Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * basebackup_incremental.c
4 : * code for incremental backup support
5 : *
6 : * This code isn't actually in charge of taking an incremental backup;
7 : * the actual construction of the incremental backup happens in
8 : * basebackup.c. Here, we're concerned with providing the necessary
9 : * supports for that operation. In particular, we need to parse the
10 : * backup manifest supplied by the user taking the incremental backup
11 : * and extract the required information from it.
12 : *
13 : * Portions Copyright (c) 2010-2025, PostgreSQL Global Development Group
14 : *
15 : * IDENTIFICATION
16 : * src/backend/backup/basebackup_incremental.c
17 : *
18 : *-------------------------------------------------------------------------
19 : */
20 : #include "postgres.h"
21 :
22 : #include "access/timeline.h"
23 : #include "access/xlog.h"
24 : #include "backup/basebackup_incremental.h"
25 : #include "backup/walsummary.h"
26 : #include "common/blkreftable.h"
27 : #include "common/hashfn.h"
28 : #include "common/int.h"
29 : #include "common/parse_manifest.h"
30 : #include "postmaster/walsummarizer.h"
31 :
32 : #define BLOCKS_PER_READ 512
33 :
34 : /*
35 : * We expect to find the last lines of the manifest, including the checksum,
36 : * in the last MIN_CHUNK bytes of the manifest. We trigger an incremental
37 : * parse step if we are about to overflow MAX_CHUNK bytes.
38 : */
39 : #define MIN_CHUNK 1024
40 : #define MAX_CHUNK (128 * 1024)
41 :
42 : /*
43 : * Details extracted from the WAL ranges present in the supplied backup manifest.
44 : */
45 : typedef struct
46 : {
47 : TimeLineID tli;
48 : XLogRecPtr start_lsn;
49 : XLogRecPtr end_lsn;
50 : } backup_wal_range;
51 :
52 : /*
53 : * Details extracted from the file list present in the supplied backup manifest.
54 : */
55 : typedef struct
56 : {
57 : uint32 status;
58 : const char *path;
59 : uint64 size;
60 : } backup_file_entry;
61 :
62 : static uint32 hash_string_pointer(const char *s);
63 : #define SH_PREFIX backup_file
64 : #define SH_ELEMENT_TYPE backup_file_entry
65 : #define SH_KEY_TYPE const char *
66 : #define SH_KEY path
67 : #define SH_HASH_KEY(tb, key) hash_string_pointer(key)
68 : #define SH_EQUAL(tb, a, b) (strcmp(a, b) == 0)
69 : #define SH_SCOPE static inline
70 : #define SH_DECLARE
71 : #define SH_DEFINE
72 : #include "lib/simplehash.h"
73 :
74 : struct IncrementalBackupInfo
75 : {
76 : /* Memory context for this object and its subsidiary objects. */
77 : MemoryContext mcxt;
78 :
79 : /* Temporary buffer for storing the manifest while parsing it. */
80 : StringInfoData buf;
81 :
82 : /* WAL ranges extracted from the backup manifest. */
83 : List *manifest_wal_ranges;
84 :
85 : /*
86 : * Files extracted from the backup manifest.
87 : *
88 : * We don't really need this information, because we use WAL summaries to
89 : * figure out what's changed. It would be unsafe to just rely on the list
90 : * of files that existed before, because it's possible for a file to be
91 : * removed and a new one created with the same name and different
92 : * contents. In such cases, the whole file must still be sent. We can tell
93 : * from the WAL summaries whether that happened, but not from the file
94 : * list.
95 : *
96 : * Nonetheless, this data is useful for sanity checking. If a file that we
97 : * think we shouldn't need to send is not present in the manifest for the
98 : * prior backup, something has gone terribly wrong. We retain the file
99 : * names and sizes, but not the checksums or last modified times, for
100 : * which we have no use.
101 : *
102 : * One significant downside of storing this data is that it consumes
103 : * memory. If that turns out to be a problem, we might have to decide not
104 : * to retain this information, or to make it optional.
105 : */
106 : backup_file_hash *manifest_files;
107 :
108 : /*
109 : * Block-reference table for the incremental backup.
110 : *
111 : * It's possible that storing the entire block-reference table in memory
112 : * will be a problem for some users. The in-memory format that we're using
113 : * here is pretty efficient, converging to little more than 1 bit per
114 : * block for relation forks with large numbers of modified blocks. It's
115 : * possible, however, that if you try to perform an incremental backup of
116 : * a database with a sufficiently large number of relations on a
117 : * sufficiently small machine, you could run out of memory here. If that
118 : * turns out to be a problem in practice, we'll need to be more clever.
119 : */
120 : BlockRefTable *brtab;
121 :
122 : /*
123 : * State object for incremental JSON parsing
124 : */
125 : JsonManifestParseIncrementalState *inc_state;
126 : };
127 :
128 : static void manifest_process_version(JsonManifestParseContext *context,
129 : int manifest_version);
130 : static void manifest_process_system_identifier(JsonManifestParseContext *context,
131 : uint64 manifest_system_identifier);
132 : static void manifest_process_file(JsonManifestParseContext *context,
133 : const char *pathname,
134 : uint64 size,
135 : pg_checksum_type checksum_type,
136 : int checksum_length,
137 : uint8 *checksum_payload);
138 : static void manifest_process_wal_range(JsonManifestParseContext *context,
139 : TimeLineID tli,
140 : XLogRecPtr start_lsn,
141 : XLogRecPtr end_lsn);
142 : pg_noreturn static void manifest_report_error(JsonManifestParseContext *context,
143 : const char *fmt,...)
