1 /*-------------------------------------------------------------------------
4 * Support functions to rewrite tables.
6 * These functions provide a facility to completely rewrite a heap, while
7 * preserving visibility information and update chains.
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.
16 * To use the facility:
19 * while (fetch next tuple)
22 * rewrite_heap_dead_tuple
25 * // do any transformations here if required
31 * The contents of the new relation shouldn't be relied on until after
32 * end_heap_rewrite is called.
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.
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.)
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
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.
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.
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.
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
91 * directly through smgr. Note, however, that any data sent to the new
92 * heap's TOAST table will go through the normal bufmgr.
95 * Portions Copyright (c) 1996-2013, PostgreSQL Global Development Group
96 * Portions Copyright (c) 1994-5, Regents of the University of California
99 * src/backend/access/heap/rewriteheap.c
101 *-------------------------------------------------------------------------
103 #include "postgres.h"
105 #include "access/heapam.h"
106 #include "access/heapam_xlog.h"
107 #include "access/rewriteheap.h"
108 #include "access/transam.h"
109 #include "access/tuptoaster.h"
110 #include "storage/bufmgr.h"
111 #include "storage/smgr.h"
112 #include "utils/memutils.h"
113 #include "utils/rel.h"
114 #include "utils/tqual.h"
118 * State associated with a rewrite operation. This is opaque to the user
119 * of the rewrite facility.
121 typedef struct RewriteStateData
123 Relation rs_new_rel; /* destination heap */
124 Page rs_buffer; /* page currently being built */
125 BlockNumber rs_blockno; /* block where page will go */
126 bool rs_buffer_valid; /* T if any tuples in buffer */
127 bool rs_use_wal; /* must we WAL-log inserts? */
128 TransactionId rs_oldest_xmin; /* oldest xmin used by caller to
129 * determine tuple visibility */
130 TransactionId rs_freeze_xid;/* Xid that will be used as freeze cutoff
132 MultiXactId rs_freeze_multi;/* MultiXactId that will be used as freeze
133 * cutoff point for multixacts */
134 MemoryContext rs_cxt; /* for hash tables and entries and tuples in
136 HTAB *rs_unresolved_tups; /* unmatched A tuples */
137 HTAB *rs_old_new_tid_map; /* unmatched B tuples */
141 * The lookup keys for the hash tables are tuple TID and xmin (we must check
142 * both to avoid false matches from dead tuples). Beware that there is
143 * probably some padding space in this struct; it must be zeroed out for
144 * correct hashtable operation.
148 TransactionId xmin; /* tuple xmin */
149 ItemPointerData tid; /* tuple location in old heap */
153 * Entry structures for the hash tables
157 TidHashKey key; /* expected xmin/old location of B tuple */
158 ItemPointerData old_tid; /* A's location in the old heap */
159 HeapTuple tuple; /* A's tuple contents */
162 typedef UnresolvedTupData *UnresolvedTup;
166 TidHashKey key; /* actual xmin/old location of B tuple */
167 ItemPointerData new_tid; /* where we put it in the new heap */
168 } OldToNewMappingData;
170 typedef OldToNewMappingData *OldToNewMapping;
173 /* prototypes for internal functions */
174 static void raw_heap_insert(RewriteState state, HeapTuple tup);
178 * Begin a rewrite of a table
180 * new_heap new, locked heap relation to insert tuples to
181 * oldest_xmin xid used by the caller to determine which tuples are dead
182 * freeze_xid xid before which tuples will be frozen
183 * freeze_multi multixact before which multis will be frozen
184 * use_wal should the inserts to the new heap be WAL-logged?
186 * Returns an opaque RewriteState, allocated in current memory context,
187 * to be used in subsequent calls to the other functions.
190 begin_heap_rewrite(Relation new_heap, TransactionId oldest_xmin,
191 TransactionId freeze_xid, MultiXactId freeze_multi,
195 MemoryContext rw_cxt;
196 MemoryContext old_cxt;
200 * To ease cleanup, make a separate context that will contain the
201 * RewriteState struct itself plus all subsidiary data.
