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) 1996-2017, 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 :
37 :
38 : static bool _hash_alloc_buckets(Relation rel, BlockNumber firstblock,
39 : uint32 nblocks);
40 : static void _hash_splitbucket(Relation rel, Buffer metabuf,
41 : Bucket obucket, Bucket nbucket,
42 : Buffer obuf,
43 : Buffer nbuf,
44 : HTAB *htab,
45 : uint32 maxbucket,
46 : uint32 highmask, uint32 lowmask);
47 : static void log_split_page(Relation rel, Buffer buf);
48 :
49 :
50 : /*
51 : * We use high-concurrency locking on hash indexes (see README for an overview
52 : * of the locking rules). However, we can skip taking lmgr locks when the
53 : * index is local to the current backend (ie, either temp or new in the
54 : * current transaction). No one else can see it, so there's no reason to
55 : * take locks. We still take buffer-level locks, but not lmgr locks.
56 : */
57 : #define USELOCKING(rel) (!RELATION_IS_LOCAL(rel))
58 :
59 :
60 : /*
61 : * _hash_getbuf() -- Get a buffer by block number for read or write.
62 : *
63 : * 'access' must be HASH_READ, HASH_WRITE, or HASH_NOLOCK.
64 : * 'flags' is a bitwise OR of the allowed page types.
65 : *
66 : * This must be used only to fetch pages that are expected to be valid
67 : * already. _hash_checkpage() is applied using the given flags.
68 : *
69 : * When this routine returns, the appropriate lock is set on the
70 : * requested buffer and its reference count has been incremented
71 : * (ie, the buffer is "locked and pinned").
72 : *
73 : * P_NEW is disallowed because this routine can only be used
74 : * to access pages that are known to be before the filesystem EOF.
75 : * Extending the index should be done with _hash_getnewbuf.
76 : */
77 : Buffer
78 195435 : _hash_getbuf(Relation rel, BlockNumber blkno, int access, int flags)
79 : {
80 : Buffer buf;
81 :
82 195435 : if (blkno == P_NEW)
83 0 : elog(ERROR, "hash AM does not use P_NEW");
84 :
85 195435 : buf = ReadBuffer(rel, blkno);
86 :
87 195435 : if (access != HASH_NOLOCK)
88 114883 : LockBuffer(buf, access);
89 :
90 : /* ref count and lock type are correct */
91 :
92 195435 : _hash_checkpage(rel, buf, flags);
93 :
94 195435 : return buf;
95 : }
96 :
97 : /*
98 : * _hash_getbuf_with_condlock_cleanup() -- Try to get a buffer for cleanup.
99 : *
100 : * We read the page and try to acquire a cleanup lock. If we get it,
101 : * we return the buffer; otherwise, we return InvalidBuffer.
102 : */
103 : Buffer
104 72 : _hash_getbuf_with_condlock_cleanup(Relation rel, BlockNumber blkno, int flags)
105 : {
106 : Buffer buf;
107 :
108 72 : if (blkno == P_NEW)
109 0 : elog(ERROR, "hash AM does not use P_NEW");
110 :
111 72 : buf = ReadBuffer(rel, blkno);
112 :
113 72 : if (!ConditionalLockBufferForCleanup(buf))
114 : {
115 0 : ReleaseBuffer(buf);
116 0 : return InvalidBuffer;
117 : }
118 :
119 : /* ref count and lock type are correct */
120 :
121 72 : _hash_checkpage(rel, buf, flags);
122 :
123 72 : return buf;
124 : }
125 :
126 : /*
127 : * _hash_getinitbuf() -- Get and initialize a buffer by block number.
128 : *
129 : * This must be used only to fetch pages that are known to be before
130 : * the index's filesystem EOF, but are to be filled from scratch.
131 : * _hash_pageinit() is applied automatically. Otherwise it has
132 : * effects similar to _hash_getbuf() with access = HASH_WRITE.
133 : *
134 : * When this routine returns, a write lock is set on the
135 : * requested buffer and its reference count has been incremented
136 : * (ie, the buffer is "locked and pinned").
137 : *
138 : * P_NEW is disallowed because this routine can only be used
139 : * to access pages that are known to be before the filesystem EOF.
140 : * Extending the index should be done with _hash_getnewbuf.
141 : */
142 : Buffer
143 8 : _hash_getinitbuf(Relation rel, BlockNumber blkno)
144 : {
145 : Buffer buf;
146 :
147 8 : if (blkno == P_NEW)
148 0 : elog(ERROR, "hash AM does not use P_NEW");
149 :
150 8 : buf = ReadBufferExtended(rel, MAIN_FORKNUM, blkno, RBM_ZERO_AND_LOCK,
151 : NULL);
152 :
153 : /* ref count and lock type are correct */
154 :
155 : /* initialize the page */
156 8 : _hash_pageinit(BufferGetPage(buf), BufferGetPageSize(buf));
157 :
158 8 : return buf;
159 : }
160 :
161 : /*
162 : * _hash_initbuf() -- Get and initialize a buffer by bucket number.
163 : */
164 : void
165 428 : _hash_initbuf(Buffer buf, uint32 max_bucket, uint32 num_bucket, uint32 flag,
166 : bool initpage)
167 : {
168 : HashPageOpaque pageopaque;
169 : Page page;
170 :
171 428 : page = BufferGetPage(buf);
172 :
173 : /* initialize the page */
174 428 : if (initpage)
175 0 : _hash_pageinit(page, BufferGetPageSize(buf));
176 :
177 428 : pageopaque = (HashPageOpaque) PageGetSpecialPointer(page);
178 :
179 : /*
180 : * Set hasho_prevblkno with current hashm_maxbucket. This value will be
181 : * used to validate cached HashMetaPageData. See
182 : * _hash_getbucketbuf_from_hashkey().
183 : */
184 428 : pageopaque->hasho_prevblkno = max_bucket;
185 428 : pageopaque->hasho_nextblkno = InvalidBlockNumber;
186 428 : pageopaque->hasho_bucket = num_bucket;
187 428 : pageopaque->hasho_flag = flag;
188 428 : pageopaque->hasho_page_id = HASHO_PAGE_ID;
189 428 : }
190 :
191 : /*
192 : * _hash_getnewbuf() -- Get a new page at the end of the index.
193 : *
194 : * This has the same API as _hash_getinitbuf, except that we are adding
195 : * a page to the index, and hence expect the page to be past the
196 : * logical EOF. (However, we have to support the case where it isn't,
197 : * since a prior try might have crashed after extending the filesystem
198 : * EOF but before updating the metapage to reflect the added page.)
199 : *
200 : * It is caller's responsibility to ensure that only one process can
201 : * extend the index at a time. In practice, this function is called
202 : * only while holding write lock on the metapage, because adding a page
203 : * is always associated with an update of metapage data.
204 : */
205 : Buffer
206 552 : _hash_getnewbuf(Relation rel, BlockNumber blkno, ForkNumber forkNum)
207 : {
208 552 : BlockNumber nblocks = RelationGetNumberOfBlocksInFork(rel, forkNum);
209 : Buffer buf;
210 :
211 552 : if (blkno == P_NEW)
212 0 : elog(ERROR, "hash AM does not use P_NEW");
213 552 : if (blkno > nblocks)
214 0 : elog(ERROR, "access to noncontiguous page in hash index \"%s\"",
215 : RelationGetRelationName(rel));
216 :
217 : /* smgr insists we use P_NEW to extend the relation */
218 552 : if (blkno == nblocks)
219 : {
220 480 : buf = ReadBufferExtended(rel, forkNum, P_NEW, RBM_NORMAL, NULL);
221 480 : if (BufferGetBlockNumber(buf) != blkno)
222 0 : elog(ERROR, "unexpected hash relation size: %u, should be %u",
223 : BufferGetBlockNumber(buf), blkno);
224 480 : LockBuffer(buf, HASH_WRITE);
225 : }
226 : else
227 : {
228 72 : buf = ReadBufferExtended(rel, forkNum, blkno, RBM_ZERO_AND_LOCK,
229 : NULL);
230 : }
231 :
232 : /* ref count and lock type are correct */
233 :
234 : /* initialize the page */
235 552 : _hash_pageinit(BufferGetPage(buf), BufferGetPageSize(buf));
236 :
237 552 : return buf;
238 : }
239 :
240 : /*
241 : * _hash_getbuf_with_strategy() -- Get a buffer with nondefault strategy.
