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 : XLogRecPtr asyncXactLSN; /* LSN of newest async commit/abort */
453 : XLogRecPtr replicationSlotMinLSN; /* oldest LSN needed by any slot */
454 :
455 : XLogSegNo lastRemovedSegNo; /* latest removed/recycled XLOG segment */
456 :
457 : /* Fake LSN counter, for unlogged relations. */
458 : pg_atomic_uint64 unloggedLSN;
459 :
460 : /* Time and LSN of last xlog segment switch. Protected by WALWriteLock. */
461 : pg_time_t lastSegSwitchTime;
462 : XLogRecPtr lastSegSwitchLSN;
463 :
464 : /* These are accessed using atomics -- info_lck not needed */
465 : pg_atomic_uint64 logInsertResult; /* last byte + 1 inserted to buffers */
466 : pg_atomic_uint64 logWriteResult; /* last byte + 1 written out */
467 : pg_atomic_uint64 logFlushResult; /* last byte + 1 flushed */
468 :
469 : /*
470 : * First initialized page in the cache (first byte position).
471 : */
472 : XLogRecPtr InitializedFrom;
473 :
474 : /*
475 : * Latest reserved for initialization page in the cache (last byte
476 : * position + 1).
477 : *
478 : * To change the identity of a buffer, you need to advance
479 : * InitializeReserved first. To change the identity of a buffer that's
480 : * still dirty, the old page needs to be written out first, and for that
481 : * you need WALWriteLock, and you need to ensure that there are no
482 : * in-progress insertions to the page by calling
483 : * WaitXLogInsertionsToFinish().
484 : */
485 : pg_atomic_uint64 InitializeReserved;
486 :
487 : /*
488 : * Latest initialized page in the cache (last byte position + 1).
489 : *
490 : * InitializedUpTo is updated after the buffer initialization. After
491 : * update, waiters got notification using InitializedUpToCondVar.
492 : */
493 : pg_atomic_uint64 InitializedUpTo;
494 : ConditionVariable InitializedUpToCondVar;
495 :
496 : /*
497 : * These values do not change after startup, although the pointed-to pages
498 : * and xlblocks values certainly do. xlblocks values are changed
499 : * lock-free according to the check for the xlog write position and are
500 : * accompanied by changes of InitializeReserved and InitializedUpTo.
501 : */
502 : char *pages; /* buffers for unwritten XLOG pages */
503 : pg_atomic_uint64 *xlblocks; /* 1st byte ptr-s + XLOG_BLCKSZ */
504 : int XLogCacheBlck; /* highest allocated xlog buffer index */
505 :
506 : /*
507 : * InsertTimeLineID is the timeline into which new WAL is being inserted
508 : * and flushed. It is zero during recovery, and does not change once set.
509 : *
510 : * If we create a new timeline when the system was started up,
511 : * PrevTimeLineID is the old timeline's ID that we forked off from.
512 : * Otherwise it's equal to InsertTimeLineID.
513 : *
514 : * We set these fields while holding info_lck. Most that reads these
515 : * values knows that recovery is no longer in progress and so can safely
516 : * read the value without a lock, but code that could be run either during
517 : * or after recovery can take info_lck while reading these values.
518 : */
519 : TimeLineID InsertTimeLineID;
520 : TimeLineID PrevTimeLineID;
521 :
522 : /*
523 : * SharedRecoveryState indicates if we're still in crash or archive
524 : * recovery. Protected by info_lck.
525 : */
526 : RecoveryState SharedRecoveryState;
527 :
528 : /*
529 : * InstallXLogFileSegmentActive indicates whether the checkpointer should
530 : * arrange for future segments by recycling and/or PreallocXlogFiles().
531 : * Protected by ControlFileLock. Only the startup process changes it. If
532 : * true, anyone can use InstallXLogFileSegment(). If false, the startup
533 : * process owns the exclusive right to install segments, by reading from
534 : * the archive and possibly replacing existing files.
535 : */
536 : bool InstallXLogFileSegmentActive;
537 :
538 : /*
539 : * WalWriterSleeping indicates whether the WAL writer is currently in
540 : * low-power mode (and hence should be nudged if an async commit occurs).
541 : * Protected by info_lck.
542 : */
543 : bool WalWriterSleeping;
544 :
545 : /*
546 : * During recovery, we keep a copy of the latest checkpoint record here.
547 : * lastCheckPointRecPtr points to start of checkpoint record and
548 : * lastCheckPointEndPtr points to end+1 of checkpoint record. Used by the
549 : * checkpointer when it wants to create a restartpoint.
550 : *
551 : * Protected by info_lck.
552 : */
553 : XLogRecPtr lastCheckPointRecPtr;
554 : XLogRecPtr lastCheckPointEndPtr;
555 : CheckPoint lastCheckPoint;
556 :
557 : /*
558 : * lastFpwDisableRecPtr points to the start of the last replayed
559 : * XLOG_FPW_CHANGE record that instructs full_page_writes is disabled.
560 : */
561 : XLogRecPtr lastFpwDisableRecPtr;
562 :
563 : slock_t info_lck; /* locks shared variables shown above */
564 : } XLogCtlData;
565 :
566 : /*
567 : * Classification of XLogInsertRecord operations.
568 : */
569 : typedef enum
570 : {
571 : WALINSERT_NORMAL,
572 : WALINSERT_SPECIAL_SWITCH,
573 : WALINSERT_SPECIAL_CHECKPOINT
574 : } WalInsertClass;
575 :
576 : static XLogCtlData *XLogCtl = NULL;
577 :
578 : /* a private copy of XLogCtl->Insert.WALInsertLocks, for convenience */
579 : static WALInsertLockPadded *WALInsertLocks = NULL;
580 :
581 : /*
582 : * We maintain an image of pg_control in shared memory.
583 : */
584 : static ControlFileData *ControlFile = NULL;
585 :
586 : /*
587 : * Calculate the amount of space left on the page after 'endptr'. Beware
588 : * multiple evaluation!
589 : */
590 : #define INSERT_FREESPACE(endptr) \
591 : (((endptr) % XLOG_BLCKSZ == 0) ? 0 : (XLOG_BLCKSZ - (endptr) % XLOG_BLCKSZ))
592 :
593 : /* Macro to advance to next buffer index. */
594 : #define NextBufIdx(idx) \
595 : (((idx) == XLogCtl->XLogCacheBlck) ? 0 : ((idx) + 1))
596 :
597 : /*
598 : * XLogRecPtrToBufIdx returns the index of the WAL buffer that holds, or
599 : * would hold if it was in cache, the page containing 'recptr'.
600 : */
601 : #define XLogRecPtrToBufIdx(recptr) \
602 : (((recptr) / XLOG_BLCKSZ) % (XLogCtl->XLogCacheBlck + 1))
603 :
604 : /*
605 : * These are the number of bytes in a WAL page usable for WAL data.
606 : */
607 : #define UsableBytesInPage (XLOG_BLCKSZ - SizeOfXLogShortPHD)
608 :
609 : /*
610 : * Convert values of GUCs measured in megabytes to equiv. segment count.
611 : * Rounds down.
612 : */
613 : #define ConvertToXSegs(x, segsize) XLogMBVarToSegs((x), (segsize))
614 :
615 : /* The number of bytes in a WAL segment usable for WAL data. */
616 : static int UsableBytesInSegment;
617 :
618 : /*
619 : * Private, possibly out-of-date copy of shared LogwrtResult.
620 : * See discussion above.
621 : */
622 : static XLogwrtResult LogwrtResult = {0, 0};
623 :
624 : /*
625 : * Update local copy of shared XLogCtl->log{Write,Flush}Result
626 : *
627 : * It's critical that Flush always trails Write, so the order of the reads is
628 : * important, as is the barrier. See also XLogWrite.
629 : */
630 : #define RefreshXLogWriteResult(_target) \
631 : do { \
632 : _target.Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult); \
633 : pg_read_barrier(); \
634 : _target.Write = pg_atomic_read_u64(&XLogCtl->logWriteResult); \
635 : } while (0)
636 :
637 : /*
638 : * openLogFile is -1 or a kernel FD for an open log file segment.
639 : * openLogSegNo identifies the segment, and openLogTLI the corresponding TLI.
640 : * These variables are only used to write the XLOG, and so will normally refer
641 : * to the active segment.
642 : *
643 : * Note: call Reserve/ReleaseExternalFD to track consumption of this FD.
644 : */
645 : static int openLogFile = -1;
646 : static XLogSegNo openLogSegNo = 0;
647 : static TimeLineID openLogTLI = 0;
648 :
649 : /*
650 : * Local copies of equivalent fields in the control file. When running
651 : * crash recovery, LocalMinRecoveryPoint is set to InvalidXLogRecPtr as we
652 : * expect to replay all the WAL available, and updateMinRecoveryPoint is
653 : * switched to false to prevent any updates while replaying records.
654 : * Those values are kept consistent as long as crash recovery runs.
655 : */
656 : static XLogRecPtr LocalMinRecoveryPoint;
657 : static TimeLineID LocalMinRecoveryPointTLI;
658 : static bool updateMinRecoveryPoint = true;
659 :
660 : /* For WALInsertLockAcquire/Release functions */
661 : static int MyLockNo = 0;
662 : static bool holdingAllLocks = false;
663 :
664 : #ifdef WAL_DEBUG
665 : static MemoryContext walDebugCxt = NULL;
666 : #endif
667 :
668 : static void CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI,
669 : XLogRecPtr EndOfLog,
670 : TimeLineID newTLI);
671 : static void CheckRequiredParameterValues(void);
672 : static void XLogReportParameters(void);
673 : static int LocalSetXLogInsertAllowed(void);
674 : static void CreateEndOfRecoveryRecord(void);
675 : static XLogRecPtr CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn,
676 : XLogRecPtr pagePtr,
677 : TimeLineID newTLI);
678 : static void CheckPointGuts(XLogRecPtr checkPointRedo, int flags);
679 : static void KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo);
680 : static XLogRecPtr XLogGetReplicationSlotMinimumLSN(void);
681 :
682 : static void AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli,
683 : bool opportunistic);
684 : static void XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible);
685 : static bool InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
686 : bool find_free, XLogSegNo max_segno,
687 : TimeLineID tli);
688 : static void XLogFileClose(void);
689 : static void PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli);
690 : static void RemoveTempXlogFiles(void);
691 : static void RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr,
692 : XLogRecPtr endptr, TimeLineID insertTLI);
693 : static void RemoveXlogFile(const struct dirent *segment_de,
694 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
695 : TimeLineID insertTLI);
696 : static void UpdateLastRemovedPtr(char *filename);
697 : static void ValidateXLOGDirectoryStructure(void);
698 : static void CleanupBackupHistory(void);
699 : static void UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force);
700 : static bool PerformRecoveryXLogAction(void);
701 : static void InitControlFile(uint64 sysidentifier, uint32 data_checksum_version);
702 : static void WriteControlFile(void);
703 : static void ReadControlFile(void);
704 : static void UpdateControlFile(void);
705 : static char *str_time(pg_time_t tnow);
706 :
707 : static int get_sync_bit(int method);
708 :
709 : static void CopyXLogRecordToWAL(int write_len, bool isLogSwitch,
710 : XLogRecData *rdata,
711 : XLogRecPtr StartPos, XLogRecPtr EndPos,
712 : TimeLineID tli);
713 : static void ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos,
714 : XLogRecPtr *EndPos, XLogRecPtr *PrevPtr);
715 : static bool ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos,
716 : XLogRecPtr *PrevPtr);
717 : static XLogRecPtr WaitXLogInsertionsToFinish(XLogRecPtr upto);
718 : static char *GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli);
719 : static XLogRecPtr XLogBytePosToRecPtr(uint64 bytepos);
720 : static XLogRecPtr XLogBytePosToEndRecPtr(uint64 bytepos);
721 : static uint64 XLogRecPtrToBytePos(XLogRecPtr ptr);
722 :
723 : static void WALInsertLockAcquire(void);
724 : static void WALInsertLockAcquireExclusive(void);
725 : static void WALInsertLockRelease(void);
726 : static void WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt);
727 :
728 : /*
729 : * Insert an XLOG record represented by an already-constructed chain of data
730 : * chunks. This is a low-level routine; to construct the WAL record header
731 : * and data, use the higher-level routines in xloginsert.c.
732 : *
733 : * If 'fpw_lsn' is valid, it is the oldest LSN among the pages that this
734 : * WAL record applies to, that were not included in the record as full page
735 : * images. If fpw_lsn <= RedoRecPtr, the function does not perform the
736 : * insertion and returns InvalidXLogRecPtr. The caller can then recalculate
737 : * which pages need a full-page image, and retry. If fpw_lsn is invalid, the
738 : * record is always inserted.
739 : *
740 : * 'flags' gives more in-depth control on the record being inserted. See
741 : * XLogSetRecordFlags() for details.
742 : *
743 : * 'topxid_included' tells whether the top-transaction id is logged along with
744 : * current subtransaction. See XLogRecordAssemble().
745 : *
746 : * The first XLogRecData in the chain must be for the record header, and its
747 : * data must be MAXALIGNed. XLogInsertRecord fills in the xl_prev and
748 : * xl_crc fields in the header, the rest of the header must already be filled
749 : * by the caller.
750 : *
751 : * Returns XLOG pointer to end of record (beginning of next record).
752 : * This can be used as LSN for data pages affected by the logged action.
753 : * (LSN is the XLOG point up to which the XLOG must be flushed to disk
754 : * before the data page can be written out. This implements the basic
755 : * WAL rule "write the log before the data".)
756 : */
757 : XLogRecPtr
758 29225442 : XLogInsertRecord(XLogRecData *rdata,
759 : XLogRecPtr fpw_lsn,
760 : uint8 flags,
761 : int num_fpi,
762 : bool topxid_included)
763 : {
764 29225442 : XLogCtlInsert *Insert = &XLogCtl->Insert;
765 : pg_crc32c rdata_crc;
766 : bool inserted;
767 29225442 : XLogRecord *rechdr = (XLogRecord *) rdata->data;
768 29225442 : uint8 info = rechdr->xl_info & ~XLR_INFO_MASK;
769 29225442 : WalInsertClass class = WALINSERT_NORMAL;
770 : XLogRecPtr StartPos;
771 : XLogRecPtr EndPos;
772 29225442 : bool prevDoPageWrites = doPageWrites;
773 : TimeLineID insertTLI;
774 :
775 : /* Does this record type require special handling? */
776 29225442 : if (unlikely(rechdr->xl_rmid == RM_XLOG_ID))
777 : {
778 444106 : if (info == XLOG_SWITCH)
779 1558 : class = WALINSERT_SPECIAL_SWITCH;
780 442548 : else if (info == XLOG_CHECKPOINT_REDO)
781 1804 : class = WALINSERT_SPECIAL_CHECKPOINT;
782 : }
783 :
784 : /* we assume that all of the record header is in the first chunk */
785 : Assert(rdata->len >= SizeOfXLogRecord);
786 :
787 : /* cross-check on whether we should be here or not */
788 29225442 : if (!XLogInsertAllowed())
789 0 : elog(ERROR, "cannot make new WAL entries during recovery");
790 :
791 : /*
792 : * Given that we're not in recovery, InsertTimeLineID is set and can't
793 : * change, so we can read it without a lock.
794 : */
795 29225442 : insertTLI = XLogCtl->InsertTimeLineID;
796 :
797 : /*----------
798 : *
799 : * We have now done all the preparatory work we can without holding a
800 : * lock or modifying shared state. From here on, inserting the new WAL
801 : * record to the shared WAL buffer cache is a two-step process:
802 : *
803 : * 1. Reserve the right amount of space from the WAL. The current head of
804 : * reserved space is kept in Insert->CurrBytePos, and is protected by
805 : * insertpos_lck.
806 : *
807 : * 2. Copy the record to the reserved WAL space. This involves finding the
808 : * correct WAL buffer containing the reserved space, and copying the
809 : * record in place. This can be done concurrently in multiple processes.
810 : *
811 : * To keep track of which insertions are still in-progress, each concurrent
812 : * inserter acquires an insertion lock. In addition to just indicating that
813 : * an insertion is in progress, the lock tells others how far the inserter
814 : * has progressed. There is a small fixed number of insertion locks,
815 : * determined by NUM_XLOGINSERT_LOCKS. When an inserter crosses a page
816 : * boundary, it updates the value stored in the lock to the how far it has
817 : * inserted, to allow the previous buffer to be flushed.
818 : *
819 : * Holding onto an insertion lock also protects RedoRecPtr and
820 : * fullPageWrites from changing until the insertion is finished.
821 : *
822 : * Step 2 can usually be done completely in parallel. If the required WAL
823 : * page is not initialized yet, you have to go through AdvanceXLInsertBuffer,
824 : * which will ensure it is initialized. But the WAL writer tries to do that
825 : * ahead of insertions to avoid that from happening in the critical path.
826 : *
827 : *----------
828 : */
829 29225442 : START_CRIT_SECTION();
830 :
831 29225442 : if (likely(class == WALINSERT_NORMAL))
832 : {
833 29222080 : WALInsertLockAcquire();
834 :
835 : /*
836 : * Check to see if my copy of RedoRecPtr is out of date. If so, may
837 : * have to go back and have the caller recompute everything. This can
838 : * only happen just after a checkpoint, so it's better to be slow in
839 : * this case and fast otherwise.
840 : *
841 : * Also check to see if fullPageWrites was just turned on or there's a
842 : * running backup (which forces full-page writes); if we weren't
843 : * already doing full-page writes then go back and recompute.
844 : *
845 : * If we aren't doing full-page writes then RedoRecPtr doesn't
846 : * actually affect the contents of the XLOG record, so we'll update
847 : * our local copy but not force a recomputation. (If doPageWrites was
848 : * just turned off, we could recompute the record without full pages,
849 : * but we choose not to bother.)
850 : */
851 29222080 : if (RedoRecPtr != Insert->RedoRecPtr)
852 : {
853 : Assert(RedoRecPtr < Insert->RedoRecPtr);
854 13650 : RedoRecPtr = Insert->RedoRecPtr;
855 : }
856 29222080 : doPageWrites = (Insert->fullPageWrites || Insert->runningBackups > 0);
857 :
858 29222080 : if (doPageWrites &&
859 28739996 : (!prevDoPageWrites ||
860 26377040 : (fpw_lsn != InvalidXLogRecPtr && fpw_lsn <= RedoRecPtr)))
861 : {
862 : /*
863 : * Oops, some buffer now needs to be backed up that the caller
864 : * didn't back up. Start over.
865 : */
866 15120 : WALInsertLockRelease();
867 15120 : END_CRIT_SECTION();
868 15120 : return InvalidXLogRecPtr;
869 : }
870 :
871 : /*
872 : * Reserve space for the record in the WAL. This also sets the xl_prev
873 : * pointer.
874 : */
875 29206960 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
876 : &rechdr->xl_prev);
877 :
878 : /* Normal records are always inserted. */
879 29206960 : inserted = true;
880 : }
881 3362 : else if (class == WALINSERT_SPECIAL_SWITCH)
882 : {
883 : /*
884 : * In order to insert an XLOG_SWITCH record, we need to hold all of
885 : * the WAL insertion locks, not just one, so that no one else can
886 : * begin inserting a record until we've figured out how much space
887 : * remains in the current WAL segment and claimed all of it.
888 : *
889 : * Nonetheless, this case is simpler than the normal cases handled
890 : * below, which must check for changes in doPageWrites and RedoRecPtr.
891 : * Those checks are only needed for records that can contain buffer
892 : * references, and an XLOG_SWITCH record never does.
893 : */
894 : Assert(fpw_lsn == InvalidXLogRecPtr);
895 1558 : WALInsertLockAcquireExclusive();
896 1558 : inserted = ReserveXLogSwitch(&StartPos, &EndPos, &rechdr->xl_prev);
897 : }
898 : else
899 : {
900 : Assert(class == WALINSERT_SPECIAL_CHECKPOINT);
901 :
902 : /*
903 : * We need to update both the local and shared copies of RedoRecPtr,
904 : * which means that we need to hold all the WAL insertion locks.
905 : * However, there can't be any buffer references, so as above, we need
906 : * not check RedoRecPtr before inserting the record; we just need to
907 : * update it afterwards.
908 : */
909 : Assert(fpw_lsn == InvalidXLogRecPtr);
910 1804 : WALInsertLockAcquireExclusive();
911 1804 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
912 : &rechdr->xl_prev);
913 1804 : RedoRecPtr = Insert->RedoRecPtr = StartPos;
914 1804 : inserted = true;
915 : }
916 :
917 29210322 : if (inserted)
918 : {
919 : /*
920 : * Now that xl_prev has been filled in, calculate CRC of the record
921 : * header.
922 : */
923 29210206 : rdata_crc = rechdr->xl_crc;
924 29210206 : COMP_CRC32C(rdata_crc, rechdr, offsetof(XLogRecord, xl_crc));
925 29210206 : FIN_CRC32C(rdata_crc);
926 29210206 : rechdr->xl_crc = rdata_crc;
927 :
928 : /*
929 : * All the record data, including the header, is now ready to be
930 : * inserted. Copy the record in the space reserved.
931 : */
932 29210206 : CopyXLogRecordToWAL(rechdr->xl_tot_len,
933 : class == WALINSERT_SPECIAL_SWITCH, rdata,
934 : StartPos, EndPos, insertTLI);
935 :
936 : /*
937 : * Unless record is flagged as not important, update LSN of last
938 : * important record in the current slot. When holding all locks, just
939 : * update the first one.
940 : */
941 29210206 : if ((flags & XLOG_MARK_UNIMPORTANT) == 0)
942 : {
943 29030418 : int lockno = holdingAllLocks ? 0 : MyLockNo;
944 :
945 29030418 : WALInsertLocks[lockno].l.lastImportantAt = StartPos;
946 : }
947 : }
948 : else
949 : {
950 : /*
951 : * This was an xlog-switch record, but the current insert location was
952 : * already exactly at the beginning of a segment, so there was no need
953 : * to do anything.
954 : */
955 : }
956 :
957 : /*
958 : * Done! Let others know that we're finished.
959 : */
960 29210322 : WALInsertLockRelease();
961 :
962 29210322 : END_CRIT_SECTION();
963 :
964 29210322 : MarkCurrentTransactionIdLoggedIfAny();
965 :
966 : /*
967 : * Mark top transaction id is logged (if needed) so that we should not try
968 : * to log it again with the next WAL record in the current subtransaction.
969 : */
970 29210322 : if (topxid_included)
971 438 : MarkSubxactTopXidLogged();
972 :
973 : /*
974 : * Update shared LogwrtRqst.Write, if we crossed page boundary.
975 : */
976 29210322 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
977 : {
978 3303170 : SpinLockAcquire(&XLogCtl->info_lck);
979 : /* advance global request to include new block(s) */
980 3303170 : if (XLogCtl->LogwrtRqst.Write < EndPos)
981 3166232 : XLogCtl->LogwrtRqst.Write = EndPos;
982 3303170 : SpinLockRelease(&XLogCtl->info_lck);
983 3303170 : RefreshXLogWriteResult(LogwrtResult);
984 : }
985 :
986 : /*
987 : * If this was an XLOG_SWITCH record, flush the record and the empty
988 : * padding space that fills the rest of the segment, and perform
989 : * end-of-segment actions (eg, notifying archiver).
990 : */
991 29210322 : if (class == WALINSERT_SPECIAL_SWITCH)
992 : {
993 : TRACE_POSTGRESQL_WAL_SWITCH();
994 1558 : XLogFlush(EndPos);
995 :
996 : /*
997 : * Even though we reserved the rest of the segment for us, which is
998 : * reflected in EndPos, we return a pointer to just the end of the
999 : * xlog-switch record.
1000 : */
1001 1558 : if (inserted)
1002 : {
1003 1442 : EndPos = StartPos + SizeOfXLogRecord;
1004 1442 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
1005 : {
1006 0 : uint64 offset = XLogSegmentOffset(EndPos, wal_segment_size);
1007 :
1008 0 : if (offset == EndPos % XLOG_BLCKSZ)
1009 0 : EndPos += SizeOfXLogLongPHD;
1010 : else
1011 0 : EndPos += SizeOfXLogShortPHD;
1012 : }
1013 : }
1014 : }
1015 :
1016 : #ifdef WAL_DEBUG
1017 : if (XLOG_DEBUG)
1018 : {
1019 : static XLogReaderState *debug_reader = NULL;
1020 : XLogRecord *record;
1021 : DecodedXLogRecord *decoded;
1022 : StringInfoData buf;
1023 : StringInfoData recordBuf;
1024 : char *errormsg = NULL;
1025 : MemoryContext oldCxt;
1026 :
1027 : oldCxt = MemoryContextSwitchTo(walDebugCxt);
1028 :
1029 : initStringInfo(&buf);
1030 : appendStringInfo(&buf, "INSERT @ %X/%08X: ", LSN_FORMAT_ARGS(EndPos));
1031 :
1032 : /*
1033 : * We have to piece together the WAL record data from the XLogRecData
1034 : * entries, so that we can pass it to the rm_desc function as one
1035 : * contiguous chunk.
1036 : */
1037 : initStringInfo(&recordBuf);
1038 : for (; rdata != NULL; rdata = rdata->next)
1039 : appendBinaryStringInfo(&recordBuf, rdata->data, rdata->len);
1040 :
1041 : /* We also need temporary space to decode the record. */
1042 : record = (XLogRecord *) recordBuf.data;
1043 : decoded = (DecodedXLogRecord *)
1044 : palloc(DecodeXLogRecordRequiredSpace(record->xl_tot_len));
1045 :
1046 : if (!debug_reader)
1047 : debug_reader = XLogReaderAllocate(wal_segment_size, NULL,
1048 : XL_ROUTINE(.page_read = NULL,
1049 : .segment_open = NULL,
1050 : .segment_close = NULL),
1051 : NULL);
1052 : if (!debug_reader)
1053 : {
1054 : appendStringInfoString(&buf, "error decoding record: out of memory while allocating a WAL reading processor");
1055 : }
1056 : else if (!DecodeXLogRecord(debug_reader,
1057 : decoded,
1058 : record,
1059 : EndPos,
1060 : &errormsg))
1061 : {
1062 : appendStringInfo(&buf, "error decoding record: %s",
1063 : errormsg ? errormsg : "no error message");
1064 : }
1065 : else
1066 : {
1067 : appendStringInfoString(&buf, " - ");
1068 :
1069 : debug_reader->record = decoded;
1070 : xlog_outdesc(&buf, debug_reader);
1071 : debug_reader->record = NULL;
1072 : }
1073 : elog(LOG, "%s", buf.data);
1074 :
1075 : pfree(decoded);
1076 : pfree(buf.data);
1077 : pfree(recordBuf.data);
1078 : MemoryContextSwitchTo(oldCxt);
1079 : }
1080 : #endif
1081 :
1082 : /*
1083 : * Update our global variables
1084 : */
1085 29210322 : ProcLastRecPtr = StartPos;
1086 29210322 : XactLastRecEnd = EndPos;
1087 :
1088 : /* Report WAL traffic to the instrumentation. */
1089 29210322 : if (inserted)
1090 : {
1091 29210206 : pgWalUsage.wal_bytes += rechdr->xl_tot_len;
1092 29210206 : pgWalUsage.wal_records++;
1093 29210206 : pgWalUsage.wal_fpi += num_fpi;
1094 : }
1095 :
1096 29210322 : return EndPos;
1097 : }
1098 :
1099 : /*
1100 : * Reserves the right amount of space for a record of given size from the WAL.
1101 : * *StartPos is set to the beginning of the reserved section, *EndPos to
1102 : * its end+1. *PrevPtr is set to the beginning of the previous record; it is
1103 : * used to set the xl_prev of this record.
1104 : *
1105 : * This is the performance critical part of XLogInsert that must be serialized
1106 : * across backends. The rest can happen mostly in parallel. Try to keep this
1107 : * section as short as possible, insertpos_lck can be heavily contended on a
1108 : * busy system.
1109 : *
1110 : * NB: The space calculation here must match the code in CopyXLogRecordToWAL,
1111 : * where we actually copy the record to the reserved space.
1112 : *
1113 : * NB: Testing shows that XLogInsertRecord runs faster if this code is inlined;
1114 : * however, because there are two call sites, the compiler is reluctant to
1115 : * inline. We use pg_attribute_always_inline here to try to convince it.
1116 : */
1117 : static pg_attribute_always_inline void
1118 29208764 : ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1119 : XLogRecPtr *PrevPtr)
1120 : {
1121 29208764 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1122 : uint64 startbytepos;
1123 : uint64 endbytepos;
1124 : uint64 prevbytepos;
1125 :
1126 29208764 : size = MAXALIGN(size);
1127 :
1128 : /* All (non xlog-switch) records should contain data. */
1129 : Assert(size > SizeOfXLogRecord);
1130 :
1131 : /*
1132 : * The duration the spinlock needs to be held is minimized by minimizing
1133 : * the calculations that have to be done while holding the lock. The
1134 : * current tip of reserved WAL is kept in CurrBytePos, as a byte position
1135 : * that only counts "usable" bytes in WAL, that is, it excludes all WAL
1136 : * page headers. The mapping between "usable" byte positions and physical
1137 : * positions (XLogRecPtrs) can be done outside the locked region, and
1138 : * because the usable byte position doesn't include any headers, reserving
1139 : * X bytes from WAL is almost as simple as "CurrBytePos += X".
1140 : */
1141 29208764 : SpinLockAcquire(&Insert->insertpos_lck);
1142 :
1143 29208764 : startbytepos = Insert->CurrBytePos;
1144 29208764 : endbytepos = startbytepos + size;
1145 29208764 : prevbytepos = Insert->PrevBytePos;
1146 29208764 : Insert->CurrBytePos = endbytepos;
1147 29208764 : Insert->PrevBytePos = startbytepos;
1148 :
1149 29208764 : SpinLockRelease(&Insert->insertpos_lck);
1150 :
1151 29208764 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1152 29208764 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1153 29208764 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1154 :
1155 : /*
1156 : * Check that the conversions between "usable byte positions" and
1157 : * XLogRecPtrs work consistently in both directions.
1158 : */
1159 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1160 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1161 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1162 29208764 : }
1163 :
1164 : /*
1165 : * Like ReserveXLogInsertLocation(), but for an xlog-switch record.
1166 : *
1167 : * A log-switch record is handled slightly differently. The rest of the
1168 : * segment will be reserved for this insertion, as indicated by the returned
1169 : * *EndPos value. However, if we are already at the beginning of the current
1170 : * segment, *StartPos and *EndPos are set to the current location without
1171 : * reserving any space, and the function returns false.
1172 : */
1173 : static bool
1174 1558 : ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1175 : {
1176 1558 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1177 : uint64 startbytepos;
1178 : uint64 endbytepos;
1179 : uint64 prevbytepos;
1180 1558 : uint32 size = MAXALIGN(SizeOfXLogRecord);
1181 : XLogRecPtr ptr;
1182 : uint32 segleft;
1183 :
1184 : /*
1185 : * These calculations are a bit heavy-weight to be done while holding a
1186 : * spinlock, but since we're holding all the WAL insertion locks, there
1187 : * are no other inserters competing for it. GetXLogInsertRecPtr() does
1188 : * compete for it, but that's not called very frequently.
1189 : */
1190 1558 : SpinLockAcquire(&Insert->insertpos_lck);
1191 :
1192 1558 : startbytepos = Insert->CurrBytePos;
1193 :
1194 1558 : ptr = XLogBytePosToEndRecPtr(startbytepos);
1195 1558 : if (XLogSegmentOffset(ptr, wal_segment_size) == 0)
1196 : {
1197 116 : SpinLockRelease(&Insert->insertpos_lck);
1198 116 : *EndPos = *StartPos = ptr;
1199 116 : return false;
1200 : }
1201 :
1202 1442 : endbytepos = startbytepos + size;
1203 1442 : prevbytepos = Insert->PrevBytePos;
1204 :
1205 1442 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1206 1442 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1207 :
1208 1442 : segleft = wal_segment_size - XLogSegmentOffset(*EndPos, wal_segment_size);
1209 1442 : if (segleft != wal_segment_size)
1210 : {
1211 : /* consume the rest of the segment */
1212 1442 : *EndPos += segleft;
1213 1442 : endbytepos = XLogRecPtrToBytePos(*EndPos);
1214 : }
1215 1442 : Insert->CurrBytePos = endbytepos;
1216 1442 : Insert->PrevBytePos = startbytepos;
1217 :
1218 1442 : SpinLockRelease(&Insert->insertpos_lck);
1219 :
1220 1442 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1221 :
1222 : Assert(XLogSegmentOffset(*EndPos, wal_segment_size) == 0);
1223 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1224 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1225 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1226 :
1227 1442 : return true;
1228 : }
1229 :
1230 : /*
1231 : * Subroutine of XLogInsertRecord. Copies a WAL record to an already-reserved
1232 : * area in the WAL.
1233 : */
1234 : static void
1235 29210206 : CopyXLogRecordToWAL(int write_len, bool isLogSwitch, XLogRecData *rdata,
1236 : XLogRecPtr StartPos, XLogRecPtr EndPos, TimeLineID tli)
1237 : {
1238 : char *currpos;
1239 : int freespace;
1240 : int written;
1241 : XLogRecPtr CurrPos;
1242 : XLogPageHeader pagehdr;
1243 :
1244 : /*
1245 : * Get a pointer to the right place in the right WAL buffer to start
1246 : * inserting to.
1247 : */
1248 29210206 : CurrPos = StartPos;
1249 29210206 : currpos = GetXLogBuffer(CurrPos, tli);
1250 29210206 : freespace = INSERT_FREESPACE(CurrPos);
1251 :
1252 : /*
1253 : * there should be enough space for at least the first field (xl_tot_len)
1254 : * on this page.
1255 : */
1256 : Assert(freespace >= sizeof(uint32));
1257 :
1258 : /* Copy record data */
1259 29210206 : written = 0;
1260 137892650 : while (rdata != NULL)
1261 : {
1262 108682444 : const char *rdata_data = rdata->data;
1263 108682444 : int rdata_len = rdata->len;
1264 :
1265 112222058 : while (rdata_len > freespace)
1266 : {
1267 : /*
1268 : * Write what fits on this page, and continue on the next page.
1269 : */
1270 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || freespace == 0);
1271 3539614 : memcpy(currpos, rdata_data, freespace);
1272 3539614 : rdata_data += freespace;
1273 3539614 : rdata_len -= freespace;
1274 3539614 : written += freespace;
1275 3539614 : CurrPos += freespace;
1276 :
1277 : /*
1278 : * Get pointer to beginning of next page, and set the xlp_rem_len
1279 : * in the page header. Set XLP_FIRST_IS_CONTRECORD.
1280 : *
1281 : * It's safe to set the contrecord flag and xlp_rem_len without a
1282 : * lock on the page. All the other flags were already set when the
1283 : * page was initialized, in AdvanceXLInsertBuffer, and we're the
1284 : * only backend that needs to set the contrecord flag.
1285 : */
1286 3539614 : currpos = GetXLogBuffer(CurrPos, tli);
1287 3539614 : pagehdr = (XLogPageHeader) currpos;
1288 3539614 : pagehdr->xlp_rem_len = write_len - written;
1289 3539614 : pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1290 :
1291 : /* skip over the page header */
1292 3539614 : if (XLogSegmentOffset(CurrPos, wal_segment_size) == 0)
1293 : {
1294 2512 : CurrPos += SizeOfXLogLongPHD;
1295 2512 : currpos += SizeOfXLogLongPHD;
1296 : }
1297 : else
1298 : {
1299 3537102 : CurrPos += SizeOfXLogShortPHD;
1300 3537102 : currpos += SizeOfXLogShortPHD;
1301 : }
1302 3539614 : freespace = INSERT_FREESPACE(CurrPos);
1303 : }
1304 :
1305 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1306 108682444 : memcpy(currpos, rdata_data, rdata_len);
1307 108682444 : currpos += rdata_len;
1308 108682444 : CurrPos += rdata_len;
1309 108682444 : freespace -= rdata_len;
1310 108682444 : written += rdata_len;
1311 :
1312 108682444 : rdata = rdata->next;
1313 : }
1314 : Assert(written == write_len);
1315 :
1316 : /*
1317 : * If this was an xlog-switch, it's not enough to write the switch record,
1318 : * we also have to consume all the remaining space in the WAL segment. We
1319 : * have already reserved that space, but we need to actually fill it.
1320 : */
1321 29210206 : if (isLogSwitch && XLogSegmentOffset(CurrPos, wal_segment_size) != 0)
1322 : {
1323 : /* An xlog-switch record doesn't contain any data besides the header */
1324 : Assert(write_len == SizeOfXLogRecord);
1325 :
1326 : /* Assert that we did reserve the right amount of space */
1327 : Assert(XLogSegmentOffset(EndPos, wal_segment_size) == 0);
1328 :
1329 : /* Use up all the remaining space on the current page */
1330 1442 : CurrPos += freespace;
1331 :
1332 : /*
1333 : * Cause all remaining pages in the segment to be flushed, leaving the
1334 : * XLog position where it should be, at the start of the next segment.
1335 : * We do this one page at a time, to make sure we don't deadlock
1336 : * against ourselves if wal_buffers < wal_segment_size.
1337 : */
1338 1383952 : while (CurrPos < EndPos)
1339 : {
1340 : /*
1341 : * The minimal action to flush the page would be to call
1342 : * WALInsertLockUpdateInsertingAt(CurrPos) followed by
1343 : * AdvanceXLInsertBuffer(...). The page would be left initialized
1344 : * mostly to zeros, except for the page header (always the short
1345 : * variant, as this is never a segment's first page).
1346 : *
1347 : * The large vistas of zeros are good for compressibility, but the
1348 : * headers interrupting them every XLOG_BLCKSZ (with values that
1349 : * differ from page to page) are not. The effect varies with
1350 : * compression tool, but bzip2 for instance compresses about an
1351 : * order of magnitude worse if those headers are left in place.
1352 : *
1353 : * Rather than complicating AdvanceXLInsertBuffer itself (which is
1354 : * called in heavily-loaded circumstances as well as this lightly-
1355 : * loaded one) with variant behavior, we just use GetXLogBuffer
1356 : * (which itself calls the two methods we need) to get the pointer
1357 : * and zero most of the page. Then we just zero the page header.
1358 : */
1359 1382510 : currpos = GetXLogBuffer(CurrPos, tli);
1360 5530040 : MemSet(currpos, 0, SizeOfXLogShortPHD);
1361 :
1362 1382510 : CurrPos += XLOG_BLCKSZ;
1363 : }
1364 : }
1365 : else
1366 : {
1367 : /* Align the end position, so that the next record starts aligned */
1368 29208764 : CurrPos = MAXALIGN64(CurrPos);
1369 : }
1370 :
1371 29210206 : if (CurrPos != EndPos)
1372 0 : ereport(PANIC,
1373 : errcode(ERRCODE_DATA_CORRUPTED),
1374 : errmsg_internal("space reserved for WAL record does not match what was written"));
1375 29210206 : }
1376 :
1377 : /*
1378 : * Acquire a WAL insertion lock, for inserting to WAL.
1379 : */
1380 : static void
1381 29222100 : WALInsertLockAcquire(void)
1382 : {
1383 : bool immed;
1384 :
1385 : /*
1386 : * It doesn't matter which of the WAL insertion locks we acquire, so try
1387 : * the one we used last time. If the system isn't particularly busy, it's
1388 : * a good bet that it's still available, and it's good to have some
1389 : * affinity to a particular lock so that you don't unnecessarily bounce
1390 : * cache lines between processes when there's no contention.
1391 : *
1392 : * If this is the first time through in this backend, pick a lock
1393 : * (semi-)randomly. This allows the locks to be used evenly if you have a
1394 : * lot of very short connections.
1395 : */
1396 : static int lockToTry = -1;
1397 :
1398 29222100 : if (lockToTry == -1)
1399 15752 : lockToTry = MyProcNumber % NUM_XLOGINSERT_LOCKS;
1400 29222100 : MyLockNo = lockToTry;
1401 :
1402 : /*
1403 : * The insertingAt value is initially set to 0, as we don't know our
1404 : * insert location yet.
1405 : */
1406 29222100 : immed = LWLockAcquire(&WALInsertLocks[MyLockNo].l.lock, LW_EXCLUSIVE);
1407 29222100 : if (!immed)
1408 : {
1409 : /*
1410 : * If we couldn't get the lock immediately, try another lock next
1411 : * time. On a system with more insertion locks than concurrent
1412 : * inserters, this causes all the inserters to eventually migrate to a
1413 : * lock that no-one else is using. On a system with more inserters
1414 : * than locks, it still helps to distribute the inserters evenly
1415 : * across the locks.
1416 : */
1417 56408 : lockToTry = (lockToTry + 1) % NUM_XLOGINSERT_LOCKS;
1418 : }
1419 29222100 : }
1420 :
1421 : /*
1422 : * Acquire all WAL insertion locks, to prevent other backends from inserting
1423 : * to WAL.
1424 : */
1425 : static void
1426 8534 : WALInsertLockAcquireExclusive(void)
1427 : {
1428 : int i;
1429 :
1430 : /*
1431 : * When holding all the locks, all but the last lock's insertingAt
1432 : * indicator is set to 0xFFFFFFFFFFFFFFFF, which is higher than any real
1433 : * XLogRecPtr value, to make sure that no-one blocks waiting on those.
1434 : */
1435 68272 : for (i = 0; i < NUM_XLOGINSERT_LOCKS - 1; i++)
1436 : {
1437 59738 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1438 59738 : LWLockUpdateVar(&WALInsertLocks[i].l.lock,
1439 59738 : &WALInsertLocks[i].l.insertingAt,
1440 : PG_UINT64_MAX);
1441 : }
1442 : /* Variable value reset to 0 at release */
1443 8534 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1444 :
1445 8534 : holdingAllLocks = true;
1446 8534 : }
1447 :
1448 : /*
1449 : * Release our insertion lock (or locks, if we're holding them all).
1450 : *
1451 : * NB: Reset all variables to 0, so they cause LWLockWaitForVar to block the
1452 : * next time the lock is acquired.
1453 : */
1454 : static void
1455 29230634 : WALInsertLockRelease(void)
1456 : {
1457 29230634 : if (holdingAllLocks)
1458 : {
1459 : int i;
1460 :
1461 76806 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1462 68272 : LWLockReleaseClearVar(&WALInsertLocks[i].l.lock,
1463 68272 : &WALInsertLocks[i].l.insertingAt,
1464 : 0);
1465 :
1466 8534 : holdingAllLocks = false;
1467 : }
1468 : else
1469 : {
1470 29222100 : LWLockReleaseClearVar(&WALInsertLocks[MyLockNo].l.lock,
1471 29222100 : &WALInsertLocks[MyLockNo].l.insertingAt,
1472 : 0);
1473 : }
1474 29230634 : }
1475 :
1476 : /*
1477 : * Update our insertingAt value, to let others know that we've finished
1478 : * inserting up to that point.
1479 : */
1480 : static void
1481 5111790 : WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt)
1482 : {
1483 5111790 : if (holdingAllLocks)
1484 : {
1485 : /*
1486 : * We use the last lock to mark our actual position, see comments in
1487 : * WALInsertLockAcquireExclusive.
1488 : */
1489 1370462 : LWLockUpdateVar(&WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.lock,
1490 1370462 : &WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.insertingAt,
1491 : insertingAt);
1492 : }
1493 : else
1494 3741328 : LWLockUpdateVar(&WALInsertLocks[MyLockNo].l.lock,
1495 3741328 : &WALInsertLocks[MyLockNo].l.insertingAt,
1496 : insertingAt);
1497 5111790 : }
1498 :
1499 : /*
1500 : * Wait for any WAL insertions < upto to finish.
1501 : *
1502 : * Returns the location of the oldest insertion that is still in-progress.
1503 : * Any WAL prior to that point has been fully copied into WAL buffers, and
1504 : * can be flushed out to disk. Because this waits for any insertions older
1505 : * than 'upto' to finish, the return value is always >= 'upto'.
1506 : *
1507 : * Note: When you are about to write out WAL, you must call this function
1508 : * *before* acquiring WALWriteLock, to avoid deadlocks. This function might
1509 : * need to wait for an insertion to finish (or at least advance to next
1510 : * uninitialized page), and the inserter might need to evict an old WAL buffer
1511 : * to make room for a new one, which in turn requires WALWriteLock.
1512 : */
1513 : static XLogRecPtr
1514 4223260 : WaitXLogInsertionsToFinish(XLogRecPtr upto)
1515 : {
1516 : uint64 bytepos;
1517 : XLogRecPtr inserted;
1518 : XLogRecPtr reservedUpto;
1519 : XLogRecPtr finishedUpto;
1520 4223260 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1521 : int i;
1522 :
1523 4223260 : if (MyProc == NULL)
1524 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1525 :
1526 : /*
1527 : * Check if there's any work to do. Use a barrier to ensure we get the
1528 : * freshest value.
1529 : */
1530 4223260 : inserted = pg_atomic_read_membarrier_u64(&XLogCtl->logInsertResult);
1531 4223260 : if (upto <= inserted)
1532 3434990 : return inserted;
1533 :
1534 : /* Read the current insert position */
1535 788270 : SpinLockAcquire(&Insert->insertpos_lck);
1536 788270 : bytepos = Insert->CurrBytePos;
1537 788270 : SpinLockRelease(&Insert->insertpos_lck);
1538 788270 : reservedUpto = XLogBytePosToEndRecPtr(bytepos);
1539 :
1540 : /*
1541 : * No-one should request to flush a piece of WAL that hasn't even been
1542 : * reserved yet. However, it can happen if there is a block with a bogus
1543 : * LSN on disk, for example. XLogFlush checks for that situation and
1544 : * complains, but only after the flush. Here we just assume that to mean
1545 : * that all WAL that has been reserved needs to be finished. In this
1546 : * corner-case, the return value can be smaller than 'upto' argument.
1547 : */
1548 788270 : if (upto > reservedUpto)
1549 : {
1550 0 : ereport(LOG,
1551 : errmsg("request to flush past end of generated WAL; request %X/%08X, current position %X/%08X",
1552 : LSN_FORMAT_ARGS(upto), LSN_FORMAT_ARGS(reservedUpto)));
1553 0 : upto = reservedUpto;
1554 : }
1555 :
1556 : /*
1557 : * Loop through all the locks, sleeping on any in-progress insert older
1558 : * than 'upto'.
1559 : *
1560 : * finishedUpto is our return value, indicating the point upto which all
1561 : * the WAL insertions have been finished. Initialize it to the head of
1562 : * reserved WAL, and as we iterate through the insertion locks, back it
1563 : * out for any insertion that's still in progress.
