Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * hashpage.c
4 : * Hash table page management code for the Postgres hash access method
5 : *
6 : * Portions Copyright (c) 1996-2026, 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 "access/xloginsert.h"
34 : #include "miscadmin.h"
35 : #include "port/pg_bitutils.h"
36 : #include "storage/predicate.h"
37 : #include "storage/smgr.h"
38 : #include "utils/rel.h"
39 :
40 : static bool _hash_alloc_buckets(Relation rel, BlockNumber firstblock,
41 : uint32 nblocks);
42 : static void _hash_splitbucket(Relation rel, Buffer metabuf,
43 : Bucket obucket, Bucket nbucket,
44 : Buffer obuf,
45 : Buffer nbuf,
46 : HTAB *htab,
47 : uint32 maxbucket,
48 : uint32 highmask, uint32 lowmask);
49 : static void log_split_page(Relation rel, Buffer buf);
50 :
51 :
52 : /*
53 : * _hash_getbuf() -- Get a buffer by block number for read or write.
54 : *
55 : * 'access' must be HASH_READ, HASH_WRITE, or HASH_NOLOCK.
56 : * 'flags' is a bitwise OR of the allowed page types.
57 : *
58 : * This must be used only to fetch pages that are expected to be valid
59 : * already. _hash_checkpage() is applied using the given flags.
60 : *
61 : * When this routine returns, the appropriate lock is set on the
62 : * requested buffer and its reference count has been incremented
63 : * (ie, the buffer is "locked and pinned").
64 : *
65 : * P_NEW is disallowed because this routine can only be used
66 : * to access pages that are known to be before the filesystem EOF.
67 : * Extending the index should be done with _hash_getnewbuf.
68 : */
69 : Buffer
70 1340027 : _hash_getbuf(Relation rel, BlockNumber blkno, int access, int flags)
71 : {
72 : Buffer buf;
73 :
74 1340027 : if (blkno == P_NEW)
75 0 : elog(ERROR, "hash AM does not use P_NEW");
76 :
77 1340027 : buf = ReadBuffer(rel, blkno);
78 :
79 1340027 : if (access != HASH_NOLOCK)
80 849234 : LockBuffer(buf, access);
81 :
82 : /* ref count and lock type are correct */
83 :
84 1340027 : _hash_checkpage(rel, buf, flags);
85 :
86 1340027 : return buf;
87 : }
88 :
89 : /*
90 : * _hash_getbuf_with_condlock_cleanup() -- Try to get a buffer for cleanup.
91 : *
92 : * We read the page and try to acquire a cleanup lock. If we get it,
93 : * we return the buffer; otherwise, we return InvalidBuffer.
94 : */
95 : Buffer
96 890 : _hash_getbuf_with_condlock_cleanup(Relation rel, BlockNumber blkno, int flags)
97 : {
98 : Buffer buf;
99 :
100 890 : if (blkno == P_NEW)
101 0 : elog(ERROR, "hash AM does not use P_NEW");
102 :
103 890 : buf = ReadBuffer(rel, blkno);
104 :
105 890 : if (!ConditionalLockBufferForCleanup(buf))
106 : {
107 0 : ReleaseBuffer(buf);
108 0 : return InvalidBuffer;
109 : }
110 :
111 : /* ref count and lock type are correct */
112 :
113 890 : _hash_checkpage(rel, buf, flags);
114 :
115 890 : return buf;
116 : }
117 :
118 : /*
119 : * _hash_getinitbuf() -- Get and initialize a buffer by block number.
120 : *
121 : * This must be used only to fetch pages that are known to be before
122 : * the index's filesystem EOF, but are to be filled from scratch.
123 : * _hash_pageinit() is applied automatically. Otherwise it has
124 : * effects similar to _hash_getbuf() with access = HASH_WRITE.
125 : *
126 : * When this routine returns, a write lock is set on the
127 : * requested buffer and its reference count has been incremented
128 : * (ie, the buffer is "locked and pinned").
129 : *
130 : * P_NEW is disallowed because this routine can only be used
131 : * to access pages that are known to be before the filesystem EOF.
132 : * Extending the index should be done with _hash_getnewbuf.
133 : */
134 : Buffer
135 46 : _hash_getinitbuf(Relation rel, BlockNumber blkno)
136 : {
137 : Buffer buf;
138 :
139 46 : if (blkno == P_NEW)
140 0 : elog(ERROR, "hash AM does not use P_NEW");
141 :
142 46 : buf = ReadBufferExtended(rel, MAIN_FORKNUM, blkno, RBM_ZERO_AND_LOCK,
143 : NULL);
144 :
145 : /* ref count and lock type are correct */
146 :
147 : /* initialize the page */
148 46 : _hash_pageinit(BufferGetPage(buf), BufferGetPageSize(buf));
149 :
150 46 : return buf;
151 : }
152 :
153 : /*
154 : * _hash_initbuf() -- Get and initialize a buffer by bucket number.
155 : */
156 : void
157 5571 : _hash_initbuf(Buffer buf, uint32 max_bucket, uint32 num_bucket, uint32 flag,
158 : bool initpage)
159 : {
160 : HashPageOpaque pageopaque;
161 : Page page;
162 :
163 5571 : page = BufferGetPage(buf);
164 :
165 : /* initialize the page */
166 5571 : if (initpage)
167 291 : _hash_pageinit(page, BufferGetPageSize(buf));
168 :
169 5571 : pageopaque = HashPageGetOpaque(page);
170 :
171 : /*
172 : * Set hasho_prevblkno with current hashm_maxbucket. This value will be
173 : * used to validate cached HashMetaPageData. See
174 : * _hash_getbucketbuf_from_hashkey().
175 : */
176 5571 : pageopaque->hasho_prevblkno = max_bucket;
177 5571 : pageopaque->hasho_nextblkno = InvalidBlockNumber;
178 5571 : pageopaque->hasho_bucket = num_bucket;
179 5571 : pageopaque->hasho_flag = flag;
180 5571 : pageopaque->hasho_page_id = HASHO_PAGE_ID;
181 5571 : }
182 :
183 : /*
184 : * _hash_getnewbuf() -- Get a new page at the end of the index.
185 : *
186 : * This has the same API as _hash_getinitbuf, except that we are adding
187 : * a page to the index, and hence expect the page to be past the
188 : * logical EOF. (However, we have to support the case where it isn't,
189 : * since a prior try might have crashed after extending the filesystem
190 : * EOF but before updating the metapage to reflect the added page.)
191 : *
192 : * It is caller's responsibility to ensure that only one process can
193 : * extend the index at a time. In practice, this function is called
194 : * only while holding write lock on the metapage, because adding a page
195 : * is always associated with an update of metapage data.
196 : */
197 : Buffer
198 6825 : _hash_getnewbuf(Relation rel, BlockNumber blkno, ForkNumber forkNum)
199 : {
200 6825 : BlockNumber nblocks = RelationGetNumberOfBlocksInFork(rel, forkNum);
201 : Buffer buf;
202 :
203 6825 : if (blkno == P_NEW)
204 0 : elog(ERROR, "hash AM does not use P_NEW");
205 6825 : if (blkno > nblocks)
206 0 : elog(ERROR, "access to noncontiguous page in hash index \"%s\"",
207 : RelationGetRelationName(rel));
208 :
209 : /* smgr insists we explicitly extend the relation */
210 6825 : if (blkno == nblocks)
211 : {
212 5935 : buf = ExtendBufferedRel(BMR_REL(rel), forkNum, NULL,
213 : EB_LOCK_FIRST | EB_SKIP_EXTENSION_LOCK);
214 5935 : if (BufferGetBlockNumber(buf) != blkno)
215 0 : elog(ERROR, "unexpected hash relation size: %u, should be %u",
216 : BufferGetBlockNumber(buf), blkno);
217 : }
218 : else
219 : {
220 890 : buf = ReadBufferExtended(rel, forkNum, blkno, RBM_ZERO_AND_LOCK,
221 : NULL);
222 : }
223 :
224 : /* ref count and lock type are correct */
225 :
226 : /* initialize the page */
227 6825 : _hash_pageinit(BufferGetPage(buf), BufferGetPageSize(buf));
228 :
229 6825 : return buf;
230 : }
231 :
232 : /*
233 : * _hash_getbuf_with_strategy() -- Get a buffer with nondefault strategy.
234 : *
235 : * This is identical to _hash_getbuf() but also allows a buffer access
236 : * strategy to be specified. We use this for VACUUM operations.
