Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * bufmgr.c
4 : * buffer manager interface routines
5 : *
6 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/storage/buffer/bufmgr.c
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 : * StartReadBuffer() -- as above, with separate wait step
23 : * StartReadBuffers() -- multiple block version
24 : * WaitReadBuffers() -- second step of above
25 : *
26 : * ReleaseBuffer() -- unpin a buffer
27 : *
28 : * MarkBufferDirty() -- mark a pinned buffer's contents as "dirty".
29 : * The disk write is delayed until buffer replacement or checkpoint.
30 : *
31 : * See also these files:
32 : * freelist.c -- chooses victim for buffer replacement
33 : * buf_table.c -- manages the buffer lookup table
34 : */
35 : #include "postgres.h"
36 :
37 : #include <sys/file.h>
38 : #include <unistd.h>
39 :
40 : #include "access/tableam.h"
41 : #include "access/xloginsert.h"
42 : #include "access/xlogutils.h"
43 : #include "catalog/storage.h"
44 : #include "catalog/storage_xlog.h"
45 : #include "executor/instrument.h"
46 : #include "lib/binaryheap.h"
47 : #include "miscadmin.h"
48 : #include "pg_trace.h"
49 : #include "pgstat.h"
50 : #include "postmaster/bgwriter.h"
51 : #include "storage/aio.h"
52 : #include "storage/buf_internals.h"
53 : #include "storage/bufmgr.h"
54 : #include "storage/fd.h"
55 : #include "storage/ipc.h"
56 : #include "storage/lmgr.h"
57 : #include "storage/proc.h"
58 : #include "storage/read_stream.h"
59 : #include "storage/smgr.h"
60 : #include "storage/standby.h"
61 : #include "utils/memdebug.h"
62 : #include "utils/ps_status.h"
63 : #include "utils/rel.h"
64 : #include "utils/resowner.h"
65 : #include "utils/timestamp.h"
66 :
67 :
68 : /* Note: these two macros only work on shared buffers, not local ones! */
69 : #define BufHdrGetBlock(bufHdr) ((Block) (BufferBlocks + ((Size) (bufHdr)->buf_id) * BLCKSZ))
70 : #define BufferGetLSN(bufHdr) (PageGetLSN(BufHdrGetBlock(bufHdr)))
71 :
72 : /* Note: this macro only works on local buffers, not shared ones! */
73 : #define LocalBufHdrGetBlock(bufHdr) \
74 : LocalBufferBlockPointers[-((bufHdr)->buf_id + 2)]
75 :
76 : /* Bits in SyncOneBuffer's return value */
77 : #define BUF_WRITTEN 0x01
78 : #define BUF_REUSABLE 0x02
79 :
80 : #define RELS_BSEARCH_THRESHOLD 20
81 :
82 : /*
83 : * This is the size (in the number of blocks) above which we scan the
84 : * entire buffer pool to remove the buffers for all the pages of relation
85 : * being dropped. For the relations with size below this threshold, we find
86 : * the buffers by doing lookups in BufMapping table.
87 : */
88 : #define BUF_DROP_FULL_SCAN_THRESHOLD (uint64) (NBuffers / 32)
89 :
90 : typedef struct PrivateRefCountEntry
91 : {
92 : Buffer buffer;
93 : int32 refcount;
94 : } PrivateRefCountEntry;
95 :
96 : /* 64 bytes, about the size of a cache line on common systems */
97 : #define REFCOUNT_ARRAY_ENTRIES 8
98 :
99 : /*
100 : * Status of buffers to checkpoint for a particular tablespace, used
101 : * internally in BufferSync.
102 : */
103 : typedef struct CkptTsStatus
104 : {
105 : /* oid of the tablespace */
106 : Oid tsId;
107 :
108 : /*
109 : * Checkpoint progress for this tablespace. To make progress comparable
110 : * between tablespaces the progress is, for each tablespace, measured as a
111 : * number between 0 and the total number of to-be-checkpointed pages. Each
112 : * page checkpointed in this tablespace increments this space's progress
113 : * by progress_slice.
114 : */
115 : float8 progress;
116 : float8 progress_slice;
117 :
118 : /* number of to-be checkpointed pages in this tablespace */
119 : int num_to_scan;
120 : /* already processed pages in this tablespace */
121 : int num_scanned;
122 :
123 : /* current offset in CkptBufferIds for this tablespace */
124 : int index;
125 : } CkptTsStatus;
126 :
127 : /*
128 : * Type for array used to sort SMgrRelations
129 : *
130 : * FlushRelationsAllBuffers shares the same comparator function with
131 : * DropRelationsAllBuffers. Pointer to this struct and RelFileLocator must be
132 : * compatible.
133 : */
134 : typedef struct SMgrSortArray
135 : {
136 : RelFileLocator rlocator; /* This must be the first member */
137 : SMgrRelation srel;
138 : } SMgrSortArray;
139 :
140 : /* GUC variables */
141 : bool zero_damaged_pages = false;
142 : int bgwriter_lru_maxpages = 100;
143 : double bgwriter_lru_multiplier = 2.0;
144 : bool track_io_timing = false;
145 :
146 : /*
147 : * How many buffers PrefetchBuffer callers should try to stay ahead of their
148 : * ReadBuffer calls by. Zero means "never prefetch". This value is only used
149 : * for buffers not belonging to tablespaces that have their
150 : * effective_io_concurrency parameter set.
151 : */
152 : int effective_io_concurrency = DEFAULT_EFFECTIVE_IO_CONCURRENCY;
153 :
154 : /*
155 : * Like effective_io_concurrency, but used by maintenance code paths that might
156 : * benefit from a higher setting because they work on behalf of many sessions.
157 : * Overridden by the tablespace setting of the same name.
158 : */
159 : int maintenance_io_concurrency = DEFAULT_MAINTENANCE_IO_CONCURRENCY;
160 :
161 : /*
162 : * Limit on how many blocks should be handled in single I/O operations.
163 : * StartReadBuffers() callers should respect it, as should other operations
164 : * that call smgr APIs directly. It is computed as the minimum of underlying
165 : * GUCs io_combine_limit_guc and io_max_combine_limit.
166 : */
167 : int io_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
168 : int io_combine_limit_guc = DEFAULT_IO_COMBINE_LIMIT;
169 : int io_max_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
170 :
171 : /*
172 : * GUC variables about triggering kernel writeback for buffers written; OS
173 : * dependent defaults are set via the GUC mechanism.
174 : */
175 : int checkpoint_flush_after = DEFAULT_CHECKPOINT_FLUSH_AFTER;
176 : int bgwriter_flush_after = DEFAULT_BGWRITER_FLUSH_AFTER;
177 : int backend_flush_after = DEFAULT_BACKEND_FLUSH_AFTER;
178 :
179 : /* local state for LockBufferForCleanup */
180 : static BufferDesc *PinCountWaitBuf = NULL;
181 :
182 : /*
183 : * Backend-Private refcount management:
184 : *
185 : * Each buffer also has a private refcount that keeps track of the number of
186 : * times the buffer is pinned in the current process. This is so that the
187 : * shared refcount needs to be modified only once if a buffer is pinned more
188 : * than once by an individual backend. It's also used to check that no buffers
189 : * are still pinned at the end of transactions and when exiting.
190 : *
191 : *
192 : * To avoid - as we used to - requiring an array with NBuffers entries to keep
193 : * track of local buffers, we use a small sequentially searched array
194 : * (PrivateRefCountArray) and an overflow hash table (PrivateRefCountHash) to
195 : * keep track of backend local pins.
196 : *
197 : * Until no more than REFCOUNT_ARRAY_ENTRIES buffers are pinned at once, all
198 : * refcounts are kept track of in the array; after that, new array entries
199 : * displace old ones into the hash table. That way a frequently used entry
200 : * can't get "stuck" in the hashtable while infrequent ones clog the array.
201 : *
202 : * Note that in most scenarios the number of pinned buffers will not exceed
203 : * REFCOUNT_ARRAY_ENTRIES.
204 : *
205 : *
206 : * To enter a buffer into the refcount tracking mechanism first reserve a free
207 : * entry using ReservePrivateRefCountEntry() and then later, if necessary,
208 : * fill it with NewPrivateRefCountEntry(). That split lets us avoid doing
209 : * memory allocations in NewPrivateRefCountEntry() which can be important
210 : * because in some scenarios it's called with a spinlock held...
211 : */
212 : static struct PrivateRefCountEntry PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES];
213 : static HTAB *PrivateRefCountHash = NULL;
214 : static int32 PrivateRefCountOverflowed = 0;
215 : static uint32 PrivateRefCountClock = 0;
216 : static PrivateRefCountEntry *ReservedRefCountEntry = NULL;
217 :
218 : static uint32 MaxProportionalPins;
219 :
220 : static void ReservePrivateRefCountEntry(void);
221 : static PrivateRefCountEntry *NewPrivateRefCountEntry(Buffer buffer);
222 : static PrivateRefCountEntry *GetPrivateRefCountEntry(Buffer buffer, bool do_move);
223 : static inline int32 GetPrivateRefCount(Buffer buffer);
224 : static void ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref);
225 :
226 : /* ResourceOwner callbacks to hold in-progress I/Os and buffer pins */
227 : static void ResOwnerReleaseBufferIO(Datum res);
228 : static char *ResOwnerPrintBufferIO(Datum res);
229 : static void ResOwnerReleaseBufferPin(Datum res);
230 : static char *ResOwnerPrintBufferPin(Datum res);
231 :
232 : const ResourceOwnerDesc buffer_io_resowner_desc =
233 : {
234 : .name = "buffer io",
235 : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
236 : .release_priority = RELEASE_PRIO_BUFFER_IOS,
237 : .ReleaseResource = ResOwnerReleaseBufferIO,
238 : .DebugPrint = ResOwnerPrintBufferIO
239 : };
240 :
241 : const ResourceOwnerDesc buffer_pin_resowner_desc =
242 : {
243 : .name = "buffer pin",
244 : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
245 : .release_priority = RELEASE_PRIO_BUFFER_PINS,
246 : .ReleaseResource = ResOwnerReleaseBufferPin,
247 : .DebugPrint = ResOwnerPrintBufferPin
248 : };
249 :
250 : /*
251 : * Ensure that the PrivateRefCountArray has sufficient space to store one more
252 : * entry. This has to be called before using NewPrivateRefCountEntry() to fill
253 : * a new entry - but it's perfectly fine to not use a reserved entry.
254 : */
255 : static void
256 124059052 : ReservePrivateRefCountEntry(void)
257 : {
258 : /* Already reserved (or freed), nothing to do */
259 124059052 : if (ReservedRefCountEntry != NULL)
260 115775282 : return;
261 :
262 : /*
263 : * First search for a free entry the array, that'll be sufficient in the
264 : * majority of cases.
265 : */
266 : {
267 : int i;
268 :
269 21325912 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
270 : {
271 : PrivateRefCountEntry *res;
272 :
273 20973420 : res = &PrivateRefCountArray[i];
274 :
275 20973420 : if (res->buffer == InvalidBuffer)
276 : {
277 7931278 : ReservedRefCountEntry = res;
278 7931278 : return;
279 : }
280 : }
281 : }
282 :
283 : /*
284 : * No luck. All array entries are full. Move one array entry into the hash
285 : * table.
286 : */
287 : {
288 : /*
289 : * Move entry from the current clock position in the array into the
290 : * hashtable. Use that slot.
291 : */
292 : PrivateRefCountEntry *hashent;
293 : bool found;
294 :
295 : /* select victim slot */
296 352492 : ReservedRefCountEntry =
297 352492 : &PrivateRefCountArray[PrivateRefCountClock++ % REFCOUNT_ARRAY_ENTRIES];
298 :
299 : /* Better be used, otherwise we shouldn't get here. */
300 : Assert(ReservedRefCountEntry->buffer != InvalidBuffer);
301 :
302 : /* enter victim array entry into hashtable */
303 352492 : hashent = hash_search(PrivateRefCountHash,
304 352492 : &(ReservedRefCountEntry->buffer),
305 : HASH_ENTER,
306 : &found);
307 : Assert(!found);
308 352492 : hashent->refcount = ReservedRefCountEntry->refcount;
309 :
310 : /* clear the now free array slot */
311 352492 : ReservedRefCountEntry->buffer = InvalidBuffer;
312 352492 : ReservedRefCountEntry->refcount = 0;
313 :
314 352492 : PrivateRefCountOverflowed++;
315 : }
316 : }
317 :
318 : /*
319 : * Fill a previously reserved refcount entry.
320 : */
321 : static PrivateRefCountEntry *
322 112825636 : NewPrivateRefCountEntry(Buffer buffer)
323 : {
324 : PrivateRefCountEntry *res;
325 :
326 : /* only allowed to be called when a reservation has been made */
327 : Assert(ReservedRefCountEntry != NULL);
328 :
329 : /* use up the reserved entry */
330 112825636 : res = ReservedRefCountEntry;
331 112825636 : ReservedRefCountEntry = NULL;
332 :
333 : /* and fill it */
334 112825636 : res->buffer = buffer;
335 112825636 : res->refcount = 0;
336 :
337 112825636 : return res;
338 : }
339 :
340 : /*
341 : * Return the PrivateRefCount entry for the passed buffer.
342 : *
343 : * Returns NULL if a buffer doesn't have a refcount entry. Otherwise, if
344 : * do_move is true, and the entry resides in the hashtable the entry is
345 : * optimized for frequent access by moving it to the array.
346 : */
347 : static PrivateRefCountEntry *
348 277769762 : GetPrivateRefCountEntry(Buffer buffer, bool do_move)
349 : {
350 : PrivateRefCountEntry *res;
351 : int i;
352 :
353 : Assert(BufferIsValid(buffer));
354 : Assert(!BufferIsLocal(buffer));
355 :
356 : /*
357 : * First search for references in the array, that'll be sufficient in the
358 : * majority of cases.
359 : */
360 1310877332 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
361 : {
362 1201970706 : res = &PrivateRefCountArray[i];
363 :
364 1201970706 : if (res->buffer == buffer)
365 168863136 : return res;
366 : }
367 :
368 : /*
369 : * By here we know that the buffer, if already pinned, isn't residing in
370 : * the array.
371 : *
372 : * Only look up the buffer in the hashtable if we've previously overflowed
373 : * into it.
374 : */
375 108906626 : if (PrivateRefCountOverflowed == 0)
376 108083998 : return NULL;
377 :
378 822628 : res = hash_search(PrivateRefCountHash, &buffer, HASH_FIND, NULL);
379 :
380 822628 : if (res == NULL)
381 467022 : return NULL;
382 355606 : else if (!do_move)
383 : {
384 : /* caller doesn't want us to move the hash entry into the array */
385 315848 : return res;
386 : }
387 : else
388 : {
389 : /* move buffer from hashtable into the free array slot */
390 : bool found;
391 : PrivateRefCountEntry *free;
392 :
393 : /* Ensure there's a free array slot */
394 39758 : ReservePrivateRefCountEntry();
395 :
396 : /* Use up the reserved slot */
397 : Assert(ReservedRefCountEntry != NULL);
398 39758 : free = ReservedRefCountEntry;
399 39758 : ReservedRefCountEntry = NULL;
400 : Assert(free->buffer == InvalidBuffer);
401 :
402 : /* and fill it */
403 39758 : free->buffer = buffer;
404 39758 : free->refcount = res->refcount;
405 :
406 : /* delete from hashtable */
407 39758 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
408 : Assert(found);
409 : Assert(PrivateRefCountOverflowed > 0);
410 39758 : PrivateRefCountOverflowed--;
411 :
412 39758 : return free;
413 : }
414 : }
415 :
416 : /*
417 : * Returns how many times the passed buffer is pinned by this backend.
418 : *
419 : * Only works for shared memory buffers!
420 : */
421 : static inline int32
422 5406714 : GetPrivateRefCount(Buffer buffer)
423 : {
424 : PrivateRefCountEntry *ref;
425 :
426 : Assert(BufferIsValid(buffer));
427 : Assert(!BufferIsLocal(buffer));
428 :
429 : /*
430 : * Not moving the entry - that's ok for the current users, but we might
431 : * want to change this one day.
432 : */
433 5406714 : ref = GetPrivateRefCountEntry(buffer, false);
434 :
435 5406714 : if (ref == NULL)
436 8084 : return 0;
437 5398630 : return ref->refcount;
438 : }
439 :
440 : /*
441 : * Release resources used to track the reference count of a buffer which we no
442 : * longer have pinned and don't want to pin again immediately.
443 : */
444 : static void
445 112825636 : ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref)
446 : {
447 : Assert(ref->refcount == 0);
448 :
449 112825636 : if (ref >= &PrivateRefCountArray[0] &&
450 : ref < &PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES])
451 : {
452 112512902 : ref->buffer = InvalidBuffer;
453 :
454 : /*
455 : * Mark the just used entry as reserved - in many scenarios that
456 : * allows us to avoid ever having to search the array/hash for free
457 : * entries.
458 : */
459 112512902 : ReservedRefCountEntry = ref;
460 : }
461 : else
462 : {
463 : bool found;
464 312734 : Buffer buffer = ref->buffer;
465 :
466 312734 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
467 : Assert(found);
468 : Assert(PrivateRefCountOverflowed > 0);
469 312734 : PrivateRefCountOverflowed--;
470 : }
471 112825636 : }
472 :
473 : /*
474 : * BufferIsPinned
475 : * True iff the buffer is pinned (also checks for valid buffer number).
476 : *
477 : * NOTE: what we check here is that *this* backend holds a pin on
478 : * the buffer. We do not care whether some other backend does.
479 : */
480 : #define BufferIsPinned(bufnum) \
481 : ( \
482 : !BufferIsValid(bufnum) ? \
483 : false \
484 : : \
485 : BufferIsLocal(bufnum) ? \
486 : (LocalRefCount[-(bufnum) - 1] > 0) \
487 : : \
488 : (GetPrivateRefCount(bufnum) > 0) \
489 : )
490 :
491 :
492 : static Buffer ReadBuffer_common(Relation rel,
493 : SMgrRelation smgr, char smgr_persistence,
494 : ForkNumber forkNum, BlockNumber blockNum,
495 : ReadBufferMode mode, BufferAccessStrategy strategy);
496 : static BlockNumber ExtendBufferedRelCommon(BufferManagerRelation bmr,
497 : ForkNumber fork,
498 : BufferAccessStrategy strategy,
499 : uint32 flags,
500 : uint32 extend_by,
501 : BlockNumber extend_upto,
502 : Buffer *buffers,
503 : uint32 *extended_by);
504 : static BlockNumber ExtendBufferedRelShared(BufferManagerRelation bmr,
505 : ForkNumber fork,
506 : BufferAccessStrategy strategy,
507 : uint32 flags,
508 : uint32 extend_by,
509 : BlockNumber extend_upto,
510 : Buffer *buffers,
511 : uint32 *extended_by);
512 : static bool PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy);
513 : static void PinBuffer_Locked(BufferDesc *buf);
514 : static void UnpinBuffer(BufferDesc *buf);
515 : static void UnpinBufferNoOwner(BufferDesc *buf);
516 : static void BufferSync(int flags);
517 : static uint32 WaitBufHdrUnlocked(BufferDesc *buf);
518 : static int SyncOneBuffer(int buf_id, bool skip_recently_used,
519 : WritebackContext *wb_context);
520 : static void WaitIO(BufferDesc *buf);
521 : static void AbortBufferIO(Buffer buffer);
522 : static void shared_buffer_write_error_callback(void *arg);
523 : static void local_buffer_write_error_callback(void *arg);
524 : static inline BufferDesc *BufferAlloc(SMgrRelation smgr,
525 : char relpersistence,
526 : ForkNumber forkNum,
527 : BlockNumber blockNum,
528 : BufferAccessStrategy strategy,
529 : bool *foundPtr, IOContext io_context);
530 : static bool AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks);
531 : static void CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete);
532 : static Buffer GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context);
533 : static void FlushBuffer(BufferDesc *buf, SMgrRelation reln,
534 : IOObject io_object, IOContext io_context);
535 : static void FindAndDropRelationBuffers(RelFileLocator rlocator,
536 : ForkNumber forkNum,
537 : BlockNumber nForkBlock,
538 : BlockNumber firstDelBlock);
539 : static void RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
540 : RelFileLocator dstlocator,
541 : ForkNumber forkNum, bool permanent);
542 : static void AtProcExit_Buffers(int code, Datum arg);
543 : static void CheckForBufferLeaks(void);
544 : static int rlocator_comparator(const void *p1, const void *p2);
545 : static inline int buffertag_comparator(const BufferTag *ba, const BufferTag *bb);
546 : static inline int ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b);
547 : static int ts_ckpt_progress_comparator(Datum a, Datum b, void *arg);
548 :
549 :
550 : /*
551 : * Implementation of PrefetchBuffer() for shared buffers.
552 : */
553 : PrefetchBufferResult
554 61566 : PrefetchSharedBuffer(SMgrRelation smgr_reln,
555 : ForkNumber forkNum,
556 : BlockNumber blockNum)
557 : {
558 61566 : PrefetchBufferResult result = {InvalidBuffer, false};
559 : BufferTag newTag; /* identity of requested block */
560 : uint32 newHash; /* hash value for newTag */
561 : LWLock *newPartitionLock; /* buffer partition lock for it */
562 : int buf_id;
563 :
564 : Assert(BlockNumberIsValid(blockNum));
565 :
566 : /* create a tag so we can lookup the buffer */
567 61566 : InitBufferTag(&newTag, &smgr_reln->smgr_rlocator.locator,
568 : forkNum, blockNum);
569 :
570 : /* determine its hash code and partition lock ID */
571 61566 : newHash = BufTableHashCode(&newTag);
572 61566 : newPartitionLock = BufMappingPartitionLock(newHash);
573 :
574 : /* see if the block is in the buffer pool already */
575 61566 : LWLockAcquire(newPartitionLock, LW_SHARED);
576 61566 : buf_id = BufTableLookup(&newTag, newHash);
577 61566 : LWLockRelease(newPartitionLock);
578 :
579 : /* If not in buffers, initiate prefetch */
580 61566 : if (buf_id < 0)
581 : {
582 : #ifdef USE_PREFETCH
583 : /*
584 : * Try to initiate an asynchronous read. This returns false in
585 : * recovery if the relation file doesn't exist.
586 : */
587 34072 : if ((io_direct_flags & IO_DIRECT_DATA) == 0 &&
588 16812 : smgrprefetch(smgr_reln, forkNum, blockNum, 1))
589 : {
590 16812 : result.initiated_io = true;
591 : }
592 : #endif /* USE_PREFETCH */
593 : }
594 : else
595 : {
596 : /*
597 : * Report the buffer it was in at that time. The caller may be able
598 : * to avoid a buffer table lookup, but it's not pinned and it must be
599 : * rechecked!
600 : */
601 44306 : result.recent_buffer = buf_id + 1;
602 : }
603 :
604 : /*
605 : * If the block *is* in buffers, we do nothing. This is not really ideal:
606 : * the block might be just about to be evicted, which would be stupid
607 : * since we know we are going to need it soon. But the only easy answer
608 : * is to bump the usage_count, which does not seem like a great solution:
609 : * when the caller does ultimately touch the block, usage_count would get
610 : * bumped again, resulting in too much favoritism for blocks that are
611 : * involved in a prefetch sequence. A real fix would involve some
612 : * additional per-buffer state, and it's not clear that there's enough of
613 : * a problem to justify that.
614 : */
615 :
616 61566 : return result;
617 : }
618 :
619 : /*
620 : * PrefetchBuffer -- initiate asynchronous read of a block of a relation
621 : *
622 : * This is named by analogy to ReadBuffer but doesn't actually allocate a
623 : * buffer. Instead it tries to ensure that a future ReadBuffer for the given
624 : * block will not be delayed by the I/O. Prefetching is optional.
625 : *
626 : * There are three possible outcomes:
627 : *
628 : * 1. If the block is already cached, the result includes a valid buffer that
629 : * could be used by the caller to avoid the need for a later buffer lookup, but
630 : * it's not pinned, so the caller must recheck it.
631 : *
632 : * 2. If the kernel has been asked to initiate I/O, the initiated_io member is
633 : * true. Currently there is no way to know if the data was already cached by
634 : * the kernel and therefore didn't really initiate I/O, and no way to know when
635 : * the I/O completes other than using synchronous ReadBuffer().
636 : *
637 : * 3. Otherwise, the buffer wasn't already cached by PostgreSQL, and
638 : * USE_PREFETCH is not defined (this build doesn't support prefetching due to
639 : * lack of a kernel facility), direct I/O is enabled, or the underlying
640 : * relation file wasn't found and we are in recovery. (If the relation file
641 : * wasn't found and we are not in recovery, an error is raised).
642 : */
643 : PrefetchBufferResult
644 41506 : PrefetchBuffer(Relation reln, ForkNumber forkNum, BlockNumber blockNum)
645 : {
646 : Assert(RelationIsValid(reln));
647 : Assert(BlockNumberIsValid(blockNum));
648 :
649 41506 : if (RelationUsesLocalBuffers(reln))
650 : {
651 : /* see comments in ReadBufferExtended */
652 1566 : if (RELATION_IS_OTHER_TEMP(reln))
653 0 : ereport(ERROR,
654 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
655 : errmsg("cannot access temporary tables of other sessions")));
656 :
657 : /* pass it off to localbuf.c */
658 1566 : return PrefetchLocalBuffer(RelationGetSmgr(reln), forkNum, blockNum);
659 : }
660 : else
661 : {
662 : /* pass it to the shared buffer version */
663 39940 : return PrefetchSharedBuffer(RelationGetSmgr(reln), forkNum, blockNum);
664 : }
665 : }
666 :
667 : /*
668 : * ReadRecentBuffer -- try to pin a block in a recently observed buffer
669 : *
670 : * Compared to ReadBuffer(), this avoids a buffer mapping lookup when it's
671 : * successful. Return true if the buffer is valid and still has the expected
672 : * tag. In that case, the buffer is pinned and the usage count is bumped.
673 : */
674 : bool
675 8100 : ReadRecentBuffer(RelFileLocator rlocator, ForkNumber forkNum, BlockNumber blockNum,
676 : Buffer recent_buffer)
677 : {
678 : BufferDesc *bufHdr;
679 : BufferTag tag;
680 : uint32 buf_state;
681 : bool have_private_ref;
682 :
683 : Assert(BufferIsValid(recent_buffer));
684 :
685 8100 : ResourceOwnerEnlarge(CurrentResourceOwner);
686 8100 : ReservePrivateRefCountEntry();
687 8100 : InitBufferTag(&tag, &rlocator, forkNum, blockNum);
688 :
689 8100 : if (BufferIsLocal(recent_buffer))
690 : {
691 64 : int b = -recent_buffer - 1;
692 :
693 64 : bufHdr = GetLocalBufferDescriptor(b);
694 64 : buf_state = pg_atomic_read_u32(&bufHdr->state);
695 :
696 : /* Is it still valid and holding the right tag? */
697 64 : if ((buf_state & BM_VALID) && BufferTagsEqual(&tag, &bufHdr->tag))
698 : {
699 64 : PinLocalBuffer(bufHdr, true);
700 :
701 64 : pgBufferUsage.local_blks_hit++;
702 :
703 64 : return true;
704 : }
705 : }
706 : else
707 : {
708 8036 : bufHdr = GetBufferDescriptor(recent_buffer - 1);
709 8036 : have_private_ref = GetPrivateRefCount(recent_buffer) > 0;
710 :
711 : /*
712 : * Do we already have this buffer pinned with a private reference? If
713 : * so, it must be valid and it is safe to check the tag without
714 : * locking. If not, we have to lock the header first and then check.
715 : */
716 8036 : if (have_private_ref)
717 0 : buf_state = pg_atomic_read_u32(&bufHdr->state);
718 : else
719 8036 : buf_state = LockBufHdr(bufHdr);
720 :
721 8036 : if ((buf_state & BM_VALID) && BufferTagsEqual(&tag, &bufHdr->tag))
722 : {
723 : /*
724 : * It's now safe to pin the buffer. We can't pin first and ask
725 : * questions later, because it might confuse code paths like
726 : * InvalidateBuffer() if we pinned a random non-matching buffer.
727 : */
728 7964 : if (have_private_ref)
729 0 : PinBuffer(bufHdr, NULL); /* bump pin count */
730 : else
731 7964 : PinBuffer_Locked(bufHdr); /* pin for first time */
732 :
733 7964 : pgBufferUsage.shared_blks_hit++;
734 :
735 7964 : return true;
736 : }
737 :
738 : /* If we locked the header above, now unlock. */
739 72 : if (!have_private_ref)
740 72 : UnlockBufHdr(bufHdr, buf_state);
741 : }
742 :
743 72 : return false;
744 : }
745 :
746 : /*
747 : * ReadBuffer -- a shorthand for ReadBufferExtended, for reading from main
748 : * fork with RBM_NORMAL mode and default strategy.
749 : */
750 : Buffer
751 82782354 : ReadBuffer(Relation reln, BlockNumber blockNum)
752 : {
753 82782354 : return ReadBufferExtended(reln, MAIN_FORKNUM, blockNum, RBM_NORMAL, NULL);
754 : }
755 :
756 : /*
757 : * ReadBufferExtended -- returns a buffer containing the requested
758 : * block of the requested relation. If the blknum
759 : * requested is P_NEW, extend the relation file and
760 : * allocate a new block. (Caller is responsible for
761 : * ensuring that only one backend tries to extend a
762 : * relation at the same time!)
763 : *
764 : * Returns: the buffer number for the buffer containing
765 : * the block read. The returned buffer has been pinned.
766 : * Does not return on error --- elog's instead.
767 : *
768 : * Assume when this function is called, that reln has been opened already.
769 : *
770 : * In RBM_NORMAL mode, the page is read from disk, and the page header is
771 : * validated. An error is thrown if the page header is not valid. (But
772 : * note that an all-zero page is considered "valid"; see
773 : * PageIsVerified().)
774 : *
775 : * RBM_ZERO_ON_ERROR is like the normal mode, but if the page header is not
776 : * valid, the page is zeroed instead of throwing an error. This is intended
777 : * for non-critical data, where the caller is prepared to repair errors.
778 : *
779 : * In RBM_ZERO_AND_LOCK mode, if the page isn't in buffer cache already, it's
780 : * filled with zeros instead of reading it from disk. Useful when the caller
781 : * is going to fill the page from scratch, since this saves I/O and avoids
782 : * unnecessary failure if the page-on-disk has corrupt page headers.
783 : * The page is returned locked to ensure that the caller has a chance to
784 : * initialize the page before it's made visible to others.
785 : * Caution: do not use this mode to read a page that is beyond the relation's
786 : * current physical EOF; that is likely to cause problems in md.c when
787 : * the page is modified and written out. P_NEW is OK, though.
788 : *
789 : * RBM_ZERO_AND_CLEANUP_LOCK is the same as RBM_ZERO_AND_LOCK, but acquires
790 : * a cleanup-strength lock on the page.
791 : *
792 : * RBM_NORMAL_NO_LOG mode is treated the same as RBM_NORMAL here.
793 : *
794 : * If strategy is not NULL, a nondefault buffer access strategy is used.
795 : * See buffer/README for details.
796 : */
797 : inline Buffer
798 99876368 : ReadBufferExtended(Relation reln, ForkNumber forkNum, BlockNumber blockNum,
799 : ReadBufferMode mode, BufferAccessStrategy strategy)
800 : {
801 : Buffer buf;
802 :
803 : /*
804 : * Reject attempts to read non-local temporary relations; we would be
805 : * likely to get wrong data since we have no visibility into the owning
806 : * session's local buffers.
807 : */
808 99876368 : if (RELATION_IS_OTHER_TEMP(reln))
809 0 : ereport(ERROR,
810 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
811 : errmsg("cannot access temporary tables of other sessions")));
812 :
813 : /*
814 : * Read the buffer, and update pgstat counters to reflect a cache hit or
815 : * miss.
816 : */
817 99876368 : buf = ReadBuffer_common(reln, RelationGetSmgr(reln), 0,
818 : forkNum, blockNum, mode, strategy);
819 :
820 99876322 : return buf;
821 : }
822 :
823 :
824 : /*
825 : * ReadBufferWithoutRelcache -- like ReadBufferExtended, but doesn't require
826 : * a relcache entry for the relation.
827 : *
828 : * Pass permanent = true for a RELPERSISTENCE_PERMANENT relation, and
829 : * permanent = false for a RELPERSISTENCE_UNLOGGED relation. This function
830 : * cannot be used for temporary relations (and making that work might be
831 : * difficult, unless we only want to read temporary relations for our own
832 : * ProcNumber).
833 : */
834 : Buffer
835 11270968 : ReadBufferWithoutRelcache(RelFileLocator rlocator, ForkNumber forkNum,
836 : BlockNumber blockNum, ReadBufferMode mode,
837 : BufferAccessStrategy strategy, bool permanent)
838 : {
839 11270968 : SMgrRelation smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
840 :
841 11270968 : return ReadBuffer_common(NULL, smgr,
842 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
843 : forkNum, blockNum,
844 : mode, strategy);
845 : }
846 :
847 : /*
848 : * Convenience wrapper around ExtendBufferedRelBy() extending by one block.
849 : */
850 : Buffer
851 93488 : ExtendBufferedRel(BufferManagerRelation bmr,
852 : ForkNumber forkNum,
853 : BufferAccessStrategy strategy,
854 : uint32 flags)
855 : {
856 : Buffer buf;
857 93488 : uint32 extend_by = 1;
858 :
859 93488 : ExtendBufferedRelBy(bmr, forkNum, strategy, flags, extend_by,
860 : &buf, &extend_by);
861 :
862 93488 : return buf;
863 : }
864 :
865 : /*
866 : * Extend relation by multiple blocks.
867 : *
868 : * Tries to extend the relation by extend_by blocks. Depending on the
869 : * availability of resources the relation may end up being extended by a
870 : * smaller number of pages (unless an error is thrown, always by at least one
871 : * page). *extended_by is updated to the number of pages the relation has been
872 : * extended to.
873 : *
874 : * buffers needs to be an array that is at least extend_by long. Upon
875 : * completion, the first extend_by array elements will point to a pinned
876 : * buffer.
877 : *
878 : * If EB_LOCK_FIRST is part of flags, the first returned buffer is
879 : * locked. This is useful for callers that want a buffer that is guaranteed to
880 : * be empty.
881 : */
882 : BlockNumber
883 318730 : ExtendBufferedRelBy(BufferManagerRelation bmr,
884 : ForkNumber fork,
885 : BufferAccessStrategy strategy,
886 : uint32 flags,
887 : uint32 extend_by,
888 : Buffer *buffers,
889 : uint32 *extended_by)
890 : {
891 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
892 : Assert(bmr.smgr == NULL || bmr.relpersistence != 0);
893 : Assert(extend_by > 0);
894 :
895 318730 : if (bmr.smgr == NULL)
896 : {
897 318730 : bmr.smgr = RelationGetSmgr(bmr.rel);
898 318730 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
899 : }
900 :
901 318730 : return ExtendBufferedRelCommon(bmr, fork, strategy, flags,
902 : extend_by, InvalidBlockNumber,
903 : buffers, extended_by);
904 : }
905 :
906 : /*
907 : * Extend the relation so it is at least extend_to blocks large, return buffer
908 : * (extend_to - 1).
909 : *
910 : * This is useful for callers that want to write a specific page, regardless
911 : * of the current size of the relation (e.g. useful for visibilitymap and for
912 : * crash recovery).
913 : */
914 : Buffer
915 100316 : ExtendBufferedRelTo(BufferManagerRelation bmr,
916 : ForkNumber fork,
917 : BufferAccessStrategy strategy,
918 : uint32 flags,
919 : BlockNumber extend_to,
920 : ReadBufferMode mode)
921 : {
922 : BlockNumber current_size;
923 100316 : uint32 extended_by = 0;
924 100316 : Buffer buffer = InvalidBuffer;
925 : Buffer buffers[64];
926 :
927 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
928 : Assert(bmr.smgr == NULL || bmr.relpersistence != 0);
929 : Assert(extend_to != InvalidBlockNumber && extend_to > 0);
930 :
931 100316 : if (bmr.smgr == NULL)
932 : {
933 13690 : bmr.smgr = RelationGetSmgr(bmr.rel);
934 13690 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
935 : }
936 :
937 : /*
938 : * If desired, create the file if it doesn't exist. If
939 : * smgr_cached_nblocks[fork] is positive then it must exist, no need for
940 : * an smgrexists call.
