Age Owner Branch data TLA 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/injection_point.h"
39 : : #include "utils/rel.h"
40 : :
41 : : static bool _hash_alloc_buckets(Relation rel, BlockNumber firstblock,
42 : : uint32 nblocks);
43 : : static void _hash_splitbucket(Relation rel, Buffer metabuf,
44 : : Bucket obucket, Bucket nbucket,
45 : : Buffer obuf,
46 : : Buffer nbuf,
47 : : HTAB *htab,
48 : : uint32 maxbucket,
49 : : uint32 highmask, uint32 lowmask);
50 : : static void log_split_page(Relation rel, Buffer buf);
51 : :
52 : :
53 : : /*
54 : : * _hash_getbuf() -- Get a buffer by block number for read or write.
55 : : *
56 : : * 'access' must be HASH_READ, HASH_WRITE, or HASH_NOLOCK.
57 : : * 'flags' is a bitwise OR of the allowed page types.
58 : : *
59 : : * This must be used only to fetch pages that are expected to be valid
60 : : * already. _hash_checkpage() is applied using the given flags.
61 : : *
62 : : * When this routine returns, the appropriate lock is set on the
63 : : * requested buffer and its reference count has been incremented
64 : : * (ie, the buffer is "locked and pinned").
65 : : *
66 : : * P_NEW is disallowed because this routine can only be used
67 : : * to access pages that are known to be before the filesystem EOF.
68 : : * Extending the index should be done with _hash_getnewbuf.
69 : : */
70 : : Buffer
7102 tgl@sss.pgh.pa.us 71 :CBC 1342025 : _hash_getbuf(Relation rel, BlockNumber blkno, int access, int flags)
72 : : {
73 : : Buffer buf;
74 : :
7116 75 [ - + ]: 1342025 : if (blkno == P_NEW)
7116 tgl@sss.pgh.pa.us 76 [ # # ]:UBC 0 : elog(ERROR, "hash AM does not use P_NEW");
77 : :
8439 tgl@sss.pgh.pa.us 78 :CBC 1342025 : buf = ReadBuffer(rel, blkno);
79 : :
80 [ + + ]: 1342025 : if (access != HASH_NOLOCK)
81 : 850255 : LockBuffer(buf, access);
82 : :
83 : : /* ref count and lock type are correct */
84 : :
7102 85 : 1342025 : _hash_checkpage(rel, buf, flags);
86 : :
87 : 1342025 : return buf;
88 : : }
89 : :
90 : : /*
91 : : * _hash_getbuf_with_condlock_cleanup() -- Try to get a buffer for cleanup.
92 : : *
93 : : * We read the page and try to acquire a cleanup lock. If we get it,
94 : : * we return the buffer; otherwise, we return InvalidBuffer.
95 : : */
96 : : Buffer
3603 rhaas@postgresql.org 97 : 892 : _hash_getbuf_with_condlock_cleanup(Relation rel, BlockNumber blkno, int flags)
98 : : {
99 : : Buffer buf;
100 : :
101 [ - + ]: 892 : if (blkno == P_NEW)
3603 rhaas@postgresql.org 102 [ # # ]:UBC 0 : elog(ERROR, "hash AM does not use P_NEW");
103 : :
3603 rhaas@postgresql.org 104 :CBC 892 : buf = ReadBuffer(rel, blkno);
105 : :
106 [ - + ]: 892 : if (!ConditionalLockBufferForCleanup(buf))
107 : : {
3603 rhaas@postgresql.org 108 :UBC 0 : ReleaseBuffer(buf);
109 : 0 : return InvalidBuffer;
110 : : }
111 : :
112 : : /* ref count and lock type are correct */
113 : :
3603 rhaas@postgresql.org 114 :CBC 892 : _hash_checkpage(rel, buf, flags);
115 : :
116 : 892 : return buf;
117 : : }
118 : :
119 : : /*
120 : : * _hash_getinitbuf() -- Get and initialize a buffer by block number.
121 : : *
122 : : * This must be used only to fetch pages that are known to be before
123 : : * the index's filesystem EOF, but are to be filled from scratch.
124 : : * _hash_pageinit() is applied automatically. Otherwise it has
125 : : * effects similar to _hash_getbuf() with access = HASH_WRITE.
126 : : *
127 : : * When this routine returns, a write lock is set on the
128 : : * requested buffer and its reference count has been incremented
129 : : * (ie, the buffer is "locked and pinned").
130 : : *
131 : : * P_NEW is disallowed because this routine can only be used
132 : : * to access pages that are known to be before the filesystem EOF.
133 : : * Extending the index should be done with _hash_getnewbuf.
134 : : */
135 : : Buffer
7102 tgl@sss.pgh.pa.us 136 : 46 : _hash_getinitbuf(Relation rel, BlockNumber blkno)
137 : : {
138 : : Buffer buf;
139 : :
140 [ - + ]: 46 : if (blkno == P_NEW)
7102 tgl@sss.pgh.pa.us 141 [ # # ]:UBC 0 : elog(ERROR, "hash AM does not use P_NEW");
142 : :
4351 heikki.linnakangas@i 143 :CBC 46 : buf = ReadBufferExtended(rel, MAIN_FORKNUM, blkno, RBM_ZERO_AND_LOCK,
144 : : NULL);
145 : :
146 : : /* ref count and lock type are correct */
147 : :
148 : : /* initialize the page */
3827 kgrittn@postgresql.o 149 : 46 : _hash_pageinit(BufferGetPage(buf), BufferGetPageSize(buf));
150 : :
8439 tgl@sss.pgh.pa.us 151 : 46 : return buf;
152 : : }
153 : :
154 : : /*
155 : : * _hash_initbuf() -- Get and initialize a buffer by bucket number.
156 : : */
157 : : void
3506 rhaas@postgresql.org 158 : 5589 : _hash_initbuf(Buffer buf, uint32 max_bucket, uint32 num_bucket, uint32 flag,
159 : : bool initpage)
160 : : {
161 : : HashPageOpaque pageopaque;
162 : : Page page;
163 : :
164 : 5589 : page = BufferGetPage(buf);
165 : :
166 : : /* initialize the page */
167 [ + + ]: 5589 : if (initpage)
168 : 291 : _hash_pageinit(page, BufferGetPageSize(buf));
169 : :
1655 michael@paquier.xyz 170 : 5589 : pageopaque = HashPageGetOpaque(page);
171 : :
172 : : /*
173 : : * Set hasho_prevblkno with current hashm_maxbucket. This value will be
174 : : * used to validate cached HashMetaPageData. See
175 : : * _hash_getbucketbuf_from_hashkey().
176 : : */
3506 rhaas@postgresql.org 177 : 5589 : pageopaque->hasho_prevblkno = max_bucket;
178 : 5589 : pageopaque->hasho_nextblkno = InvalidBlockNumber;
179 : 5589 : pageopaque->hasho_bucket = num_bucket;
180 : 5589 : pageopaque->hasho_flag = flag;
181 : 5589 : pageopaque->hasho_page_id = HASHO_PAGE_ID;
182 : 5589 : }
183 : :
184 : : /*
185 : : * _hash_getnewbuf() -- Get a new page at the end of the index.
186 : : *
187 : : * This has the same API as _hash_getinitbuf, except that we are adding
188 : : * a page to the index, and hence expect the page to be past the
189 : : * logical EOF. (However, we have to support the case where it isn't,
190 : : * since a prior try might have crashed after extending the filesystem
191 : : * EOF but before updating the metapage to reflect the added page.)
192 : : *
193 : : * It is caller's responsibility to ensure that only one process can
194 : : * extend the index at a time. In practice, this function is called
195 : : * only while holding write lock on the metapage, because adding a page
196 : : * is always associated with an update of metapage data.
197 : : */
198 : : Buffer
5766 199 : 6856 : _hash_getnewbuf(Relation rel, BlockNumber blkno, ForkNumber forkNum)
200 : : {
201 : 6856 : BlockNumber nblocks = RelationGetNumberOfBlocksInFork(rel, forkNum);
202 : : Buffer buf;
203 : :
7116 tgl@sss.pgh.pa.us 204 [ - + ]: 6856 : if (blkno == P_NEW)
7116 tgl@sss.pgh.pa.us 205 [ # # ]:UBC 0 : elog(ERROR, "hash AM does not use P_NEW");
7116 tgl@sss.pgh.pa.us 206 [ - + ]:CBC 6856 : if (blkno > nblocks)
7116 tgl@sss.pgh.pa.us 207 [ # # ]:UBC 0 : elog(ERROR, "access to noncontiguous page in hash index \"%s\"",
208 : : RelationGetRelationName(rel));
209 : :
210 : : /* smgr insists we explicitly extend the relation */
7116 tgl@sss.pgh.pa.us 211 [ + + ]:CBC 6856 : if (blkno == nblocks)
212 : : {
1146 tmunro@postgresql.or 213 : 5964 : buf = ExtendBufferedRel(BMR_REL(rel), forkNum, NULL,
214 : : EB_LOCK_FIRST | EB_SKIP_EXTENSION_LOCK);
7116 tgl@sss.pgh.pa.us 215 [ - + ]: 5964 : if (BufferGetBlockNumber(buf) != blkno)
7116 tgl@sss.pgh.pa.us 216 [ # # ]:UBC 0 : elog(ERROR, "unexpected hash relation size: %u, should be %u",
217 : : BufferGetBlockNumber(buf), blkno);
218 : : }
219 : : else
220 : : {
4351 heikki.linnakangas@i 221 :CBC 892 : buf = ReadBufferExtended(rel, forkNum, blkno, RBM_ZERO_AND_LOCK,
222 : : NULL);
223 : : }
224 : :
225 : : /* ref count and lock type are correct */
226 : :
227 : : /* initialize the page */
3827 kgrittn@postgresql.o 228 : 6856 : _hash_pageinit(BufferGetPage(buf), BufferGetPageSize(buf));
229 : :
7116 tgl@sss.pgh.pa.us 230 : 6856 : return buf;
231 : : }
232 : :
233 : : /*
234 : : * _hash_getbuf_with_strategy() -- Get a buffer with nondefault strategy.
235 : : *
236 : : * This is identical to _hash_getbuf() but also allows a buffer access
237 : : * strategy to be specified. We use this for VACUUM operations.