237 : */
238 : Buffer
239 800 : _hash_getbuf_with_strategy(Relation rel, BlockNumber blkno,
240 : int access, int flags,
241 : BufferAccessStrategy bstrategy)
242 : {
243 : Buffer buf;
244 :
245 800 : if (blkno == P_NEW)
246 0 : elog(ERROR, "hash AM does not use P_NEW");
247 :
248 800 : buf = ReadBufferExtended(rel, MAIN_FORKNUM, blkno, RBM_NORMAL, bstrategy);
249 :
250 800 : if (access != HASH_NOLOCK)
251 800 : LockBuffer(buf, access);
252 :
253 : /* ref count and lock type are correct */
254 :
255 800 : _hash_checkpage(rel, buf, flags);
256 :
257 800 : return buf;
258 : }
259 :
260 : /*
261 : * _hash_relbuf() -- release a locked buffer.
262 : *
263 : * Lock and pin (refcount) are both dropped.
264 : */
265 : void
266 787924 : _hash_relbuf(Relation rel, Buffer buf)
267 : {
268 787924 : UnlockReleaseBuffer(buf);
269 787924 : }
270 :
271 : /*
272 : * _hash_dropbuf() -- release an unlocked buffer.
273 : *
274 : * This is used to unpin a buffer on which we hold no lock.
275 : */
276 : void
277 561126 : _hash_dropbuf(Relation rel, Buffer buf)
278 : {
279 561126 : ReleaseBuffer(buf);
280 561126 : }
281 :
282 : /*
283 : * _hash_dropscanbuf() -- release buffers used in scan.
284 : *
285 : * This routine unpins the buffers used during scan on which we
286 : * hold no lock.
287 : */
288 : void
289 832 : _hash_dropscanbuf(Relation rel, HashScanOpaque so)
290 : {
291 : /* release pin we hold on primary bucket page */
292 832 : if (BufferIsValid(so->hashso_bucket_buf) &&
293 341 : so->hashso_bucket_buf != so->currPos.buf)
294 93 : _hash_dropbuf(rel, so->hashso_bucket_buf);
295 832 : so->hashso_bucket_buf = InvalidBuffer;
296 :
297 : /* release pin we hold on primary bucket page of bucket being split */
298 832 : if (BufferIsValid(so->hashso_split_bucket_buf) &&
299 0 : so->hashso_split_bucket_buf != so->currPos.buf)
300 0 : _hash_dropbuf(rel, so->hashso_split_bucket_buf);
301 832 : so->hashso_split_bucket_buf = InvalidBuffer;
302 :
303 : /* release any pin we still hold */
304 832 : if (BufferIsValid(so->currPos.buf))
305 248 : _hash_dropbuf(rel, so->currPos.buf);
306 832 : so->currPos.buf = InvalidBuffer;
307 :
308 : /* reset split scan */
309 832 : so->hashso_buc_populated = false;
310 832 : so->hashso_buc_split = false;
311 832 : }
312 :
313 :
314 : /*
315 : * _hash_init() -- Initialize the metadata page of a hash index,
316 : * the initial buckets, and the initial bitmap page.
317 : *
318 : * The initial number of buckets is dependent on num_tuples, an estimate
319 : * of the number of tuples to be loaded into the index initially. The
320 : * chosen number of buckets is returned.
321 : *
322 : * We are fairly cavalier about locking here, since we know that no one else
323 : * could be accessing this index. In particular the rule about not holding
324 : * multiple buffer locks is ignored.
325 : */
326 : uint32
327 211 : _hash_init(Relation rel, double num_tuples, ForkNumber forkNum)
328 : {
329 : Buffer metabuf;
330 : Buffer buf;
331 : Buffer bitmapbuf;
332 : Page pg;
333 : HashMetaPage metap;
334 : RegProcedure procid;
335 : int32 data_width;
336 : int32 item_width;
337 : int32 ffactor;
338 : uint32 num_buckets;
339 : uint32 i;
340 : bool use_wal;
341 :
342 : /* safety check */
343 211 : if (RelationGetNumberOfBlocksInFork(rel, forkNum) != 0)
344 0 : elog(ERROR, "cannot initialize non-empty hash index \"%s\"",
345 : RelationGetRelationName(rel));
346 :
347 : /*
348 : * WAL log creation of pages if the relation is persistent, or this is the
349 : * init fork. Init forks for unlogged relations always need to be WAL
350 : * logged.
351 : */
352 211 : use_wal = RelationNeedsWAL(rel) || forkNum == INIT_FORKNUM;
353 :
354 : /*
355 : * Determine the target fill factor (in tuples per bucket) for this index.
356 : * The idea is to make the fill factor correspond to pages about as full
357 : * as the user-settable fillfactor parameter says. We can compute it
358 : * exactly since the index datatype (i.e. uint32 hash key) is fixed-width.
359 : */
360 211 : data_width = sizeof(uint32);
361 211 : item_width = MAXALIGN(sizeof(IndexTupleData)) + MAXALIGN(data_width) +
362 : sizeof(ItemIdData); /* include the line pointer */
363 211 : ffactor = HashGetTargetPageUsage(rel) / item_width;
364 : /* keep to a sane range */
365 211 : if (ffactor < 10)
366 0 : ffactor = 10;
367 :
368 211 : procid = index_getprocid(rel, 1, HASHSTANDARD_PROC);
369 :
370 : /*
371 : * We initialize the metapage, the first N bucket pages, and the first
372 : * bitmap page in sequence, using _hash_getnewbuf to cause smgrextend()
373 : * calls to occur. This ensures that the smgr level has the right idea of
374 : * the physical index length.
375 : *
376 : * Critical section not required, because on error the creation of the
377 : * whole relation will be rolled back.
378 : */
379 211 : metabuf = _hash_getnewbuf(rel, HASH_METAPAGE, forkNum);
380 211 : _hash_init_metabuffer(metabuf, num_tuples, procid, ffactor, false);
381 211 : MarkBufferDirty(metabuf);
382 :
383 211 : pg = BufferGetPage(metabuf);
384 211 : metap = HashPageGetMeta(pg);
385 :
386 : /* XLOG stuff */
387 211 : if (use_wal)
388 : {
389 : xl_hash_init_meta_page xlrec;
390 : XLogRecPtr recptr;
391 :
392 137 : xlrec.num_tuples = num_tuples;
393 137 : xlrec.procid = metap->hashm_procid;
394 137 : xlrec.ffactor = metap->hashm_ffactor;
395 :
396 137 : XLogBeginInsert();
397 137 : XLogRegisterData(&xlrec, SizeOfHashInitMetaPage);
398 137 : XLogRegisterBuffer(0, metabuf, REGBUF_WILL_INIT | REGBUF_STANDARD);
399 :
400 137 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_INIT_META_PAGE);
401 :
402 137 : PageSetLSN(BufferGetPage(metabuf), recptr);
403 : }
404 :
405 211 : num_buckets = metap->hashm_maxbucket + 1;
406 :
407 : /*
408 : * Release buffer lock on the metapage while we initialize buckets.
409 : * Otherwise, we'll be in interrupt holdoff and the CHECK_FOR_INTERRUPTS
410 : * won't accomplish anything. It's a bad idea to hold buffer locks for
411 : * long intervals in any case, since that can block the bgwriter.
412 : */
413 211 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
414 :
415 : /*
416 : * Initialize and WAL Log the first N buckets
417 : */
418 5491 : for (i = 0; i < num_buckets; i++)
419 : {
420 : BlockNumber blkno;
421 :
422 : /* Allow interrupts, in case N is huge */
423 5280 : CHECK_FOR_INTERRUPTS();
424 :
425 5280 : blkno = BUCKET_TO_BLKNO(metap, i);
426 5280 : buf = _hash_getnewbuf(rel, blkno, forkNum);
427 5280 : _hash_initbuf(buf, metap->hashm_maxbucket, i, LH_BUCKET_PAGE, false);
428 5280 : MarkBufferDirty(buf);
429 :
430 5280 : if (use_wal)
431 3944 : log_newpage(&rel->rd_locator,
432 : forkNum,
433 : blkno,
434 : BufferGetPage(buf),
435 : true);
436 5280 : _hash_relbuf(rel, buf);
437 : }
438 :
439 : /* Now reacquire buffer lock on metapage */
440 211 : LockBuffer(metabuf, BUFFER_LOCK_EXCLUSIVE);
441 :
442 : /*
443 : * Initialize bitmap page
444 : */
445 211 : bitmapbuf = _hash_getnewbuf(rel, num_buckets + 1, forkNum);
446 211 : _hash_initbitmapbuffer(bitmapbuf, metap->hashm_bmsize, false);
447 211 : MarkBufferDirty(bitmapbuf);
448 :
449 : /* add the new bitmap page to the metapage's list of bitmaps */
450 : /* metapage already has a write lock */
451 211 : if (metap->hashm_nmaps >= HASH_MAX_BITMAPS)
452 0 : ereport(ERROR,
453 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
454 : errmsg("out of overflow pages in hash index \"%s\"",
455 : RelationGetRelationName(rel))));
456 :
457 211 : metap->hashm_mapp[metap->hashm_nmaps] = num_buckets + 1;
458 :
459 211 : metap->hashm_nmaps++;
460 211 : MarkBufferDirty(metabuf);
461 :
462 : /* XLOG stuff */
463 211 : if (use_wal)
464 : {
465 : xl_hash_init_bitmap_page xlrec;
466 : XLogRecPtr recptr;
467 :
468 137 : xlrec.bmsize = metap->hashm_bmsize;
469 :
470 137 : XLogBeginInsert();
471 137 : XLogRegisterData(&xlrec, SizeOfHashInitBitmapPage);
472 137 : XLogRegisterBuffer(0, bitmapbuf, REGBUF_WILL_INIT);
473 :
474 : /*
475 : * This is safe only because nobody else can be modifying the index at
476 : * this stage; it's only visible to the transaction that is creating
477 : * it.