242 : *
243 : * This is identical to _hash_getbuf() but also allows a buffer access
244 : * strategy to be specified. We use this for VACUUM operations.
245 : */
246 : Buffer
247 132 : _hash_getbuf_with_strategy(Relation rel, BlockNumber blkno,
248 : int access, int flags,
249 : BufferAccessStrategy bstrategy)
250 : {
251 : Buffer buf;
252 :
253 132 : if (blkno == P_NEW)
254 0 : elog(ERROR, "hash AM does not use P_NEW");
255 :
256 132 : buf = ReadBufferExtended(rel, MAIN_FORKNUM, blkno, RBM_NORMAL, bstrategy);
257 :
258 132 : if (access != HASH_NOLOCK)
259 132 : LockBuffer(buf, access);
260 :
261 : /* ref count and lock type are correct */
262 :
263 132 : _hash_checkpage(rel, buf, flags);
264 :
265 132 : return buf;
266 : }
267 :
268 : /*
269 : * _hash_relbuf() -- release a locked buffer.
270 : *
271 : * Lock and pin (refcount) are both dropped.
272 : */
273 : void
274 109431 : _hash_relbuf(Relation rel, Buffer buf)
275 : {
276 109431 : UnlockReleaseBuffer(buf);
277 109431 : }
278 :
279 : /*
280 : * _hash_dropbuf() -- release an unlocked buffer.
281 : *
282 : * This is used to unpin a buffer on which we hold no lock.
283 : */
284 : void
285 86768 : _hash_dropbuf(Relation rel, Buffer buf)
286 : {
287 86768 : ReleaseBuffer(buf);
288 86768 : }
289 :
290 : /*
291 : * _hash_dropscanbuf() -- release buffers used in scan.
292 : *
293 : * This routine unpins the buffers used during scan on which we
294 : * hold no lock.
295 : */
296 : void
297 74 : _hash_dropscanbuf(Relation rel, HashScanOpaque so)
298 : {
299 : /* release pin we hold on primary bucket page */
300 99 : if (BufferIsValid(so->hashso_bucket_buf) &&
301 25 : so->hashso_bucket_buf != so->hashso_curbuf)
302 25 : _hash_dropbuf(rel, so->hashso_bucket_buf);
303 74 : so->hashso_bucket_buf = InvalidBuffer;
304 :
305 : /* release pin we hold on primary bucket page of bucket being split */
306 74 : if (BufferIsValid(so->hashso_split_bucket_buf) &&
307 0 : so->hashso_split_bucket_buf != so->hashso_curbuf)
308 0 : _hash_dropbuf(rel, so->hashso_split_bucket_buf);
309 74 : so->hashso_split_bucket_buf = InvalidBuffer;
310 :
311 : /* release any pin we still hold */
312 74 : if (BufferIsValid(so->hashso_curbuf))
313 0 : _hash_dropbuf(rel, so->hashso_curbuf);
314 74 : so->hashso_curbuf = InvalidBuffer;
315 :
316 : /* reset split scan */
317 74 : so->hashso_buc_populated = false;
318 74 : so->hashso_buc_split = false;
319 74 : }
320 :
321 :
322 : /*
323 : * _hash_init() -- Initialize the metadata page of a hash index,
324 : * the initial buckets, and the initial bitmap page.
325 : *
326 : * The initial number of buckets is dependent on num_tuples, an estimate
327 : * of the number of tuples to be loaded into the index initially. The
328 : * chosen number of buckets is returned.
329 : *
330 : * We are fairly cavalier about locking here, since we know that no one else
331 : * could be accessing this index. In particular the rule about not holding
332 : * multiple buffer locks is ignored.
333 : */
334 : uint32
335 15 : _hash_init(Relation rel, double num_tuples, ForkNumber forkNum)
336 : {
337 : Buffer metabuf;
338 : Buffer buf;
339 : Buffer bitmapbuf;
340 : Page pg;
341 : HashMetaPage metap;
342 : RegProcedure procid;
343 : int32 data_width;
344 : int32 item_width;
345 : int32 ffactor;
346 : uint32 num_buckets;
347 : uint32 i;
348 : bool use_wal;
349 :
350 : /* safety check */
351 15 : if (RelationGetNumberOfBlocksInFork(rel, forkNum) != 0)
352 0 : elog(ERROR, "cannot initialize non-empty hash index \"%s\"",
353 : RelationGetRelationName(rel));
354 :
355 : /*
356 : * WAL log creation of pages if the relation is persistent, or this is the
357 : * init fork. Init forks for unlogged relations always need to be WAL
358 : * logged.
359 : */
360 15 : use_wal = RelationNeedsWAL(rel) || forkNum == INIT_FORKNUM;
361 :
362 : /*
363 : * Determine the target fill factor (in tuples per bucket) for this index.
364 : * The idea is to make the fill factor correspond to pages about as full
365 : * as the user-settable fillfactor parameter says. We can compute it
366 : * exactly since the index datatype (i.e. uint32 hash key) is fixed-width.
367 : */
368 15 : data_width = sizeof(uint32);
369 15 : item_width = MAXALIGN(sizeof(IndexTupleData)) + MAXALIGN(data_width) +
370 : sizeof(ItemIdData); /* include the line pointer */
371 15 : ffactor = RelationGetTargetPageUsage(rel, HASH_DEFAULT_FILLFACTOR) / item_width;
372 : /* keep to a sane range */
373 15 : if (ffactor < 10)
374 0 : ffactor = 10;
375 :
376 15 : procid = index_getprocid(rel, 1, HASHSTANDARD_PROC);
377 :
378 : /*
379 : * We initialize the metapage, the first N bucket pages, and the first
380 : * bitmap page in sequence, using _hash_getnewbuf to cause smgrextend()
381 : * calls to occur. This ensures that the smgr level has the right idea of
382 : * the physical index length.
383 : *
384 : * Critical section not required, because on error the creation of the
385 : * whole relation will be rolled back.
386 : */
387 15 : metabuf = _hash_getnewbuf(rel, HASH_METAPAGE, forkNum);
388 15 : _hash_init_metabuffer(metabuf, num_tuples, procid, ffactor, false);
389 15 : MarkBufferDirty(metabuf);
390 :
391 15 : pg = BufferGetPage(metabuf);
392 15 : metap = HashPageGetMeta(pg);
393 :
394 : /* XLOG stuff */
395 15 : if (use_wal)
396 : {
397 : xl_hash_init_meta_page xlrec;
398 : XLogRecPtr recptr;
399 :
400 13 : xlrec.num_tuples = num_tuples;
401 13 : xlrec.procid = metap->hashm_procid;
402 13 : xlrec.ffactor = metap->hashm_ffactor;
403 :
404 13 : XLogBeginInsert();
405 13 : XLogRegisterData((char *) &xlrec, SizeOfHashInitMetaPage);
406 13 : XLogRegisterBuffer(0, metabuf, REGBUF_WILL_INIT);
407 :
408 13 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_INIT_META_PAGE);
409 :
410 13 : PageSetLSN(BufferGetPage(metabuf), recptr);
411 : }
412 :
413 15 : num_buckets = metap->hashm_maxbucket + 1;
414 :
415 : /*
416 : * Release buffer lock on the metapage while we initialize buckets.
417 : * Otherwise, we'll be in interrupt holdoff and the CHECK_FOR_INTERRUPTS
418 : * won't accomplish anything. It's a bad idea to hold buffer locks for
419 : * long intervals in any case, since that can block the bgwriter.