941 : */
942 100316 : if ((flags & EB_CREATE_FORK_IF_NEEDED) &&
943 13690 : (bmr.smgr->smgr_cached_nblocks[fork] == 0 ||
944 42 : bmr.smgr->smgr_cached_nblocks[fork] == InvalidBlockNumber) &&
945 13648 : !smgrexists(bmr.smgr, fork))
946 : {
947 13610 : LockRelationForExtension(bmr.rel, ExclusiveLock);
948 :
949 : /* recheck, fork might have been created concurrently */
950 13610 : if (!smgrexists(bmr.smgr, fork))
951 13606 : smgrcreate(bmr.smgr, fork, flags & EB_PERFORMING_RECOVERY);
952 :
953 13610 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
954 : }
955 :
956 : /*
957 : * If requested, invalidate size cache, so that smgrnblocks asks the
958 : * kernel.
959 : */
960 100316 : if (flags & EB_CLEAR_SIZE_CACHE)
961 13690 : bmr.smgr->smgr_cached_nblocks[fork] = InvalidBlockNumber;
962 :
963 : /*
964 : * Estimate how many pages we'll need to extend by. This avoids acquiring
965 : * unnecessarily many victim buffers.
966 : */
967 100316 : current_size = smgrnblocks(bmr.smgr, fork);
968 :
969 : /*
970 : * Since no-one else can be looking at the page contents yet, there is no
971 : * difference between an exclusive lock and a cleanup-strength lock. Note
972 : * that we pass the original mode to ReadBuffer_common() below, when
973 : * falling back to reading the buffer to a concurrent relation extension.
974 : */
975 100316 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
976 85910 : flags |= EB_LOCK_TARGET;
977 :
978 204786 : while (current_size < extend_to)
979 : {
980 104470 : uint32 num_pages = lengthof(buffers);
981 : BlockNumber first_block;
982 :
983 104470 : if ((uint64) current_size + num_pages > extend_to)
984 104338 : num_pages = extend_to - current_size;
985 :
986 104470 : first_block = ExtendBufferedRelCommon(bmr, fork, strategy, flags,
987 : num_pages, extend_to,
988 : buffers, &extended_by);
989 :
990 104470 : current_size = first_block + extended_by;
991 : Assert(num_pages != 0 || current_size >= extend_to);
992 :
993 222886 : for (uint32 i = 0; i < extended_by; i++)
994 : {
995 118416 : if (first_block + i != extend_to - 1)
996 18126 : ReleaseBuffer(buffers[i]);
997 : else
998 100290 : buffer = buffers[i];
999 : }
1000 : }
1001 :
1002 : /*
1003 : * It's possible that another backend concurrently extended the relation.
1004 : * In that case read the buffer.
1005 : *
1006 : * XXX: Should we control this via a flag?
1007 : */
1008 100316 : if (buffer == InvalidBuffer)
1009 : {
1010 : Assert(extended_by == 0);
1011 26 : buffer = ReadBuffer_common(bmr.rel, bmr.smgr, bmr.relpersistence,
1012 : fork, extend_to - 1, mode, strategy);
1013 : }
1014 :
1015 100316 : return buffer;
1016 : }
1017 :
1018 : /*
1019 : * Lock and optionally zero a buffer, as part of the implementation of
1020 : * RBM_ZERO_AND_LOCK or RBM_ZERO_AND_CLEANUP_LOCK. The buffer must be already
1021 : * pinned. If the buffer is not already valid, it is zeroed and made valid.
1022 : */
1023 : static void
1024 603874 : ZeroAndLockBuffer(Buffer buffer, ReadBufferMode mode, bool already_valid)
1025 : {
1026 : BufferDesc *bufHdr;
1027 : bool need_to_zero;
1028 603874 : bool isLocalBuf = BufferIsLocal(buffer);
1029 :
1030 : Assert(mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK);
1031 :
1032 603874 : if (already_valid)
1033 : {
1034 : /*
1035 : * If the caller already knew the buffer was valid, we can skip some
1036 : * header interaction. The caller just wants to lock the buffer.
1037 : */
1038 74158 : need_to_zero = false;
1039 : }
1040 529716 : else if (isLocalBuf)
1041 : {
1042 : /* Simple case for non-shared buffers. */
1043 48 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
1044 48 : need_to_zero = StartLocalBufferIO(bufHdr, true, false);
1045 : }
1046 : else
1047 : {
1048 : /*
1049 : * Take BM_IO_IN_PROGRESS, or discover that BM_VALID has been set
1050 : * concurrently. Even though we aren't doing I/O, that ensures that
1051 : * we don't zero a page that someone else has pinned. An exclusive
1052 : * content lock wouldn't be enough, because readers are allowed to
1053 : * drop the content lock after determining that a tuple is visible
1054 : * (see buffer access rules in README).
1055 : */
1056 529668 : bufHdr = GetBufferDescriptor(buffer - 1);
1057 529668 : need_to_zero = StartBufferIO(bufHdr, true, false);
1058 : }
1059 :
1060 603874 : if (need_to_zero)
1061 : {
1062 529716 : memset(BufferGetPage(buffer), 0, BLCKSZ);
1063 :
1064 : /*
1065 : * Grab the buffer content lock before marking the page as valid, to
1066 : * make sure that no other backend sees the zeroed page before the
1067 : * caller has had a chance to initialize it.
1068 : *
1069 : * Since no-one else can be looking at the page contents yet, there is
1070 : * no difference between an exclusive lock and a cleanup-strength
1071 : * lock. (Note that we cannot use LockBuffer() or
1072 : * LockBufferForCleanup() here, because they assert that the buffer is
1073 : * already valid.)
1074 : */
1075 529716 : if (!isLocalBuf)
1076 529668 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_EXCLUSIVE);
1077 :
1078 : /* Set BM_VALID, terminate IO, and wake up any waiters */
1079 529716 : if (isLocalBuf)
1080 48 : TerminateLocalBufferIO(bufHdr, false, BM_VALID, false);
1081 : else
1082 529668 : TerminateBufferIO(bufHdr, false, BM_VALID, true, false);
1083 : }
1084 74158 : else if (!isLocalBuf)
1085 : {
1086 : /*
1087 : * The buffer is valid, so we can't zero it. The caller still expects
1088 : * the page to be locked on return.
1089 : */
1090 74118 : if (mode == RBM_ZERO_AND_LOCK)
1091 74036 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1092 : else
1093 82 : LockBufferForCleanup(buffer);
1094 : }
1095 603874 : }
1096 :
1097 : /*
1098 : * Pin a buffer for a given block. *foundPtr is set to true if the block was
1099 : * already present, or false if more work is required to either read it in or
1100 : * zero it.
1101 : */
1102 : static pg_attribute_always_inline Buffer
1103 118651076 : PinBufferForBlock(Relation rel,
1104 : SMgrRelation smgr,
1105 : char persistence,
1106 : ForkNumber forkNum,
1107 : BlockNumber blockNum,
1108 : BufferAccessStrategy strategy,
1109 : bool *foundPtr)
1110 : {
1111 : BufferDesc *bufHdr;
1112 : IOContext io_context;
1113 : IOObject io_object;
1114 :
1115 : Assert(blockNum != P_NEW);
1116 :
1117 : /* Persistence should be set before */
1118 : Assert((persistence == RELPERSISTENCE_TEMP ||
1119 : persistence == RELPERSISTENCE_PERMANENT ||
1120 : persistence == RELPERSISTENCE_UNLOGGED));
1121 :
1122 118651076 : if (persistence == RELPERSISTENCE_TEMP)
1123 : {
1124 2545690 : io_context = IOCONTEXT_NORMAL;
1125 2545690 : io_object = IOOBJECT_TEMP_RELATION;
1126 : }
1127 : else
1128 : {
1129 116105386 : io_context = IOContextForStrategy(strategy);
1130 116105386 : io_object = IOOBJECT_RELATION;
1131 : }
1132 :
1133 : TRACE_POSTGRESQL_BUFFER_READ_START(forkNum, blockNum,
1134 : smgr->smgr_rlocator.locator.spcOid,
1135 : smgr->smgr_rlocator.locator.dbOid,
1136 : smgr->smgr_rlocator.locator.relNumber,
1137 : smgr->smgr_rlocator.backend);
1138 :
1139 118651076 : if (persistence == RELPERSISTENCE_TEMP)
1140 : {
1141 2545690 : bufHdr = LocalBufferAlloc(smgr, forkNum, blockNum, foundPtr);
1142 2545678 : if (*foundPtr)
1143 2528936 : pgBufferUsage.local_blks_hit++;
1144 : }
1145 : else
1146 : {
1147 116105386 : bufHdr = BufferAlloc(smgr, persistence, forkNum, blockNum,
1148 : strategy, foundPtr, io_context);
1149 116105386 : if (*foundPtr)
1150 112849726 : pgBufferUsage.shared_blks_hit++;
1151 : }
1152 118651064 : if (rel)
1153 : {
1154 : /*
1155 : * While pgBufferUsage's "read" counter isn't bumped unless we reach
1156 : * WaitReadBuffers() (so, not for hits, and not for buffers that are
1157 : * zeroed instead), the per-relation stats always count them.
1158 : */
1159 106938064 : pgstat_count_buffer_read(rel);
1160 106938064 : if (*foundPtr)
1161 104546900 : pgstat_count_buffer_hit(rel);
1162 : }
1163 118651064 : if (*foundPtr)
1164 : {
1165 115378662 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1166 115378662 : if (VacuumCostActive)
1167 3451198 : VacuumCostBalance += VacuumCostPageHit;
1168 :
1169 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forkNum, blockNum,
1170 : smgr->smgr_rlocator.locator.spcOid,
1171 : smgr->smgr_rlocator.locator.dbOid,
1172 : smgr->smgr_rlocator.locator.relNumber,
1173 : smgr->smgr_rlocator.backend,
1174 : true);
1175 : }
1176 :
1177 118651064 : return BufferDescriptorGetBuffer(bufHdr);
1178 : }
1179 :
1180 : /*
1181 : * ReadBuffer_common -- common logic for all ReadBuffer variants
1182 : *
1183 : * smgr is required, rel is optional unless using P_NEW.
1184 : */
1185 : static pg_attribute_always_inline Buffer
1186 111147362 : ReadBuffer_common(Relation rel, SMgrRelation smgr, char smgr_persistence,
1187 : ForkNumber forkNum,
1188 : BlockNumber blockNum, ReadBufferMode mode,
1189 : BufferAccessStrategy strategy)
1190 : {
1191 : ReadBuffersOperation operation;
1192 : Buffer buffer;
1193 : int flags;
1194 : char persistence;
1195 :
1196 : /*
1197 : * Backward compatibility path, most code should use ExtendBufferedRel()
1198 : * instead, as acquiring the extension lock inside ExtendBufferedRel()
1199 : * scales a lot better.
1200 : */
1201 111147362 : if (unlikely(blockNum == P_NEW))
1202 : {
1203 574 : uint32 flags = EB_SKIP_EXTENSION_LOCK;
1204 :
1205 : /*
1206 : * Since no-one else can be looking at the page contents yet, there is
1207 : * no difference between an exclusive lock and a cleanup-strength
1208 : * lock.
1209 : */
1210 574 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1211 0 : flags |= EB_LOCK_FIRST;
1212 :
1213 574 : return ExtendBufferedRel(BMR_REL(rel), forkNum, strategy, flags);
1214 : }
1215 :
1216 111146788 : if (rel)
1217 99875820 : persistence = rel->rd_rel->relpersistence;
1218 : else
1219 11270968 : persistence = smgr_persistence;
1220 :
1221 111146788 : if (unlikely(mode == RBM_ZERO_AND_CLEANUP_LOCK ||
1222 : mode == RBM_ZERO_AND_LOCK))
1223 : {
1224 : bool found;
1225 :
1226 603874 : buffer = PinBufferForBlock(rel, smgr, persistence,
1227 : forkNum, blockNum, strategy, &found);
1228 603874 : ZeroAndLockBuffer(buffer, mode, found);
1229 603874 : return buffer;
1230 : }
1231 :
1232 : /*
1233 : * Signal that we are going to immediately wait. If we're immediately
1234 : * waiting, there is no benefit in actually executing the IO
1235 : * asynchronously, it would just add dispatch overhead.
1236 : */
1237 110542914 : flags = READ_BUFFERS_SYNCHRONOUSLY;
1238 110542914 : if (mode == RBM_ZERO_ON_ERROR)
1239 2392320 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1240 110542914 : operation.smgr = smgr;
1241 110542914 : operation.rel = rel;
1242 110542914 : operation.persistence = persistence;
1243 110542914 : operation.forknum = forkNum;
1244 110542914 : operation.strategy = strategy;
1245 110542914 : if (StartReadBuffer(&operation,
1246 : &buffer,
1247 : blockNum,
1248 : flags))
1249 1259824 : WaitReadBuffers(&operation);
1250 :
1251 110542868 : return buffer;
1252 : }
1253 :
1254 : static pg_attribute_always_inline bool
1255 117719194 : StartReadBuffersImpl(ReadBuffersOperation *operation,
1256 : Buffer *buffers,
1257 : BlockNumber blockNum,
1258 : int *nblocks,
1259 : int flags,
1260 : bool allow_forwarding)
1261 : {
1262 117719194 : int actual_nblocks = *nblocks;
1263 117719194 : int maxcombine = 0;
1264 : bool did_start_io;
1265 :
1266 : Assert(*nblocks == 1 || allow_forwarding);
1267 : Assert(*nblocks > 0);
1268 : Assert(*nblocks <= MAX_IO_COMBINE_LIMIT);
1269 :
1270 120461880 : for (int i = 0; i < actual_nblocks; ++i)
1271 : {
1272 : bool found;
1273 :
1274 118050450 : if (allow_forwarding && buffers[i] != InvalidBuffer)
1275 3248 : {
1276 : BufferDesc *bufHdr;
1277 :
1278 : /*
1279 : * This is a buffer that was pinned by an earlier call to
1280 : * StartReadBuffers(), but couldn't be handled in one operation at
1281 : * that time. The operation was split, and the caller has passed
1282 : * an already pinned buffer back to us to handle the rest of the
1283 : * operation. It must continue at the expected block number.
1284 : */
1285 : Assert(BufferGetBlockNumber(buffers[i]) == blockNum + i);
1286 :
1287 : /*
1288 : * It might be an already valid buffer (a hit) that followed the
1289 : * final contiguous block of an earlier I/O (a miss) marking the
1290 : * end of it, or a buffer that some other backend has since made
1291 : * valid by performing the I/O for us, in which case we can handle
1292 : * it as a hit now. It is safe to check for a BM_VALID flag with
1293 : * a relaxed load, because we got a fresh view of it while pinning
1294 : * it in the previous call.
1295 : *
1296 : * On the other hand if we don't see BM_VALID yet, it must be an
1297 : * I/O that was split by the previous call and we need to try to
1298 : * start a new I/O from this block. We're also racing against any
1299 : * other backend that might start the I/O or even manage to mark
1300 : * it BM_VALID after this check, but StartBufferIO() will handle
1301 : * those cases.
1302 : */
1303 3248 : if (BufferIsLocal(buffers[i]))
1304 4 : bufHdr = GetLocalBufferDescriptor(-buffers[i] - 1);
1305 : else
1306 3244 : bufHdr = GetBufferDescriptor(buffers[i] - 1);
1307 : Assert(pg_atomic_read_u32(&bufHdr->state) & BM_TAG_VALID);
1308 3248 : found = pg_atomic_read_u32(&bufHdr->state) & BM_VALID;
1309 : }
1310 : else
1311 : {
1312 118047190 : buffers[i] = PinBufferForBlock(operation->rel,
1313 118047202 : operation->smgr,
1314 118047202 : operation->persistence,
1315 : operation->forknum,
1316 : blockNum + i,
1317 : operation->strategy,
1318 : &found);
1319 : }
1320 :
1321 118050438 : if (found)
1322 : {
1323 : /*
1324 : * We have a hit. If it's the first block in the requested range,
1325 : * we can return it immediately and report that WaitReadBuffers()
1326 : * does not need to be called. If the initial value of *nblocks
1327 : * was larger, the caller will have to call again for the rest.
1328 : */
1329 115307752 : if (i == 0)
1330 : {
1331 115304500 : *nblocks = 1;
1332 :
1333 : #ifdef USE_ASSERT_CHECKING
1334 :
1335 : /*
1336 : * Initialize enough of ReadBuffersOperation to make
1337 : * CheckReadBuffersOperation() work. Outside of assertions
1338 : * that's not necessary when no IO is issued.
1339 : */
1340 : operation->buffers = buffers;
1341 : operation->blocknum = blockNum;
1342 : operation->nblocks = 1;
1343 : operation->nblocks_done = 1;
1344 : CheckReadBuffersOperation(operation, true);
1345 : #endif
1346 115304500 : return false;
1347 : }
1348 :
1349 : /*
1350 : * Otherwise we already have an I/O to perform, but this block
1351 : * can't be included as it is already valid. Split the I/O here.
1352 : * There may or may not be more blocks requiring I/O after this
1353 : * one, we haven't checked, but they can't be contiguous with this
1354 : * one in the way. We'll leave this buffer pinned, forwarding it
1355 : * to the next call, avoiding the need to unpin it here and re-pin
1356 : * it in the next call.
1357 : */
1358 3252 : actual_nblocks = i;
1359 3252 : break;
1360 : }
1361 : else
1362 : {
1363 : /*
1364 : * Check how many blocks we can cover with the same IO. The smgr
1365 : * implementation might e.g. be limited due to a segment boundary.
1366 : */
1367 2742686 : if (i == 0 && actual_nblocks > 1)
1368 : {
1369 66232 : maxcombine = smgrmaxcombine(operation->smgr,
1370 : operation->forknum,
1371 : blockNum);
1372 66232 : if (unlikely(maxcombine < actual_nblocks))
1373 : {
1374 0 : elog(DEBUG2, "limiting nblocks at %u from %u to %u",
1375 : blockNum, actual_nblocks, maxcombine);
1376 0 : actual_nblocks = maxcombine;
1377 : }
1378 : }
1379 : }
1380 : }
1381 2414682 : *nblocks = actual_nblocks;
1382 :
1383 : /* Populate information needed for I/O. */
1384 2414682 : operation->buffers = buffers;
1385 2414682 : operation->blocknum = blockNum;
1386 2414682 : operation->flags = flags;
1387 2414682 : operation->nblocks = actual_nblocks;
1388 2414682 : operation->nblocks_done = 0;
1389 2414682 : pgaio_wref_clear(&operation->io_wref);
1390 :
1391 : /*
1392 : * When using AIO, start the IO in the background. If not, issue prefetch
1393 : * requests if desired by the caller.
1394 : *
1395 : * The reason we have a dedicated path for IOMETHOD_SYNC here is to
1396 : * de-risk the introduction of AIO somewhat. It's a large architectural
1397 : * change, with lots of chances for unanticipated performance effects.
1398 : *
1399 : * Use of IOMETHOD_SYNC already leads to not actually performing IO
1400 : * asynchronously, but without the check here we'd execute IO earlier than
1401 : * we used to. Eventually this IOMETHOD_SYNC specific path should go away.
1402 : */
1403 2414682 : if (io_method != IOMETHOD_SYNC)
1404 : {
1405 : /*
1406 : * Try to start IO asynchronously. It's possible that no IO needs to
1407 : * be started, if another backend already performed the IO.
1408 : *
1409 : * Note that if an IO is started, it might not cover the entire
1410 : * requested range, e.g. because an intermediary block has been read
1411 : * in by another backend. In that case any "trailing" buffers we
1412 : * already pinned above will be "forwarded" by read_stream.c to the
1413 : * next call to StartReadBuffers().
1414 : *
1415 : * This is signalled to the caller by decrementing *nblocks *and*
1416 : * reducing operation->nblocks. The latter is done here, but not below
1417 : * WaitReadBuffers(), as in WaitReadBuffers() we can't "shorten" the
1418 : * overall read size anymore, we need to retry until done in its
1419 : * entirety or until failed.
1420 : */
1421 2412514 : did_start_io = AsyncReadBuffers(operation, nblocks);
1422 :
1423 2412484 : operation->nblocks = *nblocks;
1424 : }
1425 : else
1426 : {
1427 2168 : operation->flags |= READ_BUFFERS_SYNCHRONOUSLY;
1428 :
1429 2168 : if (flags & READ_BUFFERS_ISSUE_ADVICE)
1430 : {
1431 : /*
1432 : * In theory we should only do this if PinBufferForBlock() had to
1433 : * allocate new buffers above. That way, if two calls to
1434 : * StartReadBuffers() were made for the same blocks before
1435 : * WaitReadBuffers(), only the first would issue the advice.
1436 : * That'd be a better simulation of true asynchronous I/O, which
1437 : * would only start the I/O once, but isn't done here for
1438 : * simplicity.
1439 : */
1440 4 : smgrprefetch(operation->smgr,
1441 : operation->forknum,
1442 : blockNum,
1443 : actual_nblocks);
1444 : }
1445 :
1446 : /*
1447 : * Indicate that WaitReadBuffers() should be called. WaitReadBuffers()
1448 : * will initiate the necessary IO.
1449 : */
1450 2168 : did_start_io = true;
1451 : }
1452 :
1453 2414652 : CheckReadBuffersOperation(operation, !did_start_io);
1454 :
1455 2414652 : return did_start_io;
1456 : }
1457 :
1458 : /*
1459 : * Begin reading a range of blocks beginning at blockNum and extending for
1460 : * *nblocks. *nblocks and the buffers array are in/out parameters. On entry,
1461 : * the buffers elements covered by *nblocks must hold either InvalidBuffer or
1462 : * buffers forwarded by an earlier call to StartReadBuffers() that was split
1463 : * and is now being continued. On return, *nblocks holds the number of blocks
1464 : * accepted by this operation. If it is less than the original number then
1465 : * this operation has been split, but buffer elements up to the original
1466 : * requested size may hold forwarded buffers to be used for a continuing
1467 : * operation. The caller must either start a new I/O beginning at the block
1468 : * immediately following the blocks accepted by this call and pass those
1469 : * buffers back in, or release them if it chooses not to. It shouldn't make
1470 : * any other use of or assumptions about forwarded buffers.
1471 : *
1472 : * If false is returned, no I/O is necessary and the buffers covered by
1473 : * *nblocks on exit are valid and ready to be accessed. If true is returned,
1474 : * an I/O has been started, and WaitReadBuffers() must be called with the same
1475 : * operation object before the buffers covered by *nblocks on exit can be
1476 : * accessed. Along with the operation object, the caller-supplied array of
1477 : * buffers must remain valid until WaitReadBuffers() is called, and any
1478 : * forwarded buffers must also be preserved for a continuing call unless
1479 : * they are explicitly released.
1480 : *
1481 : * Currently the I/O is only started with optional operating system advice if
1482 : * requested by the caller with READ_BUFFERS_ISSUE_ADVICE, and the real I/O
1483 : * happens synchronously in WaitReadBuffers(). In future work, true I/O could
1484 : * be initiated here.
1485 : */
1486 : bool
1487 3559028 : StartReadBuffers(ReadBuffersOperation *operation,
1488 : Buffer *buffers,
1489 : BlockNumber blockNum,
1490 : int *nblocks,
1491 : int flags)
1492 : {
1493 3559028 : return StartReadBuffersImpl(operation, buffers, blockNum, nblocks, flags,
1494 : true /* expect forwarded buffers */ );
1495 : }
1496 :
1497 : /*
1498 : * Single block version of the StartReadBuffers(). This might save a few
1499 : * instructions when called from another translation unit, because it is
1500 : * specialized for nblocks == 1.
1501 : *
1502 : * This version does not support "forwarded" buffers: they cannot be created
1503 : * by reading only one block and *buffer is ignored on entry.
1504 : */
1505 : bool
1506 114160166 : StartReadBuffer(ReadBuffersOperation *operation,
1507 : Buffer *buffer,
1508 : BlockNumber blocknum,
1509 : int flags)
1510 : {
1511 114160166 : int nblocks = 1;
1512 : bool result;
1513 :
1514 114160166 : result = StartReadBuffersImpl(operation, buffer, blocknum, &nblocks, flags,
1515 : false /* single block, no forwarding */ );
1516 : Assert(nblocks == 1); /* single block can't be short */
1517 :
1518 114160136 : return result;
1519 : }
1520 :
1521 : /*
1522 : * Perform sanity checks on the ReadBuffersOperation.
1523 : */
1524 : static void
1525 7234934 : CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete)
1526 : {
1527 : #ifdef USE_ASSERT_CHECKING
1528 : Assert(operation->nblocks_done <= operation->nblocks);
1529 : Assert(!is_complete || operation->nblocks == operation->nblocks_done);
1530 :
1531 : for (int i = 0; i < operation->nblocks; i++)
1532 : {
1533 : Buffer buffer = operation->buffers[i];
1534 : BufferDesc *buf_hdr = BufferIsLocal(buffer) ?
1535 : GetLocalBufferDescriptor(-buffer - 1) :
1536 : GetBufferDescriptor(buffer - 1);
1537 :
1538 : Assert(BufferGetBlockNumber(buffer) == operation->blocknum + i);
1539 : Assert(pg_atomic_read_u32(&buf_hdr->state) & BM_TAG_VALID);
1540 :
1541 : if (i < operation->nblocks_done)
1542 : Assert(pg_atomic_read_u32(&buf_hdr->state) & BM_VALID);
1543 : }
1544 : #endif
1545 7234934 : }
1546 :
1547 : /* helper for ReadBuffersCanStartIO(), to avoid repetition */
1548 : static inline bool
1549 2742722 : ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
1550 : {
1551 2742722 : if (BufferIsLocal(buffer))
1552 16694 : return StartLocalBufferIO(GetLocalBufferDescriptor(-buffer - 1),
1553 : true, nowait);
1554 : else
1555 2726028 : return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
1556 : }
1557 :
1558 : /*
1559 : * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
1560 : */
1561 : static inline bool
1562 2742722 : ReadBuffersCanStartIO(Buffer buffer, bool nowait)
1563 : {
1564 : /*
1565 : * If this backend currently has staged IO, we need to submit the pending
1566 : * IO before waiting for the right to issue IO, to avoid the potential for
1567 : * deadlocks (and, more commonly, unnecessary delays for other backends).
1568 : */
1569 2742722 : if (!nowait && pgaio_have_staged())
1570 : {
1571 754 : if (ReadBuffersCanStartIOOnce(buffer, true))
1572 754 : return true;
1573 :
1574 : /*
1575 : * Unfortunately StartBufferIO() returning false doesn't allow to
1576 : * distinguish between the buffer already being valid and IO already
1577 : * being in progress. Since IO already being in progress is quite
1578 : * rare, this approach seems fine.
1579 : */
1580 0 : pgaio_submit_staged();
1581 : }
1582 :
1583 2741968 : return ReadBuffersCanStartIOOnce(buffer, nowait);
1584 : }
1585 :
1586 : /*
1587 : * Helper for WaitReadBuffers() that processes the results of a readv
1588 : * operation, raising an error if necessary.
1589 : */
1590 : static void
1591 2409096 : ProcessReadBuffersResult(ReadBuffersOperation *operation)
1592 : {
1593 2409096 : PgAioReturn *aio_ret = &operation->io_return;
1594 2409096 : PgAioResultStatus rs = aio_ret->result.status;
1595 2409096 : int newly_read_blocks = 0;
1596 :
1597 : Assert(pgaio_wref_valid(&operation->io_wref));
1598 : Assert(aio_ret->result.status != PGAIO_RS_UNKNOWN);
1599 :
1600 : /*
1601 : * SMGR reports the number of blocks successfully read as the result of
1602 : * the IO operation. Thus we can simply add that to ->nblocks_done.
1603 : */
1604 :
1605 2409096 : if (likely(rs != PGAIO_RS_ERROR))
1606 2409038 : newly_read_blocks = aio_ret->result.result;
1607 :
1608 2409096 : if (rs == PGAIO_RS_ERROR || rs == PGAIO_RS_WARNING)
1609 90 : pgaio_result_report(aio_ret->result, &aio_ret->target_data,
1610 : rs == PGAIO_RS_ERROR ? ERROR : WARNING);
1611 2409006 : else if (aio_ret->result.status == PGAIO_RS_PARTIAL)
1612 : {
1613 : /*
1614 : * We'll retry, so we just emit a debug message to the server log (or
1615 : * not even that in prod scenarios).
1616 : */
1617 20 : pgaio_result_report(aio_ret->result, &aio_ret->target_data, DEBUG1);
1618 20 : elog(DEBUG3, "partial read, will retry");
1619 : }
1620 :
1621 : Assert(newly_read_blocks > 0);
1622 : Assert(newly_read_blocks <= MAX_IO_COMBINE_LIMIT);
1623 :
1624 2409038 : operation->nblocks_done += newly_read_blocks;
1625 :
1626 : Assert(operation->nblocks_done <= operation->nblocks);
1627 2409038 : }
1628 :
1629 : void
1630 2409076 : WaitReadBuffers(ReadBuffersOperation *operation)
1631 : {
1632 2409076 : PgAioReturn *aio_ret = &operation->io_return;
1633 : IOContext io_context;
1634 : IOObject io_object;
1635 :
1636 2409076 : if (operation->persistence == RELPERSISTENCE_TEMP)
1637 : {
1638 2940 : io_context = IOCONTEXT_NORMAL;
1639 2940 : io_object = IOOBJECT_TEMP_RELATION;
1640 : }
1641 : else
1642 : {
1643 2406136 : io_context = IOContextForStrategy(operation->strategy);
1644 2406136 : io_object = IOOBJECT_RELATION;
1645 : }
1646 :
1647 : /*
1648 : * If we get here without an IO operation having been issued, the
1649 : * io_method == IOMETHOD_SYNC path must have been used. Otherwise the
1650 : * caller should not have called WaitReadBuffers().
1651 : *
1652 : * In the case of IOMETHOD_SYNC, we start - as we used to before the
1653 : * introducing of AIO - the IO in WaitReadBuffers(). This is done as part
1654 : * of the retry logic below, no extra code is required.
1655 : *
1656 : * This path is expected to eventually go away.
1657 : */
1658 2409076 : if (!pgaio_wref_valid(&operation->io_wref) && io_method != IOMETHOD_SYNC)
1659 0 : elog(ERROR, "waiting for read operation that didn't read");
1660 :
1661 : /*
1662 : * To handle partial reads, and IOMETHOD_SYNC, we re-issue IO until we're
1663 : * done. We may need multiple retries, not just because we could get
1664 : * multiple partial reads, but also because some of the remaining
1665 : * to-be-read buffers may have been read in by other backends, limiting
1666 : * the IO size.
1667 : */
1668 : while (true)
1669 2188 : {
1670 : int ignored_nblocks_progress;
1671 :
1672 2411264 : CheckReadBuffersOperation(operation, false);
1673 :
1674 : /*
1675 : * If there is an IO associated with the operation, we may need to
1676 : * wait for it.
1677 : */
1678 2411264 : if (pgaio_wref_valid(&operation->io_wref))
1679 : {
1680 : /*
1681 : * Track the time spent waiting for the IO to complete. As
1682 : * tracking a wait even if we don't actually need to wait
1683 : *
1684 : * a) is not cheap, due to the timestamping overhead
1685 : *
1686 : * b) reports some time as waiting, even if we never waited
1687 : *
1688 : * we first check if we already know the IO is complete.
1689 : */
1690 2409096 : if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
1691 1135672 : !pgaio_wref_check_done(&operation->io_wref))
1692 : {
1693 485164 : instr_time io_start = pgstat_prepare_io_time(track_io_timing);
1694 :
1695 485164 : pgaio_wref_wait(&operation->io_wref);
1696 :
1697 : /*
1698 : * The IO operation itself was already counted earlier, in
1699 : * AsyncReadBuffers(), this just accounts for the wait time.
1700 : */
1701 485164 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1702 : io_start, 0, 0);
1703 : }
1704 : else
1705 : {
1706 : Assert(pgaio_wref_check_done(&operation->io_wref));
1707 : }
1708 :
1709 : /*
1710 : * We now are sure the IO completed. Check the results. This
1711 : * includes reporting on errors if there were any.
1712 : */
1713 2409096 : ProcessReadBuffersResult(operation);
1714 : }
1715 :
1716 : /*
1717 : * Most of the time, the one IO we already started, will read in
1718 : * everything. But we need to deal with partial reads and buffers not
1719 : * needing IO anymore.
1720 : */
1721 2411206 : if (operation->nblocks_done == operation->nblocks)
1722 2409018 : break;
1723 :
1724 2188 : CHECK_FOR_INTERRUPTS();
1725 :
1726 : /*
1727 : * This may only complete the IO partially, either because some
1728 : * buffers were already valid, or because of a partial read.
1729 : *
1730 : * NB: In contrast to after the AsyncReadBuffers() call in
1731 : * StartReadBuffers(), we do *not* reduce
1732 : * ReadBuffersOperation->nblocks here, callers expect the full
1733 : * operation to be completed at this point (as more operations may
1734 : * have been queued).
1735 : */
1736 2188 : AsyncReadBuffers(operation, &ignored_nblocks_progress);
1737 : }
1738 :
1739 2409018 : CheckReadBuffersOperation(operation, true);
1740 :
1741 : /* NB: READ_DONE tracepoint was already executed in completion callback */
1742 2409018 : }
1743 :
1744 : /*
1745 : * Initiate IO for the ReadBuffersOperation
1746 : *
1747 : * This function only starts a single IO at a time. The size of the IO may be
1748 : * limited to below the to-be-read blocks, if one of the buffers has
1749 : * concurrently been read in. If the first to-be-read buffer is already valid,
1750 : * no IO will be issued.
1751 : *
1752 : * To support retries after partial reads, the first operation->nblocks_done
1753 : * buffers are skipped.
1754 : *
1755 : * On return *nblocks_progress is updated to reflect the number of buffers
1756 : * affected by the call. If the first buffer is valid, *nblocks_progress is
1757 : * set to 1 and operation->nblocks_done is incremented.
1758 : *
1759 : * Returns true if IO was initiated, false if no IO was necessary.
1760 : */
1761 : static bool
1762 2414702 : AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
1763 : {
1764 2414702 : Buffer *buffers = &operation->buffers[0];
1765 2414702 : int flags = operation->flags;
1766 2414702 : BlockNumber blocknum = operation->blocknum;
1767 2414702 : ForkNumber forknum = operation->forknum;
1768 2414702 : char persistence = operation->persistence;
1769 2414702 : int16 nblocks_done = operation->nblocks_done;
1770 2414702 : Buffer *io_buffers = &operation->buffers[nblocks_done];
1771 2414702 : int io_buffers_len = 0;
1772 : PgAioHandle *ioh;
1773 2414702 : uint32 ioh_flags = 0;
1774 : void *io_pages[MAX_IO_COMBINE_LIMIT];
1775 : IOContext io_context;
1776 : IOObject io_object;
1777 : bool did_start_io;
1778 :
1779 : /*
1780 : * When this IO is executed synchronously, either because the caller will
1781 : * immediately block waiting for the IO or because IOMETHOD_SYNC is used,
1782 : * the AIO subsystem needs to know.
1783 : */
1784 2414702 : if (flags & READ_BUFFERS_SYNCHRONOUSLY)
1785 1262146 : ioh_flags |= PGAIO_HF_SYNCHRONOUS;
1786 :
1787 2414702 : if (persistence == RELPERSISTENCE_TEMP)
1788 : {
1789 3528 : io_context = IOCONTEXT_NORMAL;
1790 3528 : io_object = IOOBJECT_TEMP_RELATION;
1791 3528 : ioh_flags |= PGAIO_HF_REFERENCES_LOCAL;
1792 : }
1793 : else
1794 : {
1795 2411174 : io_context = IOContextForStrategy(operation->strategy);
1796 2411174 : io_object = IOOBJECT_RELATION;
1797 : }
1798 :
1799 : /*
1800 : * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
1801 : * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
1802 : * set globally, but on a per-session basis. The completion callback,
1803 : * which may be run in other processes, e.g. in IO workers, may have a
1804 : * different value of the zero_damaged_pages GUC.
1805 : *
1806 : * XXX: We probably should eventually use a different flag for
1807 : * zero_damaged_pages, so we can report different log levels / error codes
1808 : * for zero_damaged_pages and ZERO_ON_ERROR.
1809 : */
1810 2414702 : if (zero_damaged_pages)
1811 32 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1812 :
1813 : /*
1814 : * For the same reason as with zero_damaged_pages we need to use this
1815 : * backend's ignore_checksum_failure value.
1816 : */
1817 2414702 : if (ignore_checksum_failure)
1818 16 : flags |= READ_BUFFERS_IGNORE_CHECKSUM_FAILURES;
1819 :
1820 :
1821 : /*
1822 : * To be allowed to report stats in the local completion callback we need
1823 : * to prepare to report stats now. This ensures we can safely report the
1824 : * checksum failure even in a critical section.
1825 : */
1826 2414702 : pgstat_prepare_report_checksum_failure(operation->smgr->smgr_rlocator.locator.dbOid);
1827 :
1828 : /*
1829 : * Get IO handle before ReadBuffersCanStartIO(), as pgaio_io_acquire()
1830 : * might block, which we don't want after setting IO_IN_PROGRESS.
