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 129527632 : ReservePrivateRefCountEntry(void)
260 : {
261 : /* Already reserved (or freed), nothing to do */
262 129527632 : if (ReservedRefCountEntry != NULL)
263 121328748 : 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 20881634 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
273 : {
274 : PrivateRefCountEntry *res;
275 :
276 20540378 : res = &PrivateRefCountArray[i];
277 :
278 20540378 : if (res->buffer == InvalidBuffer)
279 : {
280 7857628 : ReservedRefCountEntry = res;
281 7857628 : 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 341256 : ReservedRefCountEntry =
300 341256 : &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 341256 : hashent = hash_search(PrivateRefCountHash,
307 341256 : &(ReservedRefCountEntry->buffer),
308 : HASH_ENTER,
309 : &found);
310 : Assert(!found);
311 341256 : hashent->refcount = ReservedRefCountEntry->refcount;
312 :
313 : /* clear the now free array slot */
314 341256 : ReservedRefCountEntry->buffer = InvalidBuffer;
315 341256 : ReservedRefCountEntry->refcount = 0;
316 :
317 341256 : PrivateRefCountOverflowed++;
318 : }
319 : }
320 :
321 : /*
322 : * Fill a previously reserved refcount entry.
323 : */
324 : static PrivateRefCountEntry *
325 118101848 : 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 118101848 : res = ReservedRefCountEntry;
334 118101848 : ReservedRefCountEntry = NULL;
335 :
336 : /* and fill it */
337 118101848 : res->buffer = buffer;
338 118101848 : res->refcount = 0;
339 :
340 118101848 : 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 291836122 : 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 1372854242 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
364 : {
365 1258794844 : res = &PrivateRefCountArray[i];
366 :
367 1258794844 : if (res->buffer == buffer)
368 177776724 : 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 114059398 : if (PrivateRefCountOverflowed == 0)
379 113216232 : return NULL;
380 :
381 843166 : res = hash_search(PrivateRefCountHash, &buffer, HASH_FIND, NULL);
382 :
383 843166 : if (res == NULL)
384 497142 : return NULL;
385 346024 : else if (!do_move)
386 : {
387 : /* caller doesn't want us to move the hash entry into the array */
388 303938 : 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 42086 : ReservePrivateRefCountEntry();
398 :
399 : /* Use up the reserved slot */
400 : Assert(ReservedRefCountEntry != NULL);
401 42086 : free = ReservedRefCountEntry;
402 42086 : ReservedRefCountEntry = NULL;
403 : Assert(free->buffer == InvalidBuffer);
404 :
405 : /* and fill it */
406 42086 : free->buffer = buffer;
407 42086 : free->refcount = res->refcount;
408 :
409 : /* delete from hashtable */
410 42086 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
411 : Assert(found);
412 : Assert(PrivateRefCountOverflowed > 0);
413 42086 : PrivateRefCountOverflowed--;
414 :
415 42086 : 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 6179846 : 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 6179846 : ref = GetPrivateRefCountEntry(buffer, false);
437 :
438 6179846 : if (ref == NULL)
439 50 : return 0;
440 6179796 : 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 118101848 : ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref)
449 : {
450 : Assert(ref->refcount == 0);
451 :
452 118101848 : if (ref >= &PrivateRefCountArray[0] &&
453 : ref < &PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES])
454 : {
455 117802678 : 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 117802678 : ReservedRefCountEntry = ref;
463 : }
464 : else
465 : {
466 : bool found;
467 299170 : Buffer buffer = ref->buffer;
468 :
469 299170 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
470 : Assert(found);
471 : Assert(PrivateRefCountOverflowed > 0);
472 299170 : PrivateRefCountOverflowed--;
473 : }
474 118101848 : }
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 62428 : PrefetchSharedBuffer(SMgrRelation smgr_reln,
564 : ForkNumber forkNum,
565 : BlockNumber blockNum)
566 : {
567 62428 : 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 62428 : InitBufferTag(&newTag, &smgr_reln->smgr_rlocator.locator,
577 : forkNum, blockNum);
578 :
579 : /* determine its hash code and partition lock ID */
580 62428 : newHash = BufTableHashCode(&newTag);
581 62428 : newPartitionLock = BufMappingPartitionLock(newHash);
582 :
583 : /* see if the block is in the buffer pool already */
584 62428 : LWLockAcquire(newPartitionLock, LW_SHARED);
585 62428 : buf_id = BufTableLookup(&newTag, newHash);
586 62428 : LWLockRelease(newPartitionLock);
587 :
588 : /* If not in buffers, initiate prefetch */
589 62428 : 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 34688 : if ((io_direct_flags & IO_DIRECT_DATA) == 0 &&
597 17120 : smgrprefetch(smgr_reln, forkNum, blockNum, 1))
598 : {
599 17120 : 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 44860 : 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 62428 : 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 41068 : PrefetchBuffer(Relation reln, ForkNumber forkNum, BlockNumber blockNum)
654 : {
655 : Assert(RelationIsValid(reln));
656 : Assert(BlockNumberIsValid(blockNum));
657 :
658 41068 : 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 39502 : 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 9188 : 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 9188 : ResourceOwnerEnlarge(CurrentResourceOwner);
694 9188 : ReservePrivateRefCountEntry();
695 9188 : InitBufferTag(&tag, &rlocator, forkNum, blockNum);
696 :
697 9188 : 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 9124 : 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 18160 : if (BufferTagsEqual(&tag, &bufHdr->tag) &&
726 9036 : PinBuffer(bufHdr, NULL, true))
727 : {
728 9024 : if (BufferTagsEqual(&tag, &bufHdr->tag))
729 : {
730 9024 : pgBufferUsage.shared_blks_hit++;
731 9024 : return true;
732 : }
733 0 : UnpinBuffer(bufHdr);
734 : }
735 : }
736 :
737 100 : 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 86304144 : ReadBuffer(Relation reln, BlockNumber blockNum)
746 : {
747 86304144 : 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 104246710 : 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 104246710 : 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 104246710 : buf = ReadBuffer_common(reln, RelationGetSmgr(reln), 0,
812 : forkNum, blockNum, mode, strategy);
813 :
814 104246664 : 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 11389326 : ReadBufferWithoutRelcache(RelFileLocator rlocator, ForkNumber forkNum,
830 : BlockNumber blockNum, ReadBufferMode mode,
831 : BufferAccessStrategy strategy, bool permanent)
832 : {
833 11389326 : SMgrRelation smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
834 :
835 11389326 : 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 90670 : ExtendBufferedRel(BufferManagerRelation bmr,
846 : ForkNumber forkNum,
847 : BufferAccessStrategy strategy,
848 : uint32 flags)
849 : {
850 : Buffer buf;
851 90670 : uint32 extend_by = 1;
852 :
853 90670 : ExtendBufferedRelBy(bmr, forkNum, strategy, flags, extend_by,
854 : &buf, &extend_by);
855 :
856 90670 : 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 318884 : 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 318884 : if (bmr.smgr == NULL)
890 : {
891 318884 : bmr.smgr = RelationGetSmgr(bmr.rel);
892 318884 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
893 : }
894 :
895 318884 : return ExtendBufferedRelCommon(bmr, fork, strategy, flags,
896 : extend_by, InvalidBlockNumber,
897 : buffers, extended_by);
898 : }
899 :
900 : /*
901 : * Extend the relation so it is at least extend_to blocks large, return buffer
902 : * (extend_to - 1).