1564 : */
1565 788270 : finishedUpto = reservedUpto;
1566 7094430 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1567 : {
1568 6306160 : XLogRecPtr insertingat = InvalidXLogRecPtr;
1569 :
1570 : do
1571 : {
1572 : /*
1573 : * See if this insertion is in progress. LWLockWaitForVar will
1574 : * wait for the lock to be released, or for the 'value' to be set
1575 : * by a LWLockUpdateVar call. When a lock is initially acquired,
1576 : * its value is 0 (InvalidXLogRecPtr), which means that we don't
1577 : * know where it's inserting yet. We will have to wait for it. If
1578 : * it's a small insertion, the record will most likely fit on the
1579 : * same page and the inserter will release the lock without ever
1580 : * calling LWLockUpdateVar. But if it has to sleep, it will
1581 : * advertise the insertion point with LWLockUpdateVar before
1582 : * sleeping.
1583 : *
1584 : * In this loop we are only waiting for insertions that started
1585 : * before WaitXLogInsertionsToFinish was called. The lack of
1586 : * memory barriers in the loop means that we might see locks as
1587 : * "unused" that have since become used. This is fine because
1588 : * they only can be used for later insertions that we would not
1589 : * want to wait on anyway. Not taking a lock to acquire the
1590 : * current insertingAt value means that we might see older
1591 : * insertingAt values. This is also fine, because if we read a
1592 : * value too old, we will add ourselves to the wait queue, which
1593 : * contains atomic operations.
1594 : */
1595 6421818 : if (LWLockWaitForVar(&WALInsertLocks[i].l.lock,
1596 6421818 : &WALInsertLocks[i].l.insertingAt,
1597 : insertingat, &insertingat))
1598 : {
1599 : /* the lock was free, so no insertion in progress */
1600 4623186 : insertingat = InvalidXLogRecPtr;
1601 4623186 : break;
1602 : }
1603 :
1604 : /*
1605 : * This insertion is still in progress. Have to wait, unless the
1606 : * inserter has proceeded past 'upto'.
1607 : */
1608 1798632 : } while (insertingat < upto);
1609 :
1610 6306160 : if (insertingat != InvalidXLogRecPtr && insertingat < finishedUpto)
1611 627796 : finishedUpto = insertingat;
1612 : }
1613 :
1614 : /*
1615 : * Advance the limit we know to have been inserted and return the freshest
1616 : * value we know of, which might be beyond what we requested if somebody
1617 : * is concurrently doing this with an 'upto' pointer ahead of us.
1618 : */
1619 788270 : finishedUpto = pg_atomic_monotonic_advance_u64(&XLogCtl->logInsertResult,
1620 : finishedUpto);
1621 :
1622 788270 : return finishedUpto;
1623 : }
1624 :
1625 : /*
1626 : * Get a pointer to the right location in the WAL buffer containing the
1627 : * given XLogRecPtr.
1628 : *
1629 : * If the page is not initialized yet, it is initialized. That might require
1630 : * evicting an old dirty buffer from the buffer cache, which means I/O.
1631 : *
1632 : * The caller must ensure that the page containing the requested location
1633 : * isn't evicted yet, and won't be evicted. The way to ensure that is to
1634 : * hold onto a WAL insertion lock with the insertingAt position set to
1635 : * something <= ptr. GetXLogBuffer() will update insertingAt if it needs
1636 : * to evict an old page from the buffer. (This means that once you call
1637 : * GetXLogBuffer() with a given 'ptr', you must not access anything before
1638 : * that point anymore, and must not call GetXLogBuffer() with an older 'ptr'
1639 : * later, because older buffers might be recycled already)
1640 : */
1641 : static char *
1642 34132350 : GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli)
1643 : {
1644 : int idx;
1645 : XLogRecPtr endptr;
1646 : static uint64 cachedPage = 0;
1647 : static char *cachedPos = NULL;
1648 : XLogRecPtr expectedEndPtr;
1649 :
1650 : /*
1651 : * Fast path for the common case that we need to access again the same
1652 : * page as last time.
1653 : */
1654 34132350 : if (ptr / XLOG_BLCKSZ == cachedPage)
1655 : {
1656 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1657 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1658 27939866 : return cachedPos + ptr % XLOG_BLCKSZ;
1659 : }
1660 :
1661 : /*
1662 : * The XLog buffer cache is organized so that a page is always loaded to a
1663 : * particular buffer. That way we can easily calculate the buffer a given
1664 : * page must be loaded into, from the XLogRecPtr alone.
1665 : */
1666 6192484 : idx = XLogRecPtrToBufIdx(ptr);
1667 :
1668 : /*
1669 : * See what page is loaded in the buffer at the moment. It could be the
1670 : * page we're looking for, or something older. It can't be anything newer
1671 : * - that would imply the page we're looking for has already been written
1672 : * out to disk and evicted, and the caller is responsible for making sure
1673 : * that doesn't happen.
1674 : *
1675 : * We don't hold a lock while we read the value. If someone is just about
1676 : * to initialize or has just initialized the page, it's possible that we
1677 : * get InvalidXLogRecPtr. That's ok, we'll grab the mapping lock (in
1678 : * AdvanceXLInsertBuffer) and retry if we see anything other than the page
1679 : * we're looking for.
1680 : */
1681 6192484 : expectedEndPtr = ptr;
1682 6192484 : expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
1683 :
1684 6192484 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1685 6192484 : if (expectedEndPtr != endptr)
1686 : {
1687 : XLogRecPtr initializedUpto;
1688 :
1689 : /*
1690 : * Before calling AdvanceXLInsertBuffer(), which can block, let others
1691 : * know how far we're finished with inserting the record.
1692 : *
1693 : * NB: If 'ptr' points to just after the page header, advertise a
1694 : * position at the beginning of the page rather than 'ptr' itself. If
1695 : * there are no other insertions running, someone might try to flush
1696 : * up to our advertised location. If we advertised a position after
1697 : * the page header, someone might try to flush the page header, even
1698 : * though page might actually not be initialized yet. As the first
1699 : * inserter on the page, we are effectively responsible for making
1700 : * sure that it's initialized, before we let insertingAt to move past
1701 : * the page header.
1702 : */
1703 5111790 : if (ptr % XLOG_BLCKSZ == SizeOfXLogShortPHD &&
1704 12990 : XLogSegmentOffset(ptr, wal_segment_size) > XLOG_BLCKSZ)
1705 12990 : initializedUpto = ptr - SizeOfXLogShortPHD;
1706 5098800 : else if (ptr % XLOG_BLCKSZ == SizeOfXLogLongPHD &&
1707 2328 : XLogSegmentOffset(ptr, wal_segment_size) < XLOG_BLCKSZ)
1708 1190 : initializedUpto = ptr - SizeOfXLogLongPHD;
1709 : else
1710 5097610 : initializedUpto = ptr;
1711 :
1712 5111790 : WALInsertLockUpdateInsertingAt(initializedUpto);
1713 :
1714 5111790 : AdvanceXLInsertBuffer(ptr, tli, false);
1715 5111790 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1716 :
1717 5111790 : if (expectedEndPtr != endptr)
1718 0 : elog(PANIC, "could not find WAL buffer for %X/%08X",
1719 : LSN_FORMAT_ARGS(ptr));
1720 : }
1721 : else
1722 : {
1723 : /*
1724 : * Make sure the initialization of the page is visible to us, and
1725 : * won't arrive later to overwrite the WAL data we write on the page.
1726 : */
1727 1080694 : pg_memory_barrier();
1728 : }
1729 :
1730 : /*
1731 : * Found the buffer holding this page. Return a pointer to the right
1732 : * offset within the page.
1733 : */
1734 6192484 : cachedPage = ptr / XLOG_BLCKSZ;
1735 6192484 : cachedPos = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1736 :
1737 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1738 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1739 :
1740 6192484 : return cachedPos + ptr % XLOG_BLCKSZ;
1741 : }
1742 :
1743 : /*
1744 : * Read WAL data directly from WAL buffers, if available. Returns the number
1745 : * of bytes read successfully.
1746 : *
1747 : * Fewer than 'count' bytes may be read if some of the requested WAL data has
1748 : * already been evicted.
1749 : *
1750 : * No locks are taken.
1751 : *
1752 : * Caller should ensure that it reads no further than LogwrtResult.Write
1753 : * (which should have been updated by the caller when determining how far to
1754 : * read). The 'tli' argument is only used as a convenient safety check so that
1755 : * callers do not read from WAL buffers on a historical timeline.
1756 : */
1757 : Size
1758 193982 : WALReadFromBuffers(char *dstbuf, XLogRecPtr startptr, Size count,
1759 : TimeLineID tli)
1760 : {
1761 193982 : char *pdst = dstbuf;
1762 193982 : XLogRecPtr recptr = startptr;
1763 : XLogRecPtr inserted;
1764 193982 : Size nbytes = count;
1765 :
1766 193982 : if (RecoveryInProgress() || tli != GetWALInsertionTimeLine())
1767 1834 : return 0;
1768 :
1769 : Assert(!XLogRecPtrIsInvalid(startptr));
1770 :
1771 : /*
1772 : * Caller should ensure that the requested data has been inserted into WAL
1773 : * buffers before we try to read it.
1774 : */
1775 192148 : inserted = pg_atomic_read_u64(&XLogCtl->logInsertResult);
1776 192148 : if (startptr + count > inserted)
1777 0 : ereport(ERROR,
1778 : errmsg("cannot read past end of generated WAL: requested %X/%08X, current position %X/%08X",
1779 : LSN_FORMAT_ARGS(startptr + count),
1780 : LSN_FORMAT_ARGS(inserted)));
1781 :
1782 : /*
1783 : * Loop through the buffers without a lock. For each buffer, atomically
1784 : * read and verify the end pointer, then copy the data out, and finally
1785 : * re-read and re-verify the end pointer.
1786 : *
1787 : * Once a page is evicted, it never returns to the WAL buffers, so if the
1788 : * end pointer matches the expected end pointer before and after we copy
1789 : * the data, then the right page must have been present during the data
1790 : * copy. Read barriers are necessary to ensure that the data copy actually
1791 : * happens between the two verification steps.
1792 : *
1793 : * If either verification fails, we simply terminate the loop and return
1794 : * with the data that had been already copied out successfully.
1795 : */
1796 203900 : while (nbytes > 0)
1797 : {
1798 201202 : uint32 offset = recptr % XLOG_BLCKSZ;
1799 201202 : int idx = XLogRecPtrToBufIdx(recptr);
1800 : XLogRecPtr expectedEndPtr;
1801 : XLogRecPtr endptr;
1802 : const char *page;
1803 : const char *psrc;
1804 : Size npagebytes;
1805 :
1806 : /*
1807 : * Calculate the end pointer we expect in the xlblocks array if the
1808 : * correct page is present.
1809 : */
1810 201202 : expectedEndPtr = recptr + (XLOG_BLCKSZ - offset);
1811 :
1812 : /*
1813 : * First verification step: check that the correct page is present in
1814 : * the WAL buffers.
1815 : */
1816 201202 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1817 201202 : if (expectedEndPtr != endptr)
1818 189450 : break;
1819 :
1820 : /*
1821 : * The correct page is present (or was at the time the endptr was
1822 : * read; must re-verify later). Calculate pointer to source data and
1823 : * determine how much data to read from this page.
1824 : */
1825 11752 : page = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1826 11752 : psrc = page + offset;
1827 11752 : npagebytes = Min(nbytes, XLOG_BLCKSZ - offset);
1828 :
1829 : /*
1830 : * Ensure that the data copy and the first verification step are not
1831 : * reordered.
1832 : */
1833 11752 : pg_read_barrier();
1834 :
1835 : /* data copy */
1836 11752 : memcpy(pdst, psrc, npagebytes);
1837 :
1838 : /*
1839 : * Ensure that the data copy and the second verification step are not
1840 : * reordered.
1841 : */
1842 11752 : pg_read_barrier();
1843 :
1844 : /*
1845 : * Second verification step: check that the page we read from wasn't
1846 : * evicted while we were copying the data.
1847 : */
1848 11752 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1849 11752 : if (expectedEndPtr != endptr)
1850 0 : break;
1851 :
1852 11752 : pdst += npagebytes;
1853 11752 : recptr += npagebytes;
1854 11752 : nbytes -= npagebytes;
1855 : }
1856 :
1857 : Assert(pdst - dstbuf <= count);
1858 :
1859 192148 : return pdst - dstbuf;
1860 : }
1861 :
1862 : /*
1863 : * Converts a "usable byte position" to XLogRecPtr. A usable byte position
1864 : * is the position starting from the beginning of WAL, excluding all WAL
1865 : * page headers.
1866 : */
1867 : static XLogRecPtr
1868 58425576 : XLogBytePosToRecPtr(uint64 bytepos)
1869 : {
1870 : uint64 fullsegs;
1871 : uint64 fullpages;
1872 : uint64 bytesleft;
1873 : uint32 seg_offset;
1874 : XLogRecPtr result;
1875 :
1876 58425576 : fullsegs = bytepos / UsableBytesInSegment;
1877 58425576 : bytesleft = bytepos % UsableBytesInSegment;
1878 :
1879 58425576 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1880 : {
1881 : /* fits on first page of segment */
1882 108766 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1883 : }
1884 : else
1885 : {
1886 : /* account for the first page on segment with long header */
1887 58316810 : seg_offset = XLOG_BLCKSZ;
1888 58316810 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1889 :
1890 58316810 : fullpages = bytesleft / UsableBytesInPage;
1891 58316810 : bytesleft = bytesleft % UsableBytesInPage;
1892 :
1893 58316810 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1894 : }
1895 :
1896 58425576 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1897 :
1898 58425576 : return result;
1899 : }
1900 :
1901 : /*
1902 : * Like XLogBytePosToRecPtr, but if the position is at a page boundary,
1903 : * returns a pointer to the beginning of the page (ie. before page header),
1904 : * not to where the first xlog record on that page would go to. This is used
1905 : * when converting a pointer to the end of a record.
1906 : */
1907 : static XLogRecPtr
1908 30000034 : XLogBytePosToEndRecPtr(uint64 bytepos)
1909 : {
1910 : uint64 fullsegs;
1911 : uint64 fullpages;
1912 : uint64 bytesleft;
1913 : uint32 seg_offset;
1914 : XLogRecPtr result;
1915 :
1916 30000034 : fullsegs = bytepos / UsableBytesInSegment;
1917 30000034 : bytesleft = bytepos % UsableBytesInSegment;
1918 :
1919 30000034 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1920 : {
1921 : /* fits on first page of segment */
1922 166048 : if (bytesleft == 0)
1923 109026 : seg_offset = 0;
1924 : else
1925 57022 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1926 : }
1927 : else
1928 : {
1929 : /* account for the first page on segment with long header */
1930 29833986 : seg_offset = XLOG_BLCKSZ;
1931 29833986 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1932 :
1933 29833986 : fullpages = bytesleft / UsableBytesInPage;
1934 29833986 : bytesleft = bytesleft % UsableBytesInPage;
1935 :
1936 29833986 : if (bytesleft == 0)
1937 29254 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
1938 : else
1939 29804732 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1940 : }
1941 :
1942 30000034 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1943 :
1944 30000034 : return result;
1945 : }
1946 :
1947 : /*
1948 : * Convert an XLogRecPtr to a "usable byte position".
1949 : */
1950 : static uint64
1951 4930 : XLogRecPtrToBytePos(XLogRecPtr ptr)
1952 : {
1953 : uint64 fullsegs;
1954 : uint32 fullpages;
1955 : uint32 offset;
1956 : uint64 result;
1957 :
1958 4930 : XLByteToSeg(ptr, fullsegs, wal_segment_size);
1959 :
1960 4930 : fullpages = (XLogSegmentOffset(ptr, wal_segment_size)) / XLOG_BLCKSZ;
1961 4930 : offset = ptr % XLOG_BLCKSZ;
1962 :
1963 4930 : if (fullpages == 0)
1964 : {
1965 1952 : result = fullsegs * UsableBytesInSegment;
1966 1952 : if (offset > 0)
1967 : {
1968 : Assert(offset >= SizeOfXLogLongPHD);
1969 472 : result += offset - SizeOfXLogLongPHD;
1970 : }
1971 : }
1972 : else
1973 : {
1974 2978 : result = fullsegs * UsableBytesInSegment +
1975 2978 : (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
1976 2978 : (fullpages - 1) * UsableBytesInPage; /* full pages */
1977 2978 : if (offset > 0)
1978 : {
1979 : Assert(offset >= SizeOfXLogShortPHD);
1980 2960 : result += offset - SizeOfXLogShortPHD;
1981 : }
1982 : }
1983 :
1984 4930 : return result;
1985 : }
1986 :
1987 : /*
1988 : * Initialize XLOG buffers, writing out old buffers if they still contain
1989 : * unwritten data, upto the page containing 'upto'. Or if 'opportunistic' is
1990 : * true, initialize as many pages as we can without having to write out
1991 : * unwritten data. Any new pages are initialized to zeros, with pages headers
1992 : * initialized properly.
1993 : */
1994 : static void
1995 5121950 : AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
1996 : {
1997 5121950 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1998 : int nextidx;
1999 : XLogRecPtr OldPageRqstPtr;
2000 : XLogwrtRqst WriteRqst;
2001 5121950 : XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
2002 : XLogRecPtr NewPageBeginPtr;
2003 : XLogPageHeader NewPage;
2004 : XLogRecPtr ReservedPtr;
2005 5121950 : int npages pg_attribute_unused() = 0;
2006 :
2007 : /*
2008 : * We must run the loop below inside the critical section as we expect
2009 : * XLogCtl->InitializedUpTo to eventually keep up. The most of callers
2010 : * already run inside the critical section. Except for WAL writer, which
2011 : * passed 'opportunistic == true', and therefore we don't perform
2012 : * operations that could error out.
2013 : *
2014 : * Start an explicit critical section anyway though.
2015 : */
2016 : Assert(CritSectionCount > 0 || opportunistic);
2017 5121950 : START_CRIT_SECTION();
2018 :
2019 : /*--
2020 : * Loop till we get all the pages in WAL buffer before 'upto' reserved for
2021 : * initialization. Multiple process can initialize different buffers with
2022 : * this loop in parallel as following.
2023 : *
2024 : * 1. Reserve page for initialization using XLogCtl->InitializeReserved.
2025 : * 2. Initialize the reserved page.
2026 : * 3. Attempt to advance XLogCtl->InitializedUpTo,
2027 : */
2028 5121950 : ReservedPtr = pg_atomic_read_u64(&XLogCtl->InitializeReserved);
2029 15165176 : while (upto >= ReservedPtr || opportunistic)
2030 : {
2031 : Assert(ReservedPtr % XLOG_BLCKSZ == 0);
2032 :
2033 : /*
2034 : * Get ending-offset of the buffer page we need to replace.
2035 : *
2036 : * We don't lookup into xlblocks, but rather calculate position we
2037 : * must wait to be written. If it was written, xlblocks will have this
2038 : * position (or uninitialized)
2039 : */
2040 10053386 : if (ReservedPtr + XLOG_BLCKSZ > XLogCtl->InitializedFrom + XLOG_BLCKSZ * XLOGbuffers)
2041 9499274 : OldPageRqstPtr = ReservedPtr + XLOG_BLCKSZ - (XLogRecPtr) XLOG_BLCKSZ * XLOGbuffers;
2042 : else
2043 554112 : OldPageRqstPtr = InvalidXLogRecPtr;
2044 :
2045 10053386 : if (LogwrtResult.Write < OldPageRqstPtr && opportunistic)
2046 : {
2047 : /*
2048 : * If we just want to pre-initialize as much as we can without
2049 : * flushing, give up now.
2050 : */
2051 10160 : upto = ReservedPtr - 1;
2052 10160 : break;
2053 : }
2054 :
2055 : /*
2056 : * Attempt to reserve the page for initialization. Failure means that
2057 : * this page got reserved by another process.
2058 : */
2059 10043226 : if (!pg_atomic_compare_exchange_u64(&XLogCtl->InitializeReserved,
2060 : &ReservedPtr,
2061 : ReservedPtr + XLOG_BLCKSZ))
2062 5025290 : continue;
2063 :
2064 : /*
2065 : * Wait till page gets correctly initialized up to OldPageRqstPtr.
2066 : */
2067 5017936 : nextidx = XLogRecPtrToBufIdx(ReservedPtr);
2068 5018464 : while (pg_atomic_read_u64(&XLogCtl->InitializedUpTo) < OldPageRqstPtr)
2069 528 : ConditionVariableSleep(&XLogCtl->InitializedUpToCondVar, WAIT_EVENT_WAL_BUFFER_INIT);
2070 5017936 : ConditionVariableCancelSleep();
2071 : Assert(pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]) == OldPageRqstPtr);
2072 :
2073 : /* Fall through if it's already written out. */
2074 5017936 : if (LogwrtResult.Write < OldPageRqstPtr)
2075 : {
2076 : /* Nope, got work to do. */
2077 :
2078 : /* Advance shared memory write request position */
2079 3963928 : SpinLockAcquire(&XLogCtl->info_lck);
2080 3963928 : if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
2081 1244236 : XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
2082 3963928 : SpinLockRelease(&XLogCtl->info_lck);
2083 :
2084 : /*
2085 : * Acquire an up-to-date LogwrtResult value and see if we still
2086 : * need to write it or if someone else already did.
2087 : */
2088 3963928 : RefreshXLogWriteResult(LogwrtResult);
2089 3963928 : if (LogwrtResult.Write < OldPageRqstPtr)
2090 : {
2091 3943212 : WaitXLogInsertionsToFinish(OldPageRqstPtr);
2092 :
2093 3943212 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2094 :
2095 3943212 : RefreshXLogWriteResult(LogwrtResult);
2096 3943212 : if (LogwrtResult.Write >= OldPageRqstPtr)
2097 : {
2098 : /* OK, someone wrote it already */
2099 56798 : LWLockRelease(WALWriteLock);
2100 : }
2101 : else
2102 : {
2103 : /* Have to write it ourselves */
2104 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2105 3886414 : WriteRqst.Write = OldPageRqstPtr;
2106 3886414 : WriteRqst.Flush = 0;
2107 3886414 : XLogWrite(WriteRqst, tli, false);
2108 3886414 : LWLockRelease(WALWriteLock);
2109 3886414 : pgWalUsage.wal_buffers_full++;
2110 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_DONE();
2111 : }
2112 : }
2113 : }
2114 :
2115 : /*
2116 : * Now the next buffer slot is free and we can set it up to be the
2117 : * next output page.
2118 : */
2119 5017936 : NewPageBeginPtr = ReservedPtr;
2120 5017936 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2121 :
2122 5017936 : NewPage = (XLogPageHeader) (XLogCtl->pages + nextidx * (Size) XLOG_BLCKSZ);
2123 :
2124 : /*
2125 : * Mark the xlblock with InvalidXLogRecPtr and issue a write barrier
2126 : * before initializing. Otherwise, the old page may be partially
2127 : * zeroed but look valid.
2128 : */
2129 5017936 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], InvalidXLogRecPtr);
2130 5017936 : pg_write_barrier();
2131 :
2132 : /*
2133 : * Be sure to re-zero the buffer so that bytes beyond what we've
2134 : * written will look like zeroes and not valid XLOG records...
2135 : */
2136 5017936 : MemSet(NewPage, 0, XLOG_BLCKSZ);
2137 :
2138 : /*
2139 : * Fill the new page's header
2140 : */
2141 5017936 : NewPage->xlp_magic = XLOG_PAGE_MAGIC;
2142 :
2143 : /* NewPage->xlp_info = 0; */ /* done by memset */
2144 5017936 : NewPage->xlp_tli = tli;
2145 5017936 : NewPage->xlp_pageaddr = NewPageBeginPtr;
2146 :
2147 : /* NewPage->xlp_rem_len = 0; */ /* done by memset */
2148 :
2149 : /*
2150 : * If online backup is not in progress, mark the header to indicate
2151 : * that WAL records beginning in this page have removable backup
2152 : * blocks. This allows the WAL archiver to know whether it is safe to
2153 : * compress archived WAL data by transforming full-block records into
2154 : * the non-full-block format. It is sufficient to record this at the
2155 : * page level because we force a page switch (in fact a segment
2156 : * switch) when starting a backup, so the flag will be off before any
2157 : * records can be written during the backup. At the end of a backup,
2158 : * the last page will be marked as all unsafe when perhaps only part
2159 : * is unsafe, but at worst the archiver would miss the opportunity to
2160 : * compress a few records.
2161 : */
2162 5017936 : if (Insert->runningBackups == 0)
2163 4760422 : NewPage->xlp_info |= XLP_BKP_REMOVABLE;
2164 :
2165 : /*
2166 : * If first page of an XLOG segment file, make it a long header.
2167 : */
2168 5017936 : if ((XLogSegmentOffset(NewPage->xlp_pageaddr, wal_segment_size)) == 0)
2169 : {
2170 3710 : XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2171 :
2172 3710 : NewLongPage->xlp_sysid = ControlFile->system_identifier;
2173 3710 : NewLongPage->xlp_seg_size = wal_segment_size;
2174 3710 : NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2175 3710 : NewPage->xlp_info |= XLP_LONG_HEADER;
2176 : }
2177 :
2178 : /*
2179 : * Make sure the initialization of the page becomes visible to others
2180 : * before the xlblocks update. GetXLogBuffer() reads xlblocks without
2181 : * holding a lock.
2182 : */
2183 5017936 : pg_write_barrier();
2184 :
2185 : /*-----
2186 : * Update the value of XLogCtl->xlblocks[nextidx] and try to advance
2187 : * XLogCtl->InitializedUpTo in a lock-less manner.
2188 : *
2189 : * First, let's provide a formal proof of the algorithm. Let it be 'n'
2190 : * process with the following variables in shared memory:
2191 : * f - an array of 'n' boolean flags,
2192 : * v - atomic integer variable.
2193 : *
2194 : * Also, let
2195 : * i - a number of a process,
2196 : * j - local integer variable,
2197 : * CAS(var, oldval, newval) - compare-and-swap atomic operation
2198 : * returning true on success,
2199 : * write_barrier()/read_barrier() - memory barriers.
2200 : *
2201 : * The pseudocode for each process is the following.
2202 : *
2203 : * j := i
2204 : * f[i] := true
2205 : * write_barrier()
2206 : * while CAS(v, j, j + 1):
2207 : * j := j + 1
2208 : * read_barrier()
2209 : * if not f[j]:
2210 : * break
2211 : *
2212 : * Let's prove that v eventually reaches the value of n.
2213 : * 1. Prove by contradiction. Assume v doesn't reach n and stucks
2214 : * on k, where k < n.
2215 : * 2. Process k attempts CAS(v, k, k + 1). 1). If, as we assumed, v
2216 : * gets stuck at k, then this CAS operation must fail. Therefore,
2217 : * v < k when process k attempts CAS(v, k, k + 1).
2218 : * 3. If, as we assumed, v gets stuck at k, then the value k of v
2219 : * must be achieved by some process m, where m < k. The process
2220 : * m must observe f[k] == false. Otherwise, it will later attempt
2221 : * CAS(v, k, k + 1) with success.
2222 : * 4. Therefore, corresponding read_barrier() (while j == k) on
2223 : * process m reached before write_barrier() of process k. But then
2224 : * process k attempts CAS(v, k, k + 1) after process m successfully
2225 : * incremented v to k, and that CAS operation must succeed.
2226 : * That leads to a contradiction. So, there is no such k (k < n)
2227 : * where v gets stuck. Q.E.D.
2228 : *
2229 : * To apply this proof to the code below, we assume
2230 : * XLogCtl->InitializedUpTo will play the role of v with XLOG_BLCKSZ
2231 : * granularity. We also assume setting XLogCtl->xlblocks[nextidx] to
2232 : * NewPageEndPtr to play the role of setting f[i] to true. Also, note
2233 : * that processes can't concurrently map different xlog locations to
2234 : * the same nextidx because we previously requested that
2235 : * XLogCtl->InitializedUpTo >= OldPageRqstPtr. So, a xlog buffer can
2236 : * be taken for initialization only once the previous initialization
2237 : * takes effect on XLogCtl->InitializedUpTo.
2238 : */
2239 :
2240 5017936 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], NewPageEndPtr);
2241 :
2242 5017936 : pg_write_barrier();
2243 :
2244 5096862 : while (pg_atomic_compare_exchange_u64(&XLogCtl->InitializedUpTo, &NewPageBeginPtr, NewPageEndPtr))
2245 : {
2246 5025796 : NewPageBeginPtr = NewPageEndPtr;
2247 5025796 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2248 5025796 : nextidx = XLogRecPtrToBufIdx(NewPageBeginPtr);
2249 :
2250 5025796 : pg_read_barrier();
2251 :
2252 5025796 : if (pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]) != NewPageEndPtr)
2253 : {
2254 : /*
2255 : * Page at nextidx wasn't initialized yet, so we can't move
2256 : * InitializedUpto further. It will be moved by backend which
2257 : * will initialize nextidx.
2258 : */
2259 4946870 : ConditionVariableBroadcast(&XLogCtl->InitializedUpToCondVar);
2260 4946870 : break;
2261 : }
2262 : }
2263 :
2264 5017936 : npages++;
2265 : }
2266 :
2267 5121950 : END_CRIT_SECTION();
2268 :
2269 : /*
2270 : * All the pages in WAL buffer before 'upto' were reserved for
2271 : * initialization. However, some pages might be reserved by concurrent
2272 : * processes. Wait till they finish initialization.
2273 : */
2274 6280914 : while (upto >= pg_atomic_read_u64(&XLogCtl->InitializedUpTo))
2275 1158964 : ConditionVariableSleep(&XLogCtl->InitializedUpToCondVar, WAIT_EVENT_WAL_BUFFER_INIT);
2276 5121950 : ConditionVariableCancelSleep();
2277 :
2278 5121950 : pg_read_barrier();
2279 :
2280 : #ifdef WAL_DEBUG
2281 : if (XLOG_DEBUG && npages > 0)
2282 : {
2283 : elog(DEBUG1, "initialized %d pages, up to %X/%08X",
2284 : npages, LSN_FORMAT_ARGS(NewPageEndPtr));
2285 : }
2286 : #endif
2287 5121950 : }
2288 :
2289 : /*
2290 : * Calculate CheckPointSegments based on max_wal_size_mb and
2291 : * checkpoint_completion_target.
2292 : */
2293 : static void
2294 15264 : CalculateCheckpointSegments(void)
2295 : {
2296 : double target;
2297 :
2298 : /*-------
2299 : * Calculate the distance at which to trigger a checkpoint, to avoid
2300 : * exceeding max_wal_size_mb. This is based on two assumptions:
2301 : *
2302 : * a) we keep WAL for only one checkpoint cycle (prior to PG11 we kept
2303 : * WAL for two checkpoint cycles to allow us to recover from the
2304 : * secondary checkpoint if the first checkpoint failed, though we
2305 : * only did this on the primary anyway, not on standby. Keeping just
2306 : * one checkpoint simplifies processing and reduces disk space in
2307 : * many smaller databases.)
2308 : * b) during checkpoint, we consume checkpoint_completion_target *
2309 : * number of segments consumed between checkpoints.
2310 : *-------
2311 : */
2312 15264 : target = (double) ConvertToXSegs(max_wal_size_mb, wal_segment_size) /
2313 15264 : (1.0 + CheckPointCompletionTarget);
2314 :
2315 : /* round down */
2316 15264 : CheckPointSegments = (int) target;
2317 :
2318 15264 : if (CheckPointSegments < 1)
2319 20 : CheckPointSegments = 1;
2320 15264 : }
2321 :
2322 : void
2323 11074 : assign_max_wal_size(int newval, void *extra)
2324 : {
2325 11074 : max_wal_size_mb = newval;
2326 11074 : CalculateCheckpointSegments();
2327 11074 : }
2328 :
2329 : void
2330 2226 : assign_checkpoint_completion_target(double newval, void *extra)
2331 : {
2332 2226 : CheckPointCompletionTarget = newval;
2333 2226 : CalculateCheckpointSegments();
2334 2226 : }
2335 :
2336 : bool
2337 4294 : check_wal_segment_size(int *newval, void **extra, GucSource source)
2338 : {
2339 4294 : if (!IsValidWalSegSize(*newval))
2340 : {
2341 0 : GUC_check_errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.");
2342 0 : return false;
2343 : }
2344 :
2345 4294 : return true;
2346 : }
2347 :
2348 : /*
2349 : * At a checkpoint, how many WAL segments to recycle as preallocated future
2350 : * XLOG segments? Returns the highest segment that should be preallocated.
2351 : */
2352 : static XLogSegNo
2353 3378 : XLOGfileslop(XLogRecPtr lastredoptr)
2354 : {
2355 : XLogSegNo minSegNo;
2356 : XLogSegNo maxSegNo;
2357 : double distance;
2358 : XLogSegNo recycleSegNo;
2359 :
2360 : /*
2361 : * Calculate the segment numbers that min_wal_size_mb and max_wal_size_mb
2362 : * correspond to. Always recycle enough segments to meet the minimum, and
2363 : * remove enough segments to stay below the maximum.
2364 : */
2365 3378 : minSegNo = lastredoptr / wal_segment_size +
2366 3378 : ConvertToXSegs(min_wal_size_mb, wal_segment_size) - 1;
2367 3378 : maxSegNo = lastredoptr / wal_segment_size +
2368 3378 : ConvertToXSegs(max_wal_size_mb, wal_segment_size) - 1;
2369 :
2370 : /*
2371 : * Between those limits, recycle enough segments to get us through to the
2372 : * estimated end of next checkpoint.
2373 : *
2374 : * To estimate where the next checkpoint will finish, assume that the
2375 : * system runs steadily consuming CheckPointDistanceEstimate bytes between
2376 : * every checkpoint.
2377 : */
2378 3378 : distance = (1.0 + CheckPointCompletionTarget) * CheckPointDistanceEstimate;
2379 : /* add 10% for good measure. */
2380 3378 : distance *= 1.10;
2381 :
2382 3378 : recycleSegNo = (XLogSegNo) ceil(((double) lastredoptr + distance) /
2383 : wal_segment_size);
2384 :
2385 3378 : if (recycleSegNo < minSegNo)
2386 2386 : recycleSegNo = minSegNo;
2387 3378 : if (recycleSegNo > maxSegNo)
2388 760 : recycleSegNo = maxSegNo;
2389 :
2390 3378 : return recycleSegNo;
2391 : }
2392 :
2393 : /*
2394 : * Check whether we've consumed enough xlog space that a checkpoint is needed.
2395 : *
2396 : * new_segno indicates a log file that has just been filled up (or read
2397 : * during recovery). We measure the distance from RedoRecPtr to new_segno
2398 : * and see if that exceeds CheckPointSegments.
2399 : *
2400 : * Note: it is caller's responsibility that RedoRecPtr is up-to-date.
2401 : */
2402 : bool
2403 10286 : XLogCheckpointNeeded(XLogSegNo new_segno)
2404 : {
2405 : XLogSegNo old_segno;
2406 :
2407 10286 : XLByteToSeg(RedoRecPtr, old_segno, wal_segment_size);
2408 :
2409 10286 : if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2410 6578 : return true;
2411 3708 : return false;
2412 : }
2413 :
2414 : /*
2415 : * Write and/or fsync the log at least as far as WriteRqst indicates.
2416 : *
2417 : * If flexible == true, we don't have to write as far as WriteRqst, but
2418 : * may stop at any convenient boundary (such as a cache or logfile boundary).
2419 : * This option allows us to avoid uselessly issuing multiple writes when a
2420 : * single one would do.
2421 : *
2422 : * Must be called with WALWriteLock held. WaitXLogInsertionsToFinish(WriteRqst)
2423 : * must be called before grabbing the lock, to make sure the data is ready to
2424 : * write.
2425 : */
2426 : static void
2427 4152004 : XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible)
2428 : {
2429 : bool ispartialpage;
2430 : bool last_iteration;
2431 : bool finishing_seg;
2432 : int curridx;
2433 : int npages;
2434 : int startidx;
2435 : uint32 startoffset;
2436 :
2437 : /* We should always be inside a critical section here */
2438 : Assert(CritSectionCount > 0);
2439 :
2440 : /*
2441 : * Update local LogwrtResult (caller probably did this already, but...)
2442 : */
2443 4152004 : RefreshXLogWriteResult(LogwrtResult);
2444 :
2445 : /*
2446 : * Since successive pages in the xlog cache are consecutively allocated,
2447 : * we can usually gather multiple pages together and issue just one
2448 : * write() call. npages is the number of pages we have determined can be
2449 : * written together; startidx is the cache block index of the first one,
2450 : * and startoffset is the file offset at which it should go. The latter
2451 : * two variables are only valid when npages > 0, but we must initialize
2452 : * all of them to keep the compiler quiet.
2453 : */
2454 4152004 : npages = 0;
2455 4152004 : startidx = 0;
2456 4152004 : startoffset = 0;
2457 :
2458 : /*
2459 : * Within the loop, curridx is the cache block index of the page to
2460 : * consider writing. Begin at the buffer containing the next unwritten
2461 : * page, or last partially written page.
2462 : */
2463 4152004 : curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2464 :
2465 9074946 : while (LogwrtResult.Write < WriteRqst.Write)
2466 : {
2467 : /*
2468 : * Make sure we're not ahead of the insert process. This could happen
2469 : * if we're passed a bogus WriteRqst.Write that is past the end of the
2470 : * last page that's been initialized by AdvanceXLInsertBuffer.
2471 : */
2472 5178740 : XLogRecPtr EndPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[curridx]);
2473 :
2474 5178740 : if (LogwrtResult.Write >= EndPtr)
2475 0 : elog(PANIC, "xlog write request %X/%08X is past end of log %X/%08X",
2476 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2477 : LSN_FORMAT_ARGS(EndPtr));
2478 :
2479 : /* Advance LogwrtResult.Write to end of current buffer page */
2480 5178740 : LogwrtResult.Write = EndPtr;
2481 5178740 : ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2482 :
2483 5178740 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2484 : wal_segment_size))
2485 : {
2486 : /*
2487 : * Switch to new logfile segment. We cannot have any pending
2488 : * pages here (since we dump what we have at segment end).
2489 : */
2490 : Assert(npages == 0);
2491 26578 : if (openLogFile >= 0)
2492 12258 : XLogFileClose();
2493 26578 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2494 : wal_segment_size);
2495 26578 : openLogTLI = tli;
2496 :
2497 : /* create/use new log file */
2498 26578 : openLogFile = XLogFileInit(openLogSegNo, tli);
2499 26578 : ReserveExternalFD();
2500 : }
2501 :
2502 : /* Make sure we have the current logfile open */
2503 5178740 : if (openLogFile < 0)
2504 : {
2505 0 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2506 : wal_segment_size);
2507 0 : openLogTLI = tli;
2508 0 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2509 0 : ReserveExternalFD();
2510 : }
2511 :
2512 : /* Add current page to the set of pending pages-to-dump */
2513 5178740 : if (npages == 0)
2514 : {
2515 : /* first of group */
2516 4186390 : startidx = curridx;
2517 4186390 : startoffset = XLogSegmentOffset(LogwrtResult.Write - XLOG_BLCKSZ,
2518 : wal_segment_size);
2519 : }
2520 5178740 : npages++;
2521 :
2522 : /*
2523 : * Dump the set if this will be the last loop iteration, or if we are
2524 : * at the last page of the cache area (since the next page won't be
2525 : * contiguous in memory), or if we are at the end of the logfile
2526 : * segment.
2527 : */
2528 5178740 : last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2529 :
2530 10109366 : finishing_seg = !ispartialpage &&
2531 4930626 : (startoffset + npages * XLOG_BLCKSZ) >= wal_segment_size;
2532 :
2533 5178740 : if (last_iteration ||
2534 1028854 : curridx == XLogCtl->XLogCacheBlck ||
2535 : finishing_seg)
2536 : {
2537 : char *from;
2538 : Size nbytes;
2539 : Size nleft;
2540 : ssize_t written;
2541 : instr_time start;
2542 :
2543 : /* OK to write the page(s) */
2544 4186390 : from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2545 4186390 : nbytes = npages * (Size) XLOG_BLCKSZ;
2546 4186390 : nleft = nbytes;
2547 : do
2548 : {
2549 4186390 : errno = 0;
2550 :
2551 : /*
2552 : * Measure I/O timing to write WAL data, for pg_stat_io.
2553 : */
2554 4186390 : start = pgstat_prepare_io_time(track_wal_io_timing);
2555 :
2556 4186390 : pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE);
2557 4186390 : written = pg_pwrite(openLogFile, from, nleft, startoffset);
2558 4186390 : pgstat_report_wait_end();
2559 :
2560 4186390 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL,
2561 : IOOP_WRITE, start, 1, written);
2562 :
2563 4186390 : if (written <= 0)
2564 : {
2565 : char xlogfname[MAXFNAMELEN];
2566 : int save_errno;
2567 :
2568 0 : if (errno == EINTR)
2569 0 : continue;
2570 :
2571 0 : save_errno = errno;
2572 0 : XLogFileName(xlogfname, tli, openLogSegNo,
2573 : wal_segment_size);
2574 0 : errno = save_errno;
2575 0 : ereport(PANIC,
2576 : (errcode_for_file_access(),
2577 : errmsg("could not write to log file \"%s\" at offset %u, length %zu: %m",
2578 : xlogfname, startoffset, nleft)));
2579 : }
2580 4186390 : nleft -= written;
2581 4186390 : from += written;
2582 4186390 : startoffset += written;
2583 4186390 : } while (nleft > 0);
2584 :
2585 4186390 : npages = 0;
2586 :
2587 : /*
2588 : * If we just wrote the whole last page of a logfile segment,
2589 : * fsync the segment immediately. This avoids having to go back
2590 : * and re-open prior segments when an fsync request comes along
2591 : * later. Doing it here ensures that one and only one backend will
2592 : * perform this fsync.
2593 : *
2594 : * This is also the right place to notify the Archiver that the
2595 : * segment is ready to copy to archival storage, and to update the
2596 : * timer for archive_timeout, and to signal for a checkpoint if
2597 : * too many logfile segments have been used since the last
2598 : * checkpoint.
2599 : */
2600 4186390 : if (finishing_seg)
2601 : {
2602 3966 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2603 :
2604 : /* signal that we need to wakeup walsenders later */
2605 3966 : WalSndWakeupRequest();
2606 :
2607 3966 : LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2608 :
2609 3966 : if (XLogArchivingActive())
2610 812 : XLogArchiveNotifySeg(openLogSegNo, tli);
2611 :
2612 3966 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2613 3966 : XLogCtl->lastSegSwitchLSN = LogwrtResult.Flush;
2614 :
2615 : /*
2616 : * Request a checkpoint if we've consumed too much xlog since
2617 : * the last one. For speed, we first check using the local
2618 : * copy of RedoRecPtr, which might be out of date; if it looks
2619 : * like a checkpoint is needed, forcibly update RedoRecPtr and
2620 : * recheck.
2621 : */
2622 3966 : if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2623 : {
2624 468 : (void) GetRedoRecPtr();
2625 468 : if (XLogCheckpointNeeded(openLogSegNo))
2626 382 : RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2627 : }
2628 : }
2629 : }
2630 :
2631 5178740 : if (ispartialpage)
2632 : {
2633 : /* Only asked to write a partial page */
2634 248114 : LogwrtResult.Write = WriteRqst.Write;
2635 248114 : break;
2636 : }
2637 4930626 : curridx = NextBufIdx(curridx);
2638 :
2639 : /* If flexible, break out of loop as soon as we wrote something */
2640 4930626 : if (flexible && npages == 0)
2641 7684 : break;
2642 : }
2643 :
2644 : Assert(npages == 0);
2645 :
2646 : /*
2647 : * If asked to flush, do so
2648 : */
2649 4152004 : if (LogwrtResult.Flush < WriteRqst.Flush &&
2650 264112 : LogwrtResult.Flush < LogwrtResult.Write)
2651 : {
2652 : /*
2653 : * Could get here without iterating above loop, in which case we might
2654 : * have no open file or the wrong one. However, we do not need to
2655 : * fsync more than one file.
2656 : */
2657 264004 : if (wal_sync_method != WAL_SYNC_METHOD_OPEN &&
2658 264004 : wal_sync_method != WAL_SYNC_METHOD_OPEN_DSYNC)
2659 : {
2660 264004 : if (openLogFile >= 0 &&
2661 263980 : !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2662 : wal_segment_size))
2663 186 : XLogFileClose();
2664 264004 : if (openLogFile < 0)
2665 : {
2666 210 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2667 : wal_segment_size);
2668 210 : openLogTLI = tli;
2669 210 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2670 210 : ReserveExternalFD();
2671 : }
2672 :
2673 264004 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2674 : }
2675 :
2676 : /* signal that we need to wakeup walsenders later */
2677 264004 : WalSndWakeupRequest();
2678 :
2679 264004 : LogwrtResult.Flush = LogwrtResult.Write;
2680 : }
2681 :
2682 : /*
2683 : * Update shared-memory status
2684 : *
2685 : * We make sure that the shared 'request' values do not fall behind the
2686 : * 'result' values. This is not absolutely essential, but it saves some
2687 : * code in a couple of places.
2688 : */
2689 4152004 : SpinLockAcquire(&XLogCtl->info_lck);
2690 4152004 : if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
2691 232008 : XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
2692 4152004 : if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
2693 267046 : XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
2694 4152004 : SpinLockRelease(&XLogCtl->info_lck);
2695 :
2696 : /*
2697 : * We write Write first, bar, then Flush. When reading, the opposite must
2698 : * be done (with a matching barrier in between), so that we always see a
2699 : * Flush value that trails behind the Write value seen.
2700 : */
2701 4152004 : pg_atomic_write_u64(&XLogCtl->logWriteResult, LogwrtResult.Write);
2702 4152004 : pg_write_barrier();
2703 4152004 : pg_atomic_write_u64(&XLogCtl->logFlushResult, LogwrtResult.Flush);
2704 :
2705 : #ifdef USE_ASSERT_CHECKING
2706 : {
2707 : XLogRecPtr Flush;
2708 : XLogRecPtr Write;
2709 : XLogRecPtr Insert;
2710 :
2711 : Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult);
2712 : pg_read_barrier();
2713 : Write = pg_atomic_read_u64(&XLogCtl->logWriteResult);
2714 : pg_read_barrier();
2715 : Insert = pg_atomic_read_u64(&XLogCtl->logInsertResult);
2716 :
2717 : /* WAL written to disk is always ahead of WAL flushed */
2718 : Assert(Write >= Flush);
2719 :
2720 : /* WAL inserted to buffers is always ahead of WAL written */
2721 : Assert(Insert >= Write);
2722 : }
2723 : #endif
2724 4152004 : }
2725 :
2726 : /*
2727 : * Record the LSN for an asynchronous transaction commit/abort
2728 : * and nudge the WALWriter if there is work for it to do.
2729 : * (This should not be called for synchronous commits.)