420 : */
421 15 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
422 :
423 : /*
424 : * Initialize and WAL Log the first N buckets
425 : */
426 443 : for (i = 0; i < num_buckets; i++)
427 : {
428 : BlockNumber blkno;
429 :
430 : /* Allow interrupts, in case N is huge */
431 428 : CHECK_FOR_INTERRUPTS();
432 :
433 428 : blkno = BUCKET_TO_BLKNO(metap, i);
434 428 : buf = _hash_getnewbuf(rel, blkno, forkNum);
435 428 : _hash_initbuf(buf, metap->hashm_maxbucket, i, LH_BUCKET_PAGE, false);
436 428 : MarkBufferDirty(buf);
437 :
438 428 : if (use_wal)
439 412 : log_newpage(&rel->rd_node,
440 : forkNum,
441 : blkno,
442 412 : BufferGetPage(buf),
443 : true);
444 428 : _hash_relbuf(rel, buf);
445 : }
446 :
447 : /* Now reacquire buffer lock on metapage */
448 15 : LockBuffer(metabuf, BUFFER_LOCK_EXCLUSIVE);
449 :
450 : /*
451 : * Initialize bitmap page
452 : */
453 15 : bitmapbuf = _hash_getnewbuf(rel, num_buckets + 1, forkNum);
454 15 : _hash_initbitmapbuffer(bitmapbuf, metap->hashm_bmsize, false);
455 15 : MarkBufferDirty(bitmapbuf);
456 :
457 : /* add the new bitmap page to the metapage's list of bitmaps */
458 : /* metapage already has a write lock */
459 15 : if (metap->hashm_nmaps >= HASH_MAX_BITMAPS)
460 0 : ereport(ERROR,
461 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
462 : errmsg("out of overflow pages in hash index \"%s\"",
463 : RelationGetRelationName(rel))));
464 :
465 15 : metap->hashm_mapp[metap->hashm_nmaps] = num_buckets + 1;
466 :
467 15 : metap->hashm_nmaps++;
468 15 : MarkBufferDirty(metabuf);
469 :
470 : /* XLOG stuff */
471 15 : if (use_wal)
472 : {
473 : xl_hash_init_bitmap_page xlrec;
474 : XLogRecPtr recptr;
475 :
476 13 : xlrec.bmsize = metap->hashm_bmsize;
477 :
478 13 : XLogBeginInsert();
479 13 : XLogRegisterData((char *) &xlrec, SizeOfHashInitBitmapPage);
480 13 : XLogRegisterBuffer(0, bitmapbuf, REGBUF_WILL_INIT);
481 :
482 : /*
483 : * This is safe only because nobody else can be modifying the index at
484 : * this stage; it's only visible to the transaction that is creating
485 : * it.
486 : */
487 13 : XLogRegisterBuffer(1, metabuf, REGBUF_STANDARD);
488 :
489 13 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_INIT_BITMAP_PAGE);
490 :
491 13 : PageSetLSN(BufferGetPage(bitmapbuf), recptr);
492 13 : PageSetLSN(BufferGetPage(metabuf), recptr);
493 : }
494 :
495 : /* all done */
496 15 : _hash_relbuf(rel, bitmapbuf);
497 15 : _hash_relbuf(rel, metabuf);
498 :
499 15 : return num_buckets;
500 : }
501 :
502 : /*
503 : * _hash_init_metabuffer() -- Initialize the metadata page of a hash index.
504 : */
505 : void
506 15 : _hash_init_metabuffer(Buffer buf, double num_tuples, RegProcedure procid,
507 : uint16 ffactor, bool initpage)
508 : {
509 : HashMetaPage metap;
510 : HashPageOpaque pageopaque;
511 : Page page;
512 : double dnumbuckets;
513 : uint32 num_buckets;
514 : uint32 spare_index;
515 : uint32 i;
516 :
517 : /*
518 : * Choose the number of initial bucket pages to match the fill factor
519 : * given the estimated number of tuples. We round up the result to the
520 : * total number of buckets which has to be allocated before using its
521 : * _hashm_spare element. However always force at least 2 bucket pages. The
522 : * upper limit is determined by considerations explained in
523 : * _hash_expandtable().
524 : */
525 15 : dnumbuckets = num_tuples / ffactor;
526 15 : if (dnumbuckets <= 2.0)
527 6 : num_buckets = 2;
528 9 : else if (dnumbuckets >= (double) 0x40000000)
529 0 : num_buckets = 0x40000000;
530 : else
531 9 : num_buckets = _hash_get_totalbuckets(_hash_spareindex(dnumbuckets));
532 :
533 15 : spare_index = _hash_spareindex(num_buckets);
534 15 : Assert(spare_index < HASH_MAX_SPLITPOINTS);
535 :
536 15 : page = BufferGetPage(buf);
537 15 : if (initpage)
538 0 : _hash_pageinit(page, BufferGetPageSize(buf));
539 :
540 15 : pageopaque = (HashPageOpaque) PageGetSpecialPointer(page);
541 15 : pageopaque->hasho_prevblkno = InvalidBlockNumber;
542 15 : pageopaque->hasho_nextblkno = InvalidBlockNumber;
543 15 : pageopaque->hasho_bucket = -1;
544 15 : pageopaque->hasho_flag = LH_META_PAGE;
545 15 : pageopaque->hasho_page_id = HASHO_PAGE_ID;
546 :
547 15 : metap = HashPageGetMeta(page);
548 :
549 15 : metap->hashm_magic = HASH_MAGIC;
550 15 : metap->hashm_version = HASH_VERSION;
551 15 : metap->hashm_ntuples = 0;
552 15 : metap->hashm_nmaps = 0;
553 15 : metap->hashm_ffactor = ffactor;
554 15 : metap->hashm_bsize = HashGetMaxBitmapSize(page);
555 : /* find largest bitmap array size that will fit in page size */
556 30 : for (i = _hash_log2(metap->hashm_bsize); i > 0; --i)
557 : {
558 30 : if ((1 << i) <= metap->hashm_bsize)
559 15 : break;
560 : }
561 15 : Assert(i > 0);
562 15 : metap->hashm_bmsize = 1 << i;
563 15 : metap->hashm_bmshift = i + BYTE_TO_BIT;
564 15 : Assert((1 << BMPG_SHIFT(metap)) == (BMPG_MASK(metap) + 1));
565 :
566 : /*
567 : * Label the index with its primary hash support function's OID. This is
568 : * pretty useless for normal operation (in fact, hashm_procid is not used
569 : * anywhere), but it might be handy for forensic purposes so we keep it.
570 : */
571 15 : metap->hashm_procid = procid;
572 :
573 : /*
574 : * We initialize the index with N buckets, 0 .. N-1, occupying physical
575 : * blocks 1 to N. The first freespace bitmap page is in block N+1.
576 : */
577 15 : metap->hashm_maxbucket = num_buckets - 1;
578 :
579 : /*
580 : * Set highmask as next immediate ((2 ^ x) - 1), which should be
581 : * sufficient to cover num_buckets.
582 : */
583 15 : metap->hashm_highmask = (1 << (_hash_log2(num_buckets + 1))) - 1;
584 15 : metap->hashm_lowmask = (metap->hashm_highmask >> 1);
585 :
586 15 : MemSet(metap->hashm_spares, 0, sizeof(metap->hashm_spares));
587 15 : MemSet(metap->hashm_mapp, 0, sizeof(metap->hashm_mapp));
588 :
589 : /* Set up mapping for one spare page after the initial splitpoints */
590 15 : metap->hashm_spares[spare_index] = 1;
591 15 : metap->hashm_ovflpoint = spare_index;
592 15 : metap->hashm_firstfree = 0;
593 :
594 : /*
595 : * Set pd_lower just past the end of the metadata. This is to log full
596 : * page image of metapage in xloginsert.c.
597 : */
598 15 : ((PageHeader) page)->pd_lower =
599 15 : ((char *) metap + sizeof(HashMetaPageData)) - (char *) page;
600 15 : }
601 :
602 : /*
603 : * _hash_pageinit() -- Initialize a new hash index page.
604 : */
605 : void
606 583 : _hash_pageinit(Page page, Size size)
607 : {
608 583 : PageInit(page, size, sizeof(HashPageOpaqueData));
609 583 : }
610 :
611 : /*
612 : * Attempt to expand the hash table by creating one new bucket.
613 : *
614 : * This will silently do nothing if we don't get cleanup lock on old or
615 : * new bucket.
616 : *
617 : * Complete the pending splits and remove the tuples from old bucket,
618 : * if there are any left over from the previous split.
619 : *
620 : * The caller must hold a pin, but no lock, on the metapage buffer.
621 : * The buffer is returned in the same state.
622 : */
623 : void
624 72 : _hash_expandtable(Relation rel, Buffer metabuf)
625 : {
626 : HashMetaPage metap;
627 : Bucket old_bucket;
628 : Bucket new_bucket;
629 : uint32 spare_ndx;
630 : BlockNumber start_oblkno;
631 : BlockNumber start_nblkno;
632 : Buffer buf_nblkno;
633 : Buffer buf_oblkno;
634 : Page opage;
635 : Page npage;
636 : HashPageOpaque oopaque;
637 : HashPageOpaque nopaque;
638 : uint32 maxbucket;
639 : uint32 highmask;
640 : uint32 lowmask;
641 72 : bool metap_update_masks = false;
642 72 : bool metap_update_splitpoint = false;
643 :
644 : restart_expand:
645 :
646 : /*
647 : * Write-lock the meta page. It used to be necessary to acquire a
648 : * heavyweight lock to begin a split, but that is no longer required.
