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) 1996-2007, PostgreSQL Global Development Group
96 * Portions Copyright (c) 1994-5, Regents of the University of California
99 * $PostgreSQL: pgsql/src/backend/access/heap/rewriteheap.c,v 1.8 2007/11/15 21:14:32 momjian Exp $
101 *-------------------------------------------------------------------------
103 #include "postgres.h"
105 #include "access/heapam.h"
106 #include "access/rewriteheap.h"
107 #include "access/transam.h"
108 #include "access/tuptoaster.h"
109 #include "storage/smgr.h"
110 #include "utils/memutils.h"
114 * State associated with a rewrite operation. This is opaque to the user
115 * of the rewrite facility.
117 typedef struct RewriteStateData
119 Relation rs_new_rel; /* destination heap */
120 Page rs_buffer; /* page currently being built */
121 BlockNumber rs_blockno; /* block where page will go */
122 bool rs_buffer_valid; /* T if any tuples in buffer */
123 bool rs_use_wal; /* must we WAL-log inserts? */
124 TransactionId rs_oldest_xmin; /* oldest xmin used by caller to
125 * determine tuple visibility */
126 TransactionId rs_freeze_xid;/* Xid that will be used as freeze cutoff
128 MemoryContext rs_cxt; /* for hash tables and entries and tuples in
130 HTAB *rs_unresolved_tups; /* unmatched A tuples */
131 HTAB *rs_old_new_tid_map; /* unmatched B tuples */
135 * The lookup keys for the hash tables are tuple TID and xmin (we must check
136 * both to avoid false matches from dead tuples). Beware that there is
137 * probably some padding space in this struct; it must be zeroed out for
138 * correct hashtable operation.
142 TransactionId xmin; /* tuple xmin */
143 ItemPointerData tid; /* tuple location in old heap */
147 * Entry structures for the hash tables
151 TidHashKey key; /* expected xmin/old location of B tuple */
152 ItemPointerData old_tid; /* A's location in the old heap */
153 HeapTuple tuple; /* A's tuple contents */
156 typedef UnresolvedTupData *UnresolvedTup;
160 TidHashKey key; /* actual xmin/old location of B tuple */
161 ItemPointerData new_tid; /* where we put it in the new heap */
162 } OldToNewMappingData;
164 typedef OldToNewMappingData *OldToNewMapping;
167 /* prototypes for internal functions */
168 static void raw_heap_insert(RewriteState state, HeapTuple tup);
172 * Begin a rewrite of a table
174 * new_heap new, locked heap relation to insert tuples to
175 * oldest_xmin xid used by the caller to determine which tuples are dead
176 * freeze_xid xid before which tuples will be frozen
177 * use_wal should the inserts to the new heap be WAL-logged?
179 * Returns an opaque RewriteState, allocated in current memory context,
180 * to be used in subsequent calls to the other functions.
183 begin_heap_rewrite(Relation new_heap, TransactionId oldest_xmin,
184 TransactionId freeze_xid, bool use_wal)
187 MemoryContext rw_cxt;
188 MemoryContext old_cxt;
192 * To ease cleanup, make a separate context that will contain the
193 * RewriteState struct itself plus all subsidiary data.
195 rw_cxt = AllocSetContextCreate(CurrentMemoryContext,
197 ALLOCSET_DEFAULT_MINSIZE,
198 ALLOCSET_DEFAULT_INITSIZE,
199 ALLOCSET_DEFAULT_MAXSIZE);
200 old_cxt = MemoryContextSwitchTo(rw_cxt);
202 /* Create and fill in the state struct */
203 state = palloc0(sizeof(RewriteStateData));
205 state->rs_new_rel = new_heap;
206 state->rs_buffer = (Page) palloc(BLCKSZ);
207 /* new_heap needn't be empty, just locked */
208 state->rs_blockno = RelationGetNumberOfBlocks(new_heap);
209 state->rs_buffer_valid = false;
210 state->rs_use_wal = use_wal;
211 state->rs_oldest_xmin = oldest_xmin;
212 state->rs_freeze_xid = freeze_xid;
213 state->rs_cxt = rw_cxt;
215 /* Initialize hash tables used to track update chains */
216 memset(&hash_ctl, 0, sizeof(hash_ctl));
217 hash_ctl.keysize = sizeof(TidHashKey);
218 hash_ctl.entrysize = sizeof(UnresolvedTupData);
219 hash_ctl.hcxt = state->rs_cxt;
220 hash_ctl.hash = tag_hash;
222 state->rs_unresolved_tups =
223 hash_create("Rewrite / Unresolved ctids",
224 128, /* arbitrary initial size */
226 HASH_ELEM | HASH_FUNCTION | HASH_CONTEXT);
228 hash_ctl.entrysize = sizeof(OldToNewMappingData);
230 state->rs_old_new_tid_map =
231 hash_create("Rewrite / Old to new tid map",
232 128, /* arbitrary initial size */
234 HASH_ELEM | HASH_FUNCTION | HASH_CONTEXT);
236 MemoryContextSwitchTo(old_cxt);
244 * state and any other resources are freed.
