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 "access/xlogwait.h"
66 : #include "backup/basebackup.h"
67 : #include "catalog/catversion.h"
68 : #include "catalog/pg_control.h"
69 : #include "catalog/pg_database.h"
70 : #include "common/controldata_utils.h"
71 : #include "common/file_utils.h"
72 : #include "executor/instrument.h"
73 : #include "miscadmin.h"
74 : #include "pg_trace.h"
75 : #include "pgstat.h"
76 : #include "port/atomics.h"
77 : #include "postmaster/bgwriter.h"
78 : #include "postmaster/startup.h"
79 : #include "postmaster/walsummarizer.h"
80 : #include "postmaster/walwriter.h"
81 : #include "replication/origin.h"
82 : #include "replication/slot.h"
83 : #include "replication/snapbuild.h"
84 : #include "replication/walreceiver.h"
85 : #include "replication/walsender.h"
86 : #include "storage/bufmgr.h"
87 : #include "storage/fd.h"
88 : #include "storage/ipc.h"
89 : #include "storage/large_object.h"
90 : #include "storage/latch.h"
91 : #include "storage/predicate.h"
92 : #include "storage/proc.h"
93 : #include "storage/procarray.h"
94 : #include "storage/reinit.h"
95 : #include "storage/spin.h"
96 : #include "storage/sync.h"
97 : #include "utils/guc_hooks.h"
98 : #include "utils/guc_tables.h"
99 : #include "utils/injection_point.h"
100 : #include "utils/pgstat_internal.h"
101 : #include "utils/ps_status.h"
102 : #include "utils/relmapper.h"
103 : #include "utils/snapmgr.h"
104 : #include "utils/timeout.h"
105 : #include "utils/timestamp.h"
106 : #include "utils/varlena.h"
107 :
108 : #ifdef WAL_DEBUG
109 : #include "utils/memutils.h"
110 : #endif
111 :
112 : /* timeline ID to be used when bootstrapping */
113 : #define BootstrapTimeLineID 1
114 :
115 : /* User-settable parameters */
116 : int max_wal_size_mb = 1024; /* 1 GB */
117 : int min_wal_size_mb = 80; /* 80 MB */
118 : int wal_keep_size_mb = 0;
119 : int XLOGbuffers = -1;
120 : int XLogArchiveTimeout = 0;
121 : int XLogArchiveMode = ARCHIVE_MODE_OFF;
122 : char *XLogArchiveCommand = NULL;
123 : bool EnableHotStandby = false;
124 : bool fullPageWrites = true;
125 : bool wal_log_hints = false;
126 : int wal_compression = WAL_COMPRESSION_NONE;
127 : char *wal_consistency_checking_string = NULL;
128 : bool *wal_consistency_checking = NULL;
129 : bool wal_init_zero = true;
130 : bool wal_recycle = true;
131 : bool log_checkpoints = true;
132 : int wal_sync_method = DEFAULT_WAL_SYNC_METHOD;
133 : int wal_level = WAL_LEVEL_REPLICA;
134 : int CommitDelay = 0; /* precommit delay in microseconds */
135 : int CommitSiblings = 5; /* # concurrent xacts needed to sleep */
136 : int wal_retrieve_retry_interval = 5000;
137 : int max_slot_wal_keep_size_mb = -1;
138 : int wal_decode_buffer_size = 512 * 1024;
139 : bool track_wal_io_timing = false;
140 :
141 : #ifdef WAL_DEBUG
142 : bool XLOG_DEBUG = false;
143 : #endif
144 :
145 : int wal_segment_size = DEFAULT_XLOG_SEG_SIZE;
146 :
147 : /*
148 : * Number of WAL insertion locks to use. A higher value allows more insertions
149 : * to happen concurrently, but adds some CPU overhead to flushing the WAL,
150 : * which needs to iterate all the locks.
151 : */
152 : #define NUM_XLOGINSERT_LOCKS 8
153 :
154 : /*
155 : * Max distance from last checkpoint, before triggering a new xlog-based
156 : * checkpoint.
157 : */
158 : int CheckPointSegments;
159 :
160 : /* Estimated distance between checkpoints, in bytes */
161 : static double CheckPointDistanceEstimate = 0;
162 : static double PrevCheckPointDistance = 0;
163 :
164 : /*
165 : * Track whether there were any deferred checks for custom resource managers
166 : * specified in wal_consistency_checking.
167 : */
168 : static bool check_wal_consistency_checking_deferred = false;
169 :
170 : /*
171 : * GUC support
172 : */
173 : const struct config_enum_entry wal_sync_method_options[] = {
174 : {"fsync", WAL_SYNC_METHOD_FSYNC, false},
175 : #ifdef HAVE_FSYNC_WRITETHROUGH
176 : {"fsync_writethrough", WAL_SYNC_METHOD_FSYNC_WRITETHROUGH, false},
177 : #endif
178 : {"fdatasync", WAL_SYNC_METHOD_FDATASYNC, false},
179 : #ifdef O_SYNC
180 : {"open_sync", WAL_SYNC_METHOD_OPEN, false},
181 : #endif
182 : #ifdef O_DSYNC
183 : {"open_datasync", WAL_SYNC_METHOD_OPEN_DSYNC, false},
184 : #endif
185 : {NULL, 0, false}
186 : };
187 :
188 :
189 : /*
190 : * Although only "on", "off", and "always" are documented,
191 : * we accept all the likely variants of "on" and "off".
192 : */
193 : const struct config_enum_entry archive_mode_options[] = {
194 : {"always", ARCHIVE_MODE_ALWAYS, false},
195 : {"on", ARCHIVE_MODE_ON, false},
196 : {"off", ARCHIVE_MODE_OFF, false},
197 : {"true", ARCHIVE_MODE_ON, true},
198 : {"false", ARCHIVE_MODE_OFF, true},
199 : {"yes", ARCHIVE_MODE_ON, true},
200 : {"no", ARCHIVE_MODE_OFF, true},
201 : {"1", ARCHIVE_MODE_ON, true},
202 : {"0", ARCHIVE_MODE_OFF, true},
203 : {NULL, 0, false}
204 : };
205 :
206 : /*
207 : * Statistics for current checkpoint are collected in this global struct.
208 : * Because only the checkpointer or a stand-alone backend can perform
209 : * checkpoints, this will be unused in normal backends.
210 : */
211 : CheckpointStatsData CheckpointStats;
212 :
213 : /*
214 : * During recovery, lastFullPageWrites keeps track of full_page_writes that
215 : * the replayed WAL records indicate. It's initialized with full_page_writes
216 : * that the recovery starting checkpoint record indicates, and then updated
217 : * each time XLOG_FPW_CHANGE record is replayed.
218 : */
219 : static bool lastFullPageWrites;
220 :
221 : /*
222 : * Local copy of the state tracked by SharedRecoveryState in shared memory,
223 : * It is false if SharedRecoveryState is RECOVERY_STATE_DONE. True actually
224 : * means "not known, need to check the shared state".
225 : */
226 : static bool LocalRecoveryInProgress = true;
227 :
228 : /*
229 : * Local state for XLogInsertAllowed():
230 : * 1: unconditionally allowed to insert XLOG
231 : * 0: unconditionally not allowed to insert XLOG
232 : * -1: must check RecoveryInProgress(); disallow until it is false
233 : * Most processes start with -1 and transition to 1 after seeing that recovery
234 : * is not in progress. But we can also force the value for special cases.
235 : * The coding in XLogInsertAllowed() depends on the first two of these states
236 : * being numerically the same as bool true and false.
237 : */
238 : static int LocalXLogInsertAllowed = -1;
239 :
240 : /*
241 : * ProcLastRecPtr points to the start of the last XLOG record inserted by the
242 : * current backend. It is updated for all inserts. XactLastRecEnd points to
243 : * end+1 of the last record, and is reset when we end a top-level transaction,
244 : * or start a new one; so it can be used to tell if the current transaction has
245 : * created any XLOG records.
246 : *
247 : * While in parallel mode, this may not be fully up to date. When committing,
248 : * a transaction can assume this covers all xlog records written either by the
249 : * user backend or by any parallel worker which was present at any point during
250 : * the transaction. But when aborting, or when still in parallel mode, other
251 : * parallel backends may have written WAL records at later LSNs than the value
252 : * stored here. The parallel leader advances its own copy, when necessary,
253 : * in WaitForParallelWorkersToFinish.
254 : */
255 : XLogRecPtr ProcLastRecPtr = InvalidXLogRecPtr;
256 : XLogRecPtr XactLastRecEnd = InvalidXLogRecPtr;
257 : XLogRecPtr XactLastCommitEnd = InvalidXLogRecPtr;
258 :
259 : /*
260 : * RedoRecPtr is this backend's local copy of the REDO record pointer
261 : * (which is almost but not quite the same as a pointer to the most recent
262 : * CHECKPOINT record). We update this from the shared-memory copy,
263 : * XLogCtl->Insert.RedoRecPtr, whenever we can safely do so (ie, when we
264 : * hold an insertion lock). See XLogInsertRecord for details. We are also
265 : * allowed to update from XLogCtl->RedoRecPtr if we hold the info_lck;
266 : * see GetRedoRecPtr.
267 : *
268 : * NB: Code that uses this variable must be prepared not only for the
269 : * possibility that it may be arbitrarily out of date, but also for the
270 : * possibility that it might be set to InvalidXLogRecPtr. We used to
271 : * initialize it as a side effect of the first call to RecoveryInProgress(),
272 : * which meant that most code that might use it could assume that it had a
273 : * real if perhaps stale value. That's no longer the case.
274 : */
275 : static XLogRecPtr RedoRecPtr;
276 :
277 : /*
278 : * doPageWrites is this backend's local copy of (fullPageWrites ||
279 : * runningBackups > 0). It is used together with RedoRecPtr to decide whether
280 : * a full-page image of a page need to be taken.
281 : *
282 : * NB: Initially this is false, and there's no guarantee that it will be
283 : * initialized to any other value before it is first used. Any code that
284 : * makes use of it must recheck the value after obtaining a WALInsertLock,
285 : * and respond appropriately if it turns out that the previous value wasn't
286 : * accurate.
287 : */
288 : static bool doPageWrites;
289 :
290 : /*----------
291 : * Shared-memory data structures for XLOG control
292 : *
293 : * LogwrtRqst indicates a byte position that we need to write and/or fsync
294 : * the log up to (all records before that point must be written or fsynced).
295 : * The positions already written/fsynced are maintained in logWriteResult
296 : * and logFlushResult using atomic access.
297 : * In addition to the shared variable, each backend has a private copy of
298 : * both in LogwrtResult, which is updated when convenient.
299 : *
300 : * The request bookkeeping is simpler: there is a shared XLogCtl->LogwrtRqst
301 : * (protected by info_lck), but we don't need to cache any copies of it.
302 : *
303 : * info_lck is only held long enough to read/update the protected variables,
304 : * so it's a plain spinlock. The other locks are held longer (potentially
305 : * over I/O operations), so we use LWLocks for them. These locks are:
306 : *
307 : * WALBufMappingLock: must be held to replace a page in the WAL buffer cache.
308 : * It is only held while initializing and changing the mapping. If the
309 : * contents of the buffer being replaced haven't been written yet, the mapping
310 : * lock is released while the write is done, and reacquired afterwards.
311 : *
312 : * WALWriteLock: must be held to write WAL buffers to disk (XLogWrite or
313 : * XLogFlush).
314 : *
315 : * ControlFileLock: must be held to read/update control file or create
316 : * new log file.
317 : *
318 : *----------
319 : */
320 :
321 : typedef struct XLogwrtRqst
322 : {
323 : XLogRecPtr Write; /* last byte + 1 to write out */
324 : XLogRecPtr Flush; /* last byte + 1 to flush */
325 : } XLogwrtRqst;
326 :
327 : typedef struct XLogwrtResult
328 : {
329 : XLogRecPtr Write; /* last byte + 1 written out */
330 : XLogRecPtr Flush; /* last byte + 1 flushed */
331 : } XLogwrtResult;
332 :
333 : /*
334 : * Inserting to WAL is protected by a small fixed number of WAL insertion
335 : * locks. To insert to the WAL, you must hold one of the locks - it doesn't
336 : * matter which one. To lock out other concurrent insertions, you must hold
337 : * of them. Each WAL insertion lock consists of a lightweight lock, plus an
338 : * indicator of how far the insertion has progressed (insertingAt).
339 : *
340 : * The insertingAt values are read when a process wants to flush WAL from
341 : * the in-memory buffers to disk, to check that all the insertions to the
342 : * region the process is about to write out have finished. You could simply
343 : * wait for all currently in-progress insertions to finish, but the
344 : * insertingAt indicator allows you to ignore insertions to later in the WAL,
345 : * so that you only wait for the insertions that are modifying the buffers
346 : * you're about to write out.
347 : *
348 : * This isn't just an optimization. If all the WAL buffers are dirty, an
349 : * inserter that's holding a WAL insert lock might need to evict an old WAL
350 : * buffer, which requires flushing the WAL. If it's possible for an inserter
351 : * to block on another inserter unnecessarily, deadlock can arise when two
352 : * inserters holding a WAL insert lock wait for each other to finish their
353 : * insertion.
354 : *
355 : * Small WAL records that don't cross a page boundary never update the value,
356 : * the WAL record is just copied to the page and the lock is released. But
357 : * to avoid the deadlock-scenario explained above, the indicator is always
358 : * updated before sleeping while holding an insertion lock.
359 : *
360 : * lastImportantAt contains the LSN of the last important WAL record inserted
361 : * using a given lock. This value is used to detect if there has been
362 : * important WAL activity since the last time some action, like a checkpoint,
363 : * was performed - allowing to not repeat the action if not. The LSN is
364 : * updated for all insertions, unless the XLOG_MARK_UNIMPORTANT flag was
365 : * set. lastImportantAt is never cleared, only overwritten by the LSN of newer
366 : * records. Tracking the WAL activity directly in WALInsertLock has the
367 : * advantage of not needing any additional locks to update the value.
368 : */
369 : typedef struct
370 : {
371 : LWLock lock;
372 : pg_atomic_uint64 insertingAt;
373 : XLogRecPtr lastImportantAt;
374 : } WALInsertLock;
375 :
376 : /*
377 : * All the WAL insertion locks are allocated as an array in shared memory. We
378 : * force the array stride to be a power of 2, which saves a few cycles in
379 : * indexing, but more importantly also ensures that individual slots don't
380 : * cross cache line boundaries. (Of course, we have to also ensure that the
381 : * array start address is suitably aligned.)
382 : */
383 : typedef union WALInsertLockPadded
384 : {
385 : WALInsertLock l;
386 : char pad[PG_CACHE_LINE_SIZE];
387 : } WALInsertLockPadded;
388 :
389 : /*
390 : * Session status of running backup, used for sanity checks in SQL-callable
391 : * functions to start and stop backups.
392 : */
393 : static SessionBackupState sessionBackupState = SESSION_BACKUP_NONE;
394 :
395 : /*
396 : * Shared state data for WAL insertion.
397 : */
398 : typedef struct XLogCtlInsert
399 : {
400 : slock_t insertpos_lck; /* protects CurrBytePos and PrevBytePos */
401 :
402 : /*
403 : * CurrBytePos is the end of reserved WAL. The next record will be
404 : * inserted at that position. PrevBytePos is the start position of the
405 : * previously inserted (or rather, reserved) record - it is copied to the
406 : * prev-link of the next record. These are stored as "usable byte
407 : * positions" rather than XLogRecPtrs (see XLogBytePosToRecPtr()).
408 : */
409 : uint64 CurrBytePos;
410 : uint64 PrevBytePos;
411 :
412 : /*
413 : * Make sure the above heavily-contended spinlock and byte positions are
414 : * on their own cache line. In particular, the RedoRecPtr and full page
415 : * write variables below should be on a different cache line. They are
416 : * read on every WAL insertion, but updated rarely, and we don't want
417 : * those reads to steal the cache line containing Curr/PrevBytePos.
418 : */
419 : char pad[PG_CACHE_LINE_SIZE];
420 :
421 : /*
422 : * fullPageWrites is the authoritative value used by all backends to
423 : * determine whether to write full-page image to WAL. This shared value,
424 : * instead of the process-local fullPageWrites, is required because, when
425 : * full_page_writes is changed by SIGHUP, we must WAL-log it before it
426 : * actually affects WAL-logging by backends. Checkpointer sets at startup
427 : * or after SIGHUP.
428 : *
429 : * To read these fields, you must hold an insertion lock. To modify them,
430 : * you must hold ALL the locks.
431 : */
432 : XLogRecPtr RedoRecPtr; /* current redo point for insertions */
433 : bool fullPageWrites;
434 :
435 : /*
436 : * runningBackups is a counter indicating the number of backups currently
437 : * in progress. lastBackupStart is the latest checkpoint redo location
438 : * used as a starting point for an online backup.
439 : */
440 : int runningBackups;
441 : XLogRecPtr lastBackupStart;
442 :
443 : /*
444 : * WAL insertion locks.
445 : */
446 : WALInsertLockPadded *WALInsertLocks;
447 : } XLogCtlInsert;
448 :
449 : /*
450 : * Total shared-memory state for XLOG.
451 : */
452 : typedef struct XLogCtlData
453 : {
454 : XLogCtlInsert Insert;
455 :
456 : /* Protected by info_lck: */
457 : XLogwrtRqst LogwrtRqst;
458 : XLogRecPtr RedoRecPtr; /* a recent copy of Insert->RedoRecPtr */
459 : XLogRecPtr asyncXactLSN; /* LSN of newest async commit/abort */
460 : XLogRecPtr replicationSlotMinLSN; /* oldest LSN needed by any slot */
461 :
462 : XLogSegNo lastRemovedSegNo; /* latest removed/recycled XLOG segment */
463 :
464 : /* Fake LSN counter, for unlogged relations. */
465 : pg_atomic_uint64 unloggedLSN;
466 :
467 : /* Time and LSN of last xlog segment switch. Protected by WALWriteLock. */
468 : pg_time_t lastSegSwitchTime;
469 : XLogRecPtr lastSegSwitchLSN;
470 :
471 : /* These are accessed using atomics -- info_lck not needed */
472 : pg_atomic_uint64 logInsertResult; /* last byte + 1 inserted to buffers */
473 : pg_atomic_uint64 logWriteResult; /* last byte + 1 written out */
474 : pg_atomic_uint64 logFlushResult; /* last byte + 1 flushed */
475 :
476 : /*
477 : * Latest initialized page in the cache (last byte position + 1).
478 : *
479 : * To change the identity of a buffer (and InitializedUpTo), you need to
480 : * hold WALBufMappingLock. To change the identity of a buffer that's
481 : * still dirty, the old page needs to be written out first, and for that
482 : * you need WALWriteLock, and you need to ensure that there are no
483 : * in-progress insertions to the page by calling
484 : * WaitXLogInsertionsToFinish().
485 : */
486 : XLogRecPtr InitializedUpTo;
487 :
488 : /*
489 : * These values do not change after startup, although the pointed-to pages
490 : * and xlblocks values certainly do. xlblocks values are protected by
491 : * WALBufMappingLock.
492 : */
493 : char *pages; /* buffers for unwritten XLOG pages */
494 : pg_atomic_uint64 *xlblocks; /* 1st byte ptr-s + XLOG_BLCKSZ */
495 : int XLogCacheBlck; /* highest allocated xlog buffer index */
496 :
497 : /*
498 : * InsertTimeLineID is the timeline into which new WAL is being inserted
499 : * and flushed. It is zero during recovery, and does not change once set.
500 : *
501 : * If we create a new timeline when the system was started up,
502 : * PrevTimeLineID is the old timeline's ID that we forked off from.
503 : * Otherwise it's equal to InsertTimeLineID.
504 : *
505 : * We set these fields while holding info_lck. Most that reads these
506 : * values knows that recovery is no longer in progress and so can safely
507 : * read the value without a lock, but code that could be run either during
508 : * or after recovery can take info_lck while reading these values.
509 : */
510 : TimeLineID InsertTimeLineID;
511 : TimeLineID PrevTimeLineID;
512 :
513 : /*
514 : * SharedRecoveryState indicates if we're still in crash or archive
515 : * recovery. Protected by info_lck.
516 : */
517 : RecoveryState SharedRecoveryState;
518 :
519 : /*
520 : * InstallXLogFileSegmentActive indicates whether the checkpointer should
521 : * arrange for future segments by recycling and/or PreallocXlogFiles().
522 : * Protected by ControlFileLock. Only the startup process changes it. If
523 : * true, anyone can use InstallXLogFileSegment(). If false, the startup
524 : * process owns the exclusive right to install segments, by reading from
525 : * the archive and possibly replacing existing files.
526 : */
527 : bool InstallXLogFileSegmentActive;
528 :
529 : /*
530 : * WalWriterSleeping indicates whether the WAL writer is currently in
531 : * low-power mode (and hence should be nudged if an async commit occurs).
532 : * Protected by info_lck.
533 : */
534 : bool WalWriterSleeping;
535 :
536 : /*
537 : * During recovery, we keep a copy of the latest checkpoint record here.
538 : * lastCheckPointRecPtr points to start of checkpoint record and
539 : * lastCheckPointEndPtr points to end+1 of checkpoint record. Used by the
540 : * checkpointer when it wants to create a restartpoint.
541 : *
542 : * Protected by info_lck.
543 : */
544 : XLogRecPtr lastCheckPointRecPtr;
545 : XLogRecPtr lastCheckPointEndPtr;
546 : CheckPoint lastCheckPoint;
547 :
548 : /*
549 : * lastFpwDisableRecPtr points to the start of the last replayed
550 : * XLOG_FPW_CHANGE record that instructs full_page_writes is disabled.
551 : */
552 : XLogRecPtr lastFpwDisableRecPtr;
553 :
554 : slock_t info_lck; /* locks shared variables shown above */
555 : } XLogCtlData;
556 :
557 : /*
558 : * Classification of XLogInsertRecord operations.
559 : */
560 : typedef enum
561 : {
562 : WALINSERT_NORMAL,
563 : WALINSERT_SPECIAL_SWITCH,
564 : WALINSERT_SPECIAL_CHECKPOINT
565 : } WalInsertClass;
566 :
567 : static XLogCtlData *XLogCtl = NULL;
568 :
569 : /* a private copy of XLogCtl->Insert.WALInsertLocks, for convenience */
570 : static WALInsertLockPadded *WALInsertLocks = NULL;
571 :
572 : /*
573 : * We maintain an image of pg_control in shared memory.
574 : */
575 : static ControlFileData *ControlFile = NULL;
576 :
577 : /*
578 : * Calculate the amount of space left on the page after 'endptr'. Beware
579 : * multiple evaluation!
580 : */
581 : #define INSERT_FREESPACE(endptr) \
582 : (((endptr) % XLOG_BLCKSZ == 0) ? 0 : (XLOG_BLCKSZ - (endptr) % XLOG_BLCKSZ))
583 :
584 : /* Macro to advance to next buffer index. */
585 : #define NextBufIdx(idx) \
586 : (((idx) == XLogCtl->XLogCacheBlck) ? 0 : ((idx) + 1))
587 :
588 : /*
589 : * XLogRecPtrToBufIdx returns the index of the WAL buffer that holds, or
590 : * would hold if it was in cache, the page containing 'recptr'.
591 : */
592 : #define XLogRecPtrToBufIdx(recptr) \
593 : (((recptr) / XLOG_BLCKSZ) % (XLogCtl->XLogCacheBlck + 1))
594 :
595 : /*
596 : * These are the number of bytes in a WAL page usable for WAL data.
597 : */
598 : #define UsableBytesInPage (XLOG_BLCKSZ - SizeOfXLogShortPHD)
599 :
600 : /*
601 : * Convert values of GUCs measured in megabytes to equiv. segment count.
602 : * Rounds down.
603 : */
604 : #define ConvertToXSegs(x, segsize) XLogMBVarToSegs((x), (segsize))
605 :
606 : /* The number of bytes in a WAL segment usable for WAL data. */
607 : static int UsableBytesInSegment;
608 :
609 : /*
610 : * Private, possibly out-of-date copy of shared LogwrtResult.
611 : * See discussion above.
612 : */
613 : static XLogwrtResult LogwrtResult = {0, 0};
614 :
615 : /*
616 : * Update local copy of shared XLogCtl->log{Write,Flush}Result
617 : *
618 : * It's critical that Flush always trails Write, so the order of the reads is
619 : * important, as is the barrier. See also XLogWrite.
620 : */
621 : #define RefreshXLogWriteResult(_target) \
622 : do { \
623 : _target.Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult); \
624 : pg_read_barrier(); \
625 : _target.Write = pg_atomic_read_u64(&XLogCtl->logWriteResult); \
626 : } while (0)
627 :
628 : /*
629 : * openLogFile is -1 or a kernel FD for an open log file segment.
630 : * openLogSegNo identifies the segment, and openLogTLI the corresponding TLI.
631 : * These variables are only used to write the XLOG, and so will normally refer
632 : * to the active segment.
633 : *
634 : * Note: call Reserve/ReleaseExternalFD to track consumption of this FD.
635 : */
636 : static int openLogFile = -1;
637 : static XLogSegNo openLogSegNo = 0;
638 : static TimeLineID openLogTLI = 0;
639 :
640 : /*
641 : * Local copies of equivalent fields in the control file. When running
642 : * crash recovery, LocalMinRecoveryPoint is set to InvalidXLogRecPtr as we
643 : * expect to replay all the WAL available, and updateMinRecoveryPoint is
644 : * switched to false to prevent any updates while replaying records.
645 : * Those values are kept consistent as long as crash recovery runs.
646 : */
647 : static XLogRecPtr LocalMinRecoveryPoint;
648 : static TimeLineID LocalMinRecoveryPointTLI;
649 : static bool updateMinRecoveryPoint = true;
650 :
651 : /* For WALInsertLockAcquire/Release functions */
652 : static int MyLockNo = 0;
653 : static bool holdingAllLocks = false;
654 :
655 : #ifdef WAL_DEBUG
656 : static MemoryContext walDebugCxt = NULL;
657 : #endif
658 :
659 : static void CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI,
660 : XLogRecPtr EndOfLog,
661 : TimeLineID newTLI);
662 : static void CheckRequiredParameterValues(void);
663 : static void XLogReportParameters(void);
664 : static int LocalSetXLogInsertAllowed(void);
665 : static void CreateEndOfRecoveryRecord(void);
666 : static XLogRecPtr CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn,
667 : XLogRecPtr pagePtr,
668 : TimeLineID newTLI);
669 : static void CheckPointGuts(XLogRecPtr checkPointRedo, int flags);
670 : static void KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo);
671 : static XLogRecPtr XLogGetReplicationSlotMinimumLSN(void);
672 :
673 : static void AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli,
674 : bool opportunistic);
675 : static void XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible);
676 : static bool InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
677 : bool find_free, XLogSegNo max_segno,
678 : TimeLineID tli);
679 : static void XLogFileClose(void);
680 : static void PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli);
681 : static void RemoveTempXlogFiles(void);
682 : static void RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr,
683 : XLogRecPtr endptr, TimeLineID insertTLI);
684 : static void RemoveXlogFile(const struct dirent *segment_de,
685 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
686 : TimeLineID insertTLI);
687 : static void UpdateLastRemovedPtr(char *filename);
688 : static void ValidateXLOGDirectoryStructure(void);
689 : static void CleanupBackupHistory(void);
690 : static void UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force);
691 : static bool PerformRecoveryXLogAction(void);
692 : static void InitControlFile(uint64 sysidentifier, uint32 data_checksum_version);
693 : static void WriteControlFile(void);
694 : static void ReadControlFile(void);
695 : static void UpdateControlFile(void);
696 : static char *str_time(pg_time_t tnow, char *buf, size_t bufsize);
697 :
698 : static int get_sync_bit(int method);
699 :
700 : static void CopyXLogRecordToWAL(int write_len, bool isLogSwitch,
701 : XLogRecData *rdata,
702 : XLogRecPtr StartPos, XLogRecPtr EndPos,
703 : TimeLineID tli);
704 : static void ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos,
705 : XLogRecPtr *EndPos, XLogRecPtr *PrevPtr);
706 : static bool ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos,
707 : XLogRecPtr *PrevPtr);
708 : static XLogRecPtr WaitXLogInsertionsToFinish(XLogRecPtr upto);
709 : static char *GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli);
710 : static XLogRecPtr XLogBytePosToRecPtr(uint64 bytepos);
711 : static XLogRecPtr XLogBytePosToEndRecPtr(uint64 bytepos);
712 : static uint64 XLogRecPtrToBytePos(XLogRecPtr ptr);
713 :
714 : static void WALInsertLockAcquire(void);
715 : static void WALInsertLockAcquireExclusive(void);
716 : static void WALInsertLockRelease(void);
717 : static void WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt);
718 :
719 : /*
720 : * Insert an XLOG record represented by an already-constructed chain of data
721 : * chunks. This is a low-level routine; to construct the WAL record header
722 : * and data, use the higher-level routines in xloginsert.c.
723 : *
724 : * If 'fpw_lsn' is valid, it is the oldest LSN among the pages that this
725 : * WAL record applies to, that were not included in the record as full page
726 : * images. If fpw_lsn <= RedoRecPtr, the function does not perform the
727 : * insertion and returns InvalidXLogRecPtr. The caller can then recalculate
728 : * which pages need a full-page image, and retry. If fpw_lsn is invalid, the
729 : * record is always inserted.
730 : *
731 : * 'flags' gives more in-depth control on the record being inserted. See
732 : * XLogSetRecordFlags() for details.
733 : *
734 : * 'topxid_included' tells whether the top-transaction id is logged along with
735 : * current subtransaction. See XLogRecordAssemble().
736 : *
737 : * The first XLogRecData in the chain must be for the record header, and its
738 : * data must be MAXALIGNed. XLogInsertRecord fills in the xl_prev and
739 : * xl_crc fields in the header, the rest of the header must already be filled
740 : * by the caller.
741 : *
742 : * Returns XLOG pointer to end of record (beginning of next record).
743 : * This can be used as LSN for data pages affected by the logged action.
744 : * (LSN is the XLOG point up to which the XLOG must be flushed to disk
745 : * before the data page can be written out. This implements the basic
746 : * WAL rule "write the log before the data".)
747 : */
748 : XLogRecPtr
749 29257150 : XLogInsertRecord(XLogRecData *rdata,
750 : XLogRecPtr fpw_lsn,
751 : uint8 flags,
752 : int num_fpi,
753 : uint64 fpi_bytes,
754 : bool topxid_included)
755 : {
756 29257150 : XLogCtlInsert *Insert = &XLogCtl->Insert;
757 : pg_crc32c rdata_crc;
758 : bool inserted;
759 29257150 : XLogRecord *rechdr = (XLogRecord *) rdata->data;
760 29257150 : uint8 info = rechdr->xl_info & ~XLR_INFO_MASK;
761 29257150 : WalInsertClass class = WALINSERT_NORMAL;
762 : XLogRecPtr StartPos;
763 : XLogRecPtr EndPos;
764 29257150 : bool prevDoPageWrites = doPageWrites;
765 : TimeLineID insertTLI;
766 :
767 : /* Does this record type require special handling? */
768 29257150 : if (unlikely(rechdr->xl_rmid == RM_XLOG_ID))
769 : {
770 448266 : if (info == XLOG_SWITCH)
771 1568 : class = WALINSERT_SPECIAL_SWITCH;
772 446698 : else if (info == XLOG_CHECKPOINT_REDO)
773 1824 : class = WALINSERT_SPECIAL_CHECKPOINT;
774 : }
775 :
776 : /* we assume that all of the record header is in the first chunk */
777 : Assert(rdata->len >= SizeOfXLogRecord);
778 :
779 : /* cross-check on whether we should be here or not */
780 29257150 : if (!XLogInsertAllowed())
781 0 : elog(ERROR, "cannot make new WAL entries during recovery");
782 :
783 : /*
784 : * Given that we're not in recovery, InsertTimeLineID is set and can't
785 : * change, so we can read it without a lock.
786 : */
787 29257150 : insertTLI = XLogCtl->InsertTimeLineID;
788 :
789 : /*----------
790 : *
791 : * We have now done all the preparatory work we can without holding a
792 : * lock or modifying shared state. From here on, inserting the new WAL
793 : * record to the shared WAL buffer cache is a two-step process:
794 : *
795 : * 1. Reserve the right amount of space from the WAL. The current head of
796 : * reserved space is kept in Insert->CurrBytePos, and is protected by
797 : * insertpos_lck.
798 : *
799 : * 2. Copy the record to the reserved WAL space. This involves finding the
800 : * correct WAL buffer containing the reserved space, and copying the
801 : * record in place. This can be done concurrently in multiple processes.
802 : *
803 : * To keep track of which insertions are still in-progress, each concurrent
804 : * inserter acquires an insertion lock. In addition to just indicating that
805 : * an insertion is in progress, the lock tells others how far the inserter
806 : * has progressed. There is a small fixed number of insertion locks,
807 : * determined by NUM_XLOGINSERT_LOCKS. When an inserter crosses a page
808 : * boundary, it updates the value stored in the lock to the how far it has
809 : * inserted, to allow the previous buffer to be flushed.
810 : *
811 : * Holding onto an insertion lock also protects RedoRecPtr and
812 : * fullPageWrites from changing until the insertion is finished.
813 : *
814 : * Step 2 can usually be done completely in parallel. If the required WAL
815 : * page is not initialized yet, you have to grab WALBufMappingLock to
816 : * initialize it, but the WAL writer tries to do that ahead of insertions
817 : * to avoid that from happening in the critical path.
818 : *
819 : *----------
820 : */
821 29257150 : START_CRIT_SECTION();
822 :
823 29257150 : if (likely(class == WALINSERT_NORMAL))
824 : {
825 29253758 : WALInsertLockAcquire();
826 :
827 : /*
828 : * Check to see if my copy of RedoRecPtr is out of date. If so, may
829 : * have to go back and have the caller recompute everything. This can
830 : * only happen just after a checkpoint, so it's better to be slow in
831 : * this case and fast otherwise.
832 : *
833 : * Also check to see if fullPageWrites was just turned on or there's a
834 : * running backup (which forces full-page writes); if we weren't
835 : * already doing full-page writes then go back and recompute.
836 : *
837 : * If we aren't doing full-page writes then RedoRecPtr doesn't
838 : * actually affect the contents of the XLOG record, so we'll update
839 : * our local copy but not force a recomputation. (If doPageWrites was
840 : * just turned off, we could recompute the record without full pages,
841 : * but we choose not to bother.)
842 : */
843 29253758 : if (RedoRecPtr != Insert->RedoRecPtr)
844 : {
845 : Assert(RedoRecPtr < Insert->RedoRecPtr);
846 13922 : RedoRecPtr = Insert->RedoRecPtr;
847 : }
848 29253758 : doPageWrites = (Insert->fullPageWrites || Insert->runningBackups > 0);
849 :
850 29253758 : if (doPageWrites &&
851 28768736 : (!prevDoPageWrites ||
852 26402318 : (XLogRecPtrIsValid(fpw_lsn) && fpw_lsn <= RedoRecPtr)))
853 : {
854 : /*
855 : * Oops, some buffer now needs to be backed up that the caller
856 : * didn't back up. Start over.
857 : */
858 15452 : WALInsertLockRelease();
859 15452 : END_CRIT_SECTION();
860 15452 : return InvalidXLogRecPtr;
861 : }
862 :
863 : /*
864 : * Reserve space for the record in the WAL. This also sets the xl_prev
865 : * pointer.
866 : */
867 29238306 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
868 : &rechdr->xl_prev);
869 :
870 : /* Normal records are always inserted. */
871 29238306 : inserted = true;
872 : }
873 3392 : else if (class == WALINSERT_SPECIAL_SWITCH)
874 : {
875 : /*
876 : * In order to insert an XLOG_SWITCH record, we need to hold all of
877 : * the WAL insertion locks, not just one, so that no one else can
878 : * begin inserting a record until we've figured out how much space
879 : * remains in the current WAL segment and claimed all of it.
880 : *
881 : * Nonetheless, this case is simpler than the normal cases handled
882 : * below, which must check for changes in doPageWrites and RedoRecPtr.
883 : * Those checks are only needed for records that can contain buffer
884 : * references, and an XLOG_SWITCH record never does.
885 : */
886 : Assert(!XLogRecPtrIsValid(fpw_lsn));
887 1568 : WALInsertLockAcquireExclusive();
888 1568 : inserted = ReserveXLogSwitch(&StartPos, &EndPos, &rechdr->xl_prev);
889 : }
890 : else
891 : {
892 : Assert(class == WALINSERT_SPECIAL_CHECKPOINT);
893 :
894 : /*
895 : * We need to update both the local and shared copies of RedoRecPtr,
896 : * which means that we need to hold all the WAL insertion locks.
897 : * However, there can't be any buffer references, so as above, we need
898 : * not check RedoRecPtr before inserting the record; we just need to
899 : * update it afterwards.
900 : */
901 : Assert(!XLogRecPtrIsValid(fpw_lsn));
902 1824 : WALInsertLockAcquireExclusive();
903 1824 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
904 : &rechdr->xl_prev);
905 1824 : RedoRecPtr = Insert->RedoRecPtr = StartPos;
906 1824 : inserted = true;
907 : }
908 :
909 29241698 : if (inserted)
910 : {
911 : /*
912 : * Now that xl_prev has been filled in, calculate CRC of the record
913 : * header.
914 : */
915 29241582 : rdata_crc = rechdr->xl_crc;
916 29241582 : COMP_CRC32C(rdata_crc, rechdr, offsetof(XLogRecord, xl_crc));
917 29241582 : FIN_CRC32C(rdata_crc);
918 29241582 : rechdr->xl_crc = rdata_crc;
919 :
920 : /*
921 : * All the record data, including the header, is now ready to be
922 : * inserted. Copy the record in the space reserved.
923 : */
924 29241582 : CopyXLogRecordToWAL(rechdr->xl_tot_len,
925 : class == WALINSERT_SPECIAL_SWITCH, rdata,
926 : StartPos, EndPos, insertTLI);
927 :
928 : /*
929 : * Unless record is flagged as not important, update LSN of last
930 : * important record in the current slot. When holding all locks, just
931 : * update the first one.
932 : */
933 29241582 : if ((flags & XLOG_MARK_UNIMPORTANT) == 0)
934 : {
935 29058620 : int lockno = holdingAllLocks ? 0 : MyLockNo;
936 :
937 29058620 : WALInsertLocks[lockno].l.lastImportantAt = StartPos;
938 : }
939 : }
940 : else
941 : {
942 : /*
943 : * This was an xlog-switch record, but the current insert location was
944 : * already exactly at the beginning of a segment, so there was no need
945 : * to do anything.
946 : */
947 : }
948 :
949 : /*
950 : * Done! Let others know that we're finished.
951 : */
952 29241698 : WALInsertLockRelease();
953 :
954 29241698 : END_CRIT_SECTION();
955 :
956 29241698 : MarkCurrentTransactionIdLoggedIfAny();
957 :
958 : /*
959 : * Mark top transaction id is logged (if needed) so that we should not try
960 : * to log it again with the next WAL record in the current subtransaction.
961 : */
962 29241698 : if (topxid_included)
963 438 : MarkSubxactTopXidLogged();
964 :
965 : /*
966 : * Update shared LogwrtRqst.Write, if we crossed page boundary.
967 : */
968 29241698 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
969 : {
970 3333406 : SpinLockAcquire(&XLogCtl->info_lck);
971 : /* advance global request to include new block(s) */
972 3333406 : if (XLogCtl->LogwrtRqst.Write < EndPos)
973 3138178 : XLogCtl->LogwrtRqst.Write = EndPos;
974 3333406 : SpinLockRelease(&XLogCtl->info_lck);
975 3333406 : RefreshXLogWriteResult(LogwrtResult);
976 : }
977 :
978 : /*
979 : * If this was an XLOG_SWITCH record, flush the record and the empty
980 : * padding space that fills the rest of the segment, and perform
981 : * end-of-segment actions (eg, notifying archiver).
982 : */
983 29241698 : if (class == WALINSERT_SPECIAL_SWITCH)
984 : {
985 : TRACE_POSTGRESQL_WAL_SWITCH();
986 1568 : XLogFlush(EndPos);
987 :
988 : /*
989 : * Even though we reserved the rest of the segment for us, which is
990 : * reflected in EndPos, we return a pointer to just the end of the
991 : * xlog-switch record.
992 : */
993 1568 : if (inserted)
994 : {
995 1452 : EndPos = StartPos + SizeOfXLogRecord;
996 1452 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
997 : {
998 0 : uint64 offset = XLogSegmentOffset(EndPos, wal_segment_size);
999 :
1000 0 : if (offset == EndPos % XLOG_BLCKSZ)
1001 0 : EndPos += SizeOfXLogLongPHD;
1002 : else
1003 0 : EndPos += SizeOfXLogShortPHD;
1004 : }
1005 : }
1006 : }
1007 :
1008 : #ifdef WAL_DEBUG
1009 : if (XLOG_DEBUG)
1010 : {
1011 : static XLogReaderState *debug_reader = NULL;
1012 : XLogRecord *record;
1013 : DecodedXLogRecord *decoded;
1014 : StringInfoData buf;
1015 : StringInfoData recordBuf;
1016 : char *errormsg = NULL;
1017 : MemoryContext oldCxt;
1018 :
1019 : oldCxt = MemoryContextSwitchTo(walDebugCxt);
1020 :
1021 : initStringInfo(&buf);
1022 : appendStringInfo(&buf, "INSERT @ %X/%08X: ", LSN_FORMAT_ARGS(EndPos));
1023 :
1024 : /*
1025 : * We have to piece together the WAL record data from the XLogRecData
1026 : * entries, so that we can pass it to the rm_desc function as one
1027 : * contiguous chunk.
1028 : */
1029 : initStringInfo(&recordBuf);
1030 : for (; rdata != NULL; rdata = rdata->next)
1031 : appendBinaryStringInfo(&recordBuf, rdata->data, rdata->len);
1032 :
1033 : /* We also need temporary space to decode the record. */
1034 : record = (XLogRecord *) recordBuf.data;
1035 : decoded = (DecodedXLogRecord *)
1036 : palloc(DecodeXLogRecordRequiredSpace(record->xl_tot_len));
1037 :
1038 : if (!debug_reader)
1039 : debug_reader = XLogReaderAllocate(wal_segment_size, NULL,
1040 : XL_ROUTINE(.page_read = NULL,
1041 : .segment_open = NULL,
1042 : .segment_close = NULL),
1043 : NULL);
1044 : if (!debug_reader)
1045 : {
1046 : appendStringInfoString(&buf, "error decoding record: out of memory while allocating a WAL reading processor");
1047 : }
1048 : else if (!DecodeXLogRecord(debug_reader,
1049 : decoded,
1050 : record,
1051 : EndPos,
1052 : &errormsg))
1053 : {
1054 : appendStringInfo(&buf, "error decoding record: %s",
1055 : errormsg ? errormsg : "no error message");
1056 : }
1057 : else
1058 : {
1059 : appendStringInfoString(&buf, " - ");
1060 :
1061 : debug_reader->record = decoded;
1062 : xlog_outdesc(&buf, debug_reader);
1063 : debug_reader->record = NULL;
1064 : }
1065 : elog(LOG, "%s", buf.data);
1066 :
1067 : pfree(decoded);
1068 : pfree(buf.data);
1069 : pfree(recordBuf.data);
1070 : MemoryContextSwitchTo(oldCxt);
1071 : }
1072 : #endif
1073 :
1074 : /*
1075 : * Update our global variables
1076 : */
1077 29241698 : ProcLastRecPtr = StartPos;
1078 29241698 : XactLastRecEnd = EndPos;
1079 :
1080 : /* Report WAL traffic to the instrumentation. */
1081 29241698 : if (inserted)
1082 : {
1083 29241582 : pgWalUsage.wal_bytes += rechdr->xl_tot_len;
1084 29241582 : pgWalUsage.wal_records++;
1085 29241582 : pgWalUsage.wal_fpi += num_fpi;
1086 29241582 : pgWalUsage.wal_fpi_bytes += fpi_bytes;
1087 :
1088 : /* Required for the flush of pending stats WAL data */
1089 29241582 : pgstat_report_fixed = true;
1090 : }
1091 :
1092 29241698 : return EndPos;
1093 : }
1094 :
1095 : /*
1096 : * Reserves the right amount of space for a record of given size from the WAL.
1097 : * *StartPos is set to the beginning of the reserved section, *EndPos to
1098 : * its end+1. *PrevPtr is set to the beginning of the previous record; it is
1099 : * used to set the xl_prev of this record.
1100 : *
1101 : * This is the performance critical part of XLogInsert that must be serialized
1102 : * across backends. The rest can happen mostly in parallel. Try to keep this
1103 : * section as short as possible, insertpos_lck can be heavily contended on a
1104 : * busy system.
1105 : *
1106 : * NB: The space calculation here must match the code in CopyXLogRecordToWAL,
1107 : * where we actually copy the record to the reserved space.
1108 : *
1109 : * NB: Testing shows that XLogInsertRecord runs faster if this code is inlined;
1110 : * however, because there are two call sites, the compiler is reluctant to
1111 : * inline. We use pg_attribute_always_inline here to try to convince it.
1112 : */
1113 : static pg_attribute_always_inline void
1114 29240130 : ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1115 : XLogRecPtr *PrevPtr)
1116 : {
1117 29240130 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1118 : uint64 startbytepos;
1119 : uint64 endbytepos;
1120 : uint64 prevbytepos;
1121 :
1122 29240130 : size = MAXALIGN(size);
1123 :
1124 : /* All (non xlog-switch) records should contain data. */
1125 : Assert(size > SizeOfXLogRecord);
1126 :
1127 : /*
1128 : * The duration the spinlock needs to be held is minimized by minimizing
1129 : * the calculations that have to be done while holding the lock. The
1130 : * current tip of reserved WAL is kept in CurrBytePos, as a byte position
1131 : * that only counts "usable" bytes in WAL, that is, it excludes all WAL
1132 : * page headers. The mapping between "usable" byte positions and physical
1133 : * positions (XLogRecPtrs) can be done outside the locked region, and
1134 : * because the usable byte position doesn't include any headers, reserving
1135 : * X bytes from WAL is almost as simple as "CurrBytePos += X".
1136 : */
1137 29240130 : SpinLockAcquire(&Insert->insertpos_lck);
1138 :
1139 29240130 : startbytepos = Insert->CurrBytePos;
1140 29240130 : endbytepos = startbytepos + size;
1141 29240130 : prevbytepos = Insert->PrevBytePos;
1142 29240130 : Insert->CurrBytePos = endbytepos;
1143 29240130 : Insert->PrevBytePos = startbytepos;
1144 :
1145 29240130 : SpinLockRelease(&Insert->insertpos_lck);
1146 :
1147 29240130 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1148 29240130 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1149 29240130 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1150 :
1151 : /*
1152 : * Check that the conversions between "usable byte positions" and
1153 : * XLogRecPtrs work consistently in both directions.
1154 : */
1155 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1156 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1157 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1158 29240130 : }
1159 :
1160 : /*
1161 : * Like ReserveXLogInsertLocation(), but for an xlog-switch record.
1162 : *
1163 : * A log-switch record is handled slightly differently. The rest of the
1164 : * segment will be reserved for this insertion, as indicated by the returned
1165 : * *EndPos value. However, if we are already at the beginning of the current
1166 : * segment, *StartPos and *EndPos are set to the current location without
1167 : * reserving any space, and the function returns false.
1168 : */
1169 : static bool
1170 1568 : ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1171 : {
1172 1568 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1173 : uint64 startbytepos;
1174 : uint64 endbytepos;
1175 : uint64 prevbytepos;
1176 1568 : uint32 size = MAXALIGN(SizeOfXLogRecord);
1177 : XLogRecPtr ptr;
1178 : uint32 segleft;
1179 :
1180 : /*
1181 : * These calculations are a bit heavy-weight to be done while holding a
1182 : * spinlock, but since we're holding all the WAL insertion locks, there
1183 : * are no other inserters competing for it. GetXLogInsertRecPtr() does
1184 : * compete for it, but that's not called very frequently.
