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/ps_status.h"
100 : #include "utils/relmapper.h"
101 : #include "utils/snapmgr.h"
102 : #include "utils/timeout.h"
103 : #include "utils/timestamp.h"
104 : #include "utils/varlena.h"
105 :
106 : #ifdef WAL_DEBUG
107 : #include "utils/memutils.h"
108 : #endif
109 :
110 : /* timeline ID to be used when bootstrapping */
111 : #define BootstrapTimeLineID 1
112 :
113 : /* User-settable parameters */
114 : int max_wal_size_mb = 1024; /* 1 GB */
115 : int min_wal_size_mb = 80; /* 80 MB */
116 : int wal_keep_size_mb = 0;
117 : int XLOGbuffers = -1;
118 : int XLogArchiveTimeout = 0;
119 : int XLogArchiveMode = ARCHIVE_MODE_OFF;
120 : char *XLogArchiveCommand = NULL;
121 : bool EnableHotStandby = false;
122 : bool fullPageWrites = true;
123 : bool wal_log_hints = false;
124 : int wal_compression = WAL_COMPRESSION_NONE;
125 : char *wal_consistency_checking_string = NULL;
126 : bool *wal_consistency_checking = NULL;
127 : bool wal_init_zero = true;
128 : bool wal_recycle = true;
129 : bool log_checkpoints = true;
130 : int wal_sync_method = DEFAULT_WAL_SYNC_METHOD;
131 : int wal_level = WAL_LEVEL_REPLICA;
132 : int CommitDelay = 0; /* precommit delay in microseconds */
133 : int CommitSiblings = 5; /* # concurrent xacts needed to sleep */
134 : int wal_retrieve_retry_interval = 5000;
135 : int max_slot_wal_keep_size_mb = -1;
136 : int wal_decode_buffer_size = 512 * 1024;
137 : bool track_wal_io_timing = false;
138 :
139 : #ifdef WAL_DEBUG
140 : bool XLOG_DEBUG = false;
141 : #endif
142 :
143 : int wal_segment_size = DEFAULT_XLOG_SEG_SIZE;
144 :
145 : /*
146 : * Number of WAL insertion locks to use. A higher value allows more insertions
147 : * to happen concurrently, but adds some CPU overhead to flushing the WAL,
148 : * which needs to iterate all the locks.
149 : */
150 : #define NUM_XLOGINSERT_LOCKS 8
151 :
152 : /*
153 : * Max distance from last checkpoint, before triggering a new xlog-based
154 : * checkpoint.
155 : */
156 : int CheckPointSegments;
157 :
158 : /* Estimated distance between checkpoints, in bytes */
159 : static double CheckPointDistanceEstimate = 0;
160 : static double PrevCheckPointDistance = 0;
161 :
162 : /*
163 : * Track whether there were any deferred checks for custom resource managers
164 : * specified in wal_consistency_checking.
165 : */
166 : static bool check_wal_consistency_checking_deferred = false;
167 :
168 : /*
169 : * GUC support
170 : */
171 : const struct config_enum_entry wal_sync_method_options[] = {
172 : {"fsync", WAL_SYNC_METHOD_FSYNC, false},
173 : #ifdef HAVE_FSYNC_WRITETHROUGH
174 : {"fsync_writethrough", WAL_SYNC_METHOD_FSYNC_WRITETHROUGH, false},
175 : #endif
176 : {"fdatasync", WAL_SYNC_METHOD_FDATASYNC, false},
177 : #ifdef O_SYNC
178 : {"open_sync", WAL_SYNC_METHOD_OPEN, false},
179 : #endif
180 : #ifdef O_DSYNC
181 : {"open_datasync", WAL_SYNC_METHOD_OPEN_DSYNC, false},
182 : #endif
183 : {NULL, 0, false}
184 : };
185 :
186 :
187 : /*
188 : * Although only "on", "off", and "always" are documented,
189 : * we accept all the likely variants of "on" and "off".
190 : */
191 : const struct config_enum_entry archive_mode_options[] = {
192 : {"always", ARCHIVE_MODE_ALWAYS, false},
193 : {"on", ARCHIVE_MODE_ON, false},
194 : {"off", ARCHIVE_MODE_OFF, false},
195 : {"true", ARCHIVE_MODE_ON, true},
196 : {"false", ARCHIVE_MODE_OFF, true},
197 : {"yes", ARCHIVE_MODE_ON, true},
198 : {"no", ARCHIVE_MODE_OFF, true},
199 : {"1", ARCHIVE_MODE_ON, true},
200 : {"0", ARCHIVE_MODE_OFF, true},
201 : {NULL, 0, false}
202 : };
203 :
204 : /*
205 : * Statistics for current checkpoint are collected in this global struct.
206 : * Because only the checkpointer or a stand-alone backend can perform
207 : * checkpoints, this will be unused in normal backends.
208 : */
209 : CheckpointStatsData CheckpointStats;
210 :
211 : /*
212 : * During recovery, lastFullPageWrites keeps track of full_page_writes that
213 : * the replayed WAL records indicate. It's initialized with full_page_writes
214 : * that the recovery starting checkpoint record indicates, and then updated
215 : * each time XLOG_FPW_CHANGE record is replayed.
216 : */
217 : static bool lastFullPageWrites;
218 :
219 : /*
220 : * Local copy of the state tracked by SharedRecoveryState in shared memory,
221 : * It is false if SharedRecoveryState is RECOVERY_STATE_DONE. True actually
222 : * means "not known, need to check the shared state".
223 : */
224 : static bool LocalRecoveryInProgress = true;
225 :
226 : /*
227 : * Local state for XLogInsertAllowed():
228 : * 1: unconditionally allowed to insert XLOG
229 : * 0: unconditionally not allowed to insert XLOG
230 : * -1: must check RecoveryInProgress(); disallow until it is false
231 : * Most processes start with -1 and transition to 1 after seeing that recovery
232 : * is not in progress. But we can also force the value for special cases.
233 : * The coding in XLogInsertAllowed() depends on the first two of these states
234 : * being numerically the same as bool true and false.
235 : */
236 : static int LocalXLogInsertAllowed = -1;
237 :
238 : /*
239 : * ProcLastRecPtr points to the start of the last XLOG record inserted by the
240 : * current backend. It is updated for all inserts. XactLastRecEnd points to
241 : * end+1 of the last record, and is reset when we end a top-level transaction,
242 : * or start a new one; so it can be used to tell if the current transaction has
243 : * created any XLOG records.
244 : *
245 : * While in parallel mode, this may not be fully up to date. When committing,
246 : * a transaction can assume this covers all xlog records written either by the
247 : * user backend or by any parallel worker which was present at any point during
248 : * the transaction. But when aborting, or when still in parallel mode, other
249 : * parallel backends may have written WAL records at later LSNs than the value
250 : * stored here. The parallel leader advances its own copy, when necessary,
251 : * in WaitForParallelWorkersToFinish.
252 : */
253 : XLogRecPtr ProcLastRecPtr = InvalidXLogRecPtr;
254 : XLogRecPtr XactLastRecEnd = InvalidXLogRecPtr;
255 : XLogRecPtr XactLastCommitEnd = InvalidXLogRecPtr;
256 :
257 : /*
258 : * RedoRecPtr is this backend's local copy of the REDO record pointer
259 : * (which is almost but not quite the same as a pointer to the most recent
260 : * CHECKPOINT record). We update this from the shared-memory copy,
261 : * XLogCtl->Insert.RedoRecPtr, whenever we can safely do so (ie, when we
262 : * hold an insertion lock). See XLogInsertRecord for details. We are also
263 : * allowed to update from XLogCtl->RedoRecPtr if we hold the info_lck;
264 : * see GetRedoRecPtr.
265 : *
266 : * NB: Code that uses this variable must be prepared not only for the
267 : * possibility that it may be arbitrarily out of date, but also for the
268 : * possibility that it might be set to InvalidXLogRecPtr. We used to
269 : * initialize it as a side effect of the first call to RecoveryInProgress(),
270 : * which meant that most code that might use it could assume that it had a
271 : * real if perhaps stale value. That's no longer the case.
272 : */
273 : static XLogRecPtr RedoRecPtr;
274 :
275 : /*
276 : * doPageWrites is this backend's local copy of (fullPageWrites ||
277 : * runningBackups > 0). It is used together with RedoRecPtr to decide whether
278 : * a full-page image of a page need to be taken.
279 : *
280 : * NB: Initially this is false, and there's no guarantee that it will be
281 : * initialized to any other value before it is first used. Any code that
282 : * makes use of it must recheck the value after obtaining a WALInsertLock,
283 : * and respond appropriately if it turns out that the previous value wasn't
284 : * accurate.
285 : */
286 : static bool doPageWrites;
287 :
288 : /*----------
289 : * Shared-memory data structures for XLOG control
290 : *
291 : * LogwrtRqst indicates a byte position that we need to write and/or fsync
292 : * the log up to (all records before that point must be written or fsynced).
293 : * The positions already written/fsynced are maintained in logWriteResult
294 : * and logFlushResult using atomic access.
295 : * In addition to the shared variable, each backend has a private copy of
296 : * both in LogwrtResult, which is updated when convenient.
297 : *
298 : * The request bookkeeping is simpler: there is a shared XLogCtl->LogwrtRqst
299 : * (protected by info_lck), but we don't need to cache any copies of it.
300 : *
301 : * info_lck is only held long enough to read/update the protected variables,
302 : * so it's a plain spinlock. The other locks are held longer (potentially
303 : * over I/O operations), so we use LWLocks for them. These locks are:
304 : *
305 : * WALWriteLock: must be held to write WAL buffers to disk (XLogWrite or
306 : * XLogFlush).
307 : *
308 : * ControlFileLock: must be held to read/update control file or create
309 : * new log file.
310 : *
311 : *----------
312 : */
313 :
314 : typedef struct XLogwrtRqst
315 : {
316 : XLogRecPtr Write; /* last byte + 1 to write out */
317 : XLogRecPtr Flush; /* last byte + 1 to flush */
318 : } XLogwrtRqst;
319 :
320 : typedef struct XLogwrtResult
321 : {
322 : XLogRecPtr Write; /* last byte + 1 written out */
323 : XLogRecPtr Flush; /* last byte + 1 flushed */
324 : } XLogwrtResult;
325 :
326 : /*
327 : * Inserting to WAL is protected by a small fixed number of WAL insertion
328 : * locks. To insert to the WAL, you must hold one of the locks - it doesn't
329 : * matter which one. To lock out other concurrent insertions, you must hold
330 : * of them. Each WAL insertion lock consists of a lightweight lock, plus an
331 : * indicator of how far the insertion has progressed (insertingAt).
332 : *
333 : * The insertingAt values are read when a process wants to flush WAL from
334 : * the in-memory buffers to disk, to check that all the insertions to the
335 : * region the process is about to write out have finished. You could simply
336 : * wait for all currently in-progress insertions to finish, but the
337 : * insertingAt indicator allows you to ignore insertions to later in the WAL,
338 : * so that you only wait for the insertions that are modifying the buffers
339 : * you're about to write out.
340 : *
341 : * This isn't just an optimization. If all the WAL buffers are dirty, an
342 : * inserter that's holding a WAL insert lock might need to evict an old WAL
343 : * buffer, which requires flushing the WAL. If it's possible for an inserter
344 : * to block on another inserter unnecessarily, deadlock can arise when two
345 : * inserters holding a WAL insert lock wait for each other to finish their
346 : * insertion.
347 : *
348 : * Small WAL records that don't cross a page boundary never update the value,
349 : * the WAL record is just copied to the page and the lock is released. But
350 : * to avoid the deadlock-scenario explained above, the indicator is always
351 : * updated before sleeping while holding an insertion lock.
352 : *
353 : * lastImportantAt contains the LSN of the last important WAL record inserted
354 : * using a given lock. This value is used to detect if there has been
355 : * important WAL activity since the last time some action, like a checkpoint,
356 : * was performed - allowing to not repeat the action if not. The LSN is
357 : * updated for all insertions, unless the XLOG_MARK_UNIMPORTANT flag was
358 : * set. lastImportantAt is never cleared, only overwritten by the LSN of newer
359 : * records. Tracking the WAL activity directly in WALInsertLock has the
360 : * advantage of not needing any additional locks to update the value.
361 : */
362 : typedef struct
363 : {
364 : LWLock lock;
365 : pg_atomic_uint64 insertingAt;
366 : XLogRecPtr lastImportantAt;
367 : } WALInsertLock;
368 :
369 : /*
370 : * All the WAL insertion locks are allocated as an array in shared memory. We
371 : * force the array stride to be a power of 2, which saves a few cycles in
372 : * indexing, but more importantly also ensures that individual slots don't
373 : * cross cache line boundaries. (Of course, we have to also ensure that the
374 : * array start address is suitably aligned.)
375 : */
376 : typedef union WALInsertLockPadded
377 : {
378 : WALInsertLock l;
379 : char pad[PG_CACHE_LINE_SIZE];
380 : } WALInsertLockPadded;
381 :
382 : /*
383 : * Session status of running backup, used for sanity checks in SQL-callable
384 : * functions to start and stop backups.
385 : */
386 : static SessionBackupState sessionBackupState = SESSION_BACKUP_NONE;
387 :
388 : /*
389 : * Shared state data for WAL insertion.
390 : */
391 : typedef struct XLogCtlInsert
392 : {
393 : slock_t insertpos_lck; /* protects CurrBytePos and PrevBytePos */
394 :
395 : /*
396 : * CurrBytePos is the end of reserved WAL. The next record will be
397 : * inserted at that position. PrevBytePos is the start position of the
398 : * previously inserted (or rather, reserved) record - it is copied to the
399 : * prev-link of the next record. These are stored as "usable byte
400 : * positions" rather than XLogRecPtrs (see XLogBytePosToRecPtr()).
401 : */
402 : uint64 CurrBytePos;
403 : uint64 PrevBytePos;
404 :
405 : /*
406 : * Make sure the above heavily-contended spinlock and byte positions are
407 : * on their own cache line. In particular, the RedoRecPtr and full page
408 : * write variables below should be on a different cache line. They are
409 : * read on every WAL insertion, but updated rarely, and we don't want
410 : * those reads to steal the cache line containing Curr/PrevBytePos.
411 : */
412 : char pad[PG_CACHE_LINE_SIZE];
413 :
414 : /*
415 : * fullPageWrites is the authoritative value used by all backends to
416 : * determine whether to write full-page image to WAL. This shared value,
417 : * instead of the process-local fullPageWrites, is required because, when
418 : * full_page_writes is changed by SIGHUP, we must WAL-log it before it
419 : * actually affects WAL-logging by backends. Checkpointer sets at startup
420 : * or after SIGHUP.
421 : *
422 : * To read these fields, you must hold an insertion lock. To modify them,
423 : * you must hold ALL the locks.
424 : */
425 : XLogRecPtr RedoRecPtr; /* current redo point for insertions */
426 : bool fullPageWrites;
427 :
428 : /*
429 : * runningBackups is a counter indicating the number of backups currently
430 : * in progress. lastBackupStart is the latest checkpoint redo location
431 : * used as a starting point for an online backup.
432 : */
433 : int runningBackups;
434 : XLogRecPtr lastBackupStart;
435 :
436 : /*
437 : * WAL insertion locks.
438 : */
439 : WALInsertLockPadded *WALInsertLocks;
440 : } XLogCtlInsert;
441 :
442 : /*
443 : * Total shared-memory state for XLOG.
444 : */
445 : typedef struct XLogCtlData
446 : {
447 : XLogCtlInsert Insert;
448 :
449 : /* Protected by info_lck: */
450 : XLogwrtRqst LogwrtRqst;
451 : XLogRecPtr RedoRecPtr; /* a recent copy of Insert->RedoRecPtr */
452 : FullTransactionId ckptFullXid; /* nextXid of latest checkpoint */
453 : XLogRecPtr asyncXactLSN; /* LSN of newest async commit/abort */
454 : XLogRecPtr replicationSlotMinLSN; /* oldest LSN needed by any slot */
455 :
456 : XLogSegNo lastRemovedSegNo; /* latest removed/recycled XLOG segment */
457 :
458 : /* Fake LSN counter, for unlogged relations. */
459 : pg_atomic_uint64 unloggedLSN;
460 :
461 : /* Time and LSN of last xlog segment switch. Protected by WALWriteLock. */
462 : pg_time_t lastSegSwitchTime;
463 : XLogRecPtr lastSegSwitchLSN;
464 :
465 : /* These are accessed using atomics -- info_lck not needed */
466 : pg_atomic_uint64 logInsertResult; /* last byte + 1 inserted to buffers */
467 : pg_atomic_uint64 logWriteResult; /* last byte + 1 written out */
468 : pg_atomic_uint64 logFlushResult; /* last byte + 1 flushed */
469 :
470 : /*
471 : * First initialized page in the cache (first byte position).
472 : */
473 : XLogRecPtr InitializedFrom;
474 :
475 : /*
476 : * Latest reserved for initialization page in the cache (last byte
477 : * position + 1).
478 : *
479 : * To change the identity of a buffer, you need to advance
480 : * InitializeReserved first. To change the identity of a buffer that's
481 : * still dirty, the old page needs to be written out first, and for that
482 : * you need WALWriteLock, and you need to ensure that there are no
483 : * in-progress insertions to the page by calling
484 : * WaitXLogInsertionsToFinish().
485 : */
486 : pg_atomic_uint64 InitializeReserved;
487 :
488 : /*
489 : * Latest initialized page in the cache (last byte position + 1).
490 : *
491 : * InitializedUpTo is updated after the buffer initialization. After
492 : * update, waiters got notification using InitializedUpToCondVar.
493 : */
494 : pg_atomic_uint64 InitializedUpTo;
495 : ConditionVariable InitializedUpToCondVar;
496 :
497 : /*
498 : * These values do not change after startup, although the pointed-to pages
499 : * and xlblocks values certainly do. xlblocks values are changed
500 : * lock-free according to the check for the xlog write position and are
501 : * accompanied by changes of InitializeReserved and InitializedUpTo.
502 : */
503 : char *pages; /* buffers for unwritten XLOG pages */
504 : pg_atomic_uint64 *xlblocks; /* 1st byte ptr-s + XLOG_BLCKSZ */
505 : int XLogCacheBlck; /* highest allocated xlog buffer index */
506 :
507 : /*
508 : * InsertTimeLineID is the timeline into which new WAL is being inserted
509 : * and flushed. It is zero during recovery, and does not change once set.
510 : *
511 : * If we create a new timeline when the system was started up,
512 : * PrevTimeLineID is the old timeline's ID that we forked off from.
513 : * Otherwise it's equal to InsertTimeLineID.
514 : *
515 : * We set these fields while holding info_lck. Most that reads these
516 : * values knows that recovery is no longer in progress and so can safely
517 : * read the value without a lock, but code that could be run either during
518 : * or after recovery can take info_lck while reading these values.
519 : */
520 : TimeLineID InsertTimeLineID;
521 : TimeLineID PrevTimeLineID;
522 :
523 : /*
524 : * SharedRecoveryState indicates if we're still in crash or archive
525 : * recovery. Protected by info_lck.
526 : */
527 : RecoveryState SharedRecoveryState;
528 :
529 : /*
530 : * InstallXLogFileSegmentActive indicates whether the checkpointer should
531 : * arrange for future segments by recycling and/or PreallocXlogFiles().
532 : * Protected by ControlFileLock. Only the startup process changes it. If
533 : * true, anyone can use InstallXLogFileSegment(). If false, the startup
534 : * process owns the exclusive right to install segments, by reading from
535 : * the archive and possibly replacing existing files.
536 : */
537 : bool InstallXLogFileSegmentActive;
538 :
539 : /*
540 : * WalWriterSleeping indicates whether the WAL writer is currently in
541 : * low-power mode (and hence should be nudged if an async commit occurs).
542 : * Protected by info_lck.
543 : */
544 : bool WalWriterSleeping;
545 :
546 : /*
547 : * During recovery, we keep a copy of the latest checkpoint record here.
548 : * lastCheckPointRecPtr points to start of checkpoint record and
549 : * lastCheckPointEndPtr points to end+1 of checkpoint record. Used by the
550 : * checkpointer when it wants to create a restartpoint.
551 : *
552 : * Protected by info_lck.
553 : */
554 : XLogRecPtr lastCheckPointRecPtr;
555 : XLogRecPtr lastCheckPointEndPtr;
556 : CheckPoint lastCheckPoint;
557 :
558 : /*
559 : * lastFpwDisableRecPtr points to the start of the last replayed
560 : * XLOG_FPW_CHANGE record that instructs full_page_writes is disabled.
561 : */
562 : XLogRecPtr lastFpwDisableRecPtr;
563 :
564 : slock_t info_lck; /* locks shared variables shown above */
565 : } XLogCtlData;
566 :
567 : /*
568 : * Classification of XLogInsertRecord operations.
569 : */
570 : typedef enum
571 : {
572 : WALINSERT_NORMAL,
573 : WALINSERT_SPECIAL_SWITCH,
574 : WALINSERT_SPECIAL_CHECKPOINT
575 : } WalInsertClass;
576 :
577 : static XLogCtlData *XLogCtl = NULL;
578 :
579 : /* a private copy of XLogCtl->Insert.WALInsertLocks, for convenience */
580 : static WALInsertLockPadded *WALInsertLocks = NULL;
581 :
582 : /*
583 : * We maintain an image of pg_control in shared memory.
584 : */
585 : static ControlFileData *ControlFile = NULL;
586 :
587 : /*
588 : * Calculate the amount of space left on the page after 'endptr'. Beware
589 : * multiple evaluation!
590 : */
591 : #define INSERT_FREESPACE(endptr) \
592 : (((endptr) % XLOG_BLCKSZ == 0) ? 0 : (XLOG_BLCKSZ - (endptr) % XLOG_BLCKSZ))
593 :
594 : /* Macro to advance to next buffer index. */
595 : #define NextBufIdx(idx) \
596 : (((idx) == XLogCtl->XLogCacheBlck) ? 0 : ((idx) + 1))
597 :
598 : /*
599 : * XLogRecPtrToBufIdx returns the index of the WAL buffer that holds, or
600 : * would hold if it was in cache, the page containing 'recptr'.
601 : */
602 : #define XLogRecPtrToBufIdx(recptr) \
603 : (((recptr) / XLOG_BLCKSZ) % (XLogCtl->XLogCacheBlck + 1))
604 :
605 : /*
606 : * These are the number of bytes in a WAL page usable for WAL data.
607 : */
608 : #define UsableBytesInPage (XLOG_BLCKSZ - SizeOfXLogShortPHD)
609 :
610 : /*
611 : * Convert values of GUCs measured in megabytes to equiv. segment count.
612 : * Rounds down.
613 : */
614 : #define ConvertToXSegs(x, segsize) XLogMBVarToSegs((x), (segsize))
615 :
616 : /* The number of bytes in a WAL segment usable for WAL data. */
617 : static int UsableBytesInSegment;
618 :
619 : /*
620 : * Private, possibly out-of-date copy of shared LogwrtResult.
621 : * See discussion above.
622 : */
623 : static XLogwrtResult LogwrtResult = {0, 0};
624 :
625 : /*
626 : * Update local copy of shared XLogCtl->log{Write,Flush}Result
627 : *
628 : * It's critical that Flush always trails Write, so the order of the reads is
629 : * important, as is the barrier. See also XLogWrite.
630 : */
631 : #define RefreshXLogWriteResult(_target) \
632 : do { \
633 : _target.Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult); \
634 : pg_read_barrier(); \
635 : _target.Write = pg_atomic_read_u64(&XLogCtl->logWriteResult); \
636 : } while (0)
637 :
638 : /*
639 : * openLogFile is -1 or a kernel FD for an open log file segment.
640 : * openLogSegNo identifies the segment, and openLogTLI the corresponding TLI.
641 : * These variables are only used to write the XLOG, and so will normally refer
642 : * to the active segment.
643 : *
644 : * Note: call Reserve/ReleaseExternalFD to track consumption of this FD.
645 : */
646 : static int openLogFile = -1;
647 : static XLogSegNo openLogSegNo = 0;
648 : static TimeLineID openLogTLI = 0;
649 :
650 : /*
651 : * Local copies of equivalent fields in the control file. When running
652 : * crash recovery, LocalMinRecoveryPoint is set to InvalidXLogRecPtr as we
653 : * expect to replay all the WAL available, and updateMinRecoveryPoint is
654 : * switched to false to prevent any updates while replaying records.
655 : * Those values are kept consistent as long as crash recovery runs.
656 : */
657 : static XLogRecPtr LocalMinRecoveryPoint;
658 : static TimeLineID LocalMinRecoveryPointTLI;
659 : static bool updateMinRecoveryPoint = true;
660 :
661 : /* For WALInsertLockAcquire/Release functions */
662 : static int MyLockNo = 0;
663 : static bool holdingAllLocks = false;
664 :
665 : #ifdef WAL_DEBUG
666 : static MemoryContext walDebugCxt = NULL;
667 : #endif
668 :
669 : static void CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI,
670 : XLogRecPtr EndOfLog,
671 : TimeLineID newTLI);
672 : static void CheckRequiredParameterValues(void);
673 : static void XLogReportParameters(void);
674 : static int LocalSetXLogInsertAllowed(void);
675 : static void CreateEndOfRecoveryRecord(void);
676 : static XLogRecPtr CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn,
677 : XLogRecPtr pagePtr,
678 : TimeLineID newTLI);
679 : static void CheckPointGuts(XLogRecPtr checkPointRedo, int flags);
680 : static void KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo);
681 : static XLogRecPtr XLogGetReplicationSlotMinimumLSN(void);
682 :
683 : static void AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli,
684 : bool opportunistic);
685 : static void XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible);
686 : static bool InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
687 : bool find_free, XLogSegNo max_segno,
688 : TimeLineID tli);
689 : static void XLogFileClose(void);
690 : static void PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli);
691 : static void RemoveTempXlogFiles(void);
692 : static void RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr,
693 : XLogRecPtr endptr, TimeLineID insertTLI);
694 : static void RemoveXlogFile(const struct dirent *segment_de,
695 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
696 : TimeLineID insertTLI);
697 : static void UpdateLastRemovedPtr(char *filename);
698 : static void ValidateXLOGDirectoryStructure(void);
699 : static void CleanupBackupHistory(void);
700 : static void UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force);
701 : static bool PerformRecoveryXLogAction(void);
702 : static void InitControlFile(uint64 sysidentifier, uint32 data_checksum_version);
703 : static void WriteControlFile(void);
704 : static void ReadControlFile(void);
705 : static void UpdateControlFile(void);
706 : static char *str_time(pg_time_t tnow);
707 :
708 : static int get_sync_bit(int method);
709 :
710 : static void CopyXLogRecordToWAL(int write_len, bool isLogSwitch,
711 : XLogRecData *rdata,
712 : XLogRecPtr StartPos, XLogRecPtr EndPos,
713 : TimeLineID tli);
714 : static void ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos,
715 : XLogRecPtr *EndPos, XLogRecPtr *PrevPtr);
716 : static bool ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos,
717 : XLogRecPtr *PrevPtr);
718 : static XLogRecPtr WaitXLogInsertionsToFinish(XLogRecPtr upto);
719 : static char *GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli);
720 : static XLogRecPtr XLogBytePosToRecPtr(uint64 bytepos);
721 : static XLogRecPtr XLogBytePosToEndRecPtr(uint64 bytepos);
722 : static uint64 XLogRecPtrToBytePos(XLogRecPtr ptr);
723 :
724 : static void WALInsertLockAcquire(void);
725 : static void WALInsertLockAcquireExclusive(void);
726 : static void WALInsertLockRelease(void);
727 : static void WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt);
728 :
729 : /*
730 : * Insert an XLOG record represented by an already-constructed chain of data
731 : * chunks. This is a low-level routine; to construct the WAL record header
732 : * and data, use the higher-level routines in xloginsert.c.
733 : *
734 : * If 'fpw_lsn' is valid, it is the oldest LSN among the pages that this
735 : * WAL record applies to, that were not included in the record as full page
736 : * images. If fpw_lsn <= RedoRecPtr, the function does not perform the
737 : * insertion and returns InvalidXLogRecPtr. The caller can then recalculate
738 : * which pages need a full-page image, and retry. If fpw_lsn is invalid, the
739 : * record is always inserted.
740 : *
741 : * 'flags' gives more in-depth control on the record being inserted. See
742 : * XLogSetRecordFlags() for details.
743 : *
744 : * 'topxid_included' tells whether the top-transaction id is logged along with
745 : * current subtransaction. See XLogRecordAssemble().
746 : *
747 : * The first XLogRecData in the chain must be for the record header, and its
748 : * data must be MAXALIGNed. XLogInsertRecord fills in the xl_prev and
749 : * xl_crc fields in the header, the rest of the header must already be filled
750 : * by the caller.
751 : *
752 : * Returns XLOG pointer to end of record (beginning of next record).
753 : * This can be used as LSN for data pages affected by the logged action.
754 : * (LSN is the XLOG point up to which the XLOG must be flushed to disk
755 : * before the data page can be written out. This implements the basic
756 : * WAL rule "write the log before the data".)
757 : */
758 : XLogRecPtr
759 29240078 : XLogInsertRecord(XLogRecData *rdata,
760 : XLogRecPtr fpw_lsn,
761 : uint8 flags,
762 : int num_fpi,
763 : bool topxid_included)
764 : {
765 29240078 : XLogCtlInsert *Insert = &XLogCtl->Insert;
766 : pg_crc32c rdata_crc;
767 : bool inserted;
768 29240078 : XLogRecord *rechdr = (XLogRecord *) rdata->data;
769 29240078 : uint8 info = rechdr->xl_info & ~XLR_INFO_MASK;
770 29240078 : WalInsertClass class = WALINSERT_NORMAL;
771 : XLogRecPtr StartPos;
772 : XLogRecPtr EndPos;
773 29240078 : bool prevDoPageWrites = doPageWrites;
774 : TimeLineID insertTLI;
775 :
776 : /* Does this record type require special handling? */
777 29240078 : if (unlikely(rechdr->xl_rmid == RM_XLOG_ID))
778 : {
779 442808 : if (info == XLOG_SWITCH)
780 1464 : class = WALINSERT_SPECIAL_SWITCH;
781 441344 : else if (info == XLOG_CHECKPOINT_REDO)
782 1806 : class = WALINSERT_SPECIAL_CHECKPOINT;
783 : }
784 :
785 : /* we assume that all of the record header is in the first chunk */
786 : Assert(rdata->len >= SizeOfXLogRecord);
787 :
788 : /* cross-check on whether we should be here or not */
789 29240078 : if (!XLogInsertAllowed())
790 0 : elog(ERROR, "cannot make new WAL entries during recovery");
791 :
792 : /*
793 : * Given that we're not in recovery, InsertTimeLineID is set and can't
794 : * change, so we can read it without a lock.
795 : */
796 29240078 : insertTLI = XLogCtl->InsertTimeLineID;
797 :
798 : /*----------
799 : *
800 : * We have now done all the preparatory work we can without holding a
801 : * lock or modifying shared state. From here on, inserting the new WAL
802 : * record to the shared WAL buffer cache is a two-step process:
803 : *
804 : * 1. Reserve the right amount of space from the WAL. The current head of
805 : * reserved space is kept in Insert->CurrBytePos, and is protected by
806 : * insertpos_lck.
807 : *
808 : * 2. Copy the record to the reserved WAL space. This involves finding the
809 : * correct WAL buffer containing the reserved space, and copying the
810 : * record in place. This can be done concurrently in multiple processes.
811 : *
812 : * To keep track of which insertions are still in-progress, each concurrent
813 : * inserter acquires an insertion lock. In addition to just indicating that
814 : * an insertion is in progress, the lock tells others how far the inserter
815 : * has progressed. There is a small fixed number of insertion locks,
816 : * determined by NUM_XLOGINSERT_LOCKS. When an inserter crosses a page
817 : * boundary, it updates the value stored in the lock to the how far it has
818 : * inserted, to allow the previous buffer to be flushed.
819 : *
820 : * Holding onto an insertion lock also protects RedoRecPtr and
821 : * fullPageWrites from changing until the insertion is finished.
822 : *
823 : * Step 2 can usually be done completely in parallel. If the required WAL
824 : * page is not initialized yet, you have to go through AdvanceXLInsertBuffer,
825 : * which will ensure it is initialized. But the WAL writer tries to do that
826 : * ahead of insertions to avoid that from happening in the critical path.
827 : *
828 : *----------
829 : */
830 29240078 : START_CRIT_SECTION();
831 :
832 29240078 : if (likely(class == WALINSERT_NORMAL))
833 : {
834 29236808 : WALInsertLockAcquire();
835 :
836 : /*
837 : * Check to see if my copy of RedoRecPtr is out of date. If so, may
838 : * have to go back and have the caller recompute everything. This can
839 : * only happen just after a checkpoint, so it's better to be slow in
840 : * this case and fast otherwise.
841 : *
842 : * Also check to see if fullPageWrites was just turned on or there's a
843 : * running backup (which forces full-page writes); if we weren't
844 : * already doing full-page writes then go back and recompute.
845 : *
846 : * If we aren't doing full-page writes then RedoRecPtr doesn't
847 : * actually affect the contents of the XLOG record, so we'll update
848 : * our local copy but not force a recomputation. (If doPageWrites was
849 : * just turned off, we could recompute the record without full pages,
850 : * but we choose not to bother.)
851 : */
852 29236808 : if (RedoRecPtr != Insert->RedoRecPtr)
853 : {
854 : Assert(RedoRecPtr < Insert->RedoRecPtr);
855 13464 : RedoRecPtr = Insert->RedoRecPtr;
856 : }
857 29236808 : doPageWrites = (Insert->fullPageWrites || Insert->runningBackups > 0);
858 :
859 29236808 : if (doPageWrites &&
860 28756322 : (!prevDoPageWrites ||
861 26390898 : (fpw_lsn != InvalidXLogRecPtr && fpw_lsn <= RedoRecPtr)))
862 : {
863 : /*
864 : * Oops, some buffer now needs to be backed up that the caller
865 : * didn't back up. Start over.
866 : */
867 14804 : WALInsertLockRelease();
868 14804 : END_CRIT_SECTION();
869 14804 : return InvalidXLogRecPtr;
870 : }
871 :
872 : /*
873 : * Reserve space for the record in the WAL. This also sets the xl_prev
874 : * pointer.
875 : */
876 29222004 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
877 : &rechdr->xl_prev);
878 :
879 : /* Normal records are always inserted. */
880 29222004 : inserted = true;
881 : }
882 3270 : else if (class == WALINSERT_SPECIAL_SWITCH)
883 : {
884 : /*
885 : * In order to insert an XLOG_SWITCH record, we need to hold all of
886 : * the WAL insertion locks, not just one, so that no one else can
887 : * begin inserting a record until we've figured out how much space
888 : * remains in the current WAL segment and claimed all of it.
889 : *
890 : * Nonetheless, this case is simpler than the normal cases handled
891 : * below, which must check for changes in doPageWrites and RedoRecPtr.
892 : * Those checks are only needed for records that can contain buffer
893 : * references, and an XLOG_SWITCH record never does.
894 : */
895 : Assert(fpw_lsn == InvalidXLogRecPtr);
896 1464 : WALInsertLockAcquireExclusive();
897 1464 : inserted = ReserveXLogSwitch(&StartPos, &EndPos, &rechdr->xl_prev);
898 : }
899 : else
900 : {
901 : Assert(class == WALINSERT_SPECIAL_CHECKPOINT);
902 :
903 : /*
904 : * We need to update both the local and shared copies of RedoRecPtr,
905 : * which means that we need to hold all the WAL insertion locks.
906 : * However, there can't be any buffer references, so as above, we need
907 : * not check RedoRecPtr before inserting the record; we just need to
908 : * update it afterwards.
909 : */
910 : Assert(fpw_lsn == InvalidXLogRecPtr);
911 1806 : WALInsertLockAcquireExclusive();
912 1806 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
913 : &rechdr->xl_prev);
914 1806 : RedoRecPtr = Insert->RedoRecPtr = StartPos;
915 1806 : inserted = true;
916 : }
917 :
918 29225274 : if (inserted)
919 : {
920 : /*
921 : * Now that xl_prev has been filled in, calculate CRC of the record
922 : * header.
923 : */
924 29225164 : rdata_crc = rechdr->xl_crc;
925 29225164 : COMP_CRC32C(rdata_crc, rechdr, offsetof(XLogRecord, xl_crc));
926 29225164 : FIN_CRC32C(rdata_crc);
927 29225164 : rechdr->xl_crc = rdata_crc;
928 :
929 : /*
930 : * All the record data, including the header, is now ready to be
931 : * inserted. Copy the record in the space reserved.
932 : */
933 29225164 : CopyXLogRecordToWAL(rechdr->xl_tot_len,
934 : class == WALINSERT_SPECIAL_SWITCH, rdata,
935 : StartPos, EndPos, insertTLI);
936 :
937 : /*
938 : * Unless record is flagged as not important, update LSN of last
939 : * important record in the current slot. When holding all locks, just
940 : * update the first one.
941 : */
942 29225164 : if ((flags & XLOG_MARK_UNIMPORTANT) == 0)
943 : {
944 29045552 : int lockno = holdingAllLocks ? 0 : MyLockNo;
945 :
946 29045552 : WALInsertLocks[lockno].l.lastImportantAt = StartPos;
947 : }
948 : }
949 : else
950 : {
951 : /*
952 : * This was an xlog-switch record, but the current insert location was
953 : * already exactly at the beginning of a segment, so there was no need
954 : * to do anything.
955 : */
956 : }
957 :
958 : /*
959 : * Done! Let others know that we're finished.
960 : */
961 29225274 : WALInsertLockRelease();
962 :
963 29225274 : END_CRIT_SECTION();
964 :
965 29225274 : MarkCurrentTransactionIdLoggedIfAny();
966 :
967 : /*
968 : * Mark top transaction id is logged (if needed) so that we should not try
969 : * to log it again with the next WAL record in the current subtransaction.
970 : */
971 29225274 : if (topxid_included)
972 438 : MarkSubxactTopXidLogged();
973 :
974 : /*
975 : * Update shared LogwrtRqst.Write, if we crossed page boundary.
976 : */
977 29225274 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
978 : {
979 3307284 : SpinLockAcquire(&XLogCtl->info_lck);
980 : /* advance global request to include new block(s) */
981 3307284 : if (XLogCtl->LogwrtRqst.Write < EndPos)
982 3165688 : XLogCtl->LogwrtRqst.Write = EndPos;
983 3307284 : SpinLockRelease(&XLogCtl->info_lck);
984 3307284 : RefreshXLogWriteResult(LogwrtResult);
985 : }
986 :
987 : /*
988 : * If this was an XLOG_SWITCH record, flush the record and the empty
989 : * padding space that fills the rest of the segment, and perform
990 : * end-of-segment actions (eg, notifying archiver).
991 : */
992 29225274 : if (class == WALINSERT_SPECIAL_SWITCH)
993 : {
994 : TRACE_POSTGRESQL_WAL_SWITCH();
995 1464 : XLogFlush(EndPos);
996 :
997 : /*
998 : * Even though we reserved the rest of the segment for us, which is
999 : * reflected in EndPos, we return a pointer to just the end of the
1000 : * xlog-switch record.
1001 : */
1002 1464 : if (inserted)
1003 : {
1004 1354 : EndPos = StartPos + SizeOfXLogRecord;
1005 1354 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
1006 : {
1007 0 : uint64 offset = XLogSegmentOffset(EndPos, wal_segment_size);
1008 :
1009 0 : if (offset == EndPos % XLOG_BLCKSZ)
1010 0 : EndPos += SizeOfXLogLongPHD;
1011 : else
1012 0 : EndPos += SizeOfXLogShortPHD;
1013 : }
1014 : }
1015 : }
1016 :
1017 : #ifdef WAL_DEBUG
1018 : if (XLOG_DEBUG)
1019 : {
1020 : static XLogReaderState *debug_reader = NULL;
1021 : XLogRecord *record;
1022 : DecodedXLogRecord *decoded;
1023 : StringInfoData buf;
1024 : StringInfoData recordBuf;
1025 : char *errormsg = NULL;
1026 : MemoryContext oldCxt;
1027 :
1028 : oldCxt = MemoryContextSwitchTo(walDebugCxt);
1029 :
1030 : initStringInfo(&buf);
1031 : appendStringInfo(&buf, "INSERT @ %X/%08X: ", LSN_FORMAT_ARGS(EndPos));
1032 :
1033 : /*
1034 : * We have to piece together the WAL record data from the XLogRecData
1035 : * entries, so that we can pass it to the rm_desc function as one
1036 : * contiguous chunk.
1037 : */
1038 : initStringInfo(&recordBuf);
1039 : for (; rdata != NULL; rdata = rdata->next)
1040 : appendBinaryStringInfo(&recordBuf, rdata->data, rdata->len);
1041 :
1042 : /* We also need temporary space to decode the record. */
1043 : record = (XLogRecord *) recordBuf.data;
1044 : decoded = (DecodedXLogRecord *)
1045 : palloc(DecodeXLogRecordRequiredSpace(record->xl_tot_len));
1046 :
1047 : if (!debug_reader)
1048 : debug_reader = XLogReaderAllocate(wal_segment_size, NULL,
1049 : XL_ROUTINE(.page_read = NULL,
1050 : .segment_open = NULL,
1051 : .segment_close = NULL),
1052 : NULL);
1053 : if (!debug_reader)
1054 : {
1055 : appendStringInfoString(&buf, "error decoding record: out of memory while allocating a WAL reading processor");
1056 : }
1057 : else if (!DecodeXLogRecord(debug_reader,
1058 : decoded,
1059 : record,
1060 : EndPos,
1061 : &errormsg))
1062 : {
1063 : appendStringInfo(&buf, "error decoding record: %s",
1064 : errormsg ? errormsg : "no error message");
1065 : }
1066 : else
1067 : {
1068 : appendStringInfoString(&buf, " - ");
1069 :
1070 : debug_reader->record = decoded;
1071 : xlog_outdesc(&buf, debug_reader);
1072 : debug_reader->record = NULL;
1073 : }
1074 : elog(LOG, "%s", buf.data);
1075 :
1076 : pfree(decoded);
1077 : pfree(buf.data);
1078 : pfree(recordBuf.data);
1079 : MemoryContextSwitchTo(oldCxt);
1080 : }
1081 : #endif
1082 :
1083 : /*
1084 : * Update our global variables
1085 : */
1086 29225274 : ProcLastRecPtr = StartPos;
1087 29225274 : XactLastRecEnd = EndPos;
1088 :
1089 : /* Report WAL traffic to the instrumentation. */
1090 29225274 : if (inserted)
1091 : {
1092 29225164 : pgWalUsage.wal_bytes += rechdr->xl_tot_len;
1093 29225164 : pgWalUsage.wal_records++;
1094 29225164 : pgWalUsage.wal_fpi += num_fpi;
1095 : }
1096 :
1097 29225274 : return EndPos;
1098 : }
1099 :
1100 : /*
1101 : * Reserves the right amount of space for a record of given size from the WAL.
1102 : * *StartPos is set to the beginning of the reserved section, *EndPos to
1103 : * its end+1. *PrevPtr is set to the beginning of the previous record; it is
1104 : * used to set the xl_prev of this record.
1105 : *
1106 : * This is the performance critical part of XLogInsert that must be serialized
1107 : * across backends. The rest can happen mostly in parallel. Try to keep this
1108 : * section as short as possible, insertpos_lck can be heavily contended on a
1109 : * busy system.
1110 : *
1111 : * NB: The space calculation here must match the code in CopyXLogRecordToWAL,
1112 : * where we actually copy the record to the reserved space.
1113 : *
1114 : * NB: Testing shows that XLogInsertRecord runs faster if this code is inlined;
1115 : * however, because there are two call sites, the compiler is reluctant to
1116 : * inline. We use pg_attribute_always_inline here to try to convince it.
1117 : */
1118 : static pg_attribute_always_inline void
1119 29223810 : ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1120 : XLogRecPtr *PrevPtr)
1121 : {
1122 29223810 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1123 : uint64 startbytepos;
1124 : uint64 endbytepos;
1125 : uint64 prevbytepos;
1126 :
1127 29223810 : size = MAXALIGN(size);
1128 :
1129 : /* All (non xlog-switch) records should contain data. */
1130 : Assert(size > SizeOfXLogRecord);
1131 :
1132 : /*
1133 : * The duration the spinlock needs to be held is minimized by minimizing
1134 : * the calculations that have to be done while holding the lock. The
1135 : * current tip of reserved WAL is kept in CurrBytePos, as a byte position
1136 : * that only counts "usable" bytes in WAL, that is, it excludes all WAL
1137 : * page headers. The mapping between "usable" byte positions and physical
1138 : * positions (XLogRecPtrs) can be done outside the locked region, and
1139 : * because the usable byte position doesn't include any headers, reserving
1140 : * X bytes from WAL is almost as simple as "CurrBytePos += X".
1141 : */
1142 29223810 : SpinLockAcquire(&Insert->insertpos_lck);
1143 :
1144 29223810 : startbytepos = Insert->CurrBytePos;
1145 29223810 : endbytepos = startbytepos + size;
1146 29223810 : prevbytepos = Insert->PrevBytePos;
1147 29223810 : Insert->CurrBytePos = endbytepos;
1148 29223810 : Insert->PrevBytePos = startbytepos;
1149 :
1150 29223810 : SpinLockRelease(&Insert->insertpos_lck);
1151 :
1152 29223810 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1153 29223810 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1154 29223810 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1155 :
1156 : /*
1157 : * Check that the conversions between "usable byte positions" and
1158 : * XLogRecPtrs work consistently in both directions.
1159 : */
1160 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1161 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1162 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1163 29223810 : }
1164 :
1165 : /*
1166 : * Like ReserveXLogInsertLocation(), but for an xlog-switch record.
1167 : *
1168 : * A log-switch record is handled slightly differently. The rest of the
1169 : * segment will be reserved for this insertion, as indicated by the returned
1170 : * *EndPos value. However, if we are already at the beginning of the current
1171 : * segment, *StartPos and *EndPos are set to the current location without
1172 : * reserving any space, and the function returns false.
1173 : */
1174 : static bool
1175 1464 : ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1176 : {
1177 1464 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1178 : uint64 startbytepos;
1179 : uint64 endbytepos;
1180 : uint64 prevbytepos;
1181 1464 : uint32 size = MAXALIGN(SizeOfXLogRecord);
1182 : XLogRecPtr ptr;
1183 : uint32 segleft;
1184 :
1185 : /*
1186 : * These calculations are a bit heavy-weight to be done while holding a
1187 : * spinlock, but since we're holding all the WAL insertion locks, there
1188 : * are no other inserters competing for it. GetXLogInsertRecPtr() does
1189 : * compete for it, but that's not called very frequently.
1190 : */
1191 1464 : SpinLockAcquire(&Insert->insertpos_lck);
1192 :
1193 1464 : startbytepos = Insert->CurrBytePos;
1194 :
1195 1464 : ptr = XLogBytePosToEndRecPtr(startbytepos);
1196 1464 : if (XLogSegmentOffset(ptr, wal_segment_size) == 0)
1197 : {
1198 110 : SpinLockRelease(&Insert->insertpos_lck);
1199 110 : *EndPos = *StartPos = ptr;
1200 110 : return false;
1201 : }
1202 :
1203 1354 : endbytepos = startbytepos + size;
1204 1354 : prevbytepos = Insert->PrevBytePos;
1205 :
1206 1354 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1207 1354 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1208 :
1209 1354 : segleft = wal_segment_size - XLogSegmentOffset(*EndPos, wal_segment_size);
1210 1354 : if (segleft != wal_segment_size)
1211 : {
1212 : /* consume the rest of the segment */
1213 1354 : *EndPos += segleft;
1214 1354 : endbytepos = XLogRecPtrToBytePos(*EndPos);
1215 : }
1216 1354 : Insert->CurrBytePos = endbytepos;
1217 1354 : Insert->PrevBytePos = startbytepos;
1218 :
1219 1354 : SpinLockRelease(&Insert->insertpos_lck);
1220 :
1221 1354 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1222 :
1223 : Assert(XLogSegmentOffset(*EndPos, wal_segment_size) == 0);
1224 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1225 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1226 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1227 :
1228 1354 : return true;
1229 : }
1230 :
1231 : /*
1232 : * Subroutine of XLogInsertRecord. Copies a WAL record to an already-reserved
1233 : * area in the WAL.
1234 : */
1235 : static void
1236 29225164 : CopyXLogRecordToWAL(int write_len, bool isLogSwitch, XLogRecData *rdata,
1237 : XLogRecPtr StartPos, XLogRecPtr EndPos, TimeLineID tli)
1238 : {
1239 : char *currpos;
1240 : int freespace;
1241 : int written;
1242 : XLogRecPtr CurrPos;
1243 : XLogPageHeader pagehdr;
1244 :
1245 : /*
1246 : * Get a pointer to the right place in the right WAL buffer to start
1247 : * inserting to.
1248 : */
1249 29225164 : CurrPos = StartPos;
1250 29225164 : currpos = GetXLogBuffer(CurrPos, tli);
1251 29225164 : freespace = INSERT_FREESPACE(CurrPos);
1252 :
1253 : /*
1254 : * there should be enough space for at least the first field (xl_tot_len)
1255 : * on this page.