903 : *
904 : * This is useful for callers that want to write a specific page, regardless
905 : * of the current size of the relation (e.g. useful for visibilitymap and for
906 : * crash recovery).
907 : */
908 : Buffer
909 102820 : ExtendBufferedRelTo(BufferManagerRelation bmr,
910 : ForkNumber fork,
911 : BufferAccessStrategy strategy,
912 : uint32 flags,
913 : BlockNumber extend_to,
914 : ReadBufferMode mode)
915 : {
916 : BlockNumber current_size;
917 102820 : uint32 extended_by = 0;
918 102820 : Buffer buffer = InvalidBuffer;
919 : Buffer buffers[64];
920 :
921 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
922 : Assert(bmr.smgr == NULL || bmr.relpersistence != 0);
923 : Assert(extend_to != InvalidBlockNumber && extend_to > 0);
924 :
925 102820 : if (bmr.smgr == NULL)
926 : {
927 14084 : bmr.smgr = RelationGetSmgr(bmr.rel);
928 14084 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
929 : }
930 :
931 : /*
932 : * If desired, create the file if it doesn't exist. If
933 : * smgr_cached_nblocks[fork] is positive then it must exist, no need for
934 : * an smgrexists call.
935 : */
936 102820 : if ((flags & EB_CREATE_FORK_IF_NEEDED) &&
937 14084 : (bmr.smgr->smgr_cached_nblocks[fork] == 0 ||
938 36 : bmr.smgr->smgr_cached_nblocks[fork] == InvalidBlockNumber) &&
939 14048 : !smgrexists(bmr.smgr, fork))
940 : {
941 14016 : LockRelationForExtension(bmr.rel, ExclusiveLock);
942 :
943 : /* recheck, fork might have been created concurrently */
944 14016 : if (!smgrexists(bmr.smgr, fork))
945 14010 : smgrcreate(bmr.smgr, fork, flags & EB_PERFORMING_RECOVERY);
946 :
947 14016 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
948 : }
949 :
950 : /*
951 : * If requested, invalidate size cache, so that smgrnblocks asks the
952 : * kernel.
953 : */
954 102820 : if (flags & EB_CLEAR_SIZE_CACHE)
955 14084 : bmr.smgr->smgr_cached_nblocks[fork] = InvalidBlockNumber;
956 :
957 : /*
958 : * Estimate how many pages we'll need to extend by. This avoids acquiring
959 : * unnecessarily many victim buffers.
960 : */
961 102820 : current_size = smgrnblocks(bmr.smgr, fork);
962 :
963 : /*
964 : * Since no-one else can be looking at the page contents yet, there is no
965 : * difference between an exclusive lock and a cleanup-strength lock. Note
966 : * that we pass the original mode to ReadBuffer_common() below, when
967 : * falling back to reading the buffer to a concurrent relation extension.
968 : */
969 102820 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
970 88004 : flags |= EB_LOCK_TARGET;
971 :
972 209926 : while (current_size < extend_to)
973 : {
974 107106 : uint32 num_pages = lengthof(buffers);
975 : BlockNumber first_block;
976 :
977 107106 : if ((uint64) current_size + num_pages > extend_to)
978 106974 : num_pages = extend_to - current_size;
979 :
980 107106 : first_block = ExtendBufferedRelCommon(bmr, fork, strategy, flags,
981 : num_pages, extend_to,
982 : buffers, &extended_by);
983 :
984 107106 : current_size = first_block + extended_by;
985 : Assert(num_pages != 0 || current_size >= extend_to);
986 :
987 228622 : for (uint32 i = 0; i < extended_by; i++)
988 : {
989 121516 : if (first_block + i != extend_to - 1)
990 18722 : ReleaseBuffer(buffers[i]);
991 : else
992 102794 : buffer = buffers[i];
993 : }
994 : }
995 :
996 : /*
997 : * It's possible that another backend concurrently extended the relation.
998 : * In that case read the buffer.
999 : *
1000 : * XXX: Should we control this via a flag?
1001 : */
1002 102820 : if (buffer == InvalidBuffer)
1003 : {
1004 : Assert(extended_by == 0);
1005 26 : buffer = ReadBuffer_common(bmr.rel, bmr.smgr, bmr.relpersistence,
1006 : fork, extend_to - 1, mode, strategy);
1007 : }
1008 :
1009 102820 : return buffer;
1010 : }
1011 :
1012 : /*
1013 : * Lock and optionally zero a buffer, as part of the implementation of
1014 : * RBM_ZERO_AND_LOCK or RBM_ZERO_AND_CLEANUP_LOCK. The buffer must be already
1015 : * pinned. If the buffer is not already valid, it is zeroed and made valid.