478 : */
479 137 : XLogRegisterBuffer(1, metabuf, REGBUF_STANDARD);
480 :
481 137 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_INIT_BITMAP_PAGE);
482 :
483 137 : PageSetLSN(BufferGetPage(bitmapbuf), recptr);
484 137 : PageSetLSN(BufferGetPage(metabuf), recptr);
485 : }
486 :
487 : /* all done */
488 211 : _hash_relbuf(rel, bitmapbuf);
489 211 : _hash_relbuf(rel, metabuf);
490 :
491 211 : return num_buckets;
492 : }
493 :
494 : /*
495 : * _hash_init_metabuffer() -- Initialize the metadata page of a hash index.
496 : */
497 : void
498 240 : _hash_init_metabuffer(Buffer buf, double num_tuples, RegProcedure procid,
499 : uint16 ffactor, bool initpage)
500 : {
501 : HashMetaPage metap;
502 : HashPageOpaque pageopaque;
503 : Page page;
504 : double dnumbuckets;
505 : uint32 num_buckets;
506 : uint32 spare_index;
507 : uint32 lshift;
508 :
509 : /*
510 : * Choose the number of initial bucket pages to match the fill factor
511 : * given the estimated number of tuples. We round up the result to the
512 : * total number of buckets which has to be allocated before using its
513 : * hashm_spares element. However always force at least 2 bucket pages. The
514 : * upper limit is determined by considerations explained in
515 : * _hash_expandtable().
516 : */
517 240 : dnumbuckets = num_tuples / ffactor;
518 240 : if (dnumbuckets <= 2.0)
519 71 : num_buckets = 2;
520 169 : else if (dnumbuckets >= (double) 0x40000000)
521 0 : num_buckets = 0x40000000;
522 : else
523 169 : num_buckets = _hash_get_totalbuckets(_hash_spareindex(dnumbuckets));
524 :
525 240 : spare_index = _hash_spareindex(num_buckets);
526 : Assert(spare_index < HASH_MAX_SPLITPOINTS);
527 :
528 240 : page = BufferGetPage(buf);
529 240 : if (initpage)
530 29 : _hash_pageinit(page, BufferGetPageSize(buf));
531 :
532 240 : pageopaque = HashPageGetOpaque(page);
533 240 : pageopaque->hasho_prevblkno = InvalidBlockNumber;
534 240 : pageopaque->hasho_nextblkno = InvalidBlockNumber;
535 240 : pageopaque->hasho_bucket = InvalidBucket;
536 240 : pageopaque->hasho_flag = LH_META_PAGE;
537 240 : pageopaque->hasho_page_id = HASHO_PAGE_ID;
538 :
539 240 : metap = HashPageGetMeta(page);
540 :
541 240 : metap->hashm_magic = HASH_MAGIC;
542 240 : metap->hashm_version = HASH_VERSION;
543 240 : metap->hashm_ntuples = 0;
544 240 : metap->hashm_nmaps = 0;
545 240 : metap->hashm_ffactor = ffactor;
546 240 : metap->hashm_bsize = HashGetMaxBitmapSize(page);
547 :
548 : /* find largest bitmap array size that will fit in page size */
549 240 : lshift = pg_leftmost_one_pos32(metap->hashm_bsize);
550 : Assert(lshift > 0);
551 240 : metap->hashm_bmsize = 1 << lshift;
552 240 : metap->hashm_bmshift = lshift + BYTE_TO_BIT;
553 : Assert((1 << BMPG_SHIFT(metap)) == (BMPG_MASK(metap) + 1));
554 :
555 : /*
556 : * Label the index with its primary hash support function's OID. This is
557 : * pretty useless for normal operation (in fact, hashm_procid is not used
558 : * anywhere), but it might be handy for forensic purposes so we keep it.
559 : */
560 240 : metap->hashm_procid = procid;
561 :
562 : /*
563 : * We initialize the index with N buckets, 0 .. N-1, occupying physical
564 : * blocks 1 to N. The first freespace bitmap page is in block N+1.
565 : */
566 240 : metap->hashm_maxbucket = num_buckets - 1;
567 :
568 : /*
569 : * Set highmask as next immediate ((2 ^ x) - 1), which should be
570 : * sufficient to cover num_buckets.
571 : */
572 240 : metap->hashm_highmask = pg_nextpower2_32(num_buckets + 1) - 1;
573 240 : metap->hashm_lowmask = (metap->hashm_highmask >> 1);
574 :
575 240 : MemSet(metap->hashm_spares, 0, sizeof(metap->hashm_spares));
576 240 : MemSet(metap->hashm_mapp, 0, sizeof(metap->hashm_mapp));
577 :
578 : /* Set up mapping for one spare page after the initial splitpoints */
579 240 : metap->hashm_spares[spare_index] = 1;
580 240 : metap->hashm_ovflpoint = spare_index;
581 240 : metap->hashm_firstfree = 0;
582 :
583 : /*
584 : * Set pd_lower just past the end of the metadata. This is essential,
585 : * because without doing so, metadata will be lost if xlog.c compresses
586 : * the page.
587 : */
588 240 : ((PageHeader) page)->pd_lower =
589 240 : ((char *) metap + sizeof(HashMetaPageData)) - (char *) page;
590 240 : }
591 :
592 : /*
593 : * _hash_pageinit() -- Initialize a new hash index page.
594 : */
595 : void
596 7384 : _hash_pageinit(Page page, Size size)
597 : {
598 7384 : PageInit(page, size, sizeof(HashPageOpaqueData));
599 7384 : }
600 :
601 : /*
602 : * Attempt to expand the hash table by creating one new bucket.
603 : *
604 : * This will silently do nothing if we don't get cleanup lock on old or
605 : * new bucket.
606 : *
607 : * Complete the pending splits and remove the tuples from old bucket,
608 : * if there are any left over from the previous split.
609 : *
610 : * The caller must hold a pin, but no lock, on the metapage buffer.
611 : * The buffer is returned in the same state.
612 : */
613 : void
614 890 : _hash_expandtable(Relation rel, Buffer metabuf)
615 : {
616 : HashMetaPage metap;
617 : Bucket old_bucket;
618 : Bucket new_bucket;
619 : uint32 spare_ndx;
620 : BlockNumber start_oblkno;
621 : BlockNumber start_nblkno;
622 : Buffer buf_nblkno;
623 : Buffer buf_oblkno;
624 : Page opage;
625 : Page npage;
626 : HashPageOpaque oopaque;
627 : HashPageOpaque nopaque;
628 : uint32 maxbucket;
629 : uint32 highmask;
630 : uint32 lowmask;
631 890 : bool metap_update_masks = false;
632 890 : bool metap_update_splitpoint = false;
633 : XLogRecPtr recptr;
634 :
635 0 : restart_expand:
636 :
637 : /*
638 : * Write-lock the meta page. It used to be necessary to acquire a
639 : * heavyweight lock to begin a split, but that is no longer required.
640 : */
641 890 : LockBuffer(metabuf, BUFFER_LOCK_EXCLUSIVE);
642 :
643 890 : _hash_checkpage(rel, metabuf, LH_META_PAGE);
644 890 : metap = HashPageGetMeta(BufferGetPage(metabuf));
645 :
646 : /*
647 : * Check to see if split is still needed; someone else might have already
648 : * done one while we waited for the lock.
