l);
4718 : int rmid;
4719 :
4720 : /* Check for 'all'. */
4721 902 : if (pg_strcasecmp(tok, "all") == 0)
4722 : {
4723 230786 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4724 229888 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL)
4725 8980 : newwalconsistency[rmid] = true;
4726 : }
4727 : else
4728 : {
4729 : /* Check if the token matches any known resource manager. */
4730 4 : bool found = false;
4731 :
4732 72 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4733 : {
4734 108 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL &&
4735 36 : pg_strcasecmp(tok, GetRmgr(rmid).rm_name) == 0)
4736 : {
4737 4 : newwalconsistency[rmid] = true;
4738 4 : found = true;
4739 4 : break;
4740 : }
4741 : }
4742 4 : if (!found)
4743 : {
4744 : /*
4745 : * During startup, it might be a not-yet-loaded custom
4746 : * resource manager. Defer checking until
4747 : * InitializeWalConsistencyChecking().
4748 : */
4749 0 : if (!process_shared_preload_libraries_done)
4750 : {
4751 0 : check_wal_consistency_checking_deferred = true;
4752 : }
4753 : else
4754 : {
4755 0 : GUC_check_errdetail("Unrecognized key word: \"%s\".", tok);
4756 0 : pfree(rawstring);
4757 0 : list_free(elemlist);
4758 0 : return false;
4759 : }
4760 : }
4761 : }
4762 : }
4763 :
4764 4042 : pfree(rawstring);
4765 4042 : list_free(elemlist);
4766 :
4767 : /* assign new value */
4768 4042 : *extra = guc_malloc(LOG, (RM_MAX_ID + 1) * sizeof(bool));
4769 4042 : if (!*extra)
4770 0 : return false;
4771 4042 : memcpy(*extra, newwalconsistency, (RM_MAX_ID + 1) * sizeof(bool));
4772 4042 : return true;
4773 : }
4774 :
4775 : /*
4776 : * GUC assign_hook for wal_consistency_checking
4777 : */
4778 : void
4779 4040 : assign_wal_consistency_checking(const char *newval, void *extra)
4780 : {
4781 : /*
4782 : * If some checks were deferred, it's possible that the checks will fail
4783 : * later during InitializeWalConsistencyChecking(). But in that case, the
4784 : * postmaster will exit anyway, so it's safe to proceed with the
4785 : * assignment.
4786 : *
4787 : * Any built-in resource managers specified are assigned immediately,
4788 : * which affects WAL created before shared_preload_libraries are
4789 : * processed. Any custom resource managers specified won't be assigned
4790 : * until after shared_preload_libraries are processed, but that's OK
4791 : * because WAL for a custom resource manager can't be written before the
4792 : * module is loaded anyway.
4793 : */
4794 4040 : wal_consistency_checking = extra;
4795 4040 : }
4796 :
4797 : /*
4798 : * InitializeWalConsistencyChecking: run after loading custom resource managers
4799 : *
4800 : * If any unknown resource managers were specified in the
4801 : * wal_consistency_checking GUC, processing was deferred. Now that
4802 : * shared_preload_libraries have been loaded, process wal_consistency_checking
4803 : * again.
4804 : */
4805 : void
4806 1870 : InitializeWalConsistencyChecking(void)
4807 : {
4808 : Assert(process_shared_preload_libraries_done);
4809 :
4810 1870 : if (check_wal_consistency_checking_deferred)
4811 : {
4812 : struct config_generic *guc;
4813 :
4814 0 : guc = find_option("wal_consistency_checking", false, false, ERROR);
4815 :
4816 0 : check_wal_consistency_checking_deferred = false;
4817 :
4818 0 : set_config_option_ext("wal_consistency_checking",
4819 : wal_consistency_checking_string,
4820 : guc->scontext, guc->source, guc->srole,
4821 : GUC_ACTION_SET, true, ERROR, false);
4822 :
4823 : /* checking should not be deferred again */
4824 : Assert(!check_wal_consistency_checking_deferred);
4825 : }
4826 1870 : }
4827 :
4828 : /*
4829 : * GUC show_hook for archive_command
4830 : */
4831 : const char *
4832 3448 : show_archive_command(void)
4833 : {
4834 3448 : if (XLogArchivingActive())
4835 4 : return XLogArchiveCommand;
4836 : else
4837 3444 : return "(disabled)";
4838 : }
4839 :
4840 : /*
4841 : * GUC show_hook for in_hot_standby
4842 : */
4843 : const char *
4844 30006 : show_in_hot_standby(void)
4845 : {
4846 : /*
4847 : * We display the actual state based on shared memory, so that this GUC
4848 : * reports up-to-date state if examined intra-query. The underlying
4849 : * variable (in_hot_standby_guc) changes only when we transmit a new value
4850 : * to the client.
4851 : */
4852 30006 : return RecoveryInProgress() ? "on" : "off";
4853 : }
4854 :
4855 : /*
4856 : * Read the control file, set respective GUCs.
4857 : *
4858 : * This is to be called during startup, including a crash recovery cycle,
4859 : * unless in bootstrap mode, where no control file yet exists. As there's no
4860 : * usable shared memory yet (its sizing can depend on the contents of the
4861 : * control file!), first store the contents in local memory. XLOGShmemInit()
4862 : * will then copy it to shared memory later.
4863 : *
4864 : * reset just controls whether previous contents are to be expected (in the
4865 : * reset case, there's a dangling pointer into old shared memory), or not.
4866 : */
4867 : void
4868 1890 : LocalProcessControlFile(bool reset)
4869 : {
4870 : Assert(reset || ControlFile == NULL);
4871 1890 : ControlFile = palloc(sizeof(ControlFileData));
4872 1890 : ReadControlFile();
4873 1890 : }
4874 :
4875 : /*
4876 : * Get the wal_level from the control file. For a standby, this value should be
4877 : * considered as its active wal_level, because it may be different from what
4878 : * was originally configured on standby.
4879 : */
4880 : WalLevel
4881 138 : GetActiveWalLevelOnStandby(void)
4882 : {
4883 138 : return ControlFile->wal_level;
4884 : }
4885 :
4886 : /*
4887 : * Initialization of shared memory for XLOG
4888 : */
4889 : Size
4890 6222 : XLOGShmemSize(void)
4891 : {
4892 : Size size;
4893 :
4894 : /*
4895 : * If the value of wal_buffers is -1, use the preferred auto-tune value.
4896 : * This isn't an amazingly clean place to do this, but we must wait till
4897 : * NBuffers has received its final value, and must do it before using the
4898 : * value of XLOGbuffers to do anything important.
4899 : *
4900 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
4901 : * However, if the DBA explicitly set wal_buffers = -1 in the config file,
4902 : * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
4903 : * the matter with PGC_S_OVERRIDE.
4904 : */
4905 6222 : if (XLOGbuffers == -1)
4906 : {
4907 : char buf[32];
4908 :
4909 2170 : snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
4910 2170 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4911 : PGC_S_DYNAMIC_DEFAULT);
4912 2170 : if (XLOGbuffers == -1) /* failed to apply it? */
4913 0 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4914 : PGC_S_OVERRIDE);
4915 : }
4916 : Assert(XLOGbuffers > 0);
4917 :
4918 : /* XLogCtl */
4919 6222 : size = sizeof(XLogCtlData);
4920 :
4921 : /* WAL insertion locks, plus alignment */
4922 6222 : size = add_size(size, mul_size(sizeof(WALInsertLockPadded), NUM_XLOGINSERT_LOCKS + 1));
4923 : /* xlblocks array */
4924 6222 : size = add_size(size, mul_size(sizeof(pg_atomic_uint64), XLOGbuffers));
4925 : /* extra alignment padding for XLOG I/O buffers */
4926 6222 : size = add_size(size, Max(XLOG_BLCKSZ, PG_IO_ALIGN_SIZE));
4927 : /* and the buffers themselves */
4928 6222 : size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
4929 :
4930 : /*
4931 : * Note: we don't count ControlFileData, it comes out of the "slop factor"
4932 : * added by CreateSharedMemoryAndSemaphores. This lets us use this
4933 : * routine again below to compute the actual allocation size.
4934 : */
4935 :
4936 6222 : return size;
4937 : }
4938 :
4939 : void
4940 2174 : XLOGShmemInit(void)
4941 : {
4942 : bool foundCFile,
4943 : foundXLog;
4944 : char *allocptr;
4945 : int i;
4946 : ControlFileData *localControlFile;
4947 :
4948 : #ifdef WAL_DEBUG
4949 :
4950 : /*
4951 : * Create a memory context for WAL debugging that's exempt from the normal
4952 : * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
4953 : * an allocation fails, but wal_debug is not for production use anyway.
4954 : */
4955 : if (walDebugCxt == NULL)
4956 : {
4957 : walDebugCxt = AllocSetContextCreate(TopMemoryContext,
4958 : "WAL Debug",
4959 : ALLOCSET_DEFAULT_SIZES);
4960 : MemoryContextAllowInCriticalSection(walDebugCxt, true);
4961 : }
4962 : #endif
4963 :
4964 :
4965 2174 : XLogCtl = (XLogCtlData *)
4966 2174 : ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog);
4967 :
4968 2174 : localControlFile = ControlFile;
4969 2174 : ControlFile = (ControlFileData *)
4970 2174 : ShmemInitStruct("Control File", sizeof(ControlFileData), &foundCFile);
4971 :
4972 2174 : if (foundCFile || foundXLog)
4973 : {
4974 : /* both should be present or neither */
4975 : Assert(foundCFile && foundXLog);
4976 :
4977 : /* Initialize local copy of WALInsertLocks */
4978 0 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks;
4979 :
4980 0 : if (localControlFile)
4981 0 : pfree(localControlFile);
4982 0 : return;
4983 : }
4984 2174 : memset(XLogCtl, 0, sizeof(XLogCtlData));
4985 :
4986 : /*
4987 : * Already have read control file locally, unless in bootstrap mode. Move
4988 : * contents into shared memory.
4989 : */
4990 2174 : if (localControlFile)
4991 : {
4992 1874 : memcpy(ControlFile, localControlFile, sizeof(ControlFileData));
4993 1874 : pfree(localControlFile);
4994 : }
4995 :
4996 : /*
4997 : * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
4998 : * multiple of the alignment for same, so no extra alignment padding is
4999 : * needed here.
5000 : */
5001 2174 : allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
5002 2174 : XLogCtl->xlblocks = (pg_atomic_uint64 *) allocptr;
5003 2174 : allocptr += sizeof(pg_atomic_uint64) * XLOGbuffers;
5004 :
5005 620330 : for (i = 0; i < XLOGbuffers; i++)
5006 : {
5007 618156 : pg_atomic_init_u64(&XLogCtl->xlblocks[i], InvalidXLogRecPtr);
5008 : }
5009 :
5010 : /* WAL insertion locks. Ensure they're aligned to the full padded size */
5011 2174 : allocptr += sizeof(WALInsertLockPadded) -
5012 2174 : ((uintptr_t) allocptr) % sizeof(WALInsertLockPadded);
5013 2174 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks =
5014 : (WALInsertLockPadded *) allocptr;
5015 2174 : allocptr += sizeof(WALInsertLockPadded) * NUM_XLOGINSERT_LOCKS;
5016 :
5017 19566 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
5018 : {
5019 17392 : LWLockInitialize(&WALInsertLocks[i].l.lock, LWTRANCHE_WAL_INSERT);
5020 17392 : pg_atomic_init_u64(&WALInsertLocks[i].l.insertingAt, InvalidXLogRecPtr);
5021 17392 : WALInsertLocks[i].l.lastImportantAt = InvalidXLogRecPtr;
5022 : }
5023 :
5024 : /*
5025 : * Align the start of the page buffers to a full xlog block size boundary.
5026 : * This simplifies some calculations in XLOG insertion. It is also
5027 : * required for O_DIRECT.
5028 : */
5029 2174 : allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5030 2174 : XLogCtl->pages = allocptr;
5031 2174 : memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5032 :
5033 : /*
5034 : * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5035 : * in additional info.)
5036 : */
5037 2174 : XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5038 2174 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5039 2174 : XLogCtl->InstallXLogFileSegmentActive = false;
5040 2174 : XLogCtl->WalWriterSleeping = false;
5041 :
5042 2174 : SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5043 2174 : SpinLockInit(&XLogCtl->info_lck);
5044 2174 : pg_atomic_init_u64(&XLogCtl->logInsertResult, InvalidXLogRecPtr);
5045 2174 : pg_atomic_init_u64(&XLogCtl->logWriteResult, InvalidXLogRecPtr);
5046 2174 : pg_atomic_init_u64(&XLogCtl->logFlushResult, InvalidXLogRecPtr);
5047 2174 : pg_atomic_init_u64(&XLogCtl->unloggedLSN, InvalidXLogRecPtr);
5048 : }
5049 :
5050 : /*
5051 : * This func must be called ONCE on system install. It creates pg_control
5052 : * and the initial XLOG segment.
5053 : */
5054 : void
5055 100 : BootStrapXLOG(uint32 data_checksum_version)
5056 : {
5057 : CheckPoint checkPoint;
5058 : char *buffer;
5059 : XLogPageHeader page;
5060 : XLogLongPageHeader longpage;
5061 : XLogRecord *record;
5062 : char *recptr;
5063 : uint64 sysidentifier;
5064 : struct timeval tv;
5065 : pg_crc32c crc;
5066 :
5067 : /* allow ordinary WAL segment creation, like StartupXLOG() would */
5068 100 : SetInstallXLogFileSegmentActive();
5069 :
5070 : /*
5071 : * Select a hopefully-unique system identifier code for this installation.
5072 : * We use the result of gettimeofday(), including the fractional seconds
5073 : * field, as being about as unique as we can easily get. (Think not to
5074 : * use random(), since it hasn't been seeded and there's no portable way
5075 : * to seed it other than the system clock value...) The upper half of the
5076 : * uint64 value is just the tv_sec part, while the lower half contains the
5077 : * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
5078 : * PID for a little extra uniqueness. A person knowing this encoding can
5079 : * determine the initialization time of the installation, which could
5080 : * perhaps be useful sometimes.
5081 : */
5082 100 : gettimeofday(&tv, NULL);
5083 100 : sysidentifier = ((uint64) tv.tv_sec) << 32;
5084 100 : sysidentifier |= ((uint64) tv.tv_usec) << 12;
5085 100 : sysidentifier |= getpid() & 0xFFF;
5086 :
5087 : /* page buffer must be aligned suitably for O_DIRECT */
5088 100 : buffer = (char *) palloc(XLOG_BLCKSZ + XLOG_BLCKSZ);
5089 100 : page = (XLogPageHeader) TYPEALIGN(XLOG_BLCKSZ, buffer);
5090 100 : memset(page, 0, XLOG_BLCKSZ);
5091 :
5092 : /*
5093 : * Set up information for the initial checkpoint record
5094 : *
5095 : * The initial checkpoint record is written to the beginning of the WAL
5096 : * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5097 : * used, so that we can use 0/0 to mean "before any valid WAL segment".
5098 : */
5099 100 : checkPoint.redo = wal_segment_size + SizeOfXLogLongPHD;
5100 100 : checkPoint.ThisTimeLineID = BootstrapTimeLineID;
5101 100 : checkPoint.PrevTimeLineID = BootstrapTimeLineID;
5102 100 : checkPoint.fullPageWrites = fullPageWrites;
5103 100 : checkPoint.wal_level = wal_level;
5104 : checkPoint.nextXid =
5105 100 : FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId);
5106 100 : checkPoint.nextOid = FirstGenbkiObjectId;
5107 100 : checkPoint.nextMulti = FirstMultiXactId;
5108 100 : checkPoint.nextMultiOffset = 0;
5109 100 : checkPoint.oldestXid = FirstNormalTransactionId;
5110 100 : checkPoint.oldestXidDB = Template1DbOid;
5111 100 : checkPoint.oldestMulti = FirstMultiXactId;
5112 100 : checkPoint.oldestMultiDB = Template1DbOid;
5113 100 : checkPoint.oldestCommitTsXid = InvalidTransactionId;
5114 100 : checkPoint.newestCommitTsXid = InvalidTransactionId;
5115 100 : checkPoint.time = (pg_time_t) time(NULL);
5116 100 : checkPoint.oldestActiveXid = InvalidTransactionId;
5117 :
5118 100 : TransamVariables->nextXid = checkPoint.nextXid;
5119 100 : TransamVariables->nextOid = checkPoint.nextOid;
5120 100 : TransamVariables->oidCount = 0;
5121 100 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5122 100 : AdvanceOldestClogXid(checkPoint.oldestXid);
5123 100 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5124 100 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5125 100 : SetCommitTsLimit(InvalidTransactionId, InvalidTransactionId);
5126 :
5127 : /* Set up the XLOG page header */
5128 100 : page->xlp_magic = XLOG_PAGE_MAGIC;
5129 100 : page->xlp_info = XLP_LONG_HEADER;
5130 100 : page->xlp_tli = BootstrapTimeLineID;
5131 100 : page->xlp_pageaddr = wal_segment_size;
5132 100 : longpage = (XLogLongPageHeader) page;
5133 100 : longpage->xlp_sysid = sysidentifier;
5134 100 : longpage->xlp_seg_size = wal_segment_size;
5135 100 : longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5136 :
5137 : /* Insert the initial checkpoint record */
5138 100 : recptr = ((char *) page + SizeOfXLogLongPHD);
5139 100 : record = (XLogRecord *) recptr;
5140 100 : record->xl_prev = 0;
5141 100 : record->xl_xid = InvalidTransactionId;
5142 100 : record->xl_tot_len = SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(checkPoint);
5143 100 : record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5144 100 : record->xl_rmid = RM_XLOG_ID;
5145 100 : recptr += SizeOfXLogRecord;
5146 : /* fill the XLogRecordDataHeaderShort struct */
5147 100 : *(recptr++) = (char) XLR_BLOCK_ID_DATA_SHORT;
5148 100 : *(recptr++) = sizeof(checkPoint);
5149 100 : memcpy(recptr, &checkPoint, sizeof(checkPoint));
5150 100 : recptr += sizeof(checkPoint);
5151 : Assert(recptr - (char *) record == record->xl_tot_len);
5152 :
5153 100 : INIT_CRC32C(crc);
5154 100 : COMP_CRC32C(crc, ((char *) record) + SizeOfXLogRecord, record->xl_tot_len - SizeOfXLogRecord);
5155 100 : COMP_CRC32C(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5156 100 : FIN_CRC32C(crc);
5157 100 : record->xl_crc = crc;
5158 :
5159 : /* Create first XLOG segment file */
5160 100 : openLogTLI = BootstrapTimeLineID;
5161 100 : openLogFile = XLogFileInit(1, BootstrapTimeLineID);
5162 :
5163 : /*
5164 : * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
5165 : * close the file again in a moment.
5166 : */
5167 :
5168 : /* Write the first page with the initial record */
5169 100 : errno = 0;
5170 100 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE);
5171 100 : if (write(openLogFile, page, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5172 : {
5173 : /* if write didn't set errno, assume problem is no disk space */
5174 0 : if (errno == 0)
5175 0 : errno = ENOSPC;
5176 0 : ereport(PANIC,
5177 : (errcode_for_file_access(),
5178 : errmsg("could not write bootstrap write-ahead log file: %m")));
5179 : }
5180 100 : pgstat_report_wait_end();
5181 :
5182 100 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC);
5183 100 : if (pg_fsync(openLogFile) != 0)
5184 0 : ereport(PANIC,
5185 : (errcode_for_file_access(),
5186 : errmsg("could not fsync bootstrap write-ahead log file: %m")));
5187 100 : pgstat_report_wait_end();
5188 :
5189 100 : if (close(openLogFile) != 0)
5190 0 : ereport(PANIC,
5191 : (errcode_for_file_access(),
5192 : errmsg("could not close bootstrap write-ahead log file: %m")));
5193 :
5194 100 : openLogFile = -1;
5195 :
5196 : /* Now create pg_control */
5197 100 : InitControlFile(sysidentifier, data_checksum_version);
5198 100 : ControlFile->time = checkPoint.time;
5199 100 : ControlFile->checkPoint = checkPoint.redo;
5200 100 : ControlFile->checkPointCopy = checkPoint;
5201 :
5202 : /* some additional ControlFile fields are set in WriteControlFile() */
5203 100 : WriteControlFile();
5204 :
5205 : /* Bootstrap the commit log, too */
5206 100 : BootStrapCLOG();
5207 100 : BootStrapCommitTs();
5208 100 : BootStrapSUBTRANS();
5209 100 : BootStrapMultiXact();
5210 :
5211 100 : pfree(buffer);
5212 :
5213 : /*
5214 : * Force control file to be read - in contrast to normal processing we'd
5215 : * otherwise never run the checks and GUC related initializations therein.
5216 : */
5217 100 : ReadControlFile();
5218 100 : }
5219 :
5220 : static char *
5221 1670 : str_time(pg_time_t tnow, char *buf, size_t bufsize)
5222 : {
5223 1670 : pg_strftime(buf, bufsize,
5224 : "%Y-%m-%d %H:%M:%S %Z",
5225 1670 : pg_localtime(&tnow, log_timezone));
5226 :
5227 1670 : return buf;
5228 : }
5229 :
5230 : /*
5231 : * Initialize the first WAL segment on new timeline.
5232 : */
5233 : static void
5234 98 : XLogInitNewTimeline(TimeLineID endTLI, XLogRecPtr endOfLog, TimeLineID newTLI)
5235 : {
5236 : char xlogfname[MAXFNAMELEN];
5237 : XLogSegNo endLogSegNo;
5238 : XLogSegNo startLogSegNo;
5239 :
5240 : /* we always switch to a new timeline after archive recovery */
5241 : Assert(endTLI != newTLI);
5242 :
5243 : /*
5244 : * Update min recovery point one last time.
5245 : */
5246 98 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5247 :
5248 : /*
5249 : * Calculate the last segment on the old timeline, and the first segment
5250 : * on the new timeline. If the switch happens in the middle of a segment,
5251 : * they are the same, but if the switch happens exactly at a segment
5252 : * boundary, startLogSegNo will be endLogSegNo + 1.
5253 : */
5254 98 : XLByteToPrevSeg(endOfLog, endLogSegNo, wal_segment_size);
5255 98 : XLByteToSeg(endOfLog, startLogSegNo, wal_segment_size);
5256 :
5257 : /*
5258 : * Initialize the starting WAL segment for the new timeline. If the switch
5259 : * happens in the middle of a segment, copy data from the last WAL segment
5260 : * of the old timeline up to the switch point, to the starting WAL segment
5261 : * on the new timeline.
5262 : */
5263 98 : if (endLogSegNo == startLogSegNo)
5264 : {
5265 : /*
5266 : * Make a copy of the file on the new timeline.
5267 : *
5268 : * Writing WAL isn't allowed yet, so there are no locking
5269 : * considerations. But we should be just as tense as XLogFileInit to
5270 : * avoid emplacing a bogus file.
5271 : */
5272 78 : XLogFileCopy(newTLI, endLogSegNo, endTLI, endLogSegNo,
5273 78 : XLogSegmentOffset(endOfLog, wal_segment_size));
5274 : }
5275 : else
5276 : {
5277 : /*
5278 : * The switch happened at a segment boundary, so just create the next
5279 : * segment on the new timeline.
5280 : */
5281 : int fd;
5282 :
5283 20 : fd = XLogFileInit(startLogSegNo, newTLI);
5284 :
5285 20 : if (close(fd) != 0)
5286 : {
5287 0 : int save_errno = errno;
5288 :
5289 0 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5290 0 : errno = save_errno;
5291 0 : ereport(ERROR,
5292 : (errcode_for_file_access(),
5293 : errmsg("could not close file \"%s\": %m", xlogfname)));
5294 : }
5295 : }
5296 :
5297 : /*
5298 : * Let's just make real sure there are not .ready or .done flags posted
5299 : * for the new segment.
5300 : */
5301 98 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5302 98 : XLogArchiveCleanup(xlogfname);
5303 98 : }
5304 :
5305 : /*
5306 : * Perform cleanup actions at the conclusion of archive recovery.
5307 : */
5308 : static void
5309 98 : CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI, XLogRecPtr EndOfLog,
5310 : TimeLineID newTLI)
5311 : {
5312 : /*
5313 : * Execute the recovery_end_command, if any.
5314 : */
5315 98 : if (recoveryEndCommand && strcmp(recoveryEndCommand, "") != 0)
5316 4 : ExecuteRecoveryCommand(recoveryEndCommand,
5317 : "recovery_end_command",
5318 : true,
5319 : WAIT_EVENT_RECOVERY_END_COMMAND);
5320 :
5321 : /*
5322 : * We switched to a new timeline. Clean up segments on the old timeline.