1256 : */
1257 : Assert(freespace >= sizeof(uint32));
1258 :
1259 : /* Copy record data */
1260 29225164 : written = 0;
1261 137974570 : while (rdata != NULL)
1262 : {
1263 108749406 : const char *rdata_data = rdata->data;
1264 108749406 : int rdata_len = rdata->len;
1265 :
1266 112296460 : while (rdata_len > freespace)
1267 : {
1268 : /*
1269 : * Write what fits on this page, and continue on the next page.
1270 : */
1271 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || freespace == 0);
1272 3547054 : memcpy(currpos, rdata_data, freespace);
1273 3547054 : rdata_data += freespace;
1274 3547054 : rdata_len -= freespace;
1275 3547054 : written += freespace;
1276 3547054 : CurrPos += freespace;
1277 :
1278 : /*
1279 : * Get pointer to beginning of next page, and set the xlp_rem_len
1280 : * in the page header. Set XLP_FIRST_IS_CONTRECORD.
1281 : *
1282 : * It's safe to set the contrecord flag and xlp_rem_len without a
1283 : * lock on the page. All the other flags were already set when the
1284 : * page was initialized, in AdvanceXLInsertBuffer, and we're the
1285 : * only backend that needs to set the contrecord flag.
1286 : */
1287 3547054 : currpos = GetXLogBuffer(CurrPos, tli);
1288 3547054 : pagehdr = (XLogPageHeader) currpos;
1289 3547054 : pagehdr->xlp_rem_len = write_len - written;
1290 3547054 : pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1291 :
1292 : /* skip over the page header */
1293 3547054 : if (XLogSegmentOffset(CurrPos, wal_segment_size) == 0)
1294 : {
1295 2512 : CurrPos += SizeOfXLogLongPHD;
1296 2512 : currpos += SizeOfXLogLongPHD;
1297 : }
1298 : else
1299 : {
1300 3544542 : CurrPos += SizeOfXLogShortPHD;
1301 3544542 : currpos += SizeOfXLogShortPHD;
1302 : }
1303 3547054 : freespace = INSERT_FREESPACE(CurrPos);
1304 : }
1305 :
1306 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1307 108749406 : memcpy(currpos, rdata_data, rdata_len);
1308 108749406 : currpos += rdata_len;
1309 108749406 : CurrPos += rdata_len;
1310 108749406 : freespace -= rdata_len;
1311 108749406 : written += rdata_len;
1312 :
1313 108749406 : rdata = rdata->next;
1314 : }
1315 : Assert(written == write_len);
1316 :
1317 : /*
1318 : * If this was an xlog-switch, it's not enough to write the switch record,
1319 : * we also have to consume all the remaining space in the WAL segment. We
1320 : * have already reserved that space, but we need to actually fill it.
1321 : */
1322 29225164 : if (isLogSwitch && XLogSegmentOffset(CurrPos, wal_segment_size) != 0)
1323 : {
1324 : /* An xlog-switch record doesn't contain any data besides the header */
1325 : Assert(write_len == SizeOfXLogRecord);
1326 :
1327 : /* Assert that we did reserve the right amount of space */
1328 : Assert(XLogSegmentOffset(EndPos, wal_segment_size) == 0);
1329 :
1330 : /* Use up all the remaining space on the current page */
1331 1354 : CurrPos += freespace;
1332 :
1333 : /*
1334 : * Cause all remaining pages in the segment to be flushed, leaving the
1335 : * XLog position where it should be, at the start of the next segment.
1336 : * We do this one page at a time, to make sure we don't deadlock
1337 : * against ourselves if wal_buffers < wal_segment_size.
1338 : */
1339 1209436 : while (CurrPos < EndPos)
1340 : {
1341 : /*
1342 : * The minimal action to flush the page would be to call
1343 : * WALInsertLockUpdateInsertingAt(CurrPos) followed by
1344 : * AdvanceXLInsertBuffer(...). The page would be left initialized
1345 : * mostly to zeros, except for the page header (always the short
1346 : * variant, as this is never a segment's first page).
1347 : *
1348 : * The large vistas of zeros are good for compressibility, but the
1349 : * headers interrupting them every XLOG_BLCKSZ (with values that
1350 : * differ from page to page) are not. The effect varies with
1351 : * compression tool, but bzip2 for instance compresses about an
1352 : * order of magnitude worse if those headers are left in place.
1353 : *
1354 : * Rather than complicating AdvanceXLInsertBuffer itself (which is
1355 : * called in heavily-loaded circumstances as well as this lightly-
1356 : * loaded one) with variant behavior, we just use GetXLogBuffer
1357 : * (which itself calls the two methods we need) to get the pointer
1358 : * and zero most of the page. Then we just zero the page header.
1359 : */
1360 1208082 : currpos = GetXLogBuffer(CurrPos, tli);
1361 4832328 : MemSet(currpos, 0, SizeOfXLogShortPHD);
1362 :
1363 1208082 : CurrPos += XLOG_BLCKSZ;
1364 : }
1365 : }
1366 : else
1367 : {
1368 : /* Align the end position, so that the next record starts aligned */
1369 29223810 : CurrPos = MAXALIGN64(CurrPos);
1370 : }
1371 :
1372 29225164 : if (CurrPos != EndPos)
1373 0 : ereport(PANIC,
1374 : errcode(ERRCODE_DATA_CORRUPTED),
1375 : errmsg_internal("space reserved for WAL record does not match what was written"));
1376 29225164 : }
1377 :
1378 : /*
1379 : * Acquire a WAL insertion lock, for inserting to WAL.
1380 : */
1381 : static void
1382 29236828 : WALInsertLockAcquire(void)
1383 : {
1384 : bool immed;
1385 :
1386 : /*
1387 : * It doesn't matter which of the WAL insertion locks we acquire, so try
1388 : * the one we used last time. If the system isn't particularly busy, it's
1389 : * a good bet that it's still available, and it's good to have some
1390 : * affinity to a particular lock so that you don't unnecessarily bounce
1391 : * cache lines between processes when there's no contention.
1392 : *
1393 : * If this is the first time through in this backend, pick a lock
1394 : * (semi-)randomly. This allows the locks to be used evenly if you have a
1395 : * lot of very short connections.
1396 : */
1397 : static int lockToTry = -1;
1398 :
1399 29236828 : if (lockToTry == -1)
1400 15506 : lockToTry = MyProcNumber % NUM_XLOGINSERT_LOCKS;
1401 29236828 : MyLockNo = lockToTry;
1402 :
1403 : /*
1404 : * The insertingAt value is initially set to 0, as we don't know our
1405 : * insert location yet.
1406 : */
1407 29236828 : immed = LWLockAcquire(&WALInsertLocks[MyLockNo].l.lock, LW_EXCLUSIVE);
1408 29236828 : if (!immed)
1409 : {
1410 : /*
1411 : * If we couldn't get the lock immediately, try another lock next
1412 : * time. On a system with more insertion locks than concurrent
1413 : * inserters, this causes all the inserters to eventually migrate to a
1414 : * lock that no-one else is using. On a system with more inserters
1415 : * than locks, it still helps to distribute the inserters evenly
1416 : * across the locks.
1417 : */
1418 69906 : lockToTry = (lockToTry + 1) % NUM_XLOGINSERT_LOCKS;
1419 : }
1420 29236828 : }
1421 :
1422 : /*
1423 : * Acquire all WAL insertion locks, to prevent other backends from inserting
1424 : * to WAL.
1425 : */
1426 : static void
1427 8370 : WALInsertLockAcquireExclusive(void)
1428 : {
1429 : int i;
1430 :
1431 : /*
1432 : * When holding all the locks, all but the last lock's insertingAt
1433 : * indicator is set to 0xFFFFFFFFFFFFFFFF, which is higher than any real
1434 : * XLogRecPtr value, to make sure that no-one blocks waiting on those.
1435 : */
1436 66960 : for (i = 0; i < NUM_XLOGINSERT_LOCKS - 1; i++)
1437 : {
1438 58590 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1439 58590 : LWLockUpdateVar(&WALInsertLocks[i].l.lock,
1440 58590 : &WALInsertLocks[i].l.insertingAt,
1441 : PG_UINT64_MAX);
1442 : }
1443 : /* Variable value reset to 0 at release */
1444 8370 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1445 :
1446 8370 : holdingAllLocks = true;
1447 8370 : }
1448 :
1449 : /*
1450 : * Release our insertion lock (or locks, if we're holding them all).
1451 : *
1452 : * NB: Reset all variables to 0, so they cause LWLockWaitForVar to block the
1453 : * next time the lock is acquired.
1454 : */
1455 : static void
1456 29245198 : WALInsertLockRelease(void)
1457 : {
1458 29245198 : if (holdingAllLocks)
1459 : {
1460 : int i;
1461 :
1462 75330 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1463 66960 : LWLockReleaseClearVar(&WALInsertLocks[i].l.lock,
1464 66960 : &WALInsertLocks[i].l.insertingAt,
1465 : 0);
1466 :
1467 8370 : holdingAllLocks = false;
1468 : }
1469 : else
1470 : {
1471 29236828 : LWLockReleaseClearVar(&WALInsertLocks[MyLockNo].l.lock,
1472 29236828 : &WALInsertLocks[MyLockNo].l.insertingAt,
1473 : 0);
1474 : }
1475 29245198 : }
1476 :
1477 : /*
1478 : * Update our insertingAt value, to let others know that we've finished
1479 : * inserting up to that point.
1480 : */
1481 : static void
1482 5057672 : WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt)
1483 : {
1484 5057672 : if (holdingAllLocks)
1485 : {
1486 : /*
1487 : * We use the last lock to mark our actual position, see comments in
1488 : * WALInsertLockAcquireExclusive.
1489 : */
1490 1204020 : LWLockUpdateVar(&WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.lock,
1491 1204020 : &WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.insertingAt,
1492 : insertingAt);
1493 : }
1494 : else
1495 3853652 : LWLockUpdateVar(&WALInsertLocks[MyLockNo].l.lock,
1496 3853652 : &WALInsertLocks[MyLockNo].l.insertingAt,
1497 : insertingAt);
1498 5057672 : }
1499 :
1500 : /*
1501 : * Wait for any WAL insertions < upto to finish.
1502 : *
1503 : * Returns the location of the oldest insertion that is still in-progress.
1504 : * Any WAL prior to that point has been fully copied into WAL buffers, and
1505 : * can be flushed out to disk. Because this waits for any insertions older
1506 : * than 'upto' to finish, the return value is always >= 'upto'.
1507 : *
1508 : * Note: When you are about to write out WAL, you must call this function
1509 : * *before* acquiring WALWriteLock, to avoid deadlocks. This function might
1510 : * need to wait for an insertion to finish (or at least advance to next
1511 : * uninitialized page), and the inserter might need to evict an old WAL buffer
1512 : * to make room for a new one, which in turn requires WALWriteLock.
1513 : */
1514 : static XLogRecPtr
1515 4096772 : WaitXLogInsertionsToFinish(XLogRecPtr upto)
1516 : {
1517 : uint64 bytepos;
1518 : XLogRecPtr inserted;
1519 : XLogRecPtr reservedUpto;
1520 : XLogRecPtr finishedUpto;
1521 4096772 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1522 : int i;
1523 :
1524 4096772 : if (MyProc == NULL)
1525 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1526 :
1527 : /*
1528 : * Check if there's any work to do. Use a barrier to ensure we get the
1529 : * freshest value.
1530 : */
1531 4096772 : inserted = pg_atomic_read_membarrier_u64(&XLogCtl->logInsertResult);
1532 4096772 : if (upto <= inserted)
1533 3312402 : return inserted;
1534 :
1535 : /* Read the current insert position */
1536 784370 : SpinLockAcquire(&Insert->insertpos_lck);
1537 784370 : bytepos = Insert->CurrBytePos;
1538 784370 : SpinLockRelease(&Insert->insertpos_lck);
1539 784370 : reservedUpto = XLogBytePosToEndRecPtr(bytepos);
1540 :
1541 : /*
1542 : * No-one should request to flush a piece of WAL that hasn't even been
1543 : * reserved yet. However, it can happen if there is a block with a bogus
1544 : * LSN on disk, for example. XLogFlush checks for that situation and
1545 : * complains, but only after the flush. Here we just assume that to mean
1546 : * that all WAL that has been reserved needs to be finished. In this
1547 : * corner-case, the return value can be smaller than 'upto' argument.
1548 : */
1549 784370 : if (upto > reservedUpto)
1550 : {
1551 0 : ereport(LOG,
1552 : errmsg("request to flush past end of generated WAL; request %X/%08X, current position %X/%08X",
1553 : LSN_FORMAT_ARGS(upto), LSN_FORMAT_ARGS(reservedUpto)));
1554 0 : upto = reservedUpto;
1555 : }
1556 :
1557 : /*
1558 : * Loop through all the locks, sleeping on any in-progress insert older
1559 : * than 'upto'.
1560 : *
1561 : * finishedUpto is our return value, indicating the point upto which all
1562 : * the WAL insertions have been finished. Initialize it to the head of
1563 : * reserved WAL, and as we iterate through the insertion locks, back it
1564 : * out for any insertion that's still in progress.
1565 : */
1566 784370 : finishedUpto = reservedUpto;
1567 7059330 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1568 : {
1569 6274960 : XLogRecPtr insertingat = InvalidXLogRecPtr;
1570 :
1571 : do
1572 : {
1573 : /*
1574 : * See if this insertion is in progress. LWLockWaitForVar will
1575 : * wait for the lock to be released, or for the 'value' to be set
1576 : * by a LWLockUpdateVar call. When a lock is initially acquired,
1577 : * its value is 0 (InvalidXLogRecPtr), which means that we don't
1578 : * know where it's inserting yet. We will have to wait for it. If
1579 : * it's a small insertion, the record will most likely fit on the
1580 : * same page and the inserter will release the lock without ever
1581 : * calling LWLockUpdateVar. But if it has to sleep, it will
1582 : * advertise the insertion point with LWLockUpdateVar before
1583 : * sleeping.
1584 : *
1585 : * In this loop we are only waiting for insertions that started
1586 : * before WaitXLogInsertionsToFinish was called. The lack of
1587 : * memory barriers in the loop means that we might see locks as
1588 : * "unused" that have since become used. This is fine because
1589 : * they only can be used for later insertions that we would not
1590 : * want to wait on anyway. Not taking a lock to acquire the
1591 : * current insertingAt value means that we might see older
1592 : * insertingAt values. This is also fine, because if we read a
1593 : * value too old, we will add ourselves to the wait queue, which
1594 : * contains atomic operations.
1595 : */
1596 6379998 : if (LWLockWaitForVar(&WALInsertLocks[i].l.lock,
1597 6379998 : &WALInsertLocks[i].l.insertingAt,
1598 : insertingat, &insertingat))
1599 : {
1600 : /* the lock was free, so no insertion in progress */
1601 4605344 : insertingat = InvalidXLogRecPtr;
1602 4605344 : break;
1603 : }
1604 :
1605 : /*
1606 : * This insertion is still in progress. Have to wait, unless the
1607 : * inserter has proceeded past 'upto'.
1608 : */
1609 1774654 : } while (insertingat < upto);
1610 :
1611 6274960 : if (insertingat != InvalidXLogRecPtr && insertingat < finishedUpto)
1612 626948 : finishedUpto = insertingat;
1613 : }
1614 :
1615 : /*
1616 : * Advance the limit we know to have been inserted and return the freshest
1617 : * value we know of, which might be beyond what we requested if somebody
1618 : * is concurrently doing this with an 'upto' pointer ahead of us.
1619 : */
1620 784370 : finishedUpto = pg_atomic_monotonic_advance_u64(&XLogCtl->logInsertResult,
1621 : finishedUpto);
1622 :
1623 784370 : return finishedUpto;
1624 : }
1625 :
1626 : /*
1627 : * Get a pointer to the right location in the WAL buffer containing the
1628 : * given XLogRecPtr.
1629 : *
1630 : * If the page is not initialized yet, it is initialized. That might require
1631 : * evicting an old dirty buffer from the buffer cache, which means I/O.
1632 : *
1633 : * The caller must ensure that the page containing the requested location
1634 : * isn't evicted yet, and won't be evicted. The way to ensure that is to
1635 : * hold onto a WAL insertion lock with the insertingAt position set to
1636 : * something <= ptr. GetXLogBuffer() will update insertingAt if it needs
1637 : * to evict an old page from the buffer. (This means that once you call
1638 : * GetXLogBuffer() with a given 'ptr', you must not access anything before
1639 : * that point anymore, and must not call GetXLogBuffer() with an older 'ptr'
1640 : * later, because older buffers might be recycled already)
1641 : */
1642 : static char *
1643 33980320 : GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli)
1644 : {
1645 : int idx;
1646 : XLogRecPtr endptr;
1647 : static uint64 cachedPage = 0;
1648 : static char *cachedPos = NULL;
1649 : XLogRecPtr expectedEndPtr;
1650 :
1651 : /*
1652 : * Fast path for the common case that we need to access again the same
1653 : * page as last time.
1654 : */
1655 33980320 : if (ptr / XLOG_BLCKSZ == cachedPage)
1656 : {
1657 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1658 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1659 27783314 : return cachedPos + ptr % XLOG_BLCKSZ;
1660 : }
1661 :
1662 : /*
1663 : * The XLog buffer cache is organized so that a page is always loaded to a
1664 : * particular buffer. That way we can easily calculate the buffer a given
1665 : * page must be loaded into, from the XLogRecPtr alone.
1666 : */
1667 6197006 : idx = XLogRecPtrToBufIdx(ptr);
1668 :
1669 : /*
1670 : * See what page is loaded in the buffer at the moment. It could be the
1671 : * page we're looking for, or something older. It can't be anything newer
1672 : * - that would imply the page we're looking for has already been written
1673 : * out to disk and evicted, and the caller is responsible for making sure
1674 : * that doesn't happen.
1675 : *
1676 : * We don't hold a lock while we read the value. If someone is just about
1677 : * to initialize or has just initialized the page, it's possible that we
1678 : * get InvalidXLogRecPtr. That's ok, we'll grab the mapping lock (in
1679 : * AdvanceXLInsertBuffer) and retry if we see anything other than the page
1680 : * we're looking for.
1681 : */
1682 6197006 : expectedEndPtr = ptr;
1683 6197006 : expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
1684 :
1685 6197006 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1686 6197006 : if (expectedEndPtr != endptr)
1687 : {
1688 : XLogRecPtr initializedUpto;
1689 :
1690 : /*
1691 : * Before calling AdvanceXLInsertBuffer(), which can block, let others
1692 : * know how far we're finished with inserting the record.
1693 : *
1694 : * NB: If 'ptr' points to just after the page header, advertise a
1695 : * position at the beginning of the page rather than 'ptr' itself. If
1696 : * there are no other insertions running, someone might try to flush
1697 : * up to our advertised location. If we advertised a position after
1698 : * the page header, someone might try to flush the page header, even
1699 : * though page might actually not be initialized yet. As the first
1700 : * inserter on the page, we are effectively responsible for making
1701 : * sure that it's initialized, before we let insertingAt to move past
1702 : * the page header.
1703 : */
1704 5057672 : if (ptr % XLOG_BLCKSZ == SizeOfXLogShortPHD &&
1705 12086 : XLogSegmentOffset(ptr, wal_segment_size) > XLOG_BLCKSZ)
1706 12086 : initializedUpto = ptr - SizeOfXLogShortPHD;
1707 5045586 : else if (ptr % XLOG_BLCKSZ == SizeOfXLogLongPHD &&
1708 2434 : XLogSegmentOffset(ptr, wal_segment_size) < XLOG_BLCKSZ)
1709 1136 : initializedUpto = ptr - SizeOfXLogLongPHD;
1710 : else
1711 5044450 : initializedUpto = ptr;
1712 :
1713 5057672 : WALInsertLockUpdateInsertingAt(initializedUpto);
1714 :
1715 5057672 : AdvanceXLInsertBuffer(ptr, tli, false);
1716 5057672 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1717 :
1718 5057672 : if (expectedEndPtr != endptr)
1719 0 : elog(PANIC, "could not find WAL buffer for %X/%08X",
1720 : LSN_FORMAT_ARGS(ptr));
1721 : }
1722 : else
1723 : {
1724 : /*
1725 : * Make sure the initialization of the page is visible to us, and
1726 : * won't arrive later to overwrite the WAL data we write on the page.
1727 : */
1728 1139334 : pg_memory_barrier();
1729 : }
1730 :
1731 : /*
1732 : * Found the buffer holding this page. Return a pointer to the right
1733 : * offset within the page.
1734 : */
1735 6197006 : cachedPage = ptr / XLOG_BLCKSZ;
1736 6197006 : cachedPos = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1737 :
1738 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1739 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1740 :
1741 6197006 : return cachedPos + ptr % XLOG_BLCKSZ;
1742 : }
1743 :
1744 : /*
1745 : * Read WAL data directly from WAL buffers, if available. Returns the number
1746 : * of bytes read successfully.
1747 : *
1748 : * Fewer than 'count' bytes may be read if some of the requested WAL data has
1749 : * already been evicted.
1750 : *
1751 : * No locks are taken.
1752 : *
1753 : * Caller should ensure that it reads no further than LogwrtResult.Write
1754 : * (which should have been updated by the caller when determining how far to
1755 : * read). The 'tli' argument is only used as a convenient safety check so that
1756 : * callers do not read from WAL buffers on a historical timeline.
1757 : */
1758 : Size
1759 194126 : WALReadFromBuffers(char *dstbuf, XLogRecPtr startptr, Size count,
1760 : TimeLineID tli)
1761 : {
1762 194126 : char *pdst = dstbuf;
1763 194126 : XLogRecPtr recptr = startptr;
1764 : XLogRecPtr inserted;
1765 194126 : Size nbytes = count;
1766 :
1767 194126 : if (RecoveryInProgress() || tli != GetWALInsertionTimeLine())
1768 1834 : return 0;
1769 :
1770 : Assert(!XLogRecPtrIsInvalid(startptr));
1771 :
1772 : /*
1773 : * Caller should ensure that the requested data has been inserted into WAL
1774 : * buffers before we try to read it.
1775 : */
1776 192292 : inserted = pg_atomic_read_u64(&XLogCtl->logInsertResult);
1777 192292 : if (startptr + count > inserted)
1778 0 : ereport(ERROR,
1779 : errmsg("cannot read past end of generated WAL: requested %X/%08X, current position %X/%08X",
1780 : LSN_FORMAT_ARGS(startptr + count),
1781 : LSN_FORMAT_ARGS(inserted)));
1782 :
1783 : /*
1784 : * Loop through the buffers without a lock. For each buffer, atomically
1785 : * read and verify the end pointer, then copy the data out, and finally
1786 : * re-read and re-verify the end pointer.
1787 : *
1788 : * Once a page is evicted, it never returns to the WAL buffers, so if the
1789 : * end pointer matches the expected end pointer before and after we copy
1790 : * the data, then the right page must have been present during the data
1791 : * copy. Read barriers are necessary to ensure that the data copy actually
1792 : * happens between the two verification steps.
1793 : *
1794 : * If either verification fails, we simply terminate the loop and return
1795 : * with the data that had been already copied out successfully.
1796 : */
1797 198870 : while (nbytes > 0)
1798 : {
1799 197276 : uint32 offset = recptr % XLOG_BLCKSZ;
1800 197276 : int idx = XLogRecPtrToBufIdx(recptr);
1801 : XLogRecPtr expectedEndPtr;
1802 : XLogRecPtr endptr;
1803 : const char *page;
1804 : const char *psrc;
1805 : Size npagebytes;
1806 :
1807 : /*
1808 : * Calculate the end pointer we expect in the xlblocks array if the
1809 : * correct page is present.
1810 : */
1811 197276 : expectedEndPtr = recptr + (XLOG_BLCKSZ - offset);
1812 :
1813 : /*
1814 : * First verification step: check that the correct page is present in
1815 : * the WAL buffers.
1816 : */
1817 197276 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1818 197276 : if (expectedEndPtr != endptr)
1819 190698 : break;
1820 :
1821 : /*
1822 : * The correct page is present (or was at the time the endptr was
1823 : * read; must re-verify later). Calculate pointer to source data and
1824 : * determine how much data to read from this page.
1825 : */
1826 6578 : page = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1827 6578 : psrc = page + offset;
1828 6578 : npagebytes = Min(nbytes, XLOG_BLCKSZ - offset);
1829 :
1830 : /*
1831 : * Ensure that the data copy and the first verification step are not
1832 : * reordered.
1833 : */
1834 6578 : pg_read_barrier();
1835 :
1836 : /* data copy */
1837 6578 : memcpy(pdst, psrc, npagebytes);
1838 :
1839 : /*
1840 : * Ensure that the data copy and the second verification step are not
1841 : * reordered.
1842 : */
1843 6578 : pg_read_barrier();
1844 :
1845 : /*
1846 : * Second verification step: check that the page we read from wasn't
1847 : * evicted while we were copying the data.
1848 : */
1849 6578 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1850 6578 : if (expectedEndPtr != endptr)
1851 0 : break;
1852 :
1853 6578 : pdst += npagebytes;
1854 6578 : recptr += npagebytes;
1855 6578 : nbytes -= npagebytes;
1856 : }
1857 :
1858 : Assert(pdst - dstbuf <= count);
1859 :
1860 192292 : return pdst - dstbuf;
1861 : }
1862 :
1863 : /*
1864 : * Converts a "usable byte position" to XLogRecPtr. A usable byte position
1865 : * is the position starting from the beginning of WAL, excluding all WAL
1866 : * page headers.
1867 : */
1868 : static XLogRecPtr
1869 58455458 : XLogBytePosToRecPtr(uint64 bytepos)
1870 : {
1871 : uint64 fullsegs;
1872 : uint64 fullpages;
1873 : uint64 bytesleft;
1874 : uint32 seg_offset;
1875 : XLogRecPtr result;
1876 :
1877 58455458 : fullsegs = bytepos / UsableBytesInSegment;
1878 58455458 : bytesleft = bytepos % UsableBytesInSegment;
1879 :
1880 58455458 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1881 : {
1882 : /* fits on first page of segment */
1883 104614 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1884 : }
1885 : else
1886 : {
1887 : /* account for the first page on segment with long header */
1888 58350844 : seg_offset = XLOG_BLCKSZ;
1889 58350844 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1890 :
1891 58350844 : fullpages = bytesleft / UsableBytesInPage;
1892 58350844 : bytesleft = bytesleft % UsableBytesInPage;
1893 :
1894 58350844 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1895 : }
1896 :
1897 58455458 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1898 :
1899 58455458 : return result;
1900 : }
1901 :
1902 : /*
1903 : * Like XLogBytePosToRecPtr, but if the position is at a page boundary,
1904 : * returns a pointer to the beginning of the page (ie. before page header),
1905 : * not to where the first xlog record on that page would go to. This is used
1906 : * when converting a pointer to the end of a record.
1907 : */
1908 : static XLogRecPtr
1909 30010998 : XLogBytePosToEndRecPtr(uint64 bytepos)
1910 : {
1911 : uint64 fullsegs;
1912 : uint64 fullpages;
1913 : uint64 bytesleft;
1914 : uint32 seg_offset;
1915 : XLogRecPtr result;
1916 :
1917 30010998 : fullsegs = bytepos / UsableBytesInSegment;
1918 30010998 : bytesleft = bytepos % UsableBytesInSegment;
1919 :
1920 30010998 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1921 : {
1922 : /* fits on first page of segment */
1923 163734 : if (bytesleft == 0)
1924 108924 : seg_offset = 0;
1925 : else
1926 54810 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1927 : }
1928 : else
1929 : {
1930 : /* account for the first page on segment with long header */
1931 29847264 : seg_offset = XLOG_BLCKSZ;
1932 29847264 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1933 :
1934 29847264 : fullpages = bytesleft / UsableBytesInPage;
1935 29847264 : bytesleft = bytesleft % UsableBytesInPage;
1936 :
1937 29847264 : if (bytesleft == 0)
1938 28660 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
1939 : else
1940 29818604 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1941 : }
1942 :
1943 30010998 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1944 :
1945 30010998 : return result;
1946 : }
1947 :
1948 : /*
1949 : * Convert an XLogRecPtr to a "usable byte position".
1950 : */
1951 : static uint64
1952 4810 : XLogRecPtrToBytePos(XLogRecPtr ptr)
1953 : {
1954 : uint64 fullsegs;
1955 : uint32 fullpages;
1956 : uint32 offset;
1957 : uint64 result;
1958 :
1959 4810 : XLByteToSeg(ptr, fullsegs, wal_segment_size);
1960 :
1961 4810 : fullpages = (XLogSegmentOffset(ptr, wal_segment_size)) / XLOG_BLCKSZ;
1962 4810 : offset = ptr % XLOG_BLCKSZ;
1963 :
1964 4810 : if (fullpages == 0)
1965 : {
1966 1864 : result = fullsegs * UsableBytesInSegment;
1967 1864 : if (offset > 0)
1968 : {
1969 : Assert(offset >= SizeOfXLogLongPHD);
1970 472 : result += offset - SizeOfXLogLongPHD;
1971 : }
1972 : }
1973 : else
1974 : {
1975 2946 : result = fullsegs * UsableBytesInSegment +
1976 2946 : (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
1977 2946 : (fullpages - 1) * UsableBytesInPage; /* full pages */
1978 2946 : if (offset > 0)
1979 : {
1980 : Assert(offset >= SizeOfXLogShortPHD);
1981 2926 : result += offset - SizeOfXLogShortPHD;
1982 : }
1983 : }
1984 :
1985 4810 : return result;
1986 : }
1987 :
1988 : /*
1989 : * Initialize XLOG buffers, writing out old buffers if they still contain
1990 : * unwritten data, upto the page containing 'upto'. Or if 'opportunistic' is
1991 : * true, initialize as many pages as we can without having to write out
1992 : * unwritten data. Any new pages are initialized to zeros, with pages headers
1993 : * initialized properly.
1994 : */
1995 : static void
1996 5067994 : AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
1997 : {
1998 5067994 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1999 : int nextidx;
2000 : XLogRecPtr OldPageRqstPtr;
2001 : XLogwrtRqst WriteRqst;
2002 5067994 : XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
2003 : XLogRecPtr NewPageBeginPtr;
2004 : XLogPageHeader NewPage;
2005 : XLogRecPtr ReservedPtr;
2006 5067994 : int npages pg_attribute_unused() = 0;
2007 :
2008 : /*
2009 : * We must run the loop below inside the critical section as we expect
2010 : * XLogCtl->InitializedUpTo to eventually keep up. The most of callers
2011 : * already run inside the critical section. Except for WAL writer, which
2012 : * passed 'opportunistic == true', and therefore we don't perform
2013 : * operations that could error out.
2014 : *
2015 : * Start an explicit critical section anyway though.
2016 : */
2017 : Assert(CritSectionCount > 0 || opportunistic);
2018 5067994 : START_CRIT_SECTION();
2019 :
2020 : /*--
2021 : * Loop till we get all the pages in WAL buffer before 'upto' reserved for
2022 : * initialization. Multiple process can initialize different buffers with
2023 : * this loop in parallel as following.
2024 : *
2025 : * 1. Reserve page for initialization using XLogCtl->InitializeReserved.
2026 : * 2. Initialize the reserved page.
2027 : * 3. Attempt to advance XLogCtl->InitializedUpTo,
2028 : */
2029 5067994 : ReservedPtr = pg_atomic_read_u64(&XLogCtl->InitializeReserved);
2030 14807336 : while (upto >= ReservedPtr || opportunistic)
2031 : {
2032 : Assert(ReservedPtr % XLOG_BLCKSZ == 0);
2033 :
2034 : /*
2035 : * Get ending-offset of the buffer page we need to replace.
2036 : *
2037 : * We don't lookup into xlblocks, but rather calculate position we
2038 : * must wait to be written. If it was written, xlblocks will have this
2039 : * position (or uninitialized)
2040 : */
2041 9749664 : if (ReservedPtr + XLOG_BLCKSZ > XLogCtl->InitializedFrom + XLOG_BLCKSZ * XLOGbuffers)
2042 9180104 : OldPageRqstPtr = ReservedPtr + XLOG_BLCKSZ - (XLogRecPtr) XLOG_BLCKSZ * XLOGbuffers;
2043 : else
2044 569560 : OldPageRqstPtr = InvalidXLogRecPtr;
2045 :
2046 9749664 : if (LogwrtResult.Write < OldPageRqstPtr && opportunistic)
2047 : {
2048 : /*
2049 : * If we just want to pre-initialize as much as we can without
2050 : * flushing, give up now.
2051 : */
2052 10322 : upto = ReservedPtr - 1;
2053 10322 : break;
2054 : }
2055 :
2056 : /*
2057 : * Attempt to reserve the page for initialization. Failure means that
2058 : * this page got reserved by another process.
2059 : */
2060 9739342 : if (!pg_atomic_compare_exchange_u64(&XLogCtl->InitializeReserved,
2061 : &ReservedPtr,
2062 : ReservedPtr + XLOG_BLCKSZ))
2063 4874510 : continue;
2064 :
2065 : /*
2066 : * Wait till page gets correctly initialized up to OldPageRqstPtr.
2067 : */
2068 4864832 : nextidx = XLogRecPtrToBufIdx(ReservedPtr);
2069 4865390 : while (pg_atomic_read_u64(&XLogCtl->InitializedUpTo) < OldPageRqstPtr)
2070 558 : ConditionVariableSleep(&XLogCtl->InitializedUpToCondVar, WAIT_EVENT_WAL_BUFFER_INIT);
2071 4864832 : ConditionVariableCancelSleep();
2072 : Assert(pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]) == OldPageRqstPtr);
2073 :
2074 : /* Fall through if it's already written out. */
2075 4864832 : if (LogwrtResult.Write < OldPageRqstPtr)
2076 : {
2077 : /* Nope, got work to do. */
2078 :
2079 : /* Advance shared memory write request position */
2080 3838548 : SpinLockAcquire(&XLogCtl->info_lck);
2081 3838548 : if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
2082 1114868 : XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
2083 3838548 : SpinLockRelease(&XLogCtl->info_lck);
2084 :
2085 : /*
2086 : * Acquire an up-to-date LogwrtResult value and see if we still
2087 : * need to write it or if someone else already did.
2088 : */
2089 3838548 : RefreshXLogWriteResult(LogwrtResult);
2090 3838548 : if (LogwrtResult.Write < OldPageRqstPtr)
2091 : {
2092 3814794 : WaitXLogInsertionsToFinish(OldPageRqstPtr);
2093 :
2094 3814794 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2095 :
2096 3814794 : RefreshXLogWriteResult(LogwrtResult);
2097 3814794 : if (LogwrtResult.Write >= OldPageRqstPtr)
2098 : {
2099 : /* OK, someone wrote it already */
2100 60562 : LWLockRelease(WALWriteLock);
2101 : }
2102 : else
2103 : {
2104 : /* Have to write it ourselves */
2105 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2106 3754232 : WriteRqst.Write = OldPageRqstPtr;
2107 3754232 : WriteRqst.Flush = 0;
2108 3754232 : XLogWrite(WriteRqst, tli, false);
2109 3754232 : LWLockRelease(WALWriteLock);
2110 3754232 : pgWalUsage.wal_buffers_full++;
2111 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_DONE();
2112 : }
2113 : }
2114 : }
2115 :
2116 : /*
2117 : * Now the next buffer slot is free and we can set it up to be the
2118 : * next output page.
2119 : */
2120 4864832 : NewPageBeginPtr = ReservedPtr;
2121 4864832 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2122 :
2123 4864832 : NewPage = (XLogPageHeader) (XLogCtl->pages + nextidx * (Size) XLOG_BLCKSZ);
2124 :
2125 : /*
2126 : * Mark the xlblock with InvalidXLogRecPtr and issue a write barrier
2127 : * before initializing. Otherwise, the old page may be partially
2128 : * zeroed but look valid.
2129 : */
2130 4864832 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], InvalidXLogRecPtr);
2131 4864832 : pg_write_barrier();
2132 :
2133 : /*
2134 : * Be sure to re-zero the buffer so that bytes beyond what we've
2135 : * written will look like zeroes and not valid XLOG records...
2136 : */
2137 4864832 : MemSet(NewPage, 0, XLOG_BLCKSZ);
2138 :
2139 : /*
2140 : * Fill the new page's header
2141 : */
2142 4864832 : NewPage->xlp_magic = XLOG_PAGE_MAGIC;
2143 :
2144 : /* NewPage->xlp_info = 0; */ /* done by memset */
2145 4864832 : NewPage->xlp_tli = tli;
2146 4864832 : NewPage->xlp_pageaddr = NewPageBeginPtr;
2147 :
2148 : /* NewPage->xlp_rem_len = 0; */ /* done by memset */
2149 :
2150 : /*
2151 : * If online backup is not in progress, mark the header to indicate
2152 : * that WAL records beginning in this page have removable backup
2153 : * blocks. This allows the WAL archiver to know whether it is safe to
2154 : * compress archived WAL data by transforming full-block records into
2155 : * the non-full-block format. It is sufficient to record this at the
2156 : * page level because we force a page switch (in fact a segment
2157 : * switch) when starting a backup, so the flag will be off before any
2158 : * records can be written during the backup. At the end of a backup,
2159 : * the last page will be marked as all unsafe when perhaps only part
2160 : * is unsafe, but at worst the archiver would miss the opportunity to
2161 : * compress a few records.
2162 : */
2163 4864832 : if (Insert->runningBackups == 0)
2164 4598230 : NewPage->xlp_info |= XLP_BKP_REMOVABLE;
2165 :
2166 : /*
2167 : * If first page of an XLOG segment file, make it a long header.
2168 : */
2169 4864832 : if ((XLogSegmentOffset(NewPage->xlp_pageaddr, wal_segment_size)) == 0)
2170 : {
2171 3646 : XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2172 :
2173 3646 : NewLongPage->xlp_sysid = ControlFile->system_identifier;
2174 3646 : NewLongPage->xlp_seg_size = wal_segment_size;
2175 3646 : NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2176 3646 : NewPage->xlp_info |= XLP_LONG_HEADER;
2177 : }
2178 :
2179 : /*
2180 : * Make sure the initialization of the page becomes visible to others
2181 : * before the xlblocks update. GetXLogBuffer() reads xlblocks without
2182 : * holding a lock.
2183 : */
2184 4864832 : pg_write_barrier();
2185 :
2186 : /*-----
2187 : * Update the value of XLogCtl->xlblocks[nextidx] and try to advance
2188 : * XLogCtl->InitializedUpTo in a lock-less manner.
2189 : *
2190 : * First, let's provide a formal proof of the algorithm. Let it be 'n'
2191 : * process with the following variables in shared memory:
2192 : * f - an array of 'n' boolean flags,
2193 : * v - atomic integer variable.
2194 : *
2195 : * Also, let
2196 : * i - a number of a process,
2197 : * j - local integer variable,
2198 : * CAS(var, oldval, newval) - compare-and-swap atomic operation
2199 : * returning true on success,
2200 : * write_barrier()/read_barrier() - memory barriers.
2201 : *
2202 : * The pseudocode for each process is the following.
2203 : *
2204 : * j := i
2205 : * f[i] := true
2206 : * write_barrier()
2207 : * while CAS(v, j, j + 1):
2208 : * j := j + 1
2209 : * read_barrier()
2210 : * if not f[j]:
2211 : * break
2212 : *
2213 : * Let's prove that v eventually reaches the value of n.
2214 : * 1. Prove by contradiction. Assume v doesn't reach n and stucks
2215 : * on k, where k < n.
2216 : * 2. Process k attempts CAS(v, k, k + 1). 1). If, as we assumed, v
2217 : * gets stuck at k, then this CAS operation must fail. Therefore,
2218 : * v < k when process k attempts CAS(v, k, k + 1).
2219 : * 3. If, as we assumed, v gets stuck at k, then the value k of v
2220 : * must be achieved by some process m, where m < k. The process
2221 : * m must observe f[k] == false. Otherwise, it will later attempt
2222 : * CAS(v, k, k + 1) with success.
2223 : * 4. Therefore, corresponding read_barrier() (while j == k) on
2224 : * process m reached before write_barrier() of process k. But then
2225 : * process k attempts CAS(v, k, k + 1) after process m successfully
2226 : * incremented v to k, and that CAS operation must succeed.
2227 : * That leads to a contradiction. So, there is no such k (k < n)
2228 : * where v gets stuck. Q.E.D.
2229 : *
2230 : * To apply this proof to the code below, we assume
2231 : * XLogCtl->InitializedUpTo will play the role of v with XLOG_BLCKSZ
2232 : * granularity. We also assume setting XLogCtl->xlblocks[nextidx] to
2233 : * NewPageEndPtr to play the role of setting f[i] to true. Also, note
2234 : * that processes can't concurrently map different xlog locations to
2235 : * the same nextidx because we previously requested that
2236 : * XLogCtl->InitializedUpTo >= OldPageRqstPtr. So, a xlog buffer can
2237 : * be taken for initialization only once the previous initialization
2238 : * takes effect on XLogCtl->InitializedUpTo.
2239 : */
2240 :
2241 4864832 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], NewPageEndPtr);
2242 :
2243 4864832 : pg_write_barrier();
2244 :
2245 4938662 : while (pg_atomic_compare_exchange_u64(&XLogCtl->InitializedUpTo, &NewPageBeginPtr, NewPageEndPtr))
2246 : {
2247 4871416 : NewPageBeginPtr = NewPageEndPtr;
2248 4871416 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2249 4871416 : nextidx = XLogRecPtrToBufIdx(NewPageBeginPtr);
2250 :
2251 4871416 : pg_read_barrier();
2252 :
2253 4871416 : if (pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]) != NewPageEndPtr)
2254 : {
2255 : /*
2256 : * Page at nextidx wasn't initialized yet, so we can't move
2257 : * InitializedUpto further. It will be moved by backend which
2258 : * will initialize nextidx.
2259 : */
2260 4797586 : ConditionVariableBroadcast(&XLogCtl->InitializedUpToCondVar);
2261 4797586 : break;
2262 : }
2263 : }
2264 :
2265 4864832 : npages++;
2266 : }
2267 :
2268 5067994 : END_CRIT_SECTION();
2269 :
2270 : /*
2271 : * All the pages in WAL buffer before 'upto' were reserved for
2272 : * initialization. However, some pages might be reserved by concurrent
2273 : * processes. Wait till they finish initialization.
2274 : */
2275 6422888 : while (upto >= pg_atomic_read_u64(&XLogCtl->InitializedUpTo))
2276 1354894 : ConditionVariableSleep(&XLogCtl->InitializedUpToCondVar, WAIT_EVENT_WAL_BUFFER_INIT);
2277 5067994 : ConditionVariableCancelSleep();
2278 :
2279 5067994 : pg_read_barrier();
2280 :
2281 : #ifdef WAL_DEBUG
2282 : if (XLOG_DEBUG && npages > 0)
2283 : {
2284 : elog(DEBUG1, "initialized %d pages, up to %X/%08X",
2285 : npages, LSN_FORMAT_ARGS(NewPageEndPtr));
2286 : }
2287 : #endif
2288 5067994 : }
2289 :
2290 : /*
2291 : * Calculate CheckPointSegments based on max_wal_size_mb and
2292 : * checkpoint_completion_target.
2293 : */
2294 : static void
2295 14766 : CalculateCheckpointSegments(void)
2296 : {
2297 : double target;
2298 :
2299 : /*-------
2300 : * Calculate the distance at which to trigger a checkpoint, to avoid
2301 : * exceeding max_wal_size_mb. This is based on two assumptions:
2302 : *
2303 : * a) we keep WAL for only one checkpoint cycle (prior to PG11 we kept
2304 : * WAL for two checkpoint cycles to allow us to recover from the
2305 : * secondary checkpoint if the first checkpoint failed, though we
2306 : * only did this on the primary anyway, not on standby. Keeping just
2307 : * one checkpoint simplifies processing and reduces disk space in
2308 : * many smaller databases.)
2309 : * b) during checkpoint, we consume checkpoint_completion_target *
2310 : * number of segments consumed between checkpoints.
2311 : *-------
2312 : */
2313 14766 : target = (double) ConvertToXSegs(max_wal_size_mb, wal_segment_size) /
2314 14766 : (1.0 + CheckPointCompletionTarget);
2315 :
2316 : /* round down */
2317 14766 : CheckPointSegments = (int) target;
2318 :
2319 14766 : if (CheckPointSegments < 1)
2320 20 : CheckPointSegments = 1;
2321 14766 : }
2322 :
2323 : void
2324 10612 : assign_max_wal_size(int newval, void *extra)
2325 : {
2326 10612 : max_wal_size_mb = newval;
2327 10612 : CalculateCheckpointSegments();
2328 10612 : }
2329 :
2330 : void
2331 2208 : assign_checkpoint_completion_target(double newval, void *extra)
2332 : {
2333 2208 : CheckPointCompletionTarget = newval;
2334 2208 : CalculateCheckpointSegments();
2335 2208 : }
2336 :
2337 : bool
2338 4258 : check_wal_segment_size(int *newval, void **extra, GucSource source)
2339 : {
2340 4258 : if (!IsValidWalSegSize(*newval))
2341 : {
2342 0 : GUC_check_errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.");
2343 0 : return false;
2344 : }
2345 :
2346 4258 : return true;
2347 : }
2348 :
2349 : /*
2350 : * GUC check_hook for max_slot_wal_keep_size
2351 : *
2352 : * We don't allow the value of max_slot_wal_keep_size other than -1 during the
2353 : * binary upgrade. See start_postmaster() in pg_upgrade for more details.
2354 : */
2355 : bool
2356 2464 : check_max_slot_wal_keep_size(int *newval, void **extra, GucSource source)
2357 : {
2358 2464 : if (IsBinaryUpgrade && *newval != -1)
2359 : {
2360 0 : GUC_check_errdetail("\"%s\" must be set to -1 during binary upgrade mode.",
2361 : "max_slot_wal_keep_size");
2362 0 : return false;
2363 : }
2364 :
2365 2464 : return true;
2366 : }
2367 :
2368 : /*
2369 : * At a checkpoint, how many WAL segments to recycle as preallocated future
2370 : * XLOG segments? Returns the highest segment that should be preallocated.
2371 : */
2372 : static XLogSegNo
2373 3360 : XLOGfileslop(XLogRecPtr lastredoptr)
2374 : {
2375 : XLogSegNo minSegNo;
2376 : XLogSegNo maxSegNo;
2377 : double distance;
2378 : XLogSegNo recycleSegNo;
2379 :
2380 : /*
2381 : * Calculate the segment numbers that min_wal_size_mb and max_wal_size_mb
2382 : * correspond to. Always recycle enough segments to meet the minimum, and
2383 : * remove enough segments to stay below the maximum.
2384 : */
2385 3360 : minSegNo = lastredoptr / wal_segment_size +
2386 3360 : ConvertToXSegs(min_wal_size_mb, wal_segment_size) - 1;
2387 3360 : maxSegNo = lastredoptr / wal_segment_size +
2388 3360 : ConvertToXSegs(max_wal_size_mb, wal_segment_size) - 1;
2389 :
2390 : /*
2391 : * Between those limits, recycle enough segments to get us through to the
2392 : * estimated end of next checkpoint.
2393 : *
2394 : * To estimate where the next checkpoint will finish, assume that the
2395 : * system runs steadily consuming CheckPointDistanceEstimate bytes between
2396 : * every checkpoint.
2397 : */
2398 3360 : distance = (1.0 + CheckPointCompletionTarget) * CheckPointDistanceEstimate;
2399 : /* add 10% for good measure. */
2400 3360 : distance *= 1.10;
2401 :
2402 3360 : recycleSegNo = (XLogSegNo) ceil(((double) lastredoptr + distance) /
2403 : wal_segment_size);
2404 :
2405 3360 : if (recycleSegNo < minSegNo)
2406 2372 : recycleSegNo = minSegNo;
2407 3360 : if (recycleSegNo > maxSegNo)
2408 754 : recycleSegNo = maxSegNo;
2409 :
2410 3360 : return recycleSegNo;
2411 : }
2412 :
2413 : /*
2414 : * Check whether we've consumed enough xlog space that a checkpoint is needed.
2415 : *
2416 : * new_segno indicates a log file that has just been filled up (or read
2417 : * during recovery). We measure the distance from RedoRecPtr to new_segno
2418 : * and see if that exceeds CheckPointSegments.
2419 : *
2420 : * Note: it is caller's responsibility that RedoRecPtr is up-to-date.
2421 : */
2422 : bool
2423 10398 : XLogCheckpointNeeded(XLogSegNo new_segno)
2424 : {
2425 : XLogSegNo old_segno;
2426 :
2427 10398 : XLByteToSeg(RedoRecPtr, old_segno, wal_segment_size);
2428 :
2429 10398 : if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2430 6748 : return true;
2431 3650 : return false;
2432 : }
2433 :
2434 : /*
2435 : * Write and/or fsync the log at least as far as WriteRqst indicates.
2436 : *
2437 : * If flexible == true, we don't have to write as far as WriteRqst, but
2438 : * may stop at any convenient boundary (such as a cache or logfile boundary).