238 : : */
239 : : Buffer
7075 240 : 798 : _hash_getbuf_with_strategy(Relation rel, BlockNumber blkno,
241 : : int access, int flags,
242 : : BufferAccessStrategy bstrategy)
243 : : {
244 : : Buffer buf;
245 : :
246 [ - + ]: 798 : if (blkno == P_NEW)
7075 tgl@sss.pgh.pa.us 247 [ # # ]:UBC 0 : elog(ERROR, "hash AM does not use P_NEW");
248 : :
6555 heikki.linnakangas@i 249 :CBC 798 : buf = ReadBufferExtended(rel, MAIN_FORKNUM, blkno, RBM_NORMAL, bstrategy);
250 : :
7075 tgl@sss.pgh.pa.us 251 [ + - ]: 798 : if (access != HASH_NOLOCK)
252 : 798 : LockBuffer(buf, access);
253 : :
254 : : /* ref count and lock type are correct */
255 : :
256 : 798 : _hash_checkpage(rel, buf, flags);
257 : :
258 : 798 : return buf;
259 : : }
260 : :
261 : : /*
262 : : * _hash_relbuf() -- release a locked buffer.
263 : : *
264 : : * Lock and pin (refcount) are both dropped.
265 : : */
266 : : void
8439 267 : 788929 : _hash_relbuf(Relation rel, Buffer buf)
268 : : {
7500 269 : 788929 : UnlockReleaseBuffer(buf);
8439 270 : 788929 : }
271 : :
272 : : /*
273 : : * _hash_dropbuf() -- release an unlocked buffer.
274 : : *
275 : : * This is used to unpin a buffer on which we hold no lock.
276 : : */
277 : : void
278 : 562075 : _hash_dropbuf(Relation rel, Buffer buf)
279 : : {
280 : 562075 : ReleaseBuffer(buf);
281 : 562075 : }
282 : :
283 : : /*
284 : : * _hash_dropscanbuf() -- release buffers used in scan.
285 : : *
286 : : * This routine unpins the buffers used during scan on which we
287 : : * hold no lock.
288 : : */
289 : : void
3603 rhaas@postgresql.org 290 : 842 : _hash_dropscanbuf(Relation rel, HashScanOpaque so)
291 : : {
292 : : /* release pin we hold on primary bucket page */
293 [ + + ]: 842 : if (BufferIsValid(so->hashso_bucket_buf) &&
3307 294 [ + + ]: 345 : so->hashso_bucket_buf != so->currPos.buf)
3603 295 : 95 : _hash_dropbuf(rel, so->hashso_bucket_buf);
296 : 842 : so->hashso_bucket_buf = InvalidBuffer;
297 : :
298 : : /* release pin we hold on primary bucket page of bucket being split */
299 [ + + ]: 842 : if (BufferIsValid(so->hashso_split_bucket_buf) &&
3307 300 [ + + ]: 3 : so->hashso_split_bucket_buf != so->currPos.buf)
3603 301 : 1 : _hash_dropbuf(rel, so->hashso_split_bucket_buf);
302 : 842 : so->hashso_split_bucket_buf = InvalidBuffer;
303 : :
304 : : /* release any pin we still hold */
3307 305 [ + + ]: 842 : if (BufferIsValid(so->currPos.buf))
306 : 252 : _hash_dropbuf(rel, so->currPos.buf);
307 : 842 : so->currPos.buf = InvalidBuffer;
308 : :
309 : : /* reset split scan */
3603 310 : 842 : so->hashso_buc_populated = false;
311 : 842 : so->hashso_buc_split = false;
312 : 842 : }
313 : :
314 : :
315 : : /*
316 : : * _hash_init() -- Initialize the metadata page of a hash index,
317 : : * the initial buckets, and the initial bitmap page.
318 : : *
319 : : * The initial number of buckets is dependent on num_tuples, an estimate
320 : : * of the number of tuples to be loaded into the index initially. The
321 : : * chosen number of buckets is returned.
322 : : *
323 : : * We are fairly cavalier about locking here, since we know that no one else
324 : : * could be accessing this index. In particular the rule about not holding
325 : : * multiple buffer locks is ignored.
326 : : */
327 : : uint32
3506 328 : 216 : _hash_init(Relation rel, double num_tuples, ForkNumber forkNum)
329 : : {
330 : : Buffer metabuf;
331 : : Buffer buf;
332 : : Buffer bitmapbuf;
333 : : Page pg;
334 : : HashMetaPage metap;
335 : : RegProcedure procid;
336 : : int32 data_width;
337 : : int32 item_width;
338 : : int32 ffactor;
339 : : uint32 num_buckets;
340 : : uint32 i;
341 : : bool use_wal;
342 : :
343 : : /* safety check */
5766 344 [ - + ]: 216 : if (RelationGetNumberOfBlocksInFork(rel, forkNum) != 0)
8484 tgl@sss.pgh.pa.us 345 [ # # ]:UBC 0 : elog(ERROR, "cannot initialize non-empty hash index \"%s\"",
346 : : RelationGetRelationName(rel));
347 : :
348 : : /*
349 : : * WAL log creation of pages if the relation is persistent, or this is the
350 : : * init fork. Init forks for unlogged relations always need to be WAL
351 : : * logged.
352 : : */
3374 rhaas@postgresql.org 353 [ + + + + :CBC 216 : use_wal = RelationNeedsWAL(rel) || forkNum == INIT_FORKNUM;
- + - - +
+ ]
354 : :
355 : : /*
356 : : * Determine the target fill factor (in tuples per bucket) for this index.
357 : : * The idea is to make the fill factor correspond to pages about as full
358 : : * as the user-settable fillfactor parameter says. We can compute it
359 : : * exactly since the index datatype (i.e. uint32 hash key) is fixed-width.
360 : : */
6601 tgl@sss.pgh.pa.us 361 : 216 : data_width = sizeof(uint32);
7565 362 : 216 : item_width = MAXALIGN(sizeof(IndexTupleData)) + MAXALIGN(data_width) +
363 : : sizeof(ItemIdData); /* include the line pointer */
2513 michael@paquier.xyz 364 [ + - - + : 216 : ffactor = HashGetTargetPageUsage(rel) / item_width;
+ + ]
365 : : /* keep to a sane range */
8439 tgl@sss.pgh.pa.us 366 [ - + ]: 216 : if (ffactor < 10)
8439 tgl@sss.pgh.pa.us 367 :UBC 0 : ffactor = 10;
368 : :
3329 rhaas@postgresql.org 369 :CBC 216 : procid = index_getprocid(rel, 1, HASHSTANDARD_PROC);
370 : :
371 : : /*
372 : : * We initialize the metapage, the first N bucket pages, and the first
373 : : * bitmap page in sequence, using _hash_getnewbuf to cause smgrextend()
374 : : * calls to occur. This ensures that the smgr level has the right idea of
375 : : * the physical index length.
376 : : *
377 : : * Critical section not required, because on error the creation of the
378 : : * whole relation will be rolled back.
379 : : */
3506 380 : 216 : metabuf = _hash_getnewbuf(rel, HASH_METAPAGE, forkNum);
381 : 216 : _hash_init_metabuffer(metabuf, num_tuples, procid, ffactor, false);
382 : 216 : MarkBufferDirty(metabuf);
383 : :
384 : 216 : pg = BufferGetPage(metabuf);
385 : 216 : metap = HashPageGetMeta(pg);
386 : :
387 : : /* XLOG stuff */
3374 388 [ + + ]: 216 : if (use_wal)
389 : : {
390 : : xl_hash_init_meta_page xlrec;
391 : : XLogRecPtr recptr;
392 : :
3499 393 : 170 : xlrec.num_tuples = num_tuples;
394 : 170 : xlrec.procid = metap->hashm_procid;
395 : 170 : xlrec.ffactor = metap->hashm_ffactor;
396 : :
397 : 170 : XLogBeginInsert();
608 peter@eisentraut.org 398 : 170 : XLogRegisterData(&xlrec, SizeOfHashInitMetaPage);
3265 tgl@sss.pgh.pa.us 399 : 170 : XLogRegisterBuffer(0, metabuf, REGBUF_WILL_INIT | REGBUF_STANDARD);
400 : :
3499 rhaas@postgresql.org 401 : 170 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_INIT_META_PAGE);
402 : :
403 : 170 : PageSetLSN(BufferGetPage(metabuf), recptr);
404 : : }
405 : :
3506 406 : 216 : num_buckets = metap->hashm_maxbucket + 1;
407 : :
408 : : /*
409 : : * Release buffer lock on the metapage while we initialize buckets.
410 : : * Otherwise, we'll be in interrupt holdoff and the CHECK_FOR_INTERRUPTS
411 : : * won't accomplish anything. It's a bad idea to hold buffer locks for
412 : : * long intervals in any case, since that can block the bgwriter.
413 : : */
414 : 216 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
415 : :
416 : : /*
417 : : * Initialize and WAL Log the first N buckets
418 : : */
419 [ + + ]: 5514 : for (i = 0; i < num_buckets; i++)
420 : : {
421 : : BlockNumber blkno;
422 : :
423 : : /* Allow interrupts, in case N is huge */
424 [ - + ]: 5298 : CHECK_FOR_INTERRUPTS();
425 : :
426 [ + + ]: 5298 : blkno = BUCKET_TO_BLKNO(metap, i);
427 : 5298 : buf = _hash_getnewbuf(rel, blkno, forkNum);
428 : 5298 : _hash_initbuf(buf, metap->hashm_maxbucket, i, LH_BUCKET_PAGE, false);
429 : 5298 : MarkBufferDirty(buf);
430 : :
3374 431 [ + + ]: 5298 : if (use_wal)
1559 432 : 5152 : log_newpage(&rel->rd_locator,
433 : : forkNum,
434 : : blkno,
435 : : BufferGetPage(buf),
436 : : true);
3506 437 : 5298 : _hash_relbuf(rel, buf);
438 : : }
439 : :
440 : : /* Now reacquire buffer lock on metapage */
441 : 216 : LockBuffer(metabuf, BUFFER_LOCK_EXCLUSIVE);
442 : :
443 : : /*
444 : : * Initialize bitmap page
445 : : */
446 : 216 : bitmapbuf = _hash_getnewbuf(rel, num_buckets + 1, forkNum);
447 : 216 : _hash_initbitmapbuffer(bitmapbuf, metap->hashm_bmsize, false);
448 : 216 : MarkBufferDirty(bitmapbuf);
449 : :
450 : : /* add the new bitmap page to the metapage's list of bitmaps */
451 : : /* metapage already has a write lock */
452 [ - + ]: 216 : if (metap->hashm_nmaps >= HASH_MAX_BITMAPS)
3506 rhaas@postgresql.org 453 [ # # ]:UBC 0 : ereport(ERROR,
454 : : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
455 : : errmsg("out of overflow pages in hash index \"%s\"",
456 : : RelationGetRelationName(rel))));
457 : :
3506 rhaas@postgresql.org 458 :CBC 216 : metap->hashm_mapp[metap->hashm_nmaps] = num_buckets + 1;
459 : :
460 : 216 : metap->hashm_nmaps++;
461 : 216 : MarkBufferDirty(metabuf);
462 : :
463 : : /* XLOG stuff */
3374 464 [ + + ]: 216 : if (use_wal)
465 : : {
466 : : xl_hash_init_bitmap_page xlrec;
467 : : XLogRecPtr recptr;
468 : :
3499 469 : 170 : xlrec.bmsize = metap->hashm_bmsize;
470 : :
471 : 170 : XLogBeginInsert();
608 peter@eisentraut.org 472 : 170 : XLogRegisterData(&xlrec, SizeOfHashInitBitmapPage);
3499 rhaas@postgresql.org 473 : 170 : XLogRegisterBuffer(0, bitmapbuf, REGBUF_WILL_INIT);
474 : :
475 : : /*
476 : : * This is safe only because nobody else can be modifying the index at
477 : : * this stage; it's only visible to the transaction that is creating
478 : : * it.
