Branch data 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 : 54 : begin_heap_rewrite(Relation old_heap, Relation new_heap, TransactionId oldest_xmin,
235 : : TransactionId freeze_xid, MultiXactId cutoff_multi)
236 : : {
237 : 54 : RewriteState state;
238 : 54 : MemoryContext rw_cxt;
239 : 54 : MemoryContext old_cxt;
240 : 54 : 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 : 54 : rw_cxt = AllocSetContextCreate(CurrentMemoryContext,
247 : : "Table rewrite",
248 : : ALLOCSET_DEFAULT_SIZES);
249 : 54 : old_cxt = MemoryContextSwitchTo(rw_cxt);
250 : :
251 : : /* Create and fill in the state struct */
252 : 54 : state = palloc0_object(RewriteStateData);
253 : :
254 : 54 : state->rs_old_rel = old_heap;
255 : 54 : state->rs_new_rel = new_heap;
256 : 54 : state->rs_buffer = NULL;
257 : : /* new_heap needn't be empty, just locked */
258 : 54 : state->rs_blockno = RelationGetNumberOfBlocks(new_heap);
259 : 54 : state->rs_oldest_xmin = oldest_xmin;
260 : 54 : state->rs_freeze_xid = freeze_xid;
261 : 54 : state->rs_cutoff_multi = cutoff_multi;
262 : 54 : state->rs_cxt = rw_cxt;
263 : 54 : state->rs_bulkstate = smgr_bulk_start_rel(new_heap, MAIN_FORKNUM);
264 : :
265 : : /* Initialize hash tables used to track update chains */
266 : 54 : hash_ctl.keysize = sizeof(TidHashKey);
267 : 54 : hash_ctl.entrysize = sizeof(UnresolvedTupData);
268 : 54 : hash_ctl.hcxt = state->rs_cxt;
269 : :
270 : 54 : state->rs_unresolved_tups =
271 : 54 : hash_create("Rewrite / Unresolved ctids",
272 : : 128, /* arbitrary initial size */
273 : : &hash_ctl,
274 : : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
275 : :
276 : 54 : hash_ctl.entrysize = sizeof(OldToNewMappingData);
277 : :
278 : 54 : state->rs_old_new_tid_map =
279 : 54 : 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 : 54 : MemoryContextSwitchTo(old_cxt);
285 : :
286 : 54 : logical_begin_heap_rewrite(state);
287 : :
288 : 108 : return state;
289 : 54 : }
290 : :
291 : : /*
292 : : * End a rewrite.
293 : : *
294 : : * state and any other resources are freed.
295 : : */
296 : : void
297 : 54 : end_heap_rewrite(RewriteState state)
298 : : {
299 : 54 : HASH_SEQ_STATUS seq_status;
300 : 54 : 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 : 54 : hash_seq_init(&seq_status, state->rs_unresolved_tups);
307 : :
308 [ + - ]: 54 : 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 [ + + ]: 54 : if (state->rs_buffer)
316 : : {
317 : 39 : smgr_bulk_write(state->rs_bulkstate, state->rs_blockno, state->rs_buffer, true);
318 : 39 : state->rs_buffer = NULL;
319 : 39 : }
320 : :
321 : 54 : smgr_bulk_finish(state->rs_bulkstate);
322 : :
323 : 54 : logical_end_heap_rewrite(state);
324 : :
325 : : /* Deleting the context frees everything */
326 : 54 : MemoryContextDelete(state->rs_cxt);
327 : 54 : }
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 : 114899 : rewrite_heap_tuple(RewriteState state,
342 : : HeapTuple old_tuple, HeapTuple new_tuple)
343 : : {
344 : 114899 : MemoryContext old_cxt;
345 : 114899 : ItemPointerData old_tid;
346 : 114899 : TidHashKey hashkey;
