Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * xlogutils.c
4 : *
5 : * PostgreSQL write-ahead log manager utility routines
6 : *
7 : * This file contains support routines that are used by XLOG replay functions.
8 : * None of this code is used during normal system operation.
9 : *
10 : *
11 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
12 : * Portions Copyright (c) 1994, Regents of the University of California
13 : *
14 : * src/backend/access/transam/xlogutils.c
15 : *
16 : *-------------------------------------------------------------------------
17 : */
18 : #include "postgres.h"
19 :
20 : #include <unistd.h>
21 :
22 : #include "access/timeline.h"
23 : #include "access/xlogrecovery.h"
24 : #include "access/xlog_internal.h"
25 : #include "access/xlogutils.h"
26 : #include "miscadmin.h"
27 : #include "storage/fd.h"
28 : #include "storage/smgr.h"
29 : #include "utils/hsearch.h"
30 : #include "utils/rel.h"
31 :
32 :
33 : /* GUC variable */
34 : bool ignore_invalid_pages = false;
35 :
36 : /*
37 : * Are we doing recovery from XLOG?
38 : *
39 : * This is only ever true in the startup process; it should be read as meaning
40 : * "this process is replaying WAL records", rather than "the system is in
41 : * recovery mode". It should be examined primarily by functions that need
42 : * to act differently when called from a WAL redo function (e.g., to skip WAL
43 : * logging). To check whether the system is in recovery regardless of which
44 : * process you're running in, use RecoveryInProgress() but only after shared
45 : * memory startup and lock initialization.
46 : *
47 : * This is updated from xlog.c and xlogrecovery.c, but lives here because
48 : * it's mostly read by WAL redo functions.
49 : */
50 : bool InRecovery = false;
51 :
52 : /* Are we in Hot Standby mode? Only valid in startup process, see xlogutils.h */
53 : HotStandbyState standbyState = STANDBY_DISABLED;
54 :
55 : /*
56 : * During XLOG replay, we may see XLOG records for incremental updates of
57 : * pages that no longer exist, because their relation was later dropped or
58 : * truncated. (Note: this is only possible when full_page_writes = OFF,
59 : * since when it's ON, the first reference we see to a page should always
60 : * be a full-page rewrite not an incremental update.) Rather than simply
61 : * ignoring such records, we make a note of the referenced page, and then
62 : * complain if we don't actually see a drop or truncate covering the page
63 : * later in replay.
64 : */
65 : typedef struct xl_invalid_page_key
66 : {
67 : RelFileLocator locator; /* the relation */
68 : ForkNumber forkno; /* the fork number */
69 : BlockNumber blkno; /* the page */
70 : } xl_invalid_page_key;
71 :
72 : typedef struct xl_invalid_page
73 : {
74 : xl_invalid_page_key key; /* hash key ... must be first */
75 : bool present; /* page existed but contained zeroes */
76 : } xl_invalid_page;
77 :
78 : static HTAB *invalid_page_tab = NULL;
79 :
80 : static int read_local_xlog_page_guts(XLogReaderState *state, XLogRecPtr targetPagePtr,
81 : int reqLen, XLogRecPtr targetRecPtr,
82 : char *cur_page, bool wait_for_wal);
83 :
84 : /* Report a reference to an invalid page */
85 : static void
86 0 : report_invalid_page(int elevel, RelFileLocator locator, ForkNumber forkno,
87 : BlockNumber blkno, bool present)
88 : {
89 0 : RelPathStr path = relpathperm(locator, forkno);
90 :
91 0 : if (present)
92 0 : elog(elevel, "page %u of relation %s is uninitialized",
93 : blkno, path.str);
94 : else
95 0 : elog(elevel, "page %u of relation %s does not exist",
96 : blkno, path.str);
97 0 : }
98 :
99 : /* Log a reference to an invalid page */
100 : static void
101 0 : log_invalid_page(RelFileLocator locator, ForkNumber forkno, BlockNumber blkno,
102 : bool present)
103 : {
104 : xl_invalid_page_key key;
105 : xl_invalid_page *hentry;
106 : bool found;
107 :
108 : /*
109 : * Once recovery has reached a consistent state, the invalid-page table
110 : * should be empty and remain so. If a reference to an invalid page is
111 : * found after consistency is reached, PANIC immediately. This might seem
112 : * aggressive, but it's better than letting the invalid reference linger
113 : * in the hash table until the end of recovery and PANIC there, which
114 : * might come only much later if this is a standby server.
115 : */
116 0 : if (reachedConsistency)
117 : {
118 0 : report_invalid_page(WARNING, locator, forkno, blkno, present);
119 0 : elog(ignore_invalid_pages ? WARNING : PANIC,
120 : "WAL contains references to invalid pages");
121 : }
122 :
123 : /*
124 : * Log references to invalid pages at DEBUG1 level. This allows some
125 : * tracing of the cause (note the elog context mechanism will tell us
126 : * something about the XLOG record that generated the reference).
