Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * rewriteheap.c
4 : * Support functions to rewrite tables.
5 : *
6 : * These functions provide a facility to completely rewrite a heap, while
7 : * preserving visibility information and update chains.
8 : *
9 : * INTERFACE
10 : *
11 : * The caller is responsible for creating the new heap, all catalog
12 : * changes, supplying the tuples to be written to the new heap, and
13 : * rebuilding indexes. The caller must hold AccessExclusiveLock on the
14 : * target table, because we assume no one else is writing into it.
15 : *
16 : * To use the facility:
17 : *
18 : * begin_heap_rewrite
19 : * while (fetch next tuple)
20 : * {
21 : * if (tuple is dead)
22 : * rewrite_heap_dead_tuple
23 : * else
24 : * {
25 : * // do any transformations here if required
26 : * rewrite_heap_tuple
27 : * }
28 : * }
29 : * end_heap_rewrite
30 : *
31 : * The contents of the new relation shouldn't be relied on until after
32 : * end_heap_rewrite is called.
33 : *
34 : *
35 : * IMPLEMENTATION
36 : *
37 : * This would be a fairly trivial affair, except that we need to maintain
38 : * the ctid chains that link versions of an updated tuple together.
39 : * Since the newly stored tuples will have tids different from the original
40 : * ones, if we just copied t_ctid fields to the new table the links would
41 : * be wrong. When we are required to copy a (presumably recently-dead or
42 : * delete-in-progress) tuple whose ctid doesn't point to itself, we have
43 : * to substitute the correct ctid instead.
44 : *
45 : * For each ctid reference from A -> B, we might encounter either A first
46 : * or B first. (Note that a tuple in the middle of a chain is both A and B
47 : * of different pairs.)
48 : *
49 : * If we encounter A first, we'll store the tuple in the unresolved_tups
50 : * hash table. When we later encounter B, we remove A from the hash table,
51 : * fix the ctid to point to the new location of B, and insert both A and B
52 : * to the new heap.
53 : *
54 : * If we encounter B first, we can insert B to the new heap right away.
55 : * We then add an entry to the old_new_tid_map hash table showing B's
56 : * original tid (in the old heap) and new tid (in the new heap).
57 : * When we later encounter A, we get the new location of B from the table,
58 : * and can write A immediately with the correct ctid.
59 : *
60 : * Entries in the hash tables can be removed as soon as the later tuple
61 : * is encountered. That helps to keep the memory usage down. At the end,
62 : * both tables are usually empty; we should have encountered both A and B
63 : * of each pair. However, it's possible for A to be RECENTLY_DEAD and B
64 : * entirely DEAD according to HeapTupleSatisfiesVacuum, because the test
65 : * for deadness using OldestXmin is not exact. In such a case we might
66 : * encounter B first, and skip it, and find A later. Then A would be added
67 : * to unresolved_tups, and stay there until end of the rewrite. Since
68 : * this case is very unusual, we don't worry about the memory usage.
69 : *
70 : * Using in-memory hash tables means that we use some memory for each live
71 : * update chain in the table, from the time we find one end of the
72 : * reference until we find the other end. That shouldn't be a problem in
73 : * practice, but if you do something like an UPDATE without a where-clause
74 : * on a large table, and then run CLUSTER in the same transaction, you
75 : * could run out of memory. It doesn't seem worthwhile to add support for
76 : * spill-to-disk, as there shouldn't be that many RECENTLY_DEAD tuples in a
77 : * table under normal circumstances. Furthermore, in the typical scenario
78 : * of CLUSTERing on an unchanging key column, we'll see all the versions
79 : * of a given tuple together anyway, and so the peak memory usage is only
80 : * proportional to the number of RECENTLY_DEAD versions of a single row, not
81 : * in the whole table. Note that if we do fail halfway through a CLUSTER,
82 : * the old table is still valid, so failure is not catastrophic.
83 : *
84 : * We can't use the normal heap_insert function to insert into the new
85 : * heap, because heap_insert overwrites the visibility information.
86 : * We use a special-purpose raw_heap_insert function instead, which
87 : * is optimized for bulk inserting a lot of tuples, knowing that we have
88 : * exclusive access to the heap. raw_heap_insert builds new pages in
89 : * local storage. When a page is full, or at the end of the process,
90 : * we insert it to WAL as a single record and then write it to disk with
91 : * the bulk smgr writer. Note, however, that any data sent to the new
92 : * heap's TOAST table will go through the normal bufmgr.
93 : *
94 : *
95 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
96 : * Portions Copyright (c) 1994-5, Regents of the University of California
97 : *
98 : * IDENTIFICATION
99 : * src/backend/access/heap/rewriteheap.c
100 : *
101 : *-------------------------------------------------------------------------
102 : */
103 : #include "postgres.h"
104 :
105 : #include <unistd.h>
106 :
107 : #include "access/heapam.h"
108 : #include "access/heapam_xlog.h"
109 : #include "access/heaptoast.h"
110 : #include "access/rewriteheap.h"
111 : #include "access/transam.h"
112 : #include "access/xact.h"
113 : #include "access/xloginsert.h"
114 : #include "common/file_utils.h"
115 : #include "lib/ilist.h"
116 : #include "miscadmin.h"
117 : #include "pgstat.h"
118 : #include "replication/slot.h"
119 : #include "storage/bufmgr.h"
120 : #include "storage/bulk_write.h"
121 : #include "storage/fd.h"
122 : #include "storage/procarray.h"
123 : #include "utils/memutils.h"
124 : #include "utils/rel.h"
125 : #include "utils/wait_event.h"
126 :
127 : /*
128 : * State associated with a rewrite operation. This is opaque to the user
129 : * of the rewrite facility.
130 : */
131 : typedef struct RewriteStateData
132 : {
133 : Relation rs_old_rel; /* source heap */
134 : Relation rs_new_rel; /* destination heap */
135 : BulkWriteState *rs_bulkstate; /* writer for the destination */
136 : BulkWriteBuffer rs_buffer; /* page currently being built */
137 : BlockNumber rs_blockno; /* block where page will go */
138 : bool rs_logical_rewrite; /* do we need to do logical rewriting */
139 : TransactionId rs_oldest_xmin; /* oldest xmin used by caller to determine
140 : * tuple visibility */
141 : TransactionId rs_freeze_xid; /* Xid that will be used as freeze cutoff
142 : * point */
143 : TransactionId rs_logical_xmin; /* Xid that will be used as cutoff point
144 : * for logical rewrites */
145 : MultiXactId rs_cutoff_multi; /* MultiXactId that will be used as cutoff
146 : * point for multixacts */
147 : MemoryContext rs_cxt; /* for hash tables and entries and tuples in
148 : * them */
149 : XLogRecPtr rs_begin_lsn; /* XLogInsertLsn when starting the rewrite */
150 : HTAB *rs_unresolved_tups; /* unmatched A tuples */
151 : HTAB *rs_old_new_tid_map; /* unmatched B tuples */
152 : HTAB *rs_logical_mappings; /* logical remapping files */
153 : uint32 rs_num_rewrite_mappings; /* # in memory mappings */
154 : } RewriteStateData;
155 :
156 : /*
157 : * The lookup keys for the hash tables are tuple TID and xmin (we must check
158 : * both to avoid false matches from dead tuples). Beware that there is
159 : * probably some padding space in this struct; it must be zeroed out for
160 : * correct hashtable operation.