479 : : */
480 : 170 : XLogRegisterBuffer(1, metabuf, REGBUF_STANDARD);
481 : :
482 : 170 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_INIT_BITMAP_PAGE);
483 : :
484 : 170 : PageSetLSN(BufferGetPage(bitmapbuf), recptr);
485 : 170 : PageSetLSN(BufferGetPage(metabuf), recptr);
486 : : }
487 : :
488 : : /* all done */
3506 489 : 216 : _hash_relbuf(rel, bitmapbuf);
490 : 216 : _hash_relbuf(rel, metabuf);
491 : :
492 : 216 : return num_buckets;
493 : : }
494 : :
495 : : /*
496 : : * _hash_init_metabuffer() -- Initialize the metadata page of a hash index.
497 : : */
498 : : void
499 : 246 : _hash_init_metabuffer(Buffer buf, double num_tuples, RegProcedure procid,
500 : : uint16 ffactor, bool initpage)
501 : : {
502 : : HashMetaPage metap;
503 : : HashPageOpaque pageopaque;
504 : : Page page;
505 : : double dnumbuckets;
506 : : uint32 num_buckets;
507 : : uint32 spare_index;
508 : : uint32 lshift;
509 : :
510 : : /*
511 : : * Choose the number of initial bucket pages to match the fill factor
512 : : * given the estimated number of tuples. We round up the result to the
513 : : * total number of buckets which has to be allocated before using its
514 : : * hashm_spares element. However always force at least 2 bucket pages. The
515 : : * upper limit is determined by considerations explained in
516 : : * _hash_expandtable().
517 : : */
6785 tgl@sss.pgh.pa.us 518 : 246 : dnumbuckets = num_tuples / ffactor;
519 [ + + ]: 246 : if (dnumbuckets <= 2.0)
520 : 72 : num_buckets = 2;
521 [ - + ]: 174 : else if (dnumbuckets >= (double) 0x40000000)
6785 tgl@sss.pgh.pa.us 522 :UBC 0 : num_buckets = 0x40000000;
523 : : else
3479 rhaas@postgresql.org 524 :CBC 174 : num_buckets = _hash_get_totalbuckets(_hash_spareindex(dnumbuckets));
525 : :
526 : 246 : spare_index = _hash_spareindex(num_buckets);
527 [ - + ]: 246 : Assert(spare_index < HASH_MAX_SPLITPOINTS);
528 : :
3506 529 : 246 : page = BufferGetPage(buf);
530 [ + + ]: 246 : if (initpage)
531 : 30 : _hash_pageinit(page, BufferGetPageSize(buf));
532 : :
1655 michael@paquier.xyz 533 : 246 : pageopaque = HashPageGetOpaque(page);
8442 tgl@sss.pgh.pa.us 534 : 246 : pageopaque->hasho_prevblkno = InvalidBlockNumber;
535 : 246 : pageopaque->hasho_nextblkno = InvalidBlockNumber;
1928 peter@eisentraut.org 536 : 246 : pageopaque->hasho_bucket = InvalidBucket;
8441 tgl@sss.pgh.pa.us 537 : 246 : pageopaque->hasho_flag = LH_META_PAGE;
7126 538 : 246 : pageopaque->hasho_page_id = HASHO_PAGE_ID;
539 : :
3506 rhaas@postgresql.org 540 : 246 : metap = HashPageGetMeta(page);
541 : :
10627 bruce@momjian.us 542 : 246 : metap->hashm_magic = HASH_MAGIC;
543 : 246 : metap->hashm_version = HASH_VERSION;
8442 tgl@sss.pgh.pa.us 544 : 246 : metap->hashm_ntuples = 0;
10627 bruce@momjian.us 545 : 246 : metap->hashm_nmaps = 0;
8439 tgl@sss.pgh.pa.us 546 : 246 : metap->hashm_ffactor = ffactor;
3506 rhaas@postgresql.org 547 : 246 : metap->hashm_bsize = HashGetMaxBitmapSize(page);
548 : :
549 : : /* find largest bitmap array size that will fit in page size */
2378 drowley@postgresql.o 550 : 246 : lshift = pg_leftmost_one_pos32(metap->hashm_bsize);
551 [ - + ]: 246 : Assert(lshift > 0);
552 : 246 : metap->hashm_bmsize = 1 << lshift;
553 : 246 : metap->hashm_bmshift = lshift + BYTE_TO_BIT;
8442 tgl@sss.pgh.pa.us 554 [ - + ]: 246 : Assert((1 << BMPG_SHIFT(metap)) == (BMPG_MASK(metap) + 1));
555 : :
556 : : /*
557 : : * Label the index with its primary hash support function's OID. This is
558 : : * pretty useless for normal operation (in fact, hashm_procid is not used
559 : : * anywhere), but it might be handy for forensic purposes so we keep it.
560 : : */
3506 rhaas@postgresql.org 561 : 246 : metap->hashm_procid = procid;
562 : :
563 : : /*
564 : : * We initialize the index with N buckets, 0 .. N-1, occupying physical
565 : : * blocks 1 to N. The first freespace bitmap page is in block N+1.
566 : : */
3479 567 : 246 : metap->hashm_maxbucket = num_buckets - 1;
568 : :
569 : : /*
570 : : * Set highmask as next immediate ((2 ^ x) - 1), which should be
571 : : * sufficient to cover num_buckets.
572 : : */
2378 drowley@postgresql.o 573 : 246 : metap->hashm_highmask = pg_nextpower2_32(num_buckets + 1) - 1;
3479 rhaas@postgresql.org 574 : 246 : metap->hashm_lowmask = (metap->hashm_highmask >> 1);
575 : :
7824 neilc@samurai.com 576 [ - + - - : 246 : MemSet(metap->hashm_spares, 0, sizeof(metap->hashm_spares));
- - - - -
- ]
577 [ - + - - : 246 : MemSet(metap->hashm_mapp, 0, sizeof(metap->hashm_mapp));
- - - - -
- ]
578 : :
579 : : /* Set up mapping for one spare page after the initial splitpoints */
3479 rhaas@postgresql.org 580 : 246 : metap->hashm_spares[spare_index] = 1;
581 : 246 : metap->hashm_ovflpoint = spare_index;
8442 tgl@sss.pgh.pa.us 582 : 246 : metap->hashm_firstfree = 0;
583 : :
584 : : /*
585 : : * Set pd_lower just past the end of the metadata. This is essential,
586 : : * because without doing so, metadata will be lost if xlog.c compresses
587 : : * the page.
588 : : */
3506 rhaas@postgresql.org 589 : 246 : ((PageHeader) page)->pd_lower =
590 : 246 : ((char *) metap + sizeof(HashMetaPageData)) - (char *) page;
11052 scrappy@hub.org 591 : 246 : }
592 : :
593 : : /*
594 : : * _hash_pageinit() -- Initialize a new hash index page.
595 : : */
596 : : void
597 : 7418 : _hash_pageinit(Page page, Size size)
598 : : {
10627 bruce@momjian.us 599 : 7418 : PageInit(page, size, sizeof(HashPageOpaqueData));
11052 scrappy@hub.org 600 : 7418 : }
601 : :
602 : : /*
603 : : * Attempt to expand the hash table by creating one new bucket.
604 : : *
605 : : * This will silently do nothing if we don't get cleanup lock on old or
606 : : * new bucket.
607 : : *
608 : : * Complete the pending splits and remove the tuples from old bucket,
609 : : * if there are any left over from the previous split.
610 : : *
611 : : * The caller must hold a pin, but no lock, on the metapage buffer.
612 : : * The buffer is returned in the same state.
613 : : */
614 : : void
615 : 892 : _hash_expandtable(Relation rel, Buffer metabuf)
616 : : {
617 : : HashMetaPage metap;
618 : : Bucket old_bucket;
619 : : Bucket new_bucket;
620 : : uint32 spare_ndx;
621 : : BlockNumber start_oblkno;
622 : : BlockNumber start_nblkno;
623 : : Buffer buf_nblkno;
624 : : Buffer buf_oblkno;
625 : : Page opage;
626 : : Page npage;
627 : : HashPageOpaque oopaque;
628 : : HashPageOpaque nopaque;
629 : : uint32 maxbucket;
630 : : uint32 highmask;
631 : : uint32 lowmask;
3499 rhaas@postgresql.org 632 : 892 : bool metap_update_masks = false;
633 : 892 : bool metap_update_splitpoint = false;
634 : : XLogRecPtr recptr;
635 : :
3603 rhaas@postgresql.org 636 :UBC 0 : restart_expand:
637 : :
638 : : /*
639 : : * Write-lock the meta page. It used to be necessary to acquire a
640 : : * heavyweight lock to begin a split, but that is no longer required.
641 : : */
3580 rhaas@postgresql.org 642 :CBC 892 : LockBuffer(metabuf, BUFFER_LOCK_EXCLUSIVE);
643 : :
7645 tgl@sss.pgh.pa.us 644 : 892 : _hash_checkpage(rel, metabuf, LH_META_PAGE);
3827 kgrittn@postgresql.o 645 : 892 : metap = HashPageGetMeta(BufferGetPage(metabuf));
646 : :
647 : : /*
648 : : * Check to see if split is still needed; someone else might have already
649 : : * done one while we waited for the lock.
