1 /*-------------------------------------------------------------------------
2 *
3 * hashpage.c
4 * Hash table page management code for the Postgres hash access method
5 *
6 * Portions Copyright (c) 1996-2018, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
8 *
9 *
10 * IDENTIFICATION
11 * src/backend/access/hash/hashpage.c
12 *
13 * NOTES
14 * Postgres hash pages look like ordinary relation pages. The opaque
15 * data at high addresses includes information about the page including
16 * whether a page is an overflow page or a true bucket, the bucket
17 * number, and the block numbers of the preceding and following pages
18 * in the same bucket.
19 *
20 * The first page in a hash relation, page zero, is special -- it stores
21 * information describing the hash table; it is referred to as the
22 * "meta page." Pages one and higher store the actual data.
23 *
24 * There are also bitmap pages, which are not manipulated here;
25 * see hashovfl.c.
26 *
27 *-------------------------------------------------------------------------
28 */
29 #include "postgres.h"
30
31 #include "access/hash.h"
32 #include "access/hash_xlog.h"
33 #include "miscadmin.h"
34 #include "storage/lmgr.h"
35 #include "storage/smgr.h"
36 #include "storage/predicate.h"
37
38
39 static bool _hash_alloc_buckets(Relation rel, BlockNumber firstblock,
40 uint32 nblocks);
41 static void _hash_splitbucket(Relation rel, Buffer metabuf,
42 Bucket obucket, Bucket nbucket,
43 Buffer obuf,
44 Buffer nbuf,
45 HTAB *htab,
46 uint32 maxbucket,
47 uint32 highmask, uint32 lowmask);
48 static void log_split_page(Relation rel, Buffer buf);
49
50
51 /*
52 * We use high-concurrency locking on hash indexes (see README for an overview
53 * of the locking rules). However, we can skip taking lmgr locks when the
54 * index is local to the current backend (ie, either temp or new in the
55 * current transaction). No one else can see it, so there's no reason to
56 * take locks. We still take buffer-level locks, but not lmgr locks.
57 */
58 #define USELOCKING(rel) (!RELATION_IS_LOCAL(rel))
59
60
61 /*
62 * _hash_getbuf() -- Get a buffer by block number for read or write.
63 *
64 * 'access' must be HASH_READ, HASH_WRITE, or HASH_NOLOCK.
65 * 'flags' is a bitwise OR of the allowed page types.
66 *
67 * This must be used only to fetch pages that are expected to be valid
68 * already. _hash_checkpage() is applied using the given flags.
69 *
70 * When this routine returns, the appropriate lock is set on the
71 * requested buffer and its reference count has been incremented
72 * (ie, the buffer is "locked and pinned").
73 *
74 * P_NEW is disallowed because this routine can only be used
75 * to access pages that are known to be before the filesystem EOF.
76 * Extending the index should be done with _hash_getnewbuf.
77 */
78 Buffer
_hash_getbuf(Relation rel,BlockNumber blkno,int access,int flags)79 _hash_getbuf(Relation rel, BlockNumber blkno, int access, int flags)
80 {
81 Buffer buf;
82
83 if (blkno == P_NEW)
84 elog(ERROR, "hash AM does not use P_NEW");
85
86 buf = ReadBuffer(rel, blkno);
87
88 if (access != HASH_NOLOCK)
89 LockBuffer(buf, access);
90
91 /* ref count and lock type are correct */
92
93 _hash_checkpage(rel, buf, flags);
94
95 return buf;
96 }
97
98 /*
99 * _hash_getbuf_with_condlock_cleanup() -- Try to get a buffer for cleanup.
100 *
101 * We read the page and try to acquire a cleanup lock. If we get it,
102 * we return the buffer; otherwise, we return InvalidBuffer.
103 */
104 Buffer
_hash_getbuf_with_condlock_cleanup(Relation rel,BlockNumber blkno,int flags)105 _hash_getbuf_with_condlock_cleanup(Relation rel, BlockNumber blkno, int flags)
106 {
107 Buffer buf;
108
109 if (blkno == P_NEW)
110 elog(ERROR, "hash AM does not use P_NEW");
111
112 buf = ReadBuffer(rel, blkno);
113
114 if (!ConditionalLockBufferForCleanup(buf))
115 {
116 ReleaseBuffer(buf);
117 return InvalidBuffer;
118 }
119
120 /* ref count and lock type are correct */
121
122 _hash_checkpage(rel, buf, flags);
123
124 return buf;
125 }
126
127 /*
128 * _hash_getinitbuf() -- Get and initialize a buffer by block number.
129 *
130 * This must be used only to fetch pages that are known to be before
131 * the index's filesystem EOF, but are to be filled from scratch.
132 * _hash_pageinit() is applied automatically. Otherwise it has
133 * effects similar to _hash_getbuf() with access = HASH_WRITE.
134 *
135 * When this routine returns, a write lock is set on the
136 * requested buffer and its reference count has been incremented
137 * (ie, the buffer is "locked and pinned").
138 *
139 * P_NEW is disallowed because this routine can only be used
140 * to access pages that are known to be before the filesystem EOF.
141 * Extending the index should be done with _hash_getnewbuf.
142 */
143 Buffer
_hash_getinitbuf(Relation rel,BlockNumber blkno)144 _hash_getinitbuf(Relation rel, BlockNumber blkno)
145 {
146 Buffer buf;
147
148 if (blkno == P_NEW)
149 elog(ERROR, "hash AM does not use P_NEW");
150
151 buf = ReadBufferExtended(rel, MAIN_FORKNUM, blkno, RBM_ZERO_AND_LOCK,
152 NULL);
153
154 /* ref count and lock type are correct */
155
156 /* initialize the page */
157 _hash_pageinit(BufferGetPage(buf), BufferGetPageSize(buf));
158
159 return buf;
160 }
161
162 /*
163 * _hash_initbuf() -- Get and initialize a buffer by bucket number.
164 */
165 void
_hash_initbuf(Buffer buf,uint32 max_bucket,uint32 num_bucket,uint32 flag,bool initpage)166 _hash_initbuf(Buffer buf, uint32 max_bucket, uint32 num_bucket, uint32 flag,
167 bool initpage)
168 {
169 HashPageOpaque pageopaque;
170 Page page;
171
172 page = BufferGetPage(buf);
173
174 /* initialize the page */
175 if (initpage)
176 _hash_pageinit(page, BufferGetPageSize(buf));
177
178 pageopaque = (HashPageOpaque) PageGetSpecialPointer(page);
179
180 /*
181 * Set hasho_prevblkno with current hashm_maxbucket. This value will be
182 * used to validate cached HashMetaPageData. See
183 * _hash_getbucketbuf_from_hashkey().
184 */
185 pageopaque->hasho_prevblkno = max_bucket;
186 pageopaque->hasho_nextblkno = InvalidBlockNumber;
187 pageopaque->hasho_bucket = num_bucket;
188 pageopaque->hasho_flag = flag;
189 pageopaque->hasho_page_id = HASHO_PAGE_ID;
190 }
191
192 /*
193 * _hash_getnewbuf() -- Get a new page at the end of the index.
194 *
195 * This has the same API as _hash_getinitbuf, except that we are adding
196 * a page to the index, and hence expect the page to be past the
197 * logical EOF. (However, we have to support the case where it isn't,
198 * since a prior try might have crashed after extending the filesystem
199 * EOF but before updating the metapage to reflect the added page.)
200 *
201 * It is caller's responsibility to ensure that only one process can
202 * extend the index at a time. In practice, this function is called
203 * only while holding write lock on the metapage, because adding a page
204 * is always associated with an update of metapage data.
