Branch data 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) 1996-2026, 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 : 34915 : forget_invalid_pages(RelFileLocator locator, ForkNumber forkno,
166 : : BlockNumber minblkno)
167 : : {
168 : : HASH_SEQ_STATUS status;
169 : : xl_invalid_page *hentry;
170 : :
171 [ + - ]: 34915 : if (invalid_page_tab == NULL)
172 : 34915 : 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 : 15 : forget_invalid_pages_db(Oid dbid)
197 : : {
198 : : HASH_SEQ_STATUS status;
199 : : xl_invalid_page *hentry;
200 : :
201 [ + - ]: 15 : if (invalid_page_tab == NULL)
202 : 15 : 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 : 757 : XLogHaveInvalidPages(void)
225 : : {
226 [ - + - - ]: 757 : if (invalid_page_tab != NULL &&
227 : 0 : hash_get_num_entries(invalid_page_tab) > 0)
228 : 0 : return true;
229 : 757 : return false;
230 : : }
231 : :
232 : : /* Complain about any remaining invalid-page entries */
233 : : void
234 : 133 : XLogCheckInvalidPages(void)
235 : : {
236 : : HASH_SEQ_STATUS status;
237 : : xl_invalid_page *hentry;
238 : 133 : bool foundone = false;
239 : :
240 [ + - ]: 133 : if (invalid_page_tab == NULL)
241 : 133 : 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 : 2964391 : XLogReadBufferForRedo(XLogReaderState *record, uint8 block_id,
304 : : Buffer *buf)
305 : : {
306 : 2964391 : 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 : 53111 : XLogInitBufferForRedo(XLogReaderState *record, uint8 block_id)
316 : : {
317 : : Buffer buf;
318 : :
319 : 53111 : XLogReadBufferForRedoExtended(record, block_id, RBM_ZERO_AND_LOCK, false,
320 : : &buf);
321 : 53111 : return buf;
322 : : }
323 : :
324 : : /*
325 : : * If a redo routine modified an init fork, flush the buffer immediately.
326 : : *
327 : : * At the end of crash recovery the init forks of unlogged relations are
328 : : * copied to the main fork directly from disk, without going through shared
329 : : * buffers. Therefore, redo routines that update init forks without
330 : : * restoring a full-page image must call this after setting the page LSN and
331 : : * marking the buffer dirty.
332 : : */
333 : : void
334 : 2431 : XLogFlushBufferForRedoIfInit(XLogReaderState *record, uint8 block_id,
335 : : Buffer buffer)
336 : : {
337 : : ForkNumber forknum;
338 : :
339 : : Assert(BufferIsValid(buffer));
340 : :
341 : 2431 : XLogRecGetBlockTag(record, block_id, NULL, &forknum, NULL);
342 [ + + ]: 2431 : if (forknum == INIT_FORKNUM)
343 : 25 : FlushOneBuffer(buffer);
344 : 2431 : }
345 : :
346 : : /*
347 : : * XLogReadBufferForRedoExtended
348 : : * Like XLogReadBufferForRedo, but with extra options.
349 : : *
350 : : * In RBM_ZERO_* modes, if the page doesn't exist, the relation is extended
351 : : * with all-zeroes pages up to the referenced block number. In
352 : : * RBM_ZERO_AND_LOCK and RBM_ZERO_AND_CLEANUP_LOCK modes, the return value
353 : : * is always BLK_NEEDS_REDO.
354 : : *
355 : : * (The RBM_ZERO_AND_CLEANUP_LOCK mode is redundant with the get_cleanup_lock
356 : : * parameter. Do not use an inconsistent combination!)
357 : : *
358 : : * If 'get_cleanup_lock' is true, a "cleanup lock" is acquired on the buffer
359 : : * using LockBufferForCleanup(), instead of a regular exclusive lock.
360 : : */
361 : : XLogRedoAction
362 : 3047783 : XLogReadBufferForRedoExtended(XLogReaderState *record,
363 : : uint8 block_id,
364 : : ReadBufferMode mode, bool get_cleanup_lock,
365 : : Buffer *buf)
366 : : {
367 : 3047783 : XLogRecPtr lsn = record->EndRecPtr;
368 : : RelFileLocator rlocator;
369 : : ForkNumber forknum;
370 : : BlockNumber blkno;
371 : : Buffer prefetch_buffer;
372 : : Page page;
373 : : bool zeromode;
374 : : bool willinit;
375 : :
376 [ - + ]: 3047783 : if (!XLogRecGetBlockTagExtended(record, block_id, &rlocator, &forknum, &blkno,
377 : : &prefetch_buffer))
378 : : {
379 : : /* Caller specified a bogus block_id */
380 [ # # ]: 0 : elog(PANIC, "failed to locate backup block with ID %d in WAL record",
381 : : block_id);
382 : : }
383 : :
384 : : /*
385 : : * Make sure that if the block is marked with WILL_INIT, the caller is
386 : : * going to initialize it. And vice versa.
