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) 1996-2026, PostgreSQL Global Development Group
32 : * Portions Copyright (c) 1994, Regents of the University of California
33 : *
34 : * src/backend/access/transam/xlog.c
35 : *
36 : *-------------------------------------------------------------------------
37 : */
38 :
39 : #include "postgres.h"
40 :
41 : #include <ctype.h>
42 : #include <math.h>
43 : #include <time.h>
44 : #include <fcntl.h>
45 : #include <sys/stat.h>
46 : #include <sys/time.h>
47 : #include <unistd.h>
48 :
49 : #include "access/clog.h"
50 : #include "access/commit_ts.h"
51 : #include "access/heaptoast.h"
52 : #include "access/multixact.h"
53 : #include "access/rewriteheap.h"
54 : #include "access/subtrans.h"
55 : #include "access/timeline.h"
56 : #include "access/transam.h"
57 : #include "access/twophase.h"
58 : #include "access/xact.h"
59 : #include "access/xlog_internal.h"
60 : #include "access/xlogarchive.h"
61 : #include "access/xloginsert.h"
62 : #include "access/xlogreader.h"
63 : #include "access/xlogrecovery.h"
64 : #include "access/xlogutils.h"
65 : #include "access/xlogwait.h"
66 : #include "backup/basebackup.h"
67 : #include "catalog/catversion.h"
68 : #include "catalog/pg_control.h"
69 : #include "catalog/pg_database.h"
70 : #include "common/controldata_utils.h"
71 : #include "common/file_utils.h"
72 : #include "executor/instrument.h"
73 : #include "miscadmin.h"
74 : #include "pg_trace.h"
75 : #include "pgstat.h"
76 : #include "port/atomics.h"
77 : #include "postmaster/bgwriter.h"
78 : #include "postmaster/datachecksum_state.h"
79 : #include "postmaster/startup.h"
80 : #include "postmaster/walsummarizer.h"
81 : #include "postmaster/walwriter.h"
82 : #include "replication/origin.h"
83 : #include "replication/slot.h"
84 : #include "replication/slotsync.h"
85 : #include "replication/snapbuild.h"
86 : #include "replication/walreceiver.h"
87 : #include "replication/walsender.h"
88 : #include "storage/bufmgr.h"
89 : #include "storage/fd.h"
90 : #include "storage/ipc.h"
91 : #include "storage/large_object.h"
92 : #include "storage/latch.h"
93 : #include "storage/predicate.h"
94 : #include "storage/proc.h"
95 : #include "storage/procarray.h"
96 : #include "storage/procsignal.h"
97 : #include "storage/reinit.h"
98 : #include "storage/spin.h"
99 : #include "storage/subsystems.h"
100 : #include "storage/sync.h"
101 : #include "utils/guc_hooks.h"
102 : #include "utils/guc_tables.h"
103 : #include "utils/injection_point.h"
104 : #include "utils/pgstat_internal.h"
105 : #include "utils/ps_status.h"
106 : #include "utils/relmapper.h"
107 : #include "utils/snapmgr.h"
108 : #include "utils/timeout.h"
109 : #include "utils/timestamp.h"
110 : #include "utils/varlena.h"
111 : #include "utils/wait_event.h"
112 :
113 : #ifdef WAL_DEBUG
114 : #include "utils/memutils.h"
115 : #endif
116 :
117 : /* timeline ID to be used when bootstrapping */
118 : #define BootstrapTimeLineID 1
119 :
120 : /* User-settable parameters */
121 : int max_wal_size_mb = 1024; /* 1 GB */
122 : int min_wal_size_mb = 80; /* 80 MB */
123 : int wal_keep_size_mb = 0;
124 : int XLOGbuffers = -1;
125 : int XLogArchiveTimeout = 0;
126 : int XLogArchiveMode = ARCHIVE_MODE_OFF;
127 : char *XLogArchiveCommand = NULL;
128 : bool EnableHotStandby = false;
129 : bool fullPageWrites = true;
130 : bool wal_log_hints = false;
131 : int wal_compression = WAL_COMPRESSION_NONE;
132 : char *wal_consistency_checking_string = NULL;
133 : bool *wal_consistency_checking = NULL;
134 : bool wal_init_zero = true;
135 : bool wal_recycle = true;
136 : bool log_checkpoints = true;
137 : int wal_sync_method = DEFAULT_WAL_SYNC_METHOD;
138 : int wal_level = WAL_LEVEL_REPLICA;
139 : int CommitDelay = 0; /* precommit delay in microseconds */
140 : int CommitSiblings = 5; /* # concurrent xacts needed to sleep */
141 : int wal_retrieve_retry_interval = 5000;
142 : int max_slot_wal_keep_size_mb = -1;
143 : int wal_decode_buffer_size = 512 * 1024;
144 : bool track_wal_io_timing = false;
145 :
146 : #ifdef WAL_DEBUG
147 : bool XLOG_DEBUG = false;
148 : #endif
149 :
150 : int wal_segment_size = DEFAULT_XLOG_SEG_SIZE;
151 :
152 : /*
153 : * Number of WAL insertion locks to use. A higher value allows more insertions
154 : * to happen concurrently, but adds some CPU overhead to flushing the WAL,
155 : * which needs to iterate all the locks.
156 : */
157 : #define NUM_XLOGINSERT_LOCKS 8
158 :
159 : /*
160 : * Max distance from last checkpoint, before triggering a new xlog-based
161 : * checkpoint.
162 : */
163 : int CheckPointSegments;
164 :
165 : /* Estimated distance between checkpoints, in bytes */
166 : static double CheckPointDistanceEstimate = 0;
167 : static double PrevCheckPointDistance = 0;
168 :
169 : /*
170 : * Track whether there were any deferred checks for custom resource managers
171 : * specified in wal_consistency_checking.
172 : */
173 : static bool check_wal_consistency_checking_deferred = false;
174 :
175 : /*
176 : * GUC support
177 : */
178 : const struct config_enum_entry wal_sync_method_options[] = {
179 : {"fsync", WAL_SYNC_METHOD_FSYNC, false},
180 : #ifdef HAVE_FSYNC_WRITETHROUGH
181 : {"fsync_writethrough", WAL_SYNC_METHOD_FSYNC_WRITETHROUGH, false},
182 : #endif
183 : {"fdatasync", WAL_SYNC_METHOD_FDATASYNC, false},
184 : #ifdef O_SYNC
185 : {"open_sync", WAL_SYNC_METHOD_OPEN, false},
186 : #endif
187 : #ifdef O_DSYNC
188 : {"open_datasync", WAL_SYNC_METHOD_OPEN_DSYNC, false},
189 : #endif
190 : {NULL, 0, false}
191 : };
192 :
193 :
194 : /*
195 : * Although only "on", "off", and "always" are documented,
196 : * we accept all the likely variants of "on" and "off".
197 : */
198 : const struct config_enum_entry archive_mode_options[] = {
199 : {"always", ARCHIVE_MODE_ALWAYS, false},
200 : {"on", ARCHIVE_MODE_ON, false},
201 : {"off", ARCHIVE_MODE_OFF, false},
202 : {"true", ARCHIVE_MODE_ON, true},
203 : {"false", ARCHIVE_MODE_OFF, true},
204 : {"yes", ARCHIVE_MODE_ON, true},
205 : {"no", ARCHIVE_MODE_OFF, true},
206 : {"1", ARCHIVE_MODE_ON, true},
207 : {"0", ARCHIVE_MODE_OFF, true},
208 : {NULL, 0, false}
209 : };
210 :
211 : /*
212 : * Statistics for current checkpoint are collected in this global struct.
213 : * Because only the checkpointer or a stand-alone backend can perform
214 : * checkpoints, this will be unused in normal backends.
215 : */
216 : CheckpointStatsData CheckpointStats;
217 :
218 : /*
219 : * During recovery, lastFullPageWrites keeps track of full_page_writes that
220 : * the replayed WAL records indicate. It's initialized with full_page_writes
221 : * that the recovery starting checkpoint record indicates, and then updated
222 : * each time XLOG_FPW_CHANGE record is replayed.
223 : */
224 : static bool lastFullPageWrites;
225 :
226 : /*
227 : * Local copy of the state tracked by SharedRecoveryState in shared memory,
228 : * It is false if SharedRecoveryState is RECOVERY_STATE_DONE. True actually
229 : * means "not known, need to check the shared state".
230 : */
231 : static bool LocalRecoveryInProgress = true;
232 :
233 : /*
234 : * Local state for XLogInsertAllowed():
235 : * 1: unconditionally allowed to insert XLOG
236 : * 0: unconditionally not allowed to insert XLOG
237 : * -1: must check RecoveryInProgress(); disallow until it is false
238 : * Most processes start with -1 and transition to 1 after seeing that recovery
239 : * is not in progress. But we can also force the value for special cases.
240 : * The coding in XLogInsertAllowed() depends on the first two of these states
241 : * being numerically the same as bool true and false.
242 : */
243 : static int LocalXLogInsertAllowed = -1;
244 :
245 : /*
246 : * ProcLastRecPtr points to the start of the last XLOG record inserted by the
247 : * current backend. It is updated for all inserts. XactLastRecEnd points to
248 : * end+1 of the last record, and is reset when we end a top-level transaction,
249 : * or start a new one; so it can be used to tell if the current transaction has
250 : * created any XLOG records.
251 : *
252 : * While in parallel mode, this may not be fully up to date. When committing,
253 : * a transaction can assume this covers all xlog records written either by the
254 : * user backend or by any parallel worker which was present at any point during
255 : * the transaction. But when aborting, or when still in parallel mode, other
256 : * parallel backends may have written WAL records at later LSNs than the value
257 : * stored here. The parallel leader advances its own copy, when necessary,
258 : * in WaitForParallelWorkersToFinish.
259 : */
260 : XLogRecPtr ProcLastRecPtr = InvalidXLogRecPtr;
261 : XLogRecPtr XactLastRecEnd = InvalidXLogRecPtr;
262 : XLogRecPtr XactLastCommitEnd = InvalidXLogRecPtr;
263 :
264 : /*
265 : * RedoRecPtr is this backend's local copy of the REDO record pointer
266 : * (which is almost but not quite the same as a pointer to the most recent
267 : * CHECKPOINT record). We update this from the shared-memory copy,
268 : * XLogCtl->Insert.RedoRecPtr, whenever we can safely do so (ie, when we
269 : * hold an insertion lock). See XLogInsertRecord for details. We are also
270 : * allowed to update from XLogCtl->RedoRecPtr if we hold the info_lck;
271 : * see GetRedoRecPtr.
272 : *
273 : * NB: Code that uses this variable must be prepared not only for the
274 : * possibility that it may be arbitrarily out of date, but also for the
275 : * possibility that it might be set to InvalidXLogRecPtr. We used to
276 : * initialize it as a side effect of the first call to RecoveryInProgress(),
277 : * which meant that most code that might use it could assume that it had a
278 : * real if perhaps stale value. That's no longer the case.
279 : */
280 : static XLogRecPtr RedoRecPtr;
281 :
282 : /*
283 : * doPageWrites is this backend's local copy of (fullPageWrites ||
284 : * runningBackups > 0). It is used together with RedoRecPtr to decide whether
285 : * a full-page image of a page need to be taken.
286 : *
287 : * NB: Initially this is false, and there's no guarantee that it will be
288 : * initialized to any other value before it is first used. Any code that
289 : * makes use of it must recheck the value after obtaining a WALInsertLock,
290 : * and respond appropriately if it turns out that the previous value wasn't
291 : * accurate.
292 : */
293 : static bool doPageWrites;
294 :
295 : /*----------
296 : * Shared-memory data structures for XLOG control
297 : *
298 : * LogwrtRqst indicates a byte position that we need to write and/or fsync
299 : * the log up to (all records before that point must be written or fsynced).
300 : * The positions already written/fsynced are maintained in logWriteResult
301 : * and logFlushResult using atomic access.
302 : * In addition to the shared variable, each backend has a private copy of
303 : * both in LogwrtResult, which is updated when convenient.
304 : *
305 : * The request bookkeeping is simpler: there is a shared XLogCtl->LogwrtRqst
306 : * (protected by info_lck), but we don't need to cache any copies of it.
307 : *
308 : * info_lck is only held long enough to read/update the protected variables,
309 : * so it's a plain spinlock. The other locks are held longer (potentially
310 : * over I/O operations), so we use LWLocks for them. These locks are:
311 : *
312 : * WALBufMappingLock: must be held to replace a page in the WAL buffer cache.
313 : * It is only held while initializing and changing the mapping. If the
314 : * contents of the buffer being replaced haven't been written yet, the mapping
315 : * lock is released while the write is done, and reacquired afterwards.
316 : *
317 : * WALWriteLock: must be held to write WAL buffers to disk (XLogWrite or
318 : * XLogFlush).
319 : *
320 : * ControlFileLock: must be held to read/update control file or create
321 : * new log file.
322 : *
323 : *----------
324 : */
325 :
326 : typedef struct XLogwrtRqst
327 : {
328 : XLogRecPtr Write; /* last byte + 1 to write out */
329 : XLogRecPtr Flush; /* last byte + 1 to flush */
330 : } XLogwrtRqst;
331 :
332 : typedef struct XLogwrtResult
333 : {
334 : XLogRecPtr Write; /* last byte + 1 written out */
335 : XLogRecPtr Flush; /* last byte + 1 flushed */
336 : } XLogwrtResult;
337 :
338 : /*
339 : * Inserting to WAL is protected by a small fixed number of WAL insertion
340 : * locks. To insert to the WAL, you must hold one of the locks - it doesn't
341 : * matter which one. To lock out other concurrent insertions, you must hold
342 : * of them. Each WAL insertion lock consists of a lightweight lock, plus an
343 : * indicator of how far the insertion has progressed (insertingAt).
344 : *
345 : * The insertingAt values are read when a process wants to flush WAL from
346 : * the in-memory buffers to disk, to check that all the insertions to the
347 : * region the process is about to write out have finished. You could simply
348 : * wait for all currently in-progress insertions to finish, but the
349 : * insertingAt indicator allows you to ignore insertions to later in the WAL,
350 : * so that you only wait for the insertions that are modifying the buffers
351 : * you're about to write out.
352 : *
353 : * This isn't just an optimization. If all the WAL buffers are dirty, an
354 : * inserter that's holding a WAL insert lock might need to evict an old WAL
355 : * buffer, which requires flushing the WAL. If it's possible for an inserter
356 : * to block on another inserter unnecessarily, deadlock can arise when two
357 : * inserters holding a WAL insert lock wait for each other to finish their
358 : * insertion.
359 : *
360 : * Small WAL records that don't cross a page boundary never update the value,
361 : * the WAL record is just copied to the page and the lock is released. But
362 : * to avoid the deadlock-scenario explained above, the indicator is always
363 : * updated before sleeping while holding an insertion lock.
364 : *
365 : * lastImportantAt contains the LSN of the last important WAL record inserted
366 : * using a given lock. This value is used to detect if there has been
367 : * important WAL activity since the last time some action, like a checkpoint,
368 : * was performed - allowing to not repeat the action if not. The LSN is
369 : * updated for all insertions, unless the XLOG_MARK_UNIMPORTANT flag was
370 : * set. lastImportantAt is never cleared, only overwritten by the LSN of newer
371 : * records. Tracking the WAL activity directly in WALInsertLock has the
372 : * advantage of not needing any additional locks to update the value.
373 : */
374 : typedef struct
375 : {
376 : LWLock lock;
377 : pg_atomic_uint64 insertingAt;
378 : XLogRecPtr lastImportantAt;
379 : } WALInsertLock;
380 :
381 : /*
382 : * All the WAL insertion locks are allocated as an array in shared memory. We
383 : * force the array stride to be a power of 2, which saves a few cycles in
384 : * indexing, but more importantly also ensures that individual slots don't
385 : * cross cache line boundaries. (Of course, we have to also ensure that the
386 : * array start address is suitably aligned.)
387 : */
388 : typedef union WALInsertLockPadded
389 : {
390 : WALInsertLock l;
391 : char pad[PG_CACHE_LINE_SIZE];
392 : } WALInsertLockPadded;
393 :
394 : /*
395 : * Session status of running backup, used for sanity checks in SQL-callable
396 : * functions to start and stop backups.
397 : */
398 : static SessionBackupState sessionBackupState = SESSION_BACKUP_NONE;
399 :
400 : /*
401 : * Shared state data for WAL insertion.
402 : */
403 : typedef struct XLogCtlInsert
404 : {
405 : slock_t insertpos_lck; /* protects CurrBytePos and PrevBytePos */
406 :
407 : /*
408 : * CurrBytePos is the end of reserved WAL. The next record will be
409 : * inserted at that position. PrevBytePos is the start position of the
410 : * previously inserted (or rather, reserved) record - it is copied to the
411 : * prev-link of the next record. These are stored as "usable byte
412 : * positions" rather than XLogRecPtrs (see XLogBytePosToRecPtr()).
413 : */
414 : uint64 CurrBytePos;
415 : uint64 PrevBytePos;
416 :
417 : /*
418 : * Make sure the above heavily-contended spinlock and byte positions are
419 : * on their own cache line. In particular, the RedoRecPtr and full page
420 : * write variables below should be on a different cache line. They are
421 : * read on every WAL insertion, but updated rarely, and we don't want
422 : * those reads to steal the cache line containing Curr/PrevBytePos.
423 : */
424 : char pad[PG_CACHE_LINE_SIZE];
425 :
426 : /*
427 : * fullPageWrites is the authoritative value used by all backends to
428 : * determine whether to write full-page image to WAL. This shared value,
429 : * instead of the process-local fullPageWrites, is required because, when
430 : * full_page_writes is changed by SIGHUP, we must WAL-log it before it
431 : * actually affects WAL-logging by backends. Checkpointer sets at startup
432 : * or after SIGHUP.
433 : *
434 : * To read these fields, you must hold an insertion lock. To modify them,
435 : * you must hold ALL the locks.
436 : */
437 : XLogRecPtr RedoRecPtr; /* current redo point for insertions */
438 : bool fullPageWrites;
439 :
440 : /*
441 : * runningBackups is a counter indicating the number of backups currently
442 : * in progress. lastBackupStart is the latest checkpoint redo location
443 : * used as a starting point for an online backup.
444 : */
445 : int runningBackups;
446 : XLogRecPtr lastBackupStart;
447 :
448 : /*
449 : * WAL insertion locks.
450 : */
451 : WALInsertLockPadded *WALInsertLocks;
452 : } XLogCtlInsert;
453 :
454 : /*
455 : * Total shared-memory state for XLOG.
456 : */
457 : typedef struct XLogCtlData
458 : {
459 : XLogCtlInsert Insert;
460 :
461 : /* Protected by info_lck: */
462 : XLogwrtRqst LogwrtRqst;
463 : XLogRecPtr RedoRecPtr; /* a recent copy of Insert->RedoRecPtr */
464 : XLogRecPtr asyncXactLSN; /* LSN of newest async commit/abort */
465 : XLogRecPtr replicationSlotMinLSN; /* oldest LSN needed by any slot */
466 :
467 : XLogSegNo lastRemovedSegNo; /* latest removed/recycled XLOG segment */
468 :
469 : /* Fake LSN counter, for unlogged relations. */
470 : pg_atomic_uint64 unloggedLSN;
471 :
472 : /* Time and LSN of last xlog segment switch. Protected by WALWriteLock. */
473 : pg_time_t lastSegSwitchTime;
474 : XLogRecPtr lastSegSwitchLSN;
475 :
476 : /* These are accessed using atomics -- info_lck not needed */
477 : pg_atomic_uint64 logInsertResult; /* last byte + 1 inserted to buffers */
478 : pg_atomic_uint64 logWriteResult; /* last byte + 1 written out */
479 : pg_atomic_uint64 logFlushResult; /* last byte + 1 flushed */
480 :
481 : /*
482 : * Latest initialized page in the cache (last byte position + 1).
483 : *
484 : * To change the identity of a buffer (and InitializedUpTo), you need to
485 : * hold WALBufMappingLock. To change the identity of a buffer that's
486 : * still dirty, the old page needs to be written out first, and for that
487 : * you need WALWriteLock, and you need to ensure that there are no
488 : * in-progress insertions to the page by calling
489 : * WaitXLogInsertionsToFinish().
490 : */
491 : XLogRecPtr InitializedUpTo;
492 :
493 : /*
494 : * These values do not change after startup, although the pointed-to pages
495 : * and xlblocks values certainly do. xlblocks values are protected by
496 : * WALBufMappingLock.
497 : */
498 : char *pages; /* buffers for unwritten XLOG pages */
499 : pg_atomic_uint64 *xlblocks; /* 1st byte ptr-s + XLOG_BLCKSZ */
500 : int XLogCacheBlck; /* highest allocated xlog buffer index */
501 :
502 : /*
503 : * InsertTimeLineID is the timeline into which new WAL is being inserted
504 : * and flushed. It is zero during recovery, and does not change once set.
505 : *
506 : * If we create a new timeline when the system was started up,
507 : * PrevTimeLineID is the old timeline's ID that we forked off from.
508 : * Otherwise it's equal to InsertTimeLineID.
509 : *
510 : * We set these fields while holding info_lck. Most that reads these
511 : * values knows that recovery is no longer in progress and so can safely
512 : * read the value without a lock, but code that could be run either during
513 : * or after recovery can take info_lck while reading these values.
514 : */
515 : TimeLineID InsertTimeLineID;
516 : TimeLineID PrevTimeLineID;
517 :
518 : /*
519 : * SharedRecoveryState indicates if we're still in crash or archive
520 : * recovery. Protected by info_lck.
521 : */
522 : RecoveryState SharedRecoveryState;
523 :
524 : /*
525 : * InstallXLogFileSegmentActive indicates whether the checkpointer should
526 : * arrange for future segments by recycling and/or PreallocXlogFiles().
527 : * Protected by ControlFileLock. Only the startup process changes it. If
528 : * true, anyone can use InstallXLogFileSegment(). If false, the startup
529 : * process owns the exclusive right to install segments, by reading from
530 : * the archive and possibly replacing existing files.
531 : */
532 : bool InstallXLogFileSegmentActive;
533 :
534 : /*
535 : * WalWriterSleeping indicates whether the WAL writer is currently in
536 : * low-power mode (and hence should be nudged if an async commit occurs).
537 : * Protected by info_lck.
538 : */
539 : bool WalWriterSleeping;
540 :
541 : /*
542 : * During recovery, we keep a copy of the latest checkpoint record here.
543 : * lastCheckPointRecPtr points to start of checkpoint record and
544 : * lastCheckPointEndPtr points to end+1 of checkpoint record. Used by the
545 : * checkpointer when it wants to create a restartpoint.
546 : *
547 : * Protected by info_lck.
548 : */
549 : XLogRecPtr lastCheckPointRecPtr;
550 : XLogRecPtr lastCheckPointEndPtr;
551 : CheckPoint lastCheckPoint;
552 :
553 : /*
554 : * lastFpwDisableRecPtr points to the start of the last replayed
555 : * XLOG_FPW_CHANGE record that instructs full_page_writes is disabled.
556 : */
557 : XLogRecPtr lastFpwDisableRecPtr;
558 :
559 : /* last data_checksum_version we've seen */
560 : uint32 data_checksum_version;
561 :
562 : slock_t info_lck; /* locks shared variables shown above */
563 : } XLogCtlData;
564 :
565 : /*
566 : * Classification of XLogInsertRecord operations.
567 : */
568 : typedef enum
569 : {
570 : WALINSERT_NORMAL,
571 : WALINSERT_SPECIAL_SWITCH,
572 : WALINSERT_SPECIAL_CHECKPOINT
573 : } WalInsertClass;
574 :
575 : static XLogCtlData *XLogCtl = NULL;
576 :
577 : /* a private copy of XLogCtl->Insert.WALInsertLocks, for convenience */
578 : static WALInsertLockPadded *WALInsertLocks = NULL;
579 :
580 : /*
581 : * We maintain an image of pg_control in shared memory.
582 : */
583 : static ControlFileData *LocalControlFile = NULL;
584 : static ControlFileData *ControlFile = NULL;
585 :
586 : static void XLOGShmemRequest(void *arg);
587 : static void XLOGShmemInit(void *arg);
588 : static void XLOGShmemAttach(void *arg);
589 :
590 : const ShmemCallbacks XLOGShmemCallbacks = {
591 : .request_fn = XLOGShmemRequest,
592 : .init_fn = XLOGShmemInit,
593 : .attach_fn = XLOGShmemAttach,
594 : };
595 :
596 : /*
597 : * Calculate the amount of space left on the page after 'endptr'. Beware
598 : * multiple evaluation!
599 : */
600 : #define INSERT_FREESPACE(endptr) \
601 : (((endptr) % XLOG_BLCKSZ == 0) ? 0 : (XLOG_BLCKSZ - (endptr) % XLOG_BLCKSZ))
602 :
603 : /* Macro to advance to next buffer index. */
604 : #define NextBufIdx(idx) \
605 : (((idx) == XLogCtl->XLogCacheBlck) ? 0 : ((idx) + 1))
606 :
607 : /*
608 : * XLogRecPtrToBufIdx returns the index of the WAL buffer that holds, or
609 : * would hold if it was in cache, the page containing 'recptr'.
610 : */
611 : #define XLogRecPtrToBufIdx(recptr) \
612 : (((recptr) / XLOG_BLCKSZ) % (XLogCtl->XLogCacheBlck + 1))
613 :
614 : /*
615 : * These are the number of bytes in a WAL page usable for WAL data.
616 : */
617 : #define UsableBytesInPage (XLOG_BLCKSZ - SizeOfXLogShortPHD)
618 :
619 : /*
620 : * Convert values of GUCs measured in megabytes to equiv. segment count.
621 : * Rounds down.
622 : */
623 : #define ConvertToXSegs(x, segsize) XLogMBVarToSegs((x), (segsize))
624 :
625 : /* The number of bytes in a WAL segment usable for WAL data. */
626 : static int UsableBytesInSegment;
627 :
628 : /*
629 : * Private, possibly out-of-date copy of shared LogwrtResult.
630 : * See discussion above.
631 : */
632 : static XLogwrtResult LogwrtResult = {0, 0};
633 :
634 : /*
635 : * Update local copy of shared XLogCtl->log{Write,Flush}Result
636 : *
637 : * It's critical that Flush always trails Write, so the order of the reads is
638 : * important, as is the barrier. See also XLogWrite.
639 : */
640 : #define RefreshXLogWriteResult(_target) \
641 : do { \
642 : _target.Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult); \
643 : pg_read_barrier(); \
644 : _target.Write = pg_atomic_read_u64(&XLogCtl->logWriteResult); \
645 : } while (0)
646 :
647 : /*
648 : * openLogFile is -1 or a kernel FD for an open log file segment.
649 : * openLogSegNo identifies the segment, and openLogTLI the corresponding TLI.
650 : * These variables are only used to write the XLOG, and so will normally refer
651 : * to the active segment.
652 : *
653 : * Note: call Reserve/ReleaseExternalFD to track consumption of this FD.
654 : */
655 : static int openLogFile = -1;
656 : static XLogSegNo openLogSegNo = 0;
657 : static TimeLineID openLogTLI = 0;
658 :
659 : /*
660 : * Local copies of equivalent fields in the control file. When running
661 : * crash recovery, LocalMinRecoveryPoint is set to InvalidXLogRecPtr as we
662 : * expect to replay all the WAL available, and updateMinRecoveryPoint is
663 : * switched to false to prevent any updates while replaying records.
664 : * Those values are kept consistent as long as crash recovery runs.
665 : */
666 : static XLogRecPtr LocalMinRecoveryPoint;
667 : static TimeLineID LocalMinRecoveryPointTLI;
668 : static bool updateMinRecoveryPoint = true;
669 :
670 : /*
671 : * Local state for ControlFile data_checksum_version. After initialization
672 : * this is only updated when absorbing a procsignal barrier during interrupt
673 : * processing. The reason for keeping a copy in backend-private memory is to
674 : * avoid locking for interrogating the data checksum state. Possible values
675 : * are the data checksum versions defined in storage/checksum.h.
676 : */
677 : static ChecksumStateType LocalDataChecksumState = 0;
678 :
679 : /*
680 : * Variable backing the GUC, keep it in sync with LocalDataChecksumState.
681 : * See SetLocalDataChecksumState().
682 : */
683 : int data_checksums = 0;
684 :
685 : /* For WALInsertLockAcquire/Release functions */
686 : static int MyLockNo = 0;
687 : static bool holdingAllLocks = false;
688 :
689 : #ifdef WAL_DEBUG
690 : static MemoryContext walDebugCxt = NULL;
691 : #endif
692 :
693 : static void CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI,
694 : XLogRecPtr EndOfLog,
695 : TimeLineID newTLI);
696 : static void CheckRequiredParameterValues(void);
697 : static void XLogReportParameters(void);
698 : static int LocalSetXLogInsertAllowed(void);
699 : static void CreateEndOfRecoveryRecord(void);
700 : static XLogRecPtr CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn,
701 : XLogRecPtr pagePtr,
702 : TimeLineID newTLI);
703 : static void CheckPointGuts(XLogRecPtr checkPointRedo, int flags);
704 : static void KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo);
705 :
706 : static void AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli,
707 : bool opportunistic);
708 : static void XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible);
709 : static bool InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
710 : bool find_free, XLogSegNo max_segno,
711 : TimeLineID tli);
712 : static void XLogFileClose(void);
713 : static void PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli);
714 : static void RemoveTempXlogFiles(void);
715 : static void RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr,
716 : XLogRecPtr endptr, TimeLineID insertTLI);
717 : static void RemoveXlogFile(const struct dirent *segment_de,
718 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
719 : TimeLineID insertTLI);
720 : static void UpdateLastRemovedPtr(char *filename);
721 : static void ValidateXLOGDirectoryStructure(void);
722 : static void CleanupBackupHistory(void);
723 : static void UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force);
724 : static bool PerformRecoveryXLogAction(void);
725 : static void InitControlFile(uint64 sysidentifier, uint32 data_checksum_version);
726 : static void WriteControlFile(void);
727 : static void ReadControlFile(void);
728 : static void UpdateControlFile(void);
729 : static char *str_time(pg_time_t tnow, char *buf, size_t bufsize);
730 :
731 : static int get_sync_bit(int method);
732 :
733 : static void CopyXLogRecordToWAL(int write_len, bool isLogSwitch,
734 : XLogRecData *rdata,
735 : XLogRecPtr StartPos, XLogRecPtr EndPos,
736 : TimeLineID tli);
737 : static void ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos,
738 : XLogRecPtr *EndPos, XLogRecPtr *PrevPtr);
739 : static bool ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos,
740 : XLogRecPtr *PrevPtr);
741 : static XLogRecPtr WaitXLogInsertionsToFinish(XLogRecPtr upto);
742 : static char *GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli);
743 : static XLogRecPtr XLogBytePosToRecPtr(uint64 bytepos);
744 : static XLogRecPtr XLogBytePosToEndRecPtr(uint64 bytepos);
745 : static uint64 XLogRecPtrToBytePos(XLogRecPtr ptr);
746 :
747 : static void WALInsertLockAcquire(void);
748 : static void WALInsertLockAcquireExclusive(void);
749 : static void WALInsertLockRelease(void);
750 : static void WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt);
751 :
752 : static void XLogChecksums(uint32 new_type);
753 :
754 : /*
755 : * Insert an XLOG record represented by an already-constructed chain of data
756 : * chunks. This is a low-level routine; to construct the WAL record header
757 : * and data, use the higher-level routines in xloginsert.c.
758 : *
759 : * If 'fpw_lsn' is valid, it is the oldest LSN among the pages that this
760 : * WAL record applies to, that were not included in the record as full page
761 : * images. If fpw_lsn <= RedoRecPtr, the function does not perform the
762 : * insertion and returns InvalidXLogRecPtr. The caller can then recalculate
763 : * which pages need a full-page image, and retry. If fpw_lsn is invalid, the
764 : * record is always inserted.
765 : *
766 : * 'flags' gives more in-depth control on the record being inserted. See
767 : * XLogSetRecordFlags() for details.
768 : *
769 : * 'topxid_included' tells whether the top-transaction id is logged along with
770 : * current subtransaction. See XLogRecordAssemble().
771 : *
772 : * The first XLogRecData in the chain must be for the record header, and its
773 : * data must be MAXALIGNed. XLogInsertRecord fills in the xl_prev and
774 : * xl_crc fields in the header, the rest of the header must already be filled
775 : * by the caller.
776 : *
777 : * Returns XLOG pointer to end of record (beginning of next record).
778 : * This can be used as LSN for data pages affected by the logged action.
779 : * (LSN is the XLOG point up to which the XLOG must be flushed to disk
780 : * before the data page can be written out. This implements the basic
781 : * WAL rule "write the log before the data".)
782 : */
783 : XLogRecPtr
784 24738312 : XLogInsertRecord(XLogRecData *rdata,
785 : XLogRecPtr fpw_lsn,
786 : uint8 flags,
787 : int num_fpi,
788 : uint64 fpi_bytes,
789 : bool topxid_included)
790 : {
791 24738312 : XLogCtlInsert *Insert = &XLogCtl->Insert;
792 : pg_crc32c rdata_crc;
793 : bool inserted;
794 24738312 : XLogRecord *rechdr = (XLogRecord *) rdata->data;
795 24738312 : uint8 info = rechdr->xl_info & ~XLR_INFO_MASK;
796 24738312 : WalInsertClass class = WALINSERT_NORMAL;
797 : XLogRecPtr StartPos;
798 : XLogRecPtr EndPos;
799 24738312 : bool prevDoPageWrites = doPageWrites;
800 : TimeLineID insertTLI;
801 :
802 : /* Does this record type require special handling? */
803 24738312 : if (unlikely(rechdr->xl_rmid == RM_XLOG_ID))
804 : {
805 323591 : if (info == XLOG_SWITCH)
806 823 : class = WALINSERT_SPECIAL_SWITCH;
807 322768 : else if (info == XLOG_CHECKPOINT_REDO)
808 996 : class = WALINSERT_SPECIAL_CHECKPOINT;
809 : }
810 :
811 : /* we assume that all of the record header is in the first chunk */
812 : Assert(rdata->len >= SizeOfXLogRecord);
813 :
814 : /* cross-check on whether we should be here or not */
815 24738312 : if (!XLogInsertAllowed())
816 0 : elog(ERROR, "cannot make new WAL entries during recovery");
817 :
818 : /*
819 : * Given that we're not in recovery, InsertTimeLineID is set and can't
820 : * change, so we can read it without a lock.
821 : */
822 24738312 : insertTLI = XLogCtl->InsertTimeLineID;
823 :
824 : /*----------
825 : *
826 : * We have now done all the preparatory work we can without holding a
827 : * lock or modifying shared state. From here on, inserting the new WAL
828 : * record to the shared WAL buffer cache is a two-step process:
829 : *
830 : * 1. Reserve the right amount of space from the WAL. The current head of
831 : * reserved space is kept in Insert->CurrBytePos, and is protected by
832 : * insertpos_lck.
833 : *
834 : * 2. Copy the record to the reserved WAL space. This involves finding the
835 : * correct WAL buffer containing the reserved space, and copying the
836 : * record in place. This can be done concurrently in multiple processes.
837 : *
838 : * To keep track of which insertions are still in-progress, each concurrent
839 : * inserter acquires an insertion lock. In addition to just indicating that
840 : * an insertion is in progress, the lock tells others how far the inserter
841 : * has progressed. There is a small fixed number of insertion locks,
842 : * determined by NUM_XLOGINSERT_LOCKS. When an inserter crosses a page
843 : * boundary, it updates the value stored in the lock to the how far it has
844 : * inserted, to allow the previous buffer to be flushed.
845 : *
846 : * Holding onto an insertion lock also protects RedoRecPtr and
847 : * fullPageWrites from changing until the insertion is finished.
848 : *
849 : * Step 2 can usually be done completely in parallel. If the required WAL
850 : * page is not initialized yet, you have to grab WALBufMappingLock to
851 : * initialize it, but the WAL writer tries to do that ahead of insertions
852 : * to avoid that from happening in the critical path.
853 : *
854 : *----------
855 : */
856 24738312 : START_CRIT_SECTION();
857 :
858 24738312 : if (likely(class == WALINSERT_NORMAL))
859 : {
860 24736493 : WALInsertLockAcquire();
861 :
862 : /*
863 : * Check to see if my copy of RedoRecPtr is out of date. If so, may
864 : * have to go back and have the caller recompute everything. This can
865 : * only happen just after a checkpoint, so it's better to be slow in
866 : * this case and fast otherwise.
867 : *
868 : * Also check to see if fullPageWrites was just turned on or there's a
869 : * running backup (which forces full-page writes); if we weren't
870 : * already doing full-page writes then go back and recompute.
871 : *
872 : * If we aren't doing full-page writes then RedoRecPtr doesn't
873 : * actually affect the contents of the XLOG record, so we'll update
874 : * our local copy but not force a recomputation. (If doPageWrites was
875 : * just turned off, we could recompute the record without full pages,
876 : * but we choose not to bother.)
877 : */
878 24736493 : if (RedoRecPtr != Insert->RedoRecPtr)
879 : {
880 : Assert(RedoRecPtr < Insert->RedoRecPtr);
881 8156 : RedoRecPtr = Insert->RedoRecPtr;
882 : }
883 24736493 : doPageWrites = (Insert->fullPageWrites || Insert->runningBackups > 0);
884 :
885 24736493 : if (doPageWrites &&
886 22432949 : (!prevDoPageWrites ||
887 20980594 : (XLogRecPtrIsValid(fpw_lsn) && fpw_lsn <= RedoRecPtr)))
888 : {
889 : /*
890 : * Oops, some buffer now needs to be backed up that the caller
891 : * didn't back up. Start over.
892 : */
893 8896 : WALInsertLockRelease();
894 8896 : END_CRIT_SECTION();
895 8896 : return InvalidXLogRecPtr;
896 : }
897 :
898 : /*
899 : * Reserve space for the record in the WAL. This also sets the xl_prev
900 : * pointer.
901 : */
902 24727597 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
903 : &rechdr->xl_prev);
904 :
905 : /* Normal records are always inserted. */
906 24727597 : inserted = true;
907 : }
908 1819 : else if (class == WALINSERT_SPECIAL_SWITCH)
909 : {
910 : /*
911 : * In order to insert an XLOG_SWITCH record, we need to hold all of
912 : * the WAL insertion locks, not just one, so that no one else can
913 : * begin inserting a record until we've figured out how much space
914 : * remains in the current WAL segment and claimed all of it.
915 : *
916 : * Nonetheless, this case is simpler than the normal cases handled
917 : * below, which must check for changes in doPageWrites and RedoRecPtr.
918 : * Those checks are only needed for records that can contain buffer
919 : * references, and an XLOG_SWITCH record never does.
920 : */
921 : Assert(!XLogRecPtrIsValid(fpw_lsn));
922 823 : WALInsertLockAcquireExclusive();
923 823 : inserted = ReserveXLogSwitch(&StartPos, &EndPos, &rechdr->xl_prev);
924 : }
925 : else
926 : {
927 : Assert(class == WALINSERT_SPECIAL_CHECKPOINT);
928 :
929 : /*
930 : * We need to update both the local and shared copies of RedoRecPtr,
931 : * which means that we need to hold all the WAL insertion locks.
932 : * However, there can't be any buffer references, so as above, we need
933 : * not check RedoRecPtr before inserting the record; we just need to
934 : * update it afterwards.
935 : */
936 : Assert(!XLogRecPtrIsValid(fpw_lsn));
937 996 : WALInsertLockAcquireExclusive();
938 996 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
939 : &rechdr->xl_prev);
940 996 : RedoRecPtr = Insert->RedoRecPtr = StartPos;
941 996 : inserted = true;
942 : }
943 :
944 24729416 : if (inserted)
945 : {
946 : /*
947 : * Now that xl_prev has been filled in, calculate CRC of the record
948 : * header.
949 : */
950 24729353 : rdata_crc = rechdr->xl_crc;
951 24729353 : COMP_CRC32C(rdata_crc, rechdr, offsetof(XLogRecord, xl_crc));
952 24729353 : FIN_CRC32C(rdata_crc);
953 24729353 : rechdr->xl_crc = rdata_crc;
954 :
955 : /*
956 : * All the record data, including the header, is now ready to be
957 : * inserted. Copy the record in the space reserved.
958 : */
959 24729353 : CopyXLogRecordToWAL(rechdr->xl_tot_len,
960 : class == WALINSERT_SPECIAL_SWITCH, rdata,
961 : StartPos, EndPos, insertTLI);
962 :
963 : /*
964 : * Unless record is flagged as not important, update LSN of last
965 : * important record in the current slot. When holding all locks, just
966 : * update the first one.
967 : */
968 24729353 : if ((flags & XLOG_MARK_UNIMPORTANT) == 0)
969 : {
970 24570984 : int lockno = holdingAllLocks ? 0 : MyLockNo;
971 :
972 24570984 : WALInsertLocks[lockno].l.lastImportantAt = StartPos;
973 : }
974 : }
975 : else
976 : {
977 : /*
978 : * This was an xlog-switch record, but the current insert location was
979 : * already exactly at the beginning of a segment, so there was no need
980 : * to do anything.
981 : */
982 : }
983 :
984 : /*
985 : * Done! Let others know that we're finished.
986 : */
987 24729416 : WALInsertLockRelease();
988 :
989 24729416 : END_CRIT_SECTION();
990 :
991 24729416 : MarkCurrentTransactionIdLoggedIfAny();
992 :
993 : /*
994 : * Mark top transaction id is logged (if needed) so that we should not try
995 : * to log it again with the next WAL record in the current subtransaction.
996 : */
997 24729416 : if (topxid_included)
998 223 : MarkSubxactTopXidLogged();
999 :
1000 : /*
1001 : * Update shared LogwrtRqst.Write, if we crossed page boundary.
1002 : */
1003 24729416 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
1004 : {
1005 1890757 : SpinLockAcquire(&XLogCtl->info_lck);
1006 : /* advance global request to include new block(s) */
1007 1890757 : if (XLogCtl->LogwrtRqst.Write < EndPos)
1008 1826121 : XLogCtl->LogwrtRqst.Write = EndPos;
1009 1890757 : SpinLockRelease(&XLogCtl->info_lck);
1010 1890757 : RefreshXLogWriteResult(LogwrtResult);
1011 : }
1012 :
1013 : /*
1014 : * If this was an XLOG_SWITCH record, flush the record and the empty
1015 : * padding space that fills the rest of the segment, and perform
1016 : * end-of-segment actions (eg, notifying archiver).
1017 : */
1018 24729416 : if (class == WALINSERT_SPECIAL_SWITCH)
1019 : {
1020 : TRACE_POSTGRESQL_WAL_SWITCH();
1021 823 : XLogFlush(EndPos);
1022 :
1023 : /*
1024 : * Even though we reserved the rest of the segment for us, which is
1025 : * reflected in EndPos, we return a pointer to just the end of the
1026 : * xlog-switch record.
1027 : */
1028 823 : if (inserted)
1029 : {
1030 760 : EndPos = StartPos + SizeOfXLogRecord;
1031 760 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
1032 : {
1033 0 : uint64 offset = XLogSegmentOffset(EndPos, wal_segment_size);
1034 :
1035 0 : if (offset == EndPos % XLOG_BLCKSZ)
1036 0 : EndPos += SizeOfXLogLongPHD;
1037 : else
1038 0 : EndPos += SizeOfXLogShortPHD;
1039 : }
1040 : }
1041 : }
1042 :
1043 : #ifdef WAL_DEBUG
1044 : if (XLOG_DEBUG)
1045 : {
1046 : static XLogReaderState *debug_reader = NULL;
1047 : XLogRecord *record;
1048 : DecodedXLogRecord *decoded;
1049 : StringInfoData buf;
1050 : StringInfoData recordBuf;
1051 : char *errormsg = NULL;
1052 : MemoryContext oldCxt;
1053 :
1054 : oldCxt = MemoryContextSwitchTo(walDebugCxt);
1055 :
1056 : initStringInfo(&buf);
1057 : appendStringInfo(&buf, "INSERT @ %X/%08X: ", LSN_FORMAT_ARGS(EndPos));
1058 :
1059 : /*
1060 : * We have to piece together the WAL record data from the XLogRecData
1061 : * entries, so that we can pass it to the rm_desc function as one
1062 : * contiguous chunk.
1063 : */
1064 : initStringInfo(&recordBuf);
1065 : for (; rdata != NULL; rdata = rdata->next)
1066 : appendBinaryStringInfo(&recordBuf, rdata->data, rdata->len);
1067 :
1068 : /* We also need temporary space to decode the record. */
1069 : record = (XLogRecord *) recordBuf.data;
1070 : decoded = (DecodedXLogRecord *)
1071 : palloc(DecodeXLogRecordRequiredSpace(record->xl_tot_len));
1072 :
1073 : if (!debug_reader)
1074 : debug_reader = XLogReaderAllocate(wal_segment_size, NULL,
1075 : XL_ROUTINE(.page_read = NULL,
1076 : .segment_open = NULL,
1077 : .segment_close = NULL),
1078 : NULL);
1079 : if (!debug_reader)
1080 : {
1081 : appendStringInfoString(&buf, "error decoding record: out of memory while allocating a WAL reading processor");
1082 : }
1083 : else if (!DecodeXLogRecord(debug_reader,
1084 : decoded,
1085 : record,
1086 : EndPos,
1087 : &errormsg))
1088 : {
1089 : appendStringInfo(&buf, "error decoding record: %s",
1090 : errormsg ? errormsg : "no error message");
1091 : }
1092 : else
1093 : {
1094 : appendStringInfoString(&buf, " - ");
1095 :
1096 : debug_reader->record = decoded;
1097 : xlog_outdesc(&buf, debug_reader);
1098 : debug_reader->record = NULL;
1099 : }
1100 : elog(LOG, "%s", buf.data);
1101 :
1102 : pfree(decoded);
1103 : pfree(buf.data);
1104 : pfree(recordBuf.data);
1105 : MemoryContextSwitchTo(oldCxt);
1106 : }
1107 : #endif
1108 :
1109 : /*
1110 : * Update our global variables
1111 : */
1112 24729416 : ProcLastRecPtr = StartPos;
1113 24729416 : XactLastRecEnd = EndPos;
1114 :
1115 : /* Report WAL traffic to the instrumentation. */
1116 24729416 : if (inserted)
1117 : {
1118 24729353 : pgWalUsage.wal_bytes += rechdr->xl_tot_len;
1119 24729353 : pgWalUsage.wal_records++;
1120 24729353 : pgWalUsage.wal_fpi += num_fpi;
1121 24729353 : pgWalUsage.wal_fpi_bytes += fpi_bytes;
1122 :
1123 : /* Required for the flush of pending stats WAL data */
1124 24729353 : pgstat_report_fixed = true;
1125 : }
1126 :
1127 24729416 : return EndPos;
1128 : }
1129 :
1130 : /*
1131 : * Reserves the right amount of space for a record of given size from the WAL.
1132 : * *StartPos is set to the beginning of the reserved section, *EndPos to
1133 : * its end+1. *PrevPtr is set to the beginning of the previous record; it is
1134 : * used to set the xl_prev of this record.
1135 : *
1136 : * This is the performance critical part of XLogInsert that must be serialized
1137 : * across backends. The rest can happen mostly in parallel. Try to keep this
1138 : * section as short as possible, insertpos_lck can be heavily contended on a
1139 : * busy system.
1140 : *
1141 : * NB: The space calculation here must match the code in CopyXLogRecordToWAL,
1142 : * where we actually copy the record to the reserved space.
1143 : *
1144 : * NB: Testing shows that XLogInsertRecord runs faster if this code is inlined;
1145 : * however, because there are two call sites, the compiler is reluctant to
1146 : * inline. We use pg_attribute_always_inline here to try to convince it.
1147 : */
1148 : static pg_attribute_always_inline void
1149 24728593 : ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1150 : XLogRecPtr *PrevPtr)
1151 : {
1152 24728593 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1153 : uint64 startbytepos;
1154 : uint64 endbytepos;
1155 : uint64 prevbytepos;
1156 :
1157 24728593 : size = MAXALIGN(size);
1158 :
1159 : /* All (non xlog-switch) records should contain data. */
1160 : Assert(size > SizeOfXLogRecord);
1161 :
1162 : /*
1163 : * The duration the spinlock needs to be held is minimized by minimizing
1164 : * the calculations that have to be done while holding the lock. The
1165 : * current tip of reserved WAL is kept in CurrBytePos, as a byte position
1166 : * that only counts "usable" bytes in WAL, that is, it excludes all WAL
1167 : * page headers. The mapping between "usable" byte positions and physical
1168 : * positions (XLogRecPtrs) can be done outside the locked region, and
1169 : * because the usable byte position doesn't include any headers, reserving
1170 : * X bytes from WAL is almost as simple as "CurrBytePos += X".
1171 : */
1172 24728593 : SpinLockAcquire(&Insert->insertpos_lck);
1173 :
1174 24728593 : startbytepos = Insert->CurrBytePos;
1175 24728593 : endbytepos = startbytepos + size;
1176 24728593 : prevbytepos = Insert->PrevBytePos;
1177 24728593 : Insert->CurrBytePos = endbytepos;
1178 24728593 : Insert->PrevBytePos = startbytepos;
1179 :
1180 24728593 : SpinLockRelease(&Insert->insertpos_lck);
1181 :
1182 24728593 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1183 24728593 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1184 24728593 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1185 :
1186 : /*
1187 : * Check that the conversions between "usable byte positions" and
1188 : * XLogRecPtrs work consistently in both directions.
1189 : */
1190 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1191 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1192 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1193 24728593 : }
1194 :
1195 : /*
1196 : * Like ReserveXLogInsertLocation(), but for an xlog-switch record.
1197 : *
1198 : * A log-switch record is handled slightly differently. The rest of the
1199 : * segment will be reserved for this insertion, as indicated by the returned
1200 : * *EndPos value. However, if we are already at the beginning of the current
1201 : * segment, *StartPos and *EndPos are set to the current location without
1202 : * reserving any space, and the function returns false.
1203 : */
1204 : static bool
1205 823 : ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1206 : {
1207 823 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1208 : uint64 startbytepos;
1209 : uint64 endbytepos;
1210 : uint64 prevbytepos;
1211 823 : uint32 size = MAXALIGN(SizeOfXLogRecord);
1212 : XLogRecPtr ptr;
1213 : uint32 segleft;
1214 :
1215 : /*
1216 : * These calculations are a bit heavy-weight to be done while holding a
1217 : * spinlock, but since we're holding all the WAL insertion locks, there
1218 : * are no other inserters competing for it. GetXLogInsertRecPtr() does
1219 : * compete for it, but that's not called very frequently.
1220 : */
1221 823 : SpinLockAcquire(&Insert->insertpos_lck);
1222 :
1223 823 : startbytepos = Insert->CurrBytePos;
1224 :
1225 823 : ptr = XLogBytePosToEndRecPtr(startbytepos);
1226 823 : if (XLogSegmentOffset(ptr, wal_segment_size) == 0)
1227 : {
1228 63 : SpinLockRelease(&Insert->insertpos_lck);
1229 63 : *EndPos = *StartPos = ptr;
1230 63 : return false;
1231 : }
1232 :
1233 760 : endbytepos = startbytepos + size;
1234 760 : prevbytepos = Insert->PrevBytePos;
1235 :
1236 760 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1237 760 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1238 :
1239 760 : segleft = wal_segment_size - XLogSegmentOffset(*EndPos, wal_segment_size);
1240 760 : if (segleft != wal_segment_size)
1241 : {
1242 : /* consume the rest of the segment */
1243 760 : *EndPos += segleft;
1244 760 : endbytepos = XLogRecPtrToBytePos(*EndPos);
1245 : }
1246 760 : Insert->CurrBytePos = endbytepos;
1247 760 : Insert->PrevBytePos = startbytepos;
1248 :
1249 760 : SpinLockRelease(&Insert->insertpos_lck);
1250 :
1251 760 : *PrevPtr = XLogBytePosToRecPtr(prevbytepos);
1252 :
1253 : Assert(XLogSegmentOffset(*EndPos, wal_segment_size) == 0);
1254 : Assert(XLogRecPtrToBytePos(*EndPos) == endbytepos);
1255 : Assert(XLogRecPtrToBytePos(*StartPos) == startbytepos);
1256 : Assert(XLogRecPtrToBytePos(*PrevPtr) == prevbytepos);
1257 :
1258 760 : return true;
1259 : }
1260 :
1261 : /*
1262 : * Subroutine of XLogInsertRecord. Copies a WAL record to an already-reserved
1263 : * area in the WAL.
