Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * visibilitymap.c
4 : * bitmap for tracking visibility of heap tuples
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/heap/visibilitymap.c
12 : *
13 : * INTERFACE ROUTINES
14 : * visibilitymap_clear - clear bits for one page in the visibility map
15 : * visibilitymap_pin - pin a map page for setting a bit
16 : * visibilitymap_pin_ok - check whether correct map page is already pinned
17 : * visibilitymap_set - set bit(s) in a previously pinned page
18 : * visibilitymap_get_status - get status of bits
19 : * visibilitymap_count - count number of bits set in visibility map
20 : * visibilitymap_prepare_truncate -
21 : * prepare for truncation of the visibility map
22 : *
23 : * NOTES
24 : *
25 : * The visibility map is a bitmap with two bits (all-visible and all-frozen)
26 : * per heap page. A set all-visible bit means that all tuples on the page are
27 : * known visible to all transactions, and therefore the page doesn't need to
28 : * be vacuumed. A set all-frozen bit means that all tuples on the page are
29 : * completely frozen, and therefore the page doesn't need to be vacuumed even
30 : * if whole table scanning vacuum is required (e.g. anti-wraparound vacuum).
31 : * The all-frozen bit must be set only when the page is already all-visible.
32 : *
33 : * The map is conservative in the sense that we make sure that whenever a bit
34 : * is set, we know the condition is true, but if a bit is not set, it might or
35 : * might not be true.
36 : *
37 : * Changes to the visibility map bits are not separately WAL-logged. Callers
38 : * must make sure that whenever a visibility map bit is cleared, the bit is
39 : * cleared on WAL replay of the updating operation. And whenever a visibility
40 : * map bit is set, the bit is set on WAL replay of the operation that rendered
41 : * the page all-visible/all-frozen.
42 : *
43 : * The visibility map bits operate as a hint in one direction: if they are
44 : * clear, it may still be the case that every tuple on the page is visible to
45 : * all transactions (we just don't know that for certain). However, if they
46 : * are set, we may skip vacuuming pages and advance relfrozenxid or skip
47 : * reading heap pages for an index-only scan. If they are incorrectly set,
48 : * this can lead to data corruption and wrong results.
49 : *
50 : * Additionally, it is critical that the heap-page level PD_ALL_VISIBLE bit be
51 : * correctly set and cleared along with the VM bits.
52 : *
53 : * When clearing the VM, if a crash occurs after the heap page makes it to
54 : * disk but before the VM page makes it to disk, replay must clear the VM or
55 : * the next index-only scan can return wrong results or vacuum may incorrectly
56 : * advance relfrozenxid.
57 : *
58 : * When setting the VM, if a crash occurs after the visibility map page makes
59 : * it to disk and before the updated heap page makes it to disk, redo must set
60 : * the bit on the heap page. Otherwise, the next insert, update, or delete on
61 : * the heap page will fail to realize that the visibility map bit must be
62 : * cleared, possibly causing index-only scans to return wrong answers.
63 : *
64 : * VACUUM will normally skip pages for which the visibility map bit is set;
65 : * such pages can't contain any dead tuples and therefore don't need vacuuming.
66 : *
67 : * LOCKING
68 : *
69 : * In heapam.c, whenever a page is modified so that not all tuples on the
70 : * page are visible to everyone anymore, the corresponding bit in the
71 : * visibility map is cleared. In order to be crash-safe, we need to do this
72 : * while still holding a lock on the heap page and in the same critical
73 : * section that logs the page modification. However, we don't want to hold
74 : * the buffer lock over any I/O that may be required to read in the visibility
75 : * map page. To avoid this, we examine the heap page before locking it;
76 : * if the page-level PD_ALL_VISIBLE bit is set, we pin the visibility map
77 : * bit. Then, we lock the buffer. But this creates a race condition: there
78 : * is a possibility that in the time it takes to lock the buffer, the
79 : * PD_ALL_VISIBLE bit gets set. If that happens, we have to unlock the
80 : * buffer, pin the visibility map page, and relock the buffer. This shouldn't
81 : * happen often, because only VACUUM currently sets visibility map bits,
82 : * and the race will only occur if VACUUM processes a given page at almost
83 : * exactly the same time that someone tries to further modify it.