387 : : */
388 [ + + + + ]: 3047783 : zeromode = (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK);
389 : 3047783 : willinit = (XLogRecGetBlock(record, block_id)->flags & BKPBLOCK_WILL_INIT) != 0;
390 [ + + - + ]: 3047783 : if (willinit && !zeromode)
391 [ # # ]: 0 : elog(PANIC, "block with WILL_INIT flag in WAL record must be zeroed by redo routine");
392 [ + + - + ]: 3047783 : if (!willinit && zeromode)
393 [ # # ]: 0 : elog(PANIC, "block to be initialized in redo routine must be marked with WILL_INIT flag in the WAL record");
394 : :
395 : : /* If it has a full-page image and it should be restored, do it. */
396 [ + + ]: 3047783 : if (XLogRecBlockImageApply(record, block_id))
397 : : {
398 : : Assert(XLogRecHasBlockImage(record, block_id));
399 [ + + ]: 88309 : *buf = XLogReadBufferExtended(rlocator, forknum, blkno,
400 : : get_cleanup_lock ? RBM_ZERO_AND_CLEANUP_LOCK : RBM_ZERO_AND_LOCK,
401 : : prefetch_buffer);
402 : 88309 : page = BufferGetPage(*buf);
403 [ - + ]: 88309 : if (!RestoreBlockImage(record, block_id, page))
404 [ # # ]: 0 : ereport(ERROR,
405 : : (errcode(ERRCODE_INTERNAL_ERROR),
406 : : errmsg_internal("%s", record->errormsg_buf)));
407 : :
408 : : /*
409 : : * The page may be uninitialized. If so, we can't set the LSN because
410 : : * that would corrupt the page.
411 : : */
412 [ + + ]: 88309 : if (!PageIsNew(page))
413 : : {
414 : 88287 : PageSetLSN(page, lsn);
415 : : }
416 : :
417 : 88309 : MarkBufferDirty(*buf);
418 : :
419 : : /*
420 : : * At the end of crash recovery the init forks of unlogged relations
421 : : * are copied, without going through shared buffers. So we need to
422 : : * force the on-disk state of init forks to always be in sync with the
423 : : * state in shared buffers. Use XLogFlushBufferForRedoIfInit() for
424 : : * redo routines that dirty init-fork buffers without restoring a
425 : : * full-page image.
426 : : */
427 [ + + ]: 88309 : if (forknum == INIT_FORKNUM)
428 : 37 : FlushOneBuffer(*buf);
429 : :
430 : 88309 : return BLK_RESTORED;
431 : : }
432 : : else
433 : : {
434 : 2959474 : *buf = XLogReadBufferExtended(rlocator, forknum, blkno, mode, prefetch_buffer);
435 [ + - ]: 2959474 : if (BufferIsValid(*buf))
436 : : {
437 [ + + + + ]: 2959474 : if (mode != RBM_ZERO_AND_LOCK && mode != RBM_ZERO_AND_CLEANUP_LOCK)
438 : : {
439 [ + + ]: 2906142 : if (get_cleanup_lock)
440 : 12734 : LockBufferForCleanup(*buf);
441 : : else
442 : 2893408 : LockBuffer(*buf, BUFFER_LOCK_EXCLUSIVE);
443 : : }
444 [ - + ]: 2959474 : if (lsn <= PageGetLSN(BufferGetPage(*buf)))
445 : 0 : return BLK_DONE;
446 : : else
447 : 2959474 : return BLK_NEEDS_REDO;
448 : : }
449 : : else
450 : 0 : return BLK_NOTFOUND;
451 : : }
452 : : }
453 : :
454 : : /*
455 : : * XLogReadBufferExtended
456 : : * Read a page during XLOG replay
457 : : *
458 : : * This is functionally comparable to ReadBufferExtended. There's some
459 : : * differences in the behavior wrt. the "mode" argument:
460 : : *
461 : : * In RBM_NORMAL mode, if the page doesn't exist, or contains all-zeroes, we
462 : : * return InvalidBuffer. In this case the caller should silently skip the
463 : : * update on this page. (In this situation, we expect that the page was later
464 : : * dropped or truncated. If we don't see evidence of that later in the WAL
465 : : * sequence, we'll complain at the end of WAL replay.)