1831 : *
1832 : * If we need to wait for IO before we can get a handle, submit
1833 : * already-staged IO first, so that other backends don't need to wait.
1834 : * There wouldn't be a deadlock risk, as pgaio_io_acquire() just needs to
1835 : * wait for already submitted IO, which doesn't require additional locks,
1836 : * but it could still cause undesirable waits.
1837 : *
1838 : * A secondary benefit is that this would allow us to measure the time in
1839 : * pgaio_io_acquire() without causing undue timer overhead in the common,
1840 : * non-blocking, case. However, currently the pgstats infrastructure
1841 : * doesn't really allow that, as it a) asserts that an operation can't
1842 : * have time without operations b) doesn't have an API to report
1843 : * "accumulated" time.
1844 : */
1845 2414702 : ioh = pgaio_io_acquire_nb(CurrentResourceOwner, &operation->io_return);
1846 2414702 : if (unlikely(!ioh))
1847 : {
1848 5408 : pgaio_submit_staged();
1849 :
1850 5408 : ioh = pgaio_io_acquire(CurrentResourceOwner, &operation->io_return);
1851 : }
1852 :
1853 : /*
1854 : * Check if we can start IO on the first to-be-read buffer.
1855 : *
1856 : * If an I/O is already in progress in another backend, we want to wait
1857 : * for the outcome: either done, or something went wrong and we will
1858 : * retry.
1859 : */
1860 2414702 : if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
1861 : {
1862 : /*
1863 : * Someone else has already completed this block, we're done.
1864 : *
1865 : * When IO is necessary, ->nblocks_done is updated in
1866 : * ProcessReadBuffersResult(), but that is not called if no IO is
1867 : * necessary. Thus update here.
1868 : */
1869 4984 : operation->nblocks_done += 1;
1870 4984 : *nblocks_progress = 1;
1871 :
1872 4984 : pgaio_io_release(ioh);
1873 4984 : pgaio_wref_clear(&operation->io_wref);
1874 4984 : did_start_io = false;
1875 :
1876 : /*
1877 : * Report and track this as a 'hit' for this backend, even though it
1878 : * must have started out as a miss in PinBufferForBlock(). The other
1879 : * backend will track this as a 'read'.
1880 : */
1881 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forknum, blocknum + operation->nblocks_done,
1882 : operation->smgr->smgr_rlocator.locator.spcOid,
1883 : operation->smgr->smgr_rlocator.locator.dbOid,
1884 : operation->smgr->smgr_rlocator.locator.relNumber,
1885 : operation->smgr->smgr_rlocator.backend,
1886 : true);
1887 :
1888 4984 : if (persistence == RELPERSISTENCE_TEMP)
1889 0 : pgBufferUsage.local_blks_hit += 1;
1890 : else
1891 4984 : pgBufferUsage.shared_blks_hit += 1;
1892 :
1893 4984 : if (operation->rel)
1894 4984 : pgstat_count_buffer_hit(operation->rel);
1895 :
1896 4984 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1897 :
1898 4984 : if (VacuumCostActive)
1899 4 : VacuumCostBalance += VacuumCostPageHit;
1900 : }
1901 : else
1902 : {
1903 : instr_time io_start;
1904 :
1905 : /* We found a buffer that we need to read in. */
1906 : Assert(io_buffers[0] == buffers[nblocks_done]);
1907 2409718 : io_pages[0] = BufferGetBlock(buffers[nblocks_done]);
1908 2409718 : io_buffers_len = 1;
1909 :
1910 : /*
1911 : * How many neighboring-on-disk blocks can we scatter-read into other
1912 : * buffers at the same time? In this case we don't wait if we see an
1913 : * I/O already in progress. We already set BM_IO_IN_PROGRESS for the
1914 : * head block, so we should get on with that I/O as soon as possible.
1915 : */
1916 2737738 : for (int i = nblocks_done + 1; i < operation->nblocks; i++)
1917 : {
1918 328020 : if (!ReadBuffersCanStartIO(buffers[i], true))
1919 0 : break;
1920 : /* Must be consecutive block numbers. */
1921 : Assert(BufferGetBlockNumber(buffers[i - 1]) ==
1922 : BufferGetBlockNumber(buffers[i]) - 1);
1923 : Assert(io_buffers[io_buffers_len] == buffers[i]);
1924 :
1925 328020 : io_pages[io_buffers_len++] = BufferGetBlock(buffers[i]);
1926 : }
1927 :
1928 : /* get a reference to wait for in WaitReadBuffers() */
1929 2409718 : pgaio_io_get_wref(ioh, &operation->io_wref);
1930 :
1931 : /* provide the list of buffers to the completion callbacks */
1932 2409718 : pgaio_io_set_handle_data_32(ioh, (uint32 *) io_buffers, io_buffers_len);
1933 :
1934 2409718 : pgaio_io_register_callbacks(ioh,
1935 : persistence == RELPERSISTENCE_TEMP ?
1936 : PGAIO_HCB_LOCAL_BUFFER_READV :
1937 : PGAIO_HCB_SHARED_BUFFER_READV,
1938 : flags);
1939 :
1940 2409718 : pgaio_io_set_flag(ioh, ioh_flags);
1941 :
1942 : /* ---
1943 : * Even though we're trying to issue IO asynchronously, track the time
1944 : * in smgrstartreadv():
1945 : * - if io_method == IOMETHOD_SYNC, we will always perform the IO
1946 : * immediately
1947 : * - the io method might not support the IO (e.g. worker IO for a temp
1948 : * table)
1949 : * ---
1950 : */
1951 2409718 : io_start = pgstat_prepare_io_time(track_io_timing);
1952 2409718 : smgrstartreadv(ioh, operation->smgr, forknum,
1953 : blocknum + nblocks_done,
1954 : io_pages, io_buffers_len);
1955 2409688 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1956 2409688 : io_start, 1, io_buffers_len * BLCKSZ);
1957 :
1958 2409688 : if (persistence == RELPERSISTENCE_TEMP)
1959 3528 : pgBufferUsage.local_blks_read += io_buffers_len;
1960 : else
1961 2406160 : pgBufferUsage.shared_blks_read += io_buffers_len;
1962 :
1963 : /*
1964 : * Track vacuum cost when issuing IO, not after waiting for it.
1965 : * Otherwise we could end up issuing a lot of IO in a short timespan,
1966 : * despite a low cost limit.
1967 : */
1968 2409688 : if (VacuumCostActive)
1969 35602 : VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
1970 :
1971 2409688 : *nblocks_progress = io_buffers_len;
1972 2409688 : did_start_io = true;
1973 : }
1974 :
1975 2414672 : return did_start_io;
1976 : }
1977 :
1978 : /*
1979 : * BufferAlloc -- subroutine for PinBufferForBlock. Handles lookup of a shared
1980 : * buffer. If no buffer exists already, selects a replacement victim and
1981 : * evicts the old page, but does NOT read in new page.
1982 : *
1983 : * "strategy" can be a buffer replacement strategy object, or NULL for
1984 : * the default strategy. The selected buffer's usage_count is advanced when
1985 : * using the default strategy, but otherwise possibly not (see PinBuffer).
1986 : *
1987 : * The returned buffer is pinned and is already marked as holding the
1988 : * desired page. If it already did have the desired page, *foundPtr is
1989 : * set true. Otherwise, *foundPtr is set false.
1990 : *
1991 : * io_context is passed as an output parameter to avoid calling
1992 : * IOContextForStrategy() when there is a shared buffers hit and no IO
1993 : * statistics need be captured.
1994 : *
1995 : * No locks are held either at entry or exit.
1996 : */
1997 : static pg_attribute_always_inline BufferDesc *
1998 116105386 : BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
1999 : BlockNumber blockNum,
2000 : BufferAccessStrategy strategy,
2001 : bool *foundPtr, IOContext io_context)
2002 : {
2003 : BufferTag newTag; /* identity of requested block */
2004 : uint32 newHash; /* hash value for newTag */
2005 : LWLock *newPartitionLock; /* buffer partition lock for it */
2006 : int existing_buf_id;
2007 : Buffer victim_buffer;
2008 : BufferDesc *victim_buf_hdr;
2009 : uint32 victim_buf_state;
2010 :
2011 : /* Make sure we will have room to remember the buffer pin */
2012 116105386 : ResourceOwnerEnlarge(CurrentResourceOwner);
2013 116105386 : ReservePrivateRefCountEntry();
2014 :
2015 : /* create a tag so we can lookup the buffer */
2016 116105386 : InitBufferTag(&newTag, &smgr->smgr_rlocator.locator, forkNum, blockNum);
2017 :
2018 : /* determine its hash code and partition lock ID */
2019 116105386 : newHash = BufTableHashCode(&newTag);
2020 116105386 : newPartitionLock = BufMappingPartitionLock(newHash);
2021 :
2022 : /* see if the block is in the buffer pool already */
2023 116105386 : LWLockAcquire(newPartitionLock, LW_SHARED);
2024 116105386 : existing_buf_id = BufTableLookup(&newTag, newHash);
2025 116105386 : if (existing_buf_id >= 0)
2026 : {
2027 : BufferDesc *buf;
2028 : bool valid;
2029 :
2030 : /*
2031 : * Found it. Now, pin the buffer so no one can steal it from the
2032 : * buffer pool, and check to see if the correct data has been loaded
2033 : * into the buffer.
2034 : */
2035 112853766 : buf = GetBufferDescriptor(existing_buf_id);
2036 :
2037 112853766 : valid = PinBuffer(buf, strategy);
2038 :
2039 : /* Can release the mapping lock as soon as we've pinned it */
2040 112853766 : LWLockRelease(newPartitionLock);
2041 :
2042 112853766 : *foundPtr = true;
2043 :
2044 112853766 : if (!valid)
2045 : {
2046 : /*
2047 : * We can only get here if (a) someone else is still reading in
2048 : * the page, (b) a previous read attempt failed, or (c) someone
2049 : * called StartReadBuffers() but not yet WaitReadBuffers().
2050 : */
2051 4362 : *foundPtr = false;
2052 : }
2053 :
2054 112853766 : return buf;
2055 : }
2056 :
2057 : /*
2058 : * Didn't find it in the buffer pool. We'll have to initialize a new
2059 : * buffer. Remember to unlock the mapping lock while doing the work.
2060 : */
2061 3251620 : LWLockRelease(newPartitionLock);
2062 :
2063 : /*
2064 : * Acquire a victim buffer. Somebody else might try to do the same, we
2065 : * don't hold any conflicting locks. If so we'll have to undo our work
2066 : * later.
2067 : */
2068 3251620 : victim_buffer = GetVictimBuffer(strategy, io_context);
2069 3251620 : victim_buf_hdr = GetBufferDescriptor(victim_buffer - 1);
2070 :
2071 : /*
2072 : * Try to make a hashtable entry for the buffer under its new tag. If
2073 : * somebody else inserted another buffer for the tag, we'll release the
2074 : * victim buffer we acquired and use the already inserted one.
2075 : */
2076 3251620 : LWLockAcquire(newPartitionLock, LW_EXCLUSIVE);
2077 3251620 : existing_buf_id = BufTableInsert(&newTag, newHash, victim_buf_hdr->buf_id);
2078 3251620 : if (existing_buf_id >= 0)
2079 : {
2080 : BufferDesc *existing_buf_hdr;
2081 : bool valid;
2082 :
2083 : /*
2084 : * Got a collision. Someone has already done what we were about to do.
2085 : * We'll just handle this as if it were found in the buffer pool in
2086 : * the first place. First, give up the buffer we were planning to
2087 : * use.
2088 : *
2089 : * We could do this after releasing the partition lock, but then we'd
2090 : * have to call ResourceOwnerEnlarge() & ReservePrivateRefCountEntry()
2091 : * before acquiring the lock, for the rare case of such a collision.
2092 : */
2093 1040 : UnpinBuffer(victim_buf_hdr);
2094 :
2095 : /*
2096 : * The victim buffer we acquired previously is clean and unused, let
2097 : * it be found again quickly
2098 : */
2099 1040 : StrategyFreeBuffer(victim_buf_hdr);
2100 :
2101 : /* remaining code should match code at top of routine */
2102 :
2103 1040 : existing_buf_hdr = GetBufferDescriptor(existing_buf_id);
2104 :
2105 1040 : valid = PinBuffer(existing_buf_hdr, strategy);
2106 :
2107 : /* Can release the mapping lock as soon as we've pinned it */
2108 1040 : LWLockRelease(newPartitionLock);
2109 :
2110 1040 : *foundPtr = true;
2111 :
2112 1040 : if (!valid)
2113 : {
2114 : /*
2115 : * We can only get here if (a) someone else is still reading in
2116 : * the page, (b) a previous read attempt failed, or (c) someone
2117 : * called StartReadBuffers() but not yet WaitReadBuffers().
2118 : */
2119 718 : *foundPtr = false;
2120 : }
2121 :
2122 1040 : return existing_buf_hdr;
2123 : }
2124 :
2125 : /*
2126 : * Need to lock the buffer header too in order to change its tag.
2127 : */
2128 3250580 : victim_buf_state = LockBufHdr(victim_buf_hdr);
2129 :
2130 : /* some sanity checks while we hold the buffer header lock */
2131 : Assert(BUF_STATE_GET_REFCOUNT(victim_buf_state) == 1);
2132 : Assert(!(victim_buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY | BM_IO_IN_PROGRESS)));
2133 :
2134 3250580 : victim_buf_hdr->tag = newTag;
2135 :
2136 : /*
2137 : * Make sure BM_PERMANENT is set for buffers that must be written at every
2138 : * checkpoint. Unlogged buffers only need to be written at shutdown
2139 : * checkpoints, except for their "init" forks, which need to be treated
2140 : * just like permanent relations.
2141 : */
2142 3250580 : victim_buf_state |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2143 3250580 : if (relpersistence == RELPERSISTENCE_PERMANENT || forkNum == INIT_FORKNUM)
2144 3249898 : victim_buf_state |= BM_PERMANENT;
2145 :
2146 3250580 : UnlockBufHdr(victim_buf_hdr, victim_buf_state);
2147 :
2148 3250580 : LWLockRelease(newPartitionLock);
2149 :
2150 : /*
2151 : * Buffer contents are currently invalid.
2152 : */
2153 3250580 : *foundPtr = false;
2154 :
2155 3250580 : return victim_buf_hdr;
2156 : }
2157 :
2158 : /*
2159 : * InvalidateBuffer -- mark a shared buffer invalid and return it to the
2160 : * freelist.
2161 : *
2162 : * The buffer header spinlock must be held at entry. We drop it before
2163 : * returning. (This is sane because the caller must have locked the
2164 : * buffer in order to be sure it should be dropped.)
2165 : *
2166 : * This is used only in contexts such as dropping a relation. We assume
2167 : * that no other backend could possibly be interested in using the page,
2168 : * so the only reason the buffer might be pinned is if someone else is
2169 : * trying to write it out. We have to let them finish before we can
2170 : * reclaim the buffer.
2171 : *
2172 : * The buffer could get reclaimed by someone else while we are waiting
2173 : * to acquire the necessary locks; if so, don't mess it up.
2174 : */
2175 : static void
2176 213316 : InvalidateBuffer(BufferDesc *buf)
2177 : {
2178 : BufferTag oldTag;
2179 : uint32 oldHash; /* hash value for oldTag */
2180 : LWLock *oldPartitionLock; /* buffer partition lock for it */
2181 : uint32 oldFlags;
2182 : uint32 buf_state;
2183 :
2184 : /* Save the original buffer tag before dropping the spinlock */
2185 213316 : oldTag = buf->tag;
2186 :
2187 213316 : buf_state = pg_atomic_read_u32(&buf->state);
2188 : Assert(buf_state & BM_LOCKED);
2189 213316 : UnlockBufHdr(buf, buf_state);
2190 :
2191 : /*
2192 : * Need to compute the old tag's hashcode and partition lock ID. XXX is it
2193 : * worth storing the hashcode in BufferDesc so we need not recompute it
2194 : * here? Probably not.
2195 : */
2196 213316 : oldHash = BufTableHashCode(&oldTag);
2197 213316 : oldPartitionLock = BufMappingPartitionLock(oldHash);
2198 :
2199 213316 : retry:
2200 :
2201 : /*
2202 : * Acquire exclusive mapping lock in preparation for changing the buffer's
2203 : * association.
2204 : */
2205 213316 : LWLockAcquire(oldPartitionLock, LW_EXCLUSIVE);
2206 :
2207 : /* Re-lock the buffer header */
2208 213316 : buf_state = LockBufHdr(buf);
2209 :
2210 : /* If it's changed while we were waiting for lock, do nothing */
2211 213316 : if (!BufferTagsEqual(&buf->tag, &oldTag))
2212 : {
2213 2 : UnlockBufHdr(buf, buf_state);
2214 2 : LWLockRelease(oldPartitionLock);
2215 2 : return;
2216 : }
2217 :
2218 : /*
2219 : * We assume the reason for it to be pinned is that either we were
2220 : * asynchronously reading the page in before erroring out or someone else
2221 : * is flushing the page out. Wait for the IO to finish. (This could be
2222 : * an infinite loop if the refcount is messed up... it would be nice to
2223 : * time out after awhile, but there seems no way to be sure how many loops
2224 : * may be needed. Note that if the other guy has pinned the buffer but
2225 : * not yet done StartBufferIO, WaitIO will fall through and we'll
2226 : * effectively be busy-looping here.)
2227 : */
2228 213314 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 0)
2229 : {
2230 0 : UnlockBufHdr(buf, buf_state);
2231 0 : LWLockRelease(oldPartitionLock);
2232 : /* safety check: should definitely not be our *own* pin */
2233 0 : if (GetPrivateRefCount(BufferDescriptorGetBuffer(buf)) > 0)
2234 0 : elog(ERROR, "buffer is pinned in InvalidateBuffer");
2235 0 : WaitIO(buf);
2236 0 : goto retry;
2237 : }
2238 :
2239 : /*
2240 : * Clear out the buffer's tag and flags. We must do this to ensure that
2241 : * linear scans of the buffer array don't think the buffer is valid.
2242 : */
2243 213314 : oldFlags = buf_state & BUF_FLAG_MASK;
2244 213314 : ClearBufferTag(&buf->tag);
2245 213314 : buf_state &= ~(BUF_FLAG_MASK | BUF_USAGECOUNT_MASK);
2246 213314 : UnlockBufHdr(buf, buf_state);
2247 :
2248 : /*
2249 : * Remove the buffer from the lookup hashtable, if it was in there.
2250 : */
2251 213314 : if (oldFlags & BM_TAG_VALID)
2252 213314 : BufTableDelete(&oldTag, oldHash);
2253 :
2254 : /*
2255 : * Done with mapping lock.
2256 : */
2257 213314 : LWLockRelease(oldPartitionLock);
2258 :
2259 : /*
2260 : * Insert the buffer at the head of the list of free buffers.
2261 : */
2262 213314 : StrategyFreeBuffer(buf);
2263 : }
2264 :
2265 : /*
2266 : * Helper routine for GetVictimBuffer()
2267 : *
2268 : * Needs to be called on a buffer with a valid tag, pinned, but without the
2269 : * buffer header spinlock held.
2270 : *
2271 : * Returns true if the buffer can be reused, in which case the buffer is only
2272 : * pinned by this backend and marked as invalid, false otherwise.
2273 : */
2274 : static bool
2275 2318072 : InvalidateVictimBuffer(BufferDesc *buf_hdr)
2276 : {
2277 : uint32 buf_state;
2278 : uint32 hash;
2279 : LWLock *partition_lock;
2280 : BufferTag tag;
2281 :
2282 : Assert(GetPrivateRefCount(BufferDescriptorGetBuffer(buf_hdr)) == 1);
2283 :
2284 : /* have buffer pinned, so it's safe to read tag without lock */
2285 2318072 : tag = buf_hdr->tag;
2286 :
2287 2318072 : hash = BufTableHashCode(&tag);
2288 2318072 : partition_lock = BufMappingPartitionLock(hash);
2289 :
2290 2318072 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2291 :
2292 : /* lock the buffer header */
2293 2318072 : buf_state = LockBufHdr(buf_hdr);
2294 :
2295 : /*
2296 : * We have the buffer pinned nobody else should have been able to unset
2297 : * this concurrently.
2298 : */
2299 : Assert(buf_state & BM_TAG_VALID);
2300 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2301 : Assert(BufferTagsEqual(&buf_hdr->tag, &tag));
2302 :
2303 : /*
2304 : * If somebody else pinned the buffer since, or even worse, dirtied it,
2305 : * give up on this buffer: It's clearly in use.
2306 : */
2307 2318072 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 1 || (buf_state & BM_DIRTY))
2308 : {
2309 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2310 :
2311 568 : UnlockBufHdr(buf_hdr, buf_state);
2312 568 : LWLockRelease(partition_lock);
2313 :
2314 568 : return false;
2315 : }
2316 :
2317 : /*
2318 : * Clear out the buffer's tag and flags and usagecount. This is not
2319 : * strictly required, as BM_TAG_VALID/BM_VALID needs to be checked before
2320 : * doing anything with the buffer. But currently it's beneficial, as the
2321 : * cheaper pre-check for several linear scans of shared buffers use the
2322 : * tag (see e.g. FlushDatabaseBuffers()).
2323 : */
2324 2317504 : ClearBufferTag(&buf_hdr->tag);
2325 2317504 : buf_state &= ~(BUF_FLAG_MASK | BUF_USAGECOUNT_MASK);
2326 2317504 : UnlockBufHdr(buf_hdr, buf_state);
2327 :
2328 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2329 :
2330 : /* finally delete buffer from the buffer mapping table */
2331 2317504 : BufTableDelete(&tag, hash);
2332 :
2333 2317504 : LWLockRelease(partition_lock);
2334 :
2335 : Assert(!(buf_state & (BM_DIRTY | BM_VALID | BM_TAG_VALID)));
2336 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2337 : Assert(BUF_STATE_GET_REFCOUNT(pg_atomic_read_u32(&buf_hdr->state)) > 0);
2338 :
2339 2317504 : return true;
2340 : }
2341 :
2342 : static Buffer
2343 3703812 : GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context)
2344 : {
2345 : BufferDesc *buf_hdr;
2346 : Buffer buf;
2347 : uint32 buf_state;
2348 : bool from_ring;
2349 :
2350 : /*
2351 : * Ensure, while the spinlock's not yet held, that there's a free refcount
2352 : * entry, and a resource owner slot for the pin.
2353 : */
2354 3703812 : ReservePrivateRefCountEntry();
2355 3703812 : ResourceOwnerEnlarge(CurrentResourceOwner);
2356 :
2357 : /* we return here if a prospective victim buffer gets used concurrently */
2358 3715934 : again:
2359 :
2360 : /*
2361 : * Select a victim buffer. The buffer is returned with its header
2362 : * spinlock still held!
2363 : */
2364 3715934 : buf_hdr = StrategyGetBuffer(strategy, &buf_state, &from_ring);
2365 3715934 : buf = BufferDescriptorGetBuffer(buf_hdr);
2366 :
2367 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 0);
2368 :
2369 : /* Pin the buffer and then release the buffer spinlock */
2370 3715934 : PinBuffer_Locked(buf_hdr);
2371 :
2372 : /*
2373 : * We shouldn't have any other pins for this buffer.
2374 : */
2375 3715934 : CheckBufferIsPinnedOnce(buf);
2376 :
2377 : /*
2378 : * If the buffer was dirty, try to write it out. There is a race
2379 : * condition here, in that someone might dirty it after we released the
2380 : * buffer header lock above, or even while we are writing it out (since
2381 : * our share-lock won't prevent hint-bit updates). We will recheck the
2382 : * dirty bit after re-locking the buffer header.
2383 : */
2384 3715934 : if (buf_state & BM_DIRTY)
2385 : {
2386 : LWLock *content_lock;
2387 :
2388 : Assert(buf_state & BM_TAG_VALID);
2389 : Assert(buf_state & BM_VALID);
2390 :
2391 : /*
2392 : * We need a share-lock on the buffer contents to write it out (else
2393 : * we might write invalid data, eg because someone else is compacting
2394 : * the page contents while we write). We must use a conditional lock
2395 : * acquisition here to avoid deadlock. Even though the buffer was not
2396 : * pinned (and therefore surely not locked) when StrategyGetBuffer
2397 : * returned it, someone else could have pinned and exclusive-locked it
2398 : * by the time we get here. If we try to get the lock unconditionally,
2399 : * we'd block waiting for them; if they later block waiting for us,
2400 : * deadlock ensues. (This has been observed to happen when two
2401 : * backends are both trying to split btree index pages, and the second
2402 : * one just happens to be trying to split the page the first one got
2403 : * from StrategyGetBuffer.)
2404 : */
2405 506638 : content_lock = BufferDescriptorGetContentLock(buf_hdr);
2406 506638 : if (!LWLockConditionalAcquire(content_lock, LW_SHARED))
2407 : {
2408 : /*
2409 : * Someone else has locked the buffer, so give it up and loop back
2410 : * to get another one.
2411 : */
2412 0 : UnpinBuffer(buf_hdr);
2413 0 : goto again;
2414 : }
2415 :
2416 : /*
2417 : * If using a nondefault strategy, and writing the buffer would
2418 : * require a WAL flush, let the strategy decide whether to go ahead
2419 : * and write/reuse the buffer or to choose another victim. We need a
2420 : * lock to inspect the page LSN, so this can't be done inside
2421 : * StrategyGetBuffer.
2422 : */
2423 506638 : if (strategy != NULL)
2424 : {
2425 : XLogRecPtr lsn;
2426 :
2427 : /* Read the LSN while holding buffer header lock */
2428 149680 : buf_state = LockBufHdr(buf_hdr);
2429 149680 : lsn = BufferGetLSN(buf_hdr);
2430 149680 : UnlockBufHdr(buf_hdr, buf_state);
2431 :
2432 149680 : if (XLogNeedsFlush(lsn)
2433 17434 : && StrategyRejectBuffer(strategy, buf_hdr, from_ring))
2434 : {
2435 11554 : LWLockRelease(content_lock);
2436 11554 : UnpinBuffer(buf_hdr);
2437 11554 : goto again;
2438 : }
2439 : }
2440 :
2441 : /* OK, do the I/O */
2442 495084 : FlushBuffer(buf_hdr, NULL, IOOBJECT_RELATION, io_context);
2443 495084 : LWLockRelease(content_lock);
2444 :
2445 495084 : ScheduleBufferTagForWriteback(&BackendWritebackContext, io_context,
2446 : &buf_hdr->tag);
2447 : }
2448 :
2449 :
2450 3704380 : if (buf_state & BM_VALID)
2451 : {
2452 : /*
2453 : * When a BufferAccessStrategy is in use, blocks evicted from shared
2454 : * buffers are counted as IOOP_EVICT in the corresponding context
2455 : * (e.g. IOCONTEXT_BULKWRITE). Shared buffers are evicted by a
2456 : * strategy in two cases: 1) while initially claiming buffers for the
2457 : * strategy ring 2) to replace an existing strategy ring buffer
2458 : * because it is pinned or in use and cannot be reused.
2459 : *
2460 : * Blocks evicted from buffers already in the strategy ring are
2461 : * counted as IOOP_REUSE in the corresponding strategy context.
2462 : *
2463 : * At this point, we can accurately count evictions and reuses,
2464 : * because we have successfully claimed the valid buffer. Previously,
2465 : * we may have been forced to release the buffer due to concurrent
2466 : * pinners or erroring out.
2467 : */
2468 2317794 : pgstat_count_io_op(IOOBJECT_RELATION, io_context,
2469 2317794 : from_ring ? IOOP_REUSE : IOOP_EVICT, 1, 0);
2470 : }
2471 :
2472 : /*
2473 : * If the buffer has an entry in the buffer mapping table, delete it. This
2474 : * can fail because another backend could have pinned or dirtied the
2475 : * buffer.
2476 : */
2477 3704380 : if ((buf_state & BM_TAG_VALID) && !InvalidateVictimBuffer(buf_hdr))
2478 : {
2479 568 : UnpinBuffer(buf_hdr);
2480 568 : goto again;
2481 : }
2482 :
2483 : /* a final set of sanity checks */
2484 : #ifdef USE_ASSERT_CHECKING
2485 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
2486 :
2487 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2488 : Assert(!(buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY)));
2489 :
2490 : CheckBufferIsPinnedOnce(buf);
2491 : #endif
2492 :
2493 3703812 : return buf;
2494 : }
2495 :
2496 : /*
2497 : * Return the maximum number of buffers that a backend should try to pin once,
2498 : * to avoid exceeding its fair share. This is the highest value that
2499 : * GetAdditionalPinLimit() could ever return. Note that it may be zero on a
2500 : * system with a very small buffer pool relative to max_connections.
2501 : */
2502 : uint32
2503 985178 : GetPinLimit(void)
2504 : {
2505 985178 : return MaxProportionalPins;
2506 : }
2507 :
2508 : /*
2509 : * Return the maximum number of additional buffers that this backend should
2510 : * pin if it wants to stay under the per-backend limit, considering the number
2511 : * of buffers it has already pinned. Unlike LimitAdditionalPins(), the limit
2512 : * return by this function can be zero.
2513 : */
2514 : uint32
2515 7095920 : GetAdditionalPinLimit(void)
2516 : {
2517 : uint32 estimated_pins_held;
2518 :
2519 : /*
2520 : * We get the number of "overflowed" pins for free, but don't know the
2521 : * number of pins in PrivateRefCountArray. The cost of calculating that
2522 : * exactly doesn't seem worth it, so just assume the max.
2523 : */
2524 7095920 : estimated_pins_held = PrivateRefCountOverflowed + REFCOUNT_ARRAY_ENTRIES;
2525 :
2526 : /* Is this backend already holding more than its fair share? */
2527 7095920 : if (estimated_pins_held > MaxProportionalPins)
2528 2449822 : return 0;
2529 :
2530 4646098 : return MaxProportionalPins - estimated_pins_held;
2531 : }
2532 :
2533 : /*
2534 : * Limit the number of pins a batch operation may additionally acquire, to
2535 : * avoid running out of pinnable buffers.
2536 : *
2537 : * One additional pin is always allowed, on the assumption that the operation
2538 : * requires at least one to make progress.
2539 : */
2540 : void
2541 400496 : LimitAdditionalPins(uint32 *additional_pins)
2542 : {
2543 : uint32 limit;
2544 :
2545 400496 : if (*additional_pins <= 1)
2546 380404 : return;
2547 :
2548 20092 : limit = GetAdditionalPinLimit();
2549 20092 : limit = Max(limit, 1);
2550 20092 : if (limit < *additional_pins)
2551 10902 : *additional_pins = limit;
2552 : }
2553 :
2554 : /*
2555 : * Logic shared between ExtendBufferedRelBy(), ExtendBufferedRelTo(). Just to
2556 : * avoid duplicating the tracing and relpersistence related logic.
2557 : */
2558 : static BlockNumber
2559 423200 : ExtendBufferedRelCommon(BufferManagerRelation bmr,
2560 : ForkNumber fork,
2561 : BufferAccessStrategy strategy,
2562 : uint32 flags,
2563 : uint32 extend_by,
2564 : BlockNumber extend_upto,
2565 : Buffer *buffers,
2566 : uint32 *extended_by)
2567 : {
2568 : BlockNumber first_block;
2569 :
2570 : TRACE_POSTGRESQL_BUFFER_EXTEND_START(fork,
2571 : bmr.smgr->smgr_rlocator.locator.spcOid,
2572 : bmr.smgr->smgr_rlocator.locator.dbOid,
2573 : bmr.smgr->smgr_rlocator.locator.relNumber,
2574 : bmr.smgr->smgr_rlocator.backend,
2575 : extend_by);
2576 :
2577 423200 : if (bmr.relpersistence == RELPERSISTENCE_TEMP)
2578 22704 : first_block = ExtendBufferedRelLocal(bmr, fork, flags,
2579 : extend_by, extend_upto,
2580 : buffers, &extend_by);
2581 : else
2582 400496 : first_block = ExtendBufferedRelShared(bmr, fork, strategy, flags,
2583 : extend_by, extend_upto,
2584 : buffers, &extend_by);
2585 423200 : *extended_by = extend_by;
2586 :
2587 : TRACE_POSTGRESQL_BUFFER_EXTEND_DONE(fork,
2588 : bmr.smgr->smgr_rlocator.locator.spcOid,
2589 : bmr.smgr->smgr_rlocator.locator.dbOid,
2590 : bmr.smgr->smgr_rlocator.locator.relNumber,
2591 : bmr.smgr->smgr_rlocator.backend,
2592 : *extended_by,
2593 : first_block);
2594 :
2595 423200 : return first_block;
2596 : }
2597 :
2598 : /*
2599 : * Implementation of ExtendBufferedRelBy() and ExtendBufferedRelTo() for
2600 : * shared buffers.
2601 : */
2602 : static BlockNumber
2603 400496 : ExtendBufferedRelShared(BufferManagerRelation bmr,
2604 : ForkNumber fork,
2605 : BufferAccessStrategy strategy,
2606 : uint32 flags,
2607 : uint32 extend_by,
2608 : BlockNumber extend_upto,
2609 : Buffer *buffers,
2610 : uint32 *extended_by)
2611 : {
2612 : BlockNumber first_block;
2613 400496 : IOContext io_context = IOContextForStrategy(strategy);
2614 : instr_time io_start;
2615 :
2616 400496 : LimitAdditionalPins(&extend_by);
2617 :
2618 : /*
2619 : * Acquire victim buffers for extension without holding extension lock.
2620 : * Writing out victim buffers is the most expensive part of extending the
2621 : * relation, particularly when doing so requires WAL flushes. Zeroing out
2622 : * the buffers is also quite expensive, so do that before holding the
2623 : * extension lock as well.
2624 : *
2625 : * These pages are pinned by us and not valid. While we hold the pin they
2626 : * can't be acquired as victim buffers by another backend.
2627 : */
2628 852688 : for (uint32 i = 0; i < extend_by; i++)
2629 : {
2630 : Block buf_block;
2631 :
2632 452192 : buffers[i] = GetVictimBuffer(strategy, io_context);
2633 452192 : buf_block = BufHdrGetBlock(GetBufferDescriptor(buffers[i] - 1));
2634 :
2635 : /* new buffers are zero-filled */
2636 452192 : MemSet(buf_block, 0, BLCKSZ);
2637 : }
2638 :
2639 : /*
2640 : * Lock relation against concurrent extensions, unless requested not to.
2641 : *
2642 : * We use the same extension lock for all forks. That's unnecessarily
2643 : * restrictive, but currently extensions for forks don't happen often
2644 : * enough to make it worth locking more granularly.
2645 : *
2646 : * Note that another backend might have extended the relation by the time
2647 : * we get the lock.
2648 : */
2649 400496 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2650 298696 : LockRelationForExtension(bmr.rel, ExclusiveLock);
2651 :
2652 : /*
2653 : * If requested, invalidate size cache, so that smgrnblocks asks the
2654 : * kernel.
2655 : */
2656 400496 : if (flags & EB_CLEAR_SIZE_CACHE)
2657 14998 : bmr.smgr->smgr_cached_nblocks[fork] = InvalidBlockNumber;
2658 :
2659 400496 : first_block = smgrnblocks(bmr.smgr, fork);
2660 :
2661 : /*
2662 : * Now that we have the accurate relation size, check if the caller wants
2663 : * us to extend to only up to a specific size. If there were concurrent
2664 : * extensions, we might have acquired too many buffers and need to release
2665 : * them.
2666 : */
2667 400496 : if (extend_upto != InvalidBlockNumber)
2668 : {
2669 104146 : uint32 orig_extend_by = extend_by;
2670 :
2671 104146 : if (first_block > extend_upto)
2672 0 : extend_by = 0;
2673 104146 : else if ((uint64) first_block + extend_by > extend_upto)
2674 26 : extend_by = extend_upto - first_block;
2675 :
2676 104206 : for (uint32 i = extend_by; i < orig_extend_by; i++)
2677 : {
2678 60 : BufferDesc *buf_hdr = GetBufferDescriptor(buffers[i] - 1);
2679 :
2680 : /*
2681 : * The victim buffer we acquired previously is clean and unused,
2682 : * let it be found again quickly
2683 : */
2684 60 : StrategyFreeBuffer(buf_hdr);
2685 60 : UnpinBuffer(buf_hdr);
2686 : }
2687 :
2688 104146 : if (extend_by == 0)
2689 : {
2690 26 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2691 26 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2692 26 : *extended_by = extend_by;
2693 26 : return first_block;
2694 : }
2695 : }
2696 :
2697 : /* Fail if relation is already at maximum possible length */
2698 400470 : if ((uint64) first_block + extend_by >= MaxBlockNumber)
2699 0 : ereport(ERROR,
2700 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
2701 : errmsg("cannot extend relation %s beyond %u blocks",
2702 : relpath(bmr.smgr->smgr_rlocator, fork).str,
2703 : MaxBlockNumber)));
2704 :
2705 : /*
2706 : * Insert buffers into buffer table, mark as IO_IN_PROGRESS.