84 : *
85 : * To set a bit, you need to hold a lock on the heap page. That prevents
86 : * the race condition where VACUUM sees that all tuples on the page are
87 : * visible to everyone, but another backend modifies the page before VACUUM
88 : * sets the bit in the visibility map.
89 : *
90 : * When a bit is set, the LSN of the visibility map page is updated to make
91 : * sure that the visibility map update doesn't get written to disk before the
92 : * WAL record of the changes that made it possible to set the bit is flushed.
93 : * But when a bit is cleared, we don't have to do that because it's always
94 : * safe to clear a bit in the map from correctness point of view.
95 : *
96 : *-------------------------------------------------------------------------
97 : */
98 : #include "postgres.h"
99 :
100 : #include "access/heapam_xlog.h"
101 : #include "access/visibilitymap.h"
102 : #include "access/xloginsert.h"
103 : #include "access/xlogutils.h"
104 : #include "miscadmin.h"
105 : #include "port/pg_bitutils.h"
106 : #include "storage/bufmgr.h"
107 : #include "storage/smgr.h"
108 : #include "utils/inval.h"
109 : #include "utils/rel.h"
110 :
111 :
112 : /*#define TRACE_VISIBILITYMAP */
113 :
114 : /*
115 : * Size of the bitmap on each visibility map page, in bytes. There's no
116 : * extra headers, so the whole page minus the standard page header is
117 : * used for the bitmap.
118 : */
119 : #define MAPSIZE (BLCKSZ - MAXALIGN(SizeOfPageHeaderData))
120 :
121 : /* Number of heap blocks we can represent in one byte */
122 : #define HEAPBLOCKS_PER_BYTE (BITS_PER_BYTE / BITS_PER_HEAPBLOCK)
123 :
124 : /* Number of heap blocks we can represent in one visibility map page. */
125 : #define HEAPBLOCKS_PER_PAGE (MAPSIZE * HEAPBLOCKS_PER_BYTE)
126 :
127 : /* Mapping from heap block number to the right bit in the visibility map */
128 : #define HEAPBLK_TO_MAPBLOCK(x) ((x) / HEAPBLOCKS_PER_PAGE)
129 : #define HEAPBLK_TO_MAPBLOCK_LIMIT(x) \
130 : (((x) + HEAPBLOCKS_PER_PAGE - 1) / HEAPBLOCKS_PER_PAGE)
131 : #define HEAPBLK_TO_MAPBYTE(x) (((x) % HEAPBLOCKS_PER_PAGE) / HEAPBLOCKS_PER_BYTE)
132 : #define HEAPBLK_TO_OFFSET(x) (((x) % HEAPBLOCKS_PER_BYTE) * BITS_PER_HEAPBLOCK)
133 :
134 : /* Masks for counting subsets of bits in the visibility map. */
135 : #define VISIBLE_MASK8 (0x55) /* The lower bit of each bit pair */
136 : #define FROZEN_MASK8 (0xaa) /* The upper bit of each bit pair */
137 :
138 : /* prototypes for internal routines */
139 : static Buffer vm_readbuf(Relation rel, BlockNumber blkno, bool extend);
140 : static Buffer vm_extend(Relation rel, BlockNumber vm_nblocks);
141 :
142 :
143 : /*
144 : * visibilitymap_clear - clear specified bits for one page in visibility map
145 : *
146 : * You must pass a buffer containing the correct map page to this function.
147 : * Call visibilitymap_pin first to pin the right one. This function doesn't do
148 : * any I/O. Returns true if any bits have been cleared and false otherwise.
