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) 1996-2025, 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
234 568 : 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 : */
246 568 : rw_cxt = AllocSetContextCreate(CurrentMemoryContext,
247 : "Table rewrite",
248 : ALLOCSET_DEFAULT_SIZES);
249 568 : old_cxt = MemoryContextSwitchTo(rw_cxt);
250 :
251 : /* Create and fill in the state struct */
252 568 : state = palloc0(sizeof(RewriteStateData));
253 :
254 568 : state->rs_old_rel = old_heap;
255 568 : state->rs_new_rel = new_heap;
256 568 : state->rs_buffer = NULL;
257 : /* new_heap needn't be empty, just locked */
258 568 : state->rs_blockno = RelationGetNumberOfBlocks(new_heap);
259 568 : state->rs_oldest_xmin = oldest_xmin;
260 568 : state->rs_freeze_xid = freeze_xid;
261 568 : state->rs_cutoff_multi = cutoff_multi;
262 568 : state->rs_cxt = rw_cxt;
263 568 : state->rs_bulkstate = smgr_bulk_start_rel(new_heap, MAIN_FORKNUM);
264 :
265 : /* Initialize hash tables used to track update chains */
266 568 : hash_ctl.keysize = sizeof(TidHashKey);
267 568 : hash_ctl.entrysize = sizeof(UnresolvedTupData);
268 568 : hash_ctl.hcxt = state->rs_cxt;
269 :
270 568 : state->rs_unresolved_tups =
271 568 : hash_create("Rewrite / Unresolved ctids",
272 : 128, /* arbitrary initial size */
273 : &hash_ctl,
274 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
275 :
276 568 : hash_ctl.entrysize = sizeof(OldToNewMappingData);
277 :
278 568 : state->rs_old_new_tid_map =
279 568 : hash_create("Rewrite / Old to new tid map",
280 : 128, /* arbitrary initial size */
281 : &hash_ctl,
282 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
283 :
284 568 : MemoryContextSwitchTo(old_cxt);
285 :
286 568 : logical_begin_heap_rewrite(state);
287 :
288 568 : return state;
289 : }
290 :
291 : /*
292 : * End a rewrite.
293 : *
294 : * state and any other resources are freed.
295 : */
296 : void
297 568 : 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 : */
306 568 : hash_seq_init(&seq_status, state->rs_unresolved_tups);
307 :
308 568 : 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 */
315 568 : if (state->rs_buffer)
316 : {
317 398 : smgr_bulk_write(state->rs_bulkstate, state->rs_blockno, state->rs_buffer, true);
318 398 : state->rs_buffer = NULL;
319 : }
320 :
321 568 : smgr_bulk_finish(state->rs_bulkstate);
322 :
323 568 : logical_end_heap_rewrite(state);
324 :
325 : /* Deleting the context frees everything */
326 568 : MemoryContextDelete(state->rs_cxt);
327 568 : }
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 740332 : 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 740332 : 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 740332 : memcpy(&new_tuple->t_data->t_choice.t_heap,
359 740332 : &old_tuple->t_data->t_choice.t_heap,
360 : sizeof(HeapTupleFields));
361 :
362 740332 : new_tuple->t_data->t_infomask &= ~HEAP_XACT_MASK;
363 740332 : new_tuple->t_data->t_infomask2 &= ~HEAP2_XACT_MASK;
364 740332 : new_tuple->t_data->t_infomask |=
365 740332 : 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 740332 : heap_freeze_tuple(new_tuple->t_data,
372 740332 : state->rs_old_rel->rd_rel->relfrozenxid,
373 740332 : 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 740332 : 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 786744 : if (!((old_tuple->t_data->t_infomask & HEAP_XMAX_INVALID) ||
387 46412 : HeapTupleHeaderIsOnlyLocked(old_tuple->t_data)) &&
388 46412 : !HeapTupleHeaderIndicatesMovedPartitions(old_tuple->t_data) &&
389 46412 : !(ItemPointerEquals(&(old_tuple->t_self),
390 46412 : &(old_tuple->t_data->t_ctid))))
391 : {
392 : OldToNewMapping mapping;
393 :
394 962 : memset(&hashkey, 0, sizeof(hashkey));
395 962 : hashkey.xmin = HeapTupleHeaderGetUpdateXid(old_tuple->t_data);
396 962 : hashkey.tid = old_tuple->t_data->t_ctid;
397 :
398 : mapping = (OldToNewMapping)
399 962 : hash_search(state->rs_old_new_tid_map, &hashkey,
400 : HASH_FIND, NULL);
401 :
402 962 : if (mapping != NULL)
403 : {
404 : /*
405 : * We've already copied the tuple that t_ctid points to, so we can
406 : * set the ctid of this tuple to point to the new location, and
407 : * insert it right away.
408 : */
409 398 : new_tuple->t_data->t_ctid = mapping->new_tid;
410 :
411 : /* We don't need the mapping entry anymore */
412 398 : hash_search(state->rs_old_new_tid_map, &hashkey,
413 : HASH_REMOVE, &found);
414 : Assert(found);
415 : }
416 : else
417 : {
418 : /*
419 : * We haven't seen the tuple t_ctid points to yet. Stash this
420 : * tuple into unresolved_tups to be written later.
