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 24732721 : XLogInsertRecord(XLogRecData *rdata,
785 : XLogRecPtr fpw_lsn,
786 : uint8 flags,
787 : int num_fpi,
788 : uint64 fpi_bytes,
789 : bool topxid_included)
790 : {
791 24732721 : XLogCtlInsert *Insert = &XLogCtl->Insert;
792 : pg_crc32c rdata_crc;
793 : bool inserted;
794 24732721 : XLogRecord *rechdr = (XLogRecord *) rdata->data;
795 24732721 : uint8 info = rechdr->xl_info & ~XLR_INFO_MASK;
796 24732721 : WalInsertClass class = WALINSERT_NORMAL;
797 : XLogRecPtr StartPos;
798 : XLogRecPtr EndPos;
799 24732721 : bool prevDoPageWrites = doPageWrites;
800 : TimeLineID insertTLI;
801 :
802 : /* Does this record type require special handling? */
803 24732721 : if (unlikely(rechdr->xl_rmid == RM_XLOG_ID))
804 : {
805 319864 : if (info == XLOG_SWITCH)
806 815 : class = WALINSERT_SPECIAL_SWITCH;
807 319049 : else if (info == XLOG_CHECKPOINT_REDO)
808 990 : 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 24732721 : 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 24732721 : 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 24732721 : START_CRIT_SECTION();
857 :
858 24732721 : if (likely(class == WALINSERT_NORMAL))
859 : {
860 24730916 : 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 24730916 : if (RedoRecPtr != Insert->RedoRecPtr)
879 : {
880 : Assert(RedoRecPtr < Insert->RedoRecPtr);
881 8105 : RedoRecPtr = Insert->RedoRecPtr;
882 : }
883 24730916 : doPageWrites = (Insert->fullPageWrites || Insert->runningBackups > 0);
884 :
885 24730916 : if (doPageWrites &&
886 22427315 : (!prevDoPageWrites ||
887 20978014 : (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 8844 : WALInsertLockRelease();
894 8844 : END_CRIT_SECTION();
895 8844 : return InvalidXLogRecPtr;
896 : }
897 :
898 : /*
899 : * Reserve space for the record in the WAL. This also sets the xl_prev
900 : * pointer.
901 : */
902 24722072 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
903 : &rechdr->xl_prev);
904 :
905 : /* Normal records are always inserted. */
906 24722072 : inserted = true;
907 : }
908 1805 : 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));
922 815 : WALInsertLockAcquireExclusive();
923 815 : 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));
937 990 : WALInsertLockAcquireExclusive();
938 990 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
939 : &rechdr->xl_prev);
940 990 : RedoRecPtr = Insert->RedoRecPtr = StartPos;
941 990 : inserted = true;
942 : }
943 :
944 24723877 : if (inserted)
945 : {
946 : /*
947 : * Now that xl_prev has been filled in, calculate CRC of the record
948 : * header.
949 : */
950 24723816 : rdata_crc = rechdr->xl_crc;
951 24723816 : COMP_CRC32C(rdata_crc, rechdr, offsetof(XLogRecord, xl_crc));
952 24723816 : FIN_CRC32C(rdata_crc);
953 24723816 : 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 24723816 : 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 24723816 : if ((flags & XLOG_MARK_UNIMPORTANT) == 0)
969 : {
970 24565567 : int lockno = holdingAllLocks ? 0 : MyLockNo;
971 :
972 24565567 : 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 24723877 : WALInsertLockRelease();
988 :
989 24723877 : END_CRIT_SECTION();
990 :
991 24723877 : 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 24723877 : if (topxid_included)
998 223 : MarkSubxactTopXidLogged();
999 :
1000 : /*
1001 : * Update shared LogwrtRqst.Write, if we crossed page boundary.
1002 : */
1003 24723877 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
1004 : {
1005 1883308 : SpinLockAcquire(&XLogCtl->info_lck);
1006 : /* advance global request to include new block(s) */
1007 1883308 : if (XLogCtl->LogwrtRqst.Write < EndPos)
1008 1818407 : XLogCtl->LogwrtRqst.Write = EndPos;
1009 1883308 : SpinLockRelease(&XLogCtl->info_lck);
1010 1883308 : 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 24723877 : if (class == WALINSERT_SPECIAL_SWITCH)
1019 : {
1020 : TRACE_POSTGRESQL_WAL_SWITCH();
1021 815 : 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 815 : if (inserted)
1029 : {
1030 754 : EndPos = StartPos + SizeOfXLogRecord;
1031 754 : 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 24723877 : ProcLastRecPtr = StartPos;
1113 24723877 : XactLastRecEnd = EndPos;
1114 :
1115 : /* Report WAL traffic to the instrumentation. */
1116 24723877 : if (inserted)
1117 : {
1118 24723816 : pgWalUsage.wal_bytes += rechdr->xl_tot_len;
1119 24723816 : pgWalUsage.wal_records++;
1120 24723816 : pgWalUsage.wal_fpi += num_fpi;
1121 24723816 : pgWalUsage.wal_fpi_bytes += fpi_bytes;
1122 :
1123 : /* Required for the flush of pending stats WAL data */
1124 24723816 : pgstat_report_fixed = true;
1125 : }
1126 :
1127 24723877 : 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 24723062 : ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1150 : XLogRecPtr *PrevPtr)
1151 : {
1152 24723062 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1153 : uint64 startbytepos;
1154 : uint64 endbytepos;
1155 : uint64 prevbytepos;
1156 :
1157 24723062 : 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 24723062 : SpinLockAcquire(&Insert->insertpos_lck);
1173 :
1174 24723062 : startbytepos = Insert->CurrBytePos;
1175 24723062 : endbytepos = startbytepos + size;
1176 24723062 : prevbytepos = Insert->PrevBytePos;
1177 24723062 : Insert->CurrBytePos = endbytepos;
1178 24723062 : Insert->PrevBytePos = startbytepos;
1179 :
1180 24723062 : SpinLockRelease(&Insert->insertpos_lck);
1181 :
1182 24723062 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1183 24723062 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1184 24723062 : *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 24723062 : }
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 815 : ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1206 : {
1207 815 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1208 : uint64 startbytepos;
1209 : uint64 endbytepos;
1210 : uint64 prevbytepos;
1211 815 : 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 815 : SpinLockAcquire(&Insert->insertpos_lck);
1222 :
1223 815 : startbytepos = Insert->CurrBytePos;
1224 :
1225 815 : ptr = XLogBytePosToEndRecPtr(startbytepos);
1226 815 : if (XLogSegmentOffset(ptr, wal_segment_size) == 0)
1227 : {
1228 61 : SpinLockRelease(&Insert->insertpos_lck);
1229 61 : *EndPos = *StartPos = ptr;
1230 61 : return false;
1231 : }
1232 :
1233 754 : endbytepos = startbytepos + size;
1234 754 : prevbytepos = Insert->PrevBytePos;
1235 :
1236 754 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1237 754 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1238 :
1239 754 : segleft = wal_segment_size - XLogSegmentOffset(*EndPos, wal_segment_size);
1240 754 : if (segleft != wal_segment_size)
1241 : {
1242 : /* consume the rest of the segment */
1243 754 : *EndPos += segleft;
1244 754 : endbytepos = XLogRecPtrToBytePos(*EndPos);
1245 : }
1246 754 : Insert->CurrBytePos = endbytepos;
1247 754 : Insert->PrevBytePos = startbytepos;
1248 :
1249 754 : SpinLockRelease(&Insert->insertpos_lck);
1250 :
1251 754 : *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 754 : 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 24723816 : 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 24723816 : CurrPos = StartPos;
1280 24723816 : currpos = GetXLogBuffer(CurrPos, tli);
1281 24723816 : 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 24723816 : written = 0;
1291 112801601 : while (rdata != NULL)
1292 : {
1293 88077785 : const char *rdata_data = rdata->data;
1294 88077785 : int rdata_len = rdata->len;
1295 :
1296 90076952 : 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 1999167 : memcpy(currpos, rdata_data, freespace);
1303 1999167 : rdata_data += freespace;
1304 1999167 : rdata_len -= freespace;
1305 1999167 : written += freespace;
1306 1999167 : 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 1999167 : currpos = GetXLogBuffer(CurrPos, tli);
1318 1999167 : pagehdr = (XLogPageHeader) currpos;
1319 1999167 : pagehdr->xlp_rem_len = write_len - written;
1320 1999167 : pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1321 :
1322 : /* skip over the page header */
1323 1999167 : if (XLogSegmentOffset(CurrPos, wal_segment_size) == 0)
1324 : {
1325 1274 : CurrPos += SizeOfXLogLongPHD;
1326 1274 : currpos += SizeOfXLogLongPHD;
1327 : }
1328 : else
1329 : {
1330 1997893 : CurrPos += SizeOfXLogShortPHD;
1331 1997893 : currpos += SizeOfXLogShortPHD;
1332 : }
1333 1999167 : freespace = INSERT_FREESPACE(CurrPos);
1334 : }
1335 :
1336 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1337 88077785 : memcpy(currpos, rdata_data, rdata_len);
1338 88077785 : currpos += rdata_len;
1339 88077785 : CurrPos += rdata_len;
1340 88077785 : freespace -= rdata_len;
1341 88077785 : written += rdata_len;
1342 :
1343 88077785 : 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 24723816 : 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 754 : 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 752925 : 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 752171 : currpos = GetXLogBuffer(CurrPos, tli);
1391 3008684 : MemSet(currpos, 0, SizeOfXLogShortPHD);
1392 :
1393 752171 : CurrPos += XLOG_BLCKSZ;
1394 : }
1395 : }
1396 : else
1397 : {
1398 : /* Align the end position, so that the next record starts aligned */
1399 24723062 : CurrPos = MAXALIGN64(CurrPos);
1400 : }
1401 :
1402 24723816 : 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 24723816 : }
1407 :
1408 : /*
1409 : * Acquire a WAL insertion lock, for inserting to WAL.
1410 : */
1411 : static void
1412 24731917 : 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 24731917 : if (lockToTry == -1)
1430 9347 : lockToTry = MyProcNumber % NUM_XLOGINSERT_LOCKS;
1431 24731917 : 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 24731917 : immed = LWLockAcquire(&WALInsertLocks[MyLockNo].l.lock, LW_EXCLUSIVE);
1438 24731917 : 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 21433 : lockToTry = (lockToTry + 1) % NUM_XLOGINSERT_LOCKS;
1449 : }
1450 24731917 : }
1451 :
1452 : /*
1453 : * Acquire all WAL insertion locks, to prevent other backends from inserting
1454 : * to WAL.
1455 : */
1456 : static void
1457 4791 : 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 38328 : for (i = 0; i < NUM_XLOGINSERT_LOCKS - 1; i++)
1467 : {
1468 33537 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1469 33537 : LWLockUpdateVar(&WALInsertLocks[i].l.lock,
1470 33537 : &WALInsertLocks[i].l.insertingAt,
1471 : PG_UINT64_MAX);
1472 : }
1473 : /* Variable value reset to 0 at release */
1474 4791 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1475 :
1476 4791 : holdingAllLocks = true;
1477 4791 : }
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 24736708 : WALInsertLockRelease(void)
1487 : {
1488 24736708 : if (holdingAllLocks)
1489 : {
1490 : int i;
1491 :
1492 43119 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1493 38328 : LWLockReleaseClearVar(&WALInsertLocks[i].l.lock,
1494 38328 : &WALInsertLocks[i].l.insertingAt,
1495 : 0);
1496 :
1497 4791 : holdingAllLocks = false;
1498 : }
1499 : else
1500 : {
1501 24731917 : LWLockReleaseClearVar(&WALInsertLocks[MyLockNo].l.lock,
1502 24731917 : &WALInsertLocks[MyLockNo].l.insertingAt,
1503 : 0);
1504 : }
1505 24736708 : }
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 2711198 : WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt)
1513 : {
1514 2711198 : if (holdingAllLocks)
1515 : {
1516 : /*
1517 : * We use the last lock to mark our actual position, see comments in
1518 : * WALInsertLockAcquireExclusive.
1519 : */
1520 751499 : LWLockUpdateVar(&WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.lock,
1521 751499 : &WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.insertingAt,
1522 : insertingAt);
1523 : }
1524 : else
1525 1959699 : LWLockUpdateVar(&WALInsertLocks[MyLockNo].l.lock,
1526 1959699 : &WALInsertLocks[MyLockNo].l.insertingAt,
1527 : insertingAt);
1528 2711198 : }
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 2514305 : WaitXLogInsertionsToFinish(XLogRecPtr upto)
1546 : {
1547 : uint64 bytepos;
1548 : XLogRecPtr inserted;
1549 : XLogRecPtr reservedUpto;
1550 : XLogRecPtr finishedUpto;
1551 2514305 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1552 : int i;
1553 :
1554 2514305 : 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 2514305 : inserted = pg_atomic_read_membarrier_u64(&XLogCtl->logInsertResult);
1562 2514305 : if (upto <= inserted)
1563 2002744 : return inserted;
1564 :
1565 : /* Read the current insert position */
1566 511561 : SpinLockAcquire(&Insert->insertpos_lck);
1567 511561 : bytepos = Insert->CurrBytePos;
1568 511561 : SpinLockRelease(&Insert->insertpos_lck);
1569 511561 : 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 511561 : 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 511561 : finishedUpto = reservedUpto;
1597 4604049 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1598 : {
1599 4092488 : 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 4196000 : if (LWLockWaitForVar(&WALInsertLocks[i].l.lock,
1627 4196000 : &WALInsertLocks[i].l.insertingAt,
1628 : insertingat, &insertingat))
1629 : {
1630 : /* the lock was free, so no insertion in progress */
1631 2961684 : insertingat = InvalidXLogRecPtr;
1632 2961684 : break;
1633 : }
1634 :
1635 : /*
1636 : * This insertion is still in progress. Have to wait, unless the
1637 : * inserter has proceeded past 'upto'.
1638 : */
1639 1234316 : } while (insertingat < upto);
1640 :
1641 4092488 : if (XLogRecPtrIsValid(insertingat) && insertingat < finishedUpto)
1642 411065 : 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 511561 : finishedUpto = pg_atomic_monotonic_advance_u64(&XLogCtl->logInsertResult,
1651 : finishedUpto);
1652 :
1653 511561 : 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 27475165 : 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 27475165 : 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 24257520 : 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 3217645 : 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 3217645 : expectedEndPtr = ptr;
1713 3217645 : expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
1714 :
1715 3217645 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1716 3217645 : 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 2711198 : if (ptr % XLOG_BLCKSZ == SizeOfXLogShortPHD &&
1735 12566 : XLogSegmentOffset(ptr, wal_segment_size) > XLOG_BLCKSZ)
1736 12566 : initializedUpto = ptr - SizeOfXLogShortPHD;
1737 2698632 : else if (ptr % XLOG_BLCKSZ == SizeOfXLogLongPHD &&
1738 1112 : XLogSegmentOffset(ptr, wal_segment_size) < XLOG_BLCKSZ)
1739 624 : initializedUpto = ptr - SizeOfXLogLongPHD;
1740 : else
1741 2698008 : initializedUpto = ptr;
1742 :
1743 2711198 : WALInsertLockUpdateInsertingAt(initializedUpto);
1744 :
1745 2711198 : AdvanceXLInsertBuffer(ptr, tli, false);
1746 2711198 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1747 :
1748 2711198 : 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 506447 : 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 3217645 : cachedPage = ptr / XLOG_BLCKSZ;
1766 3217645 : 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 3217645 : 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 107382 : WALReadFromBuffers(char *dstbuf, XLogRecPtr startptr, Size count,
1790 : TimeLineID tli)
1791 : {
1792 107382 : char *pdst = dstbuf;
1793 107382 : XLogRecPtr recptr = startptr;
1794 : XLogRecPtr inserted;
1795 107382 : Size nbytes = count;
1796 :
1797 107382 : if (RecoveryInProgress() || tli != GetWALInsertionTimeLine())
1798 914 : 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 106468 : inserted = pg_atomic_read_u64(&XLogCtl->logInsertResult);
1807 106468 : 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 135823 : while (nbytes > 0)
1828 : {
1829 127194 : uint32 offset = recptr % XLOG_BLCKSZ;
1830 127194 : 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 127194 : 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 127194 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1848 127194 : if (expectedEndPtr != endptr)
1849 97835 : 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 29359 : page = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1857 29359 : psrc = page + offset;
1858 29359 : 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 29359 : pg_read_barrier();
1865 :
1866 : /* data copy */
1867 29359 : memcpy(pdst, psrc, npagebytes);
1868 :
1869 : /*
1870 : * Ensure that the data copy and the second verification step are not
1871 : * reordered.
1872 : */
1873 29359 : 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 29359 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1880 29359 : if (expectedEndPtr != endptr)
1881 4 : break;
1882 :
1883 29355 : pdst += npagebytes;
1884 29355 : recptr += npagebytes;
1885 29355 : nbytes -= npagebytes;
1886 : }
1887 :
1888 : Assert(pdst - dstbuf <= count);
1889 :
1890 106468 : 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 49450509 : XLogBytePosToRecPtr(uint64 bytepos)
1900 : {
1901 : uint64 fullsegs;
1902 : uint64 fullpages;
1903 : uint64 bytesleft;
1904 : uint32 seg_offset;
1905 : XLogRecPtr result;
1906 :
1907 49450509 : fullsegs = bytepos / UsableBytesInSegment;
1908 49450509 : bytesleft = bytepos % UsableBytesInSegment;
1909 :
1910 49450509 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1911 : {
1912 : /* fits on first page of segment */
1913 69328 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1914 : }
1915 : else
1916 : {
1917 : /* account for the first page on segment with long header */
1918 49381181 : seg_offset = XLOG_BLCKSZ;
1919 49381181 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1920 :
1921 49381181 : fullpages = bytesleft / UsableBytesInPage;
1922 49381181 : bytesleft = bytesleft % UsableBytesInPage;
1923 :
1924 49381181 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1925 : }
1926 :
1927 49450509 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1928 :
1929 49450509 : 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 25238461 : XLogBytePosToEndRecPtr(uint64 bytepos)
1940 : {
1941 : uint64 fullsegs;
1942 : uint64 fullpages;
1943 : uint64 bytesleft;
1944 : uint32 seg_offset;
1945 : XLogRecPtr result;
1946 :
1947 25238461 : fullsegs = bytepos / UsableBytesInSegment;
1948 25238461 : bytesleft = bytepos % UsableBytesInSegment;
1949 :
1950 25238461 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1951 : {
1952 : /* fits on first page of segment */
1953 107834 : if (bytesleft == 0)
1954 71777 : seg_offset = 0;
1955 : else
1956 36057 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1957 : }
1958 : else
1959 : {
1960 : /* account for the first page on segment with long header */
1961 25130627 : seg_offset = XLOG_BLCKSZ;
1962 25130627 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1963 :
1964 25130627 : fullpages = bytesleft / UsableBytesInPage;
1965 25130627 : bytesleft = bytesleft % UsableBytesInPage;
1966 :
1967 25130627 : if (bytesleft == 0)
1968 24288 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
1969 : else
1970 25106339 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1971 : }
1972 :
1973 25238461 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1974 :
1975 25238461 : return result;
1976 : }
1977 :
1978 : /*
1979 : * Convert an XLogRecPtr to a "usable byte position".
1980 : */
1981 : static uint64
1982 2778 : XLogRecPtrToBytePos(XLogRecPtr ptr)
1983 : {
1984 : uint64 fullsegs;
1985 : uint32 fullpages;
1986 : uint32 offset;
1987 : uint64 result;
1988 :
1989 2778 : XLByteToSeg(ptr, fullsegs, wal_segment_size);
1990 :
1991 2778 : fullpages = (XLogSegmentOffset(ptr, wal_segment_size)) / XLOG_BLCKSZ;
1992 2778 : offset = ptr % XLOG_BLCKSZ;
1993 :
1994 2778 : if (fullpages == 0)
1995 : {
1996 1062 : result = fullsegs * UsableBytesInSegment;
1997 1062 : if (offset > 0)
1998 : {
1999 : Assert(offset >= SizeOfXLogLongPHD);
2000 285 : result += offset - SizeOfXLogLongPHD;
2001 : }
2002 : }
2003 : else
2004 : {
2005 1716 : result = fullsegs * UsableBytesInSegment +
2006 1716 : (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
2007 1716 : (fullpages - 1) * UsableBytesInPage; /* full pages */
2008 1716 : if (offset > 0)
2009 : {
2010 : Assert(offset >= SizeOfXLogShortPHD);
2011 1706 : result += offset - SizeOfXLogShortPHD;
2012 : }
2013 : }
2014 :
2015 2778 : 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 2716145 : AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
2027 : {
2028 : int nextidx;
2029 : XLogRecPtr OldPageRqstPtr;
2030 : XLogwrtRqst WriteRqst;
2031 2716145 : XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
2032 : XLogRecPtr NewPageBeginPtr;
2033 : XLogPageHeader NewPage;
2034 2716145 : int npages pg_attribute_unused() = 0;
2035 :
2036 2716145 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2037 :
2038 : /*
2039 : * Now that we have the lock, check if someone initialized the page
2040 : * already.
2041 : */
2042 7870916 : while (upto >= XLogCtl->InitializedUpTo || opportunistic)
2043 : {
2044 5159718 : 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 5159718 : OldPageRqstPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]);
2052 5159718 : 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 2365908 : if (opportunistic)
2059 4947 : break;
2060 :
2061 : /* Advance shared memory write request position */
2062 2360961 : SpinLockAcquire(&XLogCtl->info_lck);
2063 2360961 : if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
2064 702252 : XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
2065 2360961 : 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 2360961 : RefreshXLogWriteResult(LogwrtResult);
2072 2360961 : 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 2340804 : LWLockRelease(WALBufMappingLock);
2081 :
2082 2340804 : WaitXLogInsertionsToFinish(OldPageRqstPtr);
2083 :
2084 2340804 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2085 :
2086 2340804 : RefreshXLogWriteResult(LogwrtResult);
2087 2340804 : if (LogwrtResult.Write >= OldPageRqstPtr)
2088 : {
2089 : /* OK, someone wrote it already */
2090 154328 : LWLockRelease(WALWriteLock);
2091 : }
2092 : else
2093 : {
2094 : /* Have to write it ourselves */
2095 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2096 2186476 : WriteRqst.Write = OldPageRqstPtr;
2097 2186476 : WriteRqst.Flush = InvalidXLogRecPtr;
2098 2186476 : XLogWrite(WriteRqst, tli, false);
2099 2186476 : LWLockRelease(WALWriteLock);
2100 2186476 : 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 2186476 : pgstat_report_fixed = true;
2108 : }
2109 : /* Re-acquire WALBufMappingLock and retry */
2110 2340804 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2111 2340804 : 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 2813967 : NewPageBeginPtr = XLogCtl->InitializedUpTo;
2120 2813967 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2121 :
2122 : Assert(XLogRecPtrToBufIdx(NewPageBeginPtr) == nextidx);
2123 :
2124 2813967 : 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 2813967 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], InvalidXLogRecPtr);
2132 2813967 : 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 2813967 : MemSet(NewPage, 0, XLOG_BLCKSZ);
2139 :
2140 : /*
2141 : * Fill the new page's header
2142 : */
2143 2813967 : NewPage->xlp_magic = XLOG_PAGE_MAGIC;
2144 :
2145 : /* NewPage->xlp_info = 0; */ /* done by memset */
2146 2813967 : NewPage->xlp_tli = tli;
2147 2813967 : 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 2813967 : if ((XLogSegmentOffset(NewPage->xlp_pageaddr, wal_segment_size)) == 0)
2155 : {
2156 1935 : XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2157 :
2158 1935 : NewLongPage->xlp_sysid = ControlFile->system_identifier;
2159 1935 : NewLongPage->xlp_seg_size = wal_segment_size;
2160 1935 : NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2161 1935 : 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 2813967 : pg_write_barrier();
2170 :
2171 2813967 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], NewPageEndPtr);
2172 2813967 : XLogCtl->InitializedUpTo = NewPageEndPtr;
2173 :
2174 2813967 : npages++;
2175 : }
2176 2716145 : 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 2716145 : }
2186 :
2187 : /*
2188 : * Calculate CheckPointSegments based on max_wal_size_mb and
2189 : * checkpoint_completion_target.
2190 : */
2191 : static void
2192 9678 : 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 9678 : target = (double) ConvertToXSegs(max_wal_size_mb, wal_segment_size) /
2211 9678 : (1.0 + CheckPointCompletionTarget);
2212 :
2213 : /* round down */
2214 9678 : CheckPointSegments = (int) target;
2215 :
2216 9678 : if (CheckPointSegments < 1)
2217 10 : CheckPointSegments = 1;
2218 9678 : }
2219 :
2220 : void
2221 7260 : assign_max_wal_size(int newval, void *extra)
2222 : {
2223 7260 : max_wal_size_mb = newval;
2224 7260 : CalculateCheckpointSegments();
2225 7260 : }
2226 :
2227 : void
2228 1280 : assign_checkpoint_completion_target(double newval, void *extra)
2229 : {
2230 1280 : CheckPointCompletionTarget = newval;
2231 1280 : CalculateCheckpointSegments();
2232 1280 : }
2233 :
2234 : bool
2235 2476 : check_wal_segment_size(int *newval, void **extra, GucSource source)
2236 : {
2237 2476 : 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 2476 : 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 1936 : 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 1936 : minSegNo = lastredoptr / wal_segment_size +
2264 1936 : ConvertToXSegs(min_wal_size_mb, wal_segment_size) - 1;
2265 1936 : maxSegNo = lastredoptr / wal_segment_size +
2266 1936 : 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 1936 : distance = (1.0 + CheckPointCompletionTarget) * CheckPointDistanceEstimate;
2277 : /* add 10% for good measure. */
2278 1936 : distance *= 1.10;
2279 :
2280 1936 : recycleSegNo = (XLogSegNo) ceil(((double) lastredoptr + distance) /
2281 : wal_segment_size);
2282 :
2283 1936 : if (recycleSegNo < minSegNo)
2284 1363 : recycleSegNo = minSegNo;
2285 1936 : if (recycleSegNo > maxSegNo)
2286 428 : recycleSegNo = maxSegNo;
2287 :
2288 1936 : 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 5137 : XLogCheckpointNeeded(XLogSegNo new_segno)
2302 : {
2303 : XLogSegNo old_segno;
2304 :
2305 5137 : XLByteToSeg(RedoRecPtr, old_segno, wal_segment_size);
2306 :
2307 5137 : if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2308 3225 : return true;
2309 1912 : 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 2353879 : 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 2353879 : 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 2353879 : npages = 0;
2353 2353879 : startidx = 0;
2354 2353879 : 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 2353879 : curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2362 :
2363 5118146 : 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 2926928 : XLogRecPtr EndPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[curridx]);
2371 :
2372 2926928 : 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 2926928 : LogwrtResult.Write = EndPtr;
2379 2926928 : ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2380 :
2381 2926928 : 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 14916 : if (openLogFile >= 0)
2390 6593 : XLogFileClose();
2391 14916 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2392 : wal_segment_size);
2393 14916 : openLogTLI = tli;
2394 :
2395 : /* create/use new log file */
2396 14916 : openLogFile = XLogFileInit(openLogSegNo, tli);
2397 14916 : ReserveExternalFD();
2398 : }
2399 :
2400 : /* Make sure we have the current logfile open */
2401 2926928 : 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 2926928 : if (npages == 0)
2412 : {
2413 : /* first of group */
2414 2371412 : startidx = curridx;
2415 2371412 : startoffset = XLogSegmentOffset(LogwrtResult.Write - XLOG_BLCKSZ,
2416 : wal_segment_size);
2417 : }
2418 2926928 : 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 2926928 : last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2427 :
2428 5695197 : finishing_seg = !ispartialpage &&
2429 2768269 : (startoffset + npages * XLOG_BLCKSZ) >= wal_segment_size;
2430 :
2431 2926928 : if (last_iteration ||
2432 574507 : 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 2371412 : from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2443 2371412 : nbytes = npages * (Size) XLOG_BLCKSZ;
2444 2371412 : nleft = nbytes;
2445 : do
2446 : {
2447 2371412 : errno = 0;
2448 :
2449 : /*
2450 : * Measure I/O timing to write WAL data, for pg_stat_io.
2451 : */
2452 2371412 : start = pgstat_prepare_io_time(track_wal_io_timing);
2453 :
2454 2371412 : pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE);
2455 2371412 : written = pg_pwrite(openLogFile, from, nleft, startoffset);
2456 2371412 : pgstat_report_wait_end();
2457 :
2458 2371412 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL,
2459 : IOOP_WRITE, start, 1, written);
2460 :
2461 2371412 : 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 2371412 : nleft -= written;
2479 2371412 : from += written;
2480 2371412 : startoffset += written;
2481 2371412 : } while (nleft > 0);
2482 :
2483 2371412 : 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 2371412 : if (finishing_seg)
2499 : {
2500 2049 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2501 :
2502 : /* signal that we need to wakeup walsenders later */
2503 2049 : WalSndWakeupRequest();
2504 :
2505 2049 : LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2506 :
2507 2049 : if (XLogArchivingActive())
2508 411 : XLogArchiveNotifySeg(openLogSegNo, tli);
2509 :
2510 2049 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2511 2049 : 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 2049 : if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2521 : {
2522 280 : (void) GetRedoRecPtr();
2523 280 : if (XLogCheckpointNeeded(openLogSegNo))
2524 224 : RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2525 : }
2526 : }
2527 : }
2528 :
2529 2926928 : if (ispartialpage)
2530 : {
2531 : /* Only asked to write a partial page */
2532 158659 : LogwrtResult.Write = WriteRqst.Write;
2533 158659 : break;
2534 : }
2535 2768269 : curridx = NextBufIdx(curridx);
2536 :
2537 : /* If flexible, break out of loop as soon as we wrote something */
2538 2768269 : if (flexible && npages == 0)
2539 4002 : break;
2540 : }
2541 :
2542 : Assert(npages == 0);
2543 :
2544 : /*
2545 : * If asked to flush, do so
2546 : */
2547 2353879 : if (LogwrtResult.Flush < WriteRqst.Flush &&
2548 166605 : 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 166524 : if (wal_sync_method != WAL_SYNC_METHOD_OPEN &&
2556 166524 : wal_sync_method != WAL_SYNC_METHOD_OPEN_DSYNC)
2557 : {
2558 166524 : if (openLogFile >= 0 &&
2559 166506 : !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2560 : wal_segment_size))
2561 81 : XLogFileClose();
2562 166524 : if (openLogFile < 0)
2563 : {
2564 99 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2565 : wal_segment_size);
2566 99 : openLogTLI = tli;
2567 99 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2568 99 : ReserveExternalFD();
2569 : }
2570 :
2571 166524 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2572 : }
2573 :
2574 : /* signal that we need to wakeup walsenders later */
2575 166524 : WalSndWakeupRequest();
2576 :
2577 166524 : 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 2353879 : SpinLockAcquire(&XLogCtl->info_lck);
2588 2353879 : if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
2589 146023 : XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
2590 2353879 : if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
2591 168148 : XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
2592 2353879 : 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 2353879 : pg_atomic_write_u64(&XLogCtl->logWriteResult, LogwrtResult.Write);
2600 2353879 : pg_write_barrier();
2601 2353879 : 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 2353879 : }
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 62144 : XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2631 : {
2632 62144 : XLogRecPtr WriteRqstPtr = asyncXactLSN;
2633 : bool sleeping;
2634 62144 : bool wakeup = false;
2635 : XLogRecPtr prevAsyncXactLSN;
2636 :
2637 62144 : SpinLockAcquire(&XLogCtl->info_lck);
2638 62144 : sleeping = XLogCtl->WalWriterSleeping;
2639 62144 : prevAsyncXactLSN = XLogCtl->asyncXactLSN;
2640 62144 : if (XLogCtl->asyncXactLSN < asyncXactLSN)
2641 61576 : XLogCtl->asyncXactLSN = asyncXactLSN;
2642 62144 : 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 62144 : if (asyncXactLSN <= prevAsyncXactLSN)
2649 568 : 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 61576 : if (sleeping)
2658 42 : wakeup = true;
2659 : else
2660 : {
2661 : int flushblocks;
2662 :
2663 61534 : RefreshXLogWriteResult(LogwrtResult);
2664 :
2665 61534 : flushblocks =
2666 61534 : WriteRqstPtr / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
2667 :
2668 61534 : if (WalWriterFlushAfter == 0 || flushblocks >= WalWriterFlushAfter)
2669 5160 : wakeup = true;
2670 : }
2671 :
2672 61576 : if (wakeup)
2673 : {
2674 5202 : volatile PROC_HDR *procglobal = ProcGlobal;
2675 5202 : ProcNumber walwriterProc = procglobal->walwriterProc;
2676 :
2677 5202 : if (walwriterProc != INVALID_PROC_NUMBER)
2678 871 : 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 43343 : XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn)
2688 : {
2689 43343 : SpinLockAcquire(&XLogCtl->info_lck);
2690 43343 : XLogCtl->replicationSlotMinLSN = lsn;
2691 43343 : SpinLockRelease(&XLogCtl->info_lck);
2692 43343 : }
2693 :
2694 :
2695 : /*
2696 : * Return the oldest LSN we must retain to satisfy the needs of some
2697 : * replication slot.
2698 : */
2699 : XLogRecPtr
2700 2568 : XLogGetReplicationSlotMinimumLSN(void)
2701 : {
2702 : XLogRecPtr retval;
2703 :
2704 2568 : SpinLockAcquire(&XLogCtl->info_lck);
2705 2568 : retval = XLogCtl->replicationSlotMinLSN;
2706 2568 : SpinLockRelease(&XLogCtl->info_lck);
2707 :
2708 2568 : 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 117752 : UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2722 : {
2723 : /* Quick check using our local copy of the variable */
2724 117752 : if (!updateMinRecoveryPoint || (!force && lsn <= LocalMinRecoveryPoint))
2725 110385 : 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 7367 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint) && InRecovery)
2739 : {
2740 32 : updateMinRecoveryPoint = false;
2741 32 : return;
2742 : }
2743 :
2744 7335 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2745 :
2746 : /* update local copy */
2747 7335 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
2748 7335 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2749 :
2750 7335 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
2751 1 : updateMinRecoveryPoint = false;
2752 7334 : 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 5842 : newMinRecoveryPoint = GetCurrentReplayRecPtr(&newMinRecoveryPointTLI);
2771 5842 : 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 5842 : if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2778 : {
2779 5493 : ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2780 5493 : ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2781 5493 : UpdateControlFile();
2782 5493 : LocalMinRecoveryPoint = newMinRecoveryPoint;
2783 5493 : LocalMinRecoveryPointTLI = newMinRecoveryPointTLI;
2784 :
2785 5493 : 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 7335 : 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 857710 : XLogFlush(XLogRecPtr record)
2802 : {
2803 : XLogRecPtr WriteRqstPtr;
2804 : XLogwrtRqst WriteRqst;
2805 857710 : 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 857710 : if (!XLogInsertAllowed())
2815 : {
2816 117301 : UpdateMinRecoveryPoint(record, false);
2817 679980 : return;
2818 : }
2819 :
2820 : /* Quick exit if already known flushed */
2821 740409 : if (record <= LogwrtResult.Flush)
2822 562679 : 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 177730 : 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 177730 : 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 3778 : {
2851 : XLogRecPtr insertpos;
2852 :
2853 : /* done already? */
2854 181508 : RefreshXLogWriteResult(LogwrtResult);
2855 181508 : if (record <= LogwrtResult.Flush)
2856 12954 : 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 168554 : SpinLockAcquire(&XLogCtl->info_lck);
2863 168554 : if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
2864 11552 : WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
2865 168554 : SpinLockRelease(&XLogCtl->info_lck);
2866 168554 : 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 168554 : 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 3778 : continue;
2883 : }
2884 :
2885 : /* Got the lock; recheck whether request is satisfied */
2886 164776 : RefreshXLogWriteResult(LogwrtResult);
2887 164776 : if (record <= LogwrtResult.Flush)
2888 : {
2889 2255 : LWLockRelease(WALWriteLock);
2890 2255 : 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 162521 : 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 162521 : WriteRqst.Write = insertpos;
2924 162521 : WriteRqst.Flush = insertpos;
2925 :
2926 162521 : XLogWrite(WriteRqst, insertTLI, false);
2927 :
2928 162521 : LWLockRelease(WALWriteLock);
2929 : /* done */
2930 162521 : break;
2931 : }
2932 :
2933 177730 : END_CRIT_SECTION();
2934 :
2935 : /* wake up walsenders now that we've released heavily contended locks */
2936 177730 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
2937 :
2938 : /*
2939 : * If we flushed an LSN that someone was waiting for, notify the waiters.
2940 : */
2941 355460 : if (waitLSNState &&
2942 177730 : (LogwrtResult.Flush >=
2943 177730 : pg_atomic_read_u64(&waitLSNState->minWaitedLSN[WAIT_LSN_TYPE_PRIMARY_FLUSH])))
2944 10 : WaitLSNWakeup(WAIT_LSN_TYPE_PRIMARY_FLUSH, LogwrtResult.Flush);
2945 :
2946 : /*
2947 : * If we still haven't flushed to the request point then we have a
2948 : * problem; most likely, the requested flush point is past end of XLOG.
2949 : * This has been seen to occur when a disk page has a corrupted LSN.
2950 : *
2951 : * Formerly we treated this as a PANIC condition, but that hurts the
2952 : * system's robustness rather than helping it: we do not want to take down
2953 : * the whole system due to corruption on one data page. In particular, if
2954 : * the bad page is encountered again during recovery then we would be
2955 : * unable to restart the database at all! (This scenario actually
2956 : * happened in the field several times with 7.1 releases.) As of 8.4, bad
2957 : * LSNs encountered during recovery are UpdateMinRecoveryPoint's problem;
2958 : * the only time we can reach here during recovery is while flushing the
2959 : * end-of-recovery checkpoint record, and we don't expect that to have a
2960 : * bad LSN.
2961 : *
2962 : * Note that for calls from xact.c, the ERROR will be promoted to PANIC
2963 : * since xact.c calls this routine inside a critical section. However,
2964 : * calls from bufmgr.c are not within critical sections and so we will not
2965 : * force a restart for a bad LSN on a data page.
2966 : */
2967 177730 : if (LogwrtResult.Flush < record)
2968 0 : elog(ERROR,
2969 : "xlog flush request %X/%08X is not satisfied --- flushed only to %X/%08X",
2970 : LSN_FORMAT_ARGS(record),
2971 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
2972 :
2973 : /*
2974 : * Cross-check XLogNeedsFlush(). Some of the checks of XLogFlush() and
2975 : * XLogNeedsFlush() are duplicated, and this assertion ensures that these
2976 : * remain consistent.
2977 : */
2978 : Assert(!XLogNeedsFlush(record));
2979 : }
2980 :
2981 : /*
2982 : * Write & flush xlog, but without specifying exactly where to.
2983 : *
2984 : * We normally write only completed blocks; but if there is nothing to do on
2985 : * that basis, we check for unwritten async commits in the current incomplete
2986 : * block, and write through the latest one of those. Thus, if async commits
2987 : * are not being used, we will write complete blocks only.
2988 : *
2989 : * If, based on the above, there's anything to write we do so immediately. But
2990 : * to avoid calling fsync, fdatasync et. al. at a rate that'd impact
2991 : * concurrent IO, we only flush WAL every wal_writer_delay ms, or if there's
2992 : * more than wal_writer_flush_after unflushed blocks.
2993 : *
2994 : * We can guarantee that async commits reach disk after at most three
2995 : * wal_writer_delay cycles. (When flushing complete blocks, we allow XLogWrite
2996 : * to write "flexibly", meaning it can stop at the end of the buffer ring;
2997 : * this makes a difference only with very high load or long wal_writer_delay,
2998 : * but imposes one extra cycle for the worst case for async commits.)
2999 : *
3000 : * This routine is invoked periodically by the background walwriter process.
3001 : *
3002 : * Returns true if there was any work to do, even if we skipped flushing due
3003 : * to wal_writer_delay/wal_writer_flush_after.
3004 : */
3005 : bool
3006 15617 : XLogBackgroundFlush(void)
3007 : {
3008 : XLogwrtRqst WriteRqst;
3009 15617 : bool flexible = true;
3010 : static TimestampTz lastflush;
3011 : TimestampTz now;
3012 : int flushblocks;
3013 : TimeLineID insertTLI;
3014 :
3015 : /* XLOG doesn't need flushing during recovery */
3016 15617 : if (RecoveryInProgress())
3017 0 : return false;
3018 :
3019 : /*
3020 : * Since we're not in recovery, InsertTimeLineID is set and can't change,
3021 : * so we can read it without a lock.
3022 : */
3023 15617 : insertTLI = XLogCtl->InsertTimeLineID;
3024 :
3025 : /* read updated LogwrtRqst */
3026 15617 : SpinLockAcquire(&XLogCtl->info_lck);
3027 15617 : WriteRqst = XLogCtl->LogwrtRqst;
3028 15617 : SpinLockRelease(&XLogCtl->info_lck);
3029 :
3030 : /* back off to last completed page boundary */
3031 15617 : WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
3032 :
3033 : /* if we have already flushed that far, consider async commit records */
3034 15617 : RefreshXLogWriteResult(LogwrtResult);
3035 15617 : if (WriteRqst.Write <= LogwrtResult.Flush)
3036 : {
3037 11513 : SpinLockAcquire(&XLogCtl->info_lck);
3038 11513 : WriteRqst.Write = XLogCtl->asyncXactLSN;
3039 11513 : SpinLockRelease(&XLogCtl->info_lck);
3040 11513 : flexible = false; /* ensure it all gets written */
3041 : }
3042 :
3043 : /*
3044 : * If already known flushed, we're done. Just need to check if we are
3045 : * holding an open file handle to a logfile that's no longer in use,
3046 : * preventing the file from being deleted.