347 : 114899 : bool found;
348 : 114899 : bool free_new;
349 : :
350 : 114899 : 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 : 114899 : memcpy(&new_tuple->t_data->t_choice.t_heap,
359 : : &old_tuple->t_data->t_choice.t_heap,
360 : : sizeof(HeapTupleFields));
361 : :
362 : 114899 : new_tuple->t_data->t_infomask &= ~HEAP_XACT_MASK;
363 : 114899 : new_tuple->t_data->t_infomask2 &= ~HEAP2_XACT_MASK;
364 : 114899 : new_tuple->t_data->t_infomask |=
365 : 114899 : 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 : 229798 : heap_freeze_tuple(new_tuple->t_data,
372 : 114899 : state->rs_old_rel->rd_rel->relfrozenxid,
373 : 114899 : state->rs_old_rel->rd_rel->relminmxid,
374 : 114899 : state->rs_freeze_xid,
375 : 114899 : 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 : 114899 : 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 [ + + ]: 114899 : if (!((old_tuple->t_data->t_infomask & HEAP_XMAX_INVALID) ||
390 [ + - ]: 18966 : HeapTupleHeaderIsOnlyLocked(old_tuple->t_data)) &&
391 [ + - + + ]: 18966 : !HeapTupleHeaderIndicatesMovedPartitions(old_tuple->t_data) &&
392 : 37932 : !(ItemPointerEquals(&(old_tuple->t_self),
393 : 18966 : &(old_tuple->t_data->t_ctid))))
394 : : {
395 : 128 : OldToNewMapping mapping;
396 : :
397 : 128 : memset(&hashkey, 0, sizeof(hashkey));
398 : 128 : hashkey.xmin = HeapTupleHeaderGetUpdateXid(old_tuple->t_data);
399 : 128 : hashkey.tid = old_tuple->t_data->t_ctid;
400 : :
401 : 128 : mapping = (OldToNewMapping)
402 : 128 : hash_search(state->rs_old_new_tid_map, &hashkey,
403 : : HASH_FIND, NULL);
404 : :
405 [ + + ]: 128 : 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 : 4 : new_tuple->t_data->t_ctid = mapping->new_tid;
413 : :
414 : : /* We don't need the mapping entry anymore */
415 : 4 : hash_search(state->rs_old_new_tid_map, &hashkey,
416 : : HASH_REMOVE, &found);
417 [ + - ]: 4 : Assert(found);
418 : 4 : }
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 : 124 : UnresolvedTup unresolved;
426 : :
427 : 124 : unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
428 : : HASH_ENTER, &found);
429 [ + - ]: 124 : Assert(!found);
430 : :
431 : 124 : unresolved->old_tid = old_tuple->t_self;
432 : 124 : 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 : 124 : MemoryContextSwitchTo(old_cxt);
439 : : return;
440 : 124 : }
441 [ + + ]: 128 : }
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 : 114775 : old_tid = old_tuple->t_self;
450 : 114775 : free_new = false;
451 : :
452 : 114775 : for (;;)
453 : : {
454 : 114899 : ItemPointerData new_tid;
455 : :
456 : : /* Insert the tuple and find out where it's put in new_heap */
457 : 114899 : raw_heap_insert(state, new_tuple);
458 : 114899 : new_tid = new_tuple->t_self;
459 : :
460 : 114899 : 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 [ + + + + ]: 114899 : if ((new_tuple->t_data->t_infomask & HEAP_UPDATED) &&
471 : 850 : !TransactionIdPrecedes(HeapTupleHeaderGetXmin(new_tuple->t_data),
472 : 425 : state->rs_oldest_xmin))
473 : : {
474 : : /*
475 : : * Okay, this is B in an update pair. See if we've seen A.