5323 : *
5324 : * If there are any higher-numbered segments on the old timeline, remove
5325 : * them. They might contain valid WAL, but they might also be
5326 : * pre-allocated files containing garbage. In any case, they are not part
5327 : * of the new timeline's history so we don't need them.
5328 : */
5329 98 : RemoveNonParentXlogFiles(EndOfLog, newTLI);
5330 :
5331 : /*
5332 : * If the switch happened in the middle of a segment, what to do with the
5333 : * last, partial segment on the old timeline? If we don't archive it, and
5334 : * the server that created the WAL never archives it either (e.g. because
5335 : * it was hit by a meteor), it will never make it to the archive. That's
5336 : * OK from our point of view, because the new segment that we created with
5337 : * the new TLI contains all the WAL from the old timeline up to the switch
5338 : * point. But if you later try to do PITR to the "missing" WAL on the old
5339 : * timeline, recovery won't find it in the archive. It's physically
5340 : * present in the new file with new TLI, but recovery won't look there
5341 : * when it's recovering to the older timeline. On the other hand, if we
5342 : * archive the partial segment, and the original server on that timeline
5343 : * is still running and archives the completed version of the same segment
5344 : * later, it will fail. (We used to do that in 9.4 and below, and it
5345 : * caused such problems).
5346 : *
5347 : * As a compromise, we rename the last segment with the .partial suffix,
5348 : * and archive it. Archive recovery will never try to read .partial
5349 : * segments, so they will normally go unused. But in the odd PITR case,
5350 : * the administrator can copy them manually to the pg_wal directory
5351 : * (removing the suffix). They can be useful in debugging, too.
5352 : *
5353 : * If a .done or .ready file already exists for the old timeline, however,
5354 : * we had already determined that the segment is complete, so we can let
5355 : * it be archived normally. (In particular, if it was restored from the
5356 : * archive to begin with, it's expected to have a .done file).
5357 : */
5358 98 : if (XLogSegmentOffset(EndOfLog, wal_segment_size) != 0 &&
5359 : XLogArchivingActive())
5360 : {
5361 : char origfname[MAXFNAMELEN];
5362 : XLogSegNo endLogSegNo;
5363 :
5364 18 : XLByteToPrevSeg(EndOfLog, endLogSegNo, wal_segment_size);
5365 18 : XLogFileName(origfname, EndOfLogTLI, endLogSegNo, wal_segment_size);
5366 :
5367 18 : if (!XLogArchiveIsReadyOrDone(origfname))
5368 : {
5369 : char origpath[MAXPGPATH];
5370 : char partialfname[MAXFNAMELEN];
5371 : char partialpath[MAXPGPATH];
5372 :
5373 : /*
5374 : * If we're summarizing WAL, we can't rename the partial file
5375 : * until the summarizer finishes with it, else it will fail.
5376 : */
5377 10 : if (summarize_wal)
5378 2 : WaitForWalSummarization(EndOfLog);
5379 :
5380 10 : XLogFilePath(origpath, EndOfLogTLI, endLogSegNo, wal_segment_size);
5381 10 : snprintf(partialfname, MAXFNAMELEN, "%s.partial", origfname);
5382 10 : snprintf(partialpath, MAXPGPATH, "%s.partial", origpath);
5383 :
5384 : /*
5385 : * Make sure there's no .done or .ready file for the .partial
5386 : * file.
5387 : */
5388 10 : XLogArchiveCleanup(partialfname);
5389 :
5390 10 : durable_rename(origpath, partialpath, ERROR);
5391 10 : XLogArchiveNotify(partialfname);
5392 : }
5393 : }
5394 98 : }
5395 :
5396 : /*
5397 : * Check to see if required parameters are set high enough on this server
5398 : * for various aspects of recovery operation.
5399 : *
5400 : * Note that all the parameters which this function tests need to be
5401 : * listed in Administrator's Overview section in high-availability.sgml.
5402 : * If you change them, don't forget to update the list.
5403 : */
5404 : static void
5405 492 : CheckRequiredParameterValues(void)
5406 : {
5407 : /*
5408 : * For archive recovery, the WAL must be generated with at least 'replica'
5409 : * wal_level.
5410 : */
5411 492 : if (ArchiveRecoveryRequested && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5412 : {
5413 4 : ereport(FATAL,
5414 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5415 : errmsg("WAL was generated with \"wal_level=minimal\", cannot continue recovering"),
5416 : errdetail("This happens if you temporarily set \"wal_level=minimal\" on the server."),
5417 : errhint("Use a backup taken after setting \"wal_level\" to higher than \"minimal\".")));
5418 : }
5419 :
5420 : /*
5421 : * For Hot Standby, the WAL must be generated with 'replica' mode, and we
5422 : * must have at least as many backend slots as the primary.
5423 : */
5424 488 : if (ArchiveRecoveryRequested && EnableHotStandby)
5425 : {
5426 : /* We ignore autovacuum_worker_slots when we make this test. */
5427 246 : RecoveryRequiresIntParameter("max_connections",
5428 : MaxConnections,
5429 246 : ControlFile->MaxConnections);
5430 246 : RecoveryRequiresIntParameter("max_worker_processes",
5431 : max_worker_processes,
5432 246 : ControlFile->max_worker_processes);
5433 246 : RecoveryRequiresIntParameter("max_wal_senders",
5434 : max_wal_senders,
5435 246 : ControlFile->max_wal_senders);
5436 246 : RecoveryRequiresIntParameter("max_prepared_transactions",
5437 : max_prepared_xacts,
5438 246 : ControlFile->max_prepared_xacts);
5439 246 : RecoveryRequiresIntParameter("max_locks_per_transaction",
5440 : max_locks_per_xact,
5441 246 : ControlFile->max_locks_per_xact);
5442 : }
5443 488 : }
5444 :
5445 : /*
5446 : * This must be called ONCE during postmaster or standalone-backend startup
5447 : */
5448 : void
5449 1890 : StartupXLOG(void)
5450 : {
5451 : XLogCtlInsert *Insert;
5452 : CheckPoint checkPoint;
5453 : bool wasShutdown;
5454 : bool didCrash;
5455 : bool haveTblspcMap;
5456 : bool haveBackupLabel;
5457 : XLogRecPtr EndOfLog;
5458 : TimeLineID EndOfLogTLI;
5459 : TimeLineID newTLI;
5460 : bool performedWalRecovery;
5461 : EndOfWalRecoveryInfo *endOfRecoveryInfo;
5462 : XLogRecPtr abortedRecPtr;
5463 : XLogRecPtr missingContrecPtr;
5464 : TransactionId oldestActiveXID;
5465 1890 : bool promoted = false;
5466 : char timebuf[128];
5467 :
5468 : /*
5469 : * We should have an aux process resource owner to use, and we should not
5470 : * be in a transaction that's installed some other resowner.
5471 : */
5472 : Assert(AuxProcessResourceOwner != NULL);
5473 : Assert(CurrentResourceOwner == NULL ||
5474 : CurrentResourceOwner == AuxProcessResourceOwner);
5475 1890 : CurrentResourceOwner = AuxProcessResourceOwner;
5476 :
5477 : /*
5478 : * Check that contents look valid.
5479 : */
5480 1890 : if (!XRecOffIsValid(ControlFile->checkPoint))
5481 0 : ereport(FATAL,
5482 : (errcode(ERRCODE_DATA_CORRUPTED),
5483 : errmsg("control file contains invalid checkpoint location")));
5484 :
5485 1890 : switch (ControlFile->state)
5486 : {
5487 1476 : case DB_SHUTDOWNED:
5488 :
5489 : /*
5490 : * This is the expected case, so don't be chatty in standalone
5491 : * mode
5492 : */
5493 1476 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
5494 : (errmsg("database system was shut down at %s",
5495 : str_time(ControlFile->time,
5496 : timebuf, sizeof(timebuf)))));
5497 1476 : break;
5498 :
5499 64 : case DB_SHUTDOWNED_IN_RECOVERY:
5500 64 : ereport(LOG,
5501 : (errmsg("database system was shut down in recovery at %s",
5502 : str_time(ControlFile->time,
5503 : timebuf, sizeof(timebuf)))));
5504 64 : break;
5505 :
5506 0 : case DB_SHUTDOWNING:
5507 0 : ereport(LOG,
5508 : (errmsg("database system shutdown was interrupted; last known up at %s",
5509 : str_time(ControlFile->time,
5510 : timebuf, sizeof(timebuf)))));
5511 0 : break;
5512 :
5513 0 : case DB_IN_CRASH_RECOVERY:
5514 0 : ereport(LOG,
5515 : (errmsg("database system was interrupted while in recovery at %s",
5516 : str_time(ControlFile->time,
5517 : timebuf, sizeof(timebuf))),
5518 : errhint("This probably means that some data is corrupted and"
5519 : " you will have to use the last backup for recovery.")));
5520 0 : break;
5521 :
5522 12 : case DB_IN_ARCHIVE_RECOVERY:
5523 12 : ereport(LOG,
5524 : (errmsg("database system was interrupted while in recovery at log time %s",
5525 : str_time(ControlFile->checkPointCopy.time,
5526 : timebuf, sizeof(timebuf))),
5527 : errhint("If this has occurred more than once some data might be corrupted"
5528 : " and you might need to choose an earlier recovery target.")));
5529 12 : break;
5530 :
5531 338 : case DB_IN_PRODUCTION:
5532 338 : ereport(LOG,
5533 : (errmsg("database system was interrupted; last known up at %s",
5534 : str_time(ControlFile->time,
5535 : timebuf, sizeof(timebuf)))));
5536 338 : break;
5537 :
5538 0 : default:
5539 0 : ereport(FATAL,
5540 : (errcode(ERRCODE_DATA_CORRUPTED),
5541 : errmsg("control file contains invalid database cluster state")));
5542 : }
5543 :
5544 : /* This is just to allow attaching to startup process with a debugger */
5545 : #ifdef XLOG_REPLAY_DELAY
5546 : if (ControlFile->state != DB_SHUTDOWNED)
5547 : pg_usleep(60000000L);
5548 : #endif
5549 :
5550 : /*
5551 : * Verify that pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
5552 : * In cases where someone has performed a copy for PITR, these directories
5553 : * may have been excluded and need to be re-created.
5554 : */
5555 1890 : ValidateXLOGDirectoryStructure();
5556 :
5557 : /* Set up timeout handler needed to report startup progress. */
5558 1890 : if (!IsBootstrapProcessingMode())
5559 1790 : RegisterTimeout(STARTUP_PROGRESS_TIMEOUT,
5560 : startup_progress_timeout_handler);
5561 :
5562 : /*----------
5563 : * If we previously crashed, perform a couple of actions:
5564 : *
5565 : * - The pg_wal directory may still include some temporary WAL segments
5566 : * used when creating a new segment, so perform some clean up to not
5567 : * bloat this path. This is done first as there is no point to sync
5568 : * this temporary data.
5569 : *
5570 : * - There might be data which we had written, intending to fsync it, but
5571 : * which we had not actually fsync'd yet. Therefore, a power failure in
5572 : * the near future might cause earlier unflushed writes to be lost, even
5573 : * though more recent data written to disk from here on would be
5574 : * persisted. To avoid that, fsync the entire data directory.
5575 : */
5576 1890 : if (ControlFile->state != DB_SHUTDOWNED &&
5577 414 : ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
5578 : {
5579 350 : RemoveTempXlogFiles();
5580 350 : SyncDataDirectory();
5581 350 : didCrash = true;
5582 : }
5583 : else
5584 1540 : didCrash = false;
5585 :
5586 : /*
5587 : * Prepare for WAL recovery if needed.
5588 : *
5589 : * InitWalRecovery analyzes the control file and the backup label file, if
5590 : * any. It updates the in-memory ControlFile buffer according to the
5591 : * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5592 : * It also applies the tablespace map file, if any.
5593 : */
5594 1890 : InitWalRecovery(ControlFile, &wasShutdown,
5595 : &haveBackupLabel, &haveTblspcMap);
5596 1890 : checkPoint = ControlFile->checkPointCopy;
5597 :
5598 : /* initialize shared memory variables from the checkpoint record */
5599 1890 : TransamVariables->nextXid = checkPoint.nextXid;
5600 1890 : TransamVariables->nextOid = checkPoint.nextOid;
5601 1890 : TransamVariables->oidCount = 0;
5602 1890 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5603 1890 : AdvanceOldestClogXid(checkPoint.oldestXid);
5604 1890 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5605 1890 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5606 1890 : SetCommitTsLimit(checkPoint.oldestCommitTsXid,
5607 : checkPoint.newestCommitTsXid);
5608 :
5609 : /*
5610 : * Clear out any old relcache cache files. This is *necessary* if we do
5611 : * any WAL replay, since that would probably result in the cache files
5612 : * being out of sync with database reality. In theory we could leave them
5613 : * in place if the database had been cleanly shut down, but it seems
5614 : * safest to just remove them always and let them be rebuilt during the
5615 : * first backend startup. These files needs to be removed from all
5616 : * directories including pg_tblspc, however the symlinks are created only
5617 : * after reading tablespace_map file in case of archive recovery from
5618 : * backup, so needs to clear old relcache files here after creating
5619 : * symlinks.
5620 : */
5621 1890 : RelationCacheInitFileRemove();
5622 :
5623 : /*
5624 : * Initialize replication slots, before there's a chance to remove
5625 : * required resources.
5626 : */
5627 1890 : StartupReplicationSlots();
5628 :
5629 : /*
5630 : * Startup logical state, needs to be setup now so we have proper data
5631 : * during crash recovery.
5632 : */
5633 1888 : StartupReorderBuffer();
5634 :
5635 : /*
5636 : * Startup CLOG. This must be done after TransamVariables->nextXid has
5637 : * been initialized and before we accept connections or begin WAL replay.
5638 : */
5639 1888 : StartupCLOG();
5640 :
5641 : /*
5642 : * Startup MultiXact. We need to do this early to be able to replay
5643 : * truncations.
5644 : */
5645 1888 : StartupMultiXact();
5646 :
5647 : /*
5648 : * Ditto for commit timestamps. Activate the facility if the setting is
5649 : * enabled in the control file, as there should be no tracking of commit
5650 : * timestamps done when the setting was disabled. This facility can be
5651 : * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
5652 : */
5653 1888 : if (ControlFile->track_commit_timestamp)
5654 28 : StartupCommitTs();
5655 :
5656 : /*
5657 : * Recover knowledge about replay progress of known replication partners.
5658 : */
5659 1888 : StartupReplicationOrigin();
5660 :
5661 : /*
5662 : * Initialize unlogged LSN. On a clean shutdown, it's restored from the
5663 : * control file. On recovery, all unlogged relations are blown away, so
5664 : * the unlogged LSN counter can be reset too.
5665 : */
5666 1888 : if (ControlFile->state == DB_SHUTDOWNED)
5667 1462 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5668 1462 : ControlFile->unloggedLSN);
5669 : else
5670 426 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5671 : FirstNormalUnloggedLSN);
5672 :
5673 : /*
5674 : * Copy any missing timeline history files between 'now' and the recovery
5675 : * target timeline from archive to pg_wal. While we don't need those files
5676 : * ourselves - the history file of the recovery target timeline covers all
5677 : * the previous timelines in the history too - a cascading standby server
5678 : * might be interested in them. Or, if you archive the WAL from this
5679 : * server to a different archive than the primary, it'd be good for all
5680 : * the history files to get archived there after failover, so that you can
5681 : * use one of the old timelines as a PITR target. Timeline history files
5682 : * are small, so it's better to copy them unnecessarily than not copy them
5683 : * and regret later.
5684 : */
5685 1888 : restoreTimeLineHistoryFiles(checkPoint.ThisTimeLineID, recoveryTargetTLI);
5686 :
5687 : /*
5688 : * Before running in recovery, scan pg_twophase and fill in its status to
5689 : * be able to work on entries generated by redo. Doing a scan before
5690 : * taking any recovery action has the merit to discard any 2PC files that
5691 : * are newer than the first record to replay, saving from any conflicts at
5692 : * replay. This avoids as well any subsequent scans when doing recovery
5693 : * of the on-disk two-phase data.
5694 : */
5695 1888 : restoreTwoPhaseData();
5696 :
5697 : /*
5698 : * When starting with crash recovery, reset pgstat data - it might not be
5699 : * valid. Otherwise restore pgstat data. It's safe to do this here,
5700 : * because postmaster will not yet have started any other processes.
5701 : *
5702 : * NB: Restoring replication slot stats relies on slot state to have
5703 : * already been restored from disk.
5704 : *
5705 : * TODO: With a bit of extra work we could just start with a pgstat file
5706 : * associated with the checkpoint redo location we're starting from.
5707 : */
5708 1888 : if (didCrash)
5709 350 : pgstat_discard_stats();
5710 : else
5711 1538 : pgstat_restore_stats();
5712 :
5713 1888 : lastFullPageWrites = checkPoint.fullPageWrites;
5714 :
5715 1888 : RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
5716 1888 : doPageWrites = lastFullPageWrites;
5717 :
5718 : /* REDO */
5719 1888 : if (InRecovery)
5720 : {
5721 : /* Initialize state for RecoveryInProgress() */
5722 426 : SpinLockAcquire(&XLogCtl->info_lck);
5723 426 : if (InArchiveRecovery)
5724 222 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
5725 : else
5726 204 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5727 426 : SpinLockRelease(&XLogCtl->info_lck);
5728 :
5729 : /*
5730 : * Update pg_control to show that we are recovering and to show the
5731 : * selected checkpoint as the place we are starting from. We also mark
5732 : * pg_control with any minimum recovery stop point obtained from a
5733 : * backup history file.
5734 : *
5735 : * No need to hold ControlFileLock yet, we aren't up far enough.
5736 : */
5737 426 : UpdateControlFile();
5738 :
5739 : /*
5740 : * If there was a backup label file, it's done its job and the info
5741 : * has now been propagated into pg_control. We must get rid of the
5742 : * label file so that if we crash during recovery, we'll pick up at
5743 : * the latest recovery restartpoint instead of going all the way back
5744 : * to the backup start point. It seems prudent though to just rename
5745 : * the file out of the way rather than delete it completely.
5746 : */
5747 426 : if (haveBackupLabel)
5748 : {
5749 142 : unlink(BACKUP_LABEL_OLD);
5750 142 : durable_rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD, FATAL);
5751 : }
5752 :
5753 : /*
5754 : * If there was a tablespace_map file, it's done its job and the
5755 : * symlinks have been created. We must get rid of the map file so
5756 : * that if we crash during recovery, we don't create symlinks again.
5757 : * It seems prudent though to just rename the file out of the way
5758 : * rather than delete it completely.
5759 : */
5760 426 : if (haveTblspcMap)
5761 : {
5762 4 : unlink(TABLESPACE_MAP_OLD);
5763 4 : durable_rename(TABLESPACE_MAP, TABLESPACE_MAP_OLD, FATAL);
5764 : }
5765 :
5766 : /*
5767 : * Initialize our local copy of minRecoveryPoint. When doing crash
5768 : * recovery we want to replay up to the end of WAL. Particularly, in
5769 : * the case of a promoted standby minRecoveryPoint value in the
5770 : * control file is only updated after the first checkpoint. However,
5771 : * if the instance crashes before the first post-recovery checkpoint
5772 : * is completed then recovery will use a stale location causing the
5773 : * startup process to think that there are still invalid page
5774 : * references when checking for data consistency.
5775 : */
5776 426 : if (InArchiveRecovery)
5777 : {
5778 222 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
5779 222 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
5780 : }
5781 : else
5782 : {
5783 204 : LocalMinRecoveryPoint = InvalidXLogRecPtr;
5784 204 : LocalMinRecoveryPointTLI = 0;
5785 : }
5786 :
5787 : /* Check that the GUCs used to generate the WAL allow recovery */
5788 426 : CheckRequiredParameterValues();
5789 :
5790 : /*
5791 : * We're in recovery, so unlogged relations may be trashed and must be
5792 : * reset. This should be done BEFORE allowing Hot Standby
5793 : * connections, so that read-only backends don't try to read whatever
5794 : * garbage is left over from before.
5795 : */
5796 426 : ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
5797 :
5798 : /*
5799 : * Likewise, delete any saved transaction snapshot files that got left
5800 : * behind by crashed backends.
5801 : */
5802 426 : DeleteAllExportedSnapshotFiles();
5803 :
5804 : /*
5805 : * Initialize for Hot Standby, if enabled. We won't let backends in
5806 : * yet, not until we've reached the min recovery point specified in
5807 : * control file and we've established a recovery snapshot from a
5808 : * running-xacts WAL record.
5809 : */
5810 426 : if (ArchiveRecoveryRequested && EnableHotStandby)
5811 : {
5812 : TransactionId *xids;
5813 : int nxids;
5814 :
5815 210 : ereport(DEBUG1,
5816 : (errmsg_internal("initializing for hot standby")));
5817 :
5818 210 : InitRecoveryTransactionEnvironment();
5819 :
5820 210 : if (wasShutdown)
5821 52 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
5822 : else
5823 158 : oldestActiveXID = checkPoint.oldestActiveXid;
5824 : Assert(TransactionIdIsValid(oldestActiveXID));
5825 :
5826 : /* Tell procarray about the range of xids it has to deal with */
5827 210 : ProcArrayInitRecovery(XidFromFullTransactionId(TransamVariables->nextXid));
5828 :
5829 : /*
5830 : * Startup subtrans only. CLOG, MultiXact and commit timestamp
5831 : * have already been started up and other SLRUs are not maintained
5832 : * during recovery and need not be started yet.
5833 : */
5834 210 : StartupSUBTRANS(oldestActiveXID);
5835 :
5836 : /*
5837 : * If we're beginning at a shutdown checkpoint, we know that
5838 : * nothing was running on the primary at this point. So fake-up an
5839 : * empty running-xacts record and use that here and now. Recover
5840 : * additional standby state for prepared transactions.
5841 : */
5842 210 : if (wasShutdown)
5843 : {
5844 : RunningTransactionsData running;
5845 : TransactionId latestCompletedXid;
5846 :
5847 : /* Update pg_subtrans entries for any prepared transactions */
5848 52 : StandbyRecoverPreparedTransactions();
5849 :
5850 : /*
5851 : * Construct a RunningTransactions snapshot representing a
5852 : * shut down server, with only prepared transactions still
5853 : * alive. We're never overflowed at this point because all
5854 : * subxids are listed with their parent prepared transactions.
5855 : */
5856 52 : running.xcnt = nxids;
5857 52 : running.subxcnt = 0;
5858 52 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
5859 52 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
5860 52 : running.oldestRunningXid = oldestActiveXID;
5861 52 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
5862 52 : TransactionIdRetreat(latestCompletedXid);
5863 : Assert(TransactionIdIsNormal(latestCompletedXid));
5864 52 : running.latestCompletedXid = latestCompletedXid;
5865 52 : running.xids = xids;
5866 :
5867 52 : ProcArrayApplyRecoveryInfo(&running);
5868 : }
5869 : }
5870 :
5871 : /*
5872 : * We're all set for replaying the WAL now. Do it.
5873 : */
5874 426 : PerformWalRecovery();
5875 310 : performedWalRecovery = true;
5876 : }
5877 : else
5878 1462 : performedWalRecovery = false;
5879 :
5880 : /*
5881 : * Finish WAL recovery.
5882 : */
5883 1772 : endOfRecoveryInfo = FinishWalRecovery();
5884 1772 : EndOfLog = endOfRecoveryInfo->endOfLog;
5885 1772 : EndOfLogTLI = endOfRecoveryInfo->endOfLogTLI;
5886 1772 : abortedRecPtr = endOfRecoveryInfo->abortedRecPtr;
5887 1772 : missingContrecPtr = endOfRecoveryInfo->missingContrecPtr;
5888 :
5889 : /*
5890 : * Reset ps status display, so as no information related to recovery shows
5891 : * up.
5892 : */
5893 1772 : set_ps_display("");
5894 :
5895 : /*
5896 : * When recovering from a backup (we are in recovery, and archive recovery
5897 : * was requested), complain if we did not roll forward far enough to reach
5898 : * the point where the database is consistent. For regular online
5899 : * backup-from-primary, that means reaching the end-of-backup WAL record
5900 : * (at which point we reset backupStartPoint to be Invalid), for
5901 : * backup-from-replica (which can't inject records into the WAL stream),
5902 : * that point is when we reach the minRecoveryPoint in pg_control (which
5903 : * we purposefully copy last when backing up from a replica). For
5904 : * pg_rewind (which creates a backup_label with a method of "pg_rewind")
5905 : * or snapshot-style backups (which don't), backupEndRequired will be set
5906 : * to false.
5907 : *
5908 : * Note: it is indeed okay to look at the local variable
5909 : * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
5910 : * might be further ahead --- ControlFile->minRecoveryPoint cannot have
5911 : * been advanced beyond the WAL we processed.