161 : */
162 : typedef struct
163 : {
164 : TransactionId xmin; /* tuple xmin */
165 : ItemPointerData tid; /* tuple location in old heap */
166 : } TidHashKey;
167 :
168 : /*
169 : * Entry structures for the hash tables
170 : */
171 : typedef struct
172 : {
173 : TidHashKey key; /* expected xmin/old location of B tuple */
174 : ItemPointerData old_tid; /* A's location in the old heap */
175 : HeapTuple tuple; /* A's tuple contents */
176 : } UnresolvedTupData;
177 :
178 : typedef UnresolvedTupData *UnresolvedTup;
179 :
180 : typedef struct
181 : {
182 : TidHashKey key; /* actual xmin/old location of B tuple */
183 : ItemPointerData new_tid; /* where we put it in the new heap */
184 : } OldToNewMappingData;
185 :
186 : typedef OldToNewMappingData *OldToNewMapping;
187 :
188 : /*
189 : * In-Memory data for an xid that might need logical remapping entries
190 : * to be logged.
191 : */
192 : typedef struct RewriteMappingFile
193 : {
194 : TransactionId xid; /* xid that might need to see the row */
195 : int vfd; /* fd of mappings file */
196 : off_t off; /* how far have we written yet */
197 : dclist_head mappings; /* list of in-memory mappings */
198 : char path[MAXPGPATH]; /* path, for error messages */
199 : } RewriteMappingFile;
200 :
201 : /*
202 : * A single In-Memory logical rewrite mapping, hanging off
203 : * RewriteMappingFile->mappings.
204 : */
205 : typedef struct RewriteMappingDataEntry
206 : {
207 : LogicalRewriteMappingData map; /* map between old and new location of the
208 : * tuple */
209 : dlist_node node;
210 : } RewriteMappingDataEntry;
211 :
212 :
213 : /* prototypes for internal functions */
214 : static void raw_heap_insert(RewriteState state, HeapTuple tup);
215 :
216 : /* internal logical remapping prototypes */
217 : static void logical_begin_heap_rewrite(RewriteState state);
218 : static void logical_rewrite_heap_tuple(RewriteState state, ItemPointerData old_tid, HeapTuple new_tuple);
219 : static void logical_end_heap_rewrite(RewriteState state);
220 :
221 :
222 : /*
223 : * Begin a rewrite of a table
224 : *
225 : * old_heap old, locked heap relation tuples will be read from
226 : * new_heap new, locked heap relation to insert tuples to
227 : * oldest_xmin xid used by the caller to determine which tuples are dead
228 : * freeze_xid xid before which tuples will be frozen
229 : * cutoff_multi multixact before which multis will be removed
230 : *
231 : * Returns an opaque RewriteState, allocated in current memory context,
232 : * to be used in subsequent calls to the other functions.
233 : */
234 : RewriteState
235 319 : begin_heap_rewrite(Relation old_heap, Relation new_heap, TransactionId oldest_xmin,
236 : TransactionId freeze_xid, MultiXactId cutoff_multi)
237 : {
238 : RewriteState state;
239 : MemoryContext rw_cxt;
240 : MemoryContext old_cxt;
241 : HASHCTL hash_ctl;
242 :
243 : /*
244 : * To ease cleanup, make a separate context that will contain the
245 : * RewriteState struct itself plus all subsidiary data.
246 : */
247 319 : rw_cxt = AllocSetContextCreate(CurrentMemoryContext,
248 : "Table rewrite",
249 : ALLOCSET_DEFAULT_SIZES);
250 319 : old_cxt = MemoryContextSwitchTo(rw_cxt);
251 :
252 : /* Create and fill in the state struct */
253 319 : state = palloc0_object(RewriteStateData);
254 :
255 319 : state->rs_old_rel = old_heap;
256 319 : state->rs_new_rel = new_heap;
257 319 : state->rs_buffer = NULL;
258 : /* new_heap needn't be empty, just locked */
259 319 : state->rs_blockno = RelationGetNumberOfBlocks(new_heap);
260 319 : state->rs_oldest_xmin = oldest_xmin;
261 319 : state->rs_freeze_xid = freeze_xid;
262 319 : state->rs_cutoff_multi = cutoff_multi;
263 319 : state->rs_cxt = rw_cxt;
264 319 : state->rs_bulkstate = smgr_bulk_start_rel(new_heap, MAIN_FORKNUM);
265 :
266 : /* Initialize hash tables used to track update chains */
267 319 : hash_ctl.keysize = sizeof(TidHashKey);
268 319 : hash_ctl.entrysize = sizeof(UnresolvedTupData);
269 319 : hash_ctl.hcxt = state->rs_cxt;
270 :
271 319 : state->rs_unresolved_tups =
272 319 : hash_create("Rewrite / Unresolved ctids",
273 : 128, /* arbitrary initial size */
274 : &hash_ctl,
275 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
276 :
277 319 : hash_ctl.entrysize = sizeof(OldToNewMappingData);
278 :
279 319 : state->rs_old_new_tid_map =
280 319 : hash_create("Rewrite / Old to new tid map",
281 : 128, /* arbitrary initial size */
282 : &hash_ctl,
283 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
284 :
285 319 : MemoryContextSwitchTo(old_cxt);
286 :
287 319 : logical_begin_heap_rewrite(state);
288 :
289 319 : return state;
290 : }
291 :
292 : /*
293 : * End a rewrite.
294 : *
295 : * state and any other resources are freed.
296 : */
297 : void
298 319 : end_heap_rewrite(RewriteState state)
299 : {
300 : HASH_SEQ_STATUS seq_status;
301 : UnresolvedTup unresolved;
302 :
303 : /*
304 : * Write any remaining tuples in the UnresolvedTups table. If we have any
305 : * left, they should in fact be dead, but let's err on the safe side.
