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 28007562 : XLogInsertRecord(XLogRecData *rdata,
749 : XLogRecPtr fpw_lsn,
750 : uint8 flags,
751 : int num_fpi,
752 : bool topxid_included)
753 : {
754 28007562 : XLogCtlInsert *Insert = &XLogCtl->Insert;
755 : pg_crc32c rdata_crc;
756 : bool inserted;
757 28007562 : XLogRecord *rechdr = (XLogRecord *) rdata->data;
758 28007562 : uint8 info = rechdr->xl_info & ~XLR_INFO_MASK;
759 28007562 : WalInsertClass class = WALINSERT_NORMAL;
760 : XLogRecPtr StartPos;
761 : XLogRecPtr EndPos;
762 28007562 : bool prevDoPageWrites = doPageWrites;
763 : TimeLineID insertTLI;
764 :
765 : /* Does this record type require special handling? */
766 28007562 : if (unlikely(rechdr->xl_rmid == RM_XLOG_ID))
767 : {
768 404848 : if (info == XLOG_SWITCH)
769 738 : class = WALINSERT_SPECIAL_SWITCH;
770 404110 : else if (info == XLOG_CHECKPOINT_REDO)
771 1074 : 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 28007562 : 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 28007562 : 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 28007562 : START_CRIT_SECTION();
820 :
821 28007562 : if (likely(class == WALINSERT_NORMAL))
822 : {
823 28005750 : 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 28005750 : if (RedoRecPtr != Insert->RedoRecPtr)
842 : {
843 : Assert(RedoRecPtr < Insert->RedoRecPtr);
844 12636 : RedoRecPtr = Insert->RedoRecPtr;
845 : }
846 28005750 : doPageWrites = (Insert->fullPageWrites || Insert->runningBackups > 0);
847 :
848 28005750 : if (doPageWrites &&
849 27575678 : (!prevDoPageWrites ||
850 25319822 : (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 13970 : WALInsertLockRelease();
857 13970 : END_CRIT_SECTION();
858 13970 : return InvalidXLogRecPtr;
859 : }
860 :
861 : /*
862 : * Reserve space for the record in the WAL. This also sets the xl_prev
863 : * pointer.
864 : */
865 27991780 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
866 : &rechdr->xl_prev);
867 :
868 : /* Normal records are always inserted. */
869 27991780 : inserted = true;
870 : }
871 1812 : 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);
885 738 : WALInsertLockAcquireExclusive();
886 738 : 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 1074 : WALInsertLockAcquireExclusive();
901 1074 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
902 : &rechdr->xl_prev);
903 1074 : RedoRecPtr = Insert->RedoRecPtr = StartPos;
904 1074 : inserted = true;
905 : }
906 :
907 27993592 : if (inserted)
908 : {
909 : /*
910 : * Now that xl_prev has been filled in, calculate CRC of the record
911 : * header.
912 : */
913 27993480 : rdata_crc = rechdr->xl_crc;
914 27993480 : COMP_CRC32C(rdata_crc, rechdr, offsetof(XLogRecord, xl_crc));
915 27993480 : FIN_CRC32C(rdata_crc);
916 27993480 : 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 27993480 : 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 27993480 : if ((flags & XLOG_MARK_UNIMPORTANT) == 0)
932 : {
933 27820608 : int lockno = holdingAllLocks ? 0 : MyLockNo;
934 :
935 27820608 : 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 27993592 : WALInsertLockRelease();
951 :
952 27993592 : END_CRIT_SECTION();
953 :
954 27993592 : 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 27993592 : if (topxid_included)
961 438 : MarkSubxactTopXidLogged();
962 :
963 : /*
964 : * Update shared LogwrtRqst.Write, if we crossed page boundary.
965 : */
966 27993592 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
967 : {
968 3191996 : SpinLockAcquire(&XLogCtl->info_lck);
969 : /* advance global request to include new block(s) */
970 3191996 : if (XLogCtl->LogwrtRqst.Write < EndPos)
971 3049038 : XLogCtl->LogwrtRqst.Write = EndPos;
972 3191996 : SpinLockRelease(&XLogCtl->info_lck);
973 3191996 : 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 27993592 : if (class == WALINSERT_SPECIAL_SWITCH)
982 : {
983 : TRACE_POSTGRESQL_WAL_SWITCH();
984 738 : 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 : */
991 738 : if (inserted)
992 : {
993 626 : EndPos = StartPos + SizeOfXLogRecord;
994 626 : 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 27993592 : ProcLastRecPtr = StartPos;
1076 27993592 : XactLastRecEnd = EndPos;
1077 :
1078 : /* Report WAL traffic to the instrumentation. */
1079 27993592 : if (inserted)
1080 : {
1081 27993480 : pgWalUsage.wal_bytes += rechdr->xl_tot_len;
1082 27993480 : pgWalUsage.wal_records++;
1083 27993480 : pgWalUsage.wal_fpi += num_fpi;
1084 : }
1085 :
1086 27993592 : 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 27992854 : ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1109 : XLogRecPtr *PrevPtr)
1110 : {
1111 27992854 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1112 : uint64 startbytepos;
1113 : uint64 endbytepos;
1114 : uint64 prevbytepos;
1115 :
1116 27992854 : 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 27992854 : SpinLockAcquire(&Insert->insertpos_lck);
1132 :
1133 27992854 : startbytepos = Insert->CurrBytePos;
1134 27992854 : endbytepos = startbytepos + size;
1135 27992854 : prevbytepos = Insert->PrevBytePos;
1136 27992854 : Insert->CurrBytePos = endbytepos;
1137 27992854 : Insert->PrevBytePos = startbytepos;
1138 :
1139 27992854 : SpinLockRelease(&Insert->insertpos_lck);
1140 :
1141 27992854 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1142 27992854 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1143 27992854 : *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 27992854 : }
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 738 : ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1165 : {
1166 738 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1167 : uint64 startbytepos;
1168 : uint64 endbytepos;
1169 : uint64 prevbytepos;
1170 738 : 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 738 : SpinLockAcquire(&Insert->insertpos_lck);
1181 :
1182 738 : startbytepos = Insert->CurrBytePos;
1183 :
1184 738 : ptr = XLogBytePosToEndRecPtr(startbytepos);
1185 738 : if (XLogSegmentOffset(ptr, wal_segment_size) == 0)
1186 : {
1187 112 : SpinLockRelease(&Insert->insertpos_lck);
1188 112 : *EndPos = *StartPos = ptr;
1189 112 : return false;
1190 : }
1191 :
1192 626 : endbytepos = startbytepos + size;
1193 626 : prevbytepos = Insert->PrevBytePos;
1194 :
1195 626 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1196 626 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1197 :
1198 626 : segleft = wal_segment_size - XLogSegmentOffset(*EndPos, wal_segment_size);
1199 626 : if (segleft != wal_segment_size)
1200 : {
1201 : /* consume the rest of the segment */
1202 626 : *EndPos += segleft;
1203 626 : endbytepos = XLogRecPtrToBytePos(*EndPos);
1204 : }
1205 626 : Insert->CurrBytePos = endbytepos;
1206 626 : Insert->PrevBytePos = startbytepos;
1207 :
1208 626 : SpinLockRelease(&Insert->insertpos_lck);
1209 :
1210 626 : *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 626 : 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 27993480 : 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 27993480 : CurrPos = StartPos;
1239 27993480 : currpos = GetXLogBuffer(CurrPos, tli);
1240 27993480 : 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 27993480 : written = 0;
1250 132171124 : while (rdata != NULL)
1251 : {
1252 104177644 : const char *rdata_data = rdata->data;
1253 104177644 : int rdata_len = rdata->len;
1254 :
1255 107568766 : 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 3391122 : memcpy(currpos, rdata_data, freespace);
1262 3391122 : rdata_data += freespace;
1263 3391122 : rdata_len -= freespace;
1264 3391122 : written += freespace;
1265 3391122 : 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 3391122 : currpos = GetXLogBuffer(CurrPos, tli);
1277 3391122 : pagehdr = (XLogPageHeader) currpos;
1278 3391122 : pagehdr->xlp_rem_len = write_len - written;
1279 3391122 : pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1280 :
1281 : /* skip over the page header */
1282 3391122 : if (XLogSegmentOffset(CurrPos, wal_segment_size) == 0)
1283 : {
1284 2252 : CurrPos += SizeOfXLogLongPHD;
1285 2252 : currpos += SizeOfXLogLongPHD;
1286 : }
1287 : else
1288 : {
1289 3388870 : CurrPos += SizeOfXLogShortPHD;
1290 3388870 : currpos += SizeOfXLogShortPHD;
1291 : }
1292 3391122 : freespace = INSERT_FREESPACE(CurrPos);
1293 : }
1294 :
1295 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1296 104177644 : memcpy(currpos, rdata_data, rdata_len);
1297 104177644 : currpos += rdata_len;
1298 104177644 : CurrPos += rdata_len;
1299 104177644 : freespace -= rdata_len;
1300 104177644 : written += rdata_len;
1301 :
1302 104177644 : 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 27993480 : 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 626 : 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 956724 : 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 956098 : currpos = GetXLogBuffer(CurrPos, tli);
1350 3824392 : MemSet(currpos, 0, SizeOfXLogShortPHD);
1351 :
1352 956098 : CurrPos += XLOG_BLCKSZ;
1353 : }
1354 : }
1355 : else
1356 : {
1357 : /* Align the end position, so that the next record starts aligned */
1358 27992854 : CurrPos = MAXALIGN64(CurrPos);
1359 : }
1360 :
1361 27993480 : 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 27993480 : }
1366 :
1367 : /*
1368 : * Acquire a WAL insertion lock, for inserting to WAL.
1369 : */
1370 : static void
1371 28005772 : 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 28005772 : if (lockToTry == -1)
1389 14358 : lockToTry = MyProcNumber % NUM_XLOGINSERT_LOCKS;
1390 28005772 : 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 28005772 : immed = LWLockAcquire(&WALInsertLocks[MyLockNo].l.lock, LW_EXCLUSIVE);
1397 28005772 : 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 76740 : lockToTry = (lockToTry + 1) % NUM_XLOGINSERT_LOCKS;
1408 : }
1409 28005772 : }
1410 :
1411 : /*
1412 : * Acquire all WAL insertion locks, to prevent other backends from inserting
1413 : * to WAL.
1414 : */
1415 : static void
1416 6236 : 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 49888 : for (i = 0; i < NUM_XLOGINSERT_LOCKS - 1; i++)
1426 : {
1427 43652 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1428 43652 : LWLockUpdateVar(&WALInsertLocks[i].l.lock,
1429 43652 : &WALInsertLocks[i].l.insertingAt,
1430 : PG_UINT64_MAX);
1431 : }
1432 : /* Variable value reset to 0 at release */
1433 6236 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1434 :
1435 6236 : holdingAllLocks = true;
1436 6236 : }
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 28012008 : WALInsertLockRelease(void)
1446 : {
1447 28012008 : if (holdingAllLocks)
1448 : {
1449 : int i;
1450 :
1451 56124 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1452 49888 : LWLockReleaseClearVar(&WALInsertLocks[i].l.lock,
1453 49888 : &WALInsertLocks[i].l.insertingAt,
1454 : 0);
1455 :
1456 6236 : holdingAllLocks = false;
1457 : }
1458 : else
1459 : {
1460 28005772 : LWLockReleaseClearVar(&WALInsertLocks[MyLockNo].l.lock,
1461 28005772 : &WALInsertLocks[MyLockNo].l.insertingAt,
1462 : 0);
1463 : }
1464 28012008 : }
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 4806176 : WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt)
1472 : {
1473 4806176 : if (holdingAllLocks)
1474 : {
1475 : /*
1476 : * We use the last lock to mark our actual position, see comments in
1477 : * WALInsertLockAcquireExclusive.
1478 : */
1479 941280 : LWLockUpdateVar(&WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.lock,
1480 941280 : &WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.insertingAt,
1481 : insertingAt);
1482 : }
1483 : else
1484 3864896 : LWLockUpdateVar(&WALInsertLocks[MyLockNo].l.lock,
1485 3864896 : &WALInsertLocks[MyLockNo].l.insertingAt,
1486 : insertingAt);
1487 4806176 : }
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 4416484 : WaitXLogInsertionsToFinish(XLogRecPtr upto)
1505 : {
1506 : uint64 bytepos;
1507 : XLogRecPtr inserted;
1508 : XLogRecPtr reservedUpto;
1509 : XLogRecPtr finishedUpto;
1510 4416484 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1511 : int i;
1512 :
1513 4416484 : 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 4416484 : inserted = pg_atomic_read_membarrier_u64(&XLogCtl->logInsertResult);
1521 4416484 : if (upto <= inserted)
1522 3655718 : return inserted;
1523 :
1524 : /* Read the current insert position */
1525 760766 : SpinLockAcquire(&Insert->insertpos_lck);
1526 760766 : bytepos = Insert->CurrBytePos;
1527 760766 : SpinLockRelease(&Insert->insertpos_lck);
1528 760766 : 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 760766 : 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 760766 : finishedUpto = reservedUpto;
1556 6846894 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1557 : {
1558 6086128 : 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 6161618 : if (LWLockWaitForVar(&WALInsertLocks[i].l.lock,
1586 6161618 : &WALInsertLocks[i].l.insertingAt,
1587 : insertingat, &insertingat))
1588 : {
1589 : /* the lock was free, so no insertion in progress */
1590 4558678 : insertingat = InvalidXLogRecPtr;
1591 4558678 : break;
1592 : }
1593 :
1594 : /*
1595 : * This insertion is still in progress. Have to wait, unless the
1596 : * inserter has proceeded past 'upto'.
1597 : */
1598 1602940 : } while (insertingat < upto);
1599 :
1600 6086128 : if (insertingat != InvalidXLogRecPtr && insertingat < finishedUpto)
1601 623696 : 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 760766 : finishedUpto = pg_atomic_monotonic_advance_u64(&XLogCtl->logInsertResult,
1610 : finishedUpto);
1611 :
1612 760766 : 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 32340722 : 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 32340722 : 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 26584982 : 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 5755740 : 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 5755740 : expectedEndPtr = ptr;
1672 5755740 : expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
1673 :
1674 5755740 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1675 5755740 : 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 4806176 : if (ptr % XLOG_BLCKSZ == SizeOfXLogShortPHD &&
1694 11110 : XLogSegmentOffset(ptr, wal_segment_size) > XLOG_BLCKSZ)
1695 11110 : initializedUpto = ptr - SizeOfXLogShortPHD;
1696 4795066 : else if (ptr % XLOG_BLCKSZ == SizeOfXLogLongPHD &&
1697 1702 : XLogSegmentOffset(ptr, wal_segment_size) < XLOG_BLCKSZ)
1698 414 : initializedUpto = ptr - SizeOfXLogLongPHD;
1699 : else
1700 4794652 : initializedUpto = ptr;
1701 :
1702 4806176 : WALInsertLockUpdateInsertingAt(initializedUpto);
1703 :
1704 4806176 : AdvanceXLInsertBuffer(ptr, tli, false);
1705 4806176 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1706 :
1707 4806176 : 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 949564 : 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 5755740 : cachedPage = ptr / XLOG_BLCKSZ;
1725 5755740 : 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 5755740 : 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 187160 : WALReadFromBuffers(char *dstbuf, XLogRecPtr startptr, Size count,
1749 : TimeLineID tli)
1750 : {
1751 187160 : char *pdst = dstbuf;
1752 187160 : XLogRecPtr recptr = startptr;
1753 : XLogRecPtr inserted;
1754 187160 : Size nbytes = count;
1755 :
1756 187160 : if (RecoveryInProgress() || tli != GetWALInsertionTimeLine())
1757 1814 : 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 185346 : inserted = pg_atomic_read_u64(&XLogCtl->logInsertResult);
1766 185346 : 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 205324 : while (nbytes > 0)
1787 : {
1788 199786 : uint32 offset = recptr % XLOG_BLCKSZ;
1789 199786 : 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 199786 : 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 199786 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1807 199786 : if (expectedEndPtr != endptr)
1808 179804 : 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 19982 : page = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1816 19982 : psrc = page + offset;
1817 19982 : 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 19982 : pg_read_barrier();
1824 :
1825 : /* data copy */
1826 19982 : memcpy(pdst, psrc, npagebytes);
1827 :
1828 : /*
1829 : * Ensure that the data copy and the second verification step are not
1830 : * reordered.
1831 : */
1832 19982 : 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 19982 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1839 19982 : if (expectedEndPtr != endptr)
1840 4 : break;
1841 :
1842 19978 : pdst += npagebytes;
1843 19978 : recptr += npagebytes;
1844 19978 : nbytes -= npagebytes;
1845 : }
1846 :
1847 : Assert(pdst - dstbuf <= count);
1848 :
1849 185346 : 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 55991970 : XLogBytePosToRecPtr(uint64 bytepos)
1859 : {
1860 : uint64 fullsegs;
1861 : uint64 fullpages;
1862 : uint64 bytesleft;
1863 : uint32 seg_offset;
1864 : XLogRecPtr result;
1865 :
1866 55991970 : fullsegs = bytepos / UsableBytesInSegment;
1867 55991970 : bytesleft = bytepos % UsableBytesInSegment;
1868 :
1869 55991970 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1870 : {
1871 : /* fits on first page of segment */
1872 89992 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1873 : }
1874 : else
1875 : {
1876 : /* account for the first page on segment with long header */
1877 55901978 : seg_offset = XLOG_BLCKSZ;
1878 55901978 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1879 :
1880 55901978 : fullpages = bytesleft / UsableBytesInPage;
1881 55901978 : bytesleft = bytesleft % UsableBytesInPage;
1882 :
1883 55901978 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1884 : }
1885 :
1886 55991970 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1887 :
1888 55991970 : 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 28754984 : XLogBytePosToEndRecPtr(uint64 bytepos)
1899 : {
1900 : uint64 fullsegs;
1901 : uint64 fullpages;
1902 : uint64 bytesleft;
1903 : uint32 seg_offset;
1904 : XLogRecPtr result;
1905 :
1906 28754984 : fullsegs = bytepos / UsableBytesInSegment;
1907 28754984 : bytesleft = bytepos % UsableBytesInSegment;
1908 :
1909 28754984 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1910 : {
1911 : /* fits on first page of segment */
1912 141600 : if (bytesleft == 0)
1913 95434 : seg_offset = 0;
1914 : else
1915 46166 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1916 : }
1917 : else
1918 : {
1919 : /* account for the first page on segment with long header */
1920 28613384 : seg_offset = XLOG_BLCKSZ;
1921 28613384 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1922 :
1923 28613384 : fullpages = bytesleft / UsableBytesInPage;
1924 28613384 : bytesleft = bytesleft % UsableBytesInPage;
1925 :
1926 28613384 : if (bytesleft == 0)
1927 27684 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
1928 : else
1929 28585700 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1930 : }
1931 :
1932 28754984 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1933 :
1934 28754984 : return result;
1935 : }
1936 :
1937 : /*
1938 : * Convert an XLogRecPtr to a "usable byte position".
1939 : */
1940 : static uint64
1941 3750 : XLogRecPtrToBytePos(XLogRecPtr ptr)
1942 : {
1943 : uint64 fullsegs;
1944 : uint32 fullpages;
1945 : uint32 offset;
1946 : uint64 result;
1947 :
1948 3750 : XLByteToSeg(ptr, fullsegs, wal_segment_size);
1949 :
1950 3750 : fullpages = (XLogSegmentOffset(ptr, wal_segment_size)) / XLOG_BLCKSZ;
1951 3750 : offset = ptr % XLOG_BLCKSZ;
1952 :
1953 3750 : if (fullpages == 0)
1954 : {
1955 1074 : result = fullsegs * UsableBytesInSegment;
1956 1074 : if (offset > 0)
1957 : {
1958 : Assert(offset >= SizeOfXLogLongPHD);
1959 410 : result += offset - SizeOfXLogLongPHD;
1960 : }
1961 : }
1962 : else
1963 : {
1964 2676 : result = fullsegs * UsableBytesInSegment +
1965 2676 : (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
1966 2676 : (fullpages - 1) * UsableBytesInPage; /* full pages */
1967 2676 : if (offset > 0)
1968 : {
1969 : Assert(offset >= SizeOfXLogShortPHD);
1970 2654 : result += offset - SizeOfXLogShortPHD;
1971 : }
1972 : }
1973 :
1974 3750 : 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 4813290 : AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
1986 : {
1987 4813290 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1988 : int nextidx;
1989 : XLogRecPtr OldPageRqstPtr;
1990 : XLogwrtRqst WriteRqst;
1991 4813290 : XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
1992 : XLogRecPtr NewPageBeginPtr;
1993 : XLogPageHeader NewPage;
1994 4813290 : int npages pg_attribute_unused() = 0;
1995 :
1996 4813290 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
1997 :
1998 : /*
1999 : * Now that we have the lock, check if someone initialized the page
2000 : * already.
2001 : */
2002 13416492 : while (upto >= XLogCtl->InitializedUpTo || opportunistic)
2003 : {
2004 8610316 : 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 8610316 : OldPageRqstPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]);
2012 8610316 : 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 4325938 : if (opportunistic)
2019 7114 : break;
2020 :
2021 : /* Advance shared memory write request position */
2022 4318824 : SpinLockAcquire(&XLogCtl->info_lck);
2023 4318824 : if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
2024 904468 : XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
2025 4318824 : 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 4318824 : RefreshXLogWriteResult(LogwrtResult);
2032 4318824 : 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 4159388 : LWLockRelease(WALBufMappingLock);
2041 :
2042 4159388 : WaitXLogInsertionsToFinish(OldPageRqstPtr);
2043 :
2044 4159388 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2045 :
2046 4159388 : RefreshXLogWriteResult(LogwrtResult);
2047 4159388 : if (LogwrtResult.Write >= OldPageRqstPtr)
2048 : {
2049 : /* OK, someone wrote it already */
2050 651328 : LWLockRelease(WALWriteLock);
2051 : }
2052 : else
2053 : {
2054 : /* Have to write it ourselves */
2055 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2056 3508060 : WriteRqst.Write = OldPageRqstPtr;
2057 3508060 : WriteRqst.Flush = 0;
2058 3508060 : XLogWrite(WriteRqst, tli, false);
2059 3508060 : LWLockRelease(WALWriteLock);
2060 3508060 : pgWalUsage.wal_buffers_full++;
2061 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_DONE();
2062 : }
2063 : /* Re-acquire WALBufMappingLock and retry */
2064 4159388 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2065 4159388 : 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 4443814 : NewPageBeginPtr = XLogCtl->InitializedUpTo;
2074 4443814 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2075 :
2076 : Assert(XLogRecPtrToBufIdx(NewPageBeginPtr) == nextidx);
2077 :
2078 4443814 : 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 4443814 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], InvalidXLogRecPtr);
2086 4443814 : 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 4443814 : MemSet(NewPage, 0, XLOG_BLCKSZ);
2093 :
2094 : /*
2095 : * Fill the new page's header
2096 : */
2097 4443814 : NewPage->xlp_magic = XLOG_PAGE_MAGIC;
2098 :
2099 : /* NewPage->xlp_info = 0; */ /* done by memset */
2100 4443814 : NewPage->xlp_tli = tli;
2101 4443814 : 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 4443814 : if (Insert->runningBackups == 0)
2119 4189372 : NewPage->xlp_info |= XLP_BKP_REMOVABLE;
2120 :
2121 : /*
2122 : * If first page of an XLOG segment file, make it a long header.
2123 : */
2124 4443814 : if ((XLogSegmentOffset(NewPage->xlp_pageaddr, wal_segment_size)) == 0)
2125 : {
2126 2698 : XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2127 :
2128 2698 : NewLongPage->xlp_sysid = ControlFile->system_identifier;
2129 2698 : NewLongPage->xlp_seg_size = wal_segment_size;
2130 2698 : NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2131 2698 : 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 4443814 : pg_write_barrier();
2140 :
2141 4443814 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], NewPageEndPtr);
2142 4443814 : XLogCtl->InitializedUpTo = NewPageEndPtr;
2143 :
2144 4443814 : npages++;
2145 : }
2146 4813290 : 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 4813290 : }
2156 :
2157 : /*
2158 : * Calculate CheckPointSegments based on max_wal_size_mb and
2159 : * checkpoint_completion_target.
2160 : */
2161 : static void
2162 13806 : 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 13806 : target = (double) ConvertToXSegs(max_wal_size_mb, wal_segment_size) /
2181 13806 : (1.0 + CheckPointCompletionTarget);
2182 :
2183 : /* round down */
2184 13806 : CheckPointSegments = (int) target;
2185 :
2186 13806 : if (CheckPointSegments < 1)
2187 18 : CheckPointSegments = 1;
2188 13806 : }
2189 :
2190 : void
2191 10028 : assign_max_wal_size(int newval, void *extra)
2192 : {
2193 10028 : max_wal_size_mb = newval;
2194 10028 : CalculateCheckpointSegments();
2195 10028 : }
2196 :
2197 : void
2198 2004 : assign_checkpoint_completion_target(double newval, void *extra)
2199 : {
2200 2004 : CheckPointCompletionTarget = newval;
2201 2004 : CalculateCheckpointSegments();
2202 2004 : }
2203 :
2204 : bool
2205 3870 : check_wal_segment_size(int *newval, void **extra, GucSource source)
2206 : {
2207 3870 : 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 3870 : 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 2220 : check_max_slot_wal_keep_size(int *newval, void **extra, GucSource source)
2224 : {
2225 2220 : 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 2220 : 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 2506 : 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 2506 : minSegNo = lastredoptr / wal_segment_size +
2253 2506 : ConvertToXSegs(min_wal_size_mb, wal_segment_size) - 1;
2254 2506 : maxSegNo = lastredoptr / wal_segment_size +
2255 2506 : 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 2506 : distance = (1.0 + CheckPointCompletionTarget) * CheckPointDistanceEstimate;
2266 : /* add 10% for good measure. */
2267 2506 : distance *= 1.10;
2268 :
2269 2506 : recycleSegNo = (XLogSegNo) ceil(((double) lastredoptr + distance) /
2270 : wal_segment_size);
2271 :
2272 2506 : if (recycleSegNo < minSegNo)
2273 1548 : recycleSegNo = minSegNo;
2274 2506 : if (recycleSegNo > maxSegNo)
2275 746 : recycleSegNo = maxSegNo;
2276 :
2277 2506 : 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 8012 : XLogCheckpointNeeded(XLogSegNo new_segno)
2291 : {
2292 : XLogSegNo old_segno;
2293 :
2294 8012 : XLByteToSeg(RedoRecPtr, old_segno, wal_segment_size);
2295 :
2296 8012 : if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2297 5340 : return true;
2298 2672 : 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 3751398 : 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 3751398 : 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 3751398 : npages = 0;
2342 3751398 : startidx = 0;
2343 3751398 : 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 3751398 : curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2351 :
2352 8099120 : 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 4584812 : XLogRecPtr EndPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[curridx]);
2360 :
2361 4584812 : 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 4584812 : LogwrtResult.Write = EndPtr;
2368 4584812 : ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2369 :
2370 4584812 : 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 22398 : if (openLogFile >= 0)
2379 9210 : XLogFileClose();
2380 22398 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2381 : wal_segment_size);
2382 22398 : openLogTLI = tli;
2383 :
2384 : /* create/use new log file */
2385 22398 : openLogFile = XLogFileInit(openLogSegNo, tli);
2386 22398 : ReserveExternalFD();
2387 : }
2388 :
2389 : /* Make sure we have the current logfile open */
2390 4584812 : 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 4584812 : if (npages == 0)
2401 : {
2402 : /* first of group */
2403 3781388 : startidx = curridx;
2404 3781388 : startoffset = XLogSegmentOffset(LogwrtResult.Write - XLOG_BLCKSZ,
2405 : wal_segment_size);
2406 : }
2407 4584812 : 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 4584812 : last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2416 :
2417 8938422 : finishing_seg = !ispartialpage &&
2418 4353610 : (startoffset + npages * XLOG_BLCKSZ) >= wal_segment_size;
2419 :
2420 4584812 : if (last_iteration ||
2421 834410 : 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 3781388 : from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2432 3781388 : nbytes = npages * (Size) XLOG_BLCKSZ;
2433 3781388 : nleft = nbytes;
2434 : do
2435 : {
2436 3781388 : errno = 0;
2437 :
2438 : /*
2439 : * Measure I/O timing to write WAL data, for pg_stat_io and/or
2440 : * pg_stat_wal.
2441 : */
2442 3781388 : start = pgstat_prepare_io_time(track_io_timing || track_wal_io_timing);
2443 :
2444 3781388 : pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE);
2445 3781388 : written = pg_pwrite(openLogFile, from, nleft, startoffset);
2446 3781388 : pgstat_report_wait_end();
2447 :
2448 3781388 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL,
2449 : IOOP_WRITE, start, 1, written);
2450 :
2451 : /*
2452 : * Increment the I/O timing and the number of times WAL data
2453 : * were written out to disk.
2454 : */
2455 3781388 : if (track_wal_io_timing)
2456 : {
2457 : instr_time end;
2458 :
2459 0 : INSTR_TIME_SET_CURRENT(end);
2460 0 : INSTR_TIME_ACCUM_DIFF(PendingWalStats.wal_write_time, end, start);
2461 : }
2462 :
2463 3781388 : PendingWalStats.wal_write++;
2464 :
2465 3781388 : if (written <= 0)
2466 : {
2467 : char xlogfname[MAXFNAMELEN];
2468 : int save_errno;
2469 :
2470 0 : if (errno == EINTR)
2471 0 : continue;
2472 :
2473 0 : save_errno = errno;
2474 0 : XLogFileName(xlogfname, tli, openLogSegNo,
2475 : wal_segment_size);
2476 0 : errno = save_errno;
2477 0 : ereport(PANIC,
2478 : (errcode_for_file_access(),
2479 : errmsg("could not write to log file \"%s\" at offset %u, length %zu: %m",
2480 : xlogfname, startoffset, nleft)));
2481 : }
2482 3781388 : nleft -= written;
2483 3781388 : from += written;
2484 3781388 : startoffset += written;
2485 3781388 : } while (nleft > 0);
2486 :
2487 3781388 : npages = 0;
2488 :
2489 : /*
2490 : * If we just wrote the whole last page of a logfile segment,
2491 : * fsync the segment immediately. This avoids having to go back
2492 : * and re-open prior segments when an fsync request comes along
2493 : * later. Doing it here ensures that one and only one backend will
2494 : * perform this fsync.
2495 : *
2496 : * This is also the right place to notify the Archiver that the
2497 : * segment is ready to copy to archival storage, and to update the
2498 : * timer for archive_timeout, and to signal for a checkpoint if
2499 : * too many logfile segments have been used since the last
2500 : * checkpoint.
2501 : */
2502 3781388 : if (finishing_seg)
2503 : {
2504 2906 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2505 :
2506 : /* signal that we need to wakeup walsenders later */
2507 2906 : WalSndWakeupRequest();
2508 :
2509 2906 : LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2510 :
2511 2906 : if (XLogArchivingActive())
2512 160 : XLogArchiveNotifySeg(openLogSegNo, tli);
2513 :
2514 2906 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2515 2906 : XLogCtl->lastSegSwitchLSN = LogwrtResult.Flush;
2516 :
2517 : /*
2518 : * Request a checkpoint if we've consumed too much xlog since
2519 : * the last one. For speed, we first check using the local
2520 : * copy of RedoRecPtr, which might be out of date; if it looks
2521 : * like a checkpoint is needed, forcibly update RedoRecPtr and
2522 : * recheck.
2523 : */
2524 2906 : if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2525 : {
2526 458 : (void) GetRedoRecPtr();
2527 458 : if (XLogCheckpointNeeded(openLogSegNo))
2528 384 : RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2529 : }
2530 : }
2531 : }
2532 :
2533 4584812 : if (ispartialpage)
2534 : {
2535 : /* Only asked to write a partial page */
2536 231202 : LogwrtResult.Write = WriteRqst.Write;
2537 231202 : break;
2538 : }
2539 4353610 : curridx = NextBufIdx(curridx);
2540 :
2541 : /* If flexible, break out of loop as soon as we wrote something */
2542 4353610 : if (flexible && npages == 0)
2543 5888 : break;
2544 : }
2545 :
2546 : Assert(npages == 0);
2547 :
2548 : /*
2549 : * If asked to flush, do so
2550 : */
2551 3751398 : if (LogwrtResult.Flush < WriteRqst.Flush &&
2552 242672 : LogwrtResult.Flush < LogwrtResult.Write)
2553 : {
2554 : /*
2555 : * Could get here without iterating above loop, in which case we might
2556 : * have no open file or the wrong one. However, we do not need to
2557 : * fsync more than one file.
2558 : */
2559 242568 : if (wal_sync_method != WAL_SYNC_METHOD_OPEN &&
2560 242568 : wal_sync_method != WAL_SYNC_METHOD_OPEN_DSYNC)
2561 : {
2562 242568 : if (openLogFile >= 0 &&
2563 242554 : !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2564 : wal_segment_size))
2565 34 : XLogFileClose();
2566 242568 : if (openLogFile < 0)
2567 : {
2568 48 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2569 : wal_segment_size);
2570 48 : openLogTLI = tli;
2571 48 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2572 48 : ReserveExternalFD();
2573 : }
2574 :
2575 242568 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2576 : }
2577 :
2578 : /* signal that we need to wakeup walsenders later */
2579 242568 : WalSndWakeupRequest();
2580 :
2581 242568 : LogwrtResult.Flush = LogwrtResult.Write;
2582 : }
2583 :
2584 : /*
2585 : * Update shared-memory status
2586 : *
2587 : * We make sure that the shared 'request' values do not fall behind the
2588 : * 'result' values. This is not absolutely essential, but it saves some
2589 : * code in a couple of places.
2590 : */
2591 3751398 : SpinLockAcquire(&XLogCtl->info_lck);
2592 3751398 : if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
2593 214622 : XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
2594 3751398 : if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
2595 244742 : XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
2596 3751398 : SpinLockRelease(&XLogCtl->info_lck);
2597 :
2598 : /*
2599 : * We write Write first, bar, then Flush. When reading, the opposite must
2600 : * be done (with a matching barrier in between), so that we always see a
2601 : * Flush value that trails behind the Write value seen.
2602 : */
2603 3751398 : pg_atomic_write_u64(&XLogCtl->logWriteResult, LogwrtResult.Write);
2604 3751398 : pg_write_barrier();
2605 3751398 : pg_atomic_write_u64(&XLogCtl->logFlushResult, LogwrtResult.Flush);
2606 :
2607 : #ifdef USE_ASSERT_CHECKING
2608 : {
2609 : XLogRecPtr Flush;
2610 : XLogRecPtr Write;
2611 : XLogRecPtr Insert;
2612 :
2613 : Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult);
2614 : pg_read_barrier();
2615 : Write = pg_atomic_read_u64(&XLogCtl->logWriteResult);
2616 : pg_read_barrier();
2617 : Insert = pg_atomic_read_u64(&XLogCtl->logInsertResult);
2618 :
2619 : /* WAL written to disk is always ahead of WAL flushed */
2620 : Assert(Write >= Flush);
2621 :
2622 : /* WAL inserted to buffers is always ahead of WAL written */
2623 : Assert(Insert >= Write);
2624 : }
2625 : #endif
2626 3751398 : }
2627 :
2628 : /*
2629 : * Record the LSN for an asynchronous transaction commit/abort
2630 : * and nudge the WALWriter if there is work for it to do.
2631 : * (This should not be called for synchronous commits.)
2632 : */
2633 : void
2634 88712 : XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2635 : {
2636 88712 : XLogRecPtr WriteRqstPtr = asyncXactLSN;
2637 : bool sleeping;
2638 88712 : bool wakeup = false;
2639 : XLogRecPtr prevAsyncXactLSN;
2640 :
2641 88712 : SpinLockAcquire(&XLogCtl->info_lck);
2642 88712 : sleeping = XLogCtl->WalWriterSleeping;
2643 88712 : prevAsyncXactLSN = XLogCtl->asyncXactLSN;
2644 88712 : if (XLogCtl->asyncXactLSN < asyncXactLSN)
2645 87886 : XLogCtl->asyncXactLSN = asyncXactLSN;
2646 88712 : SpinLockRelease(&XLogCtl->info_lck);
2647 :
2648 : /*
2649 : * If somebody else already called this function with a more aggressive
2650 : * LSN, they will have done what we needed (and perhaps more).
2651 : */
2652 88712 : if (asyncXactLSN <= prevAsyncXactLSN)
2653 826 : return;
2654 :
2655 : /*
2656 : * If the WALWriter is sleeping, kick it to make it come out of low-power
2657 : * mode, so that this async commit will reach disk within the expected
2658 : * amount of time. Otherwise, determine whether it has enough WAL
2659 : * available to flush, the same way that XLogBackgroundFlush() does.
2660 : */
2661 87886 : if (sleeping)
2662 34 : wakeup = true;
2663 : else
2664 : {
2665 : int flushblocks;
2666 :
2667 87852 : RefreshXLogWriteResult(LogwrtResult);
2668 :
2669 87852 : flushblocks =
2670 87852 : WriteRqstPtr / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
2671 :
2672 87852 : if (WalWriterFlushAfter == 0 || flushblocks >= WalWriterFlushAfter)
2673 6860 : wakeup = true;
2674 : }
2675 :
2676 87886 : if (wakeup)
2677 : {
2678 6894 : volatile PROC_HDR *procglobal = ProcGlobal;
2679 6894 : ProcNumber walwriterProc = procglobal->walwriterProc;
2680 :
2681 6894 : if (walwriterProc != INVALID_PROC_NUMBER)
2682 434 : SetLatch(&GetPGProcByNumber(walwriterProc)->procLatch);
2683 : }
2684 : }
2685 :
2686 : /*
2687 : * Record the LSN up to which we can remove WAL because it's not required by
2688 : * any replication slot.
2689 : */
2690 : void
2691 21002 : XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn)
2692 : {
2693 21002 : SpinLockAcquire(&XLogCtl->info_lck);
2694 21002 : XLogCtl->replicationSlotMinLSN = lsn;
2695 21002 : SpinLockRelease(&XLogCtl->info_lck);
2696 21002 : }
2697 :
2698 :
2699 : /*
2700 : * Return the oldest LSN we must retain to satisfy the needs of some
2701 : * replication slot.
2702 : */
2703 : static XLogRecPtr
2704 3336 : XLogGetReplicationSlotMinimumLSN(void)
2705 : {
2706 : XLogRecPtr retval;
2707 :
2708 3336 : SpinLockAcquire(&XLogCtl->info_lck);
2709 3336 : retval = XLogCtl->replicationSlotMinLSN;
2710 3336 : SpinLockRelease(&XLogCtl->info_lck);
2711 :
2712 3336 : return retval;
2713 : }
2714 :
2715 : /*
2716 : * Advance minRecoveryPoint in control file.
2717 : *
2718 : * If we crash during recovery, we must reach this point again before the
2719 : * database is consistent.
2720 : *
2721 : * If 'force' is true, 'lsn' argument is ignored. Otherwise, minRecoveryPoint
2722 : * is only updated if it's not already greater than or equal to 'lsn'.
2723 : */
2724 : static void
2725 188172 : UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2726 : {
2727 : /* Quick check using our local copy of the variable */
2728 188172 : if (!updateMinRecoveryPoint || (!force && lsn <= LocalMinRecoveryPoint))
2729 176396 : return;
2730 :
2731 : /*
2732 : * An invalid minRecoveryPoint means that we need to recover all the WAL,
2733 : * i.e., we're doing crash recovery. We never modify the control file's
2734 : * value in that case, so we can short-circuit future checks here too. The
2735 : * local values of minRecoveryPoint and minRecoveryPointTLI should not be
2736 : * updated until crash recovery finishes. We only do this for the startup
2737 : * process as it should not update its own reference of minRecoveryPoint
2738 : * until it has finished crash recovery to make sure that all WAL
2739 : * available is replayed in this case. This also saves from extra locks
2740 : * taken on the control file from the startup process.