247 end_heap_rewrite(RewriteState state)
249 HASH_SEQ_STATUS seq_status;
250 UnresolvedTup unresolved;
253 * Write any remaining tuples in the UnresolvedTups table. If we have any
254 * left, they should in fact be dead, but let's err on the safe side.
256 * XXX this really is a waste of code no?
258 hash_seq_init(&seq_status, state->rs_unresolved_tups);
260 while ((unresolved = hash_seq_search(&seq_status)) != NULL)
262 ItemPointerSetInvalid(&unresolved->tuple->t_data->t_ctid);
263 raw_heap_insert(state, unresolved->tuple);
266 /* Write the last page, if any */
267 if (state->rs_buffer_valid)
269 if (state->rs_use_wal)
270 log_newpage(&state->rs_new_rel->rd_node,
273 RelationOpenSmgr(state->rs_new_rel);
274 smgrextend(state->rs_new_rel->rd_smgr, state->rs_blockno,
275 (char *) state->rs_buffer, true);
279 * If the rel isn't temp, must fsync before commit. We use heap_sync to
280 * ensure that the toast table gets fsync'd too.
282 * It's obvious that we must do this when not WAL-logging. It's less
283 * obvious that we have to do it even if we did WAL-log the pages. The
284 * reason is the same as in tablecmds.c's copy_relation_data(): we're
285 * writing data that's not in shared buffers, and so a CHECKPOINT
286 * occurring during the rewriteheap operation won't have fsync'd data we
287 * wrote before the checkpoint.
289 if (!state->rs_new_rel->rd_istemp)
290 heap_sync(state->rs_new_rel);
292 /* Deleting the context frees everything */
293 MemoryContextDelete(state->rs_cxt);
297 * Add a tuple to the new heap.
299 * Visibility information is copied from the original tuple, except that
300 * we "freeze" very-old tuples. Note that since we scribble on new_tuple,
301 * it had better be temp storage not a pointer to the original tuple.
303 * state opaque state as returned by begin_heap_rewrite
304 * old_tuple original tuple in the old heap
305 * new_tuple new, rewritten tuple to be inserted to new heap
308 rewrite_heap_tuple(RewriteState state,
309 HeapTuple old_tuple, HeapTuple new_tuple)
311 MemoryContext old_cxt;
312 ItemPointerData old_tid;
317 old_cxt = MemoryContextSwitchTo(state->rs_cxt);
320 * Copy the original tuple's visibility information into new_tuple.
322 * XXX we might later need to copy some t_infomask2 bits, too? Right now,
323 * we intentionally clear the HOT status bits.
325 memcpy(&new_tuple->t_data->t_choice.t_heap,
326 &old_tuple->t_data->t_choice.t_heap,
327 sizeof(HeapTupleFields));
329 new_tuple->t_data->t_infomask &= ~HEAP_XACT_MASK;
330 new_tuple->t_data->t_infomask2 &= ~HEAP2_XACT_MASK;
331 new_tuple->t_data->t_infomask |=
332 old_tuple->t_data->t_infomask & HEAP_XACT_MASK;
335 * While we have our hands on the tuple, we may as well freeze any
336 * very-old xmin or xmax, so that future VACUUM effort can be saved.
338 * Note we abuse heap_freeze_tuple() a bit here, since it's expecting to
339 * be given a pointer to a tuple in a disk buffer. It happens though that
340 * we can get the right things to happen by passing InvalidBuffer for the
343 heap_freeze_tuple(new_tuple->t_data, state->rs_freeze_xid, InvalidBuffer);
346 * Invalid ctid means that ctid should point to the tuple itself. We'll
347 * override it later if the tuple is part of an update chain.
349 ItemPointerSetInvalid(&new_tuple->t_data->t_ctid);
352 * If the tuple has been updated, check the old-to-new mapping hash table.