3047 : */
3048 15617 : if (WriteRqst.Write <= LogwrtResult.Flush)
3049 : {
3050 10670 : if (openLogFile >= 0)
3051 : {
3052 6652 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
3053 : wal_segment_size))
3054 : {
3055 203 : XLogFileClose();
3056 : }
3057 : }
3058 10670 : return false;
3059 : }
3060 :
3061 : /*
3062 : * Determine how far to flush WAL, based on the wal_writer_delay and
3063 : * wal_writer_flush_after GUCs.
3064 : *
3065 : * Note that XLogSetAsyncXactLSN() performs similar calculation based on
3066 : * wal_writer_flush_after, to decide when to wake us up. Make sure the
3067 : * logic is the same in both places if you change this.
3068 : */
3069 4947 : now = GetCurrentTimestamp();
3070 4947 : flushblocks =
3071 4947 : WriteRqst.Write / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
3072 :
3073 4947 : if (WalWriterFlushAfter == 0 || lastflush == 0)
3074 : {
3075 : /* first call, or block based limits disabled */
3076 296 : WriteRqst.Flush = WriteRqst.Write;
3077 296 : lastflush = now;
3078 : }
3079 4651 : else if (TimestampDifferenceExceeds(lastflush, now, WalWriterDelay))
3080 : {
3081 : /*
3082 : * Flush the writes at least every WalWriterDelay ms. This is
3083 : * important to bound the amount of time it takes for an asynchronous
3084 : * commit to hit disk.
3085 : */
3086 4379 : WriteRqst.Flush = WriteRqst.Write;
3087 4379 : lastflush = now;
3088 : }
3089 272 : else if (flushblocks >= WalWriterFlushAfter)
3090 : {
3091 : /* exceeded wal_writer_flush_after blocks, flush */
3092 232 : WriteRqst.Flush = WriteRqst.Write;
3093 232 : lastflush = now;
3094 : }
3095 : else
3096 : {
3097 : /* no flushing, this time round */
3098 40 : WriteRqst.Flush = InvalidXLogRecPtr;
3099 : }
3100 :
3101 : #ifdef WAL_DEBUG
3102 : if (XLOG_DEBUG)
3103 : elog(LOG, "xlog bg flush request write %X/%08X; flush: %X/%08X, current is write %X/%08X; flush %X/%08X",
3104 : LSN_FORMAT_ARGS(WriteRqst.Write),
3105 : LSN_FORMAT_ARGS(WriteRqst.Flush),
3106 : LSN_FORMAT_ARGS(LogwrtResult.Write),
3107 : LSN_FORMAT_ARGS(LogwrtResult.Flush));
3108 : #endif
3109 :
3110 4947 : START_CRIT_SECTION();
3111 :
3112 : /* now wait for any in-progress insertions to finish and get write lock */
3113 4947 : WaitXLogInsertionsToFinish(WriteRqst.Write);
3114 4947 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3115 4947 : RefreshXLogWriteResult(LogwrtResult);
3116 4947 : if (WriteRqst.Write > LogwrtResult.Write ||
3117 113 : WriteRqst.Flush > LogwrtResult.Flush)
3118 : {
3119 4882 : XLogWrite(WriteRqst, insertTLI, flexible);
3120 : }
3121 4947 : LWLockRelease(WALWriteLock);
3122 :
3123 4947 : END_CRIT_SECTION();
3124 :
3125 : /* wake up walsenders now that we've released heavily contended locks */
3126 4947 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3127 :
3128 : /*
3129 : * If we flushed an LSN that someone was waiting for, notify the waiters.
3130 : */
3131 9894 : if (waitLSNState &&
3132 4947 : (LogwrtResult.Flush >=
3133 4947 : pg_atomic_read_u64(&waitLSNState->minWaitedLSN[WAIT_LSN_TYPE_PRIMARY_FLUSH])))
3134 1 : WaitLSNWakeup(WAIT_LSN_TYPE_PRIMARY_FLUSH, LogwrtResult.Flush);
3135 :
3136 : /*
3137 : * Great, done. To take some work off the critical path, try to initialize
3138 : * as many of the no-longer-needed WAL buffers for future use as we can.
3139 : */
3140 4947 : AdvanceXLInsertBuffer(InvalidXLogRecPtr, insertTLI, true);
3141 :
3142 : /*
3143 : * If we determined that we need to write data, but somebody else
3144 : * wrote/flushed already, it should be considered as being active, to
3145 : * avoid hibernating too early.
3146 : */
3147 4947 : return true;
3148 : }
3149 :
3150 : /*
3151 : * Test whether XLOG data has been flushed up to (at least) the given
3152 : * position, or whether the minimum recovery point has been updated past
3153 : * the given position.
3154 : *
3155 : * Returns true if a flush is still needed, or if the minimum recovery point
3156 : * must be updated.
3157 : *
3158 : * It is possible that someone else is already in the process of flushing
3159 : * that far, or has updated the minimum recovery point up to the given
3160 : * position.
3161 : */
3162 : bool
3163 16199310 : XLogNeedsFlush(XLogRecPtr record)
3164 : {
3165 : /*
3166 : * During recovery, we don't flush WAL but update minRecoveryPoint
3167 : * instead. So "needs flush" is taken to mean whether minRecoveryPoint
3168 : * would need to be updated.
3169 : *
3170 : * Using XLogInsertAllowed() rather than RecoveryInProgress() matters for
3171 : * the case of an end-of-recovery checkpoint, where WAL data is flushed.
3172 : * This check should be consistent with the one in XLogFlush().
3173 : */
3174 16199310 : if (!XLogInsertAllowed())
3175 : {
3176 : /* Quick exit if already known to be updated or cannot be updated */
3177 539084 : if (!updateMinRecoveryPoint || record <= LocalMinRecoveryPoint)
3178 523420 : return false;
3179 :
3180 : /*
3181 : * An invalid minRecoveryPoint means that we need to recover all the
3182 : * WAL, i.e., we're doing crash recovery. We never modify the control
3183 : * file's value in that case, so we can short-circuit future checks
3184 : * here too. This triggers a quick exit path for the startup process,
3185 : * which cannot update its local copy of minRecoveryPoint as long as
3186 : * it has not replayed all WAL available when doing crash recovery.
3187 : */
3188 15664 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint) && InRecovery)
3189 : {
3190 0 : updateMinRecoveryPoint = false;
3191 0 : return false;
3192 : }
3193 :
3194 : /*
3195 : * Update local copy of minRecoveryPoint. But if the lock is busy,
3196 : * just return a conservative guess.
3197 : */
3198 15664 : if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3199 0 : return true;
3200 15664 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
3201 15664 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3202 15664 : LWLockRelease(ControlFileLock);
3203 :
3204 : /*
3205 : * Check minRecoveryPoint for any other process than the startup
3206 : * process doing crash recovery, which should not update the control
3207 : * file value if crash recovery is still running.
3208 : */
3209 15664 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
3210 0 : updateMinRecoveryPoint = false;
3211 :
3212 : /* check again */
3213 15664 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3214 103 : return false;
3215 : else
3216 15561 : return true;
3217 : }
3218 :
3219 : /* Quick exit if already known flushed */
3220 15660226 : if (record <= LogwrtResult.Flush)
3221 15495999 : return false;
3222 :
3223 : /* read LogwrtResult and update local state */
3224 164227 : RefreshXLogWriteResult(LogwrtResult);
3225 :
3226 : /* check again */
3227 164227 : if (record <= LogwrtResult.Flush)
3228 2950 : return false;
3229 :
3230 161277 : return true;
3231 : }
3232 :
3233 : /*
3234 : * Try to make a given XLOG file segment exist.
3235 : *
3236 : * logsegno: identify segment.
3237 : *
3238 : * *added: on return, true if this call raised the number of extant segments.
3239 : *
3240 : * path: on return, this char[MAXPGPATH] has the path to the logsegno file.
3241 : *
3242 : * Returns -1 or FD of opened file. A -1 here is not an error; a caller
3243 : * wanting an open segment should attempt to open "path", which usually will
3244 : * succeed. (This is weird, but it's efficient for the callers.)
3245 : */
3246 : static int
3247 16087 : XLogFileInitInternal(XLogSegNo logsegno, TimeLineID logtli,
3248 : bool *added, char *path)
3249 : {
3250 : char tmppath[MAXPGPATH];
3251 : XLogSegNo installed_segno;
3252 : XLogSegNo max_segno;
3253 : int fd;
3254 : int save_errno;
3255 16087 : int open_flags = O_RDWR | O_CREAT | O_EXCL | PG_BINARY;
3256 : instr_time io_start;
3257 :
3258 : Assert(logtli != 0);
3259 :
3260 16087 : XLogFilePath(path, logtli, logsegno, wal_segment_size);
3261 :
3262 : /*
3263 : * Try to use existent file (checkpoint maker may have created it already)
3264 : */
3265 16087 : *added = false;
3266 16087 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3267 16087 : get_sync_bit(wal_sync_method));
3268 16087 : if (fd < 0)
3269 : {
3270 1589 : if (errno != ENOENT)
3271 0 : ereport(ERROR,
3272 : (errcode_for_file_access(),
3273 : errmsg("could not open file \"%s\": %m", path)));
3274 : }
3275 : else
3276 14498 : return fd;
3277 :
3278 : /*
3279 : * Initialize an empty (all zeroes) segment. NOTE: it is possible that
3280 : * another process is doing the same thing. If so, we will end up
3281 : * pre-creating an extra log segment. That seems OK, and better than
3282 : * holding the lock throughout this lengthy process.
3283 : */
3284 1589 : elog(DEBUG2, "creating and filling new WAL file");
3285 :
3286 1589 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3287 :
3288 1589 : unlink(tmppath);
3289 :
3290 1589 : if (io_direct_flags & IO_DIRECT_WAL_INIT)
3291 0 : open_flags |= PG_O_DIRECT;
3292 :
3293 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3294 1589 : fd = BasicOpenFile(tmppath, open_flags);
3295 1589 : if (fd < 0)
3296 0 : ereport(ERROR,
3297 : (errcode_for_file_access(),
3298 : errmsg("could not create file \"%s\": %m", tmppath)));
3299 :
3300 : /* Measure I/O timing when initializing segment */
3301 1589 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3302 :
3303 1589 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE);
3304 1589 : save_errno = 0;
3305 1589 : if (wal_init_zero)
3306 : {
3307 : ssize_t rc;
3308 :
3309 : /*
3310 : * Zero-fill the file. With this setting, we do this the hard way to
3311 : * ensure that all the file space has really been allocated. On
3312 : * platforms that allow "holes" in files, just seeking to the end
3313 : * doesn't allocate intermediate space. This way, we know that we
3314 : * have all the space and (after the fsync below) that all the
3315 : * indirect blocks are down on disk. Therefore, fdatasync(2) or
3316 : * O_DSYNC will be sufficient to sync future writes to the log file.
3317 : */
3318 1589 : rc = pg_pwrite_zeros(fd, wal_segment_size, 0);
3319 :
3320 1589 : if (rc < 0)
3321 0 : save_errno = errno;
3322 : }
3323 : else
3324 : {
3325 : /*
3326 : * Otherwise, seeking to the end and writing a solitary byte is
3327 : * enough.
3328 : */
3329 0 : errno = 0;
3330 0 : if (pg_pwrite(fd, "\0", 1, wal_segment_size - 1) != 1)
3331 : {
3332 : /* if write didn't set errno, assume no disk space */
3333 0 : save_errno = errno ? errno : ENOSPC;
3334 : }
3335 : }
3336 1589 : pgstat_report_wait_end();
3337 :
3338 : /*
3339 : * A full segment worth of data is written when using wal_init_zero. One
3340 : * byte is written when not using it.
3341 : */
3342 1589 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT, IOOP_WRITE,
3343 : io_start, 1,
3344 1589 : wal_init_zero ? wal_segment_size : 1);
3345 :
3346 1589 : if (save_errno)
3347 : {
3348 : /*
3349 : * If we fail to make the file, delete it to release disk space
3350 : */
3351 0 : unlink(tmppath);
3352 :
3353 0 : close(fd);
3354 :
3355 0 : errno = save_errno;
3356 :
3357 0 : ereport(ERROR,
3358 : (errcode_for_file_access(),
3359 : errmsg("could not write to file \"%s\": %m", tmppath)));
3360 : }
3361 :
3362 : /* Measure I/O timing when flushing segment */
3363 1589 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3364 :
3365 1589 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC);
3366 1589 : if (pg_fsync(fd) != 0)
3367 : {
3368 0 : save_errno = errno;
3369 0 : close(fd);
3370 0 : errno = save_errno;
3371 0 : ereport(ERROR,
3372 : (errcode_for_file_access(),
3373 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3374 : }
3375 1589 : pgstat_report_wait_end();
3376 :
3377 1589 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT,
3378 : IOOP_FSYNC, io_start, 1, 0);
3379 :
3380 1589 : if (close(fd) != 0)
3381 0 : ereport(ERROR,
3382 : (errcode_for_file_access(),
3383 : errmsg("could not close file \"%s\": %m", tmppath)));
3384 :
3385 : /*
3386 : * Now move the segment into place with its final name. Cope with
3387 : * possibility that someone else has created the file while we were
3388 : * filling ours: if so, use ours to pre-create a future log segment.
3389 : */
3390 1589 : installed_segno = logsegno;
3391 :
3392 : /*
3393 : * XXX: What should we use as max_segno? We used to use XLOGfileslop when
3394 : * that was a constant, but that was always a bit dubious: normally, at a
3395 : * checkpoint, XLOGfileslop was the offset from the checkpoint record, but
3396 : * here, it was the offset from the insert location. We can't do the
3397 : * normal XLOGfileslop calculation here because we don't have access to
3398 : * the prior checkpoint's redo location. So somewhat arbitrarily, just use
3399 : * CheckPointSegments.
3400 : */
3401 1589 : max_segno = logsegno + CheckPointSegments;
3402 1589 : if (InstallXLogFileSegment(&installed_segno, tmppath, true, max_segno,
3403 : logtli))
3404 : {
3405 1589 : *added = true;
3406 1589 : elog(DEBUG2, "done creating and filling new WAL file");
3407 : }
3408 : else
3409 : {
3410 : /*
3411 : * No need for any more future segments, or InstallXLogFileSegment()
3412 : * failed to rename the file into place. If the rename failed, a
3413 : * caller opening the file may fail.
3414 : */
3415 0 : unlink(tmppath);
3416 0 : elog(DEBUG2, "abandoned new WAL file");
3417 : }
3418 :
3419 1589 : return -1;
3420 : }
3421 :
3422 : /*
3423 : * Create a new XLOG file segment, or open a pre-existing one.
3424 : *
3425 : * logsegno: identify segment to be created/opened.
3426 : *
3427 : * Returns FD of opened file.
3428 : *
3429 : * Note: errors here are ERROR not PANIC because we might or might not be
3430 : * inside a critical section (eg, during checkpoint there is no reason to
3431 : * take down the system on failure). They will promote to PANIC if we are
3432 : * in a critical section.
3433 : */
3434 : int
3435 15865 : XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)
3436 : {
3437 : bool ignore_added;
3438 : char path[MAXPGPATH];
3439 : int fd;
3440 :
3441 : Assert(logtli != 0);
3442 :
3443 15865 : fd = XLogFileInitInternal(logsegno, logtli, &ignore_added, path);
3444 15865 : if (fd >= 0)
3445 14334 : return fd;
3446 :
3447 : /* Now open original target segment (might not be file I just made) */
3448 1531 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3449 1531 : get_sync_bit(wal_sync_method));
3450 1531 : if (fd < 0)
3451 0 : ereport(ERROR,
3452 : (errcode_for_file_access(),
3453 : errmsg("could not open file \"%s\": %m", path)));
3454 1531 : return fd;
3455 : }
3456 :
3457 : /*
3458 : * Create a new XLOG file segment by copying a pre-existing one.
3459 : *
3460 : * destsegno: identify segment to be created.
3461 : *
3462 : * srcTLI, srcsegno: identify segment to be copied (could be from
3463 : * a different timeline)
3464 : *
3465 : * upto: how much of the source file to copy (the rest is filled with
3466 : * zeros)
3467 : *
3468 : * Currently this is only used during recovery, and so there are no locking
3469 : * considerations. But we should be just as tense as XLogFileInit to avoid
3470 : * emplacing a bogus file.
3471 : */
3472 : static void
3473 43 : XLogFileCopy(TimeLineID destTLI, XLogSegNo destsegno,
3474 : TimeLineID srcTLI, XLogSegNo srcsegno,
3475 : int upto)
3476 : {
3477 : char path[MAXPGPATH];
3478 : char tmppath[MAXPGPATH];
3479 : PGAlignedXLogBlock buffer;
3480 : int srcfd;
3481 : int fd;
3482 : int nbytes;
3483 :
3484 : /*
3485 : * Open the source file
3486 : */
3487 43 : XLogFilePath(path, srcTLI, srcsegno, wal_segment_size);
3488 43 : srcfd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
3489 43 : if (srcfd < 0)
3490 0 : ereport(ERROR,
3491 : (errcode_for_file_access(),
3492 : errmsg("could not open file \"%s\": %m", path)));
3493 :
3494 : /*
3495 : * Copy into a temp file name.
3496 : */
3497 43 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3498 :
3499 43 : unlink(tmppath);
3500 :
3501 : /* do not use get_sync_bit() here --- want to fsync only at end of fill */
3502 43 : fd = OpenTransientFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
3503 43 : if (fd < 0)
3504 0 : ereport(ERROR,
3505 : (errcode_for_file_access(),
3506 : errmsg("could not create file \"%s\": %m", tmppath)));
3507 :
3508 : /*
3509 : * Do the data copying.
3510 : */
3511 88107 : for (nbytes = 0; nbytes < wal_segment_size; nbytes += sizeof(buffer))
3512 : {
3513 : int nread;
3514 :
3515 88064 : nread = upto - nbytes;
3516 :
3517 : /*
3518 : * The part that is not read from the source file is filled with
3519 : * zeros.
3520 : */
3521 88064 : if (nread < sizeof(buffer))
3522 43 : memset(buffer.data, 0, sizeof(buffer));
3523 :
3524 88064 : if (nread > 0)
3525 : {
3526 : int r;
3527 :
3528 2793 : if (nread > sizeof(buffer))
3529 2750 : nread = sizeof(buffer);
3530 2793 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ);
3531 2793 : r = read(srcfd, buffer.data, nread);
3532 2793 : if (r != nread)
3533 : {
3534 0 : if (r < 0)
3535 0 : ereport(ERROR,
3536 : (errcode_for_file_access(),
3537 : errmsg("could not read file \"%s\": %m",
3538 : path)));
3539 : else
3540 0 : ereport(ERROR,
3541 : (errcode(ERRCODE_DATA_CORRUPTED),
3542 : errmsg("could not read file \"%s\": read %d of %zu",
3543 : path, r, (Size) nread)));
3544 : }
3545 2793 : pgstat_report_wait_end();
3546 : }
3547 88064 : errno = 0;
3548 88064 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_WRITE);
3549 88064 : if ((int) write(fd, buffer.data, sizeof(buffer)) != (int) sizeof(buffer))
3550 : {
3551 0 : int save_errno = errno;
3552 :
3553 : /*
3554 : * If we fail to make the file, delete it to release disk space
3555 : */
3556 0 : unlink(tmppath);
3557 : /* if write didn't set errno, assume problem is no disk space */
3558 0 : errno = save_errno ? save_errno : ENOSPC;
3559 :
3560 0 : ereport(ERROR,
3561 : (errcode_for_file_access(),
3562 : errmsg("could not write to file \"%s\": %m", tmppath)));
3563 : }
3564 88064 : pgstat_report_wait_end();
3565 : }
3566 :
3567 43 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_SYNC);
3568 43 : if (pg_fsync(fd) != 0)
3569 0 : ereport(data_sync_elevel(ERROR),
3570 : (errcode_for_file_access(),
3571 : errmsg("could not fsync file \"%s\": %m", tmppath)));
3572 43 : pgstat_report_wait_end();
3573 :
3574 43 : if (CloseTransientFile(fd) != 0)
3575 0 : ereport(ERROR,
3576 : (errcode_for_file_access(),
3577 : errmsg("could not close file \"%s\": %m", tmppath)));
3578 :
3579 43 : if (CloseTransientFile(srcfd) != 0)
3580 0 : ereport(ERROR,
3581 : (errcode_for_file_access(),
3582 : errmsg("could not close file \"%s\": %m", path)));
3583 :
3584 : /*
3585 : * Now move the segment into place with its final name.
3586 : */
3587 43 : if (!InstallXLogFileSegment(&destsegno, tmppath, false, 0, destTLI))
3588 0 : elog(ERROR, "InstallXLogFileSegment should not have failed");
3589 43 : }
3590 :
3591 : /*
3592 : * Install a new XLOG segment file as a current or future log segment.
3593 : *
3594 : * This is used both to install a newly-created segment (which has a temp
3595 : * filename while it's being created) and to recycle an old segment.
3596 : *
3597 : * *segno: identify segment to install as (or first possible target).
3598 : * When find_free is true, this is modified on return to indicate the
3599 : * actual installation location or last segment searched.
3600 : *
3601 : * tmppath: initial name of file to install. It will be renamed into place.
3602 : *
3603 : * find_free: if true, install the new segment at the first empty segno
3604 : * number at or after the passed numbers. If false, install the new segment
3605 : * exactly where specified, deleting any existing segment file there.
3606 : *
3607 : * max_segno: maximum segment number to install the new file as. Fail if no
3608 : * free slot is found between *segno and max_segno. (Ignored when find_free
3609 : * is false.)
3610 : *
3611 : * tli: The timeline on which the new segment should be installed.
3612 : *
3613 : * Returns true if the file was installed successfully. false indicates that
3614 : * max_segno limit was exceeded, the startup process has disabled this
3615 : * function for now, or an error occurred while renaming the file into place.
3616 : */
3617 : static bool
3618 3250 : InstallXLogFileSegment(XLogSegNo *segno, char *tmppath,
3619 : bool find_free, XLogSegNo max_segno, TimeLineID tli)
3620 : {
3621 : char path[MAXPGPATH];
3622 : struct stat stat_buf;
3623 :
3624 : Assert(tli != 0);
3625 :
3626 3250 : XLogFilePath(path, tli, *segno, wal_segment_size);
3627 :
3628 3250 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3629 3250 : if (!XLogCtl->InstallXLogFileSegmentActive)
3630 : {
3631 0 : LWLockRelease(ControlFileLock);
3632 0 : return false;
3633 : }
3634 :
3635 3250 : if (!find_free)
3636 : {
3637 : /* Force installation: get rid of any pre-existing segment file */
3638 43 : durable_unlink(path, DEBUG1);
3639 : }
3640 : else
3641 : {
3642 : /* Find a free slot to put it in */
3643 4402 : while (stat(path, &stat_buf) == 0)
3644 : {
3645 1381 : if ((*segno) >= max_segno)
3646 : {
3647 : /* Failed to find a free slot within specified range */
3648 186 : LWLockRelease(ControlFileLock);
3649 186 : return false;
3650 : }
3651 1195 : (*segno)++;
3652 1195 : XLogFilePath(path, tli, *segno, wal_segment_size);
3653 : }
3654 : }
3655 :
3656 : Assert(access(path, F_OK) != 0 && errno == ENOENT);
3657 3064 : if (durable_rename(tmppath, path, LOG) != 0)
3658 : {
3659 0 : LWLockRelease(ControlFileLock);
3660 : /* durable_rename already emitted log message */
3661 0 : return false;
3662 : }
3663 :
3664 3064 : LWLockRelease(ControlFileLock);
3665 :
3666 3064 : return true;
3667 : }
3668 :
3669 : /*
3670 : * Open a pre-existing logfile segment for writing.
3671 : */
3672 : int
3673 99 : XLogFileOpen(XLogSegNo segno, TimeLineID tli)
3674 : {
3675 : char path[MAXPGPATH];
3676 : int fd;
3677 :
3678 99 : XLogFilePath(path, tli, segno, wal_segment_size);
3679 :
3680 99 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3681 99 : get_sync_bit(wal_sync_method));
3682 99 : if (fd < 0)
3683 0 : ereport(PANIC,
3684 : (errcode_for_file_access(),
3685 : errmsg("could not open file \"%s\": %m", path)));
3686 :
3687 99 : return fd;
3688 : }
3689 :
3690 : /*
3691 : * Close the current logfile segment for writing.
3692 : */
3693 : static void
3694 6877 : XLogFileClose(void)
3695 : {
3696 : Assert(openLogFile >= 0);
3697 :
3698 : /*
3699 : * WAL segment files will not be re-read in normal operation, so we advise
3700 : * the OS to release any cached pages. But do not do so if WAL archiving
3701 : * or streaming is active, because archiver and walsender process could
3702 : * use the cache to read the WAL segment.
3703 : */
3704 : #if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
3705 6877 : if (!XLogIsNeeded() && (io_direct_flags & IO_DIRECT_WAL) == 0)
3706 132 : (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3707 : #endif
3708 :
3709 6877 : if (close(openLogFile) != 0)
3710 : {
3711 : char xlogfname[MAXFNAMELEN];
3712 0 : int save_errno = errno;
3713 :
3714 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo, wal_segment_size);
3715 0 : errno = save_errno;
3716 0 : ereport(PANIC,
3717 : (errcode_for_file_access(),
3718 : errmsg("could not close file \"%s\": %m", xlogfname)));
3719 : }
3720 :
3721 6877 : openLogFile = -1;
3722 6877 : ReleaseExternalFD();
3723 6877 : }
3724 :
3725 : /*
3726 : * Preallocate log files beyond the specified log endpoint.
3727 : *
3728 : * XXX this is currently extremely conservative, since it forces only one
3729 : * future log segment to exist, and even that only if we are 75% done with
3730 : * the current one. This is only appropriate for very low-WAL-volume systems.
3731 : * High-volume systems will be OK once they've built up a sufficient set of
3732 : * recycled log segments, but the startup transient is likely to include
3733 : * a lot of segment creations by foreground processes, which is not so good.
3734 : *
3735 : * XLogFileInitInternal() can ereport(ERROR). All known causes indicate big
3736 : * trouble; for example, a full filesystem is one cause. The checkpoint WAL
3737 : * and/or ControlFile updates already completed. If a RequestCheckpoint()
3738 : * initiated the present checkpoint and an ERROR ends this function, the
3739 : * command that called RequestCheckpoint() fails. That's not ideal, but it's
3740 : * not worth contorting more functions to use caller-specified elevel values.
3741 : * (With or without RequestCheckpoint(), an ERROR forestalls some inessential
3742 : * reporting and resource reclamation.)
3743 : */
3744 : static void
3745 2214 : PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli)
3746 : {
3747 : XLogSegNo _logSegNo;
3748 : int lf;
3749 : bool added;
3750 : char path[MAXPGPATH];
3751 : uint64 offset;
3752 :
3753 2214 : if (!XLogCtl->InstallXLogFileSegmentActive)
3754 13 : return; /* unlocked check says no */
3755 :
3756 2201 : XLByteToPrevSeg(endptr, _logSegNo, wal_segment_size);
3757 2201 : offset = XLogSegmentOffset(endptr - 1, wal_segment_size);
3758 2201 : if (offset >= (uint32) (0.75 * wal_segment_size))
3759 : {
3760 222 : _logSegNo++;
3761 222 : lf = XLogFileInitInternal(_logSegNo, tli, &added, path);
3762 222 : if (lf >= 0)
3763 164 : close(lf);
3764 222 : if (added)
3765 58 : CheckpointStats.ckpt_segs_added++;
3766 : }
3767 : }
3768 :
3769 : /*
3770 : * Throws an error if the given log segment has already been removed or
3771 : * recycled. The caller should only pass a segment that it knows to have
3772 : * existed while the server has been running, as this function always
3773 : * succeeds if no WAL segments have been removed since startup.
3774 : * 'tli' is only used in the error message.
3775 : *
3776 : * Note: this function guarantees to keep errno unchanged on return.
3777 : * This supports callers that use this to possibly deliver a better
3778 : * error message about a missing file, while still being able to throw
3779 : * a normal file-access error afterwards, if this does return.
3780 : */
3781 : void
3782 133593 : CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3783 : {
3784 133593 : int save_errno = errno;
3785 : XLogSegNo lastRemovedSegNo;
3786 :
3787 133593 : SpinLockAcquire(&XLogCtl->info_lck);
3788 133593 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3789 133593 : SpinLockRelease(&XLogCtl->info_lck);
3790 :
3791 133593 : if (segno <= lastRemovedSegNo)
3792 : {
3793 : char filename[MAXFNAMELEN];
3794 :
3795 0 : XLogFileName(filename, tli, segno, wal_segment_size);
3796 0 : errno = save_errno;
3797 0 : ereport(ERROR,
3798 : (errcode_for_file_access(),
3799 : errmsg("requested WAL segment %s has already been removed",
3800 : filename)));
3801 : }
3802 133593 : errno = save_errno;
3803 133593 : }
3804 :
3805 : /*
3806 : * Return the last WAL segment removed, or 0 if no segment has been removed
3807 : * since startup.
3808 : *
3809 : * NB: the result can be out of date arbitrarily fast, the caller has to deal
3810 : * with that.
3811 : */
3812 : XLogSegNo
3813 1285 : XLogGetLastRemovedSegno(void)
3814 : {
3815 : XLogSegNo lastRemovedSegNo;
3816 :
3817 1285 : SpinLockAcquire(&XLogCtl->info_lck);
3818 1285 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3819 1285 : SpinLockRelease(&XLogCtl->info_lck);
3820 :
3821 1285 : return lastRemovedSegNo;
3822 : }
3823 :
3824 : /*
3825 : * Return the oldest WAL segment on the given TLI that still exists in
3826 : * XLOGDIR, or 0 if none.
3827 : */
3828 : XLogSegNo
3829 7 : XLogGetOldestSegno(TimeLineID tli)
3830 : {
3831 : DIR *xldir;
3832 : struct dirent *xlde;
3833 7 : XLogSegNo oldest_segno = 0;
3834 :
3835 7 : xldir = AllocateDir(XLOGDIR);
3836 50 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3837 : {
3838 : TimeLineID file_tli;
3839 : XLogSegNo file_segno;
3840 :
3841 : /* Ignore files that are not XLOG segments. */
3842 43 : if (!IsXLogFileName(xlde->d_name))
3843 29 : continue;
3844 :
3845 : /* Parse filename to get TLI and segno. */
3846 14 : XLogFromFileName(xlde->d_name, &file_tli, &file_segno,
3847 : wal_segment_size);
3848 :
3849 : /* Ignore anything that's not from the TLI of interest. */
3850 14 : if (tli != file_tli)
3851 0 : continue;
3852 :
3853 : /* If it's the oldest so far, update oldest_segno. */
3854 14 : if (oldest_segno == 0 || file_segno < oldest_segno)
3855 9 : oldest_segno = file_segno;
3856 : }
3857 :
3858 7 : FreeDir(xldir);
3859 7 : return oldest_segno;
3860 : }
3861 :
3862 : /*
3863 : * Update the last removed segno pointer in shared memory, to reflect that the
3864 : * given XLOG file has been removed.
3865 : */
3866 : static void
3867 2615 : UpdateLastRemovedPtr(char *filename)
3868 : {
3869 : uint32 tli;
3870 : XLogSegNo segno;
3871 :
3872 2615 : XLogFromFileName(filename, &tli, &segno, wal_segment_size);
3873 :
3874 2615 : SpinLockAcquire(&XLogCtl->info_lck);
3875 2615 : if (segno > XLogCtl->lastRemovedSegNo)
3876 1128 : XLogCtl->lastRemovedSegNo = segno;
3877 2615 : SpinLockRelease(&XLogCtl->info_lck);
3878 2615 : }
3879 :
3880 : /*
3881 : * Remove all temporary log files in pg_wal
3882 : *
3883 : * This is called at the beginning of recovery after a previous crash,
3884 : * at a point where no other processes write fresh WAL data.
3885 : */
3886 : static void
3887 188 : RemoveTempXlogFiles(void)
3888 : {
3889 : DIR *xldir;
3890 : struct dirent *xlde;
3891 :
3892 188 : elog(DEBUG2, "removing all temporary WAL segments");
3893 :
3894 188 : xldir = AllocateDir(XLOGDIR);
3895 1271 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3896 : {
3897 : char path[MAXPGPATH];
3898 :
3899 1083 : if (strncmp(xlde->d_name, "xlogtemp.", 9) != 0)
3900 1083 : continue;
3901 :
3902 0 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
3903 0 : unlink(path);
3904 0 : elog(DEBUG2, "removed temporary WAL segment \"%s\"", path);
3905 : }
3906 188 : FreeDir(xldir);
3907 188 : }
3908 :
3909 : /*
3910 : * Recycle or remove all log files older or equal to passed segno.
3911 : *
3912 : * endptr is current (or recent) end of xlog, and lastredoptr is the
3913 : * redo pointer of the last checkpoint. These are used to determine
3914 : * whether we want to recycle rather than delete no-longer-wanted log files.
3915 : *
3916 : * insertTLI is the current timeline for XLOG insertion. Any recycled
3917 : * segments should be reused for this timeline.
3918 : */
3919 : static void
3920 1936 : RemoveOldXlogFiles(XLogSegNo segno, XLogRecPtr lastredoptr, XLogRecPtr endptr,
3921 : TimeLineID insertTLI)
3922 : {
3923 : DIR *xldir;
3924 : struct dirent *xlde;
3925 : char lastoff[MAXFNAMELEN];
3926 : XLogSegNo endlogSegNo;
3927 : XLogSegNo recycleSegNo;
3928 :
3929 : /* Initialize info about where to try to recycle to */
3930 1936 : XLByteToSeg(endptr, endlogSegNo, wal_segment_size);
3931 1936 : recycleSegNo = XLOGfileslop(lastredoptr);
3932 :
3933 : /*
3934 : * Construct a filename of the last segment to be kept. The timeline ID
3935 : * doesn't matter, we ignore that in the comparison. (During recovery,
3936 : * InsertTimeLineID isn't set, so we can't use that.)
3937 : */
3938 1936 : XLogFileName(lastoff, 0, segno, wal_segment_size);
3939 :
3940 1936 : elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
3941 : lastoff);
3942 :
3943 1936 : xldir = AllocateDir(XLOGDIR);
3944 :
3945 54328 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3946 : {
3947 : /* Ignore files that are not XLOG segments */
3948 52392 : if (!IsXLogFileName(xlde->d_name) &&
3949 8211 : !IsPartialXLogFileName(xlde->d_name))
3950 8209 : continue;
3951 :
3952 : /*
3953 : * We ignore the timeline part of the XLOG segment identifiers in
3954 : * deciding whether a segment is still needed. This ensures that we
3955 : * won't prematurely remove a segment from a parent timeline. We could
3956 : * probably be a little more proactive about removing segments of
3957 : * non-parent timelines, but that would be a whole lot more
3958 : * complicated.
3959 : *
3960 : * We use the alphanumeric sorting property of the filenames to decide
3961 : * which ones are earlier than the lastoff segment.
3962 : */
3963 44183 : if (strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
3964 : {
3965 37685 : if (XLogArchiveCheckDone(xlde->d_name))
3966 : {
3967 : /* Update the last removed location in shared memory first */
3968 2615 : UpdateLastRemovedPtr(xlde->d_name);
3969 :
3970 2615 : RemoveXlogFile(xlde, recycleSegNo, &endlogSegNo, insertTLI);
3971 : }
3972 : }
3973 : }
3974 :
3975 1936 : FreeDir(xldir);
3976 1936 : }
3977 :
3978 : /*
3979 : * Recycle or remove WAL files that are not part of the given timeline's
3980 : * history.
3981 : *
3982 : * This is called during recovery, whenever we switch to follow a new
3983 : * timeline, and at the end of recovery when we create a new timeline. We
3984 : * wouldn't otherwise care about extra WAL files lying in pg_wal, but they
3985 : * might be leftover pre-allocated or recycled WAL segments on the old timeline
3986 : * that we haven't used yet, and contain garbage. If we just leave them in
3987 : * pg_wal, they will eventually be archived, and we can't let that happen.
3988 : * Files that belong to our timeline history are valid, because we have
3989 : * successfully replayed them, but from others we can't be sure.
3990 : *
3991 : * 'switchpoint' is the current point in WAL where we switch to new timeline,
3992 : * and 'newTLI' is the new timeline we switch to.
3993 : */
3994 : void
3995 66 : RemoveNonParentXlogFiles(XLogRecPtr switchpoint, TimeLineID newTLI)
3996 : {
3997 : DIR *xldir;
3998 : struct dirent *xlde;
3999 : char switchseg[MAXFNAMELEN];
4000 : XLogSegNo endLogSegNo;
4001 : XLogSegNo switchLogSegNo;
4002 : XLogSegNo recycleSegNo;
4003 :
4004 : /*
4005 : * Initialize info about where to begin the work. This will recycle,
4006 : * somewhat arbitrarily, 10 future segments.
4007 : */
4008 66 : XLByteToPrevSeg(switchpoint, switchLogSegNo, wal_segment_size);
4009 66 : XLByteToSeg(switchpoint, endLogSegNo, wal_segment_size);
4010 66 : recycleSegNo = endLogSegNo + 10;
4011 :
4012 : /*
4013 : * Construct a filename of the last segment to be kept.
4014 : */
4015 66 : XLogFileName(switchseg, newTLI, switchLogSegNo, wal_segment_size);
4016 :
4017 66 : elog(DEBUG2, "attempting to remove WAL segments newer than log file %s",
4018 : switchseg);
4019 :
4020 66 : xldir = AllocateDir(XLOGDIR);
4021 :
4022 625 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4023 : {
4024 : /* Ignore files that are not XLOG segments */
4025 559 : if (!IsXLogFileName(xlde->d_name))
4026 348 : continue;
4027 :
4028 : /*
4029 : * Remove files that are on a timeline older than the new one we're
4030 : * switching to, but with a segment number >= the first segment on the
4031 : * new timeline.
4032 : */
4033 211 : if (strncmp(xlde->d_name, switchseg, 8) < 0 &&
4034 138 : strcmp(xlde->d_name + 8, switchseg + 8) > 0)
4035 : {
4036 : /*
4037 : * If the file has already been marked as .ready, however, don't
4038 : * remove it yet. It should be OK to remove it - files that are
4039 : * not part of our timeline history are not required for recovery
4040 : * - but seems safer to let them be archived and removed later.
4041 : */
4042 16 : if (!XLogArchiveIsReady(xlde->d_name))
4043 16 : RemoveXlogFile(xlde, recycleSegNo, &endLogSegNo, newTLI);
4044 : }
4045 : }
4046 :
4047 66 : FreeDir(xldir);
4048 66 : }
4049 :
4050 : /*
4051 : * Recycle or remove a log file that's no longer needed.
4052 : *
4053 : * segment_de is the dirent structure of the segment to recycle or remove.
4054 : * recycleSegNo is the segment number to recycle up to. endlogSegNo is
4055 : * the segment number of the current (or recent) end of WAL.
4056 : *
4057 : * endlogSegNo gets incremented if the segment is recycled so as it is not
4058 : * checked again with future callers of this function.
4059 : *
4060 : * insertTLI is the current timeline for XLOG insertion. Any recycled segments
4061 : * should be used for this timeline.
4062 : */
4063 : static void
4064 2631 : RemoveXlogFile(const struct dirent *segment_de,
4065 : XLogSegNo recycleSegNo, XLogSegNo *endlogSegNo,
4066 : TimeLineID insertTLI)
4067 : {
4068 : char path[MAXPGPATH];
4069 : #ifdef WIN32
4070 : char newpath[MAXPGPATH];
4071 : #endif
4072 2631 : const char *segname = segment_de->d_name;
4073 :
4074 2631 : snprintf(path, MAXPGPATH, XLOGDIR "/%s", segname);
4075 :
4076 : /*
4077 : * Before deleting the file, see if it can be recycled as a future log
4078 : * segment. Only recycle normal files, because we don't want to recycle
4079 : * symbolic links pointing to a separate archive directory.
4080 : */
4081 2631 : if (wal_recycle &&
4082 2631 : *endlogSegNo <= recycleSegNo &&
4083 3566 : XLogCtl->InstallXLogFileSegmentActive && /* callee rechecks this */
4084 3236 : get_dirent_type(path, segment_de, false, DEBUG2) == PGFILETYPE_REG &&
4085 1618 : InstallXLogFileSegment(endlogSegNo, path,
4086 : true, recycleSegNo, insertTLI))
4087 : {
4088 1432 : ereport(DEBUG2,
4089 : (errmsg_internal("recycled write-ahead log file \"%s\"",
4090 : segname)));
4091 1432 : CheckpointStats.ckpt_segs_recycled++;
4092 : /* Needn't recheck that slot on future iterations */
4093 1432 : (*endlogSegNo)++;
4094 : }
4095 : else
4096 : {
4097 : /* No need for any more future segments, or recycling failed ... */
4098 : int rc;
4099 :
4100 1199 : ereport(DEBUG2,
4101 : (errmsg_internal("removing write-ahead log file \"%s\"",
4102 : segname)));
4103 :
4104 : #ifdef WIN32
4105 :
4106 : /*
4107 : * On Windows, if another process (e.g another backend) holds the file
4108 : * open in FILE_SHARE_DELETE mode, unlink will succeed, but the file
4109 : * will still show up in directory listing until the last handle is
4110 : * closed. To avoid confusing the lingering deleted file for a live
4111 : * WAL file that needs to be archived, rename it before deleting it.
4112 : *
4113 : * If another process holds the file open without FILE_SHARE_DELETE
4114 : * flag, rename will fail. We'll try again at the next checkpoint.