205 */
206 Buffer
_hash_getnewbuf(Relation rel,BlockNumber blkno,ForkNumber forkNum)207 _hash_getnewbuf(Relation rel, BlockNumber blkno, ForkNumber forkNum)
208 {
209 BlockNumber nblocks = RelationGetNumberOfBlocksInFork(rel, forkNum);
210 Buffer buf;
211
212 if (blkno == P_NEW)
213 elog(ERROR, "hash AM does not use P_NEW");
214 if (blkno > nblocks)
215 elog(ERROR, "access to noncontiguous page in hash index \"%s\"",
216 RelationGetRelationName(rel));
217
218 /* smgr insists we use P_NEW to extend the relation */
219 if (blkno == nblocks)
220 {
221 buf = ReadBufferExtended(rel, forkNum, P_NEW, RBM_NORMAL, NULL);
222 if (BufferGetBlockNumber(buf) != blkno)
223 elog(ERROR, "unexpected hash relation size: %u, should be %u",
224 BufferGetBlockNumber(buf), blkno);
225 LockBuffer(buf, HASH_WRITE);
226 }
227 else
228 {
229 buf = ReadBufferExtended(rel, forkNum, blkno, RBM_ZERO_AND_LOCK,
230 NULL);
231 }
232
233 /* ref count and lock type are correct */
234
235 /* initialize the page */
236 _hash_pageinit(BufferGetPage(buf), BufferGetPageSize(buf));
237
238 return buf;
239 }
240
241 /*
242 * _hash_getbuf_with_strategy() -- Get a buffer with nondefault strategy.
243 *
244 * This is identical to _hash_getbuf() but also allows a buffer access
245 * strategy to be specified. We use this for VACUUM operations.
246 */
247 Buffer
_hash_getbuf_with_strategy(Relation rel,BlockNumber blkno,int access,int flags,BufferAccessStrategy bstrategy)248 _hash_getbuf_with_strategy(Relation rel, BlockNumber blkno,
249 int access, int flags,
250 BufferAccessStrategy bstrategy)
251 {
252 Buffer buf;
253
254 if (blkno == P_NEW)
255 elog(ERROR, "hash AM does not use P_NEW");
256
257 buf = ReadBufferExtended(rel, MAIN_FORKNUM, blkno, RBM_NORMAL, bstrategy);
258
259 if (access != HASH_NOLOCK)
260 LockBuffer(buf, access);
261
262 /* ref count and lock type are correct */
263
264 _hash_checkpage(rel, buf, flags);
265
266 return buf;
267 }
268
269 /*
270 * _hash_relbuf() -- release a locked buffer.
271 *
272 * Lock and pin (refcount) are both dropped.
273 */
274 void
_hash_relbuf(Relation rel,Buffer buf)275 _hash_relbuf(Relation rel, Buffer buf)
276 {
277 UnlockReleaseBuffer(buf);
278 }
279
280 /*
281 * _hash_dropbuf() -- release an unlocked buffer.
282 *
283 * This is used to unpin a buffer on which we hold no lock.
284 */
285 void
_hash_dropbuf(Relation rel,Buffer buf)286 _hash_dropbuf(Relation rel, Buffer buf)
287 {
288 ReleaseBuffer(buf);
289 }
290
291 /*
292 * _hash_dropscanbuf() -- release buffers used in scan.
293 *
294 * This routine unpins the buffers used during scan on which we
295 * hold no lock.
296 */
297 void
_hash_dropscanbuf(Relation rel,HashScanOpaque so)298 _hash_dropscanbuf(Relation rel, HashScanOpaque so)
299 {
300 /* release pin we hold on primary bucket page */
301 if (BufferIsValid(so->hashso_bucket_buf) &&
302 so->hashso_bucket_buf != so->currPos.buf)
303 _hash_dropbuf(rel, so->hashso_bucket_buf);
304 so->hashso_bucket_buf = InvalidBuffer;
305
306 /* release pin we hold on primary bucket page of bucket being split */
307 if (BufferIsValid(so->hashso_split_bucket_buf) &&
308 so->hashso_split_bucket_buf != so->currPos.buf)
309 _hash_dropbuf(rel, so->hashso_split_bucket_buf);
310 so->hashso_split_bucket_buf = InvalidBuffer;
311
312 /* release any pin we still hold */
313 if (BufferIsValid(so->currPos.buf))
314 _hash_dropbuf(rel, so->currPos.buf);
315 so->currPos.buf = InvalidBuffer;
316
317 /* reset split scan */
318 so->hashso_buc_populated = false;
319 so->hashso_buc_split = false;
320 }
321
322
323 /*
324 * _hash_init() -- Initialize the metadata page of a hash index,
325 * the initial buckets, and the initial bitmap page.
326 *
327 * The initial number of buckets is dependent on num_tuples, an estimate
328 * of the number of tuples to be loaded into the index initially. The
329 * chosen number of buckets is returned.
330 *
331 * We are fairly cavalier about locking here, since we know that no one else
332 * could be accessing this index. In particular the rule about not holding
333 * multiple buffer locks is ignored.
334 */
335 uint32
_hash_init(Relation rel,double num_tuples,ForkNumber forkNum)336 _hash_init(Relation rel, double num_tuples, ForkNumber forkNum)
337 {
338 Buffer metabuf;
339 Buffer buf;
340 Buffer bitmapbuf;
341 Page pg;
342 HashMetaPage metap;
343 RegProcedure procid;
344 int32 data_width;
345 int32 item_width;
346 int32 ffactor;
347 uint32 num_buckets;
348 uint32 i;
349 bool use_wal;
350
351 /* safety check */
352 if (RelationGetNumberOfBlocksInFork(rel, forkNum) != 0)
353 elog(ERROR, "cannot initialize non-empty hash index \"%s\"",
354 RelationGetRelationName(rel));
355
356 /*
357 * WAL log creation of pages if the relation is persistent, or this is the
358 * init fork. Init forks for unlogged relations always need to be WAL
359 * logged.
360 */
361 use_wal = RelationNeedsWAL(rel) || forkNum == INIT_FORKNUM;
362
363 /*
364 * Determine the target fill factor (in tuples per bucket) for this index.
365 * The idea is to make the fill factor correspond to pages about as full
366 * as the user-settable fillfactor parameter says. We can compute it
367 * exactly since the index datatype (i.e. uint32 hash key) is fixed-width.
368 */
369 data_width = sizeof(uint32);
370 item_width = MAXALIGN(sizeof(IndexTupleData)) + MAXALIGN(data_width) +
371 sizeof(ItemIdData); /* include the line pointer */
372 ffactor = RelationGetTargetPageUsage(rel, HASH_DEFAULT_FILLFACTOR) / item_width;
373 /* keep to a sane range */
374 if (ffactor < 10)
375 ffactor = 10;
376
377 procid = index_getprocid(rel, 1, HASHSTANDARD_PROC);
378
379 /*
380 * We initialize the metapage, the first N bucket pages, and the first
381 * bitmap page in sequence, using _hash_getnewbuf to cause smgrextend()
382 * calls to occur. This ensures that the smgr level has the right idea of
383 * the physical index length.
384 *
385 * Critical section not required, because on error the creation of the
386 * whole relation will be rolled back.
387 */
388 metabuf = _hash_getnewbuf(rel, HASH_METAPAGE, forkNum);
389 _hash_init_metabuffer(metabuf, num_tuples, procid, ffactor, false);
390 MarkBufferDirty(metabuf);
391
392 pg = BufferGetPage(metabuf);
393 metap = HashPageGetMeta(pg);
394
395 /* XLOG stuff */
396 if (use_wal)
397 {
398 xl_hash_init_meta_page xlrec;
399 XLogRecPtr recptr;
400
401 xlrec.num_tuples = num_tuples;
402 xlrec.procid = metap->hashm_procid;
403 xlrec.ffactor = metap->hashm_ffactor;
404
405 XLogBeginInsert();
406 XLogRegisterData((char *) &xlrec, SizeOfHashInitMetaPage);
407 XLogRegisterBuffer(0, metabuf, REGBUF_WILL_INIT | REGBUF_STANDARD);
408
409 recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_INIT_META_PAGE);
410
411 PageSetLSN(BufferGetPage(metabuf), recptr);
412 }
413
414 num_buckets = metap->hashm_maxbucket + 1;
415
416 /*
417 * Release buffer lock on the metapage while we initialize buckets.