354 if (!(old_tuple->t_data->t_infomask & (HEAP_XMAX_INVALID |
356 !(ItemPointerEquals(&(old_tuple->t_self),
357 &(old_tuple->t_data->t_ctid))))
359 OldToNewMapping mapping;
361 memset(&hashkey, 0, sizeof(hashkey));
362 hashkey.xmin = HeapTupleHeaderGetXmax(old_tuple->t_data);
363 hashkey.tid = old_tuple->t_data->t_ctid;
365 mapping = (OldToNewMapping)
366 hash_search(state->rs_old_new_tid_map, &hashkey,
372 * We've already copied the tuple that t_ctid points to, so we can
373 * set the ctid of this tuple to point to the new location, and
374 * insert it right away.
376 new_tuple->t_data->t_ctid = mapping->new_tid;
378 /* We don't need the mapping entry anymore */
379 hash_search(state->rs_old_new_tid_map, &hashkey,
380 HASH_REMOVE, &found);
386 * We haven't seen the tuple t_ctid points to yet. Stash this
387 * tuple into unresolved_tups to be written later.
389 UnresolvedTup unresolved;
391 unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
395 unresolved->old_tid = old_tuple->t_self;
396 unresolved->tuple = heap_copytuple(new_tuple);
399 * We can't do anything more now, since we don't know where the
400 * tuple will be written.
402 MemoryContextSwitchTo(old_cxt);
408 * Now we will write the tuple, and then check to see if it is the B tuple
409 * in any new or known pair. When we resolve a known pair, we will be
410 * able to write that pair's A tuple, and then we have to check if it
411 * resolves some other pair. Hence, we need a loop here.
413 old_tid = old_tuple->t_self;
418 ItemPointerData new_tid;
420 /* Insert the tuple and find out where it's put in new_heap */
421 raw_heap_insert(state, new_tuple);
422 new_tid = new_tuple->t_self;
425 * If the tuple is the updated version of a row, and the prior version
426 * wouldn't be DEAD yet, then we need to either resolve the prior
427 * version (if it's waiting in rs_unresolved_tups), or make an entry
428 * in rs_old_new_tid_map (so we can resolve it when we do see it).
429 * The previous tuple's xmax would equal this one's xmin, so it's
430 * RECENTLY_DEAD if and only if the xmin is not before OldestXmin.
432 if ((new_tuple->t_data->t_infomask & HEAP_UPDATED) &&
433 !TransactionIdPrecedes(HeapTupleHeaderGetXmin(new_tuple->t_data),
434 state->rs_oldest_xmin))
437 * Okay, this is B in an update pair. See if we've seen A.
439 UnresolvedTup unresolved;
441 memset(&hashkey, 0, sizeof(hashkey));
442 hashkey.xmin = HeapTupleHeaderGetXmin(new_tuple->t_data);
443 hashkey.tid = old_tid;
445 unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
448 if (unresolved != NULL)
451 * We have seen and memorized the previous tuple already. Now
452 * that we know where we inserted the tuple its t_ctid points
453 * to, fix its t_ctid and insert it to the new heap.
456 heap_freetuple(new_tuple);
457 new_tuple = unresolved->tuple;
459 old_tid = unresolved->old_tid;
460 new_tuple->t_data->t_ctid = new_tid;
463 * We don't need the hash entry anymore, but don't free its
466 hash_search(state->rs_unresolved_tups, &hashkey,
467 HASH_REMOVE, &found);
470 /* loop back to insert the previous tuple in the chain */
476 * Remember the new tid of this tuple. We'll use it to set the
477 * ctid when we find the previous tuple in the chain.
479 OldToNewMapping mapping;
481 mapping = hash_search(state->rs_old_new_tid_map, &hashkey,
485 mapping->new_tid = new_tid;
489 /* Done with this (chain of) tuples, for now */
491 heap_freetuple(new_tuple);
495 MemoryContextSwitchTo(old_cxt);
499 * Register a dead tuple with an ongoing rewrite. Dead tuples are not
500 * copied to the new table, but we still make note of them so that we
501 * can release some resources earlier.
504 rewrite_heap_dead_tuple(RewriteState state, HeapTuple old_tuple)
507 * If we have already seen an earlier tuple in the update chain that
508 * points to this tuple, let's forget about that earlier tuple. It's in
509 * fact dead as well, our simple xmax < OldestXmin test in
510 * HeapTupleSatisfiesVacuum just wasn't enough to detect it. It happens
511 * when xmin of a tuple is greater than xmax, which sounds
512 * counter-intuitive but is perfectly valid.
