Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * xlog.c
4 : * PostgreSQL write-ahead log manager
5 : *
6 : * The Write-Ahead Log (WAL) functionality is split into several source
7 : * files, in addition to this one:
8 : *
9 : * xloginsert.c - Functions for constructing WAL records
10 : * xlogrecovery.c - WAL recovery and standby code
11 : * xlogreader.c - Facility for reading WAL files and parsing WAL records
12 : * xlogutils.c - Helper functions for WAL redo routines
13 : *
14 : * This file contains functions for coordinating database startup and
15 : * checkpointing, and managing the write-ahead log buffers when the
16 : * system is running.
17 : *
18 : * StartupXLOG() is the main entry point of the startup process. It
19 : * coordinates database startup, performing WAL recovery, and the
20 : * transition from WAL recovery into normal operations.
21 : *
22 : * XLogInsertRecord() inserts a WAL record into the WAL buffers. Most
23 : * callers should not call this directly, but use the functions in
24 : * xloginsert.c to construct the WAL record. XLogFlush() can be used
25 : * to force the WAL to disk.
26 : *
27 : * In addition to those, there are many other functions for interrogating
28 : * the current system state, and for starting/stopping backups.
29 : *
30 : *
31 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
32 : * Portions Copyright (c) 1994, Regents of the University of California
33 : *
34 : * src/backend/access/transam/xlog.c
35 : *
36 : *-------------------------------------------------------------------------
37 : */
38 :
39 : #include "postgres.h"
40 :
41 : #include <ctype.h>
42 : #include <math.h>
43 : #include <time.h>
44 : #include <fcntl.h>
45 : #include <sys/stat.h>
46 : #include <sys/time.h>
47 : #include <unistd.h>
48 :
49 : #include "access/clog.h"
50 : #include "access/commit_ts.h"
51 : #include "access/heaptoast.h"
52 : #include "access/multixact.h"
53 : #include "access/rewriteheap.h"
54 : #include "access/subtrans.h"
55 : #include "access/timeline.h"
56 : #include "access/transam.h"
57 : #include "access/twophase.h"
58 : #include "access/xact.h"
59 : #include "access/xlog_internal.h"
60 : #include "access/xlogarchive.h"
61 : #include "access/xloginsert.h"
62 : #include "access/xlogreader.h"
63 : #include "access/xlogrecovery.h"
64 : #include "access/xlogutils.h"
65 : #include "backup/basebackup.h"
66 : #include "catalog/catversion.h"
67 : #include "catalog/pg_control.h"
68 : #include "catalog/pg_database.h"
69 : #include "common/controldata_utils.h"
70 : #include "common/file_utils.h"
71 : #include "executor/instrument.h"
72 : #include "miscadmin.h"
73 : #include "pg_trace.h"
74 : #include "pgstat.h"
75 : #include "port/atomics.h"
76 : #include "postmaster/bgwriter.h"
77 : #include "postmaster/startup.h"
78 : #include "postmaster/walsummarizer.h"
79 : #include "postmaster/walwriter.h"
80 : #include "replication/origin.h"
81 : #include "replication/slot.h"
82 : #include "replication/snapbuild.h"
83 : #include "replication/walreceiver.h"
84 : #include "replication/walsender.h"
85 : #include "storage/bufmgr.h"
86 : #include "storage/fd.h"
87 : #include "storage/ipc.h"
88 : #include "storage/large_object.h"
89 : #include "storage/latch.h"
90 : #include "storage/predicate.h"
91 : #include "storage/proc.h"
92 : #include "storage/procarray.h"
93 : #include "storage/reinit.h"
94 : #include "storage/spin.h"
95 : #include "storage/sync.h"
96 : #include "utils/guc_hooks.h"
97 : #include "utils/guc_tables.h"
98 : #include "utils/injection_point.h"
99 : #include "utils/ps_status.h"
100 : #include "utils/relmapper.h"
101 : #include "utils/snapmgr.h"
102 : #include "utils/timeout.h"
103 : #include "utils/timestamp.h"
104 : #include "utils/varlena.h"
105 :
106 : #ifdef WAL_DEBUG
107 : #include "utils/memutils.h"
108 : #endif
109 :
110 : /* timeline ID to be used when bootstrapping */
111 : #define BootstrapTimeLineID 1
112 :
113 : /* User-settable parameters */
114 : int max_wal_size_mb = 1024; /* 1 GB */
115 : int min_wal_size_mb = 80; /* 80 MB */
116 : int wal_keep_size_mb = 0;
117 : int XLOGbuffers = -1;
118 : int XLogArchiveTimeout = 0;
119 : int XLogArchiveMode = ARCHIVE_MODE_OFF;
120 : char *XLogArchiveCommand = NULL;
121 : bool EnableHotStandby = false;
122 : bool fullPageWrites = true;
123 : bool wal_log_hints = false;
124 : int wal_compression = WAL_COMPRESSION_NONE;
125 : char *wal_consistency_checking_string = NULL;
126 : bool *wal_consistency_checking = NULL;
127 : bool wal_init_zero = true;
128 : bool wal_recycle = true;
129 : bool log_checkpoints = true;
130 : int wal_sync_method = DEFAULT_WAL_SYNC_METHOD;
131 : int wal_level = WAL_LEVEL_REPLICA;
132 : int CommitDelay = 0; /* precommit delay in microseconds */
133 : int CommitSiblings = 5; /* # concurrent xacts needed to sleep */
134 : int wal_retrieve_retry_interval = 5000;
135 : int max_slot_wal_keep_size_mb = -1;
136 : int wal_decode_buffer_size = 512 * 1024;
137 : bool track_wal_io_timing = false;
138 :
139 : #ifdef WAL_DEBUG
140 : bool XLOG_DEBUG = false;
141 : #endif
142 :
143 : int wal_segment_size = DEFAULT_XLOG_SEG_SIZE;
144 :
145 : /*
146 : * Number of WAL insertion locks to use. A higher value allows more insertions
147 : * to happen concurrently, but adds some CPU overhead to flushing the WAL,
148 : * which needs to iterate all the locks.
149 : */
150 : #define NUM_XLOGINSERT_LOCKS 8
151 :
152 : /*
153 : * Max distance from last checkpoint, before triggering a new xlog-based
154 : * checkpoint.
155 : */
156 : int CheckPointSegments;
157 :
158 : /* Estimated distance between checkpoints, in bytes */
159 : static double CheckPointDistanceEstimate = 0;
160 : static double PrevCheckPointDistance = 0;
161 :
162 : /*
163 : * Track whether there were any deferred checks for custom resource managers
164 : * specified in wal_consistency_checking.
165 : */
166 : static bool check_wal_consistency_checking_deferred = false;
167 :
168 : /*
169 : * GUC support
170 : */
171 : const struct config_enum_entry wal_sync_method_options[] = {
172 : {"fsync", WAL_SYNC_METHOD_FSYNC, false},
173 : #ifdef HAVE_FSYNC_WRITETHROUGH
174 : {"fsync_writethrough", WAL_SYNC_METHOD_FSYNC_WRITETHROUGH, false},
175 : #endif
176 : {"fdatasync", WAL_SYNC_METHOD_FDATASYNC, false},
177 : #ifdef O_SYNC
178 : {"open_sync", WAL_SYNC_METHOD_OPEN, false},
179 : #endif
180 : #ifdef O_DSYNC
181 : {"open_datasync", WAL_SYNC_METHOD_OPEN_DSYNC, false},
182 : #endif
183 : {NULL, 0, false}
184 : };
185 :
186 :
187 : /*
188 : * Although only "on", "off", and "always" are documented,
189 : * we accept all the likely variants of "on" and "off".
190 : */
191 : const struct config_enum_entry archive_mode_options[] = {
192 : {"always", ARCHIVE_MODE_ALWAYS, false},
193 : {"on", ARCHIVE_MODE_ON, false},
194 : {"off", ARCHIVE_MODE_OFF, false},
195 : {"true", ARCHIVE_MODE_ON, true},
196 : {"false", ARCHIVE_MODE_OFF, true},
197 : {"yes", ARCHIVE_MODE_ON, true},
198 : {"no", ARCHIVE_MODE_OFF, true},
199 : {"1", ARCHIVE_MODE_ON, true},
200 : {"0", ARCHIVE_MODE_OFF, true},
201 : {NULL, 0, false}
202 : };
203 :
204 : /*
205 : * Statistics for current checkpoint are collected in this global struct.
206 : * Because only the checkpointer or a stand-alone backend can perform
207 : * checkpoints, this will be unused in normal backends.
208 : */
209 : CheckpointStatsData CheckpointStats;
210 :
211 : /*
212 : * During recovery, lastFullPageWrites keeps track of full_page_writes that
213 : * the replayed WAL records indicate. It's initialized with full_page_writes
214 : * that the recovery starting checkpoint record indicates, and then updated
215 : * each time XLOG_FPW_CHANGE record is replayed.
216 : */
217 : static bool lastFullPageWrites;
218 :
219 : /*
220 : * Local copy of the state tracked by SharedRecoveryState in shared memory,
221 : * It is false if SharedRecoveryState is RECOVERY_STATE_DONE. True actually
222 : * means "not known, need to check the shared state".
223 : */
224 : static bool LocalRecoveryInProgress = true;
225 :
226 : /*
227 : * Local state for XLogInsertAllowed():
228 : * 1: unconditionally allowed to insert XLOG
229 : * 0: unconditionally not allowed to insert XLOG
230 : * -1: must check RecoveryInProgress(); disallow until it is false
231 : * Most processes start with -1 and transition to 1 after seeing that recovery
232 : * is not in progress. But we can also force the value for special cases.
233 : * The coding in XLogInsertAllowed() depends on the first two of these states
234 : * being numerically the same as bool true and false.
235 : */
236 : static int LocalXLogInsertAllowed = -1;
237 :
238 : /*
239 : * ProcLastRecPtr points to the start of the last XLOG record inserted by the
240 : * current backend. It is updated for all inserts. XactLastRecEnd points to
241 : * end+1 of the last record, and is reset when we end a top-level transaction,
242 : * or start a new one; so it can be used to tell if the current transaction has
243 : * created any XLOG records.
244 : *
245 : * While in parallel mode, this may not be fully up to date. When committing,
246 : * a transaction can assume this covers all xlog records written either by the
247 : * user backend or by any parallel worker which was present at any point during
248 : * the transaction. But when aborting, or when still in parallel mode, other
249 : * parallel backends may have written WAL records at later LSNs than the value
250 : * stored here. The parallel leader advances its own copy, when necessary,
251 : * in WaitForParallelWorkersToFinish.
252 : */
253 : XLogRecPtr ProcLastRecPtr = InvalidXLogRecPtr;
254 : XLogRecPtr XactLastRecEnd = InvalidXLogRecPtr;
255 : XLogRecPtr XactLastCommitEnd = InvalidXLogRecPtr;
256 :
257 : /*
258 : * RedoRecPtr is this backend's local copy of the REDO record pointer
259 : * (which is almost but not quite the same as a pointer to the most recent
260 : * CHECKPOINT record). We update this from the shared-memory copy,
261 : * XLogCtl->Insert.RedoRecPtr, whenever we can safely do so (ie, when we
262 : * hold an insertion lock). See XLogInsertRecord for details. We are also
263 : * allowed to update from XLogCtl->RedoRecPtr if we hold the info_lck;
264 : * see GetRedoRecPtr.
265 : *
266 : * NB: Code that uses this variable must be prepared not only for the
267 : * possibility that it may be arbitrarily out of date, but also for the
268 : * possibility that it might be set to InvalidXLogRecPtr. We used to
269 : * initialize it as a side effect of the first call to RecoveryInProgress(),
270 : * which meant that most code that might use it could assume that it had a
271 : * real if perhaps stale value. That's no longer the case.
272 : */
273 : static XLogRecPtr RedoRecPtr;
274 :
275 : /*
276 : * doPageWrites is this backend's local copy of (fullPageWrites ||
277 : * runningBackups > 0). It is used together with RedoRecPtr to decide whether
278 : * a full-page image of a page need to be taken.
279 : *
280 : * NB: Initially this is false, and there's no guarantee that it will be
281 : * initialized to any other value before it is first used. Any code that
282 : * makes use of it must recheck the value after obtaining a WALInsertLock,
283 : * and respond appropriately if it turns out that the previous value wasn't
284 : * accurate.
285 : */
286 : static bool doPageWrites;
287 :
288 : /*----------
289 : * Shared-memory data structures for XLOG control
290 : *
291 : * LogwrtRqst indicates a byte position that we need to write and/or fsync
292 : * the log up to (all records before that point must be written or fsynced).
293 : * The positions already written/fsynced are maintained in logWriteResult
294 : * and logFlushResult using atomic access.
295 : * In addition to the shared variable, each backend has a private copy of
296 : * both in LogwrtResult, which is updated when convenient.
297 : *
298 : * The request bookkeeping is simpler: there is a shared XLogCtl->LogwrtRqst
299 : * (protected by info_lck), but we don't need to cache any copies of it.
300 : *
301 : * info_lck is only held long enough to read/update the protected variables,
302 : * so it's a plain spinlock. The other locks are held longer (potentially
303 : * over I/O operations), so we use LWLocks for them. These locks are:
304 : *
305 : * WALBufMappingLock: must be held to replace a page in the WAL buffer cache.
306 : * It is only held while initializing and changing the mapping. If the
307 : * contents of the buffer being replaced haven't been written yet, the mapping
308 : * lock is released while the write is done, and reacquired afterwards.
309 : *
310 : * WALWriteLock: must be held to write WAL buffers to disk (XLogWrite or
311 : * XLogFlush).
312 : *
313 : * ControlFileLock: must be held to read/update control file or create
314 : * new log file.
315 : *
316 : *----------
317 : */
318 :
319 : typedef struct XLogwrtRqst
320 : {
321 : XLogRecPtr Write; /* last byte + 1 to write out */
322 : XLogRecPtr Flush; /* last byte + 1 to flush */
323 : } XLogwrtRqst;
324 :
325 : typedef struct XLogwrtResult
326 : {
327 : XLogRecPtr Write; /* last byte + 1 written out */
328 : XLogRecPtr Flush; /* last byte + 1 flushed */
329 : } XLogwrtResult;
330 :
331 : /*
332 : * Inserting to WAL is protected by a small fixed number of WAL insertion
333 : * locks. To insert to the WAL, you must hold one of the locks - it doesn't
334 : * matter which one. To lock out other concurrent insertions, you must hold
335 : * of them. Each WAL insertion lock consists of a lightweight lock, plus an
336 : * indicator of how far the insertion has progressed (insertingAt).
337 : *
338 : * The insertingAt values are read when a process wants to flush WAL from
339 : * the in-memory buffers to disk, to check that all the insertions to the
340 : * region the process is about to write out have finished. You could simply
341 : * wait for all currently in-progress insertions to finish, but the
342 : * insertingAt indicator allows you to ignore insertions to later in the WAL,
343 : * so that you only wait for the insertions that are modifying the buffers
344 : * you're about to write out.
345 : *
346 : * This isn't just an optimization. If all the WAL buffers are dirty, an
347 : * inserter that's holding a WAL insert lock might need to evict an old WAL
348 : * buffer, which requires flushing the WAL. If it's possible for an inserter
349 : * to block on another inserter unnecessarily, deadlock can arise when two
350 : * inserters holding a WAL insert lock wait for each other to finish their
351 : * insertion.
352 : *
353 : * Small WAL records that don't cross a page boundary never update the value,
354 : * the WAL record is just copied to the page and the lock is released. But
355 : * to avoid the deadlock-scenario explained above, the indicator is always
356 : * updated before sleeping while holding an insertion lock.
357 : *
358 : * lastImportantAt contains the LSN of the last important WAL record inserted
359 : * using a given lock. This value is used to detect if there has been
360 : * important WAL activity since the last time some action, like a checkpoint,
361 : * was performed - allowing to not repeat the action if not. The LSN is
362 : * updated for all insertions, unless the XLOG_MARK_UNIMPORTANT flag was
363 : * set. lastImportantAt is never cleared, only overwritten by the LSN of newer
364 : * records. Tracking the WAL activity directly in WALInsertLock has the
365 : * advantage of not needing any additional locks to update the value.
366 : */
367 : typedef struct
368 : {
369 : LWLock lock;
370 : pg_atomic_uint64 insertingAt;
371 : XLogRecPtr lastImportantAt;
372 : } WALInsertLock;
373 :
374 : /*
375 : * All the WAL insertion locks are allocated as an array in shared memory. We
376 : * force the array stride to be a power of 2, which saves a few cycles in
377 : * indexing, but more importantly also ensures that individual slots don't
378 : * cross cache line boundaries. (Of course, we have to also ensure that the
379 : * array start address is suitably aligned.)
380 : */
381 : typedef union WALInsertLockPadded
382 : {
383 : WALInsertLock l;
384 : char pad[PG_CACHE_LINE_SIZE];
385 : } WALInsertLockPadded;
386 :
387 : /*
388 : * Session status of running backup, used for sanity checks in SQL-callable
389 : * functions to start and stop backups.
390 : */
391 : static SessionBackupState sessionBackupState = SESSION_BACKUP_NONE;
392 :
393 : /*
394 : * Shared state data for WAL insertion.
395 : */
396 : typedef struct XLogCtlInsert
397 : {
398 : slock_t insertpos_lck; /* protects CurrBytePos and PrevBytePos */
399 :
400 : /*
401 : * CurrBytePos is the end of reserved WAL. The next record will be
402 : * inserted at that position. PrevBytePos is the start position of the
403 : * previously inserted (or rather, reserved) record - it is copied to the
404 : * prev-link of the next record. These are stored as "usable byte
405 : * positions" rather than XLogRecPtrs (see XLogBytePosToRecPtr()).
406 : */
407 : uint64 CurrBytePos;
408 : uint64 PrevBytePos;
409 :
410 : /*
411 : * Make sure the above heavily-contended spinlock and byte positions are
412 : * on their own cache line. In particular, the RedoRecPtr and full page
413 : * write variables below should be on a different cache line. They are
414 : * read on every WAL insertion, but updated rarely, and we don't want
415 : * those reads to steal the cache line containing Curr/PrevBytePos.
416 : */
417 : char pad[PG_CACHE_LINE_SIZE];
418 :
419 : /*
420 : * fullPageWrites is the authoritative value used by all backends to
421 : * determine whether to write full-page image to WAL. This shared value,
422 : * instead of the process-local fullPageWrites, is required because, when
423 : * full_page_writes is changed by SIGHUP, we must WAL-log it before it
424 : * actually affects WAL-logging by backends. Checkpointer sets at startup
425 : * or after SIGHUP.
426 : *
427 : * To read these fields, you must hold an insertion lock. To modify them,
428 : * you must hold ALL the locks.
429 : */
430 : XLogRecPtr RedoRecPtr; /* current redo point for insertions */
431 : bool fullPageWrites;
432 :
433 : /*
434 : * runningBackups is a counter indicating the number of backups currently
435 : * in progress. lastBackupStart is the latest checkpoint redo location
436 : * used as a starting point for an online backup.
437 : */
438 : int runningBackups;
439 : XLogRecPtr lastBackupStart;
440 :
441 : /*
442 : * WAL insertion locks.
443 : */
444 : WALInsertLockPadded *WALInsertLocks;
445 : } XLogCtlInsert;
446 :
447 : /*
448 : * Total shared-memory state for XLOG.
449 : */
450 : typedef struct XLogCtlData
451 : {
452 : XLogCtlInsert Insert;
453 :
454 : /* Protected by info_lck: */
455 : XLogwrtRqst LogwrtRqst;
456 : XLogRecPtr RedoRecPtr; /* a recent copy of Insert->RedoRecPtr */
457 : FullTransactionId ckptFullXid; /* nextXid of latest checkpoint */
458 : XLogRecPtr asyncXactLSN; /* LSN of newest async commit/abort */
459 : XLogRecPtr replicationSlotMinLSN; /* oldest LSN needed by any slot */
460 :
461 : XLogSegNo lastRemovedSegNo; /* latest removed/recycled XLOG segment */
462 :
463 : /* Fake LSN counter, for unlogged relations. */
464 : pg_atomic_uint64 unloggedLSN;
465 :
466 : /* Time and LSN of last xlog segment switch. Protected by WALWriteLock. */
467 : pg_time_t lastSegSwitchTime;
468 : XLogRecPtr lastSegSwitchLSN;
469 :
470 : /* These are accessed using atomics -- info_lck not needed */
471 : pg_atomic_uint64 logInsertResult; /* last byte + 1 inserted to buffers */
472 : pg_atomic_uint64 logWriteResult; /* last byte + 1 written out */
473 : pg_atomic_uint64 logFlushResult; /* last byte + 1 flushed */
474 :
475 : /*
476 : * Latest initialized page in the cache (last byte position + 1).
477 : *
478 : * To change the identity of a buffer (and InitializedUpTo), you need to
479 : * hold WALBufMappingLock. To change the identity of a buffer that's
480 : * still dirty, the old page needs to be written out first, and for that
481 : * you need WALWriteLock, and you need to ensure that there are no
482 : * in-progress insertions to the page by calling
483 : * WaitXLogInsertionsToFinish().
484 : */
485 : XLogRecPtr InitializedUpTo;
486 :
487 : /*
488 : * These values do not change after startup, although the pointed-to pages
489 : * and xlblocks values certainly do. xlblocks values are protected by
490 : * WALBufMappingLock.
491 : */
492 : char *pages; /* buffers for unwritten XLOG pages */
493 : pg_atomic_uint64 *xlblocks; /* 1st byte ptr-s + XLOG_BLCKSZ */
494 : int XLogCacheBlck; /* highest allocated xlog buffer index */
495 :
496 : /*
497 : * InsertTimeLineID is the timeline into which new WAL is being inserted
498 : * and flushed. It is zero during recovery, and does not change once set.
499 : *
500 : * If we create a new timeline when the system was started up,
501 : * PrevTimeLineID is the old timeline's ID that we forked off from.
502 : * Otherwise it's equal to InsertTimeLineID.
503 : *
504 : * We set these fields while holding info_lck. Most that reads these
505 : * values knows that recovery is no longer in progress and so can safely
506 : * read the value without a lock, but code that could be run either during
507 : * or after recovery can take info_lck while reading these values.
508 : */
509 : TimeLineID InsertTimeLineID;
510 : TimeLineID PrevTimeLineID;
511 :
512 : /*
513 : * SharedRecoveryState indicates if we're still in crash or archive
514 : * recovery. Protected by info_lck.
515 : */
516 : RecoveryState SharedRecoveryState;
517 :
518 : /*
519 : * InstallXLogFileSegmentActive indicates whether the checkpointer should
520 : * arrange for future segments by recycling and/or PreallocXlogFiles().
521 : * Protected by ControlFileLock. Only the startup process changes it. If
522 : * true, anyone can use InstallXLogFileSegment(). If false, the startup
523 : * process owns the exclusive right to install segments, by reading from
524 : * the archive and possibly replacing existing files.
525 : */
526 : bool InstallXLogFileSegmentActive;
527 :
528 : /*
529 : * WalWriterSleeping indicates whether the WAL writer is currently in
530 : * low-power mode (and hence should be nudged if an async commit occurs).
531 : * Protected by info_lck.
532 : */
533 : bool WalWriterSleeping;
534 :
535 : /*
536 : * During recovery, we keep a copy of the latest checkpoint record here.
537 : * lastCheckPointRecPtr points to start of checkpoint record and
538 : * lastCheckPointEndPtr points to end+1 of checkpoint record. Used by the
539 : * checkpointer when it wants to create a restartpoint.
540 : *
541 : * Protected by info_lck.
542 : */
543 : XLogRecPtr lastCheckPointRecPtr;
544 : XLogRecPtr lastCheckPointEndPtr;
545 : CheckPoint lastCheckPoint;
546 :
547 : /*
548 : * lastFpwDisableRecPtr points to the start of the last replayed
549 : * XLOG_FPW_CHANGE record that instructs full_page_writes is disabled.
550 : */
551 : XLogRecPtr lastFpwDisableRecPtr;
552 :
553 : slock_t info_lck; /* locks shared variables shown above */
554 : } XLogCtlData;
555 :
556 : /*
557 : * Classification of XLogInsertRecord operations.
558 : */
559 : typedef enum
560 : {
561 : WALINSERT_NORMAL,
562 : WALINSERT_SPECIAL_SWITCH,
563 : WALINSERT_SPECIAL_CHECKPOINT
564 : } WalInsertClass;
565 :
566 : static XLogCtlData *XLogCtl = NULL;
567 :
568 : /* a private copy of XLogCtl->Insert.WALInsertLocks, for convenience */
569 : static WALInsertLockPadded *WALInsertLocks = NULL;
570 :
571 : /*
572 : * We maintain an image of pg_control in shared memory.
573 : */
574 : static ControlFileData *ControlFile = NULL;
575 :
576 : /*
577 : * Calculate the amount of space left on the page after 'endptr'. Beware
578 : * multiple evaluation!
579 : */
580 : #define INSERT_FREESPACE(endptr) \
581 : (((endptr) % XLOG_BLCKSZ == 0) ? 0 : (XLOG_BLCKSZ - (endptr) % XLOG_BLCKSZ))
582 :
583 : /* Macro to advance to next buffer index. */
584 : #define NextBufIdx(idx) \
585 : (((idx) == XLogCtl->XLogCacheBlck) ? 0 : ((idx) + 1))
586 :
587 : /*
588 : * XLogRecPtrToBufIdx returns the index of the WAL buffer that holds, or
589 : * would hold if it was in cache, the page containing 'recptr'.
590 : */
591 : #define XLogRecPtrToBufIdx(recptr) \
592 : (((recptr) / XLOG_BLCKSZ) % (XLogCtl->XLogCacheBlck + 1))
593 :
594 : /*
595 : * These are the number of bytes in a WAL page usable for WAL data.
596 : */
597 : #define UsableBytesInPage (XLOG_BLCKSZ - SizeOfXLogShortPHD)
598 :
599 : /*
600 : * Convert values of GUCs measured in megabytes to equiv. segment count.
601 : * Rounds down.
602 : */
603 : #define ConvertToXSegs(x, segsize) XLogMBVarToSegs((x), (segsize))
604 :
605 : /* The number of bytes in a WAL segment usable for WAL data. */
606 : static int UsableBytesInSegment;
607 :
608 : /*
609 : * Private, possibly out-of-date copy of shared LogwrtResult.
610 : * See discussion above.
611 : */
612 : static XLogwrtResult LogwrtResult = {0, 0};
613 :
614 : /*
615 : * Update local copy of shared XLogCtl->log{Write,Flush}Result
616 : *
617 : * It's critical that Flush always trails Write, so the order of the reads is
618 : * important, as is the barrier. See also XLogWrite.
619 : */
620 : #define RefreshXLogWriteResult(_target) \
621 : do { \
622 : _target.Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult); \
623 : pg_read_barrier(); \
624 : _target.Write = pg_atomic_read_u64(&XLogCtl->logWriteResult); \
625 : } while (0)
626 :
627 : /*
628 : * openLogFile is -1 or a kernel FD for an open log file segment.
629 : * openLogSegNo identifies the segment, and openLogTLI the corresponding TLI.
630 : * These variables are only used to write the XLOG, and so will normally refer
631 : * to the active segment.
632 : *
633 : * Note: call Reserve/ReleaseExternalFD to track consumption of this FD.
634 : */
635 : static int openLogFile = -1;
636 : static XLogSegNo openLogSegNo = 0;
637 : static TimeLineID openLogTLI = 0;
638 :
639 : /*
640 : * Local copies of equivalent fields in the control file. When running
641 : * crash recovery, LocalMinRecoveryPoint is set to InvalidXLogRecPtr as we
642 : * expect to replay all the WAL available, and updateMinRecoveryPoint is
643 : * switched to false to prevent any updates while replaying records.
644 : * Those values are kept consistent as long as crash recovery runs.
645 : */
646 : static XLogRecPtr LocalMinRecoveryPoint;
647 : static TimeLineID LocalMinRecoveryPointTLI;
648 : static bool updateMinRecoveryPoint = true;
649 :
650 : /* For WALInsertLockAcquire/Release functions */
651 : static int MyLockNo = 0;
652 : static bool holdingAllLocks = false;
653 :
654 : #ifdef WAL_DEBUG
655 : static MemoryContext walDebugCxt = NULL;
656 : #endif
657 :
658 : static void CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI,
659 : XLogRecPtr EndOfLog,
660 : TimeLineID newTLI);
661 : static void CheckRequiredParameterValues(void);
662 : static void XLogReportParameters(void);
663 : static int LocalSetXLogInsertAllowed(void);
664 : static void CreateEndOfRecoveryRecord(void);
665 : static XLogRecPtr CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn,
666 : XLogRecPtr pagePtr,
667 : TimeLineID newTLI);
668 : static void CheckPointGuts(XLogRecPtr checkPointRedo, int flags);
669 : static void KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo);
670 : static XLogRecPtr XLogGetReplicationSlotMinimumLSN(void);
671 :
672 : static void AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli,
673 : bool opportunistic);
674 : static void XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible);
675 : static bool InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
676 : bool find_free, XLogSegNo max_segno,
677 : TimeLineID tli);
678 : static void XLogFileClose(void);
679 : static void PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli);
680 : static void RemoveTempXlogFiles(void);
681 : static void RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr,
682 : XLogRecPtr endptr, TimeLineID insertTLI);
683 : static void RemoveXlogFile(const struct dirent *segment_de,
684 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
685 : TimeLineID insertTLI);
686 : static void UpdateLastRemovedPtr(char *filename);
687 : static void ValidateXLOGDirectoryStructure(void);
688 : static void CleanupBackupHistory(void);
689 : static void UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force);
690 : static bool PerformRecoveryXLogAction(void);
691 : static void InitControlFile(uint64 sysidentifier, uint32 data_checksum_version);
692 : static void WriteControlFile(void);
693 : static void ReadControlFile(void);
694 : static void UpdateControlFile(void);
695 : static char *str_time(pg_time_t tnow);
696 :
697 : static int get_sync_bit(int method);
698 :
699 : static void CopyXLogRecordToWAL(int write_len, bool isLogSwitch,
700 : XLogRecData *rdata,
701 : XLogRecPtr StartPos, XLogRecPtr EndPos,
702 : TimeLineID tli);
703 : static void ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos,
704 : XLogRecPtr *EndPos, XLogRecPtr *PrevPtr);
705 : static bool ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos,
706 : XLogRecPtr *PrevPtr);
707 : static XLogRecPtr WaitXLogInsertionsToFinish(XLogRecPtr upto);
708 : static char *GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli);
709 : static XLogRecPtr XLogBytePosToRecPtr(uint64 bytepos);
710 : static XLogRecPtr XLogBytePosToEndRecPtr(uint64 bytepos);
711 : static uint64 XLogRecPtrToBytePos(XLogRecPtr ptr);
712 :
713 : static void WALInsertLockAcquire(void);
714 : static void WALInsertLockAcquireExclusive(void);
715 : static void WALInsertLockRelease(void);
716 : static void WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt);
717 :
718 : /*
719 : * Insert an XLOG record represented by an already-constructed chain of data
720 : * chunks. This is a low-level routine; to construct the WAL record header
721 : * and data, use the higher-level routines in xloginsert.c.
722 : *
723 : * If 'fpw_lsn' is valid, it is the oldest LSN among the pages that this
724 : * WAL record applies to, that were not included in the record as full page
725 : * images. If fpw_lsn <= RedoRecPtr, the function does not perform the
726 : * insertion and returns InvalidXLogRecPtr. The caller can then recalculate
727 : * which pages need a full-page image, and retry. If fpw_lsn is invalid, the
728 : * record is always inserted.
729 : *
730 : * 'flags' gives more in-depth control on the record being inserted. See
731 : * XLogSetRecordFlags() for details.
732 : *
733 : * 'topxid_included' tells whether the top-transaction id is logged along with
734 : * current subtransaction. See XLogRecordAssemble().
735 : *
736 : * The first XLogRecData in the chain must be for the record header, and its
737 : * data must be MAXALIGNed. XLogInsertRecord fills in the xl_prev and
738 : * xl_crc fields in the header, the rest of the header must already be filled
739 : * by the caller.
740 : *
741 : * Returns XLOG pointer to end of record (beginning of next record).
742 : * This can be used as LSN for data pages affected by the logged action.
743 : * (LSN is the XLOG point up to which the XLOG must be flushed to disk
744 : * before the data page can be written out. This implements the basic
745 : * WAL rule "write the log before the data".)
746 : */
747 : XLogRecPtr
748 28420966 : XLogInsertRecord(XLogRecData *rdata,
749 : XLogRecPtr fpw_lsn,
750 : uint8 flags,
751 : int num_fpi,
752 : bool topxid_included)
753 : {
754 28420966 : XLogCtlInsert *Insert = &XLogCtl->Insert;
755 : pg_crc32c rdata_crc;
756 : bool inserted;
757 28420966 : XLogRecord *rechdr = (XLogRecord *) rdata->data;
758 28420966 : uint8 info = rechdr->xl_info & ~XLR_INFO_MASK;
759 28420966 : WalInsertClass class = WALINSERT_NORMAL;
760 : XLogRecPtr StartPos;
761 : XLogRecPtr EndPos;
762 28420966 : bool prevDoPageWrites = doPageWrites;
763 : TimeLineID insertTLI;
764 :
765 : /* Does this record type require special handling? */
766 28420966 : if (unlikely(rechdr->xl_rmid == RM_XLOG_ID))
767 : {
768 414396 : if (info == XLOG_SWITCH)
769 750 : class = WALINSERT_SPECIAL_SWITCH;
770 413646 : else if (info == XLOG_CHECKPOINT_REDO)
771 1100 : class = WALINSERT_SPECIAL_CHECKPOINT;
772 : }
773 :
774 : /* we assume that all of the record header is in the first chunk */
775 : Assert(rdata->len >= SizeOfXLogRecord);
776 :
777 : /* cross-check on whether we should be here or not */
778 28420966 : if (!XLogInsertAllowed())
779 0 : elog(ERROR, "cannot make new WAL entries during recovery");
780 :
781 : /*
782 : * Given that we're not in recovery, InsertTimeLineID is set and can't
783 : * change, so we can read it without a lock.
784 : */
785 28420966 : insertTLI = XLogCtl->InsertTimeLineID;
786 :
787 : /*----------
788 : *
789 : * We have now done all the preparatory work we can without holding a
790 : * lock or modifying shared state. From here on, inserting the new WAL
791 : * record to the shared WAL buffer cache is a two-step process:
792 : *
793 : * 1. Reserve the right amount of space from the WAL. The current head of
794 : * reserved space is kept in Insert->CurrBytePos, and is protected by
795 : * insertpos_lck.
796 : *
797 : * 2. Copy the record to the reserved WAL space. This involves finding the
798 : * correct WAL buffer containing the reserved space, and copying the
799 : * record in place. This can be done concurrently in multiple processes.
800 : *
801 : * To keep track of which insertions are still in-progress, each concurrent
802 : * inserter acquires an insertion lock. In addition to just indicating that
803 : * an insertion is in progress, the lock tells others how far the inserter
804 : * has progressed. There is a small fixed number of insertion locks,
805 : * determined by NUM_XLOGINSERT_LOCKS. When an inserter crosses a page
806 : * boundary, it updates the value stored in the lock to the how far it has
807 : * inserted, to allow the previous buffer to be flushed.
808 : *
809 : * Holding onto an insertion lock also protects RedoRecPtr and
810 : * fullPageWrites from changing until the insertion is finished.
811 : *
812 : * Step 2 can usually be done completely in parallel. If the required WAL
813 : * page is not initialized yet, you have to grab WALBufMappingLock to
814 : * initialize it, but the WAL writer tries to do that ahead of insertions
815 : * to avoid that from happening in the critical path.
816 : *
817 : *----------
818 : */
819 28420966 : START_CRIT_SECTION();
820 :
821 28420966 : if (likely(class == WALINSERT_NORMAL))
822 : {
823 28419116 : WALInsertLockAcquire();
824 :
825 : /*
826 : * Check to see if my copy of RedoRecPtr is out of date. If so, may
827 : * have to go back and have the caller recompute everything. This can
828 : * only happen just after a checkpoint, so it's better to be slow in
829 : * this case and fast otherwise.
830 : *
831 : * Also check to see if fullPageWrites was just turned on or there's a
832 : * running backup (which forces full-page writes); if we weren't
833 : * already doing full-page writes then go back and recompute.
834 : *
835 : * If we aren't doing full-page writes then RedoRecPtr doesn't
836 : * actually affect the contents of the XLOG record, so we'll update
837 : * our local copy but not force a recomputation. (If doPageWrites was
838 : * just turned off, we could recompute the record without full pages,
839 : * but we choose not to bother.)
840 : */
841 28419116 : if (RedoRecPtr != Insert->RedoRecPtr)
842 : {
843 : Assert(RedoRecPtr < Insert->RedoRecPtr);
844 13114 : RedoRecPtr = Insert->RedoRecPtr;
845 : }
846 28419116 : doPageWrites = (Insert->fullPageWrites || Insert->runningBackups > 0);
847 :
848 28419116 : if (doPageWrites &&
849 27943068 : (!prevDoPageWrites ||
850 25661778 : (fpw_lsn != InvalidXLogRecPtr && fpw_lsn <= RedoRecPtr)))
851 : {
852 : /*
853 : * Oops, some buffer now needs to be backed up that the caller
854 : * didn't back up. Start over.
855 : */
856 14480 : WALInsertLockRelease();
857 14480 : END_CRIT_SECTION();
858 14480 : return InvalidXLogRecPtr;
859 : }
860 :
861 : /*
862 : * Reserve space for the record in the WAL. This also sets the xl_prev
863 : * pointer.
864 : */
865 28404636 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
866 : &rechdr->xl_prev);
867 :
868 : /* Normal records are always inserted. */
869 28404636 : inserted = true;
870 : }
871 1850 : else if (class == WALINSERT_SPECIAL_SWITCH)
872 : {
873 : /*
874 : * In order to insert an XLOG_SWITCH record, we need to hold all of
875 : * the WAL insertion locks, not just one, so that no one else can
876 : * begin inserting a record until we've figured out how much space
877 : * remains in the current WAL segment and claimed all of it.
878 : *
879 : * Nonetheless, this case is simpler than the normal cases handled
880 : * below, which must check for changes in doPageWrites and RedoRecPtr.
881 : * Those checks are only needed for records that can contain buffer
882 : * references, and an XLOG_SWITCH record never does.
883 : */
884 : Assert(fpw_lsn == InvalidXLogRecPtr);
885 750 : WALInsertLockAcquireExclusive();
886 750 : inserted = ReserveXLogSwitch(&StartPos, &EndPos, &rechdr->xl_prev);
887 : }
888 : else
889 : {
890 : Assert(class == WALINSERT_SPECIAL_CHECKPOINT);
891 :
892 : /*
893 : * We need to update both the local and shared copies of RedoRecPtr,
894 : * which means that we need to hold all the WAL insertion locks.
895 : * However, there can't be any buffer references, so as above, we need
896 : * not check RedoRecPtr before inserting the record; we just need to
897 : * update it afterwards.
898 : */
899 : Assert(fpw_lsn == InvalidXLogRecPtr);
900 1100 : WALInsertLockAcquireExclusive();
901 1100 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
902 : &rechdr->xl_prev);
903 1100 : RedoRecPtr = Insert->RedoRecPtr = StartPos;
904 1100 : inserted = true;
905 : }
906 :
907 28406486 : if (inserted)
908 : {
909 : /*
910 : * Now that xl_prev has been filled in, calculate CRC of the record
911 : * header.
912 : */
913 28406370 : rdata_crc = rechdr->xl_crc;
914 28406370 : COMP_CRC32C(rdata_crc, rechdr, offsetof(XLogRecord, xl_crc));
915 28406370 : FIN_CRC32C(rdata_crc);
916 28406370 : rechdr->xl_crc = rdata_crc;
917 :
918 : /*
919 : * All the record data, including the header, is now ready to be
920 : * inserted. Copy the record in the space reserved.
921 : */
922 28406370 : CopyXLogRecordToWAL(rechdr->xl_tot_len,
923 : class == WALINSERT_SPECIAL_SWITCH, rdata,
924 : StartPos, EndPos, insertTLI);
925 :
926 : /*
927 : * Unless record is flagged as not important, update LSN of last
928 : * important record in the current slot. When holding all locks, just
929 : * update the first one.
930 : */
931 28406370 : if ((flags & XLOG_MARK_UNIMPORTANT) == 0)
932 : {
933 28232568 : int lockno = holdingAllLocks ? 0 : MyLockNo;
934 :
935 28232568 : WALInsertLocks[lockno].l.lastImportantAt = StartPos;
936 : }
937 : }
938 : else
939 : {
940 : /*
941 : * This was an xlog-switch record, but the current insert location was
942 : * already exactly at the beginning of a segment, so there was no need
943 : * to do anything.
944 : */
945 : }
946 :
947 : /*
948 : * Done! Let others know that we're finished.
949 : */
950 28406486 : WALInsertLockRelease();
951 :
952 28406486 : END_CRIT_SECTION();
953 :
954 28406486 : MarkCurrentTransactionIdLoggedIfAny();
955 :
956 : /*
957 : * Mark top transaction id is logged (if needed) so that we should not try
958 : * to log it again with the next WAL record in the current subtransaction.
959 : */
960 28406486 : if (topxid_included)
961 438 : MarkSubxactTopXidLogged();
962 :
963 : /*
964 : * Update shared LogwrtRqst.Write, if we crossed page boundary.
965 : */
966 28406486 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
967 : {
968 3249980 : SpinLockAcquire(&XLogCtl->info_lck);
969 : /* advance global request to include new block(s) */
970 3249980 : if (XLogCtl->LogwrtRqst.Write < EndPos)
971 3092830 : XLogCtl->LogwrtRqst.Write = EndPos;
972 3249980 : SpinLockRelease(&XLogCtl->info_lck);
973 3249980 : RefreshXLogWriteResult(LogwrtResult);
974 : }
975 :
976 : /*
977 : * If this was an XLOG_SWITCH record, flush the record and the empty
978 : * padding space that fills the rest of the segment, and perform
979 : * end-of-segment actions (eg, notifying archiver).
980 : */
981 28406486 : if (class == WALINSERT_SPECIAL_SWITCH)
982 : {
983 : TRACE_POSTGRESQL_WAL_SWITCH();
984 750 : XLogFlush(EndPos);
985 :
986 : /*
987 : * Even though we reserved the rest of the segment for us, which is
988 : * reflected in EndPos, we return a pointer to just the end of the
989 : * xlog-switch record.
990 : */
991 750 : if (inserted)
992 : {
993 634 : EndPos = StartPos + SizeOfXLogRecord;
994 634 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
995 : {
996 0 : uint64 offset = XLogSegmentOffset(EndPos, wal_segment_size);
997 :
998 0 : if (offset == EndPos % XLOG_BLCKSZ)
999 0 : EndPos += SizeOfXLogLongPHD;
1000 : else
1001 0 : EndPos += SizeOfXLogShortPHD;
1002 : }
1003 : }
1004 : }
1005 :
1006 : #ifdef WAL_DEBUG
1007 : if (XLOG_DEBUG)
1008 : {
1009 : static XLogReaderState *debug_reader = NULL;
1010 : XLogRecord *record;
1011 : DecodedXLogRecord *decoded;
1012 : StringInfoData buf;
1013 : StringInfoData recordBuf;
1014 : char *errormsg = NULL;
1015 : MemoryContext oldCxt;
1016 :
1017 : oldCxt = MemoryContextSwitchTo(walDebugCxt);
1018 :
1019 : initStringInfo(&buf);
1020 : appendStringInfo(&buf, "INSERT @ %X/%X: ", LSN_FORMAT_ARGS(EndPos));
1021 :
1022 : /*
1023 : * We have to piece together the WAL record data from the XLogRecData
1024 : * entries, so that we can pass it to the rm_desc function as one
1025 : * contiguous chunk.
1026 : */
1027 : initStringInfo(&recordBuf);
1028 : for (; rdata != NULL; rdata = rdata->next)
1029 : appendBinaryStringInfo(&recordBuf, rdata->data, rdata->len);
1030 :
1031 : /* We also need temporary space to decode the record. */
1032 : record = (XLogRecord *) recordBuf.data;
1033 : decoded = (DecodedXLogRecord *)
1034 : palloc(DecodeXLogRecordRequiredSpace(record->xl_tot_len));
1035 :
1036 : if (!debug_reader)
1037 : debug_reader = XLogReaderAllocate(wal_segment_size, NULL,
1038 : XL_ROUTINE(.page_read = NULL,
1039 : .segment_open = NULL,
1040 : .segment_close = NULL),
1041 : NULL);
1042 : if (!debug_reader)
1043 : {
1044 : appendStringInfoString(&buf, "error decoding record: out of memory while allocating a WAL reading processor");
1045 : }
1046 : else if (!DecodeXLogRecord(debug_reader,
1047 : decoded,
1048 : record,
1049 : EndPos,
1050 : &errormsg))
1051 : {
1052 : appendStringInfo(&buf, "error decoding record: %s",
1053 : errormsg ? errormsg : "no error message");
1054 : }
1055 : else
1056 : {
1057 : appendStringInfoString(&buf, " - ");
1058 :
1059 : debug_reader->record = decoded;
1060 : xlog_outdesc(&buf, debug_reader);
1061 : debug_reader->record = NULL;
1062 : }
1063 : elog(LOG, "%s", buf.data);
1064 :
1065 : pfree(decoded);
1066 : pfree(buf.data);
1067 : pfree(recordBuf.data);
1068 : MemoryContextSwitchTo(oldCxt);
1069 : }
1070 : #endif
1071 :
1072 : /*
1073 : * Update our global variables
1074 : */
1075 28406486 : ProcLastRecPtr = StartPos;
1076 28406486 : XactLastRecEnd = EndPos;
1077 :
1078 : /* Report WAL traffic to the instrumentation. */
1079 28406486 : if (inserted)
1080 : {
1081 28406370 : pgWalUsage.wal_bytes += rechdr->xl_tot_len;
1082 28406370 : pgWalUsage.wal_records++;
1083 28406370 : pgWalUsage.wal_fpi += num_fpi;
1084 : }
1085 :
1086 28406486 : return EndPos;
1087 : }
1088 :
1089 : /*
1090 : * Reserves the right amount of space for a record of given size from the WAL.
1091 : * *StartPos is set to the beginning of the reserved section, *EndPos to
1092 : * its end+1. *PrevPtr is set to the beginning of the previous record; it is
1093 : * used to set the xl_prev of this record.
1094 : *
1095 : * This is the performance critical part of XLogInsert that must be serialized
1096 : * across backends. The rest can happen mostly in parallel. Try to keep this
1097 : * section as short as possible, insertpos_lck can be heavily contended on a
1098 : * busy system.
1099 : *
1100 : * NB: The space calculation here must match the code in CopyXLogRecordToWAL,
1101 : * where we actually copy the record to the reserved space.
1102 : *
1103 : * NB: Testing shows that XLogInsertRecord runs faster if this code is inlined;
1104 : * however, because there are two call sites, the compiler is reluctant to
1105 : * inline. We use pg_attribute_always_inline here to try to convince it.
1106 : */
1107 : static pg_attribute_always_inline void
1108 28405736 : ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1109 : XLogRecPtr *PrevPtr)
1110 : {
1111 28405736 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1112 : uint64 startbytepos;
1113 : uint64 endbytepos;
1114 : uint64 prevbytepos;
1115 :
1116 28405736 : size = MAXALIGN(size);
1117 :
1118 : /* All (non xlog-switch) records should contain data. */
1119 : Assert(size > SizeOfXLogRecord);
1120 :
1121 : /*
1122 : * The duration the spinlock needs to be held is minimized by minimizing
1123 : * the calculations that have to be done while holding the lock. The
1124 : * current tip of reserved WAL is kept in CurrBytePos, as a byte position
1125 : * that only counts "usable" bytes in WAL, that is, it excludes all WAL
1126 : * page headers. The mapping between "usable" byte positions and physical
1127 : * positions (XLogRecPtrs) can be done outside the locked region, and
1128 : * because the usable byte position doesn't include any headers, reserving
1129 : * X bytes from WAL is almost as simple as "CurrBytePos += X".