149 : */
150 : bool
151 19933 : visibilitymap_clear(Relation rel, BlockNumber heapBlk, Buffer vmbuf, uint8 flags)
152 : {
153 19933 : BlockNumber mapBlock = HEAPBLK_TO_MAPBLOCK(heapBlk);
154 19933 : int mapByte = HEAPBLK_TO_MAPBYTE(heapBlk);
155 19933 : int mapOffset = HEAPBLK_TO_OFFSET(heapBlk);
156 19933 : uint8 mask = flags << mapOffset;
157 : char *map;
158 19933 : bool cleared = false;
159 :
160 : /* Must never clear all_visible bit while leaving all_frozen bit set */
161 : Assert(flags & VISIBILITYMAP_VALID_BITS);
162 : Assert(flags != VISIBILITYMAP_ALL_VISIBLE);
163 :
164 : #ifdef TRACE_VISIBILITYMAP
165 : elog(DEBUG1, "vm_clear %s %d", RelationGetRelationName(rel), heapBlk);
166 : #endif
167 :
168 19933 : if (!BufferIsValid(vmbuf) || BufferGetBlockNumber(vmbuf) != mapBlock)
169 0 : elog(ERROR, "wrong buffer passed to visibilitymap_clear");
170 :
171 19933 : LockBuffer(vmbuf, BUFFER_LOCK_EXCLUSIVE);
172 19933 : map = PageGetContents(BufferGetPage(vmbuf));
173 :
174 19933 : if (map[mapByte] & mask)
175 : {
176 17529 : map[mapByte] &= ~mask;
177 :
178 17529 : MarkBufferDirty(vmbuf);
179 17529 : cleared = true;
180 : }
181 :
182 19933 : LockBuffer(vmbuf, BUFFER_LOCK_UNLOCK);
183 :
184 19933 : return cleared;
185 : }
186 :
187 : /*
188 : * visibilitymap_pin - pin a map page for setting a bit
189 : *
190 : * Setting a bit in the visibility map is a two-phase operation. First, call
191 : * visibilitymap_pin, to pin the visibility map page containing the bit for
192 : * the heap page. Because that can require I/O to read the map page, you
193 : * shouldn't hold a lock on the heap page while doing that. Then, call
194 : * visibilitymap_set to actually set the bit.
195 : *
196 : * On entry, *vmbuf should be InvalidBuffer or a valid buffer returned by
197 : * an earlier call to visibilitymap_pin or visibilitymap_get_status on the same
198 : * relation. On return, *vmbuf is a valid buffer with the map page containing
199 : * the bit for heapBlk.
200 : *
201 : * If the page doesn't exist in the map file yet, it is extended.
202 : */
203 : void
204 886779 : visibilitymap_pin(Relation rel, BlockNumber heapBlk, Buffer *vmbuf)
205 : {
206 886779 : BlockNumber mapBlock = HEAPBLK_TO_MAPBLOCK(heapBlk);
207 :
208 : /* Reuse the old pinned buffer if possible */
209 886779 : if (BufferIsValid(*vmbuf))
210 : {
211 387462 : if (BufferGetBlockNumber(*vmbuf) == mapBlock)
212 387462 : return;
213 :
214 0 : ReleaseBuffer(*vmbuf);
215 : }
216 499317 : *vmbuf = vm_readbuf(rel, mapBlock, true);
217 : }
218 :
219 : /*
220 : * visibilitymap_pin_ok - do we already have the correct page pinned?
221 : *
222 : * On entry, vmbuf should be InvalidBuffer or a valid buffer returned by
223 : * an earlier call to visibilitymap_pin or visibilitymap_get_status on the same
224 : * relation. The return value indicates whether the buffer covers the
225 : * given heapBlk.
226 : */
227 : bool
228 15535 : visibilitymap_pin_ok(BlockNumber heapBlk, Buffer vmbuf)
229 : {
230 15535 : BlockNumber mapBlock = HEAPBLK_TO_MAPBLOCK(heapBlk);
231 :
232 15535 : return BufferIsValid(vmbuf) && BufferGetBlockNumber(vmbuf) == mapBlock;
233 : }
234 :
235 : /*
236 : * Set VM (visibility map) flags in the VM block in vmBuf.
237 : *
238 : * This function is intended for callers that log VM changes together
239 : * with the heap page modifications that rendered the page all-visible.
240 : *
241 : * vmBuf must be pinned and exclusively locked, and it must cover the VM bits
242 : * corresponding to heapBlk.
243 : *
244 : * In normal operation (not recovery), this must be called inside a critical
245 : * section that also applies the necessary heap page changes and, if
246 : * applicable, emits WAL.
247 : *
248 : * The caller is responsible for ensuring consistency between the heap page
249 : * and the VM page by holding a pin and exclusive lock on the buffer
250 : * containing heapBlk.