5912 : */
5913 1772 : if (InRecovery &&
5914 310 : (EndOfLog < LocalMinRecoveryPoint ||
5915 310 : !XLogRecPtrIsInvalid(ControlFile->backupStartPoint)))
5916 : {
5917 : /*
5918 : * Ran off end of WAL before reaching end-of-backup WAL record, or
5919 : * minRecoveryPoint. That's a bad sign, indicating that you tried to
5920 : * recover from an online backup but never called pg_backup_stop(), or
5921 : * you didn't archive all the WAL needed.
5922 : */
5923 0 : if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
5924 : {
5925 0 : if (!XLogRecPtrIsInvalid(ControlFile->backupStartPoint) || ControlFile->backupEndRequired)
5926 0 : ereport(FATAL,
5927 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5928 : errmsg("WAL ends before end of online backup"),
5929 : errhint("All WAL generated while online backup was taken must be available at recovery.")));
5930 : else
5931 0 : ereport(FATAL,
5932 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5933 : errmsg("WAL ends before consistent recovery point")));
5934 : }
5935 : }
5936 :
5937 : /*
5938 : * Reset unlogged relations to the contents of their INIT fork. This is
5939 : * done AFTER recovery is complete so as to include any unlogged relations
5940 : * created during recovery, but BEFORE recovery is marked as having
5941 : * completed successfully. Otherwise we'd not retry if any of the post
5942 : * end-of-recovery steps fail.
5943 : */
5944 1772 : if (InRecovery)
5945 310 : ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
5946 :
5947 : /*
5948 : * Pre-scan prepared transactions to find out the range of XIDs present.
5949 : * This information is not quite needed yet, but it is positioned here so
5950 : * as potential problems are detected before any on-disk change is done.
5951 : */
5952 1772 : oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
5953 :
5954 : /*
5955 : * Allow ordinary WAL segment creation before possibly switching to a new
5956 : * timeline, which creates a new segment, and after the last ReadRecord().
5957 : */
5958 1772 : SetInstallXLogFileSegmentActive();
5959 :
5960 : /*
5961 : * Consider whether we need to assign a new timeline ID.
5962 : *
5963 : * If we did archive recovery, we always assign a new ID. This handles a
5964 : * couple of issues. If we stopped short of the end of WAL during
5965 : * recovery, then we are clearly generating a new timeline and must assign
5966 : * it a unique new ID. Even if we ran to the end, modifying the current
5967 : * last segment is problematic because it may result in trying to
5968 : * overwrite an already-archived copy of that segment, and we encourage
5969 : * DBAs to make their archive_commands reject that. We can dodge the
5970 : * problem by making the new active segment have a new timeline ID.
5971 : *
5972 : * In a normal crash recovery, we can just extend the timeline we were in.
5973 : */
5974 1772 : newTLI = endOfRecoveryInfo->lastRecTLI;
5975 1772 : if (ArchiveRecoveryRequested)
5976 : {
5977 98 : newTLI = findNewestTimeLine(recoveryTargetTLI) + 1;
5978 98 : ereport(LOG,
5979 : (errmsg("selected new timeline ID: %u", newTLI)));
5980 :
5981 : /*
5982 : * Make a writable copy of the last WAL segment. (Note that we also
5983 : * have a copy of the last block of the old WAL in
5984 : * endOfRecovery->lastPage; we will use that below.)
5985 : */
5986 98 : XLogInitNewTimeline(EndOfLogTLI, EndOfLog, newTLI);
5987 :
5988 : /*
5989 : * Remove the signal files out of the way, so that we don't
5990 : * accidentally re-enter archive recovery mode in a subsequent crash.
5991 : */
5992 98 : if (endOfRecoveryInfo->standby_signal_file_found)
5993 92 : durable_unlink(STANDBY_SIGNAL_FILE, FATAL);
5994 :
5995 98 : if (endOfRecoveryInfo->recovery_signal_file_found)
5996 6 : durable_unlink(RECOVERY_SIGNAL_FILE, FATAL);
5997 :
5998 : /*
5999 : * Write the timeline history file, and have it archived. After this
6000 : * point (or rather, as soon as the file is archived), the timeline
6001 : * will appear as "taken" in the WAL archive and to any standby
6002 : * servers. If we crash before actually switching to the new
6003 : * timeline, standby servers will nevertheless think that we switched
6004 : * to the new timeline, and will try to connect to the new timeline.
6005 : * To minimize the window for that, try to do as little as possible
6006 : * between here and writing the end-of-recovery record.
6007 : */
6008 98 : writeTimeLineHistory(newTLI, recoveryTargetTLI,
6009 : EndOfLog, endOfRecoveryInfo->recoveryStopReason);
6010 :
6011 98 : ereport(LOG,
6012 : (errmsg("archive recovery complete")));
6013 : }
6014 :
6015 : /* Save the selected TimeLineID in shared memory, too */
6016 1772 : SpinLockAcquire(&XLogCtl->info_lck);
6017 1772 : XLogCtl->InsertTimeLineID = newTLI;
6018 1772 : XLogCtl->PrevTimeLineID = endOfRecoveryInfo->lastRecTLI;
6019 1772 : SpinLockRelease(&XLogCtl->info_lck);
6020 :
6021 : /*
6022 : * Actually, if WAL ended in an incomplete record, skip the parts that
6023 : * made it through and start writing after the portion that persisted.
6024 : * (It's critical to first write an OVERWRITE_CONTRECORD message, which
6025 : * we'll do as soon as we're open for writing new WAL.)
6026 : */
6027 1772 : if (!XLogRecPtrIsInvalid(missingContrecPtr))
6028 : {
6029 : /*
6030 : * We should only have a missingContrecPtr if we're not switching to a
6031 : * new timeline. When a timeline switch occurs, WAL is copied from the
6032 : * old timeline to the new only up to the end of the last complete
6033 : * record, so there can't be an incomplete WAL record that we need to
6034 : * disregard.
6035 : */
6036 : Assert(newTLI == endOfRecoveryInfo->lastRecTLI);
6037 : Assert(!XLogRecPtrIsInvalid(abortedRecPtr));
6038 22 : EndOfLog = missingContrecPtr;
6039 : }
6040 :
6041 : /*
6042 : * Prepare to write WAL starting at EndOfLog location, and init xlog
6043 : * buffer cache using the block containing the last record from the
6044 : * previous incarnation.
6045 : */
6046 1772 : Insert = &XLogCtl->Insert;
6047 1772 : Insert->PrevBytePos = XLogRecPtrToBytePos(endOfRecoveryInfo->lastRec);
6048 1772 : Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
6049 :
6050 : /*
6051 : * Tricky point here: lastPage contains the *last* block that the LastRec
6052 : * record spans, not the one it starts in. The last block is indeed the
6053 : * one we want to use.
6054 : */
6055 1772 : if (EndOfLog % XLOG_BLCKSZ != 0)
6056 : {
6057 : char *page;
6058 : int len;
6059 : int firstIdx;
6060 :
6061 1716 : firstIdx = XLogRecPtrToBufIdx(EndOfLog);
6062 1716 : len = EndOfLog - endOfRecoveryInfo->lastPageBeginPtr;
6063 : Assert(len < XLOG_BLCKSZ);
6064 :
6065 : /* Copy the valid part of the last block, and zero the rest */
6066 1716 : page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
6067 1716 : memcpy(page, endOfRecoveryInfo->lastPage, len);
6068 1716 : memset(page + len, 0, XLOG_BLCKSZ - len);
6069 :
6070 1716 : pg_atomic_write_u64(&XLogCtl->xlblocks[firstIdx], endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6071 1716 : XLogCtl->InitializedUpTo = endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ;
6072 : }
6073 : else
6074 : {
6075 : /*
6076 : * There is no partial block to copy. Just set InitializedUpTo, and
6077 : * let the first attempt to insert a log record to initialize the next
6078 : * buffer.
6079 : */
6080 56 : XLogCtl->InitializedUpTo = EndOfLog;
6081 : }
6082 :
6083 : /*
6084 : * Update local and shared status. This is OK to do without any locks
6085 : * because no other process can be reading or writing WAL yet.
6086 : */
6087 1772 : LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
6088 1772 : pg_atomic_write_u64(&XLogCtl->logInsertResult, EndOfLog);
6089 1772 : pg_atomic_write_u64(&XLogCtl->logWriteResult, EndOfLog);
6090 1772 : pg_atomic_write_u64(&XLogCtl->logFlushResult, EndOfLog);
6091 1772 : XLogCtl->LogwrtRqst.Write = EndOfLog;
6092 1772 : XLogCtl->LogwrtRqst.Flush = EndOfLog;
6093 :
6094 : /*
6095 : * Preallocate additional log files, if wanted.
6096 : */
6097 1772 : PreallocXlogFiles(EndOfLog, newTLI);
6098 :
6099 : /*
6100 : * Okay, we're officially UP.
6101 : */
6102 1772 : InRecovery = false;
6103 :
6104 : /* start the archive_timeout timer and LSN running */
6105 1772 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
6106 1772 : XLogCtl->lastSegSwitchLSN = EndOfLog;
6107 :
6108 : /* also initialize latestCompletedXid, to nextXid - 1 */
6109 1772 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
6110 1772 : TransamVariables->latestCompletedXid = TransamVariables->nextXid;
6111 1772 : FullTransactionIdRetreat(&TransamVariables->latestCompletedXid);
6112 1772 : LWLockRelease(ProcArrayLock);
6113 :
6114 : /*
6115 : * Start up subtrans, if not already done for hot standby. (commit
6116 : * timestamps are started below, if necessary.)
6117 : */
6118 1772 : if (standbyState == STANDBY_DISABLED)
6119 1674 : StartupSUBTRANS(oldestActiveXID);
6120 :
6121 : /*
6122 : * Perform end of recovery actions for any SLRUs that need it.
6123 : */
6124 1772 : TrimCLOG();
6125 1772 : TrimMultiXact();
6126 :
6127 : /*
6128 : * Reload shared-memory state for prepared transactions. This needs to
6129 : * happen before renaming the last partial segment of the old timeline as
6130 : * it may be possible that we have to recover some transactions from it.
6131 : */
6132 1772 : RecoverPreparedTransactions();
6133 :
6134 : /* Shut down xlogreader */
6135 1772 : ShutdownWalRecovery();
6136 :
6137 : /* Enable WAL writes for this backend only. */
6138 1772 : LocalSetXLogInsertAllowed();
6139 :
6140 : /* If necessary, write overwrite-contrecord before doing anything else */
6141 1772 : if (!XLogRecPtrIsInvalid(abortedRecPtr))
6142 : {
6143 : Assert(!XLogRecPtrIsInvalid(missingContrecPtr));
6144 22 : CreateOverwriteContrecordRecord(abortedRecPtr, missingContrecPtr, newTLI);
6145 : }
6146 :
6147 : /*
6148 : * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
6149 : * record before resource manager writes cleanup WAL records or checkpoint
6150 : * record is written.
6151 : */
6152 1772 : Insert->fullPageWrites = lastFullPageWrites;
6153 1772 : UpdateFullPageWrites();
6154 :
6155 : /*
6156 : * Emit checkpoint or end-of-recovery record in XLOG, if required.
6157 : */
6158 1772 : if (performedWalRecovery)
6159 310 : promoted = PerformRecoveryXLogAction();
6160 :
6161 : /*
6162 : * If any of the critical GUCs have changed, log them before we allow
6163 : * backends to write WAL.
6164 : */
6165 1772 : XLogReportParameters();
6166 :
6167 : /* If this is archive recovery, perform post-recovery cleanup actions. */
6168 1772 : if (ArchiveRecoveryRequested)
6169 98 : CleanupAfterArchiveRecovery(EndOfLogTLI, EndOfLog, newTLI);
6170 :
6171 : /*
6172 : * Local WAL inserts enabled, so it's time to finish initialization of
6173 : * commit timestamp.
6174 : */
6175 1772 : CompleteCommitTsInitialization();
6176 :
6177 : /* Clean up EndOfWalRecoveryInfo data to appease Valgrind leak checking */
6178 1772 : if (endOfRecoveryInfo->lastPage)
6179 1738 : pfree(endOfRecoveryInfo->lastPage);
6180 1772 : pfree(endOfRecoveryInfo->recoveryStopReason);
6181 1772 : pfree(endOfRecoveryInfo);
6182 :
6183 : /*
6184 : * All done with end-of-recovery actions.
6185 : *
6186 : * Now allow backends to write WAL and update the control file status in
6187 : * consequence. SharedRecoveryState, that controls if backends can write
6188 : * WAL, is updated while holding ControlFileLock to prevent other backends
6189 : * to look at an inconsistent state of the control file in shared memory.
6190 : * There is still a small window during which backends can write WAL and
6191 : * the control file is still referring to a system not in DB_IN_PRODUCTION
6192 : * state while looking at the on-disk control file.
6193 : *
6194 : * Also, we use info_lck to update SharedRecoveryState to ensure that
6195 : * there are no race conditions concerning visibility of other recent
6196 : * updates to shared memory.
6197 : */
6198 1772 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6199 1772 : ControlFile->state = DB_IN_PRODUCTION;
6200 :
6201 1772 : SpinLockAcquire(&XLogCtl->info_lck);
6202 1772 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_DONE;
6203 1772 : SpinLockRelease(&XLogCtl->info_lck);
6204 :
6205 1772 : UpdateControlFile();
6206 1772 : LWLockRelease(ControlFileLock);
6207 :
6208 : /*
6209 : * Shutdown the recovery environment. This must occur after
6210 : * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
6211 : * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
6212 : * any session building a snapshot will not rely on KnownAssignedXids as
6213 : * RecoveryInProgress() would return false at this stage. This is
6214 : * particularly critical for prepared 2PC transactions, that would still
6215 : * need to be included in snapshots once recovery has ended.
6216 : */
6217 1772 : if (standbyState != STANDBY_DISABLED)
6218 98 : ShutdownRecoveryTransactionEnvironment();
6219 :
6220 : /*
6221 : * If there were cascading standby servers connected to us, nudge any wal
6222 : * sender processes to notice that we've been promoted.
6223 : */
6224 1772 : WalSndWakeup(true, true);
6225 :
6226 : /*
6227 : * If this was a promotion, request an (online) checkpoint now. This isn't
6228 : * required for consistency, but the last restartpoint might be far back,
6229 : * and in case of a crash, recovering from it might take a longer than is
6230 : * appropriate now that we're not in standby mode anymore.
6231 : */
6232 1772 : if (promoted)
6233 84 : RequestCheckpoint(CHECKPOINT_FORCE);
6234 1772 : }
6235 :
6236 : /*
6237 : * Callback from PerformWalRecovery(), called when we switch from crash
6238 : * recovery to archive recovery mode. Updates the control file accordingly.
6239 : */
6240 : void
6241 4 : SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr, TimeLineID replayTLI)
6242 : {
6243 : /* initialize minRecoveryPoint to this record */
6244 4 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6245 4 : ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6246 4 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6247 : {
6248 4 : ControlFile->minRecoveryPoint = EndRecPtr;
6249 4 : ControlFile->minRecoveryPointTLI = replayTLI;
6250 : }
6251 : /* update local copy */
6252 4 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6253 4 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6254 :
6255 : /*
6256 : * The startup process can update its local copy of minRecoveryPoint from
6257 : * this point.
6258 : */
6259 4 : updateMinRecoveryPoint = true;
6260 :
6261 4 : UpdateControlFile();
6262 :
6263 : /*
6264 : * We update SharedRecoveryState while holding the lock on ControlFileLock
6265 : * so both states are consistent in shared memory.
6266 : */
6267 4 : SpinLockAcquire(&XLogCtl->info_lck);
6268 4 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6269 4 : SpinLockRelease(&XLogCtl->info_lck);
6270 :
6271 4 : LWLockRelease(ControlFileLock);
6272 4 : }
6273 :
6274 : /*
6275 : * Callback from PerformWalRecovery(), called when we reach the end of backup.
6276 : * Updates the control file accordingly.
6277 : */
6278 : void
6279 142 : ReachedEndOfBackup(XLogRecPtr EndRecPtr, TimeLineID tli)
6280 : {
6281 : /*
6282 : * We have reached the end of base backup, as indicated by pg_control. The
6283 : * data on disk is now consistent (unless minRecoveryPoint is further
6284 : * ahead, which can happen if we crashed during previous recovery). Reset
6285 : * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
6286 : * make sure we don't allow starting up at an earlier point even if
6287 : * recovery is stopped and restarted soon after this.
6288 : */
6289 142 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6290 :
6291 142 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6292 : {
6293 134 : ControlFile->minRecoveryPoint = EndRecPtr;
6294 134 : ControlFile->minRecoveryPointTLI = tli;
6295 : }
6296 :
6297 142 : ControlFile->backupStartPoint = InvalidXLogRecPtr;
6298 142 : ControlFile->backupEndPoint = InvalidXLogRecPtr;
6299 142 : ControlFile->backupEndRequired = false;
6300 142 : UpdateControlFile();
6301 :
6302 142 : LWLockRelease(ControlFileLock);
6303 142 : }
6304 :
6305 : /*
6306 : * Perform whatever XLOG actions are necessary at end of REDO.
6307 : *
6308 : * The goal here is to make sure that we'll be able to recover properly if
6309 : * we crash again. If we choose to write a checkpoint, we'll write a shutdown
6310 : * checkpoint rather than an on-line one. This is not particularly critical,
6311 : * but since we may be assigning a new TLI, using a shutdown checkpoint allows
6312 : * us to have the rule that TLI only changes in shutdown checkpoints, which
6313 : * allows some extra error checking in xlog_redo.
6314 : */
6315 : static bool
6316 310 : PerformRecoveryXLogAction(void)
6317 : {
6318 310 : bool promoted = false;
6319 :
6320 : /*
6321 : * Perform a checkpoint to update all our recovery activity to disk.
6322 : *
6323 : * Note that we write a shutdown checkpoint rather than an on-line one.
6324 : * This is not particularly critical, but since we may be assigning a new
6325 : * TLI, using a shutdown checkpoint allows us to have the rule that TLI
6326 : * only changes in shutdown checkpoints, which allows some extra error
6327 : * checking in xlog_redo.
6328 : *
6329 : * In promotion, only create a lightweight end-of-recovery record instead
6330 : * of a full checkpoint. A checkpoint is requested later, after we're
6331 : * fully out of recovery mode and already accepting queries.
6332 : */
6333 408 : if (ArchiveRecoveryRequested && IsUnderPostmaster &&
6334 98 : PromoteIsTriggered())
6335 : {
6336 84 : promoted = true;
6337 :
6338 : /*
6339 : * Insert a special WAL record to mark the end of recovery, since we
6340 : * aren't doing a checkpoint. That means that the checkpointer process
6341 : * may likely be in the middle of a time-smoothed restartpoint and
6342 : * could continue to be for minutes after this. That sounds strange,
6343 : * but the effect is roughly the same and it would be stranger to try
6344 : * to come out of the restartpoint and then checkpoint. We request a
6345 : * checkpoint later anyway, just for safety.
6346 : */
6347 84 : CreateEndOfRecoveryRecord();
6348 : }
6349 : else
6350 : {
6351 226 : RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
6352 : CHECKPOINT_FAST |
6353 : CHECKPOINT_WAIT);
6354 : }
6355 :
6356 310 : return promoted;
6357 : }
6358 :
6359 : /*
6360 : * Is the system still in recovery?
6361 : *
6362 : * Unlike testing InRecovery, this works in any process that's connected to
6363 : * shared memory.
6364 : */
6365 : bool
6366 177840434 : RecoveryInProgress(void)
6367 : {
6368 : /*
6369 : * We check shared state each time only until we leave recovery mode. We
6370 : * can't re-enter recovery, so there's no need to keep checking after the
6371 : * shared variable has once been seen false.
6372 : */
6373 177840434 : if (!LocalRecoveryInProgress)
6374 173891154 : return false;
6375 : else
6376 : {
6377 : /*
6378 : * use volatile pointer to make sure we make a fresh read of the
6379 : * shared variable.
6380 : */
6381 3949280 : volatile XLogCtlData *xlogctl = XLogCtl;
6382 :
6383 3949280 : LocalRecoveryInProgress = (xlogctl->SharedRecoveryState != RECOVERY_STATE_DONE);
6384 :
6385 : /*
6386 : * Note: We don't need a memory barrier when we're still in recovery.
6387 : * We might exit recovery immediately after return, so the caller
6388 : * can't rely on 'true' meaning that we're still in recovery anyway.
6389 : */
6390 :
6391 3949280 : return LocalRecoveryInProgress;
6392 : }
6393 : }
6394 :
6395 : /*
6396 : * Returns current recovery state from shared memory.
6397 : *
6398 : * This returned state is kept consistent with the contents of the control
6399 : * file. See details about the possible values of RecoveryState in xlog.h.
6400 : */
6401 : RecoveryState
6402 36356 : GetRecoveryState(void)
6403 : {
6404 : RecoveryState retval;
6405 :
6406 36356 : SpinLockAcquire(&XLogCtl->info_lck);
6407 36356 : retval = XLogCtl->SharedRecoveryState;
6408 36356 : SpinLockRelease(&XLogCtl->info_lck);
6409 :
6410 36356 : return retval;
6411 : }
6412 :
6413 : /*
6414 : * Is this process allowed to insert new WAL records?
6415 : *
6416 : * Ordinarily this is essentially equivalent to !RecoveryInProgress().
6417 : * But we also have provisions for forcing the result "true" or "false"
6418 : * within specific processes regardless of the global state.
6419 : */
6420 : bool
6421 60885670 : XLogInsertAllowed(void)
6422 : {
6423 : /*
6424 : * If value is "unconditionally true" or "unconditionally false", just
6425 : * return it. This provides the normal fast path once recovery is known
6426 : * done.
6427 : */
6428 60885670 : if (LocalXLogInsertAllowed >= 0)
6429 60656546 : return (bool) LocalXLogInsertAllowed;
6430 :
6431 : /*
6432 : * Else, must check to see if we're still in recovery.
6433 : */
6434 229124 : if (RecoveryInProgress())
6435 212460 : return false;
6436 :
6437 : /*
6438 : * On exit from recovery, reset to "unconditionally true", since there is
6439 : * no need to keep checking.
6440 : */
6441 16664 : LocalXLogInsertAllowed = 1;
6442 16664 : return true;
6443 : }
6444 :
6445 : /*
6446 : * Make XLogInsertAllowed() return true in the current process only.
6447 : *
6448 : * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
6449 : * and even call LocalSetXLogInsertAllowed() again after that.
6450 : *
6451 : * Returns the previous value of LocalXLogInsertAllowed.
6452 : */
6453 : static int
6454 1830 : LocalSetXLogInsertAllowed(void)
6455 : {
6456 1830 : int oldXLogAllowed = LocalXLogInsertAllowed;
6457 :
6458 1830 : LocalXLogInsertAllowed = 1;
6459 :
6460 1830 : return oldXLogAllowed;
6461 : }
6462 :
6463 : /*
6464 : * Return the current Redo pointer from shared memory.
6465 : *
6466 : * As a side-effect, the local RedoRecPtr copy is updated.
6467 : */
6468 : XLogRecPtr
6469 589962 : GetRedoRecPtr(void)
6470 : {
6471 : XLogRecPtr ptr;
6472 :
6473 : /*
6474 : * The possibly not up-to-date copy in XlogCtl is enough. Even if we
6475 : * grabbed a WAL insertion lock to read the authoritative value in
6476 : * Insert->RedoRecPtr, someone might update it just after we've released
6477 : * the lock.
6478 : */
6479 589962 : SpinLockAcquire(&XLogCtl->info_lck);
6480 589962 : ptr = XLogCtl->RedoRecPtr;
6481 589962 : SpinLockRelease(&XLogCtl->info_lck);
6482 :
6483 589962 : if (RedoRecPtr < ptr)
6484 2976 : RedoRecPtr = ptr;
6485 :
6486 589962 : return RedoRecPtr;
6487 : }
6488 :
6489 : /*
6490 : * Return information needed to decide whether a modified block needs a
6491 : * full-page image to be included in the WAL record.
6492 : *
6493 : * The returned values are cached copies from backend-private memory, and
6494 : * possibly out-of-date or, indeed, uninitialized, in which case they will
6495 : * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6496 : * re-check them against up-to-date values, while holding the WAL insert lock.
6497 : */
6498 : void
6499 29423928 : GetFullPageWriteInfo(XLogRecPtr *RedoRecPtr_p, bool *doPageWrites_p)
6500 : {
6501 29423928 : *RedoRecPtr_p = RedoRecPtr;
6502 29423928 : *doPageWrites_p = doPageWrites;
6503 29423928 : }
6504 :
6505 : /*
6506 : * GetInsertRecPtr -- Returns the current insert position.
6507 : *
6508 : * NOTE: The value *actually* returned is the position of the last full
6509 : * xlog page. It lags behind the real insert position by at most 1 page.