421 : */
422 : UnresolvedTup unresolved;
423 :
424 564 : unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
425 : HASH_ENTER, &found);
426 : Assert(!found);
427 :
428 564 : unresolved->old_tid = old_tuple->t_self;
429 564 : unresolved->tuple = heap_copytuple(new_tuple);
430 :
431 : /*
432 : * We can't do anything more now, since we don't know where the
433 : * tuple will be written.
434 : */
435 564 : MemoryContextSwitchTo(old_cxt);
436 564 : return;
437 : }
438 : }
439 :
440 : /*
441 : * Now we will write the tuple, and then check to see if it is the B tuple
442 : * in any new or known pair. When we resolve a known pair, we will be
443 : * able to write that pair's A tuple, and then we have to check if it
444 : * resolves some other pair. Hence, we need a loop here.
445 : */
446 739768 : old_tid = old_tuple->t_self;
447 739768 : free_new = false;
448 :
449 : for (;;)
450 564 : {
451 : ItemPointerData new_tid;
452 :
453 : /* Insert the tuple and find out where it's put in new_heap */
454 740332 : raw_heap_insert(state, new_tuple);
455 740332 : new_tid = new_tuple->t_self;
456 :
457 740332 : logical_rewrite_heap_tuple(state, old_tid, new_tuple);
458 :
459 : /*
460 : * If the tuple is the updated version of a row, and the prior version
461 : * wouldn't be DEAD yet, then we need to either resolve the prior
462 : * version (if it's waiting in rs_unresolved_tups), or make an entry
463 : * in rs_old_new_tid_map (so we can resolve it when we do see it). The
464 : * previous tuple's xmax would equal this one's xmin, so it's
465 : * RECENTLY_DEAD if and only if the xmin is not before OldestXmin.
466 : */
467 740332 : if ((new_tuple->t_data->t_infomask & HEAP_UPDATED) &&
468 20774 : !TransactionIdPrecedes(HeapTupleHeaderGetXmin(new_tuple->t_data),
469 : state->rs_oldest_xmin))
470 : {
471 : /*
472 : * Okay, this is B in an update pair. See if we've seen A.
473 : */
474 : UnresolvedTup unresolved;
475 :
476 962 : memset(&hashkey, 0, sizeof(hashkey));
477 962 : hashkey.xmin = HeapTupleHeaderGetXmin(new_tuple->t_data);
478 962 : hashkey.tid = old_tid;
479 :
480 962 : unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
481 : HASH_FIND, NULL);
482 :
483 962 : if (unresolved != NULL)
484 : {
485 : /*
486 : * We have seen and memorized the previous tuple already. Now
487 : * that we know where we inserted the tuple its t_ctid points
488 : * to, fix its t_ctid and insert it to the new heap.
489 : */
490 564 : if (free_new)
491 140 : heap_freetuple(new_tuple);
492 564 : new_tuple = unresolved->tuple;
493 564 : free_new = true;
494 564 : old_tid = unresolved->old_tid;
495 564 : new_tuple->t_data->t_ctid = new_tid;
496 :
497 : /*
498 : * We don't need the hash entry anymore, but don't free its
499 : * tuple just yet.
500 : */
501 564 : hash_search(state->rs_unresolved_tups, &hashkey,
502 : HASH_REMOVE, &found);
503 : Assert(found);
504 :
505 : /* loop back to insert the previous tuple in the chain */
506 564 : continue;
507 : }
508 : else
509 : {
510 : /*
511 : * Remember the new tid of this tuple. We'll use it to set the
512 : * ctid when we find the previous tuple in the chain.
513 : */
514 : OldToNewMapping mapping;
515 :
516 398 : mapping = hash_search(state->rs_old_new_tid_map, &hashkey,
517 : HASH_ENTER, &found);
518 : Assert(!found);
519 :
520 398 : mapping->new_tid = new_tid;
521 : }
522 : }
523 :
524 : /* Done with this (chain of) tuples, for now */
525 739768 : if (free_new)
526 424 : heap_freetuple(new_tuple);
527 739768 : break;
528 : }
529 :
530 739768 : MemoryContextSwitchTo(old_cxt);
531 : }
532 :
533 : /*
534 : * Register a dead tuple with an ongoing rewrite. Dead tuples are not
535 : * copied to the new table, but we still make note of them so that we
536 : * can release some resources earlier.
537 : *
538 : * Returns true if a tuple was removed from the unresolved_tups table.
539 : * This indicates that that tuple, previously thought to be "recently dead",
540 : * is now known really dead and won't be written to the output.
541 : */
542 : bool
543 33062 : rewrite_heap_dead_tuple(RewriteState state, HeapTuple old_tuple)
544 : {
545 : /*
546 : * If we have already seen an earlier tuple in the update chain that
547 : * points to this tuple, let's forget about that earlier tuple. It's in
548 : * fact dead as well, our simple xmax < OldestXmin test in
549 : * HeapTupleSatisfiesVacuum just wasn't enough to detect it. It happens
550 : * when xmin of a tuple is greater than xmax, which sounds
551 : * counter-intuitive but is perfectly valid.