1264 : */
1265 : static void
1266 24729353 : CopyXLogRecordToWAL(int write_len, bool isLogSwitch, XLogRecData *rdata,
1267 : XLogRecPtr StartPos, XLogRecPtr EndPos, TimeLineID tli)
1268 : {
1269 : char *currpos;
1270 : int freespace;
1271 : int written;
1272 : XLogRecPtr CurrPos;
1273 : XLogPageHeader pagehdr;
1274 :
1275 : /*
1276 : * Get a pointer to the right place in the right WAL buffer to start
1277 : * inserting to.
1278 : */
1279 24729353 : CurrPos = StartPos;
1280 24729353 : currpos = GetXLogBuffer(CurrPos, tli);
1281 24729353 : freespace = INSERT_FREESPACE(CurrPos);
1282 :
1283 : /*
1284 : * there should be enough space for at least the first field (xl_tot_len)
1285 : * on this page.
1286 : */
1287 : Assert(freespace >= sizeof(uint32));
1288 :
1289 : /* Copy record data */
1290 24729353 : written = 0;
1291 112822655 : while (rdata != NULL)
1292 : {
1293 88093302 : const char *rdata_data = rdata->data;
1294 88093302 : int rdata_len = rdata->len;
1295 :
1296 90100253 : while (rdata_len > freespace)
1297 : {
1298 : /*
1299 : * Write what fits on this page, and continue on the next page.
1300 : */
1301 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || freespace == 0);
1302 2006951 : memcpy(currpos, rdata_data, freespace);
1303 2006951 : rdata_data += freespace;
1304 2006951 : rdata_len -= freespace;
1305 2006951 : written += freespace;
1306 2006951 : CurrPos += freespace;
1307 :
1308 : /*
1309 : * Get pointer to beginning of next page, and set the xlp_rem_len
1310 : * in the page header. Set XLP_FIRST_IS_CONTRECORD.
1311 : *
1312 : * It's safe to set the contrecord flag and xlp_rem_len without a
1313 : * lock on the page. All the other flags were already set when the
1314 : * page was initialized, in AdvanceXLInsertBuffer, and we're the
1315 : * only backend that needs to set the contrecord flag.
1316 : */
1317 2006951 : currpos = GetXLogBuffer(CurrPos, tli);
1318 2006951 : pagehdr = (XLogPageHeader) currpos;
1319 2006951 : pagehdr->xlp_rem_len = write_len - written;
1320 2006951 : pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1321 :
1322 : /* skip over the page header */
1323 2006951 : if (XLogSegmentOffset(CurrPos, wal_segment_size) == 0)
1324 : {
1325 1282 : CurrPos += SizeOfXLogLongPHD;
1326 1282 : currpos += SizeOfXLogLongPHD;
1327 : }
1328 : else
1329 : {
1330 2005669 : CurrPos += SizeOfXLogShortPHD;
1331 2005669 : currpos += SizeOfXLogShortPHD;
1332 : }
1333 2006951 : freespace = INSERT_FREESPACE(CurrPos);
1334 : }
1335 :
1336 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1337 88093302 : memcpy(currpos, rdata_data, rdata_len);
1338 88093302 : currpos += rdata_len;
1339 88093302 : CurrPos += rdata_len;
1340 88093302 : freespace -= rdata_len;
1341 88093302 : written += rdata_len;
1342 :
1343 88093302 : rdata = rdata->next;
1344 : }
1345 : Assert(written == write_len);
1346 :
1347 : /*
1348 : * If this was an xlog-switch, it's not enough to write the switch record,
1349 : * we also have to consume all the remaining space in the WAL segment. We
1350 : * have already reserved that space, but we need to actually fill it.
1351 : */
1352 24729353 : if (isLogSwitch && XLogSegmentOffset(CurrPos, wal_segment_size) != 0)
1353 : {
1354 : /* An xlog-switch record doesn't contain any data besides the header */
1355 : Assert(write_len == SizeOfXLogRecord);
1356 :
1357 : /* Assert that we did reserve the right amount of space */
1358 : Assert(XLogSegmentOffset(EndPos, wal_segment_size) == 0);
1359 :
1360 : /* Use up all the remaining space on the current page */
1361 760 : CurrPos += freespace;
1362 :
1363 : /*
1364 : * Cause all remaining pages in the segment to be flushed, leaving the
1365 : * XLog position where it should be, at the start of the next segment.
1366 : * We do this one page at a time, to make sure we don't deadlock
1367 : * against ourselves if wal_buffers < wal_segment_size.
1368 : */
1369 763257 : while (CurrPos < EndPos)
1370 : {
1371 : /*
1372 : * The minimal action to flush the page would be to call
1373 : * WALInsertLockUpdateInsertingAt(CurrPos) followed by
1374 : * AdvanceXLInsertBuffer(...). The page would be left initialized
1375 : * mostly to zeros, except for the page header (always the short
1376 : * variant, as this is never a segment's first page).
1377 : *
1378 : * The large vistas of zeros are good for compressibility, but the
1379 : * headers interrupting them every XLOG_BLCKSZ (with values that
1380 : * differ from page to page) are not. The effect varies with
1381 : * compression tool, but bzip2 for instance compresses about an
1382 : * order of magnitude worse if those headers are left in place.
1383 : *
1384 : * Rather than complicating AdvanceXLInsertBuffer itself (which is
1385 : * called in heavily-loaded circumstances as well as this lightly-
1386 : * loaded one) with variant behavior, we just use GetXLogBuffer
1387 : * (which itself calls the two methods we need) to get the pointer
1388 : * and zero most of the page. Then we just zero the page header.
1389 : */
1390 762497 : currpos = GetXLogBuffer(CurrPos, tli);
1391 3049988 : MemSet(currpos, 0, SizeOfXLogShortPHD);
1392 :
1393 762497 : CurrPos += XLOG_BLCKSZ;
1394 : }
1395 : }
1396 : else
1397 : {
1398 : /* Align the end position, so that the next record starts aligned */
1399 24728593 : CurrPos = MAXALIGN64(CurrPos);
1400 : }
1401 :
1402 24729353 : if (CurrPos != EndPos)
1403 0 : ereport(PANIC,
1404 : errcode(ERRCODE_DATA_CORRUPTED),
1405 : errmsg_internal("space reserved for WAL record does not match what was written"));
1406 24729353 : }
1407 :
1408 : /*
1409 : * Acquire a WAL insertion lock, for inserting to WAL.
1410 : */
1411 : static void
1412 24737500 : WALInsertLockAcquire(void)
1413 : {
1414 : bool immed;
1415 :
1416 : /*
1417 : * It doesn't matter which of the WAL insertion locks we acquire, so try
1418 : * the one we used last time. If the system isn't particularly busy, it's
1419 : * a good bet that it's still available, and it's good to have some
1420 : * affinity to a particular lock so that you don't unnecessarily bounce
1421 : * cache lines between processes when there's no contention.
1422 : *
1423 : * If this is the first time through in this backend, pick a lock
1424 : * (semi-)randomly. This allows the locks to be used evenly if you have a
1425 : * lot of very short connections.
1426 : */
1427 : static int lockToTry = -1;
1428 :
1429 24737500 : if (lockToTry == -1)
1430 9380 : lockToTry = MyProcNumber % NUM_XLOGINSERT_LOCKS;
1431 24737500 : MyLockNo = lockToTry;
1432 :
1433 : /*
1434 : * The insertingAt value is initially set to 0, as we don't know our
1435 : * insert location yet.
1436 : */
1437 24737500 : immed = LWLockAcquire(&WALInsertLocks[MyLockNo].l.lock, LW_EXCLUSIVE);
1438 24737500 : if (!immed)
1439 : {
1440 : /*
1441 : * If we couldn't get the lock immediately, try another lock next
1442 : * time. On a system with more insertion locks than concurrent
1443 : * inserters, this causes all the inserters to eventually migrate to a
1444 : * lock that no-one else is using. On a system with more inserters
1445 : * than locks, it still helps to distribute the inserters evenly
1446 : * across the locks.
1447 : */
1448 20195 : lockToTry = (lockToTry + 1) % NUM_XLOGINSERT_LOCKS;
1449 : }
1450 24737500 : }
1451 :
1452 : /*
1453 : * Acquire all WAL insertion locks, to prevent other backends from inserting
1454 : * to WAL.
1455 : */
1456 : static void
1457 4827 : WALInsertLockAcquireExclusive(void)
1458 : {
1459 : int i;
1460 :
1461 : /*
1462 : * When holding all the locks, all but the last lock's insertingAt
1463 : * indicator is set to 0xFFFFFFFFFFFFFFFF, which is higher than any real
1464 : * XLogRecPtr value, to make sure that no-one blocks waiting on those.
1465 : */
1466 38616 : for (i = 0; i < NUM_XLOGINSERT_LOCKS - 1; i++)
1467 : {
1468 33789 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1469 33789 : LWLockUpdateVar(&WALInsertLocks[i].l.lock,
1470 33789 : &WALInsertLocks[i].l.insertingAt,
1471 : PG_UINT64_MAX);
1472 : }
1473 : /* Variable value reset to 0 at release */
1474 4827 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1475 :
1476 4827 : holdingAllLocks = true;
1477 4827 : }
1478 :
1479 : /*
1480 : * Release our insertion lock (or locks, if we're holding them all).
1481 : *
1482 : * NB: Reset all variables to 0, so they cause LWLockWaitForVar to block the
1483 : * next time the lock is acquired.
1484 : */
1485 : static void
1486 24742327 : WALInsertLockRelease(void)
1487 : {
1488 24742327 : if (holdingAllLocks)
1489 : {
1490 : int i;
1491 :
1492 43443 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1493 38616 : LWLockReleaseClearVar(&WALInsertLocks[i].l.lock,
1494 38616 : &WALInsertLocks[i].l.insertingAt,
1495 : 0);
1496 :
1497 4827 : holdingAllLocks = false;
1498 : }
1499 : else
1500 : {
1501 24737500 : LWLockReleaseClearVar(&WALInsertLocks[MyLockNo].l.lock,
1502 24737500 : &WALInsertLocks[MyLockNo].l.insertingAt,
1503 : 0);
1504 : }
1505 24742327 : }
1506 :
1507 : /*
1508 : * Update our insertingAt value, to let others know that we've finished
1509 : * inserting up to that point.
1510 : */
1511 : static void
1512 2717228 : WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt)
1513 : {
1514 2717228 : if (holdingAllLocks)
1515 : {
1516 : /*
1517 : * We use the last lock to mark our actual position, see comments in
1518 : * WALInsertLockAcquireExclusive.
1519 : */
1520 760278 : LWLockUpdateVar(&WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.lock,
1521 760278 : &WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.insertingAt,
1522 : insertingAt);
1523 : }
1524 : else
1525 1956950 : LWLockUpdateVar(&WALInsertLocks[MyLockNo].l.lock,
1526 1956950 : &WALInsertLocks[MyLockNo].l.insertingAt,
1527 : insertingAt);
1528 2717228 : }
1529 :
1530 : /*
1531 : * Wait for any WAL insertions < upto to finish.
1532 : *
1533 : * Returns the location of the oldest insertion that is still in-progress.
1534 : * Any WAL prior to that point has been fully copied into WAL buffers, and
1535 : * can be flushed out to disk. Because this waits for any insertions older
1536 : * than 'upto' to finish, the return value is always >= 'upto'.
1537 : *
1538 : * Note: When you are about to write out WAL, you must call this function
1539 : * *before* acquiring WALWriteLock, to avoid deadlocks. This function might
1540 : * need to wait for an insertion to finish (or at least advance to next
1541 : * uninitialized page), and the inserter might need to evict an old WAL buffer
1542 : * to make room for a new one, which in turn requires WALWriteLock.
1543 : */
1544 : static XLogRecPtr
1545 2527191 : WaitXLogInsertionsToFinish(XLogRecPtr upto)
1546 : {
1547 : uint64 bytepos;
1548 : XLogRecPtr inserted;
1549 : XLogRecPtr reservedUpto;
1550 : XLogRecPtr finishedUpto;
1551 2527191 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1552 : int i;
1553 :
1554 2527191 : if (MyProc == NULL)
1555 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1556 :
1557 : /*
1558 : * Check if there's any work to do. Use a barrier to ensure we get the
1559 : * freshest value.
1560 : */
1561 2527191 : inserted = pg_atomic_read_membarrier_u64(&XLogCtl->logInsertResult);
1562 2527191 : if (upto <= inserted)
1563 2011312 : return inserted;
1564 :
1565 : /* Read the current insert position */
1566 515879 : SpinLockAcquire(&Insert->insertpos_lck);
1567 515879 : bytepos = Insert->CurrBytePos;
1568 515879 : SpinLockRelease(&Insert->insertpos_lck);
1569 515879 : reservedUpto = XLogBytePosToEndRecPtr(bytepos);
1570 :
1571 : /*
1572 : * No-one should request to flush a piece of WAL that hasn't even been
1573 : * reserved yet. However, it can happen if there is a block with a bogus
1574 : * LSN on disk, for example. XLogFlush checks for that situation and
1575 : * complains, but only after the flush. Here we just assume that to mean
1576 : * that all WAL that has been reserved needs to be finished. In this
1577 : * corner-case, the return value can be smaller than 'upto' argument.
1578 : */
1579 515879 : if (upto > reservedUpto)
1580 : {
1581 0 : ereport(LOG,
1582 : errmsg("request to flush past end of generated WAL; request %X/%08X, current position %X/%08X",
1583 : LSN_FORMAT_ARGS(upto), LSN_FORMAT_ARGS(reservedUpto)));
1584 0 : upto = reservedUpto;
1585 : }
1586 :
1587 : /*
1588 : * Loop through all the locks, sleeping on any in-progress insert older
1589 : * than 'upto'.
1590 : *
1591 : * finishedUpto is our return value, indicating the point upto which all
1592 : * the WAL insertions have been finished. Initialize it to the head of
1593 : * reserved WAL, and as we iterate through the insertion locks, back it
1594 : * out for any insertion that's still in progress.
1595 : */
1596 515879 : finishedUpto = reservedUpto;
1597 4642911 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1598 : {
1599 4127032 : XLogRecPtr insertingat = InvalidXLogRecPtr;
1600 :
1601 : do
1602 : {
1603 : /*
1604 : * See if this insertion is in progress. LWLockWaitForVar will
1605 : * wait for the lock to be released, or for the 'value' to be set
1606 : * by a LWLockUpdateVar call. When a lock is initially acquired,
1607 : * its value is 0 (InvalidXLogRecPtr), which means that we don't
1608 : * know where it's inserting yet. We will have to wait for it. If
1609 : * it's a small insertion, the record will most likely fit on the
1610 : * same page and the inserter will release the lock without ever
1611 : * calling LWLockUpdateVar. But if it has to sleep, it will
1612 : * advertise the insertion point with LWLockUpdateVar before
1613 : * sleeping.
1614 : *
1615 : * In this loop we are only waiting for insertions that started
1616 : * before WaitXLogInsertionsToFinish was called. The lack of
1617 : * memory barriers in the loop means that we might see locks as
1618 : * "unused" that have since become used. This is fine because
1619 : * they only can be used for later insertions that we would not
1620 : * want to wait on anyway. Not taking a lock to acquire the
1621 : * current insertingAt value means that we might see older
1622 : * insertingAt values. This is also fine, because if we read a
1623 : * value too old, we will add ourselves to the wait queue, which
1624 : * contains atomic operations.
1625 : */
1626 4232361 : if (LWLockWaitForVar(&WALInsertLocks[i].l.lock,
1627 4232361 : &WALInsertLocks[i].l.insertingAt,
1628 : insertingat, &insertingat))
1629 : {
1630 : /* the lock was free, so no insertion in progress */
1631 2970292 : insertingat = InvalidXLogRecPtr;
1632 2970292 : break;
1633 : }
1634 :
1635 : /*
1636 : * This insertion is still in progress. Have to wait, unless the
1637 : * inserter has proceeded past 'upto'.
1638 : */
1639 1262069 : } while (insertingat < upto);
1640 :
1641 4127032 : if (XLogRecPtrIsValid(insertingat) && insertingat < finishedUpto)
1642 416322 : finishedUpto = insertingat;
1643 : }
1644 :
1645 : /*
1646 : * Advance the limit we know to have been inserted and return the freshest
1647 : * value we know of, which might be beyond what we requested if somebody
1648 : * is concurrently doing this with an 'upto' pointer ahead of us.
1649 : */
1650 515879 : finishedUpto = pg_atomic_monotonic_advance_u64(&XLogCtl->logInsertResult,
1651 : finishedUpto);
1652 :
1653 515879 : return finishedUpto;
1654 : }
1655 :
1656 : /*
1657 : * Get a pointer to the right location in the WAL buffer containing the
1658 : * given XLogRecPtr.
1659 : *
1660 : * If the page is not initialized yet, it is initialized. That might require
1661 : * evicting an old dirty buffer from the buffer cache, which means I/O.
1662 : *
1663 : * The caller must ensure that the page containing the requested location
1664 : * isn't evicted yet, and won't be evicted. The way to ensure that is to
1665 : * hold onto a WAL insertion lock with the insertingAt position set to
1666 : * something <= ptr. GetXLogBuffer() will update insertingAt if it needs
1667 : * to evict an old page from the buffer. (This means that once you call
1668 : * GetXLogBuffer() with a given 'ptr', you must not access anything before
1669 : * that point anymore, and must not call GetXLogBuffer() with an older 'ptr'
1670 : * later, because older buffers might be recycled already)
1671 : */
1672 : static char *
1673 27498812 : GetXLogBuffer(XLogRecPtr ptr, TimeLineID tli)
1674 : {
1675 : int idx;
1676 : XLogRecPtr endptr;
1677 : static uint64 cachedPage = 0;
1678 : static char *cachedPos = NULL;
1679 : XLogRecPtr expectedEndPtr;
1680 :
1681 : /*
1682 : * Fast path for the common case that we need to access again the same
1683 : * page as last time.
1684 : */
1685 27498812 : if (ptr / XLOG_BLCKSZ == cachedPage)
1686 : {
1687 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1688 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1689 24275638 : return cachedPos + ptr % XLOG_BLCKSZ;
1690 : }
1691 :
1692 : /*
1693 : * The XLog buffer cache is organized so that a page is always loaded to a
1694 : * particular buffer. That way we can easily calculate the buffer a given
1695 : * page must be loaded into, from the XLogRecPtr alone.
1696 : */
1697 3223174 : idx = XLogRecPtrToBufIdx(ptr);
1698 :
1699 : /*
1700 : * See what page is loaded in the buffer at the moment. It could be the
1701 : * page we're looking for, or something older. It can't be anything newer
1702 : * - that would imply the page we're looking for has already been written
1703 : * out to disk and evicted, and the caller is responsible for making sure
1704 : * that doesn't happen.
1705 : *
1706 : * We don't hold a lock while we read the value. If someone is just about
1707 : * to initialize or has just initialized the page, it's possible that we
1708 : * get InvalidXLogRecPtr. That's ok, we'll grab the mapping lock (in
1709 : * AdvanceXLInsertBuffer) and retry if we see anything other than the page
1710 : * we're looking for.
1711 : */
1712 3223174 : expectedEndPtr = ptr;
1713 3223174 : expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
1714 :
1715 3223174 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1716 3223174 : if (expectedEndPtr != endptr)
1717 : {
1718 : XLogRecPtr initializedUpto;
1719 :
1720 : /*
1721 : * Before calling AdvanceXLInsertBuffer(), which can block, let others
1722 : * know how far we're finished with inserting the record.
1723 : *
1724 : * NB: If 'ptr' points to just after the page header, advertise a
1725 : * position at the beginning of the page rather than 'ptr' itself. If
1726 : * there are no other insertions running, someone might try to flush
1727 : * up to our advertised location. If we advertised a position after
1728 : * the page header, someone might try to flush the page header, even
1729 : * though page might actually not be initialized yet. As the first
1730 : * inserter on the page, we are effectively responsible for making
1731 : * sure that it's initialized, before we let insertingAt to move past
1732 : * the page header.
1733 : */
1734 2717228 : if (ptr % XLOG_BLCKSZ == SizeOfXLogShortPHD &&
1735 12416 : XLogSegmentOffset(ptr, wal_segment_size) > XLOG_BLCKSZ)
1736 12416 : initializedUpto = ptr - SizeOfXLogShortPHD;
1737 2704812 : else if (ptr % XLOG_BLCKSZ == SizeOfXLogLongPHD &&
1738 1065 : XLogSegmentOffset(ptr, wal_segment_size) < XLOG_BLCKSZ)
1739 633 : initializedUpto = ptr - SizeOfXLogLongPHD;
1740 : else
1741 2704179 : initializedUpto = ptr;
1742 :
1743 2717228 : WALInsertLockUpdateInsertingAt(initializedUpto);
1744 :
1745 2717228 : AdvanceXLInsertBuffer(ptr, tli, false);
1746 2717228 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1747 :
1748 2717228 : if (expectedEndPtr != endptr)
1749 0 : elog(PANIC, "could not find WAL buffer for %X/%08X",
1750 : LSN_FORMAT_ARGS(ptr));
1751 : }
1752 : else
1753 : {
1754 : /*
1755 : * Make sure the initialization of the page is visible to us, and
1756 : * won't arrive later to overwrite the WAL data we write on the page.
1757 : */
1758 505946 : pg_memory_barrier();
1759 : }
1760 :
1761 : /*
1762 : * Found the buffer holding this page. Return a pointer to the right
1763 : * offset within the page.
1764 : */
1765 3223174 : cachedPage = ptr / XLOG_BLCKSZ;
1766 3223174 : cachedPos = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1767 :
1768 : Assert(((XLogPageHeader) cachedPos)->xlp_magic == XLOG_PAGE_MAGIC);
1769 : Assert(((XLogPageHeader) cachedPos)->xlp_pageaddr == ptr - (ptr % XLOG_BLCKSZ));
1770 :
1771 3223174 : return cachedPos + ptr % XLOG_BLCKSZ;
1772 : }
1773 :
1774 : /*
1775 : * Read WAL data directly from WAL buffers, if available. Returns the number
1776 : * of bytes read successfully.
1777 : *
1778 : * Fewer than 'count' bytes may be read if some of the requested WAL data has
1779 : * already been evicted.
1780 : *
1781 : * No locks are taken.
1782 : *
1783 : * Caller should ensure that it reads no further than LogwrtResult.Write
1784 : * (which should have been updated by the caller when determining how far to
1785 : * read). The 'tli' argument is only used as a convenient safety check so that
1786 : * callers do not read from WAL buffers on a historical timeline.
1787 : */
1788 : Size
1789 108318 : WALReadFromBuffers(char *dstbuf, XLogRecPtr startptr, Size count,
1790 : TimeLineID tli)
1791 : {
1792 108318 : char *pdst = dstbuf;
1793 108318 : XLogRecPtr recptr = startptr;
1794 : XLogRecPtr inserted;
1795 108318 : Size nbytes = count;
1796 :
1797 108318 : if (RecoveryInProgress() || tli != GetWALInsertionTimeLine())
1798 1049 : return 0;
1799 :
1800 : Assert(XLogRecPtrIsValid(startptr));
1801 :
1802 : /*
1803 : * Caller should ensure that the requested data has been inserted into WAL
1804 : * buffers before we try to read it.
1805 : */
1806 107269 : inserted = pg_atomic_read_u64(&XLogCtl->logInsertResult);
1807 107269 : if (startptr + count > inserted)
1808 0 : ereport(ERROR,
1809 : errmsg("cannot read past end of generated WAL: requested %X/%08X, current position %X/%08X",
1810 : LSN_FORMAT_ARGS(startptr + count),
1811 : LSN_FORMAT_ARGS(inserted)));
1812 :
1813 : /*
1814 : * Loop through the buffers without a lock. For each buffer, atomically
1815 : * read and verify the end pointer, then copy the data out, and finally
1816 : * re-read and re-verify the end pointer.
1817 : *
1818 : * Once a page is evicted, it never returns to the WAL buffers, so if the
1819 : * end pointer matches the expected end pointer before and after we copy
1820 : * the data, then the right page must have been present during the data
1821 : * copy. Read barriers are necessary to ensure that the data copy actually
1822 : * happens between the two verification steps.
1823 : *
1824 : * If either verification fails, we simply terminate the loop and return
1825 : * with the data that had been already copied out successfully.
1826 : */
1827 137548 : while (nbytes > 0)
1828 : {
1829 128628 : uint32 offset = recptr % XLOG_BLCKSZ;
1830 128628 : int idx = XLogRecPtrToBufIdx(recptr);
1831 : XLogRecPtr expectedEndPtr;
1832 : XLogRecPtr endptr;
1833 : const char *page;
1834 : const char *psrc;
1835 : Size npagebytes;
1836 :
1837 : /*
1838 : * Calculate the end pointer we expect in the xlblocks array if the
1839 : * correct page is present.
1840 : */
1841 128628 : expectedEndPtr = recptr + (XLOG_BLCKSZ - offset);
1842 :
1843 : /*
1844 : * First verification step: check that the correct page is present in
1845 : * the WAL buffers.
1846 : */
1847 128628 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1848 128628 : if (expectedEndPtr != endptr)
1849 98339 : break;
1850 :
1851 : /*
1852 : * The correct page is present (or was at the time the endptr was
1853 : * read; must re-verify later). Calculate pointer to source data and
1854 : * determine how much data to read from this page.
1855 : */
1856 30289 : page = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1857 30289 : psrc = page + offset;
1858 30289 : npagebytes = Min(nbytes, XLOG_BLCKSZ - offset);
1859 :
1860 : /*
1861 : * Ensure that the data copy and the first verification step are not
1862 : * reordered.
1863 : */
1864 30289 : pg_read_barrier();
1865 :
1866 : /* data copy */
1867 30289 : memcpy(pdst, psrc, npagebytes);
1868 :
1869 : /*
1870 : * Ensure that the data copy and the second verification step are not
1871 : * reordered.
1872 : */
1873 30289 : pg_read_barrier();
1874 :
1875 : /*
1876 : * Second verification step: check that the page we read from wasn't
1877 : * evicted while we were copying the data.
1878 : */
1879 30289 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1880 30289 : if (expectedEndPtr != endptr)
1881 10 : break;
1882 :
1883 30279 : pdst += npagebytes;
1884 30279 : recptr += npagebytes;
1885 30279 : nbytes -= npagebytes;
1886 : }
1887 :
1888 : Assert(pdst - dstbuf <= count);
1889 :
1890 107269 : return pdst - dstbuf;
1891 : }
1892 :
1893 : /*
1894 : * Converts a "usable byte position" to XLogRecPtr. A usable byte position
1895 : * is the position starting from the beginning of WAL, excluding all WAL
1896 : * page headers.
1897 : */
1898 : static XLogRecPtr
1899 49461601 : XLogBytePosToRecPtr(uint64 bytepos)
1900 : {
1901 : uint64 fullsegs;
1902 : uint64 fullpages;
1903 : uint64 bytesleft;
1904 : uint32 seg_offset;
1905 : XLogRecPtr result;
1906 :
1907 49461601 : fullsegs = bytepos / UsableBytesInSegment;
1908 49461601 : bytesleft = bytepos % UsableBytesInSegment;
1909 :
1910 49461601 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1911 : {
1912 : /* fits on first page of segment */
1913 70771 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1914 : }
1915 : else
1916 : {
1917 : /* account for the first page on segment with long header */
1918 49390830 : seg_offset = XLOG_BLCKSZ;
1919 49390830 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1920 :
1921 49390830 : fullpages = bytesleft / UsableBytesInPage;
1922 49390830 : bytesleft = bytesleft % UsableBytesInPage;
1923 :
1924 49390830 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1925 : }
1926 :
1927 49461601 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1928 :
1929 49461601 : return result;
1930 : }
1931 :
1932 : /*
1933 : * Like XLogBytePosToRecPtr, but if the position is at a page boundary,
1934 : * returns a pointer to the beginning of the page (ie. before page header),
1935 : * not to where the first xlog record on that page would go to. This is used
1936 : * when converting a pointer to the end of a record.
1937 : */
1938 : static XLogRecPtr
1939 25249406 : XLogBytePosToEndRecPtr(uint64 bytepos)
1940 : {
1941 : uint64 fullsegs;
1942 : uint64 fullpages;
1943 : uint64 bytesleft;
1944 : uint32 seg_offset;
1945 : XLogRecPtr result;
1946 :
1947 25249406 : fullsegs = bytepos / UsableBytesInSegment;
1948 25249406 : bytesleft = bytepos % UsableBytesInSegment;
1949 :
1950 25249406 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1951 : {
1952 : /* fits on first page of segment */
1953 109906 : if (bytesleft == 0)
1954 73057 : seg_offset = 0;
1955 : else
1956 36849 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1957 : }
1958 : else
1959 : {
1960 : /* account for the first page on segment with long header */
1961 25139500 : seg_offset = XLOG_BLCKSZ;
1962 25139500 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1963 :
1964 25139500 : fullpages = bytesleft / UsableBytesInPage;
1965 25139500 : bytesleft = bytesleft % UsableBytesInPage;
1966 :
1967 25139500 : if (bytesleft == 0)
1968 24131 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
1969 : else
1970 25115369 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1971 : }
1972 :
1973 25249406 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1974 :
1975 25249406 : return result;
1976 : }
1977 :
1978 : /*
1979 : * Convert an XLogRecPtr to a "usable byte position".
1980 : */
1981 : static uint64
1982 2796 : XLogRecPtrToBytePos(XLogRecPtr ptr)
1983 : {
1984 : uint64 fullsegs;
1985 : uint32 fullpages;
1986 : uint32 offset;
1987 : uint64 result;
1988 :
1989 2796 : XLByteToSeg(ptr, fullsegs, wal_segment_size);
1990 :
1991 2796 : fullpages = (XLogSegmentOffset(ptr, wal_segment_size)) / XLOG_BLCKSZ;
1992 2796 : offset = ptr % XLOG_BLCKSZ;
1993 :
1994 2796 : if (fullpages == 0)
1995 : {
1996 1069 : result = fullsegs * UsableBytesInSegment;
1997 1069 : if (offset > 0)
1998 : {
1999 : Assert(offset >= SizeOfXLogLongPHD);
2000 286 : result += offset - SizeOfXLogLongPHD;
2001 : }
2002 : }
2003 : else
2004 : {
2005 1727 : result = fullsegs * UsableBytesInSegment +
2006 1727 : (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
2007 1727 : (fullpages - 1) * UsableBytesInPage; /* full pages */
2008 1727 : if (offset > 0)
2009 : {
2010 : Assert(offset >= SizeOfXLogShortPHD);
2011 1717 : result += offset - SizeOfXLogShortPHD;
2012 : }
2013 : }
2014 :
2015 2796 : return result;
2016 : }
2017 :
2018 : /*
2019 : * Initialize XLOG buffers, writing out old buffers if they still contain
2020 : * unwritten data, upto the page containing 'upto'. Or if 'opportunistic' is
2021 : * true, initialize as many pages as we can without having to write out
2022 : * unwritten data. Any new pages are initialized to zeros, with pages headers
2023 : * initialized properly.
2024 : */
2025 : static void
2026 2722163 : AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
2027 : {
2028 : int nextidx;
2029 : XLogRecPtr OldPageRqstPtr;
2030 : XLogwrtRqst WriteRqst;
2031 2722163 : XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
2032 : XLogRecPtr NewPageBeginPtr;
2033 : XLogPageHeader NewPage;
2034 2722163 : int npages pg_attribute_unused() = 0;
2035 :
2036 2722163 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2037 :
2038 : /*
2039 : * Now that we have the lock, check if someone initialized the page
2040 : * already.
2041 : */
2042 7901330 : while (upto >= XLogCtl->InitializedUpTo || opportunistic)
2043 : {
2044 5184102 : nextidx = XLogRecPtrToBufIdx(XLogCtl->InitializedUpTo);
2045 :
2046 : /*
2047 : * Get ending-offset of the buffer page we need to replace (this may
2048 : * be zero if the buffer hasn't been used yet). Fall through if it's
2049 : * already written out.
2050 : */
2051 5184102 : OldPageRqstPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]);
2052 5184102 : if (LogwrtResult.Write < OldPageRqstPtr)
2053 : {
2054 : /*
2055 : * Nope, got work to do. If we just want to pre-initialize as much
2056 : * as we can without flushing, give up now.
2057 : */
2058 2377968 : if (opportunistic)
2059 4935 : break;
2060 :
2061 : /* Advance shared memory write request position */
2062 2373033 : SpinLockAcquire(&XLogCtl->info_lck);
2063 2373033 : if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
2064 712517 : XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
2065 2373033 : SpinLockRelease(&XLogCtl->info_lck);
2066 :
2067 : /*
2068 : * Acquire an up-to-date LogwrtResult value and see if we still
2069 : * need to write it or if someone else already did.
2070 : */
2071 2373033 : RefreshXLogWriteResult(LogwrtResult);
2072 2373033 : if (LogwrtResult.Write < OldPageRqstPtr)
2073 : {
2074 : /*
2075 : * Must acquire write lock. Release WALBufMappingLock first,
2076 : * to make sure that all insertions that we need to wait for
2077 : * can finish (up to this same position). Otherwise we risk
2078 : * deadlock.
2079 : */
2080 2353599 : LWLockRelease(WALBufMappingLock);
2081 :
2082 2353599 : WaitXLogInsertionsToFinish(OldPageRqstPtr);
2083 :
2084 2353599 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2085 :
2086 2353599 : RefreshXLogWriteResult(LogwrtResult);
2087 2353599 : if (LogwrtResult.Write >= OldPageRqstPtr)
2088 : {
2089 : /* OK, someone wrote it already */
2090 149767 : LWLockRelease(WALWriteLock);
2091 : }
2092 : else
2093 : {
2094 : /* Have to write it ourselves */
2095 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2096 2203832 : WriteRqst.Write = OldPageRqstPtr;
2097 2203832 : WriteRqst.Flush = InvalidXLogRecPtr;
2098 2203832 : XLogWrite(WriteRqst, tli, false);
2099 2203832 : LWLockRelease(WALWriteLock);
2100 2203832 : pgWalUsage.wal_buffers_full++;
2101 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_DONE();
2102 :
2103 : /*
2104 : * Required for the flush of pending stats WAL data, per
2105 : * update of pgWalUsage.
2106 : */
2107 2203832 : pgstat_report_fixed = true;
2108 : }
2109 : /* Re-acquire WALBufMappingLock and retry */
2110 2353599 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2111 2353599 : continue;
2112 : }
2113 : }
2114 :
2115 : /*
2116 : * Now the next buffer slot is free and we can set it up to be the
2117 : * next output page.
2118 : */
2119 2825568 : NewPageBeginPtr = XLogCtl->InitializedUpTo;
2120 2825568 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2121 :
2122 : Assert(XLogRecPtrToBufIdx(NewPageBeginPtr) == nextidx);
2123 :
2124 2825568 : NewPage = (XLogPageHeader) (XLogCtl->pages + nextidx * (Size) XLOG_BLCKSZ);
2125 :
2126 : /*
2127 : * Mark the xlblock with InvalidXLogRecPtr and issue a write barrier
2128 : * before initializing. Otherwise, the old page may be partially
2129 : * zeroed but look valid.
2130 : */
2131 2825568 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], InvalidXLogRecPtr);
2132 2825568 : pg_write_barrier();
2133 :
2134 : /*
2135 : * Be sure to re-zero the buffer so that bytes beyond what we've
2136 : * written will look like zeroes and not valid XLOG records...
2137 : */
2138 2825568 : MemSet(NewPage, 0, XLOG_BLCKSZ);
2139 :
2140 : /*
2141 : * Fill the new page's header
2142 : */
2143 2825568 : NewPage->xlp_magic = XLOG_PAGE_MAGIC;
2144 :
2145 : /* NewPage->xlp_info = 0; */ /* done by memset */
2146 2825568 : NewPage->xlp_tli = tli;
2147 2825568 : NewPage->xlp_pageaddr = NewPageBeginPtr;
2148 :
2149 : /* NewPage->xlp_rem_len = 0; */ /* done by memset */
2150 :
2151 : /*
2152 : * If first page of an XLOG segment file, make it a long header.
2153 : */
2154 2825568 : if ((XLogSegmentOffset(NewPage->xlp_pageaddr, wal_segment_size)) == 0)
2155 : {
2156 1936 : XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2157 :
2158 1936 : NewLongPage->xlp_sysid = ControlFile->system_identifier;
2159 1936 : NewLongPage->xlp_seg_size = wal_segment_size;
2160 1936 : NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2161 1936 : NewPage->xlp_info |= XLP_LONG_HEADER;
2162 : }
2163 :
2164 : /*
2165 : * Make sure the initialization of the page becomes visible to others
2166 : * before the xlblocks update. GetXLogBuffer() reads xlblocks without
2167 : * holding a lock.
2168 : */
2169 2825568 : pg_write_barrier();
2170 :
2171 2825568 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], NewPageEndPtr);
2172 2825568 : XLogCtl->InitializedUpTo = NewPageEndPtr;
2173 :
2174 2825568 : npages++;
2175 : }
2176 2722163 : LWLockRelease(WALBufMappingLock);
2177 :
2178 : #ifdef WAL_DEBUG
2179 : if (XLOG_DEBUG && npages > 0)
2180 : {
2181 : elog(DEBUG1, "initialized %d pages, up to %X/%08X",
2182 : npages, LSN_FORMAT_ARGS(NewPageEndPtr));
2183 : }
2184 : #endif
2185 2722163 : }
2186 :
2187 : /*
2188 : * Calculate CheckPointSegments based on max_wal_size_mb and
2189 : * checkpoint_completion_target.
2190 : */
2191 : static void
2192 9718 : CalculateCheckpointSegments(void)
2193 : {
2194 : double target;
2195 :
2196 : /*-------
2197 : * Calculate the distance at which to trigger a checkpoint, to avoid
2198 : * exceeding max_wal_size_mb. This is based on two assumptions:
2199 : *
2200 : * a) we keep WAL for only one checkpoint cycle (prior to PG11 we kept
2201 : * WAL for two checkpoint cycles to allow us to recover from the
2202 : * secondary checkpoint if the first checkpoint failed, though we
2203 : * only did this on the primary anyway, not on standby. Keeping just
2204 : * one checkpoint simplifies processing and reduces disk space in
2205 : * many smaller databases.)
2206 : * b) during checkpoint, we consume checkpoint_completion_target *
2207 : * number of segments consumed between checkpoints.
2208 : *-------
2209 : */
2210 9718 : target = (double) ConvertToXSegs(max_wal_size_mb, wal_segment_size) /
2211 9718 : (1.0 + CheckPointCompletionTarget);
2212 :
2213 : /* round down */
2214 9718 : CheckPointSegments = (int) target;
2215 :
2216 9718 : if (CheckPointSegments < 1)
2217 10 : CheckPointSegments = 1;
2218 9718 : }
2219 :
2220 : void
2221 7282 : assign_max_wal_size(int newval, void *extra)
2222 : {
2223 7282 : max_wal_size_mb = newval;
2224 7282 : CalculateCheckpointSegments();
2225 7282 : }
2226 :
2227 : void
2228 1289 : assign_checkpoint_completion_target(double newval, void *extra)
2229 : {
2230 1289 : CheckPointCompletionTarget = newval;
2231 1289 : CalculateCheckpointSegments();
2232 1289 : }
2233 :
2234 : bool
2235 2494 : check_wal_segment_size(int *newval, void **extra, GucSource source)
2236 : {
2237 2494 : if (!IsValidWalSegSize(*newval))
2238 : {
2239 0 : GUC_check_errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.");
2240 0 : return false;
2241 : }
2242 :
2243 2494 : return true;
2244 : }
2245 :
2246 : /*
2247 : * At a checkpoint, how many WAL segments to recycle as preallocated future
2248 : * XLOG segments? Returns the highest segment that should be preallocated.
2249 : */
2250 : static XLogSegNo
2251 1951 : XLOGfileslop(XLogRecPtr lastredoptr)
2252 : {
2253 : XLogSegNo minSegNo;
2254 : XLogSegNo maxSegNo;
2255 : double distance;
2256 : XLogSegNo recycleSegNo;
2257 :
2258 : /*
2259 : * Calculate the segment numbers that min_wal_size_mb and max_wal_size_mb
2260 : * correspond to. Always recycle enough segments to meet the minimum, and
2261 : * remove enough segments to stay below the maximum.
2262 : */
2263 1951 : minSegNo = lastredoptr / wal_segment_size +
2264 1951 : ConvertToXSegs(min_wal_size_mb, wal_segment_size) - 1;
2265 1951 : maxSegNo = lastredoptr / wal_segment_size +
2266 1951 : ConvertToXSegs(max_wal_size_mb, wal_segment_size) - 1;
2267 :
2268 : /*
2269 : * Between those limits, recycle enough segments to get us through to the
2270 : * estimated end of next checkpoint.
2271 : *
2272 : * To estimate where the next checkpoint will finish, assume that the
2273 : * system runs steadily consuming CheckPointDistanceEstimate bytes between
2274 : * every checkpoint.
2275 : */
2276 1951 : distance = (1.0 + CheckPointCompletionTarget) * CheckPointDistanceEstimate;
2277 : /* add 10% for good measure. */
2278 1951 : distance *= 1.10;
2279 :
2280 1951 : recycleSegNo = (XLogSegNo) ceil(((double) lastredoptr + distance) /
2281 : wal_segment_size);
2282 :
2283 1951 : if (recycleSegNo < minSegNo)
2284 1371 : recycleSegNo = minSegNo;
2285 1951 : if (recycleSegNo > maxSegNo)
2286 429 : recycleSegNo = maxSegNo;
2287 :
2288 1951 : return recycleSegNo;
2289 : }
2290 :
2291 : /*
2292 : * Check whether we've consumed enough xlog space that a checkpoint is needed.
2293 : *
2294 : * new_segno indicates a log file that has just been filled up (or read
2295 : * during recovery). We measure the distance from RedoRecPtr to new_segno
2296 : * and see if that exceeds CheckPointSegments.
2297 : *
2298 : * Note: it is caller's responsibility that RedoRecPtr is up-to-date.
2299 : */
2300 : bool
2301 5140 : XLogCheckpointNeeded(XLogSegNo new_segno)
2302 : {
2303 : XLogSegNo old_segno;
2304 :
2305 5140 : XLByteToSeg(RedoRecPtr, old_segno, wal_segment_size);
2306 :
2307 5140 : if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2308 3212 : return true;
2309 1928 : return false;
2310 : }
2311 :
2312 : /*
2313 : * Write and/or fsync the log at least as far as WriteRqst indicates.
2314 : *
2315 : * If flexible == true, we don't have to write as far as WriteRqst, but
2316 : * may stop at any convenient boundary (such as a cache or logfile boundary).
2317 : * This option allows us to avoid uselessly issuing multiple writes when a
2318 : * single one would do.
2319 : *
2320 : * Must be called with WALWriteLock held. WaitXLogInsertionsToFinish(WriteRqst)
2321 : * must be called before grabbing the lock, to make sure the data is ready to
2322 : * write.
2323 : */
2324 : static void
2325 2371601 : XLogWrite(XLogwrtRqst WriteRqst, TimeLineID tli, bool flexible)
2326 : {
2327 : bool ispartialpage;
2328 : bool last_iteration;
2329 : bool finishing_seg;
2330 : int curridx;
2331 : int npages;
2332 : int startidx;
2333 : uint32 startoffset;
2334 :
2335 : /* We should always be inside a critical section here */
2336 : Assert(CritSectionCount > 0);
2337 :
2338 : /*
2339 : * Update local LogwrtResult (caller probably did this already, but...)
2340 : */
2341 2371601 : RefreshXLogWriteResult(LogwrtResult);
2342 :
2343 : /*
2344 : * Since successive pages in the xlog cache are consecutively allocated,
2345 : * we can usually gather multiple pages together and issue just one
2346 : * write() call. npages is the number of pages we have determined can be
2347 : * written together; startidx is the cache block index of the first one,
2348 : * and startoffset is the file offset at which it should go. The latter
2349 : * two variables are only valid when npages > 0, but we must initialize
2350 : * all of them to keep the compiler quiet.
2351 : */
2352 2371601 : npages = 0;
2353 2371601 : startidx = 0;
2354 2371601 : startoffset = 0;
2355 :
2356 : /*
2357 : * Within the loop, curridx is the cache block index of the page to
2358 : * consider writing. Begin at the buffer containing the next unwritten
2359 : * page, or last partially written page.
2360 : */
2361 2371601 : curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2362 :
2363 5152792 : while (LogwrtResult.Write < WriteRqst.Write)
2364 : {
2365 : /*
2366 : * Make sure we're not ahead of the insert process. This could happen
2367 : * if we're passed a bogus WriteRqst.Write that is past the end of the
2368 : * last page that's been initialized by AdvanceXLInsertBuffer.
2369 : */
2370 2944352 : XLogRecPtr EndPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[curridx]);
2371 :
2372 2944352 : if (LogwrtResult.Write >= EndPtr)
2373 0 : elog(PANIC, "xlog write request %X/%08X is past end of log %X/%08X",
2374 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2375 : LSN_FORMAT_ARGS(EndPtr));
2376 :
2377 : /* Advance LogwrtResult.Write to end of current buffer page */
2378 2944352 : LogwrtResult.Write = EndPtr;
2379 2944352 : ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2380 :
2381 2944352 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2382 : wal_segment_size))
2383 : {
2384 : /*
2385 : * Switch to new logfile segment. We cannot have any pending
2386 : * pages here (since we dump what we have at segment end).
2387 : */
2388 : Assert(npages == 0);
2389 14993 : if (openLogFile >= 0)
2390 6639 : XLogFileClose();
2391 14993 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2392 : wal_segment_size);
2393 14993 : openLogTLI = tli;
2394 :
2395 : /* create/use new log file */
2396 14993 : openLogFile = XLogFileInit(openLogSegNo, tli);
2397 14993 : ReserveExternalFD();
2398 : }
2399 :
2400 : /* Make sure we have the current logfile open */
2401 2944352 : if (openLogFile < 0)
2402 : {
2403 0 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2404 : wal_segment_size);
2405 0 : openLogTLI = tli;
2406 0 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2407 0 : ReserveExternalFD();
2408 : }
2409 :
2410 : /* Add current page to the set of pending pages-to-dump */
2411 2944352 : if (npages == 0)
2412 : {
2413 : /* first of group */
2414 2389215 : startidx = curridx;
2415 2389215 : startoffset = XLogSegmentOffset(LogwrtResult.Write - XLOG_BLCKSZ,
2416 : wal_segment_size);
2417 : }
2418 2944352 : npages++;
2419 :
2420 : /*
2421 : * Dump the set if this will be the last loop iteration, or if we are
2422 : * at the last page of the cache area (since the next page won't be
2423 : * contiguous in memory), or if we are at the end of the logfile
2424 : * segment.
2425 : */
2426 2944352 : last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2427 :
2428 5729505 : finishing_seg = !ispartialpage &&
2429 2785153 : (startoffset + npages * XLOG_BLCKSZ) >= wal_segment_size;
2430 :
2431 2944352 : if (last_iteration ||
2432 574259 : curridx == XLogCtl->XLogCacheBlck ||
2433 : finishing_seg)
2434 : {
2435 : char *from;
2436 : Size nbytes;
2437 : Size nleft;
2438 : ssize_t written;
2439 : instr_time start;
2440 :
2441 : /* OK to write the page(s) */
2442 2389215 : from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2443 2389215 : nbytes = npages * (Size) XLOG_BLCKSZ;
2444 2389215 : nleft = nbytes;
2445 : do
2446 : {
2447 2389215 : errno = 0;
2448 :
2449 : /*
2450 : * Measure I/O timing to write WAL data, for pg_stat_io.
2451 : */
2452 2389215 : start = pgstat_prepare_io_time(track_wal_io_timing);
2453 :
2454 2389215 : pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE);
2455 2389215 : written = pg_pwrite(openLogFile, from, nleft, startoffset);
2456 2389215 : pgstat_report_wait_end();
2457 :
2458 2389215 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL,
2459 : IOOP_WRITE, start, 1, written);
2460 :
2461 2389215 : if (written <= 0)
2462 : {
2463 : char xlogfname[MAXFNAMELEN];
2464 : int save_errno;
2465 :
2466 0 : if (errno == EINTR)
2467 0 : continue;
2468 :
2469 0 : save_errno = errno;
2470 0 : XLogFileName(xlogfname, tli, openLogSegNo,
2471 : wal_segment_size);
2472 0 : errno = save_errno;
2473 0 : ereport(PANIC,
2474 : (errcode_for_file_access(),
2475 : errmsg("could not write to log file \"%s\" at offset %u, length %zu: %m",
2476 : xlogfname, startoffset, nleft)));
2477 : }
2478 2389215 : nleft -= written;
2479 2389215 : from += written;
2480 2389215 : startoffset += written;
2481 2389215 : } while (nleft > 0);
2482 :
2483 2389215 : npages = 0;
2484 :
2485 : /*
2486 : * If we just wrote the whole last page of a logfile segment,
2487 : * fsync the segment immediately. This avoids having to go back
2488 : * and re-open prior segments when an fsync request comes along
2489 : * later. Doing it here ensures that one and only one backend will
2490 : * perform this fsync.
2491 : *
2492 : * This is also the right place to notify the Archiver that the
2493 : * segment is ready to copy to archival storage, and to update the
2494 : * timer for archive_timeout, and to signal for a checkpoint if
2495 : * too many logfile segments have been used since the last
2496 : * checkpoint.
2497 : */
2498 2389215 : if (finishing_seg)
2499 : {
2500 2048 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2501 :
2502 : /* signal that we need to wakeup walsenders later */
2503 2048 : WalSndWakeupRequest();
2504 :
2505 2048 : LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2506 :
2507 2048 : if (XLogArchivingActive())
2508 415 : XLogArchiveNotifySeg(openLogSegNo, tli);
2509 :
2510 2048 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2511 2048 : XLogCtl->lastSegSwitchLSN = LogwrtResult.Flush;
2512 :
2513 : /*
2514 : * Request a checkpoint if we've consumed too much xlog since
2515 : * the last one. For speed, we first check using the local
2516 : * copy of RedoRecPtr, which might be out of date; if it looks
2517 : * like a checkpoint is needed, forcibly update RedoRecPtr and
2518 : * recheck.
2519 : */
2520 2048 : if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2521 : {
2522 277 : (void) GetRedoRecPtr();
2523 277 : if (XLogCheckpointNeeded(openLogSegNo))
2524 223 : RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2525 : }
2526 : }
2527 : }
2528 :
2529 2944352 : if (ispartialpage)
2530 : {
2531 : /* Only asked to write a partial page */
2532 159199 : LogwrtResult.Write = WriteRqst.Write;
2533 159199 : break;
2534 : }
2535 2785153 : curridx = NextBufIdx(curridx);
2536 :
2537 : /* If flexible, break out of loop as soon as we wrote something */
2538 2785153 : if (flexible && npages == 0)
2539 3962 : break;
2540 : }
2541 :
2542 : Assert(npages == 0);
2543 :
2544 : /*
2545 : * If asked to flush, do so
2546 : */
2547 2371601 : if (LogwrtResult.Flush < WriteRqst.Flush &&
2548 166965 : LogwrtResult.Flush < LogwrtResult.Write)
2549 : {
2550 : /*
2551 : * Could get here without iterating above loop, in which case we might
2552 : * have no open file or the wrong one. However, we do not need to
2553 : * fsync more than one file.
