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 : #ifdef USE_ASSERT_CHECKING
44 : #include "catalog/pg_tablespace_d.h"
45 : #endif
46 : #include "catalog/storage.h"
47 : #include "catalog/storage_xlog.h"
48 : #include "executor/instrument.h"
49 : #include "lib/binaryheap.h"
50 : #include "miscadmin.h"
51 : #include "pg_trace.h"
52 : #include "pgstat.h"
53 : #include "postmaster/bgwriter.h"
54 : #include "storage/aio.h"
55 : #include "storage/buf_internals.h"
56 : #include "storage/bufmgr.h"
57 : #include "storage/fd.h"
58 : #include "storage/ipc.h"
59 : #include "storage/lmgr.h"
60 : #include "storage/proc.h"
61 : #include "storage/read_stream.h"
62 : #include "storage/smgr.h"
63 : #include "storage/standby.h"
64 : #include "utils/memdebug.h"
65 : #include "utils/ps_status.h"
66 : #include "utils/rel.h"
67 : #include "utils/resowner.h"
68 : #include "utils/timestamp.h"
69 :
70 :
71 : /* Note: these two macros only work on shared buffers, not local ones! */
72 : #define BufHdrGetBlock(bufHdr) ((Block) (BufferBlocks + ((Size) (bufHdr)->buf_id) * BLCKSZ))
73 : #define BufferGetLSN(bufHdr) (PageGetLSN(BufHdrGetBlock(bufHdr)))
74 :
75 : /* Note: this macro only works on local buffers, not shared ones! */
76 : #define LocalBufHdrGetBlock(bufHdr) \
77 : LocalBufferBlockPointers[-((bufHdr)->buf_id + 2)]
78 :
79 : /* Bits in SyncOneBuffer's return value */
80 : #define BUF_WRITTEN 0x01
81 : #define BUF_REUSABLE 0x02
82 :
83 : #define RELS_BSEARCH_THRESHOLD 20
84 :
85 : /*
86 : * This is the size (in the number of blocks) above which we scan the
87 : * entire buffer pool to remove the buffers for all the pages of relation
88 : * being dropped. For the relations with size below this threshold, we find
89 : * the buffers by doing lookups in BufMapping table.
90 : */
91 : #define BUF_DROP_FULL_SCAN_THRESHOLD (uint64) (NBuffers / 32)
92 :
93 : typedef struct PrivateRefCountEntry
94 : {
95 : Buffer buffer;
96 : int32 refcount;
97 : } PrivateRefCountEntry;
98 :
99 : /* 64 bytes, about the size of a cache line on common systems */
100 : #define REFCOUNT_ARRAY_ENTRIES 8
101 :
102 : /*
103 : * Status of buffers to checkpoint for a particular tablespace, used
104 : * internally in BufferSync.
105 : */
106 : typedef struct CkptTsStatus
107 : {
108 : /* oid of the tablespace */
109 : Oid tsId;
110 :
111 : /*
112 : * Checkpoint progress for this tablespace. To make progress comparable
113 : * between tablespaces the progress is, for each tablespace, measured as a
114 : * number between 0 and the total number of to-be-checkpointed pages. Each
115 : * page checkpointed in this tablespace increments this space's progress
116 : * by progress_slice.
117 : */
118 : float8 progress;
119 : float8 progress_slice;
120 :
121 : /* number of to-be checkpointed pages in this tablespace */
122 : int num_to_scan;
123 : /* already processed pages in this tablespace */
124 : int num_scanned;
125 :
126 : /* current offset in CkptBufferIds for this tablespace */
127 : int index;
128 : } CkptTsStatus;
129 :
130 : /*
131 : * Type for array used to sort SMgrRelations
132 : *
133 : * FlushRelationsAllBuffers shares the same comparator function with
134 : * DropRelationsAllBuffers. Pointer to this struct and RelFileLocator must be
135 : * compatible.
136 : */
137 : typedef struct SMgrSortArray
138 : {
139 : RelFileLocator rlocator; /* This must be the first member */
140 : SMgrRelation srel;
141 : } SMgrSortArray;
142 :
143 : /* GUC variables */
144 : bool zero_damaged_pages = false;
145 : int bgwriter_lru_maxpages = 100;
146 : double bgwriter_lru_multiplier = 2.0;
147 : bool track_io_timing = false;
148 :
149 : /*
150 : * How many buffers PrefetchBuffer callers should try to stay ahead of their
151 : * ReadBuffer calls by. Zero means "never prefetch". This value is only used
152 : * for buffers not belonging to tablespaces that have their
153 : * effective_io_concurrency parameter set.
154 : */
155 : int effective_io_concurrency = DEFAULT_EFFECTIVE_IO_CONCURRENCY;
156 :
157 : /*
158 : * Like effective_io_concurrency, but used by maintenance code paths that might
159 : * benefit from a higher setting because they work on behalf of many sessions.
160 : * Overridden by the tablespace setting of the same name.
161 : */
162 : int maintenance_io_concurrency = DEFAULT_MAINTENANCE_IO_CONCURRENCY;
163 :
164 : /*
165 : * Limit on how many blocks should be handled in single I/O operations.
166 : * StartReadBuffers() callers should respect it, as should other operations
167 : * that call smgr APIs directly. It is computed as the minimum of underlying
168 : * GUCs io_combine_limit_guc and io_max_combine_limit.
169 : */
170 : int io_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
171 : int io_combine_limit_guc = DEFAULT_IO_COMBINE_LIMIT;
172 : int io_max_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
173 :
174 : /*
175 : * GUC variables about triggering kernel writeback for buffers written; OS
176 : * dependent defaults are set via the GUC mechanism.
177 : */
178 : int checkpoint_flush_after = DEFAULT_CHECKPOINT_FLUSH_AFTER;
179 : int bgwriter_flush_after = DEFAULT_BGWRITER_FLUSH_AFTER;
180 : int backend_flush_after = DEFAULT_BACKEND_FLUSH_AFTER;
181 :
182 : /* local state for LockBufferForCleanup */
183 : static BufferDesc *PinCountWaitBuf = NULL;
184 :
185 : /*
186 : * Backend-Private refcount management:
187 : *
188 : * Each buffer also has a private refcount that keeps track of the number of
189 : * times the buffer is pinned in the current process. This is so that the
190 : * shared refcount needs to be modified only once if a buffer is pinned more
191 : * than once by an individual backend. It's also used to check that no buffers
192 : * are still pinned at the end of transactions and when exiting.
193 : *
194 : *
195 : * To avoid - as we used to - requiring an array with NBuffers entries to keep
196 : * track of local buffers, we use a small sequentially searched array
197 : * (PrivateRefCountArray) and an overflow hash table (PrivateRefCountHash) to
198 : * keep track of backend local pins.
199 : *
200 : * Until no more than REFCOUNT_ARRAY_ENTRIES buffers are pinned at once, all
201 : * refcounts are kept track of in the array; after that, new array entries
202 : * displace old ones into the hash table. That way a frequently used entry
203 : * can't get "stuck" in the hashtable while infrequent ones clog the array.
204 : *
205 : * Note that in most scenarios the number of pinned buffers will not exceed
206 : * REFCOUNT_ARRAY_ENTRIES.
207 : *
208 : *
209 : * To enter a buffer into the refcount tracking mechanism first reserve a free
210 : * entry using ReservePrivateRefCountEntry() and then later, if necessary,
211 : * fill it with NewPrivateRefCountEntry(). That split lets us avoid doing
212 : * memory allocations in NewPrivateRefCountEntry() which can be important
213 : * because in some scenarios it's called with a spinlock held...
214 : */
215 : static struct PrivateRefCountEntry PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES];
216 : static HTAB *PrivateRefCountHash = NULL;
217 : static int32 PrivateRefCountOverflowed = 0;
218 : static uint32 PrivateRefCountClock = 0;
219 : static PrivateRefCountEntry *ReservedRefCountEntry = NULL;
220 :
221 : static uint32 MaxProportionalPins;
222 :
223 : static void ReservePrivateRefCountEntry(void);
224 : static PrivateRefCountEntry *NewPrivateRefCountEntry(Buffer buffer);
225 : static PrivateRefCountEntry *GetPrivateRefCountEntry(Buffer buffer, bool do_move);
226 : static inline int32 GetPrivateRefCount(Buffer buffer);
227 : static void ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref);
228 :
229 : /* ResourceOwner callbacks to hold in-progress I/Os and buffer pins */
230 : static void ResOwnerReleaseBufferIO(Datum res);
231 : static char *ResOwnerPrintBufferIO(Datum res);
232 : static void ResOwnerReleaseBufferPin(Datum res);
233 : static char *ResOwnerPrintBufferPin(Datum res);
234 :
235 : const ResourceOwnerDesc buffer_io_resowner_desc =
236 : {
237 : .name = "buffer io",
238 : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
239 : .release_priority = RELEASE_PRIO_BUFFER_IOS,
240 : .ReleaseResource = ResOwnerReleaseBufferIO,
241 : .DebugPrint = ResOwnerPrintBufferIO
242 : };
243 :
244 : const ResourceOwnerDesc buffer_pin_resowner_desc =
245 : {
246 : .name = "buffer pin",
247 : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
248 : .release_priority = RELEASE_PRIO_BUFFER_PINS,
249 : .ReleaseResource = ResOwnerReleaseBufferPin,
250 : .DebugPrint = ResOwnerPrintBufferPin
251 : };
252 :
253 : /*
254 : * Ensure that the PrivateRefCountArray has sufficient space to store one more
255 : * entry. This has to be called before using NewPrivateRefCountEntry() to fill
256 : * a new entry - but it's perfectly fine to not use a reserved entry.
257 : */
258 : static void
259 131070498 : ReservePrivateRefCountEntry(void)
260 : {
261 : /* Already reserved (or freed), nothing to do */
262 131070498 : if (ReservedRefCountEntry != NULL)
263 122816604 : return;
264 :
265 : /*
266 : * First search for a free entry the array, that'll be sufficient in the
267 : * majority of cases.
268 : */
269 : {
270 : int i;
271 :
272 20983460 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
273 : {
274 : PrivateRefCountEntry *res;
275 :
276 20644728 : res = &PrivateRefCountArray[i];
277 :
278 20644728 : if (res->buffer == InvalidBuffer)
279 : {
280 7915162 : ReservedRefCountEntry = res;
281 7915162 : return;
282 : }
283 : }
284 : }
285 :
286 : /*
287 : * No luck. All array entries are full. Move one array entry into the hash
288 : * table.
289 : */
290 : {
291 : /*
292 : * Move entry from the current clock position in the array into the
293 : * hashtable. Use that slot.
294 : */
295 : PrivateRefCountEntry *hashent;
296 : bool found;
297 :
298 : /* select victim slot */
299 338732 : ReservedRefCountEntry =
300 338732 : &PrivateRefCountArray[PrivateRefCountClock++ % REFCOUNT_ARRAY_ENTRIES];
301 :
302 : /* Better be used, otherwise we shouldn't get here. */
303 : Assert(ReservedRefCountEntry->buffer != InvalidBuffer);
304 :
305 : /* enter victim array entry into hashtable */
306 338732 : hashent = hash_search(PrivateRefCountHash,
307 338732 : &(ReservedRefCountEntry->buffer),
308 : HASH_ENTER,
309 : &found);
310 : Assert(!found);
311 338732 : hashent->refcount = ReservedRefCountEntry->refcount;
312 :
313 : /* clear the now free array slot */
314 338732 : ReservedRefCountEntry->buffer = InvalidBuffer;
315 338732 : ReservedRefCountEntry->refcount = 0;
316 :
317 338732 : PrivateRefCountOverflowed++;
318 : }
319 : }
320 :
321 : /*
322 : * Fill a previously reserved refcount entry.
323 : */
324 : static PrivateRefCountEntry *
325 119489338 : NewPrivateRefCountEntry(Buffer buffer)
326 : {
327 : PrivateRefCountEntry *res;
328 :
329 : /* only allowed to be called when a reservation has been made */
330 : Assert(ReservedRefCountEntry != NULL);
331 :
332 : /* use up the reserved entry */
333 119489338 : res = ReservedRefCountEntry;
334 119489338 : ReservedRefCountEntry = NULL;
335 :
336 : /* and fill it */
337 119489338 : res->buffer = buffer;
338 119489338 : res->refcount = 0;
339 :
340 119489338 : return res;
341 : }
342 :
343 : /*
344 : * Return the PrivateRefCount entry for the passed buffer.
345 : *
346 : * Returns NULL if a buffer doesn't have a refcount entry. Otherwise, if
347 : * do_move is true, and the entry resides in the hashtable the entry is
348 : * optimized for frequent access by moving it to the array.
349 : */
350 : static PrivateRefCountEntry *
351 295548398 : GetPrivateRefCountEntry(Buffer buffer, bool do_move)
352 : {
353 : PrivateRefCountEntry *res;
354 : int i;
355 :
356 : Assert(BufferIsValid(buffer));
357 : Assert(!BufferIsLocal(buffer));
358 :
359 : /*
360 : * First search for references in the array, that'll be sufficient in the
361 : * majority of cases.
362 : */
363 1390149516 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
364 : {
365 1274739464 : res = &PrivateRefCountArray[i];
366 :
367 1274739464 : if (res->buffer == buffer)
368 180138346 : return res;
369 : }
370 :
371 : /*
372 : * By here we know that the buffer, if already pinned, isn't residing in
373 : * the array.
374 : *
375 : * Only look up the buffer in the hashtable if we've previously overflowed
376 : * into it.
377 : */
378 115410052 : if (PrivateRefCountOverflowed == 0)
379 114614184 : return NULL;
380 :
381 795868 : res = hash_search(PrivateRefCountHash, &buffer, HASH_FIND, NULL);
382 :
383 795868 : if (res == NULL)
384 452588 : return NULL;
385 343280 : else if (!do_move)
386 : {
387 : /* caller doesn't want us to move the hash entry into the array */
388 301934 : return res;
389 : }
390 : else
391 : {
392 : /* move buffer from hashtable into the free array slot */
393 : bool found;
394 : PrivateRefCountEntry *free;
395 :
396 : /* Ensure there's a free array slot */
397 41346 : ReservePrivateRefCountEntry();
398 :
399 : /* Use up the reserved slot */
400 : Assert(ReservedRefCountEntry != NULL);
401 41346 : free = ReservedRefCountEntry;
402 41346 : ReservedRefCountEntry = NULL;
403 : Assert(free->buffer == InvalidBuffer);
404 :
405 : /* and fill it */
406 41346 : free->buffer = buffer;
407 41346 : free->refcount = res->refcount;
408 :
409 : /* delete from hashtable */
410 41346 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
411 : Assert(found);
412 : Assert(PrivateRefCountOverflowed > 0);
413 41346 : PrivateRefCountOverflowed--;
414 :
415 41346 : return free;
416 : }
417 : }
418 :
419 : /*
420 : * Returns how many times the passed buffer is pinned by this backend.
421 : *
422 : * Only works for shared memory buffers!
423 : */
424 : static inline int32
425 6538952 : GetPrivateRefCount(Buffer buffer)
426 : {
427 : PrivateRefCountEntry *ref;
428 :
429 : Assert(BufferIsValid(buffer));
430 : Assert(!BufferIsLocal(buffer));
431 :
432 : /*
433 : * Not moving the entry - that's ok for the current users, but we might
434 : * want to change this one day.
435 : */
436 6538952 : ref = GetPrivateRefCountEntry(buffer, false);
437 :
438 6538952 : if (ref == NULL)
439 52 : return 0;
440 6538900 : return ref->refcount;
441 : }
442 :
443 : /*
444 : * Release resources used to track the reference count of a buffer which we no
445 : * longer have pinned and don't want to pin again immediately.
446 : */
447 : static void
448 119489338 : ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref)
449 : {
450 : Assert(ref->refcount == 0);
451 :
452 119489338 : if (ref >= &PrivateRefCountArray[0] &&
453 : ref < &PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES])
454 : {
455 119191952 : ref->buffer = InvalidBuffer;
456 :
457 : /*
458 : * Mark the just used entry as reserved - in many scenarios that
459 : * allows us to avoid ever having to search the array/hash for free
460 : * entries.
461 : */
462 119191952 : ReservedRefCountEntry = ref;
463 : }
464 : else
465 : {
466 : bool found;
467 297386 : Buffer buffer = ref->buffer;
468 :
469 297386 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
470 : Assert(found);
471 : Assert(PrivateRefCountOverflowed > 0);
472 297386 : PrivateRefCountOverflowed--;
473 : }
474 119489338 : }
475 :
476 : /*
477 : * BufferIsPinned
478 : * True iff the buffer is pinned (also checks for valid buffer number).
479 : *
480 : * NOTE: what we check here is that *this* backend holds a pin on
481 : * the buffer. We do not care whether some other backend does.
482 : */
483 : #define BufferIsPinned(bufnum) \
484 : ( \
485 : !BufferIsValid(bufnum) ? \
486 : false \
487 : : \
488 : BufferIsLocal(bufnum) ? \
489 : (LocalRefCount[-(bufnum) - 1] > 0) \
490 : : \
491 : (GetPrivateRefCount(bufnum) > 0) \
492 : )
493 :
494 :
495 : static Buffer ReadBuffer_common(Relation rel,
496 : SMgrRelation smgr, char smgr_persistence,
497 : ForkNumber forkNum, BlockNumber blockNum,
498 : ReadBufferMode mode, BufferAccessStrategy strategy);
499 : static BlockNumber ExtendBufferedRelCommon(BufferManagerRelation bmr,
500 : ForkNumber fork,
501 : BufferAccessStrategy strategy,
502 : uint32 flags,
503 : uint32 extend_by,
504 : BlockNumber extend_upto,
505 : Buffer *buffers,
506 : uint32 *extended_by);
507 : static BlockNumber ExtendBufferedRelShared(BufferManagerRelation bmr,
508 : ForkNumber fork,
509 : BufferAccessStrategy strategy,
510 : uint32 flags,
511 : uint32 extend_by,
512 : BlockNumber extend_upto,
513 : Buffer *buffers,
514 : uint32 *extended_by);
515 : static bool PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
516 : bool skip_if_not_valid);
517 : static void PinBuffer_Locked(BufferDesc *buf);
518 : static void UnpinBuffer(BufferDesc *buf);
519 : static void UnpinBufferNoOwner(BufferDesc *buf);
520 : static void BufferSync(int flags);
521 : static int SyncOneBuffer(int buf_id, bool skip_recently_used,
522 : WritebackContext *wb_context);
523 : static void WaitIO(BufferDesc *buf);
524 : static void AbortBufferIO(Buffer buffer);
525 : static void shared_buffer_write_error_callback(void *arg);
526 : static void local_buffer_write_error_callback(void *arg);
527 : static inline BufferDesc *BufferAlloc(SMgrRelation smgr,
528 : char relpersistence,
529 : ForkNumber forkNum,
530 : BlockNumber blockNum,
531 : BufferAccessStrategy strategy,
532 : bool *foundPtr, IOContext io_context);
533 : static bool AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress);
534 : static void CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete);
535 : static Buffer GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context);
536 : static void FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
537 : IOObject io_object, IOContext io_context);
538 : static void FlushBuffer(BufferDesc *buf, SMgrRelation reln,
539 : IOObject io_object, IOContext io_context);
540 : static void FindAndDropRelationBuffers(RelFileLocator rlocator,
541 : ForkNumber forkNum,
542 : BlockNumber nForkBlock,
543 : BlockNumber firstDelBlock);
544 : static void RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
545 : RelFileLocator dstlocator,
546 : ForkNumber forkNum, bool permanent);
547 : static void AtProcExit_Buffers(int code, Datum arg);
548 : static void CheckForBufferLeaks(void);
549 : #ifdef USE_ASSERT_CHECKING
550 : static void AssertNotCatalogBufferLock(LWLock *lock, LWLockMode mode,
551 : void *unused_context);
552 : #endif
553 : static int rlocator_comparator(const void *p1, const void *p2);
554 : static inline int buffertag_comparator(const BufferTag *ba, const BufferTag *bb);
555 : static inline int ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b);
556 : static int ts_ckpt_progress_comparator(Datum a, Datum b, void *arg);
557 :
558 :
559 : /*
560 : * Implementation of PrefetchBuffer() for shared buffers.
561 : */
562 : PrefetchBufferResult
563 61802 : PrefetchSharedBuffer(SMgrRelation smgr_reln,
564 : ForkNumber forkNum,
565 : BlockNumber blockNum)
566 : {
567 61802 : PrefetchBufferResult result = {InvalidBuffer, false};
568 : BufferTag newTag; /* identity of requested block */
569 : uint32 newHash; /* hash value for newTag */
570 : LWLock *newPartitionLock; /* buffer partition lock for it */
571 : int buf_id;
572 :
573 : Assert(BlockNumberIsValid(blockNum));
574 :
575 : /* create a tag so we can lookup the buffer */
576 61802 : InitBufferTag(&newTag, &smgr_reln->smgr_rlocator.locator,
577 : forkNum, blockNum);
578 :
579 : /* determine its hash code and partition lock ID */
580 61802 : newHash = BufTableHashCode(&newTag);
581 61802 : newPartitionLock = BufMappingPartitionLock(newHash);
582 :
583 : /* see if the block is in the buffer pool already */
584 61802 : LWLockAcquire(newPartitionLock, LW_SHARED);
585 61802 : buf_id = BufTableLookup(&newTag, newHash);
586 61802 : LWLockRelease(newPartitionLock);
587 :
588 : /* If not in buffers, initiate prefetch */
589 61802 : if (buf_id < 0)
590 : {
591 : #ifdef USE_PREFETCH
592 : /*
593 : * Try to initiate an asynchronous read. This returns false in
594 : * recovery if the relation file doesn't exist.
595 : */
596 34528 : if ((io_direct_flags & IO_DIRECT_DATA) == 0 &&
597 17040 : smgrprefetch(smgr_reln, forkNum, blockNum, 1))
598 : {
599 17040 : result.initiated_io = true;
600 : }
601 : #endif /* USE_PREFETCH */
602 : }
603 : else
604 : {
605 : /*
606 : * Report the buffer it was in at that time. The caller may be able
607 : * to avoid a buffer table lookup, but it's not pinned and it must be
608 : * rechecked!
609 : */
610 44314 : result.recent_buffer = buf_id + 1;
611 : }
612 :
613 : /*
614 : * If the block *is* in buffers, we do nothing. This is not really ideal:
615 : * the block might be just about to be evicted, which would be stupid
616 : * since we know we are going to need it soon. But the only easy answer
617 : * is to bump the usage_count, which does not seem like a great solution:
618 : * when the caller does ultimately touch the block, usage_count would get
619 : * bumped again, resulting in too much favoritism for blocks that are
620 : * involved in a prefetch sequence. A real fix would involve some
621 : * additional per-buffer state, and it's not clear that there's enough of
622 : * a problem to justify that.
623 : */
624 :
625 61802 : return result;
626 : }
627 :
628 : /*
629 : * PrefetchBuffer -- initiate asynchronous read of a block of a relation
630 : *
631 : * This is named by analogy to ReadBuffer but doesn't actually allocate a
632 : * buffer. Instead it tries to ensure that a future ReadBuffer for the given
633 : * block will not be delayed by the I/O. Prefetching is optional.
634 : *
635 : * There are three possible outcomes:
636 : *
637 : * 1. If the block is already cached, the result includes a valid buffer that
638 : * could be used by the caller to avoid the need for a later buffer lookup, but
639 : * it's not pinned, so the caller must recheck it.
640 : *
641 : * 2. If the kernel has been asked to initiate I/O, the initiated_io member is
642 : * true. Currently there is no way to know if the data was already cached by
643 : * the kernel and therefore didn't really initiate I/O, and no way to know when
644 : * the I/O completes other than using synchronous ReadBuffer().
645 : *
646 : * 3. Otherwise, the buffer wasn't already cached by PostgreSQL, and
647 : * USE_PREFETCH is not defined (this build doesn't support prefetching due to
648 : * lack of a kernel facility), direct I/O is enabled, or the underlying
649 : * relation file wasn't found and we are in recovery. (If the relation file
650 : * wasn't found and we are not in recovery, an error is raised).
651 : */
652 : PrefetchBufferResult
653 40634 : PrefetchBuffer(Relation reln, ForkNumber forkNum, BlockNumber blockNum)
654 : {
655 : Assert(RelationIsValid(reln));
656 : Assert(BlockNumberIsValid(blockNum));
657 :
658 40634 : if (RelationUsesLocalBuffers(reln))
659 : {
660 : /* see comments in ReadBufferExtended */
661 1566 : if (RELATION_IS_OTHER_TEMP(reln))
662 0 : ereport(ERROR,
663 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
664 : errmsg("cannot access temporary tables of other sessions")));
665 :
666 : /* pass it off to localbuf.c */
667 1566 : return PrefetchLocalBuffer(RelationGetSmgr(reln), forkNum, blockNum);
668 : }
669 : else
670 : {
671 : /* pass it to the shared buffer version */
672 39068 : return PrefetchSharedBuffer(RelationGetSmgr(reln), forkNum, blockNum);
673 : }
674 : }
675 :
676 : /*
677 : * ReadRecentBuffer -- try to pin a block in a recently observed buffer
678 : *
679 : * Compared to ReadBuffer(), this avoids a buffer mapping lookup when it's
680 : * successful. Return true if the buffer is valid and still has the expected
681 : * tag. In that case, the buffer is pinned and the usage count is bumped.
682 : */
683 : bool
684 8978 : ReadRecentBuffer(RelFileLocator rlocator, ForkNumber forkNum, BlockNumber blockNum,
685 : Buffer recent_buffer)
686 : {
687 : BufferDesc *bufHdr;
688 : BufferTag tag;
689 : uint32 buf_state;
690 :
691 : Assert(BufferIsValid(recent_buffer));
692 :
693 8978 : ResourceOwnerEnlarge(CurrentResourceOwner);
694 8978 : ReservePrivateRefCountEntry();
695 8978 : InitBufferTag(&tag, &rlocator, forkNum, blockNum);
696 :
697 8978 : if (BufferIsLocal(recent_buffer))
698 : {
699 64 : int b = -recent_buffer - 1;
700 :
701 64 : bufHdr = GetLocalBufferDescriptor(b);
702 64 : buf_state = pg_atomic_read_u32(&bufHdr->state);
703 :
704 : /* Is it still valid and holding the right tag? */
705 64 : if ((buf_state & BM_VALID) && BufferTagsEqual(&tag, &bufHdr->tag))
706 : {
707 64 : PinLocalBuffer(bufHdr, true);
708 :
709 64 : pgBufferUsage.local_blks_hit++;
710 :
711 64 : return true;
712 : }
713 : }
714 : else
715 : {
716 8914 : bufHdr = GetBufferDescriptor(recent_buffer - 1);
717 :
718 : /*
719 : * Is it still valid and holding the right tag? We do an unlocked tag
720 : * comparison first, to make it unlikely that we'll increment the
721 : * usage counter of the wrong buffer, if someone calls us with a very
722 : * out of date recent_buffer. Then we'll check it again if we get the
723 : * pin.
724 : */
725 17752 : if (BufferTagsEqual(&tag, &bufHdr->tag) &&
726 8838 : PinBuffer(bufHdr, NULL, true))
727 : {
728 8826 : if (BufferTagsEqual(&tag, &bufHdr->tag))
729 : {
730 8826 : pgBufferUsage.shared_blks_hit++;
731 8826 : return true;
732 : }
733 0 : UnpinBuffer(bufHdr);
734 : }
735 : }
736 :
737 88 : return false;
738 : }
739 :
740 : /*
741 : * ReadBuffer -- a shorthand for ReadBufferExtended, for reading from main
742 : * fork with RBM_NORMAL mode and default strategy.
743 : */
744 : Buffer
745 87332462 : ReadBuffer(Relation reln, BlockNumber blockNum)
746 : {
747 87332462 : return ReadBufferExtended(reln, MAIN_FORKNUM, blockNum, RBM_NORMAL, NULL);
748 : }
749 :
750 : /*
751 : * ReadBufferExtended -- returns a buffer containing the requested
752 : * block of the requested relation. If the blknum
753 : * requested is P_NEW, extend the relation file and
754 : * allocate a new block. (Caller is responsible for
755 : * ensuring that only one backend tries to extend a
756 : * relation at the same time!)
757 : *
758 : * Returns: the buffer number for the buffer containing
759 : * the block read. The returned buffer has been pinned.
760 : * Does not return on error --- elog's instead.
761 : *
762 : * Assume when this function is called, that reln has been opened already.
763 : *
764 : * In RBM_NORMAL mode, the page is read from disk, and the page header is
765 : * validated. An error is thrown if the page header is not valid. (But
766 : * note that an all-zero page is considered "valid"; see
767 : * PageIsVerified().)
768 : *
769 : * RBM_ZERO_ON_ERROR is like the normal mode, but if the page header is not
770 : * valid, the page is zeroed instead of throwing an error. This is intended
771 : * for non-critical data, where the caller is prepared to repair errors.
772 : *
773 : * In RBM_ZERO_AND_LOCK mode, if the page isn't in buffer cache already, it's
774 : * filled with zeros instead of reading it from disk. Useful when the caller
775 : * is going to fill the page from scratch, since this saves I/O and avoids
776 : * unnecessary failure if the page-on-disk has corrupt page headers.
777 : * The page is returned locked to ensure that the caller has a chance to
778 : * initialize the page before it's made visible to others.
779 : * Caution: do not use this mode to read a page that is beyond the relation's
780 : * current physical EOF; that is likely to cause problems in md.c when
781 : * the page is modified and written out. P_NEW is OK, though.
782 : *
783 : * RBM_ZERO_AND_CLEANUP_LOCK is the same as RBM_ZERO_AND_LOCK, but acquires
784 : * a cleanup-strength lock on the page.
785 : *
786 : * RBM_NORMAL_NO_LOG mode is treated the same as RBM_NORMAL here.
787 : *
788 : * If strategy is not NULL, a nondefault buffer access strategy is used.
789 : * See buffer/README for details.
790 : */
791 : inline Buffer
792 105538232 : ReadBufferExtended(Relation reln, ForkNumber forkNum, BlockNumber blockNum,
793 : ReadBufferMode mode, BufferAccessStrategy strategy)
794 : {
795 : Buffer buf;
796 :
797 : /*
798 : * Reject attempts to read non-local temporary relations; we would be
799 : * likely to get wrong data since we have no visibility into the owning
800 : * session's local buffers.
801 : */
802 105538232 : if (RELATION_IS_OTHER_TEMP(reln))
803 0 : ereport(ERROR,
804 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
805 : errmsg("cannot access temporary tables of other sessions")));
806 :
807 : /*
808 : * Read the buffer, and update pgstat counters to reflect a cache hit or
809 : * miss.
810 : */
811 105538232 : buf = ReadBuffer_common(reln, RelationGetSmgr(reln), 0,
812 : forkNum, blockNum, mode, strategy);
813 :
814 105538186 : return buf;
815 : }
816 :
817 :
818 : /*
819 : * ReadBufferWithoutRelcache -- like ReadBufferExtended, but doesn't require
820 : * a relcache entry for the relation.
821 : *
822 : * Pass permanent = true for a RELPERSISTENCE_PERMANENT relation, and
823 : * permanent = false for a RELPERSISTENCE_UNLOGGED relation. This function
824 : * cannot be used for temporary relations (and making that work might be
825 : * difficult, unless we only want to read temporary relations for our own
826 : * ProcNumber).
827 : */
828 : Buffer
829 11358346 : ReadBufferWithoutRelcache(RelFileLocator rlocator, ForkNumber forkNum,
830 : BlockNumber blockNum, ReadBufferMode mode,
831 : BufferAccessStrategy strategy, bool permanent)
832 : {
833 11358346 : SMgrRelation smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
834 :
835 11358346 : return ReadBuffer_common(NULL, smgr,
836 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
837 : forkNum, blockNum,
838 : mode, strategy);
839 : }
840 :
841 : /*
842 : * Convenience wrapper around ExtendBufferedRelBy() extending by one block.
843 : */
844 : Buffer
845 90526 : ExtendBufferedRel(BufferManagerRelation bmr,
846 : ForkNumber forkNum,
847 : BufferAccessStrategy strategy,
848 : uint32 flags)
849 : {
850 : Buffer buf;
851 90526 : uint32 extend_by = 1;
852 :
853 90526 : ExtendBufferedRelBy(bmr, forkNum, strategy, flags, extend_by,
854 : &buf, &extend_by);
855 :
856 90526 : return buf;
857 : }
858 :
859 : /*
860 : * Extend relation by multiple blocks.
861 : *
862 : * Tries to extend the relation by extend_by blocks. Depending on the
863 : * availability of resources the relation may end up being extended by a
864 : * smaller number of pages (unless an error is thrown, always by at least one
865 : * page). *extended_by is updated to the number of pages the relation has been
866 : * extended to.
867 : *
868 : * buffers needs to be an array that is at least extend_by long. Upon
869 : * completion, the first extend_by array elements will point to a pinned
870 : * buffer.
871 : *
872 : * If EB_LOCK_FIRST is part of flags, the first returned buffer is
873 : * locked. This is useful for callers that want a buffer that is guaranteed to
874 : * be empty.
875 : */
876 : BlockNumber
877 318390 : ExtendBufferedRelBy(BufferManagerRelation bmr,
878 : ForkNumber fork,
879 : BufferAccessStrategy strategy,
880 : uint32 flags,
881 : uint32 extend_by,
882 : Buffer *buffers,
883 : uint32 *extended_by)
884 : {
885 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
886 : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
887 : Assert(extend_by > 0);
888 :
889 318390 : if (bmr.relpersistence == '\0')
890 318390 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
891 :
892 318390 : return ExtendBufferedRelCommon(bmr, fork, strategy, flags,
893 : extend_by, InvalidBlockNumber,
894 : buffers, extended_by);
895 : }
896 :
897 : /*
898 : * Extend the relation so it is at least extend_to blocks large, return buffer
899 : * (extend_to - 1).
900 : *
901 : * This is useful for callers that want to write a specific page, regardless
902 : * of the current size of the relation (e.g. useful for visibilitymap and for
903 : * crash recovery).
904 : */
905 : Buffer
906 102798 : ExtendBufferedRelTo(BufferManagerRelation bmr,
907 : ForkNumber fork,
908 : BufferAccessStrategy strategy,
909 : uint32 flags,
910 : BlockNumber extend_to,
911 : ReadBufferMode mode)
912 : {
913 : BlockNumber current_size;
914 102798 : uint32 extended_by = 0;
915 102798 : Buffer buffer = InvalidBuffer;
916 : Buffer buffers[64];
917 :
918 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
919 : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
920 : Assert(extend_to != InvalidBlockNumber && extend_to > 0);
921 :
922 102798 : if (bmr.relpersistence == '\0')
923 14108 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
924 :
925 : /*
926 : * If desired, create the file if it doesn't exist. If
927 : * smgr_cached_nblocks[fork] is positive then it must exist, no need for
928 : * an smgrexists call.
929 : */
930 102798 : if ((flags & EB_CREATE_FORK_IF_NEEDED) &&
931 14108 : (BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == 0 ||
932 40 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == InvalidBlockNumber) &&
933 14068 : !smgrexists(BMR_GET_SMGR(bmr), fork))
934 : {
935 14038 : LockRelationForExtension(bmr.rel, ExclusiveLock);
936 :
937 : /* recheck, fork might have been created concurrently */
938 14038 : if (!smgrexists(BMR_GET_SMGR(bmr), fork))
939 14030 : smgrcreate(BMR_GET_SMGR(bmr), fork, flags & EB_PERFORMING_RECOVERY);
940 :
941 14038 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
942 : }
943 :
944 : /*
945 : * If requested, invalidate size cache, so that smgrnblocks asks the
946 : * kernel.
947 : */
948 102798 : if (flags & EB_CLEAR_SIZE_CACHE)
949 14108 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
950 :
951 : /*
952 : * Estimate how many pages we'll need to extend by. This avoids acquiring
953 : * unnecessarily many victim buffers.
954 : */
955 102798 : current_size = smgrnblocks(BMR_GET_SMGR(bmr), fork);
956 :
957 : /*
958 : * Since no-one else can be looking at the page contents yet, there is no
959 : * difference between an exclusive lock and a cleanup-strength lock. Note
960 : * that we pass the original mode to ReadBuffer_common() below, when
961 : * falling back to reading the buffer to a concurrent relation extension.
962 : */
963 102798 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
964 87954 : flags |= EB_LOCK_TARGET;
965 :
966 209604 : while (current_size < extend_to)
967 : {
968 106806 : uint32 num_pages = lengthof(buffers);
969 : BlockNumber first_block;
970 :
971 106806 : if ((uint64) current_size + num_pages > extend_to)
972 106804 : num_pages = extend_to - current_size;
973 :
974 106806 : first_block = ExtendBufferedRelCommon(bmr, fork, strategy, flags,
975 : num_pages, extend_to,
976 : buffers, &extended_by);
977 :
978 106806 : current_size = first_block + extended_by;
979 : Assert(num_pages != 0 || current_size >= extend_to);
980 :
981 228018 : for (uint32 i = 0; i < extended_by; i++)
982 : {
983 121212 : if (first_block + i != extend_to - 1)
984 18430 : ReleaseBuffer(buffers[i]);
985 : else
986 102782 : buffer = buffers[i];
987 : }
988 : }
989 :
990 : /*
991 : * It's possible that another backend concurrently extended the relation.
992 : * In that case read the buffer.
993 : *
994 : * XXX: Should we control this via a flag?