1016 : */
1017 : static void
1018 628800 : ZeroAndLockBuffer(Buffer buffer, ReadBufferMode mode, bool already_valid)
1019 : {
1020 : BufferDesc *bufHdr;
1021 : bool need_to_zero;
1022 628800 : bool isLocalBuf = BufferIsLocal(buffer);
1023 :
1024 : Assert(mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK);
1025 :
1026 628800 : if (already_valid)
1027 : {
1028 : /*
1029 : * If the caller already knew the buffer was valid, we can skip some
1030 : * header interaction. The caller just wants to lock the buffer.
1031 : */
1032 73144 : need_to_zero = false;
1033 : }
1034 555656 : else if (isLocalBuf)
1035 : {
1036 : /* Simple case for non-shared buffers. */
1037 48 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
1038 48 : need_to_zero = StartLocalBufferIO(bufHdr, true, false);
1039 : }
1040 : else
1041 : {
1042 : /*
1043 : * Take BM_IO_IN_PROGRESS, or discover that BM_VALID has been set
1044 : * concurrently. Even though we aren't doing I/O, that ensures that
1045 : * we don't zero a page that someone else has pinned. An exclusive
1046 : * content lock wouldn't be enough, because readers are allowed to
1047 : * drop the content lock after determining that a tuple is visible
1048 : * (see buffer access rules in README).
1049 : */
1050 555608 : bufHdr = GetBufferDescriptor(buffer - 1);
1051 555608 : need_to_zero = StartBufferIO(bufHdr, true, false);
1052 : }
1053 :
1054 628800 : if (need_to_zero)
1055 : {
1056 555656 : memset(BufferGetPage(buffer), 0, BLCKSZ);
1057 :
1058 : /*
1059 : * Grab the buffer content lock before marking the page as valid, to
1060 : * make sure that no other backend sees the zeroed page before the
1061 : * caller has had a chance to initialize it.
1062 : *
1063 : * Since no-one else can be looking at the page contents yet, there is
1064 : * no difference between an exclusive lock and a cleanup-strength
1065 : * lock. (Note that we cannot use LockBuffer() or
1066 : * LockBufferForCleanup() here, because they assert that the buffer is
1067 : * already valid.)
1068 : */
1069 555656 : if (!isLocalBuf)
1070 555608 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1071 :
1072 : /* Set BM_VALID, terminate IO, and wake up any waiters */
1073 555656 : if (isLocalBuf)
1074 48 : TerminateLocalBufferIO(bufHdr, false, BM_VALID, false);
1075 : else
1076 555608 : TerminateBufferIO(bufHdr, false, BM_VALID, true, false);
1077 : }
1078 73144 : else if (!isLocalBuf)
1079 : {
1080 : /*
1081 : * The buffer is valid, so we can't zero it. The caller still expects
1082 : * the page to be locked on return.
1083 : */
1084 73104 : if (mode == RBM_ZERO_AND_LOCK)
1085 73004 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1086 : else
1087 100 : LockBufferForCleanup(buffer);
1088 : }
1089 628800 : }
1090 :
1091 : /*
1092 : * Pin a buffer for a given block. *foundPtr is set to true if the block was
1093 : * already present, or false if more work is required to either read it in or
1094 : * zero it.
1095 : */
1096 : static pg_attribute_always_inline Buffer
1097 123914328 : PinBufferForBlock(Relation rel,
1098 : SMgrRelation smgr,
1099 : char persistence,
1100 : ForkNumber forkNum,
1101 : BlockNumber blockNum,
1102 : BufferAccessStrategy strategy,
1103 : bool *foundPtr)
1104 : {
1105 : BufferDesc *bufHdr;
1106 : IOContext io_context;
1107 : IOObject io_object;
1108 :
1109 : Assert(blockNum != P_NEW);
1110 :
1111 : /* Persistence should be set before */
1112 : Assert((persistence == RELPERSISTENCE_TEMP ||
1113 : persistence == RELPERSISTENCE_PERMANENT ||
1114 : persistence == RELPERSISTENCE_UNLOGGED));
1115 :
1116 123914328 : if (persistence == RELPERSISTENCE_TEMP)
1117 : {
1118 2554156 : io_context = IOCONTEXT_NORMAL;
1119 2554156 : io_object = IOOBJECT_TEMP_RELATION;
1120 : }
1121 : else
1122 : {
1123 121360172 : io_context = IOContextForStrategy(strategy);
1124 121360172 : io_object = IOOBJECT_RELATION;
1125 : }
1126 :
1127 : TRACE_POSTGRESQL_BUFFER_READ_START(forkNum, blockNum,
1128 : smgr->smgr_rlocator.locator.spcOid,
1129 : smgr->smgr_rlocator.locator.dbOid,
1130 : smgr->smgr_rlocator.locator.relNumber,
1131 : smgr->smgr_rlocator.backend);
1132 :
1133 123914328 : if (persistence == RELPERSISTENCE_TEMP)
1134 : {
1135 2554156 : bufHdr = LocalBufferAlloc(smgr, forkNum, blockNum, foundPtr);
1136 2554144 : if (*foundPtr)
1137 2537362 : pgBufferUsage.local_blks_hit++;
1138 : }
1139 : else
1140 : {
1141 121360172 : bufHdr = BufferAlloc(smgr, persistence, forkNum, blockNum,
1142 : strategy, foundPtr, io_context);
1143 121360172 : if (*foundPtr)
1144 118012072 : pgBufferUsage.shared_blks_hit++;
1145 : }
1146 123914316 : if (rel)
1147 : {
1148 : /*
1149 : * While pgBufferUsage's "read" counter isn't bumped unless we reach
1150 : * WaitReadBuffers() (so, not for hits, and not for buffers that are
1151 : * zeroed instead), the per-relation stats always count them.
1152 : */
1153 112058514 : pgstat_count_buffer_read(rel);
1154 112058514 : if (*foundPtr)
1155 109609122 : pgstat_count_buffer_hit(rel);
1156 : }
1157 123914316 : if (*foundPtr)
1158 : {
1159 120549434 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1160 120549434 : if (VacuumCostActive)
1161 5812544 : VacuumCostBalance += VacuumCostPageHit;
1162 :
1163 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forkNum, blockNum,
1164 : smgr->smgr_rlocator.locator.spcOid,
1165 : smgr->smgr_rlocator.locator.dbOid,
1166 : smgr->smgr_rlocator.locator.relNumber,
1167 : smgr->smgr_rlocator.backend,
1168 : true);
1169 : }
1170 :
1171 123914316 : return BufferDescriptorGetBuffer(bufHdr);
1172 : }
1173 :
1174 : /*
1175 : * ReadBuffer_common -- common logic for all ReadBuffer variants
1176 : *
1177 : * smgr is required, rel is optional unless using P_NEW.