4115 : */
4116 : snprintf(newpath, MAXPGPATH, "%s.deleted", path);
4117 : if (rename(path, newpath) != 0)
4118 : {
4119 : ereport(LOG,
4120 : (errcode_for_file_access(),
4121 : errmsg("could not rename file \"%s\": %m",
4122 : path)));
4123 : return;
4124 : }
4125 : rc = durable_unlink(newpath, LOG);
4126 : #else
4127 1199 : rc = durable_unlink(path, LOG);
4128 : #endif
4129 1199 : if (rc != 0)
4130 : {
4131 : /* Message already logged by durable_unlink() */
4132 0 : return;
4133 : }
4134 1199 : CheckpointStats.ckpt_segs_removed++;
4135 : }
4136 :
4137 2631 : XLogArchiveCleanup(segname);
4138 : }
4139 :
4140 : /*
4141 : * Verify whether pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
4142 : * If the latter do not exist, recreate them.
4143 : *
4144 : * It is not the goal of this function to verify the contents of these
4145 : * directories, but to help in cases where someone has performed a cluster
4146 : * copy for PITR purposes but omitted pg_wal from the copy.
4147 : *
4148 : * We could also recreate pg_wal if it doesn't exist, but a deliberate
4149 : * policy decision was made not to. It is fairly common for pg_wal to be
4150 : * a symlink, and if that was the DBA's intent then automatically making a
4151 : * plain directory would result in degraded performance with no notice.
4152 : */
4153 : static void
4154 1079 : ValidateXLOGDirectoryStructure(void)
4155 : {
4156 : char path[MAXPGPATH];
4157 : struct stat stat_buf;
4158 :
4159 : /* Check for pg_wal; if it doesn't exist, error out */
4160 1079 : if (stat(XLOGDIR, &stat_buf) != 0 ||
4161 1079 : !S_ISDIR(stat_buf.st_mode))
4162 0 : ereport(FATAL,
4163 : (errcode_for_file_access(),
4164 : errmsg("required WAL directory \"%s\" does not exist",
4165 : XLOGDIR)));
4166 :
4167 : /* Check for archive_status */
4168 1079 : snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4169 1079 : if (stat(path, &stat_buf) == 0)
4170 : {
4171 : /* Check for weird cases where it exists but isn't a directory */
4172 1078 : if (!S_ISDIR(stat_buf.st_mode))
4173 0 : ereport(FATAL,
4174 : (errcode_for_file_access(),
4175 : errmsg("required WAL directory \"%s\" does not exist",
4176 : path)));
4177 : }
4178 : else
4179 : {
4180 1 : ereport(LOG,
4181 : (errmsg("creating missing WAL directory \"%s\"", path)));
4182 1 : if (MakePGDirectory(path) < 0)
4183 0 : ereport(FATAL,
4184 : (errcode_for_file_access(),
4185 : errmsg("could not create missing directory \"%s\": %m",
4186 : path)));
4187 : }
4188 :
4189 : /* Check for summaries */
4190 1079 : snprintf(path, MAXPGPATH, XLOGDIR "/summaries");
4191 1079 : if (stat(path, &stat_buf) == 0)
4192 : {
4193 : /* Check for weird cases where it exists but isn't a directory */
4194 1078 : if (!S_ISDIR(stat_buf.st_mode))
4195 0 : ereport(FATAL,
4196 : (errmsg("required WAL directory \"%s\" does not exist",
4197 : path)));
4198 : }
4199 : else
4200 : {
4201 1 : ereport(LOG,
4202 : (errmsg("creating missing WAL directory \"%s\"", path)));
4203 1 : if (MakePGDirectory(path) < 0)
4204 0 : ereport(FATAL,
4205 : (errmsg("could not create missing directory \"%s\": %m",
4206 : path)));
4207 : }
4208 1079 : }
4209 :
4210 : /*
4211 : * Remove previous backup history files. This also retries creation of
4212 : * .ready files for any backup history files for which XLogArchiveNotify
4213 : * failed earlier.
4214 : */
4215 : static void
4216 162 : CleanupBackupHistory(void)
4217 : {
4218 : DIR *xldir;
4219 : struct dirent *xlde;
4220 : char path[MAXPGPATH + sizeof(XLOGDIR)];
4221 :
4222 162 : xldir = AllocateDir(XLOGDIR);
4223 :
4224 1671 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4225 : {
4226 1347 : if (IsBackupHistoryFileName(xlde->d_name))
4227 : {
4228 171 : if (XLogArchiveCheckDone(xlde->d_name))
4229 : {
4230 136 : elog(DEBUG2, "removing WAL backup history file \"%s\"",
4231 : xlde->d_name);
4232 136 : snprintf(path, sizeof(path), XLOGDIR "/%s", xlde->d_name);
4233 136 : unlink(path);
4234 136 : XLogArchiveCleanup(xlde->d_name);
4235 : }
4236 : }
4237 : }
4238 :
4239 162 : FreeDir(xldir);
4240 162 : }
4241 :
4242 : /*
4243 : * I/O routines for pg_control
4244 : *
4245 : * *ControlFile is a buffer in shared memory that holds an image of the
4246 : * contents of pg_control. WriteControlFile() initializes pg_control
4247 : * given a preloaded buffer, ReadControlFile() loads the buffer from
4248 : * the pg_control file (during postmaster or standalone-backend startup),
4249 : * and UpdateControlFile() rewrites pg_control after we modify xlog state.
4250 : * InitControlFile() fills the buffer with initial values.
4251 : *
4252 : * For simplicity, WriteControlFile() initializes the fields of pg_control
4253 : * that are related to checking backend/database compatibility, and
4254 : * ReadControlFile() verifies they are correct. We could split out the
4255 : * I/O and compatibility-check functions, but there seems no need currently.
4256 : */
4257 :
4258 : static void
4259 57 : InitControlFile(uint64 sysidentifier, uint32 data_checksum_version)
4260 : {
4261 : char mock_auth_nonce[MOCK_AUTH_NONCE_LEN];
4262 :
4263 : /*
4264 : * Generate a random nonce. This is used for authentication requests that
4265 : * will fail because the user does not exist. The nonce is used to create
4266 : * a genuine-looking password challenge for the non-existent user, in lieu
4267 : * of an actual stored password.
4268 : */
4269 57 : if (!pg_strong_random(mock_auth_nonce, MOCK_AUTH_NONCE_LEN))
4270 0 : ereport(PANIC,
4271 : (errcode(ERRCODE_INTERNAL_ERROR),
4272 : errmsg("could not generate secret authorization token")));
4273 :
4274 57 : memset(ControlFile, 0, sizeof(ControlFileData));
4275 : /* Initialize pg_control status fields */
4276 57 : ControlFile->system_identifier = sysidentifier;
4277 57 : memcpy(ControlFile->mock_authentication_nonce, mock_auth_nonce, MOCK_AUTH_NONCE_LEN);
4278 57 : ControlFile->state = DB_SHUTDOWNED;
4279 57 : ControlFile->unloggedLSN = FirstNormalUnloggedLSN;
4280 :
4281 : /* Set important parameter values for use when replaying WAL */
4282 57 : ControlFile->MaxConnections = MaxConnections;
4283 57 : ControlFile->max_worker_processes = max_worker_processes;
4284 57 : ControlFile->max_wal_senders = max_wal_senders;
4285 57 : ControlFile->max_prepared_xacts = max_prepared_xacts;
4286 57 : ControlFile->max_locks_per_xact = max_locks_per_xact;
4287 57 : ControlFile->wal_level = wal_level;
4288 57 : ControlFile->wal_log_hints = wal_log_hints;
4289 57 : ControlFile->track_commit_timestamp = track_commit_timestamp;
4290 57 : ControlFile->data_checksum_version = data_checksum_version;
4291 :
4292 : /*
4293 : * Set the data_checksum_version value into XLogCtl, which is where all
4294 : * processes get the current value from.
4295 : */
4296 57 : XLogCtl->data_checksum_version = data_checksum_version;
4297 57 : }
4298 :
4299 : static void
4300 57 : WriteControlFile(void)
4301 : {
4302 : int fd;
4303 : char buffer[PG_CONTROL_FILE_SIZE]; /* need not be aligned */
4304 :
4305 : /*
4306 : * Initialize version and compatibility-check fields
4307 : */
4308 57 : ControlFile->pg_control_version = PG_CONTROL_VERSION;
4309 57 : ControlFile->catalog_version_no = CATALOG_VERSION_NO;
4310 :
4311 57 : ControlFile->maxAlign = MAXIMUM_ALIGNOF;
4312 57 : ControlFile->floatFormat = FLOATFORMAT_VALUE;
4313 :
4314 57 : ControlFile->blcksz = BLCKSZ;
4315 57 : ControlFile->relseg_size = RELSEG_SIZE;
4316 57 : ControlFile->slru_pages_per_segment = SLRU_PAGES_PER_SEGMENT;
4317 57 : ControlFile->xlog_blcksz = XLOG_BLCKSZ;
4318 57 : ControlFile->xlog_seg_size = wal_segment_size;
4319 :
4320 57 : ControlFile->nameDataLen = NAMEDATALEN;
4321 57 : ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
4322 :
4323 57 : ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
4324 57 : ControlFile->loblksize = LOBLKSIZE;
4325 :
4326 57 : ControlFile->float8ByVal = true; /* vestigial */
4327 :
4328 : /*
4329 : * Initialize the default 'char' signedness.
4330 : *
4331 : * The signedness of the char type is implementation-defined. For instance
4332 : * on x86 architecture CPUs, the char data type is typically treated as
4333 : * signed by default, whereas on aarch architecture CPUs, it is typically
4334 : * treated as unsigned by default. In v17 or earlier, we accidentally let
4335 : * C implementation signedness affect persistent data. This led to
4336 : * inconsistent results when comparing char data across different
4337 : * platforms.
4338 : *
4339 : * This flag can be used as a hint to ensure consistent behavior for
4340 : * pre-v18 data files that store data sorted by the 'char' type on disk,
4341 : * especially in cross-platform replication scenarios.
4342 : *
4343 : * Newly created database clusters unconditionally set the default char
4344 : * signedness to true. pg_upgrade changes this flag for clusters that were
4345 : * initialized on signedness=false platforms. As a result,
4346 : * signedness=false setting will become rare over time. If we had known
4347 : * about this problem during the last development cycle that forced initdb
4348 : * (v8.3), we would have made all clusters signed or all clusters
4349 : * unsigned. Making pg_upgrade the only source of signedness=false will
4350 : * cause the population of database clusters to converge toward that
4351 : * retrospective ideal.
4352 : */
4353 57 : ControlFile->default_char_signedness = true;
4354 :
4355 : /* Contents are protected with a CRC */
4356 57 : INIT_CRC32C(ControlFile->crc);
4357 57 : COMP_CRC32C(ControlFile->crc,
4358 : ControlFile,
4359 : offsetof(ControlFileData, crc));
4360 57 : FIN_CRC32C(ControlFile->crc);
4361 :
4362 : /*
4363 : * We write out PG_CONTROL_FILE_SIZE bytes into pg_control, zero-padding
4364 : * the excess over sizeof(ControlFileData). This reduces the odds of
4365 : * premature-EOF errors when reading pg_control. We'll still fail when we
4366 : * check the contents of the file, but hopefully with a more specific
4367 : * error than "couldn't read pg_control".
4368 : */
4369 57 : memset(buffer, 0, PG_CONTROL_FILE_SIZE);
4370 57 : memcpy(buffer, ControlFile, sizeof(ControlFileData));
4371 :
4372 57 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4373 : O_RDWR | O_CREAT | O_EXCL | PG_BINARY);
4374 57 : if (fd < 0)
4375 0 : ereport(PANIC,
4376 : (errcode_for_file_access(),
4377 : errmsg("could not create file \"%s\": %m",
4378 : XLOG_CONTROL_FILE)));
4379 :
4380 57 : errno = 0;
4381 57 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_WRITE);
4382 57 : if (write(fd, buffer, PG_CONTROL_FILE_SIZE) != PG_CONTROL_FILE_SIZE)
4383 : {
4384 : /* if write didn't set errno, assume problem is no disk space */
4385 0 : if (errno == 0)
4386 0 : errno = ENOSPC;
4387 0 : ereport(PANIC,
4388 : (errcode_for_file_access(),
4389 : errmsg("could not write to file \"%s\": %m",
4390 : XLOG_CONTROL_FILE)));
4391 : }
4392 57 : pgstat_report_wait_end();
4393 :
4394 57 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_SYNC);
4395 57 : if (pg_fsync(fd) != 0)
4396 0 : ereport(PANIC,
4397 : (errcode_for_file_access(),
4398 : errmsg("could not fsync file \"%s\": %m",
4399 : XLOG_CONTROL_FILE)));
4400 57 : pgstat_report_wait_end();
4401 :
4402 57 : if (close(fd) != 0)
4403 0 : ereport(PANIC,
4404 : (errcode_for_file_access(),
4405 : errmsg("could not close file \"%s\": %m",
4406 : XLOG_CONTROL_FILE)));
4407 57 : }
4408 :
4409 : static void
4410 1138 : ReadControlFile(void)
4411 : {
4412 : pg_crc32c crc;
4413 : int fd;
4414 : char wal_segsz_str[20];
4415 : int r;
4416 :
4417 : /*
4418 : * Read data...
4419 : */
4420 1138 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4421 : O_RDWR | PG_BINARY);
4422 1138 : if (fd < 0)
4423 0 : ereport(PANIC,
4424 : (errcode_for_file_access(),
4425 : errmsg("could not open file \"%s\": %m",
4426 : XLOG_CONTROL_FILE)));
4427 :
4428 1138 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ);
4429 1138 : r = read(fd, ControlFile, sizeof(ControlFileData));
4430 1138 : if (r != sizeof(ControlFileData))
4431 : {
4432 0 : if (r < 0)
4433 0 : ereport(PANIC,
4434 : (errcode_for_file_access(),
4435 : errmsg("could not read file \"%s\": %m",
4436 : XLOG_CONTROL_FILE)));
4437 : else
4438 0 : ereport(PANIC,
4439 : (errcode(ERRCODE_DATA_CORRUPTED),
4440 : errmsg("could not read file \"%s\": read %d of %zu",
4441 : XLOG_CONTROL_FILE, r, sizeof(ControlFileData))));
4442 : }
4443 1138 : pgstat_report_wait_end();
4444 :
4445 1138 : close(fd);
4446 :
4447 : /*
4448 : * Check for expected pg_control format version. If this is wrong, the
4449 : * CRC check will likely fail because we'll be checking the wrong number
4450 : * of bytes. Complaining about wrong version will probably be more
4451 : * enlightening than complaining about wrong CRC.
4452 : */
4453 :
4454 1138 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
4455 0 : ereport(FATAL,
4456 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4457 : errmsg("database files are incompatible with server"),
4458 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
4459 : " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
4460 : ControlFile->pg_control_version, ControlFile->pg_control_version,
4461 : PG_CONTROL_VERSION, PG_CONTROL_VERSION),
4462 : errhint("This could be a problem of mismatched byte ordering. It looks like you need to initdb.")));
4463 :
4464 1138 : if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
4465 0 : ereport(FATAL,
4466 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4467 : errmsg("database files are incompatible with server"),
4468 : errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
4469 : " but the server was compiled with PG_CONTROL_VERSION %d.",
4470 : ControlFile->pg_control_version, PG_CONTROL_VERSION),
4471 : errhint("It looks like you need to initdb.")));
4472 :
4473 : /* Now check the CRC. */
4474 1138 : INIT_CRC32C(crc);
4475 1138 : COMP_CRC32C(crc,
4476 : ControlFile,
4477 : offsetof(ControlFileData, crc));
4478 1138 : FIN_CRC32C(crc);
4479 :
4480 1138 : if (!EQ_CRC32C(crc, ControlFile->crc))
4481 0 : ereport(FATAL,
4482 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4483 : errmsg("incorrect checksum in control file")));
4484 :
4485 : /*
4486 : * Do compatibility checking immediately. If the database isn't
4487 : * compatible with the backend executable, we want to abort before we can
4488 : * possibly do any damage.
4489 : */
4490 1138 : if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
4491 0 : ereport(FATAL,
4492 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4493 : errmsg("database files are incompatible with server"),
4494 : /* translator: %s is a variable name and %d is its value */
4495 : errdetail("The database cluster was initialized with %s %d,"
4496 : " but the server was compiled with %s %d.",
4497 : "CATALOG_VERSION_NO", ControlFile->catalog_version_no,
4498 : "CATALOG_VERSION_NO", CATALOG_VERSION_NO),
4499 : errhint("It looks like you need to initdb.")));
4500 1138 : if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
4501 0 : ereport(FATAL,
4502 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4503 : errmsg("database files are incompatible with server"),
4504 : /* translator: %s is a variable name and %d is its value */
4505 : errdetail("The database cluster was initialized with %s %d,"
4506 : " but the server was compiled with %s %d.",
4507 : "MAXALIGN", ControlFile->maxAlign,
4508 : "MAXALIGN", MAXIMUM_ALIGNOF),
4509 : errhint("It looks like you need to initdb.")));
4510 1138 : if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
4511 0 : ereport(FATAL,
4512 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4513 : errmsg("database files are incompatible with server"),
4514 : errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
4515 : errhint("It looks like you need to initdb.")));
4516 1138 : if (ControlFile->blcksz != BLCKSZ)
4517 0 : ereport(FATAL,
4518 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4519 : errmsg("database files are incompatible with server"),
4520 : /* translator: %s is a variable name and %d is its value */
4521 : errdetail("The database cluster was initialized with %s %d,"
4522 : " but the server was compiled with %s %d.",
4523 : "BLCKSZ", ControlFile->blcksz,
4524 : "BLCKSZ", BLCKSZ),
4525 : errhint("It looks like you need to recompile or initdb.")));
4526 1138 : if (ControlFile->relseg_size != RELSEG_SIZE)
4527 0 : ereport(FATAL,
4528 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4529 : errmsg("database files are incompatible with server"),
4530 : /* translator: %s is a variable name and %d is its value */
4531 : errdetail("The database cluster was initialized with %s %d,"
4532 : " but the server was compiled with %s %d.",
4533 : "RELSEG_SIZE", ControlFile->relseg_size,
4534 : "RELSEG_SIZE", RELSEG_SIZE),
4535 : errhint("It looks like you need to recompile or initdb.")));
4536 1138 : if (ControlFile->slru_pages_per_segment != SLRU_PAGES_PER_SEGMENT)
4537 0 : ereport(FATAL,
4538 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4539 : errmsg("database files are incompatible with server"),
4540 : /* translator: %s is a variable name and %d is its value */
4541 : errdetail("The database cluster was initialized with %s %d,"
4542 : " but the server was compiled with %s %d.",
4543 : "SLRU_PAGES_PER_SEGMENT", ControlFile->slru_pages_per_segment,
4544 : "SLRU_PAGES_PER_SEGMENT", SLRU_PAGES_PER_SEGMENT),
4545 : errhint("It looks like you need to recompile or initdb.")));
4546 1138 : if (ControlFile->xlog_blcksz != XLOG_BLCKSZ)
4547 0 : ereport(FATAL,
4548 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4549 : errmsg("database files are incompatible with server"),
4550 : /* translator: %s is a variable name and %d is its value */
4551 : errdetail("The database cluster was initialized with %s %d,"
4552 : " but the server was compiled with %s %d.",
4553 : "XLOG_BLCKSZ", ControlFile->xlog_blcksz,
4554 : "XLOG_BLCKSZ", XLOG_BLCKSZ),
4555 : errhint("It looks like you need to recompile or initdb.")));
4556 1138 : if (ControlFile->nameDataLen != NAMEDATALEN)
4557 0 : ereport(FATAL,
4558 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4559 : errmsg("database files are incompatible with server"),
4560 : /* translator: %s is a variable name and %d is its value */
4561 : errdetail("The database cluster was initialized with %s %d,"
4562 : " but the server was compiled with %s %d.",
4563 : "NAMEDATALEN", ControlFile->nameDataLen,
4564 : "NAMEDATALEN", NAMEDATALEN),
4565 : errhint("It looks like you need to recompile or initdb.")));
4566 1138 : if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
4567 0 : ereport(FATAL,
4568 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4569 : errmsg("database files are incompatible with server"),
4570 : /* translator: %s is a variable name and %d is its value */
4571 : errdetail("The database cluster was initialized with %s %d,"
4572 : " but the server was compiled with %s %d.",
4573 : "INDEX_MAX_KEYS", ControlFile->indexMaxKeys,
4574 : "INDEX_MAX_KEYS", INDEX_MAX_KEYS),
4575 : errhint("It looks like you need to recompile or initdb.")));
4576 1138 : if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
4577 0 : ereport(FATAL,
4578 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4579 : errmsg("database files are incompatible with server"),
4580 : /* translator: %s is a variable name and %d is its value */
4581 : errdetail("The database cluster was initialized with %s %d,"
4582 : " but the server was compiled with %s %d.",
4583 : "TOAST_MAX_CHUNK_SIZE", ControlFile->toast_max_chunk_size,
4584 : "TOAST_MAX_CHUNK_SIZE", (int) TOAST_MAX_CHUNK_SIZE),
4585 : errhint("It looks like you need to recompile or initdb.")));
4586 1138 : if (ControlFile->loblksize != LOBLKSIZE)
4587 0 : ereport(FATAL,
4588 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
4589 : errmsg("database files are incompatible with server"),
4590 : /* translator: %s is a variable name and %d is its value */
4591 : errdetail("The database cluster was initialized with %s %d,"
4592 : " but the server was compiled with %s %d.",
4593 : "LOBLKSIZE", ControlFile->loblksize,
4594 : "LOBLKSIZE", (int) LOBLKSIZE),
4595 : errhint("It looks like you need to recompile or initdb.")));
4596 :
4597 : Assert(ControlFile->float8ByVal); /* vestigial, not worth an error msg */
4598 :
4599 1138 : wal_segment_size = ControlFile->xlog_seg_size;
4600 :
4601 1138 : if (!IsValidWalSegSize(wal_segment_size))
4602 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4603 : errmsg_plural("invalid WAL segment size in control file (%d byte)",
4604 : "invalid WAL segment size in control file (%d bytes)",
4605 : wal_segment_size,
4606 : wal_segment_size),
4607 : errdetail("The WAL segment size must be a power of two between 1 MB and 1 GB.")));
4608 :
4609 1138 : snprintf(wal_segsz_str, sizeof(wal_segsz_str), "%d", wal_segment_size);
4610 1138 : SetConfigOption("wal_segment_size", wal_segsz_str, PGC_INTERNAL,
4611 : PGC_S_DYNAMIC_DEFAULT);
4612 :
4613 : /* check and update variables dependent on wal_segment_size */
4614 1138 : if (ConvertToXSegs(min_wal_size_mb, wal_segment_size) < 2)
4615 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4616 : /* translator: both %s are GUC names */
4617 : errmsg("\"%s\" must be at least twice \"%s\"",
4618 : "min_wal_size", "wal_segment_size")));
4619 :
4620 1138 : if (ConvertToXSegs(max_wal_size_mb, wal_segment_size) < 2)
4621 0 : ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
4622 : /* translator: both %s are GUC names */
4623 : errmsg("\"%s\" must be at least twice \"%s\"",
4624 : "max_wal_size", "wal_segment_size")));
4625 :
4626 1138 : UsableBytesInSegment =
4627 1138 : (wal_segment_size / XLOG_BLCKSZ * UsableBytesInPage) -
4628 : (SizeOfXLogLongPHD - SizeOfXLogShortPHD);
4629 :
4630 1138 : CalculateCheckpointSegments();
4631 1138 : }
4632 :
4633 : /*
4634 : * Utility wrapper to update the control file. Note that the control
4635 : * file gets flushed.
4636 : */
4637 : static void
4638 10187 : UpdateControlFile(void)
4639 : {
4640 10187 : update_controlfile(DataDir, ControlFile, true);
4641 10187 : }
4642 :
4643 : /*
4644 : * Returns the unique system identifier from control file.
4645 : */
4646 : uint64
4647 1566 : GetSystemIdentifier(void)
4648 : {
4649 : Assert(ControlFile != NULL);
4650 1566 : return ControlFile->system_identifier;
4651 : }
4652 :
4653 : /*
4654 : * Returns the random nonce from control file.
4655 : */
4656 : char *
4657 2 : GetMockAuthenticationNonce(void)
4658 : {
4659 : Assert(ControlFile != NULL);
4660 2 : return ControlFile->mock_authentication_nonce;
4661 : }
4662 :
4663 : /*
4664 : * DataChecksumsNeedWrite
4665 : * Returns whether data checksums must be written or not
4666 : *
4667 : * Returns true if data checksums are enabled, or are in the process of being
4668 : * enabled. During "inprogress-on" and "inprogress-off" states checksums must
4669 : * be written even though they are not verified (see datachecksum_state.c for
4670 : * a longer discussion).
4671 : *
4672 : * This function is intended for callsites which are about to write a data page
4673 : * to storage, and need to know whether to re-calculate the checksum for the
4674 : * page header. Calling this function must be performed as close to the write
4675 : * operation as possible to keep the critical section short.
4676 : */
4677 : bool
4678 842434 : DataChecksumsNeedWrite(void)
4679 : {
4680 932900 : return (LocalDataChecksumState == PG_DATA_CHECKSUM_VERSION ||
4681 897764 : LocalDataChecksumState == PG_DATA_CHECKSUM_INPROGRESS_ON ||
4682 55330 : LocalDataChecksumState == PG_DATA_CHECKSUM_INPROGRESS_OFF);
4683 : }
4684 :
4685 :
4686 : bool
4687 6 : DataChecksumsOff(void)
4688 : {
4689 : bool ret;
4690 :
4691 6 : SpinLockAcquire(&XLogCtl->info_lck);
4692 6 : ret = (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_OFF);
4693 6 : SpinLockRelease(&XLogCtl->info_lck);
4694 :
4695 6 : return ret;
4696 : }
4697 :
4698 : bool
4699 9 : DataChecksumsOn(void)
4700 : {
4701 : bool ret;
4702 :
4703 9 : SpinLockAcquire(&XLogCtl->info_lck);
4704 9 : ret = (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_VERSION);
4705 9 : SpinLockRelease(&XLogCtl->info_lck);
4706 :
4707 9 : return ret;
4708 : }
4709 :
4710 : bool
4711 164 : DataChecksumsInProgressOn(void)
4712 : {
4713 : bool ret;
4714 :
4715 164 : SpinLockAcquire(&XLogCtl->info_lck);
4716 164 : ret = (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_INPROGRESS_ON);
4717 164 : SpinLockRelease(&XLogCtl->info_lck);
4718 :
4719 164 : return ret;
4720 : }
4721 :
4722 : /*
4723 : * DataChecksumsNeedVerify
4724 : * Returns whether data checksums must be verified or not
4725 : *
4726 : * Data checksums are only verified if they are fully enabled in the cluster.
4727 : * During the "inprogress-on" and "inprogress-off" states they are only
4728 : * updated, not verified (see datachecksum_state.c for a longer discussion).
4729 : *
4730 : * This function is intended for callsites which have read data and are about
4731 : * to perform checksum validation based on the result of this. Calling this
4732 : * function must be performed as close to the validation call as possible to
4733 : * keep the critical section short. This is in order to protect against time of
4734 : * check/time of use situations around data checksum validation.
4735 : */
4736 : bool
4737 2629342 : DataChecksumsNeedVerify(void)
4738 : {
4739 2629342 : return (LocalDataChecksumState == PG_DATA_CHECKSUM_VERSION);
4740 : }
4741 :
4742 : /*
4743 : * SetDataChecksumsOnInProgress
4744 : * Sets the data checksum state to "inprogress-on" to enable checksums
4745 : *
4746 : * To start the process of enabling data checksums in a running cluster the
4747 : * data_checksum_version state must be changed to "inprogress-on". See
4748 : * SetDataChecksumsOn below for a description on how this state change works.
4749 : * This function blocks until all backends in the cluster have acknowledged the
4750 : * state transition.
4751 : */
4752 : void
4753 8 : SetDataChecksumsOnInProgress(void)
4754 : {
4755 : uint64 barrier;
4756 :
4757 : /*
4758 : * The state transition is performed in a critical section with
4759 : * checkpoints held off to provide crash safety.
4760 : */
4761 8 : START_CRIT_SECTION();
4762 8 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
4763 :
4764 8 : XLogChecksums(PG_DATA_CHECKSUM_INPROGRESS_ON);
4765 :
4766 8 : SpinLockAcquire(&XLogCtl->info_lck);
4767 8 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_INPROGRESS_ON;
4768 8 : SpinLockRelease(&XLogCtl->info_lck);
4769 :
4770 8 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
4771 8 : ControlFile->data_checksum_version = PG_DATA_CHECKSUM_INPROGRESS_ON;
4772 8 : UpdateControlFile();
4773 8 : LWLockRelease(ControlFileLock);
4774 :
4775 8 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_ON);
4776 :
4777 8 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
4778 8 : END_CRIT_SECTION();
4779 :
4780 8 : WaitForProcSignalBarrier(barrier);
4781 8 : }
4782 :
4783 : /*
4784 : * SetDataChecksumsOn
4785 : * Set data checksums state to 'on' cluster-wide
4786 : *
4787 : * Enabling data checksums is performed using two barriers, the first one to
4788 : * set the state to "inprogress-on" (done by SetDataChecksumsOnInProgress())
4789 : * and the second one to set the state to "on" (done here). Below is a short
4790 : * description of the processing, a more detailed write-up can be found in
4791 : * datachecksum_state.c.
4792 : *
4793 : * To start the process of enabling data checksums in a running cluster the
4794 : * data_checksum_version state must be changed to "inprogress-on". This state
4795 : * requires data checksums to be written but not verified. This ensures that
4796 : * all data pages can be checksummed without the risk of false negatives in
4797 : * validation during the process. When all existing pages are guaranteed to
4798 : * have checksums, and all new pages will be initiated with checksums, the
4799 : * state can be changed to "on". Once the state is "on" checksums will be both
4800 : * written and verified.
4801 : *
4802 : * This function blocks until all backends in the cluster have acknowledged the
4803 : * state transition.
4804 : */
4805 : void
4806 6 : SetDataChecksumsOn(void)
4807 : {
4808 : uint64 barrier;
4809 :
4810 6 : SpinLockAcquire(&XLogCtl->info_lck);
4811 :
4812 : /*
4813 : * The only allowed state transition to "on" is from "inprogress-on" since
4814 : * that state ensures that all pages will have data checksums written. No
4815 : * such state transition exists, if it does happen it's likely due to a
4816 : * programmer error.
4817 : */
4818 6 : if (XLogCtl->data_checksum_version != PG_DATA_CHECKSUM_INPROGRESS_ON)
4819 : {
4820 0 : SpinLockRelease(&XLogCtl->info_lck);
4821 0 : elog(WARNING,
4822 : "cannot set data checksums to \"on\", current state is not \"inprogress-on\", disabling");
4823 0 : SetDataChecksumsOff();
4824 0 : return;
4825 : }
4826 :
4827 6 : SpinLockRelease(&XLogCtl->info_lck);
4828 :
4829 6 : INJECTION_POINT("datachecksums-enable-checksums-delay", NULL);
4830 6 : START_CRIT_SECTION();
4831 6 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
4832 :
4833 6 : XLogChecksums(PG_DATA_CHECKSUM_VERSION);
4834 :
4835 6 : SpinLockAcquire(&XLogCtl->info_lck);
4836 6 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_VERSION;
4837 6 : SpinLockRelease(&XLogCtl->info_lck);
4838 :
4839 : /*
4840 : * Update the controlfile before waiting since if we have an immediate
4841 : * shutdown while waiting we want to come back up with checksums enabled.
4842 : */
4843 6 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
4844 6 : ControlFile->data_checksum_version = PG_DATA_CHECKSUM_VERSION;
4845 6 : UpdateControlFile();
4846 6 : LWLockRelease(ControlFileLock);
4847 :
4848 6 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_ON);
4849 :
4850 6 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
4851 6 : END_CRIT_SECTION();
4852 :
4853 6 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT | CHECKPOINT_FAST);
4854 6 : WaitForProcSignalBarrier(barrier);
4855 : }
4856 :
4857 : /*
4858 : * SetDataChecksumsOff
4859 : * Disables data checksums cluster-wide
4860 : *
4861 : * Disabling data checksums must be performed with two sets of barriers, each
4862 : * carrying a different state. The state is first set to "inprogress-off"
4863 : * during which checksums are still written but not verified. This ensures that
4864 : * backends which have yet to observe the state change from "on" won't get
4865 : * validation errors on concurrently modified pages. Once all backends have
4866 : * changed to "inprogress-off", the barrier for moving to "off" can be emitted.
4867 : * This function blocks until all backends in the cluster have acknowledged the
4868 : * state transition.
4869 : */
4870 : void
4871 6 : SetDataChecksumsOff(void)
4872 : {
4873 : uint64 barrier;
4874 :
4875 6 : SpinLockAcquire(&XLogCtl->info_lck);
4876 :
4877 : /* If data checksums are already disabled there is nothing to do */
4878 6 : if (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_OFF)
4879 : {
4880 0 : SpinLockRelease(&XLogCtl->info_lck);
4881 0 : return;
4882 : }
4883 :
4884 : /*
4885 : * If data checksums are currently enabled, or in the process of being
4886 : * enabled, we first transition to the "inprogress-off" state during which
4887 : * backends continue to write checksums without verifying them. When all
4888 : * backends are in "inprogress-off" the next transition to "off" can be
4889 : * performed, after which all data checksum processing is disabled.
4890 : */
4891 6 : if (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_VERSION ||
4892 2 : XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_INPROGRESS_ON)
4893 : {
4894 6 : SpinLockRelease(&XLogCtl->info_lck);
4895 :
4896 6 : START_CRIT_SECTION();
4897 6 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
4898 :
4899 6 : XLogChecksums(PG_DATA_CHECKSUM_INPROGRESS_OFF);
4900 :
4901 6 : SpinLockAcquire(&XLogCtl->info_lck);
4902 6 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_INPROGRESS_OFF;
4903 6 : SpinLockRelease(&XLogCtl->info_lck);
4904 :
4905 6 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
4906 6 : ControlFile->data_checksum_version = PG_DATA_CHECKSUM_INPROGRESS_OFF;
4907 6 : UpdateControlFile();
4908 6 : LWLockRelease(ControlFileLock);
4909 :
4910 6 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_OFF);
4911 :
4912 6 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
4913 6 : END_CRIT_SECTION();
4914 :
4915 6 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT | CHECKPOINT_FAST);
4916 6 : WaitForProcSignalBarrier(barrier);
4917 :
4918 : /*
4919 : * At this point we know that no backends are verifying data checksums
4920 : * during reading. Next, we can safely move to state "off" to also
4921 : * stop writing checksums.
4922 : */
4923 : }
4924 : else
4925 : {
4926 : /*
4927 : * Ending up here implies that the checksums state is "inprogress-off"
4928 : * and we can transition directly to "off" from there.
4929 : */
4930 0 : SpinLockRelease(&XLogCtl->info_lck);
4931 : }
4932 :
4933 6 : START_CRIT_SECTION();
4934 : /* Ensure that we don't incur a checkpoint during disabling checksums */
4935 6 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
4936 :
4937 6 : XLogChecksums(PG_DATA_CHECKSUM_OFF);
4938 :
4939 6 : SpinLockAcquire(&XLogCtl->info_lck);
4940 6 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_OFF;
4941 6 : SpinLockRelease(&XLogCtl->info_lck);
4942 :
4943 6 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
4944 6 : ControlFile->data_checksum_version = PG_DATA_CHECKSUM_OFF;
4945 6 : UpdateControlFile();
4946 6 : LWLockRelease(ControlFileLock);
4947 :
4948 6 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_OFF);
4949 :
4950 6 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
4951 6 : END_CRIT_SECTION();
4952 :
4953 6 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT | CHECKPOINT_FAST);
4954 6 : WaitForProcSignalBarrier(barrier);
4955 : }
4956 :
4957 : /*
4958 : * InitLocalDataChecksumState
4959 : *
4960 : * Set up backend local caches of controldata variables which may change at
4961 : * any point during runtime and thus require special cased locking. So far
4962 : * this only applies to data_checksum_version, but it's intended to be general
4963 : * purpose enough to handle future cases.
4964 : */
4965 : void
4966 24191 : InitLocalDataChecksumState(void)
4967 : {
4968 : Assert(InterruptHoldoffCount > 0);
4969 24191 : SpinLockAcquire(&XLogCtl->info_lck);
4970 24191 : SetLocalDataChecksumState(XLogCtl->data_checksum_version);
4971 24191 : SpinLockRelease(&XLogCtl->info_lck);
4972 24191 : }
4973 :
4974 : void
4975 27441 : SetLocalDataChecksumState(uint32 data_checksum_version)
4976 : {
4977 27441 : LocalDataChecksumState = data_checksum_version;
4978 :
4979 27441 : data_checksums = data_checksum_version;
4980 27441 : }
4981 :
4982 : /* guc hook */
4983 : const char *
4984 1958 : show_data_checksums(void)
4985 : {
4986 1958 : return get_checksum_state_string(LocalDataChecksumState);
4987 : }
4988 :
4989 : /*
4990 : * Return true if the cluster was initialized on a platform where the
4991 : * default signedness of char is "signed". This function exists for code
4992 : * that deals with pre-v18 data files that store data sorted by the 'char'
4993 : * type on disk (e.g., GIN and GiST indexes). See the comments in
4994 : * WriteControlFile() for details.
4995 : */
4996 : bool
4997 89903 : GetDefaultCharSignedness(void)
4998 : {
4999 89903 : return ControlFile->default_char_signedness;
5000 : }
5001 :
5002 : /*
5003 : * Returns a fake LSN for unlogged relations.
5004 : *
5005 : * Each call generates an LSN that is greater than any previous value
5006 : * returned. The current counter value is saved and restored across clean
5007 : * shutdowns, but like unlogged relations, does not survive a crash. This can
5008 : * be used in lieu of real LSN values returned by XLogInsert, if you need an
5009 : * LSN-like increasing sequence of numbers without writing any WAL.
5010 : */
5011 : XLogRecPtr
5012 201459 : GetFakeLSNForUnloggedRel(void)
5013 : {
5014 201459 : return pg_atomic_fetch_add_u64(&XLogCtl->unloggedLSN, 1);
5015 : }
5016 :
5017 : /*
5018 : * Auto-tune the number of XLOG buffers.
5019 : *
5020 : * The preferred setting for wal_buffers is about 3% of shared_buffers, with
5021 : * a maximum of one XLOG segment (there is little reason to think that more
5022 : * is helpful, at least so long as we force an fsync when switching log files)
5023 : * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
5024 : * 9.1, when auto-tuning was added).
5025 : *
5026 : * This should not be called until NBuffers has received its final value.
5027 : */
5028 : static int
5029 1234 : XLOGChooseNumBuffers(void)
5030 : {
5031 : int xbuffers;
5032 :
5033 1234 : xbuffers = NBuffers / 32;
5034 1234 : if (xbuffers > (wal_segment_size / XLOG_BLCKSZ))
5035 24 : xbuffers = (wal_segment_size / XLOG_BLCKSZ);
5036 1234 : if (xbuffers < 8)
5037 457 : xbuffers = 8;
5038 1234 : return xbuffers;
5039 : }
5040 :
5041 : /*
5042 : * GUC check_hook for wal_buffers
5043 : */
5044 : bool
5045 2514 : check_wal_buffers(int *newval, void **extra, GucSource source)
5046 : {
5047 : /*
5048 : * -1 indicates a request for auto-tune.
5049 : */
5050 2514 : if (*newval == -1)
5051 : {
5052 : /*
5053 : * If we haven't yet changed the boot_val default of -1, just let it
5054 : * be. We'll fix it when XLOGShmemRequest is called.
5055 : */
5056 1280 : if (XLOGbuffers == -1)
5057 1280 : return true;
5058 :
5059 : /* Otherwise, substitute the auto-tune value */
5060 0 : *newval = XLOGChooseNumBuffers();
5061 : }
5062 :
5063 : /*
5064 : * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
5065 : * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
5066 : * the case, we just silently treat such values as a request for the
5067 : * minimum. (We could throw an error instead, but that doesn't seem very
5068 : * helpful.)
5069 : */
5070 1234 : if (*newval < 4)
5071 0 : *newval = 4;
5072 :
5073 1234 : return true;
5074 : }
5075 :
5076 : /*
5077 : * GUC check_hook for wal_consistency_checking
5078 : */
5079 : bool
5080 2274 : check_wal_consistency_checking(char **newval, void **extra, GucSource source)
5081 : {
5082 : char *rawstring;
5083 : List *elemlist;
5084 : ListCell *l;
5085 : bool newwalconsistency[RM_MAX_ID + 1];
5086 :
5087 : /* Initialize the array */
5088 75042 : MemSet(newwalconsistency, 0, (RM_MAX_ID + 1) * sizeof(bool));
5089 :
5090 : /* Need a modifiable copy of string */
5091 2274 : rawstring = pstrdup(*newval);
5092 :
5093 : /* Parse string into list of identifiers */
5094 2274 : if (!SplitIdentifierString(rawstring, ',', &elemlist))
5095 : {
5096 : /* syntax error in list */
5097 0 : GUC_check_errdetail("List syntax is invalid.");
5098 0 : pfree(rawstring);
5099 0 : list_free(elemlist);
5100 0 : return false;
5101 : }
5102 :
5103 2774 : foreach(l, elemlist)
5104 : {
5105 500 : char *tok = (char *) lfirst(l);
5106 : int rmid;
5107 :
5108 : /* Check for 'all'. */
5109 500 : if (pg_strcasecmp(tok, "all") == 0)
5110 : {
5111 127986 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
5112 127488 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL)
5113 4980 : newwalconsistency[rmid] = true;
5114 : }
5115 : else
5116 : {
5117 : /* Check if the token matches any known resource manager. */
5118 2 : bool found = false;
5119 :
5120 36 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
5121 : {
5122 54 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL &&
5123 18 : pg_strcasecmp(tok, GetRmgr(rmid).rm_name) == 0)
5124 : {
5125 2 : newwalconsistency[rmid] = true;
5126 2 : found = true;
5127 2 : break;
5128 : }
5129 : }
5130 2 : if (!found)
5131 : {
5132 : /*
5133 : * During startup, it might be a not-yet-loaded custom
5134 : * resource manager. Defer checking until
5135 : * InitializeWalConsistencyChecking().