2741 : */
2742 11776 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint) && InRecovery)
2743 : {
2744 62 : updateMinRecoveryPoint = false;
2745 62 : return;
2746 : }
2747 :
2748 11714 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2749 :
2750 : /* update local copy */
2751 11714 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
2752 11714 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2753 :
2754 11714 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint))
2755 4 : updateMinRecoveryPoint = false;
2756 11710 : else if (force || LocalMinRecoveryPoint < lsn)
2757 : {
2758 : XLogRecPtr newMinRecoveryPoint;
2759 : TimeLineID newMinRecoveryPointTLI;
2760 :
2761 : /*
2762 : * To avoid having to update the control file too often, we update it
2763 : * all the way to the last record being replayed, even though 'lsn'
2764 : * would suffice for correctness. This also allows the 'force' case
2765 : * to not need a valid 'lsn' value.
2766 : *
2767 : * Another important reason for doing it this way is that the passed
2768 : * 'lsn' value could be bogus, i.e., past the end of available WAL, if
2769 : * the caller got it from a corrupted heap page. Accepting such a
2770 : * value as the min recovery point would prevent us from coming up at
2771 : * all. Instead, we just log a warning and continue with recovery.
2772 : * (See also the comments about corrupt LSNs in XLogFlush.)
2773 : */
2774 9578 : newMinRecoveryPoint = GetCurrentReplayRecPtr(&newMinRecoveryPointTLI);
2775 9578 : if (!force && newMinRecoveryPoint < lsn)
2776 0 : elog(WARNING,
2777 : "xlog min recovery request %X/%X is past current point %X/%X",
2778 : LSN_FORMAT_ARGS(lsn), LSN_FORMAT_ARGS(newMinRecoveryPoint));
2779 :
2780 : /* update control file */
2781 9578 : if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2782 : {
2783 9522 : ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2784 9522 : ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2785 9522 : UpdateControlFile();
2786 9522 : LocalMinRecoveryPoint = newMinRecoveryPoint;
2787 9522 : LocalMinRecoveryPointTLI = newMinRecoveryPointTLI;
2788 :
2789 9522 : ereport(DEBUG2,
2790 : (errmsg_internal("updated min recovery point to %X/%X on timeline %u",
2791 : LSN_FORMAT_ARGS(newMinRecoveryPoint),
2792 : newMinRecoveryPointTLI)));
2793 : }
2794 : }
2795 11714 : LWLockRelease(ControlFileLock);
2796 : }
2797 :
2798 : /*
2799 : * Ensure that all XLOG data through the given position is flushed to disk.
2800 : *
2801 : * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
2802 : * already held, and we try to avoid acquiring it if possible.
2803 : */
2804 : void
2805 1216402 : XLogFlush(XLogRecPtr record)
2806 : {
2807 : XLogRecPtr WriteRqstPtr;
2808 : XLogwrtRqst WriteRqst;
2809 1216402 : TimeLineID insertTLI = XLogCtl->InsertTimeLineID;
2810 :
2811 : /*
2812 : * During REDO, we are reading not writing WAL. Therefore, instead of
2813 : * trying to flush the WAL, we should update minRecoveryPoint instead. We
2814 : * test XLogInsertAllowed(), not InRecovery, because we need checkpointer
2815 : * to act this way too, and because when it tries to write the
2816 : * end-of-recovery checkpoint, it should indeed flush.
2817 : */
2818 1216402 : if (!XLogInsertAllowed())
2819 : {
2820 187934 : UpdateMinRecoveryPoint(record, false);
2821 959372 : return;
2822 : }
2823 :
2824 : /* Quick exit if already known flushed */
2825 1028468 : if (record <= LogwrtResult.Flush)
2826 771438 : return;
2827 :
2828 : #ifdef WAL_DEBUG
2829 : if (XLOG_DEBUG)
2830 : elog(LOG, "xlog flush request %X/%X; write %X/%X; flush %X/%X",
2831 : LSN_FORMAT_ARGS(record),
2832 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2833 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2834 : #endif
2835 :
2836 257030 : START_CRIT_SECTION();
2837 :
2838 : /*
2839 : * Since fsync is usually a horribly expensive operation, we try to
2840 : * piggyback as much data as we can on each fsync: if we see any more data
2841 : * entered into the xlog buffer, we'll write and fsync that too, so that
2842 : * the final value of LogwrtResult.Flush is as large as possible. This
2843 : * gives us some chance of avoiding another fsync immediately after.
2844 : */
2845 :
2846 : /* initialize to given target; may increase below */
2847 257030 : WriteRqstPtr = record;
2848 :
2849 : /*
2850 : * Now wait until we get the write lock, or someone else does the flush
2851 : * for us.
2852 : */
2853 : for (;;)
2854 11832 : {
2855 : XLogRecPtr insertpos;
2856 :
2857 : /* done already? */
2858 268862 : RefreshXLogWriteResult(LogwrtResult);
2859 268862 : if (record <= LogwrtResult.Flush)
2860 18880 : break;
2861 :
2862 : /*
2863 : * Before actually performing the write, wait for all in-flight
2864 : * insertions to the pages we're about to write to finish.
2865 : */
2866 249982 : SpinLockAcquire(&XLogCtl->info_lck);
2867 249982 : if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
2868 15168 : WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
2869 249982 : SpinLockRelease(&XLogCtl->info_lck);
2870 249982 : insertpos = WaitXLogInsertionsToFinish(WriteRqstPtr);
2871 :
2872 : /*
2873 : * Try to get the write lock. If we can't get it immediately, wait
2874 : * until it's released, and recheck if we still need to do the flush
2875 : * or if the backend that held the lock did it for us already. This
2876 : * helps to maintain a good rate of group committing when the system
2877 : * is bottlenecked by the speed of fsyncing.
2878 : */
2879 249982 : if (!LWLockAcquireOrWait(WALWriteLock, LW_EXCLUSIVE))
2880 : {
2881 : /*
2882 : * The lock is now free, but we didn't acquire it yet. Before we
2883 : * do, loop back to check if someone else flushed the record for
2884 : * us already.
2885 : */
2886 11832 : continue;
2887 : }
2888 :
2889 : /* Got the lock; recheck whether request is satisfied */
2890 238150 : RefreshXLogWriteResult(LogwrtResult);
2891 238150 : if (record <= LogwrtResult.Flush)
2892 : {
2893 1676 : LWLockRelease(WALWriteLock);
2894 1676 : break;
2895 : }
2896 :
2897 : /*
2898 : * Sleep before flush! By adding a delay here, we may give further
2899 : * backends the opportunity to join the backlog of group commit
2900 : * followers; this can significantly improve transaction throughput,
2901 : * at the risk of increasing transaction latency.
2902 : *
2903 : * We do not sleep if enableFsync is not turned on, nor if there are
2904 : * fewer than CommitSiblings other backends with active transactions.
2905 : */
2906 236474 : if (CommitDelay > 0 && enableFsync &&
2907 0 : MinimumActiveBackends(CommitSiblings))
2908 : {
2909 0 : pg_usleep(CommitDelay);
2910 :
2911 : /*
2912 : * Re-check how far we can now flush the WAL. It's generally not
2913 : * safe to call WaitXLogInsertionsToFinish while holding
2914 : * WALWriteLock, because an in-progress insertion might need to
2915 : * also grab WALWriteLock to make progress. But we know that all
2916 : * the insertions up to insertpos have already finished, because
2917 : * that's what the earlier WaitXLogInsertionsToFinish() returned.
2918 : * We're only calling it again to allow insertpos to be moved
2919 : * further forward, not to actually wait for anyone.
2920 : */
2921 0 : insertpos = WaitXLogInsertionsToFinish(insertpos);
2922 : }
2923 :
2924 : /* try to write/flush later additions to XLOG as well */
2925 236474 : WriteRqst.Write = insertpos;
2926 236474 : WriteRqst.Flush = insertpos;
2927 :
2928 236474 : XLogWrite(WriteRqst, insertTLI, false);
2929 :
2930 236474 : LWLockRelease(WALWriteLock);
2931 : /* done */
2932 236474 : break;
2933 : }
2934 :
2935 257030 : END_CRIT_SECTION();
2936 :
2937 : /* wake up walsenders now that we've released heavily contended locks */
2938 257030 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
2939 :
2940 : /*
2941 : * If we still haven't flushed to the request point then we have a
2942 : * problem; most likely, the requested flush point is past end of XLOG.
2943 : * This has been seen to occur when a disk page has a corrupted LSN.
2944 : *
2945 : * Formerly we treated this as a PANIC condition, but that hurts the
2946 : * system's robustness rather than helping it: we do not want to take down
2947 : * the whole system due to corruption on one data page. In particular, if
2948 : * the bad page is encountered again during recovery then we would be
2949 : * unable to restart the database at all! (This scenario actually
2950 : * happened in the field several times with 7.1 releases.) As of 8.4, bad
2951 : * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
2952 : * the only time we can reach here during recovery is while flushing the
2953 : * end-of-recovery checkpoint record, and we don't expect that to have a
2954 : * bad LSN.
2955 : *
2956 : * Note that for calls from xact.c, the ERROR will be promoted to PANIC
2957 : * since xact.c calls this routine inside a critical section. However,
2958 : * calls from bufmgr.c are not within critical sections and so we will not
2959 : * force a restart for a bad LSN on a data page.
2960 : */
2961 257030 : if (LogwrtResult.Flush < record)
2962 0 : elog(ERROR,
2963 : "xlog flush request %X/%X is not satisfied --- flushed only to %X/%X",
2964 : LSN_FORMAT_ARGS(record),
2965 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2966 : }
2967 :
2968 : /*
2969 : * Write & flush xlog, but without specifying exactly where to.
2970 : *
2971 : * We normally write only completed blocks; but if there is nothing to do on
2972 : * that basis, we check for unwritten async commits in the current incomplete
2973 : * block, and write through the latest one of those. Thus, if async commits
2974 : * are not being used, we will write complete blocks only.
2975 : *
2976 : * If, based on the above, there's anything to write we do so immediately. But
2977 : * to avoid calling fsync, fdatasync et. al. at a rate that'd impact
2978 : * concurrent IO, we only flush WAL every wal_writer_delay ms, or if there's
2979 : * more than wal_writer_flush_after unflushed blocks.
2980 : *
2981 : * We can guarantee that async commits reach disk after at most three
2982 : * wal_writer_delay cycles. (When flushing complete blocks, we allow XLogWrite
2983 : * to write "flexibly", meaning it can stop at the end of the buffer ring;
2984 : * this makes a difference only with very high load or long wal_writer_delay,
2985 : * but imposes one extra cycle for the worst case for async commits.)
2986 : *
2987 : * This routine is invoked periodically by the background walwriter process.
2988 : *
2989 : * Returns true if there was any work to do, even if we skipped flushing due
2990 : * to wal_writer_delay/wal_writer_flush_after.
2991 : */
2992 : bool
2993 18996 : XLogBackgroundFlush(void)
2994 : {
2995 : XLogwrtRqst WriteRqst;
2996 18996 : bool flexible = true;
2997 : static TimestampTz lastflush;
2998 : TimestampTz now;
2999 : int flushblocks;
3000 : TimeLineID insertTLI;
3001 :
3002 : /* XLOG doesn't need flushing during recovery */
3003 18996 : if (RecoveryInProgress())
3004 16 : return false;
3005 :
3006 : /*
3007 : * Since we're not in recovery, InsertTimeLineID is set and can't change,
3008 : * so we can read it without a lock.
3009 : */
3010 18980 : insertTLI = XLogCtl->InsertTimeLineID;
3011 :
3012 : /* read updated LogwrtRqst */
3013 18980 : SpinLockAcquire(&XLogCtl->info_lck);
3014 18980 : WriteRqst = XLogCtl->LogwrtRqst;
3015 18980 : SpinLockRelease(&XLogCtl->info_lck);
3016 :
3017 : /* back off to last completed page boundary */
3018 18980 : WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
3019 :
3020 : /* if we have already flushed that far, consider async commit records */
3021 18980 : RefreshXLogWriteResult(LogwrtResult);
3022 18980 : if (WriteRqst.Write <= LogwrtResult.Flush)
3023 : {
3024 12850 : SpinLockAcquire(&XLogCtl->info_lck);
3025 12850 : WriteRqst.Write = XLogCtl->asyncXactLSN;
3026 12850 : SpinLockRelease(&XLogCtl->info_lck);
3027 12850 : flexible = false; /* ensure it all gets written */
3028 : }
3029 :
3030 : /*
3031 : * If already known flushed, we're done. Just need to check if we are
3032 : * holding an open file handle to a logfile that's no longer in use,
3033 : * preventing the file from being deleted.
3034 : */
3035 18980 : if (WriteRqst.Write <= LogwrtResult.Flush)
3036 : {
3037 11866 : if (openLogFile >= 0)
3038 : {
3039 6932 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
3040 : wal_segment_size))
3041 : {
3042 272 : XLogFileClose();
3043 : }
3044 : }
3045 11866 : return false;
3046 : }
3047 :
3048 : /*
3049 : * Determine how far to flush WAL, based on the wal_writer_delay and
3050 : * wal_writer_flush_after GUCs.
3051 : *
3052 : * Note that XLogSetAsyncXactLSN() performs similar calculation based on
3053 : * wal_writer_flush_after, to decide when to wake us up. Make sure the
3054 : * logic is the same in both places if you change this.
3055 : */
3056 7114 : now = GetCurrentTimestamp();
3057 7114 : flushblocks =
3058 7114 : WriteRqst.Write / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
3059 :
3060 7114 : if (WalWriterFlushAfter == 0 || lastflush == 0)
3061 : {
3062 : /* first call, or block based limits disabled */
3063 386 : WriteRqst.Flush = WriteRqst.Write;
3064 386 : lastflush = now;
3065 : }
3066 6728 : else if (TimestampDifferenceExceeds(lastflush, now, WalWriterDelay))
3067 : {
3068 : /*
3069 : * Flush the writes at least every WalWriterDelay ms. This is
3070 : * important to bound the amount of time it takes for an asynchronous
3071 : * commit to hit disk.
3072 : */
3073 6384 : WriteRqst.Flush = WriteRqst.Write;
3074 6384 : lastflush = now;
3075 : }
3076 344 : else if (flushblocks >= WalWriterFlushAfter)
3077 : {
3078 : /* exceeded wal_writer_flush_after blocks, flush */
3079 312 : WriteRqst.Flush = WriteRqst.Write;
3080 312 : lastflush = now;
3081 : }
3082 : else
3083 : {
3084 : /* no flushing, this time round */
3085 32 : WriteRqst.Flush = 0;
3086 : }
3087 :
3088 : #ifdef WAL_DEBUG
3089 : if (XLOG_DEBUG)
3090 : elog(LOG, "xlog bg flush request write %X/%X; flush: %X/%X, current is write %X/%X; flush %X/%X",
3091 : LSN_FORMAT_ARGS(WriteRqst.Write),
3092 : LSN_FORMAT_ARGS(WriteRqst.Flush),
3093 : LSN_FORMAT_ARGS(LogwrtResult.Write),
3094 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
3095 : #endif
3096 :
3097 7114 : START_CRIT_SECTION();
3098 :
3099 : /* now wait for any in-progress insertions to finish and get write lock */
3100 7114 : WaitXLogInsertionsToFinish(WriteRqst.Write);
3101 7114 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3102 7114 : RefreshXLogWriteResult(LogwrtResult);
3103 7114 : if (WriteRqst.Write > LogwrtResult.Write ||
3104 288 : WriteRqst.Flush > LogwrtResult.Flush)
3105 : {
3106 6864 : XLogWrite(WriteRqst, insertTLI, flexible);
3107 : }
3108 7114 : LWLockRelease(WALWriteLock);
3109 :
3110 7114 : END_CRIT_SECTION();
3111 :
3112 : /* wake up walsenders now that we've released heavily contended locks */
3113 7114 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3114 :
3115 : /*
3116 : * Great, done. To take some work off the critical path, try to initialize
3117 : * as many of the no-longer-needed WAL buffers for future use as we can.
3118 : */
3119 7114 : AdvanceXLInsertBuffer(InvalidXLogRecPtr, insertTLI, true);
3120 :
3121 : /*
3122 : * If we determined that we need to write data, but somebody else
3123 : * wrote/flushed already, it should be considered as being active, to
3124 : * avoid hibernating too early.
3125 : */
3126 7114 : return true;
3127 : }
3128 :
3129 : /*
3130 : * Test whether XLOG data has been flushed up to (at least) the given position.
3131 : *
3132 : * Returns true if a flush is still needed. (It may be that someone else
3133 : * is already in process of flushing that far, however.)
3134 : */
3135 : bool
3136 17157466 : XLogNeedsFlush(XLogRecPtr record)
3137 : {
3138 : /*
3139 : * During recovery, we don't flush WAL but update minRecoveryPoint
3140 : * instead. So "needs flush" is taken to mean whether minRecoveryPoint
3141 : * would need to be updated.
3142 : */
3143 17157466 : if (RecoveryInProgress())
3144 : {
3145 : /*
3146 : * An invalid minRecoveryPoint means that we need to recover all the
3147 : * WAL, i.e., we're doing crash recovery. We never modify the control
3148 : * file's value in that case, so we can short-circuit future checks
3149 : * here too. This triggers a quick exit path for the startup process,
3150 : * which cannot update its local copy of minRecoveryPoint as long as
3151 : * it has not replayed all WAL available when doing crash recovery.
3152 : */
3153 1610412 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint) && InRecovery)
3154 0 : updateMinRecoveryPoint = false;
3155 :
3156 : /* Quick exit if already known to be updated or cannot be updated */
3157 1610412 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3158 1591516 : return false;
3159 :
3160 : /*
3161 : * Update local copy of minRecoveryPoint. But if the lock is busy,
3162 : * just return a conservative guess.
3163 : */
3164 18896 : if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3165 0 : return true;
3166 18896 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
3167 18896 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3168 18896 : LWLockRelease(ControlFileLock);
3169 :
3170 : /*
3171 : * Check minRecoveryPoint for any other process than the startup
3172 : * process doing crash recovery, which should not update the control
3173 : * file value if crash recovery is still running.
3174 : */
3175 18896 : if (XLogRecPtrIsInvalid(LocalMinRecoveryPoint))
3176 0 : updateMinRecoveryPoint = false;
3177 :
3178 : /* check again */
3179 18896 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3180 146 : return false;
3181 : else
3182 18750 : return true;
3183 : }
3184 :
3185 : /* Quick exit if already known flushed */
3186 15547054 : if (record <= LogwrtResult.Flush)
3187 15073668 : return false;
3188 :
3189 : /* read LogwrtResult and update local state */
3190 473386 : RefreshXLogWriteResult(LogwrtResult);
3191 :
3192 : /* check again */
3193 473386 : if (record <= LogwrtResult.Flush)
3194 4144 : return false;
3195 :
3196 469242 : return true;
3197 : }
3198 :
3199 : /*
3200 : * Try to make a given XLOG file segment exist.
3201 : *
3202 : * logsegno: identify segment.
3203 : *
3204 : * *added: on return, true if this call raised the number of extant segments.
3205 : *
3206 : * path: on return, this char[MAXPGPATH] has the path to the logsegno file.
3207 : *
3208 : * Returns -1 or FD of opened file. A -1 here is not an error; a caller
3209 : * wanting an open segment should attempt to open "path", which usually will
3210 : * succeed. (This is weird, but it's efficient for the callers.)
3211 : */
3212 : static int
3213 24574 : XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,
3214 : bool *added, char *path)
3215 : {
3216 : char tmppath[MAXPGPATH];
3217 : XLogSegNo installed_segno;
3218 : XLogSegNo max_segno;
3219 : int fd;
3220 : int save_errno;
3221 24574 : int open_flags = O_RDWR | O_CREAT | O_EXCL | PG_BINARY;
3222 : instr_time io_start;
3223 :
3224 : Assert(logtli != 0);
3225 :
3226 24574 : XLogFilePath(path, logtli, logsegno, wal_segment_size);
3227 :
3228 : /*
3229 : * Try to use existent file (checkpoint maker may have created it already)
3230 : */
3231 24574 : *added = false;
3232 24574 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3233 24574 : get_sync_bit(wal_sync_method));
3234 24574 : if (fd < 0)
3235 : {
3236 2548 : if (errno != ENOENT)
3237 0 : ereport(ERROR,
3238 : (errcode_for_file_access(),
3239 : errmsg("could not open file \"%s\": %m", path)));
3240 : }
3241 : else
3242 22026 : return fd;
3243 :
3244 : /*
3245 : * Initialize an empty (all zeroes) segment. NOTE: it is possible that
3246 : * another process is doing the same thing. If so, we will end up
3247 : * pre-creating an extra log segment. That seems OK, and better than
3248 : * holding the lock throughout this lengthy process.
3249 : */
3250 2548 : elog(DEBUG2, "creating and filling new WAL file");
3251 :
3252 2548 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3253 :
3254 2548 : unlink(tmppath);
3255 :
3256 2548 : if (io_direct_flags & IO_DIRECT_WAL_INIT)
3257 0 : open_flags |= PG_O_DIRECT;
3258 :
3259 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3260 2548 : fd = BasicOpenFile(tmppath, open_flags);
3261 2548 : if (fd < 0)
3262 0 : ereport(ERROR,
3263 : (errcode_for_file_access(),
3264 : errmsg("could not create file \"%s\": %m", tmppath)));
3265 :
3266 : /* Measure I/O timing when initializing segment */
3267 2548 : io_start = pgstat_prepare_io_time(track_io_timing);
3268 :
3269 2548 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE);
3270 2548 : save_errno = 0;
3271 2548 : if (wal_init_zero)
3272 : {
3273 : ssize_t rc;
3274 :
3275 : /*
3276 : * Zero-fill the file. With this setting, we do this the hard way to
3277 : * ensure that all the file space has really been allocated. On
3278 : * platforms that allow "holes" in files, just seeking to the end
3279 : * doesn't allocate intermediate space. This way, we know that we
3280 : * have all the space and (after the fsync below) that all the
3281 : * indirect blocks are down on disk. Therefore, fdatasync(2) or
3282 : * O_DSYNC will be sufficient to sync future writes to the log file.
3283 : */
3284 2548 : rc = pg_pwrite_zeros(fd, wal_segment_size, 0);
3285 :
3286 2548 : if (rc < 0)
3287 0 : save_errno = errno;
3288 : }
3289 : else
3290 : {
3291 : /*
3292 : * Otherwise, seeking to the end and writing a solitary byte is
3293 : * enough.
3294 : */
3295 0 : errno = 0;
3296 0 : if (pg_pwrite(fd, "\0", 1, wal_segment_size - 1) != 1)
3297 : {
3298 : /* if write didn't set errno, assume no disk space */
3299 0 : save_errno = errno ? errno : ENOSPC;
3300 : }
3301 : }
3302 2548 : pgstat_report_wait_end();
3303 :
3304 : /*
3305 : * A full segment worth of data is written when using wal_init_zero. One
3306 : * byte is written when not using it.
3307 : */
3308 2548 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT, IOOP_WRITE,
3309 : io_start, 1,
3310 2548 : wal_init_zero ? wal_segment_size : 1);
3311 :
3312 2548 : if (save_errno)
3313 : {
3314 : /*
3315 : * If we fail to make the file, delete it to release disk space
3316 : */
3317 0 : unlink(tmppath);
3318 :
3319 0 : close(fd);
3320 :
3321 0 : errno = save_errno;
3322 :
3323 0 : ereport(ERROR,
3324 : (errcode_for_file_access(),
3325 : errmsg("could not write to file \"%s\": %m", tmppath)));
3326 : }
3327 :
3328 : /* Measure I/O timing when flushing segment */
3329 2548 : io_start = pgstat_prepare_io_time(track_io_timing);
3330 :
3331 2548 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC);
3332 2548 : if (pg_fsync(fd) != 0)
3333 : {
3334 0 : save_errno = errno;
3335 0 : close(fd);
3336 0 : errno = save_errno;
3337 0 : ereport(ERROR,
3338 : (errcode_for_file_access(),
3339 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3340 : }
3341 2548 : pgstat_report_wait_end();
3342 :
3343 2548 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT,
3344 : IOOP_FSYNC, io_start, 1, 0);
3345 :
3346 2548 : if (close(fd) != 0)
3347 0 : ereport(ERROR,
3348 : (errcode_for_file_access(),
3349 : errmsg("could not close file \"%s\": %m", tmppath)));
3350 :
3351 : /*
3352 : * Now move the segment into place with its final name. Cope with
3353 : * possibility that someone else has created the file while we were
3354 : * filling ours: if so, use ours to pre-create a future log segment.
3355 : */
3356 2548 : installed_segno = logsegno;
3357 :
3358 : /*
3359 : * XXX: What should we use as max_segno? We used to use XLOGfileslop when
3360 : * that was a constant, but that was always a bit dubious: normally, at a
3361 : * checkpoint, XLOGfileslop was the offset from the checkpoint record, but
3362 : * here, it was the offset from the insert location. We can't do the
3363 : * normal XLOGfileslop calculation here because we don't have access to
3364 : * the prior checkpoint's redo location. So somewhat arbitrarily, just use
3365 : * CheckPointSegments.
3366 : */
3367 2548 : max_segno = logsegno + CheckPointSegments;
3368 2548 : if (InstallXLogFileSegment(&installed_segno, tmppath, true, max_segno,
3369 : logtli))
3370 : {
3371 2548 : *added = true;
3372 2548 : elog(DEBUG2, "done creating and filling new WAL file");
3373 : }
3374 : else
3375 : {
3376 : /*
3377 : * No need for any more future segments, or InstallXLogFileSegment()
3378 : * failed to rename the file into place. If the rename failed, a
3379 : * caller opening the file may fail.
3380 : */
3381 0 : unlink(tmppath);
3382 0 : elog(DEBUG2, "abandoned new WAL file");
3383 : }
3384 :
3385 2548 : return -1;
3386 : }
3387 :
3388 : /*
3389 : * Create a new XLOG file segment, or open a pre-existing one.
3390 : *
3391 : * logsegno: identify segment to be created/opened.
3392 : *
3393 : * Returns FD of opened file.
3394 : *
3395 : * Note: errors here are ERROR not PANIC because we might or might not be
3396 : * inside a critical section (eg, during checkpoint there is no reason to
3397 : * take down the system on failure). They will promote to PANIC if we are
3398 : * in a critical section.
3399 : */
3400 : int
3401 24118 : XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)
3402 : {
3403 : bool ignore_added;
3404 : char path[MAXPGPATH];
3405 : int fd;
3406 :
3407 : Assert(logtli != 0);
3408 :
3409 24118 : fd = XLogFileInitInternal(logsegno, logtli, &ignore_added, path);
3410 24118 : if (fd >= 0)
3411 21860 : return fd;
3412 :
3413 : /* Now open original target segment (might not be file I just made) */
3414 2258 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3415 2258 : get_sync_bit(wal_sync_method));
3416 2258 : if (fd < 0)
3417 0 : ereport(ERROR,
3418 : (errcode_for_file_access(),
3419 : errmsg("could not open file \"%s\": %m", path)));
3420 2258 : return fd;
3421 : }
3422 :
3423 : /*
3424 : * Create a new XLOG file segment by copying a pre-existing one.
3425 : *
3426 : * destsegno: identify segment to be created.
3427 : *
3428 : * srcTLI, srcsegno: identify segment to be copied (could be from
3429 : * a different timeline)
3430 : *
3431 : * upto: how much of the source file to copy (the rest is filled with
3432 : * zeros)
3433 : *
3434 : * Currently this is only used during recovery, and so there are no locking
3435 : * considerations. But we should be just as tense as XLogFileInit to avoid
3436 : * emplacing a bogus file.
3437 : */
3438 : static void
3439 76 : XLogFileCopy(TimeLineID destTLI, XLogSegNo destsegno,
3440 : TimeLineID srcTLI, XLogSegNo srcsegno,
3441 : int upto)
3442 : {
3443 : char path[MAXPGPATH];
3444 : char tmppath[MAXPGPATH];
3445 : PGAlignedXLogBlock buffer;
3446 : int srcfd;
3447 : int fd;
3448 : int nbytes;
3449 :
3450 : /*
3451 : * Open the source file
3452 : */
3453 76 : XLogFilePath(path, srcTLI, srcsegno, wal_segment_size);
3454 76 : srcfd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
3455 76 : if (srcfd < 0)
3456 0 : ereport(ERROR,
3457 : (errcode_for_file_access(),
3458 : errmsg("could not open file \"%s\": %m", path)));
3459 :
3460 : /*
3461 : * Copy into a temp file name.
3462 : */
3463 76 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3464 :
3465 76 : unlink(tmppath);
3466 :
3467 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3468 76 : fd = OpenTransientFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
3469 76 : if (fd < 0)
3470 0 : ereport(ERROR,
3471 : (errcode_for_file_access(),
3472 : errmsg("could not create file \"%s\": %m", tmppath)));
3473 :
3474 : /*
3475 : * Do the data copying.
3476 : */
3477 155724 : for (nbytes = 0; nbytes < wal_segment_size; nbytes += sizeof(buffer))
3478 : {
3479 : int nread;
3480 :
3481 155648 : nread = upto - nbytes;
3482 :
3483 : /*
3484 : * The part that is not read from the source file is filled with
3485 : * zeros.
3486 : */
3487 155648 : if (nread < sizeof(buffer))
3488 76 : memset(buffer.data, 0, sizeof(buffer));
3489 :
3490 155648 : if (nread > 0)
3491 : {
3492 : int r;
3493 :
3494 5356 : if (nread > sizeof(buffer))
3495 5280 : nread = sizeof(buffer);
3496 5356 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ);
3497 5356 : r = read(srcfd, buffer.data, nread);
3498 5356 : if (r != nread)
3499 : {
3500 0 : if (r < 0)
3501 0 : ereport(ERROR,
3502 : (errcode_for_file_access(),
3503 : errmsg("could not read file \"%s\": %m",
3504 : path)));
3505 : else
3506 0 : ereport(ERROR,
3507 : (errcode(ERRCODE_DATA_CORRUPTED),
3508 : errmsg("could not read file \"%s\": read %d of %zu",
3509 : path, r, (Size) nread)));
3510 : }
3511 5356 : pgstat_report_wait_end();
3512 : }
3513 155648 : errno = 0;
3514 155648 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_WRITE);
3515 155648 : if ((int) write(fd, buffer.data, sizeof(buffer)) != (int) sizeof(buffer))
3516 : {
3517 0 : int save_errno = errno;
3518 :
3519 : /*
3520 : * If we fail to make the file, delete it to release disk space
3521 : */
3522 0 : unlink(tmppath);
3523 : /* if write didn't set errno, assume problem is no disk space */
3524 0 : errno = save_errno ? save_errno : ENOSPC;
3525 :
3526 0 : ereport(ERROR,
3527 : (errcode_for_file_access(),
3528 : errmsg("could not write to file \"%s\": %m", tmppath)));
3529 : }
3530 155648 : pgstat_report_wait_end();
3531 : }
3532 :
3533 76 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_SYNC);
3534 76 : if (pg_fsync(fd) != 0)
3535 0 : ereport(data_sync_elevel(ERROR),
3536 : (errcode_for_file_access(),
3537 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3538 76 : pgstat_report_wait_end();
3539 :
3540 76 : if (CloseTransientFile(fd) != 0)
3541 0 : ereport(ERROR,
3542 : (errcode_for_file_access(),
3543 : errmsg("could not close file \"%s\": %m", tmppath)));
3544 :
3545 76 : if (CloseTransientFile(srcfd) != 0)
3546 0 : ereport(ERROR,
3547 : (errcode_for_file_access(),
3548 : errmsg("could not close file \"%s\": %m", path)));
3549 :
3550 : /*
3551 : * Now move the segment into place with its final name.
3552 : */
3553 76 : if (!InstallXLogFileSegment(&destsegno, tmppath, false, 0, destTLI))
3554 0 : elog(ERROR, "InstallXLogFileSegment should not have failed");
3555 76 : }
3556 :
3557 : /*
3558 : * Install a new XLOG segment file as a current or future log segment.
3559 : *
3560 : * This is used both to install a newly-created segment (which has a temp
3561 : * filename while it's being created) and to recycle an old segment.
3562 : *
3563 : * *segno: identify segment to install as (or first possible target).
3564 : * When find_free is true, this is modified on return to indicate the
3565 : * actual installation location or last segment searched.
3566 : *
3567 : * tmppath: initial name of file to install. It will be renamed into place.
3568 : *
3569 : * find_free: if true, install the new segment at the first empty segno
3570 : * number at or after the passed numbers. If false, install the new segment
3571 : * exactly where specified, deleting any existing segment file there.
3572 : *
3573 : * max_segno: maximum segment number to install the new file as. Fail if no
3574 : * free slot is found between *segno and max_segno. (Ignored when find_free
3575 : * is false.)
3576 : *
3577 : * tli: The timeline on which the new segment should be installed.
3578 : *
3579 : * Returns true if the file was installed successfully. false indicates that
3580 : * max_segno limit was exceeded, the startup process has disabled this
3581 : * function for now, or an error occurred while renaming the file into place.
3582 : */
3583 : static bool
3584 4756 : InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
3585 : bool find_free, XLogSegNo max_segno, TimeLineID tli)
3586 : {
3587 : char path[MAXPGPATH];
3588 : struct stat stat_buf;
3589 :
3590 : Assert(tli != 0);
3591 :
3592 4756 : XLogFilePath(path, tli, *segno, wal_segment_size);
3593 :
3594 4756 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3595 4756 : if (!XLogCtl->InstallXLogFileSegmentActive)
3596 : {
3597 0 : LWLockRelease(ControlFileLock);
3598 0 : return false;
3599 : }
3600 :
3601 4756 : if (!find_free)
3602 : {
3603 : /* Force installation: get rid of any pre-existing segment file */
3604 76 : durable_unlink(path, DEBUG1);
3605 : }
3606 : else
3607 : {
3608 : /* Find a free slot to put it in */
3609 6972 : while (stat(path, &stat_buf) == 0)
3610 : {
3611 2350 : if ((*segno) >= max_segno)
3612 : {
3613 : /* Failed to find a free slot within specified range */
3614 58 : LWLockRelease(ControlFileLock);
3615 58 : return false;
3616 : }
3617 2292 : (*segno)++;
3618 2292 : XLogFilePath(path, tli, *segno, wal_segment_size);
3619 : }
3620 : }
3621 :
3622 : Assert(access(path, F_OK) != 0 && errno == ENOENT);
3623 4698 : if (durable_rename(tmppath, path, LOG) != 0)
3624 : {
3625 0 : LWLockRelease(ControlFileLock);
3626 : /* durable_rename already emitted log message */
3627 0 : return false;
3628 : }
3629 :
3630 4698 : LWLockRelease(ControlFileLock);
3631 :
3632 4698 : return true;
3633 : }
3634 :
3635 : /*
3636 : * Open a pre-existing logfile segment for writing.
3637 : */
3638 : int
3639 48 : XLogFileOpen(XLogSegNo segno, TimeLineID tli)
3640 : {
3641 : char path[MAXPGPATH];
3642 : int fd;
3643 :
3644 48 : XLogFilePath(path, tli, segno, wal_segment_size);
3645 :
3646 48 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3647 48 : get_sync_bit(wal_sync_method));
3648 48 : if (fd < 0)
3649 0 : ereport(PANIC,
3650 : (errcode_for_file_access(),
3651 : errmsg("could not open file \"%s\": %m", path)));
3652 :
3653 48 : return fd;
3654 : }
3655 :
3656 : /*
3657 : * Close the current logfile segment for writing.
3658 : */
3659 : static void
3660 9516 : XLogFileClose(void)
3661 : {
3662 : Assert(openLogFile >= 0);
3663 :
3664 : /*
3665 : * WAL segment files will not be re-read in normal operation, so we advise
3666 : * the OS to release any cached pages. But do not do so if WAL archiving
3667 : * or streaming is active, because archiver and walsender process could
3668 : * use the cache to read the WAL segment.
3669 : */
3670 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3671 9516 : if (!XLogIsNeeded() && (io_direct_flags & IO_DIRECT_WAL) == 0)
3672 2342 : (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3673 : #endif
3674 :
3675 9516 : if (close(openLogFile) != 0)
3676 : {
3677 : char xlogfname[MAXFNAMELEN];
3678 0 : int save_errno = errno;
3679 :
3680 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo, wal_segment_size);
3681 0 : errno = save_errno;
3682 0 : ereport(PANIC,
3683 : (errcode_for_file_access(),
3684 : errmsg("could not close file \"%s\": %m", xlogfname)));
3685 : }
3686 :
3687 9516 : openLogFile = -1;
3688 9516 : ReleaseExternalFD();
3689 9516 : }
3690 :
3691 : /*
3692 : * Preallocate log files beyond the specified log endpoint.
3693 : *
3694 : * XXX this is currently extremely conservative, since it forces only one
3695 : * future log segment to exist, and even that only if we are 75% done with
3696 : * the current one. This is only appropriate for very low-WAL-volume systems.
3697 : * High-volume systems will be OK once they've built up a sufficient set of
3698 : * recycled log segments, but the startup transient is likely to include
3699 : * a lot of segment creations by foreground processes, which is not so good.
3700 : *
3701 : * XLogFileInitInternal() can ereport(ERROR). All known causes indicate big
3702 : * trouble; for example, a full filesystem is one cause. The checkpoint WAL
3703 : * and/or ControlFile updates already completed. If a RequestCheckpoint()
3704 : * initiated the present checkpoint and an ERROR ends this function, the
3705 : * command that called RequestCheckpoint() fails. That's not ideal, but it's
3706 : * not worth contorting more functions to use caller-specified elevel values.
3707 : * (With or without RequestCheckpoint(), an ERROR forestalls some inessential
3708 : * reporting and resource reclamation.)
3709 : */
3710 : static void
3711 3008 : PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli)
3712 : {
3713 : XLogSegNo _logSegNo;
3714 : int lf;
3715 : bool added;
3716 : char path[MAXPGPATH];
3717 : uint64 offset;
3718 :
3719 3008 : if (!XLogCtl->InstallXLogFileSegmentActive)
3720 10 : return; /* unlocked check says no */
3721 :
3722 2998 : XLByteToPrevSeg(endptr, _logSegNo, wal_segment_size);
3723 2998 : offset = XLogSegmentOffset(endptr - 1, wal_segment_size);
3724 2998 : if (offset >= (uint32) (0.75 * wal_segment_size))
3725 : {
3726 456 : _logSegNo++;
3727 456 : lf = XLogFileInitInternal(_logSegNo, tli, &added, path);
3728 456 : if (lf >= 0)
3729 166 : close(lf);
3730 456 : if (added)
3731 290 : CheckpointStats.ckpt_segs_added++;
3732 : }
3733 : }
3734 :
3735 : /*
3736 : * Throws an error if the given log segment has already been removed or
3737 : * recycled. The caller should only pass a segment that it knows to have
3738 : * existed while the server has been running, as this function always
3739 : * succeeds if no WAL segments have been removed since startup.
3740 : * 'tli' is only used in the error message.
3741 : *
3742 : * Note: this function guarantees to keep errno unchanged on return.
3743 : * This supports callers that use this to possibly deliver a better
3744 : * error message about a missing file, while still being able to throw
3745 : * a normal file-access error afterwards, if this does return.
3746 : */
3747 : void
3748 223334 : CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3749 : {
3750 223334 : int save_errno = errno;
3751 : XLogSegNo lastRemovedSegNo;
3752 :
3753 223334 : SpinLockAcquire(&XLogCtl->info_lck);
3754 223334 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3755 223334 : SpinLockRelease(&XLogCtl->info_lck);
3756 :
3757 223334 : if (segno <= lastRemovedSegNo)
3758 : {
3759 : char filename[MAXFNAMELEN];
3760 :
3761 0 : XLogFileName(filename, tli, segno, wal_segment_size);
3762 0 : errno = save_errno;
3763 0 : ereport(ERROR,
3764 : (errcode_for_file_access(),
3765 : errmsg("requested WAL segment %s has already been removed",
3766 : filename)));
3767 : }
3768 223334 : errno = save_errno;
3769 223334 : }
3770 :
3771 : /*
3772 : * Return the last WAL segment removed, or 0 if no segment has been removed
3773 : * since startup.