2554 : */
2555 166881 : if (wal_sync_method != WAL_SYNC_METHOD_OPEN &&
2556 166881 : wal_sync_method != WAL_SYNC_METHOD_OPEN_DSYNC)
2557 : {
2558 166881 : if (openLogFile >= 0 &&
2559 166861 : !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2560 : wal_segment_size))
2561 86 : XLogFileClose();
2562 166881 : if (openLogFile < 0)
2563 : {
2564 106 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2565 : wal_segment_size);
2566 106 : openLogTLI = tli;
2567 106 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2568 106 : ReserveExternalFD();
2569 : }
2570 :
2571 166881 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2572 : }
2573 :
2574 : /* signal that we need to wakeup walsenders later */
2575 166881 : WalSndWakeupRequest();
2576 :
2577 166881 : LogwrtResult.Flush = LogwrtResult.Write;
2578 : }
2579 :
2580 : /*
2581 : * Update shared-memory status
2582 : *
2583 : * We make sure that the shared 'request' values do not fall behind the
2584 : * 'result' values. This is not absolutely essential, but it saves some
2585 : * code in a couple of places.
2586 : */
2587 2371601 : SpinLockAcquire(&XLogCtl->info_lck);
2588 2371601 : if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
2589 146345 : XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
2590 2371601 : if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
2591 168529 : XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
2592 2371601 : SpinLockRelease(&XLogCtl->info_lck);
2593 :
2594 : /*
2595 : * We write Write first, bar, then Flush. When reading, the opposite must
2596 : * be done (with a matching barrier in between), so that we always see a
2597 : * Flush value that trails behind the Write value seen.
2598 : */
2599 2371601 : pg_atomic_write_u64(&XLogCtl->logWriteResult, LogwrtResult.Write);
2600 2371601 : pg_write_barrier();
2601 2371601 : pg_atomic_write_u64(&XLogCtl->logFlushResult, LogwrtResult.Flush);
2602 :
2603 : #ifdef USE_ASSERT_CHECKING
2604 : {
2605 : XLogRecPtr Flush;
2606 : XLogRecPtr Write;
2607 : XLogRecPtr Insert;
2608 :
2609 : Flush = pg_atomic_read_u64(&XLogCtl->logFlushResult);
2610 : pg_read_barrier();
2611 : Write = pg_atomic_read_u64(&XLogCtl->logWriteResult);
2612 : pg_read_barrier();
2613 : Insert = pg_atomic_read_u64(&XLogCtl->logInsertResult);
2614 :
2615 : /* WAL written to disk is always ahead of WAL flushed */
2616 : Assert(Write >= Flush);
2617 :
2618 : /* WAL inserted to buffers is always ahead of WAL written */
2619 : Assert(Insert >= Write);
2620 : }
2621 : #endif
2622 2371601 : }
2623 :
2624 : /*
2625 : * Record the LSN for an asynchronous transaction commit/abort
2626 : * and nudge the WALWriter if there is work for it to do.
2627 : * (This should not be called for synchronous commits.)
2628 : */
2629 : void
2630 62566 : XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2631 : {
2632 62566 : XLogRecPtr WriteRqstPtr = asyncXactLSN;
2633 : bool sleeping;
2634 62566 : bool wakeup = false;
2635 : XLogRecPtr prevAsyncXactLSN;
2636 :
2637 62566 : SpinLockAcquire(&XLogCtl->info_lck);
2638 62566 : sleeping = XLogCtl->WalWriterSleeping;
2639 62566 : prevAsyncXactLSN = XLogCtl->asyncXactLSN;
2640 62566 : if (XLogCtl->asyncXactLSN < asyncXactLSN)
2641 61994 : XLogCtl->asyncXactLSN = asyncXactLSN;
2642 62566 : SpinLockRelease(&XLogCtl->info_lck);
2643 :
2644 : /*
2645 : * If somebody else already called this function with a more aggressive
2646 : * LSN, they will have done what we needed (and perhaps more).
2647 : */
2648 62566 : if (asyncXactLSN <= prevAsyncXactLSN)
2649 572 : return;
2650 :
2651 : /*
2652 : * If the WALWriter is sleeping, kick it to make it come out of low-power
2653 : * mode, so that this async commit will reach disk within the expected
2654 : * amount of time. Otherwise, determine whether it has enough WAL
2655 : * available to flush, the same way that XLogBackgroundFlush() does.
2656 : */
2657 61994 : if (sleeping)
2658 48 : wakeup = true;
2659 : else
2660 : {
2661 : int flushblocks;
2662 :
2663 61946 : RefreshXLogWriteResult(LogwrtResult);
2664 :
2665 61946 : flushblocks =
2666 61946 : WriteRqstPtr / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
2667 :
2668 61946 : if (WalWriterFlushAfter == 0 || flushblocks >= WalWriterFlushAfter)
2669 5031 : wakeup = true;
2670 : }
2671 :
2672 61994 : if (wakeup)
2673 : {
2674 5079 : volatile PROC_HDR *procglobal = ProcGlobal;
2675 5079 : ProcNumber walwriterProc = procglobal->walwriterProc;
2676 :
2677 5079 : if (walwriterProc != INVALID_PROC_NUMBER)
2678 748 : SetLatch(&GetPGProcByNumber(walwriterProc)->procLatch);
2679 : }
2680 : }
2681 :
2682 : /*
2683 : * Record the LSN up to which we can remove WAL because it's not required by
2684 : * any replication slot.
2685 : */
2686 : void
2687 43764 : XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn)
2688 : {
2689 43764 : SpinLockAcquire(&XLogCtl->info_lck);
2690 43764 : XLogCtl->replicationSlotMinLSN = lsn;
2691 43764 : SpinLockRelease(&XLogCtl->info_lck);
2692 43764 : }
2693 :
2694 :
2695 : /*
2696 : * Return the oldest LSN we must retain to satisfy the needs of some
2697 : * replication slot.
2698 : */
2699 : XLogRecPtr
2700 2597 : XLogGetReplicationSlotMinimumLSN(void)
2701 : {
2702 : XLogRecPtr retval;
2703 :
2704 2597 : SpinLockAcquire(&XLogCtl->info_lck);
2705 2597 : retval = XLogCtl->replicationSlotMinLSN;
2706 2597 : SpinLockRelease(&XLogCtl->info_lck);
2707 :
2708 2597 : return retval;
2709 : }
2710 :
2711 : /*
2712 : * Advance minRecoveryPoint in control file.
2713 : *
2714 : * If we crash during recovery, we must reach this point again before the
2715 : * database is consistent.
2716 : *
2717 : * If 'force' is true, 'lsn' argument is ignored. Otherwise, minRecoveryPoint
2718 : * is only updated if it's not already greater than or equal to 'lsn'.
2719 : */
2720 : static void
2721 124175 : UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2722 : {
2723 : /* Quick check using our local copy of the variable */
2724 124175 : if (!updateMinRecoveryPoint || (!force && lsn <= LocalMinRecoveryPoint))
2725 116779 : return;
2726 :
2727 : /*
2728 : * An invalid minRecoveryPoint means that we need to recover all the WAL,
2729 : * i.e., we're doing crash recovery. We never modify the control file's
2730 : * value in that case, so we can short-circuit future checks here too. The
2731 : * local values of minRecoveryPoint and minRecoveryPointTLI should not be
2732 : * updated until crash recovery finishes. We only do this for the startup
2733 : * process as it should not update its own reference of minRecoveryPoint
2734 : * until it has finished crash recovery to make sure that all WAL
2735 : * available is replayed in this case. This also saves from extra locks
2736 : * taken on the control file from the startup process.
2737 : */
2738 7396 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint) && InRecovery)
2739 : {
2740 32 : updateMinRecoveryPoint = false;
2741 32 : return;
2742 : }
2743 :
2744 7364 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2745 :
2746 : /* update local copy */
2747 7364 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
2748 7364 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2749 :
2750 7364 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
2751 1 : updateMinRecoveryPoint = false;
2752 7363 : else if (force || LocalMinRecoveryPoint < lsn)
2753 : {
2754 : XLogRecPtr newMinRecoveryPoint;
2755 : TimeLineID newMinRecoveryPointTLI;
2756 :
2757 : /*
2758 : * To avoid having to update the control file too often, we update it
2759 : * all the way to the last record being replayed, even though 'lsn'
2760 : * would suffice for correctness. This also allows the 'force' case
2761 : * to not need a valid 'lsn' value.
2762 : *
2763 : * Another important reason for doing it this way is that the passed
2764 : * 'lsn' value could be bogus, i.e., past the end of available WAL, if
2765 : * the caller got it from a corrupted heap page. Accepting such a
2766 : * value as the min recovery point would prevent us from coming up at
2767 : * all. Instead, we just log a warning and continue with recovery.
2768 : * (See also the comments about corrupt LSNs in XLogFlush.)
2769 : */
2770 5891 : newMinRecoveryPoint = GetCurrentReplayRecPtr(&newMinRecoveryPointTLI);
2771 5891 : if (!force && newMinRecoveryPoint < lsn)
2772 0 : elog(WARNING,
2773 : "xlog min recovery request %X/%08X is past current point %X/%08X",
2774 : LSN_FORMAT_ARGS(lsn), LSN_FORMAT_ARGS(newMinRecoveryPoint));
2775 :
2776 : /* update control file */
2777 5891 : if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2778 : {
2779 5538 : ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2780 5538 : ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2781 5538 : UpdateControlFile();
2782 5538 : LocalMinRecoveryPoint = newMinRecoveryPoint;
2783 5538 : LocalMinRecoveryPointTLI = newMinRecoveryPointTLI;
2784 :
2785 5538 : ereport(DEBUG2,
2786 : errmsg_internal("updated min recovery point to %X/%08X on timeline %u",
2787 : LSN_FORMAT_ARGS(newMinRecoveryPoint),
2788 : newMinRecoveryPointTLI));
2789 : }
2790 : }
2791 7364 : LWLockRelease(ControlFileLock);
2792 : }
2793 :
2794 : /*
2795 : * Ensure that all XLOG data through the given position is flushed to disk.
2796 : *
2797 : * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
2798 : * already held, and we try to avoid acquiring it if possible.
2799 : */
2800 : void
2801 868803 : XLogFlush(XLogRecPtr record)
2802 : {
2803 : XLogRecPtr WriteRqstPtr;
2804 : XLogwrtRqst WriteRqst;
2805 868803 : TimeLineID insertTLI = XLogCtl->InsertTimeLineID;
2806 :
2807 : /*
2808 : * During REDO, we are reading not writing WAL. Therefore, instead of
2809 : * trying to flush the WAL, we should update minRecoveryPoint instead. We
2810 : * test XLogInsertAllowed(), not InRecovery, because we need checkpointer
2811 : * to act this way too, and because when it tries to write the
2812 : * end-of-recovery checkpoint, it should indeed flush.
2813 : */
2814 868803 : if (!XLogInsertAllowed())
2815 : {
2816 123716 : UpdateMinRecoveryPoint(record, false);
2817 690603 : return;
2818 : }
2819 :
2820 : /* Quick exit if already known flushed */
2821 745087 : if (record <= LogwrtResult.Flush)
2822 566887 : return;
2823 :
2824 : #ifdef WAL_DEBUG
2825 : if (XLOG_DEBUG)
2826 : elog(LOG, "xlog flush request %X/%08X; write %X/%08X; flush %X/%08X",
2827 : LSN_FORMAT_ARGS(record),
2828 : LSN_FORMAT_ARGS(LogwrtResult.Write),
2829 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2830 : #endif
2831 :
2832 178200 : START_CRIT_SECTION();
2833 :
2834 : /*
2835 : * Since fsync is usually a horribly expensive operation, we try to
2836 : * piggyback as much data as we can on each fsync: if we see any more data
2837 : * entered into the xlog buffer, we'll write and fsync that too, so that
2838 : * the final value of LogwrtResult.Flush is as large as possible. This
2839 : * gives us some chance of avoiding another fsync immediately after.
2840 : */
2841 :
2842 : /* initialize to given target; may increase below */
2843 178200 : WriteRqstPtr = record;
2844 :
2845 : /*
2846 : * Now wait until we get the write lock, or someone else does the flush
2847 : * for us.
2848 : */
2849 : for (;;)
2850 3382 : {
2851 : XLogRecPtr insertpos;
2852 :
2853 : /* done already? */
2854 181582 : RefreshXLogWriteResult(LogwrtResult);
2855 181582 : if (record <= LogwrtResult.Flush)
2856 12925 : break;
2857 :
2858 : /*
2859 : * Before actually performing the write, wait for all in-flight
2860 : * insertions to the pages we're about to write to finish.
2861 : */
2862 168657 : SpinLockAcquire(&XLogCtl->info_lck);
2863 168657 : if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
2864 11571 : WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
2865 168657 : SpinLockRelease(&XLogCtl->info_lck);
2866 168657 : insertpos = WaitXLogInsertionsToFinish(WriteRqstPtr);
2867 :
2868 : /*
2869 : * Try to get the write lock. If we can't get it immediately, wait
2870 : * until it's released, and recheck if we still need to do the flush
2871 : * or if the backend that held the lock did it for us already. This
2872 : * helps to maintain a good rate of group committing when the system
2873 : * is bottlenecked by the speed of fsyncing.
2874 : */
2875 168657 : if (!LWLockAcquireOrWait(WALWriteLock, LW_EXCLUSIVE))
2876 : {
2877 : /*
2878 : * The lock is now free, but we didn't acquire it yet. Before we
2879 : * do, loop back to check if someone else flushed the record for
2880 : * us already.
2881 : */
2882 3382 : continue;
2883 : }
2884 :
2885 : /* Got the lock; recheck whether request is satisfied */
2886 165275 : RefreshXLogWriteResult(LogwrtResult);
2887 165275 : if (record <= LogwrtResult.Flush)
2888 : {
2889 2368 : LWLockRelease(WALWriteLock);
2890 2368 : break;
2891 : }
2892 :
2893 : /*
2894 : * Sleep before flush! By adding a delay here, we may give further
2895 : * backends the opportunity to join the backlog of group commit
2896 : * followers; this can significantly improve transaction throughput,
2897 : * at the risk of increasing transaction latency.
2898 : *
2899 : * We do not sleep if enableFsync is not turned on, nor if there are
2900 : * fewer than CommitSiblings other backends with active transactions.
2901 : */
2902 162907 : if (CommitDelay > 0 && enableFsync &&
2903 0 : MinimumActiveBackends(CommitSiblings))
2904 : {
2905 0 : pgstat_report_wait_start(WAIT_EVENT_COMMIT_DELAY);
2906 0 : pg_usleep(CommitDelay);
2907 0 : pgstat_report_wait_end();
2908 :
2909 : /*
2910 : * Re-check how far we can now flush the WAL. It's generally not
2911 : * safe to call WaitXLogInsertionsToFinish while holding
2912 : * WALWriteLock, because an in-progress insertion might need to
2913 : * also grab WALWriteLock to make progress. But we know that all
2914 : * the insertions up to insertpos have already finished, because
2915 : * that's what the earlier WaitXLogInsertionsToFinish() returned.
2916 : * We're only calling it again to allow insertpos to be moved
2917 : * further forward, not to actually wait for anyone.
2918 : */
2919 0 : insertpos = WaitXLogInsertionsToFinish(insertpos);
2920 : }
2921 :
2922 : /* try to write/flush later additions to XLOG as well */
2923 162907 : WriteRqst.Write = insertpos;
2924 162907 : WriteRqst.Flush = insertpos;
2925 :
2926 162907 : XLogWrite(WriteRqst, insertTLI, false);
2927 :
2928 162907 : LWLockRelease(WALWriteLock);
2929 : /* done */
2930 162907 : break;
2931 : }
2932 :
2933 178200 : END_CRIT_SECTION();
2934 :
2935 : /* wake up walsenders now that we've released heavily contended locks */
2936 178200 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
2937 :
2938 : /*
2939 : * Wake up processes waiting for primary flush LSN to reach current flush
2940 : * position.
2941 : */
2942 178200 : WaitLSNWakeup(WAIT_LSN_TYPE_PRIMARY_FLUSH, LogwrtResult.Flush);
2943 :
2944 : /*
2945 : * If we still haven't flushed to the request point then we have a
2946 : * problem; most likely, the requested flush point is past end of XLOG.
2947 : * This has been seen to occur when a disk page has a corrupted LSN.
2948 : *
2949 : * Formerly we treated this as a PANIC condition, but that hurts the
2950 : * system's robustness rather than helping it: we do not want to take down
2951 : * the whole system due to corruption on one data page. In particular, if
2952 : * the bad page is encountered again during recovery then we would be
2953 : * unable to restart the database at all! (This scenario actually
2954 : * happened in the field several times with 7.1 releases.) As of 8.4, bad
2955 : * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
2956 : * the only time we can reach here during recovery is while flushing the
2957 : * end-of-recovery checkpoint record, and we don't expect that to have a
2958 : * bad LSN.
2959 : *
2960 : * Note that for calls from xact.c, the ERROR will be promoted to PANIC
2961 : * since xact.c calls this routine inside a critical section. However,
2962 : * calls from bufmgr.c are not within critical sections and so we will not
2963 : * force a restart for a bad LSN on a data page.
2964 : */
2965 178200 : if (LogwrtResult.Flush < record)
2966 0 : elog(ERROR,
2967 : "xlog flush request %X/%08X is not satisfied --- flushed only to %X/%08X",
2968 : LSN_FORMAT_ARGS(record),
2969 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2970 :
2971 : /*
2972 : * Cross-check XLogNeedsFlush(). Some of the checks of XLogFlush() and
2973 : * XLogNeedsFlush() are duplicated, and this assertion ensures that these
2974 : * remain consistent.
2975 : */
2976 : Assert(!XLogNeedsFlush(record));
2977 : }
2978 :
2979 : /*
2980 : * Write & flush xlog, but without specifying exactly where to.
2981 : *
2982 : * We normally write only completed blocks; but if there is nothing to do on
2983 : * that basis, we check for unwritten async commits in the current incomplete
2984 : * block, and write through the latest one of those. Thus, if async commits
2985 : * are not being used, we will write complete blocks only.
2986 : *
2987 : * If, based on the above, there's anything to write we do so immediately. But
2988 : * to avoid calling fsync, fdatasync et. al. at a rate that'd impact
2989 : * concurrent IO, we only flush WAL every wal_writer_delay ms, or if there's
2990 : * more than wal_writer_flush_after unflushed blocks.
2991 : *
2992 : * We can guarantee that async commits reach disk after at most three
2993 : * wal_writer_delay cycles. (When flushing complete blocks, we allow XLogWrite
2994 : * to write "flexibly", meaning it can stop at the end of the buffer ring;
2995 : * this makes a difference only with very high load or long wal_writer_delay,
2996 : * but imposes one extra cycle for the worst case for async commits.)
2997 : *
2998 : * This routine is invoked periodically by the background walwriter process.
2999 : *
3000 : * Returns true if there was any work to do, even if we skipped flushing due
3001 : * to wal_writer_delay/wal_writer_flush_after.
3002 : */
3003 : bool
3004 15820 : XLogBackgroundFlush(void)
3005 : {
3006 : XLogwrtRqst WriteRqst;
3007 15820 : bool flexible = true;
3008 : static TimestampTz lastflush;
3009 : TimestampTz now;
3010 : int flushblocks;
3011 : TimeLineID insertTLI;
3012 :
3013 : /* XLOG doesn't need flushing during recovery */
3014 15820 : if (RecoveryInProgress())
3015 0 : return false;
3016 :
3017 : /*
3018 : * Since we're not in recovery, InsertTimeLineID is set and can't change,
3019 : * so we can read it without a lock.
3020 : */
3021 15820 : insertTLI = XLogCtl->InsertTimeLineID;
3022 :
3023 : /* read updated LogwrtRqst */
3024 15820 : SpinLockAcquire(&XLogCtl->info_lck);
3025 15820 : WriteRqst = XLogCtl->LogwrtRqst;
3026 15820 : SpinLockRelease(&XLogCtl->info_lck);
3027 :
3028 : /* back off to last completed page boundary */
3029 15820 : WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
3030 :
3031 : /* if we have already flushed that far, consider async commit records */
3032 15820 : RefreshXLogWriteResult(LogwrtResult);
3033 15820 : if (WriteRqst.Write <= LogwrtResult.Flush)
3034 : {
3035 11739 : SpinLockAcquire(&XLogCtl->info_lck);
3036 11739 : WriteRqst.Write = XLogCtl->asyncXactLSN;
3037 11739 : SpinLockRelease(&XLogCtl->info_lck);
3038 11739 : flexible = false; /* ensure it all gets written */
3039 : }
3040 :
3041 : /*
3042 : * If already known flushed, we're done. Just need to check if we are
3043 : * holding an open file handle to a logfile that's no longer in use,
3044 : * preventing the file from being deleted.
3045 : */
3046 15820 : if (WriteRqst.Write <= LogwrtResult.Flush)
3047 : {
3048 10885 : if (openLogFile >= 0)
3049 : {
3050 6802 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
3051 : wal_segment_size))
3052 : {
3053 184 : XLogFileClose();
3054 : }
3055 : }
3056 10885 : return false;
3057 : }
3058 :
3059 : /*
3060 : * Determine how far to flush WAL, based on the wal_writer_delay and
3061 : * wal_writer_flush_after GUCs.
3062 : *
3063 : * Note that XLogSetAsyncXactLSN() performs similar calculation based on
3064 : * wal_writer_flush_after, to decide when to wake us up. Make sure the
3065 : * logic is the same in both places if you change this.
3066 : */
3067 4935 : now = GetCurrentTimestamp();
3068 4935 : flushblocks =
3069 4935 : WriteRqst.Write / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
3070 :
3071 4935 : if (WalWriterFlushAfter == 0 || lastflush == 0)
3072 : {
3073 : /* first call, or block based limits disabled */
3074 297 : WriteRqst.Flush = WriteRqst.Write;
3075 297 : lastflush = now;
3076 : }
3077 4638 : else if (TimestampDifferenceExceeds(lastflush, now, WalWriterDelay))
3078 : {
3079 : /*
3080 : * Flush the writes at least every WalWriterDelay ms. This is
3081 : * important to bound the amount of time it takes for an asynchronous
3082 : * commit to hit disk.
3083 : */
3084 4374 : WriteRqst.Flush = WriteRqst.Write;
3085 4374 : lastflush = now;
3086 : }
3087 264 : else if (flushblocks >= WalWriterFlushAfter)
3088 : {
3089 : /* exceeded wal_writer_flush_after blocks, flush */
3090 220 : WriteRqst.Flush = WriteRqst.Write;
3091 220 : lastflush = now;
3092 : }
3093 : else
3094 : {
3095 : /* no flushing, this time round */
3096 44 : WriteRqst.Flush = InvalidXLogRecPtr;
3097 : }
3098 :
3099 : #ifdef WAL_DEBUG
3100 : if (XLOG_DEBUG)
3101 : elog(LOG, "xlog bg flush request write %X/%08X; flush: %X/%08X, current is write %X/%08X; flush %X/%08X",
3102 : LSN_FORMAT_ARGS(WriteRqst.Write),
3103 : LSN_FORMAT_ARGS(WriteRqst.Flush),
3104 : LSN_FORMAT_ARGS(LogwrtResult.Write),
3105 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
3106 : #endif
3107 :
3108 4935 : START_CRIT_SECTION();
3109 :
3110 : /* now wait for any in-progress insertions to finish and get write lock */
3111 4935 : WaitXLogInsertionsToFinish(WriteRqst.Write);
3112 4935 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3113 4935 : RefreshXLogWriteResult(LogwrtResult);
3114 4935 : if (WriteRqst.Write > LogwrtResult.Write ||
3115 127 : WriteRqst.Flush > LogwrtResult.Flush)
3116 : {
3117 4862 : XLogWrite(WriteRqst, insertTLI, flexible);
3118 : }
3119 4935 : LWLockRelease(WALWriteLock);
3120 :
3121 4935 : END_CRIT_SECTION();
3122 :
3123 : /* wake up walsenders now that we've released heavily contended locks */
3124 4935 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3125 :
3126 : /*
3127 : * Wake up processes waiting for primary flush LSN to reach current flush
3128 : * position.
3129 : */
3130 4935 : WaitLSNWakeup(WAIT_LSN_TYPE_PRIMARY_FLUSH, LogwrtResult.Flush);
3131 :
3132 : /*
3133 : * Great, done. To take some work off the critical path, try to initialize
3134 : * as many of the no-longer-needed WAL buffers for future use as we can.
3135 : */
3136 4935 : AdvanceXLInsertBuffer(InvalidXLogRecPtr, insertTLI, true);
3137 :
3138 : /*
3139 : * If we determined that we need to write data, but somebody else
3140 : * wrote/flushed already, it should be considered as being active, to
3141 : * avoid hibernating too early.
3142 : */
3143 4935 : return true;
3144 : }
3145 :
3146 : /*
3147 : * Test whether XLOG data has been flushed up to (at least) the given
3148 : * position, or whether the minimum recovery point has been updated past
3149 : * the given position.
3150 : *
3151 : * Returns true if a flush is still needed, or if the minimum recovery point
3152 : * must be updated.
3153 : *
3154 : * It is possible that someone else is already in the process of flushing
3155 : * that far, or has updated the minimum recovery point up to the given
3156 : * position.
3157 : */
3158 : bool
3159 16281644 : XLogNeedsFlush(XLogRecPtr record)
3160 : {
3161 : /*
3162 : * During recovery, we don't flush WAL but update minRecoveryPoint
3163 : * instead. So "needs flush" is taken to mean whether minRecoveryPoint
3164 : * would need to be updated.
3165 : *
3166 : * Using XLogInsertAllowed() rather than RecoveryInProgress() matters for
3167 : * the case of an end-of-recovery checkpoint, where WAL data is flushed.
3168 : * This check should be consistent with the one in XLogFlush().
3169 : */
3170 16281644 : if (!XLogInsertAllowed())
3171 : {
3172 : /* Quick exit if already known to be updated or cannot be updated */
3173 538112 : if (!updateMinRecoveryPoint || record <= LocalMinRecoveryPoint)
3174 517063 : return false;
3175 :
3176 : /*
3177 : * An invalid minRecoveryPoint means that we need to recover all the
3178 : * WAL, i.e., we're doing crash recovery. We never modify the control
3179 : * file's value in that case, so we can short-circuit future checks
3180 : * here too. This triggers a quick exit path for the startup process,
3181 : * which cannot update its local copy of minRecoveryPoint as long as
3182 : * it has not replayed all WAL available when doing crash recovery.
3183 : */
3184 21049 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint) && InRecovery)
3185 : {
3186 0 : updateMinRecoveryPoint = false;
3187 0 : return false;
3188 : }
3189 :
3190 : /*
3191 : * Update local copy of minRecoveryPoint. But if the lock is busy,
3192 : * just return a conservative guess.
3193 : */
3194 21049 : if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3195 0 : return true;
3196 21049 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
3197 21049 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3198 21049 : LWLockRelease(ControlFileLock);
3199 :
3200 : /*
3201 : * Check minRecoveryPoint for any other process than the startup
3202 : * process doing crash recovery, which should not update the control
3203 : * file value if crash recovery is still running.
3204 : */
3205 21049 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
3206 0 : updateMinRecoveryPoint = false;
3207 :
3208 : /* check again */
3209 21049 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3210 106 : return false;
3211 : else
3212 20943 : return true;
3213 : }
3214 :
3215 : /* Quick exit if already known flushed */
3216 15743532 : if (record <= LogwrtResult.Flush)
3217 15479143 : return false;
3218 :
3219 : /* read LogwrtResult and update local state */
3220 264389 : RefreshXLogWriteResult(LogwrtResult);
3221 :
3222 : /* check again */
3223 264389 : if (record <= LogwrtResult.Flush)
3224 2915 : return false;
3225 :
3226 261474 : return true;
3227 : }
3228 :
3229 : /*
3230 : * Try to make a given XLOG file segment exist.
3231 : *
3232 : * logsegno: identify segment.
3233 : *
3234 : * *added: on return, true if this call raised the number of extant segments.
3235 : *
3236 : * path: on return, this char[MAXPGPATH] has the path to the logsegno file.
3237 : *
3238 : * Returns -1 or FD of opened file. A -1 here is not an error; a caller
3239 : * wanting an open segment should attempt to open "path", which usually will
3240 : * succeed. (This is weird, but it's efficient for the callers.)
3241 : */
3242 : static int
3243 16183 : XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,
3244 : bool *added, char *path)
3245 : {
3246 : char tmppath[MAXPGPATH];
3247 : XLogSegNo installed_segno;
3248 : XLogSegNo max_segno;
3249 : int fd;
3250 : int save_errno;
3251 16183 : int open_flags = O_RDWR | O_CREAT | O_EXCL | PG_BINARY;
3252 : instr_time io_start;
3253 :
3254 : Assert(logtli != 0);
3255 :
3256 16183 : XLogFilePath(path, logtli, logsegno, wal_segment_size);
3257 :
3258 : /*
3259 : * Try to use existent file (checkpoint maker may have created it already)
3260 : */
3261 16183 : *added = false;
3262 16183 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3263 16183 : get_sync_bit(wal_sync_method));
3264 16183 : if (fd < 0)
3265 : {
3266 1594 : if (errno != ENOENT)
3267 0 : ereport(ERROR,
3268 : (errcode_for_file_access(),
3269 : errmsg("could not open file \"%s\": %m", path)));
3270 : }
3271 : else
3272 14589 : return fd;
3273 :
3274 : /*
3275 : * Initialize an empty (all zeroes) segment. NOTE: it is possible that
3276 : * another process is doing the same thing. If so, we will end up
3277 : * pre-creating an extra log segment. That seems OK, and better than
3278 : * holding the lock throughout this lengthy process.
3279 : */
3280 1594 : elog(DEBUG2, "creating and filling new WAL file");
3281 :
3282 1594 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3283 :
3284 1594 : unlink(tmppath);
3285 :
3286 1594 : if (io_direct_flags & IO_DIRECT_WAL_INIT)
3287 0 : open_flags |= PG_O_DIRECT;
3288 :
3289 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3290 1594 : fd = BasicOpenFile(tmppath, open_flags);
3291 1594 : if (fd < 0)
3292 0 : ereport(ERROR,
3293 : (errcode_for_file_access(),
3294 : errmsg("could not create file \"%s\": %m", tmppath)));
3295 :
3296 : /* Measure I/O timing when initializing segment */
3297 1594 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3298 :
3299 1594 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE);
3300 1594 : save_errno = 0;
3301 1594 : if (wal_init_zero)
3302 : {
3303 : ssize_t rc;
3304 :
3305 : /*
3306 : * Zero-fill the file. With this setting, we do this the hard way to
3307 : * ensure that all the file space has really been allocated. On
3308 : * platforms that allow "holes" in files, just seeking to the end
3309 : * doesn't allocate intermediate space. This way, we know that we
3310 : * have all the space and (after the fsync below) that all the
3311 : * indirect blocks are down on disk. Therefore, fdatasync(2) or
3312 : * O_DSYNC will be sufficient to sync future writes to the log file.
3313 : */
3314 1594 : rc = pg_pwrite_zeros(fd, wal_segment_size, 0);
3315 :
3316 1594 : if (rc < 0)
3317 0 : save_errno = errno;
3318 : }
3319 : else
3320 : {
3321 : /*
3322 : * Otherwise, seeking to the end and writing a solitary byte is
3323 : * enough.
3324 : */
3325 0 : errno = 0;
3326 0 : if (pg_pwrite(fd, "\0", 1, wal_segment_size - 1) != 1)
3327 : {
3328 : /* if write didn't set errno, assume no disk space */
3329 0 : save_errno = errno ? errno : ENOSPC;
3330 : }
3331 : }
3332 1594 : pgstat_report_wait_end();
3333 :
3334 : /*
3335 : * A full segment worth of data is written when using wal_init_zero. One
3336 : * byte is written when not using it.
3337 : */
3338 1594 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT, IOOP_WRITE,
3339 : io_start, 1,
3340 1594 : wal_init_zero ? wal_segment_size : 1);
3341 :
3342 1594 : if (save_errno)
3343 : {
3344 : /*
3345 : * If we fail to make the file, delete it to release disk space
3346 : */
3347 0 : unlink(tmppath);
3348 :
3349 0 : close(fd);
3350 :
3351 0 : errno = save_errno;
3352 :
3353 0 : ereport(ERROR,
3354 : (errcode_for_file_access(),
3355 : errmsg("could not write to file \"%s\": %m", tmppath)));
3356 : }
3357 :
3358 : /* Measure I/O timing when flushing segment */
3359 1594 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3360 :
3361 1594 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC);
3362 1594 : if (pg_fsync(fd) != 0)
3363 : {
3364 0 : save_errno = errno;
3365 0 : close(fd);
3366 0 : errno = save_errno;
3367 0 : ereport(ERROR,
3368 : (errcode_for_file_access(),
3369 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3370 : }
3371 1594 : pgstat_report_wait_end();
3372 :
3373 1594 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT,
3374 : IOOP_FSYNC, io_start, 1, 0);
3375 :
3376 1594 : if (close(fd) != 0)
3377 0 : ereport(ERROR,
3378 : (errcode_for_file_access(),
3379 : errmsg("could not close file \"%s\": %m", tmppath)));
3380 :
3381 : /*
3382 : * Now move the segment into place with its final name. Cope with
3383 : * possibility that someone else has created the file while we were
3384 : * filling ours: if so, use ours to pre-create a future log segment.
3385 : */
3386 1594 : installed_segno = logsegno;
3387 :
3388 : /*
3389 : * XXX: What should we use as max_segno? We used to use XLOGfileslop when
3390 : * that was a constant, but that was always a bit dubious: normally, at a
3391 : * checkpoint, XLOGfileslop was the offset from the checkpoint record, but
3392 : * here, it was the offset from the insert location. We can't do the
3393 : * normal XLOGfileslop calculation here because we don't have access to
3394 : * the prior checkpoint's redo location. So somewhat arbitrarily, just use
3395 : * CheckPointSegments.
3396 : */
3397 1594 : max_segno = logsegno + CheckPointSegments;
3398 1594 : if (InstallXLogFileSegment(&installed_segno, tmppath, true, max_segno,
3399 : logtli))
3400 : {
3401 1594 : *added = true;
3402 1594 : elog(DEBUG2, "done creating and filling new WAL file");
3403 : }
3404 : else
3405 : {
3406 : /*
3407 : * No need for any more future segments, or InstallXLogFileSegment()
3408 : * failed to rename the file into place. If the rename failed, a
3409 : * caller opening the file may fail.
3410 : */
3411 0 : unlink(tmppath);
3412 0 : elog(DEBUG2, "abandoned new WAL file");
3413 : }
3414 :
3415 1594 : return -1;
3416 : }
3417 :
3418 : /*
3419 : * Create a new XLOG file segment, or open a pre-existing one.
3420 : *
3421 : * logsegno: identify segment to be created/opened.
3422 : *
3423 : * Returns FD of opened file.
3424 : *
3425 : * Note: errors here are ERROR not PANIC because we might or might not be
3426 : * inside a critical section (eg, during checkpoint there is no reason to
3427 : * take down the system on failure). They will promote to PANIC if we are
3428 : * in a critical section.
3429 : */
3430 : int
3431 15954 : XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)
3432 : {
3433 : bool ignore_added;
3434 : char path[MAXPGPATH];
3435 : int fd;
3436 :
3437 : Assert(logtli != 0);
3438 :
3439 15954 : fd = XLogFileInitInternal(logsegno, logtli, &ignore_added, path);
3440 15954 : if (fd >= 0)
3441 14429 : return fd;
3442 :
3443 : /* Now open original target segment (might not be file I just made) */
3444 1525 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3445 1525 : get_sync_bit(wal_sync_method));
3446 1525 : if (fd < 0)
3447 0 : ereport(ERROR,
3448 : (errcode_for_file_access(),
3449 : errmsg("could not open file \"%s\": %m", path)));
3450 1525 : return fd;
3451 : }
3452 :
3453 : /*
3454 : * Create a new XLOG file segment by copying a pre-existing one.
3455 : *
3456 : * destsegno: identify segment to be created.
3457 : *
3458 : * srcTLI, srcsegno: identify segment to be copied (could be from
3459 : * a different timeline)
3460 : *
3461 : * upto: how much of the source file to copy (the rest is filled with
3462 : * zeros)
3463 : *
3464 : * Currently this is only used during recovery, and so there are no locking
3465 : * considerations. But we should be just as tense as XLogFileInit to avoid
3466 : * emplacing a bogus file.
3467 : */
3468 : static void
3469 45 : XLogFileCopy(TimeLineID destTLI, XLogSegNo destsegno,
3470 : TimeLineID srcTLI, XLogSegNo srcsegno,
3471 : int upto)
3472 : {
3473 : char path[MAXPGPATH];
3474 : char tmppath[MAXPGPATH];
3475 : PGAlignedXLogBlock buffer;
3476 : int srcfd;
3477 : int fd;
3478 : int nbytes;
3479 :
3480 : /*
3481 : * Open the source file
3482 : */
3483 45 : XLogFilePath(path, srcTLI, srcsegno, wal_segment_size);
3484 45 : srcfd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
3485 45 : if (srcfd < 0)
3486 0 : ereport(ERROR,
3487 : (errcode_for_file_access(),
3488 : errmsg("could not open file \"%s\": %m", path)));
3489 :
3490 : /*
3491 : * Copy into a temp file name.
3492 : */
3493 45 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3494 :
3495 45 : unlink(tmppath);
3496 :
3497 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3498 45 : fd = OpenTransientFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
3499 45 : if (fd < 0)
3500 0 : ereport(ERROR,
3501 : (errcode_for_file_access(),
3502 : errmsg("could not create file \"%s\": %m", tmppath)));
3503 :
3504 : /*
3505 : * Do the data copying.
3506 : */
3507 92205 : for (nbytes = 0; nbytes < wal_segment_size; nbytes += sizeof(buffer))
3508 : {
3509 : int nread;
3510 :
3511 92160 : nread = upto - nbytes;
3512 :
3513 : /*
3514 : * The part that is not read from the source file is filled with
3515 : * zeros.
3516 : */
3517 92160 : if (nread < sizeof(buffer))
3518 45 : memset(buffer.data, 0, sizeof(buffer));
3519 :
3520 92160 : if (nread > 0)
3521 : {
3522 : int r;
3523 :
3524 3077 : if (nread > sizeof(buffer))
3525 3032 : nread = sizeof(buffer);
3526 3077 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ);
3527 3077 : r = read(srcfd, buffer.data, nread);
3528 3077 : if (r != nread)
3529 : {
3530 0 : if (r < 0)
3531 0 : ereport(ERROR,
3532 : (errcode_for_file_access(),
3533 : errmsg("could not read file \"%s\": %m",
3534 : path)));
3535 : else
3536 0 : ereport(ERROR,
3537 : (errcode(ERRCODE_DATA_CORRUPTED),
3538 : errmsg("could not read file \"%s\": read %d of %zu",
3539 : path, r, (Size) nread)));
3540 : }
3541 3077 : pgstat_report_wait_end();
3542 : }
3543 92160 : errno = 0;
3544 92160 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_WRITE);
3545 92160 : if ((int) write(fd, buffer.data, sizeof(buffer)) != (int) sizeof(buffer))
3546 : {
3547 0 : int save_errno = errno;
3548 :
3549 : /*
3550 : * If we fail to make the file, delete it to release disk space
3551 : */
3552 0 : unlink(tmppath);
3553 : /* if write didn't set errno, assume problem is no disk space */
3554 0 : errno = save_errno ? save_errno : ENOSPC;
3555 :
3556 0 : ereport(ERROR,
3557 : (errcode_for_file_access(),
3558 : errmsg("could not write to file \"%s\": %m", tmppath)));
3559 : }
3560 92160 : pgstat_report_wait_end();
3561 : }
3562 :
3563 45 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_SYNC);
3564 45 : if (pg_fsync(fd) != 0)
3565 0 : ereport(data_sync_elevel(ERROR),
3566 : (errcode_for_file_access(),
3567 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3568 45 : pgstat_report_wait_end();
3569 :
3570 45 : if (CloseTransientFile(fd) != 0)
3571 0 : ereport(ERROR,
3572 : (errcode_for_file_access(),
3573 : errmsg("could not close file \"%s\": %m", tmppath)));
3574 :
3575 45 : if (CloseTransientFile(srcfd) != 0)
3576 0 : ereport(ERROR,
3577 : (errcode_for_file_access(),
3578 : errmsg("could not close file \"%s\": %m", path)));
3579 :
3580 : /*
3581 : * Now move the segment into place with its final name.
3582 : */
3583 45 : if (!InstallXLogFileSegment(&destsegno, tmppath, false, 0, destTLI))
3584 0 : elog(ERROR, "InstallXLogFileSegment should not have failed");
3585 45 : }
3586 :
3587 : /*
3588 : * Install a new XLOG segment file as a current or future log segment.
3589 : *
3590 : * This is used both to install a newly-created segment (which has a temp
3591 : * filename while it's being created) and to recycle an old segment.
3592 : *
3593 : * *segno: identify segment to install as (or first possible target).
3594 : * When find_free is true, this is modified on return to indicate the
3595 : * actual installation location or last segment searched.
3596 : *
3597 : * tmppath: initial name of file to install. It will be renamed into place.
3598 : *
3599 : * find_free: if true, install the new segment at the first empty segno
3600 : * number at or after the passed numbers. If false, install the new segment
3601 : * exactly where specified, deleting any existing segment file there.
3602 : *
3603 : * max_segno: maximum segment number to install the new file as. Fail if no
3604 : * free slot is found between *segno and max_segno. (Ignored when find_free
3605 : * is false.)
3606 : *
3607 : * tli: The timeline on which the new segment should be installed.
3608 : *
3609 : * Returns true if the file was installed successfully. false indicates that
3610 : * max_segno limit was exceeded, the startup process has disabled this
3611 : * function for now, or an error occurred while renaming the file into place.
3612 : */
3613 : static bool
3614 3285 : InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
3615 : bool find_free, XLogSegNo max_segno, TimeLineID tli)
3616 : {
3617 : char path[MAXPGPATH];
3618 : struct stat stat_buf;
3619 :
3620 : Assert(tli != 0);
3621 :
3622 3285 : XLogFilePath(path, tli, *segno, wal_segment_size);
3623 :
3624 3285 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3625 3285 : if (!XLogCtl->InstallXLogFileSegmentActive)
3626 : {
3627 0 : LWLockRelease(ControlFileLock);
3628 0 : return false;
3629 : }
3630 :
3631 3285 : if (!find_free)
3632 : {
3633 : /* Force installation: get rid of any pre-existing segment file */
3634 45 : durable_unlink(path, DEBUG1);
3635 : }
3636 : else
3637 : {
3638 : /* Find a free slot to put it in */
3639 4443 : while (stat(path, &stat_buf) == 0)
3640 : {
3641 1392 : if ((*segno) >= max_segno)
3642 : {
3643 : /* Failed to find a free slot within specified range */
3644 189 : LWLockRelease(ControlFileLock);
3645 189 : return false;
3646 : }
3647 1203 : (*segno)++;
3648 1203 : XLogFilePath(path, tli, *segno, wal_segment_size);
3649 : }
3650 : }
3651 :
3652 : Assert(access(path, F_OK) != 0 && errno == ENOENT);
3653 3096 : if (durable_rename(tmppath, path, LOG) != 0)
3654 : {
3655 0 : LWLockRelease(ControlFileLock);
3656 : /* durable_rename already emitted log message */
3657 0 : return false;
3658 : }
3659 :
3660 3096 : LWLockRelease(ControlFileLock);
3661 :
3662 3096 : return true;
3663 : }
3664 :
3665 : /*
3666 : * Open a pre-existing logfile segment for writing.
3667 : */
3668 : int
3669 106 : XLogFileOpen(XLogSegNo segno, TimeLineID tli)
3670 : {
3671 : char path[MAXPGPATH];
3672 : int fd;
3673 :
3674 106 : XLogFilePath(path, tli, segno, wal_segment_size);
3675 :
3676 106 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3677 106 : get_sync_bit(wal_sync_method));
3678 106 : if (fd < 0)
3679 0 : ereport(PANIC,
3680 : (errcode_for_file_access(),
3681 : errmsg("could not open file \"%s\": %m", path)));
3682 :
3683 106 : return fd;
3684 : }
3685 :
3686 : /*
3687 : * Close the current logfile segment for writing.
3688 : */
3689 : static void
3690 6909 : XLogFileClose(void)
3691 : {
3692 : Assert(openLogFile >= 0);
3693 :
3694 : /*
3695 : * WAL segment files will not be re-read in normal operation, so we advise
3696 : * the OS to release any cached pages. But do not do so if WAL archiving
3697 : * or streaming is active, because archiver and walsender process could
3698 : * use the cache to read the WAL segment.
3699 : */
3700 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3701 6909 : if (!XLogIsNeeded() && (io_direct_flags & IO_DIRECT_WAL) == 0)
3702 142 : (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3703 : #endif
3704 :
3705 6909 : if (close(openLogFile) != 0)
3706 : {
3707 : char xlogfname[MAXFNAMELEN];
3708 0 : int save_errno = errno;
3709 :
3710 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo, wal_segment_size);
3711 0 : errno = save_errno;
3712 0 : ereport(PANIC,
3713 : (errcode_for_file_access(),
3714 : errmsg("could not close file \"%s\": %m", xlogfname)));
3715 : }
3716 :
3717 6909 : openLogFile = -1;
3718 6909 : ReleaseExternalFD();
3719 6909 : }
3720 :
3721 : /*
3722 : * Preallocate log files beyond the specified log endpoint.
3723 : *
3724 : * XXX this is currently extremely conservative, since it forces only one
3725 : * future log segment to exist, and even that only if we are 75% done with
3726 : * the current one. This is only appropriate for very low-WAL-volume systems.
3727 : * High-volume systems will be OK once they've built up a sufficient set of
3728 : * recycled log segments, but the startup transient is likely to include
3729 : * a lot of segment creations by foreground processes, which is not so good.
3730 : *
3731 : * XLogFileInitInternal() can ereport(ERROR). All known causes indicate big
3732 : * trouble; for example, a full filesystem is one cause. The checkpoint WAL
3733 : * and/or ControlFile updates already completed. If a RequestCheckpoint()
3734 : * initiated the present checkpoint and an ERROR ends this function, the
3735 : * command that called RequestCheckpoint() fails. That's not ideal, but it's
3736 : * not worth contorting more functions to use caller-specified elevel values.
3737 : * (With or without RequestCheckpoint(), an ERROR forestalls some inessential
3738 : * reporting and resource reclamation.)
3739 : */
3740 : static void
3741 2229 : PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli)
3742 : {
3743 : XLogSegNo _logSegNo;
3744 : int lf;
3745 : bool added;
3746 : char path[MAXPGPATH];
3747 : uint64 offset;
3748 :
3749 2229 : if (!XLogCtl->InstallXLogFileSegmentActive)
3750 12 : return; /* unlocked check says no */
3751 :
3752 2217 : XLByteToPrevSeg(endptr, _logSegNo, wal_segment_size);
3753 2217 : offset = XLogSegmentOffset(endptr - 1, wal_segment_size);
3754 2217 : if (offset >= (uint32) (0.75 * wal_segment_size))
3755 : {
3756 229 : _logSegNo++;
3757 229 : lf = XLogFileInitInternal(_logSegNo, tli, &added, path);
3758 229 : if (lf >= 0)
3759 160 : close(lf);
3760 229 : if (added)
3761 69 : CheckpointStats.ckpt_segs_added++;
3762 : }
3763 : }
3764 :
3765 : /*
3766 : * Throws an error if the given log segment has already been removed or
3767 : * recycled. The caller should only pass a segment that it knows to have
3768 : * existed while the server has been running, as this function always
3769 : * succeeds if no WAL segments have been removed since startup.
3770 : * 'tli' is only used in the error message.
3771 : *
3772 : * Note: this function guarantees to keep errno unchanged on return.
3773 : * This supports callers that use this to possibly deliver a better
3774 : * error message about a missing file, while still being able to throw
3775 : * a normal file-access error afterwards, if this does return.
3776 : */
3777 : void
3778 131466 : CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3779 : {
3780 131466 : int save_errno = errno;
3781 : XLogSegNo lastRemovedSegNo;
3782 :
3783 131466 : SpinLockAcquire(&XLogCtl->info_lck);
3784 131466 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3785 131466 : SpinLockRelease(&XLogCtl->info_lck);
3786 :
3787 131466 : if (segno <= lastRemovedSegNo)
3788 : {
3789 : char filename[MAXFNAMELEN];
3790 :
3791 0 : XLogFileName(filename, tli, segno, wal_segment_size);
3792 0 : errno = save_errno;
3793 0 : ereport(ERROR,
3794 : (errcode_for_file_access(),
3795 : errmsg("requested WAL segment %s has already been removed",
3796 : filename)));
3797 : }
3798 131466 : errno = save_errno;
3799 131466 : }
3800 :
3801 : /*
3802 : * Return the last WAL segment removed, or 0 if no segment has been removed
3803 : * since startup.
3804 : *
3805 : * NB: the result can be out of date arbitrarily fast, the caller has to deal
3806 : * with that.
3807 : */
3808 : XLogSegNo
3809 1303 : XLogGetLastRemovedSegno(void)
3810 : {
3811 : XLogSegNo lastRemovedSegNo;
3812 :
3813 1303 : SpinLockAcquire(&XLogCtl->info_lck);
3814 1303 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3815 1303 : SpinLockRelease(&XLogCtl->info_lck);
3816 :
3817 1303 : return lastRemovedSegNo;
3818 : }
3819 :
3820 : /*
3821 : * Return the oldest WAL segment on the given TLI that still exists in
3822 : * XLOGDIR, or 0 if none.
3823 : */
3824 : XLogSegNo
3825 7 : XLogGetOldestSegno(TimeLineID tli)
3826 : {
3827 : DIR *xldir;
3828 : struct dirent *xlde;
3829 7 : XLogSegNo oldest_segno = 0;
3830 :
3831 7 : xldir = AllocateDir(XLOGDIR);
3832 50 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3833 : {
3834 : TimeLineID file_tli;
3835 : XLogSegNo file_segno;
3836 :
3837 : /* Ignore files that are not XLOG segments. */
3838 43 : if (!IsXLogFileName(xlde->d_name))
3839 29 : continue;
3840 :
3841 : /* Parse filename to get TLI and segno. */
3842 14 : XLogFromFileName(xlde->d_name, &file_tli, &file_segno,
3843 : wal_segment_size);
3844 :
3845 : /* Ignore anything that's not from the TLI of interest. */
3846 14 : if (tli != file_tli)
3847 0 : continue;
3848 :
3849 : /* If it's the oldest so far, update oldest_segno. */
3850 14 : if (oldest_segno == 0 || file_segno < oldest_segno)
3851 9 : oldest_segno = file_segno;
3852 : }
3853 :
3854 7 : FreeDir(xldir);
3855 7 : return oldest_segno;
3856 : }
3857 :
3858 : /*
3859 : * Update the last removed segno pointer in shared memory, to reflect that the
3860 : * given XLOG file has been removed.
3861 : */
3862 : static void
3863 2676 : UpdateLastRemovedPtr(char *filename)
3864 : {
3865 : uint32 tli;
3866 : XLogSegNo segno;
3867 :
3868 2676 : XLogFromFileName(filename, &tli, &segno, wal_segment_size);
3869 :
3870 2676 : SpinLockAcquire(&XLogCtl->info_lck);
3871 2676 : if (segno > XLogCtl->lastRemovedSegNo)
3872 1171 : XLogCtl->lastRemovedSegNo = segno;
3873 2676 : SpinLockRelease(&XLogCtl->info_lck);
3874 2676 : }
3875 :
3876 : /*
3877 : * Remove all temporary log files in pg_wal
3878 : *
3879 : * This is called at the beginning of recovery after a previous crash,
3880 : * at a point where no other processes write fresh WAL data.