2707 : *
2708 : * This needs to happen before we extend the relation, because as soon as
2709 : * we do, other backends can start to read in those pages.
2710 : */
2711 852602 : for (uint32 i = 0; i < extend_by; i++)
2712 : {
2713 452132 : Buffer victim_buf = buffers[i];
2714 452132 : BufferDesc *victim_buf_hdr = GetBufferDescriptor(victim_buf - 1);
2715 : BufferTag tag;
2716 : uint32 hash;
2717 : LWLock *partition_lock;
2718 : int existing_id;
2719 :
2720 : /* in case we need to pin an existing buffer below */
2721 452132 : ResourceOwnerEnlarge(CurrentResourceOwner);
2722 452132 : ReservePrivateRefCountEntry();
2723 :
2724 452132 : InitBufferTag(&tag, &bmr.smgr->smgr_rlocator.locator, fork, first_block + i);
2725 452132 : hash = BufTableHashCode(&tag);
2726 452132 : partition_lock = BufMappingPartitionLock(hash);
2727 :
2728 452132 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2729 :
2730 452132 : existing_id = BufTableInsert(&tag, hash, victim_buf_hdr->buf_id);
2731 :
2732 : /*
2733 : * We get here only in the corner case where we are trying to extend
2734 : * the relation but we found a pre-existing buffer. This can happen
2735 : * because a prior attempt at extending the relation failed, and
2736 : * because mdread doesn't complain about reads beyond EOF (when
2737 : * zero_damaged_pages is ON) and so a previous attempt to read a block
2738 : * beyond EOF could have left a "valid" zero-filled buffer.
2739 : * Unfortunately, we have also seen this case occurring because of
2740 : * buggy Linux kernels that sometimes return an lseek(SEEK_END) result
2741 : * that doesn't account for a recent write. In that situation, the
2742 : * pre-existing buffer would contain valid data that we don't want to
2743 : * overwrite. Since the legitimate cases should always have left a
2744 : * zero-filled buffer, complain if not PageIsNew.
2745 : */
2746 452132 : if (existing_id >= 0)
2747 : {
2748 0 : BufferDesc *existing_hdr = GetBufferDescriptor(existing_id);
2749 : Block buf_block;
2750 : bool valid;
2751 :
2752 : /*
2753 : * Pin the existing buffer before releasing the partition lock,
2754 : * preventing it from being evicted.
2755 : */
2756 0 : valid = PinBuffer(existing_hdr, strategy);
2757 :
2758 0 : LWLockRelease(partition_lock);
2759 :
2760 : /*
2761 : * The victim buffer we acquired previously is clean and unused,
2762 : * let it be found again quickly
2763 : */
2764 0 : StrategyFreeBuffer(victim_buf_hdr);
2765 0 : UnpinBuffer(victim_buf_hdr);
2766 :
2767 0 : buffers[i] = BufferDescriptorGetBuffer(existing_hdr);
2768 0 : buf_block = BufHdrGetBlock(existing_hdr);
2769 :
2770 0 : if (valid && !PageIsNew((Page) buf_block))
2771 0 : ereport(ERROR,
2772 : (errmsg("unexpected data beyond EOF in block %u of relation %s",
2773 : existing_hdr->tag.blockNum,
2774 : relpath(bmr.smgr->smgr_rlocator, fork).str),
2775 : errhint("This has been seen to occur with buggy kernels; consider updating your system.")));
2776 :
2777 : /*
2778 : * We *must* do smgr[zero]extend before succeeding, else the page
2779 : * will not be reserved by the kernel, and the next P_NEW call
2780 : * will decide to return the same page. Clear the BM_VALID bit,
2781 : * do StartBufferIO() and proceed.
2782 : *
2783 : * Loop to handle the very small possibility that someone re-sets
2784 : * BM_VALID between our clearing it and StartBufferIO inspecting
2785 : * it.
2786 : */
2787 : do
2788 : {
2789 0 : uint32 buf_state = LockBufHdr(existing_hdr);
2790 :
2791 0 : buf_state &= ~BM_VALID;
2792 0 : UnlockBufHdr(existing_hdr, buf_state);
2793 0 : } while (!StartBufferIO(existing_hdr, true, false));
2794 : }
2795 : else
2796 : {
2797 : uint32 buf_state;
2798 :
2799 452132 : buf_state = LockBufHdr(victim_buf_hdr);
2800 :
2801 : /* some sanity checks while we hold the buffer header lock */
2802 : Assert(!(buf_state & (BM_VALID | BM_TAG_VALID | BM_DIRTY | BM_JUST_DIRTIED)));
2803 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2804 :
2805 452132 : victim_buf_hdr->tag = tag;
2806 :
2807 452132 : buf_state |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2808 452132 : if (bmr.relpersistence == RELPERSISTENCE_PERMANENT || fork == INIT_FORKNUM)
2809 441558 : buf_state |= BM_PERMANENT;
2810 :
2811 452132 : UnlockBufHdr(victim_buf_hdr, buf_state);
2812 :
2813 452132 : LWLockRelease(partition_lock);
2814 :
2815 : /* XXX: could combine the locked operations in it with the above */
2816 452132 : StartBufferIO(victim_buf_hdr, true, false);
2817 : }
2818 : }
2819 :
2820 400470 : io_start = pgstat_prepare_io_time(track_io_timing);
2821 :
2822 : /*
2823 : * Note: if smgrzeroextend fails, we will end up with buffers that are
2824 : * allocated but not marked BM_VALID. The next relation extension will
2825 : * still select the same block number (because the relation didn't get any
2826 : * longer on disk) and so future attempts to extend the relation will find
2827 : * the same buffers (if they have not been recycled) but come right back
2828 : * here to try smgrzeroextend again.
2829 : *
2830 : * We don't need to set checksum for all-zero pages.
2831 : */
2832 400470 : smgrzeroextend(bmr.smgr, fork, first_block, extend_by, false);
2833 :
2834 : /*
2835 : * Release the file-extension lock; it's now OK for someone else to extend
2836 : * the relation some more.
2837 : *
2838 : * We remove IO_IN_PROGRESS after this, as waking up waiting backends can
2839 : * take noticeable time.
2840 : */
2841 400470 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2842 298670 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2843 :
2844 400470 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context, IOOP_EXTEND,
2845 400470 : io_start, 1, extend_by * BLCKSZ);
2846 :
2847 : /* Set BM_VALID, terminate IO, and wake up any waiters */
2848 852602 : for (uint32 i = 0; i < extend_by; i++)
2849 : {
2850 452132 : Buffer buf = buffers[i];
2851 452132 : BufferDesc *buf_hdr = GetBufferDescriptor(buf - 1);
2852 452132 : bool lock = false;
2853 :
2854 452132 : if (flags & EB_LOCK_FIRST && i == 0)
2855 295764 : lock = true;
2856 156368 : else if (flags & EB_LOCK_TARGET)
2857 : {
2858 : Assert(extend_upto != InvalidBlockNumber);
2859 87068 : if (first_block + i + 1 == extend_upto)
2860 85910 : lock = true;
2861 : }
2862 :
2863 452132 : if (lock)
2864 381674 : LWLockAcquire(BufferDescriptorGetContentLock(buf_hdr), LW_EXCLUSIVE);
2865 :
2866 452132 : TerminateBufferIO(buf_hdr, false, BM_VALID, true, false);
2867 : }
2868 :
2869 400470 : pgBufferUsage.shared_blks_written += extend_by;
2870 :
2871 400470 : *extended_by = extend_by;
2872 :
2873 400470 : return first_block;
2874 : }
2875 :
2876 : /*
2877 : * BufferIsExclusiveLocked
2878 : *
2879 : * Checks if buffer is exclusive-locked.
2880 : *
2881 : * Buffer must be pinned.
2882 : */
2883 : bool
2884 0 : BufferIsExclusiveLocked(Buffer buffer)
2885 : {
2886 : BufferDesc *bufHdr;
2887 :
2888 : Assert(BufferIsPinned(buffer));
2889 :
2890 0 : if (BufferIsLocal(buffer))
2891 : {
2892 : /* Content locks are not maintained for local buffers. */
2893 0 : return true;
2894 : }
2895 : else
2896 : {
2897 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2898 0 : return LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
2899 : LW_EXCLUSIVE);
2900 : }
2901 : }
2902 :
2903 : /*
2904 : * BufferIsDirty
2905 : *
2906 : * Checks if buffer is already dirty.
2907 : *
2908 : * Buffer must be pinned and exclusive-locked. (Without an exclusive lock,
2909 : * the result may be stale before it's returned.)
2910 : */
2911 : bool
2912 0 : BufferIsDirty(Buffer buffer)
2913 : {
2914 : BufferDesc *bufHdr;
2915 :
2916 : Assert(BufferIsPinned(buffer));
2917 :
2918 0 : if (BufferIsLocal(buffer))
2919 : {
2920 0 : int bufid = -buffer - 1;
2921 :
2922 0 : bufHdr = GetLocalBufferDescriptor(bufid);
2923 : /* Content locks are not maintained for local buffers. */
2924 : }
2925 : else
2926 : {
2927 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2928 : Assert(LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
2929 : LW_EXCLUSIVE));
2930 : }
2931 :
2932 0 : return pg_atomic_read_u32(&bufHdr->state) & BM_DIRTY;
2933 : }
2934 :
2935 : /*
2936 : * MarkBufferDirty
2937 : *
2938 : * Marks buffer contents as dirty (actual write happens later).
2939 : *
2940 : * Buffer must be pinned and exclusive-locked. (If caller does not hold
2941 : * exclusive lock, then somebody could be in process of writing the buffer,
2942 : * leading to risk of bad data written to disk.)
2943 : */
2944 : void
2945 43079152 : MarkBufferDirty(Buffer buffer)
2946 : {
2947 : BufferDesc *bufHdr;
2948 : uint32 buf_state;
2949 : uint32 old_buf_state;
2950 :
2951 43079152 : if (!BufferIsValid(buffer))
2952 0 : elog(ERROR, "bad buffer ID: %d", buffer);
2953 :
2954 43079152 : if (BufferIsLocal(buffer))
2955 : {
2956 2436568 : MarkLocalBufferDirty(buffer);
2957 2436568 : return;
2958 : }
2959 :
2960 40642584 : bufHdr = GetBufferDescriptor(buffer - 1);
2961 :
2962 : Assert(BufferIsPinned(buffer));
2963 : Assert(LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
2964 : LW_EXCLUSIVE));
2965 :
2966 40642584 : old_buf_state = pg_atomic_read_u32(&bufHdr->state);
2967 : for (;;)
2968 : {
2969 40642774 : if (old_buf_state & BM_LOCKED)
2970 52 : old_buf_state = WaitBufHdrUnlocked(bufHdr);
2971 :
2972 40642774 : buf_state = old_buf_state;
2973 :
2974 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2975 40642774 : buf_state |= BM_DIRTY | BM_JUST_DIRTIED;
2976 :
2977 40642774 : if (pg_atomic_compare_exchange_u32(&bufHdr->state, &old_buf_state,
2978 : buf_state))
2979 40642584 : break;
2980 : }
2981 :
2982 : /*
2983 : * If the buffer was not dirty already, do vacuum accounting.
2984 : */
2985 40642584 : if (!(old_buf_state & BM_DIRTY))
2986 : {
2987 1252844 : pgBufferUsage.shared_blks_dirtied++;
2988 1252844 : if (VacuumCostActive)
2989 12090 : VacuumCostBalance += VacuumCostPageDirty;
2990 : }
2991 : }
2992 :
2993 : /*
2994 : * ReleaseAndReadBuffer -- combine ReleaseBuffer() and ReadBuffer()
2995 : *
2996 : * Formerly, this saved one cycle of acquiring/releasing the BufMgrLock
2997 : * compared to calling the two routines separately. Now it's mainly just
2998 : * a convenience function. However, if the passed buffer is valid and
2999 : * already contains the desired block, we just return it as-is; and that
3000 : * does save considerable work compared to a full release and reacquire.
3001 : *
3002 : * Note: it is OK to pass buffer == InvalidBuffer, indicating that no old
3003 : * buffer actually needs to be released. This case is the same as ReadBuffer,
3004 : * but can save some tests in the caller.
3005 : */
3006 : Buffer
3007 54441966 : ReleaseAndReadBuffer(Buffer buffer,
3008 : Relation relation,
3009 : BlockNumber blockNum)
3010 : {
3011 54441966 : ForkNumber forkNum = MAIN_FORKNUM;
3012 : BufferDesc *bufHdr;
3013 :
3014 54441966 : if (BufferIsValid(buffer))
3015 : {
3016 : Assert(BufferIsPinned(buffer));
3017 32618296 : if (BufferIsLocal(buffer))
3018 : {
3019 73716 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
3020 80760 : if (bufHdr->tag.blockNum == blockNum &&
3021 14088 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3022 7044 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3023 7044 : return buffer;
3024 66672 : UnpinLocalBuffer(buffer);
3025 : }
3026 : else
3027 : {
3028 32544580 : bufHdr = GetBufferDescriptor(buffer - 1);
3029 : /* we have pin, so it's ok to examine tag without spinlock */
3030 43845860 : if (bufHdr->tag.blockNum == blockNum &&
3031 22602560 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3032 11301280 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3033 11301280 : return buffer;
3034 21243300 : UnpinBuffer(bufHdr);
3035 : }
3036 : }
3037 :
3038 43133642 : return ReadBuffer(relation, blockNum);
3039 : }
3040 :
3041 : /*
3042 : * PinBuffer -- make buffer unavailable for replacement.
3043 : *
3044 : * For the default access strategy, the buffer's usage_count is incremented
3045 : * when we first pin it; for other strategies we just make sure the usage_count
3046 : * isn't zero. (The idea of the latter is that we don't want synchronized
3047 : * heap scans to inflate the count, but we need it to not be zero to discourage
3048 : * other backends from stealing buffers from our ring. As long as we cycle
3049 : * through the ring faster than the global clock-sweep cycles, buffers in
3050 : * our ring won't be chosen as victims for replacement by other backends.)
3051 : *
3052 : * This should be applied only to shared buffers, never local ones.
3053 : *
3054 : * Since buffers are pinned/unpinned very frequently, pin buffers without
3055 : * taking the buffer header lock; instead update the state variable in loop of
3056 : * CAS operations. Hopefully it's just a single CAS.
3057 : *
3058 : * Note that ResourceOwnerEnlarge() and ReservePrivateRefCountEntry()
3059 : * must have been done already.
3060 : *
3061 : * Returns true if buffer is BM_VALID, else false. This provision allows
3062 : * some callers to avoid an extra spinlock cycle.
3063 : */
3064 : static bool
3065 112854806 : PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy)
3066 : {
3067 112854806 : Buffer b = BufferDescriptorGetBuffer(buf);
3068 : bool result;
3069 : PrivateRefCountEntry *ref;
3070 :
3071 : Assert(!BufferIsLocal(b));
3072 : Assert(ReservedRefCountEntry != NULL);
3073 :
3074 112854806 : ref = GetPrivateRefCountEntry(b, true);
3075 :
3076 112854806 : if (ref == NULL)
3077 : {
3078 : uint32 buf_state;
3079 : uint32 old_buf_state;
3080 :
3081 108542936 : ref = NewPrivateRefCountEntry(b);
3082 :
3083 108542936 : old_buf_state = pg_atomic_read_u32(&buf->state);
3084 : for (;;)
3085 : {
3086 108569718 : if (old_buf_state & BM_LOCKED)
3087 1964 : old_buf_state = WaitBufHdrUnlocked(buf);
3088 :
3089 108569718 : buf_state = old_buf_state;
3090 :
3091 : /* increase refcount */
3092 108569718 : buf_state += BUF_REFCOUNT_ONE;
3093 :
3094 108569718 : if (strategy == NULL)
3095 : {
3096 : /* Default case: increase usagecount unless already max. */
3097 107263424 : if (BUF_STATE_GET_USAGECOUNT(buf_state) < BM_MAX_USAGE_COUNT)
3098 6137744 : buf_state += BUF_USAGECOUNT_ONE;
3099 : }
3100 : else
3101 : {
3102 : /*
3103 : * Ring buffers shouldn't evict others from pool. Thus we
3104 : * don't make usagecount more than 1.
3105 : */
3106 1306294 : if (BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3107 68830 : buf_state += BUF_USAGECOUNT_ONE;
3108 : }
3109 :
3110 108569718 : if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
3111 : buf_state))
3112 : {
3113 108542936 : result = (buf_state & BM_VALID) != 0;
3114 :
3115 : /*
3116 : * Assume that we acquired a buffer pin for the purposes of
3117 : * Valgrind buffer client checks (even in !result case) to
3118 : * keep things simple. Buffers that are unsafe to access are
3119 : * not generally guaranteed to be marked undefined or
3120 : * non-accessible in any case.
3121 : */
3122 : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(buf), BLCKSZ);
3123 108542936 : break;
3124 : }
3125 : }
3126 : }
3127 : else
3128 : {
3129 : /*
3130 : * If we previously pinned the buffer, it is likely to be valid, but
3131 : * it may not be if StartReadBuffers() was called and
3132 : * WaitReadBuffers() hasn't been called yet. We'll check by loading
3133 : * the flags without locking. This is racy, but it's OK to return
3134 : * false spuriously: when WaitReadBuffers() calls StartBufferIO(),
3135 : * it'll see that it's now valid.
3136 : *
3137 : * Note: We deliberately avoid a Valgrind client request here.
3138 : * Individual access methods can optionally superimpose buffer page
3139 : * client requests on top of our client requests to enforce that
3140 : * buffers are only accessed while locked (and pinned). It's possible
3141 : * that the buffer page is legitimately non-accessible here. We
3142 : * cannot meddle with that.
3143 : */
3144 4311870 : result = (pg_atomic_read_u32(&buf->state) & BM_VALID) != 0;
3145 : }
3146 :
3147 112854806 : ref->refcount++;
3148 : Assert(ref->refcount > 0);
3149 112854806 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
3150 112854806 : return result;
3151 : }
3152 :
3153 : /*
3154 : * PinBuffer_Locked -- as above, but caller already locked the buffer header.
3155 : * The spinlock is released before return.
3156 : *
3157 : * As this function is called with the spinlock held, the caller has to
3158 : * previously call ReservePrivateRefCountEntry() and
3159 : * ResourceOwnerEnlarge(CurrentResourceOwner);
3160 : *
3161 : * Currently, no callers of this function want to modify the buffer's
3162 : * usage_count at all, so there's no need for a strategy parameter.
3163 : * Also we don't bother with a BM_VALID test (the caller could check that for
3164 : * itself).
3165 : *
3166 : * Also all callers only ever use this function when it's known that the
3167 : * buffer can't have a preexisting pin by this backend. That allows us to skip
3168 : * searching the private refcount array & hash, which is a boon, because the
3169 : * spinlock is still held.
3170 : *
3171 : * Note: use of this routine is frequently mandatory, not just an optimization
3172 : * to save a spin lock/unlock cycle, because we need to pin a buffer before
3173 : * its state can change under us.
3174 : */
3175 : static void
3176 4282700 : PinBuffer_Locked(BufferDesc *buf)
3177 : {
3178 : Buffer b;
3179 : PrivateRefCountEntry *ref;
3180 : uint32 buf_state;
3181 :
3182 : /*
3183 : * As explained, We don't expect any preexisting pins. That allows us to
3184 : * manipulate the PrivateRefCount after releasing the spinlock
3185 : */
3186 : Assert(GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf), false) == NULL);
3187 :
3188 : /*
3189 : * Buffer can't have a preexisting pin, so mark its page as defined to
3190 : * Valgrind (this is similar to the PinBuffer() case where the backend
3191 : * doesn't already have a buffer pin)
3192 : */
3193 : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(buf), BLCKSZ);
3194 :
3195 : /*
3196 : * Since we hold the buffer spinlock, we can update the buffer state and
3197 : * release the lock in one operation.
3198 : */
3199 4282700 : buf_state = pg_atomic_read_u32(&buf->state);
3200 : Assert(buf_state & BM_LOCKED);
3201 4282700 : buf_state += BUF_REFCOUNT_ONE;
3202 4282700 : UnlockBufHdr(buf, buf_state);
3203 :
3204 4282700 : b = BufferDescriptorGetBuffer(buf);
3205 :
3206 4282700 : ref = NewPrivateRefCountEntry(b);
3207 4282700 : ref->refcount++;
3208 :
3209 4282700 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
3210 4282700 : }
3211 :
3212 : /*
3213 : * Support for waking up another backend that is waiting for the cleanup lock
3214 : * to be released using BM_PIN_COUNT_WAITER.
3215 : *
3216 : * See LockBufferForCleanup().
3217 : *
3218 : * Expected to be called just after releasing a buffer pin (in a BufferDesc,
3219 : * not just reducing the backend-local pincount for the buffer).
3220 : */
3221 : static void
3222 4 : WakePinCountWaiter(BufferDesc *buf)
3223 : {
3224 : /*
3225 : * Acquire the buffer header lock, re-check that there's a waiter. Another
3226 : * backend could have unpinned this buffer, and already woken up the
3227 : * waiter.
3228 : *
3229 : * There's no danger of the buffer being replaced after we unpinned it
3230 : * above, as it's pinned by the waiter. The waiter removes
3231 : * BM_PIN_COUNT_WAITER if it stops waiting for a reason other than this
3232 : * backend waking it up.
3233 : */
3234 4 : uint32 buf_state = LockBufHdr(buf);
3235 :
3236 4 : if ((buf_state & BM_PIN_COUNT_WAITER) &&
3237 4 : BUF_STATE_GET_REFCOUNT(buf_state) == 1)
3238 4 : {
3239 : /* we just released the last pin other than the waiter's */
3240 4 : int wait_backend_pgprocno = buf->wait_backend_pgprocno;
3241 :
3242 4 : buf_state &= ~BM_PIN_COUNT_WAITER;
3243 4 : UnlockBufHdr(buf, buf_state);
3244 4 : ProcSendSignal(wait_backend_pgprocno);
3245 : }
3246 : else
3247 0 : UnlockBufHdr(buf, buf_state);
3248 4 : }
3249 :
3250 : /*
3251 : * UnpinBuffer -- make buffer available for replacement.
3252 : *
3253 : * This should be applied only to shared buffers, never local ones. This
3254 : * always adjusts CurrentResourceOwner.
3255 : */
3256 : static void
3257 138314340 : UnpinBuffer(BufferDesc *buf)
3258 : {
3259 138314340 : Buffer b = BufferDescriptorGetBuffer(buf);
3260 :
3261 138314340 : ResourceOwnerForgetBuffer(CurrentResourceOwner, b);
3262 138314340 : UnpinBufferNoOwner(buf);
3263 138314340 : }
3264 :
3265 : static void
3266 138322874 : UnpinBufferNoOwner(BufferDesc *buf)
3267 : {
3268 : PrivateRefCountEntry *ref;
3269 138322874 : Buffer b = BufferDescriptorGetBuffer(buf);
3270 :
3271 : Assert(!BufferIsLocal(b));
3272 :
3273 : /* not moving as we're likely deleting it soon anyway */
3274 138322874 : ref = GetPrivateRefCountEntry(b, false);
3275 : Assert(ref != NULL);
3276 : Assert(ref->refcount > 0);
3277 138322874 : ref->refcount--;
3278 138322874 : if (ref->refcount == 0)
3279 : {
3280 : uint32 buf_state;
3281 : uint32 old_buf_state;
3282 :
3283 : /*
3284 : * Mark buffer non-accessible to Valgrind.
3285 : *
3286 : * Note that the buffer may have already been marked non-accessible
3287 : * within access method code that enforces that buffers are only
3288 : * accessed while a buffer lock is held.
3289 : */
3290 : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
3291 :
3292 : /* I'd better not still hold the buffer content lock */
3293 : Assert(!LWLockHeldByMe(BufferDescriptorGetContentLock(buf)));
3294 :
3295 : /*
3296 : * Decrement the shared reference count.
3297 : *
3298 : * Since buffer spinlock holder can update status using just write,
3299 : * it's not safe to use atomic decrement here; thus use a CAS loop.
3300 : */
3301 112825636 : old_buf_state = pg_atomic_read_u32(&buf->state);
3302 : for (;;)
3303 : {
3304 112850900 : if (old_buf_state & BM_LOCKED)
3305 1800 : old_buf_state = WaitBufHdrUnlocked(buf);
3306 :
3307 112850900 : buf_state = old_buf_state;
3308 :
3309 112850900 : buf_state -= BUF_REFCOUNT_ONE;
3310 :
3311 112850900 : if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
3312 : buf_state))
3313 112825636 : break;
3314 : }
3315 :
3316 : /* Support LockBufferForCleanup() */
3317 112825636 : if (buf_state & BM_PIN_COUNT_WAITER)
3318 4 : WakePinCountWaiter(buf);
3319 :
3320 112825636 : ForgetPrivateRefCountEntry(ref);
3321 : }
3322 138322874 : }
3323 :
3324 : #define ST_SORT sort_checkpoint_bufferids
3325 : #define ST_ELEMENT_TYPE CkptSortItem
3326 : #define ST_COMPARE(a, b) ckpt_buforder_comparator(a, b)
3327 : #define ST_SCOPE static
3328 : #define ST_DEFINE
3329 : #include <lib/sort_template.h>
3330 :
3331 : /*
3332 : * BufferSync -- Write out all dirty buffers in the pool.
3333 : *
3334 : * This is called at checkpoint time to write out all dirty shared buffers.
3335 : * The checkpoint request flags should be passed in. If CHECKPOINT_IMMEDIATE
3336 : * is set, we disable delays between writes; if CHECKPOINT_IS_SHUTDOWN,
3337 : * CHECKPOINT_END_OF_RECOVERY or CHECKPOINT_FLUSH_ALL is set, we write even
3338 : * unlogged buffers, which are otherwise skipped. The remaining flags
3339 : * currently have no effect here.
3340 : */
3341 : static void
3342 2642 : BufferSync(int flags)
3343 : {
3344 : uint32 buf_state;
3345 : int buf_id;
3346 : int num_to_scan;
3347 : int num_spaces;
3348 : int num_processed;
3349 : int num_written;
3350 2642 : CkptTsStatus *per_ts_stat = NULL;
3351 : Oid last_tsid;
3352 : binaryheap *ts_heap;
3353 : int i;
3354 2642 : int mask = BM_DIRTY;
3355 : WritebackContext wb_context;
3356 :
3357 : /*
3358 : * Unless this is a shutdown checkpoint or we have been explicitly told,
3359 : * we write only permanent, dirty buffers. But at shutdown or end of
3360 : * recovery, we write all dirty buffers.
3361 : */
3362 2642 : if (!((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
3363 : CHECKPOINT_FLUSH_ALL))))
3364 1260 : mask |= BM_PERMANENT;
3365 :
3366 : /*
3367 : * Loop over all buffers, and mark the ones that need to be written with
3368 : * BM_CHECKPOINT_NEEDED. Count them as we go (num_to_scan), so that we
3369 : * can estimate how much work needs to be done.
3370 : *
3371 : * This allows us to write only those pages that were dirty when the
3372 : * checkpoint began, and not those that get dirtied while it proceeds.
3373 : * Whenever a page with BM_CHECKPOINT_NEEDED is written out, either by us
3374 : * later in this function, or by normal backends or the bgwriter cleaning
3375 : * scan, the flag is cleared. Any buffer dirtied after this point won't
3376 : * have the flag set.
3377 : *
3378 : * Note that if we fail to write some buffer, we may leave buffers with
3379 : * BM_CHECKPOINT_NEEDED still set. This is OK since any such buffer would
3380 : * certainly need to be written for the next checkpoint attempt, too.
3381 : */
3382 2642 : num_to_scan = 0;
3383 23001074 : for (buf_id = 0; buf_id < NBuffers; buf_id++)
3384 : {
3385 22998432 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3386 :
3387 : /*
3388 : * Header spinlock is enough to examine BM_DIRTY, see comment in
3389 : * SyncOneBuffer.
3390 : */
3391 22998432 : buf_state = LockBufHdr(bufHdr);
3392 :
3393 22998432 : if ((buf_state & mask) == mask)
3394 : {
3395 : CkptSortItem *item;
3396 :
3397 553674 : buf_state |= BM_CHECKPOINT_NEEDED;
3398 :
3399 553674 : item = &CkptBufferIds[num_to_scan++];
3400 553674 : item->buf_id = buf_id;
3401 553674 : item->tsId = bufHdr->tag.spcOid;
3402 553674 : item->relNumber = BufTagGetRelNumber(&bufHdr->tag);
3403 553674 : item->forkNum = BufTagGetForkNum(&bufHdr->tag);
3404 553674 : item->blockNum = bufHdr->tag.blockNum;
3405 : }
3406 :
3407 22998432 : UnlockBufHdr(bufHdr, buf_state);
3408 :
3409 : /* Check for barrier events in case NBuffers is large. */
3410 22998432 : if (ProcSignalBarrierPending)
3411 0 : ProcessProcSignalBarrier();
3412 : }
3413 :
3414 2642 : if (num_to_scan == 0)
3415 610 : return; /* nothing to do */
3416 :
3417 2032 : WritebackContextInit(&wb_context, &checkpoint_flush_after);
3418 :
3419 : TRACE_POSTGRESQL_BUFFER_SYNC_START(NBuffers, num_to_scan);
3420 :
3421 : /*
3422 : * Sort buffers that need to be written to reduce the likelihood of random
3423 : * IO. The sorting is also important for the implementation of balancing
3424 : * writes between tablespaces. Without balancing writes we'd potentially
3425 : * end up writing to the tablespaces one-by-one; possibly overloading the
3426 : * underlying system.
3427 : */
3428 2032 : sort_checkpoint_bufferids(CkptBufferIds, num_to_scan);
3429 :
3430 2032 : num_spaces = 0;
3431 :
3432 : /*
3433 : * Allocate progress status for each tablespace with buffers that need to
3434 : * be flushed. This requires the to-be-flushed array to be sorted.
3435 : */
3436 2032 : last_tsid = InvalidOid;
3437 555706 : for (i = 0; i < num_to_scan; i++)
3438 : {
3439 : CkptTsStatus *s;
3440 : Oid cur_tsid;
3441 :
3442 553674 : cur_tsid = CkptBufferIds[i].tsId;
3443 :
3444 : /*
3445 : * Grow array of per-tablespace status structs, every time a new
3446 : * tablespace is found.
3447 : */
3448 553674 : if (last_tsid == InvalidOid || last_tsid != cur_tsid)
3449 3012 : {
3450 : Size sz;
3451 :
3452 3012 : num_spaces++;
3453 :
3454 : /*
3455 : * Not worth adding grow-by-power-of-2 logic here - even with a
3456 : * few hundred tablespaces this should be fine.
3457 : */
3458 3012 : sz = sizeof(CkptTsStatus) * num_spaces;
3459 :
3460 3012 : if (per_ts_stat == NULL)
3461 2032 : per_ts_stat = (CkptTsStatus *) palloc(sz);
3462 : else
3463 980 : per_ts_stat = (CkptTsStatus *) repalloc(per_ts_stat, sz);
3464 :
3465 3012 : s = &per_ts_stat[num_spaces - 1];
3466 3012 : memset(s, 0, sizeof(*s));
3467 3012 : s->tsId = cur_tsid;
3468 :
3469 : /*
3470 : * The first buffer in this tablespace. As CkptBufferIds is sorted
3471 : * by tablespace all (s->num_to_scan) buffers in this tablespace
3472 : * will follow afterwards.
3473 : */
3474 3012 : s->index = i;
3475 :
3476 : /*
3477 : * progress_slice will be determined once we know how many buffers
3478 : * are in each tablespace, i.e. after this loop.
3479 : */
3480 :
3481 3012 : last_tsid = cur_tsid;
3482 : }
3483 : else
3484 : {
3485 550662 : s = &per_ts_stat[num_spaces - 1];
3486 : }
3487 :
3488 553674 : s->num_to_scan++;
3489 :
3490 : /* Check for barrier events. */
3491 553674 : if (ProcSignalBarrierPending)
3492 0 : ProcessProcSignalBarrier();
3493 : }
3494 :
3495 : Assert(num_spaces > 0);
3496 :
3497 : /*
3498 : * Build a min-heap over the write-progress in the individual tablespaces,
3499 : * and compute how large a portion of the total progress a single
3500 : * processed buffer is.
3501 : */
3502 2032 : ts_heap = binaryheap_allocate(num_spaces,
3503 : ts_ckpt_progress_comparator,
3504 : NULL);
3505 :
3506 5044 : for (i = 0; i < num_spaces; i++)
3507 : {
3508 3012 : CkptTsStatus *ts_stat = &per_ts_stat[i];
3509 :
3510 3012 : ts_stat->progress_slice = (float8) num_to_scan / ts_stat->num_to_scan;
3511 :
3512 3012 : binaryheap_add_unordered(ts_heap, PointerGetDatum(ts_stat));
3513 : }
3514 :
3515 2032 : binaryheap_build(ts_heap);
3516 :
3517 : /*
3518 : * Iterate through to-be-checkpointed buffers and write the ones (still)
3519 : * marked with BM_CHECKPOINT_NEEDED. The writes are balanced between
3520 : * tablespaces; otherwise the sorting would lead to only one tablespace
3521 : * receiving writes at a time, making inefficient use of the hardware.
3522 : */
3523 2032 : num_processed = 0;
3524 2032 : num_written = 0;
3525 555706 : while (!binaryheap_empty(ts_heap))
3526 : {
3527 553674 : BufferDesc *bufHdr = NULL;
3528 : CkptTsStatus *ts_stat = (CkptTsStatus *)
3529 553674 : DatumGetPointer(binaryheap_first(ts_heap));
3530 :
3531 553674 : buf_id = CkptBufferIds[ts_stat->index].buf_id;
3532 : Assert(buf_id != -1);
3533 :
3534 553674 : bufHdr = GetBufferDescriptor(buf_id);
3535 :
3536 553674 : num_processed++;
3537 :
3538 : /*
3539 : * We don't need to acquire the lock here, because we're only looking
3540 : * at a single bit. It's possible that someone else writes the buffer
3541 : * and clears the flag right after we check, but that doesn't matter
3542 : * since SyncOneBuffer will then do nothing. However, there is a
3543 : * further race condition: it's conceivable that between the time we
3544 : * examine the bit here and the time SyncOneBuffer acquires the lock,
3545 : * someone else not only wrote the buffer but replaced it with another
3546 : * page and dirtied it. In that improbable case, SyncOneBuffer will
3547 : * write the buffer though we didn't need to. It doesn't seem worth
3548 : * guarding against this, though.
3549 : */
3550 553674 : if (pg_atomic_read_u32(&bufHdr->state) & BM_CHECKPOINT_NEEDED)
3551 : {
3552 515004 : if (SyncOneBuffer(buf_id, false, &wb_context) & BUF_WRITTEN)
3553 : {
3554 : TRACE_POSTGRESQL_BUFFER_SYNC_WRITTEN(buf_id);
3555 515004 : PendingCheckpointerStats.buffers_written++;
3556 515004 : num_written++;
3557 : }
3558 : }
3559 :
3560 : /*
3561 : * Measure progress independent of actually having to flush the buffer
3562 : * - otherwise writing become unbalanced.
3563 : */
3564 553674 : ts_stat->progress += ts_stat->progress_slice;
3565 553674 : ts_stat->num_scanned++;
3566 553674 : ts_stat->index++;
3567 :
3568 : /* Have all the buffers from the tablespace been processed? */
3569 553674 : if (ts_stat->num_scanned == ts_stat->num_to_scan)
3570 : {
3571 3012 : binaryheap_remove_first(ts_heap);
3572 : }
3573 : else
3574 : {
3575 : /* update heap with the new progress */
3576 550662 : binaryheap_replace_first(ts_heap, PointerGetDatum(ts_stat));
3577 : }
3578 :
3579 : /*
3580 : * Sleep to throttle our I/O rate.
3581 : *
3582 : * (This will check for barrier events even if it doesn't sleep.)
3583 : */
3584 553674 : CheckpointWriteDelay(flags, (double) num_processed / num_to_scan);
3585 : }
3586 :
3587 : /*
3588 : * Issue all pending flushes. Only checkpointer calls BufferSync(), so
3589 : * IOContext will always be IOCONTEXT_NORMAL.