552 : *
553 : * We don't bother to try to detect the situation the other way round,
554 : * when we encounter the dead tuple first and then the recently dead one
555 : * that points to it. If that happens, we'll have some unmatched entries
556 : * in the UnresolvedTups hash table at the end. That can happen anyway,
557 : * because a vacuum might have removed the dead tuple in the chain before
558 : * us.
559 : */
560 : UnresolvedTup unresolved;
561 : TidHashKey hashkey;
562 : bool found;
563 :
564 33062 : memset(&hashkey, 0, sizeof(hashkey));
565 33062 : hashkey.xmin = HeapTupleHeaderGetXmin(old_tuple->t_data);
566 33062 : hashkey.tid = old_tuple->t_self;
567 :
568 33062 : unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
569 : HASH_FIND, NULL);
570 :
571 33062 : if (unresolved != NULL)
572 : {
573 : /* Need to free the contained tuple as well as the hashtable entry */
574 0 : heap_freetuple(unresolved->tuple);
575 0 : hash_search(state->rs_unresolved_tups, &hashkey,
576 : HASH_REMOVE, &found);
577 : Assert(found);
578 0 : return true;
579 : }
580 :
581 33062 : return false;
582 : }
583 :
584 : /*
585 : * Insert a tuple to the new relation. This has to track heap_insert
586 : * and its subsidiary functions!
587 : *
588 : * t_self of the tuple is set to the new TID of the tuple. If t_ctid of the
589 : * tuple is invalid on entry, it's replaced with the new TID as well (in
590 : * the inserted data only, not in the caller's copy).
591 : */
592 : static void
593 740332 : raw_heap_insert(RewriteState state, HeapTuple tup)
594 : {
595 : Page page;
596 : Size pageFreeSpace,
597 : saveFreeSpace;
598 : Size len;
599 : OffsetNumber newoff;
600 : HeapTuple heaptup;
601 :
602 : /*
603 : * If the new tuple is too big for storage or contains already toasted
604 : * out-of-line attributes from some other relation, invoke the toaster.
605 : *
606 : * Note: below this point, heaptup is the data we actually intend to store
607 : * into the relation; tup is the caller's original untoasted data.
608 : */
609 740332 : if (state->rs_new_rel->rd_rel->relkind == RELKIND_TOASTVALUE)
610 : {
611 : /* toast table entries should never be recursively toasted */
612 : Assert(!HeapTupleHasExternal(tup));
613 0 : heaptup = tup;
614 : }
615 740332 : else if (HeapTupleHasExternal(tup) || tup->t_len > TOAST_TUPLE_THRESHOLD)
616 598 : {
617 598 : int options = HEAP_INSERT_SKIP_FSM;
618 :
619 : /*
620 : * While rewriting the heap for VACUUM FULL / CLUSTER, make sure data
621 : * for the TOAST table are not logically decoded. The main heap is
622 : * WAL-logged as XLOG FPI records, which are not logically decoded.
623 : */
624 598 : options |= HEAP_INSERT_NO_LOGICAL;
625 :
626 598 : heaptup = heap_toast_insert_or_update(state->rs_new_rel, tup, NULL,
627 : options);
628 : }
629 : else
630 739734 : heaptup = tup;
631 :
632 740332 : len = MAXALIGN(heaptup->t_len); /* be conservative */
633 :
634 : /*
635 : * If we're gonna fail for oversize tuple, do it right away
636 : */
637 740332 : if (len > MaxHeapTupleSize)
638 0 : ereport(ERROR,
639 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
640 : errmsg("row is too big: size %zu, maximum size %zu",
641 : len, MaxHeapTupleSize)));
642 :
643 : /* Compute desired extra freespace due to fillfactor option */
644 740332 : saveFreeSpace = RelationGetTargetPageFreeSpace(state->rs_new_rel,
645 : HEAP_DEFAULT_FILLFACTOR);
646 :
647 : /* Now we can check to see if there's enough free space already. */
648 740332 : page = (Page) state->rs_buffer;
649 740332 : if (page)
650 : {
651 739934 : pageFreeSpace = PageGetHeapFreeSpace(page);
652 :
653 739934 : if (len + saveFreeSpace > pageFreeSpace)
654 : {
655 : /*
656 : * Doesn't fit, so write out the existing page. It always
657 : * contains a tuple. Hence, unlike RelationGetBufferForTuple(),
658 : * enforce saveFreeSpace unconditionally.
659 : */
660 10380 : smgr_bulk_write(state->rs_bulkstate, state->rs_blockno, state->rs_buffer, true);
661 10380 : state->rs_buffer = NULL;
662 10380 : page = NULL;
663 10380 : state->rs_blockno++;
664 : }
665 : }
666 :
667 740332 : if (!page)
668 : {
669 : /* Initialize a new empty page */
670 10778 : state->rs_buffer = smgr_bulk_get_buf(state->rs_bulkstate);
671 10778 : page = (Page) state->rs_buffer;
672 10778 : PageInit(page, BLCKSZ, 0);
673 : }
674 :
675 : /* And now we can insert the tuple into the page */
676 740332 : newoff = PageAddItem(page, heaptup->t_data, heaptup->t_len, InvalidOffsetNumber, false, true);
677 740332 : if (newoff == InvalidOffsetNumber)
678 0 : elog(ERROR, "failed to add tuple");
679 :
680 : /* Update caller's t_self to the actual position where it was stored */
681 740332 : ItemPointerSet(&(tup->t_self), state->rs_blockno, newoff);
682 :
683 : /*
684 : * Insert the correct position into CTID of the stored tuple, too, if the
685 : * caller didn't supply a valid CTID.