203 rw_cxt = AllocSetContextCreate(CurrentMemoryContext,
205 ALLOCSET_DEFAULT_MINSIZE,
206 ALLOCSET_DEFAULT_INITSIZE,
207 ALLOCSET_DEFAULT_MAXSIZE);
208 old_cxt = MemoryContextSwitchTo(rw_cxt);
210 /* Create and fill in the state struct */
211 state = palloc0(sizeof(RewriteStateData));
213 state->rs_new_rel = new_heap;
214 state->rs_buffer = (Page) palloc(BLCKSZ);
215 /* new_heap needn't be empty, just locked */
216 state->rs_blockno = RelationGetNumberOfBlocks(new_heap);
217 state->rs_buffer_valid = false;
218 state->rs_use_wal = use_wal;
219 state->rs_oldest_xmin = oldest_xmin;
220 state->rs_freeze_xid = freeze_xid;
221 state->rs_freeze_multi = freeze_multi;
222 state->rs_cxt = rw_cxt;
224 /* Initialize hash tables used to track update chains */
225 memset(&hash_ctl, 0, sizeof(hash_ctl));
226 hash_ctl.keysize = sizeof(TidHashKey);
227 hash_ctl.entrysize = sizeof(UnresolvedTupData);
228 hash_ctl.hcxt = state->rs_cxt;
229 hash_ctl.hash = tag_hash;
231 state->rs_unresolved_tups =
232 hash_create("Rewrite / Unresolved ctids",
233 128, /* arbitrary initial size */
235 HASH_ELEM | HASH_FUNCTION | HASH_CONTEXT);
237 hash_ctl.entrysize = sizeof(OldToNewMappingData);
239 state->rs_old_new_tid_map =
240 hash_create("Rewrite / Old to new tid map",
241 128, /* arbitrary initial size */
243 HASH_ELEM | HASH_FUNCTION | HASH_CONTEXT);
245 MemoryContextSwitchTo(old_cxt);
253 * state and any other resources are freed.
256 end_heap_rewrite(RewriteState state)
258 HASH_SEQ_STATUS seq_status;
259 UnresolvedTup unresolved;
262 * Write any remaining tuples in the UnresolvedTups table. If we have any
263 * left, they should in fact be dead, but let's err on the safe side.
265 hash_seq_init(&seq_status, state->rs_unresolved_tups);
267 while ((unresolved = hash_seq_search(&seq_status)) != NULL)
269 ItemPointerSetInvalid(&unresolved->tuple->t_data->t_ctid);
270 raw_heap_insert(state, unresolved->tuple);
273 /* Write the last page, if any */
274 if (state->rs_buffer_valid)
276 if (state->rs_use_wal)
277 log_newpage(&state->rs_new_rel->rd_node,
281 RelationOpenSmgr(state->rs_new_rel);
282 smgrextend(state->rs_new_rel->rd_smgr, MAIN_FORKNUM, state->rs_blockno,
283 (char *) state->rs_buffer, true);
287 * If the rel is WAL-logged, must fsync before commit. We use heap_sync
288 * to ensure that the toast table gets fsync'd too.
290 * It's obvious that we must do this when not WAL-logging. It's less
291 * obvious that we have to do it even if we did WAL-log the pages. The
292 * reason is the same as in tablecmds.c's copy_relation_data(): we're
293 * writing data that's not in shared buffers, and so a CHECKPOINT
294 * occurring during the rewriteheap operation won't have fsync'd data we
295 * wrote before the checkpoint.
297 if (RelationNeedsWAL(state->rs_new_rel))
298 heap_sync(state->rs_new_rel);
300 /* Deleting the context frees everything */
301 MemoryContextDelete(state->rs_cxt);
305 * Add a tuple to the new heap.
307 * Visibility information is copied from the original tuple, except that
308 * we "freeze" very-old tuples. Note that since we scribble on new_tuple,
309 * it had better be temp storage not a pointer to the original tuple.
311 * state opaque state as returned by begin_heap_rewrite
312 * old_tuple original tuple in the old heap
313 * new_tuple new, rewritten tuple to be inserted to new heap
316 rewrite_heap_tuple(RewriteState state,
317 HeapTuple old_tuple, HeapTuple new_tuple)
319 MemoryContext old_cxt;
320 ItemPointerData old_tid;
325 old_cxt = MemoryContextSwitchTo(state->rs_cxt);
328 * Copy the original tuple's visibility information into new_tuple.
330 * XXX we might later need to copy some t_infomask2 bits, too? Right now,
331 * we intentionally clear the HOT status bits.
333 memcpy(&new_tuple->t_data->t_choice.t_heap,
334 &old_tuple->t_data->t_choice.t_heap,
335 sizeof(HeapTupleFields));
337 new_tuple->t_data->t_infomask &= ~HEAP_XACT_MASK;
338 new_tuple->t_data->t_infomask2 &= ~HEAP2_XACT_MASK;
339 new_tuple->t_data->t_infomask |=
340 old_tuple->t_data->t_infomask & HEAP_XACT_MASK;
343 * While we have our hands on the tuple, we may as well freeze any
344 * very-old xmin or xmax, so that future VACUUM effort can be saved.
346 heap_freeze_tuple(new_tuple->t_data, state->rs_freeze_xid,
347 state->rs_freeze_multi);
350 * Invalid ctid means that ctid should point to the tuple itself. We'll
351 * override it later if the tuple is part of an update chain.