418 * Otherwise, we'll be in interrupt holdoff and the CHECK_FOR_INTERRUPTS
419 * won't accomplish anything. It's a bad idea to hold buffer locks for
420 * long intervals in any case, since that can block the bgwriter.
421 */
422 LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
423
424 /*
425 * Initialize and WAL Log the first N buckets
426 */
427 for (i = 0; i < num_buckets; i++)
428 {
429 BlockNumber blkno;
430
431 /* Allow interrupts, in case N is huge */
432 CHECK_FOR_INTERRUPTS();
433
434 blkno = BUCKET_TO_BLKNO(metap, i);
435 buf = _hash_getnewbuf(rel, blkno, forkNum);
436 _hash_initbuf(buf, metap->hashm_maxbucket, i, LH_BUCKET_PAGE, false);
437 MarkBufferDirty(buf);
438
439 if (use_wal)
440 log_newpage(&rel->rd_node,
441 forkNum,
442 blkno,
443 BufferGetPage(buf),
444 true);
445 _hash_relbuf(rel, buf);
446 }
447
448 /* Now reacquire buffer lock on metapage */
449 LockBuffer(metabuf, BUFFER_LOCK_EXCLUSIVE);
450
451 /*
452 * Initialize bitmap page
453 */
454 bitmapbuf = _hash_getnewbuf(rel, num_buckets + 1, forkNum);
455 _hash_initbitmapbuffer(bitmapbuf, metap->hashm_bmsize, false);
456 MarkBufferDirty(bitmapbuf);
457
458 /* add the new bitmap page to the metapage's list of bitmaps */
459 /* metapage already has a write lock */
460 if (metap->hashm_nmaps >= HASH_MAX_BITMAPS)
461 ereport(ERROR,
462 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
463 errmsg("out of overflow pages in hash index \"%s\"",
464 RelationGetRelationName(rel))));
465
466 metap->hashm_mapp[metap->hashm_nmaps] = num_buckets + 1;
467
468 metap->hashm_nmaps++;
469 MarkBufferDirty(metabuf);
470
471 /* XLOG stuff */
472 if (use_wal)
473 {
474 xl_hash_init_bitmap_page xlrec;
475 XLogRecPtr recptr;
476
477 xlrec.bmsize = metap->hashm_bmsize;
478
479 XLogBeginInsert();
480 XLogRegisterData((char *) &xlrec, SizeOfHashInitBitmapPage);
481 XLogRegisterBuffer(0, bitmapbuf, REGBUF_WILL_INIT);
482
483 /*
484 * This is safe only because nobody else can be modifying the index at
485 * this stage; it's only visible to the transaction that is creating
486 * it.
487 */
488 XLogRegisterBuffer(1, metabuf, REGBUF_STANDARD);
489
490 recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_INIT_BITMAP_PAGE);
491
492 PageSetLSN(BufferGetPage(bitmapbuf), recptr);
493 PageSetLSN(BufferGetPage(metabuf), recptr);
494 }
495
496 /* all done */
497 _hash_relbuf(rel, bitmapbuf);
498 _hash_relbuf(rel, metabuf);
499
500 return num_buckets;
501 }
502
503 /*
504 * _hash_init_metabuffer() -- Initialize the metadata page of a hash index.
505 */
506 void
_hash_init_metabuffer(Buffer buf,double num_tuples,RegProcedure procid,uint16 ffactor,bool initpage)507 _hash_init_metabuffer(Buffer buf, double num_tuples, RegProcedure procid,
508 uint16 ffactor, bool initpage)
509 {
510 HashMetaPage metap;
511 HashPageOpaque pageopaque;
512 Page page;
513 double dnumbuckets;
514 uint32 num_buckets;
515 uint32 spare_index;
516 uint32 i;
517
518 /*
519 * Choose the number of initial bucket pages to match the fill factor
520 * given the estimated number of tuples. We round up the result to the
521 * total number of buckets which has to be allocated before using its
522 * _hashm_spare element. However always force at least 2 bucket pages. The
523 * upper limit is determined by considerations explained in
524 * _hash_expandtable().
525 */
526 dnumbuckets = num_tuples / ffactor;
527 if (dnumbuckets <= 2.0)
528 num_buckets = 2;
529 else if (dnumbuckets >= (double) 0x40000000)
530 num_buckets = 0x40000000;
531 else
532 num_buckets = _hash_get_totalbuckets(_hash_spareindex(dnumbuckets));
533
534 spare_index = _hash_spareindex(num_buckets);
535 Assert(spare_index < HASH_MAX_SPLITPOINTS);
536
537 page = BufferGetPage(buf);
538 if (initpage)
539 _hash_pageinit(page, BufferGetPageSize(buf));
540
541 pageopaque = (HashPageOpaque) PageGetSpecialPointer(page);
542 pageopaque->hasho_prevblkno = InvalidBlockNumber;
543 pageopaque->hasho_nextblkno = InvalidBlockNumber;
544 pageopaque->hasho_bucket = -1;
545 pageopaque->hasho_flag = LH_META_PAGE;
546 pageopaque->hasho_page_id = HASHO_PAGE_ID;
547
548 metap = HashPageGetMeta(page);
549
550 metap->hashm_magic = HASH_MAGIC;
551 metap->hashm_version = HASH_VERSION;
552 metap->hashm_ntuples = 0;
553 metap->hashm_nmaps = 0;
554 metap->hashm_ffactor = ffactor;
555 metap->hashm_bsize = HashGetMaxBitmapSize(page);
556 /* find largest bitmap array size that will fit in page size */
557 for (i = _hash_log2(metap->hashm_bsize); i > 0; --i)
558 {
559 if ((1 << i) <= metap->hashm_bsize)
560 break;
561 }
562 Assert(i > 0);
563 metap->hashm_bmsize = 1 << i;
564 metap->hashm_bmshift = i + BYTE_TO_BIT;
565 Assert((1 << BMPG_SHIFT(metap)) == (BMPG_MASK(metap) + 1));
566
567 /*
568 * Label the index with its primary hash support function's OID. This is
569 * pretty useless for normal operation (in fact, hashm_procid is not used
570 * anywhere), but it might be handy for forensic purposes so we keep it.
571 */
572 metap->hashm_procid = procid;
573
574 /*
575 * We initialize the index with N buckets, 0 .. N-1, occupying physical
576 * blocks 1 to N. The first freespace bitmap page is in block N+1.
577 */
578 metap->hashm_maxbucket = num_buckets - 1;
579
580 /*
581 * Set highmask as next immediate ((2 ^ x) - 1), which should be
582 * sufficient to cover num_buckets.
583 */
584 metap->hashm_highmask = (1 << (_hash_log2(num_buckets + 1))) - 1;
585 metap->hashm_lowmask = (metap->hashm_highmask >> 1);
586
587 MemSet(metap->hashm_spares, 0, sizeof(metap->hashm_spares));
588 MemSet(metap->hashm_mapp, 0, sizeof(metap->hashm_mapp));
589
590 /* Set up mapping for one spare page after the initial splitpoints */
591 metap->hashm_spares[spare_index] = 1;
592 metap->hashm_ovflpoint = spare_index;
593 metap->hashm_firstfree = 0;
594
595 /*
596 * Set pd_lower just past the end of the metadata. This is essential,
597 * because without doing so, metadata will be lost if xlog.c compresses
598 * the page.
599 */
600 ((PageHeader) page)->pd_lower =
601 ((char *) metap + sizeof(HashMetaPageData)) - (char *) page;
602 }
603
604 /*
605 * _hash_pageinit() -- Initialize a new hash index page.
606 */
607 void
_hash_pageinit(Page page,Size size)608 _hash_pageinit(Page page, Size size)
609 {
610 PageInit(page, size, sizeof(HashPageOpaqueData));
611 }
612
613 /*
614 * Attempt to expand the hash table by creating one new bucket.
615 *
616 * This will silently do nothing if we don't get cleanup lock on old or
617 * new bucket.
618 *
619 * Complete the pending splits and remove the tuples from old bucket,
620 * if there are any left over from the previous split.