466 : : *
467 : : * In RBM_ZERO_* modes, if the page doesn't exist, the relation is extended
468 : : * with all-zeroes pages up to the given block number.
469 : : *
470 : : * In RBM_NORMAL_NO_LOG mode, we return InvalidBuffer if the page doesn't
471 : : * exist, and we don't check for all-zeroes. Thus, no log entry is made
472 : : * to imply that the page should be dropped or truncated later.
473 : : *
474 : : * Optionally, recent_buffer can be used to provide a hint about the location
475 : : * of the page in the buffer pool; it does not have to be correct, but avoids
476 : : * a buffer mapping table probe if it is.
477 : : *
478 : : * NB: A redo function should normally not call this directly. To get a page
479 : : * to modify, use XLogReadBufferForRedoExtended instead. It is important that
480 : : * all pages modified by a WAL record are registered in the WAL records, or
481 : : * they will be invisible to tools that need to know which pages are modified.
482 : : */
483 : : Buffer
484 : 5775276 : XLogReadBufferExtended(RelFileLocator rlocator, ForkNumber forknum,
485 : : BlockNumber blkno, ReadBufferMode mode,
486 : : Buffer recent_buffer)
487 : : {
488 : : BlockNumber lastblock;
489 : : Buffer buffer;
490 : : SMgrRelation smgr;
491 : :
492 : : Assert(blkno != P_NEW);
493 : :
494 : : /* Do we have a clue where the buffer might be already? */
495 [ + + + - ]: 5775276 : if (BufferIsValid(recent_buffer) &&
496 [ + + ]: 4978 : mode == RBM_NORMAL &&
497 : 4978 : ReadRecentBuffer(rlocator, forknum, blkno, recent_buffer))
498 : : {
499 : 4946 : buffer = recent_buffer;
500 : 4946 : goto recent_buffer_fast_path;
501 : : }
502 : :
503 : : /* Open the relation at smgr level */
504 : 5770330 : smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
505 : :
506 : : /*
507 : : * Create the target file if it doesn't already exist. This lets us cope
508 : : * if the replay sequence contains writes to a relation that is later
509 : : * deleted. (The original coding of this routine would instead suppress
510 : : * the writes, but that seems like it risks losing valuable data if the
511 : : * filesystem loses an inode during a crash. Better to write the data
512 : : * until we are actually told to delete the file.)
513 : : */
514 : 5770330 : smgrcreate(smgr, forknum, true);
515 : :
516 : 5770330 : lastblock = smgrnblocks(smgr, forknum);
517 : :
518 [ + + ]: 5770330 : if (blkno < lastblock)
519 : : {
520 : : /* page exists in file */
521 : 5723859 : buffer = ReadBufferWithoutRelcache(rlocator, forknum, blkno,
522 : : mode, NULL, true);
523 : : }
524 : : else
525 : : {
526 : : /* hm, page doesn't exist in file */
527 [ - + ]: 46471 : if (mode == RBM_NORMAL)
528 : : {
529 : 0 : log_invalid_page(rlocator, forknum, blkno, false);
530 : 0 : return InvalidBuffer;
531 : : }
532 [ - + ]: 46471 : if (mode == RBM_NORMAL_NO_LOG)
533 : 0 : return InvalidBuffer;
534 : : /* OK to extend the file */
535 : : /* we do this in recovery only - no rel-extension lock needed */
536 : : Assert(InRecovery);
537 : 46471 : buffer = ExtendBufferedRelTo(BMR_SMGR(smgr, RELPERSISTENCE_PERMANENT),
538 : : forknum,
539 : : NULL,
540 : : EB_PERFORMING_RECOVERY |
541 : : EB_SKIP_EXTENSION_LOCK,
542 : : blkno + 1,
543 : : mode);
544 : : }
545 : :
546 : 5775276 : recent_buffer_fast_path:
547 [ + + ]: 5775276 : if (mode == RBM_NORMAL)
548 : : {
549 : : /* check that page has been initialized */
550 : 2897539 : Page page = BufferGetPage(buffer);
551 : :
552 : : /*
553 : : * We assume that PageIsNew is safe without a lock. During recovery,
554 : : * there should be no other backends that could modify the buffer at
555 : : * the same time.