251 : *
252 : * rlocator is used only for debugging messages.
253 : */
254 : void
255 71176 : visibilitymap_set(BlockNumber heapBlk,
256 : Buffer vmBuf, uint8 flags,
257 : const RelFileLocator rlocator)
258 : {
259 71176 : BlockNumber mapBlock = HEAPBLK_TO_MAPBLOCK(heapBlk);
260 71176 : uint32 mapByte = HEAPBLK_TO_MAPBYTE(heapBlk);
261 71176 : uint8 mapOffset = HEAPBLK_TO_OFFSET(heapBlk);
262 : Page page;
263 : uint8 *map;
264 : uint8 status;
265 :
266 : #ifdef TRACE_VISIBILITYMAP
267 : elog(DEBUG1, "vm_set flags 0x%02X for %s %d",
268 : flags,
269 : relpathbackend(rlocator, MyProcNumber, MAIN_FORKNUM).str,
270 : heapBlk);
271 : #endif
272 :
273 : /* Call in same critical section where WAL is emitted. */
274 : Assert(InRecovery || CritSectionCount > 0);
275 :
276 : /* Flags should be valid. Also never clear bits with this function */
277 : Assert((flags & VISIBILITYMAP_VALID_BITS) == flags);
278 :
279 : /* Must never set all_frozen bit without also setting all_visible bit */
280 : Assert(flags != VISIBILITYMAP_ALL_FROZEN);
281 :
282 : /* Check that we have the right VM page pinned */
283 71176 : if (!BufferIsValid(vmBuf) || BufferGetBlockNumber(vmBuf) != mapBlock)
284 0 : elog(ERROR, "wrong VM buffer passed to visibilitymap_set");
285 :
286 : Assert(BufferIsLockedByMeInMode(vmBuf, BUFFER_LOCK_EXCLUSIVE));
287 :
288 71176 : page = BufferGetPage(vmBuf);
289 71176 : map = (uint8 *) PageGetContents(page);
290 :
291 71176 : status = (map[mapByte] >> mapOffset) & VISIBILITYMAP_VALID_BITS;
292 71176 : if (flags != status)
293 : {
294 71176 : map[mapByte] |= (flags << mapOffset);
295 71176 : MarkBufferDirty(vmBuf);
296 : }
297 71176 : }
298 :
299 : /*
300 : * visibilitymap_get_status - get status of bits
301 : *
302 : * Are all tuples on heapBlk visible to all or are marked frozen, according
303 : * to the visibility map?
304 : *
305 : * On entry, *vmbuf should be InvalidBuffer or a valid buffer returned by an
306 : * earlier call to visibilitymap_pin or visibilitymap_get_status on the same
307 : * relation. On return, *vmbuf is a valid buffer with the map page containing
308 : * the bit for heapBlk, or InvalidBuffer. The caller is responsible for
309 : * releasing *vmbuf after it's done testing and setting bits.
310 : *
311 : * NOTE: This function is typically called without a lock on the heap page,
312 : * so somebody else could change the bit just after we look at it. In fact,
313 : * since we don't lock the visibility map page either, it's even possible that
314 : * someone else could have changed the bit just before we look at it, but yet
315 : * we might see the old value. It is the caller's responsibility to deal with
316 : * all concurrency issues!
317 : */
318 : uint8
319 4937207 : visibilitymap_get_status(Relation rel, BlockNumber heapBlk, Buffer *vmbuf)
320 : {
321 4937207 : BlockNumber mapBlock = HEAPBLK_TO_MAPBLOCK(heapBlk);
322 4937207 : uint32 mapByte = HEAPBLK_TO_MAPBYTE(heapBlk);
323 4937207 : uint8 mapOffset = HEAPBLK_TO_OFFSET(heapBlk);
324 : char *map;
325 : uint8 result;
326 :
327 : #ifdef TRACE_VISIBILITYMAP
328 : elog(DEBUG1, "vm_get_status %s %d", RelationGetRelationName(rel), heapBlk);
329 : #endif
330 :
331 : /* Reuse the old pinned buffer if possible */
332 4937207 : if (BufferIsValid(*vmbuf))
333 : {
334 3817519 : if (BufferGetBlockNumber(*vmbuf) != mapBlock)
335 : {
336 0 : ReleaseBuffer(*vmbuf);
337 0 : *vmbuf = InvalidBuffer;
338 : }
339 : }
340 :
341 4937207 : if (!BufferIsValid(*vmbuf))
342 : {
343 1119688 : *vmbuf = vm_readbuf(rel, mapBlock, false);
344 1119688 : if (!BufferIsValid(*vmbuf))
345 999474 : return (uint8) 0;
346 : }
347 :
348 3937733 : map = PageGetContents(BufferGetPage(*vmbuf));
349 :
350 : /*
351 : * A single byte read is atomic. There could be memory-ordering effects
352 : * here, but for performance reasons we make it the caller's job to worry
353 : * about that.