2439 : * This option allows us to avoid uselessly issuing multiple writes when a
2440 : * single one would do.
2441 : *
2442 : * Must be called with WALWriteLock held. WaitXLogInsertionsToFinish(WriteRqst)
2443 : * must be called before grabbing the lock, to make sure the data is ready to
2444 : * write.
2445 : */
2446 : static void
2447 4020536 : XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible)
2448 : {
2449 : bool ispartialpage;
2450 : bool last_iteration;
2451 : bool finishing_seg;
2452 : int curridx;
2453 : int npages;
2454 : int startidx;
2455 : uint32 startoffset;
2456 :
2457 : /* We should always be inside a critical section here */
2458 : Assert(CritSectionCount > 0);
2459 :
2460 : /*
2461 : * Update local LogwrtResult (caller probably did this already, but...)
2462 : */
2463 4020536 : RefreshXLogWriteResult(LogwrtResult);
2464 :
2465 : /*
2466 : * Since successive pages in the xlog cache are consecutively allocated,
2467 : * we can usually gather multiple pages together and issue just one
2468 : * write() call. npages is the number of pages we have determined can be
2469 : * written together; startidx is the cache block index of the first one,
2470 : * and startoffset is the file offset at which it should go. The latter
2471 : * two variables are only valid when npages > 0, but we must initialize
2472 : * all of them to keep the compiler quiet.
2473 : */
2474 4020536 : npages = 0;
2475 4020536 : startidx = 0;
2476 4020536 : startoffset = 0;
2477 :
2478 : /*
2479 : * Within the loop, curridx is the cache block index of the page to
2480 : * consider writing. Begin at the buffer containing the next unwritten
2481 : * page, or last partially written page.
2482 : */
2483 4020536 : curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2484 :
2485 8782094 : while (LogwrtResult.Write < WriteRqst.Write)
2486 : {
2487 : /*
2488 : * Make sure we're not ahead of the insert process. This could happen
2489 : * if we're passed a bogus WriteRqst.Write that is past the end of the
2490 : * last page that's been initialized by AdvanceXLInsertBuffer.
2491 : */
2492 5017020 : XLogRecPtr EndPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[curridx]);
2493 :
2494 5017020 : if (LogwrtResult.Write >= EndPtr)
2495 0 : elog(PANIC, "xlog write request %X/%08X is past end of log %X/%08X",
2496 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2497 : LSN_FORMAT_ARGS(EndPtr));
2498 :
2499 : /* Advance LogwrtResult.Write to end of current buffer page */
2500 5017020 : LogwrtResult.Write = EndPtr;
2501 5017020 : ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2502 :
2503 5017020 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2504 : wal_segment_size))
2505 : {
2506 : /*
2507 : * Switch to new logfile segment. We cannot have any pending
2508 : * pages here (since we dump what we have at segment end).
2509 : */
2510 : Assert(npages == 0);
2511 26294 : if (openLogFile >= 0)
2512 12240 : XLogFileClose();
2513 26294 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2514 : wal_segment_size);
2515 26294 : openLogTLI = tli;
2516 :
2517 : /* create/use new log file */
2518 26294 : openLogFile = XLogFileInit(openLogSegNo, tli);
2519 26294 : ReserveExternalFD();
2520 : }
2521 :
2522 : /* Make sure we have the current logfile open */
2523 5017020 : if (openLogFile < 0)
2524 : {
2525 0 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2526 : wal_segment_size);
2527 0 : openLogTLI = tli;
2528 0 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2529 0 : ReserveExternalFD();
2530 : }
2531 :
2532 : /* Add current page to the set of pending pages-to-dump */
2533 5017020 : if (npages == 0)
2534 : {
2535 : /* first of group */
2536 4056496 : startidx = curridx;
2537 4056496 : startoffset = XLogSegmentOffset(LogwrtResult.Write - XLOG_BLCKSZ,
2538 : wal_segment_size);
2539 : }
2540 5017020 : npages++;
2541 :
2542 : /*
2543 : * Dump the set if this will be the last loop iteration, or if we are
2544 : * at the last page of the cache area (since the next page won't be
2545 : * contiguous in memory), or if we are at the end of the logfile
2546 : * segment.
2547 : */
2548 5017020 : last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2549 :
2550 9786398 : finishing_seg = !ispartialpage &&
2551 4769378 : (startoffset + npages * XLOG_BLCKSZ) >= wal_segment_size;
2552 :
2553 5017020 : if (last_iteration ||
2554 998618 : curridx == XLogCtl->XLogCacheBlck ||
2555 : finishing_seg)
2556 : {
2557 : char *from;
2558 : Size nbytes;
2559 : Size nleft;
2560 : ssize_t written;
2561 : instr_time start;
2562 :
2563 : /* OK to write the page(s) */
2564 4056496 : from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2565 4056496 : nbytes = npages * (Size) XLOG_BLCKSZ;
2566 4056496 : nleft = nbytes;
2567 : do
2568 : {
2569 4056496 : errno = 0;
2570 :
2571 : /*
2572 : * Measure I/O timing to write WAL data, for pg_stat_io.
2573 : */
2574 4056496 : start = pgstat_prepare_io_time(track_wal_io_timing);
2575 :
2576 4056496 : pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE);
2577 4056496 : written = pg_pwrite(openLogFile, from, nleft, startoffset);
2578 4056496 : pgstat_report_wait_end();
2579 :
2580 4056496 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL,
2581 : IOOP_WRITE, start, 1, written);
2582 :
2583 4056496 : if (written <= 0)
2584 : {
2585 : char xlogfname[MAXFNAMELEN];
2586 : int save_errno;
2587 :
2588 0 : if (errno == EINTR)
2589 0 : continue;
2590 :
2591 0 : save_errno = errno;
2592 0 : XLogFileName(xlogfname, tli, openLogSegNo,
2593 : wal_segment_size);
2594 0 : errno = save_errno;
2595 0 : ereport(PANIC,
2596 : (errcode_for_file_access(),
2597 : errmsg("could not write to log file \"%s\" at offset %u, length %zu: %m",
2598 : xlogfname, startoffset, nleft)));
2599 : }
2600 4056496 : nleft -= written;
2601 4056496 : from += written;
2602 4056496 : startoffset += written;
2603 4056496 : } while (nleft > 0);
2604 :
2605 4056496 : npages = 0;
2606 :
2607 : /*
2608 : * If we just wrote the whole last page of a logfile segment,
2609 : * fsync the segment immediately. This avoids having to go back
2610 : * and re-open prior segments when an fsync request comes along
2611 : * later. Doing it here ensures that one and only one backend will
2612 : * perform this fsync.
2613 : *
2614 : * This is also the right place to notify the Archiver that the
2615 : * segment is ready to copy to archival storage, and to update the
2616 : * timer for archive_timeout, and to signal for a checkpoint if
2617 : * too many logfile segments have been used since the last
2618 : * checkpoint.
2619 : */
2620 4056496 : if (finishing_seg)
2621 : {
2622 3896 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2623 :
2624 : /* signal that we need to wakeup walsenders later */
2625 3896 : WalSndWakeupRequest();
2626 :
2627 3896 : LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2628 :
2629 3896 : if (XLogArchivingActive())
2630 810 : XLogArchiveNotifySeg(openLogSegNo, tli);
2631 :
2632 3896 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2633 3896 : XLogCtl->lastSegSwitchLSN = LogwrtResult.Flush;
2634 :
2635 : /*
2636 : * Request a checkpoint if we've consumed too much xlog since
2637 : * the last one. For speed, we first check using the local
2638 : * copy of RedoRecPtr, which might be out of date; if it looks
2639 : * like a checkpoint is needed, forcibly update RedoRecPtr and
2640 : * recheck.
2641 : */
2642 3896 : if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2643 : {
2644 474 : (void) GetRedoRecPtr();
2645 474 : if (XLogCheckpointNeeded(openLogSegNo))
2646 384 : RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2647 : }
2648 : }
2649 : }
2650 :
2651 5017020 : if (ispartialpage)
2652 : {
2653 : /* Only asked to write a partial page */
2654 247642 : LogwrtResult.Write = WriteRqst.Write;
2655 247642 : break;
2656 : }
2657 4769378 : curridx = NextBufIdx(curridx);
2658 :
2659 : /* If flexible, break out of loop as soon as we wrote something */
2660 4769378 : if (flexible && npages == 0)
2661 7820 : break;
2662 : }
2663 :
2664 : Assert(npages == 0);
2665 :
2666 : /*
2667 : * If asked to flush, do so
2668 : */
2669 4020536 : if (LogwrtResult.Flush < WriteRqst.Flush &&
2670 264912 : LogwrtResult.Flush < LogwrtResult.Write)
2671 : {
2672 : /*
2673 : * Could get here without iterating above loop, in which case we might
2674 : * have no open file or the wrong one. However, we do not need to
2675 : * fsync more than one file.
2676 : */
2677 264786 : if (wal_sync_method != WAL_SYNC_METHOD_OPEN &&
2678 264786 : wal_sync_method != WAL_SYNC_METHOD_OPEN_DSYNC)
2679 : {
2680 264786 : if (openLogFile >= 0 &&
2681 264762 : !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2682 : wal_segment_size))
2683 164 : XLogFileClose();
2684 264786 : if (openLogFile < 0)
2685 : {
2686 188 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2687 : wal_segment_size);
2688 188 : openLogTLI = tli;
2689 188 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2690 188 : ReserveExternalFD();
2691 : }
2692 :
2693 264786 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2694 : }
2695 :
2696 : /* signal that we need to wakeup walsenders later */
2697 264786 : WalSndWakeupRequest();
2698 :
2699 264786 : LogwrtResult.Flush = LogwrtResult.Write;
2700 : }
2701 :
2702 : /*
2703 : * Update shared-memory status
2704 : *
2705 : * We make sure that the shared 'request' values do not fall behind the
2706 : * 'result' values. This is not absolutely essential, but it saves some
2707 : * code in a couple of places.
2708 : */
2709 4020536 : SpinLockAcquire(&XLogCtl->info_lck);
2710 4020536 : if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
2711 230656 : XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
2712 4020536 : if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
2713 267734 : XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
2714 4020536 : SpinLockRelease(&XLogCtl->info_lck);
2715 :
2716 : /*
2717 : * We write Write first, bar, then Flush. When reading, the opposite must
2718 : * be done (with a matching barrier in between), so that we always see a
2719 : * Flush value that trails behind the Write value seen.
2720 : */
2721 4020536 : pg_atomic_write_u64(&XLogCtl->logWriteResult, LogwrtResult.Write);
2722 4020536 : pg_write_barrier();
2723 4020536 : pg_atomic_write_u64(&XLogCtl->logFlushResult, LogwrtResult.Flush);
2724 :
2725 : #ifdef USE_ASSERT_CHECKING
2726 : {
2727 : XLogRecPtr Flush;
2728 : XLogRecPtr Write;
2729 : XLogRecPtr Insert;
2730 :
2731 : Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult);
2732 : pg_read_barrier();
2733 : Write = pg_atomic_read_u64(&XLogCtl->logWriteResult);
2734 : pg_read_barrier();
2735 : Insert = pg_atomic_read_u64(&XLogCtl->logInsertResult);
2736 :
2737 : /* WAL written to disk is always ahead of WAL flushed */
2738 : Assert(Write >= Flush);
2739 :
2740 : /* WAL inserted to buffers is always ahead of WAL written */
2741 : Assert(Insert >= Write);
2742 : }
2743 : #endif
2744 4020536 : }
2745 :
2746 : /*
2747 : * Record the LSN for an asynchronous transaction commit/abort
2748 : * and nudge the WALWriter if there is work for it to do.
2749 : * (This should not be called for synchronous commits.)
2750 : */
2751 : void
2752 98848 : XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2753 : {
2754 98848 : XLogRecPtr WriteRqstPtr = asyncXactLSN;
2755 : bool sleeping;
2756 98848 : bool wakeup = false;
2757 : XLogRecPtr prevAsyncXactLSN;
2758 :
2759 98848 : SpinLockAcquire(&XLogCtl->info_lck);
2760 98848 : sleeping = XLogCtl->WalWriterSleeping;
2761 98848 : prevAsyncXactLSN = XLogCtl->asyncXactLSN;
2762 98848 : if (XLogCtl->asyncXactLSN < asyncXactLSN)
2763 97886 : XLogCtl->asyncXactLSN = asyncXactLSN;
2764 98848 : SpinLockRelease(&XLogCtl->info_lck);
2765 :
2766 : /*
2767 : * If somebody else already called this function with a more aggressive
2768 : * LSN, they will have done what we needed (and perhaps more).
2769 : */
2770 98848 : if (asyncXactLSN <= prevAsyncXactLSN)
2771 962 : return;
2772 :
2773 : /*
2774 : * If the WALWriter is sleeping, kick it to make it come out of low-power
2775 : * mode, so that this async commit will reach disk within the expected
2776 : * amount of time. Otherwise, determine whether it has enough WAL
2777 : * available to flush, the same way that XLogBackgroundFlush() does.
2778 : */
2779 97886 : if (sleeping)
2780 56 : wakeup = true;
2781 : else
2782 : {
2783 : int flushblocks;
2784 :
2785 97830 : RefreshXLogWriteResult(LogwrtResult);
2786 :
2787 97830 : flushblocks =
2788 97830 : WriteRqstPtr / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
2789 :
2790 97830 : if (WalWriterFlushAfter == 0 || flushblocks >= WalWriterFlushAfter)
2791 7572 : wakeup = true;
2792 : }
2793 :
2794 97886 : if (wakeup)
2795 : {
2796 7628 : volatile PROC_HDR *procglobal = ProcGlobal;
2797 7628 : ProcNumber walwriterProc = procglobal->walwriterProc;
2798 :
2799 7628 : if (walwriterProc != INVALID_PROC_NUMBER)
2800 404 : SetLatch(&GetPGProcByNumber(walwriterProc)->procLatch);
2801 : }
2802 : }
2803 :
2804 : /*
2805 : * Record the LSN up to which we can remove WAL because it's not required by
2806 : * any replication slot.
2807 : */
2808 : void
2809 50320 : XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn)
2810 : {
2811 50320 : SpinLockAcquire(&XLogCtl->info_lck);
2812 50320 : XLogCtl->replicationSlotMinLSN = lsn;
2813 50320 : SpinLockRelease(&XLogCtl->info_lck);
2814 50320 : }
2815 :
2816 :
2817 : /*
2818 : * Return the oldest LSN we must retain to satisfy the needs of some
2819 : * replication slot.
2820 : */
2821 : static XLogRecPtr
2822 4290 : XLogGetReplicationSlotMinimumLSN(void)
2823 : {
2824 : XLogRecPtr retval;
2825 :
2826 4290 : SpinLockAcquire(&XLogCtl->info_lck);
2827 4290 : retval = XLogCtl->replicationSlotMinLSN;
2828 4290 : SpinLockRelease(&XLogCtl->info_lck);
2829 :
2830 4290 : return retval;
2831 : }
2832 :
2833 : /*
2834 : * Advance minRecoveryPoint in control file.
2835 : *
2836 : * If we crash during recovery, we must reach this point again before the
2837 : * database is consistent.
2838 : *
2839 : * If 'force' is true, 'lsn' argument is ignored. Otherwise, minRecoveryPoint
2840 : * is only updated if it's not already greater than or equal to 'lsn'.
2841 : */
2842 : static void
2843 211928 : UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2844 : {
2845 : /* Quick check using our local copy of the variable */
2846 211928 : if (!updateMinRecoveryPoint || (!force && lsn <= LocalMinRecoveryPoint))
2847 198572 : return;
2848 :
2849 : /*
2850 : * An invalid minRecoveryPoint means that we need to recover all the WAL,
2851 : * i.e., we're doing crash recovery. We never modify the control file's
2852 : * value in that case, so we can short-circuit future checks here too. The
2853 : * local values of minRecoveryPoint and minRecoveryPointTLI should not be
2854 : * updated until crash recovery finishes. We only do this for the startup
2855 : * process as it should not update its own reference of minRecoveryPoint
2856 : * until it has finished crash recovery to make sure that all WAL
2857 : * available is replayed in this case. This also saves from extra locks
2858 : * taken on the control file from the startup process.
2859 : */
2860 13356 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint) && InRecovery)
2861 : {
2862 62 : updateMinRecoveryPoint = false;
2863 62 : return;
2864 : }
2865 :
2866 13294 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2867 :
2868 : /* update local copy */
2869 13294 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
2870 13294 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2871 :
2872 13294 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint))
2873 6 : updateMinRecoveryPoint = false;
2874 13288 : else if (force || LocalMinRecoveryPoint < lsn)
2875 : {
2876 : XLogRecPtr newMinRecoveryPoint;
2877 : TimeLineID newMinRecoveryPointTLI;
2878 :
2879 : /*
2880 : * To avoid having to update the control file too often, we update it
2881 : * all the way to the last record being replayed, even though 'lsn'
2882 : * would suffice for correctness. This also allows the 'force' case
2883 : * to not need a valid 'lsn' value.
2884 : *
2885 : * Another important reason for doing it this way is that the passed
2886 : * 'lsn' value could be bogus, i.e., past the end of available WAL, if
2887 : * the caller got it from a corrupted heap page. Accepting such a
2888 : * value as the min recovery point would prevent us from coming up at
2889 : * all. Instead, we just log a warning and continue with recovery.
2890 : * (See also the comments about corrupt LSNs in XLogFlush.)
2891 : */
2892 10706 : newMinRecoveryPoint = GetCurrentReplayRecPtr(&newMinRecoveryPointTLI);
2893 10706 : if (!force && newMinRecoveryPoint < lsn)
2894 0 : elog(WARNING,
2895 : "xlog min recovery request %X/%08X is past current point %X/%08X",
2896 : LSN_FORMAT_ARGS(lsn), LSN_FORMAT_ARGS(newMinRecoveryPoint));
2897 :
2898 : /* update control file */
2899 10706 : if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2900 : {
2901 10010 : ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2902 10010 : ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2903 10010 : UpdateControlFile();
2904 10010 : LocalMinRecoveryPoint = newMinRecoveryPoint;
2905 10010 : LocalMinRecoveryPointTLI = newMinRecoveryPointTLI;
2906 :
2907 10010 : ereport(DEBUG2,
2908 : errmsg_internal("updated min recovery point to %X/%08X on timeline %u",
2909 : LSN_FORMAT_ARGS(newMinRecoveryPoint),
2910 : newMinRecoveryPointTLI));
2911 : }
2912 : }
2913 13294 : LWLockRelease(ControlFileLock);
2914 : }
2915 :
2916 : /*
2917 : * Ensure that all XLOG data through the given position is flushed to disk.
2918 : *
2919 : * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
2920 : * already held, and we try to avoid acquiring it if possible.
2921 : */
2922 : void
2923 1371302 : XLogFlush(XLogRecPtr record)
2924 : {
2925 : XLogRecPtr WriteRqstPtr;
2926 : XLogwrtRqst WriteRqst;
2927 1371302 : TimeLineID insertTLI = XLogCtl->InsertTimeLineID;
2928 :
2929 : /*
2930 : * During REDO, we are reading not writing WAL. Therefore, instead of
2931 : * trying to flush the WAL, we should update minRecoveryPoint instead. We
2932 : * test XLogInsertAllowed(), not InRecovery, because we need checkpointer
2933 : * to act this way too, and because when it tries to write the
2934 : * end-of-recovery checkpoint, it should indeed flush.
2935 : */
2936 1371302 : if (!XLogInsertAllowed())
2937 : {
2938 211042 : UpdateMinRecoveryPoint(record, false);
2939 1089950 : return;
2940 : }
2941 :
2942 : /* Quick exit if already known flushed */
2943 1160260 : if (record <= LogwrtResult.Flush)
2944 878908 : return;
2945 :
2946 : #ifdef WAL_DEBUG
2947 : if (XLOG_DEBUG)
2948 : elog(LOG, "xlog flush request %X/%08X; write %X/%08X; flush %X/%08X",
2949 : LSN_FORMAT_ARGS(record),
2950 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2951 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2952 : #endif
2953 :
2954 281352 : START_CRIT_SECTION();
2955 :
2956 : /*
2957 : * Since fsync is usually a horribly expensive operation, we try to
2958 : * piggyback as much data as we can on each fsync: if we see any more data
2959 : * entered into the xlog buffer, we'll write and fsync that too, so that
2960 : * the final value of LogwrtResult.Flush is as large as possible. This
2961 : * gives us some chance of avoiding another fsync immediately after.
2962 : */
2963 :
2964 : /* initialize to given target; may increase below */
2965 281352 : WriteRqstPtr = record;
2966 :
2967 : /*
2968 : * Now wait until we get the write lock, or someone else does the flush
2969 : * for us.
2970 : */
2971 : for (;;)
2972 12248 : {
2973 : XLogRecPtr insertpos;
2974 :
2975 : /* done already? */
2976 293600 : RefreshXLogWriteResult(LogwrtResult);
2977 293600 : if (record <= LogwrtResult.Flush)
2978 21944 : break;
2979 :
2980 : /*
2981 : * Before actually performing the write, wait for all in-flight
2982 : * insertions to the pages we're about to write to finish.
2983 : */
2984 271656 : SpinLockAcquire(&XLogCtl->info_lck);
2985 271656 : if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
2986 19374 : WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
2987 271656 : SpinLockRelease(&XLogCtl->info_lck);
2988 271656 : insertpos = WaitXLogInsertionsToFinish(WriteRqstPtr);
2989 :
2990 : /*
2991 : * Try to get the write lock. If we can't get it immediately, wait
2992 : * until it's released, and recheck if we still need to do the flush
2993 : * or if the backend that held the lock did it for us already. This
2994 : * helps to maintain a good rate of group committing when the system
2995 : * is bottlenecked by the speed of fsyncing.
2996 : */
2997 271656 : if (!LWLockAcquireOrWait(WALWriteLock, LW_EXCLUSIVE))
2998 : {
2999 : /*
3000 : * The lock is now free, but we didn't acquire it yet. Before we
3001 : * do, loop back to check if someone else flushed the record for
3002 : * us already.
3003 : */
3004 12248 : continue;
3005 : }
3006 :
3007 : /* Got the lock; recheck whether request is satisfied */
3008 259408 : RefreshXLogWriteResult(LogwrtResult);
3009 259408 : if (record <= LogwrtResult.Flush)
3010 : {
3011 2886 : LWLockRelease(WALWriteLock);
3012 2886 : break;
3013 : }
3014 :
3015 : /*
3016 : * Sleep before flush! By adding a delay here, we may give further
3017 : * backends the opportunity to join the backlog of group commit
3018 : * followers; this can significantly improve transaction throughput,
3019 : * at the risk of increasing transaction latency.
3020 : *
3021 : * We do not sleep if enableFsync is not turned on, nor if there are
3022 : * fewer than CommitSiblings other backends with active transactions.
3023 : */
3024 256522 : if (CommitDelay > 0 && enableFsync &&
3025 0 : MinimumActiveBackends(CommitSiblings))
3026 : {
3027 0 : pg_usleep(CommitDelay);
3028 :
3029 : /*
3030 : * Re-check how far we can now flush the WAL. It's generally not
3031 : * safe to call WaitXLogInsertionsToFinish while holding
3032 : * WALWriteLock, because an in-progress insertion might need to
3033 : * also grab WALWriteLock to make progress. But we know that all
3034 : * the insertions up to insertpos have already finished, because
3035 : * that's what the earlier WaitXLogInsertionsToFinish() returned.
3036 : * We're only calling it again to allow insertpos to be moved
3037 : * further forward, not to actually wait for anyone.
3038 : */
3039 0 : insertpos = WaitXLogInsertionsToFinish(insertpos);
3040 : }
3041 :
3042 : /* try to write/flush later additions to XLOG as well */
3043 256522 : WriteRqst.Write = insertpos;
3044 256522 : WriteRqst.Flush = insertpos;
3045 :
3046 256522 : XLogWrite(WriteRqst, insertTLI, false);
3047 :
3048 256522 : LWLockRelease(WALWriteLock);
3049 : /* done */
3050 256522 : break;
3051 : }
3052 :
3053 281352 : END_CRIT_SECTION();
3054 :
3055 : /* wake up walsenders now that we've released heavily contended locks */
3056 281352 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3057 :
3058 : /*
3059 : * If we still haven't flushed to the request point then we have a
3060 : * problem; most likely, the requested flush point is past end of XLOG.
3061 : * This has been seen to occur when a disk page has a corrupted LSN.
3062 : *
3063 : * Formerly we treated this as a PANIC condition, but that hurts the
3064 : * system's robustness rather than helping it: we do not want to take down
3065 : * the whole system due to corruption on one data page. In particular, if
3066 : * the bad page is encountered again during recovery then we would be
3067 : * unable to restart the database at all! (This scenario actually
3068 : * happened in the field several times with 7.1 releases.) As of 8.4, bad
3069 : * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
3070 : * the only time we can reach here during recovery is while flushing the
3071 : * end-of-recovery checkpoint record, and we don't expect that to have a
3072 : * bad LSN.
3073 : *
3074 : * Note that for calls from xact.c, the ERROR will be promoted to PANIC
3075 : * since xact.c calls this routine inside a critical section. However,
3076 : * calls from bufmgr.c are not within critical sections and so we will not
3077 : * force a restart for a bad LSN on a data page.
3078 : */
3079 281352 : if (LogwrtResult.Flush < record)
3080 0 : elog(ERROR,
3081 : "xlog flush request %X/%08X is not satisfied --- flushed only to %X/%08X",
3082 : LSN_FORMAT_ARGS(record),
3083 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
3084 : }
3085 :
3086 : /*
3087 : * Write & flush xlog, but without specifying exactly where to.
3088 : *
3089 : * We normally write only completed blocks; but if there is nothing to do on
3090 : * that basis, we check for unwritten async commits in the current incomplete
3091 : * block, and write through the latest one of those. Thus, if async commits
3092 : * are not being used, we will write complete blocks only.
3093 : *
3094 : * If, based on the above, there's anything to write we do so immediately. But
3095 : * to avoid calling fsync, fdatasync et. al. at a rate that'd impact
3096 : * concurrent IO, we only flush WAL every wal_writer_delay ms, or if there's
3097 : * more than wal_writer_flush_after unflushed blocks.
3098 : *
3099 : * We can guarantee that async commits reach disk after at most three
3100 : * wal_writer_delay cycles. (When flushing complete blocks, we allow XLogWrite
3101 : * to write "flexibly", meaning it can stop at the end of the buffer ring;
3102 : * this makes a difference only with very high load or long wal_writer_delay,
3103 : * but imposes one extra cycle for the worst case for async commits.)
3104 : *
3105 : * This routine is invoked periodically by the background walwriter process.
3106 : *
3107 : * Returns true if there was any work to do, even if we skipped flushing due
3108 : * to wal_writer_delay/wal_writer_flush_after.
3109 : */
3110 : bool
3111 36442 : XLogBackgroundFlush(void)
3112 : {
3113 : XLogwrtRqst WriteRqst;
3114 36442 : bool flexible = true;
3115 : static TimestampTz lastflush;
3116 : TimestampTz now;
3117 : int flushblocks;
3118 : TimeLineID insertTLI;
3119 :
3120 : /* XLOG doesn't need flushing during recovery */
3121 36442 : if (RecoveryInProgress())
3122 200 : return false;
3123 :
3124 : /*
3125 : * Since we're not in recovery, InsertTimeLineID is set and can't change,
3126 : * so we can read it without a lock.
3127 : */
3128 36242 : insertTLI = XLogCtl->InsertTimeLineID;
3129 :
3130 : /* read updated LogwrtRqst */
3131 36242 : SpinLockAcquire(&XLogCtl->info_lck);
3132 36242 : WriteRqst = XLogCtl->LogwrtRqst;
3133 36242 : SpinLockRelease(&XLogCtl->info_lck);
3134 :
3135 : /* back off to last completed page boundary */
3136 36242 : WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
3137 :
3138 : /* if we have already flushed that far, consider async commit records */
3139 36242 : RefreshXLogWriteResult(LogwrtResult);
3140 36242 : if (WriteRqst.Write <= LogwrtResult.Flush)
3141 : {
3142 27644 : SpinLockAcquire(&XLogCtl->info_lck);
3143 27644 : WriteRqst.Write = XLogCtl->asyncXactLSN;
3144 27644 : SpinLockRelease(&XLogCtl->info_lck);
3145 27644 : flexible = false; /* ensure it all gets written */
3146 : }
3147 :
3148 : /*
3149 : * If already known flushed, we're done. Just need to check if we are
3150 : * holding an open file handle to a logfile that's no longer in use,
3151 : * preventing the file from being deleted.
3152 : */
3153 36242 : if (WriteRqst.Write <= LogwrtResult.Flush)
3154 : {
3155 25920 : if (openLogFile >= 0)
3156 : {
3157 16692 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
3158 : wal_segment_size))
3159 : {
3160 264 : XLogFileClose();
3161 : }
3162 : }
3163 25920 : return false;
3164 : }
3165 :
3166 : /*
3167 : * Determine how far to flush WAL, based on the wal_writer_delay and
3168 : * wal_writer_flush_after GUCs.
3169 : *
3170 : * Note that XLogSetAsyncXactLSN() performs similar calculation based on
3171 : * wal_writer_flush_after, to decide when to wake us up. Make sure the
3172 : * logic is the same in both places if you change this.
3173 : */
3174 10322 : now = GetCurrentTimestamp();
3175 10322 : flushblocks =
3176 10322 : WriteRqst.Write / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
3177 :
3178 10322 : if (WalWriterFlushAfter == 0 || lastflush == 0)
3179 : {
3180 : /* first call, or block based limits disabled */
3181 480 : WriteRqst.Flush = WriteRqst.Write;
3182 480 : lastflush = now;
3183 : }
3184 9842 : else if (TimestampDifferenceExceeds(lastflush, now, WalWriterDelay))
3185 : {
3186 : /*
3187 : * Flush the writes at least every WalWriterDelay ms. This is
3188 : * important to bound the amount of time it takes for an asynchronous
3189 : * commit to hit disk.
3190 : */
3191 9584 : WriteRqst.Flush = WriteRqst.Write;
3192 9584 : lastflush = now;
3193 : }
3194 258 : else if (flushblocks >= WalWriterFlushAfter)
3195 : {
3196 : /* exceeded wal_writer_flush_after blocks, flush */
3197 230 : WriteRqst.Flush = WriteRqst.Write;
3198 230 : lastflush = now;
3199 : }
3200 : else
3201 : {
3202 : /* no flushing, this time round */
3203 28 : WriteRqst.Flush = 0;
3204 : }
3205 :
3206 : #ifdef WAL_DEBUG
3207 : if (XLOG_DEBUG)
3208 : elog(LOG, "xlog bg flush request write %X/%08X; flush: %X/%08X, current is write %X/%08X; flush %X/%08X",
3209 : LSN_FORMAT_ARGS(WriteRqst.Write),
3210 : LSN_FORMAT_ARGS(WriteRqst.Flush),
3211 : LSN_FORMAT_ARGS(LogwrtResult.Write),
3212 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
3213 : #endif
3214 :
3215 10322 : START_CRIT_SECTION();
3216 :
3217 : /* now wait for any in-progress insertions to finish and get write lock */
3218 10322 : WaitXLogInsertionsToFinish(WriteRqst.Write);
3219 10322 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3220 10322 : RefreshXLogWriteResult(LogwrtResult);
3221 10322 : if (WriteRqst.Write > LogwrtResult.Write ||
3222 844 : WriteRqst.Flush > LogwrtResult.Flush)
3223 : {
3224 9782 : XLogWrite(WriteRqst, insertTLI, flexible);
3225 : }
3226 10322 : LWLockRelease(WALWriteLock);
3227 :
3228 10322 : END_CRIT_SECTION();
3229 :
3230 : /* wake up walsenders now that we've released heavily contended locks */
3231 10322 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3232 :
3233 : /*
3234 : * Great, done. To take some work off the critical path, try to initialize
3235 : * as many of the no-longer-needed WAL buffers for future use as we can.
3236 : */
3237 10322 : AdvanceXLInsertBuffer(InvalidXLogRecPtr, insertTLI, true);
3238 :
3239 : /*
3240 : * If we determined that we need to write data, but somebody else
3241 : * wrote/flushed already, it should be considered as being active, to
3242 : * avoid hibernating too early.
3243 : */
3244 10322 : return true;
3245 : }
3246 :
3247 : /*
3248 : * Test whether XLOG data has been flushed up to (at least) the given position.
3249 : *
3250 : * Returns true if a flush is still needed. (It may be that someone else
3251 : * is already in process of flushing that far, however.)
3252 : */
3253 : bool
3254 18203280 : XLogNeedsFlush(XLogRecPtr record)
3255 : {
3256 : /*
3257 : * During recovery, we don't flush WAL but update minRecoveryPoint
3258 : * instead. So "needs flush" is taken to mean whether minRecoveryPoint
3259 : * would need to be updated.
3260 : */
3261 18203280 : if (RecoveryInProgress())
3262 : {
3263 : /*
3264 : * An invalid minRecoveryPoint means that we need to recover all the
3265 : * WAL, i.e., we're doing crash recovery. We never modify the control
3266 : * file's value in that case, so we can short-circuit future checks
3267 : * here too. This triggers a quick exit path for the startup process,
3268 : * which cannot update its local copy of minRecoveryPoint as long as
3269 : * it has not replayed all WAL available when doing crash recovery.
3270 : */
3271 1176870 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint) && InRecovery)
3272 0 : updateMinRecoveryPoint = false;
3273 :
3274 : /* Quick exit if already known to be updated or cannot be updated */
3275 1176870 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3276 1151332 : return false;
3277 :
3278 : /*
3279 : * Update local copy of minRecoveryPoint. But if the lock is busy,
3280 : * just return a conservative guess.
3281 : */
3282 25538 : if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3283 0 : return true;
3284 25538 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
3285 25538 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3286 25538 : LWLockRelease(ControlFileLock);
3287 :
3288 : /*
3289 : * Check minRecoveryPoint for any other process than the startup
3290 : * process doing crash recovery, which should not update the control
3291 : * file value if crash recovery is still running.
3292 : */
3293 25538 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint))
3294 0 : updateMinRecoveryPoint = false;
3295 :
3296 : /* check again */
3297 25538 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3298 162 : return false;
3299 : else
3300 25376 : return true;
3301 : }
3302 :
3303 : /* Quick exit if already known flushed */
3304 17026410 : if (record <= LogwrtResult.Flush)
3305 16587316 : return false;
3306 :
3307 : /* read LogwrtResult and update local state */
3308 439094 : RefreshXLogWriteResult(LogwrtResult);
3309 :
3310 : /* check again */
3311 439094 : if (record <= LogwrtResult.Flush)
3312 6108 : return false;
3313 :
3314 432986 : return true;
3315 : }
3316 :
3317 : /*
3318 : * Try to make a given XLOG file segment exist.
3319 : *
3320 : * logsegno: identify segment.
3321 : *
3322 : * *added: on return, true if this call raised the number of extant segments.
3323 : *
3324 : * path: on return, this char[MAXPGPATH] has the path to the logsegno file.
3325 : *
3326 : * Returns -1 or FD of opened file. A -1 here is not an error; a caller
3327 : * wanting an open segment should attempt to open "path", which usually will
3328 : * succeed. (This is weird, but it's efficient for the callers.)
3329 : */
3330 : static int
3331 28600 : XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,
3332 : bool *added, char *path)
3333 : {
3334 : char tmppath[MAXPGPATH];
3335 : XLogSegNo installed_segno;
3336 : XLogSegNo max_segno;
3337 : int fd;
3338 : int save_errno;
3339 28600 : int open_flags = O_RDWR | O_CREAT | O_EXCL | PG_BINARY;
3340 : instr_time io_start;
3341 :
3342 : Assert(logtli != 0);
3343 :
3344 28600 : XLogFilePath(path, logtli, logsegno, wal_segment_size);
3345 :
3346 : /*
3347 : * Try to use existent file (checkpoint maker may have created it already)
3348 : */
3349 28600 : *added = false;
3350 28600 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3351 28600 : get_sync_bit(wal_sync_method));
3352 28600 : if (fd < 0)
3353 : {
3354 2954 : if (errno != ENOENT)
3355 0 : ereport(ERROR,
3356 : (errcode_for_file_access(),
3357 : errmsg("could not open file \"%s\": %m", path)));
3358 : }
3359 : else
3360 25646 : return fd;
3361 :
3362 : /*
3363 : * Initialize an empty (all zeroes) segment. NOTE: it is possible that
3364 : * another process is doing the same thing. If so, we will end up
3365 : * pre-creating an extra log segment. That seems OK, and better than
3366 : * holding the lock throughout this lengthy process.
3367 : */
3368 2954 : elog(DEBUG2, "creating and filling new WAL file");
3369 :
3370 2954 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3371 :
3372 2954 : unlink(tmppath);
3373 :
3374 2954 : if (io_direct_flags & IO_DIRECT_WAL_INIT)
3375 0 : open_flags |= PG_O_DIRECT;
3376 :
3377 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3378 2954 : fd = BasicOpenFile(tmppath, open_flags);
3379 2954 : if (fd < 0)
3380 0 : ereport(ERROR,
3381 : (errcode_for_file_access(),
3382 : errmsg("could not create file \"%s\": %m", tmppath)));
3383 :
3384 : /* Measure I/O timing when initializing segment */
3385 2954 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3386 :
3387 2954 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE);
3388 2954 : save_errno = 0;
3389 2954 : if (wal_init_zero)
3390 : {
3391 : ssize_t rc;
3392 :
3393 : /*
3394 : * Zero-fill the file. With this setting, we do this the hard way to
3395 : * ensure that all the file space has really been allocated. On
3396 : * platforms that allow "holes" in files, just seeking to the end
3397 : * doesn't allocate intermediate space. This way, we know that we
3398 : * have all the space and (after the fsync below) that all the
3399 : * indirect blocks are down on disk. Therefore, fdatasync(2) or
3400 : * O_DSYNC will be sufficient to sync future writes to the log file.
3401 : */
3402 2954 : rc = pg_pwrite_zeros(fd, wal_segment_size, 0);
3403 :
3404 2954 : if (rc < 0)
3405 0 : save_errno = errno;
3406 : }
3407 : else
3408 : {
3409 : /*
3410 : * Otherwise, seeking to the end and writing a solitary byte is
3411 : * enough.
3412 : */
3413 0 : errno = 0;
3414 0 : if (pg_pwrite(fd, "\0", 1, wal_segment_size - 1) != 1)
3415 : {
3416 : /* if write didn't set errno, assume no disk space */
3417 0 : save_errno = errno ? errno : ENOSPC;
3418 : }
3419 : }
3420 2954 : pgstat_report_wait_end();
3421 :
3422 : /*
3423 : * A full segment worth of data is written when using wal_init_zero. One
3424 : * byte is written when not using it.
3425 : */
3426 2954 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT, IOOP_WRITE,
3427 : io_start, 1,
3428 2954 : wal_init_zero ? wal_segment_size : 1);
3429 :
3430 2954 : if (save_errno)
3431 : {
3432 : /*
3433 : * If we fail to make the file, delete it to release disk space
3434 : */
3435 0 : unlink(tmppath);
3436 :
3437 0 : close(fd);
3438 :
3439 0 : errno = save_errno;
3440 :
3441 0 : ereport(ERROR,
3442 : (errcode_for_file_access(),
3443 : errmsg("could not write to file \"%s\": %m", tmppath)));
3444 : }
3445 :
3446 : /* Measure I/O timing when flushing segment */
3447 2954 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3448 :
3449 2954 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC);
3450 2954 : if (pg_fsync(fd) != 0)
3451 : {
3452 0 : save_errno = errno;
3453 0 : close(fd);
3454 0 : errno = save_errno;
3455 0 : ereport(ERROR,
3456 : (errcode_for_file_access(),
3457 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3458 : }
3459 2954 : pgstat_report_wait_end();
3460 :
3461 2954 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT,
3462 : IOOP_FSYNC, io_start, 1, 0);
3463 :
3464 2954 : if (close(fd) != 0)
3465 0 : ereport(ERROR,
3466 : (errcode_for_file_access(),
3467 : errmsg("could not close file \"%s\": %m", tmppath)));
3468 :
3469 : /*
3470 : * Now move the segment into place with its final name. Cope with
3471 : * possibility that someone else has created the file while we were
3472 : * filling ours: if so, use ours to pre-create a future log segment.
3473 : */
3474 2954 : installed_segno = logsegno;
3475 :
3476 : /*
3477 : * XXX: What should we use as max_segno? We used to use XLOGfileslop when
3478 : * that was a constant, but that was always a bit dubious: normally, at a
3479 : * checkpoint, XLOGfileslop was the offset from the checkpoint record, but
3480 : * here, it was the offset from the insert location. We can't do the
3481 : * normal XLOGfileslop calculation here because we don't have access to
3482 : * the prior checkpoint's redo location. So somewhat arbitrarily, just use
3483 : * CheckPointSegments.
3484 : */
3485 2954 : max_segno = logsegno + CheckPointSegments;
3486 2954 : if (InstallXLogFileSegment(&installed_segno, tmppath, true, max_segno,
3487 : logtli))
3488 : {
3489 2954 : *added = true;
3490 2954 : elog(DEBUG2, "done creating and filling new WAL file");
3491 : }
3492 : else
3493 : {
3494 : /*
3495 : * No need for any more future segments, or InstallXLogFileSegment()
3496 : * failed to rename the file into place. If the rename failed, a
3497 : * caller opening the file may fail.
3498 : */
3499 0 : unlink(tmppath);
3500 0 : elog(DEBUG2, "abandoned new WAL file");
3501 : }
3502 :
3503 2954 : return -1;
3504 : }
3505 :
3506 : /*
3507 : * Create a new XLOG file segment, or open a pre-existing one.
3508 : *
3509 : * logsegno: identify segment to be created/opened.
3510 : *
3511 : * Returns FD of opened file.
3512 : *
3513 : * Note: errors here are ERROR not PANIC because we might or might not be
3514 : * inside a critical section (eg, during checkpoint there is no reason to
3515 : * take down the system on failure). They will promote to PANIC if we are
3516 : * in a critical section.
3517 : */
3518 : int
3519 28062 : XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)
3520 : {
3521 : bool ignore_added;
3522 : char path[MAXPGPATH];
3523 : int fd;
3524 :
3525 : Assert(logtli != 0);
3526 :
3527 28062 : fd = XLogFileInitInternal(logsegno, logtli, &ignore_added, path);
3528 28062 : if (fd >= 0)
3529 25422 : return fd;
3530 :
3531 : /* Now open original target segment (might not be file I just made) */
3532 2640 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3533 2640 : get_sync_bit(wal_sync_method));
3534 2640 : if (fd < 0)
3535 0 : ereport(ERROR,
3536 : (errcode_for_file_access(),
3537 : errmsg("could not open file \"%s\": %m", path)));
3538 2640 : return fd;
3539 : }
3540 :
3541 : /*
3542 : * Create a new XLOG file segment by copying a pre-existing one.
3543 : *
3544 : * destsegno: identify segment to be created.
3545 : *
3546 : * srcTLI, srcsegno: identify segment to be copied (could be from
3547 : * a different timeline)
3548 : *
3549 : * upto: how much of the source file to copy (the rest is filled with
3550 : * zeros)
3551 : *
3552 : * Currently this is only used during recovery, and so there are no locking
3553 : * considerations. But we should be just as tense as XLogFileInit to avoid
3554 : * emplacing a bogus file.
3555 : */
3556 : static void
3557 78 : XLogFileCopy(TimeLineID destTLI, XLogSegNo destsegno,
3558 : TimeLineID srcTLI, XLogSegNo srcsegno,
3559 : int upto)
3560 : {
3561 : char path[MAXPGPATH];
3562 : char tmppath[MAXPGPATH];
3563 : PGAlignedXLogBlock buffer;
3564 : int srcfd;
3565 : int fd;
3566 : int nbytes;
3567 :
3568 : /*
3569 : * Open the source file
3570 : */
3571 78 : XLogFilePath(path, srcTLI, srcsegno, wal_segment_size);
3572 78 : srcfd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
3573 78 : if (srcfd < 0)
3574 0 : ereport(ERROR,
3575 : (errcode_for_file_access(),
3576 : errmsg("could not open file \"%s\": %m", path)));
3577 :
3578 : /*
3579 : * Copy into a temp file name.
3580 : */
3581 78 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3582 :
3583 78 : unlink(tmppath);
3584 :
3585 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3586 78 : fd = OpenTransientFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
3587 78 : if (fd < 0)
3588 0 : ereport(ERROR,
3589 : (errcode_for_file_access(),
3590 : errmsg("could not create file \"%s\": %m", tmppath)));
3591 :
3592 : /*
3593 : * Do the data copying.
3594 : */
3595 159822 : for (nbytes = 0; nbytes < wal_segment_size; nbytes += sizeof(buffer))
3596 : {
3597 : int nread;
3598 :
3599 159744 : nread = upto - nbytes;
3600 :
3601 : /*
3602 : * The part that is not read from the source file is filled with
3603 : * zeros.
3604 : */
3605 159744 : if (nread < sizeof(buffer))
3606 78 : memset(buffer.data, 0, sizeof(buffer));
3607 :
3608 159744 : if (nread > 0)
3609 : {
3610 : int r;
3611 :
3612 5388 : if (nread > sizeof(buffer))
3613 5310 : nread = sizeof(buffer);
3614 5388 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ);
3615 5388 : r = read(srcfd, buffer.data, nread);
3616 5388 : if (r != nread)
3617 : {
3618 0 : if (r < 0)
3619 0 : ereport(ERROR,
3620 : (errcode_for_file_access(),
3621 : errmsg("could not read file \"%s\": %m",
3622 : path)));
3623 : else
3624 0 : ereport(ERROR,
3625 : (errcode(ERRCODE_DATA_CORRUPTED),
3626 : errmsg("could not read file \"%s\": read %d of %zu",
3627 : path, r, (Size) nread)));
3628 : }
3629 5388 : pgstat_report_wait_end();
3630 : }
3631 159744 : errno = 0;
3632 159744 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_WRITE);
3633 159744 : if ((int) write(fd, buffer.data, sizeof(buffer)) != (int) sizeof(buffer))
3634 : {
3635 0 : int save_errno = errno;
3636 :
3637 : /*
3638 : * If we fail to make the file, delete it to release disk space
3639 : */
3640 0 : unlink(tmppath);
3641 : /* if write didn't set errno, assume problem is no disk space */
3642 0 : errno = save_errno ? save_errno : ENOSPC;
3643 :
3644 0 : ereport(ERROR,
3645 : (errcode_for_file_access(),
3646 : errmsg("could not write to file \"%s\": %m", tmppath)));
3647 : }
3648 159744 : pgstat_report_wait_end();
3649 : }
3650 :
3651 78 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_SYNC);
3652 78 : if (pg_fsync(fd) != 0)
3653 0 : ereport(data_sync_elevel(ERROR),
3654 : (errcode_for_file_access(),
3655 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3656 78 : pgstat_report_wait_end();
3657 :
3658 78 : if (CloseTransientFile(fd) != 0)
3659 0 : ereport(ERROR,
3660 : (errcode_for_file_access(),
3661 : errmsg("could not close file \"%s\": %m", tmppath)));
3662 :
3663 78 : if (CloseTransientFile(srcfd) != 0)
3664 0 : ereport(ERROR,
3665 : (errcode_for_file_access(),
3666 : errmsg("could not close file \"%s\": %m", path)));
3667 :
3668 : /*
3669 : * Now move the segment into place with its final name.
3670 : */
3671 78 : if (!InstallXLogFileSegment(&destsegno, tmppath, false, 0, destTLI))
3672 0 : elog(ERROR, "InstallXLogFileSegment should not have failed");
3673 78 : }
3674 :
3675 : /*
3676 : * Install a new XLOG segment file as a current or future log segment.
3677 : *
3678 : * This is used both to install a newly-created segment (which has a temp
3679 : * filename while it's being created) and to recycle an old segment.
3680 : *
3681 : * *segno: identify segment to install as (or first possible target).
3682 : * When find_free is true, this is modified on return to indicate the
3683 : * actual installation location or last segment searched.
3684 : *
3685 : * tmppath: initial name of file to install. It will be renamed into place.
3686 : *
3687 : * find_free: if true, install the new segment at the first empty segno
3688 : * number at or after the passed numbers. If false, install the new segment
3689 : * exactly where specified, deleting any existing segment file there.
3690 : *
3691 : * max_segno: maximum segment number to install the new file as. Fail if no
3692 : * free slot is found between *segno and max_segno. (Ignored when find_free
3693 : * is false.)
3694 : *
3695 : * tli: The timeline on which the new segment should be installed.
3696 : *
3697 : * Returns true if the file was installed successfully. false indicates that
3698 : * max_segno limit was exceeded, the startup process has disabled this
3699 : * function for now, or an error occurred while renaming the file into place.