3590 : */
3591 2032 : IssuePendingWritebacks(&wb_context, IOCONTEXT_NORMAL);
3592 :
3593 2032 : pfree(per_ts_stat);
3594 2032 : per_ts_stat = NULL;
3595 2032 : binaryheap_free(ts_heap);
3596 :
3597 : /*
3598 : * Update checkpoint statistics. As noted above, this doesn't include
3599 : * buffers written by other backends or bgwriter scan.
3600 : */
3601 2032 : CheckpointStats.ckpt_bufs_written += num_written;
3602 :
3603 : TRACE_POSTGRESQL_BUFFER_SYNC_DONE(NBuffers, num_written, num_to_scan);
3604 : }
3605 :
3606 : /*
3607 : * BgBufferSync -- Write out some dirty buffers in the pool.
3608 : *
3609 : * This is called periodically by the background writer process.
3610 : *
3611 : * Returns true if it's appropriate for the bgwriter process to go into
3612 : * low-power hibernation mode. (This happens if the strategy clock sweep
3613 : * has been "lapped" and no buffer allocations have occurred recently,
3614 : * or if the bgwriter has been effectively disabled by setting
3615 : * bgwriter_lru_maxpages to 0.)
3616 : */
3617 : bool
3618 20058 : BgBufferSync(WritebackContext *wb_context)
3619 : {
3620 : /* info obtained from freelist.c */
3621 : int strategy_buf_id;
3622 : uint32 strategy_passes;
3623 : uint32 recent_alloc;
3624 :
3625 : /*
3626 : * Information saved between calls so we can determine the strategy
3627 : * point's advance rate and avoid scanning already-cleaned buffers.
3628 : */
3629 : static bool saved_info_valid = false;
3630 : static int prev_strategy_buf_id;
3631 : static uint32 prev_strategy_passes;
3632 : static int next_to_clean;
3633 : static uint32 next_passes;
3634 :
3635 : /* Moving averages of allocation rate and clean-buffer density */
3636 : static float smoothed_alloc = 0;
3637 : static float smoothed_density = 10.0;
3638 :
3639 : /* Potentially these could be tunables, but for now, not */
3640 20058 : float smoothing_samples = 16;
3641 20058 : float scan_whole_pool_milliseconds = 120000.0;
3642 :
3643 : /* Used to compute how far we scan ahead */
3644 : long strategy_delta;
3645 : int bufs_to_lap;
3646 : int bufs_ahead;
3647 : float scans_per_alloc;
3648 : int reusable_buffers_est;
3649 : int upcoming_alloc_est;
3650 : int min_scan_buffers;
3651 :
3652 : /* Variables for the scanning loop proper */
3653 : int num_to_scan;
3654 : int num_written;
3655 : int reusable_buffers;
3656 :
3657 : /* Variables for final smoothed_density update */
3658 : long new_strategy_delta;
3659 : uint32 new_recent_alloc;
3660 :
3661 : /*
3662 : * Find out where the freelist clock sweep currently is, and how many
3663 : * buffer allocations have happened since our last call.
3664 : */
3665 20058 : strategy_buf_id = StrategySyncStart(&strategy_passes, &recent_alloc);
3666 :
3667 : /* Report buffer alloc counts to pgstat */
3668 20058 : PendingBgWriterStats.buf_alloc += recent_alloc;
3669 :
3670 : /*
3671 : * If we're not running the LRU scan, just stop after doing the stats
3672 : * stuff. We mark the saved state invalid so that we can recover sanely
3673 : * if LRU scan is turned back on later.
3674 : */
3675 20058 : if (bgwriter_lru_maxpages <= 0)
3676 : {
3677 38 : saved_info_valid = false;
3678 38 : return true;
3679 : }
3680 :
3681 : /*
3682 : * Compute strategy_delta = how many buffers have been scanned by the
3683 : * clock sweep since last time. If first time through, assume none. Then
3684 : * see if we are still ahead of the clock sweep, and if so, how many
3685 : * buffers we could scan before we'd catch up with it and "lap" it. Note:
3686 : * weird-looking coding of xxx_passes comparisons are to avoid bogus
3687 : * behavior when the passes counts wrap around.
3688 : */
3689 20020 : if (saved_info_valid)
3690 : {
3691 19038 : int32 passes_delta = strategy_passes - prev_strategy_passes;
3692 :
3693 19038 : strategy_delta = strategy_buf_id - prev_strategy_buf_id;
3694 19038 : strategy_delta += (long) passes_delta * NBuffers;
3695 :
3696 : Assert(strategy_delta >= 0);
3697 :
3698 19038 : if ((int32) (next_passes - strategy_passes) > 0)
3699 : {
3700 : /* we're one pass ahead of the strategy point */
3701 3382 : bufs_to_lap = strategy_buf_id - next_to_clean;
3702 : #ifdef BGW_DEBUG
3703 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3704 : next_passes, next_to_clean,
3705 : strategy_passes, strategy_buf_id,
3706 : strategy_delta, bufs_to_lap);
3707 : #endif
3708 : }
3709 15656 : else if (next_passes == strategy_passes &&
3710 12470 : next_to_clean >= strategy_buf_id)
3711 : {
3712 : /* on same pass, but ahead or at least not behind */
3713 12160 : bufs_to_lap = NBuffers - (next_to_clean - strategy_buf_id);
3714 : #ifdef BGW_DEBUG
3715 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3716 : next_passes, next_to_clean,
3717 : strategy_passes, strategy_buf_id,
3718 : strategy_delta, bufs_to_lap);
3719 : #endif
3720 : }
3721 : else
3722 : {
3723 : /*
3724 : * We're behind, so skip forward to the strategy point and start
3725 : * cleaning from there.
3726 : */
3727 : #ifdef BGW_DEBUG
3728 : elog(DEBUG2, "bgwriter behind: bgw %u-%u strategy %u-%u delta=%ld",
3729 : next_passes, next_to_clean,
3730 : strategy_passes, strategy_buf_id,
3731 : strategy_delta);
3732 : #endif
3733 3496 : next_to_clean = strategy_buf_id;
3734 3496 : next_passes = strategy_passes;
3735 3496 : bufs_to_lap = NBuffers;
3736 : }
3737 : }
3738 : else
3739 : {
3740 : /*
3741 : * Initializing at startup or after LRU scanning had been off. Always
3742 : * start at the strategy point.
3743 : */
3744 : #ifdef BGW_DEBUG
3745 : elog(DEBUG2, "bgwriter initializing: strategy %u-%u",
3746 : strategy_passes, strategy_buf_id);
3747 : #endif
3748 982 : strategy_delta = 0;
3749 982 : next_to_clean = strategy_buf_id;
3750 982 : next_passes = strategy_passes;
3751 982 : bufs_to_lap = NBuffers;
3752 : }
3753 :
3754 : /* Update saved info for next time */
3755 20020 : prev_strategy_buf_id = strategy_buf_id;
3756 20020 : prev_strategy_passes = strategy_passes;
3757 20020 : saved_info_valid = true;
3758 :
3759 : /*
3760 : * Compute how many buffers had to be scanned for each new allocation, ie,
3761 : * 1/density of reusable buffers, and track a moving average of that.
3762 : *
3763 : * If the strategy point didn't move, we don't update the density estimate
3764 : */
3765 20020 : if (strategy_delta > 0 && recent_alloc > 0)
3766 : {
3767 4374 : scans_per_alloc = (float) strategy_delta / (float) recent_alloc;
3768 4374 : smoothed_density += (scans_per_alloc - smoothed_density) /
3769 : smoothing_samples;
3770 : }
3771 :
3772 : /*
3773 : * Estimate how many reusable buffers there are between the current
3774 : * strategy point and where we've scanned ahead to, based on the smoothed
3775 : * density estimate.
3776 : */
3777 20020 : bufs_ahead = NBuffers - bufs_to_lap;
3778 20020 : reusable_buffers_est = (float) bufs_ahead / smoothed_density;
3779 :
3780 : /*
3781 : * Track a moving average of recent buffer allocations. Here, rather than
3782 : * a true average we want a fast-attack, slow-decline behavior: we
3783 : * immediately follow any increase.
3784 : */
3785 20020 : if (smoothed_alloc <= (float) recent_alloc)
3786 5650 : smoothed_alloc = recent_alloc;
3787 : else
3788 14370 : smoothed_alloc += ((float) recent_alloc - smoothed_alloc) /
3789 : smoothing_samples;
3790 :
3791 : /* Scale the estimate by a GUC to allow more aggressive tuning. */
3792 20020 : upcoming_alloc_est = (int) (smoothed_alloc * bgwriter_lru_multiplier);
3793 :
3794 : /*
3795 : * If recent_alloc remains at zero for many cycles, smoothed_alloc will
3796 : * eventually underflow to zero, and the underflows produce annoying
3797 : * kernel warnings on some platforms. Once upcoming_alloc_est has gone to
3798 : * zero, there's no point in tracking smaller and smaller values of
3799 : * smoothed_alloc, so just reset it to exactly zero to avoid this
3800 : * syndrome. It will pop back up as soon as recent_alloc increases.
3801 : */
3802 20020 : if (upcoming_alloc_est == 0)
3803 3336 : smoothed_alloc = 0;
3804 :
3805 : /*
3806 : * Even in cases where there's been little or no buffer allocation
3807 : * activity, we want to make a small amount of progress through the buffer
3808 : * cache so that as many reusable buffers as possible are clean after an
3809 : * idle period.
3810 : *
3811 : * (scan_whole_pool_milliseconds / BgWriterDelay) computes how many times
3812 : * the BGW will be called during the scan_whole_pool time; slice the
3813 : * buffer pool into that many sections.
3814 : */
3815 20020 : min_scan_buffers = (int) (NBuffers / (scan_whole_pool_milliseconds / BgWriterDelay));
3816 :
3817 20020 : if (upcoming_alloc_est < (min_scan_buffers + reusable_buffers_est))
3818 : {
3819 : #ifdef BGW_DEBUG
3820 : elog(DEBUG2, "bgwriter: alloc_est=%d too small, using min=%d + reusable_est=%d",
3821 : upcoming_alloc_est, min_scan_buffers, reusable_buffers_est);
3822 : #endif
3823 10790 : upcoming_alloc_est = min_scan_buffers + reusable_buffers_est;
3824 : }
3825 :
3826 : /*
3827 : * Now write out dirty reusable buffers, working forward from the
3828 : * next_to_clean point, until we have lapped the strategy scan, or cleaned
3829 : * enough buffers to match our estimate of the next cycle's allocation
3830 : * requirements, or hit the bgwriter_lru_maxpages limit.
3831 : */
3832 :
3833 20020 : num_to_scan = bufs_to_lap;
3834 20020 : num_written = 0;
3835 20020 : reusable_buffers = reusable_buffers_est;
3836 :
3837 : /* Execute the LRU scan */
3838 3242632 : while (num_to_scan > 0 && reusable_buffers < upcoming_alloc_est)
3839 : {
3840 3222614 : int sync_state = SyncOneBuffer(next_to_clean, true,
3841 : wb_context);
3842 :
3843 3222614 : if (++next_to_clean >= NBuffers)
3844 : {
3845 3940 : next_to_clean = 0;
3846 3940 : next_passes++;
3847 : }
3848 3222614 : num_to_scan--;
3849 :
3850 3222614 : if (sync_state & BUF_WRITTEN)
3851 : {
3852 33038 : reusable_buffers++;
3853 33038 : if (++num_written >= bgwriter_lru_maxpages)
3854 : {
3855 2 : PendingBgWriterStats.maxwritten_clean++;
3856 2 : break;
3857 : }
3858 : }
3859 3189576 : else if (sync_state & BUF_REUSABLE)
3860 2406096 : reusable_buffers++;
3861 : }
3862 :
3863 20020 : PendingBgWriterStats.buf_written_clean += num_written;
3864 :
3865 : #ifdef BGW_DEBUG
3866 : 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",
3867 : recent_alloc, smoothed_alloc, strategy_delta, bufs_ahead,
3868 : smoothed_density, reusable_buffers_est, upcoming_alloc_est,
3869 : bufs_to_lap - num_to_scan,
3870 : num_written,
3871 : reusable_buffers - reusable_buffers_est);
3872 : #endif
3873 :
3874 : /*
3875 : * Consider the above scan as being like a new allocation scan.
3876 : * Characterize its density and update the smoothed one based on it. This
3877 : * effectively halves the moving average period in cases where both the
3878 : * strategy and the background writer are doing some useful scanning,
3879 : * which is helpful because a long memory isn't as desirable on the
3880 : * density estimates.
3881 : */
3882 20020 : new_strategy_delta = bufs_to_lap - num_to_scan;
3883 20020 : new_recent_alloc = reusable_buffers - reusable_buffers_est;
3884 20020 : if (new_strategy_delta > 0 && new_recent_alloc > 0)
3885 : {
3886 16446 : scans_per_alloc = (float) new_strategy_delta / (float) new_recent_alloc;
3887 16446 : smoothed_density += (scans_per_alloc - smoothed_density) /
3888 : smoothing_samples;
3889 :
3890 : #ifdef BGW_DEBUG
3891 : elog(DEBUG2, "bgwriter: cleaner density alloc=%u scan=%ld density=%.2f new smoothed=%.2f",
3892 : new_recent_alloc, new_strategy_delta,
3893 : scans_per_alloc, smoothed_density);
3894 : #endif
3895 : }
3896 :
3897 : /* Return true if OK to hibernate */
3898 20020 : return (bufs_to_lap == 0 && recent_alloc == 0);
3899 : }
3900 :
3901 : /*
3902 : * SyncOneBuffer -- process a single buffer during syncing.
3903 : *
3904 : * If skip_recently_used is true, we don't write currently-pinned buffers, nor
3905 : * buffers marked recently used, as these are not replacement candidates.
3906 : *
3907 : * Returns a bitmask containing the following flag bits:
3908 : * BUF_WRITTEN: we wrote the buffer.
3909 : * BUF_REUSABLE: buffer is available for replacement, ie, it has
3910 : * pin count 0 and usage count 0.
3911 : *
3912 : * (BUF_WRITTEN could be set in error if FlushBuffer finds the buffer clean
3913 : * after locking it, but we don't care all that much.)
3914 : */
3915 : static int
3916 3737618 : SyncOneBuffer(int buf_id, bool skip_recently_used, WritebackContext *wb_context)
3917 : {
3918 3737618 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3919 3737618 : int result = 0;
3920 : uint32 buf_state;
3921 : BufferTag tag;
3922 :
3923 : /* Make sure we can handle the pin */
3924 3737618 : ReservePrivateRefCountEntry();
3925 3737618 : ResourceOwnerEnlarge(CurrentResourceOwner);
3926 :
3927 : /*
3928 : * Check whether buffer needs writing.
3929 : *
3930 : * We can make this check without taking the buffer content lock so long
3931 : * as we mark pages dirty in access methods *before* logging changes with
3932 : * XLogInsert(): if someone marks the buffer dirty just after our check we
3933 : * don't worry because our checkpoint.redo points before log record for
3934 : * upcoming changes and so we are not required to write such dirty buffer.
3935 : */
3936 3737618 : buf_state = LockBufHdr(bufHdr);
3937 :
3938 3737618 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 0 &&
3939 3732072 : BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3940 : {
3941 2442682 : result |= BUF_REUSABLE;
3942 : }
3943 1294936 : else if (skip_recently_used)
3944 : {
3945 : /* Caller told us not to write recently-used buffers */
3946 783480 : UnlockBufHdr(bufHdr, buf_state);
3947 783480 : return result;
3948 : }
3949 :
3950 2954138 : if (!(buf_state & BM_VALID) || !(buf_state & BM_DIRTY))
3951 : {
3952 : /* It's clean, so nothing to do */
3953 2406096 : UnlockBufHdr(bufHdr, buf_state);
3954 2406096 : return result;
3955 : }
3956 :
3957 : /*
3958 : * Pin it, share-lock it, write it. (FlushBuffer will do nothing if the
3959 : * buffer is clean by the time we've locked it.)
3960 : */
3961 548042 : PinBuffer_Locked(bufHdr);
3962 548042 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_SHARED);
3963 :
3964 548042 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
3965 :
3966 548042 : LWLockRelease(BufferDescriptorGetContentLock(bufHdr));
3967 :
3968 548042 : tag = bufHdr->tag;
3969 :
3970 548042 : UnpinBuffer(bufHdr);
3971 :
3972 : /*
3973 : * SyncOneBuffer() is only called by checkpointer and bgwriter, so
3974 : * IOContext will always be IOCONTEXT_NORMAL.
3975 : */
3976 548042 : ScheduleBufferTagForWriteback(wb_context, IOCONTEXT_NORMAL, &tag);
3977 :
3978 548042 : return result | BUF_WRITTEN;
3979 : }
3980 :
3981 : /*
3982 : * AtEOXact_Buffers - clean up at end of transaction.
3983 : *
3984 : * As of PostgreSQL 8.0, buffer pins should get released by the
3985 : * ResourceOwner mechanism. This routine is just a debugging
3986 : * cross-check that no pins remain.
3987 : */
3988 : void
3989 852940 : AtEOXact_Buffers(bool isCommit)
3990 : {
3991 852940 : CheckForBufferLeaks();
3992 :
3993 852940 : AtEOXact_LocalBuffers(isCommit);
3994 :
3995 : Assert(PrivateRefCountOverflowed == 0);
3996 852940 : }
3997 :
3998 : /*
3999 : * Initialize access to shared buffer pool
4000 : *
4001 : * This is called during backend startup (whether standalone or under the
4002 : * postmaster). It sets up for this backend's access to the already-existing
4003 : * buffer pool.
4004 : */
4005 : void
4006 42770 : InitBufferManagerAccess(void)
4007 : {
4008 : HASHCTL hash_ctl;
4009 :
4010 : /*
4011 : * An advisory limit on the number of pins each backend should hold, based
4012 : * on shared_buffers and the maximum number of connections possible.
4013 : * That's very pessimistic, but outside toy-sized shared_buffers it should
4014 : * allow plenty of pins. LimitAdditionalPins() and
4015 : * GetAdditionalPinLimit() can be used to check the remaining balance.
4016 : */
4017 42770 : MaxProportionalPins = NBuffers / (MaxBackends + NUM_AUXILIARY_PROCS);
4018 :
4019 42770 : memset(&PrivateRefCountArray, 0, sizeof(PrivateRefCountArray));
4020 :
4021 42770 : hash_ctl.keysize = sizeof(int32);
4022 42770 : hash_ctl.entrysize = sizeof(PrivateRefCountEntry);
4023 :
4024 42770 : PrivateRefCountHash = hash_create("PrivateRefCount", 100, &hash_ctl,
4025 : HASH_ELEM | HASH_BLOBS);
4026 :
4027 : /*
4028 : * AtProcExit_Buffers needs LWLock access, and thereby has to be called at
4029 : * the corresponding phase of backend shutdown.
4030 : */
4031 : Assert(MyProc != NULL);
4032 42770 : on_shmem_exit(AtProcExit_Buffers, 0);
4033 42770 : }
4034 :
4035 : /*
4036 : * During backend exit, ensure that we released all shared-buffer locks and
4037 : * assert that we have no remaining pins.
4038 : */
4039 : static void
4040 42770 : AtProcExit_Buffers(int code, Datum arg)
4041 : {
4042 42770 : UnlockBuffers();
4043 :
4044 42770 : CheckForBufferLeaks();
4045 :
4046 : /* localbuf.c needs a chance too */
4047 42770 : AtProcExit_LocalBuffers();
4048 42770 : }
4049 :
4050 : /*
4051 : * CheckForBufferLeaks - ensure this backend holds no buffer pins
4052 : *
4053 : * As of PostgreSQL 8.0, buffer pins should get released by the
4054 : * ResourceOwner mechanism. This routine is just a debugging
4055 : * cross-check that no pins remain.
4056 : */
4057 : static void
4058 895710 : CheckForBufferLeaks(void)
4059 : {
4060 : #ifdef USE_ASSERT_CHECKING
4061 : int RefCountErrors = 0;
4062 : PrivateRefCountEntry *res;
4063 : int i;
4064 : char *s;
4065 :
4066 : /* check the array */
4067 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4068 : {
4069 : res = &PrivateRefCountArray[i];
4070 :
4071 : if (res->buffer != InvalidBuffer)
4072 : {
4073 : s = DebugPrintBufferRefcount(res->buffer);
4074 : elog(WARNING, "buffer refcount leak: %s", s);
4075 : pfree(s);
4076 :
4077 : RefCountErrors++;
4078 : }
4079 : }
4080 :
4081 : /* if necessary search the hash */
4082 : if (PrivateRefCountOverflowed)
4083 : {
4084 : HASH_SEQ_STATUS hstat;
4085 :
4086 : hash_seq_init(&hstat, PrivateRefCountHash);
4087 : while ((res = (PrivateRefCountEntry *) hash_seq_search(&hstat)) != NULL)
4088 : {
4089 : s = DebugPrintBufferRefcount(res->buffer);
4090 : elog(WARNING, "buffer refcount leak: %s", s);
4091 : pfree(s);
4092 : RefCountErrors++;
4093 : }
4094 : }
4095 :
4096 : Assert(RefCountErrors == 0);
4097 : #endif
4098 895710 : }
4099 :
4100 : /*
4101 : * Helper routine to issue warnings when a buffer is unexpectedly pinned
4102 : */
4103 : char *
4104 80 : DebugPrintBufferRefcount(Buffer buffer)
4105 : {
4106 : BufferDesc *buf;
4107 : int32 loccount;
4108 : char *result;
4109 : ProcNumber backend;
4110 : uint32 buf_state;
4111 :
4112 : Assert(BufferIsValid(buffer));
4113 80 : if (BufferIsLocal(buffer))
4114 : {
4115 32 : buf = GetLocalBufferDescriptor(-buffer - 1);
4116 32 : loccount = LocalRefCount[-buffer - 1];
4117 32 : backend = MyProcNumber;
4118 : }
4119 : else
4120 : {
4121 48 : buf = GetBufferDescriptor(buffer - 1);
4122 48 : loccount = GetPrivateRefCount(buffer);
4123 48 : backend = INVALID_PROC_NUMBER;
4124 : }
4125 :
4126 : /* theoretically we should lock the bufhdr here */
4127 80 : buf_state = pg_atomic_read_u32(&buf->state);
4128 :
4129 80 : result = psprintf("[%03d] (rel=%s, blockNum=%u, flags=0x%x, refcount=%u %d)",
4130 : buffer,
4131 80 : relpathbackend(BufTagGetRelFileLocator(&buf->tag), backend,
4132 : BufTagGetForkNum(&buf->tag)).str,
4133 : buf->tag.blockNum, buf_state & BUF_FLAG_MASK,
4134 : BUF_STATE_GET_REFCOUNT(buf_state), loccount);
4135 80 : return result;
4136 : }
4137 :
4138 : /*
4139 : * CheckPointBuffers
4140 : *
4141 : * Flush all dirty blocks in buffer pool to disk at checkpoint time.
4142 : *
4143 : * Note: temporary relations do not participate in checkpoints, so they don't
4144 : * need to be flushed.
4145 : */
4146 : void
4147 2642 : CheckPointBuffers(int flags)
4148 : {
4149 2642 : BufferSync(flags);
4150 2642 : }
4151 :
4152 : /*
4153 : * BufferGetBlockNumber
4154 : * Returns the block number associated with a buffer.
4155 : *
4156 : * Note:
4157 : * Assumes that the buffer is valid and pinned, else the
4158 : * value may be obsolete immediately...
4159 : */
4160 : BlockNumber
4161 97451078 : BufferGetBlockNumber(Buffer buffer)
4162 : {
4163 : BufferDesc *bufHdr;
4164 :
4165 : Assert(BufferIsPinned(buffer));
4166 :
4167 97451078 : if (BufferIsLocal(buffer))
4168 3803854 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4169 : else
4170 93647224 : bufHdr = GetBufferDescriptor(buffer - 1);
4171 :
4172 : /* pinned, so OK to read tag without spinlock */
4173 97451078 : return bufHdr->tag.blockNum;
4174 : }
4175 :
4176 : /*
4177 : * BufferGetTag
4178 : * Returns the relfilelocator, fork number and block number associated with
4179 : * a buffer.
4180 : */
4181 : void
4182 29817958 : BufferGetTag(Buffer buffer, RelFileLocator *rlocator, ForkNumber *forknum,
4183 : BlockNumber *blknum)
4184 : {
4185 : BufferDesc *bufHdr;
4186 :
4187 : /* Do the same checks as BufferGetBlockNumber. */
4188 : Assert(BufferIsPinned(buffer));
4189 :
4190 29817958 : if (BufferIsLocal(buffer))
4191 0 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4192 : else
4193 29817958 : bufHdr = GetBufferDescriptor(buffer - 1);
4194 :
4195 : /* pinned, so OK to read tag without spinlock */
4196 29817958 : *rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
4197 29817958 : *forknum = BufTagGetForkNum(&bufHdr->tag);
4198 29817958 : *blknum = bufHdr->tag.blockNum;
4199 29817958 : }
4200 :
4201 : /*
4202 : * FlushBuffer
4203 : * Physically write out a shared buffer.
4204 : *
4205 : * NOTE: this actually just passes the buffer contents to the kernel; the
4206 : * real write to disk won't happen until the kernel feels like it. This
4207 : * is okay from our point of view since we can redo the changes from WAL.
4208 : * However, we will need to force the changes to disk via fsync before
4209 : * we can checkpoint WAL.
4210 : *
4211 : * The caller must hold a pin on the buffer and have share-locked the
4212 : * buffer contents. (Note: a share-lock does not prevent updates of
4213 : * hint bits in the buffer, so the page could change while the write
4214 : * is in progress, but we assume that that will not invalidate the data
4215 : * written.)
4216 : *
4217 : * If the caller has an smgr reference for the buffer's relation, pass it
4218 : * as the second parameter. If not, pass NULL.
4219 : */
4220 : static void
4221 1053802 : FlushBuffer(BufferDesc *buf, SMgrRelation reln, IOObject io_object,
4222 : IOContext io_context)
4223 : {
4224 : XLogRecPtr recptr;
4225 : ErrorContextCallback errcallback;
4226 : instr_time io_start;
4227 : Block bufBlock;
4228 : char *bufToWrite;
4229 : uint32 buf_state;
4230 :
4231 : /*
4232 : * Try to start an I/O operation. If StartBufferIO returns false, then
4233 : * someone else flushed the buffer before we could, so we need not do
4234 : * anything.
4235 : */
4236 1053802 : if (!StartBufferIO(buf, false, false))
4237 40 : return;
4238 :
4239 : /* Setup error traceback support for ereport() */
4240 1053762 : errcallback.callback = shared_buffer_write_error_callback;
4241 1053762 : errcallback.arg = buf;
4242 1053762 : errcallback.previous = error_context_stack;
4243 1053762 : error_context_stack = &errcallback;
4244 :
4245 : /* Find smgr relation for buffer */
4246 1053762 : if (reln == NULL)
4247 1043380 : reln = smgropen(BufTagGetRelFileLocator(&buf->tag), INVALID_PROC_NUMBER);
4248 :
4249 : TRACE_POSTGRESQL_BUFFER_FLUSH_START(BufTagGetForkNum(&buf->tag),
4250 : buf->tag.blockNum,
4251 : reln->smgr_rlocator.locator.spcOid,
4252 : reln->smgr_rlocator.locator.dbOid,
4253 : reln->smgr_rlocator.locator.relNumber);
4254 :
4255 1053762 : buf_state = LockBufHdr(buf);
4256 :
4257 : /*
4258 : * Run PageGetLSN while holding header lock, since we don't have the
4259 : * buffer locked exclusively in all cases.
4260 : */
4261 1053762 : recptr = BufferGetLSN(buf);
4262 :
4263 : /* To check if block content changes while flushing. - vadim 01/17/97 */
4264 1053762 : buf_state &= ~BM_JUST_DIRTIED;
4265 1053762 : UnlockBufHdr(buf, buf_state);
4266 :
4267 : /*
4268 : * Force XLOG flush up to buffer's LSN. This implements the basic WAL
4269 : * rule that log updates must hit disk before any of the data-file changes
4270 : * they describe do.
4271 : *
4272 : * However, this rule does not apply to unlogged relations, which will be
4273 : * lost after a crash anyway. Most unlogged relation pages do not bear
4274 : * LSNs since we never emit WAL records for them, and therefore flushing
4275 : * up through the buffer LSN would be useless, but harmless. However,
4276 : * GiST indexes use LSNs internally to track page-splits, and therefore
4277 : * unlogged GiST pages bear "fake" LSNs generated by
4278 : * GetFakeLSNForUnloggedRel. It is unlikely but possible that the fake
4279 : * LSN counter could advance past the WAL insertion point; and if it did
4280 : * happen, attempting to flush WAL through that location would fail, with
4281 : * disastrous system-wide consequences. To make sure that can't happen,
4282 : * skip the flush if the buffer isn't permanent.
4283 : */
4284 1053762 : if (buf_state & BM_PERMANENT)
4285 1050200 : XLogFlush(recptr);
4286 :
4287 : /*
4288 : * Now it's safe to write the buffer to disk. Note that no one else should
4289 : * have been able to write it, while we were busy with log flushing,
4290 : * because we got the exclusive right to perform I/O by setting the
4291 : * BM_IO_IN_PROGRESS bit.
4292 : */
4293 1053762 : bufBlock = BufHdrGetBlock(buf);
4294 :
4295 : /*
4296 : * Update page checksum if desired. Since we have only shared lock on the
4297 : * buffer, other processes might be updating hint bits in it, so we must
4298 : * copy the page to private storage if we do checksumming.
4299 : */
4300 1053762 : bufToWrite = PageSetChecksumCopy((Page) bufBlock, buf->tag.blockNum);
4301 :
4302 1053762 : io_start = pgstat_prepare_io_time(track_io_timing);
4303 :
4304 : /*
4305 : * bufToWrite is either the shared buffer or a copy, as appropriate.
4306 : */
4307 1053762 : smgrwrite(reln,
4308 1053762 : BufTagGetForkNum(&buf->tag),
4309 : buf->tag.blockNum,
4310 : bufToWrite,
4311 : false);
4312 :
4313 : /*
4314 : * When a strategy is in use, only flushes of dirty buffers already in the
4315 : * strategy ring are counted as strategy writes (IOCONTEXT
4316 : * [BULKREAD|BULKWRITE|VACUUM] IOOP_WRITE) for the purpose of IO
4317 : * statistics tracking.
4318 : *
4319 : * If a shared buffer initially added to the ring must be flushed before
4320 : * being used, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE.
4321 : *
4322 : * If a shared buffer which was added to the ring later because the
4323 : * current strategy buffer is pinned or in use or because all strategy
4324 : * buffers were dirty and rejected (for BAS_BULKREAD operations only)
4325 : * requires flushing, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE
4326 : * (from_ring will be false).
4327 : *
4328 : * When a strategy is not in use, the write can only be a "regular" write
4329 : * of a dirty shared buffer (IOCONTEXT_NORMAL IOOP_WRITE).
4330 : */
4331 1053762 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
4332 : IOOP_WRITE, io_start, 1, BLCKSZ);
4333 :
4334 1053762 : pgBufferUsage.shared_blks_written++;
4335 :
4336 : /*
4337 : * Mark the buffer as clean (unless BM_JUST_DIRTIED has become set) and
4338 : * end the BM_IO_IN_PROGRESS state.
4339 : */
4340 1053762 : TerminateBufferIO(buf, true, 0, true, false);
4341 :
4342 : TRACE_POSTGRESQL_BUFFER_FLUSH_DONE(BufTagGetForkNum(&buf->tag),
4343 : buf->tag.blockNum,
4344 : reln->smgr_rlocator.locator.spcOid,
4345 : reln->smgr_rlocator.locator.dbOid,
4346 : reln->smgr_rlocator.locator.relNumber);
4347 :
4348 : /* Pop the error context stack */
4349 1053762 : error_context_stack = errcallback.previous;
4350 : }
4351 :
4352 : /*
4353 : * RelationGetNumberOfBlocksInFork
4354 : * Determines the current number of pages in the specified relation fork.
4355 : *
4356 : * Note that the accuracy of the result will depend on the details of the
4357 : * relation's storage. For builtin AMs it'll be accurate, but for external AMs
4358 : * it might not be.
4359 : */
4360 : BlockNumber
4361 3502288 : RelationGetNumberOfBlocksInFork(Relation relation, ForkNumber forkNum)
4362 : {
4363 3502288 : if (RELKIND_HAS_TABLE_AM(relation->rd_rel->relkind))
4364 : {
4365 : /*
4366 : * Not every table AM uses BLCKSZ wide fixed size blocks. Therefore
4367 : * tableam returns the size in bytes - but for the purpose of this
4368 : * routine, we want the number of blocks. Therefore divide, rounding
4369 : * up.
4370 : */
4371 : uint64 szbytes;
4372 :
4373 2648804 : szbytes = table_relation_size(relation, forkNum);
4374 :
4375 2648766 : return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
4376 : }
4377 853484 : else if (RELKIND_HAS_STORAGE(relation->rd_rel->relkind))
4378 : {
4379 853484 : return smgrnblocks(RelationGetSmgr(relation), forkNum);
4380 : }
4381 : else
4382 : Assert(false);
4383 :
4384 0 : return 0; /* keep compiler quiet */
4385 : }
4386 :
4387 : /*
4388 : * BufferIsPermanent
4389 : * Determines whether a buffer will potentially still be around after
4390 : * a crash. Caller must hold a buffer pin.
4391 : */
4392 : bool
4393 19742234 : BufferIsPermanent(Buffer buffer)
4394 : {
4395 : BufferDesc *bufHdr;
4396 :
4397 : /* Local buffers are used only for temp relations. */
4398 19742234 : if (BufferIsLocal(buffer))
4399 1248026 : return false;
4400 :
4401 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4402 : Assert(BufferIsValid(buffer));
4403 : Assert(BufferIsPinned(buffer));
4404 :
4405 : /*
4406 : * BM_PERMANENT can't be changed while we hold a pin on the buffer, so we
4407 : * need not bother with the buffer header spinlock. Even if someone else
4408 : * changes the buffer header state while we're doing this, the state is
4409 : * changed atomically, so we'll read the old value or the new value, but
4410 : * not random garbage.
4411 : */
4412 18494208 : bufHdr = GetBufferDescriptor(buffer - 1);
4413 18494208 : return (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT) != 0;
4414 : }
4415 :
4416 : /*
4417 : * BufferGetLSNAtomic
4418 : * Retrieves the LSN of the buffer atomically using a buffer header lock.
4419 : * This is necessary for some callers who may not have an exclusive lock
4420 : * on the buffer.
4421 : */
4422 : XLogRecPtr
4423 17411662 : BufferGetLSNAtomic(Buffer buffer)
4424 : {
4425 17411662 : char *page = BufferGetPage(buffer);
4426 : BufferDesc *bufHdr;
4427 : XLogRecPtr lsn;
4428 : uint32 buf_state;
4429 :
4430 : /*
4431 : * If we don't need locking for correctness, fastpath out.
4432 : */
4433 17411662 : if (!XLogHintBitIsNeeded() || BufferIsLocal(buffer))
4434 571974 : return PageGetLSN(page);
4435 :
4436 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4437 : Assert(BufferIsValid(buffer));
4438 : Assert(BufferIsPinned(buffer));
4439 :
4440 16839688 : bufHdr = GetBufferDescriptor(buffer - 1);
4441 16839688 : buf_state = LockBufHdr(bufHdr);
4442 16839688 : lsn = PageGetLSN(page);
4443 16839688 : UnlockBufHdr(bufHdr, buf_state);
4444 :
4445 16839688 : return lsn;
4446 : }
4447 :
4448 : /* ---------------------------------------------------------------------
4449 : * DropRelationBuffers
4450 : *
4451 : * This function removes from the buffer pool all the pages of the
4452 : * specified relation forks that have block numbers >= firstDelBlock.
4453 : * (In particular, with firstDelBlock = 0, all pages are removed.)
4454 : * Dirty pages are simply dropped, without bothering to write them
4455 : * out first. Therefore, this is NOT rollback-able, and so should be
4456 : * used only with extreme caution!
4457 : *
4458 : * Currently, this is called only from smgr.c when the underlying file
4459 : * is about to be deleted or truncated (firstDelBlock is needed for
4460 : * the truncation case). The data in the affected pages would therefore
4461 : * be deleted momentarily anyway, and there is no point in writing it.
4462 : * It is the responsibility of higher-level code to ensure that the
4463 : * deletion or truncation does not lose any data that could be needed
4464 : * later. It is also the responsibility of higher-level code to ensure
4465 : * that no other process could be trying to load more pages of the
4466 : * relation into buffers.