6510 : * For that, we don't need to scan through WAL insertion locks, and an
6511 : * approximation is enough for the current usage of this function.
6512 : */
6513 : XLogRecPtr
6514 16464 : GetInsertRecPtr(void)
6515 : {
6516 : XLogRecPtr recptr;
6517 :
6518 16464 : SpinLockAcquire(&XLogCtl->info_lck);
6519 16464 : recptr = XLogCtl->LogwrtRqst.Write;
6520 16464 : SpinLockRelease(&XLogCtl->info_lck);
6521 :
6522 16464 : return recptr;
6523 : }
6524 :
6525 : /*
6526 : * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6527 : * position known to be fsync'd to disk. This should only be used on a
6528 : * system that is known not to be in recovery.
6529 : */
6530 : XLogRecPtr
6531 461310 : GetFlushRecPtr(TimeLineID *insertTLI)
6532 : {
6533 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6534 :
6535 461310 : RefreshXLogWriteResult(LogwrtResult);
6536 :
6537 : /*
6538 : * If we're writing and flushing WAL, the time line can't be changing, so
6539 : * no lock is required.
6540 : */
6541 461310 : if (insertTLI)
6542 42974 : *insertTLI = XLogCtl->InsertTimeLineID;
6543 :
6544 461310 : return LogwrtResult.Flush;
6545 : }
6546 :
6547 : /*
6548 : * GetWALInsertionTimeLine -- Returns the current timeline of a system that
6549 : * is not in recovery.
6550 : */
6551 : TimeLineID
6552 214078 : GetWALInsertionTimeLine(void)
6553 : {
6554 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6555 :
6556 : /* Since the value can't be changing, no lock is required. */
6557 214078 : return XLogCtl->InsertTimeLineID;
6558 : }
6559 :
6560 : /*
6561 : * GetWALInsertionTimeLineIfSet -- If the system is not in recovery, returns
6562 : * the WAL insertion timeline; else, returns 0. Wherever possible, use
6563 : * GetWALInsertionTimeLine() instead, since it's cheaper. Note that this
6564 : * function decides recovery has ended as soon as the insert TLI is set, which
6565 : * happens before we set XLogCtl->SharedRecoveryState to RECOVERY_STATE_DONE.
6566 : */
6567 : TimeLineID
6568 0 : GetWALInsertionTimeLineIfSet(void)
6569 : {
6570 : TimeLineID insertTLI;
6571 :
6572 0 : SpinLockAcquire(&XLogCtl->info_lck);
6573 0 : insertTLI = XLogCtl->InsertTimeLineID;
6574 0 : SpinLockRelease(&XLogCtl->info_lck);
6575 :
6576 0 : return insertTLI;
6577 : }
6578 :
6579 : /*
6580 : * GetLastImportantRecPtr -- Returns the LSN of the last important record
6581 : * inserted. All records not explicitly marked as unimportant are considered
6582 : * important.
6583 : *
6584 : * The LSN is determined by computing the maximum of
6585 : * WALInsertLocks[i].lastImportantAt.
6586 : */
6587 : XLogRecPtr
6588 3138 : GetLastImportantRecPtr(void)
6589 : {
6590 3138 : XLogRecPtr res = InvalidXLogRecPtr;
6591 : int i;
6592 :
6593 28242 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
6594 : {
6595 : XLogRecPtr last_important;
6596 :
6597 : /*
6598 : * Need to take a lock to prevent torn reads of the LSN, which are
6599 : * possible on some of the supported platforms. WAL insert locks only
6600 : * support exclusive mode, so we have to use that.
6601 : */
6602 25104 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
6603 25104 : last_important = WALInsertLocks[i].l.lastImportantAt;
6604 25104 : LWLockRelease(&WALInsertLocks[i].l.lock);
6605 :
6606 25104 : if (res < last_important)
6607 5392 : res = last_important;
6608 : }
6609 :
6610 3138 : return res;
6611 : }
6612 :
6613 : /*
6614 : * Get the time and LSN of the last xlog segment switch
6615 : */
6616 : pg_time_t
6617 0 : GetLastSegSwitchData(XLogRecPtr *lastSwitchLSN)
6618 : {
6619 : pg_time_t result;
6620 :
6621 : /* Need WALWriteLock, but shared lock is sufficient */
6622 0 : LWLockAcquire(WALWriteLock, LW_SHARED);
6623 0 : result = XLogCtl->lastSegSwitchTime;
6624 0 : *lastSwitchLSN = XLogCtl->lastSegSwitchLSN;
6625 0 : LWLockRelease(WALWriteLock);
6626 :
6627 0 : return result;
6628 : }
6629 :
6630 : /*
6631 : * This must be called ONCE during postmaster or standalone-backend shutdown
6632 : */
6633 : void
6634 1294 : ShutdownXLOG(int code, Datum arg)
6635 : {
6636 : /*
6637 : * We should have an aux process resource owner to use, and we should not
6638 : * be in a transaction that's installed some other resowner.
6639 : */
6640 : Assert(AuxProcessResourceOwner != NULL);
6641 : Assert(CurrentResourceOwner == NULL ||
6642 : CurrentResourceOwner == AuxProcessResourceOwner);
6643 1294 : CurrentResourceOwner = AuxProcessResourceOwner;
6644 :
6645 : /* Don't be chatty in standalone mode */
6646 1294 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
6647 : (errmsg("shutting down")));
6648 :
6649 : /*
6650 : * Signal walsenders to move to stopping state.
6651 : */
6652 1294 : WalSndInitStopping();
6653 :
6654 : /*
6655 : * Wait for WAL senders to be in stopping state. This prevents commands
6656 : * from writing new WAL.
6657 : */
6658 1294 : WalSndWaitStopping();
6659 :
6660 1294 : if (RecoveryInProgress())
6661 110 : CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
6662 : else
6663 : {
6664 : /*
6665 : * If archiving is enabled, rotate the last XLOG file so that all the
6666 : * remaining records are archived (postmaster wakes up the archiver
6667 : * process one more time at the end of shutdown). The checkpoint
6668 : * record will go to the next XLOG file and won't be archived (yet).
6669 : */
6670 1184 : if (XLogArchivingActive())
6671 28 : RequestXLogSwitch(false);
6672 :
6673 1184 : CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
6674 : }
6675 1294 : }
6676 :
6677 : /*
6678 : * Log start of a checkpoint.
6679 : */
6680 : static void
6681 2844 : LogCheckpointStart(int flags, bool restartpoint)
6682 : {
6683 2844 : if (restartpoint)
6684 380 : ereport(LOG,
6685 : /* translator: the placeholders show checkpoint options */
6686 : (errmsg("restartpoint starting:%s%s%s%s%s%s%s%s",
6687 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6688 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6689 : (flags & CHECKPOINT_FAST) ? " fast" : "",
6690 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6691 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6692 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6693 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6694 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "")));
6695 : else
6696 2464 : ereport(LOG,
6697 : /* translator: the placeholders show checkpoint options */
6698 : (errmsg("checkpoint starting:%s%s%s%s%s%s%s%s",
6699 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6700 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6701 : (flags & CHECKPOINT_FAST) ? " fast" : "",
6702 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6703 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6704 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6705 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6706 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "")));
6707 2844 : }
6708 :
6709 : /*
6710 : * Log end of a checkpoint.
6711 : */
6712 : static void
6713 3428 : LogCheckpointEnd(bool restartpoint)
6714 : {
6715 : long write_msecs,
6716 : sync_msecs,
6717 : total_msecs,
6718 : longest_msecs,
6719 : average_msecs;
6720 : uint64 average_sync_time;
6721 :
6722 3428 : CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
6723 :
6724 3428 : write_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_write_t,
6725 : CheckpointStats.ckpt_sync_t);
6726 :
6727 3428 : sync_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_sync_t,
6728 : CheckpointStats.ckpt_sync_end_t);
6729 :
6730 : /* Accumulate checkpoint timing summary data, in milliseconds. */
6731 3428 : PendingCheckpointerStats.write_time += write_msecs;
6732 3428 : PendingCheckpointerStats.sync_time += sync_msecs;
6733 :
6734 : /*
6735 : * All of the published timing statistics are accounted for. Only
6736 : * continue if a log message is to be written.
6737 : */
6738 3428 : if (!log_checkpoints)
6739 584 : return;
6740 :
6741 2844 : total_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_start_t,
6742 : CheckpointStats.ckpt_end_t);
6743 :
6744 : /*
6745 : * Timing values returned from CheckpointStats are in microseconds.
6746 : * Convert to milliseconds for consistent printing.
6747 : */
6748 2844 : longest_msecs = (long) ((CheckpointStats.ckpt_longest_sync + 999) / 1000);
6749 :
6750 2844 : average_sync_time = 0;
6751 2844 : if (CheckpointStats.ckpt_sync_rels > 0)
6752 0 : average_sync_time = CheckpointStats.ckpt_agg_sync_time /
6753 0 : CheckpointStats.ckpt_sync_rels;
6754 2844 : average_msecs = (long) ((average_sync_time + 999) / 1000);
6755 :
6756 : /*
6757 : * ControlFileLock is not required to see ControlFile->checkPoint and
6758 : * ->checkPointCopy here as we are the only updator of those variables at
6759 : * this moment.
6760 : */
6761 2844 : if (restartpoint)
6762 380 : ereport(LOG,
6763 : (errmsg("restartpoint complete: wrote %d buffers (%.1f%%), "
6764 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6765 : "%d removed, %d recycled; write=%ld.%03d s, "
6766 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6767 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6768 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6769 : CheckpointStats.ckpt_bufs_written,
6770 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6771 : CheckpointStats.ckpt_slru_written,
6772 : CheckpointStats.ckpt_segs_added,
6773 : CheckpointStats.ckpt_segs_removed,
6774 : CheckpointStats.ckpt_segs_recycled,
6775 : write_msecs / 1000, (int) (write_msecs % 1000),
6776 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6777 : total_msecs / 1000, (int) (total_msecs % 1000),
6778 : CheckpointStats.ckpt_sync_rels,
6779 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6780 : average_msecs / 1000, (int) (average_msecs % 1000),
6781 : (int) (PrevCheckPointDistance / 1024.0),
6782 : (int) (CheckPointDistanceEstimate / 1024.0),
6783 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6784 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6785 : else
6786 2464 : ereport(LOG,
6787 : (errmsg("checkpoint complete: wrote %d buffers (%.1f%%), "
6788 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6789 : "%d removed, %d recycled; write=%ld.%03d s, "
6790 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6791 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6792 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6793 : CheckpointStats.ckpt_bufs_written,
6794 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6795 : CheckpointStats.ckpt_slru_written,
6796 : CheckpointStats.ckpt_segs_added,
6797 : CheckpointStats.ckpt_segs_removed,
6798 : CheckpointStats.ckpt_segs_recycled,
6799 : write_msecs / 1000, (int) (write_msecs % 1000),
6800 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6801 : total_msecs / 1000, (int) (total_msecs % 1000),
6802 : CheckpointStats.ckpt_sync_rels,
6803 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6804 : average_msecs / 1000, (int) (average_msecs % 1000),
6805 : (int) (PrevCheckPointDistance / 1024.0),
6806 : (int) (CheckPointDistanceEstimate / 1024.0),
6807 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6808 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6809 : }
6810 :
6811 : /*
6812 : * Update the estimate of distance between checkpoints.
6813 : *
6814 : * The estimate is used to calculate the number of WAL segments to keep
6815 : * preallocated, see XLOGfileslop().
6816 : */
6817 : static void
6818 3428 : UpdateCheckPointDistanceEstimate(uint64 nbytes)
6819 : {
6820 : /*
6821 : * To estimate the number of segments consumed between checkpoints, keep a
6822 : * moving average of the amount of WAL generated in previous checkpoint
6823 : * cycles. However, if the load is bursty, with quiet periods and busy
6824 : * periods, we want to cater for the peak load. So instead of a plain
6825 : * moving average, let the average decline slowly if the previous cycle
6826 : * used less WAL than estimated, but bump it up immediately if it used
6827 : * more.
6828 : *
6829 : * When checkpoints are triggered by max_wal_size, this should converge to
6830 : * CheckpointSegments * wal_segment_size,
6831 : *
6832 : * Note: This doesn't pay any attention to what caused the checkpoint.
6833 : * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
6834 : * starting a base backup, are counted the same as those created
6835 : * automatically. The slow-decline will largely mask them out, if they are
6836 : * not frequent. If they are frequent, it seems reasonable to count them
6837 : * in as any others; if you issue a manual checkpoint every 5 minutes and
6838 : * never let a timed checkpoint happen, it makes sense to base the
6839 : * preallocation on that 5 minute interval rather than whatever
6840 : * checkpoint_timeout is set to.
6841 : */
6842 3428 : PrevCheckPointDistance = nbytes;
6843 3428 : if (CheckPointDistanceEstimate < nbytes)
6844 1458 : CheckPointDistanceEstimate = nbytes;
6845 : else
6846 1970 : CheckPointDistanceEstimate =
6847 1970 : (0.90 * CheckPointDistanceEstimate + 0.10 * (double) nbytes);
6848 3428 : }
6849 :
6850 : /*
6851 : * Update the ps display for a process running a checkpoint. Note that
6852 : * this routine should not do any allocations so as it can be called
6853 : * from a critical section.
6854 : */
6855 : static void
6856 6856 : update_checkpoint_display(int flags, bool restartpoint, bool reset)
6857 : {
6858 : /*
6859 : * The status is reported only for end-of-recovery and shutdown
6860 : * checkpoints or shutdown restartpoints. Updating the ps display is
6861 : * useful in those situations as it may not be possible to rely on
6862 : * pg_stat_activity to see the status of the checkpointer or the startup
6863 : * process.
6864 : */
6865 6856 : if ((flags & (CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IS_SHUTDOWN)) == 0)
6866 4288 : return;
6867 :
6868 2568 : if (reset)
6869 1284 : set_ps_display("");
6870 : else
6871 : {
6872 : char activitymsg[128];
6873 :
6874 3852 : snprintf(activitymsg, sizeof(activitymsg), "performing %s%s%s",
6875 1284 : (flags & CHECKPOINT_END_OF_RECOVERY) ? "end-of-recovery " : "",
6876 1284 : (flags & CHECKPOINT_IS_SHUTDOWN) ? "shutdown " : "",
6877 : restartpoint ? "restartpoint" : "checkpoint");
6878 1284 : set_ps_display(activitymsg);
6879 : }
6880 : }
6881 :
6882 :
6883 : /*
6884 : * Perform a checkpoint --- either during shutdown, or on-the-fly
6885 : *
6886 : * flags is a bitwise OR of the following:
6887 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
6888 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
6889 : * CHECKPOINT_FAST: finish the checkpoint ASAP, ignoring
6890 : * checkpoint_completion_target parameter.
6891 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
6892 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
6893 : * CHECKPOINT_END_OF_RECOVERY).
6894 : * CHECKPOINT_FLUSH_UNLOGGED: also flush buffers of unlogged tables.
6895 : *
6896 : * Note: flags contains other bits, of interest here only for logging purposes.
6897 : * In particular note that this routine is synchronous and does not pay
6898 : * attention to CHECKPOINT_WAIT.
6899 : *
6900 : * If !shutdown then we are writing an online checkpoint. An XLOG_CHECKPOINT_REDO
6901 : * record is inserted into WAL at the logical location of the checkpoint, before
6902 : * flushing anything to disk, and when the checkpoint is eventually completed,
6903 : * and it is from this point that WAL replay will begin in the case of a recovery
6904 : * from this checkpoint. Once everything is written to disk, an
6905 : * XLOG_CHECKPOINT_ONLINE record is written to complete the checkpoint, and
6906 : * points back to the earlier XLOG_CHECKPOINT_REDO record. This mechanism allows
6907 : * other write-ahead log records to be written while the checkpoint is in
6908 : * progress, but we must be very careful about order of operations. This function
6909 : * may take many minutes to execute on a busy system.
6910 : *
6911 : * On the other hand, when shutdown is true, concurrent insertion into the
6912 : * write-ahead log is impossible, so there is no need for two separate records.
6913 : * In this case, we only insert an XLOG_CHECKPOINT_SHUTDOWN record, and it's
6914 : * both the record marking the completion of the checkpoint and the location
6915 : * from which WAL replay would begin if needed.
6916 : *
6917 : * Returns true if a new checkpoint was performed, or false if it was skipped
6918 : * because the system was idle.
6919 : */
6920 : bool
6921 3052 : CreateCheckPoint(int flags)
6922 : {
6923 : bool shutdown;
6924 : CheckPoint checkPoint;
6925 : XLogRecPtr recptr;
6926 : XLogSegNo _logSegNo;
6927 3052 : XLogCtlInsert *Insert = &XLogCtl->Insert;
6928 : uint32 freespace;
6929 : XLogRecPtr PriorRedoPtr;
6930 : XLogRecPtr last_important_lsn;
6931 : VirtualTransactionId *vxids;
6932 : int nvxids;
6933 3052 : int oldXLogAllowed = 0;
6934 :
6935 : /*
6936 : * An end-of-recovery checkpoint is really a shutdown checkpoint, just
6937 : * issued at a different time.
6938 : */
6939 3052 : if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
6940 1242 : shutdown = true;
6941 : else
6942 1810 : shutdown = false;
6943 :
6944 : /* sanity check */
6945 3052 : if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
6946 0 : elog(ERROR, "can't create a checkpoint during recovery");
6947 :
6948 : /*
6949 : * Prepare to accumulate statistics.
6950 : *
6951 : * Note: because it is possible for log_checkpoints to change while a
6952 : * checkpoint proceeds, we always accumulate stats, even if
6953 : * log_checkpoints is currently off.
6954 : */
6955 33572 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
6956 3052 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
6957 :
6958 : /*
6959 : * Let smgr prepare for checkpoint; this has to happen outside the
6960 : * critical section and before we determine the REDO pointer. Note that
6961 : * smgr must not do anything that'd have to be undone if we decide no
6962 : * checkpoint is needed.
6963 : */
6964 3052 : SyncPreCheckpoint();
6965 :
6966 : /*
6967 : * Use a critical section to force system panic if we have trouble.
6968 : */
6969 3052 : START_CRIT_SECTION();
6970 :
6971 3052 : if (shutdown)
6972 : {
6973 1242 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6974 1242 : ControlFile->state = DB_SHUTDOWNING;
6975 1242 : UpdateControlFile();
6976 1242 : LWLockRelease(ControlFileLock);
6977 : }
6978 :
6979 : /* Begin filling in the checkpoint WAL record */
6980 36624 : MemSet(&checkPoint, 0, sizeof(checkPoint));
6981 3052 : checkPoint.time = (pg_time_t) time(NULL);
6982 :
6983 : /*
6984 : * For Hot Standby, derive the oldestActiveXid before we fix the redo
6985 : * pointer. This allows us to begin accumulating changes to assemble our
6986 : * starting snapshot of locks and transactions.
6987 : */
6988 3052 : if (!shutdown && XLogStandbyInfoActive())
6989 1712 : checkPoint.oldestActiveXid = GetOldestActiveTransactionId(false, true);
6990 : else
6991 1340 : checkPoint.oldestActiveXid = InvalidTransactionId;
6992 :
6993 : /*
6994 : * Get location of last important record before acquiring insert locks (as
6995 : * GetLastImportantRecPtr() also locks WAL locks).
6996 : */
6997 3052 : last_important_lsn = GetLastImportantRecPtr();
6998 :
6999 : /*
7000 : * If this isn't a shutdown or forced checkpoint, and if there has been no
7001 : * WAL activity requiring a checkpoint, skip it. The idea here is to
7002 : * avoid inserting duplicate checkpoints when the system is idle.
7003 : */
7004 3052 : if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
7005 : CHECKPOINT_FORCE)) == 0)
7006 : {
7007 378 : if (last_important_lsn == ControlFile->checkPoint)
7008 : {
7009 4 : END_CRIT_SECTION();
7010 4 : ereport(DEBUG1,
7011 : (errmsg_internal("checkpoint skipped because system is idle")));
7012 4 : return false;
7013 : }
7014 : }
7015 :
7016 : /*
7017 : * An end-of-recovery checkpoint is created before anyone is allowed to
7018 : * write WAL. To allow us to write the checkpoint record, temporarily
7019 : * enable XLogInsertAllowed.
7020 : */
7021 3048 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7022 58 : oldXLogAllowed = LocalSetXLogInsertAllowed();
7023 :
7024 3048 : checkPoint.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7025 3048 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7026 58 : checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7027 : else
7028 2990 : checkPoint.PrevTimeLineID = checkPoint.ThisTimeLineID;
7029 :
7030 : /*
7031 : * We must block concurrent insertions while examining insert state.
7032 : */
7033 3048 : WALInsertLockAcquireExclusive();
7034 :
7035 3048 : checkPoint.fullPageWrites = Insert->fullPageWrites;
7036 3048 : checkPoint.wal_level = wal_level;
7037 :
7038 3048 : if (shutdown)
7039 : {
7040 1242 : XLogRecPtr curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
7041 :
7042 : /*
7043 : * Compute new REDO record ptr = location of next XLOG record.
7044 : *
7045 : * Since this is a shutdown checkpoint, there can't be any concurrent
7046 : * WAL insertion.
7047 : */
7048 1242 : freespace = INSERT_FREESPACE(curInsert);
7049 1242 : if (freespace == 0)
7050 : {
7051 0 : if (XLogSegmentOffset(curInsert, wal_segment_size) == 0)
7052 0 : curInsert += SizeOfXLogLongPHD;
7053 : else
7054 0 : curInsert += SizeOfXLogShortPHD;
7055 : }
7056 1242 : checkPoint.redo = curInsert;
7057 :
7058 : /*
7059 : * Here we update the shared RedoRecPtr for future XLogInsert calls;
7060 : * this must be done while holding all the insertion locks.
7061 : *
7062 : * Note: if we fail to complete the checkpoint, RedoRecPtr will be
7063 : * left pointing past where it really needs to point. This is okay;
7064 : * the only consequence is that XLogInsert might back up whole buffers
7065 : * that it didn't really need to. We can't postpone advancing
7066 : * RedoRecPtr because XLogInserts that happen while we are dumping
7067 : * buffers must assume that their buffer changes are not included in
7068 : * the checkpoint.
7069 : */
7070 1242 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
7071 : }
7072 :
7073 : /*
7074 : * Now we can release the WAL insertion locks, allowing other xacts to
7075 : * proceed while we are flushing disk buffers.
7076 : */
7077 3048 : WALInsertLockRelease();
7078 :
7079 : /*
7080 : * If this is an online checkpoint, we have not yet determined the redo
7081 : * point. We do so now by inserting the special XLOG_CHECKPOINT_REDO
7082 : * record; the LSN at which it starts becomes the new redo pointer. We
7083 : * don't do this for a shutdown checkpoint, because in that case no WAL
7084 : * can be written between the redo point and the insertion of the
7085 : * checkpoint record itself, so the checkpoint record itself serves to
7086 : * mark the redo point.
7087 : */
7088 3048 : if (!shutdown)
7089 : {
7090 : /* Include WAL level in record for WAL summarizer's benefit. */
7091 1806 : XLogBeginInsert();
7092 1806 : XLogRegisterData(&wal_level, sizeof(wal_level));
7093 1806 : (void) XLogInsert(RM_XLOG_ID, XLOG_CHECKPOINT_REDO);
7094 :
7095 : /*
7096 : * XLogInsertRecord will have updated XLogCtl->Insert.RedoRecPtr in
7097 : * shared memory and RedoRecPtr in backend-local memory, but we need
7098 : * to copy that into the record that will be inserted when the
7099 : * checkpoint is complete.
7100 : */
7101 1806 : checkPoint.redo = RedoRecPtr;
7102 : }
7103 :
7104 : /* Update the info_lck-protected copy of RedoRecPtr as well */
7105 3048 : SpinLockAcquire(&XLogCtl->info_lck);
7106 3048 : XLogCtl->RedoRecPtr = checkPoint.redo;
7107 3048 : SpinLockRelease(&XLogCtl->info_lck);
7108 :
7109 : /*
7110 : * If enabled, log checkpoint start. We postpone this until now so as not
7111 : * to log anything if we decided to skip the checkpoint.
7112 : */
7113 3048 : if (log_checkpoints)
7114 2464 : LogCheckpointStart(flags, false);
7115 :
7116 : /* Update the process title */
7117 3048 : update_checkpoint_display(flags, false, false);
7118 :
7119 : TRACE_POSTGRESQL_CHECKPOINT_START(flags);
7120 :
7121 : /*
7122 : * Get the other info we need for the checkpoint record.
7123 : *
7124 : * We don't need to save oldestClogXid in the checkpoint, it only matters
7125 : * for the short period in which clog is being truncated, and if we crash
7126 : * during that we'll redo the clog truncation and fix up oldestClogXid
7127 : * there.