3774 : *
3775 : * NB: the result can be out of date arbitrarily fast, the caller has to deal
3776 : * with that.
3777 : */
3778 : XLogSegNo
3779 1972 : XLogGetLastRemovedSegno(void)
3780 : {
3781 : XLogSegNo lastRemovedSegNo;
3782 :
3783 1972 : SpinLockAcquire(&XLogCtl->info_lck);
3784 1972 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3785 1972 : SpinLockRelease(&XLogCtl->info_lck);
3786 :
3787 1972 : return lastRemovedSegNo;
3788 : }
3789 :
3790 : /*
3791 : * Return the oldest WAL segment on the given TLI that still exists in
3792 : * XLOGDIR, or 0 if none.
3793 : */
3794 : XLogSegNo
3795 10 : XLogGetOldestSegno(TimeLineID tli)
3796 : {
3797 : DIR *xldir;
3798 : struct dirent *xlde;
3799 10 : XLogSegNo oldest_segno = 0;
3800 :
3801 10 : xldir = AllocateDir(XLOGDIR);
3802 66 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3803 : {
3804 : TimeLineID file_tli;
3805 : XLogSegNo file_segno;
3806 :
3807 : /* Ignore files that are not XLOG segments. */
3808 56 : if (!IsXLogFileName(xlde->d_name))
3809 40 : continue;
3810 :
3811 : /* Parse filename to get TLI and segno. */
3812 16 : XLogFromFileName(xlde->d_name, &file_tli, &file_segno,
3813 : wal_segment_size);
3814 :
3815 : /* Ignore anything that's not from the TLI of interest. */
3816 16 : if (tli != file_tli)
3817 0 : continue;
3818 :
3819 : /* If it's the oldest so far, update oldest_segno. */
3820 16 : if (oldest_segno == 0 || file_segno < oldest_segno)
3821 14 : oldest_segno = file_segno;
3822 : }
3823 :
3824 10 : FreeDir(xldir);
3825 10 : return oldest_segno;
3826 : }
3827 :
3828 : /*
3829 : * Update the last removed segno pointer in shared memory, to reflect that the
3830 : * given XLOG file has been removed.
3831 : */
3832 : static void
3833 3832 : UpdateLastRemovedPtr(char *filename)
3834 : {
3835 : uint32 tli;
3836 : XLogSegNo segno;
3837 :
3838 3832 : XLogFromFileName(filename, &tli, &segno, wal_segment_size);
3839 :
3840 3832 : SpinLockAcquire(&XLogCtl->info_lck);
3841 3832 : if (segno > XLogCtl->lastRemovedSegNo)
3842 1654 : XLogCtl->lastRemovedSegNo = segno;
3843 3832 : SpinLockRelease(&XLogCtl->info_lck);
3844 3832 : }
3845 :
3846 : /*
3847 : * Remove all temporary log files in pg_wal
3848 : *
3849 : * This is called at the beginning of recovery after a previous crash,
3850 : * at a point where no other processes write fresh WAL data.
3851 : */
3852 : static void
3853 342 : RemoveTempXlogFiles(void)
3854 : {
3855 : DIR *xldir;
3856 : struct dirent *xlde;
3857 :
3858 342 : elog(DEBUG2, "removing all temporary WAL segments");
3859 :
3860 342 : xldir = AllocateDir(XLOGDIR);
3861 2214 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3862 : {
3863 : char path[MAXPGPATH];
3864 :
3865 1872 : if (strncmp(xlde->d_name, "xlogtemp.", 9) != 0)
3866 1872 : continue;
3867 :
3868 0 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
3869 0 : unlink(path);
3870 0 : elog(DEBUG2, "removed temporary WAL segment \"%s\"", path);
3871 : }
3872 342 : FreeDir(xldir);
3873 342 : }
3874 :
3875 : /*
3876 : * Recycle or remove all log files older or equal to passed segno.
3877 : *
3878 : * endptr is current (or recent) end of xlog, and lastredoptr is the
3879 : * redo pointer of the last checkpoint. These are used to determine
3880 : * whether we want to recycle rather than delete no-longer-wanted log files.
3881 : *
3882 : * insertTLI is the current timeline for XLOG insertion. Any recycled
3883 : * segments should be reused for this timeline.
3884 : */
3885 : static void
3886 2506 : RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr, XLogRecPtr endptr,
3887 : TimeLineID insertTLI)
3888 : {
3889 : DIR *xldir;
3890 : struct dirent *xlde;
3891 : char lastoff[MAXFNAMELEN];
3892 : XLogSegNo endlogSegNo;
3893 : XLogSegNo recycleSegNo;
3894 :
3895 : /* Initialize info about where to try to recycle to */
3896 2506 : XLByteToSeg(endptr, endlogSegNo, wal_segment_size);
3897 2506 : recycleSegNo = XLOGfileslop(lastredoptr);
3898 :
3899 : /*
3900 : * Construct a filename of the last segment to be kept. The timeline ID
3901 : * doesn't matter, we ignore that in the comparison. (During recovery,
3902 : * InsertTimeLineID isn't set, so we can't use that.)
3903 : */
3904 2506 : XLogFileName(lastoff, 0, segno, wal_segment_size);
3905 :
3906 2506 : elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
3907 : lastoff);
3908 :
3909 2506 : xldir = AllocateDir(XLOGDIR);
3910 :
3911 38374 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3912 : {
3913 : /* Ignore files that are not XLOG segments */
3914 35868 : if (!IsXLogFileName(xlde->d_name) &&
3915 10466 : !IsPartialXLogFileName(xlde->d_name))
3916 10462 : continue;
3917 :
3918 : /*
3919 : * We ignore the timeline part of the XLOG segment identifiers in
3920 : * deciding whether a segment is still needed. This ensures that we
3921 : * won't prematurely remove a segment from a parent timeline. We could
3922 : * probably be a little more proactive about removing segments of
3923 : * non-parent timelines, but that would be a whole lot more
3924 : * complicated.
3925 : *
3926 : * We use the alphanumeric sorting property of the filenames to decide
3927 : * which ones are earlier than the lastoff segment.
3928 : */
3929 25406 : if (strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
3930 : {
3931 3848 : if (XLogArchiveCheckDone(xlde->d_name))
3932 : {
3933 : /* Update the last removed location in shared memory first */
3934 3832 : UpdateLastRemovedPtr(xlde->d_name);
3935 :
3936 3832 : RemoveXlogFile(xlde, recycleSegNo, &endlogSegNo, insertTLI);
3937 : }
3938 : }
3939 : }
3940 :
3941 2506 : FreeDir(xldir);
3942 2506 : }
3943 :
3944 : /*
3945 : * Recycle or remove WAL files that are not part of the given timeline's
3946 : * history.
3947 : *
3948 : * This is called during recovery, whenever we switch to follow a new
3949 : * timeline, and at the end of recovery when we create a new timeline. We
3950 : * wouldn't otherwise care about extra WAL files lying in pg_wal, but they
3951 : * might be leftover pre-allocated or recycled WAL segments on the old timeline
3952 : * that we haven't used yet, and contain garbage. If we just leave them in
3953 : * pg_wal, they will eventually be archived, and we can't let that happen.
3954 : * Files that belong to our timeline history are valid, because we have
3955 : * successfully replayed them, but from others we can't be sure.
3956 : *
3957 : * 'switchpoint' is the current point in WAL where we switch to new timeline,
3958 : * and 'newTLI' is the new timeline we switch to.
3959 : */
3960 : void
3961 120 : RemoveNonParentXlogFiles(XLogRecPtr switchpoint, TimeLineID newTLI)
3962 : {
3963 : DIR *xldir;
3964 : struct dirent *xlde;
3965 : char switchseg[MAXFNAMELEN];
3966 : XLogSegNo endLogSegNo;
3967 : XLogSegNo switchLogSegNo;
3968 : XLogSegNo recycleSegNo;
3969 :
3970 : /*
3971 : * Initialize info about where to begin the work. This will recycle,
3972 : * somewhat arbitrarily, 10 future segments.
3973 : */
3974 120 : XLByteToPrevSeg(switchpoint, switchLogSegNo, wal_segment_size);
3975 120 : XLByteToSeg(switchpoint, endLogSegNo, wal_segment_size);
3976 120 : recycleSegNo = endLogSegNo + 10;
3977 :
3978 : /*
3979 : * Construct a filename of the last segment to be kept.
3980 : */
3981 120 : XLogFileName(switchseg, newTLI, switchLogSegNo, wal_segment_size);
3982 :
3983 120 : elog(DEBUG2, "attempting to remove WAL segments newer than log file %s",
3984 : switchseg);
3985 :
3986 120 : xldir = AllocateDir(XLOGDIR);
3987 :
3988 1140 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3989 : {
3990 : /* Ignore files that are not XLOG segments */
3991 1020 : if (!IsXLogFileName(xlde->d_name))
3992 636 : continue;
3993 :
3994 : /*
3995 : * Remove files that are on a timeline older than the new one we're
3996 : * switching to, but with a segment number >= the first segment on the
3997 : * new timeline.
3998 : */
3999 384 : if (strncmp(xlde->d_name, switchseg, 8) < 0 &&
4000 248 : strcmp(xlde->d_name + 8, switchseg + 8) > 0)
4001 : {
4002 : /*
4003 : * If the file has already been marked as .ready, however, don't
4004 : * remove it yet. It should be OK to remove it - files that are
4005 : * not part of our timeline history are not required for recovery
4006 : * - but seems safer to let them be archived and removed later.
4007 : */
4008 32 : if (!XLogArchiveIsReady(xlde->d_name))
4009 32 : RemoveXlogFile(xlde, recycleSegNo, &endLogSegNo, newTLI);
4010 : }
4011 : }
4012 :
4013 120 : FreeDir(xldir);
4014 120 : }
4015 :
4016 : /*
4017 : * Recycle or remove a log file that's no longer needed.
4018 : *
4019 : * segment_de is the dirent structure of the segment to recycle or remove.
4020 : * recycleSegNo is the segment number to recycle up to. endlogSegNo is
4021 : * the segment number of the current (or recent) end of WAL.
4022 : *
4023 : * endlogSegNo gets incremented if the segment is recycled so as it is not
4024 : * checked again with future callers of this function.
4025 : *
4026 : * insertTLI is the current timeline for XLOG insertion. Any recycled segments
4027 : * should be used for this timeline.
4028 : */
4029 : static void
4030 3864 : RemoveXlogFile(const struct dirent *segment_de,
4031 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
4032 : TimeLineID insertTLI)
4033 : {
4034 : char path[MAXPGPATH];
4035 : #ifdef WIN32
4036 : char newpath[MAXPGPATH];
4037 : #endif
4038 3864 : const char *segname = segment_de->d_name;
4039 :
4040 3864 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", segname);
4041 :
4042 : /*
4043 : * Before deleting the file, see if it can be recycled as a future log
4044 : * segment. Only recycle normal files, because we don't want to recycle
4045 : * symbolic links pointing to a separate archive directory.
4046 : */
4047 3864 : if (wal_recycle &&
4048 3864 : *endlogSegNo <= recycleSegNo &&
4049 4276 : XLogCtl->InstallXLogFileSegmentActive && /* callee rechecks this */
4050 4264 : get_dirent_type(path, segment_de, false, DEBUG2) == PGFILETYPE_REG &&
4051 2132 : InstallXLogFileSegment(endlogSegNo, path,
4052 : true, recycleSegNo, insertTLI))
4053 : {
4054 2074 : ereport(DEBUG2,
4055 : (errmsg_internal("recycled write-ahead log file \"%s\"",
4056 : segname)));
4057 2074 : CheckpointStats.ckpt_segs_recycled++;
4058 : /* Needn't recheck that slot on future iterations */
4059 2074 : (*endlogSegNo)++;
4060 : }
4061 : else
4062 : {
4063 : /* No need for any more future segments, or recycling failed ... */
4064 : int rc;
4065 :
4066 1790 : ereport(DEBUG2,
4067 : (errmsg_internal("removing write-ahead log file \"%s\"",
4068 : segname)));
4069 :
4070 : #ifdef WIN32
4071 :
4072 : /*
4073 : * On Windows, if another process (e.g another backend) holds the file
4074 : * open in FILE_SHARE_DELETE mode, unlink will succeed, but the file
4075 : * will still show up in directory listing until the last handle is
4076 : * closed. To avoid confusing the lingering deleted file for a live
4077 : * WAL file that needs to be archived, rename it before deleting it.
4078 : *
4079 : * If another process holds the file open without FILE_SHARE_DELETE
4080 : * flag, rename will fail. We'll try again at the next checkpoint.
4081 : */
4082 : snprintf(newpath, MAXPGPATH, "%s.deleted", path);
4083 : if (rename(path, newpath) != 0)
4084 : {
4085 : ereport(LOG,
4086 : (errcode_for_file_access(),
4087 : errmsg("could not rename file \"%s\": %m",
4088 : path)));
4089 : return;
4090 : }
4091 : rc = durable_unlink(newpath, LOG);
4092 : #else
4093 1790 : rc = durable_unlink(path, LOG);
4094 : #endif
4095 1790 : if (rc != 0)
4096 : {
4097 : /* Message already logged by durable_unlink() */
4098 0 : return;
4099 : }
4100 1790 : CheckpointStats.ckpt_segs_removed++;
4101 : }
4102 :
4103 3864 : XLogArchiveCleanup(segname);
4104 : }
4105 :
4106 : /*
4107 : * Verify whether pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
4108 : * If the latter do not exist, recreate them.
4109 : *
4110 : * It is not the goal of this function to verify the contents of these
4111 : * directories, but to help in cases where someone has performed a cluster
4112 : * copy for PITR purposes but omitted pg_wal from the copy.
4113 : *
4114 : * We could also recreate pg_wal if it doesn't exist, but a deliberate
4115 : * policy decision was made not to. It is fairly common for pg_wal to be
4116 : * a symlink, and if that was the DBA's intent then automatically making a
4117 : * plain directory would result in degraded performance with no notice.
4118 : */
4119 : static void
4120 1668 : ValidateXLOGDirectoryStructure(void)
4121 : {
4122 : char path[MAXPGPATH];
4123 : struct stat stat_buf;
4124 :
4125 : /* Check for pg_wal; if it doesn't exist, error out */
4126 1668 : if (stat(XLOGDIR, &stat_buf) != 0 ||
4127 1668 : !S_ISDIR(stat_buf.st_mode))
4128 0 : ereport(FATAL,
4129 : (errcode_for_file_access(),
4130 : errmsg("required WAL directory \"%s\" does not exist",
4131 : XLOGDIR)));
4132 :
4133 : /* Check for archive_status */
4134 1668 : snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4135 1668 : if (stat(path, &stat_buf) == 0)
4136 : {
4137 : /* Check for weird cases where it exists but isn't a directory */
4138 1666 : if (!S_ISDIR(stat_buf.st_mode))
4139 0 : ereport(FATAL,
4140 : (errcode_for_file_access(),
4141 : errmsg("required WAL directory \"%s\" does not exist",
4142 : path)));
4143 : }
4144 : else
4145 : {
4146 2 : ereport(LOG,
4147 : (errmsg("creating missing WAL directory \"%s\"", path)));
4148 2 : if (MakePGDirectory(path) < 0)
4149 0 : ereport(FATAL,
4150 : (errcode_for_file_access(),
4151 : errmsg("could not create missing directory \"%s\": %m",
4152 : path)));
4153 : }
4154 :
4155 : /* Check for summaries */
4156 1668 : snprintf(path, MAXPGPATH, XLOGDIR "/summaries");
4157 1668 : if (stat(path, &stat_buf) == 0)
4158 : {
4159 : /* Check for weird cases where it exists but isn't a directory */
4160 1666 : if (!S_ISDIR(stat_buf.st_mode))
4161 0 : ereport(FATAL,
4162 : (errmsg("required WAL directory \"%s\" does not exist",
4163 : path)));
4164 : }
4165 : else
4166 : {
4167 2 : ereport(LOG,
4168 : (errmsg("creating missing WAL directory \"%s\"", path)));
4169 2 : if (MakePGDirectory(path) < 0)
4170 0 : ereport(FATAL,
4171 : (errmsg("could not create missing directory \"%s\": %m",
4172 : path)));
4173 : }
4174 1668 : }
4175 :
4176 : /*
4177 : * Remove previous backup history files. This also retries creation of
4178 : * .ready files for any backup history files for which XLogArchiveNotify
4179 : * failed earlier.
4180 : */
4181 : static void
4182 290 : CleanupBackupHistory(void)
4183 : {
4184 : DIR *xldir;
4185 : struct dirent *xlde;
4186 : char path[MAXPGPATH + sizeof(XLOGDIR)];
4187 :
4188 290 : xldir = AllocateDir(XLOGDIR);
4189 :
4190 2622 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4191 : {
4192 2332 : if (IsBackupHistoryFileName(xlde->d_name))
4193 : {
4194 306 : if (XLogArchiveCheckDone(xlde->d_name))
4195 : {
4196 240 : elog(DEBUG2, "removing WAL backup history file \"%s\"",
4197 : xlde->d_name);
4198 240 : snprintf(path, sizeof(path), XLOGDIR "/%s", xlde->d_name);
4199 240 : unlink(path);
4200 240 : XLogArchiveCleanup(xlde->d_name);
4201 : }
4202 : }
4203 : }
4204 :
4205 290 : FreeDir(xldir);
4206 290 : }
4207 :
4208 : /*
4209 : * I/O routines for pg_control
4210 : *
4211 : * *ControlFile is a buffer in shared memory that holds an image of the
4212 : * contents of pg_control. WriteControlFile() initializes pg_control
4213 : * given a preloaded buffer, ReadControlFile() loads the buffer from
4214 : * the pg_control file (during postmaster or standalone-backend startup),
4215 : * and UpdateControlFile() rewrites pg_control after we modify xlog state.
4216 : * InitControlFile() fills the buffer with initial values.
4217 : *
4218 : * For simplicity, WriteControlFile() initializes the fields of pg_control
4219 : * that are related to checking backend/database compatibility, and
4220 : * ReadControlFile() verifies they are correct. We could split out the
4221 : * I/O and compatibility-check functions, but there seems no need currently.
4222 : */
4223 :
4224 : static void
4225 90 : InitControlFile(uint64 sysidentifier, uint32 data_checksum_version)
4226 : {
4227 : char mock_auth_nonce[MOCK_AUTH_NONCE_LEN];
4228 :
4229 : /*
4230 : * Generate a random nonce. This is used for authentication requests that
4231 : * will fail because the user does not exist. The nonce is used to create
4232 : * a genuine-looking password challenge for the non-existent user, in lieu
4233 : * of an actual stored password.
4234 : */
4235 90 : if (!pg_strong_random(mock_auth_nonce, MOCK_AUTH_NONCE_LEN))
4236 0 : ereport(PANIC,
4237 : (errcode(ERRCODE_INTERNAL_ERROR),
4238 : errmsg("could not generate secret authorization token")));
4239 :
4240 90 : memset(ControlFile, 0, sizeof(ControlFileData));
4241 : /* Initialize pg_control status fields */
4242 90 : ControlFile->system_identifier = sysidentifier;
4243 90 : memcpy(ControlFile->mock_authentication_nonce, mock_auth_nonce, MOCK_AUTH_NONCE_LEN);
4244 90 : ControlFile->state = DB_SHUTDOWNED;
4245 90 : ControlFile->unloggedLSN = FirstNormalUnloggedLSN;
4246 :
4247 : /* Set important parameter values for use when replaying WAL */
4248 90 : ControlFile->MaxConnections = MaxConnections;
4249 90 : ControlFile->max_worker_processes = max_worker_processes;
4250 90 : ControlFile->max_wal_senders = max_wal_senders;
4251 90 : ControlFile->max_prepared_xacts = max_prepared_xacts;
4252 90 : ControlFile->max_locks_per_xact = max_locks_per_xact;
4253 90 : ControlFile->wal_level = wal_level;
4254 90 : ControlFile->wal_log_hints = wal_log_hints;
4255 90 : ControlFile->track_commit_timestamp = track_commit_timestamp;
4256 90 : ControlFile->data_checksum_version = data_checksum_version;
4257 90 : }
4258 :
4259 : static void
4260 90 : WriteControlFile(void)
4261 : {
4262 : int fd;
4263 : char buffer[PG_CONTROL_FILE_SIZE]; /* need not be aligned */
4264 :
4265 : /*
4266 : * Initialize version and compatibility-check fields
4267 : */
4268 90 : ControlFile->pg_control_version = PG_CONTROL_VERSION;
4269 90 : ControlFile->catalog_version_no = CATALOG_VERSION_NO;
4270 :
4271 90 : ControlFile->maxAlign = MAXIMUM_ALIGNOF;
4272 90 : ControlFile->floatFormat = FLOATFORMAT_VALUE;
4273 :
4274 90 : ControlFile->blcksz = BLCKSZ;
4275 90 : ControlFile->relseg_size = RELSEG_SIZE;
4276 90 : ControlFile->xlog_blcksz = XLOG_BLCKSZ;
4277 90 : ControlFile->xlog_seg_size = wal_segment_size;
4278 :
4279 90 : ControlFile->nameDataLen = NAMEDATALEN;
4280 90 : ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
4281 :
4282 90 : ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
4283 90 : ControlFile->loblksize = LOBLKSIZE;
4284 :
4285 90 : ControlFile->float8ByVal = FLOAT8PASSBYVAL;
4286 :
4287 : /*
4288 : * Initialize the default 'char' signedness.
4289 : *
4290 : * The signedness of the char type is implementation-defined. For instance
4291 : * on x86 architecture CPUs, the char data type is typically treated as
4292 : * signed by default, whereas on aarch architecture CPUs, it is typically
4293 : * treated as unsigned by default. In v17 or earlier, we accidentally let
4294 : * C implementation signedness affect persistent data. This led to
4295 : * inconsistent results when comparing char data across different
4296 : * platforms.
4297 : *
4298 : * This flag can be used as a hint to ensure consistent behavior for
4299 : * pre-v18 data files that store data sorted by the 'char' type on disk,
4300 : * especially in cross-platform replication scenarios.
4301 : *
4302 : * Newly created database clusters unconditionally set the default char
4303 : * signedness to true. pg_upgrade changes this flag for clusters that were
4304 : * initialized on signedness=false platforms. As a result,
4305 : * signedness=false setting will become rare over time. If we had known
4306 : * about this problem during the last development cycle that forced initdb
4307 : * (v8.3), we would have made all clusters signed or all clusters
4308 : * unsigned. Making pg_upgrade the only source of signedness=false will
4309 : * cause the population of database clusters to converge toward that
4310 : * retrospective ideal.
4311 : */
4312 90 : ControlFile->default_char_signedness = true;
4313 :
4314 : /* Contents are protected with a CRC */
4315 90 : INIT_CRC32C(ControlFile->crc);
4316 90 : COMP_CRC32C(ControlFile->crc,
4317 : ControlFile,
4318 : offsetof(ControlFileData, crc));
4319 90 : FIN_CRC32C(ControlFile->crc);
4320 :
4321 : /*
4322 : * We write out PG_CONTROL_FILE_SIZE bytes into pg_control, zero-padding
4323 : * the excess over sizeof(ControlFileData). This reduces the odds of
4324 : * premature-EOF errors when reading pg_control. We'll still fail when we
4325 : * check the contents of the file, but hopefully with a more specific
4326 : * error than "couldn't read pg_control".
4327 : */
4328 90 : memset(buffer, 0, PG_CONTROL_FILE_SIZE);
4329 90 : memcpy(buffer, ControlFile, sizeof(ControlFileData));
4330 :
4331 90 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4332 : O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
4333 90 : if (fd < 0)
4334 0 : ereport(PANIC,
4335 : (errcode_for_file_access(),
4336 : errmsg("could not create file \"%s\": %m",
4337 : XLOG_CONTROL_FILE)));
4338 :
4339 90 : errno = 0;
4340 90 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_WRITE);
4341 90 : if (write(fd, buffer, PG_CONTROL_FILE_SIZE) != PG_CONTROL_FILE_SIZE)
4342 : {
4343 : /* if write didn't set errno, assume problem is no disk space */
4344 0 : if (errno == 0)
4345 0 : errno = ENOSPC;
4346 0 : ereport(PANIC,
4347 : (errcode_for_file_access(),
4348 : errmsg("could not write to file \"%s\": %m",
4349 : XLOG_CONTROL_FILE)));
4350 : }
4351 90 : pgstat_report_wait_end();
4352 :
4353 90 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_SYNC);
4354 90 : if (pg_fsync(fd) != 0)
4355 0 : ereport(PANIC,
4356 : (errcode_for_file_access(),
4357 : errmsg("could not fsync file \"%s\": %m",
4358 : XLOG_CONTROL_FILE)));
4359 90 : pgstat_report_wait_end();
4360 :
4361 90 : if (close(fd) != 0)
4362 0 : ereport(PANIC,
4363 : (errcode_for_file_access(),
4364 : errmsg("could not close file \"%s\": %m",
4365 : XLOG_CONTROL_FILE)));
4366 90 : }
4367 :
4368 : static void
4369 1774 : ReadControlFile(void)
4370 : {
4371 : pg_crc32c crc;
4372 : int fd;
4373 : char wal_segsz_str[20];
4374 : int r;
4375 :
4376 : /*
4377 : * Read data...
4378 : */
4379 1774 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4380 : O_RDWR | PG_BINARY);
4381 1774 : if (fd < 0)
4382 0 : ereport(PANIC,
4383 : (errcode_for_file_access(),
4384 : errmsg("could not open file \"%s\": %m",
4385 : XLOG_CONTROL_FILE)));
4386 :
4387 1774 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ);
4388 1774 : r = read(fd, ControlFile, sizeof(ControlFileData));
4389 1774 : if (r != sizeof(ControlFileData))
4390 : {
4391 0 : if (r < 0)
4392 0 : ereport(PANIC,
4393 : (errcode_for_file_access(),
4394 : errmsg("could not read file \"%s\": %m",
4395 : XLOG_CONTROL_FILE)));
4396 : else
4397 0 : ereport(PANIC,
4398 : (errcode(ERRCODE_DATA_CORRUPTED),
4399 : errmsg("could not read file \"%s\": read %d of %zu",
4400 : XLOG_CONTROL_FILE, r, sizeof(ControlFileData))));
4401 : }
4402 1774 : pgstat_report_wait_end();
4403 :
4404 1774 : close(fd);
4405 :
4406 : /*
4407 : * Check for expected pg_control format version. If this is wrong, the
4408 : * CRC check will likely fail because we'll be checking the wrong number
4409 : * of bytes. Complaining about wrong version will probably be more
4410 : * enlightening than complaining about wrong CRC.
4411 : */
4412 :
4413 1774 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
4414 0 : ereport(FATAL,
4415 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4416 : errmsg("database files are incompatible with server"),
4417 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4418 : " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4419 : ControlFile->pg_control_version, ControlFile->pg_control_version,
4420 : PG_CONTROL_VERSION, PG_CONTROL_VERSION),
4421 : errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4422 :
4423 1774 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
4424 0 : ereport(FATAL,
4425 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4426 : errmsg("database files are incompatible with server"),
4427 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4428 : " but the server was compiled with PG_CONTROL_VERSION %d.",
4429 : ControlFile->pg_control_version, PG_CONTROL_VERSION),
4430 : errhint("It looks like you need to initdb.")));
4431 :
4432 : /* Now check the CRC. */
4433 1774 : INIT_CRC32C(crc);
4434 1774 : COMP_CRC32C(crc,
4435 : ControlFile,
4436 : offsetof(ControlFileData, crc));
4437 1774 : FIN_CRC32C(crc);
4438 :
4439 1774 : if (!EQ_CRC32C(crc, ControlFile->crc))
4440 0 : ereport(FATAL,
4441 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4442 : errmsg("incorrect checksum in control file")));
4443 :
4444 : /*
4445 : * Do compatibility checking immediately. If the database isn't
4446 : * compatible with the backend executable, we want to abort before we can
4447 : * possibly do any damage.
4448 : */
4449 1774 : if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
4450 0 : ereport(FATAL,
4451 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4452 : errmsg("database files are incompatible with server"),
4453 : /* translator: %s is a variable name and %d is its value */
4454 : errdetail("The database cluster was initialized with %s %d,"
4455 : " but the server was compiled with %s %d.",
4456 : "CATALOG_VERSION_NO", ControlFile->catalog_version_no,
4457 : "CATALOG_VERSION_NO", CATALOG_VERSION_NO),
4458 : errhint("It looks like you need to initdb.")));
4459 1774 : if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
4460 0 : ereport(FATAL,
4461 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4462 : errmsg("database files are incompatible with server"),
4463 : /* translator: %s is a variable name and %d is its value */
4464 : errdetail("The database cluster was initialized with %s %d,"
4465 : " but the server was compiled with %s %d.",
4466 : "MAXALIGN", ControlFile->maxAlign,
4467 : "MAXALIGN", MAXIMUM_ALIGNOF),
4468 : errhint("It looks like you need to initdb.")));
4469 1774 : if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
4470 0 : ereport(FATAL,
4471 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4472 : errmsg("database files are incompatible with server"),
4473 : errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4474 : errhint("It looks like you need to initdb.")));
4475 1774 : if (ControlFile->blcksz != BLCKSZ)
4476 0 : ereport(FATAL,
4477 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4478 : errmsg("database files are incompatible with server"),
4479 : /* translator: %s is a variable name and %d is its value */
4480 : errdetail("The database cluster was initialized with %s %d,"
4481 : " but the server was compiled with %s %d.",
4482 : "BLCKSZ", ControlFile->blcksz,
4483 : "BLCKSZ", BLCKSZ),
4484 : errhint("It looks like you need to recompile or initdb.")));
4485 1774 : if (ControlFile->relseg_size != RELSEG_SIZE)
4486 0 : ereport(FATAL,
4487 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4488 : errmsg("database files are incompatible with server"),
4489 : /* translator: %s is a variable name and %d is its value */
4490 : errdetail("The database cluster was initialized with %s %d,"
4491 : " but the server was compiled with %s %d.",
4492 : "RELSEG_SIZE", ControlFile->relseg_size,
4493 : "RELSEG_SIZE", RELSEG_SIZE),
4494 : errhint("It looks like you need to recompile or initdb.")));
4495 1774 : if (ControlFile->xlog_blcksz != XLOG_BLCKSZ)
4496 0 : ereport(FATAL,
4497 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4498 : errmsg("database files are incompatible with server"),
4499 : /* translator: %s is a variable name and %d is its value */
4500 : errdetail("The database cluster was initialized with %s %d,"
4501 : " but the server was compiled with %s %d.",
4502 : "XLOG_BLCKSZ", ControlFile->xlog_blcksz,
4503 : "XLOG_BLCKSZ", XLOG_BLCKSZ),
4504 : errhint("It looks like you need to recompile or initdb.")));
4505 1774 : if (ControlFile->nameDataLen != NAMEDATALEN)
4506 0 : ereport(FATAL,
4507 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4508 : errmsg("database files are incompatible with server"),
4509 : /* translator: %s is a variable name and %d is its value */
4510 : errdetail("The database cluster was initialized with %s %d,"
4511 : " but the server was compiled with %s %d.",
4512 : "NAMEDATALEN", ControlFile->nameDataLen,
4513 : "NAMEDATALEN", NAMEDATALEN),
4514 : errhint("It looks like you need to recompile or initdb.")));
4515 1774 : if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
4516 0 : ereport(FATAL,
4517 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4518 : errmsg("database files are incompatible with server"),
4519 : /* translator: %s is a variable name and %d is its value */
4520 : errdetail("The database cluster was initialized with %s %d,"
4521 : " but the server was compiled with %s %d.",
4522 : "INDEX_MAX_KEYS", ControlFile->indexMaxKeys,
4523 : "INDEX_MAX_KEYS", INDEX_MAX_KEYS),
4524 : errhint("It looks like you need to recompile or initdb.")));
4525 1774 : if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
4526 0 : ereport(FATAL,
4527 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4528 : errmsg("database files are incompatible with server"),
4529 : /* translator: %s is a variable name and %d is its value */
4530 : errdetail("The database cluster was initialized with %s %d,"
4531 : " but the server was compiled with %s %d.",
4532 : "TOAST_MAX_CHUNK_SIZE", ControlFile->toast_max_chunk_size,
4533 : "TOAST_MAX_CHUNK_SIZE", (int) TOAST_MAX_CHUNK_SIZE),
4534 : errhint("It looks like you need to recompile or initdb.")));
4535 1774 : if (ControlFile->loblksize != LOBLKSIZE)
4536 0 : ereport(FATAL,
4537 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4538 : errmsg("database files are incompatible with server"),
4539 : /* translator: %s is a variable name and %d is its value */
4540 : errdetail("The database cluster was initialized with %s %d,"
4541 : " but the server was compiled with %s %d.",
4542 : "LOBLKSIZE", ControlFile->loblksize,
4543 : "LOBLKSIZE", (int) LOBLKSIZE),
4544 : errhint("It looks like you need to recompile or initdb.")));
4545 :
4546 : #ifdef USE_FLOAT8_BYVAL
4547 1774 : if (ControlFile->float8ByVal != true)
4548 0 : ereport(FATAL,
4549 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4550 : errmsg("database files are incompatible with server"),
4551 : errdetail("The database cluster was initialized without USE_FLOAT8_BYVAL"
4552 : " but the server was compiled with USE_FLOAT8_BYVAL."),
4553 : errhint("It looks like you need to recompile or initdb.")));
4554 : #else
4555 : if (ControlFile->float8ByVal != false)
4556 : ereport(FATAL,
4557 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4558 : errmsg("database files are incompatible with server"),
4559 : errdetail("The database cluster was initialized with USE_FLOAT8_BYVAL"
4560 : " but the server was compiled without USE_FLOAT8_BYVAL."),
4561 : errhint("It looks like you need to recompile or initdb.")));
4562 : #endif
4563 :
4564 1774 : wal_segment_size = ControlFile->xlog_seg_size;
4565 :
4566 1774 : if (!IsValidWalSegSize(wal_segment_size))
4567 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4568 : errmsg_plural("invalid WAL segment size in control file (%d byte)",
4569 : "invalid WAL segment size in control file (%d bytes)",
4570 : wal_segment_size,
4571 : wal_segment_size),
4572 : errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.")));
4573 :
4574 1774 : snprintf(wal_segsz_str, sizeof(wal_segsz_str), "%d", wal_segment_size);
4575 1774 : SetConfigOption("wal_segment_size", wal_segsz_str, PGC_INTERNAL,
4576 : PGC_S_DYNAMIC_DEFAULT);
4577 :
4578 : /* check and update variables dependent on wal_segment_size */
4579 1774 : if (ConvertToXSegs(min_wal_size_mb, wal_segment_size) < 2)
4580 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4581 : /* translator: both %s are GUC names */
4582 : errmsg("\"%s\" must be at least twice \"%s\"",
4583 : "min_wal_size", "wal_segment_size")));
4584 :
4585 1774 : if (ConvertToXSegs(max_wal_size_mb, wal_segment_size) < 2)
4586 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4587 : /* translator: both %s are GUC names */
4588 : errmsg("\"%s\" must be at least twice \"%s\"",
4589 : "max_wal_size", "wal_segment_size")));
4590 :
4591 1774 : UsableBytesInSegment =
4592 1774 : (wal_segment_size / XLOG_BLCKSZ * UsableBytesInPage) -
4593 : (SizeOfXLogLongPHD - SizeOfXLogShortPHD);
4594 :
4595 1774 : CalculateCheckpointSegments();
4596 :
4597 : /* Make the initdb settings visible as GUC variables, too */
4598 1774 : SetConfigOption("data_checksums", DataChecksumsEnabled() ? "yes" : "no",
4599 : PGC_INTERNAL, PGC_S_DYNAMIC_DEFAULT);
4600 1774 : }
4601 :
4602 : /*
4603 : * Utility wrapper to update the control file. Note that the control
4604 : * file gets flushed.
4605 : */
4606 : static void
4607 16206 : UpdateControlFile(void)
4608 : {
4609 16206 : update_controlfile(DataDir, ControlFile, true);
4610 16206 : }
4611 :
4612 : /*
4613 : * Returns the unique system identifier from control file.
4614 : */
4615 : uint64
4616 2626 : GetSystemIdentifier(void)
4617 : {
4618 : Assert(ControlFile != NULL);
4619 2626 : return ControlFile->system_identifier;
4620 : }
4621 :
4622 : /*
4623 : * Returns the random nonce from control file.
4624 : */
4625 : char *
4626 2 : GetMockAuthenticationNonce(void)
4627 : {
4628 : Assert(ControlFile != NULL);
4629 2 : return ControlFile->mock_authentication_nonce;
4630 : }
4631 :
4632 : /*
4633 : * Are checksums enabled for data pages?
4634 : */
4635 : bool
4636 22756292 : DataChecksumsEnabled(void)
4637 : {
4638 : Assert(ControlFile != NULL);
4639 22756292 : return (ControlFile->data_checksum_version > 0);
4640 : }
4641 :
4642 : /*
4643 : * Return true if the cluster was initialized on a platform where the
4644 : * default signedness of char is "signed". This function exists for code
4645 : * that deals with pre-v18 data files that store data sorted by the 'char'
4646 : * type on disk (e.g., GIN and GiST indexes). See the comments in
4647 : * WriteControlFile() for details.
4648 : */
4649 : bool
4650 6 : GetDefaultCharSignedness(void)
4651 : {
4652 6 : return ControlFile->default_char_signedness;
4653 : }
4654 :
4655 : /*
4656 : * Returns a fake LSN for unlogged relations.
4657 : *
4658 : * Each call generates an LSN that is greater than any previous value
4659 : * returned. The current counter value is saved and restored across clean
4660 : * shutdowns, but like unlogged relations, does not survive a crash. This can
4661 : * be used in lieu of real LSN values returned by XLogInsert, if you need an
4662 : * LSN-like increasing sequence of numbers without writing any WAL.
4663 : */
4664 : XLogRecPtr
4665 66 : GetFakeLSNForUnloggedRel(void)
4666 : {
4667 66 : return pg_atomic_fetch_add_u64(&XLogCtl->unloggedLSN, 1);
4668 : }
4669 :
4670 : /*
4671 : * Auto-tune the number of XLOG buffers.
4672 : *
4673 : * The preferred setting for wal_buffers is about 3% of shared_buffers, with
4674 : * a maximum of one XLOG segment (there is little reason to think that more
4675 : * is helpful, at least so long as we force an fsync when switching log files)
4676 : * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
4677 : * 9.1, when auto-tuning was added).
4678 : *
4679 : * This should not be called until NBuffers has received its final value.
4680 : */
4681 : static int
4682 1934 : XLOGChooseNumBuffers(void)
4683 : {
4684 : int xbuffers;
4685 :
4686 1934 : xbuffers = NBuffers / 32;
4687 1934 : if (xbuffers > (wal_segment_size / XLOG_BLCKSZ))
4688 44 : xbuffers = (wal_segment_size / XLOG_BLCKSZ);
4689 1934 : if (xbuffers < 8)
4690 756 : xbuffers = 8;
4691 1934 : return xbuffers;
4692 : }
4693 :
4694 : /*
4695 : * GUC check_hook for wal_buffers
4696 : */
4697 : bool
4698 3938 : check_wal_buffers(int *newval, void **extra, GucSource source)
4699 : {
4700 : /*
4701 : * -1 indicates a request for auto-tune.
4702 : */
4703 3938 : if (*newval == -1)
4704 : {
4705 : /*
4706 : * If we haven't yet changed the boot_val default of -1, just let it
4707 : * be. We'll fix it when XLOGShmemSize is called.