650 : : *
651 : : * Make sure this stays in sync with _hash_doinsert()
652 : : */
8439 tgl@sss.pgh.pa.us 653 : 892 : if (metap->hashm_ntuples <=
654 [ - + ]: 892 : (double) metap->hashm_ffactor * (metap->hashm_maxbucket + 1))
8439 tgl@sss.pgh.pa.us 655 :UBC 0 : goto fail;
656 : :
657 : : /*
658 : : * Can't split anymore if maxbucket has reached its maximum possible
659 : : * value.
660 : : *
661 : : * Ideally we'd allow bucket numbers up to UINT_MAX-1 (no higher because
662 : : * the calculation maxbucket+1 mustn't overflow). Currently we restrict
663 : : * to half that to prevent failure of pg_ceil_log2_32() and insufficient
664 : : * space in hashm_spares[]. It's moot anyway because an index with 2^32
665 : : * buckets would certainly overflow BlockNumber and hence
666 : : * _hash_alloc_buckets() would fail, but if we supported buckets smaller
667 : : * than a disk block then this would be an independent constraint.
668 : : *
669 : : * If you change this, see also the maximum initial number of buckets in
670 : : * _hash_init().
671 : : */
7267 tgl@sss.pgh.pa.us 672 [ - + ]:CBC 892 : if (metap->hashm_maxbucket >= (uint32) 0x7FFFFFFE)
7267 tgl@sss.pgh.pa.us 673 :UBC 0 : goto fail;
674 : :
675 : : /*
676 : : * Determine which bucket is to be split, and attempt to take cleanup lock
677 : : * on the old bucket. If we can't get the lock, give up.
678 : : *
679 : : * The cleanup lock protects us not only against other backends, but
680 : : * against our own backend as well.
681 : : *
682 : : * The cleanup lock is mainly to protect the split from concurrent
683 : : * inserts. See src/backend/access/hash/README, Lock Definitions for
684 : : * further details. Due to this locking restriction, if there is any
685 : : * pending scan, the split will give up which is not good, but harmless.
686 : : */
7116 tgl@sss.pgh.pa.us 687 :CBC 892 : new_bucket = metap->hashm_maxbucket + 1;
688 : :
8442 689 : 892 : old_bucket = (new_bucket & metap->hashm_lowmask);
690 : :
8439 691 [ + + ]: 892 : start_oblkno = BUCKET_TO_BLKNO(metap, old_bucket);
692 : :
3603 rhaas@postgresql.org 693 : 892 : buf_oblkno = _hash_getbuf_with_condlock_cleanup(rel, start_oblkno, LH_BUCKET_PAGE);
694 [ - + ]: 892 : if (!buf_oblkno)
8439 tgl@sss.pgh.pa.us 695 :UBC 0 : goto fail;
696 : :
3603 rhaas@postgresql.org 697 :CBC 892 : opage = BufferGetPage(buf_oblkno);
1655 michael@paquier.xyz 698 : 892 : oopaque = HashPageGetOpaque(opage);
699 : :
700 : : /*
701 : : * We want to finish the split from a bucket as there is no apparent
702 : : * benefit by not doing so and it will make the code complicated to finish
703 : : * the split that involves multiple buckets considering the case where new
704 : : * split also fails. We don't need to consider the new bucket for
705 : : * completing the split here as it is not possible that a re-split of new
706 : : * bucket starts when there is still a pending split from old bucket.
707 : : */
3603 rhaas@postgresql.org 708 [ - + ]: 892 : if (H_BUCKET_BEING_SPLIT(oopaque))
709 : : {
710 : : /*
711 : : * Copy bucket mapping info now; refer the comment in code below where
712 : : * we copy this information before calling _hash_splitbucket to see
713 : : * why this is okay.
714 : : */
3603 rhaas@postgresql.org 715 :UBC 0 : maxbucket = metap->hashm_maxbucket;
716 : 0 : highmask = metap->hashm_highmask;
717 : 0 : lowmask = metap->hashm_lowmask;
718 : :
719 : : /*
720 : : * Release the lock on metapage and old_bucket, before completing the
721 : : * split.
722 : : */
3580 723 : 0 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
724 : 0 : LockBuffer(buf_oblkno, BUFFER_LOCK_UNLOCK);
725 : :
3603 726 : 0 : _hash_finish_split(rel, metabuf, buf_oblkno, old_bucket, maxbucket,
727 : : highmask, lowmask);
728 : :
729 : : /* release the pin on old buffer and retry for expand. */
730 : 0 : _hash_dropbuf(rel, buf_oblkno);
731 : :
732 : 0 : goto restart_expand;
733 : : }
734 : :
735 : : /*
736 : : * Clean the tuples remained from the previous split. This operation
737 : : * requires cleanup lock and we already have one on the old bucket, so
738 : : * let's do it. We also don't want to allow further splits from the bucket
739 : : * till the garbage of previous split is cleaned. This has two
740 : : * advantages; first, it helps in avoiding the bloat due to garbage and
741 : : * second is, during cleanup of bucket, we are always sure that the
742 : : * garbage tuples belong to most recently split bucket. On the contrary,
743 : : * if we allow cleanup of bucket after meta page is updated to indicate
744 : : * the new split and before the actual split, the cleanup operation won't
745 : : * be able to decide whether the tuple has been moved to the newly created
746 : : * bucket and ended up deleting such tuples.
747 : : */
3603 rhaas@postgresql.org 748 [ - + ]:CBC 892 : if (H_NEEDS_SPLIT_CLEANUP(oopaque))
749 : : {
750 : : /*
751 : : * Copy bucket mapping info now; refer to the comment in code below
752 : : * where we copy this information before calling _hash_splitbucket to
753 : : * see why this is okay.
754 : : */
3598 rhaas@postgresql.org 755 :UBC 0 : maxbucket = metap->hashm_maxbucket;
756 : 0 : highmask = metap->hashm_highmask;
757 : 0 : lowmask = metap->hashm_lowmask;
758 : :
759 : : /* Release the metapage lock. */
3580 760 : 0 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
761 : :
3603 762 : 0 : hashbucketcleanup(rel, old_bucket, buf_oblkno, start_oblkno, NULL,
763 : : maxbucket, highmask, lowmask, NULL, NULL, true,
764 : : NULL, NULL);
765 : :
766 : 0 : _hash_dropbuf(rel, buf_oblkno);
767 : :
768 : 0 : goto restart_expand;
769 : : }
770 : :
771 : : /*
772 : : * There shouldn't be any active scan on new bucket.
773 : : *
774 : : * Note: it is safe to compute the new bucket's blkno here, even though we
775 : : * may still need to update the BUCKET_TO_BLKNO mapping. This is because
776 : : * the current value of hashm_spares[hashm_ovflpoint] correctly shows
777 : : * where we are going to put a new splitpoint's worth of buckets.
778 : : */
7116 tgl@sss.pgh.pa.us 779 [ + - ]:CBC 892 : start_nblkno = BUCKET_TO_BLKNO(metap, new_bucket);
780 : :
781 : : /*
782 : : * If the split point is increasing we need to allocate a new batch of
783 : : * bucket pages.
784 : : */
3479 rhaas@postgresql.org 785 : 892 : spare_ndx = _hash_spareindex(new_bucket + 1);
7116 tgl@sss.pgh.pa.us 786 [ + + ]: 892 : if (spare_ndx > metap->hashm_ovflpoint)
787 : : {
788 : : uint32 buckets_to_add;
789 : :
790 [ - + ]: 40 : Assert(spare_ndx == metap->hashm_ovflpoint + 1);
791 : :
792 : : /*
793 : : * We treat allocation of buckets as a separate WAL-logged action.
794 : : * Even if we fail after this operation, won't leak bucket pages;
795 : : * rather, the next split will consume this space. In any case, even
796 : : * without failure we don't use all the space in one split operation.
797 : : */
3479 rhaas@postgresql.org 798 : 40 : buckets_to_add = _hash_get_totalbuckets(spare_ndx) - new_bucket;
799 [ - + ]: 40 : if (!_hash_alloc_buckets(rel, start_nblkno, buckets_to_add))
800 : : {
801 : : /* can't split due to BlockNumber overflow */
3603 rhaas@postgresql.org 802 :UBC 0 : _hash_relbuf(rel, buf_oblkno);
7116 tgl@sss.pgh.pa.us 803 : 0 : goto fail;
804 : : }
805 : : }
806 : :
807 : : /*
808 : : * Physically allocate the new bucket's primary page. We want to do this
809 : : * before changing the metapage's mapping info, in case we can't get the
810 : : * disk space.
811 : : *
812 : : * XXX It doesn't make sense to call _hash_getnewbuf first, zeroing the
813 : : * buffer, and then only afterwards check whether we have a cleanup lock.
814 : : * However, since no scan can be accessing the buffer yet, any concurrent
815 : : * accesses will just be from processes like the bgwriter or checkpointer
816 : : * which don't care about its contents, so it doesn't really matter.
817 : : */
4214 tgl@sss.pgh.pa.us 818 :CBC 892 : buf_nblkno = _hash_getnewbuf(rel, start_nblkno, MAIN_FORKNUM);
3603 rhaas@postgresql.org 819 [ - + ]: 892 : if (!IsBufferCleanupOK(buf_nblkno))
820 : : {
3603 rhaas@postgresql.org 821 :UBC 0 : _hash_relbuf(rel, buf_oblkno);
822 : 0 : _hash_relbuf(rel, buf_nblkno);
823 : 0 : goto fail;
824 : : }
825 : :
826 : : /*
827 : : * Since we are scribbling on the pages in the shared buffers, establish a
828 : : * critical section. Any failure in this next code leaves us with a big
829 : : * problem: the metapage is effectively corrupt but could get written back
830 : : * to disk.
831 : : */
7795 tgl@sss.pgh.pa.us 832 :CBC 892 : START_CRIT_SECTION();
833 : :
834 : : /*
835 : : * Okay to proceed with split. Update the metapage bucket mapping info.