2730 : */
2731 : void
2732 99038 : XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2733 : {
2734 99038 : XLogRecPtr WriteRqstPtr = asyncXactLSN;
2735 : bool sleeping;
2736 99038 : bool wakeup = false;
2737 : XLogRecPtr prevAsyncXactLSN;
2738 :
2739 99038 : SpinLockAcquire(&XLogCtl->info_lck);
2740 99038 : sleeping = XLogCtl->WalWriterSleeping;
2741 99038 : prevAsyncXactLSN = XLogCtl->asyncXactLSN;
2742 99038 : if (XLogCtl->asyncXactLSN < asyncXactLSN)
2743 98134 : XLogCtl->asyncXactLSN = asyncXactLSN;
2744 99038 : SpinLockRelease(&XLogCtl->info_lck);
2745 :
2746 : /*
2747 : * If somebody else already called this function with a more aggressive
2748 : * LSN, they will have done what we needed (and perhaps more).
2749 : */
2750 99038 : if (asyncXactLSN <= prevAsyncXactLSN)
2751 904 : return;
2752 :
2753 : /*
2754 : * If the WALWriter is sleeping, kick it to make it come out of low-power
2755 : * mode, so that this async commit will reach disk within the expected
2756 : * amount of time. Otherwise, determine whether it has enough WAL
2757 : * available to flush, the same way that XLogBackgroundFlush() does.
2758 : */
2759 98134 : if (sleeping)
2760 60 : wakeup = true;
2761 : else
2762 : {
2763 : int flushblocks;
2764 :
2765 98074 : RefreshXLogWriteResult(LogwrtResult);
2766 :
2767 98074 : flushblocks =
2768 98074 : WriteRqstPtr / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
2769 :
2770 98074 : if (WalWriterFlushAfter == 0 || flushblocks >= WalWriterFlushAfter)
2771 7640 : wakeup = true;
2772 : }
2773 :
2774 98134 : if (wakeup)
2775 : {
2776 7700 : volatile PROC_HDR *procglobal = ProcGlobal;
2777 7700 : ProcNumber walwriterProc = procglobal->walwriterProc;
2778 :
2779 7700 : if (walwriterProc != INVALID_PROC_NUMBER)
2780 490 : SetLatch(&GetPGProcByNumber(walwriterProc)->procLatch);
2781 : }
2782 : }
2783 :
2784 : /*
2785 : * Record the LSN up to which we can remove WAL because it's not required by
2786 : * any replication slot.
2787 : */
2788 : void
2789 53492 : XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn)
2790 : {
2791 53492 : SpinLockAcquire(&XLogCtl->info_lck);
2792 53492 : XLogCtl->replicationSlotMinLSN = lsn;
2793 53492 : SpinLockRelease(&XLogCtl->info_lck);
2794 53492 : }
2795 :
2796 :
2797 : /*
2798 : * Return the oldest LSN we must retain to satisfy the needs of some
2799 : * replication slot.
2800 : */
2801 : static XLogRecPtr
2802 4334 : XLogGetReplicationSlotMinimumLSN(void)
2803 : {
2804 : XLogRecPtr retval;
2805 :
2806 4334 : SpinLockAcquire(&XLogCtl->info_lck);
2807 4334 : retval = XLogCtl->replicationSlotMinLSN;
2808 4334 : SpinLockRelease(&XLogCtl->info_lck);
2809 :
2810 4334 : return retval;
2811 : }
2812 :
2813 : /*
2814 : * Advance minRecoveryPoint in control file.
2815 : *
2816 : * If we crash during recovery, we must reach this point again before the
2817 : * database is consistent.
2818 : *
2819 : * If 'force' is true, 'lsn' argument is ignored. Otherwise, minRecoveryPoint
2820 : * is only updated if it's not already greater than or equal to 'lsn'.
2821 : */
2822 : static void
2823 207210 : UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2824 : {
2825 : /* Quick check using our local copy of the variable */
2826 207210 : if (!updateMinRecoveryPoint || (!force && lsn <= LocalMinRecoveryPoint))
2827 194038 : return;
2828 :
2829 : /*
2830 : * An invalid minRecoveryPoint means that we need to recover all the WAL,
2831 : * i.e., we're doing crash recovery. We never modify the control file's
2832 : * value in that case, so we can short-circuit future checks here too. The
2833 : * local values of minRecoveryPoint and minRecoveryPointTLI should not be
2834 : * updated until crash recovery finishes. We only do this for the startup
2835 : * process as it should not update its own reference of minRecoveryPoint
2836 : * until it has finished crash recovery to make sure that all WAL
2837 : * available is replayed in this case. This also saves from extra locks
2838 : * taken on the control file from the startup process.
2839 : */
2840 13172 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint) && InRecovery)
2841 : {
2842 62 : updateMinRecoveryPoint = false;
2843 62 : return;
2844 : }
2845 :
2846 13110 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2847 :
2848 : /* update local copy */
2849 13110 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
2850 13110 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2851 :
2852 13110 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint))
2853 4 : updateMinRecoveryPoint = false;
2854 13106 : else if (force || LocalMinRecoveryPoint < lsn)
2855 : {
2856 : XLogRecPtr newMinRecoveryPoint;
2857 : TimeLineID newMinRecoveryPointTLI;
2858 :
2859 : /*
2860 : * To avoid having to update the control file too often, we update it
2861 : * all the way to the last record being replayed, even though 'lsn'
2862 : * would suffice for correctness. This also allows the 'force' case
2863 : * to not need a valid 'lsn' value.
2864 : *
2865 : * Another important reason for doing it this way is that the passed
2866 : * 'lsn' value could be bogus, i.e., past the end of available WAL, if
2867 : * the caller got it from a corrupted heap page. Accepting such a
2868 : * value as the min recovery point would prevent us from coming up at
2869 : * all. Instead, we just log a warning and continue with recovery.
2870 : * (See also the comments about corrupt LSNs in XLogFlush.)
2871 : */
2872 10574 : newMinRecoveryPoint = GetCurrentReplayRecPtr(&newMinRecoveryPointTLI);
2873 10574 : if (!force && newMinRecoveryPoint < lsn)
2874 0 : elog(WARNING,
2875 : "xlog min recovery request %X/%08X is past current point %X/%08X",
2876 : LSN_FORMAT_ARGS(lsn), LSN_FORMAT_ARGS(newMinRecoveryPoint));
2877 :
2878 : /* update control file */
2879 10574 : if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2880 : {
2881 9876 : ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2882 9876 : ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2883 9876 : UpdateControlFile();
2884 9876 : LocalMinRecoveryPoint = newMinRecoveryPoint;
2885 9876 : LocalMinRecoveryPointTLI = newMinRecoveryPointTLI;
2886 :
2887 9876 : ereport(DEBUG2,
2888 : errmsg_internal("updated min recovery point to %X/%08X on timeline %u",
2889 : LSN_FORMAT_ARGS(newMinRecoveryPoint),
2890 : newMinRecoveryPointTLI));
2891 : }
2892 : }
2893 13110 : LWLockRelease(ControlFileLock);
2894 : }
2895 :
2896 : /*
2897 : * Ensure that all XLOG data through the given position is flushed to disk.
2898 : *
2899 : * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
2900 : * already held, and we try to avoid acquiring it if possible.
2901 : */
2902 : void
2903 1351430 : XLogFlush(XLogRecPtr record)
2904 : {
2905 : XLogRecPtr WriteRqstPtr;
2906 : XLogwrtRqst WriteRqst;
2907 1351430 : TimeLineID insertTLI = XLogCtl->InsertTimeLineID;
2908 :
2909 : /*
2910 : * During REDO, we are reading not writing WAL. Therefore, instead of
2911 : * trying to flush the WAL, we should update minRecoveryPoint instead. We
2912 : * test XLogInsertAllowed(), not InRecovery, because we need checkpointer
2913 : * to act this way too, and because when it tries to write the
2914 : * end-of-recovery checkpoint, it should indeed flush.
2915 : */
2916 1351430 : if (!XLogInsertAllowed())
2917 : {
2918 206328 : UpdateMinRecoveryPoint(record, false);
2919 1071968 : return;
2920 : }
2921 :
2922 : /* Quick exit if already known flushed */
2923 1145102 : if (record <= LogwrtResult.Flush)
2924 865640 : return;
2925 :
2926 : #ifdef WAL_DEBUG
2927 : if (XLOG_DEBUG)
2928 : elog(LOG, "xlog flush request %X/%08X; write %X/%08X; flush %X/%08X",
2929 : LSN_FORMAT_ARGS(record),
2930 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2931 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2932 : #endif
2933 :
2934 279462 : START_CRIT_SECTION();
2935 :
2936 : /*
2937 : * Since fsync is usually a horribly expensive operation, we try to
2938 : * piggyback as much data as we can on each fsync: if we see any more data
2939 : * entered into the xlog buffer, we'll write and fsync that too, so that
2940 : * the final value of LogwrtResult.Flush is as large as possible. This
2941 : * gives us some chance of avoiding another fsync immediately after.
2942 : */
2943 :
2944 : /* initialize to given target; may increase below */
2945 279462 : WriteRqstPtr = record;
2946 :
2947 : /*
2948 : * Now wait until we get the write lock, or someone else does the flush
2949 : * for us.
2950 : */
2951 : for (;;)
2952 10722 : {
2953 : XLogRecPtr insertpos;
2954 :
2955 : /* done already? */
2956 290184 : RefreshXLogWriteResult(LogwrtResult);
2957 290184 : if (record <= LogwrtResult.Flush)
2958 20296 : break;
2959 :
2960 : /*
2961 : * Before actually performing the write, wait for all in-flight
2962 : * insertions to the pages we're about to write to finish.
2963 : */
2964 269888 : SpinLockAcquire(&XLogCtl->info_lck);
2965 269888 : if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
2966 18344 : WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
2967 269888 : SpinLockRelease(&XLogCtl->info_lck);
2968 269888 : insertpos = WaitXLogInsertionsToFinish(WriteRqstPtr);
2969 :
2970 : /*
2971 : * Try to get the write lock. If we can't get it immediately, wait
2972 : * until it's released, and recheck if we still need to do the flush
2973 : * or if the backend that held the lock did it for us already. This
2974 : * helps to maintain a good rate of group committing when the system
2975 : * is bottlenecked by the speed of fsyncing.
2976 : */
2977 269888 : if (!LWLockAcquireOrWait(WALWriteLock, LW_EXCLUSIVE))
2978 : {
2979 : /*
2980 : * The lock is now free, but we didn't acquire it yet. Before we
2981 : * do, loop back to check if someone else flushed the record for
2982 : * us already.
2983 : */
2984 10722 : continue;
2985 : }
2986 :
2987 : /* Got the lock; recheck whether request is satisfied */
2988 259166 : RefreshXLogWriteResult(LogwrtResult);
2989 259166 : if (record <= LogwrtResult.Flush)
2990 : {
2991 3204 : LWLockRelease(WALWriteLock);
2992 3204 : break;
2993 : }
2994 :
2995 : /*
2996 : * Sleep before flush! By adding a delay here, we may give further
2997 : * backends the opportunity to join the backlog of group commit
2998 : * followers; this can significantly improve transaction throughput,
2999 : * at the risk of increasing transaction latency.
3000 : *
3001 : * We do not sleep if enableFsync is not turned on, nor if there are
3002 : * fewer than CommitSiblings other backends with active transactions.
3003 : */
3004 255962 : if (CommitDelay > 0 && enableFsync &&
3005 0 : MinimumActiveBackends(CommitSiblings))
3006 : {
3007 0 : pg_usleep(CommitDelay);
3008 :
3009 : /*
3010 : * Re-check how far we can now flush the WAL. It's generally not
3011 : * safe to call WaitXLogInsertionsToFinish while holding
3012 : * WALWriteLock, because an in-progress insertion might need to
3013 : * also grab WALWriteLock to make progress. But we know that all
3014 : * the insertions up to insertpos have already finished, because
3015 : * that's what the earlier WaitXLogInsertionsToFinish() returned.
3016 : * We're only calling it again to allow insertpos to be moved
3017 : * further forward, not to actually wait for anyone.
3018 : */
3019 0 : insertpos = WaitXLogInsertionsToFinish(insertpos);
3020 : }
3021 :
3022 : /* try to write/flush later additions to XLOG as well */
3023 255962 : WriteRqst.Write = insertpos;
3024 255962 : WriteRqst.Flush = insertpos;
3025 :
3026 255962 : XLogWrite(WriteRqst, insertTLI, false);
3027 :
3028 255962 : LWLockRelease(WALWriteLock);
3029 : /* done */
3030 255962 : break;
3031 : }
3032 :
3033 279462 : END_CRIT_SECTION();
3034 :
3035 : /* wake up walsenders now that we've released heavily contended locks */
3036 279462 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3037 :
3038 : /*
3039 : * If we still haven't flushed to the request point then we have a
3040 : * problem; most likely, the requested flush point is past end of XLOG.
3041 : * This has been seen to occur when a disk page has a corrupted LSN.
3042 : *
3043 : * Formerly we treated this as a PANIC condition, but that hurts the
3044 : * system's robustness rather than helping it: we do not want to take down
3045 : * the whole system due to corruption on one data page. In particular, if
3046 : * the bad page is encountered again during recovery then we would be
3047 : * unable to restart the database at all! (This scenario actually
3048 : * happened in the field several times with 7.1 releases.) As of 8.4, bad
3049 : * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
3050 : * the only time we can reach here during recovery is while flushing the
3051 : * end-of-recovery checkpoint record, and we don't expect that to have a
3052 : * bad LSN.
3053 : *
3054 : * Note that for calls from xact.c, the ERROR will be promoted to PANIC
3055 : * since xact.c calls this routine inside a critical section. However,
3056 : * calls from bufmgr.c are not within critical sections and so we will not
3057 : * force a restart for a bad LSN on a data page.
3058 : */
3059 279462 : if (LogwrtResult.Flush < record)
3060 0 : elog(ERROR,
3061 : "xlog flush request %X/%08X is not satisfied --- flushed only to %X/%08X",
3062 : LSN_FORMAT_ARGS(record),
3063 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
3064 : }
3065 :
3066 : /*
3067 : * Write & flush xlog, but without specifying exactly where to.
3068 : *
3069 : * We normally write only completed blocks; but if there is nothing to do on
3070 : * that basis, we check for unwritten async commits in the current incomplete
3071 : * block, and write through the latest one of those. Thus, if async commits
3072 : * are not being used, we will write complete blocks only.
3073 : *
3074 : * If, based on the above, there's anything to write we do so immediately. But
3075 : * to avoid calling fsync, fdatasync et. al. at a rate that'd impact
3076 : * concurrent IO, we only flush WAL every wal_writer_delay ms, or if there's
3077 : * more than wal_writer_flush_after unflushed blocks.
3078 : *
3079 : * We can guarantee that async commits reach disk after at most three
3080 : * wal_writer_delay cycles. (When flushing complete blocks, we allow XLogWrite
3081 : * to write "flexibly", meaning it can stop at the end of the buffer ring;
3082 : * this makes a difference only with very high load or long wal_writer_delay,
3083 : * but imposes one extra cycle for the worst case for async commits.)
3084 : *
3085 : * This routine is invoked periodically by the background walwriter process.
3086 : *
3087 : * Returns true if there was any work to do, even if we skipped flushing due
3088 : * to wal_writer_delay/wal_writer_flush_after.
3089 : */
3090 : bool
3091 36150 : XLogBackgroundFlush(void)
3092 : {
3093 : XLogwrtRqst WriteRqst;
3094 36150 : bool flexible = true;
3095 : static TimestampTz lastflush;
3096 : TimestampTz now;
3097 : int flushblocks;
3098 : TimeLineID insertTLI;
3099 :
3100 : /* XLOG doesn't need flushing during recovery */
3101 36150 : if (RecoveryInProgress())
3102 2184 : return false;
3103 :
3104 : /*
3105 : * Since we're not in recovery, InsertTimeLineID is set and can't change,
3106 : * so we can read it without a lock.
3107 : */
3108 33966 : insertTLI = XLogCtl->InsertTimeLineID;
3109 :
3110 : /* read updated LogwrtRqst */
3111 33966 : SpinLockAcquire(&XLogCtl->info_lck);
3112 33966 : WriteRqst = XLogCtl->LogwrtRqst;
3113 33966 : SpinLockRelease(&XLogCtl->info_lck);
3114 :
3115 : /* back off to last completed page boundary */
3116 33966 : WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
3117 :
3118 : /* if we have already flushed that far, consider async commit records */
3119 33966 : RefreshXLogWriteResult(LogwrtResult);
3120 33966 : if (WriteRqst.Write <= LogwrtResult.Flush)
3121 : {
3122 25536 : SpinLockAcquire(&XLogCtl->info_lck);
3123 25536 : WriteRqst.Write = XLogCtl->asyncXactLSN;
3124 25536 : SpinLockRelease(&XLogCtl->info_lck);
3125 25536 : flexible = false; /* ensure it all gets written */
3126 : }
3127 :
3128 : /*
3129 : * If already known flushed, we're done. Just need to check if we are
3130 : * holding an open file handle to a logfile that's no longer in use,
3131 : * preventing the file from being deleted.
3132 : */
3133 33966 : if (WriteRqst.Write <= LogwrtResult.Flush)
3134 : {
3135 23806 : if (openLogFile >= 0)
3136 : {
3137 17114 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
3138 : wal_segment_size))
3139 : {
3140 290 : XLogFileClose();
3141 : }
3142 : }
3143 23806 : return false;
3144 : }
3145 :
3146 : /*
3147 : * Determine how far to flush WAL, based on the wal_writer_delay and
3148 : * wal_writer_flush_after GUCs.
3149 : *
3150 : * Note that XLogSetAsyncXactLSN() performs similar calculation based on
3151 : * wal_writer_flush_after, to decide when to wake us up. Make sure the
3152 : * logic is the same in both places if you change this.
3153 : */
3154 10160 : now = GetCurrentTimestamp();
3155 10160 : flushblocks =
3156 10160 : WriteRqst.Write / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
3157 :
3158 10160 : if (WalWriterFlushAfter == 0 || lastflush == 0)
3159 : {
3160 : /* first call, or block based limits disabled */
3161 490 : WriteRqst.Flush = WriteRqst.Write;
3162 490 : lastflush = now;
3163 : }
3164 9670 : else if (TimestampDifferenceExceeds(lastflush, now, WalWriterDelay))
3165 : {
3166 : /*
3167 : * Flush the writes at least every WalWriterDelay ms. This is
3168 : * important to bound the amount of time it takes for an asynchronous
3169 : * commit to hit disk.
3170 : */
3171 9348 : WriteRqst.Flush = WriteRqst.Write;
3172 9348 : lastflush = now;
3173 : }
3174 322 : else if (flushblocks >= WalWriterFlushAfter)
3175 : {
3176 : /* exceeded wal_writer_flush_after blocks, flush */
3177 292 : WriteRqst.Flush = WriteRqst.Write;
3178 292 : lastflush = now;
3179 : }
3180 : else
3181 : {
3182 : /* no flushing, this time round */
3183 30 : WriteRqst.Flush = 0;
3184 : }
3185 :
3186 : #ifdef WAL_DEBUG
3187 : if (XLOG_DEBUG)
3188 : elog(LOG, "xlog bg flush request write %X/%08X; flush: %X/%08X, current is write %X/%08X; flush %X/%08X",
3189 : LSN_FORMAT_ARGS(WriteRqst.Write),
3190 : LSN_FORMAT_ARGS(WriteRqst.Flush),
3191 : LSN_FORMAT_ARGS(LogwrtResult.Write),
3192 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
3193 : #endif
3194 :
3195 10160 : START_CRIT_SECTION();
3196 :
3197 : /* now wait for any in-progress insertions to finish and get write lock */
3198 10160 : WaitXLogInsertionsToFinish(WriteRqst.Write);
3199 10160 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3200 10160 : RefreshXLogWriteResult(LogwrtResult);
3201 10160 : if (WriteRqst.Write > LogwrtResult.Write ||
3202 834 : WriteRqst.Flush > LogwrtResult.Flush)
3203 : {
3204 9628 : XLogWrite(WriteRqst, insertTLI, flexible);
3205 : }
3206 10160 : LWLockRelease(WALWriteLock);
3207 :
3208 10160 : END_CRIT_SECTION();
3209 :
3210 : /* wake up walsenders now that we've released heavily contended locks */
3211 10160 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3212 :
3213 : /*
3214 : * Great, done. To take some work off the critical path, try to initialize
3215 : * as many of the no-longer-needed WAL buffers for future use as we can.
3216 : */
3217 10160 : AdvanceXLInsertBuffer(InvalidXLogRecPtr, insertTLI, true);
3218 :
3219 : /*
3220 : * If we determined that we need to write data, but somebody else
3221 : * wrote/flushed already, it should be considered as being active, to
3222 : * avoid hibernating too early.
3223 : */
3224 10160 : return true;
3225 : }
3226 :
3227 : /*
3228 : * Test whether XLOG data has been flushed up to (at least) the given position.
3229 : *
3230 : * Returns true if a flush is still needed. (It may be that someone else
3231 : * is already in process of flushing that far, however.)
3232 : */
3233 : bool
3234 18227720 : XLogNeedsFlush(XLogRecPtr record)
3235 : {
3236 : /*
3237 : * During recovery, we don't flush WAL but update minRecoveryPoint
3238 : * instead. So "needs flush" is taken to mean whether minRecoveryPoint
3239 : * would need to be updated.
3240 : */
3241 18227720 : if (RecoveryInProgress())
3242 : {
3243 : /*
3244 : * An invalid minRecoveryPoint means that we need to recover all the
3245 : * WAL, i.e., we're doing crash recovery. We never modify the control
3246 : * file's value in that case, so we can short-circuit future checks
3247 : * here too. This triggers a quick exit path for the startup process,
3248 : * which cannot update its local copy of minRecoveryPoint as long as
3249 : * it has not replayed all WAL available when doing crash recovery.
3250 : */
3251 1277074 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint) && InRecovery)
3252 0 : updateMinRecoveryPoint = false;
3253 :
3254 : /* Quick exit if already known to be updated or cannot be updated */
3255 1277074 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3256 1258384 : return false;
3257 :
3258 : /*
3259 : * Update local copy of minRecoveryPoint. But if the lock is busy,
3260 : * just return a conservative guess.
3261 : */
3262 18690 : if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3263 0 : return true;
3264 18690 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
3265 18690 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3266 18690 : LWLockRelease(ControlFileLock);
3267 :
3268 : /*
3269 : * Check minRecoveryPoint for any other process than the startup
3270 : * process doing crash recovery, which should not update the control
3271 : * file value if crash recovery is still running.
3272 : */
3273 18690 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint))
3274 0 : updateMinRecoveryPoint = false;
3275 :
3276 : /* check again */
3277 18690 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3278 148 : return false;
3279 : else
3280 18542 : return true;
3281 : }
3282 :
3283 : /* Quick exit if already known flushed */
3284 16950646 : if (record <= LogwrtResult.Flush)
3285 16541940 : return false;
3286 :
3287 : /* read LogwrtResult and update local state */
3288 408706 : RefreshXLogWriteResult(LogwrtResult);
3289 :
3290 : /* check again */
3291 408706 : if (record <= LogwrtResult.Flush)
3292 5950 : return false;
3293 :
3294 402756 : return true;
3295 : }
3296 :
3297 : /*
3298 : * Try to make a given XLOG file segment exist.
3299 : *
3300 : * logsegno: identify segment.
3301 : *
3302 : * *added: on return, true if this call raised the number of extant segments.
3303 : *
3304 : * path: on return, this char[MAXPGPATH] has the path to the logsegno file.
3305 : *
3306 : * Returns -1 or FD of opened file. A -1 here is not an error; a caller
3307 : * wanting an open segment should attempt to open "path", which usually will
3308 : * succeed. (This is weird, but it's efficient for the callers.)
3309 : */
3310 : static int
3311 28868 : XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,
3312 : bool *added, char *path)
3313 : {
3314 : char tmppath[MAXPGPATH];
3315 : XLogSegNo installed_segno;
3316 : XLogSegNo max_segno;
3317 : int fd;
3318 : int save_errno;
3319 28868 : int open_flags = O_RDWR | O_CREAT | O_EXCL | PG_BINARY;
3320 : instr_time io_start;
3321 :
3322 : Assert(logtli != 0);
3323 :
3324 28868 : XLogFilePath(path, logtli, logsegno, wal_segment_size);
3325 :
3326 : /*
3327 : * Try to use existent file (checkpoint maker may have created it already)
3328 : */
3329 28868 : *added = false;
3330 28868 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3331 28868 : get_sync_bit(wal_sync_method));
3332 28868 : if (fd < 0)
3333 : {
3334 3002 : if (errno != ENOENT)
3335 0 : ereport(ERROR,
3336 : (errcode_for_file_access(),
3337 : errmsg("could not open file \"%s\": %m", path)));
3338 : }
3339 : else
3340 25866 : return fd;
3341 :
3342 : /*
3343 : * Initialize an empty (all zeroes) segment. NOTE: it is possible that
3344 : * another process is doing the same thing. If so, we will end up
3345 : * pre-creating an extra log segment. That seems OK, and better than
3346 : * holding the lock throughout this lengthy process.
3347 : */
3348 3002 : elog(DEBUG2, "creating and filling new WAL file");
3349 :
3350 3002 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3351 :
3352 3002 : unlink(tmppath);
3353 :
3354 3002 : if (io_direct_flags & IO_DIRECT_WAL_INIT)
3355 0 : open_flags |= PG_O_DIRECT;
3356 :
3357 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3358 3002 : fd = BasicOpenFile(tmppath, open_flags);
3359 3002 : if (fd < 0)
3360 0 : ereport(ERROR,
3361 : (errcode_for_file_access(),
3362 : errmsg("could not create file \"%s\": %m", tmppath)));
3363 :
3364 : /* Measure I/O timing when initializing segment */
3365 3002 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3366 :
3367 3002 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE);
3368 3002 : save_errno = 0;
3369 3002 : if (wal_init_zero)
3370 : {
3371 : ssize_t rc;
3372 :
3373 : /*
3374 : * Zero-fill the file. With this setting, we do this the hard way to
3375 : * ensure that all the file space has really been allocated. On
3376 : * platforms that allow "holes" in files, just seeking to the end
3377 : * doesn't allocate intermediate space. This way, we know that we
3378 : * have all the space and (after the fsync below) that all the
3379 : * indirect blocks are down on disk. Therefore, fdatasync(2) or
3380 : * O_DSYNC will be sufficient to sync future writes to the log file.
3381 : */
3382 3002 : rc = pg_pwrite_zeros(fd, wal_segment_size, 0);
3383 :
3384 3002 : if (rc < 0)
3385 0 : save_errno = errno;
3386 : }
3387 : else
3388 : {
3389 : /*
3390 : * Otherwise, seeking to the end and writing a solitary byte is
3391 : * enough.
3392 : */
3393 0 : errno = 0;
3394 0 : if (pg_pwrite(fd, "\0", 1, wal_segment_size - 1) != 1)
3395 : {
3396 : /* if write didn't set errno, assume no disk space */
3397 0 : save_errno = errno ? errno : ENOSPC;
3398 : }
3399 : }
3400 3002 : pgstat_report_wait_end();
3401 :
3402 : /*
3403 : * A full segment worth of data is written when using wal_init_zero. One
3404 : * byte is written when not using it.
3405 : */
3406 3002 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT, IOOP_WRITE,
3407 : io_start, 1,
3408 3002 : wal_init_zero ? wal_segment_size : 1);
3409 :
3410 3002 : if (save_errno)
3411 : {
3412 : /*
3413 : * If we fail to make the file, delete it to release disk space
3414 : */
3415 0 : unlink(tmppath);
3416 :
3417 0 : close(fd);
3418 :
3419 0 : errno = save_errno;
3420 :
3421 0 : ereport(ERROR,
3422 : (errcode_for_file_access(),
3423 : errmsg("could not write to file \"%s\": %m", tmppath)));
3424 : }
3425 :
3426 : /* Measure I/O timing when flushing segment */
3427 3002 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3428 :
3429 3002 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC);
3430 3002 : if (pg_fsync(fd) != 0)
3431 : {
3432 0 : save_errno = errno;
3433 0 : close(fd);
3434 0 : errno = save_errno;
3435 0 : ereport(ERROR,
3436 : (errcode_for_file_access(),
3437 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3438 : }
3439 3002 : pgstat_report_wait_end();
3440 :
3441 3002 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT,
3442 : IOOP_FSYNC, io_start, 1, 0);
3443 :
3444 3002 : if (close(fd) != 0)
3445 0 : ereport(ERROR,
3446 : (errcode_for_file_access(),
3447 : errmsg("could not close file \"%s\": %m", tmppath)));
3448 :
3449 : /*
3450 : * Now move the segment into place with its final name. Cope with
3451 : * possibility that someone else has created the file while we were
3452 : * filling ours: if so, use ours to pre-create a future log segment.
3453 : */
3454 3002 : installed_segno = logsegno;
3455 :
3456 : /*
3457 : * XXX: What should we use as max_segno? We used to use XLOGfileslop when
3458 : * that was a constant, but that was always a bit dubious: normally, at a
3459 : * checkpoint, XLOGfileslop was the offset from the checkpoint record, but
3460 : * here, it was the offset from the insert location. We can't do the
3461 : * normal XLOGfileslop calculation here because we don't have access to
3462 : * the prior checkpoint's redo location. So somewhat arbitrarily, just use
3463 : * CheckPointSegments.
3464 : */
3465 3002 : max_segno = logsegno + CheckPointSegments;
3466 3002 : if (InstallXLogFileSegment(&installed_segno, tmppath, true, max_segno,
3467 : logtli))
3468 : {
3469 3002 : *added = true;
3470 3002 : elog(DEBUG2, "done creating and filling new WAL file");
3471 : }
3472 : else
3473 : {
3474 : /*
3475 : * No need for any more future segments, or InstallXLogFileSegment()
3476 : * failed to rename the file into place. If the rename failed, a
3477 : * caller opening the file may fail.
3478 : */
3479 0 : unlink(tmppath);
3480 0 : elog(DEBUG2, "abandoned new WAL file");
3481 : }
3482 :
3483 3002 : return -1;
3484 : }
3485 :
3486 : /*
3487 : * Create a new XLOG file segment, or open a pre-existing one.
3488 : *
3489 : * logsegno: identify segment to be created/opened.
3490 : *
3491 : * Returns FD of opened file.
3492 : *
3493 : * Note: errors here are ERROR not PANIC because we might or might not be
3494 : * inside a critical section (eg, during checkpoint there is no reason to
3495 : * take down the system on failure). They will promote to PANIC if we are
3496 : * in a critical section.
3497 : */
3498 : int
3499 28332 : XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)
3500 : {
3501 : bool ignore_added;
3502 : char path[MAXPGPATH];
3503 : int fd;
3504 :
3505 : Assert(logtli != 0);
3506 :
3507 28332 : fd = XLogFileInitInternal(logsegno, logtli, &ignore_added, path);
3508 28332 : if (fd >= 0)
3509 25622 : return fd;
3510 :
3511 : /* Now open original target segment (might not be file I just made) */
3512 2710 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3513 2710 : get_sync_bit(wal_sync_method));
3514 2710 : if (fd < 0)
3515 0 : ereport(ERROR,
3516 : (errcode_for_file_access(),
3517 : errmsg("could not open file \"%s\": %m", path)));
3518 2710 : return fd;
3519 : }
3520 :
3521 : /*
3522 : * Create a new XLOG file segment by copying a pre-existing one.
3523 : *
3524 : * destsegno: identify segment to be created.
3525 : *
3526 : * srcTLI, srcsegno: identify segment to be copied (could be from
3527 : * a different timeline)
3528 : *
3529 : * upto: how much of the source file to copy (the rest is filled with
3530 : * zeros)
3531 : *
3532 : * Currently this is only used during recovery, and so there are no locking
3533 : * considerations. But we should be just as tense as XLogFileInit to avoid
3534 : * emplacing a bogus file.
3535 : */
3536 : static void
3537 74 : XLogFileCopy(TimeLineID destTLI, XLogSegNo destsegno,
3538 : TimeLineID srcTLI, XLogSegNo srcsegno,
3539 : int upto)
3540 : {
3541 : char path[MAXPGPATH];
3542 : char tmppath[MAXPGPATH];
3543 : PGAlignedXLogBlock buffer;
3544 : int srcfd;
3545 : int fd;
3546 : int nbytes;
3547 :
3548 : /*
3549 : * Open the source file
3550 : */
3551 74 : XLogFilePath(path, srcTLI, srcsegno, wal_segment_size);
3552 74 : srcfd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
3553 74 : if (srcfd < 0)
3554 0 : ereport(ERROR,
3555 : (errcode_for_file_access(),
3556 : errmsg("could not open file \"%s\": %m", path)));
3557 :
3558 : /*
3559 : * Copy into a temp file name.
3560 : */
3561 74 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3562 :
3563 74 : unlink(tmppath);
3564 :
3565 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3566 74 : fd = OpenTransientFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
3567 74 : if (fd < 0)
3568 0 : ereport(ERROR,
3569 : (errcode_for_file_access(),
3570 : errmsg("could not create file \"%s\": %m", tmppath)));
3571 :
3572 : /*
3573 : * Do the data copying.
3574 : */
3575 151626 : for (nbytes = 0; nbytes < wal_segment_size; nbytes += sizeof(buffer))
3576 : {
3577 : int nread;
3578 :
3579 151552 : nread = upto - nbytes;
3580 :
3581 : /*
3582 : * The part that is not read from the source file is filled with
3583 : * zeros.
3584 : */
3585 151552 : if (nread < sizeof(buffer))
3586 74 : memset(buffer.data, 0, sizeof(buffer));
3587 :
3588 151552 : if (nread > 0)
3589 : {
3590 : int r;
3591 :
3592 5382 : if (nread > sizeof(buffer))
3593 5308 : nread = sizeof(buffer);
3594 5382 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ);
3595 5382 : r = read(srcfd, buffer.data, nread);
3596 5382 : if (r != nread)
3597 : {
3598 0 : if (r < 0)
3599 0 : ereport(ERROR,
3600 : (errcode_for_file_access(),
3601 : errmsg("could not read file \"%s\": %m",
3602 : path)));
3603 : else
3604 0 : ereport(ERROR,
3605 : (errcode(ERRCODE_DATA_CORRUPTED),
3606 : errmsg("could not read file \"%s\": read %d of %zu",
3607 : path, r, (Size) nread)));
3608 : }
3609 5382 : pgstat_report_wait_end();
3610 : }
3611 151552 : errno = 0;
3612 151552 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_WRITE);
3613 151552 : if ((int) write(fd, buffer.data, sizeof(buffer)) != (int) sizeof(buffer))
3614 : {
3615 0 : int save_errno = errno;
3616 :
3617 : /*
3618 : * If we fail to make the file, delete it to release disk space
3619 : */
3620 0 : unlink(tmppath);
3621 : /* if write didn't set errno, assume problem is no disk space */
3622 0 : errno = save_errno ? save_errno : ENOSPC;
3623 :
3624 0 : ereport(ERROR,
3625 : (errcode_for_file_access(),
3626 : errmsg("could not write to file \"%s\": %m", tmppath)));
3627 : }
3628 151552 : pgstat_report_wait_end();
3629 : }
3630 :
3631 74 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_SYNC);
3632 74 : if (pg_fsync(fd) != 0)
3633 0 : ereport(data_sync_elevel(ERROR),
3634 : (errcode_for_file_access(),
3635 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3636 74 : pgstat_report_wait_end();
3637 :
3638 74 : if (CloseTransientFile(fd) != 0)
3639 0 : ereport(ERROR,
3640 : (errcode_for_file_access(),
3641 : errmsg("could not close file \"%s\": %m", tmppath)));
3642 :
3643 74 : if (CloseTransientFile(srcfd) != 0)
3644 0 : ereport(ERROR,
3645 : (errcode_for_file_access(),
3646 : errmsg("could not close file \"%s\": %m", path)));
3647 :
3648 : /*
3649 : * Now move the segment into place with its final name.
3650 : */
3651 74 : if (!InstallXLogFileSegment(&destsegno, tmppath, false, 0, destTLI))
3652 0 : elog(ERROR, "InstallXLogFileSegment should not have failed");
3653 74 : }
3654 :
3655 : /*
3656 : * Install a new XLOG segment file as a current or future log segment.
3657 : *
3658 : * This is used both to install a newly-created segment (which has a temp
3659 : * filename while it's being created) and to recycle an old segment.
3660 : *
3661 : * *segno: identify segment to install as (or first possible target).
3662 : * When find_free is true, this is modified on return to indicate the
3663 : * actual installation location or last segment searched.
3664 : *
3665 : * tmppath: initial name of file to install. It will be renamed into place.
3666 : *
3667 : * find_free: if true, install the new segment at the first empty segno
3668 : * number at or after the passed numbers. If false, install the new segment
3669 : * exactly where specified, deleting any existing segment file there.
3670 : *
3671 : * max_segno: maximum segment number to install the new file as. Fail if no
3672 : * free slot is found between *segno and max_segno. (Ignored when find_free
3673 : * is false.)
3674 : *
3675 : * tli: The timeline on which the new segment should be installed.
3676 : *
3677 : * Returns true if the file was installed successfully. false indicates that
3678 : * max_segno limit was exceeded, the startup process has disabled this
3679 : * function for now, or an error occurred while renaming the file into place.
3680 : */
3681 : static bool
3682 5818 : InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
3683 : bool find_free, XLogSegNo max_segno, TimeLineID tli)
3684 : {
3685 : char path[MAXPGPATH];
3686 : struct stat stat_buf;
3687 :
3688 : Assert(tli != 0);
3689 :
3690 5818 : XLogFilePath(path, tli, *segno, wal_segment_size);
3691 :
3692 5818 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3693 5818 : if (!XLogCtl->InstallXLogFileSegmentActive)
3694 : {
3695 0 : LWLockRelease(ControlFileLock);
3696 0 : return false;
3697 : }
3698 :
3699 5818 : if (!find_free)
3700 : {
3701 : /* Force installation: get rid of any pre-existing segment file */
3702 74 : durable_unlink(path, DEBUG1);
3703 : }
3704 : else
3705 : {
3706 : /* Find a free slot to put it in */
3707 8382 : while (stat(path, &stat_buf) == 0)
3708 : {
3709 2688 : if ((*segno) >= max_segno)
3710 : {
3711 : /* Failed to find a free slot within specified range */
3712 50 : LWLockRelease(ControlFileLock);
3713 50 : return false;
3714 : }
3715 2638 : (*segno)++;
3716 2638 : XLogFilePath(path, tli, *segno, wal_segment_size);
3717 : }
3718 : }
3719 :
3720 : Assert(access(path, F_OK) != 0 && errno == ENOENT);
3721 5768 : if (durable_rename(tmppath, path, LOG) != 0)
3722 : {
3723 0 : LWLockRelease(ControlFileLock);
3724 : /* durable_rename already emitted log message */
3725 0 : return false;
3726 : }
3727 :
3728 5768 : LWLockRelease(ControlFileLock);
3729 :
3730 5768 : return true;
3731 : }
3732 :
3733 : /*
3734 : * Open a pre-existing logfile segment for writing.
3735 : */
3736 : int
3737 210 : XLogFileOpen(XLogSegNo segno, TimeLineID tli)
3738 : {
3739 : char path[MAXPGPATH];
3740 : int fd;
3741 :
3742 210 : XLogFilePath(path, tli, segno, wal_segment_size);
3743 :
3744 210 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3745 210 : get_sync_bit(wal_sync_method));
3746 210 : if (fd < 0)
3747 0 : ereport(PANIC,
3748 : (errcode_for_file_access(),
3749 : errmsg("could not open file \"%s\": %m", path)));
3750 :
3751 210 : return fd;
3752 : }
3753 :
3754 : /*
3755 : * Close the current logfile segment for writing.
3756 : */
3757 : static void
3758 12734 : XLogFileClose(void)
3759 : {
3760 : Assert(openLogFile >= 0);
3761 :
3762 : /*
3763 : * WAL segment files will not be re-read in normal operation, so we advise
3764 : * the OS to release any cached pages. But do not do so if WAL archiving
3765 : * or streaming is active, because archiver and walsender process could
3766 : * use the cache to read the WAL segment.
3767 : */
3768 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3769 12734 : if (!XLogIsNeeded() && (io_direct_flags & IO_DIRECT_WAL) == 0)
3770 2992 : (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3771 : #endif
3772 :
3773 12734 : if (close(openLogFile) != 0)
3774 : {
3775 : char xlogfname[MAXFNAMELEN];
3776 0 : int save_errno = errno;
3777 :
3778 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo, wal_segment_size);
3779 0 : errno = save_errno;
3780 0 : ereport(PANIC,
3781 : (errcode_for_file_access(),
3782 : errmsg("could not close file \"%s\": %m", xlogfname)));
3783 : }
3784 :
3785 12734 : openLogFile = -1;
3786 12734 : ReleaseExternalFD();
3787 12734 : }
3788 :
3789 : /*
3790 : * Preallocate log files beyond the specified log endpoint.
3791 : *
3792 : * XXX this is currently extremely conservative, since it forces only one
3793 : * future log segment to exist, and even that only if we are 75% done with
3794 : * the current one. This is only appropriate for very low-WAL-volume systems.
3795 : * High-volume systems will be OK once they've built up a sufficient set of
3796 : * recycled log segments, but the startup transient is likely to include
3797 : * a lot of segment creations by foreground processes, which is not so good.
3798 : *
3799 : * XLogFileInitInternal() can ereport(ERROR). All known causes indicate big
3800 : * trouble; for example, a full filesystem is one cause. The checkpoint WAL
3801 : * and/or ControlFile updates already completed. If a RequestCheckpoint()
3802 : * initiated the present checkpoint and an ERROR ends this function, the
3803 : * command that called RequestCheckpoint() fails. That's not ideal, but it's
3804 : * not worth contorting more functions to use caller-specified elevel values.
3805 : * (With or without RequestCheckpoint(), an ERROR forestalls some inessential
3806 : * reporting and resource reclamation.)
3807 : */
3808 : static void
3809 3922 : PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli)
3810 : {
3811 : XLogSegNo _logSegNo;
3812 : int lf;
3813 : bool added;
3814 : char path[MAXPGPATH];
3815 : uint64 offset;
3816 :
3817 3922 : if (!XLogCtl->InstallXLogFileSegmentActive)
3818 18 : return; /* unlocked check says no */
3819 :
3820 3904 : XLByteToPrevSeg(endptr, _logSegNo, wal_segment_size);
3821 3904 : offset = XLogSegmentOffset(endptr - 1, wal_segment_size);
3822 3904 : if (offset >= (uint32) (0.75 * wal_segment_size))
3823 : {
3824 536 : _logSegNo++;
3825 536 : lf = XLogFileInitInternal(_logSegNo, tli, &added, path);
3826 536 : if (lf >= 0)
3827 244 : close(lf);
3828 536 : if (added)
3829 292 : CheckpointStats.ckpt_segs_added++;
3830 : }
3831 : }
3832 :
3833 : /*
3834 : * Throws an error if the given log segment has already been removed or
3835 : * recycled. The caller should only pass a segment that it knows to have
3836 : * existed while the server has been running, as this function always
3837 : * succeeds if no WAL segments have been removed since startup.
3838 : * 'tli' is only used in the error message.
3839 : *
3840 : * Note: this function guarantees to keep errno unchanged on return.
3841 : * This supports callers that use this to possibly deliver a better
3842 : * error message about a missing file, while still being able to throw
3843 : * a normal file-access error afterwards, if this does return.
3844 : */
3845 : void
3846 229990 : CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3847 : {
3848 229990 : int save_errno = errno;
3849 : XLogSegNo lastRemovedSegNo;
3850 :
3851 229990 : SpinLockAcquire(&XLogCtl->info_lck);
3852 229990 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3853 229990 : SpinLockRelease(&XLogCtl->info_lck);
3854 :
3855 229990 : if (segno <= lastRemovedSegNo)
3856 : {
3857 : char filename[MAXFNAMELEN];
3858 :
3859 0 : XLogFileName(filename, tli, segno, wal_segment_size);
3860 0 : errno = save_errno;
3861 0 : ereport(ERROR,
3862 : (errcode_for_file_access(),
3863 : errmsg("requested WAL segment %s has already been removed",
3864 : filename)));
3865 : }
3866 229990 : errno = save_errno;
3867 229990 : }
3868 :
3869 : /*
3870 : * Return the last WAL segment removed, or 0 if no segment has been removed
3871 : * since startup.
3872 : *
3873 : * NB: the result can be out of date arbitrarily fast, the caller has to deal
3874 : * with that.
3875 : */
3876 : XLogSegNo
3877 2148 : XLogGetLastRemovedSegno(void)
3878 : {
3879 : XLogSegNo lastRemovedSegNo;
3880 :
3881 2148 : SpinLockAcquire(&XLogCtl->info_lck);
3882 2148 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3883 2148 : SpinLockRelease(&XLogCtl->info_lck);
3884 :
3885 2148 : return lastRemovedSegNo;
3886 : }
3887 :
3888 : /*
3889 : * Return the oldest WAL segment on the given TLI that still exists in
3890 : * XLOGDIR, or 0 if none.
3891 : */
3892 : XLogSegNo
3893 10 : XLogGetOldestSegno(TimeLineID tli)
3894 : {
3895 : DIR *xldir;
3896 : struct dirent *xlde;
3897 10 : XLogSegNo oldest_segno = 0;
3898 :
3899 10 : xldir = AllocateDir(XLOGDIR);
3900 66 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3901 : {
3902 : TimeLineID file_tli;
3903 : XLogSegNo file_segno;
3904 :
3905 : /* Ignore files that are not XLOG segments. */
3906 56 : if (!IsXLogFileName(xlde->d_name))
3907 40 : continue;
3908 :
3909 : /* Parse filename to get TLI and segno. */
3910 16 : XLogFromFileName(xlde->d_name, &file_tli, &file_segno,
3911 : wal_segment_size);
3912 :
3913 : /* Ignore anything that's not from the TLI of interest. */
3914 16 : if (tli != file_tli)
3915 0 : continue;
3916 :
3917 : /* If it's the oldest so far, update oldest_segno. */
3918 16 : if (oldest_segno == 0 || file_segno < oldest_segno)
3919 14 : oldest_segno = file_segno;
3920 : }
3921 :
3922 10 : FreeDir(xldir);
3923 10 : return oldest_segno;
3924 : }
3925 :
3926 : /*
3927 : * Update the last removed segno pointer in shared memory, to reflect that the
3928 : * given XLOG file has been removed.