5136 : */
5137 0 : if (!process_shared_preload_libraries_done)
5138 : {
5139 0 : check_wal_consistency_checking_deferred = true;
5140 : }
5141 : else
5142 : {
5143 0 : GUC_check_errdetail("Unrecognized key word: \"%s\".", tok);
5144 0 : pfree(rawstring);
5145 0 : list_free(elemlist);
5146 0 : return false;
5147 : }
5148 : }
5149 : }
5150 : }
5151 :
5152 2274 : pfree(rawstring);
5153 2274 : list_free(elemlist);
5154 :
5155 : /* assign new value */
5156 2274 : *extra = guc_malloc(LOG, (RM_MAX_ID + 1) * sizeof(bool));
5157 2274 : if (!*extra)
5158 0 : return false;
5159 2274 : memcpy(*extra, newwalconsistency, (RM_MAX_ID + 1) * sizeof(bool));
5160 2274 : return true;
5161 : }
5162 :
5163 : /*
5164 : * GUC assign_hook for wal_consistency_checking
5165 : */
5166 : void
5167 2273 : assign_wal_consistency_checking(const char *newval, void *extra)
5168 : {
5169 : /*
5170 : * If some checks were deferred, it's possible that the checks will fail
5171 : * later during InitializeWalConsistencyChecking(). But in that case, the
5172 : * postmaster will exit anyway, so it's safe to proceed with the
5173 : * assignment.
5174 : *
5175 : * Any built-in resource managers specified are assigned immediately,
5176 : * which affects WAL created before shared_preload_libraries are
5177 : * processed. Any custom resource managers specified won't be assigned
5178 : * until after shared_preload_libraries are processed, but that's OK
5179 : * because WAL for a custom resource manager can't be written before the
5180 : * module is loaded anyway.
5181 : */
5182 2273 : wal_consistency_checking = extra;
5183 2273 : }
5184 :
5185 : /*
5186 : * InitializeWalConsistencyChecking: run after loading custom resource managers
5187 : *
5188 : * If any unknown resource managers were specified in the
5189 : * wal_consistency_checking GUC, processing was deferred. Now that
5190 : * shared_preload_libraries have been loaded, process wal_consistency_checking
5191 : * again.
5192 : */
5193 : void
5194 1063 : InitializeWalConsistencyChecking(void)
5195 : {
5196 : Assert(process_shared_preload_libraries_done);
5197 :
5198 1063 : if (check_wal_consistency_checking_deferred)
5199 : {
5200 : struct config_generic *guc;
5201 :
5202 0 : guc = find_option("wal_consistency_checking", false, false, ERROR);
5203 :
5204 0 : check_wal_consistency_checking_deferred = false;
5205 :
5206 0 : set_config_option_ext("wal_consistency_checking",
5207 : wal_consistency_checking_string,
5208 : guc->scontext, guc->source, guc->srole,
5209 : GUC_ACTION_SET, true, ERROR, false);
5210 :
5211 : /* checking should not be deferred again */
5212 : Assert(!check_wal_consistency_checking_deferred);
5213 : }
5214 1063 : }
5215 :
5216 : /*
5217 : * GUC show_hook for archive_command
5218 : */
5219 : const char *
5220 1956 : show_archive_command(void)
5221 : {
5222 1956 : if (XLogArchivingActive())
5223 109 : return XLogArchiveCommand;
5224 : else
5225 1847 : return "(disabled)";
5226 : }
5227 :
5228 : /*
5229 : * GUC show_hook for in_hot_standby
5230 : */
5231 : const char *
5232 17218 : show_in_hot_standby(void)
5233 : {
5234 : /*
5235 : * We display the actual state based on shared memory, so that this GUC
5236 : * reports up-to-date state if examined intra-query. The underlying
5237 : * variable (in_hot_standby_guc) changes only when we transmit a new value
5238 : * to the client.
5239 : */
5240 17218 : return RecoveryInProgress() ? "on" : "off";
5241 : }
5242 :
5243 : /*
5244 : * GUC show_hook for effective_wal_level
5245 : */
5246 : const char *
5247 1992 : show_effective_wal_level(void)
5248 : {
5249 1992 : if (wal_level == WAL_LEVEL_MINIMAL)
5250 299 : return "minimal";
5251 :
5252 : /*
5253 : * During recovery, effective_wal_level reflects the primary's
5254 : * configuration rather than the local wal_level value.
5255 : */
5256 1693 : if (RecoveryInProgress())
5257 27 : return IsXLogLogicalInfoEnabled() ? "logical" : "replica";
5258 :
5259 1666 : return XLogLogicalInfoActive() ? "logical" : "replica";
5260 : }
5261 :
5262 : /*
5263 : * Read the control file, set respective GUCs.
5264 : *
5265 : * This is to be called during startup, including a crash recovery cycle,
5266 : * unless in bootstrap mode, where no control file yet exists. As there's no
5267 : * usable shared memory yet (its sizing can depend on the contents of the
5268 : * control file!), first store the contents in local memory. XLOGShmemInit()
5269 : * will then copy it to shared memory later.
5270 : *
5271 : * reset just controls whether previous contents are to be expected (in the
5272 : * reset case, there's a dangling pointer into old shared memory), or not.
5273 : */
5274 : void
5275 1081 : LocalProcessControlFile(bool reset)
5276 : {
5277 : Assert(reset || ControlFile == NULL);
5278 1081 : LocalControlFile = palloc_object(ControlFileData);
5279 1081 : ControlFile = LocalControlFile;
5280 1081 : ReadControlFile();
5281 1081 : SetLocalDataChecksumState(ControlFile->data_checksum_version);
5282 1081 : }
5283 :
5284 : /*
5285 : * Get the wal_level from the control file. For a standby, this value should be
5286 : * considered as its active wal_level, because it may be different from what
5287 : * was originally configured on standby.
5288 : */
5289 : WalLevel
5290 0 : GetActiveWalLevelOnStandby(void)
5291 : {
5292 0 : return ControlFile->wal_level;
5293 : }
5294 :
5295 : /*
5296 : * Register shared memory for XLOG.
5297 : */
5298 : static void
5299 1239 : XLOGShmemRequest(void *arg)
5300 : {
5301 : Size size;
5302 :
5303 : /*
5304 : * If the value of wal_buffers is -1, use the preferred auto-tune value.
5305 : * This isn't an amazingly clean place to do this, but we must wait till
5306 : * NBuffers has received its final value, and must do it before using the
5307 : * value of XLOGbuffers to do anything important.
5308 : *
5309 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
5310 : * However, if the DBA explicitly set wal_buffers = -1 in the config file,
5311 : * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
5312 : * the matter with PGC_S_OVERRIDE.
5313 : */
5314 1239 : if (XLOGbuffers == -1)
5315 : {
5316 : char buf[32];
5317 :
5318 1234 : snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
5319 1234 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
5320 : PGC_S_DYNAMIC_DEFAULT);
5321 1234 : if (XLOGbuffers == -1) /* failed to apply it? */
5322 0 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
5323 : PGC_S_OVERRIDE);
5324 : }
5325 : Assert(XLOGbuffers > 0);
5326 :
5327 : /* XLogCtl */
5328 1239 : size = sizeof(XLogCtlData);
5329 :
5330 : /* WAL insertion locks, plus alignment */
5331 1239 : size = add_size(size, mul_size(sizeof(WALInsertLockPadded), NUM_XLOGINSERT_LOCKS + 1));
5332 : /* xlblocks array */
5333 1239 : size = add_size(size, mul_size(sizeof(pg_atomic_uint64), XLOGbuffers));
5334 : /* extra alignment padding for XLOG I/O buffers */
5335 1239 : size = add_size(size, Max(XLOG_BLCKSZ, PG_IO_ALIGN_SIZE));
5336 : /* and the buffers themselves */
5337 1239 : size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
5338 :
5339 1239 : ShmemRequestStruct(.name = "XLOG Ctl",
5340 : .size = size,
5341 : .ptr = (void **) &XLogCtl,
5342 : );
5343 1239 : ShmemRequestStruct(.name = "Control File",
5344 : .size = sizeof(ControlFileData),
5345 : .ptr = (void **) &ControlFile,
5346 : );
5347 1239 : }
5348 :
5349 : /*
5350 : * XLOGShmemInit - initialize the XLogCtl shared memory area.
5351 : */
5352 : static void
5353 1236 : XLOGShmemInit(void *arg)
5354 : {
5355 : char *allocptr;
5356 : int i;
5357 :
5358 : #ifdef WAL_DEBUG
5359 :
5360 : /*
5361 : * Create a memory context for WAL debugging that's exempt from the normal
5362 : * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
5363 : * an allocation fails, but wal_debug is not for production use anyway.
5364 : */
5365 : if (walDebugCxt == NULL)
5366 : {
5367 : walDebugCxt = AllocSetContextCreate(TopMemoryContext,
5368 : "WAL Debug",
5369 : ALLOCSET_DEFAULT_SIZES);
5370 : MemoryContextAllowInCriticalSection(walDebugCxt, true);
5371 : }
5372 : #endif
5373 :
5374 1236 : memset(XLogCtl, 0, sizeof(XLogCtlData));
5375 :
5376 : /*
5377 : * Already have read control file locally, unless in bootstrap mode. Move
5378 : * contents into shared memory.
5379 : */
5380 1236 : if (LocalControlFile)
5381 : {
5382 1065 : memcpy(ControlFile, LocalControlFile, sizeof(ControlFileData));
5383 1065 : pfree(LocalControlFile);
5384 1065 : LocalControlFile = NULL;
5385 : }
5386 :
5387 : /*
5388 : * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
5389 : * multiple of the alignment for same, so no extra alignment padding is
5390 : * needed here.
5391 : */
5392 1236 : allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
5393 1236 : XLogCtl->xlblocks = (pg_atomic_uint64 *) allocptr;
5394 1236 : allocptr += sizeof(pg_atomic_uint64) * XLOGbuffers;
5395 :
5396 365595 : for (i = 0; i < XLOGbuffers; i++)
5397 : {
5398 364359 : pg_atomic_init_u64(&XLogCtl->xlblocks[i], InvalidXLogRecPtr);
5399 : }
5400 :
5401 : /* WAL insertion locks. Ensure they're aligned to the full padded size */
5402 1236 : allocptr += sizeof(WALInsertLockPadded) -
5403 1236 : ((uintptr_t) allocptr) % sizeof(WALInsertLockPadded);
5404 1236 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks =
5405 : (WALInsertLockPadded *) allocptr;
5406 1236 : allocptr += sizeof(WALInsertLockPadded) * NUM_XLOGINSERT_LOCKS;
5407 :
5408 11124 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
5409 : {
5410 9888 : LWLockInitialize(&WALInsertLocks[i].l.lock, LWTRANCHE_WAL_INSERT);
5411 9888 : pg_atomic_init_u64(&WALInsertLocks[i].l.insertingAt, InvalidXLogRecPtr);
5412 9888 : WALInsertLocks[i].l.lastImportantAt = InvalidXLogRecPtr;
5413 : }
5414 :
5415 : /*
5416 : * Align the start of the page buffers to a full xlog block size boundary.
5417 : * This simplifies some calculations in XLOG insertion. It is also
5418 : * required for O_DIRECT.
5419 : */
5420 1236 : allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5421 1236 : XLogCtl->pages = allocptr;
5422 1236 : memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5423 :
5424 : /*
5425 : * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5426 : * in additional info.)
5427 : */
5428 1236 : XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5429 1236 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5430 1236 : XLogCtl->InstallXLogFileSegmentActive = false;
5431 1236 : XLogCtl->WalWriterSleeping = false;
5432 :
5433 : /* Use the checksum info from control file */
5434 1236 : XLogCtl->data_checksum_version = ControlFile->data_checksum_version;
5435 1236 : SetLocalDataChecksumState(XLogCtl->data_checksum_version);
5436 :
5437 1236 : SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5438 1236 : SpinLockInit(&XLogCtl->info_lck);
5439 1236 : pg_atomic_init_u64(&XLogCtl->logInsertResult, InvalidXLogRecPtr);
5440 1236 : pg_atomic_init_u64(&XLogCtl->logWriteResult, InvalidXLogRecPtr);
5441 1236 : pg_atomic_init_u64(&XLogCtl->logFlushResult, InvalidXLogRecPtr);
5442 1236 : pg_atomic_init_u64(&XLogCtl->unloggedLSN, InvalidXLogRecPtr);
5443 1236 : }
5444 :
5445 : /*
5446 : * XLOGShmemAttach - re-establish WALInsertLocks pointer after attaching.
5447 : */
5448 : static void
5449 0 : XLOGShmemAttach(void *arg)
5450 : {
5451 0 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks;
5452 0 : }
5453 :
5454 : /*
5455 : * This func must be called ONCE on system install. It creates pg_control
5456 : * and the initial XLOG segment.
5457 : */
5458 : void
5459 57 : BootStrapXLOG(uint32 data_checksum_version)
5460 : {
5461 : CheckPoint checkPoint;
5462 : PGAlignedXLogBlock buffer;
5463 : XLogPageHeader page;
5464 : XLogLongPageHeader longpage;
5465 : XLogRecord *record;
5466 : char *recptr;
5467 : uint64 sysidentifier;
5468 : struct timeval tv;
5469 : pg_crc32c crc;
5470 :
5471 : /* allow ordinary WAL segment creation, like StartupXLOG() would */
5472 57 : SetInstallXLogFileSegmentActive();
5473 :
5474 : /*
5475 : * Select a hopefully-unique system identifier code for this installation.
5476 : * We use the result of gettimeofday(), including the fractional seconds
5477 : * field, as being about as unique as we can easily get. (Think not to
5478 : * use random(), since it hasn't been seeded and there's no portable way
5479 : * to seed it other than the system clock value...) The upper half of the
5480 : * uint64 value is just the tv_sec part, while the lower half contains the
5481 : * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
5482 : * PID for a little extra uniqueness. A person knowing this encoding can
5483 : * determine the initialization time of the installation, which could
5484 : * perhaps be useful sometimes.
5485 : */
5486 57 : gettimeofday(&tv, NULL);
5487 57 : sysidentifier = ((uint64) tv.tv_sec) << 32;
5488 57 : sysidentifier |= ((uint64) tv.tv_usec) << 12;
5489 57 : sysidentifier |= getpid() & 0xFFF;
5490 :
5491 57 : memset(&buffer, 0, sizeof buffer);
5492 57 : page = (XLogPageHeader) &buffer;
5493 :
5494 : /*
5495 : * Set up information for the initial checkpoint record
5496 : *
5497 : * The initial checkpoint record is written to the beginning of the WAL
5498 : * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5499 : * used, so that we can use 0/0 to mean "before any valid WAL segment".
5500 : */
5501 57 : checkPoint.redo = wal_segment_size + SizeOfXLogLongPHD;
5502 57 : checkPoint.ThisTimeLineID = BootstrapTimeLineID;
5503 57 : checkPoint.PrevTimeLineID = BootstrapTimeLineID;
5504 57 : checkPoint.fullPageWrites = fullPageWrites;
5505 57 : checkPoint.logicalDecodingEnabled = (wal_level == WAL_LEVEL_LOGICAL);
5506 57 : checkPoint.wal_level = wal_level;
5507 : checkPoint.nextXid =
5508 57 : FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId);
5509 57 : checkPoint.nextOid = FirstGenbkiObjectId;
5510 57 : checkPoint.nextMulti = FirstMultiXactId;
5511 57 : checkPoint.nextMultiOffset = 1;
5512 57 : checkPoint.oldestXid = FirstNormalTransactionId;
5513 57 : checkPoint.oldestXidDB = Template1DbOid;
5514 57 : checkPoint.oldestMulti = FirstMultiXactId;
5515 57 : checkPoint.oldestMultiDB = Template1DbOid;
5516 57 : checkPoint.oldestCommitTsXid = InvalidTransactionId;
5517 57 : checkPoint.newestCommitTsXid = InvalidTransactionId;
5518 57 : checkPoint.time = (pg_time_t) time(NULL);
5519 57 : checkPoint.oldestActiveXid = InvalidTransactionId;
5520 57 : checkPoint.dataChecksumState = data_checksum_version;
5521 :
5522 57 : TransamVariables->nextXid = checkPoint.nextXid;
5523 57 : TransamVariables->nextOid = checkPoint.nextOid;
5524 57 : TransamVariables->oidCount = 0;
5525 57 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5526 57 : AdvanceOldestClogXid(checkPoint.oldestXid);
5527 57 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5528 57 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
5529 57 : SetCommitTsLimit(InvalidTransactionId, InvalidTransactionId);
5530 :
5531 : /* Set up the XLOG page header */
5532 57 : page->xlp_magic = XLOG_PAGE_MAGIC;
5533 57 : page->xlp_info = XLP_LONG_HEADER;
5534 57 : page->xlp_tli = BootstrapTimeLineID;
5535 57 : page->xlp_pageaddr = wal_segment_size;
5536 57 : longpage = (XLogLongPageHeader) page;
5537 57 : longpage->xlp_sysid = sysidentifier;
5538 57 : longpage->xlp_seg_size = wal_segment_size;
5539 57 : longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5540 :
5541 : /* Insert the initial checkpoint record */
5542 57 : recptr = ((char *) page + SizeOfXLogLongPHD);
5543 57 : record = (XLogRecord *) recptr;
5544 57 : record->xl_prev = InvalidXLogRecPtr;
5545 57 : record->xl_xid = InvalidTransactionId;
5546 57 : record->xl_tot_len = SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(checkPoint);
5547 57 : record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5548 57 : record->xl_rmid = RM_XLOG_ID;
5549 57 : recptr += SizeOfXLogRecord;
5550 : /* fill the XLogRecordDataHeaderShort struct */
5551 57 : *(recptr++) = (char) XLR_BLOCK_ID_DATA_SHORT;
5552 57 : *(recptr++) = sizeof(checkPoint);
5553 57 : memcpy(recptr, &checkPoint, sizeof(checkPoint));
5554 57 : recptr += sizeof(checkPoint);
5555 : Assert(recptr - (char *) record == record->xl_tot_len);
5556 :
5557 57 : INIT_CRC32C(crc);
5558 57 : COMP_CRC32C(crc, ((char *) record) + SizeOfXLogRecord, record->xl_tot_len - SizeOfXLogRecord);
5559 57 : COMP_CRC32C(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5560 57 : FIN_CRC32C(crc);
5561 57 : record->xl_crc = crc;
5562 :
5563 : /* Create first XLOG segment file */
5564 57 : openLogTLI = BootstrapTimeLineID;
5565 57 : openLogFile = XLogFileInit(1, BootstrapTimeLineID);
5566 :
5567 : /*
5568 : * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
5569 : * close the file again in a moment.
5570 : */
5571 :
5572 : /* Write the first page with the initial record */
5573 57 : errno = 0;
5574 57 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE);
5575 57 : if (write(openLogFile, &buffer, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5576 : {
5577 : /* if write didn't set errno, assume problem is no disk space */
5578 0 : if (errno == 0)
5579 0 : errno = ENOSPC;
5580 0 : ereport(PANIC,
5581 : (errcode_for_file_access(),
5582 : errmsg("could not write bootstrap write-ahead log file: %m")));
5583 : }
5584 57 : pgstat_report_wait_end();
5585 :
5586 57 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC);
5587 57 : if (pg_fsync(openLogFile) != 0)
5588 0 : ereport(PANIC,
5589 : (errcode_for_file_access(),
5590 : errmsg("could not fsync bootstrap write-ahead log file: %m")));
5591 57 : pgstat_report_wait_end();
5592 :
5593 57 : if (close(openLogFile) != 0)
5594 0 : ereport(PANIC,
5595 : (errcode_for_file_access(),
5596 : errmsg("could not close bootstrap write-ahead log file: %m")));
5597 :
5598 57 : openLogFile = -1;
5599 :
5600 : /* Now create pg_control */
5601 57 : InitControlFile(sysidentifier, data_checksum_version);
5602 57 : ControlFile->time = checkPoint.time;
5603 57 : ControlFile->checkPoint = checkPoint.redo;
5604 57 : ControlFile->checkPointCopy = checkPoint;
5605 :
5606 : /* some additional ControlFile fields are set in WriteControlFile() */
5607 57 : WriteControlFile();
5608 :
5609 : /* Bootstrap the commit log, too */
5610 57 : BootStrapCLOG();
5611 57 : BootStrapCommitTs();
5612 57 : BootStrapSUBTRANS();
5613 57 : BootStrapMultiXact();
5614 :
5615 : /*
5616 : * Force control file to be read - in contrast to normal processing we'd
5617 : * otherwise never run the checks and GUC related initializations therein.
5618 : */
5619 57 : ReadControlFile();
5620 57 : }
5621 :
5622 : static char *
5623 955 : str_time(pg_time_t tnow, char *buf, size_t bufsize)
5624 : {
5625 955 : pg_strftime(buf, bufsize,
5626 : "%Y-%m-%d %H:%M:%S %Z",
5627 955 : pg_localtime(&tnow, log_timezone));
5628 :
5629 955 : return buf;
5630 : }
5631 :
5632 : /*
5633 : * Initialize the first WAL segment on new timeline.
5634 : */
5635 : static void
5636 55 : XLogInitNewTimeline(TimeLineID endTLI, XLogRecPtr endOfLog, TimeLineID newTLI)
5637 : {
5638 : char xlogfname[MAXFNAMELEN];
5639 : XLogSegNo endLogSegNo;
5640 : XLogSegNo startLogSegNo;
5641 :
5642 : /* we always switch to a new timeline after archive recovery */
5643 : Assert(endTLI != newTLI);
5644 :
5645 : /*
5646 : * Update min recovery point one last time.
5647 : */
5648 55 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5649 :
5650 : /*
5651 : * Calculate the last segment on the old timeline, and the first segment
5652 : * on the new timeline. If the switch happens in the middle of a segment,
5653 : * they are the same, but if the switch happens exactly at a segment
5654 : * boundary, startLogSegNo will be endLogSegNo + 1.
5655 : */
5656 55 : XLByteToPrevSeg(endOfLog, endLogSegNo, wal_segment_size);
5657 55 : XLByteToSeg(endOfLog, startLogSegNo, wal_segment_size);
5658 :
5659 : /*
5660 : * Initialize the starting WAL segment for the new timeline. If the switch
5661 : * happens in the middle of a segment, copy data from the last WAL segment
5662 : * of the old timeline up to the switch point, to the starting WAL segment
5663 : * on the new timeline.
5664 : */
5665 55 : if (endLogSegNo == startLogSegNo)
5666 : {
5667 : /*
5668 : * Make a copy of the file on the new timeline.
5669 : *
5670 : * Writing WAL isn't allowed yet, so there are no locking
5671 : * considerations. But we should be just as tense as XLogFileInit to
5672 : * avoid emplacing a bogus file.
5673 : */
5674 43 : XLogFileCopy(newTLI, endLogSegNo, endTLI, endLogSegNo,
5675 43 : XLogSegmentOffset(endOfLog, wal_segment_size));
5676 : }
5677 : else
5678 : {
5679 : /*
5680 : * The switch happened at a segment boundary, so just create the next
5681 : * segment on the new timeline.
5682 : */
5683 : int fd;
5684 :
5685 12 : fd = XLogFileInit(startLogSegNo, newTLI);
5686 :
5687 12 : if (close(fd) != 0)
5688 : {
5689 0 : int save_errno = errno;
5690 :
5691 0 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5692 0 : errno = save_errno;
5693 0 : ereport(ERROR,
5694 : (errcode_for_file_access(),
5695 : errmsg("could not close file \"%s\": %m", xlogfname)));
5696 : }
5697 : }
5698 :
5699 : /*
5700 : * Let's just make real sure there are not .ready or .done flags posted
5701 : * for the new segment.
5702 : */
5703 55 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5704 55 : XLogArchiveCleanup(xlogfname);
5705 55 : }
5706 :
5707 : /*
5708 : * Perform cleanup actions at the conclusion of archive recovery.
5709 : */
5710 : static void
5711 55 : CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI, XLogRecPtr EndOfLog,
5712 : TimeLineID newTLI)
5713 : {
5714 : /*
5715 : * Execute the recovery_end_command, if any.
5716 : */
5717 55 : if (recoveryEndCommand && strcmp(recoveryEndCommand, "") != 0)
5718 2 : ExecuteRecoveryCommand(recoveryEndCommand,
5719 : "recovery_end_command",
5720 : true,
5721 : WAIT_EVENT_RECOVERY_END_COMMAND);
5722 :
5723 : /*
5724 : * We switched to a new timeline. Clean up segments on the old timeline.
5725 : *
5726 : * If there are any higher-numbered segments on the old timeline, remove
5727 : * them. They might contain valid WAL, but they might also be
5728 : * pre-allocated files containing garbage. In any case, they are not part
5729 : * of the new timeline's history so we don't need them.
5730 : */
5731 55 : RemoveNonParentXlogFiles(EndOfLog, newTLI);
5732 :
5733 : /*
5734 : * If the switch happened in the middle of a segment, what to do with the
5735 : * last, partial segment on the old timeline? If we don't archive it, and
5736 : * the server that created the WAL never archives it either (e.g. because
5737 : * it was hit by a meteor), it will never make it to the archive. That's
5738 : * OK from our point of view, because the new segment that we created with
5739 : * the new TLI contains all the WAL from the old timeline up to the switch
5740 : * point. But if you later try to do PITR to the "missing" WAL on the old
5741 : * timeline, recovery won't find it in the archive. It's physically
5742 : * present in the new file with new TLI, but recovery won't look there
5743 : * when it's recovering to the older timeline. On the other hand, if we
5744 : * archive the partial segment, and the original server on that timeline
5745 : * is still running and archives the completed version of the same segment
5746 : * later, it will fail. (We used to do that in 9.4 and below, and it
5747 : * caused such problems).
5748 : *
5749 : * As a compromise, we rename the last segment with the .partial suffix,
5750 : * and archive it. Archive recovery will never try to read .partial
5751 : * segments, so they will normally go unused. But in the odd PITR case,
5752 : * the administrator can copy them manually to the pg_wal directory
5753 : * (removing the suffix). They can be useful in debugging, too.
5754 : *
5755 : * If a .done or .ready file already exists for the old timeline, however,
5756 : * we had already determined that the segment is complete, so we can let
5757 : * it be archived normally. (In particular, if it was restored from the
5758 : * archive to begin with, it's expected to have a .done file).
5759 : */
5760 55 : if (XLogSegmentOffset(EndOfLog, wal_segment_size) != 0 &&
5761 : XLogArchivingActive())
5762 : {
5763 : char origfname[MAXFNAMELEN];
5764 : XLogSegNo endLogSegNo;
5765 :
5766 9 : XLByteToPrevSeg(EndOfLog, endLogSegNo, wal_segment_size);
5767 9 : XLogFileName(origfname, EndOfLogTLI, endLogSegNo, wal_segment_size);
5768 :
5769 9 : if (!XLogArchiveIsReadyOrDone(origfname))
5770 : {
5771 : char origpath[MAXPGPATH];
5772 : char partialfname[MAXFNAMELEN];
5773 : char partialpath[MAXPGPATH];
5774 :
5775 : /*
5776 : * If we're summarizing WAL, we can't rename the partial file
5777 : * until the summarizer finishes with it, else it will fail.
5778 : */
5779 5 : if (summarize_wal)
5780 1 : WaitForWalSummarization(EndOfLog);
5781 :
5782 5 : XLogFilePath(origpath, EndOfLogTLI, endLogSegNo, wal_segment_size);
5783 5 : snprintf(partialfname, MAXFNAMELEN, "%s.partial", origfname);
5784 5 : snprintf(partialpath, MAXPGPATH, "%s.partial", origpath);
5785 :
5786 : /*
5787 : * Make sure there's no .done or .ready file for the .partial
5788 : * file.
5789 : */
5790 5 : XLogArchiveCleanup(partialfname);
5791 :
5792 5 : durable_rename(origpath, partialpath, ERROR);
5793 5 : XLogArchiveNotify(partialfname);
5794 : }
5795 : }
5796 55 : }
5797 :
5798 : /*
5799 : * Check to see if required parameters are set high enough on this server
5800 : * for various aspects of recovery operation.
5801 : *
5802 : * Note that all the parameters which this function tests need to be
5803 : * listed in Administrator's Overview section in high-availability.sgml.
5804 : * If you change them, don't forget to update the list.
5805 : */
5806 : static void
5807 264 : CheckRequiredParameterValues(void)
5808 : {
5809 : /*
5810 : * For archive recovery, the WAL must be generated with at least 'replica'
5811 : * wal_level.
5812 : */
5813 264 : if (ArchiveRecoveryRequested && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5814 : {
5815 2 : ereport(FATAL,
5816 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5817 : errmsg("WAL was generated with \"wal_level=minimal\", cannot continue recovering"),
5818 : errdetail("This happens if you temporarily set \"wal_level=minimal\" on the server."),
5819 : errhint("Use a backup taken after setting \"wal_level\" to higher than \"minimal\".")));
5820 : }
5821 :
5822 : /*
5823 : * For Hot Standby, the WAL must be generated with 'replica' mode, and we
5824 : * must have at least as many backend slots as the primary.
5825 : */
5826 262 : if (ArchiveRecoveryRequested && EnableHotStandby)
5827 : {
5828 : /* We ignore autovacuum_worker_slots when we make this test. */
5829 138 : RecoveryRequiresIntParameter("max_connections",
5830 : MaxConnections,
5831 138 : ControlFile->MaxConnections);
5832 138 : RecoveryRequiresIntParameter("max_worker_processes",
5833 : max_worker_processes,
5834 138 : ControlFile->max_worker_processes);
5835 138 : RecoveryRequiresIntParameter("max_wal_senders",
5836 : max_wal_senders,
5837 138 : ControlFile->max_wal_senders);
5838 138 : RecoveryRequiresIntParameter("max_prepared_transactions",
5839 : max_prepared_xacts,
5840 138 : ControlFile->max_prepared_xacts);
5841 138 : RecoveryRequiresIntParameter("max_locks_per_transaction",
5842 : max_locks_per_xact,
5843 138 : ControlFile->max_locks_per_xact);
5844 : }
5845 262 : }
5846 :
5847 : /*
5848 : * This must be called ONCE during postmaster or standalone-backend startup
5849 : */
5850 : void
5851 1079 : StartupXLOG(void)
5852 : {
5853 : XLogCtlInsert *Insert;
5854 : CheckPoint checkPoint;
5855 : bool wasShutdown;
5856 : bool didCrash;
5857 : bool haveTblspcMap;
5858 : bool haveBackupLabel;
5859 : XLogRecPtr EndOfLog;
5860 : TimeLineID EndOfLogTLI;
5861 : TimeLineID newTLI;
5862 : bool performedWalRecovery;
5863 : EndOfWalRecoveryInfo *endOfRecoveryInfo;
5864 : XLogRecPtr abortedRecPtr;
5865 : XLogRecPtr missingContrecPtr;
5866 : TransactionId oldestActiveXID;
5867 1079 : bool promoted = false;
5868 : char timebuf[128];
5869 :
5870 : /*
5871 : * We should have an aux process resource owner to use, and we should not
5872 : * be in a transaction that's installed some other resowner.
5873 : */
5874 : Assert(AuxProcessResourceOwner != NULL);
5875 : Assert(CurrentResourceOwner == NULL ||
5876 : CurrentResourceOwner == AuxProcessResourceOwner);
5877 1079 : CurrentResourceOwner = AuxProcessResourceOwner;
5878 :
5879 : /*
5880 : * Check that contents look valid.
5881 : */
5882 1079 : if (!XRecOffIsValid(ControlFile->checkPoint))
5883 0 : ereport(FATAL,
5884 : (errcode(ERRCODE_DATA_CORRUPTED),
5885 : errmsg("control file contains invalid checkpoint location")));
5886 :
5887 1079 : switch (ControlFile->state)
5888 : {
5889 858 : case DB_SHUTDOWNED:
5890 :
5891 : /*
5892 : * This is the expected case, so don't be chatty in standalone
5893 : * mode
5894 : */
5895 858 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
5896 : (errmsg("database system was shut down at %s",
5897 : str_time(ControlFile->time,
5898 : timebuf, sizeof(timebuf)))));
5899 858 : break;
5900 :
5901 33 : case DB_SHUTDOWNED_IN_RECOVERY:
5902 33 : ereport(LOG,
5903 : (errmsg("database system was shut down in recovery at %s",
5904 : str_time(ControlFile->time,
5905 : timebuf, sizeof(timebuf)))));
5906 33 : break;
5907 :
5908 0 : case DB_SHUTDOWNING:
5909 0 : ereport(LOG,
5910 : (errmsg("database system shutdown was interrupted; last known up at %s",
5911 : str_time(ControlFile->time,
5912 : timebuf, sizeof(timebuf)))));
5913 0 : break;
5914 :
5915 0 : case DB_IN_CRASH_RECOVERY:
5916 0 : ereport(LOG,
5917 : (errmsg("database system was interrupted while in recovery at %s",
5918 : str_time(ControlFile->time,
5919 : timebuf, sizeof(timebuf))),
5920 : errhint("This probably means that some data is corrupted and"
5921 : " you will have to use the last backup for recovery.")));
5922 0 : break;
5923 :
5924 7 : case DB_IN_ARCHIVE_RECOVERY:
5925 7 : ereport(LOG,
5926 : (errmsg("database system was interrupted while in recovery at log time %s",
5927 : str_time(ControlFile->checkPointCopy.time,
5928 : timebuf, sizeof(timebuf))),
5929 : errhint("If this has occurred more than once some data might be corrupted"
5930 : " and you might need to choose an earlier recovery target.")));
5931 7 : break;
5932 :
5933 181 : case DB_IN_PRODUCTION:
5934 181 : ereport(LOG,
5935 : (errmsg("database system was interrupted; last known up at %s",
5936 : str_time(ControlFile->time,
5937 : timebuf, sizeof(timebuf)))));
5938 181 : break;
5939 :
5940 0 : default:
5941 0 : ereport(FATAL,
5942 : (errcode(ERRCODE_DATA_CORRUPTED),
5943 : errmsg("control file contains invalid database cluster state")));
5944 : }
5945 :
5946 : /* This is just to allow attaching to startup process with a debugger */
5947 : #ifdef XLOG_REPLAY_DELAY
5948 : if (ControlFile->state != DB_SHUTDOWNED)
5949 : pg_usleep(60000000L);
5950 : #endif
5951 :
5952 : /*
5953 : * Verify that pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
5954 : * In cases where someone has performed a copy for PITR, these directories
5955 : * may have been excluded and need to be re-created.
5956 : */
5957 1079 : ValidateXLOGDirectoryStructure();
5958 :
5959 : /* Set up timeout handler needed to report startup progress. */
5960 1079 : if (!IsBootstrapProcessingMode())
5961 1022 : RegisterTimeout(STARTUP_PROGRESS_TIMEOUT,
5962 : startup_progress_timeout_handler);
5963 :
5964 : /*----------
5965 : * If we previously crashed, perform a couple of actions:
5966 : *
5967 : * - The pg_wal directory may still include some temporary WAL segments
5968 : * used when creating a new segment, so perform some clean up to not
5969 : * bloat this path. This is done first as there is no point to sync
5970 : * this temporary data.
5971 : *
5972 : * - There might be data which we had written, intending to fsync it, but
5973 : * which we had not actually fsync'd yet. Therefore, a power failure in
5974 : * the near future might cause earlier unflushed writes to be lost, even
5975 : * though more recent data written to disk from here on would be
5976 : * persisted. To avoid that, fsync the entire data directory.
5977 : */
5978 1079 : if (ControlFile->state != DB_SHUTDOWNED &&
5979 221 : ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
5980 : {
5981 188 : RemoveTempXlogFiles();
5982 188 : SyncDataDirectory();
5983 188 : didCrash = true;
5984 : }
5985 : else
5986 891 : didCrash = false;
5987 :
5988 : /*
5989 : * Prepare for WAL recovery if needed.
5990 : *
5991 : * InitWalRecovery analyzes the control file and the backup label file, if
5992 : * any. It updates the in-memory ControlFile buffer according to the
5993 : * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5994 : * It also applies the tablespace map file, if any.
5995 : */
5996 1079 : InitWalRecovery(ControlFile, &wasShutdown,
5997 : &haveBackupLabel, &haveTblspcMap);
5998 1077 : checkPoint = ControlFile->checkPointCopy;
5999 :
6000 : /* initialize shared memory variables from the checkpoint record */
6001 1077 : TransamVariables->nextXid = checkPoint.nextXid;
6002 1077 : TransamVariables->nextOid = checkPoint.nextOid;
6003 1077 : TransamVariables->oidCount = 0;
6004 1077 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
6005 1077 : AdvanceOldestClogXid(checkPoint.oldestXid);
6006 1077 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
6007 1077 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
6008 1077 : SetCommitTsLimit(checkPoint.oldestCommitTsXid,
6009 : checkPoint.newestCommitTsXid);
6010 :
6011 : /*
6012 : * Clear out any old relcache cache files. This is *necessary* if we do
6013 : * any WAL replay, since that would probably result in the cache files
6014 : * being out of sync with database reality. In theory we could leave them
6015 : * in place if the database had been cleanly shut down, but it seems
6016 : * safest to just remove them always and let them be rebuilt during the
6017 : * first backend startup. These files needs to be removed from all
6018 : * directories including pg_tblspc, however the symlinks are created only
6019 : * after reading tablespace_map file in case of archive recovery from
6020 : * backup, so needs to clear old relcache files here after creating
6021 : * symlinks.
6022 : */
6023 1077 : RelationCacheInitFileRemove();
6024 :
6025 : /*
6026 : * Initialize replication slots, before there's a chance to remove
6027 : * required resources.
6028 : */
6029 1077 : StartupReplicationSlots();
6030 :
6031 : /*
6032 : * Startup the logical decoding status with the last status stored in the
6033 : * checkpoint record.
6034 : */
6035 1075 : StartupLogicalDecodingStatus(checkPoint.logicalDecodingEnabled);
6036 :
6037 : /*
6038 : * Startup logical state, needs to be setup now so we have proper data
6039 : * during crash recovery.
6040 : */
6041 1075 : StartupReorderBuffer();
6042 :
6043 : /*
6044 : * Startup CLOG. This must be done after TransamVariables->nextXid has
6045 : * been initialized and before we accept connections or begin WAL replay.
6046 : */
6047 1075 : StartupCLOG();
6048 :
6049 : /*
6050 : * Startup MultiXact. We need to do this early to be able to replay
6051 : * truncations.
6052 : */
6053 1075 : StartupMultiXact();
6054 :
6055 : /*
6056 : * Ditto for commit timestamps. Activate the facility if the setting is
6057 : * enabled in the control file, as there should be no tracking of commit
6058 : * timestamps done when the setting was disabled. This facility can be
6059 : * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
6060 : */
6061 1075 : if (ControlFile->track_commit_timestamp)
6062 14 : StartupCommitTs();
6063 :
6064 : /*
6065 : * Recover knowledge about replay progress of known replication partners.
6066 : */
6067 1075 : StartupReplicationOrigin();
6068 :
6069 : /*
6070 : * Initialize unlogged LSN. On a clean shutdown, it's restored from the
6071 : * control file. On recovery, all unlogged relations are blown away, so
6072 : * the unlogged LSN counter can be reset too.
6073 : */
6074 1075 : if (ControlFile->state == DB_SHUTDOWNED)
6075 849 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
6076 849 : ControlFile->unloggedLSN);
6077 : else
6078 226 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
6079 : FirstNormalUnloggedLSN);
6080 :
6081 : /*
6082 : * Copy any missing timeline history files between 'now' and the recovery
6083 : * target timeline from archive to pg_wal. While we don't need those files
6084 : * ourselves - the history file of the recovery target timeline covers all
6085 : * the previous timelines in the history too - a cascading standby server
6086 : * might be interested in them. Or, if you archive the WAL from this
6087 : * server to a different archive than the primary, it'd be good for all
6088 : * the history files to get archived there after failover, so that you can
6089 : * use one of the old timelines as a PITR target. Timeline history files
6090 : * are small, so it's better to copy them unnecessarily than not copy them
6091 : * and regret later.
6092 : */
6093 1075 : restoreTimeLineHistoryFiles(checkPoint.ThisTimeLineID, recoveryTargetTLI);
6094 :
6095 : /*
6096 : * Before running in recovery, scan pg_twophase and fill in its status to
6097 : * be able to work on entries generated by redo. Doing a scan before
6098 : * taking any recovery action has the merit to discard any 2PC files that
6099 : * are newer than the first record to replay, saving from any conflicts at
6100 : * replay. This avoids as well any subsequent scans when doing recovery
6101 : * of the on-disk two-phase data.
6102 : */
6103 1075 : restoreTwoPhaseData();
6104 :
6105 : /*
6106 : * When starting with crash recovery, reset pgstat data - it might not be
6107 : * valid. Otherwise restore pgstat data. It's safe to do this here,
6108 : * because postmaster will not yet have started any other processes.
6109 : *
6110 : * NB: Restoring replication slot stats relies on slot state to have
6111 : * already been restored from disk.
6112 : *
6113 : * TODO: With a bit of extra work we could just start with a pgstat file
6114 : * associated with the checkpoint redo location we're starting from.
