PostgreSQL Source Code  git master
storage.c
Go to the documentation of this file.
1 /*-------------------------------------------------------------------------
2  *
3  * storage.c
4  * code to create and destroy physical storage for relations
5  *
6  * Portions Copyright (c) 1996-2020, PostgreSQL Global Development Group
7  * Portions Copyright (c) 1994, Regents of the University of California
8  *
9  *
10  * IDENTIFICATION
11  * src/backend/catalog/storage.c
12  *
13  * NOTES
14  * Some of this code used to be in storage/smgr/smgr.c, and the
15  * function names still reflect that.
16  *
17  *-------------------------------------------------------------------------
18  */
19 
20 #include "postgres.h"
21 
22 #include "access/parallel.h"
23 #include "access/visibilitymap.h"
24 #include "access/xact.h"
25 #include "access/xlog.h"
26 #include "access/xloginsert.h"
27 #include "access/xlogutils.h"
28 #include "catalog/storage.h"
29 #include "catalog/storage_xlog.h"
30 #include "miscadmin.h"
31 #include "storage/freespace.h"
32 #include "storage/smgr.h"
33 #include "utils/hsearch.h"
34 #include "utils/memutils.h"
35 #include "utils/rel.h"
36 
37 /* GUC variables */
38 int wal_skip_threshold = 2048; /* in kilobytes */
39 
40 /*
41  * We keep a list of all relations (represented as RelFileNode values)
42  * that have been created or deleted in the current transaction. When
43  * a relation is created, we create the physical file immediately, but
44  * remember it so that we can delete the file again if the current
45  * transaction is aborted. Conversely, a deletion request is NOT
46  * executed immediately, but is just entered in the list. When and if
47  * the transaction commits, we can delete the physical file.
48  *
49  * To handle subtransactions, every entry is marked with its transaction
50  * nesting level. At subtransaction commit, we reassign the subtransaction's
51  * entries to the parent nesting level. At subtransaction abort, we can
52  * immediately execute the abort-time actions for all entries of the current
53  * nesting level.
54  *
55  * NOTE: the list is kept in TopMemoryContext to be sure it won't disappear
56  * unbetimes. It'd probably be OK to keep it in TopTransactionContext,
57  * but I'm being paranoid.
58  */
59 
60 typedef struct PendingRelDelete
61 {
62  RelFileNode relnode; /* relation that may need to be deleted */
63  BackendId backend; /* InvalidBackendId if not a temp rel */
64  bool atCommit; /* T=delete at commit; F=delete at abort */
65  int nestLevel; /* xact nesting level of request */
66  struct PendingRelDelete *next; /* linked-list link */
68 
69 typedef struct PendingRelSync
70 {
72  bool is_truncated; /* Has the file experienced truncation? */
74 
75 static PendingRelDelete *pendingDeletes = NULL; /* head of linked list */
77 
78 
79 /*
80  * AddPendingSync
81  * Queue an at-commit fsync.
82  */
83 static void
85 {
86  PendingRelSync *pending;
87  bool found;
88 
89  /* create the hash if not yet */
90  if (!pendingSyncHash)
91  {
92  HASHCTL ctl;
93 
94  ctl.keysize = sizeof(RelFileNode);
95  ctl.entrysize = sizeof(PendingRelSync);
97  pendingSyncHash = hash_create("pending sync hash", 16, &ctl,
99  }
100 
101  pending = hash_search(pendingSyncHash, rnode, HASH_ENTER, &found);
102  Assert(!found);
103  pending->is_truncated = false;
104 }
105 
106 /*
107  * RelationCreateStorage
108  * Create physical storage for a relation.
109  *
110  * Create the underlying disk file storage for the relation. This only
111  * creates the main fork; additional forks are created lazily by the
112  * modules that need them.
113  *
114  * This function is transactional. The creation is WAL-logged, and if the
115  * transaction aborts later on, the storage will be destroyed.
116  */
118 RelationCreateStorage(RelFileNode rnode, char relpersistence)
119 {
120  PendingRelDelete *pending;
121  SMgrRelation srel;
123  bool needs_wal;
124 
125  Assert(!IsInParallelMode()); /* couldn't update pendingSyncHash */
126 
127  switch (relpersistence)
128  {
129  case RELPERSISTENCE_TEMP:
130  backend = BackendIdForTempRelations();
131  needs_wal = false;
132  break;
133  case RELPERSISTENCE_UNLOGGED:
134  backend = InvalidBackendId;
135  needs_wal = false;
136  break;
137  case RELPERSISTENCE_PERMANENT:
138  backend = InvalidBackendId;
139  needs_wal = true;
140  break;
141  default:
142  elog(ERROR, "invalid relpersistence: %c", relpersistence);
143  return NULL; /* placate compiler */
144  }
145 
146  srel = smgropen(rnode, backend);
147  smgrcreate(srel, MAIN_FORKNUM, false);
148 
149  if (needs_wal)
151 
152  /* Add the relation to the list of stuff to delete at abort */
153  pending = (PendingRelDelete *)
155  pending->relnode = rnode;
156  pending->backend = backend;
157  pending->atCommit = false; /* delete if abort */
159  pending->next = pendingDeletes;
160  pendingDeletes = pending;
161 
162  if (relpersistence == RELPERSISTENCE_PERMANENT && !XLogIsNeeded())
163  {
164  Assert(backend == InvalidBackendId);
165  AddPendingSync(&rnode);
166  }
167 
168  return srel;
169 }
170 
171 /*
172  * Perform XLogInsert of an XLOG_SMGR_CREATE record to WAL.
