1 /*-------------------------------------------------------------------------
2  *
3  * visibilitymap.c
4  *	  bitmap for tracking visibility of heap tuples
5  *
6  * Portions Copyright (c) 1996-2021, 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 a bit 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  * Clearing visibility map bits is not separately WAL-logged.  The callers
38  * must make sure that whenever a bit is cleared, the bit is cleared on WAL
39  * replay of the updating operation as well.
40  *
41  * When we *set* a visibility map during VACUUM, we must write WAL.  This may
42  * seem counterintuitive, since the bit is basically a hint: if it is clear,
43  * it may still be the case that every tuple on the page is visible to all
44  * transactions; we just don't know that for certain.  The difficulty is that
45  * there are two bits which are typically set together: the PD_ALL_VISIBLE bit
46  * on the page itself, and the visibility map bit.  If a crash occurs after the
47  * visibility map page makes it to disk and before the updated heap page makes
48  * it to disk, redo must set the bit on the heap page.  Otherwise, the next
49  * insert, update, or delete on the heap page will fail to realize that the
50  * visibility map bit must be cleared, possibly causing index-only scans to
51  * return wrong answers.
52  *
53  * VACUUM will normally skip pages for which the visibility map bit is set;
54  * such pages can't contain any dead tuples and therefore don't need vacuuming.
55  *
56  * LOCKING
57  *
58  * In heapam.c, whenever a page is modified so that not all tuples on the
59  * page are visible to everyone anymore, the corresponding bit in the
60  * visibility map is cleared. In order to be crash-safe, we need to do this
61  * while still holding a lock on the heap page and in the same critical
62  * section that logs the page modification. However, we don't want to hold
63  * the buffer lock over any I/O that may be required to read in the visibility
64  * map page.  To avoid this, we examine the heap page before locking it;
65  * if the page-level PD_ALL_VISIBLE bit is set, we pin the visibility map
66  * bit.  Then, we lock the buffer.  But this creates a race condition: there
67  * is a possibility that in the time it takes to lock the buffer, the
68  * PD_ALL_VISIBLE bit gets set.  If that happens, we have to unlock the
69  * buffer, pin the visibility map page, and relock the buffer.  This shouldn't
70  * happen often, because only VACUUM currently sets visibility map bits,
71  * and the race will only occur if VACUUM processes a given page at almost
72  * exactly the same time that someone tries to further modify it.
73  *
74  * To set a bit, you need to hold a lock on the heap page. That prevents
75  * the race condition where VACUUM sees that all tuples on the page are
76  * visible to everyone, but another backend modifies the page before VACUUM
77  * sets the bit in the visibility map.
78  *
79  * When a bit is set, the LSN of the visibility map page is updated to make
80  * sure that the visibility map update doesn't get written to disk before the
81  * WAL record of the changes that made it possible to set the bit is flushed.
82  * But when a bit is cleared, we don't have to do that because it's always
83  * safe to clear a bit in the map from correctness point of view.
84  *
85  *-------------------------------------------------------------------------
86  */
87 #include "postgres.h"
88 
89 #include "access/heapam_xlog.h"
90 #include "access/visibilitymap.h"
91 #include "access/xlog.h"
92 #include "miscadmin.h"
93 #include "port/pg_bitutils.h"
94 #include "storage/bufmgr.h"
95 #include "storage/lmgr.h"
96 #include "storage/smgr.h"
97 #include "utils/inval.h"
98 
99 
100 /*#define TRACE_VISIBILITYMAP */
101 
102 /*
103  * Size of the bitmap on each visibility map page, in bytes. There's no
104  * extra headers, so the whole page minus the standard page header is
105  * used for the bitmap.