144 : pg_attribute_printf(2, 3);
145 : static int compare_block_numbers(const void *a, const void *b);
146 :
147 : /*
148 : * Create a new object for storing information extracted from the manifest
149 : * supplied when creating an incremental backup.
150 : */
151 : IncrementalBackupInfo *
152 22 : CreateIncrementalBackupInfo(MemoryContext mcxt)
153 : {
154 : IncrementalBackupInfo *ib;
155 : MemoryContext oldcontext;
156 : JsonManifestParseContext *context;
157 :
158 22 : oldcontext = MemoryContextSwitchTo(mcxt);
159 :
160 22 : ib = palloc0(sizeof(IncrementalBackupInfo));
161 22 : ib->mcxt = mcxt;
162 22 : initStringInfo(&ib->buf);
163 :
164 : /*
165 : * It's hard to guess how many files a "typical" installation will have in
166 : * the data directory, but a fresh initdb creates almost 1000 files as of
167 : * this writing, so it seems to make sense for our estimate to
168 : * substantially higher.
169 : */
170 22 : ib->manifest_files = backup_file_create(mcxt, 10000, NULL);
171 :
172 22 : context = palloc0(sizeof(JsonManifestParseContext));
173 : /* Parse the manifest. */
174 22 : context->private_data = ib;
175 22 : context->version_cb = manifest_process_version;
176 22 : context->system_identifier_cb = manifest_process_system_identifier;
177 22 : context->per_file_cb = manifest_process_file;
178 22 : context->per_wal_range_cb = manifest_process_wal_range;
179 22 : context->error_cb = manifest_report_error;
180 :
181 22 : ib->inc_state = json_parse_manifest_incremental_init(context);
182 :
183 22 : MemoryContextSwitchTo(oldcontext);
184 :
185 22 : return ib;
186 : }
187 :
188 : /*
189 : * Before taking an incremental backup, the caller must supply the backup
190 : * manifest from a prior backup. Each chunk of manifest data received
191 : * from the client should be passed to this function.
192 : */
193 : void
194 66 : AppendIncrementalManifestData(IncrementalBackupInfo *ib, const char *data,
195 : int len)
196 : {
197 : MemoryContext oldcontext;
198 :
199 : /* Switch to our memory context. */
200 66 : oldcontext = MemoryContextSwitchTo(ib->mcxt);
201 :
202 66 : if (ib->buf.len > MIN_CHUNK && ib->buf.len + len > MAX_CHUNK)
203 : {
204 : /*
205 : * time for an incremental parse. We'll do all but the last MIN_CHUNK
206 : * so that we have enough left for the final piece.
207 : */
208 22 : json_parse_manifest_incremental_chunk(ib->inc_state, ib->buf.data,
209 22 : ib->buf.len - MIN_CHUNK, false);
210 : /* now remove what we just parsed */
211 20 : memmove(ib->buf.data, ib->buf.data + (ib->buf.len - MIN_CHUNK),
212 : MIN_CHUNK + 1);
213 20 : ib->buf.len = MIN_CHUNK;
214 : }
215 :
216 64 : appendBinaryStringInfo(&ib->buf, data, len);
217 :
218 : /* Switch back to previous memory context. */
219 64 : MemoryContextSwitchTo(oldcontext);
220 64 : }
221 :
222 : /*
223 : * Finalize an IncrementalBackupInfo object after all manifest data has
224 : * been supplied via calls to AppendIncrementalManifestData.
225 : */
226 : void
227 20 : FinalizeIncrementalManifest(IncrementalBackupInfo *ib)
228 : {
229 : MemoryContext oldcontext;
230 :
231 : /* Switch to our memory context. */
232 20 : oldcontext = MemoryContextSwitchTo(ib->mcxt);
233 :
234 : /* Parse the last chunk of the manifest */
235 20 : json_parse_manifest_incremental_chunk(ib->inc_state, ib->buf.data,
236 20 : ib->buf.len, true);
237 :
238 : /* Done with the buffer, so release memory. */
239 20 : pfree(ib->buf.data);
240 20 : ib->buf.data = NULL;
241 :
242 : /* Done with inc_state, so release that memory too */
243 20 : json_parse_manifest_incremental_shutdown(ib->inc_state);
244 :
245 : /* Switch back to previous memory context. */
246 20 : MemoryContextSwitchTo(oldcontext);
247 20 : }
248 :
249 : /*
250 : * Prepare to take an incremental backup.
251 : *
252 : * Before this function is called, AppendIncrementalManifestData and
253 : * FinalizeIncrementalManifest should have already been called to pass all
254 : * the manifest data to this object.
255 : *
256 : * This function performs sanity checks on the data extracted from the
257 : * manifest and figures out for which WAL ranges we need summaries, and
258 : * whether those summaries are available. Then, it reads and combines the
259 : * data from those summary files. It also updates the backup_state with the
260 : * reference TLI and LSN for the prior backup.