306 : */
307 319 : hash_seq_init(&seq_status, state->rs_unresolved_tups);
308 :
309 319 : while ((unresolved = hash_seq_search(&seq_status)) != NULL)
310 : {
311 0 : ItemPointerSetInvalid(&unresolved->tuple->t_data->t_ctid);
312 0 : raw_heap_insert(state, unresolved->tuple);
313 : }
314 :
315 : /* Write the last page, if any */
316 319 : if (state->rs_buffer)
317 : {
318 213 : smgr_bulk_write(state->rs_bulkstate, state->rs_blockno, state->rs_buffer, true);
319 213 : state->rs_buffer = NULL;
320 : }
321 :
322 319 : smgr_bulk_finish(state->rs_bulkstate);
323 :
324 319 : logical_end_heap_rewrite(state);
325 :
326 : /* Deleting the context frees everything */
327 319 : MemoryContextDelete(state->rs_cxt);
328 319 : }
329 :
330 : /*
331 : * Add a tuple to the new heap.
332 : *
333 : * Visibility information is copied from the original tuple, except that
334 : * we "freeze" very-old tuples. Note that since we scribble on new_tuple,
335 : * it had better be temp storage not a pointer to the original tuple.
336 : *
337 : * state opaque state as returned by begin_heap_rewrite
338 : * old_tuple original tuple in the old heap
339 : * new_tuple new, rewritten tuple to be inserted to new heap
340 : */
341 : void
342 358768 : rewrite_heap_tuple(RewriteState state,
343 : HeapTuple old_tuple, HeapTuple new_tuple)
344 : {
345 : MemoryContext old_cxt;
346 : ItemPointerData old_tid;
347 : TidHashKey hashkey;
348 : bool found;
349 : bool free_new;
350 :
351 358768 : old_cxt = MemoryContextSwitchTo(state->rs_cxt);
352 :
353 : /*
354 : * Copy the original tuple's visibility information into new_tuple.
355 : *
356 : * XXX we might later need to copy some t_infomask2 bits, too? Right now,
357 : * we intentionally clear the HOT status bits.
358 : */
359 358768 : memcpy(&new_tuple->t_data->t_choice.t_heap,
360 358768 : &old_tuple->t_data->t_choice.t_heap,
361 : sizeof(HeapTupleFields));
362 :
363 358768 : new_tuple->t_data->t_infomask &= ~HEAP_XACT_MASK;
364 358768 : new_tuple->t_data->t_infomask2 &= ~HEAP2_XACT_MASK;
365 358768 : new_tuple->t_data->t_infomask |=
366 358768 : old_tuple->t_data->t_infomask & HEAP_XACT_MASK;
367 :
368 : /*
369 : * While we have our hands on the tuple, we may as well freeze any
370 : * eligible xmin or xmax, so that future VACUUM effort can be saved.
371 : */
372 358768 : heap_freeze_tuple(new_tuple->t_data,
373 358768 : state->rs_old_rel->rd_rel->relfrozenxid,
374 358768 : state->rs_old_rel->rd_rel->relminmxid,
375 : state->rs_freeze_xid,
376 : state->rs_cutoff_multi);
377 :
378 : /*
379 : * Invalid ctid means that ctid should point to the tuple itself. We'll
380 : * override it later if the tuple is part of an update chain.
381 : */
382 358768 : ItemPointerSetInvalid(&new_tuple->t_data->t_ctid);
383 :
384 : /*
385 : * If the tuple has been updated, check the old-to-new mapping hash table.
386 : *
387 : * Note that this check relies on the HeapTupleSatisfiesVacuum() in
388 : * heapam_relation_copy_for_cluster() to have set hint bits.
389 : */
390 369613 : if (!((old_tuple->t_data->t_infomask & HEAP_XMAX_INVALID) ||
391 10845 : HeapTupleHeaderIsOnlyLocked(old_tuple->t_data)) &&
392 10845 : !HeapTupleHeaderIndicatesMovedPartitions(old_tuple->t_data) &&
393 10845 : !(ItemPointerEquals(&(old_tuple->t_self),
394 10845 : &(old_tuple->t_data->t_ctid))))
395 : {
396 : OldToNewMapping mapping;
397 :
398 436 : memset(&hashkey, 0, sizeof(hashkey));
399 436 : hashkey.xmin = HeapTupleHeaderGetUpdateXid(old_tuple->t_data);
400 436 : hashkey.tid = old_tuple->t_data->t_ctid;
401 :
402 : mapping = (OldToNewMapping)
403 436 : hash_search(state->rs_old_new_tid_map, &hashkey,
404 : HASH_FIND, NULL);
405 :
406 436 : if (mapping != NULL)
407 : {
408 : /*
409 : * We've already copied the tuple that t_ctid points to, so we can
410 : * set the ctid of this tuple to point to the new location, and
411 : * insert it right away.
412 : */
413 208 : new_tuple->t_data->t_ctid = mapping->new_tid;
414 :
415 : /* We don't need the mapping entry anymore */
416 208 : hash_search(state->rs_old_new_tid_map, &hashkey,
417 : HASH_REMOVE, &found);
418 : Assert(found);
419 : }
420 : else
421 : {
422 : /*
423 : * We haven't seen the tuple t_ctid points to yet. Stash this
424 : * tuple into unresolved_tups to be written later.
425 : */
426 : UnresolvedTup unresolved;
427 :
428 228 : unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
429 : HASH_ENTER, &found);
430 : Assert(!found);
431 :
432 228 : unresolved->old_tid = old_tuple->t_self;
433 228 : unresolved->tuple = heap_copytuple(new_tuple);
434 :
435 : /*
436 : * We can't do anything more now, since we don't know where the
437 : * tuple will be written.
438 : */
439 228 : MemoryContextSwitchTo(old_cxt);
440 228 : return;
441 : }
442 : }
443 :
444 : /*
445 : * Now we will write the tuple, and then check to see if it is the B tuple
446 : * in any new or known pair. When we resolve a known pair, we will be
447 : * able to write that pair's A tuple, and then we have to check if it
448 : * resolves some other pair. Hence, we need a loop here.
449 : */
450 358540 : old_tid = old_tuple->t_self;
451 358540 : free_new = false;
452 :
453 : for (;;)
454 228 : {
455 : ItemPointerData new_tid;
456 :
457 : /* Insert the tuple and find out where it's put in new_heap */
458 358768 : raw_heap_insert(state, new_tuple);
459 358768 : new_tid = new_tuple->t_self;
460 :
461 358768 : logical_rewrite_heap_tuple(state, old_tid, new_tuple);
462 :
463 : /*
464 : * If the tuple is the updated version of a row, and the prior version
465 : * wouldn't be DEAD yet, then we need to either resolve the prior
466 : * version (if it's waiting in rs_unresolved_tups), or make an entry
467 : * in rs_old_new_tid_map (so we can resolve it when we do see it). The
468 : * previous tuple's xmax would equal this one's xmin, so it's
469 : * RECENTLY_DEAD if and only if the xmin is not before OldestXmin.
470 : */
471 358768 : if ((new_tuple->t_data->t_infomask & HEAP_UPDATED) &&
472 10008 : !TransactionIdPrecedes(HeapTupleHeaderGetXmin(new_tuple->t_data),
473 : state->rs_oldest_xmin))
474 : {
475 : /*
476 : * Okay, this is B in an update pair. See if we've seen A.