106  */
107 #define MAPSIZE (BLCKSZ - MAXALIGN(SizeOfPageHeaderData))
108 
109 /* Number of heap blocks we can represent in one byte */
110 #define HEAPBLOCKS_PER_BYTE (BITS_PER_BYTE / BITS_PER_HEAPBLOCK)
111 
112 /* Number of heap blocks we can represent in one visibility map page. */
113 #define HEAPBLOCKS_PER_PAGE (MAPSIZE * HEAPBLOCKS_PER_BYTE)
114 
115 /* Mapping from heap block number to the right bit in the visibility map */
116 #define HEAPBLK_TO_MAPBLOCK(x) ((x) / HEAPBLOCKS_PER_PAGE)
117 #define HEAPBLK_TO_MAPBYTE(x) (((x) % HEAPBLOCKS_PER_PAGE) / HEAPBLOCKS_PER_BYTE)
118 #define HEAPBLK_TO_OFFSET(x) (((x) % HEAPBLOCKS_PER_BYTE) * BITS_PER_HEAPBLOCK)
119 
120 /* Masks for counting subsets of bits in the visibility map. */
121 #define VISIBLE_MASK64	UINT64CONST(0x5555555555555555) /* The lower bit of each
122 														 * bit pair */
123 #define FROZEN_MASK64	UINT64CONST(0xaaaaaaaaaaaaaaaa) /* The upper bit of each
124 														 * bit pair */
125 
126 /* prototypes for internal routines */
127 static Buffer vm_readbuf(Relation rel, BlockNumber blkno, bool extend);
128 static void vm_extend(Relation rel, BlockNumber vm_nblocks);
129 
130 
131 /*
132  *	visibilitymap_clear - clear specified bits for one page in visibility map
133  *
134  * You must pass a buffer containing the correct map page to this function.
135  * Call visibilitymap_pin first to pin the right one. This function doesn't do
136  * any I/O.  Returns true if any bits have been cleared and false otherwise.
137  */
138 bool
visibilitymap_clear(Relation rel,BlockNumber heapBlk,Buffer buf,uint8 flags)139 visibilitymap_clear(Relation rel, BlockNumber heapBlk, Buffer buf, uint8 flags)
140 {
141 	BlockNumber mapBlock = HEAPBLK_TO_MAPBLOCK(heapBlk);
142 	int			mapByte = HEAPBLK_TO_MAPBYTE(heapBlk);
143 	int			mapOffset = HEAPBLK_TO_OFFSET(heapBlk);
144 	uint8		mask = flags << mapOffset;
145 	char	   *map;
146 	bool		cleared = false;
147 
148 	Assert(flags & VISIBILITYMAP_VALID_BITS);
149 
150 #ifdef TRACE_VISIBILITYMAP
151 	elog(DEBUG1, "vm_clear %s %d", RelationGetRelationName(rel), heapBlk);
152 #endif
153 
154 	if (!BufferIsValid(buf) || BufferGetBlockNumber(buf) != mapBlock)
155 		elog(ERROR, "wrong buffer passed to visibilitymap_clear");
156 
157 	LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
158 	map = PageGetContents(BufferGetPage(buf));
159 
160 	if (map[mapByte] & mask)
161 	{
162 		map[mapByte] &= ~mask;
163 
164 		MarkBufferDirty(buf);
165 		cleared = true;
166 	}
167 
168 	LockBuffer(buf, BUFFER_LOCK_UNLOCK);
169 
170 	return cleared;
171 }
172 
173 /*
174  *	visibilitymap_pin - pin a map page for setting a bit
175  *
176  * Setting a bit in the visibility map is a two-phase operation. First, call
177  * visibilitymap_pin, to pin the visibility map page containing the bit for
178  * the heap page. Because that can require I/O to read the map page, you
179  * shouldn't hold a lock on the heap page while doing that. Then, call
180  * visibilitymap_set to actually set the bit.
181  *
182  * On entry, *buf should be InvalidBuffer or a valid buffer returned by
183  * an earlier call to visibilitymap_pin or visibilitymap_get_status on the same
184  * relation. On return, *buf is a valid buffer with the map page containing
185  * the bit for heapBlk.
186  *
187  * If the page doesn't exist in the map file yet, it is extended.
188  */
189 void
visibilitymap_pin(Relation rel,BlockNumber heapBlk,Buffer * buf)190 visibilitymap_pin(Relation rel, BlockNumber heapBlk, Buffer *buf)
191 {
192 	BlockNumber mapBlock = HEAPBLK_TO_MAPBLOCK(heapBlk);
193 
194 	/* Reuse the old pinned buffer if possible */
195 	if (BufferIsValid(*buf))
196 	{
197 		if (BufferGetBlockNumber(*buf) == mapBlock)
198 			return;
199 
200 		ReleaseBuffer(*buf);
201 	}
202 	*buf = vm_readbuf(rel, mapBlock, true);
203 }
204 
205 /*
206  *	visibilitymap_pin_ok - do we already have the correct page pinned?