1130 : */
1131 28405736 : SpinLockAcquire(&Insert->insertpos_lck);
1132 :
1133 28405736 : startbytepos = Insert->CurrBytePos;
1134 28405736 : endbytepos = startbytepos + size;
1135 28405736 : prevbytepos = Insert->PrevBytePos;
1136 28405736 : Insert->CurrBytePos = endbytepos;
1137 28405736 : Insert->PrevBytePos = startbytepos;
1138 :
1139 28405736 : SpinLockRelease(&Insert->insertpos_lck);
1140 :
1141 28405736 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1142 28405736 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1143 28405736 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1144 :
1145 : /*
1146 : * Check that the conversions between "usable byte positions" and
1147 : * XLogRecPtrs work consistently in both directions.
1148 : */
1149 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1150 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1151 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1152 28405736 : }
1153 :
1154 : /*
1155 : * Like ReserveXLogInsertLocation(), but for an xlog-switch record.
1156 : *
1157 : * A log-switch record is handled slightly differently. The rest of the
1158 : * segment will be reserved for this insertion, as indicated by the returned
1159 : * *EndPos value. However, if we are already at the beginning of the current
1160 : * segment, *StartPos and *EndPos are set to the current location without
1161 : * reserving any space, and the function returns false.
1162 : */
1163 : static bool
1164 750 : ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1165 : {
1166 750 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1167 : uint64 startbytepos;
1168 : uint64 endbytepos;
1169 : uint64 prevbytepos;
1170 750 : uint32 size = MAXALIGN(SizeOfXLogRecord);
1171 : XLogRecPtr ptr;
1172 : uint32 segleft;
1173 :
1174 : /*
1175 : * These calculations are a bit heavy-weight to be done while holding a
1176 : * spinlock, but since we're holding all the WAL insertion locks, there
1177 : * are no other inserters competing for it. GetXLogInsertRecPtr() does
1178 : * compete for it, but that's not called very frequently.
1179 : */
1180 750 : SpinLockAcquire(&Insert->insertpos_lck);
1181 :
1182 750 : startbytepos = Insert->CurrBytePos;
1183 :
1184 750 : ptr = XLogBytePosToEndRecPtr(startbytepos);
1185 750 : if (XLogSegmentOffset(ptr, wal_segment_size) == 0)
1186 : {
1187 116 : SpinLockRelease(&Insert->insertpos_lck);
1188 116 : *EndPos = *StartPos = ptr;
1189 116 : return false;
1190 : }
1191 :
1192 634 : endbytepos = startbytepos + size;
1193 634 : prevbytepos = Insert->PrevBytePos;
1194 :
1195 634 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1196 634 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1197 :
1198 634 : segleft = wal_segment_size - XLogSegmentOffset(*EndPos, wal_segment_size);
1199 634 : if (segleft != wal_segment_size)
1200 : {
1201 : /* consume the rest of the segment */
1202 634 : *EndPos += segleft;
1203 634 : endbytepos = XLogRecPtrToBytePos(*EndPos);
1204 : }
1205 634 : Insert->CurrBytePos = endbytepos;
1206 634 : Insert->PrevBytePos = startbytepos;
1207 :
1208 634 : SpinLockRelease(&Insert->insertpos_lck);
1209 :
1210 634 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1211 :
1212 : Assert(XLogSegmentOffset(*EndPos, wal_segment_size) == 0);
1213 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1214 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1215 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1216 :
1217 634 : return true;
1218 : }
1219 :
1220 : /*
1221 : * Subroutine of XLogInsertRecord. Copies a WAL record to an already-reserved
1222 : * area in the WAL.
1223 : */
1224 : static void
1225 28406370 : CopyXLogRecordToWAL(int write_len, bool isLogSwitch, XLogRecData *rdata,
1226 : XLogRecPtr StartPos, XLogRecPtr EndPos, TimeLineID tli)
1227 : {
1228 : char *currpos;
1229 : int freespace;
1230 : int written;
1231 : XLogRecPtr CurrPos;
1232 : XLogPageHeader pagehdr;
1233 :
1234 : /*
1235 : * Get a pointer to the right place in the right WAL buffer to start
1236 : * inserting to.
1237 : */
1238 28406370 : CurrPos = StartPos;
1239 28406370 : currpos = GetXLogBuffer(CurrPos, tli);
1240 28406370 : freespace = INSERT_FREESPACE(CurrPos);
1241 :
1242 : /*
1243 : * there should be enough space for at least the first field (xl_tot_len)
1244 : * on this page.
1245 : */
1246 : Assert(freespace >= sizeof(uint32));
1247 :
1248 : /* Copy record data */
1249 28406370 : written = 0;
1250 134274386 : while (rdata != NULL)
1251 : {
1252 105868016 : const char *rdata_data = rdata->data;
1253 105868016 : int rdata_len = rdata->len;
1254 :
1255 109332484 : while (rdata_len > freespace)
1256 : {
1257 : /*
1258 : * Write what fits on this page, and continue on the next page.
1259 : */
1260 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || freespace == 0);
1261 3464468 : memcpy(currpos, rdata_data, freespace);
1262 3464468 : rdata_data += freespace;
1263 3464468 : rdata_len -= freespace;
1264 3464468 : written += freespace;
1265 3464468 : CurrPos += freespace;
1266 :
1267 : /*
1268 : * Get pointer to beginning of next page, and set the xlp_rem_len
1269 : * in the page header. Set XLP_FIRST_IS_CONTRECORD.
1270 : *
1271 : * It's safe to set the contrecord flag and xlp_rem_len without a
1272 : * lock on the page. All the other flags were already set when the
1273 : * page was initialized, in AdvanceXLInsertBuffer, and we're the
1274 : * only backend that needs to set the contrecord flag.
1275 : */
1276 3464468 : currpos = GetXLogBuffer(CurrPos, tli);
1277 3464468 : pagehdr = (XLogPageHeader) currpos;
1278 3464468 : pagehdr->xlp_rem_len = write_len - written;
1279 3464468 : pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1280 :
1281 : /* skip over the page header */
1282 3464468 : if (XLogSegmentOffset(CurrPos, wal_segment_size) == 0)
1283 : {
1284 2298 : CurrPos += SizeOfXLogLongPHD;
1285 2298 : currpos += SizeOfXLogLongPHD;
1286 : }
1287 : else
1288 : {
1289 3462170 : CurrPos += SizeOfXLogShortPHD;
1290 3462170 : currpos += SizeOfXLogShortPHD;
1291 : }
1292 3464468 : freespace = INSERT_FREESPACE(CurrPos);
1293 : }
1294 :
1295 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1296 105868016 : memcpy(currpos, rdata_data, rdata_len);
1297 105868016 : currpos += rdata_len;
1298 105868016 : CurrPos += rdata_len;
1299 105868016 : freespace -= rdata_len;
1300 105868016 : written += rdata_len;
1301 :
1302 105868016 : rdata = rdata->next;
1303 : }
1304 : Assert(written == write_len);
1305 :
1306 : /*
1307 : * If this was an xlog-switch, it's not enough to write the switch record,
1308 : * we also have to consume all the remaining space in the WAL segment. We
1309 : * have already reserved that space, but we need to actually fill it.
1310 : */
1311 28406370 : if (isLogSwitch && XLogSegmentOffset(CurrPos, wal_segment_size) != 0)
1312 : {
1313 : /* An xlog-switch record doesn't contain any data besides the header */
1314 : Assert(write_len == SizeOfXLogRecord);
1315 :
1316 : /* Assert that we did reserve the right amount of space */
1317 : Assert(XLogSegmentOffset(EndPos, wal_segment_size) == 0);
1318 :
1319 : /* Use up all the remaining space on the current page */
1320 634 : CurrPos += freespace;
1321 :
1322 : /*
1323 : * Cause all remaining pages in the segment to be flushed, leaving the
1324 : * XLog position where it should be, at the start of the next segment.
1325 : * We do this one page at a time, to make sure we don't deadlock
1326 : * against ourselves if wal_buffers < wal_segment_size.
1327 : */
1328 971058 : while (CurrPos < EndPos)
1329 : {
1330 : /*
1331 : * The minimal action to flush the page would be to call
1332 : * WALInsertLockUpdateInsertingAt(CurrPos) followed by
1333 : * AdvanceXLInsertBuffer(...). The page would be left initialized
1334 : * mostly to zeros, except for the page header (always the short
1335 : * variant, as this is never a segment's first page).
1336 : *
1337 : * The large vistas of zeros are good for compressibility, but the
1338 : * headers interrupting them every XLOG_BLCKSZ (with values that
1339 : * differ from page to page) are not. The effect varies with
1340 : * compression tool, but bzip2 for instance compresses about an
1341 : * order of magnitude worse if those headers are left in place.
1342 : *
1343 : * Rather than complicating AdvanceXLInsertBuffer itself (which is
1344 : * called in heavily-loaded circumstances as well as this lightly-
1345 : * loaded one) with variant behavior, we just use GetXLogBuffer
1346 : * (which itself calls the two methods we need) to get the pointer
1347 : * and zero most of the page. Then we just zero the page header.
1348 : */
1349 970424 : currpos = GetXLogBuffer(CurrPos, tli);
1350 3881696 : MemSet(currpos, 0, SizeOfXLogShortPHD);
1351 :
1352 970424 : CurrPos += XLOG_BLCKSZ;
1353 : }
1354 : }
1355 : else
1356 : {
1357 : /* Align the end position, so that the next record starts aligned */
1358 28405736 : CurrPos = MAXALIGN64(CurrPos);
1359 : }
1360 :
1361 28406370 : if (CurrPos != EndPos)
1362 0 : ereport(PANIC,
1363 : errcode(ERRCODE_DATA_CORRUPTED),
1364 : errmsg_internal("space reserved for WAL record does not match what was written"));
1365 28406370 : }
1366 :
1367 : /*
1368 : * Acquire a WAL insertion lock, for inserting to WAL.
1369 : */
1370 : static void
1371 28419138 : WALInsertLockAcquire(void)
1372 : {
1373 : bool immed;
1374 :
1375 : /*
1376 : * It doesn't matter which of the WAL insertion locks we acquire, so try
1377 : * the one we used last time. If the system isn't particularly busy, it's
1378 : * a good bet that it's still available, and it's good to have some
1379 : * affinity to a particular lock so that you don't unnecessarily bounce
1380 : * cache lines between processes when there's no contention.
1381 : *
1382 : * If this is the first time through in this backend, pick a lock
1383 : * (semi-)randomly. This allows the locks to be used evenly if you have a
1384 : * lot of very short connections.
1385 : */
1386 : static int lockToTry = -1;
1387 :
1388 28419138 : if (lockToTry == -1)
1389 15112 : lockToTry = MyProcNumber % NUM_XLOGINSERT_LOCKS;
1390 28419138 : MyLockNo = lockToTry;
1391 :
1392 : /*
1393 : * The insertingAt value is initially set to 0, as we don't know our
1394 : * insert location yet.
1395 : */
1396 28419138 : immed = LWLockAcquire(&WALInsertLocks[MyLockNo].l.lock, LW_EXCLUSIVE);
1397 28419138 : if (!immed)
1398 : {
1399 : /*
1400 : * If we couldn't get the lock immediately, try another lock next
1401 : * time. On a system with more insertion locks than concurrent
1402 : * inserters, this causes all the inserters to eventually migrate to a
1403 : * lock that no-one else is using. On a system with more inserters
1404 : * than locks, it still helps to distribute the inserters evenly
1405 : * across the locks.
1406 : */
1407 55068 : lockToTry = (lockToTry + 1) % NUM_XLOGINSERT_LOCKS;
1408 : }
1409 28419138 : }
1410 :
1411 : /*
1412 : * Acquire all WAL insertion locks, to prevent other backends from inserting
1413 : * to WAL.
1414 : */
1415 : static void
1416 6462 : WALInsertLockAcquireExclusive(void)
1417 : {
1418 : int i;
1419 :
1420 : /*
1421 : * When holding all the locks, all but the last lock's insertingAt
1422 : * indicator is set to 0xFFFFFFFFFFFFFFFF, which is higher than any real
1423 : * XLogRecPtr value, to make sure that no-one blocks waiting on those.
1424 : */
1425 51696 : for (i = 0; i < NUM_XLOGINSERT_LOCKS - 1; i++)
1426 : {
1427 45234 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1428 45234 : LWLockUpdateVar(&WALInsertLocks[i].l.lock,
1429 45234 : &WALInsertLocks[i].l.insertingAt,
1430 : PG_UINT64_MAX);
1431 : }
1432 : /* Variable value reset to 0 at release */
1433 6462 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1434 :
1435 6462 : holdingAllLocks = true;
1436 6462 : }
1437 :
1438 : /*
1439 : * Release our insertion lock (or locks, if we're holding them all).
1440 : *
1441 : * NB: Reset all variables to 0, so they cause LWLockWaitForVar to block the
1442 : * next time the lock is acquired.
1443 : */
1444 : static void
1445 28425600 : WALInsertLockRelease(void)
1446 : {
1447 28425600 : if (holdingAllLocks)
1448 : {
1449 : int i;
1450 :
1451 58158 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1452 51696 : LWLockReleaseClearVar(&WALInsertLocks[i].l.lock,
1453 51696 : &WALInsertLocks[i].l.insertingAt,
1454 : 0);
1455 :
1456 6462 : holdingAllLocks = false;
1457 : }
1458 : else
1459 : {
1460 28419138 : LWLockReleaseClearVar(&WALInsertLocks[MyLockNo].l.lock,
1461 28419138 : &WALInsertLocks[MyLockNo].l.insertingAt,
1462 : 0);
1463 : }
1464 28425600 : }
1465 :
1466 : /*
1467 : * Update our insertingAt value, to let others know that we've finished
1468 : * inserting up to that point.
1469 : */
1470 : static void
1471 4858652 : WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt)
1472 : {
1473 4858652 : if (holdingAllLocks)
1474 : {
1475 : /*
1476 : * We use the last lock to mark our actual position, see comments in
1477 : * WALInsertLockAcquireExclusive.
1478 : */
1479 967098 : LWLockUpdateVar(&WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.lock,
1480 967098 : &WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.insertingAt,
1481 : insertingAt);
1482 : }
1483 : else
1484 3891554 : LWLockUpdateVar(&WALInsertLocks[MyLockNo].l.lock,
1485 3891554 : &WALInsertLocks[MyLockNo].l.insertingAt,
1486 : insertingAt);
1487 4858652 : }
1488 :
1489 : /*
1490 : * Wait for any WAL insertions < upto to finish.
1491 : *
1492 : * Returns the location of the oldest insertion that is still in-progress.
1493 : * Any WAL prior to that point has been fully copied into WAL buffers, and
1494 : * can be flushed out to disk. Because this waits for any insertions older
1495 : * than 'upto' to finish, the return value is always >= 'upto'.
1496 : *
1497 : * Note: When you are about to write out WAL, you must call this function
1498 : * *before* acquiring WALWriteLock, to avoid deadlocks. This function might
1499 : * need to wait for an insertion to finish (or at least advance to next
1500 : * uninitialized page), and the inserter might need to evict an old WAL buffer
1501 : * to make room for a new one, which in turn requires WALWriteLock.
1502 : */
1503 : static XLogRecPtr
1504 4478002 : WaitXLogInsertionsToFinish(XLogRecPtr upto)
1505 : {
1506 : uint64 bytepos;
1507 : XLogRecPtr inserted;
1508 : XLogRecPtr reservedUpto;
1509 : XLogRecPtr finishedUpto;
1510 4478002 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1511 : int i;
1512 :
1513 4478002 : if (MyProc == NULL)
1514 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1515 :
1516 : /*
1517 : * Check if there's any work to do. Use a barrier to ensure we get the
1518 : * freshest value.
1519 : */
1520 4478002 : inserted = pg_atomic_read_membarrier_u64(&XLogCtl->logInsertResult);
1521 4478002 : if (upto <= inserted)
1522 3692670 : return inserted;
1523 :
1524 : /* Read the current insert position */
1525 785332 : SpinLockAcquire(&Insert->insertpos_lck);
1526 785332 : bytepos = Insert->CurrBytePos;
1527 785332 : SpinLockRelease(&Insert->insertpos_lck);
1528 785332 : reservedUpto = XLogBytePosToEndRecPtr(bytepos);
1529 :
1530 : /*
1531 : * No-one should request to flush a piece of WAL that hasn't even been
1532 : * reserved yet. However, it can happen if there is a block with a bogus
1533 : * LSN on disk, for example. XLogFlush checks for that situation and
1534 : * complains, but only after the flush. Here we just assume that to mean
1535 : * that all WAL that has been reserved needs to be finished. In this
1536 : * corner-case, the return value can be smaller than 'upto' argument.
1537 : */
1538 785332 : if (upto > reservedUpto)
1539 : {
1540 0 : ereport(LOG,
1541 : (errmsg("request to flush past end of generated WAL; request %X/%X, current position %X/%X",
1542 : LSN_FORMAT_ARGS(upto), LSN_FORMAT_ARGS(reservedUpto))));
1543 0 : upto = reservedUpto;
1544 : }
1545 :
1546 : /*
1547 : * Loop through all the locks, sleeping on any in-progress insert older
1548 : * than 'upto'.
1549 : *
1550 : * finishedUpto is our return value, indicating the point upto which all
1551 : * the WAL insertions have been finished. Initialize it to the head of
1552 : * reserved WAL, and as we iterate through the insertion locks, back it
1553 : * out for any insertion that's still in progress.
1554 : */
1555 785332 : finishedUpto = reservedUpto;
1556 7067988 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1557 : {
1558 6282656 : XLogRecPtr insertingat = InvalidXLogRecPtr;
1559 :
1560 : do
1561 : {
1562 : /*
1563 : * See if this insertion is in progress. LWLockWaitForVar will
1564 : * wait for the lock to be released, or for the 'value' to be set
1565 : * by a LWLockUpdateVar call. When a lock is initially acquired,
1566 : * its value is 0 (InvalidXLogRecPtr), which means that we don't
1567 : * know where it's inserting yet. We will have to wait for it. If
1568 : * it's a small insertion, the record will most likely fit on the
1569 : * same page and the inserter will release the lock without ever
1570 : * calling LWLockUpdateVar. But if it has to sleep, it will
1571 : * advertise the insertion point with LWLockUpdateVar before
1572 : * sleeping.
1573 : *
1574 : * In this loop we are only waiting for insertions that started
1575 : * before WaitXLogInsertionsToFinish was called. The lack of
1576 : * memory barriers in the loop means that we might see locks as
1577 : * "unused" that have since become used. This is fine because
1578 : * they only can be used for later insertions that we would not
1579 : * want to wait on anyway. Not taking a lock to acquire the
1580 : * current insertingAt value means that we might see older
1581 : * insertingAt values. This is also fine, because if we read a
1582 : * value too old, we will add ourselves to the wait queue, which
1583 : * contains atomic operations.
1584 : */
1585 6373334 : if (LWLockWaitForVar(&WALInsertLocks[i].l.lock,
1586 6373334 : &WALInsertLocks[i].l.insertingAt,
1587 : insertingat, &insertingat))
1588 : {
1589 : /* the lock was free, so no insertion in progress */
1590 4667578 : insertingat = InvalidXLogRecPtr;
1591 4667578 : break;
1592 : }
1593 :
1594 : /*
1595 : * This insertion is still in progress. Have to wait, unless the
1596 : * inserter has proceeded past 'upto'.
1597 : */
1598 1705756 : } while (insertingat < upto);
1599 :
1600 6282656 : if (insertingat != InvalidXLogRecPtr && insertingat < finishedUpto)
1601 659580 : finishedUpto = insertingat;
1602 : }
1603 :
1604 : /*
1605 : * Advance the limit we know to have been inserted and return the freshest
1606 : * value we know of, which might be beyond what we requested if somebody
1607 : * is concurrently doing this with an 'upto' pointer ahead of us.
1608 : */
1609 785332 : finishedUpto = pg_atomic_monotonic_advance_u64(&XLogCtl->logInsertResult,
1610 : finishedUpto);
1611 :
1612 785332 : return finishedUpto;
1613 : }
1614 :
1615 : /*
1616 : * Get a pointer to the right location in the WAL buffer containing the
1617 : * given XLogRecPtr.
1618 : *
1619 : * If the page is not initialized yet, it is initialized. That might require
1620 : * evicting an old dirty buffer from the buffer cache, which means I/O.
1621 : *
1622 : * The caller must ensure that the page containing the requested location
1623 : * isn't evicted yet, and won't be evicted. The way to ensure that is to
1624 : * hold onto a WAL insertion lock with the insertingAt position set to
1625 : * something <= ptr. GetXLogBuffer() will update insertingAt if it needs
1626 : * to evict an old page from the buffer. (This means that once you call
1627 : * GetXLogBuffer() with a given 'ptr', you must not access anything before
1628 : * that point anymore, and must not call GetXLogBuffer() with an older 'ptr'
1629 : * later, because older buffers might be recycled already)
1630 : */
1631 : static char *
1632 32841284 : GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli)
1633 : {
1634 : int idx;
1635 : XLogRecPtr endptr;
1636 : static uint64 cachedPage = 0;
1637 : static char *cachedPos = NULL;
1638 : XLogRecPtr expectedEndPtr;
1639 :
1640 : /*
1641 : * Fast path for the common case that we need to access again the same
1642 : * page as last time.
1643 : */
1644 32841284 : if (ptr / XLOG_BLCKSZ == cachedPage)
1645 : {
1646 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1647 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1648 27015098 : return cachedPos + ptr % XLOG_BLCKSZ;
1649 : }
1650 :
1651 : /*
1652 : * The XLog buffer cache is organized so that a page is always loaded to a
1653 : * particular buffer. That way we can easily calculate the buffer a given
1654 : * page must be loaded into, from the XLogRecPtr alone.
1655 : */
1656 5826186 : idx = XLogRecPtrToBufIdx(ptr);
1657 :
1658 : /*
1659 : * See what page is loaded in the buffer at the moment. It could be the
1660 : * page we're looking for, or something older. It can't be anything newer
1661 : * - that would imply the page we're looking for has already been written
1662 : * out to disk and evicted, and the caller is responsible for making sure
1663 : * that doesn't happen.
1664 : *
1665 : * We don't hold a lock while we read the value. If someone is just about
1666 : * to initialize or has just initialized the page, it's possible that we
1667 : * get InvalidXLogRecPtr. That's ok, we'll grab the mapping lock (in
1668 : * AdvanceXLInsertBuffer) and retry if we see anything other than the page
1669 : * we're looking for.
1670 : */
1671 5826186 : expectedEndPtr = ptr;
1672 5826186 : expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
1673 :
1674 5826186 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1675 5826186 : if (expectedEndPtr != endptr)
1676 : {
1677 : XLogRecPtr initializedUpto;
1678 :
1679 : /*
1680 : * Before calling AdvanceXLInsertBuffer(), which can block, let others
1681 : * know how far we're finished with inserting the record.
1682 : *
1683 : * NB: If 'ptr' points to just after the page header, advertise a
1684 : * position at the beginning of the page rather than 'ptr' itself. If
1685 : * there are no other insertions running, someone might try to flush
1686 : * up to our advertised location. If we advertised a position after
1687 : * the page header, someone might try to flush the page header, even
1688 : * though page might actually not be initialized yet. As the first
1689 : * inserter on the page, we are effectively responsible for making
1690 : * sure that it's initialized, before we let insertingAt to move past
1691 : * the page header.
1692 : */
1693 4858652 : if (ptr % XLOG_BLCKSZ == SizeOfXLogShortPHD &&
1694 11330 : XLogSegmentOffset(ptr, wal_segment_size) > XLOG_BLCKSZ)
1695 11330 : initializedUpto = ptr - SizeOfXLogShortPHD;
1696 4847322 : else if (ptr % XLOG_BLCKSZ == SizeOfXLogLongPHD &&
1697 1810 : XLogSegmentOffset(ptr, wal_segment_size) < XLOG_BLCKSZ)
1698 420 : initializedUpto = ptr - SizeOfXLogLongPHD;
1699 : else
1700 4846902 : initializedUpto = ptr;
1701 :
1702 4858652 : WALInsertLockUpdateInsertingAt(initializedUpto);
1703 :
1704 4858652 : AdvanceXLInsertBuffer(ptr, tli, false);
1705 4858652 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1706 :
1707 4858652 : if (expectedEndPtr != endptr)
1708 0 : elog(PANIC, "could not find WAL buffer for %X/%X",
1709 : LSN_FORMAT_ARGS(ptr));
1710 : }
1711 : else
1712 : {
1713 : /*
1714 : * Make sure the initialization of the page is visible to us, and
1715 : * won't arrive later to overwrite the WAL data we write on the page.
1716 : */
1717 967534 : pg_memory_barrier();
1718 : }
1719 :
1720 : /*
1721 : * Found the buffer holding this page. Return a pointer to the right
1722 : * offset within the page.
1723 : */
1724 5826186 : cachedPage = ptr / XLOG_BLCKSZ;
1725 5826186 : cachedPos = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1726 :
1727 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1728 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1729 :
1730 5826186 : return cachedPos + ptr % XLOG_BLCKSZ;
1731 : }
1732 :
1733 : /*
1734 : * Read WAL data directly from WAL buffers, if available. Returns the number
1735 : * of bytes read successfully.
1736 : *
1737 : * Fewer than 'count' bytes may be read if some of the requested WAL data has
1738 : * already been evicted.
1739 : *
1740 : * No locks are taken.
1741 : *
1742 : * Caller should ensure that it reads no further than LogwrtResult.Write
1743 : * (which should have been updated by the caller when determining how far to
1744 : * read). The 'tli' argument is only used as a convenient safety check so that
1745 : * callers do not read from WAL buffers on a historical timeline.
1746 : */
1747 : Size
1748 191314 : WALReadFromBuffers(char *dstbuf, XLogRecPtr startptr, Size count,
1749 : TimeLineID tli)
1750 : {
1751 191314 : char *pdst = dstbuf;
1752 191314 : XLogRecPtr recptr = startptr;
1753 : XLogRecPtr inserted;
1754 191314 : Size nbytes = count;
1755 :
1756 191314 : if (RecoveryInProgress() || tli != GetWALInsertionTimeLine())
1757 1818 : return 0;
1758 :
1759 : Assert(!XLogRecPtrIsInvalid(startptr));
1760 :
1761 : /*
1762 : * Caller should ensure that the requested data has been inserted into WAL
1763 : * buffers before we try to read it.
1764 : */
1765 189496 : inserted = pg_atomic_read_u64(&XLogCtl->logInsertResult);
1766 189496 : if (startptr + count > inserted)
1767 0 : ereport(ERROR,
1768 : errmsg("cannot read past end of generated WAL: requested %X/%X, current position %X/%X",
1769 : LSN_FORMAT_ARGS(startptr + count),
1770 : LSN_FORMAT_ARGS(inserted)));
1771 :
1772 : /*
1773 : * Loop through the buffers without a lock. For each buffer, atomically
1774 : * read and verify the end pointer, then copy the data out, and finally
1775 : * re-read and re-verify the end pointer.
1776 : *
1777 : * Once a page is evicted, it never returns to the WAL buffers, so if the
1778 : * end pointer matches the expected end pointer before and after we copy
1779 : * the data, then the right page must have been present during the data
1780 : * copy. Read barriers are necessary to ensure that the data copy actually
1781 : * happens between the two verification steps.
1782 : *
1783 : * If either verification fails, we simply terminate the loop and return
1784 : * with the data that had been already copied out successfully.
1785 : */
1786 207216 : while (nbytes > 0)
1787 : {
1788 202102 : uint32 offset = recptr % XLOG_BLCKSZ;
1789 202102 : int idx = XLogRecPtrToBufIdx(recptr);
1790 : XLogRecPtr expectedEndPtr;
1791 : XLogRecPtr endptr;
1792 : const char *page;
1793 : const char *psrc;
1794 : Size npagebytes;
1795 :
1796 : /*
1797 : * Calculate the end pointer we expect in the xlblocks array if the
1798 : * correct page is present.
1799 : */
1800 202102 : expectedEndPtr = recptr + (XLOG_BLCKSZ - offset);
1801 :
1802 : /*
1803 : * First verification step: check that the correct page is present in
1804 : * the WAL buffers.
1805 : */
1806 202102 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1807 202102 : if (expectedEndPtr != endptr)
1808 184380 : break;
1809 :
1810 : /*
1811 : * The correct page is present (or was at the time the endptr was
1812 : * read; must re-verify later). Calculate pointer to source data and
1813 : * determine how much data to read from this page.
1814 : */
1815 17722 : page = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1816 17722 : psrc = page + offset;
1817 17722 : npagebytes = Min(nbytes, XLOG_BLCKSZ - offset);
1818 :
1819 : /*
1820 : * Ensure that the data copy and the first verification step are not
1821 : * reordered.
1822 : */
1823 17722 : pg_read_barrier();
1824 :
1825 : /* data copy */
1826 17722 : memcpy(pdst, psrc, npagebytes);
1827 :
1828 : /*
1829 : * Ensure that the data copy and the second verification step are not
1830 : * reordered.
1831 : */
1832 17722 : pg_read_barrier();
1833 :
1834 : /*
1835 : * Second verification step: check that the page we read from wasn't
1836 : * evicted while we were copying the data.
1837 : */
1838 17722 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1839 17722 : if (expectedEndPtr != endptr)
1840 2 : break;
1841 :
1842 17720 : pdst += npagebytes;
1843 17720 : recptr += npagebytes;
1844 17720 : nbytes -= npagebytes;
1845 : }
1846 :
1847 : Assert(pdst - dstbuf <= count);
1848 :
1849 189496 : return pdst - dstbuf;
1850 : }
1851 :
1852 : /*
1853 : * Converts a "usable byte position" to XLogRecPtr. A usable byte position
1854 : * is the position starting from the beginning of WAL, excluding all WAL
1855 : * page headers.
1856 : */
1857 : static XLogRecPtr
1858 56817856 : XLogBytePosToRecPtr(uint64 bytepos)
1859 : {
1860 : uint64 fullsegs;
1861 : uint64 fullpages;
1862 : uint64 bytesleft;
1863 : uint32 seg_offset;
1864 : XLogRecPtr result;
1865 :
1866 56817856 : fullsegs = bytepos / UsableBytesInSegment;
1867 56817856 : bytesleft = bytepos % UsableBytesInSegment;
1868 :
1869 56817856 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1870 : {
1871 : /* fits on first page of segment */
1872 93612 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1873 : }
1874 : else
1875 : {
1876 : /* account for the first page on segment with long header */
1877 56724244 : seg_offset = XLOG_BLCKSZ;
1878 56724244 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1879 :
1880 56724244 : fullpages = bytesleft / UsableBytesInPage;
1881 56724244 : bytesleft = bytesleft % UsableBytesInPage;
1882 :
1883 56724244 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1884 : }
1885 :
1886 56817856 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1887 :
1888 56817856 : return result;
1889 : }
1890 :
1891 : /*
1892 : * Like XLogBytePosToRecPtr, but if the position is at a page boundary,
1893 : * returns a pointer to the beginning of the page (ie. before page header),
1894 : * not to where the first xlog record on that page would go to. This is used
1895 : * when converting a pointer to the end of a record.
1896 : */
1897 : static XLogRecPtr
1898 29192452 : XLogBytePosToEndRecPtr(uint64 bytepos)
1899 : {
1900 : uint64 fullsegs;
1901 : uint64 fullpages;
1902 : uint64 bytesleft;
1903 : uint32 seg_offset;
1904 : XLogRecPtr result;
1905 :
1906 29192452 : fullsegs = bytepos / UsableBytesInSegment;
1907 29192452 : bytesleft = bytepos % UsableBytesInSegment;
1908 :
1909 29192452 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1910 : {
1911 : /* fits on first page of segment */
1912 145254 : if (bytesleft == 0)
1913 97244 : seg_offset = 0;
1914 : else
1915 48010 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1916 : }
1917 : else
1918 : {
1919 : /* account for the first page on segment with long header */
1920 29047198 : seg_offset = XLOG_BLCKSZ;
1921 29047198 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1922 :
1923 29047198 : fullpages = bytesleft / UsableBytesInPage;
1924 29047198 : bytesleft = bytesleft % UsableBytesInPage;
1925 :
1926 29047198 : if (bytesleft == 0)
1927 28216 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
1928 : else
1929 29018982 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1930 : }
1931 :
1932 29192452 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1933 :
1934 29192452 : return result;
1935 : }
1936 :
1937 : /*
1938 : * Convert an XLogRecPtr to a "usable byte position".
1939 : */
1940 : static uint64
1941 3930 : XLogRecPtrToBytePos(XLogRecPtr ptr)
1942 : {
1943 : uint64 fullsegs;
1944 : uint32 fullpages;
1945 : uint32 offset;
1946 : uint64 result;
1947 :
1948 3930 : XLByteToSeg(ptr, fullsegs, wal_segment_size);
1949 :
1950 3930 : fullpages = (XLogSegmentOffset(ptr, wal_segment_size)) / XLOG_BLCKSZ;
1951 3930 : offset = ptr % XLOG_BLCKSZ;
1952 :
1953 3930 : if (fullpages == 0)
1954 : {
1955 1140 : result = fullsegs * UsableBytesInSegment;
1956 1140 : if (offset > 0)
1957 : {
1958 : Assert(offset >= SizeOfXLogLongPHD);
1959 464 : result += offset - SizeOfXLogLongPHD;
1960 : }
1961 : }
1962 : else
1963 : {
1964 2790 : result = fullsegs * UsableBytesInSegment +
1965 2790 : (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
1966 2790 : (fullpages - 1) * UsableBytesInPage; /* full pages */
1967 2790 : if (offset > 0)
1968 : {
1969 : Assert(offset >= SizeOfXLogShortPHD);
1970 2768 : result += offset - SizeOfXLogShortPHD;
1971 : }
1972 : }
1973 :
1974 3930 : return result;
1975 : }
1976 :
1977 : /*
1978 : * Initialize XLOG buffers, writing out old buffers if they still contain
1979 : * unwritten data, upto the page containing 'upto'. Or if 'opportunistic' is
1980 : * true, initialize as many pages as we can without having to write out
1981 : * unwritten data. Any new pages are initialized to zeros, with pages headers
1982 : * initialized properly.
1983 : */
1984 : static void
1985 4866030 : AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
1986 : {
1987 4866030 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1988 : int nextidx;
1989 : XLogRecPtr OldPageRqstPtr;
1990 : XLogwrtRqst WriteRqst;
1991 4866030 : XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
1992 : XLogRecPtr NewPageBeginPtr;
1993 : XLogPageHeader NewPage;
1994 4866030 : int npages pg_attribute_unused() = 0;
1995 :
1996 4866030 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
1997 :
1998 : /*
1999 : * Now that we have the lock, check if someone initialized the page
2000 : * already.
2001 : */
2002 13624076 : while (upto >= XLogCtl->InitializedUpTo || opportunistic)
2003 : {
2004 8765424 : nextidx = XLogRecPtrToBufIdx(XLogCtl->InitializedUpTo);
2005 :
2006 : /*
2007 : * Get ending-offset of the buffer page we need to replace (this may
2008 : * be zero if the buffer hasn't been used yet). Fall through if it's
2009 : * already written out.
2010 : */
2011 8765424 : OldPageRqstPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]);
2012 8765424 : if (LogwrtResult.Write < OldPageRqstPtr)
2013 : {
2014 : /*
2015 : * Nope, got work to do. If we just want to pre-initialize as much
2016 : * as we can without flushing, give up now.
2017 : */
2018 4384190 : if (opportunistic)
2019 7378 : break;
2020 :
2021 : /* Advance shared memory write request position */
2022 4376812 : SpinLockAcquire(&XLogCtl->info_lck);
2023 4376812 : if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
2024 918732 : XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
2025 4376812 : SpinLockRelease(&XLogCtl->info_lck);
2026 :
2027 : /*
2028 : * Acquire an up-to-date LogwrtResult value and see if we still
2029 : * need to write it or if someone else already did.
2030 : */
2031 4376812 : RefreshXLogWriteResult(LogwrtResult);
2032 4376812 : if (LogwrtResult.Write < OldPageRqstPtr)
2033 : {
2034 : /*
2035 : * Must acquire write lock. Release WALBufMappingLock first,
2036 : * to make sure that all insertions that we need to wait for
2037 : * can finish (up to this same position). Otherwise we risk
2038 : * deadlock.
2039 : */
2040 4219204 : LWLockRelease(WALBufMappingLock);
2041 :
2042 4219204 : WaitXLogInsertionsToFinish(OldPageRqstPtr);
2043 :
2044 4219204 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2045 :
2046 4219204 : RefreshXLogWriteResult(LogwrtResult);
2047 4219204 : if (LogwrtResult.Write >= OldPageRqstPtr)
2048 : {
2049 : /* OK, someone wrote it already */
2050 649192 : LWLockRelease(WALWriteLock);
2051 : }
2052 : else
2053 : {
2054 : /* Have to write it ourselves */
2055 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2056 3570012 : WriteRqst.Write = OldPageRqstPtr;
2057 3570012 : WriteRqst.Flush = 0;
2058 3570012 : XLogWrite(WriteRqst, tli, false);
2059 3570012 : LWLockRelease(WALWriteLock);
2060 3570012 : pgWalUsage.wal_buffers_full++;
2061 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_DONE();
2062 : }
2063 : /* Re-acquire WALBufMappingLock and retry */
2064 4219204 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2065 4219204 : continue;
2066 : }
2067 : }
2068 :
2069 : /*
2070 : * Now the next buffer slot is free and we can set it up to be the
2071 : * next output page.
2072 : */
2073 4538842 : NewPageBeginPtr = XLogCtl->InitializedUpTo;
2074 4538842 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2075 :
2076 : Assert(XLogRecPtrToBufIdx(NewPageBeginPtr) == nextidx);
2077 :
2078 4538842 : NewPage = (XLogPageHeader) (XLogCtl->pages + nextidx * (Size) XLOG_BLCKSZ);
2079 :
2080 : /*
2081 : * Mark the xlblock with InvalidXLogRecPtr and issue a write barrier
2082 : * before initializing. Otherwise, the old page may be partially
2083 : * zeroed but look valid.
2084 : */
2085 4538842 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], InvalidXLogRecPtr);
2086 4538842 : pg_write_barrier();
2087 :
2088 : /*
2089 : * Be sure to re-zero the buffer so that bytes beyond what we've
2090 : * written will look like zeroes and not valid XLOG records...
2091 : */
2092 4538842 : MemSet(NewPage, 0, XLOG_BLCKSZ);
2093 :
2094 : /*
2095 : * Fill the new page's header
2096 : */
2097 4538842 : NewPage->xlp_magic = XLOG_PAGE_MAGIC;
2098 :
2099 : /* NewPage->xlp_info = 0; */ /* done by memset */
2100 4538842 : NewPage->xlp_tli = tli;
2101 4538842 : NewPage->xlp_pageaddr = NewPageBeginPtr;
2102 :
2103 : /* NewPage->xlp_rem_len = 0; */ /* done by memset */
2104 :
2105 : /*
2106 : * If online backup is not in progress, mark the header to indicate
2107 : * that WAL records beginning in this page have removable backup
2108 : * blocks. This allows the WAL archiver to know whether it is safe to
2109 : * compress archived WAL data by transforming full-block records into
2110 : * the non-full-block format. It is sufficient to record this at the
2111 : * page level because we force a page switch (in fact a segment
2112 : * switch) when starting a backup, so the flag will be off before any
2113 : * records can be written during the backup. At the end of a backup,
2114 : * the last page will be marked as all unsafe when perhaps only part
2115 : * is unsafe, but at worst the archiver would miss the opportunity to
2116 : * compress a few records.
2117 : */
2118 4538842 : if (Insert->runningBackups == 0)
2119 4285426 : NewPage->xlp_info |= XLP_BKP_REMOVABLE;
2120 :
2121 : /*
2122 : * If first page of an XLOG segment file, make it a long header.
2123 : */
2124 4538842 : if ((XLogSegmentOffset(NewPage->xlp_pageaddr, wal_segment_size)) == 0)
2125 : {
2126 2746 : XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2127 :
2128 2746 : NewLongPage->xlp_sysid = ControlFile->system_identifier;
2129 2746 : NewLongPage->xlp_seg_size = wal_segment_size;
2130 2746 : NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2131 2746 : NewPage->xlp_info |= XLP_LONG_HEADER;
2132 : }
2133 :
2134 : /*
2135 : * Make sure the initialization of the page becomes visible to others
2136 : * before the xlblocks update. GetXLogBuffer() reads xlblocks without
2137 : * holding a lock.
2138 : */
2139 4538842 : pg_write_barrier();
2140 :
2141 4538842 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], NewPageEndPtr);
2142 4538842 : XLogCtl->InitializedUpTo = NewPageEndPtr;
2143 :
2144 4538842 : npages++;
2145 : }
2146 4866030 : LWLockRelease(WALBufMappingLock);
2147 :
2148 : #ifdef WAL_DEBUG
2149 : if (XLOG_DEBUG && npages > 0)
2150 : {
2151 : elog(DEBUG1, "initialized %d pages, up to %X/%X",
2152 : npages, LSN_FORMAT_ARGS(NewPageEndPtr));
2153 : }
2154 : #endif
2155 4866030 : }
2156 :
2157 : /*
2158 : * Calculate CheckPointSegments based on max_wal_size_mb and
2159 : * checkpoint_completion_target.
2160 : */
2161 : static void
2162 14248 : CalculateCheckpointSegments(void)
2163 : {
2164 : double target;
2165 :
2166 : /*-------
2167 : * Calculate the distance at which to trigger a checkpoint, to avoid
2168 : * exceeding max_wal_size_mb. This is based on two assumptions:
2169 : *
2170 : * a) we keep WAL for only one checkpoint cycle (prior to PG11 we kept
2171 : * WAL for two checkpoint cycles to allow us to recover from the
2172 : * secondary checkpoint if the first checkpoint failed, though we
2173 : * only did this on the primary anyway, not on standby. Keeping just
2174 : * one checkpoint simplifies processing and reduces disk space in
2175 : * many smaller databases.)
2176 : * b) during checkpoint, we consume checkpoint_completion_target *
2177 : * number of segments consumed between checkpoints.
2178 : *-------
2179 : */
2180 14248 : target = (double) ConvertToXSegs(max_wal_size_mb, wal_segment_size) /
2181 14248 : (1.0 + CheckPointCompletionTarget);
2182 :
2183 : /* round down */
2184 14248 : CheckPointSegments = (int) target;
2185 :
2186 14248 : if (CheckPointSegments < 1)
2187 18 : CheckPointSegments = 1;
2188 14248 : }
2189 :
2190 : void
2191 10282 : assign_max_wal_size(int newval, void *extra)
2192 : {
2193 10282 : max_wal_size_mb = newval;
2194 10282 : CalculateCheckpointSegments();
2195 10282 : }
2196 :
2197 : void
2198 2098 : assign_checkpoint_completion_target(double newval, void *extra)
2199 : {
2200 2098 : CheckPointCompletionTarget = newval;
2201 2098 : CalculateCheckpointSegments();
2202 2098 : }
2203 :
2204 : bool
2205 4058 : check_wal_segment_size(int *newval, void **extra, GucSource source)
2206 : {
2207 4058 : if (!IsValidWalSegSize(*newval))
2208 : {
2209 0 : GUC_check_errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.");
2210 0 : return false;
2211 : }
2212 :
2213 4058 : return true;
2214 : }
2215 :
2216 : /*
2217 : * GUC check_hook for max_slot_wal_keep_size
2218 : *
2219 : * We don't allow the value of max_slot_wal_keep_size other than -1 during the
2220 : * binary upgrade. See start_postmaster() in pg_upgrade for more details.
2221 : */
2222 : bool
2223 2350 : check_max_slot_wal_keep_size(int *newval, void **extra, GucSource source)
2224 : {
2225 2350 : if (IsBinaryUpgrade && *newval != -1)
2226 : {
2227 0 : GUC_check_errdetail("\"%s\" must be set to -1 during binary upgrade mode.",
2228 : "max_slot_wal_keep_size");
2229 0 : return false;
2230 : }
2231 :
2232 2350 : return true;
2233 : }
2234 :
2235 : /*
2236 : * At a checkpoint, how many WAL segments to recycle as preallocated future
2237 : * XLOG segments? Returns the highest segment that should be preallocated.
2238 : */
2239 : static XLogSegNo
2240 2606 : XLOGfileslop(XLogRecPtr lastredoptr)
2241 : {
2242 : XLogSegNo minSegNo;
2243 : XLogSegNo maxSegNo;
2244 : double distance;
2245 : XLogSegNo recycleSegNo;
2246 :
2247 : /*
2248 : * Calculate the segment numbers that min_wal_size_mb and max_wal_size_mb
2249 : * correspond to. Always recycle enough segments to meet the minimum, and
2250 : * remove enough segments to stay below the maximum.
2251 : */
2252 2606 : minSegNo = lastredoptr / wal_segment_size +
2253 2606 : ConvertToXSegs(min_wal_size_mb, wal_segment_size) - 1;
2254 2606 : maxSegNo = lastredoptr / wal_segment_size +
2255 2606 : ConvertToXSegs(max_wal_size_mb, wal_segment_size) - 1;
2256 :
2257 : /*
2258 : * Between those limits, recycle enough segments to get us through to the
2259 : * estimated end of next checkpoint.
2260 : *
2261 : * To estimate where the next checkpoint will finish, assume that the
2262 : * system runs steadily consuming CheckPointDistanceEstimate bytes between
2263 : * every checkpoint.
2264 : */
2265 2606 : distance = (1.0 + CheckPointCompletionTarget) * CheckPointDistanceEstimate;
2266 : /* add 10% for good measure. */
2267 2606 : distance *= 1.10;
2268 :
2269 2606 : recycleSegNo = (XLogSegNo) ceil(((double) lastredoptr + distance) /
2270 : wal_segment_size);
2271 :
2272 2606 : if (recycleSegNo < minSegNo)
2273 1640 : recycleSegNo = minSegNo;
2274 2606 : if (recycleSegNo > maxSegNo)
2275 760 : recycleSegNo = maxSegNo;
2276 :
2277 2606 : return recycleSegNo;
2278 : }
2279 :
2280 : /*
2281 : * Check whether we've consumed enough xlog space that a checkpoint is needed.
2282 : *
2283 : * new_segno indicates a log file that has just been filled up (or read
2284 : * during recovery). We measure the distance from RedoRecPtr to new_segno
2285 : * and see if that exceeds CheckPointSegments.
2286 : *
2287 : * Note: it is caller's responsibility that RedoRecPtr is up-to-date.
2288 : */
2289 : bool
2290 8150 : XLogCheckpointNeeded(XLogSegNo new_segno)
2291 : {
2292 : XLogSegNo old_segno;
2293 :
2294 8150 : XLByteToSeg(RedoRecPtr, old_segno, wal_segment_size);
2295 :
2296 8150 : if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2297 5422 : return true;
2298 2728 : return false;
2299 : }
2300 :
2301 : /*
2302 : * Write and/or fsync the log at least as far as WriteRqst indicates.
2303 : *
2304 : * If flexible == true, we don't have to write as far as WriteRqst, but
2305 : * may stop at any convenient boundary (such as a cache or logfile boundary).
2306 : * This option allows us to avoid uselessly issuing multiple writes when a
2307 : * single one would do.
2308 : *
2309 : * Must be called with WALWriteLock held. WaitXLogInsertionsToFinish(WriteRqst)
2310 : * must be called before grabbing the lock, to make sure the data is ready to
2311 : * write.
2312 : */
2313 : static void
2314 3815952 : XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible)
2315 : {
2316 : bool ispartialpage;
2317 : bool last_iteration;
2318 : bool finishing_seg;
2319 : int curridx;
2320 : int npages;
2321 : int startidx;
2322 : uint32 startoffset;
2323 :
2324 : /* We should always be inside a critical section here */
2325 : Assert(CritSectionCount > 0);
2326 :
2327 : /*
2328 : * Update local LogwrtResult (caller probably did this already, but...)