354 : */
355 3937733 : result = ((map[mapByte] >> mapOffset) & VISIBILITYMAP_VALID_BITS);
356 3937733 : return result;
357 : }
358 :
359 : /*
360 : * visibilitymap_count - count number of bits set in visibility map
361 : *
362 : * Note: we ignore the possibility of race conditions when the table is being
363 : * extended concurrently with the call. New pages added to the table aren't
364 : * going to be marked all-visible or all-frozen, so they won't affect the result.
365 : */
366 : void
367 132648 : visibilitymap_count(Relation rel, BlockNumber *all_visible, BlockNumber *all_frozen)
368 : {
369 : BlockNumber mapBlock;
370 132648 : BlockNumber nvisible = 0;
371 132648 : BlockNumber nfrozen = 0;
372 :
373 : /* all_visible must be specified */
374 : Assert(all_visible);
375 :
376 132648 : for (mapBlock = 0;; mapBlock++)
377 46942 : {
378 : Buffer mapBuffer;
379 : uint64 *map;
380 :
381 : /*
382 : * Read till we fall off the end of the map. We assume that any extra
383 : * bytes in the last page are zeroed, so we don't bother excluding
384 : * them from the count.
385 : */
386 179590 : mapBuffer = vm_readbuf(rel, mapBlock, false);
387 179590 : if (!BufferIsValid(mapBuffer))
388 132648 : break;
389 :
390 : /*
391 : * We choose not to lock the page, since the result is going to be
392 : * immediately stale anyway if anyone is concurrently setting or
393 : * clearing bits, and we only really need an approximate value.
394 : */
395 46942 : map = (uint64 *) PageGetContents(BufferGetPage(mapBuffer));
396 :
397 46942 : nvisible += pg_popcount_masked((const char *) map, MAPSIZE, VISIBLE_MASK8);
398 46942 : if (all_frozen)
399 46942 : nfrozen += pg_popcount_masked((const char *) map, MAPSIZE, FROZEN_MASK8);
400 :
401 46942 : ReleaseBuffer(mapBuffer);
402 : }
403 :
404 132648 : *all_visible = nvisible;
405 132648 : if (all_frozen)
406 132648 : *all_frozen = nfrozen;
407 132648 : }
408 :
409 : /*
410 : * visibilitymap_prepare_truncate -
411 : * prepare for truncation of the visibility map
412 : *
413 : * nheapblocks is the new size of the heap.
414 : *
415 : * Return the number of blocks of new visibility map.
416 : * If it's InvalidBlockNumber, there is nothing to truncate;
417 : * otherwise the caller is responsible for calling smgrtruncate()
418 : * to truncate the visibility map pages.
419 : */
420 : BlockNumber
421 230 : visibilitymap_prepare_truncate(Relation rel, BlockNumber nheapblocks)
422 : {
423 : BlockNumber newnblocks;
424 :
425 : /* last remaining block, byte, and bit */
426 230 : BlockNumber truncBlock = HEAPBLK_TO_MAPBLOCK(nheapblocks);
427 230 : uint32 truncByte = HEAPBLK_TO_MAPBYTE(nheapblocks);
428 230 : uint8 truncOffset = HEAPBLK_TO_OFFSET(nheapblocks);
429 :
430 : #ifdef TRACE_VISIBILITYMAP
431 : elog(DEBUG1, "vm_truncate %s %d", RelationGetRelationName(rel), nheapblocks);
432 : #endif
433 :
434 : /*
435 : * If no visibility map has been created yet for this relation, there's
436 : * nothing to truncate.
