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 24106330 : XLogInsertRecord(XLogRecData *rdata,
785 : XLogRecPtr fpw_lsn,
786 : uint8 flags,
787 : int num_fpi,
788 : uint64 fpi_bytes,
789 : bool topxid_included)
790 : {
791 24106330 : XLogCtlInsert *Insert = &XLogCtl->Insert;
792 : pg_crc32c rdata_crc;
793 : bool inserted;
794 24106330 : XLogRecord *rechdr = (XLogRecord *) rdata->data;
795 24106330 : uint8 info = rechdr->xl_info & ~XLR_INFO_MASK;
796 24106330 : WalInsertClass class = WALINSERT_NORMAL;
797 : XLogRecPtr StartPos;
798 : XLogRecPtr EndPos;
799 24106330 : bool prevDoPageWrites = doPageWrites;
800 : TimeLineID insertTLI;
801 :
802 : /* Does this record type require special handling? */
803 24106330 : if (unlikely(rechdr->xl_rmid == RM_XLOG_ID))
804 : {
805 437399 : if (info == XLOG_SWITCH)
806 814 : class = WALINSERT_SPECIAL_SWITCH;
807 436585 : else if (info == XLOG_CHECKPOINT_REDO)
808 992 : 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 24106330 : 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 24106330 : 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 24106330 : START_CRIT_SECTION();
857 :
858 24106330 : if (likely(class == WALINSERT_NORMAL))
859 : {
860 24104524 : 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 24104524 : if (RedoRecPtr != Insert->RedoRecPtr)
879 : {
880 : Assert(RedoRecPtr < Insert->RedoRecPtr);
881 8102 : RedoRecPtr = Insert->RedoRecPtr;
882 : }
883 24104524 : doPageWrites = (Insert->fullPageWrites || Insert->runningBackups > 0);
884 :
885 24104524 : if (doPageWrites &&
886 21800875 : (!prevDoPageWrites ||
887 20268429 : (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 8795 : WALInsertLockRelease();
894 8795 : END_CRIT_SECTION();
895 8795 : return InvalidXLogRecPtr;
896 : }
897 :
898 : /*
899 : * Reserve space for the record in the WAL. This also sets the xl_prev
900 : * pointer.
901 : */
902 24095729 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
903 : &rechdr->xl_prev);
904 :
905 : /* Normal records are always inserted. */
906 24095729 : inserted = true;
907 : }
908 1806 : 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));
922 814 : WALInsertLockAcquireExclusive();
923 814 : 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));
937 992 : WALInsertLockAcquireExclusive();
938 992 : ReserveXLogInsertLocation(rechdr->xl_tot_len, &StartPos, &EndPos,
939 : &rechdr->xl_prev);
940 992 : RedoRecPtr = Insert->RedoRecPtr = StartPos;
941 992 : inserted = true;
942 : }
943 :
944 24097535 : if (inserted)
945 : {
946 : /*
947 : * Now that xl_prev has been filled in, calculate CRC of the record
948 : * header.
949 : */
950 24097472 : rdata_crc = rechdr->xl_crc;
951 24097472 : COMP_CRC32C(rdata_crc, rechdr, offsetof(XLogRecord, xl_crc));
952 24097472 : FIN_CRC32C(rdata_crc);
953 24097472 : 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 24097472 : 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 24097472 : if ((flags & XLOG_MARK_UNIMPORTANT) == 0)
969 : {
970 23855781 : int lockno = holdingAllLocks ? 0 : MyLockNo;
971 :
972 23855781 : 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 24097535 : WALInsertLockRelease();
988 :
989 24097535 : END_CRIT_SECTION();
990 :
991 24097535 : 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 24097535 : if (topxid_included)
998 223 : MarkSubxactTopXidLogged();
999 :
1000 : /*
1001 : * Update shared LogwrtRqst.Write, if we crossed page boundary.
1002 : */
1003 24097535 : if (StartPos / XLOG_BLCKSZ != EndPos / XLOG_BLCKSZ)
1004 : {
1005 1880850 : SpinLockAcquire(&XLogCtl->info_lck);
1006 : /* advance global request to include new block(s) */
1007 1880850 : if (XLogCtl->LogwrtRqst.Write < EndPos)
1008 1816255 : XLogCtl->LogwrtRqst.Write = EndPos;
1009 1880850 : SpinLockRelease(&XLogCtl->info_lck);
1010 1880850 : 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 24097535 : if (class == WALINSERT_SPECIAL_SWITCH)
1019 : {
1020 : TRACE_POSTGRESQL_WAL_SWITCH();
1021 814 : 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 814 : if (inserted)
1029 : {
1030 751 : EndPos = StartPos + SizeOfXLogRecord;
1031 751 : 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 24097535 : ProcLastRecPtr = StartPos;
1113 24097535 : XactLastRecEnd = EndPos;
1114 :
1115 : /* Report WAL traffic to the instrumentation. */
1116 24097535 : if (inserted)
1117 : {
1118 24097472 : pgWalUsage.wal_bytes += rechdr->xl_tot_len;
1119 24097472 : pgWalUsage.wal_records++;
1120 24097472 : pgWalUsage.wal_fpi += num_fpi;
1121 24097472 : pgWalUsage.wal_fpi_bytes += fpi_bytes;
1122 :
1123 : /* Required for the flush of pending stats WAL data */
1124 24097472 : pgstat_report_fixed = true;
1125 : }
1126 :
1127 24097535 : 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 24096721 : ReserveXLogInsertLocation(int size, XLogRecPtr *StartPos, XLogRecPtr *EndPos,
1150 : XLogRecPtr *PrevPtr)
1151 : {
1152 24096721 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1153 : uint64 startbytepos;
1154 : uint64 endbytepos;
1155 : uint64 prevbytepos;
1156 :
1157 24096721 : 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 24096721 : SpinLockAcquire(&Insert->insertpos_lck);
1173 :
1174 24096721 : startbytepos = Insert->CurrBytePos;
1175 24096721 : endbytepos = startbytepos + size;
1176 24096721 : prevbytepos = Insert->PrevBytePos;
1177 24096721 : Insert->CurrBytePos = endbytepos;
1178 24096721 : Insert->PrevBytePos = startbytepos;
1179 :
1180 24096721 : SpinLockRelease(&Insert->insertpos_lck);
1181 :
1182 24096721 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1183 24096721 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1184 24096721 : *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 24096721 : }
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 814 : ReserveXLogSwitch(XLogRecPtr *StartPos, XLogRecPtr *EndPos, XLogRecPtr *PrevPtr)
1206 : {
1207 814 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1208 : uint64 startbytepos;
1209 : uint64 endbytepos;
1210 : uint64 prevbytepos;
1211 814 : 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 814 : SpinLockAcquire(&Insert->insertpos_lck);
1222 :
1223 814 : startbytepos = Insert->CurrBytePos;
1224 :
1225 814 : ptr = XLogBytePosToEndRecPtr(startbytepos);
1226 814 : if (XLogSegmentOffset(ptr, wal_segment_size) == 0)
1227 : {
1228 63 : SpinLockRelease(&Insert->insertpos_lck);
1229 63 : *EndPos = *StartPos = ptr;
1230 63 : return false;
1231 : }
1232 :
1233 751 : endbytepos = startbytepos + size;
1234 751 : prevbytepos = Insert->PrevBytePos;
1235 :
1236 751 : *StartPos = XLogBytePosToRecPtr(startbytepos);
1237 751 : *EndPos = XLogBytePosToEndRecPtr(endbytepos);
1238 :
1239 751 : segleft = wal_segment_size - XLogSegmentOffset(*EndPos, wal_segment_size);
1240 751 : if (segleft != wal_segment_size)
1241 : {
1242 : /* consume the rest of the segment */
1243 751 : *EndPos += segleft;
1244 751 : endbytepos = XLogRecPtrToBytePos(*EndPos);
1245 : }
1246 751 : Insert->CurrBytePos = endbytepos;
1247 751 : Insert->PrevBytePos = startbytepos;
1248 :
1249 751 : SpinLockRelease(&Insert->insertpos_lck);
1250 :
1251 751 : *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 751 : 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 24097472 : 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 24097472 : CurrPos = StartPos;
1280 24097472 : currpos = GetXLogBuffer(CurrPos, tli);
1281 24097472 : 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 24097472 : written = 0;
1291 109619608 : while (rdata != NULL)
1292 : {
1293 85522136 : const char *rdata_data = rdata->data;
1294 85522136 : int rdata_len = rdata->len;
1295 :
1296 87524556 : 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 2002420 : memcpy(currpos, rdata_data, freespace);
1303 2002420 : rdata_data += freespace;
1304 2002420 : rdata_len -= freespace;
1305 2002420 : written += freespace;
1306 2002420 : 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 2002420 : currpos = GetXLogBuffer(CurrPos, tli);
1318 2002420 : pagehdr = (XLogPageHeader) currpos;
1319 2002420 : pagehdr->xlp_rem_len = write_len - written;
1320 2002420 : pagehdr->xlp_info |= XLP_FIRST_IS_CONTRECORD;
1321 :
1322 : /* skip over the page header */
1323 2002420 : if (XLogSegmentOffset(CurrPos, wal_segment_size) == 0)
1324 : {
1325 1280 : CurrPos += SizeOfXLogLongPHD;
1326 1280 : currpos += SizeOfXLogLongPHD;
1327 : }
1328 : else
1329 : {
1330 2001140 : CurrPos += SizeOfXLogShortPHD;
1331 2001140 : currpos += SizeOfXLogShortPHD;
1332 : }
1333 2002420 : freespace = INSERT_FREESPACE(CurrPos);
1334 : }
1335 :
1336 : Assert(CurrPos % XLOG_BLCKSZ >= SizeOfXLogShortPHD || rdata_len == 0);
1337 85522136 : memcpy(currpos, rdata_data, rdata_len);
1338 85522136 : currpos += rdata_len;
1339 85522136 : CurrPos += rdata_len;
1340 85522136 : freespace -= rdata_len;
1341 85522136 : written += rdata_len;
1342 :
1343 85522136 : 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 24097472 : 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 751 : 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 747586 : 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 746835 : currpos = GetXLogBuffer(CurrPos, tli);
1391 2987340 : MemSet(currpos, 0, SizeOfXLogShortPHD);
1392 :
1393 746835 : CurrPos += XLOG_BLCKSZ;
1394 : }
1395 : }
1396 : else
1397 : {
1398 : /* Align the end position, so that the next record starts aligned */
1399 24096721 : CurrPos = MAXALIGN64(CurrPos);
1400 : }
1401 :
1402 24097472 : 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 24097472 : }
1407 :
1408 : /*
1409 : * Acquire a WAL insertion lock, for inserting to WAL.
1410 : */
1411 : static void
1412 24105526 : 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 24105526 : if (lockToTry == -1)
1430 9259 : lockToTry = MyProcNumber % NUM_XLOGINSERT_LOCKS;
1431 24105526 : 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 24105526 : immed = LWLockAcquire(&WALInsertLocks[MyLockNo].l.lock, LW_EXCLUSIVE);
1438 24105526 : 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 20378 : lockToTry = (lockToTry + 1) % NUM_XLOGINSERT_LOCKS;
1449 : }
1450 24105526 : }
1451 :
1452 : /*
1453 : * Acquire all WAL insertion locks, to prevent other backends from inserting
1454 : * to WAL.
1455 : */
1456 : static void
1457 4747 : 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 37976 : for (i = 0; i < NUM_XLOGINSERT_LOCKS - 1; i++)
1467 : {
1468 33229 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1469 33229 : LWLockUpdateVar(&WALInsertLocks[i].l.lock,
1470 33229 : &WALInsertLocks[i].l.insertingAt,
1471 : PG_UINT64_MAX);
1472 : }
1473 : /* Variable value reset to 0 at release */
1474 4747 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
1475 :
1476 4747 : holdingAllLocks = true;
1477 4747 : }
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 24110273 : WALInsertLockRelease(void)
1487 : {
1488 24110273 : if (holdingAllLocks)
1489 : {
1490 : int i;
1491 :
1492 42723 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1493 37976 : LWLockReleaseClearVar(&WALInsertLocks[i].l.lock,
1494 37976 : &WALInsertLocks[i].l.insertingAt,
1495 : 0);
1496 :
1497 4747 : holdingAllLocks = false;
1498 : }
1499 : else
1500 : {
1501 24105526 : LWLockReleaseClearVar(&WALInsertLocks[MyLockNo].l.lock,
1502 24105526 : &WALInsertLocks[MyLockNo].l.insertingAt,
1503 : 0);
1504 : }
1505 24110273 : }
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 2703844 : WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt)
1513 : {
1514 2703844 : if (holdingAllLocks)
1515 : {
1516 : /*
1517 : * We use the last lock to mark our actual position, see comments in
1518 : * WALInsertLockAcquireExclusive.
1519 : */
1520 744573 : LWLockUpdateVar(&WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.lock,
1521 744573 : &WALInsertLocks[NUM_XLOGINSERT_LOCKS - 1].l.insertingAt,
1522 : insertingAt);
1523 : }
1524 : else
1525 1959271 : LWLockUpdateVar(&WALInsertLocks[MyLockNo].l.lock,
1526 1959271 : &WALInsertLocks[MyLockNo].l.insertingAt,
1527 : insertingAt);
1528 2703844 : }
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 2507031 : WaitXLogInsertionsToFinish(XLogRecPtr upto)
1546 : {
1547 : uint64 bytepos;
1548 : XLogRecPtr inserted;
1549 : XLogRecPtr reservedUpto;
1550 : XLogRecPtr finishedUpto;
1551 2507031 : XLogCtlInsert *Insert = &XLogCtl->Insert;
1552 : int i;
1553 :
1554 2507031 : 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 2507031 : inserted = pg_atomic_read_membarrier_u64(&XLogCtl->logInsertResult);
1562 2507031 : if (upto <= inserted)
1563 1995798 : return inserted;
1564 :
1565 : /* Read the current insert position */
1566 511233 : SpinLockAcquire(&Insert->insertpos_lck);
1567 511233 : bytepos = Insert->CurrBytePos;
1568 511233 : SpinLockRelease(&Insert->insertpos_lck);
1569 511233 : 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 511233 : 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 511233 : finishedUpto = reservedUpto;
1597 4601097 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
1598 : {
1599 4089864 : 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 4196823 : if (LWLockWaitForVar(&WALInsertLocks[i].l.lock,
1627 4196823 : &WALInsertLocks[i].l.insertingAt,
1628 : insertingat, &insertingat))
1629 : {
1630 : /* the lock was free, so no insertion in progress */
1631 2954379 : insertingat = InvalidXLogRecPtr;
1632 2954379 : break;
1633 : }
1634 :
1635 : /*
1636 : * This insertion is still in progress. Have to wait, unless the
1637 : * inserter has proceeded past 'upto'.
1638 : */
1639 1242444 : } while (insertingat < upto);
1640 :
1641 4089864 : if (XLogRecPtrIsValid(insertingat) && insertingat < finishedUpto)
1642 413016 : 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 511233 : finishedUpto = pg_atomic_monotonic_advance_u64(&XLogCtl->logInsertResult,
1651 : finishedUpto);
1652 :
1653 511233 : 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 26846737 : 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 26846737 : 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 23639342 : 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 3207395 : 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 3207395 : expectedEndPtr = ptr;
1713 3207395 : expectedEndPtr += XLOG_BLCKSZ - ptr % XLOG_BLCKSZ;
1714 :
1715 3207395 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1716 3207395 : 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 2703844 : if (ptr % XLOG_BLCKSZ == SizeOfXLogShortPHD &&
1735 12215 : XLogSegmentOffset(ptr, wal_segment_size) > XLOG_BLCKSZ)
1736 12215 : initializedUpto = ptr - SizeOfXLogShortPHD;
1737 2691629 : else if (ptr % XLOG_BLCKSZ == SizeOfXLogLongPHD &&
1738 1059 : XLogSegmentOffset(ptr, wal_segment_size) < XLOG_BLCKSZ)
1739 626 : initializedUpto = ptr - SizeOfXLogLongPHD;
1740 : else
1741 2691003 : initializedUpto = ptr;
1742 :
1743 2703844 : WALInsertLockUpdateInsertingAt(initializedUpto);
1744 :
1745 2703844 : AdvanceXLInsertBuffer(ptr, tli, false);
1746 2703844 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1747 :
1748 2703844 : 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 503551 : 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 3207395 : cachedPage = ptr / XLOG_BLCKSZ;
1766 3207395 : 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 3207395 : 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 106594 : WALReadFromBuffers(char *dstbuf, XLogRecPtr startptr, Size count,
1790 : TimeLineID tli)
1791 : {
1792 106594 : char *pdst = dstbuf;
1793 106594 : XLogRecPtr recptr = startptr;
1794 : XLogRecPtr inserted;
1795 106594 : Size nbytes = count;
1796 :
1797 106594 : if (RecoveryInProgress() || tli != GetWALInsertionTimeLine())
1798 913 : 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 105681 : inserted = pg_atomic_read_u64(&XLogCtl->logInsertResult);
1807 105681 : 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 135589 : while (nbytes > 0)
1828 : {
1829 127004 : uint32 offset = recptr % XLOG_BLCKSZ;
1830 127004 : 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 127004 : 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 127004 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1848 127004 : if (expectedEndPtr != endptr)
1849 97092 : 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 29912 : page = XLogCtl->pages + idx * (Size) XLOG_BLCKSZ;
1857 29912 : psrc = page + offset;
1858 29912 : 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 29912 : pg_read_barrier();
1865 :
1866 : /* data copy */
1867 29912 : memcpy(pdst, psrc, npagebytes);
1868 :
1869 : /*
1870 : * Ensure that the data copy and the second verification step are not
1871 : * reordered.
1872 : */
1873 29912 : 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 29912 : endptr = pg_atomic_read_u64(&XLogCtl->xlblocks[idx]);
1880 29912 : if (expectedEndPtr != endptr)
1881 4 : break;
1882 :
1883 29908 : pdst += npagebytes;
1884 29908 : recptr += npagebytes;
1885 29908 : nbytes -= npagebytes;
1886 : }
1887 :
1888 : Assert(pdst - dstbuf <= count);
1889 :
1890 105681 : 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 48197808 : XLogBytePosToRecPtr(uint64 bytepos)
1900 : {
1901 : uint64 fullsegs;
1902 : uint64 fullpages;
1903 : uint64 bytesleft;
1904 : uint32 seg_offset;
1905 : XLogRecPtr result;
1906 :
1907 48197808 : fullsegs = bytepos / UsableBytesInSegment;
1908 48197808 : bytesleft = bytepos % UsableBytesInSegment;
1909 :
1910 48197808 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1911 : {
1912 : /* fits on first page of segment */
1913 69180 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1914 : }
1915 : else
1916 : {
1917 : /* account for the first page on segment with long header */
1918 48128628 : seg_offset = XLOG_BLCKSZ;
1919 48128628 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1920 :
1921 48128628 : fullpages = bytesleft / UsableBytesInPage;
1922 48128628 : bytesleft = bytesleft % UsableBytesInPage;
1923 :
1924 48128628 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1925 : }
1926 :
1927 48197808 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1928 :
1929 48197808 : 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 24726693 : XLogBytePosToEndRecPtr(uint64 bytepos)
1940 : {
1941 : uint64 fullsegs;
1942 : uint64 fullpages;
1943 : uint64 bytesleft;
1944 : uint32 seg_offset;
1945 : XLogRecPtr result;
1946 :
1947 24726693 : fullsegs = bytepos / UsableBytesInSegment;
1948 24726693 : bytesleft = bytepos % UsableBytesInSegment;
1949 :
1950 24726693 : if (bytesleft < XLOG_BLCKSZ - SizeOfXLogLongPHD)
1951 : {
1952 : /* fits on first page of segment */
1953 107121 : if (bytesleft == 0)
1954 71117 : seg_offset = 0;
1955 : else
1956 36004 : seg_offset = bytesleft + SizeOfXLogLongPHD;
1957 : }
1958 : else
1959 : {
1960 : /* account for the first page on segment with long header */
1961 24619572 : seg_offset = XLOG_BLCKSZ;
1962 24619572 : bytesleft -= XLOG_BLCKSZ - SizeOfXLogLongPHD;
1963 :
1964 24619572 : fullpages = bytesleft / UsableBytesInPage;
1965 24619572 : bytesleft = bytesleft % UsableBytesInPage;
1966 :
1967 24619572 : if (bytesleft == 0)
1968 23659 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft;
1969 : else
1970 24595913 : seg_offset += fullpages * XLOG_BLCKSZ + bytesleft + SizeOfXLogShortPHD;
1971 : }
1972 :
1973 24726693 : XLogSegNoOffsetToRecPtr(fullsegs, seg_offset, wal_segment_size, result);
1974 :
1975 24726693 : return result;
1976 : }
1977 :
1978 : /*
1979 : * Convert an XLogRecPtr to a "usable byte position".
1980 : */
1981 : static uint64
1982 2765 : XLogRecPtrToBytePos(XLogRecPtr ptr)
1983 : {
1984 : uint64 fullsegs;
1985 : uint32 fullpages;
1986 : uint32 offset;
1987 : uint64 result;
1988 :
1989 2765 : XLByteToSeg(ptr, fullsegs, wal_segment_size);
1990 :
1991 2765 : fullpages = (XLogSegmentOffset(ptr, wal_segment_size)) / XLOG_BLCKSZ;
1992 2765 : offset = ptr % XLOG_BLCKSZ;
1993 :
1994 2765 : if (fullpages == 0)
1995 : {
1996 1059 : result = fullsegs * UsableBytesInSegment;
1997 1059 : if (offset > 0)
1998 : {
1999 : Assert(offset >= SizeOfXLogLongPHD);
2000 285 : result += offset - SizeOfXLogLongPHD;
2001 : }
2002 : }
2003 : else
2004 : {
2005 1706 : result = fullsegs * UsableBytesInSegment +
2006 1706 : (XLOG_BLCKSZ - SizeOfXLogLongPHD) + /* account for first page */
2007 1706 : (fullpages - 1) * UsableBytesInPage; /* full pages */
2008 1706 : if (offset > 0)
2009 : {
2010 : Assert(offset >= SizeOfXLogShortPHD);
2011 1697 : result += offset - SizeOfXLogShortPHD;
2012 : }
2013 : }
2014 :
2015 2765 : 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 2708701 : AdvanceXLInsertBuffer(XLogRecPtr upto, TimeLineID tli, bool opportunistic)
2027 : {
2028 : int nextidx;
2029 : XLogRecPtr OldPageRqstPtr;
2030 : XLogwrtRqst WriteRqst;
2031 2708701 : XLogRecPtr NewPageEndPtr = InvalidXLogRecPtr;
2032 : XLogRecPtr NewPageBeginPtr;
2033 : XLogPageHeader NewPage;
2034 2708701 : int npages pg_attribute_unused() = 0;
2035 :
2036 2708701 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2037 :
2038 : /*
2039 : * Now that we have the lock, check if someone initialized the page
2040 : * already.
2041 : */
2042 7847407 : while (upto >= XLogCtl->InitializedUpTo || opportunistic)
2043 : {
2044 5143563 : 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 5143563 : OldPageRqstPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[nextidx]);
2052 5143563 : 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 2358598 : if (opportunistic)
2059 4857 : break;
2060 :
2061 : /* Advance shared memory write request position */
2062 2353741 : SpinLockAcquire(&XLogCtl->info_lck);
2063 2353741 : if (XLogCtl->LogwrtRqst.Write < OldPageRqstPtr)
2064 696853 : XLogCtl->LogwrtRqst.Write = OldPageRqstPtr;
2065 2353741 : 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 2353741 : RefreshXLogWriteResult(LogwrtResult);
2072 2353741 : 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 2335094 : LWLockRelease(WALBufMappingLock);
2081 :
2082 2335094 : WaitXLogInsertionsToFinish(OldPageRqstPtr);
2083 :
2084 2335094 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
2085 :
2086 2335094 : RefreshXLogWriteResult(LogwrtResult);
2087 2335094 : if (LogwrtResult.Write >= OldPageRqstPtr)
2088 : {
2089 : /* OK, someone wrote it already */
2090 150752 : LWLockRelease(WALWriteLock);
2091 : }
2092 : else
2093 : {
2094 : /* Have to write it ourselves */
2095 : TRACE_POSTGRESQL_WAL_BUFFER_WRITE_DIRTY_START();
2096 2184342 : WriteRqst.Write = OldPageRqstPtr;
2097 2184342 : WriteRqst.Flush = InvalidXLogRecPtr;
2098 2184342 : XLogWrite(WriteRqst, tli, false);
2099 2184342 : LWLockRelease(WALWriteLock);
2100 2184342 : 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 2184342 : pgstat_report_fixed = true;
2108 : }
2109 : /* Re-acquire WALBufMappingLock and retry */
2110 2335094 : LWLockAcquire(WALBufMappingLock, LW_EXCLUSIVE);
2111 2335094 : 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 2803612 : NewPageBeginPtr = XLogCtl->InitializedUpTo;
2120 2803612 : NewPageEndPtr = NewPageBeginPtr + XLOG_BLCKSZ;
2121 :
2122 : Assert(XLogRecPtrToBufIdx(NewPageBeginPtr) == nextidx);
2123 :
2124 2803612 : 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 2803612 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], InvalidXLogRecPtr);
2132 2803612 : 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 2803612 : MemSet(NewPage, 0, XLOG_BLCKSZ);
2139 :
2140 : /*
2141 : * Fill the new page's header
2142 : */
2143 2803612 : NewPage->xlp_magic = XLOG_PAGE_MAGIC;
2144 :
2145 : /* NewPage->xlp_info = 0; */ /* done by memset */
2146 2803612 : NewPage->xlp_tli = tli;
2147 2803612 : 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 2803612 : if ((XLogSegmentOffset(NewPage->xlp_pageaddr, wal_segment_size)) == 0)
2155 : {
2156 1927 : XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
2157 :
2158 1927 : NewLongPage->xlp_sysid = ControlFile->system_identifier;
2159 1927 : NewLongPage->xlp_seg_size = wal_segment_size;
2160 1927 : NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
2161 1927 : 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 2803612 : pg_write_barrier();
2170 :
2171 2803612 : pg_atomic_write_u64(&XLogCtl->xlblocks[nextidx], NewPageEndPtr);
2172 2803612 : XLogCtl->InitializedUpTo = NewPageEndPtr;
2173 :
2174 2803612 : npages++;
2175 : }
2176 2708701 : 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 2708701 : }
2186 :
2187 : /*
2188 : * Calculate CheckPointSegments based on max_wal_size_mb and
2189 : * checkpoint_completion_target.
2190 : */
2191 : static void
2192 9743 : 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 9743 : target = (double) ConvertToXSegs(max_wal_size_mb, wal_segment_size) /
2211 9743 : (1.0 + CheckPointCompletionTarget);
2212 :
2213 : /* round down */
2214 9743 : CheckPointSegments = (int) target;
2215 :
2216 9743 : if (CheckPointSegments < 1)
2217 10 : CheckPointSegments = 1;
2218 9743 : }
2219 :
2220 : void
2221 7335 : assign_max_wal_size(int newval, void *extra)
2222 : {
2223 7335 : max_wal_size_mb = newval;
2224 7335 : CalculateCheckpointSegments();
2225 7335 : }
2226 :
2227 : void
2228 1275 : assign_checkpoint_completion_target(double newval, void *extra)
2229 : {
2230 1275 : CheckPointCompletionTarget = newval;
2231 1275 : CalculateCheckpointSegments();
2232 1275 : }
2233 :
2234 : bool
2235 2466 : check_wal_segment_size(int *newval, void **extra, GucSource source)
2236 : {
2237 2466 : 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 2466 : 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 1931 : 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 1931 : minSegNo = lastredoptr / wal_segment_size +
2264 1931 : ConvertToXSegs(min_wal_size_mb, wal_segment_size) - 1;
2265 1931 : maxSegNo = lastredoptr / wal_segment_size +
2266 1931 : 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 1931 : distance = (1.0 + CheckPointCompletionTarget) * CheckPointDistanceEstimate;
2277 : /* add 10% for good measure. */
2278 1931 : distance *= 1.10;
2279 :
2280 1931 : recycleSegNo = (XLogSegNo) ceil(((double) lastredoptr + distance) /
2281 : wal_segment_size);
2282 :
2283 1931 : if (recycleSegNo < minSegNo)
2284 1356 : recycleSegNo = minSegNo;
2285 1931 : if (recycleSegNo > maxSegNo)
2286 426 : recycleSegNo = maxSegNo;
2287 :
2288 1931 : 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 5104 : XLogCheckpointNeeded(XLogSegNo new_segno)
2302 : {
2303 : XLogSegNo old_segno;
2304 :
2305 5104 : XLByteToSeg(RedoRecPtr, old_segno, wal_segment_size);
2306 :
2307 5104 : if (new_segno >= old_segno + (uint64) (CheckPointSegments - 1))
2308 3183 : return true;
2309 1921 : 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 2350077 : 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 2350077 : 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 2350077 : npages = 0;
2353 2350077 : startidx = 0;
2354 2350077 : 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 2350077 : curridx = XLogRecPtrToBufIdx(LogwrtResult.Write);
2362 :
2363 5112863 : 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 2923570 : XLogRecPtr EndPtr = pg_atomic_read_u64(&XLogCtl->xlblocks[curridx]);
2371 :
2372 2923570 : 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 2923570 : LogwrtResult.Write = EndPtr;
2379 2923570 : ispartialpage = WriteRqst.Write < LogwrtResult.Write;
2380 :
2381 2923570 : 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 15129 : if (openLogFile >= 0)
2390 6840 : XLogFileClose();
2391 15129 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2392 : wal_segment_size);
2393 15129 : openLogTLI = tli;
2394 :
2395 : /* create/use new log file */
2396 15129 : openLogFile = XLogFileInit(openLogSegNo, tli);
2397 15129 : ReserveExternalFD();
2398 : }
2399 :
2400 : /* Make sure we have the current logfile open */
2401 2923570 : 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 2923570 : if (npages == 0)
2412 : {
2413 : /* first of group */
2414 2367187 : startidx = curridx;
2415 2367187 : startoffset = XLogSegmentOffset(LogwrtResult.Write - XLOG_BLCKSZ,
2416 : wal_segment_size);
2417 : }
2418 2923570 : 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 2923570 : last_iteration = WriteRqst.Write <= LogwrtResult.Write;
2427 :
2428 5690240 : finishing_seg = !ispartialpage &&
2429 2766670 : (startoffset + npages * XLOG_BLCKSZ) >= wal_segment_size;
2430 :
2431 2923570 : if (last_iteration ||
2432 575146 : 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 2367187 : from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
2443 2367187 : nbytes = npages * (Size) XLOG_BLCKSZ;
2444 2367187 : nleft = nbytes;
2445 : do
2446 : {
2447 2367187 : errno = 0;
2448 :
2449 : /*
2450 : * Measure I/O timing to write WAL data, for pg_stat_io.
2451 : */
2452 2367187 : start = pgstat_prepare_io_time(track_wal_io_timing);
2453 :
2454 2367187 : pgstat_report_wait_start(WAIT_EVENT_WAL_WRITE);
2455 2367187 : written = pg_pwrite(openLogFile, from, nleft, startoffset);
2456 2367187 : pgstat_report_wait_end();
2457 :
2458 2367187 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL,
2459 : IOOP_WRITE, start, 1, written);
2460 :
2461 2367187 : 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 2367187 : nleft -= written;
2479 2367187 : from += written;
2480 2367187 : startoffset += written;
2481 2367187 : } while (nleft > 0);
2482 :
2483 2367187 : 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 2367187 : if (finishing_seg)
2499 : {
2500 2052 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2501 :
2502 : /* signal that we need to wakeup walsenders later */
2503 2052 : WalSndWakeupRequest();
2504 :
2505 2052 : LogwrtResult.Flush = LogwrtResult.Write; /* end of page */
2506 :
2507 2052 : if (XLogArchivingActive())
2508 412 : XLogArchiveNotifySeg(openLogSegNo, tli);
2509 :
2510 2052 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
2511 2052 : 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 2052 : if (IsUnderPostmaster && XLogCheckpointNeeded(openLogSegNo))
2521 : {
2522 273 : (void) GetRedoRecPtr();
2523 273 : if (XLogCheckpointNeeded(openLogSegNo))
2524 223 : RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
2525 : }
2526 : }
2527 : }
2528 :
2529 2923570 : if (ispartialpage)
2530 : {
2531 : /* Only asked to write a partial page */
2532 156900 : LogwrtResult.Write = WriteRqst.Write;
2533 156900 : break;
2534 : }
2535 2766670 : curridx = NextBufIdx(curridx);
2536 :
2537 : /* If flexible, break out of loop as soon as we wrote something */
2538 2766670 : if (flexible && npages == 0)
2539 3884 : break;
2540 : }
2541 :
2542 : Assert(npages == 0);
2543 :
2544 : /*
2545 : * If asked to flush, do so
2546 : */
2547 2350077 : if (LogwrtResult.Flush < WriteRqst.Flush &&
2548 164928 : 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 164861 : if (wal_sync_method != WAL_SYNC_METHOD_OPEN &&
2556 164861 : wal_sync_method != WAL_SYNC_METHOD_OPEN_DSYNC)
2557 : {
2558 164861 : if (openLogFile >= 0 &&
2559 164842 : !XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
2560 : wal_segment_size))
2561 110 : XLogFileClose();
2562 164861 : if (openLogFile < 0)
2563 : {
2564 129 : XLByteToPrevSeg(LogwrtResult.Write, openLogSegNo,
2565 : wal_segment_size);
2566 129 : openLogTLI = tli;
2567 129 : openLogFile = XLogFileOpen(openLogSegNo, tli);
2568 129 : ReserveExternalFD();
2569 : }
2570 :
2571 164861 : issue_xlog_fsync(openLogFile, openLogSegNo, tli);
2572 : }
2573 :
2574 : /* signal that we need to wakeup walsenders later */
2575 164861 : WalSndWakeupRequest();
2576 :
2577 164861 : 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 2350077 : SpinLockAcquire(&XLogCtl->info_lck);
2588 2350077 : if (XLogCtl->LogwrtRqst.Write < LogwrtResult.Write)
2589 145206 : XLogCtl->LogwrtRqst.Write = LogwrtResult.Write;
2590 2350077 : if (XLogCtl->LogwrtRqst.Flush < LogwrtResult.Flush)
2591 166482 : XLogCtl->LogwrtRqst.Flush = LogwrtResult.Flush;
2592 2350077 : 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 2350077 : pg_atomic_write_u64(&XLogCtl->logWriteResult, LogwrtResult.Write);
2600 2350077 : pg_write_barrier();
2601 2350077 : 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 2350077 : }
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 61591 : XLogSetAsyncXactLSN(XLogRecPtr asyncXactLSN)
2631 : {
2632 61591 : XLogRecPtr WriteRqstPtr = asyncXactLSN;
2633 : bool sleeping;
2634 61591 : bool wakeup = false;
2635 : XLogRecPtr prevAsyncXactLSN;
2636 :
2637 61591 : SpinLockAcquire(&XLogCtl->info_lck);
2638 61591 : sleeping = XLogCtl->WalWriterSleeping;
2639 61591 : prevAsyncXactLSN = XLogCtl->asyncXactLSN;
2640 61591 : if (XLogCtl->asyncXactLSN < asyncXactLSN)
2641 61031 : XLogCtl->asyncXactLSN = asyncXactLSN;
2642 61591 : 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 61591 : if (asyncXactLSN <= prevAsyncXactLSN)
2649 560 : 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 61031 : if (sleeping)
2658 45 : wakeup = true;
2659 : else
2660 : {
2661 : int flushblocks;
2662 :
2663 60986 : RefreshXLogWriteResult(LogwrtResult);
2664 :
2665 60986 : flushblocks =
2666 60986 : WriteRqstPtr / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
2667 :
2668 60986 : if (WalWriterFlushAfter == 0 || flushblocks >= WalWriterFlushAfter)
2669 5079 : wakeup = true;
2670 : }
2671 :
2672 61031 : if (wakeup)
2673 : {
2674 5124 : volatile PROC_HDR *procglobal = ProcGlobal;
2675 5124 : ProcNumber walwriterProc = procglobal->walwriterProc;
2676 :
2677 5124 : if (walwriterProc != INVALID_PROC_NUMBER)
2678 781 : 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 42558 : XLogSetReplicationSlotMinimumLSN(XLogRecPtr lsn)
2688 : {
2689 42558 : SpinLockAcquire(&XLogCtl->info_lck);
2690 42558 : XLogCtl->replicationSlotMinLSN = lsn;
2691 42558 : SpinLockRelease(&XLogCtl->info_lck);
2692 42558 : }
2693 :
2694 :
2695 : /*
2696 : * Return the oldest LSN we must retain to satisfy the needs of some
2697 : * replication slot.