2329 : */
2330 3815952 : RefreshXLogWriteResult(LogwrtResult);
2331 :
2332 : /*
2333 : * Since successive pages in the xlog cache are consecutively allocated,
2334 : * we can usually gather multiple pages together and issue just one
2335 : * write() call. npages is the number of pages we have determined can be
2336 : * written together; startidx is the cache block index of the first one,
2337 : * and startoffset is the file offset at which it should go. The latter
2338 : * two variables are only valid when npages > 0, but we must initialize
2339 : * all of them to keep the compiler quiet.
2340 : */
2341 3815952 : npages = 0;
2342 3815952 : startidx = 0;
2343 3815952 : startoffset = 0;
2344 :
2345 : /*
2346 : * Within the loop, curridx is the cache block index of the page to
2347 : * consider writing. Begin at the buffer containing the next unwritten
2348 : * page, or last partially written page.
2349 : */
2350 3815952 : curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2351 :
2352 8252892 : while (LogwrtResult.Write < WriteRqst.Write)
2353 : {
2354 : /*
2355 : * Make sure we're not ahead of the insert process. This could happen
2356 : * if we're passed a bogus WriteRqst.Write that is past the end of the
2357 : * last page that's been initialized by AdvanceXLInsertBuffer.
2358 : */
2359 4676634 : XLogRecPtr EndPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[curridx]);
2360 :
2361 4676634 : if (LogwrtResult.Write >= EndPtr)
2362 0 : elog(PANIC, "xlog write request %X/%X is past end of log %X/%X",
2363 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2364 : LSN_FORMAT_ARGS(EndPtr));
2365 :
2366 : /* Advance LogwrtResult.Write to end of current buffer page */
2367 4676634 : LogwrtResult.Write = EndPtr;
2368 4676634 : ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2369 :
2370 4676634 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2371 : wal_segment_size))
2372 : {
2373 : /*
2374 : * Switch to new logfile segment. We cannot have any pending
2375 : * pages here (since we dump what we have at segment end).
2376 : */
2377 : Assert(npages == 0);
2378 23092 : if (openLogFile >= 0)
2379 9434 : XLogFileClose();
2380 23092 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2381 : wal_segment_size);
2382 23092 : openLogTLI = tli;
2383 :
2384 : /* create/use new log file */
2385 23092 : openLogFile = XLogFileInit(openLogSegNo, tli);
2386 23092 : ReserveExternalFD();
2387 : }
2388 :
2389 : /* Make sure we have the current logfile open */
2390 4676634 : if (openLogFile < 0)
2391 : {
2392 0 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2393 : wal_segment_size);
2394 0 : openLogTLI = tli;
2395 0 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2396 0 : ReserveExternalFD();
2397 : }
2398 :
2399 : /* Add current page to the set of pending pages-to-dump */
2400 4676634 : if (npages == 0)
2401 : {
2402 : /* first of group */
2403 3845986 : startidx = curridx;
2404 3845986 : startoffset = XLogSegmentOffset(LogwrtResult.Write - XLOG_BLCKSZ,
2405 : wal_segment_size);
2406 : }
2407 4676634 : npages++;
2408 :
2409 : /*
2410 : * Dump the set if this will be the last loop iteration, or if we are
2411 : * at the last page of the cache area (since the next page won't be
2412 : * contiguous in memory), or if we are at the end of the logfile
2413 : * segment.
2414 : */
2415 4676634 : last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2416 :
2417 9119694 : finishing_seg = !ispartialpage &&
2418 4443060 : (startoffset + npages * XLOG_BLCKSZ) >= wal_segment_size;
2419 :
2420 4676634 : if (last_iteration ||
2421 861822 : curridx == XLogCtl->XLogCacheBlck ||
2422 : finishing_seg)
2423 : {
2424 : char *from;
2425 : Size nbytes;
2426 : Size nleft;
2427 : ssize_t written;
2428 : instr_time start;
2429 :
2430 : /* OK to write the page(s) */
2431 3845986 : from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2432 3845986 : nbytes = npages * (Size) XLOG_BLCKSZ;
2433 3845986 : nleft = nbytes;
2434 : do
2435 : {
2436 3845986 : errno = 0;
2437 :
2438 : /*
2439 : * Measure I/O timing to write WAL data, for pg_stat_io.
2440 : */
2441 3845986 : start = pgstat_prepare_io_time(track_wal_io_timing);
2442 :
2443 3845986 : pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE);
2444 3845986 : written = pg_pwrite(openLogFile, from, nleft, startoffset);
2445 3845986 : pgstat_report_wait_end();
2446 :
2447 3845986 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL,
2448 : IOOP_WRITE, start, 1, written);
2449 :
2450 3845986 : if (written <= 0)
2451 : {
2452 : char xlogfname[MAXFNAMELEN];
2453 : int save_errno;
2454 :
2455 0 : if (errno == EINTR)
2456 0 : continue;
2457 :
2458 0 : save_errno = errno;
2459 0 : XLogFileName(xlogfname, tli, openLogSegNo,
2460 : wal_segment_size);
2461 0 : errno = save_errno;
2462 0 : ereport(PANIC,
2463 : (errcode_for_file_access(),
2464 : errmsg("could not write to log file \"%s\" at offset %u, length %zu: %m",
2465 : xlogfname, startoffset, nleft)));
2466 : }
2467 3845986 : nleft -= written;
2468 3845986 : from += written;
2469 3845986 : startoffset += written;
2470 3845986 : } while (nleft > 0);
2471 :
2472 3845986 : npages = 0;
2473 :
2474 : /*
2475 : * If we just wrote the whole last page of a logfile segment,
2476 : * fsync the segment immediately. This avoids having to go back
2477 : * and re-open prior segments when an fsync request comes along
2478 : * later. Doing it here ensures that one and only one backend will
2479 : * perform this fsync.
2480 : *
2481 : * This is also the right place to notify the Archiver that the
2482 : * segment is ready to copy to archival storage, and to update the
2483 : * timer for archive_timeout, and to signal for a checkpoint if
2484 : * too many logfile segments have been used since the last
2485 : * checkpoint.
2486 : */
2487 3845986 : if (finishing_seg)
2488 : {
2489 2962 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2490 :
2491 : /* signal that we need to wakeup walsenders later */
2492 2962 : WalSndWakeupRequest();
2493 :
2494 2962 : LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2495 :
2496 2962 : if (XLogArchivingActive())
2497 168 : XLogArchiveNotifySeg(openLogSegNo, tli);
2498 :
2499 2962 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2500 2962 : XLogCtl->lastSegSwitchLSN = LogwrtResult.Flush;
2501 :
2502 : /*
2503 : * Request a checkpoint if we've consumed too much xlog since
2504 : * the last one. For speed, we first check using the local
2505 : * copy of RedoRecPtr, which might be out of date; if it looks
2506 : * like a checkpoint is needed, forcibly update RedoRecPtr and
2507 : * recheck.
2508 : */
2509 2962 : if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2510 : {
2511 458 : (void) GetRedoRecPtr();
2512 458 : if (XLogCheckpointNeeded(openLogSegNo))
2513 382 : RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2514 : }
2515 : }
2516 : }
2517 :
2518 4676634 : if (ispartialpage)
2519 : {
2520 : /* Only asked to write a partial page */
2521 233574 : LogwrtResult.Write = WriteRqst.Write;
2522 233574 : break;
2523 : }
2524 4443060 : curridx = NextBufIdx(curridx);
2525 :
2526 : /* If flexible, break out of loop as soon as we wrote something */
2527 4443060 : if (flexible && npages == 0)
2528 6120 : break;
2529 : }
2530 :
2531 : Assert(npages == 0);
2532 :
2533 : /*
2534 : * If asked to flush, do so
2535 : */
2536 3815952 : if (LogwrtResult.Flush < WriteRqst.Flush &&
2537 245272 : LogwrtResult.Flush < LogwrtResult.Write)
2538 : {
2539 : /*
2540 : * Could get here without iterating above loop, in which case we might
2541 : * have no open file or the wrong one. However, we do not need to
2542 : * fsync more than one file.
2543 : */
2544 245176 : if (wal_sync_method != WAL_SYNC_METHOD_OPEN &&
2545 245176 : wal_sync_method != WAL_SYNC_METHOD_OPEN_DSYNC)
2546 : {
2547 245176 : if (openLogFile >= 0 &&
2548 245158 : !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2549 : wal_segment_size))
2550 40 : XLogFileClose();
2551 245176 : if (openLogFile < 0)
2552 : {
2553 58 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2554 : wal_segment_size);
2555 58 : openLogTLI = tli;
2556 58 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2557 58 : ReserveExternalFD();
2558 : }
2559 :
2560 245176 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2561 : }
2562 :
2563 : /* signal that we need to wakeup walsenders later */
2564 245176 : WalSndWakeupRequest();
2565 :
2566 245176 : LogwrtResult.Flush = LogwrtResult.Write;
2567 : }
2568 :
2569 : /*
2570 : * Update shared-memory status
2571 : *
2572 : * We make sure that the shared 'request' values do not fall behind the
2573 : * 'result' values. This is not absolutely essential, but it saves some
2574 : * code in a couple of places.
2575 : */
2576 3815952 : SpinLockAcquire(&XLogCtl->info_lck);
2577 3815952 : if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
2578 218110 : XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
2579 3815952 : if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
2580 247352 : XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
2581 3815952 : SpinLockRelease(&XLogCtl->info_lck);
2582 :
2583 : /*
2584 : * We write Write first, bar, then Flush. When reading, the opposite must
2585 : * be done (with a matching barrier in between), so that we always see a
2586 : * Flush value that trails behind the Write value seen.
2587 : */
2588 3815952 : pg_atomic_write_u64(&XLogCtl->logWriteResult, LogwrtResult.Write);
2589 3815952 : pg_write_barrier();
2590 3815952 : pg_atomic_write_u64(&XLogCtl->logFlushResult, LogwrtResult.Flush);
2591 :
2592 : #ifdef USE_ASSERT_CHECKING
2593 : {
2594 : XLogRecPtr Flush;
2595 : XLogRecPtr Write;
2596 : XLogRecPtr Insert;
2597 :
2598 : Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult);
2599 : pg_read_barrier();
2600 : Write = pg_atomic_read_u64(&XLogCtl->logWriteResult);
2601 : pg_read_barrier();
2602 : Insert = pg_atomic_read_u64(&XLogCtl->logInsertResult);
2603 :
2604 : /* WAL written to disk is always ahead of WAL flushed */
2605 : Assert(Write >= Flush);
2606 :
2607 : /* WAL inserted to buffers is always ahead of WAL written */
2608 : Assert(Insert >= Write);
2609 : }
2610 : #endif
2611 3815952 : }
2612 :
2613 : /*
2614 : * Record the LSN for an asynchronous transaction commit/abort
2615 : * and nudge the WALWriter if there is work for it to do.
2616 : * (This should not be called for synchronous commits.)
2617 : */
2618 : void
2619 92100 : XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2620 : {
2621 92100 : XLogRecPtr WriteRqstPtr = asyncXactLSN;
2622 : bool sleeping;
2623 92100 : bool wakeup = false;
2624 : XLogRecPtr prevAsyncXactLSN;
2625 :
2626 92100 : SpinLockAcquire(&XLogCtl->info_lck);
2627 92100 : sleeping = XLogCtl->WalWriterSleeping;
2628 92100 : prevAsyncXactLSN = XLogCtl->asyncXactLSN;
2629 92100 : if (XLogCtl->asyncXactLSN < asyncXactLSN)
2630 91246 : XLogCtl->asyncXactLSN = asyncXactLSN;
2631 92100 : SpinLockRelease(&XLogCtl->info_lck);
2632 :
2633 : /*
2634 : * If somebody else already called this function with a more aggressive
2635 : * LSN, they will have done what we needed (and perhaps more).
2636 : */
2637 92100 : if (asyncXactLSN <= prevAsyncXactLSN)
2638 854 : return;
2639 :
2640 : /*
2641 : * If the WALWriter is sleeping, kick it to make it come out of low-power
2642 : * mode, so that this async commit will reach disk within the expected
2643 : * amount of time. Otherwise, determine whether it has enough WAL
2644 : * available to flush, the same way that XLogBackgroundFlush() does.
2645 : */
2646 91246 : if (sleeping)
2647 34 : wakeup = true;
2648 : else
2649 : {
2650 : int flushblocks;
2651 :
2652 91212 : RefreshXLogWriteResult(LogwrtResult);
2653 :
2654 91212 : flushblocks =
2655 91212 : WriteRqstPtr / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
2656 :
2657 91212 : if (WalWriterFlushAfter == 0 || flushblocks >= WalWriterFlushAfter)
2658 6858 : wakeup = true;
2659 : }
2660 :
2661 91246 : if (wakeup)
2662 : {
2663 6892 : volatile PROC_HDR *procglobal = ProcGlobal;
2664 6892 : ProcNumber walwriterProc = procglobal->walwriterProc;
2665 :
2666 6892 : if (walwriterProc != INVALID_PROC_NUMBER)
2667 432 : SetLatch(&GetPGProcByNumber(walwriterProc)->procLatch);
2668 : }
2669 : }
2670 :
2671 : /*
2672 : * Record the LSN up to which we can remove WAL because it's not required by
2673 : * any replication slot.
2674 : */
2675 : void
2676 23772 : XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn)
2677 : {
2678 23772 : SpinLockAcquire(&XLogCtl->info_lck);
2679 23772 : XLogCtl->replicationSlotMinLSN = lsn;
2680 23772 : SpinLockRelease(&XLogCtl->info_lck);
2681 23772 : }
2682 :
2683 :
2684 : /*
2685 : * Return the oldest LSN we must retain to satisfy the needs of some
2686 : * replication slot.
2687 : */
2688 : static XLogRecPtr
2689 3448 : XLogGetReplicationSlotMinimumLSN(void)
2690 : {
2691 : XLogRecPtr retval;
2692 :
2693 3448 : SpinLockAcquire(&XLogCtl->info_lck);
2694 3448 : retval = XLogCtl->replicationSlotMinLSN;
2695 3448 : SpinLockRelease(&XLogCtl->info_lck);
2696 :
2697 3448 : return retval;
2698 : }
2699 :
2700 : /*
2701 : * Advance minRecoveryPoint in control file.
2702 : *
2703 : * If we crash during recovery, we must reach this point again before the
2704 : * database is consistent.
2705 : *
2706 : * If 'force' is true, 'lsn' argument is ignored. Otherwise, minRecoveryPoint
2707 : * is only updated if it's not already greater than or equal to 'lsn'.
2708 : */
2709 : static void
2710 191150 : UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2711 : {
2712 : /* Quick check using our local copy of the variable */
2713 191150 : if (!updateMinRecoveryPoint || (!force && lsn <= LocalMinRecoveryPoint))
2714 179158 : return;
2715 :
2716 : /*
2717 : * An invalid minRecoveryPoint means that we need to recover all the WAL,
2718 : * i.e., we're doing crash recovery. We never modify the control file's
2719 : * value in that case, so we can short-circuit future checks here too. The
2720 : * local values of minRecoveryPoint and minRecoveryPointTLI should not be
2721 : * updated until crash recovery finishes. We only do this for the startup
2722 : * process as it should not update its own reference of minRecoveryPoint
2723 : * until it has finished crash recovery to make sure that all WAL
2724 : * available is replayed in this case. This also saves from extra locks
2725 : * taken on the control file from the startup process.
2726 : */
2727 11992 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint) && InRecovery)
2728 : {
2729 62 : updateMinRecoveryPoint = false;
2730 62 : return;
2731 : }
2732 :
2733 11930 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2734 :
2735 : /* update local copy */
2736 11930 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
2737 11930 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2738 :
2739 11930 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint))
2740 2 : updateMinRecoveryPoint = false;
2741 11928 : else if (force || LocalMinRecoveryPoint < lsn)
2742 : {
2743 : XLogRecPtr newMinRecoveryPoint;
2744 : TimeLineID newMinRecoveryPointTLI;
2745 :
2746 : /*
2747 : * To avoid having to update the control file too often, we update it
2748 : * all the way to the last record being replayed, even though 'lsn'
2749 : * would suffice for correctness. This also allows the 'force' case
2750 : * to not need a valid 'lsn' value.
2751 : *
2752 : * Another important reason for doing it this way is that the passed
2753 : * 'lsn' value could be bogus, i.e., past the end of available WAL, if
2754 : * the caller got it from a corrupted heap page. Accepting such a
2755 : * value as the min recovery point would prevent us from coming up at
2756 : * all. Instead, we just log a warning and continue with recovery.
2757 : * (See also the comments about corrupt LSNs in XLogFlush.)
2758 : */
2759 9742 : newMinRecoveryPoint = GetCurrentReplayRecPtr(&newMinRecoveryPointTLI);
2760 9742 : if (!force && newMinRecoveryPoint < lsn)
2761 0 : elog(WARNING,
2762 : "xlog min recovery request %X/%X is past current point %X/%X",
2763 : LSN_FORMAT_ARGS(lsn), LSN_FORMAT_ARGS(newMinRecoveryPoint));
2764 :
2765 : /* update control file */
2766 9742 : if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2767 : {
2768 9682 : ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2769 9682 : ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2770 9682 : UpdateControlFile();
2771 9682 : LocalMinRecoveryPoint = newMinRecoveryPoint;
2772 9682 : LocalMinRecoveryPointTLI = newMinRecoveryPointTLI;
2773 :
2774 9682 : ereport(DEBUG2,
2775 : (errmsg_internal("updated min recovery point to %X/%X on timeline %u",
2776 : LSN_FORMAT_ARGS(newMinRecoveryPoint),
2777 : newMinRecoveryPointTLI)));
2778 : }
2779 : }
2780 11930 : LWLockRelease(ControlFileLock);
2781 : }
2782 :
2783 : /*
2784 : * Ensure that all XLOG data through the given position is flushed to disk.
2785 : *
2786 : * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
2787 : * already held, and we try to avoid acquiring it if possible.
2788 : */
2789 : void
2790 1252846 : XLogFlush(XLogRecPtr record)
2791 : {
2792 : XLogRecPtr WriteRqstPtr;
2793 : XLogwrtRqst WriteRqst;
2794 1252846 : TimeLineID insertTLI = XLogCtl->InsertTimeLineID;
2795 :
2796 : /*
2797 : * During REDO, we are reading not writing WAL. Therefore, instead of
2798 : * trying to flush the WAL, we should update minRecoveryPoint instead. We
2799 : * test XLogInsertAllowed(), not InRecovery, because we need checkpointer
2800 : * to act this way too, and because when it tries to write the
2801 : * end-of-recovery checkpoint, it should indeed flush.
2802 : */
2803 1252846 : if (!XLogInsertAllowed())
2804 : {
2805 190904 : UpdateMinRecoveryPoint(record, false);
2806 993000 : return;
2807 : }
2808 :
2809 : /* Quick exit if already known flushed */
2810 1061942 : if (record <= LogwrtResult.Flush)
2811 802096 : return;
2812 :
2813 : #ifdef WAL_DEBUG
2814 : if (XLOG_DEBUG)
2815 : elog(LOG, "xlog flush request %X/%X; write %X/%X; flush %X/%X",
2816 : LSN_FORMAT_ARGS(record),
2817 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2818 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2819 : #endif
2820 :
2821 259846 : START_CRIT_SECTION();
2822 :
2823 : /*
2824 : * Since fsync is usually a horribly expensive operation, we try to
2825 : * piggyback as much data as we can on each fsync: if we see any more data
2826 : * entered into the xlog buffer, we'll write and fsync that too, so that
2827 : * the final value of LogwrtResult.Flush is as large as possible. This
2828 : * gives us some chance of avoiding another fsync immediately after.
2829 : */
2830 :
2831 : /* initialize to given target; may increase below */
2832 259846 : WriteRqstPtr = record;
2833 :
2834 : /*
2835 : * Now wait until we get the write lock, or someone else does the flush
2836 : * for us.
2837 : */
2838 : for (;;)
2839 10918 : {
2840 : XLogRecPtr insertpos;
2841 :
2842 : /* done already? */
2843 270764 : RefreshXLogWriteResult(LogwrtResult);
2844 270764 : if (record <= LogwrtResult.Flush)
2845 19344 : break;
2846 :
2847 : /*
2848 : * Before actually performing the write, wait for all in-flight
2849 : * insertions to the pages we're about to write to finish.
2850 : */
2851 251420 : SpinLockAcquire(&XLogCtl->info_lck);
2852 251420 : if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
2853 15384 : WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
2854 251420 : SpinLockRelease(&XLogCtl->info_lck);
2855 251420 : insertpos = WaitXLogInsertionsToFinish(WriteRqstPtr);
2856 :
2857 : /*
2858 : * Try to get the write lock. If we can't get it immediately, wait
2859 : * until it's released, and recheck if we still need to do the flush
2860 : * or if the backend that held the lock did it for us already. This
2861 : * helps to maintain a good rate of group committing when the system
2862 : * is bottlenecked by the speed of fsyncing.
2863 : */
2864 251420 : if (!LWLockAcquireOrWait(WALWriteLock, LW_EXCLUSIVE))
2865 : {
2866 : /*
2867 : * The lock is now free, but we didn't acquire it yet. Before we
2868 : * do, loop back to check if someone else flushed the record for
2869 : * us already.
2870 : */
2871 10918 : continue;
2872 : }
2873 :
2874 : /* Got the lock; recheck whether request is satisfied */
2875 240502 : RefreshXLogWriteResult(LogwrtResult);
2876 240502 : if (record <= LogwrtResult.Flush)
2877 : {
2878 1720 : LWLockRelease(WALWriteLock);
2879 1720 : break;
2880 : }
2881 :
2882 : /*
2883 : * Sleep before flush! By adding a delay here, we may give further
2884 : * backends the opportunity to join the backlog of group commit
2885 : * followers; this can significantly improve transaction throughput,
2886 : * at the risk of increasing transaction latency.
2887 : *
2888 : * We do not sleep if enableFsync is not turned on, nor if there are
2889 : * fewer than CommitSiblings other backends with active transactions.
2890 : */
2891 238782 : if (CommitDelay > 0 && enableFsync &&
2892 0 : MinimumActiveBackends(CommitSiblings))
2893 : {
2894 0 : pg_usleep(CommitDelay);
2895 :
2896 : /*
2897 : * Re-check how far we can now flush the WAL. It's generally not
2898 : * safe to call WaitXLogInsertionsToFinish while holding
2899 : * WALWriteLock, because an in-progress insertion might need to
2900 : * also grab WALWriteLock to make progress. But we know that all
2901 : * the insertions up to insertpos have already finished, because
2902 : * that's what the earlier WaitXLogInsertionsToFinish() returned.
2903 : * We're only calling it again to allow insertpos to be moved
2904 : * further forward, not to actually wait for anyone.
2905 : */
2906 0 : insertpos = WaitXLogInsertionsToFinish(insertpos);
2907 : }
2908 :
2909 : /* try to write/flush later additions to XLOG as well */
2910 238782 : WriteRqst.Write = insertpos;
2911 238782 : WriteRqst.Flush = insertpos;
2912 :
2913 238782 : XLogWrite(WriteRqst, insertTLI, false);
2914 :
2915 238782 : LWLockRelease(WALWriteLock);
2916 : /* done */
2917 238782 : break;
2918 : }
2919 :
2920 259846 : END_CRIT_SECTION();
2921 :
2922 : /* wake up walsenders now that we've released heavily contended locks */
2923 259846 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
2924 :
2925 : /*
2926 : * If we still haven't flushed to the request point then we have a
2927 : * problem; most likely, the requested flush point is past end of XLOG.
2928 : * This has been seen to occur when a disk page has a corrupted LSN.
2929 : *
2930 : * Formerly we treated this as a PANIC condition, but that hurts the
2931 : * system's robustness rather than helping it: we do not want to take down
2932 : * the whole system due to corruption on one data page. In particular, if
2933 : * the bad page is encountered again during recovery then we would be
2934 : * unable to restart the database at all! (This scenario actually
2935 : * happened in the field several times with 7.1 releases.) As of 8.4, bad
2936 : * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
2937 : * the only time we can reach here during recovery is while flushing the
2938 : * end-of-recovery checkpoint record, and we don't expect that to have a
2939 : * bad LSN.
2940 : *
2941 : * Note that for calls from xact.c, the ERROR will be promoted to PANIC
2942 : * since xact.c calls this routine inside a critical section. However,
2943 : * calls from bufmgr.c are not within critical sections and so we will not
2944 : * force a restart for a bad LSN on a data page.
2945 : */
2946 259846 : if (LogwrtResult.Flush < record)
2947 0 : elog(ERROR,
2948 : "xlog flush request %X/%X is not satisfied --- flushed only to %X/%X",
2949 : LSN_FORMAT_ARGS(record),
2950 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
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 21662 : XLogBackgroundFlush(void)
2979 : {
2980 : XLogwrtRqst WriteRqst;
2981 21662 : 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 21662 : if (RecoveryInProgress())
2989 16 : 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 21646 : insertTLI = XLogCtl->InsertTimeLineID;
2996 :
2997 : /* read updated LogwrtRqst */
2998 21646 : SpinLockAcquire(&XLogCtl->info_lck);
2999 21646 : WriteRqst = XLogCtl->LogwrtRqst;
3000 21646 : SpinLockRelease(&XLogCtl->info_lck);
3001 :
3002 : /* back off to last completed page boundary */
3003 21646 : WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
3004 :
3005 : /* if we have already flushed that far, consider async commit records */
3006 21646 : RefreshXLogWriteResult(LogwrtResult);
3007 21646 : if (WriteRqst.Write <= LogwrtResult.Flush)
3008 : {
3009 15282 : SpinLockAcquire(&XLogCtl->info_lck);
3010 15282 : WriteRqst.Write = XLogCtl->asyncXactLSN;
3011 15282 : SpinLockRelease(&XLogCtl->info_lck);
3012 15282 : 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 21646 : if (WriteRqst.Write <= LogwrtResult.Flush)
3021 : {
3022 14268 : if (openLogFile >= 0)
3023 : {
3024 6972 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
3025 : wal_segment_size))
3026 : {
3027 256 : XLogFileClose();
3028 : }
3029 : }
3030 14268 : 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 7378 : now = GetCurrentTimestamp();
3042 7378 : flushblocks =
3043 7378 : WriteRqst.Write / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
3044 :
3045 7378 : if (WalWriterFlushAfter == 0 || lastflush == 0)
3046 : {
3047 : /* first call, or block based limits disabled */
3048 410 : WriteRqst.Flush = WriteRqst.Write;
3049 410 : lastflush = now;
3050 : }
3051 6968 : 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 6640 : WriteRqst.Flush = WriteRqst.Write;
3059 6640 : lastflush = now;
3060 : }
3061 328 : else if (flushblocks >= WalWriterFlushAfter)
3062 : {
3063 : /* exceeded wal_writer_flush_after blocks, flush */
3064 306 : WriteRqst.Flush = WriteRqst.Write;
3065 306 : lastflush = now;
3066 : }
3067 : else
3068 : {
3069 : /* no flushing, this time round */
3070 22 : WriteRqst.Flush = 0;
3071 : }
3072 :
3073 : #ifdef WAL_DEBUG
3074 : if (XLOG_DEBUG)
3075 : elog(LOG, "xlog bg flush request write %X/%X; flush: %X/%X, current is write %X/%X; flush %X/%X",
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 7378 : START_CRIT_SECTION();
3083 :
3084 : /* now wait for any in-progress insertions to finish and get write lock */
3085 7378 : WaitXLogInsertionsToFinish(WriteRqst.Write);
3086 7378 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3087 7378 : RefreshXLogWriteResult(LogwrtResult);
3088 7378 : if (WriteRqst.Write > LogwrtResult.Write ||
3089 264 : WriteRqst.Flush > LogwrtResult.Flush)
3090 : {
3091 7158 : XLogWrite(WriteRqst, insertTLI, flexible);
3092 : }
3093 7378 : LWLockRelease(WALWriteLock);
3094 :
3095 7378 : END_CRIT_SECTION();
3096 :
3097 : /* wake up walsenders now that we've released heavily contended locks */
3098 7378 : 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 7378 : 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 7378 : return true;
3112 : }
3113 :
3114 : /*
3115 : * Test whether XLOG data has been flushed up to (at least) the given position.
3116 : *
3117 : * Returns true if a flush is still needed. (It may be that someone else
3118 : * is already in process of flushing that far, however.)
3119 : */
3120 : bool
3121 17474454 : XLogNeedsFlush(XLogRecPtr record)
3122 : {
3123 : /*
3124 : * During recovery, we don't flush WAL but update minRecoveryPoint
3125 : * instead. So "needs flush" is taken to mean whether minRecoveryPoint
3126 : * would need to be updated.
3127 : */
3128 17474454 : if (RecoveryInProgress())
3129 : {
3130 : /*
3131 : * An invalid minRecoveryPoint means that we need to recover all the
3132 : * WAL, i.e., we're doing crash recovery. We never modify the control
3133 : * file's value in that case, so we can short-circuit future checks
3134 : * here too. This triggers a quick exit path for the startup process,
3135 : * which cannot update its local copy of minRecoveryPoint as long as
3136 : * it has not replayed all WAL available when doing crash recovery.
3137 : */
3138 1653736 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint) && InRecovery)
3139 0 : updateMinRecoveryPoint = false;
3140 :
3141 : /* Quick exit if already known to be updated or cannot be updated */
3142 1653736 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3143 1625726 : return false;
3144 :
3145 : /*
3146 : * Update local copy of minRecoveryPoint. But if the lock is busy,
3147 : * just return a conservative guess.
3148 : */
3149 28010 : if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3150 0 : return true;
3151 28010 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
3152 28010 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3153 28010 : LWLockRelease(ControlFileLock);
3154 :
3155 : /*
3156 : * Check minRecoveryPoint for any other process than the startup
3157 : * process doing crash recovery, which should not update the control
3158 : * file value if crash recovery is still running.
3159 : */
3160 28010 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint))
3161 0 : updateMinRecoveryPoint = false;
3162 :
3163 : /* check again */
3164 28010 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3165 162 : return false;
3166 : else
3167 27848 : return true;
3168 : }
3169 :
3170 : /* Quick exit if already known flushed */
3171 15820718 : if (record <= LogwrtResult.Flush)
3172 15315924 : return false;
3173 :
3174 : /* read LogwrtResult and update local state */
3175 504794 : RefreshXLogWriteResult(LogwrtResult);
3176 :
3177 : /* check again */
3178 504794 : if (record <= LogwrtResult.Flush)
3179 5066 : return false;
3180 :
3181 499728 : return true;
3182 : }
3183 :
3184 : /*
3185 : * Try to make a given XLOG file segment exist.
3186 : *
3187 : * logsegno: identify segment.
3188 : *
3189 : * *added: on return, true if this call raised the number of extant segments.
3190 : *
3191 : * path: on return, this char[MAXPGPATH] has the path to the logsegno file.
3192 : *
3193 : * Returns -1 or FD of opened file. A -1 here is not an error; a caller
3194 : * wanting an open segment should attempt to open "path", which usually will
3195 : * succeed. (This is weird, but it's efficient for the callers.)
3196 : */
3197 : static int
3198 25330 : XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,
3199 : bool *added, char *path)
3200 : {
3201 : char tmppath[MAXPGPATH];
3202 : XLogSegNo installed_segno;
3203 : XLogSegNo max_segno;
3204 : int fd;
3205 : int save_errno;
3206 25330 : int open_flags = O_RDWR | O_CREAT | O_EXCL | PG_BINARY;
3207 : instr_time io_start;
3208 :
3209 : Assert(logtli != 0);
3210 :
3211 25330 : XLogFilePath(path, logtli, logsegno, wal_segment_size);
3212 :
3213 : /*
3214 : * Try to use existent file (checkpoint maker may have created it already)
3215 : */
3216 25330 : *added = false;
3217 25330 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3218 25330 : get_sync_bit(wal_sync_method));
3219 25330 : if (fd < 0)
3220 : {
3221 2592 : if (errno != ENOENT)
3222 0 : ereport(ERROR,
3223 : (errcode_for_file_access(),
3224 : errmsg("could not open file \"%s\": %m", path)));
3225 : }
3226 : else
3227 22738 : return fd;
3228 :
3229 : /*
3230 : * Initialize an empty (all zeroes) segment. NOTE: it is possible that
3231 : * another process is doing the same thing. If so, we will end up
3232 : * pre-creating an extra log segment. That seems OK, and better than
3233 : * holding the lock throughout this lengthy process.
3234 : */
3235 2592 : elog(DEBUG2, "creating and filling new WAL file");
3236 :
3237 2592 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3238 :
3239 2592 : unlink(tmppath);
3240 :
3241 2592 : if (io_direct_flags & IO_DIRECT_WAL_INIT)
3242 0 : open_flags |= PG_O_DIRECT;
3243 :
3244 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3245 2592 : fd = BasicOpenFile(tmppath, open_flags);
3246 2592 : if (fd < 0)
3247 0 : ereport(ERROR,
3248 : (errcode_for_file_access(),
3249 : errmsg("could not create file \"%s\": %m", tmppath)));
3250 :
3251 : /* Measure I/O timing when initializing segment */
3252 2592 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3253 :
3254 2592 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE);
3255 2592 : save_errno = 0;
3256 2592 : if (wal_init_zero)
3257 : {
3258 : ssize_t rc;
3259 :
3260 : /*
3261 : * Zero-fill the file. With this setting, we do this the hard way to
3262 : * ensure that all the file space has really been allocated. On
3263 : * platforms that allow "holes" in files, just seeking to the end
3264 : * doesn't allocate intermediate space. This way, we know that we
3265 : * have all the space and (after the fsync below) that all the
3266 : * indirect blocks are down on disk. Therefore, fdatasync(2) or
3267 : * O_DSYNC will be sufficient to sync future writes to the log file.
3268 : */
3269 2592 : rc = pg_pwrite_zeros(fd, wal_segment_size, 0);
3270 :
3271 2592 : if (rc < 0)
3272 0 : save_errno = errno;
3273 : }
3274 : else
3275 : {
3276 : /*
3277 : * Otherwise, seeking to the end and writing a solitary byte is
3278 : * enough.
3279 : */
3280 0 : errno = 0;
3281 0 : if (pg_pwrite(fd, "\0", 1, wal_segment_size - 1) != 1)
3282 : {
3283 : /* if write didn't set errno, assume no disk space */
3284 0 : save_errno = errno ? errno : ENOSPC;
3285 : }
3286 : }
3287 2592 : pgstat_report_wait_end();
3288 :
3289 : /*
3290 : * A full segment worth of data is written when using wal_init_zero. One
3291 : * byte is written when not using it.
3292 : */
3293 2592 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT, IOOP_WRITE,
3294 : io_start, 1,
3295 2592 : wal_init_zero ? wal_segment_size : 1);
3296 :
3297 2592 : if (save_errno)
3298 : {
3299 : /*
3300 : * If we fail to make the file, delete it to release disk space
3301 : */
3302 0 : unlink(tmppath);
3303 :
3304 0 : close(fd);
3305 :
3306 0 : errno = save_errno;
3307 :
3308 0 : ereport(ERROR,
3309 : (errcode_for_file_access(),
3310 : errmsg("could not write to file \"%s\": %m", tmppath)));
3311 : }
3312 :
3313 : /* Measure I/O timing when flushing segment */
3314 2592 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3315 :
3316 2592 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC);
3317 2592 : if (pg_fsync(fd) != 0)
3318 : {
3319 0 : save_errno = errno;
3320 0 : close(fd);
3321 0 : errno = save_errno;
3322 0 : ereport(ERROR,
3323 : (errcode_for_file_access(),
3324 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3325 : }
3326 2592 : pgstat_report_wait_end();
3327 :
3328 2592 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT,
3329 : IOOP_FSYNC, io_start, 1, 0);
3330 :
3331 2592 : if (close(fd) != 0)
3332 0 : ereport(ERROR,
3333 : (errcode_for_file_access(),
3334 : errmsg("could not close file \"%s\": %m", tmppath)));
3335 :
3336 : /*
3337 : * Now move the segment into place with its final name. Cope with
3338 : * possibility that someone else has created the file while we were
3339 : * filling ours: if so, use ours to pre-create a future log segment.
3340 : */
3341 2592 : installed_segno = logsegno;
3342 :
3343 : /*
3344 : * XXX: What should we use as max_segno? We used to use XLOGfileslop when
3345 : * that was a constant, but that was always a bit dubious: normally, at a
3346 : * checkpoint, XLOGfileslop was the offset from the checkpoint record, but
3347 : * here, it was the offset from the insert location. We can't do the
3348 : * normal XLOGfileslop calculation here because we don't have access to
3349 : * the prior checkpoint's redo location. So somewhat arbitrarily, just use
3350 : * CheckPointSegments.
3351 : */
3352 2592 : max_segno = logsegno + CheckPointSegments;
3353 2592 : if (InstallXLogFileSegment(&installed_segno, tmppath, true, max_segno,
3354 : logtli))
3355 : {
3356 2592 : *added = true;
3357 2592 : elog(DEBUG2, "done creating and filling new WAL file");
3358 : }
3359 : else
3360 : {
3361 : /*
3362 : * No need for any more future segments, or InstallXLogFileSegment()
3363 : * failed to rename the file into place. If the rename failed, a
3364 : * caller opening the file may fail.
3365 : */
3366 0 : unlink(tmppath);
3367 0 : elog(DEBUG2, "abandoned new WAL file");
3368 : }
3369 :
3370 2592 : return -1;
3371 : }
3372 :
3373 : /*
3374 : * Create a new XLOG file segment, or open a pre-existing one.
3375 : *
3376 : * logsegno: identify segment to be created/opened.
3377 : *
3378 : * Returns FD of opened file.
3379 : *
3380 : * Note: errors here are ERROR not PANIC because we might or might not be
3381 : * inside a critical section (eg, during checkpoint there is no reason to
3382 : * take down the system on failure). They will promote to PANIC if we are
3383 : * in a critical section.
3384 : */
3385 : int
3386 24834 : XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)
3387 : {
3388 : bool ignore_added;
3389 : char path[MAXPGPATH];
3390 : int fd;
3391 :
3392 : Assert(logtli != 0);
3393 :
3394 24834 : fd = XLogFileInitInternal(logsegno, logtli, &ignore_added, path);
3395 24834 : if (fd >= 0)
3396 22550 : return fd;
3397 :
3398 : /* Now open original target segment (might not be file I just made) */
3399 2284 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3400 2284 : get_sync_bit(wal_sync_method));
3401 2284 : if (fd < 0)
3402 0 : ereport(ERROR,
3403 : (errcode_for_file_access(),
3404 : errmsg("could not open file \"%s\": %m", path)));
3405 2284 : return fd;
3406 : }
3407 :
3408 : /*
3409 : * Create a new XLOG file segment by copying a pre-existing one.
3410 : *
3411 : * destsegno: identify segment to be created.
3412 : *
3413 : * srcTLI, srcsegno: identify segment to be copied (could be from
3414 : * a different timeline)
3415 : *
3416 : * upto: how much of the source file to copy (the rest is filled with
3417 : * zeros)
3418 : *
3419 : * Currently this is only used during recovery, and so there are no locking
3420 : * considerations. But we should be just as tense as XLogFileInit to avoid
3421 : * emplacing a bogus file.
3422 : */
3423 : static void
3424 76 : XLogFileCopy(TimeLineID destTLI, XLogSegNo destsegno,
3425 : TimeLineID srcTLI, XLogSegNo srcsegno,
3426 : int upto)
3427 : {
3428 : char path[MAXPGPATH];
3429 : char tmppath[MAXPGPATH];
3430 : PGAlignedXLogBlock buffer;
3431 : int srcfd;
3432 : int fd;
3433 : int nbytes;
3434 :
3435 : /*
3436 : * Open the source file
3437 : */
3438 76 : XLogFilePath(path, srcTLI, srcsegno, wal_segment_size);
3439 76 : srcfd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
3440 76 : if (srcfd < 0)
3441 0 : ereport(ERROR,
3442 : (errcode_for_file_access(),
3443 : errmsg("could not open file \"%s\": %m", path)));
3444 :
3445 : /*
3446 : * Copy into a temp file name.
3447 : */
3448 76 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3449 :
3450 76 : unlink(tmppath);
3451 :
3452 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3453 76 : fd = OpenTransientFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
3454 76 : if (fd < 0)
3455 0 : ereport(ERROR,
3456 : (errcode_for_file_access(),
3457 : errmsg("could not create file \"%s\": %m", tmppath)));
3458 :
3459 : /*
3460 : * Do the data copying.
3461 : */
3462 155724 : for (nbytes = 0; nbytes < wal_segment_size; nbytes += sizeof(buffer))
3463 : {
3464 : int nread;
3465 :
3466 155648 : nread = upto - nbytes;
3467 :
3468 : /*
3469 : * The part that is not read from the source file is filled with
3470 : * zeros.
3471 : */
3472 155648 : if (nread < sizeof(buffer))
3473 76 : memset(buffer.data, 0, sizeof(buffer));
3474 :
3475 155648 : if (nread > 0)
3476 : {
3477 : int r;
3478 :
3479 5356 : if (nread > sizeof(buffer))
3480 5280 : nread = sizeof(buffer);
3481 5356 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ);
3482 5356 : r = read(srcfd, buffer.data, nread);
3483 5356 : if (r != nread)
3484 : {
3485 0 : if (r < 0)
3486 0 : ereport(ERROR,
3487 : (errcode_for_file_access(),
3488 : errmsg("could not read file \"%s\": %m",
3489 : path)));
3490 : else
3491 0 : ereport(ERROR,
3492 : (errcode(ERRCODE_DATA_CORRUPTED),
3493 : errmsg("could not read file \"%s\": read %d of %zu",
3494 : path, r, (Size) nread)));
3495 : }
3496 5356 : pgstat_report_wait_end();
3497 : }
3498 155648 : errno = 0;
3499 155648 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_WRITE);
3500 155648 : if ((int) write(fd, buffer.data, sizeof(buffer)) != (int) sizeof(buffer))
3501 : {
3502 0 : int save_errno = errno;
3503 :
3504 : /*
3505 : * If we fail to make the file, delete it to release disk space
3506 : */
3507 0 : unlink(tmppath);
3508 : /* if write didn't set errno, assume problem is no disk space */
3509 0 : errno = save_errno ? save_errno : ENOSPC;
3510 :
3511 0 : ereport(ERROR,
3512 : (errcode_for_file_access(),
3513 : errmsg("could not write to file \"%s\": %m", tmppath)));
3514 : }
3515 155648 : pgstat_report_wait_end();
3516 : }
3517 :
3518 76 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_SYNC);
3519 76 : if (pg_fsync(fd) != 0)
3520 0 : ereport(data_sync_elevel(ERROR),
3521 : (errcode_for_file_access(),
3522 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3523 76 : pgstat_report_wait_end();
3524 :
3525 76 : if (CloseTransientFile(fd) != 0)
3526 0 : ereport(ERROR,
3527 : (errcode_for_file_access(),
3528 : errmsg("could not close file \"%s\": %m", tmppath)));
3529 :
3530 76 : if (CloseTransientFile(srcfd) != 0)
3531 0 : ereport(ERROR,
3532 : (errcode_for_file_access(),
3533 : errmsg("could not close file \"%s\": %m", path)));
3534 :
3535 : /*
3536 : * Now move the segment into place with its final name.
3537 : */
3538 76 : if (!InstallXLogFileSegment(&destsegno, tmppath, false, 0, destTLI))
3539 0 : elog(ERROR, "InstallXLogFileSegment should not have failed");
3540 76 : }
3541 :
3542 : /*
3543 : * Install a new XLOG segment file as a current or future log segment.
3544 : *
3545 : * This is used both to install a newly-created segment (which has a temp
3546 : * filename while it's being created) and to recycle an old segment.
3547 : *
3548 : * *segno: identify segment to install as (or first possible target).
3549 : * When find_free is true, this is modified on return to indicate the
3550 : * actual installation location or last segment searched.
3551 : *
3552 : * tmppath: initial name of file to install. It will be renamed into place.
3553 : *
3554 : * find_free: if true, install the new segment at the first empty segno
3555 : * number at or after the passed numbers. If false, install the new segment
3556 : * exactly where specified, deleting any existing segment file there.
3557 : *
3558 : * max_segno: maximum segment number to install the new file as. Fail if no
3559 : * free slot is found between *segno and max_segno. (Ignored when find_free
3560 : * is false.)
3561 : *
3562 : * tli: The timeline on which the new segment should be installed.
3563 : *
3564 : * Returns true if the file was installed successfully. false indicates that
3565 : * max_segno limit was exceeded, the startup process has disabled this
3566 : * function for now, or an error occurred while renaming the file into place.
3567 : */
3568 : static bool
3569 4852 : InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
3570 : bool find_free, XLogSegNo max_segno, TimeLineID tli)
3571 : {
3572 : char path[MAXPGPATH];
3573 : struct stat stat_buf;
3574 :
3575 : Assert(tli != 0);
3576 :
3577 4852 : XLogFilePath(path, tli, *segno, wal_segment_size);
3578 :
3579 4852 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3580 4852 : if (!XLogCtl->InstallXLogFileSegmentActive)
3581 : {
3582 0 : LWLockRelease(ControlFileLock);
3583 0 : return false;
3584 : }
3585 :
3586 4852 : if (!find_free)
3587 : {
3588 : /* Force installation: get rid of any pre-existing segment file */
3589 76 : durable_unlink(path, DEBUG1);
3590 : }
3591 : else
3592 : {
3593 : /* Find a free slot to put it in */
3594 7092 : while (stat(path, &stat_buf) == 0)
3595 : {
3596 2370 : if ((*segno) >= max_segno)
3597 : {
3598 : /* Failed to find a free slot within specified range */
3599 54 : LWLockRelease(ControlFileLock);
3600 54 : return false;
3601 : }
3602 2316 : (*segno)++;
3603 2316 : XLogFilePath(path, tli, *segno, wal_segment_size);
3604 : }
3605 : }
3606 :
3607 : Assert(access(path, F_OK) != 0 && errno == ENOENT);
3608 4798 : if (durable_rename(tmppath, path, LOG) != 0)
3609 : {
3610 0 : LWLockRelease(ControlFileLock);
3611 : /* durable_rename already emitted log message */
3612 0 : return false;
3613 : }
3614 :
3615 4798 : LWLockRelease(ControlFileLock);
3616 :
3617 4798 : return true;
3618 : }
3619 :
3620 : /*
3621 : * Open a pre-existing logfile segment for writing.
3622 : */
3623 : int
3624 58 : XLogFileOpen(XLogSegNo segno, TimeLineID tli)
3625 : {
3626 : char path[MAXPGPATH];
3627 : int fd;
3628 :
3629 58 : XLogFilePath(path, tli, segno, wal_segment_size);
3630 :
3631 58 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3632 58 : get_sync_bit(wal_sync_method));
3633 58 : if (fd < 0)
3634 0 : ereport(PANIC,
3635 : (errcode_for_file_access(),
3636 : errmsg("could not open file \"%s\": %m", path)));
3637 :
3638 58 : return fd;
3639 : }
3640 :
3641 : /*
3642 : * Close the current logfile segment for writing.
3643 : */
3644 : static void
3645 9730 : XLogFileClose(void)
3646 : {
3647 : Assert(openLogFile >= 0);
3648 :
3649 : /*
3650 : * WAL segment files will not be re-read in normal operation, so we advise
3651 : * the OS to release any cached pages. But do not do so if WAL archiving
3652 : * or streaming is active, because archiver and walsender process could
3653 : * use the cache to read the WAL segment.
3654 : */
3655 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3656 9730 : if (!XLogIsNeeded() && (io_direct_flags & IO_DIRECT_WAL) == 0)
3657 2318 : (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3658 : #endif
3659 :
3660 9730 : if (close(openLogFile) != 0)
3661 : {
3662 : char xlogfname[MAXFNAMELEN];
3663 0 : int save_errno = errno;
3664 :
3665 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo, wal_segment_size);
3666 0 : errno = save_errno;
3667 0 : ereport(PANIC,
3668 : (errcode_for_file_access(),
3669 : errmsg("could not close file \"%s\": %m", xlogfname)));
3670 : }
3671 :
3672 9730 : openLogFile = -1;
3673 9730 : ReleaseExternalFD();
3674 9730 : }
3675 :
3676 : /*
3677 : * Preallocate log files beyond the specified log endpoint.