437 : */
438 230 : if (!smgrexists(RelationGetSmgr(rel), VISIBILITYMAP_FORKNUM))
439 0 : return InvalidBlockNumber;
440 :
441 : /*
442 : * Unless the new size is exactly at a visibility map page boundary, the
443 : * tail bits in the last remaining map page, representing truncated heap
444 : * blocks, need to be cleared. This is not only tidy, but also necessary
445 : * because we don't get a chance to clear the bits if the heap is extended
446 : * again.
447 : */
448 230 : if (truncByte != 0 || truncOffset != 0)
449 133 : {
450 : Buffer mapBuffer;
451 : Page page;
452 : char *map;
453 :
454 133 : newnblocks = truncBlock + 1;
455 :
456 133 : mapBuffer = vm_readbuf(rel, truncBlock, false);
457 133 : if (!BufferIsValid(mapBuffer))
458 : {
459 : /* nothing to do, the file was already smaller */
460 0 : return InvalidBlockNumber;
461 : }
462 :
463 133 : page = BufferGetPage(mapBuffer);
464 133 : map = PageGetContents(page);
465 :
466 133 : LockBuffer(mapBuffer, BUFFER_LOCK_EXCLUSIVE);
467 :
468 : /* NO EREPORT(ERROR) from here till changes are logged */
469 133 : START_CRIT_SECTION();
470 :
471 : /* Clear out the unwanted bytes. */
472 133 : MemSet(&map[truncByte + 1], 0, MAPSIZE - (truncByte + 1));
473 :
474 : /*----
475 : * Mask out the unwanted bits of the last remaining byte.
476 : *
477 : * ((1 << 0) - 1) = 00000000
478 : * ((1 << 1) - 1) = 00000001
479 : * ...
480 : * ((1 << 6) - 1) = 00111111
481 : * ((1 << 7) - 1) = 01111111
482 : *----
483 : */
484 133 : map[truncByte] &= (1 << truncOffset) - 1;
485 :
486 : /*
487 : * Truncation of a relation is WAL-logged at a higher-level, and we
488 : * will be called at WAL replay. But if checksums are enabled, we need
489 : * to still write a WAL record to protect against a torn page, if the
490 : * page is flushed to disk before the truncation WAL record. We cannot
491 : * use MarkBufferDirtyHint here, because that will not dirty the page
492 : * during recovery.
493 : */
494 133 : MarkBufferDirty(mapBuffer);
495 133 : if (!InRecovery && RelationNeedsWAL(rel) && XLogHintBitIsNeeded())
496 111 : log_newpage_buffer(mapBuffer, false);
497 :
498 133 : END_CRIT_SECTION();
499 :
500 133 : UnlockReleaseBuffer(mapBuffer);
501 : }
502 : else
503 97 : newnblocks = truncBlock;
504 :
505 230 : if (smgrnblocks(RelationGetSmgr(rel), VISIBILITYMAP_FORKNUM) <= newnblocks)
506 : {
507 : /* nothing to do, the file was already smaller than requested size */
508 133 : return InvalidBlockNumber;
509 : }
510 :
511 97 : return newnblocks;
512 : }
513 :
514 : /*
515 : * visibilitymap_truncation_length -
516 : * compute truncation length for visibility map
517 : *
518 : * Given a proposed truncation length for the main fork, compute the
519 : * correct truncation length for the visibility map. Should return the
520 : * same answer as visibilitymap_prepare_truncate(), but without modifying
521 : * anything.
522 : */
523 : BlockNumber
524 1 : visibilitymap_truncation_length(BlockNumber nheapblocks)
525 : {
526 1 : return HEAPBLK_TO_MAPBLOCK_LIMIT(nheapblocks);
527 : }
528 :
529 : /*
530 : * Read a visibility map page.
531 : *
532 : * If the page doesn't exist, InvalidBuffer is returned, or if 'extend' is
533 : * true, the visibility map file is extended.
