| blob | de28374b40507ac494c85a154cc0c39547d38409 |
1 /*-------------------------------------------------------------------------
2 *
3 * bufmgr.c
4 * buffer manager interface routines
5 *
6 * Portions Copyright (c) 1996-2009, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
8 *
9 *
10 * IDENTIFICATION
11 * $PostgreSQL$
12 *
13 *-------------------------------------------------------------------------
14 */
15 /*
16 * Principal entry points:
17 *
18 * ReadBuffer() -- find or create a buffer holding the requested page,
19 * and pin it so that no one can destroy it while this process
20 * is using it.
21 *
22 * ReleaseBuffer() -- unpin a buffer
23 *
24 * MarkBufferDirty() -- mark a pinned buffer's contents as "dirty".
25 * The disk write is delayed until buffer replacement or checkpoint.
26 *
27 * See also these files:
28 * freelist.c -- chooses victim for buffer replacement
29 * buf_table.c -- manages the buffer lookup table
30 */
33 #include <sys/file.h>
34 #include <unistd.h>
50 /* Note: these two macros only work on shared buffers, not local ones! */
51 #define BufHdrGetBlock(bufHdr) ((Block) (BufferBlocks + ((Size) (bufHdr)->buf_id) * BLCKSZ))
52 #define BufferGetLSN(bufHdr) (PageGetLSN(BufHdrGetBlock(bufHdr)))
54 /* Note: this macro only works on local buffers, not shared ones! */
55 #define LocalBufHdrGetBlock(bufHdr) \
56 LocalBufferBlockPointers[-((bufHdr)->buf_id + 2)]
58 /* Bits in SyncOneBuffer's return value */
59 #define BUF_WRITTEN 0x01
60 #define BUF_REUSABLE 0x02
63 /* GUC variables */
68 /*
69 * How many buffers PrefetchBuffer callers should try to stay ahead of their
70 * ReadBuffer calls by. This is maintained by the assign hook for
71 * effective_io_concurrency. Zero means "never prefetch".
72 */
75 /* local state for StartBufferIO and related functions */
79 /* local state for LockBufferForCleanup */
98 BlockNumber blockNum,
99 BufferAccessStrategy strategy,
105 /*
106 * PrefetchBuffer -- initiate asynchronous read of a block of a relation
107 *
108 * This is named by analogy to ReadBuffer but doesn't actually allocate a
109 * buffer. Instead it tries to ensure that a future ReadBuffer for the given
110 * block will not be delayed by the I/O. Prefetching is optional.
111 * No-op if prefetching isn't compiled in.
112 */
113 void
115 {
116 #ifdef USE_PREFETCH
120 /* Open it at the smgr level if not already done */
124 {
125 /* see comments in ReadBufferExtended */
131 /* pass it off to localbuf.c */
133 }
134 else
135 {
141 /* create a tag so we can lookup the buffer */
144 /* determine its hash code and partition lock ID */
148 /* see if the block is in the buffer pool already */
153 /* If not in buffers, initiate prefetch */
157 /*
158 * If the block *is* in buffers, we do nothing. This is not really
159 * ideal: the block might be just about to be evicted, which would be
160 * stupid since we know we are going to need it soon. But the only
161 * easy answer is to bump the usage_count, which does not seem like a
162 * great solution: when the caller does ultimately touch the block,
163 * usage_count would get bumped again, resulting in too much
164 * favoritism for blocks that are involved in a prefetch sequence. A
165 * real fix would involve some additional per-buffer state, and it's
166 * not clear that there's enough of a problem to justify that.
167 */
168 }
170 }
173 /*
174 * ReadBuffer -- a shorthand for ReadBufferExtended, for reading from main
175 * fork with RBM_NORMAL mode and default strategy.
176 */
177 Buffer
179 {
181 }
183 /*
184 * ReadBufferExtended -- returns a buffer containing the requested
185 * block of the requested relation. If the blknum
186 * requested is P_NEW, extend the relation file and
187 * allocate a new block. (Caller is responsible for
188 * ensuring that only one backend tries to extend a
189 * relation at the same time!)
190 *
191 * Returns: the buffer number for the buffer containing
192 * the block read. The returned buffer has been pinned.
193 * Does not return on error --- elog's instead.
194 *
195 * Assume when this function is called, that reln has been opened already.
196 *
197 * In RBM_NORMAL mode, the page is read from disk, and the page header is
198 * validated. An error is thrown if the page header is not valid.
199 *
200 * RBM_ZERO_ON_ERROR is like the normal mode, but if the page header is not
201 * valid, the page is zeroed instead of throwing an error. This is intended
202 * for non-critical data, where the caller is prepared to repair errors.
203 *
204 * In RBM_ZERO mode, if the page isn't in buffer cache already, it's filled
205 * with zeros instead of reading it from disk. Useful when the caller is
206 * going to fill the page from scratch, since this saves I/O and avoids
207 * unnecessary failure if the page-on-disk has corrupt page headers.
208 * Caution: do not use this mode to read a page that is beyond the relation's
209 * current physical EOF; that is likely to cause problems in md.c when
210 * the page is modified and written out. P_NEW is OK, though.
211 *
212 * If strategy is not NULL, a nondefault buffer access strategy is used.
213 * See buffer/README for details.
214 */
215 Buffer
218 {
220 Buffer buf;
222 /* Open it at the smgr level if not already done */
225 /*
226 * Reject attempts to read non-local temporary relations; we would be
227 * likely to get wrong data since we have no visibility into the owning
228 * session's local buffers.
229 */
235 /*
236 * Read the buffer, and update pgstat counters to reflect a cache hit or
237 * miss.
238 */
245 }
248 /*
249 * ReadBufferWithoutRelcache -- like ReadBufferExtended, but doesn't require
250 * a relcache entry for the relation.
251 *
252 * NB: caller is assumed to know what it's doing if isTemp is true.
253 */
254 Buffer
258 {
265 }
268 /*
269 * ReadBuffer_common -- common logic for all ReadBuffer variants
270 *
271 * *hit is set to true if the request was satisfied from shared buffer cache.
272 */
273 static Buffer
277 {
279 Block bufBlock;
285 /* Make sure we will have room to remember the buffer pin */
294 isLocalBuf,
295 isExtend);
297 /* Substitute proper block number if caller asked for P_NEW */
302 {
303 ReadLocalBufferCount++;
306 LocalBufferHitCount++;
307 }
308 else
309 {
310 ReadBufferCount++;
312 /*
313 * lookup the buffer. IO_IN_PROGRESS is set if the requested block is
314 * not currently in memory.