476 : : */
477 : 128 : UnresolvedTup unresolved;
478 : :
479 : 128 : memset(&hashkey, 0, sizeof(hashkey));
480 : 128 : hashkey.xmin = HeapTupleHeaderGetXmin(new_tuple->t_data);
481 : 128 : hashkey.tid = old_tid;
482 : :
483 : 128 : unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
484 : : HASH_FIND, NULL);
485 : :
486 [ + + ]: 128 : 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 [ + + ]: 124 : if (free_new)
494 : 47 : heap_freetuple(new_tuple);
495 : 124 : new_tuple = unresolved->tuple;
496 : 124 : free_new = true;
497 : 124 : old_tid = unresolved->old_tid;
498 : 124 : 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 : 124 : hash_search(state->rs_unresolved_tups, &hashkey,
505 : : HASH_REMOVE, &found);
506 [ - + ]: 124 : Assert(found);
507 : :
508 : : /* loop back to insert the previous tuple in the chain */
509 : 124 : 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 : 4 : OldToNewMapping mapping;
518 : :
519 : 4 : mapping = hash_search(state->rs_old_new_tid_map, &hashkey,
520 : : HASH_ENTER, &found);
521 [ - + ]: 4 : Assert(!found);
522 : :
523 : 4 : mapping->new_tid = new_tid;
524 : 4 : }
525 [ + + ]: 128 : }
526 : :
527 : : /* Done with this (chain of) tuples, for now */
528 [ + + ]: 114775 : if (free_new)
529 : 77 : heap_freetuple(new_tuple);
530 : 114775 : break;
531 [ + + ]: 114899 : }
532 : :
533 : 114775 : MemoryContextSwitchTo(old_cxt);
534 : 114899 : }
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 : 432 : 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 : 432 : UnresolvedTup unresolved;
564 : 432 : TidHashKey hashkey;
565 : 432 : bool found;
566 : :
567 : 432 : memset(&hashkey, 0, sizeof(hashkey));
568 : 432 : hashkey.xmin = HeapTupleHeaderGetXmin(old_tuple->t_data);
569 : 432 : hashkey.tid = old_tuple->t_self;
570 : :
571 : 432 : unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
572 : : HASH_FIND, NULL);
573 : :
574 [ - + ]: 432 : 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 [ # # ]: 0 : Assert(found);
581 : 0 : return true;
582 : : }
583 : :
584 : 432 : return false;
585 : 432 : }
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 : 114899 : raw_heap_insert(RewriteState state, HeapTuple tup)
597 : : {
598 : 114899 : Page page;
599 : 114899 : Size pageFreeSpace,
600 : : saveFreeSpace;
601 : 114899 : Size len;
602 : 114899 : OffsetNumber newoff;
603 : 114899 : 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 [ - + ]: 114899 : if (state->rs_new_rel->rd_rel->relkind == RELKIND_TOASTVALUE)
613 : : {
614 : : /* toast table entries should never be recursively toasted */
615 [ # # ]: 0 : Assert(!HeapTupleHasExternal(tup));
616 : 0 : heaptup = tup;
617 : 0 : }
618 [ + + + + ]: 114899 : else if (HeapTupleHasExternal(tup) || tup->t_len > TOAST_TUPLE_THRESHOLD)
619 : : {
620 : 63 : 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 : 63 : options |= HEAP_INSERT_NO_LOGICAL;
628 : :
629 : 126 : heaptup = heap_toast_insert_or_update(state->rs_new_rel, tup, NULL,
630 : 63 : options);
631 : 63 : }
632 : : else
633 : 114836 : heaptup = tup;
634 : :
635 : 114899 : len = MAXALIGN(heaptup->t_len); /* be conservative */