3881 : */
3882 : static void
3883 192 : RemoveTempXlogFiles(void)
3884 : {
3885 : DIR *xldir;
3886 : struct dirent *xlde;
3887 :
3888 192 : elog(DEBUG2, "removing all temporary WAL segments");
3889 :
3890 192 : xldir = AllocateDir(XLOGDIR);
3891 1296 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3892 : {
3893 : char path[MAXPGPATH];
3894 :
3895 1104 : if (strncmp(xlde->d_name, "xlogtemp.", 9) != 0)
3896 1104 : continue;
3897 :
3898 0 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
3899 0 : unlink(path);
3900 0 : elog(DEBUG2, "removed temporary WAL segment \"%s\"", path);
3901 : }
3902 192 : FreeDir(xldir);
3903 192 : }
3904 :
3905 : /*
3906 : * Recycle or remove all log files older or equal to passed segno.
3907 : *
3908 : * endptr is current (or recent) end of xlog, and lastredoptr is the
3909 : * redo pointer of the last checkpoint. These are used to determine
3910 : * whether we want to recycle rather than delete no-longer-wanted log files.
3911 : *
3912 : * insertTLI is the current timeline for XLOG insertion. Any recycled
3913 : * segments should be reused for this timeline.
3914 : */
3915 : static void
3916 1951 : RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr, XLogRecPtr endptr,
3917 : TimeLineID insertTLI)
3918 : {
3919 : DIR *xldir;
3920 : struct dirent *xlde;
3921 : char lastoff[MAXFNAMELEN];
3922 : XLogSegNo endlogSegNo;
3923 : XLogSegNo recycleSegNo;
3924 :
3925 : /* Initialize info about where to try to recycle to */
3926 1951 : XLByteToSeg(endptr, endlogSegNo, wal_segment_size);
3927 1951 : recycleSegNo = XLOGfileslop(lastredoptr);
3928 :
3929 : /*
3930 : * Construct a filename of the last segment to be kept. The timeline ID
3931 : * doesn't matter, we ignore that in the comparison. (During recovery,
3932 : * InsertTimeLineID isn't set, so we can't use that.)
3933 : */
3934 1951 : XLogFileName(lastoff, 0, segno, wal_segment_size);
3935 :
3936 1951 : elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
3937 : lastoff);
3938 :
3939 1951 : xldir = AllocateDir(XLOGDIR);
3940 :
3941 52414 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3942 : {
3943 : /* Ignore files that are not XLOG segments */
3944 50463 : if (!IsXLogFileName(xlde->d_name) &&
3945 8271 : !IsPartialXLogFileName(xlde->d_name))
3946 8269 : continue;
3947 :
3948 : /*
3949 : * We ignore the timeline part of the XLOG segment identifiers in
3950 : * deciding whether a segment is still needed. This ensures that we
3951 : * won't prematurely remove a segment from a parent timeline. We could
3952 : * probably be a little more proactive about removing segments of
3953 : * non-parent timelines, but that would be a whole lot more
3954 : * complicated.
3955 : *
3956 : * We use the alphanumeric sorting property of the filenames to decide
3957 : * which ones are earlier than the lastoff segment.
3958 : */
3959 42194 : if (strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
3960 : {
3961 35661 : if (XLogArchiveCheckDone(xlde->d_name))
3962 : {
3963 : /* Update the last removed location in shared memory first */
3964 2676 : UpdateLastRemovedPtr(xlde->d_name);
3965 :
3966 2676 : RemoveXlogFile(xlde, recycleSegNo, &endlogSegNo, insertTLI);
3967 : }
3968 : }
3969 : }
3970 :
3971 1951 : FreeDir(xldir);
3972 1951 : }
3973 :
3974 : /*
3975 : * Recycle or remove WAL files that are not part of the given timeline's
3976 : * history.
3977 : *
3978 : * This is called during recovery, whenever we switch to follow a new
3979 : * timeline, and at the end of recovery when we create a new timeline. We
3980 : * wouldn't otherwise care about extra WAL files lying in pg_wal, but they
3981 : * might be leftover pre-allocated or recycled WAL segments on the old timeline
3982 : * that we haven't used yet, and contain garbage. If we just leave them in
3983 : * pg_wal, they will eventually be archived, and we can't let that happen.
3984 : * Files that belong to our timeline history are valid, because we have
3985 : * successfully replayed them, but from others we can't be sure.
3986 : *
3987 : * 'switchpoint' is the current point in WAL where we switch to new timeline,
3988 : * and 'newTLI' is the new timeline we switch to.
3989 : */
3990 : void
3991 70 : RemoveNonParentXlogFiles(XLogRecPtr switchpoint, TimeLineID newTLI)
3992 : {
3993 : DIR *xldir;
3994 : struct dirent *xlde;
3995 : char switchseg[MAXFNAMELEN];
3996 : XLogSegNo endLogSegNo;
3997 : XLogSegNo switchLogSegNo;
3998 : XLogSegNo recycleSegNo;
3999 :
4000 : /*
4001 : * Initialize info about where to begin the work. This will recycle,
4002 : * somewhat arbitrarily, 10 future segments.
4003 : */
4004 70 : XLByteToPrevSeg(switchpoint, switchLogSegNo, wal_segment_size);
4005 70 : XLByteToSeg(switchpoint, endLogSegNo, wal_segment_size);
4006 70 : recycleSegNo = endLogSegNo + 10;
4007 :
4008 : /*
4009 : * Construct a filename of the last segment to be kept.
4010 : */
4011 70 : XLogFileName(switchseg, newTLI, switchLogSegNo, wal_segment_size);
4012 :
4013 70 : elog(DEBUG2, "attempting to remove WAL segments newer than log file %s",
4014 : switchseg);
4015 :
4016 70 : xldir = AllocateDir(XLOGDIR);
4017 :
4018 665 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4019 : {
4020 : /* Ignore files that are not XLOG segments */
4021 595 : if (!IsXLogFileName(xlde->d_name))
4022 368 : continue;
4023 :
4024 : /*
4025 : * Remove files that are on a timeline older than the new one we're
4026 : * switching to, but with a segment number >= the first segment on the
4027 : * new timeline.
4028 : */
4029 227 : if (strncmp(xlde->d_name, switchseg, 8) < 0 &&
4030 148 : strcmp(xlde->d_name + 8, switchseg + 8) > 0)
4031 : {
4032 : /*
4033 : * If the file has already been marked as .ready, however, don't
4034 : * remove it yet. It should be OK to remove it - files that are
4035 : * not part of our timeline history are not required for recovery
4036 : * - but seems safer to let them be archived and removed later.
4037 : */
4038 16 : if (!XLogArchiveIsReady(xlde->d_name))
4039 16 : RemoveXlogFile(xlde, recycleSegNo, &endLogSegNo, newTLI);
4040 : }
4041 : }
4042 :
4043 70 : FreeDir(xldir);
4044 70 : }
4045 :
4046 : /*
4047 : * Recycle or remove a log file that's no longer needed.
4048 : *
4049 : * segment_de is the dirent structure of the segment to recycle or remove.
4050 : * recycleSegNo is the segment number to recycle up to. endlogSegNo is
4051 : * the segment number of the current (or recent) end of WAL.
4052 : *
4053 : * endlogSegNo gets incremented if the segment is recycled so as it is not
4054 : * checked again with future callers of this function.
4055 : *
4056 : * insertTLI is the current timeline for XLOG insertion. Any recycled segments
4057 : * should be used for this timeline.
4058 : */
4059 : static void
4060 2692 : RemoveXlogFile(const struct dirent *segment_de,
4061 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
4062 : TimeLineID insertTLI)
4063 : {
4064 : char path[MAXPGPATH];
4065 : #ifdef WIN32
4066 : char newpath[MAXPGPATH];
4067 : #endif
4068 2692 : const char *segname = segment_de->d_name;
4069 :
4070 2692 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", segname);
4071 :
4072 : /*
4073 : * Before deleting the file, see if it can be recycled as a future log
4074 : * segment. Only recycle normal files, because we don't want to recycle
4075 : * symbolic links pointing to a separate archive directory.
4076 : */
4077 2692 : if (wal_recycle &&
4078 2692 : *endlogSegNo <= recycleSegNo &&
4079 3620 : XLogCtl->InstallXLogFileSegmentActive && /* callee rechecks this */
4080 3292 : get_dirent_type(path, segment_de, false, DEBUG2) == PGFILETYPE_REG &&
4081 1646 : InstallXLogFileSegment(endlogSegNo, path,
4082 : true, recycleSegNo, insertTLI))
4083 : {
4084 1457 : ereport(DEBUG2,
4085 : (errmsg_internal("recycled write-ahead log file \"%s\"",
4086 : segname)));
4087 1457 : CheckpointStats.ckpt_segs_recycled++;
4088 : /* Needn't recheck that slot on future iterations */
4089 1457 : (*endlogSegNo)++;
4090 : }
4091 : else
4092 : {
4093 : /* No need for any more future segments, or recycling failed ... */
4094 : int rc;
4095 :
4096 1235 : ereport(DEBUG2,
4097 : (errmsg_internal("removing write-ahead log file \"%s\"",
4098 : segname)));
4099 :
4100 : #ifdef WIN32
4101 :
4102 : /*
4103 : * On Windows, if another process (e.g another backend) holds the file
4104 : * open in FILE_SHARE_DELETE mode, unlink will succeed, but the file
4105 : * will still show up in directory listing until the last handle is
4106 : * closed. To avoid confusing the lingering deleted file for a live
4107 : * WAL file that needs to be archived, rename it before deleting it.
4108 : *
4109 : * If another process holds the file open without FILE_SHARE_DELETE
4110 : * flag, rename will fail. We'll try again at the next checkpoint.
4111 : */
4112 : snprintf(newpath, MAXPGPATH, "%s.deleted", path);
4113 : if (rename(path, newpath) != 0)
4114 : {
4115 : ereport(LOG,
4116 : (errcode_for_file_access(),
4117 : errmsg("could not rename file \"%s\": %m",
4118 : path)));
4119 : return;
4120 : }
4121 : rc = durable_unlink(newpath, LOG);
4122 : #else
4123 1235 : rc = durable_unlink(path, LOG);
4124 : #endif
4125 1235 : if (rc != 0)
4126 : {
4127 : /* Message already logged by durable_unlink() */
4128 0 : return;
4129 : }
4130 1235 : CheckpointStats.ckpt_segs_removed++;
4131 : }
4132 :
4133 2692 : XLogArchiveCleanup(segname);
4134 : }
4135 :
4136 : /*
4137 : * Verify whether pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
4138 : * If the latter do not exist, recreate them.
4139 : *
4140 : * It is not the goal of this function to verify the contents of these
4141 : * directories, but to help in cases where someone has performed a cluster
4142 : * copy for PITR purposes but omitted pg_wal from the copy.
4143 : *
4144 : * We could also recreate pg_wal if it doesn't exist, but a deliberate
4145 : * policy decision was made not to. It is fairly common for pg_wal to be
4146 : * a symlink, and if that was the DBA's intent then automatically making a
4147 : * plain directory would result in degraded performance with no notice.
4148 : */
4149 : static void
4150 1088 : ValidateXLOGDirectoryStructure(void)
4151 : {
4152 : char path[MAXPGPATH];
4153 : struct stat stat_buf;
4154 :
4155 : /* Check for pg_wal; if it doesn't exist, error out */
4156 1088 : if (stat(XLOGDIR, &stat_buf) != 0 ||
4157 1088 : !S_ISDIR(stat_buf.st_mode))
4158 0 : ereport(FATAL,
4159 : (errcode_for_file_access(),
4160 : errmsg("required WAL directory \"%s\" does not exist",
4161 : XLOGDIR)));
4162 :
4163 : /* Check for archive_status */
4164 1088 : snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4165 1088 : if (stat(path, &stat_buf) == 0)
4166 : {
4167 : /* Check for weird cases where it exists but isn't a directory */
4168 1087 : if (!S_ISDIR(stat_buf.st_mode))
4169 0 : ereport(FATAL,
4170 : (errcode_for_file_access(),
4171 : errmsg("required WAL directory \"%s\" does not exist",
4172 : path)));
4173 : }
4174 : else
4175 : {
4176 1 : ereport(LOG,
4177 : (errmsg("creating missing WAL directory \"%s\"", path)));
4178 1 : if (MakePGDirectory(path) < 0)
4179 0 : ereport(FATAL,
4180 : (errcode_for_file_access(),
4181 : errmsg("could not create missing directory \"%s\": %m",
4182 : path)));
4183 : }
4184 :
4185 : /* Check for summaries */
4186 1088 : snprintf(path, MAXPGPATH, XLOGDIR "/summaries");
4187 1088 : if (stat(path, &stat_buf) == 0)
4188 : {
4189 : /* Check for weird cases where it exists but isn't a directory */
4190 1087 : if (!S_ISDIR(stat_buf.st_mode))
4191 0 : ereport(FATAL,
4192 : (errmsg("required WAL directory \"%s\" does not exist",
4193 : path)));
4194 : }
4195 : else
4196 : {
4197 1 : ereport(LOG,
4198 : (errmsg("creating missing WAL directory \"%s\"", path)));
4199 1 : if (MakePGDirectory(path) < 0)
4200 0 : ereport(FATAL,
4201 : (errmsg("could not create missing directory \"%s\": %m",
4202 : path)));
4203 : }
4204 1088 : }
4205 :
4206 : /*
4207 : * Remove previous backup history files. This also retries creation of
4208 : * .ready files for any backup history files for which XLogArchiveNotify
4209 : * failed earlier.
4210 : */
4211 : static void
4212 164 : CleanupBackupHistory(void)
4213 : {
4214 : DIR *xldir;
4215 : struct dirent *xlde;
4216 : char path[MAXPGPATH + sizeof(XLOGDIR)];
4217 :
4218 164 : xldir = AllocateDir(XLOGDIR);
4219 :
4220 1685 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4221 : {
4222 1357 : if (IsBackupHistoryFileName(xlde->d_name))
4223 : {
4224 173 : if (XLogArchiveCheckDone(xlde->d_name))
4225 : {
4226 137 : elog(DEBUG2, "removing WAL backup history file \"%s\"",
4227 : xlde->d_name);
4228 137 : snprintf(path, sizeof(path), XLOGDIR "/%s", xlde->d_name);
4229 137 : unlink(path);
4230 137 : XLogArchiveCleanup(xlde->d_name);
4231 : }
4232 : }
4233 : }
4234 :
4235 164 : FreeDir(xldir);
4236 164 : }
4237 :
4238 : /*
4239 : * I/O routines for pg_control
4240 : *
4241 : * *ControlFile is a buffer in shared memory that holds an image of the
4242 : * contents of pg_control. WriteControlFile() initializes pg_control
4243 : * given a preloaded buffer, ReadControlFile() loads the buffer from
4244 : * the pg_control file (during postmaster or standalone-backend startup),
4245 : * and UpdateControlFile() rewrites pg_control after we modify xlog state.
4246 : * InitControlFile() fills the buffer with initial values.
4247 : *
4248 : * For simplicity, WriteControlFile() initializes the fields of pg_control
4249 : * that are related to checking backend/database compatibility, and
4250 : * ReadControlFile() verifies they are correct. We could split out the
4251 : * I/O and compatibility-check functions, but there seems no need currently.
4252 : */
4253 :
4254 : static void
4255 57 : InitControlFile(uint64 sysidentifier, uint32 data_checksum_version)
4256 : {
4257 : char mock_auth_nonce[MOCK_AUTH_NONCE_LEN];
4258 :
4259 : /*
4260 : * Generate a random nonce. This is used for authentication requests that
4261 : * will fail because the user does not exist. The nonce is used to create
4262 : * a genuine-looking password challenge for the non-existent user, in lieu
4263 : * of an actual stored password.
4264 : */
4265 57 : if (!pg_strong_random(mock_auth_nonce, MOCK_AUTH_NONCE_LEN))
4266 0 : ereport(PANIC,
4267 : (errcode(ERRCODE_INTERNAL_ERROR),
4268 : errmsg("could not generate secret authorization token")));
4269 :
4270 57 : memset(ControlFile, 0, sizeof(ControlFileData));
4271 : /* Initialize pg_control status fields */
4272 57 : ControlFile->system_identifier = sysidentifier;
4273 57 : memcpy(ControlFile->mock_authentication_nonce, mock_auth_nonce, MOCK_AUTH_NONCE_LEN);
4274 57 : ControlFile->state = DB_SHUTDOWNED;
4275 57 : ControlFile->unloggedLSN = FirstNormalUnloggedLSN;
4276 :
4277 : /* Set important parameter values for use when replaying WAL */
4278 57 : ControlFile->MaxConnections = MaxConnections;
4279 57 : ControlFile->max_worker_processes = max_worker_processes;
4280 57 : ControlFile->max_wal_senders = max_wal_senders;
4281 57 : ControlFile->max_prepared_xacts = max_prepared_xacts;
4282 57 : ControlFile->max_locks_per_xact = max_locks_per_xact;
4283 57 : ControlFile->wal_level = wal_level;
4284 57 : ControlFile->wal_log_hints = wal_log_hints;
4285 57 : ControlFile->track_commit_timestamp = track_commit_timestamp;
4286 57 : ControlFile->data_checksum_version = data_checksum_version;
4287 :
4288 : /*
4289 : * Set the data_checksum_version value into XLogCtl, which is where all
4290 : * processes get the current value from.
4291 : */
4292 57 : XLogCtl->data_checksum_version = data_checksum_version;
4293 57 : }
4294 :
4295 : static void
4296 57 : WriteControlFile(void)
4297 : {
4298 : int fd;
4299 : char buffer[PG_CONTROL_FILE_SIZE]; /* need not be aligned */
4300 :
4301 : /*
4302 : * Initialize version and compatibility-check fields
4303 : */
4304 57 : ControlFile->pg_control_version = PG_CONTROL_VERSION;
4305 57 : ControlFile->catalog_version_no = CATALOG_VERSION_NO;
4306 :
4307 57 : ControlFile->maxAlign = MAXIMUM_ALIGNOF;
4308 57 : ControlFile->floatFormat = FLOATFORMAT_VALUE;
4309 :
4310 57 : ControlFile->blcksz = BLCKSZ;
4311 57 : ControlFile->relseg_size = RELSEG_SIZE;
4312 57 : ControlFile->slru_pages_per_segment = SLRU_PAGES_PER_SEGMENT;
4313 57 : ControlFile->xlog_blcksz = XLOG_BLCKSZ;
4314 57 : ControlFile->xlog_seg_size = wal_segment_size;
4315 :
4316 57 : ControlFile->nameDataLen = NAMEDATALEN;
4317 57 : ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
4318 :
4319 57 : ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
4320 57 : ControlFile->loblksize = LOBLKSIZE;
4321 :
4322 57 : ControlFile->float8ByVal = true; /* vestigial */
4323 :
4324 : /*
4325 : * Initialize the default 'char' signedness.
4326 : *
4327 : * The signedness of the char type is implementation-defined. For instance
4328 : * on x86 architecture CPUs, the char data type is typically treated as
4329 : * signed by default, whereas on aarch architecture CPUs, it is typically
4330 : * treated as unsigned by default. In v17 or earlier, we accidentally let
4331 : * C implementation signedness affect persistent data. This led to
4332 : * inconsistent results when comparing char data across different
4333 : * platforms.
4334 : *
4335 : * This flag can be used as a hint to ensure consistent behavior for
4336 : * pre-v18 data files that store data sorted by the 'char' type on disk,
4337 : * especially in cross-platform replication scenarios.
4338 : *
4339 : * Newly created database clusters unconditionally set the default char
4340 : * signedness to true. pg_upgrade changes this flag for clusters that were
4341 : * initialized on signedness=false platforms. As a result,
4342 : * signedness=false setting will become rare over time. If we had known
4343 : * about this problem during the last development cycle that forced initdb
4344 : * (v8.3), we would have made all clusters signed or all clusters
4345 : * unsigned. Making pg_upgrade the only source of signedness=false will
4346 : * cause the population of database clusters to converge toward that
4347 : * retrospective ideal.
4348 : */
4349 57 : ControlFile->default_char_signedness = true;
4350 :
4351 : /* Contents are protected with a CRC */
4352 57 : INIT_CRC32C(ControlFile->crc);
4353 57 : COMP_CRC32C(ControlFile->crc,
4354 : ControlFile,
4355 : offsetof(ControlFileData, crc));
4356 57 : FIN_CRC32C(ControlFile->crc);
4357 :
4358 : /*
4359 : * We write out PG_CONTROL_FILE_SIZE bytes into pg_control, zero-padding
4360 : * the excess over sizeof(ControlFileData). This reduces the odds of
4361 : * premature-EOF errors when reading pg_control. We'll still fail when we
4362 : * check the contents of the file, but hopefully with a more specific
4363 : * error than "couldn't read pg_control".
4364 : */
4365 57 : memset(buffer, 0, PG_CONTROL_FILE_SIZE);
4366 57 : memcpy(buffer, ControlFile, sizeof(ControlFileData));
4367 :
4368 57 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4369 : O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
4370 57 : if (fd < 0)
4371 0 : ereport(PANIC,
4372 : (errcode_for_file_access(),
4373 : errmsg("could not create file \"%s\": %m",
4374 : XLOG_CONTROL_FILE)));
4375 :
4376 57 : errno = 0;
4377 57 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_WRITE);
4378 57 : if (write(fd, buffer, PG_CONTROL_FILE_SIZE) != PG_CONTROL_FILE_SIZE)
4379 : {
4380 : /* if write didn't set errno, assume problem is no disk space */
4381 0 : if (errno == 0)
4382 0 : errno = ENOSPC;
4383 0 : ereport(PANIC,
4384 : (errcode_for_file_access(),
4385 : errmsg("could not write to file \"%s\": %m",
4386 : XLOG_CONTROL_FILE)));
4387 : }
4388 57 : pgstat_report_wait_end();
4389 :
4390 57 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_SYNC);
4391 57 : if (pg_fsync(fd) != 0)
4392 0 : ereport(PANIC,
4393 : (errcode_for_file_access(),
4394 : errmsg("could not fsync file \"%s\": %m",
4395 : XLOG_CONTROL_FILE)));
4396 57 : pgstat_report_wait_end();
4397 :
4398 57 : if (close(fd) != 0)
4399 0 : ereport(PANIC,
4400 : (errcode_for_file_access(),
4401 : errmsg("could not close file \"%s\": %m",
4402 : XLOG_CONTROL_FILE)));
4403 57 : }
4404 :
4405 : static void
4406 1147 : ReadControlFile(void)
4407 : {
4408 : pg_crc32c crc;
4409 : int fd;
4410 : char wal_segsz_str[20];
4411 : int r;
4412 :
4413 : /*
4414 : * Read data...
4415 : */
4416 1147 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4417 : O_RDWR | PG_BINARY);
4418 1147 : if (fd < 0)
4419 0 : ereport(PANIC,
4420 : (errcode_for_file_access(),
4421 : errmsg("could not open file \"%s\": %m",
4422 : XLOG_CONTROL_FILE)));
4423 :
4424 1147 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ);
4425 1147 : r = read(fd, ControlFile, sizeof(ControlFileData));
4426 1147 : if (r != sizeof(ControlFileData))
4427 : {
4428 0 : if (r < 0)
4429 0 : ereport(PANIC,
4430 : (errcode_for_file_access(),
4431 : errmsg("could not read file \"%s\": %m",
4432 : XLOG_CONTROL_FILE)));
4433 : else
4434 0 : ereport(PANIC,
4435 : (errcode(ERRCODE_DATA_CORRUPTED),
4436 : errmsg("could not read file \"%s\": read %d of %zu",
4437 : XLOG_CONTROL_FILE, r, sizeof(ControlFileData))));
4438 : }
4439 1147 : pgstat_report_wait_end();
4440 :
4441 1147 : close(fd);
4442 :
4443 : /*
4444 : * Check for expected pg_control format version. If this is wrong, the
4445 : * CRC check will likely fail because we'll be checking the wrong number
4446 : * of bytes. Complaining about wrong version will probably be more
4447 : * enlightening than complaining about wrong CRC.
4448 : */
4449 :
4450 1147 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
4451 0 : ereport(FATAL,
4452 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4453 : errmsg("database files are incompatible with server"),
4454 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4455 : " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4456 : ControlFile->pg_control_version, ControlFile->pg_control_version,
4457 : PG_CONTROL_VERSION, PG_CONTROL_VERSION),
4458 : errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4459 :
4460 1147 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
4461 0 : ereport(FATAL,
4462 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4463 : errmsg("database files are incompatible with server"),
4464 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4465 : " but the server was compiled with PG_CONTROL_VERSION %d.",
4466 : ControlFile->pg_control_version, PG_CONTROL_VERSION),
4467 : errhint("It looks like you need to initdb.")));
4468 :
4469 : /* Now check the CRC. */
4470 1147 : INIT_CRC32C(crc);
4471 1147 : COMP_CRC32C(crc,
4472 : ControlFile,
4473 : offsetof(ControlFileData, crc));
4474 1147 : FIN_CRC32C(crc);
4475 :
4476 1147 : if (!EQ_CRC32C(crc, ControlFile->crc))
4477 0 : ereport(FATAL,
4478 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4479 : errmsg("incorrect checksum in control file")));
4480 :
4481 : /*
4482 : * Do compatibility checking immediately. If the database isn't
4483 : * compatible with the backend executable, we want to abort before we can
4484 : * possibly do any damage.
4485 : */
4486 1147 : if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
4487 0 : ereport(FATAL,
4488 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4489 : errmsg("database files are incompatible with server"),
4490 : /* translator: %s is a variable name and %d is its value */
4491 : errdetail("The database cluster was initialized with %s %d,"
4492 : " but the server was compiled with %s %d.",
4493 : "CATALOG_VERSION_NO", ControlFile->catalog_version_no,
4494 : "CATALOG_VERSION_NO", CATALOG_VERSION_NO),
4495 : errhint("It looks like you need to initdb.")));
4496 1147 : if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
4497 0 : ereport(FATAL,
4498 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4499 : errmsg("database files are incompatible with server"),
4500 : /* translator: %s is a variable name and %d is its value */
4501 : errdetail("The database cluster was initialized with %s %d,"
4502 : " but the server was compiled with %s %d.",
4503 : "MAXALIGN", ControlFile->maxAlign,
4504 : "MAXALIGN", MAXIMUM_ALIGNOF),
4505 : errhint("It looks like you need to initdb.")));
4506 1147 : if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
4507 0 : ereport(FATAL,
4508 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4509 : errmsg("database files are incompatible with server"),
4510 : errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4511 : errhint("It looks like you need to initdb.")));
4512 1147 : if (ControlFile->blcksz != BLCKSZ)
4513 0 : ereport(FATAL,
4514 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4515 : errmsg("database files are incompatible with server"),
4516 : /* translator: %s is a variable name and %d is its value */
4517 : errdetail("The database cluster was initialized with %s %d,"
4518 : " but the server was compiled with %s %d.",
4519 : "BLCKSZ", ControlFile->blcksz,
4520 : "BLCKSZ", BLCKSZ),
4521 : errhint("It looks like you need to recompile or initdb.")));
4522 1147 : if (ControlFile->relseg_size != RELSEG_SIZE)
4523 0 : ereport(FATAL,
4524 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4525 : errmsg("database files are incompatible with server"),
4526 : /* translator: %s is a variable name and %d is its value */
4527 : errdetail("The database cluster was initialized with %s %d,"
4528 : " but the server was compiled with %s %d.",
4529 : "RELSEG_SIZE", ControlFile->relseg_size,
4530 : "RELSEG_SIZE", RELSEG_SIZE),
4531 : errhint("It looks like you need to recompile or initdb.")));
4532 1147 : if (ControlFile->slru_pages_per_segment != SLRU_PAGES_PER_SEGMENT)
4533 0 : ereport(FATAL,
4534 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4535 : errmsg("database files are incompatible with server"),
4536 : /* translator: %s is a variable name and %d is its value */
4537 : errdetail("The database cluster was initialized with %s %d,"
4538 : " but the server was compiled with %s %d.",
4539 : "SLRU_PAGES_PER_SEGMENT", ControlFile->slru_pages_per_segment,
4540 : "SLRU_PAGES_PER_SEGMENT", SLRU_PAGES_PER_SEGMENT),
4541 : errhint("It looks like you need to recompile or initdb.")));
4542 1147 : if (ControlFile->xlog_blcksz != XLOG_BLCKSZ)
4543 0 : ereport(FATAL,
4544 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4545 : errmsg("database files are incompatible with server"),
4546 : /* translator: %s is a variable name and %d is its value */
4547 : errdetail("The database cluster was initialized with %s %d,"
4548 : " but the server was compiled with %s %d.",
4549 : "XLOG_BLCKSZ", ControlFile->xlog_blcksz,
4550 : "XLOG_BLCKSZ", XLOG_BLCKSZ),
4551 : errhint("It looks like you need to recompile or initdb.")));
4552 1147 : if (ControlFile->nameDataLen != NAMEDATALEN)
4553 0 : ereport(FATAL,
4554 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4555 : errmsg("database files are incompatible with server"),
4556 : /* translator: %s is a variable name and %d is its value */
4557 : errdetail("The database cluster was initialized with %s %d,"
4558 : " but the server was compiled with %s %d.",
4559 : "NAMEDATALEN", ControlFile->nameDataLen,
4560 : "NAMEDATALEN", NAMEDATALEN),
4561 : errhint("It looks like you need to recompile or initdb.")));
4562 1147 : if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
4563 0 : ereport(FATAL,
4564 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4565 : errmsg("database files are incompatible with server"),
4566 : /* translator: %s is a variable name and %d is its value */
4567 : errdetail("The database cluster was initialized with %s %d,"
4568 : " but the server was compiled with %s %d.",
4569 : "INDEX_MAX_KEYS", ControlFile->indexMaxKeys,
4570 : "INDEX_MAX_KEYS", INDEX_MAX_KEYS),
4571 : errhint("It looks like you need to recompile or initdb.")));
4572 1147 : if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
4573 0 : ereport(FATAL,
4574 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4575 : errmsg("database files are incompatible with server"),
4576 : /* translator: %s is a variable name and %d is its value */
4577 : errdetail("The database cluster was initialized with %s %d,"
4578 : " but the server was compiled with %s %d.",
4579 : "TOAST_MAX_CHUNK_SIZE", ControlFile->toast_max_chunk_size,
4580 : "TOAST_MAX_CHUNK_SIZE", (int) TOAST_MAX_CHUNK_SIZE),
4581 : errhint("It looks like you need to recompile or initdb.")));
4582 1147 : if (ControlFile->loblksize != LOBLKSIZE)
4583 0 : ereport(FATAL,
4584 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4585 : errmsg("database files are incompatible with server"),
4586 : /* translator: %s is a variable name and %d is its value */
4587 : errdetail("The database cluster was initialized with %s %d,"
4588 : " but the server was compiled with %s %d.",
4589 : "LOBLKSIZE", ControlFile->loblksize,
4590 : "LOBLKSIZE", (int) LOBLKSIZE),
4591 : errhint("It looks like you need to recompile or initdb.")));
4592 :
4593 : Assert(ControlFile->float8ByVal); /* vestigial, not worth an error msg */
4594 :
4595 1147 : wal_segment_size = ControlFile->xlog_seg_size;
4596 :
4597 1147 : if (!IsValidWalSegSize(wal_segment_size))
4598 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4599 : errmsg_plural("invalid WAL segment size in control file (%d byte)",
4600 : "invalid WAL segment size in control file (%d bytes)",
4601 : wal_segment_size,
4602 : wal_segment_size),
4603 : errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.")));
4604 :
4605 1147 : snprintf(wal_segsz_str, sizeof(wal_segsz_str), "%d", wal_segment_size);
4606 1147 : SetConfigOption("wal_segment_size", wal_segsz_str, PGC_INTERNAL,
4607 : PGC_S_DYNAMIC_DEFAULT);
4608 :
4609 : /* check and update variables dependent on wal_segment_size */
4610 1147 : if (ConvertToXSegs(min_wal_size_mb, wal_segment_size) < 2)
4611 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4612 : /* translator: both %s are GUC names */
4613 : errmsg("\"%s\" must be at least twice \"%s\"",
4614 : "min_wal_size", "wal_segment_size")));
4615 :
4616 1147 : if (ConvertToXSegs(max_wal_size_mb, wal_segment_size) < 2)
4617 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4618 : /* translator: both %s are GUC names */
4619 : errmsg("\"%s\" must be at least twice \"%s\"",
4620 : "max_wal_size", "wal_segment_size")));
4621 :
4622 1147 : UsableBytesInSegment =
4623 1147 : (wal_segment_size / XLOG_BLCKSZ * UsableBytesInPage) -
4624 : (SizeOfXLogLongPHD - SizeOfXLogShortPHD);
4625 :
4626 1147 : CalculateCheckpointSegments();
4627 1147 : }
4628 :
4629 : /*
4630 : * Utility wrapper to update the control file. Note that the control
4631 : * file gets flushed.
4632 : */
4633 : static void
4634 10285 : UpdateControlFile(void)
4635 : {
4636 10285 : update_controlfile(DataDir, ControlFile, true);
4637 10285 : }
4638 :
4639 : /*
4640 : * Returns the unique system identifier from control file.
4641 : */
4642 : uint64
4643 1590 : GetSystemIdentifier(void)
4644 : {
4645 : Assert(ControlFile != NULL);
4646 1590 : return ControlFile->system_identifier;
4647 : }
4648 :
4649 : /*
4650 : * Returns the random nonce from control file.
4651 : */
4652 : char *
4653 2 : GetMockAuthenticationNonce(void)
4654 : {
4655 : Assert(ControlFile != NULL);
4656 2 : return ControlFile->mock_authentication_nonce;
4657 : }
4658 :
4659 : /*
4660 : * DataChecksumsNeedWrite
4661 : * Returns whether data checksums must be written or not
4662 : *
4663 : * Returns true if data checksums are enabled, or are in the process of being
4664 : * enabled. During "inprogress-on" and "inprogress-off" states checksums must
4665 : * be written even though they are not verified (see datachecksum_state.c for
4666 : * a longer discussion).
4667 : *
4668 : * This function is intended for callsites which are about to write a data page
4669 : * to storage, and need to know whether to re-calculate the checksum for the
4670 : * page header. Calling this function must be performed as close to the write
4671 : * operation as possible to keep the critical section short.
4672 : */
4673 : bool
4674 841876 : DataChecksumsNeedWrite(void)
4675 : {
4676 938334 : return (LocalDataChecksumState == PG_DATA_CHECKSUM_VERSION ||
4677 897241 : LocalDataChecksumState == PG_DATA_CHECKSUM_INPROGRESS_ON ||
4678 55365 : LocalDataChecksumState == PG_DATA_CHECKSUM_INPROGRESS_OFF);
4679 : }
4680 :
4681 :
4682 : bool
4683 6 : DataChecksumsOff(void)
4684 : {
4685 : bool ret;
4686 :
4687 6 : SpinLockAcquire(&XLogCtl->info_lck);
4688 6 : ret = (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_OFF);
4689 6 : SpinLockRelease(&XLogCtl->info_lck);
4690 :
4691 6 : return ret;
4692 : }
4693 :
4694 : bool
4695 10 : DataChecksumsOn(void)
4696 : {
4697 : bool ret;
4698 :
4699 10 : SpinLockAcquire(&XLogCtl->info_lck);
4700 10 : ret = (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_VERSION);
4701 10 : SpinLockRelease(&XLogCtl->info_lck);
4702 :
4703 10 : return ret;
4704 : }
4705 :
4706 : bool
4707 165 : DataChecksumsInProgressOn(void)
4708 : {
4709 : bool ret;
4710 :
4711 165 : SpinLockAcquire(&XLogCtl->info_lck);
4712 165 : ret = (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_INPROGRESS_ON);
4713 165 : SpinLockRelease(&XLogCtl->info_lck);
4714 :
4715 165 : return ret;
4716 : }
4717 :
4718 : /*
4719 : * DataChecksumsNeedVerify
4720 : * Returns whether data checksums must be verified or not
4721 : *
4722 : * Data checksums are only verified if they are fully enabled in the cluster.
4723 : * During the "inprogress-on" and "inprogress-off" states they are only
4724 : * updated, not verified (see datachecksum_state.c for a longer discussion).
4725 : *
4726 : * This function is intended for callsites which have read data and are about
4727 : * to perform checksum validation based on the result of this. Calling this
4728 : * function must be performed as close to the validation call as possible to
4729 : * keep the critical section short. This is in order to protect against time of
4730 : * check/time of use situations around data checksum validation.
4731 : */
4732 : bool
4733 2664906 : DataChecksumsNeedVerify(void)
4734 : {
4735 2664906 : return (LocalDataChecksumState == PG_DATA_CHECKSUM_VERSION);
4736 : }
4737 :
4738 : /*
4739 : * SetDataChecksumsOnInProgress
4740 : * Sets the data checksum state to "inprogress-on" to enable checksums
4741 : *
4742 : * To start the process of enabling data checksums in a running cluster the
4743 : * data_checksum_version state must be changed to "inprogress-on". See
4744 : * SetDataChecksumsOn below for a description on how this state change works.
4745 : * This function blocks until all backends in the cluster have acknowledged the
4746 : * state transition.
4747 : */
4748 : void
4749 9 : SetDataChecksumsOnInProgress(void)
4750 : {
4751 : uint64 barrier;
4752 :
4753 : /*
4754 : * The state transition is performed in a critical section with
4755 : * checkpoints held off to provide crash safety.
4756 : */
4757 9 : START_CRIT_SECTION();
4758 9 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
4759 :
4760 9 : XLogChecksums(PG_DATA_CHECKSUM_INPROGRESS_ON);
4761 :
4762 9 : SpinLockAcquire(&XLogCtl->info_lck);
4763 9 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_INPROGRESS_ON;
4764 9 : SpinLockRelease(&XLogCtl->info_lck);
4765 :
4766 9 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
4767 9 : ControlFile->data_checksum_version = PG_DATA_CHECKSUM_INPROGRESS_ON;
4768 9 : UpdateControlFile();
4769 9 : LWLockRelease(ControlFileLock);
4770 :
4771 9 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_ON);
4772 :
4773 9 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
4774 9 : END_CRIT_SECTION();
4775 :
4776 9 : WaitForProcSignalBarrier(barrier);
4777 9 : }
4778 :
4779 : /*
4780 : * SetDataChecksumsOn
4781 : * Set data checksums state to 'on' cluster-wide
4782 : *
4783 : * Enabling data checksums is performed using two barriers, the first one to
4784 : * set the state to "inprogress-on" (done by SetDataChecksumsOnInProgress())
4785 : * and the second one to set the state to "on" (done here). Below is a short
4786 : * description of the processing, a more detailed write-up can be found in
4787 : * datachecksum_state.c.
4788 : *
4789 : * To start the process of enabling data checksums in a running cluster the
4790 : * data_checksum_version state must be changed to "inprogress-on". This state
4791 : * requires data checksums to be written but not verified. This ensures that
4792 : * all data pages can be checksummed without the risk of false negatives in
4793 : * validation during the process. When all existing pages are guaranteed to
4794 : * have checksums, and all new pages will be initiated with checksums, the
4795 : * state can be changed to "on". Once the state is "on" checksums will be both
4796 : * written and verified.
4797 : *
4798 : * This function blocks until all backends in the cluster have acknowledged the
4799 : * state transition.
4800 : */
4801 : void
4802 7 : SetDataChecksumsOn(void)
4803 : {
4804 : uint64 barrier;
4805 :
4806 7 : SpinLockAcquire(&XLogCtl->info_lck);
4807 :
4808 : /*
4809 : * The only allowed state transition to "on" is from "inprogress-on" since
4810 : * that state ensures that all pages will have data checksums written. No
4811 : * such state transition exists, if it does happen it's likely due to a
4812 : * programmer error.
4813 : */
4814 7 : if (XLogCtl->data_checksum_version != PG_DATA_CHECKSUM_INPROGRESS_ON)
4815 : {
4816 0 : SpinLockRelease(&XLogCtl->info_lck);
4817 0 : elog(WARNING,
4818 : "cannot set data checksums to \"on\", current state is not \"inprogress-on\", disabling");
4819 0 : SetDataChecksumsOff();
4820 0 : return;
4821 : }
4822 :
4823 7 : SpinLockRelease(&XLogCtl->info_lck);
4824 :
4825 7 : INJECTION_POINT("datachecksums-enable-checksums-delay", NULL);
4826 7 : START_CRIT_SECTION();
4827 7 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
4828 :
4829 7 : XLogChecksums(PG_DATA_CHECKSUM_VERSION);
4830 :
4831 7 : SpinLockAcquire(&XLogCtl->info_lck);
4832 7 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_VERSION;
4833 7 : SpinLockRelease(&XLogCtl->info_lck);
4834 :
4835 : /*
4836 : * Update the controlfile before waiting since if we have an immediate
4837 : * shutdown while waiting we want to come back up with checksums enabled.
4838 : */
4839 7 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
4840 7 : ControlFile->data_checksum_version = PG_DATA_CHECKSUM_VERSION;
4841 7 : UpdateControlFile();
4842 7 : LWLockRelease(ControlFileLock);
4843 :
4844 7 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_ON);
4845 :
4846 7 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
4847 7 : END_CRIT_SECTION();
4848 :
4849 7 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT | CHECKPOINT_FAST);
4850 7 : WaitForProcSignalBarrier(barrier);
4851 : }
4852 :
4853 : /*
4854 : * SetDataChecksumsOff
4855 : * Disables data checksums cluster-wide
4856 : *
4857 : * Disabling data checksums must be performed with two sets of barriers, each
4858 : * carrying a different state. The state is first set to "inprogress-off"
4859 : * during which checksums are still written but not verified. This ensures that
4860 : * backends which have yet to observe the state change from "on" won't get
4861 : * validation errors on concurrently modified pages. Once all backends have
4862 : * changed to "inprogress-off", the barrier for moving to "off" can be emitted.
4863 : * This function blocks until all backends in the cluster have acknowledged the
4864 : * state transition.
4865 : */
4866 : void
4867 6 : SetDataChecksumsOff(void)
4868 : {
4869 : uint64 barrier;
4870 :
4871 6 : SpinLockAcquire(&XLogCtl->info_lck);
4872 :
4873 : /* If data checksums are already disabled there is nothing to do */
4874 6 : if (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_OFF)
4875 : {
4876 0 : SpinLockRelease(&XLogCtl->info_lck);
4877 0 : return;
4878 : }
4879 :
4880 : /*
4881 : * If data checksums are currently enabled, or in the process of being
4882 : * enabled, we first transition to the "inprogress-off" state during which
4883 : * backends continue to write checksums without verifying them. When all
4884 : * backends are in "inprogress-off" the next transition to "off" can be
4885 : * performed, after which all data checksum processing is disabled.
4886 : */
4887 6 : if (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_VERSION ||
4888 2 : XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_INPROGRESS_ON)
4889 : {
4890 6 : SpinLockRelease(&XLogCtl->info_lck);
4891 :
4892 6 : START_CRIT_SECTION();
4893 6 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
4894 :
4895 6 : XLogChecksums(PG_DATA_CHECKSUM_INPROGRESS_OFF);
4896 :
4897 6 : SpinLockAcquire(&XLogCtl->info_lck);
4898 6 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_INPROGRESS_OFF;
4899 6 : SpinLockRelease(&XLogCtl->info_lck);
4900 :
4901 6 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
4902 6 : ControlFile->data_checksum_version = PG_DATA_CHECKSUM_INPROGRESS_OFF;
4903 6 : UpdateControlFile();
4904 6 : LWLockRelease(ControlFileLock);
4905 :
4906 6 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_OFF);
4907 :
4908 6 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
4909 6 : END_CRIT_SECTION();
4910 :
4911 6 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT | CHECKPOINT_FAST);
4912 6 : WaitForProcSignalBarrier(barrier);
4913 :
4914 : /*
4915 : * At this point we know that no backends are verifying data checksums
4916 : * during reading. Next, we can safely move to state "off" to also
4917 : * stop writing checksums.
4918 : */
4919 : }
4920 : else
4921 : {
4922 : /*
4923 : * Ending up here implies that the checksums state is "inprogress-off"
4924 : * and we can transition directly to "off" from there.
4925 : */
4926 0 : SpinLockRelease(&XLogCtl->info_lck);
4927 : }
4928 :
4929 6 : START_CRIT_SECTION();
4930 : /* Ensure that we don't incur a checkpoint during disabling checksums */
4931 6 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
4932 :
4933 6 : XLogChecksums(PG_DATA_CHECKSUM_OFF);
4934 :
4935 6 : SpinLockAcquire(&XLogCtl->info_lck);
4936 6 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_OFF;
4937 6 : SpinLockRelease(&XLogCtl->info_lck);
4938 :
4939 6 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
4940 6 : ControlFile->data_checksum_version = PG_DATA_CHECKSUM_OFF;
4941 6 : UpdateControlFile();
4942 6 : LWLockRelease(ControlFileLock);
4943 :
4944 6 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_OFF);
4945 :
4946 6 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
4947 6 : END_CRIT_SECTION();
4948 :
4949 6 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT | CHECKPOINT_FAST);
4950 6 : WaitForProcSignalBarrier(barrier);
4951 : }
4952 :
4953 : /*
4954 : * InitLocalDataChecksumState
4955 : *
4956 : * Set up backend local caches of controldata variables which may change at
4957 : * any point during runtime and thus require special cased locking. So far
4958 : * this only applies to data_checksum_version, but it's intended to be general
4959 : * purpose enough to handle future cases.
4960 : */
4961 : void
4962 24401 : InitLocalDataChecksumState(void)
4963 : {
4964 : Assert(InterruptHoldoffCount > 0);
4965 24401 : SpinLockAcquire(&XLogCtl->info_lck);
4966 24401 : SetLocalDataChecksumState(XLogCtl->data_checksum_version);
4967 24401 : SpinLockRelease(&XLogCtl->info_lck);
4968 24401 : }
4969 :
4970 : void
4971 27709 : SetLocalDataChecksumState(uint32 data_checksum_version)
4972 : {
4973 27709 : LocalDataChecksumState = data_checksum_version;
4974 :
4975 27709 : data_checksums = data_checksum_version;
4976 27709 : }
4977 :
4978 : /* guc hook */
4979 : const char *
4980 1966 : show_data_checksums(void)
4981 : {
4982 1966 : return get_checksum_state_string(LocalDataChecksumState);
4983 : }
4984 :
4985 : /*
4986 : * Return true if the cluster was initialized on a platform where the
4987 : * default signedness of char is "signed". This function exists for code
4988 : * that deals with pre-v18 data files that store data sorted by the 'char'
4989 : * type on disk (e.g., GIN and GiST indexes). See the comments in
4990 : * WriteControlFile() for details.
4991 : */
4992 : bool
4993 89903 : GetDefaultCharSignedness(void)
4994 : {
4995 89903 : return ControlFile->default_char_signedness;
4996 : }
4997 :
4998 : /*
4999 : * Returns a fake LSN for unlogged relations.
5000 : *
5001 : * Each call generates an LSN that is greater than any previous value
5002 : * returned. The current counter value is saved and restored across clean
5003 : * shutdowns, but like unlogged relations, does not survive a crash. This can
5004 : * be used in lieu of real LSN values returned by XLogInsert, if you need an
5005 : * LSN-like increasing sequence of numbers without writing any WAL.
5006 : */
5007 : XLogRecPtr
5008 202665 : GetFakeLSNForUnloggedRel(void)
5009 : {
5010 202665 : return pg_atomic_fetch_add_u64(&XLogCtl->unloggedLSN, 1);
5011 : }
5012 :
5013 : /*
5014 : * Auto-tune the number of XLOG buffers.
5015 : *
5016 : * The preferred setting for wal_buffers is about 3% of shared_buffers, with
5017 : * a maximum of one XLOG segment (there is little reason to think that more
5018 : * is helpful, at least so long as we force an fsync when switching log files)
5019 : * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
5020 : * 9.1, when auto-tuning was added).
5021 : *
5022 : * This should not be called until NBuffers has received its final value.
5023 : */
5024 : static int
5025 1243 : XLOGChooseNumBuffers(void)
5026 : {
5027 : int xbuffers;
5028 :
5029 1243 : xbuffers = NBuffers / 32;
5030 1243 : if (xbuffers > (wal_segment_size / XLOG_BLCKSZ))
5031 24 : xbuffers = (wal_segment_size / XLOG_BLCKSZ);
5032 1243 : if (xbuffers < 8)
5033 466 : xbuffers = 8;
5034 1243 : return xbuffers;
5035 : }
5036 :
5037 : /*
5038 : * GUC check_hook for wal_buffers
5039 : */
5040 : bool
5041 2532 : check_wal_buffers(int *newval, void **extra, GucSource source)
5042 : {
5043 : /*
5044 : * -1 indicates a request for auto-tune.
5045 : */
5046 2532 : if (*newval == -1)
5047 : {
5048 : /*
5049 : * If we haven't yet changed the boot_val default of -1, just let it
5050 : * be. We'll fix it when XLOGShmemRequest is called.
5051 : */
5052 1289 : if (XLOGbuffers == -1)
5053 1289 : return true;
5054 :
5055 : /* Otherwise, substitute the auto-tune value */
5056 0 : *newval = XLOGChooseNumBuffers();
5057 : }
5058 :
5059 : /*
5060 : * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
5061 : * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
5062 : * the case, we just silently treat such values as a request for the
5063 : * minimum. (We could throw an error instead, but that doesn't seem very
5064 : * helpful.)
5065 : */
5066 1243 : if (*newval < 4)
5067 0 : *newval = 4;
5068 :
5069 1243 : return true;
5070 : }
5071 :
5072 : /*
5073 : * GUC check_hook for wal_consistency_checking
5074 : */
5075 : bool
5076 2281 : check_wal_consistency_checking(char **newval, void **extra, GucSource source)
5077 : {
5078 : char *rawstring;
5079 : List *elemlist;
5080 : ListCell *l;
5081 : bool newwalconsistency[RM_MAX_ID + 1];
5082 :
5083 : /* Initialize the array */
5084 75273 : MemSet(newwalconsistency, 0, (RM_MAX_ID + 1) * sizeof(bool));
5085 :
5086 : /* Need a modifiable copy of string */
5087 2281 : rawstring = pstrdup(*newval);
5088 :
5089 : /* Parse string into list of identifiers */
5090 2281 : if (!SplitIdentifierString(rawstring, ',', &elemlist))
5091 : {
5092 : /* syntax error in list */
5093 0 : GUC_check_errdetail("List syntax is invalid.");
5094 0 : pfree(rawstring);
5095 0 : list_free(elemlist);
5096 0 : return false;
5097 : }
5098 :
5099 2780 : foreach(l, elemlist)
5100 : {
5101 499 : char *tok = (char *) lfirst(l);
5102 : int rmid;
5103 :
5104 : /* Check for 'all'. */
5105 499 : if (pg_strcasecmp(tok, "all") == 0)
5106 : {
5107 127729 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
5108 127232 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL)
5109 4970 : newwalconsistency[rmid] = true;
5110 : }
5111 : else
5112 : {
5113 : /* Check if the token matches any known resource manager. */
5114 2 : bool found = false;
5115 :
5116 36 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
5117 : {
5118 54 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL &&
5119 18 : pg_strcasecmp(tok, GetRmgr(rmid).rm_name) == 0)
5120 : {
5121 2 : newwalconsistency[rmid] = true;
5122 2 : found = true;
5123 2 : break;
5124 : }
5125 : }
5126 2 : if (!found)
5127 : {
5128 : /*
5129 : * During startup, it might be a not-yet-loaded custom
5130 : * resource manager. Defer checking until
5131 : * InitializeWalConsistencyChecking().