3678 : *
3679 : * XXX this is currently extremely conservative, since it forces only one
3680 : * future log segment to exist, and even that only if we are 75% done with
3681 : * the current one. This is only appropriate for very low-WAL-volume systems.
3682 : * High-volume systems will be OK once they've built up a sufficient set of
3683 : * recycled log segments, but the startup transient is likely to include
3684 : * a lot of segment creations by foreground processes, which is not so good.
3685 : *
3686 : * XLogFileInitInternal() can ereport(ERROR). All known causes indicate big
3687 : * trouble; for example, a full filesystem is one cause. The checkpoint WAL
3688 : * and/or ControlFile updates already completed. If a RequestCheckpoint()
3689 : * initiated the present checkpoint and an ERROR ends this function, the
3690 : * command that called RequestCheckpoint() fails. That's not ideal, but it's
3691 : * not worth contorting more functions to use caller-specified elevel values.
3692 : * (With or without RequestCheckpoint(), an ERROR forestalls some inessential
3693 : * reporting and resource reclamation.)
3694 : */
3695 : static void
3696 3124 : PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli)
3697 : {
3698 : XLogSegNo _logSegNo;
3699 : int lf;
3700 : bool added;
3701 : char path[MAXPGPATH];
3702 : uint64 offset;
3703 :
3704 3124 : if (!XLogCtl->InstallXLogFileSegmentActive)
3705 16 : return; /* unlocked check says no */
3706 :
3707 3108 : XLByteToPrevSeg(endptr, _logSegNo, wal_segment_size);
3708 3108 : offset = XLogSegmentOffset(endptr - 1, wal_segment_size);
3709 3108 : if (offset >= (uint32) (0.75 * wal_segment_size))
3710 : {
3711 496 : _logSegNo++;
3712 496 : lf = XLogFileInitInternal(_logSegNo, tli, &added, path);
3713 496 : if (lf >= 0)
3714 188 : close(lf);
3715 496 : if (added)
3716 308 : CheckpointStats.ckpt_segs_added++;
3717 : }
3718 : }
3719 :
3720 : /*
3721 : * Throws an error if the given log segment has already been removed or
3722 : * recycled. The caller should only pass a segment that it knows to have
3723 : * existed while the server has been running, as this function always
3724 : * succeeds if no WAL segments have been removed since startup.
3725 : * 'tli' is only used in the error message.
3726 : *
3727 : * Note: this function guarantees to keep errno unchanged on return.
3728 : * This supports callers that use this to possibly deliver a better
3729 : * error message about a missing file, while still being able to throw
3730 : * a normal file-access error afterwards, if this does return.
3731 : */
3732 : void
3733 229240 : CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3734 : {
3735 229240 : int save_errno = errno;
3736 : XLogSegNo lastRemovedSegNo;
3737 :
3738 229240 : SpinLockAcquire(&XLogCtl->info_lck);
3739 229240 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3740 229240 : SpinLockRelease(&XLogCtl->info_lck);
3741 :
3742 229240 : if (segno <= lastRemovedSegNo)
3743 : {
3744 : char filename[MAXFNAMELEN];
3745 :
3746 0 : XLogFileName(filename, tli, segno, wal_segment_size);
3747 0 : errno = save_errno;
3748 0 : ereport(ERROR,
3749 : (errcode_for_file_access(),
3750 : errmsg("requested WAL segment %s has already been removed",
3751 : filename)));
3752 : }
3753 229240 : errno = save_errno;
3754 229240 : }
3755 :
3756 : /*
3757 : * Return the last WAL segment removed, or 0 if no segment has been removed
3758 : * since startup.
3759 : *
3760 : * NB: the result can be out of date arbitrarily fast, the caller has to deal
3761 : * with that.
3762 : */
3763 : XLogSegNo
3764 2004 : XLogGetLastRemovedSegno(void)
3765 : {
3766 : XLogSegNo lastRemovedSegNo;
3767 :
3768 2004 : SpinLockAcquire(&XLogCtl->info_lck);
3769 2004 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3770 2004 : SpinLockRelease(&XLogCtl->info_lck);
3771 :
3772 2004 : return lastRemovedSegNo;
3773 : }
3774 :
3775 : /*
3776 : * Return the oldest WAL segment on the given TLI that still exists in
3777 : * XLOGDIR, or 0 if none.
3778 : */
3779 : XLogSegNo
3780 10 : XLogGetOldestSegno(TimeLineID tli)
3781 : {
3782 : DIR *xldir;
3783 : struct dirent *xlde;
3784 10 : XLogSegNo oldest_segno = 0;
3785 :
3786 10 : xldir = AllocateDir(XLOGDIR);
3787 66 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3788 : {
3789 : TimeLineID file_tli;
3790 : XLogSegNo file_segno;
3791 :
3792 : /* Ignore files that are not XLOG segments. */
3793 56 : if (!IsXLogFileName(xlde->d_name))
3794 40 : continue;
3795 :
3796 : /* Parse filename to get TLI and segno. */
3797 16 : XLogFromFileName(xlde->d_name, &file_tli, &file_segno,
3798 : wal_segment_size);
3799 :
3800 : /* Ignore anything that's not from the TLI of interest. */
3801 16 : if (tli != file_tli)
3802 0 : continue;
3803 :
3804 : /* If it's the oldest so far, update oldest_segno. */
3805 16 : if (oldest_segno == 0 || file_segno < oldest_segno)
3806 14 : oldest_segno = file_segno;
3807 : }
3808 :
3809 10 : FreeDir(xldir);
3810 10 : return oldest_segno;
3811 : }
3812 :
3813 : /*
3814 : * Update the last removed segno pointer in shared memory, to reflect that the
3815 : * given XLOG file has been removed.
3816 : */
3817 : static void
3818 3900 : UpdateLastRemovedPtr(char *filename)
3819 : {
3820 : uint32 tli;
3821 : XLogSegNo segno;
3822 :
3823 3900 : XLogFromFileName(filename, &tli, &segno, wal_segment_size);
3824 :
3825 3900 : SpinLockAcquire(&XLogCtl->info_lck);
3826 3900 : if (segno > XLogCtl->lastRemovedSegNo)
3827 1730 : XLogCtl->lastRemovedSegNo = segno;
3828 3900 : SpinLockRelease(&XLogCtl->info_lck);
3829 3900 : }
3830 :
3831 : /*
3832 : * Remove all temporary log files in pg_wal
3833 : *
3834 : * This is called at the beginning of recovery after a previous crash,
3835 : * at a point where no other processes write fresh WAL data.
3836 : */
3837 : static void
3838 342 : RemoveTempXlogFiles(void)
3839 : {
3840 : DIR *xldir;
3841 : struct dirent *xlde;
3842 :
3843 342 : elog(DEBUG2, "removing all temporary WAL segments");
3844 :
3845 342 : xldir = AllocateDir(XLOGDIR);
3846 2208 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3847 : {
3848 : char path[MAXPGPATH];
3849 :
3850 1866 : if (strncmp(xlde->d_name, "xlogtemp.", 9) != 0)
3851 1866 : continue;
3852 :
3853 0 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
3854 0 : unlink(path);
3855 0 : elog(DEBUG2, "removed temporary WAL segment \"%s\"", path);
3856 : }
3857 342 : FreeDir(xldir);
3858 342 : }
3859 :
3860 : /*
3861 : * Recycle or remove all log files older or equal to passed segno.
3862 : *
3863 : * endptr is current (or recent) end of xlog, and lastredoptr is the
3864 : * redo pointer of the last checkpoint. These are used to determine
3865 : * whether we want to recycle rather than delete no-longer-wanted log files.
3866 : *
3867 : * insertTLI is the current timeline for XLOG insertion. Any recycled
3868 : * segments should be reused for this timeline.
3869 : */
3870 : static void
3871 2606 : RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr, XLogRecPtr endptr,
3872 : TimeLineID insertTLI)
3873 : {
3874 : DIR *xldir;
3875 : struct dirent *xlde;
3876 : char lastoff[MAXFNAMELEN];
3877 : XLogSegNo endlogSegNo;
3878 : XLogSegNo recycleSegNo;
3879 :
3880 : /* Initialize info about where to try to recycle to */
3881 2606 : XLByteToSeg(endptr, endlogSegNo, wal_segment_size);
3882 2606 : recycleSegNo = XLOGfileslop(lastredoptr);
3883 :
3884 : /*
3885 : * Construct a filename of the last segment to be kept. The timeline ID
3886 : * doesn't matter, we ignore that in the comparison. (During recovery,
3887 : * InsertTimeLineID isn't set, so we can't use that.)
3888 : */
3889 2606 : XLogFileName(lastoff, 0, segno, wal_segment_size);
3890 :
3891 2606 : elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
3892 : lastoff);
3893 :
3894 2606 : xldir = AllocateDir(XLOGDIR);
3895 :
3896 42868 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3897 : {
3898 : /* Ignore files that are not XLOG segments */
3899 40262 : if (!IsXLogFileName(xlde->d_name) &&
3900 10874 : !IsPartialXLogFileName(xlde->d_name))
3901 10870 : continue;
3902 :
3903 : /*
3904 : * We ignore the timeline part of the XLOG segment identifiers in
3905 : * deciding whether a segment is still needed. This ensures that we
3906 : * won't prematurely remove a segment from a parent timeline. We could
3907 : * probably be a little more proactive about removing segments of
3908 : * non-parent timelines, but that would be a whole lot more
3909 : * complicated.
3910 : *
3911 : * We use the alphanumeric sorting property of the filenames to decide
3912 : * which ones are earlier than the lastoff segment.
3913 : */
3914 29392 : if (strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
3915 : {
3916 3918 : if (XLogArchiveCheckDone(xlde->d_name))
3917 : {
3918 : /* Update the last removed location in shared memory first */
3919 3900 : UpdateLastRemovedPtr(xlde->d_name);
3920 :
3921 3900 : RemoveXlogFile(xlde, recycleSegNo, &endlogSegNo, insertTLI);
3922 : }
3923 : }
3924 : }
3925 :
3926 2606 : FreeDir(xldir);
3927 2606 : }
3928 :
3929 : /*
3930 : * Recycle or remove WAL files that are not part of the given timeline's
3931 : * history.
3932 : *
3933 : * This is called during recovery, whenever we switch to follow a new
3934 : * timeline, and at the end of recovery when we create a new timeline. We
3935 : * wouldn't otherwise care about extra WAL files lying in pg_wal, but they
3936 : * might be leftover pre-allocated or recycled WAL segments on the old timeline
3937 : * that we haven't used yet, and contain garbage. If we just leave them in
3938 : * pg_wal, they will eventually be archived, and we can't let that happen.
3939 : * Files that belong to our timeline history are valid, because we have
3940 : * successfully replayed them, but from others we can't be sure.
3941 : *
3942 : * 'switchpoint' is the current point in WAL where we switch to new timeline,
3943 : * and 'newTLI' is the new timeline we switch to.
3944 : */
3945 : void
3946 120 : RemoveNonParentXlogFiles(XLogRecPtr switchpoint, TimeLineID newTLI)
3947 : {
3948 : DIR *xldir;
3949 : struct dirent *xlde;
3950 : char switchseg[MAXFNAMELEN];
3951 : XLogSegNo endLogSegNo;
3952 : XLogSegNo switchLogSegNo;
3953 : XLogSegNo recycleSegNo;
3954 :
3955 : /*
3956 : * Initialize info about where to begin the work. This will recycle,
3957 : * somewhat arbitrarily, 10 future segments.
3958 : */
3959 120 : XLByteToPrevSeg(switchpoint, switchLogSegNo, wal_segment_size);
3960 120 : XLByteToSeg(switchpoint, endLogSegNo, wal_segment_size);
3961 120 : recycleSegNo = endLogSegNo + 10;
3962 :
3963 : /*
3964 : * Construct a filename of the last segment to be kept.
3965 : */
3966 120 : XLogFileName(switchseg, newTLI, switchLogSegNo, wal_segment_size);
3967 :
3968 120 : elog(DEBUG2, "attempting to remove WAL segments newer than log file %s",
3969 : switchseg);
3970 :
3971 120 : xldir = AllocateDir(XLOGDIR);
3972 :
3973 1140 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3974 : {
3975 : /* Ignore files that are not XLOG segments */
3976 1020 : if (!IsXLogFileName(xlde->d_name))
3977 636 : continue;
3978 :
3979 : /*
3980 : * Remove files that are on a timeline older than the new one we're
3981 : * switching to, but with a segment number >= the first segment on the
3982 : * new timeline.
3983 : */
3984 384 : if (strncmp(xlde->d_name, switchseg, 8) < 0 &&
3985 248 : strcmp(xlde->d_name + 8, switchseg + 8) > 0)
3986 : {
3987 : /*
3988 : * If the file has already been marked as .ready, however, don't
3989 : * remove it yet. It should be OK to remove it - files that are
3990 : * not part of our timeline history are not required for recovery
3991 : * - but seems safer to let them be archived and removed later.
3992 : */
3993 32 : if (!XLogArchiveIsReady(xlde->d_name))
3994 32 : RemoveXlogFile(xlde, recycleSegNo, &endLogSegNo, newTLI);
3995 : }
3996 : }
3997 :
3998 120 : FreeDir(xldir);
3999 120 : }
4000 :
4001 : /*
4002 : * Recycle or remove a log file that's no longer needed.
4003 : *
4004 : * segment_de is the dirent structure of the segment to recycle or remove.
4005 : * recycleSegNo is the segment number to recycle up to. endlogSegNo is
4006 : * the segment number of the current (or recent) end of WAL.
4007 : *
4008 : * endlogSegNo gets incremented if the segment is recycled so as it is not
4009 : * checked again with future callers of this function.
4010 : *
4011 : * insertTLI is the current timeline for XLOG insertion. Any recycled segments
4012 : * should be used for this timeline.
4013 : */
4014 : static void
4015 3932 : RemoveXlogFile(const struct dirent *segment_de,
4016 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
4017 : TimeLineID insertTLI)
4018 : {
4019 : char path[MAXPGPATH];
4020 : #ifdef WIN32
4021 : char newpath[MAXPGPATH];
4022 : #endif
4023 3932 : const char *segname = segment_de->d_name;
4024 :
4025 3932 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", segname);
4026 :
4027 : /*
4028 : * Before deleting the file, see if it can be recycled as a future log
4029 : * segment. Only recycle normal files, because we don't want to recycle
4030 : * symbolic links pointing to a separate archive directory.
4031 : */
4032 3932 : if (wal_recycle &&
4033 3932 : *endlogSegNo <= recycleSegNo &&
4034 4382 : XLogCtl->InstallXLogFileSegmentActive && /* callee rechecks this */
4035 4368 : get_dirent_type(path, segment_de, false, DEBUG2) == PGFILETYPE_REG &&
4036 2184 : InstallXLogFileSegment(endlogSegNo, path,
4037 : true, recycleSegNo, insertTLI))
4038 : {
4039 2130 : ereport(DEBUG2,
4040 : (errmsg_internal("recycled write-ahead log file \"%s\"",
4041 : segname)));
4042 2130 : CheckpointStats.ckpt_segs_recycled++;
4043 : /* Needn't recheck that slot on future iterations */
4044 2130 : (*endlogSegNo)++;
4045 : }
4046 : else
4047 : {
4048 : /* No need for any more future segments, or recycling failed ... */
4049 : int rc;
4050 :
4051 1802 : ereport(DEBUG2,
4052 : (errmsg_internal("removing write-ahead log file \"%s\"",
4053 : segname)));
4054 :
4055 : #ifdef WIN32
4056 :
4057 : /*
4058 : * On Windows, if another process (e.g another backend) holds the file
4059 : * open in FILE_SHARE_DELETE mode, unlink will succeed, but the file
4060 : * will still show up in directory listing until the last handle is
4061 : * closed. To avoid confusing the lingering deleted file for a live
4062 : * WAL file that needs to be archived, rename it before deleting it.
4063 : *
4064 : * If another process holds the file open without FILE_SHARE_DELETE
4065 : * flag, rename will fail. We'll try again at the next checkpoint.
4066 : */
4067 : snprintf(newpath, MAXPGPATH, "%s.deleted", path);
4068 : if (rename(path, newpath) != 0)
4069 : {
4070 : ereport(LOG,
4071 : (errcode_for_file_access(),
4072 : errmsg("could not rename file \"%s\": %m",
4073 : path)));
4074 : return;
4075 : }
4076 : rc = durable_unlink(newpath, LOG);
4077 : #else
4078 1802 : rc = durable_unlink(path, LOG);
4079 : #endif
4080 1802 : if (rc != 0)
4081 : {
4082 : /* Message already logged by durable_unlink() */
4083 0 : return;
4084 : }
4085 1802 : CheckpointStats.ckpt_segs_removed++;
4086 : }
4087 :
4088 3932 : XLogArchiveCleanup(segname);
4089 : }
4090 :
4091 : /*
4092 : * Verify whether pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
4093 : * If the latter do not exist, recreate them.
4094 : *
4095 : * It is not the goal of this function to verify the contents of these
4096 : * directories, but to help in cases where someone has performed a cluster
4097 : * copy for PITR purposes but omitted pg_wal from the copy.
4098 : *
4099 : * We could also recreate pg_wal if it doesn't exist, but a deliberate
4100 : * policy decision was made not to. It is fairly common for pg_wal to be
4101 : * a symlink, and if that was the DBA's intent then automatically making a
4102 : * plain directory would result in degraded performance with no notice.
4103 : */
4104 : static void
4105 1762 : ValidateXLOGDirectoryStructure(void)
4106 : {
4107 : char path[MAXPGPATH];
4108 : struct stat stat_buf;
4109 :
4110 : /* Check for pg_wal; if it doesn't exist, error out */
4111 1762 : if (stat(XLOGDIR, &stat_buf) != 0 ||
4112 1762 : !S_ISDIR(stat_buf.st_mode))
4113 0 : ereport(FATAL,
4114 : (errcode_for_file_access(),
4115 : errmsg("required WAL directory \"%s\" does not exist",
4116 : XLOGDIR)));
4117 :
4118 : /* Check for archive_status */
4119 1762 : snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4120 1762 : if (stat(path, &stat_buf) == 0)
4121 : {
4122 : /* Check for weird cases where it exists but isn't a directory */
4123 1760 : if (!S_ISDIR(stat_buf.st_mode))
4124 0 : ereport(FATAL,
4125 : (errcode_for_file_access(),
4126 : errmsg("required WAL directory \"%s\" does not exist",
4127 : path)));
4128 : }
4129 : else
4130 : {
4131 2 : ereport(LOG,
4132 : (errmsg("creating missing WAL directory \"%s\"", path)));
4133 2 : if (MakePGDirectory(path) < 0)
4134 0 : ereport(FATAL,
4135 : (errcode_for_file_access(),
4136 : errmsg("could not create missing directory \"%s\": %m",
4137 : path)));
4138 : }
4139 :
4140 : /* Check for summaries */
4141 1762 : snprintf(path, MAXPGPATH, XLOGDIR "/summaries");
4142 1762 : if (stat(path, &stat_buf) == 0)
4143 : {
4144 : /* Check for weird cases where it exists but isn't a directory */
4145 1760 : if (!S_ISDIR(stat_buf.st_mode))
4146 0 : ereport(FATAL,
4147 : (errmsg("required WAL directory \"%s\" does not exist",
4148 : path)));
4149 : }
4150 : else
4151 : {
4152 2 : ereport(LOG,
4153 : (errmsg("creating missing WAL directory \"%s\"", path)));
4154 2 : if (MakePGDirectory(path) < 0)
4155 0 : ereport(FATAL,
4156 : (errmsg("could not create missing directory \"%s\": %m",
4157 : path)));
4158 : }
4159 1762 : }
4160 :
4161 : /*
4162 : * Remove previous backup history files. This also retries creation of
4163 : * .ready files for any backup history files for which XLogArchiveNotify
4164 : * failed earlier.
4165 : */
4166 : static void
4167 296 : CleanupBackupHistory(void)
4168 : {
4169 : DIR *xldir;
4170 : struct dirent *xlde;
4171 : char path[MAXPGPATH + sizeof(XLOGDIR)];
4172 :
4173 296 : xldir = AllocateDir(XLOGDIR);
4174 :
4175 2672 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4176 : {
4177 2376 : if (IsBackupHistoryFileName(xlde->d_name))
4178 : {
4179 314 : if (XLogArchiveCheckDone(xlde->d_name))
4180 : {
4181 248 : elog(DEBUG2, "removing WAL backup history file \"%s\"",
4182 : xlde->d_name);
4183 248 : snprintf(path, sizeof(path), XLOGDIR "/%s", xlde->d_name);
4184 248 : unlink(path);
4185 248 : XLogArchiveCleanup(xlde->d_name);
4186 : }
4187 : }
4188 : }
4189 :
4190 296 : FreeDir(xldir);
4191 296 : }
4192 :
4193 : /*
4194 : * I/O routines for pg_control
4195 : *
4196 : * *ControlFile is a buffer in shared memory that holds an image of the
4197 : * contents of pg_control. WriteControlFile() initializes pg_control
4198 : * given a preloaded buffer, ReadControlFile() loads the buffer from
4199 : * the pg_control file (during postmaster or standalone-backend startup),
4200 : * and UpdateControlFile() rewrites pg_control after we modify xlog state.
4201 : * InitControlFile() fills the buffer with initial values.
4202 : *
4203 : * For simplicity, WriteControlFile() initializes the fields of pg_control
4204 : * that are related to checking backend/database compatibility, and
4205 : * ReadControlFile() verifies they are correct. We could split out the
4206 : * I/O and compatibility-check functions, but there seems no need currently.
4207 : */
4208 :
4209 : static void
4210 90 : InitControlFile(uint64 sysidentifier, uint32 data_checksum_version)
4211 : {
4212 : char mock_auth_nonce[MOCK_AUTH_NONCE_LEN];
4213 :
4214 : /*
4215 : * Generate a random nonce. This is used for authentication requests that
4216 : * will fail because the user does not exist. The nonce is used to create
4217 : * a genuine-looking password challenge for the non-existent user, in lieu
4218 : * of an actual stored password.
4219 : */
4220 90 : if (!pg_strong_random(mock_auth_nonce, MOCK_AUTH_NONCE_LEN))
4221 0 : ereport(PANIC,
4222 : (errcode(ERRCODE_INTERNAL_ERROR),
4223 : errmsg("could not generate secret authorization token")));
4224 :
4225 90 : memset(ControlFile, 0, sizeof(ControlFileData));
4226 : /* Initialize pg_control status fields */
4227 90 : ControlFile->system_identifier = sysidentifier;
4228 90 : memcpy(ControlFile->mock_authentication_nonce, mock_auth_nonce, MOCK_AUTH_NONCE_LEN);
4229 90 : ControlFile->state = DB_SHUTDOWNED;
4230 90 : ControlFile->unloggedLSN = FirstNormalUnloggedLSN;
4231 :
4232 : /* Set important parameter values for use when replaying WAL */
4233 90 : ControlFile->MaxConnections = MaxConnections;
4234 90 : ControlFile->max_worker_processes = max_worker_processes;
4235 90 : ControlFile->max_wal_senders = max_wal_senders;
4236 90 : ControlFile->max_prepared_xacts = max_prepared_xacts;
4237 90 : ControlFile->max_locks_per_xact = max_locks_per_xact;
4238 90 : ControlFile->wal_level = wal_level;
4239 90 : ControlFile->wal_log_hints = wal_log_hints;
4240 90 : ControlFile->track_commit_timestamp = track_commit_timestamp;
4241 90 : ControlFile->data_checksum_version = data_checksum_version;
4242 90 : }
4243 :
4244 : static void
4245 90 : WriteControlFile(void)
4246 : {
4247 : int fd;
4248 : char buffer[PG_CONTROL_FILE_SIZE]; /* need not be aligned */
4249 :
4250 : /*
4251 : * Initialize version and compatibility-check fields
4252 : */
4253 90 : ControlFile->pg_control_version = PG_CONTROL_VERSION;
4254 90 : ControlFile->catalog_version_no = CATALOG_VERSION_NO;
4255 :
4256 90 : ControlFile->maxAlign = MAXIMUM_ALIGNOF;
4257 90 : ControlFile->floatFormat = FLOATFORMAT_VALUE;
4258 :
4259 90 : ControlFile->blcksz = BLCKSZ;
4260 90 : ControlFile->relseg_size = RELSEG_SIZE;
4261 90 : ControlFile->xlog_blcksz = XLOG_BLCKSZ;
4262 90 : ControlFile->xlog_seg_size = wal_segment_size;
4263 :
4264 90 : ControlFile->nameDataLen = NAMEDATALEN;
4265 90 : ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
4266 :
4267 90 : ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
4268 90 : ControlFile->loblksize = LOBLKSIZE;
4269 :
4270 90 : ControlFile->float8ByVal = FLOAT8PASSBYVAL;
4271 :
4272 : /*
4273 : * Initialize the default 'char' signedness.
4274 : *
4275 : * The signedness of the char type is implementation-defined. For instance
4276 : * on x86 architecture CPUs, the char data type is typically treated as
4277 : * signed by default, whereas on aarch architecture CPUs, it is typically
4278 : * treated as unsigned by default. In v17 or earlier, we accidentally let
4279 : * C implementation signedness affect persistent data. This led to
4280 : * inconsistent results when comparing char data across different
4281 : * platforms.
4282 : *
4283 : * This flag can be used as a hint to ensure consistent behavior for
4284 : * pre-v18 data files that store data sorted by the 'char' type on disk,
4285 : * especially in cross-platform replication scenarios.
4286 : *
4287 : * Newly created database clusters unconditionally set the default char
4288 : * signedness to true. pg_upgrade changes this flag for clusters that were
4289 : * initialized on signedness=false platforms. As a result,
4290 : * signedness=false setting will become rare over time. If we had known
4291 : * about this problem during the last development cycle that forced initdb
4292 : * (v8.3), we would have made all clusters signed or all clusters
4293 : * unsigned. Making pg_upgrade the only source of signedness=false will
4294 : * cause the population of database clusters to converge toward that
4295 : * retrospective ideal.
4296 : */
4297 90 : ControlFile->default_char_signedness = true;
4298 :
4299 : /* Contents are protected with a CRC */
4300 90 : INIT_CRC32C(ControlFile->crc);
4301 90 : COMP_CRC32C(ControlFile->crc,
4302 : ControlFile,
4303 : offsetof(ControlFileData, crc));
4304 90 : FIN_CRC32C(ControlFile->crc);
4305 :
4306 : /*
4307 : * We write out PG_CONTROL_FILE_SIZE bytes into pg_control, zero-padding
4308 : * the excess over sizeof(ControlFileData). This reduces the odds of
4309 : * premature-EOF errors when reading pg_control. We'll still fail when we
4310 : * check the contents of the file, but hopefully with a more specific
4311 : * error than "couldn't read pg_control".
4312 : */
4313 90 : memset(buffer, 0, PG_CONTROL_FILE_SIZE);
4314 90 : memcpy(buffer, ControlFile, sizeof(ControlFileData));
4315 :
4316 90 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4317 : O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
4318 90 : if (fd < 0)
4319 0 : ereport(PANIC,
4320 : (errcode_for_file_access(),
4321 : errmsg("could not create file \"%s\": %m",
4322 : XLOG_CONTROL_FILE)));
4323 :
4324 90 : errno = 0;
4325 90 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_WRITE);
4326 90 : if (write(fd, buffer, PG_CONTROL_FILE_SIZE) != PG_CONTROL_FILE_SIZE)
4327 : {
4328 : /* if write didn't set errno, assume problem is no disk space */
4329 0 : if (errno == 0)
4330 0 : errno = ENOSPC;
4331 0 : ereport(PANIC,
4332 : (errcode_for_file_access(),
4333 : errmsg("could not write to file \"%s\": %m",
4334 : XLOG_CONTROL_FILE)));
4335 : }
4336 90 : pgstat_report_wait_end();
4337 :
4338 90 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_SYNC);
4339 90 : if (pg_fsync(fd) != 0)
4340 0 : ereport(PANIC,
4341 : (errcode_for_file_access(),
4342 : errmsg("could not fsync file \"%s\": %m",
4343 : XLOG_CONTROL_FILE)));
4344 90 : pgstat_report_wait_end();
4345 :
4346 90 : if (close(fd) != 0)
4347 0 : ereport(PANIC,
4348 : (errcode_for_file_access(),
4349 : errmsg("could not close file \"%s\": %m",
4350 : XLOG_CONTROL_FILE)));
4351 90 : }
4352 :
4353 : static void
4354 1868 : ReadControlFile(void)
4355 : {
4356 : pg_crc32c crc;
4357 : int fd;
4358 : char wal_segsz_str[20];
4359 : int r;
4360 :
4361 : /*
4362 : * Read data...
4363 : */
4364 1868 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4365 : O_RDWR | PG_BINARY);
4366 1868 : if (fd < 0)
4367 0 : ereport(PANIC,
4368 : (errcode_for_file_access(),
4369 : errmsg("could not open file \"%s\": %m",
4370 : XLOG_CONTROL_FILE)));
4371 :
4372 1868 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ);
4373 1868 : r = read(fd, ControlFile, sizeof(ControlFileData));
4374 1868 : if (r != sizeof(ControlFileData))
4375 : {
4376 0 : if (r < 0)
4377 0 : ereport(PANIC,
4378 : (errcode_for_file_access(),
4379 : errmsg("could not read file \"%s\": %m",
4380 : XLOG_CONTROL_FILE)));
4381 : else
4382 0 : ereport(PANIC,
4383 : (errcode(ERRCODE_DATA_CORRUPTED),
4384 : errmsg("could not read file \"%s\": read %d of %zu",
4385 : XLOG_CONTROL_FILE, r, sizeof(ControlFileData))));
4386 : }
4387 1868 : pgstat_report_wait_end();
4388 :
4389 1868 : close(fd);
4390 :
4391 : /*
4392 : * Check for expected pg_control format version. If this is wrong, the
4393 : * CRC check will likely fail because we'll be checking the wrong number
4394 : * of bytes. Complaining about wrong version will probably be more
4395 : * enlightening than complaining about wrong CRC.
4396 : */
4397 :
4398 1868 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
4399 0 : ereport(FATAL,
4400 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4401 : errmsg("database files are incompatible with server"),
4402 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4403 : " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4404 : ControlFile->pg_control_version, ControlFile->pg_control_version,
4405 : PG_CONTROL_VERSION, PG_CONTROL_VERSION),
4406 : errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4407 :
4408 1868 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
4409 0 : ereport(FATAL,
4410 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4411 : errmsg("database files are incompatible with server"),
4412 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4413 : " but the server was compiled with PG_CONTROL_VERSION %d.",
4414 : ControlFile->pg_control_version, PG_CONTROL_VERSION),
4415 : errhint("It looks like you need to initdb.")));
4416 :
4417 : /* Now check the CRC. */
4418 1868 : INIT_CRC32C(crc);
4419 1868 : COMP_CRC32C(crc,
4420 : ControlFile,
4421 : offsetof(ControlFileData, crc));
4422 1868 : FIN_CRC32C(crc);
4423 :
4424 1868 : if (!EQ_CRC32C(crc, ControlFile->crc))
4425 0 : ereport(FATAL,
4426 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4427 : errmsg("incorrect checksum in control file")));
4428 :
4429 : /*
4430 : * Do compatibility checking immediately. If the database isn't
4431 : * compatible with the backend executable, we want to abort before we can
4432 : * possibly do any damage.
4433 : */
4434 1868 : if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
4435 0 : ereport(FATAL,
4436 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4437 : errmsg("database files are incompatible with server"),
4438 : /* translator: %s is a variable name and %d is its value */
4439 : errdetail("The database cluster was initialized with %s %d,"
4440 : " but the server was compiled with %s %d.",
4441 : "CATALOG_VERSION_NO", ControlFile->catalog_version_no,
4442 : "CATALOG_VERSION_NO", CATALOG_VERSION_NO),
4443 : errhint("It looks like you need to initdb.")));
4444 1868 : if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
4445 0 : ereport(FATAL,
4446 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4447 : errmsg("database files are incompatible with server"),
4448 : /* translator: %s is a variable name and %d is its value */
4449 : errdetail("The database cluster was initialized with %s %d,"
4450 : " but the server was compiled with %s %d.",
4451 : "MAXALIGN", ControlFile->maxAlign,
4452 : "MAXALIGN", MAXIMUM_ALIGNOF),
4453 : errhint("It looks like you need to initdb.")));
4454 1868 : if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
4455 0 : ereport(FATAL,
4456 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4457 : errmsg("database files are incompatible with server"),
4458 : errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4459 : errhint("It looks like you need to initdb.")));
4460 1868 : if (ControlFile->blcksz != BLCKSZ)
4461 0 : ereport(FATAL,
4462 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4463 : errmsg("database files are incompatible with server"),
4464 : /* translator: %s is a variable name and %d is its value */
4465 : errdetail("The database cluster was initialized with %s %d,"
4466 : " but the server was compiled with %s %d.",
4467 : "BLCKSZ", ControlFile->blcksz,
4468 : "BLCKSZ", BLCKSZ),
4469 : errhint("It looks like you need to recompile or initdb.")));
4470 1868 : if (ControlFile->relseg_size != RELSEG_SIZE)
4471 0 : ereport(FATAL,
4472 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4473 : errmsg("database files are incompatible with server"),
4474 : /* translator: %s is a variable name and %d is its value */
4475 : errdetail("The database cluster was initialized with %s %d,"
4476 : " but the server was compiled with %s %d.",
4477 : "RELSEG_SIZE", ControlFile->relseg_size,
4478 : "RELSEG_SIZE", RELSEG_SIZE),
4479 : errhint("It looks like you need to recompile or initdb.")));
4480 1868 : if (ControlFile->xlog_blcksz != XLOG_BLCKSZ)
4481 0 : ereport(FATAL,
4482 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4483 : errmsg("database files are incompatible with server"),
4484 : /* translator: %s is a variable name and %d is its value */
4485 : errdetail("The database cluster was initialized with %s %d,"
4486 : " but the server was compiled with %s %d.",
4487 : "XLOG_BLCKSZ", ControlFile->xlog_blcksz,
4488 : "XLOG_BLCKSZ", XLOG_BLCKSZ),
4489 : errhint("It looks like you need to recompile or initdb.")));
4490 1868 : if (ControlFile->nameDataLen != NAMEDATALEN)
4491 0 : ereport(FATAL,
4492 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4493 : errmsg("database files are incompatible with server"),
4494 : /* translator: %s is a variable name and %d is its value */
4495 : errdetail("The database cluster was initialized with %s %d,"
4496 : " but the server was compiled with %s %d.",
4497 : "NAMEDATALEN", ControlFile->nameDataLen,
4498 : "NAMEDATALEN", NAMEDATALEN),
4499 : errhint("It looks like you need to recompile or initdb.")));
4500 1868 : if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
4501 0 : ereport(FATAL,
4502 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4503 : errmsg("database files are incompatible with server"),
4504 : /* translator: %s is a variable name and %d is its value */
4505 : errdetail("The database cluster was initialized with %s %d,"
4506 : " but the server was compiled with %s %d.",
4507 : "INDEX_MAX_KEYS", ControlFile->indexMaxKeys,
4508 : "INDEX_MAX_KEYS", INDEX_MAX_KEYS),
4509 : errhint("It looks like you need to recompile or initdb.")));
4510 1868 : if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
4511 0 : ereport(FATAL,
4512 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4513 : errmsg("database files are incompatible with server"),
4514 : /* translator: %s is a variable name and %d is its value */
4515 : errdetail("The database cluster was initialized with %s %d,"
4516 : " but the server was compiled with %s %d.",
4517 : "TOAST_MAX_CHUNK_SIZE", ControlFile->toast_max_chunk_size,
4518 : "TOAST_MAX_CHUNK_SIZE", (int) TOAST_MAX_CHUNK_SIZE),
4519 : errhint("It looks like you need to recompile or initdb.")));
4520 1868 : if (ControlFile->loblksize != LOBLKSIZE)
4521 0 : ereport(FATAL,
4522 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4523 : errmsg("database files are incompatible with server"),
4524 : /* translator: %s is a variable name and %d is its value */
4525 : errdetail("The database cluster was initialized with %s %d,"
4526 : " but the server was compiled with %s %d.",
4527 : "LOBLKSIZE", ControlFile->loblksize,
4528 : "LOBLKSIZE", (int) LOBLKSIZE),
4529 : errhint("It looks like you need to recompile or initdb.")));
4530 :
4531 : #ifdef USE_FLOAT8_BYVAL
4532 1868 : if (ControlFile->float8ByVal != true)
4533 0 : ereport(FATAL,
4534 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4535 : errmsg("database files are incompatible with server"),
4536 : errdetail("The database cluster was initialized without USE_FLOAT8_BYVAL"
4537 : " but the server was compiled with USE_FLOAT8_BYVAL."),
4538 : errhint("It looks like you need to recompile or initdb.")));
4539 : #else
4540 : if (ControlFile->float8ByVal != false)
4541 : ereport(FATAL,
4542 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4543 : errmsg("database files are incompatible with server"),
4544 : errdetail("The database cluster was initialized with USE_FLOAT8_BYVAL"
4545 : " but the server was compiled without USE_FLOAT8_BYVAL."),
4546 : errhint("It looks like you need to recompile or initdb.")));
4547 : #endif
4548 :
4549 1868 : wal_segment_size = ControlFile->xlog_seg_size;
4550 :
4551 1868 : if (!IsValidWalSegSize(wal_segment_size))
4552 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4553 : errmsg_plural("invalid WAL segment size in control file (%d byte)",
4554 : "invalid WAL segment size in control file (%d bytes)",
4555 : wal_segment_size,
4556 : wal_segment_size),
4557 : errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.")));
4558 :
4559 1868 : snprintf(wal_segsz_str, sizeof(wal_segsz_str), "%d", wal_segment_size);
4560 1868 : SetConfigOption("wal_segment_size", wal_segsz_str, PGC_INTERNAL,
4561 : PGC_S_DYNAMIC_DEFAULT);
4562 :
4563 : /* check and update variables dependent on wal_segment_size */
4564 1868 : if (ConvertToXSegs(min_wal_size_mb, wal_segment_size) < 2)
4565 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4566 : /* translator: both %s are GUC names */
4567 : errmsg("\"%s\" must be at least twice \"%s\"",
4568 : "min_wal_size", "wal_segment_size")));
4569 :
4570 1868 : if (ConvertToXSegs(max_wal_size_mb, wal_segment_size) < 2)
4571 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4572 : /* translator: both %s are GUC names */
4573 : errmsg("\"%s\" must be at least twice \"%s\"",
4574 : "max_wal_size", "wal_segment_size")));
4575 :
4576 1868 : UsableBytesInSegment =
4577 1868 : (wal_segment_size / XLOG_BLCKSZ * UsableBytesInPage) -
4578 : (SizeOfXLogLongPHD - SizeOfXLogShortPHD);
4579 :
4580 1868 : CalculateCheckpointSegments();
4581 :
4582 : /* Make the initdb settings visible as GUC variables, too */
4583 1868 : SetConfigOption("data_checksums", DataChecksumsEnabled() ? "yes" : "no",
4584 : PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
4585 1868 : }
4586 :
4587 : /*
4588 : * Utility wrapper to update the control file. Note that the control
4589 : * file gets flushed.
4590 : */
4591 : static void
4592 16690 : UpdateControlFile(void)
4593 : {
4594 16690 : update_controlfile(DataDir, ControlFile, true);
4595 16690 : }
4596 :
4597 : /*
4598 : * Returns the unique system identifier from control file.
4599 : */
4600 : uint64
4601 2686 : GetSystemIdentifier(void)
4602 : {
4603 : Assert(ControlFile != NULL);
4604 2686 : return ControlFile->system_identifier;
4605 : }
4606 :
4607 : /*
4608 : * Returns the random nonce from control file.
4609 : */
4610 : char *
4611 2 : GetMockAuthenticationNonce(void)
4612 : {
4613 : Assert(ControlFile != NULL);
4614 2 : return ControlFile->mock_authentication_nonce;
4615 : }
4616 :
4617 : /*
4618 : * Are checksums enabled for data pages?
4619 : */
4620 : bool
4621 23201836 : DataChecksumsEnabled(void)
4622 : {
4623 : Assert(ControlFile != NULL);
4624 23201836 : return (ControlFile->data_checksum_version > 0);
4625 : }
4626 :
4627 : /*
4628 : * Return true if the cluster was initialized on a platform where the
4629 : * default signedness of char is "signed". This function exists for code
4630 : * that deals with pre-v18 data files that store data sorted by the 'char'
4631 : * type on disk (e.g., GIN and GiST indexes). See the comments in
4632 : * WriteControlFile() for details.
4633 : */
4634 : bool
4635 6 : GetDefaultCharSignedness(void)
4636 : {
4637 6 : return ControlFile->default_char_signedness;
4638 : }
4639 :
4640 : /*
4641 : * Returns a fake LSN for unlogged relations.
4642 : *
4643 : * Each call generates an LSN that is greater than any previous value
4644 : * returned. The current counter value is saved and restored across clean
4645 : * shutdowns, but like unlogged relations, does not survive a crash. This can
4646 : * be used in lieu of real LSN values returned by XLogInsert, if you need an
4647 : * LSN-like increasing sequence of numbers without writing any WAL.
4648 : */
4649 : XLogRecPtr
4650 66 : GetFakeLSNForUnloggedRel(void)
4651 : {
4652 66 : return pg_atomic_fetch_add_u64(&XLogCtl->unloggedLSN, 1);
4653 : }
4654 :
4655 : /*
4656 : * Auto-tune the number of XLOG buffers.
4657 : *
4658 : * The preferred setting for wal_buffers is about 3% of shared_buffers, with
4659 : * a maximum of one XLOG segment (there is little reason to think that more
4660 : * is helpful, at least so long as we force an fsync when switching log files)
4661 : * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
4662 : * 9.1, when auto-tuning was added).
4663 : *
4664 : * This should not be called until NBuffers has received its final value.
4665 : */
4666 : static int
4667 2028 : XLOGChooseNumBuffers(void)
4668 : {
4669 : int xbuffers;
4670 :
4671 2028 : xbuffers = NBuffers / 32;
4672 2028 : if (xbuffers > (wal_segment_size / XLOG_BLCKSZ))
4673 44 : xbuffers = (wal_segment_size / XLOG_BLCKSZ);
4674 2028 : if (xbuffers < 8)
4675 768 : xbuffers = 8;
4676 2028 : return xbuffers;
4677 : }
4678 :
4679 : /*
4680 : * GUC check_hook for wal_buffers
4681 : */
4682 : bool
4683 4126 : check_wal_buffers(int *newval, void **extra, GucSource source)
4684 : {
4685 : /*
4686 : * -1 indicates a request for auto-tune.
4687 : */
4688 4126 : if (*newval == -1)
4689 : {
4690 : /*
4691 : * If we haven't yet changed the boot_val default of -1, just let it
4692 : * be. We'll fix it when XLOGShmemSize is called.
4693 : */
4694 2098 : if (XLOGbuffers == -1)
4695 2098 : return true;
4696 :
4697 : /* Otherwise, substitute the auto-tune value */
4698 0 : *newval = XLOGChooseNumBuffers();
4699 : }
4700 :
4701 : /*
4702 : * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
4703 : * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
4704 : * the case, we just silently treat such values as a request for the
4705 : * minimum. (We could throw an error instead, but that doesn't seem very
4706 : * helpful.)
4707 : */
4708 2028 : if (*newval < 4)
4709 0 : *newval = 4;
4710 :
4711 2028 : return true;
4712 : }
4713 :
4714 : /*
4715 : * GUC check_hook for wal_consistency_checking
4716 : */
4717 : bool
4718 3890 : check_wal_consistency_checking(char **newval, void **extra, GucSource source)
4719 : {
4720 : char *rawstring;
4721 : List *elemlist;
4722 : ListCell *l;
4723 : bool newwalconsistency[RM_MAX_ID + 1];
4724 :
4725 : /* Initialize the array */
4726 128370 : MemSet(newwalconsistency, 0, (RM_MAX_ID + 1) * sizeof(bool));
4727 :
4728 : /* Need a modifiable copy of string */
4729 3890 : rawstring = pstrdup(*newval);
4730 :
4731 : /* Parse string into list of identifiers */
4732 3890 : if (!SplitIdentifierString(rawstring, ',', &elemlist))
4733 : {
4734 : /* syntax error in list */
4735 0 : GUC_check_errdetail("List syntax is invalid.");
4736 0 : pfree(rawstring);
4737 0 : list_free(elemlist);
4738 0 : return false;
4739 : }
4740 :
4741 4792 : foreach(l, elemlist)
4742 : {
4743 902 : char *tok = (char *) lfirst(l);
4744 : int rmid;
4745 :
4746 : /* Check for 'all'. */
4747 902 : if (pg_strcasecmp(tok, "all") == 0)
4748 : {
4749 230786 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4750 229888 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL)
4751 8980 : newwalconsistency[rmid] = true;
4752 : }
4753 : else
4754 : {
4755 : /* Check if the token matches any known resource manager. */
4756 4 : bool found = false;
4757 :
4758 72 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4759 : {
4760 108 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL &&
4761 36 : pg_strcasecmp(tok, GetRmgr(rmid).rm_name) == 0)
4762 : {
4763 4 : newwalconsistency[rmid] = true;
4764 4 : found = true;
4765 4 : break;
4766 : }
4767 : }
4768 4 : if (!found)
4769 : {
4770 : /*
4771 : * During startup, it might be a not-yet-loaded custom
4772 : * resource manager. Defer checking until
4773 : * InitializeWalConsistencyChecking().
4774 : */
4775 0 : if (!process_shared_preload_libraries_done)
4776 : {
4777 0 : check_wal_consistency_checking_deferred = true;
4778 : }
4779 : else
4780 : {
4781 0 : GUC_check_errdetail("Unrecognized key word: \"%s\".", tok);
4782 0 : pfree(rawstring);
4783 0 : list_free(elemlist);
4784 0 : return false;
4785 : }
4786 : }
4787 : }
4788 : }
4789 :
4790 3890 : pfree(rawstring);
4791 3890 : list_free(elemlist);
4792 :
4793 : /* assign new value */
4794 3890 : *extra = guc_malloc(LOG, (RM_MAX_ID + 1) * sizeof(bool));
4795 3890 : if (!*extra)
4796 0 : return false;
4797 3890 : memcpy(*extra, newwalconsistency, (RM_MAX_ID + 1) * sizeof(bool));
4798 3890 : return true;
4799 : }
4800 :
4801 : /*
4802 : * GUC assign_hook for wal_consistency_checking
4803 : */
4804 : void
4805 3888 : assign_wal_consistency_checking(const char *newval, void *extra)
4806 : {
4807 : /*
4808 : * If some checks were deferred, it's possible that the checks will fail
4809 : * later during InitializeWalConsistencyChecking(). But in that case, the
4810 : * postmaster will exit anyway, so it's safe to proceed with the
4811 : * assignment.
4812 : *
4813 : * Any built-in resource managers specified are assigned immediately,
4814 : * which affects WAL created before shared_preload_libraries are
4815 : * processed. Any custom resource managers specified won't be assigned
4816 : * until after shared_preload_libraries are processed, but that's OK
4817 : * because WAL for a custom resource manager can't be written before the
4818 : * module is loaded anyway.
4819 : */
4820 3888 : wal_consistency_checking = extra;
4821 3888 : }
4822 :
4823 : /*
4824 : * InitializeWalConsistencyChecking: run after loading custom resource managers
4825 : *
4826 : * If any unknown resource managers were specified in the
4827 : * wal_consistency_checking GUC, processing was deferred. Now that
4828 : * shared_preload_libraries have been loaded, process wal_consistency_checking
4829 : * again.