3700 : */
3701 : static bool
3702 5784 : InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
3703 : bool find_free, XLogSegNo max_segno, TimeLineID tli)
3704 : {
3705 : char path[MAXPGPATH];
3706 : struct stat stat_buf;
3707 :
3708 : Assert(tli != 0);
3709 :
3710 5784 : XLogFilePath(path, tli, *segno, wal_segment_size);
3711 :
3712 5784 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3713 5784 : if (!XLogCtl->InstallXLogFileSegmentActive)
3714 : {
3715 0 : LWLockRelease(ControlFileLock);
3716 0 : return false;
3717 : }
3718 :
3719 5784 : if (!find_free)
3720 : {
3721 : /* Force installation: get rid of any pre-existing segment file */
3722 78 : durable_unlink(path, DEBUG1);
3723 : }
3724 : else
3725 : {
3726 : /* Find a free slot to put it in */
3727 8332 : while (stat(path, &stat_buf) == 0)
3728 : {
3729 2676 : if ((*segno) >= max_segno)
3730 : {
3731 : /* Failed to find a free slot within specified range */
3732 50 : LWLockRelease(ControlFileLock);
3733 50 : return false;
3734 : }
3735 2626 : (*segno)++;
3736 2626 : XLogFilePath(path, tli, *segno, wal_segment_size);
3737 : }
3738 : }
3739 :
3740 : Assert(access(path, F_OK) != 0 && errno == ENOENT);
3741 5734 : if (durable_rename(tmppath, path, LOG) != 0)
3742 : {
3743 0 : LWLockRelease(ControlFileLock);
3744 : /* durable_rename already emitted log message */
3745 0 : return false;
3746 : }
3747 :
3748 5734 : LWLockRelease(ControlFileLock);
3749 :
3750 5734 : return true;
3751 : }
3752 :
3753 : /*
3754 : * Open a pre-existing logfile segment for writing.
3755 : */
3756 : int
3757 188 : XLogFileOpen(XLogSegNo segno, TimeLineID tli)
3758 : {
3759 : char path[MAXPGPATH];
3760 : int fd;
3761 :
3762 188 : XLogFilePath(path, tli, segno, wal_segment_size);
3763 :
3764 188 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3765 188 : get_sync_bit(wal_sync_method));
3766 188 : if (fd < 0)
3767 0 : ereport(PANIC,
3768 : (errcode_for_file_access(),
3769 : errmsg("could not open file \"%s\": %m", path)));
3770 :
3771 188 : return fd;
3772 : }
3773 :
3774 : /*
3775 : * Close the current logfile segment for writing.
3776 : */
3777 : static void
3778 12668 : XLogFileClose(void)
3779 : {
3780 : Assert(openLogFile >= 0);
3781 :
3782 : /*
3783 : * WAL segment files will not be re-read in normal operation, so we advise
3784 : * the OS to release any cached pages. But do not do so if WAL archiving
3785 : * or streaming is active, because archiver and walsender process could
3786 : * use the cache to read the WAL segment.
3787 : */
3788 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3789 12668 : if (!XLogIsNeeded() && (io_direct_flags & IO_DIRECT_WAL) == 0)
3790 3044 : (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3791 : #endif
3792 :
3793 12668 : if (close(openLogFile) != 0)
3794 : {
3795 : char xlogfname[MAXFNAMELEN];
3796 0 : int save_errno = errno;
3797 :
3798 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo, wal_segment_size);
3799 0 : errno = save_errno;
3800 0 : ereport(PANIC,
3801 : (errcode_for_file_access(),
3802 : errmsg("could not close file \"%s\": %m", xlogfname)));
3803 : }
3804 :
3805 12668 : openLogFile = -1;
3806 12668 : ReleaseExternalFD();
3807 12668 : }
3808 :
3809 : /*
3810 : * Preallocate log files beyond the specified log endpoint.
3811 : *
3812 : * XXX this is currently extremely conservative, since it forces only one
3813 : * future log segment to exist, and even that only if we are 75% done with
3814 : * the current one. This is only appropriate for very low-WAL-volume systems.
3815 : * High-volume systems will be OK once they've built up a sufficient set of
3816 : * recycled log segments, but the startup transient is likely to include
3817 : * a lot of segment creations by foreground processes, which is not so good.
3818 : *
3819 : * XLogFileInitInternal() can ereport(ERROR). All known causes indicate big
3820 : * trouble; for example, a full filesystem is one cause. The checkpoint WAL
3821 : * and/or ControlFile updates already completed. If a RequestCheckpoint()
3822 : * initiated the present checkpoint and an ERROR ends this function, the
3823 : * command that called RequestCheckpoint() fails. That's not ideal, but it's
3824 : * not worth contorting more functions to use caller-specified elevel values.
3825 : * (With or without RequestCheckpoint(), an ERROR forestalls some inessential
3826 : * reporting and resource reclamation.)
3827 : */
3828 : static void
3829 3902 : PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli)
3830 : {
3831 : XLogSegNo _logSegNo;
3832 : int lf;
3833 : bool added;
3834 : char path[MAXPGPATH];
3835 : uint64 offset;
3836 :
3837 3902 : if (!XLogCtl->InstallXLogFileSegmentActive)
3838 20 : return; /* unlocked check says no */
3839 :
3840 3882 : XLByteToPrevSeg(endptr, _logSegNo, wal_segment_size);
3841 3882 : offset = XLogSegmentOffset(endptr - 1, wal_segment_size);
3842 3882 : if (offset >= (uint32) (0.75 * wal_segment_size))
3843 : {
3844 538 : _logSegNo++;
3845 538 : lf = XLogFileInitInternal(_logSegNo, tli, &added, path);
3846 538 : if (lf >= 0)
3847 224 : close(lf);
3848 538 : if (added)
3849 314 : CheckpointStats.ckpt_segs_added++;
3850 : }
3851 : }
3852 :
3853 : /*
3854 : * Throws an error if the given log segment has already been removed or
3855 : * recycled. The caller should only pass a segment that it knows to have
3856 : * existed while the server has been running, as this function always
3857 : * succeeds if no WAL segments have been removed since startup.
3858 : * 'tli' is only used in the error message.
3859 : *
3860 : * Note: this function guarantees to keep errno unchanged on return.
3861 : * This supports callers that use this to possibly deliver a better
3862 : * error message about a missing file, while still being able to throw
3863 : * a normal file-access error afterwards, if this does return.
3864 : */
3865 : void
3866 229730 : CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3867 : {
3868 229730 : int save_errno = errno;
3869 : XLogSegNo lastRemovedSegNo;
3870 :
3871 229730 : SpinLockAcquire(&XLogCtl->info_lck);
3872 229730 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3873 229730 : SpinLockRelease(&XLogCtl->info_lck);
3874 :
3875 229730 : if (segno <= lastRemovedSegNo)
3876 : {
3877 : char filename[MAXFNAMELEN];
3878 :
3879 0 : XLogFileName(filename, tli, segno, wal_segment_size);
3880 0 : errno = save_errno;
3881 0 : ereport(ERROR,
3882 : (errcode_for_file_access(),
3883 : errmsg("requested WAL segment %s has already been removed",
3884 : filename)));
3885 : }
3886 229730 : errno = save_errno;
3887 229730 : }
3888 :
3889 : /*
3890 : * Return the last WAL segment removed, or 0 if no segment has been removed
3891 : * since startup.
3892 : *
3893 : * NB: the result can be out of date arbitrarily fast, the caller has to deal
3894 : * with that.
3895 : */
3896 : XLogSegNo
3897 2114 : XLogGetLastRemovedSegno(void)
3898 : {
3899 : XLogSegNo lastRemovedSegNo;
3900 :
3901 2114 : SpinLockAcquire(&XLogCtl->info_lck);
3902 2114 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3903 2114 : SpinLockRelease(&XLogCtl->info_lck);
3904 :
3905 2114 : return lastRemovedSegNo;
3906 : }
3907 :
3908 : /*
3909 : * Return the oldest WAL segment on the given TLI that still exists in
3910 : * XLOGDIR, or 0 if none.
3911 : */
3912 : XLogSegNo
3913 10 : XLogGetOldestSegno(TimeLineID tli)
3914 : {
3915 : DIR *xldir;
3916 : struct dirent *xlde;
3917 10 : XLogSegNo oldest_segno = 0;
3918 :
3919 10 : xldir = AllocateDir(XLOGDIR);
3920 66 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3921 : {
3922 : TimeLineID file_tli;
3923 : XLogSegNo file_segno;
3924 :
3925 : /* Ignore files that are not XLOG segments. */
3926 56 : if (!IsXLogFileName(xlde->d_name))
3927 40 : continue;
3928 :
3929 : /* Parse filename to get TLI and segno. */
3930 16 : XLogFromFileName(xlde->d_name, &file_tli, &file_segno,
3931 : wal_segment_size);
3932 :
3933 : /* Ignore anything that's not from the TLI of interest. */
3934 16 : if (tli != file_tli)
3935 0 : continue;
3936 :
3937 : /* If it's the oldest so far, update oldest_segno. */
3938 16 : if (oldest_segno == 0 || file_segno < oldest_segno)
3939 14 : oldest_segno = file_segno;
3940 : }
3941 :
3942 10 : FreeDir(xldir);
3943 10 : return oldest_segno;
3944 : }
3945 :
3946 : /*
3947 : * Update the last removed segno pointer in shared memory, to reflect that the
3948 : * given XLOG file has been removed.
3949 : */
3950 : static void
3951 5058 : UpdateLastRemovedPtr(char *filename)
3952 : {
3953 : uint32 tli;
3954 : XLogSegNo segno;
3955 :
3956 5058 : XLogFromFileName(filename, &tli, &segno, wal_segment_size);
3957 :
3958 5058 : SpinLockAcquire(&XLogCtl->info_lck);
3959 5058 : if (segno > XLogCtl->lastRemovedSegNo)
3960 2104 : XLogCtl->lastRemovedSegNo = segno;
3961 5058 : SpinLockRelease(&XLogCtl->info_lck);
3962 5058 : }
3963 :
3964 : /*
3965 : * Remove all temporary log files in pg_wal
3966 : *
3967 : * This is called at the beginning of recovery after a previous crash,
3968 : * at a point where no other processes write fresh WAL data.
3969 : */
3970 : static void
3971 348 : RemoveTempXlogFiles(void)
3972 : {
3973 : DIR *xldir;
3974 : struct dirent *xlde;
3975 :
3976 348 : elog(DEBUG2, "removing all temporary WAL segments");
3977 :
3978 348 : xldir = AllocateDir(XLOGDIR);
3979 2292 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3980 : {
3981 : char path[MAXPGPATH];
3982 :
3983 1944 : if (strncmp(xlde->d_name, "xlogtemp.", 9) != 0)
3984 1942 : continue;
3985 :
3986 2 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
3987 2 : unlink(path);
3988 2 : elog(DEBUG2, "removed temporary WAL segment \"%s\"", path);
3989 : }
3990 348 : FreeDir(xldir);
3991 348 : }
3992 :
3993 : /*
3994 : * Recycle or remove all log files older or equal to passed segno.
3995 : *
3996 : * endptr is current (or recent) end of xlog, and lastredoptr is the
3997 : * redo pointer of the last checkpoint. These are used to determine
3998 : * whether we want to recycle rather than delete no-longer-wanted log files.
3999 : *
4000 : * insertTLI is the current timeline for XLOG insertion. Any recycled
4001 : * segments should be reused for this timeline.
4002 : */
4003 : static void
4004 3360 : RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr, XLogRecPtr endptr,
4005 : TimeLineID insertTLI)
4006 : {
4007 : DIR *xldir;
4008 : struct dirent *xlde;
4009 : char lastoff[MAXFNAMELEN];
4010 : XLogSegNo endlogSegNo;
4011 : XLogSegNo recycleSegNo;
4012 :
4013 : /* Initialize info about where to try to recycle to */
4014 3360 : XLByteToSeg(endptr, endlogSegNo, wal_segment_size);
4015 3360 : recycleSegNo = XLOGfileslop(lastredoptr);
4016 :
4017 : /*
4018 : * Construct a filename of the last segment to be kept. The timeline ID
4019 : * doesn't matter, we ignore that in the comparison. (During recovery,
4020 : * InsertTimeLineID isn't set, so we can't use that.)
4021 : */
4022 3360 : XLogFileName(lastoff, 0, segno, wal_segment_size);
4023 :
4024 3360 : elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
4025 : lastoff);
4026 :
4027 3360 : xldir = AllocateDir(XLOGDIR);
4028 :
4029 85838 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4030 : {
4031 : /* Ignore files that are not XLOG segments */
4032 82478 : if (!IsXLogFileName(xlde->d_name) &&
4033 14546 : !IsPartialXLogFileName(xlde->d_name))
4034 14542 : continue;
4035 :
4036 : /*
4037 : * We ignore the timeline part of the XLOG segment identifiers in
4038 : * deciding whether a segment is still needed. This ensures that we
4039 : * won't prematurely remove a segment from a parent timeline. We could
4040 : * probably be a little more proactive about removing segments of
4041 : * non-parent timelines, but that would be a whole lot more
4042 : * complicated.
4043 : *
4044 : * We use the alphanumeric sorting property of the filenames to decide
4045 : * which ones are earlier than the lastoff segment.
4046 : */
4047 67936 : if (strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
4048 : {
4049 42260 : if (XLogArchiveCheckDone(xlde->d_name))
4050 : {
4051 : /* Update the last removed location in shared memory first */
4052 5058 : UpdateLastRemovedPtr(xlde->d_name);
4053 :
4054 5058 : RemoveXlogFile(xlde, recycleSegNo, &endlogSegNo, insertTLI);
4055 : }
4056 : }
4057 : }
4058 :
4059 3360 : FreeDir(xldir);
4060 3360 : }
4061 :
4062 : /*
4063 : * Recycle or remove WAL files that are not part of the given timeline's
4064 : * history.
4065 : *
4066 : * This is called during recovery, whenever we switch to follow a new
4067 : * timeline, and at the end of recovery when we create a new timeline. We
4068 : * wouldn't otherwise care about extra WAL files lying in pg_wal, but they
4069 : * might be leftover pre-allocated or recycled WAL segments on the old timeline
4070 : * that we haven't used yet, and contain garbage. If we just leave them in
4071 : * pg_wal, they will eventually be archived, and we can't let that happen.
4072 : * Files that belong to our timeline history are valid, because we have
4073 : * successfully replayed them, but from others we can't be sure.
4074 : *
4075 : * 'switchpoint' is the current point in WAL where we switch to new timeline,
4076 : * and 'newTLI' is the new timeline we switch to.
4077 : */
4078 : void
4079 120 : RemoveNonParentXlogFiles(XLogRecPtr switchpoint, TimeLineID newTLI)
4080 : {
4081 : DIR *xldir;
4082 : struct dirent *xlde;
4083 : char switchseg[MAXFNAMELEN];
4084 : XLogSegNo endLogSegNo;
4085 : XLogSegNo switchLogSegNo;
4086 : XLogSegNo recycleSegNo;
4087 :
4088 : /*
4089 : * Initialize info about where to begin the work. This will recycle,
4090 : * somewhat arbitrarily, 10 future segments.
4091 : */
4092 120 : XLByteToPrevSeg(switchpoint, switchLogSegNo, wal_segment_size);
4093 120 : XLByteToSeg(switchpoint, endLogSegNo, wal_segment_size);
4094 120 : recycleSegNo = endLogSegNo + 10;
4095 :
4096 : /*
4097 : * Construct a filename of the last segment to be kept.
4098 : */
4099 120 : XLogFileName(switchseg, newTLI, switchLogSegNo, wal_segment_size);
4100 :
4101 120 : elog(DEBUG2, "attempting to remove WAL segments newer than log file %s",
4102 : switchseg);
4103 :
4104 120 : xldir = AllocateDir(XLOGDIR);
4105 :
4106 1142 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4107 : {
4108 : /* Ignore files that are not XLOG segments */
4109 1022 : if (!IsXLogFileName(xlde->d_name))
4110 636 : continue;
4111 :
4112 : /*
4113 : * Remove files that are on a timeline older than the new one we're
4114 : * switching to, but with a segment number >= the first segment on the
4115 : * new timeline.
4116 : */
4117 386 : if (strncmp(xlde->d_name, switchseg, 8) < 0 &&
4118 250 : strcmp(xlde->d_name + 8, switchseg + 8) > 0)
4119 : {
4120 : /*
4121 : * If the file has already been marked as .ready, however, don't
4122 : * remove it yet. It should be OK to remove it - files that are
4123 : * not part of our timeline history are not required for recovery
4124 : * - but seems safer to let them be archived and removed later.
4125 : */
4126 32 : if (!XLogArchiveIsReady(xlde->d_name))
4127 32 : RemoveXlogFile(xlde, recycleSegNo, &endLogSegNo, newTLI);
4128 : }
4129 : }
4130 :
4131 120 : FreeDir(xldir);
4132 120 : }
4133 :
4134 : /*
4135 : * Recycle or remove a log file that's no longer needed.
4136 : *
4137 : * segment_de is the dirent structure of the segment to recycle or remove.
4138 : * recycleSegNo is the segment number to recycle up to. endlogSegNo is
4139 : * the segment number of the current (or recent) end of WAL.
4140 : *
4141 : * endlogSegNo gets incremented if the segment is recycled so as it is not
4142 : * checked again with future callers of this function.
4143 : *
4144 : * insertTLI is the current timeline for XLOG insertion. Any recycled segments
4145 : * should be used for this timeline.
4146 : */
4147 : static void
4148 5090 : RemoveXlogFile(const struct dirent *segment_de,
4149 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
4150 : TimeLineID insertTLI)
4151 : {
4152 : char path[MAXPGPATH];
4153 : #ifdef WIN32
4154 : char newpath[MAXPGPATH];
4155 : #endif
4156 5090 : const char *segname = segment_de->d_name;
4157 :
4158 5090 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", segname);
4159 :
4160 : /*
4161 : * Before deleting the file, see if it can be recycled as a future log
4162 : * segment. Only recycle normal files, because we don't want to recycle
4163 : * symbolic links pointing to a separate archive directory.
4164 : */
4165 5090 : if (wal_recycle &&
4166 5090 : *endlogSegNo <= recycleSegNo &&
4167 6158 : XLogCtl->InstallXLogFileSegmentActive && /* callee rechecks this */
4168 5504 : get_dirent_type(path, segment_de, false, DEBUG2) == PGFILETYPE_REG &&
4169 2752 : InstallXLogFileSegment(endlogSegNo, path,
4170 : true, recycleSegNo, insertTLI))
4171 : {
4172 2702 : ereport(DEBUG2,
4173 : (errmsg_internal("recycled write-ahead log file \"%s\"",
4174 : segname)));
4175 2702 : CheckpointStats.ckpt_segs_recycled++;
4176 : /* Needn't recheck that slot on future iterations */
4177 2702 : (*endlogSegNo)++;
4178 : }
4179 : else
4180 : {
4181 : /* No need for any more future segments, or recycling failed ... */
4182 : int rc;
4183 :
4184 2388 : ereport(DEBUG2,
4185 : (errmsg_internal("removing write-ahead log file \"%s\"",
4186 : segname)));
4187 :
4188 : #ifdef WIN32
4189 :
4190 : /*
4191 : * On Windows, if another process (e.g another backend) holds the file
4192 : * open in FILE_SHARE_DELETE mode, unlink will succeed, but the file
4193 : * will still show up in directory listing until the last handle is
4194 : * closed. To avoid confusing the lingering deleted file for a live
4195 : * WAL file that needs to be archived, rename it before deleting it.
4196 : *
4197 : * If another process holds the file open without FILE_SHARE_DELETE
4198 : * flag, rename will fail. We'll try again at the next checkpoint.
4199 : */
4200 : snprintf(newpath, MAXPGPATH, "%s.deleted", path);
4201 : if (rename(path, newpath) != 0)
4202 : {
4203 : ereport(LOG,
4204 : (errcode_for_file_access(),
4205 : errmsg("could not rename file \"%s\": %m",
4206 : path)));
4207 : return;
4208 : }
4209 : rc = durable_unlink(newpath, LOG);
4210 : #else
4211 2388 : rc = durable_unlink(path, LOG);
4212 : #endif
4213 2388 : if (rc != 0)
4214 : {
4215 : /* Message already logged by durable_unlink() */
4216 0 : return;
4217 : }
4218 2388 : CheckpointStats.ckpt_segs_removed++;
4219 : }
4220 :
4221 5090 : XLogArchiveCleanup(segname);
4222 : }
4223 :
4224 : /*
4225 : * Verify whether pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
4226 : * If the latter do not exist, recreate them.
4227 : *
4228 : * It is not the goal of this function to verify the contents of these
4229 : * directories, but to help in cases where someone has performed a cluster
4230 : * copy for PITR purposes but omitted pg_wal from the copy.
4231 : *
4232 : * We could also recreate pg_wal if it doesn't exist, but a deliberate
4233 : * policy decision was made not to. It is fairly common for pg_wal to be
4234 : * a symlink, and if that was the DBA's intent then automatically making a
4235 : * plain directory would result in degraded performance with no notice.
4236 : */
4237 : static void
4238 1846 : ValidateXLOGDirectoryStructure(void)
4239 : {
4240 : char path[MAXPGPATH];
4241 : struct stat stat_buf;
4242 :
4243 : /* Check for pg_wal; if it doesn't exist, error out */
4244 1846 : if (stat(XLOGDIR, &stat_buf) != 0 ||
4245 1846 : !S_ISDIR(stat_buf.st_mode))
4246 0 : ereport(FATAL,
4247 : (errcode_for_file_access(),
4248 : errmsg("required WAL directory \"%s\" does not exist",
4249 : XLOGDIR)));
4250 :
4251 : /* Check for archive_status */
4252 1846 : snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4253 1846 : if (stat(path, &stat_buf) == 0)
4254 : {
4255 : /* Check for weird cases where it exists but isn't a directory */
4256 1844 : if (!S_ISDIR(stat_buf.st_mode))
4257 0 : ereport(FATAL,
4258 : (errcode_for_file_access(),
4259 : errmsg("required WAL directory \"%s\" does not exist",
4260 : path)));
4261 : }
4262 : else
4263 : {
4264 2 : ereport(LOG,
4265 : (errmsg("creating missing WAL directory \"%s\"", path)));
4266 2 : if (MakePGDirectory(path) < 0)
4267 0 : ereport(FATAL,
4268 : (errcode_for_file_access(),
4269 : errmsg("could not create missing directory \"%s\": %m",
4270 : path)));
4271 : }
4272 :
4273 : /* Check for summaries */
4274 1846 : snprintf(path, MAXPGPATH, XLOGDIR "/summaries");
4275 1846 : if (stat(path, &stat_buf) == 0)
4276 : {
4277 : /* Check for weird cases where it exists but isn't a directory */
4278 1844 : if (!S_ISDIR(stat_buf.st_mode))
4279 0 : ereport(FATAL,
4280 : (errmsg("required WAL directory \"%s\" does not exist",
4281 : path)));
4282 : }
4283 : else
4284 : {
4285 2 : ereport(LOG,
4286 : (errmsg("creating missing WAL directory \"%s\"", path)));
4287 2 : if (MakePGDirectory(path) < 0)
4288 0 : ereport(FATAL,
4289 : (errmsg("could not create missing directory \"%s\": %m",
4290 : path)));
4291 : }
4292 1846 : }
4293 :
4294 : /*
4295 : * Remove previous backup history files. This also retries creation of
4296 : * .ready files for any backup history files for which XLogArchiveNotify
4297 : * failed earlier.
4298 : */
4299 : static void
4300 296 : CleanupBackupHistory(void)
4301 : {
4302 : DIR *xldir;
4303 : struct dirent *xlde;
4304 : char path[MAXPGPATH + sizeof(XLOGDIR)];
4305 :
4306 296 : xldir = AllocateDir(XLOGDIR);
4307 :
4308 2996 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4309 : {
4310 2404 : if (IsBackupHistoryFileName(xlde->d_name))
4311 : {
4312 312 : if (XLogArchiveCheckDone(xlde->d_name))
4313 : {
4314 246 : elog(DEBUG2, "removing WAL backup history file \"%s\"",
4315 : xlde->d_name);
4316 246 : snprintf(path, sizeof(path), XLOGDIR "/%s", xlde->d_name);
4317 246 : unlink(path);
4318 246 : XLogArchiveCleanup(xlde->d_name);
4319 : }
4320 : }
4321 : }
4322 :
4323 296 : FreeDir(xldir);
4324 296 : }
4325 :
4326 : /*
4327 : * I/O routines for pg_control
4328 : *
4329 : * *ControlFile is a buffer in shared memory that holds an image of the
4330 : * contents of pg_control. WriteControlFile() initializes pg_control
4331 : * given a preloaded buffer, ReadControlFile() loads the buffer from
4332 : * the pg_control file (during postmaster or standalone-backend startup),
4333 : * and UpdateControlFile() rewrites pg_control after we modify xlog state.
4334 : * InitControlFile() fills the buffer with initial values.
4335 : *
4336 : * For simplicity, WriteControlFile() initializes the fields of pg_control
4337 : * that are related to checking backend/database compatibility, and
4338 : * ReadControlFile() verifies they are correct. We could split out the
4339 : * I/O and compatibility-check functions, but there seems no need currently.
4340 : */
4341 :
4342 : static void
4343 102 : InitControlFile(uint64 sysidentifier, uint32 data_checksum_version)
4344 : {
4345 : char mock_auth_nonce[MOCK_AUTH_NONCE_LEN];
4346 :
4347 : /*
4348 : * Generate a random nonce. This is used for authentication requests that
4349 : * will fail because the user does not exist. The nonce is used to create
4350 : * a genuine-looking password challenge for the non-existent user, in lieu
4351 : * of an actual stored password.
4352 : */
4353 102 : if (!pg_strong_random(mock_auth_nonce, MOCK_AUTH_NONCE_LEN))
4354 0 : ereport(PANIC,
4355 : (errcode(ERRCODE_INTERNAL_ERROR),
4356 : errmsg("could not generate secret authorization token")));
4357 :
4358 102 : memset(ControlFile, 0, sizeof(ControlFileData));
4359 : /* Initialize pg_control status fields */
4360 102 : ControlFile->system_identifier = sysidentifier;
4361 102 : memcpy(ControlFile->mock_authentication_nonce, mock_auth_nonce, MOCK_AUTH_NONCE_LEN);
4362 102 : ControlFile->state = DB_SHUTDOWNED;
4363 102 : ControlFile->unloggedLSN = FirstNormalUnloggedLSN;
4364 :
4365 : /* Set important parameter values for use when replaying WAL */
4366 102 : ControlFile->MaxConnections = MaxConnections;
4367 102 : ControlFile->max_worker_processes = max_worker_processes;
4368 102 : ControlFile->max_wal_senders = max_wal_senders;
4369 102 : ControlFile->max_prepared_xacts = max_prepared_xacts;
4370 102 : ControlFile->max_locks_per_xact = max_locks_per_xact;
4371 102 : ControlFile->wal_level = wal_level;
4372 102 : ControlFile->wal_log_hints = wal_log_hints;
4373 102 : ControlFile->track_commit_timestamp = track_commit_timestamp;
4374 102 : ControlFile->data_checksum_version = data_checksum_version;
4375 102 : }
4376 :
4377 : static void
4378 102 : WriteControlFile(void)
4379 : {
4380 : int fd;
4381 : char buffer[PG_CONTROL_FILE_SIZE]; /* need not be aligned */
4382 :
4383 : /*
4384 : * Initialize version and compatibility-check fields
4385 : */
4386 102 : ControlFile->pg_control_version = PG_CONTROL_VERSION;
4387 102 : ControlFile->catalog_version_no = CATALOG_VERSION_NO;
4388 :
4389 102 : ControlFile->maxAlign = MAXIMUM_ALIGNOF;
4390 102 : ControlFile->floatFormat = FLOATFORMAT_VALUE;
4391 :
4392 102 : ControlFile->blcksz = BLCKSZ;
4393 102 : ControlFile->relseg_size = RELSEG_SIZE;
4394 102 : ControlFile->xlog_blcksz = XLOG_BLCKSZ;
4395 102 : ControlFile->xlog_seg_size = wal_segment_size;
4396 :
4397 102 : ControlFile->nameDataLen = NAMEDATALEN;
4398 102 : ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
4399 :
4400 102 : ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
4401 102 : ControlFile->loblksize = LOBLKSIZE;
4402 :
4403 102 : ControlFile->float8ByVal = FLOAT8PASSBYVAL;
4404 :
4405 : /*
4406 : * Initialize the default 'char' signedness.
4407 : *
4408 : * The signedness of the char type is implementation-defined. For instance
4409 : * on x86 architecture CPUs, the char data type is typically treated as
4410 : * signed by default, whereas on aarch architecture CPUs, it is typically
4411 : * treated as unsigned by default. In v17 or earlier, we accidentally let
4412 : * C implementation signedness affect persistent data. This led to
4413 : * inconsistent results when comparing char data across different
4414 : * platforms.
4415 : *
4416 : * This flag can be used as a hint to ensure consistent behavior for
4417 : * pre-v18 data files that store data sorted by the 'char' type on disk,
4418 : * especially in cross-platform replication scenarios.
4419 : *
4420 : * Newly created database clusters unconditionally set the default char
4421 : * signedness to true. pg_upgrade changes this flag for clusters that were
4422 : * initialized on signedness=false platforms. As a result,
4423 : * signedness=false setting will become rare over time. If we had known
4424 : * about this problem during the last development cycle that forced initdb
4425 : * (v8.3), we would have made all clusters signed or all clusters
4426 : * unsigned. Making pg_upgrade the only source of signedness=false will
4427 : * cause the population of database clusters to converge toward that
4428 : * retrospective ideal.
4429 : */
4430 102 : ControlFile->default_char_signedness = true;
4431 :
4432 : /* Contents are protected with a CRC */
4433 102 : INIT_CRC32C(ControlFile->crc);
4434 102 : COMP_CRC32C(ControlFile->crc,
4435 : ControlFile,
4436 : offsetof(ControlFileData, crc));
4437 102 : FIN_CRC32C(ControlFile->crc);
4438 :
4439 : /*
4440 : * We write out PG_CONTROL_FILE_SIZE bytes into pg_control, zero-padding
4441 : * the excess over sizeof(ControlFileData). This reduces the odds of
4442 : * premature-EOF errors when reading pg_control. We'll still fail when we
4443 : * check the contents of the file, but hopefully with a more specific
4444 : * error than "couldn't read pg_control".
4445 : */
4446 102 : memset(buffer, 0, PG_CONTROL_FILE_SIZE);
4447 102 : memcpy(buffer, ControlFile, sizeof(ControlFileData));
4448 :
4449 102 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4450 : O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
4451 102 : if (fd < 0)
4452 0 : ereport(PANIC,
4453 : (errcode_for_file_access(),
4454 : errmsg("could not create file \"%s\": %m",
4455 : XLOG_CONTROL_FILE)));
4456 :
4457 102 : errno = 0;
4458 102 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_WRITE);
4459 102 : if (write(fd, buffer, PG_CONTROL_FILE_SIZE) != PG_CONTROL_FILE_SIZE)
4460 : {
4461 : /* if write didn't set errno, assume problem is no disk space */
4462 0 : if (errno == 0)
4463 0 : errno = ENOSPC;
4464 0 : ereport(PANIC,
4465 : (errcode_for_file_access(),
4466 : errmsg("could not write to file \"%s\": %m",
4467 : XLOG_CONTROL_FILE)));
4468 : }
4469 102 : pgstat_report_wait_end();
4470 :
4471 102 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_SYNC);
4472 102 : if (pg_fsync(fd) != 0)
4473 0 : ereport(PANIC,
4474 : (errcode_for_file_access(),
4475 : errmsg("could not fsync file \"%s\": %m",
4476 : XLOG_CONTROL_FILE)));
4477 102 : pgstat_report_wait_end();
4478 :
4479 102 : if (close(fd) != 0)
4480 0 : ereport(PANIC,
4481 : (errcode_for_file_access(),
4482 : errmsg("could not close file \"%s\": %m",
4483 : XLOG_CONTROL_FILE)));
4484 102 : }
4485 :
4486 : static void
4487 1946 : ReadControlFile(void)
4488 : {
4489 : pg_crc32c crc;
4490 : int fd;
4491 : char wal_segsz_str[20];
4492 : int r;
4493 :
4494 : /*
4495 : * Read data...
4496 : */
4497 1946 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4498 : O_RDWR | PG_BINARY);
4499 1946 : if (fd < 0)
4500 0 : ereport(PANIC,
4501 : (errcode_for_file_access(),
4502 : errmsg("could not open file \"%s\": %m",
4503 : XLOG_CONTROL_FILE)));
4504 :
4505 1946 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ);
4506 1946 : r = read(fd, ControlFile, sizeof(ControlFileData));
4507 1946 : if (r != sizeof(ControlFileData))
4508 : {
4509 0 : if (r < 0)
4510 0 : ereport(PANIC,
4511 : (errcode_for_file_access(),
4512 : errmsg("could not read file \"%s\": %m",
4513 : XLOG_CONTROL_FILE)));
4514 : else
4515 0 : ereport(PANIC,
4516 : (errcode(ERRCODE_DATA_CORRUPTED),
4517 : errmsg("could not read file \"%s\": read %d of %zu",
4518 : XLOG_CONTROL_FILE, r, sizeof(ControlFileData))));
4519 : }
4520 1946 : pgstat_report_wait_end();
4521 :
4522 1946 : close(fd);
4523 :
4524 : /*
4525 : * Check for expected pg_control format version. If this is wrong, the
4526 : * CRC check will likely fail because we'll be checking the wrong number
4527 : * of bytes. Complaining about wrong version will probably be more
4528 : * enlightening than complaining about wrong CRC.
4529 : */
4530 :
4531 1946 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
4532 0 : ereport(FATAL,
4533 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4534 : errmsg("database files are incompatible with server"),
4535 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4536 : " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4537 : ControlFile->pg_control_version, ControlFile->pg_control_version,
4538 : PG_CONTROL_VERSION, PG_CONTROL_VERSION),
4539 : errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4540 :
4541 1946 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
4542 0 : ereport(FATAL,
4543 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4544 : errmsg("database files are incompatible with server"),
4545 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4546 : " but the server was compiled with PG_CONTROL_VERSION %d.",
4547 : ControlFile->pg_control_version, PG_CONTROL_VERSION),
4548 : errhint("It looks like you need to initdb.")));
4549 :
4550 : /* Now check the CRC. */
4551 1946 : INIT_CRC32C(crc);
4552 1946 : COMP_CRC32C(crc,
4553 : ControlFile,
4554 : offsetof(ControlFileData, crc));
4555 1946 : FIN_CRC32C(crc);
4556 :
4557 1946 : if (!EQ_CRC32C(crc, ControlFile->crc))
4558 0 : ereport(FATAL,
4559 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4560 : errmsg("incorrect checksum in control file")));
4561 :
4562 : /*
4563 : * Do compatibility checking immediately. If the database isn't
4564 : * compatible with the backend executable, we want to abort before we can
4565 : * possibly do any damage.
4566 : */
4567 1946 : if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
4568 0 : ereport(FATAL,
4569 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4570 : errmsg("database files are incompatible with server"),
4571 : /* translator: %s is a variable name and %d is its value */
4572 : errdetail("The database cluster was initialized with %s %d,"
4573 : " but the server was compiled with %s %d.",
4574 : "CATALOG_VERSION_NO", ControlFile->catalog_version_no,
4575 : "CATALOG_VERSION_NO", CATALOG_VERSION_NO),
4576 : errhint("It looks like you need to initdb.")));
4577 1946 : if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
4578 0 : ereport(FATAL,
4579 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4580 : errmsg("database files are incompatible with server"),
4581 : /* translator: %s is a variable name and %d is its value */
4582 : errdetail("The database cluster was initialized with %s %d,"
4583 : " but the server was compiled with %s %d.",
4584 : "MAXALIGN", ControlFile->maxAlign,
4585 : "MAXALIGN", MAXIMUM_ALIGNOF),
4586 : errhint("It looks like you need to initdb.")));
4587 1946 : if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
4588 0 : ereport(FATAL,
4589 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4590 : errmsg("database files are incompatible with server"),
4591 : errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4592 : errhint("It looks like you need to initdb.")));
4593 1946 : if (ControlFile->blcksz != BLCKSZ)
4594 0 : ereport(FATAL,
4595 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4596 : errmsg("database files are incompatible with server"),
4597 : /* translator: %s is a variable name and %d is its value */
4598 : errdetail("The database cluster was initialized with %s %d,"
4599 : " but the server was compiled with %s %d.",
4600 : "BLCKSZ", ControlFile->blcksz,
4601 : "BLCKSZ", BLCKSZ),
4602 : errhint("It looks like you need to recompile or initdb.")));
4603 1946 : if (ControlFile->relseg_size != RELSEG_SIZE)
4604 0 : ereport(FATAL,
4605 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4606 : errmsg("database files are incompatible with server"),
4607 : /* translator: %s is a variable name and %d is its value */
4608 : errdetail("The database cluster was initialized with %s %d,"
4609 : " but the server was compiled with %s %d.",
4610 : "RELSEG_SIZE", ControlFile->relseg_size,
4611 : "RELSEG_SIZE", RELSEG_SIZE),
4612 : errhint("It looks like you need to recompile or initdb.")));
4613 1946 : if (ControlFile->xlog_blcksz != XLOG_BLCKSZ)
4614 0 : ereport(FATAL,
4615 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4616 : errmsg("database files are incompatible with server"),
4617 : /* translator: %s is a variable name and %d is its value */
4618 : errdetail("The database cluster was initialized with %s %d,"
4619 : " but the server was compiled with %s %d.",
4620 : "XLOG_BLCKSZ", ControlFile->xlog_blcksz,
4621 : "XLOG_BLCKSZ", XLOG_BLCKSZ),
4622 : errhint("It looks like you need to recompile or initdb.")));
4623 1946 : if (ControlFile->nameDataLen != NAMEDATALEN)
4624 0 : ereport(FATAL,
4625 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4626 : errmsg("database files are incompatible with server"),
4627 : /* translator: %s is a variable name and %d is its value */
4628 : errdetail("The database cluster was initialized with %s %d,"
4629 : " but the server was compiled with %s %d.",
4630 : "NAMEDATALEN", ControlFile->nameDataLen,
4631 : "NAMEDATALEN", NAMEDATALEN),
4632 : errhint("It looks like you need to recompile or initdb.")));
4633 1946 : if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
4634 0 : ereport(FATAL,
4635 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4636 : errmsg("database files are incompatible with server"),
4637 : /* translator: %s is a variable name and %d is its value */
4638 : errdetail("The database cluster was initialized with %s %d,"
4639 : " but the server was compiled with %s %d.",
4640 : "INDEX_MAX_KEYS", ControlFile->indexMaxKeys,
4641 : "INDEX_MAX_KEYS", INDEX_MAX_KEYS),
4642 : errhint("It looks like you need to recompile or initdb.")));
4643 1946 : if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
4644 0 : ereport(FATAL,
4645 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4646 : errmsg("database files are incompatible with server"),
4647 : /* translator: %s is a variable name and %d is its value */
4648 : errdetail("The database cluster was initialized with %s %d,"
4649 : " but the server was compiled with %s %d.",
4650 : "TOAST_MAX_CHUNK_SIZE", ControlFile->toast_max_chunk_size,
4651 : "TOAST_MAX_CHUNK_SIZE", (int) TOAST_MAX_CHUNK_SIZE),
4652 : errhint("It looks like you need to recompile or initdb.")));
4653 1946 : if (ControlFile->loblksize != LOBLKSIZE)
4654 0 : ereport(FATAL,
4655 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4656 : errmsg("database files are incompatible with server"),
4657 : /* translator: %s is a variable name and %d is its value */
4658 : errdetail("The database cluster was initialized with %s %d,"
4659 : " but the server was compiled with %s %d.",
4660 : "LOBLKSIZE", ControlFile->loblksize,
4661 : "LOBLKSIZE", (int) LOBLKSIZE),
4662 : errhint("It looks like you need to recompile or initdb.")));
4663 :
4664 : #ifdef USE_FLOAT8_BYVAL
4665 1946 : if (ControlFile->float8ByVal != true)
4666 0 : ereport(FATAL,
4667 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4668 : errmsg("database files are incompatible with server"),
4669 : errdetail("The database cluster was initialized without USE_FLOAT8_BYVAL"
4670 : " but the server was compiled with USE_FLOAT8_BYVAL."),
4671 : errhint("It looks like you need to recompile or initdb.")));
4672 : #else
4673 : if (ControlFile->float8ByVal != false)
4674 : ereport(FATAL,
4675 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4676 : errmsg("database files are incompatible with server"),
4677 : errdetail("The database cluster was initialized with USE_FLOAT8_BYVAL"
4678 : " but the server was compiled without USE_FLOAT8_BYVAL."),
4679 : errhint("It looks like you need to recompile or initdb.")));
4680 : #endif
4681 :
4682 1946 : wal_segment_size = ControlFile->xlog_seg_size;
4683 :
4684 1946 : if (!IsValidWalSegSize(wal_segment_size))
4685 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4686 : errmsg_plural("invalid WAL segment size in control file (%d byte)",
4687 : "invalid WAL segment size in control file (%d bytes)",
4688 : wal_segment_size,
4689 : wal_segment_size),
4690 : errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.")));
4691 :
4692 1946 : snprintf(wal_segsz_str, sizeof(wal_segsz_str), "%d", wal_segment_size);
4693 1946 : SetConfigOption("wal_segment_size", wal_segsz_str, PGC_INTERNAL,
4694 : PGC_S_DYNAMIC_DEFAULT);
4695 :
4696 : /* check and update variables dependent on wal_segment_size */
4697 1946 : if (ConvertToXSegs(min_wal_size_mb, wal_segment_size) < 2)
4698 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4699 : /* translator: both %s are GUC names */
4700 : errmsg("\"%s\" must be at least twice \"%s\"",
4701 : "min_wal_size", "wal_segment_size")));
4702 :
4703 1946 : if (ConvertToXSegs(max_wal_size_mb, wal_segment_size) < 2)
4704 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4705 : /* translator: both %s are GUC names */
4706 : errmsg("\"%s\" must be at least twice \"%s\"",
4707 : "max_wal_size", "wal_segment_size")));
4708 :
4709 1946 : UsableBytesInSegment =
4710 1946 : (wal_segment_size / XLOG_BLCKSZ * UsableBytesInPage) -
4711 : (SizeOfXLogLongPHD - SizeOfXLogShortPHD);
4712 :
4713 1946 : CalculateCheckpointSegments();
4714 :
4715 : /* Make the initdb settings visible as GUC variables, too */
4716 1946 : SetConfigOption("data_checksums", DataChecksumsEnabled() ? "yes" : "no",
4717 : PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
4718 1946 : }
4719 :
4720 : /*
4721 : * Utility wrapper to update the control file. Note that the control
4722 : * file gets flushed.
4723 : */
4724 : static void
4725 17958 : UpdateControlFile(void)
4726 : {
4727 17958 : update_controlfile(DataDir, ControlFile, true);
4728 17958 : }
4729 :
4730 : /*
4731 : * Returns the unique system identifier from control file.
4732 : */
4733 : uint64
4734 2712 : GetSystemIdentifier(void)
4735 : {
4736 : Assert(ControlFile != NULL);
4737 2712 : return ControlFile->system_identifier;
4738 : }
4739 :
4740 : /*
4741 : * Returns the random nonce from control file.
4742 : */
4743 : char *
4744 2 : GetMockAuthenticationNonce(void)
4745 : {
4746 : Assert(ControlFile != NULL);
4747 2 : return ControlFile->mock_authentication_nonce;
4748 : }
4749 :
4750 : /*
4751 : * Are checksums enabled for data pages?
4752 : */
4753 : bool
4754 20209898 : DataChecksumsEnabled(void)
4755 : {
4756 : Assert(ControlFile != NULL);
4757 20209898 : return (ControlFile->data_checksum_version > 0);
4758 : }
4759 :
4760 : /*
4761 : * Return true if the cluster was initialized on a platform where the
4762 : * default signedness of char is "signed". This function exists for code
4763 : * that deals with pre-v18 data files that store data sorted by the 'char'
4764 : * type on disk (e.g., GIN and GiST indexes). See the comments in
4765 : * WriteControlFile() for details.
4766 : */
4767 : bool
4768 6 : GetDefaultCharSignedness(void)
4769 : {
4770 6 : return ControlFile->default_char_signedness;
4771 : }
4772 :
4773 : /*
4774 : * Returns a fake LSN for unlogged relations.
4775 : *
4776 : * Each call generates an LSN that is greater than any previous value
4777 : * returned. The current counter value is saved and restored across clean
4778 : * shutdowns, but like unlogged relations, does not survive a crash. This can
4779 : * be used in lieu of real LSN values returned by XLogInsert, if you need an
4780 : * LSN-like increasing sequence of numbers without writing any WAL.
4781 : */
4782 : XLogRecPtr
4783 66 : GetFakeLSNForUnloggedRel(void)
4784 : {
4785 66 : return pg_atomic_fetch_add_u64(&XLogCtl->unloggedLSN, 1);
4786 : }
4787 :
4788 : /*
4789 : * Auto-tune the number of XLOG buffers.
4790 : *
4791 : * The preferred setting for wal_buffers is about 3% of shared_buffers, with
4792 : * a maximum of one XLOG segment (there is little reason to think that more
4793 : * is helpful, at least so long as we force an fsync when switching log files)
4794 : * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
4795 : * 9.1, when auto-tuning was added).
4796 : *
4797 : * This should not be called until NBuffers has received its final value.
4798 : */
4799 : static int
4800 2130 : XLOGChooseNumBuffers(void)
4801 : {
4802 : int xbuffers;
4803 :
4804 2130 : xbuffers = NBuffers / 32;
4805 2130 : if (xbuffers > (wal_segment_size / XLOG_BLCKSZ))
4806 54 : xbuffers = (wal_segment_size / XLOG_BLCKSZ);
4807 2130 : if (xbuffers < 8)
4808 778 : xbuffers = 8;
4809 2130 : return xbuffers;
4810 : }
4811 :
4812 : /*
4813 : * GUC check_hook for wal_buffers
4814 : */
4815 : bool
4816 4338 : check_wal_buffers(int *newval, void **extra, GucSource source)
4817 : {
4818 : /*
4819 : * -1 indicates a request for auto-tune.
4820 : */
4821 4338 : if (*newval == -1)
4822 : {
4823 : /*
4824 : * If we haven't yet changed the boot_val default of -1, just let it
4825 : * be. We'll fix it when XLOGShmemSize is called.
4826 : */
4827 2208 : if (XLOGbuffers == -1)
4828 2208 : return true;
4829 :
4830 : /* Otherwise, substitute the auto-tune value */
4831 0 : *newval = XLOGChooseNumBuffers();
4832 : }
4833 :
4834 : /*
4835 : * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
4836 : * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
4837 : * the case, we just silently treat such values as a request for the
4838 : * minimum. (We could throw an error instead, but that doesn't seem very
4839 : * helpful.)
4840 : */
4841 2130 : if (*newval < 4)
4842 0 : *newval = 4;
4843 :
4844 2130 : return true;
4845 : }
4846 :
4847 : /*
4848 : * GUC check_hook for wal_consistency_checking
4849 : */
4850 : bool
4851 4020 : check_wal_consistency_checking(char **newval, void **extra, GucSource source)
4852 : {
4853 : char *rawstring;
4854 : List *elemlist;
4855 : ListCell *l;
4856 : bool newwalconsistency[RM_MAX_ID + 1];
4857 :
4858 : /* Initialize the array */
4859 132660 : MemSet(newwalconsistency, 0, (RM_MAX_ID + 1) * sizeof(bool));
4860 :
4861 : /* Need a modifiable copy of string */
4862 4020 : rawstring = pstrdup(*newval);
4863 :
4864 : /* Parse string into list of identifiers */
4865 4020 : if (!SplitIdentifierString(rawstring, ',', &elemlist))
4866 : {
4867 : /* syntax error in list */
4868 0 : GUC_check_errdetail("List syntax is invalid.");
4869 0 : pfree(rawstring);
4870 0 : list_free(elemlist);
4871 0 : return false;
4872 : }
4873 :
4874 4932 : foreach(l, elemlist)
4875 : {
4876 912 : char *tok = (char *) lfirst(l);
4877 : int rmid;
4878 :
4879 : /* Check for 'all'. */
4880 912 : if (pg_strcasecmp(tok, "all") == 0)
4881 : {
4882 233356 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4883 232448 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL)
4884 9080 : newwalconsistency[rmid] = true;
4885 : }
4886 : else
4887 : {
4888 : /* Check if the token matches any known resource manager. */
4889 4 : bool found = false;
4890 :
4891 72 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4892 : {
4893 108 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL &&
4894 36 : pg_strcasecmp(tok, GetRmgr(rmid).rm_name) == 0)
4895 : {
4896 4 : newwalconsistency[rmid] = true;
4897 4 : found = true;
4898 4 : break;
4899 : }
4900 : }
4901 4 : if (!found)
4902 : {
4903 : /*
4904 : * During startup, it might be a not-yet-loaded custom
4905 : * resource manager. Defer checking until
4906 : * InitializeWalConsistencyChecking().
