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 :
126 : /*
127 : * State associated with a rewrite operation. This is opaque to the user
128 : * of the rewrite facility.
129 : */
130 : typedef struct RewriteStateData
131 : {
132 : Relation rs_old_rel; /* source heap */
133 : Relation rs_new_rel; /* destination heap */
134 : BulkWriteState *rs_bulkstate; /* writer for the destination */
135 : BulkWriteBuffer rs_buffer; /* page currently being built */
136 : BlockNumber rs_blockno; /* block where page will go */
137 : bool rs_logical_rewrite; /* do we need to do logical rewriting */
138 : TransactionId rs_oldest_xmin; /* oldest xmin used by caller to determine
139 : * tuple visibility */
140 : TransactionId rs_freeze_xid; /* Xid that will be used as freeze cutoff
141 : * point */
142 : TransactionId rs_logical_xmin; /* Xid that will be used as cutoff point
143 : * for logical rewrites */
144 : MultiXactId rs_cutoff_multi; /* MultiXactId that will be used as cutoff
145 : * point for multixacts */
146 : MemoryContext rs_cxt; /* for hash tables and entries and tuples in
147 : * them */
148 : XLogRecPtr rs_begin_lsn; /* XLogInsertLsn when starting the rewrite */
149 : HTAB *rs_unresolved_tups; /* unmatched A tuples */
150 : HTAB *rs_old_new_tid_map; /* unmatched B tuples */
151 : HTAB *rs_logical_mappings; /* logical remapping files */
152 : uint32 rs_num_rewrite_mappings; /* # in memory mappings */
153 : } RewriteStateData;
154 :
155 : /*
156 : * The lookup keys for the hash tables are tuple TID and xmin (we must check
157 : * both to avoid false matches from dead tuples). Beware that there is
158 : * probably some padding space in this struct; it must be zeroed out for
159 : * correct hashtable operation.
160 : */
161 : typedef struct
162 : {
163 : TransactionId xmin; /* tuple xmin */
164 : ItemPointerData tid; /* tuple location in old heap */
165 : } TidHashKey;
166 :
167 : /*
168 : * Entry structures for the hash tables
169 : */
170 : typedef struct
171 : {
172 : TidHashKey key; /* expected xmin/old location of B tuple */
173 : ItemPointerData old_tid; /* A's location in the old heap */
174 : HeapTuple tuple; /* A's tuple contents */
175 : } UnresolvedTupData;
176 :
177 : typedef UnresolvedTupData *UnresolvedTup;
178 :
179 : typedef struct
180 : {
181 : TidHashKey key; /* actual xmin/old location of B tuple */
182 : ItemPointerData new_tid; /* where we put it in the new heap */
183 : } OldToNewMappingData;
184 :
185 : typedef OldToNewMappingData *OldToNewMapping;
186 :
187 : /*
188 : * In-Memory data for an xid that might need logical remapping entries
189 : * to be logged.
190 : */
191 : typedef struct RewriteMappingFile
192 : {
193 : TransactionId xid; /* xid that might need to see the row */
194 : int vfd; /* fd of mappings file */
195 : off_t off; /* how far have we written yet */
196 : dclist_head mappings; /* list of in-memory mappings */
197 : char path[MAXPGPATH]; /* path, for error messages */
198 : } RewriteMappingFile;
199 :
200 : /*
201 : * A single In-Memory logical rewrite mapping, hanging off
202 : * RewriteMappingFile->mappings.
203 : */
204 : typedef struct RewriteMappingDataEntry
205 : {
206 : LogicalRewriteMappingData map; /* map between old and new location of the
207 : * tuple */
208 : dlist_node node;
209 : } RewriteMappingDataEntry;
210 :
211 :
212 : /* prototypes for internal functions */
213 : static void raw_heap_insert(RewriteState state, HeapTuple tup);
214 :
215 : /* internal logical remapping prototypes */
216 : static void logical_begin_heap_rewrite(RewriteState state);
217 : static void logical_rewrite_heap_tuple(RewriteState state, ItemPointerData old_tid, HeapTuple new_tuple);
218 : static void logical_end_heap_rewrite(RewriteState state);
219 :
220 :
221 : /*
222 : * Begin a rewrite of a table
223 : *
224 : * old_heap old, locked heap relation tuples will be read from
225 : * new_heap new, locked heap relation to insert tuples to
226 : * oldest_xmin xid used by the caller to determine which tuples are dead
227 : * freeze_xid xid before which tuples will be frozen
228 : * cutoff_multi multixact before which multis will be removed
229 : *
230 : * Returns an opaque RewriteState, allocated in current memory context,
231 : * to be used in subsequent calls to the other functions.
232 : */
233 : RewriteState
234 285 : begin_heap_rewrite(Relation old_heap, Relation new_heap, TransactionId oldest_xmin,
235 : TransactionId freeze_xid, MultiXactId cutoff_multi)
236 : {
237 : RewriteState state;
238 : MemoryContext rw_cxt;
239 : MemoryContext old_cxt;
240 : HASHCTL hash_ctl;
241 :
242 : /*
243 : * To ease cleanup, make a separate context that will contain the
244 : * RewriteState struct itself plus all subsidiary data.