477 : */
478 : UnresolvedTup unresolved;
479 :
480 436 : memset(&hashkey, 0, sizeof(hashkey));
481 436 : hashkey.xmin = HeapTupleHeaderGetXmin(new_tuple->t_data);
482 436 : hashkey.tid = old_tid;
483 :
484 436 : unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
485 : HASH_FIND, NULL);
486 :
487 436 : if (unresolved != NULL)
488 : {
489 : /*
490 : * We have seen and memorized the previous tuple already. Now
491 : * that we know where we inserted the tuple its t_ctid points
492 : * to, fix its t_ctid and insert it to the new heap.
493 : */
494 228 : if (free_new)
495 59 : heap_freetuple(new_tuple);
496 228 : new_tuple = unresolved->tuple;
497 228 : free_new = true;
498 228 : old_tid = unresolved->old_tid;
499 228 : new_tuple->t_data->t_ctid = new_tid;
500 :
501 : /*
502 : * We don't need the hash entry anymore, but don't free its
503 : * tuple just yet.
504 : */
505 228 : hash_search(state->rs_unresolved_tups, &hashkey,
506 : HASH_REMOVE, &found);
507 : Assert(found);
508 :
509 : /* loop back to insert the previous tuple in the chain */
510 228 : continue;
511 : }
512 : else
513 : {
514 : /*
515 : * Remember the new tid of this tuple. We'll use it to set the
516 : * ctid when we find the previous tuple in the chain.
517 : */
518 : OldToNewMapping mapping;
519 :
520 208 : mapping = hash_search(state->rs_old_new_tid_map, &hashkey,
521 : HASH_ENTER, &found);
522 : Assert(!found);
523 :
524 208 : mapping->new_tid = new_tid;
525 : }
526 : }
527 :
528 : /* Done with this (chain of) tuples, for now */
529 358540 : if (free_new)
530 169 : heap_freetuple(new_tuple);
531 358540 : break;
532 : }
533 :
534 358540 : MemoryContextSwitchTo(old_cxt);
535 : }
536 :
537 : /*
538 : * Register a dead tuple with an ongoing rewrite. Dead tuples are not
539 : * copied to the new table, but we still make note of them so that we
540 : * can release some resources earlier.
541 : *
542 : * Returns true if a tuple was removed from the unresolved_tups table.
543 : * This indicates that that tuple, previously thought to be "recently dead",
544 : * is now known really dead and won't be written to the output.
545 : */
546 : bool
547 16476 : rewrite_heap_dead_tuple(RewriteState state, HeapTuple old_tuple)
548 : {
549 : /*
550 : * If we have already seen an earlier tuple in the update chain that
551 : * points to this tuple, let's forget about that earlier tuple. It's in
552 : * fact dead as well, our simple xmax < OldestXmin test in
553 : * HeapTupleSatisfiesVacuum just wasn't enough to detect it. It happens
554 : * when xmin of a tuple is greater than xmax, which sounds
555 : * counter-intuitive but is perfectly valid.
556 : *
557 : * We don't bother to try to detect the situation the other way round,
558 : * when we encounter the dead tuple first and then the recently dead one
559 : * that points to it. If that happens, we'll have some unmatched entries
560 : * in the UnresolvedTups hash table at the end. That can happen anyway,
561 : * because a vacuum might have removed the dead tuple in the chain before
562 : * us.
563 : */
564 : UnresolvedTup unresolved;
565 : TidHashKey hashkey;
566 : bool found;
567 :
568 16476 : memset(&hashkey, 0, sizeof(hashkey));
569 16476 : hashkey.xmin = HeapTupleHeaderGetXmin(old_tuple->t_data);
570 16476 : hashkey.tid = old_tuple->t_self;
571 :
572 16476 : unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
573 : HASH_FIND, NULL);
574 :
575 16476 : if (unresolved != NULL)
576 : {
577 : /* Need to free the contained tuple as well as the hashtable entry */
578 0 : heap_freetuple(unresolved->tuple);
579 0 : hash_search(state->rs_unresolved_tups, &hashkey,
580 : HASH_REMOVE, &found);
581 : Assert(found);
582 0 : return true;
583 : }
584 :
585 16476 : return false;
586 : }
587 :
588 : /*
589 : * Insert a tuple to the new relation. This has to track heap_insert
590 : * and its subsidiary functions!
591 : *
592 : * t_self of the tuple is set to the new TID of the tuple. If t_ctid of the
593 : * tuple is invalid on entry, it's replaced with the new TID as well (in
594 : * the inserted data only, not in the caller's copy).
595 : */
596 : static void
597 358768 : raw_heap_insert(RewriteState state, HeapTuple tup)
598 : {
599 : Page page;
600 : Size pageFreeSpace,
601 : saveFreeSpace;
602 : Size len;
603 : OffsetNumber newoff;
604 : HeapTuple heaptup;
605 :
606 : /*
607 : * If the new tuple is too big for storage or contains already toasted
608 : * out-of-line attributes from some other relation, invoke the toaster.
609 : *
610 : * Note: below this point, heaptup is the data we actually intend to store
611 : * into the relation; tup is the caller's original untoasted data.
612 : */
613 358768 : if (state->rs_new_rel->rd_rel->relkind == RELKIND_TOASTVALUE)
614 : {
615 : /* toast table entries should never be recursively toasted */
616 : Assert(!HeapTupleHasExternal(tup));
617 0 : heaptup = tup;
618 : }
619 358768 : else if (HeapTupleHasExternal(tup) || tup->t_len > TOAST_TUPLE_THRESHOLD)
620 321 : {
621 321 : int options = HEAP_INSERT_SKIP_FSM;
622 :
623 : /*
624 : * While rewriting the heap for VACUUM FULL / CLUSTER, make sure data
625 : * for the TOAST table are not logically decoded. The main heap is
626 : * WAL-logged as XLOG FPI records, which are not logically decoded.
627 : */
628 321 : options |= HEAP_INSERT_NO_LOGICAL;
629 :
630 321 : heaptup = heap_toast_insert_or_update(state->rs_new_rel, tup, NULL,
631 : options);
632 : }
633 : else
634 358447 : heaptup = tup;
635 :
636 358768 : len = MAXALIGN(heaptup->t_len); /* be conservative */
637 :
638 : /*
639 : * If we're gonna fail for oversize tuple, do it right away
640 : */
641 358768 : if (len > MaxHeapTupleSize)
642 0 : ereport(ERROR,
643 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
644 : errmsg("row is too big: size %zu, maximum size %zu",
645 : len, MaxHeapTupleSize)));
646 :
647 : /* Compute desired extra freespace due to fillfactor option */
648 358768 : saveFreeSpace = RelationGetTargetPageFreeSpace(state->rs_new_rel,
649 : HEAP_DEFAULT_FILLFACTOR);
650 :
651 : /* Now we can check to see if there's enough free space already. */
652 358768 : page = (Page) state->rs_buffer;
653 358768 : if (page)
654 : {
655 358555 : pageFreeSpace = PageGetHeapFreeSpace(page);
656 :
657 358555 : if (len + saveFreeSpace > pageFreeSpace)
658 : {
659 : /*
660 : * Doesn't fit, so write out the existing page. It always
661 : * contains a tuple. Hence, unlike RelationGetBufferForTuple(),
662 : * enforce saveFreeSpace unconditionally.