315 */
318 BufferHitCount++;
319 }
321 /* At this point we do NOT hold any locks. */
323 /* if it was already in the buffer pool, we're done */
325 {
327 {
328 /* Just need to update stats before we exit */
338 isLocalBuf,
339 isExtend,
340 found);
343 }
345 /*
346 * We get here only in the corner case where we are trying to extend
347 * the relation but we found a pre-existing buffer marked BM_VALID.
348 * This can happen because mdread doesn't complain about reads beyond
349 * EOF (when zero_damaged_pages is ON) and so a previous attempt to
350 * read a block beyond EOF could have left a "valid" zero-filled
351 * buffer. Unfortunately, we have also seen this case occurring
352 * because of buggy Linux kernels that sometimes return an
353 * lseek(SEEK_END) result that doesn't account for a recent write. In
354 * that situation, the pre-existing buffer would contain valid data
355 * that we don't want to overwrite. Since the legitimate case should
356 * always have left a zero-filled buffer, complain if not PageIsNew.
357 */
365 /*
366 * We *must* do smgrextend before succeeding, else the page will not
367 * be reserved by the kernel, and the next P_NEW call will decide to
368 * return the same page. Clear the BM_VALID bit, do the StartBufferIO
369 * call that BufferAlloc didn't, and proceed.
370 */
372 {
373 /* Only need to adjust flags */
376 }
377 else
378 {
379 /*
380 * Loop to handle the very small possibility that someone re-sets
381 * BM_VALID between our clearing it and StartBufferIO inspecting
382 * it.
383 */
384 do
385 {
391 }
392 }
394 /*
395 * if we have gotten to this point, we have allocated a buffer for the
396 * page but its contents are not yet valid. IO_IN_PROGRESS is set for it,
397 * if it's a shared buffer.
398 *
399 * Note: if smgrextend fails, we will end up with a buffer that is
400 * allocated but not marked BM_VALID. P_NEW will still select the same
401 * block number (because the relation didn't get any longer on disk) and
402 * so future attempts to extend the relation will find the same buffer (if
403 * it's not been recycled) but come right back here to try smgrextend
404 * again.
405 */
411 {
412 /* new buffers are zero-filled */
415 }
416 else
417 {
418 /*
419 * Read in the page, unless the caller intends to overwrite it and
420 * just wants us to allocate a buffer.
421 */
424 else
425 {
428 /* check for garbage data */
430 {
432 {
436 blockNum,
439 }
440 else
444 blockNum,
446 }
447 }
448 }
451 {
452 /* Only need to adjust flags */
454 }
455 else
456 {
457 /* Set BM_VALID, terminate IO, and wake up any waiters */
459 }
468 isLocalBuf,
469 isExtend,
470 found);
473 }
475 /*
476 * BufferAlloc -- subroutine for ReadBuffer. Handles lookup of a shared
477 * buffer. If no buffer exists already, selects a replacement
478 * victim and evicts the old page, but does NOT read in new page.
479 *
480 * "strategy" can be a buffer replacement strategy object, or NULL for
481 * the default strategy. The selected buffer's usage_count is advanced when
482 * using the default strategy, but otherwise possibly not (see PinBuffer).
483 *
484 * The returned buffer is pinned and is already marked as holding the
485 * desired page. If it already did have the desired page, *foundPtr is
486 * set TRUE. Otherwise, *foundPtr is set FALSE and the buffer is marked
487 * as IO_IN_PROGRESS; ReadBuffer will now need to do I/O to fill it.
488 *
489 * *foundPtr is actually redundant with the buffer's BM_VALID flag, but
490 * we keep it for simplicity in ReadBuffer.
491 *
492 * No locks are held either at entry or exit.
493 */
496 BlockNumber blockNum,
497 BufferAccessStrategy strategy,
499 {
506 BufFlags oldFlags;
511 /* create a tag so we can lookup the buffer */
514 /* determine its hash code and partition lock ID */
518 /* see if the block is in the buffer pool already */
522 {
523 /*
524 * Found it. Now, pin the buffer so no one can steal it from the
525 * buffer pool, and check to see if the correct data has been loaded
526 * into the buffer.
527 */
532 /* Can release the mapping lock as soon as we've pinned it */
538 {
539 /*
540 * We can only get here if (a) someone else is still reading in
541 * the page, or (b) a previous read attempt failed. We have to
542 * wait for any active read attempt to finish, and then set up our
543 * own read attempt if the page is still not BM_VALID.
544 * StartBufferIO does it all.
545 */
547 {
548 /*
549 * If we get here, previous attempts to read the buffer must
550 * have failed ... but we shall bravely try again.
551 */
553 }
554 }
557 }
559 /*
560 * Didn't find it in the buffer pool. We'll have to initialize a new
561 * buffer. Remember to unlock the mapping lock while doing the work.
562 */
565 /* Loop here in case we have to try another victim buffer */
567 {
570 /*
571 * Select a victim buffer. The buffer is returned with its header
572 * spinlock still held! Also (in most cases) the BufFreelistLock is
573 * still held, since it would be bad to hold the spinlock while
574 * possibly waking up other processes.
575 */
580 /* Must copy buffer flags while we still hold the spinlock */
583 /* Pin the buffer and then release the buffer spinlock */
586 /* Now it's safe to release the freelist lock */
590 /*
591 * If the buffer was dirty, try to write it out. There is a race
592 * condition here, in that someone might dirty it after we released it
593 * above, or even while we are writing it out (since our share-lock
594 * won't prevent hint-bit updates). We will recheck the dirty bit
595 * after re-locking the buffer header.
596 */
598 {
599 /*
600 * We need a share-lock on the buffer contents to write it out
601 * (else we might write invalid data, eg because someone else is
602 * compacting the page contents while we write). We must use a
603 * conditional lock acquisition here to avoid deadlock. Even
604 * though the buffer was not pinned (and therefore surely not
605 * locked) when StrategyGetBuffer returned it, someone else could
606 * have pinned and exclusive-locked it by the time we get here. If
607 * we try to get the lock unconditionally, we'd block waiting for
608 * them; if they later block waiting for us, deadlock ensues.
609 * (This has been observed to happen when two backends are both
610 * trying to split btree index pages, and the second one just
611 * happens to be trying to split the page the first one got from
612 * StrategyGetBuffer.)