261 : */
262 : void
263 20 : PrepareForIncrementalBackup(IncrementalBackupInfo *ib,
264 : BackupState *backup_state)
265 : {
266 : MemoryContext oldcontext;
267 : List *expectedTLEs;
268 : List *all_wslist,
269 20 : *required_wslist = NIL;
270 : ListCell *lc;
271 : TimeLineHistoryEntry **tlep;
272 : int num_wal_ranges;
273 : int i;
274 20 : bool found_backup_start_tli = false;
275 20 : TimeLineID earliest_wal_range_tli = 0;
276 20 : XLogRecPtr earliest_wal_range_start_lsn = InvalidXLogRecPtr;
277 20 : TimeLineID latest_wal_range_tli = 0;
278 :
279 : Assert(ib->buf.data == NULL);
280 :
281 : /* Switch to our memory context. */
282 20 : oldcontext = MemoryContextSwitchTo(ib->mcxt);
283 :
284 : /*
285 : * A valid backup manifest must always contain at least one WAL range
286 : * (usually exactly one, unless the backup spanned a timeline switch).
287 : */
288 20 : num_wal_ranges = list_length(ib->manifest_wal_ranges);
289 20 : if (num_wal_ranges == 0)
290 0 : ereport(ERROR,
291 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
292 : errmsg("manifest contains no required WAL ranges")));
293 :
294 : /*
295 : * Match up the TLIs that appear in the WAL ranges of the backup manifest
296 : * with those that appear in this server's timeline history. We expect
297 : * every backup_wal_range to match to a TimeLineHistoryEntry; if it does
298 : * not, that's an error.
299 : *
300 : * This loop also decides which of the WAL ranges is the manifest is most
301 : * ancient and which one is the newest, according to the timeline history
302 : * of this server, and stores TLIs of those WAL ranges into
303 : * earliest_wal_range_tli and latest_wal_range_tli. It also updates
304 : * earliest_wal_range_start_lsn to the start LSN of the WAL range for
305 : * earliest_wal_range_tli.
306 : *
307 : * Note that the return value of readTimeLineHistory puts the latest
308 : * timeline at the beginning of the list, not the end. Hence, the earliest
309 : * TLI is the one that occurs nearest the end of the list returned by
310 : * readTimeLineHistory, and the latest TLI is the one that occurs closest
311 : * to the beginning.
312 : */
313 20 : expectedTLEs = readTimeLineHistory(backup_state->starttli);
314 20 : tlep = palloc0(num_wal_ranges * sizeof(TimeLineHistoryEntry *));
315 40 : for (i = 0; i < num_wal_ranges; ++i)
316 : {
317 20 : backup_wal_range *range = list_nth(ib->manifest_wal_ranges, i);
318 20 : bool saw_earliest_wal_range_tli = false;
319 20 : bool saw_latest_wal_range_tli = false;
320 :
321 : /* Search this server's history for this WAL range's TLI. */
322 22 : foreach(lc, expectedTLEs)
323 : {
324 22 : TimeLineHistoryEntry *tle = lfirst(lc);
325 :
326 22 : if (tle->tli == range->tli)
327 : {
328 20 : tlep[i] = tle;
329 20 : break;
330 : }
331 :
332 2 : if (tle->tli == earliest_wal_range_tli)
333 0 : saw_earliest_wal_range_tli = true;
334 2 : if (tle->tli == latest_wal_range_tli)
335 0 : saw_latest_wal_range_tli = true;
336 : }
337 :
338 : /*
339 : * An incremental backup can only be taken relative to a backup that
340 : * represents a previous state of this server. If the backup requires
341 : * WAL from a timeline that's not in our history, that definitely
342 : * isn't the case.
343 : */
344 20 : if (tlep[i] == NULL)
345 0 : ereport(ERROR,
346 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
347 : errmsg("timeline %u found in manifest, but not in this server's history",
348 : range->tli)));
349 :
350 : /*
351 : * If we found this TLI in the server's history before encountering
352 : * the latest TLI seen so far in the server's history, then this TLI
353 : * is the latest one seen so far.
354 : *
355 : * If on the other hand we saw the earliest TLI seen so far before
356 : * finding this TLI, this TLI is earlier than the earliest one seen so
357 : * far. And if this is the first TLI for which we've searched, it's
358 : * also the earliest one seen so far.
359 : *
360 : * On the first loop iteration, both things should necessarily be
361 : * true.
362 : */
363 20 : if (!saw_latest_wal_range_tli)
364 20 : latest_wal_range_tli = range->tli;
365 20 : if (earliest_wal_range_tli == 0 || saw_earliest_wal_range_tli)
366 : {
367 20 : earliest_wal_range_tli = range->tli;
368 20 : earliest_wal_range_start_lsn = range->start_lsn;
369 : }
370 : }
371 :
372 : /*
373 : * Propagate information about the prior backup into the backup_label that
374 : * will be generated for this backup.
375 : */
376 20 : backup_state->istartpoint = earliest_wal_range_start_lsn;
377 20 : backup_state->istarttli = earliest_wal_range_tli;
378 :
379 : /*
380 : * Sanity check start and end LSNs for the WAL ranges in the manifest.
381 : *
382 : * Commonly, there won't be any timeline switches during the prior backup
383 : * at all, but if there are, they should happen at the same LSNs that this
384 : * server switched timelines.
385 : *
386 : * Whether there are any timeline switches during the prior backup or not,
387 : * the prior backup shouldn't require any WAL from a timeline prior to the
388 : * start of that timeline. It also shouldn't require any WAL from later
389 : * than the start of this backup.