207  *
208  * On entry, buf should be InvalidBuffer or a valid buffer returned by
209  * an earlier call to visibilitymap_pin or visibilitymap_get_status on the same
210  * relation.  The return value indicates whether the buffer covers the
211  * given heapBlk.
212  */
213 bool
visibilitymap_pin_ok(BlockNumber heapBlk,Buffer buf)214 visibilitymap_pin_ok(BlockNumber heapBlk, Buffer buf)
215 {
216 	BlockNumber mapBlock = HEAPBLK_TO_MAPBLOCK(heapBlk);
217 
218 	return BufferIsValid(buf) && BufferGetBlockNumber(buf) == mapBlock;
219 }
220 
221 /*
222  *	visibilitymap_set - set bit(s) on a previously pinned page
223  *
224  * recptr is the LSN of the XLOG record we're replaying, if we're in recovery,
225  * or InvalidXLogRecPtr in normal running.  The page LSN is advanced to the
226  * one provided; in normal running, we generate a new XLOG record and set the
227  * page LSN to that value.  cutoff_xid is the largest xmin on the page being
228  * marked all-visible; it is needed for Hot Standby, and can be
229  * InvalidTransactionId if the page contains no tuples.  It can also be set
230  * to InvalidTransactionId when a page that is already all-visible is being
231  * marked all-frozen.
232  *
233  * Caller is expected to set the heap page's PD_ALL_VISIBLE bit before calling
234  * this function. Except in recovery, caller should also pass the heap
235  * buffer. When checksums are enabled and we're not in recovery, we must add
236  * the heap buffer to the WAL chain to protect it from being torn.
237  *
238  * You must pass a buffer containing the correct map page to this function.
239  * Call visibilitymap_pin first to pin the right one. This function doesn't do
240  * any I/O.
241  */
242 void
visibilitymap_set(Relation rel,BlockNumber heapBlk,Buffer heapBuf,XLogRecPtr recptr,Buffer vmBuf,TransactionId cutoff_xid,uint8 flags)243 visibilitymap_set(Relation rel, BlockNumber heapBlk, Buffer heapBuf,
244 				  XLogRecPtr recptr, Buffer vmBuf, TransactionId cutoff_xid,
245 				  uint8 flags)
246 {
247 	BlockNumber mapBlock = HEAPBLK_TO_MAPBLOCK(heapBlk);
248 	uint32		mapByte = HEAPBLK_TO_MAPBYTE(heapBlk);
249 	uint8		mapOffset = HEAPBLK_TO_OFFSET(heapBlk);
250 	Page		page;
251 	uint8	   *map;
252 
253 #ifdef TRACE_VISIBILITYMAP
254 	elog(DEBUG1, "vm_set %s %d", RelationGetRelationName(rel), heapBlk);
255 #endif
256 
257 	Assert(InRecovery || XLogRecPtrIsInvalid(recptr));
258 	Assert(InRecovery || BufferIsValid(heapBuf));
259 	Assert(flags & VISIBILITYMAP_VALID_BITS);
260 
261 	/* Check that we have the right heap page pinned, if present */
262 	if (BufferIsValid(heapBuf) && BufferGetBlockNumber(heapBuf) != heapBlk)
263 		elog(ERROR, "wrong heap buffer passed to visibilitymap_set");
264 
265 	/* Check that we have the right VM page pinned */
266 	if (!BufferIsValid(vmBuf) || BufferGetBlockNumber(vmBuf) != mapBlock)
267 		elog(ERROR, "wrong VM buffer passed to visibilitymap_set");
268 
269 	page = BufferGetPage(vmBuf);
270 	map = (uint8 *) PageGetContents(page);
271 	LockBuffer(vmBuf, BUFFER_LOCK_EXCLUSIVE);
272 
273 	if (flags != (map[mapByte] >> mapOffset & VISIBILITYMAP_VALID_BITS))
274 	{
275 		START_CRIT_SECTION();
276 
277 		map[mapByte] |= (flags << mapOffset);
278 		MarkBufferDirty(vmBuf);
279 
280 		if (RelationNeedsWAL(rel))
281 		{
282 			if (XLogRecPtrIsInvalid(recptr))
283 			{
284 				Assert(!InRecovery);
285 				recptr = log_heap_visible(rel->rd_node, heapBuf, vmBuf,
286 										  cutoff_xid, flags);
287 
288 				/*
289 				 * If data checksums are enabled (or wal_log_hints=on), we
290 				 * need to protect the heap page from being torn.