686 : */
687 740332 : if (!ItemPointerIsValid(&tup->t_data->t_ctid))
688 : {
689 : ItemId newitemid;
690 : HeapTupleHeader onpage_tup;
691 :
692 739370 : newitemid = PageGetItemId(page, newoff);
693 739370 : onpage_tup = (HeapTupleHeader) PageGetItem(page, newitemid);
694 :
695 739370 : onpage_tup->t_ctid = tup->t_self;
696 : }
697 :
698 : /* If heaptup is a private copy, release it. */
699 740332 : if (heaptup != tup)
700 598 : heap_freetuple(heaptup);
701 740332 : }
702 :
703 : /* ------------------------------------------------------------------------
704 : * Logical rewrite support
705 : *
706 : * When doing logical decoding - which relies on using cmin/cmax of catalog
707 : * tuples, via xl_heap_new_cid records - heap rewrites have to log enough
708 : * information to allow the decoding backend to update its internal mapping
709 : * of (relfilelocator,ctid) => (cmin, cmax) to be correct for the rewritten heap.
710 : *
711 : * For that, every time we find a tuple that's been modified in a catalog
712 : * relation within the xmin horizon of any decoding slot, we log a mapping
713 : * from the old to the new location.
714 : *
715 : * To deal with rewrites that abort the filename of a mapping file contains
716 : * the xid of the transaction performing the rewrite, which then can be
717 : * checked before being read in.
718 : *
719 : * For efficiency we don't immediately spill every single map mapping for a
720 : * row to disk but only do so in batches when we've collected several of them
721 : * in memory or when end_heap_rewrite() has been called.
722 : *
723 : * Crash-Safety: This module diverts from the usual patterns of doing WAL
724 : * since it cannot rely on checkpoint flushing out all buffers and thus
725 : * waiting for exclusive locks on buffers. Usually the XLogInsert() covering
726 : * buffer modifications is performed while the buffer(s) that are being
727 : * modified are exclusively locked guaranteeing that both the WAL record and
728 : * the modified heap are on either side of the checkpoint. But since the
729 : * mapping files we log aren't in shared_buffers that interlock doesn't work.
730 : *
731 : * Instead we simply write the mapping files out to disk, *before* the
732 : * XLogInsert() is performed. That guarantees that either the XLogInsert() is
733 : * inserted after the checkpoint's redo pointer or that the checkpoint (via
734 : * CheckPointLogicalRewriteHeap()) has flushed the (partial) mapping file to
735 : * disk. That leaves the tail end that has not yet been flushed open to
736 : * corruption, which is solved by including the current offset in the
737 : * xl_heap_rewrite_mapping records and truncating the mapping file to it
738 : * during replay. Every time a rewrite is finished all generated mapping files
739 : * are synced to disk.
740 : *
741 : * Note that if we were only concerned about crash safety we wouldn't have to
742 : * deal with WAL logging at all - an fsync() at the end of a rewrite would be
743 : * sufficient for crash safety. Any mapping that hasn't been safely flushed to
744 : * disk has to be by an aborted (explicitly or via a crash) transaction and is
745 : * ignored by virtue of the xid in its name being subject to a
746 : * TransactionDidCommit() check. But we want to support having standbys via
747 : * physical replication, both for availability and to do logical decoding
748 : * there.
749 : * ------------------------------------------------------------------------
750 : */
751 :
752 : /*
753 : * Do preparations for logging logical mappings during a rewrite if
754 : * necessary. If we detect that we don't need to log anything we'll prevent
755 : * any further action by the various logical rewrite functions.
756 : */
757 : static void
758 568 : logical_begin_heap_rewrite(RewriteState state)
759 : {
760 : HASHCTL hash_ctl;
761 : TransactionId logical_xmin;
762 :
763 : /*
764 : * We only need to persist these mappings if the rewritten table can be
765 : * accessed during logical decoding, if not, we can skip doing any
766 : * additional work.
767 : */
768 568 : state->rs_logical_rewrite =
769 568 : RelationIsAccessibleInLogicalDecoding(state->rs_old_rel);
770 :
771 568 : if (!state->rs_logical_rewrite)
772 528 : return;
773 :
774 44 : ProcArrayGetReplicationSlotXmin(NULL, &logical_xmin);
775 :
776 : /*
777 : * If there are no logical slots in progress we don't need to do anything,
778 : * there cannot be any remappings for relevant rows yet. The relation's
779 : * lock protects us against races.