1178 : */
1179 : static pg_attribute_always_inline Buffer
1180 115636062 : ReadBuffer_common(Relation rel, SMgrRelation smgr, char smgr_persistence,
1181 : ForkNumber forkNum,
1182 : BlockNumber blockNum, ReadBufferMode mode,
1183 : BufferAccessStrategy strategy)
1184 : {
1185 : ReadBuffersOperation operation;
1186 : Buffer buffer;
1187 : int flags;
1188 : char persistence;
1189 :
1190 : /*
1191 : * Backward compatibility path, most code should use ExtendBufferedRel()
1192 : * instead, as acquiring the extension lock inside ExtendBufferedRel()
1193 : * scales a lot better.
1194 : */
1195 115636062 : if (unlikely(blockNum == P_NEW))
1196 : {
1197 520 : uint32 flags = EB_SKIP_EXTENSION_LOCK;
1198 :
1199 : /*
1200 : * Since no-one else can be looking at the page contents yet, there is
1201 : * no difference between an exclusive lock and a cleanup-strength
1202 : * lock.
1203 : */
1204 520 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1205 0 : flags |= EB_LOCK_FIRST;
1206 :
1207 520 : return ExtendBufferedRel(BMR_REL(rel), forkNum, strategy, flags);
1208 : }
1209 :
1210 115635542 : if (rel)
1211 104246216 : persistence = rel->rd_rel->relpersistence;
1212 : else
1213 11389326 : persistence = smgr_persistence;
1214 :
1215 115635542 : if (unlikely(mode == RBM_ZERO_AND_CLEANUP_LOCK ||
1216 : mode == RBM_ZERO_AND_LOCK))
1217 : {
1218 : bool found;
1219 :
1220 628800 : buffer = PinBufferForBlock(rel, smgr, persistence,
1221 : forkNum, blockNum, strategy, &found);
1222 628800 : ZeroAndLockBuffer(buffer, mode, found);
1223 628800 : return buffer;
1224 : }
1225 :
1226 : /*
1227 : * Signal that we are going to immediately wait. If we're immediately
1228 : * waiting, there is no benefit in actually executing the IO
1229 : * asynchronously, it would just add dispatch overhead.
1230 : */
1231 115006742 : flags = READ_BUFFERS_SYNCHRONOUSLY;
1232 115006742 : if (mode == RBM_ZERO_ON_ERROR)
1233 3033370 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1234 115006742 : operation.smgr = smgr;
1235 115006742 : operation.rel = rel;
1236 115006742 : operation.persistence = persistence;
1237 115006742 : operation.forknum = forkNum;
1238 115006742 : operation.strategy = strategy;
1239 115006742 : if (StartReadBuffer(&operation,
1240 : &buffer,
1241 : blockNum,
1242 : flags))
1243 1372814 : WaitReadBuffers(&operation);
1244 :
1245 115006696 : return buffer;
1246 : }
1247 :
1248 : static pg_attribute_always_inline bool
1249 122953574 : StartReadBuffersImpl(ReadBuffersOperation *operation,
1250 : Buffer *buffers,
1251 : BlockNumber blockNum,
1252 : int *nblocks,
1253 : int flags,
1254 : bool allow_forwarding)
1255 : {
1256 122953574 : int actual_nblocks = *nblocks;
1257 122953574 : int maxcombine = 0;
1258 : bool did_start_io;
1259 :
1260 : Assert(*nblocks == 1 || allow_forwarding);
1261 : Assert(*nblocks > 0);
1262 : Assert(*nblocks <= MAX_IO_COMBINE_LIMIT);
1263 :
1264 125762800 : for (int i = 0; i < actual_nblocks; ++i)
1265 : {
1266 : bool found;
1267 :
1268 123288312 : if (allow_forwarding && buffers[i] != InvalidBuffer)
1269 2784 : {
1270 : BufferDesc *bufHdr;
1271 :
1272 : /*
1273 : * This is a buffer that was pinned by an earlier call to
1274 : * StartReadBuffers(), but couldn't be handled in one operation at
1275 : * that time. The operation was split, and the caller has passed
1276 : * an already pinned buffer back to us to handle the rest of the
1277 : * operation. It must continue at the expected block number.
1278 : */
1279 : Assert(BufferGetBlockNumber(buffers[i]) == blockNum + i);
1280 :
1281 : /*
1282 : * It might be an already valid buffer (a hit) that followed the
1283 : * final contiguous block of an earlier I/O (a miss) marking the
1284 : * end of it, or a buffer that some other backend has since made
1285 : * valid by performing the I/O for us, in which case we can handle
1286 : * it as a hit now. It is safe to check for a BM_VALID flag with
1287 : * a relaxed load, because we got a fresh view of it while pinning
1288 : * it in the previous call.
1289 : *
1290 : * On the other hand if we don't see BM_VALID yet, it must be an
1291 : * I/O that was split by the previous call and we need to try to
1292 : * start a new I/O from this block. We're also racing against any
1293 : * other backend that might start the I/O or even manage to mark
1294 : * it BM_VALID after this check, but StartBufferIO() will handle
1295 : * those cases.
1296 : */
1297 2784 : if (BufferIsLocal(buffers[i]))
1298 4 : bufHdr = GetLocalBufferDescriptor(-buffers[i] - 1);
1299 : else
1300 2780 : bufHdr = GetBufferDescriptor(buffers[i] - 1);
1301 : Assert(pg_atomic_read_u32(&bufHdr->state) & BM_TAG_VALID);
1302 2784 : found = pg_atomic_read_u32(&bufHdr->state) & BM_VALID;
1303 : }
1304 : else
1305 : {
1306 123285516 : buffers[i] = PinBufferForBlock(operation->rel,
1307 : operation->smgr,
1308 123285528 : operation->persistence,
1309 : operation->forknum,
1310 : blockNum + i,
1311 : operation->strategy,
1312 : &found);
1313 : }
1314 :
1315 123288300 : if (found)
1316 : {
1317 : /*
1318 : * We have a hit. If it's the first block in the requested range,
1319 : * we can return it immediately and report that WaitReadBuffers()
1320 : * does not need to be called. If the initial value of *nblocks
1321 : * was larger, the caller will have to call again for the rest.