127 : */
128 0 : if (message_level_is_interesting(DEBUG1))
129 0 : report_invalid_page(DEBUG1, locator, forkno, blkno, present);
130 :
131 0 : if (invalid_page_tab == NULL)
132 : {
133 : /* create hash table when first needed */
134 : HASHCTL ctl;
135 :
136 0 : ctl.keysize = sizeof(xl_invalid_page_key);
137 0 : ctl.entrysize = sizeof(xl_invalid_page);
138 :
139 0 : invalid_page_tab = hash_create("XLOG invalid-page table",
140 : 100,
141 : &ctl,
142 : HASH_ELEM | HASH_BLOBS);
143 : }
144 :
145 : /* we currently assume xl_invalid_page_key contains no padding */
146 0 : key.locator = locator;
147 0 : key.forkno = forkno;
148 0 : key.blkno = blkno;
149 : hentry = (xl_invalid_page *)
150 0 : hash_search(invalid_page_tab, &key, HASH_ENTER, &found);
151 :
152 0 : if (!found)
153 : {
154 : /* hash_search already filled in the key */
155 0 : hentry->present = present;
156 : }
157 : else
158 : {
159 : /* repeat reference ... leave "present" as it was */
160 : }
161 0 : }
162 :
163 : /* Forget any invalid pages >= minblkno, because they've been dropped */
164 : static void
165 59880 : forget_invalid_pages(RelFileLocator locator, ForkNumber forkno,
166 : BlockNumber minblkno)
167 : {
168 : HASH_SEQ_STATUS status;
169 : xl_invalid_page *hentry;
170 :
171 59880 : if (invalid_page_tab == NULL)
172 59880 : return; /* nothing to do */
173 :
174 0 : hash_seq_init(&status, invalid_page_tab);
175 :
176 0 : while ((hentry = (xl_invalid_page *) hash_seq_search(&status)) != NULL)
177 : {
178 0 : if (RelFileLocatorEquals(hentry->key.locator, locator) &&
179 0 : hentry->key.forkno == forkno &&
180 0 : hentry->key.blkno >= minblkno)
181 : {
182 0 : elog(DEBUG2, "page %u of relation %s has been dropped",
183 : hentry->key.blkno,
184 : relpathperm(hentry->key.locator, forkno).str);
185 :
186 0 : if (hash_search(invalid_page_tab,
187 0 : &hentry->key,
188 : HASH_REMOVE, NULL) == NULL)
189 0 : elog(ERROR, "hash table corrupted");
190 : }
191 : }
192 : }
193 :
194 : /* Forget any invalid pages in a whole database */
195 : static void
196 26 : forget_invalid_pages_db(Oid dbid)
197 : {
198 : HASH_SEQ_STATUS status;
199 : xl_invalid_page *hentry;
200 :
201 26 : if (invalid_page_tab == NULL)
202 26 : return; /* nothing to do */
203 :
204 0 : hash_seq_init(&status, invalid_page_tab);
205 :
206 0 : while ((hentry = (xl_invalid_page *) hash_seq_search(&status)) != NULL)
207 : {
208 0 : if (hentry->key.locator.dbOid == dbid)
209 : {
210 0 : elog(DEBUG2, "page %u of relation %s has been dropped",
211 : hentry->key.blkno,
212 : relpathperm(hentry->key.locator, hentry->key.forkno).str);
213 :
214 0 : if (hash_search(invalid_page_tab,
215 0 : &hentry->key,
216 : HASH_REMOVE, NULL) == NULL)
217 0 : elog(ERROR, "hash table corrupted");
218 : }
219 : }
220 : }
221 :
222 : /* Are there any unresolved references to invalid pages? */
223 : bool
224 1406 : XLogHaveInvalidPages(void)
225 : {
226 1406 : if (invalid_page_tab != NULL &&
227 0 : hash_get_num_entries(invalid_page_tab) > 0)
228 0 : return true;
229 1406 : return false;
230 : }
231 :
232 : /* Complain about any remaining invalid-page entries */
233 : void
234 224 : XLogCheckInvalidPages(void)
235 : {
236 : HASH_SEQ_STATUS status;
237 : xl_invalid_page *hentry;
238 224 : bool foundone = false;
239 :
240 224 : if (invalid_page_tab == NULL)
241 224 : return; /* nothing to do */
242 :
243 0 : hash_seq_init(&status, invalid_page_tab);
244 :
245 : /*
246 : * Our strategy is to emit WARNING messages for all remaining entries and
247 : * only PANIC after we've dumped all the available info.
248 : */
249 0 : while ((hentry = (xl_invalid_page *) hash_seq_search(&status)) != NULL)
250 : {
251 0 : report_invalid_page(WARNING, hentry->key.locator, hentry->key.forkno,
252 0 : hentry->key.blkno, hentry->present);
253 0 : foundone = true;
254 : }
255 :
256 0 : if (foundone)
257 0 : elog(ignore_invalid_pages ? WARNING : PANIC,
258 : "WAL contains references to invalid pages");
259 :
260 0 : hash_destroy(invalid_page_tab);
261 0 : invalid_page_tab = NULL;
262 : }
263 :
264 :
265 : /*
266 : * XLogReadBufferForRedo
267 : * Read a page during XLOG replay
268 : *
269 : * Reads a block referenced by a WAL record into shared buffer cache, and
270 : * determines what needs to be done to redo the changes to it. If the WAL
271 : * record includes a full-page image of the page, it is restored.
272 : *
273 : * 'record.EndRecPtr' is compared to the page's LSN to determine if the record
274 : * has already been replayed. 'block_id' is the ID number the block was
275 : * registered with, when the WAL record was created.