390 : *
391 : * If any of these sanity checks fail, one possible explanation is that
392 : * the user has generated WAL on the same timeline with the same LSNs more
393 : * than once. For instance, if two standbys running on timeline 1 were
394 : * both promoted and (due to a broken archiving setup) both selected new
395 : * timeline ID 2, then it's possible that one of these checks might trip.
396 : *
397 : * Note that there are lots of ways for the user to do something very bad
398 : * without tripping any of these checks, and they are not intended to be
399 : * comprehensive. It's pretty hard to see how we could be certain of
400 : * anything here. However, if there's a problem staring us right in the
401 : * face, it's best to report it, so we do.
402 : */
403 40 : for (i = 0; i < num_wal_ranges; ++i)
404 : {
405 20 : backup_wal_range *range = list_nth(ib->manifest_wal_ranges, i);
406 :
407 20 : if (range->tli == earliest_wal_range_tli)
408 : {
409 20 : if (range->start_lsn < tlep[i]->begin)
410 0 : ereport(ERROR,
411 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
412 : errmsg("manifest requires WAL from initial timeline %u starting at %X/%X, but that timeline begins at %X/%X",
413 : range->tli,
414 : LSN_FORMAT_ARGS(range->start_lsn),
415 : LSN_FORMAT_ARGS(tlep[i]->begin))));
416 : }
417 : else
418 : {
419 0 : if (range->start_lsn != tlep[i]->begin)
420 0 : ereport(ERROR,
421 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
422 : errmsg("manifest requires WAL from continuation timeline %u starting at %X/%X, but that timeline begins at %X/%X",
423 : range->tli,
424 : LSN_FORMAT_ARGS(range->start_lsn),
425 : LSN_FORMAT_ARGS(tlep[i]->begin))));
426 : }
427 :
428 20 : if (range->tli == latest_wal_range_tli)
429 : {
430 20 : if (range->end_lsn > backup_state->startpoint)
431 0 : ereport(ERROR,
432 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
433 : errmsg("manifest requires WAL from final timeline %u ending at %X/%X, but this backup starts at %X/%X",
434 : range->tli,
435 : LSN_FORMAT_ARGS(range->end_lsn),
436 : LSN_FORMAT_ARGS(backup_state->startpoint)),
437 : errhint("This can happen for incremental backups on a standby if there was little activity since the previous backup.")));
438 : }
439 : else
440 : {
441 0 : if (range->end_lsn != tlep[i]->end)
442 0 : ereport(ERROR,
443 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
444 : errmsg("manifest requires WAL from non-final timeline %u ending at %X/%X, but this server switched timelines at %X/%X",
445 : range->tli,
446 : LSN_FORMAT_ARGS(range->end_lsn),
447 : LSN_FORMAT_ARGS(tlep[i]->end))));
448 : }
449 :
450 : }
451 :
452 : /*
453 : * Wait for WAL summarization to catch up to the backup start LSN. This
454 : * will throw an error if the WAL summarizer appears to be stuck. If WAL
455 : * summarization gets disabled while we're waiting, this will return
456 : * immediately, and we'll error out further down if the WAL summaries are
457 : * incomplete.
458 : */
459 20 : WaitForWalSummarization(backup_state->startpoint);
460 :
461 : /*
462 : * Retrieve a list of all WAL summaries on any timeline that overlap with
463 : * the LSN range of interest. We could instead call GetWalSummaries() once
464 : * per timeline in the loop that follows, but that would involve reading
465 : * the directory multiple times. It should be mildly faster - and perhaps
466 : * a bit safer - to do it just once.
467 : */
468 20 : all_wslist = GetWalSummaries(0, earliest_wal_range_start_lsn,
469 : backup_state->startpoint);
470 :
471 : /*
472 : * We need WAL summaries for everything that happened during the prior
473 : * backup and everything that happened afterward up until the point where
474 : * the current backup started.
475 : */
476 22 : foreach(lc, expectedTLEs)
477 : {
478 22 : TimeLineHistoryEntry *tle = lfirst(lc);
479 22 : XLogRecPtr tli_start_lsn = tle->begin;
480 22 : XLogRecPtr tli_end_lsn = tle->end;
481 22 : XLogRecPtr tli_missing_lsn = InvalidXLogRecPtr;
482 : List *tli_wslist;
483 :
484 : /*
485 : * Working through the history of this server from the current
486 : * timeline backwards, we skip everything until we find the timeline
487 : * where this backup started. Most of the time, this means we won't
488 : * skip anything at all, as it's unlikely that the timeline has
489 : * changed since the beginning of the backup moments ago.
490 : */
491 22 : if (tle->tli == backup_state->starttli)
492 : {
493 20 : found_backup_start_tli = true;
494 20 : tli_end_lsn = backup_state->startpoint;
495 : }
496 2 : else if (!found_backup_start_tli)
497 0 : continue;
498 :
499 : /*
500 : * Find the summaries that overlap the LSN range of interest for this
501 : * timeline. If this is the earliest timeline involved, the range of
502 : * interest begins with the start LSN of the prior backup; otherwise,
503 : * it begins at the LSN at which this timeline came into existence. If
504 : * this is the latest TLI involved, the range of interest ends at the
505 : * start LSN of the current backup; otherwise, it ends at the point
506 : * where we switched from this timeline to the next one.