2698 : */
2699 : XLogRecPtr
2700 2545 : XLogGetReplicationSlotMinimumLSN(void)
2701 : {
2702 : XLogRecPtr retval;
2703 :
2704 2545 : SpinLockAcquire(&XLogCtl->info_lck);
2705 2545 : retval = XLogCtl->replicationSlotMinLSN;
2706 2545 : SpinLockRelease(&XLogCtl->info_lck);
2707 :
2708 2545 : 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 122360 : UpdateMinRecoveryPoint(XLogRecPtr lsn, bool force)
2722 : {
2723 : /* Quick check using our local copy of the variable */
2724 122360 : if (!updateMinRecoveryPoint || (!force && lsn <= LocalMinRecoveryPoint))
2725 114985 : 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 7375 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint) && InRecovery)
2739 : {
2740 32 : updateMinRecoveryPoint = false;
2741 32 : return;
2742 : }
2743 :
2744 7343 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2745 :
2746 : /* update local copy */
2747 7343 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
2748 7343 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
2749 :
2750 7343 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
2751 1 : updateMinRecoveryPoint = false;
2752 7342 : 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 5884 : newMinRecoveryPoint = GetCurrentReplayRecPtr(&newMinRecoveryPointTLI);
2771 5884 : 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 5884 : if (ControlFile->minRecoveryPoint < newMinRecoveryPoint)
2778 : {
2779 5530 : ControlFile->minRecoveryPoint = newMinRecoveryPoint;
2780 5530 : ControlFile->minRecoveryPointTLI = newMinRecoveryPointTLI;
2781 5530 : UpdateControlFile();
2782 5530 : LocalMinRecoveryPoint = newMinRecoveryPoint;
2783 5530 : LocalMinRecoveryPointTLI = newMinRecoveryPointTLI;
2784 :
2785 5530 : 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 7343 : 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 865060 : XLogFlush(XLogRecPtr record)
2802 : {
2803 : XLogRecPtr WriteRqstPtr;
2804 : XLogwrtRqst WriteRqst;
2805 865060 : 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 865060 : if (!XLogInsertAllowed())
2815 : {
2816 121901 : UpdateMinRecoveryPoint(record, false);
2817 688319 : return;
2818 : }
2819 :
2820 : /* Quick exit if already known flushed */
2821 743159 : if (record <= LogwrtResult.Flush)
2822 566418 : 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 176741 : 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 176741 : 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 3683 : {
2851 : XLogRecPtr insertpos;
2852 :
2853 : /* done already? */
2854 180424 : RefreshXLogWriteResult(LogwrtResult);
2855 180424 : if (record <= LogwrtResult.Flush)
2856 13343 : 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 167081 : SpinLockAcquire(&XLogCtl->info_lck);
2863 167081 : if (WriteRqstPtr < XLogCtl->LogwrtRqst.Write)
2864 11886 : WriteRqstPtr = XLogCtl->LogwrtRqst.Write;
2865 167081 : SpinLockRelease(&XLogCtl->info_lck);
2866 167081 : 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 167081 : 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 3683 : continue;
2883 : }
2884 :
2885 : /* Got the lock; recheck whether request is satisfied */
2886 163398 : RefreshXLogWriteResult(LogwrtResult);
2887 163398 : if (record <= LogwrtResult.Flush)
2888 : {
2889 2443 : LWLockRelease(WALWriteLock);
2890 2443 : 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 160955 : 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 160955 : WriteRqst.Write = insertpos;
2924 160955 : WriteRqst.Flush = insertpos;
2925 :
2926 160955 : XLogWrite(WriteRqst, insertTLI, false);
2927 :
2928 160955 : LWLockRelease(WALWriteLock);
2929 : /* done */
2930 160955 : break;
2931 : }
2932 :
2933 176741 : END_CRIT_SECTION();
2934 :
2935 : /* wake up walsenders now that we've released heavily contended locks */
2936 176741 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
2937 :
2938 : /*
2939 : * If we flushed an LSN that someone was waiting for, notify the waiters.
2940 : */
2941 353482 : if (waitLSNState &&
2942 176741 : (LogwrtResult.Flush >=
2943 176741 : pg_atomic_read_u64(&waitLSNState->minWaitedLSN[WAIT_LSN_TYPE_PRIMARY_FLUSH])))
2944 9 : 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 176741 : 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 15712 : XLogBackgroundFlush(void)
3007 : {
3008 : XLogwrtRqst WriteRqst;
3009 15712 : 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 15712 : 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 15712 : insertTLI = XLogCtl->InsertTimeLineID;
3024 :
3025 : /* read updated LogwrtRqst */
3026 15712 : SpinLockAcquire(&XLogCtl->info_lck);
3027 15712 : WriteRqst = XLogCtl->LogwrtRqst;
3028 15712 : SpinLockRelease(&XLogCtl->info_lck);
3029 :
3030 : /* back off to last completed page boundary */
3031 15712 : WriteRqst.Write -= WriteRqst.Write % XLOG_BLCKSZ;
3032 :
3033 : /* if we have already flushed that far, consider async commit records */
3034 15712 : RefreshXLogWriteResult(LogwrtResult);
3035 15712 : if (WriteRqst.Write <= LogwrtResult.Flush)
3036 : {
3037 11700 : SpinLockAcquire(&XLogCtl->info_lck);
3038 11700 : WriteRqst.Write = XLogCtl->asyncXactLSN;
3039 11700 : SpinLockRelease(&XLogCtl->info_lck);
3040 11700 : 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 15712 : if (WriteRqst.Write <= LogwrtResult.Flush)
3049 : {
3050 10855 : if (openLogFile >= 0)
3051 : {
3052 6696 : if (!XLByteInPrevSeg(LogwrtResult.Write, openLogSegNo,
3053 : wal_segment_size))
3054 : {
3055 175 : XLogFileClose();
3056 : }
3057 : }
3058 10855 : 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 4857 : now = GetCurrentTimestamp();
3070 4857 : flushblocks =
3071 4857 : WriteRqst.Write / XLOG_BLCKSZ - LogwrtResult.Flush / XLOG_BLCKSZ;
3072 :
3073 4857 : if (WalWriterFlushAfter == 0 || lastflush == 0)
3074 : {
3075 : /* first call, or block based limits disabled */
3076 285 : WriteRqst.Flush = WriteRqst.Write;
3077 285 : lastflush = now;
3078 : }
3079 4572 : 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 4268 : WriteRqst.Flush = WriteRqst.Write;
3087 4268 : lastflush = now;
3088 : }
3089 304 : else if (flushblocks >= WalWriterFlushAfter)
3090 : {
3091 : /* exceeded wal_writer_flush_after blocks, flush */
3092 252 : WriteRqst.Flush = WriteRqst.Write;
3093 252 : lastflush = now;
3094 : }
3095 : else
3096 : {
3097 : /* no flushing, this time round */
3098 52 : 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 4857 : START_CRIT_SECTION();
3111 :
3112 : /* now wait for any in-progress insertions to finish and get write lock */
3113 4857 : WaitXLogInsertionsToFinish(WriteRqst.Write);
3114 4857 : LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
3115 4857 : RefreshXLogWriteResult(LogwrtResult);
3116 4857 : if (WriteRqst.Write > LogwrtResult.Write ||
3117 139 : WriteRqst.Flush > LogwrtResult.Flush)
3118 : {
3119 4780 : XLogWrite(WriteRqst, insertTLI, flexible);
3120 : }
3121 4857 : LWLockRelease(WALWriteLock);
3122 :
3123 4857 : END_CRIT_SECTION();
3124 :
3125 : /* wake up walsenders now that we've released heavily contended locks */
3126 4857 : WalSndWakeupProcessRequests(true, !RecoveryInProgress());
3127 :
3128 : /*
3129 : * If we flushed an LSN that someone was waiting for, notify the waiters.
3130 : */
3131 9714 : if (waitLSNState &&
3132 4857 : (LogwrtResult.Flush >=
3133 4857 : pg_atomic_read_u64(&waitLSNState->minWaitedLSN[WAIT_LSN_TYPE_PRIMARY_FLUSH])))
3134 0 : 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 4857 : 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 4857 : 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 16187133 : 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 16187133 : if (!XLogInsertAllowed())
3175 : {
3176 : /* Quick exit if already known to be updated or cannot be updated */
3177 526211 : if (!updateMinRecoveryPoint || record <= LocalMinRecoveryPoint)
3178 510259 : 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 15952 : 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 15952 : if (!LWLockConditionalAcquire(ControlFileLock, LW_SHARED))
3199 0 : return true;
3200 15952 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
3201 15952 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
3202 15952 : 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 15952 : if (!XLogRecPtrIsValid(LocalMinRecoveryPoint))
3210 0 : updateMinRecoveryPoint = false;
3211 :
3212 : /* check again */
3213 15952 : if (record <= LocalMinRecoveryPoint || !updateMinRecoveryPoint)
3214 100 : return false;
3215 : else
3216 15852 : return true;
3217 : }
3218 :
3219 : /* Quick exit if already known flushed */
3220 15660922 : if (record <= LogwrtResult.Flush)
3221 15460757 : return false;
3222 :
3223 : /* read LogwrtResult and update local state */
3224 200165 : RefreshXLogWriteResult(LogwrtResult);
3225 :
3226 : /* check again */
3227 200165 : if (record <= LogwrtResult.Flush)
3228 2908 : return false;
3229 :
3230 197257 : 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 16306 : 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 16306 : int open_flags = O_RDWR | O_CREAT | O_EXCL | PG_BINARY;
3256 : instr_time io_start;
3257 :
3258 : Assert(logtli != 0);
3259 :
3260 16306 : XLogFilePath(path, logtli, logsegno, wal_segment_size);
3261 :
3262 : /*
3263 : * Try to use existent file (checkpoint maker may have created it already)
3264 : */
3265 16306 : *added = false;
3266 16306 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3267 16306 : get_sync_bit(wal_sync_method));
3268 16306 : if (fd < 0)
3269 : {
3270 1443 : 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 14863 : 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 1443 : elog(DEBUG2, "creating and filling new WAL file");
3285 :
3286 1443 : snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3287 :
3288 1443 : unlink(tmppath);
3289 :
3290 1443 : 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 1443 : fd = BasicOpenFile(tmppath, open_flags);
3295 1443 : 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 1443 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3302 :
3303 1443 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_WRITE);
3304 1443 : save_errno = 0;
3305 1443 : 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 1443 : rc = pg_pwrite_zeros(fd, wal_segment_size, 0);
3319 :
3320 1443 : 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 1443 : 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 1443 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT, IOOP_WRITE,
3343 : io_start, 1,
3344 1443 : wal_init_zero ? wal_segment_size : 1);
3345 :
3346 1443 : 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 1443 : io_start = pgstat_prepare_io_time(track_wal_io_timing);
3364 :
3365 1443 : pgstat_report_wait_start(WAIT_EVENT_WAL_INIT_SYNC);
3366 1443 : 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 1443 : pgstat_report_wait_end();
3376 :
3377 1443 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_INIT,
3378 : IOOP_FSYNC, io_start, 1, 0);
3379 :
3380 1443 : 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 1443 : 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 1443 : max_segno = logsegno + CheckPointSegments;
3402 1443 : if (InstallXLogFileSegment(&installed_segno, tmppath, true, max_segno,
3403 : logtli))
3404 : {
3405 1443 : *added = true;
3406 1443 : 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 1443 : 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 16081 : XLogFileInit(XLogSegNo logsegno, TimeLineID logtli)
3436 : {
3437 : bool ignore_added;
3438 : char path[MAXPGPATH];
3439 : int fd;
3440 :
3441 : Assert(logtli != 0);
3442 :
3443 16081 : fd = XLogFileInitInternal(logsegno, logtli, &ignore_added, path);
3444 16081 : if (fd >= 0)
3445 14700 : return fd;
3446 :
3447 : /* Now open original target segment (might not be file I just made) */
3448 1381 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3449 1381 : get_sync_bit(wal_sync_method));
3450 1381 : if (fd < 0)
3451 0 : ereport(ERROR,
3452 : (errcode_for_file_access(),
3453 : errmsg("could not open file \"%s\": %m", path)));
3454 1381 : 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 2802 : if (nread > sizeof(buffer))
3529 2759 : nread = sizeof(buffer);
3530 2802 : pgstat_report_wait_start(WAIT_EVENT_WAL_COPY_READ);
3531 2802 : r = read(srcfd, buffer.data, nread);
3532 2802 : 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 2802 : 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 3236 : 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 3236 : XLogFilePath(path, tli, *segno, wal_segment_size);
3627 :
3628 3236 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
3629 3236 : if (!XLogCtl->InstallXLogFileSegmentActive)
3630 : {
3631 0 : LWLockRelease(ControlFileLock);
3632 0 : return false;
3633 : }
3634 :
3635 3236 : 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 4665 : while (stat(path, &stat_buf) == 0)
3644 : {
3645 1645 : if ((*segno) >= max_segno)
3646 : {
3647 : /* Failed to find a free slot within specified range */
3648 173 : LWLockRelease(ControlFileLock);
3649 173 : return false;
3650 : }
3651 1472 : (*segno)++;
3652 1472 : XLogFilePath(path, tli, *segno, wal_segment_size);
3653 : }
3654 : }
3655 :
3656 : Assert(access(path, F_OK) != 0 && errno == ENOENT);
3657 3063 : 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 3063 : LWLockRelease(ControlFileLock);
3665 :
3666 3063 : return true;
3667 : }
3668 :
3669 : /*
3670 : * Open a pre-existing logfile segment for writing.
3671 : */
3672 : int
3673 129 : XLogFileOpen(XLogSegNo segno, TimeLineID tli)
3674 : {
3675 : char path[MAXPGPATH];
3676 : int fd;
3677 :
3678 129 : XLogFilePath(path, tli, segno, wal_segment_size);
3679 :
3680 129 : fd = BasicOpenFile(path, O_RDWR | PG_BINARY | O_CLOEXEC |
3681 129 : get_sync_bit(wal_sync_method));
3682 129 : if (fd < 0)
3683 0 : ereport(PANIC,
3684 : (errcode_for_file_access(),
3685 : errmsg("could not open file \"%s\": %m", path)));
3686 :
3687 129 : return fd;
3688 : }
3689 :
3690 : /*
3691 : * Close the current logfile segment for writing.
3692 : */
3693 : static void
3694 7125 : 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 7125 : if (!XLogIsNeeded() && (io_direct_flags & IO_DIRECT_WAL) == 0)
3706 250 : (void) posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
3707 : #endif
3708 :
3709 7125 : 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 7125 : openLogFile = -1;
3722 7125 : ReleaseExternalFD();
3723 7125 : }
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 2209 : PreallocXlogFiles(XLogRecPtr endptr, TimeLineID tli)
3746 : {
3747 : XLogSegNo _logSegNo;
3748 : int lf;
3749 : bool added;
3750 : char path[MAXPGPATH];
3751 : uint64 offset;
3752 :
3753 2209 : if (!XLogCtl->InstallXLogFileSegmentActive)
3754 11 : return; /* unlocked check says no */
3755 :
3756 2198 : XLByteToPrevSeg(endptr, _logSegNo, wal_segment_size);
3757 2198 : offset = XLogSegmentOffset(endptr - 1, wal_segment_size);
3758 2198 : if (offset >= (uint32) (0.75 * wal_segment_size))
3759 : {
3760 225 : _logSegNo++;
3761 225 : lf = XLogFileInitInternal(_logSegNo, tli, &added, path);
3762 225 : if (lf >= 0)
3763 163 : close(lf);
3764 225 : if (added)
3765 62 : 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 127743 : CheckXLogRemoved(XLogSegNo segno, TimeLineID tli)
3783 : {
3784 127743 : int save_errno = errno;
3785 : XLogSegNo lastRemovedSegNo;
3786 :
3787 127743 : SpinLockAcquire(&XLogCtl->info_lck);
3788 127743 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3789 127743 : SpinLockRelease(&XLogCtl->info_lck);
3790 :
3791 127743 : 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 127743 : errno = save_errno;
3803 127743 : }
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 1264 : XLogGetLastRemovedSegno(void)
3814 : {
3815 : XLogSegNo lastRemovedSegNo;
3816 :
3817 1264 : SpinLockAcquire(&XLogCtl->info_lck);
3818 1264 : lastRemovedSegNo = XLogCtl->lastRemovedSegNo;
3819 1264 : SpinLockRelease(&XLogCtl->info_lck);
3820 :
3821 1264 : 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 2779 : UpdateLastRemovedPtr(char *filename)
3868 : {
3869 : uint32 tli;
3870 : XLogSegNo segno;
3871 :
3872 2779 : XLogFromFileName(filename, &tli, &segno, wal_segment_size);
3873 :
3874 2779 : SpinLockAcquire(&XLogCtl->info_lck);
3875 2779 : if (segno > XLogCtl->lastRemovedSegNo)
3876 1265 : XLogCtl->lastRemovedSegNo = segno;
3877 2779 : SpinLockRelease(&XLogCtl->info_lck);
3878 2779 : }
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 1931 : 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 1931 : XLByteToSeg(endptr, endlogSegNo, wal_segment_size);
3931 1931 : 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 1931 : XLogFileName(lastoff, 0, segno, wal_segment_size);
3939 :
3940 1931 : elog(DEBUG2, "attempting to remove WAL segments older than log file %s",
3941 : lastoff);
3942 :
3943 1931 : xldir = AllocateDir(XLOGDIR);
3944 :
3945 29237 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
3946 : {
3947 : /* Ignore files that are not XLOG segments */
3948 27306 : if (!IsXLogFileName(xlde->d_name) &&
3949 8171 : !IsPartialXLogFileName(xlde->d_name))
3950 8169 : 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 19137 : if (strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
3964 : {
3965 12241 : if (XLogArchiveCheckDone(xlde->d_name))
3966 : {
3967 : /* Update the last removed location in shared memory first */
3968 2779 : UpdateLastRemovedPtr(xlde->d_name);
3969 :
3970 2779 : RemoveXlogFile(xlde, recycleSegNo, &endlogSegNo, insertTLI);
3971 : }
3972 : }
3973 : }
3974 :
3975 1931 : FreeDir(xldir);
3976 1931 : }
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 67 : 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 67 : XLByteToPrevSeg(switchpoint, switchLogSegNo, wal_segment_size);
4009 67 : XLByteToSeg(switchpoint, endLogSegNo, wal_segment_size);
4010 67 : recycleSegNo = endLogSegNo + 10;
4011 :
4012 : /*
4013 : * Construct a filename of the last segment to be kept.
4014 : */
4015 67 : XLogFileName(switchseg, newTLI, switchLogSegNo, wal_segment_size);
4016 :
4017 67 : elog(DEBUG2, "attempting to remove WAL segments newer than log file %s",
4018 : switchseg);
4019 :
4020 67 : xldir = AllocateDir(XLOGDIR);
4021 :
4022 636 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4023 : {
4024 : /* Ignore files that are not XLOG segments */
4025 569 : if (!IsXLogFileName(xlde->d_name))
4026 353 : 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 216 : if (strncmp(xlde->d_name, switchseg, 8) < 0 &&
4034 140 : 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 17 : if (!XLogArchiveIsReady(xlde->d_name))
4043 17 : RemoveXlogFile(xlde, recycleSegNo, &endLogSegNo, newTLI);
4044 : }
4045 : }
4046 :
4047 67 : FreeDir(xldir);
4048 67 : }
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 2796 : 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 2796 : const char *segname = segment_de->d_name;
4073 :
4074 2796 : 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 2796 : if (wal_recycle &&
4082 2796 : *endlogSegNo <= recycleSegNo &&
4083 3828 : XLogCtl->InstallXLogFileSegmentActive && /* callee rechecks this */
4084 3500 : get_dirent_type(path, segment_de, false, DEBUG2) == PGFILETYPE_REG &&
4085 1750 : InstallXLogFileSegment(endlogSegNo, path,
4086 : true, recycleSegNo, insertTLI))
4087 : {
4088 1577 : ereport(DEBUG2,
4089 : (errmsg_internal("recycled write-ahead log file \"%s\"",
4090 : segname)));
4091 1577 : CheckpointStats.ckpt_segs_recycled++;
4092 : /* Needn't recheck that slot on future iterations */
4093 1577 : (*endlogSegNo)++;
4094 : }
4095 : else
4096 : {
4097 : /* No need for any more future segments, or recycling failed ... */
4098 : int rc;
4099 :
4100 1219 : 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 1219 : rc = durable_unlink(path, LOG);
4128 : #endif
4129 1219 : if (rc != 0)
4130 : {
4131 : /* Message already logged by durable_unlink() */
4132 0 : return;
4133 : }
4134 1219 : CheckpointStats.ckpt_segs_removed++;
4135 : }
4136 :
4137 2796 : 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 1074 : 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 1074 : if (stat(XLOGDIR, &stat_buf) != 0 ||
4161 1074 : !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 1074 : snprintf(path, MAXPGPATH, XLOGDIR "/archive_status");
4169 1074 : if (stat(path, &stat_buf) == 0)
4170 : {
4171 : /* Check for weird cases where it exists but isn't a directory */
4172 1073 : 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 1074 : snprintf(path, MAXPGPATH, XLOGDIR "/summaries");
4191 1074 : if (stat(path, &stat_buf) == 0)
4192 : {
4193 : /* Check for weird cases where it exists but isn't a directory */
4194 1073 : 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 1074 : }
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 1669 : while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
4225 : {
4226 1345 : if (IsBackupHistoryFileName(xlde->d_name))
4227 : {
4228 172 : 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 1133 : 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 1133 : fd = BasicOpenFile(XLOG_CONTROL_FILE,
4421 : O_RDWR | PG_BINARY);
4422 1133 : 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 1133 : pgstat_report_wait_start(WAIT_EVENT_CONTROL_FILE_READ);
4429 1133 : r = read(fd, ControlFile, sizeof(ControlFileData));
4430 1133 : 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 1133 : pgstat_report_wait_end();
4444 :
4445 1133 : 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 1133 : 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 1133 : 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 1133 : INIT_CRC32C(crc);
4475 1133 : COMP_CRC32C(crc,
4476 : ControlFile,
4477 : offsetof(ControlFileData, crc));
4478 1133 : FIN_CRC32C(crc);
4479 :
4480 1133 : 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 1133 : 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 1133 : 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 1133 : 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 1133 : 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 1133 : 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 1133 : 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 1133 : 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 1133 : 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 1133 : 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 1133 : 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 1133 : 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 1133 : wal_segment_size = ControlFile->xlog_seg_size;
4600 :
4601 1133 : 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 1133 : snprintf(wal_segsz_str, sizeof(wal_segsz_str), "%d", wal_segment_size);
4610 1133 : 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 1133 : 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 1133 : 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 1133 : UsableBytesInSegment =
4627 1133 : (wal_segment_size / XLOG_BLCKSZ * UsableBytesInPage) -
4628 : (SizeOfXLogLongPHD - SizeOfXLogShortPHD);
4629 :
4630 1133 : CalculateCheckpointSegments();
4631 1133 : }
4632 :
4633 : /*
4634 : * Utility wrapper to update the control file. Note that the control
4635 : * file gets flushed.
4636 : */
4637 : static void
4638 10163 : UpdateControlFile(void)
4639 : {
4640 10163 : update_controlfile(DataDir, ControlFile, true);
4641 10163 : }
4642 :
4643 : /*
4644 : * Returns the unique system identifier from control file.
4645 : */
4646 : uint64
4647 1564 : GetSystemIdentifier(void)
4648 : {
4649 : Assert(ControlFile != NULL);
4650 1564 : 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 852354 : DataChecksumsNeedWrite(void)
4679 : {
4680 942820 : return (LocalDataChecksumState == PG_DATA_CHECKSUM_VERSION ||
4681 907664 : LocalDataChecksumState == PG_DATA_CHECKSUM_INPROGRESS_ON ||
4682 55310 : LocalDataChecksumState == PG_DATA_CHECKSUM_INPROGRESS_OFF);
4683 : }
4684 :
4685 : bool
4686 167 : DataChecksumsInProgressOn(void)
4687 : {
4688 167 : return LocalDataChecksumState == PG_DATA_CHECKSUM_INPROGRESS_ON;
4689 : }
4690 :
4691 : /*
4692 : * DataChecksumsNeedVerify
4693 : * Returns whether data checksums must be verified or not
4694 : *
4695 : * Data checksums are only verified if they are fully enabled in the cluster.
4696 : * During the "inprogress-on" and "inprogress-off" states they are only
4697 : * updated, not verified (see datachecksum_state.c for a longer discussion).
4698 : *
4699 : * This function is intended for callsites which have read data and are about
4700 : * to perform checksum validation based on the result of this. Calling this
4701 : * function must be performed as close to the validation call as possible to
4702 : * keep the critical section short. This is in order to protect against time of
4703 : * check/time of use situations around data checksum validation.
4704 : */
4705 : bool
4706 2641443 : DataChecksumsNeedVerify(void)
4707 : {
4708 2641443 : return (LocalDataChecksumState == PG_DATA_CHECKSUM_VERSION);
4709 : }
4710 :
4711 : /*
4712 : * SetDataChecksumsOnInProgress
4713 : * Sets the data checksum state to "inprogress-on" to enable checksums
4714 : *
4715 : * To start the process of enabling data checksums in a running cluster the
4716 : * data_checksum_version state must be changed to "inprogress-on". See
4717 : * SetDataChecksumsOn below for a description on how this state change works.
4718 : * This function blocks until all backends in the cluster have acknowledged the
4719 : * state transition.
4720 : */
4721 : void
4722 8 : SetDataChecksumsOnInProgress(void)
4723 : {
4724 : uint64 barrier;
4725 :
4726 : Assert(ControlFile != NULL);
4727 :
4728 : /*
4729 : * The state transition is performed in a critical section with
4730 : * checkpoints held off to provide crash safety.
4731 : */
4732 8 : START_CRIT_SECTION();
4733 8 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
4734 :
4735 8 : XLogChecksums(PG_DATA_CHECKSUM_INPROGRESS_ON);
4736 :
4737 8 : SpinLockAcquire(&XLogCtl->info_lck);
4738 8 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_INPROGRESS_ON;
4739 8 : SpinLockRelease(&XLogCtl->info_lck);
4740 :
4741 8 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_ON);
4742 :
4743 8 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
4744 8 : END_CRIT_SECTION();
4745 :
4746 : /*
4747 : * Update the controlfile before waiting since if we have an immediate
4748 : * shutdown while waiting we want to come back up with checksums enabled.
4749 : */
4750 8 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
4751 8 : ControlFile->data_checksum_version = PG_DATA_CHECKSUM_INPROGRESS_ON;
4752 8 : UpdateControlFile();
4753 8 : LWLockRelease(ControlFileLock);
4754 :
4755 : /*
4756 : * Await state change in all backends to ensure that all backends are in
4757 : * "inprogress-on". Once done we know that all backends are writing data
4758 : * checksums.
4759 : */
4760 8 : WaitForProcSignalBarrier(barrier);
4761 8 : }
4762 :
4763 : /*
4764 : * SetDataChecksumsOn
4765 : * Set data checksums state to 'on' cluster-wide
4766 : *
4767 : * Enabling data checksums is performed using two barriers, the first one to
4768 : * set the state to "inprogress-on" (done by SetDataChecksumsOnInProgress())
4769 : * and the second one to set the state to "on" (done here). Below is a short
4770 : * description of the processing, a more detailed write-up can be found in
4771 : * datachecksum_state.c.
4772 : *
4773 : * To start the process of enabling data checksums in a running cluster the
4774 : * data_checksum_version state must be changed to "inprogress-on". This state
4775 : * requires data checksums to be written but not verified. This ensures that
4776 : * all data pages can be checksummed without the risk of false negatives in
4777 : * validation during the process. When all existing pages are guaranteed to
4778 : * have checksums, and all new pages will be initiated with checksums, the
4779 : * state can be changed to "on". Once the state is "on" checksums will be both
4780 : * written and verified.
4781 : *
4782 : * This function blocks until all backends in the cluster have acknowledged the
4783 : * state transition.
4784 : */
4785 : void
4786 6 : SetDataChecksumsOn(void)
4787 : {
4788 : uint64 barrier;
4789 :
4790 : Assert(ControlFile != NULL);
4791 :
4792 6 : SpinLockAcquire(&XLogCtl->info_lck);
4793 :
4794 : /*
4795 : * The only allowed state transition to "on" is from "inprogress-on" since
4796 : * that state ensures that all pages will have data checksums written. No
4797 : * such state transition exists, if it does happen it's likely due to a
4798 : * programmer error.
4799 : */
4800 6 : if (XLogCtl->data_checksum_version != PG_DATA_CHECKSUM_INPROGRESS_ON)
4801 : {
4802 0 : SpinLockRelease(&XLogCtl->info_lck);
4803 0 : elog(WARNING,
4804 : "cannot set data checksums to \"on\", current state is not \"inprogress-on\", disabling");
4805 0 : SetDataChecksumsOff();
4806 0 : return;
4807 : }
4808 :
4809 6 : SpinLockRelease(&XLogCtl->info_lck);
4810 :
4811 6 : INJECTION_POINT("datachecksums-enable-checksums-delay", NULL);
4812 6 : START_CRIT_SECTION();
4813 6 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
4814 :
4815 6 : XLogChecksums(PG_DATA_CHECKSUM_VERSION);
4816 :
4817 6 : SpinLockAcquire(&XLogCtl->info_lck);
4818 6 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_VERSION;
4819 6 : SpinLockRelease(&XLogCtl->info_lck);
4820 :
4821 6 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_ON);
4822 :
4823 6 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
4824 6 : END_CRIT_SECTION();
4825 :
4826 : /*
4827 : * Update the controlfile before waiting since if we have an immediate
4828 : * shutdown while waiting we want to come back up with checksums enabled.
4829 : */
4830 6 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
4831 6 : ControlFile->data_checksum_version = PG_DATA_CHECKSUM_VERSION;
4832 6 : UpdateControlFile();
4833 6 : LWLockRelease(ControlFileLock);
4834 :
4835 6 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT | CHECKPOINT_FAST);
4836 :
4837 : /*
4838 : * Await state transition to "on" in all backends. When done we know that
4839 : * data checksums are both written and verified in all backends.
4840 : */
4841 6 : WaitForProcSignalBarrier(barrier);
4842 : }
4843 :
4844 : /*
4845 : * SetDataChecksumsOff
4846 : * Disables data checksums cluster-wide
4847 : *
4848 : * Disabling data checksums must be performed with two sets of barriers, each
4849 : * carrying a different state. The state is first set to "inprogress-off"
4850 : * during which checksums are still written but not verified. This ensures that
4851 : * backends which have yet to observe the state change from "on" won't get
4852 : * validation errors on concurrently modified pages. Once all backends have
4853 : * changed to "inprogress-off", the barrier for moving to "off" can be emitted.
4854 : * This function blocks until all backends in the cluster have acknowledged the
4855 : * state transition.
4856 : */
4857 : void
4858 8 : SetDataChecksumsOff(void)
4859 : {
4860 : uint64 barrier;
4861 :
4862 : Assert(ControlFile != NULL);
4863 :
4864 8 : SpinLockAcquire(&XLogCtl->info_lck);
4865 :
4866 : /* If data checksums are already disabled there is nothing to do */
4867 8 : if (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_OFF)
4868 : {
4869 2 : SpinLockRelease(&XLogCtl->info_lck);
4870 2 : return;
4871 : }
4872 :
4873 : /*
4874 : * If data checksums are currently enabled we first transition to the
4875 : * "inprogress-off" state during which backends continue to write
4876 : * checksums without verifying them. When all backends are in
4877 : * "inprogress-off" the next transition to "off" can be performed, after
4878 : * which all data checksum processing is disabled.
4879 : */
4880 6 : if (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_VERSION)
4881 : {
4882 4 : SpinLockRelease(&XLogCtl->info_lck);
4883 :
4884 4 : START_CRIT_SECTION();
4885 4 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
4886 :
4887 4 : XLogChecksums(PG_DATA_CHECKSUM_INPROGRESS_OFF);
4888 :
4889 4 : SpinLockAcquire(&XLogCtl->info_lck);
4890 4 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_INPROGRESS_OFF;
4891 4 : SpinLockRelease(&XLogCtl->info_lck);
4892 :
4893 4 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_OFF);
4894 :
4895 4 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
4896 4 : END_CRIT_SECTION();
4897 :
4898 4 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
4899 4 : ControlFile->data_checksum_version = PG_DATA_CHECKSUM_OFF;
4900 4 : UpdateControlFile();
4901 4 : LWLockRelease(ControlFileLock);
4902 :
4903 4 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT | CHECKPOINT_FAST);
4904 :
4905 : /*
4906 : * Update local state in all backends to ensure that any backend in
4907 : * "on" state is changed to "inprogress-off".
4908 : */
4909 4 : WaitForProcSignalBarrier(barrier);
4910 :
4911 : /*
4912 : * At this point we know that no backends are verifying data checksums
4913 : * during reading. Next, we can safely move to state "off" to also
4914 : * stop writing checksums.
4915 : */
4916 : }
4917 : else
4918 : {
4919 : /*
4920 : * Ending up here implies that the checksums state is "inprogress-on"
4921 : * or "inprogress-off" and we can transition directly to "off" from
4922 : * there.
4923 : */
4924 2 : SpinLockRelease(&XLogCtl->info_lck);
4925 : }
4926 :
4927 6 : START_CRIT_SECTION();
4928 : /* Ensure that we don't incur a checkpoint during disabling checksums */
4929 6 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
4930 :
4931 6 : XLogChecksums(PG_DATA_CHECKSUM_OFF);
4932 :
4933 6 : SpinLockAcquire(&XLogCtl->info_lck);
4934 6 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_OFF;
4935 6 : SpinLockRelease(&XLogCtl->info_lck);
4936 :
4937 6 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_OFF);
4938 :
4939 6 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
4940 6 : END_CRIT_SECTION();
4941 :
4942 6 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
4943 6 : ControlFile->data_checksum_version = PG_DATA_CHECKSUM_OFF;
4944 6 : UpdateControlFile();
4945 6 : LWLockRelease(ControlFileLock);
4946 :
4947 6 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT | CHECKPOINT_FAST);
4948 :
4949 6 : WaitForProcSignalBarrier(barrier);
4950 : }
4951 :
4952 : /*
4953 : * InitLocalDataChecksumState
4954 : *
4955 : * Set up backend local caches of controldata variables which may change at
4956 : * any point during runtime and thus require special cased locking. So far
4957 : * this only applies to data_checksum_version, but it's intended to be general
4958 : * purpose enough to handle future cases.
4959 : */
4960 : void
4961 24658 : InitLocalDataChecksumState(void)
4962 : {
4963 24658 : SpinLockAcquire(&XLogCtl->info_lck);
4964 24658 : SetLocalDataChecksumState(XLogCtl->data_checksum_version);
4965 24658 : SpinLockRelease(&XLogCtl->info_lck);
4966 24658 : }
4967 :
4968 : void
4969 27253 : SetLocalDataChecksumState(uint32 data_checksum_version)
4970 : {
4971 27253 : LocalDataChecksumState = data_checksum_version;
4972 :
4973 27253 : data_checksums = data_checksum_version;
4974 27253 : }
4975 :
4976 : /* guc hook */
4977 : const char *
4978 0 : show_data_checksums(void)
4979 : {
4980 0 : return get_checksum_state_string(LocalDataChecksumState);
4981 : }
4982 :
4983 : /*
4984 : * Return true if the cluster was initialized on a platform where the
4985 : * default signedness of char is "signed". This function exists for code
4986 : * that deals with pre-v18 data files that store data sorted by the 'char'
4987 : * type on disk (e.g., GIN and GiST indexes). See the comments in
4988 : * WriteControlFile() for details.
4989 : */
4990 : bool
4991 89903 : GetDefaultCharSignedness(void)
4992 : {
4993 89903 : return ControlFile->default_char_signedness;
4994 : }
4995 :
4996 : /*
4997 : * Returns a fake LSN for unlogged relations.
4998 : *
4999 : * Each call generates an LSN that is greater than any previous value
5000 : * returned. The current counter value is saved and restored across clean
5001 : * shutdowns, but like unlogged relations, does not survive a crash. This can
5002 : * be used in lieu of real LSN values returned by XLogInsert, if you need an
5003 : * LSN-like increasing sequence of numbers without writing any WAL.
5004 : */
5005 : XLogRecPtr
5006 201460 : GetFakeLSNForUnloggedRel(void)
5007 : {
5008 201460 : return pg_atomic_fetch_add_u64(&XLogCtl->unloggedLSN, 1);
5009 : }
5010 :
5011 : /*
5012 : * Auto-tune the number of XLOG buffers.
5013 : *
5014 : * The preferred setting for wal_buffers is about 3% of shared_buffers, with
5015 : * a maximum of one XLOG segment (there is little reason to think that more
5016 : * is helpful, at least so long as we force an fsync when switching log files)
5017 : * and a minimum of 8 blocks (which was the default value prior to PostgreSQL
5018 : * 9.1, when auto-tuning was added).
5019 : *
5020 : * This should not be called until NBuffers has received its final value.
5021 : */
5022 : static int
5023 1229 : XLOGChooseNumBuffers(void)
5024 : {
5025 : int xbuffers;
5026 :
5027 1229 : xbuffers = NBuffers / 32;
5028 1229 : if (xbuffers > (wal_segment_size / XLOG_BLCKSZ))
5029 24 : xbuffers = (wal_segment_size / XLOG_BLCKSZ);
5030 1229 : if (xbuffers < 8)
5031 457 : xbuffers = 8;
5032 1229 : return xbuffers;
5033 : }
5034 :
5035 : /*
5036 : * GUC check_hook for wal_buffers
5037 : */
5038 : bool
5039 2504 : check_wal_buffers(int *newval, void **extra, GucSource source)
5040 : {
5041 : /*
5042 : * -1 indicates a request for auto-tune.