276 : *
277 : * Returns one of the following:
278 : *
279 : * BLK_NEEDS_REDO - changes from the WAL record need to be applied
280 : * BLK_DONE - block doesn't need replaying
281 : * BLK_RESTORED - block was restored from a full-page image included in
282 : * the record
283 : * BLK_NOTFOUND - block was not found (because it was truncated away by
284 : * an operation later in the WAL stream)
285 : *
286 : * On return, the buffer is locked in exclusive-mode, and returned in *buf.
287 : * Note that the buffer is locked and returned even if it doesn't need
288 : * replaying. (Getting the buffer lock is not really necessary during
289 : * single-process crash recovery, but some subroutines such as MarkBufferDirty
290 : * will complain if we don't have the lock. In hot standby mode it's
291 : * definitely necessary.)
292 : *
293 : * Note: when a backup block is available in XLOG with the BKPIMAGE_APPLY flag
294 : * set, we restore it, even if the page in the database appears newer. This
295 : * is to protect ourselves against database pages that were partially or
296 : * incorrectly written during a crash. We assume that the XLOG data must be
297 : * good because it has passed a CRC check, while the database page might not
298 : * be. This will force us to replay all subsequent modifications of the page
299 : * that appear in XLOG, rather than possibly ignoring them as already
300 : * applied, but that's not a huge drawback.
301 : */
302 : XLogRedoAction
303 5588418 : XLogReadBufferForRedo(XLogReaderState *record, uint8 block_id,
304 : Buffer *buf)
305 : {
306 5588418 : return XLogReadBufferForRedoExtended(record, block_id, RBM_NORMAL,
307 : false, buf);
308 : }
309 :
310 : /*
311 : * Pin and lock a buffer referenced by a WAL record, for the purpose of
312 : * re-initializing it.
313 : */
314 : Buffer
315 103992 : XLogInitBufferForRedo(XLogReaderState *record, uint8 block_id)
316 : {
317 : Buffer buf;
318 :
319 103992 : XLogReadBufferForRedoExtended(record, block_id, RBM_ZERO_AND_LOCK, false,
320 : &buf);
321 103992 : return buf;
322 : }
323 :
324 : /*
325 : * XLogReadBufferForRedoExtended
326 : * Like XLogReadBufferForRedo, but with extra options.
327 : *
328 : * In RBM_ZERO_* modes, if the page doesn't exist, the relation is extended
329 : * with all-zeroes pages up to the referenced block number. In
330 : * RBM_ZERO_AND_LOCK and RBM_ZERO_AND_CLEANUP_LOCK modes, the return value
331 : * is always BLK_NEEDS_REDO.
332 : *
333 : * (The RBM_ZERO_AND_CLEANUP_LOCK mode is redundant with the get_cleanup_lock
334 : * parameter. Do not use an inconsistent combination!)
335 : *
336 : * If 'get_cleanup_lock' is true, a "cleanup lock" is acquired on the buffer
337 : * using LockBufferForCleanup(), instead of a regular exclusive lock.
338 : */
339 : XLogRedoAction
340 5726488 : XLogReadBufferForRedoExtended(XLogReaderState *record,
341 : uint8 block_id,
342 : ReadBufferMode mode, bool get_cleanup_lock,
343 : Buffer *buf)
344 : {
345 5726488 : XLogRecPtr lsn = record->EndRecPtr;
346 : RelFileLocator rlocator;
347 : ForkNumber forknum;
348 : BlockNumber blkno;
349 : Buffer prefetch_buffer;
350 : Page page;
351 : bool zeromode;
352 : bool willinit;
353 :
354 5726488 : if (!XLogRecGetBlockTagExtended(record, block_id, &rlocator, &forknum, &blkno,
355 : &prefetch_buffer))
356 : {
357 : /* Caller specified a bogus block_id */
358 0 : elog(PANIC, "failed to locate backup block with ID %d in WAL record",
359 : block_id);
360 : }
361 :
362 : /*
363 : * Make sure that if the block is marked with WILL_INIT, the caller is
364 : * going to initialize it. And vice versa.
365 : */
366 5726488 : zeromode = (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK);
367 5726488 : willinit = (XLogRecGetBlock(record, block_id)->flags & BKPBLOCK_WILL_INIT) != 0;
368 5726488 : if (willinit && !zeromode)
369 0 : elog(PANIC, "block with WILL_INIT flag in WAL record must be zeroed by redo routine");
370 5726488 : if (!willinit && zeromode)
371 0 : elog(PANIC, "block to be initialized in redo routine must be marked with WILL_INIT flag in the WAL record");
372 :
373 : /* If it has a full-page image and it should be restored, do it. */
374 5726488 : if (XLogRecBlockImageApply(record, block_id))
375 : {
376 : Assert(XLogRecHasBlockImage(record, block_id));
377 134458 : *buf = XLogReadBufferExtended(rlocator, forknum, blkno,
378 : get_cleanup_lock ? RBM_ZERO_AND_CLEANUP_LOCK : RBM_ZERO_AND_LOCK,
379 : prefetch_buffer);
380 134458 : page = BufferGetPage(*buf);
381 134458 : if (!RestoreBlockImage(record, block_id, page))
382 0 : ereport(ERROR,
383 : (errcode(ERRCODE_INTERNAL_ERROR),
384 : errmsg_internal("%s", record->errormsg_buf)));
385 :
386 : /*
387 : * The page may be uninitialized. If so, we can't set the LSN because
388 : * that would corrupt the page.