780 : */
781 44 : if (logical_xmin == InvalidTransactionId)
782 : {
783 4 : state->rs_logical_rewrite = false;
784 4 : return;
785 : }
786 :
787 40 : state->rs_logical_xmin = logical_xmin;
788 40 : state->rs_begin_lsn = GetXLogInsertRecPtr();
789 40 : state->rs_num_rewrite_mappings = 0;
790 :
791 40 : hash_ctl.keysize = sizeof(TransactionId);
792 40 : hash_ctl.entrysize = sizeof(RewriteMappingFile);
793 40 : hash_ctl.hcxt = state->rs_cxt;
794 :
795 40 : state->rs_logical_mappings =
796 40 : hash_create("Logical rewrite mapping",
797 : 128, /* arbitrary initial size */
798 : &hash_ctl,
799 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
800 : }
801 :
802 : /*
803 : * Flush all logical in-memory mappings to disk, but don't fsync them yet.
804 : */
805 : static void
806 18 : logical_heap_rewrite_flush_mappings(RewriteState state)
807 : {
808 : HASH_SEQ_STATUS seq_status;
809 : RewriteMappingFile *src;
810 : dlist_mutable_iter iter;
811 :
812 : Assert(state->rs_logical_rewrite);
813 :
814 : /* no logical rewrite in progress, no need to iterate over mappings */
815 18 : if (state->rs_num_rewrite_mappings == 0)
816 0 : return;
817 :
818 18 : elog(DEBUG1, "flushing %u logical rewrite mapping entries",
819 : state->rs_num_rewrite_mappings);
820 :
821 18 : hash_seq_init(&seq_status, state->rs_logical_mappings);
822 198 : while ((src = (RewriteMappingFile *) hash_seq_search(&seq_status)) != NULL)
823 : {
824 : char *waldata;
825 : char *waldata_start;
826 : xl_heap_rewrite_mapping xlrec;
827 : Oid dboid;
828 : uint32 len;
829 : int written;
830 180 : uint32 num_mappings = dclist_count(&src->mappings);
831 :
832 : /* this file hasn't got any new mappings */
833 180 : if (num_mappings == 0)
834 0 : continue;
835 :
836 180 : if (state->rs_old_rel->rd_rel->relisshared)
837 0 : dboid = InvalidOid;
838 : else
839 180 : dboid = MyDatabaseId;
840 :
841 180 : xlrec.num_mappings = num_mappings;
842 180 : xlrec.mapped_rel = RelationGetRelid(state->rs_old_rel);
843 180 : xlrec.mapped_xid = src->xid;
844 180 : xlrec.mapped_db = dboid;
845 180 : xlrec.offset = src->off;
846 180 : xlrec.start_lsn = state->rs_begin_lsn;
847 :
848 : /* write all mappings consecutively */
849 180 : len = num_mappings * sizeof(LogicalRewriteMappingData);
850 180 : waldata_start = waldata = palloc(len);
851 :
852 : /*
853 : * collect data we need to write out, but don't modify ondisk data yet
854 : */
855 1626 : dclist_foreach_modify(iter, &src->mappings)
856 : {
857 : RewriteMappingDataEntry *pmap;
858 :
859 1446 : pmap = dclist_container(RewriteMappingDataEntry, node, iter.cur);
860 :
861 1446 : memcpy(waldata, &pmap->map, sizeof(pmap->map));
862 1446 : waldata += sizeof(pmap->map);
863 :
864 : /* remove from the list and free */
865 1446 : dclist_delete_from(&src->mappings, &pmap->node);
866 1446 : pfree(pmap);
867 :
868 : /* update bookkeeping */
869 1446 : state->rs_num_rewrite_mappings--;
870 : }
871 :
872 : Assert(dclist_count(&src->mappings) == 0);
873 : Assert(waldata == waldata_start + len);
874 :
875 : /*
876 : * Note that we deviate from the usual WAL coding practices here,
877 : * check the above "Logical rewrite support" comment for reasoning.
878 : */
879 180 : written = FileWrite(src->vfd, waldata_start, len, src->off,
880 : WAIT_EVENT_LOGICAL_REWRITE_WRITE);
881 180 : if (written != len)
882 0 : ereport(ERROR,
883 : (errcode_for_file_access(),
884 : errmsg("could not write to file \"%s\", wrote %d of %d: %m", src->path,
885 : written, len)));
886 180 : src->off += len;
887 :
888 180 : XLogBeginInsert();
889 180 : XLogRegisterData(&xlrec, sizeof(xlrec));
890 180 : XLogRegisterData(waldata_start, len);
891 :
892 : /* write xlog record */
893 180 : XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_REWRITE);
894 :
895 180 : pfree(waldata_start);
896 : }
897 : Assert(state->rs_num_rewrite_mappings == 0);
898 : }
899 :
900 : /*
901 : * Logical remapping part of end_heap_rewrite().
902 : */
903 : static void
904 568 : logical_end_heap_rewrite(RewriteState state)
905 : {
906 : HASH_SEQ_STATUS seq_status;
907 : RewriteMappingFile *src;
908 :
909 : /* done, no logical rewrite in progress */
910 568 : if (!state->rs_logical_rewrite)
911 528 : return;
912 :
913 : /* writeout remaining in-memory entries */
914 40 : if (state->rs_num_rewrite_mappings > 0)
915 18 : logical_heap_rewrite_flush_mappings(state);
916 :
917 : /* Iterate over all mappings we have written and fsync the files. */
918 40 : hash_seq_init(&seq_status, state->rs_logical_mappings);
919 220 : while ((src = (RewriteMappingFile *) hash_seq_search(&seq_status)) != NULL)
920 : {
921 180 : if (FileSync(src->vfd, WAIT_EVENT_LOGICAL_REWRITE_SYNC) != 0)
922 0 : ereport(data_sync_elevel(ERROR),
923 : (errcode_for_file_access(),
924 : errmsg("could not fsync file \"%s\": %m", src->path)));
925 180 : FileClose(src->vfd);
926 : }
927 : /* memory context cleanup will deal with the rest */
928 : }
929 :
930 : /*
931 : * Log a single (old->new) mapping for 'xid'.