4830 : */
4831 : void
4832 1758 : InitializeWalConsistencyChecking(void)
4833 : {
4834 : Assert(process_shared_preload_libraries_done);
4835 :
4836 1758 : if (check_wal_consistency_checking_deferred)
4837 : {
4838 : struct config_generic *guc;
4839 :
4840 0 : guc = find_option("wal_consistency_checking", false, false, ERROR);
4841 :
4842 0 : check_wal_consistency_checking_deferred = false;
4843 :
4844 0 : set_config_option_ext("wal_consistency_checking",
4845 : wal_consistency_checking_string,
4846 : guc->scontext, guc->source, guc->srole,
4847 : GUC_ACTION_SET, true, ERROR, false);
4848 :
4849 : /* checking should not be deferred again */
4850 : Assert(!check_wal_consistency_checking_deferred);
4851 : }
4852 1758 : }
4853 :
4854 : /*
4855 : * GUC show_hook for archive_command
4856 : */
4857 : const char *
4858 3470 : show_archive_command(void)
4859 : {
4860 3470 : if (XLogArchivingActive())
4861 4 : return XLogArchiveCommand;
4862 : else
4863 3466 : return "(disabled)";
4864 : }
4865 :
4866 : /*
4867 : * GUC show_hook for in_hot_standby
4868 : */
4869 : const char *
4870 31020 : show_in_hot_standby(void)
4871 : {
4872 : /*
4873 : * We display the actual state based on shared memory, so that this GUC
4874 : * reports up-to-date state if examined intra-query. The underlying
4875 : * variable (in_hot_standby_guc) changes only when we transmit a new value
4876 : * to the client.
4877 : */
4878 31020 : return RecoveryInProgress() ? "on" : "off";
4879 : }
4880 :
4881 : /*
4882 : * Read the control file, set respective GUCs.
4883 : *
4884 : * This is to be called during startup, including a crash recovery cycle,
4885 : * unless in bootstrap mode, where no control file yet exists. As there's no
4886 : * usable shared memory yet (its sizing can depend on the contents of the
4887 : * control file!), first store the contents in local memory. XLOGShmemInit()
4888 : * will then copy it to shared memory later.
4889 : *
4890 : * reset just controls whether previous contents are to be expected (in the
4891 : * reset case, there's a dangling pointer into old shared memory), or not.
4892 : */
4893 : void
4894 1778 : LocalProcessControlFile(bool reset)
4895 : {
4896 : Assert(reset || ControlFile == NULL);
4897 1778 : ControlFile = palloc(sizeof(ControlFileData));
4898 1778 : ReadControlFile();
4899 1778 : }
4900 :
4901 : /*
4902 : * Get the wal_level from the control file. For a standby, this value should be
4903 : * considered as its active wal_level, because it may be different from what
4904 : * was originally configured on standby.
4905 : */
4906 : WalLevel
4907 136 : GetActiveWalLevelOnStandby(void)
4908 : {
4909 136 : return ControlFile->wal_level;
4910 : }
4911 :
4912 : /*
4913 : * Initialization of shared memory for XLOG
4914 : */
4915 : Size
4916 5826 : XLOGShmemSize(void)
4917 : {
4918 : Size size;
4919 :
4920 : /*
4921 : * If the value of wal_buffers is -1, use the preferred auto-tune value.
4922 : * This isn't an amazingly clean place to do this, but we must wait till
4923 : * NBuffers has received its final value, and must do it before using the
4924 : * value of XLOGbuffers to do anything important.
4925 : *
4926 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
4927 : * However, if the DBA explicitly set wal_buffers = -1 in the config file,
4928 : * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
4929 : * the matter with PGC_S_OVERRIDE.
4930 : */
4931 5826 : if (XLOGbuffers == -1)
4932 : {
4933 : char buf[32];
4934 :
4935 2028 : snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
4936 2028 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4937 : PGC_S_DYNAMIC_DEFAULT);
4938 2028 : if (XLOGbuffers == -1) /* failed to apply it? */
4939 0 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4940 : PGC_S_OVERRIDE);
4941 : }
4942 : Assert(XLOGbuffers > 0);
4943 :
4944 : /* XLogCtl */
4945 5826 : size = sizeof(XLogCtlData);
4946 :
4947 : /* WAL insertion locks, plus alignment */
4948 5826 : size = add_size(size, mul_size(sizeof(WALInsertLockPadded), NUM_XLOGINSERT_LOCKS + 1));
4949 : /* xlblocks array */
4950 5826 : size = add_size(size, mul_size(sizeof(pg_atomic_uint64), XLOGbuffers));
4951 : /* extra alignment padding for XLOG I/O buffers */
4952 5826 : size = add_size(size, Max(XLOG_BLCKSZ, PG_IO_ALIGN_SIZE));
4953 : /* and the buffers themselves */
4954 5826 : size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
4955 :
4956 : /*
4957 : * Note: we don't count ControlFileData, it comes out of the "slop factor"
4958 : * added by CreateSharedMemoryAndSemaphores. This lets us use this
4959 : * routine again below to compute the actual allocation size.
4960 : */
4961 :
4962 5826 : return size;
4963 : }
4964 :
4965 : void
4966 2032 : XLOGShmemInit(void)
4967 : {
4968 : bool foundCFile,
4969 : foundXLog;
4970 : char *allocptr;
4971 : int i;
4972 : ControlFileData *localControlFile;
4973 :
4974 : #ifdef WAL_DEBUG
4975 :
4976 : /*
4977 : * Create a memory context for WAL debugging that's exempt from the normal
4978 : * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
4979 : * an allocation fails, but wal_debug is not for production use anyway.
4980 : */
4981 : if (walDebugCxt == NULL)
4982 : {
4983 : walDebugCxt = AllocSetContextCreate(TopMemoryContext,
4984 : "WAL Debug",
4985 : ALLOCSET_DEFAULT_SIZES);
4986 : MemoryContextAllowInCriticalSection(walDebugCxt, true);
4987 : }
4988 : #endif
4989 :
4990 :
4991 2032 : XLogCtl = (XLogCtlData *)
4992 2032 : ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog);
4993 :
4994 2032 : localControlFile = ControlFile;
4995 2032 : ControlFile = (ControlFileData *)
4996 2032 : ShmemInitStruct("Control File", sizeof(ControlFileData), &foundCFile);
4997 :
4998 2032 : if (foundCFile || foundXLog)
4999 : {
5000 : /* both should be present or neither */
5001 : Assert(foundCFile && foundXLog);
5002 :
5003 : /* Initialize local copy of WALInsertLocks */
5004 0 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks;
5005 :
5006 0 : if (localControlFile)
5007 0 : pfree(localControlFile);
5008 0 : return;
5009 : }
5010 2032 : memset(XLogCtl, 0, sizeof(XLogCtlData));
5011 :
5012 : /*
5013 : * Already have read control file locally, unless in bootstrap mode. Move
5014 : * contents into shared memory.
5015 : */
5016 2032 : if (localControlFile)
5017 : {
5018 1762 : memcpy(ControlFile, localControlFile, sizeof(ControlFileData));
5019 1762 : pfree(localControlFile);
5020 : }
5021 :
5022 : /*
5023 : * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
5024 : * multiple of the alignment for same, so no extra alignment padding is
5025 : * needed here.
5026 : */
5027 2032 : allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
5028 2032 : XLogCtl->xlblocks = (pg_atomic_uint64 *) allocptr;
5029 2032 : allocptr += sizeof(pg_atomic_uint64) * XLOGbuffers;
5030 :
5031 592134 : for (i = 0; i < XLOGbuffers; i++)
5032 : {
5033 590102 : pg_atomic_init_u64(&XLogCtl->xlblocks[i], InvalidXLogRecPtr);
5034 : }
5035 :
5036 : /* WAL insertion locks. Ensure they're aligned to the full padded size */
5037 2032 : allocptr += sizeof(WALInsertLockPadded) -
5038 2032 : ((uintptr_t) allocptr) % sizeof(WALInsertLockPadded);
5039 2032 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks =
5040 : (WALInsertLockPadded *) allocptr;
5041 2032 : allocptr += sizeof(WALInsertLockPadded) * NUM_XLOGINSERT_LOCKS;
5042 :
5043 18288 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
5044 : {
5045 16256 : LWLockInitialize(&WALInsertLocks[i].l.lock, LWTRANCHE_WAL_INSERT);
5046 16256 : pg_atomic_init_u64(&WALInsertLocks[i].l.insertingAt, InvalidXLogRecPtr);
5047 16256 : WALInsertLocks[i].l.lastImportantAt = InvalidXLogRecPtr;
5048 : }
5049 :
5050 : /*
5051 : * Align the start of the page buffers to a full xlog block size boundary.
5052 : * This simplifies some calculations in XLOG insertion. It is also
5053 : * required for O_DIRECT.
5054 : */
5055 2032 : allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5056 2032 : XLogCtl->pages = allocptr;
5057 2032 : memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5058 :
5059 : /*
5060 : * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5061 : * in additional info.)
5062 : */
5063 2032 : XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5064 2032 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5065 2032 : XLogCtl->InstallXLogFileSegmentActive = false;
5066 2032 : XLogCtl->WalWriterSleeping = false;
5067 :
5068 2032 : SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5069 2032 : SpinLockInit(&XLogCtl->info_lck);
5070 2032 : pg_atomic_init_u64(&XLogCtl->logInsertResult, InvalidXLogRecPtr);
5071 2032 : pg_atomic_init_u64(&XLogCtl->logWriteResult, InvalidXLogRecPtr);
5072 2032 : pg_atomic_init_u64(&XLogCtl->logFlushResult, InvalidXLogRecPtr);
5073 2032 : pg_atomic_init_u64(&XLogCtl->unloggedLSN, InvalidXLogRecPtr);
5074 : }
5075 :
5076 : /*
5077 : * This func must be called ONCE on system install. It creates pg_control
5078 : * and the initial XLOG segment.
5079 : */
5080 : void
5081 90 : BootStrapXLOG(uint32 data_checksum_version)
5082 : {
5083 : CheckPoint checkPoint;
5084 : char *buffer;
5085 : XLogPageHeader page;
5086 : XLogLongPageHeader longpage;
5087 : XLogRecord *record;
5088 : char *recptr;
5089 : uint64 sysidentifier;
5090 : struct timeval tv;
5091 : pg_crc32c crc;
5092 :
5093 : /* allow ordinary WAL segment creation, like StartupXLOG() would */
5094 90 : SetInstallXLogFileSegmentActive();
5095 :
5096 : /*
5097 : * Select a hopefully-unique system identifier code for this installation.
5098 : * We use the result of gettimeofday(), including the fractional seconds
5099 : * field, as being about as unique as we can easily get. (Think not to
5100 : * use random(), since it hasn't been seeded and there's no portable way
5101 : * to seed it other than the system clock value...) The upper half of the
5102 : * uint64 value is just the tv_sec part, while the lower half contains the
5103 : * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
5104 : * PID for a little extra uniqueness. A person knowing this encoding can
5105 : * determine the initialization time of the installation, which could
5106 : * perhaps be useful sometimes.
5107 : */
5108 90 : gettimeofday(&tv, NULL);
5109 90 : sysidentifier = ((uint64) tv.tv_sec) << 32;
5110 90 : sysidentifier |= ((uint64) tv.tv_usec) << 12;
5111 90 : sysidentifier |= getpid() & 0xFFF;
5112 :
5113 : /* page buffer must be aligned suitably for O_DIRECT */
5114 90 : buffer = (char *) palloc(XLOG_BLCKSZ + XLOG_BLCKSZ);
5115 90 : page = (XLogPageHeader) TYPEALIGN(XLOG_BLCKSZ, buffer);
5116 90 : memset(page, 0, XLOG_BLCKSZ);
5117 :
5118 : /*
5119 : * Set up information for the initial checkpoint record
5120 : *
5121 : * The initial checkpoint record is written to the beginning of the WAL
5122 : * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5123 : * used, so that we can use 0/0 to mean "before any valid WAL segment".
5124 : */
5125 90 : checkPoint.redo = wal_segment_size + SizeOfXLogLongPHD;
5126 90 : checkPoint.ThisTimeLineID = BootstrapTimeLineID;
5127 90 : checkPoint.PrevTimeLineID = BootstrapTimeLineID;
5128 90 : checkPoint.fullPageWrites = fullPageWrites;
5129 90 : checkPoint.wal_level = wal_level;
5130 : checkPoint.nextXid =
5131 90 : FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId);
5132 90 : checkPoint.nextOid = FirstGenbkiObjectId;
5133 90 : checkPoint.nextMulti = FirstMultiXactId;
5134 90 : checkPoint.nextMultiOffset = 0;
5135 90 : checkPoint.oldestXid = FirstNormalTransactionId;
5136 90 : checkPoint.oldestXidDB = Template1DbOid;
5137 90 : checkPoint.oldestMulti = FirstMultiXactId;
5138 90 : checkPoint.oldestMultiDB = Template1DbOid;
5139 90 : checkPoint.oldestCommitTsXid = InvalidTransactionId;
5140 90 : checkPoint.newestCommitTsXid = InvalidTransactionId;
5141 90 : checkPoint.time = (pg_time_t) time(NULL);
5142 90 : checkPoint.oldestActiveXid = InvalidTransactionId;
5143 :
5144 90 : TransamVariables->nextXid = checkPoint.nextXid;
5145 90 : TransamVariables->nextOid = checkPoint.nextOid;
5146 90 : TransamVariables->oidCount = 0;
5147 90 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5148 90 : AdvanceOldestClogXid(checkPoint.oldestXid);
5149 90 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5150 90 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5151 90 : SetCommitTsLimit(InvalidTransactionId, InvalidTransactionId);
5152 :
5153 : /* Set up the XLOG page header */
5154 90 : page->xlp_magic = XLOG_PAGE_MAGIC;
5155 90 : page->xlp_info = XLP_LONG_HEADER;
5156 90 : page->xlp_tli = BootstrapTimeLineID;
5157 90 : page->xlp_pageaddr = wal_segment_size;
5158 90 : longpage = (XLogLongPageHeader) page;
5159 90 : longpage->xlp_sysid = sysidentifier;
5160 90 : longpage->xlp_seg_size = wal_segment_size;
5161 90 : longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5162 :
5163 : /* Insert the initial checkpoint record */
5164 90 : recptr = ((char *) page + SizeOfXLogLongPHD);
5165 90 : record = (XLogRecord *) recptr;
5166 90 : record->xl_prev = 0;
5167 90 : record->xl_xid = InvalidTransactionId;
5168 90 : record->xl_tot_len = SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(checkPoint);
5169 90 : record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5170 90 : record->xl_rmid = RM_XLOG_ID;
5171 90 : recptr += SizeOfXLogRecord;
5172 : /* fill the XLogRecordDataHeaderShort struct */
5173 90 : *(recptr++) = (char) XLR_BLOCK_ID_DATA_SHORT;
5174 90 : *(recptr++) = sizeof(checkPoint);
5175 90 : memcpy(recptr, &checkPoint, sizeof(checkPoint));
5176 90 : recptr += sizeof(checkPoint);
5177 : Assert(recptr - (char *) record == record->xl_tot_len);
5178 :
5179 90 : INIT_CRC32C(crc);
5180 90 : COMP_CRC32C(crc, ((char *) record) + SizeOfXLogRecord, record->xl_tot_len - SizeOfXLogRecord);
5181 90 : COMP_CRC32C(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5182 90 : FIN_CRC32C(crc);
5183 90 : record->xl_crc = crc;
5184 :
5185 : /* Create first XLOG segment file */
5186 90 : openLogTLI = BootstrapTimeLineID;
5187 90 : openLogFile = XLogFileInit(1, BootstrapTimeLineID);
5188 :
5189 : /*
5190 : * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
5191 : * close the file again in a moment.
5192 : */
5193 :
5194 : /* Write the first page with the initial record */
5195 90 : errno = 0;
5196 90 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE);
5197 90 : if (write(openLogFile, page, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5198 : {
5199 : /* if write didn't set errno, assume problem is no disk space */
5200 0 : if (errno == 0)
5201 0 : errno = ENOSPC;
5202 0 : ereport(PANIC,
5203 : (errcode_for_file_access(),
5204 : errmsg("could not write bootstrap write-ahead log file: %m")));
5205 : }
5206 90 : pgstat_report_wait_end();
5207 :
5208 90 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC);
5209 90 : if (pg_fsync(openLogFile) != 0)
5210 0 : ereport(PANIC,
5211 : (errcode_for_file_access(),
5212 : errmsg("could not fsync bootstrap write-ahead log file: %m")));
5213 90 : pgstat_report_wait_end();
5214 :
5215 90 : if (close(openLogFile) != 0)
5216 0 : ereport(PANIC,
5217 : (errcode_for_file_access(),
5218 : errmsg("could not close bootstrap write-ahead log file: %m")));
5219 :
5220 90 : openLogFile = -1;
5221 :
5222 : /* Now create pg_control */
5223 90 : InitControlFile(sysidentifier, data_checksum_version);
5224 90 : ControlFile->time = checkPoint.time;
5225 90 : ControlFile->checkPoint = checkPoint.redo;
5226 90 : ControlFile->checkPointCopy = checkPoint;
5227 :
5228 : /* some additional ControlFile fields are set in WriteControlFile() */
5229 90 : WriteControlFile();
5230 :
5231 : /* Bootstrap the commit log, too */
5232 90 : BootStrapCLOG();
5233 90 : BootStrapCommitTs();
5234 90 : BootStrapSUBTRANS();
5235 90 : BootStrapMultiXact();
5236 :
5237 90 : pfree(buffer);
5238 :
5239 : /*
5240 : * Force control file to be read - in contrast to normal processing we'd
5241 : * otherwise never run the checks and GUC related initializations therein.
5242 : */
5243 90 : ReadControlFile();
5244 90 : }
5245 :
5246 : static char *
5247 1582 : str_time(pg_time_t tnow)
5248 : {
5249 1582 : char *buf = palloc(128);
5250 :
5251 1582 : pg_strftime(buf, 128,
5252 : "%Y-%m-%d %H:%M:%S %Z",
5253 1582 : pg_localtime(&tnow, log_timezone));
5254 :
5255 1582 : return buf;
5256 : }
5257 :
5258 : /*
5259 : * Initialize the first WAL segment on new timeline.
5260 : */
5261 : static void
5262 98 : XLogInitNewTimeline(TimeLineID endTLI, XLogRecPtr endOfLog, TimeLineID newTLI)
5263 : {
5264 : char xlogfname[MAXFNAMELEN];
5265 : XLogSegNo endLogSegNo;
5266 : XLogSegNo startLogSegNo;
5267 :
5268 : /* we always switch to a new timeline after archive recovery */
5269 : Assert(endTLI != newTLI);
5270 :
5271 : /*
5272 : * Update min recovery point one last time.
5273 : */
5274 98 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5275 :
5276 : /*
5277 : * Calculate the last segment on the old timeline, and the first segment
5278 : * on the new timeline. If the switch happens in the middle of a segment,
5279 : * they are the same, but if the switch happens exactly at a segment
5280 : * boundary, startLogSegNo will be endLogSegNo + 1.
5281 : */
5282 98 : XLByteToPrevSeg(endOfLog, endLogSegNo, wal_segment_size);
5283 98 : XLByteToSeg(endOfLog, startLogSegNo, wal_segment_size);
5284 :
5285 : /*
5286 : * Initialize the starting WAL segment for the new timeline. If the switch
5287 : * happens in the middle of a segment, copy data from the last WAL segment
5288 : * of the old timeline up to the switch point, to the starting WAL segment
5289 : * on the new timeline.
5290 : */
5291 98 : if (endLogSegNo == startLogSegNo)
5292 : {
5293 : /*
5294 : * Make a copy of the file on the new timeline.
5295 : *
5296 : * Writing WAL isn't allowed yet, so there are no locking
5297 : * considerations. But we should be just as tense as XLogFileInit to
5298 : * avoid emplacing a bogus file.
5299 : */
5300 76 : XLogFileCopy(newTLI, endLogSegNo, endTLI, endLogSegNo,
5301 76 : XLogSegmentOffset(endOfLog, wal_segment_size));
5302 : }
5303 : else
5304 : {
5305 : /*
5306 : * The switch happened at a segment boundary, so just create the next
5307 : * segment on the new timeline.
5308 : */
5309 : int fd;
5310 :
5311 22 : fd = XLogFileInit(startLogSegNo, newTLI);
5312 :
5313 22 : if (close(fd) != 0)
5314 : {
5315 0 : int save_errno = errno;
5316 :
5317 0 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5318 0 : errno = save_errno;
5319 0 : ereport(ERROR,
5320 : (errcode_for_file_access(),
5321 : errmsg("could not close file \"%s\": %m", xlogfname)));
5322 : }
5323 : }
5324 :
5325 : /*
5326 : * Let's just make real sure there are not .ready or .done flags posted
5327 : * for the new segment.
5328 : */
5329 98 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5330 98 : XLogArchiveCleanup(xlogfname);
5331 98 : }
5332 :
5333 : /*
5334 : * Perform cleanup actions at the conclusion of archive recovery.
5335 : */
5336 : static void
5337 98 : CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI, XLogRecPtr EndOfLog,
5338 : TimeLineID newTLI)
5339 : {
5340 : /*
5341 : * Execute the recovery_end_command, if any.
5342 : */
5343 98 : if (recoveryEndCommand && strcmp(recoveryEndCommand, "") != 0)
5344 4 : ExecuteRecoveryCommand(recoveryEndCommand,
5345 : "recovery_end_command",
5346 : true,
5347 : WAIT_EVENT_RECOVERY_END_COMMAND);
5348 :
5349 : /*
5350 : * We switched to a new timeline. Clean up segments on the old timeline.
5351 : *
5352 : * If there are any higher-numbered segments on the old timeline, remove
5353 : * them. They might contain valid WAL, but they might also be
5354 : * pre-allocated files containing garbage. In any case, they are not part
5355 : * of the new timeline's history so we don't need them.
5356 : */
5357 98 : RemoveNonParentXlogFiles(EndOfLog, newTLI);
5358 :
5359 : /*
5360 : * If the switch happened in the middle of a segment, what to do with the
5361 : * last, partial segment on the old timeline? If we don't archive it, and
5362 : * the server that created the WAL never archives it either (e.g. because
5363 : * it was hit by a meteor), it will never make it to the archive. That's
5364 : * OK from our point of view, because the new segment that we created with
5365 : * the new TLI contains all the WAL from the old timeline up to the switch
5366 : * point. But if you later try to do PITR to the "missing" WAL on the old
5367 : * timeline, recovery won't find it in the archive. It's physically
5368 : * present in the new file with new TLI, but recovery won't look there
5369 : * when it's recovering to the older timeline. On the other hand, if we
5370 : * archive the partial segment, and the original server on that timeline
5371 : * is still running and archives the completed version of the same segment
5372 : * later, it will fail. (We used to do that in 9.4 and below, and it
5373 : * caused such problems).
5374 : *
5375 : * As a compromise, we rename the last segment with the .partial suffix,
5376 : * and archive it. Archive recovery will never try to read .partial
5377 : * segments, so they will normally go unused. But in the odd PITR case,
5378 : * the administrator can copy them manually to the pg_wal directory
5379 : * (removing the suffix). They can be useful in debugging, too.
5380 : *
5381 : * If a .done or .ready file already exists for the old timeline, however,
5382 : * we had already determined that the segment is complete, so we can let
5383 : * it be archived normally. (In particular, if it was restored from the
5384 : * archive to begin with, it's expected to have a .done file).
5385 : */
5386 98 : if (XLogSegmentOffset(EndOfLog, wal_segment_size) != 0 &&
5387 : XLogArchivingActive())
5388 : {
5389 : char origfname[MAXFNAMELEN];
5390 : XLogSegNo endLogSegNo;
5391 :
5392 18 : XLByteToPrevSeg(EndOfLog, endLogSegNo, wal_segment_size);
5393 18 : XLogFileName(origfname, EndOfLogTLI, endLogSegNo, wal_segment_size);
5394 :
5395 18 : if (!XLogArchiveIsReadyOrDone(origfname))
5396 : {
5397 : char origpath[MAXPGPATH];
5398 : char partialfname[MAXFNAMELEN];
5399 : char partialpath[MAXPGPATH];
5400 :
5401 : /*
5402 : * If we're summarizing WAL, we can't rename the partial file
5403 : * until the summarizer finishes with it, else it will fail.
5404 : */
5405 10 : if (summarize_wal)
5406 2 : WaitForWalSummarization(EndOfLog);
5407 :
5408 10 : XLogFilePath(origpath, EndOfLogTLI, endLogSegNo, wal_segment_size);
5409 10 : snprintf(partialfname, MAXFNAMELEN, "%s.partial", origfname);
5410 10 : snprintf(partialpath, MAXPGPATH, "%s.partial", origpath);
5411 :
5412 : /*
5413 : * Make sure there's no .done or .ready file for the .partial
5414 : * file.
5415 : */
5416 10 : XLogArchiveCleanup(partialfname);
5417 :
5418 10 : durable_rename(origpath, partialpath, ERROR);
5419 10 : XLogArchiveNotify(partialfname);
5420 : }
5421 : }
5422 98 : }
5423 :
5424 : /*
5425 : * Check to see if required parameters are set high enough on this server
5426 : * for various aspects of recovery operation.
5427 : *
5428 : * Note that all the parameters which this function tests need to be
5429 : * listed in Administrator's Overview section in high-availability.sgml.
5430 : * If you change them, don't forget to update the list.
5431 : */
5432 : static void
5433 482 : CheckRequiredParameterValues(void)
5434 : {
5435 : /*
5436 : * For archive recovery, the WAL must be generated with at least 'replica'
5437 : * wal_level.
5438 : */
5439 482 : if (ArchiveRecoveryRequested && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5440 : {
5441 4 : ereport(FATAL,
5442 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5443 : errmsg("WAL was generated with \"wal_level=minimal\", cannot continue recovering"),
5444 : errdetail("This happens if you temporarily set \"wal_level=minimal\" on the server."),
5445 : errhint("Use a backup taken after setting \"wal_level\" to higher than \"minimal\".")));
5446 : }
5447 :
5448 : /*
5449 : * For Hot Standby, the WAL must be generated with 'replica' mode, and we
5450 : * must have at least as many backend slots as the primary.
5451 : */
5452 478 : if (ArchiveRecoveryRequested && EnableHotStandby)
5453 : {
5454 : /* We ignore autovacuum_worker_slots when we make this test. */
5455 240 : RecoveryRequiresIntParameter("max_connections",
5456 : MaxConnections,
5457 240 : ControlFile->MaxConnections);
5458 240 : RecoveryRequiresIntParameter("max_worker_processes",
5459 : max_worker_processes,
5460 240 : ControlFile->max_worker_processes);
5461 240 : RecoveryRequiresIntParameter("max_wal_senders",
5462 : max_wal_senders,
5463 240 : ControlFile->max_wal_senders);
5464 240 : RecoveryRequiresIntParameter("max_prepared_transactions",
5465 : max_prepared_xacts,
5466 240 : ControlFile->max_prepared_xacts);
5467 240 : RecoveryRequiresIntParameter("max_locks_per_transaction",
5468 : max_locks_per_xact,
5469 240 : ControlFile->max_locks_per_xact);
5470 : }
5471 478 : }
5472 :
5473 : /*
5474 : * This must be called ONCE during postmaster or standalone-backend startup
5475 : */
5476 : void
5477 1762 : StartupXLOG(void)
5478 : {
5479 : XLogCtlInsert *Insert;
5480 : CheckPoint checkPoint;
5481 : bool wasShutdown;
5482 : bool didCrash;
5483 : bool haveTblspcMap;
5484 : bool haveBackupLabel;
5485 : XLogRecPtr EndOfLog;
5486 : TimeLineID EndOfLogTLI;
5487 : TimeLineID newTLI;
5488 : bool performedWalRecovery;
5489 : EndOfWalRecoveryInfo *endOfRecoveryInfo;
5490 : XLogRecPtr abortedRecPtr;
5491 : XLogRecPtr missingContrecPtr;
5492 : TransactionId oldestActiveXID;
5493 1762 : bool promoted = false;
5494 :
5495 : /*
5496 : * We should have an aux process resource owner to use, and we should not
5497 : * be in a transaction that's installed some other resowner.
5498 : */
5499 : Assert(AuxProcessResourceOwner != NULL);
5500 : Assert(CurrentResourceOwner == NULL ||
5501 : CurrentResourceOwner == AuxProcessResourceOwner);
5502 1762 : CurrentResourceOwner = AuxProcessResourceOwner;
5503 :
5504 : /*
5505 : * Check that contents look valid.
5506 : */
5507 1762 : if (!XRecOffIsValid(ControlFile->checkPoint))
5508 0 : ereport(FATAL,
5509 : (errcode(ERRCODE_DATA_CORRUPTED),
5510 : errmsg("control file contains invalid checkpoint location")));
5511 :
5512 1762 : switch (ControlFile->state)
5513 : {
5514 1356 : case DB_SHUTDOWNED:
5515 :
5516 : /*
5517 : * This is the expected case, so don't be chatty in standalone
5518 : * mode
5519 : */
5520 1356 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
5521 : (errmsg("database system was shut down at %s",
5522 : str_time(ControlFile->time))));
5523 1356 : break;
5524 :
5525 64 : case DB_SHUTDOWNED_IN_RECOVERY:
5526 64 : ereport(LOG,
5527 : (errmsg("database system was shut down in recovery at %s",
5528 : str_time(ControlFile->time))));
5529 64 : break;
5530 :
5531 0 : case DB_SHUTDOWNING:
5532 0 : ereport(LOG,
5533 : (errmsg("database system shutdown was interrupted; last known up at %s",
5534 : str_time(ControlFile->time))));
5535 0 : break;
5536 :
5537 0 : case DB_IN_CRASH_RECOVERY:
5538 0 : ereport(LOG,
5539 : (errmsg("database system was interrupted while in recovery at %s",
5540 : str_time(ControlFile->time)),
5541 : errhint("This probably means that some data is corrupted and"
5542 : " you will have to use the last backup for recovery.")));
5543 0 : break;
5544 :
5545 12 : case DB_IN_ARCHIVE_RECOVERY:
5546 12 : ereport(LOG,
5547 : (errmsg("database system was interrupted while in recovery at log time %s",
5548 : str_time(ControlFile->checkPointCopy.time)),
5549 : errhint("If this has occurred more than once some data might be corrupted"
5550 : " and you might need to choose an earlier recovery target.")));
5551 12 : break;
5552 :
5553 330 : case DB_IN_PRODUCTION:
5554 330 : ereport(LOG,
5555 : (errmsg("database system was interrupted; last known up at %s",
5556 : str_time(ControlFile->time))));
5557 330 : break;
5558 :
5559 0 : default:
5560 0 : ereport(FATAL,
5561 : (errcode(ERRCODE_DATA_CORRUPTED),
5562 : errmsg("control file contains invalid database cluster state")));
5563 : }
5564 :
5565 : /* This is just to allow attaching to startup process with a debugger */
5566 : #ifdef XLOG_REPLAY_DELAY
5567 : if (ControlFile->state != DB_SHUTDOWNED)
5568 : pg_usleep(60000000L);
5569 : #endif
5570 :
5571 : /*
5572 : * Verify that pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
5573 : * In cases where someone has performed a copy for PITR, these directories
5574 : * may have been excluded and need to be re-created.
5575 : */
5576 1762 : ValidateXLOGDirectoryStructure();
5577 :
5578 : /* Set up timeout handler needed to report startup progress. */
5579 1762 : if (!IsBootstrapProcessingMode())
5580 1672 : RegisterTimeout(STARTUP_PROGRESS_TIMEOUT,
5581 : startup_progress_timeout_handler);
5582 :
5583 : /*----------
5584 : * If we previously crashed, perform a couple of actions:
5585 : *
5586 : * - The pg_wal directory may still include some temporary WAL segments
5587 : * used when creating a new segment, so perform some clean up to not
5588 : * bloat this path. This is done first as there is no point to sync
5589 : * this temporary data.
5590 : *
5591 : * - There might be data which we had written, intending to fsync it, but
5592 : * which we had not actually fsync'd yet. Therefore, a power failure in
5593 : * the near future might cause earlier unflushed writes to be lost, even
5594 : * though more recent data written to disk from here on would be
5595 : * persisted. To avoid that, fsync the entire data directory.
5596 : */
5597 1762 : if (ControlFile->state != DB_SHUTDOWNED &&
5598 406 : ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
5599 : {
5600 342 : RemoveTempXlogFiles();
5601 342 : SyncDataDirectory();
5602 342 : didCrash = true;
5603 : }
5604 : else
5605 1420 : didCrash = false;
5606 :
5607 : /*
5608 : * Prepare for WAL recovery if needed.
5609 : *
5610 : * InitWalRecovery analyzes the control file and the backup label file, if
5611 : * any. It updates the in-memory ControlFile buffer according to the
5612 : * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5613 : * It also applies the tablespace map file, if any.
5614 : */
5615 1762 : InitWalRecovery(ControlFile, &wasShutdown,
5616 : &haveBackupLabel, &haveTblspcMap);
5617 1762 : checkPoint = ControlFile->checkPointCopy;
5618 :
5619 : /* initialize shared memory variables from the checkpoint record */
5620 1762 : TransamVariables->nextXid = checkPoint.nextXid;
5621 1762 : TransamVariables->nextOid = checkPoint.nextOid;
5622 1762 : TransamVariables->oidCount = 0;
5623 1762 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5624 1762 : AdvanceOldestClogXid(checkPoint.oldestXid);
5625 1762 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5626 1762 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5627 1762 : SetCommitTsLimit(checkPoint.oldestCommitTsXid,
5628 : checkPoint.newestCommitTsXid);
5629 1762 : XLogCtl->ckptFullXid = checkPoint.nextXid;
5630 :
5631 : /*
5632 : * Clear out any old relcache cache files. This is *necessary* if we do
5633 : * any WAL replay, since that would probably result in the cache files
5634 : * being out of sync with database reality. In theory we could leave them
5635 : * in place if the database had been cleanly shut down, but it seems
5636 : * safest to just remove them always and let them be rebuilt during the
5637 : * first backend startup. These files needs to be removed from all
5638 : * directories including pg_tblspc, however the symlinks are created only
5639 : * after reading tablespace_map file in case of archive recovery from
5640 : * backup, so needs to clear old relcache files here after creating
5641 : * symlinks.
5642 : */
5643 1762 : RelationCacheInitFileRemove();
5644 :
5645 : /*
5646 : * Initialize replication slots, before there's a chance to remove
5647 : * required resources.
5648 : */
5649 1762 : StartupReplicationSlots();
5650 :
5651 : /*
5652 : * Startup logical state, needs to be setup now so we have proper data
5653 : * during crash recovery.
5654 : */
5655 1760 : StartupReorderBuffer();
5656 :
5657 : /*
5658 : * Startup CLOG. This must be done after TransamVariables->nextXid has
5659 : * been initialized and before we accept connections or begin WAL replay.
5660 : */
5661 1760 : StartupCLOG();
5662 :
5663 : /*
5664 : * Startup MultiXact. We need to do this early to be able to replay
5665 : * truncations.
5666 : */
5667 1760 : StartupMultiXact();
5668 :
5669 : /*
5670 : * Ditto for commit timestamps. Activate the facility if the setting is
5671 : * enabled in the control file, as there should be no tracking of commit
5672 : * timestamps done when the setting was disabled. This facility can be
5673 : * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
5674 : */
5675 1760 : if (ControlFile->track_commit_timestamp)
5676 20 : StartupCommitTs();
5677 :
5678 : /*
5679 : * Recover knowledge about replay progress of known replication partners.
5680 : */
5681 1760 : StartupReplicationOrigin();
5682 :
5683 : /*
5684 : * Initialize unlogged LSN. On a clean shutdown, it's restored from the
5685 : * control file. On recovery, all unlogged relations are blown away, so
5686 : * the unlogged LSN counter can be reset too.
5687 : */
5688 1760 : if (ControlFile->state == DB_SHUTDOWNED)
5689 1342 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5690 1342 : ControlFile->unloggedLSN);
5691 : else
5692 418 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5693 : FirstNormalUnloggedLSN);
5694 :
5695 : /*
5696 : * Copy any missing timeline history files between 'now' and the recovery
5697 : * target timeline from archive to pg_wal. While we don't need those files
5698 : * ourselves - the history file of the recovery target timeline covers all
5699 : * the previous timelines in the history too - a cascading standby server
5700 : * might be interested in them. Or, if you archive the WAL from this
5701 : * server to a different archive than the primary, it'd be good for all
5702 : * the history files to get archived there after failover, so that you can
5703 : * use one of the old timelines as a PITR target. Timeline history files
5704 : * are small, so it's better to copy them unnecessarily than not copy them
5705 : * and regret later.
5706 : */
5707 1760 : restoreTimeLineHistoryFiles(checkPoint.ThisTimeLineID, recoveryTargetTLI);
5708 :
5709 : /*
5710 : * Before running in recovery, scan pg_twophase and fill in its status to
5711 : * be able to work on entries generated by redo. Doing a scan before
5712 : * taking any recovery action has the merit to discard any 2PC files that
5713 : * are newer than the first record to replay, saving from any conflicts at
5714 : * replay. This avoids as well any subsequent scans when doing recovery
5715 : * of the on-disk two-phase data.
5716 : */
5717 1760 : restoreTwoPhaseData();
5718 :
5719 : /*
5720 : * When starting with crash recovery, reset pgstat data - it might not be
5721 : * valid. Otherwise restore pgstat data. It's safe to do this here,
5722 : * because postmaster will not yet have started any other processes.
5723 : *
5724 : * NB: Restoring replication slot stats relies on slot state to have
5725 : * already been restored from disk.
5726 : *
5727 : * TODO: With a bit of extra work we could just start with a pgstat file
5728 : * associated with the checkpoint redo location we're starting from.
5729 : */
5730 1760 : if (didCrash)
5731 342 : pgstat_discard_stats();
5732 : else
5733 1418 : pgstat_restore_stats();
5734 :
5735 1760 : lastFullPageWrites = checkPoint.fullPageWrites;
5736 :
5737 1760 : RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
5738 1760 : doPageWrites = lastFullPageWrites;
5739 :
5740 : /* REDO */
5741 1760 : if (InRecovery)
5742 : {
5743 : /* Initialize state for RecoveryInProgress() */
5744 418 : SpinLockAcquire(&XLogCtl->info_lck);
5745 418 : if (InArchiveRecovery)
5746 218 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
5747 : else
5748 200 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5749 418 : SpinLockRelease(&XLogCtl->info_lck);
5750 :
5751 : /*
5752 : * Update pg_control to show that we are recovering and to show the
5753 : * selected checkpoint as the place we are starting from. We also mark
5754 : * pg_control with any minimum recovery stop point obtained from a
5755 : * backup history file.
5756 : *
5757 : * No need to hold ControlFileLock yet, we aren't up far enough.
5758 : */
5759 418 : UpdateControlFile();
5760 :
5761 : /*
5762 : * If there was a backup label file, it's done its job and the info
5763 : * has now been propagated into pg_control. We must get rid of the
5764 : * label file so that if we crash during recovery, we'll pick up at
5765 : * the latest recovery restartpoint instead of going all the way back
5766 : * to the backup start point. It seems prudent though to just rename
5767 : * the file out of the way rather than delete it completely.
5768 : */
5769 418 : if (haveBackupLabel)
5770 : {
5771 138 : unlink(BACKUP_LABEL_OLD);
5772 138 : durable_rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD, FATAL);
5773 : }
5774 :
5775 : /*
5776 : * If there was a tablespace_map file, it's done its job and the
5777 : * symlinks have been created. We must get rid of the map file so
5778 : * that if we crash during recovery, we don't create symlinks again.
5779 : * It seems prudent though to just rename the file out of the way
5780 : * rather than delete it completely.
5781 : */
5782 418 : if (haveTblspcMap)
5783 : {
5784 4 : unlink(TABLESPACE_MAP_OLD);
5785 4 : durable_rename(TABLESPACE_MAP, TABLESPACE_MAP_OLD, FATAL);
5786 : }
5787 :
5788 : /*
5789 : * Initialize our local copy of minRecoveryPoint. When doing crash
5790 : * recovery we want to replay up to the end of WAL. Particularly, in
5791 : * the case of a promoted standby minRecoveryPoint value in the
5792 : * control file is only updated after the first checkpoint. However,
5793 : * if the instance crashes before the first post-recovery checkpoint
5794 : * is completed then recovery will use a stale location causing the
5795 : * startup process to think that there are still invalid page
5796 : * references when checking for data consistency.
5797 : */
5798 418 : if (InArchiveRecovery)
5799 : {
5800 218 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
5801 218 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
5802 : }
5803 : else
5804 : {
5805 200 : LocalMinRecoveryPoint = InvalidXLogRecPtr;
5806 200 : LocalMinRecoveryPointTLI = 0;
5807 : }
5808 :
5809 : /* Check that the GUCs used to generate the WAL allow recovery */
5810 418 : CheckRequiredParameterValues();
5811 :
5812 : /*
5813 : * We're in recovery, so unlogged relations may be trashed and must be
5814 : * reset. This should be done BEFORE allowing Hot Standby
5815 : * connections, so that read-only backends don't try to read whatever
5816 : * garbage is left over from before.
5817 : */
5818 418 : ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
5819 :
5820 : /*
5821 : * Likewise, delete any saved transaction snapshot files that got left
5822 : * behind by crashed backends.
5823 : */
5824 418 : DeleteAllExportedSnapshotFiles();
5825 :
5826 : /*
5827 : * Initialize for Hot Standby, if enabled. We won't let backends in
5828 : * yet, not until we've reached the min recovery point specified in
5829 : * control file and we've established a recovery snapshot from a
5830 : * running-xacts WAL record.
5831 : */
5832 418 : if (ArchiveRecoveryRequested && EnableHotStandby)
5833 : {
5834 : TransactionId *xids;
5835 : int nxids;
5836 :
5837 206 : ereport(DEBUG1,
5838 : (errmsg_internal("initializing for hot standby")));
5839 :
5840 206 : InitRecoveryTransactionEnvironment();
5841 :
5842 206 : if (wasShutdown)
5843 52 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
5844 : else
5845 154 : oldestActiveXID = checkPoint.oldestActiveXid;
5846 : Assert(TransactionIdIsValid(oldestActiveXID));
5847 :
5848 : /* Tell procarray about the range of xids it has to deal with */
5849 206 : ProcArrayInitRecovery(XidFromFullTransactionId(TransamVariables->nextXid));
5850 :
5851 : /*
5852 : * Startup subtrans only. CLOG, MultiXact and commit timestamp
5853 : * have already been started up and other SLRUs are not maintained
5854 : * during recovery and need not be started yet.
5855 : */
5856 206 : StartupSUBTRANS(oldestActiveXID);
5857 :
5858 : /*
5859 : * If we're beginning at a shutdown checkpoint, we know that
5860 : * nothing was running on the primary at this point. So fake-up an
5861 : * empty running-xacts record and use that here and now. Recover
5862 : * additional standby state for prepared transactions.
5863 : */
5864 206 : if (wasShutdown)
5865 : {
5866 : RunningTransactionsData running;
5867 : TransactionId latestCompletedXid;
5868 :
5869 : /* Update pg_subtrans entries for any prepared transactions */
5870 52 : StandbyRecoverPreparedTransactions();
5871 :
5872 : /*
5873 : * Construct a RunningTransactions snapshot representing a
5874 : * shut down server, with only prepared transactions still
5875 : * alive. We're never overflowed at this point because all
5876 : * subxids are listed with their parent prepared transactions.
5877 : */
5878 52 : running.xcnt = nxids;
5879 52 : running.subxcnt = 0;
5880 52 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
5881 52 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
5882 52 : running.oldestRunningXid = oldestActiveXID;
5883 52 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
5884 52 : TransactionIdRetreat(latestCompletedXid);
5885 : Assert(TransactionIdIsNormal(latestCompletedXid));
5886 52 : running.latestCompletedXid = latestCompletedXid;
5887 52 : running.xids = xids;
5888 :
5889 52 : ProcArrayApplyRecoveryInfo(&running);
5890 : }
5891 : }
5892 :
5893 : /*
5894 : * We're all set for replaying the WAL now. Do it.
5895 : */
5896 418 : PerformWalRecovery();
5897 306 : performedWalRecovery = true;
5898 : }
5899 : else
5900 1342 : performedWalRecovery = false;
5901 :
5902 : /*
5903 : * Finish WAL recovery.
5904 : */
5905 1648 : endOfRecoveryInfo = FinishWalRecovery();
5906 1648 : EndOfLog = endOfRecoveryInfo->endOfLog;
5907 1648 : EndOfLogTLI = endOfRecoveryInfo->endOfLogTLI;
5908 1648 : abortedRecPtr = endOfRecoveryInfo->abortedRecPtr;
5909 1648 : missingContrecPtr = endOfRecoveryInfo->missingContrecPtr;
5910 :
5911 : /*
5912 : * Reset ps status display, so as no information related to recovery shows
5913 : * up.
5914 : */
5915 1648 : set_ps_display("");
5916 :
5917 : /*
5918 : * When recovering from a backup (we are in recovery, and archive recovery
5919 : * was requested), complain if we did not roll forward far enough to reach
5920 : * the point where the database is consistent. For regular online
5921 : * backup-from-primary, that means reaching the end-of-backup WAL record
5922 : * (at which point we reset backupStartPoint to be Invalid), for
5923 : * backup-from-replica (which can't inject records into the WAL stream),
5924 : * that point is when we reach the minRecoveryPoint in pg_control (which
5925 : * we purposefully copy last when backing up from a replica). For
5926 : * pg_rewind (which creates a backup_label with a method of "pg_rewind")
5927 : * or snapshot-style backups (which don't), backupEndRequired will be set
5928 : * to false.
5929 : *
5930 : * Note: it is indeed okay to look at the local variable
5931 : * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
5932 : * might be further ahead --- ControlFile->minRecoveryPoint cannot have
5933 : * been advanced beyond the WAL we processed.
5934 : */
5935 1648 : if (InRecovery &&
5936 306 : (EndOfLog < LocalMinRecoveryPoint ||
5937 306 : !XLogRecPtrIsInvalid(ControlFile->backupStartPoint)))
5938 : {
5939 : /*
5940 : * Ran off end of WAL before reaching end-of-backup WAL record, or
5941 : * minRecoveryPoint. That's a bad sign, indicating that you tried to
5942 : * recover from an online backup but never called pg_backup_stop(), or
5943 : * you didn't archive all the WAL needed.
5944 : */
5945 0 : if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
5946 : {
5947 0 : if (!XLogRecPtrIsInvalid(ControlFile->backupStartPoint) || ControlFile->backupEndRequired)
5948 0 : ereport(FATAL,
5949 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5950 : errmsg("WAL ends before end of online backup"),
5951 : errhint("All WAL generated while online backup was taken must be available at recovery.")));
5952 : else
5953 0 : ereport(FATAL,
5954 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5955 : errmsg("WAL ends before consistent recovery point")));
5956 : }
5957 : }
5958 :
5959 : /*
5960 : * Reset unlogged relations to the contents of their INIT fork. This is
5961 : * done AFTER recovery is complete so as to include any unlogged relations
5962 : * created during recovery, but BEFORE recovery is marked as having
5963 : * completed successfully. Otherwise we'd not retry if any of the post
5964 : * end-of-recovery steps fail.
5965 : */
5966 1648 : if (InRecovery)
5967 306 : ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
5968 :
5969 : /*
5970 : * Pre-scan prepared transactions to find out the range of XIDs present.
5971 : * This information is not quite needed yet, but it is positioned here so
5972 : * as potential problems are detected before any on-disk change is done.
5973 : */
5974 1648 : oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
5975 :
5976 : /*
5977 : * Allow ordinary WAL segment creation before possibly switching to a new
5978 : * timeline, which creates a new segment, and after the last ReadRecord().
5979 : */
5980 1648 : SetInstallXLogFileSegmentActive();
5981 :
5982 : /*
5983 : * Consider whether we need to assign a new timeline ID.
5984 : *
5985 : * If we did archive recovery, we always assign a new ID. This handles a
5986 : * couple of issues. If we stopped short of the end of WAL during
5987 : * recovery, then we are clearly generating a new timeline and must assign
5988 : * it a unique new ID. Even if we ran to the end, modifying the current
5989 : * last segment is problematic because it may result in trying to
5990 : * overwrite an already-archived copy of that segment, and we encourage
5991 : * DBAs to make their archive_commands reject that. We can dodge the
5992 : * problem by making the new active segment have a new timeline ID.