5043 : */
5044 2504 : if (*newval == -1)
5045 : {
5046 : /*
5047 : * If we haven't yet changed the boot_val default of -1, just let it
5048 : * be. We'll fix it when XLOGShmemSize is called.
5049 : */
5050 1275 : if (XLOGbuffers == -1)
5051 1275 : return true;
5052 :
5053 : /* Otherwise, substitute the auto-tune value */
5054 0 : *newval = XLOGChooseNumBuffers();
5055 : }
5056 :
5057 : /*
5058 : * We clamp manually-set values to at least 4 blocks. Prior to PostgreSQL
5059 : * 9.1, a minimum of 4 was enforced by guc.c, but since that is no longer
5060 : * the case, we just silently treat such values as a request for the
5061 : * minimum. (We could throw an error instead, but that doesn't seem very
5062 : * helpful.)
5063 : */
5064 1229 : if (*newval < 4)
5065 0 : *newval = 4;
5066 :
5067 1229 : return true;
5068 : }
5069 :
5070 : /*
5071 : * GUC check_hook for wal_consistency_checking
5072 : */
5073 : bool
5074 2263 : check_wal_consistency_checking(char **newval, void **extra, GucSource source)
5075 : {
5076 : char *rawstring;
5077 : List *elemlist;
5078 : ListCell *l;
5079 : bool newwalconsistency[RM_MAX_ID + 1];
5080 :
5081 : /* Initialize the array */
5082 74679 : MemSet(newwalconsistency, 0, (RM_MAX_ID + 1) * sizeof(bool));
5083 :
5084 : /* Need a modifiable copy of string */
5085 2263 : rawstring = pstrdup(*newval);
5086 :
5087 : /* Parse string into list of identifiers */
5088 2263 : if (!SplitIdentifierString(rawstring, ',', &elemlist))
5089 : {
5090 : /* syntax error in list */
5091 0 : GUC_check_errdetail("List syntax is invalid.");
5092 0 : pfree(rawstring);
5093 0 : list_free(elemlist);
5094 0 : return false;
5095 : }
5096 :
5097 2760 : foreach(l, elemlist)
5098 : {
5099 497 : char *tok = (char *) lfirst(l);
5100 : int rmid;
5101 :
5102 : /* Check for 'all'. */
5103 497 : if (pg_strcasecmp(tok, "all") == 0)
5104 : {
5105 127215 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
5106 126720 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL)
5107 4950 : newwalconsistency[rmid] = true;
5108 : }
5109 : else
5110 : {
5111 : /* Check if the token matches any known resource manager. */
5112 2 : bool found = false;
5113 :
5114 36 : for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
5115 : {
5116 54 : if (RmgrIdExists(rmid) && GetRmgr(rmid).rm_mask != NULL &&
5117 18 : pg_strcasecmp(tok, GetRmgr(rmid).rm_name) == 0)
5118 : {
5119 2 : newwalconsistency[rmid] = true;
5120 2 : found = true;
5121 2 : break;
5122 : }
5123 : }
5124 2 : if (!found)
5125 : {
5126 : /*
5127 : * During startup, it might be a not-yet-loaded custom
5128 : * resource manager. Defer checking until
5129 : * InitializeWalConsistencyChecking().
5130 : */
5131 0 : if (!process_shared_preload_libraries_done)
5132 : {
5133 0 : check_wal_consistency_checking_deferred = true;
5134 : }
5135 : else
5136 : {
5137 0 : GUC_check_errdetail("Unrecognized key word: \"%s\".", tok);
5138 0 : pfree(rawstring);
5139 0 : list_free(elemlist);
5140 0 : return false;
5141 : }
5142 : }
5143 : }
5144 : }
5145 :
5146 2263 : pfree(rawstring);
5147 2263 : list_free(elemlist);
5148 :
5149 : /* assign new value */
5150 2263 : *extra = guc_malloc(LOG, (RM_MAX_ID + 1) * sizeof(bool));
5151 2263 : if (!*extra)
5152 0 : return false;
5153 2263 : memcpy(*extra, newwalconsistency, (RM_MAX_ID + 1) * sizeof(bool));
5154 2263 : return true;
5155 : }
5156 :
5157 : /*
5158 : * GUC assign_hook for wal_consistency_checking
5159 : */
5160 : void
5161 2262 : assign_wal_consistency_checking(const char *newval, void *extra)
5162 : {
5163 : /*
5164 : * If some checks were deferred, it's possible that the checks will fail
5165 : * later during InitializeWalConsistencyChecking(). But in that case, the
5166 : * postmaster will exit anyway, so it's safe to proceed with the
5167 : * assignment.
5168 : *
5169 : * Any built-in resource managers specified are assigned immediately,
5170 : * which affects WAL created before shared_preload_libraries are
5171 : * processed. Any custom resource managers specified won't be assigned
5172 : * until after shared_preload_libraries are processed, but that's OK
5173 : * because WAL for a custom resource manager can't be written before the
5174 : * module is loaded anyway.
5175 : */
5176 2262 : wal_consistency_checking = extra;
5177 2262 : }
5178 :
5179 : /*
5180 : * InitializeWalConsistencyChecking: run after loading custom resource managers
5181 : *
5182 : * If any unknown resource managers were specified in the
5183 : * wal_consistency_checking GUC, processing was deferred. Now that
5184 : * shared_preload_libraries have been loaded, process wal_consistency_checking
5185 : * again.
5186 : */
5187 : void
5188 1058 : InitializeWalConsistencyChecking(void)
5189 : {
5190 : Assert(process_shared_preload_libraries_done);
5191 :
5192 1058 : if (check_wal_consistency_checking_deferred)
5193 : {
5194 : struct config_generic *guc;
5195 :
5196 0 : guc = find_option("wal_consistency_checking", false, false, ERROR);
5197 :
5198 0 : check_wal_consistency_checking_deferred = false;
5199 :
5200 0 : set_config_option_ext("wal_consistency_checking",
5201 : wal_consistency_checking_string,
5202 : guc->scontext, guc->source, guc->srole,
5203 : GUC_ACTION_SET, true, ERROR, false);
5204 :
5205 : /* checking should not be deferred again */
5206 : Assert(!check_wal_consistency_checking_deferred);
5207 : }
5208 1058 : }
5209 :
5210 : /*
5211 : * GUC show_hook for archive_command
5212 : */
5213 : const char *
5214 1951 : show_archive_command(void)
5215 : {
5216 1951 : if (XLogArchivingActive())
5217 100 : return XLogArchiveCommand;
5218 : else
5219 1851 : return "(disabled)";
5220 : }
5221 :
5222 : /*
5223 : * GUC show_hook for in_hot_standby
5224 : */
5225 : const char *
5226 17136 : show_in_hot_standby(void)
5227 : {
5228 : /*
5229 : * We display the actual state based on shared memory, so that this GUC
5230 : * reports up-to-date state if examined intra-query. The underlying
5231 : * variable (in_hot_standby_guc) changes only when we transmit a new value
5232 : * to the client.
5233 : */
5234 17136 : return RecoveryInProgress() ? "on" : "off";
5235 : }
5236 :
5237 : /*
5238 : * GUC show_hook for effective_wal_level
5239 : */
5240 : const char *
5241 1984 : show_effective_wal_level(void)
5242 : {
5243 1984 : if (wal_level == WAL_LEVEL_MINIMAL)
5244 306 : return "minimal";
5245 :
5246 : /*
5247 : * During recovery, effective_wal_level reflects the primary's
5248 : * configuration rather than the local wal_level value.
5249 : */
5250 1678 : if (RecoveryInProgress())
5251 27 : return IsXLogLogicalInfoEnabled() ? "logical" : "replica";
5252 :
5253 1651 : return XLogLogicalInfoActive() ? "logical" : "replica";
5254 : }
5255 :
5256 : /*
5257 : * Read the control file, set respective GUCs.
5258 : *
5259 : * This is to be called during startup, including a crash recovery cycle,
5260 : * unless in bootstrap mode, where no control file yet exists. As there's no
5261 : * usable shared memory yet (its sizing can depend on the contents of the
5262 : * control file!), first store the contents in local memory. XLOGShmemInit()
5263 : * will then copy it to shared memory later.
5264 : *
5265 : * reset just controls whether previous contents are to be expected (in the
5266 : * reset case, there's a dangling pointer into old shared memory), or not.
5267 : */
5268 : void
5269 1076 : LocalProcessControlFile(bool reset)
5270 : {
5271 : Assert(reset || ControlFile == NULL);
5272 1076 : LocalControlFile = palloc_object(ControlFileData);
5273 1076 : ControlFile = LocalControlFile;
5274 1076 : ReadControlFile();
5275 1076 : SetLocalDataChecksumState(ControlFile->data_checksum_version);
5276 1076 : }
5277 :
5278 : /*
5279 : * Get the wal_level from the control file. For a standby, this value should be
5280 : * considered as its active wal_level, because it may be different from what
5281 : * was originally configured on standby.
5282 : */
5283 : WalLevel
5284 0 : GetActiveWalLevelOnStandby(void)
5285 : {
5286 0 : return ControlFile->wal_level;
5287 : }
5288 :
5289 : /*
5290 : * Register shared memory for XLOG.
5291 : */
5292 : static void
5293 1234 : XLOGShmemRequest(void *arg)
5294 : {
5295 : Size size;
5296 :
5297 : /*
5298 : * If the value of wal_buffers is -1, use the preferred auto-tune value.
5299 : * This isn't an amazingly clean place to do this, but we must wait till
5300 : * NBuffers has received its final value, and must do it before using the
5301 : * value of XLOGbuffers to do anything important.
5302 : *
5303 : * We prefer to report this value's source as PGC_S_DYNAMIC_DEFAULT.
5304 : * However, if the DBA explicitly set wal_buffers = -1 in the config file,
5305 : * then PGC_S_DYNAMIC_DEFAULT will fail to override that and we must force
5306 : * the matter with PGC_S_OVERRIDE.
5307 : */
5308 1234 : if (XLOGbuffers == -1)
5309 : {
5310 : char buf[32];
5311 :
5312 1229 : snprintf(buf, sizeof(buf), "%d", XLOGChooseNumBuffers());
5313 1229 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
5314 : PGC_S_DYNAMIC_DEFAULT);
5315 1229 : if (XLOGbuffers == -1) /* failed to apply it? */
5316 0 : SetConfigOption("wal_buffers", buf, PGC_POSTMASTER,
5317 : PGC_S_OVERRIDE);
5318 : }
5319 : Assert(XLOGbuffers > 0);
5320 :
5321 : /* XLogCtl */
5322 1234 : size = sizeof(XLogCtlData);
5323 :
5324 : /* WAL insertion locks, plus alignment */
5325 1234 : size = add_size(size, mul_size(sizeof(WALInsertLockPadded), NUM_XLOGINSERT_LOCKS + 1));
5326 : /* xlblocks array */
5327 1234 : size = add_size(size, mul_size(sizeof(pg_atomic_uint64), XLOGbuffers));
5328 : /* extra alignment padding for XLOG I/O buffers */
5329 1234 : size = add_size(size, Max(XLOG_BLCKSZ, PG_IO_ALIGN_SIZE));
5330 : /* and the buffers themselves */
5331 1234 : size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
5332 :
5333 1234 : ShmemRequestStruct(.name = "XLOG Ctl",
5334 : .size = size,
5335 : .ptr = (void **) &XLogCtl,
5336 : );
5337 1234 : ShmemRequestStruct(.name = "Control File",
5338 : .size = sizeof(ControlFileData),
5339 : .ptr = (void **) &ControlFile,
5340 : );
5341 1234 : }
5342 :
5343 : /*
5344 : * XLOGShmemInit - initialize the XLogCtl shared memory area.
5345 : */
5346 : static void
5347 1231 : XLOGShmemInit(void *arg)
5348 : {
5349 : char *allocptr;
5350 : int i;
5351 :
5352 : #ifdef WAL_DEBUG
5353 :
5354 : /*
5355 : * Create a memory context for WAL debugging that's exempt from the normal
5356 : * "no pallocs in critical section" rule. Yes, that can lead to a PANIC if
5357 : * an allocation fails, but wal_debug is not for production use anyway.
5358 : */
5359 : if (walDebugCxt == NULL)
5360 : {
5361 : walDebugCxt = AllocSetContextCreate(TopMemoryContext,
5362 : "WAL Debug",
5363 : ALLOCSET_DEFAULT_SIZES);
5364 : MemoryContextAllowInCriticalSection(walDebugCxt, true);
5365 : }
5366 : #endif
5367 :
5368 1231 : memset(XLogCtl, 0, sizeof(XLogCtlData));
5369 :
5370 : /*
5371 : * Already have read control file locally, unless in bootstrap mode. Move
5372 : * contents into shared memory.
5373 : */
5374 1231 : if (LocalControlFile)
5375 : {
5376 1060 : memcpy(ControlFile, LocalControlFile, sizeof(ControlFileData));
5377 1060 : pfree(LocalControlFile);
5378 1060 : LocalControlFile = NULL;
5379 : }
5380 :
5381 : /*
5382 : * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
5383 : * multiple of the alignment for same, so no extra alignment padding is
5384 : * needed here.
5385 : */
5386 1231 : allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
5387 1231 : XLogCtl->xlblocks = (pg_atomic_uint64 *) allocptr;
5388 1231 : allocptr += sizeof(pg_atomic_uint64) * XLOGbuffers;
5389 :
5390 363030 : for (i = 0; i < XLOGbuffers; i++)
5391 : {
5392 361799 : pg_atomic_init_u64(&XLogCtl->xlblocks[i], InvalidXLogRecPtr);
5393 : }
5394 :
5395 : /* WAL insertion locks. Ensure they're aligned to the full padded size */
5396 1231 : allocptr += sizeof(WALInsertLockPadded) -
5397 1231 : ((uintptr_t) allocptr) % sizeof(WALInsertLockPadded);
5398 1231 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks =
5399 : (WALInsertLockPadded *) allocptr;
5400 1231 : allocptr += sizeof(WALInsertLockPadded) * NUM_XLOGINSERT_LOCKS;
5401 :
5402 11079 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
5403 : {
5404 9848 : LWLockInitialize(&WALInsertLocks[i].l.lock, LWTRANCHE_WAL_INSERT);
5405 9848 : pg_atomic_init_u64(&WALInsertLocks[i].l.insertingAt, InvalidXLogRecPtr);
5406 9848 : WALInsertLocks[i].l.lastImportantAt = InvalidXLogRecPtr;
5407 : }
5408 :
5409 : /*
5410 : * Align the start of the page buffers to a full xlog block size boundary.
5411 : * This simplifies some calculations in XLOG insertion. It is also
5412 : * required for O_DIRECT.
5413 : */
5414 1231 : allocptr = (char *) TYPEALIGN(XLOG_BLCKSZ, allocptr);
5415 1231 : XLogCtl->pages = allocptr;
5416 1231 : memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
5417 :
5418 : /*
5419 : * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
5420 : * in additional info.)
5421 : */
5422 1231 : XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
5423 1231 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
5424 1231 : XLogCtl->InstallXLogFileSegmentActive = false;
5425 1231 : XLogCtl->WalWriterSleeping = false;
5426 :
5427 : /* Use the checksum info from control file */
5428 1231 : XLogCtl->data_checksum_version = ControlFile->data_checksum_version;
5429 :
5430 1231 : SetLocalDataChecksumState(XLogCtl->data_checksum_version);
5431 :
5432 1231 : SpinLockInit(&XLogCtl->Insert.insertpos_lck);
5433 1231 : SpinLockInit(&XLogCtl->info_lck);
5434 1231 : pg_atomic_init_u64(&XLogCtl->logInsertResult, InvalidXLogRecPtr);
5435 1231 : pg_atomic_init_u64(&XLogCtl->logWriteResult, InvalidXLogRecPtr);
5436 1231 : pg_atomic_init_u64(&XLogCtl->logFlushResult, InvalidXLogRecPtr);
5437 1231 : pg_atomic_init_u64(&XLogCtl->unloggedLSN, InvalidXLogRecPtr);
5438 1231 : }
5439 :
5440 : /*
5441 : * XLOGShmemAttach - re-establish WALInsertLocks pointer after attaching.
5442 : */
5443 : static void
5444 0 : XLOGShmemAttach(void *arg)
5445 : {
5446 0 : WALInsertLocks = XLogCtl->Insert.WALInsertLocks;
5447 0 : }
5448 :
5449 : /*
5450 : * This func must be called ONCE on system install. It creates pg_control
5451 : * and the initial XLOG segment.
5452 : */
5453 : void
5454 57 : BootStrapXLOG(uint32 data_checksum_version)
5455 : {
5456 : CheckPoint checkPoint;
5457 : PGAlignedXLogBlock buffer;
5458 : XLogPageHeader page;
5459 : XLogLongPageHeader longpage;
5460 : XLogRecord *record;
5461 : char *recptr;
5462 : uint64 sysidentifier;
5463 : struct timeval tv;
5464 : pg_crc32c crc;
5465 :
5466 : /* allow ordinary WAL segment creation, like StartupXLOG() would */
5467 57 : SetInstallXLogFileSegmentActive();
5468 :
5469 : /*
5470 : * Select a hopefully-unique system identifier code for this installation.
5471 : * We use the result of gettimeofday(), including the fractional seconds
5472 : * field, as being about as unique as we can easily get. (Think not to
5473 : * use random(), since it hasn't been seeded and there's no portable way
5474 : * to seed it other than the system clock value...) The upper half of the
5475 : * uint64 value is just the tv_sec part, while the lower half contains the
5476 : * tv_usec part (which must fit in 20 bits), plus 12 bits from our current
5477 : * PID for a little extra uniqueness. A person knowing this encoding can
5478 : * determine the initialization time of the installation, which could
5479 : * perhaps be useful sometimes.
5480 : */
5481 57 : gettimeofday(&tv, NULL);
5482 57 : sysidentifier = ((uint64) tv.tv_sec) << 32;
5483 57 : sysidentifier |= ((uint64) tv.tv_usec) << 12;
5484 57 : sysidentifier |= getpid() & 0xFFF;
5485 :
5486 57 : memset(&buffer, 0, sizeof buffer);
5487 57 : page = (XLogPageHeader) &buffer;
5488 :
5489 : /*
5490 : * Set up information for the initial checkpoint record
5491 : *
5492 : * The initial checkpoint record is written to the beginning of the WAL
5493 : * segment with logid=0 logseg=1. The very first WAL segment, 0/0, is not
5494 : * used, so that we can use 0/0 to mean "before any valid WAL segment".
5495 : */
5496 57 : checkPoint.redo = wal_segment_size + SizeOfXLogLongPHD;
5497 57 : checkPoint.ThisTimeLineID = BootstrapTimeLineID;
5498 57 : checkPoint.PrevTimeLineID = BootstrapTimeLineID;
5499 57 : checkPoint.fullPageWrites = fullPageWrites;
5500 57 : checkPoint.logicalDecodingEnabled = (wal_level == WAL_LEVEL_LOGICAL);
5501 57 : checkPoint.wal_level = wal_level;
5502 : checkPoint.nextXid =
5503 57 : FullTransactionIdFromEpochAndXid(0, FirstNormalTransactionId);
5504 57 : checkPoint.nextOid = FirstGenbkiObjectId;
5505 57 : checkPoint.nextMulti = FirstMultiXactId;
5506 57 : checkPoint.nextMultiOffset = 1;
5507 57 : checkPoint.oldestXid = FirstNormalTransactionId;
5508 57 : checkPoint.oldestXidDB = Template1DbOid;
5509 57 : checkPoint.oldestMulti = FirstMultiXactId;
5510 57 : checkPoint.oldestMultiDB = Template1DbOid;
5511 57 : checkPoint.oldestCommitTsXid = InvalidTransactionId;
5512 57 : checkPoint.newestCommitTsXid = InvalidTransactionId;
5513 57 : checkPoint.time = (pg_time_t) time(NULL);
5514 57 : checkPoint.oldestActiveXid = InvalidTransactionId;
5515 57 : checkPoint.dataChecksumState = data_checksum_version;
5516 :
5517 57 : TransamVariables->nextXid = checkPoint.nextXid;
5518 57 : TransamVariables->nextOid = checkPoint.nextOid;
5519 57 : TransamVariables->oidCount = 0;
5520 57 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
5521 57 : AdvanceOldestClogXid(checkPoint.oldestXid);
5522 57 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
5523 57 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
5524 57 : SetCommitTsLimit(InvalidTransactionId, InvalidTransactionId);
5525 :
5526 : /* Set up the XLOG page header */
5527 57 : page->xlp_magic = XLOG_PAGE_MAGIC;
5528 57 : page->xlp_info = XLP_LONG_HEADER;
5529 57 : page->xlp_tli = BootstrapTimeLineID;
5530 57 : page->xlp_pageaddr = wal_segment_size;
5531 57 : longpage = (XLogLongPageHeader) page;
5532 57 : longpage->xlp_sysid = sysidentifier;
5533 57 : longpage->xlp_seg_size = wal_segment_size;
5534 57 : longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
5535 :
5536 : /* Insert the initial checkpoint record */
5537 57 : recptr = ((char *) page + SizeOfXLogLongPHD);
5538 57 : record = (XLogRecord *) recptr;
5539 57 : record->xl_prev = InvalidXLogRecPtr;
5540 57 : record->xl_xid = InvalidTransactionId;
5541 57 : record->xl_tot_len = SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(checkPoint);
5542 57 : record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
5543 57 : record->xl_rmid = RM_XLOG_ID;
5544 57 : recptr += SizeOfXLogRecord;
5545 : /* fill the XLogRecordDataHeaderShort struct */
5546 57 : *(recptr++) = (char) XLR_BLOCK_ID_DATA_SHORT;
5547 57 : *(recptr++) = sizeof(checkPoint);
5548 57 : memcpy(recptr, &checkPoint, sizeof(checkPoint));
5549 57 : recptr += sizeof(checkPoint);
5550 : Assert(recptr - (char *) record == record->xl_tot_len);
5551 :
5552 57 : INIT_CRC32C(crc);
5553 57 : COMP_CRC32C(crc, ((char *) record) + SizeOfXLogRecord, record->xl_tot_len - SizeOfXLogRecord);
5554 57 : COMP_CRC32C(crc, (char *) record, offsetof(XLogRecord, xl_crc));
5555 57 : FIN_CRC32C(crc);
5556 57 : record->xl_crc = crc;
5557 :
5558 : /* Create first XLOG segment file */
5559 57 : openLogTLI = BootstrapTimeLineID;
5560 57 : openLogFile = XLogFileInit(1, BootstrapTimeLineID);
5561 :
5562 : /*
5563 : * We needn't bother with Reserve/ReleaseExternalFD here, since we'll
5564 : * close the file again in a moment.
5565 : */
5566 :
5567 : /* Write the first page with the initial record */
5568 57 : errno = 0;
5569 57 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_WRITE);
5570 57 : if (write(openLogFile, &buffer, XLOG_BLCKSZ) != XLOG_BLCKSZ)
5571 : {
5572 : /* if write didn't set errno, assume problem is no disk space */
5573 0 : if (errno == 0)
5574 0 : errno = ENOSPC;
5575 0 : ereport(PANIC,
5576 : (errcode_for_file_access(),
5577 : errmsg("could not write bootstrap write-ahead log file: %m")));
5578 : }
5579 57 : pgstat_report_wait_end();
5580 :
5581 57 : pgstat_report_wait_start(WAIT_EVENT_WAL_BOOTSTRAP_SYNC);
5582 57 : if (pg_fsync(openLogFile) != 0)
5583 0 : ereport(PANIC,
5584 : (errcode_for_file_access(),
5585 : errmsg("could not fsync bootstrap write-ahead log file: %m")));
5586 57 : pgstat_report_wait_end();
5587 :
5588 57 : if (close(openLogFile) != 0)
5589 0 : ereport(PANIC,
5590 : (errcode_for_file_access(),
5591 : errmsg("could not close bootstrap write-ahead log file: %m")));
5592 :
5593 57 : openLogFile = -1;
5594 :
5595 : /* Now create pg_control */
5596 57 : InitControlFile(sysidentifier, data_checksum_version);
5597 57 : ControlFile->time = checkPoint.time;
5598 57 : ControlFile->checkPoint = checkPoint.redo;
5599 57 : ControlFile->checkPointCopy = checkPoint;
5600 :
5601 : /* some additional ControlFile fields are set in WriteControlFile() */
5602 57 : WriteControlFile();
5603 :
5604 : /* Bootstrap the commit log, too */
5605 57 : BootStrapCLOG();
5606 57 : BootStrapCommitTs();
5607 57 : BootStrapSUBTRANS();
5608 57 : BootStrapMultiXact();
5609 :
5610 : /*
5611 : * Force control file to be read - in contrast to normal processing we'd
5612 : * otherwise never run the checks and GUC related initializations therein.
5613 : */
5614 57 : ReadControlFile();
5615 57 : }
5616 :
5617 : static char *
5618 950 : str_time(pg_time_t tnow, char *buf, size_t bufsize)
5619 : {
5620 950 : pg_strftime(buf, bufsize,
5621 : "%Y-%m-%d %H:%M:%S %Z",
5622 950 : pg_localtime(&tnow, log_timezone));
5623 :
5624 950 : return buf;
5625 : }
5626 :
5627 : /*
5628 : * Initialize the first WAL segment on new timeline.
5629 : */
5630 : static void
5631 55 : XLogInitNewTimeline(TimeLineID endTLI, XLogRecPtr endOfLog, TimeLineID newTLI)
5632 : {
5633 : char xlogfname[MAXFNAMELEN];
5634 : XLogSegNo endLogSegNo;
5635 : XLogSegNo startLogSegNo;
5636 :
5637 : /* we always switch to a new timeline after archive recovery */
5638 : Assert(endTLI != newTLI);
5639 :
5640 : /*
5641 : * Update min recovery point one last time.
5642 : */
5643 55 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
5644 :
5645 : /*
5646 : * Calculate the last segment on the old timeline, and the first segment
5647 : * on the new timeline. If the switch happens in the middle of a segment,
5648 : * they are the same, but if the switch happens exactly at a segment
5649 : * boundary, startLogSegNo will be endLogSegNo + 1.
5650 : */
5651 55 : XLByteToPrevSeg(endOfLog, endLogSegNo, wal_segment_size);
5652 55 : XLByteToSeg(endOfLog, startLogSegNo, wal_segment_size);
5653 :
5654 : /*
5655 : * Initialize the starting WAL segment for the new timeline. If the switch
5656 : * happens in the middle of a segment, copy data from the last WAL segment
5657 : * of the old timeline up to the switch point, to the starting WAL segment
5658 : * on the new timeline.
5659 : */
5660 55 : if (endLogSegNo == startLogSegNo)
5661 : {
5662 : /*
5663 : * Make a copy of the file on the new timeline.
5664 : *
5665 : * Writing WAL isn't allowed yet, so there are no locking
5666 : * considerations. But we should be just as tense as XLogFileInit to
5667 : * avoid emplacing a bogus file.
5668 : */
5669 43 : XLogFileCopy(newTLI, endLogSegNo, endTLI, endLogSegNo,
5670 43 : XLogSegmentOffset(endOfLog, wal_segment_size));
5671 : }
5672 : else
5673 : {
5674 : /*
5675 : * The switch happened at a segment boundary, so just create the next
5676 : * segment on the new timeline.
5677 : */
5678 : int fd;
5679 :
5680 12 : fd = XLogFileInit(startLogSegNo, newTLI);
5681 :
5682 12 : if (close(fd) != 0)
5683 : {
5684 0 : int save_errno = errno;
5685 :
5686 0 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5687 0 : errno = save_errno;
5688 0 : ereport(ERROR,
5689 : (errcode_for_file_access(),
5690 : errmsg("could not close file \"%s\": %m", xlogfname)));
5691 : }
5692 : }
5693 :
5694 : /*
5695 : * Let's just make real sure there are not .ready or .done flags posted
5696 : * for the new segment.
5697 : */
5698 55 : XLogFileName(xlogfname, newTLI, startLogSegNo, wal_segment_size);
5699 55 : XLogArchiveCleanup(xlogfname);
5700 55 : }
5701 :
5702 : /*
5703 : * Perform cleanup actions at the conclusion of archive recovery.
5704 : */
5705 : static void
5706 55 : CleanupAfterArchiveRecovery(TimeLineID EndOfLogTLI, XLogRecPtr EndOfLog,
5707 : TimeLineID newTLI)
5708 : {
5709 : /*
5710 : * Execute the recovery_end_command, if any.
5711 : */
5712 55 : if (recoveryEndCommand && strcmp(recoveryEndCommand, "") != 0)
5713 2 : ExecuteRecoveryCommand(recoveryEndCommand,
5714 : "recovery_end_command",
5715 : true,
5716 : WAIT_EVENT_RECOVERY_END_COMMAND);
5717 :
5718 : /*
5719 : * We switched to a new timeline. Clean up segments on the old timeline.
5720 : *
5721 : * If there are any higher-numbered segments on the old timeline, remove
5722 : * them. They might contain valid WAL, but they might also be
5723 : * pre-allocated files containing garbage. In any case, they are not part
5724 : * of the new timeline's history so we don't need them.
5725 : */
5726 55 : RemoveNonParentXlogFiles(EndOfLog, newTLI);
5727 :
5728 : /*
5729 : * If the switch happened in the middle of a segment, what to do with the
5730 : * last, partial segment on the old timeline? If we don't archive it, and
5731 : * the server that created the WAL never archives it either (e.g. because
5732 : * it was hit by a meteor), it will never make it to the archive. That's
5733 : * OK from our point of view, because the new segment that we created with
5734 : * the new TLI contains all the WAL from the old timeline up to the switch
5735 : * point. But if you later try to do PITR to the "missing" WAL on the old
5736 : * timeline, recovery won't find it in the archive. It's physically
5737 : * present in the new file with new TLI, but recovery won't look there
5738 : * when it's recovering to the older timeline. On the other hand, if we
5739 : * archive the partial segment, and the original server on that timeline
5740 : * is still running and archives the completed version of the same segment
5741 : * later, it will fail. (We used to do that in 9.4 and below, and it
5742 : * caused such problems).
5743 : *
5744 : * As a compromise, we rename the last segment with the .partial suffix,
5745 : * and archive it. Archive recovery will never try to read .partial
5746 : * segments, so they will normally go unused. But in the odd PITR case,
5747 : * the administrator can copy them manually to the pg_wal directory
5748 : * (removing the suffix). They can be useful in debugging, too.
5749 : *
5750 : * If a .done or .ready file already exists for the old timeline, however,
5751 : * we had already determined that the segment is complete, so we can let
5752 : * it be archived normally. (In particular, if it was restored from the
5753 : * archive to begin with, it's expected to have a .done file).
5754 : */
5755 55 : if (XLogSegmentOffset(EndOfLog, wal_segment_size) != 0 &&
5756 : XLogArchivingActive())
5757 : {
5758 : char origfname[MAXFNAMELEN];
5759 : XLogSegNo endLogSegNo;
5760 :
5761 9 : XLByteToPrevSeg(EndOfLog, endLogSegNo, wal_segment_size);
5762 9 : XLogFileName(origfname, EndOfLogTLI, endLogSegNo, wal_segment_size);
5763 :
5764 9 : if (!XLogArchiveIsReadyOrDone(origfname))
5765 : {
5766 : char origpath[MAXPGPATH];
5767 : char partialfname[MAXFNAMELEN];
5768 : char partialpath[MAXPGPATH];
5769 :
5770 : /*
5771 : * If we're summarizing WAL, we can't rename the partial file
5772 : * until the summarizer finishes with it, else it will fail.
5773 : */
5774 5 : if (summarize_wal)
5775 1 : WaitForWalSummarization(EndOfLog);
5776 :
5777 5 : XLogFilePath(origpath, EndOfLogTLI, endLogSegNo, wal_segment_size);
5778 5 : snprintf(partialfname, MAXFNAMELEN, "%s.partial", origfname);
5779 5 : snprintf(partialpath, MAXPGPATH, "%s.partial", origpath);
5780 :
5781 : /*
5782 : * Make sure there's no .done or .ready file for the .partial
5783 : * file.
5784 : */
5785 5 : XLogArchiveCleanup(partialfname);
5786 :
5787 5 : durable_rename(origpath, partialpath, ERROR);
5788 5 : XLogArchiveNotify(partialfname);
5789 : }
5790 : }
5791 55 : }
5792 :
5793 : /*
5794 : * Check to see if required parameters are set high enough on this server
5795 : * for various aspects of recovery operation.
5796 : *
5797 : * Note that all the parameters which this function tests need to be
5798 : * listed in Administrator's Overview section in high-availability.sgml.
5799 : * If you change them, don't forget to update the list.
5800 : */
5801 : static void
5802 265 : CheckRequiredParameterValues(void)
5803 : {
5804 : /*
5805 : * For archive recovery, the WAL must be generated with at least 'replica'
5806 : * wal_level.
5807 : */
5808 265 : if (ArchiveRecoveryRequested && ControlFile->wal_level == WAL_LEVEL_MINIMAL)
5809 : {
5810 2 : ereport(FATAL,
5811 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5812 : errmsg("WAL was generated with \"wal_level=minimal\", cannot continue recovering"),
5813 : errdetail("This happens if you temporarily set \"wal_level=minimal\" on the server."),
5814 : errhint("Use a backup taken after setting \"wal_level\" to higher than \"minimal\".")));
5815 : }
5816 :
5817 : /*
5818 : * For Hot Standby, the WAL must be generated with 'replica' mode, and we
5819 : * must have at least as many backend slots as the primary.
5820 : */
5821 263 : if (ArchiveRecoveryRequested && EnableHotStandby)
5822 : {
5823 : /* We ignore autovacuum_worker_slots when we make this test. */
5824 139 : RecoveryRequiresIntParameter("max_connections",
5825 : MaxConnections,
5826 139 : ControlFile->MaxConnections);
5827 139 : RecoveryRequiresIntParameter("max_worker_processes",
5828 : max_worker_processes,
5829 139 : ControlFile->max_worker_processes);
5830 139 : RecoveryRequiresIntParameter("max_wal_senders",
5831 : max_wal_senders,
5832 139 : ControlFile->max_wal_senders);
5833 139 : RecoveryRequiresIntParameter("max_prepared_transactions",
5834 : max_prepared_xacts,
5835 139 : ControlFile->max_prepared_xacts);
5836 139 : RecoveryRequiresIntParameter("max_locks_per_transaction",
5837 : max_locks_per_xact,
5838 139 : ControlFile->max_locks_per_xact);
5839 : }
5840 263 : }
5841 :
5842 : /*
5843 : * This must be called ONCE during postmaster or standalone-backend startup
5844 : */
5845 : void
5846 1074 : StartupXLOG(void)
5847 : {
5848 : XLogCtlInsert *Insert;
5849 : CheckPoint checkPoint;
5850 : bool wasShutdown;
5851 : bool didCrash;
5852 : bool haveTblspcMap;
5853 : bool haveBackupLabel;
5854 : XLogRecPtr EndOfLog;
5855 : TimeLineID EndOfLogTLI;
5856 : TimeLineID newTLI;
5857 : bool performedWalRecovery;
5858 : EndOfWalRecoveryInfo *endOfRecoveryInfo;
5859 : XLogRecPtr abortedRecPtr;
5860 : XLogRecPtr missingContrecPtr;
5861 : TransactionId oldestActiveXID;
5862 1074 : bool promoted = false;
5863 : char timebuf[128];
5864 :
5865 : /*
5866 : * We should have an aux process resource owner to use, and we should not
5867 : * be in a transaction that's installed some other resowner.
5868 : */
5869 : Assert(AuxProcessResourceOwner != NULL);
5870 : Assert(CurrentResourceOwner == NULL ||
5871 : CurrentResourceOwner == AuxProcessResourceOwner);
5872 1074 : CurrentResourceOwner = AuxProcessResourceOwner;
5873 :
5874 : /*
5875 : * Check that contents look valid.