4467 : * --------------------------------------------------------------------
4468 : */
4469 : void
4470 1228 : DropRelationBuffers(SMgrRelation smgr_reln, ForkNumber *forkNum,
4471 : int nforks, BlockNumber *firstDelBlock)
4472 : {
4473 : int i;
4474 : int j;
4475 : RelFileLocatorBackend rlocator;
4476 : BlockNumber nForkBlock[MAX_FORKNUM];
4477 1228 : uint64 nBlocksToInvalidate = 0;
4478 :
4479 1228 : rlocator = smgr_reln->smgr_rlocator;
4480 :
4481 : /* If it's a local relation, it's localbuf.c's problem. */
4482 1228 : if (RelFileLocatorBackendIsTemp(rlocator))
4483 : {
4484 664 : if (rlocator.backend == MyProcNumber)
4485 : {
4486 1374 : for (j = 0; j < nforks; j++)
4487 710 : DropRelationLocalBuffers(rlocator.locator, forkNum[j],
4488 710 : firstDelBlock[j]);
4489 : }
4490 752 : return;
4491 : }
4492 :
4493 : /*
4494 : * To remove all the pages of the specified relation forks from the buffer
4495 : * pool, we need to scan the entire buffer pool but we can optimize it by
4496 : * finding the buffers from BufMapping table provided we know the exact
4497 : * size of each fork of the relation. The exact size is required to ensure
4498 : * that we don't leave any buffer for the relation being dropped as
4499 : * otherwise the background writer or checkpointer can lead to a PANIC
4500 : * error while flushing buffers corresponding to files that don't exist.
4501 : *
4502 : * To know the exact size, we rely on the size cached for each fork by us
4503 : * during recovery which limits the optimization to recovery and on
4504 : * standbys but we can easily extend it once we have shared cache for
4505 : * relation size.
4506 : *
4507 : * In recovery, we cache the value returned by the first lseek(SEEK_END)
4508 : * and the future writes keeps the cached value up-to-date. See
4509 : * smgrextend. It is possible that the value of the first lseek is smaller
4510 : * than the actual number of existing blocks in the file due to buggy
4511 : * Linux kernels that might not have accounted for the recent write. But
4512 : * that should be fine because there must not be any buffers after that
4513 : * file size.
4514 : */
4515 780 : for (i = 0; i < nforks; i++)
4516 : {
4517 : /* Get the number of blocks for a relation's fork */
4518 668 : nForkBlock[i] = smgrnblocks_cached(smgr_reln, forkNum[i]);
4519 :
4520 668 : if (nForkBlock[i] == InvalidBlockNumber)
4521 : {
4522 452 : nBlocksToInvalidate = InvalidBlockNumber;
4523 452 : break;
4524 : }
4525 :
4526 : /* calculate the number of blocks to be invalidated */
4527 216 : nBlocksToInvalidate += (nForkBlock[i] - firstDelBlock[i]);
4528 : }
4529 :
4530 : /*
4531 : * We apply the optimization iff the total number of blocks to invalidate
4532 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4533 : */
4534 564 : if (BlockNumberIsValid(nBlocksToInvalidate) &&
4535 112 : nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4536 : {
4537 250 : for (j = 0; j < nforks; j++)
4538 162 : FindAndDropRelationBuffers(rlocator.locator, forkNum[j],
4539 162 : nForkBlock[j], firstDelBlock[j]);
4540 88 : return;
4541 : }
4542 :
4543 6206172 : for (i = 0; i < NBuffers; i++)
4544 : {
4545 6205696 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4546 : uint32 buf_state;
4547 :
4548 : /*
4549 : * We can make this a tad faster by prechecking the buffer tag before
4550 : * we attempt to lock the buffer; this saves a lot of lock
4551 : * acquisitions in typical cases. It should be safe because the
4552 : * caller must have AccessExclusiveLock on the relation, or some other
4553 : * reason to be certain that no one is loading new pages of the rel
4554 : * into the buffer pool. (Otherwise we might well miss such pages
4555 : * entirely.) Therefore, while the tag might be changing while we
4556 : * look at it, it can't be changing *to* a value we care about, only
4557 : * *away* from such a value. So false negatives are impossible, and
4558 : * false positives are safe because we'll recheck after getting the
4559 : * buffer lock.
4560 : *
4561 : * We could check forkNum and blockNum as well as the rlocator, but
4562 : * the incremental win from doing so seems small.
4563 : */
4564 6205696 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator))
4565 6189606 : continue;
4566 :
4567 16090 : buf_state = LockBufHdr(bufHdr);
4568 :
4569 40780 : for (j = 0; j < nforks; j++)
4570 : {
4571 28724 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator) &&
4572 28724 : BufTagGetForkNum(&bufHdr->tag) == forkNum[j] &&
4573 15872 : bufHdr->tag.blockNum >= firstDelBlock[j])
4574 : {
4575 4034 : InvalidateBuffer(bufHdr); /* releases spinlock */
4576 4034 : break;
4577 : }
4578 : }
4579 16090 : if (j >= nforks)
4580 12056 : UnlockBufHdr(bufHdr, buf_state);
4581 : }
4582 : }
4583 :
4584 : /* ---------------------------------------------------------------------
4585 : * DropRelationsAllBuffers
4586 : *
4587 : * This function removes from the buffer pool all the pages of all
4588 : * forks of the specified relations. It's equivalent to calling
4589 : * DropRelationBuffers once per fork per relation with firstDelBlock = 0.
4590 : * --------------------------------------------------------------------
4591 : */
4592 : void
4593 27638 : DropRelationsAllBuffers(SMgrRelation *smgr_reln, int nlocators)
4594 : {
4595 : int i;
4596 27638 : int n = 0;
4597 : SMgrRelation *rels;
4598 : BlockNumber (*block)[MAX_FORKNUM + 1];
4599 27638 : uint64 nBlocksToInvalidate = 0;
4600 : RelFileLocator *locators;
4601 27638 : bool cached = true;
4602 : bool use_bsearch;
4603 :
4604 27638 : if (nlocators == 0)
4605 0 : return;
4606 :
4607 27638 : rels = palloc(sizeof(SMgrRelation) * nlocators); /* non-local relations */
4608 :
4609 : /* If it's a local relation, it's localbuf.c's problem. */
4610 119610 : for (i = 0; i < nlocators; i++)
4611 : {
4612 91972 : if (RelFileLocatorBackendIsTemp(smgr_reln[i]->smgr_rlocator))
4613 : {
4614 6168 : if (smgr_reln[i]->smgr_rlocator.backend == MyProcNumber)
4615 6168 : DropRelationAllLocalBuffers(smgr_reln[i]->smgr_rlocator.locator);
4616 : }
4617 : else
4618 85804 : rels[n++] = smgr_reln[i];
4619 : }
4620 :
4621 : /*
4622 : * If there are no non-local relations, then we're done. Release the
4623 : * memory and return.
4624 : */
4625 27638 : if (n == 0)
4626 : {
4627 1564 : pfree(rels);
4628 1564 : return;
4629 : }
4630 :
4631 : /*
4632 : * This is used to remember the number of blocks for all the relations
4633 : * forks.
4634 : */
4635 : block = (BlockNumber (*)[MAX_FORKNUM + 1])
4636 26074 : palloc(sizeof(BlockNumber) * n * (MAX_FORKNUM + 1));
4637 :
4638 : /*
4639 : * We can avoid scanning the entire buffer pool if we know the exact size
4640 : * of each of the given relation forks. See DropRelationBuffers.
4641 : */
4642 54794 : for (i = 0; i < n && cached; i++)
4643 : {
4644 45970 : for (int j = 0; j <= MAX_FORKNUM; j++)
4645 : {
4646 : /* Get the number of blocks for a relation's fork. */
4647 41680 : block[i][j] = smgrnblocks_cached(rels[i], j);
4648 :
4649 : /* We need to only consider the relation forks that exists. */
4650 41680 : if (block[i][j] == InvalidBlockNumber)
4651 : {
4652 37006 : if (!smgrexists(rels[i], j))
4653 12576 : continue;
4654 24430 : cached = false;
4655 24430 : break;
4656 : }
4657 :
4658 : /* calculate the total number of blocks to be invalidated */
4659 4674 : nBlocksToInvalidate += block[i][j];
4660 : }
4661 : }
4662 :
4663 : /*
4664 : * We apply the optimization iff the total number of blocks to invalidate
4665 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4666 : */
4667 26074 : if (cached && nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4668 : {
4669 2710 : for (i = 0; i < n; i++)
4670 : {
4671 7450 : for (int j = 0; j <= MAX_FORKNUM; j++)
4672 : {
4673 : /* ignore relation forks that doesn't exist */
4674 5960 : if (!BlockNumberIsValid(block[i][j]))
4675 4448 : continue;
4676 :
4677 : /* drop all the buffers for a particular relation fork */
4678 1512 : FindAndDropRelationBuffers(rels[i]->smgr_rlocator.locator,
4679 1512 : j, block[i][j], 0);
4680 : }
4681 : }
4682 :
4683 1220 : pfree(block);
4684 1220 : pfree(rels);
4685 1220 : return;
4686 : }
4687 :
4688 24854 : pfree(block);
4689 24854 : locators = palloc(sizeof(RelFileLocator) * n); /* non-local relations */
4690 109168 : for (i = 0; i < n; i++)
4691 84314 : locators[i] = rels[i]->smgr_rlocator.locator;
4692 :
4693 : /*
4694 : * For low number of relations to drop just use a simple walk through, to
4695 : * save the bsearch overhead. The threshold to use is rather a guess than
4696 : * an exactly determined value, as it depends on many factors (CPU and RAM
4697 : * speeds, amount of shared buffers etc.).
4698 : */
4699 24854 : use_bsearch = n > RELS_BSEARCH_THRESHOLD;
4700 :
4701 : /* sort the list of rlocators if necessary */
4702 24854 : if (use_bsearch)
4703 334 : qsort(locators, n, sizeof(RelFileLocator), rlocator_comparator);
4704 :
4705 277119766 : for (i = 0; i < NBuffers; i++)
4706 : {
4707 277094912 : RelFileLocator *rlocator = NULL;
4708 277094912 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4709 : uint32 buf_state;
4710 :
4711 : /*
4712 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4713 : * saves some cycles.
4714 : */
4715 :
4716 277094912 : if (!use_bsearch)
4717 : {
4718 : int j;
4719 :
4720 1084268978 : for (j = 0; j < n; j++)
4721 : {
4722 810871850 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &locators[j]))
4723 : {
4724 176248 : rlocator = &locators[j];
4725 176248 : break;
4726 : }
4727 : }
4728 : }
4729 : else
4730 : {
4731 : RelFileLocator locator;
4732 :
4733 3521536 : locator = BufTagGetRelFileLocator(&bufHdr->tag);
4734 3521536 : rlocator = bsearch(&locator,
4735 : locators, n, sizeof(RelFileLocator),
4736 : rlocator_comparator);
4737 : }
4738 :
4739 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
4740 277094912 : if (rlocator == NULL)
4741 276915208 : continue;
4742 :
4743 179704 : buf_state = LockBufHdr(bufHdr);
4744 179704 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, rlocator))
4745 179704 : InvalidateBuffer(bufHdr); /* releases spinlock */
4746 : else
4747 0 : UnlockBufHdr(bufHdr, buf_state);
4748 : }
4749 :
4750 24854 : pfree(locators);
4751 24854 : pfree(rels);
4752 : }
4753 :
4754 : /* ---------------------------------------------------------------------
4755 : * FindAndDropRelationBuffers
4756 : *
4757 : * This function performs look up in BufMapping table and removes from the
4758 : * buffer pool all the pages of the specified relation fork that has block
4759 : * number >= firstDelBlock. (In particular, with firstDelBlock = 0, all
4760 : * pages are removed.)
4761 : * --------------------------------------------------------------------
4762 : */
4763 : static void
4764 1674 : FindAndDropRelationBuffers(RelFileLocator rlocator, ForkNumber forkNum,
4765 : BlockNumber nForkBlock,
4766 : BlockNumber firstDelBlock)
4767 : {
4768 : BlockNumber curBlock;
4769 :
4770 4026 : for (curBlock = firstDelBlock; curBlock < nForkBlock; curBlock++)
4771 : {
4772 : uint32 bufHash; /* hash value for tag */
4773 : BufferTag bufTag; /* identity of requested block */
4774 : LWLock *bufPartitionLock; /* buffer partition lock for it */
4775 : int buf_id;
4776 : BufferDesc *bufHdr;
4777 : uint32 buf_state;
4778 :
4779 : /* create a tag so we can lookup the buffer */
4780 2352 : InitBufferTag(&bufTag, &rlocator, forkNum, curBlock);
4781 :
4782 : /* determine its hash code and partition lock ID */
4783 2352 : bufHash = BufTableHashCode(&bufTag);
4784 2352 : bufPartitionLock = BufMappingPartitionLock(bufHash);
4785 :
4786 : /* Check that it is in the buffer pool. If not, do nothing. */
4787 2352 : LWLockAcquire(bufPartitionLock, LW_SHARED);
4788 2352 : buf_id = BufTableLookup(&bufTag, bufHash);
4789 2352 : LWLockRelease(bufPartitionLock);
4790 :
4791 2352 : if (buf_id < 0)
4792 242 : continue;
4793 :
4794 2110 : bufHdr = GetBufferDescriptor(buf_id);
4795 :
4796 : /*
4797 : * We need to lock the buffer header and recheck if the buffer is
4798 : * still associated with the same block because the buffer could be
4799 : * evicted by some other backend loading blocks for a different
4800 : * relation after we release lock on the BufMapping table.
4801 : */
4802 2110 : buf_state = LockBufHdr(bufHdr);
4803 :
4804 4220 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator) &&
4805 2110 : BufTagGetForkNum(&bufHdr->tag) == forkNum &&
4806 2110 : bufHdr->tag.blockNum >= firstDelBlock)
4807 2110 : InvalidateBuffer(bufHdr); /* releases spinlock */
4808 : else
4809 0 : UnlockBufHdr(bufHdr, buf_state);
4810 : }
4811 1674 : }
4812 :
4813 : /* ---------------------------------------------------------------------
4814 : * DropDatabaseBuffers
4815 : *
4816 : * This function removes all the buffers in the buffer cache for a
4817 : * particular database. Dirty pages are simply dropped, without
4818 : * bothering to write them out first. This is used when we destroy a
4819 : * database, to avoid trying to flush data to disk when the directory
4820 : * tree no longer exists. Implementation is pretty similar to
4821 : * DropRelationBuffers() which is for destroying just one relation.
4822 : * --------------------------------------------------------------------
4823 : */
4824 : void
4825 136 : DropDatabaseBuffers(Oid dbid)
4826 : {
4827 : int i;
4828 :
4829 : /*
4830 : * We needn't consider local buffers, since by assumption the target
4831 : * database isn't our own.
4832 : */
4833 :
4834 992904 : for (i = 0; i < NBuffers; i++)
4835 : {
4836 992768 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4837 : uint32 buf_state;
4838 :
4839 : /*
4840 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4841 : * saves some cycles.
4842 : */
4843 992768 : if (bufHdr->tag.dbOid != dbid)
4844 965300 : continue;
4845 :
4846 27468 : buf_state = LockBufHdr(bufHdr);
4847 27468 : if (bufHdr->tag.dbOid == dbid)
4848 27468 : InvalidateBuffer(bufHdr); /* releases spinlock */
4849 : else
4850 0 : UnlockBufHdr(bufHdr, buf_state);
4851 : }
4852 136 : }
4853 :
4854 : /* ---------------------------------------------------------------------
4855 : * FlushRelationBuffers
4856 : *
4857 : * This function writes all dirty pages of a relation out to disk
4858 : * (or more accurately, out to kernel disk buffers), ensuring that the
4859 : * kernel has an up-to-date view of the relation.
4860 : *
4861 : * Generally, the caller should be holding AccessExclusiveLock on the
4862 : * target relation to ensure that no other backend is busy dirtying
4863 : * more blocks of the relation; the effects can't be expected to last
4864 : * after the lock is released.
4865 : *
4866 : * XXX currently it sequentially searches the buffer pool, should be
4867 : * changed to more clever ways of searching. This routine is not
4868 : * used in any performance-critical code paths, so it's not worth
4869 : * adding additional overhead to normal paths to make it go faster.
4870 : * --------------------------------------------------------------------
4871 : */
4872 : void
4873 288 : FlushRelationBuffers(Relation rel)
4874 : {
4875 : int i;
4876 : BufferDesc *bufHdr;
4877 288 : SMgrRelation srel = RelationGetSmgr(rel);
4878 :
4879 288 : if (RelationUsesLocalBuffers(rel))
4880 : {
4881 1818 : for (i = 0; i < NLocBuffer; i++)
4882 : {
4883 : uint32 buf_state;
4884 :
4885 1800 : bufHdr = GetLocalBufferDescriptor(i);
4886 1800 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
4887 600 : ((buf_state = pg_atomic_read_u32(&bufHdr->state)) &
4888 : (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
4889 : {
4890 : ErrorContextCallback errcallback;
4891 :
4892 : /* Setup error traceback support for ereport() */
4893 600 : errcallback.callback = local_buffer_write_error_callback;
4894 600 : errcallback.arg = bufHdr;
4895 600 : errcallback.previous = error_context_stack;
4896 600 : error_context_stack = &errcallback;
4897 :
4898 600 : FlushLocalBuffer(bufHdr, srel);
4899 :
4900 : /* Pop the error context stack */
4901 600 : error_context_stack = errcallback.previous;
4902 : }
4903 : }
4904 :
4905 18 : return;
4906 : }
4907 :
4908 3221006 : for (i = 0; i < NBuffers; i++)
4909 : {
4910 : uint32 buf_state;
4911 :
4912 3220736 : bufHdr = GetBufferDescriptor(i);
4913 :
4914 : /*
4915 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4916 : * saves some cycles.
4917 : */
4918 3220736 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator))
4919 3220268 : continue;
4920 :
4921 : /* Make sure we can handle the pin */
4922 468 : ReservePrivateRefCountEntry();
4923 468 : ResourceOwnerEnlarge(CurrentResourceOwner);
4924 :
4925 468 : buf_state = LockBufHdr(bufHdr);
4926 468 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
4927 468 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
4928 : {
4929 382 : PinBuffer_Locked(bufHdr);
4930 382 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_SHARED);
4931 382 : FlushBuffer(bufHdr, srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
4932 382 : LWLockRelease(BufferDescriptorGetContentLock(bufHdr));
4933 382 : UnpinBuffer(bufHdr);
4934 : }
4935 : else
4936 86 : UnlockBufHdr(bufHdr, buf_state);
4937 : }
4938 : }
4939 :
4940 : /* ---------------------------------------------------------------------
4941 : * FlushRelationsAllBuffers
4942 : *
4943 : * This function flushes out of the buffer pool all the pages of all
4944 : * forks of the specified smgr relations. It's equivalent to calling
4945 : * FlushRelationBuffers once per relation. The relations are assumed not
4946 : * to use local buffers.
4947 : * --------------------------------------------------------------------
4948 : */
4949 : void
4950 38 : FlushRelationsAllBuffers(SMgrRelation *smgrs, int nrels)
4951 : {
4952 : int i;
4953 : SMgrSortArray *srels;
4954 : bool use_bsearch;
4955 :
4956 38 : if (nrels == 0)
4957 0 : return;
4958 :
4959 : /* fill-in array for qsort */
4960 38 : srels = palloc(sizeof(SMgrSortArray) * nrels);
4961 :
4962 86 : for (i = 0; i < nrels; i++)
4963 : {
4964 : Assert(!RelFileLocatorBackendIsTemp(smgrs[i]->smgr_rlocator));
4965 :
4966 48 : srels[i].rlocator = smgrs[i]->smgr_rlocator.locator;
4967 48 : srels[i].srel = smgrs[i];
4968 : }
4969 :
4970 : /*
4971 : * Save the bsearch overhead for low number of relations to sync. See
4972 : * DropRelationsAllBuffers for details.
4973 : */
4974 38 : use_bsearch = nrels > RELS_BSEARCH_THRESHOLD;
4975 :
4976 : /* sort the list of SMgrRelations if necessary */
4977 38 : if (use_bsearch)
4978 0 : qsort(srels, nrels, sizeof(SMgrSortArray), rlocator_comparator);
4979 :
4980 622630 : for (i = 0; i < NBuffers; i++)
4981 : {
4982 622592 : SMgrSortArray *srelent = NULL;
4983 622592 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4984 : uint32 buf_state;
4985 :
4986 : /*
4987 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4988 : * saves some cycles.
4989 : */
4990 :
4991 622592 : if (!use_bsearch)
4992 : {
4993 : int j;
4994 :
4995 1397834 : for (j = 0; j < nrels; j++)
4996 : {
4997 786424 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srels[j].rlocator))
4998 : {
4999 11182 : srelent = &srels[j];
5000 11182 : break;
5001 : }
5002 : }
5003 : }
5004 : else
5005 : {
5006 : RelFileLocator rlocator;
5007 :
5008 0 : rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
5009 0 : srelent = bsearch(&rlocator,
5010 : srels, nrels, sizeof(SMgrSortArray),
5011 : rlocator_comparator);
5012 : }
5013 :
5014 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
5015 622592 : if (srelent == NULL)
5016 611410 : continue;
5017 :
5018 : /* Make sure we can handle the pin */
5019 11182 : ReservePrivateRefCountEntry();
5020 11182 : ResourceOwnerEnlarge(CurrentResourceOwner);
5021 :
5022 11182 : buf_state = LockBufHdr(bufHdr);
5023 11182 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srelent->rlocator) &&
5024 11182 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5025 : {
5026 10000 : PinBuffer_Locked(bufHdr);
5027 10000 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_SHARED);
5028 10000 : FlushBuffer(bufHdr, srelent->srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5029 10000 : LWLockRelease(BufferDescriptorGetContentLock(bufHdr));
5030 10000 : UnpinBuffer(bufHdr);
5031 : }
5032 : else
5033 1182 : UnlockBufHdr(bufHdr, buf_state);
5034 : }
5035 :
5036 38 : pfree(srels);
5037 : }
5038 :
5039 : /* ---------------------------------------------------------------------
5040 : * RelationCopyStorageUsingBuffer
5041 : *
5042 : * Copy fork's data using bufmgr. Same as RelationCopyStorage but instead
5043 : * of using smgrread and smgrextend this will copy using bufmgr APIs.
5044 : *
5045 : * Refer comments atop CreateAndCopyRelationData() for details about
5046 : * 'permanent' parameter.
5047 : * --------------------------------------------------------------------
5048 : */
5049 : static void
5050 140704 : RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
5051 : RelFileLocator dstlocator,
5052 : ForkNumber forkNum, bool permanent)
5053 : {
5054 : Buffer srcBuf;
5055 : Buffer dstBuf;
5056 : Page srcPage;
5057 : Page dstPage;
5058 : bool use_wal;
5059 : BlockNumber nblocks;
5060 : BlockNumber blkno;
5061 : PGIOAlignedBlock buf;
5062 : BufferAccessStrategy bstrategy_src;
5063 : BufferAccessStrategy bstrategy_dst;
5064 : BlockRangeReadStreamPrivate p;
5065 : ReadStream *src_stream;
5066 : SMgrRelation src_smgr;
5067 :
5068 : /*
5069 : * In general, we want to write WAL whenever wal_level > 'minimal', but we
5070 : * can skip it when copying any fork of an unlogged relation other than
5071 : * the init fork.
5072 : */
5073 140704 : use_wal = XLogIsNeeded() && (permanent || forkNum == INIT_FORKNUM);
5074 :
5075 : /* Get number of blocks in the source relation. */
5076 140704 : nblocks = smgrnblocks(smgropen(srclocator, INVALID_PROC_NUMBER),
5077 : forkNum);
5078 :
5079 : /* Nothing to copy; just return. */
5080 140704 : if (nblocks == 0)
5081 24542 : return;
5082 :
5083 : /*
5084 : * Bulk extend the destination relation of the same size as the source
5085 : * relation before starting to copy block by block.
5086 : */
5087 116162 : memset(buf.data, 0, BLCKSZ);
5088 116162 : smgrextend(smgropen(dstlocator, INVALID_PROC_NUMBER), forkNum, nblocks - 1,
5089 : buf.data, true);
5090 :
5091 : /* This is a bulk operation, so use buffer access strategies. */
5092 116162 : bstrategy_src = GetAccessStrategy(BAS_BULKREAD);
5093 116162 : bstrategy_dst = GetAccessStrategy(BAS_BULKWRITE);
5094 :
5095 : /* Initialize streaming read */
5096 116162 : p.current_blocknum = 0;
5097 116162 : p.last_exclusive = nblocks;
5098 116162 : src_smgr = smgropen(srclocator, INVALID_PROC_NUMBER);
5099 :
5100 : /*
5101 : * It is safe to use batchmode as block_range_read_stream_cb takes no
5102 : * locks.
5103 : */
5104 116162 : src_stream = read_stream_begin_smgr_relation(READ_STREAM_FULL |
5105 : READ_STREAM_USE_BATCHING,
5106 : bstrategy_src,
5107 : src_smgr,
5108 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
5109 : forkNum,
5110 : block_range_read_stream_cb,
5111 : &p,
5112 : 0);
5113 :
5114 : /* Iterate over each block of the source relation file. */
5115 558186 : for (blkno = 0; blkno < nblocks; blkno++)
5116 : {
5117 442028 : CHECK_FOR_INTERRUPTS();
5118 :
5119 : /* Read block from source relation. */
5120 442028 : srcBuf = read_stream_next_buffer(src_stream, NULL);
5121 442024 : LockBuffer(srcBuf, BUFFER_LOCK_SHARE);
5122 442024 : srcPage = BufferGetPage(srcBuf);
5123 :
5124 442024 : dstBuf = ReadBufferWithoutRelcache(dstlocator, forkNum,
5125 : BufferGetBlockNumber(srcBuf),
5126 : RBM_ZERO_AND_LOCK, bstrategy_dst,
5127 : permanent);
5128 442024 : dstPage = BufferGetPage(dstBuf);
5129 :
5130 442024 : START_CRIT_SECTION();
5131 :
5132 : /* Copy page data from the source to the destination. */
5133 442024 : memcpy(dstPage, srcPage, BLCKSZ);
5134 442024 : MarkBufferDirty(dstBuf);
5135 :
5136 : /* WAL-log the copied page. */
5137 442024 : if (use_wal)
5138 241672 : log_newpage_buffer(dstBuf, true);
5139 :
5140 442024 : END_CRIT_SECTION();
5141 :
5142 442024 : UnlockReleaseBuffer(dstBuf);
5143 442024 : UnlockReleaseBuffer(srcBuf);
5144 : }
5145 : Assert(read_stream_next_buffer(src_stream, NULL) == InvalidBuffer);
5146 116158 : read_stream_end(src_stream);
5147 :
5148 116158 : FreeAccessStrategy(bstrategy_src);
5149 116158 : FreeAccessStrategy(bstrategy_dst);
5150 : }
5151 :
5152 : /* ---------------------------------------------------------------------
5153 : * CreateAndCopyRelationData
5154 : *
5155 : * Create destination relation storage and copy all forks from the
5156 : * source relation to the destination.
5157 : *
5158 : * Pass permanent as true for permanent relations and false for
5159 : * unlogged relations. Currently this API is not supported for
5160 : * temporary relations.
5161 : * --------------------------------------------------------------------
5162 : */
5163 : void
5164 105768 : CreateAndCopyRelationData(RelFileLocator src_rlocator,
5165 : RelFileLocator dst_rlocator, bool permanent)
5166 : {
5167 : char relpersistence;
5168 : SMgrRelation src_rel;
5169 : SMgrRelation dst_rel;
5170 :
5171 : /* Set the relpersistence. */
5172 105768 : relpersistence = permanent ?
5173 : RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED;
5174 :
5175 105768 : src_rel = smgropen(src_rlocator, INVALID_PROC_NUMBER);
5176 105768 : dst_rel = smgropen(dst_rlocator, INVALID_PROC_NUMBER);
5177 :
5178 : /*
5179 : * Create and copy all forks of the relation. During create database we
5180 : * have a separate cleanup mechanism which deletes complete database
5181 : * directory. Therefore, each individual relation doesn't need to be
5182 : * registered for cleanup.
5183 : */
5184 105768 : RelationCreateStorage(dst_rlocator, relpersistence, false);
5185 :
5186 : /* copy main fork. */
5187 105768 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, MAIN_FORKNUM,
5188 : permanent);
5189 :
5190 : /* copy those extra forks that exist */
5191 423056 : for (ForkNumber forkNum = MAIN_FORKNUM + 1;
5192 317292 : forkNum <= MAX_FORKNUM; forkNum++)
5193 : {
5194 317292 : if (smgrexists(src_rel, forkNum))
5195 : {
5196 34936 : smgrcreate(dst_rel, forkNum, false);
5197 :
5198 : /*
5199 : * WAL log creation if the relation is persistent, or this is the
5200 : * init fork of an unlogged relation.
5201 : */
5202 34936 : if (permanent || forkNum == INIT_FORKNUM)
5203 34936 : log_smgrcreate(&dst_rlocator, forkNum);
5204 :
5205 : /* Copy a fork's data, block by block. */
5206 34936 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, forkNum,
5207 : permanent);
5208 : }
5209 : }
5210 105764 : }
5211 :
5212 : /* ---------------------------------------------------------------------
5213 : * FlushDatabaseBuffers
5214 : *
5215 : * This function writes all dirty pages of a database out to disk
5216 : * (or more accurately, out to kernel disk buffers), ensuring that the
5217 : * kernel has an up-to-date view of the database.
5218 : *
5219 : * Generally, the caller should be holding an appropriate lock to ensure
5220 : * no other backend is active in the target database; otherwise more
5221 : * pages could get dirtied.
5222 : *
5223 : * Note we don't worry about flushing any pages of temporary relations.
5224 : * It's assumed these wouldn't be interesting.
5225 : * --------------------------------------------------------------------
5226 : */
5227 : void
5228 8 : FlushDatabaseBuffers(Oid dbid)
5229 : {
5230 : int i;
5231 : BufferDesc *bufHdr;
5232 :
5233 1032 : for (i = 0; i < NBuffers; i++)
5234 : {
5235 : uint32 buf_state;
5236 :
5237 1024 : bufHdr = GetBufferDescriptor(i);
5238 :
5239 : /*
5240 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5241 : * saves some cycles.
5242 : */
5243 1024 : if (bufHdr->tag.dbOid != dbid)
5244 706 : continue;
5245 :
5246 : /* Make sure we can handle the pin */
5247 318 : ReservePrivateRefCountEntry();
5248 318 : ResourceOwnerEnlarge(CurrentResourceOwner);
5249 :
5250 318 : buf_state = LockBufHdr(bufHdr);
5251 318 : if (bufHdr->tag.dbOid == dbid &&
5252 318 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5253 : {
5254 100 : PinBuffer_Locked(bufHdr);
5255 100 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_SHARED);
5256 100 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5257 100 : LWLockRelease(BufferDescriptorGetContentLock(bufHdr));
5258 100 : UnpinBuffer(bufHdr);
5259 : }
5260 : else
5261 218 : UnlockBufHdr(bufHdr, buf_state);
5262 : }
5263 8 : }
5264 :
5265 : /*
5266 : * Flush a previously, shared or exclusively, locked and pinned buffer to the
5267 : * OS.
5268 : */
5269 : void
5270 158 : FlushOneBuffer(Buffer buffer)
5271 : {
5272 : BufferDesc *bufHdr;
5273 :
5274 : /* currently not needed, but no fundamental reason not to support */
5275 : Assert(!BufferIsLocal(buffer));
5276 :
5277 : Assert(BufferIsPinned(buffer));
5278 :
5279 158 : bufHdr = GetBufferDescriptor(buffer - 1);
5280 :
5281 : Assert(LWLockHeldByMe(BufferDescriptorGetContentLock(bufHdr)));
5282 :
5283 158 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5284 158 : }
5285 :
5286 : /*
5287 : * ReleaseBuffer -- release the pin on a buffer
5288 : */
5289 : void
5290 119710410 : ReleaseBuffer(Buffer buffer)
5291 : {
5292 119710410 : if (!BufferIsValid(buffer))
5293 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5294 :
5295 119710410 : if (BufferIsLocal(buffer))
5296 3211394 : UnpinLocalBuffer(buffer);
5297 : else
5298 116499016 : UnpinBuffer(GetBufferDescriptor(buffer - 1));
5299 119710410 : }
5300 :
5301 : /*
5302 : * UnlockReleaseBuffer -- release the content lock and pin on a buffer
5303 : *
5304 : * This is just a shorthand for a common combination.
5305 : */
5306 : void
5307 37494288 : UnlockReleaseBuffer(Buffer buffer)
5308 : {
5309 37494288 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5310 37494288 : ReleaseBuffer(buffer);
5311 37494288 : }
5312 :
5313 : /*
5314 : * IncrBufferRefCount
5315 : * Increment the pin count on a buffer that we have *already* pinned
5316 : * at least once.
5317 : *
5318 : * This function cannot be used on a buffer we do not have pinned,
5319 : * because it doesn't change the shared buffer state.
5320 : */
5321 : void
5322 21894394 : IncrBufferRefCount(Buffer buffer)
5323 : {
5324 : Assert(BufferIsPinned(buffer));
5325 21894394 : ResourceOwnerEnlarge(CurrentResourceOwner);
5326 21894394 : if (BufferIsLocal(buffer))
5327 709026 : LocalRefCount[-buffer - 1]++;
5328 : else
5329 : {
5330 : PrivateRefCountEntry *ref;
5331 :
5332 21185368 : ref = GetPrivateRefCountEntry(buffer, true);
5333 : Assert(ref != NULL);
5334 21185368 : ref->refcount++;
5335 : }
5336 21894394 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buffer);
5337 21894394 : }
5338 :
5339 : /*
5340 : * MarkBufferDirtyHint
5341 : *
5342 : * Mark a buffer dirty for non-critical changes.
5343 : *
5344 : * This is essentially the same as MarkBufferDirty, except:
5345 : *
5346 : * 1. The caller does not write WAL; so if checksums are enabled, we may need
5347 : * to write an XLOG_FPI_FOR_HINT WAL record to protect against torn pages.
5348 : * 2. The caller might have only share-lock instead of exclusive-lock on the
5349 : * buffer's content lock.
5350 : * 3. This function does not guarantee that the buffer is always marked dirty
5351 : * (due to a race condition), so it cannot be used for important changes.
5352 : */
5353 : void
5354 20638838 : MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
5355 : {
5356 : BufferDesc *bufHdr;
5357 20638838 : Page page = BufferGetPage(buffer);
5358 :
5359 20638838 : if (!BufferIsValid(buffer))
5360 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5361 :
5362 20638838 : if (BufferIsLocal(buffer))
5363 : {
5364 1264340 : MarkLocalBufferDirty(buffer);
5365 1264340 : return;
5366 : }
5367 :
5368 19374498 : bufHdr = GetBufferDescriptor(buffer - 1);
5369 :
5370 : Assert(GetPrivateRefCount(buffer) > 0);
5371 : /* here, either share or exclusive lock is OK */
5372 : Assert(LWLockHeldByMe(BufferDescriptorGetContentLock(bufHdr)));
5373 :
5374 : /*
5375 : * This routine might get called many times on the same page, if we are
5376 : * making the first scan after commit of an xact that added/deleted many
5377 : * tuples. So, be as quick as we can if the buffer is already dirty. We
5378 : * do this by not acquiring spinlock if it looks like the status bits are
5379 : * already set. Since we make this test unlocked, there's a chance we
5380 : * might fail to notice that the flags have just been cleared, and failed
5381 : * to reset them, due to memory-ordering issues. But since this function
5382 : * is only intended to be used in cases where failing to write out the
5383 : * data would be harmless anyway, it doesn't really matter.
5384 : */
5385 19374498 : if ((pg_atomic_read_u32(&bufHdr->state) & (BM_DIRTY | BM_JUST_DIRTIED)) !=
5386 : (BM_DIRTY | BM_JUST_DIRTIED))
5387 : {
5388 1823400 : XLogRecPtr lsn = InvalidXLogRecPtr;
5389 1823400 : bool dirtied = false;
5390 1823400 : bool delayChkptFlags = false;
5391 : uint32 buf_state;
5392 :
5393 : /*
5394 : * If we need to protect hint bit updates from torn writes, WAL-log a
5395 : * full page image of the page. This full page image is only necessary
5396 : * if the hint bit update is the first change to the page since the
5397 : * last checkpoint.
5398 : *
5399 : * We don't check full_page_writes here because that logic is included
5400 : * when we call XLogInsert() since the value changes dynamically.
5401 : */
5402 3645010 : if (XLogHintBitIsNeeded() &&
5403 1821610 : (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT))
5404 : {
5405 : /*
5406 : * If we must not write WAL, due to a relfilelocator-specific
5407 : * condition or being in recovery, don't dirty the page. We can
5408 : * set the hint, just not dirty the page as a result so the hint
5409 : * is lost when we evict the page or shutdown.