649 : *
650 : * Make sure this stays in sync with _hash_doinsert()
651 : */
652 890 : if (metap->hashm_ntuples <=
653 890 : (double) metap->hashm_ffactor * (metap->hashm_maxbucket + 1))
654 0 : goto fail;
655 :
656 : /*
657 : * Can't split anymore if maxbucket has reached its maximum possible
658 : * value.
659 : *
660 : * Ideally we'd allow bucket numbers up to UINT_MAX-1 (no higher because
661 : * the calculation maxbucket+1 mustn't overflow). Currently we restrict
662 : * to half that to prevent failure of pg_ceil_log2_32() and insufficient
663 : * space in hashm_spares[]. It's moot anyway because an index with 2^32
664 : * buckets would certainly overflow BlockNumber and hence
665 : * _hash_alloc_buckets() would fail, but if we supported buckets smaller
666 : * than a disk block then this would be an independent constraint.
667 : *
668 : * If you change this, see also the maximum initial number of buckets in
669 : * _hash_init().
670 : */
671 890 : if (metap->hashm_maxbucket >= (uint32) 0x7FFFFFFE)
672 0 : goto fail;
673 :
674 : /*
675 : * Determine which bucket is to be split, and attempt to take cleanup lock
676 : * on the old bucket. If we can't get the lock, give up.
677 : *
678 : * The cleanup lock protects us not only against other backends, but
679 : * against our own backend as well.
680 : *
681 : * The cleanup lock is mainly to protect the split from concurrent
682 : * inserts. See src/backend/access/hash/README, Lock Definitions for
683 : * further details. Due to this locking restriction, if there is any
684 : * pending scan, the split will give up which is not good, but harmless.
685 : */
686 890 : new_bucket = metap->hashm_maxbucket + 1;
687 :
688 890 : old_bucket = (new_bucket & metap->hashm_lowmask);
689 :
690 890 : start_oblkno = BUCKET_TO_BLKNO(metap, old_bucket);
691 :
692 890 : buf_oblkno = _hash_getbuf_with_condlock_cleanup(rel, start_oblkno, LH_BUCKET_PAGE);
693 890 : if (!buf_oblkno)
694 0 : goto fail;
695 :
696 890 : opage = BufferGetPage(buf_oblkno);
697 890 : oopaque = HashPageGetOpaque(opage);
698 :
699 : /*
700 : * We want to finish the split from a bucket as there is no apparent
701 : * benefit by not doing so and it will make the code complicated to finish
702 : * the split that involves multiple buckets considering the case where new
703 : * split also fails. We don't need to consider the new bucket for
704 : * completing the split here as it is not possible that a re-split of new
705 : * bucket starts when there is still a pending split from old bucket.
706 : */
707 890 : if (H_BUCKET_BEING_SPLIT(oopaque))
708 : {
709 : /*
710 : * Copy bucket mapping info now; refer the comment in code below where
711 : * we copy this information before calling _hash_splitbucket to see
712 : * why this is okay.
713 : */
714 0 : maxbucket = metap->hashm_maxbucket;
715 0 : highmask = metap->hashm_highmask;
716 0 : lowmask = metap->hashm_lowmask;
717 :
718 : /*
719 : * Release the lock on metapage and old_bucket, before completing the
720 : * split.
721 : */
722 0 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
723 0 : LockBuffer(buf_oblkno, BUFFER_LOCK_UNLOCK);
724 :
725 0 : _hash_finish_split(rel, metabuf, buf_oblkno, old_bucket, maxbucket,
726 : highmask, lowmask);
727 :
728 : /* release the pin on old buffer and retry for expand. */
729 0 : _hash_dropbuf(rel, buf_oblkno);
730 :
731 0 : goto restart_expand;
732 : }
733 :
734 : /*
735 : * Clean the tuples remained from the previous split. This operation
736 : * requires cleanup lock and we already have one on the old bucket, so
737 : * let's do it. We also don't want to allow further splits from the bucket
738 : * till the garbage of previous split is cleaned. This has two
739 : * advantages; first, it helps in avoiding the bloat due to garbage and
740 : * second is, during cleanup of bucket, we are always sure that the
741 : * garbage tuples belong to most recently split bucket. On the contrary,
742 : * if we allow cleanup of bucket after meta page is updated to indicate
743 : * the new split and before the actual split, the cleanup operation won't
744 : * be able to decide whether the tuple has been moved to the newly created
745 : * bucket and ended up deleting such tuples.
746 : */
747 890 : if (H_NEEDS_SPLIT_CLEANUP(oopaque))
748 : {
749 : /*
750 : * Copy bucket mapping info now; refer to the comment in code below
751 : * where we copy this information before calling _hash_splitbucket to
752 : * see why this is okay.
753 : */
754 0 : maxbucket = metap->hashm_maxbucket;
755 0 : highmask = metap->hashm_highmask;
756 0 : lowmask = metap->hashm_lowmask;
757 :
758 : /* Release the metapage lock. */
759 0 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
760 :
761 0 : hashbucketcleanup(rel, old_bucket, buf_oblkno, start_oblkno, NULL,
762 : maxbucket, highmask, lowmask, NULL, NULL, true,
763 : NULL, NULL);
764 :
765 0 : _hash_dropbuf(rel, buf_oblkno);
766 :
767 0 : goto restart_expand;
768 : }
769 :
770 : /*
771 : * There shouldn't be any active scan on new bucket.
772 : *
773 : * Note: it is safe to compute the new bucket's blkno here, even though we
774 : * may still need to update the BUCKET_TO_BLKNO mapping. This is because
775 : * the current value of hashm_spares[hashm_ovflpoint] correctly shows
776 : * where we are going to put a new splitpoint's worth of buckets.
777 : */
778 890 : start_nblkno = BUCKET_TO_BLKNO(metap, new_bucket);
779 :
780 : /*
781 : * If the split point is increasing we need to allocate a new batch of
782 : * bucket pages.
783 : */
784 890 : spare_ndx = _hash_spareindex(new_bucket + 1);
785 890 : if (spare_ndx > metap->hashm_ovflpoint)
786 : {
787 : uint32 buckets_to_add;
788 :
789 : Assert(spare_ndx == metap->hashm_ovflpoint + 1);
790 :
791 : /*
792 : * We treat allocation of buckets as a separate WAL-logged action.
793 : * Even if we fail after this operation, won't leak bucket pages;
794 : * rather, the next split will consume this space. In any case, even
795 : * without failure we don't use all the space in one split operation.
796 : */
797 38 : buckets_to_add = _hash_get_totalbuckets(spare_ndx) - new_bucket;
798 38 : if (!_hash_alloc_buckets(rel, start_nblkno, buckets_to_add))
799 : {
800 : /* can't split due to BlockNumber overflow */
801 0 : _hash_relbuf(rel, buf_oblkno);
802 0 : goto fail;
803 : }
804 : }
805 :
806 : /*
807 : * Physically allocate the new bucket's primary page. We want to do this
808 : * before changing the metapage's mapping info, in case we can't get the
809 : * disk space.
810 : *
811 : * XXX It doesn't make sense to call _hash_getnewbuf first, zeroing the
812 : * buffer, and then only afterwards check whether we have a cleanup lock.
813 : * However, since no scan can be accessing the buffer yet, any concurrent
814 : * accesses will just be from processes like the bgwriter or checkpointer
815 : * which don't care about its contents, so it doesn't really matter.
816 : */
817 890 : buf_nblkno = _hash_getnewbuf(rel, start_nblkno, MAIN_FORKNUM);
818 890 : if (!IsBufferCleanupOK(buf_nblkno))
819 : {
820 0 : _hash_relbuf(rel, buf_oblkno);
821 0 : _hash_relbuf(rel, buf_nblkno);
822 0 : goto fail;
823 : }
824 :
825 : /*
826 : * Since we are scribbling on the pages in the shared buffers, establish a
827 : * critical section. Any failure in this next code leaves us with a big
828 : * problem: the metapage is effectively corrupt but could get written back
829 : * to disk.
830 : */
831 890 : START_CRIT_SECTION();
832 :
833 : /*
834 : * Okay to proceed with split. Update the metapage bucket mapping info.
835 : */
836 890 : metap->hashm_maxbucket = new_bucket;
837 :
838 890 : if (new_bucket > metap->hashm_highmask)
839 : {
840 : /* Starting a new doubling */
841 14 : metap->hashm_lowmask = metap->hashm_highmask;
842 14 : metap->hashm_highmask = new_bucket | metap->hashm_lowmask;
843 14 : metap_update_masks = true;
844 : }
845 :
846 : /*
847 : * If the split point is increasing we need to adjust the hashm_spares[]
848 : * array and hashm_ovflpoint so that future overflow pages will be created
849 : * beyond this new batch of bucket pages.