4907 : */
4908 0 : if (!process_shared_preload_libraries_done)
4909 : {
4910 0 : check_wal_consistency_checking_deferred = true;
4911 : }
4912 : else
4913 : {
4914 0 : GUC_check_errdetail("Unrecognized key word: \"%s\".", tok);
4915 0 : pfree(rawstring);
4916 0 : list_free(elemlist);
4917 0 : return false;
4918 : }
4919 : }
4920 : }
4921 : }
4922 :
4923 4020 : pfree(rawstring);
4924 4020 : list_free(elemlist);
4925 :
4926 : /* assign new value */
4927 4020 : *extra = guc_malloc(LOG, (RM_MAX_ID + 1) * sizeof(bool));
4928 4020 : if (!*extra)
4929 0 : return false;
4930 4020 : memcpy(*extra, newwalconsistency, (RM_MAX_ID + 1) * sizeof(bool));
4931 4020 : return true;
4932 : }
4933 :
4934 : /*
4935 : * GUC assign_hook for wal_consistency_checking
4936 : */
4937 : void
4938 4018 : assign_wal_consistency_checking(const char *newval, void *extra)
4939 : {
4940 : /*
4941 : * If some checks were deferred, it's possible that the checks will fail
4942 : * later during InitializeWalConsistencyChecking(). But in that case, the
4943 : * postmaster will exit anyway, so it's safe to proceed with the
4944 : * assignment.
4945 : *
4946 : * Any built-in resource managers specified are assigned immediately,
4947 : * which affects WAL created before shared_preload_libraries are
4948 : * processed. Any custom resource managers specified won't be assigned
4949 : * until after shared_preload_libraries are processed, but that's OK
4950 : * because WAL for a custom resource manager can't be written before the
4951 : * module is loaded anyway.
4952 : */
4953 4018 : wal_consistency_checking = extra;
4954 4018 : }
4955 :
4956 : /*
4957 : * InitializeWalConsistencyChecking: run after loading custom resource managers
4958 : *
4959 : * If any unknown resource managers were specified in the
4960 : * wal_consistency_checking GUC, processing was deferred. Now that
4961 : * shared_preload_libraries have been loaded, process wal_consistency_checking
4962 : * again.
4963 : */
4964 : void
4965 1824 : InitializeWalConsistencyChecking(void)
4966 : {
4967 : Assert(process_shared_preload_libraries_done);
4968 :
4969 1824 : if (check_wal_consistency_checking_deferred)
4970 : {
4971 : struct config_generic *guc;
4972 :
4973 0 : guc = find_option("wal_consistency_checking", false, false, ERROR);
4974 :
4975 0 : check_wal_consistency_checking_deferred = false;
4976 :
4977 0 : set_config_option_ext("wal_consistency_checking",
4978 : wal_consistency_checking_string,
4979 : guc->scontext, guc->source, guc->srole,
4980 : GUC_ACTION_SET, true, ERROR, false);
4981 :
4982 : /* checking should not be deferred again */
4983 : Assert(!check_wal_consistency_checking_deferred);
4984 : }
4985 1824 : }
4986 :
4987 : /*
4988 : * GUC show_hook for archive_command
4989 : */
4990 : const char *
4991 3544 : show_archive_command(void)
4992 : {
4993 3544 : if (XLogArchivingActive())
4994 4 : return XLogArchiveCommand;
4995 : else
4996 3540 : return "(disabled)";
4997 : }
4998 :
4999 : /*
5000 : * GUC show_hook for in_hot_standby
5001 : */
5002 : const char *
5003 29516 : show_in_hot_standby(void)
5004 : {
5005 : /*
5006 : * We display the actual state based on shared memory, so that this GUC
5007 : * reports up-to-date state if examined intra-query. The underlying
5008 : * variable (in_hot_standby_guc) changes only when we transmit a new value
5009 : * to the client.
5010 : */
5011 29516 : return RecoveryInProgress() ? "on" : "off";
5012 : }
5013 :
5014 : /*
5015 : * Read the control file, set respective GUCs.
5016 : *
5017 : * This is to be called during startup, including a crash recovery cycle,
5018 : * unless in bootstrap mode, where no control file yet exists. As there's no
5019 : * usable shared memory yet (its sizing can depend on the contents of the
5020 : * control file!), first store the contents in local memory. XLOGShmemInit()
5021 : * will then copy it to shared memory later.
5022 : *
5023 : * reset just controls whether previous contents are to be expected (in the
5024 : * reset case, there's a dangling pointer into old shared memory), or not.
5025 : */
5026 : void
5027 1844 : LocalProcessControlFile(bool reset)
5028 : {
5029 : Assert(reset || ControlFile == NULL);
5030 1844 : ControlFile = palloc(sizeof(ControlFileData));
5031 1844 : ReadControlFile();
5032 1844 : }
5033 :
5034 : /*
5035 : * Get the wal_level from the control file. For a standby, this value should be
5036 : * considered as its active wal_level, because it may be different from what
5037 : * was originally configured on standby.
5038 : */
5039 : WalLevel
5040 138 : GetActiveWalLevelOnStandby(void)
5041 : {
5042 138 : return ControlFile->wal_level;
5043 : }
5044 :
5045 : /*
5046 : * Initialization of shared memory for XLOG
5047 : */
5048 : Size
5049 6096 : XLOGShmemSize(void)
5050 : {
5051 : Size size;
5052 :
5053 : /*
5054 : * If the value of wal_buffers is -1, use the preferred auto-tune value.
5055 : * This isn't an amazingly clean place to do this, but we must wait till
5056 : * NBuffers has received its final value, and must do it before using the
5057 : * value of XLOGbuffers to do anything important.
5058 : *
5059 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
5060 : * However, if the DBA explicitly set wal_buffers = -1 in the config file,
5061 : * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
5062 : * the matter with PGC_S_OVERRIDE.
5063 : */
5064 6096 : if (XLOGbuffers == -1)
5065 : {
5066 : char buf[32];
5067 :
5068 2130 : snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
5069 2130 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
5070 : PGC_S_DYNAMIC_DEFAULT);
5071 2130 : if (XLOGbuffers == -1) /* failed to apply it? */
5072 0 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
5073 : PGC_S_OVERRIDE);
5074 : }
5075 : Assert(XLOGbuffers > 0);
5076 :
5077 : /* XLogCtl */
5078 6096 : size = sizeof(XLogCtlData);
5079 :
5080 : /* WAL insertion locks, plus alignment */
5081 6096 : size = add_size(size, mul_size(sizeof(WALInsertLockPadded), NUM_XLOGINSERT_LOCKS + 1));
5082 : /* xlblocks array */
5083 6096 : size = add_size(size, mul_size(sizeof(pg_atomic_uint64), XLOGbuffers));
5084 : /* extra alignment padding for XLOG I/O buffers */
5085 6096 : size = add_size(size, Max(XLOG_BLCKSZ, PG_IO_ALIGN_SIZE));
5086 : /* and the buffers themselves */
5087 6096 : size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
5088 :
5089 : /*
5090 : * Note: we don't count ControlFileData, it comes out of the "slop factor"
5091 : * added by CreateSharedMemoryAndSemaphores. This lets us use this
5092 : * routine again below to compute the actual allocation size.
5093 : */
5094 :
5095 6096 : return size;
5096 : }
5097 :
5098 : void
5099 2134 : XLOGShmemInit(void)
5100 : {
5101 : bool foundCFile,
5102 : foundXLog;
5103 : char *allocptr;
5104 : int i;
5105 : ControlFileData *localControlFile;
5106 :
5107 : #ifdef WAL_DEBUG
5108 :
5109 : /*
5110 : * Create a memory context for WAL debugging that's exempt from the normal
5111 : * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
5112 : * an allocation fails, but wal_debug is not for production use anyway.
5113 : */
5114 : if (walDebugCxt == NULL)
5115 : {
5116 : walDebugCxt = AllocSetContextCreate(TopMemoryContext,
5117 : "WAL Debug",
5118 : ALLOCSET_DEFAULT_SIZES);
5119 : MemoryContextAllowInCriticalSection(walDebugCxt, true);
5120 : }
5121 : #endif
5122 :
5123 :
5124 2134 : XLogCtl = (XLogCtlData *)
5125 2134 : ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog);
5126 :
5127 2134 : localControlFile = ControlFile;
5128 2134 : ControlFile = (ControlFileData *)
5129 2134 : ShmemInitStruct("Control File", sizeof(ControlFileData), &foundCFile);
5130 :
5131 2134 : if (foundCFile || foundXLog)
5132 : {
5133 : /* both should be present or neither */
5134 : Assert(foundCFile && foundXLog);
5135 :
5136 : /* Initialize local copy of WALInsertLocks */
5137 0 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks;
5138 :
5139 0 : if (localControlFile)
5140 0 : pfree(localControlFile);
5141 0 : return;
5142 : }
5143 2134 : memset(XLogCtl, 0, sizeof(XLogCtlData));
5144 :
5145 : /*
5146 : * Already have read control file locally, unless in bootstrap mode. Move
5147 : * contents into shared memory.
5148 : */
5149 2134 : if (localControlFile)
5150 : {
5151 1828 : memcpy(ControlFile, localControlFile, sizeof(ControlFileData));
5152 1828 : pfree(localControlFile);
5153 : }
5154 :
5155 : /*
5156 : * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
5157 : * multiple of the alignment for same, so no extra alignment padding is
5158 : * needed here.
5159 : */
5160 2134 : allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
5161 2134 : XLogCtl->xlblocks = (pg_atomic_uint64 *) allocptr;
5162 2134 : allocptr += sizeof(pg_atomic_uint64) * XLOGbuffers;
5163 :
5164 629808 : for (i = 0; i < XLOGbuffers; i++)
5165 : {
5166 627674 : pg_atomic_init_u64(&XLogCtl->xlblocks[i], InvalidXLogRecPtr);
5167 : }
5168 :
5169 : /* WAL insertion locks. Ensure they're aligned to the full padded size */
5170 2134 : allocptr += sizeof(WALInsertLockPadded) -
5171 2134 : ((uintptr_t) allocptr) % sizeof(WALInsertLockPadded);
5172 2134 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks =
5173 : (WALInsertLockPadded *) allocptr;
5174 2134 : allocptr += sizeof(WALInsertLockPadded) * NUM_XLOGINSERT_LOCKS;
5175 :
5176 19206 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
5177 : {
5178 17072 : LWLockInitialize(&WALInsertLocks[i].l.lock, LWTRANCHE_WAL_INSERT);
5179 17072 : pg_atomic_init_u64(&WALInsertLocks[i].l.insertingAt, InvalidXLogRecPtr);
5180 17072 : WALInsertLocks[i].l.lastImportantAt = InvalidXLogRecPtr;
5181 : }
5182 :
5183 : /*
5184 : * Align the start of the page buffers to a full xlog block size boundary.
5185 : * This simplifies some calculations in XLOG insertion. It is also
5186 : * required for O_DIRECT.
5187 : */
5188 2134 : allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5189 2134 : XLogCtl->pages = allocptr;
5190 2134 : memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5191 :
5192 : /*
5193 : * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5194 : * in additional info.)
5195 : */
5196 2134 : XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5197 2134 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5198 2134 : XLogCtl->InstallXLogFileSegmentActive = false;
5199 2134 : XLogCtl->WalWriterSleeping = false;
5200 :
5201 2134 : SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5202 2134 : SpinLockInit(&XLogCtl->info_lck);
5203 2134 : pg_atomic_init_u64(&XLogCtl->logInsertResult, InvalidXLogRecPtr);
5204 2134 : pg_atomic_init_u64(&XLogCtl->logWriteResult, InvalidXLogRecPtr);
5205 2134 : pg_atomic_init_u64(&XLogCtl->logFlushResult, InvalidXLogRecPtr);
5206 2134 : pg_atomic_init_u64(&XLogCtl->unloggedLSN, InvalidXLogRecPtr);
5207 :
5208 2134 : pg_atomic_init_u64(&XLogCtl->InitializeReserved, InvalidXLogRecPtr);
5209 2134 : pg_atomic_init_u64(&XLogCtl->InitializedUpTo, InvalidXLogRecPtr);
5210 2134 : ConditionVariableInit(&XLogCtl->InitializedUpToCondVar);
5211 : }
5212 :
5213 : /*
5214 : * This func must be called ONCE on system install. It creates pg_control
5215 : * and the initial XLOG segment.
5216 : */
5217 : void
5218 102 : BootStrapXLOG(uint32 data_checksum_version)
5219 : {
5220 : CheckPoint checkPoint;
5221 : char *buffer;
5222 : XLogPageHeader page;
5223 : XLogLongPageHeader longpage;
5224 : XLogRecord *record;
5225 : char *recptr;
5226 : uint64 sysidentifier;
5227 : struct timeval tv;
5228 : pg_crc32c crc;
5229 :
5230 : /* allow ordinary WAL segment creation, like StartupXLOG() would */
5231 102 : SetInstallXLogFileSegmentActive();
5232 :
5233 : /*
5234 : * Select a hopefully-unique system identifier code for this installation.
5235 : * We use the result of gettimeofday(), including the fractional seconds
5236 : * field, as being about as unique as we can easily get. (Think not to
5237 : * use random(), since it hasn't been seeded and there's no portable way
5238 : * to seed it other than the system clock value...) The upper half of the
5239 : * uint64 value is just the tv_sec part, while the lower half contains the
5240 : * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
5241 : * PID for a little extra uniqueness. A person knowing this encoding can
5242 : * determine the initialization time of the installation, which could
5243 : * perhaps be useful sometimes.
5244 : */
5245 102 : gettimeofday(&tv, NULL);
5246 102 : sysidentifier = ((uint64) tv.tv_sec) << 32;
5247 102 : sysidentifier |= ((uint64) tv.tv_usec) << 12;
5248 102 : sysidentifier |= getpid() & 0xFFF;
5249 :
5250 : /* page buffer must be aligned suitably for O_DIRECT */
5251 102 : buffer = (char *) palloc(XLOG_BLCKSZ + XLOG_BLCKSZ);
5252 102 : page = (XLogPageHeader) TYPEALIGN(XLOG_BLCKSZ, buffer);
5253 102 : memset(page, 0, XLOG_BLCKSZ);
5254 :
5255 : /*
5256 : * Set up information for the initial checkpoint record
5257 : *
5258 : * The initial checkpoint record is written to the beginning of the WAL
5259 : * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5260 : * used, so that we can use 0/0 to mean "before any valid WAL segment".
5261 : */
5262 102 : checkPoint.redo = wal_segment_size + SizeOfXLogLongPHD;
5263 102 : checkPoint.ThisTimeLineID = BootstrapTimeLineID;
5264 102 : checkPoint.PrevTimeLineID = BootstrapTimeLineID;
5265 102 : checkPoint.fullPageWrites = fullPageWrites;
5266 102 : checkPoint.wal_level = wal_level;
5267 : checkPoint.nextXid =
5268 102 : FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId);
5269 102 : checkPoint.nextOid = FirstGenbkiObjectId;
5270 102 : checkPoint.nextMulti = FirstMultiXactId;
5271 102 : checkPoint.nextMultiOffset = 0;
5272 102 : checkPoint.oldestXid = FirstNormalTransactionId;
5273 102 : checkPoint.oldestXidDB = Template1DbOid;
5274 102 : checkPoint.oldestMulti = FirstMultiXactId;
5275 102 : checkPoint.oldestMultiDB = Template1DbOid;
5276 102 : checkPoint.oldestCommitTsXid = InvalidTransactionId;
5277 102 : checkPoint.newestCommitTsXid = InvalidTransactionId;
5278 102 : checkPoint.time = (pg_time_t) time(NULL);
5279 102 : checkPoint.oldestActiveXid = InvalidTransactionId;
5280 :
5281 102 : TransamVariables->nextXid = checkPoint.nextXid;
5282 102 : TransamVariables->nextOid = checkPoint.nextOid;
5283 102 : TransamVariables->oidCount = 0;
5284 102 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5285 102 : AdvanceOldestClogXid(checkPoint.oldestXid);
5286 102 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5287 102 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5288 102 : SetCommitTsLimit(InvalidTransactionId, InvalidTransactionId);
5289 :
5290 : /* Set up the XLOG page header */
5291 102 : page->xlp_magic = XLOG_PAGE_MAGIC;
5292 102 : page->xlp_info = XLP_LONG_HEADER;
5293 102 : page->xlp_tli = BootstrapTimeLineID;
5294 102 : page->xlp_pageaddr = wal_segment_size;
5295 102 : longpage = (XLogLongPageHeader) page;
5296 102 : longpage->xlp_sysid = sysidentifier;
5297 102 : longpage->xlp_seg_size = wal_segment_size;
5298 102 : longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5299 :
5300 : /* Insert the initial checkpoint record */
5301 102 : recptr = ((char *) page + SizeOfXLogLongPHD);
5302 102 : record = (XLogRecord *) recptr;
5303 102 : record->xl_prev = 0;
5304 102 : record->xl_xid = InvalidTransactionId;
5305 102 : record->xl_tot_len = SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(checkPoint);
5306 102 : record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5307 102 : record->xl_rmid = RM_XLOG_ID;
5308 102 : recptr += SizeOfXLogRecord;
5309 : /* fill the XLogRecordDataHeaderShort struct */
5310 102 : *(recptr++) = (char) XLR_BLOCK_ID_DATA_SHORT;
5311 102 : *(recptr++) = sizeof(checkPoint);
5312 102 : memcpy(recptr, &checkPoint, sizeof(checkPoint));
5313 102 : recptr += sizeof(checkPoint);
5314 : Assert(recptr - (char *) record == record->xl_tot_len);
5315 :
5316 102 : INIT_CRC32C(crc);
5317 102 : COMP_CRC32C(crc, ((char *) record) + SizeOfXLogRecord, record->xl_tot_len - SizeOfXLogRecord);
5318 102 : COMP_CRC32C(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5319 102 : FIN_CRC32C(crc);
5320 102 : record->xl_crc = crc;
5321 :
5322 : /* Create first XLOG segment file */
5323 102 : openLogTLI = BootstrapTimeLineID;
5324 102 : openLogFile = XLogFileInit(1, BootstrapTimeLineID);
5325 :
5326 : /*
5327 : * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
5328 : * close the file again in a moment.
5329 : */
5330 :
5331 : /* Write the first page with the initial record */
5332 102 : errno = 0;
5333 102 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE);
5334 102 : if (write(openLogFile, page, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5335 : {
5336 : /* if write didn't set errno, assume problem is no disk space */
5337 0 : if (errno == 0)
5338 0 : errno = ENOSPC;
5339 0 : ereport(PANIC,
5340 : (errcode_for_file_access(),
5341 : errmsg("could not write bootstrap write-ahead log file: %m")));
5342 : }
5343 102 : pgstat_report_wait_end();
5344 :
5345 102 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC);
5346 102 : if (pg_fsync(openLogFile) != 0)
5347 0 : ereport(PANIC,
5348 : (errcode_for_file_access(),
5349 : errmsg("could not fsync bootstrap write-ahead log file: %m")));
5350 102 : pgstat_report_wait_end();
5351 :
5352 102 : if (close(openLogFile) != 0)
5353 0 : ereport(PANIC,
5354 : (errcode_for_file_access(),
5355 : errmsg("could not close bootstrap write-ahead log file: %m")));
5356 :
5357 102 : openLogFile = -1;
5358 :
5359 : /* Now create pg_control */
5360 102 : InitControlFile(sysidentifier, data_checksum_version);
5361 102 : ControlFile->time = checkPoint.time;
5362 102 : ControlFile->checkPoint = checkPoint.redo;
5363 102 : ControlFile->checkPointCopy = checkPoint;
5364 :
5365 : /* some additional ControlFile fields are set in WriteControlFile() */
5366 102 : WriteControlFile();
5367 :
5368 : /* Bootstrap the commit log, too */
5369 102 : BootStrapCLOG();
5370 102 : BootStrapCommitTs();
5371 102 : BootStrapSUBTRANS();
5372 102 : BootStrapMultiXact();
5373 :
5374 102 : pfree(buffer);
5375 :
5376 : /*
5377 : * Force control file to be read - in contrast to normal processing we'd
5378 : * otherwise never run the checks and GUC related initializations therein.
5379 : */
5380 102 : ReadControlFile();
5381 102 : }
5382 :
5383 : static char *
5384 1642 : str_time(pg_time_t tnow)
5385 : {
5386 1642 : char *buf = palloc(128);
5387 :
5388 1642 : pg_strftime(buf, 128,
5389 : "%Y-%m-%d %H:%M:%S %Z",
5390 1642 : pg_localtime(&tnow, log_timezone));
5391 :
5392 1642 : return buf;
5393 : }
5394 :
5395 : /*
5396 : * Initialize the first WAL segment on new timeline.
5397 : */
5398 : static void
5399 98 : XLogInitNewTimeline(TimeLineID endTLI, XLogRecPtr endOfLog, TimeLineID newTLI)
5400 : {
5401 : char xlogfname[MAXFNAMELEN];
5402 : XLogSegNo endLogSegNo;
5403 : XLogSegNo startLogSegNo;
5404 :
5405 : /* we always switch to a new timeline after archive recovery */
5406 : Assert(endTLI != newTLI);
5407 :
5408 : /*
5409 : * Update min recovery point one last time.
5410 : */
5411 98 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5412 :
5413 : /*
5414 : * Calculate the last segment on the old timeline, and the first segment
5415 : * on the new timeline. If the switch happens in the middle of a segment,
5416 : * they are the same, but if the switch happens exactly at a segment
5417 : * boundary, startLogSegNo will be endLogSegNo + 1.
5418 : */
5419 98 : XLByteToPrevSeg(endOfLog, endLogSegNo, wal_segment_size);
5420 98 : XLByteToSeg(endOfLog, startLogSegNo, wal_segment_size);
5421 :
5422 : /*
5423 : * Initialize the starting WAL segment for the new timeline. If the switch
5424 : * happens in the middle of a segment, copy data from the last WAL segment
5425 : * of the old timeline up to the switch point, to the starting WAL segment
5426 : * on the new timeline.
5427 : */
5428 98 : if (endLogSegNo == startLogSegNo)
5429 : {
5430 : /*
5431 : * Make a copy of the file on the new timeline.
5432 : *
5433 : * Writing WAL isn't allowed yet, so there are no locking
5434 : * considerations. But we should be just as tense as XLogFileInit to
5435 : * avoid emplacing a bogus file.
5436 : */
5437 78 : XLogFileCopy(newTLI, endLogSegNo, endTLI, endLogSegNo,
5438 78 : XLogSegmentOffset(endOfLog, wal_segment_size));
5439 : }
5440 : else
5441 : {
5442 : /*
5443 : * The switch happened at a segment boundary, so just create the next
5444 : * segment on the new timeline.
5445 : */
5446 : int fd;
5447 :
5448 20 : fd = XLogFileInit(startLogSegNo, newTLI);
5449 :
5450 20 : if (close(fd) != 0)
5451 : {
5452 0 : int save_errno = errno;
5453 :
5454 0 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5455 0 : errno = save_errno;
5456 0 : ereport(ERROR,
5457 : (errcode_for_file_access(),
5458 : errmsg("could not close file \"%s\": %m", xlogfname)));
5459 : }
5460 : }
5461 :
5462 : /*
5463 : * Let's just make real sure there are not .ready or .done flags posted
5464 : * for the new segment.
5465 : */
5466 98 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5467 98 : XLogArchiveCleanup(xlogfname);
5468 98 : }
5469 :
5470 : /*
5471 : * Perform cleanup actions at the conclusion of archive recovery.
5472 : */
5473 : static void
5474 98 : CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI, XLogRecPtr EndOfLog,
5475 : TimeLineID newTLI)
5476 : {
5477 : /*
5478 : * Execute the recovery_end_command, if any.
5479 : */
5480 98 : if (recoveryEndCommand && strcmp(recoveryEndCommand, "") != 0)
5481 4 : ExecuteRecoveryCommand(recoveryEndCommand,
5482 : "recovery_end_command",
5483 : true,
5484 : WAIT_EVENT_RECOVERY_END_COMMAND);
5485 :
5486 : /*
5487 : * We switched to a new timeline. Clean up segments on the old timeline.
5488 : *
5489 : * If there are any higher-numbered segments on the old timeline, remove
5490 : * them. They might contain valid WAL, but they might also be
5491 : * pre-allocated files containing garbage. In any case, they are not part
5492 : * of the new timeline's history so we don't need them.
5493 : */
5494 98 : RemoveNonParentXlogFiles(EndOfLog, newTLI);
5495 :
5496 : /*
5497 : * If the switch happened in the middle of a segment, what to do with the
5498 : * last, partial segment on the old timeline? If we don't archive it, and
5499 : * the server that created the WAL never archives it either (e.g. because
5500 : * it was hit by a meteor), it will never make it to the archive. That's
5501 : * OK from our point of view, because the new segment that we created with
5502 : * the new TLI contains all the WAL from the old timeline up to the switch
5503 : * point. But if you later try to do PITR to the "missing" WAL on the old
5504 : * timeline, recovery won't find it in the archive. It's physically
5505 : * present in the new file with new TLI, but recovery won't look there
5506 : * when it's recovering to the older timeline. On the other hand, if we
5507 : * archive the partial segment, and the original server on that timeline
5508 : * is still running and archives the completed version of the same segment
5509 : * later, it will fail. (We used to do that in 9.4 and below, and it
5510 : * caused such problems).
5511 : *
5512 : * As a compromise, we rename the last segment with the .partial suffix,
5513 : * and archive it. Archive recovery will never try to read .partial
5514 : * segments, so they will normally go unused. But in the odd PITR case,
5515 : * the administrator can copy them manually to the pg_wal directory
5516 : * (removing the suffix). They can be useful in debugging, too.
5517 : *
5518 : * If a .done or .ready file already exists for the old timeline, however,
5519 : * we had already determined that the segment is complete, so we can let
5520 : * it be archived normally. (In particular, if it was restored from the
5521 : * archive to begin with, it's expected to have a .done file).
5522 : */
5523 98 : if (XLogSegmentOffset(EndOfLog, wal_segment_size) != 0 &&
5524 : XLogArchivingActive())
5525 : {
5526 : char origfname[MAXFNAMELEN];
5527 : XLogSegNo endLogSegNo;
5528 :
5529 18 : XLByteToPrevSeg(EndOfLog, endLogSegNo, wal_segment_size);
5530 18 : XLogFileName(origfname, EndOfLogTLI, endLogSegNo, wal_segment_size);
5531 :
5532 18 : if (!XLogArchiveIsReadyOrDone(origfname))
5533 : {
5534 : char origpath[MAXPGPATH];
5535 : char partialfname[MAXFNAMELEN];
5536 : char partialpath[MAXPGPATH];
5537 :
5538 : /*
5539 : * If we're summarizing WAL, we can't rename the partial file
5540 : * until the summarizer finishes with it, else it will fail.
5541 : */
5542 10 : if (summarize_wal)
5543 2 : WaitForWalSummarization(EndOfLog);
5544 :
5545 10 : XLogFilePath(origpath, EndOfLogTLI, endLogSegNo, wal_segment_size);
5546 10 : snprintf(partialfname, MAXFNAMELEN, "%s.partial", origfname);
5547 10 : snprintf(partialpath, MAXPGPATH, "%s.partial", origpath);
5548 :
5549 : /*
5550 : * Make sure there's no .done or .ready file for the .partial
5551 : * file.
5552 : */
5553 10 : XLogArchiveCleanup(partialfname);
5554 :
5555 10 : durable_rename(origpath, partialpath, ERROR);
5556 10 : XLogArchiveNotify(partialfname);
5557 : }
5558 : }
5559 98 : }
5560 :
5561 : /*
5562 : * Check to see if required parameters are set high enough on this server
5563 : * for various aspects of recovery operation.
5564 : *
5565 : * Note that all the parameters which this function tests need to be
5566 : * listed in Administrator's Overview section in high-availability.sgml.
5567 : * If you change them, don't forget to update the list.
5568 : */
5569 : static void
5570 488 : CheckRequiredParameterValues(void)
5571 : {
5572 : /*
5573 : * For archive recovery, the WAL must be generated with at least 'replica'
5574 : * wal_level.
5575 : */
5576 488 : if (ArchiveRecoveryRequested && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5577 : {
5578 4 : ereport(FATAL,
5579 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5580 : errmsg("WAL was generated with \"wal_level=minimal\", cannot continue recovering"),
5581 : errdetail("This happens if you temporarily set \"wal_level=minimal\" on the server."),
5582 : errhint("Use a backup taken after setting \"wal_level\" to higher than \"minimal\".")));
5583 : }
5584 :
5585 : /*
5586 : * For Hot Standby, the WAL must be generated with 'replica' mode, and we
5587 : * must have at least as many backend slots as the primary.
5588 : */
5589 484 : if (ArchiveRecoveryRequested && EnableHotStandby)
5590 : {
5591 : /* We ignore autovacuum_worker_slots when we make this test. */
5592 244 : RecoveryRequiresIntParameter("max_connections",
5593 : MaxConnections,
5594 244 : ControlFile->MaxConnections);
5595 244 : RecoveryRequiresIntParameter("max_worker_processes",
5596 : max_worker_processes,
5597 244 : ControlFile->max_worker_processes);
5598 244 : RecoveryRequiresIntParameter("max_wal_senders",
5599 : max_wal_senders,
5600 244 : ControlFile->max_wal_senders);
5601 244 : RecoveryRequiresIntParameter("max_prepared_transactions",
5602 : max_prepared_xacts,
5603 244 : ControlFile->max_prepared_xacts);
5604 244 : RecoveryRequiresIntParameter("max_locks_per_transaction",
5605 : max_locks_per_xact,
5606 244 : ControlFile->max_locks_per_xact);
5607 : }
5608 484 : }
5609 :
5610 : /*
5611 : * This must be called ONCE during postmaster or standalone-backend startup
5612 : */
5613 : void
5614 1846 : StartupXLOG(void)
5615 : {
5616 : XLogCtlInsert *Insert;
5617 : CheckPoint checkPoint;
5618 : bool wasShutdown;
5619 : bool didCrash;
5620 : bool haveTblspcMap;
5621 : bool haveBackupLabel;
5622 : XLogRecPtr EndOfLog;
5623 : TimeLineID EndOfLogTLI;
5624 : TimeLineID newTLI;
5625 : bool performedWalRecovery;
5626 : EndOfWalRecoveryInfo *endOfRecoveryInfo;
5627 : XLogRecPtr abortedRecPtr;
5628 : XLogRecPtr missingContrecPtr;
5629 : TransactionId oldestActiveXID;
5630 1846 : bool promoted = false;
5631 :
5632 : /*
5633 : * We should have an aux process resource owner to use, and we should not
5634 : * be in a transaction that's installed some other resowner.
5635 : */
5636 : Assert(AuxProcessResourceOwner != NULL);
5637 : Assert(CurrentResourceOwner == NULL ||
5638 : CurrentResourceOwner == AuxProcessResourceOwner);
5639 1846 : CurrentResourceOwner = AuxProcessResourceOwner;
5640 :
5641 : /*
5642 : * Check that contents look valid.
5643 : */
5644 1846 : if (!XRecOffIsValid(ControlFile->checkPoint))
5645 0 : ereport(FATAL,
5646 : (errcode(ERRCODE_DATA_CORRUPTED),
5647 : errmsg("control file contains invalid checkpoint location")));
5648 :
5649 1846 : switch (ControlFile->state)
5650 : {
5651 1434 : case DB_SHUTDOWNED:
5652 :
5653 : /*
5654 : * This is the expected case, so don't be chatty in standalone
5655 : * mode
5656 : */
5657 1434 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
5658 : (errmsg("database system was shut down at %s",
5659 : str_time(ControlFile->time))));
5660 1434 : break;
5661 :
5662 64 : case DB_SHUTDOWNED_IN_RECOVERY:
5663 64 : ereport(LOG,
5664 : (errmsg("database system was shut down in recovery at %s",
5665 : str_time(ControlFile->time))));
5666 64 : break;
5667 :
5668 0 : case DB_SHUTDOWNING:
5669 0 : ereport(LOG,
5670 : (errmsg("database system shutdown was interrupted; last known up at %s",
5671 : str_time(ControlFile->time))));
5672 0 : break;
5673 :
5674 0 : case DB_IN_CRASH_RECOVERY:
5675 0 : ereport(LOG,
5676 : (errmsg("database system was interrupted while in recovery at %s",
5677 : str_time(ControlFile->time)),
5678 : errhint("This probably means that some data is corrupted and"
5679 : " you will have to use the last backup for recovery.")));
5680 0 : break;
5681 :
5682 12 : case DB_IN_ARCHIVE_RECOVERY:
5683 12 : ereport(LOG,
5684 : (errmsg("database system was interrupted while in recovery at log time %s",
5685 : str_time(ControlFile->checkPointCopy.time)),
5686 : errhint("If this has occurred more than once some data might be corrupted"
5687 : " and you might need to choose an earlier recovery target.")));
5688 12 : break;
5689 :
5690 336 : case DB_IN_PRODUCTION:
5691 336 : ereport(LOG,
5692 : (errmsg("database system was interrupted; last known up at %s",
5693 : str_time(ControlFile->time))));
5694 336 : break;
5695 :
5696 0 : default:
5697 0 : ereport(FATAL,
5698 : (errcode(ERRCODE_DATA_CORRUPTED),
5699 : errmsg("control file contains invalid database cluster state")));
5700 : }
5701 :
5702 : /* This is just to allow attaching to startup process with a debugger */
5703 : #ifdef XLOG_REPLAY_DELAY
5704 : if (ControlFile->state != DB_SHUTDOWNED)
5705 : pg_usleep(60000000L);
5706 : #endif
5707 :
5708 : /*
5709 : * Verify that pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
5710 : * In cases where someone has performed a copy for PITR, these directories
5711 : * may have been excluded and need to be re-created.
5712 : */
5713 1846 : ValidateXLOGDirectoryStructure();
5714 :
5715 : /* Set up timeout handler needed to report startup progress. */
5716 1846 : if (!IsBootstrapProcessingMode())
5717 1744 : RegisterTimeout(STARTUP_PROGRESS_TIMEOUT,
5718 : startup_progress_timeout_handler);
5719 :
5720 : /*----------
5721 : * If we previously crashed, perform a couple of actions:
5722 : *
5723 : * - The pg_wal directory may still include some temporary WAL segments
5724 : * used when creating a new segment, so perform some clean up to not
5725 : * bloat this path. This is done first as there is no point to sync
5726 : * this temporary data.
5727 : *
5728 : * - There might be data which we had written, intending to fsync it, but
5729 : * which we had not actually fsync'd yet. Therefore, a power failure in
5730 : * the near future might cause earlier unflushed writes to be lost, even
5731 : * though more recent data written to disk from here on would be
5732 : * persisted. To avoid that, fsync the entire data directory.
5733 : */
5734 1846 : if (ControlFile->state != DB_SHUTDOWNED &&
5735 412 : ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
5736 : {
5737 348 : RemoveTempXlogFiles();
5738 348 : SyncDataDirectory();
5739 348 : didCrash = true;
5740 : }
5741 : else
5742 1498 : didCrash = false;
5743 :
5744 : /*
5745 : * Prepare for WAL recovery if needed.
5746 : *
5747 : * InitWalRecovery analyzes the control file and the backup label file, if
5748 : * any. It updates the in-memory ControlFile buffer according to the
5749 : * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5750 : * It also applies the tablespace map file, if any.
5751 : */
5752 1846 : InitWalRecovery(ControlFile, &wasShutdown,
5753 : &haveBackupLabel, &haveTblspcMap);
5754 1846 : checkPoint = ControlFile->checkPointCopy;
5755 :
5756 : /* initialize shared memory variables from the checkpoint record */
5757 1846 : TransamVariables->nextXid = checkPoint.nextXid;
5758 1846 : TransamVariables->nextOid = checkPoint.nextOid;
5759 1846 : TransamVariables->oidCount = 0;
5760 1846 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5761 1846 : AdvanceOldestClogXid(checkPoint.oldestXid);
5762 1846 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5763 1846 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5764 1846 : SetCommitTsLimit(checkPoint.oldestCommitTsXid,
5765 : checkPoint.newestCommitTsXid);
5766 1846 : XLogCtl->ckptFullXid = checkPoint.nextXid;
5767 :
5768 : /*
5769 : * Clear out any old relcache cache files. This is *necessary* if we do
5770 : * any WAL replay, since that would probably result in the cache files
5771 : * being out of sync with database reality. In theory we could leave them
5772 : * in place if the database had been cleanly shut down, but it seems
5773 : * safest to just remove them always and let them be rebuilt during the
5774 : * first backend startup. These files needs to be removed from all
5775 : * directories including pg_tblspc, however the symlinks are created only
5776 : * after reading tablespace_map file in case of archive recovery from
5777 : * backup, so needs to clear old relcache files here after creating
5778 : * symlinks.
5779 : */
5780 1846 : RelationCacheInitFileRemove();
5781 :
5782 : /*
5783 : * Initialize replication slots, before there's a chance to remove
5784 : * required resources.
5785 : */
5786 1846 : StartupReplicationSlots();
5787 :
5788 : /*
5789 : * Startup logical state, needs to be setup now so we have proper data
5790 : * during crash recovery.
5791 : */
5792 1844 : StartupReorderBuffer();
5793 :
5794 : /*
5795 : * Startup CLOG. This must be done after TransamVariables->nextXid has
5796 : * been initialized and before we accept connections or begin WAL replay.
5797 : */
5798 1844 : StartupCLOG();
5799 :
5800 : /*
5801 : * Startup MultiXact. We need to do this early to be able to replay
5802 : * truncations.
5803 : */
5804 1844 : StartupMultiXact();
5805 :
5806 : /*
5807 : * Ditto for commit timestamps. Activate the facility if the setting is
5808 : * enabled in the control file, as there should be no tracking of commit
5809 : * timestamps done when the setting was disabled. This facility can be
5810 : * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
5811 : */
5812 1844 : if (ControlFile->track_commit_timestamp)
5813 26 : StartupCommitTs();
5814 :
5815 : /*
5816 : * Recover knowledge about replay progress of known replication partners.
5817 : */
5818 1844 : StartupReplicationOrigin();
5819 :
5820 : /*
5821 : * Initialize unlogged LSN. On a clean shutdown, it's restored from the
5822 : * control file. On recovery, all unlogged relations are blown away, so
5823 : * the unlogged LSN counter can be reset too.
5824 : */
5825 1844 : if (ControlFile->state == DB_SHUTDOWNED)
5826 1420 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5827 1420 : ControlFile->unloggedLSN);
5828 : else
5829 424 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5830 : FirstNormalUnloggedLSN);
5831 :
5832 : /*
5833 : * Copy any missing timeline history files between 'now' and the recovery
5834 : * target timeline from archive to pg_wal. While we don't need those files
5835 : * ourselves - the history file of the recovery target timeline covers all
5836 : * the previous timelines in the history too - a cascading standby server
5837 : * might be interested in them. Or, if you archive the WAL from this
5838 : * server to a different archive than the primary, it'd be good for all
5839 : * the history files to get archived there after failover, so that you can
5840 : * use one of the old timelines as a PITR target. Timeline history files
5841 : * are small, so it's better to copy them unnecessarily than not copy them
5842 : * and regret later.
5843 : */
5844 1844 : restoreTimeLineHistoryFiles(checkPoint.ThisTimeLineID, recoveryTargetTLI);
5845 :
5846 : /*
5847 : * Before running in recovery, scan pg_twophase and fill in its status to
5848 : * be able to work on entries generated by redo. Doing a scan before
5849 : * taking any recovery action has the merit to discard any 2PC files that
5850 : * are newer than the first record to replay, saving from any conflicts at
5851 : * replay. This avoids as well any subsequent scans when doing recovery
5852 : * of the on-disk two-phase data.
5853 : */
5854 1844 : restoreTwoPhaseData();
5855 :
5856 : /*
5857 : * When starting with crash recovery, reset pgstat data - it might not be
5858 : * valid. Otherwise restore pgstat data. It's safe to do this here,
5859 : * because postmaster will not yet have started any other processes.
5860 : *
5861 : * NB: Restoring replication slot stats relies on slot state to have
5862 : * already been restored from disk.
5863 : *
5864 : * TODO: With a bit of extra work we could just start with a pgstat file
5865 : * associated with the checkpoint redo location we're starting from.
5866 : */
5867 1844 : if (didCrash)
5868 348 : pgstat_discard_stats();
5869 : else
5870 1496 : pgstat_restore_stats();
5871 :
5872 1844 : lastFullPageWrites = checkPoint.fullPageWrites;
5873 :
5874 1844 : RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
5875 1844 : doPageWrites = lastFullPageWrites;
5876 :
5877 : /* REDO */
5878 1844 : if (InRecovery)
5879 : {
5880 : /* Initialize state for RecoveryInProgress() */
5881 424 : SpinLockAcquire(&XLogCtl->info_lck);
5882 424 : if (InArchiveRecovery)
5883 222 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
5884 : else
5885 202 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5886 424 : SpinLockRelease(&XLogCtl->info_lck);
5887 :
5888 : /*
5889 : * Update pg_control to show that we are recovering and to show the
5890 : * selected checkpoint as the place we are starting from. We also mark
5891 : * pg_control with any minimum recovery stop point obtained from a
5892 : * backup history file.
5893 : *
5894 : * No need to hold ControlFileLock yet, we aren't up far enough.
5895 : */
5896 424 : UpdateControlFile();
5897 :
5898 : /*
5899 : * If there was a backup label file, it's done its job and the info
5900 : * has now been propagated into pg_control. We must get rid of the
5901 : * label file so that if we crash during recovery, we'll pick up at
5902 : * the latest recovery restartpoint instead of going all the way back
5903 : * to the backup start point. It seems prudent though to just rename
5904 : * the file out of the way rather than delete it completely.
5905 : */
5906 424 : if (haveBackupLabel)
5907 : {
5908 142 : unlink(BACKUP_LABEL_OLD);
5909 142 : durable_rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD, FATAL);
5910 : }
5911 :
5912 : /*
5913 : * If there was a tablespace_map file, it's done its job and the
5914 : * symlinks have been created. We must get rid of the map file so
5915 : * that if we crash during recovery, we don't create symlinks again.
5916 : * It seems prudent though to just rename the file out of the way
5917 : * rather than delete it completely.
5918 : */
5919 424 : if (haveTblspcMap)
5920 : {
5921 4 : unlink(TABLESPACE_MAP_OLD);
5922 4 : durable_rename(TABLESPACE_MAP, TABLESPACE_MAP_OLD, FATAL);
5923 : }
5924 :
5925 : /*
5926 : * Initialize our local copy of minRecoveryPoint. When doing crash
5927 : * recovery we want to replay up to the end of WAL. Particularly, in
5928 : * the case of a promoted standby minRecoveryPoint value in the
5929 : * control file is only updated after the first checkpoint. However,
5930 : * if the instance crashes before the first post-recovery checkpoint
5931 : * is completed then recovery will use a stale location causing the
5932 : * startup process to think that there are still invalid page
5933 : * references when checking for data consistency.
5934 : */
5935 424 : if (InArchiveRecovery)
5936 : {
5937 222 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
5938 222 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
5939 : }
5940 : else
5941 : {
5942 202 : LocalMinRecoveryPoint = InvalidXLogRecPtr;
5943 202 : LocalMinRecoveryPointTLI = 0;
5944 : }
5945 :
5946 : /* Check that the GUCs used to generate the WAL allow recovery */
5947 424 : CheckRequiredParameterValues();
5948 :
5949 : /*
5950 : * We're in recovery, so unlogged relations may be trashed and must be
5951 : * reset. This should be done BEFORE allowing Hot Standby
5952 : * connections, so that read-only backends don't try to read whatever
5953 : * garbage is left over from before.
5954 : */
5955 424 : ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
5956 :
5957 : /*
5958 : * Likewise, delete any saved transaction snapshot files that got left
5959 : * behind by crashed backends.
5960 : */
5961 424 : DeleteAllExportedSnapshotFiles();
5962 :
5963 : /*
5964 : * Initialize for Hot Standby, if enabled. We won't let backends in
5965 : * yet, not until we've reached the min recovery point specified in
5966 : * control file and we've established a recovery snapshot from a
5967 : * running-xacts WAL record.
5968 : */
5969 424 : if (ArchiveRecoveryRequested && EnableHotStandby)
5970 : {
5971 : TransactionId *xids;
5972 : int nxids;
5973 :
5974 210 : ereport(DEBUG1,
5975 : (errmsg_internal("initializing for hot standby")));
5976 :
5977 210 : InitRecoveryTransactionEnvironment();
5978 :
5979 210 : if (wasShutdown)
5980 52 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
5981 : else
5982 158 : oldestActiveXID = checkPoint.oldestActiveXid;
5983 : Assert(TransactionIdIsValid(oldestActiveXID));
5984 :
5985 : /* Tell procarray about the range of xids it has to deal with */
5986 210 : ProcArrayInitRecovery(XidFromFullTransactionId(TransamVariables->nextXid));
5987 :
5988 : /*
5989 : * Startup subtrans only. CLOG, MultiXact and commit timestamp
5990 : * have already been started up and other SLRUs are not maintained
5991 : * during recovery and need not be started yet.
5992 : */
5993 210 : StartupSUBTRANS(oldestActiveXID);
5994 :
5995 : /*
5996 : * If we're beginning at a shutdown checkpoint, we know that
5997 : * nothing was running on the primary at this point. So fake-up an
5998 : * empty running-xacts record and use that here and now. Recover
5999 : * additional standby state for prepared transactions.
6000 : */
6001 210 : if (wasShutdown)
6002 : {
6003 : RunningTransactionsData running;
6004 : TransactionId latestCompletedXid;
6005 :
6006 : /* Update pg_subtrans entries for any prepared transactions */
6007 52 : StandbyRecoverPreparedTransactions();
6008 :
6009 : /*
6010 : * Construct a RunningTransactions snapshot representing a
6011 : * shut down server, with only prepared transactions still
6012 : * alive. We're never overflowed at this point because all
6013 : * subxids are listed with their parent prepared transactions.
6014 : */
6015 52 : running.xcnt = nxids;
6016 52 : running.subxcnt = 0;
6017 52 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
6018 52 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
6019 52 : running.oldestRunningXid = oldestActiveXID;
6020 52 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
6021 52 : TransactionIdRetreat(latestCompletedXid);
6022 : Assert(TransactionIdIsNormal(latestCompletedXid));
6023 52 : running.latestCompletedXid = latestCompletedXid;
6024 52 : running.xids = xids;
6025 :
6026 52 : ProcArrayApplyRecoveryInfo(&running);
6027 : }
6028 : }
6029 :
6030 : /*
6031 : * We're all set for replaying the WAL now. Do it.
6032 : */
6033 424 : PerformWalRecovery();
6034 308 : performedWalRecovery = true;
6035 : }
6036 : else
6037 1420 : performedWalRecovery = false;
6038 :
6039 : /*
6040 : * Finish WAL recovery.
6041 : */
6042 1728 : endOfRecoveryInfo = FinishWalRecovery();
6043 1728 : EndOfLog = endOfRecoveryInfo->endOfLog;
6044 1728 : EndOfLogTLI = endOfRecoveryInfo->endOfLogTLI;
6045 1728 : abortedRecPtr = endOfRecoveryInfo->abortedRecPtr;
6046 1728 : missingContrecPtr = endOfRecoveryInfo->missingContrecPtr;
6047 :
6048 : /*
6049 : * Reset ps status display, so as no information related to recovery shows
6050 : * up.
6051 : */
6052 1728 : set_ps_display("");
6053 :
6054 : /*
6055 : * When recovering from a backup (we are in recovery, and archive recovery
6056 : * was requested), complain if we did not roll forward far enough to reach
6057 : * the point where the database is consistent. For regular online
6058 : * backup-from-primary, that means reaching the end-of-backup WAL record
6059 : * (at which point we reset backupStartPoint to be Invalid), for
6060 : * backup-from-replica (which can't inject records into the WAL stream),
6061 : * that point is when we reach the minRecoveryPoint in pg_control (which
6062 : * we purposefully copy last when backing up from a replica). For
6063 : * pg_rewind (which creates a backup_label with a method of "pg_rewind")
6064 : * or snapshot-style backups (which don't), backupEndRequired will be set
6065 : * to false.
6066 : *
6067 : * Note: it is indeed okay to look at the local variable
6068 : * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
6069 : * might be further ahead --- ControlFile->minRecoveryPoint cannot have
6070 : * been advanced beyond the WAL we processed.
6071 : */
6072 1728 : if (InRecovery &&
6073 308 : (EndOfLog < LocalMinRecoveryPoint ||
6074 308 : !XLogRecPtrIsInvalid(ControlFile->backupStartPoint)))
6075 : {
6076 : /*
6077 : * Ran off end of WAL before reaching end-of-backup WAL record, or
6078 : * minRecoveryPoint. That's a bad sign, indicating that you tried to
6079 : * recover from an online backup but never called pg_backup_stop(), or
6080 : * you didn't archive all the WAL needed.
6081 : */
6082 0 : if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
6083 : {
6084 0 : if (!XLogRecPtrIsInvalid(ControlFile->backupStartPoint) || ControlFile->backupEndRequired)
6085 0 : ereport(FATAL,
6086 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6087 : errmsg("WAL ends before end of online backup"),
6088 : errhint("All WAL generated while online backup was taken must be available at recovery.")));
6089 : else
6090 0 : ereport(FATAL,
6091 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6092 : errmsg("WAL ends before consistent recovery point")));
6093 : }
6094 : }
6095 :
6096 : /*
6097 : * Reset unlogged relations to the contents of their INIT fork. This is
6098 : * done AFTER recovery is complete so as to include any unlogged relations
6099 : * created during recovery, but BEFORE recovery is marked as having
6100 : * completed successfully. Otherwise we'd not retry if any of the post
6101 : * end-of-recovery steps fail.