173  */
174 void
175 log_smgrcreate(const RelFileNode *rnode, ForkNumber forkNum)
176 {
177  xl_smgr_create xlrec;
178 
179  /*
180  * Make an XLOG entry reporting the file creation.
181  */
182  xlrec.rnode = *rnode;
183  xlrec.forkNum = forkNum;
184 
185  XLogBeginInsert();
186  XLogRegisterData((char *) &xlrec, sizeof(xlrec));
188 }
189 
190 /*
191  * RelationDropStorage
192  * Schedule unlinking of physical storage at transaction commit.
193  */
194 void
196 {
197  PendingRelDelete *pending;
198 
199  /* Add the relation to the list of stuff to delete at commit */
200  pending = (PendingRelDelete *)
202  pending->relnode = rel->rd_node;
203  pending->backend = rel->rd_backend;
204  pending->atCommit = true; /* delete if commit */
206  pending->next = pendingDeletes;
207  pendingDeletes = pending;
208 
209  /*
210  * NOTE: if the relation was created in this transaction, it will now be
211  * present in the pending-delete list twice, once with atCommit true and
212  * once with atCommit false. Hence, it will be physically deleted at end
213  * of xact in either case (and the other entry will be ignored by
214  * smgrDoPendingDeletes, so no error will occur). We could instead remove
215  * the existing list entry and delete the physical file immediately, but
216  * for now I'll keep the logic simple.
217  */
218 
219  RelationCloseSmgr(rel);
220 }
221 
222 /*
223  * RelationPreserveStorage
224  * Mark a relation as not to be deleted after all.
225  *
226  * We need this function because relation mapping changes are committed
227  * separately from commit of the whole transaction, so it's still possible
228  * for the transaction to abort after the mapping update is done.
229  * When a new physical relation is installed in the map, it would be
230  * scheduled for delete-on-abort, so we'd delete it, and be in trouble.
231  * The relation mapper fixes this by telling us to not delete such relations
232  * after all as part of its commit.
233  *
234  * We also use this to reuse an old build of an index during ALTER TABLE, this
235  * time removing the delete-at-commit entry.
236  *
237  * No-op if the relation is not among those scheduled for deletion.
238  */
239 void
241 {
242  PendingRelDelete *pending;
243  PendingRelDelete *prev;
245 
246  prev = NULL;
247  for (pending = pendingDeletes; pending != NULL; pending = next)
248  {
249  next = pending->next;
250  if (RelFileNodeEquals(rnode, pending->relnode)
251  && pending->atCommit == atCommit)
252  {
253  /* unlink and delete list entry */
254  if (prev)
255  prev->next = next;
256  else
257  pendingDeletes = next;
258  pfree(pending);
259  /* prev does not change */
260  }
261  else
262  {
263  /* unrelated entry, don't touch it */
264  prev = pending;
265  }
266  }
267 }
268 
269 /*
270  * RelationTruncate
271  * Physically truncate a relation to the specified number of blocks.
272  *
273  * This includes getting rid of any buffers for the blocks that are to be
274  * dropped.
275  */
276 void
278 {
279  bool fsm;
280  bool vm;
281  bool need_fsm_vacuum = false;
282  ForkNumber forks[MAX_FORKNUM];
284  int nforks = 0;
285 
286  /* Open it at the smgr level if not already done */
287  RelationOpenSmgr(rel);
288 
289  /*
290  * Make sure smgr_targblock etc aren't pointing somewhere past new end
291  */
293  for (int i = 0; i <= MAX_FORKNUM; ++i)
295 
296  /* Prepare for truncation of MAIN fork of the relation */
297  forks[nforks] = MAIN_FORKNUM;
298  blocks[nforks] = nblocks;
299  nforks++;
300 
301  /* Prepare for truncation of the FSM if it exists */
302  fsm = smgrexists(rel->rd_smgr, FSM_FORKNUM);
303  if (fsm)
304  {
305  blocks[nforks] = FreeSpaceMapPrepareTruncateRel(rel, nblocks);
306  if (BlockNumberIsValid(blocks[nforks]))
307  {
308  forks[nforks] = FSM_FORKNUM;
309  nforks++;
310  need_fsm_vacuum = true;
311  }
312  }
313 
314  /* Prepare for truncation of the visibility map too if it exists */
316  if (vm)
317  {
318  blocks[nforks] = visibilitymap_prepare_truncate(rel, nblocks);
319  if (BlockNumberIsValid(blocks[nforks]))
320  {
321  forks[nforks] = VISIBILITYMAP_FORKNUM;
322  nforks++;
323  }
324  }
325 
326  RelationPreTruncate(rel);
327 
328  /*
329  * We WAL-log the truncation before actually truncating, which means
330  * trouble if the truncation fails. If we then crash, the WAL replay
331  * likely isn't going to succeed in the truncation either, and cause a
332  * PANIC. It's tempting to put a critical section here, but that cure
333  * would be worse than the disease. It would turn a usually harmless
334  * failure to truncate, that might spell trouble at WAL replay, into a
335  * certain PANIC.
336  */
337  if (RelationNeedsWAL(rel))
338  {
339  /*
340  * Make an XLOG entry reporting the file truncation.