5993 : *
5994 : * In a normal crash recovery, we can just extend the timeline we were in.
5995 : */
5996 1648 : newTLI = endOfRecoveryInfo->lastRecTLI;
5997 1648 : if (ArchiveRecoveryRequested)
5998 : {
5999 98 : newTLI = findNewestTimeLine(recoveryTargetTLI) + 1;
6000 98 : ereport(LOG,
6001 : (errmsg("selected new timeline ID: %u", newTLI)));
6002 :
6003 : /*
6004 : * Make a writable copy of the last WAL segment. (Note that we also
6005 : * have a copy of the last block of the old WAL in
6006 : * endOfRecovery->lastPage; we will use that below.)
6007 : */
6008 98 : XLogInitNewTimeline(EndOfLogTLI, EndOfLog, newTLI);
6009 :
6010 : /*
6011 : * Remove the signal files out of the way, so that we don't
6012 : * accidentally re-enter archive recovery mode in a subsequent crash.
6013 : */
6014 98 : if (endOfRecoveryInfo->standby_signal_file_found)
6015 92 : durable_unlink(STANDBY_SIGNAL_FILE, FATAL);
6016 :
6017 98 : if (endOfRecoveryInfo->recovery_signal_file_found)
6018 6 : durable_unlink(RECOVERY_SIGNAL_FILE, FATAL);
6019 :
6020 : /*
6021 : * Write the timeline history file, and have it archived. After this
6022 : * point (or rather, as soon as the file is archived), the timeline
6023 : * will appear as "taken" in the WAL archive and to any standby
6024 : * servers. If we crash before actually switching to the new
6025 : * timeline, standby servers will nevertheless think that we switched
6026 : * to the new timeline, and will try to connect to the new timeline.
6027 : * To minimize the window for that, try to do as little as possible
6028 : * between here and writing the end-of-recovery record.
6029 : */
6030 98 : writeTimeLineHistory(newTLI, recoveryTargetTLI,
6031 : EndOfLog, endOfRecoveryInfo->recoveryStopReason);
6032 :
6033 98 : ereport(LOG,
6034 : (errmsg("archive recovery complete")));
6035 : }
6036 :
6037 : /* Save the selected TimeLineID in shared memory, too */
6038 1648 : SpinLockAcquire(&XLogCtl->info_lck);
6039 1648 : XLogCtl->InsertTimeLineID = newTLI;
6040 1648 : XLogCtl->PrevTimeLineID = endOfRecoveryInfo->lastRecTLI;
6041 1648 : SpinLockRelease(&XLogCtl->info_lck);
6042 :
6043 : /*
6044 : * Actually, if WAL ended in an incomplete record, skip the parts that
6045 : * made it through and start writing after the portion that persisted.
6046 : * (It's critical to first write an OVERWRITE_CONTRECORD message, which
6047 : * we'll do as soon as we're open for writing new WAL.)
6048 : */
6049 1648 : if (!XLogRecPtrIsInvalid(missingContrecPtr))
6050 : {
6051 : /*
6052 : * We should only have a missingContrecPtr if we're not switching to a
6053 : * new timeline. When a timeline switch occurs, WAL is copied from the
6054 : * old timeline to the new only up to the end of the last complete
6055 : * record, so there can't be an incomplete WAL record that we need to
6056 : * disregard.
6057 : */
6058 : Assert(newTLI == endOfRecoveryInfo->lastRecTLI);
6059 : Assert(!XLogRecPtrIsInvalid(abortedRecPtr));
6060 22 : EndOfLog = missingContrecPtr;
6061 : }
6062 :
6063 : /*
6064 : * Prepare to write WAL starting at EndOfLog location, and init xlog
6065 : * buffer cache using the block containing the last record from the
6066 : * previous incarnation.
6067 : */
6068 1648 : Insert = &XLogCtl->Insert;
6069 1648 : Insert->PrevBytePos = XLogRecPtrToBytePos(endOfRecoveryInfo->lastRec);
6070 1648 : Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
6071 :
6072 : /*
6073 : * Tricky point here: lastPage contains the *last* block that the LastRec
6074 : * record spans, not the one it starts in. The last block is indeed the
6075 : * one we want to use.
6076 : */
6077 1648 : if (EndOfLog % XLOG_BLCKSZ != 0)
6078 : {
6079 : char *page;
6080 : int len;
6081 : int firstIdx;
6082 :
6083 1584 : firstIdx = XLogRecPtrToBufIdx(EndOfLog);
6084 1584 : len = EndOfLog - endOfRecoveryInfo->lastPageBeginPtr;
6085 : Assert(len < XLOG_BLCKSZ);
6086 :
6087 : /* Copy the valid part of the last block, and zero the rest */
6088 1584 : page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
6089 1584 : memcpy(page, endOfRecoveryInfo->lastPage, len);
6090 1584 : memset(page + len, 0, XLOG_BLCKSZ - len);
6091 :
6092 1584 : pg_atomic_write_u64(&XLogCtl->xlblocks[firstIdx], endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6093 1584 : XLogCtl->InitializedUpTo = endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ;
6094 : }
6095 : else
6096 : {
6097 : /*
6098 : * There is no partial block to copy. Just set InitializedUpTo, and
6099 : * let the first attempt to insert a log record to initialize the next
6100 : * buffer.
6101 : */
6102 64 : XLogCtl->InitializedUpTo = EndOfLog;
6103 : }
6104 :
6105 : /*
6106 : * Update local and shared status. This is OK to do without any locks
6107 : * because no other process can be reading or writing WAL yet.
6108 : */
6109 1648 : LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
6110 1648 : pg_atomic_write_u64(&XLogCtl->logInsertResult, EndOfLog);
6111 1648 : pg_atomic_write_u64(&XLogCtl->logWriteResult, EndOfLog);
6112 1648 : pg_atomic_write_u64(&XLogCtl->logFlushResult, EndOfLog);
6113 1648 : XLogCtl->LogwrtRqst.Write = EndOfLog;
6114 1648 : XLogCtl->LogwrtRqst.Flush = EndOfLog;
6115 :
6116 : /*
6117 : * Preallocate additional log files, if wanted.
6118 : */
6119 1648 : PreallocXlogFiles(EndOfLog, newTLI);
6120 :
6121 : /*
6122 : * Okay, we're officially UP.
6123 : */
6124 1648 : InRecovery = false;
6125 :
6126 : /* start the archive_timeout timer and LSN running */
6127 1648 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
6128 1648 : XLogCtl->lastSegSwitchLSN = EndOfLog;
6129 :
6130 : /* also initialize latestCompletedXid, to nextXid - 1 */
6131 1648 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
6132 1648 : TransamVariables->latestCompletedXid = TransamVariables->nextXid;
6133 1648 : FullTransactionIdRetreat(&TransamVariables->latestCompletedXid);
6134 1648 : LWLockRelease(ProcArrayLock);
6135 :
6136 : /*
6137 : * Start up subtrans, if not already done for hot standby. (commit
6138 : * timestamps are started below, if necessary.)
6139 : */
6140 1648 : if (standbyState == STANDBY_DISABLED)
6141 1550 : StartupSUBTRANS(oldestActiveXID);
6142 :
6143 : /*
6144 : * Perform end of recovery actions for any SLRUs that need it.
6145 : */
6146 1648 : TrimCLOG();
6147 1648 : TrimMultiXact();
6148 :
6149 : /*
6150 : * Reload shared-memory state for prepared transactions. This needs to
6151 : * happen before renaming the last partial segment of the old timeline as
6152 : * it may be possible that we have to recover some transactions from it.
6153 : */
6154 1648 : RecoverPreparedTransactions();
6155 :
6156 : /* Shut down xlogreader */
6157 1648 : ShutdownWalRecovery();
6158 :
6159 : /* Enable WAL writes for this backend only. */
6160 1648 : LocalSetXLogInsertAllowed();
6161 :
6162 : /* If necessary, write overwrite-contrecord before doing anything else */
6163 1648 : if (!XLogRecPtrIsInvalid(abortedRecPtr))
6164 : {
6165 : Assert(!XLogRecPtrIsInvalid(missingContrecPtr));
6166 22 : CreateOverwriteContrecordRecord(abortedRecPtr, missingContrecPtr, newTLI);
6167 : }
6168 :
6169 : /*
6170 : * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
6171 : * record before resource manager writes cleanup WAL records or checkpoint
6172 : * record is written.
6173 : */
6174 1648 : Insert->fullPageWrites = lastFullPageWrites;
6175 1648 : UpdateFullPageWrites();
6176 :
6177 : /*
6178 : * Emit checkpoint or end-of-recovery record in XLOG, if required.
6179 : */
6180 1648 : if (performedWalRecovery)
6181 306 : promoted = PerformRecoveryXLogAction();
6182 :
6183 : /*
6184 : * If any of the critical GUCs have changed, log them before we allow
6185 : * backends to write WAL.
6186 : */
6187 1648 : XLogReportParameters();
6188 :
6189 : /* If this is archive recovery, perform post-recovery cleanup actions. */
6190 1648 : if (ArchiveRecoveryRequested)
6191 98 : CleanupAfterArchiveRecovery(EndOfLogTLI, EndOfLog, newTLI);
6192 :
6193 : /*
6194 : * Local WAL inserts enabled, so it's time to finish initialization of
6195 : * commit timestamp.
6196 : */
6197 1648 : CompleteCommitTsInitialization();
6198 :
6199 : /*
6200 : * All done with end-of-recovery actions.
6201 : *
6202 : * Now allow backends to write WAL and update the control file status in
6203 : * consequence. SharedRecoveryState, that controls if backends can write
6204 : * WAL, is updated while holding ControlFileLock to prevent other backends
6205 : * to look at an inconsistent state of the control file in shared memory.
6206 : * There is still a small window during which backends can write WAL and
6207 : * the control file is still referring to a system not in DB_IN_PRODUCTION
6208 : * state while looking at the on-disk control file.
6209 : *
6210 : * Also, we use info_lck to update SharedRecoveryState to ensure that
6211 : * there are no race conditions concerning visibility of other recent
6212 : * updates to shared memory.
6213 : */
6214 1648 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6215 1648 : ControlFile->state = DB_IN_PRODUCTION;
6216 :
6217 1648 : SpinLockAcquire(&XLogCtl->info_lck);
6218 1648 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_DONE;
6219 1648 : SpinLockRelease(&XLogCtl->info_lck);
6220 :
6221 1648 : UpdateControlFile();
6222 1648 : LWLockRelease(ControlFileLock);
6223 :
6224 : /*
6225 : * Shutdown the recovery environment. This must occur after
6226 : * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
6227 : * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
6228 : * any session building a snapshot will not rely on KnownAssignedXids as
6229 : * RecoveryInProgress() would return false at this stage. This is
6230 : * particularly critical for prepared 2PC transactions, that would still
6231 : * need to be included in snapshots once recovery has ended.
6232 : */
6233 1648 : if (standbyState != STANDBY_DISABLED)
6234 98 : ShutdownRecoveryTransactionEnvironment();
6235 :
6236 : /*
6237 : * If there were cascading standby servers connected to us, nudge any wal
6238 : * sender processes to notice that we've been promoted.
6239 : */
6240 1648 : WalSndWakeup(true, true);
6241 :
6242 : /*
6243 : * If this was a promotion, request an (online) checkpoint now. This isn't
6244 : * required for consistency, but the last restartpoint might be far back,
6245 : * and in case of a crash, recovering from it might take a longer than is
6246 : * appropriate now that we're not in standby mode anymore.
6247 : */
6248 1648 : if (promoted)
6249 84 : RequestCheckpoint(CHECKPOINT_FORCE);
6250 1648 : }
6251 :
6252 : /*
6253 : * Callback from PerformWalRecovery(), called when we switch from crash
6254 : * recovery to archive recovery mode. Updates the control file accordingly.
6255 : */
6256 : void
6257 4 : SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr, TimeLineID replayTLI)
6258 : {
6259 : /* initialize minRecoveryPoint to this record */
6260 4 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6261 4 : ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6262 4 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6263 : {
6264 4 : ControlFile->minRecoveryPoint = EndRecPtr;
6265 4 : ControlFile->minRecoveryPointTLI = replayTLI;
6266 : }
6267 : /* update local copy */
6268 4 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6269 4 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6270 :
6271 : /*
6272 : * The startup process can update its local copy of minRecoveryPoint from
6273 : * this point.
6274 : */
6275 4 : updateMinRecoveryPoint = true;
6276 :
6277 4 : UpdateControlFile();
6278 :
6279 : /*
6280 : * We update SharedRecoveryState while holding the lock on ControlFileLock
6281 : * so both states are consistent in shared memory.
6282 : */
6283 4 : SpinLockAcquire(&XLogCtl->info_lck);
6284 4 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6285 4 : SpinLockRelease(&XLogCtl->info_lck);
6286 :
6287 4 : LWLockRelease(ControlFileLock);
6288 4 : }
6289 :
6290 : /*
6291 : * Callback from PerformWalRecovery(), called when we reach the end of backup.
6292 : * Updates the control file accordingly.
6293 : */
6294 : void
6295 138 : ReachedEndOfBackup(XLogRecPtr EndRecPtr, TimeLineID tli)
6296 : {
6297 : /*
6298 : * We have reached the end of base backup, as indicated by pg_control. The
6299 : * data on disk is now consistent (unless minRecoveryPoint is further
6300 : * ahead, which can happen if we crashed during previous recovery). Reset
6301 : * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
6302 : * make sure we don't allow starting up at an earlier point even if
6303 : * recovery is stopped and restarted soon after this.
6304 : */
6305 138 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6306 :
6307 138 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6308 : {
6309 130 : ControlFile->minRecoveryPoint = EndRecPtr;
6310 130 : ControlFile->minRecoveryPointTLI = tli;
6311 : }
6312 :
6313 138 : ControlFile->backupStartPoint = InvalidXLogRecPtr;
6314 138 : ControlFile->backupEndPoint = InvalidXLogRecPtr;
6315 138 : ControlFile->backupEndRequired = false;
6316 138 : UpdateControlFile();
6317 :
6318 138 : LWLockRelease(ControlFileLock);
6319 138 : }
6320 :
6321 : /*
6322 : * Perform whatever XLOG actions are necessary at end of REDO.
6323 : *
6324 : * The goal here is to make sure that we'll be able to recover properly if
6325 : * we crash again. If we choose to write a checkpoint, we'll write a shutdown
6326 : * checkpoint rather than an on-line one. This is not particularly critical,
6327 : * but since we may be assigning a new TLI, using a shutdown checkpoint allows
6328 : * us to have the rule that TLI only changes in shutdown checkpoints, which
6329 : * allows some extra error checking in xlog_redo.
6330 : */
6331 : static bool
6332 306 : PerformRecoveryXLogAction(void)
6333 : {
6334 306 : bool promoted = false;
6335 :
6336 : /*
6337 : * Perform a checkpoint to update all our recovery activity to disk.
6338 : *
6339 : * Note that we write a shutdown checkpoint rather than an on-line one.
6340 : * This is not particularly critical, but since we may be assigning a new
6341 : * TLI, using a shutdown checkpoint allows us to have the rule that TLI
6342 : * only changes in shutdown checkpoints, which allows some extra error
6343 : * checking in xlog_redo.
6344 : *
6345 : * In promotion, only create a lightweight end-of-recovery record instead
6346 : * of a full checkpoint. A checkpoint is requested later, after we're
6347 : * fully out of recovery mode and already accepting queries.
6348 : */
6349 404 : if (ArchiveRecoveryRequested && IsUnderPostmaster &&
6350 98 : PromoteIsTriggered())
6351 : {
6352 84 : promoted = true;
6353 :
6354 : /*
6355 : * Insert a special WAL record to mark the end of recovery, since we
6356 : * aren't doing a checkpoint. That means that the checkpointer process
6357 : * may likely be in the middle of a time-smoothed restartpoint and
6358 : * could continue to be for minutes after this. That sounds strange,
6359 : * but the effect is roughly the same and it would be stranger to try
6360 : * to come out of the restartpoint and then checkpoint. We request a
6361 : * checkpoint later anyway, just for safety.
6362 : */
6363 84 : CreateEndOfRecoveryRecord();
6364 : }
6365 : else
6366 : {
6367 222 : RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
6368 : CHECKPOINT_IMMEDIATE |
6369 : CHECKPOINT_WAIT);
6370 : }
6371 :
6372 306 : return promoted;
6373 : }
6374 :
6375 : /*
6376 : * Is the system still in recovery?
6377 : *
6378 : * Unlike testing InRecovery, this works in any process that's connected to
6379 : * shared memory.
6380 : */
6381 : bool
6382 169708662 : RecoveryInProgress(void)
6383 : {
6384 : /*
6385 : * We check shared state each time only until we leave recovery mode. We
6386 : * can't re-enter recovery, so there's no need to keep checking after the
6387 : * shared variable has once been seen false.
6388 : */
6389 169708662 : if (!LocalRecoveryInProgress)
6390 164502432 : return false;
6391 : else
6392 : {
6393 : /*
6394 : * use volatile pointer to make sure we make a fresh read of the
6395 : * shared variable.
6396 : */
6397 5206230 : volatile XLogCtlData *xlogctl = XLogCtl;
6398 :
6399 5206230 : LocalRecoveryInProgress = (xlogctl->SharedRecoveryState != RECOVERY_STATE_DONE);
6400 :
6401 : /*
6402 : * Note: We don't need a memory barrier when we're still in recovery.
6403 : * We might exit recovery immediately after return, so the caller
6404 : * can't rely on 'true' meaning that we're still in recovery anyway.
6405 : */
6406 :
6407 5206230 : return LocalRecoveryInProgress;
6408 : }
6409 : }
6410 :
6411 : /*
6412 : * Returns current recovery state from shared memory.
6413 : *
6414 : * This returned state is kept consistent with the contents of the control
6415 : * file. See details about the possible values of RecoveryState in xlog.h.
6416 : */
6417 : RecoveryState
6418 122 : GetRecoveryState(void)
6419 : {
6420 : RecoveryState retval;
6421 :
6422 122 : SpinLockAcquire(&XLogCtl->info_lck);
6423 122 : retval = XLogCtl->SharedRecoveryState;
6424 122 : SpinLockRelease(&XLogCtl->info_lck);
6425 :
6426 122 : return retval;
6427 : }
6428 :
6429 : /*
6430 : * Is this process allowed to insert new WAL records?
6431 : *
6432 : * Ordinarily this is essentially equivalent to !RecoveryInProgress().
6433 : * But we also have provisions for forcing the result "true" or "false"
6434 : * within specific processes regardless of the global state.
6435 : */
6436 : bool
6437 59202688 : XLogInsertAllowed(void)
6438 : {
6439 : /*
6440 : * If value is "unconditionally true" or "unconditionally false", just
6441 : * return it. This provides the normal fast path once recovery is known
6442 : * done.
6443 : */
6444 59202688 : if (LocalXLogInsertAllowed >= 0)
6445 58996166 : return (bool) LocalXLogInsertAllowed;
6446 :
6447 : /*
6448 : * Else, must check to see if we're still in recovery.
6449 : */
6450 206522 : if (RecoveryInProgress())
6451 190904 : return false;
6452 :
6453 : /*
6454 : * On exit from recovery, reset to "unconditionally true", since there is
6455 : * no need to keep checking.
6456 : */
6457 15618 : LocalXLogInsertAllowed = 1;
6458 15618 : return true;
6459 : }
6460 :
6461 : /*
6462 : * Make XLogInsertAllowed() return true in the current process only.
6463 : *
6464 : * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
6465 : * and even call LocalSetXLogInsertAllowed() again after that.
6466 : *
6467 : * Returns the previous value of LocalXLogInsertAllowed.
6468 : */
6469 : static int
6470 1706 : LocalSetXLogInsertAllowed(void)
6471 : {
6472 1706 : int oldXLogAllowed = LocalXLogInsertAllowed;
6473 :
6474 1706 : LocalXLogInsertAllowed = 1;
6475 :
6476 1706 : return oldXLogAllowed;
6477 : }
6478 :
6479 : /*
6480 : * Return the current Redo pointer from shared memory.
6481 : *
6482 : * As a side-effect, the local RedoRecPtr copy is updated.
6483 : */
6484 : XLogRecPtr
6485 574278 : GetRedoRecPtr(void)
6486 : {
6487 : XLogRecPtr ptr;
6488 :
6489 : /*
6490 : * The possibly not up-to-date copy in XlogCtl is enough. Even if we
6491 : * grabbed a WAL insertion lock to read the authoritative value in
6492 : * Insert->RedoRecPtr, someone might update it just after we've released
6493 : * the lock.
6494 : */
6495 574278 : SpinLockAcquire(&XLogCtl->info_lck);
6496 574278 : ptr = XLogCtl->RedoRecPtr;
6497 574278 : SpinLockRelease(&XLogCtl->info_lck);
6498 :
6499 574278 : if (RedoRecPtr < ptr)
6500 2728 : RedoRecPtr = ptr;
6501 :
6502 574278 : return RedoRecPtr;
6503 : }
6504 :
6505 : /*
6506 : * Return information needed to decide whether a modified block needs a
6507 : * full-page image to be included in the WAL record.
6508 : *
6509 : * The returned values are cached copies from backend-private memory, and
6510 : * possibly out-of-date or, indeed, uninitialized, in which case they will
6511 : * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6512 : * re-check them against up-to-date values, while holding the WAL insert lock.
6513 : */
6514 : void
6515 28698472 : GetFullPageWriteInfo(XLogRecPtr *RedoRecPtr_p, bool *doPageWrites_p)
6516 : {
6517 28698472 : *RedoRecPtr_p = RedoRecPtr;
6518 28698472 : *doPageWrites_p = doPageWrites;
6519 28698472 : }
6520 :
6521 : /*
6522 : * GetInsertRecPtr -- Returns the current insert position.
6523 : *
6524 : * NOTE: The value *actually* returned is the position of the last full
6525 : * xlog page. It lags behind the real insert position by at most 1 page.
6526 : * For that, we don't need to scan through WAL insertion locks, and an
6527 : * approximation is enough for the current usage of this function.
6528 : */
6529 : XLogRecPtr
6530 12384 : GetInsertRecPtr(void)
6531 : {
6532 : XLogRecPtr recptr;
6533 :
6534 12384 : SpinLockAcquire(&XLogCtl->info_lck);
6535 12384 : recptr = XLogCtl->LogwrtRqst.Write;
6536 12384 : SpinLockRelease(&XLogCtl->info_lck);
6537 :
6538 12384 : return recptr;
6539 : }
6540 :
6541 : /*
6542 : * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6543 : * position known to be fsync'd to disk. This should only be used on a
6544 : * system that is known not to be in recovery.
6545 : */
6546 : XLogRecPtr
6547 349208 : GetFlushRecPtr(TimeLineID *insertTLI)
6548 : {
6549 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6550 :
6551 349208 : RefreshXLogWriteResult(LogwrtResult);
6552 :
6553 : /*
6554 : * If we're writing and flushing WAL, the time line can't be changing, so
6555 : * no lock is required.
6556 : */
6557 349208 : if (insertTLI)
6558 52244 : *insertTLI = XLogCtl->InsertTimeLineID;
6559 :
6560 349208 : return LogwrtResult.Flush;
6561 : }
6562 :
6563 : /*
6564 : * GetWALInsertionTimeLine -- Returns the current timeline of a system that
6565 : * is not in recovery.
6566 : */
6567 : TimeLineID
6568 214236 : GetWALInsertionTimeLine(void)
6569 : {
6570 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6571 :
6572 : /* Since the value can't be changing, no lock is required. */
6573 214236 : return XLogCtl->InsertTimeLineID;
6574 : }
6575 :
6576 : /*
6577 : * GetWALInsertionTimeLineIfSet -- If the system is not in recovery, returns
6578 : * the WAL insertion timeline; else, returns 0. Wherever possible, use
6579 : * GetWALInsertionTimeLine() instead, since it's cheaper. Note that this
6580 : * function decides recovery has ended as soon as the insert TLI is set, which
6581 : * happens before we set XLogCtl->SharedRecoveryState to RECOVERY_STATE_DONE.
6582 : */
6583 : TimeLineID
6584 0 : GetWALInsertionTimeLineIfSet(void)
6585 : {
6586 : TimeLineID insertTLI;
6587 :
6588 0 : SpinLockAcquire(&XLogCtl->info_lck);
6589 0 : insertTLI = XLogCtl->InsertTimeLineID;
6590 0 : SpinLockRelease(&XLogCtl->info_lck);
6591 :
6592 0 : return insertTLI;
6593 : }
6594 :
6595 : /*
6596 : * GetLastImportantRecPtr -- Returns the LSN of the last important record
6597 : * inserted. All records not explicitly marked as unimportant are considered
6598 : * important.
6599 : *
6600 : * The LSN is determined by computing the maximum of
6601 : * WALInsertLocks[i].lastImportantAt.
6602 : */
6603 : XLogRecPtr
6604 2292 : GetLastImportantRecPtr(void)
6605 : {
6606 2292 : XLogRecPtr res = InvalidXLogRecPtr;
6607 : int i;
6608 :
6609 20628 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
6610 : {
6611 : XLogRecPtr last_important;
6612 :
6613 : /*
6614 : * Need to take a lock to prevent torn reads of the LSN, which are
6615 : * possible on some of the supported platforms. WAL insert locks only
6616 : * support exclusive mode, so we have to use that.
6617 : */
6618 18336 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
6619 18336 : last_important = WALInsertLocks[i].l.lastImportantAt;
6620 18336 : LWLockRelease(&WALInsertLocks[i].l.lock);
6621 :
6622 18336 : if (res < last_important)
6623 4416 : res = last_important;
6624 : }
6625 :
6626 2292 : return res;
6627 : }
6628 :
6629 : /*
6630 : * Get the time and LSN of the last xlog segment switch
6631 : */
6632 : pg_time_t
6633 0 : GetLastSegSwitchData(XLogRecPtr *lastSwitchLSN)
6634 : {
6635 : pg_time_t result;
6636 :
6637 : /* Need WALWriteLock, but shared lock is sufficient */
6638 0 : LWLockAcquire(WALWriteLock, LW_SHARED);
6639 0 : result = XLogCtl->lastSegSwitchTime;
6640 0 : *lastSwitchLSN = XLogCtl->lastSegSwitchLSN;
6641 0 : LWLockRelease(WALWriteLock);
6642 :
6643 0 : return result;
6644 : }
6645 :
6646 : /*
6647 : * This must be called ONCE during postmaster or standalone-backend shutdown
6648 : */
6649 : void
6650 1178 : ShutdownXLOG(int code, Datum arg)
6651 : {
6652 : /*
6653 : * We should have an aux process resource owner to use, and we should not
6654 : * be in a transaction that's installed some other resowner.
6655 : */
6656 : Assert(AuxProcessResourceOwner != NULL);
6657 : Assert(CurrentResourceOwner == NULL ||
6658 : CurrentResourceOwner == AuxProcessResourceOwner);
6659 1178 : CurrentResourceOwner = AuxProcessResourceOwner;
6660 :
6661 : /* Don't be chatty in standalone mode */
6662 1178 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
6663 : (errmsg("shutting down")));
6664 :
6665 : /*
6666 : * Signal walsenders to move to stopping state.
6667 : */
6668 1178 : WalSndInitStopping();
6669 :
6670 : /*
6671 : * Wait for WAL senders to be in stopping state. This prevents commands
6672 : * from writing new WAL.
6673 : */
6674 1178 : WalSndWaitStopping();
6675 :
6676 1178 : if (RecoveryInProgress())
6677 106 : CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_IMMEDIATE);
6678 : else
6679 : {
6680 : /*
6681 : * If archiving is enabled, rotate the last XLOG file so that all the
6682 : * remaining records are archived (postmaster wakes up the archiver
6683 : * process one more time at the end of shutdown). The checkpoint
6684 : * record will go to the next XLOG file and won't be archived (yet).
6685 : */
6686 1072 : if (XLogArchivingActive())
6687 26 : RequestXLogSwitch(false);
6688 :
6689 1072 : CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_IMMEDIATE);
6690 : }
6691 1178 : }
6692 :
6693 : /*
6694 : * Log start of a checkpoint.
6695 : */
6696 : static void
6697 2082 : LogCheckpointStart(int flags, bool restartpoint)
6698 : {
6699 2082 : if (restartpoint)
6700 376 : ereport(LOG,
6701 : /* translator: the placeholders show checkpoint options */
6702 : (errmsg("restartpoint starting:%s%s%s%s%s%s%s%s",
6703 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6704 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6705 : (flags & CHECKPOINT_IMMEDIATE) ? " immediate" : "",
6706 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6707 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6708 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6709 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6710 : (flags & CHECKPOINT_FLUSH_ALL) ? " flush-all" : "")));
6711 : else
6712 1706 : ereport(LOG,
6713 : /* translator: the placeholders show checkpoint options */
6714 : (errmsg("checkpoint starting:%s%s%s%s%s%s%s%s",
6715 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6716 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6717 : (flags & CHECKPOINT_IMMEDIATE) ? " immediate" : "",
6718 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6719 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6720 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6721 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6722 : (flags & CHECKPOINT_FLUSH_ALL) ? " flush-all" : "")));
6723 2082 : }
6724 :
6725 : /*
6726 : * Log end of a checkpoint.
6727 : */
6728 : static void
6729 2606 : LogCheckpointEnd(bool restartpoint)
6730 : {
6731 : long write_msecs,
6732 : sync_msecs,
6733 : total_msecs,
6734 : longest_msecs,
6735 : average_msecs;
6736 : uint64 average_sync_time;
6737 :
6738 2606 : CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
6739 :
6740 2606 : write_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_write_t,
6741 : CheckpointStats.ckpt_sync_t);
6742 :
6743 2606 : sync_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_sync_t,
6744 : CheckpointStats.ckpt_sync_end_t);
6745 :
6746 : /* Accumulate checkpoint timing summary data, in milliseconds. */
6747 2606 : PendingCheckpointerStats.write_time += write_msecs;
6748 2606 : PendingCheckpointerStats.sync_time += sync_msecs;
6749 :
6750 : /*
6751 : * All of the published timing statistics are accounted for. Only
6752 : * continue if a log message is to be written.
6753 : */
6754 2606 : if (!log_checkpoints)
6755 524 : return;
6756 :
6757 2082 : total_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_start_t,
6758 : CheckpointStats.ckpt_end_t);
6759 :
6760 : /*
6761 : * Timing values returned from CheckpointStats are in microseconds.
6762 : * Convert to milliseconds for consistent printing.
6763 : */
6764 2082 : longest_msecs = (long) ((CheckpointStats.ckpt_longest_sync + 999) / 1000);
6765 :
6766 2082 : average_sync_time = 0;
6767 2082 : if (CheckpointStats.ckpt_sync_rels > 0)
6768 0 : average_sync_time = CheckpointStats.ckpt_agg_sync_time /
6769 0 : CheckpointStats.ckpt_sync_rels;
6770 2082 : average_msecs = (long) ((average_sync_time + 999) / 1000);
6771 :
6772 : /*
6773 : * ControlFileLock is not required to see ControlFile->checkPoint and
6774 : * ->checkPointCopy here as we are the only updator of those variables at
6775 : * this moment.
6776 : */
6777 2082 : if (restartpoint)
6778 376 : ereport(LOG,
6779 : (errmsg("restartpoint complete: wrote %d buffers (%.1f%%), "
6780 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6781 : "%d removed, %d recycled; write=%ld.%03d s, "
6782 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6783 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6784 : "estimate=%d kB; lsn=%X/%X, redo lsn=%X/%X",
6785 : CheckpointStats.ckpt_bufs_written,
6786 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6787 : CheckpointStats.ckpt_slru_written,
6788 : CheckpointStats.ckpt_segs_added,
6789 : CheckpointStats.ckpt_segs_removed,
6790 : CheckpointStats.ckpt_segs_recycled,
6791 : write_msecs / 1000, (int) (write_msecs % 1000),
6792 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6793 : total_msecs / 1000, (int) (total_msecs % 1000),
6794 : CheckpointStats.ckpt_sync_rels,
6795 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6796 : average_msecs / 1000, (int) (average_msecs % 1000),
6797 : (int) (PrevCheckPointDistance / 1024.0),
6798 : (int) (CheckPointDistanceEstimate / 1024.0),
6799 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6800 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6801 : else
6802 1706 : ereport(LOG,
6803 : (errmsg("checkpoint 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/%X, redo lsn=%X/%X",
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 : }
6826 :
6827 : /*
6828 : * Update the estimate of distance between checkpoints.
6829 : *
6830 : * The estimate is used to calculate the number of WAL segments to keep
6831 : * preallocated, see XLOGfileslop().
6832 : */
6833 : static void
6834 2606 : UpdateCheckPointDistanceEstimate(uint64 nbytes)
6835 : {
6836 : /*
6837 : * To estimate the number of segments consumed between checkpoints, keep a
6838 : * moving average of the amount of WAL generated in previous checkpoint
6839 : * cycles. However, if the load is bursty, with quiet periods and busy
6840 : * periods, we want to cater for the peak load. So instead of a plain
6841 : * moving average, let the average decline slowly if the previous cycle
6842 : * used less WAL than estimated, but bump it up immediately if it used
6843 : * more.
6844 : *
6845 : * When checkpoints are triggered by max_wal_size, this should converge to
6846 : * CheckpointSegments * wal_segment_size,
6847 : *
6848 : * Note: This doesn't pay any attention to what caused the checkpoint.
6849 : * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
6850 : * starting a base backup, are counted the same as those created
6851 : * automatically. The slow-decline will largely mask them out, if they are
6852 : * not frequent. If they are frequent, it seems reasonable to count them
6853 : * in as any others; if you issue a manual checkpoint every 5 minutes and
6854 : * never let a timed checkpoint happen, it makes sense to base the
6855 : * preallocation on that 5 minute interval rather than whatever
6856 : * checkpoint_timeout is set to.
6857 : */
6858 2606 : PrevCheckPointDistance = nbytes;
6859 2606 : if (CheckPointDistanceEstimate < nbytes)
6860 1328 : CheckPointDistanceEstimate = nbytes;
6861 : else
6862 1278 : CheckPointDistanceEstimate =
6863 1278 : (0.90 * CheckPointDistanceEstimate + 0.10 * (double) nbytes);
6864 2606 : }
6865 :
6866 : /*
6867 : * Update the ps display for a process running a checkpoint. Note that
6868 : * this routine should not do any allocations so as it can be called
6869 : * from a critical section.
6870 : */
6871 : static void
6872 5212 : update_checkpoint_display(int flags, bool restartpoint, bool reset)
6873 : {
6874 : /*
6875 : * The status is reported only for end-of-recovery and shutdown
6876 : * checkpoints or shutdown restartpoints. Updating the ps display is
6877 : * useful in those situations as it may not be possible to rely on
6878 : * pg_stat_activity to see the status of the checkpointer or the startup
6879 : * process.
6880 : */
6881 5212 : if ((flags & (CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IS_SHUTDOWN)) == 0)
6882 2872 : return;
6883 :
6884 2340 : if (reset)
6885 1170 : set_ps_display("");
6886 : else
6887 : {
6888 : char activitymsg[128];
6889 :
6890 3510 : snprintf(activitymsg, sizeof(activitymsg), "performing %s%s%s",
6891 1170 : (flags & CHECKPOINT_END_OF_RECOVERY) ? "end-of-recovery " : "",
6892 1170 : (flags & CHECKPOINT_IS_SHUTDOWN) ? "shutdown " : "",
6893 : restartpoint ? "restartpoint" : "checkpoint");
6894 1170 : set_ps_display(activitymsg);
6895 : }
6896 : }
6897 :
6898 :
6899 : /*
6900 : * Perform a checkpoint --- either during shutdown, or on-the-fly
6901 : *
6902 : * flags is a bitwise OR of the following:
6903 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
6904 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
6905 : * CHECKPOINT_IMMEDIATE: finish the checkpoint ASAP,
6906 : * ignoring checkpoint_completion_target parameter.
6907 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
6908 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
6909 : * CHECKPOINT_END_OF_RECOVERY).
6910 : * CHECKPOINT_FLUSH_ALL: also flush buffers of unlogged tables.
6911 : *
6912 : * Note: flags contains other bits, of interest here only for logging purposes.
6913 : * In particular note that this routine is synchronous and does not pay
6914 : * attention to CHECKPOINT_WAIT.
6915 : *
6916 : * If !shutdown then we are writing an online checkpoint. An XLOG_CHECKPOINT_REDO
6917 : * record is inserted into WAL at the logical location of the checkpoint, before
6918 : * flushing anything to disk, and when the checkpoint is eventually completed,
6919 : * and it is from this point that WAL replay will begin in the case of a recovery
6920 : * from this checkpoint. Once everything is written to disk, an
6921 : * XLOG_CHECKPOINT_ONLINE record is written to complete the checkpoint, and
6922 : * points back to the earlier XLOG_CHECKPOINT_REDO record. This mechanism allows
6923 : * other write-ahead log records to be written while the checkpoint is in
6924 : * progress, but we must be very careful about order of operations. This function
6925 : * may take many minutes to execute on a busy system.
6926 : *
6927 : * On the other hand, when shutdown is true, concurrent insertion into the
6928 : * write-ahead log is impossible, so there is no need for two separate records.
6929 : * In this case, we only insert an XLOG_CHECKPOINT_SHUTDOWN record, and it's
6930 : * both the record marking the completion of the checkpoint and the location
6931 : * from which WAL replay would begin if needed.
6932 : *
6933 : * Returns true if a new checkpoint was performed, or false if it was skipped
6934 : * because the system was idle.
6935 : */
6936 : bool
6937 2230 : CreateCheckPoint(int flags)
6938 : {
6939 : bool shutdown;
6940 : CheckPoint checkPoint;
6941 : XLogRecPtr recptr;
6942 : XLogSegNo _logSegNo;
6943 2230 : XLogCtlInsert *Insert = &XLogCtl->Insert;
6944 : uint32 freespace;
6945 : XLogRecPtr PriorRedoPtr;
6946 : XLogRecPtr last_important_lsn;
6947 : VirtualTransactionId *vxids;
6948 : int nvxids;
6949 2230 : int oldXLogAllowed = 0;
6950 :
6951 : /*
6952 : * An end-of-recovery checkpoint is really a shutdown checkpoint, just
6953 : * issued at a different time.
6954 : */
6955 2230 : if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
6956 1130 : shutdown = true;
6957 : else
6958 1100 : shutdown = false;
6959 :
6960 : /* sanity check */
6961 2230 : if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
6962 0 : elog(ERROR, "can't create a checkpoint during recovery");
6963 :
6964 : /*
6965 : * Prepare to accumulate statistics.
6966 : *
6967 : * Note: because it is possible for log_checkpoints to change while a
6968 : * checkpoint proceeds, we always accumulate stats, even if
6969 : * log_checkpoints is currently off.
6970 : */
6971 24530 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
6972 2230 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
6973 :
6974 : /*
6975 : * Let smgr prepare for checkpoint; this has to happen outside the
6976 : * critical section and before we determine the REDO pointer. Note that
6977 : * smgr must not do anything that'd have to be undone if we decide no
6978 : * checkpoint is needed.
6979 : */
6980 2230 : SyncPreCheckpoint();
6981 :
6982 : /*
6983 : * Use a critical section to force system panic if we have trouble.
6984 : */
6985 2230 : START_CRIT_SECTION();
6986 :
6987 2230 : if (shutdown)
6988 : {
6989 1130 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6990 1130 : ControlFile->state = DB_SHUTDOWNING;
6991 1130 : UpdateControlFile();
6992 1130 : LWLockRelease(ControlFileLock);
6993 : }
6994 :
6995 : /* Begin filling in the checkpoint WAL record */
6996 26760 : MemSet(&checkPoint, 0, sizeof(checkPoint));
6997 2230 : checkPoint.time = (pg_time_t) time(NULL);
6998 :
6999 : /*
7000 : * For Hot Standby, derive the oldestActiveXid before we fix the redo
7001 : * pointer. This allows us to begin accumulating changes to assemble our
7002 : * starting snapshot of locks and transactions.
7003 : */
7004 2230 : if (!shutdown && XLogStandbyInfoActive())
7005 1006 : checkPoint.oldestActiveXid = GetOldestActiveTransactionId();
7006 : else
7007 1224 : checkPoint.oldestActiveXid = InvalidTransactionId;
7008 :
7009 : /*
7010 : * Get location of last important record before acquiring insert locks (as
7011 : * GetLastImportantRecPtr() also locks WAL locks).
7012 : */
7013 2230 : last_important_lsn = GetLastImportantRecPtr();
7014 :
7015 : /*
7016 : * If this isn't a shutdown or forced checkpoint, and if there has been no
7017 : * WAL activity requiring a checkpoint, skip it. The idea here is to
7018 : * avoid inserting duplicate checkpoints when the system is idle.
7019 : */
7020 2230 : if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
7021 : CHECKPOINT_FORCE)) == 0)
7022 : {
7023 372 : if (last_important_lsn == ControlFile->checkPoint)
7024 : {
7025 0 : END_CRIT_SECTION();
7026 0 : ereport(DEBUG1,
7027 : (errmsg_internal("checkpoint skipped because system is idle")));
7028 0 : return false;
7029 : }
7030 : }
7031 :
7032 : /*
7033 : * An end-of-recovery checkpoint is created before anyone is allowed to
7034 : * write WAL. To allow us to write the checkpoint record, temporarily
7035 : * enable XLogInsertAllowed.
7036 : */
7037 2230 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7038 58 : oldXLogAllowed = LocalSetXLogInsertAllowed();
7039 :
7040 2230 : checkPoint.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7041 2230 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7042 58 : checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7043 : else
7044 2172 : checkPoint.PrevTimeLineID = checkPoint.ThisTimeLineID;
7045 :
7046 : /*
7047 : * We must block concurrent insertions while examining insert state.
7048 : */
7049 2230 : WALInsertLockAcquireExclusive();
7050 :
7051 2230 : checkPoint.fullPageWrites = Insert->fullPageWrites;
7052 2230 : checkPoint.wal_level = wal_level;
7053 :
7054 2230 : if (shutdown)
7055 : {
7056 1130 : XLogRecPtr curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
7057 :
7058 : /*
7059 : * Compute new REDO record ptr = location of next XLOG record.
7060 : *
7061 : * Since this is a shutdown checkpoint, there can't be any concurrent
7062 : * WAL insertion.
7063 : */
7064 1130 : freespace = INSERT_FREESPACE(curInsert);
7065 1130 : if (freespace == 0)
7066 : {
7067 0 : if (XLogSegmentOffset(curInsert, wal_segment_size) == 0)
7068 0 : curInsert += SizeOfXLogLongPHD;
7069 : else
7070 0 : curInsert += SizeOfXLogShortPHD;
7071 : }
7072 1130 : checkPoint.redo = curInsert;
7073 :
7074 : /*
7075 : * Here we update the shared RedoRecPtr for future XLogInsert calls;
7076 : * this must be done while holding all the insertion locks.
7077 : *
7078 : * Note: if we fail to complete the checkpoint, RedoRecPtr will be
7079 : * left pointing past where it really needs to point. This is okay;
7080 : * the only consequence is that XLogInsert might back up whole buffers
7081 : * that it didn't really need to. We can't postpone advancing
7082 : * RedoRecPtr because XLogInserts that happen while we are dumping
7083 : * buffers must assume that their buffer changes are not included in
7084 : * the checkpoint.
7085 : */
7086 1130 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
7087 : }
7088 :
7089 : /*
7090 : * Now we can release the WAL insertion locks, allowing other xacts to
7091 : * proceed while we are flushing disk buffers.
7092 : */
7093 2230 : WALInsertLockRelease();
7094 :
7095 : /*
7096 : * If this is an online checkpoint, we have not yet determined the redo
7097 : * point. We do so now by inserting the special XLOG_CHECKPOINT_REDO
7098 : * record; the LSN at which it starts becomes the new redo pointer. We
7099 : * don't do this for a shutdown checkpoint, because in that case no WAL
7100 : * can be written between the redo point and the insertion of the
7101 : * checkpoint record itself, so the checkpoint record itself serves to
7102 : * mark the redo point.
7103 : */
7104 2230 : if (!shutdown)
7105 : {
7106 : /* Include WAL level in record for WAL summarizer's benefit. */
7107 1100 : XLogBeginInsert();
7108 1100 : XLogRegisterData(&wal_level, sizeof(wal_level));
7109 1100 : (void) XLogInsert(RM_XLOG_ID, XLOG_CHECKPOINT_REDO);
7110 :
7111 : /*
7112 : * XLogInsertRecord will have updated XLogCtl->Insert.RedoRecPtr in
7113 : * shared memory and RedoRecPtr in backend-local memory, but we need
7114 : * to copy that into the record that will be inserted when the
7115 : * checkpoint is complete.
7116 : */
7117 1100 : checkPoint.redo = RedoRecPtr;
7118 : }
7119 :
7120 : /* Update the info_lck-protected copy of RedoRecPtr as well */
7121 2230 : SpinLockAcquire(&XLogCtl->info_lck);
7122 2230 : XLogCtl->RedoRecPtr = checkPoint.redo;
7123 2230 : SpinLockRelease(&XLogCtl->info_lck);
7124 :
7125 : /*
7126 : * If enabled, log checkpoint start. We postpone this until now so as not
7127 : * to log anything if we decided to skip the checkpoint.
7128 : */
7129 2230 : if (log_checkpoints)
7130 1706 : LogCheckpointStart(flags, false);
7131 :
7132 : /* Update the process title */
7133 2230 : update_checkpoint_display(flags, false, false);
7134 :
7135 : TRACE_POSTGRESQL_CHECKPOINT_START(flags);
7136 :
7137 : /*
7138 : * Get the other info we need for the checkpoint record.
7139 : *
7140 : * We don't need to save oldestClogXid in the checkpoint, it only matters
7141 : * for the short period in which clog is being truncated, and if we crash
7142 : * during that we'll redo the clog truncation and fix up oldestClogXid
7143 : * there.
7144 : */
7145 2230 : LWLockAcquire(XidGenLock, LW_SHARED);
7146 2230 : checkPoint.nextXid = TransamVariables->nextXid;
7147 2230 : checkPoint.oldestXid = TransamVariables->oldestXid;
7148 2230 : checkPoint.oldestXidDB = TransamVariables->oldestXidDB;
7149 2230 : LWLockRelease(XidGenLock);
7150 :
7151 2230 : LWLockAcquire(CommitTsLock, LW_SHARED);
7152 2230 : checkPoint.oldestCommitTsXid = TransamVariables->oldestCommitTsXid;
7153 2230 : checkPoint.newestCommitTsXid = TransamVariables->newestCommitTsXid;
7154 2230 : LWLockRelease(CommitTsLock);
7155 :
7156 2230 : LWLockAcquire(OidGenLock, LW_SHARED);
7157 2230 : checkPoint.nextOid = TransamVariables->nextOid;
7158 2230 : if (!shutdown)
7159 1100 : checkPoint.nextOid += TransamVariables->oidCount;
7160 2230 : LWLockRelease(OidGenLock);
7161 :
7162 2230 : MultiXactGetCheckptMulti(shutdown,
7163 : &checkPoint.nextMulti,
7164 : &checkPoint.nextMultiOffset,
7165 : &checkPoint.oldestMulti,
7166 : &checkPoint.oldestMultiDB);
7167 :
7168 : /*
7169 : * Having constructed the checkpoint record, ensure all shmem disk buffers
7170 : * and commit-log buffers are flushed to disk.
7171 : *
7172 : * This I/O could fail for various reasons. If so, we will fail to
7173 : * complete the checkpoint, but there is no reason to force a system
7174 : * panic. Accordingly, exit critical section while doing it.
7175 : */
7176 2230 : END_CRIT_SECTION();
7177 :
7178 : /*
7179 : * In some cases there are groups of actions that must all occur on one
7180 : * side or the other of a checkpoint record. Before flushing the
7181 : * checkpoint record we must explicitly wait for any backend currently
7182 : * performing those groups of actions.
7183 : *
7184 : * One example is end of transaction, so we must wait for any transactions
7185 : * that are currently in commit critical sections. If an xact inserted
7186 : * its commit record into XLOG just before the REDO point, then a crash
7187 : * restart from the REDO point would not replay that record, which means
7188 : * that our flushing had better include the xact's update of pg_xact. So
7189 : * we wait till he's out of his commit critical section before proceeding.