245 : */
246 285 : rw_cxt = AllocSetContextCreate(CurrentMemoryContext,
247 : "Table rewrite",
248 : ALLOCSET_DEFAULT_SIZES);
249 285 : old_cxt = MemoryContextSwitchTo(rw_cxt);
250 :
251 : /* Create and fill in the state struct */
252 285 : state = palloc0_object(RewriteStateData);
253 :
254 285 : state->rs_old_rel = old_heap;
255 285 : state->rs_new_rel = new_heap;
256 285 : state->rs_buffer = NULL;
257 : /* new_heap needn't be empty, just locked */
258 285 : state->rs_blockno = RelationGetNumberOfBlocks(new_heap);
259 285 : state->rs_oldest_xmin = oldest_xmin;
260 285 : state->rs_freeze_xid = freeze_xid;
261 285 : state->rs_cutoff_multi = cutoff_multi;
262 285 : state->rs_cxt = rw_cxt;
263 285 : state->rs_bulkstate = smgr_bulk_start_rel(new_heap, MAIN_FORKNUM);
264 :
265 : /* Initialize hash tables used to track update chains */
266 285 : hash_ctl.keysize = sizeof(TidHashKey);
267 285 : hash_ctl.entrysize = sizeof(UnresolvedTupData);
268 285 : hash_ctl.hcxt = state->rs_cxt;
269 :
270 285 : state->rs_unresolved_tups =
271 285 : hash_create("Rewrite / Unresolved ctids",
272 : 128, /* arbitrary initial size */
273 : &hash_ctl,
274 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
275 :
276 285 : hash_ctl.entrysize = sizeof(OldToNewMappingData);
277 :
278 285 : state->rs_old_new_tid_map =
279 285 : hash_create("Rewrite / Old to new tid map",
280 : 128, /* arbitrary initial size */
281 : &hash_ctl,
282 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
283 :
284 285 : MemoryContextSwitchTo(old_cxt);
285 :
286 285 : logical_begin_heap_rewrite(state);
287 :
288 285 : return state;
289 : }
290 :
291 : /*
292 : * End a rewrite.
293 : *
294 : * state and any other resources are freed.
295 : */
296 : void
297 285 : end_heap_rewrite(RewriteState state)
298 : {
299 : HASH_SEQ_STATUS seq_status;
300 : UnresolvedTup unresolved;
301 :
302 : /*
303 : * Write any remaining tuples in the UnresolvedTups table. If we have any
304 : * left, they should in fact be dead, but let's err on the safe side.
305 : */
306 285 : hash_seq_init(&seq_status, state->rs_unresolved_tups);
307 :
308 285 : while ((unresolved = hash_seq_search(&seq_status)) != NULL)
309 : {
310 0 : ItemPointerSetInvalid(&unresolved->tuple->t_data->t_ctid);
311 0 : raw_heap_insert(state, unresolved->tuple);
312 : }
313 :
314 : /* Write the last page, if any */
315 285 : if (state->rs_buffer)
316 : {
317 199 : smgr_bulk_write(state->rs_bulkstate, state->rs_blockno, state->rs_buffer, true);
318 199 : state->rs_buffer = NULL;
319 : }
320 :
321 285 : smgr_bulk_finish(state->rs_bulkstate);
322 :
323 285 : logical_end_heap_rewrite(state);
324 :
325 : /* Deleting the context frees everything */
326 285 : MemoryContextDelete(state->rs_cxt);
327 285 : }
328 :
329 : /*
330 : * Add a tuple to the new heap.
331 : *
332 : * Visibility information is copied from the original tuple, except that
333 : * we "freeze" very-old tuples. Note that since we scribble on new_tuple,
334 : * it had better be temp storage not a pointer to the original tuple.
335 : *
336 : * state opaque state as returned by begin_heap_rewrite
337 : * old_tuple original tuple in the old heap
338 : * new_tuple new, rewritten tuple to be inserted to new heap
339 : */
340 : void
341 362454 : rewrite_heap_tuple(RewriteState state,
342 : HeapTuple old_tuple, HeapTuple new_tuple)
343 : {
344 : MemoryContext old_cxt;
345 : ItemPointerData old_tid;
346 : TidHashKey hashkey;
347 : bool found;
348 : bool free_new;
349 :
350 362454 : old_cxt = MemoryContextSwitchTo(state->rs_cxt);
351 :
352 : /*
353 : * Copy the original tuple's visibility information into new_tuple.
354 : *
355 : * XXX we might later need to copy some t_infomask2 bits, too? Right now,
356 : * we intentionally clear the HOT status bits.
357 : */
358 362454 : memcpy(&new_tuple->t_data->t_choice.t_heap,
359 362454 : &old_tuple->t_data->t_choice.t_heap,
360 : sizeof(HeapTupleFields));
361 :
362 362454 : new_tuple->t_data->t_infomask &= ~HEAP_XACT_MASK;
363 362454 : new_tuple->t_data->t_infomask2 &= ~HEAP2_XACT_MASK;
364 362454 : new_tuple->t_data->t_infomask |=
365 362454 : old_tuple->t_data->t_infomask & HEAP_XACT_MASK;
366 :
367 : /*
368 : * While we have our hands on the tuple, we may as well freeze any
369 : * eligible xmin or xmax, so that future VACUUM effort can be saved.
370 : */
371 362454 : heap_freeze_tuple(new_tuple->t_data,
372 362454 : state->rs_old_rel->rd_rel->relfrozenxid,
373 362454 : state->rs_old_rel->rd_rel->relminmxid,
374 : state->rs_freeze_xid,
375 : state->rs_cutoff_multi);
376 :
377 : /*
378 : * Invalid ctid means that ctid should point to the tuple itself. We'll
379 : * override it later if the tuple is part of an update chain.
380 : */
381 362454 : ItemPointerSetInvalid(&new_tuple->t_data->t_ctid);
382 :
383 : /*
384 : * If the tuple has been updated, check the old-to-new mapping hash table.
385 : *
386 : * Note that this check relies on the HeapTupleSatisfiesVacuum() in
387 : * heapam_relation_copy_for_cluster() to have set hint bits.
388 : */
389 377397 : if (!((old_tuple->t_data->t_infomask & HEAP_XMAX_INVALID) ||
390 14943 : HeapTupleHeaderIsOnlyLocked(old_tuple->t_data)) &&
391 14943 : !HeapTupleHeaderIndicatesMovedPartitions(old_tuple->t_data) &&
392 14943 : !(ItemPointerEquals(&(old_tuple->t_self),
393 14943 : &(old_tuple->t_data->t_ctid))))
394 : {
395 : OldToNewMapping mapping;
396 :
397 423 : memset(&hashkey, 0, sizeof(hashkey));
398 423 : hashkey.xmin = HeapTupleHeaderGetUpdateXid(old_tuple->t_data);
399 423 : hashkey.tid = old_tuple->t_data->t_ctid;
400 :
401 : mapping = (OldToNewMapping)
402 423 : hash_search(state->rs_old_new_tid_map, &hashkey,
403 : HASH_FIND, NULL);
404 :
405 423 : if (mapping != NULL)
406 : {
407 : /*
408 : * We've already copied the tuple that t_ctid points to, so we can
409 : * set the ctid of this tuple to point to the new location, and
410 : * insert it right away.