1185 : */
1186 1568 : SpinLockAcquire(&Insert->insertpos_lck);
1187 :
1188 1568 : startbytepos = Insert->CurrBytePos;
1189 :
1190 1568 : ptr = XLogBytePosToEndRecPtr(startbytepos);
1191 1568 : if (XLogSegmentOffset(ptr, wal_segment_size) == 0)
1192 : {
1193 116 : SpinLockRelease(&Insert->insertpos_lck);
1194 116 : *EndPos = *StartPos = ptr;
1195 116 : return false;
1196 : }
1197 :
1198 1452 : endbytepos = startbytepos + size;
1199 1452 : prevbytepos = Insert->PrevBytePos;
1200 :
1201 1452 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1202 1452 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1203 :
1204 1452 : segleft = wal_segment_size - XLogSegmentOffset(*EndPos, wal_segment_size);
1205 1452 : if (segleft != wal_segment_size)
1206 : {
1207 : /* consume the rest of the segment */
1208 1452 : *EndPos += segleft;
1209 1452 : endbytepos = XLogRecPtrToBytePos(*EndPos);
1210 : }
1211 1452 : Insert->CurrBytePos = endbytepos;
1212 1452 : Insert->PrevBytePos = startbytepos;
1213 :
1214 1452 : SpinLockRelease(&Insert->insertpos_lck);
1215 :
1216 1452 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1217 :
1218 : Assert(XLogSegmentOffset(*EndPos, wal_segment_size) == 0);
1219 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1220 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1221 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1222 :
1223 1452 : return true;
1224 : }
1225 :
1226 : /*
1227 : * Subroutine of XLogInsertRecord. Copies a WAL record to an already-reserved
1228 : * area in the WAL.
1229 : */
1230 : static void
1231 29241582 : CopyXLogRecordToWAL(int write_len, bool isLogSwitch, XLogRecData *rdata,
1232 : XLogRecPtr StartPos, XLogRecPtr EndPos, TimeLineID tli)
1233 : {
1234 : char *currpos;
1235 : int freespace;
1236 : int written;
1237 : XLogRecPtr CurrPos;
1238 : XLogPageHeader pagehdr;
1239 :
1240 : /*
1241 : * Get a pointer to the right place in the right WAL buffer to start
1242 : * inserting to.
1243 : */
1244 29241582 : CurrPos = StartPos;
1245 29241582 : currpos = GetXLogBuffer(CurrPos, tli);
1246 29241582 : freespace = INSERT_FREESPACE(CurrPos);
1247 :
1248 : /*
1249 : * there should be enough space for at least the first field (xl_tot_len)
1250 : * on this page.
1251 : */
1252 : Assert(freespace >= sizeof(uint32));
1253 :
1254 : /* Copy record data */
1255 29241582 : written = 0;
1256 138162166 : while (rdata != NULL)
1257 : {
1258 108920584 : const char *rdata_data = rdata->data;
1259 108920584 : int rdata_len = rdata->len;
1260 :
1261 112484924 : while (rdata_len > freespace)
1262 : {
1263 : /*
1264 : * Write what fits on this page, and continue on the next page.
1265 : */
1266 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || freespace == 0);
1267 3564340 : memcpy(currpos, rdata_data, freespace);
1268 3564340 : rdata_data += freespace;
1269 3564340 : rdata_len -= freespace;
1270 3564340 : written += freespace;
1271 3564340 : CurrPos += freespace;
1272 :
1273 : /*
1274 : * Get pointer to beginning of next page, and set the xlp_rem_len
1275 : * in the page header. Set XLP_FIRST_IS_CONTRECORD.
1276 : *
1277 : * It's safe to set the contrecord flag and xlp_rem_len without a
1278 : * lock on the page. All the other flags were already set when the
1279 : * page was initialized, in AdvanceXLInsertBuffer, and we're the
1280 : * only backend that needs to set the contrecord flag.
1281 : */
1282 3564340 : currpos = GetXLogBuffer(CurrPos, tli);
1283 3564340 : pagehdr = (XLogPageHeader) currpos;
1284 3564340 : pagehdr->xlp_rem_len = write_len - written;
1285 3564340 : pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1286 :
1287 : /* skip over the page header */
1288 3564340 : if (XLogSegmentOffset(CurrPos, wal_segment_size) == 0)
1289 : {
1290 2364 : CurrPos += SizeOfXLogLongPHD;
1291 2364 : currpos += SizeOfXLogLongPHD;
1292 : }
1293 : else
1294 : {
1295 3561976 : CurrPos += SizeOfXLogShortPHD;
1296 3561976 : currpos += SizeOfXLogShortPHD;
1297 : }
1298 3564340 : freespace = INSERT_FREESPACE(CurrPos);
1299 : }
1300 :
1301 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1302 108920584 : memcpy(currpos, rdata_data, rdata_len);
1303 108920584 : currpos += rdata_len;
1304 108920584 : CurrPos += rdata_len;
1305 108920584 : freespace -= rdata_len;
1306 108920584 : written += rdata_len;
1307 :
1308 108920584 : rdata = rdata->next;
1309 : }
1310 : Assert(written == write_len);
1311 :
1312 : /*
1313 : * If this was an xlog-switch, it's not enough to write the switch record,
1314 : * we also have to consume all the remaining space in the WAL segment. We
1315 : * have already reserved that space, but we need to actually fill it.
1316 : */
1317 29241582 : if (isLogSwitch && XLogSegmentOffset(CurrPos, wal_segment_size) != 0)
1318 : {
1319 : /* An xlog-switch record doesn't contain any data besides the header */
1320 : Assert(write_len == SizeOfXLogRecord);
1321 :
1322 : /* Assert that we did reserve the right amount of space */
1323 : Assert(XLogSegmentOffset(EndPos, wal_segment_size) == 0);
1324 :
1325 : /* Use up all the remaining space on the current page */
1326 1452 : CurrPos += freespace;
1327 :
1328 : /*
1329 : * Cause all remaining pages in the segment to be flushed, leaving the
1330 : * XLog position where it should be, at the start of the next segment.
1331 : * We do this one page at a time, to make sure we don't deadlock
1332 : * against ourselves if wal_buffers < wal_segment_size.
1333 : */
1334 1400114 : while (CurrPos < EndPos)
1335 : {
1336 : /*
1337 : * The minimal action to flush the page would be to call
1338 : * WALInsertLockUpdateInsertingAt(CurrPos) followed by
1339 : * AdvanceXLInsertBuffer(...). The page would be left initialized
1340 : * mostly to zeros, except for the page header (always the short
1341 : * variant, as this is never a segment's first page).
1342 : *
1343 : * The large vistas of zeros are good for compressibility, but the
1344 : * headers interrupting them every XLOG_BLCKSZ (with values that
1345 : * differ from page to page) are not. The effect varies with
1346 : * compression tool, but bzip2 for instance compresses about an
1347 : * order of magnitude worse if those headers are left in place.
1348 : *
1349 : * Rather than complicating AdvanceXLInsertBuffer itself (which is
1350 : * called in heavily-loaded circumstances as well as this lightly-
1351 : * loaded one) with variant behavior, we just use GetXLogBuffer
1352 : * (which itself calls the two methods we need) to get the pointer
1353 : * and zero most of the page. Then we just zero the page header.
1354 : */
1355 1398662 : currpos = GetXLogBuffer(CurrPos, tli);
1356 5594648 : MemSet(currpos, 0, SizeOfXLogShortPHD);
1357 :
1358 1398662 : CurrPos += XLOG_BLCKSZ;
1359 : }
1360 : }
1361 : else
1362 : {
1363 : /* Align the end position, so that the next record starts aligned */
1364 29240130 : CurrPos = MAXALIGN64(CurrPos);
1365 : }
1366 :
1367 29241582 : if (CurrPos != EndPos)
1368 0 : ereport(PANIC,
1369 : errcode(ERRCODE_DATA_CORRUPTED),
1370 : errmsg_internal("space reserved for WAL record does not match what was written"));
1371 29241582 : }
1372 :
1373 : /*
1374 : * Acquire a WAL insertion lock, for inserting to WAL.
1375 : */
1376 : static void
1377 29253778 : WALInsertLockAcquire(void)
1378 : {
1379 : bool immed;
1380 :
1381 : /*
1382 : * It doesn't matter which of the WAL insertion locks we acquire, so try
1383 : * the one we used last time. If the system isn't particularly busy, it's
1384 : * a good bet that it's still available, and it's good to have some
1385 : * affinity to a particular lock so that you don't unnecessarily bounce
1386 : * cache lines between processes when there's no contention.
1387 : *
1388 : * If this is the first time through in this backend, pick a lock
1389 : * (semi-)randomly. This allows the locks to be used evenly if you have a
1390 : * lot of very short connections.
1391 : */
1392 : static int lockToTry = -1;
1393 :
1394 29253778 : if (lockToTry == -1)
1395 16186 : lockToTry = MyProcNumber % NUM_XLOGINSERT_LOCKS;
1396 29253778 : MyLockNo = lockToTry;
1397 :
1398 : /*
1399 : * The insertingAt value is initially set to 0, as we don't know our
1400 : * insert location yet.
1401 : */
1402 29253778 : immed = LWLockAcquire(&WALInsertLocks[MyLockNo].l.lock, LW_EXCLUSIVE);
1403 29253778 : if (!immed)
1404 : {
1405 : /*
1406 : * If we couldn't get the lock immediately, try another lock next
1407 : * time. On a system with more insertion locks than concurrent
1408 : * inserters, this causes all the inserters to eventually migrate to a
1409 : * lock that no-one else is using. On a system with more inserters
1410 : * than locks, it still helps to distribute the inserters evenly
1411 : * across the locks.
1412 : */
1413 51538 : lockToTry = (lockToTry + 1) % NUM_XLOGINSERT_LOCKS;
1414 : }
1415 29253778 : }
1416 :
1417 : /*
1418 : * Acquire all WAL insertion locks, to prevent other backends from inserting
1419 : * to WAL.
1420 : */
1421 : static void
1422 8686 : WALInsertLockAcquireExclusive(void)
1423 : {
1424 : int i;
1425 :
1426 : /*
1427 : * When holding all the locks, all but the last lock's insertingAt
1428 : * indicator is set to 0xFFFFFFFFFFFFFFFF, which is higher than any real
1429 : * XLogRecPtr value, to make sure that no-one blocks waiting on those.
1430 : */
1431 69488 : for (i = 0; i < NUM_XLOGINSERT_LOCKS - 1; i++)
1432 : {
1433 60802 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1434 60802 : LWLockUpdateVar(&WALInsertLocks[i].l.lock,
1435 60802 : &WALInsertLocks[i].l.insertingAt,
1436 : PG_UINT64_MAX);
1437 : }
1438 : /* Variable value reset to 0 at release */
1439 8686 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1440 :
1441 8686 : holdingAllLocks = true;
1442 8686 : }
1443 :
1444 : /*
1445 : * Release our insertion lock (or locks, if we're holding them all).
1446 : *
1447 : * NB: Reset all variables to 0, so they cause LWLockWaitForVar to block the
1448 : * next time the lock is acquired.
1449 : */
1450 : static void
1451 29262464 : WALInsertLockRelease(void)
1452 : {
1453 29262464 : if (holdingAllLocks)
1454 : {
1455 : int i;
1456 :
1457 78174 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1458 69488 : LWLockReleaseClearVar(&WALInsertLocks[i].l.lock,
1459 69488 : &WALInsertLocks[i].l.insertingAt,
1460 : 0);
1461 :
1462 8686 : holdingAllLocks = false;
1463 : }
1464 : else
1465 : {
1466 29253778 : LWLockReleaseClearVar(&WALInsertLocks[MyLockNo].l.lock,
1467 29253778 : &WALInsertLocks[MyLockNo].l.insertingAt,
1468 : 0);
1469 : }
1470 29262464 : }
1471 :
1472 : /*
1473 : * Update our insertingAt value, to let others know that we've finished
1474 : * inserting up to that point.
1475 : */
1476 : static void
1477 5184442 : WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt)
1478 : {
1479 5184442 : if (holdingAllLocks)
1480 : {
1481 : /*
1482 : * We use the last lock to mark our actual position, see comments in
1483 : * WALInsertLockAcquireExclusive.
1484 : */
1485 1383604 : LWLockUpdateVar(&WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.lock,
1486 1383604 : &WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.insertingAt,
1487 : insertingAt);
1488 : }
1489 : else
1490 3800838 : LWLockUpdateVar(&WALInsertLocks[MyLockNo].l.lock,
1491 3800838 : &WALInsertLocks[MyLockNo].l.insertingAt,
1492 : insertingAt);
1493 5184442 : }
1494 :
1495 : /*
1496 : * Wait for any WAL insertions < upto to finish.
1497 : *
1498 : * Returns the location of the oldest insertion that is still in-progress.
1499 : * Any WAL prior to that point has been fully copied into WAL buffers, and
1500 : * can be flushed out to disk. Because this waits for any insertions older
1501 : * than 'upto' to finish, the return value is always >= 'upto'.
1502 : *
1503 : * Note: When you are about to write out WAL, you must call this function
1504 : * *before* acquiring WALWriteLock, to avoid deadlocks. This function might
1505 : * need to wait for an insertion to finish (or at least advance to next
1506 : * uninitialized page), and the inserter might need to evict an old WAL buffer
1507 : * to make room for a new one, which in turn requires WALWriteLock.
1508 : */
1509 : static XLogRecPtr
1510 4789948 : WaitXLogInsertionsToFinish(XLogRecPtr upto)
1511 : {
1512 : uint64 bytepos;
1513 : XLogRecPtr inserted;
1514 : XLogRecPtr reservedUpto;
1515 : XLogRecPtr finishedUpto;
1516 4789948 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1517 : int i;
1518 :
1519 4789948 : if (MyProc == NULL)
1520 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1521 :
1522 : /*
1523 : * Check if there's any work to do. Use a barrier to ensure we get the
1524 : * freshest value.
1525 : */
1526 4789948 : inserted = pg_atomic_read_membarrier_u64(&XLogCtl->logInsertResult);
1527 4789948 : if (upto <= inserted)
1528 3867270 : return inserted;
1529 :
1530 : /* Read the current insert position */
1531 922678 : SpinLockAcquire(&Insert->insertpos_lck);
1532 922678 : bytepos = Insert->CurrBytePos;
1533 922678 : SpinLockRelease(&Insert->insertpos_lck);
1534 922678 : reservedUpto = XLogBytePosToEndRecPtr(bytepos);
1535 :
1536 : /*
1537 : * No-one should request to flush a piece of WAL that hasn't even been
1538 : * reserved yet. However, it can happen if there is a block with a bogus
1539 : * LSN on disk, for example. XLogFlush checks for that situation and
1540 : * complains, but only after the flush. Here we just assume that to mean
1541 : * that all WAL that has been reserved needs to be finished. In this
1542 : * corner-case, the return value can be smaller than 'upto' argument.
1543 : */
1544 922678 : if (upto > reservedUpto)
1545 : {
1546 0 : ereport(LOG,
1547 : errmsg("request to flush past end of generated WAL; request %X/%08X, current position %X/%08X",
1548 : LSN_FORMAT_ARGS(upto), LSN_FORMAT_ARGS(reservedUpto)));
1549 0 : upto = reservedUpto;
1550 : }
1551 :
1552 : /*
1553 : * Loop through all the locks, sleeping on any in-progress insert older
1554 : * than 'upto'.
1555 : *
1556 : * finishedUpto is our return value, indicating the point upto which all
1557 : * the WAL insertions have been finished. Initialize it to the head of
1558 : * reserved WAL, and as we iterate through the insertion locks, back it
1559 : * out for any insertion that's still in progress.
1560 : */
1561 922678 : finishedUpto = reservedUpto;
1562 8304102 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1563 : {
1564 7381424 : XLogRecPtr insertingat = InvalidXLogRecPtr;
1565 :
1566 : do
1567 : {
1568 : /*
1569 : * See if this insertion is in progress. LWLockWaitForVar will
1570 : * wait for the lock to be released, or for the 'value' to be set
1571 : * by a LWLockUpdateVar call. When a lock is initially acquired,
1572 : * its value is 0 (InvalidXLogRecPtr), which means that we don't
1573 : * know where it's inserting yet. We will have to wait for it. If
1574 : * it's a small insertion, the record will most likely fit on the
1575 : * same page and the inserter will release the lock without ever
1576 : * calling LWLockUpdateVar. But if it has to sleep, it will
1577 : * advertise the insertion point with LWLockUpdateVar before
1578 : * sleeping.
1579 : *
1580 : * In this loop we are only waiting for insertions that started
1581 : * before WaitXLogInsertionsToFinish was called. The lack of
1582 : * memory barriers in the loop means that we might see locks as
1583 : * "unused" that have since become used. This is fine because
1584 : * they only can be used for later insertions that we would not
1585 : * want to wait on anyway. Not taking a lock to acquire the
1586 : * current insertingAt value means that we might see older
1587 : * insertingAt values. This is also fine, because if we read a
1588 : * value too old, we will add ourselves to the wait queue, which
1589 : * contains atomic operations.
1590 : */
1591 7604700 : if (LWLockWaitForVar(&WALInsertLocks[i].l.lock,
1592 7604700 : &WALInsertLocks[i].l.insertingAt,
1593 : insertingat, &insertingat))
1594 : {
1595 : /* the lock was free, so no insertion in progress */
1596 5218774 : insertingat = InvalidXLogRecPtr;
1597 5218774 : break;
1598 : }
1599 :
1600 : /*
1601 : * This insertion is still in progress. Have to wait, unless the
1602 : * inserter has proceeded past 'upto'.
1603 : */
1604 2385926 : } while (insertingat < upto);
1605 :
1606 7381424 : if (XLogRecPtrIsValid(insertingat) && insertingat < finishedUpto)
1607 839822 : finishedUpto = insertingat;
1608 : }
1609 :
1610 : /*
1611 : * Advance the limit we know to have been inserted and return the freshest
1612 : * value we know of, which might be beyond what we requested if somebody
1613 : * is concurrently doing this with an 'upto' pointer ahead of us.
1614 : */
1615 922678 : finishedUpto = pg_atomic_monotonic_advance_u64(&XLogCtl->logInsertResult,
1616 : finishedUpto);
1617 :
1618 922678 : return finishedUpto;
1619 : }
1620 :
1621 : /*
1622 : * Get a pointer to the right location in the WAL buffer containing the
1623 : * given XLogRecPtr.
1624 : *
1625 : * If the page is not initialized yet, it is initialized. That might require
1626 : * evicting an old dirty buffer from the buffer cache, which means I/O.
1627 : *
1628 : * The caller must ensure that the page containing the requested location
1629 : * isn't evicted yet, and won't be evicted. The way to ensure that is to
1630 : * hold onto a WAL insertion lock with the insertingAt position set to
1631 : * something <= ptr. GetXLogBuffer() will update insertingAt if it needs
1632 : * to evict an old page from the buffer. (This means that once you call
1633 : * GetXLogBuffer() with a given 'ptr', you must not access anything before
1634 : * that point anymore, and must not call GetXLogBuffer() with an older 'ptr'
1635 : * later, because older buffers might be recycled already)
1636 : */
1637 : static char *
1638 34204604 : GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli)
1639 : {
1640 : int idx;
1641 : XLogRecPtr endptr;
1642 : static uint64 cachedPage = 0;
1643 : static char *cachedPos = NULL;
1644 : XLogRecPtr expectedEndPtr;
1645 :
1646 : /*
1647 : * Fast path for the common case that we need to access again the same
1648 : * page as last time.
1649 : */
1650 34204604 : if (ptr / XLOG_BLCKSZ == cachedPage)
1651 : {
1652 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1653 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1654 28016434 : return cachedPos + ptr % XLOG_BLCKSZ;
1655 : }
1656 :
1657 : /*
1658 : * The XLog buffer cache is organized so that a page is always loaded to a
1659 : * particular buffer. That way we can easily calculate the buffer a given
1660 : * page must be loaded into, from the XLogRecPtr alone.
1661 : */
1662 6188170 : idx = XLogRecPtrToBufIdx(ptr);
1663 :
1664 : /*
1665 : * See what page is loaded in the buffer at the moment. It could be the
1666 : * page we're looking for, or something older. It can't be anything newer
1667 : * - that would imply the page we're looking for has already been written
1668 : * out to disk and evicted, and the caller is responsible for making sure
1669 : * that doesn't happen.
1670 : *
1671 : * We don't hold a lock while we read the value. If someone is just about
1672 : * to initialize or has just initialized the page, it's possible that we
1673 : * get InvalidXLogRecPtr. That's ok, we'll grab the mapping lock (in
1674 : * AdvanceXLInsertBuffer) and retry if we see anything other than the page
1675 : * we're looking for.
1676 : */
1677 6188170 : expectedEndPtr = ptr;
1678 6188170 : expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
1679 :
1680 6188170 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1681 6188170 : if (expectedEndPtr != endptr)
1682 : {
1683 : XLogRecPtr initializedUpto;
1684 :
1685 : /*
1686 : * Before calling AdvanceXLInsertBuffer(), which can block, let others
1687 : * know how far we're finished with inserting the record.
1688 : *
1689 : * NB: If 'ptr' points to just after the page header, advertise a
1690 : * position at the beginning of the page rather than 'ptr' itself. If
1691 : * there are no other insertions running, someone might try to flush
1692 : * up to our advertised location. If we advertised a position after
1693 : * the page header, someone might try to flush the page header, even
1694 : * though page might actually not be initialized yet. As the first
1695 : * inserter on the page, we are effectively responsible for making
1696 : * sure that it's initialized, before we let insertingAt to move past
1697 : * the page header.
1698 : */
1699 5184442 : if (ptr % XLOG_BLCKSZ == SizeOfXLogShortPHD &&
1700 12668 : XLogSegmentOffset(ptr, wal_segment_size) > XLOG_BLCKSZ)
1701 12668 : initializedUpto = ptr - SizeOfXLogShortPHD;
1702 5171774 : else if (ptr % XLOG_BLCKSZ == SizeOfXLogLongPHD &&
1703 2384 : XLogSegmentOffset(ptr, wal_segment_size) < XLOG_BLCKSZ)
1704 1200 : initializedUpto = ptr - SizeOfXLogLongPHD;
1705 : else
1706 5170574 : initializedUpto = ptr;
1707 :
1708 5184442 : WALInsertLockUpdateInsertingAt(initializedUpto);
1709 :
1710 5184442 : AdvanceXLInsertBuffer(ptr, tli, false);
1711 5184442 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1712 :
1713 5184442 : if (expectedEndPtr != endptr)
1714 0 : elog(PANIC, "could not find WAL buffer for %X/%08X",
1715 : LSN_FORMAT_ARGS(ptr));
1716 : }
1717 : else
1718 : {
1719 : /*
1720 : * Make sure the initialization of the page is visible to us, and
1721 : * won't arrive later to overwrite the WAL data we write on the page.
1722 : */
1723 1003728 : pg_memory_barrier();
1724 : }
1725 :
1726 : /*
1727 : * Found the buffer holding this page. Return a pointer to the right
1728 : * offset within the page.
1729 : */
1730 6188170 : cachedPage = ptr / XLOG_BLCKSZ;
1731 6188170 : cachedPos = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1732 :
1733 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1734 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1735 :
1736 6188170 : return cachedPos + ptr % XLOG_BLCKSZ;
1737 : }
1738 :
1739 : /*
1740 : * Read WAL data directly from WAL buffers, if available. Returns the number
1741 : * of bytes read successfully.
1742 : *
1743 : * Fewer than 'count' bytes may be read if some of the requested WAL data has
1744 : * already been evicted.
1745 : *
1746 : * No locks are taken.
1747 : *
1748 : * Caller should ensure that it reads no further than LogwrtResult.Write
1749 : * (which should have been updated by the caller when determining how far to
1750 : * read). The 'tli' argument is only used as a convenient safety check so that
1751 : * callers do not read from WAL buffers on a historical timeline.
1752 : */
1753 : Size
1754 197904 : WALReadFromBuffers(char *dstbuf, XLogRecPtr startptr, Size count,
1755 : TimeLineID tli)
1756 : {
1757 197904 : char *pdst = dstbuf;
1758 197904 : XLogRecPtr recptr = startptr;
1759 : XLogRecPtr inserted;
1760 197904 : Size nbytes = count;
1761 :
1762 197904 : if (RecoveryInProgress() || tli != GetWALInsertionTimeLine())
1763 1524 : return 0;
1764 :
1765 : Assert(XLogRecPtrIsValid(startptr));
1766 :
1767 : /*
1768 : * Caller should ensure that the requested data has been inserted into WAL
1769 : * buffers before we try to read it.
1770 : */
1771 196380 : inserted = pg_atomic_read_u64(&XLogCtl->logInsertResult);
1772 196380 : if (startptr + count > inserted)
1773 0 : ereport(ERROR,
1774 : errmsg("cannot read past end of generated WAL: requested %X/%08X, current position %X/%08X",
1775 : LSN_FORMAT_ARGS(startptr + count),
1776 : LSN_FORMAT_ARGS(inserted)));
1777 :
1778 : /*
1779 : * Loop through the buffers without a lock. For each buffer, atomically
1780 : * read and verify the end pointer, then copy the data out, and finally
1781 : * re-read and re-verify the end pointer.
1782 : *
1783 : * Once a page is evicted, it never returns to the WAL buffers, so if the
1784 : * end pointer matches the expected end pointer before and after we copy
1785 : * the data, then the right page must have been present during the data
1786 : * copy. Read barriers are necessary to ensure that the data copy actually
1787 : * happens between the two verification steps.
1788 : *
1789 : * If either verification fails, we simply terminate the loop and return
1790 : * with the data that had been already copied out successfully.
1791 : */
1792 211976 : while (nbytes > 0)
1793 : {
1794 208052 : uint32 offset = recptr % XLOG_BLCKSZ;
1795 208052 : int idx = XLogRecPtrToBufIdx(recptr);
1796 : XLogRecPtr expectedEndPtr;
1797 : XLogRecPtr endptr;
1798 : const char *page;
1799 : const char *psrc;
1800 : Size npagebytes;
1801 :
1802 : /*
1803 : * Calculate the end pointer we expect in the xlblocks array if the
1804 : * correct page is present.
1805 : */
1806 208052 : expectedEndPtr = recptr + (XLOG_BLCKSZ - offset);
1807 :
1808 : /*
1809 : * First verification step: check that the correct page is present in
1810 : * the WAL buffers.
1811 : */
1812 208052 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1813 208052 : if (expectedEndPtr != endptr)
1814 192456 : break;
1815 :
1816 : /*
1817 : * The correct page is present (or was at the time the endptr was
1818 : * read; must re-verify later). Calculate pointer to source data and
1819 : * determine how much data to read from this page.
1820 : */
1821 15596 : page = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1822 15596 : psrc = page + offset;
1823 15596 : npagebytes = Min(nbytes, XLOG_BLCKSZ - offset);
1824 :
1825 : /*
1826 : * Ensure that the data copy and the first verification step are not
1827 : * reordered.
1828 : */
1829 15596 : pg_read_barrier();
1830 :
1831 : /* data copy */
1832 15596 : memcpy(pdst, psrc, npagebytes);
1833 :
1834 : /*
1835 : * Ensure that the data copy and the second verification step are not
1836 : * reordered.
1837 : */
1838 15596 : pg_read_barrier();
1839 :
1840 : /*
1841 : * Second verification step: check that the page we read from wasn't
1842 : * evicted while we were copying the data.
1843 : */
1844 15596 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1845 15596 : if (expectedEndPtr != endptr)
1846 0 : break;
1847 :
1848 15596 : pdst += npagebytes;
1849 15596 : recptr += npagebytes;
1850 15596 : nbytes -= npagebytes;
1851 : }
1852 :
1853 : Assert(pdst - dstbuf <= count);
1854 :
1855 196380 : return pdst - dstbuf;
1856 : }
1857 :
1858 : /*
1859 : * Converts a "usable byte position" to XLogRecPtr. A usable byte position
1860 : * is the position starting from the beginning of WAL, excluding all WAL
1861 : * page headers.
1862 : */
1863 : static XLogRecPtr
1864 58488430 : XLogBytePosToRecPtr(uint64 bytepos)
1865 : {
1866 : uint64 fullsegs;
1867 : uint64 fullpages;
1868 : uint64 bytesleft;
1869 : uint32 seg_offset;
1870 : XLogRecPtr result;
1871 :
1872 58488430 : fullsegs = bytepos / UsableBytesInSegment;
1873 58488430 : bytesleft = bytepos % UsableBytesInSegment;
1874 :
1875 58488430 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1876 : {
1877 : /* fits on first page of segment */
1878 104648 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1879 : }
1880 : else
1881 : {
1882 : /* account for the first page on segment with long header */
1883 58383782 : seg_offset = XLOG_BLCKSZ;
1884 58383782 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1885 :
1886 58383782 : fullpages = bytesleft / UsableBytesInPage;
1887 58383782 : bytesleft = bytesleft % UsableBytesInPage;
1888 :
1889 58383782 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1890 : }
1891 :
1892 58488430 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1893 :
1894 58488430 : return result;
1895 : }
1896 :
1897 : /*
1898 : * Like XLogBytePosToRecPtr, but if the position is at a page boundary,
1899 : * returns a pointer to the beginning of the page (ie. before page header),
1900 : * not to where the first xlog record on that page would go to. This is used
1901 : * when converting a pointer to the end of a record.
1902 : */
1903 : static XLogRecPtr
1904 30165828 : XLogBytePosToEndRecPtr(uint64 bytepos)
1905 : {
1906 : uint64 fullsegs;
1907 : uint64 fullpages;
1908 : uint64 bytesleft;
1909 : uint32 seg_offset;
1910 : XLogRecPtr result;
1911 :
1912 30165828 : fullsegs = bytepos / UsableBytesInSegment;
1913 30165828 : bytesleft = bytepos % UsableBytesInSegment;
1914 :
1915 30165828 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1916 : {
1917 : /* fits on first page of segment */
1918 166008 : if (bytesleft == 0)
1919 111088 : seg_offset = 0;
1920 : else
1921 54920 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1922 : }
1923 : else
1924 : {
1925 : /* account for the first page on segment with long header */
1926 29999820 : seg_offset = XLOG_BLCKSZ;
1927 29999820 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1928 :
1929 29999820 : fullpages = bytesleft / UsableBytesInPage;
1930 29999820 : bytesleft = bytesleft % UsableBytesInPage;
1931 :
1932 29999820 : if (bytesleft == 0)
1933 29356 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
1934 : else
1935 29970464 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1936 : }
1937 :
1938 30165828 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1939 :
1940 30165828 : return result;
1941 : }
1942 :
1943 : /*
1944 : * Convert an XLogRecPtr to a "usable byte position".
1945 : */
1946 : static uint64
1947 5044 : XLogRecPtrToBytePos(XLogRecPtr ptr)
1948 : {
1949 : uint64 fullsegs;
1950 : uint32 fullpages;
1951 : uint32 offset;
1952 : uint64 result;
1953 :
1954 5044 : XLByteToSeg(ptr, fullsegs, wal_segment_size);
1955 :
1956 5044 : fullpages = (XLogSegmentOffset(ptr, wal_segment_size)) / XLOG_BLCKSZ;
1957 5044 : offset = ptr % XLOG_BLCKSZ;
1958 :
1959 5044 : if (fullpages == 0)
1960 : {
1961 1978 : result = fullsegs * UsableBytesInSegment;
1962 1978 : if (offset > 0)
1963 : {
1964 : Assert(offset >= SizeOfXLogLongPHD);
1965 492 : result += offset - SizeOfXLogLongPHD;
1966 : }
1967 : }
1968 : else
1969 : {
1970 3066 : result = fullsegs * UsableBytesInSegment +
1971 3066 : (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
1972 3066 : (fullpages - 1) * UsableBytesInPage; /* full pages */
1973 3066 : if (offset > 0)
1974 : {
1975 : Assert(offset >= SizeOfXLogShortPHD);
1976 3046 : result += offset - SizeOfXLogShortPHD;
1977 : }
1978 : }
1979 :
1980 5044 : return result;
1981 : }
1982 :
1983 : /*
1984 : * Initialize XLOG buffers, writing out old buffers if they still contain
1985 : * unwritten data, upto the page containing 'upto'. Or if 'opportunistic' is
1986 : * true, initialize as many pages as we can without having to write out
1987 : * unwritten data. Any new pages are initialized to zeros, with pages headers
1988 : * initialized properly.
1989 : */
1990 : static void
1991 5193890 : AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
1992 : {
1993 5193890 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1994 : int nextidx;
1995 : XLogRecPtr OldPageRqstPtr;
1996 : XLogwrtRqst WriteRqst;
1997 5193890 : XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
1998 : XLogRecPtr NewPageBeginPtr;
1999 : XLogPageHeader NewPage;
2000 5193890 : int npages pg_attribute_unused() = 0;
2001 :
2002 5193890 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2003 :
2004 : /*
2005 : * Now that we have the lock, check if someone initialized the page
2006 : * already.
2007 : */
2008 14778038 : while (upto >= XLogCtl->InitializedUpTo || opportunistic)
2009 : {
2010 9593596 : nextidx = XLogRecPtrToBufIdx(XLogCtl->InitializedUpTo);
2011 :
2012 : /*
2013 : * Get ending-offset of the buffer page we need to replace (this may
2014 : * be zero if the buffer hasn't been used yet). Fall through if it's
2015 : * already written out.
2016 : */
2017 9593596 : OldPageRqstPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]);
2018 9593596 : if (LogwrtResult.Write < OldPageRqstPtr)
2019 : {
2020 : /*
2021 : * Nope, got work to do. If we just want to pre-initialize as much
2022 : * as we can without flushing, give up now.
2023 : */
2024 4613050 : if (opportunistic)
2025 9448 : break;
2026 :
2027 : /* Advance shared memory write request position */
2028 4603602 : SpinLockAcquire(&XLogCtl->info_lck);
2029 4603602 : if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
2030 1260592 : XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
2031 4603602 : SpinLockRelease(&XLogCtl->info_lck);
2032 :
2033 : /*
2034 : * Acquire an up-to-date LogwrtResult value and see if we still
2035 : * need to write it or if someone else already did.
2036 : */
2037 4603602 : RefreshXLogWriteResult(LogwrtResult);
2038 4603602 : if (LogwrtResult.Write < OldPageRqstPtr)
2039 : {
2040 : /*
2041 : * Must acquire write lock. Release WALBufMappingLock first,
2042 : * to make sure that all insertions that we need to wait for
2043 : * can finish (up to this same position). Otherwise we risk
2044 : * deadlock.
2045 : */
2046 4511790 : LWLockRelease(WALBufMappingLock);
2047 :
2048 4511790 : WaitXLogInsertionsToFinish(OldPageRqstPtr);
2049 :
2050 4511790 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2051 :
2052 4511790 : RefreshXLogWriteResult(LogwrtResult);
2053 4511790 : if (LogwrtResult.Write >= OldPageRqstPtr)
2054 : {
2055 : /* OK, someone wrote it already */
2056 535064 : LWLockRelease(WALWriteLock);
2057 : }
2058 : else
2059 : {
2060 : /* Have to write it ourselves */
2061 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2062 3976726 : WriteRqst.Write = OldPageRqstPtr;
2063 3976726 : WriteRqst.Flush = 0;
2064 3976726 : XLogWrite(WriteRqst, tli, false);
2065 3976726 : LWLockRelease(WALWriteLock);
2066 3976726 : pgWalUsage.wal_buffers_full++;
2067 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_DONE();
2068 :
2069 : /*
2070 : * Required for the flush of pending stats WAL data, per
2071 : * update of pgWalUsage.
2072 : */
2073 3976726 : pgstat_report_fixed = true;
2074 : }
2075 : /* Re-acquire WALBufMappingLock and retry */
2076 4511790 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2077 4511790 : continue;
2078 : }
2079 : }
2080 :
2081 : /*
2082 : * Now the next buffer slot is free and we can set it up to be the
2083 : * next output page.
2084 : */
2085 5072358 : NewPageBeginPtr = XLogCtl->InitializedUpTo;
2086 5072358 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2087 :
2088 : Assert(XLogRecPtrToBufIdx(NewPageBeginPtr) == nextidx);
2089 :
2090 5072358 : NewPage = (XLogPageHeader) (XLogCtl->pages + nextidx * (Size) XLOG_BLCKSZ);
2091 :
2092 : /*
2093 : * Mark the xlblock with InvalidXLogRecPtr and issue a write barrier
2094 : * before initializing. Otherwise, the old page may be partially
2095 : * zeroed but look valid.
2096 : */
2097 5072358 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], InvalidXLogRecPtr);
2098 5072358 : pg_write_barrier();
2099 :
2100 : /*
2101 : * Be sure to re-zero the buffer so that bytes beyond what we've
2102 : * written will look like zeroes and not valid XLOG records...
2103 : */
2104 5072358 : MemSet(NewPage, 0, XLOG_BLCKSZ);
2105 :
2106 : /*
2107 : * Fill the new page's header
2108 : */
2109 5072358 : NewPage->xlp_magic = XLOG_PAGE_MAGIC;
2110 :
2111 : /* NewPage->xlp_info = 0; */ /* done by memset */
2112 5072358 : NewPage->xlp_tli = tli;
2113 5072358 : NewPage->xlp_pageaddr = NewPageBeginPtr;
2114 :
2115 : /* NewPage->xlp_rem_len = 0; */ /* done by memset */
2116 :
2117 : /*
2118 : * If online backup is not in progress, mark the header to indicate
2119 : * that WAL records beginning in this page have removable backup
2120 : * blocks. This allows the WAL archiver to know whether it is safe to
2121 : * compress archived WAL data by transforming full-block records into
2122 : * the non-full-block format. It is sufficient to record this at the
2123 : * page level because we force a page switch (in fact a segment
2124 : * switch) when starting a backup, so the flag will be off before any
2125 : * records can be written during the backup. At the end of a backup,
2126 : * the last page will be marked as all unsafe when perhaps only part
2127 : * is unsafe, but at worst the archiver would miss the opportunity to
2128 : * compress a few records.
2129 : */
2130 5072358 : if (Insert->runningBackups == 0)
2131 4815098 : NewPage->xlp_info |= XLP_BKP_REMOVABLE;
2132 :
2133 : /*
2134 : * If first page of an XLOG segment file, make it a long header.
2135 : */
2136 5072358 : if ((XLogSegmentOffset(NewPage->xlp_pageaddr, wal_segment_size)) == 0)
2137 : {
2138 3612 : XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2139 :
2140 3612 : NewLongPage->xlp_sysid = ControlFile->system_identifier;
2141 3612 : NewLongPage->xlp_seg_size = wal_segment_size;
2142 3612 : NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2143 3612 : NewPage->xlp_info |= XLP_LONG_HEADER;
2144 : }
2145 :
2146 : /*
2147 : * Make sure the initialization of the page becomes visible to others
2148 : * before the xlblocks update. GetXLogBuffer() reads xlblocks without
2149 : * holding a lock.
2150 : */
2151 5072358 : pg_write_barrier();
2152 :
2153 5072358 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], NewPageEndPtr);
2154 5072358 : XLogCtl->InitializedUpTo = NewPageEndPtr;
2155 :
2156 5072358 : npages++;
2157 : }
2158 5193890 : LWLockRelease(WALBufMappingLock);
2159 :
2160 : #ifdef WAL_DEBUG
2161 : if (XLOG_DEBUG && npages > 0)
2162 : {
2163 : elog(DEBUG1, "initialized %d pages, up to %X/%08X",
2164 : npages, LSN_FORMAT_ARGS(NewPageEndPtr));
2165 : }
2166 : #endif
2167 5193890 : }
2168 :
2169 : /*
2170 : * Calculate CheckPointSegments based on max_wal_size_mb and
2171 : * checkpoint_completion_target.
2172 : */
2173 : static void
2174 15522 : CalculateCheckpointSegments(void)
2175 : {
2176 : double target;
2177 :
2178 : /*-------
2179 : * Calculate the distance at which to trigger a checkpoint, to avoid
2180 : * exceeding max_wal_size_mb. This is based on two assumptions:
2181 : *
2182 : * a) we keep WAL for only one checkpoint cycle (prior to PG11 we kept
2183 : * WAL for two checkpoint cycles to allow us to recover from the
2184 : * secondary checkpoint if the first checkpoint failed, though we
2185 : * only did this on the primary anyway, not on standby. Keeping just
2186 : * one checkpoint simplifies processing and reduces disk space in
2187 : * many smaller databases.)
2188 : * b) during checkpoint, we consume checkpoint_completion_target *
2189 : * number of segments consumed between checkpoints.
2190 : *-------
2191 : */
2192 15522 : target = (double) ConvertToXSegs(max_wal_size_mb, wal_segment_size) /
2193 15522 : (1.0 + CheckPointCompletionTarget);
2194 :
2195 : /* round down */
2196 15522 : CheckPointSegments = (int) target;
2197 :
2198 15522 : if (CheckPointSegments < 1)
2199 20 : CheckPointSegments = 1;
2200 15522 : }
2201 :
2202 : void
2203 11230 : assign_max_wal_size(int newval, void *extra)
2204 : {
2205 11230 : max_wal_size_mb = newval;
2206 11230 : CalculateCheckpointSegments();
2207 11230 : }
2208 :
2209 : void
2210 2276 : assign_checkpoint_completion_target(double newval, void *extra)
2211 : {
2212 2276 : CheckPointCompletionTarget = newval;
2213 2276 : CalculateCheckpointSegments();
2214 2276 : }
2215 :
2216 : bool
2217 4394 : check_wal_segment_size(int *newval, void **extra, GucSource source)
2218 : {
2219 4394 : if (!IsValidWalSegSize(*newval))
2220 : {
2221 0 : GUC_check_errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.");
2222 0 : return false;
2223 : }
2224 :
2225 4394 : return true;
2226 : }
2227 :
2228 : /*
2229 : * At a checkpoint, how many WAL segments to recycle as preallocated future
2230 : * XLOG segments? Returns the highest segment that should be preallocated.
2231 : */
2232 : static XLogSegNo
2233 3468 : XLOGfileslop(XLogRecPtr lastredoptr)
2234 : {
2235 : XLogSegNo minSegNo;
2236 : XLogSegNo maxSegNo;
2237 : double distance;
2238 : XLogSegNo recycleSegNo;
2239 :
2240 : /*
2241 : * Calculate the segment numbers that min_wal_size_mb and max_wal_size_mb
2242 : * correspond to. Always recycle enough segments to meet the minimum, and
2243 : * remove enough segments to stay below the maximum.
2244 : */
2245 3468 : minSegNo = lastredoptr / wal_segment_size +
2246 3468 : ConvertToXSegs(min_wal_size_mb, wal_segment_size) - 1;
2247 3468 : maxSegNo = lastredoptr / wal_segment_size +
2248 3468 : ConvertToXSegs(max_wal_size_mb, wal_segment_size) - 1;
2249 :
2250 : /*
2251 : * Between those limits, recycle enough segments to get us through to the
2252 : * estimated end of next checkpoint.
2253 : *
2254 : * To estimate where the next checkpoint will finish, assume that the
2255 : * system runs steadily consuming CheckPointDistanceEstimate bytes between
2256 : * every checkpoint.
2257 : */
2258 3468 : distance = (1.0 + CheckPointCompletionTarget) * CheckPointDistanceEstimate;
2259 : /* add 10% for good measure. */
2260 3468 : distance *= 1.10;
2261 :
2262 3468 : recycleSegNo = (XLogSegNo) ceil(((double) lastredoptr + distance) /
2263 : wal_segment_size);
2264 :
2265 3468 : if (recycleSegNo < minSegNo)
2266 2450 : recycleSegNo = minSegNo;
2267 3468 : if (recycleSegNo > maxSegNo)
2268 770 : recycleSegNo = maxSegNo;
2269 :
2270 3468 : return recycleSegNo;
2271 : }
2272 :
2273 : /*
2274 : * Check whether we've consumed enough xlog space that a checkpoint is needed.
2275 : *
2276 : * new_segno indicates a log file that has just been filled up (or read
2277 : * during recovery). We measure the distance from RedoRecPtr to new_segno
2278 : * and see if that exceeds CheckPointSegments.
2279 : *
2280 : * Note: it is caller's responsibility that RedoRecPtr is up-to-date.
2281 : */
2282 : bool
2283 10344 : XLogCheckpointNeeded(XLogSegNo new_segno)
2284 : {
2285 : XLogSegNo old_segno;
2286 :
2287 10344 : XLByteToSeg(RedoRecPtr, old_segno, wal_segment_size);
2288 :
2289 10344 : if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2290 6730 : return true;
2291 3614 : return false;
2292 : }
2293 :
2294 : /*
2295 : * Write and/or fsync the log at least as far as WriteRqst indicates.
2296 : *
2297 : * If flexible == true, we don't have to write as far as WriteRqst, but
2298 : * may stop at any convenient boundary (such as a cache or logfile boundary).
2299 : * This option allows us to avoid uselessly issuing multiple writes when a
2300 : * single one would do.
2301 : *
2302 : * Must be called with WALWriteLock held. WaitXLogInsertionsToFinish(WriteRqst)
2303 : * must be called before grabbing the lock, to make sure the data is ready to
2304 : * write.
2305 : */
2306 : static void
2307 4239132 : XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible)
2308 : {
2309 : bool ispartialpage;
2310 : bool last_iteration;
2311 : bool finishing_seg;
2312 : int curridx;
2313 : int npages;
2314 : int startidx;
2315 : uint32 startoffset;
2316 :
2317 : /* We should always be inside a critical section here */
2318 : Assert(CritSectionCount > 0);
2319 :
2320 : /*
2321 : * Update local LogwrtResult (caller probably did this already, but...)
2322 : */
2323 4239132 : RefreshXLogWriteResult(LogwrtResult);
2324 :
2325 : /*
2326 : * Since successive pages in the xlog cache are consecutively allocated,
2327 : * we can usually gather multiple pages together and issue just one
2328 : * write() call. npages is the number of pages we have determined can be
2329 : * written together; startidx is the cache block index of the first one,
2330 : * and startoffset is the file offset at which it should go. The latter
2331 : * two variables are only valid when npages > 0, but we must initialize
2332 : * all of them to keep the compiler quiet.
2333 : */
2334 4239132 : npages = 0;
2335 4239132 : startidx = 0;
2336 4239132 : startoffset = 0;
2337 :
2338 : /*
2339 : * Within the loop, curridx is the cache block index of the page to
2340 : * consider writing. Begin at the buffer containing the next unwritten
2341 : * page, or last partially written page.
2342 : */
2343 4239132 : curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2344 :
2345 9210866 : while (LogwrtResult.Write < WriteRqst.Write)
2346 : {
2347 : /*
2348 : * Make sure we're not ahead of the insert process. This could happen
2349 : * if we're passed a bogus WriteRqst.Write that is past the end of the
2350 : * last page that's been initialized by AdvanceXLInsertBuffer.
2351 : */
2352 5223598 : XLogRecPtr EndPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[curridx]);
2353 :
2354 5223598 : if (LogwrtResult.Write >= EndPtr)
2355 0 : elog(PANIC, "xlog write request %X/%08X is past end of log %X/%08X",
2356 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2357 : LSN_FORMAT_ARGS(EndPtr));
2358 :
2359 : /* Advance LogwrtResult.Write to end of current buffer page */
2360 5223598 : LogwrtResult.Write = EndPtr;
2361 5223598 : ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2362 :
2363 5223598 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2364 : wal_segment_size))
2365 : {
2366 : /*
2367 : * Switch to new logfile segment. We cannot have any pending
2368 : * pages here (since we dump what we have at segment end).
2369 : */
2370 : Assert(npages == 0);
2371 26820 : if (openLogFile >= 0)
2372 12160 : XLogFileClose();
2373 26820 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2374 : wal_segment_size);
2375 26820 : openLogTLI = tli;
2376 :
2377 : /* create/use new log file */
2378 26820 : openLogFile = XLogFileInit(openLogSegNo, tli);
2379 26820 : ReserveExternalFD();
2380 : }
2381 :
2382 : /* Make sure we have the current logfile open */
2383 5223598 : if (openLogFile < 0)
2384 : {
2385 0 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2386 : wal_segment_size);
2387 0 : openLogTLI = tli;
2388 0 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2389 0 : ReserveExternalFD();
2390 : }
2391 :
2392 : /* Add current page to the set of pending pages-to-dump */
2393 5223598 : if (npages == 0)
2394 : {
2395 : /* first of group */
2396 4268448 : startidx = curridx;
2397 4268448 : startoffset = XLogSegmentOffset(LogwrtResult.Write - XLOG_BLCKSZ,
2398 : wal_segment_size);
2399 : }
2400 5223598 : npages++;
2401 :
2402 : /*
2403 : * Dump the set if this will be the last loop iteration, or if we are
2404 : * at the last page of the cache area (since the next page won't be
2405 : * contiguous in memory), or if we are at the end of the logfile
2406 : * segment.