341  */
342  XLogRecPtr lsn;
343  xl_smgr_truncate xlrec;
344 
345  xlrec.blkno = nblocks;
346  xlrec.rnode = rel->rd_node;
347  xlrec.flags = SMGR_TRUNCATE_ALL;
348 
349  XLogBeginInsert();
350  XLogRegisterData((char *) &xlrec, sizeof(xlrec));
351 
352  lsn = XLogInsert(RM_SMGR_ID,
354 
355  /*
356  * Flush, because otherwise the truncation of the main relation might
357  * hit the disk before the WAL record, and the truncation of the FSM
358  * or visibility map. If we crashed during that window, we'd be left
359  * with a truncated heap, but the FSM or visibility map would still
360  * contain entries for the non-existent heap pages.
361  */
362  if (fsm || vm)
363  XLogFlush(lsn);
364  }
365 
366  /* Do the real work to truncate relation forks */
367  smgrtruncate(rel->rd_smgr, forks, nforks, blocks);
368 
369  /*
370  * Update upper-level FSM pages to account for the truncation. This is
371  * important because the just-truncated pages were likely marked as
372  * all-free, and would be preferentially selected.
373  */
374  if (need_fsm_vacuum)
376 }
377 
378 /*
379  * RelationPreTruncate
380  * Perform AM-independent work before a physical truncation.
381  *
382  * If an access method's relation_nontransactional_truncate does not call
383  * RelationTruncate(), it must call this before decreasing the table size.
384  */
385 void
387 {
388  PendingRelSync *pending;
389 
390  if (!pendingSyncHash)
391  return;
392  RelationOpenSmgr(rel);
393 
394  pending = hash_search(pendingSyncHash, &(rel->rd_smgr->smgr_rnode.node),
395  HASH_FIND, NULL);
396  if (pending)
397  pending->is_truncated = true;
398 }
399 
400 /*
401  * Copy a fork's data, block by block.
402  *
403  * Note that this requires that there is no dirty data in shared buffers. If
404  * it's possible that there are, callers need to flush those using
405  * e.g. FlushRelationBuffers(rel).
406  */
407 void
409  ForkNumber forkNum, char relpersistence)
410 {
412  Page page;
413  bool use_wal;
414  bool copying_initfork;
415  BlockNumber nblocks;
416  BlockNumber blkno;
417 
418  page = (Page) buf.data;
419 
420  /*
421  * The init fork for an unlogged relation in many respects has to be
422  * treated the same as normal relation, changes need to be WAL logged and
423  * it needs to be synced to disk.
424  */
425  copying_initfork = relpersistence == RELPERSISTENCE_UNLOGGED &&
426  forkNum == INIT_FORKNUM;
427 
428  /*
429  * We need to log the copied data in WAL iff WAL archiving/streaming is
430  * enabled AND it's a permanent relation. This gives the same answer as
431  * "RelationNeedsWAL(rel) || copying_initfork", because we know the
432  * current operation created a new relfilenode.
433  */
434  use_wal = XLogIsNeeded() &&
435  (relpersistence == RELPERSISTENCE_PERMANENT || copying_initfork);
436 
437  nblocks = smgrnblocks(src, forkNum);
438 
439  for (blkno = 0; blkno < nblocks; blkno++)
440  {
441  /* If we got a cancel signal during the copy of the data, quit */
443 
444  smgrread(src, forkNum, blkno, buf.data);
445 
446  if (!PageIsVerified(page, blkno))
447  ereport(ERROR,
449  errmsg("invalid page in block %u of relation %s",
450  blkno,
452  src->smgr_rnode.backend,
453  forkNum))));
454 
455  /*
456  * WAL-log the copied page. Unfortunately we don't know what kind of a
457  * page this is, so we have to log the full page including any unused
458  * space.
459  */
460  if (use_wal)
461  log_newpage(&dst->smgr_rnode.node, forkNum, blkno, page, false);
462 
463  PageSetChecksumInplace(page, blkno);
464 
465  /*
466  * Now write the page. We say skipFsync = true because there's no
467  * need for smgr to schedule an fsync for this write; we'll do it
468  * ourselves below.
469  */
470  smgrextend(dst, forkNum, blkno, buf.data, true);
471  }
472 
473  /*
474  * When we WAL-logged rel pages, we must nonetheless fsync them. The
475  * reason is that since we're copying outside shared buffers, a CHECKPOINT
476  * occurring during the copy has no way to flush the previously written
477  * data to disk (indeed it won't know the new rel even exists). A crash
478  * later on would replay WAL from the checkpoint, therefore it wouldn't
479  * replay our earlier WAL entries. If we do not fsync those pages here,
480  * they might still not be on disk when the crash occurs.
481  */
482  if (use_wal || copying_initfork)
483  smgrimmedsync(dst, forkNum);
484 }
485 
486 /*
487  * RelFileNodeSkippingWAL
488  * Check if a BM_PERMANENT relfilenode is using WAL.
489  *
490  * Changes of certain relfilenodes must not write WAL; see "Skipping WAL for
491  * New RelFileNode" in src/backend/access/transam/README. Though it is known
492  * from Relation efficiently, this function is intended for the code paths not
493  * having access to Relation.
494  */
495 bool
497 {
498  if (!pendingSyncHash ||
499  hash_search(pendingSyncHash, &rnode, HASH_FIND, NULL) == NULL)
500  return false;
501 
502  return true;
503 }
504 
505 /*
506  * EstimatePendingSyncsSpace
507  * Estimate space needed to pass syncs to parallel workers.