5876 : */
5877 1074 : if (!XRecOffIsValid(ControlFile->checkPoint))
5878 0 : ereport(FATAL,
5879 : (errcode(ERRCODE_DATA_CORRUPTED),
5880 : errmsg("control file contains invalid checkpoint location")));
5881 :
5882 1074 : switch (ControlFile->state)
5883 : {
5884 853 : case DB_SHUTDOWNED:
5885 :
5886 : /*
5887 : * This is the expected case, so don't be chatty in standalone
5888 : * mode
5889 : */
5890 853 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
5891 : (errmsg("database system was shut down at %s",
5892 : str_time(ControlFile->time,
5893 : timebuf, sizeof(timebuf)))));
5894 853 : break;
5895 :
5896 33 : case DB_SHUTDOWNED_IN_RECOVERY:
5897 33 : ereport(LOG,
5898 : (errmsg("database system was shut down in recovery at %s",
5899 : str_time(ControlFile->time,
5900 : timebuf, sizeof(timebuf)))));
5901 33 : break;
5902 :
5903 0 : case DB_SHUTDOWNING:
5904 0 : ereport(LOG,
5905 : (errmsg("database system shutdown was interrupted; last known up at %s",
5906 : str_time(ControlFile->time,
5907 : timebuf, sizeof(timebuf)))));
5908 0 : break;
5909 :
5910 0 : case DB_IN_CRASH_RECOVERY:
5911 0 : ereport(LOG,
5912 : (errmsg("database system was interrupted while in recovery at %s",
5913 : str_time(ControlFile->time,
5914 : timebuf, sizeof(timebuf))),
5915 : errhint("This probably means that some data is corrupted and"
5916 : " you will have to use the last backup for recovery.")));
5917 0 : break;
5918 :
5919 7 : case DB_IN_ARCHIVE_RECOVERY:
5920 7 : ereport(LOG,
5921 : (errmsg("database system was interrupted while in recovery at log time %s",
5922 : str_time(ControlFile->checkPointCopy.time,
5923 : timebuf, sizeof(timebuf))),
5924 : errhint("If this has occurred more than once some data might be corrupted"
5925 : " and you might need to choose an earlier recovery target.")));
5926 7 : break;
5927 :
5928 181 : case DB_IN_PRODUCTION:
5929 181 : ereport(LOG,
5930 : (errmsg("database system was interrupted; last known up at %s",
5931 : str_time(ControlFile->time,
5932 : timebuf, sizeof(timebuf)))));
5933 181 : break;
5934 :
5935 0 : default:
5936 0 : ereport(FATAL,
5937 : (errcode(ERRCODE_DATA_CORRUPTED),
5938 : errmsg("control file contains invalid database cluster state")));
5939 : }
5940 :
5941 : /* This is just to allow attaching to startup process with a debugger */
5942 : #ifdef XLOG_REPLAY_DELAY
5943 : if (ControlFile->state != DB_SHUTDOWNED)
5944 : pg_usleep(60000000L);
5945 : #endif
5946 :
5947 : /*
5948 : * Verify that pg_wal, pg_wal/archive_status, and pg_wal/summaries exist.
5949 : * In cases where someone has performed a copy for PITR, these directories
5950 : * may have been excluded and need to be re-created.
5951 : */
5952 1074 : ValidateXLOGDirectoryStructure();
5953 :
5954 : /* Set up timeout handler needed to report startup progress. */
5955 1074 : if (!IsBootstrapProcessingMode())
5956 1017 : RegisterTimeout(STARTUP_PROGRESS_TIMEOUT,
5957 : startup_progress_timeout_handler);
5958 :
5959 : /*----------
5960 : * If we previously crashed, perform a couple of actions:
5961 : *
5962 : * - The pg_wal directory may still include some temporary WAL segments
5963 : * used when creating a new segment, so perform some clean up to not
5964 : * bloat this path. This is done first as there is no point to sync
5965 : * this temporary data.
5966 : *
5967 : * - There might be data which we had written, intending to fsync it, but
5968 : * which we had not actually fsync'd yet. Therefore, a power failure in
5969 : * the near future might cause earlier unflushed writes to be lost, even
5970 : * though more recent data written to disk from here on would be
5971 : * persisted. To avoid that, fsync the entire data directory.
5972 : */
5973 1074 : if (ControlFile->state != DB_SHUTDOWNED &&
5974 221 : ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
5975 : {
5976 188 : RemoveTempXlogFiles();
5977 188 : SyncDataDirectory();
5978 188 : didCrash = true;
5979 : }
5980 : else
5981 886 : didCrash = false;
5982 :
5983 : /*
5984 : * Prepare for WAL recovery if needed.
5985 : *
5986 : * InitWalRecovery analyzes the control file and the backup label file, if
5987 : * any. It updates the in-memory ControlFile buffer according to the
5988 : * starting checkpoint, and sets InRecovery and ArchiveRecoveryRequested.
5989 : * It also applies the tablespace map file, if any.
5990 : */
5991 1074 : InitWalRecovery(ControlFile, &wasShutdown,
5992 : &haveBackupLabel, &haveTblspcMap);
5993 1072 : checkPoint = ControlFile->checkPointCopy;
5994 :
5995 : /* initialize shared memory variables from the checkpoint record */
5996 1072 : TransamVariables->nextXid = checkPoint.nextXid;
5997 1072 : TransamVariables->nextOid = checkPoint.nextOid;
5998 1072 : TransamVariables->oidCount = 0;
5999 1072 : MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
6000 1072 : AdvanceOldestClogXid(checkPoint.oldestXid);
6001 1072 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
6002 1072 : SetMultiXactIdLimit(checkPoint.oldestMulti, checkPoint.oldestMultiDB);
6003 1072 : SetCommitTsLimit(checkPoint.oldestCommitTsXid,
6004 : checkPoint.newestCommitTsXid);
6005 :
6006 : /*
6007 : * Clear out any old relcache cache files. This is *necessary* if we do
6008 : * any WAL replay, since that would probably result in the cache files
6009 : * being out of sync with database reality. In theory we could leave them
6010 : * in place if the database had been cleanly shut down, but it seems
6011 : * safest to just remove them always and let them be rebuilt during the
6012 : * first backend startup. These files needs to be removed from all
6013 : * directories including pg_tblspc, however the symlinks are created only
6014 : * after reading tablespace_map file in case of archive recovery from
6015 : * backup, so needs to clear old relcache files here after creating
6016 : * symlinks.
6017 : */
6018 1072 : RelationCacheInitFileRemove();
6019 :
6020 : /*
6021 : * Initialize replication slots, before there's a chance to remove
6022 : * required resources.
6023 : */
6024 1072 : StartupReplicationSlots();
6025 :
6026 : /*
6027 : * Startup the logical decoding status with the last status stored in the
6028 : * checkpoint record.
6029 : */
6030 1070 : StartupLogicalDecodingStatus(checkPoint.logicalDecodingEnabled);
6031 :
6032 : /*
6033 : * Startup logical state, needs to be setup now so we have proper data
6034 : * during crash recovery.
6035 : */
6036 1070 : StartupReorderBuffer();
6037 :
6038 : /*
6039 : * Startup CLOG. This must be done after TransamVariables->nextXid has
6040 : * been initialized and before we accept connections or begin WAL replay.
6041 : */
6042 1070 : StartupCLOG();
6043 :
6044 : /*
6045 : * Startup MultiXact. We need to do this early to be able to replay
6046 : * truncations.
6047 : */
6048 1070 : StartupMultiXact();
6049 :
6050 : /*
6051 : * Ditto for commit timestamps. Activate the facility if the setting is
6052 : * enabled in the control file, as there should be no tracking of commit
6053 : * timestamps done when the setting was disabled. This facility can be
6054 : * started or stopped when replaying a XLOG_PARAMETER_CHANGE record.
6055 : */
6056 1070 : if (ControlFile->track_commit_timestamp)
6057 14 : StartupCommitTs();
6058 :
6059 : /*
6060 : * Recover knowledge about replay progress of known replication partners.
6061 : */
6062 1070 : StartupReplicationOrigin();
6063 :
6064 : /*
6065 : * Initialize unlogged LSN. On a clean shutdown, it's restored from the
6066 : * control file. On recovery, all unlogged relations are blown away, so
6067 : * the unlogged LSN counter can be reset too.
6068 : */
6069 1070 : if (ControlFile->state == DB_SHUTDOWNED)
6070 844 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
6071 844 : ControlFile->unloggedLSN);
6072 : else
6073 226 : pg_atomic_write_membarrier_u64(&XLogCtl->unloggedLSN,
6074 : FirstNormalUnloggedLSN);
6075 :
6076 : /*
6077 : * Copy any missing timeline history files between 'now' and the recovery
6078 : * target timeline from archive to pg_wal. While we don't need those files
6079 : * ourselves - the history file of the recovery target timeline covers all
6080 : * the previous timelines in the history too - a cascading standby server
6081 : * might be interested in them. Or, if you archive the WAL from this
6082 : * server to a different archive than the primary, it'd be good for all
6083 : * the history files to get archived there after failover, so that you can
6084 : * use one of the old timelines as a PITR target. Timeline history files
6085 : * are small, so it's better to copy them unnecessarily than not copy them
6086 : * and regret later.
6087 : */
6088 1070 : restoreTimeLineHistoryFiles(checkPoint.ThisTimeLineID, recoveryTargetTLI);
6089 :
6090 : /*
6091 : * Before running in recovery, scan pg_twophase and fill in its status to
6092 : * be able to work on entries generated by redo. Doing a scan before
6093 : * taking any recovery action has the merit to discard any 2PC files that
6094 : * are newer than the first record to replay, saving from any conflicts at
6095 : * replay. This avoids as well any subsequent scans when doing recovery
6096 : * of the on-disk two-phase data.
6097 : */
6098 1070 : restoreTwoPhaseData();
6099 :
6100 : /*
6101 : * When starting with crash recovery, reset pgstat data - it might not be
6102 : * valid. Otherwise restore pgstat data. It's safe to do this here,
6103 : * because postmaster will not yet have started any other processes.
6104 : *
6105 : * NB: Restoring replication slot stats relies on slot state to have
6106 : * already been restored from disk.
6107 : *
6108 : * TODO: With a bit of extra work we could just start with a pgstat file
6109 : * associated with the checkpoint redo location we're starting from.
6110 : */
6111 1070 : if (didCrash)
6112 186 : pgstat_discard_stats();
6113 : else
6114 884 : pgstat_restore_stats();
6115 :
6116 1070 : lastFullPageWrites = checkPoint.fullPageWrites;
6117 :
6118 1070 : RedoRecPtr = XLogCtl->RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
6119 1070 : doPageWrites = lastFullPageWrites;
6120 :
6121 : /* REDO */
6122 1070 : if (InRecovery)
6123 : {
6124 : /* Initialize state for RecoveryInProgress() */
6125 226 : SpinLockAcquire(&XLogCtl->info_lck);
6126 226 : if (InArchiveRecovery)
6127 124 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6128 : else
6129 102 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_CRASH;
6130 226 : SpinLockRelease(&XLogCtl->info_lck);
6131 :
6132 : /*
6133 : * Update pg_control to show that we are recovering and to show the
6134 : * selected checkpoint as the place we are starting from. We also mark
6135 : * pg_control with any minimum recovery stop point obtained from a
6136 : * backup history file.
6137 : *
6138 : * No need to hold ControlFileLock yet, we aren't up far enough.
6139 : */
6140 226 : UpdateControlFile();
6141 :
6142 : /*
6143 : * If there was a backup label file, it's done its job and the info
6144 : * has now been propagated into pg_control. We must get rid of the
6145 : * label file so that if we crash during recovery, we'll pick up at
6146 : * the latest recovery restartpoint instead of going all the way back
6147 : * to the backup start point. It seems prudent though to just rename
6148 : * the file out of the way rather than delete it completely.
6149 : */
6150 226 : if (haveBackupLabel)
6151 : {
6152 82 : unlink(BACKUP_LABEL_OLD);
6153 82 : durable_rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD, FATAL);
6154 : }
6155 :
6156 : /*
6157 : * If there was a tablespace_map file, it's done its job and the
6158 : * symlinks have been created. We must get rid of the map file so
6159 : * that if we crash during recovery, we don't create symlinks again.
6160 : * It seems prudent though to just rename the file out of the way
6161 : * rather than delete it completely.
6162 : */
6163 226 : if (haveTblspcMap)
6164 : {
6165 2 : unlink(TABLESPACE_MAP_OLD);
6166 2 : durable_rename(TABLESPACE_MAP, TABLESPACE_MAP_OLD, FATAL);
6167 : }
6168 :
6169 : /*
6170 : * Initialize our local copy of minRecoveryPoint. When doing crash
6171 : * recovery we want to replay up to the end of WAL. Particularly, in
6172 : * the case of a promoted standby minRecoveryPoint value in the
6173 : * control file is only updated after the first checkpoint. However,
6174 : * if the instance crashes before the first post-recovery checkpoint
6175 : * is completed then recovery will use a stale location causing the
6176 : * startup process to think that there are still invalid page
6177 : * references when checking for data consistency.
6178 : */
6179 226 : if (InArchiveRecovery)
6180 : {
6181 124 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6182 124 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6183 : }
6184 : else
6185 : {
6186 102 : LocalMinRecoveryPoint = InvalidXLogRecPtr;
6187 102 : LocalMinRecoveryPointTLI = 0;
6188 : }
6189 :
6190 : /* Check that the GUCs used to generate the WAL allow recovery */
6191 226 : CheckRequiredParameterValues();
6192 :
6193 : /*
6194 : * We're in recovery, so unlogged relations may be trashed and must be
6195 : * reset. This should be done BEFORE allowing Hot Standby
6196 : * connections, so that read-only backends don't try to read whatever
6197 : * garbage is left over from before.
6198 : */
6199 226 : ResetUnloggedRelations(UNLOGGED_RELATION_CLEANUP);
6200 :
6201 : /*
6202 : * Likewise, delete any saved transaction snapshot files that got left
6203 : * behind by crashed backends.
6204 : */
6205 226 : DeleteAllExportedSnapshotFiles();
6206 :
6207 : /*
6208 : * Initialize for Hot Standby, if enabled. We won't let backends in
6209 : * yet, not until we've reached the min recovery point specified in
6210 : * control file and we've established a recovery snapshot from a
6211 : * running-xacts WAL record.
6212 : */
6213 226 : if (ArchiveRecoveryRequested && EnableHotStandby)
6214 : {
6215 : TransactionId *xids;
6216 : int nxids;
6217 :
6218 116 : ereport(DEBUG1,
6219 : (errmsg_internal("initializing for hot standby")));
6220 :
6221 116 : InitRecoveryTransactionEnvironment();
6222 :
6223 116 : if (wasShutdown)
6224 26 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
6225 : else
6226 90 : oldestActiveXID = checkPoint.oldestActiveXid;
6227 : Assert(TransactionIdIsValid(oldestActiveXID));
6228 :
6229 : /* Tell procarray about the range of xids it has to deal with */
6230 116 : ProcArrayInitRecovery(XidFromFullTransactionId(TransamVariables->nextXid));
6231 :
6232 : /*
6233 : * Startup subtrans only. CLOG, MultiXact and commit timestamp
6234 : * have already been started up and other SLRUs are not maintained
6235 : * during recovery and need not be started yet.
6236 : */
6237 116 : StartupSUBTRANS(oldestActiveXID);
6238 :
6239 : /*
6240 : * If we're beginning at a shutdown checkpoint, we know that
6241 : * nothing was running on the primary at this point. So fake-up an
6242 : * empty running-xacts record and use that here and now. Recover
6243 : * additional standby state for prepared transactions.
6244 : */
6245 116 : if (wasShutdown)
6246 : {
6247 : RunningTransactionsData running;
6248 : TransactionId latestCompletedXid;
6249 :
6250 : /* Update pg_subtrans entries for any prepared transactions */
6251 26 : StandbyRecoverPreparedTransactions();
6252 :
6253 : /*
6254 : * Construct a RunningTransactions snapshot representing a
6255 : * shut down server, with only prepared transactions still
6256 : * alive. We're never overflowed at this point because all
6257 : * subxids are listed with their parent prepared transactions.
6258 : */
6259 26 : running.xcnt = nxids;
6260 26 : running.subxcnt = 0;
6261 26 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
6262 26 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
6263 26 : running.oldestRunningXid = oldestActiveXID;
6264 26 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
6265 26 : TransactionIdRetreat(latestCompletedXid);
6266 : Assert(TransactionIdIsNormal(latestCompletedXid));
6267 26 : running.latestCompletedXid = latestCompletedXid;
6268 26 : running.xids = xids;
6269 :
6270 26 : ProcArrayApplyRecoveryInfo(&running);
6271 : }
6272 : }
6273 :
6274 : /*
6275 : * We're all set for replaying the WAL now. Do it.
6276 : */
6277 226 : PerformWalRecovery();
6278 163 : performedWalRecovery = true;
6279 : }
6280 : else
6281 844 : performedWalRecovery = false;
6282 :
6283 : /*
6284 : * Finish WAL recovery.
6285 : */
6286 1007 : endOfRecoveryInfo = FinishWalRecovery();
6287 1007 : EndOfLog = endOfRecoveryInfo->endOfLog;
6288 1007 : EndOfLogTLI = endOfRecoveryInfo->endOfLogTLI;
6289 1007 : abortedRecPtr = endOfRecoveryInfo->abortedRecPtr;
6290 1007 : missingContrecPtr = endOfRecoveryInfo->missingContrecPtr;
6291 :
6292 : /*
6293 : * Reset ps status display, so as no information related to recovery shows
6294 : * up.
6295 : */
6296 1007 : set_ps_display("");
6297 :
6298 : /*
6299 : * When recovering from a backup (we are in recovery, and archive recovery
6300 : * was requested), complain if we did not roll forward far enough to reach
6301 : * the point where the database is consistent. For regular online
6302 : * backup-from-primary, that means reaching the end-of-backup WAL record
6303 : * (at which point we reset backupStartPoint to be Invalid), for
6304 : * backup-from-replica (which can't inject records into the WAL stream),
6305 : * that point is when we reach the minRecoveryPoint in pg_control (which
6306 : * we purposefully copy last when backing up from a replica). For
6307 : * pg_rewind (which creates a backup_label with a method of "pg_rewind")
6308 : * or snapshot-style backups (which don't), backupEndRequired will be set
6309 : * to false.
6310 : *
6311 : * Note: it is indeed okay to look at the local variable
6312 : * LocalMinRecoveryPoint here, even though ControlFile->minRecoveryPoint
6313 : * might be further ahead --- ControlFile->minRecoveryPoint cannot have
6314 : * been advanced beyond the WAL we processed.
6315 : */
6316 1007 : if (InRecovery &&
6317 163 : (EndOfLog < LocalMinRecoveryPoint ||
6318 163 : XLogRecPtrIsValid(ControlFile->backupStartPoint)))
6319 : {
6320 : /*
6321 : * Ran off end of WAL before reaching end-of-backup WAL record, or
6322 : * minRecoveryPoint. That's a bad sign, indicating that you tried to
6323 : * recover from an online backup but never called pg_backup_stop(), or
6324 : * you didn't archive all the WAL needed.
6325 : */
6326 0 : if (ArchiveRecoveryRequested || ControlFile->backupEndRequired)
6327 : {
6328 0 : if (XLogRecPtrIsValid(ControlFile->backupStartPoint) || ControlFile->backupEndRequired)
6329 0 : ereport(FATAL,
6330 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6331 : errmsg("WAL ends before end of online backup"),
6332 : errhint("All WAL generated while online backup was taken must be available at recovery.")));
6333 : else
6334 0 : ereport(FATAL,
6335 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6336 : errmsg("WAL ends before consistent recovery point")));
6337 : }
6338 : }
6339 :
6340 : /*
6341 : * Reset unlogged relations to the contents of their INIT fork. This is
6342 : * done AFTER recovery is complete so as to include any unlogged relations
6343 : * created during recovery, but BEFORE recovery is marked as having
6344 : * completed successfully. Otherwise we'd not retry if any of the post
6345 : * end-of-recovery steps fail.
6346 : */
6347 1007 : if (InRecovery)
6348 163 : ResetUnloggedRelations(UNLOGGED_RELATION_INIT);
6349 :
6350 : /*
6351 : * Pre-scan prepared transactions to find out the range of XIDs present.
6352 : * This information is not quite needed yet, but it is positioned here so
6353 : * as potential problems are detected before any on-disk change is done.
6354 : */
6355 1007 : oldestActiveXID = PrescanPreparedTransactions(NULL, NULL);
6356 :
6357 : /*
6358 : * Allow ordinary WAL segment creation before possibly switching to a new
6359 : * timeline, which creates a new segment, and after the last ReadRecord().
6360 : */
6361 1007 : SetInstallXLogFileSegmentActive();
6362 :
6363 : /*
6364 : * Consider whether we need to assign a new timeline ID.
6365 : *
6366 : * If we did archive recovery, we always assign a new ID. This handles a
6367 : * couple of issues. If we stopped short of the end of WAL during
6368 : * recovery, then we are clearly generating a new timeline and must assign
6369 : * it a unique new ID. Even if we ran to the end, modifying the current
6370 : * last segment is problematic because it may result in trying to
6371 : * overwrite an already-archived copy of that segment, and we encourage
6372 : * DBAs to make their archive_commands reject that. We can dodge the
6373 : * problem by making the new active segment have a new timeline ID.
6374 : *
6375 : * In a normal crash recovery, we can just extend the timeline we were in.
6376 : */
6377 1007 : newTLI = endOfRecoveryInfo->lastRecTLI;
6378 1007 : if (ArchiveRecoveryRequested)
6379 : {
6380 55 : newTLI = findNewestTimeLine(recoveryTargetTLI) + 1;
6381 55 : ereport(LOG,
6382 : (errmsg("selected new timeline ID: %u", newTLI)));
6383 :
6384 : /*
6385 : * Make a writable copy of the last WAL segment. (Note that we also
6386 : * have a copy of the last block of the old WAL in
6387 : * endOfRecovery->lastPage; we will use that below.)
6388 : */
6389 55 : XLogInitNewTimeline(EndOfLogTLI, EndOfLog, newTLI);
6390 :
6391 : /*
6392 : * Remove the signal files out of the way, so that we don't
6393 : * accidentally re-enter archive recovery mode in a subsequent crash.
6394 : */
6395 55 : if (endOfRecoveryInfo->standby_signal_file_found)
6396 52 : durable_unlink(STANDBY_SIGNAL_FILE, FATAL);
6397 :
6398 55 : if (endOfRecoveryInfo->recovery_signal_file_found)
6399 4 : durable_unlink(RECOVERY_SIGNAL_FILE, FATAL);
6400 :
6401 : /*
6402 : * Write the timeline history file, and have it archived. After this
6403 : * point (or rather, as soon as the file is archived), the timeline
6404 : * will appear as "taken" in the WAL archive and to any standby
6405 : * servers. If we crash before actually switching to the new
6406 : * timeline, standby servers will nevertheless think that we switched
6407 : * to the new timeline, and will try to connect to the new timeline.
6408 : * To minimize the window for that, try to do as little as possible
6409 : * between here and writing the end-of-recovery record.
6410 : */
6411 55 : writeTimeLineHistory(newTLI, recoveryTargetTLI,
6412 : EndOfLog, endOfRecoveryInfo->recoveryStopReason);
6413 :
6414 55 : ereport(LOG,
6415 : (errmsg("archive recovery complete")));
6416 : }
6417 :
6418 : /* Save the selected TimeLineID in shared memory, too */
6419 1007 : SpinLockAcquire(&XLogCtl->info_lck);
6420 1007 : XLogCtl->InsertTimeLineID = newTLI;
6421 1007 : XLogCtl->PrevTimeLineID = endOfRecoveryInfo->lastRecTLI;
6422 1007 : SpinLockRelease(&XLogCtl->info_lck);
6423 :
6424 : /*
6425 : * Actually, if WAL ended in an incomplete record, skip the parts that
6426 : * made it through and start writing after the portion that persisted.
6427 : * (It's critical to first write an OVERWRITE_CONTRECORD message, which
6428 : * we'll do as soon as we're open for writing new WAL.)
6429 : */
6430 1007 : if (XLogRecPtrIsValid(missingContrecPtr))
6431 : {
6432 : /*
6433 : * We should only have a missingContrecPtr if we're not switching to a
6434 : * new timeline. When a timeline switch occurs, WAL is copied from the
6435 : * old timeline to the new only up to the end of the last complete
6436 : * record, so there can't be an incomplete WAL record that we need to
6437 : * disregard.
6438 : */
6439 : Assert(newTLI == endOfRecoveryInfo->lastRecTLI);
6440 : Assert(XLogRecPtrIsValid(abortedRecPtr));
6441 10 : EndOfLog = missingContrecPtr;
6442 : }
6443 :
6444 : /*
6445 : * Prepare to write WAL starting at EndOfLog location, and init xlog
6446 : * buffer cache using the block containing the last record from the
6447 : * previous incarnation.
6448 : */
6449 1007 : Insert = &XLogCtl->Insert;
6450 1007 : Insert->PrevBytePos = XLogRecPtrToBytePos(endOfRecoveryInfo->lastRec);
6451 1007 : Insert->CurrBytePos = XLogRecPtrToBytePos(EndOfLog);
6452 :
6453 : /*
6454 : * Tricky point here: lastPage contains the *last* block that the LastRec
6455 : * record spans, not the one it starts in. The last block is indeed the
6456 : * one we want to use.
6457 : */
6458 1007 : if (EndOfLog % XLOG_BLCKSZ != 0)
6459 : {
6460 : char *page;
6461 : int len;
6462 : int firstIdx;
6463 :
6464 975 : firstIdx = XLogRecPtrToBufIdx(EndOfLog);
6465 975 : len = EndOfLog - endOfRecoveryInfo->lastPageBeginPtr;
6466 : Assert(len < XLOG_BLCKSZ);
6467 :
6468 : /* Copy the valid part of the last block, and zero the rest */
6469 975 : page = &XLogCtl->pages[firstIdx * XLOG_BLCKSZ];
6470 975 : memcpy(page, endOfRecoveryInfo->lastPage, len);
6471 975 : memset(page + len, 0, XLOG_BLCKSZ - len);
6472 :
6473 975 : pg_atomic_write_u64(&XLogCtl->xlblocks[firstIdx], endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ);
6474 975 : XLogCtl->InitializedUpTo = endOfRecoveryInfo->lastPageBeginPtr + XLOG_BLCKSZ;
6475 : }
6476 : else
6477 : {
6478 : /*
6479 : * There is no partial block to copy. Just set InitializedUpTo, and
6480 : * let the first attempt to insert a log record to initialize the next
6481 : * buffer.
6482 : */
6483 32 : XLogCtl->InitializedUpTo = EndOfLog;
6484 : }
6485 :
6486 : /*
6487 : * Update local and shared status. This is OK to do without any locks
6488 : * because no other process can be reading or writing WAL yet.
6489 : */
6490 1007 : LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
6491 1007 : pg_atomic_write_u64(&XLogCtl->logInsertResult, EndOfLog);
6492 1007 : pg_atomic_write_u64(&XLogCtl->logWriteResult, EndOfLog);
6493 1007 : pg_atomic_write_u64(&XLogCtl->logFlushResult, EndOfLog);
6494 1007 : XLogCtl->LogwrtRqst.Write = EndOfLog;
6495 1007 : XLogCtl->LogwrtRqst.Flush = EndOfLog;
6496 :
6497 : /*
6498 : * Preallocate additional log files, if wanted.
6499 : */
6500 1007 : PreallocXlogFiles(EndOfLog, newTLI);
6501 :
6502 : /*
6503 : * Okay, we're officially UP.
6504 : */
6505 1007 : InRecovery = false;
6506 :
6507 : /* start the archive_timeout timer and LSN running */
6508 1007 : XLogCtl->lastSegSwitchTime = (pg_time_t) time(NULL);
6509 1007 : XLogCtl->lastSegSwitchLSN = EndOfLog;
6510 :
6511 : /* also initialize latestCompletedXid, to nextXid - 1 */
6512 1007 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
6513 1007 : TransamVariables->latestCompletedXid = TransamVariables->nextXid;
6514 1007 : FullTransactionIdRetreat(&TransamVariables->latestCompletedXid);
6515 1007 : LWLockRelease(ProcArrayLock);
6516 :
6517 : /*
6518 : * Start up subtrans, if not already done for hot standby. (commit
6519 : * timestamps are started below, if necessary.)
6520 : */
6521 1007 : if (standbyState == STANDBY_DISABLED)
6522 952 : StartupSUBTRANS(oldestActiveXID);
6523 :
6524 : /*
6525 : * Perform end of recovery actions for any SLRUs that need it.
6526 : */
6527 1007 : TrimCLOG();
6528 1007 : TrimMultiXact();
6529 :
6530 : /*
6531 : * Reload shared-memory state for prepared transactions. This needs to
6532 : * happen before renaming the last partial segment of the old timeline as
6533 : * it may be possible that we have to recover some transactions from it.
6534 : */
6535 1007 : RecoverPreparedTransactions();
6536 :
6537 : /* Shut down xlogreader */
6538 1007 : ShutdownWalRecovery();
6539 :
6540 : /* Enable WAL writes for this backend only. */
6541 1007 : LocalSetXLogInsertAllowed();
6542 :
6543 : /* If necessary, write overwrite-contrecord before doing anything else */
6544 1007 : if (XLogRecPtrIsValid(abortedRecPtr))
6545 : {
6546 : Assert(XLogRecPtrIsValid(missingContrecPtr));
6547 10 : CreateOverwriteContrecordRecord(abortedRecPtr, missingContrecPtr, newTLI);
6548 : }
6549 :
6550 : /*
6551 : * Update full_page_writes in shared memory and write an XLOG_FPW_CHANGE
6552 : * record before resource manager writes cleanup WAL records or checkpoint
6553 : * record is written.
6554 : */
6555 1007 : Insert->fullPageWrites = lastFullPageWrites;
6556 1007 : UpdateFullPageWrites();
6557 :
6558 : /*
6559 : * Emit checkpoint or end-of-recovery record in XLOG, if required.
6560 : */
6561 1007 : if (performedWalRecovery)
6562 163 : promoted = PerformRecoveryXLogAction();
6563 :
6564 : /*
6565 : * If any of the critical GUCs have changed, log them before we allow
6566 : * backends to write WAL.
6567 : */
6568 1007 : XLogReportParameters();
6569 :
6570 : /* If this is archive recovery, perform post-recovery cleanup actions. */
6571 1007 : if (ArchiveRecoveryRequested)
6572 55 : CleanupAfterArchiveRecovery(EndOfLogTLI, EndOfLog, newTLI);
6573 :
6574 : /*
6575 : * Local WAL inserts enabled, so it's time to finish initialization of
6576 : * commit timestamp.
6577 : */
6578 1007 : CompleteCommitTsInitialization();
6579 :
6580 : /*
6581 : * Update logical decoding status in shared memory and write an
6582 : * XLOG_LOGICAL_DECODING_STATUS_CHANGE, if necessary.
6583 : */
6584 1007 : UpdateLogicalDecodingStatusEndOfRecovery();
6585 :
6586 : /* Clean up EndOfWalRecoveryInfo data to appease Valgrind leak checking */
6587 1007 : if (endOfRecoveryInfo->lastPage)
6588 985 : pfree(endOfRecoveryInfo->lastPage);
6589 1007 : pfree(endOfRecoveryInfo->recoveryStopReason);
6590 1007 : pfree(endOfRecoveryInfo);
6591 :
6592 : /*
6593 : * If we reach this point with checksums in the state inprogress-on, it
6594 : * means that data checksums were in the process of being enabled when the
6595 : * cluster shut down. Since processing didn't finish, the operation will
6596 : * have to be restarted from scratch since there is no capability to
6597 : * continue where it was when the cluster shut down. Thus, revert the
6598 : * state back to off, and inform the user with a warning message. Being
6599 : * able to restart processing is a TODO, but it wouldn't be possible to
6600 : * restart here since we cannot launch a dynamic background worker
6601 : * directly from here (it has to be from a regular backend).
6602 : */
6603 1007 : if (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_INPROGRESS_ON)
6604 : {
6605 0 : XLogChecksums(PG_DATA_CHECKSUM_OFF);
6606 :
6607 0 : SpinLockAcquire(&XLogCtl->info_lck);
6608 0 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_OFF;
6609 0 : SetLocalDataChecksumState(XLogCtl->data_checksum_version);
6610 0 : SpinLockRelease(&XLogCtl->info_lck);
6611 :
6612 0 : ereport(WARNING,
6613 : errmsg("enabling data checksums was interrupted"),
6614 : errhint("Data checksum processing must be manually restarted for checksums to be enabled"));
6615 : }
6616 :
6617 : /*
6618 : * If data checksums were being disabled when the cluster was shut down,
6619 : * we know that we have a state where all backends have stopped validating
6620 : * checksums and we can move to off instead of prompting the user to
6621 : * perform any action.
6622 : */
6623 1007 : if (XLogCtl->data_checksum_version == PG_DATA_CHECKSUM_INPROGRESS_OFF)
6624 : {
6625 0 : XLogChecksums(PG_DATA_CHECKSUM_OFF);
6626 :
6627 0 : SpinLockAcquire(&XLogCtl->info_lck);
6628 0 : XLogCtl->data_checksum_version = PG_DATA_CHECKSUM_OFF;
6629 0 : SetLocalDataChecksumState(XLogCtl->data_checksum_version);
6630 0 : SpinLockRelease(&XLogCtl->info_lck);
6631 : }
6632 :
6633 : /*
6634 : * All done with end-of-recovery actions.
6635 : *
6636 : * Now allow backends to write WAL and update the control file status in
6637 : * consequence. SharedRecoveryState, that controls if backends can write
6638 : * WAL, is updated while holding ControlFileLock to prevent other backends
6639 : * to look at an inconsistent state of the control file in shared memory.
6640 : * There is still a small window during which backends can write WAL and
6641 : * the control file is still referring to a system not in DB_IN_PRODUCTION
6642 : * state while looking at the on-disk control file.
6643 : *
6644 : * Also, we use info_lck to update SharedRecoveryState to ensure that
6645 : * there are no race conditions concerning visibility of other recent
6646 : * updates to shared memory.
6647 : */
6648 1007 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6649 1007 : ControlFile->state = DB_IN_PRODUCTION;
6650 :
6651 1007 : SpinLockAcquire(&XLogCtl->info_lck);
6652 1007 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_DONE;
6653 1007 : SpinLockRelease(&XLogCtl->info_lck);
6654 :
6655 1007 : UpdateControlFile();
6656 1007 : LWLockRelease(ControlFileLock);
6657 :
6658 : /*
6659 : * Wake up the checkpointer process as there might be a request to disable
6660 : * logical decoding by concurrent slot drop.
6661 : */
6662 1007 : WakeupCheckpointer();
6663 :
6664 : /*
6665 : * Wake up all waiters. They need to report an error that recovery was
6666 : * ended before reaching the target LSN.
6667 : */
6668 1007 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_REPLAY, InvalidXLogRecPtr);
6669 1007 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_WRITE, InvalidXLogRecPtr);
6670 1007 : WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_FLUSH, InvalidXLogRecPtr);
6671 :
6672 : /*
6673 : * Shutdown the recovery environment. This must occur after
6674 : * RecoverPreparedTransactions() (see notes in lock_twophase_recover())
6675 : * and after switching SharedRecoveryState to RECOVERY_STATE_DONE so as
6676 : * any session building a snapshot will not rely on KnownAssignedXids as
6677 : * RecoveryInProgress() would return false at this stage. This is
6678 : * particularly critical for prepared 2PC transactions, that would still
6679 : * need to be included in snapshots once recovery has ended.
6680 : */
6681 1007 : if (standbyState != STANDBY_DISABLED)
6682 55 : ShutdownRecoveryTransactionEnvironment();
6683 :
6684 : /*
6685 : * If there were cascading standby servers connected to us, nudge any wal
6686 : * sender processes to notice that we've been promoted.
6687 : */
6688 1007 : WalSndWakeup(true, true);
6689 :
6690 : /*
6691 : * If this was a promotion, request an (online) checkpoint now. This isn't
6692 : * required for consistency, but the last restartpoint might be far back,
6693 : * and in case of a crash, recovering from it might take a longer than is
6694 : * appropriate now that we're not in standby mode anymore.
6695 : */
6696 1007 : if (promoted)
6697 48 : RequestCheckpoint(CHECKPOINT_FORCE);
6698 1007 : }
6699 :
6700 : /*
6701 : * Callback from PerformWalRecovery(), called when we switch from crash
6702 : * recovery to archive recovery mode. Updates the control file accordingly.
6703 : */
6704 : void
6705 1 : SwitchIntoArchiveRecovery(XLogRecPtr EndRecPtr, TimeLineID replayTLI)
6706 : {
6707 : /* initialize minRecoveryPoint to this record */
6708 1 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6709 1 : ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
6710 1 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6711 : {
6712 1 : ControlFile->minRecoveryPoint = EndRecPtr;
6713 1 : ControlFile->minRecoveryPointTLI = replayTLI;
6714 : }
6715 : /* update local copy */
6716 1 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
6717 1 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
6718 :
6719 : /*
6720 : * The startup process can update its local copy of minRecoveryPoint from
6721 : * this point.
6722 : */
6723 1 : updateMinRecoveryPoint = true;
6724 :
6725 1 : UpdateControlFile();
6726 :
6727 : /*
6728 : * We update SharedRecoveryState while holding the lock on ControlFileLock
6729 : * so both states are consistent in shared memory.
6730 : */
6731 1 : SpinLockAcquire(&XLogCtl->info_lck);
6732 1 : XLogCtl->SharedRecoveryState = RECOVERY_STATE_ARCHIVE;
6733 1 : SpinLockRelease(&XLogCtl->info_lck);
6734 :
6735 1 : LWLockRelease(ControlFileLock);
6736 1 : }
6737 :
6738 : /*
6739 : * Callback from PerformWalRecovery(), called when we reach the end of backup.
6740 : * Updates the control file accordingly.
6741 : */
6742 : void
6743 82 : ReachedEndOfBackup(XLogRecPtr EndRecPtr, TimeLineID tli)
6744 : {
6745 : /*
6746 : * We have reached the end of base backup, as indicated by pg_control. The
6747 : * data on disk is now consistent (unless minRecoveryPoint is further
6748 : * ahead, which can happen if we crashed during previous recovery). Reset
6749 : * backupStartPoint and backupEndPoint, and update minRecoveryPoint to
6750 : * make sure we don't allow starting up at an earlier point even if
6751 : * recovery is stopped and restarted soon after this.