2407 : */
2408 5223598 : last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2409 :
2410 10202802 : finishing_seg = !ispartialpage &&
2411 4979204 : (startoffset + npages * XLOG_BLCKSZ) >= wal_segment_size;
2412 :
2413 5223598 : if (last_iteration ||
2414 987340 : curridx == XLogCtl->XLogCacheBlck ||
2415 : finishing_seg)
2416 : {
2417 : char *from;
2418 : Size nbytes;
2419 : Size nleft;
2420 : ssize_t written;
2421 : instr_time start;
2422 :
2423 : /* OK to write the page(s) */
2424 4268448 : from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2425 4268448 : nbytes = npages * (Size) XLOG_BLCKSZ;
2426 4268448 : nleft = nbytes;
2427 : do
2428 : {
2429 4268448 : errno = 0;
2430 :
2431 : /*
2432 : * Measure I/O timing to write WAL data, for pg_stat_io.
2433 : */
2434 4268448 : start = pgstat_prepare_io_time(track_wal_io_timing);
2435 :
2436 4268448 : pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE);
2437 4268448 : written = pg_pwrite(openLogFile, from, nleft, startoffset);
2438 4268448 : pgstat_report_wait_end();
2439 :
2440 4268448 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL,
2441 : IOOP_WRITE, start, 1, written);
2442 :
2443 4268448 : if (written <= 0)
2444 : {
2445 : char xlogfname[MAXFNAMELEN];
2446 : int save_errno;
2447 :
2448 0 : if (errno == EINTR)
2449 0 : continue;
2450 :
2451 0 : save_errno = errno;
2452 0 : XLogFileName(xlogfname, tli, openLogSegNo,
2453 : wal_segment_size);
2454 0 : errno = save_errno;
2455 0 : ereport(PANIC,
2456 : (errcode_for_file_access(),
2457 : errmsg("could not write to log file \"%s\" at offset %u, length %zu: %m",
2458 : xlogfname, startoffset, nleft)));
2459 : }
2460 4268448 : nleft -= written;
2461 4268448 : from += written;
2462 4268448 : startoffset += written;
2463 4268448 : } while (nleft > 0);
2464 :
2465 4268448 : npages = 0;
2466 :
2467 : /*
2468 : * If we just wrote the whole last page of a logfile segment,
2469 : * fsync the segment immediately. This avoids having to go back
2470 : * and re-open prior segments when an fsync request comes along
2471 : * later. Doing it here ensures that one and only one backend will
2472 : * perform this fsync.
2473 : *
2474 : * This is also the right place to notify the Archiver that the
2475 : * segment is ready to copy to archival storage, and to update the
2476 : * timer for archive_timeout, and to signal for a checkpoint if
2477 : * too many logfile segments have been used since the last
2478 : * checkpoint.
2479 : */
2480 4268448 : if (finishing_seg)
2481 : {
2482 3844 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2483 :
2484 : /* signal that we need to wakeup walsenders later */
2485 3844 : WalSndWakeupRequest();
2486 :
2487 3844 : LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2488 :
2489 3844 : if (XLogArchivingActive())
2490 812 : XLogArchiveNotifySeg(openLogSegNo, tli);
2491 :
2492 3844 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2493 3844 : XLogCtl->lastSegSwitchLSN = LogwrtResult.Flush;
2494 :
2495 : /*
2496 : * Request a checkpoint if we've consumed too much xlog since
2497 : * the last one. For speed, we first check using the local
2498 : * copy of RedoRecPtr, which might be out of date; if it looks
2499 : * like a checkpoint is needed, forcibly update RedoRecPtr and
2500 : * recheck.
2501 : */
2502 3844 : if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2503 : {
2504 486 : (void) GetRedoRecPtr();
2505 486 : if (XLogCheckpointNeeded(openLogSegNo))
2506 388 : RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2507 : }
2508 : }
2509 : }
2510 :
2511 5223598 : if (ispartialpage)
2512 : {
2513 : /* Only asked to write a partial page */
2514 244394 : LogwrtResult.Write = WriteRqst.Write;
2515 244394 : break;
2516 : }
2517 4979204 : curridx = NextBufIdx(curridx);
2518 :
2519 : /* If flexible, break out of loop as soon as we wrote something */
2520 4979204 : if (flexible && npages == 0)
2521 7470 : break;
2522 : }
2523 :
2524 : Assert(npages == 0);
2525 :
2526 : /*
2527 : * If asked to flush, do so
2528 : */
2529 4239132 : if (LogwrtResult.Flush < WriteRqst.Flush &&
2530 260916 : LogwrtResult.Flush < LogwrtResult.Write)
2531 : {
2532 : /*
2533 : * Could get here without iterating above loop, in which case we might
2534 : * have no open file or the wrong one. However, we do not need to
2535 : * fsync more than one file.
2536 : */
2537 260784 : if (wal_sync_method != WAL_SYNC_METHOD_OPEN &&
2538 260784 : wal_sync_method != WAL_SYNC_METHOD_OPEN_DSYNC)
2539 : {
2540 260784 : if (openLogFile >= 0 &&
2541 260754 : !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2542 : wal_segment_size))
2543 306 : XLogFileClose();
2544 260784 : if (openLogFile < 0)
2545 : {
2546 336 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2547 : wal_segment_size);
2548 336 : openLogTLI = tli;
2549 336 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2550 336 : ReserveExternalFD();
2551 : }
2552 :
2553 260784 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2554 : }
2555 :
2556 : /* signal that we need to wakeup walsenders later */
2557 260784 : WalSndWakeupRequest();
2558 :
2559 260784 : LogwrtResult.Flush = LogwrtResult.Write;
2560 : }
2561 :
2562 : /*
2563 : * Update shared-memory status
2564 : *
2565 : * We make sure that the shared 'request' values do not fall behind the
2566 : * 'result' values. This is not absolutely essential, but it saves some
2567 : * code in a couple of places.
2568 : */
2569 4239132 : SpinLockAcquire(&XLogCtl->info_lck);
2570 4239132 : if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
2571 226972 : XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
2572 4239132 : if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
2573 263822 : XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
2574 4239132 : SpinLockRelease(&XLogCtl->info_lck);
2575 :
2576 : /*
2577 : * We write Write first, bar, then Flush. When reading, the opposite must
2578 : * be done (with a matching barrier in between), so that we always see a
2579 : * Flush value that trails behind the Write value seen.
2580 : */
2581 4239132 : pg_atomic_write_u64(&XLogCtl->logWriteResult, LogwrtResult.Write);
2582 4239132 : pg_write_barrier();
2583 4239132 : pg_atomic_write_u64(&XLogCtl->logFlushResult, LogwrtResult.Flush);
2584 :
2585 : #ifdef USE_ASSERT_CHECKING
2586 : {
2587 : XLogRecPtr Flush;
2588 : XLogRecPtr Write;
2589 : XLogRecPtr Insert;
2590 :
2591 : Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult);
2592 : pg_read_barrier();
2593 : Write = pg_atomic_read_u64(&XLogCtl->logWriteResult);
2594 : pg_read_barrier();
2595 : Insert = pg_atomic_read_u64(&XLogCtl->logInsertResult);
2596 :
2597 : /* WAL written to disk is always ahead of WAL flushed */
2598 : Assert(Write >= Flush);
2599 :
2600 : /* WAL inserted to buffers is always ahead of WAL written */
2601 : Assert(Insert >= Write);
2602 : }
2603 : #endif
2604 4239132 : }
2605 :
2606 : /*
2607 : * Record the LSN for an asynchronous transaction commit/abort
2608 : * and nudge the WALWriter if there is work for it to do.
2609 : * (This should not be called for synchronous commits.)
2610 : */
2611 : void
2612 95206 : XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2613 : {
2614 95206 : XLogRecPtr WriteRqstPtr = asyncXactLSN;
2615 : bool sleeping;
2616 95206 : bool wakeup = false;
2617 : XLogRecPtr prevAsyncXactLSN;
2618 :
2619 95206 : SpinLockAcquire(&XLogCtl->info_lck);
2620 95206 : sleeping = XLogCtl->WalWriterSleeping;
2621 95206 : prevAsyncXactLSN = XLogCtl->asyncXactLSN;
2622 95206 : if (XLogCtl->asyncXactLSN < asyncXactLSN)
2623 94250 : XLogCtl->asyncXactLSN = asyncXactLSN;
2624 95206 : SpinLockRelease(&XLogCtl->info_lck);
2625 :
2626 : /*
2627 : * If somebody else already called this function with a more aggressive
2628 : * LSN, they will have done what we needed (and perhaps more).
2629 : */
2630 95206 : if (asyncXactLSN <= prevAsyncXactLSN)
2631 956 : return;
2632 :
2633 : /*
2634 : * If the WALWriter is sleeping, kick it to make it come out of low-power
2635 : * mode, so that this async commit will reach disk within the expected
2636 : * amount of time. Otherwise, determine whether it has enough WAL
2637 : * available to flush, the same way that XLogBackgroundFlush() does.
2638 : */
2639 94250 : if (sleeping)
2640 36 : wakeup = true;
2641 : else
2642 : {
2643 : int flushblocks;
2644 :
2645 94214 : RefreshXLogWriteResult(LogwrtResult);
2646 :
2647 94214 : flushblocks =
2648 94214 : WriteRqstPtr / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
2649 :
2650 94214 : if (WalWriterFlushAfter == 0 || flushblocks >= WalWriterFlushAfter)
2651 7400 : wakeup = true;
2652 : }
2653 :
2654 94250 : if (wakeup)
2655 : {
2656 7436 : volatile PROC_HDR *procglobal = ProcGlobal;
2657 7436 : ProcNumber walwriterProc = procglobal->walwriterProc;
2658 :
2659 7436 : if (walwriterProc != INVALID_PROC_NUMBER)
2660 392 : SetLatch(&GetPGProcByNumber(walwriterProc)->procLatch);
2661 : }
2662 : }
2663 :
2664 : /*
2665 : * Record the LSN up to which we can remove WAL because it's not required by
2666 : * any replication slot.
2667 : */
2668 : void
2669 51732 : XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn)
2670 : {
2671 51732 : SpinLockAcquire(&XLogCtl->info_lck);
2672 51732 : XLogCtl->replicationSlotMinLSN = lsn;
2673 51732 : SpinLockRelease(&XLogCtl->info_lck);
2674 51732 : }
2675 :
2676 :
2677 : /*
2678 : * Return the oldest LSN we must retain to satisfy the needs of some
2679 : * replication slot.
2680 : */
2681 : static XLogRecPtr
2682 4468 : XLogGetReplicationSlotMinimumLSN(void)
2683 : {
2684 : XLogRecPtr retval;
2685 :
2686 4468 : SpinLockAcquire(&XLogCtl->info_lck);
2687 4468 : retval = XLogCtl->replicationSlotMinLSN;
2688 4468 : SpinLockRelease(&XLogCtl->info_lck);
2689 :
2690 4468 : return retval;
2691 : }
2692 :
2693 : /*
2694 : * Advance minRecoveryPoint in control file.
2695 : *
2696 : * If we crash during recovery, we must reach this point again before the
2697 : * database is consistent.
2698 : *
2699 : * If 'force' is true, 'lsn' argument is ignored. Otherwise, minRecoveryPoint
2700 : * is only updated if it's not already greater than or equal to 'lsn'.
2701 : */
2702 : static void
2703 218710 : UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2704 : {
2705 : /* Quick check using our local copy of the variable */
2706 218710 : if (!updateMinRecoveryPoint || (!force && lsn <= LocalMinRecoveryPoint))
2707 205346 : return;
2708 :
2709 : /*
2710 : * An invalid minRecoveryPoint means that we need to recover all the WAL,
2711 : * i.e., we're doing crash recovery. We never modify the control file's
2712 : * value in that case, so we can short-circuit future checks here too. The
2713 : * local values of minRecoveryPoint and minRecoveryPointTLI should not be
2714 : * updated until crash recovery finishes. We only do this for the startup
2715 : * process as it should not update its own reference of minRecoveryPoint
2716 : * until it has finished crash recovery to make sure that all WAL
2717 : * available is replayed in this case. This also saves from extra locks
2718 : * taken on the control file from the startup process.
2719 : */
2720 13364 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint) && InRecovery)
2721 : {
2722 62 : updateMinRecoveryPoint = false;
2723 62 : return;
2724 : }
2725 :
2726 13302 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2727 :
2728 : /* update local copy */
2729 13302 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
2730 13302 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2731 :
2732 13302 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
2733 6 : updateMinRecoveryPoint = false;
2734 13296 : else if (force || LocalMinRecoveryPoint < lsn)
2735 : {
2736 : XLogRecPtr newMinRecoveryPoint;
2737 : TimeLineID newMinRecoveryPointTLI;
2738 :
2739 : /*
2740 : * To avoid having to update the control file too often, we update it
2741 : * all the way to the last record being replayed, even though 'lsn'
2742 : * would suffice for correctness. This also allows the 'force' case
2743 : * to not need a valid 'lsn' value.
2744 : *
2745 : * Another important reason for doing it this way is that the passed
2746 : * 'lsn' value could be bogus, i.e., past the end of available WAL, if
2747 : * the caller got it from a corrupted heap page. Accepting such a
2748 : * value as the min recovery point would prevent us from coming up at
2749 : * all. Instead, we just log a warning and continue with recovery.
2750 : * (See also the comments about corrupt LSNs in XLogFlush.)
2751 : */
2752 10796 : newMinRecoveryPoint = GetCurrentReplayRecPtr(&newMinRecoveryPointTLI);
2753 10796 : if (!force && newMinRecoveryPoint < lsn)
2754 0 : elog(WARNING,
2755 : "xlog min recovery request %X/%08X is past current point %X/%08X",
2756 : LSN_FORMAT_ARGS(lsn), LSN_FORMAT_ARGS(newMinRecoveryPoint));
2757 :
2758 : /* update control file */
2759 10796 : if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2760 : {
2761 10100 : ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2762 10100 : ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2763 10100 : UpdateControlFile();
2764 10100 : LocalMinRecoveryPoint = newMinRecoveryPoint;
2765 10100 : LocalMinRecoveryPointTLI = newMinRecoveryPointTLI;
2766 :
2767 10100 : ereport(DEBUG2,
2768 : errmsg_internal("updated min recovery point to %X/%08X on timeline %u",
2769 : LSN_FORMAT_ARGS(newMinRecoveryPoint),
2770 : newMinRecoveryPointTLI));
2771 : }
2772 : }
2773 13302 : LWLockRelease(ControlFileLock);
2774 : }
2775 :
2776 : /*
2777 : * Ensure that all XLOG data through the given position is flushed to disk.
2778 : *
2779 : * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
2780 : * already held, and we try to avoid acquiring it if possible.
2781 : */
2782 : void
2783 1390252 : XLogFlush(XLogRecPtr record)
2784 : {
2785 : XLogRecPtr WriteRqstPtr;
2786 : XLogwrtRqst WriteRqst;
2787 1390252 : TimeLineID insertTLI = XLogCtl->InsertTimeLineID;
2788 :
2789 : /*
2790 : * During REDO, we are reading not writing WAL. Therefore, instead of
2791 : * trying to flush the WAL, we should update minRecoveryPoint instead. We
2792 : * test XLogInsertAllowed(), not InRecovery, because we need checkpointer
2793 : * to act this way too, and because when it tries to write the
2794 : * end-of-recovery checkpoint, it should indeed flush.
2795 : */
2796 1390252 : if (!XLogInsertAllowed())
2797 : {
2798 217830 : UpdateMinRecoveryPoint(record, false);
2799 1111382 : return;
2800 : }
2801 :
2802 : /* Quick exit if already known flushed */
2803 1172422 : if (record <= LogwrtResult.Flush)
2804 893552 : return;
2805 :
2806 : #ifdef WAL_DEBUG
2807 : if (XLOG_DEBUG)
2808 : elog(LOG, "xlog flush request %X/%08X; write %X/%08X; flush %X/%08X",
2809 : LSN_FORMAT_ARGS(record),
2810 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2811 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2812 : #endif
2813 :
2814 278870 : START_CRIT_SECTION();
2815 :
2816 : /*
2817 : * Since fsync is usually a horribly expensive operation, we try to
2818 : * piggyback as much data as we can on each fsync: if we see any more data
2819 : * entered into the xlog buffer, we'll write and fsync that too, so that
2820 : * the final value of LogwrtResult.Flush is as large as possible. This
2821 : * gives us some chance of avoiding another fsync immediately after.
2822 : */
2823 :
2824 : /* initialize to given target; may increase below */
2825 278870 : WriteRqstPtr = record;
2826 :
2827 : /*
2828 : * Now wait until we get the write lock, or someone else does the flush
2829 : * for us.
2830 : */
2831 : for (;;)
2832 11304 : {
2833 : XLogRecPtr insertpos;
2834 :
2835 : /* done already? */
2836 290174 : RefreshXLogWriteResult(LogwrtResult);
2837 290174 : if (record <= LogwrtResult.Flush)
2838 21464 : break;
2839 :
2840 : /*
2841 : * Before actually performing the write, wait for all in-flight
2842 : * insertions to the pages we're about to write to finish.
2843 : */
2844 268710 : SpinLockAcquire(&XLogCtl->info_lck);
2845 268710 : if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
2846 20864 : WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
2847 268710 : SpinLockRelease(&XLogCtl->info_lck);
2848 268710 : insertpos = WaitXLogInsertionsToFinish(WriteRqstPtr);
2849 :
2850 : /*
2851 : * Try to get the write lock. If we can't get it immediately, wait
2852 : * until it's released, and recheck if we still need to do the flush
2853 : * or if the backend that held the lock did it for us already. This
2854 : * helps to maintain a good rate of group committing when the system
2855 : * is bottlenecked by the speed of fsyncing.
2856 : */
2857 268710 : if (!LWLockAcquireOrWait(WALWriteLock, LW_EXCLUSIVE))
2858 : {
2859 : /*
2860 : * The lock is now free, but we didn't acquire it yet. Before we
2861 : * do, loop back to check if someone else flushed the record for
2862 : * us already.
2863 : */
2864 11304 : continue;
2865 : }
2866 :
2867 : /* Got the lock; recheck whether request is satisfied */
2868 257406 : RefreshXLogWriteResult(LogwrtResult);
2869 257406 : if (record <= LogwrtResult.Flush)
2870 : {
2871 4052 : LWLockRelease(WALWriteLock);
2872 4052 : break;
2873 : }
2874 :
2875 : /*
2876 : * Sleep before flush! By adding a delay here, we may give further
2877 : * backends the opportunity to join the backlog of group commit
2878 : * followers; this can significantly improve transaction throughput,
2879 : * at the risk of increasing transaction latency.
2880 : *
2881 : * We do not sleep if enableFsync is not turned on, nor if there are
2882 : * fewer than CommitSiblings other backends with active transactions.
2883 : */
2884 253354 : if (CommitDelay > 0 && enableFsync &&
2885 0 : MinimumActiveBackends(CommitSiblings))
2886 : {
2887 0 : pg_usleep(CommitDelay);
2888 :
2889 : /*
2890 : * Re-check how far we can now flush the WAL. It's generally not
2891 : * safe to call WaitXLogInsertionsToFinish while holding
2892 : * WALWriteLock, because an in-progress insertion might need to
2893 : * also grab WALWriteLock to make progress. But we know that all
2894 : * the insertions up to insertpos have already finished, because
2895 : * that's what the earlier WaitXLogInsertionsToFinish() returned.
2896 : * We're only calling it again to allow insertpos to be moved
2897 : * further forward, not to actually wait for anyone.
2898 : */
2899 0 : insertpos = WaitXLogInsertionsToFinish(insertpos);
2900 : }
2901 :
2902 : /* try to write/flush later additions to XLOG as well */
2903 253354 : WriteRqst.Write = insertpos;
2904 253354 : WriteRqst.Flush = insertpos;
2905 :
2906 253354 : XLogWrite(WriteRqst, insertTLI, false);
2907 :
2908 253354 : LWLockRelease(WALWriteLock);
2909 : /* done */
2910 253354 : break;
2911 : }
2912 :
2913 278870 : END_CRIT_SECTION();
2914 :
2915 : /* wake up walsenders now that we've released heavily contended locks */
2916 278870 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
2917 :
2918 : /*
2919 : * If we still haven't flushed to the request point then we have a
2920 : * problem; most likely, the requested flush point is past end of XLOG.
2921 : * This has been seen to occur when a disk page has a corrupted LSN.
2922 : *
2923 : * Formerly we treated this as a PANIC condition, but that hurts the
2924 : * system's robustness rather than helping it: we do not want to take down
2925 : * the whole system due to corruption on one data page. In particular, if
2926 : * the bad page is encountered again during recovery then we would be
2927 : * unable to restart the database at all! (This scenario actually
2928 : * happened in the field several times with 7.1 releases.) As of 8.4, bad
2929 : * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
2930 : * the only time we can reach here during recovery is while flushing the
2931 : * end-of-recovery checkpoint record, and we don't expect that to have a
2932 : * bad LSN.
2933 : *
2934 : * Note that for calls from xact.c, the ERROR will be promoted to PANIC
2935 : * since xact.c calls this routine inside a critical section. However,
2936 : * calls from bufmgr.c are not within critical sections and so we will not
2937 : * force a restart for a bad LSN on a data page.
2938 : */
2939 278870 : if (LogwrtResult.Flush < record)
2940 0 : elog(ERROR,
2941 : "xlog flush request %X/%08X is not satisfied --- flushed only to %X/%08X",
2942 : LSN_FORMAT_ARGS(record),
2943 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2944 :
2945 : /*
2946 : * Cross-check XLogNeedsFlush(). Some of the checks of XLogFlush() and
2947 : * XLogNeedsFlush() are duplicated, and this assertion ensures that these
2948 : * remain consistent.
2949 : */
2950 : Assert(!XLogNeedsFlush(record));
2951 : }
2952 :
2953 : /*
2954 : * Write & flush xlog, but without specifying exactly where to.
2955 : *
2956 : * We normally write only completed blocks; but if there is nothing to do on
2957 : * that basis, we check for unwritten async commits in the current incomplete
2958 : * block, and write through the latest one of those. Thus, if async commits
2959 : * are not being used, we will write complete blocks only.
2960 : *
2961 : * If, based on the above, there's anything to write we do so immediately. But
2962 : * to avoid calling fsync, fdatasync et. al. at a rate that'd impact
2963 : * concurrent IO, we only flush WAL every wal_writer_delay ms, or if there's
2964 : * more than wal_writer_flush_after unflushed blocks.
2965 : *
2966 : * We can guarantee that async commits reach disk after at most three
2967 : * wal_writer_delay cycles. (When flushing complete blocks, we allow XLogWrite
2968 : * to write "flexibly", meaning it can stop at the end of the buffer ring;
2969 : * this makes a difference only with very high load or long wal_writer_delay,
2970 : * but imposes one extra cycle for the worst case for async commits.)
2971 : *
2972 : * This routine is invoked periodically by the background walwriter process.
2973 : *
2974 : * Returns true if there was any work to do, even if we skipped flushing due
2975 : * to wal_writer_delay/wal_writer_flush_after.
2976 : */
2977 : bool
2978 33500 : XLogBackgroundFlush(void)
2979 : {
2980 : XLogwrtRqst WriteRqst;
2981 33500 : bool flexible = true;
2982 : static TimestampTz lastflush;
2983 : TimestampTz now;
2984 : int flushblocks;
2985 : TimeLineID insertTLI;
2986 :
2987 : /* XLOG doesn't need flushing during recovery */
2988 33500 : if (RecoveryInProgress())
2989 2912 : return false;
2990 :
2991 : /*
2992 : * Since we're not in recovery, InsertTimeLineID is set and can't change,
2993 : * so we can read it without a lock.
2994 : */
2995 30588 : insertTLI = XLogCtl->InsertTimeLineID;
2996 :
2997 : /* read updated LogwrtRqst */
2998 30588 : SpinLockAcquire(&XLogCtl->info_lck);
2999 30588 : WriteRqst = XLogCtl->LogwrtRqst;
3000 30588 : SpinLockRelease(&XLogCtl->info_lck);
3001 :
3002 : /* back off to last completed page boundary */
3003 30588 : WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
3004 :
3005 : /* if we have already flushed that far, consider async commit records */
3006 30588 : RefreshXLogWriteResult(LogwrtResult);
3007 30588 : if (WriteRqst.Write <= LogwrtResult.Flush)
3008 : {
3009 22412 : SpinLockAcquire(&XLogCtl->info_lck);
3010 22412 : WriteRqst.Write = XLogCtl->asyncXactLSN;
3011 22412 : SpinLockRelease(&XLogCtl->info_lck);
3012 22412 : flexible = false; /* ensure it all gets written */
3013 : }
3014 :
3015 : /*
3016 : * If already known flushed, we're done. Just need to check if we are
3017 : * holding an open file handle to a logfile that's no longer in use,
3018 : * preventing the file from being deleted.
3019 : */
3020 30588 : if (WriteRqst.Write <= LogwrtResult.Flush)
3021 : {
3022 21140 : if (openLogFile >= 0)
3023 : {
3024 15070 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
3025 : wal_segment_size))
3026 : {
3027 254 : XLogFileClose();
3028 : }
3029 : }
3030 21140 : return false;
3031 : }
3032 :
3033 : /*
3034 : * Determine how far to flush WAL, based on the wal_writer_delay and
3035 : * wal_writer_flush_after GUCs.
3036 : *
3037 : * Note that XLogSetAsyncXactLSN() performs similar calculation based on
3038 : * wal_writer_flush_after, to decide when to wake us up. Make sure the
3039 : * logic is the same in both places if you change this.
3040 : */
3041 9448 : now = GetCurrentTimestamp();
3042 9448 : flushblocks =
3043 9448 : WriteRqst.Write / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
3044 :
3045 9448 : if (WalWriterFlushAfter == 0 || lastflush == 0)
3046 : {
3047 : /* first call, or block based limits disabled */
3048 514 : WriteRqst.Flush = WriteRqst.Write;
3049 514 : lastflush = now;
3050 : }
3051 8934 : else if (TimestampDifferenceExceeds(lastflush, now, WalWriterDelay))
3052 : {
3053 : /*
3054 : * Flush the writes at least every WalWriterDelay ms. This is
3055 : * important to bound the amount of time it takes for an asynchronous
3056 : * commit to hit disk.
3057 : */
3058 8668 : WriteRqst.Flush = WriteRqst.Write;
3059 8668 : lastflush = now;
3060 : }
3061 266 : else if (flushblocks >= WalWriterFlushAfter)
3062 : {
3063 : /* exceeded wal_writer_flush_after blocks, flush */
3064 240 : WriteRqst.Flush = WriteRqst.Write;
3065 240 : lastflush = now;
3066 : }
3067 : else
3068 : {
3069 : /* no flushing, this time round */
3070 26 : WriteRqst.Flush = 0;
3071 : }
3072 :
3073 : #ifdef WAL_DEBUG
3074 : if (XLOG_DEBUG)
3075 : elog(LOG, "xlog bg flush request write %X/%08X; flush: %X/%08X, current is write %X/%08X; flush %X/%08X",
3076 : LSN_FORMAT_ARGS(WriteRqst.Write),
3077 : LSN_FORMAT_ARGS(WriteRqst.Flush),
3078 : LSN_FORMAT_ARGS(LogwrtResult.Write),
3079 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
3080 : #endif
3081 :
3082 9448 : START_CRIT_SECTION();
3083 :
3084 : /* now wait for any in-progress insertions to finish and get write lock */
3085 9448 : WaitXLogInsertionsToFinish(WriteRqst.Write);
3086 9448 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3087 9448 : RefreshXLogWriteResult(LogwrtResult);
3088 9448 : if (WriteRqst.Write > LogwrtResult.Write ||
3089 760 : WriteRqst.Flush > LogwrtResult.Flush)
3090 : {
3091 9052 : XLogWrite(WriteRqst, insertTLI, flexible);
3092 : }
3093 9448 : LWLockRelease(WALWriteLock);
3094 :
3095 9448 : END_CRIT_SECTION();
3096 :
3097 : /* wake up walsenders now that we've released heavily contended locks */
3098 9448 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3099 :
3100 : /*
3101 : * Great, done. To take some work off the critical path, try to initialize
3102 : * as many of the no-longer-needed WAL buffers for future use as we can.
3103 : */
3104 9448 : AdvanceXLInsertBuffer(InvalidXLogRecPtr, insertTLI, true);
3105 :
3106 : /*
3107 : * If we determined that we need to write data, but somebody else
3108 : * wrote/flushed already, it should be considered as being active, to
3109 : * avoid hibernating too early.
3110 : */
3111 9448 : return true;
3112 : }
3113 :
3114 : /*
3115 : * Test whether XLOG data has been flushed up to (at least) the given
3116 : * position, or whether the minimum recovery point has been updated past
3117 : * the given position.
3118 : *
3119 : * Returns true if a flush is still needed, or if the minimum recovery point
3120 : * must be updated.
3121 : *
3122 : * It is possible that someone else is already in the process of flushing
3123 : * that far, or has updated the minimum recovery point up to the given
3124 : * position.
3125 : */
3126 : bool
3127 17177976 : XLogNeedsFlush(XLogRecPtr record)
3128 : {
3129 : /*
3130 : * During recovery, we don't flush WAL but update minRecoveryPoint
3131 : * instead. So "needs flush" is taken to mean whether minRecoveryPoint
3132 : * would need to be updated.
3133 : *
3134 : * Using XLogInsertAllowed() rather than RecoveryInProgress() matters for
3135 : * the case of an end-of-recovery checkpoint, where WAL data is flushed.
3136 : * This check should be consistent with the one in XLogFlush().
3137 : */
3138 17177976 : if (!XLogInsertAllowed())
3139 : {
3140 : /* Quick exit if already known to be updated or cannot be updated */
3141 1022656 : if (!updateMinRecoveryPoint || record <= LocalMinRecoveryPoint)
3142 1005186 : return false;
3143 :
3144 : /*
3145 : * An invalid minRecoveryPoint means that we need to recover all the
3146 : * WAL, i.e., we're doing crash recovery. We never modify the control
3147 : * file's value in that case, so we can short-circuit future checks
3148 : * here too. This triggers a quick exit path for the startup process,
3149 : * which cannot update its local copy of minRecoveryPoint as long as
3150 : * it has not replayed all WAL available when doing crash recovery.
3151 : */
3152 17470 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint) && InRecovery)
3153 : {
3154 0 : updateMinRecoveryPoint = false;
3155 0 : return false;
3156 : }
3157 :
3158 : /*
3159 : * Update local copy of minRecoveryPoint. But if the lock is busy,
3160 : * just return a conservative guess.
3161 : */
3162 17470 : if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3163 0 : return true;
3164 17470 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
3165 17470 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3166 17470 : LWLockRelease(ControlFileLock);
3167 :
3168 : /*
3169 : * Check minRecoveryPoint for any other process than the startup
3170 : * process doing crash recovery, which should not update the control
3171 : * file value if crash recovery is still running.
3172 : */
3173 17470 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
3174 0 : updateMinRecoveryPoint = false;
3175 :
3176 : /* check again */
3177 17470 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3178 186 : return false;
3179 : else
3180 17284 : return true;
3181 : }
3182 :
3183 : /* Quick exit if already known flushed */
3184 16155320 : if (record <= LogwrtResult.Flush)
3185 15706232 : return false;
3186 :
3187 : /* read LogwrtResult and update local state */
3188 449088 : RefreshXLogWriteResult(LogwrtResult);
3189 :
3190 : /* check again */
3191 449088 : if (record <= LogwrtResult.Flush)
3192 6582 : return false;
3193 :
3194 442506 : return true;
3195 : }
3196 :
3197 : /*
3198 : * Try to make a given XLOG file segment exist.
3199 : *
3200 : * logsegno: identify segment.
3201 : *
3202 : * *added: on return, true if this call raised the number of extant segments.
3203 : *
3204 : * path: on return, this char[MAXPGPATH] has the path to the logsegno file.
3205 : *
3206 : * Returns -1 or FD of opened file. A -1 here is not an error; a caller
3207 : * wanting an open segment should attempt to open "path", which usually will
3208 : * succeed. (This is weird, but it's efficient for the callers.)
3209 : */
3210 : static int
3211 29130 : XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,
3212 : bool *added, char *path)
3213 : {
3214 : char tmppath[MAXPGPATH];
3215 : XLogSegNo installed_segno;
3216 : XLogSegNo max_segno;
3217 : int fd;
3218 : int save_errno;
3219 29130 : int open_flags = O_RDWR | O_CREAT | O_EXCL | PG_BINARY;
3220 : instr_time io_start;
3221 :
3222 : Assert(logtli != 0);
3223 :
3224 29130 : XLogFilePath(path, logtli, logsegno, wal_segment_size);
3225 :
3226 : /*
3227 : * Try to use existent file (checkpoint maker may have created it already)
3228 : */
3229 29130 : *added = false;
3230 29130 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3231 29130 : get_sync_bit(wal_sync_method));
3232 29130 : if (fd < 0)
3233 : {
3234 2952 : if (errno != ENOENT)
3235 0 : ereport(ERROR,
3236 : (errcode_for_file_access(),
3237 : errmsg("could not open file \"%s\": %m", path)));
3238 : }
3239 : else
3240 26178 : return fd;
3241 :
3242 : /*
3243 : * Initialize an empty (all zeroes) segment. NOTE: it is possible that
3244 : * another process is doing the same thing. If so, we will end up
3245 : * pre-creating an extra log segment. That seems OK, and better than
3246 : * holding the lock throughout this lengthy process.
3247 : */
3248 2952 : elog(DEBUG2, "creating and filling new WAL file");
3249 :
3250 2952 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3251 :
3252 2952 : unlink(tmppath);
3253 :
3254 2952 : if (io_direct_flags & IO_DIRECT_WAL_INIT)
3255 0 : open_flags |= PG_O_DIRECT;
3256 :
3257 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3258 2952 : fd = BasicOpenFile(tmppath, open_flags);
3259 2952 : if (fd < 0)
3260 0 : ereport(ERROR,
3261 : (errcode_for_file_access(),
3262 : errmsg("could not create file \"%s\": %m", tmppath)));
3263 :
3264 : /* Measure I/O timing when initializing segment */
3265 2952 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3266 :
3267 2952 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE);
3268 2952 : save_errno = 0;
3269 2952 : if (wal_init_zero)
3270 : {
3271 : ssize_t rc;
3272 :
3273 : /*
3274 : * Zero-fill the file. With this setting, we do this the hard way to
3275 : * ensure that all the file space has really been allocated. On
3276 : * platforms that allow "holes" in files, just seeking to the end
3277 : * doesn't allocate intermediate space. This way, we know that we
3278 : * have all the space and (after the fsync below) that all the
3279 : * indirect blocks are down on disk. Therefore, fdatasync(2) or
3280 : * O_DSYNC will be sufficient to sync future writes to the log file.
3281 : */
3282 2952 : rc = pg_pwrite_zeros(fd, wal_segment_size, 0);
3283 :
3284 2952 : if (rc < 0)
3285 0 : save_errno = errno;
3286 : }
3287 : else
3288 : {
3289 : /*
3290 : * Otherwise, seeking to the end and writing a solitary byte is
3291 : * enough.
3292 : */
3293 0 : errno = 0;
3294 0 : if (pg_pwrite(fd, "\0", 1, wal_segment_size - 1) != 1)
3295 : {
3296 : /* if write didn't set errno, assume no disk space */
3297 0 : save_errno = errno ? errno : ENOSPC;
3298 : }
3299 : }
3300 2952 : pgstat_report_wait_end();
3301 :
3302 : /*
3303 : * A full segment worth of data is written when using wal_init_zero. One
3304 : * byte is written when not using it.
3305 : */
3306 2952 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT, IOOP_WRITE,
3307 : io_start, 1,
3308 2952 : wal_init_zero ? wal_segment_size : 1);
3309 :
3310 2952 : if (save_errno)
3311 : {
3312 : /*
3313 : * If we fail to make the file, delete it to release disk space
3314 : */
3315 0 : unlink(tmppath);
3316 :
3317 0 : close(fd);
3318 :
3319 0 : errno = save_errno;
3320 :
3321 0 : ereport(ERROR,
3322 : (errcode_for_file_access(),
3323 : errmsg("could not write to file \"%s\": %m", tmppath)));
3324 : }
3325 :
3326 : /* Measure I/O timing when flushing segment */
3327 2952 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3328 :
3329 2952 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC);
3330 2952 : if (pg_fsync(fd) != 0)
3331 : {
3332 0 : save_errno = errno;
3333 0 : close(fd);
3334 0 : errno = save_errno;
3335 0 : ereport(ERROR,
3336 : (errcode_for_file_access(),
3337 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3338 : }
3339 2952 : pgstat_report_wait_end();
3340 :
3341 2952 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT,
3342 : IOOP_FSYNC, io_start, 1, 0);
3343 :
3344 2952 : if (close(fd) != 0)
3345 0 : ereport(ERROR,
3346 : (errcode_for_file_access(),
3347 : errmsg("could not close file \"%s\": %m", tmppath)));
3348 :
3349 : /*
3350 : * Now move the segment into place with its final name. Cope with
3351 : * possibility that someone else has created the file while we were
3352 : * filling ours: if so, use ours to pre-create a future log segment.
3353 : */
3354 2952 : installed_segno = logsegno;
3355 :
3356 : /*
3357 : * XXX: What should we use as max_segno? We used to use XLOGfileslop when
3358 : * that was a constant, but that was always a bit dubious: normally, at a
3359 : * checkpoint, XLOGfileslop was the offset from the checkpoint record, but
3360 : * here, it was the offset from the insert location. We can't do the
3361 : * normal XLOGfileslop calculation here because we don't have access to
3362 : * the prior checkpoint's redo location. So somewhat arbitrarily, just use
3363 : * CheckPointSegments.
3364 : */
3365 2952 : max_segno = logsegno + CheckPointSegments;
3366 2952 : if (InstallXLogFileSegment(&installed_segno, tmppath, true, max_segno,
3367 : logtli))
3368 : {
3369 2952 : *added = true;
3370 2952 : elog(DEBUG2, "done creating and filling new WAL file");
3371 : }
3372 : else
3373 : {
3374 : /*
3375 : * No need for any more future segments, or InstallXLogFileSegment()
3376 : * failed to rename the file into place. If the rename failed, a
3377 : * caller opening the file may fail.
3378 : */
3379 0 : unlink(tmppath);
3380 0 : elog(DEBUG2, "abandoned new WAL file");
3381 : }
3382 :
3383 2952 : return -1;
3384 : }
3385 :
3386 : /*
3387 : * Create a new XLOG file segment, or open a pre-existing one.
3388 : *
3389 : * logsegno: identify segment to be created/opened.
3390 : *
3391 : * Returns FD of opened file.
3392 : *
3393 : * Note: errors here are ERROR not PANIC because we might or might not be
3394 : * inside a critical section (eg, during checkpoint there is no reason to
3395 : * take down the system on failure). They will promote to PANIC if we are
3396 : * in a critical section.
3397 : */
3398 : int
3399 28610 : XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)
3400 : {
3401 : bool ignore_added;
3402 : char path[MAXPGPATH];
3403 : int fd;
3404 :
3405 : Assert(logtli != 0);
3406 :
3407 28610 : fd = XLogFileInitInternal(logsegno, logtli, &ignore_added, path);
3408 28610 : if (fd >= 0)
3409 25960 : return fd;
3410 :
3411 : /* Now open original target segment (might not be file I just made) */
3412 2650 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3413 2650 : get_sync_bit(wal_sync_method));
3414 2650 : if (fd < 0)
3415 0 : ereport(ERROR,
3416 : (errcode_for_file_access(),
3417 : errmsg("could not open file \"%s\": %m", path)));
3418 2650 : return fd;
3419 : }
3420 :
3421 : /*
3422 : * Create a new XLOG file segment by copying a pre-existing one.
3423 : *
3424 : * destsegno: identify segment to be created.
3425 : *
3426 : * srcTLI, srcsegno: identify segment to be copied (could be from
3427 : * a different timeline)
3428 : *
3429 : * upto: how much of the source file to copy (the rest is filled with
3430 : * zeros)
3431 : *
3432 : * Currently this is only used during recovery, and so there are no locking
3433 : * considerations. But we should be just as tense as XLogFileInit to avoid
3434 : * emplacing a bogus file.
3435 : */
3436 : static void
3437 82 : XLogFileCopy(TimeLineID destTLI, XLogSegNo destsegno,
3438 : TimeLineID srcTLI, XLogSegNo srcsegno,
3439 : int upto)
3440 : {
3441 : char path[MAXPGPATH];
3442 : char tmppath[MAXPGPATH];
3443 : PGAlignedXLogBlock buffer;
3444 : int srcfd;
3445 : int fd;
3446 : int nbytes;
3447 :
3448 : /*
3449 : * Open the source file
3450 : */
3451 82 : XLogFilePath(path, srcTLI, srcsegno, wal_segment_size);
3452 82 : srcfd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
3453 82 : if (srcfd < 0)
3454 0 : ereport(ERROR,
3455 : (errcode_for_file_access(),
3456 : errmsg("could not open file \"%s\": %m", path)));
3457 :
3458 : /*
3459 : * Copy into a temp file name.
3460 : */
3461 82 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3462 :
3463 82 : unlink(tmppath);
3464 :
3465 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3466 82 : fd = OpenTransientFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
3467 82 : if (fd < 0)
3468 0 : ereport(ERROR,
3469 : (errcode_for_file_access(),
3470 : errmsg("could not create file \"%s\": %m", tmppath)));
3471 :
3472 : /*
3473 : * Do the data copying.
3474 : */
3475 168018 : for (nbytes = 0; nbytes < wal_segment_size; nbytes += sizeof(buffer))
3476 : {
3477 : int nread;
3478 :
3479 167936 : nread = upto - nbytes;
3480 :
3481 : /*
3482 : * The part that is not read from the source file is filled with
3483 : * zeros.
3484 : */
3485 167936 : if (nread < sizeof(buffer))
3486 82 : memset(buffer.data, 0, sizeof(buffer));
3487 :
3488 167936 : if (nread > 0)
3489 : {
3490 : int r;
3491 :
3492 4846 : if (nread > sizeof(buffer))
3493 4764 : nread = sizeof(buffer);
3494 4846 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ);
3495 4846 : r = read(srcfd, buffer.data, nread);
3496 4846 : if (r != nread)
3497 : {
3498 0 : if (r < 0)
3499 0 : ereport(ERROR,
3500 : (errcode_for_file_access(),
3501 : errmsg("could not read file \"%s\": %m",
3502 : path)));
3503 : else
3504 0 : ereport(ERROR,
3505 : (errcode(ERRCODE_DATA_CORRUPTED),
3506 : errmsg("could not read file \"%s\": read %d of %zu",
3507 : path, r, (Size) nread)));
3508 : }
3509 4846 : pgstat_report_wait_end();
3510 : }
3511 167936 : errno = 0;
3512 167936 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_WRITE);
3513 167936 : if ((int) write(fd, buffer.data, sizeof(buffer)) != (int) sizeof(buffer))
3514 : {
3515 0 : int save_errno = errno;
3516 :
3517 : /*
3518 : * If we fail to make the file, delete it to release disk space
3519 : */
3520 0 : unlink(tmppath);
3521 : /* if write didn't set errno, assume problem is no disk space */
3522 0 : errno = save_errno ? save_errno : ENOSPC;
3523 :
3524 0 : ereport(ERROR,
3525 : (errcode_for_file_access(),
3526 : errmsg("could not write to file \"%s\": %m", tmppath)));
3527 : }
3528 167936 : pgstat_report_wait_end();
3529 : }
3530 :
3531 82 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_SYNC);
3532 82 : if (pg_fsync(fd) != 0)
3533 0 : ereport(data_sync_elevel(ERROR),
3534 : (errcode_for_file_access(),
3535 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3536 82 : pgstat_report_wait_end();
3537 :
3538 82 : if (CloseTransientFile(fd) != 0)
3539 0 : ereport(ERROR,
3540 : (errcode_for_file_access(),
3541 : errmsg("could not close file \"%s\": %m", tmppath)));
3542 :
3543 82 : if (CloseTransientFile(srcfd) != 0)
3544 0 : ereport(ERROR,
3545 : (errcode_for_file_access(),
3546 : errmsg("could not close file \"%s\": %m", path)));
3547 :
3548 : /*
3549 : * Now move the segment into place with its final name.
3550 : */
3551 82 : if (!InstallXLogFileSegment(&destsegno, tmppath, false, 0, destTLI))
3552 0 : elog(ERROR, "InstallXLogFileSegment should not have failed");
3553 82 : }
3554 :
3555 : /*
3556 : * Install a new XLOG segment file as a current or future log segment.
3557 : *
3558 : * This is used both to install a newly-created segment (which has a temp
3559 : * filename while it's being created) and to recycle an old segment.
3560 : *
3561 : * *segno: identify segment to install as (or first possible target).
3562 : * When find_free is true, this is modified on return to indicate the
3563 : * actual installation location or last segment searched.
3564 : *
3565 : * tmppath: initial name of file to install. It will be renamed into place.
3566 : *
3567 : * find_free: if true, install the new segment at the first empty segno
3568 : * number at or after the passed numbers. If false, install the new segment
3569 : * exactly where specified, deleting any existing segment file there.
3570 : *
3571 : * max_segno: maximum segment number to install the new file as. Fail if no
3572 : * free slot is found between *segno and max_segno. (Ignored when find_free
3573 : * is false.)
3574 : *
3575 : * tli: The timeline on which the new segment should be installed.
3576 : *
3577 : * Returns true if the file was installed successfully. false indicates that
3578 : * max_segno limit was exceeded, the startup process has disabled this
3579 : * function for now, or an error occurred while renaming the file into place.
3580 : */
3581 : static bool
3582 5794 : InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
3583 : bool find_free, XLogSegNo max_segno, TimeLineID tli)
3584 : {
3585 : char path[MAXPGPATH];
3586 : struct stat stat_buf;
3587 :
3588 : Assert(tli != 0);
3589 :
3590 5794 : XLogFilePath(path, tli, *segno, wal_segment_size);
3591 :
3592 5794 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3593 5794 : if (!XLogCtl->InstallXLogFileSegmentActive)
3594 : {
3595 0 : LWLockRelease(ControlFileLock);
3596 0 : return false;
3597 : }
3598 :
3599 5794 : if (!find_free)
3600 : {
3601 : /* Force installation: get rid of any pre-existing segment file */
3602 82 : durable_unlink(path, DEBUG1);
3603 : }
3604 : else
3605 : {
3606 : /* Find a free slot to put it in */
3607 8294 : while (stat(path, &stat_buf) == 0)
3608 : {
3609 2668 : if ((*segno) >= max_segno)
3610 : {
3611 : /* Failed to find a free slot within specified range */
3612 86 : LWLockRelease(ControlFileLock);
3613 86 : return false;
3614 : }
3615 2582 : (*segno)++;
3616 2582 : XLogFilePath(path, tli, *segno, wal_segment_size);
3617 : }
3618 : }
3619 :
3620 : Assert(access(path, F_OK) != 0 && errno == ENOENT);
3621 5708 : if (durable_rename(tmppath, path, LOG) != 0)
3622 : {
3623 0 : LWLockRelease(ControlFileLock);
3624 : /* durable_rename already emitted log message */
3625 0 : return false;
3626 : }
3627 :
3628 5708 : LWLockRelease(ControlFileLock);
3629 :
3630 5708 : return true;
3631 : }
3632 :
3633 : /*
3634 : * Open a pre-existing logfile segment for writing.
3635 : */
3636 : int
3637 336 : XLogFileOpen(XLogSegNo segno, TimeLineID tli)
3638 : {
3639 : char path[MAXPGPATH];
3640 : int fd;
3641 :
3642 336 : XLogFilePath(path, tli, segno, wal_segment_size);
3643 :
3644 336 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3645 336 : get_sync_bit(wal_sync_method));
3646 336 : if (fd < 0)
3647 0 : ereport(PANIC,
3648 : (errcode_for_file_access(),
3649 : errmsg("could not open file \"%s\": %m", path)));
3650 :
3651 336 : return fd;
3652 : }
3653 :
3654 : /*
3655 : * Close the current logfile segment for writing.
3656 : */
3657 : static void
3658 12720 : XLogFileClose(void)
3659 : {
3660 : Assert(openLogFile >= 0);
3661 :
3662 : /*
3663 : * WAL segment files will not be re-read in normal operation, so we advise
3664 : * the OS to release any cached pages. But do not do so if WAL archiving
3665 : * or streaming is active, because archiver and walsender process could
3666 : * use the cache to read the WAL segment.