4708 : */
4709 2004 : if (XLOGbuffers == -1)
4710 2004 : return true;
4711 :
4712 : /* Otherwise, substitute the auto-tune value */
4713 0 : *newval = XLOGChooseNumBuffers();
4714 : }
4715 :
4716 : /*
4717 : * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
4718 : * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
4719 : * the case, we just silently treat such values as a request for the
4720 : * minimum. (We could throw an error instead, but that doesn't seem very
4721 : * helpful.)
4722 : */
4723 1934 : if (*newval < 4)
4724 0 : *newval = 4;
4725 :
4726 1934 : return true;
4727 : }
4728 :
4729 : /*
4730 : * GUC check_hook for wal_consistency_checking
4731 : */
4732 : bool
4733 3788 : check_wal_consistency_checking(char **newval, void **extra, GucSource source)
4734 : {
4735 : char *rawstring;
4736 : List *elemlist;
4737 : ListCell *l;
4738 : bool newwalconsistency[RM_MAX_ID + 1];
4739 :
4740 : /* Initialize the array */
4741 125004 : MemSet(newwalconsistency, 0, (RM_MAX_ID + 1) * sizeof(bool));
4742 :
4743 : /* Need a modifiable copy of string */
4744 3788 : rawstring = pstrdup(*newval);
4745 :
4746 : /* Parse string into list of identifiers */
4747 3788 : if (!SplitIdentifierString(rawstring, ',', &elemlist))
4748 : {
4749 : /* syntax error in list */
4750 0 : GUC_check_errdetail("List syntax is invalid.");
4751 0 : pfree(rawstring);
4752 0 : list_free(elemlist);
4753 0 : return false;
4754 : }
4755 :
4756 4686 : foreach(l, elemlist)
4757 : {
4758 898 : char *tok = (char *) lfirst(l);
4759 : int rmid;
4760 :
4761 : /* Check for 'all'. */
4762 898 : if (pg_strcasecmp(tok, "all") == 0)
4763 : {
4764 229758 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4765 228864 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL)
4766 8940 : newwalconsistency[rmid] = true;
4767 : }
4768 : else
4769 : {
4770 : /* Check if the token matches any known resource manager. */
4771 4 : bool found = false;
4772 :
4773 72 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
4774 : {
4775 108 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL &&
4776 36 : pg_strcasecmp(tok, GetRmgr(rmid).rm_name) == 0)
4777 : {
4778 4 : newwalconsistency[rmid] = true;
4779 4 : found = true;
4780 4 : break;
4781 : }
4782 : }
4783 4 : if (!found)
4784 : {
4785 : /*
4786 : * During startup, it might be a not-yet-loaded custom
4787 : * resource manager. Defer checking until
4788 : * InitializeWalConsistencyChecking().
4789 : */
4790 0 : if (!process_shared_preload_libraries_done)
4791 : {
4792 0 : check_wal_consistency_checking_deferred = true;
4793 : }
4794 : else
4795 : {
4796 0 : GUC_check_errdetail("Unrecognized key word: \"%s\".", tok);
4797 0 : pfree(rawstring);
4798 0 : list_free(elemlist);
4799 0 : return false;
4800 : }
4801 : }
4802 : }
4803 : }
4804 :
4805 3788 : pfree(rawstring);
4806 3788 : list_free(elemlist);
4807 :
4808 : /* assign new value */
4809 3788 : *extra = guc_malloc(ERROR, (RM_MAX_ID + 1) * sizeof(bool));
4810 3788 : memcpy(*extra, newwalconsistency, (RM_MAX_ID + 1) * sizeof(bool));
4811 3788 : return true;
4812 : }
4813 :
4814 : /*
4815 : * GUC assign_hook for wal_consistency_checking
4816 : */
4817 : void
4818 3786 : assign_wal_consistency_checking(const char *newval, void *extra)
4819 : {
4820 : /*
4821 : * If some checks were deferred, it's possible that the checks will fail
4822 : * later during InitializeWalConsistencyChecking(). But in that case, the
4823 : * postmaster will exit anyway, so it's safe to proceed with the
4824 : * assignment.
4825 : *
4826 : * Any built-in resource managers specified are assigned immediately,
4827 : * which affects WAL created before shared_preload_libraries are
4828 : * processed. Any custom resource managers specified won't be assigned
4829 : * until after shared_preload_libraries are processed, but that's OK
4830 : * because WAL for a custom resource manager can't be written before the
4831 : * module is loaded anyway.
4832 : */
4833 3786 : wal_consistency_checking = extra;
4834 3786 : }
4835 :
4836 : /*
4837 : * InitializeWalConsistencyChecking: run after loading custom resource managers
4838 : *
4839 : * If any unknown resource managers were specified in the
4840 : * wal_consistency_checking GUC, processing was deferred. Now that
4841 : * shared_preload_libraries have been loaded, process wal_consistency_checking
4842 : * again.
4843 : */
4844 : void
4845 1664 : InitializeWalConsistencyChecking(void)
4846 : {
4847 : Assert(process_shared_preload_libraries_done);
4848 :
4849 1664 : if (check_wal_consistency_checking_deferred)
4850 : {
4851 : struct config_generic *guc;
4852 :
4853 0 : guc = find_option("wal_consistency_checking", false, false, ERROR);
4854 :
4855 0 : check_wal_consistency_checking_deferred = false;
4856 :
4857 0 : set_config_option_ext("wal_consistency_checking",
4858 : wal_consistency_checking_string,
4859 : guc->scontext, guc->source, guc->srole,
4860 : GUC_ACTION_SET, true, ERROR, false);
4861 :
4862 : /* checking should not be deferred again */
4863 : Assert(!check_wal_consistency_checking_deferred);
4864 : }
4865 1664 : }
4866 :
4867 : /*
4868 : * GUC show_hook for archive_command
4869 : */
4870 : const char *
4871 3414 : show_archive_command(void)
4872 : {
4873 3414 : if (XLogArchivingActive())
4874 4 : return XLogArchiveCommand;
4875 : else
4876 3410 : return "(disabled)";
4877 : }
4878 :
4879 : /*
4880 : * GUC show_hook for in_hot_standby
4881 : */
4882 : const char *
4883 27166 : show_in_hot_standby(void)
4884 : {
4885 : /*
4886 : * We display the actual state based on shared memory, so that this GUC
4887 : * reports up-to-date state if examined intra-query. The underlying
4888 : * variable (in_hot_standby_guc) changes only when we transmit a new value
4889 : * to the client.
4890 : */
4891 27166 : return RecoveryInProgress() ? "on" : "off";
4892 : }
4893 :
4894 : /*
4895 : * Read the control file, set respective GUCs.
4896 : *
4897 : * This is to be called during startup, including a crash recovery cycle,
4898 : * unless in bootstrap mode, where no control file yet exists. As there's no
4899 : * usable shared memory yet (its sizing can depend on the contents of the
4900 : * control file!), first store the contents in local memory. XLOGShmemInit()
4901 : * will then copy it to shared memory later.
4902 : *
4903 : * reset just controls whether previous contents are to be expected (in the
4904 : * reset case, there's a dangling pointer into old shared memory), or not.
4905 : */
4906 : void
4907 1684 : LocalProcessControlFile(bool reset)
4908 : {
4909 : Assert(reset || ControlFile == NULL);
4910 1684 : ControlFile = palloc(sizeof(ControlFileData));
4911 1684 : ReadControlFile();
4912 1684 : }
4913 :
4914 : /*
4915 : * Get the wal_level from the control file. For a standby, this value should be
4916 : * considered as its active wal_level, because it may be different from what
4917 : * was originally configured on standby.
4918 : */
4919 : WalLevel
4920 132 : GetActiveWalLevelOnStandby(void)
4921 : {
4922 132 : return ControlFile->wal_level;
4923 : }
4924 :
4925 : /*
4926 : * Initialization of shared memory for XLOG
4927 : */
4928 : Size
4929 5544 : XLOGShmemSize(void)
4930 : {
4931 : Size size;
4932 :
4933 : /*
4934 : * If the value of wal_buffers is -1, use the preferred auto-tune value.
4935 : * This isn't an amazingly clean place to do this, but we must wait till
4936 : * NBuffers has received its final value, and must do it before using the
4937 : * value of XLOGbuffers to do anything important.
4938 : *
4939 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
4940 : * However, if the DBA explicitly set wal_buffers = -1 in the config file,
4941 : * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
4942 : * the matter with PGC_S_OVERRIDE.
4943 : */
4944 5544 : if (XLOGbuffers == -1)
4945 : {
4946 : char buf[32];
4947 :
4948 1934 : snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
4949 1934 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4950 : PGC_S_DYNAMIC_DEFAULT);
4951 1934 : if (XLOGbuffers == -1) /* failed to apply it? */
4952 0 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
4953 : PGC_S_OVERRIDE);
4954 : }
4955 : Assert(XLOGbuffers > 0);
4956 :
4957 : /* XLogCtl */
4958 5544 : size = sizeof(XLogCtlData);
4959 :
4960 : /* WAL insertion locks, plus alignment */
4961 5544 : size = add_size(size, mul_size(sizeof(WALInsertLockPadded), NUM_XLOGINSERT_LOCKS + 1));
4962 : /* xlblocks array */
4963 5544 : size = add_size(size, mul_size(sizeof(pg_atomic_uint64), XLOGbuffers));
4964 : /* extra alignment padding for XLOG I/O buffers */
4965 5544 : size = add_size(size, Max(XLOG_BLCKSZ, PG_IO_ALIGN_SIZE));
4966 : /* and the buffers themselves */
4967 5544 : size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
4968 :
4969 : /*
4970 : * Note: we don't count ControlFileData, it comes out of the "slop factor"
4971 : * added by CreateSharedMemoryAndSemaphores. This lets us use this
4972 : * routine again below to compute the actual allocation size.
4973 : */
4974 :
4975 5544 : return size;
4976 : }
4977 :
4978 : void
4979 1938 : XLOGShmemInit(void)
4980 : {
4981 : bool foundCFile,
4982 : foundXLog;
4983 : char *allocptr;
4984 : int i;
4985 : ControlFileData *localControlFile;
4986 :
4987 : #ifdef WAL_DEBUG
4988 :
4989 : /*
4990 : * Create a memory context for WAL debugging that's exempt from the normal
4991 : * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
4992 : * an allocation fails, but wal_debug is not for production use anyway.
4993 : */
4994 : if (walDebugCxt == NULL)
4995 : {
4996 : walDebugCxt = AllocSetContextCreate(TopMemoryContext,
4997 : "WAL Debug",
4998 : ALLOCSET_DEFAULT_SIZES);
4999 : MemoryContextAllowInCriticalSection(walDebugCxt, true);
5000 : }
5001 : #endif
5002 :
5003 :
5004 1938 : XLogCtl = (XLogCtlData *)
5005 1938 : ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog);
5006 :
5007 1938 : localControlFile = ControlFile;
5008 1938 : ControlFile = (ControlFileData *)
5009 1938 : ShmemInitStruct("Control File", sizeof(ControlFileData), &foundCFile);
5010 :
5011 1938 : if (foundCFile || foundXLog)
5012 : {
5013 : /* both should be present or neither */
5014 : Assert(foundCFile && foundXLog);
5015 :
5016 : /* Initialize local copy of WALInsertLocks */
5017 0 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks;
5018 :
5019 0 : if (localControlFile)
5020 0 : pfree(localControlFile);
5021 0 : return;
5022 : }
5023 1938 : memset(XLogCtl, 0, sizeof(XLogCtlData));
5024 :
5025 : /*
5026 : * Already have read control file locally, unless in bootstrap mode. Move
5027 : * contents into shared memory.
5028 : */
5029 1938 : if (localControlFile)
5030 : {
5031 1668 : memcpy(ControlFile, localControlFile, sizeof(ControlFileData));
5032 1668 : pfree(localControlFile);
5033 : }
5034 :
5035 : /*
5036 : * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
5037 : * multiple of the alignment for same, so no extra alignment padding is
5038 : * needed here.
5039 : */
5040 1938 : allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
5041 1938 : XLogCtl->xlblocks = (pg_atomic_uint64 *) allocptr;
5042 1938 : allocptr += sizeof(pg_atomic_uint64) * XLOGbuffers;
5043 :
5044 549960 : for (i = 0; i < XLOGbuffers; i++)
5045 : {
5046 548022 : pg_atomic_init_u64(&XLogCtl->xlblocks[i], InvalidXLogRecPtr);
5047 : }
5048 :
5049 : /* WAL insertion locks. Ensure they're aligned to the full padded size */
5050 1938 : allocptr += sizeof(WALInsertLockPadded) -
5051 1938 : ((uintptr_t) allocptr) % sizeof(WALInsertLockPadded);
5052 1938 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks =
5053 : (WALInsertLockPadded *) allocptr;
5054 1938 : allocptr += sizeof(WALInsertLockPadded) * NUM_XLOGINSERT_LOCKS;
5055 :
5056 17442 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
5057 : {
5058 15504 : LWLockInitialize(&WALInsertLocks[i].l.lock, LWTRANCHE_WAL_INSERT);
5059 15504 : pg_atomic_init_u64(&WALInsertLocks[i].l.insertingAt, InvalidXLogRecPtr);
5060 15504 : WALInsertLocks[i].l.lastImportantAt = InvalidXLogRecPtr;
5061 : }
5062 :
5063 : /*
5064 : * Align the start of the page buffers to a full xlog block size boundary.
5065 : * This simplifies some calculations in XLOG insertion. It is also
5066 : * required for O_DIRECT.
5067 : */
5068 1938 : allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5069 1938 : XLogCtl->pages = allocptr;
5070 1938 : memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5071 :
5072 : /*
5073 : * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5074 : * in additional info.)
5075 : */
5076 1938 : XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5077 1938 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5078 1938 : XLogCtl->InstallXLogFileSegmentActive = false;
5079 1938 : XLogCtl->WalWriterSleeping = false;
5080 :
5081 1938 : SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5082 1938 : SpinLockInit(&XLogCtl->info_lck);
5083 1938 : pg_atomic_init_u64(&XLogCtl->logInsertResult, InvalidXLogRecPtr);
5084 1938 : pg_atomic_init_u64(&XLogCtl->logWriteResult, InvalidXLogRecPtr);
5085 1938 : pg_atomic_init_u64(&XLogCtl->logFlushResult, InvalidXLogRecPtr);
5086 1938 : pg_atomic_init_u64(&XLogCtl->unloggedLSN, InvalidXLogRecPtr);
5087 : }
5088 :
5089 : /*
5090 : * This func must be called ONCE on system install. It creates pg_control
5091 : * and the initial XLOG segment.
5092 : */
5093 : void
5094 90 : BootStrapXLOG(uint32 data_checksum_version)
5095 : {
5096 : CheckPoint checkPoint;
5097 : char *buffer;
5098 : XLogPageHeader page;
5099 : XLogLongPageHeader longpage;
5100 : XLogRecord *record;
5101 : char *recptr;
5102 : uint64 sysidentifier;
5103 : struct timeval tv;
5104 : pg_crc32c crc;
5105 :
5106 : /* allow ordinary WAL segment creation, like StartupXLOG() would */
5107 90 : SetInstallXLogFileSegmentActive();
5108 :
5109 : /*
5110 : * Select a hopefully-unique system identifier code for this installation.
5111 : * We use the result of gettimeofday(), including the fractional seconds
5112 : * field, as being about as unique as we can easily get. (Think not to
5113 : * use random(), since it hasn't been seeded and there's no portable way
5114 : * to seed it other than the system clock value...) The upper half of the
5115 : * uint64 value is just the tv_sec part, while the lower half contains the
5116 : * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
5117 : * PID for a little extra uniqueness. A person knowing this encoding can
5118 : * determine the initialization time of the installation, which could
5119 : * perhaps be useful sometimes.
5120 : */
5121 90 : gettimeofday(&tv, NULL);
5122 90 : sysidentifier = ((uint64) tv.tv_sec) << 32;
5123 90 : sysidentifier |= ((uint64) tv.tv_usec) << 12;
5124 90 : sysidentifier |= getpid() & 0xFFF;
5125 :
5126 : /* page buffer must be aligned suitably for O_DIRECT */
5127 90 : buffer = (char *) palloc(XLOG_BLCKSZ + XLOG_BLCKSZ);
5128 90 : page = (XLogPageHeader) TYPEALIGN(XLOG_BLCKSZ, buffer);
5129 90 : memset(page, 0, XLOG_BLCKSZ);
5130 :
5131 : /*
5132 : * Set up information for the initial checkpoint record
5133 : *
5134 : * The initial checkpoint record is written to the beginning of the WAL
5135 : * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5136 : * used, so that we can use 0/0 to mean "before any valid WAL segment".
5137 : */
5138 90 : checkPoint.redo = wal_segment_size + SizeOfXLogLongPHD;
5139 90 : checkPoint.ThisTimeLineID = BootstrapTimeLineID;
5140 90 : checkPoint.PrevTimeLineID = BootstrapTimeLineID;
5141 90 : checkPoint.fullPageWrites = fullPageWrites;
5142 90 : checkPoint.wal_level = wal_level;
5143 : checkPoint.nextXid =
5144 90 : FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId);
5145 90 : checkPoint.nextOid = FirstGenbkiObjectId;
5146 90 : checkPoint.nextMulti = FirstMultiXactId;
5147 90 : checkPoint.nextMultiOffset = 0;
5148 90 : checkPoint.oldestXid = FirstNormalTransactionId;
5149 90 : checkPoint.oldestXidDB = Template1DbOid;
5150 90 : checkPoint.oldestMulti = FirstMultiXactId;
5151 90 : checkPoint.oldestMultiDB = Template1DbOid;
5152 90 : checkPoint.oldestCommitTsXid = InvalidTransactionId;
5153 90 : checkPoint.newestCommitTsXid = InvalidTransactionId;
5154 90 : checkPoint.time = (pg_time_t) time(NULL);
5155 90 : checkPoint.oldestActiveXid = InvalidTransactionId;
5156 :
5157 90 : TransamVariables->nextXid = checkPoint.nextXid;
5158 90 : TransamVariables->nextOid = checkPoint.nextOid;
5159 90 : TransamVariables->oidCount = 0;
5160 90 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5161 90 : AdvanceOldestClogXid(checkPoint.oldestXid);
5162 90 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5163 90 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5164 90 : SetCommitTsLimit(InvalidTransactionId, InvalidTransactionId);
5165 :
5166 : /* Set up the XLOG page header */
5167 90 : page->xlp_magic = XLOG_PAGE_MAGIC;
5168 90 : page->xlp_info = XLP_LONG_HEADER;
5169 90 : page->xlp_tli = BootstrapTimeLineID;
5170 90 : page->xlp_pageaddr = wal_segment_size;
5171 90 : longpage = (XLogLongPageHeader) page;
5172 90 : longpage->xlp_sysid = sysidentifier;
5173 90 : longpage->xlp_seg_size = wal_segment_size;
5174 90 : longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5175 :
5176 : /* Insert the initial checkpoint record */
5177 90 : recptr = ((char *) page + SizeOfXLogLongPHD);
5178 90 : record = (XLogRecord *) recptr;
5179 90 : record->xl_prev = 0;
5180 90 : record->xl_xid = InvalidTransactionId;
5181 90 : record->xl_tot_len = SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(checkPoint);
5182 90 : record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5183 90 : record->xl_rmid = RM_XLOG_ID;
5184 90 : recptr += SizeOfXLogRecord;
5185 : /* fill the XLogRecordDataHeaderShort struct */
5186 90 : *(recptr++) = (char) XLR_BLOCK_ID_DATA_SHORT;
5187 90 : *(recptr++) = sizeof(checkPoint);
5188 90 : memcpy(recptr, &checkPoint, sizeof(checkPoint));
5189 90 : recptr += sizeof(checkPoint);
5190 : Assert(recptr - (char *) record == record->xl_tot_len);
5191 :
5192 90 : INIT_CRC32C(crc);
5193 90 : COMP_CRC32C(crc, ((char *) record) + SizeOfXLogRecord, record->xl_tot_len - SizeOfXLogRecord);
5194 90 : COMP_CRC32C(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5195 90 : FIN_CRC32C(crc);
5196 90 : record->xl_crc = crc;
5197 :
5198 : /* Create first XLOG segment file */
5199 90 : openLogTLI = BootstrapTimeLineID;
5200 90 : openLogFile = XLogFileInit(1, BootstrapTimeLineID);
5201 :
5202 : /*
5203 : * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
5204 : * close the file again in a moment.
5205 : */
5206 :
5207 : /* Write the first page with the initial record */
5208 90 : errno = 0;
5209 90 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE);
5210 90 : if (write(openLogFile, page, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5211 : {
5212 : /* if write didn't set errno, assume problem is no disk space */
5213 0 : if (errno == 0)
5214 0 : errno = ENOSPC;
5215 0 : ereport(PANIC,
5216 : (errcode_for_file_access(),
5217 : errmsg("could not write bootstrap write-ahead log file: %m")));
5218 : }
5219 90 : pgstat_report_wait_end();
5220 :
5221 90 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC);
5222 90 : if (pg_fsync(openLogFile) != 0)
5223 0 : ereport(PANIC,
5224 : (errcode_for_file_access(),
5225 : errmsg("could not fsync bootstrap write-ahead log file: %m")));
5226 90 : pgstat_report_wait_end();
5227 :
5228 90 : if (close(openLogFile) != 0)
5229 0 : ereport(PANIC,
5230 : (errcode_for_file_access(),
5231 : errmsg("could not close bootstrap write-ahead log file: %m")));
5232 :
5233 90 : openLogFile = -1;
5234 :
5235 : /* Now create pg_control */
5236 90 : InitControlFile(sysidentifier, data_checksum_version);
5237 90 : ControlFile->time = checkPoint.time;
5238 90 : ControlFile->checkPoint = checkPoint.redo;
5239 90 : ControlFile->checkPointCopy = checkPoint;
5240 :
5241 : /* some additional ControlFile fields are set in WriteControlFile() */
5242 90 : WriteControlFile();
5243 :
5244 : /* Bootstrap the commit log, too */
5245 90 : BootStrapCLOG();
5246 90 : BootStrapCommitTs();
5247 90 : BootStrapSUBTRANS();
5248 90 : BootStrapMultiXact();
5249 :
5250 90 : pfree(buffer);
5251 :
5252 : /*
5253 : * Force control file to be read - in contrast to normal processing we'd
5254 : * otherwise never run the checks and GUC related initializations therein.
5255 : */
5256 90 : ReadControlFile();
5257 90 : }
5258 :
5259 : static char *
5260 1488 : str_time(pg_time_t tnow)
5261 : {
5262 1488 : char *buf = palloc(128);
5263 :
5264 1488 : pg_strftime(buf, 128,
5265 : "%Y-%m-%d %H:%M:%S %Z",
5266 1488 : pg_localtime(&tnow, log_timezone));
5267 :
5268 1488 : return buf;
5269 : }
5270 :
5271 : /*
5272 : * Initialize the first WAL segment on new timeline.
5273 : */
5274 : static void
5275 98 : XLogInitNewTimeline(TimeLineID endTLI, XLogRecPtr endOfLog, TimeLineID newTLI)
5276 : {
5277 : char xlogfname[MAXFNAMELEN];
5278 : XLogSegNo endLogSegNo;
5279 : XLogSegNo startLogSegNo;
5280 :
5281 : /* we always switch to a new timeline after archive recovery */
5282 : Assert(endTLI != newTLI);
5283 :
5284 : /*
5285 : * Update min recovery point one last time.
5286 : */
5287 98 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5288 :
5289 : /*
5290 : * Calculate the last segment on the old timeline, and the first segment
5291 : * on the new timeline. If the switch happens in the middle of a segment,
5292 : * they are the same, but if the switch happens exactly at a segment
5293 : * boundary, startLogSegNo will be endLogSegNo + 1.
5294 : */
5295 98 : XLByteToPrevSeg(endOfLog, endLogSegNo, wal_segment_size);
5296 98 : XLByteToSeg(endOfLog, startLogSegNo, wal_segment_size);
5297 :
5298 : /*
5299 : * Initialize the starting WAL segment for the new timeline. If the switch
5300 : * happens in the middle of a segment, copy data from the last WAL segment
5301 : * of the old timeline up to the switch point, to the starting WAL segment
5302 : * on the new timeline.
5303 : */
5304 98 : if (endLogSegNo == startLogSegNo)
5305 : {
5306 : /*
5307 : * Make a copy of the file on the new timeline.
5308 : *
5309 : * Writing WAL isn't allowed yet, so there are no locking
5310 : * considerations. But we should be just as tense as XLogFileInit to
5311 : * avoid emplacing a bogus file.
5312 : */
5313 76 : XLogFileCopy(newTLI, endLogSegNo, endTLI, endLogSegNo,
5314 76 : XLogSegmentOffset(endOfLog, wal_segment_size));
5315 : }
5316 : else
5317 : {
5318 : /*
5319 : * The switch happened at a segment boundary, so just create the next
5320 : * segment on the new timeline.
5321 : */
5322 : int fd;
5323 :
5324 22 : fd = XLogFileInit(startLogSegNo, newTLI);
5325 :
5326 22 : if (close(fd) != 0)
5327 : {
5328 0 : int save_errno = errno;
5329 :
5330 0 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5331 0 : errno = save_errno;
5332 0 : ereport(ERROR,
5333 : (errcode_for_file_access(),
5334 : errmsg("could not close file \"%s\": %m", xlogfname)));
5335 : }
5336 : }
5337 :
5338 : /*
5339 : * Let's just make real sure there are not .ready or .done flags posted
5340 : * for the new segment.
5341 : */
5342 98 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5343 98 : XLogArchiveCleanup(xlogfname);
5344 98 : }
5345 :
5346 : /*
5347 : * Perform cleanup actions at the conclusion of archive recovery.
5348 : */
5349 : static void
5350 98 : CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI, XLogRecPtr EndOfLog,
5351 : TimeLineID newTLI)
5352 : {
5353 : /*
5354 : * Execute the recovery_end_command, if any.
5355 : */
5356 98 : if (recoveryEndCommand && strcmp(recoveryEndCommand, "") != 0)
5357 4 : ExecuteRecoveryCommand(recoveryEndCommand,
5358 : "recovery_end_command",
5359 : true,
5360 : WAIT_EVENT_RECOVERY_END_COMMAND);
5361 :
5362 : /*
5363 : * We switched to a new timeline. Clean up segments on the old timeline.
5364 : *
5365 : * If there are any higher-numbered segments on the old timeline, remove
5366 : * them. They might contain valid WAL, but they might also be
5367 : * pre-allocated files containing garbage. In any case, they are not part
5368 : * of the new timeline's history so we don't need them.
5369 : */
5370 98 : RemoveNonParentXlogFiles(EndOfLog, newTLI);
5371 :
5372 : /*
5373 : * If the switch happened in the middle of a segment, what to do with the
5374 : * last, partial segment on the old timeline? If we don't archive it, and
5375 : * the server that created the WAL never archives it either (e.g. because
5376 : * it was hit by a meteor), it will never make it to the archive. That's
5377 : * OK from our point of view, because the new segment that we created with
5378 : * the new TLI contains all the WAL from the old timeline up to the switch
5379 : * point. But if you later try to do PITR to the "missing" WAL on the old
5380 : * timeline, recovery won't find it in the archive. It's physically
5381 : * present in the new file with new TLI, but recovery won't look there
5382 : * when it's recovering to the older timeline. On the other hand, if we
5383 : * archive the partial segment, and the original server on that timeline
5384 : * is still running and archives the completed version of the same segment
5385 : * later, it will fail. (We used to do that in 9.4 and below, and it
5386 : * caused such problems).
5387 : *
5388 : * As a compromise, we rename the last segment with the .partial suffix,
5389 : * and archive it. Archive recovery will never try to read .partial
5390 : * segments, so they will normally go unused. But in the odd PITR case,
5391 : * the administrator can copy them manually to the pg_wal directory
5392 : * (removing the suffix). They can be useful in debugging, too.
5393 : *
5394 : * If a .done or .ready file already exists for the old timeline, however,
5395 : * we had already determined that the segment is complete, so we can let
5396 : * it be archived normally. (In particular, if it was restored from the
5397 : * archive to begin with, it's expected to have a .done file).
5398 : */
5399 98 : if (XLogSegmentOffset(EndOfLog, wal_segment_size) != 0 &&
5400 : XLogArchivingActive())
5401 : {
5402 : char origfname[MAXFNAMELEN];
5403 : XLogSegNo endLogSegNo;
5404 :
5405 18 : XLByteToPrevSeg(EndOfLog, endLogSegNo, wal_segment_size);
5406 18 : XLogFileName(origfname, EndOfLogTLI, endLogSegNo, wal_segment_size);
5407 :
5408 18 : if (!XLogArchiveIsReadyOrDone(origfname))
5409 : {
5410 : char origpath[MAXPGPATH];
5411 : char partialfname[MAXFNAMELEN];
5412 : char partialpath[MAXPGPATH];
5413 :
5414 : /*
5415 : * If we're summarizing WAL, we can't rename the partial file
5416 : * until the summarizer finishes with it, else it will fail.
5417 : */
5418 10 : if (summarize_wal)
5419 2 : WaitForWalSummarization(EndOfLog);
5420 :
5421 10 : XLogFilePath(origpath, EndOfLogTLI, endLogSegNo, wal_segment_size);
5422 10 : snprintf(partialfname, MAXFNAMELEN, "%s.partial", origfname);
5423 10 : snprintf(partialpath, MAXPGPATH, "%s.partial", origpath);
5424 :
5425 : /*
5426 : * Make sure there's no .done or .ready file for the .partial
5427 : * file.
5428 : */
5429 10 : XLogArchiveCleanup(partialfname);
5430 :
5431 10 : durable_rename(origpath, partialpath, ERROR);
5432 10 : XLogArchiveNotify(partialfname);
5433 : }
5434 : }
5435 98 : }
5436 :
5437 : /*
5438 : * Check to see if required parameters are set high enough on this server
5439 : * for various aspects of recovery operation.
5440 : *
5441 : * Note that all the parameters which this function tests need to be
5442 : * listed in Administrator's Overview section in high-availability.sgml.
5443 : * If you change them, don't forget to update the list.
5444 : */
5445 : static void
5446 472 : CheckRequiredParameterValues(void)
5447 : {
5448 : /*
5449 : * For archive recovery, the WAL must be generated with at least 'replica'
5450 : * wal_level.
5451 : */
5452 472 : if (ArchiveRecoveryRequested && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5453 : {
5454 4 : ereport(FATAL,
5455 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5456 : errmsg("WAL was generated with \"wal_level=minimal\", cannot continue recovering"),
5457 : errdetail("This happens if you temporarily set \"wal_level=minimal\" on the server."),
5458 : errhint("Use a backup taken after setting \"wal_level\" to higher than \"minimal\".")));
5459 : }
5460 :
5461 : /*
5462 : * For Hot Standby, the WAL must be generated with 'replica' mode, and we
5463 : * must have at least as many backend slots as the primary.
5464 : */
5465 468 : if (ArchiveRecoveryRequested && EnableHotStandby)
5466 : {
5467 : /* We ignore autovacuum_worker_slots when we make this test. */
5468 230 : RecoveryRequiresIntParameter("max_connections",
5469 : MaxConnections,
5470 230 : ControlFile->MaxConnections);
5471 230 : RecoveryRequiresIntParameter("max_worker_processes",
5472 : max_worker_processes,
5473 230 : ControlFile->max_worker_processes);
5474 230 : RecoveryRequiresIntParameter("max_wal_senders",
5475 : max_wal_senders,
5476 230 : ControlFile->max_wal_senders);
5477 230 : RecoveryRequiresIntParameter("max_prepared_transactions",
5478 : max_prepared_xacts,
5479 230 : ControlFile->max_prepared_xacts);
5480 230 : RecoveryRequiresIntParameter("max_locks_per_transaction",
5481 : max_locks_per_xact,
5482 230 : ControlFile->max_locks_per_xact);
5483 : }
5484 468 : }
5485 :
5486 : /*
5487 : * This must be called ONCE during postmaster or standalone-backend startup
5488 : */
5489 : void
5490 1668 : StartupXLOG(void)
5491 : {
5492 : XLogCtlInsert *Insert;
5493 : CheckPoint checkPoint;
5494 : bool wasShutdown;
5495 : bool didCrash;
5496 : bool haveTblspcMap;
5497 : bool haveBackupLabel;
5498 : XLogRecPtr EndOfLog;
5499 : TimeLineID EndOfLogTLI;
5500 : TimeLineID newTLI;
5501 : bool performedWalRecovery;
5502 : EndOfWalRecoveryInfo *endOfRecoveryInfo;
5503 : XLogRecPtr abortedRecPtr;
5504 : XLogRecPtr missingContrecPtr;
5505 : TransactionId oldestActiveXID;
5506 1668 : bool promoted = false;
5507 :
5508 : /*
5509 : * We should have an aux process resource owner to use, and we should not
5510 : * be in a transaction that's installed some other resowner.
5511 : */
5512 : Assert(AuxProcessResourceOwner != NULL);
5513 : Assert(CurrentResourceOwner == NULL ||
5514 : CurrentResourceOwner == AuxProcessResourceOwner);
5515 1668 : CurrentResourceOwner = AuxProcessResourceOwner;
5516 :
5517 : /*
5518 : * Check that contents look valid.
5519 : */
5520 1668 : if (!XRecOffIsValid(ControlFile->checkPoint))
5521 0 : ereport(FATAL,
5522 : (errcode(ERRCODE_DATA_CORRUPTED),
5523 : errmsg("control file contains invalid checkpoint location")));
5524 :
5525 1668 : switch (ControlFile->state)
5526 : {
5527 1270 : case DB_SHUTDOWNED:
5528 :
5529 : /*
5530 : * This is the expected case, so don't be chatty in standalone
5531 : * mode
5532 : */
5533 1270 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
5534 : (errmsg("database system was shut down at %s",
5535 : str_time(ControlFile->time))));
5536 1270 : break;
5537 :
5538 56 : case DB_SHUTDOWNED_IN_RECOVERY:
5539 56 : ereport(LOG,
5540 : (errmsg("database system was shut down in recovery at %s",
5541 : str_time(ControlFile->time))));
5542 56 : break;
5543 :
5544 0 : case DB_SHUTDOWNING:
5545 0 : ereport(LOG,
5546 : (errmsg("database system shutdown was interrupted; last known up at %s",
5547 : str_time(ControlFile->time))));
5548 0 : break;
5549 :
5550 0 : case DB_IN_CRASH_RECOVERY:
5551 0 : ereport(LOG,
5552 : (errmsg("database system was interrupted while in recovery at %s",
5553 : str_time(ControlFile->time)),
5554 : errhint("This probably means that some data is corrupted and"
5555 : " you will have to use the last backup for recovery.")));
5556 0 : break;
5557 :
5558 12 : case DB_IN_ARCHIVE_RECOVERY:
5559 12 : ereport(LOG,
5560 : (errmsg("database system was interrupted while in recovery at log time %s",
5561 : str_time(ControlFile->checkPointCopy.time)),
5562 : errhint("If this has occurred more than once some data might be corrupted"
5563 : " and you might need to choose an earlier recovery target.")));
5564 12 : break;
5565 :
5566 330 : case DB_IN_PRODUCTION:
5567 330 : ereport(LOG,
5568 : (errmsg("database system was interrupted; last known up at %s",
5569 : str_time(ControlFile->time))));
5570 330 : break;
5571 :
5572 0 : default:
5573 0 : ereport(FATAL,
5574 : (errcode(ERRCODE_DATA_CORRUPTED),
5575 : errmsg("control file contains invalid database cluster state")));
5576 : }
5577 :
5578 : /* This is just to allow attaching to startup process with a debugger */
5579 : #ifdef XLOG_REPLAY_DELAY
5580 : if (ControlFile->state != DB_SHUTDOWNED)
5581 : pg_usleep(60000000L);
5582 : #endif
5583 :
5584 : /*
5585 : * Verify that pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
5586 : * In cases where someone has performed a copy for PITR, these directories
5587 : * may have been excluded and need to be re-created.
5588 : */
5589 1668 : ValidateXLOGDirectoryStructure();
5590 :
5591 : /* Set up timeout handler needed to report startup progress. */
5592 1668 : if (!IsBootstrapProcessingMode())
5593 1578 : RegisterTimeout(STARTUP_PROGRESS_TIMEOUT,
5594 : startup_progress_timeout_handler);
5595 :
5596 : /*----------
5597 : * If we previously crashed, perform a couple of actions:
5598 : *
5599 : * - The pg_wal directory may still include some temporary WAL segments
5600 : * used when creating a new segment, so perform some clean up to not
5601 : * bloat this path. This is done first as there is no point to sync
5602 : * this temporary data.
5603 : *
5604 : * - There might be data which we had written, intending to fsync it, but
5605 : * which we had not actually fsync'd yet. Therefore, a power failure in
5606 : * the near future might cause earlier unflushed writes to be lost, even
5607 : * though more recent data written to disk from here on would be
5608 : * persisted. To avoid that, fsync the entire data directory.
5609 : */
5610 1668 : if (ControlFile->state != DB_SHUTDOWNED &&
5611 398 : ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
5612 : {
5613 342 : RemoveTempXlogFiles();
5614 342 : SyncDataDirectory();
5615 342 : didCrash = true;
5616 : }
5617 : else
5618 1326 : didCrash = false;
5619 :
5620 : /*
5621 : * Prepare for WAL recovery if needed.
5622 : *
5623 : * InitWalRecovery analyzes the control file and the backup label file, if
5624 : * any. It updates the in-memory ControlFile buffer according to the
5625 : * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5626 : * It also applies the tablespace map file, if any.
5627 : */
5628 1668 : InitWalRecovery(ControlFile, &wasShutdown,
5629 : &haveBackupLabel, &haveTblspcMap);
5630 1668 : checkPoint = ControlFile->checkPointCopy;
5631 :
5632 : /* initialize shared memory variables from the checkpoint record */
5633 1668 : TransamVariables->nextXid = checkPoint.nextXid;
5634 1668 : TransamVariables->nextOid = checkPoint.nextOid;
5635 1668 : TransamVariables->oidCount = 0;
5636 1668 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5637 1668 : AdvanceOldestClogXid(checkPoint.oldestXid);
5638 1668 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5639 1668 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB, true);
5640 1668 : SetCommitTsLimit(checkPoint.oldestCommitTsXid,
5641 : checkPoint.newestCommitTsXid);
5642 1668 : XLogCtl->ckptFullXid = checkPoint.nextXid;
5643 :
5644 : /*
5645 : * Clear out any old relcache cache files. This is *necessary* if we do
5646 : * any WAL replay, since that would probably result in the cache files
5647 : * being out of sync with database reality. In theory we could leave them
5648 : * in place if the database had been cleanly shut down, but it seems
5649 : * safest to just remove them always and let them be rebuilt during the
5650 : * first backend startup. These files needs to be removed from all
5651 : * directories including pg_tblspc, however the symlinks are created only
5652 : * after reading tablespace_map file in case of archive recovery from
5653 : * backup, so needs to clear old relcache files here after creating
5654 : * symlinks.
5655 : */
5656 1668 : RelationCacheInitFileRemove();
5657 :
5658 : /*
5659 : * Initialize replication slots, before there's a chance to remove
5660 : * required resources.
5661 : */
5662 1668 : StartupReplicationSlots();
5663 :
5664 : /*
5665 : * Startup logical state, needs to be setup now so we have proper data
5666 : * during crash recovery.
5667 : */
5668 1668 : StartupReorderBuffer();
5669 :
5670 : /*
5671 : * Startup CLOG. This must be done after TransamVariables->nextXid has
5672 : * been initialized and before we accept connections or begin WAL replay.
5673 : */
5674 1668 : StartupCLOG();
5675 :
5676 : /*
5677 : * Startup MultiXact. We need to do this early to be able to replay
5678 : * truncations.