389 : */
390 134458 : if (!PageIsNew(page))
391 : {
392 134426 : PageSetLSN(page, lsn);
393 : }
394 :
395 134458 : MarkBufferDirty(*buf);
396 :
397 : /*
398 : * At the end of crash recovery the init forks of unlogged relations
399 : * are copied, without going through shared buffers. So we need to
400 : * force the on-disk state of init forks to always be in sync with the
401 : * state in shared buffers.
402 : */
403 134458 : if (forknum == INIT_FORKNUM)
404 52 : FlushOneBuffer(*buf);
405 :
406 134458 : return BLK_RESTORED;
407 : }
408 : else
409 : {
410 5592030 : *buf = XLogReadBufferExtended(rlocator, forknum, blkno, mode, prefetch_buffer);
411 5592030 : if (BufferIsValid(*buf))
412 : {
413 5592030 : if (mode != RBM_ZERO_AND_LOCK && mode != RBM_ZERO_AND_CLEANUP_LOCK)
414 : {
415 5487590 : if (get_cleanup_lock)
416 17014 : LockBufferForCleanup(*buf);
417 : else
418 5470576 : LockBuffer(*buf, BUFFER_LOCK_EXCLUSIVE);
419 : }
420 5592030 : if (lsn <= PageGetLSN(BufferGetPage(*buf)))
421 0 : return BLK_DONE;
422 : else
423 5592030 : return BLK_NEEDS_REDO;
424 : }
425 : else
426 0 : return BLK_NOTFOUND;
427 : }
428 : }
429 :
430 : /*
431 : * XLogReadBufferExtended
432 : * Read a page during XLOG replay
433 : *
434 : * This is functionally comparable to ReadBufferExtended. There's some
435 : * differences in the behavior wrt. the "mode" argument:
436 : *
437 : * In RBM_NORMAL mode, if the page doesn't exist, or contains all-zeroes, we
438 : * return InvalidBuffer. In this case the caller should silently skip the
439 : * update on this page. (In this situation, we expect that the page was later
440 : * dropped or truncated. If we don't see evidence of that later in the WAL
441 : * sequence, we'll complain at the end of WAL replay.)
442 : *
443 : * In RBM_ZERO_* modes, if the page doesn't exist, the relation is extended
444 : * with all-zeroes pages up to the given block number.
445 : *
446 : * In RBM_NORMAL_NO_LOG mode, we return InvalidBuffer if the page doesn't
447 : * exist, and we don't check for all-zeroes. Thus, no log entry is made
448 : * to imply that the page should be dropped or truncated later.
449 : *
450 : * Optionally, recent_buffer can be used to provide a hint about the location
451 : * of the page in the buffer pool; it does not have to be correct, but avoids
452 : * a buffer mapping table probe if it is.
453 : *
454 : * NB: A redo function should normally not call this directly. To get a page
455 : * to modify, use XLogReadBufferForRedoExtended instead. It is important that
456 : * all pages modified by a WAL record are registered in the WAL records, or
457 : * they will be invisible to tools that need to know which pages are modified.
458 : */
459 : Buffer
460 10872994 : XLogReadBufferExtended(RelFileLocator rlocator, ForkNumber forknum,
461 : BlockNumber blkno, ReadBufferMode mode,
462 : Buffer recent_buffer)
463 : {
464 : BlockNumber lastblock;
465 : Buffer buffer;
466 : SMgrRelation smgr;
467 :
468 : Assert(blkno != P_NEW);
469 :
470 : /* Do we have a clue where the buffer might be already? */
471 10872994 : if (BufferIsValid(recent_buffer) &&
472 8204 : mode == RBM_NORMAL &&
473 8204 : ReadRecentBuffer(rlocator, forknum, blkno, recent_buffer))
474 : {
475 8158 : buffer = recent_buffer;
476 8158 : goto recent_buffer_fast_path;
477 : }
478 :
479 : /* Open the relation at smgr level */
480 10864836 : smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
481 :
482 : /*
483 : * Create the target file if it doesn't already exist. This lets us cope
484 : * if the replay sequence contains writes to a relation that is later
485 : * deleted. (The original coding of this routine would instead suppress
486 : * the writes, but that seems like it risks losing valuable data if the
487 : * filesystem loses an inode during a crash. Better to write the data
488 : * until we are actually told to delete the file.)
489 : */
490 10864836 : smgrcreate(smgr, forknum, true);
491 :
492 10864836 : lastblock = smgrnblocks(smgr, forknum);
493 :
494 10864836 : if (blkno < lastblock)
495 : {
496 : /* page exists in file */
497 10777420 : buffer = ReadBufferWithoutRelcache(rlocator, forknum, blkno,
498 : mode, NULL, true);
499 : }
500 : else
501 : {
502 : /* hm, page doesn't exist in file */
503 87416 : if (mode == RBM_NORMAL)
504 : {
505 0 : log_invalid_page(rlocator, forknum, blkno, false);
506 0 : return InvalidBuffer;
507 : }
508 87416 : if (mode == RBM_NORMAL_NO_LOG)
509 0 : return InvalidBuffer;
510 : /* OK to extend the file */
511 : /* we do this in recovery only - no rel-extension lock needed */
512 : Assert(InRecovery);
513 87416 : buffer = ExtendBufferedRelTo(BMR_SMGR(smgr, RELPERSISTENCE_PERMANENT),
514 : forknum,
515 : NULL,
516 : EB_PERFORMING_RECOVERY |
517 : EB_SKIP_EXTENSION_LOCK,
518 : blkno + 1,
519 : mode);
520 : }
521 :
522 10872994 : recent_buffer_fast_path:
523 10872994 : if (mode == RBM_NORMAL)
524 : {
525 : /* check that page has been initialized */
526 5477948 : Page page = (Page) BufferGetPage(buffer);
527 :
528 : /*
529 : * We assume that PageIsNew is safe without a lock. During recovery,
530 : * there should be no other backends that could modify the buffer at
531 : * the same time.