636 : :
637 : : /*
638 : : * If we're gonna fail for oversize tuple, do it right away
639 : : */
640 [ + - ]: 114899 : 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 [ + + ]: 114899 : 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 : 114899 : page = (Page) state->rs_buffer;
652 [ + + ]: 114899 : if (page)
653 : : {
654 : 114860 : pageFreeSpace = PageGetHeapFreeSpace(page);
655 : :
656 [ + + ]: 114860 : 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 : 1456 : smgr_bulk_write(state->rs_bulkstate, state->rs_blockno, state->rs_buffer, true);
664 : 1456 : state->rs_buffer = NULL;
665 : 1456 : page = NULL;
666 : 1456 : state->rs_blockno++;
667 : 1456 : }
668 : 114860 : }
669 : :
670 [ + + ]: 114899 : if (!page)
671 : : {
672 : : /* Initialize a new empty page */
673 : 1495 : state->rs_buffer = smgr_bulk_get_buf(state->rs_bulkstate);
674 : 1495 : page = (Page) state->rs_buffer;
675 : 1495 : PageInit(page, BLCKSZ, 0);
676 : 1495 : }
677 : :
678 : : /* And now we can insert the tuple into the page */
679 : 114899 : newoff = PageAddItem(page, heaptup->t_data, heaptup->t_len, InvalidOffsetNumber, false, true);
680 [ + - ]: 114899 : 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 : 114899 : 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 [ + + ]: 114899 : if (!ItemPointerIsValid(&tup->t_data->t_ctid))
691 : : {
692 : 114771 : ItemId newitemid;
693 : 114771 : HeapTupleHeader onpage_tup;
694 : :
695 : 114771 : newitemid = PageGetItemId(page, newoff);
696 : 114771 : onpage_tup = (HeapTupleHeader) PageGetItem(page, newitemid);
697 : :
698 : 114771 : onpage_tup->t_ctid = tup->t_self;
699 : 114771 : }
700 : :
701 : : /* If heaptup is a private copy, release it. */
702 [ + + ]: 114899 : if (heaptup != tup)
703 : 63 : heap_freetuple(heaptup);
704 : 114899 : }
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 : 54 : logical_begin_heap_rewrite(RewriteState state)
762 : : {
763 : 54 : HASHCTL hash_ctl;
764 : 54 : 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 : 54 : state->rs_logical_rewrite =
772 [ + - - + : 54 : RelationIsAccessibleInLogicalDecoding(state->rs_old_rel);
# # # # #
# # # # #
# # # # ]
773 : :
774 [ - + ]: 54 : if (!state->rs_logical_rewrite)
775 : 54 : return;
776 : :
777 : 0 : 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 [ # # ]: 0 : if (logical_xmin == InvalidTransactionId)
785 : : {
786 : 0 : state->rs_logical_rewrite = false;
787 : 0 : return;
788 : : }
789 : :
790 : 0 : state->rs_logical_xmin = logical_xmin;
791 : 0 : state->rs_begin_lsn = GetXLogInsertRecPtr();
792 : 0 : state->rs_num_rewrite_mappings = 0;
793 : :
794 : 0 : hash_ctl.keysize = sizeof(TransactionId);
795 : 0 : hash_ctl.entrysize = sizeof(RewriteMappingFile);
796 : 0 : hash_ctl.hcxt = state->rs_cxt;
797 : :
798 : 0 : state->rs_logical_mappings =
799 : 0 : hash_create("Logical rewrite mapping",
800 : : 128, /* arbitrary initial size */
801 : : &hash_ctl,
802 : : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
803 [ - + ]: 54 : }
804 : :
805 : : /*
806 : : * Flush all logical in-memory mappings to disk, but don't fsync them yet.