508  */
509 Size
511 {
512  long entries;
513 
514  entries = pendingSyncHash ? hash_get_num_entries(pendingSyncHash) : 0;
515  return mul_size(1 + entries, sizeof(RelFileNode));
516 }
517 
518 /*
519  * SerializePendingSyncs
520  * Serialize syncs for parallel workers.
521  */
522 void
523 SerializePendingSyncs(Size maxSize, char *startAddress)
524 {
525  HTAB *tmphash;
526  HASHCTL ctl;
527  HASH_SEQ_STATUS scan;
528  PendingRelSync *sync;
529  PendingRelDelete *delete;
530  RelFileNode *src;
531  RelFileNode *dest = (RelFileNode *) startAddress;
532 
533  if (!pendingSyncHash)
534  goto terminate;
535 
536  /* Create temporary hash to collect active relfilenodes */
537  ctl.keysize = sizeof(RelFileNode);
538  ctl.entrysize = sizeof(RelFileNode);
540  tmphash = hash_create("tmp relfilenodes",
541  hash_get_num_entries(pendingSyncHash), &ctl,
543 
544  /* collect all rnodes from pending syncs */
545  hash_seq_init(&scan, pendingSyncHash);
546  while ((sync = (PendingRelSync *) hash_seq_search(&scan)))
547  (void) hash_search(tmphash, &sync->rnode, HASH_ENTER, NULL);
548 
549  /* remove deleted rnodes */
550  for (delete = pendingDeletes; delete != NULL; delete = delete->next)
551  if (delete->atCommit)
552  (void) hash_search(tmphash, (void *) &delete->relnode,
553  HASH_REMOVE, NULL);
554 
555  hash_seq_init(&scan, tmphash);
556  while ((src = (RelFileNode *) hash_seq_search(&scan)))
557  *dest++ = *src;
558 
559  hash_destroy(tmphash);
560 
561 terminate:
562  MemSet(dest, 0, sizeof(RelFileNode));
563 }
564 
565 /*
566  * RestorePendingSyncs
567  * Restore syncs within a parallel worker.
568  *
569  * RelationNeedsWAL() and RelFileNodeSkippingWAL() must offer the correct
570  * answer to parallel workers. Only smgrDoPendingSyncs() reads the
571  * is_truncated field, at end of transaction. Hence, don't restore it.
572  */
573 void
574 RestorePendingSyncs(char *startAddress)
575 {
576  RelFileNode *rnode;
577 
578  Assert(pendingSyncHash == NULL);
579  for (rnode = (RelFileNode *) startAddress; rnode->relNode != 0; rnode++)
580  AddPendingSync(rnode);
581 }
582 
583 /*
584  * smgrDoPendingDeletes() -- Take care of relation deletes at end of xact.
585  *
586  * This also runs when aborting a subxact; we want to clean up a failed
587  * subxact immediately.
588  *
589  * Note: It's possible that we're being asked to remove a relation that has
590  * no physical storage in any fork. In particular, it's possible that we're
591  * cleaning up an old temporary relation for which RemovePgTempFiles has
592  * already recovered the physical storage.
593  */
594 void
595 smgrDoPendingDeletes(bool isCommit)
596 {
598  PendingRelDelete *pending;
599  PendingRelDelete *prev;
601  int nrels = 0,
602  maxrels = 0;
603  SMgrRelation *srels = NULL;
604 
605  prev = NULL;
606  for (pending = pendingDeletes; pending != NULL; pending = next)
607  {
608  next = pending->next;
609  if (pending->nestLevel < nestLevel)
610  {
611  /* outer-level entries should not be processed yet */
612  prev = pending;
613  }
614  else
615  {
616  /* unlink list entry first, so we don't retry on failure */
617  if (prev)
618  prev->next = next;
619  else
620  pendingDeletes = next;
621  /* do deletion if called for */
622  if (pending->atCommit == isCommit)
623  {
624  SMgrRelation srel;
625 
626  srel = smgropen(pending->relnode, pending->backend);
627 
628  /* allocate the initial array, or extend it, if needed */
629  if (maxrels == 0)
630  {
631  maxrels = 8;
632  srels = palloc(sizeof(SMgrRelation) * maxrels);
633  }
634  else if (maxrels <= nrels)
635  {
636  maxrels *= 2;
637  srels = repalloc(srels, sizeof(SMgrRelation) * maxrels);
638  }
639 
640  srels[nrels++] = srel;
641  }
642  /* must explicitly free the list entry */
643  pfree(pending);
644  /* prev does not change */
645  }
646  }
647 
648  if (nrels > 0)
649  {
650  smgrdounlinkall(srels, nrels, false);
651 
652  for (int i = 0; i < nrels; i++)
653  smgrclose(srels[i]);
654 
655  pfree(srels);
656  }
657 }
658 
659 /*
660  * smgrDoPendingSyncs() -- Take care of relation syncs at end of xact.