3929 : */
3930 : static void
3931 5030 : UpdateLastRemovedPtr(char *filename)
3932 : {
3933 : uint32 tli;
3934 : XLogSegNo segno;
3935 :
3936 5030 : XLogFromFileName(filename, &tli, &segno, wal_segment_size);
3937 :
3938 5030 : SpinLockAcquire(&XLogCtl->info_lck);
3939 5030 : if (segno > XLogCtl->lastRemovedSegNo)
3940 2098 : XLogCtl->lastRemovedSegNo = segno;
3941 5030 : SpinLockRelease(&XLogCtl->info_lck);
3942 5030 : }
3943 :
3944 : /*
3945 : * Remove all temporary log files in pg_wal
3946 : *
3947 : * This is called at the beginning of recovery after a previous crash,
3948 : * at a point where no other processes write fresh WAL data.
3949 : */
3950 : static void
3951 350 : RemoveTempXlogFiles(void)
3952 : {
3953 : DIR *xldir;
3954 : struct dirent *xlde;
3955 :
3956 350 : elog(DEBUG2, "removing all temporary WAL segments");
3957 :
3958 350 : xldir = AllocateDir(XLOGDIR);
3959 2384 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3960 : {
3961 : char path[MAXPGPATH];
3962 :
3963 2034 : if (strncmp(xlde->d_name, "xlogtemp.", 9) != 0)
3964 2034 : continue;
3965 :
3966 0 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
3967 0 : unlink(path);
3968 0 : elog(DEBUG2, "removed temporary WAL segment \"%s\"", path);
3969 : }
3970 350 : FreeDir(xldir);
3971 350 : }
3972 :
3973 : /*
3974 : * Recycle or remove all log files older or equal to passed segno.
3975 : *
3976 : * endptr is current (or recent) end of xlog, and lastredoptr is the
3977 : * redo pointer of the last checkpoint. These are used to determine
3978 : * whether we want to recycle rather than delete no-longer-wanted log files.
3979 : *
3980 : * insertTLI is the current timeline for XLOG insertion. Any recycled
3981 : * segments should be reused for this timeline.
3982 : */
3983 : static void
3984 3378 : RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr, XLogRecPtr endptr,
3985 : TimeLineID insertTLI)
3986 : {
3987 : DIR *xldir;
3988 : struct dirent *xlde;
3989 : char lastoff[MAXFNAMELEN];
3990 : XLogSegNo endlogSegNo;
3991 : XLogSegNo recycleSegNo;
3992 :
3993 : /* Initialize info about where to try to recycle to */
3994 3378 : XLByteToSeg(endptr, endlogSegNo, wal_segment_size);
3995 3378 : recycleSegNo = XLOGfileslop(lastredoptr);
3996 :
3997 : /*
3998 : * Construct a filename of the last segment to be kept. The timeline ID
3999 : * doesn't matter, we ignore that in the comparison. (During recovery,
4000 : * InsertTimeLineID isn't set, so we can't use that.)
4001 : */
4002 3378 : XLogFileName(lastoff, 0, segno, wal_segment_size);
4003 :
4004 3378 : elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
4005 : lastoff);
4006 :
4007 3378 : xldir = AllocateDir(XLOGDIR);
4008 :
4009 87118 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4010 : {
4011 : /* Ignore files that are not XLOG segments */
4012 83740 : if (!IsXLogFileName(xlde->d_name) &&
4013 14596 : !IsPartialXLogFileName(xlde->d_name))
4014 14592 : continue;
4015 :
4016 : /*
4017 : * We ignore the timeline part of the XLOG segment identifiers in
4018 : * deciding whether a segment is still needed. This ensures that we
4019 : * won't prematurely remove a segment from a parent timeline. We could
4020 : * probably be a little more proactive about removing segments of
4021 : * non-parent timelines, but that would be a whole lot more
4022 : * complicated.
4023 : *
4024 : * We use the alphanumeric sorting property of the filenames to decide
4025 : * which ones are earlier than the lastoff segment.
4026 : */
4027 69148 : if (strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
4028 : {
4029 45528 : if (XLogArchiveCheckDone(xlde->d_name))
4030 : {
4031 : /* Update the last removed location in shared memory first */
4032 5030 : UpdateLastRemovedPtr(xlde->d_name);
4033 :
4034 5030 : RemoveXlogFile(xlde, recycleSegNo, &endlogSegNo, insertTLI);
4035 : }
4036 : }
4037 : }
4038 :
4039 3378 : FreeDir(xldir);
4040 3378 : }
4041 :
4042 : /*
4043 : * Recycle or remove WAL files that are not part of the given timeline's
4044 : * history.
4045 : *
4046 : * This is called during recovery, whenever we switch to follow a new
4047 : * timeline, and at the end of recovery when we create a new timeline. We
4048 : * wouldn't otherwise care about extra WAL files lying in pg_wal, but they
4049 : * might be leftover pre-allocated or recycled WAL segments on the old timeline
4050 : * that we haven't used yet, and contain garbage. If we just leave them in
4051 : * pg_wal, they will eventually be archived, and we can't let that happen.
4052 : * Files that belong to our timeline history are valid, because we have
4053 : * successfully replayed them, but from others we can't be sure.
4054 : *
4055 : * 'switchpoint' is the current point in WAL where we switch to new timeline,
4056 : * and 'newTLI' is the new timeline we switch to.
4057 : */
4058 : void
4059 118 : RemoveNonParentXlogFiles(XLogRecPtr switchpoint, TimeLineID newTLI)
4060 : {
4061 : DIR *xldir;
4062 : struct dirent *xlde;
4063 : char switchseg[MAXFNAMELEN];
4064 : XLogSegNo endLogSegNo;
4065 : XLogSegNo switchLogSegNo;
4066 : XLogSegNo recycleSegNo;
4067 :
4068 : /*
4069 : * Initialize info about where to begin the work. This will recycle,
4070 : * somewhat arbitrarily, 10 future segments.
4071 : */
4072 118 : XLByteToPrevSeg(switchpoint, switchLogSegNo, wal_segment_size);
4073 118 : XLByteToSeg(switchpoint, endLogSegNo, wal_segment_size);
4074 118 : recycleSegNo = endLogSegNo + 10;
4075 :
4076 : /*
4077 : * Construct a filename of the last segment to be kept.
4078 : */
4079 118 : XLogFileName(switchseg, newTLI, switchLogSegNo, wal_segment_size);
4080 :
4081 118 : elog(DEBUG2, "attempting to remove WAL segments newer than log file %s",
4082 : switchseg);
4083 :
4084 118 : xldir = AllocateDir(XLOGDIR);
4085 :
4086 1120 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4087 : {
4088 : /* Ignore files that are not XLOG segments */
4089 1002 : if (!IsXLogFileName(xlde->d_name))
4090 626 : continue;
4091 :
4092 : /*
4093 : * Remove files that are on a timeline older than the new one we're
4094 : * switching to, but with a segment number >= the first segment on the
4095 : * new timeline.
4096 : */
4097 376 : if (strncmp(xlde->d_name, switchseg, 8) < 0 &&
4098 240 : strcmp(xlde->d_name + 8, switchseg + 8) > 0)
4099 : {
4100 : /*
4101 : * If the file has already been marked as .ready, however, don't
4102 : * remove it yet. It should be OK to remove it - files that are
4103 : * not part of our timeline history are not required for recovery
4104 : * - but seems safer to let them be archived and removed later.
4105 : */
4106 32 : if (!XLogArchiveIsReady(xlde->d_name))
4107 32 : RemoveXlogFile(xlde, recycleSegNo, &endLogSegNo, newTLI);
4108 : }
4109 : }
4110 :
4111 118 : FreeDir(xldir);
4112 118 : }
4113 :
4114 : /*
4115 : * Recycle or remove a log file that's no longer needed.
4116 : *
4117 : * segment_de is the dirent structure of the segment to recycle or remove.
4118 : * recycleSegNo is the segment number to recycle up to. endlogSegNo is
4119 : * the segment number of the current (or recent) end of WAL.
4120 : *
4121 : * endlogSegNo gets incremented if the segment is recycled so as it is not
4122 : * checked again with future callers of this function.
4123 : *
4124 : * insertTLI is the current timeline for XLOG insertion. Any recycled segments
4125 : * should be used for this timeline.
4126 : */
4127 : static void
4128 5062 : RemoveXlogFile(const struct dirent *segment_de,
4129 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
4130 : TimeLineID insertTLI)
4131 : {
4132 : char path[MAXPGPATH];
4133 : #ifdef WIN32
4134 : char newpath[MAXPGPATH];
4135 : #endif
4136 5062 : const char *segname = segment_de->d_name;
4137 :
4138 5062 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", segname);
4139 :
4140 : /*
4141 : * Before deleting the file, see if it can be recycled as a future log
4142 : * segment. Only recycle normal files, because we don't want to recycle
4143 : * symbolic links pointing to a separate archive directory.
4144 : */
4145 5062 : if (wal_recycle &&
4146 5062 : *endlogSegNo <= recycleSegNo &&
4147 6136 : XLogCtl->InstallXLogFileSegmentActive && /* callee rechecks this */
4148 5484 : get_dirent_type(path, segment_de, false, DEBUG2) == PGFILETYPE_REG &&
4149 2742 : InstallXLogFileSegment(endlogSegNo, path,
4150 : true, recycleSegNo, insertTLI))
4151 : {
4152 2692 : ereport(DEBUG2,
4153 : (errmsg_internal("recycled write-ahead log file \"%s\"",
4154 : segname)));
4155 2692 : CheckpointStats.ckpt_segs_recycled++;
4156 : /* Needn't recheck that slot on future iterations */
4157 2692 : (*endlogSegNo)++;
4158 : }
4159 : else
4160 : {
4161 : /* No need for any more future segments, or recycling failed ... */
4162 : int rc;
4163 :
4164 2370 : ereport(DEBUG2,
4165 : (errmsg_internal("removing write-ahead log file \"%s\"",
4166 : segname)));
4167 :
4168 : #ifdef WIN32
4169 :
4170 : /*
4171 : * On Windows, if another process (e.g another backend) holds the file
4172 : * open in FILE_SHARE_DELETE mode, unlink will succeed, but the file
4173 : * will still show up in directory listing until the last handle is
4174 : * closed. To avoid confusing the lingering deleted file for a live
4175 : * WAL file that needs to be archived, rename it before deleting it.
4176 : *
4177 : * If another process holds the file open without FILE_SHARE_DELETE
4178 : * flag, rename will fail. We'll try again at the next checkpoint.
4179 : */
4180 : snprintf(newpath, MAXPGPATH, "%s.deleted", path);
4181 : if (rename(path, newpath) != 0)
4182 : {
4183 : ereport(LOG,
4184 : (errcode_for_file_access(),
4185 : errmsg("could not rename file \"%s\": %m",
4186 : path)));
4187 : return;
4188 : }
4189 : rc = durable_unlink(newpath, LOG);
4190 : #else
4191 2370 : rc = durable_unlink(path, LOG);
4192 : #endif
4193 2370 : if (rc != 0)
4194 : {
4195 : /* Message already logged by durable_unlink() */
4196 0 : return;
4197 : }
4198 2370 : CheckpointStats.ckpt_segs_removed++;
4199 : }
4200 :
4201 5062 : XLogArchiveCleanup(segname);
4202 : }
4203 :
4204 : /*
4205 : * Verify whether pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
4206 : * If the latter do not exist, recreate them.
4207 : *
4208 : * It is not the goal of this function to verify the contents of these
4209 : * directories, but to help in cases where someone has performed a cluster
4210 : * copy for PITR purposes but omitted pg_wal from the copy.
4211 : *
4212 : * We could also recreate pg_wal if it doesn't exist, but a deliberate
4213 : * policy decision was made not to. It is fairly common for pg_wal to be
4214 : * a symlink, and if that was the DBA's intent then automatically making a
4215 : * plain directory would result in degraded performance with no notice.
4216 : */
4217 : static void
4218 1864 : ValidateXLOGDirectoryStructure(void)
4219 : {
4220 : char path[MAXPGPATH];
4221 : struct stat stat_buf;
4222 :
4223 : /* Check for pg_wal; if it doesn't exist, error out */
4224 1864 : if (stat(XLOGDIR, &stat_buf) != 0 ||
4225 1864 : !S_ISDIR(stat_buf.st_mode))
4226 0 : ereport(FATAL,
4227 : (errcode_for_file_access(),
4228 : errmsg("required WAL directory \"%s\" does not exist",
4229 : XLOGDIR)));
4230 :
4231 : /* Check for archive_status */
4232 1864 : snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4233 1864 : if (stat(path, &stat_buf) == 0)
4234 : {
4235 : /* Check for weird cases where it exists but isn't a directory */
4236 1862 : if (!S_ISDIR(stat_buf.st_mode))
4237 0 : ereport(FATAL,
4238 : (errcode_for_file_access(),
4239 : errmsg("required WAL directory \"%s\" does not exist",
4240 : path)));
4241 : }
4242 : else
4243 : {
4244 2 : ereport(LOG,
4245 : (errmsg("creating missing WAL directory \"%s\"", path)));
4246 2 : if (MakePGDirectory(path) < 0)
4247 0 : ereport(FATAL,
4248 : (errcode_for_file_access(),
4249 : errmsg("could not create missing directory \"%s\": %m",
4250 : path)));
4251 : }
4252 :
4253 : /* Check for summaries */
4254 1864 : snprintf(path, MAXPGPATH, XLOGDIR "/summaries");
4255 1864 : if (stat(path, &stat_buf) == 0)
4256 : {
4257 : /* Check for weird cases where it exists but isn't a directory */
4258 1862 : if (!S_ISDIR(stat_buf.st_mode))
4259 0 : ereport(FATAL,
4260 : (errmsg("required WAL directory \"%s\" does not exist",
4261 : path)));
4262 : }
4263 : else
4264 : {
4265 2 : ereport(LOG,
4266 : (errmsg("creating missing WAL directory \"%s\"", path)));
4267 2 : if (MakePGDirectory(path) < 0)
4268 0 : ereport(FATAL,
4269 : (errmsg("could not create missing directory \"%s\": %m",
4270 : path)));
4271 : }
4272 1864 : }
4273 :
4274 : /*
4275 : * Remove previous backup history files. This also retries creation of
4276 : * .ready files for any backup history files for which XLogArchiveNotify
4277 : * failed earlier.
4278 : */
4279 : static void
4280 298 : CleanupBackupHistory(void)
4281 : {
4282 : DIR *xldir;
4283 : struct dirent *xlde;
4284 : char path[MAXPGPATH + sizeof(XLOGDIR)];
4285 :
4286 298 : xldir = AllocateDir(XLOGDIR);
4287 :
4288 3010 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4289 : {
4290 2414 : if (IsBackupHistoryFileName(xlde->d_name))
4291 : {
4292 314 : if (XLogArchiveCheckDone(xlde->d_name))
4293 : {
4294 248 : elog(DEBUG2, "removing WAL backup history file \"%s\"",
4295 : xlde->d_name);
4296 248 : snprintf(path, sizeof(path), XLOGDIR "/%s", xlde->d_name);
4297 248 : unlink(path);
4298 248 : XLogArchiveCleanup(xlde->d_name);
4299 : }
4300 : }
4301 : }
4302 :
4303 298 : FreeDir(xldir);
4304 298 : }
4305 :
4306 : /*
4307 : * I/O routines for pg_control
4308 : *
4309 : * *ControlFile is a buffer in shared memory that holds an image of the
4310 : * contents of pg_control. WriteControlFile() initializes pg_control
4311 : * given a preloaded buffer, ReadControlFile() loads the buffer from
4312 : * the pg_control file (during postmaster or standalone-backend startup),
4313 : * and UpdateControlFile() rewrites pg_control after we modify xlog state.
4314 : * InitControlFile() fills the buffer with initial values.
4315 : *
4316 : * For simplicity, WriteControlFile() initializes the fields of pg_control
4317 : * that are related to checking backend/database compatibility, and
4318 : * ReadControlFile() verifies they are correct. We could split out the
4319 : * I/O and compatibility-check functions, but there seems no need currently.
4320 : */
4321 :
4322 : static void
4323 102 : InitControlFile(uint64 sysidentifier, uint32 data_checksum_version)
4324 : {
4325 : char mock_auth_nonce[MOCK_AUTH_NONCE_LEN];
4326 :
4327 : /*
4328 : * Generate a random nonce. This is used for authentication requests that
4329 : * will fail because the user does not exist. The nonce is used to create
4330 : * a genuine-looking password challenge for the non-existent user, in lieu
4331 : * of an actual stored password.
4332 : */
4333 102 : if (!pg_strong_random(mock_auth_nonce, MOCK_AUTH_NONCE_LEN))
4334 0 : ereport(PANIC,
4335 : (errcode(ERRCODE_INTERNAL_ERROR),
4336 : errmsg("could not generate secret authorization token")));
4337 :
4338 102 : memset(ControlFile, 0, sizeof(ControlFileData));
4339 : /* Initialize pg_control status fields */
4340 102 : ControlFile->system_identifier = sysidentifier;
4341 102 : memcpy(ControlFile->mock_authentication_nonce, mock_auth_nonce, MOCK_AUTH_NONCE_LEN);
4342 102 : ControlFile->state = DB_SHUTDOWNED;
4343 102 : ControlFile->unloggedLSN = FirstNormalUnloggedLSN;
4344 :
4345 : /* Set important parameter values for use when replaying WAL */
4346 102 : ControlFile->MaxConnections = MaxConnections;
4347 102 : ControlFile->max_worker_processes = max_worker_processes;
4348 102 : ControlFile->max_wal_senders = max_wal_senders;
4349 102 : ControlFile->max_prepared_xacts = max_prepared_xacts;
4350 102 : ControlFile->max_locks_per_xact = max_locks_per_xact;
4351 102 : ControlFile->wal_level = wal_level;
4352 102 : ControlFile->wal_log_hints = wal_log_hints;
4353 102 : ControlFile->track_commit_timestamp = track_commit_timestamp;
4354 102 : ControlFile->data_checksum_version = data_checksum_version;
4355 102 : }
4356 :
4357 : static void
4358 102 : WriteControlFile(void)
4359 : {
4360 : int fd;
4361 : char buffer[PG_CONTROL_FILE_SIZE]; /* need not be aligned */
4362 :
4363 : /*
4364 : * Initialize version and compatibility-check fields
4365 : */
4366 102 : ControlFile->pg_control_version = PG_CONTROL_VERSION;
4367 102 : ControlFile->catalog_version_no = CATALOG_VERSION_NO;
4368 :
4369 102 : ControlFile->maxAlign = MAXIMUM_ALIGNOF;
4370 102 : ControlFile->floatFormat = FLOATFORMAT_VALUE;
4371 :
4372 102 : ControlFile->blcksz = BLCKSZ;
4373 102 : ControlFile->relseg_size = RELSEG_SIZE;
4374 102 : ControlFile->xlog_blcksz = XLOG_BLCKSZ;
4375 102 : ControlFile->xlog_seg_size = wal_segment_size;
4376 :
4377 102 : ControlFile->nameDataLen = NAMEDATALEN;
4378 102 : ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
4379 :
4380 102 : ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
4381 102 : ControlFile->loblksize = LOBLKSIZE;
4382 :
4383 102 : ControlFile->float8ByVal = FLOAT8PASSBYVAL;
4384 :
4385 : /*
4386 : * Initialize the default 'char' signedness.
4387 : *
4388 : * The signedness of the char type is implementation-defined. For instance
4389 : * on x86 architecture CPUs, the char data type is typically treated as
4390 : * signed by default, whereas on aarch architecture CPUs, it is typically
4391 : * treated as unsigned by default. In v17 or earlier, we accidentally let
4392 : * C implementation signedness affect persistent data. This led to
4393 : * inconsistent results when comparing char data across different
4394 : * platforms.
4395 : *
4396 : * This flag can be used as a hint to ensure consistent behavior for
4397 : * pre-v18 data files that store data sorted by the 'char' type on disk,
4398 : * especially in cross-platform replication scenarios.
4399 : *
4400 : * Newly created database clusters unconditionally set the default char
4401 : * signedness to true. pg_upgrade changes this flag for clusters that were
4402 : * initialized on signedness=false platforms. As a result,
4403 : * signedness=false setting will become rare over time. If we had known
4404 : * about this problem during the last development cycle that forced initdb
4405 : * (v8.3), we would have made all clusters signed or all clusters
4406 : * unsigned. Making pg_upgrade the only source of signedness=false will
4407 : * cause the population of database clusters to converge toward that
4408 : * retrospective ideal.
4409 : */
4410 102 : ControlFile->default_char_signedness = true;
4411 :
4412 : /* Contents are protected with a CRC */
4413 102 : INIT_CRC32C(ControlFile->crc);
4414 102 : COMP_CRC32C(ControlFile->crc,
4415 : ControlFile,
4416 : offsetof(ControlFileData, crc));
4417 102 : FIN_CRC32C(ControlFile->crc);
4418 :
4419 : /*
4420 : * We write out PG_CONTROL_FILE_SIZE bytes into pg_control, zero-padding
4421 : * the excess over sizeof(ControlFileData). This reduces the odds of
4422 : * premature-EOF errors when reading pg_control. We'll still fail when we
4423 : * check the contents of the file, but hopefully with a more specific
4424 : * error than "couldn't read pg_control".
4425 : */
4426 102 : memset(buffer, 0, PG_CONTROL_FILE_SIZE);
4427 102 : memcpy(buffer, ControlFile, sizeof(ControlFileData));
4428 :
4429 102 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4430 : O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
4431 102 : if (fd < 0)
4432 0 : ereport(PANIC,
4433 : (errcode_for_file_access(),
4434 : errmsg("could not create file \"%s\": %m",
4435 : XLOG_CONTROL_FILE)));
4436 :
4437 102 : errno = 0;
4438 102 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_WRITE);
4439 102 : if (write(fd, buffer, PG_CONTROL_FILE_SIZE) != PG_CONTROL_FILE_SIZE)
4440 : {
4441 : /* if write didn't set errno, assume problem is no disk space */
4442 0 : if (errno == 0)
4443 0 : errno = ENOSPC;
4444 0 : ereport(PANIC,
4445 : (errcode_for_file_access(),
4446 : errmsg("could not write to file \"%s\": %m",
4447 : XLOG_CONTROL_FILE)));
4448 : }
4449 102 : pgstat_report_wait_end();
4450 :
4451 102 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_SYNC);
4452 102 : if (pg_fsync(fd) != 0)
4453 0 : ereport(PANIC,
4454 : (errcode_for_file_access(),
4455 : errmsg("could not fsync file \"%s\": %m",
4456 : XLOG_CONTROL_FILE)));
4457 102 : pgstat_report_wait_end();
4458 :
4459 102 : if (close(fd) != 0)
4460 0 : ereport(PANIC,
4461 : (errcode_for_file_access(),
4462 : errmsg("could not close file \"%s\": %m",
4463 : XLOG_CONTROL_FILE)));
4464 102 : }
4465 :
4466 : static void
4467 1964 : ReadControlFile(void)
4468 : {
4469 : pg_crc32c crc;
4470 : int fd;
4471 : char wal_segsz_str[20];
4472 : int r;
4473 :
4474 : /*
4475 : * Read data...
4476 : */
4477 1964 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4478 : O_RDWR | PG_BINARY);
4479 1964 : if (fd < 0)
4480 0 : ereport(PANIC,
4481 : (errcode_for_file_access(),
4482 : errmsg("could not open file \"%s\": %m",
4483 : XLOG_CONTROL_FILE)));
4484 :
4485 1964 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ);
4486 1964 : r = read(fd, ControlFile, sizeof(ControlFileData));
4487 1964 : if (r != sizeof(ControlFileData))
4488 : {
4489 0 : if (r < 0)
4490 0 : ereport(PANIC,
4491 : (errcode_for_file_access(),
4492 : errmsg("could not read file \"%s\": %m",
4493 : XLOG_CONTROL_FILE)));
4494 : else
4495 0 : ereport(PANIC,
4496 : (errcode(ERRCODE_DATA_CORRUPTED),
4497 : errmsg("could not read file \"%s\": read %d of %zu",
4498 : XLOG_CONTROL_FILE, r, sizeof(ControlFileData))));
4499 : }
4500 1964 : pgstat_report_wait_end();
4501 :
4502 1964 : close(fd);
4503 :
4504 : /*
4505 : * Check for expected pg_control format version. If this is wrong, the
4506 : * CRC check will likely fail because we'll be checking the wrong number
4507 : * of bytes. Complaining about wrong version will probably be more
4508 : * enlightening than complaining about wrong CRC.
4509 : */
4510 :
4511 1964 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
4512 0 : ereport(FATAL,
4513 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4514 : errmsg("database files are incompatible with server"),
4515 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4516 : " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4517 : ControlFile->pg_control_version, ControlFile->pg_control_version,
4518 : PG_CONTROL_VERSION, PG_CONTROL_VERSION),
4519 : errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4520 :
4521 1964 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
4522 0 : ereport(FATAL,
4523 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4524 : errmsg("database files are incompatible with server"),
4525 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4526 : " but the server was compiled with PG_CONTROL_VERSION %d.",
4527 : ControlFile->pg_control_version, PG_CONTROL_VERSION),
4528 : errhint("It looks like you need to initdb.")));
4529 :
4530 : /* Now check the CRC. */
4531 1964 : INIT_CRC32C(crc);
4532 1964 : COMP_CRC32C(crc,
4533 : ControlFile,
4534 : offsetof(ControlFileData, crc));
4535 1964 : FIN_CRC32C(crc);
4536 :
4537 1964 : if (!EQ_CRC32C(crc, ControlFile->crc))
4538 0 : ereport(FATAL,
4539 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4540 : errmsg("incorrect checksum in control file")));
4541 :
4542 : /*
4543 : * Do compatibility checking immediately. If the database isn't
4544 : * compatible with the backend executable, we want to abort before we can
4545 : * possibly do any damage.
4546 : */
4547 1964 : if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
4548 0 : ereport(FATAL,
4549 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4550 : errmsg("database files are incompatible with server"),
4551 : /* translator: %s is a variable name and %d is its value */
4552 : errdetail("The database cluster was initialized with %s %d,"
4553 : " but the server was compiled with %s %d.",
4554 : "CATALOG_VERSION_NO", ControlFile->catalog_version_no,
4555 : "CATALOG_VERSION_NO", CATALOG_VERSION_NO),
4556 : errhint("It looks like you need to initdb.")));
4557 1964 : if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
4558 0 : ereport(FATAL,
4559 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4560 : errmsg("database files are incompatible with server"),
4561 : /* translator: %s is a variable name and %d is its value */
4562 : errdetail("The database cluster was initialized with %s %d,"
4563 : " but the server was compiled with %s %d.",
4564 : "MAXALIGN", ControlFile->maxAlign,
4565 : "MAXALIGN", MAXIMUM_ALIGNOF),
4566 : errhint("It looks like you need to initdb.")));
4567 1964 : if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
4568 0 : ereport(FATAL,
4569 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4570 : errmsg("database files are incompatible with server"),
4571 : errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4572 : errhint("It looks like you need to initdb.")));
4573 1964 : if (ControlFile->blcksz != BLCKSZ)
4574 0 : ereport(FATAL,
4575 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4576 : errmsg("database files are incompatible with server"),
4577 : /* translator: %s is a variable name and %d is its value */
4578 : errdetail("The database cluster was initialized with %s %d,"
4579 : " but the server was compiled with %s %d.",
4580 : "BLCKSZ", ControlFile->blcksz,
4581 : "BLCKSZ", BLCKSZ),
4582 : errhint("It looks like you need to recompile or initdb.")));
4583 1964 : if (ControlFile->relseg_size != RELSEG_SIZE)
4584 0 : ereport(FATAL,
4585 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4586 : errmsg("database files are incompatible with server"),
4587 : /* translator: %s is a variable name and %d is its value */
4588 : errdetail("The database cluster was initialized with %s %d,"
4589 : " but the server was compiled with %s %d.",
4590 : "RELSEG_SIZE", ControlFile->relseg_size,
4591 : "RELSEG_SIZE", RELSEG_SIZE),
4592 : errhint("It looks like you need to recompile or initdb.")));
4593 1964 : if (ControlFile->xlog_blcksz != XLOG_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 : "XLOG_BLCKSZ", ControlFile->xlog_blcksz,
4601 : "XLOG_BLCKSZ", XLOG_BLCKSZ),
4602 : errhint("It looks like you need to recompile or initdb.")));
4603 1964 : if (ControlFile->nameDataLen != NAMEDATALEN)
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 : "NAMEDATALEN", ControlFile->nameDataLen,
4611 : "NAMEDATALEN", NAMEDATALEN),
4612 : errhint("It looks like you need to recompile or initdb.")));
4613 1964 : if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
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 : "INDEX_MAX_KEYS", ControlFile->indexMaxKeys,
4621 : "INDEX_MAX_KEYS", INDEX_MAX_KEYS),
4622 : errhint("It looks like you need to recompile or initdb.")));
4623 1964 : if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
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 : "TOAST_MAX_CHUNK_SIZE", ControlFile->toast_max_chunk_size,
4631 : "TOAST_MAX_CHUNK_SIZE", (int) TOAST_MAX_CHUNK_SIZE),
4632 : errhint("It looks like you need to recompile or initdb.")));
4633 1964 : if (ControlFile->loblksize != LOBLKSIZE)
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 : "LOBLKSIZE", ControlFile->loblksize,
4641 : "LOBLKSIZE", (int) LOBLKSIZE),
4642 : errhint("It looks like you need to recompile or initdb.")));
4643 :
4644 : #ifdef USE_FLOAT8_BYVAL
4645 1964 : if (ControlFile->float8ByVal != true)
4646 0 : ereport(FATAL,
4647 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4648 : errmsg("database files are incompatible with server"),
4649 : errdetail("The database cluster was initialized without USE_FLOAT8_BYVAL"
4650 : " but the server was compiled with USE_FLOAT8_BYVAL."),
4651 : errhint("It looks like you need to recompile or initdb.")));
4652 : #else
4653 : if (ControlFile->float8ByVal != false)
4654 : ereport(FATAL,
4655 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4656 : errmsg("database files are incompatible with server"),
4657 : errdetail("The database cluster was initialized with USE_FLOAT8_BYVAL"
4658 : " but the server was compiled without USE_FLOAT8_BYVAL."),
4659 : errhint("It looks like you need to recompile or initdb.")));
4660 : #endif
4661 :
4662 1964 : wal_segment_size = ControlFile->xlog_seg_size;
4663 :
4664 1964 : if (!IsValidWalSegSize(wal_segment_size))
4665 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4666 : errmsg_plural("invalid WAL segment size in control file (%d byte)",
4667 : "invalid WAL segment size in control file (%d bytes)",
4668 : wal_segment_size,
4669 : wal_segment_size),
4670 : errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.")));
4671 :
4672 1964 : snprintf(wal_segsz_str, sizeof(wal_segsz_str), "%d", wal_segment_size);
4673 1964 : SetConfigOption("wal_segment_size", wal_segsz_str, PGC_INTERNAL,
4674 : PGC_S_DYNAMIC_DEFAULT);
4675 :
4676 : /* check and update variables dependent on wal_segment_size */
4677 1964 : if (ConvertToXSegs(min_wal_size_mb, wal_segment_size) < 2)
4678 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4679 : /* translator: both %s are GUC names */
4680 : errmsg("\"%s\" must be at least twice \"%s\"",
4681 : "min_wal_size", "wal_segment_size")));
4682 :
4683 1964 : if (ConvertToXSegs(max_wal_size_mb, wal_segment_size) < 2)
4684 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4685 : /* translator: both %s are GUC names */
4686 : errmsg("\"%s\" must be at least twice \"%s\"",
4687 : "max_wal_size", "wal_segment_size")));
4688 :
4689 1964 : UsableBytesInSegment =
4690 1964 : (wal_segment_size / XLOG_BLCKSZ * UsableBytesInPage) -
4691 : (SizeOfXLogLongPHD - SizeOfXLogShortPHD);
4692 :
4693 1964 : CalculateCheckpointSegments();
4694 :
4695 : /* Make the initdb settings visible as GUC variables, too */
4696 1964 : SetConfigOption("data_checksums", DataChecksumsEnabled() ? "yes" : "no",
4697 : PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
4698 1964 : }
4699 :
4700 : /*
4701 : * Utility wrapper to update the control file. Note that the control
4702 : * file gets flushed.
4703 : */
4704 : static void
4705 17882 : UpdateControlFile(void)
4706 : {
4707 17882 : update_controlfile(DataDir, ControlFile, true);
4708 17882 : }
4709 :
4710 : /*
4711 : * Returns the unique system identifier from control file.
4712 : */
4713 : uint64
4714 2740 : GetSystemIdentifier(void)
4715 : {
4716 : Assert(ControlFile != NULL);
4717 2740 : return ControlFile->system_identifier;
4718 : }
4719 :
4720 : /*
4721 : * Returns the random nonce from control file.
4722 : */
4723 : char *
4724 2 : GetMockAuthenticationNonce(void)
4725 : {
4726 : Assert(ControlFile != NULL);
4727 2 : return ControlFile->mock_authentication_nonce;
4728 : }
4729 :
4730 : /*
4731 : * Are checksums enabled for data pages?
4732 : */
4733 : bool
4734 20089760 : DataChecksumsEnabled(void)
4735 : {
4736 : Assert(ControlFile != NULL);
4737 20089760 : return (ControlFile->data_checksum_version > 0);
4738 : }
4739 :
4740 : /*
4741 : * Return true if the cluster was initialized on a platform where the
4742 : * default signedness of char is "signed". This function exists for code
4743 : * that deals with pre-v18 data files that store data sorted by the 'char'
4744 : * type on disk (e.g., GIN and GiST indexes). See the comments in
4745 : * WriteControlFile() for details.
4746 : */
4747 : bool
4748 6 : GetDefaultCharSignedness(void)
4749 : {
4750 6 : return ControlFile->default_char_signedness;
4751 : }
4752 :
4753 : /*
4754 : * Returns a fake LSN for unlogged relations.
4755 : *
4756 : * Each call generates an LSN that is greater than any previous value
4757 : * returned. The current counter value is saved and restored across clean
4758 : * shutdowns, but like unlogged relations, does not survive a crash. This can
4759 : * be used in lieu of real LSN values returned by XLogInsert, if you need an
4760 : * LSN-like increasing sequence of numbers without writing any WAL.
4761 : */
4762 : XLogRecPtr
4763 66 : GetFakeLSNForUnloggedRel(void)
4764 : {
4765 66 : return pg_atomic_fetch_add_u64(&XLogCtl->unloggedLSN, 1);
4766 : }
4767 :
4768 : /*
4769 : * Auto-tune the number of XLOG buffers.
4770 : *
4771 : * The preferred setting for wal_buffers is about 3% of shared_buffers, with
4772 : * a maximum of one XLOG segment (there is little reason to think that more
4773 : * is helpful, at least so long as we force an fsync when switching log files)
4774 : * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
4775 : * 9.1, when auto-tuning was added).
4776 : *
4777 : * This should not be called until NBuffers has received its final value.
4778 : */
4779 : static int
4780 2148 : XLOGChooseNumBuffers(void)
4781 : {
4782 : int xbuffers;
4783 :
4784 2148 : xbuffers = NBuffers / 32;
4785 2148 : if (xbuffers > (wal_segment_size / XLOG_BLCKSZ))
4786 54 : xbuffers = (wal_segment_size / XLOG_BLCKSZ);
4787 2148 : if (xbuffers < 8)
4788 818 : xbuffers = 8;
4789 2148 : return xbuffers;
4790 : }
4791 :
4792 : /*
4793 : * GUC check_hook for wal_buffers
4794 : */
4795 : bool
4796 4374 : check_wal_buffers(int *newval, void **extra, GucSource source)
4797 : {
4798 : /*
4799 : * -1 indicates a request for auto-tune.
4800 : */
4801 4374 : if (*newval == -1)
4802 : {
4803 : /*
4804 : * If we haven't yet changed the boot_val default of -1, just let it
4805 : * be. We'll fix it when XLOGShmemSize is called.
4806 : */
4807 2226 : if (XLOGbuffers == -1)
4808 2226 : return true;
4809 :
4810 : /* Otherwise, substitute the auto-tune value */
4811 0 : *newval = XLOGChooseNumBuffers();
4812 : }
4813 :
4814 : /*
4815 : * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
4816 : * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
4817 : * the case, we just silently treat such values as a request for the
4818 : * minimum. (We could throw an error instead, but that doesn't seem very
4819 : * helpful.)
4820 : */
4821 2148 : if (*newval < 4)
4822 0 : *newval = 4;
4823 :
4824 2148 : return true;
4825 : }
4826 :
4827 : /*
4828 : * GUC check_hook for wal_consistency_checking
4829 : */
4830 : bool
4831 4018 : check_wal_consistency_checking(char **newval, void **extra, GucSource source)
4832 : {
4833 : char *rawstring;
4834 : List *elemlist;
4835 : ListCell *l;
4836 : bool newwalconsistency[RM_MAX_ID + 1];
4837 :
4838 : /* Initialize the array */
4839 132594 : MemSet(newwalconsistency, 0, (RM_MAX_ID + 1) * sizeof(bool));
4840 :
4841 : /* Need a modifiable copy of string */
4842 4018 : rawstring = pstrdup(*newval);
4843 :
4844 : /* Parse string into list of identifiers */
4845 4018 : if (!SplitIdentifierString(rawstring, ',', &elemlist))
4846 : {
4847 : /* syntax error in list */
4848 0 : GUC_check_errdetail("List syntax is invalid.");
4849 0 : pfree(rawstring);
4850 0 : list_free(elemlist);
4851 0 : return false;
4852 : }
4853 :
4854 4920 : foreach(l, elemlist)
4855 : {
4856 902 : char *tok = (char *) lfirst(l);
4857 : int rmid;
4858 :
4859 : /* Check for 'all'. */
4860 902 : if (pg_strcasecmp(tok, "all") == 0)
4861 : {
4862 230786 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4863 229888 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL)
4864 8980 : newwalconsistency[rmid] = true;
4865 : }
4866 : else
4867 : {
4868 : /* Check if the token matches any known resource manager. */
4869 4 : bool found = false;
4870 :
4871 72 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4872 : {
4873 108 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL &&
4874 36 : pg_strcasecmp(tok, GetRmgr(rmid).rm_name) == 0)
4875 : {
4876 4 : newwalconsistency[rmid] = true;
4877 4 : found = true;
4878 4 : break;
4879 : }
4880 : }
4881 4 : if (!found)
4882 : {
4883 : /*
4884 : * During startup, it might be a not-yet-loaded custom
4885 : * resource manager. Defer checking until
4886 : * InitializeWalConsistencyChecking().
4887 : */
4888 0 : if (!process_shared_preload_libraries_done)
4889 : {
4890 0 : check_wal_consistency_checking_deferred = true;
4891 : }
4892 : else
4893 : {
4894 0 : GUC_check_errdetail("Unrecognized key word: \"%s\".", tok);
4895 0 : pfree(rawstring);
4896 0 : list_free(elemlist);
4897 0 : return false;
4898 : }
4899 : }
4900 : }
4901 : }
4902 :
4903 4018 : pfree(rawstring);
4904 4018 : list_free(elemlist);
4905 :
4906 : /* assign new value */
4907 4018 : *extra = guc_malloc(LOG, (RM_MAX_ID + 1) * sizeof(bool));
4908 4018 : if (!*extra)
4909 0 : return false;
4910 4018 : memcpy(*extra, newwalconsistency, (RM_MAX_ID + 1) * sizeof(bool));
4911 4018 : return true;
4912 : }
4913 :
4914 : /*
4915 : * GUC assign_hook for wal_consistency_checking
4916 : */
4917 : void
4918 4016 : assign_wal_consistency_checking(const char *newval, void *extra)
4919 : {
4920 : /*
4921 : * If some checks were deferred, it's possible that the checks will fail
4922 : * later during InitializeWalConsistencyChecking(). But in that case, the
4923 : * postmaster will exit anyway, so it's safe to proceed with the
4924 : * assignment.
4925 : *
4926 : * Any built-in resource managers specified are assigned immediately,
4927 : * which affects WAL created before shared_preload_libraries are
4928 : * processed. Any custom resource managers specified won't be assigned
4929 : * until after shared_preload_libraries are processed, but that's OK
4930 : * because WAL for a custom resource manager can't be written before the
4931 : * module is loaded anyway.
4932 : */
4933 4016 : wal_consistency_checking = extra;
4934 4016 : }
4935 :
4936 : /*
4937 : * InitializeWalConsistencyChecking: run after loading custom resource managers
4938 : *
4939 : * If any unknown resource managers were specified in the
4940 : * wal_consistency_checking GUC, processing was deferred. Now that
4941 : * shared_preload_libraries have been loaded, process wal_consistency_checking
4942 : * again.
4943 : */
4944 : void
4945 1842 : InitializeWalConsistencyChecking(void)
4946 : {
4947 : Assert(process_shared_preload_libraries_done);
4948 :
4949 1842 : if (check_wal_consistency_checking_deferred)
4950 : {
4951 : struct config_generic *guc;
4952 :
4953 0 : guc = find_option("wal_consistency_checking", false, false, ERROR);
4954 :
4955 0 : check_wal_consistency_checking_deferred = false;
4956 :
4957 0 : set_config_option_ext("wal_consistency_checking",
4958 : wal_consistency_checking_string,
4959 : guc->scontext, guc->source, guc->srole,
4960 : GUC_ACTION_SET, true, ERROR, false);
4961 :
4962 : /* checking should not be deferred again */
4963 : Assert(!check_wal_consistency_checking_deferred);
4964 : }
4965 1842 : }
4966 :
4967 : /*
4968 : * GUC show_hook for archive_command
4969 : */
4970 : const char *
4971 3566 : show_archive_command(void)
4972 : {
4973 3566 : if (XLogArchivingActive())
4974 4 : return XLogArchiveCommand;
4975 : else
4976 3562 : return "(disabled)";
4977 : }
4978 :
4979 : /*
4980 : * GUC show_hook for in_hot_standby
4981 : */
4982 : const char *
4983 29816 : show_in_hot_standby(void)
4984 : {
4985 : /*
4986 : * We display the actual state based on shared memory, so that this GUC
4987 : * reports up-to-date state if examined intra-query. The underlying
4988 : * variable (in_hot_standby_guc) changes only when we transmit a new value
4989 : * to the client.
4990 : */
4991 29816 : return RecoveryInProgress() ? "on" : "off";
4992 : }
4993 :
4994 : /*
4995 : * Read the control file, set respective GUCs.
4996 : *
4997 : * This is to be called during startup, including a crash recovery cycle,
4998 : * unless in bootstrap mode, where no control file yet exists. As there's no
4999 : * usable shared memory yet (its sizing can depend on the contents of the
5000 : * control file!), first store the contents in local memory. XLOGShmemInit()
5001 : * will then copy it to shared memory later.
5002 : *
5003 : * reset just controls whether previous contents are to be expected (in the
5004 : * reset case, there's a dangling pointer into old shared memory), or not.
5005 : */
5006 : void
5007 1862 : LocalProcessControlFile(bool reset)
5008 : {
5009 : Assert(reset || ControlFile == NULL);
5010 1862 : ControlFile = palloc(sizeof(ControlFileData));
5011 1862 : ReadControlFile();
5012 1862 : }
5013 :
5014 : /*
5015 : * Get the wal_level from the control file. For a standby, this value should be
5016 : * considered as its active wal_level, because it may be different from what
5017 : * was originally configured on standby.
5018 : */
5019 : WalLevel
5020 138 : GetActiveWalLevelOnStandby(void)
5021 : {
5022 138 : return ControlFile->wal_level;
5023 : }
5024 :
5025 : /*
5026 : * Initialization of shared memory for XLOG
5027 : */
5028 : Size
5029 6150 : XLOGShmemSize(void)
5030 : {
5031 : Size size;
5032 :
5033 : /*
5034 : * If the value of wal_buffers is -1, use the preferred auto-tune value.
5035 : * This isn't an amazingly clean place to do this, but we must wait till
5036 : * NBuffers has received its final value, and must do it before using the
5037 : * value of XLOGbuffers to do anything important.
5038 : *
5039 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
5040 : * However, if the DBA explicitly set wal_buffers = -1 in the config file,
5041 : * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
5042 : * the matter with PGC_S_OVERRIDE.
5043 : */
5044 6150 : if (XLOGbuffers == -1)
5045 : {
5046 : char buf[32];
5047 :
5048 2148 : snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
5049 2148 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
5050 : PGC_S_DYNAMIC_DEFAULT);
5051 2148 : if (XLOGbuffers == -1) /* failed to apply it? */
5052 0 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
5053 : PGC_S_OVERRIDE);
5054 : }
5055 : Assert(XLOGbuffers > 0);
5056 :
5057 : /* XLogCtl */
5058 6150 : size = sizeof(XLogCtlData);
5059 :
5060 : /* WAL insertion locks, plus alignment */
5061 6150 : size = add_size(size, mul_size(sizeof(WALInsertLockPadded), NUM_XLOGINSERT_LOCKS + 1));
5062 : /* xlblocks array */
5063 6150 : size = add_size(size, mul_size(sizeof(pg_atomic_uint64), XLOGbuffers));
5064 : /* extra alignment padding for XLOG I/O buffers */
5065 6150 : size = add_size(size, Max(XLOG_BLCKSZ, PG_IO_ALIGN_SIZE));
5066 : /* and the buffers themselves */
5067 6150 : size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
5068 :
5069 : /*
5070 : * Note: we don't count ControlFileData, it comes out of the "slop factor"
5071 : * added by CreateSharedMemoryAndSemaphores. This lets us use this
5072 : * routine again below to compute the actual allocation size.
5073 : */
5074 :
5075 6150 : return size;
5076 : }
5077 :
5078 : void
5079 2152 : XLOGShmemInit(void)
5080 : {
5081 : bool foundCFile,
5082 : foundXLog;
5083 : char *allocptr;
5084 : int i;
5085 : ControlFileData *localControlFile;
5086 :
5087 : #ifdef WAL_DEBUG
5088 :
5089 : /*
5090 : * Create a memory context for WAL debugging that's exempt from the normal
5091 : * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
5092 : * an allocation fails, but wal_debug is not for production use anyway.
5093 : */
5094 : if (walDebugCxt == NULL)
5095 : {
5096 : walDebugCxt = AllocSetContextCreate(TopMemoryContext,
5097 : "WAL Debug",
5098 : ALLOCSET_DEFAULT_SIZES);
5099 : MemoryContextAllowInCriticalSection(walDebugCxt, true);
5100 : }
5101 : #endif
5102 :
5103 :
5104 2152 : XLogCtl = (XLogCtlData *)
5105 2152 : ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog);
5106 :
5107 2152 : localControlFile = ControlFile;
5108 2152 : ControlFile = (ControlFileData *)
5109 2152 : ShmemInitStruct("Control File", sizeof(ControlFileData), &foundCFile);
5110 :
5111 2152 : if (foundCFile || foundXLog)
5112 : {
5113 : /* both should be present or neither */
5114 : Assert(foundCFile && foundXLog);
5115 :
5116 : /* Initialize local copy of WALInsertLocks */
5117 0 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks;
5118 :
5119 0 : if (localControlFile)
5120 0 : pfree(localControlFile);
5121 0 : return;
5122 : }
5123 2152 : memset(XLogCtl, 0, sizeof(XLogCtlData));
5124 :
5125 : /*
5126 : * Already have read control file locally, unless in bootstrap mode. Move
5127 : * contents into shared memory.