5132 : */
5133 0 : if (!process_shared_preload_libraries_done)
5134 : {
5135 0 : check_wal_consistency_checking_deferred = true;
5136 : }
5137 : else
5138 : {
5139 0 : GUC_check_errdetail("Unrecognized key word: \"%s\".", tok);
5140 0 : pfree(rawstring);
5141 0 : list_free(elemlist);
5142 0 : return false;
5143 : }
5144 : }
5145 : }
5146 : }
5147 :
5148 2281 : pfree(rawstring);
5149 2281 : list_free(elemlist);
5150 :
5151 : /* assign new value */
5152 2281 : *extra = guc_malloc(LOG, (RM_MAX_ID + 1) * sizeof(bool));
5153 2281 : if (!*extra)
5154 0 : return false;
5155 2281 : memcpy(*extra, newwalconsistency, (RM_MAX_ID + 1) * sizeof(bool));
5156 2281 : return true;
5157 : }
5158 :
5159 : /*
5160 : * GUC assign_hook for wal_consistency_checking
5161 : */
5162 : void
5163 2280 : assign_wal_consistency_checking(const char *newval, void *extra)
5164 : {
5165 : /*
5166 : * If some checks were deferred, it's possible that the checks will fail
5167 : * later during InitializeWalConsistencyChecking(). But in that case, the
5168 : * postmaster will exit anyway, so it's safe to proceed with the
5169 : * assignment.
5170 : *
5171 : * Any built-in resource managers specified are assigned immediately,
5172 : * which affects WAL created before shared_preload_libraries are
5173 : * processed. Any custom resource managers specified won't be assigned
5174 : * until after shared_preload_libraries are processed, but that's OK
5175 : * because WAL for a custom resource manager can't be written before the
5176 : * module is loaded anyway.
5177 : */
5178 2280 : wal_consistency_checking = extra;
5179 2280 : }
5180 :
5181 : /*
5182 : * InitializeWalConsistencyChecking: run after loading custom resource managers
5183 : *
5184 : * If any unknown resource managers were specified in the
5185 : * wal_consistency_checking GUC, processing was deferred. Now that
5186 : * shared_preload_libraries have been loaded, process wal_consistency_checking
5187 : * again.
5188 : */
5189 : void
5190 1072 : InitializeWalConsistencyChecking(void)
5191 : {
5192 : Assert(process_shared_preload_libraries_done);
5193 :
5194 1072 : if (check_wal_consistency_checking_deferred)
5195 : {
5196 : struct config_generic *guc;
5197 :
5198 0 : guc = find_option("wal_consistency_checking", false, false, ERROR);
5199 :
5200 0 : check_wal_consistency_checking_deferred = false;
5201 :
5202 0 : set_config_option_ext("wal_consistency_checking",
5203 : wal_consistency_checking_string,
5204 : guc->scontext, guc->source, guc->srole,
5205 : GUC_ACTION_SET, true, ERROR, false);
5206 :
5207 : /* checking should not be deferred again */
5208 : Assert(!check_wal_consistency_checking_deferred);
5209 : }
5210 1072 : }
5211 :
5212 : /*
5213 : * GUC show_hook for archive_command
5214 : */
5215 : const char *
5216 1964 : show_archive_command(void)
5217 : {
5218 1964 : if (XLogArchivingActive())
5219 106 : return XLogArchiveCommand;
5220 : else
5221 1858 : return "(disabled)";
5222 : }
5223 :
5224 : /*
5225 : * GUC show_hook for in_hot_standby
5226 : */
5227 : const char *
5228 17420 : show_in_hot_standby(void)
5229 : {
5230 : /*
5231 : * We display the actual state based on shared memory, so that this GUC
5232 : * reports up-to-date state if examined intra-query. The underlying
5233 : * variable (in_hot_standby_guc) changes only when we transmit a new value
5234 : * to the client.
5235 : */
5236 17420 : return RecoveryInProgress() ? "on" : "off";
5237 : }
5238 :
5239 : /*
5240 : * GUC show_hook for effective_wal_level
5241 : */
5242 : const char *
5243 2000 : show_effective_wal_level(void)
5244 : {
5245 2000 : if (wal_level == WAL_LEVEL_MINIMAL)
5246 295 : return "minimal";
5247 :
5248 : /*
5249 : * During recovery, effective_wal_level reflects the primary's
5250 : * configuration rather than the local wal_level value.
5251 : */
5252 1705 : if (RecoveryInProgress())
5253 30 : return IsXLogLogicalInfoEnabled() ? "logical" : "replica";
5254 :
5255 1675 : return XLogLogicalInfoActive() ? "logical" : "replica";
5256 : }
5257 :
5258 : /*
5259 : * Read the control file, set respective GUCs.
5260 : *
5261 : * This is to be called during startup, including a crash recovery cycle,
5262 : * unless in bootstrap mode, where no control file yet exists. As there's no
5263 : * usable shared memory yet (its sizing can depend on the contents of the
5264 : * control file!), first store the contents in local memory. XLOGShmemInit()
5265 : * will then copy it to shared memory later.
5266 : *
5267 : * reset just controls whether previous contents are to be expected (in the
5268 : * reset case, there's a dangling pointer into old shared memory), or not.
5269 : */
5270 : void
5271 1090 : LocalProcessControlFile(bool reset)
5272 : {
5273 : Assert(reset || ControlFile == NULL);
5274 1090 : LocalControlFile = palloc_object(ControlFileData);
5275 1090 : ControlFile = LocalControlFile;
5276 1090 : ReadControlFile();
5277 1090 : SetLocalDataChecksumState(ControlFile->data_checksum_version);
5278 1090 : }
5279 :
5280 : /*
5281 : * Get the wal_level from the control file. For a standby, this value should be
5282 : * considered as its active wal_level, because it may be different from what
5283 : * was originally configured on standby.
5284 : */
5285 : WalLevel
5286 0 : GetActiveWalLevelOnStandby(void)
5287 : {
5288 0 : return ControlFile->wal_level;
5289 : }
5290 :
5291 : /*
5292 : * Register shared memory for XLOG.
5293 : */
5294 : static void
5295 1248 : XLOGShmemRequest(void *arg)
5296 : {
5297 : Size size;
5298 :
5299 : /*
5300 : * If the value of wal_buffers is -1, use the preferred auto-tune value.
5301 : * This isn't an amazingly clean place to do this, but we must wait till
5302 : * NBuffers has received its final value, and must do it before using the
5303 : * value of XLOGbuffers to do anything important.
5304 : *
5305 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
5306 : * However, if the DBA explicitly set wal_buffers = -1 in the config file,
5307 : * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
5308 : * the matter with PGC_S_OVERRIDE.
5309 : */
5310 1248 : if (XLOGbuffers == -1)
5311 : {
5312 : char buf[32];
5313 :
5314 1243 : snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
5315 1243 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
5316 : PGC_S_DYNAMIC_DEFAULT);
5317 1243 : if (XLOGbuffers == -1) /* failed to apply it? */
5318 0 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
5319 : PGC_S_OVERRIDE);
5320 : }
5321 : Assert(XLOGbuffers > 0);
5322 :
5323 : /* XLogCtl */
5324 1248 : size = sizeof(XLogCtlData);
5325 :
5326 : /* WAL insertion locks, plus alignment */
5327 1248 : size = add_size(size, mul_size(sizeof(WALInsertLockPadded), NUM_XLOGINSERT_LOCKS + 1));
5328 : /* xlblocks array */
5329 1248 : size = add_size(size, mul_size(sizeof(pg_atomic_uint64), XLOGbuffers));
5330 : /* extra alignment padding for XLOG I/O buffers */
5331 1248 : size = add_size(size, Max(XLOG_BLCKSZ, PG_IO_ALIGN_SIZE));
5332 : /* and the buffers themselves */
5333 1248 : size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
5334 :
5335 1248 : ShmemRequestStruct(.name = "XLOG Ctl",
5336 : .size = size,
5337 : .ptr = (void **) &XLogCtl,
5338 : );
5339 1248 : ShmemRequestStruct(.name = "Control File",
5340 : .size = sizeof(ControlFileData),
5341 : .ptr = (void **) &ControlFile,
5342 : );
5343 1248 : }
5344 :
5345 : /*
5346 : * XLOGShmemInit - initialize the XLogCtl shared memory area.
5347 : */
5348 : static void
5349 1245 : XLOGShmemInit(void *arg)
5350 : {
5351 : char *allocptr;
5352 : int i;
5353 :
5354 : #ifdef WAL_DEBUG
5355 :
5356 : /*
5357 : * Create a memory context for WAL debugging that's exempt from the normal
5358 : * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
5359 : * an allocation fails, but wal_debug is not for production use anyway.
5360 : */
5361 : if (walDebugCxt == NULL)
5362 : {
5363 : walDebugCxt = AllocSetContextCreate(TopMemoryContext,
5364 : "WAL Debug",
5365 : ALLOCSET_DEFAULT_SIZES);
5366 : MemoryContextAllowInCriticalSection(walDebugCxt, true);
5367 : }
5368 : #endif
5369 :
5370 1245 : memset(XLogCtl, 0, sizeof(XLogCtlData));
5371 :
5372 : /*
5373 : * Already have read control file locally, unless in bootstrap mode. Move
5374 : * contents into shared memory.
5375 : */
5376 1245 : if (LocalControlFile)
5377 : {
5378 1074 : memcpy(ControlFile, LocalControlFile, sizeof(ControlFileData));
5379 1074 : pfree(LocalControlFile);
5380 1074 : LocalControlFile = NULL;
5381 : }
5382 :
5383 : /*
5384 : * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
5385 : * multiple of the alignment for same, so no extra alignment padding is
5386 : * needed here.
5387 : */
5388 1245 : allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
5389 1245 : XLogCtl->xlblocks = (pg_atomic_uint64 *) allocptr;
5390 1245 : allocptr += sizeof(pg_atomic_uint64) * XLOGbuffers;
5391 :
5392 365676 : for (i = 0; i < XLOGbuffers; i++)
5393 : {
5394 364431 : pg_atomic_init_u64(&XLogCtl->xlblocks[i], InvalidXLogRecPtr);
5395 : }
5396 :
5397 : /* WAL insertion locks. Ensure they're aligned to the full padded size */
5398 1245 : allocptr += sizeof(WALInsertLockPadded) -
5399 1245 : ((uintptr_t) allocptr) % sizeof(WALInsertLockPadded);
5400 1245 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks =
5401 : (WALInsertLockPadded *) allocptr;
5402 1245 : allocptr += sizeof(WALInsertLockPadded) * NUM_XLOGINSERT_LOCKS;
5403 :
5404 11205 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
5405 : {
5406 9960 : LWLockInitialize(&WALInsertLocks[i].l.lock, LWTRANCHE_WAL_INSERT);
5407 9960 : pg_atomic_init_u64(&WALInsertLocks[i].l.insertingAt, InvalidXLogRecPtr);
5408 9960 : WALInsertLocks[i].l.lastImportantAt = InvalidXLogRecPtr;
5409 : }
5410 :
5411 : /*
5412 : * Align the start of the page buffers to a full xlog block size boundary.
5413 : * This simplifies some calculations in XLOG insertion. It is also
5414 : * required for O_DIRECT.
5415 : */
5416 1245 : allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5417 1245 : XLogCtl->pages = allocptr;
5418 1245 : memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5419 :
5420 : /*
5421 : * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5422 : * in additional info.)
5423 : */
5424 1245 : XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5425 1245 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5426 1245 : XLogCtl->InstallXLogFileSegmentActive = false;
5427 1245 : XLogCtl->WalWriterSleeping = false;
5428 :
5429 : /* Use the checksum info from control file */
5430 1245 : XLogCtl->data_checksum_version = ControlFile->data_checksum_version;
5431 1245 : SetLocalDataChecksumState(XLogCtl->data_checksum_version);
5432 :
5433 1245 : SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5434 1245 : SpinLockInit(&XLogCtl->info_lck);
5435 1245 : pg_atomic_init_u64(&XLogCtl->logInsertResult, InvalidXLogRecPtr);
5436 1245 : pg_atomic_init_u64(&XLogCtl->logWriteResult, InvalidXLogRecPtr);
5437 1245 : pg_atomic_init_u64(&XLogCtl->logFlushResult, InvalidXLogRecPtr);
5438 1245 : pg_atomic_init_u64(&XLogCtl->unloggedLSN, InvalidXLogRecPtr);
5439 1245 : }
5440 :
5441 : /*
5442 : * XLOGShmemAttach - re-establish WALInsertLocks pointer after attaching.
5443 : */
5444 : static void
5445 0 : XLOGShmemAttach(void *arg)
5446 : {
5447 0 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks;
5448 0 : }
5449 :
5450 : /*
5451 : * This func must be called ONCE on system install. It creates pg_control
5452 : * and the initial XLOG segment.
5453 : */
5454 : void
5455 57 : BootStrapXLOG(uint32 data_checksum_version)
5456 : {
5457 : CheckPoint checkPoint;
5458 : PGAlignedXLogBlock buffer;
5459 : XLogPageHeader page;
5460 : XLogLongPageHeader longpage;
5461 : XLogRecord *record;
5462 : char *recptr;
5463 : uint64 sysidentifier;
5464 : struct timeval tv;
5465 : pg_crc32c crc;
5466 :
5467 : /* allow ordinary WAL segment creation, like StartupXLOG() would */
5468 57 : SetInstallXLogFileSegmentActive();
5469 :
5470 : /*
5471 : * Select a hopefully-unique system identifier code for this installation.
5472 : * We use the result of gettimeofday(), including the fractional seconds
5473 : * field, as being about as unique as we can easily get. (Think not to
5474 : * use random(), since it hasn't been seeded and there's no portable way
5475 : * to seed it other than the system clock value...) The upper half of the
5476 : * uint64 value is just the tv_sec part, while the lower half contains the
5477 : * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
5478 : * PID for a little extra uniqueness. A person knowing this encoding can
5479 : * determine the initialization time of the installation, which could
5480 : * perhaps be useful sometimes.
5481 : */
5482 57 : gettimeofday(&tv, NULL);
5483 57 : sysidentifier = ((uint64) tv.tv_sec) << 32;
5484 57 : sysidentifier |= ((uint64) tv.tv_usec) << 12;
5485 57 : sysidentifier |= getpid() & 0xFFF;
5486 :
5487 57 : memset(&buffer, 0, sizeof buffer);
5488 57 : page = (XLogPageHeader) &buffer;
5489 :
5490 : /*
5491 : * Set up information for the initial checkpoint record
5492 : *
5493 : * The initial checkpoint record is written to the beginning of the WAL
5494 : * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5495 : * used, so that we can use 0/0 to mean "before any valid WAL segment".
5496 : */
5497 57 : checkPoint.redo = wal_segment_size + SizeOfXLogLongPHD;
5498 57 : checkPoint.ThisTimeLineID = BootstrapTimeLineID;
5499 57 : checkPoint.PrevTimeLineID = BootstrapTimeLineID;
5500 57 : checkPoint.fullPageWrites = fullPageWrites;
5501 57 : checkPoint.logicalDecodingEnabled = (wal_level == WAL_LEVEL_LOGICAL);
5502 57 : checkPoint.wal_level = wal_level;
5503 : checkPoint.nextXid =
5504 57 : FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId);
5505 57 : checkPoint.nextOid = FirstGenbkiObjectId;
5506 57 : checkPoint.nextMulti = FirstMultiXactId;
5507 57 : checkPoint.nextMultiOffset = 1;
5508 57 : checkPoint.oldestXid = FirstNormalTransactionId;
5509 57 : checkPoint.oldestXidDB = Template1DbOid;
5510 57 : checkPoint.oldestMulti = FirstMultiXactId;
5511 57 : checkPoint.oldestMultiDB = Template1DbOid;
5512 57 : checkPoint.oldestCommitTsXid = InvalidTransactionId;
5513 57 : checkPoint.newestCommitTsXid = InvalidTransactionId;
5514 57 : checkPoint.time = (pg_time_t) time(NULL);
5515 57 : checkPoint.oldestActiveXid = InvalidTransactionId;
5516 57 : checkPoint.dataChecksumState = data_checksum_version;
5517 :
5518 57 : TransamVariables->nextXid = checkPoint.nextXid;
5519 57 : TransamVariables->nextOid = checkPoint.nextOid;
5520 57 : TransamVariables->oidCount = 0;
5521 57 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5522 57 : AdvanceOldestClogXid(checkPoint.oldestXid);
5523 57 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5524 57 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
5525 57 : SetCommitTsLimit(InvalidTransactionId, InvalidTransactionId);
5526 :
5527 : /* Set up the XLOG page header */
5528 57 : page->xlp_magic = XLOG_PAGE_MAGIC;
5529 57 : page->xlp_info = XLP_LONG_HEADER;
5530 57 : page->xlp_tli = BootstrapTimeLineID;
5531 57 : page->xlp_pageaddr = wal_segment_size;
5532 57 : longpage = (XLogLongPageHeader) page;
5533 57 : longpage->xlp_sysid = sysidentifier;
5534 57 : longpage->xlp_seg_size = wal_segment_size;
5535 57 : longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5536 :
5537 : /* Insert the initial checkpoint record */
5538 57 : recptr = ((char *) page + SizeOfXLogLongPHD);
5539 57 : record = (XLogRecord *) recptr;
5540 57 : record->xl_prev = InvalidXLogRecPtr;
5541 57 : record->xl_xid = InvalidTransactionId;
5542 57 : record->xl_tot_len = SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(checkPoint);
5543 57 : record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5544 57 : record->xl_rmid = RM_XLOG_ID;
5545 57 : recptr += SizeOfXLogRecord;
5546 : /* fill the XLogRecordDataHeaderShort struct */
5547 57 : *(recptr++) = (char) XLR_BLOCK_ID_DATA_SHORT;
5548 57 : *(recptr++) = sizeof(checkPoint);
5549 57 : memcpy(recptr, &checkPoint, sizeof(checkPoint));
5550 57 : recptr += sizeof(checkPoint);
5551 : Assert(recptr - (char *) record == record->xl_tot_len);
5552 :
5553 57 : INIT_CRC32C(crc);
5554 57 : COMP_CRC32C(crc, ((char *) record) + SizeOfXLogRecord, record->xl_tot_len - SizeOfXLogRecord);
5555 57 : COMP_CRC32C(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5556 57 : FIN_CRC32C(crc);
5557 57 : record->xl_crc = crc;
5558 :
5559 : /* Create first XLOG segment file */
5560 57 : openLogTLI = BootstrapTimeLineID;
5561 57 : openLogFile = XLogFileInit(1, BootstrapTimeLineID);
5562 :
5563 : /*
5564 : * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
5565 : * close the file again in a moment.
5566 : */
5567 :
5568 : /* Write the first page with the initial record */
5569 57 : errno = 0;
5570 57 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE);
5571 57 : if (write(openLogFile, &buffer, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5572 : {
5573 : /* if write didn't set errno, assume problem is no disk space */
5574 0 : if (errno == 0)
5575 0 : errno = ENOSPC;
5576 0 : ereport(PANIC,
5577 : (errcode_for_file_access(),
5578 : errmsg("could not write bootstrap write-ahead log file: %m")));
5579 : }
5580 57 : pgstat_report_wait_end();
5581 :
5582 57 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC);
5583 57 : if (pg_fsync(openLogFile) != 0)
5584 0 : ereport(PANIC,
5585 : (errcode_for_file_access(),
5586 : errmsg("could not fsync bootstrap write-ahead log file: %m")));
5587 57 : pgstat_report_wait_end();
5588 :
5589 57 : if (close(openLogFile) != 0)
5590 0 : ereport(PANIC,
5591 : (errcode_for_file_access(),
5592 : errmsg("could not close bootstrap write-ahead log file: %m")));
5593 :
5594 57 : openLogFile = -1;
5595 :
5596 : /* Now create pg_control */
5597 57 : InitControlFile(sysidentifier, data_checksum_version);
5598 57 : ControlFile->time = checkPoint.time;
5599 57 : ControlFile->checkPoint = checkPoint.redo;
5600 57 : ControlFile->checkPointCopy = checkPoint;
5601 :
5602 : /* some additional ControlFile fields are set in WriteControlFile() */
5603 57 : WriteControlFile();
5604 :
5605 : /* Bootstrap the commit log, too */
5606 57 : BootStrapCLOG();
5607 57 : BootStrapCommitTs();
5608 57 : BootStrapSUBTRANS();
5609 57 : BootStrapMultiXact();
5610 :
5611 : /*
5612 : * Force control file to be read - in contrast to normal processing we'd
5613 : * otherwise never run the checks and GUC related initializations therein.
5614 : */
5615 57 : ReadControlFile();
5616 57 : }
5617 :
5618 : static char *
5619 964 : str_time(pg_time_t tnow, char *buf, size_t bufsize)
5620 : {
5621 964 : pg_strftime(buf, bufsize,
5622 : "%Y-%m-%d %H:%M:%S %Z",
5623 964 : pg_localtime(&tnow, log_timezone));
5624 :
5625 964 : return buf;
5626 : }
5627 :
5628 : /*
5629 : * Initialize the first WAL segment on new timeline.
5630 : */
5631 : static void
5632 57 : XLogInitNewTimeline(TimeLineID endTLI, XLogRecPtr endOfLog, TimeLineID newTLI)
5633 : {
5634 : char xlogfname[MAXFNAMELEN];
5635 : XLogSegNo endLogSegNo;
5636 : XLogSegNo startLogSegNo;
5637 :
5638 : /* we always switch to a new timeline after archive recovery */
5639 : Assert(endTLI != newTLI);
5640 :
5641 : /*
5642 : * Update min recovery point one last time.
5643 : */
5644 57 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5645 :
5646 : /*
5647 : * Calculate the last segment on the old timeline, and the first segment
5648 : * on the new timeline. If the switch happens in the middle of a segment,
5649 : * they are the same, but if the switch happens exactly at a segment
5650 : * boundary, startLogSegNo will be endLogSegNo + 1.
5651 : */
5652 57 : XLByteToPrevSeg(endOfLog, endLogSegNo, wal_segment_size);
5653 57 : XLByteToSeg(endOfLog, startLogSegNo, wal_segment_size);
5654 :
5655 : /*
5656 : * Initialize the starting WAL segment for the new timeline. If the switch
5657 : * happens in the middle of a segment, copy data from the last WAL segment
5658 : * of the old timeline up to the switch point, to the starting WAL segment
5659 : * on the new timeline.
5660 : */
5661 57 : if (endLogSegNo == startLogSegNo)
5662 : {
5663 : /*
5664 : * Make a copy of the file on the new timeline.
5665 : *
5666 : * Writing WAL isn't allowed yet, so there are no locking
5667 : * considerations. But we should be just as tense as XLogFileInit to
5668 : * avoid emplacing a bogus file.
5669 : */
5670 45 : XLogFileCopy(newTLI, endLogSegNo, endTLI, endLogSegNo,
5671 45 : XLogSegmentOffset(endOfLog, wal_segment_size));
5672 : }
5673 : else
5674 : {
5675 : /*
5676 : * The switch happened at a segment boundary, so just create the next
5677 : * segment on the new timeline.
5678 : */
5679 : int fd;
5680 :
5681 12 : fd = XLogFileInit(startLogSegNo, newTLI);
5682 :
5683 12 : if (close(fd) != 0)
5684 : {
5685 0 : int save_errno = errno;
5686 :
5687 0 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5688 0 : errno = save_errno;
5689 0 : ereport(ERROR,
5690 : (errcode_for_file_access(),
5691 : errmsg("could not close file \"%s\": %m", xlogfname)));
5692 : }
5693 : }
5694 :
5695 : /*
5696 : * Let's just make real sure there are not .ready or .done flags posted
5697 : * for the new segment.
5698 : */
5699 57 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5700 57 : XLogArchiveCleanup(xlogfname);
5701 57 : }
5702 :
5703 : /*
5704 : * Perform cleanup actions at the conclusion of archive recovery.
5705 : */
5706 : static void
5707 57 : CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI, XLogRecPtr EndOfLog,
5708 : TimeLineID newTLI)
5709 : {
5710 : /*
5711 : * Execute the recovery_end_command, if any.
5712 : */
5713 57 : if (recoveryEndCommand && strcmp(recoveryEndCommand, "") != 0)
5714 2 : ExecuteRecoveryCommand(recoveryEndCommand,
5715 : "recovery_end_command",
5716 : true,
5717 : WAIT_EVENT_RECOVERY_END_COMMAND);
5718 :
5719 : /*
5720 : * We switched to a new timeline. Clean up segments on the old timeline.
5721 : *
5722 : * If there are any higher-numbered segments on the old timeline, remove
5723 : * them. They might contain valid WAL, but they might also be
5724 : * pre-allocated files containing garbage. In any case, they are not part
5725 : * of the new timeline's history so we don't need them.
5726 : */
5727 57 : RemoveNonParentXlogFiles(EndOfLog, newTLI);
5728 :
5729 : /*
5730 : * If the switch happened in the middle of a segment, what to do with the
5731 : * last, partial segment on the old timeline? If we don't archive it, and
5732 : * the server that created the WAL never archives it either (e.g. because
5733 : * it was hit by a meteor), it will never make it to the archive. That's
5734 : * OK from our point of view, because the new segment that we created with
5735 : * the new TLI contains all the WAL from the old timeline up to the switch
5736 : * point. But if you later try to do PITR to the "missing" WAL on the old
5737 : * timeline, recovery won't find it in the archive. It's physically
5738 : * present in the new file with new TLI, but recovery won't look there
5739 : * when it's recovering to the older timeline. On the other hand, if we
5740 : * archive the partial segment, and the original server on that timeline
5741 : * is still running and archives the completed version of the same segment
5742 : * later, it will fail. (We used to do that in 9.4 and below, and it
5743 : * caused such problems).
5744 : *
5745 : * As a compromise, we rename the last segment with the .partial suffix,
5746 : * and archive it. Archive recovery will never try to read .partial
5747 : * segments, so they will normally go unused. But in the odd PITR case,
5748 : * the administrator can copy them manually to the pg_wal directory
5749 : * (removing the suffix). They can be useful in debugging, too.
5750 : *
5751 : * If a .done or .ready file already exists for the old timeline, however,
5752 : * we had already determined that the segment is complete, so we can let
5753 : * it be archived normally. (In particular, if it was restored from the
5754 : * archive to begin with, it's expected to have a .done file).
5755 : */
5756 57 : if (XLogSegmentOffset(EndOfLog, wal_segment_size) != 0 &&
5757 : XLogArchivingActive())
5758 : {
5759 : char origfname[MAXFNAMELEN];
5760 : XLogSegNo endLogSegNo;
5761 :
5762 9 : XLByteToPrevSeg(EndOfLog, endLogSegNo, wal_segment_size);
5763 9 : XLogFileName(origfname, EndOfLogTLI, endLogSegNo, wal_segment_size);
5764 :
5765 9 : if (!XLogArchiveIsReadyOrDone(origfname))
5766 : {
5767 : char origpath[MAXPGPATH];
5768 : char partialfname[MAXFNAMELEN];
5769 : char partialpath[MAXPGPATH];
5770 :
5771 : /*
5772 : * If we're summarizing WAL, we can't rename the partial file
5773 : * until the summarizer finishes with it, else it will fail.
5774 : */
5775 5 : if (summarize_wal)
5776 1 : WaitForWalSummarization(EndOfLog);
5777 :
5778 5 : XLogFilePath(origpath, EndOfLogTLI, endLogSegNo, wal_segment_size);
5779 5 : snprintf(partialfname, MAXFNAMELEN, "%s.partial", origfname);
5780 5 : snprintf(partialpath, MAXPGPATH, "%s.partial", origpath);
5781 :
5782 : /*
5783 : * Make sure there's no .done or .ready file for the .partial
5784 : * file.
5785 : */
5786 5 : XLogArchiveCleanup(partialfname);
5787 :
5788 5 : durable_rename(origpath, partialpath, ERROR);
5789 5 : XLogArchiveNotify(partialfname);
5790 : }
5791 : }
5792 57 : }
5793 :
5794 : /*
5795 : * Check to see if required parameters are set high enough on this server
5796 : * for various aspects of recovery operation.
5797 : *
5798 : * Note that all the parameters which this function tests need to be
5799 : * listed in Administrator's Overview section in high-availability.sgml.
5800 : * If you change them, don't forget to update the list.
5801 : */
5802 : static void
5803 269 : CheckRequiredParameterValues(void)
5804 : {
5805 : /*
5806 : * For archive recovery, the WAL must be generated with at least 'replica'
5807 : * wal_level.
5808 : */
5809 269 : if (ArchiveRecoveryRequested && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5810 : {
5811 2 : ereport(FATAL,
5812 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5813 : errmsg("WAL was generated with \"wal_level=minimal\", cannot continue recovering"),
5814 : errdetail("This happens if you temporarily set \"wal_level=minimal\" on the server."),
5815 : errhint("Use a backup taken after setting \"wal_level\" to higher than \"minimal\".")));
5816 : }
5817 :
5818 : /*
5819 : * For Hot Standby, the WAL must be generated with 'replica' mode, and we
5820 : * must have at least as many backend slots as the primary.
5821 : */
5822 267 : if (ArchiveRecoveryRequested && EnableHotStandby)
5823 : {
5824 : /* We ignore autovacuum_worker_slots when we make this test. */
5825 143 : RecoveryRequiresIntParameter("max_connections",
5826 : MaxConnections,
5827 143 : ControlFile->MaxConnections);
5828 143 : RecoveryRequiresIntParameter("max_worker_processes",
5829 : max_worker_processes,
5830 143 : ControlFile->max_worker_processes);
5831 143 : RecoveryRequiresIntParameter("max_wal_senders",
5832 : max_wal_senders,
5833 143 : ControlFile->max_wal_senders);
5834 143 : RecoveryRequiresIntParameter("max_prepared_transactions",
5835 : max_prepared_xacts,
5836 143 : ControlFile->max_prepared_xacts);
5837 143 : RecoveryRequiresIntParameter("max_locks_per_transaction",
5838 : max_locks_per_xact,
5839 143 : ControlFile->max_locks_per_xact);
5840 : }
5841 267 : }
5842 :
5843 : /*
5844 : * This must be called ONCE during postmaster or standalone-backend startup
5845 : */
5846 : void
5847 1088 : StartupXLOG(void)
5848 : {
5849 : XLogCtlInsert *Insert;
5850 : CheckPoint checkPoint;
5851 : bool wasShutdown;
5852 : bool didCrash;
5853 : bool haveTblspcMap;
5854 : bool haveBackupLabel;
5855 : XLogRecPtr EndOfLog;
5856 : TimeLineID EndOfLogTLI;
5857 : TimeLineID newTLI;
5858 : bool performedWalRecovery;
5859 : EndOfWalRecoveryInfo *endOfRecoveryInfo;
5860 : XLogRecPtr abortedRecPtr;
5861 : XLogRecPtr missingContrecPtr;
5862 : TransactionId oldestActiveXID;
5863 1088 : bool promoted = false;
5864 : char timebuf[128];
5865 :
5866 : /*
5867 : * We should have an aux process resource owner to use, and we should not
5868 : * be in a transaction that's installed some other resowner.
5869 : */
5870 : Assert(AuxProcessResourceOwner != NULL);
5871 : Assert(CurrentResourceOwner == NULL ||
5872 : CurrentResourceOwner == AuxProcessResourceOwner);
5873 1088 : CurrentResourceOwner = AuxProcessResourceOwner;
5874 :
5875 : /*
5876 : * Check that contents look valid.
5877 : */
5878 1088 : if (!XRecOffIsValid(ControlFile->checkPoint))
5879 0 : ereport(FATAL,
5880 : (errcode(ERRCODE_DATA_CORRUPTED),
5881 : errmsg("control file contains invalid checkpoint location")));
5882 :
5883 1088 : switch (ControlFile->state)
5884 : {
5885 862 : case DB_SHUTDOWNED:
5886 :
5887 : /*
5888 : * This is the expected case, so don't be chatty in standalone
5889 : * mode
5890 : */
5891 862 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
5892 : (errmsg("database system was shut down at %s",
5893 : str_time(ControlFile->time,
5894 : timebuf, sizeof(timebuf)))));
5895 862 : break;
5896 :
5897 34 : case DB_SHUTDOWNED_IN_RECOVERY:
5898 34 : ereport(LOG,
5899 : (errmsg("database system was shut down in recovery at %s",
5900 : str_time(ControlFile->time,
5901 : timebuf, sizeof(timebuf)))));
5902 34 : break;
5903 :
5904 0 : case DB_SHUTDOWNING:
5905 0 : ereport(LOG,
5906 : (errmsg("database system shutdown was interrupted; last known up at %s",
5907 : str_time(ControlFile->time,
5908 : timebuf, sizeof(timebuf)))));
5909 0 : break;
5910 :
5911 0 : case DB_IN_CRASH_RECOVERY:
5912 0 : ereport(LOG,
5913 : (errmsg("database system was interrupted while in recovery at %s",
5914 : str_time(ControlFile->time,
5915 : timebuf, sizeof(timebuf))),
5916 : errhint("This probably means that some data is corrupted and"
5917 : " you will have to use the last backup for recovery.")));
5918 0 : break;
5919 :
5920 8 : case DB_IN_ARCHIVE_RECOVERY:
5921 8 : ereport(LOG,
5922 : (errmsg("database system was interrupted while in recovery at log time %s",
5923 : str_time(ControlFile->checkPointCopy.time,
5924 : timebuf, sizeof(timebuf))),
5925 : errhint("If this has occurred more than once some data might be corrupted"
5926 : " and you might need to choose an earlier recovery target.")));
5927 8 : break;
5928 :
5929 184 : case DB_IN_PRODUCTION:
5930 184 : ereport(LOG,
5931 : (errmsg("database system was interrupted; last known up at %s",
5932 : str_time(ControlFile->time,
5933 : timebuf, sizeof(timebuf)))));
5934 184 : break;
5935 :
5936 0 : default:
5937 0 : ereport(FATAL,
5938 : (errcode(ERRCODE_DATA_CORRUPTED),
5939 : errmsg("control file contains invalid database cluster state")));
5940 : }
5941 :
5942 : /* This is just to allow attaching to startup process with a debugger */
5943 : #ifdef XLOG_REPLAY_DELAY
5944 : if (ControlFile->state != DB_SHUTDOWNED)
5945 : pg_usleep(60000000L);
5946 : #endif
5947 :
5948 : /*
5949 : * Verify that pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
5950 : * In cases where someone has performed a copy for PITR, these directories
5951 : * may have been excluded and need to be re-created.
5952 : */
5953 1088 : ValidateXLOGDirectoryStructure();
5954 :
5955 : /* Set up timeout handler needed to report startup progress. */
5956 1088 : if (!IsBootstrapProcessingMode())
5957 1031 : RegisterTimeout(STARTUP_PROGRESS_TIMEOUT,
5958 : startup_progress_timeout_handler);
5959 :
5960 : /*----------
5961 : * If we previously crashed, perform a couple of actions:
5962 : *
5963 : * - The pg_wal directory may still include some temporary WAL segments
5964 : * used when creating a new segment, so perform some clean up to not
5965 : * bloat this path. This is done first as there is no point to sync
5966 : * this temporary data.
5967 : *
5968 : * - There might be data which we had written, intending to fsync it, but
5969 : * which we had not actually fsync'd yet. Therefore, a power failure in
5970 : * the near future might cause earlier unflushed writes to be lost, even
5971 : * though more recent data written to disk from here on would be
5972 : * persisted. To avoid that, fsync the entire data directory.
5973 : */
5974 1088 : if (ControlFile->state != DB_SHUTDOWNED &&
5975 226 : ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
5976 : {
5977 192 : RemoveTempXlogFiles();
5978 192 : SyncDataDirectory();
5979 192 : didCrash = true;
5980 : }
5981 : else
5982 896 : didCrash = false;
5983 :
5984 : /*
5985 : * Prepare for WAL recovery if needed.
5986 : *
5987 : * InitWalRecovery analyzes the control file and the backup label file, if
5988 : * any. It updates the in-memory ControlFile buffer according to the
5989 : * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5990 : * It also applies the tablespace map file, if any.
5991 : */
5992 1088 : InitWalRecovery(ControlFile, &wasShutdown,
5993 : &haveBackupLabel, &haveTblspcMap);
5994 1086 : checkPoint = ControlFile->checkPointCopy;
5995 :
5996 : /* initialize shared memory variables from the checkpoint record */
5997 1086 : TransamVariables->nextXid = checkPoint.nextXid;
5998 1086 : TransamVariables->nextOid = checkPoint.nextOid;
5999 1086 : TransamVariables->oidCount = 0;
6000 1086 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
6001 1086 : AdvanceOldestClogXid(checkPoint.oldestXid);
6002 1086 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
6003 1086 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
6004 1086 : SetCommitTsLimit(checkPoint.oldestCommitTsXid,
6005 : checkPoint.newestCommitTsXid);
6006 :
6007 : /*
6008 : * Clear out any old relcache cache files. This is *necessary* if we do
6009 : * any WAL replay, since that would probably result in the cache files
6010 : * being out of sync with database reality. In theory we could leave them
6011 : * in place if the database had been cleanly shut down, but it seems
6012 : * safest to just remove them always and let them be rebuilt during the
6013 : * first backend startup. These files needs to be removed from all
6014 : * directories including pg_tblspc, however the symlinks are created only
6015 : * after reading tablespace_map file in case of archive recovery from
6016 : * backup, so needs to clear old relcache files here after creating
6017 : * symlinks.
6018 : */
6019 1086 : RelationCacheInitFileRemove();
6020 :
6021 : /*
6022 : * Initialize replication slots, before there's a chance to remove
6023 : * required resources.
6024 : */
6025 1086 : StartupReplicationSlots();
6026 :
6027 : /*
6028 : * Startup the logical decoding status with the last status stored in the
6029 : * checkpoint record.
6030 : */
6031 1084 : StartupLogicalDecodingStatus(checkPoint.logicalDecodingEnabled);
6032 :
6033 : /*
6034 : * Startup logical state, needs to be setup now so we have proper data
6035 : * during crash recovery.
6036 : */
6037 1084 : StartupReorderBuffer();
6038 :
6039 : /*
6040 : * Startup CLOG. This must be done after TransamVariables->nextXid has
6041 : * been initialized and before we accept connections or begin WAL replay.
6042 : */
6043 1084 : StartupCLOG();
6044 :
6045 : /*
6046 : * Startup MultiXact. We need to do this early to be able to replay
6047 : * truncations.
6048 : */
6049 1084 : StartupMultiXact();
6050 :
6051 : /*
6052 : * Ditto for commit timestamps. Activate the facility if the setting is
6053 : * enabled in the control file, as there should be no tracking of commit
6054 : * timestamps done when the setting was disabled. This facility can be
6055 : * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
6056 : */
6057 1084 : if (ControlFile->track_commit_timestamp)
6058 14 : StartupCommitTs();
6059 :
6060 : /*
6061 : * Recover knowledge about replay progress of known replication partners.
6062 : */
6063 1084 : StartupReplicationOrigin();
6064 :
6065 : /*
6066 : * Initialize unlogged LSN. On a clean shutdown, it's restored from the
6067 : * control file. On recovery, all unlogged relations are blown away, so
6068 : * the unlogged LSN counter can be reset too.
6069 : */
6070 1084 : if (ControlFile->state == DB_SHUTDOWNED)
6071 853 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
6072 853 : ControlFile->unloggedLSN);
6073 : else
6074 231 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
6075 : FirstNormalUnloggedLSN);
6076 :
6077 : /*
6078 : * Copy any missing timeline history files between 'now' and the recovery
6079 : * target timeline from archive to pg_wal. While we don't need those files
6080 : * ourselves - the history file of the recovery target timeline covers all
6081 : * the previous timelines in the history too - a cascading standby server
6082 : * might be interested in them. Or, if you archive the WAL from this
6083 : * server to a different archive than the primary, it'd be good for all
6084 : * the history files to get archived there after failover, so that you can
6085 : * use one of the old timelines as a PITR target. Timeline history files
6086 : * are small, so it's better to copy them unnecessarily than not copy them
6087 : * and regret later.
6088 : */
6089 1084 : restoreTimeLineHistoryFiles(checkPoint.ThisTimeLineID, recoveryTargetTLI);
6090 :
6091 : /*
6092 : * Before running in recovery, scan pg_twophase and fill in its status to
6093 : * be able to work on entries generated by redo. Doing a scan before
6094 : * taking any recovery action has the merit to discard any 2PC files that
6095 : * are newer than the first record to replay, saving from any conflicts at
6096 : * replay. This avoids as well any subsequent scans when doing recovery
6097 : * of the on-disk two-phase data.
6098 : */
6099 1084 : restoreTwoPhaseData();
6100 :
6101 : /*
6102 : * When starting with crash recovery, reset pgstat data - it might not be
6103 : * valid. Otherwise restore pgstat data. It's safe to do this here,
6104 : * because postmaster will not yet have started any other processes.
6105 : *
6106 : * NB: Restoring replication slot stats relies on slot state to have
6107 : * already been restored from disk.
6108 : *
6109 : * TODO: With a bit of extra work we could just start with a pgstat file
6110 : * associated with the checkpoint redo location we're starting from.
6111 : */
6112 1084 : if (didCrash)
6113 190 : pgstat_discard_stats();
6114 : else
6115 894 : pgstat_restore_stats();
6116 :
6117 1084 : lastFullPageWrites = checkPoint.fullPageWrites;
6118 :
6119 1084 : RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
6120 1084 : doPageWrites = lastFullPageWrites;
6121 :
6122 : /* REDO */
6123 1084 : if (InRecovery)
6124 : {
6125 : /* Initialize state for RecoveryInProgress() */
6126 231 : SpinLockAcquire(&XLogCtl->info_lck);
6127 231 : if (InArchiveRecovery)
6128 129 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6129 : else
6130 102 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
6131 231 : SpinLockRelease(&XLogCtl->info_lck);
6132 :
6133 : /*
6134 : * Update pg_control to show that we are recovering and to show the
6135 : * selected checkpoint as the place we are starting from. We also mark
6136 : * pg_control with any minimum recovery stop point obtained from a
6137 : * backup history file.
6138 : *
6139 : * No need to hold ControlFileLock yet, we aren't up far enough.
6140 : */
6141 231 : UpdateControlFile();
6142 :
6143 : /*
6144 : * If there was a backup label file, it's done its job and the info
6145 : * has now been propagated into pg_control. We must get rid of the
6146 : * label file so that if we crash during recovery, we'll pick up at
6147 : * the latest recovery restartpoint instead of going all the way back
6148 : * to the backup start point. It seems prudent though to just rename
6149 : * the file out of the way rather than delete it completely.
6150 : */
6151 231 : if (haveBackupLabel)
6152 : {
6153 86 : unlink(BACKUP_LABEL_OLD);
6154 86 : durable_rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD, FATAL);
6155 : }
6156 :
6157 : /*
6158 : * If there was a tablespace_map file, it's done its job and the
6159 : * symlinks have been created. We must get rid of the map file so
6160 : * that if we crash during recovery, we don't create symlinks again.
6161 : * It seems prudent though to just rename the file out of the way
6162 : * rather than delete it completely.
6163 : */
6164 231 : if (haveTblspcMap)
6165 : {
6166 2 : unlink(TABLESPACE_MAP_OLD);
6167 2 : durable_rename(TABLESPACE_MAP, TABLESPACE_MAP_OLD, FATAL);
6168 : }
6169 :
6170 : /*
6171 : * Initialize our local copy of minRecoveryPoint. When doing crash
6172 : * recovery we want to replay up to the end of WAL. Particularly, in
6173 : * the case of a promoted standby minRecoveryPoint value in the
6174 : * control file is only updated after the first checkpoint. However,
6175 : * if the instance crashes before the first post-recovery checkpoint
6176 : * is completed then recovery will use a stale location causing the
6177 : * startup process to think that there are still invalid page
6178 : * references when checking for data consistency.
6179 : */
6180 231 : if (InArchiveRecovery)
6181 : {
6182 129 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6183 129 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6184 : }
6185 : else
6186 : {
6187 102 : LocalMinRecoveryPoint = InvalidXLogRecPtr;
6188 102 : LocalMinRecoveryPointTLI = 0;
6189 : }
6190 :
6191 : /* Check that the GUCs used to generate the WAL allow recovery */
6192 231 : CheckRequiredParameterValues();
6193 :
6194 : /*
6195 : * We're in recovery, so unlogged relations may be trashed and must be
6196 : * reset. This should be done BEFORE allowing Hot Standby
6197 : * connections, so that read-only backends don't try to read whatever
6198 : * garbage is left over from before.
6199 : */
6200 231 : ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
6201 :
6202 : /*
6203 : * Likewise, delete any saved transaction snapshot files that got left
6204 : * behind by crashed backends.
6205 : */
6206 231 : DeleteAllExportedSnapshotFiles();
6207 :
6208 : /*
6209 : * Initialize for Hot Standby, if enabled. We won't let backends in
6210 : * yet, not until we've reached the min recovery point specified in
6211 : * control file and we've established a recovery snapshot from a
6212 : * running-xacts WAL record.
6213 : */
6214 231 : if (ArchiveRecoveryRequested && EnableHotStandby)
6215 : {
6216 : TransactionId *xids;
6217 : int nxids;
6218 :
6219 121 : ereport(DEBUG1,
6220 : (errmsg_internal("initializing for hot standby")));
6221 :
6222 121 : InitRecoveryTransactionEnvironment();
6223 :
6224 121 : if (wasShutdown)
6225 26 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
6226 : else
6227 95 : oldestActiveXID = checkPoint.oldestActiveXid;
6228 : Assert(TransactionIdIsValid(oldestActiveXID));
6229 :
6230 : /* Tell procarray about the range of xids it has to deal with */
6231 121 : ProcArrayInitRecovery(XidFromFullTransactionId(TransamVariables->nextXid));
6232 :
6233 : /*
6234 : * Startup subtrans only. CLOG, MultiXact and commit timestamp
6235 : * have already been started up and other SLRUs are not maintained
6236 : * during recovery and need not be started yet.
6237 : */
6238 121 : StartupSUBTRANS(oldestActiveXID);
6239 :
6240 : /*
6241 : * If we're beginning at a shutdown checkpoint, we know that
6242 : * nothing was running on the primary at this point. So fake-up an
6243 : * empty running-xacts record and use that here and now. Recover
6244 : * additional standby state for prepared transactions.
6245 : */
6246 121 : if (wasShutdown)
6247 : {
6248 : RunningTransactionsData running;
6249 : TransactionId latestCompletedXid;
6250 :
6251 : /* Update pg_subtrans entries for any prepared transactions */
6252 26 : StandbyRecoverPreparedTransactions();
6253 :
6254 : /*
6255 : * Construct a RunningTransactions snapshot representing a
6256 : * shut down server, with only prepared transactions still
6257 : * alive. We're never overflowed at this point because all
6258 : * subxids are listed with their parent prepared transactions.
6259 : */
6260 26 : running.xcnt = nxids;
6261 26 : running.subxcnt = 0;
6262 26 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
6263 26 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
6264 26 : running.oldestRunningXid = oldestActiveXID;
6265 26 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
6266 26 : TransactionIdRetreat(latestCompletedXid);
6267 : Assert(TransactionIdIsNormal(latestCompletedXid));
6268 26 : running.latestCompletedXid = latestCompletedXid;
6269 26 : running.xids = xids;
6270 :
6271 26 : ProcArrayApplyRecoveryInfo(&running);
6272 : }
6273 : }
6274 :
6275 : /*
6276 : * We're all set for replaying the WAL now. Do it.
6277 : */
6278 231 : PerformWalRecovery();
6279 165 : performedWalRecovery = true;
6280 : }
6281 : else
6282 853 : performedWalRecovery = false;
6283 :
6284 : /*
6285 : * Finish WAL recovery.
6286 : */
6287 1018 : endOfRecoveryInfo = FinishWalRecovery();
6288 1018 : EndOfLog = endOfRecoveryInfo->endOfLog;
6289 1018 : EndOfLogTLI = endOfRecoveryInfo->endOfLogTLI;
6290 1018 : abortedRecPtr = endOfRecoveryInfo->abortedRecPtr;
6291 1018 : missingContrecPtr = endOfRecoveryInfo->missingContrecPtr;
6292 :
6293 : /*
6294 : * Reset ps status display, so as no information related to recovery shows
6295 : * up.
6296 : */
6297 1018 : set_ps_display("");
6298 :
6299 : /*
6300 : * When recovering from a backup (we are in recovery, and archive recovery
6301 : * was requested), complain if we did not roll forward far enough to reach
6302 : * the point where the database is consistent. For regular online
6303 : * backup-from-primary, that means reaching the end-of-backup WAL record
6304 : * (at which point we reset backupStartPoint to be Invalid), for
6305 : * backup-from-replica (which can't inject records into the WAL stream),
6306 : * that point is when we reach the minRecoveryPoint in pg_control (which
6307 : * we purposefully copy last when backing up from a replica). For
6308 : * pg_rewind (which creates a backup_label with a method of "pg_rewind")
6309 : * or snapshot-style backups (which don't), backupEndRequired will be set
6310 : * to false.
6311 : *
6312 : * Note: it is indeed okay to look at the local variable
6313 : * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
6314 : * might be further ahead --- ControlFile->minRecoveryPoint cannot have
6315 : * been advanced beyond the WAL we processed.
6316 : */
6317 1018 : if (InRecovery &&
6318 165 : (EndOfLog < LocalMinRecoveryPoint ||
6319 165 : XLogRecPtrIsValid(ControlFile->backupStartPoint)))
6320 : {
6321 : /*
6322 : * Ran off end of WAL before reaching end-of-backup WAL record, or
6323 : * minRecoveryPoint. That's a bad sign, indicating that you tried to
6324 : * recover from an online backup but never called pg_backup_stop(), or
6325 : * you didn't archive all the WAL needed.
6326 : */
6327 0 : if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
6328 : {
6329 0 : if (XLogRecPtrIsValid(ControlFile->backupStartPoint) || ControlFile->backupEndRequired)
6330 0 : ereport(FATAL,
6331 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6332 : errmsg("WAL ends before end of online backup"),
6333 : errhint("All WAL generated while online backup was taken must be available at recovery.")));
6334 : else
6335 0 : ereport(FATAL,
6336 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6337 : errmsg("WAL ends before consistent recovery point")));
6338 : }
6339 : }
6340 :
6341 : /*
6342 : * Reset unlogged relations to the contents of their INIT fork. This is
6343 : * done AFTER recovery is complete so as to include any unlogged relations
6344 : * created during recovery, but BEFORE recovery is marked as having
6345 : * completed successfully. Otherwise we'd not retry if any of the post
6346 : * end-of-recovery steps fail.
6347 : */
6348 1018 : if (InRecovery)
6349 165 : ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
6350 :
6351 : /*
6352 : * Pre-scan prepared transactions to find out the range of XIDs present.
6353 : * This information is not quite needed yet, but it is positioned here so
6354 : * as potential problems are detected before any on-disk change is done.
6355 : */
6356 1018 : oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
6357 :
6358 : /*
6359 : * Allow ordinary WAL segment creation before possibly switching to a new
6360 : * timeline, which creates a new segment, and after the last ReadRecord().
6361 : */
6362 1018 : SetInstallXLogFileSegmentActive();
6363 :
6364 : /*
6365 : * Consider whether we need to assign a new timeline ID.
6366 : *
6367 : * If we did archive recovery, we always assign a new ID. This handles a
6368 : * couple of issues. If we stopped short of the end of WAL during
6369 : * recovery, then we are clearly generating a new timeline and must assign
6370 : * it a unique new ID. Even if we ran to the end, modifying the current
6371 : * last segment is problematic because it may result in trying to
6372 : * overwrite an already-archived copy of that segment, and we encourage
6373 : * DBAs to make their archive_commands reject that. We can dodge the
6374 : * problem by making the new active segment have a new timeline ID.
6375 : *
6376 : * In a normal crash recovery, we can just extend the timeline we were in.
6377 : */
6378 1018 : newTLI = endOfRecoveryInfo->lastRecTLI;
6379 1018 : if (ArchiveRecoveryRequested)
6380 : {
6381 57 : newTLI = findNewestTimeLine(recoveryTargetTLI) + 1;
6382 57 : ereport(LOG,
6383 : (errmsg("selected new timeline ID: %u", newTLI)));
6384 :
6385 : /*
6386 : * Make a writable copy of the last WAL segment. (Note that we also
6387 : * have a copy of the last block of the old WAL in
6388 : * endOfRecovery->lastPage; we will use that below.)