514 * We don't bother to try to detect the situation the other way round,
515 * when we encounter the dead tuple first and then the recently dead one
516 * that points to it. If that happens, we'll have some unmatched entries
517 * in the UnresolvedTups hash table at the end. That can happen anyway,
518 * because a vacuum might have removed the dead tuple in the chain before
521 UnresolvedTup unresolved;
525 memset(&hashkey, 0, sizeof(hashkey));
526 hashkey.xmin = HeapTupleHeaderGetXmin(old_tuple->t_data);
527 hashkey.tid = old_tuple->t_self;
529 unresolved = hash_search(state->rs_unresolved_tups, &hashkey,
532 if (unresolved != NULL)
534 /* Need to free the contained tuple as well as the hashtable entry */
535 heap_freetuple(unresolved->tuple);
536 hash_search(state->rs_unresolved_tups, &hashkey,
537 HASH_REMOVE, &found);
543 * Insert a tuple to the new relation. This has to track heap_insert
544 * and its subsidiary functions!
546 * t_self of the tuple is set to the new TID of the tuple. If t_ctid of the
547 * tuple is invalid on entry, it's replaced with the new TID as well (in
548 * the inserted data only, not in the caller's copy).
551 raw_heap_insert(RewriteState state, HeapTuple tup)
553 Page page = state->rs_buffer;
561 * If the new tuple is too big for storage or contains already toasted
562 * out-of-line attributes from some other relation, invoke the toaster.
564 * Note: below this point, heaptup is the data we actually intend to store
565 * into the relation; tup is the caller's original untoasted data.
567 if (state->rs_new_rel->rd_rel->relkind == RELKIND_TOASTVALUE)
569 /* toast table entries should never be recursively toasted */
570 Assert(!HeapTupleHasExternal(tup));
573 else if (HeapTupleHasExternal(tup) || tup->t_len > TOAST_TUPLE_THRESHOLD)
574 heaptup = toast_insert_or_update(state->rs_new_rel, tup, NULL,
575 state->rs_use_wal, false);
579 len = MAXALIGN(heaptup->t_len); /* be conservative */
582 * If we're gonna fail for oversize tuple, do it right away
584 if (len > MaxHeapTupleSize)
586 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
587 errmsg("row is too big: size %lu, maximum size %lu",
589 (unsigned long) MaxHeapTupleSize)));
591 /* Compute desired extra freespace due to fillfactor option */
592 saveFreeSpace = RelationGetTargetPageFreeSpace(state->rs_new_rel,
593 HEAP_DEFAULT_FILLFACTOR);
595 /* Now we can check to see if there's enough free space already. */
596 if (state->rs_buffer_valid)
598 pageFreeSpace = PageGetHeapFreeSpace(page);
600 if (len + saveFreeSpace > pageFreeSpace)
602 /* Doesn't fit, so write out the existing page */
605 if (state->rs_use_wal)
606 log_newpage(&state->rs_new_rel->rd_node,
611 * Now write the page. We say isTemp = true even if it's not a
612 * temp table, because there's no need for smgr to schedule an
613 * fsync for this write; we'll do it ourselves in
616 RelationOpenSmgr(state->rs_new_rel);
617 smgrextend(state->rs_new_rel->rd_smgr, state->rs_blockno,
618 (char *) page, true);
621 state->rs_buffer_valid = false;
625 if (!state->rs_buffer_valid)
627 /* Initialize a new empty page */
628 PageInit(page, BLCKSZ, 0);
629 state->rs_buffer_valid = true;
632 /* And now we can insert the tuple into the page */
633 newoff = PageAddItem(page, (Item) heaptup->t_data, len,
634 InvalidOffsetNumber, false, true);
635 if (newoff == InvalidOffsetNumber)
636 elog(ERROR, "failed to add tuple");
638 /* Update caller's t_self to the actual position where it was stored */
639 ItemPointerSet(&(tup->t_self), state->rs_blockno, newoff);
642 * Insert the correct position into CTID of the stored tuple, too, if the
643 * caller didn't supply a valid CTID.
645 if (!ItemPointerIsValid(&tup->t_data->t_ctid))
648 HeapTupleHeader onpage_tup;
650 newitemid = PageGetItemId(page, newoff);
651 onpage_tup = (HeapTupleHeader) PageGetItem(page, newitemid);
653 onpage_tup->t_ctid = tup->t_self;
656 /* If heaptup is a private copy, release it. */
658 heap_freetuple(heaptup);