661  */
662 void
663 smgrDoPendingSyncs(bool isCommit, bool isParallelWorker)
664 {
665  PendingRelDelete *pending;
666  int nrels = 0,
667  maxrels = 0;
668  SMgrRelation *srels = NULL;
669  HASH_SEQ_STATUS scan;
670  PendingRelSync *pendingsync;
671 
673 
674  if (!pendingSyncHash)
675  return; /* no relation needs sync */
676 
677  /* Abort -- just throw away all pending syncs */
678  if (!isCommit)
679  {
680  pendingSyncHash = NULL;
681  return;
682  }
683 
685 
686  /* Parallel worker -- just throw away all pending syncs */
687  if (isParallelWorker)
688  {
689  pendingSyncHash = NULL;
690  return;
691  }
692 
693  /* Skip syncing nodes that smgrDoPendingDeletes() will delete. */
694  for (pending = pendingDeletes; pending != NULL; pending = pending->next)
695  if (pending->atCommit)
696  (void) hash_search(pendingSyncHash, (void *) &pending->relnode,
697  HASH_REMOVE, NULL);
698 
699  hash_seq_init(&scan, pendingSyncHash);
700  while ((pendingsync = (PendingRelSync *) hash_seq_search(&scan)))
701  {
702  ForkNumber fork;
703  BlockNumber nblocks[MAX_FORKNUM + 1];
704  BlockNumber total_blocks = 0;
705  SMgrRelation srel;
706 
707  srel = smgropen(pendingsync->rnode, InvalidBackendId);
708 
709  /*
710  * We emit newpage WAL records for smaller relations.
711  *
712  * Small WAL records have a chance to be emitted along with other
713  * backends' WAL records. We emit WAL records instead of syncing for
714  * files that are smaller than a certain threshold, expecting faster
715  * commit. The threshold is defined by the GUC wal_skip_threshold.
716  */
717  if (!pendingsync->is_truncated)
718  {
719  for (fork = 0; fork <= MAX_FORKNUM; fork++)
720  {
721  if (smgrexists(srel, fork))
722  {
723  BlockNumber n = smgrnblocks(srel, fork);
724 
725  /* we shouldn't come here for unlogged relations */
726  Assert(fork != INIT_FORKNUM);
727  nblocks[fork] = n;
728  total_blocks += n;
729  }
730  else
731  nblocks[fork] = InvalidBlockNumber;
732  }
733  }
734 
735  /*
736  * Sync file or emit WAL records for its contents.
737  *
738  * Although we emit WAL record if the file is small enough, do file
739  * sync regardless of the size if the file has experienced a
740  * truncation. It is because the file would be followed by trailing
741  * garbage blocks after a crash recovery if, while a past longer file
742  * had been flushed out, we omitted syncing-out of the file and
743  * emitted WAL instead. You might think that we could choose WAL if
744  * the current main fork is longer than ever, but there's a case where
745  * main fork is longer than ever but FSM fork gets shorter.
746  */
747  if (pendingsync->is_truncated ||
748  total_blocks * BLCKSZ / 1024 >= wal_skip_threshold)
749  {
750  /* allocate the initial array, or extend it, if needed */
751  if (maxrels == 0)
752  {
753  maxrels = 8;
754  srels = palloc(sizeof(SMgrRelation) * maxrels);
755  }
756  else if (maxrels <= nrels)
757  {
758  maxrels *= 2;
759  srels = repalloc(srels, sizeof(SMgrRelation) * maxrels);
760  }
761 
762  srels[nrels++] = srel;
763  }
764  else
765  {
766  /* Emit WAL records for all blocks. The file is small enough. */
767  for (fork = 0; fork <= MAX_FORKNUM; fork++)
768  {
769  int n = nblocks[fork];
770  Relation rel;
771 
772  if (!BlockNumberIsValid(n))
773  continue;
774 
775  /*
776  * Emit WAL for the whole file. Unfortunately we don't know
777  * what kind of a page this is, so we have to log the full
778  * page including any unused space. ReadBufferExtended()
779  * counts some pgstat events; unfortunately, we discard them.
780  */
782  log_newpage_range(rel, fork, 0, n, false);
784  }
785  }
786  }
787 
788  pendingSyncHash = NULL;
789 
790  if (nrels > 0)
791  {
792  smgrdosyncall(srels, nrels);
793  pfree(srels);
794  }
795 }
796 
797 /*
798  * smgrGetPendingDeletes() -- Get a list of non-temp relations to be deleted.
799  *
800  * The return value is the number of relations scheduled for termination.
801  * *ptr is set to point to a freshly-palloc'd array of RelFileNodes.
802  * If there are no relations to be deleted, *ptr is set to NULL.
803  *
804  * Only non-temporary relations are included in the returned list. This is OK
805  * because the list is used only in contexts where temporary relations don't
806  * matter: we're either writing to the two-phase state file (and transactions
807  * that have touched temp tables can't be prepared) or we're writing to xlog
808  * (and all temporary files will be zapped if we restart anyway, so no need
809  * for redo to do it also).
810  *
811  * Note that the list does not include anything scheduled for termination
812  * by upper-level transactions.
813  */
814 int
815 smgrGetPendingDeletes(bool forCommit, RelFileNode **ptr)
816 {
818  int nrels;
819  RelFileNode *rptr;
820  PendingRelDelete *pending;
821 
822  nrels = 0;
823  for (pending = pendingDeletes; pending != NULL; pending = pending->next)
824  {
825  if (pending->nestLevel >= nestLevel && pending->atCommit == forCommit
826  && pending->backend == InvalidBackendId)
827  nrels++;
828  }
829  if (nrels == 0)
830  {
831  *ptr = NULL;
832  return 0;
833  }
834  rptr = (RelFileNode *) palloc(nrels * sizeof(RelFileNode));
835  *ptr = rptr;
836  for (pending = pendingDeletes; pending != NULL; pending = pending->next)
837  {
838  if (pending->nestLevel >= nestLevel && pending->atCommit == forCommit
839  && pending->backend == InvalidBackendId)
840  {
841  *rptr = pending->relnode;
842  rptr++;
843  }
844  }
845  return nrels;
846 }
847 
848 /*
849  * PostPrepare_smgr -- Clean up after a successful PREPARE
850  *
851  * What we have to do here is throw away the in-memory state about pending
852  * relation deletes. It's all been recorded in the 2PC state file and
853  * it's no longer smgr's job to worry about it.