411 : */
412 192 : new_tuple->t_data->t_ctid = mapping->new_tid;
413 :
414 : /* We don't need the mapping entry anymore */
415 192 : hash_search(state->rs_old_new_tid_map, &hashkey,
416 : HASH_REMOVE, &found);
417 : Assert(found);
418 : }
419 : else
420 : {
421 : /*
422 : * We haven't seen the tuple t_ctid points to yet. Stash this
423 : * tuple into unresolved_tups to be written later.
424 : */
425 : UnresolvedTup unresolved;
426 :
427 231 : unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
428 : HASH_ENTER, &found);
429 : Assert(!found);
430 :
431 231 : unresolved->old_tid = old_tuple->t_self;
432 231 : unresolved->tuple = heap_copytuple(new_tuple);
433 :
434 : /*
435 : * We can't do anything more now, since we don't know where the
436 : * tuple will be written.
437 : */
438 231 : MemoryContextSwitchTo(old_cxt);
439 231 : return;
440 : }
441 : }
442 :
443 : /*
444 : * Now we will write the tuple, and then check to see if it is the B tuple
445 : * in any new or known pair. When we resolve a known pair, we will be
446 : * able to write that pair's A tuple, and then we have to check if it
447 : * resolves some other pair. Hence, we need a loop here.
448 : */
449 362223 : old_tid = old_tuple->t_self;
450 362223 : free_new = false;
451 :
452 : for (;;)
453 231 : {
454 : ItemPointerData new_tid;
455 :
456 : /* Insert the tuple and find out where it's put in new_heap */
457 362454 : raw_heap_insert(state, new_tuple);
458 362454 : new_tid = new_tuple->t_self;
459 :
460 362454 : logical_rewrite_heap_tuple(state, old_tid, new_tuple);
461 :
462 : /*
463 : * If the tuple is the updated version of a row, and the prior version
464 : * wouldn't be DEAD yet, then we need to either resolve the prior
465 : * version (if it's waiting in rs_unresolved_tups), or make an entry
466 : * in rs_old_new_tid_map (so we can resolve it when we do see it). The
467 : * previous tuple's xmax would equal this one's xmin, so it's
468 : * RECENTLY_DEAD if and only if the xmin is not before OldestXmin.
469 : */
470 362454 : if ((new_tuple->t_data->t_infomask & HEAP_UPDATED) &&
471 10217 : !TransactionIdPrecedes(HeapTupleHeaderGetXmin(new_tuple->t_data),
472 : state->rs_oldest_xmin))
473 : {
474 : /*
475 : * Okay, this is B in an update pair. See if we've seen A.
476 : */
477 : UnresolvedTup unresolved;
478 :
479 423 : memset(&hashkey, 0, sizeof(hashkey));
480 423 : hashkey.xmin = HeapTupleHeaderGetXmin(new_tuple->t_data);
481 423 : hashkey.tid = old_tid;
482 :
483 423 : unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
484 : HASH_FIND, NULL);
485 :
486 423 : if (unresolved != NULL)
487 : {
488 : /*
489 : * We have seen and memorized the previous tuple already. Now
490 : * that we know where we inserted the tuple its t_ctid points
491 : * to, fix its t_ctid and insert it to the new heap.
492 : */
493 231 : if (free_new)
494 51 : heap_freetuple(new_tuple);
495 231 : new_tuple = unresolved->tuple;
496 231 : free_new = true;
497 231 : old_tid = unresolved->old_tid;
498 231 : new_tuple->t_data->t_ctid = new_tid;
499 :
500 : /*
501 : * We don't need the hash entry anymore, but don't free its
502 : * tuple just yet.
503 : */
504 231 : hash_search(state->rs_unresolved_tups, &hashkey,
505 : HASH_REMOVE, &found);
506 : Assert(found);
507 :
508 : /* loop back to insert the previous tuple in the chain */
509 231 : continue;
510 : }
511 : else
512 : {
513 : /*
514 : * Remember the new tid of this tuple. We'll use it to set the
515 : * ctid when we find the previous tuple in the chain.
516 : */
517 : OldToNewMapping mapping;
518 :
519 192 : mapping = hash_search(state->rs_old_new_tid_map, &hashkey,
520 : HASH_ENTER, &found);
521 : Assert(!found);
522 :
523 192 : mapping->new_tid = new_tid;
524 : }
525 : }
526 :
527 : /* Done with this (chain of) tuples, for now */
528 362223 : if (free_new)
529 180 : heap_freetuple(new_tuple);
530 362223 : break;
531 : }
532 :
533 362223 : MemoryContextSwitchTo(old_cxt);
534 : }
535 :
536 : /*
537 : * Register a dead tuple with an ongoing rewrite. Dead tuples are not
538 : * copied to the new table, but we still make note of them so that we
539 : * can release some resources earlier.
540 : *
541 : * Returns true if a tuple was removed from the unresolved_tups table.
542 : * This indicates that that tuple, previously thought to be "recently dead",
543 : * is now known really dead and won't be written to the output.
544 : */
545 : bool
546 16750 : rewrite_heap_dead_tuple(RewriteState state, HeapTuple old_tuple)
547 : {
548 : /*
549 : * If we have already seen an earlier tuple in the update chain that
550 : * points to this tuple, let's forget about that earlier tuple. It's in
551 : * fact dead as well, our simple xmax < OldestXmin test in
552 : * HeapTupleSatisfiesVacuum just wasn't enough to detect it. It happens
553 : * when xmin of a tuple is greater than xmax, which sounds
554 : * counter-intuitive but is perfectly valid.
555 : *
556 : * We don't bother to try to detect the situation the other way round,
557 : * when we encounter the dead tuple first and then the recently dead one
558 : * that points to it. If that happens, we'll have some unmatched entries
559 : * in the UnresolvedTups hash table at the end. That can happen anyway,
560 : * because a vacuum might have removed the dead tuple in the chain before
561 : * us.