3667 : */
3668 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3669 12720 : if (!XLogIsNeeded() && (io_direct_flags & IO_DIRECT_WAL) == 0)
3670 2768 : (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3671 : #endif
3672 :
3673 12720 : if (close(openLogFile) != 0)
3674 : {
3675 : char xlogfname[MAXFNAMELEN];
3676 0 : int save_errno = errno;
3677 :
3678 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo, wal_segment_size);
3679 0 : errno = save_errno;
3680 0 : ereport(PANIC,
3681 : (errcode_for_file_access(),
3682 : errmsg("could not close file \"%s\": %m", xlogfname)));
3683 : }
3684 :
3685 12720 : openLogFile = -1;
3686 12720 : ReleaseExternalFD();
3687 12720 : }
3688 :
3689 : /*
3690 : * Preallocate log files beyond the specified log endpoint.
3691 : *
3692 : * XXX this is currently extremely conservative, since it forces only one
3693 : * future log segment to exist, and even that only if we are 75% done with
3694 : * the current one. This is only appropriate for very low-WAL-volume systems.
3695 : * High-volume systems will be OK once they've built up a sufficient set of
3696 : * recycled log segments, but the startup transient is likely to include
3697 : * a lot of segment creations by foreground processes, which is not so good.
3698 : *
3699 : * XLogFileInitInternal() can ereport(ERROR). All known causes indicate big
3700 : * trouble; for example, a full filesystem is one cause. The checkpoint WAL
3701 : * and/or ControlFile updates already completed. If a RequestCheckpoint()
3702 : * initiated the present checkpoint and an ERROR ends this function, the
3703 : * command that called RequestCheckpoint() fails. That's not ideal, but it's
3704 : * not worth contorting more functions to use caller-specified elevel values.
3705 : * (With or without RequestCheckpoint(), an ERROR forestalls some inessential
3706 : * reporting and resource reclamation.)
3707 : */
3708 : static void
3709 4002 : PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli)
3710 : {
3711 : XLogSegNo _logSegNo;
3712 : int lf;
3713 : bool added;
3714 : char path[MAXPGPATH];
3715 : uint64 offset;
3716 :
3717 4002 : if (!XLogCtl->InstallXLogFileSegmentActive)
3718 16 : return; /* unlocked check says no */
3719 :
3720 3986 : XLByteToPrevSeg(endptr, _logSegNo, wal_segment_size);
3721 3986 : offset = XLogSegmentOffset(endptr - 1, wal_segment_size);
3722 3986 : if (offset >= (uint32) (0.75 * wal_segment_size))
3723 : {
3724 520 : _logSegNo++;
3725 520 : lf = XLogFileInitInternal(_logSegNo, tli, &added, path);
3726 520 : if (lf >= 0)
3727 218 : close(lf);
3728 520 : if (added)
3729 302 : CheckpointStats.ckpt_segs_added++;
3730 : }
3731 : }
3732 :
3733 : /*
3734 : * Throws an error if the given log segment has already been removed or
3735 : * recycled. The caller should only pass a segment that it knows to have
3736 : * existed while the server has been running, as this function always
3737 : * succeeds if no WAL segments have been removed since startup.
3738 : * 'tli' is only used in the error message.
3739 : *
3740 : * Note: this function guarantees to keep errno unchanged on return.
3741 : * This supports callers that use this to possibly deliver a better
3742 : * error message about a missing file, while still being able to throw
3743 : * a normal file-access error afterwards, if this does return.
3744 : */
3745 : void
3746 240102 : CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3747 : {
3748 240102 : int save_errno = errno;
3749 : XLogSegNo lastRemovedSegNo;
3750 :
3751 240102 : SpinLockAcquire(&XLogCtl->info_lck);
3752 240102 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3753 240102 : SpinLockRelease(&XLogCtl->info_lck);
3754 :
3755 240102 : if (segno <= lastRemovedSegNo)
3756 : {
3757 : char filename[MAXFNAMELEN];
3758 :
3759 0 : XLogFileName(filename, tli, segno, wal_segment_size);
3760 0 : errno = save_errno;
3761 0 : ereport(ERROR,
3762 : (errcode_for_file_access(),
3763 : errmsg("requested WAL segment %s has already been removed",
3764 : filename)));
3765 : }
3766 240102 : errno = save_errno;
3767 240102 : }
3768 :
3769 : /*
3770 : * Return the last WAL segment removed, or 0 if no segment has been removed
3771 : * since startup.
3772 : *
3773 : * NB: the result can be out of date arbitrarily fast, the caller has to deal
3774 : * with that.
3775 : */
3776 : XLogSegNo
3777 2212 : XLogGetLastRemovedSegno(void)
3778 : {
3779 : XLogSegNo lastRemovedSegNo;
3780 :
3781 2212 : SpinLockAcquire(&XLogCtl->info_lck);
3782 2212 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3783 2212 : SpinLockRelease(&XLogCtl->info_lck);
3784 :
3785 2212 : return lastRemovedSegNo;
3786 : }
3787 :
3788 : /*
3789 : * Return the oldest WAL segment on the given TLI that still exists in
3790 : * XLOGDIR, or 0 if none.
3791 : */
3792 : XLogSegNo
3793 10 : XLogGetOldestSegno(TimeLineID tli)
3794 : {
3795 : DIR *xldir;
3796 : struct dirent *xlde;
3797 10 : XLogSegNo oldest_segno = 0;
3798 :
3799 10 : xldir = AllocateDir(XLOGDIR);
3800 68 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3801 : {
3802 : TimeLineID file_tli;
3803 : XLogSegNo file_segno;
3804 :
3805 : /* Ignore files that are not XLOG segments. */
3806 58 : if (!IsXLogFileName(xlde->d_name))
3807 40 : continue;
3808 :
3809 : /* Parse filename to get TLI and segno. */
3810 18 : XLogFromFileName(xlde->d_name, &file_tli, &file_segno,
3811 : wal_segment_size);
3812 :
3813 : /* Ignore anything that's not from the TLI of interest. */
3814 18 : if (tli != file_tli)
3815 0 : continue;
3816 :
3817 : /* If it's the oldest so far, update oldest_segno. */
3818 18 : if (oldest_segno == 0 || file_segno < oldest_segno)
3819 16 : oldest_segno = file_segno;
3820 : }
3821 :
3822 10 : FreeDir(xldir);
3823 10 : return oldest_segno;
3824 : }
3825 :
3826 : /*
3827 : * Update the last removed segno pointer in shared memory, to reflect that the
3828 : * given XLOG file has been removed.
3829 : */
3830 : static void
3831 5070 : UpdateLastRemovedPtr(char *filename)
3832 : {
3833 : uint32 tli;
3834 : XLogSegNo segno;
3835 :
3836 5070 : XLogFromFileName(filename, &tli, &segno, wal_segment_size);
3837 :
3838 5070 : SpinLockAcquire(&XLogCtl->info_lck);
3839 5070 : if (segno > XLogCtl->lastRemovedSegNo)
3840 2144 : XLogCtl->lastRemovedSegNo = segno;
3841 5070 : SpinLockRelease(&XLogCtl->info_lck);
3842 5070 : }
3843 :
3844 : /*
3845 : * Remove all temporary log files in pg_wal
3846 : *
3847 : * This is called at the beginning of recovery after a previous crash,
3848 : * at a point where no other processes write fresh WAL data.
3849 : */
3850 : static void
3851 354 : RemoveTempXlogFiles(void)
3852 : {
3853 : DIR *xldir;
3854 : struct dirent *xlde;
3855 :
3856 354 : elog(DEBUG2, "removing all temporary WAL segments");
3857 :
3858 354 : xldir = AllocateDir(XLOGDIR);
3859 2410 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3860 : {
3861 : char path[MAXPGPATH];
3862 :
3863 2056 : if (strncmp(xlde->d_name, "xlogtemp.", 9) != 0)
3864 2056 : continue;
3865 :
3866 0 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
3867 0 : unlink(path);
3868 0 : elog(DEBUG2, "removed temporary WAL segment \"%s\"", path);
3869 : }
3870 354 : FreeDir(xldir);
3871 354 : }
3872 :
3873 : /*
3874 : * Recycle or remove all log files older or equal to passed segno.
3875 : *
3876 : * endptr is current (or recent) end of xlog, and lastredoptr is the
3877 : * redo pointer of the last checkpoint. These are used to determine
3878 : * whether we want to recycle rather than delete no-longer-wanted log files.
3879 : *
3880 : * insertTLI is the current timeline for XLOG insertion. Any recycled
3881 : * segments should be reused for this timeline.
3882 : */
3883 : static void
3884 3468 : RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr, XLogRecPtr endptr,
3885 : TimeLineID insertTLI)
3886 : {
3887 : DIR *xldir;
3888 : struct dirent *xlde;
3889 : char lastoff[MAXFNAMELEN];
3890 : XLogSegNo endlogSegNo;
3891 : XLogSegNo recycleSegNo;
3892 :
3893 : /* Initialize info about where to try to recycle to */
3894 3468 : XLByteToSeg(endptr, endlogSegNo, wal_segment_size);
3895 3468 : recycleSegNo = XLOGfileslop(lastredoptr);
3896 :
3897 : /*
3898 : * Construct a filename of the last segment to be kept. The timeline ID
3899 : * doesn't matter, we ignore that in the comparison. (During recovery,
3900 : * InsertTimeLineID isn't set, so we can't use that.)
3901 : */
3902 3468 : XLogFileName(lastoff, 0, segno, wal_segment_size);
3903 :
3904 3468 : elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
3905 : lastoff);
3906 :
3907 3468 : xldir = AllocateDir(XLOGDIR);
3908 :
3909 87218 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3910 : {
3911 : /* Ignore files that are not XLOG segments */
3912 83750 : if (!IsXLogFileName(xlde->d_name) &&
3913 14968 : !IsPartialXLogFileName(xlde->d_name))
3914 14960 : continue;
3915 :
3916 : /*
3917 : * We ignore the timeline part of the XLOG segment identifiers in
3918 : * deciding whether a segment is still needed. This ensures that we
3919 : * won't prematurely remove a segment from a parent timeline. We could
3920 : * probably be a little more proactive about removing segments of
3921 : * non-parent timelines, but that would be a whole lot more
3922 : * complicated.
3923 : *
3924 : * We use the alphanumeric sorting property of the filenames to decide
3925 : * which ones are earlier than the lastoff segment.
3926 : */
3927 68790 : if (strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
3928 : {
3929 43720 : if (XLogArchiveCheckDone(xlde->d_name))
3930 : {
3931 : /* Update the last removed location in shared memory first */
3932 5070 : UpdateLastRemovedPtr(xlde->d_name);
3933 :
3934 5070 : RemoveXlogFile(xlde, recycleSegNo, &endlogSegNo, insertTLI);
3935 : }
3936 : }
3937 : }
3938 :
3939 3468 : FreeDir(xldir);
3940 3468 : }
3941 :
3942 : /*
3943 : * Recycle or remove WAL files that are not part of the given timeline's
3944 : * history.
3945 : *
3946 : * This is called during recovery, whenever we switch to follow a new
3947 : * timeline, and at the end of recovery when we create a new timeline. We
3948 : * wouldn't otherwise care about extra WAL files lying in pg_wal, but they
3949 : * might be leftover pre-allocated or recycled WAL segments on the old timeline
3950 : * that we haven't used yet, and contain garbage. If we just leave them in
3951 : * pg_wal, they will eventually be archived, and we can't let that happen.
3952 : * Files that belong to our timeline history are valid, because we have
3953 : * successfully replayed them, but from others we can't be sure.
3954 : *
3955 : * 'switchpoint' is the current point in WAL where we switch to new timeline,
3956 : * and 'newTLI' is the new timeline we switch to.
3957 : */
3958 : void
3959 122 : RemoveNonParentXlogFiles(XLogRecPtr switchpoint, TimeLineID newTLI)
3960 : {
3961 : DIR *xldir;
3962 : struct dirent *xlde;
3963 : char switchseg[MAXFNAMELEN];
3964 : XLogSegNo endLogSegNo;
3965 : XLogSegNo switchLogSegNo;
3966 : XLogSegNo recycleSegNo;
3967 :
3968 : /*
3969 : * Initialize info about where to begin the work. This will recycle,
3970 : * somewhat arbitrarily, 10 future segments.
3971 : */
3972 122 : XLByteToPrevSeg(switchpoint, switchLogSegNo, wal_segment_size);
3973 122 : XLByteToSeg(switchpoint, endLogSegNo, wal_segment_size);
3974 122 : recycleSegNo = endLogSegNo + 10;
3975 :
3976 : /*
3977 : * Construct a filename of the last segment to be kept.
3978 : */
3979 122 : XLogFileName(switchseg, newTLI, switchLogSegNo, wal_segment_size);
3980 :
3981 122 : elog(DEBUG2, "attempting to remove WAL segments newer than log file %s",
3982 : switchseg);
3983 :
3984 122 : xldir = AllocateDir(XLOGDIR);
3985 :
3986 1170 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3987 : {
3988 : /* Ignore files that are not XLOG segments */
3989 1048 : if (!IsXLogFileName(xlde->d_name))
3990 646 : continue;
3991 :
3992 : /*
3993 : * Remove files that are on a timeline older than the new one we're
3994 : * switching to, but with a segment number >= the first segment on the
3995 : * new timeline.
3996 : */
3997 402 : if (strncmp(xlde->d_name, switchseg, 8) < 0 &&
3998 264 : strcmp(xlde->d_name + 8, switchseg + 8) > 0)
3999 : {
4000 : /*
4001 : * If the file has already been marked as .ready, however, don't
4002 : * remove it yet. It should be OK to remove it - files that are
4003 : * not part of our timeline history are not required for recovery
4004 : * - but seems safer to let them be archived and removed later.
4005 : */
4006 32 : if (!XLogArchiveIsReady(xlde->d_name))
4007 32 : RemoveXlogFile(xlde, recycleSegNo, &endLogSegNo, newTLI);
4008 : }
4009 : }
4010 :
4011 122 : FreeDir(xldir);
4012 122 : }
4013 :
4014 : /*
4015 : * Recycle or remove a log file that's no longer needed.
4016 : *
4017 : * segment_de is the dirent structure of the segment to recycle or remove.
4018 : * recycleSegNo is the segment number to recycle up to. endlogSegNo is
4019 : * the segment number of the current (or recent) end of WAL.
4020 : *
4021 : * endlogSegNo gets incremented if the segment is recycled so as it is not
4022 : * checked again with future callers of this function.
4023 : *
4024 : * insertTLI is the current timeline for XLOG insertion. Any recycled segments
4025 : * should be used for this timeline.
4026 : */
4027 : static void
4028 5102 : RemoveXlogFile(const struct dirent *segment_de,
4029 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
4030 : TimeLineID insertTLI)
4031 : {
4032 : char path[MAXPGPATH];
4033 : #ifdef WIN32
4034 : char newpath[MAXPGPATH];
4035 : #endif
4036 5102 : const char *segname = segment_de->d_name;
4037 :
4038 5102 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", segname);
4039 :
4040 : /*
4041 : * Before deleting the file, see if it can be recycled as a future log
4042 : * segment. Only recycle normal files, because we don't want to recycle
4043 : * symbolic links pointing to a separate archive directory.
4044 : */
4045 5102 : if (wal_recycle &&
4046 5102 : *endlogSegNo <= recycleSegNo &&
4047 6186 : XLogCtl->InstallXLogFileSegmentActive && /* callee rechecks this */
4048 5520 : get_dirent_type(path, segment_de, false, DEBUG2) == PGFILETYPE_REG &&
4049 2760 : InstallXLogFileSegment(endlogSegNo, path,
4050 : true, recycleSegNo, insertTLI))
4051 : {
4052 2674 : ereport(DEBUG2,
4053 : (errmsg_internal("recycled write-ahead log file \"%s\"",
4054 : segname)));
4055 2674 : CheckpointStats.ckpt_segs_recycled++;
4056 : /* Needn't recheck that slot on future iterations */
4057 2674 : (*endlogSegNo)++;
4058 : }
4059 : else
4060 : {
4061 : /* No need for any more future segments, or recycling failed ... */
4062 : int rc;
4063 :
4064 2428 : ereport(DEBUG2,
4065 : (errmsg_internal("removing write-ahead log file \"%s\"",
4066 : segname)));
4067 :
4068 : #ifdef WIN32
4069 :
4070 : /*
4071 : * On Windows, if another process (e.g another backend) holds the file
4072 : * open in FILE_SHARE_DELETE mode, unlink will succeed, but the file
4073 : * will still show up in directory listing until the last handle is
4074 : * closed. To avoid confusing the lingering deleted file for a live
4075 : * WAL file that needs to be archived, rename it before deleting it.
4076 : *
4077 : * If another process holds the file open without FILE_SHARE_DELETE
4078 : * flag, rename will fail. We'll try again at the next checkpoint.
4079 : */
4080 : snprintf(newpath, MAXPGPATH, "%s.deleted", path);
4081 : if (rename(path, newpath) != 0)
4082 : {
4083 : ereport(LOG,
4084 : (errcode_for_file_access(),
4085 : errmsg("could not rename file \"%s\": %m",
4086 : path)));
4087 : return;
4088 : }
4089 : rc = durable_unlink(newpath, LOG);
4090 : #else
4091 2428 : rc = durable_unlink(path, LOG);
4092 : #endif
4093 2428 : if (rc != 0)
4094 : {
4095 : /* Message already logged by durable_unlink() */
4096 0 : return;
4097 : }
4098 2428 : CheckpointStats.ckpt_segs_removed++;
4099 : }
4100 :
4101 5102 : XLogArchiveCleanup(segname);
4102 : }
4103 :
4104 : /*
4105 : * Verify whether pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
4106 : * If the latter do not exist, recreate them.
4107 : *
4108 : * It is not the goal of this function to verify the contents of these
4109 : * directories, but to help in cases where someone has performed a cluster
4110 : * copy for PITR purposes but omitted pg_wal from the copy.
4111 : *
4112 : * We could also recreate pg_wal if it doesn't exist, but a deliberate
4113 : * policy decision was made not to. It is fairly common for pg_wal to be
4114 : * a symlink, and if that was the DBA's intent then automatically making a
4115 : * plain directory would result in degraded performance with no notice.
4116 : */
4117 : static void
4118 1916 : ValidateXLOGDirectoryStructure(void)
4119 : {
4120 : char path[MAXPGPATH];
4121 : struct stat stat_buf;
4122 :
4123 : /* Check for pg_wal; if it doesn't exist, error out */
4124 1916 : if (stat(XLOGDIR, &stat_buf) != 0 ||
4125 1916 : !S_ISDIR(stat_buf.st_mode))
4126 0 : ereport(FATAL,
4127 : (errcode_for_file_access(),
4128 : errmsg("required WAL directory \"%s\" does not exist",
4129 : XLOGDIR)));
4130 :
4131 : /* Check for archive_status */
4132 1916 : snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4133 1916 : if (stat(path, &stat_buf) == 0)
4134 : {
4135 : /* Check for weird cases where it exists but isn't a directory */
4136 1914 : if (!S_ISDIR(stat_buf.st_mode))
4137 0 : ereport(FATAL,
4138 : (errcode_for_file_access(),
4139 : errmsg("required WAL directory \"%s\" does not exist",
4140 : path)));
4141 : }
4142 : else
4143 : {
4144 2 : ereport(LOG,
4145 : (errmsg("creating missing WAL directory \"%s\"", path)));
4146 2 : if (MakePGDirectory(path) < 0)
4147 0 : ereport(FATAL,
4148 : (errcode_for_file_access(),
4149 : errmsg("could not create missing directory \"%s\": %m",
4150 : path)));
4151 : }
4152 :
4153 : /* Check for summaries */
4154 1916 : snprintf(path, MAXPGPATH, XLOGDIR "/summaries");
4155 1916 : if (stat(path, &stat_buf) == 0)
4156 : {
4157 : /* Check for weird cases where it exists but isn't a directory */
4158 1914 : if (!S_ISDIR(stat_buf.st_mode))
4159 0 : ereport(FATAL,
4160 : (errmsg("required WAL directory \"%s\" does not exist",
4161 : path)));
4162 : }
4163 : else
4164 : {
4165 2 : ereport(LOG,
4166 : (errmsg("creating missing WAL directory \"%s\"", path)));
4167 2 : if (MakePGDirectory(path) < 0)
4168 0 : ereport(FATAL,
4169 : (errmsg("could not create missing directory \"%s\": %m",
4170 : path)));
4171 : }
4172 1916 : }
4173 :
4174 : /*
4175 : * Remove previous backup history files. This also retries creation of
4176 : * .ready files for any backup history files for which XLogArchiveNotify
4177 : * failed earlier.
4178 : */
4179 : static void
4180 300 : CleanupBackupHistory(void)
4181 : {
4182 : DIR *xldir;
4183 : struct dirent *xlde;
4184 : char path[MAXPGPATH + sizeof(XLOGDIR)];
4185 :
4186 300 : xldir = AllocateDir(XLOGDIR);
4187 :
4188 3016 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4189 : {
4190 2416 : if (IsBackupHistoryFileName(xlde->d_name))
4191 : {
4192 316 : if (XLogArchiveCheckDone(xlde->d_name))
4193 : {
4194 250 : elog(DEBUG2, "removing WAL backup history file \"%s\"",
4195 : xlde->d_name);
4196 250 : snprintf(path, sizeof(path), XLOGDIR "/%s", xlde->d_name);
4197 250 : unlink(path);
4198 250 : XLogArchiveCleanup(xlde->d_name);
4199 : }
4200 : }
4201 : }
4202 :
4203 300 : FreeDir(xldir);
4204 300 : }
4205 :
4206 : /*
4207 : * I/O routines for pg_control
4208 : *
4209 : * *ControlFile is a buffer in shared memory that holds an image of the
4210 : * contents of pg_control. WriteControlFile() initializes pg_control
4211 : * given a preloaded buffer, ReadControlFile() loads the buffer from
4212 : * the pg_control file (during postmaster or standalone-backend startup),
4213 : * and UpdateControlFile() rewrites pg_control after we modify xlog state.
4214 : * InitControlFile() fills the buffer with initial values.
4215 : *
4216 : * For simplicity, WriteControlFile() initializes the fields of pg_control
4217 : * that are related to checking backend/database compatibility, and
4218 : * ReadControlFile() verifies they are correct. We could split out the
4219 : * I/O and compatibility-check functions, but there seems no need currently.
4220 : */
4221 :
4222 : static void
4223 100 : InitControlFile(uint64 sysidentifier, uint32 data_checksum_version)
4224 : {
4225 : char mock_auth_nonce[MOCK_AUTH_NONCE_LEN];
4226 :
4227 : /*
4228 : * Generate a random nonce. This is used for authentication requests that
4229 : * will fail because the user does not exist. The nonce is used to create
4230 : * a genuine-looking password challenge for the non-existent user, in lieu
4231 : * of an actual stored password.
4232 : */
4233 100 : if (!pg_strong_random(mock_auth_nonce, MOCK_AUTH_NONCE_LEN))
4234 0 : ereport(PANIC,
4235 : (errcode(ERRCODE_INTERNAL_ERROR),
4236 : errmsg("could not generate secret authorization token")));
4237 :
4238 100 : memset(ControlFile, 0, sizeof(ControlFileData));
4239 : /* Initialize pg_control status fields */
4240 100 : ControlFile->system_identifier = sysidentifier;
4241 100 : memcpy(ControlFile->mock_authentication_nonce, mock_auth_nonce, MOCK_AUTH_NONCE_LEN);
4242 100 : ControlFile->state = DB_SHUTDOWNED;
4243 100 : ControlFile->unloggedLSN = FirstNormalUnloggedLSN;
4244 :
4245 : /* Set important parameter values for use when replaying WAL */
4246 100 : ControlFile->MaxConnections = MaxConnections;
4247 100 : ControlFile->max_worker_processes = max_worker_processes;
4248 100 : ControlFile->max_wal_senders = max_wal_senders;
4249 100 : ControlFile->max_prepared_xacts = max_prepared_xacts;
4250 100 : ControlFile->max_locks_per_xact = max_locks_per_xact;
4251 100 : ControlFile->wal_level = wal_level;
4252 100 : ControlFile->wal_log_hints = wal_log_hints;
4253 100 : ControlFile->track_commit_timestamp = track_commit_timestamp;
4254 100 : ControlFile->data_checksum_version = data_checksum_version;
4255 100 : }
4256 :
4257 : static void
4258 100 : WriteControlFile(void)
4259 : {
4260 : int fd;
4261 : char buffer[PG_CONTROL_FILE_SIZE]; /* need not be aligned */
4262 :
4263 : /*
4264 : * Initialize version and compatibility-check fields
4265 : */
4266 100 : ControlFile->pg_control_version = PG_CONTROL_VERSION;
4267 100 : ControlFile->catalog_version_no = CATALOG_VERSION_NO;
4268 :
4269 100 : ControlFile->maxAlign = MAXIMUM_ALIGNOF;
4270 100 : ControlFile->floatFormat = FLOATFORMAT_VALUE;
4271 :
4272 100 : ControlFile->blcksz = BLCKSZ;
4273 100 : ControlFile->relseg_size = RELSEG_SIZE;
4274 100 : ControlFile->slru_pages_per_segment = SLRU_PAGES_PER_SEGMENT;
4275 100 : ControlFile->xlog_blcksz = XLOG_BLCKSZ;
4276 100 : ControlFile->xlog_seg_size = wal_segment_size;
4277 :
4278 100 : ControlFile->nameDataLen = NAMEDATALEN;
4279 100 : ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
4280 :
4281 100 : ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
4282 100 : ControlFile->loblksize = LOBLKSIZE;
4283 :
4284 100 : ControlFile->float8ByVal = true; /* vestigial */
4285 :
4286 : /*
4287 : * Initialize the default 'char' signedness.
4288 : *
4289 : * The signedness of the char type is implementation-defined. For instance
4290 : * on x86 architecture CPUs, the char data type is typically treated as
4291 : * signed by default, whereas on aarch architecture CPUs, it is typically
4292 : * treated as unsigned by default. In v17 or earlier, we accidentally let
4293 : * C implementation signedness affect persistent data. This led to
4294 : * inconsistent results when comparing char data across different
4295 : * platforms.
4296 : *
4297 : * This flag can be used as a hint to ensure consistent behavior for
4298 : * pre-v18 data files that store data sorted by the 'char' type on disk,
4299 : * especially in cross-platform replication scenarios.
4300 : *
4301 : * Newly created database clusters unconditionally set the default char
4302 : * signedness to true. pg_upgrade changes this flag for clusters that were
4303 : * initialized on signedness=false platforms. As a result,
4304 : * signedness=false setting will become rare over time. If we had known
4305 : * about this problem during the last development cycle that forced initdb
4306 : * (v8.3), we would have made all clusters signed or all clusters
4307 : * unsigned. Making pg_upgrade the only source of signedness=false will
4308 : * cause the population of database clusters to converge toward that
4309 : * retrospective ideal.
4310 : */
4311 100 : ControlFile->default_char_signedness = true;
4312 :
4313 : /* Contents are protected with a CRC */
4314 100 : INIT_CRC32C(ControlFile->crc);
4315 100 : COMP_CRC32C(ControlFile->crc,
4316 : ControlFile,
4317 : offsetof(ControlFileData, crc));
4318 100 : FIN_CRC32C(ControlFile->crc);
4319 :
4320 : /*
4321 : * We write out PG_CONTROL_FILE_SIZE bytes into pg_control, zero-padding
4322 : * the excess over sizeof(ControlFileData). This reduces the odds of
4323 : * premature-EOF errors when reading pg_control. We'll still fail when we
4324 : * check the contents of the file, but hopefully with a more specific
4325 : * error than "couldn't read pg_control".
4326 : */
4327 100 : memset(buffer, 0, PG_CONTROL_FILE_SIZE);
4328 100 : memcpy(buffer, ControlFile, sizeof(ControlFileData));
4329 :
4330 100 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4331 : O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
4332 100 : if (fd < 0)
4333 0 : ereport(PANIC,
4334 : (errcode_for_file_access(),
4335 : errmsg("could not create file \"%s\": %m",
4336 : XLOG_CONTROL_FILE)));
4337 :
4338 100 : errno = 0;
4339 100 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_WRITE);
4340 100 : if (write(fd, buffer, PG_CONTROL_FILE_SIZE) != PG_CONTROL_FILE_SIZE)
4341 : {
4342 : /* if write didn't set errno, assume problem is no disk space */
4343 0 : if (errno == 0)
4344 0 : errno = ENOSPC;
4345 0 : ereport(PANIC,
4346 : (errcode_for_file_access(),
4347 : errmsg("could not write to file \"%s\": %m",
4348 : XLOG_CONTROL_FILE)));
4349 : }
4350 100 : pgstat_report_wait_end();
4351 :
4352 100 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_SYNC);
4353 100 : if (pg_fsync(fd) != 0)
4354 0 : ereport(PANIC,
4355 : (errcode_for_file_access(),
4356 : errmsg("could not fsync file \"%s\": %m",
4357 : XLOG_CONTROL_FILE)));
4358 100 : pgstat_report_wait_end();
4359 :
4360 100 : if (close(fd) != 0)
4361 0 : ereport(PANIC,
4362 : (errcode_for_file_access(),
4363 : errmsg("could not close file \"%s\": %m",
4364 : XLOG_CONTROL_FILE)));
4365 100 : }
4366 :
4367 : static void
4368 2016 : ReadControlFile(void)
4369 : {
4370 : pg_crc32c crc;
4371 : int fd;
4372 : char wal_segsz_str[20];
4373 : int r;
4374 :
4375 : /*
4376 : * Read data...
4377 : */
4378 2016 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4379 : O_RDWR | PG_BINARY);
4380 2016 : if (fd < 0)
4381 0 : ereport(PANIC,
4382 : (errcode_for_file_access(),
4383 : errmsg("could not open file \"%s\": %m",
4384 : XLOG_CONTROL_FILE)));
4385 :
4386 2016 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ);
4387 2016 : r = read(fd, ControlFile, sizeof(ControlFileData));
4388 2016 : if (r != sizeof(ControlFileData))
4389 : {
4390 0 : if (r < 0)
4391 0 : ereport(PANIC,
4392 : (errcode_for_file_access(),
4393 : errmsg("could not read file \"%s\": %m",
4394 : XLOG_CONTROL_FILE)));
4395 : else
4396 0 : ereport(PANIC,
4397 : (errcode(ERRCODE_DATA_CORRUPTED),
4398 : errmsg("could not read file \"%s\": read %d of %zu",
4399 : XLOG_CONTROL_FILE, r, sizeof(ControlFileData))));
4400 : }
4401 2016 : pgstat_report_wait_end();
4402 :
4403 2016 : close(fd);
4404 :
4405 : /*
4406 : * Check for expected pg_control format version. If this is wrong, the
4407 : * CRC check will likely fail because we'll be checking the wrong number
4408 : * of bytes. Complaining about wrong version will probably be more
4409 : * enlightening than complaining about wrong CRC.
4410 : */
4411 :
4412 2016 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
4413 0 : ereport(FATAL,
4414 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4415 : errmsg("database files are incompatible with server"),
4416 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4417 : " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4418 : ControlFile->pg_control_version, ControlFile->pg_control_version,
4419 : PG_CONTROL_VERSION, PG_CONTROL_VERSION),
4420 : errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4421 :
4422 2016 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
4423 0 : ereport(FATAL,
4424 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4425 : errmsg("database files are incompatible with server"),
4426 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4427 : " but the server was compiled with PG_CONTROL_VERSION %d.",
4428 : ControlFile->pg_control_version, PG_CONTROL_VERSION),
4429 : errhint("It looks like you need to initdb.")));
4430 :
4431 : /* Now check the CRC. */
4432 2016 : INIT_CRC32C(crc);
4433 2016 : COMP_CRC32C(crc,
4434 : ControlFile,
4435 : offsetof(ControlFileData, crc));
4436 2016 : FIN_CRC32C(crc);
4437 :
4438 2016 : if (!EQ_CRC32C(crc, ControlFile->crc))
4439 0 : ereport(FATAL,
4440 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4441 : errmsg("incorrect checksum in control file")));
4442 :
4443 : /*
4444 : * Do compatibility checking immediately. If the database isn't
4445 : * compatible with the backend executable, we want to abort before we can
4446 : * possibly do any damage.
4447 : */
4448 2016 : if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
4449 0 : ereport(FATAL,
4450 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4451 : errmsg("database files are incompatible with server"),
4452 : /* translator: %s is a variable name and %d is its value */
4453 : errdetail("The database cluster was initialized with %s %d,"
4454 : " but the server was compiled with %s %d.",
4455 : "CATALOG_VERSION_NO", ControlFile->catalog_version_no,
4456 : "CATALOG_VERSION_NO", CATALOG_VERSION_NO),
4457 : errhint("It looks like you need to initdb.")));
4458 2016 : if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
4459 0 : ereport(FATAL,
4460 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4461 : errmsg("database files are incompatible with server"),
4462 : /* translator: %s is a variable name and %d is its value */
4463 : errdetail("The database cluster was initialized with %s %d,"
4464 : " but the server was compiled with %s %d.",
4465 : "MAXALIGN", ControlFile->maxAlign,
4466 : "MAXALIGN", MAXIMUM_ALIGNOF),
4467 : errhint("It looks like you need to initdb.")));
4468 2016 : if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
4469 0 : ereport(FATAL,
4470 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4471 : errmsg("database files are incompatible with server"),
4472 : errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4473 : errhint("It looks like you need to initdb.")));
4474 2016 : if (ControlFile->blcksz != BLCKSZ)
4475 0 : ereport(FATAL,
4476 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4477 : errmsg("database files are incompatible with server"),
4478 : /* translator: %s is a variable name and %d is its value */
4479 : errdetail("The database cluster was initialized with %s %d,"
4480 : " but the server was compiled with %s %d.",
4481 : "BLCKSZ", ControlFile->blcksz,
4482 : "BLCKSZ", BLCKSZ),
4483 : errhint("It looks like you need to recompile or initdb.")));
4484 2016 : if (ControlFile->relseg_size != RELSEG_SIZE)
4485 0 : ereport(FATAL,
4486 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4487 : errmsg("database files are incompatible with server"),
4488 : /* translator: %s is a variable name and %d is its value */
4489 : errdetail("The database cluster was initialized with %s %d,"
4490 : " but the server was compiled with %s %d.",
4491 : "RELSEG_SIZE", ControlFile->relseg_size,
4492 : "RELSEG_SIZE", RELSEG_SIZE),
4493 : errhint("It looks like you need to recompile or initdb.")));
4494 2016 : if (ControlFile->slru_pages_per_segment != SLRU_PAGES_PER_SEGMENT)
4495 0 : ereport(FATAL,
4496 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4497 : errmsg("database files are incompatible with server"),
4498 : /* translator: %s is a variable name and %d is its value */
4499 : errdetail("The database cluster was initialized with %s %d,"
4500 : " but the server was compiled with %s %d.",
4501 : "SLRU_PAGES_PER_SEGMENT", ControlFile->slru_pages_per_segment,
4502 : "SLRU_PAGES_PER_SEGMENT", SLRU_PAGES_PER_SEGMENT),
4503 : errhint("It looks like you need to recompile or initdb.")));
4504 2016 : if (ControlFile->xlog_blcksz != XLOG_BLCKSZ)
4505 0 : ereport(FATAL,
4506 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4507 : errmsg("database files are incompatible with server"),
4508 : /* translator: %s is a variable name and %d is its value */
4509 : errdetail("The database cluster was initialized with %s %d,"
4510 : " but the server was compiled with %s %d.",
4511 : "XLOG_BLCKSZ", ControlFile->xlog_blcksz,
4512 : "XLOG_BLCKSZ", XLOG_BLCKSZ),
4513 : errhint("It looks like you need to recompile or initdb.")));
4514 2016 : if (ControlFile->nameDataLen != NAMEDATALEN)
4515 0 : ereport(FATAL,
4516 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4517 : errmsg("database files are incompatible with server"),
4518 : /* translator: %s is a variable name and %d is its value */
4519 : errdetail("The database cluster was initialized with %s %d,"
4520 : " but the server was compiled with %s %d.",
4521 : "NAMEDATALEN", ControlFile->nameDataLen,
4522 : "NAMEDATALEN", NAMEDATALEN),
4523 : errhint("It looks like you need to recompile or initdb.")));
4524 2016 : if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
4525 0 : ereport(FATAL,
4526 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4527 : errmsg("database files are incompatible with server"),
4528 : /* translator: %s is a variable name and %d is its value */
4529 : errdetail("The database cluster was initialized with %s %d,"
4530 : " but the server was compiled with %s %d.",
4531 : "INDEX_MAX_KEYS", ControlFile->indexMaxKeys,
4532 : "INDEX_MAX_KEYS", INDEX_MAX_KEYS),
4533 : errhint("It looks like you need to recompile or initdb.")));
4534 2016 : if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
4535 0 : ereport(FATAL,
4536 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4537 : errmsg("database files are incompatible with server"),
4538 : /* translator: %s is a variable name and %d is its value */
4539 : errdetail("The database cluster was initialized with %s %d,"
4540 : " but the server was compiled with %s %d.",
4541 : "TOAST_MAX_CHUNK_SIZE", ControlFile->toast_max_chunk_size,
4542 : "TOAST_MAX_CHUNK_SIZE", (int) TOAST_MAX_CHUNK_SIZE),
4543 : errhint("It looks like you need to recompile or initdb.")));
4544 2016 : if (ControlFile->loblksize != LOBLKSIZE)
4545 0 : ereport(FATAL,
4546 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4547 : errmsg("database files are incompatible with server"),
4548 : /* translator: %s is a variable name and %d is its value */
4549 : errdetail("The database cluster was initialized with %s %d,"
4550 : " but the server was compiled with %s %d.",
4551 : "LOBLKSIZE", ControlFile->loblksize,
4552 : "LOBLKSIZE", (int) LOBLKSIZE),
4553 : errhint("It looks like you need to recompile or initdb.")));
4554 :
4555 : Assert(ControlFile->float8ByVal); /* vestigial, not worth an error msg */
4556 :
4557 2016 : wal_segment_size = ControlFile->xlog_seg_size;
4558 :
4559 2016 : if (!IsValidWalSegSize(wal_segment_size))
4560 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4561 : errmsg_plural("invalid WAL segment size in control file (%d byte)",
4562 : "invalid WAL segment size in control file (%d bytes)",
4563 : wal_segment_size,
4564 : wal_segment_size),
4565 : errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.")));
4566 :
4567 2016 : snprintf(wal_segsz_str, sizeof(wal_segsz_str), "%d", wal_segment_size);
4568 2016 : SetConfigOption("wal_segment_size", wal_segsz_str, PGC_INTERNAL,
4569 : PGC_S_DYNAMIC_DEFAULT);
4570 :
4571 : /* check and update variables dependent on wal_segment_size */
4572 2016 : if (ConvertToXSegs(min_wal_size_mb, wal_segment_size) < 2)
4573 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4574 : /* translator: both %s are GUC names */
4575 : errmsg("\"%s\" must be at least twice \"%s\"",
4576 : "min_wal_size", "wal_segment_size")));
4577 :
4578 2016 : if (ConvertToXSegs(max_wal_size_mb, wal_segment_size) < 2)
4579 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4580 : /* translator: both %s are GUC names */
4581 : errmsg("\"%s\" must be at least twice \"%s\"",
4582 : "max_wal_size", "wal_segment_size")));
4583 :
4584 2016 : UsableBytesInSegment =
4585 2016 : (wal_segment_size / XLOG_BLCKSZ * UsableBytesInPage) -
4586 : (SizeOfXLogLongPHD - SizeOfXLogShortPHD);
4587 :
4588 2016 : CalculateCheckpointSegments();
4589 :
4590 : /* Make the initdb settings visible as GUC variables, too */
4591 2016 : SetConfigOption("data_checksums", DataChecksumsEnabled() ? "yes" : "no",
4592 : PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
4593 2016 : }
4594 :
4595 : /*
4596 : * Utility wrapper to update the control file. Note that the control
4597 : * file gets flushed.
4598 : */
4599 : static void
4600 18330 : UpdateControlFile(void)
4601 : {
4602 18330 : update_controlfile(DataDir, ControlFile, true);
4603 18330 : }
4604 :
4605 : /*
4606 : * Returns the unique system identifier from control file.
4607 : */
4608 : uint64
4609 2820 : GetSystemIdentifier(void)
4610 : {
4611 : Assert(ControlFile != NULL);
4612 2820 : return ControlFile->system_identifier;
4613 : }
4614 :
4615 : /*
4616 : * Returns the random nonce from control file.
4617 : */
4618 : char *
4619 2 : GetMockAuthenticationNonce(void)
4620 : {
4621 : Assert(ControlFile != NULL);
4622 2 : return ControlFile->mock_authentication_nonce;
4623 : }
4624 :
4625 : /*
4626 : * Are checksums enabled for data pages?
4627 : */
4628 : bool
4629 20235984 : DataChecksumsEnabled(void)
4630 : {
4631 : Assert(ControlFile != NULL);
4632 20235984 : return (ControlFile->data_checksum_version > 0);
4633 : }
4634 :
4635 : /*
4636 : * Return true if the cluster was initialized on a platform where the
4637 : * default signedness of char is "signed". This function exists for code
4638 : * that deals with pre-v18 data files that store data sorted by the 'char'
4639 : * type on disk (e.g., GIN and GiST indexes). See the comments in
4640 : * WriteControlFile() for details.
4641 : */
4642 : bool
4643 6 : GetDefaultCharSignedness(void)
4644 : {
4645 6 : return ControlFile->default_char_signedness;
4646 : }
4647 :
4648 : /*
4649 : * Returns a fake LSN for unlogged relations.
4650 : *
4651 : * Each call generates an LSN that is greater than any previous value
4652 : * returned. The current counter value is saved and restored across clean
4653 : * shutdowns, but like unlogged relations, does not survive a crash. This can
4654 : * be used in lieu of real LSN values returned by XLogInsert, if you need an
4655 : * LSN-like increasing sequence of numbers without writing any WAL.
4656 : */
4657 : XLogRecPtr
4658 66 : GetFakeLSNForUnloggedRel(void)
4659 : {
4660 66 : return pg_atomic_fetch_add_u64(&XLogCtl->unloggedLSN, 1);
4661 : }
4662 :
4663 : /*
4664 : * Auto-tune the number of XLOG buffers.
4665 : *
4666 : * The preferred setting for wal_buffers is about 3% of shared_buffers, with
4667 : * a maximum of one XLOG segment (there is little reason to think that more
4668 : * is helpful, at least so long as we force an fsync when switching log files)
4669 : * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
4670 : * 9.1, when auto-tuning was added).
4671 : *
4672 : * This should not be called until NBuffers has received its final value.
4673 : */
4674 : static int
4675 2196 : XLOGChooseNumBuffers(void)
4676 : {
4677 : int xbuffers;
4678 :
4679 2196 : xbuffers = NBuffers / 32;
4680 2196 : if (xbuffers > (wal_segment_size / XLOG_BLCKSZ))
4681 48 : xbuffers = (wal_segment_size / XLOG_BLCKSZ);
4682 2196 : if (xbuffers < 8)
4683 854 : xbuffers = 8;
4684 2196 : return xbuffers;
4685 : }
4686 :
4687 : /*
4688 : * GUC check_hook for wal_buffers
4689 : */
4690 : bool
4691 4472 : check_wal_buffers(int *newval, void **extra, GucSource source)
4692 : {
4693 : /*
4694 : * -1 indicates a request for auto-tune.
4695 : */
4696 4472 : if (*newval == -1)
4697 : {
4698 : /*
4699 : * If we haven't yet changed the boot_val default of -1, just let it
4700 : * be. We'll fix it when XLOGShmemSize is called.
4701 : */
4702 2276 : if (XLOGbuffers == -1)
4703 2276 : return true;
4704 :
4705 : /* Otherwise, substitute the auto-tune value */
4706 0 : *newval = XLOGChooseNumBuffers();
4707 : }
4708 :
4709 : /*
4710 : * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
4711 : * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
4712 : * the case, we just silently treat such values as a request for the
4713 : * minimum. (We could throw an error instead, but that doesn't seem very
4714 : * helpful.)
4715 : */
4716 2196 : if (*newval < 4)
4717 0 : *newval = 4;
4718 :
4719 2196 : return true;
4720 : }
4721 :
4722 : /*
4723 : * GUC check_hook for wal_consistency_checking
4724 : */
4725 : bool
4726 4072 : check_wal_consistency_checking(char **newval, void **extra, GucSource source)
4727 : {
4728 : char *rawstring;
4729 : List *elemlist;
4730 : ListCell *l;
4731 : bool newwalconsistency[RM_MAX_ID + 1];
4732 :
4733 : /* Initialize the array */
4734 134376 : MemSet(newwalconsistency, 0, (RM_MAX_ID + 1) * sizeof(bool));
4735 :
4736 : /* Need a modifiable copy of string */
4737 4072 : rawstring = pstrdup(*newval);
4738 :
4739 : /* Parse string into list of identifiers */
4740 4072 : if (!SplitIdentifierString(rawstring, ',', &elemlist))
4741 : {
4742 : /* syntax error in list */
4743 0 : GUC_check_errdetail("List syntax is invalid.");
4744 0 : pfree(rawstring);
4745 0 : list_free(elemlist);
4746 0 : return false;
4747 : }
4748 :
4749 4976 : foreach(l, elemlist)
4750 : {
4751 904 : char *tok = (char *) lfirst(l);
4752 : int rmid;
4753 :
4754 : /* Check for 'all'. */
4755 904 : if (pg_strcasecmp(tok, "all") == 0)
4756 : {
4757 231300 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4758 230400 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL)
4759 9000 : newwalconsistency[rmid] = true;
4760 : }
4761 : else
4762 : {
4763 : /* Check if the token matches any known resource manager. */
4764 4 : bool found = false;
4765 :
4766 72 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4767 : {
4768 108 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL &&
4769 36 : pg_strcasecmp(tok, GetRmgr(rmid).rm_name) == 0)
4770 : {
4771 4 : newwalconsistency[rmid] = true;
4772 4 : found = true;
4773 4 : break;
4774 : }
4775 : }
4776 4 : if (!found)
4777 : {
4778 : /*
4779 : * During startup, it might be a not-yet-loaded custom
4780 : * resource manager. Defer checking until
4781 : * InitializeWalConsistencyChecking().
4782 : */
4783 0 : if (!process_shared_preload_libraries_done)
4784 : {
4785 0 : check_wal_consistency_checking_deferred = true;
4786 : }
4787 : else
4788 : {
4789 0 : GUC_check_errdetail("Unrecognized key word: \"%s\".", tok);
4790 0 : pfree(rawstring);
4791 0 : list_free(elemlist);
4792 0 : return false;
4793 : }
4794 : }
4795 : }
4796 : }
4797 :
4798 4072 : pfree(rawstring);
4799 4072 : list_free(elemlist);
4800 :
4801 : /* assign new value */
4802 4072 : *extra = guc_malloc(LOG, (RM_MAX_ID + 1) * sizeof(bool));
4803 4072 : if (!*extra)
4804 0 : return false;
4805 4072 : memcpy(*extra, newwalconsistency, (RM_MAX_ID + 1) * sizeof(bool));
4806 4072 : return true;
4807 : }
4808 :
4809 : /*
4810 : * GUC assign_hook for wal_consistency_checking
4811 : */
4812 : void
4813 4070 : assign_wal_consistency_checking(const char *newval, void *extra)
4814 : {
4815 : /*
4816 : * If some checks were deferred, it's possible that the checks will fail
4817 : * later during InitializeWalConsistencyChecking(). But in that case, the
4818 : * postmaster will exit anyway, so it's safe to proceed with the
4819 : * assignment.
4820 : *
4821 : * Any built-in resource managers specified are assigned immediately,
4822 : * which affects WAL created before shared_preload_libraries are
4823 : * processed. Any custom resource managers specified won't be assigned
4824 : * until after shared_preload_libraries are processed, but that's OK
4825 : * because WAL for a custom resource manager can't be written before the
4826 : * module is loaded anyway.
4827 : */
4828 4070 : wal_consistency_checking = extra;
4829 4070 : }
4830 :
4831 : /*
4832 : * InitializeWalConsistencyChecking: run after loading custom resource managers
4833 : *
4834 : * If any unknown resource managers were specified in the
4835 : * wal_consistency_checking GUC, processing was deferred. Now that
4836 : * shared_preload_libraries have been loaded, process wal_consistency_checking
4837 : * again.