6115 : */
6116 1075 : if (didCrash)
6117 186 : pgstat_discard_stats();
6118 : else
6119 889 : pgstat_restore_stats();
6120 :
6121 1075 : lastFullPageWrites = checkPoint.fullPageWrites;
6122 :
6123 1075 : RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
6124 1075 : doPageWrites = lastFullPageWrites;
6125 :
6126 : /* REDO */
6127 1075 : if (InRecovery)
6128 : {
6129 : /* Initialize state for RecoveryInProgress() */
6130 226 : SpinLockAcquire(&XLogCtl->info_lck);
6131 226 : if (InArchiveRecovery)
6132 124 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6133 : else
6134 102 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
6135 226 : SpinLockRelease(&XLogCtl->info_lck);
6136 :
6137 : /*
6138 : * Update pg_control to show that we are recovering and to show the
6139 : * selected checkpoint as the place we are starting from. We also mark
6140 : * pg_control with any minimum recovery stop point obtained from a
6141 : * backup history file.
6142 : *
6143 : * No need to hold ControlFileLock yet, we aren't up far enough.
6144 : */
6145 226 : UpdateControlFile();
6146 :
6147 : /*
6148 : * If there was a backup label file, it's done its job and the info
6149 : * has now been propagated into pg_control. We must get rid of the
6150 : * label file so that if we crash during recovery, we'll pick up at
6151 : * the latest recovery restartpoint instead of going all the way back
6152 : * to the backup start point. It seems prudent though to just rename
6153 : * the file out of the way rather than delete it completely.
6154 : */
6155 226 : if (haveBackupLabel)
6156 : {
6157 82 : unlink(BACKUP_LABEL_OLD);
6158 82 : durable_rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD, FATAL);
6159 : }
6160 :
6161 : /*
6162 : * If there was a tablespace_map file, it's done its job and the
6163 : * symlinks have been created. We must get rid of the map file so
6164 : * that if we crash during recovery, we don't create symlinks again.
6165 : * It seems prudent though to just rename the file out of the way
6166 : * rather than delete it completely.
6167 : */
6168 226 : if (haveTblspcMap)
6169 : {
6170 2 : unlink(TABLESPACE_MAP_OLD);
6171 2 : durable_rename(TABLESPACE_MAP, TABLESPACE_MAP_OLD, FATAL);
6172 : }
6173 :
6174 : /*
6175 : * Initialize our local copy of minRecoveryPoint. When doing crash
6176 : * recovery we want to replay up to the end of WAL. Particularly, in
6177 : * the case of a promoted standby minRecoveryPoint value in the
6178 : * control file is only updated after the first checkpoint. However,
6179 : * if the instance crashes before the first post-recovery checkpoint
6180 : * is completed then recovery will use a stale location causing the
6181 : * startup process to think that there are still invalid page
6182 : * references when checking for data consistency.
6183 : */
6184 226 : if (InArchiveRecovery)
6185 : {
6186 124 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6187 124 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6188 : }
6189 : else
6190 : {
6191 102 : LocalMinRecoveryPoint = InvalidXLogRecPtr;
6192 102 : LocalMinRecoveryPointTLI = 0;
6193 : }
6194 :
6195 : /* Check that the GUCs used to generate the WAL allow recovery */
6196 226 : CheckRequiredParameterValues();
6197 :
6198 : /*
6199 : * We're in recovery, so unlogged relations may be trashed and must be
6200 : * reset. This should be done BEFORE allowing Hot Standby
6201 : * connections, so that read-only backends don't try to read whatever
6202 : * garbage is left over from before.
6203 : */
6204 226 : ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
6205 :
6206 : /*
6207 : * Likewise, delete any saved transaction snapshot files that got left
6208 : * behind by crashed backends.
6209 : */
6210 226 : DeleteAllExportedSnapshotFiles();
6211 :
6212 : /*
6213 : * Initialize for Hot Standby, if enabled. We won't let backends in
6214 : * yet, not until we've reached the min recovery point specified in
6215 : * control file and we've established a recovery snapshot from a
6216 : * running-xacts WAL record.
6217 : */
6218 226 : if (ArchiveRecoveryRequested && EnableHotStandby)
6219 : {
6220 : TransactionId *xids;
6221 : int nxids;
6222 :
6223 116 : ereport(DEBUG1,
6224 : (errmsg_internal("initializing for hot standby")));
6225 :
6226 116 : InitRecoveryTransactionEnvironment();
6227 :
6228 116 : if (wasShutdown)
6229 26 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
6230 : else
6231 90 : oldestActiveXID = checkPoint.oldestActiveXid;
6232 : Assert(TransactionIdIsValid(oldestActiveXID));
6233 :
6234 : /* Tell procarray about the range of xids it has to deal with */
6235 116 : ProcArrayInitRecovery(XidFromFullTransactionId(TransamVariables->nextXid));
6236 :
6237 : /*
6238 : * Startup subtrans only. CLOG, MultiXact and commit timestamp
6239 : * have already been started up and other SLRUs are not maintained
6240 : * during recovery and need not be started yet.
6241 : */
6242 116 : StartupSUBTRANS(oldestActiveXID);
6243 :
6244 : /*
6245 : * If we're beginning at a shutdown checkpoint, we know that
6246 : * nothing was running on the primary at this point. So fake-up an
6247 : * empty running-xacts record and use that here and now. Recover
6248 : * additional standby state for prepared transactions.
6249 : */
6250 116 : if (wasShutdown)
6251 : {
6252 : RunningTransactionsData running;
6253 : TransactionId latestCompletedXid;
6254 :
6255 : /* Update pg_subtrans entries for any prepared transactions */
6256 26 : StandbyRecoverPreparedTransactions();
6257 :
6258 : /*
6259 : * Construct a RunningTransactions snapshot representing a
6260 : * shut down server, with only prepared transactions still
6261 : * alive. We're never overflowed at this point because all
6262 : * subxids are listed with their parent prepared transactions.
6263 : */
6264 26 : running.xcnt = nxids;
6265 26 : running.subxcnt = 0;
6266 26 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
6267 26 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
6268 26 : running.oldestRunningXid = oldestActiveXID;
6269 26 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
6270 26 : TransactionIdRetreat(latestCompletedXid);
6271 : Assert(TransactionIdIsNormal(latestCompletedXid));
6272 26 : running.latestCompletedXid = latestCompletedXid;
6273 26 : running.xids = xids;
6274 :
6275 26 : ProcArrayApplyRecoveryInfo(&running);
6276 : }
6277 : }
6278 :
6279 : /*
6280 : * We're all set for replaying the WAL now. Do it.
6281 : */
6282 226 : PerformWalRecovery();
6283 163 : performedWalRecovery = true;
6284 : }
6285 : else
6286 849 : performedWalRecovery = false;
6287 :
6288 : /*
6289 : * Finish WAL recovery.
6290 : */
6291 1012 : endOfRecoveryInfo = FinishWalRecovery();
6292 1012 : EndOfLog = endOfRecoveryInfo->endOfLog;
6293 1012 : EndOfLogTLI = endOfRecoveryInfo->endOfLogTLI;
6294 1012 : abortedRecPtr = endOfRecoveryInfo->abortedRecPtr;
6295 1012 : missingContrecPtr = endOfRecoveryInfo->missingContrecPtr;
6296 :
6297 : /*
6298 : * Reset ps status display, so as no information related to recovery shows
6299 : * up.
6300 : */
6301 1012 : set_ps_display("");
6302 :
6303 : /*
6304 : * When recovering from a backup (we are in recovery, and archive recovery
6305 : * was requested), complain if we did not roll forward far enough to reach
6306 : * the point where the database is consistent. For regular online
6307 : * backup-from-primary, that means reaching the end-of-backup WAL record
6308 : * (at which point we reset backupStartPoint to be Invalid), for
6309 : * backup-from-replica (which can't inject records into the WAL stream),
6310 : * that point is when we reach the minRecoveryPoint in pg_control (which
6311 : * we purposefully copy last when backing up from a replica). For
6312 : * pg_rewind (which creates a backup_label with a method of "pg_rewind")
6313 : * or snapshot-style backups (which don't), backupEndRequired will be set
6314 : * to false.
6315 : *
6316 : * Note: it is indeed okay to look at the local variable
6317 : * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
6318 : * might be further ahead --- ControlFile->minRecoveryPoint cannot have
6319 : * been advanced beyond the WAL we processed.
6320 : */
6321 1012 : if (InRecovery &&
6322 163 : (EndOfLog < LocalMinRecoveryPoint ||
6323 163 : XLogRecPtrIsValid(ControlFile->backupStartPoint)))
6324 : {
6325 : /*
6326 : * Ran off end of WAL before reaching end-of-backup WAL record, or
6327 : * minRecoveryPoint. That's a bad sign, indicating that you tried to
6328 : * recover from an online backup but never called pg_backup_stop(), or
6329 : * you didn't archive all the WAL needed.
6330 : */
6331 0 : if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
6332 : {
6333 0 : if (XLogRecPtrIsValid(ControlFile->backupStartPoint) || ControlFile->backupEndRequired)
6334 0 : ereport(FATAL,
6335 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6336 : errmsg("WAL ends before end of online backup"),
6337 : errhint("All WAL generated while online backup was taken must be available at recovery.")));
6338 : else
6339 0 : ereport(FATAL,
6340 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6341 : errmsg("WAL ends before consistent recovery point")));
6342 : }
6343 : }
6344 :
6345 : /*
6346 : * Reset unlogged relations to the contents of their INIT fork. This is
6347 : * done AFTER recovery is complete so as to include any unlogged relations
6348 : * created during recovery, but BEFORE recovery is marked as having
6349 : * completed successfully. Otherwise we'd not retry if any of the post
6350 : * end-of-recovery steps fail.
6351 : */
6352 1012 : if (InRecovery)
6353 163 : ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
6354 :
6355 : /*
6356 : * Pre-scan prepared transactions to find out the range of XIDs present.
6357 : * This information is not quite needed yet, but it is positioned here so
6358 : * as potential problems are detected before any on-disk change is done.
6359 : */
6360 1012 : oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
6361 :
6362 : /*
6363 : * Allow ordinary WAL segment creation before possibly switching to a new
6364 : * timeline, which creates a new segment, and after the last ReadRecord().
6365 : */
6366 1012 : SetInstallXLogFileSegmentActive();
6367 :
6368 : /*
6369 : * Consider whether we need to assign a new timeline ID.
6370 : *
6371 : * If we did archive recovery, we always assign a new ID. This handles a
6372 : * couple of issues. If we stopped short of the end of WAL during
6373 : * recovery, then we are clearly generating a new timeline and must assign
6374 : * it a unique new ID. Even if we ran to the end, modifying the current
6375 : * last segment is problematic because it may result in trying to
6376 : * overwrite an already-archived copy of that segment, and we encourage
6377 : * DBAs to make their archive_commands reject that. We can dodge the
6378 : * problem by making the new active segment have a new timeline ID.
6379 : *
6380 : * In a normal crash recovery, we can just extend the timeline we were in.
6381 : */
6382 1012 : newTLI = endOfRecoveryInfo->lastRecTLI;
6383 1012 : if (ArchiveRecoveryRequested)
6384 : {
6385 55 : newTLI = findNewestTimeLine(recoveryTargetTLI) + 1;
6386 55 : ereport(LOG,
6387 : (errmsg("selected new timeline ID: %u", newTLI)));
6388 :
6389 : /*
6390 : * Make a writable copy of the last WAL segment. (Note that we also
6391 : * have a copy of the last block of the old WAL in
6392 : * endOfRecovery->lastPage; we will use that below.)
6393 : */
6394 55 : XLogInitNewTimeline(EndOfLogTLI, EndOfLog, newTLI);
6395 :
6396 : /*
6397 : * Remove the signal files out of the way, so that we don't
6398 : * accidentally re-enter archive recovery mode in a subsequent crash.
6399 : */
6400 55 : if (endOfRecoveryInfo->standby_signal_file_found)
6401 52 : durable_unlink(STANDBY_SIGNAL_FILE, FATAL);
6402 :
6403 55 : if (endOfRecoveryInfo->recovery_signal_file_found)
6404 4 : durable_unlink(RECOVERY_SIGNAL_FILE, FATAL);
6405 :
6406 : /*
6407 : * Write the timeline history file, and have it archived. After this
6408 : * point (or rather, as soon as the file is archived), the timeline
6409 : * will appear as "taken" in the WAL archive and to any standby
6410 : * servers. If we crash before actually switching to the new
6411 : * timeline, standby servers will nevertheless think that we switched
6412 : * to the new timeline, and will try to connect to the new timeline.
6413 : * To minimize the window for that, try to do as little as possible
6414 : * between here and writing the end-of-recovery record.
6415 : */
6416 55 : writeTimeLineHistory(newTLI, recoveryTargetTLI,
6417 : EndOfLog, endOfRecoveryInfo->recoveryStopReason);
6418 :
6419 55 : ereport(LOG,
6420 : (errmsg("archive recovery complete")));
6421 : }
6422 :
6423 : /* Save the selected TimeLineID in shared memory, too */
6424 1012 : SpinLockAcquire(&XLogCtl->info_lck);
6425 1012 : XLogCtl->InsertTimeLineID = newTLI;
6426 1012 : XLogCtl->PrevTimeLineID = endOfRecoveryInfo->lastRecTLI;
6427 1012 : SpinLockRelease(&XLogCtl->info_lck);
6428 :
6429 : /*
6430 : * Actually, if WAL ended in an incomplete record, skip the parts that
6431 : * made it through and start writing after the portion that persisted.
6432 : * (It's critical to first write an OVERWRITE_CONTRECORD message, which
6433 : * we'll do as soon as we're open for writing new WAL.)
6434 : */
6435 1012 : if (XLogRecPtrIsValid(missingContrecPtr))
6436 : {
6437 : /*
6438 : * We should only have a missingContrecPtr if we're not switching to a
6439 : * new timeline. When a timeline switch occurs, WAL is copied from the
6440 : * old timeline to the new only up to the end of the last complete
6441 : * record, so there can't be an incomplete WAL record that we need to
6442 : * disregard.
6443 : */
6444 : Assert(newTLI == endOfRecoveryInfo->lastRecTLI);
6445 : Assert(XLogRecPtrIsValid(abortedRecPtr));
6446 11 : EndOfLog = missingContrecPtr;
6447 : }
6448 :
6449 : /*
6450 : * Prepare to write WAL starting at EndOfLog location, and init xlog
6451 : * buffer cache using the block containing the last record from the
6452 : * previous incarnation.
6453 : */
6454 1012 : Insert = &XLogCtl->Insert;
6455 1012 : Insert->PrevBytePos = XLogRecPtrToBytePos(endOfRecoveryInfo->lastRec);
6456 1012 : Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
6457 :
6458 : /*
6459 : * Tricky point here: lastPage contains the *last* block that the LastRec
6460 : * record spans, not the one it starts in. The last block is indeed the
6461 : * one we want to use.
6462 : */
6463 1012 : if (EndOfLog % XLOG_BLCKSZ != 0)
6464 : {
6465 : char *page;
6466 : int len;
6467 : int firstIdx;
6468 :
6469 979 : firstIdx = XLogRecPtrToBufIdx(EndOfLog);
6470 979 : len = EndOfLog - endOfRecoveryInfo->lastPageBeginPtr;
6471 : Assert(len < XLOG_BLCKSZ);
6472 :
6473 : /* Copy the valid part of the last block, and zero the rest */
6474 979 : page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
6475 979 : memcpy(page, endOfRecoveryInfo->lastPage, len);
6476 979 : memset(page + len, 0, XLOG_BLCKSZ - len);
6477 :
6478 979 : pg_atomic_write_u64(&XLogCtl->xlblocks[firstIdx], endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6479 979 : XLogCtl->InitializedUpTo = endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ;
6480 : }
6481 : else
6482 : {
6483 : /*
6484 : * There is no partial block to copy. Just set InitializedUpTo, and
6485 : * let the first attempt to insert a log record to initialize the next
6486 : * buffer.
6487 : */
6488 33 : XLogCtl->InitializedUpTo = EndOfLog;
6489 : }
6490 :
6491 : /*
6492 : * Update local and shared status. This is OK to do without any locks
6493 : * because no other process can be reading or writing WAL yet.
6494 : */
6495 1012 : LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
6496 1012 : pg_atomic_write_u64(&XLogCtl->logInsertResult, EndOfLog);
6497 1012 : pg_atomic_write_u64(&XLogCtl->logWriteResult, EndOfLog);
6498 1012 : pg_atomic_write_u64(&XLogCtl->logFlushResult, EndOfLog);
6499 1012 : XLogCtl->LogwrtRqst.Write = EndOfLog;
6500 1012 : XLogCtl->LogwrtRqst.Flush = EndOfLog;
6501 :
6502 : /*
6503 : * Preallocate additional log files, if wanted.
6504 : */
6505 1012 : PreallocXlogFiles(EndOfLog, newTLI);
6506 :
6507 : /*
6508 : * Okay, we're officially UP.
6509 : */
6510 1012 : InRecovery = false;
6511 :
6512 : /* start the archive_timeout timer and LSN running */
6513 1012 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
6514 1012 : XLogCtl->lastSegSwitchLSN = EndOfLog;
6515 :
6516 : /* also initialize latestCompletedXid, to nextXid - 1 */
6517 1012 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
6518 1012 : TransamVariables->latestCompletedXid = TransamVariables->nextXid;
6519 1012 : FullTransactionIdRetreat(&TransamVariables->latestCompletedXid);
6520 1012 : LWLockRelease(ProcArrayLock);
6521 :
6522 : /*
6523 : * Start up subtrans, if not already done for hot standby. (commit
6524 : * timestamps are started below, if necessary.)
6525 : */
6526 1012 : if (standbyState == STANDBY_DISABLED)
6527 957 : StartupSUBTRANS(oldestActiveXID);
6528 :
6529 : /*
6530 : * Perform end of recovery actions for any SLRUs that need it.
6531 : */
6532 1012 : TrimCLOG();
6533 1012 : TrimMultiXact();
6534 :
6535 : /*
6536 : * Reload shared-memory state for prepared transactions. This needs to
6537 : * happen before renaming the last partial segment of the old timeline as
6538 : * it may be possible that we have to recover some transactions from it.
6539 : */
6540 1012 : RecoverPreparedTransactions();
6541 :
6542 : /* Shut down xlogreader */
6543 1012 : ShutdownWalRecovery();
6544 :
6545 : /* Enable WAL writes for this backend only. */
6546 1012 : LocalSetXLogInsertAllowed();
6547 :
6548 : /* If necessary, write overwrite-contrecord before doing anything else */
6549 1012 : if (XLogRecPtrIsValid(abortedRecPtr))
6550 : {
6551 : Assert(XLogRecPtrIsValid(missingContrecPtr));
6552 11 : CreateOverwriteContrecordRecord(abortedRecPtr, missingContrecPtr, newTLI);
6553 : }
6554 :
6555 : /*
6556 : * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
6557 : * record before resource manager writes cleanup WAL records or checkpoint
6558 : * record is written.
6559 : */
6560 1012 : Insert->fullPageWrites = lastFullPageWrites;
6561 1012 : UpdateFullPageWrites();
6562 :
6563 : /*
6564 : * Emit checkpoint or end-of-recovery record in XLOG, if required.
6565 : */
6566 1012 : if (performedWalRecovery)
6567 163 : promoted = PerformRecoveryXLogAction();
6568 :
6569 : /*
6570 : * If any of the critical GUCs have changed, log them before we allow
6571 : * backends to write WAL.
6572 : */
6573 1012 : XLogReportParameters();
6574 :
6575 : /* If this is archive recovery, perform post-recovery cleanup actions. */
6576 1012 : if (ArchiveRecoveryRequested)
6577 55 : CleanupAfterArchiveRecovery(EndOfLogTLI, EndOfLog, newTLI);
6578 :
6579 : /*
6580 : * Local WAL inserts enabled, so it's time to finish initialization of
6581 : * commit timestamp.
6582 : */
6583 1012 : CompleteCommitTsInitialization();
6584 :
6585 : /*
6586 : * Update logical decoding status in shared memory and write an
6587 : * XLOG_LOGICAL_DECODING_STATUS_CHANGE, if necessary.
6588 : */
6589 1012 : UpdateLogicalDecodingStatusEndOfRecovery();
6590 :
6591 : /* Clean up EndOfWalRecoveryInfo data to appease Valgrind leak checking */
6592 1012 : if (endOfRecoveryInfo->lastPage)
6593 990 : pfree(endOfRecoveryInfo->lastPage);
6594 1012 : pfree(endOfRecoveryInfo->recoveryStopReason);
6595 1012 : pfree(endOfRecoveryInfo);
6596 :
6597 : /*
6598 : * If we reach this point with checksums in the state inprogress-on, it
6599 : * means that data checksums were in the process of being enabled when the
6600 : * cluster shut down. Since processing didn't finish, the operation will
6601 : * have to be restarted from scratch since there is no capability to
6602 : * continue where it was when the cluster shut down. Thus, revert the
6603 : * state back to off, and inform the user with a warning message. Being
6604 : * able to restart processing is a TODO, but it wouldn't be possible to
6605 : * restart here since we cannot launch a dynamic background worker
6606 : * directly from here (it has to be from a regular backend).
6607 : */
6608 1012 : if (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_INPROGRESS_ON)
6609 : {
6610 0 : XLogChecksums(PG_DATA_CHECKSUM_OFF);
6611 :
6612 0 : SpinLockAcquire(&XLogCtl->info_lck);
6613 0 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_OFF;
6614 0 : SetLocalDataChecksumState(XLogCtl->data_checksum_version);
6615 0 : SpinLockRelease(&XLogCtl->info_lck);
6616 :
6617 0 : ereport(WARNING,
6618 : errmsg("enabling data checksums was interrupted"),
6619 : errhint("Data checksum processing must be manually restarted for checksums to be enabled"));
6620 : }
6621 :
6622 : /*
6623 : * If data checksums were being disabled when the cluster was shut down,
6624 : * we know that we have a state where all backends have stopped validating
6625 : * checksums and we can move to off instead of prompting the user to
6626 : * perform any action.
6627 : */
6628 1012 : if (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_INPROGRESS_OFF)
6629 : {
6630 0 : XLogChecksums(PG_DATA_CHECKSUM_OFF);
6631 :
6632 0 : SpinLockAcquire(&XLogCtl->info_lck);
6633 0 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_OFF;
6634 0 : SetLocalDataChecksumState(XLogCtl->data_checksum_version);
6635 0 : SpinLockRelease(&XLogCtl->info_lck);
6636 : }
6637 :
6638 : /*
6639 : * All done with end-of-recovery actions.
6640 : *
6641 : * Now allow backends to write WAL and update the control file status in
6642 : * consequence. SharedRecoveryState, that controls if backends can write
6643 : * WAL, is updated while holding ControlFileLock to prevent other backends
6644 : * to look at an inconsistent state of the control file in shared memory.
6645 : * There is still a small window during which backends can write WAL and
6646 : * the control file is still referring to a system not in DB_IN_PRODUCTION
6647 : * state while looking at the on-disk control file.
6648 : *
6649 : * Also, we use info_lck to update SharedRecoveryState to ensure that
6650 : * there are no race conditions concerning visibility of other recent
6651 : * updates to shared memory.
6652 : */
6653 1012 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6654 1012 : ControlFile->state = DB_IN_PRODUCTION;
6655 :
6656 1012 : SpinLockAcquire(&XLogCtl->info_lck);
6657 1012 : ControlFile->data_checksum_version = XLogCtl->data_checksum_version;
6658 1012 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_DONE;
6659 1012 : SpinLockRelease(&XLogCtl->info_lck);
6660 :
6661 1012 : UpdateControlFile();
6662 1012 : LWLockRelease(ControlFileLock);
6663 :
6664 : /*
6665 : * Wake up the checkpointer process as there might be a request to disable
6666 : * logical decoding by concurrent slot drop.
6667 : */
6668 1012 : WakeupCheckpointer();
6669 :
6670 : /*
6671 : * Wake up all waiters. They need to report an error that recovery was
6672 : * ended before reaching the target LSN.
6673 : */
6674 1012 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_REPLAY, InvalidXLogRecPtr);
6675 1012 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_WRITE, InvalidXLogRecPtr);
6676 1012 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_FLUSH, InvalidXLogRecPtr);
6677 :
6678 : /*
6679 : * Shutdown the recovery environment. This must occur after
6680 : * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
6681 : * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
6682 : * any session building a snapshot will not rely on KnownAssignedXids as
6683 : * RecoveryInProgress() would return false at this stage. This is
6684 : * particularly critical for prepared 2PC transactions, that would still
6685 : * need to be included in snapshots once recovery has ended.
6686 : */
6687 1012 : if (standbyState != STANDBY_DISABLED)
6688 55 : ShutdownRecoveryTransactionEnvironment();
6689 :
6690 : /*
6691 : * If there were cascading standby servers connected to us, nudge any wal
6692 : * sender processes to notice that we've been promoted.
6693 : */
6694 1012 : WalSndWakeup(true, true);
6695 :
6696 : /*
6697 : * If this was a promotion, request an (online) checkpoint now. This isn't
6698 : * required for consistency, but the last restartpoint might be far back,
6699 : * and in case of a crash, recovering from it might take a longer than is
6700 : * appropriate now that we're not in standby mode anymore.
6701 : */
6702 1012 : if (promoted)
6703 48 : RequestCheckpoint(CHECKPOINT_FORCE);
6704 1012 : }
6705 :
6706 : /*
6707 : * Callback from PerformWalRecovery(), called when we switch from crash
6708 : * recovery to archive recovery mode. Updates the control file accordingly.
6709 : */
6710 : void
6711 1 : SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr, TimeLineID replayTLI)
6712 : {
6713 : /* initialize minRecoveryPoint to this record */
6714 1 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6715 1 : ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6716 1 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6717 : {
6718 1 : ControlFile->minRecoveryPoint = EndRecPtr;
6719 1 : ControlFile->minRecoveryPointTLI = replayTLI;
6720 : }
6721 : /* update local copy */
6722 1 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6723 1 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6724 :
6725 : /*
6726 : * The startup process can update its local copy of minRecoveryPoint from
6727 : * this point.
6728 : */
6729 1 : updateMinRecoveryPoint = true;
6730 :
6731 1 : UpdateControlFile();
6732 :
6733 : /*
6734 : * We update SharedRecoveryState while holding the lock on ControlFileLock
6735 : * so both states are consistent in shared memory.
6736 : */
6737 1 : SpinLockAcquire(&XLogCtl->info_lck);
6738 1 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6739 1 : SpinLockRelease(&XLogCtl->info_lck);
6740 :
6741 1 : LWLockRelease(ControlFileLock);
6742 1 : }
6743 :
6744 : /*
6745 : * Callback from PerformWalRecovery(), called when we reach the end of backup.
6746 : * Updates the control file accordingly.
6747 : */
6748 : void
6749 82 : ReachedEndOfBackup(XLogRecPtr EndRecPtr, TimeLineID tli)
6750 : {
6751 : /*
6752 : * We have reached the end of base backup, as indicated by pg_control. The
6753 : * data on disk is now consistent (unless minRecoveryPoint is further
6754 : * ahead, which can happen if we crashed during previous recovery). Reset
6755 : * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
6756 : * make sure we don't allow starting up at an earlier point even if
6757 : * recovery is stopped and restarted soon after this.
6758 : */
6759 82 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6760 :
6761 82 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6762 : {
6763 77 : ControlFile->minRecoveryPoint = EndRecPtr;
6764 77 : ControlFile->minRecoveryPointTLI = tli;
6765 : }
6766 :
6767 82 : ControlFile->backupStartPoint = InvalidXLogRecPtr;
6768 82 : ControlFile->backupEndPoint = InvalidXLogRecPtr;
6769 82 : ControlFile->backupEndRequired = false;
6770 82 : UpdateControlFile();
6771 :
6772 82 : LWLockRelease(ControlFileLock);
6773 82 : }
6774 :
6775 : /*
6776 : * Perform whatever XLOG actions are necessary at end of REDO.
6777 : *
6778 : * The goal here is to make sure that we'll be able to recover properly if
6779 : * we crash again. If we choose to write a checkpoint, we'll write a shutdown
6780 : * checkpoint rather than an on-line one. This is not particularly critical,
6781 : * but since we may be assigning a new TLI, using a shutdown checkpoint allows
6782 : * us to have the rule that TLI only changes in shutdown checkpoints, which
6783 : * allows some extra error checking in xlog_redo.
6784 : */
6785 : static bool
6786 163 : PerformRecoveryXLogAction(void)
6787 : {
6788 163 : bool promoted = false;
6789 :
6790 : /*
6791 : * Perform a checkpoint to update all our recovery activity to disk.
6792 : *
6793 : * Note that we write a shutdown checkpoint rather than an on-line one.
6794 : * This is not particularly critical, but since we may be assigning a new
6795 : * TLI, using a shutdown checkpoint allows us to have the rule that TLI
6796 : * only changes in shutdown checkpoints, which allows some extra error
6797 : * checking in xlog_redo.
6798 : *
6799 : * In promotion, only create a lightweight end-of-recovery record instead
6800 : * of a full checkpoint. A checkpoint is requested later, after we're
6801 : * fully out of recovery mode and already accepting queries.
6802 : */
6803 218 : if (ArchiveRecoveryRequested && IsUnderPostmaster &&
6804 55 : PromoteIsTriggered())
6805 : {
6806 48 : promoted = true;
6807 :
6808 : /*
6809 : * Insert a special WAL record to mark the end of recovery, since we
6810 : * aren't doing a checkpoint. That means that the checkpointer process
6811 : * may likely be in the middle of a time-smoothed restartpoint and
6812 : * could continue to be for minutes after this. That sounds strange,
6813 : * but the effect is roughly the same and it would be stranger to try
6814 : * to come out of the restartpoint and then checkpoint. We request a
6815 : * checkpoint later anyway, just for safety.
6816 : */
6817 48 : CreateEndOfRecoveryRecord();
6818 : }
6819 : else
6820 : {
6821 115 : RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
6822 : CHECKPOINT_FAST |
6823 : CHECKPOINT_WAIT);
6824 : }
6825 :
6826 163 : return promoted;
6827 : }
6828 :
6829 : /*
6830 : * Is the system still in recovery?
6831 : *
6832 : * Unlike testing InRecovery, this works in any process that's connected to
6833 : * shared memory.
6834 : */
6835 : bool
6836 97393455 : RecoveryInProgress(void)
6837 : {
6838 : /*
6839 : * We check shared state each time only until we leave recovery mode. We
6840 : * can't re-enter recovery, so there's no need to keep checking after the
6841 : * shared variable has once been seen false.
6842 : */
6843 97393455 : if (!LocalRecoveryInProgress)
6844 95331794 : return false;
6845 : else
6846 : {
6847 : /*
6848 : * use volatile pointer to make sure we make a fresh read of the
6849 : * shared variable.
6850 : */
6851 2061661 : volatile XLogCtlData *xlogctl = XLogCtl;
6852 :
6853 2061661 : LocalRecoveryInProgress = (xlogctl->SharedRecoveryState != RECOVERY_STATE_DONE);
6854 :
6855 : /*
6856 : * Note: We don't need a memory barrier when we're still in recovery.
6857 : * We might exit recovery immediately after return, so the caller
6858 : * can't rely on 'true' meaning that we're still in recovery anyway.
6859 : */
6860 :
6861 2061661 : return LocalRecoveryInProgress;
6862 : }
6863 : }
6864 :
6865 : /*
6866 : * Returns current recovery state from shared memory.
6867 : *
6868 : * This returned state is kept consistent with the contents of the control
6869 : * file. See details about the possible values of RecoveryState in xlog.h.
6870 : */
6871 : RecoveryState
6872 35607 : GetRecoveryState(void)
6873 : {
6874 : RecoveryState retval;
6875 :
6876 35607 : SpinLockAcquire(&XLogCtl->info_lck);
6877 35607 : retval = XLogCtl->SharedRecoveryState;
6878 35607 : SpinLockRelease(&XLogCtl->info_lck);
6879 :
6880 35607 : return retval;
6881 : }
6882 :
6883 : /*
6884 : * Is this process allowed to insert new WAL records?
6885 : *
6886 : * Ordinarily this is essentially equivalent to !RecoveryInProgress().
6887 : * But we also have provisions for forcing the result "true" or "false"
6888 : * within specific processes regardless of the global state.
6889 : */
6890 : bool
6891 67235751 : XLogInsertAllowed(void)
6892 : {
6893 : /*
6894 : * If value is "unconditionally true" or "unconditionally false", just
6895 : * return it. This provides the normal fast path once recovery is known
6896 : * done.
6897 : */
6898 67235751 : if (LocalXLogInsertAllowed >= 0)
6899 66568776 : return (bool) LocalXLogInsertAllowed;
6900 :
6901 : /*
6902 : * Else, must check to see if we're still in recovery.
6903 : */
6904 666975 : if (RecoveryInProgress())
6905 656385 : return false;
6906 :
6907 : /*
6908 : * On exit from recovery, reset to "unconditionally true", since there is
6909 : * no need to keep checking.
6910 : */
6911 10590 : LocalXLogInsertAllowed = 1;
6912 10590 : return true;
6913 : }
6914 :
6915 : /*
6916 : * Make XLogInsertAllowed() return true in the current process only.
6917 : *
6918 : * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
6919 : * and even call LocalSetXLogInsertAllowed() again after that.
6920 : *
6921 : * Returns the previous value of LocalXLogInsertAllowed.
6922 : */
6923 : static int
6924 1042 : LocalSetXLogInsertAllowed(void)
6925 : {
6926 1042 : int oldXLogAllowed = LocalXLogInsertAllowed;
6927 :
6928 1042 : LocalXLogInsertAllowed = 1;
6929 :
6930 1042 : return oldXLogAllowed;
6931 : }
6932 :
6933 : /*
6934 : * Return the current Redo pointer from shared memory.
6935 : *
6936 : * As a side-effect, the local RedoRecPtr copy is updated.
6937 : */
6938 : XLogRecPtr
6939 366279 : GetRedoRecPtr(void)
6940 : {
6941 : XLogRecPtr ptr;
6942 :
6943 : /*
6944 : * The possibly not up-to-date copy in XLogCtl is enough. Even if we
6945 : * grabbed a WAL insertion lock to read the authoritative value in
6946 : * Insert->RedoRecPtr, someone might update it just after we've released
6947 : * the lock.
6948 : */
6949 366279 : SpinLockAcquire(&XLogCtl->info_lck);
6950 366279 : ptr = XLogCtl->RedoRecPtr;
6951 366279 : SpinLockRelease(&XLogCtl->info_lck);
6952 :
6953 366279 : if (RedoRecPtr < ptr)
6954 1761 : RedoRecPtr = ptr;
6955 :
6956 366279 : return RedoRecPtr;
6957 : }
6958 :
6959 : /*
6960 : * Return information needed to decide whether a modified block needs a
6961 : * full-page image to be included in the WAL record.
6962 : *
6963 : * The returned values are cached copies from backend-private memory, and
6964 : * possibly out-of-date or, indeed, uninitialized, in which case they will
6965 : * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6966 : * re-check them against up-to-date values, while holding the WAL insert lock.
6967 : */
6968 : void
6969 24942665 : GetFullPageWriteInfo(XLogRecPtr *RedoRecPtr_p, bool *doPageWrites_p)
6970 : {
6971 24942665 : *RedoRecPtr_p = RedoRecPtr;
6972 24942665 : *doPageWrites_p = doPageWrites;
6973 24942665 : }
6974 :
6975 : /*
6976 : * GetInsertRecPtr -- Returns the current insert position.
6977 : *
6978 : * NOTE: The value *actually* returned is the position of the last full
6979 : * xlog page. It lags behind the real insert position by at most 1 page.
6980 : * For that, we don't need to scan through WAL insertion locks, and an
6981 : * approximation is enough for the current usage of this function.
6982 : */
6983 : XLogRecPtr
6984 7431 : GetInsertRecPtr(void)
6985 : {
6986 : XLogRecPtr recptr;
6987 :
6988 7431 : SpinLockAcquire(&XLogCtl->info_lck);
6989 7431 : recptr = XLogCtl->LogwrtRqst.Write;
6990 7431 : SpinLockRelease(&XLogCtl->info_lck);
6991 :
6992 7431 : return recptr;
6993 : }
6994 :
6995 : /*
6996 : * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6997 : * position known to be fsync'd to disk. This should only be used on a
6998 : * system that is known not to be in recovery.
6999 : */
7000 : XLogRecPtr
7001 235522 : GetFlushRecPtr(TimeLineID *insertTLI)
7002 : {
7003 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
7004 :
7005 235522 : RefreshXLogWriteResult(LogwrtResult);
7006 :
7007 : /*
7008 : * If we're writing and flushing WAL, the time line can't be changing, so
7009 : * no lock is required.
7010 : */
7011 235522 : if (insertTLI)
7012 37188 : *insertTLI = XLogCtl->InsertTimeLineID;
7013 :
7014 235522 : return LogwrtResult.Flush;
7015 : }
7016 :
7017 : /*
7018 : * GetWALInsertionTimeLine -- Returns the current timeline of a system that
7019 : * is not in recovery.
7020 : */
7021 : TimeLineID
7022 121587 : GetWALInsertionTimeLine(void)
7023 : {
7024 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
7025 :
7026 : /* Since the value can't be changing, no lock is required. */
7027 121587 : return XLogCtl->InsertTimeLineID;
7028 : }
7029 :
7030 : /*
7031 : * GetWALInsertionTimeLineIfSet -- If the system is not in recovery, returns
7032 : * the WAL insertion timeline; else, returns 0. Wherever possible, use
7033 : * GetWALInsertionTimeLine() instead, since it's cheaper. Note that this
7034 : * function decides recovery has ended as soon as the insert TLI is set, which
7035 : * happens before we set XLogCtl->SharedRecoveryState to RECOVERY_STATE_DONE.
7036 : */
7037 : TimeLineID
7038 1 : GetWALInsertionTimeLineIfSet(void)
7039 : {
7040 : TimeLineID insertTLI;
7041 :
7042 1 : SpinLockAcquire(&XLogCtl->info_lck);
7043 1 : insertTLI = XLogCtl->InsertTimeLineID;
7044 1 : SpinLockRelease(&XLogCtl->info_lck);
7045 :
7046 1 : return insertTLI;
7047 : }
7048 :
7049 : /*
7050 : * GetLastImportantRecPtr -- Returns the LSN of the last important record
7051 : * inserted. All records not explicitly marked as unimportant are considered
7052 : * important.
7053 : *
7054 : * The LSN is determined by computing the maximum of
7055 : * WALInsertLocks[i].lastImportantAt.
7056 : */
7057 : XLogRecPtr
7058 1776 : GetLastImportantRecPtr(void)
7059 : {
7060 1776 : XLogRecPtr res = InvalidXLogRecPtr;
7061 : int i;
7062 :
7063 15984 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
7064 : {
7065 : XLogRecPtr last_important;
7066 :
7067 : /*
7068 : * Need to take a lock to prevent torn reads of the LSN, which are
7069 : * possible on some of the supported platforms. WAL insert locks only
7070 : * support exclusive mode, so we have to use that.
7071 : */
7072 14208 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
7073 14208 : last_important = WALInsertLocks[i].l.lastImportantAt;
7074 14208 : LWLockRelease(&WALInsertLocks[i].l.lock);
7075 :
7076 14208 : if (res < last_important)
7077 3054 : res = last_important;
7078 : }
7079 :
7080 1776 : return res;
7081 : }
7082 :
7083 : /*
7084 : * Get the time and LSN of the last xlog segment switch
7085 : */
7086 : pg_time_t
7087 0 : GetLastSegSwitchData(XLogRecPtr *lastSwitchLSN)
7088 : {
7089 : pg_time_t result;
7090 :
7091 : /* Need WALWriteLock, but shared lock is sufficient */
7092 0 : LWLockAcquire(WALWriteLock, LW_SHARED);
7093 0 : result = XLogCtl->lastSegSwitchTime;
7094 0 : *lastSwitchLSN = XLogCtl->lastSegSwitchLSN;
7095 0 : LWLockRelease(WALWriteLock);
7096 :
7097 0 : return result;
7098 : }
7099 :
7100 : /*
7101 : * This must be called ONCE during postmaster or standalone-backend shutdown
7102 : */
7103 : void
7104 764 : ShutdownXLOG(int code, Datum arg)
7105 : {
7106 : /*
7107 : * We should have an aux process resource owner to use, and we should not
7108 : * be in a transaction that's installed some other resowner.
7109 : */
7110 : Assert(AuxProcessResourceOwner != NULL);
7111 : Assert(CurrentResourceOwner == NULL ||
7112 : CurrentResourceOwner == AuxProcessResourceOwner);
7113 764 : CurrentResourceOwner = AuxProcessResourceOwner;
7114 :
7115 : /* Don't be chatty in standalone mode */
7116 764 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
7117 : (errmsg("shutting down")));
7118 :
7119 : /*
7120 : * Signal walsenders to move to stopping state.
7121 : */
7122 764 : WalSndInitStopping();
7123 :
7124 : /*
7125 : * Wait for WAL senders to be in stopping state. This prevents commands
7126 : * from writing new WAL.
7127 : */
7128 764 : WalSndWaitStopping();
7129 :
7130 764 : if (RecoveryInProgress())
7131 60 : CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
7132 : else
7133 : {
7134 : /*
7135 : * If archiving is enabled, rotate the last XLOG file so that all the
7136 : * remaining records are archived (postmaster wakes up the archiver
7137 : * process one more time at the end of shutdown). The checkpoint
7138 : * record will go to the next XLOG file and won't be archived (yet).
7139 : */
7140 704 : if (XLogArchivingActive())
7141 16 : RequestXLogSwitch(false);
7142 :
7143 704 : CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
7144 : }
7145 764 : }
7146 :
7147 : /*
7148 : * Format checkpoint request flags as a space-separated string for
7149 : * log messages.