562 : */
563 : UnresolvedTup unresolved;
564 : TidHashKey hashkey;
565 : bool found;
566 :
567 16750 : memset(&hashkey, 0, sizeof(hashkey));
568 16750 : hashkey.xmin = HeapTupleHeaderGetXmin(old_tuple->t_data);
569 16750 : hashkey.tid = old_tuple->t_self;
570 :
571 16750 : unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
572 : HASH_FIND, NULL);
573 :
574 16750 : if (unresolved != NULL)
575 : {
576 : /* Need to free the contained tuple as well as the hashtable entry */
577 0 : heap_freetuple(unresolved->tuple);
578 0 : hash_search(state->rs_unresolved_tups, &hashkey,
579 : HASH_REMOVE, &found);
580 : Assert(found);
581 0 : return true;
582 : }
583 :
584 16750 : return false;
585 : }
586 :
587 : /*
588 : * Insert a tuple to the new relation. This has to track heap_insert
589 : * and its subsidiary functions!
590 : *
591 : * t_self of the tuple is set to the new TID of the tuple. If t_ctid of the
592 : * tuple is invalid on entry, it's replaced with the new TID as well (in
593 : * the inserted data only, not in the caller's copy).
594 : */
595 : static void
596 362454 : raw_heap_insert(RewriteState state, HeapTuple tup)
597 : {
598 : Page page;
599 : Size pageFreeSpace,
600 : saveFreeSpace;
601 : Size len;
602 : OffsetNumber newoff;
603 : HeapTuple heaptup;
604 :
605 : /*
606 : * If the new tuple is too big for storage or contains already toasted
607 : * out-of-line attributes from some other relation, invoke the toaster.
608 : *
609 : * Note: below this point, heaptup is the data we actually intend to store
610 : * into the relation; tup is the caller's original untoasted data.
611 : */
612 362454 : if (state->rs_new_rel->rd_rel->relkind == RELKIND_TOASTVALUE)
613 : {
614 : /* toast table entries should never be recursively toasted */
615 : Assert(!HeapTupleHasExternal(tup));
616 0 : heaptup = tup;
617 : }
618 362454 : else if (HeapTupleHasExternal(tup) || tup->t_len > TOAST_TUPLE_THRESHOLD)
619 299 : {
620 299 : int options = HEAP_INSERT_SKIP_FSM;
621 :
622 : /*
623 : * While rewriting the heap for VACUUM FULL / CLUSTER, make sure data
624 : * for the TOAST table are not logically decoded. The main heap is
625 : * WAL-logged as XLOG FPI records, which are not logically decoded.
626 : */
627 299 : options |= HEAP_INSERT_NO_LOGICAL;
628 :
629 299 : heaptup = heap_toast_insert_or_update(state->rs_new_rel, tup, NULL,
630 : options);
631 : }
632 : else
633 362155 : heaptup = tup;
634 :
635 362454 : len = MAXALIGN(heaptup->t_len); /* be conservative */
636 :
637 : /*
638 : * If we're gonna fail for oversize tuple, do it right away
639 : */
640 362454 : if (len > MaxHeapTupleSize)
641 0 : ereport(ERROR,
642 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
643 : errmsg("row is too big: size %zu, maximum size %zu",
644 : len, MaxHeapTupleSize)));
645 :
646 : /* Compute desired extra freespace due to fillfactor option */
647 362454 : saveFreeSpace = RelationGetTargetPageFreeSpace(state->rs_new_rel,
648 : HEAP_DEFAULT_FILLFACTOR);
649 :
650 : /* Now we can check to see if there's enough free space already. */
651 362454 : page = (Page) state->rs_buffer;
652 362454 : if (page)
653 : {
654 362255 : pageFreeSpace = PageGetHeapFreeSpace(page);
655 :
656 362255 : if (len + saveFreeSpace > pageFreeSpace)
657 : {
658 : /*
659 : * Doesn't fit, so write out the existing page. It always
660 : * contains a tuple. Hence, unlike RelationGetBufferForTuple(),
661 : * enforce saveFreeSpace unconditionally.
662 : */
663 5157 : smgr_bulk_write(state->rs_bulkstate, state->rs_blockno, state->rs_buffer, true);
664 5157 : state->rs_buffer = NULL;
665 5157 : page = NULL;
666 5157 : state->rs_blockno++;
667 : }
668 : }
669 :
670 362454 : if (!page)
671 : {
672 : /* Initialize a new empty page */
673 5356 : state->rs_buffer = smgr_bulk_get_buf(state->rs_bulkstate);
674 5356 : page = (Page) state->rs_buffer;
675 5356 : PageInit(page, BLCKSZ, 0);
676 : }
677 :
678 : /* And now we can insert the tuple into the page */
679 362454 : newoff = PageAddItem(page, heaptup->t_data, heaptup->t_len, InvalidOffsetNumber, false, true);
680 362454 : if (newoff == InvalidOffsetNumber)
681 0 : elog(ERROR, "failed to add tuple");
682 :
683 : /* Update caller's t_self to the actual position where it was stored */
684 362454 : ItemPointerSet(&(tup->t_self), state->rs_blockno, newoff);
685 :
686 : /*
687 : * Insert the correct position into CTID of the stored tuple, too, if the
688 : * caller didn't supply a valid CTID.
689 : */
690 362454 : if (!ItemPointerIsValid(&tup->t_data->t_ctid))
691 : {
692 : ItemId newitemid;
693 : HeapTupleHeader onpage_tup;
694 :
695 362031 : newitemid = PageGetItemId(page, newoff);
696 362031 : onpage_tup = (HeapTupleHeader) PageGetItem(page, newitemid);
697 :
698 362031 : onpage_tup->t_ctid = tup->t_self;
699 : }
700 :
701 : /* If heaptup is a private copy, release it. */
702 362454 : if (heaptup != tup)
703 299 : heap_freetuple(heaptup);
704 362454 : }
705 :
706 : /* ------------------------------------------------------------------------
707 : * Logical rewrite support
708 : *
709 : * When doing logical decoding - which relies on using cmin/cmax of catalog
710 : * tuples, via xl_heap_new_cid records - heap rewrites have to log enough
711 : * information to allow the decoding backend to update its internal mapping
712 : * of (relfilelocator,ctid) => (cmin, cmax) to be correct for the rewritten heap.