932 : */
933 : static void
934 1446 : logical_rewrite_log_mapping(RewriteState state, TransactionId xid,
935 : LogicalRewriteMappingData *map)
936 : {
937 : RewriteMappingFile *src;
938 : RewriteMappingDataEntry *pmap;
939 : Oid relid;
940 : bool found;
941 :
942 1446 : relid = RelationGetRelid(state->rs_old_rel);
943 :
944 : /* look for existing mappings for this 'mapped' xid */
945 1446 : src = hash_search(state->rs_logical_mappings, &xid,
946 : HASH_ENTER, &found);
947 :
948 : /*
949 : * We haven't yet had the need to map anything for this xid, create
950 : * per-xid data structures.
951 : */
952 1446 : if (!found)
953 : {
954 : char path[MAXPGPATH];
955 : Oid dboid;
956 :
957 180 : if (state->rs_old_rel->rd_rel->relisshared)
958 0 : dboid = InvalidOid;
959 : else
960 180 : dboid = MyDatabaseId;
961 :
962 180 : snprintf(path, MAXPGPATH,
963 : "%s/" LOGICAL_REWRITE_FORMAT,
964 : PG_LOGICAL_MAPPINGS_DIR, dboid, relid,
965 180 : LSN_FORMAT_ARGS(state->rs_begin_lsn),
966 : xid, GetCurrentTransactionId());
967 :
968 180 : dclist_init(&src->mappings);
969 180 : src->off = 0;
970 180 : memcpy(src->path, path, sizeof(path));
971 180 : src->vfd = PathNameOpenFile(path,
972 : O_CREAT | O_EXCL | O_WRONLY | PG_BINARY);
973 180 : if (src->vfd < 0)
974 0 : ereport(ERROR,
975 : (errcode_for_file_access(),
976 : errmsg("could not create file \"%s\": %m", path)));
977 : }
978 :
979 1446 : pmap = MemoryContextAlloc(state->rs_cxt,
980 : sizeof(RewriteMappingDataEntry));
981 1446 : memcpy(&pmap->map, map, sizeof(LogicalRewriteMappingData));
982 1446 : dclist_push_tail(&src->mappings, &pmap->node);
983 1446 : state->rs_num_rewrite_mappings++;
984 :
985 : /*
986 : * Write out buffer every time we've too many in-memory entries across all
987 : * mapping files.
988 : */
989 1446 : if (state->rs_num_rewrite_mappings >= 1000 /* arbitrary number */ )
990 0 : logical_heap_rewrite_flush_mappings(state);
991 1446 : }
992 :
993 : /*
994 : * Perform logical remapping for a tuple that's mapped from old_tid to
995 : * new_tuple->t_self by rewrite_heap_tuple() if necessary for the tuple.
996 : */
997 : static void
998 740332 : logical_rewrite_heap_tuple(RewriteState state, ItemPointerData old_tid,
999 : HeapTuple new_tuple)
1000 : {
1001 740332 : ItemPointerData new_tid = new_tuple->t_self;
1002 740332 : TransactionId cutoff = state->rs_logical_xmin;
1003 : TransactionId xmin;
1004 : TransactionId xmax;
1005 740332 : bool do_log_xmin = false;
1006 740332 : bool do_log_xmax = false;
1007 : LogicalRewriteMappingData map;
1008 :
1009 : /* no logical rewrite in progress, we don't need to log anything */
1010 740332 : if (!state->rs_logical_rewrite)
1011 738916 : return;
1012 :
1013 53240 : xmin = HeapTupleHeaderGetXmin(new_tuple->t_data);
1014 : /* use *GetUpdateXid to correctly deal with multixacts */
1015 53240 : xmax = HeapTupleHeaderGetUpdateXid(new_tuple->t_data);
1016 :
1017 : /*
1018 : * Log the mapping iff the tuple has been created recently.
1019 : */
1020 53240 : if (TransactionIdIsNormal(xmin) && !TransactionIdPrecedes(xmin, cutoff))
1021 1070 : do_log_xmin = true;
1022 :
1023 53240 : if (!TransactionIdIsNormal(xmax))
1024 : {
1025 : /*
1026 : * no xmax is set, can't have any permanent ones, so this check is
1027 : * sufficient
1028 : */
1029 : }
1030 1002 : else if (HEAP_XMAX_IS_LOCKED_ONLY(new_tuple->t_data->t_infomask))
1031 : {
1032 : /* only locked, we don't care */
1033 : }
1034 1002 : else if (!TransactionIdPrecedes(xmax, cutoff))
1035 : {
1036 : /* tuple has been deleted recently, log */
1037 1002 : do_log_xmax = true;
1038 : }
1039 :
1040 : /* if neither needs to be logged, we're done */
1041 53240 : if (!do_log_xmin && !do_log_xmax)
1042 51824 : return;
1043 :
1044 : /* fill out mapping information */
1045 1416 : map.old_locator = state->rs_old_rel->rd_locator;
1046 1416 : map.old_tid = old_tid;
1047 1416 : map.new_locator = state->rs_new_rel->rd_locator;
1048 1416 : map.new_tid = new_tid;
1049 :
1050 : /* ---
1051 : * Now persist the mapping for the individual xids that are affected. We
1052 : * need to log for both xmin and xmax if they aren't the same transaction
1053 : * since the mapping files are per "affected" xid.