995 : */
996 102798 : if (buffer == InvalidBuffer)
997 : {
998 : Assert(extended_by == 0);
999 16 : buffer = ReadBuffer_common(bmr.rel, BMR_GET_SMGR(bmr), bmr.relpersistence,
1000 : fork, extend_to - 1, mode, strategy);
1001 : }
1002 :
1003 102798 : return buffer;
1004 : }
1005 :
1006 : /*
1007 : * Lock and optionally zero a buffer, as part of the implementation of
1008 : * RBM_ZERO_AND_LOCK or RBM_ZERO_AND_CLEANUP_LOCK. The buffer must be already
1009 : * pinned. If the buffer is not already valid, it is zeroed and made valid.
1010 : */
1011 : static void
1012 638214 : ZeroAndLockBuffer(Buffer buffer, ReadBufferMode mode, bool already_valid)
1013 : {
1014 : BufferDesc *bufHdr;
1015 : bool need_to_zero;
1016 638214 : bool isLocalBuf = BufferIsLocal(buffer);
1017 :
1018 : Assert(mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK);
1019 :
1020 638214 : if (already_valid)
1021 : {
1022 : /*
1023 : * If the caller already knew the buffer was valid, we can skip some
1024 : * header interaction. The caller just wants to lock the buffer.
1025 : */
1026 74206 : need_to_zero = false;
1027 : }
1028 564008 : else if (isLocalBuf)
1029 : {
1030 : /* Simple case for non-shared buffers. */
1031 48 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
1032 48 : need_to_zero = StartLocalBufferIO(bufHdr, true, false);
1033 : }
1034 : else
1035 : {
1036 : /*
1037 : * Take BM_IO_IN_PROGRESS, or discover that BM_VALID has been set
1038 : * concurrently. Even though we aren't doing I/O, that ensures that
1039 : * we don't zero a page that someone else has pinned. An exclusive
1040 : * content lock wouldn't be enough, because readers are allowed to
1041 : * drop the content lock after determining that a tuple is visible
1042 : * (see buffer access rules in README).
1043 : */
1044 563960 : bufHdr = GetBufferDescriptor(buffer - 1);
1045 563960 : need_to_zero = StartBufferIO(bufHdr, true, false);
1046 : }
1047 :
1048 638214 : if (need_to_zero)
1049 : {
1050 564008 : memset(BufferGetPage(buffer), 0, BLCKSZ);
1051 :
1052 : /*
1053 : * Grab the buffer content lock before marking the page as valid, to
1054 : * make sure that no other backend sees the zeroed page before the
1055 : * caller has had a chance to initialize it.
1056 : *
1057 : * Since no-one else can be looking at the page contents yet, there is
1058 : * no difference between an exclusive lock and a cleanup-strength
1059 : * lock. (Note that we cannot use LockBuffer() or
1060 : * LockBufferForCleanup() here, because they assert that the buffer is
1061 : * already valid.)
1062 : */
1063 564008 : if (!isLocalBuf)
1064 563960 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1065 :
1066 : /* Set BM_VALID, terminate IO, and wake up any waiters */
1067 564008 : if (isLocalBuf)
1068 48 : TerminateLocalBufferIO(bufHdr, false, BM_VALID, false);
1069 : else
1070 563960 : TerminateBufferIO(bufHdr, false, BM_VALID, true, false);
1071 : }
1072 74206 : else if (!isLocalBuf)
1073 : {
1074 : /*
1075 : * The buffer is valid, so we can't zero it. The caller still expects
1076 : * the page to be locked on return.
1077 : */
1078 74166 : if (mode == RBM_ZERO_AND_LOCK)
1079 74066 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1080 : else
1081 100 : LockBufferForCleanup(buffer);
1082 : }
1083 638214 : }
1084 :
1085 : /*
1086 : * Pin a buffer for a given block. *foundPtr is set to true if the block was
1087 : * already present, or false if more work is required to either read it in or
1088 : * zero it.
1089 : */
1090 : static pg_attribute_always_inline Buffer
1091 125299926 : PinBufferForBlock(Relation rel,
1092 : SMgrRelation smgr,
1093 : char persistence,
1094 : ForkNumber forkNum,
1095 : BlockNumber blockNum,
1096 : BufferAccessStrategy strategy,
1097 : bool *foundPtr)
1098 : {
1099 : BufferDesc *bufHdr;
1100 : IOContext io_context;
1101 : IOObject io_object;
1102 :
1103 : Assert(blockNum != P_NEW);
1104 :
1105 : /* Persistence should be set before */
1106 : Assert((persistence == RELPERSISTENCE_TEMP ||
1107 : persistence == RELPERSISTENCE_PERMANENT ||
1108 : persistence == RELPERSISTENCE_UNLOGGED));
1109 :
1110 125299926 : if (persistence == RELPERSISTENCE_TEMP)
1111 : {
1112 2554438 : io_context = IOCONTEXT_NORMAL;
1113 2554438 : io_object = IOOBJECT_TEMP_RELATION;
1114 : }
1115 : else
1116 : {
1117 122745488 : io_context = IOContextForStrategy(strategy);
1118 122745488 : io_object = IOOBJECT_RELATION;
1119 : }
1120 :
1121 : TRACE_POSTGRESQL_BUFFER_READ_START(forkNum, blockNum,
1122 : smgr->smgr_rlocator.locator.spcOid,
1123 : smgr->smgr_rlocator.locator.dbOid,
1124 : smgr->smgr_rlocator.locator.relNumber,
1125 : smgr->smgr_rlocator.backend);
1126 :
1127 125299926 : if (persistence == RELPERSISTENCE_TEMP)
1128 : {
1129 2554438 : bufHdr = LocalBufferAlloc(smgr, forkNum, blockNum, foundPtr);
1130 2554426 : if (*foundPtr)
1131 2537644 : pgBufferUsage.local_blks_hit++;
1132 : }
1133 : else
1134 : {
1135 122745488 : bufHdr = BufferAlloc(smgr, persistence, forkNum, blockNum,
1136 : strategy, foundPtr, io_context);
1137 122745488 : if (*foundPtr)
1138 119375834 : pgBufferUsage.shared_blks_hit++;
1139 : }
1140 125299914 : if (rel)
1141 : {
1142 : /*
1143 : * While pgBufferUsage's "read" counter isn't bumped unless we reach
1144 : * WaitReadBuffers() (so, not for hits, and not for buffers that are
1145 : * zeroed instead), the per-relation stats always count them.
1146 : */
1147 113471322 : pgstat_count_buffer_read(rel);
1148 113471322 : if (*foundPtr)
1149 111012796 : pgstat_count_buffer_hit(rel);
1150 : }
1151 125299914 : if (*foundPtr)
1152 : {
1153 121913478 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1154 121913478 : if (VacuumCostActive)
1155 7024510 : VacuumCostBalance += VacuumCostPageHit;
1156 :
1157 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forkNum, blockNum,
1158 : smgr->smgr_rlocator.locator.spcOid,
1159 : smgr->smgr_rlocator.locator.dbOid,
1160 : smgr->smgr_rlocator.locator.relNumber,
1161 : smgr->smgr_rlocator.backend,
1162 : true);
1163 : }
1164 :
1165 125299914 : return BufferDescriptorGetBuffer(bufHdr);
1166 : }
1167 :
1168 : /*
1169 : * ReadBuffer_common -- common logic for all ReadBuffer variants
1170 : *
1171 : * smgr is required, rel is optional unless using P_NEW.
1172 : */
1173 : static pg_attribute_always_inline Buffer
1174 116896594 : ReadBuffer_common(Relation rel, SMgrRelation smgr, char smgr_persistence,
1175 : ForkNumber forkNum,
1176 : BlockNumber blockNum, ReadBufferMode mode,
1177 : BufferAccessStrategy strategy)
1178 : {
1179 : ReadBuffersOperation operation;
1180 : Buffer buffer;
1181 : int flags;
1182 : char persistence;
1183 :
1184 : /*
1185 : * Backward compatibility path, most code should use ExtendBufferedRel()
1186 : * instead, as acquiring the extension lock inside ExtendBufferedRel()
1187 : * scales a lot better.
1188 : */
1189 116896594 : if (unlikely(blockNum == P_NEW))
1190 : {
1191 520 : uint32 flags = EB_SKIP_EXTENSION_LOCK;
1192 :
1193 : /*
1194 : * Since no-one else can be looking at the page contents yet, there is
1195 : * no difference between an exclusive lock and a cleanup-strength
1196 : * lock.
1197 : */
1198 520 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1199 0 : flags |= EB_LOCK_FIRST;
1200 :
1201 520 : return ExtendBufferedRel(BMR_REL(rel), forkNum, strategy, flags);
1202 : }
1203 :
1204 116896074 : if (rel)
1205 105537728 : persistence = rel->rd_rel->relpersistence;
1206 : else
1207 11358346 : persistence = smgr_persistence;
1208 :
1209 116896074 : if (unlikely(mode == RBM_ZERO_AND_CLEANUP_LOCK ||
1210 : mode == RBM_ZERO_AND_LOCK))
1211 : {
1212 : bool found;
1213 :
1214 638214 : buffer = PinBufferForBlock(rel, smgr, persistence,
1215 : forkNum, blockNum, strategy, &found);
1216 638214 : ZeroAndLockBuffer(buffer, mode, found);
1217 638214 : return buffer;
1218 : }
1219 :
1220 : /*
1221 : * Signal that we are going to immediately wait. If we're immediately
1222 : * waiting, there is no benefit in actually executing the IO
1223 : * asynchronously, it would just add dispatch overhead.
1224 : */
1225 116257860 : flags = READ_BUFFERS_SYNCHRONOUSLY;
1226 116257860 : if (mode == RBM_ZERO_ON_ERROR)
1227 3290016 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1228 116257860 : operation.smgr = smgr;
1229 116257860 : operation.rel = rel;
1230 116257860 : operation.persistence = persistence;
1231 116257860 : operation.forknum = forkNum;
1232 116257860 : operation.strategy = strategy;
1233 116257860 : if (StartReadBuffer(&operation,
1234 : &buffer,
1235 : blockNum,
1236 : flags))
1237 1390560 : WaitReadBuffers(&operation);
1238 :
1239 116257814 : return buffer;
1240 : }
1241 :
1242 : static pg_attribute_always_inline bool
1243 124330698 : StartReadBuffersImpl(ReadBuffersOperation *operation,
1244 : Buffer *buffers,
1245 : BlockNumber blockNum,
1246 : int *nblocks,
1247 : int flags,
1248 : bool allow_forwarding)
1249 : {
1250 124330698 : int actual_nblocks = *nblocks;
1251 124330698 : int maxcombine = 0;
1252 : bool did_start_io;
1253 :
1254 : Assert(*nblocks == 1 || allow_forwarding);
1255 : Assert(*nblocks > 0);
1256 : Assert(*nblocks <= MAX_IO_COMBINE_LIMIT);
1257 :
1258 127153126 : for (int i = 0; i < actual_nblocks; ++i)
1259 : {
1260 : bool found;
1261 :
1262 124664702 : if (allow_forwarding && buffers[i] != InvalidBuffer)
1263 2990 : {
1264 : BufferDesc *bufHdr;
1265 :
1266 : /*
1267 : * This is a buffer that was pinned by an earlier call to
1268 : * StartReadBuffers(), but couldn't be handled in one operation at
1269 : * that time. The operation was split, and the caller has passed
1270 : * an already pinned buffer back to us to handle the rest of the
1271 : * operation. It must continue at the expected block number.
1272 : */
1273 : Assert(BufferGetBlockNumber(buffers[i]) == blockNum + i);
1274 :
1275 : /*
1276 : * It might be an already valid buffer (a hit) that followed the
1277 : * final contiguous block of an earlier I/O (a miss) marking the
1278 : * end of it, or a buffer that some other backend has since made
1279 : * valid by performing the I/O for us, in which case we can handle
1280 : * it as a hit now. It is safe to check for a BM_VALID flag with
1281 : * a relaxed load, because we got a fresh view of it while pinning
1282 : * it in the previous call.
1283 : *
1284 : * On the other hand if we don't see BM_VALID yet, it must be an
1285 : * I/O that was split by the previous call and we need to try to
1286 : * start a new I/O from this block. We're also racing against any
1287 : * other backend that might start the I/O or even manage to mark
1288 : * it BM_VALID after this check, but StartBufferIO() will handle
1289 : * those cases.
1290 : */
1291 2990 : if (BufferIsLocal(buffers[i]))
1292 4 : bufHdr = GetLocalBufferDescriptor(-buffers[i] - 1);
1293 : else
1294 2986 : bufHdr = GetBufferDescriptor(buffers[i] - 1);
1295 : Assert(pg_atomic_read_u32(&bufHdr->state) & BM_TAG_VALID);
1296 2990 : found = pg_atomic_read_u32(&bufHdr->state) & BM_VALID;
1297 : }
1298 : else
1299 : {
1300 124661700 : buffers[i] = PinBufferForBlock(operation->rel,
1301 : operation->smgr,
1302 124661712 : operation->persistence,
1303 : operation->forknum,
1304 : blockNum + i,
1305 : operation->strategy,
1306 : &found);
1307 : }
1308 :
1309 124664690 : if (found)
1310 : {
1311 : /*
1312 : * We have a hit. If it's the first block in the requested range,
1313 : * we can return it immediately and report that WaitReadBuffers()
1314 : * does not need to be called. If the initial value of *nblocks
1315 : * was larger, the caller will have to call again for the rest.
1316 : */
1317 121842262 : if (i == 0)
1318 : {
1319 121839268 : *nblocks = 1;
1320 :
1321 : #ifdef USE_ASSERT_CHECKING
1322 :
1323 : /*
1324 : * Initialize enough of ReadBuffersOperation to make
1325 : * CheckReadBuffersOperation() work. Outside of assertions
1326 : * that's not necessary when no IO is issued.
1327 : */
1328 : operation->buffers = buffers;
1329 : operation->blocknum = blockNum;
1330 : operation->nblocks = 1;
1331 : operation->nblocks_done = 1;
1332 : CheckReadBuffersOperation(operation, true);
1333 : #endif
1334 121839268 : return false;
1335 : }
1336 :
1337 : /*
1338 : * Otherwise we already have an I/O to perform, but this block
1339 : * can't be included as it is already valid. Split the I/O here.
1340 : * There may or may not be more blocks requiring I/O after this
1341 : * one, we haven't checked, but they can't be contiguous with this
1342 : * one in the way. We'll leave this buffer pinned, forwarding it
1343 : * to the next call, avoiding the need to unpin it here and re-pin
1344 : * it in the next call.
1345 : */
1346 2994 : actual_nblocks = i;
1347 2994 : break;
1348 : }
1349 : else
1350 : {
1351 : /*
1352 : * Check how many blocks we can cover with the same IO. The smgr
1353 : * implementation might e.g. be limited due to a segment boundary.
1354 : */
1355 2822428 : if (i == 0 && actual_nblocks > 1)
1356 : {
1357 67772 : maxcombine = smgrmaxcombine(operation->smgr,
1358 : operation->forknum,
1359 : blockNum);
1360 67772 : if (unlikely(maxcombine < actual_nblocks))
1361 : {
1362 0 : elog(DEBUG2, "limiting nblocks at %u from %u to %u",
1363 : blockNum, actual_nblocks, maxcombine);
1364 0 : actual_nblocks = maxcombine;
1365 : }
1366 : }
1367 : }
1368 : }
1369 2491418 : *nblocks = actual_nblocks;
1370 :
1371 : /* Populate information needed for I/O. */
1372 2491418 : operation->buffers = buffers;
1373 2491418 : operation->blocknum = blockNum;
1374 2491418 : operation->flags = flags;
1375 2491418 : operation->nblocks = actual_nblocks;
1376 2491418 : operation->nblocks_done = 0;
1377 2491418 : pgaio_wref_clear(&operation->io_wref);
1378 :
1379 : /*
1380 : * When using AIO, start the IO in the background. If not, issue prefetch
1381 : * requests if desired by the caller.
1382 : *
1383 : * The reason we have a dedicated path for IOMETHOD_SYNC here is to
1384 : * de-risk the introduction of AIO somewhat. It's a large architectural
1385 : * change, with lots of chances for unanticipated performance effects.
1386 : *
1387 : * Use of IOMETHOD_SYNC already leads to not actually performing IO
1388 : * asynchronously, but without the check here we'd execute IO earlier than
1389 : * we used to. Eventually this IOMETHOD_SYNC specific path should go away.
1390 : */
1391 2491418 : if (io_method != IOMETHOD_SYNC)
1392 : {
1393 : /*
1394 : * Try to start IO asynchronously. It's possible that no IO needs to
1395 : * be started, if another backend already performed the IO.
1396 : *
1397 : * Note that if an IO is started, it might not cover the entire
1398 : * requested range, e.g. because an intermediary block has been read
1399 : * in by another backend. In that case any "trailing" buffers we
1400 : * already pinned above will be "forwarded" by read_stream.c to the
1401 : * next call to StartReadBuffers().
1402 : *
1403 : * This is signalled to the caller by decrementing *nblocks *and*
1404 : * reducing operation->nblocks. The latter is done here, but not below
1405 : * WaitReadBuffers(), as in WaitReadBuffers() we can't "shorten" the
1406 : * overall read size anymore, we need to retry until done in its
1407 : * entirety or until failed.
1408 : */
1409 2489256 : did_start_io = AsyncReadBuffers(operation, nblocks);
1410 :
1411 2489226 : operation->nblocks = *nblocks;
1412 : }
1413 : else
1414 : {
1415 2162 : operation->flags |= READ_BUFFERS_SYNCHRONOUSLY;
1416 :
1417 2162 : if (flags & READ_BUFFERS_ISSUE_ADVICE)
1418 : {
1419 : /*
1420 : * In theory we should only do this if PinBufferForBlock() had to
1421 : * allocate new buffers above. That way, if two calls to
1422 : * StartReadBuffers() were made for the same blocks before
1423 : * WaitReadBuffers(), only the first would issue the advice.
1424 : * That'd be a better simulation of true asynchronous I/O, which
1425 : * would only start the I/O once, but isn't done here for
1426 : * simplicity.
1427 : */
1428 4 : smgrprefetch(operation->smgr,
1429 : operation->forknum,
1430 : blockNum,
1431 : actual_nblocks);
1432 : }
1433 :
1434 : /*
1435 : * Indicate that WaitReadBuffers() should be called. WaitReadBuffers()
1436 : * will initiate the necessary IO.
1437 : */
1438 2162 : did_start_io = true;
1439 : }
1440 :
1441 2491388 : CheckReadBuffersOperation(operation, !did_start_io);
1442 :
1443 2491388 : return did_start_io;
1444 : }
1445 :
1446 : /*
1447 : * Begin reading a range of blocks beginning at blockNum and extending for
1448 : * *nblocks. *nblocks and the buffers array are in/out parameters. On entry,
1449 : * the buffers elements covered by *nblocks must hold either InvalidBuffer or
1450 : * buffers forwarded by an earlier call to StartReadBuffers() that was split
1451 : * and is now being continued. On return, *nblocks holds the number of blocks
1452 : * accepted by this operation. If it is less than the original number then
1453 : * this operation has been split, but buffer elements up to the original
1454 : * requested size may hold forwarded buffers to be used for a continuing
1455 : * operation. The caller must either start a new I/O beginning at the block
1456 : * immediately following the blocks accepted by this call and pass those
1457 : * buffers back in, or release them if it chooses not to. It shouldn't make
1458 : * any other use of or assumptions about forwarded buffers.
1459 : *
1460 : * If false is returned, no I/O is necessary and the buffers covered by
1461 : * *nblocks on exit are valid and ready to be accessed. If true is returned,
1462 : * an I/O has been started, and WaitReadBuffers() must be called with the same
1463 : * operation object before the buffers covered by *nblocks on exit can be
1464 : * accessed. Along with the operation object, the caller-supplied array of
1465 : * buffers must remain valid until WaitReadBuffers() is called, and any
1466 : * forwarded buffers must also be preserved for a continuing call unless
1467 : * they are explicitly released.
1468 : */
1469 : bool
1470 4052452 : StartReadBuffers(ReadBuffersOperation *operation,
1471 : Buffer *buffers,
1472 : BlockNumber blockNum,
1473 : int *nblocks,
1474 : int flags)
1475 : {
1476 4052452 : return StartReadBuffersImpl(operation, buffers, blockNum, nblocks, flags,
1477 : true /* expect forwarded buffers */ );
1478 : }
1479 :
1480 : /*
1481 : * Single block version of the StartReadBuffers(). This might save a few
1482 : * instructions when called from another translation unit, because it is
1483 : * specialized for nblocks == 1.
1484 : *
1485 : * This version does not support "forwarded" buffers: they cannot be created
1486 : * by reading only one block and *buffer is ignored on entry.
1487 : */
1488 : bool
1489 120278246 : StartReadBuffer(ReadBuffersOperation *operation,
1490 : Buffer *buffer,
1491 : BlockNumber blocknum,
1492 : int flags)
1493 : {
1494 120278246 : int nblocks = 1;
1495 : bool result;
1496 :
1497 120278246 : result = StartReadBuffersImpl(operation, buffer, blocknum, &nblocks, flags,
1498 : false /* single block, no forwarding */ );
1499 : Assert(nblocks == 1); /* single block can't be short */
1500 :
1501 120278216 : return result;
1502 : }
1503 :
1504 : /*
1505 : * Perform sanity checks on the ReadBuffersOperation.
1506 : */
1507 : static void
1508 7446684 : CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete)
1509 : {
1510 : #ifdef USE_ASSERT_CHECKING
1511 : Assert(operation->nblocks_done <= operation->nblocks);
1512 : Assert(!is_complete || operation->nblocks == operation->nblocks_done);
1513 :
1514 : for (int i = 0; i < operation->nblocks; i++)
1515 : {
1516 : Buffer buffer = operation->buffers[i];
1517 : BufferDesc *buf_hdr = BufferIsLocal(buffer) ?
1518 : GetLocalBufferDescriptor(-buffer - 1) :
1519 : GetBufferDescriptor(buffer - 1);
1520 :
1521 : Assert(BufferGetBlockNumber(buffer) == operation->blocknum + i);
1522 : Assert(pg_atomic_read_u32(&buf_hdr->state) & BM_TAG_VALID);
1523 :
1524 : if (i < operation->nblocks_done)
1525 : Assert(pg_atomic_read_u32(&buf_hdr->state) & BM_VALID);
1526 : }
1527 : #endif
1528 7446684 : }
1529 :
1530 : /* helper for ReadBuffersCanStartIO(), to avoid repetition */
1531 : static inline bool
1532 2822464 : ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
1533 : {
1534 2822464 : if (BufferIsLocal(buffer))
1535 16734 : return StartLocalBufferIO(GetLocalBufferDescriptor(-buffer - 1),
1536 : true, nowait);
1537 : else
1538 2805730 : return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
1539 : }
1540 :
1541 : /*
1542 : * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
1543 : */
1544 : static inline bool
1545 2822464 : ReadBuffersCanStartIO(Buffer buffer, bool nowait)
1546 : {
1547 : /*
1548 : * If this backend currently has staged IO, we need to submit the pending
1549 : * IO before waiting for the right to issue IO, to avoid the potential for
1550 : * deadlocks (and, more commonly, unnecessary delays for other backends).
1551 : */
1552 2822464 : if (!nowait && pgaio_have_staged())
1553 : {
1554 1150 : if (ReadBuffersCanStartIOOnce(buffer, true))
1555 1150 : return true;
1556 :
1557 : /*
1558 : * Unfortunately StartBufferIO() returning false doesn't allow to
1559 : * distinguish between the buffer already being valid and IO already
1560 : * being in progress. Since IO already being in progress is quite
1561 : * rare, this approach seems fine.
1562 : */
1563 0 : pgaio_submit_staged();
1564 : }
1565 :
1566 2821314 : return ReadBuffersCanStartIOOnce(buffer, nowait);
1567 : }
1568 :
1569 : /*
1570 : * Helper for WaitReadBuffers() that processes the results of a readv
1571 : * operation, raising an error if necessary.
1572 : */
1573 : static void
1574 2476606 : ProcessReadBuffersResult(ReadBuffersOperation *operation)
1575 : {
1576 2476606 : PgAioReturn *aio_ret = &operation->io_return;
1577 2476606 : PgAioResultStatus rs = aio_ret->result.status;
1578 2476606 : int newly_read_blocks = 0;
1579 :
1580 : Assert(pgaio_wref_valid(&operation->io_wref));
1581 : Assert(aio_ret->result.status != PGAIO_RS_UNKNOWN);
1582 :
1583 : /*
1584 : * SMGR reports the number of blocks successfully read as the result of
1585 : * the IO operation. Thus we can simply add that to ->nblocks_done.
1586 : */
1587 :
1588 2476606 : if (likely(rs != PGAIO_RS_ERROR))
1589 2476548 : newly_read_blocks = aio_ret->result.result;
1590 :
1591 2476606 : if (rs == PGAIO_RS_ERROR || rs == PGAIO_RS_WARNING)
1592 90 : pgaio_result_report(aio_ret->result, &aio_ret->target_data,
1593 : rs == PGAIO_RS_ERROR ? ERROR : WARNING);
1594 2476516 : else if (aio_ret->result.status == PGAIO_RS_PARTIAL)
1595 : {
1596 : /*
1597 : * We'll retry, so we just emit a debug message to the server log (or
1598 : * not even that in prod scenarios).
1599 : */
1600 20 : pgaio_result_report(aio_ret->result, &aio_ret->target_data, DEBUG1);
1601 20 : elog(DEBUG3, "partial read, will retry");
1602 : }
1603 :
1604 : Assert(newly_read_blocks > 0);
1605 : Assert(newly_read_blocks <= MAX_IO_COMBINE_LIMIT);
1606 :
1607 2476548 : operation->nblocks_done += newly_read_blocks;
1608 :
1609 : Assert(operation->nblocks_done <= operation->nblocks);
1610 2476548 : }
1611 :
1612 : void
1613 2476586 : WaitReadBuffers(ReadBuffersOperation *operation)
1614 : {
1615 2476586 : PgAioReturn *aio_ret = &operation->io_return;
1616 : IOContext io_context;
1617 : IOObject io_object;
1618 :
1619 2476586 : if (operation->persistence == RELPERSISTENCE_TEMP)
1620 : {
1621 2980 : io_context = IOCONTEXT_NORMAL;
1622 2980 : io_object = IOOBJECT_TEMP_RELATION;
1623 : }
1624 : else
1625 : {
1626 2473606 : io_context = IOContextForStrategy(operation->strategy);
1627 2473606 : io_object = IOOBJECT_RELATION;
1628 : }
1629 :
1630 : /*
1631 : * If we get here without an IO operation having been issued, the
1632 : * io_method == IOMETHOD_SYNC path must have been used. Otherwise the
1633 : * caller should not have called WaitReadBuffers().
1634 : *
1635 : * In the case of IOMETHOD_SYNC, we start - as we used to before the
1636 : * introducing of AIO - the IO in WaitReadBuffers(). This is done as part
1637 : * of the retry logic below, no extra code is required.
1638 : *
1639 : * This path is expected to eventually go away.
1640 : */
1641 2476586 : if (!pgaio_wref_valid(&operation->io_wref) && io_method != IOMETHOD_SYNC)
1642 0 : elog(ERROR, "waiting for read operation that didn't read");
1643 :
1644 : /*
1645 : * To handle partial reads, and IOMETHOD_SYNC, we re-issue IO until we're
1646 : * done. We may need multiple retries, not just because we could get
1647 : * multiple partial reads, but also because some of the remaining
1648 : * to-be-read buffers may have been read in by other backends, limiting
1649 : * the IO size.
1650 : */
1651 : while (true)
1652 2182 : {
1653 : int ignored_nblocks_progress;
1654 :
1655 2478768 : CheckReadBuffersOperation(operation, false);
1656 :
1657 : /*
1658 : * If there is an IO associated with the operation, we may need to
1659 : * wait for it.
1660 : */
1661 2478768 : if (pgaio_wref_valid(&operation->io_wref))
1662 : {
1663 : /*
1664 : * Track the time spent waiting for the IO to complete. As
1665 : * tracking a wait even if we don't actually need to wait
1666 : *
1667 : * a) is not cheap, due to the timestamping overhead
1668 : *
1669 : * b) reports some time as waiting, even if we never waited
1670 : *
1671 : * we first check if we already know the IO is complete.
1672 : */
1673 2476606 : if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
1674 1070302 : !pgaio_wref_check_done(&operation->io_wref))
1675 : {
1676 365154 : instr_time io_start = pgstat_prepare_io_time(track_io_timing);
1677 :
1678 365154 : pgaio_wref_wait(&operation->io_wref);
1679 :
1680 : /*
1681 : * The IO operation itself was already counted earlier, in
1682 : * AsyncReadBuffers(), this just accounts for the wait time.
1683 : */
1684 365154 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1685 : io_start, 0, 0);
1686 : }
1687 : else
1688 : {
1689 : Assert(pgaio_wref_check_done(&operation->io_wref));
1690 : }
1691 :
1692 : /*
1693 : * We now are sure the IO completed. Check the results. This
1694 : * includes reporting on errors if there were any.
1695 : */
1696 2476606 : ProcessReadBuffersResult(operation);
1697 : }
1698 :
1699 : /*
1700 : * Most of the time, the one IO we already started, will read in
1701 : * everything. But we need to deal with partial reads and buffers not
1702 : * needing IO anymore.
1703 : */
1704 2478710 : if (operation->nblocks_done == operation->nblocks)
1705 2476528 : break;
1706 :
1707 2182 : CHECK_FOR_INTERRUPTS();
1708 :
1709 : /*
1710 : * This may only complete the IO partially, either because some
1711 : * buffers were already valid, or because of a partial read.
1712 : *
1713 : * NB: In contrast to after the AsyncReadBuffers() call in
1714 : * StartReadBuffers(), we do *not* reduce
1715 : * ReadBuffersOperation->nblocks here, callers expect the full
1716 : * operation to be completed at this point (as more operations may
1717 : * have been queued).
1718 : */
1719 2182 : AsyncReadBuffers(operation, &ignored_nblocks_progress);
1720 : }
1721 :
1722 2476528 : CheckReadBuffersOperation(operation, true);
1723 :
1724 : /* NB: READ_DONE tracepoint was already executed in completion callback */
1725 2476528 : }
1726 :
1727 : /*
1728 : * Initiate IO for the ReadBuffersOperation
1729 : *
1730 : * This function only starts a single IO at a time. The size of the IO may be
1731 : * limited to below the to-be-read blocks, if one of the buffers has
1732 : * concurrently been read in. If the first to-be-read buffer is already valid,
1733 : * no IO will be issued.
1734 : *
1735 : * To support retries after partial reads, the first operation->nblocks_done
1736 : * buffers are skipped.
1737 : *
1738 : * On return *nblocks_progress is updated to reflect the number of buffers
1739 : * affected by the call. If the first buffer is valid, *nblocks_progress is
1740 : * set to 1 and operation->nblocks_done is incremented.
1741 : *
1742 : * Returns true if IO was initiated, false if no IO was necessary.
1743 : */
1744 : static bool
1745 2491438 : AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
1746 : {
1747 2491438 : Buffer *buffers = &operation->buffers[0];
1748 2491438 : int flags = operation->flags;
1749 2491438 : BlockNumber blocknum = operation->blocknum;
1750 2491438 : ForkNumber forknum = operation->forknum;
1751 2491438 : char persistence = operation->persistence;
1752 2491438 : int16 nblocks_done = operation->nblocks_done;
1753 2491438 : Buffer *io_buffers = &operation->buffers[nblocks_done];
1754 2491438 : int io_buffers_len = 0;
1755 : PgAioHandle *ioh;
1756 2491438 : uint32 ioh_flags = 0;
1757 : void *io_pages[MAX_IO_COMBINE_LIMIT];
1758 : IOContext io_context;
1759 : IOObject io_object;
1760 : bool did_start_io;
1761 :
1762 : /*
1763 : * When this IO is executed synchronously, either because the caller will
1764 : * immediately block waiting for the IO or because IOMETHOD_SYNC is used,
1765 : * the AIO subsystem needs to know.
1766 : */
1767 2491438 : if (flags & READ_BUFFERS_SYNCHRONOUSLY)
1768 1399878 : ioh_flags |= PGAIO_HF_SYNCHRONOUS;
1769 :
1770 2491438 : if (persistence == RELPERSISTENCE_TEMP)
1771 : {
1772 3568 : io_context = IOCONTEXT_NORMAL;
1773 3568 : io_object = IOOBJECT_TEMP_RELATION;
1774 3568 : ioh_flags |= PGAIO_HF_REFERENCES_LOCAL;
1775 : }
1776 : else
1777 : {
1778 2487870 : io_context = IOContextForStrategy(operation->strategy);
1779 2487870 : io_object = IOOBJECT_RELATION;
1780 : }
1781 :
1782 : /*
1783 : * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
1784 : * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
1785 : * set globally, but on a per-session basis. The completion callback,
1786 : * which may be run in other processes, e.g. in IO workers, may have a
1787 : * different value of the zero_damaged_pages GUC.
1788 : *
1789 : * XXX: We probably should eventually use a different flag for
1790 : * zero_damaged_pages, so we can report different log levels / error codes
1791 : * for zero_damaged_pages and ZERO_ON_ERROR.
1792 : */
1793 2491438 : if (zero_damaged_pages)
1794 32 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1795 :
1796 : /*
1797 : * For the same reason as with zero_damaged_pages we need to use this
1798 : * backend's ignore_checksum_failure value.
1799 : */
1800 2491438 : if (ignore_checksum_failure)
1801 16 : flags |= READ_BUFFERS_IGNORE_CHECKSUM_FAILURES;
1802 :
1803 :
1804 : /*
1805 : * To be allowed to report stats in the local completion callback we need
1806 : * to prepare to report stats now. This ensures we can safely report the
1807 : * checksum failure even in a critical section.
1808 : */
1809 2491438 : pgstat_prepare_report_checksum_failure(operation->smgr->smgr_rlocator.locator.dbOid);
1810 :
1811 : /*
1812 : * Get IO handle before ReadBuffersCanStartIO(), as pgaio_io_acquire()
1813 : * might block, which we don't want after setting IO_IN_PROGRESS.
1814 : *
1815 : * If we need to wait for IO before we can get a handle, submit
1816 : * already-staged IO first, so that other backends don't need to wait.
1817 : * There wouldn't be a deadlock risk, as pgaio_io_acquire() just needs to
1818 : * wait for already submitted IO, which doesn't require additional locks,
1819 : * but it could still cause undesirable waits.
1820 : *
1821 : * A secondary benefit is that this would allow us to measure the time in
1822 : * pgaio_io_acquire() without causing undue timer overhead in the common,
1823 : * non-blocking, case. However, currently the pgstats infrastructure
1824 : * doesn't really allow that, as it a) asserts that an operation can't
1825 : * have time without operations b) doesn't have an API to report
1826 : * "accumulated" time.
1827 : */
1828 2491438 : ioh = pgaio_io_acquire_nb(CurrentResourceOwner, &operation->io_return);
1829 2491438 : if (unlikely(!ioh))
1830 : {
1831 6194 : pgaio_submit_staged();
1832 :
1833 6194 : ioh = pgaio_io_acquire(CurrentResourceOwner, &operation->io_return);
1834 : }
1835 :
1836 : /*
1837 : * Check if we can start IO on the first to-be-read buffer.
1838 : *
1839 : * If an I/O is already in progress in another backend, we want to wait
1840 : * for the outcome: either done, or something went wrong and we will
1841 : * retry.
1842 : */
1843 2491438 : if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
1844 : {
1845 : /*
1846 : * Someone else has already completed this block, we're done.
1847 : *
1848 : * When IO is necessary, ->nblocks_done is updated in
1849 : * ProcessReadBuffersResult(), but that is not called if no IO is
1850 : * necessary. Thus update here.
1851 : */
1852 14210 : operation->nblocks_done += 1;
1853 14210 : *nblocks_progress = 1;
1854 :
1855 14210 : pgaio_io_release(ioh);
1856 14210 : pgaio_wref_clear(&operation->io_wref);
1857 14210 : did_start_io = false;
1858 :
1859 : /*
1860 : * Report and track this as a 'hit' for this backend, even though it
1861 : * must have started out as a miss in PinBufferForBlock(). The other
1862 : * backend will track this as a 'read'.
1863 : */
1864 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forknum, blocknum + operation->nblocks_done,
1865 : operation->smgr->smgr_rlocator.locator.spcOid,
1866 : operation->smgr->smgr_rlocator.locator.dbOid,
1867 : operation->smgr->smgr_rlocator.locator.relNumber,
1868 : operation->smgr->smgr_rlocator.backend,
1869 : true);
1870 :
1871 14210 : if (persistence == RELPERSISTENCE_TEMP)
1872 0 : pgBufferUsage.local_blks_hit += 1;
1873 : else
1874 14210 : pgBufferUsage.shared_blks_hit += 1;
1875 :
1876 14210 : if (operation->rel)
1877 14210 : pgstat_count_buffer_hit(operation->rel);
1878 :
1879 14210 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1880 :
1881 14210 : if (VacuumCostActive)
1882 0 : VacuumCostBalance += VacuumCostPageHit;
1883 : }
1884 : else
1885 : {
1886 : instr_time io_start;
1887 :
1888 : /* We found a buffer that we need to read in. */
1889 : Assert(io_buffers[0] == buffers[nblocks_done]);
1890 2477228 : io_pages[0] = BufferGetBlock(buffers[nblocks_done]);
1891 2477228 : io_buffers_len = 1;
1892 :
1893 : /*
1894 : * How many neighboring-on-disk blocks can we scatter-read into other
1895 : * buffers at the same time? In this case we don't wait if we see an
1896 : * I/O already in progress. We already set BM_IO_IN_PROGRESS for the
1897 : * head block, so we should get on with that I/O as soon as possible.