713 : *
714 : * For that, every time we find a tuple that's been modified in a catalog
715 : * relation within the xmin horizon of any decoding slot, we log a mapping
716 : * from the old to the new location.
717 : *
718 : * To deal with rewrites that abort the filename of a mapping file contains
719 : * the xid of the transaction performing the rewrite, which then can be
720 : * checked before being read in.
721 : *
722 : * For efficiency we don't immediately spill every single map mapping for a
723 : * row to disk but only do so in batches when we've collected several of them
724 : * in memory or when end_heap_rewrite() has been called.
725 : *
726 : * Crash-Safety: This module diverts from the usual patterns of doing WAL
727 : * since it cannot rely on checkpoint flushing out all buffers and thus
728 : * waiting for exclusive locks on buffers. Usually the XLogInsert() covering
729 : * buffer modifications is performed while the buffer(s) that are being
730 : * modified are exclusively locked guaranteeing that both the WAL record and
731 : * the modified heap are on either side of the checkpoint. But since the
732 : * mapping files we log aren't in shared_buffers that interlock doesn't work.
733 : *
734 : * Instead we simply write the mapping files out to disk, *before* the
735 : * XLogInsert() is performed. That guarantees that either the XLogInsert() is
736 : * inserted after the checkpoint's redo pointer or that the checkpoint (via
737 : * CheckPointLogicalRewriteHeap()) has flushed the (partial) mapping file to
738 : * disk. That leaves the tail end that has not yet been flushed open to
739 : * corruption, which is solved by including the current offset in the
740 : * xl_heap_rewrite_mapping records and truncating the mapping file to it
741 : * during replay. Every time a rewrite is finished all generated mapping files
742 : * are synced to disk.
743 : *
744 : * Note that if we were only concerned about crash safety we wouldn't have to
745 : * deal with WAL logging at all - an fsync() at the end of a rewrite would be
746 : * sufficient for crash safety. Any mapping that hasn't been safely flushed to
747 : * disk has to be by an aborted (explicitly or via a crash) transaction and is
748 : * ignored by virtue of the xid in its name being subject to a
749 : * TransactionDidCommit() check. But we want to support having standbys via
750 : * physical replication, both for availability and to do logical decoding
751 : * there.
752 : * ------------------------------------------------------------------------
753 : */
754 :
755 : /*
756 : * Do preparations for logging logical mappings during a rewrite if
757 : * necessary. If we detect that we don't need to log anything we'll prevent
758 : * any further action by the various logical rewrite functions.
759 : */
760 : static void
761 285 : logical_begin_heap_rewrite(RewriteState state)
762 : {
763 : HASHCTL hash_ctl;
764 : TransactionId logical_xmin;
765 :
766 : /*
767 : * We only need to persist these mappings if the rewritten table can be
768 : * accessed during logical decoding, if not, we can skip doing any
769 : * additional work.
770 : */
771 285 : state->rs_logical_rewrite =
772 285 : RelationIsAccessibleInLogicalDecoding(state->rs_old_rel);
773 :
774 285 : if (!state->rs_logical_rewrite)
775 265 : return;
776 :
777 22 : ProcArrayGetReplicationSlotXmin(NULL, &logical_xmin);
778 :
779 : /*
780 : * If there are no logical slots in progress we don't need to do anything,
781 : * there cannot be any remappings for relevant rows yet. The relation's
782 : * lock protects us against races.
783 : */
784 22 : if (logical_xmin == InvalidTransactionId)
785 : {
786 2 : state->rs_logical_rewrite = false;
787 2 : return;
788 : }
789 :
790 20 : state->rs_logical_xmin = logical_xmin;
791 20 : state->rs_begin_lsn = GetXLogInsertRecPtr();
792 20 : state->rs_num_rewrite_mappings = 0;
793 :
794 20 : hash_ctl.keysize = sizeof(TransactionId);
795 20 : hash_ctl.entrysize = sizeof(RewriteMappingFile);
796 20 : hash_ctl.hcxt = state->rs_cxt;
797 :
798 20 : state->rs_logical_mappings =
799 20 : hash_create("Logical rewrite mapping",
800 : 128, /* arbitrary initial size */
801 : &hash_ctl,
802 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
803 : }
804 :
805 : /*
806 : * Flush all logical in-memory mappings to disk, but don't fsync them yet.
807 : */
808 : static void
809 9 : logical_heap_rewrite_flush_mappings(RewriteState state)
810 : {
811 : HASH_SEQ_STATUS seq_status;
812 : RewriteMappingFile *src;
813 : dlist_mutable_iter iter;
814 :
815 : Assert(state->rs_logical_rewrite);
816 :
817 : /* no logical rewrite in progress, no need to iterate over mappings */
818 9 : if (state->rs_num_rewrite_mappings == 0)
819 0 : return;
820 :
821 9 : elog(DEBUG1, "flushing %u logical rewrite mapping entries",
822 : state->rs_num_rewrite_mappings);
823 :
824 9 : hash_seq_init(&seq_status, state->rs_logical_mappings);
825 99 : while ((src = (RewriteMappingFile *) hash_seq_search(&seq_status)) != NULL)
826 : {
827 : char *waldata;
828 : char *waldata_start;
829 : xl_heap_rewrite_mapping xlrec;
830 : Oid dboid;
831 : uint32 len;
832 : int written;
833 90 : uint32 num_mappings = dclist_count(&src->mappings);
834 :
835 : /* this file hasn't got any new mappings */
836 90 : if (num_mappings == 0)
837 0 : continue;
838 :
839 90 : if (state->rs_old_rel->rd_rel->relisshared)
840 0 : dboid = InvalidOid;
841 : else
842 90 : dboid = MyDatabaseId;
843 :
844 90 : xlrec.num_mappings = num_mappings;
845 90 : xlrec.mapped_rel = RelationGetRelid(state->rs_old_rel);
846 90 : xlrec.mapped_xid = src->xid;
847 90 : xlrec.mapped_db = dboid;
848 90 : xlrec.offset = src->off;
849 90 : xlrec.start_lsn = state->rs_begin_lsn;
850 :
851 : /* write all mappings consecutively */
852 90 : len = num_mappings * sizeof(LogicalRewriteMappingData);
853 90 : waldata_start = waldata = palloc(len);
854 :
855 : /*
856 : * collect data we need to write out, but don't modify ondisk data yet
857 : */
858 813 : dclist_foreach_modify(iter, &src->mappings)
859 : {
860 : RewriteMappingDataEntry *pmap;
861 :
862 723 : pmap = dclist_container(RewriteMappingDataEntry, node, iter.cur);
863 :
864 723 : memcpy(waldata, &pmap->map, sizeof(pmap->map));
865 723 : waldata += sizeof(pmap->map);
866 :
867 : /* remove from the list and free */
868 723 : dclist_delete_from(&src->mappings, &pmap->node);
869 723 : pfree(pmap);
870 :
871 : /* update bookkeeping */
872 723 : state->rs_num_rewrite_mappings--;
873 : }
874 :
875 : Assert(dclist_count(&src->mappings) == 0);
876 : Assert(waldata == waldata_start + len);
877 :
878 : /*
879 : * Note that we deviate from the usual WAL coding practices here,
880 : * check the above "Logical rewrite support" comment for reasoning.