613 */
615 {
616 /*
617 * If using a nondefault strategy, and writing the buffer
618 * would require a WAL flush, let the strategy decide whether
619 * to go ahead and write/reuse the buffer or to choose another
620 * victim. We need lock to inspect the page LSN, so this
621 * can't be done inside StrategyGetBuffer.
622 */
626 {
627 /* Drop lock/pin and loop around for another buffer */
631 }
633 /* OK, do the I/O */
646 }
647 else
648 {
649 /*
650 * Someone else has locked the buffer, so give it up and loop
651 * back to get another one.
652 */
655 }
656 }
658 /*
659 * To change the association of a valid buffer, we'll need to have
660 * exclusive lock on both the old and new mapping partitions.
661 */
663 {
664 /*
665 * Need to compute the old tag's hashcode and partition lock ID.
666 * XXX is it worth storing the hashcode in BufferDesc so we need
667 * not recompute it here? Probably not.
668 */
673 /*
674 * Must lock the lower-numbered partition first to avoid
675 * deadlocks.
676 */
678 {
681 }
683 {
686 }
687 else
688 {
689 /* only one partition, only one lock */
691 }
692 }
693 else
694 {
695 /* if it wasn't valid, we need only the new partition */
697 /* these just keep the compiler quiet about uninit variables */
700 }
702 /*
703 * Try to make a hashtable entry for the buffer under its new tag.
704 * This could fail because while we were writing someone else
705 * allocated another buffer for the same block we want to read in.
706 * Note that we have not yet removed the hashtable entry for the old
707 * tag.
708 */
712 {
713 /*
714 * Got a collision. Someone has already done what we were about to
715 * do. We'll just handle this as if it were found in the buffer
716 * pool in the first place. First, give up the buffer we were
717 * planning to use.
718 */
721 /* Can give up that buffer's mapping partition lock now */
726 /* remaining code should match code at top of routine */
732 /* Can release the mapping lock as soon as we've pinned it */
738 {
739 /*
740 * We can only get here if (a) someone else is still reading
741 * in the page, or (b) a previous read attempt failed. We
742 * have to wait for any active read attempt to finish, and
743 * then set up our own read attempt if the page is still not
744 * BM_VALID. StartBufferIO does it all.
745 */
747 {
748 /*
749 * If we get here, previous attempts to read the buffer
750 * must have failed ... but we shall bravely try again.
751 */
753 }
754 }
757 }
759 /*
760 * Need to lock the buffer header too in order to change its tag.
761 */
764 /*
765 * Somebody could have pinned or re-dirtied the buffer while we were
766 * doing the I/O and making the new hashtable entry. If so, we can't
767 * recycle this buffer; we must undo everything we've done and start
768 * over with a new victim buffer.
769 */
781 }
783 /*
784 * Okay, it's finally safe to rename the buffer.
785 *
786 * Clearing BM_VALID here is necessary, clearing the dirtybits is just
787 * paranoia. We also reset the usage_count since any recency of use of
788 * the old content is no longer relevant. (The usage_count starts out at
789 * 1 so that the buffer can survive one clock-sweep pass.)
790 */
799 {
803 }
807 /*
808 * Buffer contents are currently invalid. Try to get the io_in_progress
809 * lock. If StartBufferIO returns false, then someone else managed to
810 * read it before we did, so there's nothing left for BufferAlloc() to do.
811 */
814 else
818 }
820 /*
821 * InvalidateBuffer -- mark a shared buffer invalid and return it to the
822 * freelist.
823 *
824 * The buffer header spinlock must be held at entry. We drop it before
825 * returning. (This is sane because the caller must have locked the
826 * buffer in order to be sure it should be dropped.)
827 *
828 * This is used only in contexts such as dropping a relation. We assume
829 * that no other backend could possibly be interested in using the page,
830 * so the only reason the buffer might be pinned is if someone else is
831 * trying to write it out. We have to let them finish before we can
832 * reclaim the buffer.
833 *
834 * The buffer could get reclaimed by someone else while we are waiting
835 * to acquire the necessary locks; if so, don't mess it up.
836 */
837 static void
839 {
840 BufferTag oldTag;
843 BufFlags oldFlags;
845 /* Save the original buffer tag before dropping the spinlock */
850 /*
851 * Need to compute the old tag's hashcode and partition lock ID. XXX is it
852 * worth storing the hashcode in BufferDesc so we need not recompute it
853 * here? Probably not.
854 */
858 retry:
860 /*
861 * Acquire exclusive mapping lock in preparation for changing the buffer's
862 * association.
863 */
866 /* Re-lock the buffer header */
869 /* If it's changed while we were waiting for lock, do nothing */
871 {
875 }
877 /*
878 * We assume the only reason for it to be pinned is that someone else is
879 * flushing the page out. Wait for them to finish. (This could be an
880 * infinite loop if the refcount is messed up... it would be nice to time
881 * out after awhile, but there seems no way to be sure how many loops may
882 * be needed. Note that if the other guy has pinned the buffer but not
883 * yet done StartBufferIO, WaitIO will fall through and we'll effectively
884 * be busy-looping here.)
885 */
887 {
890 /* safety check: should definitely not be our *own* pin */
895 }
897 /*
898 * Clear out the buffer's tag and flags. We must do this to ensure that
899 * linear scans of the buffer array don't think the buffer is valid.
900 */
908 /*
909 * Remove the buffer from the lookup hashtable, if it was in there.
910 */
914 /*
915 * Done with mapping lock.
916 */
919 /*
920 * Insert the buffer at the head of the list of free buffers.
921 */
923 }
925 /*
926 * MarkBufferDirty
927 *
928 * Marks buffer contents as dirty (actual write happens later).
929 *
930 * Buffer must be pinned and exclusive-locked. (If caller does not hold
931 * exclusive lock, then somebody could be in process of writing the buffer,
932 * leading to risk of bad data written to disk.)
933 */
934 void
936 {
943 {
946 }
951 /* unfortunately we can't check if the lock is held exclusively */
958 /*
959 * If the buffer was not dirty already, do vacuum cost accounting.
960 */
967 }
969 /*
970 * ReleaseAndReadBuffer -- combine ReleaseBuffer() and ReadBuffer()
971 *
972 * Formerly, this saved one cycle of acquiring/releasing the BufMgrLock
973 * compared to calling the two routines separately. Now it's mainly just
974 * a convenience function. However, if the passed buffer is valid and
975 * already contains the desired block, we just return it as-is; and that
976 * does save considerable work compared to a full release and reacquire.