649 : */
650 72 : LockBuffer(metabuf, BUFFER_LOCK_EXCLUSIVE);
651 :
652 72 : _hash_checkpage(rel, metabuf, LH_META_PAGE);
653 72 : metap = HashPageGetMeta(BufferGetPage(metabuf));
654 :
655 : /*
656 : * Check to see if split is still needed; someone else might have already
657 : * done one while we waited for the lock.
658 : *
659 : * Make sure this stays in sync with _hash_doinsert()
660 : */
661 216 : if (metap->hashm_ntuples <=
662 144 : (double) metap->hashm_ffactor * (metap->hashm_maxbucket + 1))
663 0 : goto fail;
664 :
665 : /*
666 : * Can't split anymore if maxbucket has reached its maximum possible
667 : * value.
668 : *
669 : * Ideally we'd allow bucket numbers up to UINT_MAX-1 (no higher because
670 : * the calculation maxbucket+1 mustn't overflow). Currently we restrict
671 : * to half that because of overflow looping in _hash_log2() and
672 : * insufficient space in hashm_spares[]. It's moot anyway because an
673 : * index with 2^32 buckets would certainly overflow BlockNumber and hence
674 : * _hash_alloc_buckets() would fail, but if we supported buckets smaller
675 : * than a disk block then this would be an independent constraint.
676 : *
677 : * If you change this, see also the maximum initial number of buckets in
678 : * _hash_init().
679 : */
680 72 : if (metap->hashm_maxbucket >= (uint32) 0x7FFFFFFE)
681 0 : goto fail;
682 :
683 : /*
684 : * Determine which bucket is to be split, and attempt to take cleanup lock
685 : * on the old bucket. If we can't get the lock, give up.
686 : *
687 : * The cleanup lock protects us not only against other backends, but
688 : * against our own backend as well.
689 : *
690 : * The cleanup lock is mainly to protect the split from concurrent
691 : * inserts. See src/backend/access/hash/README, Lock Definitions for
692 : * further details. Due to this locking restriction, if there is any
693 : * pending scan, the split will give up which is not good, but harmless.
694 : */
695 72 : new_bucket = metap->hashm_maxbucket + 1;
696 :
697 72 : old_bucket = (new_bucket & metap->hashm_lowmask);
698 :
699 72 : start_oblkno = BUCKET_TO_BLKNO(metap, old_bucket);
700 :
701 72 : buf_oblkno = _hash_getbuf_with_condlock_cleanup(rel, start_oblkno, LH_BUCKET_PAGE);
702 72 : if (!buf_oblkno)
703 0 : goto fail;
704 :
705 72 : opage = BufferGetPage(buf_oblkno);
706 72 : oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
707 :
708 : /*
709 : * We want to finish the split from a bucket as there is no apparent
710 : * benefit by not doing so and it will make the code complicated to finish
711 : * the split that involves multiple buckets considering the case where new
712 : * split also fails. We don't need to consider the new bucket for
713 : * completing the split here as it is not possible that a re-split of new
714 : * bucket starts when there is still a pending split from old bucket.
715 : */
716 72 : if (H_BUCKET_BEING_SPLIT(oopaque))
717 : {
718 : /*
719 : * Copy bucket mapping info now; refer the comment in code below where
720 : * we copy this information before calling _hash_splitbucket to see
721 : * why this is okay.
722 : */
723 0 : maxbucket = metap->hashm_maxbucket;
724 0 : highmask = metap->hashm_highmask;
725 0 : lowmask = metap->hashm_lowmask;
726 :
727 : /*
728 : * Release the lock on metapage and old_bucket, before completing the
729 : * split.
730 : */
731 0 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
732 0 : LockBuffer(buf_oblkno, BUFFER_LOCK_UNLOCK);
733 :
734 0 : _hash_finish_split(rel, metabuf, buf_oblkno, old_bucket, maxbucket,
735 : highmask, lowmask);
736 :
737 : /* release the pin on old buffer and retry for expand. */
738 0 : _hash_dropbuf(rel, buf_oblkno);
739 :
740 0 : goto restart_expand;
741 : }
742 :
743 : /*
744 : * Clean the tuples remained from the previous split. This operation
745 : * requires cleanup lock and we already have one on the old bucket, so
746 : * let's do it. We also don't want to allow further splits from the bucket
747 : * till the garbage of previous split is cleaned. This has two
748 : * advantages; first, it helps in avoiding the bloat due to garbage and
749 : * second is, during cleanup of bucket, we are always sure that the
750 : * garbage tuples belong to most recently split bucket. On the contrary,
751 : * if we allow cleanup of bucket after meta page is updated to indicate
752 : * the new split and before the actual split, the cleanup operation won't
753 : * be able to decide whether the tuple has been moved to the newly created
754 : * bucket and ended up deleting such tuples.
755 : */
756 72 : if (H_NEEDS_SPLIT_CLEANUP(oopaque))
757 : {
758 : /*
759 : * Copy bucket mapping info now; refer to the comment in code below
760 : * where we copy this information before calling _hash_splitbucket to
761 : * see why this is okay.
762 : */
763 0 : maxbucket = metap->hashm_maxbucket;
764 0 : highmask = metap->hashm_highmask;
765 0 : lowmask = metap->hashm_lowmask;
766 :
767 : /* Release the metapage lock. */
768 0 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
769 :
770 0 : hashbucketcleanup(rel, old_bucket, buf_oblkno, start_oblkno, NULL,
771 : maxbucket, highmask, lowmask, NULL, NULL, true,
772 : NULL, NULL);
773 :
774 0 : _hash_dropbuf(rel, buf_oblkno);
775 :
776 0 : goto restart_expand;
777 : }
778 :
779 : /*
780 : * There shouldn't be any active scan on new bucket.
781 : *
782 : * Note: it is safe to compute the new bucket's blkno here, even though we
783 : * may still need to update the BUCKET_TO_BLKNO mapping. This is because
784 : * the current value of hashm_spares[hashm_ovflpoint] correctly shows
785 : * where we are going to put a new splitpoint's worth of buckets.
786 : */
787 72 : start_nblkno = BUCKET_TO_BLKNO(metap, new_bucket);
788 :
789 : /*
790 : * If the split point is increasing we need to allocate a new batch of
791 : * bucket pages.
792 : */
793 72 : spare_ndx = _hash_spareindex(new_bucket + 1);
794 72 : if (spare_ndx > metap->hashm_ovflpoint)
795 : {
796 : uint32 buckets_to_add;
797 :
798 4 : Assert(spare_ndx == metap->hashm_ovflpoint + 1);
799 :
800 : /*
801 : * We treat allocation of buckets as a separate WAL-logged action.
802 : * Even if we fail after this operation, won't leak bucket pages;
803 : * rather, the next split will consume this space. In any case, even
804 : * without failure we don't use all the space in one split operation.
805 : */
806 4 : buckets_to_add = _hash_get_totalbuckets(spare_ndx) - new_bucket;
807 4 : if (!_hash_alloc_buckets(rel, start_nblkno, buckets_to_add))
808 : {
809 : /* can't split due to BlockNumber overflow */
810 0 : _hash_relbuf(rel, buf_oblkno);
811 0 : goto fail;
812 : }
813 : }
814 :
815 : /*
816 : * Physically allocate the new bucket's primary page. We want to do this
817 : * before changing the metapage's mapping info, in case we can't get the
818 : * disk space. Ideally, we don't need to check for cleanup lock on new
819 : * bucket as no other backend could find this bucket unless meta page is
820 : * updated. However, it is good to be consistent with old bucket locking.
821 : */
822 72 : buf_nblkno = _hash_getnewbuf(rel, start_nblkno, MAIN_FORKNUM);
823 72 : if (!IsBufferCleanupOK(buf_nblkno))
824 : {
825 0 : _hash_relbuf(rel, buf_oblkno);
826 0 : _hash_relbuf(rel, buf_nblkno);
827 0 : goto fail;
828 : }
829 :
830 : /*
831 : * Since we are scribbling on the pages in the shared buffers, establish a
832 : * critical section. Any failure in this next code leaves us with a big
833 : * problem: the metapage is effectively corrupt but could get written back
834 : * to disk.
835 : */
836 72 : START_CRIT_SECTION();
837 :
838 : /*
839 : * Okay to proceed with split. Update the metapage bucket mapping info.