532 : */
533 5477948 : if (PageIsNew(page))
534 : {
535 0 : ReleaseBuffer(buffer);
536 0 : log_invalid_page(rlocator, forknum, blkno, true);
537 0 : return InvalidBuffer;
538 : }
539 : }
540 :
541 10872994 : return buffer;
542 : }
543 :
544 : /*
545 : * Struct actually returned by CreateFakeRelcacheEntry, though the declared
546 : * return type is Relation.
547 : */
548 : typedef struct
549 : {
550 : RelationData reldata; /* Note: this must be first */
551 : FormData_pg_class pgc;
552 : } FakeRelCacheEntryData;
553 :
554 : typedef FakeRelCacheEntryData *FakeRelCacheEntry;
555 :
556 : /*
557 : * Create a fake relation cache entry for a physical relation
558 : *
559 : * It's often convenient to use the same functions in XLOG replay as in the
560 : * main codepath, but those functions typically work with a relcache entry.
561 : * We don't have a working relation cache during XLOG replay, but this
562 : * function can be used to create a fake relcache entry instead. Only the
563 : * fields related to physical storage, like rd_rel, are initialized, so the
564 : * fake entry is only usable in low-level operations like ReadBuffer().
565 : *
566 : * This is also used for syncing WAL-skipped files.
567 : *
568 : * Caller must free the returned entry with FreeFakeRelcacheEntry().
569 : */
570 : Relation
571 111552 : CreateFakeRelcacheEntry(RelFileLocator rlocator)
572 : {
573 : FakeRelCacheEntry fakeentry;
574 : Relation rel;
575 :
576 : /* Allocate the Relation struct and all related space in one block. */
577 111552 : fakeentry = palloc0(sizeof(FakeRelCacheEntryData));
578 111552 : rel = (Relation) fakeentry;
579 :
580 111552 : rel->rd_rel = &fakeentry->pgc;
581 111552 : rel->rd_locator = rlocator;
582 :
583 : /*
584 : * We will never be working with temp rels during recovery or while
585 : * syncing WAL-skipped files.
586 : */
587 111552 : rel->rd_backend = INVALID_PROC_NUMBER;
588 :
589 : /* It must be a permanent table here */
590 111552 : rel->rd_rel->relpersistence = RELPERSISTENCE_PERMANENT;
591 :
592 : /* We don't know the name of the relation; use relfilenumber instead */
593 111552 : sprintf(RelationGetRelationName(rel), "%u", rlocator.relNumber);
594 :
595 : /*
596 : * We set up the lockRelId in case anything tries to lock the dummy
597 : * relation. Note that this is fairly bogus since relNumber may be
598 : * different from the relation's OID. It shouldn't really matter though.
599 : * In recovery, we are running by ourselves and can't have any lock
600 : * conflicts. While syncing, we already hold AccessExclusiveLock.
601 : */
602 111552 : rel->rd_lockInfo.lockRelId.dbId = rlocator.dbOid;
603 111552 : rel->rd_lockInfo.lockRelId.relId = rlocator.relNumber;
604 :
605 : /*
606 : * Set up a non-pinned SMgrRelation reference, so that we don't need to
607 : * worry about unpinning it on error.
608 : */
609 111552 : rel->rd_smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
610 :
611 111552 : return rel;
612 : }
613 :
614 : /*
615 : * Free a fake relation cache entry.
616 : */
617 : void
618 111552 : FreeFakeRelcacheEntry(Relation fakerel)
619 : {
620 111552 : pfree(fakerel);
621 111552 : }
622 :
623 : /*
624 : * Drop a relation during XLOG replay
625 : *
626 : * This is called when the relation is about to be deleted; we need to remove
627 : * any open "invalid-page" records for the relation.
628 : */
629 : void
630 59784 : XLogDropRelation(RelFileLocator rlocator, ForkNumber forknum)
631 : {
632 59784 : forget_invalid_pages(rlocator, forknum, 0);
633 59784 : }
634 :
635 : /*
636 : * Drop a whole database during XLOG replay
637 : *
638 : * As above, but for DROP DATABASE instead of dropping a single rel
639 : */
640 : void
641 26 : XLogDropDatabase(Oid dbid)
642 : {
643 : /*
644 : * This is unnecessarily heavy-handed, as it will close SMgrRelation
645 : * objects for other databases as well. DROP DATABASE occurs seldom enough
646 : * that it's not worth introducing a variant of smgrdestroy for just this
647 : * purpose.