881 : */
882 90 : written = FileWrite(src->vfd, waldata_start, len, src->off,
883 : WAIT_EVENT_LOGICAL_REWRITE_WRITE);
884 90 : if (written != len)
885 0 : ereport(ERROR,
886 : (errcode_for_file_access(),
887 : errmsg("could not write to file \"%s\", wrote %d of %d: %m", src->path,
888 : written, len)));
889 90 : src->off += len;
890 :
891 90 : XLogBeginInsert();
892 90 : XLogRegisterData(&xlrec, sizeof(xlrec));
893 90 : XLogRegisterData(waldata_start, len);
894 :
895 : /* write xlog record */
896 90 : XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_REWRITE);
897 :
898 90 : pfree(waldata_start);
899 : }
900 : Assert(state->rs_num_rewrite_mappings == 0);
901 : }
902 :
903 : /*
904 : * Logical remapping part of end_heap_rewrite().
905 : */
906 : static void
907 285 : logical_end_heap_rewrite(RewriteState state)
908 : {
909 : HASH_SEQ_STATUS seq_status;
910 : RewriteMappingFile *src;
911 :
912 : /* done, no logical rewrite in progress */
913 285 : if (!state->rs_logical_rewrite)
914 265 : return;
915 :
916 : /* writeout remaining in-memory entries */
917 20 : if (state->rs_num_rewrite_mappings > 0)
918 9 : logical_heap_rewrite_flush_mappings(state);
919 :
920 : /* Iterate over all mappings we have written and fsync the files. */
921 20 : hash_seq_init(&seq_status, state->rs_logical_mappings);
922 110 : while ((src = (RewriteMappingFile *) hash_seq_search(&seq_status)) != NULL)
923 : {
924 90 : if (FileSync(src->vfd, WAIT_EVENT_LOGICAL_REWRITE_SYNC) != 0)
925 0 : ereport(data_sync_elevel(ERROR),
926 : (errcode_for_file_access(),
927 : errmsg("could not fsync file \"%s\": %m", src->path)));
928 90 : FileClose(src->vfd);
929 : }
930 : /* memory context cleanup will deal with the rest */
931 : }
932 :
933 : /*
934 : * Log a single (old->new) mapping for 'xid'.
935 : */
936 : static void
937 723 : logical_rewrite_log_mapping(RewriteState state, TransactionId xid,
938 : LogicalRewriteMappingData *map)
939 : {
940 : RewriteMappingFile *src;
941 : RewriteMappingDataEntry *pmap;
942 : Oid relid;
943 : bool found;
944 :
945 723 : relid = RelationGetRelid(state->rs_old_rel);
946 :
947 : /* look for existing mappings for this 'mapped' xid */
948 723 : src = hash_search(state->rs_logical_mappings, &xid,
949 : HASH_ENTER, &found);
950 :
951 : /*
952 : * We haven't yet had the need to map anything for this xid, create
953 : * per-xid data structures.
954 : */
955 723 : if (!found)
956 : {
957 : char path[MAXPGPATH];
958 : Oid dboid;
959 :
960 90 : if (state->rs_old_rel->rd_rel->relisshared)
961 0 : dboid = InvalidOid;
962 : else
963 90 : dboid = MyDatabaseId;
964 :
965 90 : snprintf(path, MAXPGPATH,
966 : "%s/" LOGICAL_REWRITE_FORMAT,
967 : PG_LOGICAL_MAPPINGS_DIR, dboid, relid,
968 90 : LSN_FORMAT_ARGS(state->rs_begin_lsn),
969 : xid, GetCurrentTransactionId());
970 :
971 90 : dclist_init(&src->mappings);
972 90 : src->off = 0;
973 90 : memcpy(src->path, path, sizeof(path));
974 90 : src->vfd = PathNameOpenFile(path,
975 : O_CREAT | O_EXCL | O_WRONLY | PG_BINARY);
976 90 : if (src->vfd < 0)
977 0 : ereport(ERROR,
978 : (errcode_for_file_access(),
979 : errmsg("could not create file \"%s\": %m", path)));
980 : }
981 :
982 723 : pmap = MemoryContextAlloc(state->rs_cxt,
983 : sizeof(RewriteMappingDataEntry));
984 723 : memcpy(&pmap->map, map, sizeof(LogicalRewriteMappingData));
985 723 : dclist_push_tail(&src->mappings, &pmap->node);
986 723 : state->rs_num_rewrite_mappings++;
987 :
988 : /*
989 : * Write out buffer every time we've too many in-memory entries across all
990 : * mapping files.
991 : */
992 723 : if (state->rs_num_rewrite_mappings >= 1000 /* arbitrary number */ )
993 0 : logical_heap_rewrite_flush_mappings(state);
994 723 : }
995 :
996 : /*
997 : * Perform logical remapping for a tuple that's mapped from old_tid to
998 : * new_tuple->t_self by rewrite_heap_tuple() if necessary for the tuple.