5679 : */
5680 1668 : StartupMultiXact();
5681 :
5682 : /*
5683 : * Ditto for commit timestamps. Activate the facility if the setting is
5684 : * enabled in the control file, as there should be no tracking of commit
5685 : * timestamps done when the setting was disabled. This facility can be
5686 : * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
5687 : */
5688 1668 : if (ControlFile->track_commit_timestamp)
5689 20 : StartupCommitTs();
5690 :
5691 : /*
5692 : * Recover knowledge about replay progress of known replication partners.
5693 : */
5694 1668 : StartupReplicationOrigin();
5695 :
5696 : /*
5697 : * Initialize unlogged LSN. On a clean shutdown, it's restored from the
5698 : * control file. On recovery, all unlogged relations are blown away, so
5699 : * the unlogged LSN counter can be reset too.
5700 : */
5701 1668 : if (ControlFile->state == DB_SHUTDOWNED)
5702 1256 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5703 1256 : ControlFile->unloggedLSN);
5704 : else
5705 412 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
5706 : FirstNormalUnloggedLSN);
5707 :
5708 : /*
5709 : * Copy any missing timeline history files between 'now' and the recovery
5710 : * target timeline from archive to pg_wal. While we don't need those files
5711 : * ourselves - the history file of the recovery target timeline covers all
5712 : * the previous timelines in the history too - a cascading standby server
5713 : * might be interested in them. Or, if you archive the WAL from this
5714 : * server to a different archive than the primary, it'd be good for all
5715 : * the history files to get archived there after failover, so that you can
5716 : * use one of the old timelines as a PITR target. Timeline history files
5717 : * are small, so it's better to copy them unnecessarily than not copy them
5718 : * and regret later.
5719 : */
5720 1668 : restoreTimeLineHistoryFiles(checkPoint.ThisTimeLineID, recoveryTargetTLI);
5721 :
5722 : /*
5723 : * Before running in recovery, scan pg_twophase and fill in its status to
5724 : * be able to work on entries generated by redo. Doing a scan before
5725 : * taking any recovery action has the merit to discard any 2PC files that
5726 : * are newer than the first record to replay, saving from any conflicts at
5727 : * replay. This avoids as well any subsequent scans when doing recovery
5728 : * of the on-disk two-phase data.
5729 : */
5730 1668 : restoreTwoPhaseData();
5731 :
5732 : /*
5733 : * When starting with crash recovery, reset pgstat data - it might not be
5734 : * valid. Otherwise restore pgstat data. It's safe to do this here,
5735 : * because postmaster will not yet have started any other processes.
5736 : *
5737 : * NB: Restoring replication slot stats relies on slot state to have
5738 : * already been restored from disk.
5739 : *
5740 : * TODO: With a bit of extra work we could just start with a pgstat file
5741 : * associated with the checkpoint redo location we're starting from.
5742 : */
5743 1668 : if (didCrash)
5744 342 : pgstat_discard_stats();
5745 : else
5746 1326 : pgstat_restore_stats(checkPoint.redo);
5747 :
5748 1668 : lastFullPageWrites = checkPoint.fullPageWrites;
5749 :
5750 1668 : RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
5751 1668 : doPageWrites = lastFullPageWrites;
5752 :
5753 : /* REDO */
5754 1668 : if (InRecovery)
5755 : {
5756 : /* Initialize state for RecoveryInProgress() */
5757 412 : SpinLockAcquire(&XLogCtl->info_lck);
5758 412 : if (InArchiveRecovery)
5759 212 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
5760 : else
5761 200 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5762 412 : SpinLockRelease(&XLogCtl->info_lck);
5763 :
5764 : /*
5765 : * Update pg_control to show that we are recovering and to show the
5766 : * selected checkpoint as the place we are starting from. We also mark
5767 : * pg_control with any minimum recovery stop point obtained from a
5768 : * backup history file.
5769 : *
5770 : * No need to hold ControlFileLock yet, we aren't up far enough.
5771 : */
5772 412 : UpdateControlFile();
5773 :
5774 : /*
5775 : * If there was a backup label file, it's done its job and the info
5776 : * has now been propagated into pg_control. We must get rid of the
5777 : * label file so that if we crash during recovery, we'll pick up at
5778 : * the latest recovery restartpoint instead of going all the way back
5779 : * to the backup start point. It seems prudent though to just rename
5780 : * the file out of the way rather than delete it completely.
5781 : */
5782 412 : if (haveBackupLabel)
5783 : {
5784 138 : unlink(BACKUP_LABEL_OLD);
5785 138 : durable_rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD, FATAL);
5786 : }
5787 :
5788 : /*
5789 : * If there was a tablespace_map file, it's done its job and the
5790 : * symlinks have been created. We must get rid of the map file so
5791 : * that if we crash during recovery, we don't create symlinks again.
5792 : * It seems prudent though to just rename the file out of the way
5793 : * rather than delete it completely.
5794 : */
5795 412 : if (haveTblspcMap)
5796 : {
5797 4 : unlink(TABLESPACE_MAP_OLD);
5798 4 : durable_rename(TABLESPACE_MAP, TABLESPACE_MAP_OLD, FATAL);
5799 : }
5800 :
5801 : /*
5802 : * Initialize our local copy of minRecoveryPoint. When doing crash
5803 : * recovery we want to replay up to the end of WAL. Particularly, in
5804 : * the case of a promoted standby minRecoveryPoint value in the
5805 : * control file is only updated after the first checkpoint. However,
5806 : * if the instance crashes before the first post-recovery checkpoint
5807 : * is completed then recovery will use a stale location causing the
5808 : * startup process to think that there are still invalid page
5809 : * references when checking for data consistency.
5810 : */
5811 412 : if (InArchiveRecovery)
5812 : {
5813 212 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
5814 212 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
5815 : }
5816 : else
5817 : {
5818 200 : LocalMinRecoveryPoint = InvalidXLogRecPtr;
5819 200 : LocalMinRecoveryPointTLI = 0;
5820 : }
5821 :
5822 : /* Check that the GUCs used to generate the WAL allow recovery */
5823 412 : CheckRequiredParameterValues();
5824 :
5825 : /*
5826 : * We're in recovery, so unlogged relations may be trashed and must be
5827 : * reset. This should be done BEFORE allowing Hot Standby
5828 : * connections, so that read-only backends don't try to read whatever
5829 : * garbage is left over from before.
5830 : */
5831 412 : ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
5832 :
5833 : /*
5834 : * Likewise, delete any saved transaction snapshot files that got left
5835 : * behind by crashed backends.
5836 : */
5837 412 : DeleteAllExportedSnapshotFiles();
5838 :
5839 : /*
5840 : * Initialize for Hot Standby, if enabled. We won't let backends in
5841 : * yet, not until we've reached the min recovery point specified in
5842 : * control file and we've established a recovery snapshot from a
5843 : * running-xacts WAL record.
5844 : */
5845 412 : if (ArchiveRecoveryRequested && EnableHotStandby)
5846 : {
5847 : TransactionId *xids;
5848 : int nxids;
5849 :
5850 200 : ereport(DEBUG1,
5851 : (errmsg_internal("initializing for hot standby")));
5852 :
5853 200 : InitRecoveryTransactionEnvironment();
5854 :
5855 200 : if (wasShutdown)
5856 48 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
5857 : else
5858 152 : oldestActiveXID = checkPoint.oldestActiveXid;
5859 : Assert(TransactionIdIsValid(oldestActiveXID));
5860 :
5861 : /* Tell procarray about the range of xids it has to deal with */
5862 200 : ProcArrayInitRecovery(XidFromFullTransactionId(TransamVariables->nextXid));
5863 :
5864 : /*
5865 : * Startup subtrans only. CLOG, MultiXact and commit timestamp
5866 : * have already been started up and other SLRUs are not maintained
5867 : * during recovery and need not be started yet.
5868 : */
5869 200 : StartupSUBTRANS(oldestActiveXID);
5870 :
5871 : /*
5872 : * If we're beginning at a shutdown checkpoint, we know that
5873 : * nothing was running on the primary at this point. So fake-up an
5874 : * empty running-xacts record and use that here and now. Recover
5875 : * additional standby state for prepared transactions.
5876 : */
5877 200 : if (wasShutdown)
5878 : {
5879 : RunningTransactionsData running;
5880 : TransactionId latestCompletedXid;
5881 :
5882 : /* Update pg_subtrans entries for any prepared transactions */
5883 48 : StandbyRecoverPreparedTransactions();
5884 :
5885 : /*
5886 : * Construct a RunningTransactions snapshot representing a
5887 : * shut down server, with only prepared transactions still
5888 : * alive. We're never overflowed at this point because all
5889 : * subxids are listed with their parent prepared transactions.
5890 : */
5891 48 : running.xcnt = nxids;
5892 48 : running.subxcnt = 0;
5893 48 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
5894 48 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
5895 48 : running.oldestRunningXid = oldestActiveXID;
5896 48 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
5897 48 : TransactionIdRetreat(latestCompletedXid);
5898 : Assert(TransactionIdIsNormal(latestCompletedXid));
5899 48 : running.latestCompletedXid = latestCompletedXid;
5900 48 : running.xids = xids;
5901 :
5902 48 : ProcArrayApplyRecoveryInfo(&running);
5903 : }
5904 : }
5905 :
5906 : /*
5907 : * We're all set for replaying the WAL now. Do it.
5908 : */
5909 412 : PerformWalRecovery();
5910 306 : performedWalRecovery = true;
5911 : }
5912 : else
5913 1256 : performedWalRecovery = false;
5914 :
5915 : /*
5916 : * Finish WAL recovery.
5917 : */
5918 1562 : endOfRecoveryInfo = FinishWalRecovery();
5919 1562 : EndOfLog = endOfRecoveryInfo->endOfLog;
5920 1562 : EndOfLogTLI = endOfRecoveryInfo->endOfLogTLI;
5921 1562 : abortedRecPtr = endOfRecoveryInfo->abortedRecPtr;
5922 1562 : missingContrecPtr = endOfRecoveryInfo->missingContrecPtr;
5923 :
5924 : /*
5925 : * Reset ps status display, so as no information related to recovery shows
5926 : * up.
5927 : */
5928 1562 : set_ps_display("");
5929 :
5930 : /*
5931 : * When recovering from a backup (we are in recovery, and archive recovery
5932 : * was requested), complain if we did not roll forward far enough to reach
5933 : * the point where the database is consistent. For regular online
5934 : * backup-from-primary, that means reaching the end-of-backup WAL record
5935 : * (at which point we reset backupStartPoint to be Invalid), for
5936 : * backup-from-replica (which can't inject records into the WAL stream),
5937 : * that point is when we reach the minRecoveryPoint in pg_control (which
5938 : * we purposefully copy last when backing up from a replica). For
5939 : * pg_rewind (which creates a backup_label with a method of "pg_rewind")
5940 : * or snapshot-style backups (which don't), backupEndRequired will be set
5941 : * to false.
5942 : *
5943 : * Note: it is indeed okay to look at the local variable
5944 : * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
5945 : * might be further ahead --- ControlFile->minRecoveryPoint cannot have
5946 : * been advanced beyond the WAL we processed.
5947 : */
5948 1562 : if (InRecovery &&
5949 306 : (EndOfLog < LocalMinRecoveryPoint ||
5950 306 : !XLogRecPtrIsInvalid(ControlFile->backupStartPoint)))
5951 : {
5952 : /*
5953 : * Ran off end of WAL before reaching end-of-backup WAL record, or
5954 : * minRecoveryPoint. That's a bad sign, indicating that you tried to
5955 : * recover from an online backup but never called pg_backup_stop(), or
5956 : * you didn't archive all the WAL needed.
5957 : */
5958 0 : if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
5959 : {
5960 0 : if (!XLogRecPtrIsInvalid(ControlFile->backupStartPoint) || ControlFile->backupEndRequired)
5961 0 : ereport(FATAL,
5962 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5963 : errmsg("WAL ends before end of online backup"),
5964 : errhint("All WAL generated while online backup was taken must be available at recovery.")));
5965 : else
5966 0 : ereport(FATAL,
5967 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5968 : errmsg("WAL ends before consistent recovery point")));
5969 : }
5970 : }
5971 :
5972 : /*
5973 : * Reset unlogged relations to the contents of their INIT fork. This is
5974 : * done AFTER recovery is complete so as to include any unlogged relations
5975 : * created during recovery, but BEFORE recovery is marked as having
5976 : * completed successfully. Otherwise we'd not retry if any of the post
5977 : * end-of-recovery steps fail.
5978 : */
5979 1562 : if (InRecovery)
5980 306 : ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
5981 :
5982 : /*
5983 : * Pre-scan prepared transactions to find out the range of XIDs present.
5984 : * This information is not quite needed yet, but it is positioned here so
5985 : * as potential problems are detected before any on-disk change is done.
5986 : */
5987 1562 : oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
5988 :
5989 : /*
5990 : * Allow ordinary WAL segment creation before possibly switching to a new
5991 : * timeline, which creates a new segment, and after the last ReadRecord().
5992 : */
5993 1562 : SetInstallXLogFileSegmentActive();
5994 :
5995 : /*
5996 : * Consider whether we need to assign a new timeline ID.
5997 : *
5998 : * If we did archive recovery, we always assign a new ID. This handles a
5999 : * couple of issues. If we stopped short of the end of WAL during
6000 : * recovery, then we are clearly generating a new timeline and must assign
6001 : * it a unique new ID. Even if we ran to the end, modifying the current
6002 : * last segment is problematic because it may result in trying to
6003 : * overwrite an already-archived copy of that segment, and we encourage
6004 : * DBAs to make their archive_commands reject that. We can dodge the
6005 : * problem by making the new active segment have a new timeline ID.
6006 : *
6007 : * In a normal crash recovery, we can just extend the timeline we were in.
6008 : */
6009 1562 : newTLI = endOfRecoveryInfo->lastRecTLI;
6010 1562 : if (ArchiveRecoveryRequested)
6011 : {
6012 98 : newTLI = findNewestTimeLine(recoveryTargetTLI) + 1;
6013 98 : ereport(LOG,
6014 : (errmsg("selected new timeline ID: %u", newTLI)));
6015 :
6016 : /*
6017 : * Make a writable copy of the last WAL segment. (Note that we also
6018 : * have a copy of the last block of the old WAL in
6019 : * endOfRecovery->lastPage; we will use that below.)
6020 : */
6021 98 : XLogInitNewTimeline(EndOfLogTLI, EndOfLog, newTLI);
6022 :
6023 : /*
6024 : * Remove the signal files out of the way, so that we don't
6025 : * accidentally re-enter archive recovery mode in a subsequent crash.
6026 : */
6027 98 : if (endOfRecoveryInfo->standby_signal_file_found)
6028 92 : durable_unlink(STANDBY_SIGNAL_FILE, FATAL);
6029 :
6030 98 : if (endOfRecoveryInfo->recovery_signal_file_found)
6031 6 : durable_unlink(RECOVERY_SIGNAL_FILE, FATAL);
6032 :
6033 : /*
6034 : * Write the timeline history file, and have it archived. After this
6035 : * point (or rather, as soon as the file is archived), the timeline
6036 : * will appear as "taken" in the WAL archive and to any standby
6037 : * servers. If we crash before actually switching to the new
6038 : * timeline, standby servers will nevertheless think that we switched
6039 : * to the new timeline, and will try to connect to the new timeline.
6040 : * To minimize the window for that, try to do as little as possible
6041 : * between here and writing the end-of-recovery record.
6042 : */
6043 98 : writeTimeLineHistory(newTLI, recoveryTargetTLI,
6044 : EndOfLog, endOfRecoveryInfo->recoveryStopReason);
6045 :
6046 98 : ereport(LOG,
6047 : (errmsg("archive recovery complete")));
6048 : }
6049 :
6050 : /* Save the selected TimeLineID in shared memory, too */
6051 1562 : SpinLockAcquire(&XLogCtl->info_lck);
6052 1562 : XLogCtl->InsertTimeLineID = newTLI;
6053 1562 : XLogCtl->PrevTimeLineID = endOfRecoveryInfo->lastRecTLI;
6054 1562 : SpinLockRelease(&XLogCtl->info_lck);
6055 :
6056 : /*
6057 : * Actually, if WAL ended in an incomplete record, skip the parts that
6058 : * made it through and start writing after the portion that persisted.
6059 : * (It's critical to first write an OVERWRITE_CONTRECORD message, which
6060 : * we'll do as soon as we're open for writing new WAL.)
6061 : */
6062 1562 : if (!XLogRecPtrIsInvalid(missingContrecPtr))
6063 : {
6064 : /*
6065 : * We should only have a missingContrecPtr if we're not switching to a
6066 : * new timeline. When a timeline switch occurs, WAL is copied from the
6067 : * old timeline to the new only up to the end of the last complete
6068 : * record, so there can't be an incomplete WAL record that we need to
6069 : * disregard.
6070 : */
6071 : Assert(newTLI == endOfRecoveryInfo->lastRecTLI);
6072 : Assert(!XLogRecPtrIsInvalid(abortedRecPtr));
6073 22 : EndOfLog = missingContrecPtr;
6074 : }
6075 :
6076 : /*
6077 : * Prepare to write WAL starting at EndOfLog location, and init xlog
6078 : * buffer cache using the block containing the last record from the
6079 : * previous incarnation.
6080 : */
6081 1562 : Insert = &XLogCtl->Insert;
6082 1562 : Insert->PrevBytePos = XLogRecPtrToBytePos(endOfRecoveryInfo->lastRec);
6083 1562 : Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
6084 :
6085 : /*
6086 : * Tricky point here: lastPage contains the *last* block that the LastRec
6087 : * record spans, not the one it starts in. The last block is indeed the
6088 : * one we want to use.
6089 : */
6090 1562 : if (EndOfLog % XLOG_BLCKSZ != 0)
6091 : {
6092 : char *page;
6093 : int len;
6094 : int firstIdx;
6095 :
6096 1502 : firstIdx = XLogRecPtrToBufIdx(EndOfLog);
6097 1502 : len = EndOfLog - endOfRecoveryInfo->lastPageBeginPtr;
6098 : Assert(len < XLOG_BLCKSZ);
6099 :
6100 : /* Copy the valid part of the last block, and zero the rest */
6101 1502 : page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
6102 1502 : memcpy(page, endOfRecoveryInfo->lastPage, len);
6103 1502 : memset(page + len, 0, XLOG_BLCKSZ - len);
6104 :
6105 1502 : pg_atomic_write_u64(&XLogCtl->xlblocks[firstIdx], endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6106 1502 : XLogCtl->InitializedUpTo = endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ;
6107 : }
6108 : else
6109 : {
6110 : /*
6111 : * There is no partial block to copy. Just set InitializedUpTo, and
6112 : * let the first attempt to insert a log record to initialize the next
6113 : * buffer.
6114 : */
6115 60 : XLogCtl->InitializedUpTo = EndOfLog;
6116 : }
6117 :
6118 : /*
6119 : * Update local and shared status. This is OK to do without any locks
6120 : * because no other process can be reading or writing WAL yet.
6121 : */
6122 1562 : LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
6123 1562 : pg_atomic_write_u64(&XLogCtl->logInsertResult, EndOfLog);
6124 1562 : pg_atomic_write_u64(&XLogCtl->logWriteResult, EndOfLog);
6125 1562 : pg_atomic_write_u64(&XLogCtl->logFlushResult, EndOfLog);
6126 1562 : XLogCtl->LogwrtRqst.Write = EndOfLog;
6127 1562 : XLogCtl->LogwrtRqst.Flush = EndOfLog;
6128 :
6129 : /*
6130 : * Preallocate additional log files, if wanted.
6131 : */
6132 1562 : PreallocXlogFiles(EndOfLog, newTLI);
6133 :
6134 : /*
6135 : * Okay, we're officially UP.
6136 : */
6137 1562 : InRecovery = false;
6138 :
6139 : /* start the archive_timeout timer and LSN running */
6140 1562 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
6141 1562 : XLogCtl->lastSegSwitchLSN = EndOfLog;
6142 :
6143 : /* also initialize latestCompletedXid, to nextXid - 1 */
6144 1562 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
6145 1562 : TransamVariables->latestCompletedXid = TransamVariables->nextXid;
6146 1562 : FullTransactionIdRetreat(&TransamVariables->latestCompletedXid);
6147 1562 : LWLockRelease(ProcArrayLock);
6148 :
6149 : /*
6150 : * Start up subtrans, if not already done for hot standby. (commit
6151 : * timestamps are started below, if necessary.)
6152 : */
6153 1562 : if (standbyState == STANDBY_DISABLED)
6154 1464 : StartupSUBTRANS(oldestActiveXID);
6155 :
6156 : /*
6157 : * Perform end of recovery actions for any SLRUs that need it.
6158 : */
6159 1562 : TrimCLOG();
6160 1562 : TrimMultiXact();
6161 :
6162 : /*
6163 : * Reload shared-memory state for prepared transactions. This needs to
6164 : * happen before renaming the last partial segment of the old timeline as
6165 : * it may be possible that we have to recover some transactions from it.
6166 : */
6167 1562 : RecoverPreparedTransactions();
6168 :
6169 : /* Shut down xlogreader */
6170 1562 : ShutdownWalRecovery();
6171 :
6172 : /* Enable WAL writes for this backend only. */
6173 1562 : LocalSetXLogInsertAllowed();
6174 :
6175 : /* If necessary, write overwrite-contrecord before doing anything else */
6176 1562 : if (!XLogRecPtrIsInvalid(abortedRecPtr))
6177 : {
6178 : Assert(!XLogRecPtrIsInvalid(missingContrecPtr));
6179 22 : CreateOverwriteContrecordRecord(abortedRecPtr, missingContrecPtr, newTLI);
6180 : }
6181 :
6182 : /*
6183 : * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
6184 : * record before resource manager writes cleanup WAL records or checkpoint
6185 : * record is written.
6186 : */
6187 1562 : Insert->fullPageWrites = lastFullPageWrites;
6188 1562 : UpdateFullPageWrites();
6189 :
6190 : /*
6191 : * Emit checkpoint or end-of-recovery record in XLOG, if required.
6192 : */
6193 1562 : if (performedWalRecovery)
6194 306 : promoted = PerformRecoveryXLogAction();
6195 :
6196 : /*
6197 : * If any of the critical GUCs have changed, log them before we allow
6198 : * backends to write WAL.
6199 : */
6200 1562 : XLogReportParameters();
6201 :
6202 : /* If this is archive recovery, perform post-recovery cleanup actions. */
6203 1562 : if (ArchiveRecoveryRequested)
6204 98 : CleanupAfterArchiveRecovery(EndOfLogTLI, EndOfLog, newTLI);
6205 :
6206 : /*
6207 : * Local WAL inserts enabled, so it's time to finish initialization of
6208 : * commit timestamp.
6209 : */
6210 1562 : CompleteCommitTsInitialization();
6211 :
6212 : /*
6213 : * All done with end-of-recovery actions.
6214 : *
6215 : * Now allow backends to write WAL and update the control file status in
6216 : * consequence. SharedRecoveryState, that controls if backends can write
6217 : * WAL, is updated while holding ControlFileLock to prevent other backends
6218 : * to look at an inconsistent state of the control file in shared memory.
6219 : * There is still a small window during which backends can write WAL and
6220 : * the control file is still referring to a system not in DB_IN_PRODUCTION
6221 : * state while looking at the on-disk control file.
6222 : *
6223 : * Also, we use info_lck to update SharedRecoveryState to ensure that
6224 : * there are no race conditions concerning visibility of other recent
6225 : * updates to shared memory.
6226 : */
6227 1562 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6228 1562 : ControlFile->state = DB_IN_PRODUCTION;
6229 :
6230 1562 : SpinLockAcquire(&XLogCtl->info_lck);
6231 1562 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_DONE;
6232 1562 : SpinLockRelease(&XLogCtl->info_lck);
6233 :
6234 1562 : UpdateControlFile();
6235 1562 : LWLockRelease(ControlFileLock);
6236 :
6237 : /*
6238 : * Shutdown the recovery environment. This must occur after
6239 : * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
6240 : * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
6241 : * any session building a snapshot will not rely on KnownAssignedXids as
6242 : * RecoveryInProgress() would return false at this stage. This is
6243 : * particularly critical for prepared 2PC transactions, that would still
6244 : * need to be included in snapshots once recovery has ended.
6245 : */
6246 1562 : if (standbyState != STANDBY_DISABLED)
6247 98 : ShutdownRecoveryTransactionEnvironment();
6248 :
6249 : /*
6250 : * If there were cascading standby servers connected to us, nudge any wal
6251 : * sender processes to notice that we've been promoted.
6252 : */
6253 1562 : WalSndWakeup(true, true);
6254 :
6255 : /*
6256 : * If this was a promotion, request an (online) checkpoint now. This isn't
6257 : * required for consistency, but the last restartpoint might be far back,
6258 : * and in case of a crash, recovering from it might take a longer than is
6259 : * appropriate now that we're not in standby mode anymore.
6260 : */
6261 1562 : if (promoted)
6262 84 : RequestCheckpoint(CHECKPOINT_FORCE);
6263 1562 : }
6264 :
6265 : /*
6266 : * Callback from PerformWalRecovery(), called when we switch from crash
6267 : * recovery to archive recovery mode. Updates the control file accordingly.
6268 : */
6269 : void
6270 4 : SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr, TimeLineID replayTLI)
6271 : {
6272 : /* initialize minRecoveryPoint to this record */
6273 4 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6274 4 : ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6275 4 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6276 : {
6277 4 : ControlFile->minRecoveryPoint = EndRecPtr;
6278 4 : ControlFile->minRecoveryPointTLI = replayTLI;
6279 : }
6280 : /* update local copy */
6281 4 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6282 4 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6283 :
6284 : /*
6285 : * The startup process can update its local copy of minRecoveryPoint from
6286 : * this point.
6287 : */
6288 4 : updateMinRecoveryPoint = true;
6289 :
6290 4 : UpdateControlFile();
6291 :
6292 : /*
6293 : * We update SharedRecoveryState while holding the lock on ControlFileLock
6294 : * so both states are consistent in shared memory.
6295 : */
6296 4 : SpinLockAcquire(&XLogCtl->info_lck);
6297 4 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6298 4 : SpinLockRelease(&XLogCtl->info_lck);
6299 :
6300 4 : LWLockRelease(ControlFileLock);
6301 4 : }
6302 :
6303 : /*
6304 : * Callback from PerformWalRecovery(), called when we reach the end of backup.
6305 : * Updates the control file accordingly.
6306 : */
6307 : void
6308 138 : ReachedEndOfBackup(XLogRecPtr EndRecPtr, TimeLineID tli)
6309 : {
6310 : /*
6311 : * We have reached the end of base backup, as indicated by pg_control. The
6312 : * data on disk is now consistent (unless minRecoveryPoint is further
6313 : * ahead, which can happen if we crashed during previous recovery). Reset
6314 : * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
6315 : * make sure we don't allow starting up at an earlier point even if
6316 : * recovery is stopped and restarted soon after this.
6317 : */
6318 138 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6319 :
6320 138 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6321 : {
6322 130 : ControlFile->minRecoveryPoint = EndRecPtr;
6323 130 : ControlFile->minRecoveryPointTLI = tli;
6324 : }
6325 :
6326 138 : ControlFile->backupStartPoint = InvalidXLogRecPtr;
6327 138 : ControlFile->backupEndPoint = InvalidXLogRecPtr;
6328 138 : ControlFile->backupEndRequired = false;
6329 138 : UpdateControlFile();
6330 :
6331 138 : LWLockRelease(ControlFileLock);
6332 138 : }
6333 :
6334 : /*
6335 : * Perform whatever XLOG actions are necessary at end of REDO.
6336 : *
6337 : * The goal here is to make sure that we'll be able to recover properly if
6338 : * we crash again. If we choose to write a checkpoint, we'll write a shutdown
6339 : * checkpoint rather than an on-line one. This is not particularly critical,
6340 : * but since we may be assigning a new TLI, using a shutdown checkpoint allows
6341 : * us to have the rule that TLI only changes in shutdown checkpoints, which
6342 : * allows some extra error checking in xlog_redo.
6343 : */
6344 : static bool
6345 306 : PerformRecoveryXLogAction(void)
6346 : {
6347 306 : bool promoted = false;
6348 :
6349 : /*
6350 : * Perform a checkpoint to update all our recovery activity to disk.
6351 : *
6352 : * Note that we write a shutdown checkpoint rather than an on-line one.
6353 : * This is not particularly critical, but since we may be assigning a new
6354 : * TLI, using a shutdown checkpoint allows us to have the rule that TLI
6355 : * only changes in shutdown checkpoints, which allows some extra error
6356 : * checking in xlog_redo.
6357 : *
6358 : * In promotion, only create a lightweight end-of-recovery record instead
6359 : * of a full checkpoint. A checkpoint is requested later, after we're
6360 : * fully out of recovery mode and already accepting queries.
6361 : */
6362 404 : if (ArchiveRecoveryRequested && IsUnderPostmaster &&
6363 98 : PromoteIsTriggered())
6364 : {
6365 84 : promoted = true;
6366 :
6367 : /*
6368 : * Insert a special WAL record to mark the end of recovery, since we
6369 : * aren't doing a checkpoint. That means that the checkpointer process
6370 : * may likely be in the middle of a time-smoothed restartpoint and
6371 : * could continue to be for minutes after this. That sounds strange,
6372 : * but the effect is roughly the same and it would be stranger to try
6373 : * to come out of the restartpoint and then checkpoint. We request a
6374 : * checkpoint later anyway, just for safety.
6375 : */
6376 84 : CreateEndOfRecoveryRecord();
6377 : }
6378 : else
6379 : {
6380 222 : RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
6381 : CHECKPOINT_IMMEDIATE |
6382 : CHECKPOINT_WAIT);
6383 : }
6384 :
6385 306 : return promoted;
6386 : }
6387 :
6388 : /*
6389 : * Is the system still in recovery?
6390 : *
6391 : * Unlike testing InRecovery, this works in any process that's connected to
6392 : * shared memory.
6393 : */
6394 : bool
6395 167399140 : RecoveryInProgress(void)
6396 : {
6397 : /*
6398 : * We check shared state each time only until we leave recovery mode. We
6399 : * can't re-enter recovery, so there's no need to keep checking after the
6400 : * shared variable has once been seen false.
6401 : */
6402 167399140 : if (!LocalRecoveryInProgress)
6403 162245280 : return false;
6404 : else
6405 : {
6406 : /*
6407 : * use volatile pointer to make sure we make a fresh read of the
6408 : * shared variable.
6409 : */
6410 5153860 : volatile XLogCtlData *xlogctl = XLogCtl;
6411 :
6412 5153860 : LocalRecoveryInProgress = (xlogctl->SharedRecoveryState != RECOVERY_STATE_DONE);
6413 :
6414 : /*
6415 : * Note: We don't need a memory barrier when we're still in recovery.
6416 : * We might exit recovery immediately after return, so the caller
6417 : * can't rely on 'true' meaning that we're still in recovery anyway.
6418 : */
6419 :
6420 5153860 : return LocalRecoveryInProgress;
6421 : }
6422 : }
6423 :
6424 : /*
6425 : * Returns current recovery state from shared memory.
6426 : *
6427 : * This returned state is kept consistent with the contents of the control
6428 : * file. See details about the possible values of RecoveryState in xlog.h.
6429 : */
6430 : RecoveryState
6431 114 : GetRecoveryState(void)
6432 : {
6433 : RecoveryState retval;
6434 :
6435 114 : SpinLockAcquire(&XLogCtl->info_lck);
6436 114 : retval = XLogCtl->SharedRecoveryState;
6437 114 : SpinLockRelease(&XLogCtl->info_lck);
6438 :
6439 114 : return retval;
6440 : }
6441 :
6442 : /*
6443 : * Is this process allowed to insert new WAL records?
6444 : *
6445 : * Ordinarily this is essentially equivalent to !RecoveryInProgress().
6446 : * But we also have provisions for forcing the result "true" or "false"
6447 : * within specific processes regardless of the global state.
6448 : */
6449 : bool
6450 58337426 : XLogInsertAllowed(void)
6451 : {
6452 : /*
6453 : * If value is "unconditionally true" or "unconditionally false", just
6454 : * return it. This provides the normal fast path once recovery is known
6455 : * done.
6456 : */
6457 58337426 : if (LocalXLogInsertAllowed >= 0)
6458 58134616 : return (bool) LocalXLogInsertAllowed;
6459 :
6460 : /*
6461 : * Else, must check to see if we're still in recovery.
6462 : */
6463 202810 : if (RecoveryInProgress())
6464 187934 : return false;
6465 :
6466 : /*
6467 : * On exit from recovery, reset to "unconditionally true", since there is
6468 : * no need to keep checking.
6469 : */
6470 14876 : LocalXLogInsertAllowed = 1;
6471 14876 : return true;
6472 : }
6473 :
6474 : /*
6475 : * Make XLogInsertAllowed() return true in the current process only.
6476 : *
6477 : * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
6478 : * and even call LocalSetXLogInsertAllowed() again after that.
6479 : *
6480 : * Returns the previous value of LocalXLogInsertAllowed.
6481 : */
6482 : static int
6483 1620 : LocalSetXLogInsertAllowed(void)
6484 : {
6485 1620 : int oldXLogAllowed = LocalXLogInsertAllowed;
6486 :
6487 1620 : LocalXLogInsertAllowed = 1;
6488 :
6489 1620 : return oldXLogAllowed;
6490 : }
6491 :
6492 : /*
6493 : * Return the current Redo pointer from shared memory.
6494 : *
6495 : * As a side-effect, the local RedoRecPtr copy is updated.
6496 : */
6497 : XLogRecPtr
6498 611056 : GetRedoRecPtr(void)
6499 : {
6500 : XLogRecPtr ptr;
6501 :
6502 : /*
6503 : * The possibly not up-to-date copy in XlogCtl is enough. Even if we
6504 : * grabbed a WAL insertion lock to read the authoritative value in
6505 : * Insert->RedoRecPtr, someone might update it just after we've released
6506 : * the lock.
6507 : */
6508 611056 : SpinLockAcquire(&XLogCtl->info_lck);
6509 611056 : ptr = XLogCtl->RedoRecPtr;
6510 611056 : SpinLockRelease(&XLogCtl->info_lck);
6511 :
6512 611056 : if (RedoRecPtr < ptr)
6513 2618 : RedoRecPtr = ptr;
6514 :
6515 611056 : return RedoRecPtr;
6516 : }
6517 :
6518 : /*
6519 : * Return information needed to decide whether a modified block needs a
6520 : * full-page image to be included in the WAL record.
6521 : *
6522 : * The returned values are cached copies from backend-private memory, and
6523 : * possibly out-of-date or, indeed, uninitialized, in which case they will
6524 : * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6525 : * re-check them against up-to-date values, while holding the WAL insert lock.
6526 : */
6527 : void
6528 28277628 : GetFullPageWriteInfo(XLogRecPtr *RedoRecPtr_p, bool *doPageWrites_p)
6529 : {
6530 28277628 : *RedoRecPtr_p = RedoRecPtr;
6531 28277628 : *doPageWrites_p = doPageWrites;
6532 28277628 : }
6533 :
6534 : /*
6535 : * GetInsertRecPtr -- Returns the current insert position.
6536 : *
6537 : * NOTE: The value *actually* returned is the position of the last full
6538 : * xlog page. It lags behind the real insert position by at most 1 page.
6539 : * For that, we don't need to scan through WAL insertion locks, and an
6540 : * approximation is enough for the current usage of this function.
6541 : */
6542 : XLogRecPtr
6543 11960 : GetInsertRecPtr(void)
6544 : {
6545 : XLogRecPtr recptr;
6546 :
6547 11960 : SpinLockAcquire(&XLogCtl->info_lck);
6548 11960 : recptr = XLogCtl->LogwrtRqst.Write;
6549 11960 : SpinLockRelease(&XLogCtl->info_lck);
6550 :
6551 11960 : return recptr;
6552 : }
6553 :
6554 : /*
6555 : * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6556 : * position known to be fsync'd to disk. This should only be used on a
6557 : * system that is known not to be in recovery.
6558 : */
6559 : XLogRecPtr
6560 334728 : GetFlushRecPtr(TimeLineID *insertTLI)
6561 : {
6562 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6563 :
6564 334728 : RefreshXLogWriteResult(LogwrtResult);
6565 :
6566 : /*
6567 : * If we're writing and flushing WAL, the time line can't be changing, so
6568 : * no lock is required.
6569 : */
6570 334728 : if (insertTLI)
6571 48898 : *insertTLI = XLogCtl->InsertTimeLineID;
6572 :
6573 334728 : return LogwrtResult.Flush;
6574 : }
6575 :
6576 : /*
6577 : * GetWALInsertionTimeLine -- Returns the current timeline of a system that
6578 : * is not in recovery.
6579 : */
6580 : TimeLineID
6581 208328 : GetWALInsertionTimeLine(void)
6582 : {
6583 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6584 :
6585 : /* Since the value can't be changing, no lock is required. */
6586 208328 : return XLogCtl->InsertTimeLineID;
6587 : }
6588 :
6589 : /*
6590 : * GetWALInsertionTimeLineIfSet -- If the system is not in recovery, returns
6591 : * the WAL insertion timeline; else, returns 0. Wherever possible, use
6592 : * GetWALInsertionTimeLine() instead, since it's cheaper. Note that this
6593 : * function decides recovery has ended as soon as the insert TLI is set, which
6594 : * happens before we set XLogCtl->SharedRecoveryState to RECOVERY_STATE_DONE.
6595 : */
6596 : TimeLineID
6597 0 : GetWALInsertionTimeLineIfSet(void)
6598 : {
6599 : TimeLineID insertTLI;
6600 :
6601 0 : SpinLockAcquire(&XLogCtl->info_lck);
6602 0 : insertTLI = XLogCtl->InsertTimeLineID;
6603 0 : SpinLockRelease(&XLogCtl->info_lck);
6604 :
6605 0 : return insertTLI;
6606 : }
6607 :
6608 : /*
6609 : * GetLastImportantRecPtr -- Returns the LSN of the last important record
6610 : * inserted. All records not explicitly marked as unimportant are considered
6611 : * important.
6612 : *
6613 : * The LSN is determined by computing the maximum of
6614 : * WALInsertLocks[i].lastImportantAt.
6615 : */
6616 : XLogRecPtr
6617 2190 : GetLastImportantRecPtr(void)
6618 : {
6619 2190 : XLogRecPtr res = InvalidXLogRecPtr;
6620 : int i;
6621 :
6622 19710 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
6623 : {
6624 : XLogRecPtr last_important;
6625 :
6626 : /*
6627 : * Need to take a lock to prevent torn reads of the LSN, which are
6628 : * possible on some of the supported platforms. WAL insert locks only
6629 : * support exclusive mode, so we have to use that.
6630 : */
6631 17520 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
6632 17520 : last_important = WALInsertLocks[i].l.lastImportantAt;
6633 17520 : LWLockRelease(&WALInsertLocks[i].l.lock);
6634 :
6635 17520 : if (res < last_important)
6636 4062 : res = last_important;
6637 : }
6638 :
6639 2190 : return res;
6640 : }
6641 :
6642 : /*
6643 : * Get the time and LSN of the last xlog segment switch
6644 : */
6645 : pg_time_t
6646 0 : GetLastSegSwitchData(XLogRecPtr *lastSwitchLSN)
6647 : {
6648 : pg_time_t result;
6649 :
6650 : /* Need WALWriteLock, but shared lock is sufficient */
6651 0 : LWLockAcquire(WALWriteLock, LW_SHARED);
6652 0 : result = XLogCtl->lastSegSwitchTime;
6653 0 : *lastSwitchLSN = XLogCtl->lastSegSwitchLSN;
6654 0 : LWLockRelease(WALWriteLock);
6655 :
6656 0 : return result;
6657 : }
6658 :
6659 : /*
6660 : * This must be called ONCE during postmaster or standalone-backend shutdown
6661 : */
6662 : void
6663 1102 : ShutdownXLOG(int code, Datum arg)
6664 : {
6665 : /*
6666 : * We should have an aux process resource owner to use, and we should not
6667 : * be in a transaction that's installed some other resowner.