1322 : */
1323 120479074 : if (i == 0)
1324 : {
1325 120476286 : *nblocks = 1;
1326 :
1327 : #ifdef USE_ASSERT_CHECKING
1328 :
1329 : /*
1330 : * Initialize enough of ReadBuffersOperation to make
1331 : * CheckReadBuffersOperation() work. Outside of assertions
1332 : * that's not necessary when no IO is issued.
1333 : */
1334 : operation->buffers = buffers;
1335 : operation->blocknum = blockNum;
1336 : operation->nblocks = 1;
1337 : operation->nblocks_done = 1;
1338 : CheckReadBuffersOperation(operation, true);
1339 : #endif
1340 120476286 : return false;
1341 : }
1342 :
1343 : /*
1344 : * Otherwise we already have an I/O to perform, but this block
1345 : * can't be included as it is already valid. Split the I/O here.
1346 : * There may or may not be more blocks requiring I/O after this
1347 : * one, we haven't checked, but they can't be contiguous with this
1348 : * one in the way. We'll leave this buffer pinned, forwarding it
1349 : * to the next call, avoiding the need to unpin it here and re-pin
1350 : * it in the next call.
1351 : */
1352 2788 : actual_nblocks = i;
1353 2788 : break;
1354 : }
1355 : else
1356 : {
1357 : /*
1358 : * Check how many blocks we can cover with the same IO. The smgr
1359 : * implementation might e.g. be limited due to a segment boundary.
1360 : */
1361 2809226 : if (i == 0 && actual_nblocks > 1)
1362 : {
1363 67298 : maxcombine = smgrmaxcombine(operation->smgr,
1364 : operation->forknum,
1365 : blockNum);
1366 67298 : if (unlikely(maxcombine < actual_nblocks))
1367 : {
1368 0 : elog(DEBUG2, "limiting nblocks at %u from %u to %u",
1369 : blockNum, actual_nblocks, maxcombine);
1370 0 : actual_nblocks = maxcombine;
1371 : }
1372 : }
1373 : }
1374 : }
1375 2477276 : *nblocks = actual_nblocks;
1376 :
1377 : /* Populate information needed for I/O. */
1378 2477276 : operation->buffers = buffers;
1379 2477276 : operation->blocknum = blockNum;
1380 2477276 : operation->flags = flags;
1381 2477276 : operation->nblocks = actual_nblocks;
1382 2477276 : operation->nblocks_done = 0;
1383 2477276 : pgaio_wref_clear(&operation->io_wref);
1384 :
1385 : /*
1386 : * When using AIO, start the IO in the background. If not, issue prefetch
1387 : * requests if desired by the caller.
1388 : *
1389 : * The reason we have a dedicated path for IOMETHOD_SYNC here is to
1390 : * de-risk the introduction of AIO somewhat. It's a large architectural
1391 : * change, with lots of chances for unanticipated performance effects.
1392 : *
1393 : * Use of IOMETHOD_SYNC already leads to not actually performing IO
1394 : * asynchronously, but without the check here we'd execute IO earlier than
1395 : * we used to. Eventually this IOMETHOD_SYNC specific path should go away.
1396 : */
1397 2477276 : if (io_method != IOMETHOD_SYNC)
1398 : {
1399 : /*
1400 : * Try to start IO asynchronously. It's possible that no IO needs to
1401 : * be started, if another backend already performed the IO.
1402 : *
1403 : * Note that if an IO is started, it might not cover the entire
1404 : * requested range, e.g. because an intermediary block has been read
1405 : * in by another backend. In that case any "trailing" buffers we
1406 : * already pinned above will be "forwarded" by read_stream.c to the
1407 : * next call to StartReadBuffers().
1408 : *
1409 : * This is signalled to the caller by decrementing *nblocks *and*
1410 : * reducing operation->nblocks. The latter is done here, but not below
1411 : * WaitReadBuffers(), as in WaitReadBuffers() we can't "shorten" the
1412 : * overall read size anymore, we need to retry until done in its
1413 : * entirety or until failed.
1414 : */
1415 2475116 : did_start_io = AsyncReadBuffers(operation, nblocks);
1416 :
1417 2475086 : operation->nblocks = *nblocks;
1418 : }
1419 : else
1420 : {
1421 2160 : operation->flags |= READ_BUFFERS_SYNCHRONOUSLY;
1422 :
1423 2160 : if (flags & READ_BUFFERS_ISSUE_ADVICE)
1424 : {
1425 : /*
1426 : * In theory we should only do this if PinBufferForBlock() had to
1427 : * allocate new buffers above. That way, if two calls to
1428 : * StartReadBuffers() were made for the same blocks before
1429 : * WaitReadBuffers(), only the first would issue the advice.
1430 : * That'd be a better simulation of true asynchronous I/O, which
1431 : * would only start the I/O once, but isn't done here for
1432 : * simplicity.
1433 : */
1434 4 : smgrprefetch(operation->smgr,
1435 : operation->forknum,
1436 : blockNum,
1437 : actual_nblocks);
1438 : }
1439 :
1440 : /*
1441 : * Indicate that WaitReadBuffers() should be called. WaitReadBuffers()
1442 : * will initiate the necessary IO.