999 : */
1000 : static void
1001 362454 : logical_rewrite_heap_tuple(RewriteState state, ItemPointerData old_tid,
1002 : HeapTuple new_tuple)
1003 : {
1004 362454 : ItemPointerData new_tid = new_tuple->t_self;
1005 362454 : TransactionId cutoff = state->rs_logical_xmin;
1006 : TransactionId xmin;
1007 : TransactionId xmax;
1008 362454 : bool do_log_xmin = false;
1009 362454 : bool do_log_xmax = false;
1010 : LogicalRewriteMappingData map;
1011 :
1012 : /* no logical rewrite in progress, we don't need to log anything */
1013 362454 : if (!state->rs_logical_rewrite)
1014 361746 : return;
1015 :
1016 26897 : xmin = HeapTupleHeaderGetXmin(new_tuple->t_data);
1017 : /* use *GetUpdateXid to correctly deal with multixacts */
1018 26897 : xmax = HeapTupleHeaderGetUpdateXid(new_tuple->t_data);
1019 :
1020 : /*
1021 : * Log the mapping iff the tuple has been created recently.
1022 : */
1023 26897 : if (TransactionIdIsNormal(xmin) && !TransactionIdPrecedes(xmin, cutoff))
1024 535 : do_log_xmin = true;
1025 :
1026 26897 : if (!TransactionIdIsNormal(xmax))
1027 : {
1028 : /*
1029 : * no xmax is set, can't have any permanent ones, so this check is
1030 : * sufficient
1031 : */
1032 : }
1033 501 : else if (HEAP_XMAX_IS_LOCKED_ONLY(new_tuple->t_data->t_infomask))
1034 : {
1035 : /* only locked, we don't care */
1036 : }
1037 501 : else if (!TransactionIdPrecedes(xmax, cutoff))
1038 : {
1039 : /* tuple has been deleted recently, log */
1040 501 : do_log_xmax = true;
1041 : }
1042 :
1043 : /* if neither needs to be logged, we're done */
1044 26897 : if (!do_log_xmin && !do_log_xmax)
1045 26189 : return;
1046 :
1047 : /* fill out mapping information */
1048 708 : map.old_locator = state->rs_old_rel->rd_locator;
1049 708 : map.old_tid = old_tid;
1050 708 : map.new_locator = state->rs_new_rel->rd_locator;
1051 708 : map.new_tid = new_tid;
1052 :
1053 : /* ---
1054 : * Now persist the mapping for the individual xids that are affected. We
1055 : * need to log for both xmin and xmax if they aren't the same transaction
1056 : * since the mapping files are per "affected" xid.
1057 : * We don't muster all that much effort detecting whether xmin and xmax
1058 : * are actually the same transaction, we just check whether the xid is the
1059 : * same disregarding subtransactions. Logging too much is relatively
1060 : * harmless and we could never do the check fully since subtransaction
1061 : * data is thrown away during restarts.
1062 : * ---
1063 : */
1064 708 : if (do_log_xmin)
1065 535 : logical_rewrite_log_mapping(state, xmin, &map);
1066 : /* separately log mapping for xmax unless it'd be redundant */
1067 708 : if (do_log_xmax && !TransactionIdEquals(xmin, xmax))
1068 188 : logical_rewrite_log_mapping(state, xmax, &map);
1069 : }
1070 :
1071 : /*
1072 : * Replay XLOG_HEAP2_REWRITE records
1073 : */
1074 : void
1075 0 : heap_xlog_logical_rewrite(XLogReaderState *r)
1076 : {
1077 : char path[MAXPGPATH];
1078 : int fd;
1079 : xl_heap_rewrite_mapping *xlrec;
1080 : uint32 len;
1081 : char *data;
1082 :
1083 0 : xlrec = (xl_heap_rewrite_mapping *) XLogRecGetData(r);
1084 :
1085 0 : snprintf(path, MAXPGPATH,
1086 : "%s/" LOGICAL_REWRITE_FORMAT,
1087 : PG_LOGICAL_MAPPINGS_DIR, xlrec->mapped_db, xlrec->mapped_rel,
1088 0 : LSN_FORMAT_ARGS(xlrec->start_lsn),
1089 0 : xlrec->mapped_xid, XLogRecGetXid(r));
1090 :
1091 0 : fd = OpenTransientFile(path,
1092 : O_CREAT | O_WRONLY | PG_BINARY);
1093 0 : if (fd < 0)
1094 0 : ereport(ERROR,
1095 : (errcode_for_file_access(),
1096 : errmsg("could not create file \"%s\": %m", path)));
1097 :
1098 : /*
1099 : * Truncate all data that's not guaranteed to have been safely fsynced (by
1100 : * previous record or by the last checkpoint).
1101 : */
1102 0 : pgstat_report_wait_start(WAIT_EVENT_LOGICAL_REWRITE_TRUNCATE);
1103 0 : if (ftruncate(fd, xlrec->offset) != 0)
1104 0 : ereport(ERROR,
1105 : (errcode_for_file_access(),
1106 : errmsg("could not truncate file \"%s\" to %u: %m",
1107 : path, (uint32) xlrec->offset)));
1108 0 : pgstat_report_wait_end();
1109 :
1110 0 : data = XLogRecGetData(r) + sizeof(*xlrec);
1111 :
1112 0 : len = xlrec->num_mappings * sizeof(LogicalRewriteMappingData);
1113 :
1114 : /* write out tail end of mapping file (again) */
1115 0 : errno = 0;
1116 0 : pgstat_report_wait_start(WAIT_EVENT_LOGICAL_REWRITE_MAPPING_WRITE);
1117 0 : if (pg_pwrite(fd, data, len, xlrec->offset) != len)
1118 : {
1119 : /* if write didn't set errno, assume problem is no disk space */
1120 0 : if (errno == 0)
1121 0 : errno = ENOSPC;
1122 0 : ereport(ERROR,
1123 : (errcode_for_file_access(),
1124 : errmsg("could not write to file \"%s\": %m", path)));
1125 : }
1126 0 : pgstat_report_wait_end();
1127 :
1128 : /*
1129 : * Now fsync all previously written data. We could improve things and only
1130 : * do this for the last write to a file, but the required bookkeeping
1131 : * doesn't seem worth the trouble.