7150 : */
7151 : static const char *
7152 3204 : CheckpointFlagsString(int flags)
7153 : {
7154 : static char buf[128];
7155 :
7156 25632 : snprintf(buf, sizeof(buf), "%s%s%s%s%s%s%s%s",
7157 3204 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
7158 3204 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
7159 3204 : (flags & CHECKPOINT_FAST) ? " fast" : "",
7160 3204 : (flags & CHECKPOINT_FORCE) ? " force" : "",
7161 3204 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
7162 3204 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
7163 3204 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
7164 3204 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "");
7165 :
7166 3204 : return buf;
7167 : }
7168 :
7169 : /*
7170 : * Log start of a checkpoint.
7171 : */
7172 : static void
7173 1602 : LogCheckpointStart(int flags, bool restartpoint)
7174 : {
7175 1602 : if (restartpoint)
7176 212 : ereport(LOG,
7177 : /* translator: the placeholder shows checkpoint options */
7178 : (errmsg("restartpoint starting:%s",
7179 : CheckpointFlagsString(flags))));
7180 : else
7181 1390 : ereport(LOG,
7182 : /* translator: the placeholder shows checkpoint options */
7183 : (errmsg("checkpoint starting:%s",
7184 : CheckpointFlagsString(flags))));
7185 1602 : }
7186 :
7187 : /*
7188 : * Log end of a checkpoint.
7189 : */
7190 : static void
7191 1936 : LogCheckpointEnd(bool restartpoint, int flags)
7192 : {
7193 : long write_msecs,
7194 : sync_msecs,
7195 : total_msecs,
7196 : longest_msecs,
7197 : average_msecs;
7198 : uint64 average_sync_time;
7199 :
7200 1936 : CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
7201 :
7202 1936 : write_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_write_t,
7203 : CheckpointStats.ckpt_sync_t);
7204 :
7205 1936 : sync_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_sync_t,
7206 : CheckpointStats.ckpt_sync_end_t);
7207 :
7208 : /* Accumulate checkpoint timing summary data, in milliseconds. */
7209 1936 : PendingCheckpointerStats.write_time += write_msecs;
7210 1936 : PendingCheckpointerStats.sync_time += sync_msecs;
7211 :
7212 : /*
7213 : * All of the published timing statistics are accounted for. Only
7214 : * continue if a log message is to be written.
7215 : */
7216 1936 : if (!log_checkpoints)
7217 334 : return;
7218 :
7219 1602 : total_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_start_t,
7220 : CheckpointStats.ckpt_end_t);
7221 :
7222 : /*
7223 : * Timing values returned from CheckpointStats are in microseconds.
7224 : * Convert to milliseconds for consistent printing.
7225 : */
7226 1602 : longest_msecs = (long) ((CheckpointStats.ckpt_longest_sync + 999) / 1000);
7227 :
7228 1602 : average_sync_time = 0;
7229 1602 : if (CheckpointStats.ckpt_sync_rels > 0)
7230 0 : average_sync_time = CheckpointStats.ckpt_agg_sync_time /
7231 0 : CheckpointStats.ckpt_sync_rels;
7232 1602 : average_msecs = (long) ((average_sync_time + 999) / 1000);
7233 :
7234 : /*
7235 : * ControlFileLock is not required to see ControlFile->checkPoint and
7236 : * ->checkPointCopy here as we are the only updator of those variables at
7237 : * this moment.
7238 : */
7239 1602 : if (restartpoint)
7240 212 : ereport(LOG,
7241 : (errmsg("restartpoint complete:%s: wrote %d buffers (%.1f%%), "
7242 : "wrote %d SLRU buffers; %d WAL file(s) added, "
7243 : "%d removed, %d recycled; write=%ld.%03d s, "
7244 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
7245 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
7246 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
7247 : CheckpointFlagsString(flags),
7248 : CheckpointStats.ckpt_bufs_written,
7249 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
7250 : CheckpointStats.ckpt_slru_written,
7251 : CheckpointStats.ckpt_segs_added,
7252 : CheckpointStats.ckpt_segs_removed,
7253 : CheckpointStats.ckpt_segs_recycled,
7254 : write_msecs / 1000, (int) (write_msecs % 1000),
7255 : sync_msecs / 1000, (int) (sync_msecs % 1000),
7256 : total_msecs / 1000, (int) (total_msecs % 1000),
7257 : CheckpointStats.ckpt_sync_rels,
7258 : longest_msecs / 1000, (int) (longest_msecs % 1000),
7259 : average_msecs / 1000, (int) (average_msecs % 1000),
7260 : (int) (PrevCheckPointDistance / 1024.0),
7261 : (int) (CheckPointDistanceEstimate / 1024.0),
7262 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
7263 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
7264 : else
7265 1390 : ereport(LOG,
7266 : (errmsg("checkpoint complete:%s: wrote %d buffers (%.1f%%), "
7267 : "wrote %d SLRU buffers; %d WAL file(s) added, "
7268 : "%d removed, %d recycled; write=%ld.%03d s, "
7269 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
7270 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
7271 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
7272 : CheckpointFlagsString(flags),
7273 : CheckpointStats.ckpt_bufs_written,
7274 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
7275 : CheckpointStats.ckpt_slru_written,
7276 : CheckpointStats.ckpt_segs_added,
7277 : CheckpointStats.ckpt_segs_removed,
7278 : CheckpointStats.ckpt_segs_recycled,
7279 : write_msecs / 1000, (int) (write_msecs % 1000),
7280 : sync_msecs / 1000, (int) (sync_msecs % 1000),
7281 : total_msecs / 1000, (int) (total_msecs % 1000),
7282 : CheckpointStats.ckpt_sync_rels,
7283 : longest_msecs / 1000, (int) (longest_msecs % 1000),
7284 : average_msecs / 1000, (int) (average_msecs % 1000),
7285 : (int) (PrevCheckPointDistance / 1024.0),
7286 : (int) (CheckPointDistanceEstimate / 1024.0),
7287 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
7288 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
7289 : }
7290 :
7291 : /*
7292 : * Update the estimate of distance between checkpoints.
7293 : *
7294 : * The estimate is used to calculate the number of WAL segments to keep
7295 : * preallocated, see XLOGfileslop().
7296 : */
7297 : static void
7298 1936 : UpdateCheckPointDistanceEstimate(uint64 nbytes)
7299 : {
7300 : /*
7301 : * To estimate the number of segments consumed between checkpoints, keep a
7302 : * moving average of the amount of WAL generated in previous checkpoint
7303 : * cycles. However, if the load is bursty, with quiet periods and busy
7304 : * periods, we want to cater for the peak load. So instead of a plain
7305 : * moving average, let the average decline slowly if the previous cycle
7306 : * used less WAL than estimated, but bump it up immediately if it used
7307 : * more.
7308 : *
7309 : * When checkpoints are triggered by max_wal_size, this should converge to
7310 : * CheckpointSegments * wal_segment_size,
7311 : *
7312 : * Note: This doesn't pay any attention to what caused the checkpoint.
7313 : * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
7314 : * starting a base backup, are counted the same as those created
7315 : * automatically. The slow-decline will largely mask them out, if they are
7316 : * not frequent. If they are frequent, it seems reasonable to count them
7317 : * in as any others; if you issue a manual checkpoint every 5 minutes and
7318 : * never let a timed checkpoint happen, it makes sense to base the
7319 : * preallocation on that 5 minute interval rather than whatever
7320 : * checkpoint_timeout is set to.
7321 : */
7322 1936 : PrevCheckPointDistance = nbytes;
7323 1936 : if (CheckPointDistanceEstimate < nbytes)
7324 860 : CheckPointDistanceEstimate = nbytes;
7325 : else
7326 1076 : CheckPointDistanceEstimate =
7327 1076 : (0.90 * CheckPointDistanceEstimate + 0.10 * (double) nbytes);
7328 1936 : }
7329 :
7330 : /*
7331 : * Update the ps display for a process running a checkpoint. Note that
7332 : * this routine should not do any allocations so as it can be called
7333 : * from a critical section.
7334 : */
7335 : static void
7336 3872 : update_checkpoint_display(int flags, bool restartpoint, bool reset)
7337 : {
7338 : /*
7339 : * The status is reported only for end-of-recovery and shutdown
7340 : * checkpoints or shutdown restartpoints. Updating the ps display is
7341 : * useful in those situations as it may not be possible to rely on
7342 : * pg_stat_activity to see the status of the checkpointer or the startup
7343 : * process.
7344 : */
7345 3872 : if ((flags & (CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IS_SHUTDOWN)) == 0)
7346 2352 : return;
7347 :
7348 1520 : if (reset)
7349 760 : set_ps_display("");
7350 : else
7351 : {
7352 : char activitymsg[128];
7353 :
7354 2280 : snprintf(activitymsg, sizeof(activitymsg), "performing %s%s%s",
7355 760 : (flags & CHECKPOINT_END_OF_RECOVERY) ? "end-of-recovery " : "",
7356 760 : (flags & CHECKPOINT_IS_SHUTDOWN) ? "shutdown " : "",
7357 : restartpoint ? "restartpoint" : "checkpoint");
7358 760 : set_ps_display(activitymsg);
7359 : }
7360 : }
7361 :
7362 :
7363 : /*
7364 : * Perform a checkpoint --- either during shutdown, or on-the-fly
7365 : *
7366 : * flags is a bitwise OR of the following:
7367 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
7368 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
7369 : * CHECKPOINT_FAST: finish the checkpoint ASAP, ignoring
7370 : * checkpoint_completion_target parameter.
7371 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
7372 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
7373 : * CHECKPOINT_END_OF_RECOVERY).
7374 : * CHECKPOINT_FLUSH_UNLOGGED: also flush buffers of unlogged tables.
7375 : *
7376 : * Note: flags contains other bits, of interest here only for logging purposes.
7377 : * In particular note that this routine is synchronous and does not pay
7378 : * attention to CHECKPOINT_WAIT.
7379 : *
7380 : * If !shutdown then we are writing an online checkpoint. An XLOG_CHECKPOINT_REDO
7381 : * record is inserted into WAL at the logical location of the checkpoint, before
7382 : * flushing anything to disk, and when the checkpoint is eventually completed,
7383 : * and it is from this point that WAL replay will begin in the case of a recovery
7384 : * from this checkpoint. Once everything is written to disk, an
7385 : * XLOG_CHECKPOINT_ONLINE record is written to complete the checkpoint, and
7386 : * points back to the earlier XLOG_CHECKPOINT_REDO record. This mechanism allows
7387 : * other write-ahead log records to be written while the checkpoint is in
7388 : * progress, but we must be very careful about order of operations. This function
7389 : * may take many minutes to execute on a busy system.
7390 : *
7391 : * On the other hand, when shutdown is true, concurrent insertion into the
7392 : * write-ahead log is impossible, so there is no need for two separate records.
7393 : * In this case, we only insert an XLOG_CHECKPOINT_SHUTDOWN record, and it's
7394 : * both the record marking the completion of the checkpoint and the location
7395 : * from which WAL replay would begin if needed.
7396 : *
7397 : * Returns true if a new checkpoint was performed, or false if it was skipped
7398 : * because the system was idle.
7399 : */
7400 : bool
7401 1727 : CreateCheckPoint(int flags)
7402 : {
7403 : bool shutdown;
7404 : CheckPoint checkPoint;
7405 : XLogRecPtr recptr;
7406 : XLogSegNo _logSegNo;
7407 1727 : XLogCtlInsert *Insert = &XLogCtl->Insert;
7408 : uint32 freespace;
7409 : XLogRecPtr PriorRedoPtr;
7410 : XLogRecPtr last_important_lsn;
7411 : VirtualTransactionId *vxids;
7412 : int nvxids;
7413 1727 : int oldXLogAllowed = 0;
7414 :
7415 : /*
7416 : * An end-of-recovery checkpoint is really a shutdown checkpoint, just
7417 : * issued at a different time.
7418 : */
7419 1727 : if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
7420 734 : shutdown = true;
7421 : else
7422 993 : shutdown = false;
7423 :
7424 : /* sanity check */
7425 1727 : if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
7426 0 : elog(ERROR, "can't create a checkpoint during recovery");
7427 :
7428 : /*
7429 : * Prepare to accumulate statistics.
7430 : *
7431 : * Note: because it is possible for log_checkpoints to change while a
7432 : * checkpoint proceeds, we always accumulate stats, even if
7433 : * log_checkpoints is currently off.
7434 : */
7435 18997 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7436 1727 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7437 :
7438 : /*
7439 : * Let smgr prepare for checkpoint; this has to happen outside the
7440 : * critical section and before we determine the REDO pointer. Note that
7441 : * smgr must not do anything that'd have to be undone if we decide no
7442 : * checkpoint is needed.
7443 : */
7444 1727 : SyncPreCheckpoint();
7445 :
7446 : /* Run these points outside the critical section. */
7447 1727 : INJECTION_POINT("create-checkpoint-initial", NULL);
7448 1727 : INJECTION_POINT_LOAD("create-checkpoint-run");
7449 :
7450 : /*
7451 : * Use a critical section to force system panic if we have trouble.
7452 : */
7453 1727 : START_CRIT_SECTION();
7454 :
7455 1727 : if (shutdown)
7456 : {
7457 734 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7458 734 : ControlFile->state = DB_SHUTDOWNING;
7459 734 : UpdateControlFile();
7460 734 : LWLockRelease(ControlFileLock);
7461 : }
7462 :
7463 : /* Begin filling in the checkpoint WAL record */
7464 22451 : MemSet(&checkPoint, 0, sizeof(checkPoint));
7465 1727 : checkPoint.time = (pg_time_t) time(NULL);
7466 :
7467 : /*
7468 : * For Hot Standby, derive the oldestActiveXid before we fix the redo
7469 : * pointer. This allows us to begin accumulating changes to assemble our
7470 : * starting snapshot of locks and transactions.
7471 : */
7472 1727 : if (!shutdown && XLogStandbyInfoActive())
7473 951 : checkPoint.oldestActiveXid = GetOldestActiveTransactionId(false, true);
7474 : else
7475 776 : checkPoint.oldestActiveXid = InvalidTransactionId;
7476 :
7477 : /*
7478 : * Get location of last important record before acquiring insert locks (as
7479 : * GetLastImportantRecPtr() also locks WAL locks).
7480 : */
7481 1727 : last_important_lsn = GetLastImportantRecPtr();
7482 :
7483 : /*
7484 : * If this isn't a shutdown or forced checkpoint, and if there has been no
7485 : * WAL activity requiring a checkpoint, skip it. The idea here is to
7486 : * avoid inserting duplicate checkpoints when the system is idle.
7487 : */
7488 1727 : if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
7489 : CHECKPOINT_FORCE)) == 0)
7490 : {
7491 208 : if (last_important_lsn == ControlFile->checkPoint)
7492 : {
7493 3 : END_CRIT_SECTION();
7494 3 : ereport(DEBUG1,
7495 : (errmsg_internal("checkpoint skipped because system is idle")));
7496 3 : return false;
7497 : }
7498 : }
7499 :
7500 : /*
7501 : * An end-of-recovery checkpoint is created before anyone is allowed to
7502 : * write WAL. To allow us to write the checkpoint record, temporarily
7503 : * enable XLogInsertAllowed.
7504 : */
7505 1724 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7506 30 : oldXLogAllowed = LocalSetXLogInsertAllowed();
7507 :
7508 1724 : checkPoint.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7509 1724 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7510 30 : checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7511 : else
7512 1694 : checkPoint.PrevTimeLineID = checkPoint.ThisTimeLineID;
7513 :
7514 : /*
7515 : * We must block concurrent insertions while examining insert state.
7516 : */
7517 1724 : WALInsertLockAcquireExclusive();
7518 :
7519 1724 : checkPoint.fullPageWrites = Insert->fullPageWrites;
7520 1724 : checkPoint.wal_level = wal_level;
7521 :
7522 : /*
7523 : * Get the current data_checksum_version value from xlogctl, valid at the
7524 : * time of the checkpoint.
7525 : */
7526 1724 : SpinLockAcquire(&XLogCtl->info_lck);
7527 1724 : checkPoint.dataChecksumState = XLogCtl->data_checksum_version;
7528 1724 : SpinLockRelease(&XLogCtl->info_lck);
7529 :
7530 1724 : if (shutdown)
7531 : {
7532 734 : XLogRecPtr curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
7533 :
7534 : /*
7535 : * Compute new REDO record ptr = location of next XLOG record.
7536 : *
7537 : * Since this is a shutdown checkpoint, there can't be any concurrent
7538 : * WAL insertion.
7539 : */
7540 734 : freespace = INSERT_FREESPACE(curInsert);
7541 734 : if (freespace == 0)
7542 : {
7543 0 : if (XLogSegmentOffset(curInsert, wal_segment_size) == 0)
7544 0 : curInsert += SizeOfXLogLongPHD;
7545 : else
7546 0 : curInsert += SizeOfXLogShortPHD;
7547 : }
7548 734 : checkPoint.redo = curInsert;
7549 :
7550 : /*
7551 : * Here we update the shared RedoRecPtr for future XLogInsert calls;
7552 : * this must be done while holding all the insertion locks.
7553 : *
7554 : * Note: if we fail to complete the checkpoint, RedoRecPtr will be
7555 : * left pointing past where it really needs to point. This is okay;
7556 : * the only consequence is that XLogInsert might back up whole buffers
7557 : * that it didn't really need to. We can't postpone advancing
7558 : * RedoRecPtr because XLogInserts that happen while we are dumping
7559 : * buffers must assume that their buffer changes are not included in
7560 : * the checkpoint.
7561 : */
7562 734 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
7563 : }
7564 :
7565 : /*
7566 : * Now we can release the WAL insertion locks, allowing other xacts to
7567 : * proceed while we are flushing disk buffers.
7568 : */
7569 1724 : WALInsertLockRelease();
7570 :
7571 : /*
7572 : * If this is an online checkpoint, we have not yet determined the redo
7573 : * point. We do so now by inserting the special XLOG_CHECKPOINT_REDO
7574 : * record; the LSN at which it starts becomes the new redo pointer. We
7575 : * don't do this for a shutdown checkpoint, because in that case no WAL
7576 : * can be written between the redo point and the insertion of the
7577 : * checkpoint record itself, so the checkpoint record itself serves to
7578 : * mark the redo point.
7579 : */
7580 1724 : if (!shutdown)
7581 : {
7582 : xl_checkpoint_redo redo_rec;
7583 :
7584 990 : WALInsertLockAcquire();
7585 990 : redo_rec.wal_level = wal_level;
7586 990 : SpinLockAcquire(&XLogCtl->info_lck);
7587 990 : redo_rec.data_checksum_version = XLogCtl->data_checksum_version;
7588 990 : SpinLockRelease(&XLogCtl->info_lck);
7589 990 : WALInsertLockRelease();
7590 :
7591 : /* Include WAL level in record for WAL summarizer's benefit. */
7592 990 : XLogBeginInsert();
7593 990 : XLogRegisterData(&redo_rec, sizeof(xl_checkpoint_redo));
7594 990 : (void) XLogInsert(RM_XLOG_ID, XLOG_CHECKPOINT_REDO);
7595 :
7596 : /*
7597 : * XLogInsertRecord will have updated XLogCtl->Insert.RedoRecPtr in
7598 : * shared memory and RedoRecPtr in backend-local memory, but we need
7599 : * to copy that into the record that will be inserted when the
7600 : * checkpoint is complete.
7601 : */
7602 990 : checkPoint.redo = RedoRecPtr;
7603 : }
7604 :
7605 : /* Update the info_lck-protected copy of RedoRecPtr as well */
7606 1724 : SpinLockAcquire(&XLogCtl->info_lck);
7607 1724 : XLogCtl->RedoRecPtr = checkPoint.redo;
7608 1724 : SpinLockRelease(&XLogCtl->info_lck);
7609 :
7610 : /*
7611 : * If enabled, log checkpoint start. We postpone this until now so as not
7612 : * to log anything if we decided to skip the checkpoint.
7613 : */
7614 1724 : if (log_checkpoints)
7615 1390 : LogCheckpointStart(flags, false);
7616 :
7617 1724 : INJECTION_POINT_CACHED("create-checkpoint-run", NULL);
7618 :
7619 : /* Update the process title */
7620 1724 : update_checkpoint_display(flags, false, false);
7621 :
7622 : TRACE_POSTGRESQL_CHECKPOINT_START(flags);
7623 :
7624 : /*
7625 : * Get the other info we need for the checkpoint record.
7626 : *
7627 : * We don't need to save oldestClogXid in the checkpoint, it only matters
7628 : * for the short period in which clog is being truncated, and if we crash
7629 : * during that we'll redo the clog truncation and fix up oldestClogXid
7630 : * there.
7631 : */
7632 1724 : LWLockAcquire(XidGenLock, LW_SHARED);
7633 1724 : checkPoint.nextXid = TransamVariables->nextXid;
7634 1724 : checkPoint.oldestXid = TransamVariables->oldestXid;
7635 1724 : checkPoint.oldestXidDB = TransamVariables->oldestXidDB;
7636 1724 : LWLockRelease(XidGenLock);
7637 :
7638 1724 : LWLockAcquire(CommitTsLock, LW_SHARED);
7639 1724 : checkPoint.oldestCommitTsXid = TransamVariables->oldestCommitTsXid;
7640 1724 : checkPoint.newestCommitTsXid = TransamVariables->newestCommitTsXid;
7641 1724 : LWLockRelease(CommitTsLock);
7642 :
7643 1724 : LWLockAcquire(OidGenLock, LW_SHARED);
7644 1724 : checkPoint.nextOid = TransamVariables->nextOid;
7645 1724 : if (!shutdown)
7646 990 : checkPoint.nextOid += TransamVariables->oidCount;
7647 1724 : LWLockRelease(OidGenLock);
7648 :
7649 1724 : checkPoint.logicalDecodingEnabled = IsLogicalDecodingEnabled();
7650 :
7651 1724 : MultiXactGetCheckptMulti(shutdown,
7652 : &checkPoint.nextMulti,
7653 : &checkPoint.nextMultiOffset,
7654 : &checkPoint.oldestMulti,
7655 : &checkPoint.oldestMultiDB);
7656 :
7657 : /*
7658 : * Having constructed the checkpoint record, ensure all shmem disk buffers
7659 : * and commit-log buffers are flushed to disk.
7660 : *
7661 : * This I/O could fail for various reasons. If so, we will fail to
7662 : * complete the checkpoint, but there is no reason to force a system
7663 : * panic. Accordingly, exit critical section while doing it.
7664 : */
7665 1724 : END_CRIT_SECTION();
7666 :
7667 : /*
7668 : * In some cases there are groups of actions that must all occur on one
7669 : * side or the other of a checkpoint record. Before flushing the
7670 : * checkpoint record we must explicitly wait for any backend currently
7671 : * performing those groups of actions.
7672 : *
7673 : * One example is end of transaction, so we must wait for any transactions
7674 : * that are currently in commit critical sections. If an xact inserted
7675 : * its commit record into XLOG just before the REDO point, then a crash
7676 : * restart from the REDO point would not replay that record, which means
7677 : * that our flushing had better include the xact's update of pg_xact. So
7678 : * we wait till he's out of his commit critical section before proceeding.
7679 : * See notes in RecordTransactionCommit().
7680 : *
7681 : * Because we've already released the insertion locks, this test is a bit
7682 : * fuzzy: it is possible that we will wait for xacts we didn't really need
7683 : * to wait for. But the delay should be short and it seems better to make
7684 : * checkpoint take a bit longer than to hold off insertions longer than
7685 : * necessary. (In fact, the whole reason we have this issue is that xact.c
7686 : * does commit record XLOG insertion and clog update as two separate steps
7687 : * protected by different locks, but again that seems best on grounds of
7688 : * minimizing lock contention.)
7689 : *
7690 : * A transaction that has not yet set delayChkptFlags when we look cannot
7691 : * be at risk, since it has not inserted its commit record yet; and one
7692 : * that's already cleared it is not at risk either, since it's done fixing
7693 : * clog and we will correctly flush the update below. So we cannot miss
7694 : * any xacts we need to wait for.
7695 : */
7696 1724 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
7697 1724 : if (nvxids > 0)
7698 : {
7699 : do
7700 : {
7701 : /*
7702 : * Keep absorbing fsync requests while we wait. There could even
7703 : * be a deadlock if we don't, if the process that prevents the
7704 : * checkpoint is trying to add a request to the queue.
7705 : */
7706 26 : AbsorbSyncRequests();
7707 :
7708 26 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
7709 26 : pg_usleep(10000L); /* wait for 10 msec */
7710 26 : pgstat_report_wait_end();
7711 26 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7712 : DELAY_CHKPT_START));
7713 : }
7714 1724 : pfree(vxids);
7715 :
7716 1724 : CheckPointGuts(checkPoint.redo, flags);
7717 :
7718 1724 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
7719 1724 : if (nvxids > 0)
7720 : {
7721 : do
7722 : {
7723 0 : AbsorbSyncRequests();
7724 :
7725 0 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_COMPLETE);
7726 0 : pg_usleep(10000L); /* wait for 10 msec */
7727 0 : pgstat_report_wait_end();
7728 0 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7729 : DELAY_CHKPT_COMPLETE));
7730 : }
7731 1724 : pfree(vxids);
7732 :
7733 : /*
7734 : * Take a snapshot of running transactions and write this to WAL. This
7735 : * allows us to reconstruct the state of running transactions during
7736 : * archive recovery, if required. Skip, if this info disabled.
7737 : *
7738 : * If we are shutting down, or Startup process is completing crash
7739 : * recovery we don't need to write running xact data.
7740 : */
7741 1724 : if (!shutdown && XLogStandbyInfoActive())
7742 948 : LogStandbySnapshot(InvalidOid);
7743 :
7744 1724 : START_CRIT_SECTION();
7745 :
7746 : /*
7747 : * Now insert the checkpoint record into XLOG.
7748 : */
7749 1724 : XLogBeginInsert();
7750 1724 : XLogRegisterData(&checkPoint, sizeof(checkPoint));
7751 1724 : recptr = XLogInsert(RM_XLOG_ID,
7752 : shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
7753 : XLOG_CHECKPOINT_ONLINE);
7754 :
7755 1724 : XLogFlush(recptr);
7756 :
7757 : /*
7758 : * We mustn't write any new WAL after a shutdown checkpoint, or it will be
7759 : * overwritten at next startup. No-one should even try, this just allows
7760 : * sanity-checking. In the case of an end-of-recovery checkpoint, we want
7761 : * to just temporarily disable writing until the system has exited
7762 : * recovery.
7763 : */
7764 1724 : if (shutdown)
7765 : {
7766 734 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7767 30 : LocalXLogInsertAllowed = oldXLogAllowed;
7768 : else
7769 704 : LocalXLogInsertAllowed = 0; /* never again write WAL */
7770 : }
7771 :
7772 : /*
7773 : * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
7774 : * = end of actual checkpoint record.
7775 : */
7776 1724 : if (shutdown && checkPoint.redo != ProcLastRecPtr)
7777 0 : ereport(PANIC,
7778 : (errmsg("concurrent write-ahead log activity while database system is shutting down")));
7779 :
7780 : /*
7781 : * Remember the prior checkpoint's redo ptr for
7782 : * UpdateCheckPointDistanceEstimate()
7783 : */
7784 1724 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7785 :
7786 : /*
7787 : * Update the control file.
7788 : */
7789 1724 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7790 1724 : if (shutdown)
7791 734 : ControlFile->state = DB_SHUTDOWNED;
7792 1724 : ControlFile->checkPoint = ProcLastRecPtr;
7793 1724 : ControlFile->checkPointCopy = checkPoint;
7794 : /* crash recovery should always recover to the end of WAL */
7795 1724 : ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
7796 1724 : ControlFile->minRecoveryPointTLI = 0;
7797 :
7798 : /*
7799 : * Persist unloggedLSN value. It's reset on crash recovery, so this goes
7800 : * unused on non-shutdown checkpoints, but seems useful to store it always
7801 : * for debugging purposes.
7802 : */
7803 1724 : ControlFile->unloggedLSN = pg_atomic_read_membarrier_u64(&XLogCtl->unloggedLSN);
7804 :
7805 1724 : UpdateControlFile();
7806 1724 : LWLockRelease(ControlFileLock);
7807 :
7808 : /*
7809 : * We are now done with critical updates; no need for system panic if we
7810 : * have trouble while fooling with old log segments.
7811 : */
7812 1724 : END_CRIT_SECTION();
7813 :
7814 : /*
7815 : * WAL summaries end when the next XLOG_CHECKPOINT_REDO or
7816 : * XLOG_CHECKPOINT_SHUTDOWN record is reached. This is the first point
7817 : * where (a) we're not inside of a critical section and (b) we can be
7818 : * certain that the relevant record has been flushed to disk, which must
7819 : * happen before it can be summarized.
7820 : *
7821 : * If this is a shutdown checkpoint, then this happens reasonably
7822 : * promptly: we've only just inserted and flushed the
7823 : * XLOG_CHECKPOINT_SHUTDOWN record. If this is not a shutdown checkpoint,
7824 : * then this might not be very prompt at all: the XLOG_CHECKPOINT_REDO
7825 : * record was written before we began flushing data to disk, and that
7826 : * could be many minutes ago at this point. However, we don't XLogFlush()
7827 : * after inserting that record, so we're not guaranteed that it's on disk
7828 : * until after the above call that flushes the XLOG_CHECKPOINT_ONLINE
7829 : * record.
7830 : */
7831 1724 : WakeupWalSummarizer();
7832 :
7833 : /*
7834 : * Let smgr do post-checkpoint cleanup (eg, deleting old files).
7835 : */
7836 1724 : SyncPostCheckpoint();
7837 :
7838 : /*
7839 : * Update the average distance between checkpoints if the prior checkpoint
7840 : * exists.
7841 : */
7842 1724 : if (XLogRecPtrIsValid(PriorRedoPtr))
7843 1724 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7844 :
7845 1724 : INJECTION_POINT("checkpoint-before-old-wal-removal", NULL);
7846 :
7847 : /*
7848 : * Delete old log files, those no longer needed for last checkpoint to
7849 : * prevent the disk holding the xlog from growing full.
7850 : */
7851 1724 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7852 1724 : KeepLogSeg(recptr, &_logSegNo);
7853 1724 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7854 : _logSegNo, InvalidOid,
7855 : InvalidTransactionId))
7856 : {
7857 : /*
7858 : * Some slots have been invalidated; recalculate the old-segment
7859 : * horizon, starting again from RedoRecPtr.
7860 : */
7861 4 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7862 4 : KeepLogSeg(recptr, &_logSegNo);
7863 : }
7864 1724 : _logSegNo--;
7865 1724 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
7866 : checkPoint.ThisTimeLineID);
7867 :
7868 : /*
7869 : * Make more log segments if needed. (Do this after recycling old log
7870 : * segments, since that may supply some of the needed files.)
7871 : */
7872 1724 : if (!shutdown)
7873 990 : PreallocXlogFiles(recptr, checkPoint.ThisTimeLineID);
7874 :
7875 : /*
7876 : * Truncate pg_subtrans if possible. We can throw away all data before
7877 : * the oldest XMIN of any running transaction. No future transaction will
7878 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7879 : * in subtrans.c). During recovery, though, we mustn't do this because
7880 : * StartupSUBTRANS hasn't been called yet.
7881 : */
7882 1724 : if (!RecoveryInProgress())
7883 1694 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7884 :
7885 : /* Real work is done; log and update stats. */
7886 1724 : LogCheckpointEnd(false, flags);
7887 :
7888 : /* Reset the process title */
7889 1724 : update_checkpoint_display(flags, false, true);
7890 :
7891 : TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
7892 : NBuffers,
7893 : CheckpointStats.ckpt_segs_added,
7894 : CheckpointStats.ckpt_segs_removed,
7895 : CheckpointStats.ckpt_segs_recycled);
7896 :
7897 1724 : return true;
7898 : }
7899 :
7900 : /*
7901 : * Mark the end of recovery in WAL though without running a full checkpoint.
7902 : * We can expect that a restartpoint is likely to be in progress as we
7903 : * do this, though we are unwilling to wait for it to complete.
7904 : *
7905 : * CreateRestartPoint() allows for the case where recovery may end before
7906 : * the restartpoint completes so there is no concern of concurrent behaviour.
7907 : */
7908 : static void
7909 48 : CreateEndOfRecoveryRecord(void)
7910 : {
7911 : xl_end_of_recovery xlrec;
7912 : XLogRecPtr recptr;
7913 :
7914 : /* sanity check */
7915 48 : if (!RecoveryInProgress())
7916 0 : elog(ERROR, "can only be used to end recovery");
7917 :
7918 48 : xlrec.end_time = GetCurrentTimestamp();
7919 48 : xlrec.wal_level = wal_level;
7920 :
7921 48 : WALInsertLockAcquireExclusive();
7922 48 : xlrec.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7923 48 : xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7924 48 : WALInsertLockRelease();
7925 :
7926 48 : START_CRIT_SECTION();
7927 :
7928 48 : XLogBeginInsert();
7929 48 : XLogRegisterData(&xlrec, sizeof(xl_end_of_recovery));
7930 48 : recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY);
7931 :
7932 48 : XLogFlush(recptr);
7933 :
7934 : /*
7935 : * Update the control file so that crash recovery can follow the timeline
7936 : * changes to this point.
7937 : */
7938 48 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7939 48 : ControlFile->minRecoveryPoint = recptr;
7940 48 : ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
7941 :
7942 : /* start with the latest checksum version (as of the end of recovery) */
7943 48 : SpinLockAcquire(&XLogCtl->info_lck);
7944 48 : ControlFile->data_checksum_version = XLogCtl->data_checksum_version;
7945 48 : SpinLockRelease(&XLogCtl->info_lck);
7946 :
7947 48 : UpdateControlFile();
7948 48 : LWLockRelease(ControlFileLock);
7949 :
7950 48 : END_CRIT_SECTION();
7951 48 : }
7952 :
7953 : /*
7954 : * Write an OVERWRITE_CONTRECORD message.
7955 : *
7956 : * When on WAL replay we expect a continuation record at the start of a page
7957 : * that is not there, recovery ends and WAL writing resumes at that point.
7958 : * But it's wrong to resume writing new WAL back at the start of the record
7959 : * that was broken, because downstream consumers of that WAL (physical
7960 : * replicas) are not prepared to "rewind". So the first action after
7961 : * finishing replay of all valid WAL must be to write a record of this type
7962 : * at the point where the contrecord was missing; to support xlogreader
7963 : * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
7964 : * to the page header where the record occurs. xlogreader has an ad-hoc
7965 : * mechanism to report metadata about the broken record, which is what we
7966 : * use here.
7967 : *
7968 : * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
7969 : * skip the record it was reading, and pass back the LSN of the skipped
7970 : * record, so that its caller can verify (on "replay" of that record) that the
7971 : * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
7972 : *
7973 : * 'aborted_lsn' is the beginning position of the record that was incomplete.
7974 : * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
7975 : * beginning of the XLOG page where the record is to be inserted. They must
7976 : * match the current WAL insert position, they're passed here just so that we
7977 : * can verify that.
7978 : */
7979 : static XLogRecPtr
7980 11 : CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn, XLogRecPtr pagePtr,
7981 : TimeLineID newTLI)
7982 : {
7983 : xl_overwrite_contrecord xlrec;
7984 : XLogRecPtr recptr;
7985 : XLogPageHeader pagehdr;
7986 : XLogRecPtr startPos;
7987 :
7988 : /* sanity checks */
7989 11 : if (!RecoveryInProgress())
7990 0 : elog(ERROR, "can only be used at end of recovery");
7991 11 : if (pagePtr % XLOG_BLCKSZ != 0)
7992 0 : elog(ERROR, "invalid position for missing continuation record %X/%08X",
7993 : LSN_FORMAT_ARGS(pagePtr));
7994 :
7995 : /* The current WAL insert position should be right after the page header */
7996 11 : startPos = pagePtr;
7997 11 : if (XLogSegmentOffset(startPos, wal_segment_size) == 0)
7998 1 : startPos += SizeOfXLogLongPHD;
7999 : else
8000 10 : startPos += SizeOfXLogShortPHD;
8001 11 : recptr = GetXLogInsertRecPtr();
8002 11 : if (recptr != startPos)
8003 0 : elog(ERROR, "invalid WAL insert position %X/%08X for OVERWRITE_CONTRECORD",
8004 : LSN_FORMAT_ARGS(recptr));
8005 :
8006 11 : START_CRIT_SECTION();
8007 :
8008 : /*
8009 : * Initialize the XLOG page header (by GetXLogBuffer), and set the
8010 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
8011 : *
8012 : * No other backend is allowed to write WAL yet, so acquiring the WAL
8013 : * insertion lock is just pro forma.
8014 : */
8015 11 : WALInsertLockAcquire();
8016 11 : pagehdr = (XLogPageHeader) GetXLogBuffer(pagePtr, newTLI);
8017 11 : pagehdr->xlp_info |= XLP_FIRST_IS_OVERWRITE_CONTRECORD;
8018 11 : WALInsertLockRelease();
8019 :
8020 : /*
8021 : * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
8022 : * page. We know it becomes the first record, because no other backend is
8023 : * allowed to write WAL yet.
8024 : */
8025 11 : XLogBeginInsert();
8026 11 : xlrec.overwritten_lsn = aborted_lsn;
8027 11 : xlrec.overwrite_time = GetCurrentTimestamp();
8028 11 : XLogRegisterData(&xlrec, sizeof(xl_overwrite_contrecord));
8029 11 : recptr = XLogInsert(RM_XLOG_ID, XLOG_OVERWRITE_CONTRECORD);
8030 :
8031 : /* check that the record was inserted to the right place */
8032 11 : if (ProcLastRecPtr != startPos)
8033 0 : elog(ERROR, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%08X",
8034 : LSN_FORMAT_ARGS(ProcLastRecPtr));
8035 :
8036 11 : XLogFlush(recptr);
8037 :
8038 11 : END_CRIT_SECTION();
8039 :
8040 11 : return recptr;
8041 : }
8042 :
8043 : /*
8044 : * Flush all data in shared memory to disk, and fsync
8045 : *
8046 : * This is the common code shared between regular checkpoints and
8047 : * recovery restartpoints.
8048 : */
8049 : static void
8050 1936 : CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
8051 : {
8052 1936 : CheckPointRelationMap();
8053 1936 : CheckPointReplicationSlots(flags & CHECKPOINT_IS_SHUTDOWN);
8054 1936 : CheckPointSnapBuild();
8055 1936 : CheckPointLogicalRewriteHeap();
8056 1936 : CheckPointReplicationOrigin();
8057 :
8058 : /* Write out all dirty data in SLRUs and the main buffer pool */
8059 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags);
8060 1936 : CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
8061 1936 : CheckPointCLOG();
8062 1936 : CheckPointCommitTs();
8063 1936 : CheckPointSUBTRANS();
8064 1936 : CheckPointMultiXact();
8065 1936 : CheckPointPredicate();
8066 1936 : CheckPointBuffers(flags);
8067 :
8068 : /* Perform all queued up fsyncs */
8069 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
8070 1936 : CheckpointStats.ckpt_sync_t = GetCurrentTimestamp();
8071 1936 : ProcessSyncRequests();
8072 1936 : CheckpointStats.ckpt_sync_end_t = GetCurrentTimestamp();
8073 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
8074 :
8075 : /* We deliberately delay 2PC checkpointing as long as possible */
8076 1936 : CheckPointTwoPhase(checkPointRedo);
8077 1936 : }
8078 :
8079 : /*
8080 : * Save a checkpoint for recovery restart if appropriate
8081 : *
8082 : * This function is called each time a checkpoint record is read from XLOG.
8083 : * It must determine whether the checkpoint represents a safe restartpoint or
8084 : * not. If so, the checkpoint record is stashed in shared memory so that
8085 : * CreateRestartPoint can consult it. (Note that the latter function is
8086 : * executed by the checkpointer, while this one will be executed by the
8087 : * startup process.)
8088 : */
8089 : static void
8090 741 : RecoveryRestartPoint(const CheckPoint *checkPoint, XLogReaderState *record)
8091 : {
8092 : /*
8093 : * Also refrain from creating a restartpoint if we have seen any
8094 : * references to non-existent pages. Restarting recovery from the
8095 : * restartpoint would not see the references, so we would lose the
8096 : * cross-check that the pages belonged to a relation that was dropped
8097 : * later.
8098 : */
8099 741 : if (XLogHaveInvalidPages())
8100 : {
8101 0 : elog(DEBUG2,
8102 : "could not record restart point at %X/%08X because there are unresolved references to invalid pages",
8103 : LSN_FORMAT_ARGS(checkPoint->redo));
8104 0 : return;
8105 : }
8106 :
8107 : /*
8108 : * Copy the checkpoint record to shared memory, so that checkpointer can
8109 : * work out the next time it wants to perform a restartpoint.
8110 : */
8111 741 : SpinLockAcquire(&XLogCtl->info_lck);
8112 741 : XLogCtl->lastCheckPointRecPtr = record->ReadRecPtr;
8113 741 : XLogCtl->lastCheckPointEndPtr = record->EndRecPtr;
8114 741 : XLogCtl->lastCheckPoint = *checkPoint;
8115 741 : SpinLockRelease(&XLogCtl->info_lck);
8116 : }
8117 :
8118 : /*
8119 : * Establish a restartpoint if possible.
8120 : *
8121 : * This is similar to CreateCheckPoint, but is used during WAL recovery
8122 : * to establish a point from which recovery can roll forward without
8123 : * replaying the entire recovery log.
8124 : *
8125 : * Returns true if a new restartpoint was established. We can only establish
8126 : * a restartpoint if we have replayed a safe checkpoint record since last
8127 : * restartpoint.