807 : : */
808 : : static void
809 : 0 : logical_heap_rewrite_flush_mappings(RewriteState state)
810 : : {
811 : 0 : HASH_SEQ_STATUS seq_status;
812 : 0 : RewriteMappingFile *src;
813 : 0 : dlist_mutable_iter iter;
814 : :
815 [ # # ]: 0 : Assert(state->rs_logical_rewrite);
816 : :
817 : : /* no logical rewrite in progress, no need to iterate over mappings */
818 [ # # ]: 0 : if (state->rs_num_rewrite_mappings == 0)
819 : 0 : return;
820 : :
821 [ # # # # ]: 0 : elog(DEBUG1, "flushing %u logical rewrite mapping entries",
822 : : state->rs_num_rewrite_mappings);
823 : :
824 : 0 : hash_seq_init(&seq_status, state->rs_logical_mappings);
825 [ # # ]: 0 : while ((src = (RewriteMappingFile *) hash_seq_search(&seq_status)) != NULL)
826 : : {
827 : 0 : char *waldata;
828 : 0 : char *waldata_start;
829 : 0 : xl_heap_rewrite_mapping xlrec;
830 : 0 : Oid dboid;
831 : 0 : uint32 len;
832 : 0 : int written;
833 : 0 : uint32 num_mappings = dclist_count(&src->mappings);
834 : :
835 : : /* this file hasn't got any new mappings */
836 [ # # ]: 0 : if (num_mappings == 0)
837 : 0 : continue;
838 : :
839 [ # # ]: 0 : if (state->rs_old_rel->rd_rel->relisshared)
840 : 0 : dboid = InvalidOid;
841 : : else
842 : 0 : dboid = MyDatabaseId;
843 : :
844 : 0 : xlrec.num_mappings = num_mappings;
845 : 0 : xlrec.mapped_rel = RelationGetRelid(state->rs_old_rel);
846 : 0 : xlrec.mapped_xid = src->xid;
847 : 0 : xlrec.mapped_db = dboid;
848 : 0 : xlrec.offset = src->off;
849 : 0 : xlrec.start_lsn = state->rs_begin_lsn;
850 : :
851 : : /* write all mappings consecutively */
852 : 0 : len = num_mappings * sizeof(LogicalRewriteMappingData);
853 : 0 : waldata_start = waldata = palloc(len);
854 : :
855 : : /*
856 : : * collect data we need to write out, but don't modify ondisk data yet
857 : : */
858 [ # # # # ]: 0 : dclist_foreach_modify(iter, &src->mappings)
859 : : {
860 : 0 : RewriteMappingDataEntry *pmap;
861 : :
862 : 0 : pmap = dclist_container(RewriteMappingDataEntry, node, iter.cur);
863 : :
864 : 0 : memcpy(waldata, &pmap->map, sizeof(pmap->map));
865 : 0 : waldata += sizeof(pmap->map);
866 : :
867 : : /* remove from the list and free */
868 : 0 : dclist_delete_from(&src->mappings, &pmap->node);
869 : 0 : pfree(pmap);
870 : :
871 : : /* update bookkeeping */
872 : 0 : state->rs_num_rewrite_mappings--;
873 : 0 : }
874 : :
875 [ # # ]: 0 : Assert(dclist_count(&src->mappings) == 0);
876 [ # # ]: 0 : 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 : 0 : written = FileWrite(src->vfd, waldata_start, len, src->off,
883 : : WAIT_EVENT_LOGICAL_REWRITE_WRITE);
884 [ # # ]: 0 : 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 : 0 : src->off += len;
890 : :
891 : 0 : XLogBeginInsert();
892 : 0 : XLogRegisterData(&xlrec, sizeof(xlrec));
893 : 0 : XLogRegisterData(waldata_start, len);
894 : :
895 : : /* write xlog record */
896 : 0 : XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_REWRITE);
897 : :
898 : 0 : pfree(waldata_start);
899 [ # # ]: 0 : }
900 [ # # ]: 0 : Assert(state->rs_num_rewrite_mappings == 0);
901 : 0 : }
902 : :
903 : : /*
904 : : * Logical remapping part of end_heap_rewrite().