5128 : */
5129 2152 : if (localControlFile)
5130 : {
5131 1846 : memcpy(ControlFile, localControlFile, sizeof(ControlFileData));
5132 1846 : pfree(localControlFile);
5133 : }
5134 :
5135 : /*
5136 : * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
5137 : * multiple of the alignment for same, so no extra alignment padding is
5138 : * needed here.
5139 : */
5140 2152 : allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
5141 2152 : XLogCtl->xlblocks = (pg_atomic_uint64 *) allocptr;
5142 2152 : allocptr += sizeof(pg_atomic_uint64) * XLOGbuffers;
5143 :
5144 618882 : for (i = 0; i < XLOGbuffers; i++)
5145 : {
5146 616730 : pg_atomic_init_u64(&XLogCtl->xlblocks[i], InvalidXLogRecPtr);
5147 : }
5148 :
5149 : /* WAL insertion locks. Ensure they're aligned to the full padded size */
5150 2152 : allocptr += sizeof(WALInsertLockPadded) -
5151 2152 : ((uintptr_t) allocptr) % sizeof(WALInsertLockPadded);
5152 2152 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks =
5153 : (WALInsertLockPadded *) allocptr;
5154 2152 : allocptr += sizeof(WALInsertLockPadded) * NUM_XLOGINSERT_LOCKS;
5155 :
5156 19368 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
5157 : {
5158 17216 : LWLockInitialize(&WALInsertLocks[i].l.lock, LWTRANCHE_WAL_INSERT);
5159 17216 : pg_atomic_init_u64(&WALInsertLocks[i].l.insertingAt, InvalidXLogRecPtr);
5160 17216 : WALInsertLocks[i].l.lastImportantAt = InvalidXLogRecPtr;
5161 : }
5162 :
5163 : /*
5164 : * Align the start of the page buffers to a full xlog block size boundary.
5165 : * This simplifies some calculations in XLOG insertion. It is also
5166 : * required for O_DIRECT.
5167 : */
5168 2152 : allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5169 2152 : XLogCtl->pages = allocptr;
5170 2152 : memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5171 :
5172 : /*
5173 : * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5174 : * in additional info.)
5175 : */
5176 2152 : XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5177 2152 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5178 2152 : XLogCtl->InstallXLogFileSegmentActive = false;
5179 2152 : XLogCtl->WalWriterSleeping = false;
5180 :
5181 2152 : SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5182 2152 : SpinLockInit(&XLogCtl->info_lck);
5183 2152 : pg_atomic_init_u64(&XLogCtl->logInsertResult, InvalidXLogRecPtr);
5184 2152 : pg_atomic_init_u64(&XLogCtl->logWriteResult, InvalidXLogRecPtr);
5185 2152 : pg_atomic_init_u64(&XLogCtl->logFlushResult, InvalidXLogRecPtr);
5186 2152 : pg_atomic_init_u64(&XLogCtl->unloggedLSN, InvalidXLogRecPtr);
5187 :
5188 2152 : pg_atomic_init_u64(&XLogCtl->InitializeReserved, InvalidXLogRecPtr);
5189 2152 : pg_atomic_init_u64(&XLogCtl->InitializedUpTo, InvalidXLogRecPtr);
5190 2152 : ConditionVariableInit(&XLogCtl->InitializedUpToCondVar);
5191 : }
5192 :
5193 : /*
5194 : * This func must be called ONCE on system install. It creates pg_control
5195 : * and the initial XLOG segment.
5196 : */
5197 : void
5198 102 : BootStrapXLOG(uint32 data_checksum_version)
5199 : {
5200 : CheckPoint checkPoint;
5201 : char *buffer;
5202 : XLogPageHeader page;
5203 : XLogLongPageHeader longpage;
5204 : XLogRecord *record;
5205 : char *recptr;
5206 : uint64 sysidentifier;
5207 : struct timeval tv;
5208 : pg_crc32c crc;
5209 :
5210 : /* allow ordinary WAL segment creation, like StartupXLOG() would */
5211 102 : SetInstallXLogFileSegmentActive();
5212 :
5213 : /*
5214 : * Select a hopefully-unique system identifier code for this installation.
5215 : * We use the result of gettimeofday(), including the fractional seconds
5216 : * field, as being about as unique as we can easily get. (Think not to
5217 : * use random(), since it hasn't been seeded and there's no portable way
5218 : * to seed it other than the system clock value...) The upper half of the
5219 : * uint64 value is just the tv_sec part, while the lower half contains the
5220 : * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
5221 : * PID for a little extra uniqueness. A person knowing this encoding can
5222 : * determine the initialization time of the installation, which could
5223 : * perhaps be useful sometimes.
5224 : */
5225 102 : gettimeofday(&tv, NULL);
5226 102 : sysidentifier = ((uint64) tv.tv_sec) << 32;
5227 102 : sysidentifier |= ((uint64) tv.tv_usec) << 12;
5228 102 : sysidentifier |= getpid() & 0xFFF;
5229 :
5230 : /* page buffer must be aligned suitably for O_DIRECT */
5231 102 : buffer = (char *) palloc(XLOG_BLCKSZ + XLOG_BLCKSZ);
5232 102 : page = (XLogPageHeader) TYPEALIGN(XLOG_BLCKSZ, buffer);
5233 102 : memset(page, 0, XLOG_BLCKSZ);
5234 :
5235 : /*
5236 : * Set up information for the initial checkpoint record
5237 : *
5238 : * The initial checkpoint record is written to the beginning of the WAL
5239 : * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5240 : * used, so that we can use 0/0 to mean "before any valid WAL segment".
5241 : */
5242 102 : checkPoint.redo = wal_segment_size + SizeOfXLogLongPHD;
5243 102 : checkPoint.ThisTimeLineID = BootstrapTimeLineID;
5244 102 : checkPoint.PrevTimeLineID = BootstrapTimeLineID;
5245 102 : checkPoint.fullPageWrites = fullPageWrites;
5246 102 : checkPoint.wal_level = wal_level;
5247 : checkPoint.nextXid =
5248 102 : FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId);
5249 102 : checkPoint.nextOid = FirstGenbkiObjectId;
5250 102 : checkPoint.nextMulti = FirstMultiXactId;
5251 102 : checkPoint.nextMultiOffset = 0;
5252 102 : checkPoint.oldestXid = FirstNormalTransactionId;
5253 102 : checkPoint.oldestXidDB = Template1DbOid;
5254 102 : checkPoint.oldestMulti = FirstMultiXactId;
5255 102 : checkPoint.oldestMultiDB = Template1DbOid;
5256 102 : checkPoint.oldestCommitTsXid = InvalidTransactionId;
5257 102 : checkPoint.newestCommitTsXid = InvalidTransactionId;
5258 102 : checkPoint.time = (pg_time_t) time(NULL);
5259 102 : checkPoint.oldestActiveXid = InvalidTransactionId;
5260 :
5261 102 : TransamVariables->nextXid = checkPoint.nextXid;
5262 102 : TransamVariables->nextOid = checkPoint.nextOid;
5263 102 : TransamVariables->oidCount = 0;
5264 102 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5265 102 : AdvanceOldestClogXid(checkPoint.oldestXid);
5266 102 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5267 102 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5268 102 : SetCommitTsLimit(InvalidTransactionId, InvalidTransactionId);
5269 :
5270 : /* Set up the XLOG page header */
5271 102 : page->xlp_magic = XLOG_PAGE_MAGIC;
5272 102 : page->xlp_info = XLP_LONG_HEADER;
5273 102 : page->xlp_tli = BootstrapTimeLineID;
5274 102 : page->xlp_pageaddr = wal_segment_size;
5275 102 : longpage = (XLogLongPageHeader) page;
5276 102 : longpage->xlp_sysid = sysidentifier;
5277 102 : longpage->xlp_seg_size = wal_segment_size;
5278 102 : longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5279 :
5280 : /* Insert the initial checkpoint record */
5281 102 : recptr = ((char *) page + SizeOfXLogLongPHD);
5282 102 : record = (XLogRecord *) recptr;
5283 102 : record->xl_prev = 0;
5284 102 : record->xl_xid = InvalidTransactionId;
5285 102 : record->xl_tot_len = SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(checkPoint);
5286 102 : record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5287 102 : record->xl_rmid = RM_XLOG_ID;
5288 102 : recptr += SizeOfXLogRecord;
5289 : /* fill the XLogRecordDataHeaderShort struct */
5290 102 : *(recptr++) = (char) XLR_BLOCK_ID_DATA_SHORT;
5291 102 : *(recptr++) = sizeof(checkPoint);
5292 102 : memcpy(recptr, &checkPoint, sizeof(checkPoint));
5293 102 : recptr += sizeof(checkPoint);
5294 : Assert(recptr - (char *) record == record->xl_tot_len);
5295 :
5296 102 : INIT_CRC32C(crc);
5297 102 : COMP_CRC32C(crc, ((char *) record) + SizeOfXLogRecord, record->xl_tot_len - SizeOfXLogRecord);
5298 102 : COMP_CRC32C(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5299 102 : FIN_CRC32C(crc);
5300 102 : record->xl_crc = crc;
5301 :
5302 : /* Create first XLOG segment file */
5303 102 : openLogTLI = BootstrapTimeLineID;
5304 102 : openLogFile = XLogFileInit(1, BootstrapTimeLineID);
5305 :
5306 : /*
5307 : * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
5308 : * close the file again in a moment.
5309 : */
5310 :
5311 : /* Write the first page with the initial record */
5312 102 : errno = 0;
5313 102 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE);
5314 102 : if (write(openLogFile, page, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5315 : {
5316 : /* if write didn't set errno, assume problem is no disk space */
5317 0 : if (errno == 0)
5318 0 : errno = ENOSPC;
5319 0 : ereport(PANIC,
5320 : (errcode_for_file_access(),
5321 : errmsg("could not write bootstrap write-ahead log file: %m")));
5322 : }
5323 102 : pgstat_report_wait_end();
5324 :
5325 102 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC);
5326 102 : if (pg_fsync(openLogFile) != 0)
5327 0 : ereport(PANIC,
5328 : (errcode_for_file_access(),
5329 : errmsg("could not fsync bootstrap write-ahead log file: %m")));
5330 102 : pgstat_report_wait_end();
5331 :
5332 102 : if (close(openLogFile) != 0)
5333 0 : ereport(PANIC,
5334 : (errcode_for_file_access(),
5335 : errmsg("could not close bootstrap write-ahead log file: %m")));
5336 :
5337 102 : openLogFile = -1;
5338 :
5339 : /* Now create pg_control */
5340 102 : InitControlFile(sysidentifier, data_checksum_version);
5341 102 : ControlFile->time = checkPoint.time;
5342 102 : ControlFile->checkPoint = checkPoint.redo;
5343 102 : ControlFile->checkPointCopy = checkPoint;
5344 :
5345 : /* some additional ControlFile fields are set in WriteControlFile() */
5346 102 : WriteControlFile();
5347 :
5348 : /* Bootstrap the commit log, too */
5349 102 : BootStrapCLOG();
5350 102 : BootStrapCommitTs();
5351 102 : BootStrapSUBTRANS();
5352 102 : BootStrapMultiXact();
5353 :
5354 102 : pfree(buffer);
5355 :
5356 : /*
5357 : * Force control file to be read - in contrast to normal processing we'd
5358 : * otherwise never run the checks and GUC related initializations therein.
5359 : */
5360 102 : ReadControlFile();
5361 102 : }
5362 :
5363 : static char *
5364 1660 : str_time(pg_time_t tnow)
5365 : {
5366 1660 : char *buf = palloc(128);
5367 :
5368 1660 : pg_strftime(buf, 128,
5369 : "%Y-%m-%d %H:%M:%S %Z",
5370 1660 : pg_localtime(&tnow, log_timezone));
5371 :
5372 1660 : return buf;
5373 : }
5374 :
5375 : /*
5376 : * Initialize the first WAL segment on new timeline.
5377 : */
5378 : static void
5379 96 : XLogInitNewTimeline(TimeLineID endTLI, XLogRecPtr endOfLog, TimeLineID newTLI)
5380 : {
5381 : char xlogfname[MAXFNAMELEN];
5382 : XLogSegNo endLogSegNo;
5383 : XLogSegNo startLogSegNo;
5384 :
5385 : /* we always switch to a new timeline after archive recovery */
5386 : Assert(endTLI != newTLI);
5387 :
5388 : /*
5389 : * Update min recovery point one last time.
5390 : */
5391 96 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5392 :
5393 : /*
5394 : * Calculate the last segment on the old timeline, and the first segment
5395 : * on the new timeline. If the switch happens in the middle of a segment,
5396 : * they are the same, but if the switch happens exactly at a segment
5397 : * boundary, startLogSegNo will be endLogSegNo + 1.
5398 : */
5399 96 : XLByteToPrevSeg(endOfLog, endLogSegNo, wal_segment_size);
5400 96 : XLByteToSeg(endOfLog, startLogSegNo, wal_segment_size);
5401 :
5402 : /*
5403 : * Initialize the starting WAL segment for the new timeline. If the switch
5404 : * happens in the middle of a segment, copy data from the last WAL segment
5405 : * of the old timeline up to the switch point, to the starting WAL segment
5406 : * on the new timeline.
5407 : */
5408 96 : if (endLogSegNo == startLogSegNo)
5409 : {
5410 : /*
5411 : * Make a copy of the file on the new timeline.
5412 : *
5413 : * Writing WAL isn't allowed yet, so there are no locking
5414 : * considerations. But we should be just as tense as XLogFileInit to
5415 : * avoid emplacing a bogus file.
5416 : */
5417 74 : XLogFileCopy(newTLI, endLogSegNo, endTLI, endLogSegNo,
5418 74 : XLogSegmentOffset(endOfLog, wal_segment_size));
5419 : }
5420 : else
5421 : {
5422 : /*
5423 : * The switch happened at a segment boundary, so just create the next
5424 : * segment on the new timeline.
5425 : */
5426 : int fd;
5427 :
5428 22 : fd = XLogFileInit(startLogSegNo, newTLI);
5429 :
5430 22 : if (close(fd) != 0)
5431 : {
5432 0 : int save_errno = errno;
5433 :
5434 0 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5435 0 : errno = save_errno;
5436 0 : ereport(ERROR,
5437 : (errcode_for_file_access(),
5438 : errmsg("could not close file \"%s\": %m", xlogfname)));
5439 : }
5440 : }
5441 :
5442 : /*
5443 : * Let's just make real sure there are not .ready or .done flags posted
5444 : * for the new segment.
5445 : */
5446 96 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5447 96 : XLogArchiveCleanup(xlogfname);
5448 96 : }
5449 :
5450 : /*
5451 : * Perform cleanup actions at the conclusion of archive recovery.
5452 : */
5453 : static void
5454 96 : CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI, XLogRecPtr EndOfLog,
5455 : TimeLineID newTLI)
5456 : {
5457 : /*
5458 : * Execute the recovery_end_command, if any.
5459 : */
5460 96 : if (recoveryEndCommand && strcmp(recoveryEndCommand, "") != 0)
5461 4 : ExecuteRecoveryCommand(recoveryEndCommand,
5462 : "recovery_end_command",
5463 : true,
5464 : WAIT_EVENT_RECOVERY_END_COMMAND);
5465 :
5466 : /*
5467 : * We switched to a new timeline. Clean up segments on the old timeline.
5468 : *
5469 : * If there are any higher-numbered segments on the old timeline, remove
5470 : * them. They might contain valid WAL, but they might also be
5471 : * pre-allocated files containing garbage. In any case, they are not part
5472 : * of the new timeline's history so we don't need them.
5473 : */
5474 96 : RemoveNonParentXlogFiles(EndOfLog, newTLI);
5475 :
5476 : /*
5477 : * If the switch happened in the middle of a segment, what to do with the
5478 : * last, partial segment on the old timeline? If we don't archive it, and
5479 : * the server that created the WAL never archives it either (e.g. because
5480 : * it was hit by a meteor), it will never make it to the archive. That's
5481 : * OK from our point of view, because the new segment that we created with
5482 : * the new TLI contains all the WAL from the old timeline up to the switch
5483 : * point. But if you later try to do PITR to the "missing" WAL on the old
5484 : * timeline, recovery won't find it in the archive. It's physically
5485 : * present in the new file with new TLI, but recovery won't look there
5486 : * when it's recovering to the older timeline. On the other hand, if we
5487 : * archive the partial segment, and the original server on that timeline
5488 : * is still running and archives the completed version of the same segment
5489 : * later, it will fail. (We used to do that in 9.4 and below, and it
5490 : * caused such problems).
5491 : *
5492 : * As a compromise, we rename the last segment with the .partial suffix,
5493 : * and archive it. Archive recovery will never try to read .partial
5494 : * segments, so they will normally go unused. But in the odd PITR case,
5495 : * the administrator can copy them manually to the pg_wal directory
5496 : * (removing the suffix). They can be useful in debugging, too.
5497 : *
5498 : * If a .done or .ready file already exists for the old timeline, however,
5499 : * we had already determined that the segment is complete, so we can let
5500 : * it be archived normally. (In particular, if it was restored from the
5501 : * archive to begin with, it's expected to have a .done file).
5502 : */
5503 96 : if (XLogSegmentOffset(EndOfLog, wal_segment_size) != 0 &&
5504 : XLogArchivingActive())
5505 : {
5506 : char origfname[MAXFNAMELEN];
5507 : XLogSegNo endLogSegNo;
5508 :
5509 18 : XLByteToPrevSeg(EndOfLog, endLogSegNo, wal_segment_size);
5510 18 : XLogFileName(origfname, EndOfLogTLI, endLogSegNo, wal_segment_size);
5511 :
5512 18 : if (!XLogArchiveIsReadyOrDone(origfname))
5513 : {
5514 : char origpath[MAXPGPATH];
5515 : char partialfname[MAXFNAMELEN];
5516 : char partialpath[MAXPGPATH];
5517 :
5518 : /*
5519 : * If we're summarizing WAL, we can't rename the partial file
5520 : * until the summarizer finishes with it, else it will fail.
5521 : */
5522 10 : if (summarize_wal)
5523 2 : WaitForWalSummarization(EndOfLog);
5524 :
5525 10 : XLogFilePath(origpath, EndOfLogTLI, endLogSegNo, wal_segment_size);
5526 10 : snprintf(partialfname, MAXFNAMELEN, "%s.partial", origfname);
5527 10 : snprintf(partialpath, MAXPGPATH, "%s.partial", origpath);
5528 :
5529 : /*
5530 : * Make sure there's no .done or .ready file for the .partial
5531 : * file.
5532 : */
5533 10 : XLogArchiveCleanup(partialfname);
5534 :
5535 10 : durable_rename(origpath, partialpath, ERROR);
5536 10 : XLogArchiveNotify(partialfname);
5537 : }
5538 : }
5539 96 : }
5540 :
5541 : /*
5542 : * Check to see if required parameters are set high enough on this server
5543 : * for various aspects of recovery operation.
5544 : *
5545 : * Note that all the parameters which this function tests need to be
5546 : * listed in Administrator's Overview section in high-availability.sgml.
5547 : * If you change them, don't forget to update the list.
5548 : */
5549 : static void
5550 490 : CheckRequiredParameterValues(void)
5551 : {
5552 : /*
5553 : * For archive recovery, the WAL must be generated with at least 'replica'
5554 : * wal_level.
5555 : */
5556 490 : if (ArchiveRecoveryRequested && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5557 : {
5558 4 : ereport(FATAL,
5559 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5560 : errmsg("WAL was generated with \"wal_level=minimal\", cannot continue recovering"),
5561 : errdetail("This happens if you temporarily set \"wal_level=minimal\" on the server."),
5562 : errhint("Use a backup taken after setting \"wal_level\" to higher than \"minimal\".")));
5563 : }
5564 :
5565 : /*
5566 : * For Hot Standby, the WAL must be generated with 'replica' mode, and we
5567 : * must have at least as many backend slots as the primary.
5568 : */
5569 486 : if (ArchiveRecoveryRequested && EnableHotStandby)
5570 : {
5571 : /* We ignore autovacuum_worker_slots when we make this test. */
5572 244 : RecoveryRequiresIntParameter("max_connections",
5573 : MaxConnections,
5574 244 : ControlFile->MaxConnections);
5575 244 : RecoveryRequiresIntParameter("max_worker_processes",
5576 : max_worker_processes,
5577 244 : ControlFile->max_worker_processes);
5578 244 : RecoveryRequiresIntParameter("max_wal_senders",
5579 : max_wal_senders,
5580 244 : ControlFile->max_wal_senders);
5581 244 : RecoveryRequiresIntParameter("max_prepared_transactions",
5582 : max_prepared_xacts,
5583 244 : ControlFile->max_prepared_xacts);
5584 244 : RecoveryRequiresIntParameter("max_locks_per_transaction",
5585 : max_locks_per_xact,
5586 244 : ControlFile->max_locks_per_xact);
5587 : }
5588 486 : }
5589 :
5590 : /*
5591 : * This must be called ONCE during postmaster or standalone-backend startup
5592 : */
5593 : void
5594 1864 : StartupXLOG(void)
5595 : {
5596 : XLogCtlInsert *Insert;
5597 : CheckPoint checkPoint;
5598 : bool wasShutdown;
5599 : bool didCrash;
5600 : bool haveTblspcMap;
5601 : bool haveBackupLabel;
5602 : XLogRecPtr EndOfLog;
5603 : TimeLineID EndOfLogTLI;
5604 : TimeLineID newTLI;
5605 : bool performedWalRecovery;
5606 : EndOfWalRecoveryInfo *endOfRecoveryInfo;
5607 : XLogRecPtr abortedRecPtr;
5608 : XLogRecPtr missingContrecPtr;
5609 : TransactionId oldestActiveXID;
5610 1864 : bool promoted = false;
5611 :
5612 : /*
5613 : * We should have an aux process resource owner to use, and we should not
5614 : * be in a transaction that's installed some other resowner.
5615 : */
5616 : Assert(AuxProcessResourceOwner != NULL);
5617 : Assert(CurrentResourceOwner == NULL ||
5618 : CurrentResourceOwner == AuxProcessResourceOwner);
5619 1864 : CurrentResourceOwner = AuxProcessResourceOwner;
5620 :
5621 : /*
5622 : * Check that contents look valid.
5623 : */
5624 1864 : if (!XRecOffIsValid(ControlFile->checkPoint))
5625 0 : ereport(FATAL,
5626 : (errcode(ERRCODE_DATA_CORRUPTED),
5627 : errmsg("control file contains invalid checkpoint location")));
5628 :
5629 1864 : switch (ControlFile->state)
5630 : {
5631 1450 : case DB_SHUTDOWNED:
5632 :
5633 : /*
5634 : * This is the expected case, so don't be chatty in standalone
5635 : * mode
5636 : */
5637 1450 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
5638 : (errmsg("database system was shut down at %s",
5639 : str_time(ControlFile->time))));
5640 1450 : break;
5641 :
5642 64 : case DB_SHUTDOWNED_IN_RECOVERY:
5643 64 : ereport(LOG,
5644 : (errmsg("database system was shut down in recovery at %s",
5645 : str_time(ControlFile->time))));
5646 64 : break;
5647 :
5648 0 : case DB_SHUTDOWNING:
5649 0 : ereport(LOG,
5650 : (errmsg("database system shutdown was interrupted; last known up at %s",
5651 : str_time(ControlFile->time))));
5652 0 : break;
5653 :
5654 0 : case DB_IN_CRASH_RECOVERY:
5655 0 : ereport(LOG,
5656 : (errmsg("database system was interrupted while in recovery at %s",
5657 : str_time(ControlFile->time)),
5658 : errhint("This probably means that some data is corrupted and"
5659 : " you will have to use the last backup for recovery.")));
5660 0 : break;
5661 :
5662 12 : case DB_IN_ARCHIVE_RECOVERY:
5663 12 : ereport(LOG,
5664 : (errmsg("database system was interrupted while in recovery at log time %s",
5665 : str_time(ControlFile->checkPointCopy.time)),
5666 : errhint("If this has occurred more than once some data might be corrupted"
5667 : " and you might need to choose an earlier recovery target.")));
5668 12 : break;
5669 :
5670 338 : case DB_IN_PRODUCTION:
5671 338 : ereport(LOG,
5672 : (errmsg("database system was interrupted; last known up at %s",
5673 : str_time(ControlFile->time))));
5674 338 : break;
5675 :
5676 0 : default:
5677 0 : ereport(FATAL,
5678 : (errcode(ERRCODE_DATA_CORRUPTED),
5679 : errmsg("control file contains invalid database cluster state")));
5680 : }
5681 :
5682 : /* This is just to allow attaching to startup process with a debugger */
5683 : #ifdef XLOG_REPLAY_DELAY
5684 : if (ControlFile->state != DB_SHUTDOWNED)
5685 : pg_usleep(60000000L);
5686 : #endif
5687 :
5688 : /*
5689 : * Verify that pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
5690 : * In cases where someone has performed a copy for PITR, these directories
5691 : * may have been excluded and need to be re-created.
5692 : */
5693 1864 : ValidateXLOGDirectoryStructure();
5694 :
5695 : /* Set up timeout handler needed to report startup progress. */
5696 1864 : if (!IsBootstrapProcessingMode())
5697 1762 : RegisterTimeout(STARTUP_PROGRESS_TIMEOUT,
5698 : startup_progress_timeout_handler);
5699 :
5700 : /*----------
5701 : * If we previously crashed, perform a couple of actions:
5702 : *
5703 : * - The pg_wal directory may still include some temporary WAL segments
5704 : * used when creating a new segment, so perform some clean up to not
5705 : * bloat this path. This is done first as there is no point to sync
5706 : * this temporary data.
5707 : *
5708 : * - There might be data which we had written, intending to fsync it, but
5709 : * which we had not actually fsync'd yet. Therefore, a power failure in
5710 : * the near future might cause earlier unflushed writes to be lost, even
5711 : * though more recent data written to disk from here on would be
5712 : * persisted. To avoid that, fsync the entire data directory.
5713 : */
5714 1864 : if (ControlFile->state != DB_SHUTDOWNED &&
5715 414 : ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
5716 : {
5717 350 : RemoveTempXlogFiles();
5718 350 : SyncDataDirectory();
5719 350 : didCrash = true;
5720 : }
5721 : else
5722 1514 : didCrash = false;
5723 :
5724 : /*
5725 : * Prepare for WAL recovery if needed.
5726 : *
5727 : * InitWalRecovery analyzes the control file and the backup label file, if
5728 : * any. It updates the in-memory ControlFile buffer according to the
5729 : * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5730 : * It also applies the tablespace map file, if any.
5731 : */
5732 1864 : InitWalRecovery(ControlFile, &wasShutdown,
5733 : &haveBackupLabel, &haveTblspcMap);
5734 1864 : checkPoint = ControlFile->checkPointCopy;
5735 :
5736 : /* initialize shared memory variables from the checkpoint record */
5737 1864 : TransamVariables->nextXid = checkPoint.nextXid;
5738 1864 : TransamVariables->nextOid = checkPoint.nextOid;
5739 1864 : TransamVariables->oidCount = 0;
5740 1864 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5741 1864 : AdvanceOldestClogXid(checkPoint.oldestXid);
5742 1864 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5743 1864 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5744 1864 : SetCommitTsLimit(checkPoint.oldestCommitTsXid,
5745 : checkPoint.newestCommitTsXid);
5746 :
5747 : /*
5748 : * Clear out any old relcache cache files. This is *necessary* if we do
5749 : * any WAL replay, since that would probably result in the cache files
5750 : * being out of sync with database reality. In theory we could leave them
5751 : * in place if the database had been cleanly shut down, but it seems
5752 : * safest to just remove them always and let them be rebuilt during the
5753 : * first backend startup. These files needs to be removed from all
5754 : * directories including pg_tblspc, however the symlinks are created only
5755 : * after reading tablespace_map file in case of archive recovery from
5756 : * backup, so needs to clear old relcache files here after creating
5757 : * symlinks.
5758 : */
5759 1864 : RelationCacheInitFileRemove();
5760 :
5761 : /*
5762 : * Initialize replication slots, before there's a chance to remove
5763 : * required resources.
5764 : */
5765 1864 : StartupReplicationSlots();
5766 :
5767 : /*
5768 : * Startup logical state, needs to be setup now so we have proper data
5769 : * during crash recovery.
5770 : */
5771 1862 : StartupReorderBuffer();
5772 :
5773 : /*
5774 : * Startup CLOG. This must be done after TransamVariables->nextXid has
5775 : * been initialized and before we accept connections or begin WAL replay.
5776 : */
5777 1862 : StartupCLOG();
5778 :
5779 : /*
5780 : * Startup MultiXact. We need to do this early to be able to replay
5781 : * truncations.
5782 : */
5783 1862 : StartupMultiXact();
5784 :
5785 : /*
5786 : * Ditto for commit timestamps. Activate the facility if the setting is
5787 : * enabled in the control file, as there should be no tracking of commit
5788 : * timestamps done when the setting was disabled. This facility can be
5789 : * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
5790 : */
5791 1862 : if (ControlFile->track_commit_timestamp)
5792 26 : StartupCommitTs();
5793 :
5794 : /*
5795 : * Recover knowledge about replay progress of known replication partners.
5796 : */
5797 1862 : StartupReplicationOrigin();
5798 :
5799 : /*
5800 : * Initialize unlogged LSN. On a clean shutdown, it's restored from the
5801 : * control file. On recovery, all unlogged relations are blown away, so
5802 : * the unlogged LSN counter can be reset too.
5803 : */
5804 1862 : if (ControlFile->state == DB_SHUTDOWNED)
5805 1436 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5806 1436 : ControlFile->unloggedLSN);
5807 : else
5808 426 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5809 : FirstNormalUnloggedLSN);
5810 :
5811 : /*
5812 : * Copy any missing timeline history files between 'now' and the recovery
5813 : * target timeline from archive to pg_wal. While we don't need those files
5814 : * ourselves - the history file of the recovery target timeline covers all
5815 : * the previous timelines in the history too - a cascading standby server
5816 : * might be interested in them. Or, if you archive the WAL from this
5817 : * server to a different archive than the primary, it'd be good for all
5818 : * the history files to get archived there after failover, so that you can
5819 : * use one of the old timelines as a PITR target. Timeline history files
5820 : * are small, so it's better to copy them unnecessarily than not copy them
5821 : * and regret later.
5822 : */
5823 1862 : restoreTimeLineHistoryFiles(checkPoint.ThisTimeLineID, recoveryTargetTLI);
5824 :
5825 : /*
5826 : * Before running in recovery, scan pg_twophase and fill in its status to
5827 : * be able to work on entries generated by redo. Doing a scan before
5828 : * taking any recovery action has the merit to discard any 2PC files that
5829 : * are newer than the first record to replay, saving from any conflicts at
5830 : * replay. This avoids as well any subsequent scans when doing recovery
5831 : * of the on-disk two-phase data.
5832 : */
5833 1862 : restoreTwoPhaseData();
5834 :
5835 : /*
5836 : * When starting with crash recovery, reset pgstat data - it might not be
5837 : * valid. Otherwise restore pgstat data. It's safe to do this here,
5838 : * because postmaster will not yet have started any other processes.
5839 : *
5840 : * NB: Restoring replication slot stats relies on slot state to have
5841 : * already been restored from disk.
5842 : *
5843 : * TODO: With a bit of extra work we could just start with a pgstat file
5844 : * associated with the checkpoint redo location we're starting from.
5845 : */
5846 1862 : if (didCrash)
5847 350 : pgstat_discard_stats();
5848 : else
5849 1512 : pgstat_restore_stats();
5850 :
5851 1862 : lastFullPageWrites = checkPoint.fullPageWrites;
5852 :
5853 1862 : RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
5854 1862 : doPageWrites = lastFullPageWrites;
5855 :
5856 : /* REDO */
5857 1862 : if (InRecovery)
5858 : {
5859 : /* Initialize state for RecoveryInProgress() */
5860 426 : SpinLockAcquire(&XLogCtl->info_lck);
5861 426 : if (InArchiveRecovery)
5862 222 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
5863 : else
5864 204 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5865 426 : SpinLockRelease(&XLogCtl->info_lck);
5866 :
5867 : /*
5868 : * Update pg_control to show that we are recovering and to show the
5869 : * selected checkpoint as the place we are starting from. We also mark
5870 : * pg_control with any minimum recovery stop point obtained from a
5871 : * backup history file.
5872 : *
5873 : * No need to hold ControlFileLock yet, we aren't up far enough.
5874 : */
5875 426 : UpdateControlFile();
5876 :
5877 : /*
5878 : * If there was a backup label file, it's done its job and the info
5879 : * has now been propagated into pg_control. We must get rid of the
5880 : * label file so that if we crash during recovery, we'll pick up at
5881 : * the latest recovery restartpoint instead of going all the way back
5882 : * to the backup start point. It seems prudent though to just rename
5883 : * the file out of the way rather than delete it completely.
5884 : */
5885 426 : if (haveBackupLabel)
5886 : {
5887 142 : unlink(BACKUP_LABEL_OLD);
5888 142 : durable_rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD, FATAL);
5889 : }
5890 :
5891 : /*
5892 : * If there was a tablespace_map file, it's done its job and the
5893 : * symlinks have been created. We must get rid of the map file so
5894 : * that if we crash during recovery, we don't create symlinks again.
5895 : * It seems prudent though to just rename the file out of the way
5896 : * rather than delete it completely.
5897 : */
5898 426 : if (haveTblspcMap)
5899 : {
5900 4 : unlink(TABLESPACE_MAP_OLD);
5901 4 : durable_rename(TABLESPACE_MAP, TABLESPACE_MAP_OLD, FATAL);
5902 : }
5903 :
5904 : /*
5905 : * Initialize our local copy of minRecoveryPoint. When doing crash
5906 : * recovery we want to replay up to the end of WAL. Particularly, in
5907 : * the case of a promoted standby minRecoveryPoint value in the
5908 : * control file is only updated after the first checkpoint. However,
5909 : * if the instance crashes before the first post-recovery checkpoint
5910 : * is completed then recovery will use a stale location causing the
5911 : * startup process to think that there are still invalid page
5912 : * references when checking for data consistency.
5913 : */
5914 426 : if (InArchiveRecovery)
5915 : {
5916 222 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
5917 222 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
5918 : }
5919 : else
5920 : {
5921 204 : LocalMinRecoveryPoint = InvalidXLogRecPtr;
5922 204 : LocalMinRecoveryPointTLI = 0;
5923 : }
5924 :
5925 : /* Check that the GUCs used to generate the WAL allow recovery */
5926 426 : CheckRequiredParameterValues();
5927 :
5928 : /*
5929 : * We're in recovery, so unlogged relations may be trashed and must be
5930 : * reset. This should be done BEFORE allowing Hot Standby
5931 : * connections, so that read-only backends don't try to read whatever
5932 : * garbage is left over from before.
5933 : */
5934 426 : ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
5935 :
5936 : /*
5937 : * Likewise, delete any saved transaction snapshot files that got left
5938 : * behind by crashed backends.
5939 : */
5940 426 : DeleteAllExportedSnapshotFiles();
5941 :
5942 : /*
5943 : * Initialize for Hot Standby, if enabled. We won't let backends in
5944 : * yet, not until we've reached the min recovery point specified in
5945 : * control file and we've established a recovery snapshot from a
5946 : * running-xacts WAL record.
5947 : */
5948 426 : if (ArchiveRecoveryRequested && EnableHotStandby)
5949 : {
5950 : TransactionId *xids;
5951 : int nxids;
5952 :
5953 210 : ereport(DEBUG1,
5954 : (errmsg_internal("initializing for hot standby")));
5955 :
5956 210 : InitRecoveryTransactionEnvironment();
5957 :
5958 210 : if (wasShutdown)
5959 52 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
5960 : else
5961 158 : oldestActiveXID = checkPoint.oldestActiveXid;
5962 : Assert(TransactionIdIsValid(oldestActiveXID));
5963 :
5964 : /* Tell procarray about the range of xids it has to deal with */
5965 210 : ProcArrayInitRecovery(XidFromFullTransactionId(TransamVariables->nextXid));
5966 :
5967 : /*
5968 : * Startup subtrans only. CLOG, MultiXact and commit timestamp
5969 : * have already been started up and other SLRUs are not maintained
5970 : * during recovery and need not be started yet.
5971 : */
5972 210 : StartupSUBTRANS(oldestActiveXID);
5973 :
5974 : /*
5975 : * If we're beginning at a shutdown checkpoint, we know that
5976 : * nothing was running on the primary at this point. So fake-up an
5977 : * empty running-xacts record and use that here and now. Recover
5978 : * additional standby state for prepared transactions.
5979 : */
5980 210 : if (wasShutdown)
5981 : {
5982 : RunningTransactionsData running;
5983 : TransactionId latestCompletedXid;
5984 :
5985 : /* Update pg_subtrans entries for any prepared transactions */
5986 52 : StandbyRecoverPreparedTransactions();
5987 :
5988 : /*
5989 : * Construct a RunningTransactions snapshot representing a
5990 : * shut down server, with only prepared transactions still
5991 : * alive. We're never overflowed at this point because all
5992 : * subxids are listed with their parent prepared transactions.
5993 : */
5994 52 : running.xcnt = nxids;
5995 52 : running.subxcnt = 0;
5996 52 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
5997 52 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
5998 52 : running.oldestRunningXid = oldestActiveXID;
5999 52 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
6000 52 : TransactionIdRetreat(latestCompletedXid);
6001 : Assert(TransactionIdIsNormal(latestCompletedXid));
6002 52 : running.latestCompletedXid = latestCompletedXid;
6003 52 : running.xids = xids;
6004 :
6005 52 : ProcArrayApplyRecoveryInfo(&running);
6006 : }
6007 : }
6008 :
6009 : /*
6010 : * We're all set for replaying the WAL now. Do it.
6011 : */
6012 426 : PerformWalRecovery();
6013 308 : performedWalRecovery = true;
6014 : }
6015 : else
6016 1436 : performedWalRecovery = false;
6017 :
6018 : /*
6019 : * Finish WAL recovery.
6020 : */
6021 1744 : endOfRecoveryInfo = FinishWalRecovery();
6022 1744 : EndOfLog = endOfRecoveryInfo->endOfLog;
6023 1744 : EndOfLogTLI = endOfRecoveryInfo->endOfLogTLI;
6024 1744 : abortedRecPtr = endOfRecoveryInfo->abortedRecPtr;
6025 1744 : missingContrecPtr = endOfRecoveryInfo->missingContrecPtr;
6026 :
6027 : /*
6028 : * Reset ps status display, so as no information related to recovery shows
6029 : * up.
6030 : */
6031 1744 : set_ps_display("");
6032 :
6033 : /*
6034 : * When recovering from a backup (we are in recovery, and archive recovery
6035 : * was requested), complain if we did not roll forward far enough to reach
6036 : * the point where the database is consistent. For regular online
6037 : * backup-from-primary, that means reaching the end-of-backup WAL record
6038 : * (at which point we reset backupStartPoint to be Invalid), for
6039 : * backup-from-replica (which can't inject records into the WAL stream),
6040 : * that point is when we reach the minRecoveryPoint in pg_control (which
6041 : * we purposefully copy last when backing up from a replica). For
6042 : * pg_rewind (which creates a backup_label with a method of "pg_rewind")
6043 : * or snapshot-style backups (which don't), backupEndRequired will be set
6044 : * to false.
6045 : *
6046 : * Note: it is indeed okay to look at the local variable
6047 : * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
6048 : * might be further ahead --- ControlFile->minRecoveryPoint cannot have
6049 : * been advanced beyond the WAL we processed.
6050 : */
6051 1744 : if (InRecovery &&
6052 308 : (EndOfLog < LocalMinRecoveryPoint ||
6053 308 : !XLogRecPtrIsInvalid(ControlFile->backupStartPoint)))
6054 : {
6055 : /*
6056 : * Ran off end of WAL before reaching end-of-backup WAL record, or
6057 : * minRecoveryPoint. That's a bad sign, indicating that you tried to
6058 : * recover from an online backup but never called pg_backup_stop(), or
6059 : * you didn't archive all the WAL needed.
6060 : */
6061 0 : if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
6062 : {
6063 0 : if (!XLogRecPtrIsInvalid(ControlFile->backupStartPoint) || ControlFile->backupEndRequired)
6064 0 : ereport(FATAL,
6065 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6066 : errmsg("WAL ends before end of online backup"),
6067 : errhint("All WAL generated while online backup was taken must be available at recovery.")));
6068 : else
6069 0 : ereport(FATAL,
6070 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6071 : errmsg("WAL ends before consistent recovery point")));
6072 : }
6073 : }
6074 :
6075 : /*
6076 : * Reset unlogged relations to the contents of their INIT fork. This is
6077 : * done AFTER recovery is complete so as to include any unlogged relations
6078 : * created during recovery, but BEFORE recovery is marked as having
6079 : * completed successfully. Otherwise we'd not retry if any of the post
6080 : * end-of-recovery steps fail.
6081 : */
6082 1744 : if (InRecovery)
6083 308 : ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
6084 :
6085 : /*
6086 : * Pre-scan prepared transactions to find out the range of XIDs present.
6087 : * This information is not quite needed yet, but it is positioned here so
6088 : * as potential problems are detected before any on-disk change is done.
6089 : */
6090 1744 : oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
6091 :
6092 : /*
6093 : * Allow ordinary WAL segment creation before possibly switching to a new
6094 : * timeline, which creates a new segment, and after the last ReadRecord().
6095 : */
6096 1744 : SetInstallXLogFileSegmentActive();
6097 :
6098 : /*
6099 : * Consider whether we need to assign a new timeline ID.
6100 : *
6101 : * If we did archive recovery, we always assign a new ID. This handles a
6102 : * couple of issues. If we stopped short of the end of WAL during
6103 : * recovery, then we are clearly generating a new timeline and must assign
6104 : * it a unique new ID. Even if we ran to the end, modifying the current
6105 : * last segment is problematic because it may result in trying to
6106 : * overwrite an already-archived copy of that segment, and we encourage
6107 : * DBAs to make their archive_commands reject that. We can dodge the
6108 : * problem by making the new active segment have a new timeline ID.
6109 : *
6110 : * In a normal crash recovery, we can just extend the timeline we were in.
6111 : */
6112 1744 : newTLI = endOfRecoveryInfo->lastRecTLI;
6113 1744 : if (ArchiveRecoveryRequested)
6114 : {
6115 96 : newTLI = findNewestTimeLine(recoveryTargetTLI) + 1;
6116 96 : ereport(LOG,
6117 : (errmsg("selected new timeline ID: %u", newTLI)));
6118 :
6119 : /*
6120 : * Make a writable copy of the last WAL segment. (Note that we also
6121 : * have a copy of the last block of the old WAL in
6122 : * endOfRecovery->lastPage; we will use that below.)
6123 : */
6124 96 : XLogInitNewTimeline(EndOfLogTLI, EndOfLog, newTLI);
6125 :
6126 : /*
6127 : * Remove the signal files out of the way, so that we don't
6128 : * accidentally re-enter archive recovery mode in a subsequent crash.
6129 : */
6130 96 : if (endOfRecoveryInfo->standby_signal_file_found)
6131 90 : durable_unlink(STANDBY_SIGNAL_FILE, FATAL);
6132 :
6133 96 : if (endOfRecoveryInfo->recovery_signal_file_found)
6134 6 : durable_unlink(RECOVERY_SIGNAL_FILE, FATAL);
6135 :
6136 : /*
6137 : * Write the timeline history file, and have it archived. After this
6138 : * point (or rather, as soon as the file is archived), the timeline
6139 : * will appear as "taken" in the WAL archive and to any standby
6140 : * servers. If we crash before actually switching to the new
6141 : * timeline, standby servers will nevertheless think that we switched
6142 : * to the new timeline, and will try to connect to the new timeline.
6143 : * To minimize the window for that, try to do as little as possible
6144 : * between here and writing the end-of-recovery record.
6145 : */
6146 96 : writeTimeLineHistory(newTLI, recoveryTargetTLI,
6147 : EndOfLog, endOfRecoveryInfo->recoveryStopReason);
6148 :
6149 96 : ereport(LOG,
6150 : (errmsg("archive recovery complete")));
6151 : }
6152 :
6153 : /* Save the selected TimeLineID in shared memory, too */
6154 1744 : SpinLockAcquire(&XLogCtl->info_lck);
6155 1744 : XLogCtl->InsertTimeLineID = newTLI;
6156 1744 : XLogCtl->PrevTimeLineID = endOfRecoveryInfo->lastRecTLI;
6157 1744 : SpinLockRelease(&XLogCtl->info_lck);
6158 :
6159 : /*
6160 : * Actually, if WAL ended in an incomplete record, skip the parts that
6161 : * made it through and start writing after the portion that persisted.
6162 : * (It's critical to first write an OVERWRITE_CONTRECORD message, which
6163 : * we'll do as soon as we're open for writing new WAL.)
6164 : */
6165 1744 : if (!XLogRecPtrIsInvalid(missingContrecPtr))
6166 : {
6167 : /*
6168 : * We should only have a missingContrecPtr if we're not switching to a
6169 : * new timeline. When a timeline switch occurs, WAL is copied from the
6170 : * old timeline to the new only up to the end of the last complete
6171 : * record, so there can't be an incomplete WAL record that we need to
6172 : * disregard.
6173 : */
6174 : Assert(newTLI == endOfRecoveryInfo->lastRecTLI);
6175 : Assert(!XLogRecPtrIsInvalid(abortedRecPtr));
6176 20 : EndOfLog = missingContrecPtr;
6177 : }
6178 :
6179 : /*
6180 : * Prepare to write WAL starting at EndOfLog location, and init xlog
6181 : * buffer cache using the block containing the last record from the
6182 : * previous incarnation.
6183 : */
6184 1744 : Insert = &XLogCtl->Insert;
6185 1744 : Insert->PrevBytePos = XLogRecPtrToBytePos(endOfRecoveryInfo->lastRec);
6186 1744 : Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
6187 :
6188 : /*
6189 : * Tricky point here: lastPage contains the *last* block that the LastRec
6190 : * record spans, not the one it starts in. The last block is indeed the
6191 : * one we want to use.