507 : */
508 22 : if (tle->tli == earliest_wal_range_tli)
509 20 : tli_start_lsn = earliest_wal_range_start_lsn;
510 22 : tli_wslist = FilterWalSummaries(all_wslist, tle->tli,
511 : tli_start_lsn, tli_end_lsn);
512 :
513 : /*
514 : * There is no guarantee that the WAL summaries we found cover the
515 : * entire range of LSNs for which summaries are required, or indeed
516 : * that we found any WAL summaries at all. Check whether we have a
517 : * problem of that sort.
518 : */
519 22 : if (!WalSummariesAreComplete(tli_wslist, tli_start_lsn, tli_end_lsn,
520 : &tli_missing_lsn))
521 : {
522 0 : if (XLogRecPtrIsInvalid(tli_missing_lsn))
523 0 : ereport(ERROR,
524 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
525 : errmsg("WAL summaries are required on timeline %u from %X/%X to %X/%X, but no summaries for that timeline and LSN range exist",
526 : tle->tli,
527 : LSN_FORMAT_ARGS(tli_start_lsn),
528 : LSN_FORMAT_ARGS(tli_end_lsn))));
529 : else
530 0 : ereport(ERROR,
531 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
532 : errmsg("WAL summaries are required on timeline %u from %X/%X to %X/%X, but the summaries for that timeline and LSN range are incomplete",
533 : tle->tli,
534 : LSN_FORMAT_ARGS(tli_start_lsn),
535 : LSN_FORMAT_ARGS(tli_end_lsn)),
536 : errdetail("The first unsummarized LSN in this range is %X/%X.",
537 : LSN_FORMAT_ARGS(tli_missing_lsn))));
538 : }
539 :
540 : /*
541 : * Remember that we need to read these summaries.
542 : *
543 : * Technically, it's possible that this could read more files than
544 : * required, since tli_wslist in theory could contain redundant
545 : * summaries. For instance, if we have a summary from 0/10000000 to
546 : * 0/20000000 and also one from 0/00000000 to 0/30000000, then the
547 : * latter subsumes the former and the former could be ignored.
548 : *
549 : * We ignore this possibility because the WAL summarizer only tries to
550 : * generate summaries that do not overlap. If somehow they exist,
551 : * we'll do a bit of extra work but the results should still be
552 : * correct.
553 : */
554 22 : required_wslist = list_concat(required_wslist, tli_wslist);
555 :
556 : /*
557 : * Timelines earlier than the one in which the prior backup began are
558 : * not relevant.
559 : */
560 22 : if (tle->tli == earliest_wal_range_tli)
561 20 : break;
562 : }
563 :
564 : /*
565 : * Read all of the required block reference table files and merge all of
566 : * the data into a single in-memory block reference table.
567 : *
568 : * See the comments for struct IncrementalBackupInfo for some thoughts on
569 : * memory usage.
570 : */
571 20 : ib->brtab = CreateEmptyBlockRefTable();
572 56 : foreach(lc, required_wslist)
573 : {
574 36 : WalSummaryFile *ws = lfirst(lc);
575 : WalSummaryIO wsio;
576 : BlockRefTableReader *reader;
577 : RelFileLocator rlocator;
578 : ForkNumber forknum;
579 : BlockNumber limit_block;
580 : BlockNumber blocks[BLOCKS_PER_READ];
581 :
582 36 : wsio.file = OpenWalSummaryFile(ws, false);
583 36 : wsio.filepos = 0;
584 36 : ereport(DEBUG1,
585 : (errmsg_internal("reading WAL summary file \"%s\"",
586 : FilePathName(wsio.file))));
587 36 : reader = CreateBlockRefTableReader(ReadWalSummary, &wsio,
588 : FilePathName(wsio.file),
589 : ReportWalSummaryError, NULL);
590 738 : while (BlockRefTableReaderNextRelation(reader, &rlocator, &forknum,
591 : &limit_block))
592 : {
593 702 : BlockRefTableSetLimitBlock(ib->brtab, &rlocator,
594 : forknum, limit_block);
595 :
596 : while (1)
597 508 : {
598 : unsigned nblocks;
599 : unsigned i;
600 :
601 1210 : nblocks = BlockRefTableReaderGetBlocks(reader, blocks,
602 : BLOCKS_PER_READ);
603 1210 : if (nblocks == 0)
604 702 : break;
605 :
606 2342 : for (i = 0; i < nblocks; ++i)
607 1834 : BlockRefTableMarkBlockModified(ib->brtab, &rlocator,
608 : forknum, blocks[i]);
609 : }
610 : }
611 36 : DestroyBlockRefTableReader(reader);
612 36 : FileClose(wsio.file);
613 : }
614 :
615 : /* Switch back to previous memory context. */
616 20 : MemoryContextSwitchTo(oldcontext);
617 20 : }
618 :
619 : /*
620 : * Get the pathname that should be used when a file is sent incrementally.
621 : *
622 : * The result is a palloc'd string.
623 : */
624 : char *
625 3214 : GetIncrementalFilePath(Oid dboid, Oid spcoid, RelFileNumber relfilenumber,
626 : ForkNumber forknum, unsigned segno)
627 : {
628 : RelPathStr path;
629 : char *lastslash;
630 : char *ipath;
631 :
632 3214 : path = GetRelationPath(dboid, spcoid, relfilenumber, INVALID_PROC_NUMBER,
633 : forknum);
634 :
635 3214 : lastslash = strrchr(path.str, '/');
636 : Assert(lastslash != NULL);
637 3214 : *lastslash = '\0';
638 :
639 3214 : if (segno > 0)
640 0 : ipath = psprintf("%s/INCREMENTAL.%s.%u", path.str, lastslash + 1, segno);
641 : else
642 3214 : ipath = psprintf("%s/INCREMENTAL.%s", path.str, lastslash + 1);
643 :
644 3214 : return ipath;
645 : }
646 :
647 : /*
648 : * How should we back up a particular file as part of an incremental backup?