556 : : */
557 [ - + ]: 2897539 : if (PageIsNew(page))
558 : : {
559 : 0 : ReleaseBuffer(buffer);
560 : 0 : log_invalid_page(rlocator, forknum, blkno, true);
561 : 0 : return InvalidBuffer;
562 : : }
563 : : }
564 : :
565 : 5775276 : return buffer;
566 : : }
567 : :
568 : : /*
569 : : * Struct actually returned by CreateFakeRelcacheEntry, though the declared
570 : : * return type is Relation.
571 : : */
572 : : typedef struct
573 : : {
574 : : RelationData reldata; /* Note: this must be first */
575 : : FormData_pg_class pgc;
576 : : } FakeRelCacheEntryData;
577 : :
578 : : typedef FakeRelCacheEntryData *FakeRelCacheEntry;
579 : :
580 : : /*
581 : : * Create a fake relation cache entry for a physical relation
582 : : *
583 : : * It's often convenient to use the same functions in XLOG replay as in the
584 : : * main codepath, but those functions typically work with a relcache entry.
585 : : * We don't have a working relation cache during XLOG replay, but this
586 : : * function can be used to create a fake relcache entry instead. Only the
587 : : * fields related to physical storage, like rd_rel, are initialized, so the
588 : : * fake entry is only usable in low-level operations like ReadBuffer().
589 : : *
590 : : * This is also used for syncing WAL-skipped files.
591 : : *
592 : : * Caller must free the returned entry with FreeFakeRelcacheEntry().
593 : : */
594 : : Relation
595 : 43252 : CreateFakeRelcacheEntry(RelFileLocator rlocator)
596 : : {
597 : : FakeRelCacheEntry fakeentry;
598 : : Relation rel;
599 : :
600 : : /* Allocate the Relation struct and all related space in one block. */
601 : 43252 : fakeentry = palloc0_object(FakeRelCacheEntryData);
602 : 43252 : rel = (Relation) fakeentry;
603 : :
604 : 43252 : rel->rd_rel = &fakeentry->pgc;
605 : 43252 : rel->rd_locator = rlocator;
606 : :
607 : : /*
608 : : * We will never be working with temp rels during recovery or while
609 : : * syncing WAL-skipped files.
610 : : */
611 : 43252 : rel->rd_backend = INVALID_PROC_NUMBER;
612 : :
613 : : /* It must be a permanent table here */
614 : 43252 : rel->rd_rel->relpersistence = RELPERSISTENCE_PERMANENT;
615 : :
616 : : /* We don't know the name of the relation; use relfilenumber instead */
617 : 43252 : sprintf(RelationGetRelationName(rel), "%u", rlocator.relNumber);
618 : :
619 : : /*
620 : : * We set up the lockRelId in case anything tries to lock the dummy
621 : : * relation. Note that this is fairly bogus since relNumber may be
622 : : * different from the relation's OID. It shouldn't really matter though.
623 : : * In recovery, we are running by ourselves and can't have any lock
624 : : * conflicts. While syncing, we already hold AccessExclusiveLock.
625 : : */
626 : 43252 : rel->rd_lockInfo.lockRelId.dbId = rlocator.dbOid;
627 : 43252 : rel->rd_lockInfo.lockRelId.relId = rlocator.relNumber;
628 : :
629 : : /*
630 : : * Set up a non-pinned SMgrRelation reference, so that we don't need to
631 : : * worry about unpinning it on error.
632 : : */
633 : 43252 : rel->rd_smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
634 : :
635 : 43252 : return rel;
636 : : }
637 : :
638 : : /*
639 : : * Free a fake relation cache entry.
640 : : */
641 : : void
642 : 43252 : FreeFakeRelcacheEntry(Relation fakerel)
643 : : {
644 : 43252 : pfree(fakerel);
645 : 43252 : }
646 : :
647 : : /*
648 : : * Drop a relation during XLOG replay
649 : : *
650 : : * This is called when the relation is about to be deleted; we need to remove
651 : : * any open "invalid-page" records for the relation.
652 : : */
653 : : void
654 : 34860 : XLogDropRelation(RelFileLocator rlocator, ForkNumber forknum)
655 : : {
656 : 34860 : forget_invalid_pages(rlocator, forknum, 0);
657 : 34860 : }
658 : :
659 : : /*
660 : : * Drop a whole database during XLOG replay
661 : : *
662 : : * As above, but for DROP DATABASE instead of dropping a single rel
663 : : */
664 : : void
665 : 15 : XLogDropDatabase(Oid dbid)
666 : : {
667 : : /*
668 : : * This is unnecessarily heavy-handed, as it will close SMgrRelation
669 : : * objects for other databases as well. DROP DATABASE occurs seldom enough
670 : : * that it's not worth introducing a variant of smgrdestroy for just this
671 : : * purpose.