6192 : */
6193 1744 : if (EndOfLog % XLOG_BLCKSZ != 0)
6194 : {
6195 : char *page;
6196 : int len;
6197 : int firstIdx;
6198 :
6199 1688 : firstIdx = XLogRecPtrToBufIdx(EndOfLog);
6200 1688 : len = EndOfLog - endOfRecoveryInfo->lastPageBeginPtr;
6201 : Assert(len < XLOG_BLCKSZ);
6202 :
6203 : /* Copy the valid part of the last block, and zero the rest */
6204 1688 : page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
6205 1688 : memcpy(page, endOfRecoveryInfo->lastPage, len);
6206 1688 : memset(page + len, 0, XLOG_BLCKSZ - len);
6207 :
6208 1688 : pg_atomic_write_u64(&XLogCtl->xlblocks[firstIdx], endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6209 1688 : pg_atomic_write_u64(&XLogCtl->InitializedUpTo, endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6210 1688 : XLogCtl->InitializedFrom = endOfRecoveryInfo->lastPageBeginPtr;
6211 : }
6212 : else
6213 : {
6214 : /*
6215 : * There is no partial block to copy. Just set InitializedUpTo, and
6216 : * let the first attempt to insert a log record to initialize the next
6217 : * buffer.
6218 : */
6219 56 : pg_atomic_write_u64(&XLogCtl->InitializedUpTo, EndOfLog);
6220 56 : XLogCtl->InitializedFrom = EndOfLog;
6221 : }
6222 1744 : pg_atomic_write_u64(&XLogCtl->InitializeReserved, pg_atomic_read_u64(&XLogCtl->InitializedUpTo));
6223 :
6224 : /*
6225 : * Update local and shared status. This is OK to do without any locks
6226 : * because no other process can be reading or writing WAL yet.
6227 : */
6228 1744 : LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
6229 1744 : pg_atomic_write_u64(&XLogCtl->logInsertResult, EndOfLog);
6230 1744 : pg_atomic_write_u64(&XLogCtl->logWriteResult, EndOfLog);
6231 1744 : pg_atomic_write_u64(&XLogCtl->logFlushResult, EndOfLog);
6232 1744 : XLogCtl->LogwrtRqst.Write = EndOfLog;
6233 1744 : XLogCtl->LogwrtRqst.Flush = EndOfLog;
6234 :
6235 : /*
6236 : * Preallocate additional log files, if wanted.
6237 : */
6238 1744 : PreallocXlogFiles(EndOfLog, newTLI);
6239 :
6240 : /*
6241 : * Okay, we're officially UP.
6242 : */
6243 1744 : InRecovery = false;
6244 :
6245 : /* start the archive_timeout timer and LSN running */
6246 1744 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
6247 1744 : XLogCtl->lastSegSwitchLSN = EndOfLog;
6248 :
6249 : /* also initialize latestCompletedXid, to nextXid - 1 */
6250 1744 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
6251 1744 : TransamVariables->latestCompletedXid = TransamVariables->nextXid;
6252 1744 : FullTransactionIdRetreat(&TransamVariables->latestCompletedXid);
6253 1744 : LWLockRelease(ProcArrayLock);
6254 :
6255 : /*
6256 : * Start up subtrans, if not already done for hot standby. (commit
6257 : * timestamps are started below, if necessary.)
6258 : */
6259 1744 : if (standbyState == STANDBY_DISABLED)
6260 1648 : StartupSUBTRANS(oldestActiveXID);
6261 :
6262 : /*
6263 : * Perform end of recovery actions for any SLRUs that need it.
6264 : */
6265 1744 : TrimCLOG();
6266 1744 : TrimMultiXact();
6267 :
6268 : /*
6269 : * Reload shared-memory state for prepared transactions. This needs to
6270 : * happen before renaming the last partial segment of the old timeline as
6271 : * it may be possible that we have to recover some transactions from it.
6272 : */
6273 1744 : RecoverPreparedTransactions();
6274 :
6275 : /* Shut down xlogreader */
6276 1744 : ShutdownWalRecovery();
6277 :
6278 : /* Enable WAL writes for this backend only. */
6279 1744 : LocalSetXLogInsertAllowed();
6280 :
6281 : /* If necessary, write overwrite-contrecord before doing anything else */
6282 1744 : if (!XLogRecPtrIsInvalid(abortedRecPtr))
6283 : {
6284 : Assert(!XLogRecPtrIsInvalid(missingContrecPtr));
6285 20 : CreateOverwriteContrecordRecord(abortedRecPtr, missingContrecPtr, newTLI);
6286 : }
6287 :
6288 : /*
6289 : * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
6290 : * record before resource manager writes cleanup WAL records or checkpoint
6291 : * record is written.
6292 : */
6293 1744 : Insert->fullPageWrites = lastFullPageWrites;
6294 1744 : UpdateFullPageWrites();
6295 :
6296 : /*
6297 : * Emit checkpoint or end-of-recovery record in XLOG, if required.
6298 : */
6299 1744 : if (performedWalRecovery)
6300 308 : promoted = PerformRecoveryXLogAction();
6301 :
6302 : /*
6303 : * If any of the critical GUCs have changed, log them before we allow
6304 : * backends to write WAL.
6305 : */
6306 1744 : XLogReportParameters();
6307 :
6308 : /* If this is archive recovery, perform post-recovery cleanup actions. */
6309 1744 : if (ArchiveRecoveryRequested)
6310 96 : CleanupAfterArchiveRecovery(EndOfLogTLI, EndOfLog, newTLI);
6311 :
6312 : /*
6313 : * Local WAL inserts enabled, so it's time to finish initialization of
6314 : * commit timestamp.
6315 : */
6316 1744 : CompleteCommitTsInitialization();
6317 :
6318 : /*
6319 : * All done with end-of-recovery actions.
6320 : *
6321 : * Now allow backends to write WAL and update the control file status in
6322 : * consequence. SharedRecoveryState, that controls if backends can write
6323 : * WAL, is updated while holding ControlFileLock to prevent other backends
6324 : * to look at an inconsistent state of the control file in shared memory.
6325 : * There is still a small window during which backends can write WAL and
6326 : * the control file is still referring to a system not in DB_IN_PRODUCTION
6327 : * state while looking at the on-disk control file.
6328 : *
6329 : * Also, we use info_lck to update SharedRecoveryState to ensure that
6330 : * there are no race conditions concerning visibility of other recent
6331 : * updates to shared memory.
6332 : */
6333 1744 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6334 1744 : ControlFile->state = DB_IN_PRODUCTION;
6335 :
6336 1744 : SpinLockAcquire(&XLogCtl->info_lck);
6337 1744 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_DONE;
6338 1744 : SpinLockRelease(&XLogCtl->info_lck);
6339 :
6340 1744 : UpdateControlFile();
6341 1744 : LWLockRelease(ControlFileLock);
6342 :
6343 : /*
6344 : * Shutdown the recovery environment. This must occur after
6345 : * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
6346 : * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
6347 : * any session building a snapshot will not rely on KnownAssignedXids as
6348 : * RecoveryInProgress() would return false at this stage. This is
6349 : * particularly critical for prepared 2PC transactions, that would still
6350 : * need to be included in snapshots once recovery has ended.
6351 : */
6352 1744 : if (standbyState != STANDBY_DISABLED)
6353 96 : ShutdownRecoveryTransactionEnvironment();
6354 :
6355 : /*
6356 : * If there were cascading standby servers connected to us, nudge any wal
6357 : * sender processes to notice that we've been promoted.
6358 : */
6359 1744 : WalSndWakeup(true, true);
6360 :
6361 : /*
6362 : * If this was a promotion, request an (online) checkpoint now. This isn't
6363 : * required for consistency, but the last restartpoint might be far back,
6364 : * and in case of a crash, recovering from it might take a longer than is
6365 : * appropriate now that we're not in standby mode anymore.
6366 : */
6367 1744 : if (promoted)
6368 82 : RequestCheckpoint(CHECKPOINT_FORCE);
6369 1744 : }
6370 :
6371 : /*
6372 : * Callback from PerformWalRecovery(), called when we switch from crash
6373 : * recovery to archive recovery mode. Updates the control file accordingly.
6374 : */
6375 : void
6376 4 : SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr, TimeLineID replayTLI)
6377 : {
6378 : /* initialize minRecoveryPoint to this record */
6379 4 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6380 4 : ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6381 4 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6382 : {
6383 4 : ControlFile->minRecoveryPoint = EndRecPtr;
6384 4 : ControlFile->minRecoveryPointTLI = replayTLI;
6385 : }
6386 : /* update local copy */
6387 4 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6388 4 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6389 :
6390 : /*
6391 : * The startup process can update its local copy of minRecoveryPoint from
6392 : * this point.
6393 : */
6394 4 : updateMinRecoveryPoint = true;
6395 :
6396 4 : UpdateControlFile();
6397 :
6398 : /*
6399 : * We update SharedRecoveryState while holding the lock on ControlFileLock
6400 : * so both states are consistent in shared memory.
6401 : */
6402 4 : SpinLockAcquire(&XLogCtl->info_lck);
6403 4 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6404 4 : SpinLockRelease(&XLogCtl->info_lck);
6405 :
6406 4 : LWLockRelease(ControlFileLock);
6407 4 : }
6408 :
6409 : /*
6410 : * Callback from PerformWalRecovery(), called when we reach the end of backup.
6411 : * Updates the control file accordingly.
6412 : */
6413 : void
6414 142 : ReachedEndOfBackup(XLogRecPtr EndRecPtr, TimeLineID tli)
6415 : {
6416 : /*
6417 : * We have reached the end of base backup, as indicated by pg_control. The
6418 : * data on disk is now consistent (unless minRecoveryPoint is further
6419 : * ahead, which can happen if we crashed during previous recovery). Reset
6420 : * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
6421 : * make sure we don't allow starting up at an earlier point even if
6422 : * recovery is stopped and restarted soon after this.
6423 : */
6424 142 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6425 :
6426 142 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6427 : {
6428 134 : ControlFile->minRecoveryPoint = EndRecPtr;
6429 134 : ControlFile->minRecoveryPointTLI = tli;
6430 : }
6431 :
6432 142 : ControlFile->backupStartPoint = InvalidXLogRecPtr;
6433 142 : ControlFile->backupEndPoint = InvalidXLogRecPtr;
6434 142 : ControlFile->backupEndRequired = false;
6435 142 : UpdateControlFile();
6436 :
6437 142 : LWLockRelease(ControlFileLock);
6438 142 : }
6439 :
6440 : /*
6441 : * Perform whatever XLOG actions are necessary at end of REDO.
6442 : *
6443 : * The goal here is to make sure that we'll be able to recover properly if
6444 : * we crash again. If we choose to write a checkpoint, we'll write a shutdown
6445 : * checkpoint rather than an on-line one. This is not particularly critical,
6446 : * but since we may be assigning a new TLI, using a shutdown checkpoint allows
6447 : * us to have the rule that TLI only changes in shutdown checkpoints, which
6448 : * allows some extra error checking in xlog_redo.
6449 : */
6450 : static bool
6451 308 : PerformRecoveryXLogAction(void)
6452 : {
6453 308 : bool promoted = false;
6454 :
6455 : /*
6456 : * Perform a checkpoint to update all our recovery activity to disk.
6457 : *
6458 : * Note that we write a shutdown checkpoint rather than an on-line one.
6459 : * This is not particularly critical, but since we may be assigning a new
6460 : * TLI, using a shutdown checkpoint allows us to have the rule that TLI
6461 : * only changes in shutdown checkpoints, which allows some extra error
6462 : * checking in xlog_redo.
6463 : *
6464 : * In promotion, only create a lightweight end-of-recovery record instead
6465 : * of a full checkpoint. A checkpoint is requested later, after we're
6466 : * fully out of recovery mode and already accepting queries.
6467 : */
6468 404 : if (ArchiveRecoveryRequested && IsUnderPostmaster &&
6469 96 : PromoteIsTriggered())
6470 : {
6471 82 : promoted = true;
6472 :
6473 : /*
6474 : * Insert a special WAL record to mark the end of recovery, since we
6475 : * aren't doing a checkpoint. That means that the checkpointer process
6476 : * may likely be in the middle of a time-smoothed restartpoint and
6477 : * could continue to be for minutes after this. That sounds strange,
6478 : * but the effect is roughly the same and it would be stranger to try
6479 : * to come out of the restartpoint and then checkpoint. We request a
6480 : * checkpoint later anyway, just for safety.
6481 : */
6482 82 : CreateEndOfRecoveryRecord();
6483 : }
6484 : else
6485 : {
6486 226 : RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
6487 : CHECKPOINT_FAST |
6488 : CHECKPOINT_WAIT);
6489 : }
6490 :
6491 308 : return promoted;
6492 : }
6493 :
6494 : /*
6495 : * Is the system still in recovery?
6496 : *
6497 : * Unlike testing InRecovery, this works in any process that's connected to
6498 : * shared memory.
6499 : */
6500 : bool
6501 178382206 : RecoveryInProgress(void)
6502 : {
6503 : /*
6504 : * We check shared state each time only until we leave recovery mode. We
6505 : * can't re-enter recovery, so there's no need to keep checking after the
6506 : * shared variable has once been seen false.
6507 : */
6508 178382206 : if (!LocalRecoveryInProgress)
6509 173881926 : return false;
6510 : else
6511 : {
6512 : /*
6513 : * use volatile pointer to make sure we make a fresh read of the
6514 : * shared variable.
6515 : */
6516 4500280 : volatile XLogCtlData *xlogctl = XLogCtl;
6517 :
6518 4500280 : LocalRecoveryInProgress = (xlogctl->SharedRecoveryState != RECOVERY_STATE_DONE);
6519 :
6520 : /*
6521 : * Note: We don't need a memory barrier when we're still in recovery.
6522 : * We might exit recovery immediately after return, so the caller
6523 : * can't rely on 'true' meaning that we're still in recovery anyway.
6524 : */
6525 :
6526 4500280 : return LocalRecoveryInProgress;
6527 : }
6528 : }
6529 :
6530 : /*
6531 : * Returns current recovery state from shared memory.
6532 : *
6533 : * This returned state is kept consistent with the contents of the control
6534 : * file. See details about the possible values of RecoveryState in xlog.h.
6535 : */
6536 : RecoveryState
6537 41664 : GetRecoveryState(void)
6538 : {
6539 : RecoveryState retval;
6540 :
6541 41664 : SpinLockAcquire(&XLogCtl->info_lck);
6542 41664 : retval = XLogCtl->SharedRecoveryState;
6543 41664 : SpinLockRelease(&XLogCtl->info_lck);
6544 :
6545 41664 : return retval;
6546 : }
6547 :
6548 : /*
6549 : * Is this process allowed to insert new WAL records?
6550 : *
6551 : * Ordinarily this is essentially equivalent to !RecoveryInProgress().
6552 : * But we also have provisions for forcing the result "true" or "false"
6553 : * within specific processes regardless of the global state.
6554 : */
6555 : bool
6556 61064540 : XLogInsertAllowed(void)
6557 : {
6558 : /*
6559 : * If value is "unconditionally true" or "unconditionally false", just
6560 : * return it. This provides the normal fast path once recovery is known
6561 : * done.
6562 : */
6563 61064540 : if (LocalXLogInsertAllowed >= 0)
6564 60841782 : return (bool) LocalXLogInsertAllowed;
6565 :
6566 : /*
6567 : * Else, must check to see if we're still in recovery.
6568 : */
6569 222758 : if (RecoveryInProgress())
6570 206328 : return false;
6571 :
6572 : /*
6573 : * On exit from recovery, reset to "unconditionally true", since there is
6574 : * no need to keep checking.
6575 : */
6576 16430 : LocalXLogInsertAllowed = 1;
6577 16430 : return true;
6578 : }
6579 :
6580 : /*
6581 : * Make XLogInsertAllowed() return true in the current process only.
6582 : *
6583 : * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
6584 : * and even call LocalSetXLogInsertAllowed() again after that.
6585 : *
6586 : * Returns the previous value of LocalXLogInsertAllowed.
6587 : */
6588 : static int
6589 1802 : LocalSetXLogInsertAllowed(void)
6590 : {
6591 1802 : int oldXLogAllowed = LocalXLogInsertAllowed;
6592 :
6593 1802 : LocalXLogInsertAllowed = 1;
6594 :
6595 1802 : return oldXLogAllowed;
6596 : }
6597 :
6598 : /*
6599 : * Return the current Redo pointer from shared memory.
6600 : *
6601 : * As a side-effect, the local RedoRecPtr copy is updated.
6602 : */
6603 : XLogRecPtr
6604 593360 : GetRedoRecPtr(void)
6605 : {
6606 : XLogRecPtr ptr;
6607 :
6608 : /*
6609 : * The possibly not up-to-date copy in XlogCtl is enough. Even if we
6610 : * grabbed a WAL insertion lock to read the authoritative value in
6611 : * Insert->RedoRecPtr, someone might update it just after we've released
6612 : * the lock.
6613 : */
6614 593360 : SpinLockAcquire(&XLogCtl->info_lck);
6615 593360 : ptr = XLogCtl->RedoRecPtr;
6616 593360 : SpinLockRelease(&XLogCtl->info_lck);
6617 :
6618 593360 : if (RedoRecPtr < ptr)
6619 2922 : RedoRecPtr = ptr;
6620 :
6621 593360 : return RedoRecPtr;
6622 : }
6623 :
6624 : /*
6625 : * Return information needed to decide whether a modified block needs a
6626 : * full-page image to be included in the WAL record.
6627 : *
6628 : * The returned values are cached copies from backend-private memory, and
6629 : * possibly out-of-date or, indeed, uninitialized, in which case they will
6630 : * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6631 : * re-check them against up-to-date values, while holding the WAL insert lock.
6632 : */
6633 : void
6634 29514322 : GetFullPageWriteInfo(XLogRecPtr *RedoRecPtr_p, bool *doPageWrites_p)
6635 : {
6636 29514322 : *RedoRecPtr_p = RedoRecPtr;
6637 29514322 : *doPageWrites_p = doPageWrites;
6638 29514322 : }
6639 :
6640 : /*
6641 : * GetInsertRecPtr -- Returns the current insert position.
6642 : *
6643 : * NOTE: The value *actually* returned is the position of the last full
6644 : * xlog page. It lags behind the real insert position by at most 1 page.
6645 : * For that, we don't need to scan through WAL insertion locks, and an
6646 : * approximation is enough for the current usage of this function.
6647 : */
6648 : XLogRecPtr
6649 17218 : GetInsertRecPtr(void)
6650 : {
6651 : XLogRecPtr recptr;
6652 :
6653 17218 : SpinLockAcquire(&XLogCtl->info_lck);
6654 17218 : recptr = XLogCtl->LogwrtRqst.Write;
6655 17218 : SpinLockRelease(&XLogCtl->info_lck);
6656 :
6657 17218 : return recptr;
6658 : }
6659 :
6660 : /*
6661 : * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6662 : * position known to be fsync'd to disk. This should only be used on a
6663 : * system that is known not to be in recovery.
6664 : */
6665 : XLogRecPtr
6666 419820 : GetFlushRecPtr(TimeLineID *insertTLI)
6667 : {
6668 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6669 :
6670 419820 : RefreshXLogWriteResult(LogwrtResult);
6671 :
6672 : /*
6673 : * If we're writing and flushing WAL, the time line can't be changing, so
6674 : * no lock is required.
6675 : */
6676 419820 : if (insertTLI)
6677 43642 : *insertTLI = XLogCtl->InsertTimeLineID;
6678 :
6679 419820 : return LogwrtResult.Flush;
6680 : }
6681 :
6682 : /*
6683 : * GetWALInsertionTimeLine -- Returns the current timeline of a system that
6684 : * is not in recovery.
6685 : */
6686 : TimeLineID
6687 212122 : GetWALInsertionTimeLine(void)
6688 : {
6689 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6690 :
6691 : /* Since the value can't be changing, no lock is required. */
6692 212122 : return XLogCtl->InsertTimeLineID;
6693 : }
6694 :
6695 : /*
6696 : * GetWALInsertionTimeLineIfSet -- If the system is not in recovery, returns
6697 : * the WAL insertion timeline; else, returns 0. Wherever possible, use
6698 : * GetWALInsertionTimeLine() instead, since it's cheaper. Note that this
6699 : * function decides recovery has ended as soon as the insert TLI is set, which
6700 : * happens before we set XLogCtl->SharedRecoveryState to RECOVERY_STATE_DONE.
6701 : */
6702 : TimeLineID
6703 0 : GetWALInsertionTimeLineIfSet(void)
6704 : {
6705 : TimeLineID insertTLI;
6706 :
6707 0 : SpinLockAcquire(&XLogCtl->info_lck);
6708 0 : insertTLI = XLogCtl->InsertTimeLineID;
6709 0 : SpinLockRelease(&XLogCtl->info_lck);
6710 :
6711 0 : return insertTLI;
6712 : }
6713 :
6714 : /*
6715 : * GetLastImportantRecPtr -- Returns the LSN of the last important record
6716 : * inserted. All records not explicitly marked as unimportant are considered
6717 : * important.
6718 : *
6719 : * The LSN is determined by computing the maximum of
6720 : * WALInsertLocks[i].lastImportantAt.
6721 : */
6722 : XLogRecPtr
6723 3094 : GetLastImportantRecPtr(void)
6724 : {
6725 3094 : XLogRecPtr res = InvalidXLogRecPtr;
6726 : int i;
6727 :
6728 27846 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
6729 : {
6730 : XLogRecPtr last_important;
6731 :
6732 : /*
6733 : * Need to take a lock to prevent torn reads of the LSN, which are
6734 : * possible on some of the supported platforms. WAL insert locks only
6735 : * support exclusive mode, so we have to use that.
6736 : */
6737 24752 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
6738 24752 : last_important = WALInsertLocks[i].l.lastImportantAt;
6739 24752 : LWLockRelease(&WALInsertLocks[i].l.lock);
6740 :
6741 24752 : if (res < last_important)
6742 5434 : res = last_important;
6743 : }
6744 :
6745 3094 : return res;
6746 : }
6747 :
6748 : /*
6749 : * Get the time and LSN of the last xlog segment switch
6750 : */
6751 : pg_time_t
6752 0 : GetLastSegSwitchData(XLogRecPtr *lastSwitchLSN)
6753 : {
6754 : pg_time_t result;
6755 :
6756 : /* Need WALWriteLock, but shared lock is sufficient */
6757 0 : LWLockAcquire(WALWriteLock, LW_SHARED);
6758 0 : result = XLogCtl->lastSegSwitchTime;
6759 0 : *lastSwitchLSN = XLogCtl->lastSegSwitchLSN;
6760 0 : LWLockRelease(WALWriteLock);
6761 :
6762 0 : return result;
6763 : }
6764 :
6765 : /*
6766 : * This must be called ONCE during postmaster or standalone-backend shutdown
6767 : */
6768 : void
6769 1252 : ShutdownXLOG(int code, Datum arg)
6770 : {
6771 : /*
6772 : * We should have an aux process resource owner to use, and we should not
6773 : * be in a transaction that's installed some other resowner.
6774 : */
6775 : Assert(AuxProcessResourceOwner != NULL);
6776 : Assert(CurrentResourceOwner == NULL ||
6777 : CurrentResourceOwner == AuxProcessResourceOwner);
6778 1252 : CurrentResourceOwner = AuxProcessResourceOwner;
6779 :
6780 : /* Don't be chatty in standalone mode */
6781 1252 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
6782 : (errmsg("shutting down")));
6783 :
6784 : /*
6785 : * Signal walsenders to move to stopping state.
6786 : */
6787 1252 : WalSndInitStopping();
6788 :
6789 : /*
6790 : * Wait for WAL senders to be in stopping state. This prevents commands
6791 : * from writing new WAL.
6792 : */
6793 1252 : WalSndWaitStopping();
6794 :
6795 1252 : if (RecoveryInProgress())
6796 110 : CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
6797 : else
6798 : {
6799 : /*
6800 : * If archiving is enabled, rotate the last XLOG file so that all the
6801 : * remaining records are archived (postmaster wakes up the archiver
6802 : * process one more time at the end of shutdown). The checkpoint
6803 : * record will go to the next XLOG file and won't be archived (yet).
6804 : */
6805 1142 : if (XLogArchivingActive())
6806 28 : RequestXLogSwitch(false);
6807 :
6808 1142 : CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
6809 : }
6810 1252 : }
6811 :
6812 : /*
6813 : * Log start of a checkpoint.
6814 : */
6815 : static void
6816 2782 : LogCheckpointStart(int flags, bool restartpoint)
6817 : {
6818 2782 : if (restartpoint)
6819 374 : ereport(LOG,
6820 : /* translator: the placeholders show checkpoint options */
6821 : (errmsg("restartpoint starting:%s%s%s%s%s%s%s%s",
6822 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6823 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6824 : (flags & CHECKPOINT_FAST) ? " fast" : "",
6825 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6826 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6827 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6828 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6829 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "")));
6830 : else
6831 2408 : ereport(LOG,
6832 : /* translator: the placeholders show checkpoint options */
6833 : (errmsg("checkpoint starting:%s%s%s%s%s%s%s%s",
6834 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6835 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6836 : (flags & CHECKPOINT_FAST) ? " fast" : "",
6837 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6838 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6839 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6840 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6841 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "")));
6842 2782 : }
6843 :
6844 : /*
6845 : * Log end of a checkpoint.
6846 : */
6847 : static void
6848 3378 : LogCheckpointEnd(bool restartpoint)
6849 : {
6850 : long write_msecs,
6851 : sync_msecs,
6852 : total_msecs,
6853 : longest_msecs,
6854 : average_msecs;
6855 : uint64 average_sync_time;
6856 :
6857 3378 : CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
6858 :
6859 3378 : write_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_write_t,
6860 : CheckpointStats.ckpt_sync_t);
6861 :
6862 3378 : sync_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_sync_t,
6863 : CheckpointStats.ckpt_sync_end_t);
6864 :
6865 : /* Accumulate checkpoint timing summary data, in milliseconds. */
6866 3378 : PendingCheckpointerStats.write_time += write_msecs;
6867 3378 : PendingCheckpointerStats.sync_time += sync_msecs;
6868 :
6869 : /*
6870 : * All of the published timing statistics are accounted for. Only
6871 : * continue if a log message is to be written.
6872 : */
6873 3378 : if (!log_checkpoints)
6874 596 : return;
6875 :
6876 2782 : total_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_start_t,
6877 : CheckpointStats.ckpt_end_t);
6878 :
6879 : /*
6880 : * Timing values returned from CheckpointStats are in microseconds.
6881 : * Convert to milliseconds for consistent printing.
6882 : */
6883 2782 : longest_msecs = (long) ((CheckpointStats.ckpt_longest_sync + 999) / 1000);
6884 :
6885 2782 : average_sync_time = 0;
6886 2782 : if (CheckpointStats.ckpt_sync_rels > 0)
6887 0 : average_sync_time = CheckpointStats.ckpt_agg_sync_time /
6888 0 : CheckpointStats.ckpt_sync_rels;
6889 2782 : average_msecs = (long) ((average_sync_time + 999) / 1000);
6890 :
6891 : /*
6892 : * ControlFileLock is not required to see ControlFile->checkPoint and
6893 : * ->checkPointCopy here as we are the only updator of those variables at
6894 : * this moment.
6895 : */
6896 2782 : if (restartpoint)
6897 374 : ereport(LOG,
6898 : (errmsg("restartpoint complete: wrote %d buffers (%.1f%%), "
6899 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6900 : "%d removed, %d recycled; write=%ld.%03d s, "
6901 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6902 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6903 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6904 : CheckpointStats.ckpt_bufs_written,
6905 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6906 : CheckpointStats.ckpt_slru_written,
6907 : CheckpointStats.ckpt_segs_added,
6908 : CheckpointStats.ckpt_segs_removed,
6909 : CheckpointStats.ckpt_segs_recycled,
6910 : write_msecs / 1000, (int) (write_msecs % 1000),
6911 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6912 : total_msecs / 1000, (int) (total_msecs % 1000),
6913 : CheckpointStats.ckpt_sync_rels,
6914 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6915 : average_msecs / 1000, (int) (average_msecs % 1000),
6916 : (int) (PrevCheckPointDistance / 1024.0),
6917 : (int) (CheckPointDistanceEstimate / 1024.0),
6918 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6919 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6920 : else
6921 2408 : ereport(LOG,
6922 : (errmsg("checkpoint complete: wrote %d buffers (%.1f%%), "
6923 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6924 : "%d removed, %d recycled; write=%ld.%03d s, "
6925 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6926 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6927 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6928 : CheckpointStats.ckpt_bufs_written,
6929 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6930 : CheckpointStats.ckpt_slru_written,
6931 : CheckpointStats.ckpt_segs_added,
6932 : CheckpointStats.ckpt_segs_removed,
6933 : CheckpointStats.ckpt_segs_recycled,
6934 : write_msecs / 1000, (int) (write_msecs % 1000),
6935 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6936 : total_msecs / 1000, (int) (total_msecs % 1000),
6937 : CheckpointStats.ckpt_sync_rels,
6938 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6939 : average_msecs / 1000, (int) (average_msecs % 1000),
6940 : (int) (PrevCheckPointDistance / 1024.0),
6941 : (int) (CheckPointDistanceEstimate / 1024.0),
6942 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6943 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6944 : }
6945 :
6946 : /*
6947 : * Update the estimate of distance between checkpoints.
6948 : *
6949 : * The estimate is used to calculate the number of WAL segments to keep
6950 : * preallocated, see XLOGfileslop().
6951 : */
6952 : static void
6953 3378 : UpdateCheckPointDistanceEstimate(uint64 nbytes)
6954 : {
6955 : /*
6956 : * To estimate the number of segments consumed between checkpoints, keep a
6957 : * moving average of the amount of WAL generated in previous checkpoint
6958 : * cycles. However, if the load is bursty, with quiet periods and busy
6959 : * periods, we want to cater for the peak load. So instead of a plain
6960 : * moving average, let the average decline slowly if the previous cycle
6961 : * used less WAL than estimated, but bump it up immediately if it used
6962 : * more.
6963 : *
6964 : * When checkpoints are triggered by max_wal_size, this should converge to
6965 : * CheckpointSegments * wal_segment_size,
6966 : *
6967 : * Note: This doesn't pay any attention to what caused the checkpoint.
6968 : * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
6969 : * starting a base backup, are counted the same as those created
6970 : * automatically. The slow-decline will largely mask them out, if they are
6971 : * not frequent. If they are frequent, it seems reasonable to count them
6972 : * in as any others; if you issue a manual checkpoint every 5 minutes and
6973 : * never let a timed checkpoint happen, it makes sense to base the
6974 : * preallocation on that 5 minute interval rather than whatever
6975 : * checkpoint_timeout is set to.
6976 : */
6977 3378 : PrevCheckPointDistance = nbytes;
6978 3378 : if (CheckPointDistanceEstimate < nbytes)
6979 1416 : CheckPointDistanceEstimate = nbytes;
6980 : else
6981 1962 : CheckPointDistanceEstimate =
6982 1962 : (0.90 * CheckPointDistanceEstimate + 0.10 * (double) nbytes);
6983 3378 : }
6984 :
6985 : /*
6986 : * Update the ps display for a process running a checkpoint. Note that
6987 : * this routine should not do any allocations so as it can be called
6988 : * from a critical section.
6989 : */
6990 : static void
6991 6756 : update_checkpoint_display(int flags, bool restartpoint, bool reset)
6992 : {
6993 : /*
6994 : * The status is reported only for end-of-recovery and shutdown
6995 : * checkpoints or shutdown restartpoints. Updating the ps display is
6996 : * useful in those situations as it may not be possible to rely on
6997 : * pg_stat_activity to see the status of the checkpointer or the startup
6998 : * process.
6999 : */
7000 6756 : if ((flags & (CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IS_SHUTDOWN)) == 0)
7001 4280 : return;
7002 :
7003 2476 : if (reset)
7004 1238 : set_ps_display("");
7005 : else
7006 : {
7007 : char activitymsg[128];
7008 :
7009 3714 : snprintf(activitymsg, sizeof(activitymsg), "performing %s%s%s",
7010 1238 : (flags & CHECKPOINT_END_OF_RECOVERY) ? "end-of-recovery " : "",
7011 1238 : (flags & CHECKPOINT_IS_SHUTDOWN) ? "shutdown " : "",
7012 : restartpoint ? "restartpoint" : "checkpoint");
7013 1238 : set_ps_display(activitymsg);
7014 : }
7015 : }
7016 :
7017 :
7018 : /*
7019 : * Perform a checkpoint --- either during shutdown, or on-the-fly
7020 : *
7021 : * flags is a bitwise OR of the following:
7022 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
7023 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
7024 : * CHECKPOINT_FAST: finish the checkpoint ASAP, ignoring
7025 : * checkpoint_completion_target parameter.
7026 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
7027 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
7028 : * CHECKPOINT_END_OF_RECOVERY).
7029 : * CHECKPOINT_FLUSH_UNLOGGED: also flush buffers of unlogged tables.
7030 : *
7031 : * Note: flags contains other bits, of interest here only for logging purposes.
7032 : * In particular note that this routine is synchronous and does not pay
7033 : * attention to CHECKPOINT_WAIT.
7034 : *
7035 : * If !shutdown then we are writing an online checkpoint. An XLOG_CHECKPOINT_REDO
7036 : * record is inserted into WAL at the logical location of the checkpoint, before
7037 : * flushing anything to disk, and when the checkpoint is eventually completed,
7038 : * and it is from this point that WAL replay will begin in the case of a recovery
7039 : * from this checkpoint. Once everything is written to disk, an
7040 : * XLOG_CHECKPOINT_ONLINE record is written to complete the checkpoint, and
7041 : * points back to the earlier XLOG_CHECKPOINT_REDO record. This mechanism allows
7042 : * other write-ahead log records to be written while the checkpoint is in
7043 : * progress, but we must be very careful about order of operations. This function
7044 : * may take many minutes to execute on a busy system.
7045 : *
7046 : * On the other hand, when shutdown is true, concurrent insertion into the
7047 : * write-ahead log is impossible, so there is no need for two separate records.
7048 : * In this case, we only insert an XLOG_CHECKPOINT_SHUTDOWN record, and it's
7049 : * both the record marking the completion of the checkpoint and the location
7050 : * from which WAL replay would begin if needed.
7051 : *
7052 : * Returns true if a new checkpoint was performed, or false if it was skipped
7053 : * because the system was idle.
7054 : */
7055 : bool
7056 3004 : CreateCheckPoint(int flags)
7057 : {
7058 : bool shutdown;
7059 : CheckPoint checkPoint;
7060 : XLogRecPtr recptr;
7061 : XLogSegNo _logSegNo;
7062 3004 : XLogCtlInsert *Insert = &XLogCtl->Insert;
7063 : uint32 freespace;
7064 : XLogRecPtr PriorRedoPtr;
7065 : XLogRecPtr last_important_lsn;
7066 : VirtualTransactionId *vxids;
7067 : int nvxids;
7068 3004 : int oldXLogAllowed = 0;
7069 :
7070 : /*
7071 : * An end-of-recovery checkpoint is really a shutdown checkpoint, just
7072 : * issued at a different time.
7073 : */
7074 3004 : if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
7075 1200 : shutdown = true;
7076 : else
7077 1804 : shutdown = false;
7078 :
7079 : /* sanity check */
7080 3004 : if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
7081 0 : elog(ERROR, "can't create a checkpoint during recovery");
7082 :
7083 : /*
7084 : * Prepare to accumulate statistics.
7085 : *
7086 : * Note: because it is possible for log_checkpoints to change while a
7087 : * checkpoint proceeds, we always accumulate stats, even if
7088 : * log_checkpoints is currently off.
7089 : */
7090 33044 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7091 3004 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7092 :
7093 : /*
7094 : * Let smgr prepare for checkpoint; this has to happen outside the
7095 : * critical section and before we determine the REDO pointer. Note that
7096 : * smgr must not do anything that'd have to be undone if we decide no
7097 : * checkpoint is needed.
7098 : */
7099 3004 : SyncPreCheckpoint();
7100 :
7101 : /*
7102 : * Use a critical section to force system panic if we have trouble.
7103 : */
7104 3004 : START_CRIT_SECTION();
7105 :
7106 3004 : if (shutdown)
7107 : {
7108 1200 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7109 1200 : ControlFile->state = DB_SHUTDOWNING;
7110 1200 : UpdateControlFile();
7111 1200 : LWLockRelease(ControlFileLock);
7112 : }
7113 :
7114 : /* Begin filling in the checkpoint WAL record */
7115 36048 : MemSet(&checkPoint, 0, sizeof(checkPoint));
7116 3004 : checkPoint.time = (pg_time_t) time(NULL);
7117 :
7118 : /*
7119 : * For Hot Standby, derive the oldestActiveXid before we fix the redo
7120 : * pointer. This allows us to begin accumulating changes to assemble our
7121 : * starting snapshot of locks and transactions.
7122 : */
7123 3004 : if (!shutdown && XLogStandbyInfoActive())
7124 1708 : checkPoint.oldestActiveXid = GetOldestActiveTransactionId(false, true);
7125 : else
7126 1296 : checkPoint.oldestActiveXid = InvalidTransactionId;
7127 :
7128 : /*
7129 : * Get location of last important record before acquiring insert locks (as
7130 : * GetLastImportantRecPtr() also locks WAL locks).
7131 : */
7132 3004 : last_important_lsn = GetLastImportantRecPtr();
7133 :
7134 : /*
7135 : * If this isn't a shutdown or forced checkpoint, and if there has been no
7136 : * WAL activity requiring a checkpoint, skip it. The idea here is to
7137 : * avoid inserting duplicate checkpoints when the system is idle.
7138 : */
7139 3004 : if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
7140 : CHECKPOINT_FORCE)) == 0)
7141 : {
7142 378 : if (last_important_lsn == ControlFile->checkPoint)
7143 : {
7144 0 : END_CRIT_SECTION();
7145 0 : ereport(DEBUG1,
7146 : (errmsg_internal("checkpoint skipped because system is idle")));
7147 0 : return false;
7148 : }
7149 : }
7150 :
7151 : /*
7152 : * An end-of-recovery checkpoint is created before anyone is allowed to
7153 : * write WAL. To allow us to write the checkpoint record, temporarily
7154 : * enable XLogInsertAllowed.
7155 : */
7156 3004 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7157 58 : oldXLogAllowed = LocalSetXLogInsertAllowed();
7158 :
7159 3004 : checkPoint.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7160 3004 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7161 58 : checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7162 : else
7163 2946 : checkPoint.PrevTimeLineID = checkPoint.ThisTimeLineID;
7164 :
7165 : /*
7166 : * We must block concurrent insertions while examining insert state.
7167 : */
7168 3004 : WALInsertLockAcquireExclusive();
7169 :
7170 3004 : checkPoint.fullPageWrites = Insert->fullPageWrites;
7171 3004 : checkPoint.wal_level = wal_level;
7172 :
7173 3004 : if (shutdown)
7174 : {
7175 1200 : XLogRecPtr curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
7176 :
7177 : /*
7178 : * Compute new REDO record ptr = location of next XLOG record.
7179 : *
7180 : * Since this is a shutdown checkpoint, there can't be any concurrent
7181 : * WAL insertion.
7182 : */
7183 1200 : freespace = INSERT_FREESPACE(curInsert);
7184 1200 : if (freespace == 0)
7185 : {
7186 0 : if (XLogSegmentOffset(curInsert, wal_segment_size) == 0)
7187 0 : curInsert += SizeOfXLogLongPHD;
7188 : else
7189 0 : curInsert += SizeOfXLogShortPHD;
7190 : }
7191 1200 : checkPoint.redo = curInsert;
7192 :
7193 : /*
7194 : * Here we update the shared RedoRecPtr for future XLogInsert calls;
7195 : * this must be done while holding all the insertion locks.
7196 : *
7197 : * Note: if we fail to complete the checkpoint, RedoRecPtr will be
7198 : * left pointing past where it really needs to point. This is okay;
7199 : * the only consequence is that XLogInsert might back up whole buffers
7200 : * that it didn't really need to. We can't postpone advancing
7201 : * RedoRecPtr because XLogInserts that happen while we are dumping
7202 : * buffers must assume that their buffer changes are not included in
7203 : * the checkpoint.
7204 : */
7205 1200 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
7206 : }
7207 :
7208 : /*
7209 : * Now we can release the WAL insertion locks, allowing other xacts to
7210 : * proceed while we are flushing disk buffers.
7211 : */
7212 3004 : WALInsertLockRelease();
7213 :
7214 : /*
7215 : * If this is an online checkpoint, we have not yet determined the redo
7216 : * point. We do so now by inserting the special XLOG_CHECKPOINT_REDO
7217 : * record; the LSN at which it starts becomes the new redo pointer. We
7218 : * don't do this for a shutdown checkpoint, because in that case no WAL
7219 : * can be written between the redo point and the insertion of the
7220 : * checkpoint record itself, so the checkpoint record itself serves to
7221 : * mark the redo point.
7222 : */
7223 3004 : if (!shutdown)
7224 : {
7225 : /* Include WAL level in record for WAL summarizer's benefit. */
7226 1804 : XLogBeginInsert();
7227 1804 : XLogRegisterData(&wal_level, sizeof(wal_level));
7228 1804 : (void) XLogInsert(RM_XLOG_ID, XLOG_CHECKPOINT_REDO);
7229 :
7230 : /*
7231 : * XLogInsertRecord will have updated XLogCtl->Insert.RedoRecPtr in
7232 : * shared memory and RedoRecPtr in backend-local memory, but we need
7233 : * to copy that into the record that will be inserted when the
7234 : * checkpoint is complete.
7235 : */
7236 1804 : checkPoint.redo = RedoRecPtr;
7237 : }
7238 :
7239 : /* Update the info_lck-protected copy of RedoRecPtr as well */
7240 3004 : SpinLockAcquire(&XLogCtl->info_lck);
7241 3004 : XLogCtl->RedoRecPtr = checkPoint.redo;
7242 3004 : SpinLockRelease(&XLogCtl->info_lck);
7243 :
7244 : /*
7245 : * If enabled, log checkpoint start. We postpone this until now so as not
7246 : * to log anything if we decided to skip the checkpoint.
7247 : */
7248 3004 : if (log_checkpoints)
7249 2408 : LogCheckpointStart(flags, false);
7250 :
7251 : /* Update the process title */
7252 3004 : update_checkpoint_display(flags, false, false);
7253 :
7254 : TRACE_POSTGRESQL_CHECKPOINT_START(flags);
7255 :
7256 : /*
7257 : * Get the other info we need for the checkpoint record.
7258 : *
7259 : * We don't need to save oldestClogXid in the checkpoint, it only matters
7260 : * for the short period in which clog is being truncated, and if we crash
7261 : * during that we'll redo the clog truncation and fix up oldestClogXid
7262 : * there.
7263 : */
7264 3004 : LWLockAcquire(XidGenLock, LW_SHARED);
7265 3004 : checkPoint.nextXid = TransamVariables->nextXid;
7266 3004 : checkPoint.oldestXid = TransamVariables->oldestXid;
7267 3004 : checkPoint.oldestXidDB = TransamVariables->oldestXidDB;
7268 3004 : LWLockRelease(XidGenLock);
7269 :
7270 3004 : LWLockAcquire(CommitTsLock, LW_SHARED);
7271 3004 : checkPoint.oldestCommitTsXid = TransamVariables->oldestCommitTsXid;
7272 3004 : checkPoint.newestCommitTsXid = TransamVariables->newestCommitTsXid;
7273 3004 : LWLockRelease(CommitTsLock);
7274 :
7275 3004 : LWLockAcquire(OidGenLock, LW_SHARED);
7276 3004 : checkPoint.nextOid = TransamVariables->nextOid;
7277 3004 : if (!shutdown)
7278 1804 : checkPoint.nextOid += TransamVariables->oidCount;
7279 3004 : LWLockRelease(OidGenLock);
7280 :
7281 3004 : MultiXactGetCheckptMulti(shutdown,
7282 : &checkPoint.nextMulti,
7283 : &checkPoint.nextMultiOffset,
7284 : &checkPoint.oldestMulti,
7285 : &checkPoint.oldestMultiDB);
7286 :
7287 : /*
7288 : * Having constructed the checkpoint record, ensure all shmem disk buffers
7289 : * and commit-log buffers are flushed to disk.
7290 : *
7291 : * This I/O could fail for various reasons. If so, we will fail to
7292 : * complete the checkpoint, but there is no reason to force a system
7293 : * panic. Accordingly, exit critical section while doing it.
7294 : */
7295 3004 : END_CRIT_SECTION();
7296 :
7297 : /*
7298 : * In some cases there are groups of actions that must all occur on one
7299 : * side or the other of a checkpoint record. Before flushing the
7300 : * checkpoint record we must explicitly wait for any backend currently
7301 : * performing those groups of actions.
7302 : *
7303 : * One example is end of transaction, so we must wait for any transactions
7304 : * that are currently in commit critical sections. If an xact inserted
7305 : * its commit record into XLOG just before the REDO point, then a crash
7306 : * restart from the REDO point would not replay that record, which means
7307 : * that our flushing had better include the xact's update of pg_xact. So
7308 : * we wait till he's out of his commit critical section before proceeding.
7309 : * See notes in RecordTransactionCommit().
7310 : *
7311 : * Because we've already released the insertion locks, this test is a bit
7312 : * fuzzy: it is possible that we will wait for xacts we didn't really need
7313 : * to wait for. But the delay should be short and it seems better to make
7314 : * checkpoint take a bit longer than to hold off insertions longer than
7315 : * necessary. (In fact, the whole reason we have this issue is that xact.c
7316 : * does commit record XLOG insertion and clog update as two separate steps
7317 : * protected by different locks, but again that seems best on grounds of
7318 : * minimizing lock contention.)
7319 : *
7320 : * A transaction that has not yet set delayChkptFlags when we look cannot
7321 : * be at risk, since it has not inserted its commit record yet; and one
7322 : * that's already cleared it is not at risk either, since it's done fixing
7323 : * clog and we will correctly flush the update below. So we cannot miss
7324 : * any xacts we need to wait for.
7325 : */
7326 3004 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
7327 3004 : if (nvxids > 0)
7328 : {
7329 : do
7330 : {
7331 : /*
7332 : * Keep absorbing fsync requests while we wait. There could even
7333 : * be a deadlock if we don't, if the process that prevents the
7334 : * checkpoint is trying to add a request to the queue.
7335 : */
7336 124 : AbsorbSyncRequests();
7337 :
7338 124 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
7339 124 : pg_usleep(10000L); /* wait for 10 msec */
7340 124 : pgstat_report_wait_end();
7341 124 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7342 : DELAY_CHKPT_START));
7343 : }
7344 3004 : pfree(vxids);
7345 :
7346 3004 : CheckPointGuts(checkPoint.redo, flags);
7347 :
7348 3004 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
7349 3004 : if (nvxids > 0)
7350 : {
7351 : do
7352 : {
7353 0 : AbsorbSyncRequests();
7354 :
7355 0 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_COMPLETE);
7356 0 : pg_usleep(10000L); /* wait for 10 msec */
7357 0 : pgstat_report_wait_end();
7358 0 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7359 : DELAY_CHKPT_COMPLETE));
7360 : }
7361 3004 : pfree(vxids);
7362 :
7363 : /*
7364 : * Take a snapshot of running transactions and write this to WAL. This
7365 : * allows us to reconstruct the state of running transactions during
7366 : * archive recovery, if required. Skip, if this info disabled.