648 : */
649 26 : smgrdestroyall();
650 :
651 26 : forget_invalid_pages_db(dbid);
652 26 : }
653 :
654 : /*
655 : * Truncate a relation during XLOG replay
656 : *
657 : * We need to clean up any open "invalid-page" records for the dropped pages.
658 : */
659 : void
660 96 : XLogTruncateRelation(RelFileLocator rlocator, ForkNumber forkNum,
661 : BlockNumber nblocks)
662 : {
663 96 : forget_invalid_pages(rlocator, forkNum, nblocks);
664 96 : }
665 :
666 : /*
667 : * Determine which timeline to read an xlog page from and set the
668 : * XLogReaderState's currTLI to that timeline ID.
669 : *
670 : * We care about timelines in xlogreader when we might be reading xlog
671 : * generated prior to a promotion, either if we're currently a standby in
672 : * recovery or if we're a promoted primary reading xlogs generated by the old
673 : * primary before our promotion.
674 : *
675 : * wantPage must be set to the start address of the page to read and
676 : * wantLength to the amount of the page that will be read, up to
677 : * XLOG_BLCKSZ. If the amount to be read isn't known, pass XLOG_BLCKSZ.
678 : *
679 : * The currTLI argument should be the system-wide current timeline.
680 : * Note that this may be different from state->currTLI, which is the timeline
681 : * from which the caller is currently reading previous xlog records.
682 : *
683 : * We switch to an xlog segment from the new timeline eagerly when on a
684 : * historical timeline, as soon as we reach the start of the xlog segment
685 : * containing the timeline switch. The server copied the segment to the new
686 : * timeline so all the data up to the switch point is the same, but there's no
687 : * guarantee the old segment will still exist. It may have been deleted or
688 : * renamed with a .partial suffix so we can't necessarily keep reading from
689 : * the old TLI even though tliSwitchPoint says it's OK.
690 : *
691 : * We can't just check the timeline when we read a page on a different segment
692 : * to the last page. We could've received a timeline switch from a cascading
693 : * upstream, so the current segment ends abruptly (possibly getting renamed to
694 : * .partial) and we have to switch to a new one. Even in the middle of reading
695 : * a page we could have to dump the cached page and switch to a new TLI.
696 : *
697 : * Because of this, callers MAY NOT assume that currTLI is the timeline that
698 : * will be in a page's xlp_tli; the page may begin on an older timeline or we
699 : * might be reading from historical timeline data on a segment that's been
700 : * copied to a new timeline.
701 : *
702 : * The caller must also make sure it doesn't read past the current replay
703 : * position (using GetXLogReplayRecPtr) if executing in recovery, so it
704 : * doesn't fail to notice that the current timeline became historical.
705 : */
706 : void
707 99628 : XLogReadDetermineTimeline(XLogReaderState *state, XLogRecPtr wantPage,
708 : uint32 wantLength, TimeLineID currTLI)
709 : {
710 99628 : const XLogRecPtr lastReadPage = (state->seg.ws_segno *
711 99628 : state->segcxt.ws_segsize + state->segoff);
712 :
713 : Assert(wantPage != InvalidXLogRecPtr && wantPage % XLOG_BLCKSZ == 0);
714 : Assert(wantLength <= XLOG_BLCKSZ);
715 : Assert(state->readLen == 0 || state->readLen <= XLOG_BLCKSZ);
716 : Assert(currTLI != 0);
717 :
718 : /*
719 : * If the desired page is currently read in and valid, we have nothing to
720 : * do.
721 : *
722 : * The caller should've ensured that it didn't previously advance readOff
723 : * past the valid limit of this timeline, so it doesn't matter if the
724 : * current TLI has since become historical.
725 : */
726 99628 : if (lastReadPage == wantPage &&
727 3860 : state->readLen != 0 &&
728 0 : lastReadPage + state->readLen >= wantPage + Min(wantLength, XLOG_BLCKSZ - 1))
729 0 : return;
730 :
731 : /*
732 : * If we're reading from the current timeline, it hasn't become historical
733 : * and the page we're reading is after the last page read, we can again
734 : * just carry on. (Seeking backwards requires a check to make sure the
735 : * older page isn't on a prior timeline).
736 : *
737 : * currTLI might've become historical since the caller obtained the value,
738 : * but the caller is required not to read past the flush limit it saw at
739 : * the time it looked up the timeline. There's nothing we can do about it
740 : * if StartupXLOG() renames it to .partial concurrently.
741 : */
742 99628 : if (state->currTLI == currTLI && wantPage >= lastReadPage)
743 : {
744 : Assert(state->currTLIValidUntil == InvalidXLogRecPtr);
745 93520 : return;
746 : }
747 :
748 : /*
749 : * If we're just reading pages from a previously validated historical
750 : * timeline and the timeline we're reading from is valid until the end of
751 : * the current segment we can just keep reading.
752 : */
753 6108 : if (state->currTLIValidUntil != InvalidXLogRecPtr &&
754 3362 : state->currTLI != currTLI &&
755 3362 : state->currTLI != 0 &&
756 3362 : ((wantPage + wantLength) / state->segcxt.ws_segsize) <
757 3362 : (state->currTLIValidUntil / state->segcxt.ws_segsize))
758 3356 : return;
759 :
760 : /*
761 : * If we reach this point we're either looking up a page for random
762 : * access, the current timeline just became historical, or we're reading
763 : * from a new segment containing a timeline switch. In all cases we need
764 : * to determine the newest timeline on the segment.