4838 : */
4839 : void
4840 1896 : InitializeWalConsistencyChecking(void)
4841 : {
4842 : Assert(process_shared_preload_libraries_done);
4843 :
4844 1896 : if (check_wal_consistency_checking_deferred)
4845 : {
4846 : struct config_generic *guc;
4847 :
4848 0 : guc = find_option("wal_consistency_checking", false, false, ERROR);
4849 :
4850 0 : check_wal_consistency_checking_deferred = false;
4851 :
4852 0 : set_config_option_ext("wal_consistency_checking",
4853 : wal_consistency_checking_string,
4854 : guc->scontext, guc->source, guc->srole,
4855 : GUC_ACTION_SET, true, ERROR, false);
4856 :
4857 : /* checking should not be deferred again */
4858 : Assert(!check_wal_consistency_checking_deferred);
4859 : }
4860 1896 : }
4861 :
4862 : /*
4863 : * GUC show_hook for archive_command
4864 : */
4865 : const char *
4866 3478 : show_archive_command(void)
4867 : {
4868 3478 : if (XLogArchivingActive())
4869 4 : return XLogArchiveCommand;
4870 : else
4871 3474 : return "(disabled)";
4872 : }
4873 :
4874 : /*
4875 : * GUC show_hook for in_hot_standby
4876 : */
4877 : const char *
4878 30668 : show_in_hot_standby(void)
4879 : {
4880 : /*
4881 : * We display the actual state based on shared memory, so that this GUC
4882 : * reports up-to-date state if examined intra-query. The underlying
4883 : * variable (in_hot_standby_guc) changes only when we transmit a new value
4884 : * to the client.
4885 : */
4886 30668 : return RecoveryInProgress() ? "on" : "off";
4887 : }
4888 :
4889 : /*
4890 : * Read the control file, set respective GUCs.
4891 : *
4892 : * This is to be called during startup, including a crash recovery cycle,
4893 : * unless in bootstrap mode, where no control file yet exists. As there's no
4894 : * usable shared memory yet (its sizing can depend on the contents of the
4895 : * control file!), first store the contents in local memory. XLOGShmemInit()
4896 : * will then copy it to shared memory later.
4897 : *
4898 : * reset just controls whether previous contents are to be expected (in the
4899 : * reset case, there's a dangling pointer into old shared memory), or not.
4900 : */
4901 : void
4902 1916 : LocalProcessControlFile(bool reset)
4903 : {
4904 : Assert(reset || ControlFile == NULL);
4905 1916 : ControlFile = palloc(sizeof(ControlFileData));
4906 1916 : ReadControlFile();
4907 1916 : }
4908 :
4909 : /*
4910 : * Get the wal_level from the control file. For a standby, this value should be
4911 : * considered as its active wal_level, because it may be different from what
4912 : * was originally configured on standby.
4913 : */
4914 : WalLevel
4915 138 : GetActiveWalLevelOnStandby(void)
4916 : {
4917 138 : return ControlFile->wal_level;
4918 : }
4919 :
4920 : /*
4921 : * Initialization of shared memory for XLOG
4922 : */
4923 : Size
4924 6300 : XLOGShmemSize(void)
4925 : {
4926 : Size size;
4927 :
4928 : /*
4929 : * If the value of wal_buffers is -1, use the preferred auto-tune value.
4930 : * This isn't an amazingly clean place to do this, but we must wait till
4931 : * NBuffers has received its final value, and must do it before using the
4932 : * value of XLOGbuffers to do anything important.
4933 : *
4934 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
4935 : * However, if the DBA explicitly set wal_buffers = -1 in the config file,
4936 : * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
4937 : * the matter with PGC_S_OVERRIDE.
4938 : */
4939 6300 : if (XLOGbuffers == -1)
4940 : {
4941 : char buf[32];
4942 :
4943 2196 : snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
4944 2196 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4945 : PGC_S_DYNAMIC_DEFAULT);
4946 2196 : if (XLOGbuffers == -1) /* failed to apply it? */
4947 0 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4948 : PGC_S_OVERRIDE);
4949 : }
4950 : Assert(XLOGbuffers > 0);
4951 :
4952 : /* XLogCtl */
4953 6300 : size = sizeof(XLogCtlData);
4954 :
4955 : /* WAL insertion locks, plus alignment */
4956 6300 : size = add_size(size, mul_size(sizeof(WALInsertLockPadded), NUM_XLOGINSERT_LOCKS + 1));
4957 : /* xlblocks array */
4958 6300 : size = add_size(size, mul_size(sizeof(pg_atomic_uint64), XLOGbuffers));
4959 : /* extra alignment padding for XLOG I/O buffers */
4960 6300 : size = add_size(size, Max(XLOG_BLCKSZ, PG_IO_ALIGN_SIZE));
4961 : /* and the buffers themselves */
4962 6300 : size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
4963 :
4964 : /*
4965 : * Note: we don't count ControlFileData, it comes out of the "slop factor"
4966 : * added by CreateSharedMemoryAndSemaphores. This lets us use this
4967 : * routine again below to compute the actual allocation size.
4968 : */
4969 :
4970 6300 : return size;
4971 : }
4972 :
4973 : void
4974 2200 : XLOGShmemInit(void)
4975 : {
4976 : bool foundCFile,
4977 : foundXLog;
4978 : char *allocptr;
4979 : int i;
4980 : ControlFileData *localControlFile;
4981 :
4982 : #ifdef WAL_DEBUG
4983 :
4984 : /*
4985 : * Create a memory context for WAL debugging that's exempt from the normal
4986 : * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
4987 : * an allocation fails, but wal_debug is not for production use anyway.
4988 : */
4989 : if (walDebugCxt == NULL)
4990 : {
4991 : walDebugCxt = AllocSetContextCreate(TopMemoryContext,
4992 : "WAL Debug",
4993 : ALLOCSET_DEFAULT_SIZES);
4994 : MemoryContextAllowInCriticalSection(walDebugCxt, true);
4995 : }
4996 : #endif
4997 :
4998 :
4999 2200 : XLogCtl = (XLogCtlData *)
5000 2200 : ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog);
5001 :
5002 2200 : localControlFile = ControlFile;
5003 2200 : ControlFile = (ControlFileData *)
5004 2200 : ShmemInitStruct("Control File", sizeof(ControlFileData), &foundCFile);
5005 :
5006 2200 : if (foundCFile || foundXLog)
5007 : {
5008 : /* both should be present or neither */
5009 : Assert(foundCFile && foundXLog);
5010 :
5011 : /* Initialize local copy of WALInsertLocks */
5012 0 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks;
5013 :
5014 0 : if (localControlFile)
5015 0 : pfree(localControlFile);
5016 0 : return;
5017 : }
5018 2200 : memset(XLogCtl, 0, sizeof(XLogCtlData));
5019 :
5020 : /*
5021 : * Already have read control file locally, unless in bootstrap mode. Move
5022 : * contents into shared memory.
5023 : */
5024 2200 : if (localControlFile)
5025 : {
5026 1900 : memcpy(ControlFile, localControlFile, sizeof(ControlFileData));
5027 1900 : pfree(localControlFile);
5028 : }
5029 :
5030 : /*
5031 : * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
5032 : * multiple of the alignment for same, so no extra alignment padding is
5033 : * needed here.
5034 : */
5035 2200 : allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
5036 2200 : XLogCtl->xlblocks = (pg_atomic_uint64 *) allocptr;
5037 2200 : allocptr += sizeof(pg_atomic_uint64) * XLOGbuffers;
5038 :
5039 628628 : for (i = 0; i < XLOGbuffers; i++)
5040 : {
5041 626428 : pg_atomic_init_u64(&XLogCtl->xlblocks[i], InvalidXLogRecPtr);
5042 : }
5043 :
5044 : /* WAL insertion locks. Ensure they're aligned to the full padded size */
5045 2200 : allocptr += sizeof(WALInsertLockPadded) -
5046 2200 : ((uintptr_t) allocptr) % sizeof(WALInsertLockPadded);
5047 2200 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks =
5048 : (WALInsertLockPadded *) allocptr;
5049 2200 : allocptr += sizeof(WALInsertLockPadded) * NUM_XLOGINSERT_LOCKS;
5050 :
5051 19800 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
5052 : {
5053 17600 : LWLockInitialize(&WALInsertLocks[i].l.lock, LWTRANCHE_WAL_INSERT);
5054 17600 : pg_atomic_init_u64(&WALInsertLocks[i].l.insertingAt, InvalidXLogRecPtr);
5055 17600 : WALInsertLocks[i].l.lastImportantAt = InvalidXLogRecPtr;
5056 : }
5057 :
5058 : /*
5059 : * Align the start of the page buffers to a full xlog block size boundary.
5060 : * This simplifies some calculations in XLOG insertion. It is also
5061 : * required for O_DIRECT.
5062 : */
5063 2200 : allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5064 2200 : XLogCtl->pages = allocptr;
5065 2200 : memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5066 :
5067 : /*
5068 : * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5069 : * in additional info.)
5070 : */
5071 2200 : XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5072 2200 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5073 2200 : XLogCtl->InstallXLogFileSegmentActive = false;
5074 2200 : XLogCtl->WalWriterSleeping = false;
5075 :
5076 2200 : SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5077 2200 : SpinLockInit(&XLogCtl->info_lck);
5078 2200 : pg_atomic_init_u64(&XLogCtl->logInsertResult, InvalidXLogRecPtr);
5079 2200 : pg_atomic_init_u64(&XLogCtl->logWriteResult, InvalidXLogRecPtr);
5080 2200 : pg_atomic_init_u64(&XLogCtl->logFlushResult, InvalidXLogRecPtr);
5081 2200 : pg_atomic_init_u64(&XLogCtl->unloggedLSN, InvalidXLogRecPtr);
5082 : }
5083 :
5084 : /*
5085 : * This func must be called ONCE on system install. It creates pg_control
5086 : * and the initial XLOG segment.
5087 : */
5088 : void
5089 100 : BootStrapXLOG(uint32 data_checksum_version)
5090 : {
5091 : CheckPoint checkPoint;
5092 : char *buffer;
5093 : XLogPageHeader page;
5094 : XLogLongPageHeader longpage;
5095 : XLogRecord *record;
5096 : char *recptr;
5097 : uint64 sysidentifier;
5098 : struct timeval tv;
5099 : pg_crc32c crc;
5100 :
5101 : /* allow ordinary WAL segment creation, like StartupXLOG() would */
5102 100 : SetInstallXLogFileSegmentActive();
5103 :
5104 : /*
5105 : * Select a hopefully-unique system identifier code for this installation.
5106 : * We use the result of gettimeofday(), including the fractional seconds
5107 : * field, as being about as unique as we can easily get. (Think not to
5108 : * use random(), since it hasn't been seeded and there's no portable way
5109 : * to seed it other than the system clock value...) The upper half of the
5110 : * uint64 value is just the tv_sec part, while the lower half contains the
5111 : * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
5112 : * PID for a little extra uniqueness. A person knowing this encoding can
5113 : * determine the initialization time of the installation, which could
5114 : * perhaps be useful sometimes.
5115 : */
5116 100 : gettimeofday(&tv, NULL);
5117 100 : sysidentifier = ((uint64) tv.tv_sec) << 32;
5118 100 : sysidentifier |= ((uint64) tv.tv_usec) << 12;
5119 100 : sysidentifier |= getpid() & 0xFFF;
5120 :
5121 : /* page buffer must be aligned suitably for O_DIRECT */
5122 100 : buffer = (char *) palloc(XLOG_BLCKSZ + XLOG_BLCKSZ);
5123 100 : page = (XLogPageHeader) TYPEALIGN(XLOG_BLCKSZ, buffer);
5124 100 : memset(page, 0, XLOG_BLCKSZ);
5125 :
5126 : /*
5127 : * Set up information for the initial checkpoint record
5128 : *
5129 : * The initial checkpoint record is written to the beginning of the WAL
5130 : * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5131 : * used, so that we can use 0/0 to mean "before any valid WAL segment".
5132 : */
5133 100 : checkPoint.redo = wal_segment_size + SizeOfXLogLongPHD;
5134 100 : checkPoint.ThisTimeLineID = BootstrapTimeLineID;
5135 100 : checkPoint.PrevTimeLineID = BootstrapTimeLineID;
5136 100 : checkPoint.fullPageWrites = fullPageWrites;
5137 100 : checkPoint.wal_level = wal_level;
5138 : checkPoint.nextXid =
5139 100 : FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId);
5140 100 : checkPoint.nextOid = FirstGenbkiObjectId;
5141 100 : checkPoint.nextMulti = FirstMultiXactId;
5142 100 : checkPoint.nextMultiOffset = 0;
5143 100 : checkPoint.oldestXid = FirstNormalTransactionId;
5144 100 : checkPoint.oldestXidDB = Template1DbOid;
5145 100 : checkPoint.oldestMulti = FirstMultiXactId;
5146 100 : checkPoint.oldestMultiDB = Template1DbOid;
5147 100 : checkPoint.oldestCommitTsXid = InvalidTransactionId;
5148 100 : checkPoint.newestCommitTsXid = InvalidTransactionId;
5149 100 : checkPoint.time = (pg_time_t) time(NULL);
5150 100 : checkPoint.oldestActiveXid = InvalidTransactionId;
5151 :
5152 100 : TransamVariables->nextXid = checkPoint.nextXid;
5153 100 : TransamVariables->nextOid = checkPoint.nextOid;
5154 100 : TransamVariables->oidCount = 0;
5155 100 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5156 100 : AdvanceOldestClogXid(checkPoint.oldestXid);
5157 100 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5158 100 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5159 100 : SetCommitTsLimit(InvalidTransactionId, InvalidTransactionId);
5160 :
5161 : /* Set up the XLOG page header */
5162 100 : page->xlp_magic = XLOG_PAGE_MAGIC;
5163 100 : page->xlp_info = XLP_LONG_HEADER;
5164 100 : page->xlp_tli = BootstrapTimeLineID;
5165 100 : page->xlp_pageaddr = wal_segment_size;
5166 100 : longpage = (XLogLongPageHeader) page;
5167 100 : longpage->xlp_sysid = sysidentifier;
5168 100 : longpage->xlp_seg_size = wal_segment_size;
5169 100 : longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5170 :
5171 : /* Insert the initial checkpoint record */
5172 100 : recptr = ((char *) page + SizeOfXLogLongPHD);
5173 100 : record = (XLogRecord *) recptr;
5174 100 : record->xl_prev = 0;
5175 100 : record->xl_xid = InvalidTransactionId;
5176 100 : record->xl_tot_len = SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(checkPoint);
5177 100 : record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5178 100 : record->xl_rmid = RM_XLOG_ID;
5179 100 : recptr += SizeOfXLogRecord;
5180 : /* fill the XLogRecordDataHeaderShort struct */
5181 100 : *(recptr++) = (char) XLR_BLOCK_ID_DATA_SHORT;
5182 100 : *(recptr++) = sizeof(checkPoint);
5183 100 : memcpy(recptr, &checkPoint, sizeof(checkPoint));
5184 100 : recptr += sizeof(checkPoint);
5185 : Assert(recptr - (char *) record == record->xl_tot_len);
5186 :
5187 100 : INIT_CRC32C(crc);
5188 100 : COMP_CRC32C(crc, ((char *) record) + SizeOfXLogRecord, record->xl_tot_len - SizeOfXLogRecord);
5189 100 : COMP_CRC32C(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5190 100 : FIN_CRC32C(crc);
5191 100 : record->xl_crc = crc;
5192 :
5193 : /* Create first XLOG segment file */
5194 100 : openLogTLI = BootstrapTimeLineID;
5195 100 : openLogFile = XLogFileInit(1, BootstrapTimeLineID);
5196 :
5197 : /*
5198 : * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
5199 : * close the file again in a moment.
5200 : */
5201 :
5202 : /* Write the first page with the initial record */
5203 100 : errno = 0;
5204 100 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE);
5205 100 : if (write(openLogFile, page, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5206 : {
5207 : /* if write didn't set errno, assume problem is no disk space */
5208 0 : if (errno == 0)
5209 0 : errno = ENOSPC;
5210 0 : ereport(PANIC,
5211 : (errcode_for_file_access(),
5212 : errmsg("could not write bootstrap write-ahead log file: %m")));
5213 : }
5214 100 : pgstat_report_wait_end();
5215 :
5216 100 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC);
5217 100 : if (pg_fsync(openLogFile) != 0)
5218 0 : ereport(PANIC,
5219 : (errcode_for_file_access(),
5220 : errmsg("could not fsync bootstrap write-ahead log file: %m")));
5221 100 : pgstat_report_wait_end();
5222 :
5223 100 : if (close(openLogFile) != 0)
5224 0 : ereport(PANIC,
5225 : (errcode_for_file_access(),
5226 : errmsg("could not close bootstrap write-ahead log file: %m")));
5227 :
5228 100 : openLogFile = -1;
5229 :
5230 : /* Now create pg_control */
5231 100 : InitControlFile(sysidentifier, data_checksum_version);
5232 100 : ControlFile->time = checkPoint.time;
5233 100 : ControlFile->checkPoint = checkPoint.redo;
5234 100 : ControlFile->checkPointCopy = checkPoint;
5235 :
5236 : /* some additional ControlFile fields are set in WriteControlFile() */
5237 100 : WriteControlFile();
5238 :
5239 : /* Bootstrap the commit log, too */
5240 100 : BootStrapCLOG();
5241 100 : BootStrapCommitTs();
5242 100 : BootStrapSUBTRANS();
5243 100 : BootStrapMultiXact();
5244 :
5245 100 : pfree(buffer);
5246 :
5247 : /*
5248 : * Force control file to be read - in contrast to normal processing we'd
5249 : * otherwise never run the checks and GUC related initializations therein.
5250 : */
5251 100 : ReadControlFile();
5252 100 : }
5253 :
5254 : static char *
5255 1696 : str_time(pg_time_t tnow, char *buf, size_t bufsize)
5256 : {
5257 1696 : pg_strftime(buf, bufsize,
5258 : "%Y-%m-%d %H:%M:%S %Z",
5259 1696 : pg_localtime(&tnow, log_timezone));
5260 :
5261 1696 : return buf;
5262 : }
5263 :
5264 : /*
5265 : * Initialize the first WAL segment on new timeline.
5266 : */
5267 : static void
5268 100 : XLogInitNewTimeline(TimeLineID endTLI, XLogRecPtr endOfLog, TimeLineID newTLI)
5269 : {
5270 : char xlogfname[MAXFNAMELEN];
5271 : XLogSegNo endLogSegNo;
5272 : XLogSegNo startLogSegNo;
5273 :
5274 : /* we always switch to a new timeline after archive recovery */
5275 : Assert(endTLI != newTLI);
5276 :
5277 : /*
5278 : * Update min recovery point one last time.
5279 : */
5280 100 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5281 :
5282 : /*
5283 : * Calculate the last segment on the old timeline, and the first segment
5284 : * on the new timeline. If the switch happens in the middle of a segment,
5285 : * they are the same, but if the switch happens exactly at a segment
5286 : * boundary, startLogSegNo will be endLogSegNo + 1.
5287 : */
5288 100 : XLByteToPrevSeg(endOfLog, endLogSegNo, wal_segment_size);
5289 100 : XLByteToSeg(endOfLog, startLogSegNo, wal_segment_size);
5290 :
5291 : /*
5292 : * Initialize the starting WAL segment for the new timeline. If the switch
5293 : * happens in the middle of a segment, copy data from the last WAL segment
5294 : * of the old timeline up to the switch point, to the starting WAL segment
5295 : * on the new timeline.
5296 : */
5297 100 : if (endLogSegNo == startLogSegNo)
5298 : {
5299 : /*
5300 : * Make a copy of the file on the new timeline.
5301 : *
5302 : * Writing WAL isn't allowed yet, so there are no locking
5303 : * considerations. But we should be just as tense as XLogFileInit to
5304 : * avoid emplacing a bogus file.
5305 : */
5306 82 : XLogFileCopy(newTLI, endLogSegNo, endTLI, endLogSegNo,
5307 82 : XLogSegmentOffset(endOfLog, wal_segment_size));
5308 : }
5309 : else
5310 : {
5311 : /*
5312 : * The switch happened at a segment boundary, so just create the next
5313 : * segment on the new timeline.
5314 : */
5315 : int fd;
5316 :
5317 18 : fd = XLogFileInit(startLogSegNo, newTLI);
5318 :
5319 18 : if (close(fd) != 0)
5320 : {
5321 0 : int save_errno = errno;
5322 :
5323 0 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5324 0 : errno = save_errno;
5325 0 : ereport(ERROR,
5326 : (errcode_for_file_access(),
5327 : errmsg("could not close file \"%s\": %m", xlogfname)));
5328 : }
5329 : }
5330 :
5331 : /*
5332 : * Let's just make real sure there are not .ready or .done flags posted
5333 : * for the new segment.
5334 : */
5335 100 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5336 100 : XLogArchiveCleanup(xlogfname);
5337 100 : }
5338 :
5339 : /*
5340 : * Perform cleanup actions at the conclusion of archive recovery.
5341 : */
5342 : static void
5343 100 : CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI, XLogRecPtr EndOfLog,
5344 : TimeLineID newTLI)
5345 : {
5346 : /*
5347 : * Execute the recovery_end_command, if any.
5348 : */
5349 100 : if (recoveryEndCommand && strcmp(recoveryEndCommand, "") != 0)
5350 4 : ExecuteRecoveryCommand(recoveryEndCommand,
5351 : "recovery_end_command",
5352 : true,
5353 : WAIT_EVENT_RECOVERY_END_COMMAND);
5354 :
5355 : /*
5356 : * We switched to a new timeline. Clean up segments on the old timeline.
5357 : *
5358 : * If there are any higher-numbered segments on the old timeline, remove
5359 : * them. They might contain valid WAL, but they might also be
5360 : * pre-allocated files containing garbage. In any case, they are not part
5361 : * of the new timeline's history so we don't need them.
5362 : */
5363 100 : RemoveNonParentXlogFiles(EndOfLog, newTLI);
5364 :
5365 : /*
5366 : * If the switch happened in the middle of a segment, what to do with the
5367 : * last, partial segment on the old timeline? If we don't archive it, and
5368 : * the server that created the WAL never archives it either (e.g. because
5369 : * it was hit by a meteor), it will never make it to the archive. That's
5370 : * OK from our point of view, because the new segment that we created with
5371 : * the new TLI contains all the WAL from the old timeline up to the switch
5372 : * point. But if you later try to do PITR to the "missing" WAL on the old
5373 : * timeline, recovery won't find it in the archive. It's physically
5374 : * present in the new file with new TLI, but recovery won't look there
5375 : * when it's recovering to the older timeline. On the other hand, if we
5376 : * archive the partial segment, and the original server on that timeline
5377 : * is still running and archives the completed version of the same segment
5378 : * later, it will fail. (We used to do that in 9.4 and below, and it
5379 : * caused such problems).
5380 : *
5381 : * As a compromise, we rename the last segment with the .partial suffix,
5382 : * and archive it. Archive recovery will never try to read .partial
5383 : * segments, so they will normally go unused. But in the odd PITR case,
5384 : * the administrator can copy them manually to the pg_wal directory
5385 : * (removing the suffix). They can be useful in debugging, too.
5386 : *
5387 : * If a .done or .ready file already exists for the old timeline, however,
5388 : * we had already determined that the segment is complete, so we can let
5389 : * it be archived normally. (In particular, if it was restored from the
5390 : * archive to begin with, it's expected to have a .done file).
5391 : */
5392 100 : if (XLogSegmentOffset(EndOfLog, wal_segment_size) != 0 &&
5393 : XLogArchivingActive())
5394 : {
5395 : char origfname[MAXFNAMELEN];
5396 : XLogSegNo endLogSegNo;
5397 :
5398 20 : XLByteToPrevSeg(EndOfLog, endLogSegNo, wal_segment_size);
5399 20 : XLogFileName(origfname, EndOfLogTLI, endLogSegNo, wal_segment_size);
5400 :
5401 20 : if (!XLogArchiveIsReadyOrDone(origfname))
5402 : {
5403 : char origpath[MAXPGPATH];
5404 : char partialfname[MAXFNAMELEN];
5405 : char partialpath[MAXPGPATH];
5406 :
5407 : /*
5408 : * If we're summarizing WAL, we can't rename the partial file
5409 : * until the summarizer finishes with it, else it will fail.
5410 : */
5411 12 : if (summarize_wal)
5412 2 : WaitForWalSummarization(EndOfLog);
5413 :
5414 12 : XLogFilePath(origpath, EndOfLogTLI, endLogSegNo, wal_segment_size);
5415 12 : snprintf(partialfname, MAXFNAMELEN, "%s.partial", origfname);
5416 12 : snprintf(partialpath, MAXPGPATH, "%s.partial", origpath);
5417 :
5418 : /*
5419 : * Make sure there's no .done or .ready file for the .partial
5420 : * file.
5421 : */
5422 12 : XLogArchiveCleanup(partialfname);
5423 :
5424 12 : durable_rename(origpath, partialpath, ERROR);
5425 12 : XLogArchiveNotify(partialfname);
5426 : }
5427 : }
5428 100 : }
5429 :
5430 : /*
5431 : * Check to see if required parameters are set high enough on this server
5432 : * for various aspects of recovery operation.
5433 : *
5434 : * Note that all the parameters which this function tests need to be
5435 : * listed in Administrator's Overview section in high-availability.sgml.
5436 : * If you change them, don't forget to update the list.
5437 : */
5438 : static void
5439 494 : CheckRequiredParameterValues(void)
5440 : {
5441 : /*
5442 : * For archive recovery, the WAL must be generated with at least 'replica'
5443 : * wal_level.
5444 : */
5445 494 : if (ArchiveRecoveryRequested && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5446 : {
5447 4 : ereport(FATAL,
5448 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5449 : errmsg("WAL was generated with \"wal_level=minimal\", cannot continue recovering"),
5450 : errdetail("This happens if you temporarily set \"wal_level=minimal\" on the server."),
5451 : errhint("Use a backup taken after setting \"wal_level\" to higher than \"minimal\".")));
5452 : }
5453 :
5454 : /*
5455 : * For Hot Standby, the WAL must be generated with 'replica' mode, and we
5456 : * must have at least as many backend slots as the primary.
5457 : */
5458 490 : if (ArchiveRecoveryRequested && EnableHotStandby)
5459 : {
5460 : /* We ignore autovacuum_worker_slots when we make this test. */
5461 248 : RecoveryRequiresIntParameter("max_connections",
5462 : MaxConnections,
5463 248 : ControlFile->MaxConnections);
5464 248 : RecoveryRequiresIntParameter("max_worker_processes",
5465 : max_worker_processes,
5466 248 : ControlFile->max_worker_processes);
5467 248 : RecoveryRequiresIntParameter("max_wal_senders",
5468 : max_wal_senders,
5469 248 : ControlFile->max_wal_senders);
5470 248 : RecoveryRequiresIntParameter("max_prepared_transactions",
5471 : max_prepared_xacts,
5472 248 : ControlFile->max_prepared_xacts);
5473 248 : RecoveryRequiresIntParameter("max_locks_per_transaction",
5474 : max_locks_per_xact,
5475 248 : ControlFile->max_locks_per_xact);
5476 : }
5477 490 : }
5478 :
5479 : /*
5480 : * This must be called ONCE during postmaster or standalone-backend startup
5481 : */
5482 : void
5483 1916 : StartupXLOG(void)
5484 : {
5485 : XLogCtlInsert *Insert;
5486 : CheckPoint checkPoint;
5487 : bool wasShutdown;
5488 : bool didCrash;
5489 : bool haveTblspcMap;
5490 : bool haveBackupLabel;
5491 : XLogRecPtr EndOfLog;
5492 : TimeLineID EndOfLogTLI;
5493 : TimeLineID newTLI;
5494 : bool performedWalRecovery;
5495 : EndOfWalRecoveryInfo *endOfRecoveryInfo;
5496 : XLogRecPtr abortedRecPtr;
5497 : XLogRecPtr missingContrecPtr;
5498 : TransactionId oldestActiveXID;
5499 1916 : bool promoted = false;
5500 : char timebuf[128];
5501 :
5502 : /*
5503 : * We should have an aux process resource owner to use, and we should not
5504 : * be in a transaction that's installed some other resowner.
5505 : */
5506 : Assert(AuxProcessResourceOwner != NULL);
5507 : Assert(CurrentResourceOwner == NULL ||
5508 : CurrentResourceOwner == AuxProcessResourceOwner);
5509 1916 : CurrentResourceOwner = AuxProcessResourceOwner;
5510 :
5511 : /*
5512 : * Check that contents look valid.
5513 : */
5514 1916 : if (!XRecOffIsValid(ControlFile->checkPoint))
5515 0 : ereport(FATAL,
5516 : (errcode(ERRCODE_DATA_CORRUPTED),
5517 : errmsg("control file contains invalid checkpoint location")));
5518 :
5519 1916 : switch (ControlFile->state)
5520 : {
5521 1498 : case DB_SHUTDOWNED:
5522 :
5523 : /*
5524 : * This is the expected case, so don't be chatty in standalone
5525 : * mode
5526 : */
5527 1498 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
5528 : (errmsg("database system was shut down at %s",
5529 : str_time(ControlFile->time,
5530 : timebuf, sizeof(timebuf)))));
5531 1498 : break;
5532 :
5533 64 : case DB_SHUTDOWNED_IN_RECOVERY:
5534 64 : ereport(LOG,
5535 : (errmsg("database system was shut down in recovery at %s",
5536 : str_time(ControlFile->time,
5537 : timebuf, sizeof(timebuf)))));
5538 64 : break;
5539 :
5540 0 : case DB_SHUTDOWNING:
5541 0 : ereport(LOG,
5542 : (errmsg("database system shutdown was interrupted; last known up at %s",
5543 : str_time(ControlFile->time,
5544 : timebuf, sizeof(timebuf)))));
5545 0 : break;
5546 :
5547 0 : case DB_IN_CRASH_RECOVERY:
5548 0 : ereport(LOG,
5549 : (errmsg("database system was interrupted while in recovery at %s",
5550 : str_time(ControlFile->time,
5551 : timebuf, sizeof(timebuf))),
5552 : errhint("This probably means that some data is corrupted and"
5553 : " you will have to use the last backup for recovery.")));
5554 0 : break;
5555 :
5556 12 : case DB_IN_ARCHIVE_RECOVERY:
5557 12 : ereport(LOG,
5558 : (errmsg("database system was interrupted while in recovery at log time %s",
5559 : str_time(ControlFile->checkPointCopy.time,
5560 : timebuf, sizeof(timebuf))),
5561 : errhint("If this has occurred more than once some data might be corrupted"
5562 : " and you might need to choose an earlier recovery target.")));
5563 12 : break;
5564 :
5565 342 : case DB_IN_PRODUCTION:
5566 342 : ereport(LOG,
5567 : (errmsg("database system was interrupted; last known up at %s",
5568 : str_time(ControlFile->time,
5569 : timebuf, sizeof(timebuf)))));
5570 342 : break;
5571 :
5572 0 : default:
5573 0 : ereport(FATAL,
5574 : (errcode(ERRCODE_DATA_CORRUPTED),
5575 : errmsg("control file contains invalid database cluster state")));
5576 : }
5577 :
5578 : /* This is just to allow attaching to startup process with a debugger */
5579 : #ifdef XLOG_REPLAY_DELAY
5580 : if (ControlFile->state != DB_SHUTDOWNED)
5581 : pg_usleep(60000000L);
5582 : #endif
5583 :
5584 : /*
5585 : * Verify that pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
5586 : * In cases where someone has performed a copy for PITR, these directories
5587 : * may have been excluded and need to be re-created.
5588 : */
5589 1916 : ValidateXLOGDirectoryStructure();
5590 :
5591 : /* Set up timeout handler needed to report startup progress. */
5592 1916 : if (!IsBootstrapProcessingMode())
5593 1816 : RegisterTimeout(STARTUP_PROGRESS_TIMEOUT,
5594 : startup_progress_timeout_handler);
5595 :
5596 : /*----------
5597 : * If we previously crashed, perform a couple of actions:
5598 : *
5599 : * - The pg_wal directory may still include some temporary WAL segments
5600 : * used when creating a new segment, so perform some clean up to not
5601 : * bloat this path. This is done first as there is no point to sync
5602 : * this temporary data.
5603 : *
5604 : * - There might be data which we had written, intending to fsync it, but
5605 : * which we had not actually fsync'd yet. Therefore, a power failure in
5606 : * the near future might cause earlier unflushed writes to be lost, even
5607 : * though more recent data written to disk from here on would be
5608 : * persisted. To avoid that, fsync the entire data directory.
5609 : */
5610 1916 : if (ControlFile->state != DB_SHUTDOWNED &&
5611 418 : ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
5612 : {
5613 354 : RemoveTempXlogFiles();
5614 354 : SyncDataDirectory();
5615 354 : didCrash = true;
5616 : }
5617 : else
5618 1562 : didCrash = false;
5619 :
5620 : /*
5621 : * Prepare for WAL recovery if needed.
5622 : *
5623 : * InitWalRecovery analyzes the control file and the backup label file, if
5624 : * any. It updates the in-memory ControlFile buffer according to the
5625 : * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5626 : * It also applies the tablespace map file, if any.
5627 : */
5628 1916 : InitWalRecovery(ControlFile, &wasShutdown,
5629 : &haveBackupLabel, &haveTblspcMap);
5630 1916 : checkPoint = ControlFile->checkPointCopy;
5631 :
5632 : /* initialize shared memory variables from the checkpoint record */
5633 1916 : TransamVariables->nextXid = checkPoint.nextXid;
5634 1916 : TransamVariables->nextOid = checkPoint.nextOid;
5635 1916 : TransamVariables->oidCount = 0;
5636 1916 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5637 1916 : AdvanceOldestClogXid(checkPoint.oldestXid);
5638 1916 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5639 1916 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5640 1916 : SetCommitTsLimit(checkPoint.oldestCommitTsXid,
5641 : checkPoint.newestCommitTsXid);
5642 :
5643 : /*
5644 : * Clear out any old relcache cache files. This is *necessary* if we do
5645 : * any WAL replay, since that would probably result in the cache files
5646 : * being out of sync with database reality. In theory we could leave them
5647 : * in place if the database had been cleanly shut down, but it seems
5648 : * safest to just remove them always and let them be rebuilt during the
5649 : * first backend startup. These files needs to be removed from all
5650 : * directories including pg_tblspc, however the symlinks are created only
5651 : * after reading tablespace_map file in case of archive recovery from
5652 : * backup, so needs to clear old relcache files here after creating
5653 : * symlinks.
5654 : */
5655 1916 : RelationCacheInitFileRemove();
5656 :
5657 : /*
5658 : * Initialize replication slots, before there's a chance to remove
5659 : * required resources.
5660 : */
5661 1916 : StartupReplicationSlots();
5662 :
5663 : /*
5664 : * Startup logical state, needs to be setup now so we have proper data
5665 : * during crash recovery.
5666 : */
5667 1914 : StartupReorderBuffer();
5668 :
5669 : /*
5670 : * Startup CLOG. This must be done after TransamVariables->nextXid has
5671 : * been initialized and before we accept connections or begin WAL replay.
5672 : */
5673 1914 : StartupCLOG();
5674 :
5675 : /*
5676 : * Startup MultiXact. We need to do this early to be able to replay
5677 : * truncations.
5678 : */
5679 1914 : StartupMultiXact();
5680 :
5681 : /*
5682 : * Ditto for commit timestamps. Activate the facility if the setting is
5683 : * enabled in the control file, as there should be no tracking of commit
5684 : * timestamps done when the setting was disabled. This facility can be
5685 : * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
5686 : */
5687 1914 : if (ControlFile->track_commit_timestamp)
5688 28 : StartupCommitTs();
5689 :
5690 : /*
5691 : * Recover knowledge about replay progress of known replication partners.
5692 : */
5693 1914 : StartupReplicationOrigin();
5694 :
5695 : /*
5696 : * Initialize unlogged LSN. On a clean shutdown, it's restored from the
5697 : * control file. On recovery, all unlogged relations are blown away, so
5698 : * the unlogged LSN counter can be reset too.
5699 : */
5700 1914 : if (ControlFile->state == DB_SHUTDOWNED)
5701 1484 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5702 1484 : ControlFile->unloggedLSN);
5703 : else
5704 430 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5705 : FirstNormalUnloggedLSN);
5706 :
5707 : /*
5708 : * Copy any missing timeline history files between 'now' and the recovery
5709 : * target timeline from archive to pg_wal. While we don't need those files
5710 : * ourselves - the history file of the recovery target timeline covers all
5711 : * the previous timelines in the history too - a cascading standby server
5712 : * might be interested in them. Or, if you archive the WAL from this
5713 : * server to a different archive than the primary, it'd be good for all
5714 : * the history files to get archived there after failover, so that you can
5715 : * use one of the old timelines as a PITR target. Timeline history files
5716 : * are small, so it's better to copy them unnecessarily than not copy them
5717 : * and regret later.
5718 : */
5719 1914 : restoreTimeLineHistoryFiles(checkPoint.ThisTimeLineID, recoveryTargetTLI);
5720 :
5721 : /*
5722 : * Before running in recovery, scan pg_twophase and fill in its status to
5723 : * be able to work on entries generated by redo. Doing a scan before
5724 : * taking any recovery action has the merit to discard any 2PC files that
5725 : * are newer than the first record to replay, saving from any conflicts at
5726 : * replay. This avoids as well any subsequent scans when doing recovery
5727 : * of the on-disk two-phase data.
5728 : */
5729 1914 : restoreTwoPhaseData();
5730 :
5731 : /*
5732 : * When starting with crash recovery, reset pgstat data - it might not be
5733 : * valid. Otherwise restore pgstat data. It's safe to do this here,
5734 : * because postmaster will not yet have started any other processes.
5735 : *
5736 : * NB: Restoring replication slot stats relies on slot state to have
5737 : * already been restored from disk.
5738 : *
5739 : * TODO: With a bit of extra work we could just start with a pgstat file
5740 : * associated with the checkpoint redo location we're starting from.
5741 : */
5742 1914 : if (didCrash)
5743 354 : pgstat_discard_stats();
5744 : else
5745 1560 : pgstat_restore_stats();
5746 :
5747 1914 : lastFullPageWrites = checkPoint.fullPageWrites;
5748 :
5749 1914 : RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
5750 1914 : doPageWrites = lastFullPageWrites;
5751 :
5752 : /* REDO */
5753 1914 : if (InRecovery)
5754 : {
5755 : /* Initialize state for RecoveryInProgress() */
5756 430 : SpinLockAcquire(&XLogCtl->info_lck);
5757 430 : if (InArchiveRecovery)
5758 224 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
5759 : else
5760 206 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5761 430 : SpinLockRelease(&XLogCtl->info_lck);
5762 :
5763 : /*
5764 : * Update pg_control to show that we are recovering and to show the
5765 : * selected checkpoint as the place we are starting from. We also mark
5766 : * pg_control with any minimum recovery stop point obtained from a
5767 : * backup history file.
5768 : *
5769 : * No need to hold ControlFileLock yet, we aren't up far enough.
5770 : */
5771 430 : UpdateControlFile();
5772 :
5773 : /*
5774 : * If there was a backup label file, it's done its job and the info
5775 : * has now been propagated into pg_control. We must get rid of the
5776 : * label file so that if we crash during recovery, we'll pick up at
5777 : * the latest recovery restartpoint instead of going all the way back
5778 : * to the backup start point. It seems prudent though to just rename
5779 : * the file out of the way rather than delete it completely.
5780 : */
5781 430 : if (haveBackupLabel)
5782 : {
5783 144 : unlink(BACKUP_LABEL_OLD);
5784 144 : durable_rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD, FATAL);
5785 : }
5786 :
5787 : /*
5788 : * If there was a tablespace_map file, it's done its job and the
5789 : * symlinks have been created. We must get rid of the map file so
5790 : * that if we crash during recovery, we don't create symlinks again.
5791 : * It seems prudent though to just rename the file out of the way
5792 : * rather than delete it completely.
5793 : */
5794 430 : if (haveTblspcMap)
5795 : {
5796 4 : unlink(TABLESPACE_MAP_OLD);
5797 4 : durable_rename(TABLESPACE_MAP, TABLESPACE_MAP_OLD, FATAL);
5798 : }
5799 :
5800 : /*
5801 : * Initialize our local copy of minRecoveryPoint. When doing crash
5802 : * recovery we want to replay up to the end of WAL. Particularly, in
5803 : * the case of a promoted standby minRecoveryPoint value in the
5804 : * control file is only updated after the first checkpoint. However,
5805 : * if the instance crashes before the first post-recovery checkpoint
5806 : * is completed then recovery will use a stale location causing the
5807 : * startup process to think that there are still invalid page
5808 : * references when checking for data consistency.
5809 : */
5810 430 : if (InArchiveRecovery)
5811 : {
5812 224 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
5813 224 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
5814 : }
5815 : else
5816 : {
5817 206 : LocalMinRecoveryPoint = InvalidXLogRecPtr;
5818 206 : LocalMinRecoveryPointTLI = 0;
5819 : }
5820 :
5821 : /* Check that the GUCs used to generate the WAL allow recovery */
5822 430 : CheckRequiredParameterValues();
5823 :
5824 : /*
5825 : * We're in recovery, so unlogged relations may be trashed and must be
5826 : * reset. This should be done BEFORE allowing Hot Standby
5827 : * connections, so that read-only backends don't try to read whatever
5828 : * garbage is left over from before.
5829 : */
5830 430 : ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
5831 :
5832 : /*
5833 : * Likewise, delete any saved transaction snapshot files that got left
5834 : * behind by crashed backends.
5835 : */
5836 430 : DeleteAllExportedSnapshotFiles();
5837 :
5838 : /*
5839 : * Initialize for Hot Standby, if enabled. We won't let backends in
5840 : * yet, not until we've reached the min recovery point specified in
5841 : * control file and we've established a recovery snapshot from a
5842 : * running-xacts WAL record.
5843 : */
5844 430 : if (ArchiveRecoveryRequested && EnableHotStandby)
5845 : {
5846 : TransactionId *xids;
5847 : int nxids;
5848 :
5849 214 : ereport(DEBUG1,
5850 : (errmsg_internal("initializing for hot standby")));
5851 :
5852 214 : InitRecoveryTransactionEnvironment();
5853 :
5854 214 : if (wasShutdown)
5855 52 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
5856 : else
5857 162 : oldestActiveXID = checkPoint.oldestActiveXid;
5858 : Assert(TransactionIdIsValid(oldestActiveXID));
5859 :
5860 : /* Tell procarray about the range of xids it has to deal with */
5861 214 : ProcArrayInitRecovery(XidFromFullTransactionId(TransamVariables->nextXid));
5862 :
5863 : /*
5864 : * Startup subtrans only. CLOG, MultiXact and commit timestamp
5865 : * have already been started up and other SLRUs are not maintained
5866 : * during recovery and need not be started yet.
5867 : */
5868 214 : StartupSUBTRANS(oldestActiveXID);
5869 :
5870 : /*
5871 : * If we're beginning at a shutdown checkpoint, we know that
5872 : * nothing was running on the primary at this point. So fake-up an
5873 : * empty running-xacts record and use that here and now. Recover
5874 : * additional standby state for prepared transactions.
5875 : */
5876 214 : if (wasShutdown)
5877 : {
5878 : RunningTransactionsData running;
5879 : TransactionId latestCompletedXid;
5880 :
5881 : /* Update pg_subtrans entries for any prepared transactions */
5882 52 : StandbyRecoverPreparedTransactions();
5883 :
5884 : /*
5885 : * Construct a RunningTransactions snapshot representing a
5886 : * shut down server, with only prepared transactions still
5887 : * alive. We're never overflowed at this point because all
5888 : * subxids are listed with their parent prepared transactions.
5889 : */
5890 52 : running.xcnt = nxids;
5891 52 : running.subxcnt = 0;
5892 52 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
5893 52 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
5894 52 : running.oldestRunningXid = oldestActiveXID;
5895 52 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
5896 52 : TransactionIdRetreat(latestCompletedXid);
5897 : Assert(TransactionIdIsNormal(latestCompletedXid));
5898 52 : running.latestCompletedXid = latestCompletedXid;
5899 52 : running.xids = xids;
5900 :
5901 52 : ProcArrayApplyRecoveryInfo(&running);
5902 : }
5903 : }
5904 :
5905 : /*
5906 : * We're all set for replaying the WAL now. Do it.
5907 : */
5908 430 : PerformWalRecovery();
5909 312 : performedWalRecovery = true;
5910 : }
5911 : else
5912 1484 : performedWalRecovery = false;
5913 :
5914 : /*
5915 : * Finish WAL recovery.
5916 : */
5917 1796 : endOfRecoveryInfo = FinishWalRecovery();
5918 1796 : EndOfLog = endOfRecoveryInfo->endOfLog;
5919 1796 : EndOfLogTLI = endOfRecoveryInfo->endOfLogTLI;
5920 1796 : abortedRecPtr = endOfRecoveryInfo->abortedRecPtr;
5921 1796 : missingContrecPtr = endOfRecoveryInfo->missingContrecPtr;
5922 :
5923 : /*
5924 : * Reset ps status display, so as no information related to recovery shows
5925 : * up.
5926 : */
5927 1796 : set_ps_display("");
5928 :
5929 : /*
5930 : * When recovering from a backup (we are in recovery, and archive recovery
5931 : * was requested), complain if we did not roll forward far enough to reach
5932 : * the point where the database is consistent. For regular online
5933 : * backup-from-primary, that means reaching the end-of-backup WAL record
5934 : * (at which point we reset backupStartPoint to be Invalid), for
5935 : * backup-from-replica (which can't inject records into the WAL stream),
5936 : * that point is when we reach the minRecoveryPoint in pg_control (which
5937 : * we purposefully copy last when backing up from a replica). For
5938 : * pg_rewind (which creates a backup_label with a method of "pg_rewind")
5939 : * or snapshot-style backups (which don't), backupEndRequired will be set
5940 : * to false.
5941 : *
5942 : * Note: it is indeed okay to look at the local variable
5943 : * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
5944 : * might be further ahead --- ControlFile->minRecoveryPoint cannot have
5945 : * been advanced beyond the WAL we processed.
5946 : */
5947 1796 : if (InRecovery &&
5948 312 : (EndOfLog < LocalMinRecoveryPoint ||
5949 312 : XLogRecPtrIsValid(ControlFile->backupStartPoint)))
5950 : {
5951 : /*
5952 : * Ran off end of WAL before reaching end-of-backup WAL record, or
5953 : * minRecoveryPoint. That's a bad sign, indicating that you tried to
5954 : * recover from an online backup but never called pg_backup_stop(), or
5955 : * you didn't archive all the WAL needed.
5956 : */
5957 0 : if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
5958 : {
5959 0 : if (XLogRecPtrIsValid(ControlFile->backupStartPoint) || ControlFile->backupEndRequired)
5960 0 : ereport(FATAL,
5961 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5962 : errmsg("WAL ends before end of online backup"),
5963 : errhint("All WAL generated while online backup was taken must be available at recovery.")));
5964 : else
5965 0 : ereport(FATAL,
5966 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5967 : errmsg("WAL ends before consistent recovery point")));
5968 : }
5969 : }
5970 :
5971 : /*
5972 : * Reset unlogged relations to the contents of their INIT fork. This is
5973 : * done AFTER recovery is complete so as to include any unlogged relations
5974 : * created during recovery, but BEFORE recovery is marked as having
5975 : * completed successfully. Otherwise we'd not retry if any of the post
5976 : * end-of-recovery steps fail.
5977 : */
5978 1796 : if (InRecovery)
5979 312 : ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
5980 :
5981 : /*
5982 : * Pre-scan prepared transactions to find out the range of XIDs present.
5983 : * This information is not quite needed yet, but it is positioned here so
5984 : * as potential problems are detected before any on-disk change is done.
5985 : */
5986 1796 : oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
5987 :
5988 : /*
5989 : * Allow ordinary WAL segment creation before possibly switching to a new
5990 : * timeline, which creates a new segment, and after the last ReadRecord().
5991 : */
5992 1796 : SetInstallXLogFileSegmentActive();
5993 :
5994 : /*
5995 : * Consider whether we need to assign a new timeline ID.
5996 : *
5997 : * If we did archive recovery, we always assign a new ID. This handles a
5998 : * couple of issues. If we stopped short of the end of WAL during
5999 : * recovery, then we are clearly generating a new timeline and must assign
6000 : * it a unique new ID. Even if we ran to the end, modifying the current
6001 : * last segment is problematic because it may result in trying to
6002 : * overwrite an already-archived copy of that segment, and we encourage
6003 : * DBAs to make their archive_commands reject that. We can dodge the
6004 : * problem by making the new active segment have a new timeline ID.
6005 : *
6006 : * In a normal crash recovery, we can just extend the timeline we were in.