672 : : */
673 : 15 : smgrdestroyall();
674 : :
675 : 15 : forget_invalid_pages_db(dbid);
676 : 15 : }
677 : :
678 : : /*
679 : : * Truncate a relation during XLOG replay
680 : : *
681 : : * We need to clean up any open "invalid-page" records for the dropped pages.
682 : : */
683 : : void
684 : 55 : XLogTruncateRelation(RelFileLocator rlocator, ForkNumber forkNum,
685 : : BlockNumber nblocks)
686 : : {
687 : 55 : forget_invalid_pages(rlocator, forkNum, nblocks);
688 : 55 : }
689 : :
690 : : /*
691 : : * Determine which timeline to read an xlog page from and set the
692 : : * XLogReaderState's currTLI to that timeline ID.
693 : : *
694 : : * We care about timelines in xlogreader when we might be reading xlog
695 : : * generated prior to a promotion, either if we're currently a standby in
696 : : * recovery or if we're a promoted primary reading xlogs generated by the old
697 : : * primary before our promotion.
698 : : *
699 : : * wantPage must be set to the start address of the page to read and
700 : : * wantLength to the amount of the page that will be read, up to
701 : : * XLOG_BLCKSZ. If the amount to be read isn't known, pass XLOG_BLCKSZ.
702 : : *
703 : : * The currTLI argument should be the system-wide current timeline.
704 : : * Note that this may be different from state->currTLI, which is the timeline
705 : : * from which the caller is currently reading previous xlog records.
706 : : *
707 : : * We switch to an xlog segment from the new timeline eagerly when on a
708 : : * historical timeline, as soon as we reach the start of the xlog segment
709 : : * containing the timeline switch. The server copied the segment to the new
710 : : * timeline so all the data up to the switch point is the same, but there's no
711 : : * guarantee the old segment will still exist. It may have been deleted or
712 : : * renamed with a .partial suffix so we can't necessarily keep reading from
713 : : * the old TLI even though tliSwitchPoint says it's OK.
714 : : *
715 : : * We can't just check the timeline when we read a page on a different segment
716 : : * to the last page. We could've received a timeline switch from a cascading
717 : : * upstream, so the current segment ends abruptly (possibly getting renamed to
718 : : * .partial) and we have to switch to a new one. Even in the middle of reading
719 : : * a page we could have to dump the cached page and switch to a new TLI.
720 : : *
721 : : * Because of this, callers MAY NOT assume that currTLI is the timeline that
722 : : * will be in a page's xlp_tli; the page may begin on an older timeline or we
723 : : * might be reading from historical timeline data on a segment that's been
724 : : * copied to a new timeline.
725 : : *
726 : : * The caller must also make sure it doesn't read past the current replay
727 : : * position (using GetXLogReplayRecPtr) if executing in recovery, so it
728 : : * doesn't fail to notice that the current timeline became historical.
729 : : */
730 : : void
731 : 36174 : XLogReadDetermineTimeline(XLogReaderState *state, XLogRecPtr wantPage,
732 : : uint32 wantLength, TimeLineID currTLI)
733 : : {
734 : 36174 : const XLogRecPtr lastReadPage = (state->seg.ws_segno *
735 : 36174 : state->segcxt.ws_segsize + state->segoff);
736 : :
737 : : Assert(XLogRecPtrIsValid(wantPage) && wantPage % XLOG_BLCKSZ == 0);
738 : : Assert(wantLength <= XLOG_BLCKSZ);
739 : : Assert(state->readLen == 0 || state->readLen <= XLOG_BLCKSZ);
740 : : Assert(currTLI != 0);
741 : :
742 : : /*
743 : : * If the desired page is currently read in and valid, we have nothing to
744 : : * do.
745 : : *
746 : : * The caller should've ensured that it didn't previously advance readOff
747 : : * past the valid limit of this timeline, so it doesn't matter if the
748 : : * current TLI has since become historical.
749 : : */
750 [ + + ]: 36174 : if (lastReadPage == wantPage &&
751 [ - + ]: 2088 : state->readLen != 0 &&
752 [ # # ]: 0 : lastReadPage + state->readLen >= wantPage + Min(wantLength, XLOG_BLCKSZ - 1))
753 : 0 : return;
754 : :
755 : : /*
756 : : * If we're reading from the current timeline, it hasn't become historical
757 : : * and the page we're reading is after the last page read, we can again
758 : : * just carry on. (Seeking backwards requires a check to make sure the
759 : : * older page isn't on a prior timeline).