6389 : */
6390 57 : XLogInitNewTimeline(EndOfLogTLI, EndOfLog, newTLI);
6391 :
6392 : /*
6393 : * Remove the signal files out of the way, so that we don't
6394 : * accidentally re-enter archive recovery mode in a subsequent crash.
6395 : */
6396 57 : if (endOfRecoveryInfo->standby_signal_file_found)
6397 54 : durable_unlink(STANDBY_SIGNAL_FILE, FATAL);
6398 :
6399 57 : if (endOfRecoveryInfo->recovery_signal_file_found)
6400 4 : durable_unlink(RECOVERY_SIGNAL_FILE, FATAL);
6401 :
6402 : /*
6403 : * Write the timeline history file, and have it archived. After this
6404 : * point (or rather, as soon as the file is archived), the timeline
6405 : * will appear as "taken" in the WAL archive and to any standby
6406 : * servers. If we crash before actually switching to the new
6407 : * timeline, standby servers will nevertheless think that we switched
6408 : * to the new timeline, and will try to connect to the new timeline.
6409 : * To minimize the window for that, try to do as little as possible
6410 : * between here and writing the end-of-recovery record.
6411 : */
6412 57 : writeTimeLineHistory(newTLI, recoveryTargetTLI,
6413 : EndOfLog, endOfRecoveryInfo->recoveryStopReason);
6414 :
6415 57 : ereport(LOG,
6416 : (errmsg("archive recovery complete")));
6417 : }
6418 :
6419 : /* Save the selected TimeLineID in shared memory, too */
6420 1018 : SpinLockAcquire(&XLogCtl->info_lck);
6421 1018 : XLogCtl->InsertTimeLineID = newTLI;
6422 1018 : XLogCtl->PrevTimeLineID = endOfRecoveryInfo->lastRecTLI;
6423 1018 : SpinLockRelease(&XLogCtl->info_lck);
6424 :
6425 : /*
6426 : * Actually, if WAL ended in an incomplete record, skip the parts that
6427 : * made it through and start writing after the portion that persisted.
6428 : * (It's critical to first write an OVERWRITE_CONTRECORD message, which
6429 : * we'll do as soon as we're open for writing new WAL.)
6430 : */
6431 1018 : if (XLogRecPtrIsValid(missingContrecPtr))
6432 : {
6433 : /*
6434 : * We should only have a missingContrecPtr if we're not switching to a
6435 : * new timeline. When a timeline switch occurs, WAL is copied from the
6436 : * old timeline to the new only up to the end of the last complete
6437 : * record, so there can't be an incomplete WAL record that we need to
6438 : * disregard.
6439 : */
6440 : Assert(newTLI == endOfRecoveryInfo->lastRecTLI);
6441 : Assert(XLogRecPtrIsValid(abortedRecPtr));
6442 11 : EndOfLog = missingContrecPtr;
6443 : }
6444 :
6445 : /*
6446 : * Prepare to write WAL starting at EndOfLog location, and init xlog
6447 : * buffer cache using the block containing the last record from the
6448 : * previous incarnation.
6449 : */
6450 1018 : Insert = &XLogCtl->Insert;
6451 1018 : Insert->PrevBytePos = XLogRecPtrToBytePos(endOfRecoveryInfo->lastRec);
6452 1018 : Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
6453 :
6454 : /*
6455 : * Tricky point here: lastPage contains the *last* block that the LastRec
6456 : * record spans, not the one it starts in. The last block is indeed the
6457 : * one we want to use.
6458 : */
6459 1018 : if (EndOfLog % XLOG_BLCKSZ != 0)
6460 : {
6461 : char *page;
6462 : int len;
6463 : int firstIdx;
6464 :
6465 985 : firstIdx = XLogRecPtrToBufIdx(EndOfLog);
6466 985 : len = EndOfLog - endOfRecoveryInfo->lastPageBeginPtr;
6467 : Assert(len < XLOG_BLCKSZ);
6468 :
6469 : /* Copy the valid part of the last block, and zero the rest */
6470 985 : page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
6471 985 : memcpy(page, endOfRecoveryInfo->lastPage, len);
6472 985 : memset(page + len, 0, XLOG_BLCKSZ - len);
6473 :
6474 985 : pg_atomic_write_u64(&XLogCtl->xlblocks[firstIdx], endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6475 985 : XLogCtl->InitializedUpTo = endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ;
6476 : }
6477 : else
6478 : {
6479 : /*
6480 : * There is no partial block to copy. Just set InitializedUpTo, and
6481 : * let the first attempt to insert a log record to initialize the next
6482 : * buffer.
6483 : */
6484 33 : XLogCtl->InitializedUpTo = EndOfLog;
6485 : }
6486 :
6487 : /*
6488 : * Update local and shared status. This is OK to do without any locks
6489 : * because no other process can be reading or writing WAL yet.
6490 : */
6491 1018 : LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
6492 1018 : pg_atomic_write_u64(&XLogCtl->logInsertResult, EndOfLog);
6493 1018 : pg_atomic_write_u64(&XLogCtl->logWriteResult, EndOfLog);
6494 1018 : pg_atomic_write_u64(&XLogCtl->logFlushResult, EndOfLog);
6495 1018 : XLogCtl->LogwrtRqst.Write = EndOfLog;
6496 1018 : XLogCtl->LogwrtRqst.Flush = EndOfLog;
6497 :
6498 : /*
6499 : * Preallocate additional log files, if wanted.
6500 : */
6501 1018 : PreallocXlogFiles(EndOfLog, newTLI);
6502 :
6503 : /*
6504 : * Okay, we're officially UP.
6505 : */
6506 1018 : InRecovery = false;
6507 :
6508 : /* start the archive_timeout timer and LSN running */
6509 1018 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
6510 1018 : XLogCtl->lastSegSwitchLSN = EndOfLog;
6511 :
6512 : /* also initialize latestCompletedXid, to nextXid - 1 */
6513 1018 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
6514 1018 : TransamVariables->latestCompletedXid = TransamVariables->nextXid;
6515 1018 : FullTransactionIdRetreat(&TransamVariables->latestCompletedXid);
6516 1018 : LWLockRelease(ProcArrayLock);
6517 :
6518 : /*
6519 : * Start up subtrans, if not already done for hot standby. (commit
6520 : * timestamps are started below, if necessary.)
6521 : */
6522 1018 : if (standbyState == STANDBY_DISABLED)
6523 961 : StartupSUBTRANS(oldestActiveXID);
6524 :
6525 : /*
6526 : * Perform end of recovery actions for any SLRUs that need it.
6527 : */
6528 1018 : TrimCLOG();
6529 1018 : TrimMultiXact();
6530 :
6531 : /*
6532 : * Reload shared-memory state for prepared transactions. This needs to
6533 : * happen before renaming the last partial segment of the old timeline as
6534 : * it may be possible that we have to recover some transactions from it.
6535 : */
6536 1018 : RecoverPreparedTransactions();
6537 :
6538 : /* Shut down xlogreader */
6539 1018 : ShutdownWalRecovery();
6540 :
6541 : /* Enable WAL writes for this backend only. */
6542 1018 : LocalSetXLogInsertAllowed();
6543 :
6544 : /* If necessary, write overwrite-contrecord before doing anything else */
6545 1018 : if (XLogRecPtrIsValid(abortedRecPtr))
6546 : {
6547 : Assert(XLogRecPtrIsValid(missingContrecPtr));
6548 11 : CreateOverwriteContrecordRecord(abortedRecPtr, missingContrecPtr, newTLI);
6549 : }
6550 :
6551 : /*
6552 : * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
6553 : * record before resource manager writes cleanup WAL records or checkpoint
6554 : * record is written.
6555 : */
6556 1018 : Insert->fullPageWrites = lastFullPageWrites;
6557 1018 : UpdateFullPageWrites();
6558 :
6559 : /*
6560 : * Emit checkpoint or end-of-recovery record in XLOG, if required.
6561 : */
6562 1018 : if (performedWalRecovery)
6563 165 : promoted = PerformRecoveryXLogAction();
6564 :
6565 : /*
6566 : * If any of the critical GUCs have changed, log them before we allow
6567 : * backends to write WAL.
6568 : */
6569 1018 : XLogReportParameters();
6570 :
6571 : /* If this is archive recovery, perform post-recovery cleanup actions. */
6572 1018 : if (ArchiveRecoveryRequested)
6573 57 : CleanupAfterArchiveRecovery(EndOfLogTLI, EndOfLog, newTLI);
6574 :
6575 : /*
6576 : * Local WAL inserts enabled, so it's time to finish initialization of
6577 : * commit timestamp.
6578 : */
6579 1018 : CompleteCommitTsInitialization();
6580 :
6581 : /*
6582 : * Update logical decoding status in shared memory and write an
6583 : * XLOG_LOGICAL_DECODING_STATUS_CHANGE, if necessary.
6584 : */
6585 1018 : UpdateLogicalDecodingStatusEndOfRecovery();
6586 :
6587 : /* Clean up EndOfWalRecoveryInfo data to appease Valgrind leak checking */
6588 1018 : if (endOfRecoveryInfo->lastPage)
6589 996 : pfree(endOfRecoveryInfo->lastPage);
6590 1018 : pfree(endOfRecoveryInfo->recoveryStopReason);
6591 1018 : pfree(endOfRecoveryInfo);
6592 :
6593 : /*
6594 : * If we reach this point with checksums in the state inprogress-on, it
6595 : * means that data checksums were in the process of being enabled when the
6596 : * cluster shut down. Since processing didn't finish, the operation will
6597 : * have to be restarted from scratch since there is no capability to
6598 : * continue where it was when the cluster shut down. Thus, revert the
6599 : * state back to off, and inform the user with a warning message. Being
6600 : * able to restart processing is a TODO, but it wouldn't be possible to
6601 : * restart here since we cannot launch a dynamic background worker
6602 : * directly from here (it has to be from a regular backend).
6603 : */
6604 1018 : if (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_INPROGRESS_ON)
6605 : {
6606 0 : XLogChecksums(PG_DATA_CHECKSUM_OFF);
6607 :
6608 0 : SpinLockAcquire(&XLogCtl->info_lck);
6609 0 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_OFF;
6610 0 : SetLocalDataChecksumState(XLogCtl->data_checksum_version);
6611 0 : SpinLockRelease(&XLogCtl->info_lck);
6612 :
6613 0 : ereport(WARNING,
6614 : errmsg("enabling data checksums was interrupted"),
6615 : errhint("Data checksum processing must be manually restarted for checksums to be enabled"));
6616 : }
6617 :
6618 : /*
6619 : * If data checksums were being disabled when the cluster was shut down,
6620 : * we know that we have a state where all backends have stopped validating
6621 : * checksums and we can move to off instead of prompting the user to
6622 : * perform any action.
6623 : */
6624 1018 : if (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_INPROGRESS_OFF)
6625 : {
6626 0 : XLogChecksums(PG_DATA_CHECKSUM_OFF);
6627 :
6628 0 : SpinLockAcquire(&XLogCtl->info_lck);
6629 0 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_OFF;
6630 0 : SetLocalDataChecksumState(XLogCtl->data_checksum_version);
6631 0 : SpinLockRelease(&XLogCtl->info_lck);
6632 : }
6633 :
6634 : /*
6635 : * All done with end-of-recovery actions.
6636 : *
6637 : * Now allow backends to write WAL and update the control file status in
6638 : * consequence. SharedRecoveryState, that controls if backends can write
6639 : * WAL, is updated while holding ControlFileLock to prevent other backends
6640 : * to look at an inconsistent state of the control file in shared memory.
6641 : * There is still a small window during which backends can write WAL and
6642 : * the control file is still referring to a system not in DB_IN_PRODUCTION
6643 : * state while looking at the on-disk control file.
6644 : *
6645 : * Also, we use info_lck to update SharedRecoveryState to ensure that
6646 : * there are no race conditions concerning visibility of other recent
6647 : * updates to shared memory.
6648 : */
6649 1018 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6650 1018 : ControlFile->state = DB_IN_PRODUCTION;
6651 :
6652 1018 : SpinLockAcquire(&XLogCtl->info_lck);
6653 1018 : ControlFile->data_checksum_version = XLogCtl->data_checksum_version;
6654 1018 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_DONE;
6655 1018 : SpinLockRelease(&XLogCtl->info_lck);
6656 :
6657 1018 : UpdateControlFile();
6658 1018 : LWLockRelease(ControlFileLock);
6659 :
6660 : /*
6661 : * Wake up the checkpointer process as there might be a request to disable
6662 : * logical decoding by concurrent slot drop.
6663 : */
6664 1018 : WakeupCheckpointer();
6665 :
6666 : /*
6667 : * Wake up all waiters. They need to report an error that recovery was
6668 : * ended before reaching the target LSN.
6669 : */
6670 1018 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_REPLAY, InvalidXLogRecPtr);
6671 1018 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_WRITE, InvalidXLogRecPtr);
6672 1018 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_FLUSH, InvalidXLogRecPtr);
6673 :
6674 : /*
6675 : * Shutdown the recovery environment. This must occur after
6676 : * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
6677 : * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
6678 : * any session building a snapshot will not rely on KnownAssignedXids as
6679 : * RecoveryInProgress() would return false at this stage. This is
6680 : * particularly critical for prepared 2PC transactions, that would still
6681 : * need to be included in snapshots once recovery has ended.
6682 : */
6683 1018 : if (standbyState != STANDBY_DISABLED)
6684 57 : ShutdownRecoveryTransactionEnvironment();
6685 :
6686 : /*
6687 : * If there were cascading standby servers connected to us, nudge any wal
6688 : * sender processes to notice that we've been promoted.
6689 : */
6690 1018 : WalSndWakeup(true, true);
6691 :
6692 : /*
6693 : * If this was a promotion, request an (online) checkpoint now. This isn't
6694 : * required for consistency, but the last restartpoint might be far back,
6695 : * and in case of a crash, recovering from it might take a longer than is
6696 : * appropriate now that we're not in standby mode anymore.
6697 : */
6698 1018 : if (promoted)
6699 50 : RequestCheckpoint(CHECKPOINT_FORCE);
6700 1018 : }
6701 :
6702 : /*
6703 : * Callback from PerformWalRecovery(), called when we switch from crash
6704 : * recovery to archive recovery mode. Updates the control file accordingly.
6705 : */
6706 : void
6707 1 : SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr, TimeLineID replayTLI)
6708 : {
6709 : /* initialize minRecoveryPoint to this record */
6710 1 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6711 1 : ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6712 1 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6713 : {
6714 1 : ControlFile->minRecoveryPoint = EndRecPtr;
6715 1 : ControlFile->minRecoveryPointTLI = replayTLI;
6716 : }
6717 : /* update local copy */
6718 1 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6719 1 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6720 :
6721 : /*
6722 : * The startup process can update its local copy of minRecoveryPoint from
6723 : * this point.
6724 : */
6725 1 : updateMinRecoveryPoint = true;
6726 :
6727 1 : UpdateControlFile();
6728 :
6729 : /*
6730 : * We update SharedRecoveryState while holding the lock on ControlFileLock
6731 : * so both states are consistent in shared memory.
6732 : */
6733 1 : SpinLockAcquire(&XLogCtl->info_lck);
6734 1 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6735 1 : SpinLockRelease(&XLogCtl->info_lck);
6736 :
6737 1 : LWLockRelease(ControlFileLock);
6738 1 : }
6739 :
6740 : /*
6741 : * Callback from PerformWalRecovery(), called when we reach the end of backup.
6742 : * Updates the control file accordingly.
6743 : */
6744 : void
6745 86 : ReachedEndOfBackup(XLogRecPtr EndRecPtr, TimeLineID tli)
6746 : {
6747 : /*
6748 : * We have reached the end of base backup, as indicated by pg_control. The
6749 : * data on disk is now consistent (unless minRecoveryPoint is further
6750 : * ahead, which can happen if we crashed during previous recovery). Reset
6751 : * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
6752 : * make sure we don't allow starting up at an earlier point even if
6753 : * recovery is stopped and restarted soon after this.
6754 : */
6755 86 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6756 :
6757 86 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6758 : {
6759 80 : ControlFile->minRecoveryPoint = EndRecPtr;
6760 80 : ControlFile->minRecoveryPointTLI = tli;
6761 : }
6762 :
6763 86 : ControlFile->backupStartPoint = InvalidXLogRecPtr;
6764 86 : ControlFile->backupEndPoint = InvalidXLogRecPtr;
6765 86 : ControlFile->backupEndRequired = false;
6766 86 : UpdateControlFile();
6767 :
6768 86 : LWLockRelease(ControlFileLock);
6769 86 : }
6770 :
6771 : /*
6772 : * Perform whatever XLOG actions are necessary at end of REDO.
6773 : *
6774 : * The goal here is to make sure that we'll be able to recover properly if
6775 : * we crash again. If we choose to write a checkpoint, we'll write a shutdown
6776 : * checkpoint rather than an on-line one. This is not particularly critical,
6777 : * but since we may be assigning a new TLI, using a shutdown checkpoint allows
6778 : * us to have the rule that TLI only changes in shutdown checkpoints, which
6779 : * allows some extra error checking in xlog_redo.
6780 : */
6781 : static bool
6782 165 : PerformRecoveryXLogAction(void)
6783 : {
6784 165 : bool promoted = false;
6785 :
6786 : /*
6787 : * Perform a checkpoint to update all our recovery activity to disk.
6788 : *
6789 : * Note that we write a shutdown checkpoint rather than an on-line one.
6790 : * This is not particularly critical, but since we may be assigning a new
6791 : * TLI, using a shutdown checkpoint allows us to have the rule that TLI
6792 : * only changes in shutdown checkpoints, which allows some extra error
6793 : * checking in xlog_redo.
6794 : *
6795 : * In promotion, only create a lightweight end-of-recovery record instead
6796 : * of a full checkpoint. A checkpoint is requested later, after we're
6797 : * fully out of recovery mode and already accepting queries.
6798 : */
6799 222 : if (ArchiveRecoveryRequested && IsUnderPostmaster &&
6800 57 : PromoteIsTriggered())
6801 : {
6802 50 : promoted = true;
6803 :
6804 : /*
6805 : * Insert a special WAL record to mark the end of recovery, since we
6806 : * aren't doing a checkpoint. That means that the checkpointer process
6807 : * may likely be in the middle of a time-smoothed restartpoint and
6808 : * could continue to be for minutes after this. That sounds strange,
6809 : * but the effect is roughly the same and it would be stranger to try
6810 : * to come out of the restartpoint and then checkpoint. We request a
6811 : * checkpoint later anyway, just for safety.
6812 : */
6813 50 : CreateEndOfRecoveryRecord();
6814 : }
6815 : else
6816 : {
6817 115 : RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
6818 : CHECKPOINT_FAST |
6819 : CHECKPOINT_WAIT);
6820 : }
6821 :
6822 165 : return promoted;
6823 : }
6824 :
6825 : /*
6826 : * Is the system still in recovery?
6827 : *
6828 : * Unlike testing InRecovery, this works in any process that's connected to
6829 : * shared memory.
6830 : */
6831 : bool
6832 97232459 : RecoveryInProgress(void)
6833 : {
6834 : /*
6835 : * We check shared state each time only until we leave recovery mode. We
6836 : * can't re-enter recovery, so there's no need to keep checking after the
6837 : * shared variable has once been seen false.
6838 : */
6839 97232459 : if (!LocalRecoveryInProgress)
6840 95391510 : return false;
6841 : else
6842 : {
6843 : /*
6844 : * use volatile pointer to make sure we make a fresh read of the
6845 : * shared variable.
6846 : */
6847 1840949 : volatile XLogCtlData *xlogctl = XLogCtl;
6848 :
6849 1840949 : LocalRecoveryInProgress = (xlogctl->SharedRecoveryState != RECOVERY_STATE_DONE);
6850 :
6851 : /*
6852 : * Note: We don't need a memory barrier when we're still in recovery.
6853 : * We might exit recovery immediately after return, so the caller
6854 : * can't rely on 'true' meaning that we're still in recovery anyway.
6855 : */
6856 :
6857 1840949 : return LocalRecoveryInProgress;
6858 : }
6859 : }
6860 :
6861 : /*
6862 : * Returns current recovery state from shared memory.
6863 : *
6864 : * This returned state is kept consistent with the contents of the control
6865 : * file. See details about the possible values of RecoveryState in xlog.h.
6866 : */
6867 : RecoveryState
6868 33573 : GetRecoveryState(void)
6869 : {
6870 : RecoveryState retval;
6871 :
6872 33573 : SpinLockAcquire(&XLogCtl->info_lck);
6873 33573 : retval = XLogCtl->SharedRecoveryState;
6874 33573 : SpinLockRelease(&XLogCtl->info_lck);
6875 :
6876 33573 : return retval;
6877 : }
6878 :
6879 : /*
6880 : * Is this process allowed to insert new WAL records?
6881 : *
6882 : * Ordinarily this is essentially equivalent to !RecoveryInProgress().
6883 : * But we also have provisions for forcing the result "true" or "false"
6884 : * within specific processes regardless of the global state.
6885 : */
6886 : bool
6887 67340308 : XLogInsertAllowed(void)
6888 : {
6889 : /*
6890 : * If value is "unconditionally true" or "unconditionally false", just
6891 : * return it. This provides the normal fast path once recovery is known
6892 : * done.
6893 : */
6894 67340308 : if (LocalXLogInsertAllowed >= 0)
6895 66667899 : return (bool) LocalXLogInsertAllowed;
6896 :
6897 : /*
6898 : * Else, must check to see if we're still in recovery.
6899 : */
6900 672409 : if (RecoveryInProgress())
6901 661828 : return false;
6902 :
6903 : /*
6904 : * On exit from recovery, reset to "unconditionally true", since there is
6905 : * no need to keep checking.
6906 : */
6907 10581 : LocalXLogInsertAllowed = 1;
6908 10581 : return true;
6909 : }
6910 :
6911 : /*
6912 : * Make XLogInsertAllowed() return true in the current process only.
6913 : *
6914 : * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
6915 : * and even call LocalSetXLogInsertAllowed() again after that.
6916 : *
6917 : * Returns the previous value of LocalXLogInsertAllowed.
6918 : */
6919 : static int
6920 1048 : LocalSetXLogInsertAllowed(void)
6921 : {
6922 1048 : int oldXLogAllowed = LocalXLogInsertAllowed;
6923 :
6924 1048 : LocalXLogInsertAllowed = 1;
6925 :
6926 1048 : return oldXLogAllowed;
6927 : }
6928 :
6929 : /*
6930 : * Return the current Redo pointer from shared memory.
6931 : *
6932 : * As a side-effect, the local RedoRecPtr copy is updated.
6933 : */
6934 : XLogRecPtr
6935 366064 : GetRedoRecPtr(void)
6936 : {
6937 : XLogRecPtr ptr;
6938 :
6939 : /*
6940 : * The possibly not up-to-date copy in XLogCtl is enough. Even if we
6941 : * grabbed a WAL insertion lock to read the authoritative value in
6942 : * Insert->RedoRecPtr, someone might update it just after we've released
6943 : * the lock.
6944 : */
6945 366064 : SpinLockAcquire(&XLogCtl->info_lck);
6946 366064 : ptr = XLogCtl->RedoRecPtr;
6947 366064 : SpinLockRelease(&XLogCtl->info_lck);
6948 :
6949 366064 : if (RedoRecPtr < ptr)
6950 1775 : RedoRecPtr = ptr;
6951 :
6952 366064 : return RedoRecPtr;
6953 : }
6954 :
6955 : /*
6956 : * Return information needed to decide whether a modified block needs a
6957 : * full-page image to be included in the WAL record.
6958 : *
6959 : * The returned values are cached copies from backend-private memory, and
6960 : * possibly out-of-date or, indeed, uninitialized, in which case they will
6961 : * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6962 : * re-check them against up-to-date values, while holding the WAL insert lock.
6963 : */
6964 : void
6965 24948637 : GetFullPageWriteInfo(XLogRecPtr *RedoRecPtr_p, bool *doPageWrites_p)
6966 : {
6967 24948637 : *RedoRecPtr_p = RedoRecPtr;
6968 24948637 : *doPageWrites_p = doPageWrites;
6969 24948637 : }
6970 :
6971 : /*
6972 : * GetInsertRecPtr -- Returns the current insert position.
6973 : *
6974 : * NOTE: The value *actually* returned is the position of the last full
6975 : * xlog page. It lags behind the real insert position by at most 1 page.
6976 : * For that, we don't need to scan through WAL insertion locks, and an
6977 : * approximation is enough for the current usage of this function.
6978 : */
6979 : XLogRecPtr
6980 7353 : GetInsertRecPtr(void)
6981 : {
6982 : XLogRecPtr recptr;
6983 :
6984 7353 : SpinLockAcquire(&XLogCtl->info_lck);
6985 7353 : recptr = XLogCtl->LogwrtRqst.Write;
6986 7353 : SpinLockRelease(&XLogCtl->info_lck);
6987 :
6988 7353 : return recptr;
6989 : }
6990 :
6991 : /*
6992 : * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6993 : * position known to be fsync'd to disk. This should only be used on a
6994 : * system that is known not to be in recovery.
6995 : */
6996 : XLogRecPtr
6997 261503 : GetFlushRecPtr(TimeLineID *insertTLI)
6998 : {
6999 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
7000 :
7001 261503 : RefreshXLogWriteResult(LogwrtResult);
7002 :
7003 : /*
7004 : * If we're writing and flushing WAL, the time line can't be changing, so
7005 : * no lock is required.
7006 : */
7007 261503 : if (insertTLI)
7008 59015 : *insertTLI = XLogCtl->InsertTimeLineID;
7009 :
7010 261503 : return LogwrtResult.Flush;
7011 : }
7012 :
7013 : /*
7014 : * GetWALInsertionTimeLine -- Returns the current timeline of a system that
7015 : * is not in recovery.
7016 : */
7017 : TimeLineID
7018 121978 : GetWALInsertionTimeLine(void)
7019 : {
7020 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
7021 :
7022 : /* Since the value can't be changing, no lock is required. */
7023 121978 : return XLogCtl->InsertTimeLineID;
7024 : }
7025 :
7026 : /*
7027 : * GetWALInsertionTimeLineIfSet -- If the system is not in recovery, returns
7028 : * the WAL insertion timeline; else, returns 0. Wherever possible, use
7029 : * GetWALInsertionTimeLine() instead, since it's cheaper. Note that this
7030 : * function decides recovery has ended as soon as the insert TLI is set, which
7031 : * happens before we set XLogCtl->SharedRecoveryState to RECOVERY_STATE_DONE.
7032 : */
7033 : TimeLineID
7034 1 : GetWALInsertionTimeLineIfSet(void)
7035 : {
7036 : TimeLineID insertTLI;
7037 :
7038 1 : SpinLockAcquire(&XLogCtl->info_lck);
7039 1 : insertTLI = XLogCtl->InsertTimeLineID;
7040 1 : SpinLockRelease(&XLogCtl->info_lck);
7041 :
7042 1 : return insertTLI;
7043 : }
7044 :
7045 : /*
7046 : * GetLastImportantRecPtr -- Returns the LSN of the last important record
7047 : * inserted. All records not explicitly marked as unimportant are considered
7048 : * important.
7049 : *
7050 : * The LSN is determined by computing the maximum of
7051 : * WALInsertLocks[i].lastImportantAt.
7052 : */
7053 : XLogRecPtr
7054 1785 : GetLastImportantRecPtr(void)
7055 : {
7056 1785 : XLogRecPtr res = InvalidXLogRecPtr;
7057 : int i;
7058 :
7059 16065 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
7060 : {
7061 : XLogRecPtr last_important;
7062 :
7063 : /*
7064 : * Need to take a lock to prevent torn reads of the LSN, which are
7065 : * possible on some of the supported platforms. WAL insert locks only
7066 : * support exclusive mode, so we have to use that.
7067 : */
7068 14280 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
7069 14280 : last_important = WALInsertLocks[i].l.lastImportantAt;
7070 14280 : LWLockRelease(&WALInsertLocks[i].l.lock);
7071 :
7072 14280 : if (res < last_important)
7073 3097 : res = last_important;
7074 : }
7075 :
7076 1785 : return res;
7077 : }
7078 :
7079 : /*
7080 : * Get the time and LSN of the last xlog segment switch
7081 : */
7082 : pg_time_t
7083 0 : GetLastSegSwitchData(XLogRecPtr *lastSwitchLSN)
7084 : {
7085 : pg_time_t result;
7086 :
7087 : /* Need WALWriteLock, but shared lock is sufficient */
7088 0 : LWLockAcquire(WALWriteLock, LW_SHARED);
7089 0 : result = XLogCtl->lastSegSwitchTime;
7090 0 : *lastSwitchLSN = XLogCtl->lastSegSwitchLSN;
7091 0 : LWLockRelease(WALWriteLock);
7092 :
7093 0 : return result;
7094 : }
7095 :
7096 : /*
7097 : * This must be called ONCE during postmaster or standalone-backend shutdown
7098 : */
7099 : void
7100 773 : ShutdownXLOG(int code, Datum arg)
7101 : {
7102 : /*
7103 : * We should have an aux process resource owner to use, and we should not
7104 : * be in a transaction that's installed some other resowner.
7105 : */
7106 : Assert(AuxProcessResourceOwner != NULL);
7107 : Assert(CurrentResourceOwner == NULL ||
7108 : CurrentResourceOwner == AuxProcessResourceOwner);
7109 773 : CurrentResourceOwner = AuxProcessResourceOwner;
7110 :
7111 : /* Don't be chatty in standalone mode */
7112 773 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
7113 : (errmsg("shutting down")));
7114 :
7115 : /*
7116 : * Signal walsenders to move to stopping state.
7117 : */
7118 773 : WalSndInitStopping();
7119 :
7120 : /*
7121 : * Wait for WAL senders to be in stopping state. This prevents commands
7122 : * from writing new WAL.
7123 : */
7124 773 : WalSndWaitStopping();
7125 :
7126 773 : if (RecoveryInProgress())
7127 63 : CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
7128 : else
7129 : {
7130 : /*
7131 : * If archiving is enabled, rotate the last XLOG file so that all the
7132 : * remaining records are archived (postmaster wakes up the archiver
7133 : * process one more time at the end of shutdown). The checkpoint
7134 : * record will go to the next XLOG file and won't be archived (yet).
7135 : */
7136 710 : if (XLogArchivingActive())
7137 17 : RequestXLogSwitch(false);
7138 :
7139 710 : CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
7140 : }
7141 773 : }
7142 :
7143 : /*
7144 : * Format checkpoint request flags as a space-separated string for
7145 : * log messages.
7146 : */
7147 : static const char *
7148 3234 : CheckpointFlagsString(int flags)
7149 : {
7150 : static char buf[128];
7151 :
7152 25872 : snprintf(buf, sizeof(buf), "%s%s%s%s%s%s%s%s",
7153 3234 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
7154 3234 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
7155 3234 : (flags & CHECKPOINT_FAST) ? " fast" : "",
7156 3234 : (flags & CHECKPOINT_FORCE) ? " force" : "",
7157 3234 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
7158 3234 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
7159 3234 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
7160 3234 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "");
7161 :
7162 3234 : return buf;
7163 : }
7164 :
7165 : /*
7166 : * Log start of a checkpoint.
7167 : */
7168 : static void
7169 1617 : LogCheckpointStart(int flags, bool restartpoint)
7170 : {
7171 1617 : if (restartpoint)
7172 215 : ereport(LOG,
7173 : /* translator: the placeholder shows checkpoint options */
7174 : (errmsg("restartpoint starting:%s",
7175 : CheckpointFlagsString(flags))));
7176 : else
7177 1402 : ereport(LOG,
7178 : /* translator: the placeholder shows checkpoint options */
7179 : (errmsg("checkpoint starting:%s",
7180 : CheckpointFlagsString(flags))));
7181 1617 : }
7182 :
7183 : /*
7184 : * Log end of a checkpoint.
7185 : */
7186 : static void
7187 1951 : LogCheckpointEnd(bool restartpoint, int flags)
7188 : {
7189 : long write_msecs,
7190 : sync_msecs,
7191 : total_msecs,
7192 : longest_msecs,
7193 : average_msecs;
7194 : uint64 average_sync_time;
7195 :
7196 1951 : CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
7197 :
7198 1951 : write_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_write_t,
7199 : CheckpointStats.ckpt_sync_t);
7200 :
7201 1951 : sync_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_sync_t,
7202 : CheckpointStats.ckpt_sync_end_t);
7203 :
7204 : /* Accumulate checkpoint timing summary data, in milliseconds. */
7205 1951 : PendingCheckpointerStats.write_time += write_msecs;
7206 1951 : PendingCheckpointerStats.sync_time += sync_msecs;
7207 :
7208 : /*
7209 : * All of the published timing statistics are accounted for. Only
7210 : * continue if a log message is to be written.
7211 : */
7212 1951 : if (!log_checkpoints)
7213 334 : return;
7214 :
7215 1617 : total_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_start_t,
7216 : CheckpointStats.ckpt_end_t);
7217 :
7218 : /*
7219 : * Timing values returned from CheckpointStats are in microseconds.
7220 : * Convert to milliseconds for consistent printing.
7221 : */
7222 1617 : longest_msecs = (long) ((CheckpointStats.ckpt_longest_sync + 999) / 1000);
7223 :
7224 1617 : average_sync_time = 0;
7225 1617 : if (CheckpointStats.ckpt_sync_rels > 0)
7226 0 : average_sync_time = CheckpointStats.ckpt_agg_sync_time /
7227 0 : CheckpointStats.ckpt_sync_rels;
7228 1617 : average_msecs = (long) ((average_sync_time + 999) / 1000);
7229 :
7230 : /*
7231 : * ControlFileLock is not required to see ControlFile->checkPoint and
7232 : * ->checkPointCopy here as we are the only updator of those variables at
7233 : * this moment.
7234 : */
7235 1617 : if (restartpoint)
7236 215 : ereport(LOG,
7237 : (errmsg("restartpoint complete:%s: wrote %d buffers (%.1f%%), "
7238 : "wrote %d SLRU buffers; %d WAL file(s) added, "
7239 : "%d removed, %d recycled; write=%ld.%03d s, "
7240 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
7241 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
7242 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
7243 : CheckpointFlagsString(flags),
7244 : CheckpointStats.ckpt_bufs_written,
7245 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
7246 : CheckpointStats.ckpt_slru_written,
7247 : CheckpointStats.ckpt_segs_added,
7248 : CheckpointStats.ckpt_segs_removed,
7249 : CheckpointStats.ckpt_segs_recycled,
7250 : write_msecs / 1000, (int) (write_msecs % 1000),
7251 : sync_msecs / 1000, (int) (sync_msecs % 1000),
7252 : total_msecs / 1000, (int) (total_msecs % 1000),
7253 : CheckpointStats.ckpt_sync_rels,
7254 : longest_msecs / 1000, (int) (longest_msecs % 1000),
7255 : average_msecs / 1000, (int) (average_msecs % 1000),
7256 : (int) (PrevCheckPointDistance / 1024.0),
7257 : (int) (CheckPointDistanceEstimate / 1024.0),
7258 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
7259 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
7260 : else
7261 1402 : ereport(LOG,
7262 : (errmsg("checkpoint complete:%s: wrote %d buffers (%.1f%%), "
7263 : "wrote %d SLRU buffers; %d WAL file(s) added, "
7264 : "%d removed, %d recycled; write=%ld.%03d s, "
7265 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
7266 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
7267 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
7268 : CheckpointFlagsString(flags),
7269 : CheckpointStats.ckpt_bufs_written,
7270 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
7271 : CheckpointStats.ckpt_slru_written,
7272 : CheckpointStats.ckpt_segs_added,
7273 : CheckpointStats.ckpt_segs_removed,
7274 : CheckpointStats.ckpt_segs_recycled,
7275 : write_msecs / 1000, (int) (write_msecs % 1000),
7276 : sync_msecs / 1000, (int) (sync_msecs % 1000),
7277 : total_msecs / 1000, (int) (total_msecs % 1000),
7278 : CheckpointStats.ckpt_sync_rels,
7279 : longest_msecs / 1000, (int) (longest_msecs % 1000),
7280 : average_msecs / 1000, (int) (average_msecs % 1000),
7281 : (int) (PrevCheckPointDistance / 1024.0),
7282 : (int) (CheckPointDistanceEstimate / 1024.0),
7283 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
7284 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
7285 : }
7286 :
7287 : /*
7288 : * Update the estimate of distance between checkpoints.
7289 : *
7290 : * The estimate is used to calculate the number of WAL segments to keep
7291 : * preallocated, see XLOGfileslop().
7292 : */
7293 : static void
7294 1951 : UpdateCheckPointDistanceEstimate(uint64 nbytes)
7295 : {
7296 : /*
7297 : * To estimate the number of segments consumed between checkpoints, keep a
7298 : * moving average of the amount of WAL generated in previous checkpoint
7299 : * cycles. However, if the load is bursty, with quiet periods and busy
7300 : * periods, we want to cater for the peak load. So instead of a plain
7301 : * moving average, let the average decline slowly if the previous cycle
7302 : * used less WAL than estimated, but bump it up immediately if it used
7303 : * more.
7304 : *
7305 : * When checkpoints are triggered by max_wal_size, this should converge to
7306 : * CheckpointSegments * wal_segment_size,
7307 : *
7308 : * Note: This doesn't pay any attention to what caused the checkpoint.
7309 : * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
7310 : * starting a base backup, are counted the same as those created
7311 : * automatically. The slow-decline will largely mask them out, if they are
7312 : * not frequent. If they are frequent, it seems reasonable to count them
7313 : * in as any others; if you issue a manual checkpoint every 5 minutes and
7314 : * never let a timed checkpoint happen, it makes sense to base the
7315 : * preallocation on that 5 minute interval rather than whatever
7316 : * checkpoint_timeout is set to.
7317 : */
7318 1951 : PrevCheckPointDistance = nbytes;
7319 1951 : if (CheckPointDistanceEstimate < nbytes)
7320 880 : CheckPointDistanceEstimate = nbytes;
7321 : else
7322 1071 : CheckPointDistanceEstimate =
7323 1071 : (0.90 * CheckPointDistanceEstimate + 0.10 * (double) nbytes);
7324 1951 : }
7325 :
7326 : /*
7327 : * Update the ps display for a process running a checkpoint. Note that
7328 : * this routine should not do any allocations so as it can be called
7329 : * from a critical section.
7330 : */
7331 : static void
7332 3902 : update_checkpoint_display(int flags, bool restartpoint, bool reset)
7333 : {
7334 : /*
7335 : * The status is reported only for end-of-recovery and shutdown
7336 : * checkpoints or shutdown restartpoints. Updating the ps display is
7337 : * useful in those situations as it may not be possible to rely on
7338 : * pg_stat_activity to see the status of the checkpointer or the startup
7339 : * process.
7340 : */
7341 3902 : if ((flags & (CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IS_SHUTDOWN)) == 0)
7342 2374 : return;
7343 :
7344 1528 : if (reset)
7345 764 : set_ps_display("");
7346 : else
7347 : {
7348 : char activitymsg[128];
7349 :
7350 2292 : snprintf(activitymsg, sizeof(activitymsg), "performing %s%s%s",
7351 764 : (flags & CHECKPOINT_END_OF_RECOVERY) ? "end-of-recovery " : "",
7352 764 : (flags & CHECKPOINT_IS_SHUTDOWN) ? "shutdown " : "",
7353 : restartpoint ? "restartpoint" : "checkpoint");
7354 764 : set_ps_display(activitymsg);
7355 : }
7356 : }
7357 :
7358 :
7359 : /*
7360 : * Perform a checkpoint --- either during shutdown, or on-the-fly
7361 : *
7362 : * flags is a bitwise OR of the following:
7363 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
7364 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
7365 : * CHECKPOINT_FAST: finish the checkpoint ASAP, ignoring
7366 : * checkpoint_completion_target parameter.
7367 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
7368 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
7369 : * CHECKPOINT_END_OF_RECOVERY).
7370 : * CHECKPOINT_FLUSH_UNLOGGED: also flush buffers of unlogged tables.
7371 : *
7372 : * Note: flags contains other bits, of interest here only for logging purposes.
7373 : * In particular note that this routine is synchronous and does not pay
7374 : * attention to CHECKPOINT_WAIT.
7375 : *
7376 : * If !shutdown then we are writing an online checkpoint. An XLOG_CHECKPOINT_REDO
7377 : * record is inserted into WAL at the logical location of the checkpoint, before
7378 : * flushing anything to disk, and when the checkpoint is eventually completed,
7379 : * and it is from this point that WAL replay will begin in the case of a recovery
7380 : * from this checkpoint. Once everything is written to disk, an
7381 : * XLOG_CHECKPOINT_ONLINE record is written to complete the checkpoint, and
7382 : * points back to the earlier XLOG_CHECKPOINT_REDO record. This mechanism allows
7383 : * other write-ahead log records to be written while the checkpoint is in
7384 : * progress, but we must be very careful about order of operations. This function
7385 : * may take many minutes to execute on a busy system.
7386 : *
7387 : * On the other hand, when shutdown is true, concurrent insertion into the
7388 : * write-ahead log is impossible, so there is no need for two separate records.
7389 : * In this case, we only insert an XLOG_CHECKPOINT_SHUTDOWN record, and it's
7390 : * both the record marking the completion of the checkpoint and the location
7391 : * from which WAL replay would begin if needed.
7392 : *
7393 : * Returns true if a new checkpoint was performed, or false if it was skipped
7394 : * because the system was idle.
7395 : */
7396 : bool
7397 1736 : CreateCheckPoint(int flags)
7398 : {
7399 : bool shutdown;
7400 : CheckPoint checkPoint;
7401 : XLogRecPtr recptr;
7402 : XLogSegNo _logSegNo;
7403 1736 : XLogCtlInsert *Insert = &XLogCtl->Insert;
7404 : uint32 freespace;
7405 : XLogRecPtr PriorRedoPtr;
7406 : XLogRecPtr last_important_lsn;
7407 : VirtualTransactionId *vxids;
7408 : int nvxids;
7409 1736 : int oldXLogAllowed = 0;
7410 :
7411 : /*
7412 : * An end-of-recovery checkpoint is really a shutdown checkpoint, just
7413 : * issued at a different time.
7414 : */
7415 1736 : if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
7416 740 : shutdown = true;
7417 : else
7418 996 : shutdown = false;
7419 :
7420 : /* sanity check */
7421 1736 : if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
7422 0 : elog(ERROR, "can't create a checkpoint during recovery");
7423 :
7424 : /*
7425 : * Prepare to accumulate statistics.
7426 : *
7427 : * Note: because it is possible for log_checkpoints to change while a
7428 : * checkpoint proceeds, we always accumulate stats, even if
7429 : * log_checkpoints is currently off.
7430 : */
7431 19096 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7432 1736 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7433 :
7434 : /*
7435 : * Let smgr prepare for checkpoint; this has to happen outside the
7436 : * critical section and before we determine the REDO pointer. Note that
7437 : * smgr must not do anything that'd have to be undone if we decide no
7438 : * checkpoint is needed.
7439 : */
7440 1736 : SyncPreCheckpoint();
7441 :
7442 : /* Run these points outside the critical section. */
7443 1736 : INJECTION_POINT("create-checkpoint-initial", NULL);
7444 1736 : INJECTION_POINT_LOAD("create-checkpoint-run");
7445 :
7446 : /*
7447 : * Use a critical section to force system panic if we have trouble.
7448 : */
7449 1736 : START_CRIT_SECTION();
7450 :
7451 1736 : if (shutdown)
7452 : {
7453 740 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7454 740 : ControlFile->state = DB_SHUTDOWNING;
7455 740 : UpdateControlFile();
7456 740 : LWLockRelease(ControlFileLock);
7457 : }
7458 :
7459 : /* Begin filling in the checkpoint WAL record */
7460 22568 : MemSet(&checkPoint, 0, sizeof(checkPoint));
7461 1736 : checkPoint.time = (pg_time_t) time(NULL);
7462 :
7463 : /*
7464 : * For Hot Standby, derive the oldestActiveXid before we fix the redo
7465 : * pointer. This allows us to begin accumulating changes to assemble our
7466 : * starting snapshot of locks and transactions.
7467 : */
7468 1736 : if (!shutdown && XLogStandbyInfoActive())
7469 954 : checkPoint.oldestActiveXid = GetOldestActiveTransactionId(false, true);
7470 : else
7471 782 : checkPoint.oldestActiveXid = InvalidTransactionId;
7472 :
7473 : /*
7474 : * Get location of last important record before acquiring insert locks (as
7475 : * GetLastImportantRecPtr() also locks WAL locks).
7476 : */
7477 1736 : last_important_lsn = GetLastImportantRecPtr();
7478 :
7479 : /*
7480 : * If this isn't a shutdown or forced checkpoint, and if there has been no
7481 : * WAL activity requiring a checkpoint, skip it. The idea here is to
7482 : * avoid inserting duplicate checkpoints when the system is idle.
7483 : */
7484 1736 : if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
7485 : CHECKPOINT_FORCE)) == 0)
7486 : {
7487 206 : if (last_important_lsn == ControlFile->checkPoint)
7488 : {
7489 0 : END_CRIT_SECTION();
7490 0 : ereport(DEBUG1,
7491 : (errmsg_internal("checkpoint skipped because system is idle")));
7492 0 : return false;
7493 : }
7494 : }
7495 :
7496 : /*
7497 : * An end-of-recovery checkpoint is created before anyone is allowed to
7498 : * write WAL. To allow us to write the checkpoint record, temporarily
7499 : * enable XLogInsertAllowed.
7500 : */
7501 1736 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7502 30 : oldXLogAllowed = LocalSetXLogInsertAllowed();
7503 :
7504 1736 : checkPoint.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7505 1736 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7506 30 : checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7507 : else
7508 1706 : checkPoint.PrevTimeLineID = checkPoint.ThisTimeLineID;
7509 :
7510 : /*
7511 : * We must block concurrent insertions while examining insert state.
7512 : */
7513 1736 : WALInsertLockAcquireExclusive();
7514 :
7515 1736 : checkPoint.fullPageWrites = Insert->fullPageWrites;
7516 1736 : checkPoint.wal_level = wal_level;
7517 :
7518 : /*
7519 : * Get the current data_checksum_version value from xlogctl, valid at the
7520 : * time of the checkpoint.
7521 : */
7522 1736 : SpinLockAcquire(&XLogCtl->info_lck);
7523 1736 : checkPoint.dataChecksumState = XLogCtl->data_checksum_version;
7524 1736 : SpinLockRelease(&XLogCtl->info_lck);
7525 :
7526 1736 : if (shutdown)
7527 : {
7528 740 : XLogRecPtr curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
7529 :
7530 : /*
7531 : * Compute new REDO record ptr = location of next XLOG record.
7532 : *
7533 : * Since this is a shutdown checkpoint, there can't be any concurrent
7534 : * WAL insertion.
7535 : */
7536 740 : freespace = INSERT_FREESPACE(curInsert);
7537 740 : if (freespace == 0)
7538 : {
7539 0 : if (XLogSegmentOffset(curInsert, wal_segment_size) == 0)
7540 0 : curInsert += SizeOfXLogLongPHD;
7541 : else
7542 0 : curInsert += SizeOfXLogShortPHD;
7543 : }
7544 740 : checkPoint.redo = curInsert;
7545 :
7546 : /*
7547 : * Here we update the shared RedoRecPtr for future XLogInsert calls;
7548 : * this must be done while holding all the insertion locks.
7549 : *
7550 : * Note: if we fail to complete the checkpoint, RedoRecPtr will be
7551 : * left pointing past where it really needs to point. This is okay;
7552 : * the only consequence is that XLogInsert might back up whole buffers
7553 : * that it didn't really need to. We can't postpone advancing
7554 : * RedoRecPtr because XLogInserts that happen while we are dumping
7555 : * buffers must assume that their buffer changes are not included in
7556 : * the checkpoint.
7557 : */
7558 740 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
7559 : }
7560 :
7561 : /*
7562 : * Now we can release the WAL insertion locks, allowing other xacts to
7563 : * proceed while we are flushing disk buffers.
7564 : */
7565 1736 : WALInsertLockRelease();
7566 :
7567 : /*
7568 : * If this is an online checkpoint, we have not yet determined the redo
7569 : * point. We do so now by inserting the special XLOG_CHECKPOINT_REDO
7570 : * record; the LSN at which it starts becomes the new redo pointer. We
7571 : * don't do this for a shutdown checkpoint, because in that case no WAL
7572 : * can be written between the redo point and the insertion of the
7573 : * checkpoint record itself, so the checkpoint record itself serves to
7574 : * mark the redo point.
7575 : */
7576 1736 : if (!shutdown)
7577 : {
7578 : xl_checkpoint_redo redo_rec;
7579 :
7580 996 : WALInsertLockAcquire();
7581 996 : redo_rec.wal_level = wal_level;
7582 996 : SpinLockAcquire(&XLogCtl->info_lck);
7583 996 : redo_rec.data_checksum_version = XLogCtl->data_checksum_version;
7584 996 : SpinLockRelease(&XLogCtl->info_lck);
7585 996 : WALInsertLockRelease();
7586 :
7587 : /* Include WAL level in record for WAL summarizer's benefit. */
7588 996 : XLogBeginInsert();
7589 996 : XLogRegisterData(&redo_rec, sizeof(xl_checkpoint_redo));
7590 996 : (void) XLogInsert(RM_XLOG_ID, XLOG_CHECKPOINT_REDO);
7591 :
7592 : /*
7593 : * XLogInsertRecord will have updated XLogCtl->Insert.RedoRecPtr in
7594 : * shared memory and RedoRecPtr in backend-local memory, but we need
7595 : * to copy that into the record that will be inserted when the
7596 : * checkpoint is complete.
7597 : */
7598 996 : checkPoint.redo = RedoRecPtr;
7599 : }
7600 :
7601 : /* Update the info_lck-protected copy of RedoRecPtr as well */
7602 1736 : SpinLockAcquire(&XLogCtl->info_lck);
7603 1736 : XLogCtl->RedoRecPtr = checkPoint.redo;
7604 1736 : SpinLockRelease(&XLogCtl->info_lck);
7605 :
7606 : /*
7607 : * If enabled, log checkpoint start. We postpone this until now so as not
7608 : * to log anything if we decided to skip the checkpoint.
7609 : */
7610 1736 : if (log_checkpoints)
7611 1402 : LogCheckpointStart(flags, false);
7612 :
7613 1736 : INJECTION_POINT_CACHED("create-checkpoint-run", NULL);
7614 :
7615 : /* Update the process title */
7616 1736 : update_checkpoint_display(flags, false, false);
7617 :
7618 : TRACE_POSTGRESQL_CHECKPOINT_START(flags);
7619 :
7620 : /*
7621 : * Get the other info we need for the checkpoint record.
7622 : *
7623 : * We don't need to save oldestClogXid in the checkpoint, it only matters
7624 : * for the short period in which clog is being truncated, and if we crash
7625 : * during that we'll redo the clog truncation and fix up oldestClogXid
7626 : * there.
7627 : */
7628 1736 : LWLockAcquire(XidGenLock, LW_SHARED);
7629 1736 : checkPoint.nextXid = TransamVariables->nextXid;
7630 1736 : checkPoint.oldestXid = TransamVariables->oldestXid;
7631 1736 : checkPoint.oldestXidDB = TransamVariables->oldestXidDB;
7632 1736 : LWLockRelease(XidGenLock);
7633 :
7634 1736 : LWLockAcquire(CommitTsLock, LW_SHARED);
7635 1736 : checkPoint.oldestCommitTsXid = TransamVariables->oldestCommitTsXid;
7636 1736 : checkPoint.newestCommitTsXid = TransamVariables->newestCommitTsXid;
7637 1736 : LWLockRelease(CommitTsLock);
7638 :
7639 1736 : LWLockAcquire(OidGenLock, LW_SHARED);
7640 1736 : checkPoint.nextOid = TransamVariables->nextOid;
7641 1736 : if (!shutdown)
7642 996 : checkPoint.nextOid += TransamVariables->oidCount;
7643 1736 : LWLockRelease(OidGenLock);
7644 :
7645 1736 : checkPoint.logicalDecodingEnabled = IsLogicalDecodingEnabled();
7646 :
7647 1736 : MultiXactGetCheckptMulti(shutdown,
7648 : &checkPoint.nextMulti,
7649 : &checkPoint.nextMultiOffset,
7650 : &checkPoint.oldestMulti,
7651 : &checkPoint.oldestMultiDB);
7652 :
7653 : /*
7654 : * Having constructed the checkpoint record, ensure all shmem disk buffers
7655 : * and commit-log buffers are flushed to disk.