836 : : */
8439 837 : 892 : metap->hashm_maxbucket = new_bucket;
838 : :
8442 839 [ + + ]: 892 : if (new_bucket > metap->hashm_highmask)
840 : : {
841 : : /* Starting a new doubling */
842 : 14 : metap->hashm_lowmask = metap->hashm_highmask;
843 : 14 : metap->hashm_highmask = new_bucket | metap->hashm_lowmask;
3499 rhaas@postgresql.org 844 : 14 : metap_update_masks = true;
845 : : }
846 : :
847 : : /*
848 : : * If the split point is increasing we need to adjust the hashm_spares[]
849 : : * array and hashm_ovflpoint so that future overflow pages will be created
850 : : * beyond this new batch of bucket pages.
851 : : */
9220 tgl@sss.pgh.pa.us 852 [ + + ]: 892 : if (spare_ndx > metap->hashm_ovflpoint)
853 : : {
854 : 40 : metap->hashm_spares[spare_ndx] = metap->hashm_spares[metap->hashm_ovflpoint];
855 : 40 : metap->hashm_ovflpoint = spare_ndx;
3499 rhaas@postgresql.org 856 : 40 : metap_update_splitpoint = true;
857 : : }
858 : :
3512 859 : 892 : MarkBufferDirty(metabuf);
860 : :
861 : : /*
862 : : * Copy bucket mapping info now; this saves re-accessing the meta page
863 : : * inside _hash_splitbucket's inner loop. Note that once we drop the
864 : : * split lock, other splits could begin, so these values might be out of
865 : : * date before _hash_splitbucket finishes. That's okay, since all it
866 : : * needs is to tell which of these two buckets to map hashkeys into.
867 : : */
8439 tgl@sss.pgh.pa.us 868 : 892 : maxbucket = metap->hashm_maxbucket;
869 : 892 : highmask = metap->hashm_highmask;
870 : 892 : lowmask = metap->hashm_lowmask;
871 : :
3512 rhaas@postgresql.org 872 : 892 : opage = BufferGetPage(buf_oblkno);
1655 michael@paquier.xyz 873 : 892 : oopaque = HashPageGetOpaque(opage);
874 : :
875 : : /*
876 : : * Mark the old bucket to indicate that split is in progress. (At
877 : : * operation end, we will clear the split-in-progress flag.) Also, for a
878 : : * primary bucket page, hasho_prevblkno stores the number of buckets that
879 : : * existed as of the last split, so we must update that value here.
880 : : */
3512 rhaas@postgresql.org 881 : 892 : oopaque->hasho_flag |= LH_BUCKET_BEING_SPLIT;
882 : 892 : oopaque->hasho_prevblkno = maxbucket;
883 : :
884 : 892 : MarkBufferDirty(buf_oblkno);
885 : :
886 : 892 : npage = BufferGetPage(buf_nblkno);
887 : :
888 : : /*
889 : : * initialize the new bucket's primary page and mark it to indicate that
890 : : * split is in progress.
891 : : */
1655 michael@paquier.xyz 892 : 892 : nopaque = HashPageGetOpaque(npage);
3512 rhaas@postgresql.org 893 : 892 : nopaque->hasho_prevblkno = maxbucket;
894 : 892 : nopaque->hasho_nextblkno = InvalidBlockNumber;
895 : 892 : nopaque->hasho_bucket = new_bucket;
896 : 892 : nopaque->hasho_flag = LH_BUCKET_PAGE | LH_BUCKET_BEING_POPULATED;
897 : 892 : nopaque->hasho_page_id = HASHO_PAGE_ID;
898 : :
899 : 892 : MarkBufferDirty(buf_nblkno);
900 : :
901 : : /* XLOG stuff */
3499 902 [ + - - + : 892 : if (RelationNeedsWAL(rel))
- - - - ]
903 : 892 : {
904 : : xl_hash_split_allocate_page xlrec;
905 : :
906 : 892 : xlrec.new_bucket = maxbucket;
907 : 892 : xlrec.old_bucket_flag = oopaque->hasho_flag;
908 : 892 : xlrec.new_bucket_flag = nopaque->hasho_flag;
909 : 892 : xlrec.flags = 0;
910 : :
911 : 892 : XLogBeginInsert();
912 : :
913 : 892 : XLogRegisterBuffer(0, buf_oblkno, REGBUF_STANDARD);
914 : 892 : XLogRegisterBuffer(1, buf_nblkno, REGBUF_WILL_INIT);
915 : 892 : XLogRegisterBuffer(2, metabuf, REGBUF_STANDARD);
916 : :
917 [ + + ]: 892 : if (metap_update_masks)
918 : : {
919 : 14 : xlrec.flags |= XLH_SPLIT_META_UPDATE_MASKS;
608 peter@eisentraut.org 920 : 14 : XLogRegisterBufData(2, &metap->hashm_lowmask, sizeof(uint32));
921 : 14 : XLogRegisterBufData(2, &metap->hashm_highmask, sizeof(uint32));
922 : : }
923 : :
3499 rhaas@postgresql.org 924 [ + + ]: 892 : if (metap_update_splitpoint)
925 : : {
926 : 40 : xlrec.flags |= XLH_SPLIT_META_UPDATE_SPLITPOINT;
608 peter@eisentraut.org 927 : 40 : XLogRegisterBufData(2, &metap->hashm_ovflpoint,
928 : : sizeof(uint32));
3499 rhaas@postgresql.org 929 : 40 : XLogRegisterBufData(2,
608 peter@eisentraut.org 930 : 40 : &metap->hashm_spares[metap->hashm_ovflpoint],
931 : : sizeof(uint32));
932 : : }
933 : :
934 : 892 : XLogRegisterData(&xlrec, SizeOfHashSplitAllocPage);
935 : :
3499 rhaas@postgresql.org 936 : 892 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_ALLOCATE_PAGE);
937 : : }
938 : : else
204 pg@bowt.ie 939 :UBC 0 : recptr = XLogGetFakeLSN(rel);
940 : :
204 pg@bowt.ie 941 :CBC 892 : PageSetLSN(BufferGetPage(buf_oblkno), recptr);
942 : 892 : PageSetLSN(BufferGetPage(buf_nblkno), recptr);
943 : 892 : PageSetLSN(BufferGetPage(metabuf), recptr);
944 : :
3512 rhaas@postgresql.org 945 [ - + ]: 892 : END_CRIT_SECTION();
946 : :
947 : : /* drop lock, but keep pin */
3580 948 : 892 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
949 : :
950 : : /* Relocate records to the new bucket */
4210 tgl@sss.pgh.pa.us 951 : 892 : _hash_splitbucket(rel, metabuf,
952 : : old_bucket, new_bucket,
953 : : buf_oblkno, buf_nblkno, NULL,
954 : : maxbucket, highmask, lowmask);
955 : :
956 : : /* all done, now release the pins on primary buckets. */
3356 rhaas@postgresql.org 957 : 891 : _hash_dropbuf(rel, buf_oblkno);
958 : 891 : _hash_dropbuf(rel, buf_nblkno);
959 : :
8439 tgl@sss.pgh.pa.us 960 : 891 : return;
961 : :
962 : : /* Here if decide not to split or fail to acquire old bucket lock */
8439 tgl@sss.pgh.pa.us 963 :UBC 0 : fail:
964 : :
965 : : /* We didn't write the metapage, so just drop lock */
3580 rhaas@postgresql.org 966 : 0 : LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
967 : : }
968 : :
969 : :
970 : : /*
971 : : * _hash_alloc_buckets -- allocate a new splitpoint's worth of bucket pages
972 : : *
973 : : * This does not need to initialize the new bucket pages; we'll do that as
974 : : * each one is used by _hash_expandtable(). But we have to extend the logical
975 : : * EOF to the end of the splitpoint; this keeps smgr's idea of the EOF in
976 : : * sync with ours, so that we don't get complaints from smgr.
977 : : *
978 : : * We do this by writing a page of zeroes at the end of the splitpoint range.
979 : : * We expect that the filesystem will ensure that the intervening pages read
980 : : * as zeroes too. On many filesystems this "hole" will not be allocated
981 : : * immediately, which means that the index file may end up more fragmented
982 : : * than if we forced it all to be allocated now; but since we don't scan
983 : : * hash indexes sequentially anyway, that probably doesn't matter.
984 : : *
985 : : * XXX It's annoying that this code is executed with the metapage lock held.
986 : : * We need to interlock against _hash_addovflpage() adding a new overflow page
987 : : * concurrently, but it'd likely be better to use LockRelationForExtension
988 : : * for the purpose. OTOH, adding a splitpoint is a very infrequent operation,
989 : : * so it may not be worth worrying about.
990 : : *
991 : : * Returns true if successful, or false if allocation failed due to
992 : : * BlockNumber overflow.
993 : : */
994 : : static bool
7116 tgl@sss.pgh.pa.us 995 :CBC 40 : _hash_alloc_buckets(Relation rel, BlockNumber firstblock, uint32 nblocks)
996 : : {
997 : : BlockNumber lastblock;
998 : : PGIOAlignedBlock zerobuf;
999 : : Page page;
1000 : : HashPageOpaque ovflopaque;
1001 : :
7267 1002 : 40 : lastblock = firstblock + nblocks - 1;
1003 : :
1004 : : /*
1005 : : * Check for overflow in block number calculation; if so, we cannot extend
1006 : : * the index anymore.
1007 : : */
1008 [ + - - + ]: 40 : if (lastblock < firstblock || lastblock == InvalidBlockNumber)
7116 tgl@sss.pgh.pa.us 1009 :UBC 0 : return false;
1010 : :
2963 tgl@sss.pgh.pa.us 1011 :CBC 40 : page = (Page) zerobuf.data;
1012 : :
1013 : : /*
1014 : : * Initialize the page. Just zeroing the page won't work; see
1015 : : * _hash_freeovflpage for similar usage. We take care to make the special
1016 : : * space valid for the benefit of tools such as pageinspect.