760 : : *
761 : : * currTLI might've become historical since the caller obtained the value,
762 : : * but the caller is required not to read past the flush limit it saw at
763 : : * the time it looked up the timeline. There's nothing we can do about it
764 : : * if StartupXLOG() renames it to .partial concurrently.
765 : : */
766 [ + + + + ]: 36174 : if (state->currTLI == currTLI && wantPage >= lastReadPage)
767 : : {
768 : : Assert(!XLogRecPtrIsValid(state->currTLIValidUntil));
769 : 33491 : return;
770 : : }
771 : :
772 : : /*
773 : : * If we're just reading pages from a previously validated historical
774 : : * timeline and the timeline we're reading from is valid until the end of
775 : : * the current segment we can just keep reading.
776 : : */
777 [ + + ]: 2683 : if (XLogRecPtrIsValid(state->currTLIValidUntil) &&
778 [ + - ]: 1159 : state->currTLI != currTLI &&
779 [ + - ]: 1159 : state->currTLI != 0 &&
780 : 1159 : ((wantPage + wantLength) / state->segcxt.ws_segsize) <
781 [ + + ]: 1159 : (state->currTLIValidUntil / state->segcxt.ws_segsize))
782 : 1157 : return;
783 : :
784 : : /*
785 : : * If we reach this point we're either looking up a page for random
786 : : * access, the current timeline just became historical, or we're reading
787 : : * from a new segment containing a timeline switch. In all cases we need
788 : : * to determine the newest timeline on the segment.
789 : : *
790 : : * If it's the current timeline we can just keep reading from here unless
791 : : * we detect a timeline switch that makes the current timeline historical.
792 : : * If it's a historical timeline we can read all the segment on the newest
793 : : * timeline because it contains all the old timelines' data too. So only
794 : : * one switch check is required.
795 : : */
796 : : {
797 : : /*
798 : : * We need to re-read the timeline history in case it's been changed
799 : : * by a promotion or replay from a cascaded replica.
800 : : */
801 : 1526 : List *timelineHistory = readTimeLineHistory(currTLI);
802 : : XLogRecPtr endOfSegment;
803 : :
804 : 1526 : endOfSegment = ((wantPage / state->segcxt.ws_segsize) + 1) *
805 : 1526 : state->segcxt.ws_segsize - 1;
806 : : Assert(wantPage / state->segcxt.ws_segsize ==
807 : : endOfSegment / state->segcxt.ws_segsize);
808 : :
809 : : /*
810 : : * Find the timeline of the last LSN on the segment containing
811 : : * wantPage.
812 : : */
813 : 1526 : state->currTLI = tliOfPointInHistory(endOfSegment, timelineHistory);
814 : 1526 : state->currTLIValidUntil = tliSwitchPoint(state->currTLI, timelineHistory,
815 : : &state->nextTLI);
816 : :
817 : : Assert(!XLogRecPtrIsValid(state->currTLIValidUntil) ||
818 : : wantPage + wantLength < state->currTLIValidUntil);
819 : :
820 : 1526 : list_free_deep(timelineHistory);
821 : :
822 [ - + ]: 1526 : elog(DEBUG3, "switched to timeline %u valid until %X/%08X",
823 : : state->currTLI,
824 : : LSN_FORMAT_ARGS(state->currTLIValidUntil));
825 : : }
826 : : }
827 : :
828 : : /* XLogReaderRoutine->segment_open callback for local pg_wal files */
829 : : void
830 : 881 : wal_segment_open(XLogReaderState *state, XLogSegNo nextSegNo,
831 : : TimeLineID *tli_p)
832 : : {
833 : 881 : TimeLineID tli = *tli_p;
834 : : char path[MAXPGPATH];
835 : :
836 : 881 : XLogFilePath(path, tli, nextSegNo, state->segcxt.ws_segsize);
837 : 881 : state->seg.ws_file = BasicOpenFile(path, O_RDONLY | PG_BINARY);
838 [ + - ]: 881 : if (state->seg.ws_file >= 0)
839 : 881 : return;
840 : :
841 [ # # ]: 0 : if (errno == ENOENT)
842 [ # # ]: 0 : ereport(ERROR,
843 : : (errcode_for_file_access(),
844 : : errmsg("requested WAL segment %s has already been removed",
845 : : path)));
846 : : else
847 [ # # ]: 0 : ereport(ERROR,
848 : : (errcode_for_file_access(),
849 : : errmsg("could not open file \"%s\": %m",
850 : : path)));
851 : : }
852 : :
853 : : /* stock XLogReaderRoutine->segment_close callback */
854 : : void
855 : 5335 : wal_segment_close(XLogReaderState *state)
856 : : {
857 : 5335 : close(state->seg.ws_file);
858 : : /* need to check errno? */
859 : 5335 : state->seg.ws_file = -1;
860 : 5335 : }
861 : :
862 : : /*
863 : : * XLogReaderRoutine->page_read callback for reading local xlog files
864 : : *
865 : : * Public because it would likely be very helpful for someone writing another
866 : : * output method outside walsender, e.g. in a bgworker.