663 : */
664 5134 : smgr_bulk_write(state->rs_bulkstate, state->rs_blockno, state->rs_buffer, true);
665 5134 : state->rs_buffer = NULL;
666 5134 : page = NULL;
667 5134 : state->rs_blockno++;
668 : }
669 : }
670 :
671 358768 : if (!page)
672 : {
673 : /* Initialize a new empty page */
674 5347 : state->rs_buffer = smgr_bulk_get_buf(state->rs_bulkstate);
675 5347 : page = (Page) state->rs_buffer;
676 5347 : PageInit(page, BLCKSZ, 0);
677 : }
678 :
679 : /* And now we can insert the tuple into the page */
680 358768 : newoff = PageAddItem(page, heaptup->t_data, heaptup->t_len, InvalidOffsetNumber, false, true);
681 358768 : if (newoff == InvalidOffsetNumber)
682 0 : elog(ERROR, "failed to add tuple");
683 :
684 : /* Update caller's t_self to the actual position where it was stored */
685 358768 : ItemPointerSet(&(tup->t_self), state->rs_blockno, newoff);
686 :
687 : /*
688 : * Insert the correct position into CTID of the stored tuple, too, if the
689 : * caller didn't supply a valid CTID.
690 : */
691 358768 : if (!ItemPointerIsValid(&tup->t_data->t_ctid))
692 : {
693 : ItemId newitemid;
694 : HeapTupleHeader onpage_tup;
695 :
696 358332 : newitemid = PageGetItemId(page, newoff);
697 358332 : onpage_tup = (HeapTupleHeader) PageGetItem(page, newitemid);
698 :
699 358332 : onpage_tup->t_ctid = tup->t_self;
700 : }
701 :
702 : /* If heaptup is a private copy, release it. */
703 358768 : if (heaptup != tup)
704 321 : heap_freetuple(heaptup);
705 358768 : }
706 :
707 : /* ------------------------------------------------------------------------
708 : * Logical rewrite support
709 : *
710 : * When doing logical decoding - which relies on using cmin/cmax of catalog
711 : * tuples, via xl_heap_new_cid records - heap rewrites have to log enough
712 : * information to allow the decoding backend to update its internal mapping
713 : * of (relfilelocator,ctid) => (cmin, cmax) to be correct for the rewritten heap.
714 : *
715 : * For that, every time we find a tuple that's been modified in a catalog
716 : * relation within the xmin horizon of any decoding slot, we log a mapping
717 : * from the old to the new location.
718 : *
719 : * To deal with rewrites that abort the filename of a mapping file contains
720 : * the xid of the transaction performing the rewrite, which then can be
721 : * checked before being read in.
722 : *
723 : * For efficiency we don't immediately spill every single map mapping for a
724 : * row to disk but only do so in batches when we've collected several of them
725 : * in memory or when end_heap_rewrite() has been called.
726 : *
727 : * Crash-Safety: This module diverts from the usual patterns of doing WAL
728 : * since it cannot rely on checkpoint flushing out all buffers and thus
729 : * waiting for exclusive locks on buffers. Usually the XLogInsert() covering
730 : * buffer modifications is performed while the buffer(s) that are being
731 : * modified are exclusively locked guaranteeing that both the WAL record and
732 : * the modified heap are on either side of the checkpoint. But since the
733 : * mapping files we log aren't in shared_buffers that interlock doesn't work.
734 : *
735 : * Instead we simply write the mapping files out to disk, *before* the
736 : * XLogInsert() is performed. That guarantees that either the XLogInsert() is
737 : * inserted after the checkpoint's redo pointer or that the checkpoint (via
738 : * CheckPointLogicalRewriteHeap()) has flushed the (partial) mapping file to
739 : * disk. That leaves the tail end that has not yet been flushed open to
740 : * corruption, which is solved by including the current offset in the
741 : * xl_heap_rewrite_mapping records and truncating the mapping file to it
742 : * during replay. Every time a rewrite is finished all generated mapping files
743 : * are synced to disk.
744 : *
745 : * Note that if we were only concerned about crash safety we wouldn't have to
746 : * deal with WAL logging at all - an fsync() at the end of a rewrite would be
747 : * sufficient for crash safety. Any mapping that hasn't been safely flushed to
748 : * disk has to be by an aborted (explicitly or via a crash) transaction and is
749 : * ignored by virtue of the xid in its name being subject to a
750 : * TransactionDidCommit() check. But we want to support having standbys via
751 : * physical replication, both for availability and to do logical decoding
752 : * there.
753 : * ------------------------------------------------------------------------
754 : */
755 :
756 : /*
757 : * Do preparations for logging logical mappings during a rewrite if
758 : * necessary. If we detect that we don't need to log anything we'll prevent
759 : * any further action by the various logical rewrite functions.
760 : */
761 : static void
762 319 : logical_begin_heap_rewrite(RewriteState state)
763 : {
764 : HASHCTL hash_ctl;
765 : TransactionId logical_xmin;
766 :
767 : /*
768 : * We only need to persist these mappings if the rewritten table can be
769 : * accessed during logical decoding, if not, we can skip doing any
770 : * additional work.
771 : */
772 319 : state->rs_logical_rewrite =
773 319 : RelationIsAccessibleInLogicalDecoding(state->rs_old_rel);
774 :
775 319 : if (!state->rs_logical_rewrite)
776 299 : return;
777 :
778 22 : ProcArrayGetReplicationSlotXmin(NULL, &logical_xmin);
779 :
780 : /*
781 : * If there are no logical slots in progress we don't need to do anything,
782 : * there cannot be any remappings for relevant rows yet. The relation's
783 : * lock protects us against races.
784 : */
785 22 : if (logical_xmin == InvalidTransactionId)
786 : {
787 2 : state->rs_logical_rewrite = false;
788 2 : return;
789 : }
790 :
791 20 : state->rs_logical_xmin = logical_xmin;
792 20 : state->rs_begin_lsn = GetXLogInsertRecPtr();
793 20 : state->rs_num_rewrite_mappings = 0;
794 :
795 20 : hash_ctl.keysize = sizeof(TransactionId);
796 20 : hash_ctl.entrysize = sizeof(RewriteMappingFile);
797 20 : hash_ctl.hcxt = state->rs_cxt;
798 :
799 20 : state->rs_logical_mappings =
800 20 : hash_create("Logical rewrite mapping",
801 : 128, /* arbitrary initial size */
802 : &hash_ctl,
803 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
804 : }
805 :
806 : /*
807 : * Flush all logical in-memory mappings to disk, but don't fsync them yet.