6007 : */
6008 1796 : newTLI = endOfRecoveryInfo->lastRecTLI;
6009 1796 : if (ArchiveRecoveryRequested)
6010 : {
6011 100 : newTLI = findNewestTimeLine(recoveryTargetTLI) + 1;
6012 100 : ereport(LOG,
6013 : (errmsg("selected new timeline ID: %u", newTLI)));
6014 :
6015 : /*
6016 : * Make a writable copy of the last WAL segment. (Note that we also
6017 : * have a copy of the last block of the old WAL in
6018 : * endOfRecovery->lastPage; we will use that below.)
6019 : */
6020 100 : XLogInitNewTimeline(EndOfLogTLI, EndOfLog, newTLI);
6021 :
6022 : /*
6023 : * Remove the signal files out of the way, so that we don't
6024 : * accidentally re-enter archive recovery mode in a subsequent crash.
6025 : */
6026 100 : if (endOfRecoveryInfo->standby_signal_file_found)
6027 94 : durable_unlink(STANDBY_SIGNAL_FILE, FATAL);
6028 :
6029 100 : if (endOfRecoveryInfo->recovery_signal_file_found)
6030 6 : durable_unlink(RECOVERY_SIGNAL_FILE, FATAL);
6031 :
6032 : /*
6033 : * Write the timeline history file, and have it archived. After this
6034 : * point (or rather, as soon as the file is archived), the timeline
6035 : * will appear as "taken" in the WAL archive and to any standby
6036 : * servers. If we crash before actually switching to the new
6037 : * timeline, standby servers will nevertheless think that we switched
6038 : * to the new timeline, and will try to connect to the new timeline.
6039 : * To minimize the window for that, try to do as little as possible
6040 : * between here and writing the end-of-recovery record.
6041 : */
6042 100 : writeTimeLineHistory(newTLI, recoveryTargetTLI,
6043 : EndOfLog, endOfRecoveryInfo->recoveryStopReason);
6044 :
6045 100 : ereport(LOG,
6046 : (errmsg("archive recovery complete")));
6047 : }
6048 :
6049 : /* Save the selected TimeLineID in shared memory, too */
6050 1796 : SpinLockAcquire(&XLogCtl->info_lck);
6051 1796 : XLogCtl->InsertTimeLineID = newTLI;
6052 1796 : XLogCtl->PrevTimeLineID = endOfRecoveryInfo->lastRecTLI;
6053 1796 : SpinLockRelease(&XLogCtl->info_lck);
6054 :
6055 : /*
6056 : * Actually, if WAL ended in an incomplete record, skip the parts that
6057 : * made it through and start writing after the portion that persisted.
6058 : * (It's critical to first write an OVERWRITE_CONTRECORD message, which
6059 : * we'll do as soon as we're open for writing new WAL.)
6060 : */
6061 1796 : if (XLogRecPtrIsValid(missingContrecPtr))
6062 : {
6063 : /*
6064 : * We should only have a missingContrecPtr if we're not switching to a
6065 : * new timeline. When a timeline switch occurs, WAL is copied from the
6066 : * old timeline to the new only up to the end of the last complete
6067 : * record, so there can't be an incomplete WAL record that we need to
6068 : * disregard.
6069 : */
6070 : Assert(newTLI == endOfRecoveryInfo->lastRecTLI);
6071 : Assert(XLogRecPtrIsValid(abortedRecPtr));
6072 20 : EndOfLog = missingContrecPtr;
6073 : }
6074 :
6075 : /*
6076 : * Prepare to write WAL starting at EndOfLog location, and init xlog
6077 : * buffer cache using the block containing the last record from the
6078 : * previous incarnation.
6079 : */
6080 1796 : Insert = &XLogCtl->Insert;
6081 1796 : Insert->PrevBytePos = XLogRecPtrToBytePos(endOfRecoveryInfo->lastRec);
6082 1796 : Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
6083 :
6084 : /*
6085 : * Tricky point here: lastPage contains the *last* block that the LastRec
6086 : * record spans, not the one it starts in. The last block is indeed the
6087 : * one we want to use.
6088 : */
6089 1796 : if (EndOfLog % XLOG_BLCKSZ != 0)
6090 : {
6091 : char *page;
6092 : int len;
6093 : int firstIdx;
6094 :
6095 1742 : firstIdx = XLogRecPtrToBufIdx(EndOfLog);
6096 1742 : len = EndOfLog - endOfRecoveryInfo->lastPageBeginPtr;
6097 : Assert(len < XLOG_BLCKSZ);
6098 :
6099 : /* Copy the valid part of the last block, and zero the rest */
6100 1742 : page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
6101 1742 : memcpy(page, endOfRecoveryInfo->lastPage, len);
6102 1742 : memset(page + len, 0, XLOG_BLCKSZ - len);
6103 :
6104 1742 : pg_atomic_write_u64(&XLogCtl->xlblocks[firstIdx], endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6105 1742 : XLogCtl->InitializedUpTo = endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ;
6106 : }
6107 : else
6108 : {
6109 : /*
6110 : * There is no partial block to copy. Just set InitializedUpTo, and
6111 : * let the first attempt to insert a log record to initialize the next
6112 : * buffer.
6113 : */
6114 54 : XLogCtl->InitializedUpTo = EndOfLog;
6115 : }
6116 :
6117 : /*
6118 : * Update local and shared status. This is OK to do without any locks
6119 : * because no other process can be reading or writing WAL yet.
6120 : */
6121 1796 : LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
6122 1796 : pg_atomic_write_u64(&XLogCtl->logInsertResult, EndOfLog);
6123 1796 : pg_atomic_write_u64(&XLogCtl->logWriteResult, EndOfLog);
6124 1796 : pg_atomic_write_u64(&XLogCtl->logFlushResult, EndOfLog);
6125 1796 : XLogCtl->LogwrtRqst.Write = EndOfLog;
6126 1796 : XLogCtl->LogwrtRqst.Flush = EndOfLog;
6127 :
6128 : /*
6129 : * Preallocate additional log files, if wanted.
6130 : */
6131 1796 : PreallocXlogFiles(EndOfLog, newTLI);
6132 :
6133 : /*
6134 : * Okay, we're officially UP.
6135 : */
6136 1796 : InRecovery = false;
6137 :
6138 : /* start the archive_timeout timer and LSN running */
6139 1796 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
6140 1796 : XLogCtl->lastSegSwitchLSN = EndOfLog;
6141 :
6142 : /* also initialize latestCompletedXid, to nextXid - 1 */
6143 1796 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
6144 1796 : TransamVariables->latestCompletedXid = TransamVariables->nextXid;
6145 1796 : FullTransactionIdRetreat(&TransamVariables->latestCompletedXid);
6146 1796 : LWLockRelease(ProcArrayLock);
6147 :
6148 : /*
6149 : * Start up subtrans, if not already done for hot standby. (commit
6150 : * timestamps are started below, if necessary.)
6151 : */
6152 1796 : if (standbyState == STANDBY_DISABLED)
6153 1696 : StartupSUBTRANS(oldestActiveXID);
6154 :
6155 : /*
6156 : * Perform end of recovery actions for any SLRUs that need it.
6157 : */
6158 1796 : TrimCLOG();
6159 1796 : TrimMultiXact();
6160 :
6161 : /*
6162 : * Reload shared-memory state for prepared transactions. This needs to
6163 : * happen before renaming the last partial segment of the old timeline as
6164 : * it may be possible that we have to recover some transactions from it.
6165 : */
6166 1796 : RecoverPreparedTransactions();
6167 :
6168 : /* Shut down xlogreader */
6169 1796 : ShutdownWalRecovery();
6170 :
6171 : /* Enable WAL writes for this backend only. */
6172 1796 : LocalSetXLogInsertAllowed();
6173 :
6174 : /* If necessary, write overwrite-contrecord before doing anything else */
6175 1796 : if (XLogRecPtrIsValid(abortedRecPtr))
6176 : {
6177 : Assert(XLogRecPtrIsValid(missingContrecPtr));
6178 20 : CreateOverwriteContrecordRecord(abortedRecPtr, missingContrecPtr, newTLI);
6179 : }
6180 :
6181 : /*
6182 : * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
6183 : * record before resource manager writes cleanup WAL records or checkpoint
6184 : * record is written.
6185 : */
6186 1796 : Insert->fullPageWrites = lastFullPageWrites;
6187 1796 : UpdateFullPageWrites();
6188 :
6189 : /*
6190 : * Emit checkpoint or end-of-recovery record in XLOG, if required.
6191 : */
6192 1796 : if (performedWalRecovery)
6193 312 : promoted = PerformRecoveryXLogAction();
6194 :
6195 : /*
6196 : * If any of the critical GUCs have changed, log them before we allow
6197 : * backends to write WAL.
6198 : */
6199 1796 : XLogReportParameters();
6200 :
6201 : /* If this is archive recovery, perform post-recovery cleanup actions. */
6202 1796 : if (ArchiveRecoveryRequested)
6203 100 : CleanupAfterArchiveRecovery(EndOfLogTLI, EndOfLog, newTLI);
6204 :
6205 : /*
6206 : * Local WAL inserts enabled, so it's time to finish initialization of
6207 : * commit timestamp.
6208 : */
6209 1796 : CompleteCommitTsInitialization();
6210 :
6211 : /* Clean up EndOfWalRecoveryInfo data to appease Valgrind leak checking */
6212 1796 : if (endOfRecoveryInfo->lastPage)
6213 1762 : pfree(endOfRecoveryInfo->lastPage);
6214 1796 : pfree(endOfRecoveryInfo->recoveryStopReason);
6215 1796 : pfree(endOfRecoveryInfo);
6216 :
6217 : /*
6218 : * All done with end-of-recovery actions.
6219 : *
6220 : * Now allow backends to write WAL and update the control file status in
6221 : * consequence. SharedRecoveryState, that controls if backends can write
6222 : * WAL, is updated while holding ControlFileLock to prevent other backends
6223 : * to look at an inconsistent state of the control file in shared memory.
6224 : * There is still a small window during which backends can write WAL and
6225 : * the control file is still referring to a system not in DB_IN_PRODUCTION
6226 : * state while looking at the on-disk control file.
6227 : *
6228 : * Also, we use info_lck to update SharedRecoveryState to ensure that
6229 : * there are no race conditions concerning visibility of other recent
6230 : * updates to shared memory.
6231 : */
6232 1796 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6233 1796 : ControlFile->state = DB_IN_PRODUCTION;
6234 :
6235 1796 : SpinLockAcquire(&XLogCtl->info_lck);
6236 1796 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_DONE;
6237 1796 : SpinLockRelease(&XLogCtl->info_lck);
6238 :
6239 1796 : UpdateControlFile();
6240 1796 : LWLockRelease(ControlFileLock);
6241 :
6242 : /*
6243 : * Wake up all waiters for replay LSN. They need to report an error that
6244 : * recovery was ended before reaching the target LSN.
6245 : */
6246 1796 : WaitLSNWakeup(WAIT_LSN_TYPE_REPLAY, InvalidXLogRecPtr);
6247 :
6248 : /*
6249 : * Shutdown the recovery environment. This must occur after
6250 : * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
6251 : * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
6252 : * any session building a snapshot will not rely on KnownAssignedXids as
6253 : * RecoveryInProgress() would return false at this stage. This is
6254 : * particularly critical for prepared 2PC transactions, that would still
6255 : * need to be included in snapshots once recovery has ended.
6256 : */
6257 1796 : if (standbyState != STANDBY_DISABLED)
6258 100 : ShutdownRecoveryTransactionEnvironment();
6259 :
6260 : /*
6261 : * If there were cascading standby servers connected to us, nudge any wal
6262 : * sender processes to notice that we've been promoted.
6263 : */
6264 1796 : WalSndWakeup(true, true);
6265 :
6266 : /*
6267 : * If this was a promotion, request an (online) checkpoint now. This isn't
6268 : * required for consistency, but the last restartpoint might be far back,
6269 : * and in case of a crash, recovering from it might take a longer than is
6270 : * appropriate now that we're not in standby mode anymore.
6271 : */
6272 1796 : if (promoted)
6273 86 : RequestCheckpoint(CHECKPOINT_FORCE);
6274 1796 : }
6275 :
6276 : /*
6277 : * Callback from PerformWalRecovery(), called when we switch from crash
6278 : * recovery to archive recovery mode. Updates the control file accordingly.
6279 : */
6280 : void
6281 6 : SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr, TimeLineID replayTLI)
6282 : {
6283 : /* initialize minRecoveryPoint to this record */
6284 6 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6285 6 : ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6286 6 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6287 : {
6288 6 : ControlFile->minRecoveryPoint = EndRecPtr;
6289 6 : ControlFile->minRecoveryPointTLI = replayTLI;
6290 : }
6291 : /* update local copy */
6292 6 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6293 6 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6294 :
6295 : /*
6296 : * The startup process can update its local copy of minRecoveryPoint from
6297 : * this point.
6298 : */
6299 6 : updateMinRecoveryPoint = true;
6300 :
6301 6 : UpdateControlFile();
6302 :
6303 : /*
6304 : * We update SharedRecoveryState while holding the lock on ControlFileLock
6305 : * so both states are consistent in shared memory.
6306 : */
6307 6 : SpinLockAcquire(&XLogCtl->info_lck);
6308 6 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6309 6 : SpinLockRelease(&XLogCtl->info_lck);
6310 :
6311 6 : LWLockRelease(ControlFileLock);
6312 6 : }
6313 :
6314 : /*
6315 : * Callback from PerformWalRecovery(), called when we reach the end of backup.
6316 : * Updates the control file accordingly.
6317 : */
6318 : void
6319 144 : ReachedEndOfBackup(XLogRecPtr EndRecPtr, TimeLineID tli)
6320 : {
6321 : /*
6322 : * We have reached the end of base backup, as indicated by pg_control. The
6323 : * data on disk is now consistent (unless minRecoveryPoint is further
6324 : * ahead, which can happen if we crashed during previous recovery). Reset
6325 : * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
6326 : * make sure we don't allow starting up at an earlier point even if
6327 : * recovery is stopped and restarted soon after this.
6328 : */
6329 144 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6330 :
6331 144 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6332 : {
6333 136 : ControlFile->minRecoveryPoint = EndRecPtr;
6334 136 : ControlFile->minRecoveryPointTLI = tli;
6335 : }
6336 :
6337 144 : ControlFile->backupStartPoint = InvalidXLogRecPtr;
6338 144 : ControlFile->backupEndPoint = InvalidXLogRecPtr;
6339 144 : ControlFile->backupEndRequired = false;
6340 144 : UpdateControlFile();
6341 :
6342 144 : LWLockRelease(ControlFileLock);
6343 144 : }
6344 :
6345 : /*
6346 : * Perform whatever XLOG actions are necessary at end of REDO.
6347 : *
6348 : * The goal here is to make sure that we'll be able to recover properly if
6349 : * we crash again. If we choose to write a checkpoint, we'll write a shutdown
6350 : * checkpoint rather than an on-line one. This is not particularly critical,
6351 : * but since we may be assigning a new TLI, using a shutdown checkpoint allows
6352 : * us to have the rule that TLI only changes in shutdown checkpoints, which
6353 : * allows some extra error checking in xlog_redo.
6354 : */
6355 : static bool
6356 312 : PerformRecoveryXLogAction(void)
6357 : {
6358 312 : bool promoted = false;
6359 :
6360 : /*
6361 : * Perform a checkpoint to update all our recovery activity to disk.
6362 : *
6363 : * Note that we write a shutdown checkpoint rather than an on-line one.
6364 : * This is not particularly critical, but since we may be assigning a new
6365 : * TLI, using a shutdown checkpoint allows us to have the rule that TLI
6366 : * only changes in shutdown checkpoints, which allows some extra error
6367 : * checking in xlog_redo.
6368 : *
6369 : * In promotion, only create a lightweight end-of-recovery record instead
6370 : * of a full checkpoint. A checkpoint is requested later, after we're
6371 : * fully out of recovery mode and already accepting queries.
6372 : */
6373 412 : if (ArchiveRecoveryRequested && IsUnderPostmaster &&
6374 100 : PromoteIsTriggered())
6375 : {
6376 86 : promoted = true;
6377 :
6378 : /*
6379 : * Insert a special WAL record to mark the end of recovery, since we
6380 : * aren't doing a checkpoint. That means that the checkpointer process
6381 : * may likely be in the middle of a time-smoothed restartpoint and
6382 : * could continue to be for minutes after this. That sounds strange,
6383 : * but the effect is roughly the same and it would be stranger to try
6384 : * to come out of the restartpoint and then checkpoint. We request a
6385 : * checkpoint later anyway, just for safety.
6386 : */
6387 86 : CreateEndOfRecoveryRecord();
6388 : }
6389 : else
6390 : {
6391 226 : RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
6392 : CHECKPOINT_FAST |
6393 : CHECKPOINT_WAIT);
6394 : }
6395 :
6396 312 : return promoted;
6397 : }
6398 :
6399 : /*
6400 : * Is the system still in recovery?
6401 : *
6402 : * Unlike testing InRecovery, this works in any process that's connected to
6403 : * shared memory.
6404 : */
6405 : bool
6406 161950512 : RecoveryInProgress(void)
6407 : {
6408 : /*
6409 : * We check shared state each time only until we leave recovery mode. We
6410 : * can't re-enter recovery, so there's no need to keep checking after the
6411 : * shared variable has once been seen false.
6412 : */
6413 161950512 : if (!LocalRecoveryInProgress)
6414 157949756 : return false;
6415 : else
6416 : {
6417 : /*
6418 : * use volatile pointer to make sure we make a fresh read of the
6419 : * shared variable.
6420 : */
6421 4000756 : volatile XLogCtlData *xlogctl = XLogCtl;
6422 :
6423 4000756 : LocalRecoveryInProgress = (xlogctl->SharedRecoveryState != RECOVERY_STATE_DONE);
6424 :
6425 : /*
6426 : * Note: We don't need a memory barrier when we're still in recovery.
6427 : * We might exit recovery immediately after return, so the caller
6428 : * can't rely on 'true' meaning that we're still in recovery anyway.
6429 : */
6430 :
6431 4000756 : return LocalRecoveryInProgress;
6432 : }
6433 : }
6434 :
6435 : /*
6436 : * Returns current recovery state from shared memory.
6437 : *
6438 : * This returned state is kept consistent with the contents of the control
6439 : * file. See details about the possible values of RecoveryState in xlog.h.
6440 : */
6441 : RecoveryState
6442 39812 : GetRecoveryState(void)
6443 : {
6444 : RecoveryState retval;
6445 :
6446 39812 : SpinLockAcquire(&XLogCtl->info_lck);
6447 39812 : retval = XLogCtl->SharedRecoveryState;
6448 39812 : SpinLockRelease(&XLogCtl->info_lck);
6449 :
6450 39812 : return retval;
6451 : }
6452 :
6453 : /*
6454 : * Is this process allowed to insert new WAL records?
6455 : *
6456 : * Ordinarily this is essentially equivalent to !RecoveryInProgress().
6457 : * But we also have provisions for forcing the result "true" or "false"
6458 : * within specific processes regardless of the global state.
6459 : */
6460 : bool
6461 78321124 : XLogInsertAllowed(void)
6462 : {
6463 : /*
6464 : * If value is "unconditionally true" or "unconditionally false", just
6465 : * return it. This provides the normal fast path once recovery is known
6466 : * done.
6467 : */
6468 78321124 : if (LocalXLogInsertAllowed >= 0)
6469 77061660 : return (bool) LocalXLogInsertAllowed;
6470 :
6471 : /*
6472 : * Else, must check to see if we're still in recovery.
6473 : */
6474 1259464 : if (RecoveryInProgress())
6475 1240486 : return false;
6476 :
6477 : /*
6478 : * On exit from recovery, reset to "unconditionally true", since there is
6479 : * no need to keep checking.
6480 : */
6481 18978 : LocalXLogInsertAllowed = 1;
6482 18978 : return true;
6483 : }
6484 :
6485 : /*
6486 : * Make XLogInsertAllowed() return true in the current process only.
6487 : *
6488 : * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
6489 : * and even call LocalSetXLogInsertAllowed() again after that.
6490 : *
6491 : * Returns the previous value of LocalXLogInsertAllowed.
6492 : */
6493 : static int
6494 1854 : LocalSetXLogInsertAllowed(void)
6495 : {
6496 1854 : int oldXLogAllowed = LocalXLogInsertAllowed;
6497 :
6498 1854 : LocalXLogInsertAllowed = 1;
6499 :
6500 1854 : return oldXLogAllowed;
6501 : }
6502 :
6503 : /*
6504 : * Return the current Redo pointer from shared memory.
6505 : *
6506 : * As a side-effect, the local RedoRecPtr copy is updated.
6507 : */
6508 : XLogRecPtr
6509 593112 : GetRedoRecPtr(void)
6510 : {
6511 : XLogRecPtr ptr;
6512 :
6513 : /*
6514 : * The possibly not up-to-date copy in XlogCtl is enough. Even if we
6515 : * grabbed a WAL insertion lock to read the authoritative value in
6516 : * Insert->RedoRecPtr, someone might update it just after we've released
6517 : * the lock.
6518 : */
6519 593112 : SpinLockAcquire(&XLogCtl->info_lck);
6520 593112 : ptr = XLogCtl->RedoRecPtr;
6521 593112 : SpinLockRelease(&XLogCtl->info_lck);
6522 :
6523 593112 : if (RedoRecPtr < ptr)
6524 3128 : RedoRecPtr = ptr;
6525 :
6526 593112 : return RedoRecPtr;
6527 : }
6528 :
6529 : /*
6530 : * Return information needed to decide whether a modified block needs a
6531 : * full-page image to be included in the WAL record.
6532 : *
6533 : * The returned values are cached copies from backend-private memory, and
6534 : * possibly out-of-date or, indeed, uninitialized, in which case they will
6535 : * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6536 : * re-check them against up-to-date values, while holding the WAL insert lock.
6537 : */
6538 : void
6539 29544374 : GetFullPageWriteInfo(XLogRecPtr *RedoRecPtr_p, bool *doPageWrites_p)
6540 : {
6541 29544374 : *RedoRecPtr_p = RedoRecPtr;
6542 29544374 : *doPageWrites_p = doPageWrites;
6543 29544374 : }
6544 :
6545 : /*
6546 : * GetInsertRecPtr -- Returns the current insert position.
6547 : *
6548 : * NOTE: The value *actually* returned is the position of the last full
6549 : * xlog page. It lags behind the real insert position by at most 1 page.
6550 : * For that, we don't need to scan through WAL insertion locks, and an
6551 : * approximation is enough for the current usage of this function.
6552 : */
6553 : XLogRecPtr
6554 15910 : GetInsertRecPtr(void)
6555 : {
6556 : XLogRecPtr recptr;
6557 :
6558 15910 : SpinLockAcquire(&XLogCtl->info_lck);
6559 15910 : recptr = XLogCtl->LogwrtRqst.Write;
6560 15910 : SpinLockRelease(&XLogCtl->info_lck);
6561 :
6562 15910 : return recptr;
6563 : }
6564 :
6565 : /*
6566 : * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6567 : * position known to be fsync'd to disk. This should only be used on a
6568 : * system that is known not to be in recovery.
6569 : */
6570 : XLogRecPtr
6571 426348 : GetFlushRecPtr(TimeLineID *insertTLI)
6572 : {
6573 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6574 :
6575 426348 : RefreshXLogWriteResult(LogwrtResult);
6576 :
6577 : /*
6578 : * If we're writing and flushing WAL, the time line can't be changing, so
6579 : * no lock is required.
6580 : */
6581 426348 : if (insertTLI)
6582 45530 : *insertTLI = XLogCtl->InsertTimeLineID;
6583 :
6584 426348 : return LogwrtResult.Flush;
6585 : }
6586 :
6587 : /*
6588 : * GetWALInsertionTimeLine -- Returns the current timeline of a system that
6589 : * is not in recovery.
6590 : */
6591 : TimeLineID
6592 222476 : GetWALInsertionTimeLine(void)
6593 : {
6594 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6595 :
6596 : /* Since the value can't be changing, no lock is required. */
6597 222476 : return XLogCtl->InsertTimeLineID;
6598 : }
6599 :
6600 : /*
6601 : * GetWALInsertionTimeLineIfSet -- If the system is not in recovery, returns
6602 : * the WAL insertion timeline; else, returns 0. Wherever possible, use
6603 : * GetWALInsertionTimeLine() instead, since it's cheaper. Note that this
6604 : * function decides recovery has ended as soon as the insert TLI is set, which
6605 : * happens before we set XLogCtl->SharedRecoveryState to RECOVERY_STATE_DONE.
6606 : */
6607 : TimeLineID
6608 0 : GetWALInsertionTimeLineIfSet(void)
6609 : {
6610 : TimeLineID insertTLI;
6611 :
6612 0 : SpinLockAcquire(&XLogCtl->info_lck);
6613 0 : insertTLI = XLogCtl->InsertTimeLineID;
6614 0 : SpinLockRelease(&XLogCtl->info_lck);
6615 :
6616 0 : return insertTLI;
6617 : }
6618 :
6619 : /*
6620 : * GetLastImportantRecPtr -- Returns the LSN of the last important record
6621 : * inserted. All records not explicitly marked as unimportant are considered
6622 : * important.
6623 : *
6624 : * The LSN is determined by computing the maximum of
6625 : * WALInsertLocks[i].lastImportantAt.
6626 : */
6627 : XLogRecPtr
6628 3158 : GetLastImportantRecPtr(void)
6629 : {
6630 3158 : XLogRecPtr res = InvalidXLogRecPtr;
6631 : int i;
6632 :
6633 28422 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
6634 : {
6635 : XLogRecPtr last_important;
6636 :
6637 : /*
6638 : * Need to take a lock to prevent torn reads of the LSN, which are
6639 : * possible on some of the supported platforms. WAL insert locks only
6640 : * support exclusive mode, so we have to use that.
6641 : */
6642 25264 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
6643 25264 : last_important = WALInsertLocks[i].l.lastImportantAt;
6644 25264 : LWLockRelease(&WALInsertLocks[i].l.lock);
6645 :
6646 25264 : if (res < last_important)
6647 5432 : res = last_important;
6648 : }
6649 :
6650 3158 : return res;
6651 : }
6652 :
6653 : /*
6654 : * Get the time and LSN of the last xlog segment switch
6655 : */
6656 : pg_time_t
6657 0 : GetLastSegSwitchData(XLogRecPtr *lastSwitchLSN)
6658 : {
6659 : pg_time_t result;
6660 :
6661 : /* Need WALWriteLock, but shared lock is sufficient */
6662 0 : LWLockAcquire(WALWriteLock, LW_SHARED);
6663 0 : result = XLogCtl->lastSegSwitchTime;
6664 0 : *lastSwitchLSN = XLogCtl->lastSegSwitchLSN;
6665 0 : LWLockRelease(WALWriteLock);
6666 :
6667 0 : return result;
6668 : }
6669 :
6670 : /*
6671 : * This must be called ONCE during postmaster or standalone-backend shutdown
6672 : */
6673 : void
6674 1314 : ShutdownXLOG(int code, Datum arg)
6675 : {
6676 : /*
6677 : * We should have an aux process resource owner to use, and we should not
6678 : * be in a transaction that's installed some other resowner.
6679 : */
6680 : Assert(AuxProcessResourceOwner != NULL);
6681 : Assert(CurrentResourceOwner == NULL ||
6682 : CurrentResourceOwner == AuxProcessResourceOwner);
6683 1314 : CurrentResourceOwner = AuxProcessResourceOwner;
6684 :
6685 : /* Don't be chatty in standalone mode */
6686 1314 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
6687 : (errmsg("shutting down")));
6688 :
6689 : /*
6690 : * Signal walsenders to move to stopping state.
6691 : */
6692 1314 : WalSndInitStopping();
6693 :
6694 : /*
6695 : * Wait for WAL senders to be in stopping state. This prevents commands
6696 : * from writing new WAL.
6697 : */
6698 1314 : WalSndWaitStopping();
6699 :
6700 1314 : if (RecoveryInProgress())
6701 110 : CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
6702 : else
6703 : {
6704 : /*
6705 : * If archiving is enabled, rotate the last XLOG file so that all the
6706 : * remaining records are archived (postmaster wakes up the archiver
6707 : * process one more time at the end of shutdown). The checkpoint
6708 : * record will go to the next XLOG file and won't be archived (yet).
6709 : */
6710 1204 : if (XLogArchivingActive())
6711 28 : RequestXLogSwitch(false);
6712 :
6713 1204 : CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
6714 : }
6715 1314 : }
6716 :
6717 : /*
6718 : * Log start of a checkpoint.
6719 : */
6720 : static void
6721 2884 : LogCheckpointStart(int flags, bool restartpoint)
6722 : {
6723 2884 : if (restartpoint)
6724 382 : ereport(LOG,
6725 : /* translator: the placeholders show checkpoint options */
6726 : (errmsg("restartpoint starting:%s%s%s%s%s%s%s%s",
6727 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6728 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6729 : (flags & CHECKPOINT_FAST) ? " fast" : "",
6730 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6731 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6732 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6733 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6734 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "")));
6735 : else
6736 2502 : ereport(LOG,
6737 : /* translator: the placeholders show checkpoint options */
6738 : (errmsg("checkpoint starting:%s%s%s%s%s%s%s%s",
6739 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6740 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6741 : (flags & CHECKPOINT_FAST) ? " fast" : "",
6742 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6743 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6744 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6745 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6746 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "")));
6747 2884 : }
6748 :
6749 : /*
6750 : * Log end of a checkpoint.
6751 : */
6752 : static void
6753 3468 : LogCheckpointEnd(bool restartpoint)
6754 : {
6755 : long write_msecs,
6756 : sync_msecs,
6757 : total_msecs,
6758 : longest_msecs,
6759 : average_msecs;
6760 : uint64 average_sync_time;
6761 :
6762 3468 : CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
6763 :
6764 3468 : write_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_write_t,
6765 : CheckpointStats.ckpt_sync_t);
6766 :
6767 3468 : sync_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_sync_t,
6768 : CheckpointStats.ckpt_sync_end_t);
6769 :
6770 : /* Accumulate checkpoint timing summary data, in milliseconds. */
6771 3468 : PendingCheckpointerStats.write_time += write_msecs;
6772 3468 : PendingCheckpointerStats.sync_time += sync_msecs;
6773 :
6774 : /*
6775 : * All of the published timing statistics are accounted for. Only
6776 : * continue if a log message is to be written.
6777 : */
6778 3468 : if (!log_checkpoints)
6779 584 : return;
6780 :
6781 2884 : total_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_start_t,
6782 : CheckpointStats.ckpt_end_t);
6783 :
6784 : /*
6785 : * Timing values returned from CheckpointStats are in microseconds.
6786 : * Convert to milliseconds for consistent printing.
6787 : */
6788 2884 : longest_msecs = (long) ((CheckpointStats.ckpt_longest_sync + 999) / 1000);
6789 :
6790 2884 : average_sync_time = 0;
6791 2884 : if (CheckpointStats.ckpt_sync_rels > 0)
6792 0 : average_sync_time = CheckpointStats.ckpt_agg_sync_time /
6793 0 : CheckpointStats.ckpt_sync_rels;
6794 2884 : average_msecs = (long) ((average_sync_time + 999) / 1000);
6795 :
6796 : /*
6797 : * ControlFileLock is not required to see ControlFile->checkPoint and
6798 : * ->checkPointCopy here as we are the only updator of those variables at
6799 : * this moment.
6800 : */
6801 2884 : if (restartpoint)
6802 382 : ereport(LOG,
6803 : (errmsg("restartpoint complete: wrote %d buffers (%.1f%%), "
6804 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6805 : "%d removed, %d recycled; write=%ld.%03d s, "
6806 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6807 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6808 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6809 : CheckpointStats.ckpt_bufs_written,
6810 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6811 : CheckpointStats.ckpt_slru_written,
6812 : CheckpointStats.ckpt_segs_added,
6813 : CheckpointStats.ckpt_segs_removed,
6814 : CheckpointStats.ckpt_segs_recycled,
6815 : write_msecs / 1000, (int) (write_msecs % 1000),
6816 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6817 : total_msecs / 1000, (int) (total_msecs % 1000),
6818 : CheckpointStats.ckpt_sync_rels,
6819 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6820 : average_msecs / 1000, (int) (average_msecs % 1000),
6821 : (int) (PrevCheckPointDistance / 1024.0),
6822 : (int) (CheckPointDistanceEstimate / 1024.0),
6823 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6824 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6825 : else
6826 2502 : ereport(LOG,
6827 : (errmsg("checkpoint complete: wrote %d buffers (%.1f%%), "
6828 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6829 : "%d removed, %d recycled; write=%ld.%03d s, "
6830 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6831 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6832 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
6833 : CheckpointStats.ckpt_bufs_written,
6834 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6835 : CheckpointStats.ckpt_slru_written,
6836 : CheckpointStats.ckpt_segs_added,
6837 : CheckpointStats.ckpt_segs_removed,
6838 : CheckpointStats.ckpt_segs_recycled,
6839 : write_msecs / 1000, (int) (write_msecs % 1000),
6840 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6841 : total_msecs / 1000, (int) (total_msecs % 1000),
6842 : CheckpointStats.ckpt_sync_rels,
6843 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6844 : average_msecs / 1000, (int) (average_msecs % 1000),
6845 : (int) (PrevCheckPointDistance / 1024.0),
6846 : (int) (CheckPointDistanceEstimate / 1024.0),
6847 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6848 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6849 : }
6850 :
6851 : /*
6852 : * Update the estimate of distance between checkpoints.
6853 : *
6854 : * The estimate is used to calculate the number of WAL segments to keep
6855 : * preallocated, see XLOGfileslop().
6856 : */
6857 : static void
6858 3468 : UpdateCheckPointDistanceEstimate(uint64 nbytes)
6859 : {
6860 : /*
6861 : * To estimate the number of segments consumed between checkpoints, keep a
6862 : * moving average of the amount of WAL generated in previous checkpoint
6863 : * cycles. However, if the load is bursty, with quiet periods and busy
6864 : * periods, we want to cater for the peak load. So instead of a plain
6865 : * moving average, let the average decline slowly if the previous cycle
6866 : * used less WAL than estimated, but bump it up immediately if it used
6867 : * more.
6868 : *
6869 : * When checkpoints are triggered by max_wal_size, this should converge to
6870 : * CheckpointSegments * wal_segment_size,
6871 : *
6872 : * Note: This doesn't pay any attention to what caused the checkpoint.
6873 : * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
6874 : * starting a base backup, are counted the same as those created
6875 : * automatically. The slow-decline will largely mask them out, if they are
6876 : * not frequent. If they are frequent, it seems reasonable to count them
6877 : * in as any others; if you issue a manual checkpoint every 5 minutes and
6878 : * never let a timed checkpoint happen, it makes sense to base the
6879 : * preallocation on that 5 minute interval rather than whatever
6880 : * checkpoint_timeout is set to.
6881 : */
6882 3468 : PrevCheckPointDistance = nbytes;
6883 3468 : if (CheckPointDistanceEstimate < nbytes)
6884 1490 : CheckPointDistanceEstimate = nbytes;
6885 : else
6886 1978 : CheckPointDistanceEstimate =
6887 1978 : (0.90 * CheckPointDistanceEstimate + 0.10 * (double) nbytes);
6888 3468 : }
6889 :
6890 : /*
6891 : * Update the ps display for a process running a checkpoint. Note that
6892 : * this routine should not do any allocations so as it can be called
6893 : * from a critical section.
6894 : */
6895 : static void
6896 6936 : update_checkpoint_display(int flags, bool restartpoint, bool reset)
6897 : {
6898 : /*
6899 : * The status is reported only for end-of-recovery and shutdown
6900 : * checkpoints or shutdown restartpoints. Updating the ps display is
6901 : * useful in those situations as it may not be possible to rely on
6902 : * pg_stat_activity to see the status of the checkpointer or the startup
6903 : * process.
6904 : */
6905 6936 : if ((flags & (CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IS_SHUTDOWN)) == 0)
6906 4328 : return;
6907 :
6908 2608 : if (reset)
6909 1304 : set_ps_display("");
6910 : else
6911 : {
6912 : char activitymsg[128];
6913 :
6914 3912 : snprintf(activitymsg, sizeof(activitymsg), "performing %s%s%s",
6915 1304 : (flags & CHECKPOINT_END_OF_RECOVERY) ? "end-of-recovery " : "",
6916 1304 : (flags & CHECKPOINT_IS_SHUTDOWN) ? "shutdown " : "",
6917 : restartpoint ? "restartpoint" : "checkpoint");
6918 1304 : set_ps_display(activitymsg);
6919 : }
6920 : }
6921 :
6922 :
6923 : /*
6924 : * Perform a checkpoint --- either during shutdown, or on-the-fly
6925 : *
6926 : * flags is a bitwise OR of the following:
6927 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
6928 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
6929 : * CHECKPOINT_FAST: finish the checkpoint ASAP, ignoring
6930 : * checkpoint_completion_target parameter.
6931 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
6932 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
6933 : * CHECKPOINT_END_OF_RECOVERY).
6934 : * CHECKPOINT_FLUSH_UNLOGGED: also flush buffers of unlogged tables.
6935 : *
6936 : * Note: flags contains other bits, of interest here only for logging purposes.
6937 : * In particular note that this routine is synchronous and does not pay
6938 : * attention to CHECKPOINT_WAIT.
6939 : *
6940 : * If !shutdown then we are writing an online checkpoint. An XLOG_CHECKPOINT_REDO
6941 : * record is inserted into WAL at the logical location of the checkpoint, before
6942 : * flushing anything to disk, and when the checkpoint is eventually completed,
6943 : * and it is from this point that WAL replay will begin in the case of a recovery
6944 : * from this checkpoint. Once everything is written to disk, an
6945 : * XLOG_CHECKPOINT_ONLINE record is written to complete the checkpoint, and
6946 : * points back to the earlier XLOG_CHECKPOINT_REDO record. This mechanism allows
6947 : * other write-ahead log records to be written while the checkpoint is in
6948 : * progress, but we must be very careful about order of operations. This function
6949 : * may take many minutes to execute on a busy system.
6950 : *
6951 : * On the other hand, when shutdown is true, concurrent insertion into the
6952 : * write-ahead log is impossible, so there is no need for two separate records.
6953 : * In this case, we only insert an XLOG_CHECKPOINT_SHUTDOWN record, and it's
6954 : * both the record marking the completion of the checkpoint and the location
6955 : * from which WAL replay would begin if needed.
6956 : *
6957 : * Returns true if a new checkpoint was performed, or false if it was skipped
6958 : * because the system was idle.
6959 : */
6960 : bool
6961 3086 : CreateCheckPoint(int flags)
6962 : {
6963 : bool shutdown;
6964 : CheckPoint checkPoint;
6965 : XLogRecPtr recptr;
6966 : XLogSegNo _logSegNo;
6967 3086 : XLogCtlInsert *Insert = &XLogCtl->Insert;
6968 : uint32 freespace;
6969 : XLogRecPtr PriorRedoPtr;
6970 : XLogRecPtr last_important_lsn;
6971 : VirtualTransactionId *vxids;
6972 : int nvxids;
6973 3086 : int oldXLogAllowed = 0;
6974 :
6975 : /*
6976 : * An end-of-recovery checkpoint is really a shutdown checkpoint, just
6977 : * issued at a different time.
6978 : */
6979 3086 : if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
6980 1262 : shutdown = true;
6981 : else
6982 1824 : shutdown = false;
6983 :
6984 : /* sanity check */
6985 3086 : if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
6986 0 : elog(ERROR, "can't create a checkpoint during recovery");
6987 :
6988 : /*
6989 : * Prepare to accumulate statistics.
6990 : *
6991 : * Note: because it is possible for log_checkpoints to change while a
6992 : * checkpoint proceeds, we always accumulate stats, even if
6993 : * log_checkpoints is currently off.
6994 : */
6995 33946 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
6996 3086 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
6997 :
6998 : /*
6999 : * Let smgr prepare for checkpoint; this has to happen outside the
7000 : * critical section and before we determine the REDO pointer. Note that
7001 : * smgr must not do anything that'd have to be undone if we decide no
7002 : * checkpoint is needed.
7003 : */
7004 3086 : SyncPreCheckpoint();
7005 :
7006 : /*
7007 : * Use a critical section to force system panic if we have trouble.
7008 : */
7009 3086 : START_CRIT_SECTION();
7010 :
7011 3086 : if (shutdown)
7012 : {
7013 1262 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7014 1262 : ControlFile->state = DB_SHUTDOWNING;
7015 1262 : UpdateControlFile();
7016 1262 : LWLockRelease(ControlFileLock);
7017 : }
7018 :
7019 : /* Begin filling in the checkpoint WAL record */
7020 37032 : MemSet(&checkPoint, 0, sizeof(checkPoint));
7021 3086 : checkPoint.time = (pg_time_t) time(NULL);
7022 :
7023 : /*
7024 : * For Hot Standby, derive the oldestActiveXid before we fix the redo
7025 : * pointer. This allows us to begin accumulating changes to assemble our
7026 : * starting snapshot of locks and transactions.
7027 : */
7028 3086 : if (!shutdown && XLogStandbyInfoActive())
7029 1728 : checkPoint.oldestActiveXid = GetOldestActiveTransactionId(false, true);
7030 : else
7031 1358 : checkPoint.oldestActiveXid = InvalidTransactionId;
7032 :
7033 : /*
7034 : * Get location of last important record before acquiring insert locks (as
7035 : * GetLastImportantRecPtr() also locks WAL locks).
7036 : */
7037 3086 : last_important_lsn = GetLastImportantRecPtr();
7038 :
7039 : /*
7040 : * If this isn't a shutdown or forced checkpoint, and if there has been no
7041 : * WAL activity requiring a checkpoint, skip it. The idea here is to
7042 : * avoid inserting duplicate checkpoints when the system is idle.
7043 : */
7044 3086 : if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
7045 : CHECKPOINT_FORCE)) == 0)
7046 : {
7047 382 : if (last_important_lsn == ControlFile->checkPoint)
7048 : {
7049 0 : END_CRIT_SECTION();
7050 0 : ereport(DEBUG1,
7051 : (errmsg_internal("checkpoint skipped because system is idle")));
7052 0 : return false;
7053 : }
7054 : }
7055 :
7056 : /*
7057 : * An end-of-recovery checkpoint is created before anyone is allowed to
7058 : * write WAL. To allow us to write the checkpoint record, temporarily
7059 : * enable XLogInsertAllowed.
7060 : */
7061 3086 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7062 58 : oldXLogAllowed = LocalSetXLogInsertAllowed();
7063 :
7064 3086 : checkPoint.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7065 3086 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7066 58 : checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7067 : else
7068 3028 : checkPoint.PrevTimeLineID = checkPoint.ThisTimeLineID;
7069 :
7070 : /*
7071 : * We must block concurrent insertions while examining insert state.
7072 : */
7073 3086 : WALInsertLockAcquireExclusive();
7074 :
7075 3086 : checkPoint.fullPageWrites = Insert->fullPageWrites;
7076 3086 : checkPoint.wal_level = wal_level;
7077 :
7078 3086 : if (shutdown)
7079 : {
7080 1262 : XLogRecPtr curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
7081 :
7082 : /*
7083 : * Compute new REDO record ptr = location of next XLOG record.
7084 : *
7085 : * Since this is a shutdown checkpoint, there can't be any concurrent
7086 : * WAL insertion.
7087 : */
7088 1262 : freespace = INSERT_FREESPACE(curInsert);
7089 1262 : if (freespace == 0)
7090 : {
7091 0 : if (XLogSegmentOffset(curInsert, wal_segment_size) == 0)
7092 0 : curInsert += SizeOfXLogLongPHD;
7093 : else
7094 0 : curInsert += SizeOfXLogShortPHD;
7095 : }
7096 1262 : checkPoint.redo = curInsert;
7097 :
7098 : /*
7099 : * Here we update the shared RedoRecPtr for future XLogInsert calls;
7100 : * this must be done while holding all the insertion locks.
7101 : *
7102 : * Note: if we fail to complete the checkpoint, RedoRecPtr will be
7103 : * left pointing past where it really needs to point. This is okay;
7104 : * the only consequence is that XLogInsert might back up whole buffers
7105 : * that it didn't really need to. We can't postpone advancing
7106 : * RedoRecPtr because XLogInserts that happen while we are dumping
7107 : * buffers must assume that their buffer changes are not included in
7108 : * the checkpoint.
7109 : */
7110 1262 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
7111 : }
7112 :
7113 : /*
7114 : * Now we can release the WAL insertion locks, allowing other xacts to
7115 : * proceed while we are flushing disk buffers.
7116 : */
7117 3086 : WALInsertLockRelease();
7118 :
7119 : /*
7120 : * If this is an online checkpoint, we have not yet determined the redo
7121 : * point. We do so now by inserting the special XLOG_CHECKPOINT_REDO
7122 : * record; the LSN at which it starts becomes the new redo pointer. We
7123 : * don't do this for a shutdown checkpoint, because in that case no WAL
7124 : * can be written between the redo point and the insertion of the
7125 : * checkpoint record itself, so the checkpoint record itself serves to
7126 : * mark the redo point.
7127 : */
7128 3086 : if (!shutdown)
7129 : {
7130 : /* Include WAL level in record for WAL summarizer's benefit. */
7131 1824 : XLogBeginInsert();
7132 1824 : XLogRegisterData(&wal_level, sizeof(wal_level));
7133 1824 : (void) XLogInsert(RM_XLOG_ID, XLOG_CHECKPOINT_REDO);
7134 :
7135 : /*
7136 : * XLogInsertRecord will have updated XLogCtl->Insert.RedoRecPtr in
7137 : * shared memory and RedoRecPtr in backend-local memory, but we need
7138 : * to copy that into the record that will be inserted when the
7139 : * checkpoint is complete.
7140 : */
7141 1824 : checkPoint.redo = RedoRecPtr;
7142 : }
7143 :
7144 : /* Update the info_lck-protected copy of RedoRecPtr as well */
7145 3086 : SpinLockAcquire(&XLogCtl->info_lck);
7146 3086 : XLogCtl->RedoRecPtr = checkPoint.redo;
7147 3086 : SpinLockRelease(&XLogCtl->info_lck);
7148 :
7149 : /*
7150 : * If enabled, log checkpoint start. We postpone this until now so as not
7151 : * to log anything if we decided to skip the checkpoint.
7152 : */
7153 3086 : if (log_checkpoints)
7154 2502 : LogCheckpointStart(flags, false);
7155 :
7156 : /* Update the process title */
7157 3086 : update_checkpoint_display(flags, false, false);
7158 :
7159 : TRACE_POSTGRESQL_CHECKPOINT_START(flags);
7160 :
7161 : /*
7162 : * Get the other info we need for the checkpoint record.
7163 : *
7164 : * We don't need to save oldestClogXid in the checkpoint, it only matters
7165 : * for the short period in which clog is being truncated, and if we crash
7166 : * during that we'll redo the clog truncation and fix up oldestClogXid
7167 : * there.
7168 : */
7169 3086 : LWLockAcquire(XidGenLock, LW_SHARED);
7170 3086 : checkPoint.nextXid = TransamVariables->nextXid;
7171 3086 : checkPoint.oldestXid = TransamVariables->oldestXid;
7172 3086 : checkPoint.oldestXidDB = TransamVariables->oldestXidDB;
7173 3086 : LWLockRelease(XidGenLock);
7174 :
7175 3086 : LWLockAcquire(CommitTsLock, LW_SHARED);
7176 3086 : checkPoint.oldestCommitTsXid = TransamVariables->oldestCommitTsXid;
7177 3086 : checkPoint.newestCommitTsXid = TransamVariables->newestCommitTsXid;
7178 3086 : LWLockRelease(CommitTsLock);
7179 :
7180 3086 : LWLockAcquire(OidGenLock, LW_SHARED);
7181 3086 : checkPoint.nextOid = TransamVariables->nextOid;
7182 3086 : if (!shutdown)
7183 1824 : checkPoint.nextOid += TransamVariables->oidCount;
7184 3086 : LWLockRelease(OidGenLock);
7185 :
7186 3086 : MultiXactGetCheckptMulti(shutdown,
7187 : &checkPoint.nextMulti,
7188 : &checkPoint.nextMultiOffset,
7189 : &checkPoint.oldestMulti,
7190 : &checkPoint.oldestMultiDB);
7191 :
7192 : /*
7193 : * Having constructed the checkpoint record, ensure all shmem disk buffers
7194 : * and commit-log buffers are flushed to disk.
7195 : *
7196 : * This I/O could fail for various reasons. If so, we will fail to
7197 : * complete the checkpoint, but there is no reason to force a system
7198 : * panic. Accordingly, exit critical section while doing it.