905 : : */
906 : : static void
907 : 54 : logical_end_heap_rewrite(RewriteState state)
908 : : {
909 : 54 : HASH_SEQ_STATUS seq_status;
910 : 54 : RewriteMappingFile *src;
911 : :
912 : : /* done, no logical rewrite in progress */
913 [ - + ]: 54 : if (!state->rs_logical_rewrite)
914 : 54 : return;
915 : :
916 : : /* writeout remaining in-memory entries */
917 [ # # ]: 0 : if (state->rs_num_rewrite_mappings > 0)
918 : 0 : logical_heap_rewrite_flush_mappings(state);
919 : :
920 : : /* Iterate over all mappings we have written and fsync the files. */
921 : 0 : hash_seq_init(&seq_status, state->rs_logical_mappings);
922 [ # # ]: 0 : while ((src = (RewriteMappingFile *) hash_seq_search(&seq_status)) != NULL)
923 : : {
924 [ # # ]: 0 : 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 : 0 : FileClose(src->vfd);
929 : : }
930 : : /* memory context cleanup will deal with the rest */
931 [ - + ]: 54 : }
932 : :
933 : : /*
934 : : * Log a single (old->new) mapping for 'xid'.
935 : : */
936 : : static void
937 : 0 : logical_rewrite_log_mapping(RewriteState state, TransactionId xid,
938 : : LogicalRewriteMappingData *map)
939 : : {
940 : 0 : RewriteMappingFile *src;
941 : 0 : RewriteMappingDataEntry *pmap;
942 : 0 : Oid relid;
943 : 0 : bool found;
944 : :
945 : 0 : relid = RelationGetRelid(state->rs_old_rel);
946 : :
947 : : /* look for existing mappings for this 'mapped' xid */
948 : 0 : 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 [ # # ]: 0 : if (!found)
956 : : {
957 : 0 : char path[MAXPGPATH];
958 : 0 : Oid dboid;
959 : :
960 [ # # ]: 0 : if (state->rs_old_rel->rd_rel->relisshared)
961 : 0 : dboid = InvalidOid;
962 : : else
963 : 0 : dboid = MyDatabaseId;
964 : :
965 : 0 : snprintf(path, MAXPGPATH,
966 : : "%s/" LOGICAL_REWRITE_FORMAT,
967 : 0 : PG_LOGICAL_MAPPINGS_DIR, dboid, relid,
968 : 0 : LSN_FORMAT_ARGS(state->rs_begin_lsn),
969 : 0 : xid, GetCurrentTransactionId());
970 : :
971 : 0 : dclist_init(&src->mappings);
972 : 0 : src->off = 0;
973 : 0 : memcpy(src->path, path, sizeof(path));
974 : 0 : src->vfd = PathNameOpenFile(path,
975 : : O_CREAT | O_EXCL | O_WRONLY | PG_BINARY);
976 [ # # ]: 0 : if (src->vfd < 0)
977 [ # # # # ]: 0 : ereport(ERROR,
978 : : (errcode_for_file_access(),
979 : : errmsg("could not create file \"%s\": %m", path)));
980 : 0 : }
981 : :
982 : 0 : pmap = MemoryContextAlloc(state->rs_cxt,
983 : : sizeof(RewriteMappingDataEntry));
984 : 0 : memcpy(&pmap->map, map, sizeof(LogicalRewriteMappingData));
985 : 0 : dclist_push_tail(&src->mappings, &pmap->node);
986 : 0 : 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 [ # # ]: 0 : if (state->rs_num_rewrite_mappings >= 1000 /* arbitrary number */ )
993 : 0 : logical_heap_rewrite_flush_mappings(state);
994 : 0 : }
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 : 114899 : logical_rewrite_heap_tuple(RewriteState state, ItemPointerData old_tid,
1002 : : HeapTuple new_tuple)
1003 : : {
1004 : 114899 : ItemPointerData new_tid = new_tuple->t_self;
1005 : 114899 : TransactionId cutoff = state->rs_logical_xmin;
1006 : 114899 : TransactionId xmin;
1007 : 114899 : TransactionId xmax;
1008 : 114899 : bool do_log_xmin = false;
1009 : 114899 : bool do_log_xmax = false;
1010 : 114899 : LogicalRewriteMappingData map;
1011 : :
1012 : : /* no logical rewrite in progress, we don't need to log anything */
1013 [ - + ]: 114899 : if (!state->rs_logical_rewrite)
1014 : 114899 : return;
1015 : :
1016 : 0 : xmin = HeapTupleHeaderGetXmin(new_tuple->t_data);
1017 : : /* use *GetUpdateXid to correctly deal with multixacts */
1018 : 0 : xmax = HeapTupleHeaderGetUpdateXid(new_tuple->t_data);
1019 : :
1020 : : /*
1021 : : * Log the mapping iff the tuple has been created recently.