854  */
855 void
857 {
858  PendingRelDelete *pending;
860 
861  for (pending = pendingDeletes; pending != NULL; pending = next)
862  {
863  next = pending->next;
864  pendingDeletes = next;
865  /* must explicitly free the list entry */
866  pfree(pending);
867  }
868 }
869 
870 
871 /*
872  * AtSubCommit_smgr() --- Take care of subtransaction commit.
873  *
874  * Reassign all items in the pending-deletes list to the parent transaction.
875  */
876 void
878 {
880  PendingRelDelete *pending;
881 
882  for (pending = pendingDeletes; pending != NULL; pending = pending->next)
883  {
884  if (pending->nestLevel >= nestLevel)
885  pending->nestLevel = nestLevel - 1;
886  }
887 }
888 
889 /*
890  * AtSubAbort_smgr() --- Take care of subtransaction abort.
891  *
892  * Delete created relations and forget about deleted relations.
893  * We can execute these operations immediately because we know this
894  * subtransaction will not commit.
895  */
896 void
898 {
899  smgrDoPendingDeletes(false);
900 }
901 
902 void
904 {
905  XLogRecPtr lsn = record->EndRecPtr;
906  uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
907 
908  /* Backup blocks are not used in smgr records */
909  Assert(!XLogRecHasAnyBlockRefs(record));
910 
911  if (info == XLOG_SMGR_CREATE)
912  {
913  xl_smgr_create *xlrec = (xl_smgr_create *) XLogRecGetData(record);
914  SMgrRelation reln;
915 
916  reln = smgropen(xlrec->rnode, InvalidBackendId);
917  smgrcreate(reln, xlrec->forkNum, true);
918  }
919  else if (info == XLOG_SMGR_TRUNCATE)
920  {
921  xl_smgr_truncate *xlrec = (xl_smgr_truncate *) XLogRecGetData(record);
922  SMgrRelation reln;
923  Relation rel;
924  ForkNumber forks[MAX_FORKNUM];
926  int nforks = 0;
927  bool need_fsm_vacuum = false;
928 
929  reln = smgropen(xlrec->rnode, InvalidBackendId);
930 
931  /*
932  * Forcibly create relation if it doesn't exist (which suggests that
933  * it was dropped somewhere later in the WAL sequence). As in
934  * XLogReadBufferForRedo, we prefer to recreate the rel and replay the
935  * log as best we can until the drop is seen.
936  */
937  smgrcreate(reln, MAIN_FORKNUM, true);
938 
939  /*
940  * Before we perform the truncation, update minimum recovery point to
941  * cover this WAL record. Once the relation is truncated, there's no
942  * going back. The buffer manager enforces the WAL-first rule for
943  * normal updates to relation files, so that the minimum recovery
944  * point is always updated before the corresponding change in the data
945  * file is flushed to disk. We have to do the same manually here.
946  *
947  * Doing this before the truncation means that if the truncation fails
948  * for some reason, you cannot start up the system even after restart,
949  * until you fix the underlying situation so that the truncation will
950  * succeed. Alternatively, we could update the minimum recovery point
951  * after truncation, but that would leave a small window where the
952  * WAL-first rule could be violated.
953  */
954  XLogFlush(lsn);
955 
956  /* Prepare for truncation of MAIN fork */
957  if ((xlrec->flags & SMGR_TRUNCATE_HEAP) != 0)
958  {
959  forks[nforks] = MAIN_FORKNUM;
960  blocks[nforks] = xlrec->blkno;
961  nforks++;
962 
963  /* Also tell xlogutils.c about it */
964  XLogTruncateRelation(xlrec->rnode, MAIN_FORKNUM, xlrec->blkno);
965  }
966 
967  /* Prepare for truncation of FSM and VM too */
968  rel = CreateFakeRelcacheEntry(xlrec->rnode);
969 
970  if ((xlrec->flags & SMGR_TRUNCATE_FSM) != 0 &&
971  smgrexists(reln, FSM_FORKNUM))
972  {
973  blocks[nforks] = FreeSpaceMapPrepareTruncateRel(rel, xlrec->blkno);
974  if (BlockNumberIsValid(blocks[nforks]))
975  {
976  forks[nforks] = FSM_FORKNUM;
977  nforks++;
978  need_fsm_vacuum = true;
979  }
980  }
981  if ((xlrec->flags & SMGR_TRUNCATE_VM) != 0 &&
983  {
984  blocks[nforks] = visibilitymap_prepare_truncate(rel, xlrec->blkno);
985  if (BlockNumberIsValid(blocks[nforks]))
986  {
987  forks[nforks] = VISIBILITYMAP_FORKNUM;
988  nforks++;
989  }
990  }
991 
992  /* Do the real work to truncate relation forks */
993  if (nforks > 0)
994  smgrtruncate(reln, forks, nforks, blocks);
995 
996  /*
997  * Update upper-level FSM pages to account for the truncation. This is
998  * important because the just-truncated pages were likely marked as
999  * all-free, and would be preferentially selected.