6102 : */
6103 1728 : if (InRecovery)
6104 308 : ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
6105 :
6106 : /*
6107 : * Pre-scan prepared transactions to find out the range of XIDs present.
6108 : * This information is not quite needed yet, but it is positioned here so
6109 : * as potential problems are detected before any on-disk change is done.
6110 : */
6111 1728 : oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
6112 :
6113 : /*
6114 : * Allow ordinary WAL segment creation before possibly switching to a new
6115 : * timeline, which creates a new segment, and after the last ReadRecord().
6116 : */
6117 1728 : SetInstallXLogFileSegmentActive();
6118 :
6119 : /*
6120 : * Consider whether we need to assign a new timeline ID.
6121 : *
6122 : * If we did archive recovery, we always assign a new ID. This handles a
6123 : * couple of issues. If we stopped short of the end of WAL during
6124 : * recovery, then we are clearly generating a new timeline and must assign
6125 : * it a unique new ID. Even if we ran to the end, modifying the current
6126 : * last segment is problematic because it may result in trying to
6127 : * overwrite an already-archived copy of that segment, and we encourage
6128 : * DBAs to make their archive_commands reject that. We can dodge the
6129 : * problem by making the new active segment have a new timeline ID.
6130 : *
6131 : * In a normal crash recovery, we can just extend the timeline we were in.
6132 : */
6133 1728 : newTLI = endOfRecoveryInfo->lastRecTLI;
6134 1728 : if (ArchiveRecoveryRequested)
6135 : {
6136 98 : newTLI = findNewestTimeLine(recoveryTargetTLI) + 1;
6137 98 : ereport(LOG,
6138 : (errmsg("selected new timeline ID: %u", newTLI)));
6139 :
6140 : /*
6141 : * Make a writable copy of the last WAL segment. (Note that we also
6142 : * have a copy of the last block of the old WAL in
6143 : * endOfRecovery->lastPage; we will use that below.)
6144 : */
6145 98 : XLogInitNewTimeline(EndOfLogTLI, EndOfLog, newTLI);
6146 :
6147 : /*
6148 : * Remove the signal files out of the way, so that we don't
6149 : * accidentally re-enter archive recovery mode in a subsequent crash.
6150 : */
6151 98 : if (endOfRecoveryInfo->standby_signal_file_found)
6152 92 : durable_unlink(STANDBY_SIGNAL_FILE, FATAL);
6153 :
6154 98 : if (endOfRecoveryInfo->recovery_signal_file_found)
6155 6 : durable_unlink(RECOVERY_SIGNAL_FILE, FATAL);
6156 :
6157 : /*
6158 : * Write the timeline history file, and have it archived. After this
6159 : * point (or rather, as soon as the file is archived), the timeline
6160 : * will appear as "taken" in the WAL archive and to any standby
6161 : * servers. If we crash before actually switching to the new
6162 : * timeline, standby servers will nevertheless think that we switched
6163 : * to the new timeline, and will try to connect to the new timeline.
6164 : * To minimize the window for that, try to do as little as possible
6165 : * between here and writing the end-of-recovery record.
6166 : */
6167 98 : writeTimeLineHistory(newTLI, recoveryTargetTLI,
6168 : EndOfLog, endOfRecoveryInfo->recoveryStopReason);
6169 :
6170 98 : ereport(LOG,
6171 : (errmsg("archive recovery complete")));
6172 : }
6173 :
6174 : /* Save the selected TimeLineID in shared memory, too */
6175 1728 : SpinLockAcquire(&XLogCtl->info_lck);
6176 1728 : XLogCtl->InsertTimeLineID = newTLI;
6177 1728 : XLogCtl->PrevTimeLineID = endOfRecoveryInfo->lastRecTLI;
6178 1728 : SpinLockRelease(&XLogCtl->info_lck);
6179 :
6180 : /*
6181 : * Actually, if WAL ended in an incomplete record, skip the parts that
6182 : * made it through and start writing after the portion that persisted.
6183 : * (It's critical to first write an OVERWRITE_CONTRECORD message, which
6184 : * we'll do as soon as we're open for writing new WAL.)
6185 : */
6186 1728 : if (!XLogRecPtrIsInvalid(missingContrecPtr))
6187 : {
6188 : /*
6189 : * We should only have a missingContrecPtr if we're not switching to a
6190 : * new timeline. When a timeline switch occurs, WAL is copied from the
6191 : * old timeline to the new only up to the end of the last complete
6192 : * record, so there can't be an incomplete WAL record that we need to
6193 : * disregard.
6194 : */
6195 : Assert(newTLI == endOfRecoveryInfo->lastRecTLI);
6196 : Assert(!XLogRecPtrIsInvalid(abortedRecPtr));
6197 20 : EndOfLog = missingContrecPtr;
6198 : }
6199 :
6200 : /*
6201 : * Prepare to write WAL starting at EndOfLog location, and init xlog
6202 : * buffer cache using the block containing the last record from the
6203 : * previous incarnation.
6204 : */
6205 1728 : Insert = &XLogCtl->Insert;
6206 1728 : Insert->PrevBytePos = XLogRecPtrToBytePos(endOfRecoveryInfo->lastRec);
6207 1728 : Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
6208 :
6209 : /*
6210 : * Tricky point here: lastPage contains the *last* block that the LastRec
6211 : * record spans, not the one it starts in. The last block is indeed the
6212 : * one we want to use.
6213 : */
6214 1728 : if (EndOfLog % XLOG_BLCKSZ != 0)
6215 : {
6216 : char *page;
6217 : int len;
6218 : int firstIdx;
6219 :
6220 1670 : firstIdx = XLogRecPtrToBufIdx(EndOfLog);
6221 1670 : len = EndOfLog - endOfRecoveryInfo->lastPageBeginPtr;
6222 : Assert(len < XLOG_BLCKSZ);
6223 :
6224 : /* Copy the valid part of the last block, and zero the rest */
6225 1670 : page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
6226 1670 : memcpy(page, endOfRecoveryInfo->lastPage, len);
6227 1670 : memset(page + len, 0, XLOG_BLCKSZ - len);
6228 :
6229 1670 : pg_atomic_write_u64(&XLogCtl->xlblocks[firstIdx], endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6230 1670 : pg_atomic_write_u64(&XLogCtl->InitializedUpTo, endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6231 1670 : XLogCtl->InitializedFrom = endOfRecoveryInfo->lastPageBeginPtr;
6232 : }
6233 : else
6234 : {
6235 : /*
6236 : * There is no partial block to copy. Just set InitializedUpTo, and
6237 : * let the first attempt to insert a log record to initialize the next
6238 : * buffer.
6239 : */
6240 58 : pg_atomic_write_u64(&XLogCtl->InitializedUpTo, EndOfLog);
6241 58 : XLogCtl->InitializedFrom = EndOfLog;
6242 : }
6243 1728 : pg_atomic_write_u64(&XLogCtl->InitializeReserved, pg_atomic_read_u64(&XLogCtl->InitializedUpTo));
6244 :
6245 : /*
6246 : * Update local and shared status. This is OK to do without any locks
6247 : * because no other process can be reading or writing WAL yet.
6248 : */
6249 1728 : LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
6250 1728 : pg_atomic_write_u64(&XLogCtl->logInsertResult, EndOfLog);
6251 1728 : pg_atomic_write_u64(&XLogCtl->logWriteResult, EndOfLog);
6252 1728 : pg_atomic_write_u64(&XLogCtl->logFlushResult, EndOfLog);
6253 1728 : XLogCtl->LogwrtRqst.Write = EndOfLog;
6254 1728 : XLogCtl->LogwrtRqst.Flush = EndOfLog;
6255 :
6256 : /*
6257 : * Preallocate additional log files, if wanted.
6258 : */
6259 1728 : PreallocXlogFiles(EndOfLog, newTLI);
6260 :
6261 : /*
6262 : * Okay, we're officially UP.
6263 : */
6264 1728 : InRecovery = false;
6265 :
6266 : /* start the archive_timeout timer and LSN running */
6267 1728 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
6268 1728 : XLogCtl->lastSegSwitchLSN = EndOfLog;
6269 :
6270 : /* also initialize latestCompletedXid, to nextXid - 1 */
6271 1728 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
6272 1728 : TransamVariables->latestCompletedXid = TransamVariables->nextXid;
6273 1728 : FullTransactionIdRetreat(&TransamVariables->latestCompletedXid);
6274 1728 : LWLockRelease(ProcArrayLock);
6275 :
6276 : /*
6277 : * Start up subtrans, if not already done for hot standby. (commit
6278 : * timestamps are started below, if necessary.)
6279 : */
6280 1728 : if (standbyState == STANDBY_DISABLED)
6281 1630 : StartupSUBTRANS(oldestActiveXID);
6282 :
6283 : /*
6284 : * Perform end of recovery actions for any SLRUs that need it.
6285 : */
6286 1728 : TrimCLOG();
6287 1728 : TrimMultiXact();
6288 :
6289 : /*
6290 : * Reload shared-memory state for prepared transactions. This needs to
6291 : * happen before renaming the last partial segment of the old timeline as
6292 : * it may be possible that we have to recover some transactions from it.
6293 : */
6294 1728 : RecoverPreparedTransactions();
6295 :
6296 : /* Shut down xlogreader */
6297 1728 : ShutdownWalRecovery();
6298 :
6299 : /* Enable WAL writes for this backend only. */
6300 1728 : LocalSetXLogInsertAllowed();
6301 :
6302 : /* If necessary, write overwrite-contrecord before doing anything else */
6303 1728 : if (!XLogRecPtrIsInvalid(abortedRecPtr))
6304 : {
6305 : Assert(!XLogRecPtrIsInvalid(missingContrecPtr));
6306 20 : CreateOverwriteContrecordRecord(abortedRecPtr, missingContrecPtr, newTLI);
6307 : }
6308 :
6309 : /*
6310 : * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
6311 : * record before resource manager writes cleanup WAL records or checkpoint
6312 : * record is written.
6313 : */
6314 1728 : Insert->fullPageWrites = lastFullPageWrites;
6315 1728 : UpdateFullPageWrites();
6316 :
6317 : /*
6318 : * Emit checkpoint or end-of-recovery record in XLOG, if required.
6319 : */
6320 1728 : if (performedWalRecovery)
6321 308 : promoted = PerformRecoveryXLogAction();
6322 :
6323 : /*
6324 : * If any of the critical GUCs have changed, log them before we allow
6325 : * backends to write WAL.
6326 : */
6327 1728 : XLogReportParameters();
6328 :
6329 : /* If this is archive recovery, perform post-recovery cleanup actions. */
6330 1728 : if (ArchiveRecoveryRequested)
6331 98 : CleanupAfterArchiveRecovery(EndOfLogTLI, EndOfLog, newTLI);
6332 :
6333 : /*
6334 : * Local WAL inserts enabled, so it's time to finish initialization of
6335 : * commit timestamp.
6336 : */
6337 1728 : CompleteCommitTsInitialization();
6338 :
6339 : /*
6340 : * All done with end-of-recovery actions.
6341 : *
6342 : * Now allow backends to write WAL and update the control file status in
6343 : * consequence. SharedRecoveryState, that controls if backends can write
6344 : * WAL, is updated while holding ControlFileLock to prevent other backends
6345 : * to look at an inconsistent state of the control file in shared memory.
6346 : * There is still a small window during which backends can write WAL and
6347 : * the control file is still referring to a system not in DB_IN_PRODUCTION
6348 : * state while looking at the on-disk control file.
6349 : *
6350 : * Also, we use info_lck to update SharedRecoveryState to ensure that
6351 : * there are no race conditions concerning visibility of other recent
6352 : * updates to shared memory.
6353 : */
6354 1728 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6355 1728 : ControlFile->state = DB_IN_PRODUCTION;
6356 :
6357 1728 : SpinLockAcquire(&XLogCtl->info_lck);
6358 1728 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_DONE;
6359 1728 : SpinLockRelease(&XLogCtl->info_lck);
6360 :
6361 1728 : UpdateControlFile();
6362 1728 : LWLockRelease(ControlFileLock);
6363 :
6364 : /*
6365 : * Shutdown the recovery environment. This must occur after
6366 : * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
6367 : * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
6368 : * any session building a snapshot will not rely on KnownAssignedXids as
6369 : * RecoveryInProgress() would return false at this stage. This is
6370 : * particularly critical for prepared 2PC transactions, that would still
6371 : * need to be included in snapshots once recovery has ended.
6372 : */
6373 1728 : if (standbyState != STANDBY_DISABLED)
6374 98 : ShutdownRecoveryTransactionEnvironment();
6375 :
6376 : /*
6377 : * If there were cascading standby servers connected to us, nudge any wal
6378 : * sender processes to notice that we've been promoted.
6379 : */
6380 1728 : WalSndWakeup(true, true);
6381 :
6382 : /*
6383 : * If this was a promotion, request an (online) checkpoint now. This isn't
6384 : * required for consistency, but the last restartpoint might be far back,
6385 : * and in case of a crash, recovering from it might take a longer than is
6386 : * appropriate now that we're not in standby mode anymore.
6387 : */
6388 1728 : if (promoted)
6389 84 : RequestCheckpoint(CHECKPOINT_FORCE);
6390 1728 : }
6391 :
6392 : /*
6393 : * Callback from PerformWalRecovery(), called when we switch from crash
6394 : * recovery to archive recovery mode. Updates the control file accordingly.
6395 : */
6396 : void
6397 4 : SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr, TimeLineID replayTLI)
6398 : {
6399 : /* initialize minRecoveryPoint to this record */
6400 4 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6401 4 : ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6402 4 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6403 : {
6404 4 : ControlFile->minRecoveryPoint = EndRecPtr;
6405 4 : ControlFile->minRecoveryPointTLI = replayTLI;
6406 : }
6407 : /* update local copy */
6408 4 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6409 4 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6410 :
6411 : /*
6412 : * The startup process can update its local copy of minRecoveryPoint from
6413 : * this point.
6414 : */
6415 4 : updateMinRecoveryPoint = true;
6416 :
6417 4 : UpdateControlFile();
6418 :
6419 : /*
6420 : * We update SharedRecoveryState while holding the lock on ControlFileLock
6421 : * so both states are consistent in shared memory.
6422 : */
6423 4 : SpinLockAcquire(&XLogCtl->info_lck);
6424 4 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6425 4 : SpinLockRelease(&XLogCtl->info_lck);
6426 :
6427 4 : LWLockRelease(ControlFileLock);
6428 4 : }
6429 :
6430 : /*
6431 : * Callback from PerformWalRecovery(), called when we reach the end of backup.
6432 : * Updates the control file accordingly.
6433 : */
6434 : void
6435 142 : ReachedEndOfBackup(XLogRecPtr EndRecPtr, TimeLineID tli)
6436 : {
6437 : /*
6438 : * We have reached the end of base backup, as indicated by pg_control. The
6439 : * data on disk is now consistent (unless minRecoveryPoint is further
6440 : * ahead, which can happen if we crashed during previous recovery). Reset
6441 : * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
6442 : * make sure we don't allow starting up at an earlier point even if
6443 : * recovery is stopped and restarted soon after this.
6444 : */
6445 142 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6446 :
6447 142 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6448 : {
6449 134 : ControlFile->minRecoveryPoint = EndRecPtr;
6450 134 : ControlFile->minRecoveryPointTLI = tli;
6451 : }
6452 :
6453 142 : ControlFile->backupStartPoint = InvalidXLogRecPtr;
6454 142 : ControlFile->backupEndPoint = InvalidXLogRecPtr;
6455 142 : ControlFile->backupEndRequired = false;
6456 142 : UpdateControlFile();
6457 :
6458 142 : LWLockRelease(ControlFileLock);
6459 142 : }
6460 :
6461 : /*
6462 : * Perform whatever XLOG actions are necessary at end of REDO.
6463 : *
6464 : * The goal here is to make sure that we'll be able to recover properly if
6465 : * we crash again. If we choose to write a checkpoint, we'll write a shutdown
6466 : * checkpoint rather than an on-line one. This is not particularly critical,
6467 : * but since we may be assigning a new TLI, using a shutdown checkpoint allows
6468 : * us to have the rule that TLI only changes in shutdown checkpoints, which
6469 : * allows some extra error checking in xlog_redo.
6470 : */
6471 : static bool
6472 308 : PerformRecoveryXLogAction(void)
6473 : {
6474 308 : bool promoted = false;
6475 :
6476 : /*
6477 : * Perform a checkpoint to update all our recovery activity to disk.
6478 : *
6479 : * Note that we write a shutdown checkpoint rather than an on-line one.
6480 : * This is not particularly critical, but since we may be assigning a new
6481 : * TLI, using a shutdown checkpoint allows us to have the rule that TLI
6482 : * only changes in shutdown checkpoints, which allows some extra error
6483 : * checking in xlog_redo.
6484 : *
6485 : * In promotion, only create a lightweight end-of-recovery record instead
6486 : * of a full checkpoint. A checkpoint is requested later, after we're
6487 : * fully out of recovery mode and already accepting queries.
6488 : */
6489 406 : if (ArchiveRecoveryRequested && IsUnderPostmaster &&
6490 98 : PromoteIsTriggered())
6491 : {
6492 84 : promoted = true;
6493 :
6494 : /*
6495 : * Insert a special WAL record to mark the end of recovery, since we
6496 : * aren't doing a checkpoint. That means that the checkpointer process
6497 : * may likely be in the middle of a time-smoothed restartpoint and
6498 : * could continue to be for minutes after this. That sounds strange,
6499 : * but the effect is roughly the same and it would be stranger to try
6500 : * to come out of the restartpoint and then checkpoint. We request a
6501 : * checkpoint later anyway, just for safety.
6502 : */
6503 84 : CreateEndOfRecoveryRecord();
6504 : }
6505 : else
6506 : {
6507 224 : RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
6508 : CHECKPOINT_IMMEDIATE |
6509 : CHECKPOINT_WAIT);
6510 : }
6511 :
6512 308 : return promoted;
6513 : }
6514 :
6515 : /*
6516 : * Is the system still in recovery?
6517 : *
6518 : * Unlike testing InRecovery, this works in any process that's connected to
6519 : * shared memory.
6520 : */
6521 : bool
6522 178715664 : RecoveryInProgress(void)
6523 : {
6524 : /*
6525 : * We check shared state each time only until we leave recovery mode. We
6526 : * can't re-enter recovery, so there's no need to keep checking after the
6527 : * shared variable has once been seen false.
6528 : */
6529 178715664 : if (!LocalRecoveryInProgress)
6530 174435938 : return false;
6531 : else
6532 : {
6533 : /*
6534 : * use volatile pointer to make sure we make a fresh read of the
6535 : * shared variable.
6536 : */
6537 4279726 : volatile XLogCtlData *xlogctl = XLogCtl;
6538 :
6539 4279726 : LocalRecoveryInProgress = (xlogctl->SharedRecoveryState != RECOVERY_STATE_DONE);
6540 :
6541 : /*
6542 : * Note: We don't need a memory barrier when we're still in recovery.
6543 : * We might exit recovery immediately after return, so the caller
6544 : * can't rely on 'true' meaning that we're still in recovery anyway.
6545 : */
6546 :
6547 4279726 : return LocalRecoveryInProgress;
6548 : }
6549 : }
6550 :
6551 : /*
6552 : * Returns current recovery state from shared memory.
6553 : *
6554 : * This returned state is kept consistent with the contents of the control
6555 : * file. See details about the possible values of RecoveryState in xlog.h.
6556 : */
6557 : RecoveryState
6558 38378 : GetRecoveryState(void)
6559 : {
6560 : RecoveryState retval;
6561 :
6562 38378 : SpinLockAcquire(&XLogCtl->info_lck);
6563 38378 : retval = XLogCtl->SharedRecoveryState;
6564 38378 : SpinLockRelease(&XLogCtl->info_lck);
6565 :
6566 38378 : return retval;
6567 : }
6568 :
6569 : /*
6570 : * Is this process allowed to insert new WAL records?
6571 : *
6572 : * Ordinarily this is essentially equivalent to !RecoveryInProgress().
6573 : * But we also have provisions for forcing the result "true" or "false"
6574 : * within specific processes regardless of the global state.
6575 : */
6576 : bool
6577 61114306 : XLogInsertAllowed(void)
6578 : {
6579 : /*
6580 : * If value is "unconditionally true" or "unconditionally false", just
6581 : * return it. This provides the normal fast path once recovery is known
6582 : * done.
6583 : */
6584 61114306 : if (LocalXLogInsertAllowed >= 0)
6585 60887118 : return (bool) LocalXLogInsertAllowed;
6586 :
6587 : /*
6588 : * Else, must check to see if we're still in recovery.
6589 : */
6590 227188 : if (RecoveryInProgress())
6591 211042 : return false;
6592 :
6593 : /*
6594 : * On exit from recovery, reset to "unconditionally true", since there is
6595 : * no need to keep checking.
6596 : */
6597 16146 : LocalXLogInsertAllowed = 1;
6598 16146 : return true;
6599 : }
6600 :
6601 : /*
6602 : * Make XLogInsertAllowed() return true in the current process only.
6603 : *
6604 : * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
6605 : * and even call LocalSetXLogInsertAllowed() again after that.
6606 : *
6607 : * Returns the previous value of LocalXLogInsertAllowed.
6608 : */
6609 : static int
6610 1786 : LocalSetXLogInsertAllowed(void)
6611 : {
6612 1786 : int oldXLogAllowed = LocalXLogInsertAllowed;
6613 :
6614 1786 : LocalXLogInsertAllowed = 1;
6615 :
6616 1786 : return oldXLogAllowed;
6617 : }
6618 :
6619 : /*
6620 : * Return the current Redo pointer from shared memory.
6621 : *
6622 : * As a side-effect, the local RedoRecPtr copy is updated.
6623 : */
6624 : XLogRecPtr
6625 594546 : GetRedoRecPtr(void)
6626 : {
6627 : XLogRecPtr ptr;
6628 :
6629 : /*
6630 : * The possibly not up-to-date copy in XlogCtl is enough. Even if we
6631 : * grabbed a WAL insertion lock to read the authoritative value in
6632 : * Insert->RedoRecPtr, someone might update it just after we've released
6633 : * the lock.
6634 : */
6635 594546 : SpinLockAcquire(&XLogCtl->info_lck);
6636 594546 : ptr = XLogCtl->RedoRecPtr;
6637 594546 : SpinLockRelease(&XLogCtl->info_lck);
6638 :
6639 594546 : if (RedoRecPtr < ptr)
6640 2834 : RedoRecPtr = ptr;
6641 :
6642 594546 : return RedoRecPtr;
6643 : }
6644 :
6645 : /*
6646 : * Return information needed to decide whether a modified block needs a
6647 : * full-page image to be included in the WAL record.
6648 : *
6649 : * The returned values are cached copies from backend-private memory, and
6650 : * possibly out-of-date or, indeed, uninitialized, in which case they will
6651 : * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6652 : * re-check them against up-to-date values, while holding the WAL insert lock.
6653 : */
6654 : void
6655 29528688 : GetFullPageWriteInfo(XLogRecPtr *RedoRecPtr_p, bool *doPageWrites_p)
6656 : {
6657 29528688 : *RedoRecPtr_p = RedoRecPtr;
6658 29528688 : *doPageWrites_p = doPageWrites;
6659 29528688 : }
6660 :
6661 : /*
6662 : * GetInsertRecPtr -- Returns the current insert position.
6663 : *
6664 : * NOTE: The value *actually* returned is the position of the last full
6665 : * xlog page. It lags behind the real insert position by at most 1 page.
6666 : * For that, we don't need to scan through WAL insertion locks, and an
6667 : * approximation is enough for the current usage of this function.
6668 : */
6669 : XLogRecPtr
6670 17586 : GetInsertRecPtr(void)
6671 : {
6672 : XLogRecPtr recptr;
6673 :
6674 17586 : SpinLockAcquire(&XLogCtl->info_lck);
6675 17586 : recptr = XLogCtl->LogwrtRqst.Write;
6676 17586 : SpinLockRelease(&XLogCtl->info_lck);
6677 :
6678 17586 : return recptr;
6679 : }
6680 :
6681 : /*
6682 : * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6683 : * position known to be fsync'd to disk. This should only be used on a
6684 : * system that is known not to be in recovery.
6685 : */
6686 : XLogRecPtr
6687 411804 : GetFlushRecPtr(TimeLineID *insertTLI)
6688 : {
6689 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6690 :
6691 411804 : RefreshXLogWriteResult(LogwrtResult);
6692 :
6693 : /*
6694 : * If we're writing and flushing WAL, the time line can't be changing, so
6695 : * no lock is required.
6696 : */
6697 411804 : if (insertTLI)
6698 43784 : *insertTLI = XLogCtl->InsertTimeLineID;
6699 :
6700 411804 : return LogwrtResult.Flush;
6701 : }
6702 :
6703 : /*
6704 : * GetWALInsertionTimeLine -- Returns the current timeline of a system that
6705 : * is not in recovery.
6706 : */
6707 : TimeLineID
6708 213782 : GetWALInsertionTimeLine(void)
6709 : {
6710 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6711 :
6712 : /* Since the value can't be changing, no lock is required. */
6713 213782 : return XLogCtl->InsertTimeLineID;
6714 : }
6715 :
6716 : /*
6717 : * GetWALInsertionTimeLineIfSet -- If the system is not in recovery, returns
6718 : * the WAL insertion timeline; else, returns 0. Wherever possible, use
6719 : * GetWALInsertionTimeLine() instead, since it's cheaper. Note that this
6720 : * function decides recovery has ended as soon as the insert TLI is set, which
6721 : * happens before we set XLogCtl->SharedRecoveryState to RECOVERY_STATE_DONE.
6722 : */
6723 : TimeLineID
6724 0 : GetWALInsertionTimeLineIfSet(void)
6725 : {
6726 : TimeLineID insertTLI;
6727 :
6728 0 : SpinLockAcquire(&XLogCtl->info_lck);
6729 0 : insertTLI = XLogCtl->InsertTimeLineID;
6730 0 : SpinLockRelease(&XLogCtl->info_lck);
6731 :
6732 0 : return insertTLI;
6733 : }
6734 :
6735 : /*
6736 : * GetLastImportantRecPtr -- Returns the LSN of the last important record
6737 : * inserted. All records not explicitly marked as unimportant are considered
6738 : * important.
6739 : *
6740 : * The LSN is determined by computing the maximum of
6741 : * WALInsertLocks[i].lastImportantAt.
6742 : */
6743 : XLogRecPtr
6744 3084 : GetLastImportantRecPtr(void)
6745 : {
6746 3084 : XLogRecPtr res = InvalidXLogRecPtr;
6747 : int i;
6748 :
6749 27756 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
6750 : {
6751 : XLogRecPtr last_important;
6752 :
6753 : /*
6754 : * Need to take a lock to prevent torn reads of the LSN, which are
6755 : * possible on some of the supported platforms. WAL insert locks only
6756 : * support exclusive mode, so we have to use that.
6757 : */
6758 24672 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
6759 24672 : last_important = WALInsertLocks[i].l.lastImportantAt;
6760 24672 : LWLockRelease(&WALInsertLocks[i].l.lock);
6761 :
6762 24672 : if (res < last_important)
6763 5302 : res = last_important;
6764 : }
6765 :
6766 3084 : return res;
6767 : }
6768 :
6769 : /*
6770 : * Get the time and LSN of the last xlog segment switch
6771 : */
6772 : pg_time_t
6773 0 : GetLastSegSwitchData(XLogRecPtr *lastSwitchLSN)
6774 : {
6775 : pg_time_t result;
6776 :
6777 : /* Need WALWriteLock, but shared lock is sufficient */
6778 0 : LWLockAcquire(WALWriteLock, LW_SHARED);
6779 0 : result = XLogCtl->lastSegSwitchTime;
6780 0 : *lastSwitchLSN = XLogCtl->lastSegSwitchLSN;
6781 0 : LWLockRelease(WALWriteLock);
6782 :
6783 0 : return result;
6784 : }
6785 :
6786 : /*
6787 : * This must be called ONCE during postmaster or standalone-backend shutdown
6788 : */
6789 : void
6790 1238 : ShutdownXLOG(int code, Datum arg)
6791 : {
6792 : /*
6793 : * We should have an aux process resource owner to use, and we should not
6794 : * be in a transaction that's installed some other resowner.
6795 : */
6796 : Assert(AuxProcessResourceOwner != NULL);
6797 : Assert(CurrentResourceOwner == NULL ||
6798 : CurrentResourceOwner == AuxProcessResourceOwner);
6799 1238 : CurrentResourceOwner = AuxProcessResourceOwner;
6800 :
6801 : /* Don't be chatty in standalone mode */
6802 1238 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
6803 : (errmsg("shutting down")));
6804 :
6805 : /*
6806 : * Signal walsenders to move to stopping state.
6807 : */
6808 1238 : WalSndInitStopping();
6809 :
6810 : /*
6811 : * Wait for WAL senders to be in stopping state. This prevents commands
6812 : * from writing new WAL.
6813 : */
6814 1238 : WalSndWaitStopping();
6815 :
6816 1238 : if (RecoveryInProgress())
6817 110 : CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_IMMEDIATE);
6818 : else
6819 : {
6820 : /*
6821 : * If archiving is enabled, rotate the last XLOG file so that all the
6822 : * remaining records are archived (postmaster wakes up the archiver
6823 : * process one more time at the end of shutdown). The checkpoint
6824 : * record will go to the next XLOG file and won't be archived (yet).
6825 : */
6826 1128 : if (XLogArchivingActive())
6827 28 : RequestXLogSwitch(false);
6828 :
6829 1128 : CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_IMMEDIATE);
6830 : }
6831 1238 : }
6832 :
6833 : /*
6834 : * Log start of a checkpoint.
6835 : */
6836 : static void
6837 2764 : LogCheckpointStart(int flags, bool restartpoint)
6838 : {
6839 2764 : if (restartpoint)
6840 368 : ereport(LOG,
6841 : /* translator: the placeholders show checkpoint options */
6842 : (errmsg("restartpoint starting:%s%s%s%s%s%s%s%s",
6843 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6844 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6845 : (flags & CHECKPOINT_IMMEDIATE) ? " immediate" : "",
6846 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6847 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6848 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6849 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6850 : (flags & CHECKPOINT_FLUSH_ALL) ? " flush-all" : "")));
6851 : else
6852 2396 : ereport(LOG,
6853 : /* translator: the placeholders show checkpoint options */
6854 : (errmsg("checkpoint starting:%s%s%s%s%s%s%s%s",
6855 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6856 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6857 : (flags & CHECKPOINT_IMMEDIATE) ? " immediate" : "",
6858 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6859 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6860 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6861 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6862 : (flags & CHECKPOINT_FLUSH_ALL) ? " flush-all" : "")));
6863 2764 : }
6864 :
6865 : /*
6866 : * Log end of a checkpoint.
6867 : */
6868 : static void
6869 3360 : LogCheckpointEnd(bool restartpoint)
6870 : {
6871 : long write_msecs,
6872 : sync_msecs,
6873 : total_msecs,
6874 : longest_msecs,
6875 : average_msecs;
6876 : uint64 average_sync_time;
6877 :
6878 3360 : CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
6879 :
6880 3360 : write_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_write_t,
6881 : CheckpointStats.ckpt_sync_t);
6882 :
6883 3360 : sync_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_sync_t,
6884 : CheckpointStats.ckpt_sync_end_t);
6885 :
6886 : /* Accumulate checkpoint timing summary data, in milliseconds. */
6887 3360 : PendingCheckpointerStats.write_time += write_msecs;
6888 3360 : PendingCheckpointerStats.sync_time += sync_msecs;
6889 :
6890 : /*
6891 : * All of the published timing statistics are accounted for. Only
6892 : * continue if a log message is to be written.
6893 : */
6894 3360 : if (!log_checkpoints)
6895 596 : return;
6896 :
6897 2764 : total_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_start_t,
6898 : CheckpointStats.ckpt_end_t);
6899 :
6900 : /*
6901 : * Timing values returned from CheckpointStats are in microseconds.
6902 : * Convert to milliseconds for consistent printing.
6903 : */
6904 2764 : longest_msecs = (long) ((CheckpointStats.ckpt_longest_sync + 999) / 1000);
6905 :
6906 2764 : average_sync_time = 0;
6907 2764 : if (CheckpointStats.ckpt_sync_rels > 0)
6908 0 : average_sync_time = CheckpointStats.ckpt_agg_sync_time /
6909 0 : CheckpointStats.ckpt_sync_rels;
6910 2764 : average_msecs = (long) ((average_sync_time + 999) / 1000);
6911 :
6912 : /*
6913 : * ControlFileLock is not required to see ControlFile->checkPoint and
6914 : * ->checkPointCopy here as we are the only updator of those variables at
6915 : * this moment.
6916 : */
6917 2764 : if (restartpoint)
6918 368 : ereport(LOG,
6919 : (errmsg("restartpoint complete: wrote %d buffers (%.1f%%), "
6920 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6921 : "%d removed, %d recycled; write=%ld.%03d s, "
6922 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6923 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6924 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6925 : CheckpointStats.ckpt_bufs_written,
6926 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6927 : CheckpointStats.ckpt_slru_written,
6928 : CheckpointStats.ckpt_segs_added,
6929 : CheckpointStats.ckpt_segs_removed,
6930 : CheckpointStats.ckpt_segs_recycled,
6931 : write_msecs / 1000, (int) (write_msecs % 1000),
6932 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6933 : total_msecs / 1000, (int) (total_msecs % 1000),
6934 : CheckpointStats.ckpt_sync_rels,
6935 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6936 : average_msecs / 1000, (int) (average_msecs % 1000),
6937 : (int) (PrevCheckPointDistance / 1024.0),
6938 : (int) (CheckPointDistanceEstimate / 1024.0),
6939 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6940 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6941 : else
6942 2396 : ereport(LOG,
6943 : (errmsg("checkpoint complete: wrote %d buffers (%.1f%%), "
6944 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6945 : "%d removed, %d recycled; write=%ld.%03d s, "
6946 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6947 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6948 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6949 : CheckpointStats.ckpt_bufs_written,
6950 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6951 : CheckpointStats.ckpt_slru_written,
6952 : CheckpointStats.ckpt_segs_added,
6953 : CheckpointStats.ckpt_segs_removed,
6954 : CheckpointStats.ckpt_segs_recycled,
6955 : write_msecs / 1000, (int) (write_msecs % 1000),
6956 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6957 : total_msecs / 1000, (int) (total_msecs % 1000),
6958 : CheckpointStats.ckpt_sync_rels,
6959 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6960 : average_msecs / 1000, (int) (average_msecs % 1000),
6961 : (int) (PrevCheckPointDistance / 1024.0),
6962 : (int) (CheckPointDistanceEstimate / 1024.0),
6963 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6964 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6965 : }
6966 :
6967 : /*
6968 : * Update the estimate of distance between checkpoints.
6969 : *
6970 : * The estimate is used to calculate the number of WAL segments to keep
6971 : * preallocated, see XLOGfileslop().
6972 : */
6973 : static void
6974 3360 : UpdateCheckPointDistanceEstimate(uint64 nbytes)
6975 : {
6976 : /*
6977 : * To estimate the number of segments consumed between checkpoints, keep a
6978 : * moving average of the amount of WAL generated in previous checkpoint
6979 : * cycles. However, if the load is bursty, with quiet periods and busy
6980 : * periods, we want to cater for the peak load. So instead of a plain
6981 : * moving average, let the average decline slowly if the previous cycle
6982 : * used less WAL than estimated, but bump it up immediately if it used
6983 : * more.
6984 : *
6985 : * When checkpoints are triggered by max_wal_size, this should converge to
6986 : * CheckpointSegments * wal_segment_size,
6987 : *
6988 : * Note: This doesn't pay any attention to what caused the checkpoint.
6989 : * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
6990 : * starting a base backup, are counted the same as those created
6991 : * automatically. The slow-decline will largely mask them out, if they are
6992 : * not frequent. If they are frequent, it seems reasonable to count them
6993 : * in as any others; if you issue a manual checkpoint every 5 minutes and
6994 : * never let a timed checkpoint happen, it makes sense to base the
6995 : * preallocation on that 5 minute interval rather than whatever
6996 : * checkpoint_timeout is set to.
6997 : */
6998 3360 : PrevCheckPointDistance = nbytes;
6999 3360 : if (CheckPointDistanceEstimate < nbytes)
7000 1408 : CheckPointDistanceEstimate = nbytes;
7001 : else
7002 1952 : CheckPointDistanceEstimate =
7003 1952 : (0.90 * CheckPointDistanceEstimate + 0.10 * (double) nbytes);
7004 3360 : }
7005 :
7006 : /*
7007 : * Update the ps display for a process running a checkpoint. Note that
7008 : * this routine should not do any allocations so as it can be called
7009 : * from a critical section.
7010 : */
7011 : static void
7012 6720 : update_checkpoint_display(int flags, bool restartpoint, bool reset)
7013 : {
7014 : /*
7015 : * The status is reported only for end-of-recovery and shutdown
7016 : * checkpoints or shutdown restartpoints. Updating the ps display is
7017 : * useful in those situations as it may not be possible to rely on
7018 : * pg_stat_activity to see the status of the checkpointer or the startup
7019 : * process.
7020 : */
7021 6720 : if ((flags & (CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IS_SHUTDOWN)) == 0)
7022 4264 : return;
7023 :
7024 2456 : if (reset)
7025 1228 : set_ps_display("");
7026 : else
7027 : {
7028 : char activitymsg[128];
7029 :
7030 3684 : snprintf(activitymsg, sizeof(activitymsg), "performing %s%s%s",
7031 1228 : (flags & CHECKPOINT_END_OF_RECOVERY) ? "end-of-recovery " : "",
7032 1228 : (flags & CHECKPOINT_IS_SHUTDOWN) ? "shutdown " : "",
7033 : restartpoint ? "restartpoint" : "checkpoint");
7034 1228 : set_ps_display(activitymsg);
7035 : }
7036 : }
7037 :
7038 :
7039 : /*
7040 : * Perform a checkpoint --- either during shutdown, or on-the-fly
7041 : *
7042 : * flags is a bitwise OR of the following:
7043 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
7044 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
7045 : * CHECKPOINT_IMMEDIATE: finish the checkpoint ASAP,
7046 : * ignoring checkpoint_completion_target parameter.
7047 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
7048 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
7049 : * CHECKPOINT_END_OF_RECOVERY).
7050 : * CHECKPOINT_FLUSH_ALL: also flush buffers of unlogged tables.
7051 : *
7052 : * Note: flags contains other bits, of interest here only for logging purposes.
7053 : * In particular note that this routine is synchronous and does not pay
7054 : * attention to CHECKPOINT_WAIT.
7055 : *
7056 : * If !shutdown then we are writing an online checkpoint. An XLOG_CHECKPOINT_REDO
7057 : * record is inserted into WAL at the logical location of the checkpoint, before
7058 : * flushing anything to disk, and when the checkpoint is eventually completed,
7059 : * and it is from this point that WAL replay will begin in the case of a recovery
7060 : * from this checkpoint. Once everything is written to disk, an
7061 : * XLOG_CHECKPOINT_ONLINE record is written to complete the checkpoint, and
7062 : * points back to the earlier XLOG_CHECKPOINT_REDO record. This mechanism allows
7063 : * other write-ahead log records to be written while the checkpoint is in
7064 : * progress, but we must be very careful about order of operations. This function
7065 : * may take many minutes to execute on a busy system.
7066 : *
7067 : * On the other hand, when shutdown is true, concurrent insertion into the
7068 : * write-ahead log is impossible, so there is no need for two separate records.
7069 : * In this case, we only insert an XLOG_CHECKPOINT_SHUTDOWN record, and it's
7070 : * both the record marking the completion of the checkpoint and the location
7071 : * from which WAL replay would begin if needed.
7072 : *
7073 : * Returns true if a new checkpoint was performed, or false if it was skipped
7074 : * because the system was idle.
7075 : */
7076 : bool
7077 2992 : CreateCheckPoint(int flags)
7078 : {
7079 : bool shutdown;
7080 : CheckPoint checkPoint;
7081 : XLogRecPtr recptr;
7082 : XLogSegNo _logSegNo;
7083 2992 : XLogCtlInsert *Insert = &XLogCtl->Insert;
7084 : uint32 freespace;
7085 : XLogRecPtr PriorRedoPtr;
7086 : XLogRecPtr last_important_lsn;
7087 : VirtualTransactionId *vxids;
7088 : int nvxids;
7089 2992 : int oldXLogAllowed = 0;
7090 :
7091 : /*
7092 : * An end-of-recovery checkpoint is really a shutdown checkpoint, just
7093 : * issued at a different time.
7094 : */
7095 2992 : if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
7096 1186 : shutdown = true;
7097 : else
7098 1806 : shutdown = false;
7099 :
7100 : /* sanity check */
7101 2992 : if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
7102 0 : elog(ERROR, "can't create a checkpoint during recovery");
7103 :
7104 : /*
7105 : * Prepare to accumulate statistics.
7106 : *
7107 : * Note: because it is possible for log_checkpoints to change while a
7108 : * checkpoint proceeds, we always accumulate stats, even if
7109 : * log_checkpoints is currently off.
7110 : */
7111 32912 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7112 2992 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7113 :
7114 : /*
7115 : * Let smgr prepare for checkpoint; this has to happen outside the
7116 : * critical section and before we determine the REDO pointer. Note that
7117 : * smgr must not do anything that'd have to be undone if we decide no
7118 : * checkpoint is needed.
7119 : */
7120 2992 : SyncPreCheckpoint();
7121 :
7122 : /*
7123 : * Use a critical section to force system panic if we have trouble.
7124 : */
7125 2992 : START_CRIT_SECTION();
7126 :
7127 2992 : if (shutdown)
7128 : {
7129 1186 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7130 1186 : ControlFile->state = DB_SHUTDOWNING;
7131 1186 : UpdateControlFile();
7132 1186 : LWLockRelease(ControlFileLock);
7133 : }
7134 :
7135 : /* Begin filling in the checkpoint WAL record */
7136 35904 : MemSet(&checkPoint, 0, sizeof(checkPoint));
7137 2992 : checkPoint.time = (pg_time_t) time(NULL);
7138 :
7139 : /*
7140 : * For Hot Standby, derive the oldestActiveXid before we fix the redo
7141 : * pointer. This allows us to begin accumulating changes to assemble our
7142 : * starting snapshot of locks and transactions.
7143 : */
7144 2992 : if (!shutdown && XLogStandbyInfoActive())
7145 1704 : checkPoint.oldestActiveXid = GetOldestActiveTransactionId();
7146 : else
7147 1288 : checkPoint.oldestActiveXid = InvalidTransactionId;
7148 :
7149 : /*
7150 : * Get location of last important record before acquiring insert locks (as
7151 : * GetLastImportantRecPtr() also locks WAL locks).
7152 : */
7153 2992 : last_important_lsn = GetLastImportantRecPtr();
7154 :
7155 : /*
7156 : * If this isn't a shutdown or forced checkpoint, and if there has been no
7157 : * WAL activity requiring a checkpoint, skip it. The idea here is to
7158 : * avoid inserting duplicate checkpoints when the system is idle.
7159 : */
7160 2992 : if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
7161 : CHECKPOINT_FORCE)) == 0)
7162 : {
7163 380 : if (last_important_lsn == ControlFile->checkPoint)
7164 : {
7165 0 : END_CRIT_SECTION();
7166 0 : ereport(DEBUG1,
7167 : (errmsg_internal("checkpoint skipped because system is idle")));
7168 0 : return false;
7169 : }
7170 : }
7171 :
7172 : /*
7173 : * An end-of-recovery checkpoint is created before anyone is allowed to
7174 : * write WAL. To allow us to write the checkpoint record, temporarily
7175 : * enable XLogInsertAllowed.
7176 : */
7177 2992 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7178 58 : oldXLogAllowed = LocalSetXLogInsertAllowed();
7179 :
7180 2992 : checkPoint.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7181 2992 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7182 58 : checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7183 : else
7184 2934 : checkPoint.PrevTimeLineID = checkPoint.ThisTimeLineID;
7185 :
7186 : /*
7187 : * We must block concurrent insertions while examining insert state.
7188 : */
7189 2992 : WALInsertLockAcquireExclusive();
7190 :
7191 2992 : checkPoint.fullPageWrites = Insert->fullPageWrites;
7192 2992 : checkPoint.wal_level = wal_level;
7193 :
7194 2992 : if (shutdown)
7195 : {
7196 1186 : XLogRecPtr curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
7197 :
7198 : /*
7199 : * Compute new REDO record ptr = location of next XLOG record.
7200 : *
7201 : * Since this is a shutdown checkpoint, there can't be any concurrent
7202 : * WAL insertion.
7203 : */
7204 1186 : freespace = INSERT_FREESPACE(curInsert);
7205 1186 : if (freespace == 0)
7206 : {
7207 0 : if (XLogSegmentOffset(curInsert, wal_segment_size) == 0)
7208 0 : curInsert += SizeOfXLogLongPHD;
7209 : else
7210 0 : curInsert += SizeOfXLogShortPHD;
7211 : }
7212 1186 : checkPoint.redo = curInsert;
7213 :
7214 : /*
7215 : * Here we update the shared RedoRecPtr for future XLogInsert calls;
7216 : * this must be done while holding all the insertion locks.
7217 : *
7218 : * Note: if we fail to complete the checkpoint, RedoRecPtr will be
7219 : * left pointing past where it really needs to point. This is okay;
7220 : * the only consequence is that XLogInsert might back up whole buffers
7221 : * that it didn't really need to. We can't postpone advancing
7222 : * RedoRecPtr because XLogInserts that happen while we are dumping
7223 : * buffers must assume that their buffer changes are not included in
7224 : * the checkpoint.
7225 : */
7226 1186 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
7227 : }
7228 :
7229 : /*
7230 : * Now we can release the WAL insertion locks, allowing other xacts to
7231 : * proceed while we are flushing disk buffers.
7232 : */
7233 2992 : WALInsertLockRelease();
7234 :
7235 : /*
7236 : * If this is an online checkpoint, we have not yet determined the redo
7237 : * point. We do so now by inserting the special XLOG_CHECKPOINT_REDO
7238 : * record; the LSN at which it starts becomes the new redo pointer. We
7239 : * don't do this for a shutdown checkpoint, because in that case no WAL
7240 : * can be written between the redo point and the insertion of the
7241 : * checkpoint record itself, so the checkpoint record itself serves to
7242 : * mark the redo point.
7243 : */
7244 2992 : if (!shutdown)
7245 : {
7246 : /* Include WAL level in record for WAL summarizer's benefit. */
7247 1806 : XLogBeginInsert();
7248 1806 : XLogRegisterData(&wal_level, sizeof(wal_level));
7249 1806 : (void) XLogInsert(RM_XLOG_ID, XLOG_CHECKPOINT_REDO);
7250 :
7251 : /*
7252 : * XLogInsertRecord will have updated XLogCtl->Insert.RedoRecPtr in
7253 : * shared memory and RedoRecPtr in backend-local memory, but we need
7254 : * to copy that into the record that will be inserted when the
7255 : * checkpoint is complete.
7256 : */
7257 1806 : checkPoint.redo = RedoRecPtr;
7258 : }
7259 :
7260 : /* Update the info_lck-protected copy of RedoRecPtr as well */
7261 2992 : SpinLockAcquire(&XLogCtl->info_lck);
7262 2992 : XLogCtl->RedoRecPtr = checkPoint.redo;
7263 2992 : SpinLockRelease(&XLogCtl->info_lck);
7264 :
7265 : /*
7266 : * If enabled, log checkpoint start. We postpone this until now so as not
7267 : * to log anything if we decided to skip the checkpoint.
7268 : */
7269 2992 : if (log_checkpoints)
7270 2396 : LogCheckpointStart(flags, false);
7271 :
7272 : /* Update the process title */
7273 2992 : update_checkpoint_display(flags, false, false);
7274 :
7275 : TRACE_POSTGRESQL_CHECKPOINT_START(flags);
7276 :
7277 : /*
7278 : * Get the other info we need for the checkpoint record.
7279 : *
7280 : * We don't need to save oldestClogXid in the checkpoint, it only matters
7281 : * for the short period in which clog is being truncated, and if we crash
7282 : * during that we'll redo the clog truncation and fix up oldestClogXid
7283 : * there.