840 : */
841 72 : metap->hashm_maxbucket = new_bucket;
842 :
843 72 : if (new_bucket > metap->hashm_highmask)
844 : {
845 : /* Starting a new doubling */
846 3 : metap->hashm_lowmask = metap->hashm_highmask;
847 3 : metap->hashm_highmask = new_bucket | metap->hashm_lowmask;
848 3 : metap_update_masks = true;
849 : }
850 :
851 : /*
852 : * If the split point is increasing we need to adjust the hashm_spares[]
853 : * array and hashm_ovflpoint so that future overflow pages will be created
854 : * beyond this new batch of bucket pages.
855 : */
856 72 : if (spare_ndx > metap->hashm_ovflpoint)
857 : {
858 4 : metap->hashm_spares[spare_ndx] = metap->hashm_spares[metap->hashm_ovflpoint];
859 4 : metap->hashm_ovflpoint = spare_ndx;
860 4 : metap_update_splitpoint = true;
861 : }
862 :
863 72 : MarkBufferDirty(metabuf);
864 :
865 : /*
866 : * Copy bucket mapping info now; this saves re-accessing the meta page
867 : * inside _hash_splitbucket's inner loop. Note that once we drop the
868 : * split lock, other splits could begin, so these values might be out of
869 : * date before _hash_splitbucket finishes. That's okay, since all it
870 : * needs is to tell which of these two buckets to map hashkeys into.
871 : */
872 72 : maxbucket = metap->hashm_maxbucket;
873 72 : highmask = metap->hashm_highmask;
874 72 : lowmask = metap->hashm_lowmask;
875 :
876 72 : opage = BufferGetPage(buf_oblkno);
877 72 : oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
878 :
879 : /*
880 : * Mark the old bucket to indicate that split is in progress. (At
881 : * operation end, we will clear the split-in-progress flag.) Also, for a
882 : * primary bucket page, hasho_prevblkno stores the number of buckets that
883 : * existed as of the last split, so we must update that value here.
884 : */
885 72 : oopaque->hasho_flag |= LH_BUCKET_BEING_SPLIT;
886 72 : oopaque->hasho_prevblkno = maxbucket;
887 :
888 72 : MarkBufferDirty(buf_oblkno);
889 :
890 72 : npage = BufferGetPage(buf_nblkno);
891 :
892 : /*
893 : * initialize the new bucket's primary page and mark it to indicate that
894 : * split is in progress.
895 : */
896 72 : nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
897 72 : nopaque->hasho_prevblkno = maxbucket;
898 72 : nopaque->hasho_nextblkno = InvalidBlockNumber;
899 72 : nopaque->hasho_bucket = new_bucket;
900 72 : nopaque->hasho_flag = LH_BUCKET_PAGE | LH_BUCKET_BEING_POPULATED;
901 72 : nopaque->hasho_page_id = HASHO_PAGE_ID;
902 :
903 72 : MarkBufferDirty(buf_nblkno);
904 :
905 : /* XLOG stuff */
906 72 : if (RelationNeedsWAL(rel))
907 : {
908 : xl_hash_split_allocate_page xlrec;
909 : XLogRecPtr recptr;
910 :
911 72 : xlrec.new_bucket = maxbucket;
912 72 : xlrec.old_bucket_flag = oopaque->hasho_flag;
913 72 : xlrec.new_bucket_flag = nopaque->hasho_flag;
914 72 : xlrec.flags = 0;
915 :
916 72 : XLogBeginInsert();
917 :
918 72 : XLogRegisterBuffer(0, buf_oblkno, REGBUF_STANDARD);
919 72 : XLogRegisterBuffer(1, buf_nblkno, REGBUF_WILL_INIT);
920 72 : XLogRegisterBuffer(2, metabuf, REGBUF_STANDARD);
921 :
922 72 : if (metap_update_masks)
923 : {
924 3 : xlrec.flags |= XLH_SPLIT_META_UPDATE_MASKS;
925 3 : XLogRegisterBufData(2, (char *) &metap->hashm_lowmask, sizeof(uint32));
926 3 : XLogRegisterBufData(2, (char *) &metap->hashm_highmask, sizeof(uint32));
927 : }
928 :
929 72 : if (metap_update_splitpoint)
930 : {
931 4 : xlrec.flags |= XLH_SPLIT_META_UPDATE_SPLITPOINT;
932 4 : XLogRegisterBufData(2, (char *) &metap->hashm_ovflpoint,
933 : sizeof(uint32));
934 4 : XLogRegisterBufData(2,
935 4 : (char *) &metap->hashm_spares[metap->hashm_ovflpoint],
936 : sizeof(uint32));
937 : }
938 :
939 72 : XLogRegisterData((char *) &xlrec, SizeOfHashSplitAllocPage);
940 :
941 72 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_ALLOCATE_PAGE);
942 :
943 72 : PageSetLSN(BufferGetPage(buf_oblkno), recptr);
944 72 : PageSetLSN(BufferGetPage(buf_nblkno), recptr);
945 72 : PageSetLSN(BufferGetPage(metabuf), recptr);
946 : }
947 :
948 72 : END_CRIT_SECTION();
949 :
950 : /* drop lock, but keep pin */
951 72 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
952 :
953 : /* Relocate records to the new bucket */
954 72 : _hash_splitbucket(rel, metabuf,
955 : old_bucket, new_bucket,
956 : buf_oblkno, buf_nblkno, NULL,
957 : maxbucket, highmask, lowmask);
958 :
959 : /* all done, now release the pins on primary buckets. */
960 72 : _hash_dropbuf(rel, buf_oblkno);
961 72 : _hash_dropbuf(rel, buf_nblkno);
962 :
963 144 : return;
964 :
965 : /* Here if decide not to split or fail to acquire old bucket lock */
966 : fail:
967 :
968 : /* We didn't write the metapage, so just drop lock */
969 0 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
970 : }
971 :
972 :
973 : /*
974 : * _hash_alloc_buckets -- allocate a new splitpoint's worth of bucket pages
975 : *
976 : * This does not need to initialize the new bucket pages; we'll do that as
977 : * each one is used by _hash_expandtable(). But we have to extend the logical
978 : * EOF to the end of the splitpoint; this keeps smgr's idea of the EOF in
979 : * sync with ours, so that we don't get complaints from smgr.
980 : *
981 : * We do this by writing a page of zeroes at the end of the splitpoint range.
982 : * We expect that the filesystem will ensure that the intervening pages read
983 : * as zeroes too. On many filesystems this "hole" will not be allocated
984 : * immediately, which means that the index file may end up more fragmented
985 : * than if we forced it all to be allocated now; but since we don't scan
986 : * hash indexes sequentially anyway, that probably doesn't matter.
987 : *
988 : * XXX It's annoying that this code is executed with the metapage lock held.
989 : * We need to interlock against _hash_addovflpage() adding a new overflow page
990 : * concurrently, but it'd likely be better to use LockRelationForExtension
991 : * for the purpose. OTOH, adding a splitpoint is a very infrequent operation,
992 : * so it may not be worth worrying about.
993 : *
994 : * Returns TRUE if successful, or FALSE if allocation failed due to
995 : * BlockNumber overflow.
996 : */
997 : static bool
998 4 : _hash_alloc_buckets(Relation rel, BlockNumber firstblock, uint32 nblocks)
999 : {
1000 : BlockNumber lastblock;
1001 : char zerobuf[BLCKSZ];
1002 : Page page;
1003 : HashPageOpaque ovflopaque;
1004 :
1005 4 : lastblock = firstblock + nblocks - 1;
1006 :
1007 : /*
1008 : * Check for overflow in block number calculation; if so, we cannot extend
1009 : * the index anymore.
1010 : */
1011 4 : if (lastblock < firstblock || lastblock == InvalidBlockNumber)
1012 0 : return false;
1013 :
1014 4 : page = (Page) zerobuf;
1015 :
1016 : /*
1017 : * Initialize the page. Just zeroing the page won't work; see
1018 : * _hash_freeovflpage for similar usage. We take care to make the special
1019 : * space valid for the benefit of tools such as pageinspect.