621 *
622 * The caller must hold a pin, but no lock, on the metapage buffer.
623 * The buffer is returned in the same state.
624 */
625 void
_hash_expandtable(Relation rel,Buffer metabuf)626 _hash_expandtable(Relation rel, Buffer metabuf)
627 {
628 HashMetaPage metap;
629 Bucket old_bucket;
630 Bucket new_bucket;
631 uint32 spare_ndx;
632 BlockNumber start_oblkno;
633 BlockNumber start_nblkno;
634 Buffer buf_nblkno;
635 Buffer buf_oblkno;
636 Page opage;
637 Page npage;
638 HashPageOpaque oopaque;
639 HashPageOpaque nopaque;
640 uint32 maxbucket;
641 uint32 highmask;
642 uint32 lowmask;
643 bool metap_update_masks = false;
644 bool metap_update_splitpoint = false;
645
646 restart_expand:
647
648 /*
649 * Write-lock the meta page. It used to be necessary to acquire a
650 * heavyweight lock to begin a split, but that is no longer required.
651 */
652 LockBuffer(metabuf, BUFFER_LOCK_EXCLUSIVE);
653
654 _hash_checkpage(rel, metabuf, LH_META_PAGE);
655 metap = HashPageGetMeta(BufferGetPage(metabuf));
656
657 /*
658 * Check to see if split is still needed; someone else might have already
659 * done one while we waited for the lock.
660 *
661 * Make sure this stays in sync with _hash_doinsert()
662 */
663 if (metap->hashm_ntuples <=
664 (double) metap->hashm_ffactor * (metap->hashm_maxbucket + 1))
665 goto fail;
666
667 /*
668 * Can't split anymore if maxbucket has reached its maximum possible
669 * value.
670 *
671 * Ideally we'd allow bucket numbers up to UINT_MAX-1 (no higher because
672 * the calculation maxbucket+1 mustn't overflow). Currently we restrict
673 * to half that because of overflow looping in _hash_log2() and
674 * insufficient space in hashm_spares[]. It's moot anyway because an
675 * index with 2^32 buckets would certainly overflow BlockNumber and hence
676 * _hash_alloc_buckets() would fail, but if we supported buckets smaller
677 * than a disk block then this would be an independent constraint.
678 *
679 * If you change this, see also the maximum initial number of buckets in
680 * _hash_init().
681 */
682 if (metap->hashm_maxbucket >= (uint32) 0x7FFFFFFE)
683 goto fail;
684
685 /*
686 * Determine which bucket is to be split, and attempt to take cleanup lock
687 * on the old bucket. If we can't get the lock, give up.
688 *
689 * The cleanup lock protects us not only against other backends, but
690 * against our own backend as well.
691 *
692 * The cleanup lock is mainly to protect the split from concurrent
693 * inserts. See src/backend/access/hash/README, Lock Definitions for
694 * further details. Due to this locking restriction, if there is any
695 * pending scan, the split will give up which is not good, but harmless.
696 */
697 new_bucket = metap->hashm_maxbucket + 1;
698
699 old_bucket = (new_bucket & metap->hashm_lowmask);
700
701 start_oblkno = BUCKET_TO_BLKNO(metap, old_bucket);
702
703 buf_oblkno = _hash_getbuf_with_condlock_cleanup(rel, start_oblkno, LH_BUCKET_PAGE);
704 if (!buf_oblkno)
705 goto fail;
706
707 opage = BufferGetPage(buf_oblkno);
708 oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
709
710 /*
711 * We want to finish the split from a bucket as there is no apparent
712 * benefit by not doing so and it will make the code complicated to finish
713 * the split that involves multiple buckets considering the case where new
714 * split also fails. We don't need to consider the new bucket for
715 * completing the split here as it is not possible that a re-split of new
716 * bucket starts when there is still a pending split from old bucket.
717 */
718 if (H_BUCKET_BEING_SPLIT(oopaque))
719 {
720 /*
721 * Copy bucket mapping info now; refer the comment in code below where
722 * we copy this information before calling _hash_splitbucket to see
723 * why this is okay.
724 */
725 maxbucket = metap->hashm_maxbucket;
726 highmask = metap->hashm_highmask;
727 lowmask = metap->hashm_lowmask;
728
729 /*
730 * Release the lock on metapage and old_bucket, before completing the
731 * split.
732 */
733 LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
734 LockBuffer(buf_oblkno, BUFFER_LOCK_UNLOCK);
735
736 _hash_finish_split(rel, metabuf, buf_oblkno, old_bucket, maxbucket,
737 highmask, lowmask);
738
739 /* release the pin on old buffer and retry for expand. */
740 _hash_dropbuf(rel, buf_oblkno);
741
742 goto restart_expand;
743 }
744
745 /*
746 * Clean the tuples remained from the previous split. This operation
747 * requires cleanup lock and we already have one on the old bucket, so
748 * let's do it. We also don't want to allow further splits from the bucket
749 * till the garbage of previous split is cleaned. This has two
750 * advantages; first, it helps in avoiding the bloat due to garbage and
751 * second is, during cleanup of bucket, we are always sure that the
752 * garbage tuples belong to most recently split bucket. On the contrary,
753 * if we allow cleanup of bucket after meta page is updated to indicate
754 * the new split and before the actual split, the cleanup operation won't
755 * be able to decide whether the tuple has been moved to the newly created
756 * bucket and ended up deleting such tuples.
757 */
758 if (H_NEEDS_SPLIT_CLEANUP(oopaque))
759 {
760 /*
761 * Copy bucket mapping info now; refer to the comment in code below
762 * where we copy this information before calling _hash_splitbucket to
763 * see why this is okay.
764 */
765 maxbucket = metap->hashm_maxbucket;
766 highmask = metap->hashm_highmask;
767 lowmask = metap->hashm_lowmask;
768
769 /* Release the metapage lock. */
770 LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
771
772 hashbucketcleanup(rel, old_bucket, buf_oblkno, start_oblkno, NULL,
773 maxbucket, highmask, lowmask, NULL, NULL, true,
774 NULL, NULL);
775
776 _hash_dropbuf(rel, buf_oblkno);
777
778 goto restart_expand;
779 }
780
781 /*
782 * There shouldn't be any active scan on new bucket.
783 *
784 * Note: it is safe to compute the new bucket's blkno here, even though we
785 * may still need to update the BUCKET_TO_BLKNO mapping. This is because
786 * the current value of hashm_spares[hashm_ovflpoint] correctly shows
787 * where we are going to put a new splitpoint's worth of buckets.
788 */
789 start_nblkno = BUCKET_TO_BLKNO(metap, new_bucket);
790
791 /*
792 * If the split point is increasing we need to allocate a new batch of
793 * bucket pages.
794 */
795 spare_ndx = _hash_spareindex(new_bucket + 1);
796 if (spare_ndx > metap->hashm_ovflpoint)
797 {
798 uint32 buckets_to_add;
799
800 Assert(spare_ndx == metap->hashm_ovflpoint + 1);
801
802 /*
803 * We treat allocation of buckets as a separate WAL-logged action.
804 * Even if we fail after this operation, won't leak bucket pages;
805 * rather, the next split will consume this space. In any case, even
806 * without failure we don't use all the space in one split operation.
807 */
808 buckets_to_add = _hash_get_totalbuckets(spare_ndx) - new_bucket;
809 if (!_hash_alloc_buckets(rel, start_nblkno, buckets_to_add))
810 {
811 /* can't split due to BlockNumber overflow */
812 _hash_relbuf(rel, buf_oblkno);
813 goto fail;
814 }
815 }
816
817 /*
818 * Physically allocate the new bucket's primary page. We want to do this
819 * before changing the metapage's mapping info, in case we can't get the
820 * disk space. Ideally, we don't need to check for cleanup lock on new
821 * bucket as no other backend could find this bucket unless meta page is
822 * updated. However, it is good to be consistent with old bucket locking.