6752 : */
6753 82 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6754 :
6755 82 : if (ControlFile->minRecoveryPoint < EndRecPtr)
6756 : {
6757 77 : ControlFile->minRecoveryPoint = EndRecPtr;
6758 77 : ControlFile->minRecoveryPointTLI = tli;
6759 : }
6760 :
6761 82 : ControlFile->backupStartPoint = InvalidXLogRecPtr;
6762 82 : ControlFile->backupEndPoint = InvalidXLogRecPtr;
6763 82 : ControlFile->backupEndRequired = false;
6764 82 : UpdateControlFile();
6765 :
6766 82 : LWLockRelease(ControlFileLock);
6767 82 : }
6768 :
6769 : /*
6770 : * Perform whatever XLOG actions are necessary at end of REDO.
6771 : *
6772 : * The goal here is to make sure that we'll be able to recover properly if
6773 : * we crash again. If we choose to write a checkpoint, we'll write a shutdown
6774 : * checkpoint rather than an on-line one. This is not particularly critical,
6775 : * but since we may be assigning a new TLI, using a shutdown checkpoint allows
6776 : * us to have the rule that TLI only changes in shutdown checkpoints, which
6777 : * allows some extra error checking in xlog_redo.
6778 : */
6779 : static bool
6780 163 : PerformRecoveryXLogAction(void)
6781 : {
6782 163 : bool promoted = false;
6783 :
6784 : /*
6785 : * Perform a checkpoint to update all our recovery activity to disk.
6786 : *
6787 : * Note that we write a shutdown checkpoint rather than an on-line one.
6788 : * This is not particularly critical, but since we may be assigning a new
6789 : * TLI, using a shutdown checkpoint allows us to have the rule that TLI
6790 : * only changes in shutdown checkpoints, which allows some extra error
6791 : * checking in xlog_redo.
6792 : *
6793 : * In promotion, only create a lightweight end-of-recovery record instead
6794 : * of a full checkpoint. A checkpoint is requested later, after we're
6795 : * fully out of recovery mode and already accepting queries.
6796 : */
6797 218 : if (ArchiveRecoveryRequested && IsUnderPostmaster &&
6798 55 : PromoteIsTriggered())
6799 : {
6800 48 : promoted = true;
6801 :
6802 : /*
6803 : * Insert a special WAL record to mark the end of recovery, since we
6804 : * aren't doing a checkpoint. That means that the checkpointer process
6805 : * may likely be in the middle of a time-smoothed restartpoint and
6806 : * could continue to be for minutes after this. That sounds strange,
6807 : * but the effect is roughly the same and it would be stranger to try
6808 : * to come out of the restartpoint and then checkpoint. We request a
6809 : * checkpoint later anyway, just for safety.
6810 : */
6811 48 : CreateEndOfRecoveryRecord();
6812 : }
6813 : else
6814 : {
6815 115 : RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
6816 : CHECKPOINT_FAST |
6817 : CHECKPOINT_WAIT);
6818 : }
6819 :
6820 163 : return promoted;
6821 : }
6822 :
6823 : /*
6824 : * Is the system still in recovery?
6825 : *
6826 : * Unlike testing InRecovery, this works in any process that's connected to
6827 : * shared memory.
6828 : */
6829 : bool
6830 96982476 : RecoveryInProgress(void)
6831 : {
6832 : /*
6833 : * We check shared state each time only until we leave recovery mode. We
6834 : * can't re-enter recovery, so there's no need to keep checking after the
6835 : * shared variable has once been seen false.
6836 : */
6837 96982476 : if (!LocalRecoveryInProgress)
6838 94930848 : return false;
6839 : else
6840 : {
6841 : /*
6842 : * use volatile pointer to make sure we make a fresh read of the
6843 : * shared variable.
6844 : */
6845 2051628 : volatile XLogCtlData *xlogctl = XLogCtl;
6846 :
6847 2051628 : LocalRecoveryInProgress = (xlogctl->SharedRecoveryState != RECOVERY_STATE_DONE);
6848 :
6849 : /*
6850 : * Note: We don't need a memory barrier when we're still in recovery.
6851 : * We might exit recovery immediately after return, so the caller
6852 : * can't rely on 'true' meaning that we're still in recovery anyway.
6853 : */
6854 :
6855 2051628 : return LocalRecoveryInProgress;
6856 : }
6857 : }
6858 :
6859 : /*
6860 : * Returns current recovery state from shared memory.
6861 : *
6862 : * This returned state is kept consistent with the contents of the control
6863 : * file. See details about the possible values of RecoveryState in xlog.h.
6864 : */
6865 : RecoveryState
6866 10164 : GetRecoveryState(void)
6867 : {
6868 : RecoveryState retval;
6869 :
6870 10164 : SpinLockAcquire(&XLogCtl->info_lck);
6871 10164 : retval = XLogCtl->SharedRecoveryState;
6872 10164 : SpinLockRelease(&XLogCtl->info_lck);
6873 :
6874 10164 : return retval;
6875 : }
6876 :
6877 : /*
6878 : * Is this process allowed to insert new WAL records?
6879 : *
6880 : * Ordinarily this is essentially equivalent to !RecoveryInProgress().
6881 : * But we also have provisions for forcing the result "true" or "false"
6882 : * within specific processes regardless of the global state.
6883 : */
6884 : bool
6885 65985211 : XLogInsertAllowed(void)
6886 : {
6887 : /*
6888 : * If value is "unconditionally true" or "unconditionally false", just
6889 : * return it. This provides the normal fast path once recovery is known
6890 : * done.
6891 : */
6892 65985211 : if (LocalXLogInsertAllowed >= 0)
6893 65326632 : return (bool) LocalXLogInsertAllowed;
6894 :
6895 : /*
6896 : * Else, must check to see if we're still in recovery.
6897 : */
6898 658579 : if (RecoveryInProgress())
6899 648112 : return false;
6900 :
6901 : /*
6902 : * On exit from recovery, reset to "unconditionally true", since there is
6903 : * no need to keep checking.
6904 : */
6905 10467 : LocalXLogInsertAllowed = 1;
6906 10467 : return true;
6907 : }
6908 :
6909 : /*
6910 : * Make XLogInsertAllowed() return true in the current process only.
6911 : *
6912 : * Note: it is allowed to switch LocalXLogInsertAllowed back to -1 later,
6913 : * and even call LocalSetXLogInsertAllowed() again after that.
6914 : *
6915 : * Returns the previous value of LocalXLogInsertAllowed.
6916 : */
6917 : static int
6918 1037 : LocalSetXLogInsertAllowed(void)
6919 : {
6920 1037 : int oldXLogAllowed = LocalXLogInsertAllowed;
6921 :
6922 1037 : LocalXLogInsertAllowed = 1;
6923 :
6924 1037 : return oldXLogAllowed;
6925 : }
6926 :
6927 : /*
6928 : * Return the current Redo pointer from shared memory.
6929 : *
6930 : * As a side-effect, the local RedoRecPtr copy is updated.
6931 : */
6932 : XLogRecPtr
6933 383179 : GetRedoRecPtr(void)
6934 : {
6935 : XLogRecPtr ptr;
6936 :
6937 : /*
6938 : * The possibly not up-to-date copy in XLogCtl is enough. Even if we
6939 : * grabbed a WAL insertion lock to read the authoritative value in
6940 : * Insert->RedoRecPtr, someone might update it just after we've released
6941 : * the lock.
6942 : */
6943 383179 : SpinLockAcquire(&XLogCtl->info_lck);
6944 383179 : ptr = XLogCtl->RedoRecPtr;
6945 383179 : SpinLockRelease(&XLogCtl->info_lck);
6946 :
6947 383179 : if (RedoRecPtr < ptr)
6948 1745 : RedoRecPtr = ptr;
6949 :
6950 383179 : return RedoRecPtr;
6951 : }
6952 :
6953 : /*
6954 : * Return information needed to decide whether a modified block needs a
6955 : * full-page image to be included in the WAL record.
6956 : *
6957 : * The returned values are cached copies from backend-private memory, and
6958 : * possibly out-of-date or, indeed, uninitialized, in which case they will
6959 : * be InvalidXLogRecPtr and false, respectively. XLogInsertRecord will
6960 : * re-check them against up-to-date values, while holding the WAL insert lock.
6961 : */
6962 : void
6963 24314649 : GetFullPageWriteInfo(XLogRecPtr *RedoRecPtr_p, bool *doPageWrites_p)
6964 : {
6965 24314649 : *RedoRecPtr_p = RedoRecPtr;
6966 24314649 : *doPageWrites_p = doPageWrites;
6967 24314649 : }
6968 :
6969 : /*
6970 : * GetInsertRecPtr -- Returns the current insert position.
6971 : *
6972 : * NOTE: The value *actually* returned is the position of the last full
6973 : * xlog page. It lags behind the real insert position by at most 1 page.
6974 : * For that, we don't need to scan through WAL insertion locks, and an
6975 : * approximation is enough for the current usage of this function.
6976 : */
6977 : XLogRecPtr
6978 7263 : GetInsertRecPtr(void)
6979 : {
6980 : XLogRecPtr recptr;
6981 :
6982 7263 : SpinLockAcquire(&XLogCtl->info_lck);
6983 7263 : recptr = XLogCtl->LogwrtRqst.Write;
6984 7263 : SpinLockRelease(&XLogCtl->info_lck);
6985 :
6986 7263 : return recptr;
6987 : }
6988 :
6989 : /*
6990 : * GetFlushRecPtr -- Returns the current flush position, ie, the last WAL
6991 : * position known to be fsync'd to disk. This should only be used on a
6992 : * system that is known not to be in recovery.
6993 : */
6994 : XLogRecPtr
6995 215850 : GetFlushRecPtr(TimeLineID *insertTLI)
6996 : {
6997 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
6998 :
6999 215850 : RefreshXLogWriteResult(LogwrtResult);
7000 :
7001 : /*
7002 : * If we're writing and flushing WAL, the time line can't be changing, so
7003 : * no lock is required.
7004 : */
7005 215850 : if (insertTLI)
7006 24561 : *insertTLI = XLogCtl->InsertTimeLineID;
7007 :
7008 215850 : return LogwrtResult.Flush;
7009 : }
7010 :
7011 : /*
7012 : * GetWALInsertionTimeLine -- Returns the current timeline of a system that
7013 : * is not in recovery.
7014 : */
7015 : TimeLineID
7016 118516 : GetWALInsertionTimeLine(void)
7017 : {
7018 : Assert(XLogCtl->SharedRecoveryState == RECOVERY_STATE_DONE);
7019 :
7020 : /* Since the value can't be changing, no lock is required. */
7021 118516 : return XLogCtl->InsertTimeLineID;
7022 : }
7023 :
7024 : /*
7025 : * GetWALInsertionTimeLineIfSet -- If the system is not in recovery, returns
7026 : * the WAL insertion timeline; else, returns 0. Wherever possible, use
7027 : * GetWALInsertionTimeLine() instead, since it's cheaper. Note that this
7028 : * function decides recovery has ended as soon as the insert TLI is set, which
7029 : * happens before we set XLogCtl->SharedRecoveryState to RECOVERY_STATE_DONE.
7030 : */
7031 : TimeLineID
7032 1 : GetWALInsertionTimeLineIfSet(void)
7033 : {
7034 : TimeLineID insertTLI;
7035 :
7036 1 : SpinLockAcquire(&XLogCtl->info_lck);
7037 1 : insertTLI = XLogCtl->InsertTimeLineID;
7038 1 : SpinLockRelease(&XLogCtl->info_lck);
7039 :
7040 1 : return insertTLI;
7041 : }
7042 :
7043 : /*
7044 : * GetLastImportantRecPtr -- Returns the LSN of the last important record
7045 : * inserted. All records not explicitly marked as unimportant are considered
7046 : * important.
7047 : *
7048 : * The LSN is determined by computing the maximum of
7049 : * WALInsertLocks[i].lastImportantAt.
7050 : */
7051 : XLogRecPtr
7052 1774 : GetLastImportantRecPtr(void)
7053 : {
7054 1774 : XLogRecPtr res = InvalidXLogRecPtr;
7055 : int i;
7056 :
7057 15966 : for (i = 0; i < NUM_XLOGINSERT_LOCKS; i++)
7058 : {
7059 : XLogRecPtr last_important;
7060 :
7061 : /*
7062 : * Need to take a lock to prevent torn reads of the LSN, which are
7063 : * possible on some of the supported platforms. WAL insert locks only
7064 : * support exclusive mode, so we have to use that.
7065 : */
7066 14192 : LWLockAcquire(&WALInsertLocks[i].l.lock, LW_EXCLUSIVE);
7067 14192 : last_important = WALInsertLocks[i].l.lastImportantAt;
7068 14192 : LWLockRelease(&WALInsertLocks[i].l.lock);
7069 :
7070 14192 : if (res < last_important)
7071 2994 : res = last_important;
7072 : }
7073 :
7074 1774 : return res;
7075 : }
7076 :
7077 : /*
7078 : * Get the time and LSN of the last xlog segment switch
7079 : */
7080 : pg_time_t
7081 0 : GetLastSegSwitchData(XLogRecPtr *lastSwitchLSN)
7082 : {
7083 : pg_time_t result;
7084 :
7085 : /* Need WALWriteLock, but shared lock is sufficient */
7086 0 : LWLockAcquire(WALWriteLock, LW_SHARED);
7087 0 : result = XLogCtl->lastSegSwitchTime;
7088 0 : *lastSwitchLSN = XLogCtl->lastSegSwitchLSN;
7089 0 : LWLockRelease(WALWriteLock);
7090 :
7091 0 : return result;
7092 : }
7093 :
7094 : /*
7095 : * This must be called ONCE during postmaster or standalone-backend shutdown
7096 : */
7097 : void
7098 759 : ShutdownXLOG(int code, Datum arg)
7099 : {
7100 : /*
7101 : * We should have an aux process resource owner to use, and we should not
7102 : * be in a transaction that's installed some other resowner.
7103 : */
7104 : Assert(AuxProcessResourceOwner != NULL);
7105 : Assert(CurrentResourceOwner == NULL ||
7106 : CurrentResourceOwner == AuxProcessResourceOwner);
7107 759 : CurrentResourceOwner = AuxProcessResourceOwner;
7108 :
7109 : /* Don't be chatty in standalone mode */
7110 759 : ereport(IsPostmasterEnvironment ? LOG : NOTICE,
7111 : (errmsg("shutting down")));
7112 :
7113 : /*
7114 : * Signal walsenders to move to stopping state.
7115 : */
7116 759 : WalSndInitStopping();
7117 :
7118 : /*
7119 : * Wait for WAL senders to be in stopping state. This prevents commands
7120 : * from writing new WAL.
7121 : */
7122 759 : WalSndWaitStopping();
7123 :
7124 759 : if (RecoveryInProgress())
7125 60 : CreateRestartPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
7126 : else
7127 : {
7128 : /*
7129 : * If archiving is enabled, rotate the last XLOG file so that all the
7130 : * remaining records are archived (postmaster wakes up the archiver
7131 : * process one more time at the end of shutdown). The checkpoint
7132 : * record will go to the next XLOG file and won't be archived (yet).
7133 : */
7134 699 : if (XLogArchivingActive())
7135 16 : RequestXLogSwitch(false);
7136 :
7137 699 : CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FAST);
7138 : }
7139 759 : }
7140 :
7141 : /*
7142 : * Format checkpoint request flags as a space-separated string for
7143 : * log messages.
7144 : */
7145 : static const char *
7146 3194 : CheckpointFlagsString(int flags)
7147 : {
7148 : static char buf[128];
7149 :
7150 25552 : snprintf(buf, sizeof(buf), "%s%s%s%s%s%s%s%s",
7151 3194 : (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
7152 3194 : (flags & CHECKPOINT_END_OF_RECOVERY) ? " end-of-recovery" : "",
7153 3194 : (flags & CHECKPOINT_FAST) ? " fast" : "",
7154 3194 : (flags & CHECKPOINT_FORCE) ? " force" : "",
7155 3194 : (flags & CHECKPOINT_WAIT) ? " wait" : "",
7156 3194 : (flags & CHECKPOINT_CAUSE_XLOG) ? " wal" : "",
7157 3194 : (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "",
7158 3194 : (flags & CHECKPOINT_FLUSH_UNLOGGED) ? " flush-unlogged" : "");
7159 :
7160 3194 : return buf;
7161 : }
7162 :
7163 : /*
7164 : * Log start of a checkpoint.
7165 : */
7166 : static void
7167 1597 : LogCheckpointStart(int flags, bool restartpoint)
7168 : {
7169 1597 : if (restartpoint)
7170 210 : ereport(LOG,
7171 : /* translator: the placeholder shows checkpoint options */
7172 : (errmsg("restartpoint starting:%s",
7173 : CheckpointFlagsString(flags))));
7174 : else
7175 1387 : ereport(LOG,
7176 : /* translator: the placeholder shows checkpoint options */
7177 : (errmsg("checkpoint starting:%s",
7178 : CheckpointFlagsString(flags))));
7179 1597 : }
7180 :
7181 : /*
7182 : * Log end of a checkpoint.
7183 : */
7184 : static void
7185 1931 : LogCheckpointEnd(bool restartpoint, int flags)
7186 : {
7187 : long write_msecs,
7188 : sync_msecs,
7189 : total_msecs,
7190 : longest_msecs,
7191 : average_msecs;
7192 : uint64 average_sync_time;
7193 :
7194 1931 : CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
7195 :
7196 1931 : write_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_write_t,
7197 : CheckpointStats.ckpt_sync_t);
7198 :
7199 1931 : sync_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_sync_t,
7200 : CheckpointStats.ckpt_sync_end_t);
7201 :
7202 : /* Accumulate checkpoint timing summary data, in milliseconds. */
7203 1931 : PendingCheckpointerStats.write_time += write_msecs;
7204 1931 : PendingCheckpointerStats.sync_time += sync_msecs;
7205 :
7206 : /*
7207 : * All of the published timing statistics are accounted for. Only
7208 : * continue if a log message is to be written.
7209 : */
7210 1931 : if (!log_checkpoints)
7211 334 : return;
7212 :
7213 1597 : total_msecs = TimestampDifferenceMilliseconds(CheckpointStats.ckpt_start_t,
7214 : CheckpointStats.ckpt_end_t);
7215 :
7216 : /*
7217 : * Timing values returned from CheckpointStats are in microseconds.
7218 : * Convert to milliseconds for consistent printing.
7219 : */
7220 1597 : longest_msecs = (long) ((CheckpointStats.ckpt_longest_sync + 999) / 1000);
7221 :
7222 1597 : average_sync_time = 0;
7223 1597 : if (CheckpointStats.ckpt_sync_rels > 0)
7224 0 : average_sync_time = CheckpointStats.ckpt_agg_sync_time /
7225 0 : CheckpointStats.ckpt_sync_rels;
7226 1597 : average_msecs = (long) ((average_sync_time + 999) / 1000);
7227 :
7228 : /*
7229 : * ControlFileLock is not required to see ControlFile->checkPoint and
7230 : * ->checkPointCopy here as we are the only updator of those variables at
7231 : * this moment.
7232 : */
7233 1597 : if (restartpoint)
7234 210 : ereport(LOG,
7235 : (errmsg("restartpoint complete:%s: wrote %d buffers (%.1f%%), "
7236 : "wrote %d SLRU buffers; %d WAL file(s) added, "
7237 : "%d removed, %d recycled; write=%ld.%03d s, "
7238 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
7239 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
7240 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
7241 : CheckpointFlagsString(flags),
7242 : CheckpointStats.ckpt_bufs_written,
7243 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
7244 : CheckpointStats.ckpt_slru_written,
7245 : CheckpointStats.ckpt_segs_added,
7246 : CheckpointStats.ckpt_segs_removed,
7247 : CheckpointStats.ckpt_segs_recycled,
7248 : write_msecs / 1000, (int) (write_msecs % 1000),
7249 : sync_msecs / 1000, (int) (sync_msecs % 1000),
7250 : total_msecs / 1000, (int) (total_msecs % 1000),
7251 : CheckpointStats.ckpt_sync_rels,
7252 : longest_msecs / 1000, (int) (longest_msecs % 1000),
7253 : average_msecs / 1000, (int) (average_msecs % 1000),
7254 : (int) (PrevCheckPointDistance / 1024.0),
7255 : (int) (CheckPointDistanceEstimate / 1024.0),
7256 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
7257 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
7258 : else
7259 1387 : ereport(LOG,
7260 : (errmsg("checkpoint complete:%s: wrote %d buffers (%.1f%%), "
7261 : "wrote %d SLRU buffers; %d WAL file(s) added, "
7262 : "%d removed, %d recycled; write=%ld.%03d s, "
7263 : "sync=%ld.%03d s, total=%ld.%03d s; sync files=%d, "
7264 : "longest=%ld.%03d s, average=%ld.%03d s; distance=%d kB, "
7265 : "estimate=%d kB; lsn=%X/%08X, redo lsn=%X/%08X",
7266 : CheckpointFlagsString(flags),
7267 : CheckpointStats.ckpt_bufs_written,
7268 : (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
7269 : CheckpointStats.ckpt_slru_written,
7270 : CheckpointStats.ckpt_segs_added,
7271 : CheckpointStats.ckpt_segs_removed,
7272 : CheckpointStats.ckpt_segs_recycled,
7273 : write_msecs / 1000, (int) (write_msecs % 1000),
7274 : sync_msecs / 1000, (int) (sync_msecs % 1000),
7275 : total_msecs / 1000, (int) (total_msecs % 1000),
7276 : CheckpointStats.ckpt_sync_rels,
7277 : longest_msecs / 1000, (int) (longest_msecs % 1000),
7278 : average_msecs / 1000, (int) (average_msecs % 1000),
7279 : (int) (PrevCheckPointDistance / 1024.0),
7280 : (int) (CheckPointDistanceEstimate / 1024.0),
7281 : LSN_FORMAT_ARGS(ControlFile->checkPoint),
7282 : LSN_FORMAT_ARGS(ControlFile->checkPointCopy.redo))));
7283 : }
7284 :
7285 : /*
7286 : * Update the estimate of distance between checkpoints.
7287 : *
7288 : * The estimate is used to calculate the number of WAL segments to keep
7289 : * preallocated, see XLOGfileslop().
7290 : */
7291 : static void
7292 1931 : UpdateCheckPointDistanceEstimate(uint64 nbytes)
7293 : {
7294 : /*
7295 : * To estimate the number of segments consumed between checkpoints, keep a
7296 : * moving average of the amount of WAL generated in previous checkpoint
7297 : * cycles. However, if the load is bursty, with quiet periods and busy
7298 : * periods, we want to cater for the peak load. So instead of a plain
7299 : * moving average, let the average decline slowly if the previous cycle
7300 : * used less WAL than estimated, but bump it up immediately if it used
7301 : * more.
7302 : *
7303 : * When checkpoints are triggered by max_wal_size, this should converge to
7304 : * CheckpointSegments * wal_segment_size,
7305 : *
7306 : * Note: This doesn't pay any attention to what caused the checkpoint.
7307 : * Checkpoints triggered manually with CHECKPOINT command, or by e.g.
7308 : * starting a base backup, are counted the same as those created
7309 : * automatically. The slow-decline will largely mask them out, if they are
7310 : * not frequent. If they are frequent, it seems reasonable to count them
7311 : * in as any others; if you issue a manual checkpoint every 5 minutes and
7312 : * never let a timed checkpoint happen, it makes sense to base the
7313 : * preallocation on that 5 minute interval rather than whatever
7314 : * checkpoint_timeout is set to.
7315 : */
7316 1931 : PrevCheckPointDistance = nbytes;
7317 1931 : if (CheckPointDistanceEstimate < nbytes)
7318 855 : CheckPointDistanceEstimate = nbytes;
7319 : else
7320 1076 : CheckPointDistanceEstimate =
7321 1076 : (0.90 * CheckPointDistanceEstimate + 0.10 * (double) nbytes);
7322 1931 : }
7323 :
7324 : /*
7325 : * Update the ps display for a process running a checkpoint. Note that
7326 : * this routine should not do any allocations so as it can be called
7327 : * from a critical section.
7328 : */
7329 : static void
7330 3862 : update_checkpoint_display(int flags, bool restartpoint, bool reset)
7331 : {
7332 : /*
7333 : * The status is reported only for end-of-recovery and shutdown
7334 : * checkpoints or shutdown restartpoints. Updating the ps display is
7335 : * useful in those situations as it may not be possible to rely on
7336 : * pg_stat_activity to see the status of the checkpointer or the startup
7337 : * process.
7338 : */
7339 3862 : if ((flags & (CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IS_SHUTDOWN)) == 0)
7340 2356 : return;
7341 :
7342 1506 : if (reset)
7343 753 : set_ps_display("");
7344 : else
7345 : {
7346 : char activitymsg[128];
7347 :
7348 2259 : snprintf(activitymsg, sizeof(activitymsg), "performing %s%s%s",
7349 753 : (flags & CHECKPOINT_END_OF_RECOVERY) ? "end-of-recovery " : "",
7350 753 : (flags & CHECKPOINT_IS_SHUTDOWN) ? "shutdown " : "",
7351 : restartpoint ? "restartpoint" : "checkpoint");
7352 753 : set_ps_display(activitymsg);
7353 : }
7354 : }
7355 :
7356 :
7357 : /*
7358 : * Perform a checkpoint --- either during shutdown, or on-the-fly
7359 : *
7360 : * flags is a bitwise OR of the following:
7361 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
7362 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
7363 : * CHECKPOINT_FAST: finish the checkpoint ASAP, ignoring
7364 : * checkpoint_completion_target parameter.
7365 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
7366 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
7367 : * CHECKPOINT_END_OF_RECOVERY).
7368 : * CHECKPOINT_FLUSH_UNLOGGED: also flush buffers of unlogged tables.
7369 : *
7370 : * Note: flags contains other bits, of interest here only for logging purposes.
7371 : * In particular note that this routine is synchronous and does not pay
7372 : * attention to CHECKPOINT_WAIT.
7373 : *
7374 : * If !shutdown then we are writing an online checkpoint. An XLOG_CHECKPOINT_REDO
7375 : * record is inserted into WAL at the logical location of the checkpoint, before
7376 : * flushing anything to disk, and when the checkpoint is eventually completed,
7377 : * and it is from this point that WAL replay will begin in the case of a recovery
7378 : * from this checkpoint. Once everything is written to disk, an
7379 : * XLOG_CHECKPOINT_ONLINE record is written to complete the checkpoint, and
7380 : * points back to the earlier XLOG_CHECKPOINT_REDO record. This mechanism allows
7381 : * other write-ahead log records to be written while the checkpoint is in
7382 : * progress, but we must be very careful about order of operations. This function
7383 : * may take many minutes to execute on a busy system.
7384 : *
7385 : * On the other hand, when shutdown is true, concurrent insertion into the
7386 : * write-ahead log is impossible, so there is no need for two separate records.
7387 : * In this case, we only insert an XLOG_CHECKPOINT_SHUTDOWN record, and it's
7388 : * both the record marking the completion of the checkpoint and the location
7389 : * from which WAL replay would begin if needed.
7390 : *
7391 : * Returns true if a new checkpoint was performed, or false if it was skipped
7392 : * because the system was idle.
7393 : */
7394 : bool
7395 1725 : CreateCheckPoint(int flags)
7396 : {
7397 : bool shutdown;
7398 : CheckPoint checkPoint;
7399 : XLogRecPtr recptr;
7400 : XLogSegNo _logSegNo;
7401 1725 : XLogCtlInsert *Insert = &XLogCtl->Insert;
7402 : uint32 freespace;
7403 : XLogRecPtr PriorRedoPtr;
7404 : XLogRecPtr last_important_lsn;
7405 : VirtualTransactionId *vxids;
7406 : int nvxids;
7407 1725 : int oldXLogAllowed = 0;
7408 :
7409 : /*
7410 : * An end-of-recovery checkpoint is really a shutdown checkpoint, just
7411 : * issued at a different time.
7412 : */
7413 1725 : if (flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY))
7414 729 : shutdown = true;
7415 : else
7416 996 : shutdown = false;
7417 :
7418 : /* sanity check */
7419 1725 : if (RecoveryInProgress() && (flags & CHECKPOINT_END_OF_RECOVERY) == 0)
7420 0 : elog(ERROR, "can't create a checkpoint during recovery");
7421 :
7422 : /*
7423 : * Prepare to accumulate statistics.
7424 : *
7425 : * Note: because it is possible for log_checkpoints to change while a
7426 : * checkpoint proceeds, we always accumulate stats, even if
7427 : * log_checkpoints is currently off.
7428 : */
7429 18975 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
7430 1725 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
7431 :
7432 : /*
7433 : * Let smgr prepare for checkpoint; this has to happen outside the
7434 : * critical section and before we determine the REDO pointer. Note that
7435 : * smgr must not do anything that'd have to be undone if we decide no
7436 : * checkpoint is needed.
7437 : */
7438 1725 : SyncPreCheckpoint();
7439 :
7440 : /* Run these points outside the critical section. */
7441 1725 : INJECTION_POINT("create-checkpoint-initial", NULL);
7442 1725 : INJECTION_POINT_LOAD("create-checkpoint-run");
7443 :
7444 : /*
7445 : * Use a critical section to force system panic if we have trouble.
7446 : */
7447 1725 : START_CRIT_SECTION();
7448 :
7449 1725 : if (shutdown)
7450 : {
7451 729 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7452 729 : ControlFile->state = DB_SHUTDOWNING;
7453 729 : UpdateControlFile();
7454 729 : LWLockRelease(ControlFileLock);
7455 : }
7456 :
7457 : /* Begin filling in the checkpoint WAL record */
7458 22425 : MemSet(&checkPoint, 0, sizeof(checkPoint));
7459 1725 : checkPoint.time = (pg_time_t) time(NULL);
7460 :
7461 : /*
7462 : * For Hot Standby, derive the oldestActiveXid before we fix the redo
7463 : * pointer. This allows us to begin accumulating changes to assemble our
7464 : * starting snapshot of locks and transactions.
7465 : */
7466 1725 : if (!shutdown && XLogStandbyInfoActive())
7467 956 : checkPoint.oldestActiveXid = GetOldestActiveTransactionId(false, true);
7468 : else
7469 769 : checkPoint.oldestActiveXid = InvalidTransactionId;
7470 :
7471 : /*
7472 : * Get location of last important record before acquiring insert locks (as
7473 : * GetLastImportantRecPtr() also locks WAL locks).
7474 : */
7475 1725 : last_important_lsn = GetLastImportantRecPtr();
7476 :
7477 : /*
7478 : * If this isn't a shutdown or forced checkpoint, and if there has been no
7479 : * WAL activity requiring a checkpoint, skip it. The idea here is to
7480 : * avoid inserting duplicate checkpoints when the system is idle.
7481 : */
7482 1725 : if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
7483 : CHECKPOINT_FORCE)) == 0)
7484 : {
7485 207 : if (last_important_lsn == ControlFile->checkPoint)
7486 : {
7487 4 : END_CRIT_SECTION();
7488 4 : ereport(DEBUG1,
7489 : (errmsg_internal("checkpoint skipped because system is idle")));
7490 4 : return false;
7491 : }
7492 : }
7493 :
7494 : /*
7495 : * An end-of-recovery checkpoint is created before anyone is allowed to
7496 : * write WAL. To allow us to write the checkpoint record, temporarily
7497 : * enable XLogInsertAllowed.
7498 : */
7499 1721 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7500 30 : oldXLogAllowed = LocalSetXLogInsertAllowed();
7501 :
7502 1721 : checkPoint.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7503 1721 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7504 30 : checkPoint.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7505 : else
7506 1691 : checkPoint.PrevTimeLineID = checkPoint.ThisTimeLineID;
7507 :
7508 : /*
7509 : * We must block concurrent insertions while examining insert state.
7510 : */
7511 1721 : WALInsertLockAcquireExclusive();
7512 :
7513 1721 : checkPoint.fullPageWrites = Insert->fullPageWrites;
7514 1721 : checkPoint.wal_level = wal_level;
7515 :
7516 : /*
7517 : * Get the current data_checksum_version value from xlogctl, valid at the
7518 : * time of the checkpoint.
7519 : */
7520 1721 : checkPoint.dataChecksumState = XLogCtl->data_checksum_version;
7521 :
7522 1721 : if (shutdown)
7523 : {
7524 729 : XLogRecPtr curInsert = XLogBytePosToRecPtr(Insert->CurrBytePos);
7525 :
7526 : /*
7527 : * Compute new REDO record ptr = location of next XLOG record.
7528 : *
7529 : * Since this is a shutdown checkpoint, there can't be any concurrent
7530 : * WAL insertion.
7531 : */
7532 729 : freespace = INSERT_FREESPACE(curInsert);
7533 729 : if (freespace == 0)
7534 : {
7535 0 : if (XLogSegmentOffset(curInsert, wal_segment_size) == 0)
7536 0 : curInsert += SizeOfXLogLongPHD;
7537 : else
7538 0 : curInsert += SizeOfXLogShortPHD;
7539 : }
7540 729 : checkPoint.redo = curInsert;
7541 :
7542 : /*
7543 : * Here we update the shared RedoRecPtr for future XLogInsert calls;
7544 : * this must be done while holding all the insertion locks.
7545 : *
7546 : * Note: if we fail to complete the checkpoint, RedoRecPtr will be
7547 : * left pointing past where it really needs to point. This is okay;
7548 : * the only consequence is that XLogInsert might back up whole buffers
7549 : * that it didn't really need to. We can't postpone advancing
7550 : * RedoRecPtr because XLogInserts that happen while we are dumping
7551 : * buffers must assume that their buffer changes are not included in
7552 : * the checkpoint.
7553 : */
7554 729 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
7555 : }
7556 :
7557 : /*
7558 : * Now we can release the WAL insertion locks, allowing other xacts to
7559 : * proceed while we are flushing disk buffers.
7560 : */
7561 1721 : WALInsertLockRelease();
7562 :
7563 : /*
7564 : * If this is an online checkpoint, we have not yet determined the redo
7565 : * point. We do so now by inserting the special XLOG_CHECKPOINT_REDO
7566 : * record; the LSN at which it starts becomes the new redo pointer. We
7567 : * don't do this for a shutdown checkpoint, because in that case no WAL
7568 : * can be written between the redo point and the insertion of the
7569 : * checkpoint record itself, so the checkpoint record itself serves to
7570 : * mark the redo point.
7571 : */
7572 1721 : if (!shutdown)
7573 : {
7574 : xl_checkpoint_redo redo_rec;
7575 :
7576 992 : WALInsertLockAcquire();
7577 992 : redo_rec.wal_level = wal_level;
7578 992 : SpinLockAcquire(&XLogCtl->info_lck);
7579 992 : redo_rec.data_checksum_version = XLogCtl->data_checksum_version;
7580 992 : SpinLockRelease(&XLogCtl->info_lck);
7581 992 : WALInsertLockRelease();
7582 :
7583 : /* Include WAL level in record for WAL summarizer's benefit. */
7584 992 : XLogBeginInsert();
7585 992 : XLogRegisterData(&redo_rec, sizeof(xl_checkpoint_redo));
7586 992 : (void) XLogInsert(RM_XLOG_ID, XLOG_CHECKPOINT_REDO);
7587 :
7588 : /*
7589 : * XLogInsertRecord will have updated XLogCtl->Insert.RedoRecPtr in
7590 : * shared memory and RedoRecPtr in backend-local memory, but we need
7591 : * to copy that into the record that will be inserted when the
7592 : * checkpoint is complete.
7593 : */
7594 992 : checkPoint.redo = RedoRecPtr;
7595 : }
7596 :
7597 : /* Update the info_lck-protected copy of RedoRecPtr as well */
7598 1721 : SpinLockAcquire(&XLogCtl->info_lck);
7599 1721 : XLogCtl->RedoRecPtr = checkPoint.redo;
7600 1721 : SpinLockRelease(&XLogCtl->info_lck);
7601 :
7602 : /*
7603 : * If enabled, log checkpoint start. We postpone this until now so as not
7604 : * to log anything if we decided to skip the checkpoint.
7605 : */
7606 1721 : if (log_checkpoints)
7607 1387 : LogCheckpointStart(flags, false);
7608 :
7609 1721 : INJECTION_POINT_CACHED("create-checkpoint-run", NULL);
7610 :
7611 : /* Update the process title */
7612 1721 : update_checkpoint_display(flags, false, false);
7613 :
7614 : TRACE_POSTGRESQL_CHECKPOINT_START(flags);
7615 :
7616 : /*
7617 : * Get the other info we need for the checkpoint record.
7618 : *
7619 : * We don't need to save oldestClogXid in the checkpoint, it only matters
7620 : * for the short period in which clog is being truncated, and if we crash
7621 : * during that we'll redo the clog truncation and fix up oldestClogXid
7622 : * there.