7284 : */
7285 2992 : LWLockAcquire(XidGenLock, LW_SHARED);
7286 2992 : checkPoint.nextXid = TransamVariables->nextXid;
7287 2992 : checkPoint.oldestXid = TransamVariables->oldestXid;
7288 2992 : checkPoint.oldestXidDB = TransamVariables->oldestXidDB;
7289 2992 : LWLockRelease(XidGenLock);
7290 :
7291 2992 : LWLockAcquire(CommitTsLock, LW_SHARED);
7292 2992 : checkPoint.oldestCommitTsXid = TransamVariables->oldestCommitTsXid;
7293 2992 : checkPoint.newestCommitTsXid = TransamVariables->newestCommitTsXid;
7294 2992 : LWLockRelease(CommitTsLock);
7295 :
7296 2992 : LWLockAcquire(OidGenLock, LW_SHARED);
7297 2992 : checkPoint.nextOid = TransamVariables->nextOid;
7298 2992 : if (!shutdown)
7299 1806 : checkPoint.nextOid += TransamVariables->oidCount;
7300 2992 : LWLockRelease(OidGenLock);
7301 :
7302 2992 : MultiXactGetCheckptMulti(shutdown,
7303 : &checkPoint.nextMulti,
7304 : &checkPoint.nextMultiOffset,
7305 : &checkPoint.oldestMulti,
7306 : &checkPoint.oldestMultiDB);
7307 :
7308 : /*
7309 : * Having constructed the checkpoint record, ensure all shmem disk buffers
7310 : * and commit-log buffers are flushed to disk.
7311 : *
7312 : * This I/O could fail for various reasons. If so, we will fail to
7313 : * complete the checkpoint, but there is no reason to force a system
7314 : * panic. Accordingly, exit critical section while doing it.
7315 : */
7316 2992 : END_CRIT_SECTION();
7317 :
7318 : /*
7319 : * In some cases there are groups of actions that must all occur on one
7320 : * side or the other of a checkpoint record. Before flushing the
7321 : * checkpoint record we must explicitly wait for any backend currently
7322 : * performing those groups of actions.
7323 : *
7324 : * One example is end of transaction, so we must wait for any transactions
7325 : * that are currently in commit critical sections. If an xact inserted
7326 : * its commit record into XLOG just before the REDO point, then a crash
7327 : * restart from the REDO point would not replay that record, which means
7328 : * that our flushing had better include the xact's update of pg_xact. So
7329 : * we wait till he's out of his commit critical section before proceeding.
7330 : * See notes in RecordTransactionCommit().
7331 : *
7332 : * Because we've already released the insertion locks, this test is a bit
7333 : * fuzzy: it is possible that we will wait for xacts we didn't really need
7334 : * to wait for. But the delay should be short and it seems better to make
7335 : * checkpoint take a bit longer than to hold off insertions longer than
7336 : * necessary. (In fact, the whole reason we have this issue is that xact.c
7337 : * does commit record XLOG insertion and clog update as two separate steps
7338 : * protected by different locks, but again that seems best on grounds of
7339 : * minimizing lock contention.)
7340 : *
7341 : * A transaction that has not yet set delayChkptFlags when we look cannot
7342 : * be at risk, since it has not inserted its commit record yet; and one
7343 : * that's already cleared it is not at risk either, since it's done fixing
7344 : * clog and we will correctly flush the update below. So we cannot miss
7345 : * any xacts we need to wait for.
7346 : */
7347 2992 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
7348 2992 : if (nvxids > 0)
7349 : {
7350 : do
7351 : {
7352 : /*
7353 : * Keep absorbing fsync requests while we wait. There could even
7354 : * be a deadlock if we don't, if the process that prevents the
7355 : * checkpoint is trying to add a request to the queue.
7356 : */
7357 222 : AbsorbSyncRequests();
7358 :
7359 222 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
7360 222 : pg_usleep(10000L); /* wait for 10 msec */
7361 222 : pgstat_report_wait_end();
7362 222 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7363 : DELAY_CHKPT_START));
7364 : }
7365 2992 : pfree(vxids);
7366 :
7367 2992 : CheckPointGuts(checkPoint.redo, flags);
7368 :
7369 2992 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
7370 2992 : if (nvxids > 0)
7371 : {
7372 : do
7373 : {
7374 0 : AbsorbSyncRequests();
7375 :
7376 0 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_COMPLETE);
7377 0 : pg_usleep(10000L); /* wait for 10 msec */
7378 0 : pgstat_report_wait_end();
7379 0 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7380 : DELAY_CHKPT_COMPLETE));
7381 : }
7382 2992 : pfree(vxids);
7383 :
7384 : /*
7385 : * Take a snapshot of running transactions and write this to WAL. This
7386 : * allows us to reconstruct the state of running transactions during
7387 : * archive recovery, if required. Skip, if this info disabled.
7388 : *
7389 : * If we are shutting down, or Startup process is completing crash
7390 : * recovery we don't need to write running xact data.
7391 : */
7392 2992 : if (!shutdown && XLogStandbyInfoActive())
7393 1704 : LogStandbySnapshot();
7394 :
7395 2992 : START_CRIT_SECTION();
7396 :
7397 : /*
7398 : * Now insert the checkpoint record into XLOG.
7399 : */
7400 2992 : XLogBeginInsert();
7401 2992 : XLogRegisterData(&checkPoint, sizeof(checkPoint));
7402 2992 : recptr = XLogInsert(RM_XLOG_ID,
7403 : shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
7404 : XLOG_CHECKPOINT_ONLINE);
7405 :
7406 2992 : XLogFlush(recptr);
7407 :
7408 : /*
7409 : * We mustn't write any new WAL after a shutdown checkpoint, or it will be
7410 : * overwritten at next startup. No-one should even try, this just allows
7411 : * sanity-checking. In the case of an end-of-recovery checkpoint, we want
7412 : * to just temporarily disable writing until the system has exited
7413 : * recovery.
7414 : */
7415 2992 : if (shutdown)
7416 : {
7417 1186 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7418 58 : LocalXLogInsertAllowed = oldXLogAllowed;
7419 : else
7420 1128 : LocalXLogInsertAllowed = 0; /* never again write WAL */
7421 : }
7422 :
7423 : /*
7424 : * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
7425 : * = end of actual checkpoint record.
7426 : */
7427 2992 : if (shutdown && checkPoint.redo != ProcLastRecPtr)
7428 0 : ereport(PANIC,
7429 : (errmsg("concurrent write-ahead log activity while database system is shutting down")));
7430 :
7431 : /*
7432 : * Remember the prior checkpoint's redo ptr for
7433 : * UpdateCheckPointDistanceEstimate()
7434 : */
7435 2992 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7436 :
7437 : /*
7438 : * Update the control file.
7439 : */
7440 2992 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7441 2992 : if (shutdown)
7442 1186 : ControlFile->state = DB_SHUTDOWNED;
7443 2992 : ControlFile->checkPoint = ProcLastRecPtr;
7444 2992 : ControlFile->checkPointCopy = checkPoint;
7445 : /* crash recovery should always recover to the end of WAL */
7446 2992 : ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
7447 2992 : ControlFile->minRecoveryPointTLI = 0;
7448 :
7449 : /*
7450 : * Persist unloggedLSN value. It's reset on crash recovery, so this goes
7451 : * unused on non-shutdown checkpoints, but seems useful to store it always
7452 : * for debugging purposes.
7453 : */
7454 2992 : ControlFile->unloggedLSN = pg_atomic_read_membarrier_u64(&XLogCtl->unloggedLSN);
7455 :
7456 2992 : UpdateControlFile();
7457 2992 : LWLockRelease(ControlFileLock);
7458 :
7459 : /* Update shared-memory copy of checkpoint XID/epoch */
7460 2992 : SpinLockAcquire(&XLogCtl->info_lck);
7461 2992 : XLogCtl->ckptFullXid = checkPoint.nextXid;
7462 2992 : SpinLockRelease(&XLogCtl->info_lck);
7463 :
7464 : /*
7465 : * We are now done with critical updates; no need for system panic if we
7466 : * have trouble while fooling with old log segments.
7467 : */
7468 2992 : END_CRIT_SECTION();
7469 :
7470 : /*
7471 : * WAL summaries end when the next XLOG_CHECKPOINT_REDO or
7472 : * XLOG_CHECKPOINT_SHUTDOWN record is reached. This is the first point
7473 : * where (a) we're not inside of a critical section and (b) we can be
7474 : * certain that the relevant record has been flushed to disk, which must
7475 : * happen before it can be summarized.
7476 : *
7477 : * If this is a shutdown checkpoint, then this happens reasonably
7478 : * promptly: we've only just inserted and flushed the
7479 : * XLOG_CHECKPOINT_SHUTDOWN record. If this is not a shutdown checkpoint,
7480 : * then this might not be very prompt at all: the XLOG_CHECKPOINT_REDO
7481 : * record was written before we began flushing data to disk, and that
7482 : * could be many minutes ago at this point. However, we don't XLogFlush()
7483 : * after inserting that record, so we're not guaranteed that it's on disk
7484 : * until after the above call that flushes the XLOG_CHECKPOINT_ONLINE
7485 : * record.
7486 : */
7487 2992 : WakeupWalSummarizer();
7488 :
7489 : /*
7490 : * Let smgr do post-checkpoint cleanup (eg, deleting old files).
7491 : */
7492 2992 : SyncPostCheckpoint();
7493 :
7494 : /*
7495 : * Update the average distance between checkpoints if the prior checkpoint
7496 : * exists.
7497 : */
7498 2992 : if (PriorRedoPtr != InvalidXLogRecPtr)
7499 2992 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7500 :
7501 : #ifdef USE_INJECTION_POINTS
7502 2992 : INJECTION_POINT("checkpoint-before-old-wal-removal", NULL);
7503 : #endif
7504 :
7505 : /*
7506 : * Delete old log files, those no longer needed for last checkpoint to
7507 : * prevent the disk holding the xlog from growing full.
7508 : */
7509 2992 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7510 2992 : KeepLogSeg(recptr, &_logSegNo);
7511 2992 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7512 : _logSegNo, InvalidOid,
7513 : InvalidTransactionId))
7514 : {
7515 : /*
7516 : * Some slots have been invalidated; recalculate the old-segment
7517 : * horizon, starting again from RedoRecPtr.
7518 : */
7519 6 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7520 6 : KeepLogSeg(recptr, &_logSegNo);
7521 : }
7522 2992 : _logSegNo--;
7523 2992 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
7524 : checkPoint.ThisTimeLineID);
7525 :
7526 : /*
7527 : * Make more log segments if needed. (Do this after recycling old log
7528 : * segments, since that may supply some of the needed files.)
7529 : */
7530 2992 : if (!shutdown)
7531 1806 : PreallocXlogFiles(recptr, checkPoint.ThisTimeLineID);
7532 :
7533 : /*
7534 : * Truncate pg_subtrans if possible. We can throw away all data before
7535 : * the oldest XMIN of any running transaction. No future transaction will
7536 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7537 : * in subtrans.c). During recovery, though, we mustn't do this because
7538 : * StartupSUBTRANS hasn't been called yet.
7539 : */
7540 2992 : if (!RecoveryInProgress())
7541 2934 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7542 :
7543 : /* Real work is done; log and update stats. */
7544 2992 : LogCheckpointEnd(false);
7545 :
7546 : /* Reset the process title */
7547 2992 : update_checkpoint_display(flags, false, true);
7548 :
7549 : TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
7550 : NBuffers,
7551 : CheckpointStats.ckpt_segs_added,
7552 : CheckpointStats.ckpt_segs_removed,
7553 : CheckpointStats.ckpt_segs_recycled);
7554 :
7555 2992 : return true;
7556 : }
7557 :
7558 : /*
7559 : * Mark the end of recovery in WAL though without running a full checkpoint.
7560 : * We can expect that a restartpoint is likely to be in progress as we
7561 : * do this, though we are unwilling to wait for it to complete.
7562 : *
7563 : * CreateRestartPoint() allows for the case where recovery may end before
7564 : * the restartpoint completes so there is no concern of concurrent behaviour.
7565 : */
7566 : static void
7567 84 : CreateEndOfRecoveryRecord(void)
7568 : {
7569 : xl_end_of_recovery xlrec;
7570 : XLogRecPtr recptr;
7571 :
7572 : /* sanity check */
7573 84 : if (!RecoveryInProgress())
7574 0 : elog(ERROR, "can only be used to end recovery");
7575 :
7576 84 : xlrec.end_time = GetCurrentTimestamp();
7577 84 : xlrec.wal_level = wal_level;
7578 :
7579 84 : WALInsertLockAcquireExclusive();
7580 84 : xlrec.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7581 84 : xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7582 84 : WALInsertLockRelease();
7583 :
7584 84 : START_CRIT_SECTION();
7585 :
7586 84 : XLogBeginInsert();
7587 84 : XLogRegisterData(&xlrec, sizeof(xl_end_of_recovery));
7588 84 : recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY);
7589 :
7590 84 : XLogFlush(recptr);
7591 :
7592 : /*
7593 : * Update the control file so that crash recovery can follow the timeline
7594 : * changes to this point.
7595 : */
7596 84 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7597 84 : ControlFile->minRecoveryPoint = recptr;
7598 84 : ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
7599 84 : UpdateControlFile();
7600 84 : LWLockRelease(ControlFileLock);
7601 :
7602 84 : END_CRIT_SECTION();
7603 84 : }
7604 :
7605 : /*
7606 : * Write an OVERWRITE_CONTRECORD message.
7607 : *
7608 : * When on WAL replay we expect a continuation record at the start of a page
7609 : * that is not there, recovery ends and WAL writing resumes at that point.
7610 : * But it's wrong to resume writing new WAL back at the start of the record
7611 : * that was broken, because downstream consumers of that WAL (physical
7612 : * replicas) are not prepared to "rewind". So the first action after
7613 : * finishing replay of all valid WAL must be to write a record of this type
7614 : * at the point where the contrecord was missing; to support xlogreader
7615 : * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
7616 : * to the page header where the record occurs. xlogreader has an ad-hoc
7617 : * mechanism to report metadata about the broken record, which is what we
7618 : * use here.
7619 : *
7620 : * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
7621 : * skip the record it was reading, and pass back the LSN of the skipped
7622 : * record, so that its caller can verify (on "replay" of that record) that the
7623 : * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
7624 : *
7625 : * 'aborted_lsn' is the beginning position of the record that was incomplete.
7626 : * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
7627 : * beginning of the XLOG page where the record is to be inserted. They must
7628 : * match the current WAL insert position, they're passed here just so that we
7629 : * can verify that.
7630 : */
7631 : static XLogRecPtr
7632 20 : CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn, XLogRecPtr pagePtr,
7633 : TimeLineID newTLI)
7634 : {
7635 : xl_overwrite_contrecord xlrec;
7636 : XLogRecPtr recptr;
7637 : XLogPageHeader pagehdr;
7638 : XLogRecPtr startPos;
7639 :
7640 : /* sanity checks */
7641 20 : if (!RecoveryInProgress())
7642 0 : elog(ERROR, "can only be used at end of recovery");
7643 20 : if (pagePtr % XLOG_BLCKSZ != 0)
7644 0 : elog(ERROR, "invalid position for missing continuation record %X/%08X",
7645 : LSN_FORMAT_ARGS(pagePtr));
7646 :
7647 : /* The current WAL insert position should be right after the page header */
7648 20 : startPos = pagePtr;
7649 20 : if (XLogSegmentOffset(startPos, wal_segment_size) == 0)
7650 2 : startPos += SizeOfXLogLongPHD;
7651 : else
7652 18 : startPos += SizeOfXLogShortPHD;
7653 20 : recptr = GetXLogInsertRecPtr();
7654 20 : if (recptr != startPos)
7655 0 : elog(ERROR, "invalid WAL insert position %X/%08X for OVERWRITE_CONTRECORD",
7656 : LSN_FORMAT_ARGS(recptr));
7657 :
7658 20 : START_CRIT_SECTION();
7659 :
7660 : /*
7661 : * Initialize the XLOG page header (by GetXLogBuffer), and set the
7662 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
7663 : *
7664 : * No other backend is allowed to write WAL yet, so acquiring the WAL
7665 : * insertion lock is just pro forma.
7666 : */
7667 20 : WALInsertLockAcquire();
7668 20 : pagehdr = (XLogPageHeader) GetXLogBuffer(pagePtr, newTLI);
7669 20 : pagehdr->xlp_info |= XLP_FIRST_IS_OVERWRITE_CONTRECORD;
7670 20 : WALInsertLockRelease();
7671 :
7672 : /*
7673 : * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
7674 : * page. We know it becomes the first record, because no other backend is
7675 : * allowed to write WAL yet.
7676 : */
7677 20 : XLogBeginInsert();
7678 20 : xlrec.overwritten_lsn = aborted_lsn;
7679 20 : xlrec.overwrite_time = GetCurrentTimestamp();
7680 20 : XLogRegisterData(&xlrec, sizeof(xl_overwrite_contrecord));
7681 20 : recptr = XLogInsert(RM_XLOG_ID, XLOG_OVERWRITE_CONTRECORD);
7682 :
7683 : /* check that the record was inserted to the right place */
7684 20 : if (ProcLastRecPtr != startPos)
7685 0 : elog(ERROR, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%08X",
7686 : LSN_FORMAT_ARGS(ProcLastRecPtr));
7687 :
7688 20 : XLogFlush(recptr);
7689 :
7690 20 : END_CRIT_SECTION();
7691 :
7692 20 : return recptr;
7693 : }
7694 :
7695 : /*
7696 : * Flush all data in shared memory to disk, and fsync
7697 : *
7698 : * This is the common code shared between regular checkpoints and
7699 : * recovery restartpoints.
7700 : */
7701 : static void
7702 3360 : CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
7703 : {
7704 3360 : CheckPointRelationMap();
7705 3360 : CheckPointReplicationSlots(flags & CHECKPOINT_IS_SHUTDOWN);
7706 3360 : CheckPointSnapBuild();
7707 3360 : CheckPointLogicalRewriteHeap();
7708 3360 : CheckPointReplicationOrigin();
7709 :
7710 : /* Write out all dirty data in SLRUs and the main buffer pool */
7711 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags);
7712 3360 : CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
7713 3360 : CheckPointCLOG();
7714 3360 : CheckPointCommitTs();
7715 3360 : CheckPointSUBTRANS();
7716 3360 : CheckPointMultiXact();
7717 3360 : CheckPointPredicate();
7718 3360 : CheckPointBuffers(flags);
7719 :
7720 : /* Perform all queued up fsyncs */
7721 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
7722 3360 : CheckpointStats.ckpt_sync_t = GetCurrentTimestamp();
7723 3360 : ProcessSyncRequests();
7724 3360 : CheckpointStats.ckpt_sync_end_t = GetCurrentTimestamp();
7725 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
7726 :
7727 : /* We deliberately delay 2PC checkpointing as long as possible */
7728 3360 : CheckPointTwoPhase(checkPointRedo);
7729 3360 : }
7730 :
7731 : /*
7732 : * Save a checkpoint for recovery restart if appropriate
7733 : *
7734 : * This function is called each time a checkpoint record is read from XLOG.
7735 : * It must determine whether the checkpoint represents a safe restartpoint or
7736 : * not. If so, the checkpoint record is stashed in shared memory so that
7737 : * CreateRestartPoint can consult it. (Note that the latter function is
7738 : * executed by the checkpointer, while this one will be executed by the
7739 : * startup process.)
7740 : */
7741 : static void
7742 1414 : RecoveryRestartPoint(const CheckPoint *checkPoint, XLogReaderState *record)
7743 : {
7744 : /*
7745 : * Also refrain from creating a restartpoint if we have seen any
7746 : * references to non-existent pages. Restarting recovery from the
7747 : * restartpoint would not see the references, so we would lose the
7748 : * cross-check that the pages belonged to a relation that was dropped
7749 : * later.
7750 : */
7751 1414 : if (XLogHaveInvalidPages())
7752 : {
7753 0 : elog(DEBUG2,
7754 : "could not record restart point at %X/%08X because there are unresolved references to invalid pages",
7755 : LSN_FORMAT_ARGS(checkPoint->redo));
7756 0 : return;
7757 : }
7758 :
7759 : /*
7760 : * Copy the checkpoint record to shared memory, so that checkpointer can
7761 : * work out the next time it wants to perform a restartpoint.
7762 : */
7763 1414 : SpinLockAcquire(&XLogCtl->info_lck);
7764 1414 : XLogCtl->lastCheckPointRecPtr = record->ReadRecPtr;
7765 1414 : XLogCtl->lastCheckPointEndPtr = record->EndRecPtr;
7766 1414 : XLogCtl->lastCheckPoint = *checkPoint;
7767 1414 : SpinLockRelease(&XLogCtl->info_lck);
7768 : }
7769 :
7770 : /*
7771 : * Establish a restartpoint if possible.
7772 : *
7773 : * This is similar to CreateCheckPoint, but is used during WAL recovery
7774 : * to establish a point from which recovery can roll forward without
7775 : * replaying the entire recovery log.
7776 : *
7777 : * Returns true if a new restartpoint was established. We can only establish
7778 : * a restartpoint if we have replayed a safe checkpoint record since last
7779 : * restartpoint.
7780 : */
7781 : bool
7782 1156 : CreateRestartPoint(int flags)
7783 : {
7784 : XLogRecPtr lastCheckPointRecPtr;
7785 : XLogRecPtr lastCheckPointEndPtr;
7786 : CheckPoint lastCheckPoint;
7787 : XLogRecPtr PriorRedoPtr;
7788 : XLogRecPtr receivePtr;
7789 : XLogRecPtr replayPtr;
7790 : TimeLineID replayTLI;
7791 : XLogRecPtr endptr;
7792 : XLogSegNo _logSegNo;
7793 : TimestampTz xtime;
7794 :
7795 : /* Concurrent checkpoint/restartpoint cannot happen */
7796 : Assert(!IsUnderPostmaster || MyBackendType == B_CHECKPOINTER);
7797 :
7798 : /* Get a local copy of the last safe checkpoint record. */
7799 1156 : SpinLockAcquire(&XLogCtl->info_lck);
7800 1156 : lastCheckPointRecPtr = XLogCtl->lastCheckPointRecPtr;
7801 1156 : lastCheckPointEndPtr = XLogCtl->lastCheckPointEndPtr;
7802 1156 : lastCheckPoint = XLogCtl->lastCheckPoint;
7803 1156 : SpinLockRelease(&XLogCtl->info_lck);
7804 :
7805 : /*
7806 : * Check that we're still in recovery mode. It's ok if we exit recovery
7807 : * mode after this check, the restart point is valid anyway.
7808 : */
7809 1156 : if (!RecoveryInProgress())
7810 : {
7811 0 : ereport(DEBUG2,
7812 : (errmsg_internal("skipping restartpoint, recovery has already ended")));
7813 0 : return false;
7814 : }
7815 :
7816 : /*
7817 : * If the last checkpoint record we've replayed is already our last
7818 : * restartpoint, we can't perform a new restart point. We still update
7819 : * minRecoveryPoint in that case, so that if this is a shutdown restart
7820 : * point, we won't start up earlier than before. That's not strictly
7821 : * necessary, but when hot standby is enabled, it would be rather weird if
7822 : * the database opened up for read-only connections at a point-in-time
7823 : * before the last shutdown. Such time travel is still possible in case of
7824 : * immediate shutdown, though.
7825 : *
7826 : * We don't explicitly advance minRecoveryPoint when we do create a
7827 : * restartpoint. It's assumed that flushing the buffers will do that as a
7828 : * side-effect.
7829 : */
7830 1156 : if (XLogRecPtrIsInvalid(lastCheckPointRecPtr) ||
7831 492 : lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
7832 : {
7833 788 : ereport(DEBUG2,
7834 : errmsg_internal("skipping restartpoint, already performed at %X/%08X",
7835 : LSN_FORMAT_ARGS(lastCheckPoint.redo)));
7836 :
7837 788 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
7838 788 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7839 : {
7840 68 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7841 68 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7842 68 : UpdateControlFile();
7843 68 : LWLockRelease(ControlFileLock);
7844 : }
7845 788 : return false;
7846 : }
7847 :
7848 : /*
7849 : * Update the shared RedoRecPtr so that the startup process can calculate
7850 : * the number of segments replayed since last restartpoint, and request a
7851 : * restartpoint if it exceeds CheckPointSegments.
7852 : *
7853 : * Like in CreateCheckPoint(), hold off insertions to update it, although
7854 : * during recovery this is just pro forma, because no WAL insertions are
7855 : * happening.
7856 : */
7857 368 : WALInsertLockAcquireExclusive();
7858 368 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = lastCheckPoint.redo;
7859 368 : WALInsertLockRelease();
7860 :
7861 : /* Also update the info_lck-protected copy */
7862 368 : SpinLockAcquire(&XLogCtl->info_lck);
7863 368 : XLogCtl->RedoRecPtr = lastCheckPoint.redo;
7864 368 : SpinLockRelease(&XLogCtl->info_lck);
7865 :
7866 : /*
7867 : * Prepare to accumulate statistics.
7868 : *
7869 : * Note: because it is possible for log_checkpoints to change while a
7870 : * checkpoint proceeds, we always accumulate stats, even if
7871 : * log_checkpoints is currently off.
7872 : */
7873 4048 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7874 368 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7875 :
7876 368 : if (log_checkpoints)
7877 368 : LogCheckpointStart(flags, true);
7878 :
7879 : /* Update the process title */
7880 368 : update_checkpoint_display(flags, true, false);
7881 :
7882 368 : CheckPointGuts(lastCheckPoint.redo, flags);
7883 :
7884 : /*
7885 : * This location needs to be after CheckPointGuts() to ensure that some
7886 : * work has already happened during this checkpoint.
7887 : */
7888 368 : INJECTION_POINT("create-restart-point", NULL);
7889 :
7890 : /*
7891 : * Remember the prior checkpoint's redo ptr for
7892 : * UpdateCheckPointDistanceEstimate()
7893 : */
7894 368 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7895 :
7896 : /*
7897 : * Update pg_control, using current time. Check that it still shows an
7898 : * older checkpoint, else do nothing; this is a quick hack to make sure
7899 : * nothing really bad happens if somehow we get here after the
7900 : * end-of-recovery checkpoint.
7901 : */
7902 368 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7903 368 : if (ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
7904 : {
7905 : /*
7906 : * Update the checkpoint information. We do this even if the cluster
7907 : * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
7908 : * segments recycled below.
7909 : */
7910 368 : ControlFile->checkPoint = lastCheckPointRecPtr;
7911 368 : ControlFile->checkPointCopy = lastCheckPoint;
7912 :
7913 : /*
7914 : * Ensure minRecoveryPoint is past the checkpoint record and update it
7915 : * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
7916 : * this will have happened already while writing out dirty buffers,
7917 : * but not necessarily - e.g. because no buffers were dirtied. We do
7918 : * this because a backup performed in recovery uses minRecoveryPoint
7919 : * to determine which WAL files must be included in the backup, and
7920 : * the file (or files) containing the checkpoint record must be
7921 : * included, at a minimum. Note that for an ordinary restart of
7922 : * recovery there's no value in having the minimum recovery point any
7923 : * earlier than this anyway, because redo will begin just after the
7924 : * checkpoint record.
7925 : */
7926 368 : if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
7927 : {
7928 368 : if (ControlFile->minRecoveryPoint < lastCheckPointEndPtr)
7929 : {
7930 38 : ControlFile->minRecoveryPoint = lastCheckPointEndPtr;
7931 38 : ControlFile->minRecoveryPointTLI = lastCheckPoint.ThisTimeLineID;
7932 :
7933 : /* update local copy */
7934 38 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
7935 38 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
7936 : }
7937 368 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7938 42 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7939 : }
7940 368 : UpdateControlFile();
7941 : }
7942 368 : LWLockRelease(ControlFileLock);
7943 :
7944 : /*
7945 : * Update the average distance between checkpoints/restartpoints if the
7946 : * prior checkpoint exists.
7947 : */
7948 368 : if (PriorRedoPtr != InvalidXLogRecPtr)
7949 368 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7950 :
7951 : /*
7952 : * Delete old log files, those no longer needed for last restartpoint to
7953 : * prevent the disk holding the xlog from growing full.
7954 : */
7955 368 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7956 :
7957 : /*
7958 : * Retreat _logSegNo using the current end of xlog replayed or received,
7959 : * whichever is later.
7960 : */
7961 368 : receivePtr = GetWalRcvFlushRecPtr(NULL, NULL);
7962 368 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
7963 368 : endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
7964 368 : KeepLogSeg(endptr, &_logSegNo);
7965 368 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7966 : _logSegNo, InvalidOid,
7967 : InvalidTransactionId))
7968 : {
7969 : /*
7970 : * Some slots have been invalidated; recalculate the old-segment
7971 : * horizon, starting again from RedoRecPtr.
7972 : */
7973 2 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7974 2 : KeepLogSeg(endptr, &_logSegNo);
7975 : }
7976 368 : _logSegNo--;
7977 :
7978 : /*
7979 : * Try to recycle segments on a useful timeline. If we've been promoted
7980 : * since the beginning of this restartpoint, use the new timeline chosen
7981 : * at end of recovery. If we're still in recovery, use the timeline we're
7982 : * currently replaying.
7983 : *
7984 : * There is no guarantee that the WAL segments will be useful on the
7985 : * current timeline; if recovery proceeds to a new timeline right after
7986 : * this, the pre-allocated WAL segments on this timeline will not be used,
7987 : * and will go wasted until recycled on the next restartpoint. We'll live
7988 : * with that.
7989 : */
7990 368 : if (!RecoveryInProgress())
7991 0 : replayTLI = XLogCtl->InsertTimeLineID;
7992 :
7993 368 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, endptr, replayTLI);
7994 :
7995 : /*
7996 : * Make more log segments if needed. (Do this after recycling old log
7997 : * segments, since that may supply some of the needed files.)
7998 : */
7999 368 : PreallocXlogFiles(endptr, replayTLI);
8000 :
8001 : /*
8002 : * Truncate pg_subtrans if possible. We can throw away all data before
8003 : * the oldest XMIN of any running transaction. No future transaction will
8004 : * attempt to reference any pg_subtrans entry older than that (see Asserts
8005 : * in subtrans.c). When hot standby is disabled, though, we mustn't do
8006 : * this because StartupSUBTRANS hasn't been called yet.
8007 : */
8008 368 : if (EnableHotStandby)
8009 368 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
8010 :
8011 : /* Real work is done; log and update stats. */
8012 368 : LogCheckpointEnd(true);
8013 :
8014 : /* Reset the process title */
8015 368 : update_checkpoint_display(flags, true, true);
8016 :
8017 368 : xtime = GetLatestXTime();
8018 368 : ereport((log_checkpoints ? LOG : DEBUG2),
8019 : errmsg("recovery restart point at %X/%08X",
8020 : LSN_FORMAT_ARGS(lastCheckPoint.redo)),
8021 : xtime ? errdetail("Last completed transaction was at log time %s.",
8022 : timestamptz_to_str(xtime)) : 0);
8023 :
8024 : /*
8025 : * Finally, execute archive_cleanup_command, if any.
8026 : */
8027 368 : if (archiveCleanupCommand && strcmp(archiveCleanupCommand, "") != 0)
8028 0 : ExecuteRecoveryCommand(archiveCleanupCommand,
8029 : "archive_cleanup_command",
8030 : false,
8031 : WAIT_EVENT_ARCHIVE_CLEANUP_COMMAND);
8032 :
8033 368 : return true;
8034 : }
8035 :
8036 : /*
8037 : * Report availability of WAL for the given target LSN
8038 : * (typically a slot's restart_lsn)
8039 : *
8040 : * Returns one of the following enum values:
8041 : *
8042 : * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
8043 : * max_wal_size.
8044 : *
8045 : * * WALAVAIL_EXTENDED means it is still available by preserving extra
8046 : * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
8047 : * than max_wal_size, this state is not returned.
8048 : *
8049 : * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
8050 : * remove reserved segments. The walsender using this slot may return to the
8051 : * above.
8052 : *
8053 : * * WALAVAIL_REMOVED means it has been removed. A replication stream on
8054 : * a slot with this LSN cannot continue. (Any associated walsender
8055 : * processes should have been terminated already.)
8056 : *
8057 : * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
8058 : */
8059 : WALAvailability
8060 956 : GetWALAvailability(XLogRecPtr targetLSN)
8061 : {
8062 : XLogRecPtr currpos; /* current write LSN */
8063 : XLogSegNo currSeg; /* segid of currpos */
8064 : XLogSegNo targetSeg; /* segid of targetLSN */
8065 : XLogSegNo oldestSeg; /* actual oldest segid */
8066 : XLogSegNo oldestSegMaxWalSize; /* oldest segid kept by max_wal_size */
8067 : XLogSegNo oldestSlotSeg; /* oldest segid kept by slot */
8068 : uint64 keepSegs;
8069 :
8070 : /*
8071 : * slot does not reserve WAL. Either deactivated, or has never been active
8072 : */
8073 956 : if (XLogRecPtrIsInvalid(targetLSN))
8074 34 : return WALAVAIL_INVALID_LSN;
8075 :
8076 : /*
8077 : * Calculate the oldest segment currently reserved by all slots,
8078 : * considering wal_keep_size and max_slot_wal_keep_size. Initialize
8079 : * oldestSlotSeg to the current segment.
8080 : */
8081 922 : currpos = GetXLogWriteRecPtr();
8082 922 : XLByteToSeg(currpos, oldestSlotSeg, wal_segment_size);
8083 922 : KeepLogSeg(currpos, &oldestSlotSeg);
8084 :
8085 : /*
8086 : * Find the oldest extant segment file. We get 1 until checkpoint removes
8087 : * the first WAL segment file since startup, which causes the status being
8088 : * wrong under certain abnormal conditions but that doesn't actually harm.
8089 : */
8090 922 : oldestSeg = XLogGetLastRemovedSegno() + 1;
8091 :
8092 : /* calculate oldest segment by max_wal_size */
8093 922 : XLByteToSeg(currpos, currSeg, wal_segment_size);
8094 922 : keepSegs = ConvertToXSegs(max_wal_size_mb, wal_segment_size) + 1;
8095 :
8096 922 : if (currSeg > keepSegs)
8097 16 : oldestSegMaxWalSize = currSeg - keepSegs;
8098 : else
8099 906 : oldestSegMaxWalSize = 1;
8100 :
8101 : /* the segment we care about */
8102 922 : XLByteToSeg(targetLSN, targetSeg, wal_segment_size);
8103 :
8104 : /*
8105 : * No point in returning reserved or extended status values if the
8106 : * targetSeg is known to be lost.
8107 : */
8108 922 : if (targetSeg >= oldestSlotSeg)
8109 : {
8110 : /* show "reserved" when targetSeg is within max_wal_size */
8111 920 : if (targetSeg >= oldestSegMaxWalSize)
8112 916 : return WALAVAIL_RESERVED;
8113 :
8114 : /* being retained by slots exceeding max_wal_size */
8115 4 : return WALAVAIL_EXTENDED;
8116 : }
8117 :
8118 : /* WAL segments are no longer retained but haven't been removed yet */
8119 2 : if (targetSeg >= oldestSeg)
8120 2 : return WALAVAIL_UNRESERVED;
8121 :
8122 : /* Definitely lost */
8123 0 : return WALAVAIL_REMOVED;
8124 : }
8125 :
8126 :
8127 : /*
8128 : * Retreat *logSegNo to the last segment that we need to retain because of
8129 : * either wal_keep_size or replication slots.
8130 : *
8131 : * This is calculated by subtracting wal_keep_size from the given xlog
8132 : * location, recptr and by making sure that that result is below the
8133 : * requirement of replication slots. For the latter criterion we do consider
8134 : * the effects of max_slot_wal_keep_size: reserve at most that much space back
8135 : * from recptr.
8136 : *
8137 : * Note about replication slots: if this function calculates a value
8138 : * that's further ahead than what slots need reserved, then affected
8139 : * slots need to be invalidated and this function invoked again.
8140 : * XXX it might be a good idea to rewrite this function so that
8141 : * invalidation is optionally done here, instead.
8142 : */
8143 : static void
8144 4290 : KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
8145 : {
8146 : XLogSegNo currSegNo;
8147 : XLogSegNo segno;
8148 : XLogRecPtr keep;
8149 :
8150 4290 : XLByteToSeg(recptr, currSegNo, wal_segment_size);
8151 4290 : segno = currSegNo;
8152 :
8153 : /*
8154 : * Calculate how many segments are kept by slots first, adjusting for
8155 : * max_slot_wal_keep_size.
8156 : */
8157 4290 : keep = XLogGetReplicationSlotMinimumLSN();
8158 4290 : if (keep != InvalidXLogRecPtr && keep < recptr)
8159 : {
8160 1184 : XLByteToSeg(keep, segno, wal_segment_size);
8161 :
8162 : /* Cap by max_slot_wal_keep_size ... */
8163 1184 : if (max_slot_wal_keep_size_mb >= 0)
8164 : {
8165 : uint64 slot_keep_segs;
8166 :
8167 42 : slot_keep_segs =
8168 42 : ConvertToXSegs(max_slot_wal_keep_size_mb, wal_segment_size);
8169 :
8170 42 : if (currSegNo - segno > slot_keep_segs)
8171 10 : segno = currSegNo - slot_keep_segs;
8172 : }
8173 : }
8174 :
8175 : /*
8176 : * If WAL summarization is in use, don't remove WAL that has yet to be
8177 : * summarized.
8178 : */
8179 4290 : keep = GetOldestUnsummarizedLSN(NULL, NULL);
8180 4290 : if (keep != InvalidXLogRecPtr)
8181 : {
8182 : XLogSegNo unsummarized_segno;
8183 :
8184 4 : XLByteToSeg(keep, unsummarized_segno, wal_segment_size);
8185 4 : if (unsummarized_segno < segno)
8186 4 : segno = unsummarized_segno;
8187 : }
8188 :
8189 : /* but, keep at least wal_keep_size if that's set */
8190 4290 : if (wal_keep_size_mb > 0)
8191 : {
8192 : uint64 keep_segs;
8193 :
8194 138 : keep_segs = ConvertToXSegs(wal_keep_size_mb, wal_segment_size);
8195 138 : if (currSegNo - segno < keep_segs)
8196 : {
8197 : /* avoid underflow, don't go below 1 */
8198 138 : if (currSegNo <= keep_segs)
8199 130 : segno = 1;
8200 : else
8201 8 : segno = currSegNo - keep_segs;
8202 : }
8203 : }
8204 :
8205 : /* don't delete WAL segments newer than the calculated segment */
8206 4290 : if (segno < *logSegNo)
8207 748 : *logSegNo = segno;
8208 4290 : }
8209 :
8210 : /*
8211 : * Write a NEXTOID log record
8212 : */
8213 : void
8214 1192 : XLogPutNextOid(Oid nextOid)
8215 : {
8216 1192 : XLogBeginInsert();
8217 1192 : XLogRegisterData(&nextOid, sizeof(Oid));
8218 1192 : (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID);
8219 :
8220 : /*
8221 : * We need not flush the NEXTOID record immediately, because any of the
8222 : * just-allocated OIDs could only reach disk as part of a tuple insert or
8223 : * update that would have its own XLOG record that must follow the NEXTOID
8224 : * record. Therefore, the standard buffer LSN interlock applied to those
8225 : * records will ensure no such OID reaches disk before the NEXTOID record
8226 : * does.
8227 : *
8228 : * Note, however, that the above statement only covers state "within" the
8229 : * database. When we use a generated OID as a file or directory name, we
8230 : * are in a sense violating the basic WAL rule, because that filesystem
8231 : * change may reach disk before the NEXTOID WAL record does. The impact
8232 : * of this is that if a database crash occurs immediately afterward, we
8233 : * might after restart re-generate the same OID and find that it conflicts
8234 : * with the leftover file or directory. But since for safety's sake we
8235 : * always loop until finding a nonconflicting filename, this poses no real
8236 : * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8237 : */
8238 1192 : }
8239 :
8240 : /*
8241 : * Write an XLOG SWITCH record.
8242 : *
8243 : * Here we just blindly issue an XLogInsert request for the record.
8244 : * All the magic happens inside XLogInsert.
8245 : *
8246 : * The return value is either the end+1 address of the switch record,
8247 : * or the end+1 address of the prior segment if we did not need to
8248 : * write a switch record because we are already at segment start.
8249 : */
8250 : XLogRecPtr
8251 1464 : RequestXLogSwitch(bool mark_unimportant)
8252 : {
8253 : XLogRecPtr RecPtr;
8254 :
8255 : /* XLOG SWITCH has no data */
8256 1464 : XLogBeginInsert();
8257 :
8258 1464 : if (mark_unimportant)
8259 0 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8260 1464 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH);
8261 :
8262 1464 : return RecPtr;
8263 : }
8264 :
8265 : /*
8266 : * Write a RESTORE POINT record
8267 : */
8268 : XLogRecPtr
8269 6 : XLogRestorePoint(const char *rpName)
8270 : {
8271 : XLogRecPtr RecPtr;
8272 : xl_restore_point xlrec;
8273 :
8274 6 : xlrec.rp_time = GetCurrentTimestamp();
8275 6 : strlcpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8276 :
8277 6 : XLogBeginInsert();
8278 6 : XLogRegisterData(&xlrec, sizeof(xl_restore_point));
8279 :
8280 6 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT);
8281 :
8282 6 : ereport(LOG,
8283 : errmsg("restore point \"%s\" created at %X/%08X",
8284 : rpName, LSN_FORMAT_ARGS(RecPtr)));
8285 :
8286 6 : return RecPtr;
8287 : }
8288 :
8289 : /*
8290 : * Check if any of the GUC parameters that are critical for hot standby
8291 : * have changed, and update the value in pg_control file if necessary.
8292 : */
8293 : static void
8294 1728 : XLogReportParameters(void)
8295 : {
8296 1728 : if (wal_level != ControlFile->wal_level ||
8297 1252 : wal_log_hints != ControlFile->wal_log_hints ||
8298 1090 : MaxConnections != ControlFile->MaxConnections ||
8299 1088 : max_worker_processes != ControlFile->max_worker_processes ||
8300 1086 : max_wal_senders != ControlFile->max_wal_senders ||
8301 1040 : max_prepared_xacts != ControlFile->max_prepared_xacts ||
8302 860 : max_locks_per_xact != ControlFile->max_locks_per_xact ||
8303 860 : track_commit_timestamp != ControlFile->track_commit_timestamp)
8304 : {
8305 : /*
8306 : * The change in number of backend slots doesn't need to be WAL-logged
8307 : * if archiving is not enabled, as you can't start archive recovery
8308 : * with wal_level=minimal anyway. We don't really care about the
8309 : * values in pg_control either if wal_level=minimal, but seems better
8310 : * to keep them up-to-date to avoid confusion.
8311 : */
8312 888 : if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8313 : {
8314 : xl_parameter_change xlrec;
8315 : XLogRecPtr recptr;
8316 :
8317 844 : xlrec.MaxConnections = MaxConnections;
8318 844 : xlrec.max_worker_processes = max_worker_processes;
8319 844 : xlrec.max_wal_senders = max_wal_senders;
8320 844 : xlrec.max_prepared_xacts = max_prepared_xacts;
8321 844 : xlrec.max_locks_per_xact = max_locks_per_xact;
8322 844 : xlrec.wal_level = wal_level;
8323 844 : xlrec.wal_log_hints = wal_log_hints;
8324 844 : xlrec.track_commit_timestamp = track_commit_timestamp;
8325 :
8326 844 : XLogBeginInsert();
8327 844 : XLogRegisterData(&xlrec, sizeof(xlrec));
8328 :
8329 844 : recptr = XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE);
8330 844 : XLogFlush(recptr);
8331 : }
8332 :
8333 888 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8334 :
8335 888 : ControlFile->MaxConnections = MaxConnections;
8336 888 : ControlFile->max_worker_processes = max_worker_processes;
8337 888 : ControlFile->max_wal_senders = max_wal_senders;
8338 888 : ControlFile->max_prepared_xacts = max_prepared_xacts;
8339 888 : ControlFile->max_locks_per_xact = max_locks_per_xact;
8340 888 : ControlFile->wal_level = wal_level;
8341 888 : ControlFile->wal_log_hints = wal_log_hints;
8342 888 : ControlFile->track_commit_timestamp = track_commit_timestamp;
8343 888 : UpdateControlFile();
8344 :
8345 888 : LWLockRelease(ControlFileLock);
8346 : }
8347 1728 : }
8348 :
8349 : /*
8350 : * Update full_page_writes in shared memory, and write an
8351 : * XLOG_FPW_CHANGE record if necessary.
8352 : *
8353 : * Note: this function assumes there is no other process running
8354 : * concurrently that could update it.
8355 : */
8356 : void
8357 2870 : UpdateFullPageWrites(void)
8358 : {
8359 2870 : XLogCtlInsert *Insert = &XLogCtl->Insert;
8360 : bool recoveryInProgress;
8361 :
8362 : /*
8363 : * Do nothing if full_page_writes has not been changed.
8364 : *
8365 : * It's safe to check the shared full_page_writes without the lock,
8366 : * because we assume that there is no concurrently running process which
8367 : * can update it.
8368 : */
8369 2870 : if (fullPageWrites == Insert->fullPageWrites)
8370 2180 : return;
8371 :
8372 : /*
8373 : * Perform this outside critical section so that the WAL insert
8374 : * initialization done by RecoveryInProgress() doesn't trigger an
8375 : * assertion failure.
8376 : */
8377 690 : recoveryInProgress = RecoveryInProgress();
8378 :
8379 690 : START_CRIT_SECTION();
8380 :
8381 : /*
8382 : * It's always safe to take full page images, even when not strictly
8383 : * required, but not the other round. So if we're setting full_page_writes
8384 : * to true, first set it true and then write the WAL record. If we're
8385 : * setting it to false, first write the WAL record and then set the global
8386 : * flag.
8387 : */
8388 690 : if (fullPageWrites)
8389 : {
8390 670 : WALInsertLockAcquireExclusive();
8391 670 : Insert->fullPageWrites = true;
8392 670 : WALInsertLockRelease();
8393 : }
8394 :
8395 : /*
8396 : * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8397 : * full_page_writes during archive recovery, if required.
8398 : */
8399 690 : if (XLogStandbyInfoActive() && !recoveryInProgress)
8400 : {
8401 0 : XLogBeginInsert();
8402 0 : XLogRegisterData(&fullPageWrites, sizeof(bool));
8403 :
8404 0 : XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
8405 : }
8406 :
8407 690 : if (!fullPageWrites)
8408 : {
8409 20 : WALInsertLockAcquireExclusive();
8410 20 : Insert->fullPageWrites = false;
8411 20 : WALInsertLockRelease();
8412 : }
8413 690 : END_CRIT_SECTION();
8414 : }
8415 :
8416 : /*
8417 : * XLOG resource manager's routines
8418 : *
8419 : * Definitions of info values are in include/catalog/pg_control.h, though
8420 : * not all record types are related to control file updates.
8421 : *
8422 : * NOTE: Some XLOG record types that are directly related to WAL recovery
8423 : * are handled in xlogrecovery_redo().
8424 : */
8425 : void
8426 87178 : xlog_redo(XLogReaderState *record)
8427 : {
8428 87178 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8429 87178 : XLogRecPtr lsn = record->EndRecPtr;
8430 :
8431 : /*
8432 : * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
8433 : * XLOG_FPI_FOR_HINT records.
8434 : */
8435 : Assert(info == XLOG_FPI || info == XLOG_FPI_FOR_HINT ||
8436 : !XLogRecHasAnyBlockRefs(record));
8437 :
8438 87178 : if (info == XLOG_NEXTOID)
8439 : {
8440 : Oid nextOid;
8441 :
8442 : /*
8443 : * We used to try to take the maximum of TransamVariables->nextOid and
8444 : * the recorded nextOid, but that fails if the OID counter wraps
8445 : * around. Since no OID allocation should be happening during replay
8446 : * anyway, better to just believe the record exactly. We still take
8447 : * OidGenLock while setting the variable, just in case.
8448 : */
8449 186 : memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
8450 186 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8451 186 : TransamVariables->nextOid = nextOid;
8452 186 : TransamVariables->oidCount = 0;
8453 186 : LWLockRelease(OidGenLock);
8454 : }
8455 86992 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
8456 : {
8457 : CheckPoint checkPoint;
8458 : TimeLineID replayTLI;
8459 :
8460 68 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8461 : /* In a SHUTDOWN checkpoint, believe the counters exactly */
8462 68 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8463 68 : TransamVariables->nextXid = checkPoint.nextXid;
8464 68 : LWLockRelease(XidGenLock);
8465 68 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8466 68 : TransamVariables->nextOid = checkPoint.nextOid;
8467 68 : TransamVariables->oidCount = 0;
8468 68 : LWLockRelease(OidGenLock);
8469 68 : MultiXactSetNextMXact(checkPoint.nextMulti,
8470 : checkPoint.nextMultiOffset);
8471 :
8472 68 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8473 : checkPoint.oldestMultiDB);
8474 :
8475 : /*
8476 : * No need to set oldestClogXid here as well; it'll be set when we
8477 : * redo an xl_clog_truncate if it changed since initialization.
8478 : */
8479 68 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
8480 :
8481 : /*
8482 : * If we see a shutdown checkpoint while waiting for an end-of-backup
8483 : * record, the backup was canceled and the end-of-backup record will
8484 : * never arrive.