1898 : */
1899 2808254 : for (int i = nblocks_done + 1; i < operation->nblocks; i++)
1900 : {
1901 331026 : if (!ReadBuffersCanStartIO(buffers[i], true))
1902 0 : break;
1903 : /* Must be consecutive block numbers. */
1904 : Assert(BufferGetBlockNumber(buffers[i - 1]) ==
1905 : BufferGetBlockNumber(buffers[i]) - 1);
1906 : Assert(io_buffers[io_buffers_len] == buffers[i]);
1907 :
1908 331026 : io_pages[io_buffers_len++] = BufferGetBlock(buffers[i]);
1909 : }
1910 :
1911 : /* get a reference to wait for in WaitReadBuffers() */
1912 2477228 : pgaio_io_get_wref(ioh, &operation->io_wref);
1913 :
1914 : /* provide the list of buffers to the completion callbacks */
1915 2477228 : pgaio_io_set_handle_data_32(ioh, (uint32 *) io_buffers, io_buffers_len);
1916 :
1917 2477228 : pgaio_io_register_callbacks(ioh,
1918 : persistence == RELPERSISTENCE_TEMP ?
1919 : PGAIO_HCB_LOCAL_BUFFER_READV :
1920 : PGAIO_HCB_SHARED_BUFFER_READV,
1921 : flags);
1922 :
1923 2477228 : pgaio_io_set_flag(ioh, ioh_flags);
1924 :
1925 : /* ---
1926 : * Even though we're trying to issue IO asynchronously, track the time
1927 : * in smgrstartreadv():
1928 : * - if io_method == IOMETHOD_SYNC, we will always perform the IO
1929 : * immediately
1930 : * - the io method might not support the IO (e.g. worker IO for a temp
1931 : * table)
1932 : * ---
1933 : */
1934 2477228 : io_start = pgstat_prepare_io_time(track_io_timing);
1935 2477228 : smgrstartreadv(ioh, operation->smgr, forknum,
1936 : blocknum + nblocks_done,
1937 : io_pages, io_buffers_len);
1938 2477198 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1939 2477198 : io_start, 1, io_buffers_len * BLCKSZ);
1940 :
1941 2477198 : if (persistence == RELPERSISTENCE_TEMP)
1942 3568 : pgBufferUsage.local_blks_read += io_buffers_len;
1943 : else
1944 2473630 : pgBufferUsage.shared_blks_read += io_buffers_len;
1945 :
1946 : /*
1947 : * Track vacuum cost when issuing IO, not after waiting for it.
1948 : * Otherwise we could end up issuing a lot of IO in a short timespan,
1949 : * despite a low cost limit.
1950 : */
1951 2477198 : if (VacuumCostActive)
1952 44598 : VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
1953 :
1954 2477198 : *nblocks_progress = io_buffers_len;
1955 2477198 : did_start_io = true;
1956 : }
1957 :
1958 2491408 : return did_start_io;
1959 : }
1960 :
1961 : /*
1962 : * BufferAlloc -- subroutine for PinBufferForBlock. Handles lookup of a shared
1963 : * buffer. If no buffer exists already, selects a replacement victim and
1964 : * evicts the old page, but does NOT read in new page.
1965 : *
1966 : * "strategy" can be a buffer replacement strategy object, or NULL for
1967 : * the default strategy. The selected buffer's usage_count is advanced when
1968 : * using the default strategy, but otherwise possibly not (see PinBuffer).
1969 : *
1970 : * The returned buffer is pinned and is already marked as holding the
1971 : * desired page. If it already did have the desired page, *foundPtr is
1972 : * set true. Otherwise, *foundPtr is set false.
1973 : *
1974 : * io_context is passed as an output parameter to avoid calling
1975 : * IOContextForStrategy() when there is a shared buffers hit and no IO
1976 : * statistics need be captured.
1977 : *
1978 : * No locks are held either at entry or exit.
1979 : */
1980 : static pg_attribute_always_inline BufferDesc *
1981 122745488 : BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
1982 : BlockNumber blockNum,
1983 : BufferAccessStrategy strategy,
1984 : bool *foundPtr, IOContext io_context)
1985 : {
1986 : BufferTag newTag; /* identity of requested block */
1987 : uint32 newHash; /* hash value for newTag */
1988 : LWLock *newPartitionLock; /* buffer partition lock for it */
1989 : int existing_buf_id;
1990 : Buffer victim_buffer;
1991 : BufferDesc *victim_buf_hdr;
1992 : uint32 victim_buf_state;
1993 :
1994 : /* Make sure we will have room to remember the buffer pin */
1995 122745488 : ResourceOwnerEnlarge(CurrentResourceOwner);
1996 122745488 : ReservePrivateRefCountEntry();
1997 :
1998 : /* create a tag so we can lookup the buffer */
1999 122745488 : InitBufferTag(&newTag, &smgr->smgr_rlocator.locator, forkNum, blockNum);
2000 :
2001 : /* determine its hash code and partition lock ID */
2002 122745488 : newHash = BufTableHashCode(&newTag);
2003 122745488 : newPartitionLock = BufMappingPartitionLock(newHash);
2004 :
2005 : /* see if the block is in the buffer pool already */
2006 122745488 : LWLockAcquire(newPartitionLock, LW_SHARED);
2007 122745488 : existing_buf_id = BufTableLookup(&newTag, newHash);
2008 122745488 : if (existing_buf_id >= 0)
2009 : {
2010 : BufferDesc *buf;
2011 : bool valid;
2012 :
2013 : /*
2014 : * Found it. Now, pin the buffer so no one can steal it from the
2015 : * buffer pool, and check to see if the correct data has been loaded
2016 : * into the buffer.
2017 : */
2018 119387180 : buf = GetBufferDescriptor(existing_buf_id);
2019 :
2020 119387180 : valid = PinBuffer(buf, strategy, false);
2021 :
2022 : /* Can release the mapping lock as soon as we've pinned it */
2023 119387180 : LWLockRelease(newPartitionLock);
2024 :
2025 119387180 : *foundPtr = true;
2026 :
2027 119387180 : if (!valid)
2028 : {
2029 : /*
2030 : * We can only get here if (a) someone else is still reading in
2031 : * the page, (b) a previous read attempt failed, or (c) someone
2032 : * called StartReadBuffers() but not yet WaitReadBuffers().
2033 : */
2034 12148 : *foundPtr = false;
2035 : }
2036 :
2037 119387180 : return buf;
2038 : }
2039 :
2040 : /*
2041 : * Didn't find it in the buffer pool. We'll have to initialize a new
2042 : * buffer. Remember to unlock the mapping lock while doing the work.
2043 : */
2044 3358308 : LWLockRelease(newPartitionLock);
2045 :
2046 : /*
2047 : * Acquire a victim buffer. Somebody else might try to do the same, we
2048 : * don't hold any conflicting locks. If so we'll have to undo our work
2049 : * later.
2050 : */
2051 3358308 : victim_buffer = GetVictimBuffer(strategy, io_context);
2052 3358308 : victim_buf_hdr = GetBufferDescriptor(victim_buffer - 1);
2053 :
2054 : /*
2055 : * Try to make a hashtable entry for the buffer under its new tag. If
2056 : * somebody else inserted another buffer for the tag, we'll release the
2057 : * victim buffer we acquired and use the already inserted one.
2058 : */
2059 3358308 : LWLockAcquire(newPartitionLock, LW_EXCLUSIVE);
2060 3358308 : existing_buf_id = BufTableInsert(&newTag, newHash, victim_buf_hdr->buf_id);
2061 3358308 : if (existing_buf_id >= 0)
2062 : {
2063 : BufferDesc *existing_buf_hdr;
2064 : bool valid;
2065 :
2066 : /*
2067 : * Got a collision. Someone has already done what we were about to do.
2068 : * We'll just handle this as if it were found in the buffer pool in
2069 : * the first place. First, give up the buffer we were planning to
2070 : * use.
2071 : *
2072 : * We could do this after releasing the partition lock, but then we'd
2073 : * have to call ResourceOwnerEnlarge() & ReservePrivateRefCountEntry()
2074 : * before acquiring the lock, for the rare case of such a collision.
2075 : */
2076 2960 : UnpinBuffer(victim_buf_hdr);
2077 :
2078 : /* remaining code should match code at top of routine */
2079 :
2080 2960 : existing_buf_hdr = GetBufferDescriptor(existing_buf_id);
2081 :
2082 2960 : valid = PinBuffer(existing_buf_hdr, strategy, false);
2083 :
2084 : /* Can release the mapping lock as soon as we've pinned it */
2085 2960 : LWLockRelease(newPartitionLock);
2086 :
2087 2960 : *foundPtr = true;
2088 :
2089 2960 : if (!valid)
2090 : {
2091 : /*
2092 : * We can only get here if (a) someone else is still reading in
2093 : * the page, (b) a previous read attempt failed, or (c) someone
2094 : * called StartReadBuffers() but not yet WaitReadBuffers().
2095 : */
2096 2158 : *foundPtr = false;
2097 : }
2098 :
2099 2960 : return existing_buf_hdr;
2100 : }
2101 :
2102 : /*
2103 : * Need to lock the buffer header too in order to change its tag.
2104 : */
2105 3355348 : victim_buf_state = LockBufHdr(victim_buf_hdr);
2106 :
2107 : /* some sanity checks while we hold the buffer header lock */
2108 : Assert(BUF_STATE_GET_REFCOUNT(victim_buf_state) == 1);
2109 : Assert(!(victim_buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY | BM_IO_IN_PROGRESS)));
2110 :
2111 3355348 : victim_buf_hdr->tag = newTag;
2112 :
2113 : /*
2114 : * Make sure BM_PERMANENT is set for buffers that must be written at every
2115 : * checkpoint. Unlogged buffers only need to be written at shutdown
2116 : * checkpoints, except for their "init" forks, which need to be treated
2117 : * just like permanent relations.
2118 : */
2119 3355348 : victim_buf_state |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2120 3355348 : if (relpersistence == RELPERSISTENCE_PERMANENT || forkNum == INIT_FORKNUM)
2121 3354662 : victim_buf_state |= BM_PERMANENT;
2122 :
2123 3355348 : UnlockBufHdr(victim_buf_hdr, victim_buf_state);
2124 :
2125 3355348 : LWLockRelease(newPartitionLock);
2126 :
2127 : /*
2128 : * Buffer contents are currently invalid.
2129 : */
2130 3355348 : *foundPtr = false;
2131 :
2132 3355348 : return victim_buf_hdr;
2133 : }
2134 :
2135 : /*
2136 : * InvalidateBuffer -- mark a shared buffer invalid.
2137 : *
2138 : * The buffer header spinlock must be held at entry. We drop it before
2139 : * returning. (This is sane because the caller must have locked the
2140 : * buffer in order to be sure it should be dropped.)
2141 : *
2142 : * This is used only in contexts such as dropping a relation. We assume
2143 : * that no other backend could possibly be interested in using the page,
2144 : * so the only reason the buffer might be pinned is if someone else is
2145 : * trying to write it out. We have to let them finish before we can
2146 : * reclaim the buffer.
2147 : *
2148 : * The buffer could get reclaimed by someone else while we are waiting
2149 : * to acquire the necessary locks; if so, don't mess it up.
2150 : */
2151 : static void
2152 210490 : InvalidateBuffer(BufferDesc *buf)
2153 : {
2154 : BufferTag oldTag;
2155 : uint32 oldHash; /* hash value for oldTag */
2156 : LWLock *oldPartitionLock; /* buffer partition lock for it */
2157 : uint32 oldFlags;
2158 : uint32 buf_state;
2159 :
2160 : /* Save the original buffer tag before dropping the spinlock */
2161 210490 : oldTag = buf->tag;
2162 :
2163 210490 : buf_state = pg_atomic_read_u32(&buf->state);
2164 : Assert(buf_state & BM_LOCKED);
2165 210490 : UnlockBufHdr(buf, buf_state);
2166 :
2167 : /*
2168 : * Need to compute the old tag's hashcode and partition lock ID. XXX is it
2169 : * worth storing the hashcode in BufferDesc so we need not recompute it
2170 : * here? Probably not.
2171 : */
2172 210490 : oldHash = BufTableHashCode(&oldTag);
2173 210490 : oldPartitionLock = BufMappingPartitionLock(oldHash);
2174 :
2175 210494 : retry:
2176 :
2177 : /*
2178 : * Acquire exclusive mapping lock in preparation for changing the buffer's
2179 : * association.
2180 : */
2181 210494 : LWLockAcquire(oldPartitionLock, LW_EXCLUSIVE);
2182 :
2183 : /* Re-lock the buffer header */
2184 210494 : buf_state = LockBufHdr(buf);
2185 :
2186 : /* If it's changed while we were waiting for lock, do nothing */
2187 210494 : if (!BufferTagsEqual(&buf->tag, &oldTag))
2188 : {
2189 4 : UnlockBufHdr(buf, buf_state);
2190 4 : LWLockRelease(oldPartitionLock);
2191 4 : return;
2192 : }
2193 :
2194 : /*
2195 : * We assume the reason for it to be pinned is that either we were
2196 : * asynchronously reading the page in before erroring out or someone else
2197 : * is flushing the page out. Wait for the IO to finish. (This could be
2198 : * an infinite loop if the refcount is messed up... it would be nice to
2199 : * time out after awhile, but there seems no way to be sure how many loops
2200 : * may be needed. Note that if the other guy has pinned the buffer but
2201 : * not yet done StartBufferIO, WaitIO will fall through and we'll
2202 : * effectively be busy-looping here.)
2203 : */
2204 210490 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 0)
2205 : {
2206 4 : UnlockBufHdr(buf, buf_state);
2207 4 : LWLockRelease(oldPartitionLock);
2208 : /* safety check: should definitely not be our *own* pin */
2209 4 : if (GetPrivateRefCount(BufferDescriptorGetBuffer(buf)) > 0)
2210 0 : elog(ERROR, "buffer is pinned in InvalidateBuffer");
2211 4 : WaitIO(buf);
2212 4 : goto retry;
2213 : }
2214 :
2215 : /*
2216 : * Clear out the buffer's tag and flags. We must do this to ensure that
2217 : * linear scans of the buffer array don't think the buffer is valid.
2218 : */
2219 210486 : oldFlags = buf_state & BUF_FLAG_MASK;
2220 210486 : ClearBufferTag(&buf->tag);
2221 210486 : buf_state &= ~(BUF_FLAG_MASK | BUF_USAGECOUNT_MASK);
2222 210486 : UnlockBufHdr(buf, buf_state);
2223 :
2224 : /*
2225 : * Remove the buffer from the lookup hashtable, if it was in there.
2226 : */
2227 210486 : if (oldFlags & BM_TAG_VALID)
2228 210486 : BufTableDelete(&oldTag, oldHash);
2229 :
2230 : /*
2231 : * Done with mapping lock.
2232 : */
2233 210486 : LWLockRelease(oldPartitionLock);
2234 : }
2235 :
2236 : /*
2237 : * Helper routine for GetVictimBuffer()
2238 : *
2239 : * Needs to be called on a buffer with a valid tag, pinned, but without the
2240 : * buffer header spinlock held.
2241 : *
2242 : * Returns true if the buffer can be reused, in which case the buffer is only
2243 : * pinned by this backend and marked as invalid, false otherwise.
2244 : */
2245 : static bool
2246 2353904 : InvalidateVictimBuffer(BufferDesc *buf_hdr)
2247 : {
2248 : uint32 buf_state;
2249 : uint32 hash;
2250 : LWLock *partition_lock;
2251 : BufferTag tag;
2252 :
2253 : Assert(GetPrivateRefCount(BufferDescriptorGetBuffer(buf_hdr)) == 1);
2254 :
2255 : /* have buffer pinned, so it's safe to read tag without lock */
2256 2353904 : tag = buf_hdr->tag;
2257 :
2258 2353904 : hash = BufTableHashCode(&tag);
2259 2353904 : partition_lock = BufMappingPartitionLock(hash);
2260 :
2261 2353904 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2262 :
2263 : /* lock the buffer header */
2264 2353904 : buf_state = LockBufHdr(buf_hdr);
2265 :
2266 : /*
2267 : * We have the buffer pinned nobody else should have been able to unset
2268 : * this concurrently.
2269 : */
2270 : Assert(buf_state & BM_TAG_VALID);
2271 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2272 : Assert(BufferTagsEqual(&buf_hdr->tag, &tag));
2273 :
2274 : /*
2275 : * If somebody else pinned the buffer since, or even worse, dirtied it,
2276 : * give up on this buffer: It's clearly in use.
2277 : */
2278 2353904 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 1 || (buf_state & BM_DIRTY))
2279 : {
2280 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2281 :
2282 920 : UnlockBufHdr(buf_hdr, buf_state);
2283 920 : LWLockRelease(partition_lock);
2284 :
2285 920 : return false;
2286 : }
2287 :
2288 : /*
2289 : * Clear out the buffer's tag and flags and usagecount. This is not
2290 : * strictly required, as BM_TAG_VALID/BM_VALID needs to be checked before
2291 : * doing anything with the buffer. But currently it's beneficial, as the
2292 : * cheaper pre-check for several linear scans of shared buffers use the
2293 : * tag (see e.g. FlushDatabaseBuffers()).
2294 : */
2295 2352984 : ClearBufferTag(&buf_hdr->tag);
2296 2352984 : buf_state &= ~(BUF_FLAG_MASK | BUF_USAGECOUNT_MASK);
2297 2352984 : UnlockBufHdr(buf_hdr, buf_state);
2298 :
2299 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2300 :
2301 : /* finally delete buffer from the buffer mapping table */
2302 2352984 : BufTableDelete(&tag, hash);
2303 :
2304 2352984 : LWLockRelease(partition_lock);
2305 :
2306 : Assert(!(buf_state & (BM_DIRTY | BM_VALID | BM_TAG_VALID)));
2307 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2308 : Assert(BUF_STATE_GET_REFCOUNT(pg_atomic_read_u32(&buf_hdr->state)) > 0);
2309 :
2310 2352984 : return true;
2311 : }
2312 :
2313 : static Buffer
2314 3803020 : GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context)
2315 : {
2316 : BufferDesc *buf_hdr;
2317 : Buffer buf;
2318 : uint32 buf_state;
2319 : bool from_ring;
2320 :
2321 : /*
2322 : * Ensure, before we pin a victim buffer, that there's a free refcount
2323 : * entry and resource owner slot for the pin.
2324 : */
2325 3803020 : ReservePrivateRefCountEntry();
2326 3803020 : ResourceOwnerEnlarge(CurrentResourceOwner);
2327 :
2328 : /* we return here if a prospective victim buffer gets used concurrently */
2329 12288 : again:
2330 :
2331 : /*
2332 : * Select a victim buffer. The buffer is returned pinned and owned by
2333 : * this backend.
2334 : */
2335 3815308 : buf_hdr = StrategyGetBuffer(strategy, &buf_state, &from_ring);
2336 3815308 : buf = BufferDescriptorGetBuffer(buf_hdr);
2337 :
2338 : /*
2339 : * We shouldn't have any other pins for this buffer.
2340 : */
2341 3815308 : CheckBufferIsPinnedOnce(buf);
2342 :
2343 : /*
2344 : * If the buffer was dirty, try to write it out. There is a race
2345 : * condition here, in that someone might dirty it after we released the
2346 : * buffer header lock above, or even while we are writing it out (since
2347 : * our share-lock won't prevent hint-bit updates). We will recheck the
2348 : * dirty bit after re-locking the buffer header.
2349 : */
2350 3815308 : if (buf_state & BM_DIRTY)
2351 : {
2352 : LWLock *content_lock;
2353 :
2354 : Assert(buf_state & BM_TAG_VALID);
2355 : Assert(buf_state & BM_VALID);
2356 :
2357 : /*
2358 : * We need a share-lock on the buffer contents to write it out (else
2359 : * we might write invalid data, eg because someone else is compacting
2360 : * the page contents while we write). We must use a conditional lock
2361 : * acquisition here to avoid deadlock. Even though the buffer was not
2362 : * pinned (and therefore surely not locked) when StrategyGetBuffer
2363 : * returned it, someone else could have pinned and exclusive-locked it
2364 : * by the time we get here. If we try to get the lock unconditionally,
2365 : * we'd block waiting for them; if they later block waiting for us,
2366 : * deadlock ensues. (This has been observed to happen when two
2367 : * backends are both trying to split btree index pages, and the second
2368 : * one just happens to be trying to split the page the first one got
2369 : * from StrategyGetBuffer.)
2370 : */
2371 528690 : content_lock = BufferDescriptorGetContentLock(buf_hdr);
2372 528690 : if (!LWLockConditionalAcquire(content_lock, LW_SHARED))
2373 : {
2374 : /*
2375 : * Someone else has locked the buffer, so give it up and loop back
2376 : * to get another one.
2377 : */
2378 0 : UnpinBuffer(buf_hdr);
2379 0 : goto again;
2380 : }
2381 :
2382 : /*
2383 : * If using a nondefault strategy, and writing the buffer would
2384 : * require a WAL flush, let the strategy decide whether to go ahead
2385 : * and write/reuse the buffer or to choose another victim. We need a
2386 : * lock to inspect the page LSN, so this can't be done inside
2387 : * StrategyGetBuffer.
2388 : */
2389 528690 : if (strategy != NULL)
2390 : {
2391 : XLogRecPtr lsn;
2392 :
2393 : /* Read the LSN while holding buffer header lock */
2394 155508 : buf_state = LockBufHdr(buf_hdr);
2395 155508 : lsn = BufferGetLSN(buf_hdr);
2396 155508 : UnlockBufHdr(buf_hdr, buf_state);
2397 :
2398 155508 : if (XLogNeedsFlush(lsn)
2399 18190 : && StrategyRejectBuffer(strategy, buf_hdr, from_ring))
2400 : {
2401 11368 : LWLockRelease(content_lock);
2402 11368 : UnpinBuffer(buf_hdr);
2403 11368 : goto again;
2404 : }
2405 : }
2406 :
2407 : /* OK, do the I/O */
2408 517322 : FlushBuffer(buf_hdr, NULL, IOOBJECT_RELATION, io_context);
2409 517322 : LWLockRelease(content_lock);
2410 :
2411 517322 : ScheduleBufferTagForWriteback(&BackendWritebackContext, io_context,
2412 : &buf_hdr->tag);
2413 : }
2414 :
2415 :
2416 3803940 : if (buf_state & BM_VALID)
2417 : {
2418 : /*
2419 : * When a BufferAccessStrategy is in use, blocks evicted from shared
2420 : * buffers are counted as IOOP_EVICT in the corresponding context
2421 : * (e.g. IOCONTEXT_BULKWRITE). Shared buffers are evicted by a
2422 : * strategy in two cases: 1) while initially claiming buffers for the
2423 : * strategy ring 2) to replace an existing strategy ring buffer
2424 : * because it is pinned or in use and cannot be reused.
2425 : *
2426 : * Blocks evicted from buffers already in the strategy ring are
2427 : * counted as IOOP_REUSE in the corresponding strategy context.
2428 : *
2429 : * At this point, we can accurately count evictions and reuses,
2430 : * because we have successfully claimed the valid buffer. Previously,
2431 : * we may have been forced to release the buffer due to concurrent
2432 : * pinners or erroring out.
2433 : */
2434 2349626 : pgstat_count_io_op(IOOBJECT_RELATION, io_context,
2435 2349626 : from_ring ? IOOP_REUSE : IOOP_EVICT, 1, 0);
2436 : }
2437 :
2438 : /*
2439 : * If the buffer has an entry in the buffer mapping table, delete it. This
2440 : * can fail because another backend could have pinned or dirtied the
2441 : * buffer.
2442 : */
2443 3803940 : if ((buf_state & BM_TAG_VALID) && !InvalidateVictimBuffer(buf_hdr))
2444 : {
2445 920 : UnpinBuffer(buf_hdr);
2446 920 : goto again;
2447 : }
2448 :
2449 : /* a final set of sanity checks */
2450 : #ifdef USE_ASSERT_CHECKING
2451 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
2452 :
2453 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2454 : Assert(!(buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY)));
2455 :
2456 : CheckBufferIsPinnedOnce(buf);
2457 : #endif
2458 :
2459 3803020 : return buf;
2460 : }
2461 :
2462 : /*
2463 : * Return the maximum number of buffers that a backend should try to pin once,
2464 : * to avoid exceeding its fair share. This is the highest value that
2465 : * GetAdditionalPinLimit() could ever return. Note that it may be zero on a
2466 : * system with a very small buffer pool relative to max_connections.
2467 : */
2468 : uint32
2469 1336990 : GetPinLimit(void)
2470 : {
2471 1336990 : return MaxProportionalPins;
2472 : }
2473 :
2474 : /*
2475 : * Return the maximum number of additional buffers that this backend should
2476 : * pin if it wants to stay under the per-backend limit, considering the number
2477 : * of buffers it has already pinned. Unlike LimitAdditionalPins(), the limit
2478 : * return by this function can be zero.
2479 : */
2480 : uint32
2481 8080602 : GetAdditionalPinLimit(void)
2482 : {
2483 : uint32 estimated_pins_held;
2484 :
2485 : /*
2486 : * We get the number of "overflowed" pins for free, but don't know the
2487 : * number of pins in PrivateRefCountArray. The cost of calculating that
2488 : * exactly doesn't seem worth it, so just assume the max.
2489 : */
2490 8080602 : estimated_pins_held = PrivateRefCountOverflowed + REFCOUNT_ARRAY_ENTRIES;
2491 :
2492 : /* Is this backend already holding more than its fair share? */
2493 8080602 : if (estimated_pins_held > MaxProportionalPins)
2494 2368510 : return 0;
2495 :
2496 5712092 : return MaxProportionalPins - estimated_pins_held;
2497 : }
2498 :
2499 : /*
2500 : * Limit the number of pins a batch operation may additionally acquire, to
2501 : * avoid running out of pinnable buffers.
2502 : *
2503 : * One additional pin is always allowed, on the assumption that the operation
2504 : * requires at least one to make progress.
2505 : */
2506 : void
2507 402376 : LimitAdditionalPins(uint32 *additional_pins)
2508 : {
2509 : uint32 limit;
2510 :
2511 402376 : if (*additional_pins <= 1)
2512 382678 : return;
2513 :
2514 19698 : limit = GetAdditionalPinLimit();
2515 19698 : limit = Max(limit, 1);
2516 19698 : if (limit < *additional_pins)
2517 10750 : *additional_pins = limit;
2518 : }
2519 :
2520 : /*
2521 : * Logic shared between ExtendBufferedRelBy(), ExtendBufferedRelTo(). Just to
2522 : * avoid duplicating the tracing and relpersistence related logic.
2523 : */
2524 : static BlockNumber
2525 425196 : ExtendBufferedRelCommon(BufferManagerRelation bmr,
2526 : ForkNumber fork,
2527 : BufferAccessStrategy strategy,
2528 : uint32 flags,
2529 : uint32 extend_by,
2530 : BlockNumber extend_upto,
2531 : Buffer *buffers,
2532 : uint32 *extended_by)
2533 : {
2534 : BlockNumber first_block;
2535 :
2536 : TRACE_POSTGRESQL_BUFFER_EXTEND_START(fork,
2537 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2538 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2539 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2540 : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2541 : extend_by);
2542 :
2543 425196 : if (bmr.relpersistence == RELPERSISTENCE_TEMP)
2544 22820 : first_block = ExtendBufferedRelLocal(bmr, fork, flags,
2545 : extend_by, extend_upto,
2546 : buffers, &extend_by);
2547 : else
2548 402376 : first_block = ExtendBufferedRelShared(bmr, fork, strategy, flags,
2549 : extend_by, extend_upto,
2550 : buffers, &extend_by);
2551 425196 : *extended_by = extend_by;
2552 :
2553 : TRACE_POSTGRESQL_BUFFER_EXTEND_DONE(fork,
2554 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2555 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2556 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2557 : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2558 : *extended_by,
2559 : first_block);
2560 :
2561 425196 : return first_block;
2562 : }
2563 :
2564 : /*
2565 : * Implementation of ExtendBufferedRelBy() and ExtendBufferedRelTo() for
2566 : * shared buffers.
2567 : */
2568 : static BlockNumber
2569 402376 : ExtendBufferedRelShared(BufferManagerRelation bmr,
2570 : ForkNumber fork,
2571 : BufferAccessStrategy strategy,
2572 : uint32 flags,
2573 : uint32 extend_by,
2574 : BlockNumber extend_upto,
2575 : Buffer *buffers,
2576 : uint32 *extended_by)
2577 : {
2578 : BlockNumber first_block;
2579 402376 : IOContext io_context = IOContextForStrategy(strategy);
2580 : instr_time io_start;
2581 :
2582 402376 : LimitAdditionalPins(&extend_by);
2583 :
2584 : /*
2585 : * Acquire victim buffers for extension without holding extension lock.
2586 : * Writing out victim buffers is the most expensive part of extending the
2587 : * relation, particularly when doing so requires WAL flushes. Zeroing out
2588 : * the buffers is also quite expensive, so do that before holding the
2589 : * extension lock as well.
2590 : *
2591 : * These pages are pinned by us and not valid. While we hold the pin they
2592 : * can't be acquired as victim buffers by another backend.
2593 : */
2594 847088 : for (uint32 i = 0; i < extend_by; i++)
2595 : {
2596 : Block buf_block;
2597 :
2598 444712 : buffers[i] = GetVictimBuffer(strategy, io_context);
2599 444712 : buf_block = BufHdrGetBlock(GetBufferDescriptor(buffers[i] - 1));
2600 :
2601 : /* new buffers are zero-filled */
2602 444712 : MemSet(buf_block, 0, BLCKSZ);
2603 : }
2604 :
2605 : /*
2606 : * Lock relation against concurrent extensions, unless requested not to.
2607 : *
2608 : * We use the same extension lock for all forks. That's unnecessarily
2609 : * restrictive, but currently extensions for forks don't happen often
2610 : * enough to make it worth locking more granularly.
2611 : *
2612 : * Note that another backend might have extended the relation by the time
2613 : * we get the lock.
2614 : */
2615 402376 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2616 299568 : LockRelationForExtension(bmr.rel, ExclusiveLock);
2617 :
2618 : /*
2619 : * If requested, invalidate size cache, so that smgrnblocks asks the
2620 : * kernel.
2621 : */
2622 402376 : if (flags & EB_CLEAR_SIZE_CACHE)
2623 15478 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
2624 :
2625 402376 : first_block = smgrnblocks(BMR_GET_SMGR(bmr), fork);
2626 :
2627 : /*
2628 : * Now that we have the accurate relation size, check if the caller wants
2629 : * us to extend to only up to a specific size. If there were concurrent
2630 : * extensions, we might have acquired too many buffers and need to release
2631 : * them.
2632 : */
2633 402376 : if (extend_upto != InvalidBlockNumber)
2634 : {
2635 106474 : uint32 orig_extend_by = extend_by;
2636 :
2637 106474 : if (first_block > extend_upto)
2638 0 : extend_by = 0;
2639 106474 : else if ((uint64) first_block + extend_by > extend_upto)
2640 14 : extend_by = extend_upto - first_block;
2641 :
2642 106514 : for (uint32 i = extend_by; i < orig_extend_by; i++)
2643 : {
2644 40 : BufferDesc *buf_hdr = GetBufferDescriptor(buffers[i] - 1);
2645 :
2646 40 : UnpinBuffer(buf_hdr);
2647 : }
2648 :
2649 106474 : if (extend_by == 0)
2650 : {
2651 14 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2652 14 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2653 14 : *extended_by = extend_by;
2654 14 : return first_block;
2655 : }
2656 : }
2657 :
2658 : /* Fail if relation is already at maximum possible length */
2659 402362 : if ((uint64) first_block + extend_by >= MaxBlockNumber)
2660 0 : ereport(ERROR,
2661 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
2662 : errmsg("cannot extend relation %s beyond %u blocks",
2663 : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str,
2664 : MaxBlockNumber)));
2665 :
2666 : /*
2667 : * Insert buffers into buffer table, mark as IO_IN_PROGRESS.
2668 : *
2669 : * This needs to happen before we extend the relation, because as soon as
2670 : * we do, other backends can start to read in those pages.
2671 : */
2672 847034 : for (uint32 i = 0; i < extend_by; i++)
2673 : {
2674 444672 : Buffer victim_buf = buffers[i];
2675 444672 : BufferDesc *victim_buf_hdr = GetBufferDescriptor(victim_buf - 1);
2676 : BufferTag tag;
2677 : uint32 hash;
2678 : LWLock *partition_lock;
2679 : int existing_id;
2680 :
2681 : /* in case we need to pin an existing buffer below */
2682 444672 : ResourceOwnerEnlarge(CurrentResourceOwner);
2683 444672 : ReservePrivateRefCountEntry();
2684 :
2685 444672 : InitBufferTag(&tag, &BMR_GET_SMGR(bmr)->smgr_rlocator.locator, fork,
2686 : first_block + i);
2687 444672 : hash = BufTableHashCode(&tag);
2688 444672 : partition_lock = BufMappingPartitionLock(hash);
2689 :
2690 444672 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2691 :
2692 444672 : existing_id = BufTableInsert(&tag, hash, victim_buf_hdr->buf_id);
2693 :
2694 : /*
2695 : * We get here only in the corner case where we are trying to extend
2696 : * the relation but we found a pre-existing buffer. This can happen
2697 : * because a prior attempt at extending the relation failed, and
2698 : * because mdread doesn't complain about reads beyond EOF (when
2699 : * zero_damaged_pages is ON) and so a previous attempt to read a block
2700 : * beyond EOF could have left a "valid" zero-filled buffer.
2701 : *
2702 : * This has also been observed when relation was overwritten by
2703 : * external process. Since the legitimate cases should always have
2704 : * left a zero-filled buffer, complain if not PageIsNew.
2705 : */
2706 444672 : if (existing_id >= 0)
2707 : {
2708 0 : BufferDesc *existing_hdr = GetBufferDescriptor(existing_id);
2709 : Block buf_block;
2710 : bool valid;
2711 :
2712 : /*
2713 : * Pin the existing buffer before releasing the partition lock,
2714 : * preventing it from being evicted.
2715 : */
2716 0 : valid = PinBuffer(existing_hdr, strategy, false);
2717 :
2718 0 : LWLockRelease(partition_lock);
2719 0 : UnpinBuffer(victim_buf_hdr);
2720 :
2721 0 : buffers[i] = BufferDescriptorGetBuffer(existing_hdr);
2722 0 : buf_block = BufHdrGetBlock(existing_hdr);
2723 :
2724 0 : if (valid && !PageIsNew((Page) buf_block))
2725 0 : ereport(ERROR,
2726 : (errmsg("unexpected data beyond EOF in block %u of relation \"%s\"",
2727 : existing_hdr->tag.blockNum,
2728 : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str)));
2729 :
2730 : /*
2731 : * We *must* do smgr[zero]extend before succeeding, else the page
2732 : * will not be reserved by the kernel, and the next P_NEW call
2733 : * will decide to return the same page. Clear the BM_VALID bit,
2734 : * do StartBufferIO() and proceed.
2735 : *
2736 : * Loop to handle the very small possibility that someone re-sets
2737 : * BM_VALID between our clearing it and StartBufferIO inspecting
2738 : * it.
2739 : */
2740 : do
2741 : {
2742 0 : uint32 buf_state = LockBufHdr(existing_hdr);
2743 :
2744 0 : buf_state &= ~BM_VALID;
2745 0 : UnlockBufHdr(existing_hdr, buf_state);
2746 0 : } while (!StartBufferIO(existing_hdr, true, false));
2747 : }
2748 : else
2749 : {
2750 : uint32 buf_state;
2751 :
2752 444672 : buf_state = LockBufHdr(victim_buf_hdr);
2753 :
2754 : /* some sanity checks while we hold the buffer header lock */
2755 : Assert(!(buf_state & (BM_VALID | BM_TAG_VALID | BM_DIRTY | BM_JUST_DIRTIED)));
2756 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2757 :
2758 444672 : victim_buf_hdr->tag = tag;
2759 :
2760 444672 : buf_state |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2761 444672 : if (bmr.relpersistence == RELPERSISTENCE_PERMANENT || fork == INIT_FORKNUM)
2762 434064 : buf_state |= BM_PERMANENT;
2763 :
2764 444672 : UnlockBufHdr(victim_buf_hdr, buf_state);
2765 :
2766 444672 : LWLockRelease(partition_lock);
2767 :
2768 : /* XXX: could combine the locked operations in it with the above */
2769 444672 : StartBufferIO(victim_buf_hdr, true, false);
2770 : }
2771 : }
2772 :
2773 402362 : io_start = pgstat_prepare_io_time(track_io_timing);
2774 :
2775 : /*
2776 : * Note: if smgrzeroextend fails, we will end up with buffers that are
2777 : * allocated but not marked BM_VALID. The next relation extension will
2778 : * still select the same block number (because the relation didn't get any
2779 : * longer on disk) and so future attempts to extend the relation will find
2780 : * the same buffers (if they have not been recycled) but come right back
2781 : * here to try smgrzeroextend again.
2782 : *
2783 : * We don't need to set checksum for all-zero pages.
2784 : */
2785 402362 : smgrzeroextend(BMR_GET_SMGR(bmr), fork, first_block, extend_by, false);
2786 :
2787 : /*
2788 : * Release the file-extension lock; it's now OK for someone else to extend
2789 : * the relation some more.