291 				 */
292 				if (XLogHintBitIsNeeded())
293 				{
294 					Page		heapPage = BufferGetPage(heapBuf);
295 
296 					/* caller is expected to set PD_ALL_VISIBLE first */
297 					Assert(PageIsAllVisible(heapPage));
298 					PageSetLSN(heapPage, recptr);
299 				}
300 			}
301 			PageSetLSN(page, recptr);
302 		}
303 
304 		END_CRIT_SECTION();
305 	}
306 
307 	LockBuffer(vmBuf, BUFFER_LOCK_UNLOCK);
308 }
309 
310 /*
311  *	visibilitymap_get_status - get status of bits
312  *
313  * Are all tuples on heapBlk visible to all or are marked frozen, according
314  * to the visibility map?
315  *
316  * On entry, *buf should be InvalidBuffer or a valid buffer returned by an
317  * earlier call to visibilitymap_pin or visibilitymap_get_status on the same
318  * relation. On return, *buf is a valid buffer with the map page containing
319  * the bit for heapBlk, or InvalidBuffer. The caller is responsible for
320  * releasing *buf after it's done testing and setting bits.
321  *
322  * NOTE: This function is typically called without a lock on the heap page,
323  * so somebody else could change the bit just after we look at it.  In fact,
324  * since we don't lock the visibility map page either, it's even possible that
325  * someone else could have changed the bit just before we look at it, but yet
326  * we might see the old value.  It is the caller's responsibility to deal with
327  * all concurrency issues!
328  */
329 uint8
visibilitymap_get_status(Relation rel,BlockNumber heapBlk,Buffer * buf)330 visibilitymap_get_status(Relation rel, BlockNumber heapBlk, Buffer *buf)
331 {
332 	BlockNumber mapBlock = HEAPBLK_TO_MAPBLOCK(heapBlk);
333 	uint32		mapByte = HEAPBLK_TO_MAPBYTE(heapBlk);
334 	uint8		mapOffset = HEAPBLK_TO_OFFSET(heapBlk);
335 	char	   *map;
336 	uint8		result;
337 
338 #ifdef TRACE_VISIBILITYMAP
339 	elog(DEBUG1, "vm_get_status %s %d", RelationGetRelationName(rel), heapBlk);
340 #endif
341 
342 	/* Reuse the old pinned buffer if possible */
343 	if (BufferIsValid(*buf))
344 	{
345 		if (BufferGetBlockNumber(*buf) != mapBlock)
346 		{
347 			ReleaseBuffer(*buf);
348 			*buf = InvalidBuffer;
349 		}
350 	}
351 
352 	if (!BufferIsValid(*buf))
353 	{
354 		*buf = vm_readbuf(rel, mapBlock, false);
355 		if (!BufferIsValid(*buf))
356 			return false;
357 	}
358 
359 	map = PageGetContents(BufferGetPage(*buf));
360 
361 	/*
362 	 * A single byte read is atomic.  There could be memory-ordering effects
363 	 * here, but for performance reasons we make it the caller's job to worry
364 	 * about that.
365 	 */
366 	result = ((map[mapByte] >> mapOffset) & VISIBILITYMAP_VALID_BITS);
367 	return result;
368 }
369 
370 /*
371  *	visibilitymap_count  - count number of bits set in visibility map
372  *
373  * Note: we ignore the possibility of race conditions when the table is being
374  * extended concurrently with the call.  New pages added to the table aren't
375  * going to be marked all-visible or all-frozen, so they won't affect the result.