7199 : */
7200 3086 : END_CRIT_SECTION();
7201 :
7202 : /*
7203 : * In some cases there are groups of actions that must all occur on one
7204 : * side or the other of a checkpoint record. Before flushing the
7205 : * checkpoint record we must explicitly wait for any backend currently
7206 : * performing those groups of actions.
7207 : *
7208 : * One example is end of transaction, so we must wait for any transactions
7209 : * that are currently in commit critical sections. If an xact inserted
7210 : * its commit record into XLOG just before the REDO point, then a crash
7211 : * restart from the REDO point would not replay that record, which means
7212 : * that our flushing had better include the xact's update of pg_xact. So
7213 : * we wait till he's out of his commit critical section before proceeding.
7214 : * See notes in RecordTransactionCommit().
7215 : *
7216 : * Because we've already released the insertion locks, this test is a bit
7217 : * fuzzy: it is possible that we will wait for xacts we didn't really need
7218 : * to wait for. But the delay should be short and it seems better to make
7219 : * checkpoint take a bit longer than to hold off insertions longer than
7220 : * necessary. (In fact, the whole reason we have this issue is that xact.c
7221 : * does commit record XLOG insertion and clog update as two separate steps
7222 : * protected by different locks, but again that seems best on grounds of
7223 : * minimizing lock contention.)
7224 : *
7225 : * A transaction that has not yet set delayChkptFlags when we look cannot
7226 : * be at risk, since it has not inserted its commit record yet; and one
7227 : * that's already cleared it is not at risk either, since it's done fixing
7228 : * clog and we will correctly flush the update below. So we cannot miss
7229 : * any xacts we need to wait for.
7230 : */
7231 3086 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
7232 3086 : if (nvxids > 0)
7233 : {
7234 : do
7235 : {
7236 : /*
7237 : * Keep absorbing fsync requests while we wait. There could even
7238 : * be a deadlock if we don't, if the process that prevents the
7239 : * checkpoint is trying to add a request to the queue.
7240 : */
7241 64 : AbsorbSyncRequests();
7242 :
7243 64 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
7244 64 : pg_usleep(10000L); /* wait for 10 msec */
7245 64 : pgstat_report_wait_end();
7246 64 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7247 : DELAY_CHKPT_START));
7248 : }
7249 3086 : pfree(vxids);
7250 :
7251 3086 : CheckPointGuts(checkPoint.redo, flags);
7252 :
7253 3086 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
7254 3086 : if (nvxids > 0)
7255 : {
7256 : do
7257 : {
7258 2 : AbsorbSyncRequests();
7259 :
7260 2 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_COMPLETE);
7261 2 : pg_usleep(10000L); /* wait for 10 msec */
7262 2 : pgstat_report_wait_end();
7263 2 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7264 : DELAY_CHKPT_COMPLETE));
7265 : }
7266 3086 : pfree(vxids);
7267 :
7268 : /*
7269 : * Take a snapshot of running transactions and write this to WAL. This
7270 : * allows us to reconstruct the state of running transactions during
7271 : * archive recovery, if required. Skip, if this info disabled.
7272 : *
7273 : * If we are shutting down, or Startup process is completing crash
7274 : * recovery we don't need to write running xact data.
7275 : */
7276 3086 : if (!shutdown && XLogStandbyInfoActive())
7277 1728 : LogStandbySnapshot();
7278 :
7279 3086 : START_CRIT_SECTION();
7280 :
7281 : /*
7282 : * Now insert the checkpoint record into XLOG.
7283 : */
7284 3086 : XLogBeginInsert();
7285 3086 : XLogRegisterData(&checkPoint, sizeof(checkPoint));
7286 3086 : recptr = XLogInsert(RM_XLOG_ID,
7287 : shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
7288 : XLOG_CHECKPOINT_ONLINE);
7289 :
7290 3086 : XLogFlush(recptr);
7291 :
7292 : /*
7293 : * We mustn't write any new WAL after a shutdown checkpoint, or it will be
7294 : * overwritten at next startup. No-one should even try, this just allows
7295 : * sanity-checking. In the case of an end-of-recovery checkpoint, we want
7296 : * to just temporarily disable writing until the system has exited
7297 : * recovery.
7298 : */
7299 3086 : if (shutdown)
7300 : {
7301 1262 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7302 58 : LocalXLogInsertAllowed = oldXLogAllowed;
7303 : else
7304 1204 : LocalXLogInsertAllowed = 0; /* never again write WAL */
7305 : }
7306 :
7307 : /*
7308 : * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
7309 : * = end of actual checkpoint record.
7310 : */
7311 3086 : if (shutdown && checkPoint.redo != ProcLastRecPtr)
7312 0 : ereport(PANIC,
7313 : (errmsg("concurrent write-ahead log activity while database system is shutting down")));
7314 :
7315 : /*
7316 : * Remember the prior checkpoint's redo ptr for
7317 : * UpdateCheckPointDistanceEstimate()
7318 : */
7319 3086 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7320 :
7321 : /*
7322 : * Update the control file.
7323 : */
7324 3086 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7325 3086 : if (shutdown)
7326 1262 : ControlFile->state = DB_SHUTDOWNED;
7327 3086 : ControlFile->checkPoint = ProcLastRecPtr;
7328 3086 : ControlFile->checkPointCopy = checkPoint;
7329 : /* crash recovery should always recover to the end of WAL */
7330 3086 : ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
7331 3086 : ControlFile->minRecoveryPointTLI = 0;
7332 :
7333 : /*
7334 : * Persist unloggedLSN value. It's reset on crash recovery, so this goes
7335 : * unused on non-shutdown checkpoints, but seems useful to store it always
7336 : * for debugging purposes.
7337 : */
7338 3086 : ControlFile->unloggedLSN = pg_atomic_read_membarrier_u64(&XLogCtl->unloggedLSN);
7339 :
7340 3086 : UpdateControlFile();
7341 3086 : LWLockRelease(ControlFileLock);
7342 :
7343 : /*
7344 : * We are now done with critical updates; no need for system panic if we
7345 : * have trouble while fooling with old log segments.
7346 : */
7347 3086 : END_CRIT_SECTION();
7348 :
7349 : /*
7350 : * WAL summaries end when the next XLOG_CHECKPOINT_REDO or
7351 : * XLOG_CHECKPOINT_SHUTDOWN record is reached. This is the first point
7352 : * where (a) we're not inside of a critical section and (b) we can be
7353 : * certain that the relevant record has been flushed to disk, which must
7354 : * happen before it can be summarized.
7355 : *
7356 : * If this is a shutdown checkpoint, then this happens reasonably
7357 : * promptly: we've only just inserted and flushed the
7358 : * XLOG_CHECKPOINT_SHUTDOWN record. If this is not a shutdown checkpoint,
7359 : * then this might not be very prompt at all: the XLOG_CHECKPOINT_REDO
7360 : * record was written before we began flushing data to disk, and that
7361 : * could be many minutes ago at this point. However, we don't XLogFlush()
7362 : * after inserting that record, so we're not guaranteed that it's on disk
7363 : * until after the above call that flushes the XLOG_CHECKPOINT_ONLINE
7364 : * record.
7365 : */
7366 3086 : WakeupWalSummarizer();
7367 :
7368 : /*
7369 : * Let smgr do post-checkpoint cleanup (eg, deleting old files).
7370 : */
7371 3086 : SyncPostCheckpoint();
7372 :
7373 : /*
7374 : * Update the average distance between checkpoints if the prior checkpoint
7375 : * exists.
7376 : */
7377 3086 : if (XLogRecPtrIsValid(PriorRedoPtr))
7378 3086 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7379 :
7380 3086 : INJECTION_POINT("checkpoint-before-old-wal-removal", NULL);
7381 :
7382 : /*
7383 : * Delete old log files, those no longer needed for last checkpoint to
7384 : * prevent the disk holding the xlog from growing full.
7385 : */
7386 3086 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7387 3086 : KeepLogSeg(recptr, &_logSegNo);
7388 3086 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7389 : _logSegNo, InvalidOid,
7390 : InvalidTransactionId))
7391 : {
7392 : /*
7393 : * Some slots have been invalidated; recalculate the old-segment
7394 : * horizon, starting again from RedoRecPtr.
7395 : */
7396 6 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7397 6 : KeepLogSeg(recptr, &_logSegNo);
7398 : }
7399 3086 : _logSegNo--;
7400 3086 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
7401 : checkPoint.ThisTimeLineID);
7402 :
7403 : /*
7404 : * Make more log segments if needed. (Do this after recycling old log
7405 : * segments, since that may supply some of the needed files.)
7406 : */
7407 3086 : if (!shutdown)
7408 1824 : PreallocXlogFiles(recptr, checkPoint.ThisTimeLineID);
7409 :
7410 : /*
7411 : * Truncate pg_subtrans if possible. We can throw away all data before
7412 : * the oldest XMIN of any running transaction. No future transaction will
7413 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7414 : * in subtrans.c). During recovery, though, we mustn't do this because
7415 : * StartupSUBTRANS hasn't been called yet.
7416 : */
7417 3086 : if (!RecoveryInProgress())
7418 3028 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7419 :
7420 : /* Real work is done; log and update stats. */
7421 3086 : LogCheckpointEnd(false);
7422 :
7423 : /* Reset the process title */
7424 3086 : update_checkpoint_display(flags, false, true);
7425 :
7426 : TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
7427 : NBuffers,
7428 : CheckpointStats.ckpt_segs_added,
7429 : CheckpointStats.ckpt_segs_removed,
7430 : CheckpointStats.ckpt_segs_recycled);
7431 :
7432 3086 : return true;
7433 : }
7434 :
7435 : /*
7436 : * Mark the end of recovery in WAL though without running a full checkpoint.
7437 : * We can expect that a restartpoint is likely to be in progress as we
7438 : * do this, though we are unwilling to wait for it to complete.
7439 : *
7440 : * CreateRestartPoint() allows for the case where recovery may end before
7441 : * the restartpoint completes so there is no concern of concurrent behaviour.
7442 : */
7443 : static void
7444 86 : CreateEndOfRecoveryRecord(void)
7445 : {
7446 : xl_end_of_recovery xlrec;
7447 : XLogRecPtr recptr;
7448 :
7449 : /* sanity check */
7450 86 : if (!RecoveryInProgress())
7451 0 : elog(ERROR, "can only be used to end recovery");
7452 :
7453 86 : xlrec.end_time = GetCurrentTimestamp();
7454 86 : xlrec.wal_level = wal_level;
7455 :
7456 86 : WALInsertLockAcquireExclusive();
7457 86 : xlrec.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7458 86 : xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7459 86 : WALInsertLockRelease();
7460 :
7461 86 : START_CRIT_SECTION();
7462 :
7463 86 : XLogBeginInsert();
7464 86 : XLogRegisterData(&xlrec, sizeof(xl_end_of_recovery));
7465 86 : recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY);
7466 :
7467 86 : XLogFlush(recptr);
7468 :
7469 : /*
7470 : * Update the control file so that crash recovery can follow the timeline
7471 : * changes to this point.
7472 : */
7473 86 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7474 86 : ControlFile->minRecoveryPoint = recptr;
7475 86 : ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
7476 86 : UpdateControlFile();
7477 86 : LWLockRelease(ControlFileLock);
7478 :
7479 86 : END_CRIT_SECTION();
7480 86 : }
7481 :
7482 : /*
7483 : * Write an OVERWRITE_CONTRECORD message.
7484 : *
7485 : * When on WAL replay we expect a continuation record at the start of a page
7486 : * that is not there, recovery ends and WAL writing resumes at that point.
7487 : * But it's wrong to resume writing new WAL back at the start of the record
7488 : * that was broken, because downstream consumers of that WAL (physical
7489 : * replicas) are not prepared to "rewind". So the first action after
7490 : * finishing replay of all valid WAL must be to write a record of this type
7491 : * at the point where the contrecord was missing; to support xlogreader
7492 : * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
7493 : * to the page header where the record occurs. xlogreader has an ad-hoc
7494 : * mechanism to report metadata about the broken record, which is what we
7495 : * use here.
7496 : *
7497 : * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
7498 : * skip the record it was reading, and pass back the LSN of the skipped
7499 : * record, so that its caller can verify (on "replay" of that record) that the
7500 : * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
7501 : *
7502 : * 'aborted_lsn' is the beginning position of the record that was incomplete.
7503 : * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
7504 : * beginning of the XLOG page where the record is to be inserted. They must
7505 : * match the current WAL insert position, they're passed here just so that we
7506 : * can verify that.
7507 : */
7508 : static XLogRecPtr
7509 20 : CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn, XLogRecPtr pagePtr,
7510 : TimeLineID newTLI)
7511 : {
7512 : xl_overwrite_contrecord xlrec;
7513 : XLogRecPtr recptr;
7514 : XLogPageHeader pagehdr;
7515 : XLogRecPtr startPos;
7516 :
7517 : /* sanity checks */
7518 20 : if (!RecoveryInProgress())
7519 0 : elog(ERROR, "can only be used at end of recovery");
7520 20 : if (pagePtr % XLOG_BLCKSZ != 0)
7521 0 : elog(ERROR, "invalid position for missing continuation record %X/%08X",
7522 : LSN_FORMAT_ARGS(pagePtr));
7523 :
7524 : /* The current WAL insert position should be right after the page header */
7525 20 : startPos = pagePtr;
7526 20 : if (XLogSegmentOffset(startPos, wal_segment_size) == 0)
7527 2 : startPos += SizeOfXLogLongPHD;
7528 : else
7529 18 : startPos += SizeOfXLogShortPHD;
7530 20 : recptr = GetXLogInsertRecPtr();
7531 20 : if (recptr != startPos)
7532 0 : elog(ERROR, "invalid WAL insert position %X/%08X for OVERWRITE_CONTRECORD",
7533 : LSN_FORMAT_ARGS(recptr));
7534 :
7535 20 : START_CRIT_SECTION();
7536 :
7537 : /*
7538 : * Initialize the XLOG page header (by GetXLogBuffer), and set the
7539 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
7540 : *
7541 : * No other backend is allowed to write WAL yet, so acquiring the WAL
7542 : * insertion lock is just pro forma.
7543 : */
7544 20 : WALInsertLockAcquire();
7545 20 : pagehdr = (XLogPageHeader) GetXLogBuffer(pagePtr, newTLI);
7546 20 : pagehdr->xlp_info |= XLP_FIRST_IS_OVERWRITE_CONTRECORD;
7547 20 : WALInsertLockRelease();
7548 :
7549 : /*
7550 : * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
7551 : * page. We know it becomes the first record, because no other backend is
7552 : * allowed to write WAL yet.
7553 : */
7554 20 : XLogBeginInsert();
7555 20 : xlrec.overwritten_lsn = aborted_lsn;
7556 20 : xlrec.overwrite_time = GetCurrentTimestamp();
7557 20 : XLogRegisterData(&xlrec, sizeof(xl_overwrite_contrecord));
7558 20 : recptr = XLogInsert(RM_XLOG_ID, XLOG_OVERWRITE_CONTRECORD);
7559 :
7560 : /* check that the record was inserted to the right place */
7561 20 : if (ProcLastRecPtr != startPos)
7562 0 : elog(ERROR, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%08X",
7563 : LSN_FORMAT_ARGS(ProcLastRecPtr));
7564 :
7565 20 : XLogFlush(recptr);
7566 :
7567 20 : END_CRIT_SECTION();
7568 :
7569 20 : return recptr;
7570 : }
7571 :
7572 : /*
7573 : * Flush all data in shared memory to disk, and fsync
7574 : *
7575 : * This is the common code shared between regular checkpoints and
7576 : * recovery restartpoints.
7577 : */
7578 : static void
7579 3468 : CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
7580 : {
7581 3468 : CheckPointRelationMap();
7582 3468 : CheckPointReplicationSlots(flags & CHECKPOINT_IS_SHUTDOWN);
7583 3468 : CheckPointSnapBuild();
7584 3468 : CheckPointLogicalRewriteHeap();
7585 3468 : CheckPointReplicationOrigin();
7586 :
7587 : /* Write out all dirty data in SLRUs and the main buffer pool */
7588 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags);
7589 3468 : CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
7590 3468 : CheckPointCLOG();
7591 3468 : CheckPointCommitTs();
7592 3468 : CheckPointSUBTRANS();
7593 3468 : CheckPointMultiXact();
7594 3468 : CheckPointPredicate();
7595 3468 : CheckPointBuffers(flags);
7596 :
7597 : /* Perform all queued up fsyncs */
7598 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
7599 3468 : CheckpointStats.ckpt_sync_t = GetCurrentTimestamp();
7600 3468 : ProcessSyncRequests();
7601 3468 : CheckpointStats.ckpt_sync_end_t = GetCurrentTimestamp();
7602 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
7603 :
7604 : /* We deliberately delay 2PC checkpointing as long as possible */
7605 3468 : CheckPointTwoPhase(checkPointRedo);
7606 3468 : }
7607 :
7608 : /*
7609 : * Save a checkpoint for recovery restart if appropriate
7610 : *
7611 : * This function is called each time a checkpoint record is read from XLOG.
7612 : * It must determine whether the checkpoint represents a safe restartpoint or
7613 : * not. If so, the checkpoint record is stashed in shared memory so that
7614 : * CreateRestartPoint can consult it. (Note that the latter function is
7615 : * executed by the checkpointer, while this one will be executed by the
7616 : * startup process.)
7617 : */
7618 : static void
7619 1434 : RecoveryRestartPoint(const CheckPoint *checkPoint, XLogReaderState *record)
7620 : {
7621 : /*
7622 : * Also refrain from creating a restartpoint if we have seen any
7623 : * references to non-existent pages. Restarting recovery from the
7624 : * restartpoint would not see the references, so we would lose the
7625 : * cross-check that the pages belonged to a relation that was dropped
7626 : * later.
7627 : */
7628 1434 : if (XLogHaveInvalidPages())
7629 : {
7630 0 : elog(DEBUG2,
7631 : "could not record restart point at %X/%08X because there are unresolved references to invalid pages",
7632 : LSN_FORMAT_ARGS(checkPoint->redo));
7633 0 : return;
7634 : }
7635 :
7636 : /*
7637 : * Copy the checkpoint record to shared memory, so that checkpointer can
7638 : * work out the next time it wants to perform a restartpoint.
7639 : */
7640 1434 : SpinLockAcquire(&XLogCtl->info_lck);
7641 1434 : XLogCtl->lastCheckPointRecPtr = record->ReadRecPtr;
7642 1434 : XLogCtl->lastCheckPointEndPtr = record->EndRecPtr;
7643 1434 : XLogCtl->lastCheckPoint = *checkPoint;
7644 1434 : SpinLockRelease(&XLogCtl->info_lck);
7645 : }
7646 :
7647 : /*
7648 : * Establish a restartpoint if possible.
7649 : *
7650 : * This is similar to CreateCheckPoint, but is used during WAL recovery
7651 : * to establish a point from which recovery can roll forward without
7652 : * replaying the entire recovery log.
7653 : *
7654 : * Returns true if a new restartpoint was established. We can only establish
7655 : * a restartpoint if we have replayed a safe checkpoint record since last
7656 : * restartpoint.
7657 : */
7658 : bool
7659 1162 : CreateRestartPoint(int flags)
7660 : {
7661 : XLogRecPtr lastCheckPointRecPtr;
7662 : XLogRecPtr lastCheckPointEndPtr;
7663 : CheckPoint lastCheckPoint;
7664 : XLogRecPtr PriorRedoPtr;
7665 : XLogRecPtr receivePtr;
7666 : XLogRecPtr replayPtr;
7667 : TimeLineID replayTLI;
7668 : XLogRecPtr endptr;
7669 : XLogSegNo _logSegNo;
7670 : TimestampTz xtime;
7671 :
7672 : /* Concurrent checkpoint/restartpoint cannot happen */
7673 : Assert(!IsUnderPostmaster || MyBackendType == B_CHECKPOINTER);
7674 :
7675 : /* Get a local copy of the last safe checkpoint record. */
7676 1162 : SpinLockAcquire(&XLogCtl->info_lck);
7677 1162 : lastCheckPointRecPtr = XLogCtl->lastCheckPointRecPtr;
7678 1162 : lastCheckPointEndPtr = XLogCtl->lastCheckPointEndPtr;
7679 1162 : lastCheckPoint = XLogCtl->lastCheckPoint;
7680 1162 : SpinLockRelease(&XLogCtl->info_lck);
7681 :
7682 : /*
7683 : * Check that we're still in recovery mode. It's ok if we exit recovery
7684 : * mode after this check, the restart point is valid anyway.
7685 : */
7686 1162 : if (!RecoveryInProgress())
7687 : {
7688 0 : ereport(DEBUG2,
7689 : (errmsg_internal("skipping restartpoint, recovery has already ended")));
7690 0 : return false;
7691 : }
7692 :
7693 : /*
7694 : * If the last checkpoint record we've replayed is already our last
7695 : * restartpoint, we can't perform a new restart point. We still update
7696 : * minRecoveryPoint in that case, so that if this is a shutdown restart
7697 : * point, we won't start up earlier than before. That's not strictly
7698 : * necessary, but when hot standby is enabled, it would be rather weird if
7699 : * the database opened up for read-only connections at a point-in-time
7700 : * before the last shutdown. Such time travel is still possible in case of
7701 : * immediate shutdown, though.
7702 : *
7703 : * We don't explicitly advance minRecoveryPoint when we do create a
7704 : * restartpoint. It's assumed that flushing the buffers will do that as a
7705 : * side-effect.
7706 : */
7707 1162 : if (!XLogRecPtrIsValid(lastCheckPointRecPtr) ||
7708 496 : lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
7709 : {
7710 780 : ereport(DEBUG2,
7711 : errmsg_internal("skipping restartpoint, already performed at %X/%08X",
7712 : LSN_FORMAT_ARGS(lastCheckPoint.redo)));
7713 :
7714 780 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
7715 780 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7716 : {
7717 68 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7718 68 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7719 68 : UpdateControlFile();
7720 68 : LWLockRelease(ControlFileLock);
7721 : }
7722 780 : return false;
7723 : }
7724 :
7725 : /*
7726 : * Update the shared RedoRecPtr so that the startup process can calculate
7727 : * the number of segments replayed since last restartpoint, and request a
7728 : * restartpoint if it exceeds CheckPointSegments.
7729 : *
7730 : * Like in CreateCheckPoint(), hold off insertions to update it, although
7731 : * during recovery this is just pro forma, because no WAL insertions are
7732 : * happening.
7733 : */
7734 382 : WALInsertLockAcquireExclusive();
7735 382 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = lastCheckPoint.redo;
7736 382 : WALInsertLockRelease();
7737 :
7738 : /* Also update the info_lck-protected copy */
7739 382 : SpinLockAcquire(&XLogCtl->info_lck);
7740 382 : XLogCtl->RedoRecPtr = lastCheckPoint.redo;
7741 382 : SpinLockRelease(&XLogCtl->info_lck);
7742 :
7743 : /*
7744 : * Prepare to accumulate statistics.
7745 : *
7746 : * Note: because it is possible for log_checkpoints to change while a
7747 : * checkpoint proceeds, we always accumulate stats, even if
7748 : * log_checkpoints is currently off.
7749 : */
7750 4202 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7751 382 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7752 :
7753 382 : if (log_checkpoints)
7754 382 : LogCheckpointStart(flags, true);
7755 :
7756 : /* Update the process title */
7757 382 : update_checkpoint_display(flags, true, false);
7758 :
7759 382 : CheckPointGuts(lastCheckPoint.redo, flags);
7760 :
7761 : /*
7762 : * This location needs to be after CheckPointGuts() to ensure that some
7763 : * work has already happened during this checkpoint.
7764 : */
7765 382 : INJECTION_POINT("create-restart-point", NULL);
7766 :
7767 : /*
7768 : * Remember the prior checkpoint's redo ptr for
7769 : * UpdateCheckPointDistanceEstimate()
7770 : */
7771 382 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7772 :
7773 : /*
7774 : * Update pg_control, using current time. Check that it still shows an
7775 : * older checkpoint, else do nothing; this is a quick hack to make sure
7776 : * nothing really bad happens if somehow we get here after the
7777 : * end-of-recovery checkpoint.
7778 : */
7779 382 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7780 382 : if (ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
7781 : {
7782 : /*
7783 : * Update the checkpoint information. We do this even if the cluster
7784 : * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
7785 : * segments recycled below.
7786 : */
7787 382 : ControlFile->checkPoint = lastCheckPointRecPtr;
7788 382 : ControlFile->checkPointCopy = lastCheckPoint;
7789 :
7790 : /*
7791 : * Ensure minRecoveryPoint is past the checkpoint record and update it
7792 : * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
7793 : * this will have happened already while writing out dirty buffers,
7794 : * but not necessarily - e.g. because no buffers were dirtied. We do
7795 : * this because a backup performed in recovery uses minRecoveryPoint
7796 : * to determine which WAL files must be included in the backup, and
7797 : * the file (or files) containing the checkpoint record must be
7798 : * included, at a minimum. Note that for an ordinary restart of
7799 : * recovery there's no value in having the minimum recovery point any
7800 : * earlier than this anyway, because redo will begin just after the
7801 : * checkpoint record.
7802 : */
7803 382 : if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
7804 : {
7805 382 : if (ControlFile->minRecoveryPoint < lastCheckPointEndPtr)
7806 : {
7807 38 : ControlFile->minRecoveryPoint = lastCheckPointEndPtr;
7808 38 : ControlFile->minRecoveryPointTLI = lastCheckPoint.ThisTimeLineID;
7809 :
7810 : /* update local copy */
7811 38 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
7812 38 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
7813 : }
7814 382 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7815 42 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7816 : }
7817 382 : UpdateControlFile();
7818 : }
7819 382 : LWLockRelease(ControlFileLock);
7820 :
7821 : /*
7822 : * Update the average distance between checkpoints/restartpoints if the
7823 : * prior checkpoint exists.
7824 : */
7825 382 : if (XLogRecPtrIsValid(PriorRedoPtr))
7826 382 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7827 :
7828 : /*
7829 : * Delete old log files, those no longer needed for last restartpoint to
7830 : * prevent the disk holding the xlog from growing full.
7831 : */
7832 382 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7833 :
7834 : /*
7835 : * Retreat _logSegNo using the current end of xlog replayed or received,
7836 : * whichever is later.
7837 : */
7838 382 : receivePtr = GetWalRcvFlushRecPtr(NULL, NULL);
7839 382 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
7840 382 : endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
7841 382 : KeepLogSeg(endptr, &_logSegNo);
7842 382 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7843 : _logSegNo, InvalidOid,
7844 : InvalidTransactionId))
7845 : {
7846 : /*
7847 : * Some slots have been invalidated; recalculate the old-segment
7848 : * horizon, starting again from RedoRecPtr.
7849 : */
7850 2 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7851 2 : KeepLogSeg(endptr, &_logSegNo);
7852 : }
7853 382 : _logSegNo--;
7854 :
7855 : /*
7856 : * Try to recycle segments on a useful timeline. If we've been promoted
7857 : * since the beginning of this restartpoint, use the new timeline chosen
7858 : * at end of recovery. If we're still in recovery, use the timeline we're
7859 : * currently replaying.
7860 : *
7861 : * There is no guarantee that the WAL segments will be useful on the
7862 : * current timeline; if recovery proceeds to a new timeline right after
7863 : * this, the pre-allocated WAL segments on this timeline will not be used,
7864 : * and will go wasted until recycled on the next restartpoint. We'll live
7865 : * with that.
7866 : */
7867 382 : if (!RecoveryInProgress())
7868 0 : replayTLI = XLogCtl->InsertTimeLineID;
7869 :
7870 382 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, endptr, replayTLI);
7871 :
7872 : /*
7873 : * Make more log segments if needed. (Do this after recycling old log
7874 : * segments, since that may supply some of the needed files.)
7875 : */
7876 382 : PreallocXlogFiles(endptr, replayTLI);
7877 :
7878 : /*
7879 : * Truncate pg_subtrans if possible. We can throw away all data before
7880 : * the oldest XMIN of any running transaction. No future transaction will
7881 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7882 : * in subtrans.c). When hot standby is disabled, though, we mustn't do
7883 : * this because StartupSUBTRANS hasn't been called yet.
7884 : */
7885 382 : if (EnableHotStandby)
7886 382 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7887 :
7888 : /* Real work is done; log and update stats. */
7889 382 : LogCheckpointEnd(true);
7890 :
7891 : /* Reset the process title */
7892 382 : update_checkpoint_display(flags, true, true);
7893 :
7894 382 : xtime = GetLatestXTime();
7895 382 : ereport((log_checkpoints ? LOG : DEBUG2),
7896 : errmsg("recovery restart point at %X/%08X",
7897 : LSN_FORMAT_ARGS(lastCheckPoint.redo)),
7898 : xtime ? errdetail("Last completed transaction was at log time %s.",
7899 : timestamptz_to_str(xtime)) : 0);
7900 :
7901 : /*
7902 : * Finally, execute archive_cleanup_command, if any.
7903 : */
7904 382 : if (archiveCleanupCommand && strcmp(archiveCleanupCommand, "") != 0)
7905 0 : ExecuteRecoveryCommand(archiveCleanupCommand,
7906 : "archive_cleanup_command",
7907 : false,
7908 : WAIT_EVENT_ARCHIVE_CLEANUP_COMMAND);
7909 :
7910 382 : return true;
7911 : }
7912 :
7913 : /*
7914 : * Report availability of WAL for the given target LSN
7915 : * (typically a slot's restart_lsn)
7916 : *
7917 : * Returns one of the following enum values:
7918 : *
7919 : * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
7920 : * max_wal_size.
7921 : *
7922 : * * WALAVAIL_EXTENDED means it is still available by preserving extra
7923 : * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
7924 : * than max_wal_size, this state is not returned.
7925 : *
7926 : * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
7927 : * remove reserved segments. The walsender using this slot may return to the
7928 : * above.
7929 : *
7930 : * * WALAVAIL_REMOVED means it has been removed. A replication stream on
7931 : * a slot with this LSN cannot continue. (Any associated walsender
7932 : * processes should have been terminated already.)
7933 : *
7934 : * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
7935 : */
7936 : WALAvailability
7937 1068 : GetWALAvailability(XLogRecPtr targetLSN)
7938 : {
7939 : XLogRecPtr currpos; /* current write LSN */
7940 : XLogSegNo currSeg; /* segid of currpos */
7941 : XLogSegNo targetSeg; /* segid of targetLSN */
7942 : XLogSegNo oldestSeg; /* actual oldest segid */
7943 : XLogSegNo oldestSegMaxWalSize; /* oldest segid kept by max_wal_size */
7944 : XLogSegNo oldestSlotSeg; /* oldest segid kept by slot */
7945 : uint64 keepSegs;
7946 :
7947 : /*
7948 : * slot does not reserve WAL. Either deactivated, or has never been active
7949 : */
7950 1068 : if (!XLogRecPtrIsValid(targetLSN))
7951 76 : return WALAVAIL_INVALID_LSN;
7952 :
7953 : /*
7954 : * Calculate the oldest segment currently reserved by all slots,
7955 : * considering wal_keep_size and max_slot_wal_keep_size. Initialize
7956 : * oldestSlotSeg to the current segment.
7957 : */
7958 992 : currpos = GetXLogWriteRecPtr();
7959 992 : XLByteToSeg(currpos, oldestSlotSeg, wal_segment_size);
7960 992 : KeepLogSeg(currpos, &oldestSlotSeg);
7961 :
7962 : /*
7963 : * Find the oldest extant segment file. We get 1 until checkpoint removes
7964 : * the first WAL segment file since startup, which causes the status being
7965 : * wrong under certain abnormal conditions but that doesn't actually harm.
7966 : */
7967 992 : oldestSeg = XLogGetLastRemovedSegno() + 1;
7968 :
7969 : /* calculate oldest segment by max_wal_size */
7970 992 : XLByteToSeg(currpos, currSeg, wal_segment_size);
7971 992 : keepSegs = ConvertToXSegs(max_wal_size_mb, wal_segment_size) + 1;
7972 :
7973 992 : if (currSeg > keepSegs)
7974 16 : oldestSegMaxWalSize = currSeg - keepSegs;
7975 : else
7976 976 : oldestSegMaxWalSize = 1;
7977 :
7978 : /* the segment we care about */
7979 992 : XLByteToSeg(targetLSN, targetSeg, wal_segment_size);
7980 :
7981 : /*
7982 : * No point in returning reserved or extended status values if the
7983 : * targetSeg is known to be lost.
7984 : */
7985 992 : if (targetSeg >= oldestSlotSeg)
7986 : {
7987 : /* show "reserved" when targetSeg is within max_wal_size */
7988 990 : if (targetSeg >= oldestSegMaxWalSize)
7989 986 : return WALAVAIL_RESERVED;
7990 :
7991 : /* being retained by slots exceeding max_wal_size */
7992 4 : return WALAVAIL_EXTENDED;
7993 : }
7994 :
7995 : /* WAL segments are no longer retained but haven't been removed yet */
7996 2 : if (targetSeg >= oldestSeg)
7997 2 : return WALAVAIL_UNRESERVED;
7998 :
7999 : /* Definitely lost */
8000 0 : return WALAVAIL_REMOVED;
8001 : }
8002 :
8003 :
8004 : /*
8005 : * Retreat *logSegNo to the last segment that we need to retain because of
8006 : * either wal_keep_size or replication slots.
8007 : *
8008 : * This is calculated by subtracting wal_keep_size from the given xlog
8009 : * location, recptr and by making sure that that result is below the
8010 : * requirement of replication slots. For the latter criterion we do consider
8011 : * the effects of max_slot_wal_keep_size: reserve at most that much space back
8012 : * from recptr.
8013 : *
8014 : * Note about replication slots: if this function calculates a value
8015 : * that's further ahead than what slots need reserved, then affected
8016 : * slots need to be invalidated and this function invoked again.
8017 : * XXX it might be a good idea to rewrite this function so that
8018 : * invalidation is optionally done here, instead.
8019 : */
8020 : static void
8021 4468 : KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
8022 : {
8023 : XLogSegNo currSegNo;
8024 : XLogSegNo segno;
8025 : XLogRecPtr keep;
8026 :
8027 4468 : XLByteToSeg(recptr, currSegNo, wal_segment_size);
8028 4468 : segno = currSegNo;
8029 :
8030 : /* Calculate how many segments are kept by slots. */
8031 4468 : keep = XLogGetReplicationSlotMinimumLSN();
8032 4468 : if (XLogRecPtrIsValid(keep) && keep < recptr)
8033 : {
8034 1310 : XLByteToSeg(keep, segno, wal_segment_size);
8035 :
8036 : /*
8037 : * Account for max_slot_wal_keep_size to avoid keeping more than
8038 : * configured. However, don't do that during a binary upgrade: if
8039 : * slots were to be invalidated because of this, it would not be
8040 : * possible to preserve logical ones during the upgrade.
8041 : */
8042 1310 : if (max_slot_wal_keep_size_mb >= 0 && !IsBinaryUpgrade)
8043 : {
8044 : uint64 slot_keep_segs;
8045 :
8046 42 : slot_keep_segs =
8047 42 : ConvertToXSegs(max_slot_wal_keep_size_mb, wal_segment_size);
8048 :
8049 42 : if (currSegNo - segno > slot_keep_segs)
8050 10 : segno = currSegNo - slot_keep_segs;
8051 : }
8052 : }
8053 :
8054 : /*
8055 : * If WAL summarization is in use, don't remove WAL that has yet to be
8056 : * summarized.
8057 : */
8058 4468 : keep = GetOldestUnsummarizedLSN(NULL, NULL);
8059 4468 : if (XLogRecPtrIsValid(keep))
8060 : {
8061 : XLogSegNo unsummarized_segno;
8062 :
8063 4 : XLByteToSeg(keep, unsummarized_segno, wal_segment_size);
8064 4 : if (unsummarized_segno < segno)
8065 4 : segno = unsummarized_segno;
8066 : }
8067 :
8068 : /* but, keep at least wal_keep_size if that's set */
8069 4468 : if (wal_keep_size_mb > 0)
8070 : {
8071 : uint64 keep_segs;
8072 :
8073 148 : keep_segs = ConvertToXSegs(wal_keep_size_mb, wal_segment_size);
8074 148 : if (currSegNo - segno < keep_segs)
8075 : {
8076 : /* avoid underflow, don't go below 1 */
8077 148 : if (currSegNo <= keep_segs)
8078 140 : segno = 1;
8079 : else
8080 8 : segno = currSegNo - keep_segs;
8081 : }
8082 : }
8083 :
8084 : /* don't delete WAL segments newer than the calculated segment */
8085 4468 : if (segno < *logSegNo)
8086 760 : *logSegNo = segno;
8087 4468 : }
8088 :
8089 : /*
8090 : * Write a NEXTOID log record
8091 : */
8092 : void
8093 1228 : XLogPutNextOid(Oid nextOid)
8094 : {
8095 1228 : XLogBeginInsert();
8096 1228 : XLogRegisterData(&nextOid, sizeof(Oid));
8097 1228 : (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID);
8098 :
8099 : /*
8100 : * We need not flush the NEXTOID record immediately, because any of the
8101 : * just-allocated OIDs could only reach disk as part of a tuple insert or
8102 : * update that would have its own XLOG record that must follow the NEXTOID
8103 : * record. Therefore, the standard buffer LSN interlock applied to those
8104 : * records will ensure no such OID reaches disk before the NEXTOID record
8105 : * does.
8106 : *
8107 : * Note, however, that the above statement only covers state "within" the
8108 : * database. When we use a generated OID as a file or directory name, we
8109 : * are in a sense violating the basic WAL rule, because that filesystem
8110 : * change may reach disk before the NEXTOID WAL record does. The impact
8111 : * of this is that if a database crash occurs immediately afterward, we
8112 : * might after restart re-generate the same OID and find that it conflicts
8113 : * with the leftover file or directory. But since for safety's sake we
8114 : * always loop until finding a nonconflicting filename, this poses no real
8115 : * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8116 : */
8117 1228 : }
8118 :
8119 : /*
8120 : * Write an XLOG SWITCH record.
8121 : *
8122 : * Here we just blindly issue an XLogInsert request for the record.
8123 : * All the magic happens inside XLogInsert.
8124 : *
8125 : * The return value is either the end+1 address of the switch record,
8126 : * or the end+1 address of the prior segment if we did not need to
8127 : * write a switch record because we are already at segment start.
8128 : */
8129 : XLogRecPtr
8130 1568 : RequestXLogSwitch(bool mark_unimportant)
8131 : {
8132 : XLogRecPtr RecPtr;
8133 :
8134 : /* XLOG SWITCH has no data */
8135 1568 : XLogBeginInsert();
8136 :
8137 1568 : if (mark_unimportant)
8138 0 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8139 1568 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH);
8140 :
8141 1568 : return RecPtr;
8142 : }
8143 :
8144 : /*
8145 : * Write a RESTORE POINT record
8146 : */
8147 : XLogRecPtr
8148 6 : XLogRestorePoint(const char *rpName)
8149 : {
8150 : XLogRecPtr RecPtr;
8151 : xl_restore_point xlrec;
8152 :
8153 6 : xlrec.rp_time = GetCurrentTimestamp();
8154 6 : strlcpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8155 :
8156 6 : XLogBeginInsert();
8157 6 : XLogRegisterData(&xlrec, sizeof(xl_restore_point));
8158 :
8159 6 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT);
8160 :
8161 6 : ereport(LOG,
8162 : errmsg("restore point \"%s\" created at %X/%08X",
8163 : rpName, LSN_FORMAT_ARGS(RecPtr)));
8164 :
8165 6 : return RecPtr;
8166 : }
8167 :
8168 : /*
8169 : * Check if any of the GUC parameters that are critical for hot standby
8170 : * have changed, and update the value in pg_control file if necessary.
8171 : */
8172 : static void
8173 1798 : XLogReportParameters(void)
8174 : {
8175 1798 : if (wal_level != ControlFile->wal_level ||
8176 1324 : wal_log_hints != ControlFile->wal_log_hints ||
8177 1154 : MaxConnections != ControlFile->MaxConnections ||
8178 1152 : max_worker_processes != ControlFile->max_worker_processes ||
8179 1150 : max_wal_senders != ControlFile->max_wal_senders ||
8180 1108 : max_prepared_xacts != ControlFile->max_prepared_xacts ||
8181 916 : max_locks_per_xact != ControlFile->max_locks_per_xact ||
8182 916 : track_commit_timestamp != ControlFile->track_commit_timestamp)
8183 : {
8184 : /*
8185 : * The change in number of backend slots doesn't need to be WAL-logged
8186 : * if archiving is not enabled, as you can't start archive recovery
8187 : * with wal_level=minimal anyway. We don't really care about the
8188 : * values in pg_control either if wal_level=minimal, but seems better
8189 : * to keep them up-to-date to avoid confusion.
8190 : */
8191 906 : if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8192 : {
8193 : xl_parameter_change xlrec;
8194 : XLogRecPtr recptr;
8195 :
8196 866 : xlrec.MaxConnections = MaxConnections;
8197 866 : xlrec.max_worker_processes = max_worker_processes;
8198 866 : xlrec.max_wal_senders = max_wal_senders;
8199 866 : xlrec.max_prepared_xacts = max_prepared_xacts;
8200 866 : xlrec.max_locks_per_xact = max_locks_per_xact;
8201 866 : xlrec.wal_level = wal_level;
8202 866 : xlrec.wal_log_hints = wal_log_hints;
8203 866 : xlrec.track_commit_timestamp = track_commit_timestamp;
8204 :
8205 866 : XLogBeginInsert();
8206 866 : XLogRegisterData(&xlrec, sizeof(xlrec));
8207 :
8208 866 : recptr = XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE);
8209 866 : XLogFlush(recptr);
8210 : }
8211 :
8212 906 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8213 :
8214 906 : ControlFile->MaxConnections = MaxConnections;
8215 906 : ControlFile->max_worker_processes = max_worker_processes;
8216 906 : ControlFile->max_wal_senders = max_wal_senders;
8217 906 : ControlFile->max_prepared_xacts = max_prepared_xacts;
8218 906 : ControlFile->max_locks_per_xact = max_locks_per_xact;
8219 906 : ControlFile->wal_level = wal_level;
8220 906 : ControlFile->wal_log_hints = wal_log_hints;
8221 906 : ControlFile->track_commit_timestamp = track_commit_timestamp;
8222 906 : UpdateControlFile();
8223 :
8224 906 : LWLockRelease(ControlFileLock);
8225 : }
8226 1798 : }
8227 :
8228 : /*
8229 : * Update full_page_writes in shared memory, and write an
8230 : * XLOG_FPW_CHANGE record if necessary.
8231 : *
8232 : * Note: this function assumes there is no other process running
8233 : * concurrently that could update it.
8234 : */
8235 : void
8236 3000 : UpdateFullPageWrites(void)
8237 : {
8238 3000 : XLogCtlInsert *Insert = &XLogCtl->Insert;
8239 : bool recoveryInProgress;
8240 :
8241 : /*
8242 : * Do nothing if full_page_writes has not been changed.
8243 : *
8244 : * It's safe to check the shared full_page_writes without the lock,
8245 : * because we assume that there is no concurrently running process which
8246 : * can update it.
8247 : */
8248 3000 : if (fullPageWrites == Insert->fullPageWrites)
8249 2238 : return;
8250 :
8251 : /*
8252 : * Perform this outside critical section so that the WAL insert
8253 : * initialization done by RecoveryInProgress() doesn't trigger an
8254 : * assertion failure.
8255 : */
8256 762 : recoveryInProgress = RecoveryInProgress();
8257 :
8258 762 : START_CRIT_SECTION();
8259 :
8260 : /*
8261 : * It's always safe to take full page images, even when not strictly
8262 : * required, but not the other round. So if we're setting full_page_writes
8263 : * to true, first set it true and then write the WAL record. If we're
8264 : * setting it to false, first write the WAL record and then set the global
8265 : * flag.
8266 : */
8267 762 : if (fullPageWrites)
8268 : {
8269 742 : WALInsertLockAcquireExclusive();
8270 742 : Insert->fullPageWrites = true;
8271 742 : WALInsertLockRelease();
8272 : }
8273 :
8274 : /*
8275 : * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8276 : * full_page_writes during archive recovery, if required.
8277 : */
8278 762 : if (XLogStandbyInfoActive() && !recoveryInProgress)
8279 : {
8280 0 : XLogBeginInsert();
8281 0 : XLogRegisterData(&fullPageWrites, sizeof(bool));
8282 :
8283 0 : XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
8284 : }
8285 :
8286 762 : if (!fullPageWrites)
8287 : {
8288 20 : WALInsertLockAcquireExclusive();
8289 20 : Insert->fullPageWrites = false;
8290 20 : WALInsertLockRelease();
8291 : }
8292 762 : END_CRIT_SECTION();
8293 : }
8294 :
8295 : /*
8296 : * XLOG resource manager's routines
8297 : *
8298 : * Definitions of info values are in include/catalog/pg_control.h, though
8299 : * not all record types are related to control file updates.
8300 : *
8301 : * NOTE: Some XLOG record types that are directly related to WAL recovery
8302 : * are handled in xlogrecovery_redo().
8303 : */
8304 : void
8305 87696 : xlog_redo(XLogReaderState *record)
8306 : {
8307 87696 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8308 87696 : XLogRecPtr lsn = record->EndRecPtr;
8309 :
8310 : /*
8311 : * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
8312 : * XLOG_FPI_FOR_HINT records.
8313 : */
8314 : Assert(info == XLOG_FPI || info == XLOG_FPI_FOR_HINT ||
8315 : !XLogRecHasAnyBlockRefs(record));
8316 :
8317 87696 : if (info == XLOG_NEXTOID)
8318 : {
8319 : Oid nextOid;
8320 :
8321 : /*
8322 : * We used to try to take the maximum of TransamVariables->nextOid and
8323 : * the recorded nextOid, but that fails if the OID counter wraps
8324 : * around. Since no OID allocation should be happening during replay
8325 : * anyway, better to just believe the record exactly. We still take
8326 : * OidGenLock while setting the variable, just in case.
8327 : */
8328 186 : memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
8329 186 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8330 186 : TransamVariables->nextOid = nextOid;
8331 186 : TransamVariables->oidCount = 0;
8332 186 : LWLockRelease(OidGenLock);
8333 : }
8334 87510 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
8335 : {
8336 : CheckPoint checkPoint;
8337 : TimeLineID replayTLI;
8338 :
8339 70 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8340 : /* In a SHUTDOWN checkpoint, believe the counters exactly */
8341 70 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8342 70 : TransamVariables->nextXid = checkPoint.nextXid;
8343 70 : LWLockRelease(XidGenLock);
8344 70 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8345 70 : TransamVariables->nextOid = checkPoint.nextOid;
8346 70 : TransamVariables->oidCount = 0;
8347 70 : LWLockRelease(OidGenLock);
8348 70 : MultiXactSetNextMXact(checkPoint.nextMulti,
8349 : checkPoint.nextMultiOffset);
8350 :
8351 70 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8352 : checkPoint.oldestMultiDB);
8353 :
8354 : /*
8355 : * No need to set oldestClogXid here as well; it'll be set when we
8356 : * redo an xl_clog_truncate if it changed since initialization.
8357 : */
8358 70 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
8359 :
8360 : /*
8361 : * If we see a shutdown checkpoint while waiting for an end-of-backup
8362 : * record, the backup was canceled and the end-of-backup record will
8363 : * never arrive.
8364 : */
8365 70 : if (ArchiveRecoveryRequested &&
8366 68 : XLogRecPtrIsValid(ControlFile->backupStartPoint) &&
8367 0 : !XLogRecPtrIsValid(ControlFile->backupEndPoint))
8368 0 : ereport(PANIC,
8369 : (errmsg("online backup was canceled, recovery cannot continue")));
8370 :
8371 : /*
8372 : * If we see a shutdown checkpoint, we know that nothing was running
8373 : * on the primary at this point. So fake-up an empty running-xacts
8374 : * record and use that here and now. Recover additional standby state
8375 : * for prepared transactions.
8376 : */
8377 70 : if (standbyState >= STANDBY_INITIALIZED)
8378 : {
8379 : TransactionId *xids;
8380 : int nxids;
8381 : TransactionId oldestActiveXID;
8382 : TransactionId latestCompletedXid;
8383 : RunningTransactionsData running;
8384 :
8385 64 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
8386 :
8387 : /* Update pg_subtrans entries for any prepared transactions */
8388 64 : StandbyRecoverPreparedTransactions();
8389 :
8390 : /*
8391 : * Construct a RunningTransactions snapshot representing a shut
8392 : * down server, with only prepared transactions still alive. We're
8393 : * never overflowed at this point because all subxids are listed
8394 : * with their parent prepared transactions.