5410 : *
5411 : * See src/backend/storage/page/README for longer discussion.
5412 : */
5413 1942152 : if (RecoveryInProgress() ||
5414 120604 : RelFileLocatorSkippingWAL(BufTagGetRelFileLocator(&bufHdr->tag)))
5415 1706756 : return;
5416 :
5417 : /*
5418 : * If the block is already dirty because we either made a change
5419 : * or set a hint already, then we don't need to write a full page
5420 : * image. Note that aggressive cleaning of blocks dirtied by hint
5421 : * bit setting would increase the call rate. Bulk setting of hint
5422 : * bits would reduce the call rate...
5423 : *
5424 : * We must issue the WAL record before we mark the buffer dirty.
5425 : * Otherwise we might write the page before we write the WAL. That
5426 : * causes a race condition, since a checkpoint might occur between
5427 : * writing the WAL record and marking the buffer dirty. We solve
5428 : * that with a kluge, but one that is already in use during
5429 : * transaction commit to prevent race conditions. Basically, we
5430 : * simply prevent the checkpoint WAL record from being written
5431 : * until we have marked the buffer dirty. We don't start the
5432 : * checkpoint flush until we have marked dirty, so our checkpoint
5433 : * must flush the change to disk successfully or the checkpoint
5434 : * never gets written, so crash recovery will fix.
5435 : *
5436 : * It's possible we may enter here without an xid, so it is
5437 : * essential that CreateCheckPoint waits for virtual transactions
5438 : * rather than full transactionids.
5439 : */
5440 : Assert((MyProc->delayChkptFlags & DELAY_CHKPT_START) == 0);
5441 114792 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
5442 114792 : delayChkptFlags = true;
5443 114792 : lsn = XLogSaveBufferForHint(buffer, buffer_std);
5444 : }
5445 :
5446 116644 : buf_state = LockBufHdr(bufHdr);
5447 :
5448 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5449 :
5450 116644 : if (!(buf_state & BM_DIRTY))
5451 : {
5452 116598 : dirtied = true; /* Means "will be dirtied by this action" */
5453 :
5454 : /*
5455 : * Set the page LSN if we wrote a backup block. We aren't supposed
5456 : * to set this when only holding a share lock but as long as we
5457 : * serialise it somehow we're OK. We choose to set LSN while
5458 : * holding the buffer header lock, which causes any reader of an
5459 : * LSN who holds only a share lock to also obtain a buffer header
5460 : * lock before using PageGetLSN(), which is enforced in
5461 : * BufferGetLSNAtomic().
5462 : *
5463 : * If checksums are enabled, you might think we should reset the
5464 : * checksum here. That will happen when the page is written
5465 : * sometime later in this checkpoint cycle.
5466 : */
5467 116598 : if (!XLogRecPtrIsInvalid(lsn))
5468 62284 : PageSetLSN(page, lsn);
5469 : }
5470 :
5471 116644 : buf_state |= BM_DIRTY | BM_JUST_DIRTIED;
5472 116644 : UnlockBufHdr(bufHdr, buf_state);
5473 :
5474 116644 : if (delayChkptFlags)
5475 114792 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
5476 :
5477 116644 : if (dirtied)
5478 : {
5479 116598 : pgBufferUsage.shared_blks_dirtied++;
5480 116598 : if (VacuumCostActive)
5481 2388 : VacuumCostBalance += VacuumCostPageDirty;
5482 : }
5483 : }
5484 : }
5485 :
5486 : /*
5487 : * Release buffer content locks for shared buffers.
5488 : *
5489 : * Used to clean up after errors.
5490 : *
5491 : * Currently, we can expect that lwlock.c's LWLockReleaseAll() took care
5492 : * of releasing buffer content locks per se; the only thing we need to deal
5493 : * with here is clearing any PIN_COUNT request that was in progress.
5494 : */
5495 : void
5496 100598 : UnlockBuffers(void)
5497 : {
5498 100598 : BufferDesc *buf = PinCountWaitBuf;
5499 :
5500 100598 : if (buf)
5501 : {
5502 : uint32 buf_state;
5503 :
5504 0 : buf_state = LockBufHdr(buf);
5505 :
5506 : /*
5507 : * Don't complain if flag bit not set; it could have been reset but we
5508 : * got a cancel/die interrupt before getting the signal.
5509 : */
5510 0 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5511 0 : buf->wait_backend_pgprocno == MyProcNumber)
5512 0 : buf_state &= ~BM_PIN_COUNT_WAITER;
5513 :
5514 0 : UnlockBufHdr(buf, buf_state);
5515 :
5516 0 : PinCountWaitBuf = NULL;
5517 : }
5518 100598 : }
5519 :
5520 : /*
5521 : * Acquire or release the content_lock for the buffer.
5522 : */
5523 : void
5524 337208610 : LockBuffer(Buffer buffer, int mode)
5525 : {
5526 : BufferDesc *buf;
5527 :
5528 : Assert(BufferIsPinned(buffer));
5529 337208610 : if (BufferIsLocal(buffer))
5530 19770838 : return; /* local buffers need no lock */
5531 :
5532 317437772 : buf = GetBufferDescriptor(buffer - 1);
5533 :
5534 317437772 : if (mode == BUFFER_LOCK_UNLOCK)
5535 160637730 : LWLockRelease(BufferDescriptorGetContentLock(buf));
5536 156800042 : else if (mode == BUFFER_LOCK_SHARE)
5537 110781228 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_SHARED);
5538 46018814 : else if (mode == BUFFER_LOCK_EXCLUSIVE)
5539 46018814 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_EXCLUSIVE);
5540 : else
5541 0 : elog(ERROR, "unrecognized buffer lock mode: %d", mode);
5542 : }
5543 :
5544 : /*
5545 : * Acquire the content_lock for the buffer, but only if we don't have to wait.
5546 : *
5547 : * This assumes the caller wants BUFFER_LOCK_EXCLUSIVE mode.
5548 : */
5549 : bool
5550 3057016 : ConditionalLockBuffer(Buffer buffer)
5551 : {
5552 : BufferDesc *buf;
5553 :
5554 : Assert(BufferIsPinned(buffer));
5555 3057016 : if (BufferIsLocal(buffer))
5556 129548 : return true; /* act as though we got it */
5557 :
5558 2927468 : buf = GetBufferDescriptor(buffer - 1);
5559 :
5560 2927468 : return LWLockConditionalAcquire(BufferDescriptorGetContentLock(buf),
5561 : LW_EXCLUSIVE);
5562 : }
5563 :
5564 : /*
5565 : * Verify that this backend is pinning the buffer exactly once.
5566 : *
5567 : * NOTE: Like in BufferIsPinned(), what we check here is that *this* backend
5568 : * holds a pin on the buffer. We do not care whether some other backend does.
5569 : */
5570 : void
5571 4534216 : CheckBufferIsPinnedOnce(Buffer buffer)
5572 : {
5573 4534216 : if (BufferIsLocal(buffer))
5574 : {
5575 1578 : if (LocalRefCount[-buffer - 1] != 1)
5576 0 : elog(ERROR, "incorrect local pin count: %d",
5577 : LocalRefCount[-buffer - 1]);
5578 : }
5579 : else
5580 : {
5581 4532638 : if (GetPrivateRefCount(buffer) != 1)
5582 0 : elog(ERROR, "incorrect local pin count: %d",
5583 : GetPrivateRefCount(buffer));
5584 : }
5585 4534216 : }
5586 :
5587 : /*
5588 : * LockBufferForCleanup - lock a buffer in preparation for deleting items
5589 : *
5590 : * Items may be deleted from a disk page only when the caller (a) holds an
5591 : * exclusive lock on the buffer and (b) has observed that no other backend
5592 : * holds a pin on the buffer. If there is a pin, then the other backend
5593 : * might have a pointer into the buffer (for example, a heapscan reference
5594 : * to an item --- see README for more details). It's OK if a pin is added
5595 : * after the cleanup starts, however; the newly-arrived backend will be
5596 : * unable to look at the page until we release the exclusive lock.
5597 : *
5598 : * To implement this protocol, a would-be deleter must pin the buffer and
5599 : * then call LockBufferForCleanup(). LockBufferForCleanup() is similar to
5600 : * LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE), except that it loops until
5601 : * it has successfully observed pin count = 1.
5602 : */
5603 : void
5604 41388 : LockBufferForCleanup(Buffer buffer)
5605 : {
5606 : BufferDesc *bufHdr;
5607 41388 : TimestampTz waitStart = 0;
5608 41388 : bool waiting = false;
5609 41388 : bool logged_recovery_conflict = false;
5610 :
5611 : Assert(BufferIsPinned(buffer));
5612 : Assert(PinCountWaitBuf == NULL);
5613 :
5614 41388 : CheckBufferIsPinnedOnce(buffer);
5615 :
5616 : /*
5617 : * We do not yet need to be worried about in-progress AIOs holding a pin,
5618 : * as we, so far, only support doing reads via AIO and this function can
5619 : * only be called once the buffer is valid (i.e. no read can be in
5620 : * flight).
5621 : */
5622 :
5623 : /* Nobody else to wait for */
5624 41388 : if (BufferIsLocal(buffer))
5625 32 : return;
5626 :
5627 41356 : bufHdr = GetBufferDescriptor(buffer - 1);
5628 :
5629 : for (;;)
5630 20 : {
5631 : uint32 buf_state;
5632 :
5633 : /* Try to acquire lock */
5634 41376 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
5635 41376 : buf_state = LockBufHdr(bufHdr);
5636 :
5637 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5638 41376 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
5639 : {
5640 : /* Successfully acquired exclusive lock with pincount 1 */
5641 41356 : UnlockBufHdr(bufHdr, buf_state);
5642 :
5643 : /*
5644 : * Emit the log message if recovery conflict on buffer pin was
5645 : * resolved but the startup process waited longer than
5646 : * deadlock_timeout for it.
5647 : */
5648 41356 : if (logged_recovery_conflict)
5649 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
5650 : waitStart, GetCurrentTimestamp(),
5651 : NULL, false);
5652 :
5653 41356 : if (waiting)
5654 : {
5655 : /* reset ps display to remove the suffix if we added one */
5656 4 : set_ps_display_remove_suffix();
5657 4 : waiting = false;
5658 : }
5659 41356 : return;
5660 : }
5661 : /* Failed, so mark myself as waiting for pincount 1 */
5662 20 : if (buf_state & BM_PIN_COUNT_WAITER)
5663 : {
5664 0 : UnlockBufHdr(bufHdr, buf_state);
5665 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5666 0 : elog(ERROR, "multiple backends attempting to wait for pincount 1");
5667 : }
5668 20 : bufHdr->wait_backend_pgprocno = MyProcNumber;
5669 20 : PinCountWaitBuf = bufHdr;
5670 20 : buf_state |= BM_PIN_COUNT_WAITER;
5671 20 : UnlockBufHdr(bufHdr, buf_state);
5672 20 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5673 :
5674 : /* Wait to be signaled by UnpinBuffer() */
5675 20 : if (InHotStandby)
5676 : {
5677 20 : if (!waiting)
5678 : {
5679 : /* adjust the process title to indicate that it's waiting */
5680 4 : set_ps_display_suffix("waiting");
5681 4 : waiting = true;
5682 : }
5683 :
5684 : /*
5685 : * Emit the log message if the startup process is waiting longer
5686 : * than deadlock_timeout for recovery conflict on buffer pin.
5687 : *
5688 : * Skip this if first time through because the startup process has
5689 : * not started waiting yet in this case. So, the wait start
5690 : * timestamp is set after this logic.
5691 : */
5692 20 : if (waitStart != 0 && !logged_recovery_conflict)
5693 : {
5694 6 : TimestampTz now = GetCurrentTimestamp();
5695 :
5696 6 : if (TimestampDifferenceExceeds(waitStart, now,
5697 : DeadlockTimeout))
5698 : {
5699 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
5700 : waitStart, now, NULL, true);
5701 4 : logged_recovery_conflict = true;
5702 : }
5703 : }
5704 :
5705 : /*
5706 : * Set the wait start timestamp if logging is enabled and first
5707 : * time through.
5708 : */
5709 20 : if (log_recovery_conflict_waits && waitStart == 0)
5710 4 : waitStart = GetCurrentTimestamp();
5711 :
5712 : /* Publish the bufid that Startup process waits on */
5713 20 : SetStartupBufferPinWaitBufId(buffer - 1);
5714 : /* Set alarm and then wait to be signaled by UnpinBuffer() */
5715 20 : ResolveRecoveryConflictWithBufferPin();
5716 : /* Reset the published bufid */
5717 20 : SetStartupBufferPinWaitBufId(-1);
5718 : }
5719 : else
5720 0 : ProcWaitForSignal(WAIT_EVENT_BUFFER_PIN);
5721 :
5722 : /*
5723 : * Remove flag marking us as waiter. Normally this will not be set
5724 : * anymore, but ProcWaitForSignal() can return for other signals as
5725 : * well. We take care to only reset the flag if we're the waiter, as
5726 : * theoretically another backend could have started waiting. That's
5727 : * impossible with the current usages due to table level locking, but
5728 : * better be safe.
5729 : */
5730 20 : buf_state = LockBufHdr(bufHdr);
5731 20 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5732 16 : bufHdr->wait_backend_pgprocno == MyProcNumber)
5733 16 : buf_state &= ~BM_PIN_COUNT_WAITER;
5734 20 : UnlockBufHdr(bufHdr, buf_state);
5735 :
5736 20 : PinCountWaitBuf = NULL;
5737 : /* Loop back and try again */
5738 : }
5739 : }
5740 :
5741 : /*
5742 : * Check called from ProcessRecoveryConflictInterrupts() when Startup process
5743 : * requests cancellation of all pin holders that are blocking it.
5744 : */
5745 : bool
5746 8 : HoldingBufferPinThatDelaysRecovery(void)
5747 : {
5748 8 : int bufid = GetStartupBufferPinWaitBufId();
5749 :
5750 : /*
5751 : * If we get woken slowly then it's possible that the Startup process was
5752 : * already woken by other backends before we got here. Also possible that
5753 : * we get here by multiple interrupts or interrupts at inappropriate
5754 : * times, so make sure we do nothing if the bufid is not set.
5755 : */
5756 8 : if (bufid < 0)
5757 4 : return false;
5758 :
5759 4 : if (GetPrivateRefCount(bufid + 1) > 0)
5760 4 : return true;
5761 :
5762 0 : return false;
5763 : }
5764 :
5765 : /*
5766 : * ConditionalLockBufferForCleanup - as above, but don't wait to get the lock
5767 : *
5768 : * We won't loop, but just check once to see if the pin count is OK. If
5769 : * not, return false with no lock held.
5770 : */
5771 : bool
5772 863526 : ConditionalLockBufferForCleanup(Buffer buffer)
5773 : {
5774 : BufferDesc *bufHdr;
5775 : uint32 buf_state,
5776 : refcount;
5777 :
5778 : Assert(BufferIsValid(buffer));
5779 :
5780 : /* see AIO related comment in LockBufferForCleanup() */
5781 :
5782 863526 : if (BufferIsLocal(buffer))
5783 : {
5784 1600 : refcount = LocalRefCount[-buffer - 1];
5785 : /* There should be exactly one pin */
5786 : Assert(refcount > 0);
5787 1600 : if (refcount != 1)
5788 42 : return false;
5789 : /* Nobody else to wait for */
5790 1558 : return true;
5791 : }
5792 :
5793 : /* There should be exactly one local pin */
5794 861926 : refcount = GetPrivateRefCount(buffer);
5795 : Assert(refcount);
5796 861926 : if (refcount != 1)
5797 382 : return false;
5798 :
5799 : /* Try to acquire lock */
5800 861544 : if (!ConditionalLockBuffer(buffer))
5801 58 : return false;
5802 :
5803 861486 : bufHdr = GetBufferDescriptor(buffer - 1);
5804 861486 : buf_state = LockBufHdr(bufHdr);
5805 861486 : refcount = BUF_STATE_GET_REFCOUNT(buf_state);
5806 :
5807 : Assert(refcount > 0);
5808 861486 : if (refcount == 1)
5809 : {
5810 : /* Successfully acquired exclusive lock with pincount 1 */
5811 861264 : UnlockBufHdr(bufHdr, buf_state);
5812 861264 : return true;
5813 : }
5814 :
5815 : /* Failed, so release the lock */
5816 222 : UnlockBufHdr(bufHdr, buf_state);
5817 222 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5818 222 : return false;
5819 : }
5820 :
5821 : /*
5822 : * IsBufferCleanupOK - as above, but we already have the lock
5823 : *
5824 : * Check whether it's OK to perform cleanup on a buffer we've already
5825 : * locked. If we observe that the pin count is 1, our exclusive lock
5826 : * happens to be a cleanup lock, and we can proceed with anything that
5827 : * would have been allowable had we sought a cleanup lock originally.
5828 : */
5829 : bool
5830 4062 : IsBufferCleanupOK(Buffer buffer)
5831 : {
5832 : BufferDesc *bufHdr;
5833 : uint32 buf_state;
5834 :
5835 : Assert(BufferIsValid(buffer));
5836 :
5837 : /* see AIO related comment in LockBufferForCleanup() */
5838 :
5839 4062 : if (BufferIsLocal(buffer))
5840 : {
5841 : /* There should be exactly one pin */
5842 0 : if (LocalRefCount[-buffer - 1] != 1)
5843 0 : return false;
5844 : /* Nobody else to wait for */
5845 0 : return true;
5846 : }
5847 :
5848 : /* There should be exactly one local pin */
5849 4062 : if (GetPrivateRefCount(buffer) != 1)
5850 0 : return false;
5851 :
5852 4062 : bufHdr = GetBufferDescriptor(buffer - 1);
5853 :
5854 : /* caller must hold exclusive lock on buffer */
5855 : Assert(LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
5856 : LW_EXCLUSIVE));
5857 :
5858 4062 : buf_state = LockBufHdr(bufHdr);
5859 :
5860 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5861 4062 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
5862 : {
5863 : /* pincount is OK. */
5864 4062 : UnlockBufHdr(bufHdr, buf_state);
5865 4062 : return true;
5866 : }
5867 :
5868 0 : UnlockBufHdr(bufHdr, buf_state);
5869 0 : return false;
5870 : }
5871 :
5872 :
5873 : /*
5874 : * Functions for buffer I/O handling
5875 : *
5876 : * Also note that these are used only for shared buffers, not local ones.
5877 : */
5878 :
5879 : /*
5880 : * WaitIO -- Block until the IO_IN_PROGRESS flag on 'buf' is cleared.
5881 : */
5882 : static void
5883 4670 : WaitIO(BufferDesc *buf)
5884 : {
5885 4670 : ConditionVariable *cv = BufferDescriptorGetIOCV(buf);
5886 :
5887 4670 : ConditionVariablePrepareToSleep(cv);
5888 : for (;;)
5889 4642 : {
5890 : uint32 buf_state;
5891 : PgAioWaitRef iow;
5892 :
5893 : /*
5894 : * It may not be necessary to acquire the spinlock to check the flag
5895 : * here, but since this test is essential for correctness, we'd better
5896 : * play it safe.
5897 : */
5898 9312 : buf_state = LockBufHdr(buf);
5899 :
5900 : /*
5901 : * Copy the wait reference while holding the spinlock. This protects
5902 : * against a concurrent TerminateBufferIO() in another backend from
5903 : * clearing the wref while it's being read.
5904 : */
5905 9312 : iow = buf->io_wref;
5906 9312 : UnlockBufHdr(buf, buf_state);
5907 :
5908 : /* no IO in progress, we don't need to wait */
5909 9312 : if (!(buf_state & BM_IO_IN_PROGRESS))
5910 4670 : break;
5911 :
5912 : /*
5913 : * The buffer has asynchronous IO in progress, wait for it to
5914 : * complete.
5915 : */
5916 4642 : if (pgaio_wref_valid(&iow))
5917 : {
5918 4256 : pgaio_wref_wait(&iow);
5919 :
5920 : /*
5921 : * The AIO subsystem internally uses condition variables and thus
5922 : * might remove this backend from the BufferDesc's CV. While that
5923 : * wouldn't cause a correctness issue (the first CV sleep just
5924 : * immediately returns if not already registered), it seems worth
5925 : * avoiding unnecessary loop iterations, given that we take care
5926 : * to do so at the start of the function.
5927 : */
5928 4256 : ConditionVariablePrepareToSleep(cv);
5929 4256 : continue;
5930 : }
5931 :
5932 : /* wait on BufferDesc->cv, e.g. for concurrent synchronous IO */
5933 386 : ConditionVariableSleep(cv, WAIT_EVENT_BUFFER_IO);
5934 : }
5935 4670 : ConditionVariableCancelSleep();
5936 4670 : }
5937 :
5938 : /*
5939 : * StartBufferIO: begin I/O on this buffer
5940 : * (Assumptions)
5941 : * My process is executing no IO on this buffer
5942 : * The buffer is Pinned
5943 : *
5944 : * In some scenarios multiple backends could attempt the same I/O operation
5945 : * concurrently. If someone else has already started I/O on this buffer then
5946 : * we will wait for completion of the IO using WaitIO().
5947 : *
5948 : * Input operations are only attempted on buffers that are not BM_VALID,
5949 : * and output operations only on buffers that are BM_VALID and BM_DIRTY,
5950 : * so we can always tell if the work is already done.
5951 : *
5952 : * Returns true if we successfully marked the buffer as I/O busy,
5953 : * false if someone else already did the work.
5954 : *
5955 : * If nowait is true, then we don't wait for an I/O to be finished by another
5956 : * backend. In that case, false indicates either that the I/O was already
5957 : * finished, or is still in progress. This is useful for callers that want to
5958 : * find out if they can perform the I/O as part of a larger operation, without
5959 : * waiting for the answer or distinguishing the reasons why not.
5960 : */
5961 : bool
5962 4761818 : StartBufferIO(BufferDesc *buf, bool forInput, bool nowait)
5963 : {
5964 : uint32 buf_state;
5965 :
5966 4761818 : ResourceOwnerEnlarge(CurrentResourceOwner);
5967 :
5968 : for (;;)
5969 : {
5970 4766488 : buf_state = LockBufHdr(buf);
5971 :
5972 4766488 : if (!(buf_state & BM_IO_IN_PROGRESS))
5973 4761810 : break;
5974 4678 : UnlockBufHdr(buf, buf_state);
5975 4678 : if (nowait)
5976 8 : return false;
5977 4670 : WaitIO(buf);
5978 : }
5979 :
5980 : /* Once we get here, there is definitely no I/O active on this buffer */
5981 :
5982 : /* Check if someone else already did the I/O */
5983 4761810 : if (forInput ? (buf_state & BM_VALID) : !(buf_state & BM_DIRTY))
5984 : {
5985 5028 : UnlockBufHdr(buf, buf_state);
5986 5028 : return false;
5987 : }
5988 :
5989 4756782 : buf_state |= BM_IO_IN_PROGRESS;
5990 4756782 : UnlockBufHdr(buf, buf_state);
5991 :
5992 4756782 : ResourceOwnerRememberBufferIO(CurrentResourceOwner,
5993 : BufferDescriptorGetBuffer(buf));
5994 :
5995 4756782 : return true;
5996 : }
5997 :
5998 : /*
5999 : * TerminateBufferIO: release a buffer we were doing I/O on
6000 : * (Assumptions)
6001 : * My process is executing IO for the buffer
6002 : * BM_IO_IN_PROGRESS bit is set for the buffer
6003 : * The buffer is Pinned
6004 : *
6005 : * If clear_dirty is true and BM_JUST_DIRTIED is not set, we clear the
6006 : * buffer's BM_DIRTY flag. This is appropriate when terminating a
6007 : * successful write. The check on BM_JUST_DIRTIED is necessary to avoid
6008 : * marking the buffer clean if it was re-dirtied while we were writing.
6009 : *
6010 : * set_flag_bits gets ORed into the buffer's flags. It must include
6011 : * BM_IO_ERROR in a failure case. For successful completion it could
6012 : * be 0, or BM_VALID if we just finished reading in the page.
6013 : *
6014 : * If forget_owner is true, we release the buffer I/O from the current
6015 : * resource owner. (forget_owner=false is used when the resource owner itself
6016 : * is being released)
6017 : */
6018 : void
6019 4475630 : TerminateBufferIO(BufferDesc *buf, bool clear_dirty, uint32 set_flag_bits,
6020 : bool forget_owner, bool release_aio)
6021 : {
6022 : uint32 buf_state;
6023 :
6024 4475630 : buf_state = LockBufHdr(buf);
6025 :
6026 : Assert(buf_state & BM_IO_IN_PROGRESS);
6027 4475630 : buf_state &= ~BM_IO_IN_PROGRESS;
6028 :
6029 : /* Clear earlier errors, if this IO failed, it'll be marked again */
6030 4475630 : buf_state &= ~BM_IO_ERROR;
6031 :
6032 4475630 : if (clear_dirty && !(buf_state & BM_JUST_DIRTIED))
6033 1053730 : buf_state &= ~(BM_DIRTY | BM_CHECKPOINT_NEEDED);
6034 :
6035 4475630 : if (release_aio)
6036 : {
6037 : /* release ownership by the AIO subsystem */
6038 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6039 2440026 : buf_state -= BUF_REFCOUNT_ONE;
6040 2440026 : pgaio_wref_clear(&buf->io_wref);
6041 : }
6042 :
6043 4475630 : buf_state |= set_flag_bits;
6044 4475630 : UnlockBufHdr(buf, buf_state);
6045 :
6046 4475630 : if (forget_owner)
6047 2035562 : ResourceOwnerForgetBufferIO(CurrentResourceOwner,
6048 : BufferDescriptorGetBuffer(buf));
6049 :
6050 4475630 : ConditionVariableBroadcast(BufferDescriptorGetIOCV(buf));
6051 :
6052 : /*
6053 : * Support LockBufferForCleanup()
6054 : *
6055 : * We may have just released the last pin other than the waiter's. In most
6056 : * cases, this backend holds another pin on the buffer. But, if, for
6057 : * example, this backend is completing an IO issued by another backend, it
6058 : * may be time to wake the waiter.
6059 : */
6060 4475630 : if (release_aio && (buf_state & BM_PIN_COUNT_WAITER))
6061 0 : WakePinCountWaiter(buf);
6062 4475630 : }
6063 :
6064 : /*
6065 : * AbortBufferIO: Clean up active buffer I/O after an error.
6066 : *
6067 : * All LWLocks we might have held have been released,
6068 : * but we haven't yet released buffer pins, so the buffer is still pinned.
6069 : *
6070 : * If I/O was in progress, we always set BM_IO_ERROR, even though it's
6071 : * possible the error condition wasn't related to the I/O.
6072 : *
6073 : * Note: this does not remove the buffer I/O from the resource owner.
6074 : * That's correct when we're releasing the whole resource owner, but
6075 : * beware if you use this in other contexts.
6076 : */
6077 : static void
6078 30 : AbortBufferIO(Buffer buffer)
6079 : {
6080 30 : BufferDesc *buf_hdr = GetBufferDescriptor(buffer - 1);
6081 : uint32 buf_state;
6082 :
6083 30 : buf_state = LockBufHdr(buf_hdr);
6084 : Assert(buf_state & (BM_IO_IN_PROGRESS | BM_TAG_VALID));
6085 :
6086 30 : if (!(buf_state & BM_VALID))
6087 : {
6088 : Assert(!(buf_state & BM_DIRTY));
6089 30 : UnlockBufHdr(buf_hdr, buf_state);
6090 : }
6091 : else
6092 : {
6093 : Assert(buf_state & BM_DIRTY);
6094 0 : UnlockBufHdr(buf_hdr, buf_state);
6095 :
6096 : /* Issue notice if this is not the first failure... */
6097 0 : if (buf_state & BM_IO_ERROR)
6098 : {
6099 : /* Buffer is pinned, so we can read tag without spinlock */
6100 0 : ereport(WARNING,
6101 : (errcode(ERRCODE_IO_ERROR),
6102 : errmsg("could not write block %u of %s",
6103 : buf_hdr->tag.blockNum,
6104 : relpathperm(BufTagGetRelFileLocator(&buf_hdr->tag),
6105 : BufTagGetForkNum(&buf_hdr->tag)).str),
6106 : errdetail("Multiple failures --- write error might be permanent.")));
6107 : }
6108 : }
6109 :
6110 30 : TerminateBufferIO(buf_hdr, false, BM_IO_ERROR, false, false);
6111 30 : }
6112 :
6113 : /*
6114 : * Error context callback for errors occurring during shared buffer writes.
6115 : */
6116 : static void
6117 70 : shared_buffer_write_error_callback(void *arg)
6118 : {
6119 70 : BufferDesc *bufHdr = (BufferDesc *) arg;
6120 :
6121 : /* Buffer is pinned, so we can read the tag without locking the spinlock */
6122 70 : if (bufHdr != NULL)
6123 140 : errcontext("writing block %u of relation %s",
6124 : bufHdr->tag.blockNum,
6125 70 : relpathperm(BufTagGetRelFileLocator(&bufHdr->tag),
6126 : BufTagGetForkNum(&bufHdr->tag)).str);
6127 70 : }
6128 :
6129 : /*
6130 : * Error context callback for errors occurring during local buffer writes.
6131 : */
6132 : static void
6133 0 : local_buffer_write_error_callback(void *arg)
6134 : {
6135 0 : BufferDesc *bufHdr = (BufferDesc *) arg;
6136 :
6137 0 : if (bufHdr != NULL)
6138 0 : errcontext("writing block %u of relation %s",
6139 : bufHdr->tag.blockNum,
6140 0 : relpathbackend(BufTagGetRelFileLocator(&bufHdr->tag),
6141 : MyProcNumber,
6142 : BufTagGetForkNum(&bufHdr->tag)).str);
6143 0 : }
6144 :
6145 : /*
6146 : * RelFileLocator qsort/bsearch comparator; see RelFileLocatorEquals.
6147 : */
6148 : static int
6149 18431232 : rlocator_comparator(const void *p1, const void *p2)
6150 : {
6151 18431232 : RelFileLocator n1 = *(const RelFileLocator *) p1;
6152 18431232 : RelFileLocator n2 = *(const RelFileLocator *) p2;
6153 :
6154 18431232 : if (n1.relNumber < n2.relNumber)
6155 18355048 : return -1;
6156 76184 : else if (n1.relNumber > n2.relNumber)
6157 72728 : return 1;
6158 :
6159 3456 : if (n1.dbOid < n2.dbOid)
6160 0 : return -1;
6161 3456 : else if (n1.dbOid > n2.dbOid)
6162 0 : return 1;
6163 :
6164 3456 : if (n1.spcOid < n2.spcOid)
6165 0 : return -1;
6166 3456 : else if (n1.spcOid > n2.spcOid)
6167 0 : return 1;
6168 : else
6169 3456 : return 0;
6170 : }
6171 :
6172 : /*
6173 : * Lock buffer header - set BM_LOCKED in buffer state.
6174 : */
6175 : uint32
6176 73790178 : LockBufHdr(BufferDesc *desc)
6177 : {
6178 : SpinDelayStatus delayStatus;
6179 : uint32 old_buf_state;
6180 :
6181 : Assert(!BufferIsLocal(BufferDescriptorGetBuffer(desc)));
6182 :
6183 73790178 : init_local_spin_delay(&delayStatus);
6184 :
6185 : while (true)
6186 : {
6187 : /* set BM_LOCKED flag */
6188 73868658 : old_buf_state = pg_atomic_fetch_or_u32(&desc->state, BM_LOCKED);
6189 : /* if it wasn't set before we're OK */
6190 73868658 : if (!(old_buf_state & BM_LOCKED))
6191 73790178 : break;
6192 78480 : perform_spin_delay(&delayStatus);
6193 : }
6194 73790178 : finish_spin_delay(&delayStatus);
6195 73790178 : return old_buf_state | BM_LOCKED;
6196 : }
6197 :
6198 : /*
6199 : * Wait until the BM_LOCKED flag isn't set anymore and return the buffer's
6200 : * state at that point.
6201 : *
6202 : * Obviously the buffer could be locked by the time the value is returned, so
6203 : * this is primarily useful in CAS style loops.
6204 : */
6205 : static uint32
6206 3816 : WaitBufHdrUnlocked(BufferDesc *buf)
6207 : {
6208 : SpinDelayStatus delayStatus;
6209 : uint32 buf_state;
6210 :
6211 3816 : init_local_spin_delay(&delayStatus);
6212 :
6213 3816 : buf_state = pg_atomic_read_u32(&buf->state);
6214 :
6215 29540 : while (buf_state & BM_LOCKED)
6216 : {
6217 25724 : perform_spin_delay(&delayStatus);
6218 25724 : buf_state = pg_atomic_read_u32(&buf->state);
6219 : }
6220 :
6221 3816 : finish_spin_delay(&delayStatus);
6222 :
6223 3816 : return buf_state;
6224 : }
6225 :
6226 : /*
6227 : * BufferTag comparator.
6228 : */
6229 : static inline int
6230 0 : buffertag_comparator(const BufferTag *ba, const BufferTag *bb)
6231 : {
6232 : int ret;
6233 : RelFileLocator rlocatora;
6234 : RelFileLocator rlocatorb;
6235 :
6236 0 : rlocatora = BufTagGetRelFileLocator(ba);
6237 0 : rlocatorb = BufTagGetRelFileLocator(bb);
6238 :
6239 0 : ret = rlocator_comparator(&rlocatora, &rlocatorb);
6240 :
6241 0 : if (ret != 0)
6242 0 : return ret;
6243 :
6244 0 : if (BufTagGetForkNum(ba) < BufTagGetForkNum(bb))
6245 0 : return -1;
6246 0 : if (BufTagGetForkNum(ba) > BufTagGetForkNum(bb))
6247 0 : return 1;
6248 :
6249 0 : if (ba->blockNum < bb->blockNum)
6250 0 : return -1;
6251 0 : if (ba->blockNum > bb->blockNum)
6252 0 : return 1;
6253 :
6254 0 : return 0;
6255 : }
6256 :
6257 : /*
6258 : * Comparator determining the writeout order in a checkpoint.
6259 : *
6260 : * It is important that tablespaces are compared first, the logic balancing
6261 : * writes between tablespaces relies on it.
6262 : */
6263 : static inline int
6264 5670228 : ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b)
6265 : {
6266 : /* compare tablespace */
6267 5670228 : if (a->tsId < b->tsId)
6268 11984 : return -1;
6269 5658244 : else if (a->tsId > b->tsId)
6270 36682 : return 1;
6271 : /* compare relation */
6272 5621562 : if (a->relNumber < b->relNumber)
6273 1579154 : return -1;
6274 4042408 : else if (a->relNumber > b->relNumber)
6275 1517432 : return 1;
6276 : /* compare fork */
6277 2524976 : else if (a->forkNum < b->forkNum)
6278 107546 : return -1;
6279 2417430 : else if (a->forkNum > b->forkNum)
6280 112556 : return 1;
6281 : /* compare block number */
6282 2304874 : else if (a->blockNum < b->blockNum)
6283 1128032 : return -1;
6284 1176842 : else if (a->blockNum > b->blockNum)
6285 1104134 : return 1;
6286 : /* equal page IDs are unlikely, but not impossible */
6287 72708 : return 0;
6288 : }
6289 :
6290 : /*
6291 : * Comparator for a Min-Heap over the per-tablespace checkpoint completion
6292 : * progress.
6293 : */
6294 : static int
6295 437928 : ts_ckpt_progress_comparator(Datum a, Datum b, void *arg)
6296 : {
6297 437928 : CkptTsStatus *sa = (CkptTsStatus *) a;
6298 437928 : CkptTsStatus *sb = (CkptTsStatus *) b;
6299 :
6300 : /* we want a min-heap, so return 1 for the a < b */
6301 437928 : if (sa->progress < sb->progress)
6302 417174 : return 1;
6303 20754 : else if (sa->progress == sb->progress)
6304 1448 : return 0;
6305 : else
6306 19306 : return -1;
6307 : }
6308 :
6309 : /*
6310 : * Initialize a writeback context, discarding potential previous state.
6311 : *
6312 : * *max_pending is a pointer instead of an immediate value, so the coalesce
6313 : * limits can easily changed by the GUC mechanism, and so calling code does
6314 : * not have to check the current configuration. A value of 0 means that no
6315 : * writeback control will be performed.
6316 : */
6317 : void
6318 5090 : WritebackContextInit(WritebackContext *context, int *max_pending)
6319 : {
6320 : Assert(*max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
6321 :
6322 5090 : context->max_pending = max_pending;
6323 5090 : context->nr_pending = 0;
6324 5090 : }
6325 :
6326 : /*
6327 : * Add buffer to list of pending writeback requests.