1132 : */
1133 0 : pgstat_report_wait_start(WAIT_EVENT_LOGICAL_REWRITE_MAPPING_SYNC);
1134 0 : if (pg_fsync(fd) != 0)
1135 0 : ereport(data_sync_elevel(ERROR),
1136 : (errcode_for_file_access(),
1137 : errmsg("could not fsync file \"%s\": %m", path)));
1138 0 : pgstat_report_wait_end();
1139 :
1140 0 : if (CloseTransientFile(fd) != 0)
1141 0 : ereport(ERROR,
1142 : (errcode_for_file_access(),
1143 : errmsg("could not close file \"%s\": %m", path)));
1144 0 : }
1145 :
1146 : /* ---
1147 : * Perform a checkpoint for logical rewrite mappings
1148 : *
1149 : * This serves two tasks:
1150 : * 1) Remove all mappings not needed anymore based on the logical restart LSN
1151 : * 2) Flush all remaining mappings to disk, so that replay after a checkpoint
1152 : * only has to deal with the parts of a mapping that have been written out
1153 : * after the checkpoint started.
1154 : * ---
1155 : */
1156 : void
1157 1793 : CheckPointLogicalRewriteHeap(void)
1158 : {
1159 : XLogRecPtr cutoff;
1160 : XLogRecPtr redo;
1161 : DIR *mappings_dir;
1162 : struct dirent *mapping_de;
1163 : char path[MAXPGPATH + sizeof(PG_LOGICAL_MAPPINGS_DIR)];
1164 :
1165 : /*
1166 : * We start of with a minimum of the last redo pointer. No new decoding
1167 : * slot will start before that, so that's a safe upper bound for removal.
1168 : */
1169 1793 : redo = GetRedoRecPtr();
1170 :
1171 : /* now check for the restart ptrs from existing slots */
1172 1793 : cutoff = ReplicationSlotsComputeLogicalRestartLSN();
1173 :
1174 : /* don't start earlier than the restart lsn */
1175 1793 : if (XLogRecPtrIsValid(cutoff) && redo < cutoff)
1176 1 : cutoff = redo;
1177 :
1178 1793 : mappings_dir = AllocateDir(PG_LOGICAL_MAPPINGS_DIR);
1179 5559 : while ((mapping_de = ReadDir(mappings_dir, PG_LOGICAL_MAPPINGS_DIR)) != NULL)
1180 : {
1181 : Oid dboid;
1182 : Oid relid;
1183 : XLogRecPtr lsn;
1184 : TransactionId rewrite_xid;
1185 : TransactionId create_xid;
1186 : uint32 hi,
1187 : lo;
1188 : PGFileType de_type;
1189 :
1190 3766 : if (strcmp(mapping_de->d_name, ".") == 0 ||
1191 1973 : strcmp(mapping_de->d_name, "..") == 0)
1192 3586 : continue;
1193 :
1194 180 : snprintf(path, sizeof(path), "%s/%s", PG_LOGICAL_MAPPINGS_DIR, mapping_de->d_name);
1195 180 : de_type = get_dirent_type(path, mapping_de, false, DEBUG1);
1196 :
1197 180 : if (de_type != PGFILETYPE_ERROR && de_type != PGFILETYPE_REG)
1198 0 : continue;
1199 :
1200 : /* Skip over files that cannot be ours. */
1201 180 : if (strncmp(mapping_de->d_name, "map-", 4) != 0)
1202 0 : continue;
1203 :
1204 180 : if (sscanf(mapping_de->d_name, LOGICAL_REWRITE_FORMAT,
1205 : &dboid, &relid, &hi, &lo, &rewrite_xid, &create_xid) != 6)
1206 0 : elog(ERROR, "could not parse filename \"%s\"", mapping_de->d_name);
1207 :
1208 180 : lsn = ((uint64) hi) << 32 | lo;
1209 :
1210 180 : if (lsn < cutoff || !XLogRecPtrIsValid(cutoff))
1211 : {
1212 90 : elog(DEBUG1, "removing logical rewrite file \"%s\"", path);
1213 90 : if (unlink(path) < 0)
1214 0 : ereport(ERROR,
1215 : (errcode_for_file_access(),
1216 : errmsg("could not remove file \"%s\": %m", path)));
1217 : }
1218 : else
1219 : {
1220 : /* on some operating systems fsyncing a file requires O_RDWR */
1221 90 : int fd = OpenTransientFile(path, O_RDWR | PG_BINARY);
1222 :
1223 : /*
1224 : * The file cannot vanish due to concurrency since this function
1225 : * is the only one removing logical mappings and only one
1226 : * checkpoint can be in progress at a time.
1227 : */
1228 90 : if (fd < 0)
1229 0 : ereport(ERROR,
1230 : (errcode_for_file_access(),
1231 : errmsg("could not open file \"%s\": %m", path)));
1232 :
1233 : /*
1234 : * We could try to avoid fsyncing files that either haven't
1235 : * changed or have only been created since the checkpoint's start,
1236 : * but it's currently not deemed worth the effort.
1237 : */
1238 90 : pgstat_report_wait_start(WAIT_EVENT_LOGICAL_REWRITE_CHECKPOINT_SYNC);
1239 90 : if (pg_fsync(fd) != 0)
1240 0 : ereport(data_sync_elevel(ERROR),
1241 : (errcode_for_file_access(),
1242 : errmsg("could not fsync file \"%s\": %m", path)));
1243 90 : pgstat_report_wait_end();
1244 :
1245 90 : if (CloseTransientFile(fd) != 0)
1246 0 : ereport(ERROR,
1247 : (errcode_for_file_access(),
1248 : errmsg("could not close file \"%s\": %m", path)));
1249 : }
1250 : }
1251 1793 : FreeDir(mappings_dir);
1252 :
1253 : /* persist directory entries to disk */
1254 1793 : fsync_fname(PG_LOGICAL_MAPPINGS_DIR, true);
1255 1793 : }
|