1017 : : */
3499 rhaas@postgresql.org 1018 : 40 : _hash_pageinit(page, BLCKSZ);
1019 : :
1655 michael@paquier.xyz 1020 : 40 : ovflopaque = HashPageGetOpaque(page);
1021 : :
3477 rhaas@postgresql.org 1022 : 40 : ovflopaque->hasho_prevblkno = InvalidBlockNumber;
1023 : 40 : ovflopaque->hasho_nextblkno = InvalidBlockNumber;
1928 peter@eisentraut.org 1024 : 40 : ovflopaque->hasho_bucket = InvalidBucket;
3477 rhaas@postgresql.org 1025 : 40 : ovflopaque->hasho_flag = LH_UNUSED_PAGE;
1026 : 40 : ovflopaque->hasho_page_id = HASHO_PAGE_ID;
1027 : :
3499 1028 [ + - - + : 40 : if (RelationNeedsWAL(rel))
- - - - ]
1559 1029 : 40 : log_newpage(&rel->rd_locator,
1030 : : MAIN_FORKNUM,
1031 : : lastblock,
1032 : : zerobuf.data,
1033 : : true);
1034 : :
199 andres@anarazel.de 1035 : 40 : PageSetChecksum(page, lastblock);
1918 tgl@sss.pgh.pa.us 1036 : 40 : smgrextend(RelationGetSmgr(rel), MAIN_FORKNUM, lastblock, zerobuf.data,
1037 : : false);
1038 : :
7116 1039 : 40 : return true;
1040 : : }
1041 : :
1042 : :
1043 : : /*
1044 : : * _hash_splitbucket -- split 'obucket' into 'obucket' and 'nbucket'
1045 : : *
1046 : : * This routine is used to partition the tuples between old and new bucket and
1047 : : * is used to finish the incomplete split operations. To finish the previously
1048 : : * interrupted split operation, the caller needs to fill htab. If htab is set,
1049 : : * then we skip the movement of tuples that exists in htab, otherwise NULL
1050 : : * value of htab indicates movement of all the tuples that belong to the new
1051 : : * bucket.
1052 : : *
1053 : : * We are splitting a bucket that consists of a base bucket page and zero
1054 : : * or more overflow (bucket chain) pages. We must relocate tuples that
1055 : : * belong in the new bucket.
1056 : : *
1057 : : * The caller must hold cleanup locks on both buckets to ensure that
1058 : : * no one else is trying to access them (see README).
1059 : : *
1060 : : * The caller must hold a pin, but no lock, on the metapage buffer.
1061 : : * The buffer is returned in the same state. (The metapage is only
1062 : : * touched if it becomes necessary to add or remove overflow pages.)
1063 : : *
1064 : : * Split needs to retain pin on primary bucket pages of both old and new
1065 : : * buckets till end of operation. This is to prevent vacuum from starting
1066 : : * while a split is in progress.
1067 : : *
1068 : : * In addition, the caller must have created the new bucket's base page,
1069 : : * which is passed in buffer nbuf, pinned and write-locked. The lock will be
1070 : : * released here and pin must be released by the caller. (The API is set up
1071 : : * this way because we must do _hash_getnewbuf() before releasing the metapage
1072 : : * write lock. So instead of passing the new bucket's start block number, we
1073 : : * pass an actual buffer.)
1074 : : */
1075 : : static void
8442 1076 : 893 : _hash_splitbucket(Relation rel,
1077 : : Buffer metabuf,
1078 : : Bucket obucket,
1079 : : Bucket nbucket,
1080 : : Buffer obuf,
1081 : : Buffer nbuf,
1082 : : HTAB *htab,
1083 : : uint32 maxbucket,
1084 : : uint32 highmask,
1085 : : uint32 lowmask)
1086 : : {
1087 : : Buffer bucket_obuf;
1088 : : Buffer bucket_nbuf;
1089 : : Page opage;
1090 : : Page npage;
1091 : : HashPageOpaque oopaque;
1092 : : HashPageOpaque nopaque;
1093 : : OffsetNumber itup_offsets[MaxIndexTuplesPerPage];
1094 : : IndexTuple itups[MaxIndexTuplesPerPage];
3499 rhaas@postgresql.org 1095 : 893 : Size all_tups_size = 0;
1096 : : int i;
1097 : 893 : uint16 nitups = 0;
1098 : : XLogRecPtr recptr;
1099 : :
3603 1100 : 893 : bucket_obuf = obuf;
1101 : 893 : opage = BufferGetPage(obuf);
1655 michael@paquier.xyz 1102 : 893 : oopaque = HashPageGetOpaque(opage);
1103 : :
3603 rhaas@postgresql.org 1104 : 893 : bucket_nbuf = nbuf;
1105 : 893 : npage = BufferGetPage(nbuf);
1655 michael@paquier.xyz 1106 : 893 : nopaque = HashPageGetOpaque(npage);
1107 : :
1108 : : /* Copy the predicate locks from old bucket to new bucket. */
3110 teodor@sigaev.ru 1109 : 893 : PredicateLockPageSplit(rel,
1110 : : BufferGetBlockNumber(bucket_obuf),
1111 : : BufferGetBlockNumber(bucket_nbuf));
1112 : :
2 pg@bowt.ie 1113 : 893 : INJECTION_POINT("hash-split-before-relocation", NULL);
1114 : :
1115 : : /*
1116 : : * Partition the tuples in the old bucket between the old bucket and the
1117 : : * new bucket, advancing along the old bucket's overflow bucket chain and
1118 : : * adding overflow pages to the new bucket as needed. Outer loop iterates
1119 : : * once per page in old bucket.
1120 : : */
1121 : : for (;;)
10627 bruce@momjian.us 1122 : 226 : {
1123 : : BlockNumber oblkno;
1124 : : OffsetNumber ooffnum;
1125 : : OffsetNumber omaxoffnum;
1126 : :
1127 : : /* Scan each tuple in old page */
6189 tgl@sss.pgh.pa.us 1128 : 1118 : omaxoffnum = PageGetMaxOffsetNumber(opage);
1129 : 1118 : for (ooffnum = FirstOffsetNumber;
1130 [ + + ]: 205786 : ooffnum <= omaxoffnum;
1131 : 204668 : ooffnum = OffsetNumberNext(ooffnum))
1132 : : {
1133 : : IndexTuple itup;
1134 : : Size itemsz;
1135 : : Bucket bucket;
3603 rhaas@postgresql.org 1136 : 204668 : bool found = false;
1137 : :
1138 : : /* skip dead tuples */
3625 1139 [ - + ]: 204668 : if (ItemIdIsDead(PageGetItemId(opage, ooffnum)))
3625 rhaas@postgresql.org 1140 :UBC 0 : continue;
1141 : :
1142 : : /*
1143 : : * Before inserting a tuple, probe the hash table containing TIDs
1144 : : * of tuples belonging to new bucket, if we find a match, then
1145 : : * skip that tuple, else fetch the item's hash key (conveniently
1146 : : * stored in the item) and determine which bucket it now belongs
1147 : : * in.
1148 : : */
6189 tgl@sss.pgh.pa.us 1149 :CBC 204668 : itup = (IndexTuple) PageGetItem(opage,
1150 : 204668 : PageGetItemId(opage, ooffnum));
1151 : :
3603 rhaas@postgresql.org 1152 [ + + ]: 204668 : if (htab)
1153 : 318 : (void) hash_search(htab, &itup->t_tid, HASH_FIND, &found);
1154 : :
1155 [ - + ]: 204668 : if (found)
3603 rhaas@postgresql.org 1156 :UBC 0 : continue;
1157 : :
6189 tgl@sss.pgh.pa.us 1158 :CBC 204668 : bucket = _hash_hashkey2bucket(_hash_get_indextuple_hashkey(itup),
1159 : : maxbucket, highmask, lowmask);
1160 : :
1161 [ + + ]: 204668 : if (bucket == nbucket)
1162 : : {
1163 : : IndexTuple new_itup;
1164 : :
1165 : : /*
1166 : : * make a copy of index tuple as we have to scribble on it.
1167 : : */
3603 rhaas@postgresql.org 1168 : 83862 : new_itup = CopyIndexTuple(itup);
1169 : :
1170 : : /*
1171 : : * mark the index tuple as moved by split, such tuples are
1172 : : * skipped by scan if there is split in progress for a bucket.
1173 : : */
1174 : 83862 : new_itup->t_info |= INDEX_MOVED_BY_SPLIT_MASK;
1175 : :
1176 : : /*
1177 : : * insert the tuple into the new bucket. if it doesn't fit on
1178 : : * the current page in the new bucket, we must allocate a new
1179 : : * overflow page and place the tuple on that page instead.
1180 : : */
3148 tgl@sss.pgh.pa.us 1181 : 83862 : itemsz = IndexTupleSize(new_itup);
6189 1182 : 83862 : itemsz = MAXALIGN(itemsz);
1183 : :
3499 rhaas@postgresql.org 1184 [ + + ]: 83862 : if (PageGetFreeSpaceForMultipleTuples(npage, nitups + 1) < (all_tups_size + itemsz))
1185 : : {
1186 : : /*
1187 : : * Change the shared buffer state in critical section,
1188 : : * otherwise any error could make it unrecoverable.
1189 : : */
1190 : 52 : START_CRIT_SECTION();
1191 : :
1192 : 52 : _hash_pgaddmultitup(rel, nbuf, itups, itup_offsets, nitups);
3580 1193 : 52 : MarkBufferDirty(nbuf);
1194 : : /* log the split operation before releasing the lock */
3499 1195 : 52 : log_split_page(rel, nbuf);
1196 : :
1197 [ - + ]: 52 : END_CRIT_SECTION();
1198 : :
1199 : : /* drop lock, but keep pin */
3580 1200 : 52 : LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
1201 : :
1202 : : /* be tidy */
3499 1203 [ + + ]: 21216 : for (i = 0; i < nitups; i++)
1204 : 21164 : pfree(itups[i]);
1205 : 52 : nitups = 0;
1206 : 52 : all_tups_size = 0;
1207 : :
1208 : : /* chain to a new overflow page */
1860 michael@paquier.xyz 1209 : 52 : nbuf = _hash_addovflpage(rel, metabuf, nbuf, (nbuf == bucket_nbuf));
3827 kgrittn@postgresql.o 1210 : 52 : npage = BufferGetPage(nbuf);
1655 michael@paquier.xyz 1211 : 52 : nopaque = HashPageGetOpaque(npage);
1212 : : }
1213 : :
3499 rhaas@postgresql.org 1214 : 83862 : itups[nitups++] = new_itup;
1215 : 83862 : all_tups_size += itemsz;
1216 : : }
1217 : : else
1218 : : {
1219 : : /*
1220 : : * the tuple stays on this page, so nothing to do.
1221 : : */
6189 tgl@sss.pgh.pa.us 1222 [ - + ]: 120806 : Assert(bucket == obucket);
1223 : : }
1224 : : }
1225 : :
1226 : 1118 : oblkno = oopaque->hasho_nextblkno;
1227 : :
1228 : : /* retain the pin on the old primary bucket */
3603 rhaas@postgresql.org 1229 [ + + ]: 1118 : if (obuf == bucket_obuf)
3580 1230 : 892 : LockBuffer(obuf, BUFFER_LOCK_UNLOCK);
1231 : : else
6189 tgl@sss.pgh.pa.us 1232 : 226 : _hash_relbuf(rel, obuf);
1233 : :
1234 : : /* Exit loop if no more overflow pages in old bucket */
1235 [ + + ]: 1118 : if (!BlockNumberIsValid(oblkno))
1236 : : {
1237 : : /*
1238 : : * Change the shared buffer state in critical section, otherwise
1239 : : * any error could make it unrecoverable.