6668 : */
6669 : Assert(AuxProcessResourceOwner != NULL);
6670 : Assert(CurrentResourceOwner == NULL ||
6671 : CurrentResourceOwner == AuxProcessResourceOwner);
6672 1102 : CurrentResourceOwner = AuxProcessResourceOwner;
6673 :
6674 : /* Don't be chatty in standalone mode */
6675 1102 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
6676 : (errmsg("shutting down")));
6677 :
6678 : /*
6679 : * Signal walsenders to move to stopping state.
6680 : */
6681 1102 : WalSndInitStopping();
6682 :
6683 : /*
6684 : * Wait for WAL senders to be in stopping state. This prevents commands
6685 : * from writing new WAL.
6686 : */
6687 1102 : WalSndWaitStopping();
6688 :
6689 1102 : if (RecoveryInProgress())
6690 100 : CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_IMMEDIATE);
6691 : else
6692 : {
6693 : /*
6694 : * If archiving is enabled, rotate the last XLOG file so that all the
6695 : * remaining records are archived (postmaster wakes up the archiver
6696 : * process one more time at the end of shutdown). The checkpoint
6697 : * record will go to the next XLOG file and won't be archived (yet).
6698 : */
6699 1002 : if (XLogArchivingActive())
6700 26 : RequestXLogSwitch(false);
6701 :
6702 1002 : CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_IMMEDIATE);
6703 : }
6704 1102 : }
6705 :
6706 : /*
6707 : * Log start of a checkpoint.
6708 : */
6709 : static void
6710 1982 : LogCheckpointStart(int flags, bool restartpoint)
6711 : {
6712 1982 : if (restartpoint)
6713 372 : ereport(LOG,
6714 : /* translator: the placeholders show checkpoint options */
6715 : (errmsg("restartpoint starting:%s%s%s%s%s%s%s%s",
6716 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6717 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6718 : (flags & CHECKPOINT_IMMEDIATE) ? " immediate" : "",
6719 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6720 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6721 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6722 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6723 : (flags & CHECKPOINT_FLUSH_ALL) ? " flush-all" : "")));
6724 : else
6725 1610 : ereport(LOG,
6726 : /* translator: the placeholders show checkpoint options */
6727 : (errmsg("checkpoint starting:%s%s%s%s%s%s%s%s",
6728 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
6729 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
6730 : (flags & CHECKPOINT_IMMEDIATE) ? " immediate" : "",
6731 : (flags & CHECKPOINT_FORCE) ? " force" : "",
6732 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
6733 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
6734 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
6735 : (flags & CHECKPOINT_FLUSH_ALL) ? " flush-all" : "")));
6736 1982 : }
6737 :
6738 : /*
6739 : * Log end of a checkpoint.
6740 : */
6741 : static void
6742 2506 : LogCheckpointEnd(bool restartpoint)
6743 : {
6744 : long write_msecs,
6745 : sync_msecs,
6746 : total_msecs,
6747 : longest_msecs,
6748 : average_msecs;
6749 : uint64 average_sync_time;
6750 :
6751 2506 : CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
6752 :
6753 2506 : write_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_write_t,
6754 : CheckpointStats.ckpt_sync_t);
6755 :
6756 2506 : sync_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_sync_t,
6757 : CheckpointStats.ckpt_sync_end_t);
6758 :
6759 : /* Accumulate checkpoint timing summary data, in milliseconds. */
6760 2506 : PendingCheckpointerStats.write_time += write_msecs;
6761 2506 : PendingCheckpointerStats.sync_time += sync_msecs;
6762 :
6763 : /*
6764 : * All of the published timing statistics are accounted for. Only
6765 : * continue if a log message is to be written.
6766 : */
6767 2506 : if (!log_checkpoints)
6768 524 : return;
6769 :
6770 1982 : total_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_start_t,
6771 : CheckpointStats.ckpt_end_t);
6772 :
6773 : /*
6774 : * Timing values returned from CheckpointStats are in microseconds.
6775 : * Convert to milliseconds for consistent printing.
6776 : */
6777 1982 : longest_msecs = (long) ((CheckpointStats.ckpt_longest_sync + 999) / 1000);
6778 :
6779 1982 : average_sync_time = 0;
6780 1982 : if (CheckpointStats.ckpt_sync_rels > 0)
6781 0 : average_sync_time = CheckpointStats.ckpt_agg_sync_time /
6782 0 : CheckpointStats.ckpt_sync_rels;
6783 1982 : average_msecs = (long) ((average_sync_time + 999) / 1000);
6784 :
6785 : /*
6786 : * ControlFileLock is not required to see ControlFile->checkPoint and
6787 : * ->checkPointCopy here as we are the only updator of those variables at
6788 : * this moment.
6789 : */
6790 1982 : if (restartpoint)
6791 372 : ereport(LOG,
6792 : (errmsg("restartpoint complete: wrote %d buffers (%.1f%%), "
6793 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6794 : "%d removed, %d recycled; write=%ld.%03d s, "
6795 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6796 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6797 : "estimate=%d kB; lsn=%X/%X, redo lsn=%X/%X",
6798 : CheckpointStats.ckpt_bufs_written,
6799 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6800 : CheckpointStats.ckpt_slru_written,
6801 : CheckpointStats.ckpt_segs_added,
6802 : CheckpointStats.ckpt_segs_removed,
6803 : CheckpointStats.ckpt_segs_recycled,
6804 : write_msecs / 1000, (int) (write_msecs % 1000),
6805 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6806 : total_msecs / 1000, (int) (total_msecs % 1000),
6807 : CheckpointStats.ckpt_sync_rels,
6808 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6809 : average_msecs / 1000, (int) (average_msecs % 1000),
6810 : (int) (PrevCheckPointDistance / 1024.0),
6811 : (int) (CheckPointDistanceEstimate / 1024.0),
6812 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6813 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6814 : else
6815 1610 : ereport(LOG,
6816 : (errmsg("checkpoint complete: wrote %d buffers (%.1f%%), "
6817 : "wrote %d SLRU buffers; %d WAL file(s) added, "
6818 : "%d removed, %d recycled; write=%ld.%03d s, "
6819 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
6820 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
6821 : "estimate=%d kB; lsn=%X/%X, redo lsn=%X/%X",
6822 : CheckpointStats.ckpt_bufs_written,
6823 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
6824 : CheckpointStats.ckpt_slru_written,
6825 : CheckpointStats.ckpt_segs_added,
6826 : CheckpointStats.ckpt_segs_removed,
6827 : CheckpointStats.ckpt_segs_recycled,
6828 : write_msecs / 1000, (int) (write_msecs % 1000),
6829 : sync_msecs / 1000, (int) (sync_msecs % 1000),
6830 : total_msecs / 1000, (int) (total_msecs % 1000),
6831 : CheckpointStats.ckpt_sync_rels,
6832 : longest_msecs / 1000, (int) (longest_msecs % 1000),
6833 : average_msecs / 1000, (int) (average_msecs % 1000),
6834 : (int) (PrevCheckPointDistance / 1024.0),
6835 : (int) (CheckPointDistanceEstimate / 1024.0),
6836 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
6837 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
6838 : }
6839 :
6840 : /*
6841 : * Update the estimate of distance between checkpoints.
6842 : *
6843 : * The estimate is used to calculate the number of WAL segments to keep
6844 : * preallocated, see XLOGfileslop().
6845 : */
6846 : static void
6847 2506 : UpdateCheckPointDistanceEstimate(uint64 nbytes)
6848 : {
6849 : /*
6850 : * To estimate the number of segments consumed between checkpoints, keep a
6851 : * moving average of the amount of WAL generated in previous checkpoint
6852 : * cycles. However, if the load is bursty, with quiet periods and busy
6853 : * periods, we want to cater for the peak load. So instead of a plain
6854 : * moving average, let the average decline slowly if the previous cycle
6855 : * used less WAL than estimated, but bump it up immediately if it used
6856 : * more.
6857 : *
6858 : * When checkpoints are triggered by max_wal_size, this should converge to
6859 : * CheckpointSegments * wal_segment_size,
6860 : *
6861 : * Note: This doesn't pay any attention to what caused the checkpoint.
6862 : * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
6863 : * starting a base backup, are counted the same as those created
6864 : * automatically. The slow-decline will largely mask them out, if they are
6865 : * not frequent. If they are frequent, it seems reasonable to count them
6866 : * in as any others; if you issue a manual checkpoint every 5 minutes and
6867 : * never let a timed checkpoint happen, it makes sense to base the
6868 : * preallocation on that 5 minute interval rather than whatever
6869 : * checkpoint_timeout is set to.
6870 : */
6871 2506 : PrevCheckPointDistance = nbytes;
6872 2506 : if (CheckPointDistanceEstimate < nbytes)
6873 1254 : CheckPointDistanceEstimate = nbytes;
6874 : else
6875 1252 : CheckPointDistanceEstimate =
6876 1252 : (0.90 * CheckPointDistanceEstimate + 0.10 * (double) nbytes);
6877 2506 : }
6878 :
6879 : /*
6880 : * Update the ps display for a process running a checkpoint. Note that
6881 : * this routine should not do any allocations so as it can be called
6882 : * from a critical section.
6883 : */
6884 : static void
6885 5012 : update_checkpoint_display(int flags, bool restartpoint, bool reset)
6886 : {
6887 : /*
6888 : * The status is reported only for end-of-recovery and shutdown
6889 : * checkpoints or shutdown restartpoints. Updating the ps display is
6890 : * useful in those situations as it may not be possible to rely on
6891 : * pg_stat_activity to see the status of the checkpointer or the startup
6892 : * process.
6893 : */
6894 5012 : if ((flags & (CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IS_SHUTDOWN)) == 0)
6895 2820 : return;
6896 :
6897 2192 : if (reset)
6898 1096 : set_ps_display("");
6899 : else
6900 : {
6901 : char activitymsg[128];
6902 :
6903 3288 : snprintf(activitymsg, sizeof(activitymsg), "performing %s%s%s",
6904 1096 : (flags & CHECKPOINT_END_OF_RECOVERY) ? "end-of-recovery " : "",
6905 1096 : (flags & CHECKPOINT_IS_SHUTDOWN) ? "shutdown " : "",
6906 : restartpoint ? "restartpoint" : "checkpoint");
6907 1096 : set_ps_display(activitymsg);
6908 : }
6909 : }
6910 :
6911 :
6912 : /*
6913 : * Perform a checkpoint --- either during shutdown, or on-the-fly
6914 : *
6915 : * flags is a bitwise OR of the following:
6916 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
6917 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
6918 : * CHECKPOINT_IMMEDIATE: finish the checkpoint ASAP,
6919 : * ignoring checkpoint_completion_target parameter.
6920 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
6921 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
6922 : * CHECKPOINT_END_OF_RECOVERY).
6923 : * CHECKPOINT_FLUSH_ALL: also flush buffers of unlogged tables.
6924 : *
6925 : * Note: flags contains other bits, of interest here only for logging purposes.
6926 : * In particular note that this routine is synchronous and does not pay
6927 : * attention to CHECKPOINT_WAIT.
6928 : *
6929 : * If !shutdown then we are writing an online checkpoint. An XLOG_CHECKPOINT_REDO
6930 : * record is inserted into WAL at the logical location of the checkpoint, before
6931 : * flushing anything to disk, and when the checkpoint is eventually completed,
6932 : * and it is from this point that WAL replay will begin in the case of a recovery
6933 : * from this checkpoint. Once everything is written to disk, an
6934 : * XLOG_CHECKPOINT_ONLINE record is written to complete the checkpoint, and
6935 : * points back to the earlier XLOG_CHECKPOINT_REDO record. This mechanism allows
6936 : * other write-ahead log records to be written while the checkpoint is in
6937 : * progress, but we must be very careful about order of operations. This function
6938 : * may take many minutes to execute on a busy system.
6939 : *
6940 : * On the other hand, when shutdown is true, concurrent insertion into the
6941 : * write-ahead log is impossible, so there is no need for two separate records.
6942 : * In this case, we only insert an XLOG_CHECKPOINT_SHUTDOWN record, and it's
6943 : * both the record marking the completion of the checkpoint and the location
6944 : * from which WAL replay would begin if needed.
6945 : *
6946 : * Returns true if a new checkpoint was performed, or false if it was skipped
6947 : * because the system was idle.
6948 : */
6949 : bool
6950 2134 : CreateCheckPoint(int flags)
6951 : {
6952 : bool shutdown;
6953 : CheckPoint checkPoint;
6954 : XLogRecPtr recptr;
6955 : XLogSegNo _logSegNo;
6956 2134 : XLogCtlInsert *Insert = &XLogCtl->Insert;
6957 : uint32 freespace;
6958 : XLogRecPtr PriorRedoPtr;
6959 : XLogRecPtr last_important_lsn;
6960 : VirtualTransactionId *vxids;
6961 : int nvxids;
6962 2134 : int oldXLogAllowed = 0;
6963 :
6964 : /*
6965 : * An end-of-recovery checkpoint is really a shutdown checkpoint, just
6966 : * issued at a different time.
6967 : */
6968 2134 : if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
6969 1060 : shutdown = true;
6970 : else
6971 1074 : shutdown = false;
6972 :
6973 : /* sanity check */
6974 2134 : if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
6975 0 : elog(ERROR, "can't create a checkpoint during recovery");
6976 :
6977 : /*
6978 : * Prepare to accumulate statistics.
6979 : *
6980 : * Note: because it is possible for log_checkpoints to change while a
6981 : * checkpoint proceeds, we always accumulate stats, even if
6982 : * log_checkpoints is currently off.
6983 : */
6984 23474 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
6985 2134 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
6986 :
6987 : /*
6988 : * Let smgr prepare for checkpoint; this has to happen outside the
6989 : * critical section and before we determine the REDO pointer. Note that
6990 : * smgr must not do anything that'd have to be undone if we decide no
6991 : * checkpoint is needed.
6992 : */
6993 2134 : SyncPreCheckpoint();
6994 :
6995 : /*
6996 : * Use a critical section to force system panic if we have trouble.
6997 : */
6998 2134 : START_CRIT_SECTION();
6999 :
7000 2134 : if (shutdown)
7001 : {
7002 1060 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7003 1060 : ControlFile->state = DB_SHUTDOWNING;
7004 1060 : UpdateControlFile();
7005 1060 : LWLockRelease(ControlFileLock);
7006 : }
7007 :
7008 : /* Begin filling in the checkpoint WAL record */
7009 25608 : MemSet(&checkPoint, 0, sizeof(checkPoint));
7010 2134 : checkPoint.time = (pg_time_t) time(NULL);
7011 :
7012 : /*
7013 : * For Hot Standby, derive the oldestActiveXid before we fix the redo
7014 : * pointer. This allows us to begin accumulating changes to assemble our
7015 : * starting snapshot of locks and transactions.
7016 : */
7017 2134 : if (!shutdown && XLogStandbyInfoActive())
7018 1004 : checkPoint.oldestActiveXid = GetOldestActiveTransactionId();
7019 : else
7020 1130 : checkPoint.oldestActiveXid = InvalidTransactionId;
7021 :
7022 : /*
7023 : * Get location of last important record before acquiring insert locks (as
7024 : * GetLastImportantRecPtr() also locks WAL locks).
7025 : */
7026 2134 : last_important_lsn = GetLastImportantRecPtr();
7027 :
7028 : /*
7029 : * If this isn't a shutdown or forced checkpoint, and if there has been no
7030 : * WAL activity requiring a checkpoint, skip it. The idea here is to
7031 : * avoid inserting duplicate checkpoints when the system is idle.
7032 : */
7033 2134 : if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
7034 : CHECKPOINT_FORCE)) == 0)
7035 : {
7036 368 : if (last_important_lsn == ControlFile->checkPoint)
7037 : {
7038 0 : END_CRIT_SECTION();
7039 0 : ereport(DEBUG1,
7040 : (errmsg_internal("checkpoint skipped because system is idle")));
7041 0 : return false;
7042 : }
7043 : }
7044 :
7045 : /*
7046 : * An end-of-recovery checkpoint is created before anyone is allowed to
7047 : * write WAL. To allow us to write the checkpoint record, temporarily
7048 : * enable XLogInsertAllowed.
7049 : */
7050 2134 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7051 58 : oldXLogAllowed = LocalSetXLogInsertAllowed();
7052 :
7053 2134 : checkPoint.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7054 2134 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7055 58 : checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7056 : else
7057 2076 : checkPoint.PrevTimeLineID = checkPoint.ThisTimeLineID;
7058 :
7059 : /*
7060 : * We must block concurrent insertions while examining insert state.
7061 : */
7062 2134 : WALInsertLockAcquireExclusive();
7063 :
7064 2134 : checkPoint.fullPageWrites = Insert->fullPageWrites;
7065 2134 : checkPoint.wal_level = wal_level;
7066 :
7067 2134 : if (shutdown)
7068 : {
7069 1060 : XLogRecPtr curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
7070 :
7071 : /*
7072 : * Compute new REDO record ptr = location of next XLOG record.
7073 : *
7074 : * Since this is a shutdown checkpoint, there can't be any concurrent
7075 : * WAL insertion.
7076 : */
7077 1060 : freespace = INSERT_FREESPACE(curInsert);
7078 1060 : if (freespace == 0)
7079 : {
7080 0 : if (XLogSegmentOffset(curInsert, wal_segment_size) == 0)
7081 0 : curInsert += SizeOfXLogLongPHD;
7082 : else
7083 0 : curInsert += SizeOfXLogShortPHD;
7084 : }
7085 1060 : checkPoint.redo = curInsert;
7086 :
7087 : /*
7088 : * Here we update the shared RedoRecPtr for future XLogInsert calls;
7089 : * this must be done while holding all the insertion locks.
7090 : *
7091 : * Note: if we fail to complete the checkpoint, RedoRecPtr will be
7092 : * left pointing past where it really needs to point. This is okay;
7093 : * the only consequence is that XLogInsert might back up whole buffers
7094 : * that it didn't really need to. We can't postpone advancing
7095 : * RedoRecPtr because XLogInserts that happen while we are dumping
7096 : * buffers must assume that their buffer changes are not included in
7097 : * the checkpoint.
7098 : */
7099 1060 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
7100 : }
7101 :
7102 : /*
7103 : * Now we can release the WAL insertion locks, allowing other xacts to
7104 : * proceed while we are flushing disk buffers.
7105 : */
7106 2134 : WALInsertLockRelease();
7107 :
7108 : /*
7109 : * If this is an online checkpoint, we have not yet determined the redo
7110 : * point. We do so now by inserting the special XLOG_CHECKPOINT_REDO
7111 : * record; the LSN at which it starts becomes the new redo pointer. We
7112 : * don't do this for a shutdown checkpoint, because in that case no WAL
7113 : * can be written between the redo point and the insertion of the
7114 : * checkpoint record itself, so the checkpoint record itself serves to
7115 : * mark the redo point.
7116 : */
7117 2134 : if (!shutdown)
7118 : {
7119 : /* Include WAL level in record for WAL summarizer's benefit. */
7120 1074 : XLogBeginInsert();
7121 1074 : XLogRegisterData(&wal_level, sizeof(wal_level));
7122 1074 : (void) XLogInsert(RM_XLOG_ID, XLOG_CHECKPOINT_REDO);
7123 :
7124 : /*
7125 : * XLogInsertRecord will have updated XLogCtl->Insert.RedoRecPtr in
7126 : * shared memory and RedoRecPtr in backend-local memory, but we need
7127 : * to copy that into the record that will be inserted when the
7128 : * checkpoint is complete.
7129 : */
7130 1074 : checkPoint.redo = RedoRecPtr;
7131 : }
7132 :
7133 : /* Update the info_lck-protected copy of RedoRecPtr as well */
7134 2134 : SpinLockAcquire(&XLogCtl->info_lck);
7135 2134 : XLogCtl->RedoRecPtr = checkPoint.redo;
7136 2134 : SpinLockRelease(&XLogCtl->info_lck);
7137 :
7138 : /*
7139 : * If enabled, log checkpoint start. We postpone this until now so as not
7140 : * to log anything if we decided to skip the checkpoint.
7141 : */
7142 2134 : if (log_checkpoints)
7143 1610 : LogCheckpointStart(flags, false);
7144 :
7145 : /* Update the process title */
7146 2134 : update_checkpoint_display(flags, false, false);
7147 :
7148 : TRACE_POSTGRESQL_CHECKPOINT_START(flags);
7149 :
7150 : /*
7151 : * Get the other info we need for the checkpoint record.
7152 : *
7153 : * We don't need to save oldestClogXid in the checkpoint, it only matters
7154 : * for the short period in which clog is being truncated, and if we crash
7155 : * during that we'll redo the clog truncation and fix up oldestClogXid
7156 : * there.
7157 : */
7158 2134 : LWLockAcquire(XidGenLock, LW_SHARED);
7159 2134 : checkPoint.nextXid = TransamVariables->nextXid;
7160 2134 : checkPoint.oldestXid = TransamVariables->oldestXid;
7161 2134 : checkPoint.oldestXidDB = TransamVariables->oldestXidDB;
7162 2134 : LWLockRelease(XidGenLock);
7163 :
7164 2134 : LWLockAcquire(CommitTsLock, LW_SHARED);
7165 2134 : checkPoint.oldestCommitTsXid = TransamVariables->oldestCommitTsXid;
7166 2134 : checkPoint.newestCommitTsXid = TransamVariables->newestCommitTsXid;
7167 2134 : LWLockRelease(CommitTsLock);
7168 :
7169 2134 : LWLockAcquire(OidGenLock, LW_SHARED);
7170 2134 : checkPoint.nextOid = TransamVariables->nextOid;
7171 2134 : if (!shutdown)
7172 1074 : checkPoint.nextOid += TransamVariables->oidCount;
7173 2134 : LWLockRelease(OidGenLock);
7174 :
7175 2134 : MultiXactGetCheckptMulti(shutdown,
7176 : &checkPoint.nextMulti,
7177 : &checkPoint.nextMultiOffset,
7178 : &checkPoint.oldestMulti,
7179 : &checkPoint.oldestMultiDB);
7180 :
7181 : /*
7182 : * Having constructed the checkpoint record, ensure all shmem disk buffers
7183 : * and commit-log buffers are flushed to disk.
7184 : *
7185 : * This I/O could fail for various reasons. If so, we will fail to
7186 : * complete the checkpoint, but there is no reason to force a system
7187 : * panic. Accordingly, exit critical section while doing it.
7188 : */
7189 2134 : END_CRIT_SECTION();
7190 :
7191 : /*
7192 : * In some cases there are groups of actions that must all occur on one
7193 : * side or the other of a checkpoint record. Before flushing the
7194 : * checkpoint record we must explicitly wait for any backend currently
7195 : * performing those groups of actions.
7196 : *
7197 : * One example is end of transaction, so we must wait for any transactions
7198 : * that are currently in commit critical sections. If an xact inserted
7199 : * its commit record into XLOG just before the REDO point, then a crash
7200 : * restart from the REDO point would not replay that record, which means
7201 : * that our flushing had better include the xact's update of pg_xact. So
7202 : * we wait till he's out of his commit critical section before proceeding.
7203 : * See notes in RecordTransactionCommit().
7204 : *
7205 : * Because we've already released the insertion locks, this test is a bit
7206 : * fuzzy: it is possible that we will wait for xacts we didn't really need
7207 : * to wait for. But the delay should be short and it seems better to make
7208 : * checkpoint take a bit longer than to hold off insertions longer than
7209 : * necessary. (In fact, the whole reason we have this issue is that xact.c
7210 : * does commit record XLOG insertion and clog update as two separate steps
7211 : * protected by different locks, but again that seems best on grounds of
7212 : * minimizing lock contention.)
7213 : *
7214 : * A transaction that has not yet set delayChkptFlags when we look cannot
7215 : * be at risk, since it has not inserted its commit record yet; and one
7216 : * that's already cleared it is not at risk either, since it's done fixing
7217 : * clog and we will correctly flush the update below. So we cannot miss
7218 : * any xacts we need to wait for.
7219 : */
7220 2134 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
7221 2134 : if (nvxids > 0)
7222 : {
7223 : do
7224 : {
7225 : /*
7226 : * Keep absorbing fsync requests while we wait. There could even
7227 : * be a deadlock if we don't, if the process that prevents the
7228 : * checkpoint is trying to add a request to the queue.
7229 : */
7230 60 : AbsorbSyncRequests();
7231 :
7232 60 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
7233 60 : pg_usleep(10000L); /* wait for 10 msec */
7234 60 : pgstat_report_wait_end();
7235 60 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7236 60 : DELAY_CHKPT_START));
7237 : }
7238 2134 : pfree(vxids);
7239 :
7240 2134 : CheckPointGuts(checkPoint.redo, flags);
7241 :
7242 2134 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
7243 2134 : if (nvxids > 0)
7244 : {
7245 : do
7246 : {
7247 0 : AbsorbSyncRequests();
7248 :
7249 0 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_COMPLETE);
7250 0 : pg_usleep(10000L); /* wait for 10 msec */
7251 0 : pgstat_report_wait_end();
7252 0 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7253 0 : DELAY_CHKPT_COMPLETE));
7254 : }
7255 2134 : pfree(vxids);
7256 :
7257 : /*
7258 : * Take a snapshot of running transactions and write this to WAL. This
7259 : * allows us to reconstruct the state of running transactions during
7260 : * archive recovery, if required. Skip, if this info disabled.
7261 : *
7262 : * If we are shutting down, or Startup process is completing crash
7263 : * recovery we don't need to write running xact data.
7264 : */
7265 2134 : if (!shutdown && XLogStandbyInfoActive())
7266 1004 : LogStandbySnapshot();
7267 :
7268 2134 : START_CRIT_SECTION();
7269 :
7270 : /*
7271 : * Now insert the checkpoint record into XLOG.
7272 : */
7273 2134 : XLogBeginInsert();
7274 2134 : XLogRegisterData(&checkPoint, sizeof(checkPoint));
7275 2134 : recptr = XLogInsert(RM_XLOG_ID,
7276 : shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
7277 : XLOG_CHECKPOINT_ONLINE);
7278 :
7279 2134 : XLogFlush(recptr);
7280 :
7281 : /*
7282 : * We mustn't write any new WAL after a shutdown checkpoint, or it will be
7283 : * overwritten at next startup. No-one should even try, this just allows
7284 : * sanity-checking. In the case of an end-of-recovery checkpoint, we want
7285 : * to just temporarily disable writing until the system has exited
7286 : * recovery.
7287 : */
7288 2134 : if (shutdown)
7289 : {
7290 1060 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7291 58 : LocalXLogInsertAllowed = oldXLogAllowed;
7292 : else
7293 1002 : LocalXLogInsertAllowed = 0; /* never again write WAL */
7294 : }
7295 :
7296 : /*
7297 : * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
7298 : * = end of actual checkpoint record.
7299 : */
7300 2134 : if (shutdown && checkPoint.redo != ProcLastRecPtr)
7301 0 : ereport(PANIC,
7302 : (errmsg("concurrent write-ahead log activity while database system is shutting down")));
7303 :
7304 : /*
7305 : * Remember the prior checkpoint's redo ptr for
7306 : * UpdateCheckPointDistanceEstimate()
7307 : */
7308 2134 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7309 :
7310 : /*
7311 : * Update the control file.
7312 : */
7313 2134 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7314 2134 : if (shutdown)
7315 1060 : ControlFile->state = DB_SHUTDOWNED;
7316 2134 : ControlFile->checkPoint = ProcLastRecPtr;
7317 2134 : ControlFile->checkPointCopy = checkPoint;
7318 : /* crash recovery should always recover to the end of WAL */
7319 2134 : ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
7320 2134 : ControlFile->minRecoveryPointTLI = 0;
7321 :
7322 : /*
7323 : * Persist unloggedLSN value. It's reset on crash recovery, so this goes
7324 : * unused on non-shutdown checkpoints, but seems useful to store it always
7325 : * for debugging purposes.
7326 : */
7327 2134 : ControlFile->unloggedLSN = pg_atomic_read_membarrier_u64(&XLogCtl->unloggedLSN);
7328 :
7329 2134 : UpdateControlFile();
7330 2134 : LWLockRelease(ControlFileLock);
7331 :
7332 : /* Update shared-memory copy of checkpoint XID/epoch */
7333 2134 : SpinLockAcquire(&XLogCtl->info_lck);
7334 2134 : XLogCtl->ckptFullXid = checkPoint.nextXid;
7335 2134 : SpinLockRelease(&XLogCtl->info_lck);
7336 :
7337 : /*
7338 : * We are now done with critical updates; no need for system panic if we
7339 : * have trouble while fooling with old log segments.
7340 : */
7341 2134 : END_CRIT_SECTION();
7342 :
7343 : /*
7344 : * WAL summaries end when the next XLOG_CHECKPOINT_REDO or
7345 : * XLOG_CHECKPOINT_SHUTDOWN record is reached. This is the first point
7346 : * where (a) we're not inside of a critical section and (b) we can be
7347 : * certain that the relevant record has been flushed to disk, which must
7348 : * happen before it can be summarized.
7349 : *
7350 : * If this is a shutdown checkpoint, then this happens reasonably
7351 : * promptly: we've only just inserted and flushed the
7352 : * XLOG_CHECKPOINT_SHUTDOWN record. If this is not a shutdown checkpoint,
7353 : * then this might not be very prompt at all: the XLOG_CHECKPOINT_REDO
7354 : * record was written before we began flushing data to disk, and that
7355 : * could be many minutes ago at this point. However, we don't XLogFlush()
7356 : * after inserting that record, so we're not guaranteed that it's on disk
7357 : * until after the above call that flushes the XLOG_CHECKPOINT_ONLINE
7358 : * record.
7359 : */
7360 2134 : WakeupWalSummarizer();
7361 :
7362 : /*
7363 : * Let smgr do post-checkpoint cleanup (eg, deleting old files).
7364 : */
7365 2134 : SyncPostCheckpoint();
7366 :
7367 : /*
7368 : * Update the average distance between checkpoints if the prior checkpoint
7369 : * exists.
7370 : */
7371 2134 : if (PriorRedoPtr != InvalidXLogRecPtr)
7372 2134 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7373 :
7374 : /*
7375 : * Delete old log files, those no longer needed for last checkpoint to
7376 : * prevent the disk holding the xlog from growing full.
7377 : */
7378 2134 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7379 2134 : KeepLogSeg(recptr, &_logSegNo);
7380 2134 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7381 : _logSegNo, InvalidOid,
7382 : InvalidTransactionId))
7383 : {
7384 : /*
7385 : * Some slots have been invalidated; recalculate the old-segment
7386 : * horizon, starting again from RedoRecPtr.
7387 : */
7388 6 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7389 6 : KeepLogSeg(recptr, &_logSegNo);
7390 : }
7391 2134 : _logSegNo--;
7392 2134 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
7393 : checkPoint.ThisTimeLineID);
7394 :
7395 : /*
7396 : * Make more log segments if needed. (Do this after recycling old log
7397 : * segments, since that may supply some of the needed files.)
7398 : */
7399 2134 : if (!shutdown)
7400 1074 : PreallocXlogFiles(recptr, checkPoint.ThisTimeLineID);
7401 :
7402 : /*
7403 : * Truncate pg_subtrans if possible. We can throw away all data before
7404 : * the oldest XMIN of any running transaction. No future transaction will
7405 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7406 : * in subtrans.c). During recovery, though, we mustn't do this because
7407 : * StartupSUBTRANS hasn't been called yet.
7408 : */
7409 2134 : if (!RecoveryInProgress())
7410 2076 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7411 :
7412 : /* Real work is done; log and update stats. */
7413 2134 : LogCheckpointEnd(false);
7414 :
7415 : /* Reset the process title */
7416 2134 : update_checkpoint_display(flags, false, true);
7417 :
7418 : TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
7419 : NBuffers,
7420 : CheckpointStats.ckpt_segs_added,
7421 : CheckpointStats.ckpt_segs_removed,
7422 : CheckpointStats.ckpt_segs_recycled);
7423 :
7424 2134 : return true;
7425 : }
7426 :
7427 : /*
7428 : * Mark the end of recovery in WAL though without running a full checkpoint.
7429 : * We can expect that a restartpoint is likely to be in progress as we
7430 : * do this, though we are unwilling to wait for it to complete.
7431 : *
7432 : * CreateRestartPoint() allows for the case where recovery may end before
7433 : * the restartpoint completes so there is no concern of concurrent behaviour.
7434 : */
7435 : static void
7436 84 : CreateEndOfRecoveryRecord(void)
7437 : {
7438 : xl_end_of_recovery xlrec;
7439 : XLogRecPtr recptr;
7440 :
7441 : /* sanity check */
7442 84 : if (!RecoveryInProgress())
7443 0 : elog(ERROR, "can only be used to end recovery");
7444 :
7445 84 : xlrec.end_time = GetCurrentTimestamp();
7446 84 : xlrec.wal_level = wal_level;
7447 :
7448 84 : WALInsertLockAcquireExclusive();
7449 84 : xlrec.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7450 84 : xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7451 84 : WALInsertLockRelease();
7452 :
7453 84 : START_CRIT_SECTION();
7454 :
7455 84 : XLogBeginInsert();
7456 84 : XLogRegisterData(&xlrec, sizeof(xl_end_of_recovery));
7457 84 : recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY);
7458 :
7459 84 : XLogFlush(recptr);
7460 :
7461 : /*
7462 : * Update the control file so that crash recovery can follow the timeline
7463 : * changes to this point.
7464 : */
7465 84 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7466 84 : ControlFile->minRecoveryPoint = recptr;
7467 84 : ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
7468 84 : UpdateControlFile();
7469 84 : LWLockRelease(ControlFileLock);
7470 :
7471 84 : END_CRIT_SECTION();
7472 84 : }
7473 :
7474 : /*
7475 : * Write an OVERWRITE_CONTRECORD message.
7476 : *
7477 : * When on WAL replay we expect a continuation record at the start of a page
7478 : * that is not there, recovery ends and WAL writing resumes at that point.
7479 : * But it's wrong to resume writing new WAL back at the start of the record
7480 : * that was broken, because downstream consumers of that WAL (physical
7481 : * replicas) are not prepared to "rewind". So the first action after
7482 : * finishing replay of all valid WAL must be to write a record of this type
7483 : * at the point where the contrecord was missing; to support xlogreader
7484 : * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
7485 : * to the page header where the record occurs. xlogreader has an ad-hoc
7486 : * mechanism to report metadata about the broken record, which is what we
7487 : * use here.
7488 : *
7489 : * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
7490 : * skip the record it was reading, and pass back the LSN of the skipped
7491 : * record, so that its caller can verify (on "replay" of that record) that the
7492 : * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
7493 : *
7494 : * 'aborted_lsn' is the beginning position of the record that was incomplete.
7495 : * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
7496 : * beginning of the XLOG page where the record is to be inserted. They must
7497 : * match the current WAL insert position, they're passed here just so that we
7498 : * can verify that.
7499 : */
7500 : static XLogRecPtr
7501 22 : CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn, XLogRecPtr pagePtr,
7502 : TimeLineID newTLI)
7503 : {
7504 : xl_overwrite_contrecord xlrec;
7505 : XLogRecPtr recptr;
7506 : XLogPageHeader pagehdr;
7507 : XLogRecPtr startPos;
7508 :
7509 : /* sanity checks */
7510 22 : if (!RecoveryInProgress())
7511 0 : elog(ERROR, "can only be used at end of recovery");
7512 22 : if (pagePtr % XLOG_BLCKSZ != 0)
7513 0 : elog(ERROR, "invalid position for missing continuation record %X/%X",
7514 : LSN_FORMAT_ARGS(pagePtr));
7515 :
7516 : /* The current WAL insert position should be right after the page header */
7517 22 : startPos = pagePtr;
7518 22 : if (XLogSegmentOffset(startPos, wal_segment_size) == 0)
7519 2 : startPos += SizeOfXLogLongPHD;
7520 : else
7521 20 : startPos += SizeOfXLogShortPHD;
7522 22 : recptr = GetXLogInsertRecPtr();
7523 22 : if (recptr != startPos)
7524 0 : elog(ERROR, "invalid WAL insert position %X/%X for OVERWRITE_CONTRECORD",
7525 : LSN_FORMAT_ARGS(recptr));
7526 :
7527 22 : START_CRIT_SECTION();
7528 :
7529 : /*
7530 : * Initialize the XLOG page header (by GetXLogBuffer), and set the
7531 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
7532 : *
7533 : * No other backend is allowed to write WAL yet, so acquiring the WAL
7534 : * insertion lock is just pro forma.
7535 : */
7536 22 : WALInsertLockAcquire();
7537 22 : pagehdr = (XLogPageHeader) GetXLogBuffer(pagePtr, newTLI);
7538 22 : pagehdr->xlp_info |= XLP_FIRST_IS_OVERWRITE_CONTRECORD;
7539 22 : WALInsertLockRelease();
7540 :
7541 : /*
7542 : * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
7543 : * page. We know it becomes the first record, because no other backend is
7544 : * allowed to write WAL yet.
7545 : */
7546 22 : XLogBeginInsert();
7547 22 : xlrec.overwritten_lsn = aborted_lsn;
7548 22 : xlrec.overwrite_time = GetCurrentTimestamp();
7549 22 : XLogRegisterData(&xlrec, sizeof(xl_overwrite_contrecord));
7550 22 : recptr = XLogInsert(RM_XLOG_ID, XLOG_OVERWRITE_CONTRECORD);
7551 :
7552 : /* check that the record was inserted to the right place */
7553 22 : if (ProcLastRecPtr != startPos)
7554 0 : elog(ERROR, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%X",
7555 : LSN_FORMAT_ARGS(ProcLastRecPtr));
7556 :
7557 22 : XLogFlush(recptr);
7558 :
7559 22 : END_CRIT_SECTION();
7560 :
7561 22 : return recptr;
7562 : }
7563 :
7564 : /*
7565 : * Flush all data in shared memory to disk, and fsync
7566 : *
7567 : * This is the common code shared between regular checkpoints and
7568 : * recovery restartpoints.
7569 : */
7570 : static void
7571 2506 : CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
7572 : {
7573 2506 : CheckPointRelationMap();
7574 2506 : CheckPointReplicationSlots(flags & CHECKPOINT_IS_SHUTDOWN);
7575 2506 : CheckPointSnapBuild();
7576 2506 : CheckPointLogicalRewriteHeap();
7577 2506 : CheckPointReplicationOrigin();
7578 :
7579 : /* Write out all dirty data in SLRUs and the main buffer pool */
7580 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags);
7581 2506 : CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
7582 2506 : CheckPointCLOG();
7583 2506 : CheckPointCommitTs();
7584 2506 : CheckPointSUBTRANS();
7585 2506 : CheckPointMultiXact();
7586 2506 : CheckPointPredicate();
7587 2506 : CheckPointBuffers(flags);
7588 :
7589 : /* Perform all queued up fsyncs */
7590 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
7591 2506 : CheckpointStats.ckpt_sync_t = GetCurrentTimestamp();
7592 2506 : ProcessSyncRequests();
7593 2506 : CheckpointStats.ckpt_sync_end_t = GetCurrentTimestamp();
7594 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
7595 :
7596 : /* We deliberately delay 2PC checkpointing as long as possible */
7597 2506 : CheckPointTwoPhase(checkPointRedo);
7598 2506 : }
7599 :
7600 : /*
7601 : * Save a checkpoint for recovery restart if appropriate
7602 : *
7603 : * This function is called each time a checkpoint record is read from XLOG.
7604 : * It must determine whether the checkpoint represents a safe restartpoint or
7605 : * not. If so, the checkpoint record is stashed in shared memory so that
7606 : * CreateRestartPoint can consult it. (Note that the latter function is
7607 : * executed by the checkpointer, while this one will be executed by the
7608 : * startup process.)
7609 : */
7610 : static void
7611 758 : RecoveryRestartPoint(const CheckPoint *checkPoint, XLogReaderState *record)
7612 : {
7613 : /*
7614 : * Also refrain from creating a restartpoint if we have seen any
7615 : * references to non-existent pages. Restarting recovery from the
7616 : * restartpoint would not see the references, so we would lose the
7617 : * cross-check that the pages belonged to a relation that was dropped
7618 : * later.