867 : : */
868 : : int
869 : 17868 : read_local_xlog_page(XLogReaderState *state, XLogRecPtr targetPagePtr,
870 : : int reqLen, XLogRecPtr targetRecPtr, char *cur_page)
871 : : {
872 : 17868 : return read_local_xlog_page_guts(state, targetPagePtr, reqLen,
873 : : targetRecPtr, cur_page, true);
874 : : }
875 : :
876 : : /*
877 : : * Same as read_local_xlog_page except that it doesn't wait for future WAL
878 : : * to be available.
879 : : */
880 : : int
881 : 3956 : read_local_xlog_page_no_wait(XLogReaderState *state, XLogRecPtr targetPagePtr,
882 : : int reqLen, XLogRecPtr targetRecPtr,
883 : : char *cur_page)
884 : : {
885 : 3956 : return read_local_xlog_page_guts(state, targetPagePtr, reqLen,
886 : : targetRecPtr, cur_page, false);
887 : : }
888 : :
889 : : /*
890 : : * Implementation of read_local_xlog_page and its no wait version.
891 : : */
892 : : static int
893 : 21824 : read_local_xlog_page_guts(XLogReaderState *state, XLogRecPtr targetPagePtr,
894 : : int reqLen, XLogRecPtr targetRecPtr,
895 : : char *cur_page, bool wait_for_wal)
896 : : {
897 : : XLogRecPtr read_upto,
898 : : loc;
899 : : TimeLineID tli;
900 : : int count;
901 : : WALReadError errinfo;
902 : : TimeLineID currTLI;
903 : :
904 : 21824 : loc = targetPagePtr + reqLen;
905 : :
906 : : /*
907 : : * Loop waiting for xlog to be available if necessary
908 : : *
909 : : * TODO: The walsender has its own version of this function, which uses a
910 : : * condition variable to wake up whenever WAL is flushed. We could use the
911 : : * same infrastructure here, instead of the check/sleep/repeat style of
912 : : * loop.
913 : : */
914 : : while (1)
915 : : {
916 : : /*
917 : : * Determine the limit of xlog we can currently read to, and what the
918 : : * most recent timeline is.
919 : : */
920 [ + + ]: 22678 : if (!RecoveryInProgress())
921 : 21870 : read_upto = GetFlushRecPtr(&currTLI);
922 : : else
923 : : {
924 : : TimeLineID insertTLI;
925 : :
926 : 808 : read_upto = GetXLogReplayRecPtr(&currTLI);
927 : :
928 : : /*
929 : : * If the insertion timeline has already been set, use it. See
930 : : * logical_read_xlog_page() for details.
931 : : */
932 : 808 : insertTLI = GetWALInsertionTimeLineIfSet();
933 [ + + ]: 808 : if (insertTLI != 0)
934 : 55 : currTLI = insertTLI;
935 : : }
936 : 22678 : tli = currTLI;
937 : :
938 : : /*
939 : : * Check which timeline to get the record from.
940 : : *
941 : : * We have to do it each time through the loop because if we're in
942 : : * recovery as a cascading standby, the current timeline might've
943 : : * become historical. We can't rely on RecoveryInProgress() because in
944 : : * a standby configuration like
945 : : *
946 : : * A => B => C
947 : : *
948 : : * if we're a logical decoding session on C, and B gets promoted, our
949 : : * timeline will change while we remain in recovery.
950 : : *
951 : : * We can't just keep reading from the old timeline as the last WAL
952 : : * archive in the timeline will get renamed to .partial by
953 : : * StartupXLOG().
954 : : *
955 : : * If that happens after our caller determined the TLI but before we
956 : : * actually read the xlog page, we might still try to read from the
957 : : * old (now renamed) segment and fail. There's not much we can do
958 : : * about this, but it can only happen when we're a leaf of a cascading
959 : : * standby whose primary gets promoted while we're decoding, so a
960 : : * one-off ERROR isn't too bad.