765 : *
766 : * If it's the current timeline we can just keep reading from here unless
767 : * we detect a timeline switch that makes the current timeline historical.
768 : * If it's a historical timeline we can read all the segment on the newest
769 : * timeline because it contains all the old timelines' data too. So only
770 : * one switch check is required.
771 : */
772 : {
773 : /*
774 : * We need to re-read the timeline history in case it's been changed
775 : * by a promotion or replay from a cascaded replica.
776 : */
777 2752 : List *timelineHistory = readTimeLineHistory(currTLI);
778 : XLogRecPtr endOfSegment;
779 :
780 2752 : endOfSegment = ((wantPage / state->segcxt.ws_segsize) + 1) *
781 2752 : state->segcxt.ws_segsize - 1;
782 : Assert(wantPage / state->segcxt.ws_segsize ==
783 : endOfSegment / state->segcxt.ws_segsize);
784 :
785 : /*
786 : * Find the timeline of the last LSN on the segment containing
787 : * wantPage.
788 : */
789 2752 : state->currTLI = tliOfPointInHistory(endOfSegment, timelineHistory);
790 2752 : state->currTLIValidUntil = tliSwitchPoint(state->currTLI, timelineHistory,
791 : &state->nextTLI);
792 :
793 : Assert(state->currTLIValidUntil == InvalidXLogRecPtr ||
794 : wantPage + wantLength < state->currTLIValidUntil);
795 :
796 2752 : list_free_deep(timelineHistory);
797 :
798 2752 : elog(DEBUG3, "switched to timeline %u valid until %X/%X",
799 : state->currTLI,
800 : LSN_FORMAT_ARGS(state->currTLIValidUntil));
801 : }
802 : }
803 :
804 : /* XLogReaderRoutine->segment_open callback for local pg_wal files */
805 : void
806 1492 : wal_segment_open(XLogReaderState *state, XLogSegNo nextSegNo,
807 : TimeLineID *tli_p)
808 : {
809 1492 : TimeLineID tli = *tli_p;
810 : char path[MAXPGPATH];
811 :
812 1492 : XLogFilePath(path, tli, nextSegNo, state->segcxt.ws_segsize);
813 1492 : state->seg.ws_file = BasicOpenFile(path, O_RDONLY | PG_BINARY);
814 1492 : if (state->seg.ws_file >= 0)
815 1492 : return;
816 :
817 0 : if (errno == ENOENT)
818 0 : ereport(ERROR,
819 : (errcode_for_file_access(),
820 : errmsg("requested WAL segment %s has already been removed",
821 : path)));
822 : else
823 0 : ereport(ERROR,
824 : (errcode_for_file_access(),
825 : errmsg("could not open file \"%s\": %m",
826 : path)));
827 : }
828 :
829 : /* stock XLogReaderRoutine->segment_close callback */
830 : void
831 32036 : wal_segment_close(XLogReaderState *state)
832 : {
833 32036 : close(state->seg.ws_file);
834 : /* need to check errno? */
835 32036 : state->seg.ws_file = -1;
836 32036 : }
837 :
838 : /*
839 : * XLogReaderRoutine->page_read callback for reading local xlog files
840 : *
841 : * Public because it would likely be very helpful for someone writing another
842 : * output method outside walsender, e.g. in a bgworker.
843 : */
844 : int
845 39472 : read_local_xlog_page(XLogReaderState *state, XLogRecPtr targetPagePtr,
846 : int reqLen, XLogRecPtr targetRecPtr, char *cur_page)
847 : {
848 39472 : return read_local_xlog_page_guts(state, targetPagePtr, reqLen,
849 : targetRecPtr, cur_page, true);
850 : }
851 :
852 : /*
853 : * Same as read_local_xlog_page except that it doesn't wait for future WAL
854 : * to be available.
855 : */
856 : int
857 7658 : read_local_xlog_page_no_wait(XLogReaderState *state, XLogRecPtr targetPagePtr,
858 : int reqLen, XLogRecPtr targetRecPtr,
859 : char *cur_page)
860 : {
861 7658 : return read_local_xlog_page_guts(state, targetPagePtr, reqLen,
862 : targetRecPtr, cur_page, false);
863 : }
864 :
865 : /*
866 : * Implementation of read_local_xlog_page and its no wait version.
867 : */
868 : static int
869 47130 : read_local_xlog_page_guts(XLogReaderState *state, XLogRecPtr targetPagePtr,
870 : int reqLen, XLogRecPtr targetRecPtr,
871 : char *cur_page, bool wait_for_wal)
872 : {
873 : XLogRecPtr read_upto,
874 : loc;
875 : TimeLineID tli;
876 : int count;
877 : WALReadError errinfo;
878 : TimeLineID currTLI;
879 :
880 47130 : loc = targetPagePtr + reqLen;
881 :
882 : /*
883 : * Loop waiting for xlog to be available if necessary
884 : *
885 : * TODO: The walsender has its own version of this function, which uses a
886 : * condition variable to wake up whenever WAL is flushed. We could use the
887 : * same infrastructure here, instead of the check/sleep/repeat style of
888 : * loop.