8128 : */
8129 : bool
8130 608 : CreateRestartPoint(int flags)
8131 : {
8132 : XLogRecPtr lastCheckPointRecPtr;
8133 : XLogRecPtr lastCheckPointEndPtr;
8134 : CheckPoint lastCheckPoint;
8135 : XLogRecPtr PriorRedoPtr;
8136 : XLogRecPtr receivePtr;
8137 : XLogRecPtr replayPtr;
8138 : TimeLineID replayTLI;
8139 : XLogRecPtr endptr;
8140 : XLogSegNo _logSegNo;
8141 : TimestampTz xtime;
8142 :
8143 : /* Concurrent checkpoint/restartpoint cannot happen */
8144 : Assert(!IsUnderPostmaster || MyBackendType == B_CHECKPOINTER);
8145 :
8146 : /* Get a local copy of the last safe checkpoint record. */
8147 608 : SpinLockAcquire(&XLogCtl->info_lck);
8148 608 : lastCheckPointRecPtr = XLogCtl->lastCheckPointRecPtr;
8149 608 : lastCheckPointEndPtr = XLogCtl->lastCheckPointEndPtr;
8150 608 : lastCheckPoint = XLogCtl->lastCheckPoint;
8151 608 : SpinLockRelease(&XLogCtl->info_lck);
8152 :
8153 : /*
8154 : * Check that we're still in recovery mode. It's ok if we exit recovery
8155 : * mode after this check, the restart point is valid anyway.
8156 : */
8157 608 : if (!RecoveryInProgress())
8158 : {
8159 0 : ereport(DEBUG2,
8160 : (errmsg_internal("skipping restartpoint, recovery has already ended")));
8161 0 : return false;
8162 : }
8163 :
8164 : /*
8165 : * If the last checkpoint record we've replayed is already our last
8166 : * restartpoint, we can't perform a new restart point. We still update
8167 : * minRecoveryPoint in that case, so that if this is a shutdown restart
8168 : * point, we won't start up earlier than before. That's not strictly
8169 : * necessary, but when hot standby is enabled, it would be rather weird if
8170 : * the database opened up for read-only connections at a point-in-time
8171 : * before the last shutdown. Such time travel is still possible in case of
8172 : * immediate shutdown, though.
8173 : *
8174 : * We don't explicitly advance minRecoveryPoint when we do create a
8175 : * restartpoint. It's assumed that flushing the buffers will do that as a
8176 : * side-effect.
8177 : */
8178 608 : if (!XLogRecPtrIsValid(lastCheckPointRecPtr) ||
8179 277 : lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
8180 : {
8181 396 : ereport(DEBUG2,
8182 : errmsg_internal("skipping restartpoint, already performed at %X/%08X",
8183 : LSN_FORMAT_ARGS(lastCheckPoint.redo)));
8184 :
8185 396 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
8186 396 : if (flags & CHECKPOINT_IS_SHUTDOWN)
8187 : {
8188 34 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8189 34 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
8190 34 : UpdateControlFile();
8191 34 : LWLockRelease(ControlFileLock);
8192 : }
8193 396 : return false;
8194 : }
8195 :
8196 : /*
8197 : * Update the shared RedoRecPtr so that the startup process can calculate
8198 : * the number of segments replayed since last restartpoint, and request a
8199 : * restartpoint if it exceeds CheckPointSegments.
8200 : *
8201 : * Like in CreateCheckPoint(), hold off insertions to update it, although
8202 : * during recovery this is just pro forma, because no WAL insertions are
8203 : * happening.
8204 : */
8205 212 : WALInsertLockAcquireExclusive();
8206 212 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = lastCheckPoint.redo;
8207 212 : WALInsertLockRelease();
8208 :
8209 : /* Also update the info_lck-protected copy */
8210 212 : SpinLockAcquire(&XLogCtl->info_lck);
8211 212 : XLogCtl->RedoRecPtr = lastCheckPoint.redo;
8212 212 : SpinLockRelease(&XLogCtl->info_lck);
8213 :
8214 : /*
8215 : * Prepare to accumulate statistics.
8216 : *
8217 : * Note: because it is possible for log_checkpoints to change while a
8218 : * checkpoint proceeds, we always accumulate stats, even if
8219 : * log_checkpoints is currently off.
8220 : */
8221 2332 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
8222 212 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
8223 :
8224 212 : if (log_checkpoints)
8225 212 : LogCheckpointStart(flags, true);
8226 :
8227 : /* Update the process title */
8228 212 : update_checkpoint_display(flags, true, false);
8229 :
8230 212 : CheckPointGuts(lastCheckPoint.redo, flags);
8231 :
8232 : /*
8233 : * This location needs to be after CheckPointGuts() to ensure that some
8234 : * work has already happened during this checkpoint.
8235 : */
8236 212 : INJECTION_POINT("create-restart-point", NULL);
8237 :
8238 : /*
8239 : * Remember the prior checkpoint's redo ptr for
8240 : * UpdateCheckPointDistanceEstimate()
8241 : */
8242 212 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
8243 :
8244 : /*
8245 : * Update pg_control, using current time. Check that it still shows an
8246 : * older checkpoint, else do nothing; this is a quick hack to make sure
8247 : * nothing really bad happens if somehow we get here after the
8248 : * end-of-recovery checkpoint.
8249 : */
8250 212 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8251 212 : if (ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
8252 : {
8253 : /*
8254 : * Update the checkpoint information. We do this even if the cluster
8255 : * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
8256 : * segments recycled below.
8257 : */
8258 212 : ControlFile->checkPoint = lastCheckPointRecPtr;
8259 212 : ControlFile->checkPointCopy = lastCheckPoint;
8260 :
8261 : /*
8262 : * Ensure minRecoveryPoint is past the checkpoint record and update it
8263 : * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
8264 : * this will have happened already while writing out dirty buffers,
8265 : * but not necessarily - e.g. because no buffers were dirtied. We do
8266 : * this because a backup performed in recovery uses minRecoveryPoint
8267 : * to determine which WAL files must be included in the backup, and
8268 : * the file (or files) containing the checkpoint record must be
8269 : * included, at a minimum. Note that for an ordinary restart of
8270 : * recovery there's no value in having the minimum recovery point any
8271 : * earlier than this anyway, because redo will begin just after the
8272 : * checkpoint record.
8273 : */
8274 212 : if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
8275 : {
8276 212 : if (ControlFile->minRecoveryPoint < lastCheckPointEndPtr)
8277 : {
8278 19 : ControlFile->minRecoveryPoint = lastCheckPointEndPtr;
8279 19 : ControlFile->minRecoveryPointTLI = lastCheckPoint.ThisTimeLineID;
8280 :
8281 : /* update local copy */
8282 19 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
8283 19 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
8284 : }
8285 212 : if (flags & CHECKPOINT_IS_SHUTDOWN)
8286 26 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
8287 : }
8288 :
8289 : /* we shall start with the latest checksum version */
8290 212 : ControlFile->data_checksum_version = lastCheckPoint.dataChecksumState;
8291 :
8292 212 : UpdateControlFile();
8293 : }
8294 212 : LWLockRelease(ControlFileLock);
8295 :
8296 : /*
8297 : * Update the average distance between checkpoints/restartpoints if the
8298 : * prior checkpoint exists.
8299 : */
8300 212 : if (XLogRecPtrIsValid(PriorRedoPtr))
8301 212 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
8302 :
8303 : /*
8304 : * Delete old log files, those no longer needed for last restartpoint to
8305 : * prevent the disk holding the xlog from growing full.
8306 : */
8307 212 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
8308 :
8309 : /*
8310 : * Retreat _logSegNo using the current end of xlog replayed or received,
8311 : * whichever is later.
8312 : */
8313 212 : receivePtr = GetWalRcvFlushRecPtr(NULL, NULL);
8314 212 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
8315 212 : endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
8316 212 : KeepLogSeg(endptr, &_logSegNo);
8317 :
8318 212 : INJECTION_POINT("restartpoint-before-slot-invalidation", NULL);
8319 :
8320 212 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
8321 : _logSegNo, InvalidOid,
8322 : InvalidTransactionId))
8323 : {
8324 : /*
8325 : * Some slots have been invalidated; recalculate the old-segment
8326 : * horizon, starting again from RedoRecPtr.
8327 : */
8328 1 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
8329 1 : KeepLogSeg(endptr, &_logSegNo);
8330 : }
8331 212 : _logSegNo--;
8332 :
8333 : /*
8334 : * Try to recycle segments on a useful timeline. If we've been promoted
8335 : * since the beginning of this restartpoint, use the new timeline chosen
8336 : * at end of recovery. If we're still in recovery, use the timeline we're
8337 : * currently replaying.
8338 : *
8339 : * There is no guarantee that the WAL segments will be useful on the
8340 : * current timeline; if recovery proceeds to a new timeline right after
8341 : * this, the pre-allocated WAL segments on this timeline will not be used,
8342 : * and will go wasted until recycled on the next restartpoint. We'll live
8343 : * with that.
8344 : */
8345 212 : if (!RecoveryInProgress())
8346 0 : replayTLI = XLogCtl->InsertTimeLineID;
8347 :
8348 212 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, endptr, replayTLI);
8349 :
8350 : /*
8351 : * Make more log segments if needed. (Do this after recycling old log
8352 : * segments, since that may supply some of the needed files.)
8353 : */
8354 212 : PreallocXlogFiles(endptr, replayTLI);
8355 :
8356 : /*
8357 : * Truncate pg_subtrans if possible. We can throw away all data before
8358 : * the oldest XMIN of any running transaction. No future transaction will
8359 : * attempt to reference any pg_subtrans entry older than that (see Asserts
8360 : * in subtrans.c). When hot standby is disabled, though, we mustn't do
8361 : * this because StartupSUBTRANS hasn't been called yet.
8362 : */
8363 212 : if (EnableHotStandby)
8364 212 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
8365 :
8366 : /* Real work is done; log and update stats. */
8367 212 : LogCheckpointEnd(true, flags);
8368 :
8369 : /* Reset the process title */
8370 212 : update_checkpoint_display(flags, true, true);
8371 :
8372 212 : xtime = GetLatestXTime();
8373 212 : ereport((log_checkpoints ? LOG : DEBUG2),
8374 : errmsg("recovery restart point at %X/%08X",
8375 : LSN_FORMAT_ARGS(lastCheckPoint.redo)),
8376 : xtime ? errdetail("Last completed transaction was at log time %s.",
8377 : timestamptz_to_str(xtime)) : 0);
8378 :
8379 : /*
8380 : * Finally, execute archive_cleanup_command, if any.
8381 : */
8382 212 : if (archiveCleanupCommand && strcmp(archiveCleanupCommand, "") != 0)
8383 0 : ExecuteRecoveryCommand(archiveCleanupCommand,
8384 : "archive_cleanup_command",
8385 : false,
8386 : WAIT_EVENT_ARCHIVE_CLEANUP_COMMAND);
8387 :
8388 212 : return true;
8389 : }
8390 :
8391 : /*
8392 : * Report availability of WAL for the given target LSN
8393 : * (typically a slot's restart_lsn)
8394 : *
8395 : * Returns one of the following enum values:
8396 : *
8397 : * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
8398 : * max_wal_size.
8399 : *
8400 : * * WALAVAIL_EXTENDED means it is still available by preserving extra
8401 : * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
8402 : * than max_wal_size, this state is not returned.
8403 : *
8404 : * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
8405 : * remove reserved segments. The walsender using this slot may return to the
8406 : * above.
8407 : *
8408 : * * WALAVAIL_REMOVED means it has been removed. A replication stream on
8409 : * a slot with this LSN cannot continue. (Any associated walsender
8410 : * processes should have been terminated already.)
8411 : *
8412 : * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
8413 : */
8414 : WALAvailability
8415 649 : GetWALAvailability(XLogRecPtr targetLSN)
8416 : {
8417 : XLogRecPtr currpos; /* current write LSN */
8418 : XLogSegNo currSeg; /* segid of currpos */
8419 : XLogSegNo targetSeg; /* segid of targetLSN */
8420 : XLogSegNo oldestSeg; /* actual oldest segid */
8421 : XLogSegNo oldestSegMaxWalSize; /* oldest segid kept by max_wal_size */
8422 : XLogSegNo oldestSlotSeg; /* oldest segid kept by slot */
8423 : uint64 keepSegs;
8424 :
8425 : /*
8426 : * slot does not reserve WAL. Either deactivated, or has never been active
8427 : */
8428 649 : if (!XLogRecPtrIsValid(targetLSN))
8429 30 : return WALAVAIL_INVALID_LSN;
8430 :
8431 : /*
8432 : * Calculate the oldest segment currently reserved by all slots,
8433 : * considering wal_keep_size and max_slot_wal_keep_size. Initialize
8434 : * oldestSlotSeg to the current segment.
8435 : */
8436 619 : currpos = GetXLogWriteRecPtr();
8437 619 : XLByteToSeg(currpos, oldestSlotSeg, wal_segment_size);
8438 619 : KeepLogSeg(currpos, &oldestSlotSeg);
8439 :
8440 : /*
8441 : * Find the oldest extant segment file. We get 1 until checkpoint removes
8442 : * the first WAL segment file since startup, which causes the status being
8443 : * wrong under certain abnormal conditions but that doesn't actually harm.
8444 : */
8445 619 : oldestSeg = XLogGetLastRemovedSegno() + 1;
8446 :
8447 : /* calculate oldest segment by max_wal_size */
8448 619 : XLByteToSeg(currpos, currSeg, wal_segment_size);
8449 619 : keepSegs = ConvertToXSegs(max_wal_size_mb, wal_segment_size) + 1;
8450 :
8451 619 : if (currSeg > keepSegs)
8452 8 : oldestSegMaxWalSize = currSeg - keepSegs;
8453 : else
8454 611 : oldestSegMaxWalSize = 1;
8455 :
8456 : /* the segment we care about */
8457 619 : XLByteToSeg(targetLSN, targetSeg, wal_segment_size);
8458 :
8459 : /*
8460 : * No point in returning reserved or extended status values if the
8461 : * targetSeg is known to be lost.
8462 : */
8463 619 : if (targetSeg >= oldestSlotSeg)
8464 : {
8465 : /* show "reserved" when targetSeg is within max_wal_size */
8466 618 : if (targetSeg >= oldestSegMaxWalSize)
8467 616 : return WALAVAIL_RESERVED;
8468 :
8469 : /* being retained by slots exceeding max_wal_size */
8470 2 : return WALAVAIL_EXTENDED;
8471 : }
8472 :
8473 : /* WAL segments are no longer retained but haven't been removed yet */
8474 1 : if (targetSeg >= oldestSeg)
8475 1 : return WALAVAIL_UNRESERVED;
8476 :
8477 : /* Definitely lost */
8478 0 : return WALAVAIL_REMOVED;
8479 : }
8480 :
8481 :
8482 : /*
8483 : * Retreat *logSegNo to the last segment that we need to retain because of
8484 : * either wal_keep_size or replication slots.
8485 : *
8486 : * This is calculated by subtracting wal_keep_size from the given xlog
8487 : * location, recptr and by making sure that that result is below the
8488 : * requirement of replication slots. For the latter criterion we do consider
8489 : * the effects of max_slot_wal_keep_size: reserve at most that much space back
8490 : * from recptr.
8491 : *
8492 : * Note about replication slots: if this function calculates a value
8493 : * that's further ahead than what slots need reserved, then affected
8494 : * slots need to be invalidated and this function invoked again.
8495 : * XXX it might be a good idea to rewrite this function so that
8496 : * invalidation is optionally done here, instead.
8497 : */
8498 : static void
8499 2560 : KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
8500 : {
8501 : XLogSegNo currSegNo;
8502 : XLogSegNo segno;
8503 : XLogRecPtr keep;
8504 :
8505 2560 : XLByteToSeg(recptr, currSegNo, wal_segment_size);
8506 2560 : segno = currSegNo;
8507 :
8508 : /* Calculate how many segments are kept by slots. */
8509 2560 : keep = XLogGetReplicationSlotMinimumLSN();
8510 2560 : if (XLogRecPtrIsValid(keep) && keep < recptr)
8511 : {
8512 784 : XLByteToSeg(keep, segno, wal_segment_size);
8513 :
8514 : /*
8515 : * Account for max_slot_wal_keep_size to avoid keeping more than
8516 : * configured. However, don't do that during a binary upgrade: if
8517 : * slots were to be invalidated because of this, it would not be
8518 : * possible to preserve logical ones during the upgrade.
8519 : */
8520 784 : if (max_slot_wal_keep_size_mb >= 0 && !IsBinaryUpgrade)
8521 : {
8522 : uint64 slot_keep_segs;
8523 :
8524 23 : slot_keep_segs =
8525 23 : ConvertToXSegs(max_slot_wal_keep_size_mb, wal_segment_size);
8526 :
8527 23 : if (currSegNo - segno > slot_keep_segs)
8528 6 : segno = currSegNo - slot_keep_segs;
8529 : }
8530 : }
8531 :
8532 : /*
8533 : * If WAL summarization is in use, don't remove WAL that has yet to be
8534 : * summarized.
8535 : */
8536 2560 : keep = GetOldestUnsummarizedLSN(NULL, NULL);
8537 2560 : if (XLogRecPtrIsValid(keep))
8538 : {
8539 : XLogSegNo unsummarized_segno;
8540 :
8541 7 : XLByteToSeg(keep, unsummarized_segno, wal_segment_size);
8542 7 : if (unsummarized_segno < segno)
8543 7 : segno = unsummarized_segno;
8544 : }
8545 :
8546 : /* but, keep at least wal_keep_size if that's set */
8547 2560 : if (wal_keep_size_mb > 0)
8548 : {
8549 : uint64 keep_segs;
8550 :
8551 74 : keep_segs = ConvertToXSegs(wal_keep_size_mb, wal_segment_size);
8552 74 : if (currSegNo - segno < keep_segs)
8553 : {
8554 : /* avoid underflow, don't go below 1 */
8555 74 : if (currSegNo <= keep_segs)
8556 70 : segno = 1;
8557 : else
8558 4 : segno = currSegNo - keep_segs;
8559 : }
8560 : }
8561 :
8562 : /* don't delete WAL segments newer than the calculated segment */
8563 2560 : if (segno < *logSegNo)
8564 370 : *logSegNo = segno;
8565 2560 : }
8566 :
8567 : /*
8568 : * Write a NEXTOID log record
8569 : */
8570 : void
8571 693 : XLogPutNextOid(Oid nextOid)
8572 : {
8573 693 : XLogBeginInsert();
8574 693 : XLogRegisterData(&nextOid, sizeof(Oid));
8575 693 : (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID);
8576 :
8577 : /*
8578 : * We need not flush the NEXTOID record immediately, because any of the
8579 : * just-allocated OIDs could only reach disk as part of a tuple insert or
8580 : * update that would have its own XLOG record that must follow the NEXTOID
8581 : * record. Therefore, the standard buffer LSN interlock applied to those
8582 : * records will ensure no such OID reaches disk before the NEXTOID record
8583 : * does.
8584 : *
8585 : * Note, however, that the above statement only covers state "within" the
8586 : * database. When we use a generated OID as a file or directory name, we
8587 : * are in a sense violating the basic WAL rule, because that filesystem
8588 : * change may reach disk before the NEXTOID WAL record does. The impact
8589 : * of this is that if a database crash occurs immediately afterward, we
8590 : * might after restart re-generate the same OID and find that it conflicts
8591 : * with the leftover file or directory. But since for safety's sake we
8592 : * always loop until finding a nonconflicting filename, this poses no real
8593 : * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8594 : */
8595 693 : }
8596 :
8597 : /*
8598 : * Write an XLOG SWITCH record.
8599 : *
8600 : * Here we just blindly issue an XLogInsert request for the record.
8601 : * All the magic happens inside XLogInsert.
8602 : *
8603 : * The return value is either the end+1 address of the switch record,
8604 : * or the end+1 address of the prior segment if we did not need to
8605 : * write a switch record because we are already at segment start.
8606 : */
8607 : XLogRecPtr
8608 815 : RequestXLogSwitch(bool mark_unimportant)
8609 : {
8610 : XLogRecPtr RecPtr;
8611 :
8612 : /* XLOG SWITCH has no data */
8613 815 : XLogBeginInsert();
8614 :
8615 815 : if (mark_unimportant)
8616 0 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8617 815 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH);
8618 :
8619 815 : return RecPtr;
8620 : }
8621 :
8622 : /*
8623 : * Write a RESTORE POINT record
8624 : */
8625 : XLogRecPtr
8626 3 : XLogRestorePoint(const char *rpName)
8627 : {
8628 : XLogRecPtr RecPtr;
8629 : xl_restore_point xlrec;
8630 :
8631 3 : xlrec.rp_time = GetCurrentTimestamp();
8632 3 : strlcpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8633 :
8634 3 : XLogBeginInsert();
8635 3 : XLogRegisterData(&xlrec, sizeof(xl_restore_point));
8636 :
8637 3 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT);
8638 :
8639 3 : ereport(LOG,
8640 : errmsg("restore point \"%s\" created at %X/%08X",
8641 : rpName, LSN_FORMAT_ARGS(RecPtr)));
8642 :
8643 3 : return RecPtr;
8644 : }
8645 :
8646 : /*
8647 : * Write an empty XLOG record to assign a distinct LSN.
8648 : *
8649 : * This is used by some index AMs when building indexes on permanent relations
8650 : * with wal_level=minimal. In that scenario, WAL-logging will start after
8651 : * commit, but the index AM needs distinct LSNs to detect concurrent page
8652 : * modifications. When the current WAL insert position hasn't advanced since
8653 : * the last call, we emit a dummy record to ensure we get a new, distinct LSN.
8654 : */
8655 : XLogRecPtr
8656 455 : XLogAssignLSN(void)
8657 : {
8658 455 : int dummy = 0;
8659 :
8660 : /*
8661 : * Records other than XLOG_SWITCH must have content. We use an integer 0
8662 : * to satisfy this restriction.
8663 : */
8664 455 : XLogBeginInsert();
8665 455 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8666 455 : XLogRegisterData(&dummy, sizeof(dummy));
8667 455 : return XLogInsert(RM_XLOG_ID, XLOG_ASSIGN_LSN);
8668 : }
8669 :
8670 : /*
8671 : * Check if any of the GUC parameters that are critical for hot standby
8672 : * have changed, and update the value in pg_control file if necessary.
8673 : */
8674 : static void
8675 1012 : XLogReportParameters(void)
8676 : {
8677 1012 : if (wal_level != ControlFile->wal_level ||
8678 737 : wal_log_hints != ControlFile->wal_log_hints ||
8679 648 : MaxConnections != ControlFile->MaxConnections ||
8680 647 : max_worker_processes != ControlFile->max_worker_processes ||
8681 644 : max_wal_senders != ControlFile->max_wal_senders ||
8682 616 : max_prepared_xacts != ControlFile->max_prepared_xacts ||
8683 512 : max_locks_per_xact != ControlFile->max_locks_per_xact ||
8684 512 : track_commit_timestamp != ControlFile->track_commit_timestamp)
8685 : {
8686 : /*
8687 : * The change in number of backend slots doesn't need to be WAL-logged
8688 : * if archiving is not enabled, as you can't start archive recovery
8689 : * with wal_level=minimal anyway. We don't really care about the
8690 : * values in pg_control either if wal_level=minimal, but seems better
8691 : * to keep them up-to-date to avoid confusion.
8692 : */
8693 512 : if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8694 : {
8695 : xl_parameter_change xlrec;
8696 : XLogRecPtr recptr;
8697 :
8698 485 : xlrec.MaxConnections = MaxConnections;
8699 485 : xlrec.max_worker_processes = max_worker_processes;
8700 485 : xlrec.max_wal_senders = max_wal_senders;
8701 485 : xlrec.max_prepared_xacts = max_prepared_xacts;
8702 485 : xlrec.max_locks_per_xact = max_locks_per_xact;
8703 485 : xlrec.wal_level = wal_level;
8704 485 : xlrec.wal_log_hints = wal_log_hints;
8705 485 : xlrec.track_commit_timestamp = track_commit_timestamp;
8706 :
8707 485 : XLogBeginInsert();
8708 485 : XLogRegisterData(&xlrec, sizeof(xlrec));
8709 :
8710 485 : recptr = XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE);
8711 485 : XLogFlush(recptr);
8712 : }
8713 :
8714 512 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8715 :
8716 512 : ControlFile->MaxConnections = MaxConnections;
8717 512 : ControlFile->max_worker_processes = max_worker_processes;
8718 512 : ControlFile->max_wal_senders = max_wal_senders;
8719 512 : ControlFile->max_prepared_xacts = max_prepared_xacts;
8720 512 : ControlFile->max_locks_per_xact = max_locks_per_xact;
8721 512 : ControlFile->wal_level = wal_level;
8722 512 : ControlFile->wal_log_hints = wal_log_hints;
8723 512 : ControlFile->track_commit_timestamp = track_commit_timestamp;
8724 512 : UpdateControlFile();
8725 :
8726 512 : LWLockRelease(ControlFileLock);
8727 : }
8728 1012 : }
8729 :
8730 : /*
8731 : * Log the new state of checksums
8732 : */
8733 : static void
8734 26 : XLogChecksums(uint32 new_type)
8735 : {
8736 : xl_checksum_state xlrec;
8737 : XLogRecPtr recptr;
8738 :
8739 26 : xlrec.new_checksum_state = new_type;
8740 :
8741 26 : XLogBeginInsert();
8742 26 : XLogRegisterData((char *) &xlrec, sizeof(xl_checksum_state));
8743 :
8744 26 : recptr = XLogInsert(RM_XLOG2_ID, XLOG2_CHECKSUMS);
8745 26 : XLogFlush(recptr);
8746 26 : }
8747 :
8748 : /*
8749 : * Update full_page_writes in shared memory, and write an
8750 : * XLOG_FPW_CHANGE record if necessary.
8751 : *
8752 : * Note: this function assumes there is no other process running
8753 : * concurrently that could update it.
8754 : */
8755 : void
8756 1726 : UpdateFullPageWrites(void)
8757 : {
8758 1726 : XLogCtlInsert *Insert = &XLogCtl->Insert;
8759 : bool recoveryInProgress;
8760 :
8761 : /*
8762 : * Do nothing if full_page_writes has not been changed.
8763 : *
8764 : * It's safe to check the shared full_page_writes without the lock,
8765 : * because we assume that there is no concurrently running process which
8766 : * can update it.
8767 : */
8768 1726 : if (fullPageWrites == Insert->fullPageWrites)
8769 1255 : return;
8770 :
8771 : /*
8772 : * Perform this outside critical section so that the WAL insert
8773 : * initialization done by RecoveryInProgress() doesn't trigger an
8774 : * assertion failure.
8775 : */
8776 471 : recoveryInProgress = RecoveryInProgress();
8777 :
8778 471 : START_CRIT_SECTION();
8779 :
8780 : /*
8781 : * It's always safe to take full page images, even when not strictly
8782 : * required, but not the other round. So if we're setting full_page_writes
8783 : * to true, first set it true and then write the WAL record. If we're
8784 : * setting it to false, first write the WAL record and then set the global
8785 : * flag.
8786 : */
8787 471 : if (fullPageWrites)
8788 : {
8789 458 : WALInsertLockAcquireExclusive();
8790 458 : Insert->fullPageWrites = true;
8791 458 : WALInsertLockRelease();
8792 : }
8793 :
8794 : /*
8795 : * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8796 : * full_page_writes during archive recovery, if required.
8797 : */
8798 471 : if (XLogStandbyInfoActive() && !recoveryInProgress)
8799 : {
8800 0 : XLogBeginInsert();
8801 0 : XLogRegisterData(&fullPageWrites, sizeof(bool));
8802 :
8803 0 : XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
8804 : }
8805 :
8806 471 : if (!fullPageWrites)
8807 : {
8808 13 : WALInsertLockAcquireExclusive();
8809 13 : Insert->fullPageWrites = false;
8810 13 : WALInsertLockRelease();
8811 : }
8812 471 : END_CRIT_SECTION();
8813 : }
8814 :
8815 : /*
8816 : * XLOG resource manager's routines
8817 : *
8818 : * Definitions of info values are in include/catalog/pg_control.h, though
8819 : * not all record types are related to control file updates.
8820 : *
8821 : * NOTE: Some XLOG record types that are directly related to WAL recovery
8822 : * are handled in xlogrecovery_redo().
8823 : */
8824 : void
8825 111100 : xlog_redo(XLogReaderState *record)
8826 : {
8827 111100 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8828 111100 : XLogRecPtr lsn = record->EndRecPtr;
8829 :
8830 : /*
8831 : * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
8832 : * XLOG_FPI_FOR_HINT records.
8833 : */
8834 : Assert(info == XLOG_FPI || info == XLOG_FPI_FOR_HINT ||
8835 : !XLogRecHasAnyBlockRefs(record));
8836 :
8837 111100 : if (info == XLOG_NEXTOID)
8838 : {
8839 : Oid nextOid;
8840 :
8841 : /*
8842 : * We used to try to take the maximum of TransamVariables->nextOid and
8843 : * the recorded nextOid, but that fails if the OID counter wraps
8844 : * around. Since no OID allocation should be happening during replay
8845 : * anyway, better to just believe the record exactly. We still take
8846 : * OidGenLock while setting the variable, just in case.
8847 : */
8848 96 : memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
8849 96 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8850 96 : TransamVariables->nextOid = nextOid;
8851 96 : TransamVariables->oidCount = 0;
8852 96 : LWLockRelease(OidGenLock);
8853 : }
8854 111004 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
8855 : {
8856 : CheckPoint checkPoint;
8857 : TimeLineID replayTLI;
8858 :
8859 41 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8860 : /* In a SHUTDOWN checkpoint, believe the counters exactly */
8861 41 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8862 41 : TransamVariables->nextXid = checkPoint.nextXid;
8863 41 : LWLockRelease(XidGenLock);
8864 41 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8865 41 : TransamVariables->nextOid = checkPoint.nextOid;
8866 41 : TransamVariables->oidCount = 0;
8867 41 : LWLockRelease(OidGenLock);
8868 41 : MultiXactSetNextMXact(checkPoint.nextMulti,
8869 : checkPoint.nextMultiOffset);
8870 :
8871 41 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8872 : checkPoint.oldestMultiDB);
8873 :
8874 : /*
8875 : * No need to set oldestClogXid here as well; it'll be set when we
8876 : * redo an xl_clog_truncate if it changed since initialization.
8877 : */
8878 41 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
8879 :
8880 : /*
8881 : * If we see a shutdown checkpoint while waiting for an end-of-backup
8882 : * record, the backup was canceled and the end-of-backup record will
8883 : * never arrive.
8884 : */
8885 41 : if (ArchiveRecoveryRequested &&
8886 41 : XLogRecPtrIsValid(ControlFile->backupStartPoint) &&
8887 0 : !XLogRecPtrIsValid(ControlFile->backupEndPoint))
8888 0 : ereport(PANIC,
8889 : (errmsg("online backup was canceled, recovery cannot continue")));
8890 :
8891 : /*
8892 : * If we see a shutdown checkpoint, we know that nothing was running
8893 : * on the primary at this point. So fake-up an empty running-xacts
8894 : * record and use that here and now. Recover additional standby state
8895 : * for prepared transactions.
8896 : */
8897 41 : if (standbyState >= STANDBY_INITIALIZED)
8898 : {
8899 : TransactionId *xids;
8900 : int nxids;
8901 : TransactionId oldestActiveXID;
8902 : TransactionId latestCompletedXid;
8903 : RunningTransactionsData running;
8904 :
8905 39 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
8906 :
8907 : /* Update pg_subtrans entries for any prepared transactions */
8908 39 : StandbyRecoverPreparedTransactions();
8909 :
8910 : /*
8911 : * Construct a RunningTransactions snapshot representing a shut
8912 : * down server, with only prepared transactions still alive. We're
8913 : * never overflowed at this point because all subxids are listed
8914 : * with their parent prepared transactions.
8915 : */
8916 39 : running.xcnt = nxids;
8917 39 : running.subxcnt = 0;
8918 39 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
8919 39 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
8920 39 : running.oldestRunningXid = oldestActiveXID;
8921 39 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
8922 39 : TransactionIdRetreat(latestCompletedXid);
8923 : Assert(TransactionIdIsNormal(latestCompletedXid));
8924 39 : running.latestCompletedXid = latestCompletedXid;
8925 39 : running.xids = xids;
8926 :
8927 39 : ProcArrayApplyRecoveryInfo(&running);
8928 : }
8929 :
8930 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8931 41 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8932 41 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8933 41 : ControlFile->data_checksum_version = checkPoint.dataChecksumState;
8934 :
8935 41 : UpdateControlFile();
8936 41 : LWLockRelease(ControlFileLock);
8937 :
8938 : /*
8939 : * We should've already switched to the new TLI before replaying this
8940 : * record.
8941 : */
8942 41 : (void) GetCurrentReplayRecPtr(&replayTLI);
8943 41 : if (checkPoint.ThisTimeLineID != replayTLI)
8944 0 : ereport(PANIC,
8945 : (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
8946 : checkPoint.ThisTimeLineID, replayTLI)));
8947 :
8948 41 : RecoveryRestartPoint(&checkPoint, record);
8949 :
8950 : /*
8951 : * After replaying a checkpoint record, free all smgr objects.
8952 : * Otherwise we would never do so for dropped relations, as the
8953 : * startup does not process shared invalidation messages or call
8954 : * AtEOXact_SMgr().
8955 : */
8956 41 : smgrdestroyall();
8957 : }
8958 110963 : else if (info == XLOG_CHECKPOINT_ONLINE)
8959 : {
8960 : CheckPoint checkPoint;
8961 : TimeLineID replayTLI;
8962 :
8963 700 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8964 : /* In an ONLINE checkpoint, treat the XID counter as a minimum */
8965 700 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8966 700 : if (FullTransactionIdPrecedes(TransamVariables->nextXid,
8967 : checkPoint.nextXid))
8968 0 : TransamVariables->nextXid = checkPoint.nextXid;
8969 700 : LWLockRelease(XidGenLock);
8970 :
8971 : /*
8972 : * We ignore the nextOid counter in an ONLINE checkpoint, preferring
8973 : * to track OID assignment through XLOG_NEXTOID records. The nextOid
8974 : * counter is from the start of the checkpoint and might well be stale
8975 : * compared to later XLOG_NEXTOID records. We could try to take the
8976 : * maximum of the nextOid counter and our latest value, but since
8977 : * there's no particular guarantee about the speed with which the OID
8978 : * counter wraps around, that's a risky thing to do. In any case,
8979 : * users of the nextOid counter are required to avoid assignment of
8980 : * duplicates, so that a somewhat out-of-date value should be safe.
8981 : */
8982 :
8983 : /* Handle multixact */
8984 700 : MultiXactAdvanceNextMXact(checkPoint.nextMulti,
8985 : checkPoint.nextMultiOffset);
8986 :
8987 : /*
8988 : * NB: This may perform multixact truncation when replaying WAL
8989 : * generated by an older primary.
8990 : */
8991 700 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8992 : checkPoint.oldestMultiDB);
8993 700 : if (TransactionIdPrecedes(TransamVariables->oldestXid,
8994 : checkPoint.oldestXid))
8995 0 : SetTransactionIdLimit(checkPoint.oldestXid,
8996 : checkPoint.oldestXidDB);
8997 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8998 700 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8999 700 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
9000 700 : LWLockRelease(ControlFileLock);
9001 :
9002 : /* TLI should not change in an on-line checkpoint */
9003 700 : (void) GetCurrentReplayRecPtr(&replayTLI);
9004 700 : if (checkPoint.ThisTimeLineID != replayTLI)
9005 0 : ereport(PANIC,
9006 : (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
9007 : checkPoint.ThisTimeLineID, replayTLI)));
9008 :
9009 700 : RecoveryRestartPoint(&checkPoint, record);
9010 :
9011 : /*
9012 : * After replaying a checkpoint record, free all smgr objects.
9013 : * Otherwise we would never do so for dropped relations, as the
9014 : * startup does not process shared invalidation messages or call
9015 : * AtEOXact_SMgr().
9016 : */
9017 700 : smgrdestroyall();
9018 : }
9019 110263 : else if (info == XLOG_OVERWRITE_CONTRECORD)
9020 : {
9021 : /* nothing to do here, handled in xlogrecovery_redo() */
9022 : }
9023 110262 : else if (info == XLOG_END_OF_RECOVERY)
9024 : {
9025 : xl_end_of_recovery xlrec;
9026 : TimeLineID replayTLI;
9027 :
9028 10 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
9029 :
9030 : /*
9031 : * For Hot Standby, we could treat this like a Shutdown Checkpoint,
9032 : * but this case is rarer and harder to test, so the benefit doesn't
9033 : * outweigh the potential extra cost of maintenance.
9034 : */
9035 :
9036 : /*
9037 : * We should've already switched to the new TLI before replaying this
9038 : * record.
9039 : */
9040 10 : (void) GetCurrentReplayRecPtr(&replayTLI);
9041 10 : if (xlrec.ThisTimeLineID != replayTLI)
9042 0 : ereport(PANIC,
9043 : (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
9044 : xlrec.ThisTimeLineID, replayTLI)));
9045 : }
9046 110252 : else if (info == XLOG_NOOP)
9047 : {
9048 : /* nothing to do here */
9049 : }
9050 110252 : else if (info == XLOG_SWITCH)
9051 : {
9052 : /* nothing to do here */
9053 : }
9054 109797 : else if (info == XLOG_RESTORE_POINT)
9055 : {
9056 : /* nothing to do here, handled in xlogrecovery.c */
9057 : }
9058 109792 : else if (info == XLOG_ASSIGN_LSN)
9059 : {
9060 : /* nothing to do here, see XLogGetFakeLSN() */
9061 : }
9062 47465 : else if (info == XLOG_FPI || info == XLOG_FPI_FOR_HINT)
9063 : {
9064 : /*
9065 : * XLOG_FPI records contain nothing else but one or more block
9066 : * references. Every block reference must include a full-page image
9067 : * even if full_page_writes was disabled when the record was generated
9068 : * - otherwise there would be no point in this record.
9069 : *
9070 : * XLOG_FPI_FOR_HINT records are generated when a page needs to be
9071 : * WAL-logged because of a hint bit update. They are only generated
9072 : * when checksums and/or wal_log_hints are enabled. They may include
9073 : * no full-page images if full_page_writes was disabled when they were
9074 : * generated. In this case there is nothing to do here.
9075 : *
9076 : * No recovery conflicts are generated by these generic records - if a
9077 : * resource manager needs to generate conflicts, it has to define a
9078 : * separate WAL record type and redo routine.
9079 : */
9080 98098 : for (uint8 block_id = 0; block_id <= XLogRecMaxBlockId(record); block_id++)
9081 : {
9082 : Buffer buffer;
9083 :
9084 51487 : if (!XLogRecHasBlockImage(record, block_id))
9085 : {
9086 66 : if (info == XLOG_FPI)
9087 0 : elog(ERROR, "XLOG_FPI record did not contain a full-page image");
9088 66 : continue;
9089 : }
9090 :
9091 51421 : if (XLogReadBufferForRedo(record, block_id, &buffer) != BLK_RESTORED)
9092 0 : elog(ERROR, "unexpected XLogReadBufferForRedo result when restoring backup block");
9093 51421 : UnlockReleaseBuffer(buffer);
9094 : }
9095 : }
9096 854 : else if (info == XLOG_BACKUP_END)
9097 : {
9098 : /* nothing to do here, handled in xlogrecovery_redo() */
9099 : }
9100 757 : else if (info == XLOG_PARAMETER_CHANGE)
9101 : {
9102 : xl_parameter_change xlrec;
9103 :
9104 : /* Update our copy of the parameters in pg_control */
9105 38 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
9106 :
9107 38 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9108 38 : ControlFile->MaxConnections = xlrec.MaxConnections;
9109 38 : ControlFile->max_worker_processes = xlrec.max_worker_processes;
9110 38 : ControlFile->max_wal_senders = xlrec.max_wal_senders;
9111 38 : ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
9112 38 : ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
9113 38 : ControlFile->wal_level = xlrec.wal_level;
9114 38 : ControlFile->wal_log_hints = xlrec.wal_log_hints;
9115 :
9116 : /*
9117 : * Update minRecoveryPoint to ensure that if recovery is aborted, we
9118 : * recover back up to this point before allowing hot standby again.
9119 : * This is important if the max_* settings are decreased, to ensure
9120 : * you don't run queries against the WAL preceding the change. The
9121 : * local copies cannot be updated as long as crash recovery is
9122 : * happening and we expect all the WAL to be replayed.
9123 : */
9124 38 : if (InArchiveRecovery)
9125 : {
9126 23 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
9127 23 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
9128 : }
9129 38 : if (XLogRecPtrIsValid(LocalMinRecoveryPoint) && LocalMinRecoveryPoint < lsn)
9130 : {
9131 : TimeLineID replayTLI;
9132 :
9133 12 : (void) GetCurrentReplayRecPtr(&replayTLI);
9134 12 : ControlFile->minRecoveryPoint = lsn;
9135 12 : ControlFile->minRecoveryPointTLI = replayTLI;
9136 : }
9137 :
9138 38 : CommitTsParameterChange(xlrec.track_commit_timestamp,
9139 38 : ControlFile->track_commit_timestamp);
9140 38 : ControlFile->track_commit_timestamp = xlrec.track_commit_timestamp;
9141 :
9142 38 : UpdateControlFile();
9143 38 : LWLockRelease(ControlFileLock);
9144 :
9145 : /* Check to see if any parameter change gives a problem on recovery */
9146 38 : CheckRequiredParameterValues();
9147 : }
9148 719 : else if (info == XLOG_FPW_CHANGE)
9149 : {
9150 : bool fpw;
9151 :
9152 0 : memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
9153 :
9154 : /*
9155 : * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
9156 : * do_pg_backup_start() and do_pg_backup_stop() can check whether
9157 : * full_page_writes has been disabled during online backup.