353 ItemPointerSetInvalid(&new_tuple->t_data->t_ctid);
356 * If the tuple has been updated, check the old-to-new mapping hash table.
358 if (!((old_tuple->t_data->t_infomask & HEAP_XMAX_INVALID) ||
359 HeapTupleHeaderIsOnlyLocked(old_tuple->t_data)) &&
360 !(ItemPointerEquals(&(old_tuple->t_self),
361 &(old_tuple->t_data->t_ctid))))
363 OldToNewMapping mapping;
365 memset(&hashkey, 0, sizeof(hashkey));
366 hashkey.xmin = HeapTupleHeaderGetUpdateXid(old_tuple->t_data);
367 hashkey.tid = old_tuple->t_data->t_ctid;
369 mapping = (OldToNewMapping)
370 hash_search(state->rs_old_new_tid_map, &hashkey,
376 * We've already copied the tuple that t_ctid points to, so we can
377 * set the ctid of this tuple to point to the new location, and
378 * insert it right away.
380 new_tuple->t_data->t_ctid = mapping->new_tid;
382 /* We don't need the mapping entry anymore */
383 hash_search(state->rs_old_new_tid_map, &hashkey,
384 HASH_REMOVE, &found);
390 * We haven't seen the tuple t_ctid points to yet. Stash this
391 * tuple into unresolved_tups to be written later.
393 UnresolvedTup unresolved;
395 unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
399 unresolved->old_tid = old_tuple->t_self;
400 unresolved->tuple = heap_copytuple(new_tuple);
403 * We can't do anything more now, since we don't know where the
404 * tuple will be written.
406 MemoryContextSwitchTo(old_cxt);
412 * Now we will write the tuple, and then check to see if it is the B tuple
413 * in any new or known pair. When we resolve a known pair, we will be
414 * able to write that pair's A tuple, and then we have to check if it
415 * resolves some other pair. Hence, we need a loop here.
417 old_tid = old_tuple->t_self;
422 ItemPointerData new_tid;
424 /* Insert the tuple and find out where it's put in new_heap */
425 raw_heap_insert(state, new_tuple);
426 new_tid = new_tuple->t_self;
429 * If the tuple is the updated version of a row, and the prior version
430 * wouldn't be DEAD yet, then we need to either resolve the prior
431 * version (if it's waiting in rs_unresolved_tups), or make an entry
432 * in rs_old_new_tid_map (so we can resolve it when we do see it). The
433 * previous tuple's xmax would equal this one's xmin, so it's
434 * RECENTLY_DEAD if and only if the xmin is not before OldestXmin.
436 if ((new_tuple->t_data->t_infomask & HEAP_UPDATED) &&
437 !TransactionIdPrecedes(HeapTupleHeaderGetXmin(new_tuple->t_data),
438 state->rs_oldest_xmin))
441 * Okay, this is B in an update pair. See if we've seen A.
443 UnresolvedTup unresolved;
445 memset(&hashkey, 0, sizeof(hashkey));
446 hashkey.xmin = HeapTupleHeaderGetXmin(new_tuple->t_data);
447 hashkey.tid = old_tid;
449 unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
452 if (unresolved != NULL)
455 * We have seen and memorized the previous tuple already. Now
456 * that we know where we inserted the tuple its t_ctid points
457 * to, fix its t_ctid and insert it to the new heap.
460 heap_freetuple(new_tuple);
461 new_tuple = unresolved->tuple;
463 old_tid = unresolved->old_tid;
464 new_tuple->t_data->t_ctid = new_tid;
467 * We don't need the hash entry anymore, but don't free its
470 hash_search(state->rs_unresolved_tups, &hashkey,
471 HASH_REMOVE, &found);
474 /* loop back to insert the previous tuple in the chain */
480 * Remember the new tid of this tuple. We'll use it to set the
481 * ctid when we find the previous tuple in the chain.
483 OldToNewMapping mapping;
485 mapping = hash_search(state->rs_old_new_tid_map, &hashkey,
489 mapping->new_tid = new_tid;
493 /* Done with this (chain of) tuples, for now */
495 heap_freetuple(new_tuple);
499 MemoryContextSwitchTo(old_cxt);
503 * Register a dead tuple with an ongoing rewrite. Dead tuples are not
504 * copied to the new table, but we still make note of them so that we
505 * can release some resources earlier.
507 * Returns true if a tuple was removed from the unresolved_tups table.
508 * This indicates that that tuple, previously thought to be "recently dead",
509 * is now known really dead and won't be written to the output.
512 rewrite_heap_dead_tuple(RewriteState state, HeapTuple old_tuple)
515 * If we have already seen an earlier tuple in the update chain that
516 * points to this tuple, let's forget about that earlier tuple. It's in
517 * fact dead as well, our simple xmax < OldestXmin test in
518 * HeapTupleSatisfiesVacuum just wasn't enough to detect it. It happens
519 * when xmin of a tuple is greater than xmax, which sounds
520 * counter-intuitive but is perfectly valid.