376  */
377 void
visibilitymap_count(Relation rel,BlockNumber * all_visible,BlockNumber * all_frozen)378 visibilitymap_count(Relation rel, BlockNumber *all_visible, BlockNumber *all_frozen)
379 {
380 	BlockNumber mapBlock;
381 	BlockNumber nvisible = 0;
382 	BlockNumber nfrozen = 0;
383 
384 	/* all_visible must be specified */
385 	Assert(all_visible);
386 
387 	for (mapBlock = 0;; mapBlock++)
388 	{
389 		Buffer		mapBuffer;
390 		uint64	   *map;
391 		int			i;
392 
393 		/*
394 		 * Read till we fall off the end of the map.  We assume that any extra
395 		 * bytes in the last page are zeroed, so we don't bother excluding
396 		 * them from the count.
397 		 */
398 		mapBuffer = vm_readbuf(rel, mapBlock, false);
399 		if (!BufferIsValid(mapBuffer))
400 			break;
401 
402 		/*
403 		 * We choose not to lock the page, since the result is going to be
404 		 * immediately stale anyway if anyone is concurrently setting or
405 		 * clearing bits, and we only really need an approximate value.
406 		 */
407 		map = (uint64 *) PageGetContents(BufferGetPage(mapBuffer));
408 
409 		StaticAssertStmt(MAPSIZE % sizeof(uint64) == 0,
410 						 "unsupported MAPSIZE");
411 		if (all_frozen == NULL)
412 		{
413 			for (i = 0; i < MAPSIZE / sizeof(uint64); i++)
414 				nvisible += pg_popcount64(map[i] & VISIBLE_MASK64);
415 		}
416 		else
417 		{
418 			for (i = 0; i < MAPSIZE / sizeof(uint64); i++)
419 			{
420 				nvisible += pg_popcount64(map[i] & VISIBLE_MASK64);
421 				nfrozen += pg_popcount64(map[i] & FROZEN_MASK64);
422 			}
423 		}
424 
425 		ReleaseBuffer(mapBuffer);
426 	}
427 
428 	*all_visible = nvisible;
429 	if (all_frozen)
430 		*all_frozen = nfrozen;
431 }
432 
433 /*
434  *	visibilitymap_prepare_truncate -
435  *			prepare for truncation of the visibility map
436  *
437  * nheapblocks is the new size of the heap.
438  *
439  * Return the number of blocks of new visibility map.
440  * If it's InvalidBlockNumber, there is nothing to truncate;
441  * otherwise the caller is responsible for calling smgrtruncate()
442  * to truncate the visibility map pages.
443  */
444 BlockNumber
visibilitymap_prepare_truncate(Relation rel,BlockNumber nheapblocks)445 visibilitymap_prepare_truncate(Relation rel, BlockNumber nheapblocks)
446 {
447 	BlockNumber newnblocks;
448 
449 	/* last remaining block, byte, and bit */
450 	BlockNumber truncBlock = HEAPBLK_TO_MAPBLOCK(nheapblocks);
451 	uint32		truncByte = HEAPBLK_TO_MAPBYTE(nheapblocks);
452 	uint8		truncOffset = HEAPBLK_TO_OFFSET(nheapblocks);
453 
454 #ifdef TRACE_VISIBILITYMAP
455 	elog(DEBUG1, "vm_truncate %s %d", RelationGetRelationName(rel), nheapblocks);
456 #endif
457 
458 	RelationOpenSmgr(rel);
459 
460 	/*
461 	 * If no visibility map has been created yet for this relation, there's
462 	 * nothing to truncate.
463 	 */
464 	if (!smgrexists(rel->rd_smgr, VISIBILITYMAP_FORKNUM))
465 		return InvalidBlockNumber;
466 
467 	/*
468 	 * Unless the new size is exactly at a visibility map page boundary, the
469 	 * tail bits in the last remaining map page, representing truncated heap
470 	 * blocks, need to be cleared. This is not only tidy, but also necessary
471 	 * because we don't get a chance to clear the bits if the heap is extended
472 	 * again.