823 */
824 buf_nblkno = _hash_getnewbuf(rel, start_nblkno, MAIN_FORKNUM);
825 if (!IsBufferCleanupOK(buf_nblkno))
826 {
827 _hash_relbuf(rel, buf_oblkno);
828 _hash_relbuf(rel, buf_nblkno);
829 goto fail;
830 }
831
832 /*
833 * Since we are scribbling on the pages in the shared buffers, establish a
834 * critical section. Any failure in this next code leaves us with a big
835 * problem: the metapage is effectively corrupt but could get written back
836 * to disk.
837 */
838 START_CRIT_SECTION();
839
840 /*
841 * Okay to proceed with split. Update the metapage bucket mapping info.
842 */
843 metap->hashm_maxbucket = new_bucket;
844
845 if (new_bucket > metap->hashm_highmask)
846 {
847 /* Starting a new doubling */
848 metap->hashm_lowmask = metap->hashm_highmask;
849 metap->hashm_highmask = new_bucket | metap->hashm_lowmask;
850 metap_update_masks = true;
851 }
852
853 /*
854 * If the split point is increasing we need to adjust the hashm_spares[]
855 * array and hashm_ovflpoint so that future overflow pages will be created
856 * beyond this new batch of bucket pages.
857 */
858 if (spare_ndx > metap->hashm_ovflpoint)
859 {
860 metap->hashm_spares[spare_ndx] = metap->hashm_spares[metap->hashm_ovflpoint];
861 metap->hashm_ovflpoint = spare_ndx;
862 metap_update_splitpoint = true;
863 }
864
865 MarkBufferDirty(metabuf);
866
867 /*
868 * Copy bucket mapping info now; this saves re-accessing the meta page
869 * inside _hash_splitbucket's inner loop. Note that once we drop the
870 * split lock, other splits could begin, so these values might be out of
871 * date before _hash_splitbucket finishes. That's okay, since all it
872 * needs is to tell which of these two buckets to map hashkeys into.
873 */
874 maxbucket = metap->hashm_maxbucket;
875 highmask = metap->hashm_highmask;
876 lowmask = metap->hashm_lowmask;
877
878 opage = BufferGetPage(buf_oblkno);
879 oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
880
881 /*
882 * Mark the old bucket to indicate that split is in progress. (At
883 * operation end, we will clear the split-in-progress flag.) Also, for a
884 * primary bucket page, hasho_prevblkno stores the number of buckets that
885 * existed as of the last split, so we must update that value here.
886 */
887 oopaque->hasho_flag |= LH_BUCKET_BEING_SPLIT;
888 oopaque->hasho_prevblkno = maxbucket;
889
890 MarkBufferDirty(buf_oblkno);
891
892 npage = BufferGetPage(buf_nblkno);
893
894 /*
895 * initialize the new bucket's primary page and mark it to indicate that
896 * split is in progress.
897 */
898 nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
899 nopaque->hasho_prevblkno = maxbucket;
900 nopaque->hasho_nextblkno = InvalidBlockNumber;
901 nopaque->hasho_bucket = new_bucket;
902 nopaque->hasho_flag = LH_BUCKET_PAGE | LH_BUCKET_BEING_POPULATED;
903 nopaque->hasho_page_id = HASHO_PAGE_ID;
904
905 MarkBufferDirty(buf_nblkno);
906
907 /* XLOG stuff */
908 if (RelationNeedsWAL(rel))
909 {
910 xl_hash_split_allocate_page xlrec;
911 XLogRecPtr recptr;
912
913 xlrec.new_bucket = maxbucket;
914 xlrec.old_bucket_flag = oopaque->hasho_flag;
915 xlrec.new_bucket_flag = nopaque->hasho_flag;
916 xlrec.flags = 0;
917
918 XLogBeginInsert();
919
920 XLogRegisterBuffer(0, buf_oblkno, REGBUF_STANDARD);
921 XLogRegisterBuffer(1, buf_nblkno, REGBUF_WILL_INIT);
922 XLogRegisterBuffer(2, metabuf, REGBUF_STANDARD);
923
924 if (metap_update_masks)
925 {
926 xlrec.flags |= XLH_SPLIT_META_UPDATE_MASKS;
927 XLogRegisterBufData(2, (char *) &metap->hashm_lowmask, sizeof(uint32));
928 XLogRegisterBufData(2, (char *) &metap->hashm_highmask, sizeof(uint32));
929 }
930
931 if (metap_update_splitpoint)
932 {
933 xlrec.flags |= XLH_SPLIT_META_UPDATE_SPLITPOINT;
934 XLogRegisterBufData(2, (char *) &metap->hashm_ovflpoint,
935 sizeof(uint32));
936 XLogRegisterBufData(2,
937 (char *) &metap->hashm_spares[metap->hashm_ovflpoint],
938 sizeof(uint32));
939 }
940
941 XLogRegisterData((char *) &xlrec, SizeOfHashSplitAllocPage);
942
943 recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_ALLOCATE_PAGE);
944
945 PageSetLSN(BufferGetPage(buf_oblkno), recptr);
946 PageSetLSN(BufferGetPage(buf_nblkno), recptr);
947 PageSetLSN(BufferGetPage(metabuf), recptr);
948 }
949
950 END_CRIT_SECTION();
951
952 /* drop lock, but keep pin */
953 LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
954
955 /* Relocate records to the new bucket */
956 _hash_splitbucket(rel, metabuf,
957 old_bucket, new_bucket,
958 buf_oblkno, buf_nblkno, NULL,
959 maxbucket, highmask, lowmask);
960
961 /* all done, now release the pins on primary buckets. */
962 _hash_dropbuf(rel, buf_oblkno);
963 _hash_dropbuf(rel, buf_nblkno);
964
965 return;
966
967 /* Here if decide not to split or fail to acquire old bucket lock */
968 fail:
969
970 /* We didn't write the metapage, so just drop lock */
971 LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
972 }
973
974
975 /*
976 * _hash_alloc_buckets -- allocate a new splitpoint's worth of bucket pages
977 *
978 * This does not need to initialize the new bucket pages; we'll do that as
979 * each one is used by _hash_expandtable(). But we have to extend the logical
980 * EOF to the end of the splitpoint; this keeps smgr's idea of the EOF in
981 * sync with ours, so that we don't get complaints from smgr.
982 *
983 * We do this by writing a page of zeroes at the end of the splitpoint range.
984 * We expect that the filesystem will ensure that the intervening pages read
985 * as zeroes too. On many filesystems this "hole" will not be allocated
986 * immediately, which means that the index file may end up more fragmented
987 * than if we forced it all to be allocated now; but since we don't scan
988 * hash indexes sequentially anyway, that probably doesn't matter.
989 *
990 * XXX It's annoying that this code is executed with the metapage lock held.
991 * We need to interlock against _hash_addovflpage() adding a new overflow page
992 * concurrently, but it'd likely be better to use LockRelationForExtension
993 * for the purpose. OTOH, adding a splitpoint is a very infrequent operation,
994 * so it may not be worth worrying about.
995 *
996 * Returns true if successful, or false if allocation failed due to
997 * BlockNumber overflow.
998 */
999 static bool
_hash_alloc_buckets(Relation rel,BlockNumber firstblock,uint32 nblocks)1000 _hash_alloc_buckets(Relation rel, BlockNumber firstblock, uint32 nblocks)
1001 {
1002 BlockNumber lastblock;
1003 PGAlignedBlock zerobuf;
1004 Page page;
1005 HashPageOpaque ovflopaque;
1006
1007 lastblock = firstblock + nblocks - 1;
1008
1009 /*
1010 * Check for overflow in block number calculation; if so, we cannot extend
1011 * the index anymore.
1012 */
1013 if (lastblock < firstblock || lastblock == InvalidBlockNumber)
1014 return false;
1015
1016 page = (Page) zerobuf.data;
1017
1018 /*
1019 * Initialize the page. Just zeroing the page won't work; see
1020 * _hash_freeovflpage for similar usage. We take care to make the special
1021 * space valid for the benefit of tools such as pageinspect.