1054 : * We don't muster all that much effort detecting whether xmin and xmax
1055 : * are actually the same transaction, we just check whether the xid is the
1056 : * same disregarding subtransactions. Logging too much is relatively
1057 : * harmless and we could never do the check fully since subtransaction
1058 : * data is thrown away during restarts.
1059 : * ---
1060 : */
1061 1416 : if (do_log_xmin)
1062 1070 : logical_rewrite_log_mapping(state, xmin, &map);
1063 : /* separately log mapping for xmax unless it'd be redundant */
1064 1416 : if (do_log_xmax && !TransactionIdEquals(xmin, xmax))
1065 376 : logical_rewrite_log_mapping(state, xmax, &map);
1066 : }
1067 :
1068 : /*
1069 : * Replay XLOG_HEAP2_REWRITE records
1070 : */
1071 : void
1072 0 : heap_xlog_logical_rewrite(XLogReaderState *r)
1073 : {
1074 : char path[MAXPGPATH];
1075 : int fd;
1076 : xl_heap_rewrite_mapping *xlrec;
1077 : uint32 len;
1078 : char *data;
1079 :
1080 0 : xlrec = (xl_heap_rewrite_mapping *) XLogRecGetData(r);
1081 :
1082 0 : snprintf(path, MAXPGPATH,
1083 : "%s/" LOGICAL_REWRITE_FORMAT,
1084 : PG_LOGICAL_MAPPINGS_DIR, xlrec->mapped_db, xlrec->mapped_rel,
1085 0 : LSN_FORMAT_ARGS(xlrec->start_lsn),
1086 0 : xlrec->mapped_xid, XLogRecGetXid(r));
1087 :
1088 0 : fd = OpenTransientFile(path,
1089 : O_CREAT | O_WRONLY | PG_BINARY);
1090 0 : if (fd < 0)
1091 0 : ereport(ERROR,
1092 : (errcode_for_file_access(),
1093 : errmsg("could not create file \"%s\": %m", path)));
1094 :
1095 : /*
1096 : * Truncate all data that's not guaranteed to have been safely fsynced (by
1097 : * previous record or by the last checkpoint).
1098 : */
1099 0 : pgstat_report_wait_start(WAIT_EVENT_LOGICAL_REWRITE_TRUNCATE);
1100 0 : if (ftruncate(fd, xlrec->offset) != 0)
1101 0 : ereport(ERROR,
1102 : (errcode_for_file_access(),
1103 : errmsg("could not truncate file \"%s\" to %u: %m",
1104 : path, (uint32) xlrec->offset)));
1105 0 : pgstat_report_wait_end();
1106 :
1107 0 : data = XLogRecGetData(r) + sizeof(*xlrec);
1108 :
1109 0 : len = xlrec->num_mappings * sizeof(LogicalRewriteMappingData);
1110 :
1111 : /* write out tail end of mapping file (again) */
1112 0 : errno = 0;
1113 0 : pgstat_report_wait_start(WAIT_EVENT_LOGICAL_REWRITE_MAPPING_WRITE);
1114 0 : if (pg_pwrite(fd, data, len, xlrec->offset) != len)
1115 : {
1116 : /* if write didn't set errno, assume problem is no disk space */
1117 0 : if (errno == 0)
1118 0 : errno = ENOSPC;
1119 0 : ereport(ERROR,
1120 : (errcode_for_file_access(),
1121 : errmsg("could not write to file \"%s\": %m", path)));
1122 : }
1123 0 : pgstat_report_wait_end();
1124 :
1125 : /*
1126 : * Now fsync all previously written data. We could improve things and only
1127 : * do this for the last write to a file, but the required bookkeeping
1128 : * doesn't seem worth the trouble.
1129 : */
1130 0 : pgstat_report_wait_start(WAIT_EVENT_LOGICAL_REWRITE_MAPPING_SYNC);
1131 0 : if (pg_fsync(fd) != 0)
1132 0 : ereport(data_sync_elevel(ERROR),
1133 : (errcode_for_file_access(),
1134 : errmsg("could not fsync file \"%s\": %m", path)));
1135 0 : pgstat_report_wait_end();
1136 :
1137 0 : if (CloseTransientFile(fd) != 0)
1138 0 : ereport(ERROR,
1139 : (errcode_for_file_access(),
1140 : errmsg("could not close file \"%s\": %m", path)));
1141 0 : }
1142 :
1143 : /* ---
1144 : * Perform a checkpoint for logical rewrite mappings
1145 : *
1146 : * This serves two tasks:
1147 : * 1) Remove all mappings not needed anymore based on the logical restart LSN
1148 : * 2) Flush all remaining mappings to disk, so that replay after a checkpoint
1149 : * only has to deal with the parts of a mapping that have been written out
1150 : * after the checkpoint started.