649 : *
650 : * If the return value is BACK_UP_FILE_FULLY, caller should back up the whole
651 : * file just as if this were not an incremental backup. The contents of the
652 : * relative_block_numbers array are unspecified in this case.
653 : *
654 : * If the return value is BACK_UP_FILE_INCREMENTALLY, caller should include
655 : * an incremental file in the backup instead of the entire file. On return,
656 : * *num_blocks_required will be set to the number of blocks that need to be
657 : * sent, and the actual block numbers will have been stored in
658 : * relative_block_numbers, which should be an array of at least RELSEG_SIZE.
659 : * In addition, *truncation_block_length will be set to the value that should
660 : * be included in the incremental file.
661 : */
662 : FileBackupMethod
663 20324 : GetFileBackupMethod(IncrementalBackupInfo *ib, const char *path,
664 : Oid dboid, Oid spcoid,
665 : RelFileNumber relfilenumber, ForkNumber forknum,
666 : unsigned segno, size_t size,
667 : unsigned *num_blocks_required,
668 : BlockNumber *relative_block_numbers,
669 : unsigned *truncation_block_length)
670 : {
671 : BlockNumber limit_block;
672 : BlockNumber start_blkno;
673 : BlockNumber stop_blkno;
674 : RelFileLocator rlocator;
675 : BlockRefTableEntry *brtentry;
676 : unsigned i;
677 : unsigned nblocks;
678 :
679 : /* Should only be called after PrepareForIncrementalBackup. */
680 : Assert(ib->buf.data == NULL);
681 :
682 : /*
683 : * dboid could be InvalidOid if shared rel, but spcoid and relfilenumber
684 : * should have legal values.
685 : */
686 : Assert(OidIsValid(spcoid));
687 : Assert(RelFileNumberIsValid(relfilenumber));
688 :
689 : /*
690 : * If the file size is too large or not a multiple of BLCKSZ, then
691 : * something weird is happening, so give up and send the whole file.
692 : */
693 20324 : if ((size % BLCKSZ) != 0 || size / BLCKSZ > RELSEG_SIZE)
694 0 : return BACK_UP_FILE_FULLY;
695 :
696 : /*
697 : * The free-space map fork is not properly WAL-logged, so we need to
698 : * backup the entire file every time.
699 : */
700 20324 : if (forknum == FSM_FORKNUM)
701 2474 : return BACK_UP_FILE_FULLY;
702 :
703 : /*
704 : * If this file was not part of the prior backup, back it up fully.
705 : *
706 : * If this file was created after the prior backup and before the start of
707 : * the current backup, then the WAL summary information will tell us to
708 : * back up the whole file. However, if this file was created after the
709 : * start of the current backup, then the WAL summary won't know anything
710 : * about it. Without this logic, we would erroneously conclude that it was
711 : * OK to send it incrementally.
712 : *
713 : * Note that the file could have existed at the time of the prior backup,
714 : * gotten deleted, and then a new file with the same name could have been
715 : * created. In that case, this logic won't prevent the file from being
716 : * backed up incrementally. But, if the deletion happened before the start
717 : * of the current backup, the limit block will be 0, inducing a full
718 : * backup. If the deletion happened after the start of the current backup,
719 : * reconstruction will erroneously combine blocks from the current
720 : * lifespan of the file with blocks from the previous lifespan -- but in
721 : * this type of case, WAL replay to reach backup consistency should remove
722 : * and recreate the file anyway, so the initial bogus contents should not
723 : * matter.
724 : */
725 17850 : if (backup_file_lookup(ib->manifest_files, path) == NULL)
726 : {
727 : char *ipath;
728 :
729 3214 : ipath = GetIncrementalFilePath(dboid, spcoid, relfilenumber,
730 : forknum, segno);
731 3214 : if (backup_file_lookup(ib->manifest_files, ipath) == NULL)
732 548 : return BACK_UP_FILE_FULLY;
733 : }
734 :
735 : /*
736 : * Look up the special block reference table entry for the database as a
737 : * whole.
738 : */
739 17302 : rlocator.spcOid = spcoid;
740 17302 : rlocator.dbOid = dboid;
741 17302 : rlocator.relNumber = 0;
742 17302 : if (BlockRefTableGetEntry(ib->brtab, &rlocator, MAIN_FORKNUM,
743 : &limit_block) != NULL)
744 : {
745 : /*
746 : * According to the WAL summary, this database OID/tablespace OID
747 : * pairing has been created since the previous backup. So, everything
748 : * in it must be backed up fully.
749 : */
750 522 : return BACK_UP_FILE_FULLY;
751 : }
752 :
753 : /* Look up the block reference table entry for this relfilenode. */
754 16780 : rlocator.relNumber = relfilenumber;
755 16780 : brtentry = BlockRefTableGetEntry(ib->brtab, &rlocator, forknum,
756 : &limit_block);
757 :
758 : /*
759 : * If there is no entry, then there have been no WAL-logged changes to the
760 : * relation since the predecessor backup was taken, so we can back it up
761 : * incrementally and need not include any modified blocks.