6328 : */
6329 : void
6330 1043126 : ScheduleBufferTagForWriteback(WritebackContext *wb_context, IOContext io_context,
6331 : BufferTag *tag)
6332 : {
6333 : PendingWriteback *pending;
6334 :
6335 : /*
6336 : * As pg_flush_data() doesn't do anything with fsync disabled, there's no
6337 : * point in tracking in that case.
6338 : */
6339 1043126 : if (io_direct_flags & IO_DIRECT_DATA ||
6340 1042072 : !enableFsync)
6341 1043124 : return;
6342 :
6343 : /*
6344 : * Add buffer to the pending writeback array, unless writeback control is
6345 : * disabled.
6346 : */
6347 2 : if (*wb_context->max_pending > 0)
6348 : {
6349 : Assert(*wb_context->max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
6350 :
6351 0 : pending = &wb_context->pending_writebacks[wb_context->nr_pending++];
6352 :
6353 0 : pending->tag = *tag;
6354 : }
6355 :
6356 : /*
6357 : * Perform pending flushes if the writeback limit is exceeded. This
6358 : * includes the case where previously an item has been added, but control
6359 : * is now disabled.
6360 : */
6361 2 : if (wb_context->nr_pending >= *wb_context->max_pending)
6362 2 : IssuePendingWritebacks(wb_context, io_context);
6363 : }
6364 :
6365 : #define ST_SORT sort_pending_writebacks
6366 : #define ST_ELEMENT_TYPE PendingWriteback
6367 : #define ST_COMPARE(a, b) buffertag_comparator(&a->tag, &b->tag)
6368 : #define ST_SCOPE static
6369 : #define ST_DEFINE
6370 : #include <lib/sort_template.h>
6371 :
6372 : /*
6373 : * Issue all pending writeback requests, previously scheduled with
6374 : * ScheduleBufferTagForWriteback, to the OS.
6375 : *
6376 : * Because this is only used to improve the OSs IO scheduling we try to never
6377 : * error out - it's just a hint.
6378 : */
6379 : void
6380 2034 : IssuePendingWritebacks(WritebackContext *wb_context, IOContext io_context)
6381 : {
6382 : instr_time io_start;
6383 : int i;
6384 :
6385 2034 : if (wb_context->nr_pending == 0)
6386 2034 : return;
6387 :
6388 : /*
6389 : * Executing the writes in-order can make them a lot faster, and allows to
6390 : * merge writeback requests to consecutive blocks into larger writebacks.
6391 : */
6392 0 : sort_pending_writebacks(wb_context->pending_writebacks,
6393 0 : wb_context->nr_pending);
6394 :
6395 0 : io_start = pgstat_prepare_io_time(track_io_timing);
6396 :
6397 : /*
6398 : * Coalesce neighbouring writes, but nothing else. For that we iterate
6399 : * through the, now sorted, array of pending flushes, and look forward to
6400 : * find all neighbouring (or identical) writes.
6401 : */
6402 0 : for (i = 0; i < wb_context->nr_pending; i++)
6403 : {
6404 : PendingWriteback *cur;
6405 : PendingWriteback *next;
6406 : SMgrRelation reln;
6407 : int ahead;
6408 : BufferTag tag;
6409 : RelFileLocator currlocator;
6410 0 : Size nblocks = 1;
6411 :
6412 0 : cur = &wb_context->pending_writebacks[i];
6413 0 : tag = cur->tag;
6414 0 : currlocator = BufTagGetRelFileLocator(&tag);
6415 :
6416 : /*
6417 : * Peek ahead, into following writeback requests, to see if they can
6418 : * be combined with the current one.
6419 : */
6420 0 : for (ahead = 0; i + ahead + 1 < wb_context->nr_pending; ahead++)
6421 : {
6422 :
6423 0 : next = &wb_context->pending_writebacks[i + ahead + 1];
6424 :
6425 : /* different file, stop */
6426 0 : if (!RelFileLocatorEquals(currlocator,
6427 0 : BufTagGetRelFileLocator(&next->tag)) ||
6428 0 : BufTagGetForkNum(&cur->tag) != BufTagGetForkNum(&next->tag))
6429 : break;
6430 :
6431 : /* ok, block queued twice, skip */
6432 0 : if (cur->tag.blockNum == next->tag.blockNum)
6433 0 : continue;
6434 :
6435 : /* only merge consecutive writes */
6436 0 : if (cur->tag.blockNum + 1 != next->tag.blockNum)
6437 0 : break;
6438 :
6439 0 : nblocks++;
6440 0 : cur = next;
6441 : }
6442 :
6443 0 : i += ahead;
6444 :
6445 : /* and finally tell the kernel to write the data to storage */
6446 0 : reln = smgropen(currlocator, INVALID_PROC_NUMBER);
6447 0 : smgrwriteback(reln, BufTagGetForkNum(&tag), tag.blockNum, nblocks);
6448 : }
6449 :
6450 : /*
6451 : * Assume that writeback requests are only issued for buffers containing
6452 : * blocks of permanent relations.
6453 : */
6454 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
6455 0 : IOOP_WRITEBACK, io_start, wb_context->nr_pending, 0);
6456 :
6457 0 : wb_context->nr_pending = 0;
6458 : }
6459 :
6460 : /* ResourceOwner callbacks */
6461 :
6462 : static void
6463 30 : ResOwnerReleaseBufferIO(Datum res)
6464 : {
6465 30 : Buffer buffer = DatumGetInt32(res);
6466 :
6467 30 : AbortBufferIO(buffer);
6468 30 : }
6469 :
6470 : static char *
6471 0 : ResOwnerPrintBufferIO(Datum res)
6472 : {
6473 0 : Buffer buffer = DatumGetInt32(res);
6474 :
6475 0 : return psprintf("lost track of buffer IO on buffer %d", buffer);
6476 : }
6477 :
6478 : static void
6479 14504 : ResOwnerReleaseBufferPin(Datum res)
6480 : {
6481 14504 : Buffer buffer = DatumGetInt32(res);
6482 :
6483 : /* Like ReleaseBuffer, but don't call ResourceOwnerForgetBuffer */
6484 14504 : if (!BufferIsValid(buffer))
6485 0 : elog(ERROR, "bad buffer ID: %d", buffer);
6486 :
6487 14504 : if (BufferIsLocal(buffer))
6488 5970 : UnpinLocalBufferNoOwner(buffer);
6489 : else
6490 8534 : UnpinBufferNoOwner(GetBufferDescriptor(buffer - 1));
6491 14504 : }
6492 :
6493 : static char *
6494 0 : ResOwnerPrintBufferPin(Datum res)
6495 : {
6496 0 : return DebugPrintBufferRefcount(DatumGetInt32(res));
6497 : }
6498 :
6499 : /*
6500 : * Try to evict the current block in a shared buffer.
6501 : *
6502 : * This function is intended for testing/development use only!
6503 : *
6504 : * To succeed, the buffer must not be pinned on entry, so if the caller had a
6505 : * particular block in mind, it might already have been replaced by some other
6506 : * block by the time this function runs. It's also unpinned on return, so the
6507 : * buffer might be occupied again by the time control is returned, potentially
6508 : * even by the same block. This inherent raciness without other interlocking
6509 : * makes the function unsuitable for non-testing usage.
6510 : *
6511 : * Returns true if the buffer was valid and it has now been made invalid.
6512 : * Returns false if it wasn't valid, if it couldn't be evicted due to a pin,
6513 : * or if the buffer becomes dirty again while we're trying to write it out.
6514 : */
6515 : bool
6516 278 : EvictUnpinnedBuffer(Buffer buf)
6517 : {
6518 : BufferDesc *desc;
6519 : uint32 buf_state;
6520 : bool result;
6521 :
6522 : /* Make sure we can pin the buffer. */
6523 278 : ResourceOwnerEnlarge(CurrentResourceOwner);
6524 278 : ReservePrivateRefCountEntry();
6525 :
6526 : Assert(!BufferIsLocal(buf));
6527 278 : desc = GetBufferDescriptor(buf - 1);
6528 :
6529 : /* Lock the header and check if it's valid. */
6530 278 : buf_state = LockBufHdr(desc);
6531 278 : if ((buf_state & BM_VALID) == 0)
6532 : {
6533 0 : UnlockBufHdr(desc, buf_state);
6534 0 : return false;
6535 : }
6536 :
6537 : /* Check that it's not pinned already. */
6538 278 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
6539 : {
6540 0 : UnlockBufHdr(desc, buf_state);
6541 0 : return false;
6542 : }
6543 :
6544 278 : PinBuffer_Locked(desc); /* releases spinlock */
6545 :
6546 : /* If it was dirty, try to clean it once. */
6547 278 : if (buf_state & BM_DIRTY)
6548 : {
6549 36 : LWLockAcquire(BufferDescriptorGetContentLock(desc), LW_SHARED);
6550 36 : FlushBuffer(desc, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
6551 36 : LWLockRelease(BufferDescriptorGetContentLock(desc));
6552 : }
6553 :
6554 : /* This will return false if it becomes dirty or someone else pins it. */
6555 278 : result = InvalidateVictimBuffer(desc);
6556 :
6557 278 : UnpinBuffer(desc);
6558 :
6559 278 : return result;
6560 : }
6561 :
6562 : /*
6563 : * Generic implementation of the AIO handle staging callback for readv/writev
6564 : * on local/shared buffers.
6565 : *
6566 : * Each readv/writev can target multiple buffers. The buffers have already
6567 : * been registered with the IO handle.
6568 : *
6569 : * To make the IO ready for execution ("staging"), we need to ensure that the
6570 : * targeted buffers are in an appropriate state while the IO is ongoing. For
6571 : * that the AIO subsystem needs to have its own buffer pin, otherwise an error
6572 : * in this backend could lead to this backend's buffer pin being released as
6573 : * part of error handling, which in turn could lead to the buffer being
6574 : * replaced while IO is ongoing.
6575 : */
6576 : static pg_attribute_always_inline void
6577 2409832 : buffer_stage_common(PgAioHandle *ioh, bool is_write, bool is_temp)
6578 : {
6579 : uint64 *io_data;
6580 : uint8 handle_data_len;
6581 : PgAioWaitRef io_ref;
6582 2409832 : BufferTag first PG_USED_FOR_ASSERTS_ONLY = {0};
6583 :
6584 2409832 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
6585 :
6586 2409832 : pgaio_io_get_wref(ioh, &io_ref);
6587 :
6588 : /* iterate over all buffers affected by the vectored readv/writev */
6589 5147788 : for (int i = 0; i < handle_data_len; i++)
6590 : {
6591 2737956 : Buffer buffer = (Buffer) io_data[i];
6592 2737956 : BufferDesc *buf_hdr = is_temp ?
6593 16778 : GetLocalBufferDescriptor(-buffer - 1)
6594 2737956 : : GetBufferDescriptor(buffer - 1);
6595 : uint32 buf_state;
6596 :
6597 : /*
6598 : * Check that all the buffers are actually ones that could conceivably
6599 : * be done in one IO, i.e. are sequential. This is the last
6600 : * buffer-aware code before IO is actually executed and confusion
6601 : * about which buffers are targeted by IO can be hard to debug, making
6602 : * it worth doing extra-paranoid checks.
6603 : */
6604 2737956 : if (i == 0)
6605 2409832 : first = buf_hdr->tag;
6606 : else
6607 : {
6608 : Assert(buf_hdr->tag.relNumber == first.relNumber);
6609 : Assert(buf_hdr->tag.blockNum == first.blockNum + i);
6610 : }
6611 :
6612 2737956 : if (is_temp)
6613 16778 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
6614 : else
6615 2721178 : buf_state = LockBufHdr(buf_hdr);
6616 :
6617 : /* verify the buffer is in the expected state */
6618 : Assert(buf_state & BM_TAG_VALID);
6619 : if (is_write)
6620 : {
6621 : Assert(buf_state & BM_VALID);
6622 : Assert(buf_state & BM_DIRTY);
6623 : }
6624 : else
6625 : {
6626 : Assert(!(buf_state & BM_VALID));
6627 : Assert(!(buf_state & BM_DIRTY));
6628 : }
6629 :
6630 : /* temp buffers don't use BM_IO_IN_PROGRESS */
6631 2737956 : if (!is_temp)
6632 : Assert(buf_state & BM_IO_IN_PROGRESS);
6633 :
6634 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) >= 1);
6635 :
6636 : /*
6637 : * Reflect that the buffer is now owned by the AIO subsystem.
6638 : *
6639 : * For local buffers: This can't be done just via LocalRefCount, as
6640 : * one might initially think, as this backend could error out while
6641 : * AIO is still in progress, releasing all the pins by the backend
6642 : * itself.
6643 : *
6644 : * This pin is released again in TerminateBufferIO().
6645 : */
6646 2737956 : buf_state += BUF_REFCOUNT_ONE;
6647 2737956 : buf_hdr->io_wref = io_ref;
6648 :
6649 2737956 : if (is_temp)
6650 16778 : pg_atomic_unlocked_write_u32(&buf_hdr->state, buf_state);
6651 : else
6652 2721178 : UnlockBufHdr(buf_hdr, buf_state);
6653 :
6654 : /*
6655 : * Ensure the content lock that prevents buffer modifications while
6656 : * the buffer is being written out is not released early due to an
6657 : * error.
6658 : */
6659 2737956 : if (is_write && !is_temp)
6660 : {
6661 : LWLock *content_lock;
6662 :
6663 0 : content_lock = BufferDescriptorGetContentLock(buf_hdr);
6664 :
6665 : Assert(LWLockHeldByMe(content_lock));
6666 :
6667 : /*
6668 : * Lock is now owned by AIO subsystem.
6669 : */
6670 0 : LWLockDisown(content_lock);
6671 : }
6672 :
6673 : /*
6674 : * Stop tracking this buffer via the resowner - the AIO system now
6675 : * keeps track.
6676 : */
6677 2737956 : if (!is_temp)
6678 2721178 : ResourceOwnerForgetBufferIO(CurrentResourceOwner, buffer);
6679 : }
6680 2409832 : }
6681 :
6682 : /*
6683 : * Decode readv errors as encoded by buffer_readv_encode_error().
6684 : */
6685 : static inline void
6686 698 : buffer_readv_decode_error(PgAioResult result,
6687 : bool *zeroed_any,
6688 : bool *ignored_any,
6689 : uint8 *zeroed_or_error_count,
6690 : uint8 *checkfail_count,
6691 : uint8 *first_off)
6692 : {
6693 698 : uint32 rem_error = result.error_data;
6694 :
6695 : /* see static asserts in buffer_readv_encode_error */
6696 : #define READV_COUNT_BITS 7
6697 : #define READV_COUNT_MASK ((1 << READV_COUNT_BITS) - 1)
6698 :
6699 698 : *zeroed_any = rem_error & 1;
6700 698 : rem_error >>= 1;
6701 :
6702 698 : *ignored_any = rem_error & 1;
6703 698 : rem_error >>= 1;
6704 :
6705 698 : *zeroed_or_error_count = rem_error & READV_COUNT_MASK;
6706 698 : rem_error >>= READV_COUNT_BITS;
6707 :
6708 698 : *checkfail_count = rem_error & READV_COUNT_MASK;
6709 698 : rem_error >>= READV_COUNT_BITS;
6710 :
6711 698 : *first_off = rem_error & READV_COUNT_MASK;
6712 698 : rem_error >>= READV_COUNT_BITS;
6713 698 : }
6714 :
6715 : /*
6716 : * Helper to encode errors for buffer_readv_complete()
6717 : *
6718 : * Errors are encoded as follows:
6719 : * - bit 0 indicates whether any page was zeroed (1) or not (0)
6720 : * - bit 1 indicates whether any checksum failure was ignored (1) or not (0)
6721 : * - next READV_COUNT_BITS bits indicate the number of errored or zeroed pages
6722 : * - next READV_COUNT_BITS bits indicate the number of checksum failures
6723 : * - next READV_COUNT_BITS bits indicate the first offset of the first page
6724 : * that was errored or zeroed or, if no errors/zeroes, the first ignored
6725 : * checksum
6726 : */
6727 : static inline void
6728 384 : buffer_readv_encode_error(PgAioResult *result,
6729 : bool is_temp,
6730 : bool zeroed_any,
6731 : bool ignored_any,
6732 : uint8 error_count,
6733 : uint8 zeroed_count,
6734 : uint8 checkfail_count,
6735 : uint8 first_error_off,
6736 : uint8 first_zeroed_off,
6737 : uint8 first_ignored_off)
6738 : {
6739 :
6740 384 : uint8 shift = 0;
6741 384 : uint8 zeroed_or_error_count =
6742 : error_count > 0 ? error_count : zeroed_count;
6743 : uint8 first_off;
6744 :
6745 : StaticAssertStmt(PG_IOV_MAX <= 1 << READV_COUNT_BITS,
6746 : "PG_IOV_MAX is bigger than reserved space for error data");
6747 : StaticAssertStmt((1 + 1 + 3 * READV_COUNT_BITS) <= PGAIO_RESULT_ERROR_BITS,
6748 : "PGAIO_RESULT_ERROR_BITS is insufficient for buffer_readv");
6749 :
6750 : /*
6751 : * We only have space to encode one offset - but luckily that's good
6752 : * enough. If there is an error, the error is the interesting offset, same
6753 : * with a zeroed buffer vs an ignored buffer.
6754 : */
6755 384 : if (error_count > 0)
6756 188 : first_off = first_error_off;
6757 196 : else if (zeroed_count > 0)
6758 160 : first_off = first_zeroed_off;
6759 : else
6760 36 : first_off = first_ignored_off;
6761 :
6762 : Assert(!zeroed_any || error_count == 0);
6763 :
6764 384 : result->error_data = 0;
6765 :
6766 384 : result->error_data |= zeroed_any << shift;
6767 384 : shift += 1;
6768 :
6769 384 : result->error_data |= ignored_any << shift;
6770 384 : shift += 1;
6771 :
6772 384 : result->error_data |= ((uint32) zeroed_or_error_count) << shift;
6773 384 : shift += READV_COUNT_BITS;
6774 :
6775 384 : result->error_data |= ((uint32) checkfail_count) << shift;
6776 384 : shift += READV_COUNT_BITS;
6777 :
6778 384 : result->error_data |= ((uint32) first_off) << shift;
6779 384 : shift += READV_COUNT_BITS;
6780 :
6781 384 : result->id = is_temp ? PGAIO_HCB_LOCAL_BUFFER_READV :
6782 : PGAIO_HCB_SHARED_BUFFER_READV;
6783 :
6784 384 : if (error_count > 0)
6785 188 : result->status = PGAIO_RS_ERROR;
6786 : else
6787 196 : result->status = PGAIO_RS_WARNING;
6788 :
6789 : /*
6790 : * The encoding is complicated enough to warrant cross-checking it against
6791 : * the decode function.
6792 : */
6793 : #ifdef USE_ASSERT_CHECKING
6794 : {
6795 : bool zeroed_any_2,
6796 : ignored_any_2;
6797 : uint8 zeroed_or_error_count_2,
6798 : checkfail_count_2,
6799 : first_off_2;
6800 :
6801 : buffer_readv_decode_error(*result,
6802 : &zeroed_any_2, &ignored_any_2,
6803 : &zeroed_or_error_count_2,
6804 : &checkfail_count_2,
6805 : &first_off_2);
6806 : Assert(zeroed_any == zeroed_any_2);
6807 : Assert(ignored_any == ignored_any_2);
6808 : Assert(zeroed_or_error_count == zeroed_or_error_count_2);
6809 : Assert(checkfail_count == checkfail_count_2);
6810 : Assert(first_off == first_off_2);
6811 : }
6812 : #endif
6813 :
6814 : #undef READV_COUNT_BITS
6815 : #undef READV_COUNT_MASK
6816 384 : }
6817 :
6818 : /*
6819 : * Helper for AIO readv completion callbacks, supporting both shared and temp
6820 : * buffers. Gets called once for each buffer in a multi-page read.
6821 : */
6822 : static pg_attribute_always_inline void
6823 2456804 : buffer_readv_complete_one(PgAioTargetData *td, uint8 buf_off, Buffer buffer,
6824 : uint8 flags, bool failed, bool is_temp,
6825 : bool *buffer_invalid,
6826 : bool *failed_checksum,
6827 : bool *ignored_checksum,
6828 : bool *zeroed_buffer)
6829 : {
6830 2456804 : BufferDesc *buf_hdr = is_temp ?
6831 16778 : GetLocalBufferDescriptor(-buffer - 1)
6832 2456804 : : GetBufferDescriptor(buffer - 1);
6833 2456804 : BufferTag tag = buf_hdr->tag;
6834 2456804 : char *bufdata = BufferGetBlock(buffer);
6835 : uint32 set_flag_bits;
6836 : int piv_flags;
6837 :
6838 : /* check that the buffer is in the expected state for a read */
6839 : #ifdef USE_ASSERT_CHECKING
6840 : {
6841 : uint32 buf_state = pg_atomic_read_u32(&buf_hdr->state);
6842 :
6843 : Assert(buf_state & BM_TAG_VALID);
6844 : Assert(!(buf_state & BM_VALID));
6845 : /* temp buffers don't use BM_IO_IN_PROGRESS */
6846 : if (!is_temp)
6847 : Assert(buf_state & BM_IO_IN_PROGRESS);
6848 : Assert(!(buf_state & BM_DIRTY));
6849 : }
6850 : #endif
6851 :
6852 2456804 : *buffer_invalid = false;
6853 2456804 : *failed_checksum = false;
6854 2456804 : *ignored_checksum = false;
6855 2456804 : *zeroed_buffer = false;
6856 :
6857 : /*
6858 : * We ask PageIsVerified() to only log the message about checksum errors,
6859 : * as the completion might be run in any backend (or IO workers). We will
6860 : * report checksum errors in buffer_readv_report().
6861 : */
6862 2456804 : piv_flags = PIV_LOG_LOG;
6863 :
6864 : /* the local zero_damaged_pages may differ from the definer's */
6865 2456804 : if (flags & READ_BUFFERS_IGNORE_CHECKSUM_FAILURES)
6866 76 : piv_flags |= PIV_IGNORE_CHECKSUM_FAILURE;
6867 :
6868 : /* Check for garbage data. */
6869 2456804 : if (!failed)
6870 : {
6871 : PgAioResult result_one;
6872 :
6873 2456746 : if (!PageIsVerified((Page) bufdata, tag.blockNum, piv_flags,
6874 : failed_checksum))
6875 : {
6876 192 : if (flags & READ_BUFFERS_ZERO_ON_ERROR)
6877 : {
6878 92 : memset(bufdata, 0, BLCKSZ);
6879 92 : *zeroed_buffer = true;
6880 : }
6881 : else
6882 : {
6883 100 : *buffer_invalid = true;
6884 : /* mark buffer as having failed */
6885 100 : failed = true;
6886 : }
6887 : }
6888 2456554 : else if (*failed_checksum)
6889 24 : *ignored_checksum = true;
6890 :
6891 : /*
6892 : * Immediately log a message about the invalid page, but only to the
6893 : * server log. The reason to do so immediately is that this may be
6894 : * executed in a different backend than the one that originated the
6895 : * request. The reason to do so immediately is that the originator
6896 : * might not process the query result immediately (because it is busy
6897 : * doing another part of query processing) or at all (e.g. if it was
6898 : * cancelled or errored out due to another IO also failing). The
6899 : * definer of the IO will emit an ERROR or WARNING when processing the
6900 : * IO's results
6901 : *
6902 : * To avoid duplicating the code to emit these log messages, we reuse
6903 : * buffer_readv_report().
6904 : */
6905 2456746 : if (*buffer_invalid || *failed_checksum || *zeroed_buffer)
6906 : {
6907 216 : buffer_readv_encode_error(&result_one, is_temp,
6908 216 : *zeroed_buffer,
6909 216 : *ignored_checksum,
6910 216 : *buffer_invalid,
6911 216 : *zeroed_buffer ? 1 : 0,
6912 216 : *failed_checksum ? 1 : 0,
6913 : buf_off, buf_off, buf_off);
6914 216 : pgaio_result_report(result_one, td, LOG_SERVER_ONLY);
6915 : }
6916 : }
6917 :
6918 : /* Terminate I/O and set BM_VALID. */
6919 2456804 : set_flag_bits = failed ? BM_IO_ERROR : BM_VALID;
6920 2456804 : if (is_temp)
6921 16778 : TerminateLocalBufferIO(buf_hdr, false, set_flag_bits, true);
6922 : else
6923 2440026 : TerminateBufferIO(buf_hdr, false, set_flag_bits, false, true);
6924 :
6925 : /*
6926 : * Call the BUFFER_READ_DONE tracepoint in the callback, even though the
6927 : * callback may not be executed in the same backend that called
6928 : * BUFFER_READ_START. The alternative would be to defer calling the
6929 : * tracepoint to a later point (e.g. the local completion callback for
6930 : * shared buffer reads), which seems even less helpful.
6931 : */
6932 : TRACE_POSTGRESQL_BUFFER_READ_DONE(tag.forkNum,
6933 : tag.blockNum,
6934 : tag.spcOid,
6935 : tag.dbOid,
6936 : tag.relNumber,
6937 : is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
6938 : false);
6939 2456804 : }
6940 :
6941 : /*
6942 : * Perform completion handling of a single AIO read. This read may cover
6943 : * multiple blocks / buffers.
6944 : *
6945 : * Shared between shared and local buffers, to reduce code duplication.
6946 : */
6947 : static pg_attribute_always_inline PgAioResult
6948 2199768 : buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
6949 : uint8 cb_data, bool is_temp)
6950 : {
6951 2199768 : PgAioResult result = prior_result;
6952 2199768 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
6953 2199768 : uint8 first_error_off = 0;
6954 2199768 : uint8 first_zeroed_off = 0;
6955 2199768 : uint8 first_ignored_off = 0;
6956 2199768 : uint8 error_count = 0;
6957 2199768 : uint8 zeroed_count = 0;
6958 2199768 : uint8 ignored_count = 0;
6959 2199768 : uint8 checkfail_count = 0;
6960 : uint64 *io_data;
6961 : uint8 handle_data_len;
6962 :
6963 : if (is_temp)
6964 : {
6965 : Assert(td->smgr.is_temp);
6966 : Assert(pgaio_io_get_owner(ioh) == MyProcNumber);
6967 : }
6968 : else
6969 : Assert(!td->smgr.is_temp);
6970 :
6971 : /*
6972 : * Iterate over all the buffers affected by this IO and call the
6973 : * per-buffer completion function for each buffer.
6974 : */
6975 2199768 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
6976 4656572 : for (uint8 buf_off = 0; buf_off < handle_data_len; buf_off++)
6977 : {
6978 2456804 : Buffer buf = io_data[buf_off];
6979 : bool failed;
6980 2456804 : bool failed_verification = false;
6981 2456804 : bool failed_checksum = false;
6982 2456804 : bool zeroed_buffer = false;
6983 2456804 : bool ignored_checksum = false;
6984 :
6985 : Assert(BufferIsValid(buf));
6986 :
6987 : /*
6988 : * If the entire I/O failed on a lower-level, each buffer needs to be
6989 : * marked as failed. In case of a partial read, the first few buffers
6990 : * may be ok.
6991 : */
6992 2456804 : failed =
6993 2456804 : prior_result.status == PGAIO_RS_ERROR
6994 2456804 : || prior_result.result <= buf_off;
6995 :
6996 2456804 : buffer_readv_complete_one(td, buf_off, buf, cb_data, failed, is_temp,
6997 : &failed_verification,
6998 : &failed_checksum,
6999 : &ignored_checksum,
7000 : &zeroed_buffer);
7001 :
7002 : /*
7003 : * Track information about the number of different kinds of error
7004 : * conditions across all pages, as there can be multiple pages failing
7005 : * verification as part of one IO.
7006 : */
7007 2456804 : if (failed_verification && !zeroed_buffer && error_count++ == 0)
7008 88 : first_error_off = buf_off;
7009 2456804 : if (zeroed_buffer && zeroed_count++ == 0)
7010 68 : first_zeroed_off = buf_off;
7011 2456804 : if (ignored_checksum && ignored_count++ == 0)
7012 20 : first_ignored_off = buf_off;
7013 2456804 : if (failed_checksum)
7014 64 : checkfail_count++;
7015 : }
7016 :
7017 : /*
7018 : * If the smgr read succeeded [partially] and page verification failed for
7019 : * some of the pages, adjust the IO's result state appropriately.
7020 : */
7021 2199768 : if (prior_result.status != PGAIO_RS_ERROR &&
7022 2199662 : (error_count > 0 || ignored_count > 0 || zeroed_count > 0))
7023 : {
7024 168 : buffer_readv_encode_error(&result, is_temp,
7025 : zeroed_count > 0, ignored_count > 0,
7026 : error_count, zeroed_count, checkfail_count,
7027 : first_error_off, first_zeroed_off,
7028 : first_ignored_off);
7029 168 : pgaio_result_report(result, td, DEBUG1);
7030 : }
7031 :
7032 : /*
7033 : * For shared relations this reporting is done in
7034 : * shared_buffer_readv_complete_local().
7035 : */
7036 2199768 : if (is_temp && checkfail_count > 0)
7037 4 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
7038 : checkfail_count);
7039 :
7040 2199768 : return result;
7041 : }
7042 :
7043 : /*
7044 : * AIO error reporting callback for aio_shared_buffer_readv_cb and
7045 : * aio_local_buffer_readv_cb.
7046 : *
7047 : * The error is encoded / decoded in buffer_readv_encode_error() /
7048 : * buffer_readv_decode_error().
7049 : */
7050 : static void
7051 544 : buffer_readv_report(PgAioResult result, const PgAioTargetData *td,
7052 : int elevel)
7053 : {
7054 544 : int nblocks = td->smgr.nblocks;
7055 544 : BlockNumber first = td->smgr.blockNum;
7056 544 : BlockNumber last = first + nblocks - 1;
7057 544 : ProcNumber errProc =
7058 544 : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER;
7059 : RelPathStr rpath =
7060 544 : relpathbackend(td->smgr.rlocator, errProc, td->smgr.forkNum);
7061 : bool zeroed_any,
7062 : ignored_any;
7063 : uint8 zeroed_or_error_count,
7064 : checkfail_count,
7065 : first_off;
7066 : uint8 affected_count;
7067 : const char *msg_one,
7068 : *msg_mult,
7069 : *det_mult,
7070 : *hint_mult;
7071 :
7072 544 : buffer_readv_decode_error(result, &zeroed_any, &ignored_any,
7073 : &zeroed_or_error_count,
7074 : &checkfail_count,
7075 : &first_off);
7076 :
7077 : /*
7078 : * Treat a read that had both zeroed buffers *and* ignored checksums as a
7079 : * special case, it's too irregular to be emitted the same way as the
7080 : * other cases.
7081 : */
7082 544 : if (zeroed_any && ignored_any)
7083 : {
7084 : Assert(zeroed_any && ignored_any);
7085 : Assert(nblocks > 1); /* same block can't be both zeroed and ignored */
7086 : Assert(result.status != PGAIO_RS_ERROR);
7087 8 : affected_count = zeroed_or_error_count;
7088 :
7089 8 : ereport(elevel,
7090 : errcode(ERRCODE_DATA_CORRUPTED),
7091 : errmsg("zeroing %u page(s) and ignoring %u checksum failure(s) among blocks %u..%u of relation %s",
7092 : affected_count, checkfail_count, first, last, rpath.str),
7093 : affected_count > 1 ?
7094 : errdetail("Block %u held first zeroed page.",
7095 : first + first_off) : 0,
7096 : errhint("See server log for details about the other %u invalid block(s).",
7097 : affected_count + checkfail_count - 1));
7098 8 : return;
7099 : }
7100 :
7101 : /*
7102 : * The other messages are highly repetitive. To avoid duplicating a long
7103 : * and complicated ereport(), gather the translated format strings
7104 : * separately and then do one common ereport.
7105 : */
7106 536 : if (result.status == PGAIO_RS_ERROR)
7107 : {
7108 : Assert(!zeroed_any); /* can't have invalid pages when zeroing them */
7109 272 : affected_count = zeroed_or_error_count;
7110 272 : msg_one = _("invalid page in block %u of relation %s");
7111 272 : msg_mult = _("%u invalid pages among blocks %u..%u of relation %s");
7112 272 : det_mult = _("Block %u held first invalid page.");
7113 272 : hint_mult = _("See server log for the other %u invalid block(s).");
7114 : }
7115 264 : else if (zeroed_any && !ignored_any)
7116 : {
7117 216 : affected_count = zeroed_or_error_count;
7118 216 : msg_one = _("invalid page in block %u of relation %s; zeroing out page");
7119 216 : msg_mult = _("zeroing out %u invalid pages among blocks %u..%u of relation %s");
7120 216 : det_mult = _("Block %u held first zeroed page.");
7121 216 : hint_mult = _("See server log for the other %u zeroed block(s).");
7122 : }
7123 48 : else if (!zeroed_any && ignored_any)
7124 : {
7125 48 : affected_count = checkfail_count;
7126 48 : msg_one = _("ignoring checksum failure in block %u of relation %s");
7127 48 : msg_mult = _("ignoring %u checksum failures among blocks %u..%u of relation %s");
7128 48 : det_mult = _("Block %u held first ignored page.");
7129 48 : hint_mult = _("See server log for the other %u ignored block(s).");
7130 : }
7131 : else
7132 0 : pg_unreachable();
7133 :
7134 536 : ereport(elevel,
7135 : errcode(ERRCODE_DATA_CORRUPTED),
7136 : affected_count == 1 ?
7137 : errmsg_internal(msg_one, first + first_off, rpath.str) :
7138 : errmsg_internal(msg_mult, affected_count, first, last, rpath.str),
7139 : affected_count > 1 ? errdetail_internal(det_mult, first + first_off) : 0,
7140 : affected_count > 1 ? errhint_internal(hint_mult, affected_count - 1) : 0);
7141 : }
7142 :
7143 : static void
7144 2406260 : shared_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
7145 : {
7146 2406260 : buffer_stage_common(ioh, false, false);
7147 2406260 : }
7148 :
7149 : static PgAioResult
7150 2196196 : shared_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7151 : uint8 cb_data)
7152 : {
7153 2196196 : return buffer_readv_complete(ioh, prior_result, cb_data, false);
7154 : }
7155 :
7156 : /*
7157 : * We need a backend-local completion callback for shared buffers, to be able
7158 : * to report checksum errors correctly. Unfortunately that can only safely
7159 : * happen if the reporting backend has previously called
7160 : * pgstat_prepare_report_checksum_failure(), which we can only guarantee in
7161 : * the backend that started the IO. Hence this callback.
7162 : */
7163 : static PgAioResult
7164 2406260 : shared_buffer_readv_complete_local(PgAioHandle *ioh, PgAioResult prior_result,
7165 : uint8 cb_data)
7166 : {
7167 : bool zeroed_any,
7168 : ignored_any;
7169 : uint8 zeroed_or_error_count,
7170 : checkfail_count,
7171 : first_off;
7172 :
7173 2406260 : if (prior_result.status == PGAIO_RS_OK)
7174 2406106 : return prior_result;
7175 :
7176 154 : buffer_readv_decode_error(prior_result,
7177 : &zeroed_any,
7178 : &ignored_any,
7179 : &zeroed_or_error_count,
7180 : &checkfail_count,
7181 : &first_off);
7182 :
7183 154 : if (checkfail_count)
7184 : {
7185 48 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
7186 :
7187 48 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
7188 : checkfail_count);
7189 : }
7190 :
7191 154 : return prior_result;
7192 : }
7193 :
7194 : static void
7195 3572 : local_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
7196 : {
7197 3572 : buffer_stage_common(ioh, false, true);
7198 3572 : }
7199 :
7200 : static PgAioResult
7201 3572 : local_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7202 : uint8 cb_data)
7203 : {
7204 3572 : return buffer_readv_complete(ioh, prior_result, cb_data, true);
7205 : }
7206 :
7207 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
7208 : const PgAioHandleCallbacks aio_shared_buffer_readv_cb = {
7209 : .stage = shared_buffer_readv_stage,
7210 : .complete_shared = shared_buffer_readv_complete,
7211 : /* need a local callback to report checksum failures */
7212 : .complete_local = shared_buffer_readv_complete_local,
7213 : .report = buffer_readv_report,
7214 : };
7215 :
7216 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
7217 : const PgAioHandleCallbacks aio_local_buffer_readv_cb = {
7218 : .stage = local_buffer_readv_stage,
7219 :
7220 : /*
7221 : * Note that this, in contrast to the shared_buffers case, uses
7222 : * complete_local, as only the issuing backend has access to the required
7223 : * datastructures. This is important in case the IO completion may be
7224 : * consumed incidentally by another backend.
7225 : */
7226 : .complete_local = local_buffer_readv_complete,
7227 : .report = buffer_readv_report,
7228 : };
|