1240 : : */
3499 rhaas@postgresql.org 1241 : 892 : START_CRIT_SECTION();
1242 : :
1243 : 892 : _hash_pgaddmultitup(rel, nbuf, itups, itup_offsets, nitups);
3512 1244 : 892 : MarkBufferDirty(nbuf);
1245 : : /* log the split operation before releasing the lock */
3499 1246 : 892 : log_split_page(rel, nbuf);
1247 : :
1248 [ - + ]: 892 : END_CRIT_SECTION();
1249 : :
3512 1250 [ + + ]: 892 : if (nbuf == bucket_nbuf)
1251 : 888 : LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
1252 : : else
1253 : 4 : _hash_relbuf(rel, nbuf);
1254 : :
1255 : : /* be tidy */
3499 1256 [ + + ]: 63590 : for (i = 0; i < nitups; i++)
1257 : 62698 : pfree(itups[i]);
6189 tgl@sss.pgh.pa.us 1258 : 892 : break;
1259 : : }
1260 : :
1261 : : /* Else, advance to next old page */
3603 rhaas@postgresql.org 1262 : 226 : obuf = _hash_getbuf(rel, oblkno, HASH_READ, LH_OVERFLOW_PAGE);
3827 kgrittn@postgresql.o 1263 : 226 : opage = BufferGetPage(obuf);
1655 michael@paquier.xyz 1264 : 226 : oopaque = HashPageGetOpaque(opage);
1265 : : }
1266 : :
1267 : : /*
1268 : : * We're at the end of the old bucket chain, so we're done partitioning
1269 : : * the tuples. Mark the old and new buckets to indicate split is
1270 : : * finished.
1271 : : *
1272 : : * To avoid deadlocks due to locking order of buckets, first lock the old
1273 : : * bucket and then the new bucket.
1274 : : */
3580 rhaas@postgresql.org 1275 : 892 : LockBuffer(bucket_obuf, BUFFER_LOCK_EXCLUSIVE);
3603 1276 : 892 : opage = BufferGetPage(bucket_obuf);
1655 michael@paquier.xyz 1277 : 892 : oopaque = HashPageGetOpaque(opage);
1278 : :
3580 rhaas@postgresql.org 1279 : 892 : LockBuffer(bucket_nbuf, BUFFER_LOCK_EXCLUSIVE);
3603 1280 : 892 : npage = BufferGetPage(bucket_nbuf);
1655 michael@paquier.xyz 1281 : 892 : nopaque = HashPageGetOpaque(npage);
1282 : :
3499 rhaas@postgresql.org 1283 : 892 : START_CRIT_SECTION();
1284 : :
3603 1285 : 892 : oopaque->hasho_flag &= ~LH_BUCKET_BEING_SPLIT;
1286 : 892 : nopaque->hasho_flag &= ~LH_BUCKET_BEING_POPULATED;
1287 : :
1288 : : /*
1289 : : * After the split is finished, mark the old bucket to indicate that it
1290 : : * contains deletable tuples. We will clear split-cleanup flag after
1291 : : * deleting such tuples either at the end of split or at the next split
1292 : : * from old bucket or at the time of vacuum.
1293 : : */
1294 : 892 : oopaque->hasho_flag |= LH_BUCKET_NEEDS_SPLIT_CLEANUP;
1295 : :
1296 : : /*
1297 : : * now write the buffers, here we don't release the locks as caller is
1298 : : * responsible to release locks.
1299 : : */
1300 : 892 : MarkBufferDirty(bucket_obuf);
1301 : 892 : MarkBufferDirty(bucket_nbuf);
1302 : :
3499 1303 [ + - - + : 892 : if (RelationNeedsWAL(rel))
- - - - ]
1304 : 892 : {
1305 : : xl_hash_split_complete xlrec;
1306 : :
1307 : 892 : xlrec.old_bucket_flag = oopaque->hasho_flag;
1308 : 892 : xlrec.new_bucket_flag = nopaque->hasho_flag;
1309 : :
1310 : 892 : XLogBeginInsert();
1311 : :
608 peter@eisentraut.org 1312 : 892 : XLogRegisterData(&xlrec, SizeOfHashSplitComplete);
1313 : :
3499 rhaas@postgresql.org 1314 : 892 : XLogRegisterBuffer(0, bucket_obuf, REGBUF_STANDARD);
1315 : 892 : XLogRegisterBuffer(1, bucket_nbuf, REGBUF_STANDARD);
1316 : :
1317 : 892 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_COMPLETE);
1318 : : }
1319 : : else
204 pg@bowt.ie 1320 :UBC 0 : recptr = XLogGetFakeLSN(rel);
1321 : :
204 pg@bowt.ie 1322 :CBC 892 : PageSetLSN(BufferGetPage(bucket_obuf), recptr);
1323 : 892 : PageSetLSN(BufferGetPage(bucket_nbuf), recptr);
1324 : :
3499 rhaas@postgresql.org 1325 [ - + ]: 892 : END_CRIT_SECTION();
1326 : :
1327 : : /*
1328 : : * If possible, clean up the old bucket. We might not be able to do this
1329 : : * if someone else has a pin on it, but if not then we can go ahead. This
1330 : : * isn't absolutely necessary, but it reduces bloat; if we don't do it
1331 : : * now, VACUUM will do it eventually, but maybe not until new overflow
1332 : : * pages have been allocated. Note that there's no need to clean up the
1333 : : * new bucket.
1334 : : */
3356 1335 [ + - ]: 892 : if (IsBufferCleanupOK(bucket_obuf))
1336 : : {
1337 : 892 : LockBuffer(bucket_nbuf, BUFFER_LOCK_UNLOCK);
1338 : 892 : hashbucketcleanup(rel, obucket, bucket_obuf,
1339 : : BufferGetBlockNumber(bucket_obuf), NULL,
1340 : : maxbucket, highmask, lowmask, NULL, NULL, true,
1341 : : NULL, NULL);
1342 : : }
1343 : : else
1344 : : {
3356 rhaas@postgresql.org 1345 :UBC 0 : LockBuffer(bucket_nbuf, BUFFER_LOCK_UNLOCK);
1346 : 0 : LockBuffer(bucket_obuf, BUFFER_LOCK_UNLOCK);
1347 : : }
3603 rhaas@postgresql.org 1348 :CBC 892 : }
1349 : :
1350 : : /*
1351 : : * _hash_finish_split() -- Finish the previously interrupted split operation
1352 : : *
1353 : : * To complete the split operation, we form the hash table of TIDs in new
1354 : : * bucket which is then used by split operation to skip tuples that are
1355 : : * already moved before the split operation was previously interrupted.
1356 : : *
1357 : : * The caller must hold a pin, but no lock, on the metapage and old bucket's
1358 : : * primary page buffer. The buffers are returned in the same state. (The
1359 : : * metapage is only touched if it becomes necessary to add or remove overflow
1360 : : * pages.)
1361 : : */
1362 : : void
1363 : 1 : _hash_finish_split(Relation rel, Buffer metabuf, Buffer obuf, Bucket obucket,
1364 : : uint32 maxbucket, uint32 highmask, uint32 lowmask)
1365 : : {
1366 : : HASHCTL hash_ctl;
1367 : : HTAB *tidhtab;
1368 : 1 : Buffer bucket_nbuf = InvalidBuffer;
1369 : : Buffer nbuf;
1370 : : Page npage;
1371 : : BlockNumber nblkno;
1372 : : BlockNumber bucket_nblkno;
1373 : : HashPageOpaque npageopaque;
1374 : : Bucket nbucket;
1375 : : bool found;
1376 : :
1377 : : /* Initialize hash tables used to track TIDs */
1378 : 1 : hash_ctl.keysize = sizeof(ItemPointerData);
1379 : 1 : hash_ctl.entrysize = sizeof(ItemPointerData);
1380 : 1 : hash_ctl.hcxt = CurrentMemoryContext;
1381 : :
1382 : : tidhtab =
1383 : 1 : hash_create("bucket ctids",
1384 : : 256, /* arbitrary initial size */
1385 : : &hash_ctl,
1386 : : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1387 : :
1388 : 1 : bucket_nblkno = nblkno = _hash_get_newblock_from_oldbucket(rel, obucket);
1389 : :
1390 : : /*
1391 : : * Scan the new bucket and build hash table of TIDs
1392 : : */
1393 : : for (;;)
3603 rhaas@postgresql.org 1394 :UBC 0 : {
1395 : : OffsetNumber noffnum;
1396 : : OffsetNumber nmaxoffnum;
1397 : :
3603 rhaas@postgresql.org 1398 :CBC 1 : nbuf = _hash_getbuf(rel, nblkno, HASH_READ,
1399 : : LH_BUCKET_PAGE | LH_OVERFLOW_PAGE);
1400 : :
1401 : : /* remember the primary bucket buffer to acquire cleanup lock on it. */
1402 [ + - ]: 1 : if (nblkno == bucket_nblkno)
1403 : 1 : bucket_nbuf = nbuf;
1404 : :
1405 : 1 : npage = BufferGetPage(nbuf);
1655 michael@paquier.xyz 1406 : 1 : npageopaque = HashPageGetOpaque(npage);
1407 : :
1408 : : /* Scan each tuple in new page */
3603 rhaas@postgresql.org 1409 : 1 : nmaxoffnum = PageGetMaxOffsetNumber(npage);
1410 : 1 : for (noffnum = FirstOffsetNumber;
1411 [ + + ]: 2 : noffnum <= nmaxoffnum;
1412 : 1 : noffnum = OffsetNumberNext(noffnum))
1413 : : {
1414 : : IndexTuple itup;
1415 : :
1416 : : /* Fetch the item's TID and insert it in hash table. */
1417 : 1 : itup = (IndexTuple) PageGetItem(npage,
1418 : 1 : PageGetItemId(npage, noffnum));
1419 : :
1420 : 1 : (void) hash_search(tidhtab, &itup->t_tid, HASH_ENTER, &found);
1421 : :
1422 [ - + ]: 1 : Assert(!found);
1423 : : }
1424 : :
1425 : 1 : nblkno = npageopaque->hasho_nextblkno;
1426 : :
1427 : : /*
1428 : : * release our write lock without modifying buffer and ensure to
1429 : : * retain the pin on primary bucket.