762 : *
763 : * However, if the file is zero-length, we should do a full backup,
764 : * because an incremental file is always more than zero length, and it's
765 : * silly to take an incremental backup when a full backup would be
766 : * smaller.
767 : */
768 16780 : if (brtentry == NULL)
769 : {
770 16712 : if (size == 0)
771 3420 : return BACK_UP_FILE_FULLY;
772 13292 : *num_blocks_required = 0;
773 13292 : *truncation_block_length = size / BLCKSZ;
774 13292 : return BACK_UP_FILE_INCREMENTALLY;
775 : }
776 :
777 : /*
778 : * If the limit_block is less than or equal to the point where this
779 : * segment starts, send the whole file.
780 : */
781 68 : if (limit_block <= segno * RELSEG_SIZE)
782 0 : return BACK_UP_FILE_FULLY;
783 :
784 : /*
785 : * Get relevant entries from the block reference table entry.
786 : *
787 : * We shouldn't overflow computing the start or stop block numbers, but if
788 : * it manages to happen somehow, detect it and throw an error.
789 : */
790 68 : start_blkno = segno * RELSEG_SIZE;
791 68 : stop_blkno = start_blkno + (size / BLCKSZ);
792 68 : if (start_blkno / RELSEG_SIZE != segno || stop_blkno < start_blkno)
793 0 : ereport(ERROR,
794 : errcode(ERRCODE_INTERNAL_ERROR),
795 : errmsg_internal("overflow computing block number bounds for segment %u with size %zu",
796 : segno, size));
797 :
798 : /*
799 : * This will write *absolute* block numbers into the output array, but
800 : * we'll transpose them below.
801 : */
802 68 : nblocks = BlockRefTableEntryGetBlocks(brtentry, start_blkno, stop_blkno,
803 : relative_block_numbers, RELSEG_SIZE);
804 : Assert(nblocks <= RELSEG_SIZE);
805 :
806 : /*
807 : * If we're going to have to send nearly all of the blocks, then just send
808 : * the whole file, because that won't require much extra storage or
809 : * transfer and will speed up and simplify backup restoration. It's not
810 : * clear what threshold is most appropriate here and perhaps it ought to
811 : * be configurable, but for now we're just going to say that if we'd need
812 : * to send 90% of the blocks anyway, give up and send the whole file.
813 : *
814 : * NB: If you change the threshold here, at least make sure to back up the
815 : * file fully when every single block must be sent, because there's
816 : * nothing good about sending an incremental file in that case.
817 : */
818 68 : if (nblocks * BLCKSZ > size * 0.9)
819 18 : return BACK_UP_FILE_FULLY;
820 :
821 : /*
822 : * Looks like we can send an incremental file, so sort the block numbers
823 : * and then transpose them from absolute block numbers to relative block
824 : * numbers if necessary.
825 : *
826 : * NB: If the block reference table was using the bitmap representation
827 : * for a given chunk, the block numbers in that chunk will already be
828 : * sorted, but when the array-of-offsets representation is used, we can
829 : * receive block numbers here out of order.
830 : */
831 50 : qsort(relative_block_numbers, nblocks, sizeof(BlockNumber),
832 : compare_block_numbers);
833 50 : if (start_blkno != 0)
834 : {
835 0 : for (i = 0; i < nblocks; ++i)
836 0 : relative_block_numbers[i] -= start_blkno;
837 : }
838 50 : *num_blocks_required = nblocks;
839 :
840 : /*
841 : * The truncation block length is the minimum length of the reconstructed
842 : * file. Any block numbers below this threshold that are not present in
843 : * the backup need to be fetched from the prior backup. At or above this
844 : * threshold, blocks should only be included in the result if they are
845 : * present in the backup. (This may require inserting zero blocks if the
846 : * blocks included in the backup are non-consecutive.)
847 : */
848 50 : *truncation_block_length = size / BLCKSZ;
849 50 : if (BlockNumberIsValid(limit_block))
850 : {
851 0 : unsigned relative_limit = limit_block - segno * RELSEG_SIZE;
852 :
853 0 : if (*truncation_block_length < relative_limit)
854 0 : *truncation_block_length = relative_limit;
855 : }
856 :
857 : /* Send it incrementally. */
858 50 : return BACK_UP_FILE_INCREMENTALLY;
859 : }
860 :
861 : /*
862 : * Compute the size for a header of an incremental file containing a given
863 : * number of blocks. The header is rounded to a multiple of BLCKSZ, but
864 : * only if the file will store some block data.
865 : */
866 : size_t
867 13342 : GetIncrementalHeaderSize(unsigned num_blocks_required)
868 : {
869 : size_t result;
870 :
871 : /* Make sure we're not going to overflow. */
872 : Assert(num_blocks_required <= RELSEG_SIZE);
873 :
874 : /*
875 : * Three four byte quantities (magic number, truncation block length,
876 : * block count) followed by block numbers.
877 : */
878 13342 : result = 3 * sizeof(uint32) + (sizeof(BlockNumber) * num_blocks_required);
879 :
880 : /*
881 : * Round the header size to a multiple of BLCKSZ - when not a multiple of
882 : * BLCKSZ, add the missing fraction of a block. But do this only if the
883 : * file will store data for some blocks, otherwise keep it small.
884 : */
885 13342 : if ((num_blocks_required > 0) && (result % BLCKSZ != 0))
886 50 : result += BLCKSZ - (result % BLCKSZ);
887 :
888 13342 : return result;
889 : }
890 :
891 : /*
892 : * Compute the size for an incremental file containing a given number of blocks.