977 *
978 * Note: it is OK to pass buffer == InvalidBuffer, indicating that no old
979 * buffer actually needs to be released. This case is the same as ReadBuffer,
980 * but can save some tests in the caller.
981 */
982 Buffer
984 Relation relation,
985 BlockNumber blockNum)
986 {
991 {
993 {
1002 }
1003 else
1004 {
1007 /* we have pin, so it's ok to examine tag without spinlock */
1013 }
1014 }
1017 }
1019 /*
1020 * PinBuffer -- make buffer unavailable for replacement.
1021 *
1022 * For the default access strategy, the buffer's usage_count is incremented
1023 * when we first pin it; for other strategies we just make sure the usage_count
1024 * isn't zero. (The idea of the latter is that we don't want synchronized
1025 * heap scans to inflate the count, but we need it to not be zero to discourage
1026 * other backends from stealing buffers from our ring. As long as we cycle
1027 * through the ring faster than the global clock-sweep cycles, buffers in
1028 * our ring won't be chosen as victims for replacement by other backends.)
1029 *
1030 * This should be applied only to shared buffers, never local ones.
1031 *
1032 * Note that ResourceOwnerEnlargeBuffers must have been done already.
1033 *
1034 * Returns TRUE if buffer is BM_VALID, else FALSE. This provision allows
1035 * some callers to avoid an extra spinlock cycle.
1036 */
1037 static bool
1039 {
1044 {
1048 {
1051 }
1052 else
1053 {
1056 }
1059 }
1060 else
1061 {
1062 /* If we previously pinned the buffer, it must surely be valid */
1064 }
1070 }
1072 /*
1073 * PinBuffer_Locked -- as above, but caller already locked the buffer header.
1074 * The spinlock is released before return.
1075 *
1076 * Currently, no callers of this function want to modify the buffer's
1077 * usage_count at all, so there's no need for a strategy parameter.
1078 * Also we don't bother with a BM_VALID test (the caller could check that for
1079 * itself).
1080 *
1081 * Note: use of this routine is frequently mandatory, not just an optimization
1082 * to save a spin lock/unlock cycle, because we need to pin a buffer before
1083 * its state can change under us.
1084 */
1085 static void
1087 {
1097 }
1099 /*
1100 * UnpinBuffer -- make buffer available for replacement.
1101 *
1102 * This should be applied only to shared buffers, never local ones.
1103 *
1104 * Most but not all callers want CurrentResourceOwner to be adjusted.
1105 * Those that don't should pass fixOwner = FALSE.
1106 */
1107 static void
1109 {
1119 {
1120 /* I'd better not still hold any locks on the buffer */
1126 /* Decrement the shared reference count */
1130 /* Support LockBufferForCleanup() */
1133 {
1134 /* we just released the last pin other than the waiter's */
1140 }
1141 else
1143 }
1144 }
1146 /*
1147 * BufferSync -- Write out all dirty buffers in the pool.
1148 *
1149 * This is called at checkpoint time to write out all dirty shared buffers.
1150 * The checkpoint request flags should be passed in; currently the only one
1151 * examined is CHECKPOINT_IMMEDIATE, which disables delays between writes.
1152 */
1153 static void
1155 {
1161 /* Make sure we can handle the pin inside SyncOneBuffer */
1164 /*
1165 * Loop over all buffers, and mark the ones that need to be written with
1166 * BM_CHECKPOINT_NEEDED. Count them as we go (num_to_write), so that we
1167 * can estimate how much work needs to be done.
1168 *
1169 * This allows us to write only those pages that were dirty when the
1170 * checkpoint began, and not those that get dirtied while it proceeds.
1171 * Whenever a page with BM_CHECKPOINT_NEEDED is written out, either by us
1172 * later in this function, or by normal backends or the bgwriter cleaning
1173 * scan, the flag is cleared. Any buffer dirtied after this point won't
1174 * have the flag set.
1175 *
1176 * Note that if we fail to write some buffer, we may leave buffers with
1177 * BM_CHECKPOINT_NEEDED still set. This is OK since any such buffer would
1178 * certainly need to be written for the next checkpoint attempt, too.
1179 */
1182 {
1185 /*
1186 * Header spinlock is enough to examine BM_DIRTY, see comment in
1187 * SyncOneBuffer.
1188 */
1192 {
1194 num_to_write++;
1195 }
1198 }
1205 /*
1206 * Loop over all buffers again, and write the ones (still) marked with
1207 * BM_CHECKPOINT_NEEDED. In this loop, we start at the clock sweep point
1208 * since we might as well dump soon-to-be-recycled buffers first.
1209 *
1210 * Note that we don't read the buffer alloc count here --- that should be
1211 * left untouched till the next BgBufferSync() call.
1212 */
1217 {
1220 /*
1221 * We don't need to acquire the lock here, because we're only looking
1222 * at a single bit. It's possible that someone else writes the buffer
1223 * and clears the flag right after we check, but that doesn't matter
1224 * since SyncOneBuffer will then do nothing. However, there is a
1225 * further race condition: it's conceivable that between the time we
1226 * examine the bit here and the time SyncOneBuffer acquires lock,
1227 * someone else not only wrote the buffer but replaced it with another
1228 * page and dirtied it. In that improbable case, SyncOneBuffer will
1229 * write the buffer though we didn't need to. It doesn't seem worth
1230 * guarding against this, though.
1231 */
1233 {
1235 {
1238 num_written++;
1240 /*
1241 * We know there are at most num_to_write buffers with
1242 * BM_CHECKPOINT_NEEDED set; so we can stop scanning if
1243 * num_written reaches num_to_write.
1244 *
1245 * Note that num_written doesn't include buffers written by
1246 * other backends, or by the bgwriter cleaning scan. That
1247 * means that the estimate of how much progress we've made is
1248 * conservative, and also that this test will often fail to
1249 * trigger. But it seems worth making anyway.
1250 */
1254 /*
1255 * Perform normal bgwriter duties and sleep to throttle our
1256 * I/O rate.
1257 */
1260 }
1261 }
1265 }
1267 /*
1268 * Update checkpoint statistics. As noted above, this doesn't include
1269 * buffers written by other backends or bgwriter scan.
1270 */
1274 }
1276 /*
1277 * BgBufferSync -- Write out some dirty buffers in the pool.
1278 *
1279 * This is called periodically by the background writer process.