7623 : */
7624 1721 : LWLockAcquire(XidGenLock, LW_SHARED);
7625 1721 : checkPoint.nextXid = TransamVariables->nextXid;
7626 1721 : checkPoint.oldestXid = TransamVariables->oldestXid;
7627 1721 : checkPoint.oldestXidDB = TransamVariables->oldestXidDB;
7628 1721 : LWLockRelease(XidGenLock);
7629 :
7630 1721 : LWLockAcquire(CommitTsLock, LW_SHARED);
7631 1721 : checkPoint.oldestCommitTsXid = TransamVariables->oldestCommitTsXid;
7632 1721 : checkPoint.newestCommitTsXid = TransamVariables->newestCommitTsXid;
7633 1721 : LWLockRelease(CommitTsLock);
7634 :
7635 1721 : LWLockAcquire(OidGenLock, LW_SHARED);
7636 1721 : checkPoint.nextOid = TransamVariables->nextOid;
7637 1721 : if (!shutdown)
7638 992 : checkPoint.nextOid += TransamVariables->oidCount;
7639 1721 : LWLockRelease(OidGenLock);
7640 :
7641 1721 : SpinLockAcquire(&XLogCtl->info_lck);
7642 1721 : checkPoint.dataChecksumState = XLogCtl->data_checksum_version;
7643 1721 : SpinLockRelease(&XLogCtl->info_lck);
7644 :
7645 1721 : checkPoint.logicalDecodingEnabled = IsLogicalDecodingEnabled();
7646 :
7647 1721 : MultiXactGetCheckptMulti(shutdown,
7648 : &checkPoint.nextMulti,
7649 : &checkPoint.nextMultiOffset,
7650 : &checkPoint.oldestMulti,
7651 : &checkPoint.oldestMultiDB);
7652 :
7653 : /*
7654 : * Having constructed the checkpoint record, ensure all shmem disk buffers
7655 : * and commit-log buffers are flushed to disk.
7656 : *
7657 : * This I/O could fail for various reasons. If so, we will fail to
7658 : * complete the checkpoint, but there is no reason to force a system
7659 : * panic. Accordingly, exit critical section while doing it.
7660 : */
7661 1721 : END_CRIT_SECTION();
7662 :
7663 : /*
7664 : * In some cases there are groups of actions that must all occur on one
7665 : * side or the other of a checkpoint record. Before flushing the
7666 : * checkpoint record we must explicitly wait for any backend currently
7667 : * performing those groups of actions.
7668 : *
7669 : * One example is end of transaction, so we must wait for any transactions
7670 : * that are currently in commit critical sections. If an xact inserted
7671 : * its commit record into XLOG just before the REDO point, then a crash
7672 : * restart from the REDO point would not replay that record, which means
7673 : * that our flushing had better include the xact's update of pg_xact. So
7674 : * we wait till he's out of his commit critical section before proceeding.
7675 : * See notes in RecordTransactionCommit().
7676 : *
7677 : * Because we've already released the insertion locks, this test is a bit
7678 : * fuzzy: it is possible that we will wait for xacts we didn't really need
7679 : * to wait for. But the delay should be short and it seems better to make
7680 : * checkpoint take a bit longer than to hold off insertions longer than
7681 : * necessary. (In fact, the whole reason we have this issue is that xact.c
7682 : * does commit record XLOG insertion and clog update as two separate steps
7683 : * protected by different locks, but again that seems best on grounds of
7684 : * minimizing lock contention.)
7685 : *
7686 : * A transaction that has not yet set delayChkptFlags when we look cannot
7687 : * be at risk, since it has not inserted its commit record yet; and one
7688 : * that's already cleared it is not at risk either, since it's done fixing
7689 : * clog and we will correctly flush the update below. So we cannot miss
7690 : * any xacts we need to wait for.
7691 : */
7692 1721 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_START);
7693 1721 : if (nvxids > 0)
7694 : {
7695 : do
7696 : {
7697 : /*
7698 : * Keep absorbing fsync requests while we wait. There could even
7699 : * be a deadlock if we don't, if the process that prevents the
7700 : * checkpoint is trying to add a request to the queue.
7701 : */
7702 26 : AbsorbSyncRequests();
7703 :
7704 26 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
7705 26 : pg_usleep(10000L); /* wait for 10 msec */
7706 26 : pgstat_report_wait_end();
7707 26 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7708 : DELAY_CHKPT_START));
7709 : }
7710 1721 : pfree(vxids);
7711 :
7712 1721 : CheckPointGuts(checkPoint.redo, flags);
7713 :
7714 1721 : vxids = GetVirtualXIDsDelayingChkpt(&nvxids, DELAY_CHKPT_COMPLETE);
7715 1721 : if (nvxids > 0)
7716 : {
7717 : do
7718 : {
7719 0 : AbsorbSyncRequests();
7720 :
7721 0 : pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_COMPLETE);
7722 0 : pg_usleep(10000L); /* wait for 10 msec */
7723 0 : pgstat_report_wait_end();
7724 0 : } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
7725 : DELAY_CHKPT_COMPLETE));
7726 : }
7727 1721 : pfree(vxids);
7728 :
7729 : /*
7730 : * Take a snapshot of running transactions and write this to WAL. This
7731 : * allows us to reconstruct the state of running transactions during
7732 : * archive recovery, if required. Skip, if this info disabled.
7733 : *
7734 : * If we are shutting down, or Startup process is completing crash
7735 : * recovery we don't need to write running xact data.
7736 : */
7737 1721 : if (!shutdown && XLogStandbyInfoActive())
7738 952 : LogStandbySnapshot(InvalidOid);
7739 :
7740 1721 : START_CRIT_SECTION();
7741 :
7742 : /*
7743 : * Now insert the checkpoint record into XLOG.
7744 : */
7745 1721 : XLogBeginInsert();
7746 1721 : XLogRegisterData(&checkPoint, sizeof(checkPoint));
7747 1721 : recptr = XLogInsert(RM_XLOG_ID,
7748 : shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
7749 : XLOG_CHECKPOINT_ONLINE);
7750 :
7751 1721 : XLogFlush(recptr);
7752 :
7753 : /*
7754 : * We mustn't write any new WAL after a shutdown checkpoint, or it will be
7755 : * overwritten at next startup. No-one should even try, this just allows
7756 : * sanity-checking. In the case of an end-of-recovery checkpoint, we want
7757 : * to just temporarily disable writing until the system has exited
7758 : * recovery.
7759 : */
7760 1721 : if (shutdown)
7761 : {
7762 729 : if (flags & CHECKPOINT_END_OF_RECOVERY)
7763 30 : LocalXLogInsertAllowed = oldXLogAllowed;
7764 : else
7765 699 : LocalXLogInsertAllowed = 0; /* never again write WAL */
7766 : }
7767 :
7768 : /*
7769 : * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
7770 : * = end of actual checkpoint record.
7771 : */
7772 1721 : if (shutdown && checkPoint.redo != ProcLastRecPtr)
7773 0 : ereport(PANIC,
7774 : (errmsg("concurrent write-ahead log activity while database system is shutting down")));
7775 :
7776 : /*
7777 : * Remember the prior checkpoint's redo ptr for
7778 : * UpdateCheckPointDistanceEstimate()
7779 : */
7780 1721 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
7781 :
7782 : /*
7783 : * Update the control file.
7784 : */
7785 1721 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7786 1721 : if (shutdown)
7787 729 : ControlFile->state = DB_SHUTDOWNED;
7788 1721 : ControlFile->checkPoint = ProcLastRecPtr;
7789 1721 : ControlFile->checkPointCopy = checkPoint;
7790 : /* crash recovery should always recover to the end of WAL */
7791 1721 : ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
7792 1721 : ControlFile->minRecoveryPointTLI = 0;
7793 :
7794 : /* make sure we start with the checksum version as of the checkpoint */
7795 1721 : ControlFile->data_checksum_version = checkPoint.dataChecksumState;
7796 :
7797 : /*
7798 : * Persist unloggedLSN value. It's reset on crash recovery, so this goes
7799 : * unused on non-shutdown checkpoints, but seems useful to store it always
7800 : * for debugging purposes.
7801 : */
7802 1721 : ControlFile->unloggedLSN = pg_atomic_read_membarrier_u64(&XLogCtl->unloggedLSN);
7803 :
7804 1721 : UpdateControlFile();
7805 1721 : LWLockRelease(ControlFileLock);
7806 :
7807 : /*
7808 : * We are now done with critical updates; no need for system panic if we
7809 : * have trouble while fooling with old log segments.
7810 : */
7811 1721 : END_CRIT_SECTION();
7812 :
7813 : /*
7814 : * WAL summaries end when the next XLOG_CHECKPOINT_REDO or
7815 : * XLOG_CHECKPOINT_SHUTDOWN record is reached. This is the first point
7816 : * where (a) we're not inside of a critical section and (b) we can be
7817 : * certain that the relevant record has been flushed to disk, which must
7818 : * happen before it can be summarized.
7819 : *
7820 : * If this is a shutdown checkpoint, then this happens reasonably
7821 : * promptly: we've only just inserted and flushed the
7822 : * XLOG_CHECKPOINT_SHUTDOWN record. If this is not a shutdown checkpoint,
7823 : * then this might not be very prompt at all: the XLOG_CHECKPOINT_REDO
7824 : * record was written before we began flushing data to disk, and that
7825 : * could be many minutes ago at this point. However, we don't XLogFlush()
7826 : * after inserting that record, so we're not guaranteed that it's on disk
7827 : * until after the above call that flushes the XLOG_CHECKPOINT_ONLINE
7828 : * record.
7829 : */
7830 1721 : WakeupWalSummarizer();
7831 :
7832 : /*
7833 : * Let smgr do post-checkpoint cleanup (eg, deleting old files).
7834 : */
7835 1721 : SyncPostCheckpoint();
7836 :
7837 : /*
7838 : * Update the average distance between checkpoints if the prior checkpoint
7839 : * exists.
7840 : */
7841 1721 : if (XLogRecPtrIsValid(PriorRedoPtr))
7842 1721 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
7843 :
7844 1721 : INJECTION_POINT("checkpoint-before-old-wal-removal", NULL);
7845 :
7846 : /*
7847 : * Delete old log files, those no longer needed for last checkpoint to
7848 : * prevent the disk holding the xlog from growing full.
7849 : */
7850 1721 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7851 1721 : KeepLogSeg(recptr, &_logSegNo);
7852 1721 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
7853 : _logSegNo, InvalidOid,
7854 : InvalidTransactionId))
7855 : {
7856 : /*
7857 : * Some slots have been invalidated; recalculate the old-segment
7858 : * horizon, starting again from RedoRecPtr.
7859 : */
7860 4 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
7861 4 : KeepLogSeg(recptr, &_logSegNo);
7862 : }
7863 1721 : _logSegNo--;
7864 1721 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
7865 : checkPoint.ThisTimeLineID);
7866 :
7867 : /*
7868 : * Make more log segments if needed. (Do this after recycling old log
7869 : * segments, since that may supply some of the needed files.)
7870 : */
7871 1721 : if (!shutdown)
7872 992 : PreallocXlogFiles(recptr, checkPoint.ThisTimeLineID);
7873 :
7874 : /*
7875 : * Truncate pg_subtrans if possible. We can throw away all data before
7876 : * the oldest XMIN of any running transaction. No future transaction will
7877 : * attempt to reference any pg_subtrans entry older than that (see Asserts
7878 : * in subtrans.c). During recovery, though, we mustn't do this because
7879 : * StartupSUBTRANS hasn't been called yet.
7880 : */
7881 1721 : if (!RecoveryInProgress())
7882 1691 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
7883 :
7884 : /* Real work is done; log and update stats. */
7885 1721 : LogCheckpointEnd(false, flags);
7886 :
7887 : /* Reset the process title */
7888 1721 : update_checkpoint_display(flags, false, true);
7889 :
7890 : TRACE_POSTGRESQL_CHECKPOINT_DONE(CheckpointStats.ckpt_bufs_written,
7891 : NBuffers,
7892 : CheckpointStats.ckpt_segs_added,
7893 : CheckpointStats.ckpt_segs_removed,
7894 : CheckpointStats.ckpt_segs_recycled);
7895 :
7896 1721 : return true;
7897 : }
7898 :
7899 : /*
7900 : * Mark the end of recovery in WAL though without running a full checkpoint.
7901 : * We can expect that a restartpoint is likely to be in progress as we
7902 : * do this, though we are unwilling to wait for it to complete.
7903 : *
7904 : * CreateRestartPoint() allows for the case where recovery may end before
7905 : * the restartpoint completes so there is no concern of concurrent behaviour.
7906 : */
7907 : static void
7908 48 : CreateEndOfRecoveryRecord(void)
7909 : {
7910 : xl_end_of_recovery xlrec;
7911 : XLogRecPtr recptr;
7912 :
7913 : /* sanity check */
7914 48 : if (!RecoveryInProgress())
7915 0 : elog(ERROR, "can only be used to end recovery");
7916 :
7917 48 : xlrec.end_time = GetCurrentTimestamp();
7918 48 : xlrec.wal_level = wal_level;
7919 :
7920 48 : WALInsertLockAcquireExclusive();
7921 48 : xlrec.ThisTimeLineID = XLogCtl->InsertTimeLineID;
7922 48 : xlrec.PrevTimeLineID = XLogCtl->PrevTimeLineID;
7923 48 : WALInsertLockRelease();
7924 :
7925 48 : START_CRIT_SECTION();
7926 :
7927 48 : XLogBeginInsert();
7928 48 : XLogRegisterData(&xlrec, sizeof(xl_end_of_recovery));
7929 48 : recptr = XLogInsert(RM_XLOG_ID, XLOG_END_OF_RECOVERY);
7930 :
7931 48 : XLogFlush(recptr);
7932 :
7933 : /*
7934 : * Update the control file so that crash recovery can follow the timeline
7935 : * changes to this point.
7936 : */
7937 48 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
7938 48 : ControlFile->minRecoveryPoint = recptr;
7939 48 : ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
7940 :
7941 : /* start with the latest checksum version (as of the end of recovery) */
7942 48 : SpinLockAcquire(&XLogCtl->info_lck);
7943 48 : ControlFile->data_checksum_version = XLogCtl->data_checksum_version;
7944 48 : SpinLockRelease(&XLogCtl->info_lck);
7945 :
7946 48 : UpdateControlFile();
7947 48 : LWLockRelease(ControlFileLock);
7948 :
7949 48 : END_CRIT_SECTION();
7950 48 : }
7951 :
7952 : /*
7953 : * Write an OVERWRITE_CONTRECORD message.
7954 : *
7955 : * When on WAL replay we expect a continuation record at the start of a page
7956 : * that is not there, recovery ends and WAL writing resumes at that point.
7957 : * But it's wrong to resume writing new WAL back at the start of the record
7958 : * that was broken, because downstream consumers of that WAL (physical
7959 : * replicas) are not prepared to "rewind". So the first action after
7960 : * finishing replay of all valid WAL must be to write a record of this type
7961 : * at the point where the contrecord was missing; to support xlogreader
7962 : * detecting the special case, XLP_FIRST_IS_OVERWRITE_CONTRECORD is also added
7963 : * to the page header where the record occurs. xlogreader has an ad-hoc
7964 : * mechanism to report metadata about the broken record, which is what we
7965 : * use here.
7966 : *
7967 : * At replay time, XLP_FIRST_IS_OVERWRITE_CONTRECORD instructs xlogreader to
7968 : * skip the record it was reading, and pass back the LSN of the skipped
7969 : * record, so that its caller can verify (on "replay" of that record) that the
7970 : * XLOG_OVERWRITE_CONTRECORD matches what was effectively overwritten.
7971 : *
7972 : * 'aborted_lsn' is the beginning position of the record that was incomplete.
7973 : * It is included in the WAL record. 'pagePtr' and 'newTLI' point to the
7974 : * beginning of the XLOG page where the record is to be inserted. They must
7975 : * match the current WAL insert position, they're passed here just so that we
7976 : * can verify that.
7977 : */
7978 : static XLogRecPtr
7979 10 : CreateOverwriteContrecordRecord(XLogRecPtr aborted_lsn, XLogRecPtr pagePtr,
7980 : TimeLineID newTLI)
7981 : {
7982 : xl_overwrite_contrecord xlrec;
7983 : XLogRecPtr recptr;
7984 : XLogPageHeader pagehdr;
7985 : XLogRecPtr startPos;
7986 :
7987 : /* sanity checks */
7988 10 : if (!RecoveryInProgress())
7989 0 : elog(ERROR, "can only be used at end of recovery");
7990 10 : if (pagePtr % XLOG_BLCKSZ != 0)
7991 0 : elog(ERROR, "invalid position for missing continuation record %X/%08X",
7992 : LSN_FORMAT_ARGS(pagePtr));
7993 :
7994 : /* The current WAL insert position should be right after the page header */
7995 10 : startPos = pagePtr;
7996 10 : if (XLogSegmentOffset(startPos, wal_segment_size) == 0)
7997 1 : startPos += SizeOfXLogLongPHD;
7998 : else
7999 9 : startPos += SizeOfXLogShortPHD;
8000 10 : recptr = GetXLogInsertRecPtr();
8001 10 : if (recptr != startPos)
8002 0 : elog(ERROR, "invalid WAL insert position %X/%08X for OVERWRITE_CONTRECORD",
8003 : LSN_FORMAT_ARGS(recptr));
8004 :
8005 10 : START_CRIT_SECTION();
8006 :
8007 : /*
8008 : * Initialize the XLOG page header (by GetXLogBuffer), and set the
8009 : * XLP_FIRST_IS_OVERWRITE_CONTRECORD flag.
8010 : *
8011 : * No other backend is allowed to write WAL yet, so acquiring the WAL
8012 : * insertion lock is just pro forma.
8013 : */
8014 10 : WALInsertLockAcquire();
8015 10 : pagehdr = (XLogPageHeader) GetXLogBuffer(pagePtr, newTLI);
8016 10 : pagehdr->xlp_info |= XLP_FIRST_IS_OVERWRITE_CONTRECORD;
8017 10 : WALInsertLockRelease();
8018 :
8019 : /*
8020 : * Insert the XLOG_OVERWRITE_CONTRECORD record as the first record on the
8021 : * page. We know it becomes the first record, because no other backend is
8022 : * allowed to write WAL yet.
8023 : */
8024 10 : XLogBeginInsert();
8025 10 : xlrec.overwritten_lsn = aborted_lsn;
8026 10 : xlrec.overwrite_time = GetCurrentTimestamp();
8027 10 : XLogRegisterData(&xlrec, sizeof(xl_overwrite_contrecord));
8028 10 : recptr = XLogInsert(RM_XLOG_ID, XLOG_OVERWRITE_CONTRECORD);
8029 :
8030 : /* check that the record was inserted to the right place */
8031 10 : if (ProcLastRecPtr != startPos)
8032 0 : elog(ERROR, "OVERWRITE_CONTRECORD was inserted to unexpected position %X/%08X",
8033 : LSN_FORMAT_ARGS(ProcLastRecPtr));
8034 :
8035 10 : XLogFlush(recptr);
8036 :
8037 10 : END_CRIT_SECTION();
8038 :
8039 10 : return recptr;
8040 : }
8041 :
8042 : /*
8043 : * Flush all data in shared memory to disk, and fsync
8044 : *
8045 : * This is the common code shared between regular checkpoints and
8046 : * recovery restartpoints.
8047 : */
8048 : static void
8049 1931 : CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
8050 : {
8051 1931 : CheckPointRelationMap();
8052 1931 : CheckPointReplicationSlots(flags & CHECKPOINT_IS_SHUTDOWN);
8053 1931 : CheckPointSnapBuild();
8054 1931 : CheckPointLogicalRewriteHeap();
8055 1931 : CheckPointReplicationOrigin();
8056 :
8057 : /* Write out all dirty data in SLRUs and the main buffer pool */
8058 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags);
8059 1931 : CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
8060 1931 : CheckPointCLOG();
8061 1931 : CheckPointCommitTs();
8062 1931 : CheckPointSUBTRANS();
8063 1931 : CheckPointMultiXact();
8064 1931 : CheckPointPredicate();
8065 1931 : CheckPointBuffers(flags);
8066 :
8067 : /* Perform all queued up fsyncs */
8068 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_SYNC_START();
8069 1931 : CheckpointStats.ckpt_sync_t = GetCurrentTimestamp();
8070 1931 : ProcessSyncRequests();
8071 1931 : CheckpointStats.ckpt_sync_end_t = GetCurrentTimestamp();
8072 : TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE();
8073 :
8074 : /* We deliberately delay 2PC checkpointing as long as possible */
8075 1931 : CheckPointTwoPhase(checkPointRedo);
8076 1931 : }
8077 :
8078 : /*
8079 : * Save a checkpoint for recovery restart if appropriate
8080 : *
8081 : * This function is called each time a checkpoint record is read from XLOG.
8082 : * It must determine whether the checkpoint represents a safe restartpoint or
8083 : * not. If so, the checkpoint record is stashed in shared memory so that
8084 : * CreateRestartPoint can consult it. (Note that the latter function is
8085 : * executed by the checkpointer, while this one will be executed by the
8086 : * startup process.)
8087 : */
8088 : static void
8089 743 : RecoveryRestartPoint(const CheckPoint *checkPoint, XLogReaderState *record)
8090 : {
8091 : /*
8092 : * Also refrain from creating a restartpoint if we have seen any
8093 : * references to non-existent pages. Restarting recovery from the
8094 : * restartpoint would not see the references, so we would lose the
8095 : * cross-check that the pages belonged to a relation that was dropped
8096 : * later.
8097 : */
8098 743 : if (XLogHaveInvalidPages())
8099 : {
8100 0 : elog(DEBUG2,
8101 : "could not record restart point at %X/%08X because there are unresolved references to invalid pages",
8102 : LSN_FORMAT_ARGS(checkPoint->redo));
8103 0 : return;
8104 : }
8105 :
8106 : /*
8107 : * Copy the checkpoint record to shared memory, so that checkpointer can
8108 : * work out the next time it wants to perform a restartpoint.
8109 : */
8110 743 : SpinLockAcquire(&XLogCtl->info_lck);
8111 743 : XLogCtl->lastCheckPointRecPtr = record->ReadRecPtr;
8112 743 : XLogCtl->lastCheckPointEndPtr = record->EndRecPtr;
8113 743 : XLogCtl->lastCheckPoint = *checkPoint;
8114 743 : SpinLockRelease(&XLogCtl->info_lck);
8115 : }
8116 :
8117 : /*
8118 : * Establish a restartpoint if possible.
8119 : *
8120 : * This is similar to CreateCheckPoint, but is used during WAL recovery
8121 : * to establish a point from which recovery can roll forward without
8122 : * replaying the entire recovery log.
8123 : *
8124 : * Returns true if a new restartpoint was established. We can only establish
8125 : * a restartpoint if we have replayed a safe checkpoint record since last
8126 : * restartpoint.
8127 : */
8128 : bool
8129 614 : CreateRestartPoint(int flags)
8130 : {
8131 : XLogRecPtr lastCheckPointRecPtr;
8132 : XLogRecPtr lastCheckPointEndPtr;
8133 : CheckPoint lastCheckPoint;
8134 : XLogRecPtr PriorRedoPtr;
8135 : XLogRecPtr receivePtr;
8136 : XLogRecPtr replayPtr;
8137 : TimeLineID replayTLI;
8138 : XLogRecPtr endptr;
8139 : XLogSegNo _logSegNo;
8140 : TimestampTz xtime;
8141 :
8142 : /* Concurrent checkpoint/restartpoint cannot happen */
8143 : Assert(!IsUnderPostmaster || MyBackendType == B_CHECKPOINTER);
8144 :
8145 : /* Get a local copy of the last safe checkpoint record. */
8146 614 : SpinLockAcquire(&XLogCtl->info_lck);
8147 614 : lastCheckPointRecPtr = XLogCtl->lastCheckPointRecPtr;
8148 614 : lastCheckPointEndPtr = XLogCtl->lastCheckPointEndPtr;
8149 614 : lastCheckPoint = XLogCtl->lastCheckPoint;
8150 614 : SpinLockRelease(&XLogCtl->info_lck);
8151 :
8152 : /*
8153 : * Check that we're still in recovery mode. It's ok if we exit recovery
8154 : * mode after this check, the restart point is valid anyway.
8155 : */
8156 614 : if (!RecoveryInProgress())
8157 : {
8158 0 : ereport(DEBUG2,
8159 : (errmsg_internal("skipping restartpoint, recovery has already ended")));
8160 0 : return false;
8161 : }
8162 :
8163 : /*
8164 : * If the last checkpoint record we've replayed is already our last
8165 : * restartpoint, we can't perform a new restart point. We still update
8166 : * minRecoveryPoint in that case, so that if this is a shutdown restart
8167 : * point, we won't start up earlier than before. That's not strictly
8168 : * necessary, but when hot standby is enabled, it would be rather weird if
8169 : * the database opened up for read-only connections at a point-in-time
8170 : * before the last shutdown. Such time travel is still possible in case of
8171 : * immediate shutdown, though.
8172 : *
8173 : * We don't explicitly advance minRecoveryPoint when we do create a
8174 : * restartpoint. It's assumed that flushing the buffers will do that as a
8175 : * side-effect.
8176 : */
8177 614 : if (!XLogRecPtrIsValid(lastCheckPointRecPtr) ||
8178 283 : lastCheckPoint.redo <= ControlFile->checkPointCopy.redo)
8179 : {
8180 404 : ereport(DEBUG2,
8181 : errmsg_internal("skipping restartpoint, already performed at %X/%08X",
8182 : LSN_FORMAT_ARGS(lastCheckPoint.redo)));
8183 :
8184 404 : UpdateMinRecoveryPoint(InvalidXLogRecPtr, true);
8185 404 : if (flags & CHECKPOINT_IS_SHUTDOWN)
8186 : {
8187 36 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8188 36 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
8189 36 : UpdateControlFile();
8190 36 : LWLockRelease(ControlFileLock);
8191 : }
8192 404 : return false;
8193 : }
8194 :
8195 : /*
8196 : * Update the shared RedoRecPtr so that the startup process can calculate
8197 : * the number of segments replayed since last restartpoint, and request a
8198 : * restartpoint if it exceeds CheckPointSegments.
8199 : *
8200 : * Like in CreateCheckPoint(), hold off insertions to update it, although
8201 : * during recovery this is just pro forma, because no WAL insertions are
8202 : * happening.
8203 : */
8204 210 : WALInsertLockAcquireExclusive();
8205 210 : RedoRecPtr = XLogCtl->Insert.RedoRecPtr = lastCheckPoint.redo;
8206 210 : WALInsertLockRelease();
8207 :
8208 : /* Also update the info_lck-protected copy */
8209 210 : SpinLockAcquire(&XLogCtl->info_lck);
8210 210 : XLogCtl->RedoRecPtr = lastCheckPoint.redo;
8211 210 : SpinLockRelease(&XLogCtl->info_lck);
8212 :
8213 : /*
8214 : * Prepare to accumulate statistics.
8215 : *
8216 : * Note: because it is possible for log_checkpoints to change while a
8217 : * checkpoint proceeds, we always accumulate stats, even if
8218 : * log_checkpoints is currently off.
8219 : */
8220 2310 : MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
8221 210 : CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
8222 :
8223 210 : if (log_checkpoints)
8224 210 : LogCheckpointStart(flags, true);
8225 :
8226 : /* Update the process title */
8227 210 : update_checkpoint_display(flags, true, false);
8228 :
8229 210 : CheckPointGuts(lastCheckPoint.redo, flags);
8230 :
8231 : /*
8232 : * This location needs to be after CheckPointGuts() to ensure that some
8233 : * work has already happened during this checkpoint.
8234 : */
8235 210 : INJECTION_POINT("create-restart-point", NULL);
8236 :
8237 : /*
8238 : * Remember the prior checkpoint's redo ptr for
8239 : * UpdateCheckPointDistanceEstimate()
8240 : */
8241 210 : PriorRedoPtr = ControlFile->checkPointCopy.redo;
8242 :
8243 : /*
8244 : * Update pg_control, using current time. Check that it still shows an
8245 : * older checkpoint, else do nothing; this is a quick hack to make sure
8246 : * nothing really bad happens if somehow we get here after the
8247 : * end-of-recovery checkpoint.
8248 : */
8249 210 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8250 210 : if (ControlFile->checkPointCopy.redo < lastCheckPoint.redo)
8251 : {
8252 : /*
8253 : * Update the checkpoint information. We do this even if the cluster
8254 : * does not show DB_IN_ARCHIVE_RECOVERY to match with the set of WAL
8255 : * segments recycled below.
8256 : */
8257 210 : ControlFile->checkPoint = lastCheckPointRecPtr;
8258 210 : ControlFile->checkPointCopy = lastCheckPoint;
8259 :
8260 : /*
8261 : * Ensure minRecoveryPoint is past the checkpoint record and update it
8262 : * if the control file still shows DB_IN_ARCHIVE_RECOVERY. Normally,
8263 : * this will have happened already while writing out dirty buffers,
8264 : * but not necessarily - e.g. because no buffers were dirtied. We do
8265 : * this because a backup performed in recovery uses minRecoveryPoint
8266 : * to determine which WAL files must be included in the backup, and
8267 : * the file (or files) containing the checkpoint record must be
8268 : * included, at a minimum. Note that for an ordinary restart of
8269 : * recovery there's no value in having the minimum recovery point any
8270 : * earlier than this anyway, because redo will begin just after the
8271 : * checkpoint record.
8272 : */
8273 210 : if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
8274 : {
8275 210 : if (ControlFile->minRecoveryPoint < lastCheckPointEndPtr)
8276 : {
8277 20 : ControlFile->minRecoveryPoint = lastCheckPointEndPtr;
8278 20 : ControlFile->minRecoveryPointTLI = lastCheckPoint.ThisTimeLineID;
8279 :
8280 : /* update local copy */
8281 20 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
8282 20 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
8283 : }
8284 210 : if (flags & CHECKPOINT_IS_SHUTDOWN)
8285 24 : ControlFile->state = DB_SHUTDOWNED_IN_RECOVERY;
8286 : }
8287 :
8288 : /* we shall start with the latest checksum version */
8289 210 : ControlFile->data_checksum_version = lastCheckPoint.dataChecksumState;
8290 :
8291 210 : UpdateControlFile();
8292 : }
8293 210 : LWLockRelease(ControlFileLock);
8294 :
8295 : /*
8296 : * Update the average distance between checkpoints/restartpoints if the
8297 : * prior checkpoint exists.
8298 : */
8299 210 : if (XLogRecPtrIsValid(PriorRedoPtr))
8300 210 : UpdateCheckPointDistanceEstimate(RedoRecPtr - PriorRedoPtr);
8301 :
8302 : /*
8303 : * Delete old log files, those no longer needed for last restartpoint to
8304 : * prevent the disk holding the xlog from growing full.
8305 : */
8306 210 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
8307 :
8308 : /*
8309 : * Retreat _logSegNo using the current end of xlog replayed or received,
8310 : * whichever is later.
8311 : */
8312 210 : receivePtr = GetWalRcvFlushRecPtr(NULL, NULL);
8313 210 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
8314 210 : endptr = (receivePtr < replayPtr) ? replayPtr : receivePtr;
8315 210 : KeepLogSeg(endptr, &_logSegNo);
8316 :
8317 210 : INJECTION_POINT("restartpoint-before-slot-invalidation", NULL);
8318 :
8319 210 : if (InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_REMOVED | RS_INVAL_IDLE_TIMEOUT,
8320 : _logSegNo, InvalidOid,
8321 : InvalidTransactionId))
8322 : {
8323 : /*
8324 : * Some slots have been invalidated; recalculate the old-segment
8325 : * horizon, starting again from RedoRecPtr.
8326 : */
8327 1 : XLByteToSeg(RedoRecPtr, _logSegNo, wal_segment_size);
8328 1 : KeepLogSeg(endptr, &_logSegNo);
8329 : }
8330 210 : _logSegNo--;
8331 :
8332 : /*
8333 : * Try to recycle segments on a useful timeline. If we've been promoted
8334 : * since the beginning of this restartpoint, use the new timeline chosen
8335 : * at end of recovery. If we're still in recovery, use the timeline we're
8336 : * currently replaying.
8337 : *
8338 : * There is no guarantee that the WAL segments will be useful on the
8339 : * current timeline; if recovery proceeds to a new timeline right after
8340 : * this, the pre-allocated WAL segments on this timeline will not be used,
8341 : * and will go wasted until recycled on the next restartpoint. We'll live
8342 : * with that.
8343 : */
8344 210 : if (!RecoveryInProgress())
8345 0 : replayTLI = XLogCtl->InsertTimeLineID;
8346 :
8347 210 : RemoveOldXlogFiles(_logSegNo, RedoRecPtr, endptr, replayTLI);
8348 :
8349 : /*
8350 : * Make more log segments if needed. (Do this after recycling old log
8351 : * segments, since that may supply some of the needed files.)
8352 : */
8353 210 : PreallocXlogFiles(endptr, replayTLI);
8354 :
8355 : /*
8356 : * Truncate pg_subtrans if possible. We can throw away all data before
8357 : * the oldest XMIN of any running transaction. No future transaction will
8358 : * attempt to reference any pg_subtrans entry older than that (see Asserts
8359 : * in subtrans.c). When hot standby is disabled, though, we mustn't do
8360 : * this because StartupSUBTRANS hasn't been called yet.
8361 : */
8362 210 : if (EnableHotStandby)
8363 210 : TruncateSUBTRANS(GetOldestTransactionIdConsideredRunning());
8364 :
8365 : /* Real work is done; log and update stats. */
8366 210 : LogCheckpointEnd(true, flags);
8367 :
8368 : /* Reset the process title */
8369 210 : update_checkpoint_display(flags, true, true);
8370 :
8371 210 : xtime = GetLatestXTime();
8372 210 : ereport((log_checkpoints ? LOG : DEBUG2),
8373 : errmsg("recovery restart point at %X/%08X",
8374 : LSN_FORMAT_ARGS(lastCheckPoint.redo)),
8375 : xtime ? errdetail("Last completed transaction was at log time %s.",
8376 : timestamptz_to_str(xtime)) : 0);
8377 :
8378 : /*
8379 : * Finally, execute archive_cleanup_command, if any.
8380 : */
8381 210 : if (archiveCleanupCommand && strcmp(archiveCleanupCommand, "") != 0)
8382 0 : ExecuteRecoveryCommand(archiveCleanupCommand,
8383 : "archive_cleanup_command",
8384 : false,
8385 : WAIT_EVENT_ARCHIVE_CLEANUP_COMMAND);
8386 :
8387 210 : return true;
8388 : }
8389 :
8390 : /*
8391 : * Report availability of WAL for the given target LSN
8392 : * (typically a slot's restart_lsn)
8393 : *
8394 : * Returns one of the following enum values:
8395 : *
8396 : * * WALAVAIL_RESERVED means targetLSN is available and it is in the range of
8397 : * max_wal_size.
8398 : *
8399 : * * WALAVAIL_EXTENDED means it is still available by preserving extra
8400 : * segments beyond max_wal_size. If max_slot_wal_keep_size is smaller
8401 : * than max_wal_size, this state is not returned.
8402 : *
8403 : * * WALAVAIL_UNRESERVED means it is being lost and the next checkpoint will
8404 : * remove reserved segments. The walsender using this slot may return to the
8405 : * above.
8406 : *
8407 : * * WALAVAIL_REMOVED means it has been removed. A replication stream on
8408 : * a slot with this LSN cannot continue. (Any associated walsender
8409 : * processes should have been terminated already.)
8410 : *
8411 : * * WALAVAIL_INVALID_LSN means the slot hasn't been set to reserve WAL.
8412 : */
8413 : WALAvailability
8414 628 : GetWALAvailability(XLogRecPtr targetLSN)
8415 : {
8416 : XLogRecPtr currpos; /* current write LSN */
8417 : XLogSegNo currSeg; /* segid of currpos */
8418 : XLogSegNo targetSeg; /* segid of targetLSN */
8419 : XLogSegNo oldestSeg; /* actual oldest segid */
8420 : XLogSegNo oldestSegMaxWalSize; /* oldest segid kept by max_wal_size */
8421 : XLogSegNo oldestSlotSeg; /* oldest segid kept by slot */
8422 : uint64 keepSegs;
8423 :
8424 : /*
8425 : * slot does not reserve WAL. Either deactivated, or has never been active
8426 : */
8427 628 : if (!XLogRecPtrIsValid(targetLSN))
8428 27 : return WALAVAIL_INVALID_LSN;
8429 :
8430 : /*
8431 : * Calculate the oldest segment currently reserved by all slots,
8432 : * considering wal_keep_size and max_slot_wal_keep_size. Initialize
8433 : * oldestSlotSeg to the current segment.
8434 : */
8435 601 : currpos = GetXLogWriteRecPtr();
8436 601 : XLByteToSeg(currpos, oldestSlotSeg, wal_segment_size);
8437 601 : KeepLogSeg(currpos, &oldestSlotSeg);
8438 :
8439 : /*
8440 : * Find the oldest extant segment file. We get 1 until checkpoint removes
8441 : * the first WAL segment file since startup, which causes the status being
8442 : * wrong under certain abnormal conditions but that doesn't actually harm.