808 : */
809 : static void
810 9 : logical_heap_rewrite_flush_mappings(RewriteState state)
811 : {
812 : HASH_SEQ_STATUS seq_status;
813 : RewriteMappingFile *src;
814 : dlist_mutable_iter iter;
815 :
816 : Assert(state->rs_logical_rewrite);
817 :
818 : /* no logical rewrite in progress, no need to iterate over mappings */
819 9 : if (state->rs_num_rewrite_mappings == 0)
820 0 : return;
821 :
822 9 : elog(DEBUG1, "flushing %u logical rewrite mapping entries",
823 : state->rs_num_rewrite_mappings);
824 :
825 9 : hash_seq_init(&seq_status, state->rs_logical_mappings);
826 99 : while ((src = (RewriteMappingFile *) hash_seq_search(&seq_status)) != NULL)
827 : {
828 : char *waldata;
829 : char *waldata_start;
830 : xl_heap_rewrite_mapping xlrec;
831 : Oid dboid;
832 : uint32 len;
833 : int written;
834 90 : uint32 num_mappings = dclist_count(&src->mappings);
835 :
836 : /* this file hasn't got any new mappings */
837 90 : if (num_mappings == 0)
838 0 : continue;
839 :
840 90 : if (state->rs_old_rel->rd_rel->relisshared)
841 0 : dboid = InvalidOid;
842 : else
843 90 : dboid = MyDatabaseId;
844 :
845 90 : xlrec.num_mappings = num_mappings;
846 90 : xlrec.mapped_rel = RelationGetRelid(state->rs_old_rel);
847 90 : xlrec.mapped_xid = src->xid;
848 90 : xlrec.mapped_db = dboid;
849 90 : xlrec.offset = src->off;
850 90 : xlrec.start_lsn = state->rs_begin_lsn;
851 :
852 : /* write all mappings consecutively */
853 90 : len = num_mappings * sizeof(LogicalRewriteMappingData);
854 90 : waldata_start = waldata = palloc(len);
855 :
856 : /*
857 : * collect data we need to write out, but don't modify ondisk data yet
858 : */
859 813 : dclist_foreach_modify(iter, &src->mappings)
860 : {
861 : RewriteMappingDataEntry *pmap;
862 :
863 723 : pmap = dclist_container(RewriteMappingDataEntry, node, iter.cur);
864 :
865 723 : memcpy(waldata, &pmap->map, sizeof(pmap->map));
866 723 : waldata += sizeof(pmap->map);
867 :
868 : /* remove from the list and free */
869 723 : dclist_delete_from(&src->mappings, &pmap->node);
870 723 : pfree(pmap);
871 :
872 : /* update bookkeeping */
873 723 : state->rs_num_rewrite_mappings--;
874 : }
875 :
876 : Assert(dclist_count(&src->mappings) == 0);
877 : Assert(waldata == waldata_start + len);
878 :
879 : /*
880 : * Note that we deviate from the usual WAL coding practices here,
881 : * check the above "Logical rewrite support" comment for reasoning.
882 : */
883 90 : written = FileWrite(src->vfd, waldata_start, len, src->off,
884 : WAIT_EVENT_LOGICAL_REWRITE_WRITE);
885 90 : if (written != len)
886 0 : ereport(ERROR,
887 : (errcode_for_file_access(),
888 : errmsg("could not write to file \"%s\", wrote %d of %d: %m", src->path,
889 : written, len)));
890 90 : src->off += len;
891 :
892 90 : XLogBeginInsert();
893 90 : XLogRegisterData(&xlrec, sizeof(xlrec));
894 90 : XLogRegisterData(waldata_start, len);
895 :
896 : /* write xlog record */
897 90 : XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_REWRITE);
898 :
899 90 : pfree(waldata_start);
900 : }
901 : Assert(state->rs_num_rewrite_mappings == 0);
902 : }
903 :
904 : /*
905 : * Logical remapping part of end_heap_rewrite().
906 : */
907 : static void
908 319 : logical_end_heap_rewrite(RewriteState state)
909 : {
910 : HASH_SEQ_STATUS seq_status;
911 : RewriteMappingFile *src;
912 :
913 : /* done, no logical rewrite in progress */
914 319 : if (!state->rs_logical_rewrite)
915 299 : return;
916 :
917 : /* writeout remaining in-memory entries */
918 20 : if (state->rs_num_rewrite_mappings > 0)
919 9 : logical_heap_rewrite_flush_mappings(state);
920 :
921 : /* Iterate over all mappings we have written and fsync the files. */
922 20 : hash_seq_init(&seq_status, state->rs_logical_mappings);
923 110 : while ((src = (RewriteMappingFile *) hash_seq_search(&seq_status)) != NULL)
924 : {
925 90 : if (FileSync(src->vfd, WAIT_EVENT_LOGICAL_REWRITE_SYNC) != 0)
926 0 : ereport(data_sync_elevel(ERROR),
927 : (errcode_for_file_access(),
928 : errmsg("could not fsync file \"%s\": %m", src->path)));
929 90 : FileClose(src->vfd);
930 : }
931 : /* memory context cleanup will deal with the rest */
932 : }
933 :
934 : /*
935 : * Log a single (old->new) mapping for 'xid'.
936 : */
937 : static void
938 723 : logical_rewrite_log_mapping(RewriteState state, TransactionId xid,
939 : LogicalRewriteMappingData *map)
940 : {
941 : RewriteMappingFile *src;
942 : RewriteMappingDataEntry *pmap;
943 : Oid relid;
944 : bool found;
945 :
946 723 : relid = RelationGetRelid(state->rs_old_rel);
947 :
948 : /* look for existing mappings for this 'mapped' xid */
949 723 : src = hash_search(state->rs_logical_mappings, &xid,
950 : HASH_ENTER, &found);
951 :
952 : /*
953 : * We haven't yet had the need to map anything for this xid, create
954 : * per-xid data structures.