1280 */
1281 void
1283 {
1284 /* info obtained from freelist.c */
1286 uint32 strategy_passes;
1287 uint32 recent_alloc;
1289 /*
1290 * Information saved between calls so we can determine the strategy
1291 * point's advance rate and avoid scanning already-cleaned buffers.
1292 */
1299 /* Moving averages of allocation rate and clean-buffer density */
1303 /* Potentially these could be tunables, but for now, not */
1307 /* Used to compute how far we scan ahead */
1316 /* Variables for the scanning loop proper */
1321 /*
1322 * Find out where the freelist clock sweep currently is, and how many
1323 * buffer allocations have happened since our last call.
1324 */
1327 /* Report buffer alloc counts to pgstat */
1330 /*
1331 * If we're not running the LRU scan, just stop after doing the stats
1332 * stuff. We mark the saved state invalid so that we can recover sanely
1333 * if LRU scan is turned back on later.
1334 */
1336 {
1339 }
1341 /*
1342 * Compute strategy_delta = how many buffers have been scanned by the
1343 * clock sweep since last time. If first time through, assume none. Then
1344 * see if we are still ahead of the clock sweep, and if so, how many
1345 * buffers we could scan before we'd catch up with it and "lap" it. Note:
1346 * weird-looking coding of xxx_passes comparisons are to avoid bogus
1347 * behavior when the passes counts wrap around.
1348 */
1350 {
1359 {
1360 /* we're one pass ahead of the strategy point */
1362 #ifdef BGW_DEBUG
1367 #endif
1368 }
1371 {
1372 /* on same pass, but ahead or at least not behind */
1374 #ifdef BGW_DEBUG
1379 #endif
1380 }
1381 else
1382 {
1383 /*
1384 * We're behind, so skip forward to the strategy point and start
1385 * cleaning from there.
1386 */
1387 #ifdef BGW_DEBUG
1391 strategy_delta);
1392 #endif
1396 }
1397 }
1398 else
1399 {
1400 /*
1401 * Initializing at startup or after LRU scanning had been off. Always
1402 * start at the strategy point.
1403 */
1404 #ifdef BGW_DEBUG
1407 #endif
1412 }
1414 /* Update saved info for next time */
1419 /*
1420 * Compute how many buffers had to be scanned for each new allocation, ie,
1421 * 1/density of reusable buffers, and track a moving average of that.
1422 *
1423 * If the strategy point didn't move, we don't update the density estimate
1424 */
1426 {
1429 smoothing_samples;
1430 }
1432 /*
1433 * Estimate how many reusable buffers there are between the current
1434 * strategy point and where we've scanned ahead to, based on the smoothed
1435 * density estimate.
1436 */
1440 /*
1441 * Track a moving average of recent buffer allocations. Here, rather than
1442 * a true average we want a fast-attack, slow-decline behavior: we
1443 * immediately follow any increase.
1444 */
1447 else
1449 smoothing_samples;
1451 /* Scale the estimate by a GUC to allow more aggressive tuning. */
1454 /*
1455 * Even in cases where there's been little or no buffer allocation
1456 * activity, we want to make a small amount of progress through the buffer
1457 * cache so that as many reusable buffers as possible are clean after an
1458 * idle period.
1459 *
1460 * (scan_whole_pool_milliseconds / BgWriterDelay) computes how many times
1461 * the BGW will be called during the scan_whole_pool time; slice the
1462 * buffer pool into that many sections.
1463 */
1467 {
1468 #ifdef BGW_DEBUG
1471 #endif
1473 }
1475 /*
1476 * Now write out dirty reusable buffers, working forward from the
1477 * next_to_clean point, until we have lapped the strategy scan, or cleaned
1478 * enough buffers to match our estimate of the next cycle's allocation
1479 * requirements, or hit the bgwriter_lru_maxpages limit.
1480 */
1482 /* Make sure we can handle the pin inside SyncOneBuffer */
1489 /* Execute the LRU scan */
1491 {
1495 {
1497 next_passes++;
1498 }
1499 num_to_scan--;
1502 {
1503 reusable_buffers++;
1505 {
1508 }
1509 }
1511 reusable_buffers++;
1512 }
1516 #ifdef BGW_DEBUG
1517 elog(DEBUG1, "bgwriter: recent_alloc=%u smoothed=%.2f delta=%ld ahead=%d density=%.2f reusable_est=%d upcoming_est=%d scanned=%d wrote=%d reusable=%d",
1521 num_written,
1523 #endif
1525 /*
1526 * Consider the above scan as being like a new allocation scan.
1527 * Characterize its density and update the smoothed one based on it. This
1528 * effectively halves the moving average period in cases where both the
1529 * strategy and the background writer are doing some useful scanning,
1530 * which is helpful because a long memory isn't as desirable on the
1531 * density estimates.
1532 */
1536 {
1539 smoothing_samples;
1541 #ifdef BGW_DEBUG
1544 #endif
1545 }
1546 }
1548 /*
1549 * SyncOneBuffer -- process a single buffer during syncing.
1550 *
1551 * If skip_recently_used is true, we don't write currently-pinned buffers, nor
1552 * buffers marked recently used, as these are not replacement candidates.
1553 *
1554 * Returns a bitmask containing the following flag bits:
1555 * BUF_WRITTEN: we wrote the buffer.
1556 * BUF_REUSABLE: buffer is available for replacement, ie, it has
1557 * pin count 0 and usage count 0.
1558 *
1559 * (BUF_WRITTEN could be set in error if FlushBuffers finds the buffer clean
1560 * after locking it, but we don't care all that much.)
1561 *
1562 * Note: caller must have done ResourceOwnerEnlargeBuffers.
1563 */
1564 static int
1566 {
1570 /*
1571 * Check whether buffer needs writing.
1572 *
1573 * We can make this check without taking the buffer content lock so long
1574 * as we mark pages dirty in access methods *before* logging changes with
1575 * XLogInsert(): if someone marks the buffer dirty just after our check we
1576 * don't worry because our checkpoint.redo points before log record for
1577 * upcoming changes and so we are not required to write such dirty buffer.
1578 */
1584 {
1585 /* Caller told us not to write recently-used buffers */
1588 }
1591 {
1592 /* It's clean, so nothing to do */
1595 }
1597 /*
1598 * Pin it, share-lock it, write it. (FlushBuffer will do nothing if the
1599 * buffer is clean by the time we've locked it.)
1600 */
1610 }
1613 /*
1614 * Return a palloc'd string containing buffer usage statistics.