1022 */
1023 _hash_pageinit(page, BLCKSZ);
1024
1025 ovflopaque = (HashPageOpaque) PageGetSpecialPointer(page);
1026
1027 ovflopaque->hasho_prevblkno = InvalidBlockNumber;
1028 ovflopaque->hasho_nextblkno = InvalidBlockNumber;
1029 ovflopaque->hasho_bucket = -1;
1030 ovflopaque->hasho_flag = LH_UNUSED_PAGE;
1031 ovflopaque->hasho_page_id = HASHO_PAGE_ID;
1032
1033 if (RelationNeedsWAL(rel))
1034 log_newpage(&rel->rd_node,
1035 MAIN_FORKNUM,
1036 lastblock,
1037 zerobuf.data,
1038 true);
1039
1040 RelationOpenSmgr(rel);
1041 PageSetChecksumInplace(page, lastblock);
1042 smgrextend(rel->rd_smgr, MAIN_FORKNUM, lastblock, zerobuf.data, false);
1043
1044 return true;
1045 }
1046
1047
1048 /*
1049 * _hash_splitbucket -- split 'obucket' into 'obucket' and 'nbucket'
1050 *
1051 * This routine is used to partition the tuples between old and new bucket and
1052 * is used to finish the incomplete split operations. To finish the previously
1053 * interrupted split operation, the caller needs to fill htab. If htab is set,
1054 * then we skip the movement of tuples that exists in htab, otherwise NULL
1055 * value of htab indicates movement of all the tuples that belong to the new
1056 * bucket.
1057 *
1058 * We are splitting a bucket that consists of a base bucket page and zero
1059 * or more overflow (bucket chain) pages. We must relocate tuples that
1060 * belong in the new bucket.
1061 *
1062 * The caller must hold cleanup locks on both buckets to ensure that
1063 * no one else is trying to access them (see README).
1064 *
1065 * The caller must hold a pin, but no lock, on the metapage buffer.
1066 * The buffer is returned in the same state. (The metapage is only
1067 * touched if it becomes necessary to add or remove overflow pages.)
1068 *
1069 * Split needs to retain pin on primary bucket pages of both old and new
1070 * buckets till end of operation. This is to prevent vacuum from starting
1071 * while a split is in progress.
1072 *
1073 * In addition, the caller must have created the new bucket's base page,
1074 * which is passed in buffer nbuf, pinned and write-locked. The lock will be
1075 * released here and pin must be released by the caller. (The API is set up
1076 * this way because we must do _hash_getnewbuf() before releasing the metapage
1077 * write lock. So instead of passing the new bucket's start block number, we
1078 * pass an actual buffer.)
1079 */
1080 static void
_hash_splitbucket(Relation rel,Buffer metabuf,Bucket obucket,Bucket nbucket,Buffer obuf,Buffer nbuf,HTAB * htab,uint32 maxbucket,uint32 highmask,uint32 lowmask)1081 _hash_splitbucket(Relation rel,
1082 Buffer metabuf,
1083 Bucket obucket,
1084 Bucket nbucket,
1085 Buffer obuf,
1086 Buffer nbuf,
1087 HTAB *htab,
1088 uint32 maxbucket,
1089 uint32 highmask,
1090 uint32 lowmask)
1091 {
1092 Buffer bucket_obuf;
1093 Buffer bucket_nbuf;
1094 Page opage;
1095 Page npage;
1096 HashPageOpaque oopaque;
1097 HashPageOpaque nopaque;
1098 OffsetNumber itup_offsets[MaxIndexTuplesPerPage];
1099 IndexTuple itups[MaxIndexTuplesPerPage];
1100 Size all_tups_size = 0;
1101 int i;
1102 uint16 nitups = 0;
1103
1104 bucket_obuf = obuf;
1105 opage = BufferGetPage(obuf);
1106 oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
1107
1108 bucket_nbuf = nbuf;
1109 npage = BufferGetPage(nbuf);
1110 nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
1111
1112 /* Copy the predicate locks from old bucket to new bucket. */
1113 PredicateLockPageSplit(rel,
1114 BufferGetBlockNumber(bucket_obuf),
1115 BufferGetBlockNumber(bucket_nbuf));
1116
1117 /*
1118 * Partition the tuples in the old bucket between the old bucket and the
1119 * new bucket, advancing along the old bucket's overflow bucket chain and
1120 * adding overflow pages to the new bucket as needed. Outer loop iterates
1121 * once per page in old bucket.
1122 */
1123 for (;;)
1124 {
1125 BlockNumber oblkno;
1126 OffsetNumber ooffnum;
1127 OffsetNumber omaxoffnum;
1128
1129 /* Scan each tuple in old page */
1130 omaxoffnum = PageGetMaxOffsetNumber(opage);
1131 for (ooffnum = FirstOffsetNumber;
1132 ooffnum <= omaxoffnum;
1133 ooffnum = OffsetNumberNext(ooffnum))
1134 {
1135 IndexTuple itup;
1136 Size itemsz;
1137 Bucket bucket;
1138 bool found = false;
1139
1140 /* skip dead tuples */
1141 if (ItemIdIsDead(PageGetItemId(opage, ooffnum)))
1142 continue;
1143
1144 /*
1145 * Before inserting a tuple, probe the hash table containing TIDs
1146 * of tuples belonging to new bucket, if we find a match, then
1147 * skip that tuple, else fetch the item's hash key (conveniently
1148 * stored in the item) and determine which bucket it now belongs
1149 * in.
1150 */
1151 itup = (IndexTuple) PageGetItem(opage,
1152 PageGetItemId(opage, ooffnum));
1153
1154 if (htab)
1155 (void) hash_search(htab, &itup->t_tid, HASH_FIND, &found);
1156
1157 if (found)
1158 continue;
1159
1160 bucket = _hash_hashkey2bucket(_hash_get_indextuple_hashkey(itup),
1161 maxbucket, highmask, lowmask);
1162
1163 if (bucket == nbucket)
1164 {
1165 IndexTuple new_itup;
1166
1167 /*
1168 * make a copy of index tuple as we have to scribble on it.
1169 */
1170 new_itup = CopyIndexTuple(itup);
1171
1172 /*
1173 * mark the index tuple as moved by split, such tuples are
1174 * skipped by scan if there is split in progress for a bucket.
1175 */
1176 new_itup->t_info |= INDEX_MOVED_BY_SPLIT_MASK;
1177
1178 /*
1179 * insert the tuple into the new bucket. if it doesn't fit on
1180 * the current page in the new bucket, we must allocate a new
1181 * overflow page and place the tuple on that page instead.
1182 */
1183 itemsz = IndexTupleSize(new_itup);
1184 itemsz = MAXALIGN(itemsz);
1185
1186 if (PageGetFreeSpaceForMultipleTuples(npage, nitups + 1) < (all_tups_size + itemsz))
1187 {
1188 /*
1189 * Change the shared buffer state in critical section,
1190 * otherwise any error could make it unrecoverable.
1191 */
1192 START_CRIT_SECTION();
1193
1194 _hash_pgaddmultitup(rel, nbuf, itups, itup_offsets, nitups);
1195 MarkBufferDirty(nbuf);
1196 /* log the split operation before releasing the lock */
1197 log_split_page(rel, nbuf);
1198
1199 END_CRIT_SECTION();
1200
1201 /* drop lock, but keep pin */
1202 LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
1203
1204 /* be tidy */
1205 for (i = 0; i < nitups; i++)
1206 pfree(itups[i]);
1207 nitups = 0;
1208 all_tups_size = 0;
1209
1210 /* chain to a new overflow page */
1211 nbuf = _hash_addovflpage(rel, metabuf, nbuf, (nbuf == bucket_nbuf) ? true : false);
1212 npage = BufferGetPage(nbuf);
1213 nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
1214 }
1215
1216 itups[nitups++] = new_itup;
1217 all_tups_size += itemsz;
1218 }
1219 else
1220 {
1221 /*
1222 * the tuple stays on this page, so nothing to do.