7128 : */
7129 3048 : LWLockAcquire(XidGenLock, LW_SHARED);
7130 3048 : checkPoint.nextXid = TransamVariables->nextXid;
7131 3048 : checkPoint.oldestXid = TransamVariables->oldestXid;
7132 3048 : checkPoint.oldestXidDB = TransamVariables->oldestXidDB;
7133 3048 : LWLockRelease(XidGenLock);
7134 :
7135 3048 : LWLockAcquire(CommitTsLock, LW_SHARED);
7136 3048 : checkPoint.oldestCommitTsXid = TransamVariables->oldestCommitTsXid;
7137 3048 : checkPoint.newestCommitTsXid = TransamVariables->newestCommitTsXid;
7138 3048 : LWLockRelease(CommitTsLock);
7139 :
7140 3048 : LWLockAcquire(OidGenLock, LW_SHARED);
7141 3048 : checkPoint.nextOid = TransamVariables->nextOid;
7142 3048 : if (!shutdown)
7143 1806 : checkPoint.nextOid += TransamVariables->oidCount;
7144 3048 : LWLockRelease(OidGenLock);
7145 :
7146 3048 : MultiXactGetCheckptMulti(shutdown,
7147 : &checkPoint.nextMulti,
7148 : &checkPoint.nextMultiOffset,
7149 : &checkPoint.oldestMulti,
7150 : &checkPoint.oldestMultiDB);
7151 :
7152 : /*
7153 : * Having constructed the checkpoint record, ensure all shmem disk buffers
7154 : * and commit-log buffers are flushed to disk.
7155 : *
7156 : * This I/O could fail for various reasons. If so, we will fail to
7157 : * complete the checkpoint, but there is no reason to force a system
7158 : * panic. Accordingly, exit critical section while doing it.
7159 : */
7160 3048 : END_CRIT_SECTION();
7161 :
7162 : /*
7163 : * In some cases there are groups of actions that must all occur on one
7164 : * side or the other of a checkpoint record. Before flushing the
7165 : * checkpoint record we must explicitly wait for any backend currently
7166 : * performing those groups of actions.
7167 : *
7168 : * One example is end of transaction, so we must wait for any transactions
7169 : * that are currently in commit critical sections. If an xact inserted
7170 : * its commit record into XLOG just before the REDO point, then a crash
7171 : * restart from the REDO point would not replay that record, which means
7172 : * that our flushing had better include the xact's update of pg_xact. So
7173 : * we wait till he's out of his commit critical section before proceeding.
7174 : * See notes in RecordTransactionCommit().
7175 : *
7176 : * Because we've already released the insertion locks, this test is a bit
7177 : * fuzzy: it is possible that we will wait for xacts we didn't really need
7178 : * to wait for. But the delay should be short and it seems better to make
7179 : * checkpoint take a bit longer than to hold off insertions longer than
7180 : * necessary. (In fact, the whole reason we have this issue is that xact.c
7181 : * does commit record XLOG insertion and clog update as two separate steps
7182 : * protected by different locks, but again that seems best on grounds of
7183 : * minimizing lock contention.)
7184 : *
7185 : * A transaction that has not yet set delayChkptFlags when we look cannot
7186 : * be at risk, since it has not inserted its commit record yet; and one
7187 : * that's already cleared it is not at risk either, since it's done fixing
7188 : * clog and we will correctly flush the update below. So we cannot miss
7189 : * any xacts we need to wait for.
7190 : */
7191 3048 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
7192 3048 : if (nvxids > 0)
7193 : {
7194 : do
7195 : {
7196 : /*
7197 : * Keep absorbing fsync requests while we wait. There could even
7198 : * be a deadlock if we don't, if the process that prevents the
7199 : * checkpoint is trying to add a request to the queue.
7200 : */
7201 106 : AbsorbSyncRequests();
7202 :
7203 106 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
7204 106 : pg_usleep(10000L); /* wait for 10 msec */
7205 106 : pgstat_report_wait_end();
7206 106 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7207 : DELAY_CHKPT_START));
7208 : }
7209 3048 : pfree(vxids);
7210 :
7211 3048 : CheckPointGuts(checkPoint.redo, flags);
7212 :
7213 3048 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
7214 3048 : if (nvxids > 0)
7215 : {
7216 : do
7217 : {
7218 0 : AbsorbSyncRequests();
7219 :
7220 0 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_COMPLETE);
7221 0 : pg_usleep(10000L); /* wait for 10 msec */
7222 0 : pgstat_report_wait_end();
7223 0 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7224 : DELAY_CHKPT_COMPLETE));
7225 : }
7226 3048 : pfree(vxids);
7227 :
7228 : /*
7229 : * Take a snapshot of running transactions and write this to WAL. This
7230 : * allows us to reconstruct the state of running transactions during
7231 : * archive recovery, if required. Skip, if this info disabled.
7232 : *
7233 : * If we are shutting down, or Startup process is completing crash
7234 : * recovery we don't need to write running xact data.
7235 : */
7236 3048 : if (!shutdown && XLogStandbyInfoActive())
7237 1708 : LogStandbySnapshot();
7238 :
7239 3048 : START_CRIT_SECTION();
7240 :
7241 : /*
7242 : * Now insert the checkpoint record into XLOG.
7243 : */
7244 3048 : XLogBeginInsert();
7245 3048 : XLogRegisterData(&checkPoint, sizeof(checkPoint));
7246 3048 : recptr = XLogInsert(RM_XLOG_ID,
7247 : shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
7248 : XLOG_CHECKPOINT_ONLINE);
7249 :
7250 3048 : XLogFlush(recptr);
7251 :
7252 : /*
7253 : * We mustn't write any new WAL after a shutdown checkpoint, or it will be
7254 : * overwritten at next startup. No-one should even try, this just allows
7255 : * sanity-checking. In the case of an end-of-recovery checkpoint, we want
7256 : * to just temporarily disable writing until the system has exited
7257 : * recovery.
7258 : */
7259 3048 : if (shutdown)
7260 : {
7261 1242 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7262 58 : LocalXLogInsertAllowed = oldXLogAllowed;
7263 : else
7264 1184 : LocalXLogInsertAllowed = 0; /* never again write WAL */
7265 : }
7266 :
7267 : /*
7268 : * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
7269 : * = end of actual checkpoint record.
7270 : */
7271 3048 : if (shutdown && checkPoint.redo != ProcLastRecPtr)
7272 0 : ereport(PANIC,
7273 : (errmsg("concurrent write-ahead log activity while database system is shutting down")));
7274 :
7275 : /*
7276 : * Remember the prior checkpoint's redo ptr for
7277 : * UpdateCheckPointDistanceEstimate()
7278 : */
7279 3048 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7280 :
7281 : /*
7282 : * Update the control file.
7283 : */
7284 3048 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7285 3048 : if (shutdown)
7286 1242 : ControlFile->state = DB_SHUTDOWNED;
7287 3048 : ControlFile->checkPoint = ProcLastRecPtr;
7288 3048 : ControlFile->checkPointCopy = checkPoint;
7289 : /* crash recovery should always recover to the end of WAL */
7290 3048 : ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
7291 3048 : ControlFile->minRecoveryPointTLI = 0;
7292 :
7293 : /*
7294 : * Persist unloggedLSN value. It's reset on crash recovery, so this goes
7295 : * unused on non-shutdown checkpoints, but seems useful to store it always
7296 : * for debugging purposes.
7297 : */
7298 3048 : ControlFile->unloggedLSN = pg_atomic_read_membarrier_u64(&XLogCtl->unloggedLSN);
7299 :
7300 3048 : UpdateControlFile();
7301 3048 : LWLockRelease(ControlFileLock);
7302 :
7303 : /*
7304 : * We are now done with critical updates; no need for system panic if we
7305 : * have trouble while fooling with old log segments.
7306 : */
7307 3048 : END_CRIT_SECTION();
7308 :
7309 : /*
7310 : * WAL summaries end when the next XLOG_CHECKPOINT_REDO or
7311 : * XLOG_CHECKPOINT_SHUTDOWN record is reached. This is the first point
7312 : * where (a) we're not inside of a critical section and (b) we can be
7313 : * certain that the relevant record has been flushed to disk, which must
7314 : * happen before it can be summarized.
7315 : *
7316 : * If this is a shutdown checkpoint, then this happens reasonably
7317 : * promptly: we've only just inserted and flushed the
7318 : * XLOG_CHECKPOINT_SHUTDOWN record. If this is not a shutdown checkpoint,
7319 : * then this might not be very prompt at all: the XLOG_CHECKPOINT_REDO
7320 : * record was written before we began flushing data to disk, and that
7321 : * could be many minutes ago at this point. However, we don't XLogFlush()
7322 : * after inserting that record, so we're not guaranteed that it's on disk
7323 : * until after the above call that flushes the XLOG_CHECKPOINT_ONLINE
7324 : * record.
7325 : */
7326 3048 : WakeupWalSummarizer();
7327 :
7328 : /*
7329 : * Let smgr do post-checkpoint cleanup (eg, deleting old files).
7330 : */
7331 3048 : SyncPostCheckpoint();
7332 :
7333 : /*
7334 : * Update the average distance between checkpoints if the prior checkpoint
7335 : * exists.
7336 : */
7337 3048 : if (PriorRedoPtr != InvalidXLogRecPtr)
7338 3048 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7339 :
7340 : #ifdef USE_INJECTION_POINTS
7341 3048 : INJECTION_POINT("checkpoint-before-old-wal-removal", NULL);
7342 : #endif
7343 :
7344 : /*
7345 : * Delete old log files, those no longer needed for last checkpoint to
7346 : * prevent the disk holding the xlog from growing full.
7347 : */
7348 3048 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7349 3048 : KeepLogSeg(recptr, &_logSegNo);
7350 3048 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7351 : _logSegNo, InvalidOid,
7352 : InvalidTransactionId))
7353 : {
7354 : /*
7355 : * Some slots have been invalidated; recalculate the old-segment
7356 : * horizon, starting again from RedoRecPtr.
7357 : */
7358 6 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7359 6 : KeepLogSeg(recptr, &_logSegNo);
7360 : }
7361 3048 : _logSegNo--;
7362 3048 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
7363 : checkPoint.ThisTimeLineID);
7364 :
7365 : /*
7366 : * Make more log segments if needed. (Do this after recycling old log
7367 : * segments, since that may supply some of the needed files.)
7368 : */
7369 3048 : if (!shutdown)
7370 1806 : PreallocXlogFiles(recptr, checkPoint.ThisTimeLineID);
7371 :
7372 : /*
7373 : * Truncate pg_subtrans if possible. We can throw away all data before
7374 : * the oldest XMIN of any running transaction. No future transaction will
7375 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7376 : * in subtrans.c). During recovery, though, we mustn't do this because
7377 : * StartupSUBTRANS hasn't been called yet.
7378 : */
7379 3048 : if (!RecoveryInProgress())
7380 2990 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7381 :
7382 : /* Real work is done; log and update stats. */
7383 3048 : LogCheckpointEnd(false);
7384 :
7385 : /* Reset the process title */
7386 3048 : update_checkpoint_display(flags, false, true);
7387 :
7388 : TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
7389 : NBuffers,
7390 : CheckpointStats.ckpt_segs_added,
7391 : CheckpointStats.ckpt_segs_removed,
7392 : CheckpointStats.ckpt_segs_recycled);
7393 :
7394 3048 : return true;
7395 : }
7396 :
7397 : /*
7398 : * Mark the end of recovery in WAL though without running a full checkpoint.
7399 : * We can expect that a restartpoint is likely to be in progress as we
7400 : * do this, though we are unwilling to wait for it to complete.
7401 : *
7402 : * CreateRestartPoint() allows for the case where recovery may end before
7403 : * the restartpoint completes so there is no concern of concurrent behaviour.
7404 : */
7405 : static void
7406 84 : CreateEndOfRecoveryRecord(void)
7407 : {
7408 : xl_end_of_recovery xlrec;
7409 : XLogRecPtr recptr;
7410 :
7411 : /* sanity check */
7412 84 : if (!RecoveryInProgress())
7413 0 : elog(ERROR, "can only be used to end recovery");
7414 :
7415 84 : xlrec.end_time = GetCurrentTimestamp();
7416 84 : xlrec.wal_level = wal_level;
7417 :
7418 84 : WALInsertLockAcquireExclusive();
7419 84 : xlrec.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7420 84 : xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7421 84 : WALInsertLockRelease();
7422 :
7423 84 : START_CRIT_SECTION();
7424 :
7425 84 : XLogBeginInsert();
7426 84 : XLogRegisterData(&xlrec, sizeof(xl_end_of_recovery));
7427 84 : recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY);
7428 :
7429 84 : XLogFlush(recptr);
7430 :
7431 : /*
7432 : * Update the control file so that crash recovery can follow the timeline
7433 : * changes to this point.
7434 : */
7435 84 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7436 84 : ControlFile->minRecoveryPoint = recptr;
7437 84 : ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
7438 84 : UpdateControlFile();
7439 84 : LWLockRelease(ControlFileLock);
7440 :
7441 84 : END_CRIT_SECTION();
7442 84 : }
7443 :
7444 : /*
7445 : * Write an OVERWRITE_CONTRECORD message.
7446 : *
7447 : * When on WAL replay we expect a continuation record at the start of a page
7448 : * that is not there, recovery ends and WAL writing resumes at that point.
7449 : * But it's wrong to resume writing new WAL back at the start of the record
7450 : * that was broken, because downstream consumers of that WAL (physical
7451 : * replicas) are not prepared to "rewind". So the first action after
7452 : * finishing replay of all valid WAL must be to write a record of this type
7453 : * at the point where the contrecord was missing; to support xlogreader
7454 : * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
7455 : * to the page header where the record occurs. xlogreader has an ad-hoc
7456 : * mechanism to report metadata about the broken record, which is what we
7457 : * use here.
7458 : *
7459 : * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
7460 : * skip the record it was reading, and pass back the LSN of the skipped
7461 : * record, so that its caller can verify (on "replay" of that record) that the
7462 : * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
7463 : *
7464 : * 'aborted_lsn' is the beginning position of the record that was incomplete.
7465 : * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
7466 : * beginning of the XLOG page where the record is to be inserted. They must
7467 : * match the current WAL insert position, they're passed here just so that we
7468 : * can verify that.
7469 : */
7470 : static XLogRecPtr
7471 22 : CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn, XLogRecPtr pagePtr,
7472 : TimeLineID newTLI)
7473 : {
7474 : xl_overwrite_contrecord xlrec;
7475 : XLogRecPtr recptr;
7476 : XLogPageHeader pagehdr;
7477 : XLogRecPtr startPos;
7478 :
7479 : /* sanity checks */
7480 22 : if (!RecoveryInProgress())
7481 0 : elog(ERROR, "can only be used at end of recovery");
7482 22 : if (pagePtr % XLOG_BLCKSZ != 0)
7483 0 : elog(ERROR, "invalid position for missing continuation record %X/%08X",
7484 : LSN_FORMAT_ARGS(pagePtr));
7485 :
7486 : /* The current WAL insert position should be right after the page header */
7487 22 : startPos = pagePtr;
7488 22 : if (XLogSegmentOffset(startPos, wal_segment_size) == 0)
7489 2 : startPos += SizeOfXLogLongPHD;
7490 : else
7491 20 : startPos += SizeOfXLogShortPHD;
7492 22 : recptr = GetXLogInsertRecPtr();
7493 22 : if (recptr != startPos)
7494 0 : elog(ERROR, "invalid WAL insert position %X/%08X for OVERWRITE_CONTRECORD",
7495 : LSN_FORMAT_ARGS(recptr));
7496 :
7497 22 : START_CRIT_SECTION();
7498 :
7499 : /*
7500 : * Initialize the XLOG page header (by GetXLogBuffer), and set the
7501 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
7502 : *
7503 : * No other backend is allowed to write WAL yet, so acquiring the WAL
7504 : * insertion lock is just pro forma.
7505 : */
7506 22 : WALInsertLockAcquire();
7507 22 : pagehdr = (XLogPageHeader) GetXLogBuffer(pagePtr, newTLI);
7508 22 : pagehdr->xlp_info |= XLP_FIRST_IS_OVERWRITE_CONTRECORD;
7509 22 : WALInsertLockRelease();
7510 :
7511 : /*
7512 : * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
7513 : * page. We know it becomes the first record, because no other backend is
7514 : * allowed to write WAL yet.
7515 : */
7516 22 : XLogBeginInsert();
7517 22 : xlrec.overwritten_lsn = aborted_lsn;
7518 22 : xlrec.overwrite_time = GetCurrentTimestamp();
7519 22 : XLogRegisterData(&xlrec, sizeof(xl_overwrite_contrecord));
7520 22 : recptr = XLogInsert(RM_XLOG_ID, XLOG_OVERWRITE_CONTRECORD);
7521 :
7522 : /* check that the record was inserted to the right place */
7523 22 : if (ProcLastRecPtr != startPos)
7524 0 : elog(ERROR, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%08X",
7525 : LSN_FORMAT_ARGS(ProcLastRecPtr));
7526 :
7527 22 : XLogFlush(recptr);
7528 :
7529 22 : END_CRIT_SECTION();
7530 :
7531 22 : return recptr;
7532 : }
7533 :
7534 : /*
7535 : * Flush all data in shared memory to disk, and fsync
7536 : *
7537 : * This is the common code shared between regular checkpoints and
7538 : * recovery restartpoints.
7539 : */
7540 : static void
7541 3428 : CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
7542 : {
7543 3428 : CheckPointRelationMap();
7544 3428 : CheckPointReplicationSlots(flags & CHECKPOINT_IS_SHUTDOWN);
7545 3428 : CheckPointSnapBuild();
7546 3428 : CheckPointLogicalRewriteHeap();
7547 3428 : CheckPointReplicationOrigin();
7548 :
7549 : /* Write out all dirty data in SLRUs and the main buffer pool */
7550 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags);
7551 3428 : CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
7552 3428 : CheckPointCLOG();
7553 3428 : CheckPointCommitTs();
7554 3428 : CheckPointSUBTRANS();
7555 3428 : CheckPointMultiXact();
7556 3428 : CheckPointPredicate();
7557 3428 : CheckPointBuffers(flags);
7558 :
7559 : /* Perform all queued up fsyncs */
7560 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
7561 3428 : CheckpointStats.ckpt_sync_t = GetCurrentTimestamp();
7562 3428 : ProcessSyncRequests();
7563 3428 : CheckpointStats.ckpt_sync_end_t = GetCurrentTimestamp();
7564 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
7565 :
7566 : /* We deliberately delay 2PC checkpointing as long as possible */
7567 3428 : CheckPointTwoPhase(checkPointRedo);
7568 3428 : }
7569 :
7570 : /*
7571 : * Save a checkpoint for recovery restart if appropriate
7572 : *
7573 : * This function is called each time a checkpoint record is read from XLOG.
7574 : * It must determine whether the checkpoint represents a safe restartpoint or
7575 : * not. If so, the checkpoint record is stashed in shared memory so that
7576 : * CreateRestartPoint can consult it. (Note that the latter function is
7577 : * executed by the checkpointer, while this one will be executed by the
7578 : * startup process.)
7579 : */
7580 : static void
7581 1416 : RecoveryRestartPoint(const CheckPoint *checkPoint, XLogReaderState *record)
7582 : {
7583 : /*
7584 : * Also refrain from creating a restartpoint if we have seen any
7585 : * references to non-existent pages. Restarting recovery from the
7586 : * restartpoint would not see the references, so we would lose the
7587 : * cross-check that the pages belonged to a relation that was dropped
7588 : * later.
7589 : */
7590 1416 : if (XLogHaveInvalidPages())
7591 : {
7592 0 : elog(DEBUG2,
7593 : "could not record restart point at %X/%08X because there are unresolved references to invalid pages",
7594 : LSN_FORMAT_ARGS(checkPoint->redo));
7595 0 : return;
7596 : }
7597 :
7598 : /*
7599 : * Copy the checkpoint record to shared memory, so that checkpointer can
7600 : * work out the next time it wants to perform a restartpoint.
7601 : */
7602 1416 : SpinLockAcquire(&XLogCtl->info_lck);
7603 1416 : XLogCtl->lastCheckPointRecPtr = record->ReadRecPtr;
7604 1416 : XLogCtl->lastCheckPointEndPtr = record->EndRecPtr;
7605 1416 : XLogCtl->lastCheckPoint = *checkPoint;
7606 1416 : SpinLockRelease(&XLogCtl->info_lck);
7607 : }
7608 :
7609 : /*
7610 : * Establish a restartpoint if possible.
7611 : *
7612 : * This is similar to CreateCheckPoint, but is used during WAL recovery
7613 : * to establish a point from which recovery can roll forward without
7614 : * replaying the entire recovery log.
7615 : *
7616 : * Returns true if a new restartpoint was established. We can only establish
7617 : * a restartpoint if we have replayed a safe checkpoint record since last
7618 : * restartpoint.
7619 : */
7620 : bool
7621 1174 : CreateRestartPoint(int flags)
7622 : {
7623 : XLogRecPtr lastCheckPointRecPtr;
7624 : XLogRecPtr lastCheckPointEndPtr;
7625 : CheckPoint lastCheckPoint;
7626 : XLogRecPtr PriorRedoPtr;
7627 : XLogRecPtr receivePtr;
7628 : XLogRecPtr replayPtr;
7629 : TimeLineID replayTLI;
7630 : XLogRecPtr endptr;
7631 : XLogSegNo _logSegNo;
7632 : TimestampTz xtime;
7633 :
7634 : /* Concurrent checkpoint/restartpoint cannot happen */
7635 : Assert(!IsUnderPostmaster || MyBackendType == B_CHECKPOINTER);
7636 :
7637 : /* Get a local copy of the last safe checkpoint record. */
7638 1174 : SpinLockAcquire(&XLogCtl->info_lck);
7639 1174 : lastCheckPointRecPtr = XLogCtl->lastCheckPointRecPtr;
7640 1174 : lastCheckPointEndPtr = XLogCtl->lastCheckPointEndPtr;
7641 1174 : lastCheckPoint = XLogCtl->lastCheckPoint;
7642 1174 : SpinLockRelease(&XLogCtl->info_lck);
7643 :
7644 : /*
7645 : * Check that we're still in recovery mode. It's ok if we exit recovery
7646 : * mode after this check, the restart point is valid anyway.
7647 : */
7648 1174 : if (!RecoveryInProgress())
7649 : {
7650 0 : ereport(DEBUG2,
7651 : (errmsg_internal("skipping restartpoint, recovery has already ended")));
7652 0 : return false;
7653 : }
7654 :
7655 : /*
7656 : * If the last checkpoint record we've replayed is already our last
7657 : * restartpoint, we can't perform a new restart point. We still update
7658 : * minRecoveryPoint in that case, so that if this is a shutdown restart
7659 : * point, we won't start up earlier than before. That's not strictly
7660 : * necessary, but when hot standby is enabled, it would be rather weird if
7661 : * the database opened up for read-only connections at a point-in-time
7662 : * before the last shutdown. Such time travel is still possible in case of
7663 : * immediate shutdown, though.
7664 : *
7665 : * We don't explicitly advance minRecoveryPoint when we do create a
7666 : * restartpoint. It's assumed that flushing the buffers will do that as a
7667 : * side-effect.
7668 : */
7669 1174 : if (XLogRecPtrIsInvalid(lastCheckPointRecPtr) ||
7670 510 : lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
7671 : {
7672 794 : ereport(DEBUG2,
7673 : errmsg_internal("skipping restartpoint, already performed at %X/%08X",
7674 : LSN_FORMAT_ARGS(lastCheckPoint.redo)));
7675 :
7676 794 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
7677 794 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7678 : {
7679 68 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7680 68 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7681 68 : UpdateControlFile();
7682 68 : LWLockRelease(ControlFileLock);
7683 : }
7684 794 : return false;
7685 : }
7686 :
7687 : /*
7688 : * Update the shared RedoRecPtr so that the startup process can calculate
7689 : * the number of segments replayed since last restartpoint, and request a
7690 : * restartpoint if it exceeds CheckPointSegments.
7691 : *
7692 : * Like in CreateCheckPoint(), hold off insertions to update it, although
7693 : * during recovery this is just pro forma, because no WAL insertions are
7694 : * happening.
7695 : */
7696 380 : WALInsertLockAcquireExclusive();
7697 380 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = lastCheckPoint.redo;
7698 380 : WALInsertLockRelease();
7699 :
7700 : /* Also update the info_lck-protected copy */
7701 380 : SpinLockAcquire(&XLogCtl->info_lck);
7702 380 : XLogCtl->RedoRecPtr = lastCheckPoint.redo;
7703 380 : SpinLockRelease(&XLogCtl->info_lck);
7704 :
7705 : /*
7706 : * Prepare to accumulate statistics.