850 : */
851 890 : if (spare_ndx > metap->hashm_ovflpoint)
852 : {
853 38 : metap->hashm_spares[spare_ndx] = metap->hashm_spares[metap->hashm_ovflpoint];
854 38 : metap->hashm_ovflpoint = spare_ndx;
855 38 : metap_update_splitpoint = true;
856 : }
857 :
858 890 : MarkBufferDirty(metabuf);
859 :
860 : /*
861 : * Copy bucket mapping info now; this saves re-accessing the meta page
862 : * inside _hash_splitbucket's inner loop. Note that once we drop the
863 : * split lock, other splits could begin, so these values might be out of
864 : * date before _hash_splitbucket finishes. That's okay, since all it
865 : * needs is to tell which of these two buckets to map hashkeys into.
866 : */
867 890 : maxbucket = metap->hashm_maxbucket;
868 890 : highmask = metap->hashm_highmask;
869 890 : lowmask = metap->hashm_lowmask;
870 :
871 890 : opage = BufferGetPage(buf_oblkno);
872 890 : oopaque = HashPageGetOpaque(opage);
873 :
874 : /*
875 : * Mark the old bucket to indicate that split is in progress. (At
876 : * operation end, we will clear the split-in-progress flag.) Also, for a
877 : * primary bucket page, hasho_prevblkno stores the number of buckets that
878 : * existed as of the last split, so we must update that value here.
879 : */
880 890 : oopaque->hasho_flag |= LH_BUCKET_BEING_SPLIT;
881 890 : oopaque->hasho_prevblkno = maxbucket;
882 :
883 890 : MarkBufferDirty(buf_oblkno);
884 :
885 890 : npage = BufferGetPage(buf_nblkno);
886 :
887 : /*
888 : * initialize the new bucket's primary page and mark it to indicate that
889 : * split is in progress.
890 : */
891 890 : nopaque = HashPageGetOpaque(npage);
892 890 : nopaque->hasho_prevblkno = maxbucket;
893 890 : nopaque->hasho_nextblkno = InvalidBlockNumber;
894 890 : nopaque->hasho_bucket = new_bucket;
895 890 : nopaque->hasho_flag = LH_BUCKET_PAGE | LH_BUCKET_BEING_POPULATED;
896 890 : nopaque->hasho_page_id = HASHO_PAGE_ID;
897 :
898 890 : MarkBufferDirty(buf_nblkno);
899 :
900 : /* XLOG stuff */
901 890 : if (RelationNeedsWAL(rel))
902 764 : {
903 : xl_hash_split_allocate_page xlrec;
904 :
905 764 : xlrec.new_bucket = maxbucket;
906 764 : xlrec.old_bucket_flag = oopaque->hasho_flag;
907 764 : xlrec.new_bucket_flag = nopaque->hasho_flag;
908 764 : xlrec.flags = 0;
909 :
910 764 : XLogBeginInsert();
911 :
912 764 : XLogRegisterBuffer(0, buf_oblkno, REGBUF_STANDARD);
913 764 : XLogRegisterBuffer(1, buf_nblkno, REGBUF_WILL_INIT);
914 764 : XLogRegisterBuffer(2, metabuf, REGBUF_STANDARD);
915 :
916 764 : if (metap_update_masks)
917 : {
918 14 : xlrec.flags |= XLH_SPLIT_META_UPDATE_MASKS;
919 14 : XLogRegisterBufData(2, &metap->hashm_lowmask, sizeof(uint32));
920 14 : XLogRegisterBufData(2, &metap->hashm_highmask, sizeof(uint32));
921 : }
922 :
923 764 : if (metap_update_splitpoint)
924 : {
925 34 : xlrec.flags |= XLH_SPLIT_META_UPDATE_SPLITPOINT;
926 34 : XLogRegisterBufData(2, &metap->hashm_ovflpoint,
927 : sizeof(uint32));
928 34 : XLogRegisterBufData(2,
929 34 : &metap->hashm_spares[metap->hashm_ovflpoint],
930 : sizeof(uint32));
931 : }
932 :
933 764 : XLogRegisterData(&xlrec, SizeOfHashSplitAllocPage);
934 :
935 764 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_ALLOCATE_PAGE);
936 : }
937 : else
938 126 : recptr = XLogGetFakeLSN(rel);
939 :
940 890 : PageSetLSN(BufferGetPage(buf_oblkno), recptr);
941 890 : PageSetLSN(BufferGetPage(buf_nblkno), recptr);
942 890 : PageSetLSN(BufferGetPage(metabuf), recptr);
943 :
944 890 : END_CRIT_SECTION();
945 :
946 : /* drop lock, but keep pin */
947 890 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
948 :
949 : /* Relocate records to the new bucket */
950 890 : _hash_splitbucket(rel, metabuf,
951 : old_bucket, new_bucket,
952 : buf_oblkno, buf_nblkno, NULL,
953 : maxbucket, highmask, lowmask);
954 :
955 : /* all done, now release the pins on primary buckets. */
956 890 : _hash_dropbuf(rel, buf_oblkno);
957 890 : _hash_dropbuf(rel, buf_nblkno);
958 :
959 890 : return;
960 :
961 : /* Here if decide not to split or fail to acquire old bucket lock */
962 0 : fail:
963 :
964 : /* We didn't write the metapage, so just drop lock */
965 0 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
966 : }
967 :
968 :
969 : /*
970 : * _hash_alloc_buckets -- allocate a new splitpoint's worth of bucket pages
971 : *
972 : * This does not need to initialize the new bucket pages; we'll do that as
973 : * each one is used by _hash_expandtable(). But we have to extend the logical
974 : * EOF to the end of the splitpoint; this keeps smgr's idea of the EOF in
975 : * sync with ours, so that we don't get complaints from smgr.
976 : *
977 : * We do this by writing a page of zeroes at the end of the splitpoint range.
978 : * We expect that the filesystem will ensure that the intervening pages read
979 : * as zeroes too. On many filesystems this "hole" will not be allocated
980 : * immediately, which means that the index file may end up more fragmented
981 : * than if we forced it all to be allocated now; but since we don't scan
982 : * hash indexes sequentially anyway, that probably doesn't matter.
983 : *
984 : * XXX It's annoying that this code is executed with the metapage lock held.
985 : * We need to interlock against _hash_addovflpage() adding a new overflow page
986 : * concurrently, but it'd likely be better to use LockRelationForExtension
987 : * for the purpose. OTOH, adding a splitpoint is a very infrequent operation,
988 : * so it may not be worth worrying about.
989 : *
990 : * Returns true if successful, or false if allocation failed due to
991 : * BlockNumber overflow.
992 : */
993 : static bool
994 38 : _hash_alloc_buckets(Relation rel, BlockNumber firstblock, uint32 nblocks)
995 : {
996 : BlockNumber lastblock;
997 : PGIOAlignedBlock zerobuf;
998 : Page page;
999 : HashPageOpaque ovflopaque;
1000 :
1001 38 : lastblock = firstblock + nblocks - 1;
1002 :
1003 : /*
1004 : * Check for overflow in block number calculation; if so, we cannot extend
1005 : * the index anymore.
1006 : */
1007 38 : if (lastblock < firstblock || lastblock == InvalidBlockNumber)
1008 0 : return false;
1009 :
1010 38 : page = (Page) zerobuf.data;
1011 :
1012 : /*
1013 : * Initialize the page. Just zeroing the page won't work; see
1014 : * _hash_freeovflpage for similar usage. We take care to make the special
1015 : * space valid for the benefit of tools such as pageinspect.