473 	 */
474 	if (truncByte != 0 || truncOffset != 0)
475 	{
476 		Buffer		mapBuffer;
477 		Page		page;
478 		char	   *map;
479 
480 		newnblocks = truncBlock + 1;
481 
482 		mapBuffer = vm_readbuf(rel, truncBlock, false);
483 		if (!BufferIsValid(mapBuffer))
484 		{
485 			/* nothing to do, the file was already smaller */
486 			return InvalidBlockNumber;
487 		}
488 
489 		page = BufferGetPage(mapBuffer);
490 		map = PageGetContents(page);
491 
492 		LockBuffer(mapBuffer, BUFFER_LOCK_EXCLUSIVE);
493 
494 		/* NO EREPORT(ERROR) from here till changes are logged */
495 		START_CRIT_SECTION();
496 
497 		/* Clear out the unwanted bytes. */
498 		MemSet(&map[truncByte + 1], 0, MAPSIZE - (truncByte + 1));
499 
500 		/*----
501 		 * Mask out the unwanted bits of the last remaining byte.
502 		 *
503 		 * ((1 << 0) - 1) = 00000000
504 		 * ((1 << 1) - 1) = 00000001
505 		 * ...
506 		 * ((1 << 6) - 1) = 00111111
507 		 * ((1 << 7) - 1) = 01111111
508 		 *----
509 		 */
510 		map[truncByte] &= (1 << truncOffset) - 1;
511 
512 		/*
513 		 * Truncation of a relation is WAL-logged at a higher-level, and we
514 		 * will be called at WAL replay. But if checksums are enabled, we need
515 		 * to still write a WAL record to protect against a torn page, if the
516 		 * page is flushed to disk before the truncation WAL record. We cannot
517 		 * use MarkBufferDirtyHint here, because that will not dirty the page
518 		 * during recovery.
519 		 */
520 		MarkBufferDirty(mapBuffer);
521 		if (!InRecovery && RelationNeedsWAL(rel) && XLogHintBitIsNeeded())
522 			log_newpage_buffer(mapBuffer, false);
523 
524 		END_CRIT_SECTION();
525 
526 		UnlockReleaseBuffer(mapBuffer);
527 	}
528 	else
529 		newnblocks = truncBlock;
530 
531 	if (smgrnblocks(rel->rd_smgr, VISIBILITYMAP_FORKNUM) <= newnblocks)
532 	{
533 		/* nothing to do, the file was already smaller than requested size */
534 		return InvalidBlockNumber;
535 	}
536 
537 	return newnblocks;
538 }
539 
540 /*
541  * Read a visibility map page.
542  *
543  * If the page doesn't exist, InvalidBuffer is returned, or if 'extend' is
544  * true, the visibility map file is extended.
545  */
546 static Buffer
vm_readbuf(Relation rel,BlockNumber blkno,bool extend)547 vm_readbuf(Relation rel, BlockNumber blkno, bool extend)
548 {
549 	Buffer		buf;
550 
551 	/*
552 	 * We might not have opened the relation at the smgr level yet, or we
553 	 * might have been forced to close it by a sinval message.  The code below
554 	 * won't necessarily notice relation extension immediately when extend =
555 	 * false, so we rely on sinval messages to ensure that our ideas about the
556 	 * size of the map aren't too far out of date.
557 	 */
558 	RelationOpenSmgr(rel);
559 
560 	/*
561 	 * If we haven't cached the size of the visibility map fork yet, check it
562 	 * first.
563 	 */
564 	if (rel->rd_smgr->smgr_cached_nblocks[VISIBILITYMAP_FORKNUM] == InvalidBlockNumber)
565 	{
566 		if (smgrexists(rel->rd_smgr, VISIBILITYMAP_FORKNUM))
567 			smgrnblocks(rel->rd_smgr, VISIBILITYMAP_FORKNUM);
568 		else
569 			rel->rd_smgr->smgr_cached_nblocks[VISIBILITYMAP_FORKNUM] = 0;
570 	}
571 
572 	/* Handle requests beyond EOF */
573 	if (blkno >= rel->rd_smgr->smgr_cached_nblocks[VISIBILITYMAP_FORKNUM])
574 	{
575 		if (extend)
576 			vm_extend(rel, blkno + 1);
577 		else
578 			return InvalidBuffer;
579 	}
580 
581 	/*
582 	 * Use ZERO_ON_ERROR mode, and initialize the page if necessary. It's
583 	 * always safe to clear bits, so it's better to clear corrupt pages than
584 	 * error out.