2790 : *
2791 : * We remove IO_IN_PROGRESS after this, as waking up waiting backends can
2792 : * take noticeable time.
2793 : */
2794 402362 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2795 299554 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2796 :
2797 402362 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context, IOOP_EXTEND,
2798 402362 : io_start, 1, extend_by * BLCKSZ);
2799 :
2800 : /* Set BM_VALID, terminate IO, and wake up any waiters */
2801 847034 : for (uint32 i = 0; i < extend_by; i++)
2802 : {
2803 444672 : Buffer buf = buffers[i];
2804 444672 : BufferDesc *buf_hdr = GetBufferDescriptor(buf - 1);
2805 444672 : bool lock = false;
2806 :
2807 444672 : if (flags & EB_LOCK_FIRST && i == 0)
2808 295370 : lock = true;
2809 149302 : else if (flags & EB_LOCK_TARGET)
2810 : {
2811 : Assert(extend_upto != InvalidBlockNumber);
2812 88864 : if (first_block + i + 1 == extend_upto)
2813 87954 : lock = true;
2814 : }
2815 :
2816 444672 : if (lock)
2817 383324 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
2818 :
2819 444672 : TerminateBufferIO(buf_hdr, false, BM_VALID, true, false);
2820 : }
2821 :
2822 402362 : pgBufferUsage.shared_blks_written += extend_by;
2823 :
2824 402362 : *extended_by = extend_by;
2825 :
2826 402362 : return first_block;
2827 : }
2828 :
2829 : /*
2830 : * BufferIsLockedByMe
2831 : *
2832 : * Checks if this backend has the buffer locked in any mode.
2833 : *
2834 : * Buffer must be pinned.
2835 : */
2836 : bool
2837 0 : BufferIsLockedByMe(Buffer buffer)
2838 : {
2839 : BufferDesc *bufHdr;
2840 :
2841 : Assert(BufferIsPinned(buffer));
2842 :
2843 0 : if (BufferIsLocal(buffer))
2844 : {
2845 : /* Content locks are not maintained for local buffers. */
2846 0 : return true;
2847 : }
2848 : else
2849 : {
2850 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2851 0 : return LWLockHeldByMe(BufferDescriptorGetContentLock(bufHdr));
2852 : }
2853 : }
2854 :
2855 : /*
2856 : * BufferIsLockedByMeInMode
2857 : *
2858 : * Checks if this backend has the buffer locked in the specified mode.
2859 : *
2860 : * Buffer must be pinned.
2861 : */
2862 : bool
2863 0 : BufferIsLockedByMeInMode(Buffer buffer, int mode)
2864 : {
2865 : BufferDesc *bufHdr;
2866 :
2867 : Assert(BufferIsPinned(buffer));
2868 :
2869 0 : if (BufferIsLocal(buffer))
2870 : {
2871 : /* Content locks are not maintained for local buffers. */
2872 0 : return true;
2873 : }
2874 : else
2875 : {
2876 : LWLockMode lw_mode;
2877 :
2878 0 : switch (mode)
2879 : {
2880 0 : case BUFFER_LOCK_EXCLUSIVE:
2881 0 : lw_mode = LW_EXCLUSIVE;
2882 0 : break;
2883 0 : case BUFFER_LOCK_SHARE:
2884 0 : lw_mode = LW_SHARED;
2885 0 : break;
2886 0 : default:
2887 0 : pg_unreachable();
2888 : }
2889 :
2890 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2891 0 : return LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
2892 : lw_mode);
2893 : }
2894 : }
2895 :
2896 : /*
2897 : * BufferIsDirty
2898 : *
2899 : * Checks if buffer is already dirty.
2900 : *
2901 : * Buffer must be pinned and exclusive-locked. (Without an exclusive lock,
2902 : * the result may be stale before it's returned.)
2903 : */
2904 : bool
2905 0 : BufferIsDirty(Buffer buffer)
2906 : {
2907 : BufferDesc *bufHdr;
2908 :
2909 : Assert(BufferIsPinned(buffer));
2910 :
2911 0 : if (BufferIsLocal(buffer))
2912 : {
2913 0 : int bufid = -buffer - 1;
2914 :
2915 0 : bufHdr = GetLocalBufferDescriptor(bufid);
2916 : /* Content locks are not maintained for local buffers. */
2917 : }
2918 : else
2919 : {
2920 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2921 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
2922 : }
2923 :
2924 0 : return pg_atomic_read_u32(&bufHdr->state) & BM_DIRTY;
2925 : }
2926 :
2927 : /*
2928 : * MarkBufferDirty
2929 : *
2930 : * Marks buffer contents as dirty (actual write happens later).
2931 : *
2932 : * Buffer must be pinned and exclusive-locked. (If caller does not hold
2933 : * exclusive lock, then somebody could be in process of writing the buffer,
2934 : * leading to risk of bad data written to disk.)
2935 : */
2936 : void
2937 42952808 : MarkBufferDirty(Buffer buffer)
2938 : {
2939 : BufferDesc *bufHdr;
2940 : uint32 buf_state;
2941 : uint32 old_buf_state;
2942 :
2943 42952808 : if (!BufferIsValid(buffer))
2944 0 : elog(ERROR, "bad buffer ID: %d", buffer);
2945 :
2946 42952808 : if (BufferIsLocal(buffer))
2947 : {
2948 2444640 : MarkLocalBufferDirty(buffer);
2949 2444640 : return;
2950 : }
2951 :
2952 40508168 : bufHdr = GetBufferDescriptor(buffer - 1);
2953 :
2954 : Assert(BufferIsPinned(buffer));
2955 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
2956 :
2957 40508168 : old_buf_state = pg_atomic_read_u32(&bufHdr->state);
2958 : for (;;)
2959 : {
2960 40508438 : if (old_buf_state & BM_LOCKED)
2961 6 : old_buf_state = WaitBufHdrUnlocked(bufHdr);
2962 :
2963 40508438 : buf_state = old_buf_state;
2964 :
2965 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2966 40508438 : buf_state |= BM_DIRTY | BM_JUST_DIRTIED;
2967 :
2968 40508438 : if (pg_atomic_compare_exchange_u32(&bufHdr->state, &old_buf_state,
2969 : buf_state))
2970 40508168 : break;
2971 : }
2972 :
2973 : /*
2974 : * If the buffer was not dirty already, do vacuum accounting.
2975 : */
2976 40508168 : if (!(old_buf_state & BM_DIRTY))
2977 : {
2978 1301390 : pgBufferUsage.shared_blks_dirtied++;
2979 1301390 : if (VacuumCostActive)
2980 15436 : VacuumCostBalance += VacuumCostPageDirty;
2981 : }
2982 : }
2983 :
2984 : /*
2985 : * ReleaseAndReadBuffer -- combine ReleaseBuffer() and ReadBuffer()
2986 : *
2987 : * Formerly, this saved one cycle of acquiring/releasing the BufMgrLock
2988 : * compared to calling the two routines separately. Now it's mainly just
2989 : * a convenience function. However, if the passed buffer is valid and
2990 : * already contains the desired block, we just return it as-is; and that
2991 : * does save considerable work compared to a full release and reacquire.
2992 : *
2993 : * Note: it is OK to pass buffer == InvalidBuffer, indicating that no old
2994 : * buffer actually needs to be released. This case is the same as ReadBuffer,
2995 : * but can save some tests in the caller.
2996 : */
2997 : Buffer
2998 58767140 : ReleaseAndReadBuffer(Buffer buffer,
2999 : Relation relation,
3000 : BlockNumber blockNum)
3001 : {
3002 58767140 : ForkNumber forkNum = MAIN_FORKNUM;
3003 : BufferDesc *bufHdr;
3004 :
3005 58767140 : if (BufferIsValid(buffer))
3006 : {
3007 : Assert(BufferIsPinned(buffer));
3008 35223256 : if (BufferIsLocal(buffer))
3009 : {
3010 73728 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
3011 80772 : if (bufHdr->tag.blockNum == blockNum &&
3012 14088 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3013 7044 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3014 7044 : return buffer;
3015 66684 : UnpinLocalBuffer(buffer);
3016 : }
3017 : else
3018 : {
3019 35149528 : bufHdr = GetBufferDescriptor(buffer - 1);
3020 : /* we have pin, so it's ok to examine tag without spinlock */
3021 47391374 : if (bufHdr->tag.blockNum == blockNum &&
3022 24483692 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3023 12241846 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3024 12241846 : return buffer;
3025 22907682 : UnpinBuffer(bufHdr);
3026 : }
3027 : }
3028 :
3029 46518250 : return ReadBuffer(relation, blockNum);
3030 : }
3031 :
3032 : /*
3033 : * PinBuffer -- make buffer unavailable for replacement.
3034 : *
3035 : * For the default access strategy, the buffer's usage_count is incremented
3036 : * when we first pin it; for other strategies we just make sure the usage_count
3037 : * isn't zero. (The idea of the latter is that we don't want synchronized
3038 : * heap scans to inflate the count, but we need it to not be zero to discourage
3039 : * other backends from stealing buffers from our ring. As long as we cycle
3040 : * through the ring faster than the global clock-sweep cycles, buffers in
3041 : * our ring won't be chosen as victims for replacement by other backends.)
3042 : *
3043 : * This should be applied only to shared buffers, never local ones.
3044 : *
3045 : * Since buffers are pinned/unpinned very frequently, pin buffers without
3046 : * taking the buffer header lock; instead update the state variable in loop of
3047 : * CAS operations. Hopefully it's just a single CAS.
3048 : *
3049 : * Note that ResourceOwnerEnlarge() and ReservePrivateRefCountEntry()
3050 : * must have been done already.
3051 : *
3052 : * Returns true if buffer is BM_VALID, else false. This provision allows
3053 : * some callers to avoid an extra spinlock cycle. If skip_if_not_valid is
3054 : * true, then a false return value also indicates that the buffer was
3055 : * (recently) invalid and has not been pinned.
3056 : */
3057 : static bool
3058 119398978 : PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
3059 : bool skip_if_not_valid)
3060 : {
3061 119398978 : Buffer b = BufferDescriptorGetBuffer(buf);
3062 : bool result;
3063 : PrivateRefCountEntry *ref;
3064 :
3065 : Assert(!BufferIsLocal(b));
3066 : Assert(ReservedRefCountEntry != NULL);
3067 :
3068 119398978 : ref = GetPrivateRefCountEntry(b, true);
3069 :
3070 119398978 : if (ref == NULL)
3071 : {
3072 : uint32 buf_state;
3073 : uint32 old_buf_state;
3074 :
3075 115066720 : old_buf_state = pg_atomic_read_u32(&buf->state);
3076 : for (;;)
3077 : {
3078 115124170 : if (unlikely(skip_if_not_valid && !(old_buf_state & BM_VALID)))
3079 12 : return false;
3080 :
3081 115124158 : if (old_buf_state & BM_LOCKED)
3082 1604 : old_buf_state = WaitBufHdrUnlocked(buf);
3083 :
3084 115124158 : buf_state = old_buf_state;
3085 :
3086 : /* increase refcount */
3087 115124158 : buf_state += BUF_REFCOUNT_ONE;
3088 :
3089 115124158 : if (strategy == NULL)
3090 : {
3091 : /* Default case: increase usagecount unless already max. */
3092 113376356 : if (BUF_STATE_GET_USAGECOUNT(buf_state) < BM_MAX_USAGE_COUNT)
3093 6562358 : buf_state += BUF_USAGECOUNT_ONE;
3094 : }
3095 : else
3096 : {
3097 : /*
3098 : * Ring buffers shouldn't evict others from pool. Thus we
3099 : * don't make usagecount more than 1.
3100 : */
3101 1747802 : if (BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3102 72090 : buf_state += BUF_USAGECOUNT_ONE;
3103 : }
3104 :
3105 115124158 : if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
3106 : buf_state))
3107 : {
3108 115066708 : result = (buf_state & BM_VALID) != 0;
3109 :
3110 115066708 : TrackNewBufferPin(b);
3111 115066708 : break;
3112 : }
3113 : }
3114 : }
3115 : else
3116 : {
3117 : /*
3118 : * If we previously pinned the buffer, it is likely to be valid, but
3119 : * it may not be if StartReadBuffers() was called and
3120 : * WaitReadBuffers() hasn't been called yet. We'll check by loading
3121 : * the flags without locking. This is racy, but it's OK to return
3122 : * false spuriously: when WaitReadBuffers() calls StartBufferIO(),
3123 : * it'll see that it's now valid.
3124 : *
3125 : * Note: We deliberately avoid a Valgrind client request here.
3126 : * Individual access methods can optionally superimpose buffer page
3127 : * client requests on top of our client requests to enforce that
3128 : * buffers are only accessed while locked (and pinned). It's possible
3129 : * that the buffer page is legitimately non-accessible here. We
3130 : * cannot meddle with that.
3131 : */
3132 4332258 : result = (pg_atomic_read_u32(&buf->state) & BM_VALID) != 0;
3133 :
3134 : Assert(ref->refcount > 0);
3135 4332258 : ref->refcount++;
3136 4332258 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
3137 : }
3138 :
3139 119398966 : return result;
3140 : }
3141 :
3142 : /*
3143 : * PinBuffer_Locked -- as above, but caller already locked the buffer header.
3144 : * The spinlock is released before return.
3145 : *
3146 : * As this function is called with the spinlock held, the caller has to
3147 : * previously call ReservePrivateRefCountEntry() and
3148 : * ResourceOwnerEnlarge(CurrentResourceOwner);
3149 : *
3150 : * Currently, no callers of this function want to modify the buffer's
3151 : * usage_count at all, so there's no need for a strategy parameter.
3152 : * Also we don't bother with a BM_VALID test (the caller could check that for
3153 : * itself).
3154 : *
3155 : * Also all callers only ever use this function when it's known that the
3156 : * buffer can't have a preexisting pin by this backend. That allows us to skip
3157 : * searching the private refcount array & hash, which is a boon, because the
3158 : * spinlock is still held.
3159 : *
3160 : * Note: use of this routine is frequently mandatory, not just an optimization
3161 : * to save a spin lock/unlock cycle, because we need to pin a buffer before
3162 : * its state can change under us.
3163 : */
3164 : static void
3165 607322 : PinBuffer_Locked(BufferDesc *buf)
3166 : {
3167 : uint32 buf_state;
3168 :
3169 : /*
3170 : * As explained, We don't expect any preexisting pins. That allows us to
3171 : * manipulate the PrivateRefCount after releasing the spinlock
3172 : */
3173 : Assert(GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf), false) == NULL);
3174 :
3175 : /*
3176 : * Since we hold the buffer spinlock, we can update the buffer state and
3177 : * release the lock in one operation.
3178 : */
3179 607322 : buf_state = pg_atomic_read_u32(&buf->state);
3180 : Assert(buf_state & BM_LOCKED);
3181 607322 : buf_state += BUF_REFCOUNT_ONE;
3182 607322 : UnlockBufHdr(buf, buf_state);
3183 :
3184 607322 : TrackNewBufferPin(BufferDescriptorGetBuffer(buf));
3185 607322 : }
3186 :
3187 : /*
3188 : * Support for waking up another backend that is waiting for the cleanup lock
3189 : * to be released using BM_PIN_COUNT_WAITER.
3190 : *
3191 : * See LockBufferForCleanup().
3192 : *
3193 : * Expected to be called just after releasing a buffer pin (in a BufferDesc,
3194 : * not just reducing the backend-local pincount for the buffer).
3195 : */
3196 : static void
3197 138 : WakePinCountWaiter(BufferDesc *buf)
3198 : {
3199 : /*
3200 : * Acquire the buffer header lock, re-check that there's a waiter. Another
3201 : * backend could have unpinned this buffer, and already woken up the
3202 : * waiter.
3203 : *
3204 : * There's no danger of the buffer being replaced after we unpinned it
3205 : * above, as it's pinned by the waiter. The waiter removes
3206 : * BM_PIN_COUNT_WAITER if it stops waiting for a reason other than this
3207 : * backend waking it up.
3208 : */
3209 138 : uint32 buf_state = LockBufHdr(buf);
3210 :
3211 138 : if ((buf_state & BM_PIN_COUNT_WAITER) &&
3212 138 : BUF_STATE_GET_REFCOUNT(buf_state) == 1)
3213 138 : {
3214 : /* we just released the last pin other than the waiter's */
3215 138 : int wait_backend_pgprocno = buf->wait_backend_pgprocno;
3216 :
3217 138 : buf_state &= ~BM_PIN_COUNT_WAITER;
3218 138 : UnlockBufHdr(buf, buf_state);
3219 138 : ProcSendSignal(wait_backend_pgprocno);
3220 : }
3221 : else
3222 0 : UnlockBufHdr(buf, buf_state);
3223 138 : }
3224 :
3225 : /*
3226 : * UnpinBuffer -- make buffer available for replacement.
3227 : *
3228 : * This should be applied only to shared buffers, never local ones. This
3229 : * always adjusts CurrentResourceOwner.
3230 : */
3231 : static void
3232 146706944 : UnpinBuffer(BufferDesc *buf)
3233 : {
3234 146706944 : Buffer b = BufferDescriptorGetBuffer(buf);
3235 :
3236 146706944 : ResourceOwnerForgetBuffer(CurrentResourceOwner, b);
3237 146706944 : UnpinBufferNoOwner(buf);
3238 146706944 : }
3239 :
3240 : static void
3241 146716032 : UnpinBufferNoOwner(BufferDesc *buf)
3242 : {
3243 : PrivateRefCountEntry *ref;
3244 146716032 : Buffer b = BufferDescriptorGetBuffer(buf);
3245 :
3246 : Assert(!BufferIsLocal(b));
3247 :
3248 : /* not moving as we're likely deleting it soon anyway */
3249 146716032 : ref = GetPrivateRefCountEntry(b, false);
3250 : Assert(ref != NULL);
3251 : Assert(ref->refcount > 0);
3252 146716032 : ref->refcount--;
3253 146716032 : if (ref->refcount == 0)
3254 : {
3255 : uint32 buf_state;
3256 : uint32 old_buf_state;
3257 :
3258 : /*
3259 : * Mark buffer non-accessible to Valgrind.
3260 : *
3261 : * Note that the buffer may have already been marked non-accessible
3262 : * within access method code that enforces that buffers are only
3263 : * accessed while a buffer lock is held.
3264 : */
3265 : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
3266 :
3267 : /*
3268 : * I'd better not still hold the buffer content lock. Can't use
3269 : * BufferIsLockedByMe(), as that asserts the buffer is pinned.
3270 : */
3271 : Assert(!LWLockHeldByMe(BufferDescriptorGetContentLock(buf)));
3272 :
3273 : /*
3274 : * Decrement the shared reference count.
3275 : *
3276 : * Since buffer spinlock holder can update status using just write,
3277 : * it's not safe to use atomic decrement here; thus use a CAS loop.
3278 : */
3279 119489338 : old_buf_state = pg_atomic_read_u32(&buf->state);
3280 : for (;;)
3281 : {
3282 119540086 : if (old_buf_state & BM_LOCKED)
3283 1488 : old_buf_state = WaitBufHdrUnlocked(buf);
3284 :
3285 119540086 : buf_state = old_buf_state;
3286 :
3287 119540086 : buf_state -= BUF_REFCOUNT_ONE;
3288 :
3289 119540086 : if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
3290 : buf_state))
3291 119489338 : break;
3292 : }
3293 :
3294 : /* Support LockBufferForCleanup() */
3295 119489338 : if (buf_state & BM_PIN_COUNT_WAITER)
3296 138 : WakePinCountWaiter(buf);
3297 :
3298 119489338 : ForgetPrivateRefCountEntry(ref);
3299 : }
3300 146716032 : }
3301 :
3302 : /*
3303 : * Set up backend-local tracking of a buffer pinned the first time by this
3304 : * backend.
3305 : */
3306 : inline void
3307 119489338 : TrackNewBufferPin(Buffer buf)
3308 : {
3309 : PrivateRefCountEntry *ref;
3310 :
3311 119489338 : ref = NewPrivateRefCountEntry(buf);
3312 119489338 : ref->refcount++;
3313 :
3314 119489338 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buf);
3315 :
3316 : /*
3317 : * This is the first pin for this page by this backend, mark its page as
3318 : * defined to valgrind. While the page contents might not actually be
3319 : * valid yet, we don't currently guarantee that such pages are marked
3320 : * undefined or non-accessible.
3321 : *
3322 : * It's not necessarily the prettiest to do this here, but otherwise we'd
3323 : * need this block of code in multiple places.
3324 : */
3325 : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(GetBufferDescriptor(buf - 1)),
3326 : BLCKSZ);
3327 119489338 : }
3328 :
3329 : #define ST_SORT sort_checkpoint_bufferids
3330 : #define ST_ELEMENT_TYPE CkptSortItem
3331 : #define ST_COMPARE(a, b) ckpt_buforder_comparator(a, b)
3332 : #define ST_SCOPE static
3333 : #define ST_DEFINE
3334 : #include "lib/sort_template.h"
3335 :
3336 : /*
3337 : * BufferSync -- Write out all dirty buffers in the pool.
3338 : *
3339 : * This is called at checkpoint time to write out all dirty shared buffers.
3340 : * The checkpoint request flags should be passed in. If CHECKPOINT_FAST is
3341 : * set, we disable delays between writes; if CHECKPOINT_IS_SHUTDOWN,
3342 : * CHECKPOINT_END_OF_RECOVERY or CHECKPOINT_FLUSH_UNLOGGED is set, we write
3343 : * even unlogged buffers, which are otherwise skipped. The remaining flags
3344 : * currently have no effect here.
3345 : */
3346 : static void
3347 3454 : BufferSync(int flags)
3348 : {
3349 : uint32 buf_state;
3350 : int buf_id;
3351 : int num_to_scan;
3352 : int num_spaces;
3353 : int num_processed;
3354 : int num_written;
3355 3454 : CkptTsStatus *per_ts_stat = NULL;
3356 : Oid last_tsid;
3357 : binaryheap *ts_heap;
3358 : int i;
3359 3454 : uint32 mask = BM_DIRTY;
3360 : WritebackContext wb_context;
3361 :
3362 : /*
3363 : * Unless this is a shutdown checkpoint or we have been explicitly told,
3364 : * we write only permanent, dirty buffers. But at shutdown or end of
3365 : * recovery, we write all dirty buffers.
3366 : */
3367 3454 : if (!((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
3368 : CHECKPOINT_FLUSH_UNLOGGED))))
3369 1940 : mask |= BM_PERMANENT;
3370 :
3371 : /*
3372 : * Loop over all buffers, and mark the ones that need to be written with
3373 : * BM_CHECKPOINT_NEEDED. Count them as we go (num_to_scan), so that we
3374 : * can estimate how much work needs to be done.
3375 : *
3376 : * This allows us to write only those pages that were dirty when the
3377 : * checkpoint began, and not those that get dirtied while it proceeds.
3378 : * Whenever a page with BM_CHECKPOINT_NEEDED is written out, either by us
3379 : * later in this function, or by normal backends or the bgwriter cleaning
3380 : * scan, the flag is cleared. Any buffer dirtied after this point won't
3381 : * have the flag set.
3382 : *
3383 : * Note that if we fail to write some buffer, we may leave buffers with
3384 : * BM_CHECKPOINT_NEEDED still set. This is OK since any such buffer would
3385 : * certainly need to be written for the next checkpoint attempt, too.
3386 : */
3387 3454 : num_to_scan = 0;
3388 24081182 : for (buf_id = 0; buf_id < NBuffers; buf_id++)
3389 : {
3390 24077728 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3391 :
3392 : /*
3393 : * Header spinlock is enough to examine BM_DIRTY, see comment in
3394 : * SyncOneBuffer.
3395 : */
3396 24077728 : buf_state = LockBufHdr(bufHdr);
3397 :
3398 24077728 : if ((buf_state & mask) == mask)
3399 : {
3400 : CkptSortItem *item;
3401 :
3402 585468 : buf_state |= BM_CHECKPOINT_NEEDED;
3403 :
3404 585468 : item = &CkptBufferIds[num_to_scan++];
3405 585468 : item->buf_id = buf_id;
3406 585468 : item->tsId = bufHdr->tag.spcOid;
3407 585468 : item->relNumber = BufTagGetRelNumber(&bufHdr->tag);
3408 585468 : item->forkNum = BufTagGetForkNum(&bufHdr->tag);
3409 585468 : item->blockNum = bufHdr->tag.blockNum;
3410 : }
3411 :
3412 24077728 : UnlockBufHdr(bufHdr, buf_state);
3413 :
3414 : /* Check for barrier events in case NBuffers is large. */
3415 24077728 : if (ProcSignalBarrierPending)
3416 0 : ProcessProcSignalBarrier();
3417 : }
3418 :
3419 3454 : if (num_to_scan == 0)
3420 1332 : return; /* nothing to do */
3421 :
3422 2122 : WritebackContextInit(&wb_context, &checkpoint_flush_after);
3423 :
3424 : TRACE_POSTGRESQL_BUFFER_SYNC_START(NBuffers, num_to_scan);
3425 :
3426 : /*
3427 : * Sort buffers that need to be written to reduce the likelihood of random
3428 : * IO. The sorting is also important for the implementation of balancing
3429 : * writes between tablespaces. Without balancing writes we'd potentially
3430 : * end up writing to the tablespaces one-by-one; possibly overloading the
3431 : * underlying system.
3432 : */
3433 2122 : sort_checkpoint_bufferids(CkptBufferIds, num_to_scan);
3434 :
3435 2122 : num_spaces = 0;
3436 :
3437 : /*
3438 : * Allocate progress status for each tablespace with buffers that need to
3439 : * be flushed. This requires the to-be-flushed array to be sorted.
3440 : */
3441 2122 : last_tsid = InvalidOid;
3442 587590 : for (i = 0; i < num_to_scan; i++)
3443 : {
3444 : CkptTsStatus *s;
3445 : Oid cur_tsid;
3446 :
3447 585468 : cur_tsid = CkptBufferIds[i].tsId;
3448 :
3449 : /*
3450 : * Grow array of per-tablespace status structs, every time a new
3451 : * tablespace is found.
3452 : */
3453 585468 : if (last_tsid == InvalidOid || last_tsid != cur_tsid)
3454 3212 : {
3455 : Size sz;
3456 :
3457 3212 : num_spaces++;
3458 :
3459 : /*
3460 : * Not worth adding grow-by-power-of-2 logic here - even with a
3461 : * few hundred tablespaces this should be fine.
3462 : */
3463 3212 : sz = sizeof(CkptTsStatus) * num_spaces;
3464 :
3465 3212 : if (per_ts_stat == NULL)
3466 2122 : per_ts_stat = (CkptTsStatus *) palloc(sz);
3467 : else
3468 1090 : per_ts_stat = (CkptTsStatus *) repalloc(per_ts_stat, sz);
3469 :
3470 3212 : s = &per_ts_stat[num_spaces - 1];
3471 3212 : memset(s, 0, sizeof(*s));
3472 3212 : s->tsId = cur_tsid;
3473 :
3474 : /*
3475 : * The first buffer in this tablespace. As CkptBufferIds is sorted
3476 : * by tablespace all (s->num_to_scan) buffers in this tablespace
3477 : * will follow afterwards.
3478 : */
3479 3212 : s->index = i;
3480 :
3481 : /*
3482 : * progress_slice will be determined once we know how many buffers
3483 : * are in each tablespace, i.e. after this loop.
3484 : */
3485 :
3486 3212 : last_tsid = cur_tsid;
3487 : }
3488 : else
3489 : {
3490 582256 : s = &per_ts_stat[num_spaces - 1];
3491 : }
3492 :
3493 585468 : s->num_to_scan++;
3494 :
3495 : /* Check for barrier events. */
3496 585468 : if (ProcSignalBarrierPending)
3497 0 : ProcessProcSignalBarrier();
3498 : }
3499 :
3500 : Assert(num_spaces > 0);
3501 :
3502 : /*
3503 : * Build a min-heap over the write-progress in the individual tablespaces,
3504 : * and compute how large a portion of the total progress a single
3505 : * processed buffer is.
3506 : */
3507 2122 : ts_heap = binaryheap_allocate(num_spaces,
3508 : ts_ckpt_progress_comparator,
3509 : NULL);
3510 :
3511 5334 : for (i = 0; i < num_spaces; i++)
3512 : {
3513 3212 : CkptTsStatus *ts_stat = &per_ts_stat[i];
3514 :
3515 3212 : ts_stat->progress_slice = (float8) num_to_scan / ts_stat->num_to_scan;
3516 :
3517 3212 : binaryheap_add_unordered(ts_heap, PointerGetDatum(ts_stat));
3518 : }
3519 :
3520 2122 : binaryheap_build(ts_heap);
3521 :
3522 : /*
3523 : * Iterate through to-be-checkpointed buffers and write the ones (still)
3524 : * marked with BM_CHECKPOINT_NEEDED. The writes are balanced between
3525 : * tablespaces; otherwise the sorting would lead to only one tablespace
3526 : * receiving writes at a time, making inefficient use of the hardware.
3527 : */
3528 2122 : num_processed = 0;
3529 2122 : num_written = 0;
3530 587590 : while (!binaryheap_empty(ts_heap))
3531 : {
3532 585468 : BufferDesc *bufHdr = NULL;
3533 : CkptTsStatus *ts_stat = (CkptTsStatus *)
3534 585468 : DatumGetPointer(binaryheap_first(ts_heap));
3535 :
3536 585468 : buf_id = CkptBufferIds[ts_stat->index].buf_id;
3537 : Assert(buf_id != -1);
3538 :
3539 585468 : bufHdr = GetBufferDescriptor(buf_id);
3540 :
3541 585468 : num_processed++;
3542 :
3543 : /*
3544 : * We don't need to acquire the lock here, because we're only looking
3545 : * at a single bit. It's possible that someone else writes the buffer
3546 : * and clears the flag right after we check, but that doesn't matter
3547 : * since SyncOneBuffer will then do nothing. However, there is a
3548 : * further race condition: it's conceivable that between the time we
3549 : * examine the bit here and the time SyncOneBuffer acquires the lock,
3550 : * someone else not only wrote the buffer but replaced it with another
3551 : * page and dirtied it. In that improbable case, SyncOneBuffer will
3552 : * write the buffer though we didn't need to. It doesn't seem worth
3553 : * guarding against this, though.
3554 : */
3555 585468 : if (pg_atomic_read_u32(&bufHdr->state) & BM_CHECKPOINT_NEEDED)
3556 : {
3557 547384 : if (SyncOneBuffer(buf_id, false, &wb_context) & BUF_WRITTEN)
3558 : {
3559 : TRACE_POSTGRESQL_BUFFER_SYNC_WRITTEN(buf_id);
3560 547384 : PendingCheckpointerStats.buffers_written++;
3561 547384 : num_written++;
3562 : }
3563 : }
3564 :
3565 : /*
3566 : * Measure progress independent of actually having to flush the buffer
3567 : * - otherwise writing become unbalanced.
3568 : */
3569 585468 : ts_stat->progress += ts_stat->progress_slice;
3570 585468 : ts_stat->num_scanned++;
3571 585468 : ts_stat->index++;
3572 :
3573 : /* Have all the buffers from the tablespace been processed? */
3574 585468 : if (ts_stat->num_scanned == ts_stat->num_to_scan)
3575 : {
3576 3212 : binaryheap_remove_first(ts_heap);
3577 : }
3578 : else
3579 : {
3580 : /* update heap with the new progress */
3581 582256 : binaryheap_replace_first(ts_heap, PointerGetDatum(ts_stat));
3582 : }
3583 :
3584 : /*
3585 : * Sleep to throttle our I/O rate.
3586 : *
3587 : * (This will check for barrier events even if it doesn't sleep.)
3588 : */
3589 585468 : CheckpointWriteDelay(flags, (double) num_processed / num_to_scan);
3590 : }
3591 :
3592 : /*
3593 : * Issue all pending flushes. Only checkpointer calls BufferSync(), so
3594 : * IOContext will always be IOCONTEXT_NORMAL.
3595 : */
3596 2122 : IssuePendingWritebacks(&wb_context, IOCONTEXT_NORMAL);
3597 :
3598 2122 : pfree(per_ts_stat);
3599 2122 : per_ts_stat = NULL;
3600 2122 : binaryheap_free(ts_heap);
3601 :
3602 : /*
3603 : * Update checkpoint statistics. As noted above, this doesn't include
3604 : * buffers written by other backends or bgwriter scan.
3605 : */
3606 2122 : CheckpointStats.ckpt_bufs_written += num_written;
3607 :
3608 : TRACE_POSTGRESQL_BUFFER_SYNC_DONE(NBuffers, num_written, num_to_scan);
3609 : }
3610 :
3611 : /*
3612 : * BgBufferSync -- Write out some dirty buffers in the pool.
3613 : *
3614 : * This is called periodically by the background writer process.
3615 : *
3616 : * Returns true if it's appropriate for the bgwriter process to go into
3617 : * low-power hibernation mode. (This happens if the strategy clock-sweep
3618 : * has been "lapped" and no buffer allocations have occurred recently,
3619 : * or if the bgwriter has been effectively disabled by setting
3620 : * bgwriter_lru_maxpages to 0.)
3621 : */
3622 : bool
3623 24770 : BgBufferSync(WritebackContext *wb_context)
3624 : {
3625 : /* info obtained from freelist.c */
3626 : int strategy_buf_id;
3627 : uint32 strategy_passes;
3628 : uint32 recent_alloc;
3629 :
3630 : /*
3631 : * Information saved between calls so we can determine the strategy
3632 : * point's advance rate and avoid scanning already-cleaned buffers.
3633 : */
3634 : static bool saved_info_valid = false;
3635 : static int prev_strategy_buf_id;
3636 : static uint32 prev_strategy_passes;
3637 : static int next_to_clean;
3638 : static uint32 next_passes;
3639 :
3640 : /* Moving averages of allocation rate and clean-buffer density */
3641 : static float smoothed_alloc = 0;
3642 : static float smoothed_density = 10.0;
3643 :
3644 : /* Potentially these could be tunables, but for now, not */
3645 24770 : float smoothing_samples = 16;
3646 24770 : float scan_whole_pool_milliseconds = 120000.0;
3647 :
3648 : /* Used to compute how far we scan ahead */
3649 : long strategy_delta;
3650 : int bufs_to_lap;
3651 : int bufs_ahead;
3652 : float scans_per_alloc;
3653 : int reusable_buffers_est;
3654 : int upcoming_alloc_est;
3655 : int min_scan_buffers;
3656 :
3657 : /* Variables for the scanning loop proper */
3658 : int num_to_scan;
3659 : int num_written;
3660 : int reusable_buffers;
3661 :
3662 : /* Variables for final smoothed_density update */
3663 : long new_strategy_delta;
3664 : uint32 new_recent_alloc;
3665 :
3666 : /*
3667 : * Find out where the clock-sweep currently is, and how many buffer
3668 : * allocations have happened since our last call.
3669 : */
3670 24770 : strategy_buf_id = StrategySyncStart(&strategy_passes, &recent_alloc);
3671 :
3672 : /* Report buffer alloc counts to pgstat */
3673 24770 : PendingBgWriterStats.buf_alloc += recent_alloc;
3674 :
3675 : /*
3676 : * If we're not running the LRU scan, just stop after doing the stats
3677 : * stuff. We mark the saved state invalid so that we can recover sanely
3678 : * if LRU scan is turned back on later.
3679 : */
3680 24770 : if (bgwriter_lru_maxpages <= 0)
3681 : {
3682 86 : saved_info_valid = false;
3683 86 : return true;
3684 : }
3685 :
3686 : /*
3687 : * Compute strategy_delta = how many buffers have been scanned by the
3688 : * clock-sweep since last time. If first time through, assume none. Then
3689 : * see if we are still ahead of the clock-sweep, and if so, how many
3690 : * buffers we could scan before we'd catch up with it and "lap" it. Note:
3691 : * weird-looking coding of xxx_passes comparisons are to avoid bogus
3692 : * behavior when the passes counts wrap around.
3693 : */
3694 24684 : if (saved_info_valid)
3695 : {
3696 23616 : int32 passes_delta = strategy_passes - prev_strategy_passes;
3697 :
3698 23616 : strategy_delta = strategy_buf_id - prev_strategy_buf_id;
3699 23616 : strategy_delta += (long) passes_delta * NBuffers;
3700 :
3701 : Assert(strategy_delta >= 0);
3702 :
3703 23616 : if ((int32) (next_passes - strategy_passes) > 0)
3704 : {
3705 : /* we're one pass ahead of the strategy point */
3706 4662 : bufs_to_lap = strategy_buf_id - next_to_clean;
3707 : #ifdef BGW_DEBUG
3708 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3709 : next_passes, next_to_clean,
3710 : strategy_passes, strategy_buf_id,
3711 : strategy_delta, bufs_to_lap);
3712 : #endif
3713 : }
3714 18954 : else if (next_passes == strategy_passes &&
3715 14594 : next_to_clean >= strategy_buf_id)
3716 : {
3717 : /* on same pass, but ahead or at least not behind */
3718 13224 : bufs_to_lap = NBuffers - (next_to_clean - strategy_buf_id);
3719 : #ifdef BGW_DEBUG
3720 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3721 : next_passes, next_to_clean,
3722 : strategy_passes, strategy_buf_id,
3723 : strategy_delta, bufs_to_lap);
3724 : #endif
3725 : }
3726 : else
3727 : {
3728 : /*
3729 : * We're behind, so skip forward to the strategy point and start
3730 : * cleaning from there.