7190 : * See notes in RecordTransactionCommit().
7191 : *
7192 : * Because we've already released the insertion locks, this test is a bit
7193 : * fuzzy: it is possible that we will wait for xacts we didn't really need
7194 : * to wait for. But the delay should be short and it seems better to make
7195 : * checkpoint take a bit longer than to hold off insertions longer than
7196 : * necessary. (In fact, the whole reason we have this issue is that xact.c
7197 : * does commit record XLOG insertion and clog update as two separate steps
7198 : * protected by different locks, but again that seems best on grounds of
7199 : * minimizing lock contention.)
7200 : *
7201 : * A transaction that has not yet set delayChkptFlags when we look cannot
7202 : * be at risk, since it has not inserted its commit record yet; and one
7203 : * that's already cleared it is not at risk either, since it's done fixing
7204 : * clog and we will correctly flush the update below. So we cannot miss
7205 : * any xacts we need to wait for.
7206 : */
7207 2230 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
7208 2230 : if (nvxids > 0)
7209 : {
7210 : do
7211 : {
7212 : /*
7213 : * Keep absorbing fsync requests while we wait. There could even
7214 : * be a deadlock if we don't, if the process that prevents the
7215 : * checkpoint is trying to add a request to the queue.
7216 : */
7217 54 : AbsorbSyncRequests();
7218 :
7219 54 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
7220 54 : pg_usleep(10000L); /* wait for 10 msec */
7221 54 : pgstat_report_wait_end();
7222 54 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7223 54 : DELAY_CHKPT_START));
7224 : }
7225 2230 : pfree(vxids);
7226 :
7227 2230 : CheckPointGuts(checkPoint.redo, flags);
7228 :
7229 2230 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
7230 2230 : if (nvxids > 0)
7231 : {
7232 : do
7233 : {
7234 0 : AbsorbSyncRequests();
7235 :
7236 0 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_COMPLETE);
7237 0 : pg_usleep(10000L); /* wait for 10 msec */
7238 0 : pgstat_report_wait_end();
7239 0 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7240 0 : DELAY_CHKPT_COMPLETE));
7241 : }
7242 2230 : pfree(vxids);
7243 :
7244 : /*
7245 : * Take a snapshot of running transactions and write this to WAL. This
7246 : * allows us to reconstruct the state of running transactions during
7247 : * archive recovery, if required. Skip, if this info disabled.
7248 : *
7249 : * If we are shutting down, or Startup process is completing crash
7250 : * recovery we don't need to write running xact data.
7251 : */
7252 2230 : if (!shutdown && XLogStandbyInfoActive())
7253 1006 : LogStandbySnapshot();
7254 :
7255 2230 : START_CRIT_SECTION();
7256 :
7257 : /*
7258 : * Now insert the checkpoint record into XLOG.
7259 : */
7260 2230 : XLogBeginInsert();
7261 2230 : XLogRegisterData(&checkPoint, sizeof(checkPoint));
7262 2230 : recptr = XLogInsert(RM_XLOG_ID,
7263 : shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
7264 : XLOG_CHECKPOINT_ONLINE);
7265 :
7266 2230 : XLogFlush(recptr);
7267 :
7268 : /*
7269 : * We mustn't write any new WAL after a shutdown checkpoint, or it will be
7270 : * overwritten at next startup. No-one should even try, this just allows
7271 : * sanity-checking. In the case of an end-of-recovery checkpoint, we want
7272 : * to just temporarily disable writing until the system has exited
7273 : * recovery.
7274 : */
7275 2230 : if (shutdown)
7276 : {
7277 1130 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7278 58 : LocalXLogInsertAllowed = oldXLogAllowed;
7279 : else
7280 1072 : LocalXLogInsertAllowed = 0; /* never again write WAL */
7281 : }
7282 :
7283 : /*
7284 : * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
7285 : * = end of actual checkpoint record.
7286 : */
7287 2230 : if (shutdown && checkPoint.redo != ProcLastRecPtr)
7288 0 : ereport(PANIC,
7289 : (errmsg("concurrent write-ahead log activity while database system is shutting down")));
7290 :
7291 : /*
7292 : * Remember the prior checkpoint's redo ptr for
7293 : * UpdateCheckPointDistanceEstimate()
7294 : */
7295 2230 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7296 :
7297 : /*
7298 : * Update the control file.
7299 : */
7300 2230 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7301 2230 : if (shutdown)
7302 1130 : ControlFile->state = DB_SHUTDOWNED;
7303 2230 : ControlFile->checkPoint = ProcLastRecPtr;
7304 2230 : ControlFile->checkPointCopy = checkPoint;
7305 : /* crash recovery should always recover to the end of WAL */
7306 2230 : ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
7307 2230 : ControlFile->minRecoveryPointTLI = 0;
7308 :
7309 : /*
7310 : * Persist unloggedLSN value. It's reset on crash recovery, so this goes
7311 : * unused on non-shutdown checkpoints, but seems useful to store it always
7312 : * for debugging purposes.
7313 : */
7314 2230 : ControlFile->unloggedLSN = pg_atomic_read_membarrier_u64(&XLogCtl->unloggedLSN);
7315 :
7316 2230 : UpdateControlFile();
7317 2230 : LWLockRelease(ControlFileLock);
7318 :
7319 : /* Update shared-memory copy of checkpoint XID/epoch */
7320 2230 : SpinLockAcquire(&XLogCtl->info_lck);
7321 2230 : XLogCtl->ckptFullXid = checkPoint.nextXid;
7322 2230 : SpinLockRelease(&XLogCtl->info_lck);
7323 :
7324 : /*
7325 : * We are now done with critical updates; no need for system panic if we
7326 : * have trouble while fooling with old log segments.
7327 : */
7328 2230 : END_CRIT_SECTION();
7329 :
7330 : /*
7331 : * WAL summaries end when the next XLOG_CHECKPOINT_REDO or
7332 : * XLOG_CHECKPOINT_SHUTDOWN record is reached. This is the first point
7333 : * where (a) we're not inside of a critical section and (b) we can be
7334 : * certain that the relevant record has been flushed to disk, which must
7335 : * happen before it can be summarized.
7336 : *
7337 : * If this is a shutdown checkpoint, then this happens reasonably
7338 : * promptly: we've only just inserted and flushed the
7339 : * XLOG_CHECKPOINT_SHUTDOWN record. If this is not a shutdown checkpoint,
7340 : * then this might not be very prompt at all: the XLOG_CHECKPOINT_REDO
7341 : * record was written before we began flushing data to disk, and that
7342 : * could be many minutes ago at this point. However, we don't XLogFlush()
7343 : * after inserting that record, so we're not guaranteed that it's on disk
7344 : * until after the above call that flushes the XLOG_CHECKPOINT_ONLINE
7345 : * record.
7346 : */
7347 2230 : WakeupWalSummarizer();
7348 :
7349 : /*
7350 : * Let smgr do post-checkpoint cleanup (eg, deleting old files).
7351 : */
7352 2230 : SyncPostCheckpoint();
7353 :
7354 : /*
7355 : * Update the average distance between checkpoints if the prior checkpoint
7356 : * exists.
7357 : */
7358 2230 : if (PriorRedoPtr != InvalidXLogRecPtr)
7359 2230 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7360 :
7361 : /*
7362 : * Delete old log files, those no longer needed for last checkpoint to
7363 : * prevent the disk holding the xlog from growing full.
7364 : */
7365 2230 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7366 2230 : KeepLogSeg(recptr, &_logSegNo);
7367 2230 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7368 : _logSegNo, InvalidOid,
7369 : InvalidTransactionId))
7370 : {
7371 : /*
7372 : * Some slots have been invalidated; recalculate the old-segment
7373 : * horizon, starting again from RedoRecPtr.
7374 : */
7375 6 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7376 6 : KeepLogSeg(recptr, &_logSegNo);
7377 : }
7378 2230 : _logSegNo--;
7379 2230 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
7380 : checkPoint.ThisTimeLineID);
7381 :
7382 : /*
7383 : * Make more log segments if needed. (Do this after recycling old log
7384 : * segments, since that may supply some of the needed files.)
7385 : */
7386 2230 : if (!shutdown)
7387 1100 : PreallocXlogFiles(recptr, checkPoint.ThisTimeLineID);
7388 :
7389 : /*
7390 : * Truncate pg_subtrans if possible. We can throw away all data before
7391 : * the oldest XMIN of any running transaction. No future transaction will
7392 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7393 : * in subtrans.c). During recovery, though, we mustn't do this because
7394 : * StartupSUBTRANS hasn't been called yet.
7395 : */
7396 2230 : if (!RecoveryInProgress())
7397 2172 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7398 :
7399 : /* Real work is done; log and update stats. */
7400 2230 : LogCheckpointEnd(false);
7401 :
7402 : /* Reset the process title */
7403 2230 : update_checkpoint_display(flags, false, true);
7404 :
7405 : TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
7406 : NBuffers,
7407 : CheckpointStats.ckpt_segs_added,
7408 : CheckpointStats.ckpt_segs_removed,
7409 : CheckpointStats.ckpt_segs_recycled);
7410 :
7411 2230 : return true;
7412 : }
7413 :
7414 : /*
7415 : * Mark the end of recovery in WAL though without running a full checkpoint.
7416 : * We can expect that a restartpoint is likely to be in progress as we
7417 : * do this, though we are unwilling to wait for it to complete.
7418 : *
7419 : * CreateRestartPoint() allows for the case where recovery may end before
7420 : * the restartpoint completes so there is no concern of concurrent behaviour.
7421 : */
7422 : static void
7423 84 : CreateEndOfRecoveryRecord(void)
7424 : {
7425 : xl_end_of_recovery xlrec;
7426 : XLogRecPtr recptr;
7427 :
7428 : /* sanity check */
7429 84 : if (!RecoveryInProgress())
7430 0 : elog(ERROR, "can only be used to end recovery");
7431 :
7432 84 : xlrec.end_time = GetCurrentTimestamp();
7433 84 : xlrec.wal_level = wal_level;
7434 :
7435 84 : WALInsertLockAcquireExclusive();
7436 84 : xlrec.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7437 84 : xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7438 84 : WALInsertLockRelease();
7439 :
7440 84 : START_CRIT_SECTION();
7441 :
7442 84 : XLogBeginInsert();
7443 84 : XLogRegisterData(&xlrec, sizeof(xl_end_of_recovery));
7444 84 : recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY);
7445 :
7446 84 : XLogFlush(recptr);
7447 :
7448 : /*
7449 : * Update the control file so that crash recovery can follow the timeline
7450 : * changes to this point.
7451 : */
7452 84 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7453 84 : ControlFile->minRecoveryPoint = recptr;
7454 84 : ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
7455 84 : UpdateControlFile();
7456 84 : LWLockRelease(ControlFileLock);
7457 :
7458 84 : END_CRIT_SECTION();
7459 84 : }
7460 :
7461 : /*
7462 : * Write an OVERWRITE_CONTRECORD message.
7463 : *
7464 : * When on WAL replay we expect a continuation record at the start of a page
7465 : * that is not there, recovery ends and WAL writing resumes at that point.
7466 : * But it's wrong to resume writing new WAL back at the start of the record
7467 : * that was broken, because downstream consumers of that WAL (physical
7468 : * replicas) are not prepared to "rewind". So the first action after
7469 : * finishing replay of all valid WAL must be to write a record of this type
7470 : * at the point where the contrecord was missing; to support xlogreader
7471 : * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
7472 : * to the page header where the record occurs. xlogreader has an ad-hoc
7473 : * mechanism to report metadata about the broken record, which is what we
7474 : * use here.
7475 : *
7476 : * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
7477 : * skip the record it was reading, and pass back the LSN of the skipped
7478 : * record, so that its caller can verify (on "replay" of that record) that the
7479 : * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
7480 : *
7481 : * 'aborted_lsn' is the beginning position of the record that was incomplete.
7482 : * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
7483 : * beginning of the XLOG page where the record is to be inserted. They must
7484 : * match the current WAL insert position, they're passed here just so that we
7485 : * can verify that.
7486 : */
7487 : static XLogRecPtr
7488 22 : CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn, XLogRecPtr pagePtr,
7489 : TimeLineID newTLI)
7490 : {
7491 : xl_overwrite_contrecord xlrec;
7492 : XLogRecPtr recptr;
7493 : XLogPageHeader pagehdr;
7494 : XLogRecPtr startPos;
7495 :
7496 : /* sanity checks */
7497 22 : if (!RecoveryInProgress())
7498 0 : elog(ERROR, "can only be used at end of recovery");
7499 22 : if (pagePtr % XLOG_BLCKSZ != 0)
7500 0 : elog(ERROR, "invalid position for missing continuation record %X/%X",
7501 : LSN_FORMAT_ARGS(pagePtr));
7502 :
7503 : /* The current WAL insert position should be right after the page header */
7504 22 : startPos = pagePtr;
7505 22 : if (XLogSegmentOffset(startPos, wal_segment_size) == 0)
7506 2 : startPos += SizeOfXLogLongPHD;
7507 : else
7508 20 : startPos += SizeOfXLogShortPHD;
7509 22 : recptr = GetXLogInsertRecPtr();
7510 22 : if (recptr != startPos)
7511 0 : elog(ERROR, "invalid WAL insert position %X/%X for OVERWRITE_CONTRECORD",
7512 : LSN_FORMAT_ARGS(recptr));
7513 :
7514 22 : START_CRIT_SECTION();
7515 :
7516 : /*
7517 : * Initialize the XLOG page header (by GetXLogBuffer), and set the
7518 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
7519 : *
7520 : * No other backend is allowed to write WAL yet, so acquiring the WAL
7521 : * insertion lock is just pro forma.
7522 : */
7523 22 : WALInsertLockAcquire();
7524 22 : pagehdr = (XLogPageHeader) GetXLogBuffer(pagePtr, newTLI);
7525 22 : pagehdr->xlp_info |= XLP_FIRST_IS_OVERWRITE_CONTRECORD;
7526 22 : WALInsertLockRelease();
7527 :
7528 : /*
7529 : * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
7530 : * page. We know it becomes the first record, because no other backend is
7531 : * allowed to write WAL yet.
7532 : */
7533 22 : XLogBeginInsert();
7534 22 : xlrec.overwritten_lsn = aborted_lsn;
7535 22 : xlrec.overwrite_time = GetCurrentTimestamp();
7536 22 : XLogRegisterData(&xlrec, sizeof(xl_overwrite_contrecord));
7537 22 : recptr = XLogInsert(RM_XLOG_ID, XLOG_OVERWRITE_CONTRECORD);
7538 :
7539 : /* check that the record was inserted to the right place */
7540 22 : if (ProcLastRecPtr != startPos)
7541 0 : elog(ERROR, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%X",
7542 : LSN_FORMAT_ARGS(ProcLastRecPtr));
7543 :
7544 22 : XLogFlush(recptr);
7545 :
7546 22 : END_CRIT_SECTION();
7547 :
7548 22 : return recptr;
7549 : }
7550 :
7551 : /*
7552 : * Flush all data in shared memory to disk, and fsync
7553 : *
7554 : * This is the common code shared between regular checkpoints and
7555 : * recovery restartpoints.
7556 : */
7557 : static void
7558 2606 : CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
7559 : {
7560 2606 : CheckPointRelationMap();
7561 2606 : CheckPointReplicationSlots(flags & CHECKPOINT_IS_SHUTDOWN);
7562 2606 : CheckPointSnapBuild();
7563 2606 : CheckPointLogicalRewriteHeap();
7564 2606 : CheckPointReplicationOrigin();
7565 :
7566 : /* Write out all dirty data in SLRUs and the main buffer pool */
7567 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags);
7568 2606 : CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
7569 2606 : CheckPointCLOG();
7570 2606 : CheckPointCommitTs();
7571 2606 : CheckPointSUBTRANS();
7572 2606 : CheckPointMultiXact();
7573 2606 : CheckPointPredicate();
7574 2606 : CheckPointBuffers(flags);
7575 :
7576 : /* Perform all queued up fsyncs */
7577 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
7578 2606 : CheckpointStats.ckpt_sync_t = GetCurrentTimestamp();
7579 2606 : ProcessSyncRequests();
7580 2606 : CheckpointStats.ckpt_sync_end_t = GetCurrentTimestamp();
7581 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
7582 :
7583 : /* We deliberately delay 2PC checkpointing as long as possible */
7584 2606 : CheckPointTwoPhase(checkPointRedo);
7585 2606 : }
7586 :
7587 : /*
7588 : * Save a checkpoint for recovery restart if appropriate
7589 : *
7590 : * This function is called each time a checkpoint record is read from XLOG.
7591 : * It must determine whether the checkpoint represents a safe restartpoint or
7592 : * not. If so, the checkpoint record is stashed in shared memory so that
7593 : * CreateRestartPoint can consult it. (Note that the latter function is
7594 : * executed by the checkpointer, while this one will be executed by the
7595 : * startup process.)
7596 : */
7597 : static void
7598 764 : RecoveryRestartPoint(const CheckPoint *checkPoint, XLogReaderState *record)
7599 : {
7600 : /*
7601 : * Also refrain from creating a restartpoint if we have seen any
7602 : * references to non-existent pages. Restarting recovery from the
7603 : * restartpoint would not see the references, so we would lose the
7604 : * cross-check that the pages belonged to a relation that was dropped
7605 : * later.
7606 : */
7607 764 : if (XLogHaveInvalidPages())
7608 : {
7609 0 : elog(DEBUG2,
7610 : "could not record restart point at %X/%X because there "
7611 : "are unresolved references to invalid pages",
7612 : LSN_FORMAT_ARGS(checkPoint->redo));
7613 0 : return;
7614 : }
7615 :
7616 : /*
7617 : * Copy the checkpoint record to shared memory, so that checkpointer can
7618 : * work out the next time it wants to perform a restartpoint.
7619 : */
7620 764 : SpinLockAcquire(&XLogCtl->info_lck);
7621 764 : XLogCtl->lastCheckPointRecPtr = record->ReadRecPtr;
7622 764 : XLogCtl->lastCheckPointEndPtr = record->EndRecPtr;
7623 764 : XLogCtl->lastCheckPoint = *checkPoint;
7624 764 : SpinLockRelease(&XLogCtl->info_lck);
7625 : }
7626 :
7627 : /*
7628 : * Establish a restartpoint if possible.
7629 : *
7630 : * This is similar to CreateCheckPoint, but is used during WAL recovery
7631 : * to establish a point from which recovery can roll forward without
7632 : * replaying the entire recovery log.
7633 : *
7634 : * Returns true if a new restartpoint was established. We can only establish
7635 : * a restartpoint if we have replayed a safe checkpoint record since last
7636 : * restartpoint.
7637 : */
7638 : bool
7639 524 : CreateRestartPoint(int flags)
7640 : {
7641 : XLogRecPtr lastCheckPointRecPtr;
7642 : XLogRecPtr lastCheckPointEndPtr;
7643 : CheckPoint lastCheckPoint;
7644 : XLogRecPtr PriorRedoPtr;
7645 : XLogRecPtr receivePtr;
7646 : XLogRecPtr replayPtr;
7647 : TimeLineID replayTLI;
7648 : XLogRecPtr endptr;
7649 : XLogSegNo _logSegNo;
7650 : TimestampTz xtime;
7651 :
7652 : /* Concurrent checkpoint/restartpoint cannot happen */
7653 : Assert(!IsUnderPostmaster || MyBackendType == B_CHECKPOINTER);
7654 :
7655 : /* Get a local copy of the last safe checkpoint record. */
7656 524 : SpinLockAcquire(&XLogCtl->info_lck);
7657 524 : lastCheckPointRecPtr = XLogCtl->lastCheckPointRecPtr;
7658 524 : lastCheckPointEndPtr = XLogCtl->lastCheckPointEndPtr;
7659 524 : lastCheckPoint = XLogCtl->lastCheckPoint;
7660 524 : SpinLockRelease(&XLogCtl->info_lck);
7661 :
7662 : /*
7663 : * Check that we're still in recovery mode. It's ok if we exit recovery
7664 : * mode after this check, the restart point is valid anyway.
7665 : */
7666 524 : if (!RecoveryInProgress())
7667 : {
7668 0 : ereport(DEBUG2,
7669 : (errmsg_internal("skipping restartpoint, recovery has already ended")));
7670 0 : return false;
7671 : }
7672 :
7673 : /*
7674 : * If the last checkpoint record we've replayed is already our last
7675 : * restartpoint, we can't perform a new restart point. We still update
7676 : * minRecoveryPoint in that case, so that if this is a shutdown restart
7677 : * point, we won't start up earlier than before. That's not strictly
7678 : * necessary, but when hot standby is enabled, it would be rather weird if
7679 : * the database opened up for read-only connections at a point-in-time
7680 : * before the last shutdown. Such time travel is still possible in case of
7681 : * immediate shutdown, though.
7682 : *
7683 : * We don't explicitly advance minRecoveryPoint when we do create a
7684 : * restartpoint. It's assumed that flushing the buffers will do that as a
7685 : * side-effect.
7686 : */
7687 524 : if (XLogRecPtrIsInvalid(lastCheckPointRecPtr) ||
7688 502 : lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
7689 : {
7690 148 : ereport(DEBUG2,
7691 : (errmsg_internal("skipping restartpoint, already performed at %X/%X",
7692 : LSN_FORMAT_ARGS(lastCheckPoint.redo))));
7693 :
7694 148 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
7695 148 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7696 : {
7697 66 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7698 66 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7699 66 : UpdateControlFile();
7700 66 : LWLockRelease(ControlFileLock);
7701 : }
7702 148 : return false;
7703 : }
7704 :
7705 : /*
7706 : * Update the shared RedoRecPtr so that the startup process can calculate
7707 : * the number of segments replayed since last restartpoint, and request a
7708 : * restartpoint if it exceeds CheckPointSegments.
7709 : *
7710 : * Like in CreateCheckPoint(), hold off insertions to update it, although
7711 : * during recovery this is just pro forma, because no WAL insertions are
7712 : * happening.
7713 : */
7714 376 : WALInsertLockAcquireExclusive();
7715 376 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = lastCheckPoint.redo;
7716 376 : WALInsertLockRelease();
7717 :
7718 : /* Also update the info_lck-protected copy */
7719 376 : SpinLockAcquire(&XLogCtl->info_lck);
7720 376 : XLogCtl->RedoRecPtr = lastCheckPoint.redo;
7721 376 : SpinLockRelease(&XLogCtl->info_lck);
7722 :
7723 : /*
7724 : * Prepare to accumulate statistics.
7725 : *
7726 : * Note: because it is possible for log_checkpoints to change while a
7727 : * checkpoint proceeds, we always accumulate stats, even if
7728 : * log_checkpoints is currently off.
7729 : */
7730 4136 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7731 376 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7732 :
7733 376 : if (log_checkpoints)
7734 376 : LogCheckpointStart(flags, true);
7735 :
7736 : /* Update the process title */
7737 376 : update_checkpoint_display(flags, true, false);
7738 :
7739 376 : CheckPointGuts(lastCheckPoint.redo, flags);
7740 :
7741 : /*
7742 : * This location needs to be after CheckPointGuts() to ensure that some
7743 : * work has already happened during this checkpoint.
7744 : */
7745 376 : INJECTION_POINT("create-restart-point");
7746 :
7747 : /*
7748 : * Remember the prior checkpoint's redo ptr for
7749 : * UpdateCheckPointDistanceEstimate()
7750 : */
7751 376 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7752 :
7753 : /*
7754 : * Update pg_control, using current time. Check that it still shows an
7755 : * older checkpoint, else do nothing; this is a quick hack to make sure
7756 : * nothing really bad happens if somehow we get here after the
7757 : * end-of-recovery checkpoint.
7758 : */
7759 376 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7760 376 : if (ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
7761 : {
7762 : /*
7763 : * Update the checkpoint information. We do this even if the cluster
7764 : * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
7765 : * segments recycled below.
7766 : */
7767 376 : ControlFile->checkPoint = lastCheckPointRecPtr;
7768 376 : ControlFile->checkPointCopy = lastCheckPoint;
7769 :
7770 : /*
7771 : * Ensure minRecoveryPoint is past the checkpoint record and update it
7772 : * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
7773 : * this will have happened already while writing out dirty buffers,
7774 : * but not necessarily - e.g. because no buffers were dirtied. We do
7775 : * this because a backup performed in recovery uses minRecoveryPoint
7776 : * to determine which WAL files must be included in the backup, and
7777 : * the file (or files) containing the checkpoint record must be
7778 : * included, at a minimum. Note that for an ordinary restart of
7779 : * recovery there's no value in having the minimum recovery point any
7780 : * earlier than this anyway, because redo will begin just after the
7781 : * checkpoint record.
7782 : */
7783 376 : if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
7784 : {
7785 376 : if (ControlFile->minRecoveryPoint < lastCheckPointEndPtr)
7786 : {
7787 40 : ControlFile->minRecoveryPoint = lastCheckPointEndPtr;
7788 40 : ControlFile->minRecoveryPointTLI = lastCheckPoint.ThisTimeLineID;
7789 :
7790 : /* update local copy */
7791 40 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
7792 40 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
7793 : }
7794 376 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7795 40 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7796 : }
7797 376 : UpdateControlFile();
7798 : }
7799 376 : LWLockRelease(ControlFileLock);
7800 :
7801 : /*
7802 : * Update the average distance between checkpoints/restartpoints if the
7803 : * prior checkpoint exists.
7804 : */
7805 376 : if (PriorRedoPtr != InvalidXLogRecPtr)
7806 376 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7807 :
7808 : /*
7809 : * Delete old log files, those no longer needed for last restartpoint to
7810 : * prevent the disk holding the xlog from growing full.
7811 : */
7812 376 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7813 :
7814 : /*
7815 : * Retreat _logSegNo using the current end of xlog replayed or received,
7816 : * whichever is later.
7817 : */
7818 376 : receivePtr = GetWalRcvFlushRecPtr(NULL, NULL);
7819 376 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
7820 376 : endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
7821 376 : KeepLogSeg(endptr, &_logSegNo);
7822 376 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7823 : _logSegNo, InvalidOid,
7824 : InvalidTransactionId))
7825 : {
7826 : /*
7827 : * Some slots have been invalidated; recalculate the old-segment
7828 : * horizon, starting again from RedoRecPtr.
7829 : */
7830 2 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7831 2 : KeepLogSeg(endptr, &_logSegNo);
7832 : }
7833 376 : _logSegNo--;
7834 :
7835 : /*
7836 : * Try to recycle segments on a useful timeline. If we've been promoted
7837 : * since the beginning of this restartpoint, use the new timeline chosen
7838 : * at end of recovery. If we're still in recovery, use the timeline we're
7839 : * currently replaying.
7840 : *
7841 : * There is no guarantee that the WAL segments will be useful on the
7842 : * current timeline; if recovery proceeds to a new timeline right after
7843 : * this, the pre-allocated WAL segments on this timeline will not be used,
7844 : * and will go wasted until recycled on the next restartpoint. We'll live
7845 : * with that.
7846 : */
7847 376 : if (!RecoveryInProgress())
7848 0 : replayTLI = XLogCtl->InsertTimeLineID;
7849 :
7850 376 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, endptr, replayTLI);
7851 :
7852 : /*
7853 : * Make more log segments if needed. (Do this after recycling old log
7854 : * segments, since that may supply some of the needed files.)
7855 : */
7856 376 : PreallocXlogFiles(endptr, replayTLI);
7857 :
7858 : /*
7859 : * Truncate pg_subtrans if possible. We can throw away all data before
7860 : * the oldest XMIN of any running transaction. No future transaction will
7861 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7862 : * in subtrans.c). When hot standby is disabled, though, we mustn't do
7863 : * this because StartupSUBTRANS hasn't been called yet.
7864 : */
7865 376 : if (EnableHotStandby)
7866 376 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7867 :
7868 : /* Real work is done; log and update stats. */
7869 376 : LogCheckpointEnd(true);
7870 :
7871 : /* Reset the process title */
7872 376 : update_checkpoint_display(flags, true, true);
7873 :
7874 376 : xtime = GetLatestXTime();
7875 376 : ereport((log_checkpoints ? LOG : DEBUG2),
7876 : (errmsg("recovery restart point at %X/%X",
7877 : LSN_FORMAT_ARGS(lastCheckPoint.redo)),
7878 : xtime ? errdetail("Last completed transaction was at log time %s.",
7879 : timestamptz_to_str(xtime)) : 0));
7880 :
7881 : /*
7882 : * Finally, execute archive_cleanup_command, if any.
7883 : */
7884 376 : if (archiveCleanupCommand && strcmp(archiveCleanupCommand, "") != 0)
7885 0 : ExecuteRecoveryCommand(archiveCleanupCommand,
7886 : "archive_cleanup_command",
7887 : false,
7888 : WAIT_EVENT_ARCHIVE_CLEANUP_COMMAND);
7889 :
7890 376 : return true;
7891 : }
7892 :
7893 : /*
7894 : * Report availability of WAL for the given target LSN
7895 : * (typically a slot's restart_lsn)
7896 : *
7897 : * Returns one of the following enum values:
7898 : *
7899 : * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
7900 : * max_wal_size.
7901 : *
7902 : * * WALAVAIL_EXTENDED means it is still available by preserving extra
7903 : * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
7904 : * than max_wal_size, this state is not returned.
7905 : *
7906 : * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
7907 : * remove reserved segments. The walsender using this slot may return to the
7908 : * above.
7909 : *
7910 : * * WALAVAIL_REMOVED means it has been removed. A replication stream on
7911 : * a slot with this LSN cannot continue. (Any associated walsender
7912 : * processes should have been terminated already.)
7913 : *
7914 : * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
7915 : */
7916 : WALAvailability
7917 860 : GetWALAvailability(XLogRecPtr targetLSN)
7918 : {
7919 : XLogRecPtr currpos; /* current write LSN */
7920 : XLogSegNo currSeg; /* segid of currpos */
7921 : XLogSegNo targetSeg; /* segid of targetLSN */
7922 : XLogSegNo oldestSeg; /* actual oldest segid */
7923 : XLogSegNo oldestSegMaxWalSize; /* oldest segid kept by max_wal_size */
7924 : XLogSegNo oldestSlotSeg; /* oldest segid kept by slot */
7925 : uint64 keepSegs;
7926 :
7927 : /*
7928 : * slot does not reserve WAL. Either deactivated, or has never been active
7929 : */
7930 860 : if (XLogRecPtrIsInvalid(targetLSN))
7931 26 : return WALAVAIL_INVALID_LSN;
7932 :
7933 : /*
7934 : * Calculate the oldest segment currently reserved by all slots,
7935 : * considering wal_keep_size and max_slot_wal_keep_size. Initialize
7936 : * oldestSlotSeg to the current segment.
7937 : */
7938 834 : currpos = GetXLogWriteRecPtr();
7939 834 : XLByteToSeg(currpos, oldestSlotSeg, wal_segment_size);
7940 834 : KeepLogSeg(currpos, &oldestSlotSeg);
7941 :
7942 : /*
7943 : * Find the oldest extant segment file. We get 1 until checkpoint removes
7944 : * the first WAL segment file since startup, which causes the status being
7945 : * wrong under certain abnormal conditions but that doesn't actually harm.
7946 : */
7947 834 : oldestSeg = XLogGetLastRemovedSegno() + 1;
7948 :
7949 : /* calculate oldest segment by max_wal_size */
7950 834 : XLByteToSeg(currpos, currSeg, wal_segment_size);
7951 834 : keepSegs = ConvertToXSegs(max_wal_size_mb, wal_segment_size) + 1;
7952 :
7953 834 : if (currSeg > keepSegs)
7954 16 : oldestSegMaxWalSize = currSeg - keepSegs;
7955 : else
7956 818 : oldestSegMaxWalSize = 1;
7957 :
7958 : /* the segment we care about */
7959 834 : XLByteToSeg(targetLSN, targetSeg, wal_segment_size);
7960 :
7961 : /*
7962 : * No point in returning reserved or extended status values if the
7963 : * targetSeg is known to be lost.
7964 : */
7965 834 : if (targetSeg >= oldestSlotSeg)
7966 : {
7967 : /* show "reserved" when targetSeg is within max_wal_size */
7968 832 : if (targetSeg >= oldestSegMaxWalSize)
7969 828 : return WALAVAIL_RESERVED;
7970 :
7971 : /* being retained by slots exceeding max_wal_size */
7972 4 : return WALAVAIL_EXTENDED;
7973 : }
7974 :
7975 : /* WAL segments are no longer retained but haven't been removed yet */
7976 2 : if (targetSeg >= oldestSeg)
7977 2 : return WALAVAIL_UNRESERVED;
7978 :
7979 : /* Definitely lost */
7980 0 : return WALAVAIL_REMOVED;
7981 : }
7982 :
7983 :
7984 : /*
7985 : * Retreat *logSegNo to the last segment that we need to retain because of
7986 : * either wal_keep_size or replication slots.
7987 : *
7988 : * This is calculated by subtracting wal_keep_size from the given xlog
7989 : * location, recptr and by making sure that that result is below the
7990 : * requirement of replication slots. For the latter criterion we do consider
7991 : * the effects of max_slot_wal_keep_size: reserve at most that much space back
7992 : * from recptr.
7993 : *
7994 : * Note about replication slots: if this function calculates a value
7995 : * that's further ahead than what slots need reserved, then affected
7996 : * slots need to be invalidated and this function invoked again.
7997 : * XXX it might be a good idea to rewrite this function so that
7998 : * invalidation is optionally done here, instead.
7999 : */
8000 : static void
8001 3448 : KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
8002 : {
8003 : XLogSegNo currSegNo;
8004 : XLogSegNo segno;
8005 : XLogRecPtr keep;
8006 :
8007 3448 : XLByteToSeg(recptr, currSegNo, wal_segment_size);
8008 3448 : segno = currSegNo;
8009 :
8010 : /*
8011 : * Calculate how many segments are kept by slots first, adjusting for
8012 : * max_slot_wal_keep_size.
8013 : */
8014 3448 : keep = XLogGetReplicationSlotMinimumLSN();
8015 3448 : if (keep != InvalidXLogRecPtr && keep < recptr)
8016 : {
8017 1070 : XLByteToSeg(keep, segno, wal_segment_size);
8018 :
8019 : /* Cap by max_slot_wal_keep_size ... */
8020 1070 : if (max_slot_wal_keep_size_mb >= 0)
8021 : {
8022 : uint64 slot_keep_segs;
8023 :
8024 38 : slot_keep_segs =
8025 38 : ConvertToXSegs(max_slot_wal_keep_size_mb, wal_segment_size);
8026 :
8027 38 : if (currSegNo - segno > slot_keep_segs)
8028 10 : segno = currSegNo - slot_keep_segs;
8029 : }
8030 : }
8031 :
8032 : /*
8033 : * If WAL summarization is in use, don't remove WAL that has yet to be
8034 : * summarized.
8035 : */
8036 3448 : keep = GetOldestUnsummarizedLSN(NULL, NULL);
8037 3448 : if (keep != InvalidXLogRecPtr)
8038 : {
8039 : XLogSegNo unsummarized_segno;
8040 :
8041 4 : XLByteToSeg(keep, unsummarized_segno, wal_segment_size);
8042 4 : if (unsummarized_segno < segno)
8043 4 : segno = unsummarized_segno;
8044 : }
8045 :
8046 : /* but, keep at least wal_keep_size if that's set */
8047 3448 : if (wal_keep_size_mb > 0)
8048 : {
8049 : uint64 keep_segs;
8050 :
8051 138 : keep_segs = ConvertToXSegs(wal_keep_size_mb, wal_segment_size);
8052 138 : if (currSegNo - segno < keep_segs)
8053 : {
8054 : /* avoid underflow, don't go below 1 */
8055 138 : if (currSegNo <= keep_segs)
8056 130 : segno = 1;
8057 : else
8058 8 : segno = currSegNo - keep_segs;
8059 : }
8060 : }
8061 :
8062 : /* don't delete WAL segments newer than the calculated segment */
8063 3448 : if (segno < *logSegNo)
8064 618 : *logSegNo = segno;
8065 3448 : }
8066 :
8067 : /*
8068 : * Write a NEXTOID log record
8069 : */
8070 : void
8071 1128 : XLogPutNextOid(Oid nextOid)
8072 : {
8073 1128 : XLogBeginInsert();
8074 1128 : XLogRegisterData(&nextOid, sizeof(Oid));
8075 1128 : (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID);
8076 :
8077 : /*
8078 : * We need not flush the NEXTOID record immediately, because any of the
8079 : * just-allocated OIDs could only reach disk as part of a tuple insert or
8080 : * update that would have its own XLOG record that must follow the NEXTOID
8081 : * record. Therefore, the standard buffer LSN interlock applied to those
8082 : * records will ensure no such OID reaches disk before the NEXTOID record
8083 : * does.
8084 : *
8085 : * Note, however, that the above statement only covers state "within" the
8086 : * database. When we use a generated OID as a file or directory name, we
8087 : * are in a sense violating the basic WAL rule, because that filesystem
8088 : * change may reach disk before the NEXTOID WAL record does. The impact
8089 : * of this is that if a database crash occurs immediately afterward, we
8090 : * might after restart re-generate the same OID and find that it conflicts
8091 : * with the leftover file or directory. But since for safety's sake we
8092 : * always loop until finding a nonconflicting filename, this poses no real
8093 : * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8094 : */
8095 1128 : }
8096 :
8097 : /*
8098 : * Write an XLOG SWITCH record.
8099 : *
8100 : * Here we just blindly issue an XLogInsert request for the record.
8101 : * All the magic happens inside XLogInsert.
8102 : *
8103 : * The return value is either the end+1 address of the switch record,
8104 : * or the end+1 address of the prior segment if we did not need to
8105 : * write a switch record because we are already at segment start.
8106 : */
8107 : XLogRecPtr
8108 750 : RequestXLogSwitch(bool mark_unimportant)
8109 : {
8110 : XLogRecPtr RecPtr;
8111 :
8112 : /* XLOG SWITCH has no data */
8113 750 : XLogBeginInsert();
8114 :
8115 750 : if (mark_unimportant)
8116 0 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8117 750 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH);
8118 :
8119 750 : return RecPtr;
8120 : }
8121 :
8122 : /*
8123 : * Write a RESTORE POINT record
8124 : */
8125 : XLogRecPtr
8126 6 : XLogRestorePoint(const char *rpName)
8127 : {
8128 : XLogRecPtr RecPtr;
8129 : xl_restore_point xlrec;
8130 :
8131 6 : xlrec.rp_time = GetCurrentTimestamp();
8132 6 : strlcpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8133 :
8134 6 : XLogBeginInsert();
8135 6 : XLogRegisterData(&xlrec, sizeof(xl_restore_point));
8136 :
8137 6 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT);
8138 :
8139 6 : ereport(LOG,
8140 : (errmsg("restore point \"%s\" created at %X/%X",
8141 : rpName, LSN_FORMAT_ARGS(RecPtr))));
8142 :
8143 6 : return RecPtr;
8144 : }
8145 :
8146 : /*
8147 : * Check if any of the GUC parameters that are critical for hot standby
8148 : * have changed, and update the value in pg_control file if necessary.
8149 : */
8150 : static void
8151 1648 : XLogReportParameters(void)
8152 : {
8153 1648 : if (wal_level != ControlFile->wal_level ||
8154 1200 : wal_log_hints != ControlFile->wal_log_hints ||
8155 1042 : MaxConnections != ControlFile->MaxConnections ||
8156 1040 : max_worker_processes != ControlFile->max_worker_processes ||
8157 1038 : max_wal_senders != ControlFile->max_wal_senders ||
8158 994 : max_prepared_xacts != ControlFile->max_prepared_xacts ||
8159 818 : max_locks_per_xact != ControlFile->max_locks_per_xact ||
8160 818 : track_commit_timestamp != ControlFile->track_commit_timestamp)
8161 : {
8162 : /*
8163 : * The change in number of backend slots doesn't need to be WAL-logged
8164 : * if archiving is not enabled, as you can't start archive recovery
8165 : * with wal_level=minimal anyway. We don't really care about the
8166 : * values in pg_control either if wal_level=minimal, but seems better
8167 : * to keep them up-to-date to avoid confusion.
8168 : */
8169 850 : if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8170 : {
8171 : xl_parameter_change xlrec;
8172 : XLogRecPtr recptr;
8173 :
8174 806 : xlrec.MaxConnections = MaxConnections;
8175 806 : xlrec.max_worker_processes = max_worker_processes;
8176 806 : xlrec.max_wal_senders = max_wal_senders;
8177 806 : xlrec.max_prepared_xacts = max_prepared_xacts;
8178 806 : xlrec.max_locks_per_xact = max_locks_per_xact;
8179 806 : xlrec.wal_level = wal_level;
8180 806 : xlrec.wal_log_hints = wal_log_hints;
8181 806 : xlrec.track_commit_timestamp = track_commit_timestamp;
8182 :
8183 806 : XLogBeginInsert();
8184 806 : XLogRegisterData(&xlrec, sizeof(xlrec));
8185 :
8186 806 : recptr = XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE);
8187 806 : XLogFlush(recptr);
8188 : }
8189 :
8190 850 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8191 :
8192 850 : ControlFile->MaxConnections = MaxConnections;
8193 850 : ControlFile->max_worker_processes = max_worker_processes;
8194 850 : ControlFile->max_wal_senders = max_wal_senders;
8195 850 : ControlFile->max_prepared_xacts = max_prepared_xacts;
8196 850 : ControlFile->max_locks_per_xact = max_locks_per_xact;
8197 850 : ControlFile->wal_level = wal_level;
8198 850 : ControlFile->wal_log_hints = wal_log_hints;
8199 850 : ControlFile->track_commit_timestamp = track_commit_timestamp;
8200 850 : UpdateControlFile();
8201 :
8202 850 : LWLockRelease(ControlFileLock);
8203 : }
8204 1648 : }
8205 :
8206 : /*
8207 : * Update full_page_writes in shared memory, and write an
8208 : * XLOG_FPW_CHANGE record if necessary.
8209 : *
8210 : * Note: this function assumes there is no other process running
8211 : * concurrently that could update it.
8212 : */
8213 : void
8214 2738 : UpdateFullPageWrites(void)
8215 : {
8216 2738 : XLogCtlInsert *Insert = &XLogCtl->Insert;
8217 : bool recoveryInProgress;
8218 :
8219 : /*
8220 : * Do nothing if full_page_writes has not been changed.
8221 : *
8222 : * It's safe to check the shared full_page_writes without the lock,
8223 : * because we assume that there is no concurrently running process which
8224 : * can update it.
8225 : */
8226 2738 : if (fullPageWrites == Insert->fullPageWrites)
8227 1782 : return;
8228 :
8229 : /*
8230 : * Perform this outside critical section so that the WAL insert
8231 : * initialization done by RecoveryInProgress() doesn't trigger an
8232 : * assertion failure.
8233 : */
8234 956 : recoveryInProgress = RecoveryInProgress();
8235 :
8236 956 : START_CRIT_SECTION();
8237 :
8238 : /*
8239 : * It's always safe to take full page images, even when not strictly
8240 : * required, but not the other round. So if we're setting full_page_writes
8241 : * to true, first set it true and then write the WAL record. If we're
8242 : * setting it to false, first write the WAL record and then set the global
8243 : * flag.
8244 : */
8245 956 : if (fullPageWrites)
8246 : {
8247 936 : WALInsertLockAcquireExclusive();
8248 936 : Insert->fullPageWrites = true;
8249 936 : WALInsertLockRelease();
8250 : }
8251 :
8252 : /*
8253 : * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8254 : * full_page_writes during archive recovery, if required.
8255 : */
8256 956 : if (XLogStandbyInfoActive() && !recoveryInProgress)
8257 : {
8258 0 : XLogBeginInsert();
8259 0 : XLogRegisterData(&fullPageWrites, sizeof(bool));
8260 :
8261 0 : XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
8262 : }
8263 :
8264 956 : if (!fullPageWrites)
8265 : {
8266 20 : WALInsertLockAcquireExclusive();
8267 20 : Insert->fullPageWrites = false;
8268 20 : WALInsertLockRelease();
8269 : }
8270 956 : END_CRIT_SECTION();
8271 : }
8272 :
8273 : /*
8274 : * XLOG resource manager's routines
8275 : *
8276 : * Definitions of info values are in include/catalog/pg_control.h, though
8277 : * not all record types are related to control file updates.
8278 : *
8279 : * NOTE: Some XLOG record types that are directly related to WAL recovery
8280 : * are handled in xlogrecovery_redo().
8281 : */
8282 : void
8283 79542 : xlog_redo(XLogReaderState *record)
8284 : {
8285 79542 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8286 79542 : XLogRecPtr lsn = record->EndRecPtr;
8287 :
8288 : /*
8289 : * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
8290 : * XLOG_FPI_FOR_HINT records.
8291 : */
8292 : Assert(info == XLOG_FPI || info == XLOG_FPI_FOR_HINT ||
8293 : !XLogRecHasAnyBlockRefs(record));
8294 :
8295 79542 : if (info == XLOG_NEXTOID)
8296 : {
8297 : Oid nextOid;
8298 :
8299 : /*
8300 : * We used to try to take the maximum of TransamVariables->nextOid and
8301 : * the recorded nextOid, but that fails if the OID counter wraps
8302 : * around. Since no OID allocation should be happening during replay
8303 : * anyway, better to just believe the record exactly. We still take
8304 : * OidGenLock while setting the variable, just in case.
8305 : */
8306 184 : memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
8307 184 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8308 184 : TransamVariables->nextOid = nextOid;
8309 184 : TransamVariables->oidCount = 0;
8310 184 : LWLockRelease(OidGenLock);
8311 : }
8312 79358 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
8313 : {
8314 : CheckPoint checkPoint;
8315 : TimeLineID replayTLI;
8316 :
8317 68 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8318 : /* In a SHUTDOWN checkpoint, believe the counters exactly */
8319 68 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8320 68 : TransamVariables->nextXid = checkPoint.nextXid;
8321 68 : LWLockRelease(XidGenLock);
8322 68 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8323 68 : TransamVariables->nextOid = checkPoint.nextOid;
8324 68 : TransamVariables->oidCount = 0;
8325 68 : LWLockRelease(OidGenLock);
8326 68 : MultiXactSetNextMXact(checkPoint.nextMulti,
8327 : checkPoint.nextMultiOffset);
8328 :
8329 68 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8330 : checkPoint.oldestMultiDB);
8331 :
8332 : /*
8333 : * No need to set oldestClogXid here as well; it'll be set when we
8334 : * redo an xl_clog_truncate if it changed since initialization.
8335 : */
8336 68 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
8337 :
8338 : /*
8339 : * If we see a shutdown checkpoint while waiting for an end-of-backup
8340 : * record, the backup was canceled and the end-of-backup record will
8341 : * never arrive.
8342 : */
8343 68 : if (ArchiveRecoveryRequested &&
8344 66 : !XLogRecPtrIsInvalid(ControlFile->backupStartPoint) &&
8345 0 : XLogRecPtrIsInvalid(ControlFile->backupEndPoint))
8346 0 : ereport(PANIC,
8347 : (errmsg("online backup was canceled, recovery cannot continue")));
8348 :
8349 : /*
8350 : * If we see a shutdown checkpoint, we know that nothing was running
8351 : * on the primary at this point. So fake-up an empty running-xacts
8352 : * record and use that here and now. Recover additional standby state
8353 : * for prepared transactions.
8354 : */
8355 68 : if (standbyState >= STANDBY_INITIALIZED)
8356 : {
8357 : TransactionId *xids;
8358 : int nxids;
8359 : TransactionId oldestActiveXID;
8360 : TransactionId latestCompletedXid;
8361 : RunningTransactionsData running;
8362 :
8363 62 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
8364 :
8365 : /* Update pg_subtrans entries for any prepared transactions */
8366 62 : StandbyRecoverPreparedTransactions();
8367 :
8368 : /*
8369 : * Construct a RunningTransactions snapshot representing a shut
8370 : * down server, with only prepared transactions still alive. We're
8371 : * never overflowed at this point because all subxids are listed
8372 : * with their parent prepared transactions.