1016 : */
1017 38 : _hash_pageinit(page, BLCKSZ);
1018 :
1019 38 : ovflopaque = HashPageGetOpaque(page);
1020 :
1021 38 : ovflopaque->hasho_prevblkno = InvalidBlockNumber;
1022 38 : ovflopaque->hasho_nextblkno = InvalidBlockNumber;
1023 38 : ovflopaque->hasho_bucket = InvalidBucket;
1024 38 : ovflopaque->hasho_flag = LH_UNUSED_PAGE;
1025 38 : ovflopaque->hasho_page_id = HASHO_PAGE_ID;
1026 :
1027 38 : if (RelationNeedsWAL(rel))
1028 34 : log_newpage(&rel->rd_locator,
1029 : MAIN_FORKNUM,
1030 : lastblock,
1031 : zerobuf.data,
1032 : true);
1033 :
1034 38 : PageSetChecksumInplace(page, lastblock);
1035 38 : smgrextend(RelationGetSmgr(rel), MAIN_FORKNUM, lastblock, zerobuf.data,
1036 : false);
1037 :
1038 38 : return true;
1039 : }
1040 :
1041 :
1042 : /*
1043 : * _hash_splitbucket -- split 'obucket' into 'obucket' and 'nbucket'
1044 : *
1045 : * This routine is used to partition the tuples between old and new bucket and
1046 : * is used to finish the incomplete split operations. To finish the previously
1047 : * interrupted split operation, the caller needs to fill htab. If htab is set,
1048 : * then we skip the movement of tuples that exists in htab, otherwise NULL
1049 : * value of htab indicates movement of all the tuples that belong to the new
1050 : * bucket.
1051 : *
1052 : * We are splitting a bucket that consists of a base bucket page and zero
1053 : * or more overflow (bucket chain) pages. We must relocate tuples that
1054 : * belong in the new bucket.
1055 : *
1056 : * The caller must hold cleanup locks on both buckets to ensure that
1057 : * no one else is trying to access them (see README).
1058 : *
1059 : * The caller must hold a pin, but no lock, on the metapage buffer.
1060 : * The buffer is returned in the same state. (The metapage is only
1061 : * touched if it becomes necessary to add or remove overflow pages.)
1062 : *
1063 : * Split needs to retain pin on primary bucket pages of both old and new
1064 : * buckets till end of operation. This is to prevent vacuum from starting
1065 : * while a split is in progress.
1066 : *
1067 : * In addition, the caller must have created the new bucket's base page,
1068 : * which is passed in buffer nbuf, pinned and write-locked. The lock will be
1069 : * released here and pin must be released by the caller. (The API is set up
1070 : * this way because we must do _hash_getnewbuf() before releasing the metapage
1071 : * write lock. So instead of passing the new bucket's start block number, we
1072 : * pass an actual buffer.)
1073 : */
1074 : static void
1075 890 : _hash_splitbucket(Relation rel,
1076 : Buffer metabuf,
1077 : Bucket obucket,
1078 : Bucket nbucket,
1079 : Buffer obuf,
1080 : Buffer nbuf,
1081 : HTAB *htab,
1082 : uint32 maxbucket,
1083 : uint32 highmask,
1084 : uint32 lowmask)
1085 : {
1086 : Buffer bucket_obuf;
1087 : Buffer bucket_nbuf;
1088 : Page opage;
1089 : Page npage;
1090 : HashPageOpaque oopaque;
1091 : HashPageOpaque nopaque;
1092 : OffsetNumber itup_offsets[MaxIndexTuplesPerPage];
1093 : IndexTuple itups[MaxIndexTuplesPerPage];
1094 890 : Size all_tups_size = 0;
1095 : int i;
1096 890 : uint16 nitups = 0;
1097 : XLogRecPtr recptr;
1098 :
1099 890 : bucket_obuf = obuf;
1100 890 : opage = BufferGetPage(obuf);
1101 890 : oopaque = HashPageGetOpaque(opage);
1102 :
1103 890 : bucket_nbuf = nbuf;
1104 890 : npage = BufferGetPage(nbuf);
1105 890 : nopaque = HashPageGetOpaque(npage);
1106 :
1107 : /* Copy the predicate locks from old bucket to new bucket. */
1108 890 : PredicateLockPageSplit(rel,
1109 : BufferGetBlockNumber(bucket_obuf),
1110 : BufferGetBlockNumber(bucket_nbuf));
1111 :
1112 : /*
1113 : * Partition the tuples in the old bucket between the old bucket and the
1114 : * new bucket, advancing along the old bucket's overflow bucket chain and
1115 : * adding overflow pages to the new bucket as needed. Outer loop iterates
1116 : * once per page in old bucket.
1117 : */
1118 : for (;;)
1119 226 : {
1120 : BlockNumber oblkno;
1121 : OffsetNumber ooffnum;
1122 : OffsetNumber omaxoffnum;
1123 :
1124 : /* Scan each tuple in old page */
1125 1116 : omaxoffnum = PageGetMaxOffsetNumber(opage);
1126 1116 : for (ooffnum = FirstOffsetNumber;
1127 205183 : ooffnum <= omaxoffnum;
1128 204067 : ooffnum = OffsetNumberNext(ooffnum))
1129 : {
1130 : IndexTuple itup;
1131 : Size itemsz;
1132 : Bucket bucket;
1133 204067 : bool found = false;
1134 :
1135 : /* skip dead tuples */
1136 204067 : if (ItemIdIsDead(PageGetItemId(opage, ooffnum)))
1137 0 : continue;
1138 :
1139 : /*
1140 : * Before inserting a tuple, probe the hash table containing TIDs
1141 : * of tuples belonging to new bucket, if we find a match, then
1142 : * skip that tuple, else fetch the item's hash key (conveniently
1143 : * stored in the item) and determine which bucket it now belongs
1144 : * in.
1145 : */
1146 204067 : itup = (IndexTuple) PageGetItem(opage,
1147 204067 : PageGetItemId(opage, ooffnum));
1148 :
1149 204067 : if (htab)
1150 0 : (void) hash_search(htab, &itup->t_tid, HASH_FIND, &found);
1151 :
1152 204067 : if (found)
1153 0 : continue;
1154 :
1155 204067 : bucket = _hash_hashkey2bucket(_hash_get_indextuple_hashkey(itup),
1156 : maxbucket, highmask, lowmask);
1157 :
1158 204067 : if (bucket == nbucket)
1159 : {
1160 : IndexTuple new_itup;
1161 :
1162 : /*
1163 : * make a copy of index tuple as we have to scribble on it.
1164 : */
1165 83591 : new_itup = CopyIndexTuple(itup);
1166 :
1167 : /*
1168 : * mark the index tuple as moved by split, such tuples are
1169 : * skipped by scan if there is split in progress for a bucket.
1170 : */
1171 83591 : new_itup->t_info |= INDEX_MOVED_BY_SPLIT_MASK;
1172 :
1173 : /*
1174 : * insert the tuple into the new bucket. if it doesn't fit on
1175 : * the current page in the new bucket, we must allocate a new
1176 : * overflow page and place the tuple on that page instead.
1177 : */
1178 83591 : itemsz = IndexTupleSize(new_itup);
1179 83591 : itemsz = MAXALIGN(itemsz);
1180 :
1181 83591 : if (PageGetFreeSpaceForMultipleTuples(npage, nitups + 1) < (all_tups_size + itemsz))
1182 : {
1183 : /*
1184 : * Change the shared buffer state in critical section,
1185 : * otherwise any error could make it unrecoverable.
1186 : */
1187 52 : START_CRIT_SECTION();
1188 :
1189 52 : _hash_pgaddmultitup(rel, nbuf, itups, itup_offsets, nitups);
1190 52 : MarkBufferDirty(nbuf);
1191 : /* log the split operation before releasing the lock */
1192 52 : log_split_page(rel, nbuf);
1193 :
1194 52 : END_CRIT_SECTION();
1195 :
1196 : /* drop lock, but keep pin */
1197 52 : LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
1198 :
1199 : /* be tidy */
1200 21216 : for (i = 0; i < nitups; i++)
1201 21164 : pfree(itups[i]);
1202 52 : nitups = 0;
1203 52 : all_tups_size = 0;
1204 :
1205 : /* chain to a new overflow page */
1206 52 : nbuf = _hash_addovflpage(rel, metabuf, nbuf, (nbuf == bucket_nbuf));
1207 52 : npage = BufferGetPage(nbuf);
1208 52 : nopaque = HashPageGetOpaque(npage);
1209 : }
1210 :
1211 83591 : itups[nitups++] = new_itup;
1212 83591 : all_tups_size += itemsz;
1213 : }
1214 : else
1215 : {
1216 : /*
1217 : * the tuple stays on this page, so nothing to do.
1218 : */
1219 : Assert(bucket == obucket);
1220 : }
1221 : }
1222 :
1223 1116 : oblkno = oopaque->hasho_nextblkno;
1224 :
1225 : /* retain the pin on the old primary bucket */
1226 1116 : if (obuf == bucket_obuf)
1227 890 : LockBuffer(obuf, BUFFER_LOCK_UNLOCK);
1228 : else
1229 226 : _hash_relbuf(rel, obuf);
1230 :
1231 : /* Exit loop if no more overflow pages in old bucket */
1232 1116 : if (!BlockNumberIsValid(oblkno))
1233 : {
1234 : /*
1235 : * Change the shared buffer state in critical section, otherwise
1236 : * any error could make it unrecoverable.