1615 */
1618 {
1619 StringInfoData str;
1627 else
1632 else
1646 }
1648 void
1650 {
1659 }
1661 /*
1662 * AtEOXact_Buffers - clean up at end of transaction.
1663 *
1664 * As of PostgreSQL 8.0, buffer pins should get released by the
1665 * ResourceOwner mechanism. This routine is just a debugging
1666 * cross-check that no pins remain.
1667 */
1668 void
1670 {
1671 #ifdef USE_ASSERT_CHECKING
1673 {
1677 {
1679 }
1680 }
1681 #endif
1684 }
1686 /*
1687 * InitBufferPoolBackend --- second-stage initialization of a new backend
1688 *
1689 * This is called after we have acquired a PGPROC and so can safely get
1690 * LWLocks. We don't currently need to do anything at this stage ...
1691 * except register a shmem-exit callback. AtProcExit_Buffers needs LWLock
1692 * access, and thereby has to be called at the corresponding phase of
1693 * backend shutdown.
1694 */
1695 void
1697 {
1699 }
1701 /*
1702 * Ensure we have released all shared-buffer locks and pins during backend exit
1703 */
1704 static void
1706 {
1713 {
1715 {
1718 /*
1719 * We don't worry about updating ResourceOwner; if we even got
1720 * here, it suggests that ResourceOwners are messed up.
1721 */
1725 }
1726 }
1728 /* localbuf.c needs a chance too */
1730 }
1732 /*
1733 * Helper routine to issue warnings when a buffer is unexpectedly pinned
1734 */
1735 void
1737 {
1739 int32 loccount;
1744 {
1747 }
1748 else
1749 {
1752 }
1754 /* theoretically we should lock the bufhdr here */
1757 "buffer refcount leak: [%03d] "
1763 }
1765 /*
1766 * CheckPointBuffers
1767 *
1768 * Flush all dirty blocks in buffer pool to disk at checkpoint time.
1769 *
1770 * Note: temporary relations do not participate in checkpoints, so they don't
1771 * need to be flushed.
1772 */
1773 void
1775 {
1784 }
1787 /*
1788 * Do whatever is needed to prepare for commit at the bufmgr and smgr levels
1789 */
1790 void
1792 {
1793 /* Nothing to do in bufmgr anymore... */
1794 }
1796 /*
1797 * BufferGetBlockNumber
1798 * Returns the block number associated with a buffer.
1799 *
1800 * Note:
1801 * Assumes that the buffer is valid and pinned, else the
1802 * value may be obsolete immediately...
1803 */
1804 BlockNumber
1806 {
1813 else
1816 /* pinned, so OK to read tag without spinlock */
1818 }
1820 /*
1821 * BufferGetTag
1822 * Returns the relfilenode, fork number and block number associated with
1823 * a buffer.
1824 */
1825 void
1828 {
1831 /* Do the same checks as BufferGetBlockNumber. */
1836 else
1839 /* pinned, so OK to read tag without spinlock */
1843 }
1845 /*
1846 * FlushBuffer
1847 * Physically write out a shared buffer.
1848 *
1849 * NOTE: this actually just passes the buffer contents to the kernel; the
1850 * real write to disk won't happen until the kernel feels like it. This
1851 * is okay from our point of view since we can redo the changes from WAL.
1852 * However, we will need to force the changes to disk via fsync before
1853 * we can checkpoint WAL.
1854 *
1855 * The caller must hold a pin on the buffer and have share-locked the
1856 * buffer contents. (Note: a share-lock does not prevent updates of
1857 * hint bits in the buffer, so the page could change while the write
1858 * is in progress, but we assume that that will not invalidate the data
1859 * written.)
1860 *
1861 * If the caller has an smgr reference for the buffer's relation, pass it
1862 * as the second parameter. If not, pass NULL.
1863 */
1864 static void
1866 {
1867 XLogRecPtr recptr;
1868 ErrorContextCallback errcontext;
1870 /*
1871 * Acquire the buffer's io_in_progress lock. If StartBufferIO returns
1872 * false, then someone else flushed the buffer before we could, so we need
1873 * not do anything.
1874 */
1878 /* Setup error traceback support for ereport() */
1884 /* Find smgr relation for buffer */
1894 /*
1895 * Force XLOG flush up to buffer's LSN. This implements the basic WAL
1896 * rule that log updates must hit disk before any of the data-file changes
1897 * they describe do.
1898 */
1902 /*
1903 * Now it's safe to write buffer to disk. Note that no one else should
1904 * have been able to write it while we were busy with log flushing because
1905 * we have the io_in_progress lock.
1906 */
1908 /* To check if block content changes while flushing. - vadim 01/17/97 */
1919 BufferFlushCount++;
1921 /*
1922 * Mark the buffer as clean (unless BM_JUST_DIRTIED has become set) and
1923 * end the io_in_progress state.
1924 */
1933 /* Pop the error context stack */
1935 }
1937 /*
1938 * RelationGetNumberOfBlocks
1939 * Determines the current number of pages in the relation.
1940 */
1941 BlockNumber
1943 {
1944 /* Open it at the smgr level if not already done */
1948 }
1950 /* ---------------------------------------------------------------------
1951 * DropRelFileNodeBuffers
1952 *
1953 * This function removes from the buffer pool all the pages of the
1954 * specified relation that have block numbers >= firstDelBlock.
1955 * (In particular, with firstDelBlock = 0, all pages are removed.)
1956 * Dirty pages are simply dropped, without bothering to write them
1957 * out first. Therefore, this is NOT rollback-able, and so should be
1958 * used only with extreme caution!
1959 *
1960 * Currently, this is called only from smgr.c when the underlying file
1961 * is about to be deleted or truncated (firstDelBlock is needed for
1962 * the truncation case). The data in the affected pages would therefore
1963 * be deleted momentarily anyway, and there is no point in writing it.
1964 * It is the responsibility of higher-level code to ensure that the
1965 * deletion or truncation does not lose any data that could be needed
1966 * later. It is also the responsibility of higher-level code to ensure
1967 * that no other process could be trying to load more pages of the
1968 * relation into buffers.
1969 *
1970 * XXX currently it sequentially searches the buffer pool, should be
1971 * changed to more clever ways of searching. However, this routine
1972 * is used only in code paths that aren't very performance-critical,
1973 * and we shouldn't slow down the hot paths to make it faster ...