1000  */
1001  if (need_fsm_vacuum)
1002  FreeSpaceMapVacuumRange(rel, xlrec->blkno,
1004 
1005  FreeFakeRelcacheEntry(rel);
1006  }
1007  else
1008  elog(PANIC, "smgr_redo: unknown op code %u", info);
1009 }
void RelationPreserveStorage(RelFileNode rnode, bool atCommit)
Definition: storage.c:240
void XLogTruncateRelation(RelFileNode rnode, ForkNumber forkNum, BlockNumber nblocks)
Definition: xlogutils.c:644
void smgrDoPendingDeletes(bool isCommit)
Definition: storage.c:595
void hash_destroy(HTAB *hashp)
Definition: dynahash.c:827
BackendId backend
Definition: storage.c:63
void smgrclose(SMgrRelation reln)
Definition: smgr.c:257
#define SMGR_TRUNCATE_HEAP
Definition: storage_xlog.h:40
void smgrcreate(SMgrRelation reln, ForkNumber forknum, bool isRedo)
Definition: smgr.c:334
MemoryContext TopTransactionContext
Definition: mcxt.c:49
#define XLR_SPECIAL_REL_UPDATE
Definition: xlogrecord.h:71
#define HASH_CONTEXT
Definition: hsearch.h:91
#define HASH_ELEM
Definition: hsearch.h:85
void RelationPreTruncate(Relation rel)
Definition: storage.c:386
bool PageIsVerified(Page page, BlockNumber blkno)
Definition: bufpage.c:82
MemoryContext hcxt
Definition: hsearch.h:77
struct PendingRelSync PendingRelSync
struct PendingRelDelete PendingRelDelete
struct SMgrRelationData * rd_smgr
Definition: rel.h:57
bool is_truncated
Definition: storage.c:72
#define XLogIsNeeded()
Definition: xlog.h:191
unsigned char uint8
Definition: c.h:372
#define SMGR_TRUNCATE_FSM
Definition: storage_xlog.h:42
void AtSubCommit_smgr(void)
Definition: storage.c:877
Size entrysize
Definition: hsearch.h:72
int errcode(int sqlerrcode)
Definition: elog.c:610
void smgrdounlinkall(SMgrRelation *rels, int nrels, bool isRedo)
Definition: smgr.c:385
RelFileNode rnode
Definition: storage.c:71
#define MemSet(start, val, len)
Definition: c.h:949
#define RelationCloseSmgr(relation)
Definition: rel.h:525
long hash_get_num_entries(HTAB *hashp)
Definition: dynahash.c:1347
uint32 BlockNumber
Definition: block.h:31
RelFileNode rnode
Definition: storage_xlog.h:49
bool smgrexists(SMgrRelation reln, ForkNumber forknum)
Definition: smgr.c:248
void * hash_search(HTAB *hashp, const void *keyPtr, HASHACTION action, bool *foundPtr)
Definition: dynahash.c:919
#define PANIC
Definition: elog.h:53
void XLogFlush(XLogRecPtr record)
Definition: xlog.c:2845
BlockNumber smgr_cached_nblocks[MAX_FORKNUM+1]
Definition: smgr.h:54
#define XLOG_SMGR_CREATE
Definition: storage_xlog.h:30
XLogRecPtr EndRecPtr
Definition: xlogreader.h:176
void smgrread(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, char *buffer)
Definition: smgr.c:502
void RestorePendingSyncs(char *startAddress)
Definition: storage.c:574
#define RelationOpenSmgr(relation)
Definition: rel.h:513
static void AddPendingSync(const RelFileNode *rnode)
Definition: storage.c:84
Definition: dynahash.c:218
char data[BLCKSZ]
Definition: c.h:1082
void pfree(void *pointer)
Definition: mcxt.c:1057
SMgrRelation RelationCreateStorage(RelFileNode rnode, char relpersistence)
Definition: storage.c:118
bool IsInParallelMode(void)
Definition: xact.c:1012
#define XLogRecGetData(decoder)
Definition: xlogreader.h:310
HTAB * pendingSyncHash
Definition: storage.c:76
#define ERROR
Definition: elog.h:43
void smgr_redo(XLogReaderState *record)
Definition: storage.c:903
#define AssertPendingSyncs_RelationCache()
Definition: relcache.h:131
RelFileNodeBackend smgr_rnode
Definition: smgr.h:42
int smgrGetPendingDeletes(bool forCommit, RelFileNode **ptr)
Definition: storage.c:815
Relation CreateFakeRelcacheEntry(RelFileNode rnode)
Definition: xlogutils.c:556
#define BackendIdForTempRelations()
Definition: backendid.h:34
struct RelFileNode RelFileNode
static char * buf
Definition: pg_test_fsync.c:68
BlockNumber FreeSpaceMapPrepareTruncateRel(Relation rel, BlockNumber nblocks)
Definition: freespace.c:261
MemoryContext CurrentMemoryContext
Definition: mcxt.c:38
#define XLogRecGetInfo(decoder)
Definition: xlogreader.h:305
MemoryContext TopMemoryContext
Definition: mcxt.c:44
SMgrRelation smgropen(RelFileNode rnode, BackendId backend)
Definition: smgr.c:146
ForkNumber
Definition: relpath.h:40
#define ERRCODE_DATA_CORRUPTED
Definition: pg_basebackup.c:45
#define XLOG_SMGR_TRUNCATE
Definition: storage_xlog.h:31
void FreeFakeRelcacheEntry(Relation fakerel)
Definition: xlogutils.c:599
struct PendingRelDelete * next
Definition: storage.c:66
void XLogRegisterData(char *data, int len)
Definition: xloginsert.c:330
XLogRecPtr XLogInsert(RmgrId rmid, uint8 info)
Definition: xloginsert.c:422
#define HASH_BLOBS
Definition: hsearch.h:86
Size mul_size(Size s1, Size s2)
Definition: shmem.c:515
#define InvalidBackendId
Definition: backendid.h:23
void RelationDropStorage(Relation rel)
Definition: storage.c:195
HTAB * hash_create(const char *tabname, long nelem, HASHCTL *info, int flags)
Definition: dynahash.c:326
Size EstimatePendingSyncsSpace(void)
Definition: storage.c:510
void PostPrepare_smgr(void)
Definition: storage.c:856
int BackendId
Definition: backendid.h:21
void SerializePendingSyncs(Size maxSize, char *startAddress)
Definition: storage.c:523
Size keysize
Definition: hsearch.h:71
void smgrtruncate(SMgrRelation reln, ForkNumber *forknum, int nforks, BlockNumber *nblocks)
Definition: smgr.c:578
void RelationCopyStorage(SMgrRelation src, SMgrRelation dst, ForkNumber forkNum, char relpersistence)
Definition: storage.c:408
#define ereport(elevel,...)