1237 : */
1238 890 : START_CRIT_SECTION();
1239 :
1240 890 : _hash_pgaddmultitup(rel, nbuf, itups, itup_offsets, nitups);
1241 890 : MarkBufferDirty(nbuf);
1242 : /* log the split operation before releasing the lock */
1243 890 : log_split_page(rel, nbuf);
1244 :
1245 890 : END_CRIT_SECTION();
1246 :
1247 890 : if (nbuf == bucket_nbuf)
1248 886 : LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
1249 : else
1250 4 : _hash_relbuf(rel, nbuf);
1251 :
1252 : /* be tidy */
1253 63317 : for (i = 0; i < nitups; i++)
1254 62427 : pfree(itups[i]);
1255 890 : break;
1256 : }
1257 :
1258 : /* Else, advance to next old page */
1259 226 : obuf = _hash_getbuf(rel, oblkno, HASH_READ, LH_OVERFLOW_PAGE);
1260 226 : opage = BufferGetPage(obuf);
1261 226 : oopaque = HashPageGetOpaque(opage);
1262 : }
1263 :
1264 : /*
1265 : * We're at the end of the old bucket chain, so we're done partitioning
1266 : * the tuples. Mark the old and new buckets to indicate split is
1267 : * finished.
1268 : *
1269 : * To avoid deadlocks due to locking order of buckets, first lock the old
1270 : * bucket and then the new bucket.
1271 : */
1272 890 : LockBuffer(bucket_obuf, BUFFER_LOCK_EXCLUSIVE);
1273 890 : opage = BufferGetPage(bucket_obuf);
1274 890 : oopaque = HashPageGetOpaque(opage);
1275 :
1276 890 : LockBuffer(bucket_nbuf, BUFFER_LOCK_EXCLUSIVE);
1277 890 : npage = BufferGetPage(bucket_nbuf);
1278 890 : nopaque = HashPageGetOpaque(npage);
1279 :
1280 890 : START_CRIT_SECTION();
1281 :
1282 890 : oopaque->hasho_flag &= ~LH_BUCKET_BEING_SPLIT;
1283 890 : nopaque->hasho_flag &= ~LH_BUCKET_BEING_POPULATED;
1284 :
1285 : /*
1286 : * After the split is finished, mark the old bucket to indicate that it
1287 : * contains deletable tuples. We will clear split-cleanup flag after
1288 : * deleting such tuples either at the end of split or at the next split
1289 : * from old bucket or at the time of vacuum.
1290 : */
1291 890 : oopaque->hasho_flag |= LH_BUCKET_NEEDS_SPLIT_CLEANUP;
1292 :
1293 : /*
1294 : * now write the buffers, here we don't release the locks as caller is
1295 : * responsible to release locks.
1296 : */
1297 890 : MarkBufferDirty(bucket_obuf);
1298 890 : MarkBufferDirty(bucket_nbuf);
1299 :
1300 890 : if (RelationNeedsWAL(rel))
1301 764 : {
1302 : xl_hash_split_complete xlrec;
1303 :
1304 764 : xlrec.old_bucket_flag = oopaque->hasho_flag;
1305 764 : xlrec.new_bucket_flag = nopaque->hasho_flag;
1306 :
1307 764 : XLogBeginInsert();
1308 :
1309 764 : XLogRegisterData(&xlrec, SizeOfHashSplitComplete);
1310 :
1311 764 : XLogRegisterBuffer(0, bucket_obuf, REGBUF_STANDARD);
1312 764 : XLogRegisterBuffer(1, bucket_nbuf, REGBUF_STANDARD);
1313 :
1314 764 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_COMPLETE);
1315 : }
1316 : else
1317 126 : recptr = XLogGetFakeLSN(rel);
1318 :
1319 890 : PageSetLSN(BufferGetPage(bucket_obuf), recptr);
1320 890 : PageSetLSN(BufferGetPage(bucket_nbuf), recptr);
1321 :
1322 890 : END_CRIT_SECTION();
1323 :
1324 : /*
1325 : * If possible, clean up the old bucket. We might not be able to do this
1326 : * if someone else has a pin on it, but if not then we can go ahead. This
1327 : * isn't absolutely necessary, but it reduces bloat; if we don't do it
1328 : * now, VACUUM will do it eventually, but maybe not until new overflow
1329 : * pages have been allocated. Note that there's no need to clean up the
1330 : * new bucket.
1331 : */
1332 890 : if (IsBufferCleanupOK(bucket_obuf))
1333 : {
1334 890 : LockBuffer(bucket_nbuf, BUFFER_LOCK_UNLOCK);
1335 890 : hashbucketcleanup(rel, obucket, bucket_obuf,
1336 : BufferGetBlockNumber(bucket_obuf), NULL,
1337 : maxbucket, highmask, lowmask, NULL, NULL, true,
1338 : NULL, NULL);
1339 : }
1340 : else
1341 : {
1342 0 : LockBuffer(bucket_nbuf, BUFFER_LOCK_UNLOCK);
1343 0 : LockBuffer(bucket_obuf, BUFFER_LOCK_UNLOCK);
1344 : }
1345 890 : }
1346 :
1347 : /*
1348 : * _hash_finish_split() -- Finish the previously interrupted split operation
1349 : *
1350 : * To complete the split operation, we form the hash table of TIDs in new
1351 : * bucket which is then used by split operation to skip tuples that are
1352 : * already moved before the split operation was previously interrupted.
1353 : *
1354 : * The caller must hold a pin, but no lock, on the metapage and old bucket's
1355 : * primary page buffer. The buffers are returned in the same state. (The
1356 : * metapage is only touched if it becomes necessary to add or remove overflow
1357 : * pages.)
1358 : */
1359 : void
1360 0 : _hash_finish_split(Relation rel, Buffer metabuf, Buffer obuf, Bucket obucket,
1361 : uint32 maxbucket, uint32 highmask, uint32 lowmask)
1362 : {
1363 : HASHCTL hash_ctl;
1364 : HTAB *tidhtab;
1365 0 : Buffer bucket_nbuf = InvalidBuffer;
1366 : Buffer nbuf;
1367 : Page npage;
1368 : BlockNumber nblkno;
1369 : BlockNumber bucket_nblkno;
1370 : HashPageOpaque npageopaque;
1371 : Bucket nbucket;
1372 : bool found;
1373 :
1374 : /* Initialize hash tables used to track TIDs */
1375 0 : hash_ctl.keysize = sizeof(ItemPointerData);
1376 0 : hash_ctl.entrysize = sizeof(ItemPointerData);
1377 0 : hash_ctl.hcxt = CurrentMemoryContext;
1378 :
1379 : tidhtab =
1380 0 : hash_create("bucket ctids",
1381 : 256, /* arbitrary initial size */
1382 : &hash_ctl,
1383 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1384 :
1385 0 : bucket_nblkno = nblkno = _hash_get_newblock_from_oldbucket(rel, obucket);
1386 :
1387 : /*
1388 : * Scan the new bucket and build hash table of TIDs
1389 : */
1390 : for (;;)
1391 0 : {
1392 : OffsetNumber noffnum;
1393 : OffsetNumber nmaxoffnum;
1394 :
1395 0 : nbuf = _hash_getbuf(rel, nblkno, HASH_READ,
1396 : LH_BUCKET_PAGE | LH_OVERFLOW_PAGE);
1397 :
1398 : /* remember the primary bucket buffer to acquire cleanup lock on it. */
1399 0 : if (nblkno == bucket_nblkno)
1400 0 : bucket_nbuf = nbuf;
1401 :
1402 0 : npage = BufferGetPage(nbuf);
1403 0 : npageopaque = HashPageGetOpaque(npage);
1404 :
1405 : /* Scan each tuple in new page */
1406 0 : nmaxoffnum = PageGetMaxOffsetNumber(npage);
1407 0 : for (noffnum = FirstOffsetNumber;
1408 0 : noffnum <= nmaxoffnum;
1409 0 : noffnum = OffsetNumberNext(noffnum))
1410 : {
1411 : IndexTuple itup;
1412 :
1413 : /* Fetch the item's TID and insert it in hash table. */
1414 0 : itup = (IndexTuple) PageGetItem(npage,
1415 0 : PageGetItemId(npage, noffnum));
1416 :
1417 0 : (void) hash_search(tidhtab, &itup->t_tid, HASH_ENTER, &found);
1418 :
1419 : Assert(!found);
1420 : }
1421 :
1422 0 : nblkno = npageopaque->hasho_nextblkno;
1423 :
1424 : /*
1425 : * release our write lock without modifying buffer and ensure to
1426 : * retain the pin on primary bucket.