3731 : */
3732 : #ifdef BGW_DEBUG
3733 : elog(DEBUG2, "bgwriter behind: bgw %u-%u strategy %u-%u delta=%ld",
3734 : next_passes, next_to_clean,
3735 : strategy_passes, strategy_buf_id,
3736 : strategy_delta);
3737 : #endif
3738 5730 : next_to_clean = strategy_buf_id;
3739 5730 : next_passes = strategy_passes;
3740 5730 : bufs_to_lap = NBuffers;
3741 : }
3742 : }
3743 : else
3744 : {
3745 : /*
3746 : * Initializing at startup or after LRU scanning had been off. Always
3747 : * start at the strategy point.
3748 : */
3749 : #ifdef BGW_DEBUG
3750 : elog(DEBUG2, "bgwriter initializing: strategy %u-%u",
3751 : strategy_passes, strategy_buf_id);
3752 : #endif
3753 1068 : strategy_delta = 0;
3754 1068 : next_to_clean = strategy_buf_id;
3755 1068 : next_passes = strategy_passes;
3756 1068 : bufs_to_lap = NBuffers;
3757 : }
3758 :
3759 : /* Update saved info for next time */
3760 24684 : prev_strategy_buf_id = strategy_buf_id;
3761 24684 : prev_strategy_passes = strategy_passes;
3762 24684 : saved_info_valid = true;
3763 :
3764 : /*
3765 : * Compute how many buffers had to be scanned for each new allocation, ie,
3766 : * 1/density of reusable buffers, and track a moving average of that.
3767 : *
3768 : * If the strategy point didn't move, we don't update the density estimate
3769 : */
3770 24684 : if (strategy_delta > 0 && recent_alloc > 0)
3771 : {
3772 13136 : scans_per_alloc = (float) strategy_delta / (float) recent_alloc;
3773 13136 : smoothed_density += (scans_per_alloc - smoothed_density) /
3774 : smoothing_samples;
3775 : }
3776 :
3777 : /*
3778 : * Estimate how many reusable buffers there are between the current
3779 : * strategy point and where we've scanned ahead to, based on the smoothed
3780 : * density estimate.
3781 : */
3782 24684 : bufs_ahead = NBuffers - bufs_to_lap;
3783 24684 : reusable_buffers_est = (float) bufs_ahead / smoothed_density;
3784 :
3785 : /*
3786 : * Track a moving average of recent buffer allocations. Here, rather than
3787 : * a true average we want a fast-attack, slow-decline behavior: we
3788 : * immediately follow any increase.
3789 : */
3790 24684 : if (smoothed_alloc <= (float) recent_alloc)
3791 6854 : smoothed_alloc = recent_alloc;
3792 : else
3793 17830 : smoothed_alloc += ((float) recent_alloc - smoothed_alloc) /
3794 : smoothing_samples;
3795 :
3796 : /* Scale the estimate by a GUC to allow more aggressive tuning. */
3797 24684 : upcoming_alloc_est = (int) (smoothed_alloc * bgwriter_lru_multiplier);
3798 :
3799 : /*
3800 : * If recent_alloc remains at zero for many cycles, smoothed_alloc will
3801 : * eventually underflow to zero, and the underflows produce annoying
3802 : * kernel warnings on some platforms. Once upcoming_alloc_est has gone to
3803 : * zero, there's no point in tracking smaller and smaller values of
3804 : * smoothed_alloc, so just reset it to exactly zero to avoid this
3805 : * syndrome. It will pop back up as soon as recent_alloc increases.
3806 : */
3807 24684 : if (upcoming_alloc_est == 0)
3808 4154 : smoothed_alloc = 0;
3809 :
3810 : /*
3811 : * Even in cases where there's been little or no buffer allocation
3812 : * activity, we want to make a small amount of progress through the buffer
3813 : * cache so that as many reusable buffers as possible are clean after an
3814 : * idle period.
3815 : *
3816 : * (scan_whole_pool_milliseconds / BgWriterDelay) computes how many times
3817 : * the BGW will be called during the scan_whole_pool time; slice the
3818 : * buffer pool into that many sections.
3819 : */
3820 24684 : min_scan_buffers = (int) (NBuffers / (scan_whole_pool_milliseconds / BgWriterDelay));
3821 :
3822 24684 : if (upcoming_alloc_est < (min_scan_buffers + reusable_buffers_est))
3823 : {
3824 : #ifdef BGW_DEBUG
3825 : elog(DEBUG2, "bgwriter: alloc_est=%d too small, using min=%d + reusable_est=%d",
3826 : upcoming_alloc_est, min_scan_buffers, reusable_buffers_est);
3827 : #endif
3828 11356 : upcoming_alloc_est = min_scan_buffers + reusable_buffers_est;
3829 : }
3830 :
3831 : /*
3832 : * Now write out dirty reusable buffers, working forward from the
3833 : * next_to_clean point, until we have lapped the strategy scan, or cleaned
3834 : * enough buffers to match our estimate of the next cycle's allocation
3835 : * requirements, or hit the bgwriter_lru_maxpages limit.
3836 : */
3837 :
3838 24684 : num_to_scan = bufs_to_lap;
3839 24684 : num_written = 0;
3840 24684 : reusable_buffers = reusable_buffers_est;
3841 :
3842 : /* Execute the LRU scan */
3843 3489828 : while (num_to_scan > 0 && reusable_buffers < upcoming_alloc_est)
3844 : {
3845 3465146 : int sync_state = SyncOneBuffer(next_to_clean, true,
3846 : wb_context);
3847 :
3848 3465146 : if (++next_to_clean >= NBuffers)
3849 : {
3850 5382 : next_to_clean = 0;
3851 5382 : next_passes++;
3852 : }
3853 3465146 : num_to_scan--;
3854 :
3855 3465146 : if (sync_state & BUF_WRITTEN)
3856 : {
3857 47350 : reusable_buffers++;
3858 47350 : if (++num_written >= bgwriter_lru_maxpages)
3859 : {
3860 2 : PendingBgWriterStats.maxwritten_clean++;
3861 2 : break;
3862 : }
3863 : }
3864 3417796 : else if (sync_state & BUF_REUSABLE)
3865 2694196 : reusable_buffers++;
3866 : }
3867 :
3868 24684 : PendingBgWriterStats.buf_written_clean += num_written;
3869 :
3870 : #ifdef BGW_DEBUG
3871 : 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",
3872 : recent_alloc, smoothed_alloc, strategy_delta, bufs_ahead,
3873 : smoothed_density, reusable_buffers_est, upcoming_alloc_est,
3874 : bufs_to_lap - num_to_scan,
3875 : num_written,
3876 : reusable_buffers - reusable_buffers_est);
3877 : #endif
3878 :
3879 : /*
3880 : * Consider the above scan as being like a new allocation scan.
3881 : * Characterize its density and update the smoothed one based on it. This
3882 : * effectively halves the moving average period in cases where both the
3883 : * strategy and the background writer are doing some useful scanning,
3884 : * which is helpful because a long memory isn't as desirable on the
3885 : * density estimates.
3886 : */
3887 24684 : new_strategy_delta = bufs_to_lap - num_to_scan;
3888 24684 : new_recent_alloc = reusable_buffers - reusable_buffers_est;
3889 24684 : if (new_strategy_delta > 0 && new_recent_alloc > 0)
3890 : {
3891 20212 : scans_per_alloc = (float) new_strategy_delta / (float) new_recent_alloc;
3892 20212 : smoothed_density += (scans_per_alloc - smoothed_density) /
3893 : smoothing_samples;
3894 :
3895 : #ifdef BGW_DEBUG
3896 : elog(DEBUG2, "bgwriter: cleaner density alloc=%u scan=%ld density=%.2f new smoothed=%.2f",
3897 : new_recent_alloc, new_strategy_delta,
3898 : scans_per_alloc, smoothed_density);
3899 : #endif
3900 : }
3901 :
3902 : /* Return true if OK to hibernate */
3903 24684 : return (bufs_to_lap == 0 && recent_alloc == 0);
3904 : }
3905 :
3906 : /*
3907 : * SyncOneBuffer -- process a single buffer during syncing.
3908 : *
3909 : * If skip_recently_used is true, we don't write currently-pinned buffers, nor
3910 : * buffers marked recently used, as these are not replacement candidates.
3911 : *
3912 : * Returns a bitmask containing the following flag bits:
3913 : * BUF_WRITTEN: we wrote the buffer.
3914 : * BUF_REUSABLE: buffer is available for replacement, ie, it has
3915 : * pin count 0 and usage count 0.
3916 : *
3917 : * (BUF_WRITTEN could be set in error if FlushBuffer finds the buffer clean
3918 : * after locking it, but we don't care all that much.)
3919 : */
3920 : static int
3921 4012530 : SyncOneBuffer(int buf_id, bool skip_recently_used, WritebackContext *wb_context)
3922 : {
3923 4012530 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3924 4012530 : int result = 0;
3925 : uint32 buf_state;
3926 : BufferTag tag;
3927 :
3928 : /* Make sure we can handle the pin */
3929 4012530 : ReservePrivateRefCountEntry();
3930 4012530 : ResourceOwnerEnlarge(CurrentResourceOwner);
3931 :
3932 : /*
3933 : * Check whether buffer needs writing.
3934 : *
3935 : * We can make this check without taking the buffer content lock so long
3936 : * as we mark pages dirty in access methods *before* logging changes with
3937 : * XLogInsert(): if someone marks the buffer dirty just after our check we
3938 : * don't worry because our checkpoint.redo points before log record for
3939 : * upcoming changes and so we are not required to write such dirty buffer.
3940 : */
3941 4012530 : buf_state = LockBufHdr(bufHdr);
3942 :
3943 4012530 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 0 &&
3944 4002888 : BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3945 : {
3946 2745378 : result |= BUF_REUSABLE;
3947 : }
3948 1267152 : else if (skip_recently_used)
3949 : {
3950 : /* Caller told us not to write recently-used buffers */
3951 723600 : UnlockBufHdr(bufHdr, buf_state);
3952 723600 : return result;
3953 : }
3954 :
3955 3288930 : if (!(buf_state & BM_VALID) || !(buf_state & BM_DIRTY))
3956 : {
3957 : /* It's clean, so nothing to do */
3958 2694196 : UnlockBufHdr(bufHdr, buf_state);
3959 2694196 : return result;
3960 : }
3961 :
3962 : /*
3963 : * Pin it, share-lock it, write it. (FlushBuffer will do nothing if the
3964 : * buffer is clean by the time we've locked it.)
3965 : */
3966 594734 : PinBuffer_Locked(bufHdr);
3967 :
3968 594734 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
3969 :
3970 594734 : tag = bufHdr->tag;
3971 :
3972 594734 : UnpinBuffer(bufHdr);
3973 :
3974 : /*
3975 : * SyncOneBuffer() is only called by checkpointer and bgwriter, so
3976 : * IOContext will always be IOCONTEXT_NORMAL.
3977 : */
3978 594734 : ScheduleBufferTagForWriteback(wb_context, IOCONTEXT_NORMAL, &tag);
3979 :
3980 594734 : return result | BUF_WRITTEN;
3981 : }
3982 :
3983 : /*
3984 : * AtEOXact_Buffers - clean up at end of transaction.
3985 : *
3986 : * As of PostgreSQL 8.0, buffer pins should get released by the
3987 : * ResourceOwner mechanism. This routine is just a debugging
3988 : * cross-check that no pins remain.
3989 : */
3990 : void
3991 1166350 : AtEOXact_Buffers(bool isCommit)
3992 : {
3993 1166350 : CheckForBufferLeaks();
3994 :
3995 1166350 : AtEOXact_LocalBuffers(isCommit);
3996 :
3997 : Assert(PrivateRefCountOverflowed == 0);
3998 1166350 : }
3999 :
4000 : /*
4001 : * Initialize access to shared buffer pool
4002 : *
4003 : * This is called during backend startup (whether standalone or under the
4004 : * postmaster). It sets up for this backend's access to the already-existing
4005 : * buffer pool.
4006 : */
4007 : void
4008 45062 : InitBufferManagerAccess(void)
4009 : {
4010 : HASHCTL hash_ctl;
4011 :
4012 : /*
4013 : * An advisory limit on the number of pins each backend should hold, based
4014 : * on shared_buffers and the maximum number of connections possible.
4015 : * That's very pessimistic, but outside toy-sized shared_buffers it should
4016 : * allow plenty of pins. LimitAdditionalPins() and
4017 : * GetAdditionalPinLimit() can be used to check the remaining balance.
4018 : */
4019 45062 : MaxProportionalPins = NBuffers / (MaxBackends + NUM_AUXILIARY_PROCS);
4020 :
4021 45062 : memset(&PrivateRefCountArray, 0, sizeof(PrivateRefCountArray));
4022 :
4023 45062 : hash_ctl.keysize = sizeof(int32);
4024 45062 : hash_ctl.entrysize = sizeof(PrivateRefCountEntry);
4025 :
4026 45062 : PrivateRefCountHash = hash_create("PrivateRefCount", 100, &hash_ctl,
4027 : HASH_ELEM | HASH_BLOBS);
4028 :
4029 : /*
4030 : * AtProcExit_Buffers needs LWLock access, and thereby has to be called at
4031 : * the corresponding phase of backend shutdown.
4032 : */
4033 : Assert(MyProc != NULL);
4034 45062 : on_shmem_exit(AtProcExit_Buffers, 0);
4035 45062 : }
4036 :
4037 : /*
4038 : * During backend exit, ensure that we released all shared-buffer locks and
4039 : * assert that we have no remaining pins.
4040 : */
4041 : static void
4042 45062 : AtProcExit_Buffers(int code, Datum arg)
4043 : {
4044 45062 : UnlockBuffers();
4045 :
4046 45062 : CheckForBufferLeaks();
4047 :
4048 : /* localbuf.c needs a chance too */
4049 45062 : AtProcExit_LocalBuffers();
4050 45062 : }
4051 :
4052 : /*
4053 : * CheckForBufferLeaks - ensure this backend holds no buffer pins
4054 : *
4055 : * As of PostgreSQL 8.0, buffer pins should get released by the
4056 : * ResourceOwner mechanism. This routine is just a debugging
4057 : * cross-check that no pins remain.
4058 : */
4059 : static void
4060 1211412 : CheckForBufferLeaks(void)
4061 : {
4062 : #ifdef USE_ASSERT_CHECKING
4063 : int RefCountErrors = 0;
4064 : PrivateRefCountEntry *res;
4065 : int i;
4066 : char *s;
4067 :
4068 : /* check the array */
4069 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4070 : {
4071 : res = &PrivateRefCountArray[i];
4072 :
4073 : if (res->buffer != InvalidBuffer)
4074 : {
4075 : s = DebugPrintBufferRefcount(res->buffer);
4076 : elog(WARNING, "buffer refcount leak: %s", s);
4077 : pfree(s);
4078 :
4079 : RefCountErrors++;
4080 : }
4081 : }
4082 :
4083 : /* if necessary search the hash */
4084 : if (PrivateRefCountOverflowed)
4085 : {
4086 : HASH_SEQ_STATUS hstat;
4087 :
4088 : hash_seq_init(&hstat, PrivateRefCountHash);
4089 : while ((res = (PrivateRefCountEntry *) hash_seq_search(&hstat)) != NULL)
4090 : {
4091 : s = DebugPrintBufferRefcount(res->buffer);
4092 : elog(WARNING, "buffer refcount leak: %s", s);
4093 : pfree(s);
4094 : RefCountErrors++;
4095 : }
4096 : }
4097 :
4098 : Assert(RefCountErrors == 0);
4099 : #endif
4100 1211412 : }
4101 :
4102 : #ifdef USE_ASSERT_CHECKING
4103 : /*
4104 : * Check for exclusive-locked catalog buffers. This is the core of
4105 : * AssertCouldGetRelation().
4106 : *
4107 : * A backend would self-deadlock on LWLocks if the catalog scan read the
4108 : * exclusive-locked buffer. The main threat is exclusive-locked buffers of
4109 : * catalogs used in relcache, because a catcache search on any catalog may
4110 : * build that catalog's relcache entry. We don't have an inventory of
4111 : * catalogs relcache uses, so just check buffers of most catalogs.
4112 : *
4113 : * It's better to minimize waits while holding an exclusive buffer lock, so it
4114 : * would be nice to broaden this check not to be catalog-specific. However,
4115 : * bttextcmp() accesses pg_collation, and non-core opclasses might similarly
4116 : * read tables. That is deadlock-free as long as there's no loop in the
4117 : * dependency graph: modifying table A may cause an opclass to read table B,
4118 : * but it must not cause a read of table A.
4119 : */
4120 : void
4121 : AssertBufferLocksPermitCatalogRead(void)
4122 : {
4123 : ForEachLWLockHeldByMe(AssertNotCatalogBufferLock, NULL);
4124 : }
4125 :
4126 : static void
4127 : AssertNotCatalogBufferLock(LWLock *lock, LWLockMode mode,
4128 : void *unused_context)
4129 : {
4130 : BufferDesc *bufHdr;
4131 : BufferTag tag;
4132 : Oid relid;
4133 :
4134 : if (mode != LW_EXCLUSIVE)
4135 : return;
4136 :
4137 : if (!((BufferDescPadded *) lock > BufferDescriptors &&
4138 : (BufferDescPadded *) lock < BufferDescriptors + NBuffers))
4139 : return; /* not a buffer lock */
4140 :
4141 : bufHdr = (BufferDesc *)
4142 : ((char *) lock - offsetof(BufferDesc, content_lock));
4143 : tag = bufHdr->tag;
4144 :
4145 : /*
4146 : * This relNumber==relid assumption holds until a catalog experiences
4147 : * VACUUM FULL or similar. After a command like that, relNumber will be
4148 : * in the normal (non-catalog) range, and we lose the ability to detect
4149 : * hazardous access to that catalog. Calling RelidByRelfilenumber() would
4150 : * close that gap, but RelidByRelfilenumber() might then deadlock with a
4151 : * held lock.
4152 : */
4153 : relid = tag.relNumber;
4154 :
4155 : if (IsCatalogTextUniqueIndexOid(relid)) /* see comments at the callee */
4156 : return;
4157 :
4158 : Assert(!IsCatalogRelationOid(relid));
4159 : }
4160 : #endif
4161 :
4162 :
4163 : /*
4164 : * Helper routine to issue warnings when a buffer is unexpectedly pinned
4165 : */
4166 : char *
4167 80 : DebugPrintBufferRefcount(Buffer buffer)
4168 : {
4169 : BufferDesc *buf;
4170 : int32 loccount;
4171 : char *result;
4172 : ProcNumber backend;
4173 : uint32 buf_state;
4174 :
4175 : Assert(BufferIsValid(buffer));
4176 80 : if (BufferIsLocal(buffer))
4177 : {
4178 32 : buf = GetLocalBufferDescriptor(-buffer - 1);
4179 32 : loccount = LocalRefCount[-buffer - 1];
4180 32 : backend = MyProcNumber;
4181 : }
4182 : else
4183 : {
4184 48 : buf = GetBufferDescriptor(buffer - 1);
4185 48 : loccount = GetPrivateRefCount(buffer);
4186 48 : backend = INVALID_PROC_NUMBER;
4187 : }
4188 :
4189 : /* theoretically we should lock the bufhdr here */
4190 80 : buf_state = pg_atomic_read_u32(&buf->state);
4191 :
4192 80 : result = psprintf("[%03d] (rel=%s, blockNum=%u, flags=0x%x, refcount=%u %d)",
4193 : buffer,
4194 80 : relpathbackend(BufTagGetRelFileLocator(&buf->tag), backend,
4195 : BufTagGetForkNum(&buf->tag)).str,
4196 : buf->tag.blockNum, buf_state & BUF_FLAG_MASK,
4197 : BUF_STATE_GET_REFCOUNT(buf_state), loccount);
4198 80 : return result;
4199 : }
4200 :
4201 : /*
4202 : * CheckPointBuffers
4203 : *
4204 : * Flush all dirty blocks in buffer pool to disk at checkpoint time.
4205 : *
4206 : * Note: temporary relations do not participate in checkpoints, so they don't
4207 : * need to be flushed.
4208 : */
4209 : void
4210 3454 : CheckPointBuffers(int flags)
4211 : {
4212 3454 : BufferSync(flags);
4213 3454 : }
4214 :
4215 : /*
4216 : * BufferGetBlockNumber
4217 : * Returns the block number associated with a buffer.
4218 : *
4219 : * Note:
4220 : * Assumes that the buffer is valid and pinned, else the
4221 : * value may be obsolete immediately...
4222 : */
4223 : BlockNumber
4224 101756100 : BufferGetBlockNumber(Buffer buffer)
4225 : {
4226 : BufferDesc *bufHdr;
4227 :
4228 : Assert(BufferIsPinned(buffer));
4229 :
4230 101756100 : if (BufferIsLocal(buffer))
4231 3808938 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4232 : else
4233 97947162 : bufHdr = GetBufferDescriptor(buffer - 1);
4234 :
4235 : /* pinned, so OK to read tag without spinlock */
4236 101756100 : return bufHdr->tag.blockNum;
4237 : }
4238 :
4239 : /*
4240 : * BufferGetTag
4241 : * Returns the relfilelocator, fork number and block number associated with
4242 : * a buffer.
4243 : */
4244 : void
4245 30089786 : BufferGetTag(Buffer buffer, RelFileLocator *rlocator, ForkNumber *forknum,
4246 : BlockNumber *blknum)
4247 : {
4248 : BufferDesc *bufHdr;
4249 :
4250 : /* Do the same checks as BufferGetBlockNumber. */
4251 : Assert(BufferIsPinned(buffer));
4252 :
4253 30089786 : if (BufferIsLocal(buffer))
4254 0 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4255 : else
4256 30089786 : bufHdr = GetBufferDescriptor(buffer - 1);
4257 :
4258 : /* pinned, so OK to read tag without spinlock */
4259 30089786 : *rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
4260 30089786 : *forknum = BufTagGetForkNum(&bufHdr->tag);
4261 30089786 : *blknum = bufHdr->tag.blockNum;
4262 30089786 : }
4263 :
4264 : /*
4265 : * FlushBuffer
4266 : * Physically write out a shared buffer.
4267 : *
4268 : * NOTE: this actually just passes the buffer contents to the kernel; the
4269 : * real write to disk won't happen until the kernel feels like it. This
4270 : * is okay from our point of view since we can redo the changes from WAL.
4271 : * However, we will need to force the changes to disk via fsync before
4272 : * we can checkpoint WAL.
4273 : *
4274 : * The caller must hold a pin on the buffer and have share-locked the
4275 : * buffer contents. (Note: a share-lock does not prevent updates of
4276 : * hint bits in the buffer, so the page could change while the write
4277 : * is in progress, but we assume that that will not invalidate the data
4278 : * written.)
4279 : *
4280 : * If the caller has an smgr reference for the buffer's relation, pass it
4281 : * as the second parameter. If not, pass NULL.
4282 : */
4283 : static void
4284 1122466 : FlushBuffer(BufferDesc *buf, SMgrRelation reln, IOObject io_object,
4285 : IOContext io_context)
4286 : {
4287 : XLogRecPtr recptr;
4288 : ErrorContextCallback errcallback;
4289 : instr_time io_start;
4290 : Block bufBlock;
4291 : char *bufToWrite;
4292 : uint32 buf_state;
4293 :
4294 : /*
4295 : * Try to start an I/O operation. If StartBufferIO returns false, then
4296 : * someone else flushed the buffer before we could, so we need not do
4297 : * anything.
4298 : */
4299 1122466 : if (!StartBufferIO(buf, false, false))
4300 6 : return;
4301 :
4302 : /* Setup error traceback support for ereport() */
4303 1122460 : errcallback.callback = shared_buffer_write_error_callback;
4304 1122460 : errcallback.arg = buf;
4305 1122460 : errcallback.previous = error_context_stack;
4306 1122460 : error_context_stack = &errcallback;
4307 :
4308 : /* Find smgr relation for buffer */
4309 1122460 : if (reln == NULL)
4310 1114186 : reln = smgropen(BufTagGetRelFileLocator(&buf->tag), INVALID_PROC_NUMBER);
4311 :
4312 : TRACE_POSTGRESQL_BUFFER_FLUSH_START(BufTagGetForkNum(&buf->tag),
4313 : buf->tag.blockNum,
4314 : reln->smgr_rlocator.locator.spcOid,
4315 : reln->smgr_rlocator.locator.dbOid,
4316 : reln->smgr_rlocator.locator.relNumber);
4317 :
4318 1122460 : buf_state = LockBufHdr(buf);
4319 :
4320 : /*
4321 : * Run PageGetLSN while holding header lock, since we don't have the
4322 : * buffer locked exclusively in all cases.
4323 : */
4324 1122460 : recptr = BufferGetLSN(buf);
4325 :
4326 : /* To check if block content changes while flushing. - vadim 01/17/97 */
4327 1122460 : buf_state &= ~BM_JUST_DIRTIED;
4328 1122460 : UnlockBufHdr(buf, buf_state);
4329 :
4330 : /*
4331 : * Force XLOG flush up to buffer's LSN. This implements the basic WAL
4332 : * rule that log updates must hit disk before any of the data-file changes
4333 : * they describe do.
4334 : *
4335 : * However, this rule does not apply to unlogged relations, which will be
4336 : * lost after a crash anyway. Most unlogged relation pages do not bear
4337 : * LSNs since we never emit WAL records for them, and therefore flushing
4338 : * up through the buffer LSN would be useless, but harmless. However,
4339 : * GiST indexes use LSNs internally to track page-splits, and therefore
4340 : * unlogged GiST pages bear "fake" LSNs generated by
4341 : * GetFakeLSNForUnloggedRel. It is unlikely but possible that the fake
4342 : * LSN counter could advance past the WAL insertion point; and if it did
4343 : * happen, attempting to flush WAL through that location would fail, with
4344 : * disastrous system-wide consequences. To make sure that can't happen,
4345 : * skip the flush if the buffer isn't permanent.
4346 : */
4347 1122460 : if (buf_state & BM_PERMANENT)
4348 1118832 : XLogFlush(recptr);
4349 :
4350 : /*
4351 : * Now it's safe to write the buffer to disk. Note that no one else should
4352 : * have been able to write it, while we were busy with log flushing,
4353 : * because we got the exclusive right to perform I/O by setting the
4354 : * BM_IO_IN_PROGRESS bit.
4355 : */
4356 1122460 : bufBlock = BufHdrGetBlock(buf);
4357 :
4358 : /*
4359 : * Update page checksum if desired. Since we have only shared lock on the
4360 : * buffer, other processes might be updating hint bits in it, so we must
4361 : * copy the page to private storage if we do checksumming.
4362 : */
4363 1122460 : bufToWrite = PageSetChecksumCopy((Page) bufBlock, buf->tag.blockNum);
4364 :
4365 1122460 : io_start = pgstat_prepare_io_time(track_io_timing);
4366 :
4367 : /*
4368 : * bufToWrite is either the shared buffer or a copy, as appropriate.
4369 : */
4370 1122460 : smgrwrite(reln,
4371 1122460 : BufTagGetForkNum(&buf->tag),
4372 : buf->tag.blockNum,
4373 : bufToWrite,
4374 : false);
4375 :
4376 : /*
4377 : * When a strategy is in use, only flushes of dirty buffers already in the
4378 : * strategy ring are counted as strategy writes (IOCONTEXT
4379 : * [BULKREAD|BULKWRITE|VACUUM] IOOP_WRITE) for the purpose of IO
4380 : * statistics tracking.
4381 : *
4382 : * If a shared buffer initially added to the ring must be flushed before
4383 : * being used, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE.
4384 : *
4385 : * If a shared buffer which was added to the ring later because the
4386 : * current strategy buffer is pinned or in use or because all strategy
4387 : * buffers were dirty and rejected (for BAS_BULKREAD operations only)
4388 : * requires flushing, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE
4389 : * (from_ring will be false).
4390 : *
4391 : * When a strategy is not in use, the write can only be a "regular" write
4392 : * of a dirty shared buffer (IOCONTEXT_NORMAL IOOP_WRITE).
4393 : */
4394 1122460 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
4395 : IOOP_WRITE, io_start, 1, BLCKSZ);
4396 :
4397 1122460 : pgBufferUsage.shared_blks_written++;
4398 :
4399 : /*
4400 : * Mark the buffer as clean (unless BM_JUST_DIRTIED has become set) and
4401 : * end the BM_IO_IN_PROGRESS state.
4402 : */
4403 1122460 : TerminateBufferIO(buf, true, 0, true, false);
4404 :
4405 : TRACE_POSTGRESQL_BUFFER_FLUSH_DONE(BufTagGetForkNum(&buf->tag),
4406 : buf->tag.blockNum,
4407 : reln->smgr_rlocator.locator.spcOid,
4408 : reln->smgr_rlocator.locator.dbOid,
4409 : reln->smgr_rlocator.locator.relNumber);
4410 :
4411 : /* Pop the error context stack */
4412 1122460 : error_context_stack = errcallback.previous;
4413 : }
4414 :
4415 : /*
4416 : * Convenience wrapper around FlushBuffer() that locks/unlocks the buffer
4417 : * before/after calling FlushBuffer().
4418 : */
4419 : static void
4420 604984 : FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
4421 : IOObject io_object, IOContext io_context)
4422 : {
4423 604984 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_SHARED);
4424 604984 : FlushBuffer(buf, reln, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
4425 604984 : LWLockRelease(BufferDescriptorGetContentLock(buf));
4426 604984 : }
4427 :
4428 : /*
4429 : * RelationGetNumberOfBlocksInFork
4430 : * Determines the current number of pages in the specified relation fork.
4431 : *
4432 : * Note that the accuracy of the result will depend on the details of the
4433 : * relation's storage. For builtin AMs it'll be accurate, but for external AMs
4434 : * it might not be.
4435 : */
4436 : BlockNumber
4437 3788182 : RelationGetNumberOfBlocksInFork(Relation relation, ForkNumber forkNum)
4438 : {
4439 3788182 : if (RELKIND_HAS_TABLE_AM(relation->rd_rel->relkind))
4440 : {
4441 : /*
4442 : * Not every table AM uses BLCKSZ wide fixed size blocks. Therefore
4443 : * tableam returns the size in bytes - but for the purpose of this
4444 : * routine, we want the number of blocks. Therefore divide, rounding
4445 : * up.
4446 : */
4447 : uint64 szbytes;
4448 :
4449 2891120 : szbytes = table_relation_size(relation, forkNum);
4450 :
4451 2891082 : return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
4452 : }
4453 897062 : else if (RELKIND_HAS_STORAGE(relation->rd_rel->relkind))
4454 : {
4455 897062 : return smgrnblocks(RelationGetSmgr(relation), forkNum);
4456 : }
4457 : else
4458 : Assert(false);
4459 :
4460 0 : return 0; /* keep compiler quiet */
4461 : }
4462 :
4463 : /*
4464 : * BufferIsPermanent
4465 : * Determines whether a buffer will potentially still be around after
4466 : * a crash. Caller must hold a buffer pin.
4467 : */
4468 : bool
4469 18909776 : BufferIsPermanent(Buffer buffer)
4470 : {
4471 : BufferDesc *bufHdr;
4472 :
4473 : /* Local buffers are used only for temp relations. */
4474 18909776 : if (BufferIsLocal(buffer))
4475 1253940 : return false;
4476 :
4477 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4478 : Assert(BufferIsValid(buffer));
4479 : Assert(BufferIsPinned(buffer));
4480 :
4481 : /*
4482 : * BM_PERMANENT can't be changed while we hold a pin on the buffer, so we
4483 : * need not bother with the buffer header spinlock. Even if someone else
4484 : * changes the buffer header state while we're doing this, the state is
4485 : * changed atomically, so we'll read the old value or the new value, but
4486 : * not random garbage.
4487 : */
4488 17655836 : bufHdr = GetBufferDescriptor(buffer - 1);
4489 17655836 : return (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT) != 0;
4490 : }
4491 :
4492 : /*
4493 : * BufferGetLSNAtomic
4494 : * Retrieves the LSN of the buffer atomically using a buffer header lock.
4495 : * This is necessary for some callers who may not have an exclusive lock
4496 : * on the buffer.
4497 : */
4498 : XLogRecPtr
4499 14566036 : BufferGetLSNAtomic(Buffer buffer)
4500 : {
4501 14566036 : char *page = BufferGetPage(buffer);
4502 : BufferDesc *bufHdr;
4503 : XLogRecPtr lsn;
4504 : uint32 buf_state;
4505 :
4506 : /*
4507 : * If we don't need locking for correctness, fastpath out.
4508 : */
4509 14566036 : if (!XLogHintBitIsNeeded() || BufferIsLocal(buffer))
4510 475562 : return PageGetLSN(page);
4511 :
4512 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4513 : Assert(BufferIsValid(buffer));
4514 : Assert(BufferIsPinned(buffer));
4515 :
4516 14090474 : bufHdr = GetBufferDescriptor(buffer - 1);
4517 14090474 : buf_state = LockBufHdr(bufHdr);
4518 14090474 : lsn = PageGetLSN(page);
4519 14090474 : UnlockBufHdr(bufHdr, buf_state);
4520 :
4521 14090474 : return lsn;
4522 : }
4523 :
4524 : /* ---------------------------------------------------------------------
4525 : * DropRelationBuffers
4526 : *
4527 : * This function removes from the buffer pool all the pages of the
4528 : * specified relation forks that have block numbers >= firstDelBlock.
4529 : * (In particular, with firstDelBlock = 0, all pages are removed.)
4530 : * Dirty pages are simply dropped, without bothering to write them
4531 : * out first. Therefore, this is NOT rollback-able, and so should be
4532 : * used only with extreme caution!
4533 : *
4534 : * Currently, this is called only from smgr.c when the underlying file
4535 : * is about to be deleted or truncated (firstDelBlock is needed for
4536 : * the truncation case). The data in the affected pages would therefore
4537 : * be deleted momentarily anyway, and there is no point in writing it.
4538 : * It is the responsibility of higher-level code to ensure that the
4539 : * deletion or truncation does not lose any data that could be needed
4540 : * later. It is also the responsibility of higher-level code to ensure
4541 : * that no other process could be trying to load more pages of the
4542 : * relation into buffers.
4543 : * --------------------------------------------------------------------
4544 : */
4545 : void
4546 1282 : DropRelationBuffers(SMgrRelation smgr_reln, ForkNumber *forkNum,
4547 : int nforks, BlockNumber *firstDelBlock)
4548 : {
4549 : int i;
4550 : int j;
4551 : RelFileLocatorBackend rlocator;
4552 : BlockNumber nForkBlock[MAX_FORKNUM];
4553 1282 : uint64 nBlocksToInvalidate = 0;
4554 :
4555 1282 : rlocator = smgr_reln->smgr_rlocator;
4556 :
4557 : /* If it's a local relation, it's localbuf.c's problem. */
4558 1282 : if (RelFileLocatorBackendIsTemp(rlocator))
4559 : {
4560 748 : if (rlocator.backend == MyProcNumber)
4561 748 : DropRelationLocalBuffers(rlocator.locator, forkNum, nforks,
4562 : firstDelBlock);
4563 :
4564 836 : return;
4565 : }
4566 :
4567 : /*
4568 : * To remove all the pages of the specified relation forks from the buffer
4569 : * pool, we need to scan the entire buffer pool but we can optimize it by
4570 : * finding the buffers from BufMapping table provided we know the exact
4571 : * size of each fork of the relation. The exact size is required to ensure
4572 : * that we don't leave any buffer for the relation being dropped as
4573 : * otherwise the background writer or checkpointer can lead to a PANIC
4574 : * error while flushing buffers corresponding to files that don't exist.
4575 : *
4576 : * To know the exact size, we rely on the size cached for each fork by us
4577 : * during recovery which limits the optimization to recovery and on
4578 : * standbys but we can easily extend it once we have shared cache for
4579 : * relation size.
4580 : *
4581 : * In recovery, we cache the value returned by the first lseek(SEEK_END)
4582 : * and the future writes keeps the cached value up-to-date. See
4583 : * smgrextend. It is possible that the value of the first lseek is smaller
4584 : * than the actual number of existing blocks in the file due to buggy
4585 : * Linux kernels that might not have accounted for the recent write. But
4586 : * that should be fine because there must not be any buffers after that
4587 : * file size.
4588 : */
4589 746 : for (i = 0; i < nforks; i++)
4590 : {
4591 : /* Get the number of blocks for a relation's fork */
4592 634 : nForkBlock[i] = smgrnblocks_cached(smgr_reln, forkNum[i]);
4593 :
4594 634 : if (nForkBlock[i] == InvalidBlockNumber)
4595 : {
4596 422 : nBlocksToInvalidate = InvalidBlockNumber;
4597 422 : break;
4598 : }
4599 :
4600 : /* calculate the number of blocks to be invalidated */
4601 212 : nBlocksToInvalidate += (nForkBlock[i] - firstDelBlock[i]);
4602 : }
4603 :
4604 : /*
4605 : * We apply the optimization iff the total number of blocks to invalidate
4606 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4607 : */
4608 534 : if (BlockNumberIsValid(nBlocksToInvalidate) &&
4609 112 : nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4610 : {
4611 244 : for (j = 0; j < nforks; j++)
4612 156 : FindAndDropRelationBuffers(rlocator.locator, forkNum[j],
4613 156 : nForkBlock[j], firstDelBlock[j]);
4614 88 : return;
4615 : }
4616 :
4617 5714622 : for (i = 0; i < NBuffers; i++)
4618 : {
4619 5714176 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4620 : uint32 buf_state;
4621 :
4622 : /*
4623 : * We can make this a tad faster by prechecking the buffer tag before
4624 : * we attempt to lock the buffer; this saves a lot of lock
4625 : * acquisitions in typical cases. It should be safe because the
4626 : * caller must have AccessExclusiveLock on the relation, or some other
4627 : * reason to be certain that no one is loading new pages of the rel
4628 : * into the buffer pool. (Otherwise we might well miss such pages
4629 : * entirely.) Therefore, while the tag might be changing while we
4630 : * look at it, it can't be changing *to* a value we care about, only
4631 : * *away* from such a value. So false negatives are impossible, and
4632 : * false positives are safe because we'll recheck after getting the
4633 : * buffer lock.