1443 : */
1444 2160 : did_start_io = true;
1445 : }
1446 :
1447 2477246 : CheckReadBuffersOperation(operation, !did_start_io);
1448 :
1449 2477246 : return did_start_io;
1450 : }
1451 :
1452 : /*
1453 : * Begin reading a range of blocks beginning at blockNum and extending for
1454 : * *nblocks. *nblocks and the buffers array are in/out parameters. On entry,
1455 : * the buffers elements covered by *nblocks must hold either InvalidBuffer or
1456 : * buffers forwarded by an earlier call to StartReadBuffers() that was split
1457 : * and is now being continued. On return, *nblocks holds the number of blocks
1458 : * accepted by this operation. If it is less than the original number then
1459 : * this operation has been split, but buffer elements up to the original
1460 : * requested size may hold forwarded buffers to be used for a continuing
1461 : * operation. The caller must either start a new I/O beginning at the block
1462 : * immediately following the blocks accepted by this call and pass those
1463 : * buffers back in, or release them if it chooses not to. It shouldn't make
1464 : * any other use of or assumptions about forwarded buffers.
1465 : *
1466 : * If false is returned, no I/O is necessary and the buffers covered by
1467 : * *nblocks on exit are valid and ready to be accessed. If true is returned,
1468 : * an I/O has been started, and WaitReadBuffers() must be called with the same
1469 : * operation object before the buffers covered by *nblocks on exit can be
1470 : * accessed. Along with the operation object, the caller-supplied array of
1471 : * buffers must remain valid until WaitReadBuffers() is called, and any
1472 : * forwarded buffers must also be preserved for a continuing call unless
1473 : * they are explicitly released.
1474 : */
1475 : bool
1476 3922122 : StartReadBuffers(ReadBuffersOperation *operation,
1477 : Buffer *buffers,
1478 : BlockNumber blockNum,
1479 : int *nblocks,
1480 : int flags)
1481 : {
1482 3922122 : return StartReadBuffersImpl(operation, buffers, blockNum, nblocks, flags,
1483 : true /* expect forwarded buffers */ );
1484 : }
1485 :
1486 : /*
1487 : * Single block version of the StartReadBuffers(). This might save a few
1488 : * instructions when called from another translation unit, because it is
1489 : * specialized for nblocks == 1.
1490 : *
1491 : * This version does not support "forwarded" buffers: they cannot be created
1492 : * by reading only one block and *buffer is ignored on entry.
1493 : */
1494 : bool
1495 119031452 : StartReadBuffer(ReadBuffersOperation *operation,
1496 : Buffer *buffer,
1497 : BlockNumber blocknum,
1498 : int flags)
1499 : {
1500 119031452 : int nblocks = 1;
1501 : bool result;
1502 :
1503 119031452 : result = StartReadBuffersImpl(operation, buffer, blocknum, &nblocks, flags,
1504 : false /* single block, no forwarding */ );
1505 : Assert(nblocks == 1); /* single block can't be short */
1506 :
1507 119031422 : return result;
1508 : }
1509 :
1510 : /*
1511 : * Perform sanity checks on the ReadBuffersOperation.
1512 : */
1513 : static void
1514 7402496 : CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete)
1515 : {
1516 : #ifdef USE_ASSERT_CHECKING
1517 : Assert(operation->nblocks_done <= operation->nblocks);
1518 : Assert(!is_complete || operation->nblocks == operation->nblocks_done);
1519 :
1520 : for (int i = 0; i < operation->nblocks; i++)
1521 : {
1522 : Buffer buffer = operation->buffers[i];
1523 : BufferDesc *buf_hdr = BufferIsLocal(buffer) ?
1524 : GetLocalBufferDescriptor(-buffer - 1) :
1525 : GetBufferDescriptor(buffer - 1);
1526 :
1527 : Assert(BufferGetBlockNumber(buffer) == operation->blocknum + i);
1528 : Assert(pg_atomic_read_u32(&buf_hdr->state) & BM_TAG_VALID);
1529 :
1530 : if (i < operation->nblocks_done)
1531 : Assert(pg_atomic_read_u32(&buf_hdr->state) & BM_VALID);
1532 : }
1533 : #endif
1534 7402496 : }
1535 :
1536 : /* helper for ReadBuffersCanStartIO(), to avoid repetition */
1537 : static inline bool
1538 2809262 : ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
1539 : {
1540 2809262 : if (BufferIsLocal(buffer))
1541 16734 : return StartLocalBufferIO(GetLocalBufferDescriptor(-buffer - 1),
1542 : true, nowait);
1543 : else
1544 2792528 : return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
1545 : }
1546 :
1547 : /*
1548 : * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
1549 : */
1550 : static inline bool
1551 2809262 : ReadBuffersCanStartIO(Buffer buffer, bool nowait)
1552 : {
1553 : /*
1554 : * If this backend currently has staged IO, we need to submit the pending
1555 : * IO before waiting for the right to issue IO, to avoid the potential for
1556 : * deadlocks (and, more commonly, unnecessary delays for other backends).
1557 : */
1558 2809262 : if (!nowait && pgaio_have_staged())
1559 : {
1560 1132 : if (ReadBuffersCanStartIOOnce(buffer, true))
1561 1132 : return true;
1562 :
1563 : /*
1564 : * Unfortunately StartBufferIO() returning false doesn't allow to
1565 : * distinguish between the buffer already being valid and IO already
1566 : * being in progress. Since IO already being in progress is quite
1567 : * rare, this approach seems fine.
1568 : */
1569 0 : pgaio_submit_staged();
1570 : }
1571 :
1572 2808130 : return ReadBuffersCanStartIOOnce(buffer, nowait);
1573 : }
1574 :
1575 : /*
1576 : * Helper for WaitReadBuffers() that processes the results of a readv
1577 : * operation, raising an error if necessary.
1578 : */
1579 : static void
1580 2461584 : ProcessReadBuffersResult(ReadBuffersOperation *operation)
1581 : {
1582 2461584 : PgAioReturn *aio_ret = &operation->io_return;
1583 2461584 : PgAioResultStatus rs = aio_ret->result.status;
1584 2461584 : int newly_read_blocks = 0;
1585 :
1586 : Assert(pgaio_wref_valid(&operation->io_wref));
1587 : Assert(aio_ret->result.status != PGAIO_RS_UNKNOWN);
1588 :
1589 : /*
1590 : * SMGR reports the number of blocks successfully read as the result of
1591 : * the IO operation. Thus we can simply add that to ->nblocks_done.