Definition: elog.h:144
#define BlockNumberIsValid(blockNumber)
Definition: block.h:70
RelFileNode node
Definition: relfilenode.h:74
int GetCurrentTransactionNestLevel(void)
Definition: xact.c:857
RelFileNode rd_node
Definition: rel.h:55
ForkNumber forkNum
Definition: storage_xlog.h:36
BlockNumber smgrnblocks(SMgrRelation reln, ForkNumber forknum)
Definition: smgr.c:549
uint64 XLogRecPtr
Definition: xlogdefs.h:21
BackendId rd_backend
Definition: rel.h:59
#define Assert(condition)
Definition: c.h:745
#define XLR_INFO_MASK
Definition: xlogrecord.h:62
int wal_skip_threshold
Definition: storage.c:38
BlockNumber smgr_targblock
Definition: smgr.h:53
void AtSubAbort_smgr(void)
Definition: storage.c:897
size_t Size
Definition: c.h:473
BackendId backend
Definition: relfilenode.h:75
#define InvalidBlockNumber
Definition: block.h:33
#define SMGR_TRUNCATE_VM
Definition: storage_xlog.h:41
#define MAX_FORKNUM
Definition: relpath.h:55
void PageSetChecksumInplace(Page page, BlockNumber blkno)
Definition: bufpage.c:1414
void * hash_seq_search(HASH_SEQ_STATUS *status)
Definition: dynahash.c:1401
#define RelationNeedsWAL(relation)
Definition: rel.h:562
bool RelFileNodeSkippingWAL(RelFileNode rnode)
Definition: storage.c:496
void * repalloc(void *pointer, Size size)
Definition: mcxt.c:1070
void hash_seq_init(HASH_SEQ_STATUS *status, HTAB *hashp)
Definition: dynahash.c:1391
static int64 blocks
Definition: pg_checksums.c:35
void smgrextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, char *buffer, bool skipFsync)
Definition: smgr.c:463
void log_newpage_range(Relation rel, ForkNumber forkNum, BlockNumber startblk, BlockNumber endblk, bool page_std)
Definition: xloginsert.c:1123
static PendingRelDelete * pendingDeletes
Definition: storage.c:75
void * palloc(Size size)
Definition: mcxt.c:950
int errmsg(const char *fmt,...)
Definition: elog.c:824
RelFileNode relnode
Definition: storage.c:62
void * MemoryContextAlloc(MemoryContext context, Size size)
Definition: mcxt.c:797
XLogRecPtr log_newpage(RelFileNode *rnode, ForkNumber forkNum, BlockNumber blkno, Page page, bool page_std)
Definition: xloginsert.c:996
#define elog(elevel,...)
Definition: elog.h:214
int i
BlockNumber visibilitymap_prepare_truncate(Relation rel, BlockNumber nheapblocks)
#define XLogRecHasAnyBlockRefs(decoder)
Definition: xlogreader.h:312
BlockNumber blkno
Definition: storage_xlog.h:48
#define CHECK_FOR_INTERRUPTS()
Definition: miscadmin.h:99
void smgrDoPendingSyncs(bool isCommit, bool isParallelWorker)
Definition: storage.c:663
#define SMGR_TRUNCATE_ALL
Definition: storage_xlog.h:43
void XLogBeginInsert(void)
Definition: xloginsert.c:123
#define relpathbackend(rnode, backend, forknum)
Definition: relpath.h:78
void smgrimmedsync(SMgrRelation reln, ForkNumber forknum)
Definition: smgr.c:643
void FreeSpaceMapVacuumRange(Relation rel, BlockNumber start, BlockNumber end)
Definition: freespace.c:354
RelFileNode rnode
Definition: storage_xlog.h:35
Pointer Page
Definition: bufpage.h:78
#define RelFileNodeEquals(node1, node2)
Definition: relfilenode.h:88
void RelationTruncate(Relation rel, BlockNumber nblocks)
Definition: storage.c:277
void log_smgrcreate(const RelFileNode *rnode, ForkNumber forkNum)
Definition: storage.c:175
void smgrdosyncall(SMgrRelation *rels, int nrels)
Definition: smgr.c:349