1022 : : */
1023 [ # # # # ]: 0 : if (TransactionIdIsNormal(xmin) && !TransactionIdPrecedes(xmin, cutoff))
1024 : 0 : do_log_xmin = true;
1025 : :
1026 [ # # ]: 0 : if (!TransactionIdIsNormal(xmax))
1027 : : {
1028 : : /*
1029 : : * no xmax is set, can't have any permanent ones, so this check is
1030 : : * sufficient
1031 : : */
1032 : 0 : }
1033 [ # # ]: 0 : else if (HEAP_XMAX_IS_LOCKED_ONLY(new_tuple->t_data->t_infomask))
1034 : : {
1035 : : /* only locked, we don't care */
1036 : 0 : }
1037 [ # # ]: 0 : else if (!TransactionIdPrecedes(xmax, cutoff))
1038 : : {
1039 : : /* tuple has been deleted recently, log */
1040 : 0 : do_log_xmax = true;
1041 : 0 : }
1042 : :
1043 : : /* if neither needs to be logged, we're done */
1044 [ # # # # ]: 0 : if (!do_log_xmin && !do_log_xmax)
1045 : 0 : return;
1046 : :
1047 : : /* fill out mapping information */
1048 : 0 : map.old_locator = state->rs_old_rel->rd_locator;
1049 : 0 : map.old_tid = old_tid;
1050 : 0 : map.new_locator = state->rs_new_rel->rd_locator;
1051 : 0 : 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 [ # # ]: 0 : if (do_log_xmin)
1065 : 0 : logical_rewrite_log_mapping(state, xmin, &map);
1066 : : /* separately log mapping for xmax unless it'd be redundant */
1067 [ # # # # ]: 0 : if (do_log_xmax && !TransactionIdEquals(xmin, xmax))
1068 : 0 : logical_rewrite_log_mapping(state, xmax, &map);
1069 [ - + ]: 114899 : }
1070 : :
1071 : : /*
1072 : : * Replay XLOG_HEAP2_REWRITE records
1073 : : */
1074 : : void
1075 : 0 : heap_xlog_logical_rewrite(XLogReaderState *r)
1076 : : {
1077 : 0 : char path[MAXPGPATH];
1078 : 0 : int fd;
1079 : 0 : xl_heap_rewrite_mapping *xlrec;
1080 : 0 : uint32 len;
1081 : 0 : char *data;
1082 : :
1083 : 0 : xlrec = (xl_heap_rewrite_mapping *) XLogRecGetData(r);
1084 : :
1085 : 0 : snprintf(path, MAXPGPATH,
1086 : : "%s/" LOGICAL_REWRITE_FORMAT,
1087 : 0 : 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 : 0 : }
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 : 7 : CheckPointLogicalRewriteHeap(void)
1158 : : {
1159 : 7 : XLogRecPtr cutoff;
1160 : 7 : XLogRecPtr redo;
1161 : 7 : DIR *mappings_dir;
1162 : 7 : struct dirent *mapping_de;
1163 : 7 : 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 : 7 : redo = GetRedoRecPtr();
1170 : :
1171 : : /* now check for the restart ptrs from existing slots */
1172 : 7 : cutoff = ReplicationSlotsComputeLogicalRestartLSN();
1173 : :
1174 : : /* don't start earlier than the restart lsn */
1175 [ - + # # ]: 7 : if (XLogRecPtrIsValid(cutoff) && redo < cutoff)