7619 : */
7620 758 : if (XLogHaveInvalidPages())
7621 : {
7622 0 : elog(DEBUG2,
7623 : "could not record restart point at %X/%X because there "
7624 : "are unresolved references to invalid pages",
7625 : LSN_FORMAT_ARGS(checkPoint->redo));
7626 0 : return;
7627 : }
7628 :
7629 : /*
7630 : * Copy the checkpoint record to shared memory, so that checkpointer can
7631 : * work out the next time it wants to perform a restartpoint.
7632 : */
7633 758 : SpinLockAcquire(&XLogCtl->info_lck);
7634 758 : XLogCtl->lastCheckPointRecPtr = record->ReadRecPtr;
7635 758 : XLogCtl->lastCheckPointEndPtr = record->EndRecPtr;
7636 758 : XLogCtl->lastCheckPoint = *checkPoint;
7637 758 : SpinLockRelease(&XLogCtl->info_lck);
7638 : }
7639 :
7640 : /*
7641 : * Establish a restartpoint if possible.
7642 : *
7643 : * This is similar to CreateCheckPoint, but is used during WAL recovery
7644 : * to establish a point from which recovery can roll forward without
7645 : * replaying the entire recovery log.
7646 : *
7647 : * Returns true if a new restartpoint was established. We can only establish
7648 : * a restartpoint if we have replayed a safe checkpoint record since last
7649 : * restartpoint.
7650 : */
7651 : bool
7652 512 : CreateRestartPoint(int flags)
7653 : {
7654 : XLogRecPtr lastCheckPointRecPtr;
7655 : XLogRecPtr lastCheckPointEndPtr;
7656 : CheckPoint lastCheckPoint;
7657 : XLogRecPtr PriorRedoPtr;
7658 : XLogRecPtr receivePtr;
7659 : XLogRecPtr replayPtr;
7660 : TimeLineID replayTLI;
7661 : XLogRecPtr endptr;
7662 : XLogSegNo _logSegNo;
7663 : TimestampTz xtime;
7664 :
7665 : /* Concurrent checkpoint/restartpoint cannot happen */
7666 : Assert(!IsUnderPostmaster || MyBackendType == B_CHECKPOINTER);
7667 :
7668 : /* Get a local copy of the last safe checkpoint record. */
7669 512 : SpinLockAcquire(&XLogCtl->info_lck);
7670 512 : lastCheckPointRecPtr = XLogCtl->lastCheckPointRecPtr;
7671 512 : lastCheckPointEndPtr = XLogCtl->lastCheckPointEndPtr;
7672 512 : lastCheckPoint = XLogCtl->lastCheckPoint;
7673 512 : SpinLockRelease(&XLogCtl->info_lck);
7674 :
7675 : /*
7676 : * Check that we're still in recovery mode. It's ok if we exit recovery
7677 : * mode after this check, the restart point is valid anyway.
7678 : */
7679 512 : if (!RecoveryInProgress())
7680 : {
7681 0 : ereport(DEBUG2,
7682 : (errmsg_internal("skipping restartpoint, recovery has already ended")));
7683 0 : return false;
7684 : }
7685 :
7686 : /*
7687 : * If the last checkpoint record we've replayed is already our last
7688 : * restartpoint, we can't perform a new restart point. We still update
7689 : * minRecoveryPoint in that case, so that if this is a shutdown restart
7690 : * point, we won't start up earlier than before. That's not strictly
7691 : * necessary, but when hot standby is enabled, it would be rather weird if
7692 : * the database opened up for read-only connections at a point-in-time
7693 : * before the last shutdown. Such time travel is still possible in case of
7694 : * immediate shutdown, though.
7695 : *
7696 : * We don't explicitly advance minRecoveryPoint when we do create a
7697 : * restartpoint. It's assumed that flushing the buffers will do that as a
7698 : * side-effect.
7699 : */
7700 512 : if (XLogRecPtrIsInvalid(lastCheckPointRecPtr) ||
7701 496 : lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
7702 : {
7703 140 : ereport(DEBUG2,
7704 : (errmsg_internal("skipping restartpoint, already performed at %X/%X",
7705 : LSN_FORMAT_ARGS(lastCheckPoint.redo))));
7706 :
7707 140 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
7708 140 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7709 : {
7710 64 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7711 64 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7712 64 : UpdateControlFile();
7713 64 : LWLockRelease(ControlFileLock);
7714 : }
7715 140 : return false;
7716 : }
7717 :
7718 : /*
7719 : * Update the shared RedoRecPtr so that the startup process can calculate
7720 : * the number of segments replayed since last restartpoint, and request a
7721 : * restartpoint if it exceeds CheckPointSegments.
7722 : *
7723 : * Like in CreateCheckPoint(), hold off insertions to update it, although
7724 : * during recovery this is just pro forma, because no WAL insertions are
7725 : * happening.
7726 : */
7727 372 : WALInsertLockAcquireExclusive();
7728 372 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = lastCheckPoint.redo;
7729 372 : WALInsertLockRelease();
7730 :
7731 : /* Also update the info_lck-protected copy */
7732 372 : SpinLockAcquire(&XLogCtl->info_lck);
7733 372 : XLogCtl->RedoRecPtr = lastCheckPoint.redo;
7734 372 : SpinLockRelease(&XLogCtl->info_lck);
7735 :
7736 : /*
7737 : * Prepare to accumulate statistics.
7738 : *
7739 : * Note: because it is possible for log_checkpoints to change while a
7740 : * checkpoint proceeds, we always accumulate stats, even if
7741 : * log_checkpoints is currently off.
7742 : */
7743 4092 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7744 372 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7745 :
7746 372 : if (log_checkpoints)
7747 372 : LogCheckpointStart(flags, true);
7748 :
7749 : /* Update the process title */
7750 372 : update_checkpoint_display(flags, true, false);
7751 :
7752 372 : CheckPointGuts(lastCheckPoint.redo, flags);
7753 :
7754 : /*
7755 : * This location needs to be after CheckPointGuts() to ensure that some
7756 : * work has already happened during this checkpoint.
7757 : */
7758 372 : INJECTION_POINT("create-restart-point");
7759 :
7760 : /*
7761 : * Remember the prior checkpoint's redo ptr for
7762 : * UpdateCheckPointDistanceEstimate()
7763 : */
7764 372 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7765 :
7766 : /*
7767 : * Update pg_control, using current time. Check that it still shows an
7768 : * older checkpoint, else do nothing; this is a quick hack to make sure
7769 : * nothing really bad happens if somehow we get here after the
7770 : * end-of-recovery checkpoint.
7771 : */
7772 372 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7773 372 : if (ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
7774 : {
7775 : /*
7776 : * Update the checkpoint information. We do this even if the cluster
7777 : * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
7778 : * segments recycled below.
7779 : */
7780 372 : ControlFile->checkPoint = lastCheckPointRecPtr;
7781 372 : ControlFile->checkPointCopy = lastCheckPoint;
7782 :
7783 : /*
7784 : * Ensure minRecoveryPoint is past the checkpoint record and update it
7785 : * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
7786 : * this will have happened already while writing out dirty buffers,
7787 : * but not necessarily - e.g. because no buffers were dirtied. We do
7788 : * this because a backup performed in recovery uses minRecoveryPoint
7789 : * to determine which WAL files must be included in the backup, and
7790 : * the file (or files) containing the checkpoint record must be
7791 : * included, at a minimum. Note that for an ordinary restart of
7792 : * recovery there's no value in having the minimum recovery point any
7793 : * earlier than this anyway, because redo will begin just after the
7794 : * checkpoint record.
7795 : */
7796 372 : if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
7797 : {
7798 372 : if (ControlFile->minRecoveryPoint < lastCheckPointEndPtr)
7799 : {
7800 36 : ControlFile->minRecoveryPoint = lastCheckPointEndPtr;
7801 36 : ControlFile->minRecoveryPointTLI = lastCheckPoint.ThisTimeLineID;
7802 :
7803 : /* update local copy */
7804 36 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
7805 36 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
7806 : }
7807 372 : if (flags & CHECKPOINT_IS_SHUTDOWN)
7808 36 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
7809 : }
7810 372 : UpdateControlFile();
7811 : }
7812 372 : LWLockRelease(ControlFileLock);
7813 :
7814 : /*
7815 : * Update the average distance between checkpoints/restartpoints if the
7816 : * prior checkpoint exists.
7817 : */
7818 372 : if (PriorRedoPtr != InvalidXLogRecPtr)
7819 372 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7820 :
7821 : /*
7822 : * Delete old log files, those no longer needed for last restartpoint to
7823 : * prevent the disk holding the xlog from growing full.
7824 : */
7825 372 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7826 :
7827 : /*
7828 : * Retreat _logSegNo using the current end of xlog replayed or received,
7829 : * whichever is later.
7830 : */
7831 372 : receivePtr = GetWalRcvFlushRecPtr(NULL, NULL);
7832 372 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
7833 372 : endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
7834 372 : KeepLogSeg(endptr, &_logSegNo);
7835 372 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7836 : _logSegNo, InvalidOid,
7837 : InvalidTransactionId))
7838 : {
7839 : /*
7840 : * Some slots have been invalidated; recalculate the old-segment
7841 : * horizon, starting again from RedoRecPtr.
7842 : */
7843 2 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7844 2 : KeepLogSeg(endptr, &_logSegNo);
7845 : }
7846 372 : _logSegNo--;
7847 :
7848 : /*
7849 : * Try to recycle segments on a useful timeline. If we've been promoted
7850 : * since the beginning of this restartpoint, use the new timeline chosen
7851 : * at end of recovery. If we're still in recovery, use the timeline we're
7852 : * currently replaying.
7853 : *
7854 : * There is no guarantee that the WAL segments will be useful on the
7855 : * current timeline; if recovery proceeds to a new timeline right after
7856 : * this, the pre-allocated WAL segments on this timeline will not be used,
7857 : * and will go wasted until recycled on the next restartpoint. We'll live
7858 : * with that.
7859 : */
7860 372 : if (!RecoveryInProgress())
7861 0 : replayTLI = XLogCtl->InsertTimeLineID;
7862 :
7863 372 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, endptr, replayTLI);
7864 :
7865 : /*
7866 : * Make more log segments if needed. (Do this after recycling old log
7867 : * segments, since that may supply some of the needed files.)
7868 : */
7869 372 : PreallocXlogFiles(endptr, replayTLI);
7870 :
7871 : /*
7872 : * Truncate pg_subtrans if possible. We can throw away all data before
7873 : * the oldest XMIN of any running transaction. No future transaction will
7874 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7875 : * in subtrans.c). When hot standby is disabled, though, we mustn't do
7876 : * this because StartupSUBTRANS hasn't been called yet.
7877 : */
7878 372 : if (EnableHotStandby)
7879 372 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7880 :
7881 : /* Real work is done; log and update stats. */
7882 372 : LogCheckpointEnd(true);
7883 :
7884 : /* Reset the process title */
7885 372 : update_checkpoint_display(flags, true, true);
7886 :
7887 372 : xtime = GetLatestXTime();
7888 372 : ereport((log_checkpoints ? LOG : DEBUG2),
7889 : (errmsg("recovery restart point at %X/%X",
7890 : LSN_FORMAT_ARGS(lastCheckPoint.redo)),
7891 : xtime ? errdetail("Last completed transaction was at log time %s.",
7892 : timestamptz_to_str(xtime)) : 0));
7893 :
7894 : /*
7895 : * Finally, execute archive_cleanup_command, if any.
7896 : */
7897 372 : if (archiveCleanupCommand && strcmp(archiveCleanupCommand, "") != 0)
7898 0 : ExecuteRecoveryCommand(archiveCleanupCommand,
7899 : "archive_cleanup_command",
7900 : false,
7901 : WAIT_EVENT_ARCHIVE_CLEANUP_COMMAND);
7902 :
7903 372 : return true;
7904 : }
7905 :
7906 : /*
7907 : * Report availability of WAL for the given target LSN
7908 : * (typically a slot's restart_lsn)
7909 : *
7910 : * Returns one of the following enum values:
7911 : *
7912 : * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
7913 : * max_wal_size.
7914 : *
7915 : * * WALAVAIL_EXTENDED means it is still available by preserving extra
7916 : * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
7917 : * than max_wal_size, this state is not returned.
7918 : *
7919 : * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
7920 : * remove reserved segments. The walsender using this slot may return to the
7921 : * above.
7922 : *
7923 : * * WALAVAIL_REMOVED means it has been removed. A replication stream on
7924 : * a slot with this LSN cannot continue. (Any associated walsender
7925 : * processes should have been terminated already.)
7926 : *
7927 : * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
7928 : */
7929 : WALAvailability
7930 850 : GetWALAvailability(XLogRecPtr targetLSN)
7931 : {
7932 : XLogRecPtr currpos; /* current write LSN */
7933 : XLogSegNo currSeg; /* segid of currpos */
7934 : XLogSegNo targetSeg; /* segid of targetLSN */
7935 : XLogSegNo oldestSeg; /* actual oldest segid */
7936 : XLogSegNo oldestSegMaxWalSize; /* oldest segid kept by max_wal_size */
7937 : XLogSegNo oldestSlotSeg; /* oldest segid kept by slot */
7938 : uint64 keepSegs;
7939 :
7940 : /*
7941 : * slot does not reserve WAL. Either deactivated, or has never been active
7942 : */
7943 850 : if (XLogRecPtrIsInvalid(targetLSN))
7944 28 : return WALAVAIL_INVALID_LSN;
7945 :
7946 : /*
7947 : * Calculate the oldest segment currently reserved by all slots,
7948 : * considering wal_keep_size and max_slot_wal_keep_size. Initialize
7949 : * oldestSlotSeg to the current segment.
7950 : */
7951 822 : currpos = GetXLogWriteRecPtr();
7952 822 : XLByteToSeg(currpos, oldestSlotSeg, wal_segment_size);
7953 822 : KeepLogSeg(currpos, &oldestSlotSeg);
7954 :
7955 : /*
7956 : * Find the oldest extant segment file. We get 1 until checkpoint removes
7957 : * the first WAL segment file since startup, which causes the status being
7958 : * wrong under certain abnormal conditions but that doesn't actually harm.
7959 : */
7960 822 : oldestSeg = XLogGetLastRemovedSegno() + 1;
7961 :
7962 : /* calculate oldest segment by max_wal_size */
7963 822 : XLByteToSeg(currpos, currSeg, wal_segment_size);
7964 822 : keepSegs = ConvertToXSegs(max_wal_size_mb, wal_segment_size) + 1;
7965 :
7966 822 : if (currSeg > keepSegs)
7967 16 : oldestSegMaxWalSize = currSeg - keepSegs;
7968 : else
7969 806 : oldestSegMaxWalSize = 1;
7970 :
7971 : /* the segment we care about */
7972 822 : XLByteToSeg(targetLSN, targetSeg, wal_segment_size);
7973 :
7974 : /*
7975 : * No point in returning reserved or extended status values if the
7976 : * targetSeg is known to be lost.
7977 : */
7978 822 : if (targetSeg >= oldestSlotSeg)
7979 : {
7980 : /* show "reserved" when targetSeg is within max_wal_size */
7981 820 : if (targetSeg >= oldestSegMaxWalSize)
7982 816 : return WALAVAIL_RESERVED;
7983 :
7984 : /* being retained by slots exceeding max_wal_size */
7985 4 : return WALAVAIL_EXTENDED;
7986 : }
7987 :
7988 : /* WAL segments are no longer retained but haven't been removed yet */
7989 2 : if (targetSeg >= oldestSeg)
7990 2 : return WALAVAIL_UNRESERVED;
7991 :
7992 : /* Definitely lost */
7993 0 : return WALAVAIL_REMOVED;
7994 : }
7995 :
7996 :
7997 : /*
7998 : * Retreat *logSegNo to the last segment that we need to retain because of
7999 : * either wal_keep_size or replication slots.
8000 : *
8001 : * This is calculated by subtracting wal_keep_size from the given xlog
8002 : * location, recptr and by making sure that that result is below the
8003 : * requirement of replication slots. For the latter criterion we do consider
8004 : * the effects of max_slot_wal_keep_size: reserve at most that much space back
8005 : * from recptr.
8006 : *
8007 : * Note about replication slots: if this function calculates a value
8008 : * that's further ahead than what slots need reserved, then affected
8009 : * slots need to be invalidated and this function invoked again.
8010 : * XXX it might be a good idea to rewrite this function so that
8011 : * invalidation is optionally done here, instead.
8012 : */
8013 : static void
8014 3336 : KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
8015 : {
8016 : XLogSegNo currSegNo;
8017 : XLogSegNo segno;
8018 : XLogRecPtr keep;
8019 :
8020 3336 : XLByteToSeg(recptr, currSegNo, wal_segment_size);
8021 3336 : segno = currSegNo;
8022 :
8023 : /*
8024 : * Calculate how many segments are kept by slots first, adjusting for
8025 : * max_slot_wal_keep_size.
8026 : */
8027 3336 : keep = XLogGetReplicationSlotMinimumLSN();
8028 3336 : if (keep != InvalidXLogRecPtr && keep < recptr)
8029 : {
8030 1064 : XLByteToSeg(keep, segno, wal_segment_size);
8031 :
8032 : /* Cap by max_slot_wal_keep_size ... */
8033 1064 : if (max_slot_wal_keep_size_mb >= 0)
8034 : {
8035 : uint64 slot_keep_segs;
8036 :
8037 40 : slot_keep_segs =
8038 40 : ConvertToXSegs(max_slot_wal_keep_size_mb, wal_segment_size);
8039 :
8040 40 : if (currSegNo - segno > slot_keep_segs)
8041 10 : segno = currSegNo - slot_keep_segs;
8042 : }
8043 : }
8044 :
8045 : /*
8046 : * If WAL summarization is in use, don't remove WAL that has yet to be
8047 : * summarized.
8048 : */
8049 3336 : keep = GetOldestUnsummarizedLSN(NULL, NULL);
8050 3336 : if (keep != InvalidXLogRecPtr)
8051 : {
8052 : XLogSegNo unsummarized_segno;
8053 :
8054 4 : XLByteToSeg(keep, unsummarized_segno, wal_segment_size);
8055 4 : if (unsummarized_segno < segno)
8056 4 : segno = unsummarized_segno;
8057 : }
8058 :
8059 : /* but, keep at least wal_keep_size if that's set */
8060 3336 : if (wal_keep_size_mb > 0)
8061 : {
8062 : uint64 keep_segs;
8063 :
8064 138 : keep_segs = ConvertToXSegs(wal_keep_size_mb, wal_segment_size);
8065 138 : if (currSegNo - segno < keep_segs)
8066 : {
8067 : /* avoid underflow, don't go below 1 */
8068 138 : if (currSegNo <= keep_segs)
8069 130 : segno = 1;
8070 : else
8071 8 : segno = currSegNo - keep_segs;
8072 : }
8073 : }
8074 :
8075 : /* don't delete WAL segments newer than the calculated segment */
8076 3336 : if (segno < *logSegNo)
8077 590 : *logSegNo = segno;
8078 3336 : }
8079 :
8080 : /*
8081 : * Write a NEXTOID log record
8082 : */
8083 : void
8084 1094 : XLogPutNextOid(Oid nextOid)
8085 : {
8086 1094 : XLogBeginInsert();
8087 1094 : XLogRegisterData(&nextOid, sizeof(Oid));
8088 1094 : (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID);
8089 :
8090 : /*
8091 : * We need not flush the NEXTOID record immediately, because any of the
8092 : * just-allocated OIDs could only reach disk as part of a tuple insert or
8093 : * update that would have its own XLOG record that must follow the NEXTOID
8094 : * record. Therefore, the standard buffer LSN interlock applied to those
8095 : * records will ensure no such OID reaches disk before the NEXTOID record
8096 : * does.
8097 : *
8098 : * Note, however, that the above statement only covers state "within" the
8099 : * database. When we use a generated OID as a file or directory name, we
8100 : * are in a sense violating the basic WAL rule, because that filesystem
8101 : * change may reach disk before the NEXTOID WAL record does. The impact
8102 : * of this is that if a database crash occurs immediately afterward, we
8103 : * might after restart re-generate the same OID and find that it conflicts
8104 : * with the leftover file or directory. But since for safety's sake we
8105 : * always loop until finding a nonconflicting filename, this poses no real
8106 : * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8107 : */
8108 1094 : }
8109 :
8110 : /*
8111 : * Write an XLOG SWITCH record.
8112 : *
8113 : * Here we just blindly issue an XLogInsert request for the record.
8114 : * All the magic happens inside XLogInsert.
8115 : *
8116 : * The return value is either the end+1 address of the switch record,
8117 : * or the end+1 address of the prior segment if we did not need to
8118 : * write a switch record because we are already at segment start.
8119 : */
8120 : XLogRecPtr
8121 738 : RequestXLogSwitch(bool mark_unimportant)
8122 : {
8123 : XLogRecPtr RecPtr;
8124 :
8125 : /* XLOG SWITCH has no data */
8126 738 : XLogBeginInsert();
8127 :
8128 738 : if (mark_unimportant)
8129 0 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8130 738 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH);
8131 :
8132 738 : return RecPtr;
8133 : }
8134 :
8135 : /*
8136 : * Write a RESTORE POINT record
8137 : */
8138 : XLogRecPtr
8139 6 : XLogRestorePoint(const char *rpName)
8140 : {
8141 : XLogRecPtr RecPtr;
8142 : xl_restore_point xlrec;
8143 :
8144 6 : xlrec.rp_time = GetCurrentTimestamp();
8145 6 : strlcpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8146 :
8147 6 : XLogBeginInsert();
8148 6 : XLogRegisterData(&xlrec, sizeof(xl_restore_point));
8149 :
8150 6 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT);
8151 :
8152 6 : ereport(LOG,
8153 : (errmsg("restore point \"%s\" created at %X/%X",
8154 : rpName, LSN_FORMAT_ARGS(RecPtr))));
8155 :
8156 6 : return RecPtr;
8157 : }
8158 :
8159 : /*
8160 : * Check if any of the GUC parameters that are critical for hot standby
8161 : * have changed, and update the value in pg_control file if necessary.
8162 : */
8163 : static void
8164 1562 : XLogReportParameters(void)
8165 : {
8166 1562 : if (wal_level != ControlFile->wal_level ||
8167 1148 : wal_log_hints != ControlFile->wal_log_hints ||
8168 992 : MaxConnections != ControlFile->MaxConnections ||
8169 990 : max_worker_processes != ControlFile->max_worker_processes ||
8170 988 : max_wal_senders != ControlFile->max_wal_senders ||
8171 960 : max_prepared_xacts != ControlFile->max_prepared_xacts ||
8172 788 : max_locks_per_xact != ControlFile->max_locks_per_xact ||
8173 788 : track_commit_timestamp != ControlFile->track_commit_timestamp)
8174 : {
8175 : /*
8176 : * The change in number of backend slots doesn't need to be WAL-logged
8177 : * if archiving is not enabled, as you can't start archive recovery
8178 : * with wal_level=minimal anyway. We don't really care about the
8179 : * values in pg_control either if wal_level=minimal, but seems better
8180 : * to keep them up-to-date to avoid confusion.
8181 : */
8182 794 : if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8183 : {
8184 : xl_parameter_change xlrec;
8185 : XLogRecPtr recptr;
8186 :
8187 766 : xlrec.MaxConnections = MaxConnections;
8188 766 : xlrec.max_worker_processes = max_worker_processes;
8189 766 : xlrec.max_wal_senders = max_wal_senders;
8190 766 : xlrec.max_prepared_xacts = max_prepared_xacts;
8191 766 : xlrec.max_locks_per_xact = max_locks_per_xact;
8192 766 : xlrec.wal_level = wal_level;
8193 766 : xlrec.wal_log_hints = wal_log_hints;
8194 766 : xlrec.track_commit_timestamp = track_commit_timestamp;
8195 :
8196 766 : XLogBeginInsert();
8197 766 : XLogRegisterData(&xlrec, sizeof(xlrec));
8198 :
8199 766 : recptr = XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE);
8200 766 : XLogFlush(recptr);
8201 : }
8202 :
8203 794 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8204 :
8205 794 : ControlFile->MaxConnections = MaxConnections;
8206 794 : ControlFile->max_worker_processes = max_worker_processes;
8207 794 : ControlFile->max_wal_senders = max_wal_senders;
8208 794 : ControlFile->max_prepared_xacts = max_prepared_xacts;
8209 794 : ControlFile->max_locks_per_xact = max_locks_per_xact;
8210 794 : ControlFile->wal_level = wal_level;
8211 794 : ControlFile->wal_log_hints = wal_log_hints;
8212 794 : ControlFile->track_commit_timestamp = track_commit_timestamp;
8213 794 : UpdateControlFile();
8214 :
8215 794 : LWLockRelease(ControlFileLock);
8216 : }
8217 1562 : }
8218 :
8219 : /*
8220 : * Update full_page_writes in shared memory, and write an
8221 : * XLOG_FPW_CHANGE record if necessary.
8222 : *
8223 : * Note: this function assumes there is no other process running
8224 : * concurrently that could update it.
8225 : */
8226 : void
8227 2574 : UpdateFullPageWrites(void)
8228 : {
8229 2574 : XLogCtlInsert *Insert = &XLogCtl->Insert;
8230 : bool recoveryInProgress;
8231 :
8232 : /*
8233 : * Do nothing if full_page_writes has not been changed.
8234 : *
8235 : * It's safe to check the shared full_page_writes without the lock,
8236 : * because we assume that there is no concurrently running process which
8237 : * can update it.
8238 : */
8239 2574 : if (fullPageWrites == Insert->fullPageWrites)
8240 1688 : return;
8241 :
8242 : /*
8243 : * Perform this outside critical section so that the WAL insert
8244 : * initialization done by RecoveryInProgress() doesn't trigger an
8245 : * assertion failure.
8246 : */
8247 886 : recoveryInProgress = RecoveryInProgress();
8248 :
8249 886 : START_CRIT_SECTION();
8250 :
8251 : /*
8252 : * It's always safe to take full page images, even when not strictly
8253 : * required, but not the other round. So if we're setting full_page_writes
8254 : * to true, first set it true and then write the WAL record. If we're
8255 : * setting it to false, first write the WAL record and then set the global
8256 : * flag.
8257 : */
8258 886 : if (fullPageWrites)
8259 : {
8260 876 : WALInsertLockAcquireExclusive();
8261 876 : Insert->fullPageWrites = true;
8262 876 : WALInsertLockRelease();
8263 : }
8264 :
8265 : /*
8266 : * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8267 : * full_page_writes during archive recovery, if required.
8268 : */
8269 886 : if (XLogStandbyInfoActive() && !recoveryInProgress)
8270 : {
8271 0 : XLogBeginInsert();
8272 0 : XLogRegisterData(&fullPageWrites, sizeof(bool));
8273 :
8274 0 : XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
8275 : }
8276 :
8277 886 : if (!fullPageWrites)
8278 : {
8279 10 : WALInsertLockAcquireExclusive();
8280 10 : Insert->fullPageWrites = false;
8281 10 : WALInsertLockRelease();
8282 : }
8283 886 : END_CRIT_SECTION();
8284 : }
8285 :
8286 : /*
8287 : * XLOG resource manager's routines
8288 : *
8289 : * Definitions of info values are in include/catalog/pg_control.h, though
8290 : * not all record types are related to control file updates.
8291 : *
8292 : * NOTE: Some XLOG record types that are directly related to WAL recovery
8293 : * are handled in xlogrecovery_redo().
8294 : */
8295 : void
8296 80116 : xlog_redo(XLogReaderState *record)
8297 : {
8298 80116 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8299 80116 : XLogRecPtr lsn = record->EndRecPtr;
8300 :
8301 : /*
8302 : * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
8303 : * XLOG_FPI_FOR_HINT records.
8304 : */
8305 : Assert(info == XLOG_FPI || info == XLOG_FPI_FOR_HINT ||
8306 : !XLogRecHasAnyBlockRefs(record));
8307 :
8308 80116 : if (info == XLOG_NEXTOID)
8309 : {
8310 : Oid nextOid;
8311 :
8312 : /*
8313 : * We used to try to take the maximum of TransamVariables->nextOid and
8314 : * the recorded nextOid, but that fails if the OID counter wraps
8315 : * around. Since no OID allocation should be happening during replay
8316 : * anyway, better to just believe the record exactly. We still take
8317 : * OidGenLock while setting the variable, just in case.
8318 : */
8319 180 : memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
8320 180 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8321 180 : TransamVariables->nextOid = nextOid;
8322 180 : TransamVariables->oidCount = 0;
8323 180 : LWLockRelease(OidGenLock);
8324 : }
8325 79936 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
8326 : {
8327 : CheckPoint checkPoint;
8328 : TimeLineID replayTLI;
8329 :
8330 68 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8331 : /* In a SHUTDOWN checkpoint, believe the counters exactly */
8332 68 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8333 68 : TransamVariables->nextXid = checkPoint.nextXid;
8334 68 : LWLockRelease(XidGenLock);
8335 68 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8336 68 : TransamVariables->nextOid = checkPoint.nextOid;
8337 68 : TransamVariables->oidCount = 0;
8338 68 : LWLockRelease(OidGenLock);
8339 68 : MultiXactSetNextMXact(checkPoint.nextMulti,
8340 : checkPoint.nextMultiOffset);
8341 :
8342 68 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8343 : checkPoint.oldestMultiDB);
8344 :
8345 : /*
8346 : * No need to set oldestClogXid here as well; it'll be set when we
8347 : * redo an xl_clog_truncate if it changed since initialization.
8348 : */
8349 68 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
8350 :
8351 : /*
8352 : * If we see a shutdown checkpoint while waiting for an end-of-backup
8353 : * record, the backup was canceled and the end-of-backup record will
8354 : * never arrive.
8355 : */
8356 68 : if (ArchiveRecoveryRequested &&
8357 66 : !XLogRecPtrIsInvalid(ControlFile->backupStartPoint) &&
8358 0 : XLogRecPtrIsInvalid(ControlFile->backupEndPoint))
8359 0 : ereport(PANIC,
8360 : (errmsg("online backup was canceled, recovery cannot continue")));
8361 :
8362 : /*
8363 : * If we see a shutdown checkpoint, we know that nothing was running
8364 : * on the primary at this point. So fake-up an empty running-xacts
8365 : * record and use that here and now. Recover additional standby state
8366 : * for prepared transactions.
8367 : */
8368 68 : if (standbyState >= STANDBY_INITIALIZED)
8369 : {
8370 : TransactionId *xids;
8371 : int nxids;
8372 : TransactionId oldestActiveXID;
8373 : TransactionId latestCompletedXid;
8374 : RunningTransactionsData running;
8375 :
8376 62 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
8377 :
8378 : /* Update pg_subtrans entries for any prepared transactions */
8379 62 : StandbyRecoverPreparedTransactions();
8380 :
8381 : /*
8382 : * Construct a RunningTransactions snapshot representing a shut
8383 : * down server, with only prepared transactions still alive. We're
8384 : * never overflowed at this point because all subxids are listed
8385 : * with their parent prepared transactions.
8386 : */
8387 62 : running.xcnt = nxids;
8388 62 : running.subxcnt = 0;
8389 62 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
8390 62 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
8391 62 : running.oldestRunningXid = oldestActiveXID;
8392 62 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
8393 62 : TransactionIdRetreat(latestCompletedXid);
8394 : Assert(TransactionIdIsNormal(latestCompletedXid));
8395 62 : running.latestCompletedXid = latestCompletedXid;
8396 62 : running.xids = xids;
8397 :
8398 62 : ProcArrayApplyRecoveryInfo(&running);
8399 : }
8400 :
8401 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8402 68 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8403 68 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8404 68 : LWLockRelease(ControlFileLock);
8405 :
8406 : /* Update shared-memory copy of checkpoint XID/epoch */
8407 68 : SpinLockAcquire(&XLogCtl->info_lck);
8408 68 : XLogCtl->ckptFullXid = checkPoint.nextXid;
8409 68 : SpinLockRelease(&XLogCtl->info_lck);
8410 :
8411 : /*
8412 : * We should've already switched to the new TLI before replaying this
8413 : * record.
8414 : */
8415 68 : (void) GetCurrentReplayRecPtr(&replayTLI);
8416 68 : if (checkPoint.ThisTimeLineID != replayTLI)
8417 0 : ereport(PANIC,
8418 : (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
8419 : checkPoint.ThisTimeLineID, replayTLI)));
8420 :
8421 68 : RecoveryRestartPoint(&checkPoint, record);
8422 : }
8423 79868 : else if (info == XLOG_CHECKPOINT_ONLINE)
8424 : {
8425 : CheckPoint checkPoint;
8426 : TimeLineID replayTLI;
8427 :
8428 690 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8429 : /* In an ONLINE checkpoint, treat the XID counter as a minimum */
8430 690 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8431 690 : if (FullTransactionIdPrecedes(TransamVariables->nextXid,
8432 : checkPoint.nextXid))
8433 0 : TransamVariables->nextXid = checkPoint.nextXid;
8434 690 : LWLockRelease(XidGenLock);
8435 :
8436 : /*
8437 : * We ignore the nextOid counter in an ONLINE checkpoint, preferring
8438 : * to track OID assignment through XLOG_NEXTOID records. The nextOid
8439 : * counter is from the start of the checkpoint and might well be stale
8440 : * compared to later XLOG_NEXTOID records. We could try to take the
8441 : * maximum of the nextOid counter and our latest value, but since
8442 : * there's no particular guarantee about the speed with which the OID
8443 : * counter wraps around, that's a risky thing to do. In any case,
8444 : * users of the nextOid counter are required to avoid assignment of
8445 : * duplicates, so that a somewhat out-of-date value should be safe.
8446 : */
8447 :
8448 : /* Handle multixact */
8449 690 : MultiXactAdvanceNextMXact(checkPoint.nextMulti,
8450 : checkPoint.nextMultiOffset);
8451 :
8452 : /*
8453 : * NB: This may perform multixact truncation when replaying WAL
8454 : * generated by an older primary.
8455 : */
8456 690 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8457 : checkPoint.oldestMultiDB);
8458 690 : if (TransactionIdPrecedes(TransamVariables->oldestXid,
8459 : checkPoint.oldestXid))
8460 0 : SetTransactionIdLimit(checkPoint.oldestXid,
8461 : checkPoint.oldestXidDB);
8462 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8463 690 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8464 690 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8465 690 : LWLockRelease(ControlFileLock);
8466 :
8467 : /* Update shared-memory copy of checkpoint XID/epoch */
8468 690 : SpinLockAcquire(&XLogCtl->info_lck);
8469 690 : XLogCtl->ckptFullXid = checkPoint.nextXid;
8470 690 : SpinLockRelease(&XLogCtl->info_lck);
8471 :
8472 : /* TLI should not change in an on-line checkpoint */
8473 690 : (void) GetCurrentReplayRecPtr(&replayTLI);
8474 690 : if (checkPoint.ThisTimeLineID != replayTLI)
8475 0 : ereport(PANIC,
8476 : (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
8477 : checkPoint.ThisTimeLineID, replayTLI)));
8478 :
8479 690 : RecoveryRestartPoint(&checkPoint, record);
8480 : }
8481 79178 : else if (info == XLOG_OVERWRITE_CONTRECORD)
8482 : {
8483 : /* nothing to do here, handled in xlogrecovery_redo() */
8484 : }
8485 79176 : else if (info == XLOG_END_OF_RECOVERY)
8486 : {
8487 : xl_end_of_recovery xlrec;
8488 : TimeLineID replayTLI;
8489 :
8490 20 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
8491 :
8492 : /*
8493 : * For Hot Standby, we could treat this like a Shutdown Checkpoint,
8494 : * but this case is rarer and harder to test, so the benefit doesn't
8495 : * outweigh the potential extra cost of maintenance.
8496 : */
8497 :
8498 : /*
8499 : * We should've already switched to the new TLI before replaying this
8500 : * record.
8501 : */
8502 20 : (void) GetCurrentReplayRecPtr(&replayTLI);
8503 20 : if (xlrec.ThisTimeLineID != replayTLI)
8504 0 : ereport(PANIC,
8505 : (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
8506 : xlrec.ThisTimeLineID, replayTLI)));
8507 : }
8508 79156 : else if (info == XLOG_NOOP)
8509 : {
8510 : /* nothing to do here */
8511 : }
8512 79156 : else if (info == XLOG_SWITCH)
8513 : {
8514 : /* nothing to do here */
8515 : }
8516 78918 : else if (info == XLOG_RESTORE_POINT)
8517 : {
8518 : /* nothing to do here, handled in xlogrecovery.c */
8519 : }
8520 78908 : else if (info == XLOG_FPI || info == XLOG_FPI_FOR_HINT)
8521 : {
8522 : /*
8523 : * XLOG_FPI records contain nothing else but one or more block
8524 : * references. Every block reference must include a full-page image
8525 : * even if full_page_writes was disabled when the record was generated
8526 : * - otherwise there would be no point in this record.
8527 : *
8528 : * XLOG_FPI_FOR_HINT records are generated when a page needs to be
8529 : * WAL-logged because of a hint bit update. They are only generated
8530 : * when checksums and/or wal_log_hints are enabled. They may include
8531 : * no full-page images if full_page_writes was disabled when they were
8532 : * generated. In this case there is nothing to do here.
8533 : *
8534 : * No recovery conflicts are generated by these generic records - if a
8535 : * resource manager needs to generate conflicts, it has to define a
8536 : * separate WAL record type and redo routine.
8537 : */
8538 165606 : for (uint8 block_id = 0; block_id <= XLogRecMaxBlockId(record); block_id++)
8539 : {
8540 : Buffer buffer;
8541 :
8542 87614 : if (!XLogRecHasBlockImage(record, block_id))
8543 : {
8544 134 : if (info == XLOG_FPI)
8545 0 : elog(ERROR, "XLOG_FPI record did not contain a full-page image");
8546 134 : continue;
8547 : }
8548 :
8549 87480 : if (XLogReadBufferForRedo(record, block_id, &buffer) != BLK_RESTORED)
8550 0 : elog(ERROR, "unexpected XLogReadBufferForRedo result when restoring backup block");
8551 87480 : UnlockReleaseBuffer(buffer);
8552 : }
8553 : }
8554 916 : else if (info == XLOG_BACKUP_END)
8555 : {
8556 : /* nothing to do here, handled in xlogrecovery_redo() */
8557 : }
8558 752 : else if (info == XLOG_PARAMETER_CHANGE)
8559 : {
8560 : xl_parameter_change xlrec;
8561 :
8562 : /* Update our copy of the parameters in pg_control */
8563 60 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
8564 :
8565 : /*
8566 : * Invalidate logical slots if we are in hot standby and the primary
8567 : * does not have a WAL level sufficient for logical decoding. No need
8568 : * to search for potentially conflicting logically slots if standby is
8569 : * running with wal_level lower than logical, because in that case, we
8570 : * would have either disallowed creation of logical slots or
8571 : * invalidated existing ones.
8572 : */
8573 60 : if (InRecovery && InHotStandby &&
8574 30 : xlrec.wal_level < WAL_LEVEL_LOGICAL &&
8575 12 : wal_level >= WAL_LEVEL_LOGICAL)
8576 8 : InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_LEVEL,
8577 : 0, InvalidOid,
8578 : InvalidTransactionId);
8579 :
8580 60 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8581 60 : ControlFile->MaxConnections = xlrec.MaxConnections;
8582 60 : ControlFile->max_worker_processes = xlrec.max_worker_processes;
8583 60 : ControlFile->max_wal_senders = xlrec.max_wal_senders;
8584 60 : ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
8585 60 : ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
8586 60 : ControlFile->wal_level = xlrec.wal_level;
8587 60 : ControlFile->wal_log_hints = xlrec.wal_log_hints;
8588 :
8589 : /*
8590 : * Update minRecoveryPoint to ensure that if recovery is aborted, we
8591 : * recover back up to this point before allowing hot standby again.