4634 : *
4635 : * We could check forkNum and blockNum as well as the rlocator, but
4636 : * the incremental win from doing so seems small.
4637 : */
4638 5714176 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator))
4639 5700380 : continue;
4640 :
4641 13796 : buf_state = LockBufHdr(bufHdr);
4642 :
4643 35608 : for (j = 0; j < nforks; j++)
4644 : {
4645 24968 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator) &&
4646 24968 : BufTagGetForkNum(&bufHdr->tag) == forkNum[j] &&
4647 13594 : bufHdr->tag.blockNum >= firstDelBlock[j])
4648 : {
4649 3156 : InvalidateBuffer(bufHdr); /* releases spinlock */
4650 3156 : break;
4651 : }
4652 : }
4653 13796 : if (j >= nforks)
4654 10640 : UnlockBufHdr(bufHdr, buf_state);
4655 : }
4656 : }
4657 :
4658 : /* ---------------------------------------------------------------------
4659 : * DropRelationsAllBuffers
4660 : *
4661 : * This function removes from the buffer pool all the pages of all
4662 : * forks of the specified relations. It's equivalent to calling
4663 : * DropRelationBuffers once per fork per relation with firstDelBlock = 0.
4664 : * --------------------------------------------------------------------
4665 : */
4666 : void
4667 27182 : DropRelationsAllBuffers(SMgrRelation *smgr_reln, int nlocators)
4668 : {
4669 : int i;
4670 27182 : int n = 0;
4671 : SMgrRelation *rels;
4672 : BlockNumber (*block)[MAX_FORKNUM + 1];
4673 27182 : uint64 nBlocksToInvalidate = 0;
4674 : RelFileLocator *locators;
4675 27182 : bool cached = true;
4676 : bool use_bsearch;
4677 :
4678 27182 : if (nlocators == 0)
4679 0 : return;
4680 :
4681 27182 : rels = palloc(sizeof(SMgrRelation) * nlocators); /* non-local relations */
4682 :
4683 : /* If it's a local relation, it's localbuf.c's problem. */
4684 119066 : for (i = 0; i < nlocators; i++)
4685 : {
4686 91884 : if (RelFileLocatorBackendIsTemp(smgr_reln[i]->smgr_rlocator))
4687 : {
4688 6348 : if (smgr_reln[i]->smgr_rlocator.backend == MyProcNumber)
4689 6348 : DropRelationAllLocalBuffers(smgr_reln[i]->smgr_rlocator.locator);
4690 : }
4691 : else
4692 85536 : rels[n++] = smgr_reln[i];
4693 : }
4694 :
4695 : /*
4696 : * If there are no non-local relations, then we're done. Release the
4697 : * memory and return.
4698 : */
4699 27182 : if (n == 0)
4700 : {
4701 1654 : pfree(rels);
4702 1654 : return;
4703 : }
4704 :
4705 : /*
4706 : * This is used to remember the number of blocks for all the relations
4707 : * forks.
4708 : */
4709 : block = (BlockNumber (*)[MAX_FORKNUM + 1])
4710 25528 : palloc(sizeof(BlockNumber) * n * (MAX_FORKNUM + 1));
4711 :
4712 : /*
4713 : * We can avoid scanning the entire buffer pool if we know the exact size
4714 : * of each of the given relation forks. See DropRelationBuffers.
4715 : */
4716 53380 : for (i = 0; i < n && cached; i++)
4717 : {
4718 43194 : for (int j = 0; j <= MAX_FORKNUM; j++)
4719 : {
4720 : /* Get the number of blocks for a relation's fork. */
4721 39388 : block[i][j] = smgrnblocks_cached(rels[i], j);
4722 :
4723 : /* We need to only consider the relation forks that exists. */
4724 39388 : if (block[i][j] == InvalidBlockNumber)
4725 : {
4726 35250 : if (!smgrexists(rels[i], j))
4727 11204 : continue;
4728 24046 : cached = false;
4729 24046 : break;
4730 : }
4731 :
4732 : /* calculate the total number of blocks to be invalidated */
4733 4138 : nBlocksToInvalidate += block[i][j];
4734 : }
4735 : }
4736 :
4737 : /*
4738 : * We apply the optimization iff the total number of blocks to invalidate
4739 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4740 : */
4741 25528 : if (cached && nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4742 : {
4743 2458 : for (i = 0; i < n; i++)
4744 : {
4745 6740 : for (int j = 0; j <= MAX_FORKNUM; j++)
4746 : {
4747 : /* ignore relation forks that doesn't exist */
4748 5392 : if (!BlockNumberIsValid(block[i][j]))
4749 4026 : continue;
4750 :
4751 : /* drop all the buffers for a particular relation fork */
4752 1366 : FindAndDropRelationBuffers(rels[i]->smgr_rlocator.locator,
4753 1366 : j, block[i][j], 0);
4754 : }
4755 : }
4756 :
4757 1110 : pfree(block);
4758 1110 : pfree(rels);
4759 1110 : return;
4760 : }
4761 :
4762 24418 : pfree(block);
4763 24418 : locators = palloc(sizeof(RelFileLocator) * n); /* non-local relations */
4764 108606 : for (i = 0; i < n; i++)
4765 84188 : locators[i] = rels[i]->smgr_rlocator.locator;
4766 :
4767 : /*
4768 : * For low number of relations to drop just use a simple walk through, to
4769 : * save the bsearch overhead. The threshold to use is rather a guess than
4770 : * an exactly determined value, as it depends on many factors (CPU and RAM
4771 : * speeds, amount of shared buffers etc.).
4772 : */
4773 24418 : use_bsearch = n > RELS_BSEARCH_THRESHOLD;
4774 :
4775 : /* sort the list of rlocators if necessary */
4776 24418 : if (use_bsearch)
4777 346 : qsort(locators, n, sizeof(RelFileLocator), rlocator_comparator);
4778 :
4779 264286306 : for (i = 0; i < NBuffers; i++)
4780 : {
4781 264261888 : RelFileLocator *rlocator = NULL;
4782 264261888 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4783 : uint32 buf_state;
4784 :
4785 : /*
4786 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4787 : * saves some cycles.
4788 : */
4789 :
4790 264261888 : if (!use_bsearch)
4791 : {
4792 : int j;
4793 :
4794 1053725440 : for (j = 0; j < n; j++)
4795 : {
4796 793358338 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &locators[j]))
4797 : {
4798 176642 : rlocator = &locators[j];
4799 176642 : break;
4800 : }
4801 : }
4802 : }
4803 : else
4804 : {
4805 : RelFileLocator locator;
4806 :
4807 3718144 : locator = BufTagGetRelFileLocator(&bufHdr->tag);
4808 3718144 : rlocator = bsearch(&locator,
4809 : locators, n, sizeof(RelFileLocator),
4810 : rlocator_comparator);
4811 : }
4812 :
4813 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
4814 264261888 : if (rlocator == NULL)
4815 264081664 : continue;
4816 :
4817 180224 : buf_state = LockBufHdr(bufHdr);
4818 180224 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, rlocator))
4819 180224 : InvalidateBuffer(bufHdr); /* releases spinlock */
4820 : else
4821 0 : UnlockBufHdr(bufHdr, buf_state);
4822 : }
4823 :
4824 24418 : pfree(locators);
4825 24418 : pfree(rels);
4826 : }
4827 :
4828 : /* ---------------------------------------------------------------------
4829 : * FindAndDropRelationBuffers
4830 : *
4831 : * This function performs look up in BufMapping table and removes from the
4832 : * buffer pool all the pages of the specified relation fork that has block
4833 : * number >= firstDelBlock. (In particular, with firstDelBlock = 0, all
4834 : * pages are removed.)
4835 : * --------------------------------------------------------------------
4836 : */
4837 : static void
4838 1522 : FindAndDropRelationBuffers(RelFileLocator rlocator, ForkNumber forkNum,
4839 : BlockNumber nForkBlock,
4840 : BlockNumber firstDelBlock)
4841 : {
4842 : BlockNumber curBlock;
4843 :
4844 3700 : for (curBlock = firstDelBlock; curBlock < nForkBlock; curBlock++)
4845 : {
4846 : uint32 bufHash; /* hash value for tag */
4847 : BufferTag bufTag; /* identity of requested block */
4848 : LWLock *bufPartitionLock; /* buffer partition lock for it */
4849 : int buf_id;
4850 : BufferDesc *bufHdr;
4851 : uint32 buf_state;
4852 :
4853 : /* create a tag so we can lookup the buffer */
4854 2178 : InitBufferTag(&bufTag, &rlocator, forkNum, curBlock);
4855 :
4856 : /* determine its hash code and partition lock ID */
4857 2178 : bufHash = BufTableHashCode(&bufTag);
4858 2178 : bufPartitionLock = BufMappingPartitionLock(bufHash);
4859 :
4860 : /* Check that it is in the buffer pool. If not, do nothing. */
4861 2178 : LWLockAcquire(bufPartitionLock, LW_SHARED);
4862 2178 : buf_id = BufTableLookup(&bufTag, bufHash);
4863 2178 : LWLockRelease(bufPartitionLock);
4864 :
4865 2178 : if (buf_id < 0)
4866 178 : continue;
4867 :
4868 2000 : bufHdr = GetBufferDescriptor(buf_id);
4869 :
4870 : /*
4871 : * We need to lock the buffer header and recheck if the buffer is
4872 : * still associated with the same block because the buffer could be
4873 : * evicted by some other backend loading blocks for a different
4874 : * relation after we release lock on the BufMapping table.
4875 : */
4876 2000 : buf_state = LockBufHdr(bufHdr);
4877 :
4878 4000 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator) &&
4879 2000 : BufTagGetForkNum(&bufHdr->tag) == forkNum &&
4880 2000 : bufHdr->tag.blockNum >= firstDelBlock)
4881 2000 : InvalidateBuffer(bufHdr); /* releases spinlock */
4882 : else
4883 0 : UnlockBufHdr(bufHdr, buf_state);
4884 : }
4885 1522 : }
4886 :
4887 : /* ---------------------------------------------------------------------
4888 : * DropDatabaseBuffers
4889 : *
4890 : * This function removes all the buffers in the buffer cache for a
4891 : * particular database. Dirty pages are simply dropped, without
4892 : * bothering to write them out first. This is used when we destroy a
4893 : * database, to avoid trying to flush data to disk when the directory
4894 : * tree no longer exists. Implementation is pretty similar to
4895 : * DropRelationBuffers() which is for destroying just one relation.
4896 : * --------------------------------------------------------------------
4897 : */
4898 : void
4899 144 : DropDatabaseBuffers(Oid dbid)
4900 : {
4901 : int i;
4902 :
4903 : /*
4904 : * We needn't consider local buffers, since by assumption the target
4905 : * database isn't our own.
4906 : */
4907 :
4908 961424 : for (i = 0; i < NBuffers; i++)
4909 : {
4910 961280 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4911 : uint32 buf_state;
4912 :
4913 : /*
4914 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4915 : * saves some cycles.
4916 : */
4917 961280 : if (bufHdr->tag.dbOid != dbid)
4918 936170 : continue;
4919 :
4920 25110 : buf_state = LockBufHdr(bufHdr);
4921 25110 : if (bufHdr->tag.dbOid == dbid)
4922 25110 : InvalidateBuffer(bufHdr); /* releases spinlock */
4923 : else
4924 0 : UnlockBufHdr(bufHdr, buf_state);
4925 : }
4926 144 : }
4927 :
4928 : /* ---------------------------------------------------------------------
4929 : * FlushRelationBuffers
4930 : *
4931 : * This function writes all dirty pages of a relation out to disk
4932 : * (or more accurately, out to kernel disk buffers), ensuring that the
4933 : * kernel has an up-to-date view of the relation.
4934 : *
4935 : * Generally, the caller should be holding AccessExclusiveLock on the
4936 : * target relation to ensure that no other backend is busy dirtying
4937 : * more blocks of the relation; the effects can't be expected to last
4938 : * after the lock is released.
4939 : *
4940 : * XXX currently it sequentially searches the buffer pool, should be
4941 : * changed to more clever ways of searching. This routine is not
4942 : * used in any performance-critical code paths, so it's not worth
4943 : * adding additional overhead to normal paths to make it go faster.
4944 : * --------------------------------------------------------------------
4945 : */
4946 : void
4947 276 : FlushRelationBuffers(Relation rel)
4948 : {
4949 : int i;
4950 : BufferDesc *bufHdr;
4951 276 : SMgrRelation srel = RelationGetSmgr(rel);
4952 :
4953 276 : if (RelationUsesLocalBuffers(rel))
4954 : {
4955 1818 : for (i = 0; i < NLocBuffer; i++)
4956 : {
4957 : uint32 buf_state;
4958 :
4959 1800 : bufHdr = GetLocalBufferDescriptor(i);
4960 1800 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
4961 600 : ((buf_state = pg_atomic_read_u32(&bufHdr->state)) &
4962 : (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
4963 : {
4964 : ErrorContextCallback errcallback;
4965 :
4966 : /* Setup error traceback support for ereport() */
4967 600 : errcallback.callback = local_buffer_write_error_callback;
4968 600 : errcallback.arg = bufHdr;
4969 600 : errcallback.previous = error_context_stack;
4970 600 : error_context_stack = &errcallback;
4971 :
4972 : /* Make sure we can handle the pin */
4973 600 : ReservePrivateRefCountEntry();
4974 600 : ResourceOwnerEnlarge(CurrentResourceOwner);
4975 :
4976 : /*
4977 : * Pin/unpin mostly to make valgrind work, but it also seems
4978 : * like the right thing to do.
4979 : */
4980 600 : PinLocalBuffer(bufHdr, false);
4981 :
4982 :
4983 600 : FlushLocalBuffer(bufHdr, srel);
4984 :
4985 600 : UnpinLocalBuffer(BufferDescriptorGetBuffer(bufHdr));
4986 :
4987 : /* Pop the error context stack */
4988 600 : error_context_stack = errcallback.previous;
4989 : }
4990 : }
4991 :
4992 18 : return;
4993 : }
4994 :
4995 3024386 : for (i = 0; i < NBuffers; i++)
4996 : {
4997 : uint32 buf_state;
4998 :
4999 3024128 : bufHdr = GetBufferDescriptor(i);
5000 :
5001 : /*
5002 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5003 : * saves some cycles.
5004 : */
5005 3024128 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator))
5006 3023682 : continue;
5007 :
5008 : /* Make sure we can handle the pin */
5009 446 : ReservePrivateRefCountEntry();
5010 446 : ResourceOwnerEnlarge(CurrentResourceOwner);
5011 :
5012 446 : buf_state = LockBufHdr(bufHdr);
5013 446 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
5014 446 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5015 : {
5016 360 : PinBuffer_Locked(bufHdr);
5017 360 : FlushUnlockedBuffer(bufHdr, srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5018 360 : UnpinBuffer(bufHdr);
5019 : }
5020 : else
5021 86 : UnlockBufHdr(bufHdr, buf_state);
5022 : }
5023 : }
5024 :
5025 : /* ---------------------------------------------------------------------
5026 : * FlushRelationsAllBuffers
5027 : *
5028 : * This function flushes out of the buffer pool all the pages of all
5029 : * forks of the specified smgr relations. It's equivalent to calling
5030 : * FlushRelationBuffers once per relation. The relations are assumed not
5031 : * to use local buffers.
5032 : * --------------------------------------------------------------------
5033 : */
5034 : void
5035 24 : FlushRelationsAllBuffers(SMgrRelation *smgrs, int nrels)
5036 : {
5037 : int i;
5038 : SMgrSortArray *srels;
5039 : bool use_bsearch;
5040 :
5041 24 : if (nrels == 0)
5042 0 : return;
5043 :
5044 : /* fill-in array for qsort */
5045 24 : srels = palloc(sizeof(SMgrSortArray) * nrels);
5046 :
5047 48 : for (i = 0; i < nrels; i++)
5048 : {
5049 : Assert(!RelFileLocatorBackendIsTemp(smgrs[i]->smgr_rlocator));
5050 :
5051 24 : srels[i].rlocator = smgrs[i]->smgr_rlocator.locator;
5052 24 : srels[i].srel = smgrs[i];
5053 : }
5054 :
5055 : /*
5056 : * Save the bsearch overhead for low number of relations to sync. See
5057 : * DropRelationsAllBuffers for details.
5058 : */
5059 24 : use_bsearch = nrels > RELS_BSEARCH_THRESHOLD;
5060 :
5061 : /* sort the list of SMgrRelations if necessary */
5062 24 : if (use_bsearch)
5063 0 : qsort(srels, nrels, sizeof(SMgrSortArray), rlocator_comparator);
5064 :
5065 393240 : for (i = 0; i < NBuffers; i++)
5066 : {
5067 393216 : SMgrSortArray *srelent = NULL;
5068 393216 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5069 : uint32 buf_state;
5070 :
5071 : /*
5072 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5073 : * saves some cycles.
5074 : */
5075 :
5076 393216 : if (!use_bsearch)
5077 : {
5078 : int j;
5079 :
5080 777604 : for (j = 0; j < nrels; j++)
5081 : {
5082 393216 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srels[j].rlocator))
5083 : {
5084 8828 : srelent = &srels[j];
5085 8828 : break;
5086 : }
5087 : }
5088 : }
5089 : else
5090 : {
5091 : RelFileLocator rlocator;
5092 :
5093 0 : rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
5094 0 : srelent = bsearch(&rlocator,
5095 : srels, nrels, sizeof(SMgrSortArray),
5096 : rlocator_comparator);
5097 : }
5098 :
5099 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
5100 393216 : if (srelent == NULL)
5101 384388 : continue;
5102 :
5103 : /* Make sure we can handle the pin */
5104 8828 : ReservePrivateRefCountEntry();
5105 8828 : ResourceOwnerEnlarge(CurrentResourceOwner);
5106 :
5107 8828 : buf_state = LockBufHdr(bufHdr);
5108 8828 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srelent->rlocator) &&
5109 8828 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5110 : {
5111 7914 : PinBuffer_Locked(bufHdr);
5112 7914 : FlushUnlockedBuffer(bufHdr, srelent->srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5113 7914 : UnpinBuffer(bufHdr);
5114 : }
5115 : else
5116 914 : UnlockBufHdr(bufHdr, buf_state);
5117 : }
5118 :
5119 24 : pfree(srels);
5120 : }
5121 :
5122 : /* ---------------------------------------------------------------------
5123 : * RelationCopyStorageUsingBuffer
5124 : *
5125 : * Copy fork's data using bufmgr. Same as RelationCopyStorage but instead
5126 : * of using smgrread and smgrextend this will copy using bufmgr APIs.
5127 : *
5128 : * Refer comments atop CreateAndCopyRelationData() for details about
5129 : * 'permanent' parameter.
5130 : * --------------------------------------------------------------------
5131 : */
5132 : static void
5133 149048 : RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
5134 : RelFileLocator dstlocator,
5135 : ForkNumber forkNum, bool permanent)
5136 : {
5137 : Buffer srcBuf;
5138 : Buffer dstBuf;
5139 : Page srcPage;
5140 : Page dstPage;
5141 : bool use_wal;
5142 : BlockNumber nblocks;
5143 : BlockNumber blkno;
5144 : PGIOAlignedBlock buf;
5145 : BufferAccessStrategy bstrategy_src;
5146 : BufferAccessStrategy bstrategy_dst;
5147 : BlockRangeReadStreamPrivate p;
5148 : ReadStream *src_stream;
5149 : SMgrRelation src_smgr;
5150 :
5151 : /*
5152 : * In general, we want to write WAL whenever wal_level > 'minimal', but we
5153 : * can skip it when copying any fork of an unlogged relation other than
5154 : * the init fork.
5155 : */
5156 149048 : use_wal = XLogIsNeeded() && (permanent || forkNum == INIT_FORKNUM);
5157 :
5158 : /* Get number of blocks in the source relation. */
5159 149048 : nblocks = smgrnblocks(smgropen(srclocator, INVALID_PROC_NUMBER),
5160 : forkNum);
5161 :
5162 : /* Nothing to copy; just return. */
5163 149048 : if (nblocks == 0)
5164 25998 : return;
5165 :
5166 : /*
5167 : * Bulk extend the destination relation of the same size as the source
5168 : * relation before starting to copy block by block.
5169 : */
5170 123050 : memset(buf.data, 0, BLCKSZ);
5171 123050 : smgrextend(smgropen(dstlocator, INVALID_PROC_NUMBER), forkNum, nblocks - 1,
5172 : buf.data, true);
5173 :
5174 : /* This is a bulk operation, so use buffer access strategies. */
5175 123050 : bstrategy_src = GetAccessStrategy(BAS_BULKREAD);
5176 123050 : bstrategy_dst = GetAccessStrategy(BAS_BULKWRITE);
5177 :
5178 : /* Initialize streaming read */
5179 123050 : p.current_blocknum = 0;
5180 123050 : p.last_exclusive = nblocks;
5181 123050 : src_smgr = smgropen(srclocator, INVALID_PROC_NUMBER);
5182 :
5183 : /*
5184 : * It is safe to use batchmode as block_range_read_stream_cb takes no
5185 : * locks.
5186 : */
5187 123050 : src_stream = read_stream_begin_smgr_relation(READ_STREAM_FULL |
5188 : READ_STREAM_USE_BATCHING,
5189 : bstrategy_src,
5190 : src_smgr,
5191 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
5192 : forkNum,
5193 : block_range_read_stream_cb,
5194 : &p,
5195 : 0);
5196 :
5197 : /* Iterate over each block of the source relation file. */
5198 593288 : for (blkno = 0; blkno < nblocks; blkno++)
5199 : {
5200 470242 : CHECK_FOR_INTERRUPTS();
5201 :
5202 : /* Read block from source relation. */
5203 470242 : srcBuf = read_stream_next_buffer(src_stream, NULL);
5204 470238 : LockBuffer(srcBuf, BUFFER_LOCK_SHARE);
5205 470238 : srcPage = BufferGetPage(srcBuf);
5206 :
5207 470238 : dstBuf = ReadBufferWithoutRelcache(dstlocator, forkNum,
5208 : BufferGetBlockNumber(srcBuf),
5209 : RBM_ZERO_AND_LOCK, bstrategy_dst,
5210 : permanent);
5211 470238 : dstPage = BufferGetPage(dstBuf);
5212 :
5213 470238 : START_CRIT_SECTION();
5214 :
5215 : /* Copy page data from the source to the destination. */
5216 470238 : memcpy(dstPage, srcPage, BLCKSZ);
5217 470238 : MarkBufferDirty(dstBuf);
5218 :
5219 : /* WAL-log the copied page. */
5220 470238 : if (use_wal)
5221 259620 : log_newpage_buffer(dstBuf, true);
5222 :
5223 470238 : END_CRIT_SECTION();
5224 :
5225 470238 : UnlockReleaseBuffer(dstBuf);
5226 470238 : UnlockReleaseBuffer(srcBuf);
5227 : }
5228 : Assert(read_stream_next_buffer(src_stream, NULL) == InvalidBuffer);
5229 123046 : read_stream_end(src_stream);
5230 :
5231 123046 : FreeAccessStrategy(bstrategy_src);
5232 123046 : FreeAccessStrategy(bstrategy_dst);
5233 : }
5234 :
5235 : /* ---------------------------------------------------------------------
5236 : * CreateAndCopyRelationData
5237 : *
5238 : * Create destination relation storage and copy all forks from the
5239 : * source relation to the destination.
5240 : *
5241 : * Pass permanent as true for permanent relations and false for
5242 : * unlogged relations. Currently this API is not supported for
5243 : * temporary relations.
5244 : * --------------------------------------------------------------------
5245 : */
5246 : void
5247 112040 : CreateAndCopyRelationData(RelFileLocator src_rlocator,
5248 : RelFileLocator dst_rlocator, bool permanent)
5249 : {
5250 : char relpersistence;
5251 : SMgrRelation src_rel;
5252 : SMgrRelation dst_rel;
5253 :
5254 : /* Set the relpersistence. */
5255 112040 : relpersistence = permanent ?
5256 : RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED;
5257 :
5258 112040 : src_rel = smgropen(src_rlocator, INVALID_PROC_NUMBER);
5259 112040 : dst_rel = smgropen(dst_rlocator, INVALID_PROC_NUMBER);
5260 :
5261 : /*
5262 : * Create and copy all forks of the relation. During create database we
5263 : * have a separate cleanup mechanism which deletes complete database
5264 : * directory. Therefore, each individual relation doesn't need to be
5265 : * registered for cleanup.
5266 : */
5267 112040 : RelationCreateStorage(dst_rlocator, relpersistence, false);
5268 :
5269 : /* copy main fork. */
5270 112040 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, MAIN_FORKNUM,
5271 : permanent);
5272 :
5273 : /* copy those extra forks that exist */
5274 112036 : for (ForkNumber forkNum = MAIN_FORKNUM + 1;
5275 448144 : forkNum <= MAX_FORKNUM; forkNum++)
5276 : {
5277 336108 : if (smgrexists(src_rel, forkNum))
5278 : {
5279 37008 : smgrcreate(dst_rel, forkNum, false);
5280 :
5281 : /*
5282 : * WAL log creation if the relation is persistent, or this is the
5283 : * init fork of an unlogged relation.
5284 : */
5285 37008 : if (permanent || forkNum == INIT_FORKNUM)
5286 37008 : log_smgrcreate(&dst_rlocator, forkNum);
5287 :
5288 : /* Copy a fork's data, block by block. */
5289 37008 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, forkNum,
5290 : permanent);
5291 : }
5292 : }
5293 112036 : }
5294 :
5295 : /* ---------------------------------------------------------------------
5296 : * FlushDatabaseBuffers
5297 : *
5298 : * This function writes all dirty pages of a database out to disk
5299 : * (or more accurately, out to kernel disk buffers), ensuring that the
5300 : * kernel has an up-to-date view of the database.
5301 : *
5302 : * Generally, the caller should be holding an appropriate lock to ensure
5303 : * no other backend is active in the target database; otherwise more
5304 : * pages could get dirtied.
5305 : *
5306 : * Note we don't worry about flushing any pages of temporary relations.
5307 : * It's assumed these wouldn't be interesting.
5308 : * --------------------------------------------------------------------
5309 : */
5310 : void
5311 10 : FlushDatabaseBuffers(Oid dbid)
5312 : {
5313 : int i;
5314 : BufferDesc *bufHdr;
5315 :
5316 1290 : for (i = 0; i < NBuffers; i++)
5317 : {
5318 : uint32 buf_state;
5319 :
5320 1280 : bufHdr = GetBufferDescriptor(i);
5321 :
5322 : /*
5323 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5324 : * saves some cycles.
5325 : */
5326 1280 : if (bufHdr->tag.dbOid != dbid)
5327 968 : continue;
5328 :
5329 : /* Make sure we can handle the pin */
5330 312 : ReservePrivateRefCountEntry();
5331 312 : ResourceOwnerEnlarge(CurrentResourceOwner);
5332 :
5333 312 : buf_state = LockBufHdr(bufHdr);
5334 312 : if (bufHdr->tag.dbOid == dbid &&
5335 312 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5336 : {
5337 36 : PinBuffer_Locked(bufHdr);
5338 36 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5339 36 : UnpinBuffer(bufHdr);
5340 : }
5341 : else
5342 276 : UnlockBufHdr(bufHdr, buf_state);
5343 : }
5344 10 : }
5345 :
5346 : /*
5347 : * Flush a previously, shared or exclusively, locked and pinned buffer to the
5348 : * OS.
5349 : */
5350 : void
5351 160 : FlushOneBuffer(Buffer buffer)
5352 : {
5353 : BufferDesc *bufHdr;
5354 :
5355 : /* currently not needed, but no fundamental reason not to support */
5356 : Assert(!BufferIsLocal(buffer));
5357 :
5358 : Assert(BufferIsPinned(buffer));
5359 :
5360 160 : bufHdr = GetBufferDescriptor(buffer - 1);
5361 :
5362 : Assert(BufferIsLockedByMe(buffer));
5363 :
5364 160 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5365 160 : }
5366 :
5367 : /*
5368 : * ReleaseBuffer -- release the pin on a buffer
5369 : */
5370 : void
5371 126397164 : ReleaseBuffer(Buffer buffer)
5372 : {
5373 126397164 : if (!BufferIsValid(buffer))
5374 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5375 :
5376 126397164 : if (BufferIsLocal(buffer))
5377 3220512 : UnpinLocalBuffer(buffer);
5378 : else
5379 123176652 : UnpinBuffer(GetBufferDescriptor(buffer - 1));
5380 126397164 : }
5381 :
5382 : /*
5383 : * UnlockReleaseBuffer -- release the content lock and pin on a buffer
5384 : *
5385 : * This is just a shorthand for a common combination.
5386 : */
5387 : void
5388 38191382 : UnlockReleaseBuffer(Buffer buffer)
5389 : {
5390 38191382 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5391 38191382 : ReleaseBuffer(buffer);
5392 38191382 : }
5393 :
5394 : /*
5395 : * IncrBufferRefCount
5396 : * Increment the pin count on a buffer that we have *already* pinned
5397 : * at least once.
5398 : *
5399 : * This function cannot be used on a buffer we do not have pinned,
5400 : * because it doesn't change the shared buffer state.
5401 : */
5402 : void
5403 23603808 : IncrBufferRefCount(Buffer buffer)
5404 : {
5405 : Assert(BufferIsPinned(buffer));
5406 23603808 : ResourceOwnerEnlarge(CurrentResourceOwner);
5407 23603808 : if (BufferIsLocal(buffer))
5408 709372 : LocalRefCount[-buffer - 1]++;
5409 : else
5410 : {
5411 : PrivateRefCountEntry *ref;
5412 :
5413 22894436 : ref = GetPrivateRefCountEntry(buffer, true);
5414 : Assert(ref != NULL);
5415 22894436 : ref->refcount++;
5416 : }
5417 23603808 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buffer);
5418 23603808 : }
5419 :
5420 : /*
5421 : * MarkBufferDirtyHint
5422 : *
5423 : * Mark a buffer dirty for non-critical changes.
5424 : *
5425 : * This is essentially the same as MarkBufferDirty, except:
5426 : *
5427 : * 1. The caller does not write WAL; so if checksums are enabled, we may need
5428 : * to write an XLOG_FPI_FOR_HINT WAL record to protect against torn pages.
5429 : * 2. The caller might have only share-lock instead of exclusive-lock on the
5430 : * buffer's content lock.
5431 : * 3. This function does not guarantee that the buffer is always marked dirty
5432 : * (due to a race condition), so it cannot be used for important changes.
5433 : */
5434 : void
5435 19819098 : MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
5436 : {
5437 : BufferDesc *bufHdr;
5438 19819098 : Page page = BufferGetPage(buffer);
5439 :
5440 19819098 : if (!BufferIsValid(buffer))
5441 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5442 :
5443 19819098 : if (BufferIsLocal(buffer))
5444 : {
5445 1270290 : MarkLocalBufferDirty(buffer);
5446 1270290 : return;
5447 : }
5448 :
5449 18548808 : bufHdr = GetBufferDescriptor(buffer - 1);
5450 :
5451 : Assert(GetPrivateRefCount(buffer) > 0);
5452 : /* here, either share or exclusive lock is OK */
5453 : Assert(BufferIsLockedByMe(buffer));
5454 :
5455 : /*
5456 : * This routine might get called many times on the same page, if we are
5457 : * making the first scan after commit of an xact that added/deleted many
5458 : * tuples. So, be as quick as we can if the buffer is already dirty. We
5459 : * do this by not acquiring spinlock if it looks like the status bits are
5460 : * already set. Since we make this test unlocked, there's a chance we
5461 : * might fail to notice that the flags have just been cleared, and failed
5462 : * to reset them, due to memory-ordering issues. But since this function
5463 : * is only intended to be used in cases where failing to write out the
5464 : * data would be harmless anyway, it doesn't really matter.
5465 : */
5466 18548808 : if ((pg_atomic_read_u32(&bufHdr->state) & (BM_DIRTY | BM_JUST_DIRTIED)) !=
5467 : (BM_DIRTY | BM_JUST_DIRTIED))
5468 : {
5469 1602946 : XLogRecPtr lsn = InvalidXLogRecPtr;
5470 1602946 : bool dirtied = false;
5471 1602946 : bool delayChkptFlags = false;
5472 : uint32 buf_state;
5473 :
5474 : /*
5475 : * If we need to protect hint bit updates from torn writes, WAL-log a
5476 : * full page image of the page. This full page image is only necessary
5477 : * if the hint bit update is the first change to the page since the
5478 : * last checkpoint.
5479 : *
5480 : * We don't check full_page_writes here because that logic is included
5481 : * when we call XLogInsert() since the value changes dynamically.
5482 : */
5483 3203704 : if (XLogHintBitIsNeeded() &&
5484 1600758 : (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT))
5485 : {
5486 : /*
5487 : * If we must not write WAL, due to a relfilelocator-specific
5488 : * condition or being in recovery, don't dirty the page. We can
5489 : * set the hint, just not dirty the page as a result so the hint
5490 : * is lost when we evict the page or shutdown.
5491 : *
5492 : * See src/backend/storage/page/README for longer discussion.
5493 : */
5494 1724914 : if (RecoveryInProgress() ||
5495 124220 : RelFileLocatorSkippingWAL(BufTagGetRelFileLocator(&bufHdr->tag)))
5496 1482160 : return;
5497 :
5498 : /*
5499 : * If the block is already dirty because we either made a change
5500 : * or set a hint already, then we don't need to write a full page
5501 : * image. Note that aggressive cleaning of blocks dirtied by hint
5502 : * bit setting would increase the call rate. Bulk setting of hint
5503 : * bits would reduce the call rate...
5504 : *
5505 : * We must issue the WAL record before we mark the buffer dirty.
5506 : * Otherwise we might write the page before we write the WAL. That
5507 : * causes a race condition, since a checkpoint might occur between
5508 : * writing the WAL record and marking the buffer dirty. We solve
5509 : * that with a kluge, but one that is already in use during
5510 : * transaction commit to prevent race conditions. Basically, we
5511 : * simply prevent the checkpoint WAL record from being written
5512 : * until we have marked the buffer dirty. We don't start the
5513 : * checkpoint flush until we have marked dirty, so our checkpoint
5514 : * must flush the change to disk successfully or the checkpoint
5515 : * never gets written, so crash recovery will fix.
5516 : *
5517 : * It's possible we may enter here without an xid, so it is
5518 : * essential that CreateCheckPoint waits for virtual transactions
5519 : * rather than full transactionids.
5520 : */
5521 : Assert((MyProc->delayChkptFlags & DELAY_CHKPT_START) == 0);
5522 118534 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
5523 118534 : delayChkptFlags = true;
5524 118534 : lsn = XLogSaveBufferForHint(buffer, buffer_std);
5525 : }
5526 :
5527 120786 : buf_state = LockBufHdr(bufHdr);
5528 :
5529 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5530 :
5531 120786 : if (!(buf_state & BM_DIRTY))
5532 : {
5533 120694 : dirtied = true; /* Means "will be dirtied by this action" */
5534 :
5535 : /*
5536 : * Set the page LSN if we wrote a backup block. We aren't supposed
5537 : * to set this when only holding a share lock but as long as we
5538 : * serialise it somehow we're OK. We choose to set LSN while
5539 : * holding the buffer header lock, which causes any reader of an
5540 : * LSN who holds only a share lock to also obtain a buffer header
5541 : * lock before using PageGetLSN(), which is enforced in
5542 : * BufferGetLSNAtomic().
5543 : *
5544 : * If checksums are enabled, you might think we should reset the
5545 : * checksum here. That will happen when the page is written
5546 : * sometime later in this checkpoint cycle.
5547 : */
5548 120694 : if (!XLogRecPtrIsInvalid(lsn))
5549 62618 : PageSetLSN(page, lsn);
5550 : }
5551 :
5552 120786 : buf_state |= BM_DIRTY | BM_JUST_DIRTIED;
5553 120786 : UnlockBufHdr(bufHdr, buf_state);
5554 :
5555 120786 : if (delayChkptFlags)
5556 118534 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
5557 :
5558 120786 : if (dirtied)
5559 : {
5560 120694 : pgBufferUsage.shared_blks_dirtied++;
5561 120694 : if (VacuumCostActive)
5562 3534 : VacuumCostBalance += VacuumCostPageDirty;
5563 : }
5564 : }
5565 : }
5566 :
5567 : /*
5568 : * Release buffer content locks for shared buffers.
5569 : *
5570 : * Used to clean up after errors.
5571 : *
5572 : * Currently, we can expect that lwlock.c's LWLockReleaseAll() took care
5573 : * of releasing buffer content locks per se; the only thing we need to deal
5574 : * with here is clearing any PIN_COUNT request that was in progress.
5575 : */
5576 : void
5577 104732 : UnlockBuffers(void)
5578 : {
5579 104732 : BufferDesc *buf = PinCountWaitBuf;
5580 :
5581 104732 : if (buf)
5582 : {
5583 : uint32 buf_state;
5584 :
5585 0 : buf_state = LockBufHdr(buf);
5586 :
5587 : /*
5588 : * Don't complain if flag bit not set; it could have been reset but we
5589 : * got a cancel/die interrupt before getting the signal.