1223 */
1224 Assert(bucket == obucket);
1225 }
1226 }
1227
1228 oblkno = oopaque->hasho_nextblkno;
1229
1230 /* retain the pin on the old primary bucket */
1231 if (obuf == bucket_obuf)
1232 LockBuffer(obuf, BUFFER_LOCK_UNLOCK);
1233 else
1234 _hash_relbuf(rel, obuf);
1235
1236 /* Exit loop if no more overflow pages in old bucket */
1237 if (!BlockNumberIsValid(oblkno))
1238 {
1239 /*
1240 * Change the shared buffer state in critical section, otherwise
1241 * any error could make it unrecoverable.
1242 */
1243 START_CRIT_SECTION();
1244
1245 _hash_pgaddmultitup(rel, nbuf, itups, itup_offsets, nitups);
1246 MarkBufferDirty(nbuf);
1247 /* log the split operation before releasing the lock */
1248 log_split_page(rel, nbuf);
1249
1250 END_CRIT_SECTION();
1251
1252 if (nbuf == bucket_nbuf)
1253 LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
1254 else
1255 _hash_relbuf(rel, nbuf);
1256
1257 /* be tidy */
1258 for (i = 0; i < nitups; i++)
1259 pfree(itups[i]);
1260 break;
1261 }
1262
1263 /* Else, advance to next old page */
1264 obuf = _hash_getbuf(rel, oblkno, HASH_READ, LH_OVERFLOW_PAGE);
1265 opage = BufferGetPage(obuf);
1266 oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
1267 }
1268
1269 /*
1270 * We're at the end of the old bucket chain, so we're done partitioning
1271 * the tuples. Mark the old and new buckets to indicate split is
1272 * finished.
1273 *
1274 * To avoid deadlocks due to locking order of buckets, first lock the old
1275 * bucket and then the new bucket.
1276 */
1277 LockBuffer(bucket_obuf, BUFFER_LOCK_EXCLUSIVE);
1278 opage = BufferGetPage(bucket_obuf);
1279 oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
1280
1281 LockBuffer(bucket_nbuf, BUFFER_LOCK_EXCLUSIVE);
1282 npage = BufferGetPage(bucket_nbuf);
1283 nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
1284
1285 START_CRIT_SECTION();
1286
1287 oopaque->hasho_flag &= ~LH_BUCKET_BEING_SPLIT;
1288 nopaque->hasho_flag &= ~LH_BUCKET_BEING_POPULATED;
1289
1290 /*
1291 * After the split is finished, mark the old bucket to indicate that it
1292 * contains deletable tuples. We will clear split-cleanup flag after
1293 * deleting such tuples either at the end of split or at the next split
1294 * from old bucket or at the time of vacuum.
1295 */
1296 oopaque->hasho_flag |= LH_BUCKET_NEEDS_SPLIT_CLEANUP;
1297
1298 /*
1299 * now write the buffers, here we don't release the locks as caller is
1300 * responsible to release locks.
1301 */
1302 MarkBufferDirty(bucket_obuf);
1303 MarkBufferDirty(bucket_nbuf);
1304
1305 if (RelationNeedsWAL(rel))
1306 {
1307 XLogRecPtr recptr;
1308 xl_hash_split_complete xlrec;
1309
1310 xlrec.old_bucket_flag = oopaque->hasho_flag;
1311 xlrec.new_bucket_flag = nopaque->hasho_flag;
1312
1313 XLogBeginInsert();
1314
1315 XLogRegisterData((char *) &xlrec, SizeOfHashSplitComplete);
1316
1317 XLogRegisterBuffer(0, bucket_obuf, REGBUF_STANDARD);
1318 XLogRegisterBuffer(1, bucket_nbuf, REGBUF_STANDARD);
1319
1320 recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_COMPLETE);
1321
1322 PageSetLSN(BufferGetPage(bucket_obuf), recptr);
1323 PageSetLSN(BufferGetPage(bucket_nbuf), recptr);
1324 }
1325
1326 END_CRIT_SECTION();
1327
1328 /*
1329 * If possible, clean up the old bucket. We might not be able to do this
1330 * if someone else has a pin on it, but if not then we can go ahead. This
1331 * isn't absolutely necessary, but it reduces bloat; if we don't do it
1332 * now, VACUUM will do it eventually, but maybe not until new overflow
1333 * pages have been allocated. Note that there's no need to clean up the
1334 * new bucket.
1335 */
1336 if (IsBufferCleanupOK(bucket_obuf))
1337 {
1338 LockBuffer(bucket_nbuf, BUFFER_LOCK_UNLOCK);
1339 hashbucketcleanup(rel, obucket, bucket_obuf,
1340 BufferGetBlockNumber(bucket_obuf), NULL,
1341 maxbucket, highmask, lowmask, NULL, NULL, true,
1342 NULL, NULL);
1343 }
1344 else
1345 {
1346 LockBuffer(bucket_nbuf, BUFFER_LOCK_UNLOCK);
1347 LockBuffer(bucket_obuf, BUFFER_LOCK_UNLOCK);
1348 }
1349 }
1350
1351 /*
1352 * _hash_finish_split() -- Finish the previously interrupted split operation
1353 *
1354 * To complete the split operation, we form the hash table of TIDs in new
1355 * bucket which is then used by split operation to skip tuples that are
1356 * already moved before the split operation was previously interrupted.
1357 *
1358 * The caller must hold a pin, but no lock, on the metapage and old bucket's
1359 * primary page buffer. The buffers are returned in the same state. (The
1360 * metapage is only touched if it becomes necessary to add or remove overflow
1361 * pages.)
1362 */
1363 void
_hash_finish_split(Relation rel,Buffer metabuf,Buffer obuf,Bucket obucket,uint32 maxbucket,uint32 highmask,uint32 lowmask)1364 _hash_finish_split(Relation rel, Buffer metabuf, Buffer obuf, Bucket obucket,
1365 uint32 maxbucket, uint32 highmask, uint32 lowmask)
1366 {
1367 HASHCTL hash_ctl;
1368 HTAB *tidhtab;
1369 Buffer bucket_nbuf = InvalidBuffer;
1370 Buffer nbuf;
1371 Page npage;
1372 BlockNumber nblkno;
1373 BlockNumber bucket_nblkno;
1374 HashPageOpaque npageopaque;
1375 Bucket nbucket;
1376 bool found;
1377
1378 /* Initialize hash tables used to track TIDs */
1379 memset(&hash_ctl, 0, sizeof(hash_ctl));
1380 hash_ctl.keysize = sizeof(ItemPointerData);
1381 hash_ctl.entrysize = sizeof(ItemPointerData);
1382 hash_ctl.hcxt = CurrentMemoryContext;
1383
1384 tidhtab =
1385 hash_create("bucket ctids",
1386 256, /* arbitrary initial size */
1387 &hash_ctl,
1388 HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1389
1390 bucket_nblkno = nblkno = _hash_get_newblock_from_oldbucket(rel, obucket);
1391
1392 /*
1393 * Scan the new bucket and build hash table of TIDs
1394 */
1395 for (;;)
1396 {
1397 OffsetNumber noffnum;
1398 OffsetNumber nmaxoffnum;
1399
1400 nbuf = _hash_getbuf(rel, nblkno, HASH_READ,
1401 LH_BUCKET_PAGE | LH_OVERFLOW_PAGE);
1402
1403 /* remember the primary bucket buffer to acquire cleanup lock on it. */
1404 if (nblkno == bucket_nblkno)
1405 bucket_nbuf = nbuf;
1406
1407 npage = BufferGetPage(nbuf);
1408 npageopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
1409
1410 /* Scan each tuple in new page */
1411 nmaxoffnum = PageGetMaxOffsetNumber(npage);
1412 for (noffnum = FirstOffsetNumber;
1413 noffnum <= nmaxoffnum;
1414 noffnum = OffsetNumberNext(noffnum))
1415 {
1416 IndexTuple itup;
1417
1418 /* Fetch the item's TID and insert it in hash table. */
1419 itup = (IndexTuple) PageGetItem(npage,
1420 PageGetItemId(npage, noffnum));
1421
1422 (void) hash_search(tidhtab, &itup->t_tid, HASH_ENTER, &found);
1423
1424 Assert(!found);
1425 }
1426
1427 nblkno = npageopaque->hasho_nextblkno;
1428
1429 /*
1430 * release our write lock without modifying buffer and ensure to
1431 * retain the pin on primary bucket.