1592 : */
1593 :
1594 2461584 : if (likely(rs != PGAIO_RS_ERROR))
1595 2461526 : newly_read_blocks = aio_ret->result.result;
1596 :
1597 2461584 : if (rs == PGAIO_RS_ERROR || rs == PGAIO_RS_WARNING)
1598 90 : pgaio_result_report(aio_ret->result, &aio_ret->target_data,
1599 : rs == PGAIO_RS_ERROR ? ERROR : WARNING);
1600 2461494 : else if (aio_ret->result.status == PGAIO_RS_PARTIAL)
1601 : {
1602 : /*
1603 : * We'll retry, so we just emit a debug message to the server log (or
1604 : * not even that in prod scenarios).
1605 : */
1606 20 : pgaio_result_report(aio_ret->result, &aio_ret->target_data, DEBUG1);
1607 20 : elog(DEBUG3, "partial read, will retry");
1608 : }
1609 :
1610 : Assert(newly_read_blocks > 0);
1611 : Assert(newly_read_blocks <= MAX_IO_COMBINE_LIMIT);
1612 :
1613 2461526 : operation->nblocks_done += newly_read_blocks;
1614 :
1615 : Assert(operation->nblocks_done <= operation->nblocks);
1616 2461526 : }
1617 :
1618 : void
1619 2461564 : WaitReadBuffers(ReadBuffersOperation *operation)
1620 : {
1621 2461564 : PgAioReturn *aio_ret = &operation->io_return;
1622 : IOContext io_context;
1623 : IOObject io_object;
1624 :
1625 2461564 : if (operation->persistence == RELPERSISTENCE_TEMP)
1626 : {
1627 2980 : io_context = IOCONTEXT_NORMAL;
1628 2980 : io_object = IOOBJECT_TEMP_RELATION;
1629 : }
1630 : else
1631 : {
1632 2458584 : io_context = IOContextForStrategy(operation->strategy);
1633 2458584 : io_object = IOOBJECT_RELATION;
1634 : }
1635 :
1636 : /*
1637 : * If we get here without an IO operation having been issued, the
1638 : * io_method == IOMETHOD_SYNC path must have been used. Otherwise the
1639 : * caller should not have called WaitReadBuffers().
1640 : *
1641 : * In the case of IOMETHOD_SYNC, we start - as we used to before the
1642 : * introducing of AIO - the IO in WaitReadBuffers(). This is done as part
1643 : * of the retry logic below, no extra code is required.
1644 : *
1645 : * This path is expected to eventually go away.
1646 : */
1647 2461564 : if (!pgaio_wref_valid(&operation->io_wref) && io_method != IOMETHOD_SYNC)
1648 0 : elog(ERROR, "waiting for read operation that didn't read");
1649 :
1650 : /*
1651 : * To handle partial reads, and IOMETHOD_SYNC, we re-issue IO until we're
1652 : * done. We may need multiple retries, not just because we could get
1653 : * multiple partial reads, but also because some of the remaining
1654 : * to-be-read buffers may have been read in by other backends, limiting
1655 : * the IO size.
1656 : */
1657 : while (true)
1658 2180 : {
1659 : int ignored_nblocks_progress;
1660 :
1661 2463744 : CheckReadBuffersOperation(operation, false);
1662 :
1663 : /*
1664 : * If there is an IO associated with the operation, we may need to
1665 : * wait for it.
1666 : */
1667 2463744 : if (pgaio_wref_valid(&operation->io_wref))
1668 : {
1669 : /*
1670 : * Track the time spent waiting for the IO to complete. As
1671 : * tracking a wait even if we don't actually need to wait
1672 : *
1673 : * a) is not cheap, due to the timestamping overhead
1674 : *
1675 : * b) reports some time as waiting, even if we never waited
1676 : *
1677 : * we first check if we already know the IO is complete.
1678 : */
1679 2461584 : if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
1680 1073204 : !pgaio_wref_check_done(&operation->io_wref))
1681 : {
1682 395110 : instr_time io_start = pgstat_prepare_io_time(track_io_timing);
1683 :
1684 395110 : pgaio_wref_wait(&operation->io_wref);
1685 :
1686 : /*
1687 : * The IO operation itself was already counted earlier, in
1688 : * AsyncReadBuffers(), this just accounts for the wait time.
1689 : */
1690 395110 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1691 : io_start, 0, 0);
1692 : }
1693 : else
1694 : {
1695 : Assert(pgaio_wref_check_done(&operation->io_wref));
1696 : }
1697 :
1698 : /*
1699 : * We now are sure the IO completed. Check the results. This
1700 : * includes reporting on errors if there were any.
1701 : */
1702 2461584 : ProcessReadBuffersResult(operation);
1703 : }
1704 :
1705 : /*
1706 : * Most of the time, the one IO we already started, will read in
1707 : * everything. But we need to deal with partial reads and buffers not
1708 : * needing IO anymore.
1709 : */
1710 2463686 : if (operation->nblocks_done == operation->nblocks)
1711 2461506 : break;
1712 :
1713 2180 : CHECK_FOR_INTERRUPTS();
1714 :
1715 : /*
1716 : * This may only complete the IO partially, either because some
1717 : * buffers were already valid, or because of a partial read.
1718 : *
1719 : * NB: In contrast to after the AsyncReadBuffers() call in
1720 : * StartReadBuffers(), we do *not* reduce
1721 : * ReadBuffersOperation->nblocks here, callers expect the full
1722 : * operation to be completed at this point (as more operations may
1723 : * have been queued).
1724 : */
1725 2180 : AsyncReadBuffers(operation, &ignored_nblocks_progress);
1726 : }
1727 :
1728 2461506 : CheckReadBuffersOperation(operation, true);
1729 :
1730 : /* NB: READ_DONE tracepoint was already executed in completion callback */
1731 2461506 : }
1732 :
1733 : /*
1734 : * Initiate IO for the ReadBuffersOperation
1735 : *
1736 : * This function only starts a single IO at a time. The size of the IO may be
1737 : * limited to below the to-be-read blocks, if one of the buffers has
1738 : * concurrently been read in. If the first to-be-read buffer is already valid,
1739 : * no IO will be issued.
1740 : *
1741 : * To support retries after partial reads, the first operation->nblocks_done
1742 : * buffers are skipped.