5590 : */
5591 0 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5592 0 : buf->wait_backend_pgprocno == MyProcNumber)
5593 0 : buf_state &= ~BM_PIN_COUNT_WAITER;
5594 :
5595 0 : UnlockBufHdr(buf, buf_state);
5596 :
5597 0 : PinCountWaitBuf = NULL;
5598 : }
5599 104732 : }
5600 :
5601 : /*
5602 : * Acquire or release the content_lock for the buffer.
5603 : */
5604 : void
5605 349181306 : LockBuffer(Buffer buffer, int mode)
5606 : {
5607 : BufferDesc *buf;
5608 :
5609 : Assert(BufferIsPinned(buffer));
5610 349181306 : if (BufferIsLocal(buffer))
5611 19800538 : return; /* local buffers need no lock */
5612 :
5613 329380768 : buf = GetBufferDescriptor(buffer - 1);
5614 :
5615 329380768 : if (mode == BUFFER_LOCK_UNLOCK)
5616 166229640 : LWLockRelease(BufferDescriptorGetContentLock(buf));
5617 163151128 : else if (mode == BUFFER_LOCK_SHARE)
5618 115662124 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_SHARED);
5619 47489004 : else if (mode == BUFFER_LOCK_EXCLUSIVE)
5620 47489004 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_EXCLUSIVE);
5621 : else
5622 0 : elog(ERROR, "unrecognized buffer lock mode: %d", mode);
5623 : }
5624 :
5625 : /*
5626 : * Acquire the content_lock for the buffer, but only if we don't have to wait.
5627 : *
5628 : * This assumes the caller wants BUFFER_LOCK_EXCLUSIVE mode.
5629 : */
5630 : bool
5631 3213200 : ConditionalLockBuffer(Buffer buffer)
5632 : {
5633 : BufferDesc *buf;
5634 :
5635 : Assert(BufferIsPinned(buffer));
5636 3213200 : if (BufferIsLocal(buffer))
5637 129436 : return true; /* act as though we got it */
5638 :
5639 3083764 : buf = GetBufferDescriptor(buffer - 1);
5640 :
5641 3083764 : return LWLockConditionalAcquire(BufferDescriptorGetContentLock(buf),
5642 : LW_EXCLUSIVE);
5643 : }
5644 :
5645 : /*
5646 : * Verify that this backend is pinning the buffer exactly once.
5647 : *
5648 : * NOTE: Like in BufferIsPinned(), what we check here is that *this* backend
5649 : * holds a pin on the buffer. We do not care whether some other backend does.
5650 : */
5651 : void
5652 5156660 : CheckBufferIsPinnedOnce(Buffer buffer)
5653 : {
5654 5156660 : if (BufferIsLocal(buffer))
5655 : {
5656 1578 : if (LocalRefCount[-buffer - 1] != 1)
5657 0 : elog(ERROR, "incorrect local pin count: %d",
5658 : LocalRefCount[-buffer - 1]);
5659 : }
5660 : else
5661 : {
5662 5155082 : if (GetPrivateRefCount(buffer) != 1)
5663 0 : elog(ERROR, "incorrect local pin count: %d",
5664 : GetPrivateRefCount(buffer));
5665 : }
5666 5156660 : }
5667 :
5668 : /*
5669 : * LockBufferForCleanup - lock a buffer in preparation for deleting items
5670 : *
5671 : * Items may be deleted from a disk page only when the caller (a) holds an
5672 : * exclusive lock on the buffer and (b) has observed that no other backend
5673 : * holds a pin on the buffer. If there is a pin, then the other backend
5674 : * might have a pointer into the buffer (for example, a heapscan reference
5675 : * to an item --- see README for more details). It's OK if a pin is added
5676 : * after the cleanup starts, however; the newly-arrived backend will be
5677 : * unable to look at the page until we release the exclusive lock.
5678 : *
5679 : * To implement this protocol, a would-be deleter must pin the buffer and
5680 : * then call LockBufferForCleanup(). LockBufferForCleanup() is similar to
5681 : * LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE), except that it loops until
5682 : * it has successfully observed pin count = 1.
5683 : */
5684 : void
5685 44066 : LockBufferForCleanup(Buffer buffer)
5686 : {
5687 : BufferDesc *bufHdr;
5688 44066 : TimestampTz waitStart = 0;
5689 44066 : bool waiting = false;
5690 44066 : bool logged_recovery_conflict = false;
5691 :
5692 : Assert(BufferIsPinned(buffer));
5693 : Assert(PinCountWaitBuf == NULL);
5694 :
5695 44066 : CheckBufferIsPinnedOnce(buffer);
5696 :
5697 : /*
5698 : * We do not yet need to be worried about in-progress AIOs holding a pin,
5699 : * as we, so far, only support doing reads via AIO and this function can
5700 : * only be called once the buffer is valid (i.e. no read can be in
5701 : * flight).
5702 : */
5703 :
5704 : /* Nobody else to wait for */
5705 44066 : if (BufferIsLocal(buffer))
5706 32 : return;
5707 :
5708 44034 : bufHdr = GetBufferDescriptor(buffer - 1);
5709 :
5710 : for (;;)
5711 154 : {
5712 : uint32 buf_state;
5713 :
5714 : /* Try to acquire lock */
5715 44188 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
5716 44188 : buf_state = LockBufHdr(bufHdr);
5717 :
5718 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5719 44188 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
5720 : {
5721 : /* Successfully acquired exclusive lock with pincount 1 */
5722 44034 : UnlockBufHdr(bufHdr, buf_state);
5723 :
5724 : /*
5725 : * Emit the log message if recovery conflict on buffer pin was
5726 : * resolved but the startup process waited longer than
5727 : * deadlock_timeout for it.
5728 : */
5729 44034 : if (logged_recovery_conflict)
5730 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
5731 : waitStart, GetCurrentTimestamp(),
5732 : NULL, false);
5733 :
5734 44034 : if (waiting)
5735 : {
5736 : /* reset ps display to remove the suffix if we added one */
5737 4 : set_ps_display_remove_suffix();
5738 4 : waiting = false;
5739 : }
5740 44034 : return;
5741 : }
5742 : /* Failed, so mark myself as waiting for pincount 1 */
5743 154 : if (buf_state & BM_PIN_COUNT_WAITER)
5744 : {
5745 0 : UnlockBufHdr(bufHdr, buf_state);
5746 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5747 0 : elog(ERROR, "multiple backends attempting to wait for pincount 1");
5748 : }
5749 154 : bufHdr->wait_backend_pgprocno = MyProcNumber;
5750 154 : PinCountWaitBuf = bufHdr;
5751 154 : buf_state |= BM_PIN_COUNT_WAITER;
5752 154 : UnlockBufHdr(bufHdr, buf_state);
5753 154 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5754 :
5755 : /* Wait to be signaled by UnpinBuffer() */
5756 154 : if (InHotStandby)
5757 : {
5758 20 : if (!waiting)
5759 : {
5760 : /* adjust the process title to indicate that it's waiting */
5761 4 : set_ps_display_suffix("waiting");
5762 4 : waiting = true;
5763 : }
5764 :
5765 : /*
5766 : * Emit the log message if the startup process is waiting longer
5767 : * than deadlock_timeout for recovery conflict on buffer pin.
5768 : *
5769 : * Skip this if first time through because the startup process has
5770 : * not started waiting yet in this case. So, the wait start
5771 : * timestamp is set after this logic.
5772 : */
5773 20 : if (waitStart != 0 && !logged_recovery_conflict)
5774 : {
5775 6 : TimestampTz now = GetCurrentTimestamp();
5776 :
5777 6 : if (TimestampDifferenceExceeds(waitStart, now,
5778 : DeadlockTimeout))
5779 : {
5780 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
5781 : waitStart, now, NULL, true);
5782 4 : logged_recovery_conflict = true;
5783 : }
5784 : }
5785 :
5786 : /*
5787 : * Set the wait start timestamp if logging is enabled and first
5788 : * time through.
5789 : */
5790 20 : if (log_recovery_conflict_waits && waitStart == 0)
5791 4 : waitStart = GetCurrentTimestamp();
5792 :
5793 : /* Publish the bufid that Startup process waits on */
5794 20 : SetStartupBufferPinWaitBufId(buffer - 1);
5795 : /* Set alarm and then wait to be signaled by UnpinBuffer() */
5796 20 : ResolveRecoveryConflictWithBufferPin();
5797 : /* Reset the published bufid */
5798 20 : SetStartupBufferPinWaitBufId(-1);
5799 : }
5800 : else
5801 134 : ProcWaitForSignal(WAIT_EVENT_BUFFER_PIN);
5802 :
5803 : /*
5804 : * Remove flag marking us as waiter. Normally this will not be set
5805 : * anymore, but ProcWaitForSignal() can return for other signals as
5806 : * well. We take care to only reset the flag if we're the waiter, as
5807 : * theoretically another backend could have started waiting. That's
5808 : * impossible with the current usages due to table level locking, but
5809 : * better be safe.
5810 : */
5811 154 : buf_state = LockBufHdr(bufHdr);
5812 154 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5813 16 : bufHdr->wait_backend_pgprocno == MyProcNumber)
5814 16 : buf_state &= ~BM_PIN_COUNT_WAITER;
5815 154 : UnlockBufHdr(bufHdr, buf_state);
5816 :
5817 154 : PinCountWaitBuf = NULL;
5818 : /* Loop back and try again */
5819 : }
5820 : }
5821 :
5822 : /*
5823 : * Check called from ProcessRecoveryConflictInterrupts() when Startup process
5824 : * requests cancellation of all pin holders that are blocking it.
5825 : */
5826 : bool
5827 8 : HoldingBufferPinThatDelaysRecovery(void)
5828 : {
5829 8 : int bufid = GetStartupBufferPinWaitBufId();
5830 :
5831 : /*
5832 : * If we get woken slowly then it's possible that the Startup process was
5833 : * already woken by other backends before we got here. Also possible that
5834 : * we get here by multiple interrupts or interrupts at inappropriate
5835 : * times, so make sure we do nothing if the bufid is not set.
5836 : */
5837 8 : if (bufid < 0)
5838 4 : return false;
5839 :
5840 4 : if (GetPrivateRefCount(bufid + 1) > 0)
5841 4 : return true;
5842 :
5843 0 : return false;
5844 : }
5845 :
5846 : /*
5847 : * ConditionalLockBufferForCleanup - as above, but don't wait to get the lock
5848 : *
5849 : * We won't loop, but just check once to see if the pin count is OK. If
5850 : * not, return false with no lock held.
5851 : */
5852 : bool
5853 1382132 : ConditionalLockBufferForCleanup(Buffer buffer)
5854 : {
5855 : BufferDesc *bufHdr;
5856 : uint32 buf_state,
5857 : refcount;
5858 :
5859 : Assert(BufferIsValid(buffer));
5860 :
5861 : /* see AIO related comment in LockBufferForCleanup() */
5862 :
5863 1382132 : if (BufferIsLocal(buffer))
5864 : {
5865 1608 : refcount = LocalRefCount[-buffer - 1];
5866 : /* There should be exactly one pin */
5867 : Assert(refcount > 0);
5868 1608 : if (refcount != 1)
5869 42 : return false;
5870 : /* Nobody else to wait for */
5871 1566 : return true;
5872 : }
5873 :
5874 : /* There should be exactly one local pin */
5875 1380524 : refcount = GetPrivateRefCount(buffer);
5876 : Assert(refcount);
5877 1380524 : if (refcount != 1)
5878 488 : return false;
5879 :
5880 : /* Try to acquire lock */
5881 1380036 : if (!ConditionalLockBuffer(buffer))
5882 168 : return false;
5883 :
5884 1379868 : bufHdr = GetBufferDescriptor(buffer - 1);
5885 1379868 : buf_state = LockBufHdr(bufHdr);
5886 1379868 : refcount = BUF_STATE_GET_REFCOUNT(buf_state);
5887 :
5888 : Assert(refcount > 0);
5889 1379868 : if (refcount == 1)
5890 : {
5891 : /* Successfully acquired exclusive lock with pincount 1 */
5892 1379322 : UnlockBufHdr(bufHdr, buf_state);
5893 1379322 : return true;
5894 : }
5895 :
5896 : /* Failed, so release the lock */
5897 546 : UnlockBufHdr(bufHdr, buf_state);
5898 546 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5899 546 : return false;
5900 : }
5901 :
5902 : /*
5903 : * IsBufferCleanupOK - as above, but we already have the lock
5904 : *
5905 : * Check whether it's OK to perform cleanup on a buffer we've already
5906 : * locked. If we observe that the pin count is 1, our exclusive lock
5907 : * happens to be a cleanup lock, and we can proceed with anything that
5908 : * would have been allowable had we sought a cleanup lock originally.
5909 : */
5910 : bool
5911 3290 : IsBufferCleanupOK(Buffer buffer)
5912 : {
5913 : BufferDesc *bufHdr;
5914 : uint32 buf_state;
5915 :
5916 : Assert(BufferIsValid(buffer));
5917 :
5918 : /* see AIO related comment in LockBufferForCleanup() */
5919 :
5920 3290 : if (BufferIsLocal(buffer))
5921 : {
5922 : /* There should be exactly one pin */
5923 0 : if (LocalRefCount[-buffer - 1] != 1)
5924 0 : return false;
5925 : /* Nobody else to wait for */
5926 0 : return true;
5927 : }
5928 :
5929 : /* There should be exactly one local pin */
5930 3290 : if (GetPrivateRefCount(buffer) != 1)
5931 0 : return false;
5932 :
5933 3290 : bufHdr = GetBufferDescriptor(buffer - 1);
5934 :
5935 : /* caller must hold exclusive lock on buffer */
5936 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
5937 :
5938 3290 : buf_state = LockBufHdr(bufHdr);
5939 :
5940 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5941 3290 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
5942 : {
5943 : /* pincount is OK. */
5944 3290 : UnlockBufHdr(bufHdr, buf_state);
5945 3290 : return true;
5946 : }
5947 :
5948 0 : UnlockBufHdr(bufHdr, buf_state);
5949 0 : return false;
5950 : }
5951 :
5952 :
5953 : /*
5954 : * Functions for buffer I/O handling
5955 : *
5956 : * Also note that these are used only for shared buffers, not local ones.
5957 : */
5958 :
5959 : /*
5960 : * WaitIO -- Block until the IO_IN_PROGRESS flag on 'buf' is cleared.
5961 : */
5962 : static void
5963 13224 : WaitIO(BufferDesc *buf)
5964 : {
5965 13224 : ConditionVariable *cv = BufferDescriptorGetIOCV(buf);
5966 :
5967 13224 : ConditionVariablePrepareToSleep(cv);
5968 : for (;;)
5969 13146 : {
5970 : uint32 buf_state;
5971 : PgAioWaitRef iow;
5972 :
5973 : /*
5974 : * It may not be necessary to acquire the spinlock to check the flag
5975 : * here, but since this test is essential for correctness, we'd better
5976 : * play it safe.
5977 : */
5978 26370 : buf_state = LockBufHdr(buf);
5979 :
5980 : /*
5981 : * Copy the wait reference while holding the spinlock. This protects
5982 : * against a concurrent TerminateBufferIO() in another backend from
5983 : * clearing the wref while it's being read.
5984 : */
5985 26370 : iow = buf->io_wref;
5986 26370 : UnlockBufHdr(buf, buf_state);
5987 :
5988 : /* no IO in progress, we don't need to wait */
5989 26370 : if (!(buf_state & BM_IO_IN_PROGRESS))
5990 13224 : break;
5991 :
5992 : /*
5993 : * The buffer has asynchronous IO in progress, wait for it to
5994 : * complete.
5995 : */
5996 13146 : if (pgaio_wref_valid(&iow))
5997 : {
5998 11030 : pgaio_wref_wait(&iow);
5999 :
6000 : /*
6001 : * The AIO subsystem internally uses condition variables and thus
6002 : * might remove this backend from the BufferDesc's CV. While that
6003 : * wouldn't cause a correctness issue (the first CV sleep just
6004 : * immediately returns if not already registered), it seems worth
6005 : * avoiding unnecessary loop iterations, given that we take care
6006 : * to do so at the start of the function.
6007 : */
6008 11030 : ConditionVariablePrepareToSleep(cv);
6009 11030 : continue;
6010 : }
6011 :
6012 : /* wait on BufferDesc->cv, e.g. for concurrent synchronous IO */
6013 2116 : ConditionVariableSleep(cv, WAIT_EVENT_BUFFER_IO);
6014 : }
6015 13224 : ConditionVariableCancelSleep();
6016 13224 : }
6017 :
6018 : /*
6019 : * StartBufferIO: begin I/O on this buffer
6020 : * (Assumptions)
6021 : * My process is executing no IO on this buffer
6022 : * The buffer is Pinned
6023 : *
6024 : * In some scenarios multiple backends could attempt the same I/O operation
6025 : * concurrently. If someone else has already started I/O on this buffer then
6026 : * we will wait for completion of the IO using WaitIO().
6027 : *
6028 : * Input operations are only attempted on buffers that are not BM_VALID,
6029 : * and output operations only on buffers that are BM_VALID and BM_DIRTY,
6030 : * so we can always tell if the work is already done.
6031 : *
6032 : * Returns true if we successfully marked the buffer as I/O busy,
6033 : * false if someone else already did the work.
6034 : *
6035 : * If nowait is true, then we don't wait for an I/O to be finished by another
6036 : * backend. In that case, false indicates either that the I/O was already
6037 : * finished, or is still in progress. This is useful for callers that want to
6038 : * find out if they can perform the I/O as part of a larger operation, without
6039 : * waiting for the answer or distinguishing the reasons why not.
6040 : */
6041 : bool
6042 4937016 : StartBufferIO(BufferDesc *buf, bool forInput, bool nowait)
6043 : {
6044 : uint32 buf_state;
6045 :
6046 4937016 : ResourceOwnerEnlarge(CurrentResourceOwner);
6047 :
6048 : for (;;)
6049 : {
6050 4950236 : buf_state = LockBufHdr(buf);
6051 :
6052 4950236 : if (!(buf_state & BM_IO_IN_PROGRESS))
6053 4937008 : break;
6054 13228 : UnlockBufHdr(buf, buf_state);
6055 13228 : if (nowait)
6056 8 : return false;
6057 13220 : WaitIO(buf);
6058 : }
6059 :
6060 : /* Once we get here, there is definitely no I/O active on this buffer */
6061 :
6062 : /* Check if someone else already did the I/O */
6063 4937008 : if (forInput ? (buf_state & BM_VALID) : !(buf_state & BM_DIRTY))
6064 : {
6065 14220 : UnlockBufHdr(buf, buf_state);
6066 14220 : return false;
6067 : }
6068 :
6069 4922788 : buf_state |= BM_IO_IN_PROGRESS;
6070 4922788 : UnlockBufHdr(buf, buf_state);
6071 :
6072 4922788 : ResourceOwnerRememberBufferIO(CurrentResourceOwner,
6073 : BufferDescriptorGetBuffer(buf));
6074 :
6075 4922788 : return true;
6076 : }
6077 :
6078 : /*
6079 : * TerminateBufferIO: release a buffer we were doing I/O on
6080 : * (Assumptions)
6081 : * My process is executing IO for the buffer
6082 : * BM_IO_IN_PROGRESS bit is set for the buffer
6083 : * The buffer is Pinned
6084 : *
6085 : * If clear_dirty is true and BM_JUST_DIRTIED is not set, we clear the
6086 : * buffer's BM_DIRTY flag. This is appropriate when terminating a
6087 : * successful write. The check on BM_JUST_DIRTIED is necessary to avoid
6088 : * marking the buffer clean if it was re-dirtied while we were writing.
6089 : *
6090 : * set_flag_bits gets ORed into the buffer's flags. It must include
6091 : * BM_IO_ERROR in a failure case. For successful completion it could
6092 : * be 0, or BM_VALID if we just finished reading in the page.
6093 : *
6094 : * If forget_owner is true, we release the buffer I/O from the current
6095 : * resource owner. (forget_owner=false is used when the resource owner itself
6096 : * is being released)
6097 : */
6098 : void
6099 4647628 : TerminateBufferIO(BufferDesc *buf, bool clear_dirty, uint32 set_flag_bits,
6100 : bool forget_owner, bool release_aio)
6101 : {
6102 : uint32 buf_state;
6103 :
6104 4647628 : buf_state = LockBufHdr(buf);
6105 :
6106 : Assert(buf_state & BM_IO_IN_PROGRESS);
6107 4647628 : buf_state &= ~BM_IO_IN_PROGRESS;
6108 :
6109 : /* Clear earlier errors, if this IO failed, it'll be marked again */
6110 4647628 : buf_state &= ~BM_IO_ERROR;
6111 :
6112 4647628 : if (clear_dirty && !(buf_state & BM_JUST_DIRTIED))
6113 1122378 : buf_state &= ~(BM_DIRTY | BM_CHECKPOINT_NEEDED);
6114 :
6115 4647628 : if (release_aio)
6116 : {
6117 : /* release ownership by the AIO subsystem */
6118 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6119 2516494 : buf_state -= BUF_REFCOUNT_ONE;
6120 2516494 : pgaio_wref_clear(&buf->io_wref);
6121 : }
6122 :
6123 4647628 : buf_state |= set_flag_bits;
6124 4647628 : UnlockBufHdr(buf, buf_state);
6125 :
6126 4647628 : if (forget_owner)
6127 2131092 : ResourceOwnerForgetBufferIO(CurrentResourceOwner,
6128 : BufferDescriptorGetBuffer(buf));
6129 :
6130 4647628 : ConditionVariableBroadcast(BufferDescriptorGetIOCV(buf));
6131 :
6132 : /*
6133 : * Support LockBufferForCleanup()
6134 : *
6135 : * We may have just released the last pin other than the waiter's. In most
6136 : * cases, this backend holds another pin on the buffer. But, if, for
6137 : * example, this backend is completing an IO issued by another backend, it
6138 : * may be time to wake the waiter.
6139 : */
6140 4647628 : if (release_aio && (buf_state & BM_PIN_COUNT_WAITER))
6141 0 : WakePinCountWaiter(buf);
6142 4647628 : }
6143 :
6144 : /*
6145 : * AbortBufferIO: Clean up active buffer I/O after an error.
6146 : *
6147 : * All LWLocks we might have held have been released,
6148 : * but we haven't yet released buffer pins, so the buffer is still pinned.
6149 : *
6150 : * If I/O was in progress, we always set BM_IO_ERROR, even though it's
6151 : * possible the error condition wasn't related to the I/O.
6152 : *
6153 : * Note: this does not remove the buffer I/O from the resource owner.
6154 : * That's correct when we're releasing the whole resource owner, but
6155 : * beware if you use this in other contexts.
6156 : */
6157 : static void
6158 30 : AbortBufferIO(Buffer buffer)
6159 : {
6160 30 : BufferDesc *buf_hdr = GetBufferDescriptor(buffer - 1);
6161 : uint32 buf_state;
6162 :
6163 30 : buf_state = LockBufHdr(buf_hdr);
6164 : Assert(buf_state & (BM_IO_IN_PROGRESS | BM_TAG_VALID));
6165 :
6166 30 : if (!(buf_state & BM_VALID))
6167 : {
6168 : Assert(!(buf_state & BM_DIRTY));
6169 30 : UnlockBufHdr(buf_hdr, buf_state);
6170 : }
6171 : else
6172 : {
6173 : Assert(buf_state & BM_DIRTY);
6174 0 : UnlockBufHdr(buf_hdr, buf_state);
6175 :
6176 : /* Issue notice if this is not the first failure... */
6177 0 : if (buf_state & BM_IO_ERROR)
6178 : {
6179 : /* Buffer is pinned, so we can read tag without spinlock */
6180 0 : ereport(WARNING,
6181 : (errcode(ERRCODE_IO_ERROR),
6182 : errmsg("could not write block %u of %s",
6183 : buf_hdr->tag.blockNum,
6184 : relpathperm(BufTagGetRelFileLocator(&buf_hdr->tag),
6185 : BufTagGetForkNum(&buf_hdr->tag)).str),
6186 : errdetail("Multiple failures --- write error might be permanent.")));
6187 : }
6188 : }
6189 :
6190 30 : TerminateBufferIO(buf_hdr, false, BM_IO_ERROR, false, false);
6191 30 : }
6192 :
6193 : /*
6194 : * Error context callback for errors occurring during shared buffer writes.
6195 : */
6196 : static void
6197 92 : shared_buffer_write_error_callback(void *arg)
6198 : {
6199 92 : BufferDesc *bufHdr = (BufferDesc *) arg;
6200 :
6201 : /* Buffer is pinned, so we can read the tag without locking the spinlock */
6202 92 : if (bufHdr != NULL)
6203 184 : errcontext("writing block %u of relation \"%s\"",
6204 : bufHdr->tag.blockNum,
6205 92 : relpathperm(BufTagGetRelFileLocator(&bufHdr->tag),
6206 : BufTagGetForkNum(&bufHdr->tag)).str);
6207 92 : }
6208 :
6209 : /*
6210 : * Error context callback for errors occurring during local buffer writes.
6211 : */
6212 : static void
6213 0 : local_buffer_write_error_callback(void *arg)
6214 : {
6215 0 : BufferDesc *bufHdr = (BufferDesc *) arg;
6216 :
6217 0 : if (bufHdr != NULL)
6218 0 : errcontext("writing block %u of relation \"%s\"",
6219 : bufHdr->tag.blockNum,
6220 0 : relpathbackend(BufTagGetRelFileLocator(&bufHdr->tag),
6221 : MyProcNumber,
6222 : BufTagGetForkNum(&bufHdr->tag)).str);
6223 0 : }
6224 :
6225 : /*
6226 : * RelFileLocator qsort/bsearch comparator; see RelFileLocatorEquals.
6227 : */
6228 : static int
6229 19416502 : rlocator_comparator(const void *p1, const void *p2)
6230 : {
6231 19416502 : RelFileLocator n1 = *(const RelFileLocator *) p1;
6232 19416502 : RelFileLocator n2 = *(const RelFileLocator *) p2;
6233 :
6234 19416502 : if (n1.relNumber < n2.relNumber)
6235 19343702 : return -1;
6236 72800 : else if (n1.relNumber > n2.relNumber)
6237 69218 : return 1;
6238 :
6239 3582 : if (n1.dbOid < n2.dbOid)
6240 0 : return -1;
6241 3582 : else if (n1.dbOid > n2.dbOid)
6242 0 : return 1;
6243 :
6244 3582 : if (n1.spcOid < n2.spcOid)
6245 0 : return -1;
6246 3582 : else if (n1.spcOid > n2.spcOid)
6247 0 : return 1;
6248 : else
6249 3582 : return 0;
6250 : }
6251 :
6252 : /*
6253 : * Lock buffer header - set BM_LOCKED in buffer state.
6254 : */
6255 : uint32
6256 64186466 : LockBufHdr(BufferDesc *desc)
6257 : {
6258 : SpinDelayStatus delayStatus;
6259 : uint32 old_buf_state;
6260 :
6261 : Assert(!BufferIsLocal(BufferDescriptorGetBuffer(desc)));
6262 :
6263 64186466 : init_local_spin_delay(&delayStatus);
6264 :
6265 : while (true)
6266 : {
6267 : /* set BM_LOCKED flag */
6268 64196486 : old_buf_state = pg_atomic_fetch_or_u32(&desc->state, BM_LOCKED);
6269 : /* if it wasn't set before we're OK */
6270 64196486 : if (!(old_buf_state & BM_LOCKED))
6271 64186466 : break;
6272 10020 : perform_spin_delay(&delayStatus);
6273 : }
6274 64186466 : finish_spin_delay(&delayStatus);
6275 64186466 : return old_buf_state | BM_LOCKED;
6276 : }
6277 :
6278 : /*
6279 : * Wait until the BM_LOCKED flag isn't set anymore and return the buffer's
6280 : * state at that point.
6281 : *
6282 : * Obviously the buffer could be locked by the time the value is returned, so
6283 : * this is primarily useful in CAS style loops.
6284 : */
6285 : pg_noinline uint32
6286 3114 : WaitBufHdrUnlocked(BufferDesc *buf)
6287 : {
6288 : SpinDelayStatus delayStatus;
6289 : uint32 buf_state;
6290 :
6291 3114 : init_local_spin_delay(&delayStatus);
6292 :
6293 3114 : buf_state = pg_atomic_read_u32(&buf->state);
6294 :
6295 116480 : while (buf_state & BM_LOCKED)
6296 : {
6297 113366 : perform_spin_delay(&delayStatus);
6298 113366 : buf_state = pg_atomic_read_u32(&buf->state);
6299 : }
6300 :
6301 3114 : finish_spin_delay(&delayStatus);
6302 :
6303 3114 : return buf_state;
6304 : }
6305 :
6306 : /*
6307 : * BufferTag comparator.
6308 : */
6309 : static inline int
6310 0 : buffertag_comparator(const BufferTag *ba, const BufferTag *bb)
6311 : {
6312 : int ret;
6313 : RelFileLocator rlocatora;
6314 : RelFileLocator rlocatorb;
6315 :
6316 0 : rlocatora = BufTagGetRelFileLocator(ba);
6317 0 : rlocatorb = BufTagGetRelFileLocator(bb);
6318 :
6319 0 : ret = rlocator_comparator(&rlocatora, &rlocatorb);
6320 :
6321 0 : if (ret != 0)
6322 0 : return ret;
6323 :
6324 0 : if (BufTagGetForkNum(ba) < BufTagGetForkNum(bb))
6325 0 : return -1;
6326 0 : if (BufTagGetForkNum(ba) > BufTagGetForkNum(bb))
6327 0 : return 1;
6328 :
6329 0 : if (ba->blockNum < bb->blockNum)
6330 0 : return -1;
6331 0 : if (ba->blockNum > bb->blockNum)
6332 0 : return 1;
6333 :
6334 0 : return 0;
6335 : }
6336 :
6337 : /*
6338 : * Comparator determining the writeout order in a checkpoint.
6339 : *
6340 : * It is important that tablespaces are compared first, the logic balancing
6341 : * writes between tablespaces relies on it.
6342 : */
6343 : static inline int
6344 5998910 : ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b)
6345 : {
6346 : /* compare tablespace */
6347 5998910 : if (a->tsId < b->tsId)
6348 14530 : return -1;
6349 5984380 : else if (a->tsId > b->tsId)
6350 52612 : return 1;
6351 : /* compare relation */
6352 5931768 : if (a->relNumber < b->relNumber)
6353 1672144 : return -1;
6354 4259624 : else if (a->relNumber > b->relNumber)
6355 1628164 : return 1;
6356 : /* compare fork */
6357 2631460 : else if (a->forkNum < b->forkNum)
6358 109832 : return -1;
6359 2521628 : else if (a->forkNum > b->forkNum)
6360 118784 : return 1;
6361 : /* compare block number */
6362 2402844 : else if (a->blockNum < b->blockNum)
6363 1178860 : return -1;
6364 1223984 : else if (a->blockNum > b->blockNum)
6365 1150814 : return 1;
6366 : /* equal page IDs are unlikely, but not impossible */
6367 73170 : return 0;
6368 : }
6369 :
6370 : /*
6371 : * Comparator for a Min-Heap over the per-tablespace checkpoint completion
6372 : * progress.
6373 : */
6374 : static int
6375 489040 : ts_ckpt_progress_comparator(Datum a, Datum b, void *arg)
6376 : {
6377 489040 : CkptTsStatus *sa = (CkptTsStatus *) DatumGetPointer(a);
6378 489040 : CkptTsStatus *sb = (CkptTsStatus *) DatumGetPointer(b);
6379 :
6380 : /* we want a min-heap, so return 1 for the a < b */
6381 489040 : if (sa->progress < sb->progress)
6382 439340 : return 1;
6383 49700 : else if (sa->progress == sb->progress)
6384 1658 : return 0;
6385 : else
6386 48042 : return -1;
6387 : }
6388 :
6389 : /*
6390 : * Initialize a writeback context, discarding potential previous state.
6391 : *
6392 : * *max_pending is a pointer instead of an immediate value, so the coalesce
6393 : * limits can easily changed by the GUC mechanism, and so calling code does
6394 : * not have to check the current configuration. A value of 0 means that no
6395 : * writeback control will be performed.
6396 : */
6397 : void
6398 5396 : WritebackContextInit(WritebackContext *context, int *max_pending)
6399 : {
6400 : Assert(*max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
6401 :
6402 5396 : context->max_pending = max_pending;
6403 5396 : context->nr_pending = 0;
6404 5396 : }
6405 :
6406 : /*
6407 : * Add buffer to list of pending writeback requests.
6408 : */
6409 : void
6410 1112056 : ScheduleBufferTagForWriteback(WritebackContext *wb_context, IOContext io_context,
6411 : BufferTag *tag)
6412 : {
6413 : PendingWriteback *pending;
6414 :
6415 : /*
6416 : * As pg_flush_data() doesn't do anything with fsync disabled, there's no
6417 : * point in tracking in that case.
6418 : */
6419 1112056 : if (io_direct_flags & IO_DIRECT_DATA ||
6420 1111042 : !enableFsync)
6421 1112052 : return;
6422 :
6423 : /*
6424 : * Add buffer to the pending writeback array, unless writeback control is
6425 : * disabled.
6426 : */
6427 4 : if (*wb_context->max_pending > 0)
6428 : {
6429 : Assert(*wb_context->max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
6430 :
6431 0 : pending = &wb_context->pending_writebacks[wb_context->nr_pending++];
6432 :
6433 0 : pending->tag = *tag;
6434 : }
6435 :
6436 : /*
6437 : * Perform pending flushes if the writeback limit is exceeded. This
6438 : * includes the case where previously an item has been added, but control
6439 : * is now disabled.
6440 : */
6441 4 : if (wb_context->nr_pending >= *wb_context->max_pending)
6442 4 : IssuePendingWritebacks(wb_context, io_context);
6443 : }
6444 :
6445 : #define ST_SORT sort_pending_writebacks
6446 : #define ST_ELEMENT_TYPE PendingWriteback
6447 : #define ST_COMPARE(a, b) buffertag_comparator(&a->tag, &b->tag)
6448 : #define ST_SCOPE static
6449 : #define ST_DEFINE
6450 : #include "lib/sort_template.h"
6451 :
6452 : /*
6453 : * Issue all pending writeback requests, previously scheduled with
6454 : * ScheduleBufferTagForWriteback, to the OS.
6455 : *
6456 : * Because this is only used to improve the OSs IO scheduling we try to never
6457 : * error out - it's just a hint.
6458 : */
6459 : void
6460 2126 : IssuePendingWritebacks(WritebackContext *wb_context, IOContext io_context)
6461 : {
6462 : instr_time io_start;
6463 : int i;
6464 :
6465 2126 : if (wb_context->nr_pending == 0)
6466 2126 : return;
6467 :
6468 : /*
6469 : * Executing the writes in-order can make them a lot faster, and allows to
6470 : * merge writeback requests to consecutive blocks into larger writebacks.
6471 : */
6472 0 : sort_pending_writebacks(wb_context->pending_writebacks,
6473 0 : wb_context->nr_pending);
6474 :
6475 0 : io_start = pgstat_prepare_io_time(track_io_timing);
6476 :
6477 : /*
6478 : * Coalesce neighbouring writes, but nothing else. For that we iterate
6479 : * through the, now sorted, array of pending flushes, and look forward to
6480 : * find all neighbouring (or identical) writes.
6481 : */
6482 0 : for (i = 0; i < wb_context->nr_pending; i++)
6483 : {
6484 : PendingWriteback *cur;
6485 : PendingWriteback *next;
6486 : SMgrRelation reln;
6487 : int ahead;
6488 : BufferTag tag;
6489 : RelFileLocator currlocator;
6490 0 : Size nblocks = 1;
6491 :
6492 0 : cur = &wb_context->pending_writebacks[i];
6493 0 : tag = cur->tag;
6494 0 : currlocator = BufTagGetRelFileLocator(&tag);
6495 :
6496 : /*
6497 : * Peek ahead, into following writeback requests, to see if they can
6498 : * be combined with the current one.
6499 : */
6500 0 : for (ahead = 0; i + ahead + 1 < wb_context->nr_pending; ahead++)
6501 : {
6502 :
6503 0 : next = &wb_context->pending_writebacks[i + ahead + 1];
6504 :
6505 : /* different file, stop */
6506 0 : if (!RelFileLocatorEquals(currlocator,
6507 0 : BufTagGetRelFileLocator(&next->tag)) ||
6508 0 : BufTagGetForkNum(&cur->tag) != BufTagGetForkNum(&next->tag))
6509 : break;
6510 :
6511 : /* ok, block queued twice, skip */
6512 0 : if (cur->tag.blockNum == next->tag.blockNum)
6513 0 : continue;
6514 :
6515 : /* only merge consecutive writes */
6516 0 : if (cur->tag.blockNum + 1 != next->tag.blockNum)
6517 0 : break;
6518 :
6519 0 : nblocks++;
6520 0 : cur = next;
6521 : }
6522 :
6523 0 : i += ahead;
6524 :
6525 : /* and finally tell the kernel to write the data to storage */
6526 0 : reln = smgropen(currlocator, INVALID_PROC_NUMBER);
6527 0 : smgrwriteback(reln, BufTagGetForkNum(&tag), tag.blockNum, nblocks);
6528 : }
6529 :
6530 : /*
6531 : * Assume that writeback requests are only issued for buffers containing
6532 : * blocks of permanent relations.