534 : */
535 : static Buffer
536 1798728 : vm_readbuf(Relation rel, BlockNumber blkno, bool extend)
537 : {
538 : Buffer buf;
539 : SMgrRelation reln;
540 :
541 : /*
542 : * Caution: re-using this smgr pointer could fail if the relcache entry
543 : * gets closed. It's safe as long as we only do smgr-level operations
544 : * between here and the last use of the pointer.
545 : */
546 1798728 : reln = RelationGetSmgr(rel);
547 :
548 : /*
549 : * If we haven't cached the size of the visibility map fork yet, check it
550 : * first.
551 : */
552 1798728 : if (reln->smgr_cached_nblocks[VISIBILITYMAP_FORKNUM] == InvalidBlockNumber)
553 : {
554 151201 : if (smgrexists(reln, VISIBILITYMAP_FORKNUM))
555 63276 : smgrnblocks(reln, VISIBILITYMAP_FORKNUM);
556 : else
557 87925 : reln->smgr_cached_nblocks[VISIBILITYMAP_FORKNUM] = 0;
558 : }
559 :
560 : /*
561 : * For reading we use ZERO_ON_ERROR mode, and initialize the page if
562 : * necessary. It's always safe to clear bits, so it's better to clear
563 : * corrupt pages than error out.
564 : *
565 : * We use the same path below to initialize pages when extending the
566 : * relation, as a concurrent extension can end up with vm_extend()
567 : * returning an already-initialized page.
568 : */
569 1798728 : if (blkno >= reln->smgr_cached_nblocks[VISIBILITYMAP_FORKNUM])
570 : {
571 1135586 : if (extend)
572 3464 : buf = vm_extend(rel, blkno + 1);
573 : else
574 1132122 : return InvalidBuffer;
575 : }
576 : else
577 663142 : buf = ReadBufferExtended(rel, VISIBILITYMAP_FORKNUM, blkno,
578 : RBM_ZERO_ON_ERROR, NULL);
579 :
580 : /*
581 : * Initializing the page when needed is trickier than it looks, because of
582 : * the possibility of multiple backends doing this concurrently, and our
583 : * desire to not uselessly take the buffer lock in the normal path where
584 : * the page is OK. We must take the lock to initialize the page, so
585 : * recheck page newness after we have the lock, in case someone else
586 : * already did it. Also, because we initially check PageIsNew with no
587 : * lock, it's possible to fall through and return the buffer while someone
588 : * else is still initializing the page (i.e., we might see pd_upper as set
589 : * but other page header fields are still zeroes). This is harmless for
590 : * callers that will take a buffer lock themselves, but some callers
591 : * inspect the page without any lock at all. The latter is OK only so
592 : * long as it doesn't depend on the page header having correct contents.
593 : * Current usage is safe because PageGetContents() does not require that.
594 : */
595 666606 : if (PageIsNew(BufferGetPage(buf)))
596 : {
597 3751 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
598 3751 : if (PageIsNew(BufferGetPage(buf)))
599 3751 : PageInit(BufferGetPage(buf), BLCKSZ, 0);
600 3751 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
601 : }
602 666606 : return buf;
603 : }
604 :
605 : /*
606 : * Ensure that the visibility map fork is at least vm_nblocks long, extending
607 : * it if necessary with zeroed pages.
608 : */
609 : static Buffer
610 3464 : vm_extend(Relation rel, BlockNumber vm_nblocks)
611 : {
612 : Buffer buf;
613 :
614 3464 : buf = ExtendBufferedRelTo(BMR_REL(rel), VISIBILITYMAP_FORKNUM, NULL,
615 : EB_CREATE_FORK_IF_NEEDED |
616 : EB_CLEAR_SIZE_CACHE,
617 : vm_nblocks,
618 : RBM_ZERO_ON_ERROR);
619 :
620 : /*
621 : * Send a shared-inval message to force other backends to close any smgr
622 : * references they may have for this rel, which we are about to change.
623 : * This is a useful optimization because it means that backends don't have
624 : * to keep checking for creation or extension of the file, which happens
625 : * infrequently.
626 : */
627 3464 : CacheInvalidateSmgr(RelationGetSmgr(rel)->smgr_rlocator);
628 :
629 3464 : return buf;
630 : }
|