522 * We don't bother to try to detect the situation the other way round,
523 * when we encounter the dead tuple first and then the recently dead one
524 * that points to it. If that happens, we'll have some unmatched entries
525 * in the UnresolvedTups hash table at the end. That can happen anyway,
526 * because a vacuum might have removed the dead tuple in the chain before
529 UnresolvedTup unresolved;
533 memset(&hashkey, 0, sizeof(hashkey));
534 hashkey.xmin = HeapTupleHeaderGetXmin(old_tuple->t_data);
535 hashkey.tid = old_tuple->t_self;
537 unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
540 if (unresolved != NULL)
542 /* Need to free the contained tuple as well as the hashtable entry */
543 heap_freetuple(unresolved->tuple);
544 hash_search(state->rs_unresolved_tups, &hashkey,
545 HASH_REMOVE, &found);
554 * Insert a tuple to the new relation. This has to track heap_insert
555 * and its subsidiary functions!
557 * t_self of the tuple is set to the new TID of the tuple. If t_ctid of the
558 * tuple is invalid on entry, it's replaced with the new TID as well (in
559 * the inserted data only, not in the caller's copy).
562 raw_heap_insert(RewriteState state, HeapTuple tup)
564 Page page = state->rs_buffer;
572 * If the new tuple is too big for storage or contains already toasted
573 * out-of-line attributes from some other relation, invoke the toaster.
575 * Note: below this point, heaptup is the data we actually intend to store
576 * into the relation; tup is the caller's original untoasted data.
578 if (state->rs_new_rel->rd_rel->relkind == RELKIND_TOASTVALUE)
580 /* toast table entries should never be recursively toasted */
581 Assert(!HeapTupleHasExternal(tup));
584 else if (HeapTupleHasExternal(tup) || tup->t_len > TOAST_TUPLE_THRESHOLD)
585 heaptup = toast_insert_or_update(state->rs_new_rel, tup, NULL,
586 HEAP_INSERT_SKIP_FSM |
588 0 : HEAP_INSERT_SKIP_WAL));
592 len = MAXALIGN(heaptup->t_len); /* be conservative */
595 * If we're gonna fail for oversize tuple, do it right away
597 if (len > MaxHeapTupleSize)
599 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
600 errmsg("row is too big: size %lu, maximum size %lu",
602 (unsigned long) MaxHeapTupleSize)));
604 /* Compute desired extra freespace due to fillfactor option */
605 saveFreeSpace = RelationGetTargetPageFreeSpace(state->rs_new_rel,
606 HEAP_DEFAULT_FILLFACTOR);
608 /* Now we can check to see if there's enough free space already. */
609 if (state->rs_buffer_valid)
611 pageFreeSpace = PageGetHeapFreeSpace(page);
613 if (len + saveFreeSpace > pageFreeSpace)
615 /* Doesn't fit, so write out the existing page */
618 if (state->rs_use_wal)
619 log_newpage(&state->rs_new_rel->rd_node,
625 * Now write the page. We say isTemp = true even if it's not a
626 * temp table, because there's no need for smgr to schedule an
627 * fsync for this write; we'll do it ourselves in
630 RelationOpenSmgr(state->rs_new_rel);
631 smgrextend(state->rs_new_rel->rd_smgr, MAIN_FORKNUM,
632 state->rs_blockno, (char *) page, true);
635 state->rs_buffer_valid = false;
639 if (!state->rs_buffer_valid)
641 /* Initialize a new empty page */
642 PageInit(page, BLCKSZ, 0);
643 state->rs_buffer_valid = true;
646 /* And now we can insert the tuple into the page */
647 newoff = PageAddItem(page, (Item) heaptup->t_data, heaptup->t_len,
648 InvalidOffsetNumber, false, true);
649 if (newoff == InvalidOffsetNumber)
650 elog(ERROR, "failed to add tuple");
652 /* Update caller's t_self to the actual position where it was stored */
653 ItemPointerSet(&(tup->t_self), state->rs_blockno, newoff);
656 * Insert the correct position into CTID of the stored tuple, too, if the
657 * caller didn't supply a valid CTID.
659 if (!ItemPointerIsValid(&tup->t_data->t_ctid))
662 HeapTupleHeader onpage_tup;
664 newitemid = PageGetItemId(page, newoff);
665 onpage_tup = (HeapTupleHeader) PageGetItem(page, newitemid);
667 onpage_tup->t_ctid = tup->t_self;
670 /* If heaptup is a private copy, release it. */
672 heap_freetuple(heaptup);