8443 : */
8444 601 : oldestSeg = XLogGetLastRemovedSegno() + 1;
8445 :
8446 : /* calculate oldest segment by max_wal_size */
8447 601 : XLByteToSeg(currpos, currSeg, wal_segment_size);
8448 601 : keepSegs = ConvertToXSegs(max_wal_size_mb, wal_segment_size) + 1;
8449 :
8450 601 : if (currSeg > keepSegs)
8451 8 : oldestSegMaxWalSize = currSeg - keepSegs;
8452 : else
8453 593 : oldestSegMaxWalSize = 1;
8454 :
8455 : /* the segment we care about */
8456 601 : XLByteToSeg(targetLSN, targetSeg, wal_segment_size);
8457 :
8458 : /*
8459 : * No point in returning reserved or extended status values if the
8460 : * targetSeg is known to be lost.
8461 : */
8462 601 : if (targetSeg >= oldestSlotSeg)
8463 : {
8464 : /* show "reserved" when targetSeg is within max_wal_size */
8465 600 : if (targetSeg >= oldestSegMaxWalSize)
8466 598 : return WALAVAIL_RESERVED;
8467 :
8468 : /* being retained by slots exceeding max_wal_size */
8469 2 : return WALAVAIL_EXTENDED;
8470 : }
8471 :
8472 : /* WAL segments are no longer retained but haven't been removed yet */
8473 1 : if (targetSeg >= oldestSeg)
8474 1 : return WALAVAIL_UNRESERVED;
8475 :
8476 : /* Definitely lost */
8477 0 : return WALAVAIL_REMOVED;
8478 : }
8479 :
8480 :
8481 : /*
8482 : * Retreat *logSegNo to the last segment that we need to retain because of
8483 : * either wal_keep_size or replication slots.
8484 : *
8485 : * This is calculated by subtracting wal_keep_size from the given xlog
8486 : * location, recptr and by making sure that that result is below the
8487 : * requirement of replication slots. For the latter criterion we do consider
8488 : * the effects of max_slot_wal_keep_size: reserve at most that much space back
8489 : * from recptr.
8490 : *
8491 : * Note about replication slots: if this function calculates a value
8492 : * that's further ahead than what slots need reserved, then affected
8493 : * slots need to be invalidated and this function invoked again.
8494 : * XXX it might be a good idea to rewrite this function so that
8495 : * invalidation is optionally done here, instead.
8496 : */
8497 : static void
8498 2537 : KeepLogSeg(XLogRecPtr recptr, XLogSegNo *logSegNo)
8499 : {
8500 : XLogSegNo currSegNo;
8501 : XLogSegNo segno;
8502 : XLogRecPtr keep;
8503 :
8504 2537 : XLByteToSeg(recptr, currSegNo, wal_segment_size);
8505 2537 : segno = currSegNo;
8506 :
8507 : /* Calculate how many segments are kept by slots. */
8508 2537 : keep = XLogGetReplicationSlotMinimumLSN();
8509 2537 : if (XLogRecPtrIsValid(keep) && keep < recptr)
8510 : {
8511 770 : XLByteToSeg(keep, segno, wal_segment_size);
8512 :
8513 : /*
8514 : * Account for max_slot_wal_keep_size to avoid keeping more than
8515 : * configured. However, don't do that during a binary upgrade: if
8516 : * slots were to be invalidated because of this, it would not be
8517 : * possible to preserve logical ones during the upgrade.
8518 : */
8519 770 : if (max_slot_wal_keep_size_mb >= 0 && !IsBinaryUpgrade)
8520 : {
8521 : uint64 slot_keep_segs;
8522 :
8523 23 : slot_keep_segs =
8524 23 : ConvertToXSegs(max_slot_wal_keep_size_mb, wal_segment_size);
8525 :
8526 23 : if (currSegNo - segno > slot_keep_segs)
8527 6 : segno = currSegNo - slot_keep_segs;
8528 : }
8529 : }
8530 :
8531 : /*
8532 : * If WAL summarization is in use, don't remove WAL that has yet to be
8533 : * summarized.
8534 : */
8535 2537 : keep = GetOldestUnsummarizedLSN(NULL, NULL);
8536 2537 : if (XLogRecPtrIsValid(keep))
8537 : {
8538 : XLogSegNo unsummarized_segno;
8539 :
8540 7 : XLByteToSeg(keep, unsummarized_segno, wal_segment_size);
8541 7 : if (unsummarized_segno < segno)
8542 7 : segno = unsummarized_segno;
8543 : }
8544 :
8545 : /* but, keep at least wal_keep_size if that's set */
8546 2537 : if (wal_keep_size_mb > 0)
8547 : {
8548 : uint64 keep_segs;
8549 :
8550 74 : keep_segs = ConvertToXSegs(wal_keep_size_mb, wal_segment_size);
8551 74 : if (currSegNo - segno < keep_segs)
8552 : {
8553 : /* avoid underflow, don't go below 1 */
8554 74 : if (currSegNo <= keep_segs)
8555 70 : segno = 1;
8556 : else
8557 4 : segno = currSegNo - keep_segs;
8558 : }
8559 : }
8560 :
8561 : /* don't delete WAL segments newer than the calculated segment */
8562 2537 : if (segno < *logSegNo)
8563 362 : *logSegNo = segno;
8564 2537 : }
8565 :
8566 : /*
8567 : * Write a NEXTOID log record
8568 : */
8569 : void
8570 690 : XLogPutNextOid(Oid nextOid)
8571 : {
8572 690 : XLogBeginInsert();
8573 690 : XLogRegisterData(&nextOid, sizeof(Oid));
8574 690 : (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID);
8575 :
8576 : /*
8577 : * We need not flush the NEXTOID record immediately, because any of the
8578 : * just-allocated OIDs could only reach disk as part of a tuple insert or
8579 : * update that would have its own XLOG record that must follow the NEXTOID
8580 : * record. Therefore, the standard buffer LSN interlock applied to those
8581 : * records will ensure no such OID reaches disk before the NEXTOID record
8582 : * does.
8583 : *
8584 : * Note, however, that the above statement only covers state "within" the
8585 : * database. When we use a generated OID as a file or directory name, we
8586 : * are in a sense violating the basic WAL rule, because that filesystem
8587 : * change may reach disk before the NEXTOID WAL record does. The impact
8588 : * of this is that if a database crash occurs immediately afterward, we
8589 : * might after restart re-generate the same OID and find that it conflicts
8590 : * with the leftover file or directory. But since for safety's sake we
8591 : * always loop until finding a nonconflicting filename, this poses no real
8592 : * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
8593 : */
8594 690 : }
8595 :
8596 : /*
8597 : * Write an XLOG SWITCH record.
8598 : *
8599 : * Here we just blindly issue an XLogInsert request for the record.
8600 : * All the magic happens inside XLogInsert.
8601 : *
8602 : * The return value is either the end+1 address of the switch record,
8603 : * or the end+1 address of the prior segment if we did not need to
8604 : * write a switch record because we are already at segment start.
8605 : */
8606 : XLogRecPtr
8607 814 : RequestXLogSwitch(bool mark_unimportant)
8608 : {
8609 : XLogRecPtr RecPtr;
8610 :
8611 : /* XLOG SWITCH has no data */
8612 814 : XLogBeginInsert();
8613 :
8614 814 : if (mark_unimportant)
8615 0 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8616 814 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH);
8617 :
8618 814 : return RecPtr;
8619 : }
8620 :
8621 : /*
8622 : * Write a RESTORE POINT record
8623 : */
8624 : XLogRecPtr
8625 3 : XLogRestorePoint(const char *rpName)
8626 : {
8627 : XLogRecPtr RecPtr;
8628 : xl_restore_point xlrec;
8629 :
8630 3 : xlrec.rp_time = GetCurrentTimestamp();
8631 3 : strlcpy(xlrec.rp_name, rpName, MAXFNAMELEN);
8632 :
8633 3 : XLogBeginInsert();
8634 3 : XLogRegisterData(&xlrec, sizeof(xl_restore_point));
8635 :
8636 3 : RecPtr = XLogInsert(RM_XLOG_ID, XLOG_RESTORE_POINT);
8637 :
8638 3 : ereport(LOG,
8639 : errmsg("restore point \"%s\" created at %X/%08X",
8640 : rpName, LSN_FORMAT_ARGS(RecPtr)));
8641 :
8642 3 : return RecPtr;
8643 : }
8644 :
8645 : /*
8646 : * Write an empty XLOG record to assign a distinct LSN.
8647 : *
8648 : * This is used by some index AMs when building indexes on permanent relations
8649 : * with wal_level=minimal. In that scenario, WAL-logging will start after
8650 : * commit, but the index AM needs distinct LSNs to detect concurrent page
8651 : * modifications. When the current WAL insert position hasn't advanced since
8652 : * the last call, we emit a dummy record to ensure we get a new, distinct LSN.
8653 : */
8654 : XLogRecPtr
8655 112021 : XLogAssignLSN(void)
8656 : {
8657 112021 : int dummy = 0;
8658 :
8659 : /*
8660 : * Records other than XLOG_SWITCH must have content. We use an integer 0
8661 : * to satisfy this restriction.
8662 : */
8663 112021 : XLogBeginInsert();
8664 112021 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
8665 112021 : XLogRegisterData(&dummy, sizeof(dummy));
8666 112021 : return XLogInsert(RM_XLOG_ID, XLOG_ASSIGN_LSN);
8667 : }
8668 :
8669 : /*
8670 : * Check if any of the GUC parameters that are critical for hot standby
8671 : * have changed, and update the value in pg_control file if necessary.
8672 : */
8673 : static void
8674 1007 : XLogReportParameters(void)
8675 : {
8676 1007 : if (wal_level != ControlFile->wal_level ||
8677 733 : wal_log_hints != ControlFile->wal_log_hints ||
8678 644 : MaxConnections != ControlFile->MaxConnections ||
8679 643 : max_worker_processes != ControlFile->max_worker_processes ||
8680 640 : max_wal_senders != ControlFile->max_wal_senders ||
8681 613 : max_prepared_xacts != ControlFile->max_prepared_xacts ||
8682 509 : max_locks_per_xact != ControlFile->max_locks_per_xact ||
8683 509 : track_commit_timestamp != ControlFile->track_commit_timestamp)
8684 : {
8685 : /*
8686 : * The change in number of backend slots doesn't need to be WAL-logged
8687 : * if archiving is not enabled, as you can't start archive recovery
8688 : * with wal_level=minimal anyway. We don't really care about the
8689 : * values in pg_control either if wal_level=minimal, but seems better
8690 : * to keep them up-to-date to avoid confusion.
8691 : */
8692 510 : if (wal_level != ControlFile->wal_level || XLogIsNeeded())
8693 : {
8694 : xl_parameter_change xlrec;
8695 : XLogRecPtr recptr;
8696 :
8697 483 : xlrec.MaxConnections = MaxConnections;
8698 483 : xlrec.max_worker_processes = max_worker_processes;
8699 483 : xlrec.max_wal_senders = max_wal_senders;
8700 483 : xlrec.max_prepared_xacts = max_prepared_xacts;
8701 483 : xlrec.max_locks_per_xact = max_locks_per_xact;
8702 483 : xlrec.wal_level = wal_level;
8703 483 : xlrec.wal_log_hints = wal_log_hints;
8704 483 : xlrec.track_commit_timestamp = track_commit_timestamp;
8705 :
8706 483 : XLogBeginInsert();
8707 483 : XLogRegisterData(&xlrec, sizeof(xlrec));
8708 :
8709 483 : recptr = XLogInsert(RM_XLOG_ID, XLOG_PARAMETER_CHANGE);
8710 483 : XLogFlush(recptr);
8711 : }
8712 :
8713 510 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8714 :
8715 510 : ControlFile->MaxConnections = MaxConnections;
8716 510 : ControlFile->max_worker_processes = max_worker_processes;
8717 510 : ControlFile->max_wal_senders = max_wal_senders;
8718 510 : ControlFile->max_prepared_xacts = max_prepared_xacts;
8719 510 : ControlFile->max_locks_per_xact = max_locks_per_xact;
8720 510 : ControlFile->wal_level = wal_level;
8721 510 : ControlFile->wal_log_hints = wal_log_hints;
8722 510 : ControlFile->track_commit_timestamp = track_commit_timestamp;
8723 510 : UpdateControlFile();
8724 :
8725 510 : LWLockRelease(ControlFileLock);
8726 : }
8727 1007 : }
8728 :
8729 : /*
8730 : * Log the new state of checksums
8731 : */
8732 : static void
8733 24 : XLogChecksums(uint32 new_type)
8734 : {
8735 : xl_checksum_state xlrec;
8736 : XLogRecPtr recptr;
8737 :
8738 24 : xlrec.new_checksum_state = new_type;
8739 :
8740 24 : XLogBeginInsert();
8741 24 : XLogRegisterData((char *) &xlrec, sizeof(xl_checksum_state));
8742 :
8743 24 : recptr = XLogInsert(RM_XLOG2_ID, XLOG2_CHECKSUMS);
8744 24 : XLogFlush(recptr);
8745 24 : }
8746 :
8747 : /*
8748 : * Update full_page_writes in shared memory, and write an
8749 : * XLOG_FPW_CHANGE record if necessary.
8750 : *
8751 : * Note: this function assumes there is no other process running
8752 : * concurrently that could update it.
8753 : */
8754 : void
8755 1720 : UpdateFullPageWrites(void)
8756 : {
8757 1720 : XLogCtlInsert *Insert = &XLogCtl->Insert;
8758 : bool recoveryInProgress;
8759 :
8760 : /*
8761 : * Do nothing if full_page_writes has not been changed.
8762 : *
8763 : * It's safe to check the shared full_page_writes without the lock,
8764 : * because we assume that there is no concurrently running process which
8765 : * can update it.
8766 : */
8767 1720 : if (fullPageWrites == Insert->fullPageWrites)
8768 1286 : return;
8769 :
8770 : /*
8771 : * Perform this outside critical section so that the WAL insert
8772 : * initialization done by RecoveryInProgress() doesn't trigger an
8773 : * assertion failure.
8774 : */
8775 434 : recoveryInProgress = RecoveryInProgress();
8776 :
8777 434 : START_CRIT_SECTION();
8778 :
8779 : /*
8780 : * It's always safe to take full page images, even when not strictly
8781 : * required, but not the other round. So if we're setting full_page_writes
8782 : * to true, first set it true and then write the WAL record. If we're
8783 : * setting it to false, first write the WAL record and then set the global
8784 : * flag.
8785 : */
8786 434 : if (fullPageWrites)
8787 : {
8788 421 : WALInsertLockAcquireExclusive();
8789 421 : Insert->fullPageWrites = true;
8790 421 : WALInsertLockRelease();
8791 : }
8792 :
8793 : /*
8794 : * Write an XLOG_FPW_CHANGE record. This allows us to keep track of
8795 : * full_page_writes during archive recovery, if required.
8796 : */
8797 434 : if (XLogStandbyInfoActive() && !recoveryInProgress)
8798 : {
8799 0 : XLogBeginInsert();
8800 0 : XLogRegisterData(&fullPageWrites, sizeof(bool));
8801 :
8802 0 : XLogInsert(RM_XLOG_ID, XLOG_FPW_CHANGE);
8803 : }
8804 :
8805 434 : if (!fullPageWrites)
8806 : {
8807 13 : WALInsertLockAcquireExclusive();
8808 13 : Insert->fullPageWrites = false;
8809 13 : WALInsertLockRelease();
8810 : }
8811 434 : END_CRIT_SECTION();
8812 : }
8813 :
8814 : /*
8815 : * XLOG resource manager's routines
8816 : *
8817 : * Definitions of info values are in include/catalog/pg_control.h, though
8818 : * not all record types are related to control file updates.
8819 : *
8820 : * NOTE: Some XLOG record types that are directly related to WAL recovery
8821 : * are handled in xlogrecovery_redo().
8822 : */
8823 : void
8824 114398 : xlog_redo(XLogReaderState *record)
8825 : {
8826 114398 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8827 114398 : XLogRecPtr lsn = record->EndRecPtr;
8828 :
8829 : /*
8830 : * In XLOG rmgr, backup blocks are only used by XLOG_FPI and
8831 : * XLOG_FPI_FOR_HINT records.
8832 : */
8833 : Assert(info == XLOG_FPI || info == XLOG_FPI_FOR_HINT ||
8834 : !XLogRecHasAnyBlockRefs(record));
8835 :
8836 114398 : if (info == XLOG_NEXTOID)
8837 : {
8838 : Oid nextOid;
8839 :
8840 : /*
8841 : * We used to try to take the maximum of TransamVariables->nextOid and
8842 : * the recorded nextOid, but that fails if the OID counter wraps
8843 : * around. Since no OID allocation should be happening during replay
8844 : * anyway, better to just believe the record exactly. We still take
8845 : * OidGenLock while setting the variable, just in case.
8846 : */
8847 97 : memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
8848 97 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8849 97 : TransamVariables->nextOid = nextOid;
8850 97 : TransamVariables->oidCount = 0;
8851 97 : LWLockRelease(OidGenLock);
8852 : }
8853 114301 : else if (info == XLOG_CHECKPOINT_SHUTDOWN)
8854 : {
8855 : CheckPoint checkPoint;
8856 : TimeLineID replayTLI;
8857 :
8858 42 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8859 : /* In a SHUTDOWN checkpoint, believe the counters exactly */
8860 42 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8861 42 : TransamVariables->nextXid = checkPoint.nextXid;
8862 42 : LWLockRelease(XidGenLock);
8863 42 : LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
8864 42 : TransamVariables->nextOid = checkPoint.nextOid;
8865 42 : TransamVariables->oidCount = 0;
8866 42 : LWLockRelease(OidGenLock);
8867 42 : MultiXactSetNextMXact(checkPoint.nextMulti,
8868 : checkPoint.nextMultiOffset);
8869 :
8870 42 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8871 : checkPoint.oldestMultiDB);
8872 :
8873 42 : SpinLockAcquire(&XLogCtl->info_lck);
8874 42 : XLogCtl->data_checksum_version = checkPoint.dataChecksumState;
8875 42 : SetLocalDataChecksumState(checkPoint.dataChecksumState);
8876 42 : SpinLockRelease(&XLogCtl->info_lck);
8877 :
8878 : /*
8879 : * No need to set oldestClogXid here as well; it'll be set when we
8880 : * redo an xl_clog_truncate if it changed since initialization.
8881 : */
8882 42 : SetTransactionIdLimit(checkPoint.oldestXid, checkPoint.oldestXidDB);
8883 :
8884 : /*
8885 : * If we see a shutdown checkpoint while waiting for an end-of-backup
8886 : * record, the backup was canceled and the end-of-backup record will
8887 : * never arrive.
8888 : */
8889 42 : if (ArchiveRecoveryRequested &&
8890 41 : XLogRecPtrIsValid(ControlFile->backupStartPoint) &&
8891 0 : !XLogRecPtrIsValid(ControlFile->backupEndPoint))
8892 0 : ereport(PANIC,
8893 : (errmsg("online backup was canceled, recovery cannot continue")));
8894 :
8895 : /*
8896 : * If we see a shutdown checkpoint, we know that nothing was running
8897 : * on the primary at this point. So fake-up an empty running-xacts
8898 : * record and use that here and now. Recover additional standby state
8899 : * for prepared transactions.
8900 : */
8901 42 : if (standbyState >= STANDBY_INITIALIZED)
8902 : {
8903 : TransactionId *xids;
8904 : int nxids;
8905 : TransactionId oldestActiveXID;
8906 : TransactionId latestCompletedXid;
8907 : RunningTransactionsData running;
8908 :
8909 39 : oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
8910 :
8911 : /* Update pg_subtrans entries for any prepared transactions */
8912 39 : StandbyRecoverPreparedTransactions();
8913 :
8914 : /*
8915 : * Construct a RunningTransactions snapshot representing a shut
8916 : * down server, with only prepared transactions still alive. We're
8917 : * never overflowed at this point because all subxids are listed
8918 : * with their parent prepared transactions.
8919 : */
8920 39 : running.xcnt = nxids;
8921 39 : running.subxcnt = 0;
8922 39 : running.subxid_status = SUBXIDS_IN_SUBTRANS;
8923 39 : running.nextXid = XidFromFullTransactionId(checkPoint.nextXid);
8924 39 : running.oldestRunningXid = oldestActiveXID;
8925 39 : latestCompletedXid = XidFromFullTransactionId(checkPoint.nextXid);
8926 39 : TransactionIdRetreat(latestCompletedXid);
8927 : Assert(TransactionIdIsNormal(latestCompletedXid));
8928 39 : running.latestCompletedXid = latestCompletedXid;
8929 39 : running.xids = xids;
8930 :
8931 39 : ProcArrayApplyRecoveryInfo(&running);
8932 : }
8933 :
8934 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
8935 42 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
8936 42 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
8937 42 : ControlFile->data_checksum_version = checkPoint.dataChecksumState;
8938 42 : LWLockRelease(ControlFileLock);
8939 :
8940 : /*
8941 : * We should've already switched to the new TLI before replaying this
8942 : * record.
8943 : */
8944 42 : (void) GetCurrentReplayRecPtr(&replayTLI);
8945 42 : if (checkPoint.ThisTimeLineID != replayTLI)
8946 0 : ereport(PANIC,
8947 : (errmsg("unexpected timeline ID %u (should be %u) in shutdown checkpoint record",
8948 : checkPoint.ThisTimeLineID, replayTLI)));
8949 :
8950 42 : RecoveryRestartPoint(&checkPoint, record);
8951 :
8952 : /*
8953 : * After replaying a checkpoint record, free all smgr objects.
8954 : * Otherwise we would never do so for dropped relations, as the
8955 : * startup does not process shared invalidation messages or call
8956 : * AtEOXact_SMgr().
8957 : */
8958 42 : smgrdestroyall();
8959 : }
8960 114259 : else if (info == XLOG_CHECKPOINT_ONLINE)
8961 : {
8962 : CheckPoint checkPoint;
8963 : TimeLineID replayTLI;
8964 701 : bool new_state = false;
8965 : int old_state;
8966 :
8967 701 : memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
8968 : /* In an ONLINE checkpoint, treat the XID counter as a minimum */
8969 701 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
8970 701 : if (FullTransactionIdPrecedes(TransamVariables->nextXid,
8971 : checkPoint.nextXid))
8972 0 : TransamVariables->nextXid = checkPoint.nextXid;
8973 701 : LWLockRelease(XidGenLock);
8974 :
8975 : /*
8976 : * We ignore the nextOid counter in an ONLINE checkpoint, preferring
8977 : * to track OID assignment through XLOG_NEXTOID records. The nextOid
8978 : * counter is from the start of the checkpoint and might well be stale
8979 : * compared to later XLOG_NEXTOID records. We could try to take the
8980 : * maximum of the nextOid counter and our latest value, but since
8981 : * there's no particular guarantee about the speed with which the OID
8982 : * counter wraps around, that's a risky thing to do. In any case,
8983 : * users of the nextOid counter are required to avoid assignment of
8984 : * duplicates, so that a somewhat out-of-date value should be safe.
8985 : */
8986 :
8987 : /* Handle multixact */
8988 701 : MultiXactAdvanceNextMXact(checkPoint.nextMulti,
8989 : checkPoint.nextMultiOffset);
8990 :
8991 : /*
8992 : * NB: This may perform multixact truncation when replaying WAL
8993 : * generated by an older primary.
8994 : */
8995 701 : MultiXactAdvanceOldest(checkPoint.oldestMulti,
8996 : checkPoint.oldestMultiDB);
8997 701 : if (TransactionIdPrecedes(TransamVariables->oldestXid,
8998 : checkPoint.oldestXid))
8999 0 : SetTransactionIdLimit(checkPoint.oldestXid,
9000 : checkPoint.oldestXidDB);
9001 : /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
9002 701 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9003 701 : ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
9004 701 : old_state = ControlFile->data_checksum_version;
9005 701 : ControlFile->data_checksum_version = checkPoint.dataChecksumState;
9006 701 : LWLockRelease(ControlFileLock);
9007 :
9008 : /* TLI should not change in an on-line checkpoint */
9009 701 : (void) GetCurrentReplayRecPtr(&replayTLI);
9010 701 : if (checkPoint.ThisTimeLineID != replayTLI)
9011 0 : ereport(PANIC,
9012 : (errmsg("unexpected timeline ID %u (should be %u) in online checkpoint record",
9013 : checkPoint.ThisTimeLineID, replayTLI)));
9014 :
9015 701 : RecoveryRestartPoint(&checkPoint, record);
9016 :
9017 : /*
9018 : * If the data checksum state change we need to emit a barrier.
9019 : */
9020 701 : SpinLockAcquire(&XLogCtl->info_lck);
9021 701 : XLogCtl->data_checksum_version = checkPoint.dataChecksumState;
9022 701 : if (checkPoint.dataChecksumState != old_state)
9023 3 : new_state = true;
9024 701 : SpinLockRelease(&XLogCtl->info_lck);
9025 :
9026 701 : if (new_state)
9027 3 : EmitAndWaitDataChecksumsBarrier(checkPoint.dataChecksumState);
9028 :
9029 : /*
9030 : * After replaying a checkpoint record, free all smgr objects.
9031 : * Otherwise we would never do so for dropped relations, as the
9032 : * startup does not process shared invalidation messages or call
9033 : * AtEOXact_SMgr().
9034 : */
9035 701 : smgrdestroyall();
9036 : }
9037 113558 : else if (info == XLOG_OVERWRITE_CONTRECORD)
9038 : {
9039 : /* nothing to do here, handled in xlogrecovery_redo() */
9040 : }
9041 113557 : else if (info == XLOG_END_OF_RECOVERY)
9042 : {
9043 : xl_end_of_recovery xlrec;
9044 : TimeLineID replayTLI;
9045 :
9046 11 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_end_of_recovery));
9047 :
9048 : /*
9049 : * For Hot Standby, we could treat this like a Shutdown Checkpoint,
9050 : * but this case is rarer and harder to test, so the benefit doesn't
9051 : * outweigh the potential extra cost of maintenance.
9052 : */
9053 :
9054 : /*
9055 : * We should've already switched to the new TLI before replaying this
9056 : * record.
9057 : */
9058 11 : (void) GetCurrentReplayRecPtr(&replayTLI);
9059 11 : if (xlrec.ThisTimeLineID != replayTLI)
9060 0 : ereport(PANIC,
9061 : (errmsg("unexpected timeline ID %u (should be %u) in end-of-recovery record",
9062 : xlrec.ThisTimeLineID, replayTLI)));
9063 : }
9064 113546 : else if (info == XLOG_NOOP)
9065 : {
9066 : /* nothing to do here */
9067 : }
9068 113546 : else if (info == XLOG_SWITCH)
9069 : {
9070 : /* nothing to do here */
9071 : }
9072 113089 : else if (info == XLOG_RESTORE_POINT)
9073 : {
9074 : /* nothing to do here, handled in xlogrecovery.c */
9075 : }
9076 113084 : else if (info == XLOG_ASSIGN_LSN)
9077 : {
9078 : /* nothing to do here, see XLogGetFakeLSN() */
9079 : }
9080 50757 : else if (info == XLOG_FPI || info == XLOG_FPI_FOR_HINT)
9081 : {
9082 : /*
9083 : * XLOG_FPI records contain nothing else but one or more block
9084 : * references. Every block reference must include a full-page image
9085 : * even if full_page_writes was disabled when the record was generated
9086 : * - otherwise there would be no point in this record.
9087 : *
9088 : * XLOG_FPI_FOR_HINT records are generated when a page needs to be
9089 : * WAL-logged because of a hint bit update. They are only generated
9090 : * when checksums and/or wal_log_hints are enabled. They may include
9091 : * no full-page images if full_page_writes was disabled when they were
9092 : * generated. In this case there is nothing to do here.
9093 : *
9094 : * No recovery conflicts are generated by these generic records - if a
9095 : * resource manager needs to generate conflicts, it has to define a
9096 : * separate WAL record type and redo routine.
9097 : */
9098 104675 : for (uint8 block_id = 0; block_id <= XLogRecMaxBlockId(record); block_id++)
9099 : {
9100 : Buffer buffer;
9101 :
9102 54774 : if (!XLogRecHasBlockImage(record, block_id))
9103 : {
9104 66 : if (info == XLOG_FPI)
9105 0 : elog(ERROR, "XLOG_FPI record did not contain a full-page image");
9106 66 : continue;
9107 : }
9108 :
9109 54708 : if (XLogReadBufferForRedo(record, block_id, &buffer) != BLK_RESTORED)
9110 0 : elog(ERROR, "unexpected XLogReadBufferForRedo result when restoring backup block");
9111 54708 : UnlockReleaseBuffer(buffer);
9112 : }
9113 : }
9114 856 : else if (info == XLOG_BACKUP_END)
9115 : {
9116 : /* nothing to do here, handled in xlogrecovery_redo() */
9117 : }
9118 759 : else if (info == XLOG_PARAMETER_CHANGE)
9119 : {
9120 : xl_parameter_change xlrec;
9121 :
9122 : /* Update our copy of the parameters in pg_control */
9123 39 : memcpy(&xlrec, XLogRecGetData(record), sizeof(xl_parameter_change));
9124 :
9125 39 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
9126 39 : ControlFile->MaxConnections = xlrec.MaxConnections;
9127 39 : ControlFile->max_worker_processes = xlrec.max_worker_processes;
9128 39 : ControlFile->max_wal_senders = xlrec.max_wal_senders;
9129 39 : ControlFile->max_prepared_xacts = xlrec.max_prepared_xacts;
9130 39 : ControlFile->max_locks_per_xact = xlrec.max_locks_per_xact;
9131 39 : ControlFile->wal_level = xlrec.wal_level;
9132 39 : ControlFile->wal_log_hints = xlrec.wal_log_hints;
9133 :
9134 : /*
9135 : * Update minRecoveryPoint to ensure that if recovery is aborted, we
9136 : * recover back up to this point before allowing hot standby again.
9137 : * This is important if the max_* settings are decreased, to ensure
9138 : * you don't run queries against the WAL preceding the change. The
9139 : * local copies cannot be updated as long as crash recovery is
9140 : * happening and we expect all the WAL to be replayed.
9141 : */
9142 39 : if (InArchiveRecovery)
9143 : {
9144 24 : LocalMinRecoveryPoint = ControlFile->minRecoveryPoint;
9145 24 : LocalMinRecoveryPointTLI = ControlFile->minRecoveryPointTLI;
9146 : }
9147 39 : if (XLogRecPtrIsValid(LocalMinRecoveryPoint) && LocalMinRecoveryPoint < lsn)
9148 : {
9149 : TimeLineID replayTLI;
9150 :
9151 12 : (void) GetCurrentReplayRecPtr(&replayTLI);
9152 12 : ControlFile->minRecoveryPoint = lsn;
9153 12 : ControlFile->minRecoveryPointTLI = replayTLI;
9154 : }
9155 :
9156 39 : CommitTsParameterChange(xlrec.track_commit_timestamp,
9157 39 : ControlFile->track_commit_timestamp);
9158 39 : ControlFile->track_commit_timestamp = xlrec.track_commit_timestamp;
9159 :
9160 39 : UpdateControlFile();
9161 39 : LWLockRelease(ControlFileLock);
9162 :
9163 : /* Check to see if any parameter change gives a problem on recovery */
9164 39 : CheckRequiredParameterValues();
9165 : }
9166 720 : else if (info == XLOG_FPW_CHANGE)
9167 : {
9168 : bool fpw;
9169 :
9170 0 : memcpy(&fpw, XLogRecGetData(record), sizeof(bool));
9171 :
9172 : /*
9173 : * Update the LSN of the last replayed XLOG_FPW_CHANGE record so that
9174 : * do_pg_backup_start() and do_pg_backup_stop() can check whether
9175 : * full_page_writes has been disabled during online backup.
9176 : */
9177 0 : if (!fpw)
9178 : {
9179 0 : SpinLockAcquire(&XLogCtl->info_lck);
9180 0 : if (XLogCtl->lastFpwDisableRecPtr < record->ReadRecPtr)
9181 0 : XLogCtl->lastFpwDisableRecPtr = record->ReadRecPtr;
9182 0 : SpinLockRelease(&XLogCtl->info_lck);
9183 : }
9184 :
9185 : /* Keep track of full_page_writes */
9186 0 : lastFullPageWrites = fpw;
9187 : }
9188 720 : else if (info == XLOG_CHECKPOINT_REDO)
9189 : {
9190 : xl_checkpoint_redo redo_rec;
9191 702 : bool new_state = false;
9192 :
9193 702 : memcpy(&redo_rec, XLogRecGetData(record), sizeof(xl_checkpoint_redo));
9194 :
9195 702 : SpinLockAcquire(&XLogCtl->info_lck);
9196 702 : XLogCtl->data_checksum_version = redo_rec.data_checksum_version;
9197 702 : if (redo_rec.data_checksum_version != ControlFile->data_checksum_version)
9198 3 : new_state = true;
9199 702 : SpinLockRelease(&XLogCtl->info_lck);
9200 :
9201 702 : if (new_state)
9202 3 : EmitAndWaitDataChecksumsBarrier(redo_rec.data_checksum_version);
9203 : }
9204 18 : else if (info == XLOG_LOGICAL_DECODING_STATUS_CHANGE)
9205 : {
9206 : bool status;
9207 :
9208 18 : memcpy(&status, XLogRecGetData(record), sizeof(bool));
9209 :
9210 : /*
9211 : * We need to toggle the logical decoding status and update the
9212 : * XLogLogicalInfo cache of processes synchronously because
9213 : * XLogLogicalInfoActive() is used even during read-only queries
9214 : * (e.g., via RelationIsAccessibleInLogicalDecoding()). In the
9215 : * 'disable' case, it is safe to invalidate existing slots after
9216 : * disabling logical decoding because logical decoding cannot process
9217 : * subsequent WAL records, which may not contain logical information.
9218 : */
9219 18 : if (status)
9220 9 : EnableLogicalDecoding();
9221 : else
9222 9 : DisableLogicalDecoding();
9223 :
9224 18 : elog(DEBUG1, "update logical decoding status to %d during recovery",
9225 : status);
9226 :
9227 18 : if (InRecovery && InHotStandby)
9228 : {
9229 16 : if (!status)
9230 : {
9231 : /*
9232 : * Invalidate logical slots if we are in hot standby and the
9233 : * primary disabled logical decoding.
9234 : */
9235 9 : InvalidateObsoleteReplicationSlots(RS_INVAL_WAL_LEVEL,
9236 : 0, InvalidOid,
9237 : InvalidTransactionId);
9238 : }
9239 7 : else if (sync_replication_slots)
9240 : {
9241 : /*
9242 : * Signal the postmaster to launch the slotsync worker.
9243 : *
9244 : * XXX: For simplicity, we keep the slotsync worker running
9245 : * even after logical decoding is disabled. A future
9246 : * improvement can consider starting and stopping the worker
9247 : * based on logical decoding status change.
9248 : */
9249 0 : kill(PostmasterPid, SIGUSR1);
9250 : }
9251 : }
9252 : }
9253 114396 : }
9254 :
9255 : void
9256 4 : xlog2_redo(XLogReaderState *record)
9257 : {
9258 4 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
9259 :
9260 4 : if (info == XLOG2_CHECKSUMS)
9261 : {
9262 : xl_checksum_state state;
9263 :
9264 4 : memcpy(&state, XLogRecGetData(record), sizeof(xl_checksum_state));
9265 :
9266 4 : SpinLockAcquire(&XLogCtl->info_lck);
9267 4 : XLogCtl->data_checksum_version = state.new_checksum_state;
9268 4 : SpinLockRelease(&XLogCtl->info_lck);
9269 :
9270 : /*
9271 : * Block on a procsignalbarrier to await all processes having seen the
9272 : * change to checksum status. Once the barrier has been passed we can
9273 : * initiate the corresponding processing.
9274 : */
9275 4 : EmitAndWaitDataChecksumsBarrier(state.new_checksum_state);
9276 : }
9277 4 : }
9278 :
9279 : /*
9280 : * Return the extra open flags used for opening a file, depending on the
9281 : * value of the GUCs wal_sync_method, fsync and debug_io_direct.
9282 : */
9283 : static int
9284 17816 : get_sync_bit(int method)
9285 : {
9286 17816 : int o_direct_flag = 0;
9287 :
9288 : /*
9289 : * Use O_DIRECT if requested, except in walreceiver process. The WAL
9290 : * written by walreceiver is normally read by the startup process soon
9291 : * after it's written. Also, walreceiver performs unaligned writes, which
9292 : * don't work with O_DIRECT, so it is required for correctness too.
9293 : */
9294 17816 : if ((io_direct_flags & IO_DIRECT_WAL) && !AmWalReceiverProcess())
9295 9 : o_direct_flag = PG_O_DIRECT;
9296 :
9297 : /* If fsync is disabled, never open in sync mode */
9298 17816 : if (!enableFsync)
9299 17816 : return o_direct_flag;
9300 :
9301 0 : switch (method)
9302 : {
9303 : /*
9304 : * enum values for all sync options are defined even if they are
9305 : * not supported on the current platform. But if not, they are
9306 : * not included in the enum option array, and therefore will never
9307 : * be seen here.