893 : */
894 : size_t
895 13342 : GetIncrementalFileSize(unsigned num_blocks_required)
896 : {
897 : size_t result;
898 :
899 : /* Make sure we're not going to overflow. */
900 : Assert(num_blocks_required <= RELSEG_SIZE);
901 :
902 : /*
903 : * Header with three four byte quantities (magic number, truncation block
904 : * length, block count) followed by block numbers, rounded to a multiple
905 : * of BLCKSZ (for files with block data), followed by block contents.
906 : */
907 13342 : result = GetIncrementalHeaderSize(num_blocks_required);
908 13342 : result += BLCKSZ * num_blocks_required;
909 :
910 13342 : return result;
911 : }
912 :
913 : /*
914 : * Helper function for filemap hash table.
915 : */
916 : static uint32
917 42406 : hash_string_pointer(const char *s)
918 : {
919 42406 : unsigned char *ss = (unsigned char *) s;
920 :
921 42406 : return hash_bytes(ss, strlen(s));
922 : }
923 :
924 : /*
925 : * This callback to validate the manifest version for incremental backup.
926 : */
927 : static void
928 22 : manifest_process_version(JsonManifestParseContext *context,
929 : int manifest_version)
930 : {
931 : /* Incremental backups don't work with manifest version 1 */
932 22 : if (manifest_version == 1)
933 0 : context->error_cb(context,
934 : "backup manifest version 1 does not support incremental backup");
935 22 : }
936 :
937 : /*
938 : * This callback to validate the manifest system identifier against the current
939 : * database server.
940 : */
941 : static void
942 22 : manifest_process_system_identifier(JsonManifestParseContext *context,
943 : uint64 manifest_system_identifier)
944 : {
945 : uint64 system_identifier;
946 :
947 : /* Get system identifier of current system */
948 22 : system_identifier = GetSystemIdentifier();
949 :
950 22 : if (manifest_system_identifier != system_identifier)
951 2 : context->error_cb(context,
952 : "system identifier in backup manifest is %" PRIu64 ", but database system identifier is %" PRIu64,
953 : manifest_system_identifier,
954 : system_identifier);
955 20 : }
956 :
957 : /*
958 : * This callback is invoked for each file mentioned in the backup manifest.
959 : *
960 : * We store the path to each file and the size of each file for sanity-checking
961 : * purposes. For further details, see comments for IncrementalBackupInfo.
962 : */
963 : static void
964 20690 : manifest_process_file(JsonManifestParseContext *context,
965 : const char *pathname, uint64 size,
966 : pg_checksum_type checksum_type,
967 : int checksum_length,
968 : uint8 *checksum_payload)
969 : {
970 20690 : IncrementalBackupInfo *ib = context->private_data;
971 : backup_file_entry *entry;
972 : bool found;
973 :
974 20690 : entry = backup_file_insert(ib->manifest_files, pathname, &found);
975 20690 : if (!found)
976 : {
977 20690 : entry->path = MemoryContextStrdup(ib->manifest_files->ctx,
978 : pathname);
979 20690 : entry->size = size;
980 : }
981 20690 : }
982 :
983 : /*
984 : * This callback is invoked for each WAL range mentioned in the backup
985 : * manifest.
986 : *
987 : * We're just interested in learning the oldest LSN and the corresponding TLI
988 : * that appear in any WAL range.
989 : */
990 : static void
991 20 : manifest_process_wal_range(JsonManifestParseContext *context,
992 : TimeLineID tli, XLogRecPtr start_lsn,
993 : XLogRecPtr end_lsn)
994 : {
995 20 : IncrementalBackupInfo *ib = context->private_data;
996 20 : backup_wal_range *range = palloc(sizeof(backup_wal_range));
997 :
998 20 : range->tli = tli;
999 20 : range->start_lsn = start_lsn;
1000 20 : range->end_lsn = end_lsn;
1001 20 : ib->manifest_wal_ranges = lappend(ib->manifest_wal_ranges, range);
1002 20 : }
1003 :
1004 : /*
1005 : * This callback is invoked if an error occurs while parsing the backup
1006 : * manifest.
1007 : */
1008 : static void
1009 2 : manifest_report_error(JsonManifestParseContext *context, const char *fmt,...)
1010 : {
1011 : StringInfoData errbuf;
1012 :
1013 2 : initStringInfo(&errbuf);
1014 :
1015 : for (;;)
1016 0 : {
1017 : va_list ap;
1018 : int needed;
1019 :
1020 2 : va_start(ap, fmt);
1021 2 : needed = appendStringInfoVA(&errbuf, fmt, ap);
1022 2 : va_end(ap);
1023 2 : if (needed == 0)
1024 2 : break;
1025 0 : enlargeStringInfo(&errbuf, needed);
1026 : }
1027 :
1028 2 : ereport(ERROR,
1029 : errmsg_internal("%s", errbuf.data));
1030 : }
1031 :
1032 : /*
1033 : * Quicksort comparator for block numbers.
1034 : */
1035 : static int
1036 54 : compare_block_numbers(const void *a, const void *b)
1037 : {
1038 54 : BlockNumber aa = *(BlockNumber *) a;
1039 54 : BlockNumber bb = *(BlockNumber *) b;
1040 :
1041 54 : return pg_cmp_u32(aa, bb);
1042 : }
|