8373 : */
8374 62 : running.xcnt = nxids;
8375 62 : running.subxcnt = 0;
8376 62 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
8377 62 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
8378 62 : running.oldestRunningXid = oldestActiveXID;
8379 62 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
8380 62 : TransactionIdRetreat(latestCompletedXid);
8381 : Assert(TransactionIdIsNormal(latestCompletedXid));
8382 62 : running.latestCompletedXid = latestCompletedXid;
8383 62 : running.xids = xids;
8384 :
8385 62 : ProcArrayApplyRecoveryInfo(&running);
8386 : }
8387 :
8388 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8389 68 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8390 68 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8391 68 : LWLockRelease(ControlFileLock);
8392 :
8393 : /* Update shared-memory copy of checkpoint XID/epoch */
8394 68 : SpinLockAcquire(&XLogCtl->info_lck);
8395 68 : XLogCtl->ckptFullXid = checkPoint.nextXid;
8396 68 : SpinLockRelease(&XLogCtl->info_lck);
8397 :
8398 : /*
8399 : * We should've already switched to the new TLI before replaying this
8400 : * record.
8401 : */
8402 68 : (void) GetCurrentReplayRecPtr(&replayTLI);
8403 68 : if (checkPoint.ThisTimeLineID != replayTLI)
8404 0 : ereport(PANIC,
8405 : (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
8406 : checkPoint.ThisTimeLineID, replayTLI)));
8407 :
8408 68 : RecoveryRestartPoint(&checkPoint, record);
8409 : }
8410 79290 : else if (info == XLOG_CHECKPOINT_ONLINE)
8411 : {
8412 : CheckPoint checkPoint;
8413 : TimeLineID replayTLI;
8414 :
8415 696 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8416 : /* In an ONLINE checkpoint, treat the XID counter as a minimum */
8417 696 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8418 696 : if (FullTransactionIdPrecedes(TransamVariables->nextXid,
8419 : checkPoint.nextXid))
8420 0 : TransamVariables->nextXid = checkPoint.nextXid;
8421 696 : LWLockRelease(XidGenLock);
8422 :
8423 : /*
8424 : * We ignore the nextOid counter in an ONLINE checkpoint, preferring
8425 : * to track OID assignment through XLOG_NEXTOID records. The nextOid
8426 : * counter is from the start of the checkpoint and might well be stale
8427 : * compared to later XLOG_NEXTOID records. We could try to take the
8428 : * maximum of the nextOid counter and our latest value, but since
8429 : * there's no particular guarantee about the speed with which the OID
8430 : * counter wraps around, that's a risky thing to do. In any case,
8431 : * users of the nextOid counter are required to avoid assignment of
8432 : * duplicates, so that a somewhat out-of-date value should be safe.
8433 : */
8434 :
8435 : /* Handle multixact */
8436 696 : MultiXactAdvanceNextMXact(checkPoint.nextMulti,
8437 : checkPoint.nextMultiOffset);
8438 :
8439 : /*
8440 : * NB: This may perform multixact truncation when replaying WAL
8441 : * generated by an older primary.
8442 : */
8443 696 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8444 : checkPoint.oldestMultiDB);
8445 696 : if (TransactionIdPrecedes(TransamVariables->oldestXid,
8446 : checkPoint.oldestXid))
8447 0 : SetTransactionIdLimit(checkPoint.oldestXid,
8448 : checkPoint.oldestXidDB);
8449 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8450 696 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8451 696 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8452 696 : LWLockRelease(ControlFileLock);
8453 :
8454 : /* Update shared-memory copy of checkpoint XID/epoch */
8455 696 : SpinLockAcquire(&XLogCtl->info_lck);
8456 696 : XLogCtl->ckptFullXid = checkPoint.nextXid;
8457 696 : SpinLockRelease(&XLogCtl->info_lck);
8458 :
8459 : /* TLI should not change in an on-line checkpoint */
8460 696 : (void) GetCurrentReplayRecPtr(&replayTLI);
8461 696 : if (checkPoint.ThisTimeLineID != replayTLI)
8462 0 : ereport(PANIC,
8463 : (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
8464 : checkPoint.ThisTimeLineID, replayTLI)));
8465 :
8466 696 : RecoveryRestartPoint(&checkPoint, record);
8467 : }
8468 78594 : else if (info == XLOG_OVERWRITE_CONTRECORD)
8469 : {
8470 : /* nothing to do here, handled in xlogrecovery_redo() */
8471 : }
8472 78592 : else if (info == XLOG_END_OF_RECOVERY)
8473 : {
8474 : xl_end_of_recovery xlrec;
8475 : TimeLineID replayTLI;
8476 :
8477 20 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
8478 :
8479 : /*
8480 : * For Hot Standby, we could treat this like a Shutdown Checkpoint,
8481 : * but this case is rarer and harder to test, so the benefit doesn't
8482 : * outweigh the potential extra cost of maintenance.
8483 : */
8484 :
8485 : /*
8486 : * We should've already switched to the new TLI before replaying this
8487 : * record.
8488 : */
8489 20 : (void) GetCurrentReplayRecPtr(&replayTLI);
8490 20 : if (xlrec.ThisTimeLineID != replayTLI)
8491 0 : ereport(PANIC,
8492 : (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
8493 : xlrec.ThisTimeLineID, replayTLI)));
8494 : }
8495 78572 : else if (info == XLOG_NOOP)
8496 : {
8497 : /* nothing to do here */
8498 : }
8499 78572 : else if (info == XLOG_SWITCH)
8500 : {
8501 : /* nothing to do here */
8502 : }
8503 78332 : else if (info == XLOG_RESTORE_POINT)
8504 : {
8505 : /* nothing to do here, handled in xlogrecovery.c */
8506 : }
8507 78322 : else if (info == XLOG_FPI || info == XLOG_FPI_FOR_HINT)
8508 : {
8509 : /*
8510 : * XLOG_FPI records contain nothing else but one or more block
8511 : * references. Every block reference must include a full-page image
8512 : * even if full_page_writes was disabled when the record was generated
8513 : * - otherwise there would be no point in this record.
8514 : *
8515 : * XLOG_FPI_FOR_HINT records are generated when a page needs to be
8516 : * WAL-logged because of a hint bit update. They are only generated
8517 : * when checksums and/or wal_log_hints are enabled. They may include
8518 : * no full-page images if full_page_writes was disabled when they were
8519 : * generated. In this case there is nothing to do here.
8520 : *
8521 : * No recovery conflicts are generated by these generic records - if a
8522 : * resource manager needs to generate conflicts, it has to define a
8523 : * separate WAL record type and redo routine.
8524 : */
8525 164402 : for (uint8 block_id = 0; block_id <= XLogRecMaxBlockId(record); block_id++)
8526 : {
8527 : Buffer buffer;
8528 :
8529 87008 : if (!XLogRecHasBlockImage(record, block_id))
8530 : {
8531 132 : if (info == XLOG_FPI)
8532 0 : elog(ERROR, "XLOG_FPI record did not contain a full-page image");
8533 132 : continue;
8534 : }
8535 :
8536 86876 : if (XLogReadBufferForRedo(record, block_id, &buffer) != BLK_RESTORED)
8537 0 : elog(ERROR, "unexpected XLogReadBufferForRedo result when restoring backup block");
8538 86876 : UnlockReleaseBuffer(buffer);
8539 : }
8540 : }
8541 928 : else if (info == XLOG_BACKUP_END)
8542 : {
8543 : /* nothing to do here, handled in xlogrecovery_redo() */
8544 : }
8545 762 : else if (info == XLOG_PARAMETER_CHANGE)
8546 : {
8547 : xl_parameter_change xlrec;
8548 :
8549 : /* Update our copy of the parameters in pg_control */
8550 64 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
8551 :
8552 : /*
8553 : * Invalidate logical slots if we are in hot standby and the primary
8554 : * does not have a WAL level sufficient for logical decoding. No need
8555 : * to search for potentially conflicting logically slots if standby is
8556 : * running with wal_level lower than logical, because in that case, we
8557 : * would have either disallowed creation of logical slots or
8558 : * invalidated existing ones.
8559 : */
8560 64 : if (InRecovery && InHotStandby &&
8561 34 : xlrec.wal_level < WAL_LEVEL_LOGICAL &&
8562 12 : wal_level >= WAL_LEVEL_LOGICAL)
8563 8 : InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_LEVEL,
8564 : 0, InvalidOid,
8565 : InvalidTransactionId);
8566 :
8567 64 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8568 64 : ControlFile->MaxConnections = xlrec.MaxConnections;
8569 64 : ControlFile->max_worker_processes = xlrec.max_worker_processes;
8570 64 : ControlFile->max_wal_senders = xlrec.max_wal_senders;
8571 64 : ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
8572 64 : ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
8573 64 : ControlFile->wal_level = xlrec.wal_level;
8574 64 : ControlFile->wal_log_hints = xlrec.wal_log_hints;
8575 :
8576 : /*
8577 : * Update minRecoveryPoint to ensure that if recovery is aborted, we
8578 : * recover back up to this point before allowing hot standby again.
8579 : * This is important if the max_* settings are decreased, to ensure
8580 : * you don't run queries against the WAL preceding the change. The
8581 : * local copies cannot be updated as long as crash recovery is
8582 : * happening and we expect all the WAL to be replayed.
8583 : */
8584 64 : if (InArchiveRecovery)
8585 : {
8586 36 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
8587 36 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
8588 : }
8589 64 : if (LocalMinRecoveryPoint != InvalidXLogRecPtr && LocalMinRecoveryPoint < lsn)
8590 : {
8591 : TimeLineID replayTLI;
8592 :
8593 14 : (void) GetCurrentReplayRecPtr(&replayTLI);
8594 14 : ControlFile->minRecoveryPoint = lsn;
8595 14 : ControlFile->minRecoveryPointTLI = replayTLI;
8596 : }
8597 :
8598 64 : CommitTsParameterChange(xlrec.track_commit_timestamp,
8599 64 : ControlFile->track_commit_timestamp);
8600 64 : ControlFile->track_commit_timestamp = xlrec.track_commit_timestamp;
8601 :
8602 64 : UpdateControlFile();
8603 64 : LWLockRelease(ControlFileLock);
8604 :
8605 : /* Check to see if any parameter change gives a problem on recovery */
8606 64 : CheckRequiredParameterValues();
8607 : }
8608 698 : else if (info == XLOG_FPW_CHANGE)
8609 : {
8610 : bool fpw;
8611 :
8612 0 : memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
8613 :
8614 : /*
8615 : * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
8616 : * do_pg_backup_start() and do_pg_backup_stop() can check whether
8617 : * full_page_writes has been disabled during online backup.
8618 : */
8619 0 : if (!fpw)
8620 : {
8621 0 : SpinLockAcquire(&XLogCtl->info_lck);
8622 0 : if (XLogCtl->lastFpwDisableRecPtr < record->ReadRecPtr)
8623 0 : XLogCtl->lastFpwDisableRecPtr = record->ReadRecPtr;
8624 0 : SpinLockRelease(&XLogCtl->info_lck);
8625 : }
8626 :
8627 : /* Keep track of full_page_writes */
8628 0 : lastFullPageWrites = fpw;
8629 : }
8630 : else if (info == XLOG_CHECKPOINT_REDO)
8631 : {
8632 : /* nothing to do here, just for informational purposes */
8633 : }
8634 79538 : }
8635 :
8636 : /*
8637 : * Return the extra open flags used for opening a file, depending on the
8638 : * value of the GUCs wal_sync_method, fsync and debug_io_direct.
8639 : */
8640 : static int
8641 27672 : get_sync_bit(int method)
8642 : {
8643 27672 : int o_direct_flag = 0;
8644 :
8645 : /*
8646 : * Use O_DIRECT if requested, except in walreceiver process. The WAL
8647 : * written by walreceiver is normally read by the startup process soon
8648 : * after it's written. Also, walreceiver performs unaligned writes, which
8649 : * don't work with O_DIRECT, so it is required for correctness too.
8650 : */
8651 27672 : if ((io_direct_flags & IO_DIRECT_WAL) && !AmWalReceiverProcess())
8652 18 : o_direct_flag = PG_O_DIRECT;
8653 :
8654 : /* If fsync is disabled, never open in sync mode */
8655 27672 : if (!enableFsync)
8656 27672 : return o_direct_flag;
8657 :
8658 0 : switch (method)
8659 : {
8660 : /*
8661 : * enum values for all sync options are defined even if they are
8662 : * not supported on the current platform. But if not, they are
8663 : * not included in the enum option array, and therefore will never
8664 : * be seen here.
8665 : */
8666 0 : case WAL_SYNC_METHOD_FSYNC:
8667 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8668 : case WAL_SYNC_METHOD_FDATASYNC:
8669 0 : return o_direct_flag;
8670 : #ifdef O_SYNC
8671 0 : case WAL_SYNC_METHOD_OPEN:
8672 0 : return O_SYNC | o_direct_flag;
8673 : #endif
8674 : #ifdef O_DSYNC
8675 0 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8676 0 : return O_DSYNC | o_direct_flag;
8677 : #endif
8678 0 : default:
8679 : /* can't happen (unless we are out of sync with option array) */
8680 0 : elog(ERROR, "unrecognized \"wal_sync_method\": %d", method);
8681 : return 0; /* silence warning */
8682 : }
8683 : }
8684 :
8685 : /*
8686 : * GUC support
8687 : */
8688 : void
8689 2098 : assign_wal_sync_method(int new_wal_sync_method, void *extra)
8690 : {
8691 2098 : if (wal_sync_method != new_wal_sync_method)
8692 : {
8693 : /*
8694 : * To ensure that no blocks escape unsynced, force an fsync on the
8695 : * currently open log segment (if any). Also, if the open flag is
8696 : * changing, close the log file so it will be reopened (with new flag
8697 : * bit) at next use.
8698 : */
8699 0 : if (openLogFile >= 0)
8700 : {
8701 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN);
8702 0 : if (pg_fsync(openLogFile) != 0)
8703 : {
8704 : char xlogfname[MAXFNAMELEN];
8705 : int save_errno;
8706 :
8707 0 : save_errno = errno;
8708 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo,
8709 : wal_segment_size);
8710 0 : errno = save_errno;
8711 0 : ereport(PANIC,
8712 : (errcode_for_file_access(),
8713 : errmsg("could not fsync file \"%s\": %m", xlogfname)));
8714 : }
8715 :
8716 0 : pgstat_report_wait_end();
8717 0 : if (get_sync_bit(wal_sync_method) != get_sync_bit(new_wal_sync_method))
8718 0 : XLogFileClose();
8719 : }
8720 : }
8721 2098 : }
8722 :
8723 :
8724 : /*
8725 : * Issue appropriate kind of fsync (if any) for an XLOG output file.
8726 : *
8727 : * 'fd' is a file descriptor for the XLOG file to be fsync'd.
8728 : * 'segno' is for error reporting purposes.
8729 : */
8730 : void
8731 269464 : issue_xlog_fsync(int fd, XLogSegNo segno, TimeLineID tli)
8732 : {
8733 269464 : char *msg = NULL;
8734 : instr_time start;
8735 :
8736 : Assert(tli != 0);
8737 :
8738 : /*
8739 : * Quick exit if fsync is disabled or write() has already synced the WAL
8740 : * file.
8741 : */
8742 269464 : if (!enableFsync ||
8743 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN ||
8744 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN_DSYNC)
8745 269464 : return;
8746 :
8747 : /*
8748 : * Measure I/O timing to sync the WAL file for pg_stat_io.
8749 : */
8750 0 : start = pgstat_prepare_io_time(track_wal_io_timing);
8751 :
8752 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC);
8753 0 : switch (wal_sync_method)
8754 : {
8755 0 : case WAL_SYNC_METHOD_FSYNC:
8756 0 : if (pg_fsync_no_writethrough(fd) != 0)
8757 0 : msg = _("could not fsync file \"%s\": %m");
8758 0 : break;
8759 : #ifdef HAVE_FSYNC_WRITETHROUGH
8760 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8761 : if (pg_fsync_writethrough(fd) != 0)
8762 : msg = _("could not fsync write-through file \"%s\": %m");
8763 : break;
8764 : #endif
8765 0 : case WAL_SYNC_METHOD_FDATASYNC:
8766 0 : if (pg_fdatasync(fd) != 0)
8767 0 : msg = _("could not fdatasync file \"%s\": %m");
8768 0 : break;
8769 0 : case WAL_SYNC_METHOD_OPEN:
8770 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8771 : /* not reachable */
8772 : Assert(false);
8773 0 : break;
8774 0 : default:
8775 0 : ereport(PANIC,
8776 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8777 : errmsg_internal("unrecognized \"wal_sync_method\": %d", wal_sync_method));
8778 : break;
8779 : }
8780 :
8781 : /* PANIC if failed to fsync */
8782 0 : if (msg)
8783 : {
8784 : char xlogfname[MAXFNAMELEN];
8785 0 : int save_errno = errno;
8786 :
8787 0 : XLogFileName(xlogfname, tli, segno, wal_segment_size);
8788 0 : errno = save_errno;
8789 0 : ereport(PANIC,
8790 : (errcode_for_file_access(),
8791 : errmsg(msg, xlogfname)));
8792 : }
8793 :
8794 0 : pgstat_report_wait_end();
8795 :
8796 0 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL, IOOP_FSYNC,
8797 : start, 1, 0);
8798 : }
8799 :
8800 : /*
8801 : * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
8802 : * function. It creates the necessary starting checkpoint and constructs the
8803 : * backup state and tablespace map.
8804 : *
8805 : * Input parameters are "state" (the backup state), "fast" (if true, we do
8806 : * the checkpoint in immediate mode to make it faster), and "tablespaces"
8807 : * (if non-NULL, indicates a list of tablespaceinfo structs describing the
8808 : * cluster's tablespaces.).
8809 : *
8810 : * The tablespace map contents are appended to passed-in parameter
8811 : * tablespace_map and the caller is responsible for including it in the backup
8812 : * archive as 'tablespace_map'. The tablespace_map file is required mainly for
8813 : * tar format in windows as native windows utilities are not able to create
8814 : * symlinks while extracting files from tar. However for consistency and
8815 : * platform-independence, we do it the same way everywhere.
8816 : *
8817 : * It fills in "state" with the information required for the backup, such
8818 : * as the minimum WAL location that must be present to restore from this
8819 : * backup (starttli) and the corresponding timeline ID (starttli).
8820 : *
8821 : * Every successfully started backup must be stopped by calling
8822 : * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
8823 : * backups active at the same time.
8824 : *
8825 : * It is the responsibility of the caller of this function to verify the
8826 : * permissions of the calling user!
8827 : */
8828 : void
8829 324 : do_pg_backup_start(const char *backupidstr, bool fast, List **tablespaces,
8830 : BackupState *state, StringInfo tblspcmapfile)
8831 : {
8832 : bool backup_started_in_recovery;
8833 :
8834 : Assert(state != NULL);
8835 324 : backup_started_in_recovery = RecoveryInProgress();
8836 :
8837 : /*
8838 : * During recovery, we don't need to check WAL level. Because, if WAL
8839 : * level is not sufficient, it's impossible to get here during recovery.
8840 : */
8841 324 : if (!backup_started_in_recovery && !XLogIsNeeded())
8842 0 : ereport(ERROR,
8843 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8844 : errmsg("WAL level not sufficient for making an online backup"),
8845 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
8846 :
8847 324 : if (strlen(backupidstr) > MAXPGPATH)
8848 2 : ereport(ERROR,
8849 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8850 : errmsg("backup label too long (max %d bytes)",
8851 : MAXPGPATH)));
8852 :
8853 322 : strlcpy(state->name, backupidstr, sizeof(state->name));
8854 :
8855 : /*
8856 : * Mark backup active in shared memory. We must do full-page WAL writes
8857 : * during an on-line backup even if not doing so at other times, because
8858 : * it's quite possible for the backup dump to obtain a "torn" (partially
8859 : * written) copy of a database page if it reads the page concurrently with
8860 : * our write to the same page. This can be fixed as long as the first
8861 : * write to the page in the WAL sequence is a full-page write. Hence, we
8862 : * increment runningBackups then force a CHECKPOINT, to ensure there are
8863 : * no dirty pages in shared memory that might get dumped while the backup
8864 : * is in progress without having a corresponding WAL record. (Once the
8865 : * backup is complete, we need not force full-page writes anymore, since
8866 : * we expect that any pages not modified during the backup interval must
8867 : * have been correctly captured by the backup.)
8868 : *
8869 : * Note that forcing full-page writes has no effect during an online
8870 : * backup from the standby.
8871 : *
8872 : * We must hold all the insertion locks to change the value of
8873 : * runningBackups, to ensure adequate interlocking against
8874 : * XLogInsertRecord().
8875 : */
8876 322 : WALInsertLockAcquireExclusive();
8877 322 : XLogCtl->Insert.runningBackups++;
8878 322 : WALInsertLockRelease();
8879 :
8880 : /*
8881 : * Ensure we decrement runningBackups if we fail below. NB -- for this to
8882 : * work correctly, it is critical that sessionBackupState is only updated
8883 : * after this block is over.
8884 : */
8885 322 : PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, DatumGetBool(true));
8886 : {
8887 322 : bool gotUniqueStartpoint = false;
8888 : DIR *tblspcdir;
8889 : struct dirent *de;
8890 : tablespaceinfo *ti;
8891 : int datadirpathlen;
8892 :
8893 : /*
8894 : * Force an XLOG file switch before the checkpoint, to ensure that the
8895 : * WAL segment the checkpoint is written to doesn't contain pages with
8896 : * old timeline IDs. That would otherwise happen if you called
8897 : * pg_backup_start() right after restoring from a PITR archive: the
8898 : * first WAL segment containing the startup checkpoint has pages in
8899 : * the beginning with the old timeline ID. That can cause trouble at
8900 : * recovery: we won't have a history file covering the old timeline if
8901 : * pg_wal directory was not included in the base backup and the WAL
8902 : * archive was cleared too before starting the backup.
8903 : *
8904 : * This also ensures that we have emitted a WAL page header that has
8905 : * XLP_BKP_REMOVABLE off before we emit the checkpoint record.
8906 : * Therefore, if a WAL archiver (such as pglesslog) is trying to
8907 : * compress out removable backup blocks, it won't remove any that
8908 : * occur after this point.
8909 : *
8910 : * During recovery, we skip forcing XLOG file switch, which means that
8911 : * the backup taken during recovery is not available for the special
8912 : * recovery case described above.
8913 : */
8914 322 : if (!backup_started_in_recovery)
8915 308 : RequestXLogSwitch(false);
8916 :
8917 : do
8918 : {
8919 : bool checkpointfpw;
8920 :
8921 : /*
8922 : * Force a CHECKPOINT. Aside from being necessary to prevent torn
8923 : * page problems, this guarantees that two successive backup runs
8924 : * will have different checkpoint positions and hence different
8925 : * history file names, even if nothing happened in between.
8926 : *
8927 : * During recovery, establish a restartpoint if possible. We use
8928 : * the last restartpoint as the backup starting checkpoint. This
8929 : * means that two successive backup runs can have same checkpoint
8930 : * positions.
8931 : *
8932 : * Since the fact that we are executing do_pg_backup_start()
8933 : * during recovery means that checkpointer is running, we can use
8934 : * RequestCheckpoint() to establish a restartpoint.
8935 : *
8936 : * We use CHECKPOINT_IMMEDIATE only if requested by user (via
8937 : * passing fast = true). Otherwise this can take awhile.
8938 : */
8939 322 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
8940 : (fast ? CHECKPOINT_IMMEDIATE : 0));
8941 :
8942 : /*
8943 : * Now we need to fetch the checkpoint record location, and also
8944 : * its REDO pointer. The oldest point in WAL that would be needed
8945 : * to restore starting from the checkpoint is precisely the REDO
8946 : * pointer.
8947 : */
8948 322 : LWLockAcquire(ControlFileLock, LW_SHARED);
8949 322 : state->checkpointloc = ControlFile->checkPoint;
8950 322 : state->startpoint = ControlFile->checkPointCopy.redo;
8951 322 : state->starttli = ControlFile->checkPointCopy.ThisTimeLineID;
8952 322 : checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
8953 322 : LWLockRelease(ControlFileLock);
8954 :
8955 322 : if (backup_started_in_recovery)
8956 : {
8957 : XLogRecPtr recptr;
8958 :
8959 : /*
8960 : * Check to see if all WAL replayed during online backup
8961 : * (i.e., since last restartpoint used as backup starting
8962 : * checkpoint) contain full-page writes.
8963 : */
8964 14 : SpinLockAcquire(&XLogCtl->info_lck);
8965 14 : recptr = XLogCtl->lastFpwDisableRecPtr;
8966 14 : SpinLockRelease(&XLogCtl->info_lck);
8967 :
8968 14 : if (!checkpointfpw || state->startpoint <= recptr)
8969 0 : ereport(ERROR,
8970 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8971 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
8972 : "since last restartpoint"),
8973 : errhint("This means that the backup being taken on the standby "
8974 : "is corrupt and should not be used. "
8975 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
8976 : "and then try an online backup again.")));
8977 :
8978 : /*
8979 : * During recovery, since we don't use the end-of-backup WAL
8980 : * record and don't write the backup history file, the
8981 : * starting WAL location doesn't need to be unique. This means
8982 : * that two base backups started at the same time might use
8983 : * the same checkpoint as starting locations.
8984 : */
8985 14 : gotUniqueStartpoint = true;
8986 : }
8987 :
8988 : /*
8989 : * If two base backups are started at the same time (in WAL sender
8990 : * processes), we need to make sure that they use different
8991 : * checkpoints as starting locations, because we use the starting
8992 : * WAL location as a unique identifier for the base backup in the
8993 : * end-of-backup WAL record and when we write the backup history
8994 : * file. Perhaps it would be better generate a separate unique ID
8995 : * for each backup instead of forcing another checkpoint, but
8996 : * taking a checkpoint right after another is not that expensive
8997 : * either because only few buffers have been dirtied yet.
8998 : */
8999 322 : WALInsertLockAcquireExclusive();
9000 322 : if (XLogCtl->Insert.lastBackupStart < state->startpoint)
9001 : {
9002 322 : XLogCtl->Insert.lastBackupStart = state->startpoint;
9003 322 : gotUniqueStartpoint = true;
9004 : }
9005 322 : WALInsertLockRelease();
9006 322 : } while (!gotUniqueStartpoint);
9007 :
9008 : /*
9009 : * Construct tablespace_map file.
9010 : */
9011 322 : datadirpathlen = strlen(DataDir);
9012 :
9013 : /* Collect information about all tablespaces */
9014 322 : tblspcdir = AllocateDir(PG_TBLSPC_DIR);
9015 1038 : while ((de = ReadDir(tblspcdir, PG_TBLSPC_DIR)) != NULL)
9016 : {
9017 : char fullpath[MAXPGPATH + sizeof(PG_TBLSPC_DIR)];
9018 : char linkpath[MAXPGPATH];
9019 716 : char *relpath = NULL;
9020 : char *s;
9021 : PGFileType de_type;
9022 : char *badp;
9023 : Oid tsoid;
9024 :
9025 : /*
9026 : * Try to parse the directory name as an unsigned integer.
9027 : *
9028 : * Tablespace directories should be positive integers that can be
9029 : * represented in 32 bits, with no leading zeroes or trailing
9030 : * garbage. If we come across a name that doesn't meet those
9031 : * criteria, skip it.
9032 : */
9033 716 : if (de->d_name[0] < '1' || de->d_name[1] > '9')
9034 644 : continue;
9035 72 : errno = 0;
9036 72 : tsoid = strtoul(de->d_name, &badp, 10);
9037 72 : if (*badp != '\0' || errno == EINVAL || errno == ERANGE)
9038 0 : continue;
9039 :
9040 72 : snprintf(fullpath, sizeof(fullpath), "%s/%s", PG_TBLSPC_DIR, de->d_name);
9041 :
9042 72 : de_type = get_dirent_type(fullpath, de, false, ERROR);
9043 :
9044 72 : if (de_type == PGFILETYPE_LNK)
9045 : {
9046 : StringInfoData escapedpath;
9047 : int rllen;
9048 :
9049 44 : rllen = readlink(fullpath, linkpath, sizeof(linkpath));
9050 44 : if (rllen < 0)
9051 : {
9052 0 : ereport(WARNING,
9053 : (errmsg("could not read symbolic link \"%s\": %m",
9054 : fullpath)));
9055 0 : continue;
9056 : }
9057 44 : else if (rllen >= sizeof(linkpath))
9058 : {
9059 0 : ereport(WARNING,
9060 : (errmsg("symbolic link \"%s\" target is too long",
9061 : fullpath)));
9062 0 : continue;
9063 : }
9064 44 : linkpath[rllen] = '\0';
9065 :
9066 : /*
9067 : * Relpath holds the relative path of the tablespace directory
9068 : * when it's located within PGDATA, or NULL if it's located
9069 : * elsewhere.
9070 : */
9071 44 : if (rllen > datadirpathlen &&
9072 2 : strncmp(linkpath, DataDir, datadirpathlen) == 0 &&
9073 0 : IS_DIR_SEP(linkpath[datadirpathlen]))
9074 0 : relpath = pstrdup(linkpath + datadirpathlen + 1);
9075 :
9076 : /*
9077 : * Add a backslash-escaped version of the link path to the
9078 : * tablespace map file.
9079 : */
9080 44 : initStringInfo(&escapedpath);
9081 1092 : for (s = linkpath; *s; s++)
9082 : {
9083 1048 : if (*s == '\n' || *s == '\r' || *s == '\\')
9084 0 : appendStringInfoChar(&escapedpath, '\\');
9085 1048 : appendStringInfoChar(&escapedpath, *s);
9086 : }
9087 44 : appendStringInfo(tblspcmapfile, "%s %s\n",
9088 44 : de->d_name, escapedpath.data);
9089 44 : pfree(escapedpath.data);
9090 : }
9091 28 : else if (de_type == PGFILETYPE_DIR)
9092 : {
9093 : /*
9094 : * It's possible to use allow_in_place_tablespaces to create
9095 : * directories directly under pg_tblspc, for testing purposes
9096 : * only.
9097 : *
9098 : * In this case, we store a relative path rather than an
9099 : * absolute path into the tablespaceinfo.
9100 : */
9101 28 : snprintf(linkpath, sizeof(linkpath), "%s/%s",
9102 28 : PG_TBLSPC_DIR, de->d_name);
9103 28 : relpath = pstrdup(linkpath);
9104 : }
9105 : else
9106 : {
9107 : /* Skip any other file type that appears here. */
9108 0 : continue;
9109 : }
9110 :
9111 72 : ti = palloc(sizeof(tablespaceinfo));
9112 72 : ti->oid = tsoid;
9113 72 : ti->path = pstrdup(linkpath);
9114 72 : ti->rpath = relpath;
9115 72 : ti->size = -1;
9116 :
9117 72 : if (tablespaces)
9118 72 : *tablespaces = lappend(*tablespaces, ti);
9119 : }
9120 322 : FreeDir(tblspcdir);
9121 :
9122 322 : state->starttime = (pg_time_t) time(NULL);
9123 : }
9124 322 : PG_END_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, DatumGetBool(true));
9125 :
9126 322 : state->started_in_recovery = backup_started_in_recovery;
9127 :
9128 : /*
9129 : * Mark that the start phase has correctly finished for the backup.
9130 : */
9131 322 : sessionBackupState = SESSION_BACKUP_RUNNING;
9132 322 : }
9133 :
9134 : /*
9135 : * Utility routine to fetch the session-level status of a backup running.
9136 : */
9137 : SessionBackupState
9138 366 : get_backup_status(void)
9139 : {
9140 366 : return sessionBackupState;
9141 : }
9142 :
9143 : /*
9144 : * do_pg_backup_stop
9145 : *
9146 : * Utility function called at the end of an online backup. It creates history
9147 : * file (if required), resets sessionBackupState and so on. It can optionally
9148 : * wait for WAL segments to be archived.
9149 : *
9150 : * "state" is filled with the information necessary to restore from this
9151 : * backup with its stop LSN (stoppoint), its timeline ID (stoptli), etc.
9152 : *
9153 : * It is the responsibility of the caller of this function to verify the
9154 : * permissions of the calling user!
9155 : */
9156 : void
9157 310 : do_pg_backup_stop(BackupState *state, bool waitforarchive)
9158 : {
9159 310 : bool backup_stopped_in_recovery = false;
9160 : char histfilepath[MAXPGPATH];
9161 : char lastxlogfilename[MAXFNAMELEN];
9162 : char histfilename[MAXFNAMELEN];
9163 : XLogSegNo _logSegNo;
9164 : FILE *fp;
9165 : int seconds_before_warning;
9166 310 : int waits = 0;
9167 310 : bool reported_waiting = false;
9168 :
9169 : Assert(state != NULL);
9170 :
9171 310 : backup_stopped_in_recovery = RecoveryInProgress();
9172 :
9173 : /*
9174 : * During recovery, we don't need to check WAL level. Because, if WAL
9175 : * level is not sufficient, it's impossible to get here during recovery.
9176 : */
9177 310 : if (!backup_stopped_in_recovery && !XLogIsNeeded())
9178 0 : ereport(ERROR,
9179 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9180 : errmsg("WAL level not sufficient for making an online backup"),
9181 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9182 :
9183 : /*
9184 : * OK to update backup counter and session-level lock.
9185 : *
9186 : * Note that CHECK_FOR_INTERRUPTS() must not occur while updating them,
9187 : * otherwise they can be updated inconsistently, which might cause
9188 : * do_pg_abort_backup() to fail.
9189 : */
9190 310 : WALInsertLockAcquireExclusive();
9191 :
9192 : /*
9193 : * It is expected that each do_pg_backup_start() call is matched by
9194 : * exactly one do_pg_backup_stop() call.
9195 : */
9196 : Assert(XLogCtl->Insert.runningBackups > 0);
9197 310 : XLogCtl->Insert.runningBackups--;
9198 :
9199 : /*
9200 : * Clean up session-level lock.
9201 : *
9202 : * You might think that WALInsertLockRelease() can be called before
9203 : * cleaning up session-level lock because session-level lock doesn't need
9204 : * to be protected with WAL insertion lock. But since
9205 : * CHECK_FOR_INTERRUPTS() can occur in it, session-level lock must be
9206 : * cleaned up before it.
9207 : */
9208 310 : sessionBackupState = SESSION_BACKUP_NONE;
9209 :
9210 310 : WALInsertLockRelease();
9211 :
9212 : /*
9213 : * If we are taking an online backup from the standby, we confirm that the
9214 : * standby has not been promoted during the backup.
9215 : */
9216 310 : if (state->started_in_recovery && !backup_stopped_in_recovery)
9217 0 : ereport(ERROR,
9218 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9219 : errmsg("the standby was promoted during online backup"),
9220 : errhint("This means that the backup being taken is corrupt "
9221 : "and should not be used. "
9222 : "Try taking another online backup.")));
9223 :
9224 : /*
9225 : * During recovery, we don't write an end-of-backup record. We assume that
9226 : * pg_control was backed up last and its minimum recovery point can be
9227 : * available as the backup end location. Since we don't have an
9228 : * end-of-backup record, we use the pg_control value to check whether
9229 : * we've reached the end of backup when starting recovery from this
9230 : * backup. We have no way of checking if pg_control wasn't backed up last
9231 : * however.
9232 : *
9233 : * We don't force a switch to new WAL file but it is still possible to
9234 : * wait for all the required files to be archived if waitforarchive is
9235 : * true. This is okay if we use the backup to start a standby and fetch
9236 : * the missing WAL using streaming replication. But in the case of an
9237 : * archive recovery, a user should set waitforarchive to true and wait for
9238 : * them to be archived to ensure that all the required files are
9239 : * available.
9240 : *
9241 : * We return the current minimum recovery point as the backup end
9242 : * location. Note that it can be greater than the exact backup end
9243 : * location if the minimum recovery point is updated after the backup of
9244 : * pg_control. This is harmless for current uses.
9245 : *
9246 : * XXX currently a backup history file is for informational and debug
9247 : * purposes only. It's not essential for an online backup. Furthermore,
9248 : * even if it's created, it will not be archived during recovery because
9249 : * an archiver is not invoked. So it doesn't seem worthwhile to write a
9250 : * backup history file during recovery.
9251 : */
9252 310 : if (backup_stopped_in_recovery)
9253 : {
9254 : XLogRecPtr recptr;
9255 :
9256 : /*
9257 : * Check to see if all WAL replayed during online backup contain
9258 : * full-page writes.
9259 : */
9260 14 : SpinLockAcquire(&XLogCtl->info_lck);
9261 14 : recptr = XLogCtl->lastFpwDisableRecPtr;
9262 14 : SpinLockRelease(&XLogCtl->info_lck);
9263 :
9264 14 : if (state->startpoint <= recptr)
9265 0 : ereport(ERROR,
9266 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9267 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9268 : "during online backup"),
9269 : errhint("This means that the backup being taken on the standby "
9270 : "is corrupt and should not be used. "
9271 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9272 : "and then try an online backup again.")));
9273 :
9274 :
9275 14 : LWLockAcquire(ControlFileLock, LW_SHARED);
9276 14 : state->stoppoint = ControlFile->minRecoveryPoint;
9277 14 : state->stoptli = ControlFile->minRecoveryPointTLI;
9278 14 : LWLockRelease(ControlFileLock);
9279 : }
9280 : else
9281 : {
9282 : char *history_file;
9283 :
9284 : /*
9285 : * Write the backup-end xlog record
9286 : */
9287 296 : XLogBeginInsert();
9288 296 : XLogRegisterData(&state->startpoint,
9289 : sizeof(state->startpoint));
9290 296 : state->stoppoint = XLogInsert(RM_XLOG_ID, XLOG_BACKUP_END);
9291 :
9292 : /*
9293 : * Given that we're not in recovery, InsertTimeLineID is set and can't
9294 : * change, so we can read it without a lock.
9295 : */
9296 296 : state->stoptli = XLogCtl->InsertTimeLineID;
9297 :
9298 : /*
9299 : * Force a switch to a new xlog segment file, so that the backup is
9300 : * valid as soon as archiver moves out the current segment file.
9301 : */
9302 296 : RequestXLogSwitch(false);
9303 :
9304 296 : state->stoptime = (pg_time_t) time(NULL);
9305 :
9306 : /*
9307 : * Write the backup history file
9308 : */
9309 296 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9310 296 : BackupHistoryFilePath(histfilepath, state->stoptli, _logSegNo,
9311 : state->startpoint, wal_segment_size);
9312 296 : fp = AllocateFile(histfilepath, "w");
9313 296 : if (!fp)
9314 0 : ereport(ERROR,
9315 : (errcode_for_file_access(),
9316 : errmsg("could not create file \"%s\": %m",
9317 : histfilepath)));
9318 :
9319 : /* Build and save the contents of the backup history file */
9320 296 : history_file = build_backup_content(state, true);
9321 296 : fprintf(fp, "%s", history_file);
9322 296 : pfree(history_file);
9323 :
9324 296 : if (fflush(fp) || ferror(fp) || FreeFile(fp))
9325 0 : ereport(ERROR,
9326 : (errcode_for_file_access(),
9327 : errmsg("could not write file \"%s\": %m",
9328 : histfilepath)));
9329 :
9330 : /*
9331 : * Clean out any no-longer-needed history files. As a side effect,
9332 : * this will post a .ready file for the newly created history file,
9333 : * notifying the archiver that history file may be archived
9334 : * immediately.
9335 : */
9336 296 : CleanupBackupHistory();
9337 : }
9338 :
9339 : /*
9340 : * If archiving is enabled, wait for all the required WAL files to be
9341 : * archived before returning. If archiving isn't enabled, the required WAL
9342 : * needs to be transported via streaming replication (hopefully with
9343 : * wal_keep_size set high enough), or some more exotic mechanism like
9344 : * polling and copying files from pg_wal with script. We have no knowledge
9345 : * of those mechanisms, so it's up to the user to ensure that he gets all
9346 : * the required WAL.
9347 : *
9348 : * We wait until both the last WAL file filled during backup and the
9349 : * history file have been archived, and assume that the alphabetic sorting
9350 : * property of the WAL files ensures any earlier WAL files are safely
9351 : * archived as well.
9352 : *
9353 : * We wait forever, since archive_command is supposed to work and we
9354 : * assume the admin wanted his backup to work completely. If you don't
9355 : * wish to wait, then either waitforarchive should be passed in as false,
9356 : * or you can set statement_timeout. Also, some notices are issued to
9357 : * clue in anyone who might be doing this interactively.
9358 : */
9359 :
9360 310 : if (waitforarchive &&
9361 22 : ((!backup_stopped_in_recovery && XLogArchivingActive()) ||
9362 2 : (backup_stopped_in_recovery && XLogArchivingAlways())))
9363 : {
9364 10 : XLByteToPrevSeg(state->stoppoint, _logSegNo, wal_segment_size);
9365 10 : XLogFileName(lastxlogfilename, state->stoptli, _logSegNo,
9366 : wal_segment_size);
9367 :
9368 10 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9369 10 : BackupHistoryFileName(histfilename, state->stoptli, _logSegNo,
9370 : state->startpoint, wal_segment_size);
9371 :
9372 10 : seconds_before_warning = 60;
9373 10 : waits = 0;
9374 :
9375 30 : while (XLogArchiveIsBusy(lastxlogfilename) ||
9376 10 : XLogArchiveIsBusy(histfilename))
9377 : {
9378 10 : CHECK_FOR_INTERRUPTS();
9379 :
9380 10 : if (!reported_waiting && waits > 5)
9381 : {
9382 0 : ereport(NOTICE,
9383 : (errmsg("base backup done, waiting for required WAL segments to be archived")));
9384 0 : reported_waiting = true;
9385 : }
9386 :
9387 10 : (void) WaitLatch(MyLatch,
9388 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
9389 : 1000L,
9390 : WAIT_EVENT_BACKUP_WAIT_WAL_ARCHIVE);
9391 10 : ResetLatch(MyLatch);
9392 :
9393 10 : if (++waits >= seconds_before_warning)
9394 : {
9395 0 : seconds_before_warning *= 2; /* This wraps in >10 years... */
9396 0 : ereport(WARNING,
9397 : (errmsg("still waiting for all required WAL segments to be archived (%d seconds elapsed)",
9398 : waits),
9399 : errhint("Check that your \"archive_command\" is executing properly. "
9400 : "You can safely cancel this backup, "
9401 : "but the database backup will not be usable without all the WAL segments.")));
9402 : }
9403 : }
9404 :
9405 10 : ereport(NOTICE,
9406 : (errmsg("all required WAL segments have been archived")));
9407 : }
9408 300 : else if (waitforarchive)
9409 12 : ereport(NOTICE,
9410 : (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
9411 310 : }
9412 :
9413 :
9414 : /*
9415 : * do_pg_abort_backup: abort a running backup
9416 : *
9417 : * This does just the most basic steps of do_pg_backup_stop(), by taking the
9418 : * system out of backup mode, thus making it a lot more safe to call from
9419 : * an error handler.
9420 : *
9421 : * 'arg' indicates that it's being called during backup setup; so
9422 : * sessionBackupState has not been modified yet, but runningBackups has
9423 : * already been incremented. When it's false, then it's invoked as a
9424 : * before_shmem_exit handler, and therefore we must not change state
9425 : * unless sessionBackupState indicates that a backup is actually running.
9426 : *
9427 : * NB: This gets used as a PG_ENSURE_ERROR_CLEANUP callback and
9428 : * before_shmem_exit handler, hence the odd-looking signature.
9429 : */
9430 : void
9431 18 : do_pg_abort_backup(int code, Datum arg)
9432 : {
9433 18 : bool during_backup_start = DatumGetBool(arg);
9434 :
9435 : /* If called during backup start, there shouldn't be one already running */
9436 : Assert(!during_backup_start || sessionBackupState == SESSION_BACKUP_NONE);
9437 :
9438 18 : if (during_backup_start || sessionBackupState != SESSION_BACKUP_NONE)
9439 : {
9440 12 : WALInsertLockAcquireExclusive();
9441 : Assert(XLogCtl->Insert.runningBackups > 0);
9442 12 : XLogCtl->Insert.runningBackups--;
9443 :
9444 12 : sessionBackupState = SESSION_BACKUP_NONE;
9445 12 : WALInsertLockRelease();
9446 :
9447 12 : if (!during_backup_start)
9448 12 : ereport(WARNING,
9449 : errmsg("aborting backup due to backend exiting before pg_backup_stop was called"));
9450 : }
9451 18 : }
9452 :
9453 : /*
9454 : * Register a handler that will warn about unterminated backups at end of
9455 : * session, unless this has already been done.
9456 : */
9457 : void
9458 10 : register_persistent_abort_backup_handler(void)
9459 : {
9460 : static bool already_done = false;
9461 :
9462 10 : if (already_done)
9463 2 : return;
9464 8 : before_shmem_exit(do_pg_abort_backup, DatumGetBool(false));
9465 8 : already_done = true;
9466 : }
9467 :
9468 : /*
9469 : * Get latest WAL insert pointer
9470 : */
9471 : XLogRecPtr
9472 3986 : GetXLogInsertRecPtr(void)
9473 : {
9474 3986 : XLogCtlInsert *Insert = &XLogCtl->Insert;
9475 : uint64 current_bytepos;
9476 :
9477 3986 : SpinLockAcquire(&Insert->insertpos_lck);
9478 3986 : current_bytepos = Insert->CurrBytePos;
9479 3986 : SpinLockRelease(&Insert->insertpos_lck);
9480 :
9481 3986 : return XLogBytePosToRecPtr(current_bytepos);
9482 : }
9483 :
9484 : /*
9485 : * Get latest WAL write pointer
9486 : */
9487 : XLogRecPtr
9488 2744 : GetXLogWriteRecPtr(void)
9489 : {
9490 2744 : RefreshXLogWriteResult(LogwrtResult);
9491 :
9492 2744 : return LogwrtResult.Write;
9493 : }
9494 :
9495 : /*
9496 : * Returns the redo pointer of the last checkpoint or restartpoint. This is
9497 : * the oldest point in WAL that we still need, if we have to restart recovery.
9498 : */
9499 : void
9500 140 : GetOldestRestartPoint(XLogRecPtr *oldrecptr, TimeLineID *oldtli)
9501 : {
9502 140 : LWLockAcquire(ControlFileLock, LW_SHARED);
9503 140 : *oldrecptr = ControlFile->checkPointCopy.redo;
9504 140 : *oldtli = ControlFile->checkPointCopy.ThisTimeLineID;
9505 140 : LWLockRelease(ControlFileLock);
9506 140 : }
9507 :
9508 : /* Thin wrapper around ShutdownWalRcv(). */
9509 : void
9510 1964 : XLogShutdownWalRcv(void)
9511 : {
9512 1964 : ShutdownWalRcv();
9513 :
9514 1964 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9515 1964 : XLogCtl->InstallXLogFileSegmentActive = false;
9516 1964 : LWLockRelease(ControlFileLock);
9517 1964 : }
9518 :
9519 : /* Enable WAL file recycling and preallocation. */
9520 : void
9521 2052 : SetInstallXLogFileSegmentActive(void)
9522 : {
9523 2052 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9524 2052 : XLogCtl->InstallXLogFileSegmentActive = true;
9525 2052 : LWLockRelease(ControlFileLock);
9526 2052 : }
9527 :
9528 : bool
9529 0 : IsInstallXLogFileSegmentActive(void)
9530 : {
9531 : bool result;
9532 :
9533 0 : LWLockAcquire(ControlFileLock, LW_SHARED);
9534 0 : result = XLogCtl->InstallXLogFileSegmentActive;
9535 0 : LWLockRelease(ControlFileLock);
9536 :
9537 0 : return result;
9538 : }
9539 :
9540 : /*
9541 : * Update the WalWriterSleeping flag.
9542 : */
9543 : void
9544 908 : SetWalWriterSleeping(bool sleeping)
9545 : {
9546 908 : SpinLockAcquire(&XLogCtl->info_lck);
9547 908 : XLogCtl->WalWriterSleeping = sleeping;
9548 908 : SpinLockRelease(&XLogCtl->info_lck);
9549 908 : }
|