585 	 *
586 	 * The initialize-the-page part is trickier than it looks, because of the
587 	 * possibility of multiple backends doing this concurrently, and our
588 	 * desire to not uselessly take the buffer lock in the normal path where
589 	 * the page is OK.  We must take the lock to initialize the page, so
590 	 * recheck page newness after we have the lock, in case someone else
591 	 * already did it.  Also, because we initially check PageIsNew with no
592 	 * lock, it's possible to fall through and return the buffer while someone
593 	 * else is still initializing the page (i.e., we might see pd_upper as set
594 	 * but other page header fields are still zeroes).  This is harmless for
595 	 * callers that will take a buffer lock themselves, but some callers
596 	 * inspect the page without any lock at all.  The latter is OK only so
597 	 * long as it doesn't depend on the page header having correct contents.
598 	 * Current usage is safe because PageGetContents() does not require that.
599 	 */
600 	buf = ReadBufferExtended(rel, VISIBILITYMAP_FORKNUM, blkno,
601 							 RBM_ZERO_ON_ERROR, NULL);
602 	if (PageIsNew(BufferGetPage(buf)))
603 	{
604 		LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
605 		if (PageIsNew(BufferGetPage(buf)))
606 			PageInit(BufferGetPage(buf), BLCKSZ, 0);
607 		LockBuffer(buf, BUFFER_LOCK_UNLOCK);
608 	}
609 	return buf;
610 }
611 
612 /*
613  * Ensure that the visibility map fork is at least vm_nblocks long, extending
614  * it if necessary with zeroed pages.
615  */
616 static void
vm_extend(Relation rel,BlockNumber vm_nblocks)617 vm_extend(Relation rel, BlockNumber vm_nblocks)
618 {
619 	BlockNumber vm_nblocks_now;
620 	PGAlignedBlock pg;
621 
622 	PageInit((Page) pg.data, BLCKSZ, 0);
623 
624 	/*
625 	 * We use the relation extension lock to lock out other backends trying to
626 	 * extend the visibility map at the same time. It also locks out extension
627 	 * of the main fork, unnecessarily, but extending the visibility map
628 	 * happens seldom enough that it doesn't seem worthwhile to have a
629 	 * separate lock tag type for it.
630 	 *
631 	 * Note that another backend might have extended or created the relation
632 	 * by the time we get the lock.
633 	 */
634 	LockRelationForExtension(rel, ExclusiveLock);
635 
636 	/* Might have to re-open if a cache flush happened */
637 	RelationOpenSmgr(rel);
638 
639 	/*
640 	 * Create the file first if it doesn't exist.  If smgr_vm_nblocks is
641 	 * positive then it must exist, no need for an smgrexists call.
642 	 */
643 	if ((rel->rd_smgr->smgr_cached_nblocks[VISIBILITYMAP_FORKNUM] == 0 ||
644 		 rel->rd_smgr->smgr_cached_nblocks[VISIBILITYMAP_FORKNUM] == InvalidBlockNumber) &&
645 		!smgrexists(rel->rd_smgr, VISIBILITYMAP_FORKNUM))
646 		smgrcreate(rel->rd_smgr, VISIBILITYMAP_FORKNUM, false);
647 
648 	/* Invalidate cache so that smgrnblocks() asks the kernel. */
649 	rel->rd_smgr->smgr_cached_nblocks[VISIBILITYMAP_FORKNUM] = InvalidBlockNumber;
650 	vm_nblocks_now = smgrnblocks(rel->rd_smgr, VISIBILITYMAP_FORKNUM);
651 
652 	/* Now extend the file */
653 	while (vm_nblocks_now < vm_nblocks)
654 	{
655 		PageSetChecksumInplace((Page) pg.data, vm_nblocks_now);
656 
657 		smgrextend(rel->rd_smgr, VISIBILITYMAP_FORKNUM, vm_nblocks_now,
658 				   pg.data, false);
659 		vm_nblocks_now++;
660 	}
661 
662 	/*
663 	 * Send a shared-inval message to force other backends to close any smgr
664 	 * references they may have for this rel, which we are about to change.
665 	 * This is a useful optimization because it means that backends don't have
666 	 * to keep checking for creation or extension of the file, which happens
667 	 * infrequently.
668 	 */
669 	CacheInvalidateSmgr(rel->rd_smgr->smgr_rnode);
670 
671 	UnlockRelationForExtension(rel, ExclusiveLock);
672 }
673