1020 : */
1021 4 : _hash_pageinit(page, BLCKSZ);
1022 :
1023 4 : ovflopaque = (HashPageOpaque) PageGetSpecialPointer(page);
1024 :
1025 4 : ovflopaque->hasho_prevblkno = InvalidBlockNumber;
1026 4 : ovflopaque->hasho_nextblkno = InvalidBlockNumber;
1027 4 : ovflopaque->hasho_bucket = -1;
1028 4 : ovflopaque->hasho_flag = LH_UNUSED_PAGE;
1029 4 : ovflopaque->hasho_page_id = HASHO_PAGE_ID;
1030 :
1031 4 : if (RelationNeedsWAL(rel))
1032 4 : log_newpage(&rel->rd_node,
1033 : MAIN_FORKNUM,
1034 : lastblock,
1035 : zerobuf,
1036 : true);
1037 :
1038 4 : RelationOpenSmgr(rel);
1039 4 : smgrextend(rel->rd_smgr, MAIN_FORKNUM, lastblock, zerobuf, false);
1040 :
1041 4 : return true;
1042 : }
1043 :
1044 :
1045 : /*
1046 : * _hash_splitbucket -- split 'obucket' into 'obucket' and 'nbucket'
1047 : *
1048 : * This routine is used to partition the tuples between old and new bucket and
1049 : * is used to finish the incomplete split operations. To finish the previously
1050 : * interrupted split operation, the caller needs to fill htab. If htab is set,
1051 : * then we skip the movement of tuples that exists in htab, otherwise NULL
1052 : * value of htab indicates movement of all the tuples that belong to the new
1053 : * bucket.
1054 : *
1055 : * We are splitting a bucket that consists of a base bucket page and zero
1056 : * or more overflow (bucket chain) pages. We must relocate tuples that
1057 : * belong in the new bucket.
1058 : *
1059 : * The caller must hold cleanup locks on both buckets to ensure that
1060 : * no one else is trying to access them (see README).
1061 : *
1062 : * The caller must hold a pin, but no lock, on the metapage buffer.
1063 : * The buffer is returned in the same state. (The metapage is only
1064 : * touched if it becomes necessary to add or remove overflow pages.)
1065 : *
1066 : * Split needs to retain pin on primary bucket pages of both old and new
1067 : * buckets till end of operation. This is to prevent vacuum from starting
1068 : * while a split is in progress.
1069 : *
1070 : * In addition, the caller must have created the new bucket's base page,
1071 : * which is passed in buffer nbuf, pinned and write-locked. The lock will be
1072 : * released here and pin must be released by the caller. (The API is set up
1073 : * this way because we must do _hash_getnewbuf() before releasing the metapage
1074 : * write lock. So instead of passing the new bucket's start block number, we
1075 : * pass an actual buffer.)
1076 : */
1077 : static void
1078 72 : _hash_splitbucket(Relation rel,
1079 : Buffer metabuf,
1080 : Bucket obucket,
1081 : Bucket nbucket,
1082 : Buffer obuf,
1083 : Buffer nbuf,
1084 : HTAB *htab,
1085 : uint32 maxbucket,
1086 : uint32 highmask,
1087 : uint32 lowmask)
1088 : {
1089 : Buffer bucket_obuf;
1090 : Buffer bucket_nbuf;
1091 : Page opage;
1092 : Page npage;
1093 : HashPageOpaque oopaque;
1094 : HashPageOpaque nopaque;
1095 : OffsetNumber itup_offsets[MaxIndexTuplesPerPage];
1096 : IndexTuple itups[MaxIndexTuplesPerPage];
1097 72 : Size all_tups_size = 0;
1098 : int i;
1099 72 : uint16 nitups = 0;
1100 :
1101 72 : bucket_obuf = obuf;
1102 72 : opage = BufferGetPage(obuf);
1103 72 : oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
1104 :
1105 72 : bucket_nbuf = nbuf;
1106 72 : npage = BufferGetPage(nbuf);
1107 72 : nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
1108 :
1109 : /*
1110 : * Partition the tuples in the old bucket between the old bucket and the
1111 : * new bucket, advancing along the old bucket's overflow bucket chain and
1112 : * adding overflow pages to the new bucket as needed. Outer loop iterates
1113 : * once per page in old bucket.
1114 : */
1115 : for (;;)
1116 : {
1117 : BlockNumber oblkno;
1118 : OffsetNumber ooffnum;
1119 : OffsetNumber omaxoffnum;
1120 :
1121 : /* Scan each tuple in old page */
1122 123 : omaxoffnum = PageGetMaxOffsetNumber(opage);
1123 44114 : for (ooffnum = FirstOffsetNumber;
1124 : ooffnum <= omaxoffnum;
1125 43868 : ooffnum = OffsetNumberNext(ooffnum))
1126 : {
1127 : IndexTuple itup;
1128 : Size itemsz;
1129 : Bucket bucket;
1130 43868 : bool found = false;
1131 :
1132 : /* skip dead tuples */
1133 43868 : if (ItemIdIsDead(PageGetItemId(opage, ooffnum)))
1134 0 : continue;
1135 :
1136 : /*
1137 : * Before inserting a tuple, probe the hash table containing TIDs
1138 : * of tuples belonging to new bucket, if we find a match, then
1139 : * skip that tuple, else fetch the item's hash key (conveniently
1140 : * stored in the item) and determine which bucket it now belongs
1141 : * in.
1142 : */
1143 43868 : itup = (IndexTuple) PageGetItem(opage,
1144 : PageGetItemId(opage, ooffnum));
1145 :
1146 43868 : if (htab)
1147 0 : (void) hash_search(htab, &itup->t_tid, HASH_FIND, &found);
1148 :
1149 43868 : if (found)
1150 0 : continue;
1151 :
1152 43868 : bucket = _hash_hashkey2bucket(_hash_get_indextuple_hashkey(itup),
1153 : maxbucket, highmask, lowmask);
1154 :
1155 43868 : if (bucket == nbucket)
1156 : {
1157 : IndexTuple new_itup;
1158 :
1159 : /*
1160 : * make a copy of index tuple as we have to scribble on it.
1161 : */
1162 16530 : new_itup = CopyIndexTuple(itup);
1163 :
1164 : /*
1165 : * mark the index tuple as moved by split, such tuples are
1166 : * skipped by scan if there is split in progress for a bucket.
1167 : */
1168 16530 : new_itup->t_info |= INDEX_MOVED_BY_SPLIT_MASK;
1169 :
1170 : /*
1171 : * insert the tuple into the new bucket. if it doesn't fit on
1172 : * the current page in the new bucket, we must allocate a new
1173 : * overflow page and place the tuple on that page instead.
1174 : */
1175 16530 : itemsz = IndexTupleDSize(*new_itup);
1176 16530 : itemsz = MAXALIGN(itemsz);
1177 :
1178 16530 : if (PageGetFreeSpaceForMultipleTuples(npage, nitups + 1) < (all_tups_size + itemsz))
1179 : {
1180 : /*
1181 : * Change the shared buffer state in critical section,
1182 : * otherwise any error could make it unrecoverable.
1183 : */
1184 11 : START_CRIT_SECTION();
1185 :
1186 11 : _hash_pgaddmultitup(rel, nbuf, itups, itup_offsets, nitups);
1187 11 : MarkBufferDirty(nbuf);
1188 : /* log the split operation before releasing the lock */
1189 11 : log_split_page(rel, nbuf);
1190 :
1191 11 : END_CRIT_SECTION();
1192 :
1193 : /* drop lock, but keep pin */
1194 11 : LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
1195 :
1196 : /* be tidy */
1197 5610 : for (i = 0; i < nitups; i++)
1198 5599 : pfree(itups[i]);
1199 11 : nitups = 0;
1200 11 : all_tups_size = 0;
1201 :
1202 : /* chain to a new overflow page */
1203 11 : nbuf = _hash_addovflpage(rel, metabuf, nbuf, (nbuf == bucket_nbuf) ? true : false);
1204 11 : npage = BufferGetPage(nbuf);
1205 11 : nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
1206 : }
1207 :
1208 16530 : itups[nitups++] = new_itup;
1209 16530 : all_tups_size += itemsz;
1210 : }
1211 : else
1212 : {
1213 : /*
1214 : * the tuple stays on this page, so nothing to do.
1215 : */
1216 27338 : Assert(bucket == obucket);
1217 : }
1218 : }
1219 :
1220 123 : oblkno = oopaque->hasho_nextblkno;
1221 :
1222 : /* retain the pin on the old primary bucket */
1223 123 : if (obuf == bucket_obuf)
1224 72 : LockBuffer(obuf, BUFFER_LOCK_UNLOCK);
1225 : else
1226 51 : _hash_relbuf(rel, obuf);
1227 :
1228 : /* Exit loop if no more overflow pages in old bucket */
1229 123 : if (!BlockNumberIsValid(oblkno))
1230 : {
1231 : /*
1232 : * Change the shared buffer state in critical section, otherwise
1233 : * any error could make it unrecoverable.