7367 : *
7368 : * If we are shutting down, or Startup process is completing crash
7369 : * recovery we don't need to write running xact data.
7370 : */
7371 3004 : if (!shutdown && XLogStandbyInfoActive())
7372 1708 : LogStandbySnapshot();
7373 :
7374 3004 : START_CRIT_SECTION();
7375 :
7376 : /*
7377 : * Now insert the checkpoint record into XLOG.
7378 : */
7379 3004 : XLogBeginInsert();
7380 3004 : XLogRegisterData(&checkPoint, sizeof(checkPoint));
7381 3004 : recptr = XLogInsert(RM_XLOG_ID,
7382 : shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
7383 : XLOG_CHECKPOINT_ONLINE);
7384 :
7385 3004 : XLogFlush(recptr);
7386 :
7387 : /*
7388 : * We mustn't write any new WAL after a shutdown checkpoint, or it will be
7389 : * overwritten at next startup. No-one should even try, this just allows
7390 : * sanity-checking. In the case of an end-of-recovery checkpoint, we want
7391 : * to just temporarily disable writing until the system has exited
7392 : * recovery.
7393 : */
7394 3004 : if (shutdown)
7395 : {
7396 1200 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7397 58 : LocalXLogInsertAllowed = oldXLogAllowed;
7398 : else
7399 1142 : LocalXLogInsertAllowed = 0; /* never again write WAL */
7400 : }
7401 :
7402 : /*
7403 : * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
7404 : * = end of actual checkpoint record.
7405 : */
7406 3004 : if (shutdown && checkPoint.redo != ProcLastRecPtr)
7407 0 : ereport(PANIC,
7408 : (errmsg("concurrent write-ahead log activity while database system is shutting down")));
7409 :
7410 : /*
7411 : * Remember the prior checkpoint's redo ptr for
7412 : * UpdateCheckPointDistanceEstimate()
7413 : */
7414 3004 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7415 :
7416 : /*
7417 : * Update the control file.
7418 : */
7419 3004 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7420 3004 : if (shutdown)
7421 1200 : ControlFile->state = DB_SHUTDOWNED;
7422 3004 : ControlFile->checkPoint = ProcLastRecPtr;
7423 3004 : ControlFile->checkPointCopy = checkPoint;
7424 : /* crash recovery should always recover to the end of WAL */
7425 3004 : ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
7426 3004 : ControlFile->minRecoveryPointTLI = 0;
7427 :
7428 : /*
7429 : * Persist unloggedLSN value. It's reset on crash recovery, so this goes
7430 : * unused on non-shutdown checkpoints, but seems useful to store it always
7431 : * for debugging purposes.
7432 : */
7433 3004 : ControlFile->unloggedLSN = pg_atomic_read_membarrier_u64(&XLogCtl->unloggedLSN);
7434 :
7435 3004 : UpdateControlFile();
7436 3004 : LWLockRelease(ControlFileLock);
7437 :
7438 : /*
7439 : * We are now done with critical updates; no need for system panic if we
7440 : * have trouble while fooling with old log segments.
7441 : */
7442 3004 : END_CRIT_SECTION();
7443 :
7444 : /*
7445 : * WAL summaries end when the next XLOG_CHECKPOINT_REDO or
7446 : * XLOG_CHECKPOINT_SHUTDOWN record is reached. This is the first point
7447 : * where (a) we're not inside of a critical section and (b) we can be
7448 : * certain that the relevant record has been flushed to disk, which must
7449 : * happen before it can be summarized.
7450 : *
7451 : * If this is a shutdown checkpoint, then this happens reasonably
7452 : * promptly: we've only just inserted and flushed the
7453 : * XLOG_CHECKPOINT_SHUTDOWN record. If this is not a shutdown checkpoint,
7454 : * then this might not be very prompt at all: the XLOG_CHECKPOINT_REDO
7455 : * record was written before we began flushing data to disk, and that
7456 : * could be many minutes ago at this point. However, we don't XLogFlush()
7457 : * after inserting that record, so we're not guaranteed that it's on disk
7458 : * until after the above call that flushes the XLOG_CHECKPOINT_ONLINE
7459 : * record.
7460 : */
7461 3004 : WakeupWalSummarizer();
7462 :
7463 : /*
7464 : * Let smgr do post-checkpoint cleanup (eg, deleting old files).
7465 : */
7466 3004 : SyncPostCheckpoint();
7467 :
7468 : /*
7469 : * Update the average distance between checkpoints if the prior checkpoint
7470 : * exists.
7471 : */
7472 3004 : if (PriorRedoPtr != InvalidXLogRecPtr)
7473 3004 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7474 :
7475 : #ifdef USE_INJECTION_POINTS
7476 3004 : INJECTION_POINT("checkpoint-before-old-wal-removal", NULL);
7477 : #endif
7478 :
7479 : /*
7480 : * Delete old log files, those no longer needed for last checkpoint to
7481 : * prevent the disk holding the xlog from growing full.
7482 : */
7483 3004 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7484 3004 : KeepLogSeg(recptr, &_logSegNo);
7485 3004 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7486 : _logSegNo, InvalidOid,
7487 : InvalidTransactionId))
7488 : {
7489 : /*
7490 : * Some slots have been invalidated; recalculate the old-segment
7491 : * horizon, starting again from RedoRecPtr.
7492 : */
7493 6 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7494 6 : KeepLogSeg(recptr, &_logSegNo);
7495 : }
7496 3004 : _logSegNo--;
7497 3004 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
7498 : checkPoint.ThisTimeLineID);
7499 :
7500 : /*
7501 : * Make more log segments if needed. (Do this after recycling old log
7502 : * segments, since that may supply some of the needed files.)
7503 : */
7504 3004 : if (!shutdown)
7505 1804 : PreallocXlogFiles(recptr, checkPoint.ThisTimeLineID);
7506 :
7507 : /*
7508 : * Truncate pg_subtrans if possible. We can throw away all data before
7509 : * the oldest XMIN of any running transaction. No future transaction will
7510 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7511 : * in subtrans.c). During recovery, though, we mustn't do this because
7512 : * StartupSUBTRANS hasn't been called yet.
7513 : */
7514 3004 : if (!RecoveryInProgress())
7515 2946 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7516 :
7517 : /* Real work is done; log and update stats. */
7518 3004 : LogCheckpointEnd(false);
7519 :
7520 : /* Reset the process title */
7521 3004 : update_checkpoint_display(flags, false, true);
7522 :
7523 : TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
7524 : NBuffers,
7525 : CheckpointStats.ckpt_segs_added,
7526 : CheckpointStats.ckpt_segs_removed,
7527 : CheckpointStats.ckpt_segs_recycled);
7528 :
7529 3004 : return true;
7530 : }
7531 :
7532 : /*
7533 : * Mark the end of recovery in WAL though without running a full checkpoint.
7534 : * We can expect that a restartpoint is likely to be in progress as we
7535 : * do this, though we are unwilling to wait for it to complete.
7536 : *
7537 : * CreateRestartPoint() allows for the case where recovery may end before
7538 : * the restartpoint completes so there is no concern of concurrent behaviour.
7539 : */
7540 : static void
7541 82 : CreateEndOfRecoveryRecord(void)
7542 : {
7543 : xl_end_of_recovery xlrec;
7544 : XLogRecPtr recptr;
7545 :
7546 : /* sanity check */
7547 82 : if (!RecoveryInProgress())
7548 0 : elog(ERROR, "can only be used to end recovery");
7549 :
7550 82 : xlrec.end_time = GetCurrentTimestamp();
7551 82 : xlrec.wal_level = wal_level;
7552 :
7553 82 : WALInsertLockAcquireExclusive();
7554 82 : xlrec.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7555 82 : xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7556 82 : WALInsertLockRelease();
7557 :
7558 82 : START_CRIT_SECTION();
7559 :
7560 82 : XLogBeginInsert();
7561 82 : XLogRegisterData(&xlrec, sizeof(xl_end_of_recovery));
7562 82 : recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY);
7563 :
7564 82 : XLogFlush(recptr);
7565 :
7566 : /*
7567 : * Update the control file so that crash recovery can follow the timeline
7568 : * changes to this point.
7569 : */
7570 82 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7571 82 : ControlFile->minRecoveryPoint = recptr;
7572 82 : ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
7573 82 : UpdateControlFile();
7574 82 : LWLockRelease(ControlFileLock);
7575 :
7576 82 : END_CRIT_SECTION();
7577 82 : }
7578 :
7579 : /*
7580 : * Write an OVERWRITE_CONTRECORD message.
7581 : *
7582 : * When on WAL replay we expect a continuation record at the start of a page
7583 : * that is not there, recovery ends and WAL writing resumes at that point.
7584 : * But it's wrong to resume writing new WAL back at the start of the record
7585 : * that was broken, because downstream consumers of that WAL (physical
7586 : * replicas) are not prepared to "rewind". So the first action after
7587 : * finishing replay of all valid WAL must be to write a record of this type
7588 : * at the point where the contrecord was missing; to support xlogreader
7589 : * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
7590 : * to the page header where the record occurs. xlogreader has an ad-hoc
7591 : * mechanism to report metadata about the broken record, which is what we
7592 : * use here.
7593 : *
7594 : * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
7595 : * skip the record it was reading, and pass back the LSN of the skipped
7596 : * record, so that its caller can verify (on "replay" of that record) that the
7597 : * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
7598 : *
7599 : * 'aborted_lsn' is the beginning position of the record that was incomplete.
7600 : * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
7601 : * beginning of the XLOG page where the record is to be inserted. They must
7602 : * match the current WAL insert position, they're passed here just so that we
7603 : * can verify that.
7604 : */
7605 : static XLogRecPtr
7606 20 : CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn, XLogRecPtr pagePtr,
7607 : TimeLineID newTLI)
7608 : {
7609 : xl_overwrite_contrecord xlrec;
7610 : XLogRecPtr recptr;
7611 : XLogPageHeader pagehdr;
7612 : XLogRecPtr startPos;
7613 :
7614 : /* sanity checks */
7615 20 : if (!RecoveryInProgress())
7616 0 : elog(ERROR, "can only be used at end of recovery");
7617 20 : if (pagePtr % XLOG_BLCKSZ != 0)
7618 0 : elog(ERROR, "invalid position for missing continuation record %X/%08X",
7619 : LSN_FORMAT_ARGS(pagePtr));
7620 :
7621 : /* The current WAL insert position should be right after the page header */
7622 20 : startPos = pagePtr;
7623 20 : if (XLogSegmentOffset(startPos, wal_segment_size) == 0)
7624 2 : startPos += SizeOfXLogLongPHD;
7625 : else
7626 18 : startPos += SizeOfXLogShortPHD;
7627 20 : recptr = GetXLogInsertRecPtr();
7628 20 : if (recptr != startPos)
7629 0 : elog(ERROR, "invalid WAL insert position %X/%08X for OVERWRITE_CONTRECORD",
7630 : LSN_FORMAT_ARGS(recptr));
7631 :
7632 20 : START_CRIT_SECTION();
7633 :
7634 : /*
7635 : * Initialize the XLOG page header (by GetXLogBuffer), and set the
7636 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
7637 : *
7638 : * No other backend is allowed to write WAL yet, so acquiring the WAL
7639 : * insertion lock is just pro forma.
7640 : */
7641 20 : WALInsertLockAcquire();
7642 20 : pagehdr = (XLogPageHeader) GetXLogBuffer(pagePtr, newTLI);
7643 20 : pagehdr->xlp_info |= XLP_FIRST_IS_OVERWRITE_CONTRECORD;
7644 20 : WALInsertLockRelease();
7645 :
7646 : /*
7647 : * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
7648 : * page. We know it becomes the first record, because no other backend is
7649 : * allowed to write WAL yet.
7650 : */
7651 20 : XLogBeginInsert();
7652 20 : xlrec.overwritten_lsn = aborted_lsn;
7653 20 : xlrec.overwrite_time = GetCurrentTimestamp();
7654 20 : XLogRegisterData(&xlrec, sizeof(xl_overwrite_contrecord));
7655 20 : recptr = XLogInsert(RM_XLOG_ID, XLOG_OVERWRITE_CONTRECORD);
7656 :
7657 : /* check that the record was inserted to the right place */
7658 20 : if (ProcLastRecPtr != startPos)
7659 0 : elog(ERROR, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%08X",
7660 : LSN_FORMAT_ARGS(ProcLastRecPtr));
7661 :
7662 20 : XLogFlush(recptr);
7663 :
7664 20 : END_CRIT_SECTION();
7665 :
7666 20 : return recptr;
7667 : }
7668 :
7669 : /*
7670 : * Flush all data in shared memory to disk, and fsync
7671 : *
7672 : * This is the common code shared between regular checkpoints and
7673 : * recovery restartpoints.
7674 : */
7675 : static void
7676 3378 : CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
7677 : {
7678 3378 : CheckPointRelationMap();
7679 3378 : CheckPointReplicationSlots(flags & CHECKPOINT_IS_SHUTDOWN);
7680 3378 : CheckPointSnapBuild();
7681 3378 : CheckPointLogicalRewriteHeap();
7682 3378 : CheckPointReplicationOrigin();
7683 :
7684 : /* Write out all dirty data in SLRUs and the main buffer pool */
7685 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags);
7686 3378 : CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
7687 3378 : CheckPointCLOG();
7688 3378 : CheckPointCommitTs();
7689 3378 : CheckPointSUBTRANS();
7690 3378 : CheckPointMultiXact();
7691 3378 : CheckPointPredicate();
7692 3378 : CheckPointBuffers(flags);
7693 :
7694 : /* Perform all queued up fsyncs */
7695 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
7696 3378 : CheckpointStats.ckpt_sync_t = GetCurrentTimestamp();
7697 3378 : ProcessSyncRequests();
7698 3378 : CheckpointStats.ckpt_sync_end_t = GetCurrentTimestamp();
7699 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
7700 :
7701 : /* We deliberately delay 2PC checkpointing as long as possible */
7702 3378 : CheckPointTwoPhase(checkPointRedo);
7703 3378 : }
7704 :
7705 : /*
7706 : * Save a checkpoint for recovery restart if appropriate
7707 : *
7708 : * This function is called each time a checkpoint record is read from XLOG.
7709 : * It must determine whether the checkpoint represents a safe restartpoint or
7710 : * not. If so, the checkpoint record is stashed in shared memory so that
7711 : * CreateRestartPoint can consult it. (Note that the latter function is
7712 : * executed by the checkpointer, while this one will be executed by the
7713 : * startup process.)
7714 : */
7715 : static void
7716 1410 : RecoveryRestartPoint(const CheckPoint *checkPoint, XLogReaderState *record)
7717 : {
7718 : /*
7719 : * Also refrain from creating a restartpoint if we have seen any
7720 : * references to non-existent pages. Restarting recovery from the
7721 : * restartpoint would not see the references, so we would lose the
7722 : * cross-check that the pages belonged to a relation that was dropped
7723 : * later.
7724 : */
7725 1410 : if (XLogHaveInvalidPages())
7726 : {
7727 0 : elog(DEBUG2,
7728 : "could not record restart point at %X/%08X because there are unresolved references to invalid pages",
7729 : LSN_FORMAT_ARGS(checkPoint->redo));
7730 0 : return;
7731 : }
7732 :
7733 : /*
7734 : * Copy the checkpoint record to shared memory, so that checkpointer can
7735 : * work out the next time it wants to perform a restartpoint.
7736 : */
7737 1410 : SpinLockAcquire(&XLogCtl->info_lck);
7738 1410 : XLogCtl->lastCheckPointRecPtr = record->ReadRecPtr;
7739 1410 : XLogCtl->lastCheckPointEndPtr = record->EndRecPtr;
7740 1410 : XLogCtl->lastCheckPoint = *checkPoint;
7741 1410 : SpinLockRelease(&XLogCtl->info_lck);
7742 : }
7743 :
7744 : /*
7745 : * Establish a restartpoint if possible.
7746 : *
7747 : * This is similar to CreateCheckPoint, but is used during WAL recovery
7748 : * to establish a point from which recovery can roll forward without
7749 : * replaying the entire recovery log.
7750 : *
7751 : * Returns true if a new restartpoint was established. We can only establish
7752 : * a restartpoint if we have replayed a safe checkpoint record since last
7753 : * restartpoint.
7754 : */
7755 : bool
7756 1160 : CreateRestartPoint(int flags)
7757 : {
7758 : XLogRecPtr lastCheckPointRecPtr;
7759 : XLogRecPtr lastCheckPointEndPtr;
7760 : CheckPoint lastCheckPoint;
7761 : XLogRecPtr PriorRedoPtr;
7762 : XLogRecPtr receivePtr;
7763 : XLogRecPtr replayPtr;
7764 : TimeLineID replayTLI;
7765 : XLogRecPtr endptr;
7766 : XLogSegNo _logSegNo;
7767 : TimestampTz xtime;
7768 :
7769 : /* Concurrent checkpoint/restartpoint cannot happen */
7770 : Assert(!IsUnderPostmaster || MyBackendType == B_CHECKPOINTER);
7771 :
7772 : /* Get a local copy of the last safe checkpoint record. */
7773 1160 : SpinLockAcquire(&XLogCtl->info_lck);
7774 1160 : lastCheckPointRecPtr = XLogCtl->lastCheckPointRecPtr;
7775 1160 : lastCheckPointEndPtr = XLogCtl->lastCheckPointEndPtr;
7776 1160 : lastCheckPoint = XLogCtl->lastCheckPoint;
7777 1160 : SpinLockRelease(&XLogCtl->info_lck);
7778 :
7779 : /*
7780 : * Check that we're still in recovery mode. It's ok if we exit recovery
7781 : * mode after this check, the restart point is valid anyway.
7782 : */
7783 1160 : if (!RecoveryInProgress())
7784 : {
7785 0 : ereport(DEBUG2,
7786 : (errmsg_internal("skipping restartpoint, recovery has already ended")));
7787 0 : return false;
7788 : }
7789 :
7790 : /*
7791 : * If the last checkpoint record we've replayed is already our last
7792 : * restartpoint, we can't perform a new restart point. We still update
7793 : * minRecoveryPoint in that case, so that if this is a shutdown restart
7794 : * point, we won't start up earlier than before. That's not strictly
7795 : * necessary, but when hot standby is enabled, it would be rather weird if
7796 : * the database opened up for read-only connections at a point-in-time
7797 : * before the last shutdown. Such time travel is still possible in case of
7798 : * immediate shutdown, though.
7799 : *
7800 : * We don't explicitly advance minRecoveryPoint when we do create a
7801 : * restartpoint. It's assumed that flushing the buffers will do that as a
7802 : * side-effect.
7803 : */
7804 1160 : if (XLogRecPtrIsInvalid(lastCheckPointRecPtr) ||
7805 498 : lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
7806 : {
7807 786 : ereport(DEBUG2,
7808 : errmsg_internal("skipping restartpoint, already performed at %X/%08X",
7809 : LSN_FORMAT_ARGS(lastCheckPoint.redo)));
7810 :
7811 786 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
7812 786 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7813 : {
7814 72 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7815 72 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7816 72 : UpdateControlFile();
7817 72 : LWLockRelease(ControlFileLock);
7818 : }
7819 786 : return false;
7820 : }
7821 :
7822 : /*
7823 : * Update the shared RedoRecPtr so that the startup process can calculate
7824 : * the number of segments replayed since last restartpoint, and request a
7825 : * restartpoint if it exceeds CheckPointSegments.
7826 : *
7827 : * Like in CreateCheckPoint(), hold off insertions to update it, although
7828 : * during recovery this is just pro forma, because no WAL insertions are
7829 : * happening.
7830 : */
7831 374 : WALInsertLockAcquireExclusive();
7832 374 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = lastCheckPoint.redo;
7833 374 : WALInsertLockRelease();
7834 :
7835 : /* Also update the info_lck-protected copy */
7836 374 : SpinLockAcquire(&XLogCtl->info_lck);
7837 374 : XLogCtl->RedoRecPtr = lastCheckPoint.redo;
7838 374 : SpinLockRelease(&XLogCtl->info_lck);
7839 :
7840 : /*
7841 : * Prepare to accumulate statistics.
7842 : *
7843 : * Note: because it is possible for log_checkpoints to change while a
7844 : * checkpoint proceeds, we always accumulate stats, even if
7845 : * log_checkpoints is currently off.
7846 : */
7847 4114 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7848 374 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7849 :
7850 374 : if (log_checkpoints)
7851 374 : LogCheckpointStart(flags, true);
7852 :
7853 : /* Update the process title */
7854 374 : update_checkpoint_display(flags, true, false);
7855 :
7856 374 : CheckPointGuts(lastCheckPoint.redo, flags);
7857 :
7858 : /*
7859 : * This location needs to be after CheckPointGuts() to ensure that some
7860 : * work has already happened during this checkpoint.
7861 : */
7862 374 : INJECTION_POINT("create-restart-point", NULL);
7863 :
7864 : /*
7865 : * Remember the prior checkpoint's redo ptr for
7866 : * UpdateCheckPointDistanceEstimate()
7867 : */
7868 374 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7869 :
7870 : /*
7871 : * Update pg_control, using current time. Check that it still shows an
7872 : * older checkpoint, else do nothing; this is a quick hack to make sure
7873 : * nothing really bad happens if somehow we get here after the
7874 : * end-of-recovery checkpoint.
7875 : */
7876 374 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7877 374 : if (ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
7878 : {
7879 : /*
7880 : * Update the checkpoint information. We do this even if the cluster
7881 : * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
7882 : * segments recycled below.
7883 : */
7884 374 : ControlFile->checkPoint = lastCheckPointRecPtr;
7885 374 : ControlFile->checkPointCopy = lastCheckPoint;
7886 :
7887 : /*
7888 : * Ensure minRecoveryPoint is past the checkpoint record and update it
7889 : * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
7890 : * this will have happened already while writing out dirty buffers,
7891 : * but not necessarily - e.g. because no buffers were dirtied. We do
7892 : * this because a backup performed in recovery uses minRecoveryPoint
7893 : * to determine which WAL files must be included in the backup, and
7894 : * the file (or files) containing the checkpoint record must be
7895 : * included, at a minimum. Note that for an ordinary restart of
7896 : * recovery there's no value in having the minimum recovery point any
7897 : * earlier than this anyway, because redo will begin just after the
7898 : * checkpoint record.
7899 : */
7900 374 : if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
7901 : {
7902 374 : if (ControlFile->minRecoveryPoint < lastCheckPointEndPtr)
7903 : {
7904 40 : ControlFile->minRecoveryPoint = lastCheckPointEndPtr;
7905 40 : ControlFile->minRecoveryPointTLI = lastCheckPoint.ThisTimeLineID;
7906 :
7907 : /* update local copy */
7908 40 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
7909 40 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
7910 : }
7911 374 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7912 38 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7913 : }
7914 374 : UpdateControlFile();
7915 : }
7916 374 : LWLockRelease(ControlFileLock);
7917 :
7918 : /*
7919 : * Update the average distance between checkpoints/restartpoints if the
7920 : * prior checkpoint exists.
7921 : */
7922 374 : if (PriorRedoPtr != InvalidXLogRecPtr)
7923 374 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7924 :
7925 : /*
7926 : * Delete old log files, those no longer needed for last restartpoint to
7927 : * prevent the disk holding the xlog from growing full.
7928 : */
7929 374 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7930 :
7931 : /*
7932 : * Retreat _logSegNo using the current end of xlog replayed or received,
7933 : * whichever is later.
7934 : */
7935 374 : receivePtr = GetWalRcvFlushRecPtr(NULL, NULL);
7936 374 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
7937 374 : endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
7938 374 : KeepLogSeg(endptr, &_logSegNo);
7939 374 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7940 : _logSegNo, InvalidOid,
7941 : InvalidTransactionId))
7942 : {
7943 : /*
7944 : * Some slots have been invalidated; recalculate the old-segment
7945 : * horizon, starting again from RedoRecPtr.
7946 : */
7947 2 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7948 2 : KeepLogSeg(endptr, &_logSegNo);
7949 : }
7950 374 : _logSegNo--;
7951 :
7952 : /*
7953 : * Try to recycle segments on a useful timeline. If we've been promoted
7954 : * since the beginning of this restartpoint, use the new timeline chosen
7955 : * at end of recovery. If we're still in recovery, use the timeline we're
7956 : * currently replaying.
7957 : *
7958 : * There is no guarantee that the WAL segments will be useful on the
7959 : * current timeline; if recovery proceeds to a new timeline right after
7960 : * this, the pre-allocated WAL segments on this timeline will not be used,
7961 : * and will go wasted until recycled on the next restartpoint. We'll live
7962 : * with that.
7963 : */
7964 374 : if (!RecoveryInProgress())
7965 0 : replayTLI = XLogCtl->InsertTimeLineID;
7966 :
7967 374 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, endptr, replayTLI);
7968 :
7969 : /*
7970 : * Make more log segments if needed. (Do this after recycling old log
7971 : * segments, since that may supply some of the needed files.)
7972 : */
7973 374 : PreallocXlogFiles(endptr, replayTLI);
7974 :
7975 : /*
7976 : * Truncate pg_subtrans if possible. We can throw away all data before
7977 : * the oldest XMIN of any running transaction. No future transaction will
7978 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7979 : * in subtrans.c). When hot standby is disabled, though, we mustn't do
7980 : * this because StartupSUBTRANS hasn't been called yet.
7981 : */
7982 374 : if (EnableHotStandby)
7983 374 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7984 :
7985 : /* Real work is done; log and update stats. */
7986 374 : LogCheckpointEnd(true);
7987 :
7988 : /* Reset the process title */
7989 374 : update_checkpoint_display(flags, true, true);
7990 :
7991 374 : xtime = GetLatestXTime();
7992 374 : ereport((log_checkpoints ? LOG : DEBUG2),
7993 : errmsg("recovery restart point at %X/%08X",
7994 : LSN_FORMAT_ARGS(lastCheckPoint.redo)),
7995 : xtime ? errdetail("Last completed transaction was at log time %s.",
7996 : timestamptz_to_str(xtime)) : 0);
7997 :
7998 : /*
7999 : * Finally, execute archive_cleanup_command, if any.
8000 : */
8001 374 : if (archiveCleanupCommand && strcmp(archiveCleanupCommand, "") != 0)
8002 0 : ExecuteRecoveryCommand(archiveCleanupCommand,
8003 : "archive_cleanup_command",
8004 : false,
8005 : WAIT_EVENT_ARCHIVE_CLEANUP_COMMAND);
8006 :
8007 374 : return true;
8008 : }
8009 :
8010 : /*
8011 : * Report availability of WAL for the given target LSN
8012 : * (typically a slot's restart_lsn)
8013 : *
8014 : * Returns one of the following enum values:
8015 : *
8016 : * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
8017 : * max_wal_size.
8018 : *
8019 : * * WALAVAIL_EXTENDED means it is still available by preserving extra
8020 : * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
8021 : * than max_wal_size, this state is not returned.
8022 : *
8023 : * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
8024 : * remove reserved segments. The walsender using this slot may return to the
8025 : * above.
8026 : *
8027 : * * WALAVAIL_REMOVED means it has been removed. A replication stream on
8028 : * a slot with this LSN cannot continue. (Any associated walsender
8029 : * processes should have been terminated already.)
8030 : *
8031 : * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
8032 : */
8033 : WALAvailability
8034 992 : GetWALAvailability(XLogRecPtr targetLSN)
8035 : {
8036 : XLogRecPtr currpos; /* current write LSN */
8037 : XLogSegNo currSeg; /* segid of currpos */
8038 : XLogSegNo targetSeg; /* segid of targetLSN */
8039 : XLogSegNo oldestSeg; /* actual oldest segid */
8040 : XLogSegNo oldestSegMaxWalSize; /* oldest segid kept by max_wal_size */
8041 : XLogSegNo oldestSlotSeg; /* oldest segid kept by slot */
8042 : uint64 keepSegs;
8043 :
8044 : /*
8045 : * slot does not reserve WAL. Either deactivated, or has never been active
8046 : */
8047 992 : if (XLogRecPtrIsInvalid(targetLSN))
8048 44 : return WALAVAIL_INVALID_LSN;
8049 :
8050 : /*
8051 : * Calculate the oldest segment currently reserved by all slots,
8052 : * considering wal_keep_size and max_slot_wal_keep_size. Initialize
8053 : * oldestSlotSeg to the current segment.
8054 : */
8055 948 : currpos = GetXLogWriteRecPtr();
8056 948 : XLByteToSeg(currpos, oldestSlotSeg, wal_segment_size);
8057 948 : KeepLogSeg(currpos, &oldestSlotSeg);
8058 :
8059 : /*
8060 : * Find the oldest extant segment file. We get 1 until checkpoint removes
8061 : * the first WAL segment file since startup, which causes the status being
8062 : * wrong under certain abnormal conditions but that doesn't actually harm.
8063 : */
8064 948 : oldestSeg = XLogGetLastRemovedSegno() + 1;
8065 :
8066 : /* calculate oldest segment by max_wal_size */
8067 948 : XLByteToSeg(currpos, currSeg, wal_segment_size);
8068 948 : keepSegs = ConvertToXSegs(max_wal_size_mb, wal_segment_size) + 1;
8069 :
8070 948 : if (currSeg > keepSegs)
8071 16 : oldestSegMaxWalSize = currSeg - keepSegs;
8072 : else
8073 932 : oldestSegMaxWalSize = 1;
8074 :
8075 : /* the segment we care about */
8076 948 : XLByteToSeg(targetLSN, targetSeg, wal_segment_size);
8077 :
8078 : /*
8079 : * No point in returning reserved or extended status values if the
8080 : * targetSeg is known to be lost.
8081 : */
8082 948 : if (targetSeg >= oldestSlotSeg)
8083 : {
8084 : /* show "reserved" when targetSeg is within max_wal_size */
8085 946 : if (targetSeg >= oldestSegMaxWalSize)
8086 942 : return WALAVAIL_RESERVED;
8087 :
8088 : /* being retained by slots exceeding max_wal_size */
8089 4 : return WALAVAIL_EXTENDED;
8090 : }
8091 :
8092 : /* WAL segments are no longer retained but haven't been removed yet */
8093 2 : if (targetSeg >= oldestSeg)
8094 2 : return WALAVAIL_UNRESERVED;
8095 :
8096 : /* Definitely lost */
8097 0 : return WALAVAIL_REMOVED;
8098 : }
8099 :
8100 :
8101 : /*
8102 : * Retreat *logSegNo to the last segment that we need to retain because of
8103 : * either wal_keep_size or replication slots.
8104 : *
8105 : * This is calculated by subtracting wal_keep_size from the given xlog
8106 : * location, recptr and by making sure that that result is below the
8107 : * requirement of replication slots. For the latter criterion we do consider
8108 : * the effects of max_slot_wal_keep_size: reserve at most that much space back
8109 : * from recptr.
8110 : *
8111 : * Note about replication slots: if this function calculates a value
8112 : * that's further ahead than what slots need reserved, then affected
8113 : * slots need to be invalidated and this function invoked again.
8114 : * XXX it might be a good idea to rewrite this function so that
8115 : * invalidation is optionally done here, instead.
8116 : */
8117 : static void
8118 4334 : KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
8119 : {
8120 : XLogSegNo currSegNo;
8121 : XLogSegNo segno;
8122 : XLogRecPtr keep;
8123 :
8124 4334 : XLByteToSeg(recptr, currSegNo, wal_segment_size);
8125 4334 : segno = currSegNo;
8126 :
8127 : /* Calculate how many segments are kept by slots. */
8128 4334 : keep = XLogGetReplicationSlotMinimumLSN();
8129 4334 : if (keep != InvalidXLogRecPtr && keep < recptr)
8130 : {
8131 1222 : XLByteToSeg(keep, segno, wal_segment_size);
8132 :
8133 : /*
8134 : * Account for max_slot_wal_keep_size to avoid keeping more than
8135 : * configured. However, don't do that during a binary upgrade: if
8136 : * slots were to be invalidated because of this, it would not be
8137 : * possible to preserve logical ones during the upgrade.
8138 : */
8139 1222 : if (max_slot_wal_keep_size_mb >= 0 && !IsBinaryUpgrade)
8140 : {
8141 : uint64 slot_keep_segs;
8142 :
8143 42 : slot_keep_segs =
8144 42 : ConvertToXSegs(max_slot_wal_keep_size_mb, wal_segment_size);
8145 :
8146 42 : if (currSegNo - segno > slot_keep_segs)
8147 10 : segno = currSegNo - slot_keep_segs;
8148 : }
8149 : }
8150 :
8151 : /*
8152 : * If WAL summarization is in use, don't remove WAL that has yet to be
8153 : * summarized.
8154 : */
8155 4334 : keep = GetOldestUnsummarizedLSN(NULL, NULL);
8156 4334 : if (keep != InvalidXLogRecPtr)
8157 : {
8158 : XLogSegNo unsummarized_segno;
8159 :
8160 4 : XLByteToSeg(keep, unsummarized_segno, wal_segment_size);
8161 4 : if (unsummarized_segno < segno)
8162 4 : segno = unsummarized_segno;
8163 : }
8164 :
8165 : /* but, keep at least wal_keep_size if that's set */
8166 4334 : if (wal_keep_size_mb > 0)
8167 : {
8168 : uint64 keep_segs;
8169 :
8170 138 : keep_segs = ConvertToXSegs(wal_keep_size_mb, wal_segment_size);
8171 138 : if (currSegNo - segno < keep_segs)
8172 : {
8173 : /* avoid underflow, don't go below 1 */
8174 138 : if (currSegNo <= keep_segs)
8175 130 : segno = 1;
8176 : else
8177 8 : segno = currSegNo - keep_segs;
8178 : }
8179 : }
8180 :
8181 : /* don't delete WAL segments newer than the calculated segment */
8182 4334 : if (segno < *logSegNo)
8183 750 : *logSegNo = segno;
8184 4334 : }
8185 :
8186 : /*
8187 : * Write a NEXTOID log record
8188 : */
8189 : void
8190 1202 : XLogPutNextOid(Oid nextOid)
8191 : {
8192 1202 : XLogBeginInsert();
8193 1202 : XLogRegisterData(&nextOid, sizeof(Oid));
8194 1202 : (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID);
8195 :
8196 : /*
8197 : * We need not flush the NEXTOID record immediately, because any of the
8198 : * just-allocated OIDs could only reach disk as part of a tuple insert or
8199 : * update that would have its own XLOG record that must follow the NEXTOID
8200 : * record. Therefore, the standard buffer LSN interlock applied to those
8201 : * records will ensure no such OID reaches disk before the NEXTOID record
8202 : * does.
8203 : *
8204 : * Note, however, that the above statement only covers state "within" the
8205 : * database. When we use a generated OID as a file or directory name, we
8206 : * are in a sense violating the basic WAL rule, because that filesystem
8207 : * change may reach disk before the NEXTOID WAL record does. The impact
8208 : * of this is that if a database crash occurs immediately afterward, we
8209 : * might after restart re-generate the same OID and find that it conflicts
8210 : * with the leftover file or directory. But since for safety's sake we
8211 : * always loop until finding a nonconflicting filename, this poses no real
8212 : * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8213 : */
8214 1202 : }
8215 :
8216 : /*
8217 : * Write an XLOG SWITCH record.
8218 : *
8219 : * Here we just blindly issue an XLogInsert request for the record.
8220 : * All the magic happens inside XLogInsert.
8221 : *
8222 : * The return value is either the end+1 address of the switch record,
8223 : * or the end+1 address of the prior segment if we did not need to
8224 : * write a switch record because we are already at segment start.
8225 : */
8226 : XLogRecPtr
8227 1558 : RequestXLogSwitch(bool mark_unimportant)
8228 : {
8229 : XLogRecPtr RecPtr;
8230 :
8231 : /* XLOG SWITCH has no data */
8232 1558 : XLogBeginInsert();
8233 :
8234 1558 : if (mark_unimportant)
8235 0 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8236 1558 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH);
8237 :
8238 1558 : return RecPtr;
8239 : }
8240 :
8241 : /*
8242 : * Write a RESTORE POINT record
8243 : */
8244 : XLogRecPtr
8245 6 : XLogRestorePoint(const char *rpName)
8246 : {
8247 : XLogRecPtr RecPtr;
8248 : xl_restore_point xlrec;
8249 :
8250 6 : xlrec.rp_time = GetCurrentTimestamp();
8251 6 : strlcpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8252 :
8253 6 : XLogBeginInsert();
8254 6 : XLogRegisterData(&xlrec, sizeof(xl_restore_point));
8255 :
8256 6 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT);
8257 :
8258 6 : ereport(LOG,
8259 : errmsg("restore point \"%s\" created at %X/%08X",
8260 : rpName, LSN_FORMAT_ARGS(RecPtr)));
8261 :
8262 6 : return RecPtr;
8263 : }
8264 :
8265 : /*
8266 : * Check if any of the GUC parameters that are critical for hot standby
8267 : * have changed, and update the value in pg_control file if necessary.
8268 : */
8269 : static void
8270 1744 : XLogReportParameters(void)
8271 : {
8272 1744 : if (wal_level != ControlFile->wal_level ||
8273 1266 : wal_log_hints != ControlFile->wal_log_hints ||
8274 1100 : MaxConnections != ControlFile->MaxConnections ||
8275 1098 : max_worker_processes != ControlFile->max_worker_processes ||
8276 1096 : max_wal_senders != ControlFile->max_wal_senders ||
8277 1054 : max_prepared_xacts != ControlFile->max_prepared_xacts ||
8278 872 : max_locks_per_xact != ControlFile->max_locks_per_xact ||
8279 872 : track_commit_timestamp != ControlFile->track_commit_timestamp)
8280 : {
8281 : /*
8282 : * The change in number of backend slots doesn't need to be WAL-logged
8283 : * if archiving is not enabled, as you can't start archive recovery
8284 : * with wal_level=minimal anyway. We don't really care about the
8285 : * values in pg_control either if wal_level=minimal, but seems better
8286 : * to keep them up-to-date to avoid confusion.
8287 : */
8288 894 : if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8289 : {
8290 : xl_parameter_change xlrec;
8291 : XLogRecPtr recptr;
8292 :
8293 854 : xlrec.MaxConnections = MaxConnections;
8294 854 : xlrec.max_worker_processes = max_worker_processes;
8295 854 : xlrec.max_wal_senders = max_wal_senders;
8296 854 : xlrec.max_prepared_xacts = max_prepared_xacts;
8297 854 : xlrec.max_locks_per_xact = max_locks_per_xact;
8298 854 : xlrec.wal_level = wal_level;
8299 854 : xlrec.wal_log_hints = wal_log_hints;
8300 854 : xlrec.track_commit_timestamp = track_commit_timestamp;
8301 :
8302 854 : XLogBeginInsert();
8303 854 : XLogRegisterData(&xlrec, sizeof(xlrec));
8304 :
8305 854 : recptr = XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE);
8306 854 : XLogFlush(recptr);
8307 : }
8308 :
8309 894 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8310 :
8311 894 : ControlFile->MaxConnections = MaxConnections;
8312 894 : ControlFile->max_worker_processes = max_worker_processes;
8313 894 : ControlFile->max_wal_senders = max_wal_senders;
8314 894 : ControlFile->max_prepared_xacts = max_prepared_xacts;
8315 894 : ControlFile->max_locks_per_xact = max_locks_per_xact;
8316 894 : ControlFile->wal_level = wal_level;
8317 894 : ControlFile->wal_log_hints = wal_log_hints;
8318 894 : ControlFile->track_commit_timestamp = track_commit_timestamp;
8319 894 : UpdateControlFile();
8320 :
8321 894 : LWLockRelease(ControlFileLock);
8322 : }
8323 1744 : }
8324 :
8325 : /*
8326 : * Update full_page_writes in shared memory, and write an
8327 : * XLOG_FPW_CHANGE record if necessary.
8328 : *
8329 : * Note: this function assumes there is no other process running
8330 : * concurrently that could update it.
8331 : */
8332 : void
8333 2900 : UpdateFullPageWrites(void)
8334 : {
8335 2900 : XLogCtlInsert *Insert = &XLogCtl->Insert;
8336 : bool recoveryInProgress;
8337 :
8338 : /*
8339 : * Do nothing if full_page_writes has not been changed.
8340 : *
8341 : * It's safe to check the shared full_page_writes without the lock,
8342 : * because we assume that there is no concurrently running process which
8343 : * can update it.
8344 : */
8345 2900 : if (fullPageWrites == Insert->fullPageWrites)
8346 2160 : return;
8347 :
8348 : /*
8349 : * Perform this outside critical section so that the WAL insert
8350 : * initialization done by RecoveryInProgress() doesn't trigger an
8351 : * assertion failure.
8352 : */
8353 740 : recoveryInProgress = RecoveryInProgress();
8354 :
8355 740 : START_CRIT_SECTION();
8356 :
8357 : /*
8358 : * It's always safe to take full page images, even when not strictly
8359 : * required, but not the other round. So if we're setting full_page_writes
8360 : * to true, first set it true and then write the WAL record. If we're
8361 : * setting it to false, first write the WAL record and then set the global
8362 : * flag.
8363 : */
8364 740 : if (fullPageWrites)
8365 : {
8366 720 : WALInsertLockAcquireExclusive();
8367 720 : Insert->fullPageWrites = true;
8368 720 : WALInsertLockRelease();
8369 : }
8370 :
8371 : /*
8372 : * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8373 : * full_page_writes during archive recovery, if required.
8374 : */
8375 740 : if (XLogStandbyInfoActive() && !recoveryInProgress)
8376 : {
8377 0 : XLogBeginInsert();
8378 0 : XLogRegisterData(&fullPageWrites, sizeof(bool));
8379 :
8380 0 : XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
8381 : }
8382 :
8383 740 : if (!fullPageWrites)
8384 : {
8385 20 : WALInsertLockAcquireExclusive();
8386 20 : Insert->fullPageWrites = false;
8387 20 : WALInsertLockRelease();
8388 : }
8389 740 : END_CRIT_SECTION();
8390 : }
8391 :
8392 : /*
8393 : * XLOG resource manager's routines
8394 : *
8395 : * Definitions of info values are in include/catalog/pg_control.h, though
8396 : * not all record types are related to control file updates.
8397 : *
8398 : * NOTE: Some XLOG record types that are directly related to WAL recovery
8399 : * are handled in xlogrecovery_redo().
8400 : */
8401 : void
8402 86792 : xlog_redo(XLogReaderState *record)
8403 : {
8404 86792 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8405 86792 : XLogRecPtr lsn = record->EndRecPtr;
8406 :
8407 : /*
8408 : * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
8409 : * XLOG_FPI_FOR_HINT records.
8410 : */
8411 : Assert(info == XLOG_FPI || info == XLOG_FPI_FOR_HINT ||
8412 : !XLogRecHasAnyBlockRefs(record));
8413 :
8414 86792 : if (info == XLOG_NEXTOID)
8415 : {
8416 : Oid nextOid;
8417 :
8418 : /*
8419 : * We used to try to take the maximum of TransamVariables->nextOid and
8420 : * the recorded nextOid, but that fails if the OID counter wraps
8421 : * around. Since no OID allocation should be happening during replay
8422 : * anyway, better to just believe the record exactly. We still take
8423 : * OidGenLock while setting the variable, just in case.
8424 : */
8425 184 : memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
8426 184 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8427 184 : TransamVariables->nextOid = nextOid;
8428 184 : TransamVariables->oidCount = 0;
8429 184 : LWLockRelease(OidGenLock);
8430 : }
8431 86608 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
8432 : {
8433 : CheckPoint checkPoint;
8434 : TimeLineID replayTLI;
8435 :
8436 66 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8437 : /* In a SHUTDOWN checkpoint, believe the counters exactly */
8438 66 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8439 66 : TransamVariables->nextXid = checkPoint.nextXid;
8440 66 : LWLockRelease(XidGenLock);
8441 66 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8442 66 : TransamVariables->nextOid = checkPoint.nextOid;
8443 66 : TransamVariables->oidCount = 0;
8444 66 : LWLockRelease(OidGenLock);
8445 66 : MultiXactSetNextMXact(checkPoint.nextMulti,
8446 : checkPoint.nextMultiOffset);
8447 :
8448 66 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8449 : checkPoint.oldestMultiDB);
8450 :
8451 : /*
8452 : * No need to set oldestClogXid here as well; it'll be set when we
8453 : * redo an xl_clog_truncate if it changed since initialization.
8454 : */
8455 66 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
8456 :
8457 : /*
8458 : * If we see a shutdown checkpoint while waiting for an end-of-backup
8459 : * record, the backup was canceled and the end-of-backup record will
8460 : * never arrive.
8461 : */
8462 66 : if (ArchiveRecoveryRequested &&
8463 64 : !XLogRecPtrIsInvalid(ControlFile->backupStartPoint) &&
8464 0 : XLogRecPtrIsInvalid(ControlFile->backupEndPoint))
8465 0 : ereport(PANIC,
8466 : (errmsg("online backup was canceled, recovery cannot continue")));
8467 :
8468 : /*
8469 : * If we see a shutdown checkpoint, we know that nothing was running
8470 : * on the primary at this point. So fake-up an empty running-xacts
8471 : * record and use that here and now. Recover additional standby state
8472 : * for prepared transactions.
8473 : */
8474 66 : if (standbyState >= STANDBY_INITIALIZED)
8475 : {
8476 : TransactionId *xids;
8477 : int nxids;
8478 : TransactionId oldestActiveXID;
8479 : TransactionId latestCompletedXid;
8480 : RunningTransactionsData running;
8481 :
8482 60 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
8483 :
8484 : /* Update pg_subtrans entries for any prepared transactions */
8485 60 : StandbyRecoverPreparedTransactions();
8486 :
8487 : /*
8488 : * Construct a RunningTransactions snapshot representing a shut
8489 : * down server, with only prepared transactions still alive. We're
8490 : * never overflowed at this point because all subxids are listed
8491 : * with their parent prepared transactions.
8492 : */
8493 60 : running.xcnt = nxids;
8494 60 : running.subxcnt = 0;
8495 60 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
8496 60 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
8497 60 : running.oldestRunningXid = oldestActiveXID;
8498 60 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
8499 60 : TransactionIdRetreat(latestCompletedXid);
8500 : Assert(TransactionIdIsNormal(latestCompletedXid));
8501 60 : running.latestCompletedXid = latestCompletedXid;
8502 60 : running.xids = xids;
8503 :
8504 60 : ProcArrayApplyRecoveryInfo(&running);
8505 : }
8506 :
8507 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8508 66 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8509 66 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8510 66 : LWLockRelease(ControlFileLock);
8511 :
8512 : /*
8513 : * We should've already switched to the new TLI before replaying this
8514 : * record.