9158 : */
9159 0 : if (!fpw)
9160 : {
9161 0 : SpinLockAcquire(&XLogCtl->info_lck);
9162 0 : if (XLogCtl->lastFpwDisableRecPtr < record->ReadRecPtr)
9163 0 : XLogCtl->lastFpwDisableRecPtr = record->ReadRecPtr;
9164 0 : SpinLockRelease(&XLogCtl->info_lck);
9165 : }
9166 :
9167 : /* Keep track of full_page_writes */
9168 0 : lastFullPageWrites = fpw;
9169 : }
9170 719 : else if (info == XLOG_CHECKPOINT_REDO)
9171 : {
9172 : xl_checkpoint_redo redo_rec;
9173 701 : bool new_state = false;
9174 :
9175 701 : memcpy(&redo_rec, XLogRecGetData(record), sizeof(xl_checkpoint_redo));
9176 :
9177 701 : SpinLockAcquire(&XLogCtl->info_lck);
9178 701 : XLogCtl->data_checksum_version = redo_rec.data_checksum_version;
9179 701 : SetLocalDataChecksumState(redo_rec.data_checksum_version);
9180 701 : if (redo_rec.data_checksum_version != ControlFile->data_checksum_version)
9181 0 : new_state = true;
9182 701 : SpinLockRelease(&XLogCtl->info_lck);
9183 :
9184 701 : if (new_state)
9185 0 : EmitAndWaitDataChecksumsBarrier(redo_rec.data_checksum_version);
9186 : }
9187 18 : else if (info == XLOG_LOGICAL_DECODING_STATUS_CHANGE)
9188 : {
9189 : bool status;
9190 :
9191 18 : memcpy(&status, XLogRecGetData(record), sizeof(bool));
9192 :
9193 : /*
9194 : * We need to toggle the logical decoding status and update the
9195 : * XLogLogicalInfo cache of processes synchronously because
9196 : * XLogLogicalInfoActive() is used even during read-only queries
9197 : * (e.g., via RelationIsAccessibleInLogicalDecoding()). In the
9198 : * 'disable' case, it is safe to invalidate existing slots after
9199 : * disabling logical decoding because logical decoding cannot process
9200 : * subsequent WAL records, which may not contain logical information.
9201 : */
9202 18 : if (status)
9203 9 : EnableLogicalDecoding();
9204 : else
9205 9 : DisableLogicalDecoding();
9206 :
9207 18 : elog(DEBUG1, "update logical decoding status to %d during recovery",
9208 : status);
9209 :
9210 18 : if (InRecovery && InHotStandby)
9211 : {
9212 16 : if (!status)
9213 : {
9214 : /*
9215 : * Invalidate logical slots if we are in hot standby and the
9216 : * primary disabled logical decoding.
9217 : */
9218 9 : InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_LEVEL,
9219 : 0, InvalidOid,
9220 : InvalidTransactionId);
9221 : }
9222 7 : else if (sync_replication_slots)
9223 : {
9224 : /*
9225 : * Signal the postmaster to launch the slotsync worker.
9226 : *
9227 : * XXX: For simplicity, we keep the slotsync worker running
9228 : * even after logical decoding is disabled. A future
9229 : * improvement can consider starting and stopping the worker
9230 : * based on logical decoding status change.
9231 : */
9232 0 : kill(PostmasterPid, SIGUSR1);
9233 : }
9234 : }
9235 : }
9236 111098 : }
9237 :
9238 : void
9239 4 : xlog2_redo(XLogReaderState *record)
9240 : {
9241 4 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
9242 :
9243 4 : if (info == XLOG2_CHECKSUMS)
9244 : {
9245 : xl_checksum_state state;
9246 :
9247 4 : memcpy(&state, XLogRecGetData(record), sizeof(xl_checksum_state));
9248 :
9249 4 : SpinLockAcquire(&XLogCtl->info_lck);
9250 4 : XLogCtl->data_checksum_version = state.new_checksum_state;
9251 4 : SpinLockRelease(&XLogCtl->info_lck);
9252 :
9253 4 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9254 4 : ControlFile->data_checksum_version = state.new_checksum_state;
9255 4 : UpdateControlFile();
9256 4 : LWLockRelease(ControlFileLock);
9257 :
9258 : /*
9259 : * Block on a procsignalbarrier to await all processes having seen the
9260 : * change to checksum status. Once the barrier has been passed we can
9261 : * initiate the corresponding processing.
9262 : */
9263 4 : EmitAndWaitDataChecksumsBarrier(state.new_checksum_state);
9264 : }
9265 4 : }
9266 :
9267 : /*
9268 : * Return the extra open flags used for opening a file, depending on the
9269 : * value of the GUCs wal_sync_method, fsync and debug_io_direct.
9270 : */
9271 : static int
9272 17717 : get_sync_bit(int method)
9273 : {
9274 17717 : int o_direct_flag = 0;
9275 :
9276 : /*
9277 : * Use O_DIRECT if requested, except in walreceiver process. The WAL
9278 : * written by walreceiver is normally read by the startup process soon
9279 : * after it's written. Also, walreceiver performs unaligned writes, which
9280 : * don't work with O_DIRECT, so it is required for correctness too.
9281 : */
9282 17717 : if ((io_direct_flags & IO_DIRECT_WAL) && !AmWalReceiverProcess())
9283 9 : o_direct_flag = PG_O_DIRECT;
9284 :
9285 : /* If fsync is disabled, never open in sync mode */
9286 17717 : if (!enableFsync)
9287 17716 : return o_direct_flag;
9288 :
9289 1 : switch (method)
9290 : {
9291 : /*
9292 : * enum values for all sync options are defined even if they are
9293 : * not supported on the current platform. But if not, they are
9294 : * not included in the enum option array, and therefore will never
9295 : * be seen here.
9296 : */
9297 1 : case WAL_SYNC_METHOD_FSYNC:
9298 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
9299 : case WAL_SYNC_METHOD_FDATASYNC:
9300 1 : return o_direct_flag;
9301 : #ifdef O_SYNC
9302 0 : case WAL_SYNC_METHOD_OPEN:
9303 0 : return O_SYNC | o_direct_flag;
9304 : #endif
9305 : #ifdef O_DSYNC
9306 0 : case WAL_SYNC_METHOD_OPEN_DSYNC:
9307 0 : return O_DSYNC | o_direct_flag;
9308 : #endif
9309 0 : default:
9310 : /* can't happen (unless we are out of sync with option array) */
9311 0 : elog(ERROR, "unrecognized \"wal_sync_method\": %d", method);
9312 : return 0; /* silence warning */
9313 : }
9314 : }
9315 :
9316 : /*
9317 : * GUC support
9318 : */
9319 : void
9320 1280 : assign_wal_sync_method(int new_wal_sync_method, void *extra)
9321 : {
9322 1280 : if (wal_sync_method != new_wal_sync_method)
9323 : {
9324 : /*
9325 : * To ensure that no blocks escape unsynced, force an fsync on the
9326 : * currently open log segment (if any). Also, if the open flag is
9327 : * changing, close the log file so it will be reopened (with new flag
9328 : * bit) at next use.
9329 : */
9330 0 : if (openLogFile >= 0)
9331 : {
9332 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN);
9333 0 : if (pg_fsync(openLogFile) != 0)
9334 : {
9335 : char xlogfname[MAXFNAMELEN];
9336 : int save_errno;
9337 :
9338 0 : save_errno = errno;
9339 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo,
9340 : wal_segment_size);
9341 0 : errno = save_errno;
9342 0 : ereport(PANIC,
9343 : (errcode_for_file_access(),
9344 : errmsg("could not fsync file \"%s\": %m", xlogfname)));
9345 : }
9346 :
9347 0 : pgstat_report_wait_end();
9348 0 : if (get_sync_bit(wal_sync_method) != get_sync_bit(new_wal_sync_method))
9349 0 : XLogFileClose();
9350 : }
9351 : }
9352 1280 : }
9353 :
9354 :
9355 : /*
9356 : * Issue appropriate kind of fsync (if any) for an XLOG output file.
9357 : *
9358 : * 'fd' is a file descriptor for the XLOG file to be fsync'd.
9359 : * 'segno' is for error reporting purposes.
9360 : */
9361 : void
9362 210564 : issue_xlog_fsync(int fd, XLogSegNo segno, TimeLineID tli)
9363 : {
9364 210564 : char *msg = NULL;
9365 : instr_time start;
9366 :
9367 : Assert(tli != 0);
9368 :
9369 : /*
9370 : * Quick exit if fsync is disabled or write() has already synced the WAL
9371 : * file.
9372 : */
9373 210564 : if (!enableFsync ||
9374 1 : wal_sync_method == WAL_SYNC_METHOD_OPEN ||
9375 1 : wal_sync_method == WAL_SYNC_METHOD_OPEN_DSYNC)
9376 210563 : return;
9377 :
9378 : /*
9379 : * Measure I/O timing to sync the WAL file for pg_stat_io.
9380 : */
9381 1 : start = pgstat_prepare_io_time(track_wal_io_timing);
9382 :
9383 1 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC);
9384 1 : switch (wal_sync_method)
9385 : {
9386 0 : case WAL_SYNC_METHOD_FSYNC:
9387 0 : if (pg_fsync_no_writethrough(fd) != 0)
9388 0 : msg = _("could not fsync file \"%s\": %m");
9389 0 : break;
9390 : #ifdef HAVE_FSYNC_WRITETHROUGH
9391 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
9392 : if (pg_fsync_writethrough(fd) != 0)
9393 : msg = _("could not fsync write-through file \"%s\": %m");
9394 : break;
9395 : #endif
9396 1 : case WAL_SYNC_METHOD_FDATASYNC:
9397 1 : if (pg_fdatasync(fd) != 0)
9398 0 : msg = _("could not fdatasync file \"%s\": %m");
9399 1 : break;
9400 0 : case WAL_SYNC_METHOD_OPEN:
9401 : case WAL_SYNC_METHOD_OPEN_DSYNC:
9402 : /* not reachable */
9403 : Assert(false);
9404 0 : break;
9405 0 : default:
9406 0 : ereport(PANIC,
9407 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
9408 : errmsg_internal("unrecognized \"wal_sync_method\": %d", wal_sync_method));
9409 : break;
9410 : }
9411 :
9412 : /* PANIC if failed to fsync */
9413 1 : if (msg)
9414 : {
9415 : char xlogfname[MAXFNAMELEN];
9416 0 : int save_errno = errno;
9417 :
9418 0 : XLogFileName(xlogfname, tli, segno, wal_segment_size);
9419 0 : errno = save_errno;
9420 0 : ereport(PANIC,
9421 : (errcode_for_file_access(),
9422 : errmsg(msg, xlogfname)));
9423 : }
9424 :
9425 1 : pgstat_report_wait_end();
9426 :
9427 1 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL, IOOP_FSYNC,
9428 : start, 1, 0);
9429 : }
9430 :
9431 : /*
9432 : * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
9433 : * function. It creates the necessary starting checkpoint and constructs the
9434 : * backup state and tablespace map.
9435 : *
9436 : * Input parameters are "state" (the backup state), "fast" (if true, we do
9437 : * the checkpoint in fast mode), and "tablespaces" (if non-NULL, indicates a
9438 : * list of tablespaceinfo structs describing the cluster's tablespaces.).
9439 : *
9440 : * The tablespace map contents are appended to passed-in parameter
9441 : * tablespace_map and the caller is responsible for including it in the backup
9442 : * archive as 'tablespace_map'. The tablespace_map file is required mainly for
9443 : * tar format in windows as native windows utilities are not able to create
9444 : * symlinks while extracting files from tar. However for consistency and
9445 : * platform-independence, we do it the same way everywhere.
9446 : *
9447 : * It fills in "state" with the information required for the backup, such
9448 : * as the minimum WAL location that must be present to restore from this
9449 : * backup (starttli) and the corresponding timeline ID (starttli).
9450 : *
9451 : * Every successfully started backup must be stopped by calling
9452 : * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
9453 : * backups active at the same time.
9454 : *
9455 : * It is the responsibility of the caller of this function to verify the
9456 : * permissions of the calling user!
9457 : */
9458 : void
9459 178 : do_pg_backup_start(const char *backupidstr, bool fast, List **tablespaces,
9460 : BackupState *state, StringInfo tblspcmapfile)
9461 : {
9462 : bool backup_started_in_recovery;
9463 :
9464 : Assert(state != NULL);
9465 178 : backup_started_in_recovery = RecoveryInProgress();
9466 :
9467 : /*
9468 : * During recovery, we don't need to check WAL level. Because, if WAL
9469 : * level is not sufficient, it's impossible to get here during recovery.
9470 : */
9471 178 : if (!backup_started_in_recovery && !XLogIsNeeded())
9472 0 : ereport(ERROR,
9473 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9474 : errmsg("WAL level not sufficient for making an online backup"),
9475 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9476 :
9477 178 : if (strlen(backupidstr) > MAXPGPATH)
9478 1 : ereport(ERROR,
9479 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
9480 : errmsg("backup label too long (max %d bytes)",
9481 : MAXPGPATH)));
9482 :
9483 177 : strlcpy(state->name, backupidstr, sizeof(state->name));
9484 :
9485 : /*
9486 : * Mark backup active in shared memory. We must do full-page WAL writes
9487 : * during an on-line backup even if not doing so at other times, because
9488 : * it's quite possible for the backup dump to obtain a "torn" (partially
9489 : * written) copy of a database page if it reads the page concurrently with
9490 : * our write to the same page. This can be fixed as long as the first
9491 : * write to the page in the WAL sequence is a full-page write. Hence, we
9492 : * increment runningBackups then force a CHECKPOINT, to ensure there are
9493 : * no dirty pages in shared memory that might get dumped while the backup
9494 : * is in progress without having a corresponding WAL record. (Once the
9495 : * backup is complete, we need not force full-page writes anymore, since
9496 : * we expect that any pages not modified during the backup interval must
9497 : * have been correctly captured by the backup.)
9498 : *
9499 : * Note that forcing full-page writes has no effect during an online
9500 : * backup from the standby.
9501 : *
9502 : * We must hold all the insertion locks to change the value of
9503 : * runningBackups, to ensure adequate interlocking against
9504 : * XLogInsertRecord().
9505 : */
9506 177 : WALInsertLockAcquireExclusive();
9507 177 : XLogCtl->Insert.runningBackups++;
9508 177 : WALInsertLockRelease();
9509 :
9510 : /*
9511 : * Ensure we decrement runningBackups if we fail below. NB -- for this to
9512 : * work correctly, it is critical that sessionBackupState is only updated
9513 : * after this block is over.
9514 : */
9515 177 : PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9516 : {
9517 177 : bool gotUniqueStartpoint = false;
9518 : DIR *tblspcdir;
9519 : struct dirent *de;
9520 : tablespaceinfo *ti;
9521 : int datadirpathlen;
9522 :
9523 : /*
9524 : * Force an XLOG file switch before the checkpoint, to ensure that the
9525 : * WAL segment the checkpoint is written to doesn't contain pages with
9526 : * old timeline IDs. That would otherwise happen if you called
9527 : * pg_backup_start() right after restoring from a PITR archive: the
9528 : * first WAL segment containing the startup checkpoint has pages in
9529 : * the beginning with the old timeline ID. That can cause trouble at
9530 : * recovery: we won't have a history file covering the old timeline if
9531 : * pg_wal directory was not included in the base backup and the WAL
9532 : * archive was cleared too before starting the backup.
9533 : *
9534 : * During recovery, we skip forcing XLOG file switch, which means that
9535 : * the backup taken during recovery is not available for the special
9536 : * recovery case described above.
9537 : */
9538 177 : if (!backup_started_in_recovery)
9539 169 : RequestXLogSwitch(false);
9540 :
9541 : do
9542 : {
9543 : bool checkpointfpw;
9544 :
9545 : /*
9546 : * Force a CHECKPOINT. Aside from being necessary to prevent torn
9547 : * page problems, this guarantees that two successive backup runs
9548 : * will have different checkpoint positions and hence different
9549 : * history file names, even if nothing happened in between.
9550 : *
9551 : * During recovery, establish a restartpoint if possible. We use
9552 : * the last restartpoint as the backup starting checkpoint. This
9553 : * means that two successive backup runs can have same checkpoint
9554 : * positions.
9555 : *
9556 : * Since the fact that we are executing do_pg_backup_start()
9557 : * during recovery means that checkpointer is running, we can use
9558 : * RequestCheckpoint() to establish a restartpoint.
9559 : *
9560 : * We use CHECKPOINT_FAST only if requested by user (via passing
9561 : * fast = true). Otherwise this can take awhile.
9562 : */
9563 177 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
9564 : (fast ? CHECKPOINT_FAST : 0));
9565 :
9566 : /*
9567 : * Now we need to fetch the checkpoint record location, and also
9568 : * its REDO pointer. The oldest point in WAL that would be needed
9569 : * to restore starting from the checkpoint is precisely the REDO
9570 : * pointer.
9571 : */
9572 177 : LWLockAcquire(ControlFileLock, LW_SHARED);
9573 177 : state->checkpointloc = ControlFile->checkPoint;
9574 177 : state->startpoint = ControlFile->checkPointCopy.redo;
9575 177 : state->starttli = ControlFile->checkPointCopy.ThisTimeLineID;
9576 177 : checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
9577 177 : LWLockRelease(ControlFileLock);
9578 :
9579 177 : if (backup_started_in_recovery)
9580 : {
9581 : XLogRecPtr recptr;
9582 :
9583 : /*
9584 : * Check to see if all WAL replayed during online backup
9585 : * (i.e., since last restartpoint used as backup starting
9586 : * checkpoint) contain full-page writes.
9587 : */
9588 8 : SpinLockAcquire(&XLogCtl->info_lck);
9589 8 : recptr = XLogCtl->lastFpwDisableRecPtr;
9590 8 : SpinLockRelease(&XLogCtl->info_lck);
9591 :
9592 8 : if (!checkpointfpw || state->startpoint <= recptr)
9593 0 : ereport(ERROR,
9594 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9595 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9596 : "since last restartpoint"),
9597 : errhint("This means that the backup being taken on the standby "
9598 : "is corrupt and should not be used. "
9599 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9600 : "and then try an online backup again.")));
9601 :
9602 : /*
9603 : * During recovery, since we don't use the end-of-backup WAL
9604 : * record and don't write the backup history file, the
9605 : * starting WAL location doesn't need to be unique. This means
9606 : * that two base backups started at the same time might use
9607 : * the same checkpoint as starting locations.
9608 : */
9609 8 : gotUniqueStartpoint = true;
9610 : }
9611 :
9612 : /*
9613 : * If two base backups are started at the same time (in WAL sender
9614 : * processes), we need to make sure that they use different
9615 : * checkpoints as starting locations, because we use the starting
9616 : * WAL location as a unique identifier for the base backup in the
9617 : * end-of-backup WAL record and when we write the backup history
9618 : * file. Perhaps it would be better generate a separate unique ID
9619 : * for each backup instead of forcing another checkpoint, but
9620 : * taking a checkpoint right after another is not that expensive
9621 : * either because only few buffers have been dirtied yet.
9622 : */
9623 177 : WALInsertLockAcquireExclusive();
9624 177 : if (XLogCtl->Insert.lastBackupStart < state->startpoint)
9625 : {
9626 177 : XLogCtl->Insert.lastBackupStart = state->startpoint;
9627 177 : gotUniqueStartpoint = true;
9628 : }
9629 177 : WALInsertLockRelease();
9630 177 : } while (!gotUniqueStartpoint);
9631 :
9632 : /*
9633 : * Construct tablespace_map file.
9634 : */
9635 177 : datadirpathlen = strlen(DataDir);
9636 :
9637 : /* Collect information about all tablespaces */
9638 177 : tblspcdir = AllocateDir(PG_TBLSPC_DIR);
9639 568 : while ((de = ReadDir(tblspcdir, PG_TBLSPC_DIR)) != NULL)
9640 : {
9641 : char fullpath[MAXPGPATH + sizeof(PG_TBLSPC_DIR)];
9642 : char linkpath[MAXPGPATH];
9643 391 : char *relpath = NULL;
9644 : char *s;
9645 : PGFileType de_type;
9646 : char *badp;
9647 : Oid tsoid;
9648 :
9649 : /*
9650 : * Try to parse the directory name as an unsigned integer.
9651 : *
9652 : * Tablespace directories should be positive integers that can be
9653 : * represented in 32 bits, with no leading zeroes or trailing
9654 : * garbage. If we come across a name that doesn't meet those
9655 : * criteria, skip it.
9656 : */
9657 391 : if (de->d_name[0] < '1' || de->d_name[1] > '9')
9658 354 : continue;
9659 37 : errno = 0;
9660 37 : tsoid = strtoul(de->d_name, &badp, 10);
9661 37 : if (*badp != '\0' || errno == EINVAL || errno == ERANGE)
9662 0 : continue;
9663 :
9664 37 : snprintf(fullpath, sizeof(fullpath), "%s/%s", PG_TBLSPC_DIR, de->d_name);
9665 :
9666 37 : de_type = get_dirent_type(fullpath, de, false, ERROR);
9667 :
9668 37 : if (de_type == PGFILETYPE_LNK)
9669 : {
9670 : StringInfoData escapedpath;
9671 : int rllen;
9672 :
9673 23 : rllen = readlink(fullpath, linkpath, sizeof(linkpath));
9674 23 : if (rllen < 0)
9675 : {
9676 0 : ereport(WARNING,
9677 : (errmsg("could not read symbolic link \"%s\": %m",
9678 : fullpath)));
9679 0 : continue;
9680 : }
9681 23 : else if (rllen >= sizeof(linkpath))
9682 : {
9683 0 : ereport(WARNING,
9684 : (errmsg("symbolic link \"%s\" target is too long",
9685 : fullpath)));
9686 0 : continue;
9687 : }
9688 23 : linkpath[rllen] = '\0';
9689 :
9690 : /*
9691 : * Relpath holds the relative path of the tablespace directory
9692 : * when it's located within PGDATA, or NULL if it's located
9693 : * elsewhere.
9694 : */
9695 23 : if (rllen > datadirpathlen &&
9696 1 : strncmp(linkpath, DataDir, datadirpathlen) == 0 &&
9697 0 : IS_DIR_SEP(linkpath[datadirpathlen]))
9698 0 : relpath = pstrdup(linkpath + datadirpathlen + 1);
9699 :
9700 : /*
9701 : * Add a backslash-escaped version of the link path to the
9702 : * tablespace map file.
9703 : */
9704 23 : initStringInfo(&escapedpath);
9705 562 : for (s = linkpath; *s; s++)
9706 : {
9707 539 : if (*s == '\n' || *s == '\r' || *s == '\\')
9708 0 : appendStringInfoChar(&escapedpath, '\\');
9709 539 : appendStringInfoChar(&escapedpath, *s);
9710 : }
9711 23 : appendStringInfo(tblspcmapfile, "%s %s\n",
9712 23 : de->d_name, escapedpath.data);
9713 23 : pfree(escapedpath.data);
9714 : }
9715 14 : else if (de_type == PGFILETYPE_DIR)
9716 : {
9717 : /*
9718 : * It's possible to use allow_in_place_tablespaces to create
9719 : * directories directly under pg_tblspc, for testing purposes
9720 : * only.
9721 : *
9722 : * In this case, we store a relative path rather than an
9723 : * absolute path into the tablespaceinfo.
9724 : */
9725 14 : snprintf(linkpath, sizeof(linkpath), "%s/%s",
9726 14 : PG_TBLSPC_DIR, de->d_name);
9727 14 : relpath = pstrdup(linkpath);
9728 : }
9729 : else
9730 : {
9731 : /* Skip any other file type that appears here. */
9732 0 : continue;
9733 : }
9734 :
9735 37 : ti = palloc_object(tablespaceinfo);
9736 37 : ti->oid = tsoid;
9737 37 : ti->path = pstrdup(linkpath);
9738 37 : ti->rpath = relpath;
9739 37 : ti->size = -1;
9740 :
9741 37 : if (tablespaces)
9742 37 : *tablespaces = lappend(*tablespaces, ti);
9743 : }
9744 177 : FreeDir(tblspcdir);
9745 :
9746 177 : state->starttime = (pg_time_t) time(NULL);
9747 : }
9748 177 : PG_END_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9749 :
9750 177 : state->started_in_recovery = backup_started_in_recovery;
9751 :
9752 : /*
9753 : * Mark that the start phase has correctly finished for the backup.
9754 : */
9755 177 : sessionBackupState = SESSION_BACKUP_RUNNING;
9756 177 : }
9757 :
9758 : /*
9759 : * Utility routine to fetch the session-level status of a backup running.
9760 : */
9761 : SessionBackupState
9762 199 : get_backup_status(void)
9763 : {
9764 199 : return sessionBackupState;
9765 : }
9766 :
9767 : /*
9768 : * do_pg_backup_stop
9769 : *
9770 : * Utility function called at the end of an online backup. It creates history
9771 : * file (if required), resets sessionBackupState and so on. It can optionally
9772 : * wait for WAL segments to be archived.
9773 : *
9774 : * "state" is filled with the information necessary to restore from this
9775 : * backup with its stop LSN (stoppoint), its timeline ID (stoptli), etc.
9776 : *
9777 : * It is the responsibility of the caller of this function to verify the
9778 : * permissions of the calling user!
9779 : */
9780 : void
9781 170 : do_pg_backup_stop(BackupState *state, bool waitforarchive)
9782 : {
9783 170 : bool backup_stopped_in_recovery = false;
9784 : char histfilepath[MAXPGPATH];
9785 : char lastxlogfilename[MAXFNAMELEN];
9786 : char histfilename[MAXFNAMELEN];
9787 : XLogSegNo _logSegNo;
9788 : FILE *fp;
9789 : int seconds_before_warning;
9790 170 : int waits = 0;
9791 170 : bool reported_waiting = false;
9792 :
9793 : Assert(state != NULL);
9794 :
9795 170 : backup_stopped_in_recovery = RecoveryInProgress();
9796 :
9797 : /*
9798 : * During recovery, we don't need to check WAL level. Because, if WAL
9799 : * level is not sufficient, it's impossible to get here during recovery.
9800 : */
9801 170 : if (!backup_stopped_in_recovery && !XLogIsNeeded())
9802 0 : ereport(ERROR,
9803 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9804 : errmsg("WAL level not sufficient for making an online backup"),
9805 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9806 :
9807 : /*
9808 : * OK to update backup counter and session-level lock.
9809 : *
9810 : * Note that CHECK_FOR_INTERRUPTS() must not occur while updating them,
9811 : * otherwise they can be updated inconsistently, which might cause
9812 : * do_pg_abort_backup() to fail.
9813 : */
9814 170 : WALInsertLockAcquireExclusive();
9815 :
9816 : /*
9817 : * It is expected that each do_pg_backup_start() call is matched by
9818 : * exactly one do_pg_backup_stop() call.
9819 : */
9820 : Assert(XLogCtl->Insert.runningBackups > 0);
9821 170 : XLogCtl->Insert.runningBackups--;
9822 :
9823 : /*
9824 : * Clean up session-level lock.
9825 : *
9826 : * You might think that WALInsertLockRelease() can be called before
9827 : * cleaning up session-level lock because session-level lock doesn't need
9828 : * to be protected with WAL insertion lock. But since
9829 : * CHECK_FOR_INTERRUPTS() can occur in it, session-level lock must be
9830 : * cleaned up before it.
9831 : */
9832 170 : sessionBackupState = SESSION_BACKUP_NONE;
9833 :
9834 170 : WALInsertLockRelease();
9835 :
9836 : /*
9837 : * If we are taking an online backup from the standby, we confirm that the
9838 : * standby has not been promoted during the backup.
9839 : */
9840 170 : if (state->started_in_recovery && !backup_stopped_in_recovery)
9841 0 : ereport(ERROR,
9842 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9843 : errmsg("the standby was promoted during online backup"),
9844 : errhint("This means that the backup being taken is corrupt "
9845 : "and should not be used. "
9846 : "Try taking another online backup.")));
9847 :
9848 : /*
9849 : * During recovery, we don't write an end-of-backup record. We assume that
9850 : * pg_control was backed up last and its minimum recovery point can be
9851 : * available as the backup end location. Since we don't have an
9852 : * end-of-backup record, we use the pg_control value to check whether
9853 : * we've reached the end of backup when starting recovery from this
9854 : * backup. We have no way of checking if pg_control wasn't backed up last
9855 : * however.
9856 : *
9857 : * We don't force a switch to new WAL file but it is still possible to
9858 : * wait for all the required files to be archived if waitforarchive is
9859 : * true. This is okay if we use the backup to start a standby and fetch
9860 : * the missing WAL using streaming replication. But in the case of an
9861 : * archive recovery, a user should set waitforarchive to true and wait for
9862 : * them to be archived to ensure that all the required files are
9863 : * available.
9864 : *
9865 : * We return the current minimum recovery point as the backup end
9866 : * location. Note that it can be greater than the exact backup end
9867 : * location if the minimum recovery point is updated after the backup of
9868 : * pg_control. This is harmless for current uses.
9869 : *
9870 : * XXX currently a backup history file is for informational and debug
9871 : * purposes only. It's not essential for an online backup. Furthermore,
9872 : * even if it's created, it will not be archived during recovery because
9873 : * an archiver is not invoked. So it doesn't seem worthwhile to write a
9874 : * backup history file during recovery.
9875 : */
9876 170 : if (backup_stopped_in_recovery)
9877 : {
9878 : XLogRecPtr recptr;
9879 :
9880 : /*
9881 : * Check to see if all WAL replayed during online backup contain
9882 : * full-page writes.
9883 : */
9884 8 : SpinLockAcquire(&XLogCtl->info_lck);
9885 8 : recptr = XLogCtl->lastFpwDisableRecPtr;
9886 8 : SpinLockRelease(&XLogCtl->info_lck);
9887 :
9888 8 : if (state->startpoint <= recptr)
9889 0 : ereport(ERROR,
9890 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9891 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9892 : "during online backup"),
9893 : errhint("This means that the backup being taken on the standby "
9894 : "is corrupt and should not be used. "
9895 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9896 : "and then try an online backup again.")));
9897 :
9898 :
9899 8 : LWLockAcquire(ControlFileLock, LW_SHARED);
9900 8 : state->stoppoint = ControlFile->minRecoveryPoint;
9901 8 : state->stoptli = ControlFile->minRecoveryPointTLI;
9902 8 : LWLockRelease(ControlFileLock);
9903 : }
9904 : else
9905 : {
9906 : char *history_file;
9907 :
9908 : /*
9909 : * Write the backup-end xlog record
9910 : */
9911 162 : XLogBeginInsert();
9912 162 : XLogRegisterData(&state->startpoint,
9913 : sizeof(state->startpoint));
9914 162 : state->stoppoint = XLogInsert(RM_XLOG_ID, XLOG_BACKUP_END);
9915 :
9916 : /*
9917 : * Given that we're not in recovery, InsertTimeLineID is set and can't
9918 : * change, so we can read it without a lock.
9919 : */
9920 162 : state->stoptli = XLogCtl->InsertTimeLineID;
9921 :
9922 : /*
9923 : * Force a switch to a new xlog segment file, so that the backup is
9924 : * valid as soon as archiver moves out the current segment file.
9925 : */
9926 162 : RequestXLogSwitch(false);
9927 :
9928 162 : state->stoptime = (pg_time_t) time(NULL);
9929 :
9930 : /*
9931 : * Write the backup history file
9932 : */
9933 162 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9934 162 : BackupHistoryFilePath(histfilepath, state->stoptli, _logSegNo,
9935 : state->startpoint, wal_segment_size);
9936 162 : fp = AllocateFile(histfilepath, "w");
9937 162 : if (!fp)
9938 0 : ereport(ERROR,
9939 : (errcode_for_file_access(),
9940 : errmsg("could not create file \"%s\": %m",
9941 : histfilepath)));
9942 :
9943 : /* Build and save the contents of the backup history file */
9944 162 : history_file = build_backup_content(state, true);
9945 162 : fprintf(fp, "%s", history_file);
9946 162 : pfree(history_file);
9947 :
9948 162 : if (fflush(fp) || ferror(fp) || FreeFile(fp))
9949 0 : ereport(ERROR,
9950 : (errcode_for_file_access(),
9951 : errmsg("could not write file \"%s\": %m",
9952 : histfilepath)));
9953 :
9954 : /*
9955 : * Clean out any no-longer-needed history files. As a side effect,
9956 : * this will post a .ready file for the newly created history file,
9957 : * notifying the archiver that history file may be archived
9958 : * immediately.
9959 : */
9960 162 : CleanupBackupHistory();
9961 : }
9962 :
9963 : /*
9964 : * If archiving is enabled, wait for all the required WAL files to be
9965 : * archived before returning. If archiving isn't enabled, the required WAL
9966 : * needs to be transported via streaming replication (hopefully with
9967 : * wal_keep_size set high enough), or some more exotic mechanism like
9968 : * polling and copying files from pg_wal with script. We have no knowledge
9969 : * of those mechanisms, so it's up to the user to ensure that he gets all
9970 : * the required WAL.
9971 : *
9972 : * We wait until both the last WAL file filled during backup and the
9973 : * history file have been archived, and assume that the alphabetic sorting
9974 : * property of the WAL files ensures any earlier WAL files are safely
9975 : * archived as well.
9976 : *
9977 : * We wait forever, since archive_command is supposed to work and we
9978 : * assume the admin wanted his backup to work completely. If you don't
9979 : * wish to wait, then either waitforarchive should be passed in as false,
9980 : * or you can set statement_timeout. Also, some notices are issued to
9981 : * clue in anyone who might be doing this interactively.
9982 : */
9983 :
9984 170 : if (waitforarchive &&
9985 10 : ((!backup_stopped_in_recovery && XLogArchivingActive()) ||
9986 1 : (backup_stopped_in_recovery && XLogArchivingAlways())))
9987 : {
9988 4 : XLByteToPrevSeg(state->stoppoint, _logSegNo, wal_segment_size);
9989 4 : XLogFileName(lastxlogfilename, state->stoptli, _logSegNo,
9990 : wal_segment_size);
9991 :
9992 4 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9993 4 : BackupHistoryFileName(histfilename, state->stoptli, _logSegNo,
9994 : state->startpoint, wal_segment_size);
9995 :
9996 4 : seconds_before_warning = 60;
9997 4 : waits = 0;
9998 :
9999 12 : while (XLogArchiveIsBusy(lastxlogfilename) ||
10000 4 : XLogArchiveIsBusy(histfilename))
10001 : {
10002 4 : CHECK_FOR_INTERRUPTS();
10003 :
10004 4 : if (!reported_waiting && waits > 5)
10005 : {
10006 0 : ereport(NOTICE,
10007 : (errmsg("base backup done, waiting for required WAL segments to be archived")));
10008 0 : reported_waiting = true;
10009 : }
10010 :
10011 4 : (void) WaitLatch(MyLatch,
10012 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
10013 : 1000L,
10014 : WAIT_EVENT_BACKUP_WAIT_WAL_ARCHIVE);
10015 4 : ResetLatch(MyLatch);
10016 :
10017 4 : if (++waits >= seconds_before_warning)
10018 : {
10019 0 : seconds_before_warning *= 2; /* This wraps in >10 years... */
10020 0 : ereport(WARNING,
10021 : (errmsg("still waiting for all required WAL segments to be archived (%d seconds elapsed)",
10022 : waits),
10023 : errhint("Check that your \"archive_command\" is executing properly. "
10024 : "You can safely cancel this backup, "
10025 : "but the database backup will not be usable without all the WAL segments.")));
10026 : }
10027 : }
10028 :
10029 4 : ereport(NOTICE,
10030 : (errmsg("all required WAL segments have been archived")));
10031 : }
10032 166 : else if (waitforarchive)
10033 6 : ereport(NOTICE,
10034 : (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
10035 170 : }
10036 :
10037 :
10038 : /*
10039 : * do_pg_abort_backup: abort a running backup
10040 : *
10041 : * This does just the most basic steps of do_pg_backup_stop(), by taking the
10042 : * system out of backup mode, thus making it a lot more safe to call from
10043 : * an error handler.
10044 : *
10045 : * 'arg' indicates that it's being called during backup setup; so
10046 : * sessionBackupState has not been modified yet, but runningBackups has
10047 : * already been incremented. When it's false, then it's invoked as a
10048 : * before_shmem_exit handler, and therefore we must not change state
10049 : * unless sessionBackupState indicates that a backup is actually running.
10050 : *
10051 : * NB: This gets used as a PG_ENSURE_ERROR_CLEANUP callback and
10052 : * before_shmem_exit handler, hence the odd-looking signature.
10053 : */
10054 : void
10055 9 : do_pg_abort_backup(int code, Datum arg)
10056 : {
10057 9 : bool during_backup_start = DatumGetBool(arg);
10058 :
10059 : /* If called during backup start, there shouldn't be one already running */
10060 : Assert(!during_backup_start || sessionBackupState == SESSION_BACKUP_NONE);
10061 :
10062 9 : if (during_backup_start || sessionBackupState != SESSION_BACKUP_NONE)
10063 : {
10064 7 : WALInsertLockAcquireExclusive();
10065 : Assert(XLogCtl->Insert.runningBackups > 0);
10066 7 : XLogCtl->Insert.runningBackups--;
10067 :
10068 7 : sessionBackupState = SESSION_BACKUP_NONE;
10069 7 : WALInsertLockRelease();
10070 :
10071 7 : if (!during_backup_start)
10072 7 : ereport(WARNING,
10073 : errmsg("aborting backup due to backend exiting before pg_backup_stop was called"));
10074 : }
10075 9 : }
10076 :
10077 : /*
10078 : * Register a handler that will warn about unterminated backups at end of
10079 : * session, unless this has already been done.
10080 : */
10081 : void
10082 4 : register_persistent_abort_backup_handler(void)
10083 : {
10084 : static bool already_done = false;
10085 :
10086 4 : if (already_done)
10087 1 : return;
10088 3 : before_shmem_exit(do_pg_abort_backup, BoolGetDatum(false));
10089 3 : already_done = true;
10090 : }
10091 :
10092 : /*
10093 : * Get latest WAL insert pointer
10094 : */
10095 : XLogRecPtr
10096 2143 : GetXLogInsertRecPtr(void)
10097 : {
10098 2143 : XLogCtlInsert *Insert = &XLogCtl->Insert;
10099 : uint64 current_bytepos;
10100 :
10101 2143 : SpinLockAcquire(&Insert->insertpos_lck);
10102 2143 : current_bytepos = Insert->CurrBytePos;
10103 2143 : SpinLockRelease(&Insert->insertpos_lck);
10104 :
10105 2143 : return XLogBytePosToRecPtr(current_bytepos);
10106 : }
10107 :
10108 : /*
10109 : * Get latest WAL record end pointer
10110 : */
10111 : XLogRecPtr
10112 2269 : GetXLogInsertEndRecPtr(void)
10113 : {
10114 2269 : XLogCtlInsert *Insert = &XLogCtl->Insert;
10115 : uint64 current_bytepos;
10116 :
10117 2269 : SpinLockAcquire(&Insert->insertpos_lck);
10118 2269 : current_bytepos = Insert->CurrBytePos;
10119 2269 : SpinLockRelease(&Insert->insertpos_lck);
10120 :
10121 2269 : return XLogBytePosToEndRecPtr(current_bytepos);
10122 : }
10123 :
10124 : /*
10125 : * Get latest WAL write pointer
10126 : */
10127 : XLogRecPtr
10128 15016 : GetXLogWriteRecPtr(void)
10129 : {
10130 15016 : RefreshXLogWriteResult(LogwrtResult);
10131 :
10132 15016 : return LogwrtResult.Write;
10133 : }
10134 :
10135 : /*
10136 : * Returns the redo pointer of the last checkpoint or restartpoint. This is
10137 : * the oldest point in WAL that we still need, if we have to restart recovery.
10138 : */
10139 : void
10140 390 : GetOldestRestartPoint(XLogRecPtr *oldrecptr, TimeLineID *oldtli)
10141 : {
10142 390 : LWLockAcquire(ControlFileLock, LW_SHARED);
10143 390 : *oldrecptr = ControlFile->checkPointCopy.redo;
10144 390 : *oldtli = ControlFile->checkPointCopy.ThisTimeLineID;
10145 390 : LWLockRelease(ControlFileLock);
10146 390 : }
10147 :
10148 : /* Thin wrapper around ShutdownWalRcv(). */
10149 : void
10150 1073 : XLogShutdownWalRcv(void)
10151 : {
10152 : Assert(AmStartupProcess() || !IsUnderPostmaster);
10153 :
10154 1073 : ShutdownWalRcv();
10155 1073 : ResetInstallXLogFileSegmentActive();
10156 1073 : }
10157 :
10158 : /* Enable WAL file recycling and preallocation. */
10159 : void
10160 1261 : SetInstallXLogFileSegmentActive(void)
10161 : {
10162 1261 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
10163 1261 : XLogCtl->InstallXLogFileSegmentActive = true;
10164 1261 : LWLockRelease(ControlFileLock);
10165 1261 : }
10166 :
10167 : /* Disable WAL file recycling and preallocation. */
10168 : void
10169 1219 : ResetInstallXLogFileSegmentActive(void)
10170 : {
10171 1219 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
10172 1219 : XLogCtl->InstallXLogFileSegmentActive = false;
10173 1219 : LWLockRelease(ControlFileLock);
10174 1219 : }
10175 :
10176 : bool
10177 0 : IsInstallXLogFileSegmentActive(void)
10178 : {
10179 : bool result;
10180 :
10181 0 : LWLockAcquire(ControlFileLock, LW_SHARED);
10182 0 : result = XLogCtl->InstallXLogFileSegmentActive;
10183 0 : LWLockRelease(ControlFileLock);
10184 :
10185 0 : return result;
10186 : }
10187 :
10188 : /*
10189 : * Update the WalWriterSleeping flag.
10190 : */
10191 : void
10192 609 : SetWalWriterSleeping(bool sleeping)
10193 : {
10194 609 : SpinLockAcquire(&XLogCtl->info_lck);
10195 609 : XLogCtl->WalWriterSleeping = sleeping;
10196 609 : SpinLockRelease(&XLogCtl->info_lck);
10197 609 : }
|