1974 * --------------------------------------------------------------------
1975 */
1976 void
1978 BlockNumber firstDelBlock)
1979 {
1983 {
1986 }
1989 {
1997 else
1999 }
2000 }
2002 /* ---------------------------------------------------------------------
2003 * DropDatabaseBuffers
2004 *
2005 * This function removes all the buffers in the buffer cache for a
2006 * particular database. Dirty pages are simply dropped, without
2007 * bothering to write them out first. This is used when we destroy a
2008 * database, to avoid trying to flush data to disk when the directory
2009 * tree no longer exists. Implementation is pretty similar to
2010 * DropRelFileNodeBuffers() which is for destroying just one relation.
2011 * --------------------------------------------------------------------
2012 */
2013 void
2015 {
2019 /*
2020 * We needn't consider local buffers, since by assumption the target
2021 * database isn't our own.
2022 */
2025 {
2030 else
2032 }
2033 }
2035 /* -----------------------------------------------------------------
2036 * PrintBufferDescs
2037 *
2038 * this function prints all the buffer descriptors, for debugging
2039 * use only.
2040 * -----------------------------------------------------------------
2041 */
2042 #ifdef NOT_USED
2043 void
2045 {
2050 {
2051 /* theoretically we should lock the bufhdr here */
2053 "[%02d] (freeNext=%d, rel=%s, "
2059 }
2060 }
2061 #endif
2063 #ifdef NOT_USED
2064 void
2066 {
2071 {
2073 {
2074 /* theoretically we should lock the bufhdr here */
2076 "[%02d] (freeNext=%d, rel=%s, "
2082 }
2083 }
2084 }
2085 #endif
2087 /* ---------------------------------------------------------------------
2088 * FlushRelationBuffers
2089 *
2090 * This function writes all dirty pages of a relation out to disk
2091 * (or more accurately, out to kernel disk buffers), ensuring that the
2092 * kernel has an up-to-date view of the relation.
2093 *
2094 * Generally, the caller should be holding AccessExclusiveLock on the
2095 * target relation to ensure that no other backend is busy dirtying
2096 * more blocks of the relation; the effects can't be expected to last
2097 * after the lock is released.
2098 *
2099 * XXX currently it sequentially searches the buffer pool, should be
2100 * changed to more clever ways of searching. This routine is not
2101 * used in any performance-critical code paths, so it's not worth
2102 * adding additional overhead to normal paths to make it go faster;
2103 * but see also DropRelFileNodeBuffers.
2104 * --------------------------------------------------------------------
2105 */
2106 void
2108 {
2112 /* Open rel at the smgr level if not already done */
2116 {
2118 {
2122 {
2123 ErrorContextCallback errcontext;
2125 /* Setup error traceback support for ereport() */
2139 /* Pop the error context stack */
2141 }
2142 }
2145 }
2147 /* Make sure we can handle the pin inside the loop */
2151 {
2156 {
2162 }
2163 else
2165 }
2166 }
2168 /* ---------------------------------------------------------------------
2169 * FlushDatabaseBuffers
2170 *
2171 * This function writes all dirty pages of a database out to disk
2172 * (or more accurately, out to kernel disk buffers), ensuring that the
2173 * kernel has an up-to-date view of the database.
2174 *
2175 * Generally, the caller should be holding an appropriate lock to ensure
2176 * no other backend is active in the target database; otherwise more
2177 * pages could get dirtied.
2178 *
2179 * Note we don't worry about flushing any pages of temporary relations.
2180 * It's assumed these wouldn't be interesting.
2181 * --------------------------------------------------------------------
2182 */
2183 void
2185 {
2189 /* Make sure we can handle the pin inside the loop */
2193 {
2198 {
2204 }
2205 else
2207 }
2208 }
2210 /*
2211 * ReleaseBuffer -- release the pin on a buffer
2212 */
2213 void
2215 {
2224 {
2228 }
2236 else
2238 }
2240 /*
2241 * UnlockReleaseBuffer -- release the content lock and pin on a buffer
2242 *
2243 * This is just a shorthand for a common combination.
2244 */
2245 void
2247 {
2250 }
2252 /*
2253 * IncrBufferRefCount
2254 * Increment the pin count on a buffer that we have *already* pinned
2255 * at least once.
2256 *
2257 * This function cannot be used on a buffer we do not have pinned,
2258 * because it doesn't change the shared buffer state.
2259 */
2260 void
2262 {
2268 else
2270 }
2272 /*
2273 * SetBufferCommitInfoNeedsSave
2274 *
2275 * Mark a buffer dirty when we have updated tuple commit-status bits in it.
2276 *
2277 * This is essentially the same as MarkBufferDirty, except that the caller
2278 * might have only share-lock instead of exclusive-lock on the buffer's
2279 * content lock. We preserve the distinction mainly as a way of documenting
2280 * that the caller has not made a critical data change --- the status-bit
2281 * update could be redone by someone else just as easily. Therefore, no WAL
2282 * log record need be generated, whereas calls to MarkBufferDirty really ought
2283 * to be associated with a WAL-entry-creating action.
2284 */
2285 void
2287 {
2294 {
2297 }
2302 /* here, either share or exclusive lock is OK */
2305 /*
2306 * This routine might get called many times on the same page, if we are
2307 * making the first scan after commit of an xact that added/deleted many
2308 * tuples. So, be as quick as we can if the buffer is already dirty. We
2309 * do this by not acquiring spinlock if it looks like the status bits are
2310 * already OK. (Note it is okay if someone else clears BM_JUST_DIRTIED
2311 * immediately after we look, because the buffer content update is already
2312 * done and will be reflected in the I/O.)
2313 */
2316 {
2323 }
2324 }
2326 /*
2327 * Release buffer content locks for shared buffers.
2328 *
2329 * Used to clean up after errors.
2330 *
2331 * Currently, we can expect that lwlock.c's LWLockReleaseAll() took care
2332 * of releasing buffer content locks per se; the only thing we need to deal
2333 * with here is clearing any PIN_COUNT request that was in progress.
2334 */
2335 void
2337 {
2341 {
2344 /*
2345 * Don't complain if flag bit not set; it could have been reset but we
2346 * got a cancel/die interrupt before getting the signal.
2347 */
2355 }
2356 }
2358 /*
2359 * Acquire or release the content_lock for the buffer.
2360 */
2361 void
2363 {
2378 else
2380 }
2382 /*
2383 * Acquire the content_lock for the buffer, but only if we don't have to wait.
2384 *
2385 * This assumes the caller wants BUFFER_LOCK_EXCLUSIVE mode.