1234 : */
1235 72 : START_CRIT_SECTION();
1236 :
1237 72 : _hash_pgaddmultitup(rel, nbuf, itups, itup_offsets, nitups);
1238 72 : MarkBufferDirty(nbuf);
1239 : /* log the split operation before releasing the lock */
1240 72 : log_split_page(rel, nbuf);
1241 :
1242 72 : END_CRIT_SECTION();
1243 :
1244 72 : if (nbuf == bucket_nbuf)
1245 71 : LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
1246 : else
1247 1 : _hash_relbuf(rel, nbuf);
1248 :
1249 : /* be tidy */
1250 11003 : for (i = 0; i < nitups; i++)
1251 10931 : pfree(itups[i]);
1252 72 : break;
1253 : }
1254 :
1255 : /* Else, advance to next old page */
1256 51 : obuf = _hash_getbuf(rel, oblkno, HASH_READ, LH_OVERFLOW_PAGE);
1257 51 : opage = BufferGetPage(obuf);
1258 51 : oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
1259 51 : }
1260 :
1261 : /*
1262 : * We're at the end of the old bucket chain, so we're done partitioning
1263 : * the tuples. Mark the old and new buckets to indicate split is
1264 : * finished.
1265 : *
1266 : * To avoid deadlocks due to locking order of buckets, first lock the old
1267 : * bucket and then the new bucket.
1268 : */
1269 72 : LockBuffer(bucket_obuf, BUFFER_LOCK_EXCLUSIVE);
1270 72 : opage = BufferGetPage(bucket_obuf);
1271 72 : oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
1272 :
1273 72 : LockBuffer(bucket_nbuf, BUFFER_LOCK_EXCLUSIVE);
1274 72 : npage = BufferGetPage(bucket_nbuf);
1275 72 : nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
1276 :
1277 72 : START_CRIT_SECTION();
1278 :
1279 72 : oopaque->hasho_flag &= ~LH_BUCKET_BEING_SPLIT;
1280 72 : nopaque->hasho_flag &= ~LH_BUCKET_BEING_POPULATED;
1281 :
1282 : /*
1283 : * After the split is finished, mark the old bucket to indicate that it
1284 : * contains deletable tuples. We will clear split-cleanup flag after
1285 : * deleting such tuples either at the end of split or at the next split
1286 : * from old bucket or at the time of vacuum.
1287 : */
1288 72 : oopaque->hasho_flag |= LH_BUCKET_NEEDS_SPLIT_CLEANUP;
1289 :
1290 : /*
1291 : * now write the buffers, here we don't release the locks as caller is
1292 : * responsible to release locks.
1293 : */
1294 72 : MarkBufferDirty(bucket_obuf);
1295 72 : MarkBufferDirty(bucket_nbuf);
1296 :
1297 72 : if (RelationNeedsWAL(rel))
1298 : {
1299 : XLogRecPtr recptr;
1300 : xl_hash_split_complete xlrec;
1301 :
1302 72 : xlrec.old_bucket_flag = oopaque->hasho_flag;
1303 72 : xlrec.new_bucket_flag = nopaque->hasho_flag;
1304 :
1305 72 : XLogBeginInsert();
1306 :
1307 72 : XLogRegisterData((char *) &xlrec, SizeOfHashSplitComplete);
1308 :
1309 72 : XLogRegisterBuffer(0, bucket_obuf, REGBUF_STANDARD);
1310 72 : XLogRegisterBuffer(1, bucket_nbuf, REGBUF_STANDARD);
1311 :
1312 72 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_COMPLETE);
1313 :
1314 72 : PageSetLSN(BufferGetPage(bucket_obuf), recptr);
1315 72 : PageSetLSN(BufferGetPage(bucket_nbuf), recptr);
1316 : }
1317 :
1318 72 : END_CRIT_SECTION();
1319 :
1320 : /*
1321 : * If possible, clean up the old bucket. We might not be able to do this
1322 : * if someone else has a pin on it, but if not then we can go ahead. This
1323 : * isn't absolutely necessary, but it reduces bloat; if we don't do it
1324 : * now, VACUUM will do it eventually, but maybe not until new overflow
1325 : * pages have been allocated. Note that there's no need to clean up the
1326 : * new bucket.
1327 : */
1328 72 : if (IsBufferCleanupOK(bucket_obuf))
1329 : {
1330 72 : LockBuffer(bucket_nbuf, BUFFER_LOCK_UNLOCK);
1331 72 : hashbucketcleanup(rel, obucket, bucket_obuf,
1332 : BufferGetBlockNumber(bucket_obuf), NULL,
1333 : maxbucket, highmask, lowmask, NULL, NULL, true,
1334 : NULL, NULL);
1335 : }
1336 : else
1337 : {
1338 0 : LockBuffer(bucket_nbuf, BUFFER_LOCK_UNLOCK);
1339 0 : LockBuffer(bucket_obuf, BUFFER_LOCK_UNLOCK);
1340 : }
1341 72 : }
1342 :
1343 : /*
1344 : * _hash_finish_split() -- Finish the previously interrupted split operation
1345 : *
1346 : * To complete the split operation, we form the hash table of TIDs in new
1347 : * bucket which is then used by split operation to skip tuples that are
1348 : * already moved before the split operation was previously interrupted.
1349 : *
1350 : * The caller must hold a pin, but no lock, on the metapage and old bucket's
1351 : * primary page buffer. The buffers are returned in the same state. (The
1352 : * metapage is only touched if it becomes necessary to add or remove overflow
1353 : * pages.)
1354 : */
1355 : void
1356 0 : _hash_finish_split(Relation rel, Buffer metabuf, Buffer obuf, Bucket obucket,
1357 : uint32 maxbucket, uint32 highmask, uint32 lowmask)
1358 : {
1359 : HASHCTL hash_ctl;
1360 : HTAB *tidhtab;
1361 0 : Buffer bucket_nbuf = InvalidBuffer;
1362 : Buffer nbuf;
1363 : Page npage;
1364 : BlockNumber nblkno;
1365 : BlockNumber bucket_nblkno;
1366 : HashPageOpaque npageopaque;
1367 : Bucket nbucket;
1368 : bool found;
1369 :
1370 : /* Initialize hash tables used to track TIDs */
1371 0 : memset(&hash_ctl, 0, sizeof(hash_ctl));
1372 0 : hash_ctl.keysize = sizeof(ItemPointerData);
1373 0 : hash_ctl.entrysize = sizeof(ItemPointerData);
1374 0 : hash_ctl.hcxt = CurrentMemoryContext;
1375 :
1376 0 : tidhtab =
1377 : hash_create("bucket ctids",
1378 : 256, /* arbitrary initial size */
1379 : &hash_ctl,
1380 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1381 :
1382 0 : bucket_nblkno = nblkno = _hash_get_newblock_from_oldbucket(rel, obucket);
1383 :
1384 : /*
1385 : * Scan the new bucket and build hash table of TIDs
1386 : */
1387 : for (;;)
1388 : {
1389 : OffsetNumber noffnum;
1390 : OffsetNumber nmaxoffnum;
1391 :
1392 0 : nbuf = _hash_getbuf(rel, nblkno, HASH_READ,
1393 : LH_BUCKET_PAGE | LH_OVERFLOW_PAGE);
1394 :
1395 : /* remember the primary bucket buffer to acquire cleanup lock on it. */
1396 0 : if (nblkno == bucket_nblkno)
1397 0 : bucket_nbuf = nbuf;
1398 :
1399 0 : npage = BufferGetPage(nbuf);
1400 0 : npageopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
1401 :
1402 : /* Scan each tuple in new page */
1403 0 : nmaxoffnum = PageGetMaxOffsetNumber(npage);
1404 0 : for (noffnum = FirstOffsetNumber;
1405 : noffnum <= nmaxoffnum;
1406 0 : noffnum = OffsetNumberNext(noffnum))
1407 : {
1408 : IndexTuple itup;
1409 :
1410 : /* Fetch the item's TID and insert it in hash table. */
1411 0 : itup = (IndexTuple) PageGetItem(npage,
1412 : PageGetItemId(npage, noffnum));
1413 :
1414 0 : (void) hash_search(tidhtab, &itup->t_tid, HASH_ENTER, &found);
1415 :
1416 0 : Assert(!found);
1417 : }
1418 :
1419 0 : nblkno = npageopaque->hasho_nextblkno;
1420 :
1421 : /*
1422 : * release our write lock without modifying buffer and ensure to
1423 : * retain the pin on primary bucket.