8592 : * This is important if the max_* settings are decreased, to ensure
8593 : * you don't run queries against the WAL preceding the change. The
8594 : * local copies cannot be updated as long as crash recovery is
8595 : * happening and we expect all the WAL to be replayed.
8596 : */
8597 60 : if (InArchiveRecovery)
8598 : {
8599 32 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
8600 32 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
8601 : }
8602 60 : if (LocalMinRecoveryPoint != InvalidXLogRecPtr && LocalMinRecoveryPoint < lsn)
8603 : {
8604 : TimeLineID replayTLI;
8605 :
8606 14 : (void) GetCurrentReplayRecPtr(&replayTLI);
8607 14 : ControlFile->minRecoveryPoint = lsn;
8608 14 : ControlFile->minRecoveryPointTLI = replayTLI;
8609 : }
8610 :
8611 60 : CommitTsParameterChange(xlrec.track_commit_timestamp,
8612 60 : ControlFile->track_commit_timestamp);
8613 60 : ControlFile->track_commit_timestamp = xlrec.track_commit_timestamp;
8614 :
8615 60 : UpdateControlFile();
8616 60 : LWLockRelease(ControlFileLock);
8617 :
8618 : /* Check to see if any parameter change gives a problem on recovery */
8619 60 : CheckRequiredParameterValues();
8620 : }
8621 692 : else if (info == XLOG_FPW_CHANGE)
8622 : {
8623 : bool fpw;
8624 :
8625 0 : memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
8626 :
8627 : /*
8628 : * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
8629 : * do_pg_backup_start() and do_pg_backup_stop() can check whether
8630 : * full_page_writes has been disabled during online backup.
8631 : */
8632 0 : if (!fpw)
8633 : {
8634 0 : SpinLockAcquire(&XLogCtl->info_lck);
8635 0 : if (XLogCtl->lastFpwDisableRecPtr < record->ReadRecPtr)
8636 0 : XLogCtl->lastFpwDisableRecPtr = record->ReadRecPtr;
8637 0 : SpinLockRelease(&XLogCtl->info_lck);
8638 : }
8639 :
8640 : /* Keep track of full_page_writes */
8641 0 : lastFullPageWrites = fpw;
8642 : }
8643 : else if (info == XLOG_CHECKPOINT_REDO)
8644 : {
8645 : /* nothing to do here, just for informational purposes */
8646 : }
8647 80112 : }
8648 :
8649 : /*
8650 : * Return the extra open flags used for opening a file, depending on the
8651 : * value of the GUCs wal_sync_method, fsync and debug_io_direct.
8652 : */
8653 : static int
8654 26880 : get_sync_bit(int method)
8655 : {
8656 26880 : int o_direct_flag = 0;
8657 :
8658 : /*
8659 : * Use O_DIRECT if requested, except in walreceiver process. The WAL
8660 : * written by walreceiver is normally read by the startup process soon
8661 : * after it's written. Also, walreceiver performs unaligned writes, which
8662 : * don't work with O_DIRECT, so it is required for correctness too.
8663 : */
8664 26880 : if ((io_direct_flags & IO_DIRECT_WAL) && !AmWalReceiverProcess())
8665 18 : o_direct_flag = PG_O_DIRECT;
8666 :
8667 : /* If fsync is disabled, never open in sync mode */
8668 26880 : if (!enableFsync)
8669 26880 : return o_direct_flag;
8670 :
8671 0 : switch (method)
8672 : {
8673 : /*
8674 : * enum values for all sync options are defined even if they are
8675 : * not supported on the current platform. But if not, they are
8676 : * not included in the enum option array, and therefore will never
8677 : * be seen here.
8678 : */
8679 0 : case WAL_SYNC_METHOD_FSYNC:
8680 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8681 : case WAL_SYNC_METHOD_FDATASYNC:
8682 0 : return o_direct_flag;
8683 : #ifdef O_SYNC
8684 0 : case WAL_SYNC_METHOD_OPEN:
8685 0 : return O_SYNC | o_direct_flag;
8686 : #endif
8687 : #ifdef O_DSYNC
8688 0 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8689 0 : return O_DSYNC | o_direct_flag;
8690 : #endif
8691 0 : default:
8692 : /* can't happen (unless we are out of sync with option array) */
8693 0 : elog(ERROR, "unrecognized \"wal_sync_method\": %d", method);
8694 : return 0; /* silence warning */
8695 : }
8696 : }
8697 :
8698 : /*
8699 : * GUC support
8700 : */
8701 : void
8702 2004 : assign_wal_sync_method(int new_wal_sync_method, void *extra)
8703 : {
8704 2004 : if (wal_sync_method != new_wal_sync_method)
8705 : {
8706 : /*
8707 : * To ensure that no blocks escape unsynced, force an fsync on the
8708 : * currently open log segment (if any). Also, if the open flag is
8709 : * changing, close the log file so it will be reopened (with new flag
8710 : * bit) at next use.
8711 : */
8712 0 : if (openLogFile >= 0)
8713 : {
8714 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN);
8715 0 : if (pg_fsync(openLogFile) != 0)
8716 : {
8717 : char xlogfname[MAXFNAMELEN];
8718 : int save_errno;
8719 :
8720 0 : save_errno = errno;
8721 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo,
8722 : wal_segment_size);
8723 0 : errno = save_errno;
8724 0 : ereport(PANIC,
8725 : (errcode_for_file_access(),
8726 : errmsg("could not fsync file \"%s\": %m", xlogfname)));
8727 : }
8728 :
8729 0 : pgstat_report_wait_end();
8730 0 : if (get_sync_bit(wal_sync_method) != get_sync_bit(new_wal_sync_method))
8731 0 : XLogFileClose();
8732 : }
8733 : }
8734 2004 : }
8735 :
8736 :
8737 : /*
8738 : * Issue appropriate kind of fsync (if any) for an XLOG output file.
8739 : *
8740 : * 'fd' is a file descriptor for the XLOG file to be fsync'd.
8741 : * 'segno' is for error reporting purposes.
8742 : */
8743 : void
8744 264312 : issue_xlog_fsync(int fd, XLogSegNo segno, TimeLineID tli)
8745 : {
8746 264312 : char *msg = NULL;
8747 : instr_time start;
8748 :
8749 : Assert(tli != 0);
8750 :
8751 : /*
8752 : * Quick exit if fsync is disabled or write() has already synced the WAL
8753 : * file.
8754 : */
8755 264312 : if (!enableFsync ||
8756 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN ||
8757 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN_DSYNC)
8758 264312 : return;
8759 :
8760 : /*
8761 : * Measure I/O timing to sync the WAL file for pg_stat_io and/or
8762 : * pg_stat_wal.
8763 : */
8764 0 : start = pgstat_prepare_io_time(track_io_timing || track_wal_io_timing);
8765 :
8766 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC);
8767 0 : switch (wal_sync_method)
8768 : {
8769 0 : case WAL_SYNC_METHOD_FSYNC:
8770 0 : if (pg_fsync_no_writethrough(fd) != 0)
8771 0 : msg = _("could not fsync file \"%s\": %m");
8772 0 : break;
8773 : #ifdef HAVE_FSYNC_WRITETHROUGH
8774 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
8775 : if (pg_fsync_writethrough(fd) != 0)
8776 : msg = _("could not fsync write-through file \"%s\": %m");
8777 : break;
8778 : #endif
8779 0 : case WAL_SYNC_METHOD_FDATASYNC:
8780 0 : if (pg_fdatasync(fd) != 0)
8781 0 : msg = _("could not fdatasync file \"%s\": %m");
8782 0 : break;
8783 0 : case WAL_SYNC_METHOD_OPEN:
8784 : case WAL_SYNC_METHOD_OPEN_DSYNC:
8785 : /* not reachable */
8786 : Assert(false);
8787 0 : break;
8788 0 : default:
8789 0 : ereport(PANIC,
8790 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8791 : errmsg_internal("unrecognized \"wal_sync_method\": %d", wal_sync_method));
8792 : break;
8793 : }
8794 :
8795 : /* PANIC if failed to fsync */
8796 0 : if (msg)
8797 : {
8798 : char xlogfname[MAXFNAMELEN];
8799 0 : int save_errno = errno;
8800 :
8801 0 : XLogFileName(xlogfname, tli, segno, wal_segment_size);
8802 0 : errno = save_errno;
8803 0 : ereport(PANIC,
8804 : (errcode_for_file_access(),
8805 : errmsg(msg, xlogfname)));
8806 : }
8807 :
8808 0 : pgstat_report_wait_end();
8809 :
8810 : /*
8811 : * Increment the I/O timing and the number of times WAL files were synced.
8812 : */
8813 0 : if (track_wal_io_timing)
8814 : {
8815 : instr_time end;
8816 :
8817 0 : INSTR_TIME_SET_CURRENT(end);
8818 0 : INSTR_TIME_ACCUM_DIFF(PendingWalStats.wal_sync_time, end, start);
8819 : }
8820 :
8821 0 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL, IOOP_FSYNC,
8822 : start, 1, 0);
8823 :
8824 0 : PendingWalStats.wal_sync++;
8825 : }
8826 :
8827 : /*
8828 : * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
8829 : * function. It creates the necessary starting checkpoint and constructs the
8830 : * backup state and tablespace map.
8831 : *
8832 : * Input parameters are "state" (the backup state), "fast" (if true, we do
8833 : * the checkpoint in immediate mode to make it faster), and "tablespaces"
8834 : * (if non-NULL, indicates a list of tablespaceinfo structs describing the
8835 : * cluster's tablespaces.).
8836 : *
8837 : * The tablespace map contents are appended to passed-in parameter
8838 : * tablespace_map and the caller is responsible for including it in the backup
8839 : * archive as 'tablespace_map'. The tablespace_map file is required mainly for
8840 : * tar format in windows as native windows utilities are not able to create
8841 : * symlinks while extracting files from tar. However for consistency and
8842 : * platform-independence, we do it the same way everywhere.
8843 : *
8844 : * It fills in "state" with the information required for the backup, such
8845 : * as the minimum WAL location that must be present to restore from this
8846 : * backup (starttli) and the corresponding timeline ID (starttli).
8847 : *
8848 : * Every successfully started backup must be stopped by calling
8849 : * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
8850 : * backups active at the same time.
8851 : *
8852 : * It is the responsibility of the caller of this function to verify the
8853 : * permissions of the calling user!
8854 : */
8855 : void
8856 318 : do_pg_backup_start(const char *backupidstr, bool fast, List **tablespaces,
8857 : BackupState *state, StringInfo tblspcmapfile)
8858 : {
8859 : bool backup_started_in_recovery;
8860 :
8861 : Assert(state != NULL);
8862 318 : backup_started_in_recovery = RecoveryInProgress();
8863 :
8864 : /*
8865 : * During recovery, we don't need to check WAL level. Because, if WAL
8866 : * level is not sufficient, it's impossible to get here during recovery.
8867 : */
8868 318 : if (!backup_started_in_recovery && !XLogIsNeeded())
8869 0 : ereport(ERROR,
8870 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8871 : errmsg("WAL level not sufficient for making an online backup"),
8872 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
8873 :
8874 318 : if (strlen(backupidstr) > MAXPGPATH)
8875 2 : ereport(ERROR,
8876 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
8877 : errmsg("backup label too long (max %d bytes)",
8878 : MAXPGPATH)));
8879 :
8880 316 : strlcpy(state->name, backupidstr, sizeof(state->name));
8881 :
8882 : /*
8883 : * Mark backup active in shared memory. We must do full-page WAL writes
8884 : * during an on-line backup even if not doing so at other times, because
8885 : * it's quite possible for the backup dump to obtain a "torn" (partially
8886 : * written) copy of a database page if it reads the page concurrently with
8887 : * our write to the same page. This can be fixed as long as the first
8888 : * write to the page in the WAL sequence is a full-page write. Hence, we
8889 : * increment runningBackups then force a CHECKPOINT, to ensure there are
8890 : * no dirty pages in shared memory that might get dumped while the backup
8891 : * is in progress without having a corresponding WAL record. (Once the
8892 : * backup is complete, we need not force full-page writes anymore, since
8893 : * we expect that any pages not modified during the backup interval must
8894 : * have been correctly captured by the backup.)
8895 : *
8896 : * Note that forcing full-page writes has no effect during an online
8897 : * backup from the standby.
8898 : *
8899 : * We must hold all the insertion locks to change the value of
8900 : * runningBackups, to ensure adequate interlocking against
8901 : * XLogInsertRecord().
8902 : */
8903 316 : WALInsertLockAcquireExclusive();
8904 316 : XLogCtl->Insert.runningBackups++;
8905 316 : WALInsertLockRelease();
8906 :
8907 : /*
8908 : * Ensure we decrement runningBackups if we fail below. NB -- for this to
8909 : * work correctly, it is critical that sessionBackupState is only updated
8910 : * after this block is over.
8911 : */
8912 316 : PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, DatumGetBool(true));
8913 : {
8914 316 : bool gotUniqueStartpoint = false;
8915 : DIR *tblspcdir;
8916 : struct dirent *de;
8917 : tablespaceinfo *ti;
8918 : int datadirpathlen;
8919 :
8920 : /*
8921 : * Force an XLOG file switch before the checkpoint, to ensure that the
8922 : * WAL segment the checkpoint is written to doesn't contain pages with
8923 : * old timeline IDs. That would otherwise happen if you called
8924 : * pg_backup_start() right after restoring from a PITR archive: the
8925 : * first WAL segment containing the startup checkpoint has pages in
8926 : * the beginning with the old timeline ID. That can cause trouble at
8927 : * recovery: we won't have a history file covering the old timeline if
8928 : * pg_wal directory was not included in the base backup and the WAL
8929 : * archive was cleared too before starting the backup.
8930 : *
8931 : * This also ensures that we have emitted a WAL page header that has
8932 : * XLP_BKP_REMOVABLE off before we emit the checkpoint record.
8933 : * Therefore, if a WAL archiver (such as pglesslog) is trying to
8934 : * compress out removable backup blocks, it won't remove any that
8935 : * occur after this point.
8936 : *
8937 : * During recovery, we skip forcing XLOG file switch, which means that
8938 : * the backup taken during recovery is not available for the special
8939 : * recovery case described above.
8940 : */
8941 316 : if (!backup_started_in_recovery)
8942 302 : RequestXLogSwitch(false);
8943 :
8944 : do
8945 : {
8946 : bool checkpointfpw;
8947 :
8948 : /*
8949 : * Force a CHECKPOINT. Aside from being necessary to prevent torn
8950 : * page problems, this guarantees that two successive backup runs
8951 : * will have different checkpoint positions and hence different
8952 : * history file names, even if nothing happened in between.
8953 : *
8954 : * During recovery, establish a restartpoint if possible. We use
8955 : * the last restartpoint as the backup starting checkpoint. This
8956 : * means that two successive backup runs can have same checkpoint
8957 : * positions.
8958 : *
8959 : * Since the fact that we are executing do_pg_backup_start()
8960 : * during recovery means that checkpointer is running, we can use
8961 : * RequestCheckpoint() to establish a restartpoint.
8962 : *
8963 : * We use CHECKPOINT_IMMEDIATE only if requested by user (via
8964 : * passing fast = true). Otherwise this can take awhile.
8965 : */
8966 316 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
8967 : (fast ? CHECKPOINT_IMMEDIATE : 0));
8968 :
8969 : /*
8970 : * Now we need to fetch the checkpoint record location, and also
8971 : * its REDO pointer. The oldest point in WAL that would be needed
8972 : * to restore starting from the checkpoint is precisely the REDO
8973 : * pointer.
8974 : */
8975 316 : LWLockAcquire(ControlFileLock, LW_SHARED);
8976 316 : state->checkpointloc = ControlFile->checkPoint;
8977 316 : state->startpoint = ControlFile->checkPointCopy.redo;
8978 316 : state->starttli = ControlFile->checkPointCopy.ThisTimeLineID;
8979 316 : checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
8980 316 : LWLockRelease(ControlFileLock);
8981 :
8982 316 : if (backup_started_in_recovery)
8983 : {
8984 : XLogRecPtr recptr;
8985 :
8986 : /*
8987 : * Check to see if all WAL replayed during online backup
8988 : * (i.e., since last restartpoint used as backup starting
8989 : * checkpoint) contain full-page writes.
8990 : */
8991 14 : SpinLockAcquire(&XLogCtl->info_lck);
8992 14 : recptr = XLogCtl->lastFpwDisableRecPtr;
8993 14 : SpinLockRelease(&XLogCtl->info_lck);
8994 :
8995 14 : if (!checkpointfpw || state->startpoint <= recptr)
8996 0 : ereport(ERROR,
8997 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
8998 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
8999 : "since last restartpoint"),
9000 : errhint("This means that the backup being taken on the standby "
9001 : "is corrupt and should not be used. "
9002 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9003 : "and then try an online backup again.")));
9004 :
9005 : /*
9006 : * During recovery, since we don't use the end-of-backup WAL
9007 : * record and don't write the backup history file, the
9008 : * starting WAL location doesn't need to be unique. This means
9009 : * that two base backups started at the same time might use
9010 : * the same checkpoint as starting locations.
9011 : */
9012 14 : gotUniqueStartpoint = true;
9013 : }
9014 :
9015 : /*
9016 : * If two base backups are started at the same time (in WAL sender
9017 : * processes), we need to make sure that they use different
9018 : * checkpoints as starting locations, because we use the starting
9019 : * WAL location as a unique identifier for the base backup in the
9020 : * end-of-backup WAL record and when we write the backup history
9021 : * file. Perhaps it would be better generate a separate unique ID
9022 : * for each backup instead of forcing another checkpoint, but
9023 : * taking a checkpoint right after another is not that expensive
9024 : * either because only few buffers have been dirtied yet.
9025 : */
9026 316 : WALInsertLockAcquireExclusive();
9027 316 : if (XLogCtl->Insert.lastBackupStart < state->startpoint)
9028 : {
9029 316 : XLogCtl->Insert.lastBackupStart = state->startpoint;
9030 316 : gotUniqueStartpoint = true;
9031 : }
9032 316 : WALInsertLockRelease();
9033 316 : } while (!gotUniqueStartpoint);
9034 :
9035 : /*
9036 : * Construct tablespace_map file.
9037 : */
9038 316 : datadirpathlen = strlen(DataDir);
9039 :
9040 : /* Collect information about all tablespaces */
9041 316 : tblspcdir = AllocateDir(PG_TBLSPC_DIR);
9042 1020 : while ((de = ReadDir(tblspcdir, PG_TBLSPC_DIR)) != NULL)
9043 : {
9044 : char fullpath[MAXPGPATH + sizeof(PG_TBLSPC_DIR)];
9045 : char linkpath[MAXPGPATH];
9046 704 : char *relpath = NULL;
9047 : char *s;
9048 : PGFileType de_type;
9049 : char *badp;
9050 : Oid tsoid;
9051 :
9052 : /*
9053 : * Try to parse the directory name as an unsigned integer.
9054 : *
9055 : * Tablespace directories should be positive integers that can be
9056 : * represented in 32 bits, with no leading zeroes or trailing
9057 : * garbage. If we come across a name that doesn't meet those
9058 : * criteria, skip it.
9059 : */
9060 704 : if (de->d_name[0] < '1' || de->d_name[1] > '9')
9061 632 : continue;
9062 72 : errno = 0;
9063 72 : tsoid = strtoul(de->d_name, &badp, 10);
9064 72 : if (*badp != '\0' || errno == EINVAL || errno == ERANGE)
9065 0 : continue;
9066 :
9067 72 : snprintf(fullpath, sizeof(fullpath), "%s/%s", PG_TBLSPC_DIR, de->d_name);
9068 :
9069 72 : de_type = get_dirent_type(fullpath, de, false, ERROR);
9070 :
9071 72 : if (de_type == PGFILETYPE_LNK)
9072 : {
9073 : StringInfoData escapedpath;
9074 : int rllen;
9075 :
9076 44 : rllen = readlink(fullpath, linkpath, sizeof(linkpath));
9077 44 : if (rllen < 0)
9078 : {
9079 0 : ereport(WARNING,
9080 : (errmsg("could not read symbolic link \"%s\": %m",
9081 : fullpath)));
9082 0 : continue;
9083 : }
9084 44 : else if (rllen >= sizeof(linkpath))
9085 : {
9086 0 : ereport(WARNING,
9087 : (errmsg("symbolic link \"%s\" target is too long",
9088 : fullpath)));
9089 0 : continue;
9090 : }
9091 44 : linkpath[rllen] = '\0';
9092 :
9093 : /*
9094 : * Relpath holds the relative path of the tablespace directory
9095 : * when it's located within PGDATA, or NULL if it's located
9096 : * elsewhere.
9097 : */
9098 44 : if (rllen > datadirpathlen &&
9099 2 : strncmp(linkpath, DataDir, datadirpathlen) == 0 &&
9100 0 : IS_DIR_SEP(linkpath[datadirpathlen]))
9101 0 : relpath = pstrdup(linkpath + datadirpathlen + 1);
9102 :
9103 : /*
9104 : * Add a backslash-escaped version of the link path to the
9105 : * tablespace map file.
9106 : */
9107 44 : initStringInfo(&escapedpath);
9108 1092 : for (s = linkpath; *s; s++)
9109 : {
9110 1048 : if (*s == '\n' || *s == '\r' || *s == '\\')
9111 0 : appendStringInfoChar(&escapedpath, '\\');
9112 1048 : appendStringInfoChar(&escapedpath, *s);
9113 : }
9114 44 : appendStringInfo(tblspcmapfile, "%s %s\n",
9115 44 : de->d_name, escapedpath.data);
9116 44 : pfree(escapedpath.data);
9117 : }
9118 28 : else if (de_type == PGFILETYPE_DIR)
9119 : {
9120 : /*
9121 : * It's possible to use allow_in_place_tablespaces to create
9122 : * directories directly under pg_tblspc, for testing purposes
9123 : * only.
9124 : *
9125 : * In this case, we store a relative path rather than an
9126 : * absolute path into the tablespaceinfo.
9127 : */
9128 28 : snprintf(linkpath, sizeof(linkpath), "%s/%s",
9129 28 : PG_TBLSPC_DIR, de->d_name);
9130 28 : relpath = pstrdup(linkpath);
9131 : }
9132 : else
9133 : {
9134 : /* Skip any other file type that appears here. */
9135 0 : continue;
9136 : }
9137 :
9138 72 : ti = palloc(sizeof(tablespaceinfo));
9139 72 : ti->oid = tsoid;
9140 72 : ti->path = pstrdup(linkpath);
9141 72 : ti->rpath = relpath;
9142 72 : ti->size = -1;
9143 :
9144 72 : if (tablespaces)
9145 72 : *tablespaces = lappend(*tablespaces, ti);
9146 : }
9147 316 : FreeDir(tblspcdir);
9148 :
9149 316 : state->starttime = (pg_time_t) time(NULL);
9150 : }
9151 316 : PG_END_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, DatumGetBool(true));
9152 :
9153 316 : state->started_in_recovery = backup_started_in_recovery;
9154 :
9155 : /*
9156 : * Mark that the start phase has correctly finished for the backup.
9157 : */
9158 316 : sessionBackupState = SESSION_BACKUP_RUNNING;
9159 316 : }
9160 :
9161 : /*
9162 : * Utility routine to fetch the session-level status of a backup running.
9163 : */
9164 : SessionBackupState
9165 360 : get_backup_status(void)
9166 : {
9167 360 : return sessionBackupState;
9168 : }
9169 :
9170 : /*
9171 : * do_pg_backup_stop
9172 : *
9173 : * Utility function called at the end of an online backup. It creates history
9174 : * file (if required), resets sessionBackupState and so on. It can optionally
9175 : * wait for WAL segments to be archived.
9176 : *
9177 : * "state" is filled with the information necessary to restore from this
9178 : * backup with its stop LSN (stoppoint), its timeline ID (stoptli), etc.
9179 : *
9180 : * It is the responsibility of the caller of this function to verify the
9181 : * permissions of the calling user!
9182 : */
9183 : void
9184 304 : do_pg_backup_stop(BackupState *state, bool waitforarchive)
9185 : {
9186 304 : bool backup_stopped_in_recovery = false;
9187 : char histfilepath[MAXPGPATH];
9188 : char lastxlogfilename[MAXFNAMELEN];
9189 : char histfilename[MAXFNAMELEN];
9190 : XLogSegNo _logSegNo;
9191 : FILE *fp;
9192 : int seconds_before_warning;
9193 304 : int waits = 0;
9194 304 : bool reported_waiting = false;
9195 :
9196 : Assert(state != NULL);
9197 :
9198 304 : backup_stopped_in_recovery = RecoveryInProgress();
9199 :
9200 : /*
9201 : * During recovery, we don't need to check WAL level. Because, if WAL
9202 : * level is not sufficient, it's impossible to get here during recovery.
9203 : */
9204 304 : if (!backup_stopped_in_recovery && !XLogIsNeeded())
9205 0 : ereport(ERROR,
9206 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9207 : errmsg("WAL level not sufficient for making an online backup"),
9208 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9209 :
9210 : /*
9211 : * OK to update backup counter and session-level lock.
9212 : *
9213 : * Note that CHECK_FOR_INTERRUPTS() must not occur while updating them,
9214 : * otherwise they can be updated inconsistently, which might cause
9215 : * do_pg_abort_backup() to fail.
9216 : */
9217 304 : WALInsertLockAcquireExclusive();
9218 :
9219 : /*
9220 : * It is expected that each do_pg_backup_start() call is matched by
9221 : * exactly one do_pg_backup_stop() call.
9222 : */
9223 : Assert(XLogCtl->Insert.runningBackups > 0);
9224 304 : XLogCtl->Insert.runningBackups--;
9225 :
9226 : /*
9227 : * Clean up session-level lock.
9228 : *
9229 : * You might think that WALInsertLockRelease() can be called before
9230 : * cleaning up session-level lock because session-level lock doesn't need
9231 : * to be protected with WAL insertion lock. But since
9232 : * CHECK_FOR_INTERRUPTS() can occur in it, session-level lock must be
9233 : * cleaned up before it.
9234 : */
9235 304 : sessionBackupState = SESSION_BACKUP_NONE;
9236 :
9237 304 : WALInsertLockRelease();
9238 :
9239 : /*
9240 : * If we are taking an online backup from the standby, we confirm that the
9241 : * standby has not been promoted during the backup.
9242 : */
9243 304 : if (state->started_in_recovery && !backup_stopped_in_recovery)
9244 0 : ereport(ERROR,
9245 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9246 : errmsg("the standby was promoted during online backup"),
9247 : errhint("This means that the backup being taken is corrupt "
9248 : "and should not be used. "
9249 : "Try taking another online backup.")));
9250 :
9251 : /*
9252 : * During recovery, we don't write an end-of-backup record. We assume that
9253 : * pg_control was backed up last and its minimum recovery point can be
9254 : * available as the backup end location. Since we don't have an
9255 : * end-of-backup record, we use the pg_control value to check whether
9256 : * we've reached the end of backup when starting recovery from this
9257 : * backup. We have no way of checking if pg_control wasn't backed up last
9258 : * however.
9259 : *
9260 : * We don't force a switch to new WAL file but it is still possible to
9261 : * wait for all the required files to be archived if waitforarchive is
9262 : * true. This is okay if we use the backup to start a standby and fetch
9263 : * the missing WAL using streaming replication. But in the case of an
9264 : * archive recovery, a user should set waitforarchive to true and wait for
9265 : * them to be archived to ensure that all the required files are
9266 : * available.
9267 : *
9268 : * We return the current minimum recovery point as the backup end
9269 : * location. Note that it can be greater than the exact backup end
9270 : * location if the minimum recovery point is updated after the backup of
9271 : * pg_control. This is harmless for current uses.
9272 : *
9273 : * XXX currently a backup history file is for informational and debug
9274 : * purposes only. It's not essential for an online backup. Furthermore,
9275 : * even if it's created, it will not be archived during recovery because
9276 : * an archiver is not invoked. So it doesn't seem worthwhile to write a
9277 : * backup history file during recovery.
9278 : */
9279 304 : if (backup_stopped_in_recovery)
9280 : {
9281 : XLogRecPtr recptr;
9282 :
9283 : /*
9284 : * Check to see if all WAL replayed during online backup contain
9285 : * full-page writes.
9286 : */
9287 14 : SpinLockAcquire(&XLogCtl->info_lck);
9288 14 : recptr = XLogCtl->lastFpwDisableRecPtr;
9289 14 : SpinLockRelease(&XLogCtl->info_lck);
9290 :
9291 14 : if (state->startpoint <= recptr)
9292 0 : ereport(ERROR,
9293 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9294 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9295 : "during online backup"),
9296 : errhint("This means that the backup being taken on the standby "
9297 : "is corrupt and should not be used. "
9298 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9299 : "and then try an online backup again.")));
9300 :
9301 :
9302 14 : LWLockAcquire(ControlFileLock, LW_SHARED);
9303 14 : state->stoppoint = ControlFile->minRecoveryPoint;
9304 14 : state->stoptli = ControlFile->minRecoveryPointTLI;
9305 14 : LWLockRelease(ControlFileLock);
9306 : }
9307 : else
9308 : {
9309 : char *history_file;
9310 :
9311 : /*
9312 : * Write the backup-end xlog record
9313 : */
9314 290 : XLogBeginInsert();
9315 290 : XLogRegisterData(&state->startpoint,
9316 : sizeof(state->startpoint));
9317 290 : state->stoppoint = XLogInsert(RM_XLOG_ID, XLOG_BACKUP_END);
9318 :
9319 : /*
9320 : * Given that we're not in recovery, InsertTimeLineID is set and can't
9321 : * change, so we can read it without a lock.
9322 : */
9323 290 : state->stoptli = XLogCtl->InsertTimeLineID;
9324 :
9325 : /*
9326 : * Force a switch to a new xlog segment file, so that the backup is
9327 : * valid as soon as archiver moves out the current segment file.
9328 : */
9329 290 : RequestXLogSwitch(false);
9330 :
9331 290 : state->stoptime = (pg_time_t) time(NULL);
9332 :
9333 : /*
9334 : * Write the backup history file
9335 : */
9336 290 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9337 290 : BackupHistoryFilePath(histfilepath, state->stoptli, _logSegNo,
9338 : state->startpoint, wal_segment_size);
9339 290 : fp = AllocateFile(histfilepath, "w");
9340 290 : if (!fp)
9341 0 : ereport(ERROR,
9342 : (errcode_for_file_access(),
9343 : errmsg("could not create file \"%s\": %m",
9344 : histfilepath)));
9345 :
9346 : /* Build and save the contents of the backup history file */
9347 290 : history_file = build_backup_content(state, true);
9348 290 : fprintf(fp, "%s", history_file);
9349 290 : pfree(history_file);
9350 :
9351 290 : if (fflush(fp) || ferror(fp) || FreeFile(fp))
9352 0 : ereport(ERROR,
9353 : (errcode_for_file_access(),
9354 : errmsg("could not write file \"%s\": %m",
9355 : histfilepath)));
9356 :
9357 : /*
9358 : * Clean out any no-longer-needed history files. As a side effect,
9359 : * this will post a .ready file for the newly created history file,
9360 : * notifying the archiver that history file may be archived
9361 : * immediately.
9362 : */
9363 290 : CleanupBackupHistory();
9364 : }
9365 :
9366 : /*
9367 : * If archiving is enabled, wait for all the required WAL files to be
9368 : * archived before returning. If archiving isn't enabled, the required WAL
9369 : * needs to be transported via streaming replication (hopefully with
9370 : * wal_keep_size set high enough), or some more exotic mechanism like
9371 : * polling and copying files from pg_wal with script. We have no knowledge
9372 : * of those mechanisms, so it's up to the user to ensure that he gets all
9373 : * the required WAL.
9374 : *
9375 : * We wait until both the last WAL file filled during backup and the
9376 : * history file have been archived, and assume that the alphabetic sorting
9377 : * property of the WAL files ensures any earlier WAL files are safely
9378 : * archived as well.
9379 : *
9380 : * We wait forever, since archive_command is supposed to work and we
9381 : * assume the admin wanted his backup to work completely. If you don't
9382 : * wish to wait, then either waitforarchive should be passed in as false,
9383 : * or you can set statement_timeout. Also, some notices are issued to
9384 : * clue in anyone who might be doing this interactively.
9385 : */
9386 :
9387 304 : if (waitforarchive &&
9388 18 : ((!backup_stopped_in_recovery && XLogArchivingActive()) ||
9389 2 : (backup_stopped_in_recovery && XLogArchivingAlways())))
9390 : {
9391 6 : XLByteToPrevSeg(state->stoppoint, _logSegNo, wal_segment_size);
9392 6 : XLogFileName(lastxlogfilename, state->stoptli, _logSegNo,
9393 : wal_segment_size);
9394 :
9395 6 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9396 6 : BackupHistoryFileName(histfilename, state->stoptli, _logSegNo,
9397 : state->startpoint, wal_segment_size);
9398 :
9399 6 : seconds_before_warning = 60;
9400 6 : waits = 0;
9401 :
9402 18 : while (XLogArchiveIsBusy(lastxlogfilename) ||
9403 6 : XLogArchiveIsBusy(histfilename))
9404 : {
9405 6 : CHECK_FOR_INTERRUPTS();
9406 :
9407 6 : if (!reported_waiting && waits > 5)
9408 : {
9409 0 : ereport(NOTICE,
9410 : (errmsg("base backup done, waiting for required WAL segments to be archived")));
9411 0 : reported_waiting = true;
9412 : }
9413 :
9414 6 : (void) WaitLatch(MyLatch,
9415 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
9416 : 1000L,
9417 : WAIT_EVENT_BACKUP_WAIT_WAL_ARCHIVE);
9418 6 : ResetLatch(MyLatch);
9419 :
9420 6 : if (++waits >= seconds_before_warning)
9421 : {
9422 0 : seconds_before_warning *= 2; /* This wraps in >10 years... */
9423 0 : ereport(WARNING,
9424 : (errmsg("still waiting for all required WAL segments to be archived (%d seconds elapsed)",
9425 : waits),
9426 : errhint("Check that your \"archive_command\" is executing properly. "
9427 : "You can safely cancel this backup, "
9428 : "but the database backup will not be usable without all the WAL segments.")));
9429 : }
9430 : }
9431 :
9432 6 : ereport(NOTICE,
9433 : (errmsg("all required WAL segments have been archived")));
9434 : }
9435 298 : else if (waitforarchive)
9436 12 : ereport(NOTICE,
9437 : (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
9438 304 : }
9439 :
9440 :
9441 : /*
9442 : * do_pg_abort_backup: abort a running backup
9443 : *
9444 : * This does just the most basic steps of do_pg_backup_stop(), by taking the
9445 : * system out of backup mode, thus making it a lot more safe to call from
9446 : * an error handler.
9447 : *
9448 : * 'arg' indicates that it's being called during backup setup; so
9449 : * sessionBackupState has not been modified yet, but runningBackups has
9450 : * already been incremented. When it's false, then it's invoked as a
9451 : * before_shmem_exit handler, and therefore we must not change state
9452 : * unless sessionBackupState indicates that a backup is actually running.
9453 : *
9454 : * NB: This gets used as a PG_ENSURE_ERROR_CLEANUP callback and
9455 : * before_shmem_exit handler, hence the odd-looking signature.
9456 : */
9457 : void
9458 18 : do_pg_abort_backup(int code, Datum arg)
9459 : {
9460 18 : bool during_backup_start = DatumGetBool(arg);
9461 :
9462 : /* If called during backup start, there shouldn't be one already running */
9463 : Assert(!during_backup_start || sessionBackupState == SESSION_BACKUP_NONE);
9464 :
9465 18 : if (during_backup_start || sessionBackupState != SESSION_BACKUP_NONE)
9466 : {
9467 12 : WALInsertLockAcquireExclusive();
9468 : Assert(XLogCtl->Insert.runningBackups > 0);
9469 12 : XLogCtl->Insert.runningBackups--;
9470 :
9471 12 : sessionBackupState = SESSION_BACKUP_NONE;
9472 12 : WALInsertLockRelease();
9473 :
9474 12 : if (!during_backup_start)
9475 12 : ereport(WARNING,
9476 : errmsg("aborting backup due to backend exiting before pg_backup_stop was called"));
9477 : }
9478 18 : }
9479 :
9480 : /*
9481 : * Register a handler that will warn about unterminated backups at end of
9482 : * session, unless this has already been done.
9483 : */
9484 : void
9485 10 : register_persistent_abort_backup_handler(void)
9486 : {
9487 : static bool already_done = false;
9488 :
9489 10 : if (already_done)
9490 2 : return;
9491 8 : before_shmem_exit(do_pg_abort_backup, DatumGetBool(false));
9492 8 : already_done = true;
9493 : }
9494 :
9495 : /*
9496 : * Get latest WAL insert pointer
9497 : */
9498 : XLogRecPtr
9499 3950 : GetXLogInsertRecPtr(void)
9500 : {
9501 3950 : XLogCtlInsert *Insert = &XLogCtl->Insert;
9502 : uint64 current_bytepos;
9503 :
9504 3950 : SpinLockAcquire(&Insert->insertpos_lck);
9505 3950 : current_bytepos = Insert->CurrBytePos;
9506 3950 : SpinLockRelease(&Insert->insertpos_lck);
9507 :
9508 3950 : return XLogBytePosToRecPtr(current_bytepos);
9509 : }
9510 :
9511 : /*
9512 : * Get latest WAL write pointer
9513 : */
9514 : XLogRecPtr
9515 2668 : GetXLogWriteRecPtr(void)
9516 : {
9517 2668 : RefreshXLogWriteResult(LogwrtResult);
9518 :
9519 2668 : return LogwrtResult.Write;
9520 : }
9521 :
9522 : /*
9523 : * Returns the redo pointer of the last checkpoint or restartpoint. This is
9524 : * the oldest point in WAL that we still need, if we have to restart recovery.
9525 : */
9526 : void
9527 136 : GetOldestRestartPoint(XLogRecPtr *oldrecptr, TimeLineID *oldtli)
9528 : {
9529 136 : LWLockAcquire(ControlFileLock, LW_SHARED);
9530 136 : *oldrecptr = ControlFile->checkPointCopy.redo;
9531 136 : *oldtli = ControlFile->checkPointCopy.ThisTimeLineID;
9532 136 : LWLockRelease(ControlFileLock);
9533 136 : }
9534 :
9535 : /* Thin wrapper around ShutdownWalRcv(). */
9536 : void
9537 1864 : XLogShutdownWalRcv(void)
9538 : {
9539 1864 : ShutdownWalRcv();
9540 :
9541 1864 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9542 1864 : XLogCtl->InstallXLogFileSegmentActive = false;
9543 1864 : LWLockRelease(ControlFileLock);
9544 1864 : }
9545 :
9546 : /* Enable WAL file recycling and preallocation. */
9547 : void
9548 1950 : SetInstallXLogFileSegmentActive(void)
9549 : {
9550 1950 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9551 1950 : XLogCtl->InstallXLogFileSegmentActive = true;
9552 1950 : LWLockRelease(ControlFileLock);
9553 1950 : }
9554 :
9555 : bool
9556 0 : IsInstallXLogFileSegmentActive(void)
9557 : {
9558 : bool result;
9559 :
9560 0 : LWLockAcquire(ControlFileLock, LW_SHARED);
9561 0 : result = XLogCtl->InstallXLogFileSegmentActive;
9562 0 : LWLockRelease(ControlFileLock);
9563 :
9564 0 : return result;
9565 : }
9566 :
9567 : /*
9568 : * Update the WalWriterSleeping flag.
9569 : */
9570 : void
9571 846 : SetWalWriterSleeping(bool sleeping)
9572 : {
9573 846 : SpinLockAcquire(&XLogCtl->info_lck);
9574 846 : XLogCtl->WalWriterSleeping = sleeping;
9575 846 : SpinLockRelease(&XLogCtl->info_lck);
9576 846 : }
|