1151 : * ---
1152 : */
1153 : void
1154 3456 : CheckPointLogicalRewriteHeap(void)
1155 : {
1156 : XLogRecPtr cutoff;
1157 : XLogRecPtr redo;
1158 : DIR *mappings_dir;
1159 : struct dirent *mapping_de;
1160 : char path[MAXPGPATH + sizeof(PG_LOGICAL_MAPPINGS_DIR)];
1161 :
1162 : /*
1163 : * We start of with a minimum of the last redo pointer. No new decoding
1164 : * slot will start before that, so that's a safe upper bound for removal.
1165 : */
1166 3456 : redo = GetRedoRecPtr();
1167 :
1168 : /* now check for the restart ptrs from existing slots */
1169 3456 : cutoff = ReplicationSlotsComputeLogicalRestartLSN();
1170 :
1171 : /* don't start earlier than the restart lsn */
1172 3456 : if (cutoff != InvalidXLogRecPtr && redo < cutoff)
1173 2 : cutoff = redo;
1174 :
1175 3456 : mappings_dir = AllocateDir(PG_LOGICAL_MAPPINGS_DIR);
1176 10728 : while ((mapping_de = ReadDir(mappings_dir, PG_LOGICAL_MAPPINGS_DIR)) != NULL)
1177 : {
1178 : Oid dboid;
1179 : Oid relid;
1180 : XLogRecPtr lsn;
1181 : TransactionId rewrite_xid;
1182 : TransactionId create_xid;
1183 : uint32 hi,
1184 : lo;
1185 : PGFileType de_type;
1186 :
1187 7272 : if (strcmp(mapping_de->d_name, ".") == 0 ||
1188 3816 : strcmp(mapping_de->d_name, "..") == 0)
1189 6912 : continue;
1190 :
1191 360 : snprintf(path, sizeof(path), "%s/%s", PG_LOGICAL_MAPPINGS_DIR, mapping_de->d_name);
1192 360 : de_type = get_dirent_type(path, mapping_de, false, DEBUG1);
1193 :
1194 360 : if (de_type != PGFILETYPE_ERROR && de_type != PGFILETYPE_REG)
1195 0 : continue;
1196 :
1197 : /* Skip over files that cannot be ours. */
1198 360 : if (strncmp(mapping_de->d_name, "map-", 4) != 0)
1199 0 : continue;
1200 :
1201 360 : if (sscanf(mapping_de->d_name, LOGICAL_REWRITE_FORMAT,
1202 : &dboid, &relid, &hi, &lo, &rewrite_xid, &create_xid) != 6)
1203 0 : elog(ERROR, "could not parse filename \"%s\"", mapping_de->d_name);
1204 :
1205 360 : lsn = ((uint64) hi) << 32 | lo;
1206 :
1207 360 : if (lsn < cutoff || cutoff == InvalidXLogRecPtr)
1208 : {
1209 180 : elog(DEBUG1, "removing logical rewrite file \"%s\"", path);
1210 180 : if (unlink(path) < 0)
1211 0 : ereport(ERROR,
1212 : (errcode_for_file_access(),
1213 : errmsg("could not remove file \"%s\": %m", path)));
1214 : }
1215 : else
1216 : {
1217 : /* on some operating systems fsyncing a file requires O_RDWR */
1218 180 : int fd = OpenTransientFile(path, O_RDWR | PG_BINARY);
1219 :
1220 : /*
1221 : * The file cannot vanish due to concurrency since this function
1222 : * is the only one removing logical mappings and only one
1223 : * checkpoint can be in progress at a time.
1224 : */
1225 180 : if (fd < 0)
1226 0 : ereport(ERROR,
1227 : (errcode_for_file_access(),
1228 : errmsg("could not open file \"%s\": %m", path)));
1229 :
1230 : /*
1231 : * We could try to avoid fsyncing files that either haven't
1232 : * changed or have only been created since the checkpoint's start,
1233 : * but it's currently not deemed worth the effort.
1234 : */
1235 180 : pgstat_report_wait_start(WAIT_EVENT_LOGICAL_REWRITE_CHECKPOINT_SYNC);
1236 180 : if (pg_fsync(fd) != 0)
1237 0 : ereport(data_sync_elevel(ERROR),
1238 : (errcode_for_file_access(),
1239 : errmsg("could not fsync file \"%s\": %m", path)));
1240 180 : pgstat_report_wait_end();
1241 :
1242 180 : if (CloseTransientFile(fd) != 0)
1243 0 : ereport(ERROR,
1244 : (errcode_for_file_access(),
1245 : errmsg("could not close file \"%s\": %m", path)));
1246 : }
1247 : }
1248 3456 : FreeDir(mappings_dir);
1249 :
1250 : /* persist directory entries to disk */
1251 3456 : fsync_fname(PG_LOGICAL_MAPPINGS_DIR, true);
1252 3456 : }
|