1427 : */
1428 0 : if (nbuf == bucket_nbuf)
1429 0 : LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
1430 : else
1431 0 : _hash_relbuf(rel, nbuf);
1432 :
1433 : /* Exit loop if no more overflow pages in new bucket */
1434 0 : if (!BlockNumberIsValid(nblkno))
1435 0 : break;
1436 : }
1437 :
1438 : /*
1439 : * Conditionally get the cleanup lock on old and new buckets to perform
1440 : * the split operation. If we don't get the cleanup locks, silently give
1441 : * up and next insertion on old bucket will try again to complete the
1442 : * split.
1443 : */
1444 0 : if (!ConditionalLockBufferForCleanup(obuf))
1445 : {
1446 0 : hash_destroy(tidhtab);
1447 0 : return;
1448 : }
1449 0 : if (!ConditionalLockBufferForCleanup(bucket_nbuf))
1450 : {
1451 0 : LockBuffer(obuf, BUFFER_LOCK_UNLOCK);
1452 0 : hash_destroy(tidhtab);
1453 0 : return;
1454 : }
1455 :
1456 0 : npage = BufferGetPage(bucket_nbuf);
1457 0 : npageopaque = HashPageGetOpaque(npage);
1458 0 : nbucket = npageopaque->hasho_bucket;
1459 :
1460 0 : _hash_splitbucket(rel, metabuf, obucket,
1461 : nbucket, obuf, bucket_nbuf, tidhtab,
1462 : maxbucket, highmask, lowmask);
1463 :
1464 0 : _hash_dropbuf(rel, bucket_nbuf);
1465 0 : hash_destroy(tidhtab);
1466 : }
1467 :
1468 : /*
1469 : * log_split_page() -- Log the split operation
1470 : *
1471 : * We log the split operation when the new page in new bucket gets full,
1472 : * so we log the entire page.
1473 : *
1474 : * 'buf' must be locked by the caller which is also responsible for unlocking
1475 : * it.
1476 : */
1477 : static void
1478 942 : log_split_page(Relation rel, Buffer buf)
1479 : {
1480 942 : if (RelationNeedsWAL(rel))
1481 : {
1482 : XLogRecPtr recptr;
1483 :
1484 816 : XLogBeginInsert();
1485 :
1486 816 : XLogRegisterBuffer(0, buf, REGBUF_FORCE_IMAGE | REGBUF_STANDARD);
1487 :
1488 816 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_PAGE);
1489 :
1490 816 : PageSetLSN(BufferGetPage(buf), recptr);
1491 : }
1492 942 : }
1493 :
1494 : /*
1495 : * _hash_getcachedmetap() -- Returns cached metapage data.
1496 : *
1497 : * If metabuf is not InvalidBuffer, caller must hold a pin, but no lock, on
1498 : * the metapage. If not set, we'll set it before returning if we have to
1499 : * refresh the cache, and return with a pin but no lock on it; caller is
1500 : * responsible for releasing the pin.
1501 : *
1502 : * We refresh the cache if it's not initialized yet or force_refresh is true.
1503 : */
1504 : HashMetaPage
1505 491553 : _hash_getcachedmetap(Relation rel, Buffer *metabuf, bool force_refresh)
1506 : {
1507 : Page page;
1508 :
1509 : Assert(metabuf);
1510 491553 : if (force_refresh || rel->rd_amcache == NULL)
1511 : {
1512 731 : char *cache = NULL;
1513 :
1514 : /*
1515 : * It's important that we don't set rd_amcache to an invalid value.
1516 : * Either MemoryContextAlloc or _hash_getbuf could fail, so don't
1517 : * install a pointer to the newly-allocated storage in the actual
1518 : * relcache entry until both have succeeded.
1519 : */
1520 731 : if (rel->rd_amcache == NULL)
1521 320 : cache = MemoryContextAlloc(rel->rd_indexcxt,
1522 : sizeof(HashMetaPageData));
1523 :
1524 : /* Read the metapage. */
1525 731 : if (BufferIsValid(*metabuf))
1526 0 : LockBuffer(*metabuf, BUFFER_LOCK_SHARE);
1527 : else
1528 731 : *metabuf = _hash_getbuf(rel, HASH_METAPAGE, HASH_READ,
1529 : LH_META_PAGE);
1530 731 : page = BufferGetPage(*metabuf);
1531 :
1532 : /* Populate the cache. */
1533 731 : if (rel->rd_amcache == NULL)
1534 320 : rel->rd_amcache = cache;
1535 731 : memcpy(rel->rd_amcache, HashPageGetMeta(page),
1536 : sizeof(HashMetaPageData));
1537 :
1538 : /* Release metapage lock, but keep the pin. */
1539 731 : LockBuffer(*metabuf, BUFFER_LOCK_UNLOCK);
1540 : }
1541 :
1542 491553 : return (HashMetaPage) rel->rd_amcache;
1543 : }
1544 :
1545 : /*
1546 : * _hash_getbucketbuf_from_hashkey() -- Get the bucket's buffer for the given
1547 : * hashkey.
1548 : *
1549 : * Bucket pages do not move or get removed once they are allocated. This give
1550 : * us an opportunity to use the previously saved metapage contents to reach
1551 : * the target bucket buffer, instead of reading from the metapage every time.
1552 : * This saves one buffer access every time we want to reach the target bucket
1553 : * buffer, which is very helpful savings in bufmgr traffic and contention.
1554 : *
1555 : * The access type parameter (HASH_READ or HASH_WRITE) indicates whether the
1556 : * bucket buffer has to be locked for reading or writing.
1557 : *
1558 : * The out parameter cachedmetap is set with metapage contents used for
1559 : * hashkey to bucket buffer mapping. Some callers need this info to reach the
1560 : * old bucket in case of bucket split, see _hash_doinsert().
1561 : */
1562 : Buffer
1563 491114 : _hash_getbucketbuf_from_hashkey(Relation rel, uint32 hashkey, int access,
1564 : HashMetaPage *cachedmetap)
1565 : {
1566 : HashMetaPage metap;
1567 : Buffer buf;
1568 491114 : Buffer metabuf = InvalidBuffer;
1569 : Page page;
1570 : Bucket bucket;
1571 : BlockNumber blkno;
1572 : HashPageOpaque opaque;
1573 :
1574 : /* We read from target bucket buffer, hence locking is must. */
1575 : Assert(access == HASH_READ || access == HASH_WRITE);
1576 :
1577 491114 : metap = _hash_getcachedmetap(rel, &metabuf, false);
1578 : Assert(metap != NULL);
1579 :
1580 : /*
1581 : * Loop until we get a lock on the correct target bucket.
1582 : */
1583 : for (;;)
1584 : {
1585 : /*
1586 : * Compute the target bucket number, and convert to block number.
1587 : */
1588 491525 : bucket = _hash_hashkey2bucket(hashkey,
1589 : metap->hashm_maxbucket,
1590 : metap->hashm_highmask,
1591 : metap->hashm_lowmask);
1592 :
1593 491525 : blkno = BUCKET_TO_BLKNO(metap, bucket);
1594 :
1595 : /* Fetch the primary bucket page for the bucket */
1596 491525 : buf = _hash_getbuf(rel, blkno, access, LH_BUCKET_PAGE);
1597 491525 : page = BufferGetPage(buf);
1598 491525 : opaque = HashPageGetOpaque(page);
1599 : Assert(opaque->hasho_bucket == bucket);
1600 : Assert(opaque->hasho_prevblkno != InvalidBlockNumber);
1601 :
1602 : /*
1603 : * If this bucket hasn't been split, we're done.
1604 : */
1605 491525 : if (opaque->hasho_prevblkno <= metap->hashm_maxbucket)
1606 491114 : break;
1607 :
1608 : /* Drop lock on this buffer, update cached metapage, and retry. */
1609 411 : _hash_relbuf(rel, buf);
1610 411 : metap = _hash_getcachedmetap(rel, &metabuf, true);
1611 : Assert(metap != NULL);
1612 : }
1613 :
1614 491114 : if (BufferIsValid(metabuf))
1615 723 : _hash_dropbuf(rel, metabuf);
1616 :
1617 491114 : if (cachedmetap)
1618 490773 : *cachedmetap = metap;
1619 :
1620 491114 : return buf;
1621 : }
|