1432 */
1433 if (nbuf == bucket_nbuf)
1434 LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
1435 else
1436 _hash_relbuf(rel, nbuf);
1437
1438 /* Exit loop if no more overflow pages in new bucket */
1439 if (!BlockNumberIsValid(nblkno))
1440 break;
1441 }
1442
1443 /*
1444 * Conditionally get the cleanup lock on old and new buckets to perform
1445 * the split operation. If we don't get the cleanup locks, silently give
1446 * up and next insertion on old bucket will try again to complete the
1447 * split.
1448 */
1449 if (!ConditionalLockBufferForCleanup(obuf))
1450 {
1451 hash_destroy(tidhtab);
1452 return;
1453 }
1454 if (!ConditionalLockBufferForCleanup(bucket_nbuf))
1455 {
1456 LockBuffer(obuf, BUFFER_LOCK_UNLOCK);
1457 hash_destroy(tidhtab);
1458 return;
1459 }
1460
1461 npage = BufferGetPage(bucket_nbuf);
1462 npageopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
1463 nbucket = npageopaque->hasho_bucket;
1464
1465 _hash_splitbucket(rel, metabuf, obucket,
1466 nbucket, obuf, bucket_nbuf, tidhtab,
1467 maxbucket, highmask, lowmask);
1468
1469 _hash_dropbuf(rel, bucket_nbuf);
1470 hash_destroy(tidhtab);
1471 }
1472
1473 /*
1474 * log_split_page() -- Log the split operation
1475 *
1476 * We log the split operation when the new page in new bucket gets full,
1477 * so we log the entire page.
1478 *
1479 * 'buf' must be locked by the caller which is also responsible for unlocking
1480 * it.
1481 */
1482 static void
log_split_page(Relation rel,Buffer buf)1483 log_split_page(Relation rel, Buffer buf)
1484 {
1485 if (RelationNeedsWAL(rel))
1486 {
1487 XLogRecPtr recptr;
1488
1489 XLogBeginInsert();
1490
1491 XLogRegisterBuffer(0, buf, REGBUF_FORCE_IMAGE | REGBUF_STANDARD);
1492
1493 recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_PAGE);
1494
1495 PageSetLSN(BufferGetPage(buf), recptr);
1496 }
1497 }
1498
1499 /*
1500 * _hash_getcachedmetap() -- Returns cached metapage data.
1501 *
1502 * If metabuf is not InvalidBuffer, caller must hold a pin, but no lock, on
1503 * the metapage. If not set, we'll set it before returning if we have to
1504 * refresh the cache, and return with a pin but no lock on it; caller is
1505 * responsible for releasing the pin.
1506 *
1507 * We refresh the cache if it's not initialized yet or force_refresh is true.
1508 */
1509 HashMetaPage
_hash_getcachedmetap(Relation rel,Buffer * metabuf,bool force_refresh)1510 _hash_getcachedmetap(Relation rel, Buffer *metabuf, bool force_refresh)
1511 {
1512 Page page;
1513
1514 Assert(metabuf);
1515 if (force_refresh || rel->rd_amcache == NULL)
1516 {
1517 char *cache = NULL;
1518
1519 /*
1520 * It's important that we don't set rd_amcache to an invalid value.
1521 * Either MemoryContextAlloc or _hash_getbuf could fail, so don't
1522 * install a pointer to the newly-allocated storage in the actual
1523 * relcache entry until both have succeeeded.
1524 */
1525 if (rel->rd_amcache == NULL)
1526 cache = MemoryContextAlloc(rel->rd_indexcxt,
1527 sizeof(HashMetaPageData));
1528
1529 /* Read the metapage. */
1530 if (BufferIsValid(*metabuf))
1531 LockBuffer(*metabuf, BUFFER_LOCK_SHARE);
1532 else
1533 *metabuf = _hash_getbuf(rel, HASH_METAPAGE, HASH_READ,
1534 LH_META_PAGE);
1535 page = BufferGetPage(*metabuf);
1536
1537 /* Populate the cache. */
1538 if (rel->rd_amcache == NULL)
1539 rel->rd_amcache = cache;
1540 memcpy(rel->rd_amcache, HashPageGetMeta(page),
1541 sizeof(HashMetaPageData));
1542
1543 /* Release metapage lock, but keep the pin. */
1544 LockBuffer(*metabuf, BUFFER_LOCK_UNLOCK);
1545 }
1546
1547 return (HashMetaPage) rel->rd_amcache;
1548 }
1549
1550 /*
1551 * _hash_getbucketbuf_from_hashkey() -- Get the bucket's buffer for the given
1552 * hashkey.
1553 *
1554 * Bucket pages do not move or get removed once they are allocated. This give
1555 * us an opportunity to use the previously saved metapage contents to reach
1556 * the target bucket buffer, instead of reading from the metapage every time.
1557 * This saves one buffer access every time we want to reach the target bucket
1558 * buffer, which is very helpful savings in bufmgr traffic and contention.
1559 *
1560 * The access type parameter (HASH_READ or HASH_WRITE) indicates whether the
1561 * bucket buffer has to be locked for reading or writing.
1562 *
1563 * The out parameter cachedmetap is set with metapage contents used for
1564 * hashkey to bucket buffer mapping. Some callers need this info to reach the
1565 * old bucket in case of bucket split, see _hash_doinsert().
1566 */
1567 Buffer
_hash_getbucketbuf_from_hashkey(Relation rel,uint32 hashkey,int access,HashMetaPage * cachedmetap)1568 _hash_getbucketbuf_from_hashkey(Relation rel, uint32 hashkey, int access,
1569 HashMetaPage *cachedmetap)
1570 {
1571 HashMetaPage metap;
1572 Buffer buf;
1573 Buffer metabuf = InvalidBuffer;
1574 Page page;
1575 Bucket bucket;
1576 BlockNumber blkno;
1577 HashPageOpaque opaque;
1578
1579 /* We read from target bucket buffer, hence locking is must. */
1580 Assert(access == HASH_READ || access == HASH_WRITE);
1581
1582 metap = _hash_getcachedmetap(rel, &metabuf, false);
1583 Assert(metap != NULL);
1584
1585 /*
1586 * Loop until we get a lock on the correct target bucket.
1587 */
1588 for (;;)
1589 {
1590 /*
1591 * Compute the target bucket number, and convert to block number.
1592 */
1593 bucket = _hash_hashkey2bucket(hashkey,
1594 metap->hashm_maxbucket,
1595 metap->hashm_highmask,
1596 metap->hashm_lowmask);
1597
1598 blkno = BUCKET_TO_BLKNO(metap, bucket);
1599
1600 /* Fetch the primary bucket page for the bucket */
1601 buf = _hash_getbuf(rel, blkno, access, LH_BUCKET_PAGE);
1602 page = BufferGetPage(buf);
1603 opaque = (HashPageOpaque) PageGetSpecialPointer(page);
1604 Assert(opaque->hasho_bucket == bucket);
1605 Assert(opaque->hasho_prevblkno != InvalidBlockNumber);
1606
1607 /*
1608 * If this bucket hasn't been split, we're done.
1609 */
1610 if (opaque->hasho_prevblkno <= metap->hashm_maxbucket)
1611 break;
1612
1613 /* Drop lock on this buffer, update cached metapage, and retry. */
1614 _hash_relbuf(rel, buf);
1615 metap = _hash_getcachedmetap(rel, &metabuf, true);
1616 Assert(metap != NULL);
1617 }
1618
1619 if (BufferIsValid(metabuf))
1620 _hash_dropbuf(rel, metabuf);
1621
1622 if (cachedmetap)
1623 *cachedmetap = metap;
1624
1625 return buf;
1626 }
1627