8485 : */
8486 68 : if (ArchiveRecoveryRequested &&
8487 66 : !XLogRecPtrIsInvalid(ControlFile->backupStartPoint) &&
8488 0 : XLogRecPtrIsInvalid(ControlFile->backupEndPoint))
8489 0 : ereport(PANIC,
8490 : (errmsg("online backup was canceled, recovery cannot continue")));
8491 :
8492 : /*
8493 : * If we see a shutdown checkpoint, we know that nothing was running
8494 : * on the primary at this point. So fake-up an empty running-xacts
8495 : * record and use that here and now. Recover additional standby state
8496 : * for prepared transactions.
8497 : */
8498 68 : if (standbyState >= STANDBY_INITIALIZED)
8499 : {
8500 : TransactionId *xids;
8501 : int nxids;
8502 : TransactionId oldestActiveXID;
8503 : TransactionId latestCompletedXid;
8504 : RunningTransactionsData running;
8505 :
8506 62 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
8507 :
8508 : /* Update pg_subtrans entries for any prepared transactions */
8509 62 : StandbyRecoverPreparedTransactions();
8510 :
8511 : /*
8512 : * Construct a RunningTransactions snapshot representing a shut
8513 : * down server, with only prepared transactions still alive. We're
8514 : * never overflowed at this point because all subxids are listed
8515 : * with their parent prepared transactions.
8516 : */
8517 62 : running.xcnt = nxids;
8518 62 : running.subxcnt = 0;
8519 62 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
8520 62 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
8521 62 : running.oldestRunningXid = oldestActiveXID;
8522 62 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
8523 62 : TransactionIdRetreat(latestCompletedXid);
8524 : Assert(TransactionIdIsNormal(latestCompletedXid));
8525 62 : running.latestCompletedXid = latestCompletedXid;
8526 62 : running.xids = xids;
8527 :
8528 62 : ProcArrayApplyRecoveryInfo(&running);
8529 : }
8530 :
8531 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8532 68 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8533 68 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8534 68 : LWLockRelease(ControlFileLock);
8535 :
8536 : /* Update shared-memory copy of checkpoint XID/epoch */
8537 68 : SpinLockAcquire(&XLogCtl->info_lck);
8538 68 : XLogCtl->ckptFullXid = checkPoint.nextXid;
8539 68 : SpinLockRelease(&XLogCtl->info_lck);
8540 :
8541 : /*
8542 : * We should've already switched to the new TLI before replaying this
8543 : * record.
8544 : */
8545 68 : (void) GetCurrentReplayRecPtr(&replayTLI);
8546 68 : if (checkPoint.ThisTimeLineID != replayTLI)
8547 0 : ereport(PANIC,
8548 : (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
8549 : checkPoint.ThisTimeLineID, replayTLI)));
8550 :
8551 68 : RecoveryRestartPoint(&checkPoint, record);
8552 : }
8553 86924 : else if (info == XLOG_CHECKPOINT_ONLINE)
8554 : {
8555 : CheckPoint checkPoint;
8556 : TimeLineID replayTLI;
8557 :
8558 1346 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8559 : /* In an ONLINE checkpoint, treat the XID counter as a minimum */
8560 1346 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8561 1346 : if (FullTransactionIdPrecedes(TransamVariables->nextXid,
8562 : checkPoint.nextXid))
8563 0 : TransamVariables->nextXid = checkPoint.nextXid;
8564 1346 : LWLockRelease(XidGenLock);
8565 :
8566 : /*
8567 : * We ignore the nextOid counter in an ONLINE checkpoint, preferring
8568 : * to track OID assignment through XLOG_NEXTOID records. The nextOid
8569 : * counter is from the start of the checkpoint and might well be stale
8570 : * compared to later XLOG_NEXTOID records. We could try to take the
8571 : * maximum of the nextOid counter and our latest value, but since
8572 : * there's no particular guarantee about the speed with which the OID
8573 : * counter wraps around, that's a risky thing to do. In any case,
8574 : * users of the nextOid counter are required to avoid assignment of
8575 : * duplicates, so that a somewhat out-of-date value should be safe.
8576 : */
8577 :
8578 : /* Handle multixact */
8579 1346 : MultiXactAdvanceNextMXact(checkPoint.nextMulti,
8580 : checkPoint.nextMultiOffset);
8581 :
8582 : /*
8583 : * NB: This may perform multixact truncation when replaying WAL
8584 : * generated by an older primary.
8585 : */
8586 1346 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8587 : checkPoint.oldestMultiDB);
8588 1346 : if (TransactionIdPrecedes(TransamVariables->oldestXid,
8589 : checkPoint.oldestXid))
8590 0 : SetTransactionIdLimit(checkPoint.oldestXid,
8591 : checkPoint.oldestXidDB);
8592 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8593 1346 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8594 1346 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8595 1346 : LWLockRelease(ControlFileLock);
8596 :
8597 : /* Update shared-memory copy of checkpoint XID/epoch */
8598 1346 : SpinLockAcquire(&XLogCtl->info_lck);
8599 1346 : XLogCtl->ckptFullXid = checkPoint.nextXid;
8600 1346 : SpinLockRelease(&XLogCtl->info_lck);
8601 :
8602 : /* TLI should not change in an on-line checkpoint */
8603 1346 : (void) GetCurrentReplayRecPtr(&replayTLI);
8604 1346 : if (checkPoint.ThisTimeLineID != replayTLI)
8605 0 : ereport(PANIC,
8606 : (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
8607 : checkPoint.ThisTimeLineID, replayTLI)));
8608 :
8609 1346 : RecoveryRestartPoint(&checkPoint, record);
8610 : }
8611 85578 : else if (info == XLOG_OVERWRITE_CONTRECORD)
8612 : {
8613 : /* nothing to do here, handled in xlogrecovery_redo() */
8614 : }
8615 85576 : else if (info == XLOG_END_OF_RECOVERY)
8616 : {
8617 : xl_end_of_recovery xlrec;
8618 : TimeLineID replayTLI;
8619 :
8620 20 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
8621 :
8622 : /*
8623 : * For Hot Standby, we could treat this like a Shutdown Checkpoint,
8624 : * but this case is rarer and harder to test, so the benefit doesn't
8625 : * outweigh the potential extra cost of maintenance.
8626 : */
8627 :
8628 : /*
8629 : * We should've already switched to the new TLI before replaying this
8630 : * record.
8631 : */
8632 20 : (void) GetCurrentReplayRecPtr(&replayTLI);
8633 20 : if (xlrec.ThisTimeLineID != replayTLI)
8634 0 : ereport(PANIC,
8635 : (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
8636 : xlrec.ThisTimeLineID, replayTLI)));
8637 : }
8638 85556 : else if (info == XLOG_NOOP)
8639 : {
8640 : /* nothing to do here */
8641 : }
8642 85556 : else if (info == XLOG_SWITCH)
8643 : {
8644 : /* nothing to do here */
8645 : }
8646 84672 : else if (info == XLOG_RESTORE_POINT)
8647 : {
8648 : /* nothing to do here, handled in xlogrecovery.c */
8649 : }
8650 84662 : else if (info == XLOG_FPI || info == XLOG_FPI_FOR_HINT)
8651 : {
8652 : /*
8653 : * XLOG_FPI records contain nothing else but one or more block
8654 : * references. Every block reference must include a full-page image
8655 : * even if full_page_writes was disabled when the record was generated
8656 : * - otherwise there would be no point in this record.
8657 : *
8658 : * XLOG_FPI_FOR_HINT records are generated when a page needs to be
8659 : * WAL-logged because of a hint bit update. They are only generated
8660 : * when checksums and/or wal_log_hints are enabled. They may include
8661 : * no full-page images if full_page_writes was disabled when they were
8662 : * generated. In this case there is nothing to do here.
8663 : *
8664 : * No recovery conflicts are generated by these generic records - if a
8665 : * resource manager needs to generate conflicts, it has to define a
8666 : * separate WAL record type and redo routine.
8667 : */
8668 175794 : for (uint8 block_id = 0; block_id <= XLogRecMaxBlockId(record); block_id++)
8669 : {
8670 : Buffer buffer;
8671 :
8672 92714 : if (!XLogRecHasBlockImage(record, block_id))
8673 : {
8674 134 : if (info == XLOG_FPI)
8675 0 : elog(ERROR, "XLOG_FPI record did not contain a full-page image");
8676 134 : continue;
8677 : }
8678 :
8679 92580 : if (XLogReadBufferForRedo(record, block_id, &buffer) != BLK_RESTORED)
8680 0 : elog(ERROR, "unexpected XLogReadBufferForRedo result when restoring backup block");
8681 92580 : UnlockReleaseBuffer(buffer);
8682 : }
8683 : }
8684 1582 : else if (info == XLOG_BACKUP_END)
8685 : {
8686 : /* nothing to do here, handled in xlogrecovery_redo() */
8687 : }
8688 1412 : else if (info == XLOG_PARAMETER_CHANGE)
8689 : {
8690 : xl_parameter_change xlrec;
8691 :
8692 : /* Update our copy of the parameters in pg_control */
8693 64 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
8694 :
8695 : /*
8696 : * Invalidate logical slots if we are in hot standby and the primary
8697 : * does not have a WAL level sufficient for logical decoding. No need
8698 : * to search for potentially conflicting logically slots if standby is
8699 : * running with wal_level lower than logical, because in that case, we
8700 : * would have either disallowed creation of logical slots or
8701 : * invalidated existing ones.
8702 : */
8703 64 : if (InRecovery && InHotStandby &&
8704 34 : xlrec.wal_level < WAL_LEVEL_LOGICAL &&
8705 12 : wal_level >= WAL_LEVEL_LOGICAL)
8706 8 : InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_LEVEL,
8707 : 0, InvalidOid,
8708 : InvalidTransactionId);
8709 :
8710 64 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8711 64 : ControlFile->MaxConnections = xlrec.MaxConnections;
8712 64 : ControlFile->max_worker_processes = xlrec.max_worker_processes;
8713 64 : ControlFile->max_wal_senders = xlrec.max_wal_senders;
8714 64 : ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
8715 64 : ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
8716 64 : ControlFile->wal_level = xlrec.wal_level;
8717 64 : ControlFile->wal_log_hints = xlrec.wal_log_hints;
8718 :
8719 : /*
8720 : * Update minRecoveryPoint to ensure that if recovery is aborted, we
8721 : * recover back up to this point before allowing hot standby again.
8722 : * This is important if the max_* settings are decreased, to ensure
8723 : * you don't run queries against the WAL preceding the change. The
8724 : * local copies cannot be updated as long as crash recovery is
8725 : * happening and we expect all the WAL to be replayed.
8726 : */
8727 64 : if (InArchiveRecovery)
8728 : {
8729 36 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
8730 36 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
8731 : }
8732 64 : if (LocalMinRecoveryPoint != InvalidXLogRecPtr && LocalMinRecoveryPoint < lsn)
8733 : {
8734 : TimeLineID replayTLI;
8735 :
8736 14 : (void) GetCurrentReplayRecPtr(&replayTLI);
8737 14 : ControlFile->minRecoveryPoint = lsn;
8738 14 : ControlFile->minRecoveryPointTLI = replayTLI;
8739 : }
8740 :
8741 64 : CommitTsParameterChange(xlrec.track_commit_timestamp,
8742 64 : ControlFile->track_commit_timestamp);
8743 64 : ControlFile->track_commit_timestamp = xlrec.track_commit_timestamp;
8744 :
8745 64 : UpdateControlFile();
8746 64 : LWLockRelease(ControlFileLock);
8747 :
8748 : /* Check to see if any parameter change gives a problem on recovery */
8749 64 : CheckRequiredParameterValues();
8750 : }
8751 1348 : else if (info == XLOG_FPW_CHANGE)
8752 : {
8753 : bool fpw;
8754 :
8755 0 : memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
8756 :
8757 : /*
8758 : * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
8759 : * do_pg_backup_start() and do_pg_backup_stop() can check whether
8760 : * full_page_writes has been disabled during online backup.
8761 : */
8762 0 : if (!fpw)
8763 : {
8764 0 : SpinLockAcquire(&XLogCtl->info_lck);
8765 0 : if (XLogCtl->lastFpwDisableRecPtr < record->ReadRecPtr)
8766 0 : XLogCtl->lastFpwDisableRecPtr = record->ReadRecPtr;
8767 0 : SpinLockRelease(&XLogCtl->info_lck);
8768 : }
8769 :
8770 : /* Keep track of full_page_writes */
8771 0 : lastFullPageWrites = fpw;
8772 : }
8773 : else if (info == XLOG_CHECKPOINT_REDO)
8774 : {
8775 : /* nothing to do here, just for informational purposes */
8776 : }
8777 87174 : }
8778 :
8779 : /*
8780 : * Return the extra open flags used for opening a file, depending on the
8781 : * value of the GUCs wal_sync_method, fsync and debug_io_direct.
8782 : */
8783 : static int
8784 31428 : get_sync_bit(int method)
8785 : {
8786 31428 : int o_direct_flag = 0;
8787 :
8788 : /*
8789 : * Use O_DIRECT if requested, except in walreceiver process. The WAL
8790 : * written by walreceiver is normally read by the startup process soon
8791 : * after it's written. Also, walreceiver performs unaligned writes, which
8792 : * don't work with O_DIRECT, so it is required for correctness too.
8793 : */
8794 31428 : if ((io_direct_flags & IO_DIRECT_WAL) && !AmWalReceiverProcess())
8795 16 : o_direct_flag = PG_O_DIRECT;
8796 :
8797 : /* If fsync is disabled, never open in sync mode */
8798 31428 : if (!enableFsync)
8799 31428 : return o_direct_flag;
8800 :
8801 0 : switch (method)
8802 : {
8803 : /*
8804 : * enum values for all sync options are defined even if they are
8805 : * not supported on the current platform. But if not, they are
8806 : * not included in the enum option array, and therefore will never
8807 : * be seen here.
8808 : */
8809 0 : case WAL_SYNC_METHOD_FSYNC:
8810 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8811 : case WAL_SYNC_METHOD_FDATASYNC:
8812 0 : return o_direct_flag;
8813 : #ifdef O_SYNC
8814 0 : case WAL_SYNC_METHOD_OPEN:
8815 0 : return O_SYNC | o_direct_flag;
8816 : #endif
8817 : #ifdef O_DSYNC
8818 0 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8819 0 : return O_DSYNC | o_direct_flag;
8820 : #endif
8821 0 : default:
8822 : /* can't happen (unless we are out of sync with option array) */
8823 0 : elog(ERROR, "unrecognized \"wal_sync_method\": %d", method);
8824 : return 0; /* silence warning */
8825 : }
8826 : }
8827 :
8828 : /*
8829 : * GUC support
8830 : */
8831 : void
8832 2208 : assign_wal_sync_method(int new_wal_sync_method, void *extra)
8833 : {
8834 2208 : if (wal_sync_method != new_wal_sync_method)
8835 : {
8836 : /*
8837 : * To ensure that no blocks escape unsynced, force an fsync on the
8838 : * currently open log segment (if any). Also, if the open flag is
8839 : * changing, close the log file so it will be reopened (with new flag
8840 : * bit) at next use.
8841 : */
8842 0 : if (openLogFile >= 0)
8843 : {
8844 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN);
8845 0 : if (pg_fsync(openLogFile) != 0)
8846 : {
8847 : char xlogfname[MAXFNAMELEN];
8848 : int save_errno;
8849 :
8850 0 : save_errno = errno;
8851 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo,
8852 : wal_segment_size);
8853 0 : errno = save_errno;
8854 0 : ereport(PANIC,
8855 : (errcode_for_file_access(),
8856 : errmsg("could not fsync file \"%s\": %m", xlogfname)));
8857 : }
8858 :
8859 0 : pgstat_report_wait_end();
8860 0 : if (get_sync_bit(wal_sync_method) != get_sync_bit(new_wal_sync_method))
8861 0 : XLogFileClose();
8862 : }
8863 : }
8864 2208 : }
8865 :
8866 :
8867 : /*
8868 : * Issue appropriate kind of fsync (if any) for an XLOG output file.
8869 : *
8870 : * 'fd' is a file descriptor for the XLOG file to be fsync'd.
8871 : * 'segno' is for error reporting purposes.
8872 : */
8873 : void
8874 316628 : issue_xlog_fsync(int fd, XLogSegNo segno, TimeLineID tli)
8875 : {
8876 316628 : char *msg = NULL;
8877 : instr_time start;
8878 :
8879 : Assert(tli != 0);
8880 :
8881 : /*
8882 : * Quick exit if fsync is disabled or write() has already synced the WAL
8883 : * file.
8884 : */
8885 316628 : if (!enableFsync ||
8886 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN ||
8887 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN_DSYNC)
8888 316628 : return;
8889 :
8890 : /*
8891 : * Measure I/O timing to sync the WAL file for pg_stat_io.
8892 : */
8893 0 : start = pgstat_prepare_io_time(track_wal_io_timing);
8894 :
8895 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC);
8896 0 : switch (wal_sync_method)
8897 : {
8898 0 : case WAL_SYNC_METHOD_FSYNC:
8899 0 : if (pg_fsync_no_writethrough(fd) != 0)
8900 0 : msg = _("could not fsync file \"%s\": %m");
8901 0 : break;
8902 : #ifdef HAVE_FSYNC_WRITETHROUGH
8903 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8904 : if (pg_fsync_writethrough(fd) != 0)
8905 : msg = _("could not fsync write-through file \"%s\": %m");
8906 : break;
8907 : #endif
8908 0 : case WAL_SYNC_METHOD_FDATASYNC:
8909 0 : if (pg_fdatasync(fd) != 0)
8910 0 : msg = _("could not fdatasync file \"%s\": %m");
8911 0 : break;
8912 0 : case WAL_SYNC_METHOD_OPEN:
8913 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8914 : /* not reachable */
8915 : Assert(false);
8916 0 : break;
8917 0 : default:
8918 0 : ereport(PANIC,
8919 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8920 : errmsg_internal("unrecognized \"wal_sync_method\": %d", wal_sync_method));
8921 : break;
8922 : }
8923 :
8924 : /* PANIC if failed to fsync */
8925 0 : if (msg)
8926 : {
8927 : char xlogfname[MAXFNAMELEN];
8928 0 : int save_errno = errno;
8929 :
8930 0 : XLogFileName(xlogfname, tli, segno, wal_segment_size);
8931 0 : errno = save_errno;
8932 0 : ereport(PANIC,
8933 : (errcode_for_file_access(),
8934 : errmsg(msg, xlogfname)));
8935 : }
8936 :
8937 0 : pgstat_report_wait_end();
8938 :
8939 0 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL, IOOP_FSYNC,
8940 : start, 1, 0);
8941 : }
8942 :
8943 : /*
8944 : * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
8945 : * function. It creates the necessary starting checkpoint and constructs the
8946 : * backup state and tablespace map.
8947 : *
8948 : * Input parameters are "state" (the backup state), "fast" (if true, we do
8949 : * the checkpoint in immediate mode to make it faster), and "tablespaces"
8950 : * (if non-NULL, indicates a list of tablespaceinfo structs describing the
8951 : * cluster's tablespaces.).
8952 : *
8953 : * The tablespace map contents are appended to passed-in parameter
8954 : * tablespace_map and the caller is responsible for including it in the backup
8955 : * archive as 'tablespace_map'. The tablespace_map file is required mainly for
8956 : * tar format in windows as native windows utilities are not able to create
8957 : * symlinks while extracting files from tar. However for consistency and
8958 : * platform-independence, we do it the same way everywhere.
8959 : *
8960 : * It fills in "state" with the information required for the backup, such
8961 : * as the minimum WAL location that must be present to restore from this
8962 : * backup (starttli) and the corresponding timeline ID (starttli).
8963 : *
8964 : * Every successfully started backup must be stopped by calling
8965 : * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
8966 : * backups active at the same time.
8967 : *
8968 : * It is the responsibility of the caller of this function to verify the
8969 : * permissions of the calling user!
8970 : */
8971 : void
8972 324 : do_pg_backup_start(const char *backupidstr, bool fast, List **tablespaces,
8973 : BackupState *state, StringInfo tblspcmapfile)
8974 : {
8975 : bool backup_started_in_recovery;
8976 :
8977 : Assert(state != NULL);
8978 324 : backup_started_in_recovery = RecoveryInProgress();
8979 :
8980 : /*
8981 : * During recovery, we don't need to check WAL level. Because, if WAL
8982 : * level is not sufficient, it's impossible to get here during recovery.
8983 : */
8984 324 : if (!backup_started_in_recovery && !XLogIsNeeded())
8985 0 : ereport(ERROR,
8986 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8987 : errmsg("WAL level not sufficient for making an online backup"),
8988 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
8989 :
8990 324 : if (strlen(backupidstr) > MAXPGPATH)
8991 2 : ereport(ERROR,
8992 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8993 : errmsg("backup label too long (max %d bytes)",
8994 : MAXPGPATH)));
8995 :
8996 322 : strlcpy(state->name, backupidstr, sizeof(state->name));
8997 :
8998 : /*
8999 : * Mark backup active in shared memory. We must do full-page WAL writes
9000 : * during an on-line backup even if not doing so at other times, because
9001 : * it's quite possible for the backup dump to obtain a "torn" (partially
9002 : * written) copy of a database page if it reads the page concurrently with
9003 : * our write to the same page. This can be fixed as long as the first
9004 : * write to the page in the WAL sequence is a full-page write. Hence, we
9005 : * increment runningBackups then force a CHECKPOINT, to ensure there are
9006 : * no dirty pages in shared memory that might get dumped while the backup
9007 : * is in progress without having a corresponding WAL record. (Once the
9008 : * backup is complete, we need not force full-page writes anymore, since
9009 : * we expect that any pages not modified during the backup interval must
9010 : * have been correctly captured by the backup.)
9011 : *
9012 : * Note that forcing full-page writes has no effect during an online
9013 : * backup from the standby.
9014 : *
9015 : * We must hold all the insertion locks to change the value of
9016 : * runningBackups, to ensure adequate interlocking against
9017 : * XLogInsertRecord().
9018 : */
9019 322 : WALInsertLockAcquireExclusive();
9020 322 : XLogCtl->Insert.runningBackups++;
9021 322 : WALInsertLockRelease();
9022 :
9023 : /*
9024 : * Ensure we decrement runningBackups if we fail below. NB -- for this to
9025 : * work correctly, it is critical that sessionBackupState is only updated
9026 : * after this block is over.
9027 : */
9028 322 : PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, DatumGetBool(true));
9029 : {
9030 322 : bool gotUniqueStartpoint = false;
9031 : DIR *tblspcdir;
9032 : struct dirent *de;
9033 : tablespaceinfo *ti;
9034 : int datadirpathlen;
9035 :
9036 : /*
9037 : * Force an XLOG file switch before the checkpoint, to ensure that the
9038 : * WAL segment the checkpoint is written to doesn't contain pages with
9039 : * old timeline IDs. That would otherwise happen if you called
9040 : * pg_backup_start() right after restoring from a PITR archive: the
9041 : * first WAL segment containing the startup checkpoint has pages in
9042 : * the beginning with the old timeline ID. That can cause trouble at
9043 : * recovery: we won't have a history file covering the old timeline if
9044 : * pg_wal directory was not included in the base backup and the WAL
9045 : * archive was cleared too before starting the backup.
9046 : *
9047 : * This also ensures that we have emitted a WAL page header that has
9048 : * XLP_BKP_REMOVABLE off before we emit the checkpoint record.
9049 : * Therefore, if a WAL archiver (such as pglesslog) is trying to
9050 : * compress out removable backup blocks, it won't remove any that
9051 : * occur after this point.
9052 : *
9053 : * During recovery, we skip forcing XLOG file switch, which means that
9054 : * the backup taken during recovery is not available for the special
9055 : * recovery case described above.
9056 : */
9057 322 : if (!backup_started_in_recovery)
9058 308 : RequestXLogSwitch(false);
9059 :
9060 : do
9061 : {
9062 : bool checkpointfpw;
9063 :
9064 : /*
9065 : * Force a CHECKPOINT. Aside from being necessary to prevent torn
9066 : * page problems, this guarantees that two successive backup runs
9067 : * will have different checkpoint positions and hence different
9068 : * history file names, even if nothing happened in between.
9069 : *
9070 : * During recovery, establish a restartpoint if possible. We use
9071 : * the last restartpoint as the backup starting checkpoint. This
9072 : * means that two successive backup runs can have same checkpoint
9073 : * positions.
9074 : *
9075 : * Since the fact that we are executing do_pg_backup_start()
9076 : * during recovery means that checkpointer is running, we can use
9077 : * RequestCheckpoint() to establish a restartpoint.
9078 : *
9079 : * We use CHECKPOINT_IMMEDIATE only if requested by user (via
9080 : * passing fast = true). Otherwise this can take awhile.
9081 : */
9082 322 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
9083 : (fast ? CHECKPOINT_IMMEDIATE : 0));
9084 :
9085 : /*
9086 : * Now we need to fetch the checkpoint record location, and also
9087 : * its REDO pointer. The oldest point in WAL that would be needed
9088 : * to restore starting from the checkpoint is precisely the REDO
9089 : * pointer.
9090 : */
9091 322 : LWLockAcquire(ControlFileLock, LW_SHARED);
9092 322 : state->checkpointloc = ControlFile->checkPoint;
9093 322 : state->startpoint = ControlFile->checkPointCopy.redo;
9094 322 : state->starttli = ControlFile->checkPointCopy.ThisTimeLineID;
9095 322 : checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
9096 322 : LWLockRelease(ControlFileLock);
9097 :
9098 322 : if (backup_started_in_recovery)
9099 : {
9100 : XLogRecPtr recptr;
9101 :
9102 : /*
9103 : * Check to see if all WAL replayed during online backup
9104 : * (i.e., since last restartpoint used as backup starting
9105 : * checkpoint) contain full-page writes.
9106 : */
9107 14 : SpinLockAcquire(&XLogCtl->info_lck);
9108 14 : recptr = XLogCtl->lastFpwDisableRecPtr;
9109 14 : SpinLockRelease(&XLogCtl->info_lck);
9110 :
9111 14 : if (!checkpointfpw || state->startpoint <= recptr)
9112 0 : ereport(ERROR,
9113 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9114 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9115 : "since last restartpoint"),
9116 : errhint("This means that the backup being taken on the standby "
9117 : "is corrupt and should not be used. "
9118 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9119 : "and then try an online backup again.")));
9120 :
9121 : /*
9122 : * During recovery, since we don't use the end-of-backup WAL
9123 : * record and don't write the backup history file, the
9124 : * starting WAL location doesn't need to be unique. This means
9125 : * that two base backups started at the same time might use
9126 : * the same checkpoint as starting locations.
9127 : */
9128 14 : gotUniqueStartpoint = true;
9129 : }
9130 :
9131 : /*
9132 : * If two base backups are started at the same time (in WAL sender
9133 : * processes), we need to make sure that they use different
9134 : * checkpoints as starting locations, because we use the starting
9135 : * WAL location as a unique identifier for the base backup in the
9136 : * end-of-backup WAL record and when we write the backup history
9137 : * file. Perhaps it would be better generate a separate unique ID
9138 : * for each backup instead of forcing another checkpoint, but
9139 : * taking a checkpoint right after another is not that expensive
9140 : * either because only few buffers have been dirtied yet.
9141 : */
9142 322 : WALInsertLockAcquireExclusive();
9143 322 : if (XLogCtl->Insert.lastBackupStart < state->startpoint)
9144 : {
9145 322 : XLogCtl->Insert.lastBackupStart = state->startpoint;
9146 322 : gotUniqueStartpoint = true;
9147 : }
9148 322 : WALInsertLockRelease();
9149 322 : } while (!gotUniqueStartpoint);
9150 :
9151 : /*
9152 : * Construct tablespace_map file.
9153 : */
9154 322 : datadirpathlen = strlen(DataDir);
9155 :
9156 : /* Collect information about all tablespaces */
9157 322 : tblspcdir = AllocateDir(PG_TBLSPC_DIR);
9158 1038 : while ((de = ReadDir(tblspcdir, PG_TBLSPC_DIR)) != NULL)
9159 : {
9160 : char fullpath[MAXPGPATH + sizeof(PG_TBLSPC_DIR)];
9161 : char linkpath[MAXPGPATH];
9162 716 : char *relpath = NULL;
9163 : char *s;
9164 : PGFileType de_type;
9165 : char *badp;
9166 : Oid tsoid;
9167 :
9168 : /*
9169 : * Try to parse the directory name as an unsigned integer.
9170 : *
9171 : * Tablespace directories should be positive integers that can be
9172 : * represented in 32 bits, with no leading zeroes or trailing
9173 : * garbage. If we come across a name that doesn't meet those
9174 : * criteria, skip it.
9175 : */
9176 716 : if (de->d_name[0] < '1' || de->d_name[1] > '9')
9177 644 : continue;
9178 72 : errno = 0;
9179 72 : tsoid = strtoul(de->d_name, &badp, 10);
9180 72 : if (*badp != '\0' || errno == EINVAL || errno == ERANGE)
9181 0 : continue;
9182 :
9183 72 : snprintf(fullpath, sizeof(fullpath), "%s/%s", PG_TBLSPC_DIR, de->d_name);
9184 :
9185 72 : de_type = get_dirent_type(fullpath, de, false, ERROR);
9186 :
9187 72 : if (de_type == PGFILETYPE_LNK)
9188 : {
9189 : StringInfoData escapedpath;
9190 : int rllen;
9191 :
9192 44 : rllen = readlink(fullpath, linkpath, sizeof(linkpath));
9193 44 : if (rllen < 0)
9194 : {
9195 0 : ereport(WARNING,
9196 : (errmsg("could not read symbolic link \"%s\": %m",
9197 : fullpath)));
9198 0 : continue;
9199 : }
9200 44 : else if (rllen >= sizeof(linkpath))
9201 : {
9202 0 : ereport(WARNING,
9203 : (errmsg("symbolic link \"%s\" target is too long",
9204 : fullpath)));
9205 0 : continue;
9206 : }
9207 44 : linkpath[rllen] = '\0';
9208 :
9209 : /*
9210 : * Relpath holds the relative path of the tablespace directory
9211 : * when it's located within PGDATA, or NULL if it's located
9212 : * elsewhere.
9213 : */
9214 44 : if (rllen > datadirpathlen &&
9215 2 : strncmp(linkpath, DataDir, datadirpathlen) == 0 &&
9216 0 : IS_DIR_SEP(linkpath[datadirpathlen]))
9217 0 : relpath = pstrdup(linkpath + datadirpathlen + 1);
9218 :
9219 : /*
9220 : * Add a backslash-escaped version of the link path to the
9221 : * tablespace map file.
9222 : */
9223 44 : initStringInfo(&escapedpath);
9224 1092 : for (s = linkpath; *s; s++)
9225 : {
9226 1048 : if (*s == '\n' || *s == '\r' || *s == '\\')
9227 0 : appendStringInfoChar(&escapedpath, '\\');
9228 1048 : appendStringInfoChar(&escapedpath, *s);
9229 : }
9230 44 : appendStringInfo(tblspcmapfile, "%s %s\n",
9231 44 : de->d_name, escapedpath.data);
9232 44 : pfree(escapedpath.data);
9233 : }
9234 28 : else if (de_type == PGFILETYPE_DIR)
9235 : {
9236 : /*
9237 : * It's possible to use allow_in_place_tablespaces to create
9238 : * directories directly under pg_tblspc, for testing purposes
9239 : * only.
9240 : *
9241 : * In this case, we store a relative path rather than an
9242 : * absolute path into the tablespaceinfo.
9243 : */
9244 28 : snprintf(linkpath, sizeof(linkpath), "%s/%s",
9245 28 : PG_TBLSPC_DIR, de->d_name);
9246 28 : relpath = pstrdup(linkpath);
9247 : }
9248 : else
9249 : {
9250 : /* Skip any other file type that appears here. */
9251 0 : continue;
9252 : }
9253 :
9254 72 : ti = palloc(sizeof(tablespaceinfo));
9255 72 : ti->oid = tsoid;
9256 72 : ti->path = pstrdup(linkpath);
9257 72 : ti->rpath = relpath;
9258 72 : ti->size = -1;
9259 :
9260 72 : if (tablespaces)
9261 72 : *tablespaces = lappend(*tablespaces, ti);
9262 : }
9263 322 : FreeDir(tblspcdir);
9264 :
9265 322 : state->starttime = (pg_time_t) time(NULL);
9266 : }
9267 322 : PG_END_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, DatumGetBool(true));
9268 :
9269 322 : state->started_in_recovery = backup_started_in_recovery;
9270 :
9271 : /*
9272 : * Mark that the start phase has correctly finished for the backup.
9273 : */
9274 322 : sessionBackupState = SESSION_BACKUP_RUNNING;
9275 322 : }
9276 :
9277 : /*
9278 : * Utility routine to fetch the session-level status of a backup running.
9279 : */
9280 : SessionBackupState
9281 364 : get_backup_status(void)
9282 : {
9283 364 : return sessionBackupState;
9284 : }
9285 :
9286 : /*
9287 : * do_pg_backup_stop
9288 : *
9289 : * Utility function called at the end of an online backup. It creates history
9290 : * file (if required), resets sessionBackupState and so on. It can optionally
9291 : * wait for WAL segments to be archived.
9292 : *
9293 : * "state" is filled with the information necessary to restore from this
9294 : * backup with its stop LSN (stoppoint), its timeline ID (stoptli), etc.
9295 : *
9296 : * It is the responsibility of the caller of this function to verify the
9297 : * permissions of the calling user!
9298 : */
9299 : void
9300 310 : do_pg_backup_stop(BackupState *state, bool waitforarchive)
9301 : {
9302 310 : bool backup_stopped_in_recovery = false;
9303 : char histfilepath[MAXPGPATH];
9304 : char lastxlogfilename[MAXFNAMELEN];
9305 : char histfilename[MAXFNAMELEN];
9306 : XLogSegNo _logSegNo;
9307 : FILE *fp;
9308 : int seconds_before_warning;
9309 310 : int waits = 0;
9310 310 : bool reported_waiting = false;
9311 :
9312 : Assert(state != NULL);
9313 :
9314 310 : backup_stopped_in_recovery = RecoveryInProgress();
9315 :
9316 : /*
9317 : * During recovery, we don't need to check WAL level. Because, if WAL
9318 : * level is not sufficient, it's impossible to get here during recovery.
9319 : */
9320 310 : if (!backup_stopped_in_recovery && !XLogIsNeeded())
9321 0 : ereport(ERROR,
9322 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9323 : errmsg("WAL level not sufficient for making an online backup"),
9324 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9325 :
9326 : /*
9327 : * OK to update backup counter and session-level lock.
9328 : *
9329 : * Note that CHECK_FOR_INTERRUPTS() must not occur while updating them,
9330 : * otherwise they can be updated inconsistently, which might cause
9331 : * do_pg_abort_backup() to fail.
9332 : */
9333 310 : WALInsertLockAcquireExclusive();
9334 :
9335 : /*
9336 : * It is expected that each do_pg_backup_start() call is matched by
9337 : * exactly one do_pg_backup_stop() call.
9338 : */
9339 : Assert(XLogCtl->Insert.runningBackups > 0);
9340 310 : XLogCtl->Insert.runningBackups--;
9341 :
9342 : /*
9343 : * Clean up session-level lock.
9344 : *
9345 : * You might think that WALInsertLockRelease() can be called before
9346 : * cleaning up session-level lock because session-level lock doesn't need
9347 : * to be protected with WAL insertion lock. But since
9348 : * CHECK_FOR_INTERRUPTS() can occur in it, session-level lock must be
9349 : * cleaned up before it.
9350 : */
9351 310 : sessionBackupState = SESSION_BACKUP_NONE;
9352 :
9353 310 : WALInsertLockRelease();
9354 :
9355 : /*
9356 : * If we are taking an online backup from the standby, we confirm that the
9357 : * standby has not been promoted during the backup.
9358 : */
9359 310 : if (state->started_in_recovery && !backup_stopped_in_recovery)
9360 0 : ereport(ERROR,
9361 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9362 : errmsg("the standby was promoted during online backup"),
9363 : errhint("This means that the backup being taken is corrupt "
9364 : "and should not be used. "
9365 : "Try taking another online backup.")));
9366 :
9367 : /*
9368 : * During recovery, we don't write an end-of-backup record. We assume that
9369 : * pg_control was backed up last and its minimum recovery point can be
9370 : * available as the backup end location. Since we don't have an
9371 : * end-of-backup record, we use the pg_control value to check whether
9372 : * we've reached the end of backup when starting recovery from this
9373 : * backup. We have no way of checking if pg_control wasn't backed up last
9374 : * however.
9375 : *
9376 : * We don't force a switch to new WAL file but it is still possible to
9377 : * wait for all the required files to be archived if waitforarchive is
9378 : * true. This is okay if we use the backup to start a standby and fetch
9379 : * the missing WAL using streaming replication. But in the case of an
9380 : * archive recovery, a user should set waitforarchive to true and wait for
9381 : * them to be archived to ensure that all the required files are
9382 : * available.
9383 : *
9384 : * We return the current minimum recovery point as the backup end
9385 : * location. Note that it can be greater than the exact backup end
9386 : * location if the minimum recovery point is updated after the backup of
9387 : * pg_control. This is harmless for current uses.
9388 : *
9389 : * XXX currently a backup history file is for informational and debug
9390 : * purposes only. It's not essential for an online backup. Furthermore,
9391 : * even if it's created, it will not be archived during recovery because
9392 : * an archiver is not invoked. So it doesn't seem worthwhile to write a
9393 : * backup history file during recovery.
9394 : */
9395 310 : if (backup_stopped_in_recovery)
9396 : {
9397 : XLogRecPtr recptr;
9398 :
9399 : /*
9400 : * Check to see if all WAL replayed during online backup contain
9401 : * full-page writes.
9402 : */
9403 14 : SpinLockAcquire(&XLogCtl->info_lck);
9404 14 : recptr = XLogCtl->lastFpwDisableRecPtr;
9405 14 : SpinLockRelease(&XLogCtl->info_lck);
9406 :
9407 14 : if (state->startpoint <= recptr)
9408 0 : ereport(ERROR,
9409 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9410 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9411 : "during online backup"),
9412 : errhint("This means that the backup being taken on the standby "
9413 : "is corrupt and should not be used. "
9414 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9415 : "and then try an online backup again.")));
9416 :
9417 :
9418 14 : LWLockAcquire(ControlFileLock, LW_SHARED);
9419 14 : state->stoppoint = ControlFile->minRecoveryPoint;
9420 14 : state->stoptli = ControlFile->minRecoveryPointTLI;
9421 14 : LWLockRelease(ControlFileLock);
9422 : }
9423 : else
9424 : {
9425 : char *history_file;
9426 :
9427 : /*
9428 : * Write the backup-end xlog record
9429 : */
9430 296 : XLogBeginInsert();
9431 296 : XLogRegisterData(&state->startpoint,
9432 : sizeof(state->startpoint));
9433 296 : state->stoppoint = XLogInsert(RM_XLOG_ID, XLOG_BACKUP_END);
9434 :
9435 : /*
9436 : * Given that we're not in recovery, InsertTimeLineID is set and can't
9437 : * change, so we can read it without a lock.
9438 : */
9439 296 : state->stoptli = XLogCtl->InsertTimeLineID;
9440 :
9441 : /*
9442 : * Force a switch to a new xlog segment file, so that the backup is
9443 : * valid as soon as archiver moves out the current segment file.
9444 : */
9445 296 : RequestXLogSwitch(false);
9446 :
9447 296 : state->stoptime = (pg_time_t) time(NULL);
9448 :
9449 : /*
9450 : * Write the backup history file
9451 : */
9452 296 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9453 296 : BackupHistoryFilePath(histfilepath, state->stoptli, _logSegNo,
9454 : state->startpoint, wal_segment_size);
9455 296 : fp = AllocateFile(histfilepath, "w");
9456 296 : if (!fp)
9457 0 : ereport(ERROR,
9458 : (errcode_for_file_access(),
9459 : errmsg("could not create file \"%s\": %m",
9460 : histfilepath)));
9461 :
9462 : /* Build and save the contents of the backup history file */
9463 296 : history_file = build_backup_content(state, true);
9464 296 : fprintf(fp, "%s", history_file);
9465 296 : pfree(history_file);
9466 :
9467 296 : if (fflush(fp) || ferror(fp) || FreeFile(fp))
9468 0 : ereport(ERROR,
9469 : (errcode_for_file_access(),
9470 : errmsg("could not write file \"%s\": %m",
9471 : histfilepath)));
9472 :
9473 : /*
9474 : * Clean out any no-longer-needed history files. As a side effect,
9475 : * this will post a .ready file for the newly created history file,
9476 : * notifying the archiver that history file may be archived
9477 : * immediately.
9478 : */
9479 296 : CleanupBackupHistory();
9480 : }
9481 :
9482 : /*
9483 : * If archiving is enabled, wait for all the required WAL files to be
9484 : * archived before returning. If archiving isn't enabled, the required WAL
9485 : * needs to be transported via streaming replication (hopefully with
9486 : * wal_keep_size set high enough), or some more exotic mechanism like
9487 : * polling and copying files from pg_wal with script. We have no knowledge
9488 : * of those mechanisms, so it's up to the user to ensure that he gets all
9489 : * the required WAL.
9490 : *
9491 : * We wait until both the last WAL file filled during backup and the
9492 : * history file have been archived, and assume that the alphabetic sorting
9493 : * property of the WAL files ensures any earlier WAL files are safely
9494 : * archived as well.
9495 : *
9496 : * We wait forever, since archive_command is supposed to work and we
9497 : * assume the admin wanted his backup to work completely. If you don't
9498 : * wish to wait, then either waitforarchive should be passed in as false,
9499 : * or you can set statement_timeout. Also, some notices are issued to
9500 : * clue in anyone who might be doing this interactively.
9501 : */
9502 :
9503 310 : if (waitforarchive &&
9504 20 : ((!backup_stopped_in_recovery && XLogArchivingActive()) ||
9505 2 : (backup_stopped_in_recovery && XLogArchivingAlways())))
9506 : {
9507 8 : XLByteToPrevSeg(state->stoppoint, _logSegNo, wal_segment_size);
9508 8 : XLogFileName(lastxlogfilename, state->stoptli, _logSegNo,
9509 : wal_segment_size);
9510 :
9511 8 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9512 8 : BackupHistoryFileName(histfilename, state->stoptli, _logSegNo,
9513 : state->startpoint, wal_segment_size);
9514 :
9515 8 : seconds_before_warning = 60;
9516 8 : waits = 0;
9517 :
9518 24 : while (XLogArchiveIsBusy(lastxlogfilename) ||
9519 8 : XLogArchiveIsBusy(histfilename))
9520 : {
9521 8 : CHECK_FOR_INTERRUPTS();
9522 :
9523 8 : if (!reported_waiting && waits > 5)
9524 : {
9525 0 : ereport(NOTICE,
9526 : (errmsg("base backup done, waiting for required WAL segments to be archived")));
9527 0 : reported_waiting = true;
9528 : }
9529 :
9530 8 : (void) WaitLatch(MyLatch,
9531 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
9532 : 1000L,
9533 : WAIT_EVENT_BACKUP_WAIT_WAL_ARCHIVE);
9534 8 : ResetLatch(MyLatch);
9535 :
9536 8 : if (++waits >= seconds_before_warning)
9537 : {
9538 0 : seconds_before_warning *= 2; /* This wraps in >10 years... */
9539 0 : ereport(WARNING,
9540 : (errmsg("still waiting for all required WAL segments to be archived (%d seconds elapsed)",
9541 : waits),
9542 : errhint("Check that your \"archive_command\" is executing properly. "
9543 : "You can safely cancel this backup, "
9544 : "but the database backup will not be usable without all the WAL segments.")));
9545 : }
9546 : }
9547 :
9548 8 : ereport(NOTICE,
9549 : (errmsg("all required WAL segments have been archived")));
9550 : }
9551 302 : else if (waitforarchive)
9552 12 : ereport(NOTICE,
9553 : (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
9554 310 : }
9555 :
9556 :
9557 : /*
9558 : * do_pg_abort_backup: abort a running backup
9559 : *
9560 : * This does just the most basic steps of do_pg_backup_stop(), by taking the
9561 : * system out of backup mode, thus making it a lot more safe to call from
9562 : * an error handler.
9563 : *
9564 : * 'arg' indicates that it's being called during backup setup; so
9565 : * sessionBackupState has not been modified yet, but runningBackups has
9566 : * already been incremented. When it's false, then it's invoked as a
9567 : * before_shmem_exit handler, and therefore we must not change state
9568 : * unless sessionBackupState indicates that a backup is actually running.
9569 : *
9570 : * NB: This gets used as a PG_ENSURE_ERROR_CLEANUP callback and
9571 : * before_shmem_exit handler, hence the odd-looking signature.
9572 : */
9573 : void
9574 16 : do_pg_abort_backup(int code, Datum arg)
9575 : {
9576 16 : bool during_backup_start = DatumGetBool(arg);
9577 :
9578 : /* If called during backup start, there shouldn't be one already running */
9579 : Assert(!during_backup_start || sessionBackupState == SESSION_BACKUP_NONE);
9580 :
9581 16 : if (during_backup_start || sessionBackupState != SESSION_BACKUP_NONE)
9582 : {
9583 12 : WALInsertLockAcquireExclusive();
9584 : Assert(XLogCtl->Insert.runningBackups > 0);
9585 12 : XLogCtl->Insert.runningBackups--;
9586 :
9587 12 : sessionBackupState = SESSION_BACKUP_NONE;
9588 12 : WALInsertLockRelease();
9589 :
9590 12 : if (!during_backup_start)
9591 12 : ereport(WARNING,
9592 : errmsg("aborting backup due to backend exiting before pg_backup_stop was called"));
9593 : }
9594 16 : }
9595 :
9596 : /*
9597 : * Register a handler that will warn about unterminated backups at end of
9598 : * session, unless this has already been done.
9599 : */
9600 : void
9601 8 : register_persistent_abort_backup_handler(void)
9602 : {
9603 : static bool already_done = false;
9604 :
9605 8 : if (already_done)
9606 2 : return;
9607 6 : before_shmem_exit(do_pg_abort_backup, DatumGetBool(false));
9608 6 : already_done = true;
9609 : }
9610 :
9611 : /*
9612 : * Get latest WAL insert pointer
9613 : */
9614 : XLogRecPtr
9615 3944 : GetXLogInsertRecPtr(void)
9616 : {
9617 3944 : XLogCtlInsert *Insert = &XLogCtl->Insert;
9618 : uint64 current_bytepos;
9619 :
9620 3944 : SpinLockAcquire(&Insert->insertpos_lck);
9621 3944 : current_bytepos = Insert->CurrBytePos;
9622 3944 : SpinLockRelease(&Insert->insertpos_lck);
9623 :
9624 3944 : return XLogBytePosToRecPtr(current_bytepos);
9625 : }
9626 :
9627 : /*
9628 : * Get latest WAL write pointer
9629 : */
9630 : XLogRecPtr
9631 2888 : GetXLogWriteRecPtr(void)
9632 : {
9633 2888 : RefreshXLogWriteResult(LogwrtResult);
9634 :
9635 2888 : return LogwrtResult.Write;
9636 : }
9637 :
9638 : /*
9639 : * Returns the redo pointer of the last checkpoint or restartpoint. This is
9640 : * the oldest point in WAL that we still need, if we have to restart recovery.
9641 : */
9642 : void
9643 796 : GetOldestRestartPoint(XLogRecPtr *oldrecptr, TimeLineID *oldtli)
9644 : {
9645 796 : LWLockAcquire(ControlFileLock, LW_SHARED);
9646 796 : *oldrecptr = ControlFile->checkPointCopy.redo;
9647 796 : *oldtli = ControlFile->checkPointCopy.ThisTimeLineID;
9648 796 : LWLockRelease(ControlFileLock);
9649 796 : }
9650 :
9651 : /* Thin wrapper around ShutdownWalRcv(). */
9652 : void
9653 2122 : XLogShutdownWalRcv(void)
9654 : {
9655 2122 : ShutdownWalRcv();
9656 :
9657 2122 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9658 2122 : XLogCtl->InstallXLogFileSegmentActive = false;
9659 2122 : LWLockRelease(ControlFileLock);
9660 2122 : }
9661 :
9662 : /* Enable WAL file recycling and preallocation. */
9663 : void
9664 2216 : SetInstallXLogFileSegmentActive(void)
9665 : {
9666 2216 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9667 2216 : XLogCtl->InstallXLogFileSegmentActive = true;
9668 2216 : LWLockRelease(ControlFileLock);
9669 2216 : }
9670 :
9671 : bool
9672 0 : IsInstallXLogFileSegmentActive(void)
9673 : {
9674 : bool result;
9675 :
9676 0 : LWLockAcquire(ControlFileLock, LW_SHARED);
9677 0 : result = XLogCtl->InstallXLogFileSegmentActive;
9678 0 : LWLockRelease(ControlFileLock);
9679 :
9680 0 : return result;
9681 : }
9682 :
9683 : /*
9684 : * Update the WalWriterSleeping flag.
9685 : */
9686 : void
9687 966 : SetWalWriterSleeping(bool sleeping)
9688 : {
9689 966 : SpinLockAcquire(&XLogCtl->info_lck);
9690 966 : XLogCtl->WalWriterSleeping = sleeping;
9691 966 : SpinLockRelease(&XLogCtl->info_lck);
9692 966 : }
|