7707 : *
7708 : * Note: because it is possible for log_checkpoints to change while a
7709 : * checkpoint proceeds, we always accumulate stats, even if
7710 : * log_checkpoints is currently off.
7711 : */
7712 4180 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7713 380 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7714 :
7715 380 : if (log_checkpoints)
7716 380 : LogCheckpointStart(flags, true);
7717 :
7718 : /* Update the process title */
7719 380 : update_checkpoint_display(flags, true, false);
7720 :
7721 380 : CheckPointGuts(lastCheckPoint.redo, flags);
7722 :
7723 : /*
7724 : * This location needs to be after CheckPointGuts() to ensure that some
7725 : * work has already happened during this checkpoint.
7726 : */
7727 380 : INJECTION_POINT("create-restart-point", NULL);
7728 :
7729 : /*
7730 : * Remember the prior checkpoint's redo ptr for
7731 : * UpdateCheckPointDistanceEstimate()
7732 : */
7733 380 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7734 :
7735 : /*
7736 : * Update pg_control, using current time. Check that it still shows an
7737 : * older checkpoint, else do nothing; this is a quick hack to make sure
7738 : * nothing really bad happens if somehow we get here after the
7739 : * end-of-recovery checkpoint.
7740 : */
7741 380 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7742 380 : if (ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
7743 : {
7744 : /*
7745 : * Update the checkpoint information. We do this even if the cluster
7746 : * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
7747 : * segments recycled below.
7748 : */
7749 380 : ControlFile->checkPoint = lastCheckPointRecPtr;
7750 380 : ControlFile->checkPointCopy = lastCheckPoint;
7751 :
7752 : /*
7753 : * Ensure minRecoveryPoint is past the checkpoint record and update it
7754 : * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
7755 : * this will have happened already while writing out dirty buffers,
7756 : * but not necessarily - e.g. because no buffers were dirtied. We do
7757 : * this because a backup performed in recovery uses minRecoveryPoint
7758 : * to determine which WAL files must be included in the backup, and
7759 : * the file (or files) containing the checkpoint record must be
7760 : * included, at a minimum. Note that for an ordinary restart of
7761 : * recovery there's no value in having the minimum recovery point any
7762 : * earlier than this anyway, because redo will begin just after the
7763 : * checkpoint record.
7764 : */
7765 380 : if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
7766 : {
7767 380 : if (ControlFile->minRecoveryPoint < lastCheckPointEndPtr)
7768 : {
7769 36 : ControlFile->minRecoveryPoint = lastCheckPointEndPtr;
7770 36 : ControlFile->minRecoveryPointTLI = lastCheckPoint.ThisTimeLineID;
7771 :
7772 : /* update local copy */
7773 36 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
7774 36 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
7775 : }
7776 380 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7777 42 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7778 : }
7779 380 : UpdateControlFile();
7780 : }
7781 380 : LWLockRelease(ControlFileLock);
7782 :
7783 : /*
7784 : * Update the average distance between checkpoints/restartpoints if the
7785 : * prior checkpoint exists.
7786 : */
7787 380 : if (PriorRedoPtr != InvalidXLogRecPtr)
7788 380 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7789 :
7790 : /*
7791 : * Delete old log files, those no longer needed for last restartpoint to
7792 : * prevent the disk holding the xlog from growing full.
7793 : */
7794 380 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7795 :
7796 : /*
7797 : * Retreat _logSegNo using the current end of xlog replayed or received,
7798 : * whichever is later.
7799 : */
7800 380 : receivePtr = GetWalRcvFlushRecPtr(NULL, NULL);
7801 380 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
7802 380 : endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
7803 380 : KeepLogSeg(endptr, &_logSegNo);
7804 380 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7805 : _logSegNo, InvalidOid,
7806 : InvalidTransactionId))
7807 : {
7808 : /*
7809 : * Some slots have been invalidated; recalculate the old-segment
7810 : * horizon, starting again from RedoRecPtr.
7811 : */
7812 2 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7813 2 : KeepLogSeg(endptr, &_logSegNo);
7814 : }
7815 380 : _logSegNo--;
7816 :
7817 : /*
7818 : * Try to recycle segments on a useful timeline. If we've been promoted
7819 : * since the beginning of this restartpoint, use the new timeline chosen
7820 : * at end of recovery. If we're still in recovery, use the timeline we're
7821 : * currently replaying.
7822 : *
7823 : * There is no guarantee that the WAL segments will be useful on the
7824 : * current timeline; if recovery proceeds to a new timeline right after
7825 : * this, the pre-allocated WAL segments on this timeline will not be used,
7826 : * and will go wasted until recycled on the next restartpoint. We'll live
7827 : * with that.
7828 : */
7829 380 : if (!RecoveryInProgress())
7830 0 : replayTLI = XLogCtl->InsertTimeLineID;
7831 :
7832 380 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, endptr, replayTLI);
7833 :
7834 : /*
7835 : * Make more log segments if needed. (Do this after recycling old log
7836 : * segments, since that may supply some of the needed files.)
7837 : */
7838 380 : PreallocXlogFiles(endptr, replayTLI);
7839 :
7840 : /*
7841 : * Truncate pg_subtrans if possible. We can throw away all data before
7842 : * the oldest XMIN of any running transaction. No future transaction will
7843 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7844 : * in subtrans.c). When hot standby is disabled, though, we mustn't do
7845 : * this because StartupSUBTRANS hasn't been called yet.
7846 : */
7847 380 : if (EnableHotStandby)
7848 380 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7849 :
7850 : /* Real work is done; log and update stats. */
7851 380 : LogCheckpointEnd(true);
7852 :
7853 : /* Reset the process title */
7854 380 : update_checkpoint_display(flags, true, true);
7855 :
7856 380 : xtime = GetLatestXTime();
7857 380 : ereport((log_checkpoints ? LOG : DEBUG2),
7858 : errmsg("recovery restart point at %X/%08X",
7859 : LSN_FORMAT_ARGS(lastCheckPoint.redo)),
7860 : xtime ? errdetail("Last completed transaction was at log time %s.",
7861 : timestamptz_to_str(xtime)) : 0);
7862 :
7863 : /*
7864 : * Finally, execute archive_cleanup_command, if any.
7865 : */
7866 380 : if (archiveCleanupCommand && strcmp(archiveCleanupCommand, "") != 0)
7867 0 : ExecuteRecoveryCommand(archiveCleanupCommand,
7868 : "archive_cleanup_command",
7869 : false,
7870 : WAIT_EVENT_ARCHIVE_CLEANUP_COMMAND);
7871 :
7872 380 : return true;
7873 : }
7874 :
7875 : /*
7876 : * Report availability of WAL for the given target LSN
7877 : * (typically a slot's restart_lsn)
7878 : *
7879 : * Returns one of the following enum values:
7880 : *
7881 : * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
7882 : * max_wal_size.
7883 : *
7884 : * * WALAVAIL_EXTENDED means it is still available by preserving extra
7885 : * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
7886 : * than max_wal_size, this state is not returned.
7887 : *
7888 : * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
7889 : * remove reserved segments. The walsender using this slot may return to the
7890 : * above.
7891 : *
7892 : * * WALAVAIL_REMOVED means it has been removed. A replication stream on
7893 : * a slot with this LSN cannot continue. (Any associated walsender
7894 : * processes should have been terminated already.)
7895 : *
7896 : * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
7897 : */
7898 : WALAvailability
7899 1028 : GetWALAvailability(XLogRecPtr targetLSN)
7900 : {
7901 : XLogRecPtr currpos; /* current write LSN */
7902 : XLogSegNo currSeg; /* segid of currpos */
7903 : XLogSegNo targetSeg; /* segid of targetLSN */
7904 : XLogSegNo oldestSeg; /* actual oldest segid */
7905 : XLogSegNo oldestSegMaxWalSize; /* oldest segid kept by max_wal_size */
7906 : XLogSegNo oldestSlotSeg; /* oldest segid kept by slot */
7907 : uint64 keepSegs;
7908 :
7909 : /*
7910 : * slot does not reserve WAL. Either deactivated, or has never been active
7911 : */
7912 1028 : if (XLogRecPtrIsInvalid(targetLSN))
7913 70 : return WALAVAIL_INVALID_LSN;
7914 :
7915 : /*
7916 : * Calculate the oldest segment currently reserved by all slots,
7917 : * considering wal_keep_size and max_slot_wal_keep_size. Initialize
7918 : * oldestSlotSeg to the current segment.
7919 : */
7920 958 : currpos = GetXLogWriteRecPtr();
7921 958 : XLByteToSeg(currpos, oldestSlotSeg, wal_segment_size);
7922 958 : KeepLogSeg(currpos, &oldestSlotSeg);
7923 :
7924 : /*
7925 : * Find the oldest extant segment file. We get 1 until checkpoint removes
7926 : * the first WAL segment file since startup, which causes the status being
7927 : * wrong under certain abnormal conditions but that doesn't actually harm.
7928 : */
7929 958 : oldestSeg = XLogGetLastRemovedSegno() + 1;
7930 :
7931 : /* calculate oldest segment by max_wal_size */
7932 958 : XLByteToSeg(currpos, currSeg, wal_segment_size);
7933 958 : keepSegs = ConvertToXSegs(max_wal_size_mb, wal_segment_size) + 1;
7934 :
7935 958 : if (currSeg > keepSegs)
7936 16 : oldestSegMaxWalSize = currSeg - keepSegs;
7937 : else
7938 942 : oldestSegMaxWalSize = 1;
7939 :
7940 : /* the segment we care about */
7941 958 : XLByteToSeg(targetLSN, targetSeg, wal_segment_size);
7942 :
7943 : /*
7944 : * No point in returning reserved or extended status values if the
7945 : * targetSeg is known to be lost.
7946 : */
7947 958 : if (targetSeg >= oldestSlotSeg)
7948 : {
7949 : /* show "reserved" when targetSeg is within max_wal_size */
7950 956 : if (targetSeg >= oldestSegMaxWalSize)
7951 952 : return WALAVAIL_RESERVED;
7952 :
7953 : /* being retained by slots exceeding max_wal_size */
7954 4 : return WALAVAIL_EXTENDED;
7955 : }
7956 :
7957 : /* WAL segments are no longer retained but haven't been removed yet */
7958 2 : if (targetSeg >= oldestSeg)
7959 2 : return WALAVAIL_UNRESERVED;
7960 :
7961 : /* Definitely lost */
7962 0 : return WALAVAIL_REMOVED;
7963 : }
7964 :
7965 :
7966 : /*
7967 : * Retreat *logSegNo to the last segment that we need to retain because of
7968 : * either wal_keep_size or replication slots.
7969 : *
7970 : * This is calculated by subtracting wal_keep_size from the given xlog
7971 : * location, recptr and by making sure that that result is below the
7972 : * requirement of replication slots. For the latter criterion we do consider
7973 : * the effects of max_slot_wal_keep_size: reserve at most that much space back
7974 : * from recptr.
7975 : *
7976 : * Note about replication slots: if this function calculates a value
7977 : * that's further ahead than what slots need reserved, then affected
7978 : * slots need to be invalidated and this function invoked again.
7979 : * XXX it might be a good idea to rewrite this function so that
7980 : * invalidation is optionally done here, instead.
7981 : */
7982 : static void
7983 4394 : KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
7984 : {
7985 : XLogSegNo currSegNo;
7986 : XLogSegNo segno;
7987 : XLogRecPtr keep;
7988 :
7989 4394 : XLByteToSeg(recptr, currSegNo, wal_segment_size);
7990 4394 : segno = currSegNo;
7991 :
7992 : /* Calculate how many segments are kept by slots. */
7993 4394 : keep = XLogGetReplicationSlotMinimumLSN();
7994 4394 : if (keep != InvalidXLogRecPtr && keep < recptr)
7995 : {
7996 1248 : XLByteToSeg(keep, segno, wal_segment_size);
7997 :
7998 : /*
7999 : * Account for max_slot_wal_keep_size to avoid keeping more than
8000 : * configured. However, don't do that during a binary upgrade: if
8001 : * slots were to be invalidated because of this, it would not be
8002 : * possible to preserve logical ones during the upgrade.
8003 : */
8004 1248 : if (max_slot_wal_keep_size_mb >= 0 && !IsBinaryUpgrade)
8005 : {
8006 : uint64 slot_keep_segs;
8007 :
8008 40 : slot_keep_segs =
8009 40 : ConvertToXSegs(max_slot_wal_keep_size_mb, wal_segment_size);
8010 :
8011 40 : if (currSegNo - segno > slot_keep_segs)
8012 10 : segno = currSegNo - slot_keep_segs;
8013 : }
8014 : }
8015 :
8016 : /*
8017 : * If WAL summarization is in use, don't remove WAL that has yet to be
8018 : * summarized.
8019 : */
8020 4394 : keep = GetOldestUnsummarizedLSN(NULL, NULL);
8021 4394 : if (keep != InvalidXLogRecPtr)
8022 : {
8023 : XLogSegNo unsummarized_segno;
8024 :
8025 4 : XLByteToSeg(keep, unsummarized_segno, wal_segment_size);
8026 4 : if (unsummarized_segno < segno)
8027 2 : segno = unsummarized_segno;
8028 : }
8029 :
8030 : /* but, keep at least wal_keep_size if that's set */
8031 4394 : if (wal_keep_size_mb > 0)
8032 : {
8033 : uint64 keep_segs;
8034 :
8035 138 : keep_segs = ConvertToXSegs(wal_keep_size_mb, wal_segment_size);
8036 138 : if (currSegNo - segno < keep_segs)
8037 : {
8038 : /* avoid underflow, don't go below 1 */
8039 138 : if (currSegNo <= keep_segs)
8040 130 : segno = 1;
8041 : else
8042 8 : segno = currSegNo - keep_segs;
8043 : }
8044 : }
8045 :
8046 : /* don't delete WAL segments newer than the calculated segment */
8047 4394 : if (segno < *logSegNo)
8048 710 : *logSegNo = segno;
8049 4394 : }
8050 :
8051 : /*
8052 : * Write a NEXTOID log record
8053 : */
8054 : void
8055 1206 : XLogPutNextOid(Oid nextOid)
8056 : {
8057 1206 : XLogBeginInsert();
8058 1206 : XLogRegisterData(&nextOid, sizeof(Oid));
8059 1206 : (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID);
8060 :
8061 : /*
8062 : * We need not flush the NEXTOID record immediately, because any of the
8063 : * just-allocated OIDs could only reach disk as part of a tuple insert or
8064 : * update that would have its own XLOG record that must follow the NEXTOID
8065 : * record. Therefore, the standard buffer LSN interlock applied to those
8066 : * records will ensure no such OID reaches disk before the NEXTOID record
8067 : * does.
8068 : *
8069 : * Note, however, that the above statement only covers state "within" the
8070 : * database. When we use a generated OID as a file or directory name, we
8071 : * are in a sense violating the basic WAL rule, because that filesystem
8072 : * change may reach disk before the NEXTOID WAL record does. The impact
8073 : * of this is that if a database crash occurs immediately afterward, we
8074 : * might after restart re-generate the same OID and find that it conflicts
8075 : * with the leftover file or directory. But since for safety's sake we
8076 : * always loop until finding a nonconflicting filename, this poses no real
8077 : * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8078 : */
8079 1206 : }
8080 :
8081 : /*
8082 : * Write an XLOG SWITCH record.
8083 : *
8084 : * Here we just blindly issue an XLogInsert request for the record.
8085 : * All the magic happens inside XLogInsert.
8086 : *
8087 : * The return value is either the end+1 address of the switch record,
8088 : * or the end+1 address of the prior segment if we did not need to
8089 : * write a switch record because we are already at segment start.
8090 : */
8091 : XLogRecPtr
8092 1550 : RequestXLogSwitch(bool mark_unimportant)
8093 : {
8094 : XLogRecPtr RecPtr;
8095 :
8096 : /* XLOG SWITCH has no data */
8097 1550 : XLogBeginInsert();
8098 :
8099 1550 : if (mark_unimportant)
8100 0 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8101 1550 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH);
8102 :
8103 1550 : return RecPtr;
8104 : }
8105 :
8106 : /*
8107 : * Write a RESTORE POINT record
8108 : */
8109 : XLogRecPtr
8110 6 : XLogRestorePoint(const char *rpName)
8111 : {
8112 : XLogRecPtr RecPtr;
8113 : xl_restore_point xlrec;
8114 :
8115 6 : xlrec.rp_time = GetCurrentTimestamp();
8116 6 : strlcpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8117 :
8118 6 : XLogBeginInsert();
8119 6 : XLogRegisterData(&xlrec, sizeof(xl_restore_point));
8120 :
8121 6 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT);
8122 :
8123 6 : ereport(LOG,
8124 : errmsg("restore point \"%s\" created at %X/%08X",
8125 : rpName, LSN_FORMAT_ARGS(RecPtr)));
8126 :
8127 6 : return RecPtr;
8128 : }
8129 :
8130 : /*
8131 : * Check if any of the GUC parameters that are critical for hot standby
8132 : * have changed, and update the value in pg_control file if necessary.
8133 : */
8134 : static void
8135 1772 : XLogReportParameters(void)
8136 : {
8137 1772 : if (wal_level != ControlFile->wal_level ||
8138 1310 : wal_log_hints != ControlFile->wal_log_hints ||
8139 1144 : MaxConnections != ControlFile->MaxConnections ||
8140 1142 : max_worker_processes != ControlFile->max_worker_processes ||
8141 1140 : max_wal_senders != ControlFile->max_wal_senders ||
8142 1098 : max_prepared_xacts != ControlFile->max_prepared_xacts ||
8143 912 : max_locks_per_xact != ControlFile->max_locks_per_xact ||
8144 912 : track_commit_timestamp != ControlFile->track_commit_timestamp)
8145 : {
8146 : /*
8147 : * The change in number of backend slots doesn't need to be WAL-logged
8148 : * if archiving is not enabled, as you can't start archive recovery
8149 : * with wal_level=minimal anyway. We don't really care about the
8150 : * values in pg_control either if wal_level=minimal, but seems better
8151 : * to keep them up-to-date to avoid confusion.
8152 : */
8153 884 : if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8154 : {
8155 : xl_parameter_change xlrec;
8156 : XLogRecPtr recptr;
8157 :
8158 844 : xlrec.MaxConnections = MaxConnections;
8159 844 : xlrec.max_worker_processes = max_worker_processes;
8160 844 : xlrec.max_wal_senders = max_wal_senders;
8161 844 : xlrec.max_prepared_xacts = max_prepared_xacts;
8162 844 : xlrec.max_locks_per_xact = max_locks_per_xact;
8163 844 : xlrec.wal_level = wal_level;
8164 844 : xlrec.wal_log_hints = wal_log_hints;
8165 844 : xlrec.track_commit_timestamp = track_commit_timestamp;
8166 :
8167 844 : XLogBeginInsert();
8168 844 : XLogRegisterData(&xlrec, sizeof(xlrec));
8169 :
8170 844 : recptr = XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE);
8171 844 : XLogFlush(recptr);
8172 : }
8173 :
8174 884 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8175 :
8176 884 : ControlFile->MaxConnections = MaxConnections;
8177 884 : ControlFile->max_worker_processes = max_worker_processes;
8178 884 : ControlFile->max_wal_senders = max_wal_senders;
8179 884 : ControlFile->max_prepared_xacts = max_prepared_xacts;
8180 884 : ControlFile->max_locks_per_xact = max_locks_per_xact;
8181 884 : ControlFile->wal_level = wal_level;
8182 884 : ControlFile->wal_log_hints = wal_log_hints;
8183 884 : ControlFile->track_commit_timestamp = track_commit_timestamp;
8184 884 : UpdateControlFile();
8185 :
8186 884 : LWLockRelease(ControlFileLock);
8187 : }
8188 1772 : }
8189 :
8190 : /*
8191 : * Update full_page_writes in shared memory, and write an
8192 : * XLOG_FPW_CHANGE record if necessary.
8193 : *
8194 : * Note: this function assumes there is no other process running
8195 : * concurrently that could update it.
8196 : */
8197 : void
8198 2954 : UpdateFullPageWrites(void)
8199 : {
8200 2954 : XLogCtlInsert *Insert = &XLogCtl->Insert;
8201 : bool recoveryInProgress;
8202 :
8203 : /*
8204 : * Do nothing if full_page_writes has not been changed.
8205 : *
8206 : * It's safe to check the shared full_page_writes without the lock,
8207 : * because we assume that there is no concurrently running process which
8208 : * can update it.
8209 : */
8210 2954 : if (fullPageWrites == Insert->fullPageWrites)
8211 2234 : return;
8212 :
8213 : /*
8214 : * Perform this outside critical section so that the WAL insert
8215 : * initialization done by RecoveryInProgress() doesn't trigger an
8216 : * assertion failure.
8217 : */
8218 720 : recoveryInProgress = RecoveryInProgress();
8219 :
8220 720 : START_CRIT_SECTION();
8221 :
8222 : /*
8223 : * It's always safe to take full page images, even when not strictly
8224 : * required, but not the other round. So if we're setting full_page_writes
8225 : * to true, first set it true and then write the WAL record. If we're
8226 : * setting it to false, first write the WAL record and then set the global
8227 : * flag.
8228 : */
8229 720 : if (fullPageWrites)
8230 : {
8231 700 : WALInsertLockAcquireExclusive();
8232 700 : Insert->fullPageWrites = true;
8233 700 : WALInsertLockRelease();
8234 : }
8235 :
8236 : /*
8237 : * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8238 : * full_page_writes during archive recovery, if required.
8239 : */
8240 720 : if (XLogStandbyInfoActive() && !recoveryInProgress)
8241 : {
8242 0 : XLogBeginInsert();
8243 0 : XLogRegisterData(&fullPageWrites, sizeof(bool));
8244 :
8245 0 : XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
8246 : }
8247 :
8248 720 : if (!fullPageWrites)
8249 : {
8250 20 : WALInsertLockAcquireExclusive();
8251 20 : Insert->fullPageWrites = false;
8252 20 : WALInsertLockRelease();
8253 : }
8254 720 : END_CRIT_SECTION();
8255 : }
8256 :
8257 : /*
8258 : * XLOG resource manager's routines
8259 : *
8260 : * Definitions of info values are in include/catalog/pg_control.h, though
8261 : * not all record types are related to control file updates.
8262 : *
8263 : * NOTE: Some XLOG record types that are directly related to WAL recovery
8264 : * are handled in xlogrecovery_redo().
8265 : */
8266 : void
8267 86802 : xlog_redo(XLogReaderState *record)
8268 : {
8269 86802 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8270 86802 : XLogRecPtr lsn = record->EndRecPtr;
8271 :
8272 : /*
8273 : * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
8274 : * XLOG_FPI_FOR_HINT records.
8275 : */
8276 : Assert(info == XLOG_FPI || info == XLOG_FPI_FOR_HINT ||
8277 : !XLogRecHasAnyBlockRefs(record));
8278 :
8279 86802 : if (info == XLOG_NEXTOID)
8280 : {
8281 : Oid nextOid;
8282 :
8283 : /*
8284 : * We used to try to take the maximum of TransamVariables->nextOid and
8285 : * the recorded nextOid, but that fails if the OID counter wraps
8286 : * around. Since no OID allocation should be happening during replay
8287 : * anyway, better to just believe the record exactly. We still take
8288 : * OidGenLock while setting the variable, just in case.
8289 : */
8290 186 : memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
8291 186 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8292 186 : TransamVariables->nextOid = nextOid;
8293 186 : TransamVariables->oidCount = 0;
8294 186 : LWLockRelease(OidGenLock);
8295 : }
8296 86616 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
8297 : {
8298 : CheckPoint checkPoint;
8299 : TimeLineID replayTLI;
8300 :
8301 68 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8302 : /* In a SHUTDOWN checkpoint, believe the counters exactly */
8303 68 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8304 68 : TransamVariables->nextXid = checkPoint.nextXid;
8305 68 : LWLockRelease(XidGenLock);
8306 68 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8307 68 : TransamVariables->nextOid = checkPoint.nextOid;
8308 68 : TransamVariables->oidCount = 0;
8309 68 : LWLockRelease(OidGenLock);
8310 68 : MultiXactSetNextMXact(checkPoint.nextMulti,
8311 : checkPoint.nextMultiOffset);
8312 :
8313 68 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8314 : checkPoint.oldestMultiDB);
8315 :
8316 : /*
8317 : * No need to set oldestClogXid here as well; it'll be set when we
8318 : * redo an xl_clog_truncate if it changed since initialization.
8319 : */
8320 68 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
8321 :
8322 : /*
8323 : * If we see a shutdown checkpoint while waiting for an end-of-backup
8324 : * record, the backup was canceled and the end-of-backup record will
8325 : * never arrive.