2386 */
2387 bool
2389 {
2399 }
2401 /*
2402 * LockBufferForCleanup - lock a buffer in preparation for deleting items
2403 *
2404 * Items may be deleted from a disk page only when the caller (a) holds an
2405 * exclusive lock on the buffer and (b) has observed that no other backend
2406 * holds a pin on the buffer. If there is a pin, then the other backend
2407 * might have a pointer into the buffer (for example, a heapscan reference
2408 * to an item --- see README for more details). It's OK if a pin is added
2409 * after the cleanup starts, however; the newly-arrived backend will be
2410 * unable to look at the page until we release the exclusive lock.
2411 *
2412 * To implement this protocol, a would-be deleter must pin the buffer and
2413 * then call LockBufferForCleanup(). LockBufferForCleanup() is similar to
2414 * LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE), except that it loops until
2415 * it has successfully observed pin count = 1.
2416 */
2417 void
2419 {
2426 {
2427 /* There should be exactly one pin */
2431 /* Nobody else to wait for */
2433 }
2435 /* There should be exactly one local pin */
2443 {
2444 /* Try to acquire lock */
2449 {
2450 /* Successfully acquired exclusive lock with pincount 1 */
2453 }
2454 /* Failed, so mark myself as waiting for pincount 1 */
2456 {
2460 }
2466 /* Wait to be signaled by UnpinBuffer() */
2469 /* Loop back and try again */
2470 }
2471 }
2473 /*
2474 * ConditionalLockBufferForCleanup - as above, but don't wait to get the lock
2475 *
2476 * We won't loop, but just check once to see if the pin count is OK. If
2477 * not, return FALSE with no lock held.
2478 */
2479 bool
2481 {
2487 {
2488 /* There should be exactly one pin */
2492 /* Nobody else to wait for */
2494 }
2496 /* There should be exactly one local pin */
2501 /* Try to acquire lock */
2509 {
2510 /* Successfully acquired exclusive lock with pincount 1 */
2513 }
2515 /* Failed, so release the lock */
2519 }
2522 /*
2523 * Functions for buffer I/O handling
2524 *
2525 * Note: We assume that nested buffer I/O never occurs.
2526 * i.e at most one io_in_progress lock is held per proc.
2527 *
2528 * Also note that these are used only for shared buffers, not local ones.
2529 */
2531 /*
2532 * WaitIO -- Block until the IO_IN_PROGRESS flag on 'buf' is cleared.
2533 */
2534 static void
2536 {
2537 /*
2538 * Changed to wait until there's no IO - Inoue 01/13/2000
2539 *
2540 * Note this is *necessary* because an error abort in the process doing
2541 * I/O could release the io_in_progress_lock prematurely. See
2542 * AbortBufferIO.
2543 */
2545 {
2546 BufFlags sv_flags;
2548 /*
2549 * It may not be necessary to acquire the spinlock to check the flag
2550 * here, but since this test is essential for correctness, we'd better
2551 * play it safe.
2552 */
2560 }
2561 }
2563 /*
2564 * StartBufferIO: begin I/O on this buffer
2565 * (Assumptions)
2566 * My process is executing no IO
2567 * The buffer is Pinned
2568 *
2569 * In some scenarios there are race conditions in which multiple backends
2570 * could attempt the same I/O operation concurrently. If someone else
2571 * has already started I/O on this buffer then we will block on the
2572 * io_in_progress lock until he's done.
2573 *
2574 * Input operations are only attempted on buffers that are not BM_VALID,
2575 * and output operations only on buffers that are BM_VALID and BM_DIRTY,
2576 * so we can always tell if the work is already done.
2577 *
2578 * Returns TRUE if we successfully marked the buffer as I/O busy,
2579 * FALSE if someone else already did the work.
2580 */
2581 static bool
2583 {
2587 {
2588 /*
2589 * Grab the io_in_progress lock so that other processes can wait for
2590 * me to finish the I/O.
2591 */
2599 /*
2600 * The only way BM_IO_IN_PROGRESS could be set when the io_in_progress
2601 * lock isn't held is if the process doing the I/O is recovering from
2602 * an error (see AbortBufferIO). If that's the case, we must wait for
2603 * him to get unwedged.
2604 */
2608 }
2610 /* Once we get here, there is definitely no I/O active on this buffer */
2613 {
2614 /* someone else already did the I/O */
2618 }
2628 }
2630 /*
2631 * TerminateBufferIO: release a buffer we were doing I/O on
2632 * (Assumptions)
2633 * My process is executing IO for the buffer
2634 * BM_IO_IN_PROGRESS bit is set for the buffer
2635 * We hold the buffer's io_in_progress lock
2636 * The buffer is Pinned
2637 *
2638 * If clear_dirty is TRUE and BM_JUST_DIRTIED is not set, we clear the
2639 * buffer's BM_DIRTY flag. This is appropriate when terminating a
2640 * successful write. The check on BM_JUST_DIRTIED is necessary to avoid
2641 * marking the buffer clean if it was re-dirtied while we were writing.
2642 *
2643 * set_flag_bits gets ORed into the buffer's flags. It must include
2644 * BM_IO_ERROR in a failure case. For successful completion it could
2645 * be 0, or BM_VALID if we just finished reading in the page.
2646 */
2647 static void
2650 {
2666 }
2668 /*
2669 * AbortBufferIO: Clean up any active buffer I/O after an error.
2670 *
2671 * All LWLocks we might have held have been released,
2672 * but we haven't yet released buffer pins, so the buffer is still pinned.
2673 *
2674 * If I/O was in progress, we always set BM_IO_ERROR, even though it's
2675 * possible the error condition wasn't related to the I/O.
2676 */
2677 void
2679 {
2683 {
2684 /*
2685 * Since LWLockReleaseAll has already been called, we're not holding
2686 * the buffer's io_in_progress_lock. We have to re-acquire it so that
2687 * we can use TerminateBufferIO. Anyone who's executing WaitIO on the
2688 * buffer will be in a busy spin until we succeed in doing this.
2689 */
2695 {
2697 /* We'd better not think buffer is valid yet */
2700 }
2701 else
2702 {
2703 BufFlags sv_flags;
2708 /* Issue notice if this is not the first failure... */
2710 {
2711 /* Buffer is pinned, so we can read tag without spinlock */
2720 }
2721 }
2723 }
2724 }
2726 /*
2727 * Error context callback for errors occurring during buffer writes.
2728 */
2729 static void
2731 {
2734 /* Buffer is pinned, so we can read the tag without locking the spinlock */
2736 {
2742 }
2743 }