1430 : : */
1431 [ + - ]: 1 : if (nbuf == bucket_nbuf)
3580 1432 : 1 : LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
1433 : : else
3603 rhaas@postgresql.org 1434 :UBC 0 : _hash_relbuf(rel, nbuf);
1435 : :
1436 : : /* Exit loop if no more overflow pages in new bucket */
3603 rhaas@postgresql.org 1437 [ + - ]:CBC 1 : if (!BlockNumberIsValid(nblkno))
1438 : 1 : break;
1439 : : }
1440 : :
1441 : : /*
1442 : : * Conditionally get the cleanup lock on old and new buckets to perform
1443 : : * the split operation. If we don't get the cleanup locks, silently give
1444 : : * up and next insertion on old bucket will try again to complete the
1445 : : * split.
1446 : : */
1447 [ - + ]: 1 : if (!ConditionalLockBufferForCleanup(obuf))
1448 : : {
3603 rhaas@postgresql.org 1449 :UBC 0 : hash_destroy(tidhtab);
1450 : 0 : return;
1451 : : }
3603 rhaas@postgresql.org 1452 [ - + ]:CBC 1 : if (!ConditionalLockBufferForCleanup(bucket_nbuf))
1453 : : {
3580 rhaas@postgresql.org 1454 :UBC 0 : LockBuffer(obuf, BUFFER_LOCK_UNLOCK);
3603 1455 : 0 : hash_destroy(tidhtab);
1456 : 0 : return;
1457 : : }
1458 : :
3603 rhaas@postgresql.org 1459 :CBC 1 : npage = BufferGetPage(bucket_nbuf);
1655 michael@paquier.xyz 1460 : 1 : npageopaque = HashPageGetOpaque(npage);
3603 rhaas@postgresql.org 1461 : 1 : nbucket = npageopaque->hasho_bucket;
1462 : :
2 pg@bowt.ie 1463 : 1 : INJECTION_POINT("hash-finish-incomplete-split", NULL);
1464 : :
3512 rhaas@postgresql.org 1465 : 1 : _hash_splitbucket(rel, metabuf, obucket,
1466 : : nbucket, obuf, bucket_nbuf, tidhtab,
1467 : : maxbucket, highmask, lowmask);
1468 : :
3356 1469 : 1 : _hash_dropbuf(rel, bucket_nbuf);
3603 1470 : 1 : hash_destroy(tidhtab);
1471 : : }
1472 : :
1473 : : /*
1474 : : * log_split_page() -- Log the split operation
1475 : : *
1476 : : * We log the split operation when the new page in new bucket gets full,
1477 : : * so we log the entire page.
1478 : : *
1479 : : * 'buf' must be locked by the caller which is also responsible for unlocking
1480 : : * it.
1481 : : */
1482 : : static void
3499 1483 : 944 : log_split_page(Relation rel, Buffer buf)
1484 : : {
1485 : : XLogRecPtr recptr;
1486 : :
1487 [ + - - + : 944 : if (RelationNeedsWAL(rel))
- - - - ]
1488 : : {
1489 : 944 : XLogBeginInsert();
1490 : :
1491 : 944 : XLogRegisterBuffer(0, buf, REGBUF_FORCE_IMAGE | REGBUF_STANDARD);
1492 : :
1493 : 944 : recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_PAGE);
1494 : : }
1495 : : else
88 pg@bowt.ie 1496 :UBC 0 : recptr = XLogGetFakeLSN(rel);
1497 : :
88 pg@bowt.ie 1498 :CBC 944 : PageSetLSN(BufferGetPage(buf), recptr);
3499 rhaas@postgresql.org 1499 : 944 : }
1500 : :
1501 : : /*
1502 : : * _hash_getcachedmetap() -- Returns cached metapage data.
1503 : : *
1504 : : * If metabuf is not InvalidBuffer, caller must hold a pin, but no lock, on
1505 : : * the metapage. If not set, we'll set it before returning if we have to
1506 : : * refresh the cache, and return with a pin but no lock on it; caller is
1507 : : * responsible for releasing the pin.
1508 : : *
1509 : : * We refresh the cache if it's not initialized yet or force_refresh is true.
1510 : : */
1511 : : HashMetaPage
3534 1512 : 492534 : _hash_getcachedmetap(Relation rel, Buffer *metabuf, bool force_refresh)
1513 : : {
1514 : : Page page;
1515 : :
1516 [ - + ]: 492534 : Assert(metabuf);
1517 [ + + + + ]: 492534 : if (force_refresh || rel->rd_amcache == NULL)
1518 : : {
3435 bruce@momjian.us 1519 : 733 : char *cache = NULL;
1520 : :
1521 : : /*
1522 : : * It's important that we don't set rd_amcache to an invalid value.
1523 : : * Either MemoryContextAlloc or _hash_getbuf could fail, so don't
1524 : : * install a pointer to the newly-allocated storage in the actual
1525 : : * relcache entry until both have succeeded.
1526 : : */
3534 rhaas@postgresql.org 1527 [ + + ]: 733 : if (rel->rd_amcache == NULL)
1528 : 320 : cache = MemoryContextAlloc(rel->rd_indexcxt,
1529 : : sizeof(HashMetaPageData));
1530 : :
1531 : : /* Read the metapage. */
1532 [ - + ]: 733 : if (BufferIsValid(*metabuf))
3534 rhaas@postgresql.org 1533 :UBC 0 : LockBuffer(*metabuf, BUFFER_LOCK_SHARE);
1534 : : else
3534 rhaas@postgresql.org 1535 :CBC 733 : *metabuf = _hash_getbuf(rel, HASH_METAPAGE, HASH_READ,
1536 : : LH_META_PAGE);
1537 : 733 : page = BufferGetPage(*metabuf);
1538 : :
1539 : : /* Populate the cache. */
1540 [ + + ]: 733 : if (rel->rd_amcache == NULL)
1541 : 320 : rel->rd_amcache = cache;
1542 : 733 : memcpy(rel->rd_amcache, HashPageGetMeta(page),
1543 : : sizeof(HashMetaPageData));
1544 : :
1545 : : /* Release metapage lock, but keep the pin. */
1546 : 733 : LockBuffer(*metabuf, BUFFER_LOCK_UNLOCK);
1547 : : }
1548 : :
1549 : 492534 : return (HashMetaPage) rel->rd_amcache;
1550 : : }
1551 : :
1552 : : /*
1553 : : * _hash_getbucketbuf_from_hashkey() -- Get the bucket's buffer for the given
1554 : : * hashkey.
1555 : : *
1556 : : * Bucket pages do not move or get removed once they are allocated. This give
1557 : : * us an opportunity to use the previously saved metapage contents to reach
1558 : : * the target bucket buffer, instead of reading from the metapage every time.
1559 : : * This saves one buffer access every time we want to reach the target bucket
1560 : : * buffer, which is very helpful savings in bufmgr traffic and contention.
1561 : : *
1562 : : * The access type parameter (HASH_READ or HASH_WRITE) indicates whether the
1563 : : * bucket buffer has to be locked for reading or writing.
1564 : : *
1565 : : * The out parameter cachedmetap is set with metapage contents used for
1566 : : * hashkey to bucket buffer mapping. Some callers need this info to reach the
1567 : : * old bucket in case of bucket split, see _hash_doinsert().
1568 : : */
1569 : : Buffer
1570 : 492096 : _hash_getbucketbuf_from_hashkey(Relation rel, uint32 hashkey, int access,
1571 : : HashMetaPage *cachedmetap)
1572 : : {
1573 : : HashMetaPage metap;
1574 : : Buffer buf;
1575 : 492096 : Buffer metabuf = InvalidBuffer;
1576 : : Page page;
1577 : : Bucket bucket;
1578 : : BlockNumber blkno;
1579 : : HashPageOpaque opaque;
1580 : :
1581 : : /* We read from target bucket buffer, hence locking is must. */
1582 [ + + - + ]: 492096 : Assert(access == HASH_READ || access == HASH_WRITE);
1583 : :
1584 : 492096 : metap = _hash_getcachedmetap(rel, &metabuf, false);
1585 [ + - ]: 492096 : Assert(metap != NULL);
1586 : :
1587 : : /*
1588 : : * Loop until we get a lock on the correct target bucket.
1589 : : */
1590 : : for (;;)
1591 : : {
1592 : : /*
1593 : : * Compute the target bucket number, and convert to block number.
1594 : : */
1595 : 492509 : bucket = _hash_hashkey2bucket(hashkey,
1596 : : metap->hashm_maxbucket,
1597 : : metap->hashm_highmask,
1598 : : metap->hashm_lowmask);
1599 : :
1600 [ + + ]: 492509 : blkno = BUCKET_TO_BLKNO(metap, bucket);
1601 : :
1602 : : /* Fetch the primary bucket page for the bucket */
1603 : 492509 : buf = _hash_getbuf(rel, blkno, access, LH_BUCKET_PAGE);
1604 : 492509 : page = BufferGetPage(buf);
1655 michael@paquier.xyz 1605 : 492509 : opaque = HashPageGetOpaque(page);
3534 rhaas@postgresql.org 1606 [ - + ]: 492509 : Assert(opaque->hasho_bucket == bucket);
3443 1607 [ - + ]: 492509 : Assert(opaque->hasho_prevblkno != InvalidBlockNumber);
1608 : :
1609 : : /*
1610 : : * If this bucket hasn't been split, we're done.
1611 : : */
1612 [ + + ]: 492509 : if (opaque->hasho_prevblkno <= metap->hashm_maxbucket)
3534 1613 : 492096 : break;
1614 : :
1615 : : /* Drop lock on this buffer, update cached metapage, and retry. */
1616 : 413 : _hash_relbuf(rel, buf);
1617 : 413 : metap = _hash_getcachedmetap(rel, &metabuf, true);
1618 [ - + ]: 413 : Assert(metap != NULL);
1619 : : }
1620 : :
1621 [ + + ]: 492096 : if (BufferIsValid(metabuf))
1622 : 727 : _hash_dropbuf(rel, metabuf);
1623 : :
1624 [ + + ]: 492096 : if (cachedmetap)
1625 : 491751 : *cachedmetap = metap;
1626 : :
1627 : 492096 : return buf;
1628 : : }
|