8515 : */
8516 66 : (void) GetCurrentReplayRecPtr(&replayTLI);
8517 66 : if (checkPoint.ThisTimeLineID != replayTLI)
8518 0 : ereport(PANIC,
8519 : (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
8520 : checkPoint.ThisTimeLineID, replayTLI)));
8521 :
8522 66 : RecoveryRestartPoint(&checkPoint, record);
8523 : }
8524 86542 : else if (info == XLOG_CHECKPOINT_ONLINE)
8525 : {
8526 : CheckPoint checkPoint;
8527 : TimeLineID replayTLI;
8528 :
8529 1344 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8530 : /* In an ONLINE checkpoint, treat the XID counter as a minimum */
8531 1344 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8532 1344 : if (FullTransactionIdPrecedes(TransamVariables->nextXid,
8533 : checkPoint.nextXid))
8534 0 : TransamVariables->nextXid = checkPoint.nextXid;
8535 1344 : LWLockRelease(XidGenLock);
8536 :
8537 : /*
8538 : * We ignore the nextOid counter in an ONLINE checkpoint, preferring
8539 : * to track OID assignment through XLOG_NEXTOID records. The nextOid
8540 : * counter is from the start of the checkpoint and might well be stale
8541 : * compared to later XLOG_NEXTOID records. We could try to take the
8542 : * maximum of the nextOid counter and our latest value, but since
8543 : * there's no particular guarantee about the speed with which the OID
8544 : * counter wraps around, that's a risky thing to do. In any case,
8545 : * users of the nextOid counter are required to avoid assignment of
8546 : * duplicates, so that a somewhat out-of-date value should be safe.
8547 : */
8548 :
8549 : /* Handle multixact */
8550 1344 : MultiXactAdvanceNextMXact(checkPoint.nextMulti,
8551 : checkPoint.nextMultiOffset);
8552 :
8553 : /*
8554 : * NB: This may perform multixact truncation when replaying WAL
8555 : * generated by an older primary.
8556 : */
8557 1344 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8558 : checkPoint.oldestMultiDB);
8559 1344 : if (TransactionIdPrecedes(TransamVariables->oldestXid,
8560 : checkPoint.oldestXid))
8561 0 : SetTransactionIdLimit(checkPoint.oldestXid,
8562 : checkPoint.oldestXidDB);
8563 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8564 1344 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8565 1344 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8566 1344 : LWLockRelease(ControlFileLock);
8567 :
8568 : /* TLI should not change in an on-line checkpoint */
8569 1344 : (void) GetCurrentReplayRecPtr(&replayTLI);
8570 1344 : if (checkPoint.ThisTimeLineID != replayTLI)
8571 0 : ereport(PANIC,
8572 : (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
8573 : checkPoint.ThisTimeLineID, replayTLI)));
8574 :
8575 1344 : RecoveryRestartPoint(&checkPoint, record);
8576 : }
8577 85198 : else if (info == XLOG_OVERWRITE_CONTRECORD)
8578 : {
8579 : /* nothing to do here, handled in xlogrecovery_redo() */
8580 : }
8581 85196 : else if (info == XLOG_END_OF_RECOVERY)
8582 : {
8583 : xl_end_of_recovery xlrec;
8584 : TimeLineID replayTLI;
8585 :
8586 20 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
8587 :
8588 : /*
8589 : * For Hot Standby, we could treat this like a Shutdown Checkpoint,
8590 : * but this case is rarer and harder to test, so the benefit doesn't
8591 : * outweigh the potential extra cost of maintenance.
8592 : */
8593 :
8594 : /*
8595 : * We should've already switched to the new TLI before replaying this
8596 : * record.
8597 : */
8598 20 : (void) GetCurrentReplayRecPtr(&replayTLI);
8599 20 : if (xlrec.ThisTimeLineID != replayTLI)
8600 0 : ereport(PANIC,
8601 : (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
8602 : xlrec.ThisTimeLineID, replayTLI)));
8603 : }
8604 85176 : else if (info == XLOG_NOOP)
8605 : {
8606 : /* nothing to do here */
8607 : }
8608 85176 : else if (info == XLOG_SWITCH)
8609 : {
8610 : /* nothing to do here */
8611 : }
8612 84294 : else if (info == XLOG_RESTORE_POINT)
8613 : {
8614 : /* nothing to do here, handled in xlogrecovery.c */
8615 : }
8616 84284 : else if (info == XLOG_FPI || info == XLOG_FPI_FOR_HINT)
8617 : {
8618 : /*
8619 : * XLOG_FPI records contain nothing else but one or more block
8620 : * references. Every block reference must include a full-page image
8621 : * even if full_page_writes was disabled when the record was generated
8622 : * - otherwise there would be no point in this record.
8623 : *
8624 : * XLOG_FPI_FOR_HINT records are generated when a page needs to be
8625 : * WAL-logged because of a hint bit update. They are only generated
8626 : * when checksums and/or wal_log_hints are enabled. They may include
8627 : * no full-page images if full_page_writes was disabled when they were
8628 : * generated. In this case there is nothing to do here.
8629 : *
8630 : * No recovery conflicts are generated by these generic records - if a
8631 : * resource manager needs to generate conflicts, it has to define a
8632 : * separate WAL record type and redo routine.
8633 : */
8634 174828 : for (uint8 block_id = 0; block_id <= XLogRecMaxBlockId(record); block_id++)
8635 : {
8636 : Buffer buffer;
8637 :
8638 92124 : if (!XLogRecHasBlockImage(record, block_id))
8639 : {
8640 132 : if (info == XLOG_FPI)
8641 0 : elog(ERROR, "XLOG_FPI record did not contain a full-page image");
8642 132 : continue;
8643 : }
8644 :
8645 91992 : if (XLogReadBufferForRedo(record, block_id, &buffer) != BLK_RESTORED)
8646 0 : elog(ERROR, "unexpected XLogReadBufferForRedo result when restoring backup block");
8647 91992 : UnlockReleaseBuffer(buffer);
8648 : }
8649 : }
8650 1580 : else if (info == XLOG_BACKUP_END)
8651 : {
8652 : /* nothing to do here, handled in xlogrecovery_redo() */
8653 : }
8654 1410 : else if (info == XLOG_PARAMETER_CHANGE)
8655 : {
8656 : xl_parameter_change xlrec;
8657 :
8658 : /* Update our copy of the parameters in pg_control */
8659 64 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
8660 :
8661 : /*
8662 : * Invalidate logical slots if we are in hot standby and the primary
8663 : * does not have a WAL level sufficient for logical decoding. No need
8664 : * to search for potentially conflicting logically slots if standby is
8665 : * running with wal_level lower than logical, because in that case, we
8666 : * would have either disallowed creation of logical slots or
8667 : * invalidated existing ones.
8668 : */
8669 64 : if (InRecovery && InHotStandby &&
8670 34 : xlrec.wal_level < WAL_LEVEL_LOGICAL &&
8671 12 : wal_level >= WAL_LEVEL_LOGICAL)
8672 8 : InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_LEVEL,
8673 : 0, InvalidOid,
8674 : InvalidTransactionId);
8675 :
8676 64 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8677 64 : ControlFile->MaxConnections = xlrec.MaxConnections;
8678 64 : ControlFile->max_worker_processes = xlrec.max_worker_processes;
8679 64 : ControlFile->max_wal_senders = xlrec.max_wal_senders;
8680 64 : ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
8681 64 : ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
8682 64 : ControlFile->wal_level = xlrec.wal_level;
8683 64 : ControlFile->wal_log_hints = xlrec.wal_log_hints;
8684 :
8685 : /*
8686 : * Update minRecoveryPoint to ensure that if recovery is aborted, we
8687 : * recover back up to this point before allowing hot standby again.
8688 : * This is important if the max_* settings are decreased, to ensure
8689 : * you don't run queries against the WAL preceding the change. The
8690 : * local copies cannot be updated as long as crash recovery is
8691 : * happening and we expect all the WAL to be replayed.
8692 : */
8693 64 : if (InArchiveRecovery)
8694 : {
8695 36 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
8696 36 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
8697 : }
8698 64 : if (LocalMinRecoveryPoint != InvalidXLogRecPtr && LocalMinRecoveryPoint < lsn)
8699 : {
8700 : TimeLineID replayTLI;
8701 :
8702 14 : (void) GetCurrentReplayRecPtr(&replayTLI);
8703 14 : ControlFile->minRecoveryPoint = lsn;
8704 14 : ControlFile->minRecoveryPointTLI = replayTLI;
8705 : }
8706 :
8707 64 : CommitTsParameterChange(xlrec.track_commit_timestamp,
8708 64 : ControlFile->track_commit_timestamp);
8709 64 : ControlFile->track_commit_timestamp = xlrec.track_commit_timestamp;
8710 :
8711 64 : UpdateControlFile();
8712 64 : LWLockRelease(ControlFileLock);
8713 :
8714 : /* Check to see if any parameter change gives a problem on recovery */
8715 64 : CheckRequiredParameterValues();
8716 : }
8717 1346 : else if (info == XLOG_FPW_CHANGE)
8718 : {
8719 : bool fpw;
8720 :
8721 0 : memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
8722 :
8723 : /*
8724 : * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
8725 : * do_pg_backup_start() and do_pg_backup_stop() can check whether
8726 : * full_page_writes has been disabled during online backup.
8727 : */
8728 0 : if (!fpw)
8729 : {
8730 0 : SpinLockAcquire(&XLogCtl->info_lck);
8731 0 : if (XLogCtl->lastFpwDisableRecPtr < record->ReadRecPtr)
8732 0 : XLogCtl->lastFpwDisableRecPtr = record->ReadRecPtr;
8733 0 : SpinLockRelease(&XLogCtl->info_lck);
8734 : }
8735 :
8736 : /* Keep track of full_page_writes */
8737 0 : lastFullPageWrites = fpw;
8738 : }
8739 : else if (info == XLOG_CHECKPOINT_REDO)
8740 : {
8741 : /* nothing to do here, just for informational purposes */
8742 : }
8743 86788 : }
8744 :
8745 : /*
8746 : * Return the extra open flags used for opening a file, depending on the
8747 : * value of the GUCs wal_sync_method, fsync and debug_io_direct.
8748 : */
8749 : static int
8750 31788 : get_sync_bit(int method)
8751 : {
8752 31788 : int o_direct_flag = 0;
8753 :
8754 : /*
8755 : * Use O_DIRECT if requested, except in walreceiver process. The WAL
8756 : * written by walreceiver is normally read by the startup process soon
8757 : * after it's written. Also, walreceiver performs unaligned writes, which
8758 : * don't work with O_DIRECT, so it is required for correctness too.
8759 : */
8760 31788 : if ((io_direct_flags & IO_DIRECT_WAL) && !AmWalReceiverProcess())
8761 16 : o_direct_flag = PG_O_DIRECT;
8762 :
8763 : /* If fsync is disabled, never open in sync mode */
8764 31788 : if (!enableFsync)
8765 31788 : return o_direct_flag;
8766 :
8767 0 : switch (method)
8768 : {
8769 : /*
8770 : * enum values for all sync options are defined even if they are
8771 : * not supported on the current platform. But if not, they are
8772 : * not included in the enum option array, and therefore will never
8773 : * be seen here.
8774 : */
8775 0 : case WAL_SYNC_METHOD_FSYNC:
8776 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8777 : case WAL_SYNC_METHOD_FDATASYNC:
8778 0 : return o_direct_flag;
8779 : #ifdef O_SYNC
8780 0 : case WAL_SYNC_METHOD_OPEN:
8781 0 : return O_SYNC | o_direct_flag;
8782 : #endif
8783 : #ifdef O_DSYNC
8784 0 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8785 0 : return O_DSYNC | o_direct_flag;
8786 : #endif
8787 0 : default:
8788 : /* can't happen (unless we are out of sync with option array) */
8789 0 : elog(ERROR, "unrecognized \"wal_sync_method\": %d", method);
8790 : return 0; /* silence warning */
8791 : }
8792 : }
8793 :
8794 : /*
8795 : * GUC support
8796 : */
8797 : void
8798 2226 : assign_wal_sync_method(int new_wal_sync_method, void *extra)
8799 : {
8800 2226 : if (wal_sync_method != new_wal_sync_method)
8801 : {
8802 : /*
8803 : * To ensure that no blocks escape unsynced, force an fsync on the
8804 : * currently open log segment (if any). Also, if the open flag is
8805 : * changing, close the log file so it will be reopened (with new flag
8806 : * bit) at next use.
8807 : */
8808 0 : if (openLogFile >= 0)
8809 : {
8810 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN);
8811 0 : if (pg_fsync(openLogFile) != 0)
8812 : {
8813 : char xlogfname[MAXFNAMELEN];
8814 : int save_errno;
8815 :
8816 0 : save_errno = errno;
8817 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo,
8818 : wal_segment_size);
8819 0 : errno = save_errno;
8820 0 : ereport(PANIC,
8821 : (errcode_for_file_access(),
8822 : errmsg("could not fsync file \"%s\": %m", xlogfname)));
8823 : }
8824 :
8825 0 : pgstat_report_wait_end();
8826 0 : if (get_sync_bit(wal_sync_method) != get_sync_bit(new_wal_sync_method))
8827 0 : XLogFileClose();
8828 : }
8829 : }
8830 2226 : }
8831 :
8832 :
8833 : /*
8834 : * Issue appropriate kind of fsync (if any) for an XLOG output file.
8835 : *
8836 : * 'fd' is a file descriptor for the XLOG file to be fsync'd.
8837 : * 'segno' is for error reporting purposes.
8838 : */
8839 : void
8840 318276 : issue_xlog_fsync(int fd, XLogSegNo segno, TimeLineID tli)
8841 : {
8842 318276 : char *msg = NULL;
8843 : instr_time start;
8844 :
8845 : Assert(tli != 0);
8846 :
8847 : /*
8848 : * Quick exit if fsync is disabled or write() has already synced the WAL
8849 : * file.
8850 : */
8851 318276 : if (!enableFsync ||
8852 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN ||
8853 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN_DSYNC)
8854 318276 : return;
8855 :
8856 : /*
8857 : * Measure I/O timing to sync the WAL file for pg_stat_io.
8858 : */
8859 0 : start = pgstat_prepare_io_time(track_wal_io_timing);
8860 :
8861 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC);
8862 0 : switch (wal_sync_method)
8863 : {
8864 0 : case WAL_SYNC_METHOD_FSYNC:
8865 0 : if (pg_fsync_no_writethrough(fd) != 0)
8866 0 : msg = _("could not fsync file \"%s\": %m");
8867 0 : break;
8868 : #ifdef HAVE_FSYNC_WRITETHROUGH
8869 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8870 : if (pg_fsync_writethrough(fd) != 0)
8871 : msg = _("could not fsync write-through file \"%s\": %m");
8872 : break;
8873 : #endif
8874 0 : case WAL_SYNC_METHOD_FDATASYNC:
8875 0 : if (pg_fdatasync(fd) != 0)
8876 0 : msg = _("could not fdatasync file \"%s\": %m");
8877 0 : break;
8878 0 : case WAL_SYNC_METHOD_OPEN:
8879 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8880 : /* not reachable */
8881 : Assert(false);
8882 0 : break;
8883 0 : default:
8884 0 : ereport(PANIC,
8885 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8886 : errmsg_internal("unrecognized \"wal_sync_method\": %d", wal_sync_method));
8887 : break;
8888 : }
8889 :
8890 : /* PANIC if failed to fsync */
8891 0 : if (msg)
8892 : {
8893 : char xlogfname[MAXFNAMELEN];
8894 0 : int save_errno = errno;
8895 :
8896 0 : XLogFileName(xlogfname, tli, segno, wal_segment_size);
8897 0 : errno = save_errno;
8898 0 : ereport(PANIC,
8899 : (errcode_for_file_access(),
8900 : errmsg(msg, xlogfname)));
8901 : }
8902 :
8903 0 : pgstat_report_wait_end();
8904 :
8905 0 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL, IOOP_FSYNC,
8906 : start, 1, 0);
8907 : }
8908 :
8909 : /*
8910 : * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
8911 : * function. It creates the necessary starting checkpoint and constructs the
8912 : * backup state and tablespace map.
8913 : *
8914 : * Input parameters are "state" (the backup state), "fast" (if true, we do
8915 : * the checkpoint in fast mode), and "tablespaces" (if non-NULL, indicates a
8916 : * list of tablespaceinfo structs describing the cluster's tablespaces.).
8917 : *
8918 : * The tablespace map contents are appended to passed-in parameter
8919 : * tablespace_map and the caller is responsible for including it in the backup
8920 : * archive as 'tablespace_map'. The tablespace_map file is required mainly for
8921 : * tar format in windows as native windows utilities are not able to create
8922 : * symlinks while extracting files from tar. However for consistency and
8923 : * platform-independence, we do it the same way everywhere.
8924 : *
8925 : * It fills in "state" with the information required for the backup, such
8926 : * as the minimum WAL location that must be present to restore from this
8927 : * backup (starttli) and the corresponding timeline ID (starttli).
8928 : *
8929 : * Every successfully started backup must be stopped by calling
8930 : * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
8931 : * backups active at the same time.
8932 : *
8933 : * It is the responsibility of the caller of this function to verify the
8934 : * permissions of the calling user!
8935 : */
8936 : void
8937 326 : do_pg_backup_start(const char *backupidstr, bool fast, List **tablespaces,
8938 : BackupState *state, StringInfo tblspcmapfile)
8939 : {
8940 : bool backup_started_in_recovery;
8941 :
8942 : Assert(state != NULL);
8943 326 : backup_started_in_recovery = RecoveryInProgress();
8944 :
8945 : /*
8946 : * During recovery, we don't need to check WAL level. Because, if WAL
8947 : * level is not sufficient, it's impossible to get here during recovery.
8948 : */
8949 326 : if (!backup_started_in_recovery && !XLogIsNeeded())
8950 0 : ereport(ERROR,
8951 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8952 : errmsg("WAL level not sufficient for making an online backup"),
8953 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
8954 :
8955 326 : if (strlen(backupidstr) > MAXPGPATH)
8956 2 : ereport(ERROR,
8957 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8958 : errmsg("backup label too long (max %d bytes)",
8959 : MAXPGPATH)));
8960 :
8961 324 : strlcpy(state->name, backupidstr, sizeof(state->name));
8962 :
8963 : /*
8964 : * Mark backup active in shared memory. We must do full-page WAL writes
8965 : * during an on-line backup even if not doing so at other times, because
8966 : * it's quite possible for the backup dump to obtain a "torn" (partially
8967 : * written) copy of a database page if it reads the page concurrently with
8968 : * our write to the same page. This can be fixed as long as the first
8969 : * write to the page in the WAL sequence is a full-page write. Hence, we
8970 : * increment runningBackups then force a CHECKPOINT, to ensure there are
8971 : * no dirty pages in shared memory that might get dumped while the backup
8972 : * is in progress without having a corresponding WAL record. (Once the
8973 : * backup is complete, we need not force full-page writes anymore, since
8974 : * we expect that any pages not modified during the backup interval must
8975 : * have been correctly captured by the backup.)
8976 : *
8977 : * Note that forcing full-page writes has no effect during an online
8978 : * backup from the standby.
8979 : *
8980 : * We must hold all the insertion locks to change the value of
8981 : * runningBackups, to ensure adequate interlocking against
8982 : * XLogInsertRecord().
8983 : */
8984 324 : WALInsertLockAcquireExclusive();
8985 324 : XLogCtl->Insert.runningBackups++;
8986 324 : WALInsertLockRelease();
8987 :
8988 : /*
8989 : * Ensure we decrement runningBackups if we fail below. NB -- for this to
8990 : * work correctly, it is critical that sessionBackupState is only updated
8991 : * after this block is over.
8992 : */
8993 324 : PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, DatumGetBool(true));
8994 : {
8995 324 : bool gotUniqueStartpoint = false;
8996 : DIR *tblspcdir;
8997 : struct dirent *de;
8998 : tablespaceinfo *ti;
8999 : int datadirpathlen;
9000 :
9001 : /*
9002 : * Force an XLOG file switch before the checkpoint, to ensure that the
9003 : * WAL segment the checkpoint is written to doesn't contain pages with
9004 : * old timeline IDs. That would otherwise happen if you called
9005 : * pg_backup_start() right after restoring from a PITR archive: the
9006 : * first WAL segment containing the startup checkpoint has pages in
9007 : * the beginning with the old timeline ID. That can cause trouble at
9008 : * recovery: we won't have a history file covering the old timeline if
9009 : * pg_wal directory was not included in the base backup and the WAL
9010 : * archive was cleared too before starting the backup.
9011 : *
9012 : * This also ensures that we have emitted a WAL page header that has
9013 : * XLP_BKP_REMOVABLE off before we emit the checkpoint record.
9014 : * Therefore, if a WAL archiver (such as pglesslog) is trying to
9015 : * compress out removable backup blocks, it won't remove any that
9016 : * occur after this point.
9017 : *
9018 : * During recovery, we skip forcing XLOG file switch, which means that
9019 : * the backup taken during recovery is not available for the special
9020 : * recovery case described above.
9021 : */
9022 324 : if (!backup_started_in_recovery)
9023 310 : RequestXLogSwitch(false);
9024 :
9025 : do
9026 : {
9027 : bool checkpointfpw;
9028 :
9029 : /*
9030 : * Force a CHECKPOINT. Aside from being necessary to prevent torn
9031 : * page problems, this guarantees that two successive backup runs
9032 : * will have different checkpoint positions and hence different
9033 : * history file names, even if nothing happened in between.
9034 : *
9035 : * During recovery, establish a restartpoint if possible. We use
9036 : * the last restartpoint as the backup starting checkpoint. This
9037 : * means that two successive backup runs can have same checkpoint
9038 : * positions.
9039 : *
9040 : * Since the fact that we are executing do_pg_backup_start()
9041 : * during recovery means that checkpointer is running, we can use
9042 : * RequestCheckpoint() to establish a restartpoint.
9043 : *
9044 : * We use CHECKPOINT_FAST only if requested by user (via passing
9045 : * fast = true). Otherwise this can take awhile.
9046 : */
9047 324 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
9048 : (fast ? CHECKPOINT_FAST : 0));
9049 :
9050 : /*
9051 : * Now we need to fetch the checkpoint record location, and also
9052 : * its REDO pointer. The oldest point in WAL that would be needed
9053 : * to restore starting from the checkpoint is precisely the REDO
9054 : * pointer.
9055 : */
9056 324 : LWLockAcquire(ControlFileLock, LW_SHARED);
9057 324 : state->checkpointloc = ControlFile->checkPoint;
9058 324 : state->startpoint = ControlFile->checkPointCopy.redo;
9059 324 : state->starttli = ControlFile->checkPointCopy.ThisTimeLineID;
9060 324 : checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
9061 324 : LWLockRelease(ControlFileLock);
9062 :
9063 324 : if (backup_started_in_recovery)
9064 : {
9065 : XLogRecPtr recptr;
9066 :
9067 : /*
9068 : * Check to see if all WAL replayed during online backup
9069 : * (i.e., since last restartpoint used as backup starting
9070 : * checkpoint) contain full-page writes.
9071 : */
9072 14 : SpinLockAcquire(&XLogCtl->info_lck);
9073 14 : recptr = XLogCtl->lastFpwDisableRecPtr;
9074 14 : SpinLockRelease(&XLogCtl->info_lck);
9075 :
9076 14 : if (!checkpointfpw || state->startpoint <= recptr)
9077 0 : ereport(ERROR,
9078 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9079 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9080 : "since last restartpoint"),
9081 : errhint("This means that the backup being taken on the standby "
9082 : "is corrupt and should not be used. "
9083 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9084 : "and then try an online backup again.")));
9085 :
9086 : /*
9087 : * During recovery, since we don't use the end-of-backup WAL
9088 : * record and don't write the backup history file, the
9089 : * starting WAL location doesn't need to be unique. This means
9090 : * that two base backups started at the same time might use
9091 : * the same checkpoint as starting locations.
9092 : */
9093 14 : gotUniqueStartpoint = true;
9094 : }
9095 :
9096 : /*
9097 : * If two base backups are started at the same time (in WAL sender
9098 : * processes), we need to make sure that they use different
9099 : * checkpoints as starting locations, because we use the starting
9100 : * WAL location as a unique identifier for the base backup in the
9101 : * end-of-backup WAL record and when we write the backup history
9102 : * file. Perhaps it would be better generate a separate unique ID
9103 : * for each backup instead of forcing another checkpoint, but
9104 : * taking a checkpoint right after another is not that expensive
9105 : * either because only few buffers have been dirtied yet.
9106 : */
9107 324 : WALInsertLockAcquireExclusive();
9108 324 : if (XLogCtl->Insert.lastBackupStart < state->startpoint)
9109 : {
9110 324 : XLogCtl->Insert.lastBackupStart = state->startpoint;
9111 324 : gotUniqueStartpoint = true;
9112 : }
9113 324 : WALInsertLockRelease();
9114 324 : } while (!gotUniqueStartpoint);
9115 :
9116 : /*
9117 : * Construct tablespace_map file.
9118 : */
9119 324 : datadirpathlen = strlen(DataDir);
9120 :
9121 : /* Collect information about all tablespaces */
9122 324 : tblspcdir = AllocateDir(PG_TBLSPC_DIR);
9123 1046 : while ((de = ReadDir(tblspcdir, PG_TBLSPC_DIR)) != NULL)
9124 : {
9125 : char fullpath[MAXPGPATH + sizeof(PG_TBLSPC_DIR)];
9126 : char linkpath[MAXPGPATH];
9127 722 : char *relpath = NULL;
9128 : char *s;
9129 : PGFileType de_type;
9130 : char *badp;
9131 : Oid tsoid;
9132 :
9133 : /*
9134 : * Try to parse the directory name as an unsigned integer.
9135 : *
9136 : * Tablespace directories should be positive integers that can be
9137 : * represented in 32 bits, with no leading zeroes or trailing
9138 : * garbage. If we come across a name that doesn't meet those
9139 : * criteria, skip it.
9140 : */
9141 722 : if (de->d_name[0] < '1' || de->d_name[1] > '9')
9142 648 : continue;
9143 74 : errno = 0;
9144 74 : tsoid = strtoul(de->d_name, &badp, 10);
9145 74 : if (*badp != '\0' || errno == EINVAL || errno == ERANGE)
9146 0 : continue;
9147 :
9148 74 : snprintf(fullpath, sizeof(fullpath), "%s/%s", PG_TBLSPC_DIR, de->d_name);
9149 :
9150 74 : de_type = get_dirent_type(fullpath, de, false, ERROR);
9151 :
9152 74 : if (de_type == PGFILETYPE_LNK)
9153 : {
9154 : StringInfoData escapedpath;
9155 : int rllen;
9156 :
9157 46 : rllen = readlink(fullpath, linkpath, sizeof(linkpath));
9158 46 : if (rllen < 0)
9159 : {
9160 0 : ereport(WARNING,
9161 : (errmsg("could not read symbolic link \"%s\": %m",
9162 : fullpath)));
9163 0 : continue;
9164 : }
9165 46 : else if (rllen >= sizeof(linkpath))
9166 : {
9167 0 : ereport(WARNING,
9168 : (errmsg("symbolic link \"%s\" target is too long",
9169 : fullpath)));
9170 0 : continue;
9171 : }
9172 46 : linkpath[rllen] = '\0';
9173 :
9174 : /*
9175 : * Relpath holds the relative path of the tablespace directory
9176 : * when it's located within PGDATA, or NULL if it's located
9177 : * elsewhere.
9178 : */
9179 46 : if (rllen > datadirpathlen &&
9180 2 : strncmp(linkpath, DataDir, datadirpathlen) == 0 &&
9181 0 : IS_DIR_SEP(linkpath[datadirpathlen]))
9182 0 : relpath = pstrdup(linkpath + datadirpathlen + 1);
9183 :
9184 : /*
9185 : * Add a backslash-escaped version of the link path to the
9186 : * tablespace map file.
9187 : */
9188 46 : initStringInfo(&escapedpath);
9189 1124 : for (s = linkpath; *s; s++)
9190 : {
9191 1078 : if (*s == '\n' || *s == '\r' || *s == '\\')
9192 0 : appendStringInfoChar(&escapedpath, '\\');
9193 1078 : appendStringInfoChar(&escapedpath, *s);
9194 : }
9195 46 : appendStringInfo(tblspcmapfile, "%s %s\n",
9196 46 : de->d_name, escapedpath.data);
9197 46 : pfree(escapedpath.data);
9198 : }
9199 28 : else if (de_type == PGFILETYPE_DIR)
9200 : {
9201 : /*
9202 : * It's possible to use allow_in_place_tablespaces to create
9203 : * directories directly under pg_tblspc, for testing purposes
9204 : * only.
9205 : *
9206 : * In this case, we store a relative path rather than an
9207 : * absolute path into the tablespaceinfo.
9208 : */
9209 28 : snprintf(linkpath, sizeof(linkpath), "%s/%s",
9210 28 : PG_TBLSPC_DIR, de->d_name);
9211 28 : relpath = pstrdup(linkpath);
9212 : }
9213 : else
9214 : {
9215 : /* Skip any other file type that appears here. */
9216 0 : continue;
9217 : }
9218 :
9219 74 : ti = palloc(sizeof(tablespaceinfo));
9220 74 : ti->oid = tsoid;
9221 74 : ti->path = pstrdup(linkpath);
9222 74 : ti->rpath = relpath;
9223 74 : ti->size = -1;
9224 :
9225 74 : if (tablespaces)
9226 74 : *tablespaces = lappend(*tablespaces, ti);
9227 : }
9228 324 : FreeDir(tblspcdir);
9229 :
9230 324 : state->starttime = (pg_time_t) time(NULL);
9231 : }
9232 324 : PG_END_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, DatumGetBool(true));
9233 :
9234 324 : state->started_in_recovery = backup_started_in_recovery;
9235 :
9236 : /*
9237 : * Mark that the start phase has correctly finished for the backup.
9238 : */
9239 324 : sessionBackupState = SESSION_BACKUP_RUNNING;
9240 324 : }
9241 :
9242 : /*
9243 : * Utility routine to fetch the session-level status of a backup running.
9244 : */
9245 : SessionBackupState
9246 366 : get_backup_status(void)
9247 : {
9248 366 : return sessionBackupState;
9249 : }
9250 :
9251 : /*
9252 : * do_pg_backup_stop
9253 : *
9254 : * Utility function called at the end of an online backup. It creates history
9255 : * file (if required), resets sessionBackupState and so on. It can optionally
9256 : * wait for WAL segments to be archived.
9257 : *
9258 : * "state" is filled with the information necessary to restore from this
9259 : * backup with its stop LSN (stoppoint), its timeline ID (stoptli), etc.
9260 : *
9261 : * It is the responsibility of the caller of this function to verify the
9262 : * permissions of the calling user!
9263 : */
9264 : void
9265 312 : do_pg_backup_stop(BackupState *state, bool waitforarchive)
9266 : {
9267 312 : bool backup_stopped_in_recovery = false;
9268 : char histfilepath[MAXPGPATH];
9269 : char lastxlogfilename[MAXFNAMELEN];
9270 : char histfilename[MAXFNAMELEN];
9271 : XLogSegNo _logSegNo;
9272 : FILE *fp;
9273 : int seconds_before_warning;
9274 312 : int waits = 0;
9275 312 : bool reported_waiting = false;
9276 :
9277 : Assert(state != NULL);
9278 :
9279 312 : backup_stopped_in_recovery = RecoveryInProgress();
9280 :
9281 : /*
9282 : * During recovery, we don't need to check WAL level. Because, if WAL
9283 : * level is not sufficient, it's impossible to get here during recovery.
9284 : */
9285 312 : if (!backup_stopped_in_recovery && !XLogIsNeeded())
9286 0 : ereport(ERROR,
9287 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9288 : errmsg("WAL level not sufficient for making an online backup"),
9289 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9290 :
9291 : /*
9292 : * OK to update backup counter and session-level lock.
9293 : *
9294 : * Note that CHECK_FOR_INTERRUPTS() must not occur while updating them,
9295 : * otherwise they can be updated inconsistently, which might cause
9296 : * do_pg_abort_backup() to fail.
9297 : */
9298 312 : WALInsertLockAcquireExclusive();
9299 :
9300 : /*
9301 : * It is expected that each do_pg_backup_start() call is matched by
9302 : * exactly one do_pg_backup_stop() call.
9303 : */
9304 : Assert(XLogCtl->Insert.runningBackups > 0);
9305 312 : XLogCtl->Insert.runningBackups--;
9306 :
9307 : /*
9308 : * Clean up session-level lock.
9309 : *
9310 : * You might think that WALInsertLockRelease() can be called before
9311 : * cleaning up session-level lock because session-level lock doesn't need
9312 : * to be protected with WAL insertion lock. But since
9313 : * CHECK_FOR_INTERRUPTS() can occur in it, session-level lock must be
9314 : * cleaned up before it.
9315 : */
9316 312 : sessionBackupState = SESSION_BACKUP_NONE;
9317 :
9318 312 : WALInsertLockRelease();
9319 :
9320 : /*
9321 : * If we are taking an online backup from the standby, we confirm that the
9322 : * standby has not been promoted during the backup.
9323 : */
9324 312 : if (state->started_in_recovery && !backup_stopped_in_recovery)
9325 0 : ereport(ERROR,
9326 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9327 : errmsg("the standby was promoted during online backup"),
9328 : errhint("This means that the backup being taken is corrupt "
9329 : "and should not be used. "
9330 : "Try taking another online backup.")));
9331 :
9332 : /*
9333 : * During recovery, we don't write an end-of-backup record. We assume that
9334 : * pg_control was backed up last and its minimum recovery point can be
9335 : * available as the backup end location. Since we don't have an
9336 : * end-of-backup record, we use the pg_control value to check whether
9337 : * we've reached the end of backup when starting recovery from this
9338 : * backup. We have no way of checking if pg_control wasn't backed up last
9339 : * however.
9340 : *
9341 : * We don't force a switch to new WAL file but it is still possible to
9342 : * wait for all the required files to be archived if waitforarchive is
9343 : * true. This is okay if we use the backup to start a standby and fetch
9344 : * the missing WAL using streaming replication. But in the case of an
9345 : * archive recovery, a user should set waitforarchive to true and wait for
9346 : * them to be archived to ensure that all the required files are
9347 : * available.
9348 : *
9349 : * We return the current minimum recovery point as the backup end
9350 : * location. Note that it can be greater than the exact backup end
9351 : * location if the minimum recovery point is updated after the backup of
9352 : * pg_control. This is harmless for current uses.
9353 : *
9354 : * XXX currently a backup history file is for informational and debug
9355 : * purposes only. It's not essential for an online backup. Furthermore,
9356 : * even if it's created, it will not be archived during recovery because
9357 : * an archiver is not invoked. So it doesn't seem worthwhile to write a
9358 : * backup history file during recovery.
9359 : */
9360 312 : if (backup_stopped_in_recovery)
9361 : {
9362 : XLogRecPtr recptr;
9363 :
9364 : /*
9365 : * Check to see if all WAL replayed during online backup contain
9366 : * full-page writes.
9367 : */
9368 14 : SpinLockAcquire(&XLogCtl->info_lck);
9369 14 : recptr = XLogCtl->lastFpwDisableRecPtr;
9370 14 : SpinLockRelease(&XLogCtl->info_lck);
9371 :
9372 14 : if (state->startpoint <= recptr)
9373 0 : ereport(ERROR,
9374 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9375 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9376 : "during online backup"),
9377 : errhint("This means that the backup being taken on the standby "
9378 : "is corrupt and should not be used. "
9379 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9380 : "and then try an online backup again.")));
9381 :
9382 :
9383 14 : LWLockAcquire(ControlFileLock, LW_SHARED);
9384 14 : state->stoppoint = ControlFile->minRecoveryPoint;
9385 14 : state->stoptli = ControlFile->minRecoveryPointTLI;
9386 14 : LWLockRelease(ControlFileLock);
9387 : }
9388 : else
9389 : {
9390 : char *history_file;
9391 :
9392 : /*
9393 : * Write the backup-end xlog record
9394 : */
9395 298 : XLogBeginInsert();
9396 298 : XLogRegisterData(&state->startpoint,
9397 : sizeof(state->startpoint));
9398 298 : state->stoppoint = XLogInsert(RM_XLOG_ID, XLOG_BACKUP_END);
9399 :
9400 : /*
9401 : * Given that we're not in recovery, InsertTimeLineID is set and can't
9402 : * change, so we can read it without a lock.
9403 : */
9404 298 : state->stoptli = XLogCtl->InsertTimeLineID;
9405 :
9406 : /*
9407 : * Force a switch to a new xlog segment file, so that the backup is
9408 : * valid as soon as archiver moves out the current segment file.
9409 : */
9410 298 : RequestXLogSwitch(false);
9411 :
9412 298 : state->stoptime = (pg_time_t) time(NULL);
9413 :
9414 : /*
9415 : * Write the backup history file
9416 : */
9417 298 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9418 298 : BackupHistoryFilePath(histfilepath, state->stoptli, _logSegNo,
9419 : state->startpoint, wal_segment_size);
9420 298 : fp = AllocateFile(histfilepath, "w");
9421 298 : if (!fp)
9422 0 : ereport(ERROR,
9423 : (errcode_for_file_access(),
9424 : errmsg("could not create file \"%s\": %m",
9425 : histfilepath)));
9426 :
9427 : /* Build and save the contents of the backup history file */
9428 298 : history_file = build_backup_content(state, true);
9429 298 : fprintf(fp, "%s", history_file);
9430 298 : pfree(history_file);
9431 :
9432 298 : if (fflush(fp) || ferror(fp) || FreeFile(fp))
9433 0 : ereport(ERROR,
9434 : (errcode_for_file_access(),
9435 : errmsg("could not write file \"%s\": %m",
9436 : histfilepath)));
9437 :
9438 : /*
9439 : * Clean out any no-longer-needed history files. As a side effect,
9440 : * this will post a .ready file for the newly created history file,
9441 : * notifying the archiver that history file may be archived
9442 : * immediately.
9443 : */
9444 298 : CleanupBackupHistory();
9445 : }
9446 :
9447 : /*
9448 : * If archiving is enabled, wait for all the required WAL files to be
9449 : * archived before returning. If archiving isn't enabled, the required WAL
9450 : * needs to be transported via streaming replication (hopefully with
9451 : * wal_keep_size set high enough), or some more exotic mechanism like
9452 : * polling and copying files from pg_wal with script. We have no knowledge
9453 : * of those mechanisms, so it's up to the user to ensure that he gets all
9454 : * the required WAL.
9455 : *
9456 : * We wait until both the last WAL file filled during backup and the
9457 : * history file have been archived, and assume that the alphabetic sorting
9458 : * property of the WAL files ensures any earlier WAL files are safely
9459 : * archived as well.
9460 : *
9461 : * We wait forever, since archive_command is supposed to work and we
9462 : * assume the admin wanted his backup to work completely. If you don't
9463 : * wish to wait, then either waitforarchive should be passed in as false,
9464 : * or you can set statement_timeout. Also, some notices are issued to
9465 : * clue in anyone who might be doing this interactively.
9466 : */
9467 :
9468 312 : if (waitforarchive &&
9469 20 : ((!backup_stopped_in_recovery && XLogArchivingActive()) ||
9470 2 : (backup_stopped_in_recovery && XLogArchivingAlways())))
9471 : {
9472 8 : XLByteToPrevSeg(state->stoppoint, _logSegNo, wal_segment_size);
9473 8 : XLogFileName(lastxlogfilename, state->stoptli, _logSegNo,
9474 : wal_segment_size);
9475 :
9476 8 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9477 8 : BackupHistoryFileName(histfilename, state->stoptli, _logSegNo,
9478 : state->startpoint, wal_segment_size);
9479 :
9480 8 : seconds_before_warning = 60;
9481 8 : waits = 0;
9482 :
9483 28 : while (XLogArchiveIsBusy(lastxlogfilename) ||
9484 8 : XLogArchiveIsBusy(histfilename))
9485 : {
9486 12 : CHECK_FOR_INTERRUPTS();
9487 :
9488 12 : if (!reported_waiting && waits > 5)
9489 : {
9490 0 : ereport(NOTICE,
9491 : (errmsg("base backup done, waiting for required WAL segments to be archived")));
9492 0 : reported_waiting = true;
9493 : }
9494 :
9495 12 : (void) WaitLatch(MyLatch,
9496 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
9497 : 1000L,
9498 : WAIT_EVENT_BACKUP_WAIT_WAL_ARCHIVE);
9499 12 : ResetLatch(MyLatch);
9500 :
9501 12 : if (++waits >= seconds_before_warning)
9502 : {
9503 0 : seconds_before_warning *= 2; /* This wraps in >10 years... */
9504 0 : ereport(WARNING,
9505 : (errmsg("still waiting for all required WAL segments to be archived (%d seconds elapsed)",
9506 : waits),
9507 : errhint("Check that your \"archive_command\" is executing properly. "
9508 : "You can safely cancel this backup, "
9509 : "but the database backup will not be usable without all the WAL segments.")));
9510 : }
9511 : }
9512 :
9513 8 : ereport(NOTICE,
9514 : (errmsg("all required WAL segments have been archived")));
9515 : }
9516 304 : else if (waitforarchive)
9517 12 : ereport(NOTICE,
9518 : (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
9519 312 : }
9520 :
9521 :
9522 : /*
9523 : * do_pg_abort_backup: abort a running backup
9524 : *
9525 : * This does just the most basic steps of do_pg_backup_stop(), by taking the
9526 : * system out of backup mode, thus making it a lot more safe to call from
9527 : * an error handler.
9528 : *
9529 : * 'arg' indicates that it's being called during backup setup; so
9530 : * sessionBackupState has not been modified yet, but runningBackups has
9531 : * already been incremented. When it's false, then it's invoked as a
9532 : * before_shmem_exit handler, and therefore we must not change state
9533 : * unless sessionBackupState indicates that a backup is actually running.
9534 : *
9535 : * NB: This gets used as a PG_ENSURE_ERROR_CLEANUP callback and
9536 : * before_shmem_exit handler, hence the odd-looking signature.
9537 : */
9538 : void
9539 16 : do_pg_abort_backup(int code, Datum arg)
9540 : {
9541 16 : bool during_backup_start = DatumGetBool(arg);
9542 :
9543 : /* If called during backup start, there shouldn't be one already running */
9544 : Assert(!during_backup_start || sessionBackupState == SESSION_BACKUP_NONE);
9545 :
9546 16 : if (during_backup_start || sessionBackupState != SESSION_BACKUP_NONE)
9547 : {
9548 12 : WALInsertLockAcquireExclusive();
9549 : Assert(XLogCtl->Insert.runningBackups > 0);
9550 12 : XLogCtl->Insert.runningBackups--;
9551 :
9552 12 : sessionBackupState = SESSION_BACKUP_NONE;
9553 12 : WALInsertLockRelease();
9554 :
9555 12 : if (!during_backup_start)
9556 12 : ereport(WARNING,
9557 : errmsg("aborting backup due to backend exiting before pg_backup_stop was called"));
9558 : }
9559 16 : }
9560 :
9561 : /*
9562 : * Register a handler that will warn about unterminated backups at end of
9563 : * session, unless this has already been done.
9564 : */
9565 : void
9566 8 : register_persistent_abort_backup_handler(void)
9567 : {
9568 : static bool already_done = false;
9569 :
9570 8 : if (already_done)
9571 2 : return;
9572 6 : before_shmem_exit(do_pg_abort_backup, DatumGetBool(false));
9573 6 : already_done = true;
9574 : }
9575 :
9576 : /*
9577 : * Get latest WAL insert pointer
9578 : */
9579 : XLogRecPtr
9580 3964 : GetXLogInsertRecPtr(void)
9581 : {
9582 3964 : XLogCtlInsert *Insert = &XLogCtl->Insert;
9583 : uint64 current_bytepos;
9584 :
9585 3964 : SpinLockAcquire(&Insert->insertpos_lck);
9586 3964 : current_bytepos = Insert->CurrBytePos;
9587 3964 : SpinLockRelease(&Insert->insertpos_lck);
9588 :
9589 3964 : return XLogBytePosToRecPtr(current_bytepos);
9590 : }
9591 :
9592 : /*
9593 : * Get latest WAL write pointer
9594 : */
9595 : XLogRecPtr
9596 2992 : GetXLogWriteRecPtr(void)
9597 : {
9598 2992 : RefreshXLogWriteResult(LogwrtResult);
9599 :
9600 2992 : return LogwrtResult.Write;
9601 : }
9602 :
9603 : /*
9604 : * Returns the redo pointer of the last checkpoint or restartpoint. This is
9605 : * the oldest point in WAL that we still need, if we have to restart recovery.
9606 : */
9607 : void
9608 780 : GetOldestRestartPoint(XLogRecPtr *oldrecptr, TimeLineID *oldtli)
9609 : {
9610 780 : LWLockAcquire(ControlFileLock, LW_SHARED);
9611 780 : *oldrecptr = ControlFile->checkPointCopy.redo;
9612 780 : *oldtli = ControlFile->checkPointCopy.ThisTimeLineID;
9613 780 : LWLockRelease(ControlFileLock);
9614 780 : }
9615 :
9616 : /* Thin wrapper around ShutdownWalRcv(). */
9617 : void
9618 2118 : XLogShutdownWalRcv(void)
9619 : {
9620 2118 : ShutdownWalRcv();
9621 :
9622 2118 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9623 2118 : XLogCtl->InstallXLogFileSegmentActive = false;
9624 2118 : LWLockRelease(ControlFileLock);
9625 2118 : }
9626 :
9627 : /* Enable WAL file recycling and preallocation. */
9628 : void
9629 2224 : SetInstallXLogFileSegmentActive(void)
9630 : {
9631 2224 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9632 2224 : XLogCtl->InstallXLogFileSegmentActive = true;
9633 2224 : LWLockRelease(ControlFileLock);
9634 2224 : }
9635 :
9636 : bool
9637 0 : IsInstallXLogFileSegmentActive(void)
9638 : {
9639 : bool result;
9640 :
9641 0 : LWLockAcquire(ControlFileLock, LW_SHARED);
9642 0 : result = XLogCtl->InstallXLogFileSegmentActive;
9643 0 : LWLockRelease(ControlFileLock);
9644 :
9645 0 : return result;
9646 : }
9647 :
9648 : /*
9649 : * Update the WalWriterSleeping flag.
9650 : */
9651 : void
9652 968 : SetWalWriterSleeping(bool sleeping)
9653 : {
9654 968 : SpinLockAcquire(&XLogCtl->info_lck);
9655 968 : XLogCtl->WalWriterSleeping = sleeping;
9656 968 : SpinLockRelease(&XLogCtl->info_lck);
9657 968 : }
|