8326 : */
8327 68 : if (ArchiveRecoveryRequested &&
8328 66 : !XLogRecPtrIsInvalid(ControlFile->backupStartPoint) &&
8329 0 : XLogRecPtrIsInvalid(ControlFile->backupEndPoint))
8330 0 : ereport(PANIC,
8331 : (errmsg("online backup was canceled, recovery cannot continue")));
8332 :
8333 : /*
8334 : * If we see a shutdown checkpoint, we know that nothing was running
8335 : * on the primary at this point. So fake-up an empty running-xacts
8336 : * record and use that here and now. Recover additional standby state
8337 : * for prepared transactions.
8338 : */
8339 68 : if (standbyState >= STANDBY_INITIALIZED)
8340 : {
8341 : TransactionId *xids;
8342 : int nxids;
8343 : TransactionId oldestActiveXID;
8344 : TransactionId latestCompletedXid;
8345 : RunningTransactionsData running;
8346 :
8347 62 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
8348 :
8349 : /* Update pg_subtrans entries for any prepared transactions */
8350 62 : StandbyRecoverPreparedTransactions();
8351 :
8352 : /*
8353 : * Construct a RunningTransactions snapshot representing a shut
8354 : * down server, with only prepared transactions still alive. We're
8355 : * never overflowed at this point because all subxids are listed
8356 : * with their parent prepared transactions.
8357 : */
8358 62 : running.xcnt = nxids;
8359 62 : running.subxcnt = 0;
8360 62 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
8361 62 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
8362 62 : running.oldestRunningXid = oldestActiveXID;
8363 62 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
8364 62 : TransactionIdRetreat(latestCompletedXid);
8365 : Assert(TransactionIdIsNormal(latestCompletedXid));
8366 62 : running.latestCompletedXid = latestCompletedXid;
8367 62 : running.xids = xids;
8368 :
8369 62 : ProcArrayApplyRecoveryInfo(&running);
8370 : }
8371 :
8372 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8373 68 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8374 68 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8375 68 : LWLockRelease(ControlFileLock);
8376 :
8377 : /*
8378 : * We should've already switched to the new TLI before replaying this
8379 : * record.
8380 : */
8381 68 : (void) GetCurrentReplayRecPtr(&replayTLI);
8382 68 : if (checkPoint.ThisTimeLineID != replayTLI)
8383 0 : ereport(PANIC,
8384 : (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
8385 : checkPoint.ThisTimeLineID, replayTLI)));
8386 :
8387 68 : RecoveryRestartPoint(&checkPoint, record);
8388 :
8389 : /*
8390 : * After replaying a checkpoint record, free all smgr objects.
8391 : * Otherwise we would never do so for dropped relations, as the
8392 : * startup does not process shared invalidation messages or call
8393 : * AtEOXact_SMgr().
8394 : */
8395 68 : smgrdestroyall();
8396 : }
8397 86548 : else if (info == XLOG_CHECKPOINT_ONLINE)
8398 : {
8399 : CheckPoint checkPoint;
8400 : TimeLineID replayTLI;
8401 :
8402 1348 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8403 : /* In an ONLINE checkpoint, treat the XID counter as a minimum */
8404 1348 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8405 1348 : if (FullTransactionIdPrecedes(TransamVariables->nextXid,
8406 : checkPoint.nextXid))
8407 0 : TransamVariables->nextXid = checkPoint.nextXid;
8408 1348 : LWLockRelease(XidGenLock);
8409 :
8410 : /*
8411 : * We ignore the nextOid counter in an ONLINE checkpoint, preferring
8412 : * to track OID assignment through XLOG_NEXTOID records. The nextOid
8413 : * counter is from the start of the checkpoint and might well be stale
8414 : * compared to later XLOG_NEXTOID records. We could try to take the
8415 : * maximum of the nextOid counter and our latest value, but since
8416 : * there's no particular guarantee about the speed with which the OID
8417 : * counter wraps around, that's a risky thing to do. In any case,
8418 : * users of the nextOid counter are required to avoid assignment of
8419 : * duplicates, so that a somewhat out-of-date value should be safe.
8420 : */
8421 :
8422 : /* Handle multixact */
8423 1348 : MultiXactAdvanceNextMXact(checkPoint.nextMulti,
8424 : checkPoint.nextMultiOffset);
8425 :
8426 : /*
8427 : * NB: This may perform multixact truncation when replaying WAL
8428 : * generated by an older primary.
8429 : */
8430 1348 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8431 : checkPoint.oldestMultiDB);
8432 1348 : if (TransactionIdPrecedes(TransamVariables->oldestXid,
8433 : checkPoint.oldestXid))
8434 0 : SetTransactionIdLimit(checkPoint.oldestXid,
8435 : checkPoint.oldestXidDB);
8436 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8437 1348 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8438 1348 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8439 1348 : LWLockRelease(ControlFileLock);
8440 :
8441 : /* TLI should not change in an on-line checkpoint */
8442 1348 : (void) GetCurrentReplayRecPtr(&replayTLI);
8443 1348 : if (checkPoint.ThisTimeLineID != replayTLI)
8444 0 : ereport(PANIC,
8445 : (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
8446 : checkPoint.ThisTimeLineID, replayTLI)));
8447 :
8448 1348 : RecoveryRestartPoint(&checkPoint, record);
8449 :
8450 : /*
8451 : * After replaying a checkpoint record, free all smgr objects.
8452 : * Otherwise we would never do so for dropped relations, as the
8453 : * startup does not process shared invalidation messages or call
8454 : * AtEOXact_SMgr().
8455 : */
8456 1348 : smgrdestroyall();
8457 : }
8458 85200 : else if (info == XLOG_OVERWRITE_CONTRECORD)
8459 : {
8460 : /* nothing to do here, handled in xlogrecovery_redo() */
8461 : }
8462 85198 : else if (info == XLOG_END_OF_RECOVERY)
8463 : {
8464 : xl_end_of_recovery xlrec;
8465 : TimeLineID replayTLI;
8466 :
8467 20 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
8468 :
8469 : /*
8470 : * For Hot Standby, we could treat this like a Shutdown Checkpoint,
8471 : * but this case is rarer and harder to test, so the benefit doesn't
8472 : * outweigh the potential extra cost of maintenance.
8473 : */
8474 :
8475 : /*
8476 : * We should've already switched to the new TLI before replaying this
8477 : * record.
8478 : */
8479 20 : (void) GetCurrentReplayRecPtr(&replayTLI);
8480 20 : if (xlrec.ThisTimeLineID != replayTLI)
8481 0 : ereport(PANIC,
8482 : (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
8483 : xlrec.ThisTimeLineID, replayTLI)));
8484 : }
8485 85178 : else if (info == XLOG_NOOP)
8486 : {
8487 : /* nothing to do here */
8488 : }
8489 85178 : else if (info == XLOG_SWITCH)
8490 : {
8491 : /* nothing to do here */
8492 : }
8493 84294 : else if (info == XLOG_RESTORE_POINT)
8494 : {
8495 : /* nothing to do here, handled in xlogrecovery.c */
8496 : }
8497 84284 : else if (info == XLOG_FPI || info == XLOG_FPI_FOR_HINT)
8498 : {
8499 : /*
8500 : * XLOG_FPI records contain nothing else but one or more block
8501 : * references. Every block reference must include a full-page image
8502 : * even if full_page_writes was disabled when the record was generated
8503 : * - otherwise there would be no point in this record.
8504 : *
8505 : * XLOG_FPI_FOR_HINT records are generated when a page needs to be
8506 : * WAL-logged because of a hint bit update. They are only generated
8507 : * when checksums and/or wal_log_hints are enabled. They may include
8508 : * no full-page images if full_page_writes was disabled when they were
8509 : * generated. In this case there is nothing to do here.
8510 : *
8511 : * No recovery conflicts are generated by these generic records - if a
8512 : * resource manager needs to generate conflicts, it has to define a
8513 : * separate WAL record type and redo routine.
8514 : */
8515 174860 : for (uint8 block_id = 0; block_id <= XLogRecMaxBlockId(record); block_id++)
8516 : {
8517 : Buffer buffer;
8518 :
8519 92162 : if (!XLogRecHasBlockImage(record, block_id))
8520 : {
8521 132 : if (info == XLOG_FPI)
8522 0 : elog(ERROR, "XLOG_FPI record did not contain a full-page image");
8523 132 : continue;
8524 : }
8525 :
8526 92030 : if (XLogReadBufferForRedo(record, block_id, &buffer) != BLK_RESTORED)
8527 0 : elog(ERROR, "unexpected XLogReadBufferForRedo result when restoring backup block");
8528 92030 : UnlockReleaseBuffer(buffer);
8529 : }
8530 : }
8531 1586 : else if (info == XLOG_BACKUP_END)
8532 : {
8533 : /* nothing to do here, handled in xlogrecovery_redo() */
8534 : }
8535 1416 : else if (info == XLOG_PARAMETER_CHANGE)
8536 : {
8537 : xl_parameter_change xlrec;
8538 :
8539 : /* Update our copy of the parameters in pg_control */
8540 66 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
8541 :
8542 : /*
8543 : * Invalidate logical slots if we are in hot standby and the primary
8544 : * does not have a WAL level sufficient for logical decoding. No need
8545 : * to search for potentially conflicting logically slots if standby is
8546 : * running with wal_level lower than logical, because in that case, we
8547 : * would have either disallowed creation of logical slots or
8548 : * invalidated existing ones.
8549 : */
8550 66 : if (InRecovery && InHotStandby &&
8551 36 : xlrec.wal_level < WAL_LEVEL_LOGICAL &&
8552 14 : wal_level >= WAL_LEVEL_LOGICAL)
8553 10 : InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_LEVEL,
8554 : 0, InvalidOid,
8555 : InvalidTransactionId);
8556 :
8557 66 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8558 66 : ControlFile->MaxConnections = xlrec.MaxConnections;
8559 66 : ControlFile->max_worker_processes = xlrec.max_worker_processes;
8560 66 : ControlFile->max_wal_senders = xlrec.max_wal_senders;
8561 66 : ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
8562 66 : ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
8563 66 : ControlFile->wal_level = xlrec.wal_level;
8564 66 : ControlFile->wal_log_hints = xlrec.wal_log_hints;
8565 :
8566 : /*
8567 : * Update minRecoveryPoint to ensure that if recovery is aborted, we
8568 : * recover back up to this point before allowing hot standby again.
8569 : * This is important if the max_* settings are decreased, to ensure
8570 : * you don't run queries against the WAL preceding the change. The
8571 : * local copies cannot be updated as long as crash recovery is
8572 : * happening and we expect all the WAL to be replayed.
8573 : */
8574 66 : if (InArchiveRecovery)
8575 : {
8576 38 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
8577 38 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
8578 : }
8579 66 : if (LocalMinRecoveryPoint != InvalidXLogRecPtr && LocalMinRecoveryPoint < lsn)
8580 : {
8581 : TimeLineID replayTLI;
8582 :
8583 14 : (void) GetCurrentReplayRecPtr(&replayTLI);
8584 14 : ControlFile->minRecoveryPoint = lsn;
8585 14 : ControlFile->minRecoveryPointTLI = replayTLI;
8586 : }
8587 :
8588 66 : CommitTsParameterChange(xlrec.track_commit_timestamp,
8589 66 : ControlFile->track_commit_timestamp);
8590 66 : ControlFile->track_commit_timestamp = xlrec.track_commit_timestamp;
8591 :
8592 66 : UpdateControlFile();
8593 66 : LWLockRelease(ControlFileLock);
8594 :
8595 : /* Check to see if any parameter change gives a problem on recovery */
8596 66 : CheckRequiredParameterValues();
8597 : }
8598 1350 : else if (info == XLOG_FPW_CHANGE)
8599 : {
8600 : bool fpw;
8601 :
8602 0 : memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
8603 :
8604 : /*
8605 : * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
8606 : * do_pg_backup_start() and do_pg_backup_stop() can check whether
8607 : * full_page_writes has been disabled during online backup.
8608 : */
8609 0 : if (!fpw)
8610 : {
8611 0 : SpinLockAcquire(&XLogCtl->info_lck);
8612 0 : if (XLogCtl->lastFpwDisableRecPtr < record->ReadRecPtr)
8613 0 : XLogCtl->lastFpwDisableRecPtr = record->ReadRecPtr;
8614 0 : SpinLockRelease(&XLogCtl->info_lck);
8615 : }
8616 :
8617 : /* Keep track of full_page_writes */
8618 0 : lastFullPageWrites = fpw;
8619 : }
8620 : else if (info == XLOG_CHECKPOINT_REDO)
8621 : {
8622 : /* nothing to do here, just for informational purposes */
8623 : }
8624 86798 : }
8625 :
8626 : /*
8627 : * Return the extra open flags used for opening a file, depending on the
8628 : * value of the GUCs wal_sync_method, fsync and debug_io_direct.
8629 : */
8630 : static int
8631 31436 : get_sync_bit(int method)
8632 : {
8633 31436 : int o_direct_flag = 0;
8634 :
8635 : /*
8636 : * Use O_DIRECT if requested, except in walreceiver process. The WAL
8637 : * written by walreceiver is normally read by the startup process soon
8638 : * after it's written. Also, walreceiver performs unaligned writes, which
8639 : * don't work with O_DIRECT, so it is required for correctness too.
8640 : */
8641 31436 : if ((io_direct_flags & IO_DIRECT_WAL) && !AmWalReceiverProcess())
8642 16 : o_direct_flag = PG_O_DIRECT;
8643 :
8644 : /* If fsync is disabled, never open in sync mode */
8645 31436 : if (!enableFsync)
8646 31436 : return o_direct_flag;
8647 :
8648 0 : switch (method)
8649 : {
8650 : /*
8651 : * enum values for all sync options are defined even if they are
8652 : * not supported on the current platform. But if not, they are
8653 : * not included in the enum option array, and therefore will never
8654 : * be seen here.
8655 : */
8656 0 : case WAL_SYNC_METHOD_FSYNC:
8657 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8658 : case WAL_SYNC_METHOD_FDATASYNC:
8659 0 : return o_direct_flag;
8660 : #ifdef O_SYNC
8661 0 : case WAL_SYNC_METHOD_OPEN:
8662 0 : return O_SYNC | o_direct_flag;
8663 : #endif
8664 : #ifdef O_DSYNC
8665 0 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8666 0 : return O_DSYNC | o_direct_flag;
8667 : #endif
8668 0 : default:
8669 : /* can't happen (unless we are out of sync with option array) */
8670 0 : elog(ERROR, "unrecognized \"wal_sync_method\": %d", method);
8671 : return 0; /* silence warning */
8672 : }
8673 : }
8674 :
8675 : /*
8676 : * GUC support
8677 : */
8678 : void
8679 2250 : assign_wal_sync_method(int new_wal_sync_method, void *extra)
8680 : {
8681 2250 : if (wal_sync_method != new_wal_sync_method)
8682 : {
8683 : /*
8684 : * To ensure that no blocks escape unsynced, force an fsync on the
8685 : * currently open log segment (if any). Also, if the open flag is
8686 : * changing, close the log file so it will be reopened (with new flag
8687 : * bit) at next use.
8688 : */
8689 0 : if (openLogFile >= 0)
8690 : {
8691 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN);
8692 0 : if (pg_fsync(openLogFile) != 0)
8693 : {
8694 : char xlogfname[MAXFNAMELEN];
8695 : int save_errno;
8696 :
8697 0 : save_errno = errno;
8698 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo,
8699 : wal_segment_size);
8700 0 : errno = save_errno;
8701 0 : ereport(PANIC,
8702 : (errcode_for_file_access(),
8703 : errmsg("could not fsync file \"%s\": %m", xlogfname)));
8704 : }
8705 :
8706 0 : pgstat_report_wait_end();
8707 0 : if (get_sync_bit(wal_sync_method) != get_sync_bit(new_wal_sync_method))
8708 0 : XLogFileClose();
8709 : }
8710 : }
8711 2250 : }
8712 :
8713 :
8714 : /*
8715 : * Issue appropriate kind of fsync (if any) for an XLOG output file.
8716 : *
8717 : * 'fd' is a file descriptor for the XLOG file to be fsync'd.
8718 : * 'segno' is for error reporting purposes.
8719 : */
8720 : void
8721 311942 : issue_xlog_fsync(int fd, XLogSegNo segno, TimeLineID tli)
8722 : {
8723 311942 : char *msg = NULL;
8724 : instr_time start;
8725 :
8726 : Assert(tli != 0);
8727 :
8728 : /*
8729 : * Quick exit if fsync is disabled or write() has already synced the WAL
8730 : * file.
8731 : */
8732 311942 : if (!enableFsync ||
8733 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN ||
8734 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN_DSYNC)
8735 311942 : return;
8736 :
8737 : /*
8738 : * Measure I/O timing to sync the WAL file for pg_stat_io.
8739 : */
8740 0 : start = pgstat_prepare_io_time(track_wal_io_timing);
8741 :
8742 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC);
8743 0 : switch (wal_sync_method)
8744 : {
8745 0 : case WAL_SYNC_METHOD_FSYNC:
8746 0 : if (pg_fsync_no_writethrough(fd) != 0)
8747 0 : msg = _("could not fsync file \"%s\": %m");
8748 0 : break;
8749 : #ifdef HAVE_FSYNC_WRITETHROUGH
8750 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8751 : if (pg_fsync_writethrough(fd) != 0)
8752 : msg = _("could not fsync write-through file \"%s\": %m");
8753 : break;
8754 : #endif
8755 0 : case WAL_SYNC_METHOD_FDATASYNC:
8756 0 : if (pg_fdatasync(fd) != 0)
8757 0 : msg = _("could not fdatasync file \"%s\": %m");
8758 0 : break;
8759 0 : case WAL_SYNC_METHOD_OPEN:
8760 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8761 : /* not reachable */
8762 : Assert(false);
8763 0 : break;
8764 0 : default:
8765 0 : ereport(PANIC,
8766 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8767 : errmsg_internal("unrecognized \"wal_sync_method\": %d", wal_sync_method));
8768 : break;
8769 : }
8770 :
8771 : /* PANIC if failed to fsync */
8772 0 : if (msg)
8773 : {
8774 : char xlogfname[MAXFNAMELEN];
8775 0 : int save_errno = errno;
8776 :
8777 0 : XLogFileName(xlogfname, tli, segno, wal_segment_size);
8778 0 : errno = save_errno;
8779 0 : ereport(PANIC,
8780 : (errcode_for_file_access(),
8781 : errmsg(msg, xlogfname)));
8782 : }
8783 :
8784 0 : pgstat_report_wait_end();
8785 :
8786 0 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL, IOOP_FSYNC,
8787 : start, 1, 0);
8788 : }
8789 :
8790 : /*
8791 : * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
8792 : * function. It creates the necessary starting checkpoint and constructs the
8793 : * backup state and tablespace map.
8794 : *
8795 : * Input parameters are "state" (the backup state), "fast" (if true, we do
8796 : * the checkpoint in fast mode), and "tablespaces" (if non-NULL, indicates a
8797 : * list of tablespaceinfo structs describing the cluster's tablespaces.).
8798 : *
8799 : * The tablespace map contents are appended to passed-in parameter
8800 : * tablespace_map and the caller is responsible for including it in the backup
8801 : * archive as 'tablespace_map'. The tablespace_map file is required mainly for
8802 : * tar format in windows as native windows utilities are not able to create
8803 : * symlinks while extracting files from tar. However for consistency and
8804 : * platform-independence, we do it the same way everywhere.
8805 : *
8806 : * It fills in "state" with the information required for the backup, such
8807 : * as the minimum WAL location that must be present to restore from this
8808 : * backup (starttli) and the corresponding timeline ID (starttli).
8809 : *
8810 : * Every successfully started backup must be stopped by calling
8811 : * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
8812 : * backups active at the same time.
8813 : *
8814 : * It is the responsibility of the caller of this function to verify the
8815 : * permissions of the calling user!
8816 : */
8817 : void
8818 324 : do_pg_backup_start(const char *backupidstr, bool fast, List **tablespaces,
8819 : BackupState *state, StringInfo tblspcmapfile)
8820 : {
8821 : bool backup_started_in_recovery;
8822 :
8823 : Assert(state != NULL);
8824 324 : backup_started_in_recovery = RecoveryInProgress();
8825 :
8826 : /*
8827 : * During recovery, we don't need to check WAL level. Because, if WAL
8828 : * level is not sufficient, it's impossible to get here during recovery.
8829 : */
8830 324 : if (!backup_started_in_recovery && !XLogIsNeeded())
8831 0 : ereport(ERROR,
8832 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8833 : errmsg("WAL level not sufficient for making an online backup"),
8834 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
8835 :
8836 324 : if (strlen(backupidstr) > MAXPGPATH)
8837 2 : ereport(ERROR,
8838 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8839 : errmsg("backup label too long (max %d bytes)",
8840 : MAXPGPATH)));
8841 :
8842 322 : strlcpy(state->name, backupidstr, sizeof(state->name));
8843 :
8844 : /*
8845 : * Mark backup active in shared memory. We must do full-page WAL writes
8846 : * during an on-line backup even if not doing so at other times, because
8847 : * it's quite possible for the backup dump to obtain a "torn" (partially
8848 : * written) copy of a database page if it reads the page concurrently with
8849 : * our write to the same page. This can be fixed as long as the first
8850 : * write to the page in the WAL sequence is a full-page write. Hence, we
8851 : * increment runningBackups then force a CHECKPOINT, to ensure there are
8852 : * no dirty pages in shared memory that might get dumped while the backup
8853 : * is in progress without having a corresponding WAL record. (Once the
8854 : * backup is complete, we need not force full-page writes anymore, since
8855 : * we expect that any pages not modified during the backup interval must
8856 : * have been correctly captured by the backup.)
8857 : *
8858 : * Note that forcing full-page writes has no effect during an online
8859 : * backup from the standby.
8860 : *
8861 : * We must hold all the insertion locks to change the value of
8862 : * runningBackups, to ensure adequate interlocking against
8863 : * XLogInsertRecord().
8864 : */
8865 322 : WALInsertLockAcquireExclusive();
8866 322 : XLogCtl->Insert.runningBackups++;
8867 322 : WALInsertLockRelease();
8868 :
8869 : /*
8870 : * Ensure we decrement runningBackups if we fail below. NB -- for this to
8871 : * work correctly, it is critical that sessionBackupState is only updated
8872 : * after this block is over.
8873 : */
8874 322 : PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
8875 : {
8876 322 : bool gotUniqueStartpoint = false;
8877 : DIR *tblspcdir;
8878 : struct dirent *de;
8879 : tablespaceinfo *ti;
8880 : int datadirpathlen;
8881 :
8882 : /*
8883 : * Force an XLOG file switch before the checkpoint, to ensure that the
8884 : * WAL segment the checkpoint is written to doesn't contain pages with
8885 : * old timeline IDs. That would otherwise happen if you called
8886 : * pg_backup_start() right after restoring from a PITR archive: the
8887 : * first WAL segment containing the startup checkpoint has pages in
8888 : * the beginning with the old timeline ID. That can cause trouble at
8889 : * recovery: we won't have a history file covering the old timeline if
8890 : * pg_wal directory was not included in the base backup and the WAL
8891 : * archive was cleared too before starting the backup.
8892 : *
8893 : * This also ensures that we have emitted a WAL page header that has
8894 : * XLP_BKP_REMOVABLE off before we emit the checkpoint record.
8895 : * Therefore, if a WAL archiver (such as pglesslog) is trying to
8896 : * compress out removable backup blocks, it won't remove any that
8897 : * occur after this point.
8898 : *
8899 : * During recovery, we skip forcing XLOG file switch, which means that
8900 : * the backup taken during recovery is not available for the special
8901 : * recovery case described above.
8902 : */
8903 322 : if (!backup_started_in_recovery)
8904 308 : RequestXLogSwitch(false);
8905 :
8906 : do
8907 : {
8908 : bool checkpointfpw;
8909 :
8910 : /*
8911 : * Force a CHECKPOINT. Aside from being necessary to prevent torn
8912 : * page problems, this guarantees that two successive backup runs
8913 : * will have different checkpoint positions and hence different
8914 : * history file names, even if nothing happened in between.
8915 : *
8916 : * During recovery, establish a restartpoint if possible. We use
8917 : * the last restartpoint as the backup starting checkpoint. This
8918 : * means that two successive backup runs can have same checkpoint
8919 : * positions.
8920 : *
8921 : * Since the fact that we are executing do_pg_backup_start()
8922 : * during recovery means that checkpointer is running, we can use
8923 : * RequestCheckpoint() to establish a restartpoint.
8924 : *
8925 : * We use CHECKPOINT_FAST only if requested by user (via passing
8926 : * fast = true). Otherwise this can take awhile.