961 : : */
962 : 22678 : XLogReadDetermineTimeline(state, targetPagePtr, reqLen, tli);
963 : :
964 [ + + ]: 22678 : if (state->currTLI == currTLI)
965 : : {
966 : :
967 [ + + ]: 21519 : if (loc <= read_upto)
968 : 20623 : break;
969 : :
970 : : /* If asked, let's not wait for future WAL. */
971 [ + + ]: 896 : if (!wait_for_wal)
972 : : {
973 : : ReadLocalXLogPageNoWaitPrivate *private_data;
974 : :
975 : : /*
976 : : * Inform the caller of read_local_xlog_page_no_wait that the
977 : : * end of WAL has been reached.
978 : : */
979 : 42 : private_data = (ReadLocalXLogPageNoWaitPrivate *)
980 : : state->private_data;
981 : 42 : private_data->end_of_wal = true;
982 : 42 : break;
983 : : }
984 : :
985 [ - + ]: 854 : CHECK_FOR_INTERRUPTS();
986 : 854 : pg_usleep(1000L);
987 : : }
988 : : else
989 : : {
990 : : /*
991 : : * We're on a historical timeline, so limit reading to the switch
992 : : * point where we moved to the next timeline.
993 : : *
994 : : * We don't need to GetFlushRecPtr or GetXLogReplayRecPtr. We know
995 : : * about the new timeline, so we must've received past the end of
996 : : * it.
997 : : */
998 : 1159 : read_upto = state->currTLIValidUntil;
999 : :
1000 : : /*
1001 : : * Setting tli to our wanted record's TLI is slightly wrong; the
1002 : : * page might begin on an older timeline if it contains a timeline
1003 : : * switch, since its xlog segment will have been copied from the
1004 : : * prior timeline. This is pretty harmless though, as nothing
1005 : : * cares so long as the timeline doesn't go backwards. We should
1006 : : * read the page header instead; FIXME someday.
1007 : : */
1008 : 1159 : tli = state->currTLI;
1009 : :
1010 : : /* No need to wait on a historical timeline */
1011 : 1159 : break;
1012 : : }
1013 : : }
1014 : :
1015 [ + + ]: 21824 : if (targetPagePtr + XLOG_BLCKSZ <= read_upto)
1016 : : {
1017 : : /*
1018 : : * more than one block available; read only that block, have caller
1019 : : * come back if they need more.
1020 : : */
1021 : 21039 : count = XLOG_BLCKSZ;
1022 : : }
1023 [ + + ]: 785 : else if (targetPagePtr + reqLen > read_upto)
1024 : : {
1025 : : /* not enough data there */
1026 : 42 : return -1;
1027 : : }
1028 : : else
1029 : : {
1030 : : /* enough bytes available to satisfy the request */
1031 : 743 : count = read_upto - targetPagePtr;
1032 : : }
1033 : :
1034 [ - + ]: 21782 : if (!WALRead(state, cur_page, targetPagePtr, count, tli,
1035 : : &errinfo))
1036 : 0 : WALReadRaiseError(&errinfo);
1037 : :
1038 : : /* number of valid bytes in the buffer */
1039 : 21782 : return count;
1040 : : }
1041 : :
1042 : : /*
1043 : : * Backend-specific convenience code to handle read errors encountered by
1044 : : * WALRead().
1045 : : */
1046 : : void
1047 : 0 : WALReadRaiseError(WALReadError *errinfo)
1048 : : {
1049 : 0 : WALOpenSegment *seg = &errinfo->wre_seg;
1050 : : char fname[MAXFNAMELEN];
1051 : :
1052 : 0 : XLogFileName(fname, seg->ws_tli, seg->ws_segno, wal_segment_size);
1053 : :
1054 [ # # ]: 0 : if (errinfo->wre_read < 0)
1055 : : {
1056 : 0 : errno = errinfo->wre_errno;
1057 [ # # ]: 0 : ereport(ERROR,
1058 : : (errcode_for_file_access(),
1059 : : errmsg("could not read from WAL segment %s, offset %d: %m",
1060 : : fname, errinfo->wre_off)));
1061 : : }
1062 [ # # ]: 0 : else if (errinfo->wre_read == 0)
1063 : : {
1064 [ # # ]: 0 : ereport(ERROR,
1065 : : (errcode(ERRCODE_DATA_CORRUPTED),
1066 : : errmsg("could not read from WAL segment %s, offset %d: read %d of %d",
1067 : : fname, errinfo->wre_off, errinfo->wre_read,
1068 : : errinfo->wre_req)));
1069 : : }
1070 : 0 : }
|