955 : */
956 723 : if (!found)
957 : {
958 : char path[MAXPGPATH];
959 : Oid dboid;
960 :
961 90 : if (state->rs_old_rel->rd_rel->relisshared)
962 0 : dboid = InvalidOid;
963 : else
964 90 : dboid = MyDatabaseId;
965 :
966 90 : snprintf(path, MAXPGPATH,
967 : "%s/" LOGICAL_REWRITE_FORMAT,
968 : PG_LOGICAL_MAPPINGS_DIR, dboid, relid,
969 90 : LSN_FORMAT_ARGS(state->rs_begin_lsn),
970 : xid, GetCurrentTransactionId());
971 :
972 90 : dclist_init(&src->mappings);
973 90 : src->off = 0;
974 90 : memcpy(src->path, path, sizeof(path));
975 90 : src->vfd = PathNameOpenFile(path,
976 : O_CREAT | O_EXCL | O_WRONLY | PG_BINARY);
977 90 : if (src->vfd < 0)
978 0 : ereport(ERROR,
979 : (errcode_for_file_access(),
980 : errmsg("could not create file \"%s\": %m", path)));
981 : }
982 :
983 723 : pmap = MemoryContextAlloc(state->rs_cxt,
984 : sizeof(RewriteMappingDataEntry));
985 723 : memcpy(&pmap->map, map, sizeof(LogicalRewriteMappingData));
986 723 : dclist_push_tail(&src->mappings, &pmap->node);
987 723 : state->rs_num_rewrite_mappings++;
988 :
989 : /*
990 : * Write out buffer every time we've too many in-memory entries across all
991 : * mapping files.
992 : */
993 723 : if (state->rs_num_rewrite_mappings >= 1000 /* arbitrary number */ )
994 0 : logical_heap_rewrite_flush_mappings(state);
995 723 : }
996 :
997 : /*
998 : * Perform logical remapping for a tuple that's mapped from old_tid to
999 : * new_tuple->t_self by rewrite_heap_tuple() if necessary for the tuple.
1000 : */
1001 : static void
1002 358768 : logical_rewrite_heap_tuple(RewriteState state, ItemPointerData old_tid,
1003 : HeapTuple new_tuple)
1004 : {
1005 358768 : ItemPointerData new_tid = new_tuple->t_self;
1006 358768 : TransactionId cutoff = state->rs_logical_xmin;
1007 : TransactionId xmin;
1008 : TransactionId xmax;
1009 358768 : bool do_log_xmin = false;
1010 358768 : bool do_log_xmax = false;
1011 : LogicalRewriteMappingData map;
1012 :
1013 : /* no logical rewrite in progress, we don't need to log anything */
1014 358768 : if (!state->rs_logical_rewrite)
1015 358060 : return;
1016 :
1017 26920 : xmin = HeapTupleHeaderGetXmin(new_tuple->t_data);
1018 : /* use *GetUpdateXid to correctly deal with multixacts */
1019 26920 : xmax = HeapTupleHeaderGetUpdateXid(new_tuple->t_data);
1020 :
1021 : /*
1022 : * Log the mapping iff the tuple has been created recently.
1023 : */
1024 26920 : if (TransactionIdIsNormal(xmin) && !TransactionIdPrecedes(xmin, cutoff))
1025 535 : do_log_xmin = true;
1026 :
1027 26920 : if (!TransactionIdIsNormal(xmax))
1028 : {
1029 : /*
1030 : * no xmax is set, can't have any permanent ones, so this check is
1031 : * sufficient
1032 : */
1033 : }
1034 501 : else if (HEAP_XMAX_IS_LOCKED_ONLY(new_tuple->t_data->t_infomask))
1035 : {
1036 : /* only locked, we don't care */
1037 : }
1038 501 : else if (!TransactionIdPrecedes(xmax, cutoff))
1039 : {
1040 : /* tuple has been deleted recently, log */
1041 501 : do_log_xmax = true;
1042 : }
1043 :
1044 : /* if neither needs to be logged, we're done */
1045 26920 : if (!do_log_xmin && !do_log_xmax)
1046 26212 : return;
1047 :
1048 : /* fill out mapping information */
1049 708 : map.old_locator = state->rs_old_rel->rd_locator;
1050 708 : map.old_tid = old_tid;
1051 708 : map.new_locator = state->rs_new_rel->rd_locator;
1052 708 : map.new_tid = new_tid;
1053 :
1054 : /* ---
1055 : * Now persist the mapping for the individual xids that are affected. We
1056 : * need to log for both xmin and xmax if they aren't the same transaction
1057 : * since the mapping files are per "affected" xid.
1058 : * We don't muster all that much effort detecting whether xmin and xmax
1059 : * are actually the same transaction, we just check whether the xid is the
1060 : * same disregarding subtransactions. Logging too much is relatively
1061 : * harmless and we could never do the check fully since subtransaction
1062 : * data is thrown away during restarts.
1063 : * ---
1064 : */
1065 708 : if (do_log_xmin)
1066 535 : logical_rewrite_log_mapping(state, xmin, &map);
1067 : /* separately log mapping for xmax unless it'd be redundant */
1068 708 : if (do_log_xmax && !TransactionIdEquals(xmin, xmax))
1069 188 : logical_rewrite_log_mapping(state, xmax, &map);
1070 : }
1071 :
1072 : /*
1073 : * Replay XLOG_HEAP2_REWRITE records
1074 : */
1075 : void
1076 0 : heap_xlog_logical_rewrite(XLogReaderState *r)
1077 : {
1078 : char path[MAXPGPATH];
1079 : int fd;
1080 : xl_heap_rewrite_mapping *xlrec;
1081 : uint32 len;
1082 : char *data;
1083 :
1084 0 : xlrec = (xl_heap_rewrite_mapping *) XLogRecGetData(r);
1085 :
1086 0 : snprintf(path, MAXPGPATH,
1087 : "%s/" LOGICAL_REWRITE_FORMAT,
1088 : PG_LOGICAL_MAPPINGS_DIR, xlrec->mapped_db, xlrec->mapped_rel,
1089 0 : LSN_FORMAT_ARGS(xlrec->start_lsn),
1090 0 : xlrec->mapped_xid, XLogRecGetXid(r));
1091 :
1092 0 : fd = OpenTransientFile(path,
1093 : O_CREAT | O_WRONLY | PG_BINARY);
1094 0 : if (fd < 0)
1095 0 : ereport(ERROR,
1096 : (errcode_for_file_access(),
1097 : errmsg("could not create file \"%s\": %m", path)));
1098 :
1099 : /*
1100 : * Truncate all data that's not guaranteed to have been safely fsynced (by
1101 : * previous record or by the last checkpoint).