8927 : */
8928 322 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
8929 : (fast ? CHECKPOINT_FAST : 0));
8930 :
8931 : /*
8932 : * Now we need to fetch the checkpoint record location, and also
8933 : * its REDO pointer. The oldest point in WAL that would be needed
8934 : * to restore starting from the checkpoint is precisely the REDO
8935 : * pointer.
8936 : */
8937 322 : LWLockAcquire(ControlFileLock, LW_SHARED);
8938 322 : state->checkpointloc = ControlFile->checkPoint;
8939 322 : state->startpoint = ControlFile->checkPointCopy.redo;
8940 322 : state->starttli = ControlFile->checkPointCopy.ThisTimeLineID;
8941 322 : checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
8942 322 : LWLockRelease(ControlFileLock);
8943 :
8944 322 : if (backup_started_in_recovery)
8945 : {
8946 : XLogRecPtr recptr;
8947 :
8948 : /*
8949 : * Check to see if all WAL replayed during online backup
8950 : * (i.e., since last restartpoint used as backup starting
8951 : * checkpoint) contain full-page writes.
8952 : */
8953 14 : SpinLockAcquire(&XLogCtl->info_lck);
8954 14 : recptr = XLogCtl->lastFpwDisableRecPtr;
8955 14 : SpinLockRelease(&XLogCtl->info_lck);
8956 :
8957 14 : if (!checkpointfpw || state->startpoint <= recptr)
8958 0 : ereport(ERROR,
8959 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8960 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
8961 : "since last restartpoint"),
8962 : errhint("This means that the backup being taken on the standby "
8963 : "is corrupt and should not be used. "
8964 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
8965 : "and then try an online backup again.")));
8966 :
8967 : /*
8968 : * During recovery, since we don't use the end-of-backup WAL
8969 : * record and don't write the backup history file, the
8970 : * starting WAL location doesn't need to be unique. This means
8971 : * that two base backups started at the same time might use
8972 : * the same checkpoint as starting locations.
8973 : */
8974 14 : gotUniqueStartpoint = true;
8975 : }
8976 :
8977 : /*
8978 : * If two base backups are started at the same time (in WAL sender
8979 : * processes), we need to make sure that they use different
8980 : * checkpoints as starting locations, because we use the starting
8981 : * WAL location as a unique identifier for the base backup in the
8982 : * end-of-backup WAL record and when we write the backup history
8983 : * file. Perhaps it would be better generate a separate unique ID
8984 : * for each backup instead of forcing another checkpoint, but
8985 : * taking a checkpoint right after another is not that expensive
8986 : * either because only few buffers have been dirtied yet.
8987 : */
8988 322 : WALInsertLockAcquireExclusive();
8989 322 : if (XLogCtl->Insert.lastBackupStart < state->startpoint)
8990 : {
8991 322 : XLogCtl->Insert.lastBackupStart = state->startpoint;
8992 322 : gotUniqueStartpoint = true;
8993 : }
8994 322 : WALInsertLockRelease();
8995 322 : } while (!gotUniqueStartpoint);
8996 :
8997 : /*
8998 : * Construct tablespace_map file.
8999 : */
9000 322 : datadirpathlen = strlen(DataDir);
9001 :
9002 : /* Collect information about all tablespaces */
9003 322 : tblspcdir = AllocateDir(PG_TBLSPC_DIR);
9004 1038 : while ((de = ReadDir(tblspcdir, PG_TBLSPC_DIR)) != NULL)
9005 : {
9006 : char fullpath[MAXPGPATH + sizeof(PG_TBLSPC_DIR)];
9007 : char linkpath[MAXPGPATH];
9008 716 : char *relpath = NULL;
9009 : char *s;
9010 : PGFileType de_type;
9011 : char *badp;
9012 : Oid tsoid;
9013 :
9014 : /*
9015 : * Try to parse the directory name as an unsigned integer.
9016 : *
9017 : * Tablespace directories should be positive integers that can be
9018 : * represented in 32 bits, with no leading zeroes or trailing
9019 : * garbage. If we come across a name that doesn't meet those
9020 : * criteria, skip it.
9021 : */
9022 716 : if (de->d_name[0] < '1' || de->d_name[1] > '9')
9023 644 : continue;
9024 72 : errno = 0;
9025 72 : tsoid = strtoul(de->d_name, &badp, 10);
9026 72 : if (*badp != '\0' || errno == EINVAL || errno == ERANGE)
9027 0 : continue;
9028 :
9029 72 : snprintf(fullpath, sizeof(fullpath), "%s/%s", PG_TBLSPC_DIR, de->d_name);
9030 :
9031 72 : de_type = get_dirent_type(fullpath, de, false, ERROR);
9032 :
9033 72 : if (de_type == PGFILETYPE_LNK)
9034 : {
9035 : StringInfoData escapedpath;
9036 : int rllen;
9037 :
9038 44 : rllen = readlink(fullpath, linkpath, sizeof(linkpath));
9039 44 : if (rllen < 0)
9040 : {
9041 0 : ereport(WARNING,
9042 : (errmsg("could not read symbolic link \"%s\": %m",
9043 : fullpath)));
9044 0 : continue;
9045 : }
9046 44 : else if (rllen >= sizeof(linkpath))
9047 : {
9048 0 : ereport(WARNING,
9049 : (errmsg("symbolic link \"%s\" target is too long",
9050 : fullpath)));
9051 0 : continue;
9052 : }
9053 44 : linkpath[rllen] = '\0';
9054 :
9055 : /*
9056 : * Relpath holds the relative path of the tablespace directory
9057 : * when it's located within PGDATA, or NULL if it's located
9058 : * elsewhere.
9059 : */
9060 44 : if (rllen > datadirpathlen &&
9061 2 : strncmp(linkpath, DataDir, datadirpathlen) == 0 &&
9062 0 : IS_DIR_SEP(linkpath[datadirpathlen]))
9063 0 : relpath = pstrdup(linkpath + datadirpathlen + 1);
9064 :
9065 : /*
9066 : * Add a backslash-escaped version of the link path to the
9067 : * tablespace map file.
9068 : */
9069 44 : initStringInfo(&escapedpath);
9070 1092 : for (s = linkpath; *s; s++)
9071 : {
9072 1048 : if (*s == '\n' || *s == '\r' || *s == '\\')
9073 0 : appendStringInfoChar(&escapedpath, '\\');
9074 1048 : appendStringInfoChar(&escapedpath, *s);
9075 : }
9076 44 : appendStringInfo(tblspcmapfile, "%s %s\n",
9077 44 : de->d_name, escapedpath.data);
9078 44 : pfree(escapedpath.data);
9079 : }
9080 28 : else if (de_type == PGFILETYPE_DIR)
9081 : {
9082 : /*
9083 : * It's possible to use allow_in_place_tablespaces to create
9084 : * directories directly under pg_tblspc, for testing purposes
9085 : * only.
9086 : *
9087 : * In this case, we store a relative path rather than an
9088 : * absolute path into the tablespaceinfo.
9089 : */
9090 28 : snprintf(linkpath, sizeof(linkpath), "%s/%s",
9091 28 : PG_TBLSPC_DIR, de->d_name);
9092 28 : relpath = pstrdup(linkpath);
9093 : }
9094 : else
9095 : {
9096 : /* Skip any other file type that appears here. */
9097 0 : continue;
9098 : }
9099 :
9100 72 : ti = palloc(sizeof(tablespaceinfo));
9101 72 : ti->oid = tsoid;
9102 72 : ti->path = pstrdup(linkpath);
9103 72 : ti->rpath = relpath;
9104 72 : ti->size = -1;
9105 :
9106 72 : if (tablespaces)
9107 72 : *tablespaces = lappend(*tablespaces, ti);
9108 : }
9109 322 : FreeDir(tblspcdir);
9110 :
9111 322 : state->starttime = (pg_time_t) time(NULL);
9112 : }
9113 322 : PG_END_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9114 :
9115 322 : state->started_in_recovery = backup_started_in_recovery;
9116 :
9117 : /*
9118 : * Mark that the start phase has correctly finished for the backup.
9119 : */
9120 322 : sessionBackupState = SESSION_BACKUP_RUNNING;
9121 322 : }
9122 :
9123 : /*
9124 : * Utility routine to fetch the session-level status of a backup running.
9125 : */
9126 : SessionBackupState
9127 364 : get_backup_status(void)
9128 : {
9129 364 : return sessionBackupState;
9130 : }
9131 :
9132 : /*
9133 : * do_pg_backup_stop
9134 : *
9135 : * Utility function called at the end of an online backup. It creates history
9136 : * file (if required), resets sessionBackupState and so on. It can optionally
9137 : * wait for WAL segments to be archived.
9138 : *
9139 : * "state" is filled with the information necessary to restore from this
9140 : * backup with its stop LSN (stoppoint), its timeline ID (stoptli), etc.
9141 : *
9142 : * It is the responsibility of the caller of this function to verify the
9143 : * permissions of the calling user!
9144 : */
9145 : void
9146 310 : do_pg_backup_stop(BackupState *state, bool waitforarchive)
9147 : {
9148 310 : bool backup_stopped_in_recovery = false;
9149 : char histfilepath[MAXPGPATH];
9150 : char lastxlogfilename[MAXFNAMELEN];
9151 : char histfilename[MAXFNAMELEN];
9152 : XLogSegNo _logSegNo;
9153 : FILE *fp;
9154 : int seconds_before_warning;
9155 310 : int waits = 0;
9156 310 : bool reported_waiting = false;
9157 :
9158 : Assert(state != NULL);
9159 :
9160 310 : backup_stopped_in_recovery = RecoveryInProgress();
9161 :
9162 : /*
9163 : * During recovery, we don't need to check WAL level. Because, if WAL
9164 : * level is not sufficient, it's impossible to get here during recovery.
9165 : */
9166 310 : if (!backup_stopped_in_recovery && !XLogIsNeeded())
9167 0 : ereport(ERROR,
9168 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9169 : errmsg("WAL level not sufficient for making an online backup"),
9170 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9171 :
9172 : /*
9173 : * OK to update backup counter and session-level lock.
9174 : *
9175 : * Note that CHECK_FOR_INTERRUPTS() must not occur while updating them,
9176 : * otherwise they can be updated inconsistently, which might cause
9177 : * do_pg_abort_backup() to fail.
9178 : */
9179 310 : WALInsertLockAcquireExclusive();
9180 :
9181 : /*
9182 : * It is expected that each do_pg_backup_start() call is matched by
9183 : * exactly one do_pg_backup_stop() call.
9184 : */
9185 : Assert(XLogCtl->Insert.runningBackups > 0);
9186 310 : XLogCtl->Insert.runningBackups--;
9187 :
9188 : /*
9189 : * Clean up session-level lock.
9190 : *
9191 : * You might think that WALInsertLockRelease() can be called before
9192 : * cleaning up session-level lock because session-level lock doesn't need
9193 : * to be protected with WAL insertion lock. But since
9194 : * CHECK_FOR_INTERRUPTS() can occur in it, session-level lock must be
9195 : * cleaned up before it.
9196 : */
9197 310 : sessionBackupState = SESSION_BACKUP_NONE;
9198 :
9199 310 : WALInsertLockRelease();
9200 :
9201 : /*
9202 : * If we are taking an online backup from the standby, we confirm that the
9203 : * standby has not been promoted during the backup.
9204 : */
9205 310 : if (state->started_in_recovery && !backup_stopped_in_recovery)
9206 0 : ereport(ERROR,
9207 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9208 : errmsg("the standby was promoted during online backup"),
9209 : errhint("This means that the backup being taken is corrupt "
9210 : "and should not be used. "
9211 : "Try taking another online backup.")));
9212 :
9213 : /*
9214 : * During recovery, we don't write an end-of-backup record. We assume that
9215 : * pg_control was backed up last and its minimum recovery point can be
9216 : * available as the backup end location. Since we don't have an
9217 : * end-of-backup record, we use the pg_control value to check whether
9218 : * we've reached the end of backup when starting recovery from this
9219 : * backup. We have no way of checking if pg_control wasn't backed up last
9220 : * however.
9221 : *
9222 : * We don't force a switch to new WAL file but it is still possible to
9223 : * wait for all the required files to be archived if waitforarchive is
9224 : * true. This is okay if we use the backup to start a standby and fetch
9225 : * the missing WAL using streaming replication. But in the case of an
9226 : * archive recovery, a user should set waitforarchive to true and wait for
9227 : * them to be archived to ensure that all the required files are
9228 : * available.
9229 : *
9230 : * We return the current minimum recovery point as the backup end
9231 : * location. Note that it can be greater than the exact backup end
9232 : * location if the minimum recovery point is updated after the backup of
9233 : * pg_control. This is harmless for current uses.
9234 : *
9235 : * XXX currently a backup history file is for informational and debug
9236 : * purposes only. It's not essential for an online backup. Furthermore,
9237 : * even if it's created, it will not be archived during recovery because
9238 : * an archiver is not invoked. So it doesn't seem worthwhile to write a
9239 : * backup history file during recovery.
9240 : */
9241 310 : if (backup_stopped_in_recovery)
9242 : {
9243 : XLogRecPtr recptr;
9244 :
9245 : /*
9246 : * Check to see if all WAL replayed during online backup contain
9247 : * full-page writes.
9248 : */
9249 14 : SpinLockAcquire(&XLogCtl->info_lck);
9250 14 : recptr = XLogCtl->lastFpwDisableRecPtr;
9251 14 : SpinLockRelease(&XLogCtl->info_lck);
9252 :
9253 14 : if (state->startpoint <= recptr)
9254 0 : ereport(ERROR,
9255 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9256 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9257 : "during online backup"),
9258 : errhint("This means that the backup being taken on the standby "
9259 : "is corrupt and should not be used. "
9260 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9261 : "and then try an online backup again.")));
9262 :
9263 :
9264 14 : LWLockAcquire(ControlFileLock, LW_SHARED);
9265 14 : state->stoppoint = ControlFile->minRecoveryPoint;
9266 14 : state->stoptli = ControlFile->minRecoveryPointTLI;
9267 14 : LWLockRelease(ControlFileLock);
9268 : }
9269 : else
9270 : {
9271 : char *history_file;
9272 :
9273 : /*
9274 : * Write the backup-end xlog record
9275 : */
9276 296 : XLogBeginInsert();
9277 296 : XLogRegisterData(&state->startpoint,
9278 : sizeof(state->startpoint));
9279 296 : state->stoppoint = XLogInsert(RM_XLOG_ID, XLOG_BACKUP_END);
9280 :
9281 : /*
9282 : * Given that we're not in recovery, InsertTimeLineID is set and can't
9283 : * change, so we can read it without a lock.
9284 : */
9285 296 : state->stoptli = XLogCtl->InsertTimeLineID;
9286 :
9287 : /*
9288 : * Force a switch to a new xlog segment file, so that the backup is
9289 : * valid as soon as archiver moves out the current segment file.
9290 : */
9291 296 : RequestXLogSwitch(false);
9292 :
9293 296 : state->stoptime = (pg_time_t) time(NULL);
9294 :
9295 : /*
9296 : * Write the backup history file
9297 : */
9298 296 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9299 296 : BackupHistoryFilePath(histfilepath, state->stoptli, _logSegNo,
9300 : state->startpoint, wal_segment_size);
9301 296 : fp = AllocateFile(histfilepath, "w");
9302 296 : if (!fp)
9303 0 : ereport(ERROR,
9304 : (errcode_for_file_access(),
9305 : errmsg("could not create file \"%s\": %m",
9306 : histfilepath)));
9307 :
9308 : /* Build and save the contents of the backup history file */
9309 296 : history_file = build_backup_content(state, true);
9310 296 : fprintf(fp, "%s", history_file);
9311 296 : pfree(history_file);
9312 :
9313 296 : if (fflush(fp) || ferror(fp) || FreeFile(fp))
9314 0 : ereport(ERROR,
9315 : (errcode_for_file_access(),
9316 : errmsg("could not write file \"%s\": %m",
9317 : histfilepath)));
9318 :
9319 : /*
9320 : * Clean out any no-longer-needed history files. As a side effect,
9321 : * this will post a .ready file for the newly created history file,
9322 : * notifying the archiver that history file may be archived
9323 : * immediately.
9324 : */
9325 296 : CleanupBackupHistory();
9326 : }
9327 :
9328 : /*
9329 : * If archiving is enabled, wait for all the required WAL files to be
9330 : * archived before returning. If archiving isn't enabled, the required WAL
9331 : * needs to be transported via streaming replication (hopefully with
9332 : * wal_keep_size set high enough), or some more exotic mechanism like
9333 : * polling and copying files from pg_wal with script. We have no knowledge
9334 : * of those mechanisms, so it's up to the user to ensure that he gets all
9335 : * the required WAL.
9336 : *
9337 : * We wait until both the last WAL file filled during backup and the
9338 : * history file have been archived, and assume that the alphabetic sorting
9339 : * property of the WAL files ensures any earlier WAL files are safely
9340 : * archived as well.
9341 : *
9342 : * We wait forever, since archive_command is supposed to work and we
9343 : * assume the admin wanted his backup to work completely. If you don't
9344 : * wish to wait, then either waitforarchive should be passed in as false,
9345 : * or you can set statement_timeout. Also, some notices are issued to
9346 : * clue in anyone who might be doing this interactively.
9347 : */
9348 :
9349 310 : if (waitforarchive &&
9350 20 : ((!backup_stopped_in_recovery && XLogArchivingActive()) ||
9351 2 : (backup_stopped_in_recovery && XLogArchivingAlways())))
9352 : {
9353 8 : XLByteToPrevSeg(state->stoppoint, _logSegNo, wal_segment_size);
9354 8 : XLogFileName(lastxlogfilename, state->stoptli, _logSegNo,
9355 : wal_segment_size);
9356 :
9357 8 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9358 8 : BackupHistoryFileName(histfilename, state->stoptli, _logSegNo,
9359 : state->startpoint, wal_segment_size);
9360 :
9361 8 : seconds_before_warning = 60;
9362 8 : waits = 0;
9363 :
9364 24 : while (XLogArchiveIsBusy(lastxlogfilename) ||
9365 8 : XLogArchiveIsBusy(histfilename))
9366 : {
9367 8 : CHECK_FOR_INTERRUPTS();
9368 :
9369 8 : if (!reported_waiting && waits > 5)
9370 : {
9371 0 : ereport(NOTICE,
9372 : (errmsg("base backup done, waiting for required WAL segments to be archived")));
9373 0 : reported_waiting = true;
9374 : }
9375 :
9376 8 : (void) WaitLatch(MyLatch,
9377 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
9378 : 1000L,
9379 : WAIT_EVENT_BACKUP_WAIT_WAL_ARCHIVE);
9380 8 : ResetLatch(MyLatch);
9381 :
9382 8 : if (++waits >= seconds_before_warning)
9383 : {
9384 0 : seconds_before_warning *= 2; /* This wraps in >10 years... */
9385 0 : ereport(WARNING,
9386 : (errmsg("still waiting for all required WAL segments to be archived (%d seconds elapsed)",
9387 : waits),
9388 : errhint("Check that your \"archive_command\" is executing properly. "
9389 : "You can safely cancel this backup, "
9390 : "but the database backup will not be usable without all the WAL segments.")));
9391 : }
9392 : }
9393 :
9394 8 : ereport(NOTICE,
9395 : (errmsg("all required WAL segments have been archived")));
9396 : }
9397 302 : else if (waitforarchive)
9398 12 : ereport(NOTICE,
9399 : (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
9400 310 : }
9401 :
9402 :
9403 : /*
9404 : * do_pg_abort_backup: abort a running backup
9405 : *
9406 : * This does just the most basic steps of do_pg_backup_stop(), by taking the
9407 : * system out of backup mode, thus making it a lot more safe to call from
9408 : * an error handler.
9409 : *
9410 : * 'arg' indicates that it's being called during backup setup; so
9411 : * sessionBackupState has not been modified yet, but runningBackups has
9412 : * already been incremented. When it's false, then it's invoked as a
9413 : * before_shmem_exit handler, and therefore we must not change state
9414 : * unless sessionBackupState indicates that a backup is actually running.
9415 : *
9416 : * NB: This gets used as a PG_ENSURE_ERROR_CLEANUP callback and
9417 : * before_shmem_exit handler, hence the odd-looking signature.
9418 : */
9419 : void
9420 16 : do_pg_abort_backup(int code, Datum arg)
9421 : {
9422 16 : bool during_backup_start = DatumGetBool(arg);
9423 :
9424 : /* If called during backup start, there shouldn't be one already running */
9425 : Assert(!during_backup_start || sessionBackupState == SESSION_BACKUP_NONE);
9426 :
9427 16 : if (during_backup_start || sessionBackupState != SESSION_BACKUP_NONE)
9428 : {
9429 12 : WALInsertLockAcquireExclusive();
9430 : Assert(XLogCtl->Insert.runningBackups > 0);
9431 12 : XLogCtl->Insert.runningBackups--;
9432 :
9433 12 : sessionBackupState = SESSION_BACKUP_NONE;
9434 12 : WALInsertLockRelease();
9435 :
9436 12 : if (!during_backup_start)
9437 12 : ereport(WARNING,
9438 : errmsg("aborting backup due to backend exiting before pg_backup_stop was called"));
9439 : }
9440 16 : }
9441 :
9442 : /*
9443 : * Register a handler that will warn about unterminated backups at end of
9444 : * session, unless this has already been done.
9445 : */
9446 : void
9447 8 : register_persistent_abort_backup_handler(void)
9448 : {
9449 : static bool already_done = false;
9450 :
9451 8 : if (already_done)
9452 2 : return;
9453 6 : before_shmem_exit(do_pg_abort_backup, BoolGetDatum(false));
9454 6 : already_done = true;
9455 : }
9456 :
9457 : /*
9458 : * Get latest WAL insert pointer
9459 : */
9460 : XLogRecPtr
9461 3952 : GetXLogInsertRecPtr(void)
9462 : {
9463 3952 : XLogCtlInsert *Insert = &XLogCtl->Insert;
9464 : uint64 current_bytepos;
9465 :
9466 3952 : SpinLockAcquire(&Insert->insertpos_lck);
9467 3952 : current_bytepos = Insert->CurrBytePos;
9468 3952 : SpinLockRelease(&Insert->insertpos_lck);
9469 :
9470 3952 : return XLogBytePosToRecPtr(current_bytepos);
9471 : }
9472 :
9473 : /*
9474 : * Get latest WAL write pointer
9475 : */
9476 : XLogRecPtr
9477 25660 : GetXLogWriteRecPtr(void)
9478 : {
9479 25660 : RefreshXLogWriteResult(LogwrtResult);
9480 :
9481 25660 : return LogwrtResult.Write;
9482 : }
9483 :
9484 : /*
9485 : * Returns the redo pointer of the last checkpoint or restartpoint. This is
9486 : * the oldest point in WAL that we still need, if we have to restart recovery.
9487 : */
9488 : void
9489 778 : GetOldestRestartPoint(XLogRecPtr *oldrecptr, TimeLineID *oldtli)
9490 : {
9491 778 : LWLockAcquire(ControlFileLock, LW_SHARED);
9492 778 : *oldrecptr = ControlFile->checkPointCopy.redo;
9493 778 : *oldtli = ControlFile->checkPointCopy.ThisTimeLineID;
9494 778 : LWLockRelease(ControlFileLock);
9495 778 : }
9496 :
9497 : /* Thin wrapper around ShutdownWalRcv(). */
9498 : void
9499 2144 : XLogShutdownWalRcv(void)
9500 : {
9501 2144 : ShutdownWalRcv();
9502 :
9503 2144 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9504 2144 : XLogCtl->InstallXLogFileSegmentActive = false;
9505 2144 : LWLockRelease(ControlFileLock);
9506 2144 : }
9507 :
9508 : /* Enable WAL file recycling and preallocation. */
9509 : void
9510 2254 : SetInstallXLogFileSegmentActive(void)
9511 : {
9512 2254 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9513 2254 : XLogCtl->InstallXLogFileSegmentActive = true;
9514 2254 : LWLockRelease(ControlFileLock);
9515 2254 : }
9516 :
9517 : bool
9518 0 : IsInstallXLogFileSegmentActive(void)
9519 : {
9520 : bool result;
9521 :
9522 0 : LWLockAcquire(ControlFileLock, LW_SHARED);
9523 0 : result = XLogCtl->InstallXLogFileSegmentActive;
9524 0 : LWLockRelease(ControlFileLock);
9525 :
9526 0 : return result;
9527 : }
9528 :
9529 : /*
9530 : * Update the WalWriterSleeping flag.
9531 : */
9532 : void
9533 992 : SetWalWriterSleeping(bool sleeping)
9534 : {
9535 992 : SpinLockAcquire(&XLogCtl->info_lck);
9536 992 : XLogCtl->WalWriterSleeping = sleeping;
9537 992 : SpinLockRelease(&XLogCtl->info_lck);
9538 992 : }
|