8395 : */
8396 64 : running.xcnt = nxids;
8397 64 : running.subxcnt = 0;
8398 64 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
8399 64 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
8400 64 : running.oldestRunningXid = oldestActiveXID;
8401 64 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
8402 64 : TransactionIdRetreat(latestCompletedXid);
8403 : Assert(TransactionIdIsNormal(latestCompletedXid));
8404 64 : running.latestCompletedXid = latestCompletedXid;
8405 64 : running.xids = xids;
8406 :
8407 64 : ProcArrayApplyRecoveryInfo(&running);
8408 : }
8409 :
8410 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8411 70 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8412 70 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8413 70 : LWLockRelease(ControlFileLock);
8414 :
8415 : /*
8416 : * We should've already switched to the new TLI before replaying this
8417 : * record.
8418 : */
8419 70 : (void) GetCurrentReplayRecPtr(&replayTLI);
8420 70 : if (checkPoint.ThisTimeLineID != replayTLI)
8421 0 : ereport(PANIC,
8422 : (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
8423 : checkPoint.ThisTimeLineID, replayTLI)));
8424 :
8425 70 : RecoveryRestartPoint(&checkPoint, record);
8426 :
8427 : /*
8428 : * After replaying a checkpoint record, free all smgr objects.
8429 : * Otherwise we would never do so for dropped relations, as the
8430 : * startup does not process shared invalidation messages or call
8431 : * AtEOXact_SMgr().
8432 : */
8433 70 : smgrdestroyall();
8434 : }
8435 87440 : else if (info == XLOG_CHECKPOINT_ONLINE)
8436 : {
8437 : CheckPoint checkPoint;
8438 : TimeLineID replayTLI;
8439 :
8440 1368 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8441 : /* In an ONLINE checkpoint, treat the XID counter as a minimum */
8442 1368 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8443 1368 : if (FullTransactionIdPrecedes(TransamVariables->nextXid,
8444 : checkPoint.nextXid))
8445 0 : TransamVariables->nextXid = checkPoint.nextXid;
8446 1368 : LWLockRelease(XidGenLock);
8447 :
8448 : /*
8449 : * We ignore the nextOid counter in an ONLINE checkpoint, preferring
8450 : * to track OID assignment through XLOG_NEXTOID records. The nextOid
8451 : * counter is from the start of the checkpoint and might well be stale
8452 : * compared to later XLOG_NEXTOID records. We could try to take the
8453 : * maximum of the nextOid counter and our latest value, but since
8454 : * there's no particular guarantee about the speed with which the OID
8455 : * counter wraps around, that's a risky thing to do. In any case,
8456 : * users of the nextOid counter are required to avoid assignment of
8457 : * duplicates, so that a somewhat out-of-date value should be safe.
8458 : */
8459 :
8460 : /* Handle multixact */
8461 1368 : MultiXactAdvanceNextMXact(checkPoint.nextMulti,
8462 : checkPoint.nextMultiOffset);
8463 :
8464 : /*
8465 : * NB: This may perform multixact truncation when replaying WAL
8466 : * generated by an older primary.
8467 : */
8468 1368 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8469 : checkPoint.oldestMultiDB);
8470 1368 : if (TransactionIdPrecedes(TransamVariables->oldestXid,
8471 : checkPoint.oldestXid))
8472 0 : SetTransactionIdLimit(checkPoint.oldestXid,
8473 : checkPoint.oldestXidDB);
8474 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8475 1368 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8476 1368 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8477 1368 : LWLockRelease(ControlFileLock);
8478 :
8479 : /* TLI should not change in an on-line checkpoint */
8480 1368 : (void) GetCurrentReplayRecPtr(&replayTLI);
8481 1368 : if (checkPoint.ThisTimeLineID != replayTLI)
8482 0 : ereport(PANIC,
8483 : (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
8484 : checkPoint.ThisTimeLineID, replayTLI)));
8485 :
8486 1368 : RecoveryRestartPoint(&checkPoint, record);
8487 :
8488 : /*
8489 : * After replaying a checkpoint record, free all smgr objects.
8490 : * Otherwise we would never do so for dropped relations, as the
8491 : * startup does not process shared invalidation messages or call
8492 : * AtEOXact_SMgr().
8493 : */
8494 1368 : smgrdestroyall();
8495 : }
8496 86072 : else if (info == XLOG_OVERWRITE_CONTRECORD)
8497 : {
8498 : /* nothing to do here, handled in xlogrecovery_redo() */
8499 : }
8500 86070 : else if (info == XLOG_END_OF_RECOVERY)
8501 : {
8502 : xl_end_of_recovery xlrec;
8503 : TimeLineID replayTLI;
8504 :
8505 20 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
8506 :
8507 : /*
8508 : * For Hot Standby, we could treat this like a Shutdown Checkpoint,
8509 : * but this case is rarer and harder to test, so the benefit doesn't
8510 : * outweigh the potential extra cost of maintenance.
8511 : */
8512 :
8513 : /*
8514 : * We should've already switched to the new TLI before replaying this
8515 : * record.
8516 : */
8517 20 : (void) GetCurrentReplayRecPtr(&replayTLI);
8518 20 : if (xlrec.ThisTimeLineID != replayTLI)
8519 0 : ereport(PANIC,
8520 : (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
8521 : xlrec.ThisTimeLineID, replayTLI)));
8522 : }
8523 86050 : else if (info == XLOG_NOOP)
8524 : {
8525 : /* nothing to do here */
8526 : }
8527 86050 : else if (info == XLOG_SWITCH)
8528 : {
8529 : /* nothing to do here */
8530 : }
8531 85154 : else if (info == XLOG_RESTORE_POINT)
8532 : {
8533 : /* nothing to do here, handled in xlogrecovery.c */
8534 : }
8535 85144 : else if (info == XLOG_FPI || info == XLOG_FPI_FOR_HINT)
8536 : {
8537 : /*
8538 : * XLOG_FPI records contain nothing else but one or more block
8539 : * references. Every block reference must include a full-page image
8540 : * even if full_page_writes was disabled when the record was generated
8541 : * - otherwise there would be no point in this record.
8542 : *
8543 : * XLOG_FPI_FOR_HINT records are generated when a page needs to be
8544 : * WAL-logged because of a hint bit update. They are only generated
8545 : * when checksums and/or wal_log_hints are enabled. They may include
8546 : * no full-page images if full_page_writes was disabled when they were
8547 : * generated. In this case there is nothing to do here.
8548 : *
8549 : * No recovery conflicts are generated by these generic records - if a
8550 : * resource manager needs to generate conflicts, it has to define a
8551 : * separate WAL record type and redo routine.
8552 : */
8553 176748 : for (uint8 block_id = 0; block_id <= XLogRecMaxBlockId(record); block_id++)
8554 : {
8555 : Buffer buffer;
8556 :
8557 93212 : if (!XLogRecHasBlockImage(record, block_id))
8558 : {
8559 132 : if (info == XLOG_FPI)
8560 0 : elog(ERROR, "XLOG_FPI record did not contain a full-page image");
8561 132 : continue;
8562 : }
8563 :
8564 93080 : if (XLogReadBufferForRedo(record, block_id, &buffer) != BLK_RESTORED)
8565 0 : elog(ERROR, "unexpected XLogReadBufferForRedo result when restoring backup block");
8566 93080 : UnlockReleaseBuffer(buffer);
8567 : }
8568 : }
8569 1608 : else if (info == XLOG_BACKUP_END)
8570 : {
8571 : /* nothing to do here, handled in xlogrecovery_redo() */
8572 : }
8573 1434 : else if (info == XLOG_PARAMETER_CHANGE)
8574 : {
8575 : xl_parameter_change xlrec;
8576 :
8577 : /* Update our copy of the parameters in pg_control */
8578 64 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
8579 :
8580 : /*
8581 : * Invalidate logical slots if we are in hot standby and the primary
8582 : * does not have a WAL level sufficient for logical decoding. No need
8583 : * to search for potentially conflicting logically slots if standby is
8584 : * running with wal_level lower than logical, because in that case, we
8585 : * would have either disallowed creation of logical slots or
8586 : * invalidated existing ones.
8587 : */
8588 64 : if (InRecovery && InHotStandby &&
8589 34 : xlrec.wal_level < WAL_LEVEL_LOGICAL &&
8590 12 : wal_level >= WAL_LEVEL_LOGICAL)
8591 8 : InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_LEVEL,
8592 : 0, InvalidOid,
8593 : InvalidTransactionId);
8594 :
8595 64 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8596 64 : ControlFile->MaxConnections = xlrec.MaxConnections;
8597 64 : ControlFile->max_worker_processes = xlrec.max_worker_processes;
8598 64 : ControlFile->max_wal_senders = xlrec.max_wal_senders;
8599 64 : ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
8600 64 : ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
8601 64 : ControlFile->wal_level = xlrec.wal_level;
8602 64 : ControlFile->wal_log_hints = xlrec.wal_log_hints;
8603 :
8604 : /*
8605 : * Update minRecoveryPoint to ensure that if recovery is aborted, we
8606 : * recover back up to this point before allowing hot standby again.
8607 : * This is important if the max_* settings are decreased, to ensure
8608 : * you don't run queries against the WAL preceding the change. The
8609 : * local copies cannot be updated as long as crash recovery is
8610 : * happening and we expect all the WAL to be replayed.
8611 : */
8612 64 : if (InArchiveRecovery)
8613 : {
8614 36 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
8615 36 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
8616 : }
8617 64 : if (XLogRecPtrIsValid(LocalMinRecoveryPoint) && LocalMinRecoveryPoint < lsn)
8618 : {
8619 : TimeLineID replayTLI;
8620 :
8621 14 : (void) GetCurrentReplayRecPtr(&replayTLI);
8622 14 : ControlFile->minRecoveryPoint = lsn;
8623 14 : ControlFile->minRecoveryPointTLI = replayTLI;
8624 : }
8625 :
8626 64 : CommitTsParameterChange(xlrec.track_commit_timestamp,
8627 64 : ControlFile->track_commit_timestamp);
8628 64 : ControlFile->track_commit_timestamp = xlrec.track_commit_timestamp;
8629 :
8630 64 : UpdateControlFile();
8631 64 : LWLockRelease(ControlFileLock);
8632 :
8633 : /* Check to see if any parameter change gives a problem on recovery */
8634 64 : CheckRequiredParameterValues();
8635 : }
8636 1370 : else if (info == XLOG_FPW_CHANGE)
8637 : {
8638 : bool fpw;
8639 :
8640 0 : memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
8641 :
8642 : /*
8643 : * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
8644 : * do_pg_backup_start() and do_pg_backup_stop() can check whether
8645 : * full_page_writes has been disabled during online backup.
8646 : */
8647 0 : if (!fpw)
8648 : {
8649 0 : SpinLockAcquire(&XLogCtl->info_lck);
8650 0 : if (XLogCtl->lastFpwDisableRecPtr < record->ReadRecPtr)
8651 0 : XLogCtl->lastFpwDisableRecPtr = record->ReadRecPtr;
8652 0 : SpinLockRelease(&XLogCtl->info_lck);
8653 : }
8654 :
8655 : /* Keep track of full_page_writes */
8656 0 : lastFullPageWrites = fpw;
8657 : }
8658 : else if (info == XLOG_CHECKPOINT_REDO)
8659 : {
8660 : /* nothing to do here, just for informational purposes */
8661 : }
8662 87692 : }
8663 :
8664 : /*
8665 : * Return the extra open flags used for opening a file, depending on the
8666 : * value of the GUCs wal_sync_method, fsync and debug_io_direct.
8667 : */
8668 : static int
8669 32120 : get_sync_bit(int method)
8670 : {
8671 32120 : int o_direct_flag = 0;
8672 :
8673 : /*
8674 : * Use O_DIRECT if requested, except in walreceiver process. The WAL
8675 : * written by walreceiver is normally read by the startup process soon
8676 : * after it's written. Also, walreceiver performs unaligned writes, which
8677 : * don't work with O_DIRECT, so it is required for correctness too.
8678 : */
8679 32120 : if ((io_direct_flags & IO_DIRECT_WAL) && !AmWalReceiverProcess())
8680 16 : o_direct_flag = PG_O_DIRECT;
8681 :
8682 : /* If fsync is disabled, never open in sync mode */
8683 32120 : if (!enableFsync)
8684 32120 : return o_direct_flag;
8685 :
8686 0 : switch (method)
8687 : {
8688 : /*
8689 : * enum values for all sync options are defined even if they are
8690 : * not supported on the current platform. But if not, they are
8691 : * not included in the enum option array, and therefore will never
8692 : * be seen here.
8693 : */
8694 0 : case WAL_SYNC_METHOD_FSYNC:
8695 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8696 : case WAL_SYNC_METHOD_FDATASYNC:
8697 0 : return o_direct_flag;
8698 : #ifdef O_SYNC
8699 0 : case WAL_SYNC_METHOD_OPEN:
8700 0 : return O_SYNC | o_direct_flag;
8701 : #endif
8702 : #ifdef O_DSYNC
8703 0 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8704 0 : return O_DSYNC | o_direct_flag;
8705 : #endif
8706 0 : default:
8707 : /* can't happen (unless we are out of sync with option array) */
8708 0 : elog(ERROR, "unrecognized \"wal_sync_method\": %d", method);
8709 : return 0; /* silence warning */
8710 : }
8711 : }
8712 :
8713 : /*
8714 : * GUC support
8715 : */
8716 : void
8717 2276 : assign_wal_sync_method(int new_wal_sync_method, void *extra)
8718 : {
8719 2276 : if (wal_sync_method != new_wal_sync_method)
8720 : {
8721 : /*
8722 : * To ensure that no blocks escape unsynced, force an fsync on the
8723 : * currently open log segment (if any). Also, if the open flag is
8724 : * changing, close the log file so it will be reopened (with new flag
8725 : * bit) at next use.
8726 : */
8727 0 : if (openLogFile >= 0)
8728 : {
8729 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN);
8730 0 : if (pg_fsync(openLogFile) != 0)
8731 : {
8732 : char xlogfname[MAXFNAMELEN];
8733 : int save_errno;
8734 :
8735 0 : save_errno = errno;
8736 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo,
8737 : wal_segment_size);
8738 0 : errno = save_errno;
8739 0 : ereport(PANIC,
8740 : (errcode_for_file_access(),
8741 : errmsg("could not fsync file \"%s\": %m", xlogfname)));
8742 : }
8743 :
8744 0 : pgstat_report_wait_end();
8745 0 : if (get_sync_bit(wal_sync_method) != get_sync_bit(new_wal_sync_method))
8746 0 : XLogFileClose();
8747 : }
8748 : }
8749 2276 : }
8750 :
8751 :
8752 : /*
8753 : * Issue appropriate kind of fsync (if any) for an XLOG output file.
8754 : *
8755 : * 'fd' is a file descriptor for the XLOG file to be fsync'd.
8756 : * 'segno' is for error reporting purposes.
8757 : */
8758 : void
8759 314076 : issue_xlog_fsync(int fd, XLogSegNo segno, TimeLineID tli)
8760 : {
8761 314076 : char *msg = NULL;
8762 : instr_time start;
8763 :
8764 : Assert(tli != 0);
8765 :
8766 : /*
8767 : * Quick exit if fsync is disabled or write() has already synced the WAL
8768 : * file.
8769 : */
8770 314076 : if (!enableFsync ||
8771 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN ||
8772 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN_DSYNC)
8773 314076 : return;
8774 :
8775 : /*
8776 : * Measure I/O timing to sync the WAL file for pg_stat_io.
8777 : */
8778 0 : start = pgstat_prepare_io_time(track_wal_io_timing);
8779 :
8780 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC);
8781 0 : switch (wal_sync_method)
8782 : {
8783 0 : case WAL_SYNC_METHOD_FSYNC:
8784 0 : if (pg_fsync_no_writethrough(fd) != 0)
8785 0 : msg = _("could not fsync file \"%s\": %m");
8786 0 : break;
8787 : #ifdef HAVE_FSYNC_WRITETHROUGH
8788 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8789 : if (pg_fsync_writethrough(fd) != 0)
8790 : msg = _("could not fsync write-through file \"%s\": %m");
8791 : break;
8792 : #endif
8793 0 : case WAL_SYNC_METHOD_FDATASYNC:
8794 0 : if (pg_fdatasync(fd) != 0)
8795 0 : msg = _("could not fdatasync file \"%s\": %m");
8796 0 : break;
8797 0 : case WAL_SYNC_METHOD_OPEN:
8798 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8799 : /* not reachable */
8800 : Assert(false);
8801 0 : break;
8802 0 : default:
8803 0 : ereport(PANIC,
8804 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8805 : errmsg_internal("unrecognized \"wal_sync_method\": %d", wal_sync_method));
8806 : break;
8807 : }
8808 :
8809 : /* PANIC if failed to fsync */
8810 0 : if (msg)
8811 : {
8812 : char xlogfname[MAXFNAMELEN];
8813 0 : int save_errno = errno;
8814 :
8815 0 : XLogFileName(xlogfname, tli, segno, wal_segment_size);
8816 0 : errno = save_errno;
8817 0 : ereport(PANIC,
8818 : (errcode_for_file_access(),
8819 : errmsg(msg, xlogfname)));
8820 : }
8821 :
8822 0 : pgstat_report_wait_end();
8823 :
8824 0 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL, IOOP_FSYNC,
8825 : start, 1, 0);
8826 : }
8827 :
8828 : /*
8829 : * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
8830 : * function. It creates the necessary starting checkpoint and constructs the
8831 : * backup state and tablespace map.
8832 : *
8833 : * Input parameters are "state" (the backup state), "fast" (if true, we do
8834 : * the checkpoint in fast mode), and "tablespaces" (if non-NULL, indicates a
8835 : * list of tablespaceinfo structs describing the cluster's tablespaces.).
8836 : *
8837 : * The tablespace map contents are appended to passed-in parameter
8838 : * tablespace_map and the caller is responsible for including it in the backup
8839 : * archive as 'tablespace_map'. The tablespace_map file is required mainly for
8840 : * tar format in windows as native windows utilities are not able to create
8841 : * symlinks while extracting files from tar. However for consistency and
8842 : * platform-independence, we do it the same way everywhere.
8843 : *
8844 : * It fills in "state" with the information required for the backup, such
8845 : * as the minimum WAL location that must be present to restore from this
8846 : * backup (starttli) and the corresponding timeline ID (starttli).
8847 : *
8848 : * Every successfully started backup must be stopped by calling
8849 : * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
8850 : * backups active at the same time.
8851 : *
8852 : * It is the responsibility of the caller of this function to verify the
8853 : * permissions of the calling user!
8854 : */
8855 : void
8856 328 : do_pg_backup_start(const char *backupidstr, bool fast, List **tablespaces,
8857 : BackupState *state, StringInfo tblspcmapfile)
8858 : {
8859 : bool backup_started_in_recovery;
8860 :
8861 : Assert(state != NULL);
8862 328 : backup_started_in_recovery = RecoveryInProgress();
8863 :
8864 : /*
8865 : * During recovery, we don't need to check WAL level. Because, if WAL
8866 : * level is not sufficient, it's impossible to get here during recovery.
8867 : */
8868 328 : if (!backup_started_in_recovery && !XLogIsNeeded())
8869 0 : ereport(ERROR,
8870 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8871 : errmsg("WAL level not sufficient for making an online backup"),
8872 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
8873 :
8874 328 : if (strlen(backupidstr) > MAXPGPATH)
8875 2 : ereport(ERROR,
8876 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8877 : errmsg("backup label too long (max %d bytes)",
8878 : MAXPGPATH)));
8879 :
8880 326 : strlcpy(state->name, backupidstr, sizeof(state->name));
8881 :
8882 : /*
8883 : * Mark backup active in shared memory. We must do full-page WAL writes
8884 : * during an on-line backup even if not doing so at other times, because
8885 : * it's quite possible for the backup dump to obtain a "torn" (partially
8886 : * written) copy of a database page if it reads the page concurrently with
8887 : * our write to the same page. This can be fixed as long as the first
8888 : * write to the page in the WAL sequence is a full-page write. Hence, we
8889 : * increment runningBackups then force a CHECKPOINT, to ensure there are
8890 : * no dirty pages in shared memory that might get dumped while the backup
8891 : * is in progress without having a corresponding WAL record. (Once the
8892 : * backup is complete, we need not force full-page writes anymore, since
8893 : * we expect that any pages not modified during the backup interval must
8894 : * have been correctly captured by the backup.)
8895 : *
8896 : * Note that forcing full-page writes has no effect during an online
8897 : * backup from the standby.
8898 : *
8899 : * We must hold all the insertion locks to change the value of
8900 : * runningBackups, to ensure adequate interlocking against
8901 : * XLogInsertRecord().
8902 : */
8903 326 : WALInsertLockAcquireExclusive();
8904 326 : XLogCtl->Insert.runningBackups++;
8905 326 : WALInsertLockRelease();
8906 :
8907 : /*
8908 : * Ensure we decrement runningBackups if we fail below. NB -- for this to
8909 : * work correctly, it is critical that sessionBackupState is only updated
8910 : * after this block is over.
8911 : */
8912 326 : PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
8913 : {
8914 326 : bool gotUniqueStartpoint = false;
8915 : DIR *tblspcdir;
8916 : struct dirent *de;
8917 : tablespaceinfo *ti;
8918 : int datadirpathlen;
8919 :
8920 : /*
8921 : * Force an XLOG file switch before the checkpoint, to ensure that the
8922 : * WAL segment the checkpoint is written to doesn't contain pages with
8923 : * old timeline IDs. That would otherwise happen if you called
8924 : * pg_backup_start() right after restoring from a PITR archive: the
8925 : * first WAL segment containing the startup checkpoint has pages in
8926 : * the beginning with the old timeline ID. That can cause trouble at
8927 : * recovery: we won't have a history file covering the old timeline if
8928 : * pg_wal directory was not included in the base backup and the WAL
8929 : * archive was cleared too before starting the backup.
8930 : *
8931 : * This also ensures that we have emitted a WAL page header that has
8932 : * XLP_BKP_REMOVABLE off before we emit the checkpoint record.
8933 : * Therefore, if a WAL archiver (such as pglesslog) is trying to
8934 : * compress out removable backup blocks, it won't remove any that
8935 : * occur after this point.
8936 : *
8937 : * During recovery, we skip forcing XLOG file switch, which means that
8938 : * the backup taken during recovery is not available for the special
8939 : * recovery case described above.
8940 : */
8941 326 : if (!backup_started_in_recovery)
8942 312 : RequestXLogSwitch(false);
8943 :
8944 : do
8945 : {
8946 : bool checkpointfpw;
8947 :
8948 : /*
8949 : * Force a CHECKPOINT. Aside from being necessary to prevent torn
8950 : * page problems, this guarantees that two successive backup runs
8951 : * will have different checkpoint positions and hence different
8952 : * history file names, even if nothing happened in between.
8953 : *
8954 : * During recovery, establish a restartpoint if possible. We use
8955 : * the last restartpoint as the backup starting checkpoint. This
8956 : * means that two successive backup runs can have same checkpoint
8957 : * positions.
8958 : *
8959 : * Since the fact that we are executing do_pg_backup_start()
8960 : * during recovery means that checkpointer is running, we can use
8961 : * RequestCheckpoint() to establish a restartpoint.
8962 : *
8963 : * We use CHECKPOINT_FAST only if requested by user (via passing
8964 : * fast = true). Otherwise this can take awhile.
8965 : */
8966 326 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
8967 : (fast ? CHECKPOINT_FAST : 0));
8968 :
8969 : /*
8970 : * Now we need to fetch the checkpoint record location, and also
8971 : * its REDO pointer. The oldest point in WAL that would be needed
8972 : * to restore starting from the checkpoint is precisely the REDO
8973 : * pointer.
8974 : */
8975 326 : LWLockAcquire(ControlFileLock, LW_SHARED);
8976 326 : state->checkpointloc = ControlFile->checkPoint;
8977 326 : state->startpoint = ControlFile->checkPointCopy.redo;
8978 326 : state->starttli = ControlFile->checkPointCopy.ThisTimeLineID;
8979 326 : checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
8980 326 : LWLockRelease(ControlFileLock);
8981 :
8982 326 : if (backup_started_in_recovery)
8983 : {
8984 : XLogRecPtr recptr;
8985 :
8986 : /*
8987 : * Check to see if all WAL replayed during online backup
8988 : * (i.e., since last restartpoint used as backup starting
8989 : * checkpoint) contain full-page writes.
8990 : */
8991 14 : SpinLockAcquire(&XLogCtl->info_lck);
8992 14 : recptr = XLogCtl->lastFpwDisableRecPtr;
8993 14 : SpinLockRelease(&XLogCtl->info_lck);
8994 :
8995 14 : if (!checkpointfpw || state->startpoint <= recptr)
8996 0 : ereport(ERROR,
8997 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8998 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
8999 : "since last restartpoint"),
9000 : errhint("This means that the backup being taken on the standby "
9001 : "is corrupt and should not be used. "
9002 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9003 : "and then try an online backup again.")));
9004 :
9005 : /*
9006 : * During recovery, since we don't use the end-of-backup WAL
9007 : * record and don't write the backup history file, the
9008 : * starting WAL location doesn't need to be unique. This means
9009 : * that two base backups started at the same time might use
9010 : * the same checkpoint as starting locations.
9011 : */
9012 14 : gotUniqueStartpoint = true;
9013 : }
9014 :
9015 : /*
9016 : * If two base backups are started at the same time (in WAL sender
9017 : * processes), we need to make sure that they use different
9018 : * checkpoints as starting locations, because we use the starting
9019 : * WAL location as a unique identifier for the base backup in the
9020 : * end-of-backup WAL record and when we write the backup history
9021 : * file. Perhaps it would be better generate a separate unique ID
9022 : * for each backup instead of forcing another checkpoint, but
9023 : * taking a checkpoint right after another is not that expensive
9024 : * either because only few buffers have been dirtied yet.
9025 : */
9026 326 : WALInsertLockAcquireExclusive();
9027 326 : if (XLogCtl->Insert.lastBackupStart < state->startpoint)
9028 : {
9029 326 : XLogCtl->Insert.lastBackupStart = state->startpoint;
9030 326 : gotUniqueStartpoint = true;
9031 : }
9032 326 : WALInsertLockRelease();
9033 326 : } while (!gotUniqueStartpoint);
9034 :
9035 : /*
9036 : * Construct tablespace_map file.
9037 : */
9038 326 : datadirpathlen = strlen(DataDir);
9039 :
9040 : /* Collect information about all tablespaces */
9041 326 : tblspcdir = AllocateDir(PG_TBLSPC_DIR);
9042 1050 : while ((de = ReadDir(tblspcdir, PG_TBLSPC_DIR)) != NULL)
9043 : {
9044 : char fullpath[MAXPGPATH + sizeof(PG_TBLSPC_DIR)];
9045 : char linkpath[MAXPGPATH];
9046 724 : char *relpath = NULL;
9047 : char *s;
9048 : PGFileType de_type;
9049 : char *badp;
9050 : Oid tsoid;
9051 :
9052 : /*
9053 : * Try to parse the directory name as an unsigned integer.
9054 : *
9055 : * Tablespace directories should be positive integers that can be
9056 : * represented in 32 bits, with no leading zeroes or trailing
9057 : * garbage. If we come across a name that doesn't meet those
9058 : * criteria, skip it.
9059 : */
9060 724 : if (de->d_name[0] < '1' || de->d_name[1] > '9')
9061 652 : continue;
9062 72 : errno = 0;
9063 72 : tsoid = strtoul(de->d_name, &badp, 10);
9064 72 : if (*badp != '\0' || errno == EINVAL || errno == ERANGE)
9065 0 : continue;
9066 :
9067 72 : snprintf(fullpath, sizeof(fullpath), "%s/%s", PG_TBLSPC_DIR, de->d_name);
9068 :
9069 72 : de_type = get_dirent_type(fullpath, de, false, ERROR);
9070 :
9071 72 : if (de_type == PGFILETYPE_LNK)
9072 : {
9073 : StringInfoData escapedpath;
9074 : int rllen;
9075 :
9076 44 : rllen = readlink(fullpath, linkpath, sizeof(linkpath));
9077 44 : if (rllen < 0)
9078 : {
9079 0 : ereport(WARNING,
9080 : (errmsg("could not read symbolic link \"%s\": %m",
9081 : fullpath)));
9082 0 : continue;
9083 : }
9084 44 : else if (rllen >= sizeof(linkpath))
9085 : {
9086 0 : ereport(WARNING,
9087 : (errmsg("symbolic link \"%s\" target is too long",
9088 : fullpath)));
9089 0 : continue;
9090 : }
9091 44 : linkpath[rllen] = '\0';
9092 :
9093 : /*
9094 : * Relpath holds the relative path of the tablespace directory
9095 : * when it's located within PGDATA, or NULL if it's located
9096 : * elsewhere.
9097 : */
9098 44 : if (rllen > datadirpathlen &&
9099 2 : strncmp(linkpath, DataDir, datadirpathlen) == 0 &&
9100 0 : IS_DIR_SEP(linkpath[datadirpathlen]))
9101 0 : relpath = pstrdup(linkpath + datadirpathlen + 1);
9102 :
9103 : /*
9104 : * Add a backslash-escaped version of the link path to the
9105 : * tablespace map file.
9106 : */
9107 44 : initStringInfo(&escapedpath);
9108 1092 : for (s = linkpath; *s; s++)
9109 : {
9110 1048 : if (*s == '\n' || *s == '\r' || *s == '\\')
9111 0 : appendStringInfoChar(&escapedpath, '\\');
9112 1048 : appendStringInfoChar(&escapedpath, *s);
9113 : }
9114 44 : appendStringInfo(tblspcmapfile, "%s %s\n",
9115 44 : de->d_name, escapedpath.data);
9116 44 : pfree(escapedpath.data);
9117 : }
9118 28 : else if (de_type == PGFILETYPE_DIR)
9119 : {
9120 : /*
9121 : * It's possible to use allow_in_place_tablespaces to create
9122 : * directories directly under pg_tblspc, for testing purposes
9123 : * only.
9124 : *
9125 : * In this case, we store a relative path rather than an
9126 : * absolute path into the tablespaceinfo.
9127 : */
9128 28 : snprintf(linkpath, sizeof(linkpath), "%s/%s",
9129 28 : PG_TBLSPC_DIR, de->d_name);
9130 28 : relpath = pstrdup(linkpath);
9131 : }
9132 : else
9133 : {
9134 : /* Skip any other file type that appears here. */
9135 0 : continue;
9136 : }
9137 :
9138 72 : ti = palloc(sizeof(tablespaceinfo));
9139 72 : ti->oid = tsoid;
9140 72 : ti->path = pstrdup(linkpath);
9141 72 : ti->rpath = relpath;
9142 72 : ti->size = -1;
9143 :
9144 72 : if (tablespaces)
9145 72 : *tablespaces = lappend(*tablespaces, ti);
9146 : }
9147 326 : FreeDir(tblspcdir);
9148 :
9149 326 : state->starttime = (pg_time_t) time(NULL);
9150 : }
9151 326 : PG_END_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9152 :
9153 326 : state->started_in_recovery = backup_started_in_recovery;
9154 :
9155 : /*
9156 : * Mark that the start phase has correctly finished for the backup.
9157 : */
9158 326 : sessionBackupState = SESSION_BACKUP_RUNNING;
9159 326 : }
9160 :
9161 : /*
9162 : * Utility routine to fetch the session-level status of a backup running.
9163 : */
9164 : SessionBackupState
9165 368 : get_backup_status(void)
9166 : {
9167 368 : return sessionBackupState;
9168 : }
9169 :
9170 : /*
9171 : * do_pg_backup_stop
9172 : *
9173 : * Utility function called at the end of an online backup. It creates history
9174 : * file (if required), resets sessionBackupState and so on. It can optionally
9175 : * wait for WAL segments to be archived.
9176 : *
9177 : * "state" is filled with the information necessary to restore from this
9178 : * backup with its stop LSN (stoppoint), its timeline ID (stoptli), etc.
9179 : *
9180 : * It is the responsibility of the caller of this function to verify the
9181 : * permissions of the calling user!
9182 : */
9183 : void
9184 314 : do_pg_backup_stop(BackupState *state, bool waitforarchive)
9185 : {
9186 314 : bool backup_stopped_in_recovery = false;
9187 : char histfilepath[MAXPGPATH];
9188 : char lastxlogfilename[MAXFNAMELEN];
9189 : char histfilename[MAXFNAMELEN];
9190 : XLogSegNo _logSegNo;
9191 : FILE *fp;
9192 : int seconds_before_warning;
9193 314 : int waits = 0;
9194 314 : bool reported_waiting = false;
9195 :
9196 : Assert(state != NULL);
9197 :
9198 314 : backup_stopped_in_recovery = RecoveryInProgress();
9199 :
9200 : /*
9201 : * During recovery, we don't need to check WAL level. Because, if WAL
9202 : * level is not sufficient, it's impossible to get here during recovery.
9203 : */
9204 314 : if (!backup_stopped_in_recovery && !XLogIsNeeded())
9205 0 : ereport(ERROR,
9206 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9207 : errmsg("WAL level not sufficient for making an online backup"),
9208 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9209 :
9210 : /*
9211 : * OK to update backup counter and session-level lock.
9212 : *
9213 : * Note that CHECK_FOR_INTERRUPTS() must not occur while updating them,
9214 : * otherwise they can be updated inconsistently, which might cause
9215 : * do_pg_abort_backup() to fail.
9216 : */
9217 314 : WALInsertLockAcquireExclusive();
9218 :
9219 : /*
9220 : * It is expected that each do_pg_backup_start() call is matched by
9221 : * exactly one do_pg_backup_stop() call.
9222 : */
9223 : Assert(XLogCtl->Insert.runningBackups > 0);
9224 314 : XLogCtl->Insert.runningBackups--;
9225 :
9226 : /*
9227 : * Clean up session-level lock.
9228 : *
9229 : * You might think that WALInsertLockRelease() can be called before
9230 : * cleaning up session-level lock because session-level lock doesn't need
9231 : * to be protected with WAL insertion lock. But since
9232 : * CHECK_FOR_INTERRUPTS() can occur in it, session-level lock must be
9233 : * cleaned up before it.
9234 : */
9235 314 : sessionBackupState = SESSION_BACKUP_NONE;
9236 :
9237 314 : WALInsertLockRelease();
9238 :
9239 : /*
9240 : * If we are taking an online backup from the standby, we confirm that the
9241 : * standby has not been promoted during the backup.
9242 : */
9243 314 : if (state->started_in_recovery && !backup_stopped_in_recovery)
9244 0 : ereport(ERROR,
9245 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9246 : errmsg("the standby was promoted during online backup"),
9247 : errhint("This means that the backup being taken is corrupt "
9248 : "and should not be used. "
9249 : "Try taking another online backup.")));
9250 :
9251 : /*
9252 : * During recovery, we don't write an end-of-backup record. We assume that
9253 : * pg_control was backed up last and its minimum recovery point can be
9254 : * available as the backup end location. Since we don't have an
9255 : * end-of-backup record, we use the pg_control value to check whether
9256 : * we've reached the end of backup when starting recovery from this
9257 : * backup. We have no way of checking if pg_control wasn't backed up last
9258 : * however.
9259 : *
9260 : * We don't force a switch to new WAL file but it is still possible to
9261 : * wait for all the required files to be archived if waitforarchive is
9262 : * true. This is okay if we use the backup to start a standby and fetch
9263 : * the missing WAL using streaming replication. But in the case of an
9264 : * archive recovery, a user should set waitforarchive to true and wait for
9265 : * them to be archived to ensure that all the required files are
9266 : * available.
9267 : *
9268 : * We return the current minimum recovery point as the backup end
9269 : * location. Note that it can be greater than the exact backup end
9270 : * location if the minimum recovery point is updated after the backup of
9271 : * pg_control. This is harmless for current uses.
9272 : *
9273 : * XXX currently a backup history file is for informational and debug
9274 : * purposes only. It's not essential for an online backup. Furthermore,
9275 : * even if it's created, it will not be archived during recovery because
9276 : * an archiver is not invoked. So it doesn't seem worthwhile to write a
9277 : * backup history file during recovery.
9278 : */
9279 314 : if (backup_stopped_in_recovery)
9280 : {
9281 : XLogRecPtr recptr;
9282 :
9283 : /*
9284 : * Check to see if all WAL replayed during online backup contain
9285 : * full-page writes.
9286 : */
9287 14 : SpinLockAcquire(&XLogCtl->info_lck);
9288 14 : recptr = XLogCtl->lastFpwDisableRecPtr;
9289 14 : SpinLockRelease(&XLogCtl->info_lck);
9290 :
9291 14 : if (state->startpoint <= recptr)
9292 0 : ereport(ERROR,
9293 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9294 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9295 : "during online backup"),
9296 : errhint("This means that the backup being taken on the standby "
9297 : "is corrupt and should not be used. "
9298 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9299 : "and then try an online backup again.")));
9300 :
9301 :
9302 14 : LWLockAcquire(ControlFileLock, LW_SHARED);
9303 14 : state->stoppoint = ControlFile->minRecoveryPoint;
9304 14 : state->stoptli = ControlFile->minRecoveryPointTLI;
9305 14 : LWLockRelease(ControlFileLock);
9306 : }
9307 : else
9308 : {
9309 : char *history_file;
9310 :
9311 : /*
9312 : * Write the backup-end xlog record
9313 : */
9314 300 : XLogBeginInsert();
9315 300 : XLogRegisterData(&state->startpoint,
9316 : sizeof(state->startpoint));
9317 300 : state->stoppoint = XLogInsert(RM_XLOG_ID, XLOG_BACKUP_END);
9318 :
9319 : /*
9320 : * Given that we're not in recovery, InsertTimeLineID is set and can't
9321 : * change, so we can read it without a lock.
9322 : */
9323 300 : state->stoptli = XLogCtl->InsertTimeLineID;
9324 :
9325 : /*
9326 : * Force a switch to a new xlog segment file, so that the backup is
9327 : * valid as soon as archiver moves out the current segment file.
9328 : */
9329 300 : RequestXLogSwitch(false);
9330 :
9331 300 : state->stoptime = (pg_time_t) time(NULL);
9332 :
9333 : /*
9334 : * Write the backup history file
9335 : */
9336 300 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9337 300 : BackupHistoryFilePath(histfilepath, state->stoptli, _logSegNo,
9338 : state->startpoint, wal_segment_size);
9339 300 : fp = AllocateFile(histfilepath, "w");
9340 300 : if (!fp)
9341 0 : ereport(ERROR,
9342 : (errcode_for_file_access(),
9343 : errmsg("could not create file \"%s\": %m",
9344 : histfilepath)));
9345 :
9346 : /* Build and save the contents of the backup history file */
9347 300 : history_file = build_backup_content(state, true);
9348 300 : fprintf(fp, "%s", history_file);
9349 300 : pfree(history_file);
9350 :
9351 300 : if (fflush(fp) || ferror(fp) || FreeFile(fp))
9352 0 : ereport(ERROR,
9353 : (errcode_for_file_access(),
9354 : errmsg("could not write file \"%s\": %m",
9355 : histfilepath)));
9356 :
9357 : /*
9358 : * Clean out any no-longer-needed history files. As a side effect,
9359 : * this will post a .ready file for the newly created history file,
9360 : * notifying the archiver that history file may be archived
9361 : * immediately.
9362 : */
9363 300 : CleanupBackupHistory();
9364 : }
9365 :
9366 : /*
9367 : * If archiving is enabled, wait for all the required WAL files to be
9368 : * archived before returning. If archiving isn't enabled, the required WAL
9369 : * needs to be transported via streaming replication (hopefully with
9370 : * wal_keep_size set high enough), or some more exotic mechanism like
9371 : * polling and copying files from pg_wal with script. We have no knowledge
9372 : * of those mechanisms, so it's up to the user to ensure that he gets all
9373 : * the required WAL.
9374 : *
9375 : * We wait until both the last WAL file filled during backup and the
9376 : * history file have been archived, and assume that the alphabetic sorting
9377 : * property of the WAL files ensures any earlier WAL files are safely
9378 : * archived as well.
9379 : *
9380 : * We wait forever, since archive_command is supposed to work and we
9381 : * assume the admin wanted his backup to work completely. If you don't
9382 : * wish to wait, then either waitforarchive should be passed in as false,
9383 : * or you can set statement_timeout. Also, some notices are issued to
9384 : * clue in anyone who might be doing this interactively.
9385 : */
9386 :
9387 314 : if (waitforarchive &&
9388 20 : ((!backup_stopped_in_recovery && XLogArchivingActive()) ||
9389 2 : (backup_stopped_in_recovery && XLogArchivingAlways())))
9390 : {
9391 8 : XLByteToPrevSeg(state->stoppoint, _logSegNo, wal_segment_size);
9392 8 : XLogFileName(lastxlogfilename, state->stoptli, _logSegNo,
9393 : wal_segment_size);
9394 :
9395 8 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9396 8 : BackupHistoryFileName(histfilename, state->stoptli, _logSegNo,
9397 : state->startpoint, wal_segment_size);
9398 :
9399 8 : seconds_before_warning = 60;
9400 8 : waits = 0;
9401 :
9402 24 : while (XLogArchiveIsBusy(lastxlogfilename) ||
9403 8 : XLogArchiveIsBusy(histfilename))
9404 : {
9405 8 : CHECK_FOR_INTERRUPTS();
9406 :
9407 8 : if (!reported_waiting && waits > 5)
9408 : {
9409 0 : ereport(NOTICE,
9410 : (errmsg("base backup done, waiting for required WAL segments to be archived")));
9411 0 : reported_waiting = true;
9412 : }
9413 :
9414 8 : (void) WaitLatch(MyLatch,
9415 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
9416 : 1000L,
9417 : WAIT_EVENT_BACKUP_WAIT_WAL_ARCHIVE);
9418 8 : ResetLatch(MyLatch);
9419 :
9420 8 : if (++waits >= seconds_before_warning)
9421 : {
9422 0 : seconds_before_warning *= 2; /* This wraps in >10 years... */
9423 0 : ereport(WARNING,
9424 : (errmsg("still waiting for all required WAL segments to be archived (%d seconds elapsed)",
9425 : waits),
9426 : errhint("Check that your \"archive_command\" is executing properly. "
9427 : "You can safely cancel this backup, "
9428 : "but the database backup will not be usable without all the WAL segments.")));
9429 : }
9430 : }
9431 :
9432 8 : ereport(NOTICE,
9433 : (errmsg("all required WAL segments have been archived")));
9434 : }
9435 306 : else if (waitforarchive)
9436 12 : ereport(NOTICE,
9437 : (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
9438 314 : }
9439 :
9440 :
9441 : /*
9442 : * do_pg_abort_backup: abort a running backup
9443 : *
9444 : * This does just the most basic steps of do_pg_backup_stop(), by taking the
9445 : * system out of backup mode, thus making it a lot more safe to call from
9446 : * an error handler.
9447 : *
9448 : * 'arg' indicates that it's being called during backup setup; so
9449 : * sessionBackupState has not been modified yet, but runningBackups has
9450 : * already been incremented. When it's false, then it's invoked as a
9451 : * before_shmem_exit handler, and therefore we must not change state
9452 : * unless sessionBackupState indicates that a backup is actually running.
9453 : *
9454 : * NB: This gets used as a PG_ENSURE_ERROR_CLEANUP callback and
9455 : * before_shmem_exit handler, hence the odd-looking signature.
9456 : */
9457 : void
9458 16 : do_pg_abort_backup(int code, Datum arg)
9459 : {
9460 16 : bool during_backup_start = DatumGetBool(arg);
9461 :
9462 : /* If called during backup start, there shouldn't be one already running */
9463 : Assert(!during_backup_start || sessionBackupState == SESSION_BACKUP_NONE);
9464 :
9465 16 : if (during_backup_start || sessionBackupState != SESSION_BACKUP_NONE)
9466 : {
9467 12 : WALInsertLockAcquireExclusive();
9468 : Assert(XLogCtl->Insert.runningBackups > 0);
9469 12 : XLogCtl->Insert.runningBackups--;
9470 :
9471 12 : sessionBackupState = SESSION_BACKUP_NONE;
9472 12 : WALInsertLockRelease();
9473 :
9474 12 : if (!during_backup_start)
9475 12 : ereport(WARNING,
9476 : errmsg("aborting backup due to backend exiting before pg_backup_stop was called"));
9477 : }
9478 16 : }
9479 :
9480 : /*
9481 : * Register a handler that will warn about unterminated backups at end of
9482 : * session, unless this has already been done.
9483 : */
9484 : void
9485 8 : register_persistent_abort_backup_handler(void)
9486 : {
9487 : static bool already_done = false;
9488 :
9489 8 : if (already_done)
9490 2 : return;
9491 6 : before_shmem_exit(do_pg_abort_backup, BoolGetDatum(false));
9492 6 : already_done = true;
9493 : }
9494 :
9495 : /*
9496 : * Get latest WAL insert pointer
9497 : */
9498 : XLogRecPtr
9499 4004 : GetXLogInsertRecPtr(void)
9500 : {
9501 4004 : XLogCtlInsert *Insert = &XLogCtl->Insert;
9502 : uint64 current_bytepos;
9503 :
9504 4004 : SpinLockAcquire(&Insert->insertpos_lck);
9505 4004 : current_bytepos = Insert->CurrBytePos;
9506 4004 : SpinLockRelease(&Insert->insertpos_lck);
9507 :
9508 4004 : return XLogBytePosToRecPtr(current_bytepos);
9509 : }
9510 :
9511 : /*
9512 : * Get latest WAL write pointer
9513 : */
9514 : XLogRecPtr
9515 9426 : GetXLogWriteRecPtr(void)
9516 : {
9517 9426 : RefreshXLogWriteResult(LogwrtResult);
9518 :
9519 9426 : return LogwrtResult.Write;
9520 : }
9521 :
9522 : /*
9523 : * Returns the redo pointer of the last checkpoint or restartpoint. This is
9524 : * the oldest point in WAL that we still need, if we have to restart recovery.
9525 : */
9526 : void
9527 784 : GetOldestRestartPoint(XLogRecPtr *oldrecptr, TimeLineID *oldtli)
9528 : {
9529 784 : LWLockAcquire(ControlFileLock, LW_SHARED);
9530 784 : *oldrecptr = ControlFile->checkPointCopy.redo;
9531 784 : *oldtli = ControlFile->checkPointCopy.ThisTimeLineID;
9532 784 : LWLockRelease(ControlFileLock);
9533 784 : }
9534 :
9535 : /* Thin wrapper around ShutdownWalRcv(). */
9536 : void
9537 1904 : XLogShutdownWalRcv(void)
9538 : {
9539 : Assert(AmStartupProcess() || !IsUnderPostmaster);
9540 :
9541 1904 : ShutdownWalRcv();
9542 1904 : ResetInstallXLogFileSegmentActive();
9543 1904 : }
9544 :
9545 : /* Enable WAL file recycling and preallocation. */
9546 : void
9547 2272 : SetInstallXLogFileSegmentActive(void)
9548 : {
9549 2272 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9550 2272 : XLogCtl->InstallXLogFileSegmentActive = true;
9551 2272 : LWLockRelease(ControlFileLock);
9552 2272 : }
9553 :
9554 : /* Disable WAL file recycling and preallocation. */
9555 : void
9556 2164 : ResetInstallXLogFileSegmentActive(void)
9557 : {
9558 2164 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9559 2164 : XLogCtl->InstallXLogFileSegmentActive = false;
9560 2164 : LWLockRelease(ControlFileLock);
9561 2164 : }
9562 :
9563 : bool
9564 0 : IsInstallXLogFileSegmentActive(void)
9565 : {
9566 : bool result;
9567 :
9568 0 : LWLockAcquire(ControlFileLock, LW_SHARED);
9569 0 : result = XLogCtl->InstallXLogFileSegmentActive;
9570 0 : LWLockRelease(ControlFileLock);
9571 :
9572 0 : return result;
9573 : }
9574 :
9575 : /*
9576 : * Update the WalWriterSleeping flag.
9577 : */
9578 : void
9579 1010 : SetWalWriterSleeping(bool sleeping)
9580 : {
9581 1010 : SpinLockAcquire(&XLogCtl->info_lck);
9582 1010 : XLogCtl->WalWriterSleeping = sleeping;
9583 1010 : SpinLockRelease(&XLogCtl->info_lck);
9584 1010 : }
|