1102 : */
1103 0 : pgstat_report_wait_start(WAIT_EVENT_LOGICAL_REWRITE_TRUNCATE);
1104 0 : if (ftruncate(fd, xlrec->offset) != 0)
1105 0 : ereport(ERROR,
1106 : (errcode_for_file_access(),
1107 : errmsg("could not truncate file \"%s\" to %u: %m",
1108 : path, (uint32) xlrec->offset)));
1109 0 : pgstat_report_wait_end();
1110 :
1111 0 : data = XLogRecGetData(r) + sizeof(*xlrec);
1112 :
1113 0 : len = xlrec->num_mappings * sizeof(LogicalRewriteMappingData);
1114 :
1115 : /* write out tail end of mapping file (again) */
1116 0 : errno = 0;
1117 0 : pgstat_report_wait_start(WAIT_EVENT_LOGICAL_REWRITE_MAPPING_WRITE);
1118 0 : if (pg_pwrite(fd, data, len, xlrec->offset) != len)
1119 : {
1120 : /* if write didn't set errno, assume problem is no disk space */
1121 0 : if (errno == 0)
1122 0 : errno = ENOSPC;
1123 0 : ereport(ERROR,
1124 : (errcode_for_file_access(),
1125 : errmsg("could not write to file \"%s\": %m", path)));
1126 : }
1127 0 : pgstat_report_wait_end();
1128 :
1129 : /*
1130 : * Now fsync all previously written data. We could improve things and only
1131 : * do this for the last write to a file, but the required bookkeeping
1132 : * doesn't seem worth the trouble.
1133 : */
1134 0 : pgstat_report_wait_start(WAIT_EVENT_LOGICAL_REWRITE_MAPPING_SYNC);
1135 0 : if (pg_fsync(fd) != 0)
1136 0 : ereport(data_sync_elevel(ERROR),
1137 : (errcode_for_file_access(),
1138 : errmsg("could not fsync file \"%s\": %m", path)));
1139 0 : pgstat_report_wait_end();
1140 :
1141 0 : if (CloseTransientFile(fd) != 0)
1142 0 : ereport(ERROR,
1143 : (errcode_for_file_access(),
1144 : errmsg("could not close file \"%s\": %m", path)));
1145 0 : }
1146 :
1147 : /* ---
1148 : * Perform a checkpoint for logical rewrite mappings
1149 : *
1150 : * This serves two tasks:
1151 : * 1) Remove all mappings not needed anymore based on the logical restart LSN
1152 : * 2) Flush all remaining mappings to disk, so that replay after a checkpoint
1153 : * only has to deal with the parts of a mapping that have been written out
1154 : * after the checkpoint started.
1155 : * ---
1156 : */
1157 : void
1158 1809 : CheckPointLogicalRewriteHeap(void)
1159 : {
1160 : XLogRecPtr cutoff;
1161 : XLogRecPtr redo;
1162 : DIR *mappings_dir;
1163 : struct dirent *mapping_de;
1164 : char path[MAXPGPATH + sizeof(PG_LOGICAL_MAPPINGS_DIR)];
1165 :
1166 : /*
1167 : * We start of with a minimum of the last redo pointer. No new decoding
1168 : * slot will start before that, so that's a safe upper bound for removal.
1169 : */
1170 1809 : redo = GetRedoRecPtr();
1171 :
1172 : /* now check for the restart ptrs from existing slots */
1173 1809 : cutoff = ReplicationSlotsComputeLogicalRestartLSN();
1174 :
1175 : /* don't start earlier than the restart lsn */
1176 1809 : if (XLogRecPtrIsValid(cutoff) && redo < cutoff)
1177 1 : cutoff = redo;
1178 :
1179 1809 : mappings_dir = AllocateDir(PG_LOGICAL_MAPPINGS_DIR);
1180 5607 : while ((mapping_de = ReadDir(mappings_dir, PG_LOGICAL_MAPPINGS_DIR)) != NULL)
1181 : {
1182 : Oid dboid;
1183 : Oid relid;
1184 : XLogRecPtr lsn;
1185 : TransactionId rewrite_xid;
1186 : TransactionId create_xid;
1187 : uint32 hi,
1188 : lo;
1189 : PGFileType de_type;
1190 :
1191 3798 : if (strcmp(mapping_de->d_name, ".") == 0 ||
1192 1989 : strcmp(mapping_de->d_name, "..") == 0)
1193 3618 : continue;
1194 :
1195 180 : snprintf(path, sizeof(path), "%s/%s", PG_LOGICAL_MAPPINGS_DIR, mapping_de->d_name);
1196 180 : de_type = get_dirent_type(path, mapping_de, false, DEBUG1);
1197 :
1198 180 : if (de_type != PGFILETYPE_ERROR && de_type != PGFILETYPE_REG)
1199 0 : continue;
1200 :
1201 : /* Skip over files that cannot be ours. */
1202 180 : if (strncmp(mapping_de->d_name, "map-", 4) != 0)
1203 0 : continue;
1204 :
1205 180 : if (sscanf(mapping_de->d_name, LOGICAL_REWRITE_FORMAT,
1206 : &dboid, &relid, &hi, &lo, &rewrite_xid, &create_xid) != 6)
1207 0 : elog(ERROR, "could not parse filename \"%s\"", mapping_de->d_name);
1208 :
1209 180 : lsn = ((uint64) hi) << 32 | lo;
1210 :
1211 180 : if (lsn < cutoff || !XLogRecPtrIsValid(cutoff))
1212 : {
1213 90 : elog(DEBUG1, "removing logical rewrite file \"%s\"", path);
1214 90 : if (unlink(path) < 0)
1215 0 : ereport(ERROR,
1216 : (errcode_for_file_access(),
1217 : errmsg("could not remove file \"%s\": %m", path)));
1218 : }
1219 : else
1220 : {
1221 : /* on some operating systems fsyncing a file requires O_RDWR */
1222 90 : int fd = OpenTransientFile(path, O_RDWR | PG_BINARY);
1223 :
1224 : /*
1225 : * The file cannot vanish due to concurrency since this function
1226 : * is the only one removing logical mappings and only one
1227 : * checkpoint can be in progress at a time.
1228 : */
1229 90 : if (fd < 0)
1230 0 : ereport(ERROR,
1231 : (errcode_for_file_access(),
1232 : errmsg("could not open file \"%s\": %m", path)));
1233 :
1234 : /*
1235 : * We could try to avoid fsyncing files that either haven't
1236 : * changed or have only been created since the checkpoint's start,
1237 : * but it's currently not deemed worth the effort.
1238 : */
1239 90 : pgstat_report_wait_start(WAIT_EVENT_LOGICAL_REWRITE_CHECKPOINT_SYNC);
1240 90 : if (pg_fsync(fd) != 0)
1241 0 : ereport(data_sync_elevel(ERROR),
1242 : (errcode_for_file_access(),
1243 : errmsg("could not fsync file \"%s\": %m", path)));
1244 90 : pgstat_report_wait_end();
1245 :
1246 90 : if (CloseTransientFile(fd) != 0)
1247 0 : ereport(ERROR,
1248 : (errcode_for_file_access(),
1249 : errmsg("could not close file \"%s\": %m", path)));
1250 : }
1251 : }
1252 1809 : FreeDir(mappings_dir);
1253 :
1254 : /* persist directory entries to disk */
1255 1809 : fsync_fname(PG_LOGICAL_MAPPINGS_DIR, true);
1256 1809 : }
|