1176 : 0 : cutoff = redo;
1177 : :
1178 : 7 : mappings_dir = AllocateDir(PG_LOGICAL_MAPPINGS_DIR);
1179 [ + + ]: 21 : while ((mapping_de = ReadDir(mappings_dir, PG_LOGICAL_MAPPINGS_DIR)) != NULL)
1180 : : {
1181 : 14 : Oid dboid;
1182 : 14 : Oid relid;
1183 : 14 : XLogRecPtr lsn;
1184 : 14 : TransactionId rewrite_xid;
1185 : 14 : TransactionId create_xid;
1186 : 14 : uint32 hi,
1187 : : lo;
1188 : 14 : PGFileType de_type;
1189 : :
1190 [ + + + - ]: 14 : if (strcmp(mapping_de->d_name, ".") == 0 ||
1191 : 7 : strcmp(mapping_de->d_name, "..") == 0)
1192 : 14 : continue;
1193 : :
1194 : 0 : snprintf(path, sizeof(path), "%s/%s", PG_LOGICAL_MAPPINGS_DIR, mapping_de->d_name);
1195 : 0 : de_type = get_dirent_type(path, mapping_de, false, DEBUG1);
1196 : :
1197 [ # # # # ]: 0 : if (de_type != PGFILETYPE_ERROR && de_type != PGFILETYPE_REG)
1198 : 0 : continue;
1199 : :
1200 : : /* Skip over files that cannot be ours. */
1201 [ # # ]: 0 : if (strncmp(mapping_de->d_name, "map-", 4) != 0)
1202 : 0 : continue;
1203 : :
1204 : 0 : if (sscanf(mapping_de->d_name, LOGICAL_REWRITE_FORMAT,
1205 [ # # ]: 0 : &dboid, &relid, &hi, &lo, &rewrite_xid, &create_xid) != 6)
1206 [ # # # # ]: 0 : elog(ERROR, "could not parse filename \"%s\"", mapping_de->d_name);
1207 : :
1208 : 0 : lsn = ((uint64) hi) << 32 | lo;
1209 : :
1210 [ # # # # ]: 0 : if (lsn < cutoff || !XLogRecPtrIsValid(cutoff))
1211 : : {
1212 [ # # # # ]: 0 : elog(DEBUG1, "removing logical rewrite file \"%s\"", path);
1213 [ # # ]: 0 : if (unlink(path) < 0)
1214 [ # # # # ]: 0 : ereport(ERROR,
1215 : : (errcode_for_file_access(),
1216 : : errmsg("could not remove file \"%s\": %m", path)));
1217 : 0 : }
1218 : : else
1219 : : {
1220 : : /* on some operating systems fsyncing a file requires O_RDWR */
1221 : 0 : 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 [ # # ]: 0 : 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 : 0 : pgstat_report_wait_start(WAIT_EVENT_LOGICAL_REWRITE_CHECKPOINT_SYNC);
1239 [ # # ]: 0 : 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 : 0 : pgstat_report_wait_end();
1244 : :
1245 [ # # ]: 0 : if (CloseTransientFile(fd) != 0)
1246 [ # # # # ]: 0 : ereport(ERROR,
1247 : : (errcode_for_file_access(),
1248 : : errmsg("could not close file \"%s\": %m", path)));
1249 : 0 : }
1250 [ - + - ]: 14 : }
1251 : 7 : FreeDir(mappings_dir);
1252 : :
1253 : : /* persist directory entries to disk */
1254 : 7 : fsync_fname(PG_LOGICAL_MAPPINGS_DIR, true);
1255 : 7 : }
|