889 : */
890 : while (1)
891 : {
892 : /*
893 : * Determine the limit of xlog we can currently read to, and what the
894 : * most recent timeline is.
895 : */
896 48524 : if (!RecoveryInProgress())
897 46976 : read_upto = GetFlushRecPtr(&currTLI);
898 : else
899 1548 : read_upto = GetXLogReplayRecPtr(&currTLI);
900 48524 : tli = currTLI;
901 :
902 : /*
903 : * Check which timeline to get the record from.
904 : *
905 : * We have to do it each time through the loop because if we're in
906 : * recovery as a cascading standby, the current timeline might've
907 : * become historical. We can't rely on RecoveryInProgress() because in
908 : * a standby configuration like
909 : *
910 : * A => B => C
911 : *
912 : * if we're a logical decoding session on C, and B gets promoted, our
913 : * timeline will change while we remain in recovery.
914 : *
915 : * We can't just keep reading from the old timeline as the last WAL
916 : * archive in the timeline will get renamed to .partial by
917 : * StartupXLOG().
918 : *
919 : * If that happens after our caller determined the TLI but before we
920 : * actually read the xlog page, we might still try to read from the
921 : * old (now renamed) segment and fail. There's not much we can do
922 : * about this, but it can only happen when we're a leaf of a cascading
923 : * standby whose primary gets promoted while we're decoding, so a
924 : * one-off ERROR isn't too bad.
925 : */
926 48524 : XLogReadDetermineTimeline(state, targetPagePtr, reqLen, tli);
927 :
928 48524 : if (state->currTLI == currTLI)
929 : {
930 :
931 45162 : if (loc <= read_upto)
932 43748 : break;
933 :
934 : /* If asked, let's not wait for future WAL. */
935 1414 : if (!wait_for_wal)
936 : {
937 : ReadLocalXLogPageNoWaitPrivate *private_data;
938 :
939 : /*
940 : * Inform the caller of read_local_xlog_page_no_wait that the
941 : * end of WAL has been reached.
942 : */
943 20 : private_data = (ReadLocalXLogPageNoWaitPrivate *)
944 : state->private_data;
945 20 : private_data->end_of_wal = true;
946 20 : break;
947 : }
948 :
949 1394 : CHECK_FOR_INTERRUPTS();
950 1394 : pg_usleep(1000L);
951 : }
952 : else
953 : {
954 : /*
955 : * We're on a historical timeline, so limit reading to the switch
956 : * point where we moved to the next timeline.
957 : *
958 : * We don't need to GetFlushRecPtr or GetXLogReplayRecPtr. We know
959 : * about the new timeline, so we must've received past the end of
960 : * it.
961 : */
962 3362 : read_upto = state->currTLIValidUntil;
963 :
964 : /*
965 : * Setting tli to our wanted record's TLI is slightly wrong; the
966 : * page might begin on an older timeline if it contains a timeline
967 : * switch, since its xlog segment will have been copied from the
968 : * prior timeline. This is pretty harmless though, as nothing
969 : * cares so long as the timeline doesn't go backwards. We should
970 : * read the page header instead; FIXME someday.
971 : */
972 3362 : tli = state->currTLI;
973 :
974 : /* No need to wait on a historical timeline */
975 3362 : break;
976 : }
977 : }
978 :
979 47130 : if (targetPagePtr + XLOG_BLCKSZ <= read_upto)
980 : {
981 : /*
982 : * more than one block available; read only that block, have caller
983 : * come back if they need more.
984 : */
985 45714 : count = XLOG_BLCKSZ;
986 : }
987 1416 : else if (targetPagePtr + reqLen > read_upto)
988 : {
989 : /* not enough data there */
990 20 : return -1;
991 : }
992 : else
993 : {
994 : /* enough bytes available to satisfy the request */
995 1396 : count = read_upto - targetPagePtr;
996 : }
997 :
998 47110 : if (!WALRead(state, cur_page, targetPagePtr, count, tli,
999 : &errinfo))
1000 0 : WALReadRaiseError(&errinfo);
1001 :
1002 : /* number of valid bytes in the buffer */
1003 47110 : return count;
1004 : }
1005 :
1006 : /*
1007 : * Backend-specific convenience code to handle read errors encountered by
1008 : * WALRead().
1009 : */
1010 : void
1011 0 : WALReadRaiseError(WALReadError *errinfo)
1012 : {
1013 0 : WALOpenSegment *seg = &errinfo->wre_seg;
1014 : char fname[MAXFNAMELEN];
1015 :
1016 0 : XLogFileName(fname, seg->ws_tli, seg->ws_segno, wal_segment_size);
1017 :
1018 0 : if (errinfo->wre_read < 0)
1019 : {
1020 0 : errno = errinfo->wre_errno;
1021 0 : ereport(ERROR,
1022 : (errcode_for_file_access(),
1023 : errmsg("could not read from WAL segment %s, offset %d: %m",
1024 : fname, errinfo->wre_off)));
1025 : }
1026 0 : else if (errinfo->wre_read == 0)
1027 : {
1028 0 : ereport(ERROR,
1029 : (errcode(ERRCODE_DATA_CORRUPTED),
1030 : errmsg("could not read from WAL segment %s, offset %d: read %d of %d",
1031 : fname, errinfo->wre_off, errinfo->wre_read,
1032 : errinfo->wre_req)));
1033 : }
1034 0 : }
|