6533 : */
6534 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
6535 0 : IOOP_WRITEBACK, io_start, wb_context->nr_pending, 0);
6536 :
6537 0 : wb_context->nr_pending = 0;
6538 : }
6539 :
6540 : /* ResourceOwner callbacks */
6541 :
6542 : static void
6543 30 : ResOwnerReleaseBufferIO(Datum res)
6544 : {
6545 30 : Buffer buffer = DatumGetInt32(res);
6546 :
6547 30 : AbortBufferIO(buffer);
6548 30 : }
6549 :
6550 : static char *
6551 0 : ResOwnerPrintBufferIO(Datum res)
6552 : {
6553 0 : Buffer buffer = DatumGetInt32(res);
6554 :
6555 0 : return psprintf("lost track of buffer IO on buffer %d", buffer);
6556 : }
6557 :
6558 : static void
6559 15154 : ResOwnerReleaseBufferPin(Datum res)
6560 : {
6561 15154 : Buffer buffer = DatumGetInt32(res);
6562 :
6563 : /* Like ReleaseBuffer, but don't call ResourceOwnerForgetBuffer */
6564 15154 : if (!BufferIsValid(buffer))
6565 0 : elog(ERROR, "bad buffer ID: %d", buffer);
6566 :
6567 15154 : if (BufferIsLocal(buffer))
6568 6066 : UnpinLocalBufferNoOwner(buffer);
6569 : else
6570 9088 : UnpinBufferNoOwner(GetBufferDescriptor(buffer - 1));
6571 15154 : }
6572 :
6573 : static char *
6574 0 : ResOwnerPrintBufferPin(Datum res)
6575 : {
6576 0 : return DebugPrintBufferRefcount(DatumGetInt32(res));
6577 : }
6578 :
6579 : /*
6580 : * Helper function to evict unpinned buffer whose buffer header lock is
6581 : * already acquired.
6582 : */
6583 : static bool
6584 4278 : EvictUnpinnedBufferInternal(BufferDesc *desc, bool *buffer_flushed)
6585 : {
6586 : uint32 buf_state;
6587 : bool result;
6588 :
6589 4278 : *buffer_flushed = false;
6590 :
6591 4278 : buf_state = pg_atomic_read_u32(&(desc->state));
6592 : Assert(buf_state & BM_LOCKED);
6593 :
6594 4278 : if ((buf_state & BM_VALID) == 0)
6595 : {
6596 0 : UnlockBufHdr(desc, buf_state);
6597 0 : return false;
6598 : }
6599 :
6600 : /* Check that it's not pinned already. */
6601 4278 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
6602 : {
6603 0 : UnlockBufHdr(desc, buf_state);
6604 0 : return false;
6605 : }
6606 :
6607 4278 : PinBuffer_Locked(desc); /* releases spinlock */
6608 :
6609 : /* If it was dirty, try to clean it once. */
6610 4278 : if (buf_state & BM_DIRTY)
6611 : {
6612 1940 : FlushUnlockedBuffer(desc, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
6613 1940 : *buffer_flushed = true;
6614 : }
6615 :
6616 : /* This will return false if it becomes dirty or someone else pins it. */
6617 4278 : result = InvalidateVictimBuffer(desc);
6618 :
6619 4278 : UnpinBuffer(desc);
6620 :
6621 4278 : return result;
6622 : }
6623 :
6624 : /*
6625 : * Try to evict the current block in a shared buffer.
6626 : *
6627 : * This function is intended for testing/development use only!
6628 : *
6629 : * To succeed, the buffer must not be pinned on entry, so if the caller had a
6630 : * particular block in mind, it might already have been replaced by some other
6631 : * block by the time this function runs. It's also unpinned on return, so the
6632 : * buffer might be occupied again by the time control is returned, potentially
6633 : * even by the same block. This inherent raciness without other interlocking
6634 : * makes the function unsuitable for non-testing usage.
6635 : *
6636 : * *buffer_flushed is set to true if the buffer was dirty and has been
6637 : * flushed, false otherwise. However, *buffer_flushed=true does not
6638 : * necessarily mean that we flushed the buffer, it could have been flushed by
6639 : * someone else.
6640 : *
6641 : * Returns true if the buffer was valid and it has now been made invalid.
6642 : * Returns false if it wasn't valid, if it couldn't be evicted due to a pin,
6643 : * or if the buffer becomes dirty again while we're trying to write it out.
6644 : */
6645 : bool
6646 280 : EvictUnpinnedBuffer(Buffer buf, bool *buffer_flushed)
6647 : {
6648 : BufferDesc *desc;
6649 :
6650 : Assert(BufferIsValid(buf) && !BufferIsLocal(buf));
6651 :
6652 : /* Make sure we can pin the buffer. */
6653 280 : ResourceOwnerEnlarge(CurrentResourceOwner);
6654 280 : ReservePrivateRefCountEntry();
6655 :
6656 280 : desc = GetBufferDescriptor(buf - 1);
6657 280 : LockBufHdr(desc);
6658 :
6659 280 : return EvictUnpinnedBufferInternal(desc, buffer_flushed);
6660 : }
6661 :
6662 : /*
6663 : * Try to evict all the shared buffers.
6664 : *
6665 : * This function is intended for testing/development use only! See
6666 : * EvictUnpinnedBuffer().
6667 : *
6668 : * The buffers_* parameters are mandatory and indicate the total count of
6669 : * buffers that:
6670 : * - buffers_evicted - were evicted
6671 : * - buffers_flushed - were flushed
6672 : * - buffers_skipped - could not be evicted
6673 : */
6674 : void
6675 2 : EvictAllUnpinnedBuffers(int32 *buffers_evicted, int32 *buffers_flushed,
6676 : int32 *buffers_skipped)
6677 : {
6678 2 : *buffers_evicted = 0;
6679 2 : *buffers_skipped = 0;
6680 2 : *buffers_flushed = 0;
6681 :
6682 32770 : for (int buf = 1; buf <= NBuffers; buf++)
6683 : {
6684 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
6685 : uint32 buf_state;
6686 : bool buffer_flushed;
6687 :
6688 32768 : buf_state = pg_atomic_read_u32(&desc->state);
6689 32768 : if (!(buf_state & BM_VALID))
6690 28770 : continue;
6691 :
6692 3998 : ResourceOwnerEnlarge(CurrentResourceOwner);
6693 3998 : ReservePrivateRefCountEntry();
6694 :
6695 3998 : LockBufHdr(desc);
6696 :
6697 3998 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
6698 3998 : (*buffers_evicted)++;
6699 : else
6700 0 : (*buffers_skipped)++;
6701 :
6702 3998 : if (buffer_flushed)
6703 1902 : (*buffers_flushed)++;
6704 : }
6705 2 : }
6706 :
6707 : /*
6708 : * Try to evict all the shared buffers containing provided relation's pages.
6709 : *
6710 : * This function is intended for testing/development use only! See
6711 : * EvictUnpinnedBuffer().
6712 : *
6713 : * The caller must hold at least AccessShareLock on the relation to prevent
6714 : * the relation from being dropped.
6715 : *
6716 : * The buffers_* parameters are mandatory and indicate the total count of
6717 : * buffers that:
6718 : * - buffers_evicted - were evicted
6719 : * - buffers_flushed - were flushed
6720 : * - buffers_skipped - could not be evicted
6721 : */
6722 : void
6723 2 : EvictRelUnpinnedBuffers(Relation rel, int32 *buffers_evicted,
6724 : int32 *buffers_flushed, int32 *buffers_skipped)
6725 : {
6726 : Assert(!RelationUsesLocalBuffers(rel));
6727 :
6728 2 : *buffers_skipped = 0;
6729 2 : *buffers_evicted = 0;
6730 2 : *buffers_flushed = 0;
6731 :
6732 32770 : for (int buf = 1; buf <= NBuffers; buf++)
6733 : {
6734 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
6735 32768 : uint32 buf_state = pg_atomic_read_u32(&(desc->state));
6736 : bool buffer_flushed;
6737 :
6738 : /* An unlocked precheck should be safe and saves some cycles. */
6739 32768 : if ((buf_state & BM_VALID) == 0 ||
6740 54 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
6741 32768 : continue;
6742 :
6743 : /* Make sure we can pin the buffer. */
6744 0 : ResourceOwnerEnlarge(CurrentResourceOwner);
6745 0 : ReservePrivateRefCountEntry();
6746 :
6747 0 : buf_state = LockBufHdr(desc);
6748 :
6749 : /* recheck, could have changed without the lock */
6750 0 : if ((buf_state & BM_VALID) == 0 ||
6751 0 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
6752 : {
6753 0 : UnlockBufHdr(desc, buf_state);
6754 0 : continue;
6755 : }
6756 :
6757 0 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
6758 0 : (*buffers_evicted)++;
6759 : else
6760 0 : (*buffers_skipped)++;
6761 :
6762 0 : if (buffer_flushed)
6763 0 : (*buffers_flushed)++;
6764 : }
6765 2 : }
6766 :
6767 : /*
6768 : * Generic implementation of the AIO handle staging callback for readv/writev
6769 : * on local/shared buffers.
6770 : *
6771 : * Each readv/writev can target multiple buffers. The buffers have already
6772 : * been registered with the IO handle.
6773 : *
6774 : * To make the IO ready for execution ("staging"), we need to ensure that the
6775 : * targeted buffers are in an appropriate state while the IO is ongoing. For
6776 : * that the AIO subsystem needs to have its own buffer pin, otherwise an error
6777 : * in this backend could lead to this backend's buffer pin being released as
6778 : * part of error handling, which in turn could lead to the buffer being
6779 : * replaced while IO is ongoing.
6780 : */
6781 : static pg_attribute_always_inline void
6782 2477342 : buffer_stage_common(PgAioHandle *ioh, bool is_write, bool is_temp)
6783 : {
6784 : uint64 *io_data;
6785 : uint8 handle_data_len;
6786 : PgAioWaitRef io_ref;
6787 2477342 : BufferTag first PG_USED_FOR_ASSERTS_ONLY = {0};
6788 :
6789 2477342 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
6790 :
6791 2477342 : pgaio_io_get_wref(ioh, &io_ref);
6792 :
6793 : /* iterate over all buffers affected by the vectored readv/writev */
6794 5285814 : for (int i = 0; i < handle_data_len; i++)
6795 : {
6796 2808472 : Buffer buffer = (Buffer) io_data[i];
6797 2808472 : BufferDesc *buf_hdr = is_temp ?
6798 16818 : GetLocalBufferDescriptor(-buffer - 1)
6799 2808472 : : GetBufferDescriptor(buffer - 1);
6800 : uint32 buf_state;
6801 :
6802 : /*
6803 : * Check that all the buffers are actually ones that could conceivably
6804 : * be done in one IO, i.e. are sequential. This is the last
6805 : * buffer-aware code before IO is actually executed and confusion
6806 : * about which buffers are targeted by IO can be hard to debug, making
6807 : * it worth doing extra-paranoid checks.
6808 : */
6809 2808472 : if (i == 0)
6810 2477342 : first = buf_hdr->tag;
6811 : else
6812 : {
6813 : Assert(buf_hdr->tag.relNumber == first.relNumber);
6814 : Assert(buf_hdr->tag.blockNum == first.blockNum + i);
6815 : }
6816 :
6817 2808472 : if (is_temp)
6818 16818 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
6819 : else
6820 2791654 : buf_state = LockBufHdr(buf_hdr);
6821 :
6822 : /* verify the buffer is in the expected state */
6823 : Assert(buf_state & BM_TAG_VALID);
6824 : if (is_write)
6825 : {
6826 : Assert(buf_state & BM_VALID);
6827 : Assert(buf_state & BM_DIRTY);
6828 : }
6829 : else
6830 : {
6831 : Assert(!(buf_state & BM_VALID));
6832 : Assert(!(buf_state & BM_DIRTY));
6833 : }
6834 :
6835 : /* temp buffers don't use BM_IO_IN_PROGRESS */
6836 2808472 : if (!is_temp)
6837 : Assert(buf_state & BM_IO_IN_PROGRESS);
6838 :
6839 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) >= 1);
6840 :
6841 : /*
6842 : * Reflect that the buffer is now owned by the AIO subsystem.
6843 : *
6844 : * For local buffers: This can't be done just via LocalRefCount, as
6845 : * one might initially think, as this backend could error out while
6846 : * AIO is still in progress, releasing all the pins by the backend
6847 : * itself.
6848 : *
6849 : * This pin is released again in TerminateBufferIO().
6850 : */
6851 2808472 : buf_state += BUF_REFCOUNT_ONE;
6852 2808472 : buf_hdr->io_wref = io_ref;
6853 :
6854 2808472 : if (is_temp)
6855 16818 : pg_atomic_unlocked_write_u32(&buf_hdr->state, buf_state);
6856 : else
6857 2791654 : UnlockBufHdr(buf_hdr, buf_state);
6858 :
6859 : /*
6860 : * Ensure the content lock that prevents buffer modifications while
6861 : * the buffer is being written out is not released early due to an
6862 : * error.
6863 : */
6864 2808472 : if (is_write && !is_temp)
6865 : {
6866 : LWLock *content_lock;
6867 :
6868 0 : content_lock = BufferDescriptorGetContentLock(buf_hdr);
6869 :
6870 : Assert(LWLockHeldByMe(content_lock));
6871 :
6872 : /*
6873 : * Lock is now owned by AIO subsystem.
6874 : */
6875 0 : LWLockDisown(content_lock);
6876 : }
6877 :
6878 : /*
6879 : * Stop tracking this buffer via the resowner - the AIO system now
6880 : * keeps track.
6881 : */
6882 2808472 : if (!is_temp)
6883 2791654 : ResourceOwnerForgetBufferIO(CurrentResourceOwner, buffer);
6884 : }
6885 2477342 : }
6886 :
6887 : /*
6888 : * Decode readv errors as encoded by buffer_readv_encode_error().
6889 : */
6890 : static inline void
6891 698 : buffer_readv_decode_error(PgAioResult result,
6892 : bool *zeroed_any,
6893 : bool *ignored_any,
6894 : uint8 *zeroed_or_error_count,
6895 : uint8 *checkfail_count,
6896 : uint8 *first_off)
6897 : {
6898 698 : uint32 rem_error = result.error_data;
6899 :
6900 : /* see static asserts in buffer_readv_encode_error */
6901 : #define READV_COUNT_BITS 7
6902 : #define READV_COUNT_MASK ((1 << READV_COUNT_BITS) - 1)
6903 :
6904 698 : *zeroed_any = rem_error & 1;
6905 698 : rem_error >>= 1;
6906 :
6907 698 : *ignored_any = rem_error & 1;
6908 698 : rem_error >>= 1;
6909 :
6910 698 : *zeroed_or_error_count = rem_error & READV_COUNT_MASK;
6911 698 : rem_error >>= READV_COUNT_BITS;
6912 :
6913 698 : *checkfail_count = rem_error & READV_COUNT_MASK;
6914 698 : rem_error >>= READV_COUNT_BITS;
6915 :
6916 698 : *first_off = rem_error & READV_COUNT_MASK;
6917 698 : rem_error >>= READV_COUNT_BITS;
6918 698 : }
6919 :
6920 : /*
6921 : * Helper to encode errors for buffer_readv_complete()
6922 : *
6923 : * Errors are encoded as follows:
6924 : * - bit 0 indicates whether any page was zeroed (1) or not (0)
6925 : * - bit 1 indicates whether any checksum failure was ignored (1) or not (0)
6926 : * - next READV_COUNT_BITS bits indicate the number of errored or zeroed pages
6927 : * - next READV_COUNT_BITS bits indicate the number of checksum failures
6928 : * - next READV_COUNT_BITS bits indicate the first offset of the first page
6929 : * that was errored or zeroed or, if no errors/zeroes, the first ignored
6930 : * checksum
6931 : */
6932 : static inline void
6933 384 : buffer_readv_encode_error(PgAioResult *result,
6934 : bool is_temp,
6935 : bool zeroed_any,
6936 : bool ignored_any,
6937 : uint8 error_count,
6938 : uint8 zeroed_count,
6939 : uint8 checkfail_count,
6940 : uint8 first_error_off,
6941 : uint8 first_zeroed_off,
6942 : uint8 first_ignored_off)
6943 : {
6944 :
6945 384 : uint8 shift = 0;
6946 384 : uint8 zeroed_or_error_count =
6947 : error_count > 0 ? error_count : zeroed_count;
6948 : uint8 first_off;
6949 :
6950 : StaticAssertStmt(PG_IOV_MAX <= 1 << READV_COUNT_BITS,
6951 : "PG_IOV_MAX is bigger than reserved space for error data");
6952 : StaticAssertStmt((1 + 1 + 3 * READV_COUNT_BITS) <= PGAIO_RESULT_ERROR_BITS,
6953 : "PGAIO_RESULT_ERROR_BITS is insufficient for buffer_readv");
6954 :
6955 : /*
6956 : * We only have space to encode one offset - but luckily that's good
6957 : * enough. If there is an error, the error is the interesting offset, same
6958 : * with a zeroed buffer vs an ignored buffer.
6959 : */
6960 384 : if (error_count > 0)
6961 188 : first_off = first_error_off;
6962 196 : else if (zeroed_count > 0)
6963 160 : first_off = first_zeroed_off;
6964 : else
6965 36 : first_off = first_ignored_off;
6966 :
6967 : Assert(!zeroed_any || error_count == 0);
6968 :
6969 384 : result->error_data = 0;
6970 :
6971 384 : result->error_data |= zeroed_any << shift;
6972 384 : shift += 1;
6973 :
6974 384 : result->error_data |= ignored_any << shift;
6975 384 : shift += 1;
6976 :
6977 384 : result->error_data |= ((uint32) zeroed_or_error_count) << shift;
6978 384 : shift += READV_COUNT_BITS;
6979 :
6980 384 : result->error_data |= ((uint32) checkfail_count) << shift;
6981 384 : shift += READV_COUNT_BITS;
6982 :
6983 384 : result->error_data |= ((uint32) first_off) << shift;
6984 384 : shift += READV_COUNT_BITS;
6985 :
6986 384 : result->id = is_temp ? PGAIO_HCB_LOCAL_BUFFER_READV :
6987 : PGAIO_HCB_SHARED_BUFFER_READV;
6988 :
6989 384 : if (error_count > 0)
6990 188 : result->status = PGAIO_RS_ERROR;
6991 : else
6992 196 : result->status = PGAIO_RS_WARNING;
6993 :
6994 : /*
6995 : * The encoding is complicated enough to warrant cross-checking it against
6996 : * the decode function.
6997 : */
6998 : #ifdef USE_ASSERT_CHECKING
6999 : {
7000 : bool zeroed_any_2,
7001 : ignored_any_2;
7002 : uint8 zeroed_or_error_count_2,
7003 : checkfail_count_2,
7004 : first_off_2;
7005 :
7006 : buffer_readv_decode_error(*result,
7007 : &zeroed_any_2, &ignored_any_2,
7008 : &zeroed_or_error_count_2,
7009 : &checkfail_count_2,
7010 : &first_off_2);
7011 : Assert(zeroed_any == zeroed_any_2);
7012 : Assert(ignored_any == ignored_any_2);
7013 : Assert(zeroed_or_error_count == zeroed_or_error_count_2);
7014 : Assert(checkfail_count == checkfail_count_2);
7015 : Assert(first_off == first_off_2);
7016 : }
7017 : #endif
7018 :
7019 : #undef READV_COUNT_BITS
7020 : #undef READV_COUNT_MASK
7021 384 : }
7022 :
7023 : /*
7024 : * Helper for AIO readv completion callbacks, supporting both shared and temp
7025 : * buffers. Gets called once for each buffer in a multi-page read.
7026 : */
7027 : static pg_attribute_always_inline void
7028 2533312 : buffer_readv_complete_one(PgAioTargetData *td, uint8 buf_off, Buffer buffer,
7029 : uint8 flags, bool failed, bool is_temp,
7030 : bool *buffer_invalid,
7031 : bool *failed_checksum,
7032 : bool *ignored_checksum,
7033 : bool *zeroed_buffer)
7034 : {
7035 2533312 : BufferDesc *buf_hdr = is_temp ?
7036 16818 : GetLocalBufferDescriptor(-buffer - 1)
7037 2533312 : : GetBufferDescriptor(buffer - 1);
7038 2533312 : BufferTag tag = buf_hdr->tag;
7039 2533312 : char *bufdata = BufferGetBlock(buffer);
7040 : uint32 set_flag_bits;
7041 : int piv_flags;
7042 :
7043 : /* check that the buffer is in the expected state for a read */
7044 : #ifdef USE_ASSERT_CHECKING
7045 : {
7046 : uint32 buf_state = pg_atomic_read_u32(&buf_hdr->state);
7047 :
7048 : Assert(buf_state & BM_TAG_VALID);
7049 : Assert(!(buf_state & BM_VALID));
7050 : /* temp buffers don't use BM_IO_IN_PROGRESS */
7051 : if (!is_temp)
7052 : Assert(buf_state & BM_IO_IN_PROGRESS);
7053 : Assert(!(buf_state & BM_DIRTY));
7054 : }
7055 : #endif
7056 :
7057 2533312 : *buffer_invalid = false;
7058 2533312 : *failed_checksum = false;
7059 2533312 : *ignored_checksum = false;
7060 2533312 : *zeroed_buffer = false;
7061 :
7062 : /*
7063 : * We ask PageIsVerified() to only log the message about checksum errors,
7064 : * as the completion might be run in any backend (or IO workers). We will
7065 : * report checksum errors in buffer_readv_report().
7066 : */
7067 2533312 : piv_flags = PIV_LOG_LOG;
7068 :
7069 : /* the local zero_damaged_pages may differ from the definer's */
7070 2533312 : if (flags & READ_BUFFERS_IGNORE_CHECKSUM_FAILURES)
7071 76 : piv_flags |= PIV_IGNORE_CHECKSUM_FAILURE;
7072 :
7073 : /* Check for garbage data. */
7074 2533312 : if (!failed)
7075 : {
7076 : /*
7077 : * If the buffer is not currently pinned by this backend, e.g. because
7078 : * we're completing this IO after an error, the buffer data will have
7079 : * been marked as inaccessible when the buffer was unpinned. The AIO
7080 : * subsystem holds a pin, but that doesn't prevent the buffer from
7081 : * having been marked as inaccessible. The completion might also be
7082 : * executed in a different process.
7083 : */
7084 : #ifdef USE_VALGRIND
7085 : if (!BufferIsPinned(buffer))
7086 : VALGRIND_MAKE_MEM_DEFINED(bufdata, BLCKSZ);
7087 : #endif
7088 :
7089 2533254 : if (!PageIsVerified((Page) bufdata, tag.blockNum, piv_flags,
7090 : failed_checksum))
7091 : {
7092 192 : if (flags & READ_BUFFERS_ZERO_ON_ERROR)
7093 : {
7094 92 : memset(bufdata, 0, BLCKSZ);
7095 92 : *zeroed_buffer = true;
7096 : }
7097 : else
7098 : {
7099 100 : *buffer_invalid = true;
7100 : /* mark buffer as having failed */
7101 100 : failed = true;
7102 : }
7103 : }
7104 2533062 : else if (*failed_checksum)
7105 24 : *ignored_checksum = true;
7106 :
7107 : /* undo what we did above */
7108 : #ifdef USE_VALGRIND
7109 : if (!BufferIsPinned(buffer))
7110 : VALGRIND_MAKE_MEM_NOACCESS(bufdata, BLCKSZ);
7111 : #endif
7112 :
7113 : /*
7114 : * Immediately log a message about the invalid page, but only to the
7115 : * server log. The reason to do so immediately is that this may be
7116 : * executed in a different backend than the one that originated the
7117 : * request. The reason to do so immediately is that the originator
7118 : * might not process the query result immediately (because it is busy
7119 : * doing another part of query processing) or at all (e.g. if it was
7120 : * cancelled or errored out due to another IO also failing). The
7121 : * definer of the IO will emit an ERROR or WARNING when processing the
7122 : * IO's results
7123 : *
7124 : * To avoid duplicating the code to emit these log messages, we reuse
7125 : * buffer_readv_report().
7126 : */
7127 2533254 : if (*buffer_invalid || *failed_checksum || *zeroed_buffer)
7128 : {
7129 216 : PgAioResult result_one = {0};
7130 :
7131 216 : buffer_readv_encode_error(&result_one, is_temp,
7132 216 : *zeroed_buffer,
7133 216 : *ignored_checksum,
7134 216 : *buffer_invalid,
7135 216 : *zeroed_buffer ? 1 : 0,
7136 216 : *failed_checksum ? 1 : 0,
7137 : buf_off, buf_off, buf_off);
7138 216 : pgaio_result_report(result_one, td, LOG_SERVER_ONLY);
7139 : }
7140 : }
7141 :
7142 : /* Terminate I/O and set BM_VALID. */
7143 2533312 : set_flag_bits = failed ? BM_IO_ERROR : BM_VALID;
7144 2533312 : if (is_temp)
7145 16818 : TerminateLocalBufferIO(buf_hdr, false, set_flag_bits, true);
7146 : else
7147 2516494 : TerminateBufferIO(buf_hdr, false, set_flag_bits, false, true);
7148 :
7149 : /*
7150 : * Call the BUFFER_READ_DONE tracepoint in the callback, even though the
7151 : * callback may not be executed in the same backend that called
7152 : * BUFFER_READ_START. The alternative would be to defer calling the
7153 : * tracepoint to a later point (e.g. the local completion callback for
7154 : * shared buffer reads), which seems even less helpful.
7155 : */
7156 : TRACE_POSTGRESQL_BUFFER_READ_DONE(tag.forkNum,
7157 : tag.blockNum,
7158 : tag.spcOid,
7159 : tag.dbOid,
7160 : tag.relNumber,
7161 : is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
7162 : false);
7163 2533312 : }
7164 :
7165 : /*
7166 : * Perform completion handling of a single AIO read. This read may cover
7167 : * multiple blocks / buffers.
7168 : *
7169 : * Shared between shared and local buffers, to reduce code duplication.
7170 : */
7171 : static pg_attribute_always_inline PgAioResult
7172 2267922 : buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7173 : uint8 cb_data, bool is_temp)
7174 : {
7175 2267922 : PgAioResult result = prior_result;
7176 2267922 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
7177 2267922 : uint8 first_error_off = 0;
7178 2267922 : uint8 first_zeroed_off = 0;
7179 2267922 : uint8 first_ignored_off = 0;
7180 2267922 : uint8 error_count = 0;
7181 2267922 : uint8 zeroed_count = 0;
7182 2267922 : uint8 ignored_count = 0;
7183 2267922 : uint8 checkfail_count = 0;
7184 : uint64 *io_data;
7185 : uint8 handle_data_len;
7186 :
7187 : if (is_temp)
7188 : {
7189 : Assert(td->smgr.is_temp);
7190 : Assert(pgaio_io_get_owner(ioh) == MyProcNumber);
7191 : }
7192 : else
7193 : Assert(!td->smgr.is_temp);
7194 :
7195 : /*
7196 : * Iterate over all the buffers affected by this IO and call the
7197 : * per-buffer completion function for each buffer.
7198 : */
7199 2267922 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
7200 4801234 : for (uint8 buf_off = 0; buf_off < handle_data_len; buf_off++)
7201 : {
7202 2533312 : Buffer buf = io_data[buf_off];
7203 : bool failed;
7204 2533312 : bool failed_verification = false;
7205 2533312 : bool failed_checksum = false;
7206 2533312 : bool zeroed_buffer = false;
7207 2533312 : bool ignored_checksum = false;
7208 :
7209 : Assert(BufferIsValid(buf));
7210 :
7211 : /*
7212 : * If the entire I/O failed on a lower-level, each buffer needs to be
7213 : * marked as failed. In case of a partial read, the first few buffers
7214 : * may be ok.
7215 : */
7216 2533312 : failed =
7217 2533312 : prior_result.status == PGAIO_RS_ERROR
7218 2533312 : || prior_result.result <= buf_off;
7219 :
7220 2533312 : buffer_readv_complete_one(td, buf_off, buf, cb_data, failed, is_temp,
7221 : &failed_verification,
7222 : &failed_checksum,
7223 : &ignored_checksum,
7224 : &zeroed_buffer);
7225 :
7226 : /*
7227 : * Track information about the number of different kinds of error
7228 : * conditions across all pages, as there can be multiple pages failing
7229 : * verification as part of one IO.
7230 : */
7231 2533312 : if (failed_verification && !zeroed_buffer && error_count++ == 0)
7232 88 : first_error_off = buf_off;
7233 2533312 : if (zeroed_buffer && zeroed_count++ == 0)
7234 68 : first_zeroed_off = buf_off;
7235 2533312 : if (ignored_checksum && ignored_count++ == 0)
7236 20 : first_ignored_off = buf_off;
7237 2533312 : if (failed_checksum)
7238 64 : checkfail_count++;
7239 : }
7240 :
7241 : /*
7242 : * If the smgr read succeeded [partially] and page verification failed for
7243 : * some of the pages, adjust the IO's result state appropriately.
7244 : */
7245 2267922 : if (prior_result.status != PGAIO_RS_ERROR &&
7246 2267816 : (error_count > 0 || ignored_count > 0 || zeroed_count > 0))
7247 : {
7248 168 : buffer_readv_encode_error(&result, is_temp,
7249 : zeroed_count > 0, ignored_count > 0,
7250 : error_count, zeroed_count, checkfail_count,
7251 : first_error_off, first_zeroed_off,
7252 : first_ignored_off);
7253 168 : pgaio_result_report(result, td, DEBUG1);
7254 : }
7255 :
7256 : /*
7257 : * For shared relations this reporting is done in
7258 : * shared_buffer_readv_complete_local().
7259 : */
7260 2267922 : if (is_temp && checkfail_count > 0)
7261 4 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
7262 : checkfail_count);
7263 :
7264 2267922 : return result;
7265 : }
7266 :
7267 : /*
7268 : * AIO error reporting callback for aio_shared_buffer_readv_cb and
7269 : * aio_local_buffer_readv_cb.
7270 : *
7271 : * The error is encoded / decoded in buffer_readv_encode_error() /
7272 : * buffer_readv_decode_error().
7273 : */
7274 : static void
7275 544 : buffer_readv_report(PgAioResult result, const PgAioTargetData *td,
7276 : int elevel)
7277 : {
7278 544 : int nblocks = td->smgr.nblocks;
7279 544 : BlockNumber first = td->smgr.blockNum;
7280 544 : BlockNumber last = first + nblocks - 1;
7281 544 : ProcNumber errProc =
7282 544 : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER;
7283 : RelPathStr rpath =
7284 544 : relpathbackend(td->smgr.rlocator, errProc, td->smgr.forkNum);
7285 : bool zeroed_any,
7286 : ignored_any;
7287 : uint8 zeroed_or_error_count,
7288 : checkfail_count,
7289 : first_off;
7290 : uint8 affected_count;
7291 : const char *msg_one,
7292 : *msg_mult,
7293 : *det_mult,
7294 : *hint_mult;
7295 :
7296 544 : buffer_readv_decode_error(result, &zeroed_any, &ignored_any,
7297 : &zeroed_or_error_count,
7298 : &checkfail_count,
7299 : &first_off);
7300 :
7301 : /*
7302 : * Treat a read that had both zeroed buffers *and* ignored checksums as a
7303 : * special case, it's too irregular to be emitted the same way as the
7304 : * other cases.
7305 : */
7306 544 : if (zeroed_any && ignored_any)
7307 : {
7308 : Assert(zeroed_any && ignored_any);
7309 : Assert(nblocks > 1); /* same block can't be both zeroed and ignored */
7310 : Assert(result.status != PGAIO_RS_ERROR);
7311 8 : affected_count = zeroed_or_error_count;
7312 :
7313 8 : ereport(elevel,
7314 : errcode(ERRCODE_DATA_CORRUPTED),
7315 : errmsg("zeroing %u page(s) and ignoring %u checksum failure(s) among blocks %u..%u of relation \"%s\"",
7316 : affected_count, checkfail_count, first, last, rpath.str),
7317 : affected_count > 1 ?
7318 : errdetail("Block %u held the first zeroed page.",
7319 : first + first_off) : 0,
7320 : errhint_plural("See server log for details about the other %d invalid block.",
7321 : "See server log for details about the other %d invalid blocks.",
7322 : affected_count + checkfail_count - 1,
7323 : affected_count + checkfail_count - 1));
7324 8 : return;
7325 : }
7326 :
7327 : /*
7328 : * The other messages are highly repetitive. To avoid duplicating a long
7329 : * and complicated ereport(), gather the translated format strings
7330 : * separately and then do one common ereport.
7331 : */
7332 536 : if (result.status == PGAIO_RS_ERROR)
7333 : {
7334 : Assert(!zeroed_any); /* can't have invalid pages when zeroing them */
7335 272 : affected_count = zeroed_or_error_count;
7336 272 : msg_one = _("invalid page in block %u of relation \"%s\"");
7337 272 : msg_mult = _("%u invalid pages among blocks %u..%u of relation \"%s\"");
7338 272 : det_mult = _("Block %u held the first invalid page.");
7339 272 : hint_mult = _("See server log for the other %u invalid block(s).");
7340 : }
7341 264 : else if (zeroed_any && !ignored_any)
7342 : {
7343 216 : affected_count = zeroed_or_error_count;
7344 216 : msg_one = _("invalid page in block %u of relation \"%s\"; zeroing out page");
7345 216 : msg_mult = _("zeroing out %u invalid pages among blocks %u..%u of relation \"%s\"");
7346 216 : det_mult = _("Block %u held the first zeroed page.");
7347 216 : hint_mult = _("See server log for the other %u zeroed block(s).");
7348 : }
7349 48 : else if (!zeroed_any && ignored_any)
7350 : {
7351 48 : affected_count = checkfail_count;
7352 48 : msg_one = _("ignoring checksum failure in block %u of relation \"%s\"");
7353 48 : msg_mult = _("ignoring %u checksum failures among blocks %u..%u of relation \"%s\"");
7354 48 : det_mult = _("Block %u held the first ignored page.");
7355 48 : hint_mult = _("See server log for the other %u ignored block(s).");
7356 : }
7357 : else
7358 0 : pg_unreachable();
7359 :
7360 536 : ereport(elevel,
7361 : errcode(ERRCODE_DATA_CORRUPTED),
7362 : affected_count == 1 ?
7363 : errmsg_internal(msg_one, first + first_off, rpath.str) :
7364 : errmsg_internal(msg_mult, affected_count, first, last, rpath.str),
7365 : affected_count > 1 ? errdetail_internal(det_mult, first + first_off) : 0,
7366 : affected_count > 1 ? errhint_internal(hint_mult, affected_count - 1) : 0);
7367 : }
7368 :
7369 : static void
7370 2473730 : shared_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
7371 : {
7372 2473730 : buffer_stage_common(ioh, false, false);
7373 2473730 : }
7374 :
7375 : static PgAioResult
7376 2264310 : shared_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7377 : uint8 cb_data)
7378 : {
7379 2264310 : return buffer_readv_complete(ioh, prior_result, cb_data, false);
7380 : }
7381 :
7382 : /*
7383 : * We need a backend-local completion callback for shared buffers, to be able
7384 : * to report checksum errors correctly. Unfortunately that can only safely
7385 : * happen if the reporting backend has previously called
7386 : * pgstat_prepare_report_checksum_failure(), which we can only guarantee in
7387 : * the backend that started the IO. Hence this callback.
7388 : */
7389 : static PgAioResult
7390 2473730 : shared_buffer_readv_complete_local(PgAioHandle *ioh, PgAioResult prior_result,
7391 : uint8 cb_data)
7392 : {
7393 : bool zeroed_any,
7394 : ignored_any;
7395 : uint8 zeroed_or_error_count,
7396 : checkfail_count,
7397 : first_off;
7398 :
7399 2473730 : if (prior_result.status == PGAIO_RS_OK)
7400 2473576 : return prior_result;
7401 :
7402 154 : buffer_readv_decode_error(prior_result,
7403 : &zeroed_any,
7404 : &ignored_any,
7405 : &zeroed_or_error_count,
7406 : &checkfail_count,
7407 : &first_off);
7408 :
7409 154 : if (checkfail_count)
7410 : {
7411 48 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
7412 :
7413 48 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
7414 : checkfail_count);
7415 : }
7416 :
7417 154 : return prior_result;
7418 : }
7419 :
7420 : static void
7421 3612 : local_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
7422 : {
7423 3612 : buffer_stage_common(ioh, false, true);
7424 3612 : }
7425 :
7426 : static PgAioResult
7427 3612 : local_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7428 : uint8 cb_data)
7429 : {
7430 3612 : return buffer_readv_complete(ioh, prior_result, cb_data, true);
7431 : }
7432 :
7433 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
7434 : const PgAioHandleCallbacks aio_shared_buffer_readv_cb = {
7435 : .stage = shared_buffer_readv_stage,
7436 : .complete_shared = shared_buffer_readv_complete,
7437 : /* need a local callback to report checksum failures */
7438 : .complete_local = shared_buffer_readv_complete_local,
7439 : .report = buffer_readv_report,
7440 : };
7441 :
7442 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
7443 : const PgAioHandleCallbacks aio_local_buffer_readv_cb = {
7444 : .stage = local_buffer_readv_stage,
7445 :
7446 : /*
7447 : * Note that this, in contrast to the shared_buffers case, uses
7448 : * complete_local, as only the issuing backend has access to the required
7449 : * datastructures. This is important in case the IO completion may be
7450 : * consumed incidentally by another backend.
7451 : */
7452 : .complete_local = local_buffer_readv_complete,
7453 : .report = buffer_readv_report,
7454 : };
|