1424 : */
1425 0 : if (nbuf == bucket_nbuf)
1426 0 : LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
1427 : else
1428 0 : _hash_relbuf(rel, nbuf);
1429 :
1430 : /* Exit loop if no more overflow pages in new bucket */
1431 0 : if (!BlockNumberIsValid(nblkno))
1432 0 : break;
1433 0 : }
1434 :
1435 : /*
1436 : * Conditionally get the cleanup lock on old and new buckets to perform
1437 : * the split operation. If we don't get the cleanup locks, silently give
1438 : * up and next insertion on old bucket will try again to complete the
1439 : * split.
1440 : */
1441 0 : if (!ConditionalLockBufferForCleanup(obuf))
1442 : {
1443 0 : hash_destroy(tidhtab);
1444 0 : return;
1445 : }
1446 0 : if (!ConditionalLockBufferForCleanup(bucket_nbuf))
1447 : {
1448 0 : LockBuffer(obuf, BUFFER_LOCK_UNLOCK);
1449 0 : hash_destroy(tidhtab);
1450 0 : return;
1451 : }
1452 :
1453 0 : npage = BufferGetPage(bucket_nbuf);
1454 0 : npageopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
1455 0 : nbucket = npageopaque->hasho_bucket;
1456 :
1457 0 : _hash_splitbucket(rel, metabuf, obucket,
1458 : nbucket, obuf, bucket_nbuf, tidhtab,
1459 : maxbucket, highmask, lowmask);
1460 :
1461 0 : _hash_dropbuf(rel, bucket_nbuf);
1462 0 : hash_destroy(tidhtab);
1463 : }
1464 :
1465 : /*
1466 : * log_split_page() -- Log the split operation
1467 : *
1468 : * We log the split operation when the new page in new bucket gets full,
1469 : * so we log the entire page.
1470 : *
1471 : * 'buf' must be locked by the caller which is also responsible for unlocking
1472 : * it.
1473 : */
1474 : static void
1475 83 : log_split_page(Relation rel, Buffer buf)
1476 : {
1477 83 : if (RelationNeedsWAL(rel))
1478 : {
1479 : XLogRecPtr recptr;
1480 :
1481 83 : XLogBeginInsert();
1482 :
1483 83 : XLogRegisterBuffer(0, buf, REGBUF_FORCE_IMAGE | REGBUF_STANDARD);
1484 :
1485 83 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_PAGE);
1486 :
1487 83 : PageSetLSN(BufferGetPage(buf), recptr);
1488 : }
1489 83 : }
1490 :
1491 : /*
1492 : * _hash_getcachedmetap() -- Returns cached metapage data.
1493 : *
1494 : * If metabuf is not InvalidBuffer, caller must hold a pin, but no lock, on
1495 : * the metapage. If not set, we'll set it before returning if we have to
1496 : * refresh the cache, and return with a pin but no lock on it; caller is
1497 : * responsible for releasing the pin.
1498 : *
1499 : * We refresh the cache if it's not initialized yet or force_refresh is true.
1500 : */
1501 : HashMetaPage
1502 80642 : _hash_getcachedmetap(Relation rel, Buffer *metabuf, bool force_refresh)
1503 : {
1504 : Page page;
1505 :
1506 80642 : Assert(metabuf);
1507 80642 : if (force_refresh || rel->rd_amcache == NULL)
1508 : {
1509 85 : char *cache = NULL;
1510 :
1511 : /*
1512 : * It's important that we don't set rd_amcache to an invalid value.
1513 : * Either MemoryContextAlloc or _hash_getbuf could fail, so don't
1514 : * install a pointer to the newly-allocated storage in the actual
1515 : * relcache entry until both have succeeeded.
1516 : */
1517 85 : if (rel->rd_amcache == NULL)
1518 20 : cache = MemoryContextAlloc(rel->rd_indexcxt,
1519 : sizeof(HashMetaPageData));
1520 :
1521 : /* Read the metapage. */
1522 85 : if (BufferIsValid(*metabuf))
1523 0 : LockBuffer(*metabuf, BUFFER_LOCK_SHARE);
1524 : else
1525 85 : *metabuf = _hash_getbuf(rel, HASH_METAPAGE, HASH_READ,
1526 : LH_META_PAGE);
1527 85 : page = BufferGetPage(*metabuf);
1528 :
1529 : /* Populate the cache. */
1530 85 : if (rel->rd_amcache == NULL)
1531 20 : rel->rd_amcache = cache;
1532 85 : memcpy(rel->rd_amcache, HashPageGetMeta(page),
1533 : sizeof(HashMetaPageData));
1534 :
1535 : /* Release metapage lock, but keep the pin. */
1536 85 : LockBuffer(*metabuf, BUFFER_LOCK_UNLOCK);
1537 : }
1538 :
1539 80642 : return (HashMetaPage) rel->rd_amcache;
1540 : }
1541 :
1542 : /*
1543 : * _hash_getbucketbuf_from_hashkey() -- Get the bucket's buffer for the given
1544 : * hashkey.
1545 : *
1546 : * Bucket pages do not move or get removed once they are allocated. This give
1547 : * us an opportunity to use the previously saved metapage contents to reach
1548 : * the target bucket buffer, instead of reading from the metapage every time.
1549 : * This saves one buffer access every time we want to reach the target bucket
1550 : * buffer, which is very helpful savings in bufmgr traffic and contention.
1551 : *
1552 : * The access type parameter (HASH_READ or HASH_WRITE) indicates whether the
1553 : * bucket buffer has to be locked for reading or writing.
1554 : *
1555 : * The out parameter cachedmetap is set with metapage contents used for
1556 : * hashkey to bucket buffer mapping. Some callers need this info to reach the
1557 : * old bucket in case of bucket split, see _hash_doinsert().
1558 : */
1559 : Buffer
1560 80577 : _hash_getbucketbuf_from_hashkey(Relation rel, uint32 hashkey, int access,
1561 : HashMetaPage *cachedmetap)
1562 : {
1563 : HashMetaPage metap;
1564 : Buffer buf;
1565 80577 : Buffer metabuf = InvalidBuffer;
1566 : Page page;
1567 : Bucket bucket;
1568 : BlockNumber blkno;
1569 : HashPageOpaque opaque;
1570 :
1571 : /* We read from target bucket buffer, hence locking is must. */
1572 80577 : Assert(access == HASH_READ || access == HASH_WRITE);
1573 :
1574 80577 : metap = _hash_getcachedmetap(rel, &metabuf, false);
1575 80577 : Assert(metap != NULL);
1576 :
1577 : /*
1578 : * Loop until we get a lock on the correct target bucket.
1579 : */
1580 : for (;;)
1581 : {
1582 : /*
1583 : * Compute the target bucket number, and convert to block number.
1584 : */
1585 80642 : bucket = _hash_hashkey2bucket(hashkey,
1586 : metap->hashm_maxbucket,
1587 : metap->hashm_highmask,
1588 : metap->hashm_lowmask);
1589 :
1590 80642 : blkno = BUCKET_TO_BLKNO(metap, bucket);
1591 :
1592 : /* Fetch the primary bucket page for the bucket */
1593 80642 : buf = _hash_getbuf(rel, blkno, access, LH_BUCKET_PAGE);
1594 80642 : page = BufferGetPage(buf);
1595 80642 : opaque = (HashPageOpaque) PageGetSpecialPointer(page);
1596 80642 : Assert(opaque->hasho_bucket == bucket);
1597 80642 : Assert(opaque->hasho_prevblkno != InvalidBlockNumber);
1598 :
1599 : /*
1600 : * If this bucket hasn't been split, we're done.
1601 : */
1602 80642 : if (opaque->hasho_prevblkno <= metap->hashm_maxbucket)
1603 80577 : break;
1604 :
1605 : /* Drop lock on this buffer, update cached metapage, and retry. */
1606 65 : _hash_relbuf(rel, buf);
1607 65 : metap = _hash_getcachedmetap(rel, &metabuf, true);
1608 65 : Assert(metap != NULL);
1609 65 : }
1610 :
1611 80577 : if (BufferIsValid(metabuf))
1612 85 : _hash_dropbuf(rel, metabuf);
1613 :
1614 80577 : if (cachedmetap)
1615 80552 : *cachedmetap = metap;
1616 :
1617 80577 : return buf;
1618 : }
|