7656 : *
7657 : * This I/O could fail for various reasons. If so, we will fail to
7658 : * complete the checkpoint, but there is no reason to force a system
7659 : * panic. Accordingly, exit critical section while doing it.
7660 : */
7661 1736 : END_CRIT_SECTION();
7662 :
7663 : /*
7664 : * In some cases there are groups of actions that must all occur on one
7665 : * side or the other of a checkpoint record. Before flushing the
7666 : * checkpoint record we must explicitly wait for any backend currently
7667 : * performing those groups of actions.
7668 : *
7669 : * One example is end of transaction, so we must wait for any transactions
7670 : * that are currently in commit critical sections. If an xact inserted
7671 : * its commit record into XLOG just before the REDO point, then a crash
7672 : * restart from the REDO point would not replay that record, which means
7673 : * that our flushing had better include the xact's update of pg_xact. So
7674 : * we wait till he's out of his commit critical section before proceeding.
7675 : * See notes in RecordTransactionCommit().
7676 : *
7677 : * Because we've already released the insertion locks, this test is a bit
7678 : * fuzzy: it is possible that we will wait for xacts we didn't really need
7679 : * to wait for. But the delay should be short and it seems better to make
7680 : * checkpoint take a bit longer than to hold off insertions longer than
7681 : * necessary. (In fact, the whole reason we have this issue is that xact.c
7682 : * does commit record XLOG insertion and clog update as two separate steps
7683 : * protected by different locks, but again that seems best on grounds of
7684 : * minimizing lock contention.)
7685 : *
7686 : * A transaction that has not yet set delayChkptFlags when we look cannot
7687 : * be at risk, since it has not inserted its commit record yet; and one
7688 : * that's already cleared it is not at risk either, since it's done fixing
7689 : * clog and we will correctly flush the update below. So we cannot miss
7690 : * any xacts we need to wait for.
7691 : */
7692 1736 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
7693 1736 : if (nvxids > 0)
7694 : {
7695 : do
7696 : {
7697 : /*
7698 : * Keep absorbing fsync requests while we wait. There could even
7699 : * be a deadlock if we don't, if the process that prevents the
7700 : * checkpoint is trying to add a request to the queue.
7701 : */
7702 21 : AbsorbSyncRequests();
7703 :
7704 21 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
7705 21 : pg_usleep(10000L); /* wait for 10 msec */
7706 21 : pgstat_report_wait_end();
7707 21 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7708 : DELAY_CHKPT_START));
7709 : }
7710 1736 : pfree(vxids);
7711 :
7712 1736 : CheckPointGuts(checkPoint.redo, flags);
7713 :
7714 1736 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
7715 1736 : if (nvxids > 0)
7716 : {
7717 : do
7718 : {
7719 0 : AbsorbSyncRequests();
7720 :
7721 0 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_COMPLETE);
7722 0 : pg_usleep(10000L); /* wait for 10 msec */
7723 0 : pgstat_report_wait_end();
7724 0 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7725 : DELAY_CHKPT_COMPLETE));
7726 : }
7727 1736 : pfree(vxids);
7728 :
7729 : /*
7730 : * Take a snapshot of running transactions and write this to WAL. This
7731 : * allows us to reconstruct the state of running transactions during
7732 : * archive recovery, if required. Skip, if this info disabled.
7733 : *
7734 : * If we are shutting down, or Startup process is completing crash
7735 : * recovery we don't need to write running xact data.
7736 : */
7737 1736 : if (!shutdown && XLogStandbyInfoActive())
7738 954 : LogStandbySnapshot(InvalidOid);
7739 :
7740 1736 : START_CRIT_SECTION();
7741 :
7742 : /*
7743 : * Now insert the checkpoint record into XLOG.
7744 : */
7745 1736 : XLogBeginInsert();
7746 1736 : XLogRegisterData(&checkPoint, sizeof(checkPoint));
7747 1736 : recptr = XLogInsert(RM_XLOG_ID,
7748 : shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
7749 : XLOG_CHECKPOINT_ONLINE);
7750 :
7751 1736 : XLogFlush(recptr);
7752 :
7753 : /*
7754 : * We mustn't write any new WAL after a shutdown checkpoint, or it will be
7755 : * overwritten at next startup. No-one should even try, this just allows
7756 : * sanity-checking. In the case of an end-of-recovery checkpoint, we want
7757 : * to just temporarily disable writing until the system has exited
7758 : * recovery.
7759 : */
7760 1736 : if (shutdown)
7761 : {
7762 740 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7763 30 : LocalXLogInsertAllowed = oldXLogAllowed;
7764 : else
7765 710 : LocalXLogInsertAllowed = 0; /* never again write WAL */
7766 : }
7767 :
7768 : /*
7769 : * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
7770 : * = end of actual checkpoint record.
7771 : */
7772 1736 : if (shutdown && checkPoint.redo != ProcLastRecPtr)
7773 0 : ereport(PANIC,
7774 : (errmsg("concurrent write-ahead log activity while database system is shutting down")));
7775 :
7776 : /*
7777 : * Remember the prior checkpoint's redo ptr for
7778 : * UpdateCheckPointDistanceEstimate()
7779 : */
7780 1736 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7781 :
7782 : /*
7783 : * Update the control file.
7784 : */
7785 1736 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7786 1736 : if (shutdown)
7787 740 : ControlFile->state = DB_SHUTDOWNED;
7788 1736 : ControlFile->checkPoint = ProcLastRecPtr;
7789 1736 : ControlFile->checkPointCopy = checkPoint;
7790 : /* crash recovery should always recover to the end of WAL */
7791 1736 : ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
7792 1736 : ControlFile->minRecoveryPointTLI = 0;
7793 :
7794 : /*
7795 : * Persist unloggedLSN value. It's reset on crash recovery, so this goes
7796 : * unused on non-shutdown checkpoints, but seems useful to store it always
7797 : * for debugging purposes.
7798 : */
7799 1736 : ControlFile->unloggedLSN = pg_atomic_read_membarrier_u64(&XLogCtl->unloggedLSN);
7800 :
7801 1736 : UpdateControlFile();
7802 1736 : LWLockRelease(ControlFileLock);
7803 :
7804 : /*
7805 : * We are now done with critical updates; no need for system panic if we
7806 : * have trouble while fooling with old log segments.
7807 : */
7808 1736 : END_CRIT_SECTION();
7809 :
7810 : /*
7811 : * WAL summaries end when the next XLOG_CHECKPOINT_REDO or
7812 : * XLOG_CHECKPOINT_SHUTDOWN record is reached. This is the first point
7813 : * where (a) we're not inside of a critical section and (b) we can be
7814 : * certain that the relevant record has been flushed to disk, which must
7815 : * happen before it can be summarized.
7816 : *
7817 : * If this is a shutdown checkpoint, then this happens reasonably
7818 : * promptly: we've only just inserted and flushed the
7819 : * XLOG_CHECKPOINT_SHUTDOWN record. If this is not a shutdown checkpoint,
7820 : * then this might not be very prompt at all: the XLOG_CHECKPOINT_REDO
7821 : * record was written before we began flushing data to disk, and that
7822 : * could be many minutes ago at this point. However, we don't XLogFlush()
7823 : * after inserting that record, so we're not guaranteed that it's on disk
7824 : * until after the above call that flushes the XLOG_CHECKPOINT_ONLINE
7825 : * record.
7826 : */
7827 1736 : WakeupWalSummarizer();
7828 :
7829 : /*
7830 : * Let smgr do post-checkpoint cleanup (eg, deleting old files).
7831 : */
7832 1736 : SyncPostCheckpoint();
7833 :
7834 : /*
7835 : * Update the average distance between checkpoints if the prior checkpoint
7836 : * exists.
7837 : */
7838 1736 : if (XLogRecPtrIsValid(PriorRedoPtr))
7839 1736 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7840 :
7841 1736 : INJECTION_POINT("checkpoint-before-old-wal-removal", NULL);
7842 :
7843 : /*
7844 : * Delete old log files, those no longer needed for last checkpoint to
7845 : * prevent the disk holding the xlog from growing full.
7846 : */
7847 1736 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7848 1736 : KeepLogSeg(recptr, &_logSegNo);
7849 1736 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7850 : _logSegNo, InvalidOid,
7851 : InvalidTransactionId))
7852 : {
7853 : /*
7854 : * Some slots have been invalidated; recalculate the old-segment
7855 : * horizon, starting again from RedoRecPtr.
7856 : */
7857 4 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7858 4 : KeepLogSeg(recptr, &_logSegNo);
7859 : }
7860 1736 : _logSegNo--;
7861 1736 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
7862 : checkPoint.ThisTimeLineID);
7863 :
7864 : /*
7865 : * Make more log segments if needed. (Do this after recycling old log
7866 : * segments, since that may supply some of the needed files.)
7867 : */
7868 1736 : if (!shutdown)
7869 996 : PreallocXlogFiles(recptr, checkPoint.ThisTimeLineID);
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). During recovery, though, we mustn't do this because
7876 : * StartupSUBTRANS hasn't been called yet.
7877 : */
7878 1736 : if (!RecoveryInProgress())
7879 1706 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7880 :
7881 : /* Real work is done; log and update stats. */
7882 1736 : LogCheckpointEnd(false, flags);
7883 :
7884 : /* Reset the process title */
7885 1736 : update_checkpoint_display(flags, false, true);
7886 :
7887 : TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
7888 : NBuffers,
7889 : CheckpointStats.ckpt_segs_added,
7890 : CheckpointStats.ckpt_segs_removed,
7891 : CheckpointStats.ckpt_segs_recycled);
7892 :
7893 1736 : return true;
7894 : }
7895 :
7896 : /*
7897 : * Mark the end of recovery in WAL though without running a full checkpoint.
7898 : * We can expect that a restartpoint is likely to be in progress as we
7899 : * do this, though we are unwilling to wait for it to complete.
7900 : *
7901 : * CreateRestartPoint() allows for the case where recovery may end before
7902 : * the restartpoint completes so there is no concern of concurrent behaviour.
7903 : */
7904 : static void
7905 50 : CreateEndOfRecoveryRecord(void)
7906 : {
7907 : xl_end_of_recovery xlrec;
7908 : XLogRecPtr recptr;
7909 :
7910 : /* sanity check */
7911 50 : if (!RecoveryInProgress())
7912 0 : elog(ERROR, "can only be used to end recovery");
7913 :
7914 50 : xlrec.end_time = GetCurrentTimestamp();
7915 50 : xlrec.wal_level = wal_level;
7916 :
7917 50 : WALInsertLockAcquireExclusive();
7918 50 : xlrec.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7919 50 : xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7920 50 : WALInsertLockRelease();
7921 :
7922 50 : START_CRIT_SECTION();
7923 :
7924 50 : XLogBeginInsert();
7925 50 : XLogRegisterData(&xlrec, sizeof(xl_end_of_recovery));
7926 50 : recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY);
7927 :
7928 50 : XLogFlush(recptr);
7929 :
7930 : /*
7931 : * Update the control file so that crash recovery can follow the timeline
7932 : * changes to this point.
7933 : */
7934 50 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7935 50 : ControlFile->minRecoveryPoint = recptr;
7936 50 : ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
7937 :
7938 : /* start with the latest checksum version (as of the end of recovery) */
7939 50 : SpinLockAcquire(&XLogCtl->info_lck);
7940 50 : ControlFile->data_checksum_version = XLogCtl->data_checksum_version;
7941 50 : SpinLockRelease(&XLogCtl->info_lck);
7942 :
7943 50 : UpdateControlFile();
7944 50 : LWLockRelease(ControlFileLock);
7945 :
7946 50 : END_CRIT_SECTION();
7947 50 : }
7948 :
7949 : /*
7950 : * Write an OVERWRITE_CONTRECORD message.
7951 : *
7952 : * When on WAL replay we expect a continuation record at the start of a page
7953 : * that is not there, recovery ends and WAL writing resumes at that point.
7954 : * But it's wrong to resume writing new WAL back at the start of the record
7955 : * that was broken, because downstream consumers of that WAL (physical
7956 : * replicas) are not prepared to "rewind". So the first action after
7957 : * finishing replay of all valid WAL must be to write a record of this type
7958 : * at the point where the contrecord was missing; to support xlogreader
7959 : * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
7960 : * to the page header where the record occurs. xlogreader has an ad-hoc
7961 : * mechanism to report metadata about the broken record, which is what we
7962 : * use here.
7963 : *
7964 : * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
7965 : * skip the record it was reading, and pass back the LSN of the skipped
7966 : * record, so that its caller can verify (on "replay" of that record) that the
7967 : * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
7968 : *
7969 : * 'aborted_lsn' is the beginning position of the record that was incomplete.
7970 : * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
7971 : * beginning of the XLOG page where the record is to be inserted. They must
7972 : * match the current WAL insert position, they're passed here just so that we
7973 : * can verify that.
7974 : */
7975 : static XLogRecPtr
7976 11 : CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn, XLogRecPtr pagePtr,
7977 : TimeLineID newTLI)
7978 : {
7979 : xl_overwrite_contrecord xlrec;
7980 : XLogRecPtr recptr;
7981 : XLogPageHeader pagehdr;
7982 : XLogRecPtr startPos;
7983 :
7984 : /* sanity checks */
7985 11 : if (!RecoveryInProgress())
7986 0 : elog(ERROR, "can only be used at end of recovery");
7987 11 : if (pagePtr % XLOG_BLCKSZ != 0)
7988 0 : elog(ERROR, "invalid position for missing continuation record %X/%08X",
7989 : LSN_FORMAT_ARGS(pagePtr));
7990 :
7991 : /* The current WAL insert position should be right after the page header */
7992 11 : startPos = pagePtr;
7993 11 : if (XLogSegmentOffset(startPos, wal_segment_size) == 0)
7994 1 : startPos += SizeOfXLogLongPHD;
7995 : else
7996 10 : startPos += SizeOfXLogShortPHD;
7997 11 : recptr = GetXLogInsertRecPtr();
7998 11 : if (recptr != startPos)
7999 0 : elog(ERROR, "invalid WAL insert position %X/%08X for OVERWRITE_CONTRECORD",
8000 : LSN_FORMAT_ARGS(recptr));
8001 :
8002 11 : START_CRIT_SECTION();
8003 :
8004 : /*
8005 : * Initialize the XLOG page header (by GetXLogBuffer), and set the
8006 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
8007 : *
8008 : * No other backend is allowed to write WAL yet, so acquiring the WAL
8009 : * insertion lock is just pro forma.
8010 : */
8011 11 : WALInsertLockAcquire();
8012 11 : pagehdr = (XLogPageHeader) GetXLogBuffer(pagePtr, newTLI);
8013 11 : pagehdr->xlp_info |= XLP_FIRST_IS_OVERWRITE_CONTRECORD;
8014 11 : WALInsertLockRelease();
8015 :
8016 : /*
8017 : * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
8018 : * page. We know it becomes the first record, because no other backend is
8019 : * allowed to write WAL yet.
8020 : */
8021 11 : XLogBeginInsert();
8022 11 : xlrec.overwritten_lsn = aborted_lsn;
8023 11 : xlrec.overwrite_time = GetCurrentTimestamp();
8024 11 : XLogRegisterData(&xlrec, sizeof(xl_overwrite_contrecord));
8025 11 : recptr = XLogInsert(RM_XLOG_ID, XLOG_OVERWRITE_CONTRECORD);
8026 :
8027 : /* check that the record was inserted to the right place */
8028 11 : if (ProcLastRecPtr != startPos)
8029 0 : elog(ERROR, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%08X",
8030 : LSN_FORMAT_ARGS(ProcLastRecPtr));
8031 :
8032 11 : XLogFlush(recptr);
8033 :
8034 11 : END_CRIT_SECTION();
8035 :
8036 11 : return recptr;
8037 : }
8038 :
8039 : /*
8040 : * Flush all data in shared memory to disk, and fsync
8041 : *
8042 : * This is the common code shared between regular checkpoints and
8043 : * recovery restartpoints.
8044 : */
8045 : static void
8046 1951 : CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
8047 : {
8048 1951 : CheckPointRelationMap();
8049 1951 : CheckPointReplicationSlots(flags & CHECKPOINT_IS_SHUTDOWN);
8050 1951 : CheckPointSnapBuild();
8051 1951 : CheckPointLogicalRewriteHeap();
8052 1951 : CheckPointReplicationOrigin();
8053 :
8054 : /* Write out all dirty data in SLRUs and the main buffer pool */
8055 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags);
8056 1951 : CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
8057 1951 : CheckPointCLOG();
8058 1951 : CheckPointCommitTs();
8059 1951 : CheckPointSUBTRANS();
8060 1951 : CheckPointMultiXact();
8061 1951 : CheckPointPredicate();
8062 1951 : CheckPointBuffers(flags);
8063 :
8064 : /* Perform all queued up fsyncs */
8065 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
8066 1951 : CheckpointStats.ckpt_sync_t = GetCurrentTimestamp();
8067 1951 : ProcessSyncRequests();
8068 1951 : CheckpointStats.ckpt_sync_end_t = GetCurrentTimestamp();
8069 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
8070 :
8071 : /* We deliberately delay 2PC checkpointing as long as possible */
8072 1951 : CheckPointTwoPhase(checkPointRedo);
8073 1951 : }
8074 :
8075 : /*
8076 : * Save a checkpoint for recovery restart if appropriate
8077 : *
8078 : * This function is called each time a checkpoint record is read from XLOG.
8079 : * It must determine whether the checkpoint represents a safe restartpoint or
8080 : * not. If so, the checkpoint record is stashed in shared memory so that
8081 : * CreateRestartPoint can consult it. (Note that the latter function is
8082 : * executed by the checkpointer, while this one will be executed by the
8083 : * startup process.)
8084 : */
8085 : static void
8086 753 : RecoveryRestartPoint(const CheckPoint *checkPoint, XLogReaderState *record)
8087 : {
8088 : /*
8089 : * Also refrain from creating a restartpoint if we have seen any
8090 : * references to non-existent pages. Restarting recovery from the
8091 : * restartpoint would not see the references, so we would lose the
8092 : * cross-check that the pages belonged to a relation that was dropped
8093 : * later.
8094 : */
8095 753 : if (XLogHaveInvalidPages())
8096 : {
8097 0 : elog(DEBUG2,
8098 : "could not record restart point at %X/%08X because there are unresolved references to invalid pages",
8099 : LSN_FORMAT_ARGS(checkPoint->redo));
8100 0 : return;
8101 : }
8102 :
8103 : /*
8104 : * Copy the checkpoint record to shared memory, so that checkpointer can
8105 : * work out the next time it wants to perform a restartpoint.
8106 : */
8107 753 : SpinLockAcquire(&XLogCtl->info_lck);
8108 753 : XLogCtl->lastCheckPointRecPtr = record->ReadRecPtr;
8109 753 : XLogCtl->lastCheckPointEndPtr = record->EndRecPtr;
8110 753 : XLogCtl->lastCheckPoint = *checkPoint;
8111 753 : SpinLockRelease(&XLogCtl->info_lck);
8112 : }
8113 :
8114 : /*
8115 : * Establish a restartpoint if possible.
8116 : *
8117 : * This is similar to CreateCheckPoint, but is used during WAL recovery
8118 : * to establish a point from which recovery can roll forward without
8119 : * replaying the entire recovery log.
8120 : *
8121 : * Returns true if a new restartpoint was established. We can only establish
8122 : * a restartpoint if we have replayed a safe checkpoint record since last
8123 : * restartpoint.
8124 : */
8125 : bool
8126 617 : CreateRestartPoint(int flags)
8127 : {
8128 : XLogRecPtr lastCheckPointRecPtr;
8129 : XLogRecPtr lastCheckPointEndPtr;
8130 : CheckPoint lastCheckPoint;
8131 : XLogRecPtr PriorRedoPtr;
8132 : XLogRecPtr receivePtr;
8133 : XLogRecPtr replayPtr;
8134 : TimeLineID replayTLI;
8135 : XLogRecPtr endptr;
8136 : XLogSegNo _logSegNo;
8137 : TimestampTz xtime;
8138 :
8139 : /* Concurrent checkpoint/restartpoint cannot happen */
8140 : Assert(!IsUnderPostmaster || MyBackendType == B_CHECKPOINTER);
8141 :
8142 : /* Get a local copy of the last safe checkpoint record. */
8143 617 : SpinLockAcquire(&XLogCtl->info_lck);
8144 617 : lastCheckPointRecPtr = XLogCtl->lastCheckPointRecPtr;
8145 617 : lastCheckPointEndPtr = XLogCtl->lastCheckPointEndPtr;
8146 617 : lastCheckPoint = XLogCtl->lastCheckPoint;
8147 617 : SpinLockRelease(&XLogCtl->info_lck);
8148 :
8149 : /*
8150 : * Check that we're still in recovery mode. It's ok if we exit recovery
8151 : * mode after this check, the restart point is valid anyway.
8152 : */
8153 617 : if (!RecoveryInProgress())
8154 : {
8155 0 : ereport(DEBUG2,
8156 : (errmsg_internal("skipping restartpoint, recovery has already ended")));
8157 0 : return false;
8158 : }
8159 :
8160 : /*
8161 : * If the last checkpoint record we've replayed is already our last
8162 : * restartpoint, we can't perform a new restart point. We still update
8163 : * minRecoveryPoint in that case, so that if this is a shutdown restart
8164 : * point, we won't start up earlier than before. That's not strictly
8165 : * necessary, but when hot standby is enabled, it would be rather weird if
8166 : * the database opened up for read-only connections at a point-in-time
8167 : * before the last shutdown. Such time travel is still possible in case of
8168 : * immediate shutdown, though.
8169 : *
8170 : * We don't explicitly advance minRecoveryPoint when we do create a
8171 : * restartpoint. It's assumed that flushing the buffers will do that as a
8172 : * side-effect.
8173 : */
8174 617 : if (!XLogRecPtrIsValid(lastCheckPointRecPtr) ||
8175 286 : lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
8176 : {
8177 402 : ereport(DEBUG2,
8178 : errmsg_internal("skipping restartpoint, already performed at %X/%08X",
8179 : LSN_FORMAT_ARGS(lastCheckPoint.redo)));
8180 :
8181 402 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
8182 402 : if (flags & CHECKPOINT_IS_SHUTDOWN)
8183 : {
8184 39 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8185 39 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
8186 39 : UpdateControlFile();
8187 39 : LWLockRelease(ControlFileLock);
8188 : }
8189 402 : return false;
8190 : }
8191 :
8192 : /*
8193 : * Update the shared RedoRecPtr so that the startup process can calculate
8194 : * the number of segments replayed since last restartpoint, and request a
8195 : * restartpoint if it exceeds CheckPointSegments.
8196 : *
8197 : * Like in CreateCheckPoint(), hold off insertions to update it, although
8198 : * during recovery this is just pro forma, because no WAL insertions are
8199 : * happening.
8200 : */
8201 215 : WALInsertLockAcquireExclusive();
8202 215 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = lastCheckPoint.redo;
8203 215 : WALInsertLockRelease();
8204 :
8205 : /* Also update the info_lck-protected copy */
8206 215 : SpinLockAcquire(&XLogCtl->info_lck);
8207 215 : XLogCtl->RedoRecPtr = lastCheckPoint.redo;
8208 215 : SpinLockRelease(&XLogCtl->info_lck);
8209 :
8210 : /*
8211 : * Prepare to accumulate statistics.
8212 : *
8213 : * Note: because it is possible for log_checkpoints to change while a
8214 : * checkpoint proceeds, we always accumulate stats, even if
8215 : * log_checkpoints is currently off.
8216 : */
8217 2365 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
8218 215 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
8219 :
8220 215 : if (log_checkpoints)
8221 215 : LogCheckpointStart(flags, true);
8222 :
8223 : /* Update the process title */
8224 215 : update_checkpoint_display(flags, true, false);
8225 :
8226 215 : CheckPointGuts(lastCheckPoint.redo, flags);
8227 :
8228 : /*
8229 : * This location needs to be after CheckPointGuts() to ensure that some
8230 : * work has already happened during this checkpoint.
8231 : */
8232 215 : INJECTION_POINT("create-restart-point", NULL);
8233 :
8234 : /*
8235 : * Remember the prior checkpoint's redo ptr for
8236 : * UpdateCheckPointDistanceEstimate()
8237 : */
8238 215 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
8239 :
8240 : /*
8241 : * Update pg_control, using current time. Check that it still shows an
8242 : * older checkpoint, else do nothing; this is a quick hack to make sure
8243 : * nothing really bad happens if somehow we get here after the
8244 : * end-of-recovery checkpoint.
8245 : */
8246 215 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8247 215 : if (ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
8248 : {
8249 : /*
8250 : * Update the checkpoint information. We do this even if the cluster
8251 : * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
8252 : * segments recycled below.
8253 : */
8254 215 : ControlFile->checkPoint = lastCheckPointRecPtr;
8255 215 : ControlFile->checkPointCopy = lastCheckPoint;
8256 :
8257 : /*
8258 : * Ensure minRecoveryPoint is past the checkpoint record and update it
8259 : * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
8260 : * this will have happened already while writing out dirty buffers,
8261 : * but not necessarily - e.g. because no buffers were dirtied. We do
8262 : * this because a backup performed in recovery uses minRecoveryPoint
8263 : * to determine which WAL files must be included in the backup, and
8264 : * the file (or files) containing the checkpoint record must be
8265 : * included, at a minimum. Note that for an ordinary restart of
8266 : * recovery there's no value in having the minimum recovery point any
8267 : * earlier than this anyway, because redo will begin just after the
8268 : * checkpoint record.
8269 : */
8270 215 : if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
8271 : {
8272 214 : if (ControlFile->minRecoveryPoint < lastCheckPointEndPtr)
8273 : {
8274 21 : ControlFile->minRecoveryPoint = lastCheckPointEndPtr;
8275 21 : ControlFile->minRecoveryPointTLI = lastCheckPoint.ThisTimeLineID;
8276 :
8277 : /* update local copy */
8278 21 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
8279 21 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
8280 : }
8281 214 : if (flags & CHECKPOINT_IS_SHUTDOWN)
8282 24 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
8283 : }
8284 :
8285 : /* we shall start with the latest checksum version */
8286 215 : ControlFile->data_checksum_version = lastCheckPoint.dataChecksumState;
8287 :
8288 215 : UpdateControlFile();
8289 : }
8290 215 : LWLockRelease(ControlFileLock);
8291 :
8292 : /*
8293 : * Update the average distance between checkpoints/restartpoints if the
8294 : * prior checkpoint exists.
8295 : */
8296 215 : if (XLogRecPtrIsValid(PriorRedoPtr))
8297 215 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
8298 :
8299 : /*
8300 : * Delete old log files, those no longer needed for last restartpoint to
8301 : * prevent the disk holding the xlog from growing full.
8302 : */
8303 215 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
8304 :
8305 : /*
8306 : * Retreat _logSegNo using the current end of xlog replayed or received,
8307 : * whichever is later.
8308 : */
8309 215 : receivePtr = GetWalRcvFlushRecPtr(NULL, NULL);
8310 215 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
8311 215 : endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
8312 215 : KeepLogSeg(endptr, &_logSegNo);
8313 :
8314 215 : INJECTION_POINT("restartpoint-before-slot-invalidation", NULL);
8315 :
8316 215 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
8317 : _logSegNo, InvalidOid,
8318 : InvalidTransactionId))
8319 : {
8320 : /*
8321 : * Some slots have been invalidated; recalculate the old-segment
8322 : * horizon, starting again from RedoRecPtr.
8323 : */
8324 1 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
8325 1 : KeepLogSeg(endptr, &_logSegNo);
8326 : }
8327 215 : _logSegNo--;
8328 :
8329 : /*
8330 : * Try to recycle segments on a useful timeline. If we've been promoted
8331 : * since the beginning of this restartpoint, use the new timeline chosen
8332 : * at end of recovery. If we're still in recovery, use the timeline we're
8333 : * currently replaying.
8334 : *
8335 : * There is no guarantee that the WAL segments will be useful on the
8336 : * current timeline; if recovery proceeds to a new timeline right after
8337 : * this, the pre-allocated WAL segments on this timeline will not be used,
8338 : * and will go wasted until recycled on the next restartpoint. We'll live
8339 : * with that.
8340 : */
8341 215 : if (!RecoveryInProgress())
8342 1 : replayTLI = XLogCtl->InsertTimeLineID;
8343 :
8344 215 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, endptr, replayTLI);
8345 :
8346 : /*
8347 : * Make more log segments if needed. (Do this after recycling old log
8348 : * segments, since that may supply some of the needed files.)
8349 : */
8350 215 : PreallocXlogFiles(endptr, replayTLI);
8351 :
8352 : /*
8353 : * Truncate pg_subtrans if possible. We can throw away all data before
8354 : * the oldest XMIN of any running transaction. No future transaction will
8355 : * attempt to reference any pg_subtrans entry older than that (see Asserts
8356 : * in subtrans.c). When hot standby is disabled, though, we mustn't do
8357 : * this because StartupSUBTRANS hasn't been called yet.
8358 : */
8359 215 : if (EnableHotStandby)
8360 215 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
8361 :
8362 : /* Real work is done; log and update stats. */
8363 215 : LogCheckpointEnd(true, flags);
8364 :
8365 : /* Reset the process title */
8366 215 : update_checkpoint_display(flags, true, true);
8367 :
8368 215 : xtime = GetLatestXTime();
8369 215 : ereport((log_checkpoints ? LOG : DEBUG2),
8370 : errmsg("recovery restart point at %X/%08X",
8371 : LSN_FORMAT_ARGS(lastCheckPoint.redo)),
8372 : xtime ? errdetail("Last completed transaction was at log time %s.",
8373 : timestamptz_to_str(xtime)) : 0);
8374 :
8375 : /*
8376 : * Finally, execute archive_cleanup_command, if any.
8377 : */
8378 215 : if (archiveCleanupCommand && strcmp(archiveCleanupCommand, "") != 0)
8379 0 : ExecuteRecoveryCommand(archiveCleanupCommand,
8380 : "archive_cleanup_command",
8381 : false,
8382 : WAIT_EVENT_ARCHIVE_CLEANUP_COMMAND);
8383 :
8384 215 : return true;
8385 : }
8386 :
8387 : /*
8388 : * Report availability of WAL for the given target LSN
8389 : * (typically a slot's restart_lsn)
8390 : *
8391 : * Returns one of the following enum values:
8392 : *
8393 : * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
8394 : * max_wal_size.
8395 : *
8396 : * * WALAVAIL_EXTENDED means it is still available by preserving extra
8397 : * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
8398 : * than max_wal_size, this state is not returned.
8399 : *
8400 : * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
8401 : * remove reserved segments. The walsender using this slot may return to the
8402 : * above.
8403 : *
8404 : * * WALAVAIL_REMOVED means it has been removed. A replication stream on
8405 : * a slot with this LSN cannot continue. (Any associated walsender
8406 : * processes should have been terminated already.)
8407 : *
8408 : * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
8409 : */
8410 : WALAvailability
8411 664 : GetWALAvailability(XLogRecPtr targetLSN)
8412 : {
8413 : XLogRecPtr currpos; /* current write LSN */
8414 : XLogSegNo currSeg; /* segid of currpos */
8415 : XLogSegNo targetSeg; /* segid of targetLSN */
8416 : XLogSegNo oldestSeg; /* actual oldest segid */
8417 : XLogSegNo oldestSegMaxWalSize; /* oldest segid kept by max_wal_size */
8418 : XLogSegNo oldestSlotSeg; /* oldest segid kept by slot */
8419 : uint64 keepSegs;
8420 :
8421 : /*
8422 : * slot does not reserve WAL. Either deactivated, or has never been active
8423 : */
8424 664 : if (!XLogRecPtrIsValid(targetLSN))
8425 31 : return WALAVAIL_INVALID_LSN;
8426 :
8427 : /*
8428 : * Calculate the oldest segment currently reserved by all slots,
8429 : * considering wal_keep_size and max_slot_wal_keep_size. Initialize
8430 : * oldestSlotSeg to the current segment.
8431 : */
8432 633 : currpos = GetXLogWriteRecPtr();
8433 633 : XLByteToSeg(currpos, oldestSlotSeg, wal_segment_size);
8434 633 : KeepLogSeg(currpos, &oldestSlotSeg);
8435 :
8436 : /*
8437 : * Find the oldest extant segment file. We get 1 until checkpoint removes
8438 : * the first WAL segment file since startup, which causes the status being
8439 : * wrong under certain abnormal conditions but that doesn't actually harm.
8440 : */
8441 633 : oldestSeg = XLogGetLastRemovedSegno() + 1;
8442 :
8443 : /* calculate oldest segment by max_wal_size */
8444 633 : XLByteToSeg(currpos, currSeg, wal_segment_size);
8445 633 : keepSegs = ConvertToXSegs(max_wal_size_mb, wal_segment_size) + 1;
8446 :
8447 633 : if (currSeg > keepSegs)
8448 8 : oldestSegMaxWalSize = currSeg - keepSegs;
8449 : else
8450 625 : oldestSegMaxWalSize = 1;
8451 :
8452 : /* the segment we care about */
8453 633 : XLByteToSeg(targetLSN, targetSeg, wal_segment_size);
8454 :
8455 : /*
8456 : * No point in returning reserved or extended status values if the
8457 : * targetSeg is known to be lost.
8458 : */
8459 633 : if (targetSeg >= oldestSlotSeg)
8460 : {
8461 : /* show "reserved" when targetSeg is within max_wal_size */
8462 632 : if (targetSeg >= oldestSegMaxWalSize)
8463 630 : return WALAVAIL_RESERVED;
8464 :
8465 : /* being retained by slots exceeding max_wal_size */
8466 2 : return WALAVAIL_EXTENDED;
8467 : }
8468 :
8469 : /* WAL segments are no longer retained but haven't been removed yet */
8470 1 : if (targetSeg >= oldestSeg)
8471 1 : return WALAVAIL_UNRESERVED;
8472 :
8473 : /* Definitely lost */
8474 0 : return WALAVAIL_REMOVED;
8475 : }
8476 :
8477 :
8478 : /*
8479 : * Retreat *logSegNo to the last segment that we need to retain because of
8480 : * either wal_keep_size or replication slots.
8481 : *
8482 : * This is calculated by subtracting wal_keep_size from the given xlog
8483 : * location, recptr and by making sure that that result is below the
8484 : * requirement of replication slots. For the latter criterion we do consider
8485 : * the effects of max_slot_wal_keep_size: reserve at most that much space back
8486 : * from recptr.
8487 : *
8488 : * Note about replication slots: if this function calculates a value
8489 : * that's further ahead than what slots need reserved, then affected
8490 : * slots need to be invalidated and this function invoked again.
8491 : * XXX it might be a good idea to rewrite this function so that
8492 : * invalidation is optionally done here, instead.
8493 : */
8494 : static void
8495 2589 : KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
8496 : {
8497 : XLogSegNo currSegNo;
8498 : XLogSegNo segno;
8499 : XLogRecPtr keep;
8500 :
8501 2589 : XLByteToSeg(recptr, currSegNo, wal_segment_size);
8502 2589 : segno = currSegNo;
8503 :
8504 : /* Calculate how many segments are kept by slots. */
8505 2589 : keep = XLogGetReplicationSlotMinimumLSN();
8506 2589 : if (XLogRecPtrIsValid(keep) && keep < recptr)
8507 : {
8508 797 : XLByteToSeg(keep, segno, wal_segment_size);
8509 :
8510 : /*
8511 : * Account for max_slot_wal_keep_size to avoid keeping more than
8512 : * configured. However, don't do that during a binary upgrade: if
8513 : * slots were to be invalidated because of this, it would not be
8514 : * possible to preserve logical ones during the upgrade.
8515 : */
8516 797 : if (max_slot_wal_keep_size_mb >= 0 && !IsBinaryUpgrade)
8517 : {
8518 : uint64 slot_keep_segs;
8519 :
8520 23 : slot_keep_segs =
8521 23 : ConvertToXSegs(max_slot_wal_keep_size_mb, wal_segment_size);
8522 :
8523 23 : if (currSegNo - segno > slot_keep_segs)
8524 6 : segno = currSegNo - slot_keep_segs;
8525 : }
8526 : }
8527 :
8528 : /*
8529 : * If WAL summarization is in use, don't remove WAL that has yet to be
8530 : * summarized.
8531 : */
8532 2589 : keep = GetOldestUnsummarizedLSN(NULL, NULL);
8533 2589 : if (XLogRecPtrIsValid(keep))
8534 : {
8535 : XLogSegNo unsummarized_segno;
8536 :
8537 7 : XLByteToSeg(keep, unsummarized_segno, wal_segment_size);
8538 7 : if (unsummarized_segno < segno)
8539 7 : segno = unsummarized_segno;
8540 : }
8541 :
8542 : /* but, keep at least wal_keep_size if that's set */
8543 2589 : if (wal_keep_size_mb > 0)
8544 : {
8545 : uint64 keep_segs;
8546 :
8547 74 : keep_segs = ConvertToXSegs(wal_keep_size_mb, wal_segment_size);
8548 74 : if (currSegNo - segno < keep_segs)
8549 : {
8550 : /* avoid underflow, don't go below 1 */
8551 74 : if (currSegNo <= keep_segs)
8552 70 : segno = 1;
8553 : else
8554 4 : segno = currSegNo - keep_segs;
8555 : }
8556 : }
8557 :
8558 : /* don't delete WAL segments newer than the calculated segment */
8559 2589 : if (segno < *logSegNo)
8560 364 : *logSegNo = segno;
8561 2589 : }
8562 :
8563 : /*
8564 : * Write a NEXTOID log record
8565 : */
8566 : void
8567 691 : XLogPutNextOid(Oid nextOid)
8568 : {
8569 691 : XLogBeginInsert();
8570 691 : XLogRegisterData(&nextOid, sizeof(Oid));
8571 691 : (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID);
8572 :
8573 : /*
8574 : * We need not flush the NEXTOID record immediately, because any of the
8575 : * just-allocated OIDs could only reach disk as part of a tuple insert or
8576 : * update that would have its own XLOG record that must follow the NEXTOID
8577 : * record. Therefore, the standard buffer LSN interlock applied to those
8578 : * records will ensure no such OID reaches disk before the NEXTOID record
8579 : * does.
8580 : *
8581 : * Note, however, that the above statement only covers state "within" the
8582 : * database. When we use a generated OID as a file or directory name, we
8583 : * are in a sense violating the basic WAL rule, because that filesystem
8584 : * change may reach disk before the NEXTOID WAL record does. The impact
8585 : * of this is that if a database crash occurs immediately afterward, we
8586 : * might after restart re-generate the same OID and find that it conflicts
8587 : * with the leftover file or directory. But since for safety's sake we
8588 : * always loop until finding a nonconflicting filename, this poses no real
8589 : * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8590 : */
8591 691 : }
8592 :
8593 : /*
8594 : * Write an XLOG SWITCH record.
8595 : *
8596 : * Here we just blindly issue an XLogInsert request for the record.
8597 : * All the magic happens inside XLogInsert.
8598 : *
8599 : * The return value is either the end+1 address of the switch record,
8600 : * or the end+1 address of the prior segment if we did not need to
8601 : * write a switch record because we are already at segment start.
8602 : */
8603 : XLogRecPtr
8604 823 : RequestXLogSwitch(bool mark_unimportant)
8605 : {
8606 : XLogRecPtr RecPtr;
8607 :
8608 : /* XLOG SWITCH has no data */
8609 823 : XLogBeginInsert();
8610 :
8611 823 : if (mark_unimportant)
8612 0 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8613 823 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH);
8614 :
8615 823 : return RecPtr;
8616 : }
8617 :
8618 : /*
8619 : * Write a RESTORE POINT record
8620 : */
8621 : XLogRecPtr
8622 3 : XLogRestorePoint(const char *rpName)
8623 : {
8624 : XLogRecPtr RecPtr;
8625 : xl_restore_point xlrec;
8626 :
8627 3 : xlrec.rp_time = GetCurrentTimestamp();
8628 3 : strlcpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8629 :
8630 3 : XLogBeginInsert();
8631 3 : XLogRegisterData(&xlrec, sizeof(xl_restore_point));
8632 :
8633 3 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT);
8634 :
8635 3 : ereport(LOG,
8636 : errmsg("restore point \"%s\" created at %X/%08X",
8637 : rpName, LSN_FORMAT_ARGS(RecPtr)));
8638 :
8639 3 : return RecPtr;
8640 : }
8641 :
8642 : /*
8643 : * Write an empty XLOG record to assign a distinct LSN.
8644 : *
8645 : * This is used by some index AMs when building indexes on permanent relations
8646 : * with wal_level=minimal. In that scenario, WAL-logging will start after
8647 : * commit, but the index AM needs distinct LSNs to detect concurrent page
8648 : * modifications. When the current WAL insert position hasn't advanced since
8649 : * the last call, we emit a dummy record to ensure we get a new, distinct LSN.
8650 : */
8651 : XLogRecPtr
8652 441 : XLogAssignLSN(void)
8653 : {
8654 441 : int dummy = 0;
8655 :
8656 : /*
8657 : * Records other than XLOG_SWITCH must have content. We use an integer 0
8658 : * to satisfy this restriction.
8659 : */
8660 441 : XLogBeginInsert();
8661 441 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8662 441 : XLogRegisterData(&dummy, sizeof(dummy));
8663 441 : return XLogInsert(RM_XLOG_ID, XLOG_ASSIGN_LSN);
8664 : }
8665 :
8666 : /*
8667 : * Check if any of the GUC parameters that are critical for hot standby
8668 : * have changed, and update the value in pg_control file if necessary.
8669 : */
8670 : static void
8671 1018 : XLogReportParameters(void)
8672 : {
8673 1018 : if (wal_level != ControlFile->wal_level ||
8674 743 : wal_log_hints != ControlFile->wal_log_hints ||
8675 651 : MaxConnections != ControlFile->MaxConnections ||
8676 650 : max_worker_processes != ControlFile->max_worker_processes ||
8677 647 : max_wal_senders != ControlFile->max_wal_senders ||
8678 619 : max_prepared_xacts != ControlFile->max_prepared_xacts ||
8679 515 : max_locks_per_xact != ControlFile->max_locks_per_xact ||
8680 515 : track_commit_timestamp != ControlFile->track_commit_timestamp)
8681 : {
8682 : /*
8683 : * The change in number of backend slots doesn't need to be WAL-logged
8684 : * if archiving is not enabled, as you can't start archive recovery
8685 : * with wal_level=minimal anyway. We don't really care about the
8686 : * values in pg_control either if wal_level=minimal, but seems better
8687 : * to keep them up-to-date to avoid confusion.
8688 : */
8689 515 : if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8690 : {
8691 : xl_parameter_change xlrec;
8692 : XLogRecPtr recptr;
8693 :
8694 488 : xlrec.MaxConnections = MaxConnections;
8695 488 : xlrec.max_worker_processes = max_worker_processes;
8696 488 : xlrec.max_wal_senders = max_wal_senders;
8697 488 : xlrec.max_prepared_xacts = max_prepared_xacts;
8698 488 : xlrec.max_locks_per_xact = max_locks_per_xact;
8699 488 : xlrec.wal_level = wal_level;
8700 488 : xlrec.wal_log_hints = wal_log_hints;
8701 488 : xlrec.track_commit_timestamp = track_commit_timestamp;
8702 :
8703 488 : XLogBeginInsert();
8704 488 : XLogRegisterData(&xlrec, sizeof(xlrec));
8705 :
8706 488 : recptr = XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE);
8707 488 : XLogFlush(recptr);
8708 : }
8709 :
8710 515 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8711 :
8712 515 : ControlFile->MaxConnections = MaxConnections;
8713 515 : ControlFile->max_worker_processes = max_worker_processes;
8714 515 : ControlFile->max_wal_senders = max_wal_senders;
8715 515 : ControlFile->max_prepared_xacts = max_prepared_xacts;
8716 515 : ControlFile->max_locks_per_xact = max_locks_per_xact;
8717 515 : ControlFile->wal_level = wal_level;
8718 515 : ControlFile->wal_log_hints = wal_log_hints;
8719 515 : ControlFile->track_commit_timestamp = track_commit_timestamp;
8720 515 : UpdateControlFile();
8721 :
8722 515 : LWLockRelease(ControlFileLock);
8723 : }
8724 1018 : }
8725 :
8726 : /*
8727 : * Log the new state of checksums
8728 : */
8729 : static void
8730 28 : XLogChecksums(uint32 new_type)
8731 : {
8732 : xl_checksum_state xlrec;
8733 : XLogRecPtr recptr;
8734 :
8735 28 : xlrec.new_checksum_state = new_type;
8736 :
8737 28 : XLogBeginInsert();
8738 28 : XLogRegisterData((char *) &xlrec, sizeof(xl_checksum_state));
8739 :
8740 28 : recptr = XLogInsert(RM_XLOG2_ID, XLOG2_CHECKSUMS);
8741 28 : XLogFlush(recptr);
8742 28 : }
8743 :
8744 : /*
8745 : * Update full_page_writes in shared memory, and write an
8746 : * XLOG_FPW_CHANGE record if necessary.
8747 : *
8748 : * Note: this function assumes there is no other process running
8749 : * concurrently that could update it.
8750 : */
8751 : void
8752 1743 : UpdateFullPageWrites(void)
8753 : {
8754 1743 : XLogCtlInsert *Insert = &XLogCtl->Insert;
8755 : bool recoveryInProgress;
8756 :
8757 : /*
8758 : * Do nothing if full_page_writes has not been changed.
8759 : *
8760 : * It's safe to check the shared full_page_writes without the lock,
8761 : * because we assume that there is no concurrently running process which
8762 : * can update it.
8763 : */
8764 1743 : if (fullPageWrites == Insert->fullPageWrites)
8765 1273 : return;
8766 :
8767 : /*
8768 : * Perform this outside critical section so that the WAL insert
8769 : * initialization done by RecoveryInProgress() doesn't trigger an
8770 : * assertion failure.
8771 : */
8772 470 : recoveryInProgress = RecoveryInProgress();
8773 :
8774 470 : START_CRIT_SECTION();
8775 :
8776 : /*
8777 : * It's always safe to take full page images, even when not strictly
8778 : * required, but not the other round. So if we're setting full_page_writes
8779 : * to true, first set it true and then write the WAL record. If we're
8780 : * setting it to false, first write the WAL record and then set the global
8781 : * flag.
8782 : */
8783 470 : if (fullPageWrites)
8784 : {
8785 457 : WALInsertLockAcquireExclusive();
8786 457 : Insert->fullPageWrites = true;
8787 457 : WALInsertLockRelease();
8788 : }
8789 :
8790 : /*
8791 : * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8792 : * full_page_writes during archive recovery, if required.
8793 : */
8794 470 : if (XLogStandbyInfoActive() && !recoveryInProgress)
8795 : {
8796 0 : XLogBeginInsert();
8797 0 : XLogRegisterData(&fullPageWrites, sizeof(bool));
8798 :
8799 0 : XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
8800 : }
8801 :
8802 470 : if (!fullPageWrites)
8803 : {
8804 13 : WALInsertLockAcquireExclusive();
8805 13 : Insert->fullPageWrites = false;
8806 13 : WALInsertLockRelease();
8807 : }
8808 470 : END_CRIT_SECTION();
8809 : }
8810 :
8811 : /*
8812 : * XLOG resource manager's routines
8813 : *
8814 : * Definitions of info values are in include/catalog/pg_control.h, though
8815 : * not all record types are related to control file updates.
8816 : *
8817 : * NOTE: Some XLOG record types that are directly related to WAL recovery
8818 : * are handled in xlogrecovery_redo().
8819 : */
8820 : void
8821 115286 : xlog_redo(XLogReaderState *record)
8822 : {
8823 115286 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8824 115286 : XLogRecPtr lsn = record->EndRecPtr;
8825 :
8826 : /*
8827 : * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
8828 : * XLOG_FPI_FOR_HINT records.
8829 : */
8830 : Assert(info == XLOG_FPI || info == XLOG_FPI_FOR_HINT ||
8831 : !XLogRecHasAnyBlockRefs(record));
8832 :
8833 115286 : if (info == XLOG_NEXTOID)
8834 : {
8835 : Oid nextOid;
8836 :
8837 : /*
8838 : * We used to try to take the maximum of TransamVariables->nextOid and
8839 : * the recorded nextOid, but that fails if the OID counter wraps
8840 : * around. Since no OID allocation should be happening during replay
8841 : * anyway, better to just believe the record exactly. We still take
8842 : * OidGenLock while setting the variable, just in case.
8843 : */
8844 97 : memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
8845 97 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8846 97 : TransamVariables->nextOid = nextOid;
8847 97 : TransamVariables->oidCount = 0;
8848 97 : LWLockRelease(OidGenLock);
8849 : }
8850 115189 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
8851 : {
8852 : CheckPoint checkPoint;
8853 : TimeLineID replayTLI;
8854 :
8855 44 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8856 : /* In a SHUTDOWN checkpoint, believe the counters exactly */
8857 44 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8858 44 : TransamVariables->nextXid = checkPoint.nextXid;
8859 44 : LWLockRelease(XidGenLock);
8860 44 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8861 44 : TransamVariables->nextOid = checkPoint.nextOid;
8862 44 : TransamVariables->oidCount = 0;
8863 44 : LWLockRelease(OidGenLock);
8864 44 : MultiXactSetNextMXact(checkPoint.nextMulti,
8865 : checkPoint.nextMultiOffset);
8866 :
8867 44 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8868 : checkPoint.oldestMultiDB);
8869 :
8870 : /*
8871 : * No need to set oldestClogXid here as well; it'll be set when we
8872 : * redo an xl_clog_truncate if it changed since initialization.
8873 : */
8874 44 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
8875 :
8876 : /*
8877 : * If we see a shutdown checkpoint while waiting for an end-of-backup
8878 : * record, the backup was canceled and the end-of-backup record will
8879 : * never arrive.
8880 : */
8881 44 : if (ArchiveRecoveryRequested &&
8882 43 : XLogRecPtrIsValid(ControlFile->backupStartPoint) &&
8883 0 : !XLogRecPtrIsValid(ControlFile->backupEndPoint))
8884 0 : ereport(PANIC,
8885 : (errmsg("online backup was canceled, recovery cannot continue")));
8886 :
8887 : /*
8888 : * If we see a shutdown checkpoint, we know that nothing was running
8889 : * on the primary at this point. So fake-up an empty running-xacts
8890 : * record and use that here and now. Recover additional standby state
8891 : * for prepared transactions.
8892 : */
8893 44 : if (standbyState >= STANDBY_INITIALIZED)
8894 : {
8895 : TransactionId *xids;
8896 : int nxids;
8897 : TransactionId oldestActiveXID;
8898 : TransactionId latestCompletedXid;
8899 : RunningTransactionsData running;
8900 :
8901 41 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
8902 :
8903 : /* Update pg_subtrans entries for any prepared transactions */
8904 41 : StandbyRecoverPreparedTransactions();
8905 :
8906 : /*
8907 : * Construct a RunningTransactions snapshot representing a shut
8908 : * down server, with only prepared transactions still alive. We're
8909 : * never overflowed at this point because all subxids are listed
8910 : * with their parent prepared transactions.