9308 : */
9309 0 : case WAL_SYNC_METHOD_FSYNC:
9310 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
9311 : case WAL_SYNC_METHOD_FDATASYNC:
9312 0 : return o_direct_flag;
9313 : #ifdef O_SYNC
9314 0 : case WAL_SYNC_METHOD_OPEN:
9315 0 : return O_SYNC | o_direct_flag;
9316 : #endif
9317 : #ifdef O_DSYNC
9318 0 : case WAL_SYNC_METHOD_OPEN_DSYNC:
9319 0 : return O_DSYNC | o_direct_flag;
9320 : #endif
9321 0 : default:
9322 : /* can't happen (unless we are out of sync with option array) */
9323 0 : elog(ERROR, "unrecognized \"wal_sync_method\": %d", method);
9324 : return 0; /* silence warning */
9325 : }
9326 : }
9327 :
9328 : /*
9329 : * GUC support
9330 : */
9331 : void
9332 1275 : assign_wal_sync_method(int new_wal_sync_method, void *extra)
9333 : {
9334 1275 : if (wal_sync_method != new_wal_sync_method)
9335 : {
9336 : /*
9337 : * To ensure that no blocks escape unsynced, force an fsync on the
9338 : * currently open log segment (if any). Also, if the open flag is
9339 : * changing, close the log file so it will be reopened (with new flag
9340 : * bit) at next use.
9341 : */
9342 0 : if (openLogFile >= 0)
9343 : {
9344 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN);
9345 0 : if (pg_fsync(openLogFile) != 0)
9346 : {
9347 : char xlogfname[MAXFNAMELEN];
9348 : int save_errno;
9349 :
9350 0 : save_errno = errno;
9351 0 : XLogFileName(xlogfname, openLogTLI, openLogSegNo,
9352 : wal_segment_size);
9353 0 : errno = save_errno;
9354 0 : ereport(PANIC,
9355 : (errcode_for_file_access(),
9356 : errmsg("could not fsync file \"%s\": %m", xlogfname)));
9357 : }
9358 :
9359 0 : pgstat_report_wait_end();
9360 0 : if (get_sync_bit(wal_sync_method) != get_sync_bit(new_wal_sync_method))
9361 0 : XLogFileClose();
9362 : }
9363 : }
9364 1275 : }
9365 :
9366 :
9367 : /*
9368 : * Issue appropriate kind of fsync (if any) for an XLOG output file.
9369 : *
9370 : * 'fd' is a file descriptor for the XLOG file to be fsync'd.
9371 : * 'segno' is for error reporting purposes.
9372 : */
9373 : void
9374 208090 : issue_xlog_fsync(int fd, XLogSegNo segno, TimeLineID tli)
9375 : {
9376 208090 : char *msg = NULL;
9377 : instr_time start;
9378 :
9379 : Assert(tli != 0);
9380 :
9381 : /*
9382 : * Quick exit if fsync is disabled or write() has already synced the WAL
9383 : * file.
9384 : */
9385 208090 : if (!enableFsync ||
9386 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN ||
9387 0 : wal_sync_method == WAL_SYNC_METHOD_OPEN_DSYNC)
9388 208090 : return;
9389 :
9390 : /*
9391 : * Measure I/O timing to sync the WAL file for pg_stat_io.
9392 : */
9393 0 : start = pgstat_prepare_io_time(track_wal_io_timing);
9394 :
9395 0 : pgstat_report_wait_start(WAIT_EVENT_WAL_SYNC);
9396 0 : switch (wal_sync_method)
9397 : {
9398 0 : case WAL_SYNC_METHOD_FSYNC:
9399 0 : if (pg_fsync_no_writethrough(fd) != 0)
9400 0 : msg = _("could not fsync file \"%s\": %m");
9401 0 : break;
9402 : #ifdef HAVE_FSYNC_WRITETHROUGH
9403 : case WAL_SYNC_METHOD_FSYNC_WRITETHROUGH:
9404 : if (pg_fsync_writethrough(fd) != 0)
9405 : msg = _("could not fsync write-through file \"%s\": %m");
9406 : break;
9407 : #endif
9408 0 : case WAL_SYNC_METHOD_FDATASYNC:
9409 0 : if (pg_fdatasync(fd) != 0)
9410 0 : msg = _("could not fdatasync file \"%s\": %m");
9411 0 : break;
9412 0 : case WAL_SYNC_METHOD_OPEN:
9413 : case WAL_SYNC_METHOD_OPEN_DSYNC:
9414 : /* not reachable */
9415 : Assert(false);
9416 0 : break;
9417 0 : default:
9418 0 : ereport(PANIC,
9419 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
9420 : errmsg_internal("unrecognized \"wal_sync_method\": %d", wal_sync_method));
9421 : break;
9422 : }
9423 :
9424 : /* PANIC if failed to fsync */
9425 0 : if (msg)
9426 : {
9427 : char xlogfname[MAXFNAMELEN];
9428 0 : int save_errno = errno;
9429 :
9430 0 : XLogFileName(xlogfname, tli, segno, wal_segment_size);
9431 0 : errno = save_errno;
9432 0 : ereport(PANIC,
9433 : (errcode_for_file_access(),
9434 : errmsg(msg, xlogfname)));
9435 : }
9436 :
9437 0 : pgstat_report_wait_end();
9438 :
9439 0 : pgstat_count_io_op_time(IOOBJECT_WAL, IOCONTEXT_NORMAL, IOOP_FSYNC,
9440 : start, 1, 0);
9441 : }
9442 :
9443 : /*
9444 : * do_pg_backup_start is the workhorse of the user-visible pg_backup_start()
9445 : * function. It creates the necessary starting checkpoint and constructs the
9446 : * backup state and tablespace map.
9447 : *
9448 : * Input parameters are "state" (the backup state), "fast" (if true, we do
9449 : * the checkpoint in fast mode), and "tablespaces" (if non-NULL, indicates a
9450 : * list of tablespaceinfo structs describing the cluster's tablespaces.).
9451 : *
9452 : * The tablespace map contents are appended to passed-in parameter
9453 : * tablespace_map and the caller is responsible for including it in the backup
9454 : * archive as 'tablespace_map'. The tablespace_map file is required mainly for
9455 : * tar format in windows as native windows utilities are not able to create
9456 : * symlinks while extracting files from tar. However for consistency and
9457 : * platform-independence, we do it the same way everywhere.
9458 : *
9459 : * It fills in "state" with the information required for the backup, such
9460 : * as the minimum WAL location that must be present to restore from this
9461 : * backup (starttli) and the corresponding timeline ID (starttli).
9462 : *
9463 : * Every successfully started backup must be stopped by calling
9464 : * do_pg_backup_stop() or do_pg_abort_backup(). There can be many
9465 : * backups active at the same time.
9466 : *
9467 : * It is the responsibility of the caller of this function to verify the
9468 : * permissions of the calling user!
9469 : */
9470 : void
9471 177 : do_pg_backup_start(const char *backupidstr, bool fast, List **tablespaces,
9472 : BackupState *state, StringInfo tblspcmapfile)
9473 : {
9474 : bool backup_started_in_recovery;
9475 :
9476 : Assert(state != NULL);
9477 177 : backup_started_in_recovery = RecoveryInProgress();
9478 :
9479 : /*
9480 : * During recovery, we don't need to check WAL level. Because, if WAL
9481 : * level is not sufficient, it's impossible to get here during recovery.
9482 : */
9483 177 : if (!backup_started_in_recovery && !XLogIsNeeded())
9484 0 : ereport(ERROR,
9485 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9486 : errmsg("WAL level not sufficient for making an online backup"),
9487 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9488 :
9489 177 : if (strlen(backupidstr) > MAXPGPATH)
9490 1 : ereport(ERROR,
9491 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
9492 : errmsg("backup label too long (max %d bytes)",
9493 : MAXPGPATH)));
9494 :
9495 176 : strlcpy(state->name, backupidstr, sizeof(state->name));
9496 :
9497 : /*
9498 : * Mark backup active in shared memory. We must do full-page WAL writes
9499 : * during an on-line backup even if not doing so at other times, because
9500 : * it's quite possible for the backup dump to obtain a "torn" (partially
9501 : * written) copy of a database page if it reads the page concurrently with
9502 : * our write to the same page. This can be fixed as long as the first
9503 : * write to the page in the WAL sequence is a full-page write. Hence, we
9504 : * increment runningBackups then force a CHECKPOINT, to ensure there are
9505 : * no dirty pages in shared memory that might get dumped while the backup
9506 : * is in progress without having a corresponding WAL record. (Once the
9507 : * backup is complete, we need not force full-page writes anymore, since
9508 : * we expect that any pages not modified during the backup interval must
9509 : * have been correctly captured by the backup.)
9510 : *
9511 : * Note that forcing full-page writes has no effect during an online
9512 : * backup from the standby.
9513 : *
9514 : * We must hold all the insertion locks to change the value of
9515 : * runningBackups, to ensure adequate interlocking against
9516 : * XLogInsertRecord().
9517 : */
9518 176 : WALInsertLockAcquireExclusive();
9519 176 : XLogCtl->Insert.runningBackups++;
9520 176 : WALInsertLockRelease();
9521 :
9522 : /*
9523 : * Ensure we decrement runningBackups if we fail below. NB -- for this to
9524 : * work correctly, it is critical that sessionBackupState is only updated
9525 : * after this block is over.
9526 : */
9527 176 : PG_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9528 : {
9529 176 : bool gotUniqueStartpoint = false;
9530 : DIR *tblspcdir;
9531 : struct dirent *de;
9532 : tablespaceinfo *ti;
9533 : int datadirpathlen;
9534 :
9535 : /*
9536 : * Force an XLOG file switch before the checkpoint, to ensure that the
9537 : * WAL segment the checkpoint is written to doesn't contain pages with
9538 : * old timeline IDs. That would otherwise happen if you called
9539 : * pg_backup_start() right after restoring from a PITR archive: the
9540 : * first WAL segment containing the startup checkpoint has pages in
9541 : * the beginning with the old timeline ID. That can cause trouble at
9542 : * recovery: we won't have a history file covering the old timeline if
9543 : * pg_wal directory was not included in the base backup and the WAL
9544 : * archive was cleared too before starting the backup.
9545 : *
9546 : * During recovery, we skip forcing XLOG file switch, which means that
9547 : * the backup taken during recovery is not available for the special
9548 : * recovery case described above.
9549 : */
9550 176 : if (!backup_started_in_recovery)
9551 168 : RequestXLogSwitch(false);
9552 :
9553 : do
9554 : {
9555 : bool checkpointfpw;
9556 :
9557 : /*
9558 : * Force a CHECKPOINT. Aside from being necessary to prevent torn
9559 : * page problems, this guarantees that two successive backup runs
9560 : * will have different checkpoint positions and hence different
9561 : * history file names, even if nothing happened in between.
9562 : *
9563 : * During recovery, establish a restartpoint if possible. We use
9564 : * the last restartpoint as the backup starting checkpoint. This
9565 : * means that two successive backup runs can have same checkpoint
9566 : * positions.
9567 : *
9568 : * Since the fact that we are executing do_pg_backup_start()
9569 : * during recovery means that checkpointer is running, we can use
9570 : * RequestCheckpoint() to establish a restartpoint.
9571 : *
9572 : * We use CHECKPOINT_FAST only if requested by user (via passing
9573 : * fast = true). Otherwise this can take awhile.
9574 : */
9575 176 : RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT |
9576 : (fast ? CHECKPOINT_FAST : 0));
9577 :
9578 : /*
9579 : * Now we need to fetch the checkpoint record location, and also
9580 : * its REDO pointer. The oldest point in WAL that would be needed
9581 : * to restore starting from the checkpoint is precisely the REDO
9582 : * pointer.
9583 : */
9584 176 : LWLockAcquire(ControlFileLock, LW_SHARED);
9585 176 : state->checkpointloc = ControlFile->checkPoint;
9586 176 : state->startpoint = ControlFile->checkPointCopy.redo;
9587 176 : state->starttli = ControlFile->checkPointCopy.ThisTimeLineID;
9588 176 : checkpointfpw = ControlFile->checkPointCopy.fullPageWrites;
9589 176 : LWLockRelease(ControlFileLock);
9590 :
9591 176 : if (backup_started_in_recovery)
9592 : {
9593 : XLogRecPtr recptr;
9594 :
9595 : /*
9596 : * Check to see if all WAL replayed during online backup
9597 : * (i.e., since last restartpoint used as backup starting
9598 : * checkpoint) contain full-page writes.
9599 : */
9600 8 : SpinLockAcquire(&XLogCtl->info_lck);
9601 8 : recptr = XLogCtl->lastFpwDisableRecPtr;
9602 8 : SpinLockRelease(&XLogCtl->info_lck);
9603 :
9604 8 : if (!checkpointfpw || state->startpoint <= recptr)
9605 0 : ereport(ERROR,
9606 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9607 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9608 : "since last restartpoint"),
9609 : errhint("This means that the backup being taken on the standby "
9610 : "is corrupt and should not be used. "
9611 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9612 : "and then try an online backup again.")));
9613 :
9614 : /*
9615 : * During recovery, since we don't use the end-of-backup WAL
9616 : * record and don't write the backup history file, the
9617 : * starting WAL location doesn't need to be unique. This means
9618 : * that two base backups started at the same time might use
9619 : * the same checkpoint as starting locations.
9620 : */
9621 8 : gotUniqueStartpoint = true;
9622 : }
9623 :
9624 : /*
9625 : * If two base backups are started at the same time (in WAL sender
9626 : * processes), we need to make sure that they use different
9627 : * checkpoints as starting locations, because we use the starting
9628 : * WAL location as a unique identifier for the base backup in the
9629 : * end-of-backup WAL record and when we write the backup history
9630 : * file. Perhaps it would be better generate a separate unique ID
9631 : * for each backup instead of forcing another checkpoint, but
9632 : * taking a checkpoint right after another is not that expensive
9633 : * either because only few buffers have been dirtied yet.
9634 : */
9635 176 : WALInsertLockAcquireExclusive();
9636 176 : if (XLogCtl->Insert.lastBackupStart < state->startpoint)
9637 : {
9638 176 : XLogCtl->Insert.lastBackupStart = state->startpoint;
9639 176 : gotUniqueStartpoint = true;
9640 : }
9641 176 : WALInsertLockRelease();
9642 176 : } while (!gotUniqueStartpoint);
9643 :
9644 : /*
9645 : * Construct tablespace_map file.
9646 : */
9647 176 : datadirpathlen = strlen(DataDir);
9648 :
9649 : /* Collect information about all tablespaces */
9650 176 : tblspcdir = AllocateDir(PG_TBLSPC_DIR);
9651 565 : while ((de = ReadDir(tblspcdir, PG_TBLSPC_DIR)) != NULL)
9652 : {
9653 : char fullpath[MAXPGPATH + sizeof(PG_TBLSPC_DIR)];
9654 : char linkpath[MAXPGPATH];
9655 389 : char *relpath = NULL;
9656 : char *s;
9657 : PGFileType de_type;
9658 : char *badp;
9659 : Oid tsoid;
9660 :
9661 : /*
9662 : * Try to parse the directory name as an unsigned integer.
9663 : *
9664 : * Tablespace directories should be positive integers that can be
9665 : * represented in 32 bits, with no leading zeroes or trailing
9666 : * garbage. If we come across a name that doesn't meet those
9667 : * criteria, skip it.
9668 : */
9669 389 : if (de->d_name[0] < '1' || de->d_name[1] > '9')
9670 352 : continue;
9671 37 : errno = 0;
9672 37 : tsoid = strtoul(de->d_name, &badp, 10);
9673 37 : if (*badp != '\0' || errno == EINVAL || errno == ERANGE)
9674 0 : continue;
9675 :
9676 37 : snprintf(fullpath, sizeof(fullpath), "%s/%s", PG_TBLSPC_DIR, de->d_name);
9677 :
9678 37 : de_type = get_dirent_type(fullpath, de, false, ERROR);
9679 :
9680 37 : if (de_type == PGFILETYPE_LNK)
9681 : {
9682 : StringInfoData escapedpath;
9683 : int rllen;
9684 :
9685 23 : rllen = readlink(fullpath, linkpath, sizeof(linkpath));
9686 23 : if (rllen < 0)
9687 : {
9688 0 : ereport(WARNING,
9689 : (errmsg("could not read symbolic link \"%s\": %m",
9690 : fullpath)));
9691 0 : continue;
9692 : }
9693 23 : else if (rllen >= sizeof(linkpath))
9694 : {
9695 0 : ereport(WARNING,
9696 : (errmsg("symbolic link \"%s\" target is too long",
9697 : fullpath)));
9698 0 : continue;
9699 : }
9700 23 : linkpath[rllen] = '\0';
9701 :
9702 : /*
9703 : * Relpath holds the relative path of the tablespace directory
9704 : * when it's located within PGDATA, or NULL if it's located
9705 : * elsewhere.
9706 : */
9707 23 : if (rllen > datadirpathlen &&
9708 1 : strncmp(linkpath, DataDir, datadirpathlen) == 0 &&
9709 0 : IS_DIR_SEP(linkpath[datadirpathlen]))
9710 0 : relpath = pstrdup(linkpath + datadirpathlen + 1);
9711 :
9712 : /*
9713 : * Add a backslash-escaped version of the link path to the
9714 : * tablespace map file.
9715 : */
9716 23 : initStringInfo(&escapedpath);
9717 562 : for (s = linkpath; *s; s++)
9718 : {
9719 539 : if (*s == '\n' || *s == '\r' || *s == '\\')
9720 0 : appendStringInfoChar(&escapedpath, '\\');
9721 539 : appendStringInfoChar(&escapedpath, *s);
9722 : }
9723 23 : appendStringInfo(tblspcmapfile, "%s %s\n",
9724 23 : de->d_name, escapedpath.data);
9725 23 : pfree(escapedpath.data);
9726 : }
9727 14 : else if (de_type == PGFILETYPE_DIR)
9728 : {
9729 : /*
9730 : * It's possible to use allow_in_place_tablespaces to create
9731 : * directories directly under pg_tblspc, for testing purposes
9732 : * only.
9733 : *
9734 : * In this case, we store a relative path rather than an
9735 : * absolute path into the tablespaceinfo.
9736 : */
9737 14 : snprintf(linkpath, sizeof(linkpath), "%s/%s",
9738 14 : PG_TBLSPC_DIR, de->d_name);
9739 14 : relpath = pstrdup(linkpath);
9740 : }
9741 : else
9742 : {
9743 : /* Skip any other file type that appears here. */
9744 0 : continue;
9745 : }
9746 :
9747 37 : ti = palloc_object(tablespaceinfo);
9748 37 : ti->oid = tsoid;
9749 37 : ti->path = pstrdup(linkpath);
9750 37 : ti->rpath = relpath;
9751 37 : ti->size = -1;
9752 :
9753 37 : if (tablespaces)
9754 37 : *tablespaces = lappend(*tablespaces, ti);
9755 : }
9756 176 : FreeDir(tblspcdir);
9757 :
9758 176 : state->starttime = (pg_time_t) time(NULL);
9759 : }
9760 176 : PG_END_ENSURE_ERROR_CLEANUP(do_pg_abort_backup, BoolGetDatum(true));
9761 :
9762 176 : state->started_in_recovery = backup_started_in_recovery;
9763 :
9764 : /*
9765 : * Mark that the start phase has correctly finished for the backup.
9766 : */
9767 176 : sessionBackupState = SESSION_BACKUP_RUNNING;
9768 176 : }
9769 :
9770 : /*
9771 : * Utility routine to fetch the session-level status of a backup running.
9772 : */
9773 : SessionBackupState
9774 198 : get_backup_status(void)
9775 : {
9776 198 : return sessionBackupState;
9777 : }
9778 :
9779 : /*
9780 : * do_pg_backup_stop
9781 : *
9782 : * Utility function called at the end of an online backup. It creates history
9783 : * file (if required), resets sessionBackupState and so on. It can optionally
9784 : * wait for WAL segments to be archived.
9785 : *
9786 : * "state" is filled with the information necessary to restore from this
9787 : * backup with its stop LSN (stoppoint), its timeline ID (stoptli), etc.
9788 : *
9789 : * It is the responsibility of the caller of this function to verify the
9790 : * permissions of the calling user!
9791 : */
9792 : void
9793 170 : do_pg_backup_stop(BackupState *state, bool waitforarchive)
9794 : {
9795 170 : bool backup_stopped_in_recovery = false;
9796 : char histfilepath[MAXPGPATH];
9797 : char lastxlogfilename[MAXFNAMELEN];
9798 : char histfilename[MAXFNAMELEN];
9799 : XLogSegNo _logSegNo;
9800 : FILE *fp;
9801 : int seconds_before_warning;
9802 170 : int waits = 0;
9803 170 : bool reported_waiting = false;
9804 :
9805 : Assert(state != NULL);
9806 :
9807 170 : backup_stopped_in_recovery = RecoveryInProgress();
9808 :
9809 : /*
9810 : * During recovery, we don't need to check WAL level. Because, if WAL
9811 : * level is not sufficient, it's impossible to get here during recovery.
9812 : */
9813 170 : if (!backup_stopped_in_recovery && !XLogIsNeeded())
9814 0 : ereport(ERROR,
9815 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9816 : errmsg("WAL level not sufficient for making an online backup"),
9817 : errhint("\"wal_level\" must be set to \"replica\" or \"logical\" at server start.")));
9818 :
9819 : /*
9820 : * OK to update backup counter and session-level lock.
9821 : *
9822 : * Note that CHECK_FOR_INTERRUPTS() must not occur while updating them,
9823 : * otherwise they can be updated inconsistently, which might cause
9824 : * do_pg_abort_backup() to fail.
9825 : */
9826 170 : WALInsertLockAcquireExclusive();
9827 :
9828 : /*
9829 : * It is expected that each do_pg_backup_start() call is matched by
9830 : * exactly one do_pg_backup_stop() call.
9831 : */
9832 : Assert(XLogCtl->Insert.runningBackups > 0);
9833 170 : XLogCtl->Insert.runningBackups--;
9834 :
9835 : /*
9836 : * Clean up session-level lock.
9837 : *
9838 : * You might think that WALInsertLockRelease() can be called before
9839 : * cleaning up session-level lock because session-level lock doesn't need
9840 : * to be protected with WAL insertion lock. But since
9841 : * CHECK_FOR_INTERRUPTS() can occur in it, session-level lock must be
9842 : * cleaned up before it.
9843 : */
9844 170 : sessionBackupState = SESSION_BACKUP_NONE;
9845 :
9846 170 : WALInsertLockRelease();
9847 :
9848 : /*
9849 : * If we are taking an online backup from the standby, we confirm that the
9850 : * standby has not been promoted during the backup.
9851 : */
9852 170 : if (state->started_in_recovery && !backup_stopped_in_recovery)
9853 0 : ereport(ERROR,
9854 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9855 : errmsg("the standby was promoted during online backup"),
9856 : errhint("This means that the backup being taken is corrupt "
9857 : "and should not be used. "
9858 : "Try taking another online backup.")));
9859 :
9860 : /*
9861 : * During recovery, we don't write an end-of-backup record. We assume that
9862 : * pg_control was backed up last and its minimum recovery point can be
9863 : * available as the backup end location. Since we don't have an
9864 : * end-of-backup record, we use the pg_control value to check whether
9865 : * we've reached the end of backup when starting recovery from this
9866 : * backup. We have no way of checking if pg_control wasn't backed up last
9867 : * however.
9868 : *
9869 : * We don't force a switch to new WAL file but it is still possible to
9870 : * wait for all the required files to be archived if waitforarchive is
9871 : * true. This is okay if we use the backup to start a standby and fetch
9872 : * the missing WAL using streaming replication. But in the case of an
9873 : * archive recovery, a user should set waitforarchive to true and wait for
9874 : * them to be archived to ensure that all the required files are
9875 : * available.
9876 : *
9877 : * We return the current minimum recovery point as the backup end
9878 : * location. Note that it can be greater than the exact backup end
9879 : * location if the minimum recovery point is updated after the backup of
9880 : * pg_control. This is harmless for current uses.
9881 : *
9882 : * XXX currently a backup history file is for informational and debug
9883 : * purposes only. It's not essential for an online backup. Furthermore,
9884 : * even if it's created, it will not be archived during recovery because
9885 : * an archiver is not invoked. So it doesn't seem worthwhile to write a
9886 : * backup history file during recovery.
9887 : */
9888 170 : if (backup_stopped_in_recovery)
9889 : {
9890 : XLogRecPtr recptr;
9891 :
9892 : /*
9893 : * Check to see if all WAL replayed during online backup contain
9894 : * full-page writes.
9895 : */
9896 8 : SpinLockAcquire(&XLogCtl->info_lck);
9897 8 : recptr = XLogCtl->lastFpwDisableRecPtr;
9898 8 : SpinLockRelease(&XLogCtl->info_lck);
9899 :
9900 8 : if (state->startpoint <= recptr)
9901 0 : ereport(ERROR,
9902 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
9903 : errmsg("WAL generated with \"full_page_writes=off\" was replayed "
9904 : "during online backup"),
9905 : errhint("This means that the backup being taken on the standby "
9906 : "is corrupt and should not be used. "
9907 : "Enable \"full_page_writes\" and run CHECKPOINT on the primary, "
9908 : "and then try an online backup again.")));
9909 :
9910 :
9911 8 : LWLockAcquire(ControlFileLock, LW_SHARED);
9912 8 : state->stoppoint = ControlFile->minRecoveryPoint;
9913 8 : state->stoptli = ControlFile->minRecoveryPointTLI;
9914 8 : LWLockRelease(ControlFileLock);
9915 : }
9916 : else
9917 : {
9918 : char *history_file;
9919 :
9920 : /*
9921 : * Write the backup-end xlog record
9922 : */
9923 162 : XLogBeginInsert();
9924 162 : XLogRegisterData(&state->startpoint,
9925 : sizeof(state->startpoint));
9926 162 : state->stoppoint = XLogInsert(RM_XLOG_ID, XLOG_BACKUP_END);
9927 :
9928 : /*
9929 : * Given that we're not in recovery, InsertTimeLineID is set and can't
9930 : * change, so we can read it without a lock.
9931 : */
9932 162 : state->stoptli = XLogCtl->InsertTimeLineID;
9933 :
9934 : /*
9935 : * Force a switch to a new xlog segment file, so that the backup is
9936 : * valid as soon as archiver moves out the current segment file.
9937 : */
9938 162 : RequestXLogSwitch(false);
9939 :
9940 162 : state->stoptime = (pg_time_t) time(NULL);
9941 :
9942 : /*
9943 : * Write the backup history file
9944 : */
9945 162 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
9946 162 : BackupHistoryFilePath(histfilepath, state->stoptli, _logSegNo,
9947 : state->startpoint, wal_segment_size);
9948 162 : fp = AllocateFile(histfilepath, "w");
9949 162 : if (!fp)
9950 0 : ereport(ERROR,
9951 : (errcode_for_file_access(),
9952 : errmsg("could not create file \"%s\": %m",
9953 : histfilepath)));
9954 :
9955 : /* Build and save the contents of the backup history file */
9956 162 : history_file = build_backup_content(state, true);
9957 162 : fprintf(fp, "%s", history_file);
9958 162 : pfree(history_file);
9959 :
9960 162 : if (fflush(fp) || ferror(fp) || FreeFile(fp))
9961 0 : ereport(ERROR,
9962 : (errcode_for_file_access(),
9963 : errmsg("could not write file \"%s\": %m",
9964 : histfilepath)));
9965 :
9966 : /*
9967 : * Clean out any no-longer-needed history files. As a side effect,
9968 : * this will post a .ready file for the newly created history file,
9969 : * notifying the archiver that history file may be archived
9970 : * immediately.
9971 : */
9972 162 : CleanupBackupHistory();
9973 : }
9974 :
9975 : /*
9976 : * If archiving is enabled, wait for all the required WAL files to be
9977 : * archived before returning. If archiving isn't enabled, the required WAL
9978 : * needs to be transported via streaming replication (hopefully with
9979 : * wal_keep_size set high enough), or some more exotic mechanism like
9980 : * polling and copying files from pg_wal with script. We have no knowledge
9981 : * of those mechanisms, so it's up to the user to ensure that he gets all
9982 : * the required WAL.
9983 : *
9984 : * We wait until both the last WAL file filled during backup and the
9985 : * history file have been archived, and assume that the alphabetic sorting
9986 : * property of the WAL files ensures any earlier WAL files are safely
9987 : * archived as well.
9988 : *
9989 : * We wait forever, since archive_command is supposed to work and we
9990 : * assume the admin wanted his backup to work completely. If you don't
9991 : * wish to wait, then either waitforarchive should be passed in as false,
9992 : * or you can set statement_timeout. Also, some notices are issued to
9993 : * clue in anyone who might be doing this interactively.
9994 : */
9995 :
9996 170 : if (waitforarchive &&
9997 11 : ((!backup_stopped_in_recovery && XLogArchivingActive()) ||
9998 1 : (backup_stopped_in_recovery && XLogArchivingAlways())))
9999 : {
10000 5 : XLByteToPrevSeg(state->stoppoint, _logSegNo, wal_segment_size);
10001 5 : XLogFileName(lastxlogfilename, state->stoptli, _logSegNo,
10002 : wal_segment_size);
10003 :
10004 5 : XLByteToSeg(state->startpoint, _logSegNo, wal_segment_size);
10005 5 : BackupHistoryFileName(histfilename, state->stoptli, _logSegNo,
10006 : state->startpoint, wal_segment_size);
10007 :
10008 5 : seconds_before_warning = 60;
10009 5 : waits = 0;
10010 :
10011 15 : while (XLogArchiveIsBusy(lastxlogfilename) ||
10012 5 : XLogArchiveIsBusy(histfilename))
10013 : {
10014 5 : CHECK_FOR_INTERRUPTS();
10015 :
10016 5 : if (!reported_waiting && waits > 5)
10017 : {
10018 0 : ereport(NOTICE,
10019 : (errmsg("base backup done, waiting for required WAL segments to be archived")));
10020 0 : reported_waiting = true;
10021 : }
10022 :
10023 5 : (void) WaitLatch(MyLatch,
10024 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
10025 : 1000L,
10026 : WAIT_EVENT_BACKUP_WAIT_WAL_ARCHIVE);
10027 5 : ResetLatch(MyLatch);
10028 :
10029 5 : if (++waits >= seconds_before_warning)
10030 : {
10031 0 : seconds_before_warning *= 2; /* This wraps in >10 years... */
10032 0 : ereport(WARNING,
10033 : (errmsg("still waiting for all required WAL segments to be archived (%d seconds elapsed)",
10034 : waits),
10035 : errhint("Check that your \"archive_command\" is executing properly. "
10036 : "You can safely cancel this backup, "
10037 : "but the database backup will not be usable without all the WAL segments.")));
10038 : }
10039 : }
10040 :
10041 5 : ereport(NOTICE,
10042 : (errmsg("all required WAL segments have been archived")));
10043 : }
10044 165 : else if (waitforarchive)
10045 6 : ereport(NOTICE,
10046 : (errmsg("WAL archiving is not enabled; you must ensure that all required WAL segments are copied through other means to complete the backup")));
10047 170 : }
10048 :
10049 :
10050 : /*
10051 : * do_pg_abort_backup: abort a running backup
10052 : *
10053 : * This does just the most basic steps of do_pg_backup_stop(), by taking the
10054 : * system out of backup mode, thus making it a lot more safe to call from
10055 : * an error handler.
10056 : *
10057 : * 'arg' indicates that it's being called during backup setup; so
10058 : * sessionBackupState has not been modified yet, but runningBackups has
10059 : * already been incremented. When it's false, then it's invoked as a
10060 : * before_shmem_exit handler, and therefore we must not change state
10061 : * unless sessionBackupState indicates that a backup is actually running.
10062 : *
10063 : * NB: This gets used as a PG_ENSURE_ERROR_CLEANUP callback and
10064 : * before_shmem_exit handler, hence the odd-looking signature.
10065 : */
10066 : void
10067 9 : do_pg_abort_backup(int code, Datum arg)
10068 : {
10069 9 : bool during_backup_start = DatumGetBool(arg);
10070 :
10071 : /* If called during backup start, there shouldn't be one already running */
10072 : Assert(!during_backup_start || sessionBackupState == SESSION_BACKUP_NONE);
10073 :
10074 9 : if (during_backup_start || sessionBackupState != SESSION_BACKUP_NONE)
10075 : {
10076 6 : WALInsertLockAcquireExclusive();
10077 : Assert(XLogCtl->Insert.runningBackups > 0);
10078 6 : XLogCtl->Insert.runningBackups--;
10079 :
10080 6 : sessionBackupState = SESSION_BACKUP_NONE;
10081 6 : WALInsertLockRelease();
10082 :
10083 6 : if (!during_backup_start)
10084 6 : ereport(WARNING,
10085 : errmsg("aborting backup due to backend exiting before pg_backup_stop was called"));
10086 : }
10087 9 : }
10088 :
10089 : /*
10090 : * Register a handler that will warn about unterminated backups at end of
10091 : * session, unless this has already been done.
10092 : */
10093 : void
10094 5 : register_persistent_abort_backup_handler(void)
10095 : {
10096 : static bool already_done = false;
10097 :
10098 5 : if (already_done)
10099 1 : return;
10100 4 : before_shmem_exit(do_pg_abort_backup, BoolGetDatum(false));
10101 4 : already_done = true;
10102 : }
10103 :
10104 : /*
10105 : * Get latest WAL insert pointer
10106 : */
10107 : XLogRecPtr
10108 2135 : GetXLogInsertRecPtr(void)
10109 : {
10110 2135 : XLogCtlInsert *Insert = &XLogCtl->Insert;
10111 : uint64 current_bytepos;
10112 :
10113 2135 : SpinLockAcquire(&Insert->insertpos_lck);
10114 2135 : current_bytepos = Insert->CurrBytePos;
10115 2135 : SpinLockRelease(&Insert->insertpos_lck);
10116 :
10117 2135 : return XLogBytePosToRecPtr(current_bytepos);
10118 : }
10119 :
10120 : /*
10121 : * Get latest WAL record end pointer
10122 : */
10123 : XLogRecPtr
10124 117174 : GetXLogInsertEndRecPtr(void)
10125 : {
10126 117174 : XLogCtlInsert *Insert = &XLogCtl->Insert;
10127 : uint64 current_bytepos;
10128 :
10129 117174 : SpinLockAcquire(&Insert->insertpos_lck);
10130 117174 : current_bytepos = Insert->CurrBytePos;
10131 117174 : SpinLockRelease(&Insert->insertpos_lck);
10132 :
10133 117174 : return XLogBytePosToEndRecPtr(current_bytepos);
10134 : }
10135 :
10136 : /*
10137 : * Get latest WAL write pointer
10138 : */
10139 : XLogRecPtr
10140 10554 : GetXLogWriteRecPtr(void)
10141 : {
10142 10554 : RefreshXLogWriteResult(LogwrtResult);
10143 :
10144 10554 : return LogwrtResult.Write;
10145 : }
10146 :
10147 : /*
10148 : * Returns the redo pointer of the last checkpoint or restartpoint. This is
10149 : * the oldest point in WAL that we still need, if we have to restart recovery.
10150 : */
10151 : void
10152 391 : GetOldestRestartPoint(XLogRecPtr *oldrecptr, TimeLineID *oldtli)
10153 : {
10154 391 : LWLockAcquire(ControlFileLock, LW_SHARED);
10155 391 : *oldrecptr = ControlFile->checkPointCopy.redo;
10156 391 : *oldtli = ControlFile->checkPointCopy.ThisTimeLineID;
10157 391 : LWLockRelease(ControlFileLock);
10158 391 : }
10159 :
10160 : /* Thin wrapper around ShutdownWalRcv(). */
10161 : void
10162 1068 : XLogShutdownWalRcv(void)
10163 : {
10164 : Assert(AmStartupProcess() || !IsUnderPostmaster);
10165 :
10166 1068 : ShutdownWalRcv();
10167 1068 : ResetInstallXLogFileSegmentActive();
10168 1068 : }
10169 :
10170 : /* Enable WAL file recycling and preallocation. */
10171 : void
10172 1262 : SetInstallXLogFileSegmentActive(void)
10173 : {
10174 1262 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
10175 1262 : XLogCtl->InstallXLogFileSegmentActive = true;
10176 1262 : LWLockRelease(ControlFileLock);
10177 1262 : }
10178 :
10179 : /* Disable WAL file recycling and preallocation. */
10180 : void
10181 1222 : ResetInstallXLogFileSegmentActive(void)
10182 : {
10183 1222 : LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
10184 1222 : XLogCtl->InstallXLogFileSegmentActive = false;
10185 1222 : LWLockRelease(ControlFileLock);
10186 1222 : }
10187 :
10188 : bool
10189 0 : IsInstallXLogFileSegmentActive(void)
10190 : {
10191 : bool result;
10192 :
10193 0 : LWLockAcquire(ControlFileLock, LW_SHARED);
10194 0 : result = XLogCtl->InstallXLogFileSegmentActive;
10195 0 : LWLockRelease(ControlFileLock);
10196 :
10197 0 : return result;
10198 : }
10199 :
10200 : /*
10201 : * Update the WalWriterSleeping flag.
10202 : */
10203 : void
10204 612 : SetWalWriterSleeping(bool sleeping)
10205 : {
10206 612 : SpinLockAcquire(&XLogCtl->info_lck);
10207 612 : XLogCtl->WalWriterSleeping = sleeping;
10208 612 : SpinLockRelease(&XLogCtl->info_lck);
10209 612 : }
|