8911 : */
8912 41 : running.xcnt = nxids;
8913 41 : running.subxcnt = 0;
8914 41 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
8915 41 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
8916 41 : running.oldestRunningXid = oldestActiveXID;
8917 41 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
8918 41 : TransactionIdRetreat(latestCompletedXid);
8919 : Assert(TransactionIdIsNormal(latestCompletedXid));
8920 41 : running.latestCompletedXid = latestCompletedXid;
8921 41 : running.xids = xids;
8922 :
8923 41 : ProcArrayApplyRecoveryInfo(&running);
8924 : }
8925 :
8926 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8927 44 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8928 44 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8929 44 : ControlFile->data_checksum_version = checkPoint.dataChecksumState;
8930 :
8931 44 : UpdateControlFile();
8932 44 : LWLockRelease(ControlFileLock);
8933 :
8934 : /*
8935 : * We should've already switched to the new TLI before replaying this
8936 : * record.
8937 : */
8938 44 : (void) GetCurrentReplayRecPtr(&replayTLI);
8939 44 : if (checkPoint.ThisTimeLineID != replayTLI)
8940 0 : ereport(PANIC,
8941 : (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
8942 : checkPoint.ThisTimeLineID, replayTLI)));
8943 :
8944 44 : RecoveryRestartPoint(&checkPoint, record);
8945 :
8946 : /*
8947 : * After replaying a checkpoint record, free all smgr objects.
8948 : * Otherwise we would never do so for dropped relations, as the
8949 : * startup does not process shared invalidation messages or call
8950 : * AtEOXact_SMgr().
8951 : */
8952 44 : smgrdestroyall();
8953 : }
8954 115145 : else if (info == XLOG_CHECKPOINT_ONLINE)
8955 : {
8956 : CheckPoint checkPoint;
8957 : TimeLineID replayTLI;
8958 :
8959 709 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8960 : /* In an ONLINE checkpoint, treat the XID counter as a minimum */
8961 709 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8962 709 : if (FullTransactionIdPrecedes(TransamVariables->nextXid,
8963 : checkPoint.nextXid))
8964 0 : TransamVariables->nextXid = checkPoint.nextXid;
8965 709 : LWLockRelease(XidGenLock);
8966 :
8967 : /*
8968 : * We ignore the nextOid counter in an ONLINE checkpoint, preferring
8969 : * to track OID assignment through XLOG_NEXTOID records. The nextOid
8970 : * counter is from the start of the checkpoint and might well be stale
8971 : * compared to later XLOG_NEXTOID records. We could try to take the
8972 : * maximum of the nextOid counter and our latest value, but since
8973 : * there's no particular guarantee about the speed with which the OID
8974 : * counter wraps around, that's a risky thing to do. In any case,
8975 : * users of the nextOid counter are required to avoid assignment of
8976 : * duplicates, so that a somewhat out-of-date value should be safe.
8977 : */
8978 :
8979 : /* Handle multixact */
8980 709 : MultiXactAdvanceNextMXact(checkPoint.nextMulti,
8981 : checkPoint.nextMultiOffset);
8982 :
8983 : /*
8984 : * NB: This may perform multixact truncation when replaying WAL
8985 : * generated by an older primary.
8986 : */
8987 709 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8988 : checkPoint.oldestMultiDB);
8989 709 : if (TransactionIdPrecedes(TransamVariables->oldestXid,
8990 : checkPoint.oldestXid))
8991 0 : SetTransactionIdLimit(checkPoint.oldestXid,
8992 : checkPoint.oldestXidDB);
8993 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8994 709 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8995 709 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8996 709 : LWLockRelease(ControlFileLock);
8997 :
8998 : /* TLI should not change in an on-line checkpoint */
8999 709 : (void) GetCurrentReplayRecPtr(&replayTLI);
9000 709 : if (checkPoint.ThisTimeLineID != replayTLI)
9001 0 : ereport(PANIC,
9002 : (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
9003 : checkPoint.ThisTimeLineID, replayTLI)));
9004 :
9005 709 : RecoveryRestartPoint(&checkPoint, record);
9006 :
9007 : /*
9008 : * After replaying a checkpoint record, free all smgr objects.
9009 : * Otherwise we would never do so for dropped relations, as the
9010 : * startup does not process shared invalidation messages or call
9011 : * AtEOXact_SMgr().
9012 : */
9013 709 : smgrdestroyall();
9014 : }
9015 114436 : else if (info == XLOG_OVERWRITE_CONTRECORD)
9016 : {
9017 : /* nothing to do here, handled in xlogrecovery_redo() */
9018 : }
9019 114435 : else if (info == XLOG_END_OF_RECOVERY)
9020 : {
9021 : xl_end_of_recovery xlrec;
9022 : TimeLineID replayTLI;
9023 :
9024 12 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
9025 :
9026 : /*
9027 : * For Hot Standby, we could treat this like a Shutdown Checkpoint,
9028 : * but this case is rarer and harder to test, so the benefit doesn't
9029 : * outweigh the potential extra cost of maintenance.
9030 : */
9031 :
9032 : /*
9033 : * We should've already switched to the new TLI before replaying this
9034 : * record.
9035 : */
9036 12 : (void) GetCurrentReplayRecPtr(&replayTLI);
9037 12 : if (xlrec.ThisTimeLineID != replayTLI)
9038 0 : ereport(PANIC,
9039 : (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
9040 : xlrec.ThisTimeLineID, replayTLI)));
9041 : }
9042 114423 : else if (info == XLOG_NOOP)
9043 : {
9044 : /* nothing to do here */
9045 : }
9046 114423 : else if (info == XLOG_SWITCH)
9047 : {
9048 : /* nothing to do here */
9049 : }
9050 113956 : else if (info == XLOG_RESTORE_POINT)
9051 : {
9052 : /* nothing to do here, handled in xlogrecovery.c */
9053 : }
9054 113951 : else if (info == XLOG_ASSIGN_LSN)
9055 : {
9056 : /* nothing to do here, see XLogGetFakeLSN() */
9057 : }
9058 51624 : else if (info == XLOG_FPI || info == XLOG_FPI_FOR_HINT)
9059 : {
9060 : /*
9061 : * XLOG_FPI records contain nothing else but one or more block
9062 : * references. Every block reference must include a full-page image
9063 : * even if full_page_writes was disabled when the record was generated
9064 : * - otherwise there would be no point in this record.
9065 : *
9066 : * XLOG_FPI_FOR_HINT records are generated when a page needs to be
9067 : * WAL-logged because of a hint bit update. They are only generated
9068 : * when checksums and/or wal_log_hints are enabled. They may include
9069 : * no full-page images if full_page_writes was disabled when they were
9070 : * generated. In this case there is nothing to do here.
9071 : *
9072 : * No recovery conflicts are generated by these generic records - if a
9073 : * resource manager needs to generate conflicts, it has to define a
9074 : * separate WAL record type and redo routine.
9075 : */
9076 106400 : for (uint8 block_id = 0; block_id <= XLogRecMaxBlockId(record); block_id++)
9077 : {
9078 : Buffer buffer;
9079 :
9080 55644 : if (!XLogRecHasBlockImage(record, block_id))
9081 : {
9082 66 : if (info == XLOG_FPI)
9083 0 : elog(ERROR, "XLOG_FPI record did not contain a full-page image");
9084 66 : continue;
9085 : }
9086 :
9087 55578 : if (XLogReadBufferForRedo(record, block_id, &buffer) != BLK_RESTORED)
9088 0 : elog(ERROR, "unexpected XLogReadBufferForRedo result when restoring backup block");
9089 55578 : UnlockReleaseBuffer(buffer);
9090 : }
9091 : }
9092 868 : else if (info == XLOG_BACKUP_END)
9093 : {
9094 : /* nothing to do here, handled in xlogrecovery_redo() */
9095 : }
9096 766 : else if (info == XLOG_PARAMETER_CHANGE)
9097 : {
9098 : xl_parameter_change xlrec;
9099 :
9100 : /* Update our copy of the parameters in pg_control */
9101 38 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
9102 :
9103 38 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9104 38 : ControlFile->MaxConnections = xlrec.MaxConnections;
9105 38 : ControlFile->max_worker_processes = xlrec.max_worker_processes;
9106 38 : ControlFile->max_wal_senders = xlrec.max_wal_senders;
9107 38 : ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
9108 38 : ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
9109 38 : ControlFile->wal_level = xlrec.wal_level;
9110 38 : ControlFile->wal_log_hints = xlrec.wal_log_hints;
9111 :
9112 : /*
9113 : * Update minRecoveryPoint to ensure that if recovery is aborted, we
9114 : * recover back up to this point before allowing hot standby again.
9115 : * This is important if the max_* settings are decreased, to ensure
9116 : * you don't run queries against the WAL preceding the change. The
9117 : * local copies cannot be updated as long as crash recovery is
9118 : * happening and we expect all the WAL to be replayed.
9119 : */
9120 38 : if (InArchiveRecovery)
9121 : {
9122 23 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
9123 23 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
9124 : }
9125 38 : if (XLogRecPtrIsValid(LocalMinRecoveryPoint) && LocalMinRecoveryPoint < lsn)
9126 : {
9127 : TimeLineID replayTLI;
9128 :
9129 12 : (void) GetCurrentReplayRecPtr(&replayTLI);
9130 12 : ControlFile->minRecoveryPoint = lsn;
9131 12 : ControlFile->minRecoveryPointTLI = replayTLI;
9132 : }
9133 :
9134 38 : CommitTsParameterChange(xlrec.track_commit_timestamp,
9135 38 : ControlFile->track_commit_timestamp);
9136 38 : ControlFile->track_commit_timestamp = xlrec.track_commit_timestamp;
9137 :
9138 38 : UpdateControlFile();
9139 38 : LWLockRelease(ControlFileLock);
9140 :
9141 : /* Check to see if any parameter change gives a problem on recovery */
9142 38 : CheckRequiredParameterValues();
9143 : }
9144 728 : else if (info == XLOG_FPW_CHANGE)
9145 : {
9146 : bool fpw;
9147 :
9148 0 : memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
9149 :
9150 : /*
9151 : * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
9152 : * do_pg_backup_start() and do_pg_backup_stop() can check whether
9153 : * full_page_writes has been disabled during online backup.
9154 : */
9155 0 : if (!fpw)
9156 : {
9157 0 : SpinLockAcquire(&XLogCtl->info_lck);
9158 0 : if (XLogCtl->lastFpwDisableRecPtr < record->ReadRecPtr)
9159 0 : XLogCtl->lastFpwDisableRecPtr = record->ReadRecPtr;
9160 0 : SpinLockRelease(&XLogCtl->info_lck);
9161 : }
9162 :
9163 : /* Keep track of full_page_writes */
9164 0 : lastFullPageWrites = fpw;
9165 : }
9166 728 : else if (info == XLOG_CHECKPOINT_REDO)
9167 : {
9168 : xl_checkpoint_redo redo_rec;
9169 710 : bool new_state = false;
9170 :
9171 710 : memcpy(&redo_rec, XLogRecGetData(record), sizeof(xl_checkpoint_redo));
9172 :
9173 710 : SpinLockAcquire(&XLogCtl->info_lck);
9174 710 : XLogCtl->data_checksum_version = redo_rec.data_checksum_version;
9175 710 : SetLocalDataChecksumState(redo_rec.data_checksum_version);
9176 710 : if (redo_rec.data_checksum_version != ControlFile->data_checksum_version)
9177 0 : new_state = true;
9178 710 : SpinLockRelease(&XLogCtl->info_lck);
9179 :
9180 710 : if (new_state)
9181 0 : EmitAndWaitDataChecksumsBarrier(redo_rec.data_checksum_version);
9182 : }
9183 18 : else if (info == XLOG_LOGICAL_DECODING_STATUS_CHANGE)
9184 : {
9185 : bool status;
9186 :
9187 18 : memcpy(&status, XLogRecGetData(record), sizeof(bool));
9188 :
9189 : /*
9190 : * We need to toggle the logical decoding status and update the
9191 : * XLogLogicalInfo cache of processes synchronously because
9192 : * XLogLogicalInfoActive() is used even during read-only queries
9193 : * (e.g., via RelationIsAccessibleInLogicalDecoding()). In the
9194 : * 'disable' case, it is safe to invalidate existing slots after
9195 : * disabling logical decoding because logical decoding cannot process
9196 : * subsequent WAL records, which may not contain logical information.
9197 : */
9198 18 : if (status)
9199 9 : EnableLogicalDecoding();
9200 : else
9201 9 : DisableLogicalDecoding();
9202 :
9203 18 : elog(DEBUG1, "update logical decoding status to %d during recovery",
9204 : status);
9205 :
9206 18 : if (InRecovery && InHotStandby)
9207 : {
9208 16 : if (!status)
9209 : {
9210 : /*
9211 : * Invalidate logical slots if we are in hot standby and the
9212 : * primary disabled logical decoding.
9213 : */
9214 9 : InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_LEVEL,
9215 : 0, InvalidOid,
9216 : InvalidTransactionId);
9217 : }
9218 7 : else if (sync_replication_slots)
9219 : {
9220 : /*
9221 : * Signal the postmaster to launch the slotsync worker.
9222 : *
9223 : * XXX: For simplicity, we keep the slotsync worker running
9224 : * even after logical decoding is disabled. A future
9225 : * improvement can consider starting and stopping the worker
9226 : * based on logical decoding status change.
9227 : */
9228 0 : kill(PostmasterPid, SIGUSR1);
9229 : }
9230 : }
9231 : }
9232 115284 : }
9233 :
9234 : void
9235 6 : xlog2_redo(XLogReaderState *record)
9236 : {
9237 6 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
9238 :
9239 6 : if (info == XLOG2_CHECKSUMS)
9240 : {
9241 : xl_checksum_state state;
9242 :
9243 6 : memcpy(&state, XLogRecGetData(record), sizeof(xl_checksum_state));
9244 :
9245 6 : SpinLockAcquire(&XLogCtl->info_lck);
9246 6 : XLogCtl->data_checksum_version = state.new_checksum_state;
9247 6 : SpinLockRelease(&XLogCtl->info_lck);
9248 :
9249 6 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9250 6 : ControlFile->data_checksum_version = state.new_checksum_state;
9251 6 : UpdateControlFile();
9252 6 : LWLockRelease(ControlFileLock);
9253 :
9254 : /*
9255 : * Block on a procsignalbarrier to await all processes having seen the
9256 : * change to checksum status. Once the barrier has been passed we can
9257 : * initiate the corresponding processing.
9258 : */
9259 6 : EmitAndWaitDataChecksumsBarrier(state.new_checksum_state);
9260 : }
9261 6 : }
9262 :
9263 : /*
9264 : * Return the extra open flags used for opening a file, depending on the
9265 : * value of the GUCs wal_sync_method, fsync and debug_io_direct.
9266 : */
9267 : static int
9268 17814 : get_sync_bit(int method)
9269 : {
9270 17814 : int o_direct_flag = 0;
9271 :
9272 : /*
9273 : * Use O_DIRECT if requested, except in walreceiver process. The WAL
9274 : * written by walreceiver is normally read by the startup process soon
9275 : * after it's written. Also, walreceiver performs unaligned writes, which
9276 : * don't work with O_DIRECT, so it is required for correctness too.
9277 : */
9278 17814 : if ((io_direct_flags & IO_DIRECT_WAL) && !AmWalReceiverProcess())
9279 8 : o_direct_flag = PG_O_DIRECT;
9280 :
9281 : /* If fsync is disabled, never open in sync mode */
9282 17814 : if (!enableFsync)
9283 17814 : return o_direct_flag;
9284 :
9285 0 : switch (method)
9286 : {
9287 : /*
9288 : * enum values for all sync options are defined even if they are
9289 : * not supported on the current platform. But if not, they are
9290 : * not included in the enum option array, and therefore will never
9291 : * be seen here.
9292 : */
9293 0 : case WAL_SYNC_METHOD_FSYNC:
9294 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
9295 : case WAL_SYNC_METHOD_FDATASYNC:
9296 0 : return o_direct_flag;
9297 : #ifdef O_SYNC
9298 0 : case WAL_SYNC_METHOD_OPEN:
9299 0 : return O_SYNC | o_direct_flag;
9300 : #endif
9301 : #ifdef O_DSYNC
9302 0 : case WAL_SYNC_METHOD_OPEN_DSYNC:
9303 0 : return O_DSYNC | o_direct_flag;
9304 : #endif
9305 0 : default:
9306 : /* can't happen (unless we are out of sync with option array) */
9307 0 : elog(ERROR, "unrecognized \"wal_sync_method\": %d", method);
9308 : return 0; /* silence warning */
9309 : }
9310 : }
9311 :
9312 : /*
9313 : * GUC support
9314 : */
9315 : void
9316 1289 : assign_wal_sync_method(int new_wal_sync_method, void *extra)
9317 : {
9318 1289 : if (wal_sync_method != new_wal_sync_method)
9319 : {
9320 : /*
9321 : * To ensure that no blocks escape unsynced, force an fsync on the
9322 : * currently open log segment (if any). Also, if the open flag is
9323 : * changing, close the log file so it will be reopened (with new flag
9324 : * bit) at next use.
9325 : */
9326 0 : if (openLogFile >= 0)
9327 : {
9328 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN);
9329 0 : if (pg_fsync(openLogFile) != 0)
9330 : {
9331 : char xlogfname[MAXFNAMELEN];
9332 : int save_errno;
9333 :
9334 0 : save_errno = errno;
9335 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo,
9336 : wal_segment_size);
9337 0 : errno = save_errno;
9338 0 : ereport(PANIC,
9339 : (errcode_for_file_access(),
9340 : errmsg("could not fsync file \"%s\": %m", xlogfname)));
9341 : }
9342 :
9343 0 : pgstat_report_wait_end();
9344 0 : if (get_sync_bit(wal_sync_method) != get_sync_bit(new_wal_sync_method))
9345 0 : XLogFileClose();
9346 : }
9347 : }
9348 1289 : }
9349 :
9350 :
9351 : /*
9352 : * Issue appropriate kind of fsync (if any) for an XLOG output file.
9353 : *
9354 : * 'fd' is a file descriptor for the XLOG file to be fsync'd.
9355 : * 'segno' is for error reporting purposes.
9356 : */
9357 : void
9358 211343 : issue_xlog_fsync(int fd, XLogSegNo segno, TimeLineID tli)
9359 : {
9360 211343 : char *msg = NULL;
9361 : instr_time start;
9362 :
9363 : Assert(tli != 0);
9364 :
9365 : /*
9366 : * Quick exit if fsync is disabled or write() has already synced the WAL
9367 : * file.
9368 : */
9369 211343 : if (!enableFsync ||
9370 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN ||
9371 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN_DSYNC)
9372 211343 : return;
9373 :
9374 : /*
9375 : * Measure I/O timing to sync the WAL file for pg_stat_io.
9376 : */
9377 0 : start = pgstat_prepare_io_time(track_wal_io_timing);
9378 :
9379 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC);
9380 0 : switch (wal_sync_method)
9381 : {
9382 0 : case WAL_SYNC_METHOD_FSYNC:
9383 0 : if (pg_fsync_no_writethrough(fd) != 0)
9384 0 : msg = _("could not fsync file \"%s\": %m");
9385 0 : break;
9386 : #ifdef HAVE_FSYNC_WRITETHROUGH
9387 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
9388 : if (pg_fsync_writethrough(fd) != 0)
9389 : msg = _("could not fsync write-through file \"%s\": %m");
9390 : break;
9391 : #endif
9392 0 : case WAL_SYNC_METHOD_FDATASYNC:
9393 0 : if (pg_fdatasync(fd) != 0)
9394 0 : msg = _("could not fdatasync file \"%s\": %m");
9395 0 : break;
9396 0 : case WAL_SYNC_METHOD_OPEN:
9397 : case WAL_SYNC_METHOD_OPEN_DSYNC:
9398 : /* not reachable */
9399 : Assert(false);
9400 0 : break;
9401 0 : default:
9402 0 : ereport(PANIC,
9403 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
9404 : errmsg_internal("unrecognized \"wal_sync_method\": %d", wal_sync_method));
9405 : break;
9406 : }
9407 :
9408 : /* PANIC if failed to fsync */
9409 0 : if (msg)
9410 : {
9411 : char xlogfname[MAXFNAMELEN];
9412 0 : int save_errno = errno;
9413 :
9414 0 : XLogFileName(xlogfname, tli, segno, wal_segment_size);
9415 0 : errno = save_errno;
9416 0 : ereport(PANIC,
9417 : (errcode_for_file_access(),
9418 : errmsg(msg, xlogfname)));
9419 : }
9420 :
9421 0 : pgstat_report_wait_end();
9422 :
9423 0 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL, IOOP_FSYNC,
9424 : start, 1, 0);
9425 : }
9426 :
9427 : /*
9428 : * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
9429 : * function. It creates the necessary starting checkpoint and constructs the
9430 : * backup state and tablespace map.
9431 : *
9432 : * Input parameters are "state" (the backup state), "fast" (if true, we do
9433 : * the checkpoint in fast mode), and "tablespaces" (if non-NULL, indicates a
9434 : * list of tablespaceinfo structs describing the cluster's tablespaces.).
9435 : *
9436 : * The tablespace map contents are appended to passed-in parameter
9437 : * tablespace_map and the caller is responsible for including it in the backup
9438 : * archive as 'tablespace_map'. The tablespace_map file is required mainly for
9439 : * tar format in windows as native windows utilities are not able to create
9440 : * symlinks while extracting files from tar. However for consistency and
9441 : * platform-independence, we do it the same way everywhere.
9442 : *
9443 : * It fills in "state" with the information required for the backup, such
9444 : * as the minimum WAL location that must be present to restore from this
9445 : * backup (starttli) and the corresponding timeline ID (starttli).
9446 : *
9447 : * Every successfully started backup must be stopped by calling
9448 : * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
9449 : * backups active at the same time.
9450 : *
9451 : * It is the responsibility of the caller of this function to verify the
9452 : * permissions of the calling user!
9453 : */
9454 : void
9455 180 : do_pg_backup_start(const char *backupidstr, bool fast, List **tablespaces,
9456 : BackupState *state, StringInfo tblspcmapfile)
9457 : {
9458 : bool backup_started_in_recovery;
9459 :
9460 : Assert(state != NULL);
9461 180 : backup_started_in_recovery = RecoveryInProgress();
9462 :
9463 : /*
9464 : * During recovery, we don't need to check WAL level. Because, if WAL
9465 : * level is not sufficient, it's impossible to get here during recovery.
9466 : */
9467 180 : if (!backup_started_in_recovery && !XLogIsNeeded())
9468 0 : ereport(ERROR,
9469 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9470 : errmsg("WAL level not sufficient for making an online backup"),
9471 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9472 :
9473 180 : if (strlen(backupidstr) > MAXPGPATH)
9474 1 : ereport(ERROR,
9475 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
9476 : errmsg("backup label too long (max %d bytes)",
9477 : MAXPGPATH)));
9478 :
9479 179 : strlcpy(state->name, backupidstr, sizeof(state->name));
9480 :
9481 : /*
9482 : * Mark backup active in shared memory. We must do full-page WAL writes
9483 : * during an on-line backup even if not doing so at other times, because
9484 : * it's quite possible for the backup dump to obtain a "torn" (partially
9485 : * written) copy of a database page if it reads the page concurrently with
9486 : * our write to the same page. This can be fixed as long as the first
9487 : * write to the page in the WAL sequence is a full-page write. Hence, we
9488 : * increment runningBackups then force a CHECKPOINT, to ensure there are
9489 : * no dirty pages in shared memory that might get dumped while the backup
9490 : * is in progress without having a corresponding WAL record. (Once the
9491 : * backup is complete, we need not force full-page writes anymore, since
9492 : * we expect that any pages not modified during the backup interval must
9493 : * have been correctly captured by the backup.)
9494 : *
9495 : * Note that forcing full-page writes has no effect during an online
9496 : * backup from the standby.
9497 : *
9498 : * We must hold all the insertion locks to change the value of
9499 : * runningBackups, to ensure adequate interlocking against
9500 : * XLogInsertRecord().
9501 : */
9502 179 : WALInsertLockAcquireExclusive();
9503 179 : XLogCtl->Insert.runningBackups++;
9504 179 : WALInsertLockRelease();
9505 :
9506 : /*
9507 : * Ensure we decrement runningBackups if we fail below. NB -- for this to
9508 : * work correctly, it is critical that sessionBackupState is only updated
9509 : * after this block is over.
9510 : */
9511 179 : PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9512 : {
9513 179 : bool gotUniqueStartpoint = false;
9514 : DIR *tblspcdir;
9515 : struct dirent *de;
9516 : tablespaceinfo *ti;
9517 : int datadirpathlen;
9518 :
9519 : /*
9520 : * Force an XLOG file switch before the checkpoint, to ensure that the
9521 : * WAL segment the checkpoint is written to doesn't contain pages with
9522 : * old timeline IDs. That would otherwise happen if you called
9523 : * pg_backup_start() right after restoring from a PITR archive: the
9524 : * first WAL segment containing the startup checkpoint has pages in
9525 : * the beginning with the old timeline ID. That can cause trouble at
9526 : * recovery: we won't have a history file covering the old timeline if
9527 : * pg_wal directory was not included in the base backup and the WAL
9528 : * archive was cleared too before starting the backup.
9529 : *
9530 : * During recovery, we skip forcing XLOG file switch, which means that
9531 : * the backup taken during recovery is not available for the special
9532 : * recovery case described above.
9533 : */
9534 179 : if (!backup_started_in_recovery)
9535 170 : RequestXLogSwitch(false);
9536 :
9537 : do
9538 : {
9539 : bool checkpointfpw;
9540 :
9541 : /*
9542 : * Force a CHECKPOINT. Aside from being necessary to prevent torn
9543 : * page problems, this guarantees that two successive backup runs
9544 : * will have different checkpoint positions and hence different
9545 : * history file names, even if nothing happened in between.
9546 : *
9547 : * During recovery, establish a restartpoint if possible. We use
9548 : * the last restartpoint as the backup starting checkpoint. This
9549 : * means that two successive backup runs can have same checkpoint
9550 : * positions.
9551 : *
9552 : * Since the fact that we are executing do_pg_backup_start()
9553 : * during recovery means that checkpointer is running, we can use
9554 : * RequestCheckpoint() to establish a restartpoint.
9555 : *
9556 : * We use CHECKPOINT_FAST only if requested by user (via passing
9557 : * fast = true). Otherwise this can take awhile.
9558 : */
9559 179 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
9560 : (fast ? CHECKPOINT_FAST : 0));
9561 :
9562 : /*
9563 : * Now we need to fetch the checkpoint record location, and also
9564 : * its REDO pointer. The oldest point in WAL that would be needed
9565 : * to restore starting from the checkpoint is precisely the REDO
9566 : * pointer.
9567 : */
9568 179 : LWLockAcquire(ControlFileLock, LW_SHARED);
9569 179 : state->checkpointloc = ControlFile->checkPoint;
9570 179 : state->startpoint = ControlFile->checkPointCopy.redo;
9571 179 : state->starttli = ControlFile->checkPointCopy.ThisTimeLineID;
9572 179 : checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
9573 179 : LWLockRelease(ControlFileLock);
9574 :
9575 179 : if (backup_started_in_recovery)
9576 : {
9577 : XLogRecPtr recptr;
9578 :
9579 : /*
9580 : * Check to see if all WAL replayed during online backup
9581 : * (i.e., since last restartpoint used as backup starting
9582 : * checkpoint) contain full-page writes.
9583 : */
9584 9 : SpinLockAcquire(&XLogCtl->info_lck);
9585 9 : recptr = XLogCtl->lastFpwDisableRecPtr;
9586 9 : SpinLockRelease(&XLogCtl->info_lck);
9587 :
9588 9 : if (!checkpointfpw || state->startpoint <= recptr)
9589 0 : ereport(ERROR,
9590 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9591 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9592 : "since last restartpoint"),
9593 : errhint("This means that the backup being taken on the standby "
9594 : "is corrupt and should not be used. "
9595 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9596 : "and then try an online backup again.")));
9597 :
9598 : /*
9599 : * During recovery, since we don't use the end-of-backup WAL
9600 : * record and don't write the backup history file, the
9601 : * starting WAL location doesn't need to be unique. This means
9602 : * that two base backups started at the same time might use
9603 : * the same checkpoint as starting locations.
9604 : */
9605 9 : gotUniqueStartpoint = true;
9606 : }
9607 :
9608 : /*
9609 : * If two base backups are started at the same time (in WAL sender
9610 : * processes), we need to make sure that they use different
9611 : * checkpoints as starting locations, because we use the starting
9612 : * WAL location as a unique identifier for the base backup in the
9613 : * end-of-backup WAL record and when we write the backup history
9614 : * file. Perhaps it would be better generate a separate unique ID
9615 : * for each backup instead of forcing another checkpoint, but
9616 : * taking a checkpoint right after another is not that expensive
9617 : * either because only few buffers have been dirtied yet.
9618 : */
9619 179 : WALInsertLockAcquireExclusive();
9620 179 : if (XLogCtl->Insert.lastBackupStart < state->startpoint)
9621 : {
9622 179 : XLogCtl->Insert.lastBackupStart = state->startpoint;
9623 179 : gotUniqueStartpoint = true;
9624 : }
9625 179 : WALInsertLockRelease();
9626 179 : } while (!gotUniqueStartpoint);
9627 :
9628 : /*
9629 : * Construct tablespace_map file.
9630 : */
9631 179 : datadirpathlen = strlen(DataDir);
9632 :
9633 : /* Collect information about all tablespaces */
9634 179 : tblspcdir = AllocateDir(PG_TBLSPC_DIR);
9635 574 : while ((de = ReadDir(tblspcdir, PG_TBLSPC_DIR)) != NULL)
9636 : {
9637 : char fullpath[MAXPGPATH + sizeof(PG_TBLSPC_DIR)];
9638 : char linkpath[MAXPGPATH];
9639 395 : char *relpath = NULL;
9640 : char *s;
9641 : PGFileType de_type;
9642 : char *badp;
9643 : Oid tsoid;
9644 :
9645 : /*
9646 : * Try to parse the directory name as an unsigned integer.
9647 : *
9648 : * Tablespace directories should be positive integers that can be
9649 : * represented in 32 bits, with no leading zeroes or trailing
9650 : * garbage. If we come across a name that doesn't meet those
9651 : * criteria, skip it.
9652 : */
9653 395 : if (de->d_name[0] < '1' || de->d_name[1] > '9')
9654 358 : continue;
9655 37 : errno = 0;
9656 37 : tsoid = strtoul(de->d_name, &badp, 10);
9657 37 : if (*badp != '\0' || errno == EINVAL || errno == ERANGE)
9658 0 : continue;
9659 :
9660 37 : snprintf(fullpath, sizeof(fullpath), "%s/%s", PG_TBLSPC_DIR, de->d_name);
9661 :
9662 37 : de_type = get_dirent_type(fullpath, de, false, ERROR);
9663 :
9664 37 : if (de_type == PGFILETYPE_LNK)
9665 : {
9666 : StringInfoData escapedpath;
9667 : int rllen;
9668 :
9669 23 : rllen = readlink(fullpath, linkpath, sizeof(linkpath));
9670 23 : if (rllen < 0)
9671 : {
9672 0 : ereport(WARNING,
9673 : (errmsg("could not read symbolic link \"%s\": %m",
9674 : fullpath)));
9675 0 : continue;
9676 : }
9677 23 : else if (rllen >= sizeof(linkpath))
9678 : {
9679 0 : ereport(WARNING,
9680 : (errmsg("symbolic link \"%s\" target is too long",
9681 : fullpath)));
9682 0 : continue;
9683 : }
9684 23 : linkpath[rllen] = '\0';
9685 :
9686 : /*
9687 : * Relpath holds the relative path of the tablespace directory
9688 : * when it's located within PGDATA, or NULL if it's located
9689 : * elsewhere.
9690 : */
9691 23 : if (rllen > datadirpathlen &&
9692 1 : strncmp(linkpath, DataDir, datadirpathlen) == 0 &&
9693 0 : IS_DIR_SEP(linkpath[datadirpathlen]))
9694 0 : relpath = pstrdup(linkpath + datadirpathlen + 1);
9695 :
9696 : /*
9697 : * Add a backslash-escaped version of the link path to the
9698 : * tablespace map file.
9699 : */
9700 23 : initStringInfo(&escapedpath);
9701 562 : for (s = linkpath; *s; s++)
9702 : {
9703 539 : if (*s == '\n' || *s == '\r' || *s == '\\')
9704 0 : appendStringInfoChar(&escapedpath, '\\');
9705 539 : appendStringInfoChar(&escapedpath, *s);
9706 : }
9707 23 : appendStringInfo(tblspcmapfile, "%s %s\n",
9708 23 : de->d_name, escapedpath.data);
9709 23 : pfree(escapedpath.data);
9710 : }
9711 14 : else if (de_type == PGFILETYPE_DIR)
9712 : {
9713 : /*
9714 : * It's possible to use allow_in_place_tablespaces to create
9715 : * directories directly under pg_tblspc, for testing purposes
9716 : * only.
9717 : *
9718 : * In this case, we store a relative path rather than an
9719 : * absolute path into the tablespaceinfo.
9720 : */
9721 14 : snprintf(linkpath, sizeof(linkpath), "%s/%s",
9722 14 : PG_TBLSPC_DIR, de->d_name);
9723 14 : relpath = pstrdup(linkpath);
9724 : }
9725 : else
9726 : {
9727 : /* Skip any other file type that appears here. */
9728 0 : continue;
9729 : }
9730 :
9731 37 : ti = palloc_object(tablespaceinfo);
9732 37 : ti->oid = tsoid;
9733 37 : ti->path = pstrdup(linkpath);
9734 37 : ti->rpath = relpath;
9735 37 : ti->size = -1;
9736 :
9737 37 : if (tablespaces)
9738 37 : *tablespaces = lappend(*tablespaces, ti);
9739 : }
9740 179 : FreeDir(tblspcdir);
9741 :
9742 179 : state->starttime = (pg_time_t) time(NULL);
9743 : }
9744 179 : PG_END_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9745 :
9746 179 : state->started_in_recovery = backup_started_in_recovery;
9747 :
9748 : /*
9749 : * Mark that the start phase has correctly finished for the backup.
9750 : */
9751 179 : sessionBackupState = SESSION_BACKUP_RUNNING;
9752 179 : }
9753 :
9754 : /*
9755 : * Utility routine to fetch the session-level status of a backup running.
9756 : */
9757 : SessionBackupState
9758 200 : get_backup_status(void)
9759 : {
9760 200 : return sessionBackupState;
9761 : }
9762 :
9763 : /*
9764 : * do_pg_backup_stop
9765 : *
9766 : * Utility function called at the end of an online backup. It creates history
9767 : * file (if required), resets sessionBackupState and so on. It can optionally
9768 : * wait for WAL segments to be archived.
9769 : *
9770 : * "state" is filled with the information necessary to restore from this
9771 : * backup with its stop LSN (stoppoint), its timeline ID (stoptli), etc.
9772 : *
9773 : * It is the responsibility of the caller of this function to verify the
9774 : * permissions of the calling user!
9775 : */
9776 : void
9777 173 : do_pg_backup_stop(BackupState *state, bool waitforarchive)
9778 : {
9779 173 : bool backup_stopped_in_recovery = false;
9780 : char histfilepath[MAXPGPATH];
9781 : char lastxlogfilename[MAXFNAMELEN];
9782 : char histfilename[MAXFNAMELEN];
9783 : XLogSegNo _logSegNo;
9784 : FILE *fp;
9785 : int seconds_before_warning;
9786 173 : int waits = 0;
9787 173 : bool reported_waiting = false;
9788 :
9789 : Assert(state != NULL);
9790 :
9791 173 : backup_stopped_in_recovery = RecoveryInProgress();
9792 :
9793 : /*
9794 : * During recovery, we don't need to check WAL level. Because, if WAL
9795 : * level is not sufficient, it's impossible to get here during recovery.
9796 : */
9797 173 : if (!backup_stopped_in_recovery && !XLogIsNeeded())
9798 0 : ereport(ERROR,
9799 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9800 : errmsg("WAL level not sufficient for making an online backup"),
9801 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9802 :
9803 : /*
9804 : * OK to update backup counter and session-level lock.
9805 : *
9806 : * Note that CHECK_FOR_INTERRUPTS() must not occur while updating them,
9807 : * otherwise they can be updated inconsistently, which might cause
9808 : * do_pg_abort_backup() to fail.
9809 : */
9810 173 : WALInsertLockAcquireExclusive();
9811 :
9812 : /*
9813 : * It is expected that each do_pg_backup_start() call is matched by
9814 : * exactly one do_pg_backup_stop() call.
9815 : */
9816 : Assert(XLogCtl->Insert.runningBackups > 0);
9817 173 : XLogCtl->Insert.runningBackups--;
9818 :
9819 : /*
9820 : * Clean up session-level lock.
9821 : *
9822 : * You might think that WALInsertLockRelease() can be called before
9823 : * cleaning up session-level lock because session-level lock doesn't need
9824 : * to be protected with WAL insertion lock. But since
9825 : * CHECK_FOR_INTERRUPTS() can occur in it, session-level lock must be
9826 : * cleaned up before it.
9827 : */
9828 173 : sessionBackupState = SESSION_BACKUP_NONE;
9829 :
9830 173 : WALInsertLockRelease();
9831 :
9832 : /*
9833 : * If we are taking an online backup from the standby, we confirm that the
9834 : * standby has not been promoted during the backup.
9835 : */
9836 173 : if (state->started_in_recovery && !backup_stopped_in_recovery)
9837 0 : ereport(ERROR,
9838 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9839 : errmsg("the standby was promoted during online backup"),
9840 : errhint("This means that the backup being taken is corrupt "
9841 : "and should not be used. "
9842 : "Try taking another online backup.")));
9843 :
9844 : /*
9845 : * During recovery, we don't write an end-of-backup record. We assume that
9846 : * pg_control was backed up last and its minimum recovery point can be
9847 : * available as the backup end location. Since we don't have an
9848 : * end-of-backup record, we use the pg_control value to check whether
9849 : * we've reached the end of backup when starting recovery from this
9850 : * backup. We have no way of checking if pg_control wasn't backed up last
9851 : * however.
9852 : *
9853 : * We don't force a switch to new WAL file but it is still possible to
9854 : * wait for all the required files to be archived if waitforarchive is
9855 : * true. This is okay if we use the backup to start a standby and fetch
9856 : * the missing WAL using streaming replication. But in the case of an
9857 : * archive recovery, a user should set waitforarchive to true and wait for
9858 : * them to be archived to ensure that all the required files are
9859 : * available.
9860 : *
9861 : * We return the current minimum recovery point as the backup end
9862 : * location. Note that it can be greater than the exact backup end
9863 : * location if the minimum recovery point is updated after the backup of
9864 : * pg_control. This is harmless for current uses.
9865 : *
9866 : * XXX currently a backup history file is for informational and debug
9867 : * purposes only. It's not essential for an online backup. Furthermore,
9868 : * even if it's created, it will not be archived during recovery because
9869 : * an archiver is not invoked. So it doesn't seem worthwhile to write a
9870 : * backup history file during recovery.
9871 : */
9872 173 : if (backup_stopped_in_recovery)
9873 : {
9874 : XLogRecPtr recptr;
9875 :
9876 : /*
9877 : * Check to see if all WAL replayed during online backup contain
9878 : * full-page writes.
9879 : */
9880 9 : SpinLockAcquire(&XLogCtl->info_lck);
9881 9 : recptr = XLogCtl->lastFpwDisableRecPtr;
9882 9 : SpinLockRelease(&XLogCtl->info_lck);
9883 :
9884 9 : if (state->startpoint <= recptr)
9885 0 : ereport(ERROR,
9886 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9887 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9888 : "during online backup"),
9889 : errhint("This means that the backup being taken on the standby "
9890 : "is corrupt and should not be used. "
9891 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9892 : "and then try an online backup again.")));
9893 :
9894 :
9895 9 : LWLockAcquire(ControlFileLock, LW_SHARED);
9896 9 : state->stoppoint = ControlFile->minRecoveryPoint;
9897 9 : state->stoptli = ControlFile->minRecoveryPointTLI;
9898 9 : LWLockRelease(ControlFileLock);
9899 : }
9900 : else
9901 : {
9902 : char *history_file;
9903 :
9904 : /*
9905 : * Write the backup-end xlog record
9906 : */
9907 164 : XLogBeginInsert();
9908 164 : XLogRegisterData(&state->startpoint,
9909 : sizeof(state->startpoint));
9910 164 : state->stoppoint = XLogInsert(RM_XLOG_ID, XLOG_BACKUP_END);
9911 :
9912 : /*
9913 : * Given that we're not in recovery, InsertTimeLineID is set and can't
9914 : * change, so we can read it without a lock.
9915 : */
9916 164 : state->stoptli = XLogCtl->InsertTimeLineID;
9917 :
9918 : /*
9919 : * Force a switch to a new xlog segment file, so that the backup is
9920 : * valid as soon as archiver moves out the current segment file.
9921 : */
9922 164 : RequestXLogSwitch(false);
9923 :
9924 164 : state->stoptime = (pg_time_t) time(NULL);
9925 :
9926 : /*
9927 : * Write the backup history file
9928 : */
9929 164 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9930 164 : BackupHistoryFilePath(histfilepath, state->stoptli, _logSegNo,
9931 : state->startpoint, wal_segment_size);
9932 164 : fp = AllocateFile(histfilepath, "w");
9933 164 : if (!fp)
9934 0 : ereport(ERROR,
9935 : (errcode_for_file_access(),
9936 : errmsg("could not create file \"%s\": %m",
9937 : histfilepath)));
9938 :
9939 : /* Build and save the contents of the backup history file */
9940 164 : history_file = build_backup_content(state, true);
9941 164 : fprintf(fp, "%s", history_file);
9942 164 : pfree(history_file);
9943 :
9944 164 : if (fflush(fp) || ferror(fp) || FreeFile(fp))
9945 0 : ereport(ERROR,
9946 : (errcode_for_file_access(),
9947 : errmsg("could not write file \"%s\": %m",
9948 : histfilepath)));
9949 :
9950 : /*
9951 : * Clean out any no-longer-needed history files. As a side effect,
9952 : * this will post a .ready file for the newly created history file,
9953 : * notifying the archiver that history file may be archived
9954 : * immediately.
9955 : */
9956 164 : CleanupBackupHistory();
9957 : }
9958 :
9959 : /*
9960 : * If archiving is enabled, wait for all the required WAL files to be
9961 : * archived before returning. If archiving isn't enabled, the required WAL
9962 : * needs to be transported via streaming replication (hopefully with
9963 : * wal_keep_size set high enough), or some more exotic mechanism like
9964 : * polling and copying files from pg_wal with script. We have no knowledge
9965 : * of those mechanisms, so it's up to the user to ensure that he gets all
9966 : * the required WAL.
9967 : *
9968 : * We wait until both the last WAL file filled during backup and the
9969 : * history file have been archived, and assume that the alphabetic sorting
9970 : * property of the WAL files ensures any earlier WAL files are safely
9971 : * archived as well.
9972 : *
9973 : * We wait forever, since archive_command is supposed to work and we
9974 : * assume the admin wanted his backup to work completely. If you don't
9975 : * wish to wait, then either waitforarchive should be passed in as false,
9976 : * or you can set statement_timeout. Also, some notices are issued to
9977 : * clue in anyone who might be doing this interactively.
9978 : */
9979 :
9980 173 : if (waitforarchive &&
9981 10 : ((!backup_stopped_in_recovery && XLogArchivingActive()) ||
9982 1 : (backup_stopped_in_recovery && XLogArchivingAlways())))
9983 : {
9984 4 : XLByteToPrevSeg(state->stoppoint, _logSegNo, wal_segment_size);
9985 4 : XLogFileName(lastxlogfilename, state->stoptli, _logSegNo,
9986 : wal_segment_size);
9987 :
9988 4 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9989 4 : BackupHistoryFileName(histfilename, state->stoptli, _logSegNo,
9990 : state->startpoint, wal_segment_size);
9991 :
9992 4 : seconds_before_warning = 60;
9993 4 : waits = 0;
9994 :
9995 12 : while (XLogArchiveIsBusy(lastxlogfilename) ||
9996 4 : XLogArchiveIsBusy(histfilename))
9997 : {
9998 4 : CHECK_FOR_INTERRUPTS();
9999 :
10000 4 : if (!reported_waiting && waits > 5)
10001 : {
10002 0 : ereport(NOTICE,
10003 : (errmsg("base backup done, waiting for required WAL segments to be archived")));
10004 0 : reported_waiting = true;
10005 : }
10006 :
10007 4 : (void) WaitLatch(MyLatch,
10008 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
10009 : 1000L,
10010 : WAIT_EVENT_BACKUP_WAIT_WAL_ARCHIVE);
10011 4 : ResetLatch(MyLatch);
10012 :
10013 4 : if (++waits >= seconds_before_warning)
10014 : {
10015 0 : seconds_before_warning *= 2; /* This wraps in >10 years... */
10016 0 : ereport(WARNING,
10017 : (errmsg("still waiting for all required WAL segments to be archived (%d seconds elapsed)",
10018 : waits),
10019 : errhint("Check that your \"archive_command\" is executing properly. "
10020 : "You can safely cancel this backup, "
10021 : "but the database backup will not be usable without all the WAL segments.")));
10022 : }
10023 : }
10024 :
10025 4 : ereport(NOTICE,
10026 : (errmsg("all required WAL segments have been archived")));
10027 : }
10028 169 : else if (waitforarchive)
10029 6 : ereport(NOTICE,
10030 : (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
10031 173 : }
10032 :
10033 :
10034 : /*
10035 : * do_pg_abort_backup: abort a running backup
10036 : *
10037 : * This does just the most basic steps of do_pg_backup_stop(), by taking the
10038 : * system out of backup mode, thus making it a lot more safe to call from
10039 : * an error handler.
10040 : *
10041 : * 'arg' indicates that it's being called during backup setup; so
10042 : * sessionBackupState has not been modified yet, but runningBackups has
10043 : * already been incremented. When it's false, then it's invoked as a
10044 : * before_shmem_exit handler, and therefore we must not change state
10045 : * unless sessionBackupState indicates that a backup is actually running.
10046 : *
10047 : * NB: This gets used as a PG_ENSURE_ERROR_CLEANUP callback and
10048 : * before_shmem_exit handler, hence the odd-looking signature.
10049 : */
10050 : void
10051 8 : do_pg_abort_backup(int code, Datum arg)
10052 : {
10053 8 : bool during_backup_start = DatumGetBool(arg);
10054 :
10055 : /* If called during backup start, there shouldn't be one already running */
10056 : Assert(!during_backup_start || sessionBackupState == SESSION_BACKUP_NONE);
10057 :
10058 8 : if (during_backup_start || sessionBackupState != SESSION_BACKUP_NONE)
10059 : {
10060 6 : WALInsertLockAcquireExclusive();
10061 : Assert(XLogCtl->Insert.runningBackups > 0);
10062 6 : XLogCtl->Insert.runningBackups--;
10063 :
10064 6 : sessionBackupState = SESSION_BACKUP_NONE;
10065 6 : WALInsertLockRelease();
10066 :
10067 6 : if (!during_backup_start)
10068 6 : ereport(WARNING,
10069 : errmsg("aborting backup due to backend exiting before pg_backup_stop was called"));
10070 : }
10071 8 : }
10072 :
10073 : /*
10074 : * Register a handler that will warn about unterminated backups at end of
10075 : * session, unless this has already been done.
10076 : */
10077 : void
10078 4 : register_persistent_abort_backup_handler(void)
10079 : {
10080 : static bool already_done = false;
10081 :
10082 4 : if (already_done)
10083 1 : return;
10084 3 : before_shmem_exit(do_pg_abort_backup, BoolGetDatum(false));
10085 3 : already_done = true;
10086 : }
10087 :
10088 : /*
10089 : * Get latest WAL insert pointer
10090 : */
10091 : XLogRecPtr
10092 2155 : GetXLogInsertRecPtr(void)
10093 : {
10094 2155 : XLogCtlInsert *Insert = &XLogCtl->Insert;
10095 : uint64 current_bytepos;
10096 :
10097 2155 : SpinLockAcquire(&Insert->insertpos_lck);
10098 2155 : current_bytepos = Insert->CurrBytePos;
10099 2155 : SpinLockRelease(&Insert->insertpos_lck);
10100 :
10101 2155 : return XLogBytePosToRecPtr(current_bytepos);
10102 : }
10103 :
10104 : /*
10105 : * Get latest WAL record end pointer
10106 : */
10107 : XLogRecPtr
10108 3351 : GetXLogInsertEndRecPtr(void)
10109 : {
10110 3351 : XLogCtlInsert *Insert = &XLogCtl->Insert;
10111 : uint64 current_bytepos;
10112 :
10113 3351 : SpinLockAcquire(&Insert->insertpos_lck);
10114 3351 : current_bytepos = Insert->CurrBytePos;
10115 3351 : SpinLockRelease(&Insert->insertpos_lck);
10116 :
10117 3351 : return XLogBytePosToEndRecPtr(current_bytepos);
10118 : }
10119 :
10120 : /*
10121 : * Get latest WAL write pointer
10122 : */
10123 : XLogRecPtr
10124 8335 : GetXLogWriteRecPtr(void)
10125 : {
10126 8335 : RefreshXLogWriteResult(LogwrtResult);
10127 :
10128 8335 : return LogwrtResult.Write;
10129 : }
10130 :
10131 : /*
10132 : * Returns the redo pointer of the last checkpoint or restartpoint. This is
10133 : * the oldest point in WAL that we still need, if we have to restart recovery.
10134 : */
10135 : void
10136 401 : GetOldestRestartPoint(XLogRecPtr *oldrecptr, TimeLineID *oldtli)
10137 : {
10138 401 : LWLockAcquire(ControlFileLock, LW_SHARED);
10139 401 : *oldrecptr = ControlFile->checkPointCopy.redo;
10140 401 : *oldtli = ControlFile->checkPointCopy.ThisTimeLineID;
10141 401 : LWLockRelease(ControlFileLock);
10142 401 : }
10143 :
10144 : /* Thin wrapper around ShutdownWalRcv(). */
10145 : void
10146 1080 : XLogShutdownWalRcv(void)
10147 : {
10148 : Assert(AmStartupProcess() || !IsUnderPostmaster);
10149 :
10150 1080 : ShutdownWalRcv();
10151 1080 : ResetInstallXLogFileSegmentActive();
10152 1080 : }
10153 :
10154 : /* Enable WAL file recycling and preallocation. */
10155 : void
10156 1280 : SetInstallXLogFileSegmentActive(void)
10157 : {
10158 1280 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
10159 1280 : XLogCtl->InstallXLogFileSegmentActive = true;
10160 1280 : LWLockRelease(ControlFileLock);
10161 1280 : }
10162 :
10163 : /* Disable WAL file recycling and preallocation. */
10164 : void
10165 1258 : ResetInstallXLogFileSegmentActive(void)
10166 : {
10167 1258 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
10168 1258 : XLogCtl->InstallXLogFileSegmentActive = false;
10169 1258 : LWLockRelease(ControlFileLock);
10170 1258 : }
10171 :
10172 : bool
10173 0 : IsInstallXLogFileSegmentActive(void)
10174 : {
10175 : bool result;
10176 :
10177 0 : LWLockAcquire(ControlFileLock, LW_SHARED);
10178 0 : result = XLogCtl->InstallXLogFileSegmentActive;
10179 0 : LWLockRelease(ControlFileLock);
10180 :
10181 0 : return result;
10182 : }
10183 :
10184 : /*
10185 : * Update the WalWriterSleeping flag.
10186 : */
10187 : void
10188 620 : SetWalWriterSleeping(bool sleeping)
10189 : {
10190 620 : SpinLockAcquire(&XLogCtl->info_lck);
10191 620 : XLogCtl->WalWriterSleeping = sleeping;
10192 620 : SpinLockRelease(&XLogCtl->info_lck);
10193 620 : }
|