1743 : *
1744 : * On return *nblocks_progress is updated to reflect the number of buffers
1745 : * affected by the call. If the first buffer is valid, *nblocks_progress is
1746 : * set to 1 and operation->nblocks_done is incremented.
1747 : *
1748 : * Returns true if IO was initiated, false if no IO was necessary.
1749 : */
1750 : static bool
1751 2477296 : AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
1752 : {
1753 2477296 : Buffer *buffers = &operation->buffers[0];
1754 2477296 : int flags = operation->flags;
1755 2477296 : BlockNumber blocknum = operation->blocknum;
1756 2477296 : ForkNumber forknum = operation->forknum;
1757 2477296 : char persistence = operation->persistence;
1758 2477296 : int16 nblocks_done = operation->nblocks_done;
1759 2477296 : Buffer *io_buffers = &operation->buffers[nblocks_done];
1760 2477296 : int io_buffers_len = 0;
1761 : PgAioHandle *ioh;
1762 2477296 : uint32 ioh_flags = 0;
1763 : void *io_pages[MAX_IO_COMBINE_LIMIT];
1764 : IOContext io_context;
1765 : IOObject io_object;
1766 : bool did_start_io;
1767 :
1768 : /*
1769 : * When this IO is executed synchronously, either because the caller will
1770 : * immediately block waiting for the IO or because IOMETHOD_SYNC is used,
1771 : * the AIO subsystem needs to know.
1772 : */
1773 2477296 : if (flags & READ_BUFFERS_SYNCHRONOUSLY)
1774 1382896 : ioh_flags |= PGAIO_HF_SYNCHRONOUS;
1775 :
1776 2477296 : if (persistence == RELPERSISTENCE_TEMP)
1777 : {
1778 3568 : io_context = IOCONTEXT_NORMAL;
1779 3568 : io_object = IOOBJECT_TEMP_RELATION;
1780 3568 : ioh_flags |= PGAIO_HF_REFERENCES_LOCAL;
1781 : }
1782 : else
1783 : {
1784 2473728 : io_context = IOContextForStrategy(operation->strategy);
1785 2473728 : io_object = IOOBJECT_RELATION;
1786 : }
1787 :
1788 : /*
1789 : * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
1790 : * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
1791 : * set globally, but on a per-session basis. The completion callback,
1792 : * which may be run in other processes, e.g. in IO workers, may have a
1793 : * different value of the zero_damaged_pages GUC.
1794 : *
1795 : * XXX: We probably should eventually use a different flag for
1796 : * zero_damaged_pages, so we can report different log levels / error codes
1797 : * for zero_damaged_pages and ZERO_ON_ERROR.
1798 : */
1799 2477296 : if (zero_damaged_pages)
1800 32 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1801 :
1802 : /*
1803 : * For the same reason as with zero_damaged_pages we need to use this
1804 : * backend's ignore_checksum_failure value.
1805 : */
1806 2477296 : if (ignore_checksum_failure)
1807 16 : flags |= READ_BUFFERS_IGNORE_CHECKSUM_FAILURES;
1808 :
1809 :
1810 : /*
1811 : * To be allowed to report stats in the local completion callback we need
1812 : * to prepare to report stats now. This ensures we can safely report the
1813 : * checksum failure even in a critical section.
1814 : */
1815 2477296 : pgstat_prepare_report_checksum_failure(operation->smgr->smgr_rlocator.locator.dbOid);
1816 :
1817 : /*
1818 : * Get IO handle before ReadBuffersCanStartIO(), as pgaio_io_acquire()
1819 : * might block, which we don't want after setting IO_IN_PROGRESS.
1820 : *
1821 : * If we need to wait for IO before we can get a handle, submit
1822 : * already-staged IO first, so that other backends don't need to wait.
1823 : * There wouldn't be a deadlock risk, as pgaio_io_acquire() just needs to
1824 : * wait for already submitted IO, which doesn't require additional locks,
1825 : * but it could still cause undesirable waits.
1826 : *
1827 : * A secondary benefit is that this would allow us to measure the time in
1828 : * pgaio_io_acquire() without causing undue timer overhead in the common,
1829 : * non-blocking, case. However, currently the pgstats infrastructure
1830 : * doesn't really allow that, as it a) asserts that an operation can't
1831 : * have time without operations b) doesn't have an API to report
1832 : * "accumulated" time.
1833 : */
1834 2477296 : ioh = pgaio_io_acquire_nb(CurrentResourceOwner, &operation->io_return);
1835 2477296 : if (unlikely(!ioh))
1836 : {
1837 6030 : pgaio_submit_staged();
1838 :
1839 6030 : ioh = pgaio_io_acquire(CurrentResourceOwner, &operation->io_return);
1840 : }
1841 :
1842 : /*
1843 : * Check if we can start IO on the first to-be-read buffer.
1844 : *
1845 : * If an I/O is already in progress in another backend, we want to wait
1846 : * for the outcome: either done, or something went wrong and we will
1847 : * retry.
1848 : */
1849 2477296 : if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
1850 : {
1851 : /*
1852 : * Someone else has already completed this block, we're done.
1853 : *
1854 : * When IO is necessary, ->nblocks_done is updated in
1855 : * ProcessReadBuffersResult(), but that is not called if no IO is
1856 : * necessary. Thus update here.
1857 : */
1858 15084 : operation->nblocks_done += 1;
1859 15084 : *nblocks_progress = 1;
1860 :
1861 15084 : pgaio_io_release(ioh);
1862 15084 : pgaio_wref_clear(&operation->io_wref);
1863 15084 : did_start_io = false;
1864 :
1865 : /*
1866 : * Report and track this as a 'hit' for this backend, even though it
1867 : * must have started out as a miss in PinBufferForBlock(). The other
1868 : * backend will track this as a 'read'.
1869 : */
1870 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forknum, blocknum + operation->nblocks_done,
1871 : operation->smgr->smgr_rlocator.locator.spcOid,
1872 : operation->smgr->smgr_rlocator.locator.dbOid,
1873 : operation->smgr->smgr_rlocator.locator.relNumber,
1874 : operation->smgr->smgr_rlocator.backend,
1875 : true);
1876 :
1877 15084 : if (persistence == RELPERSISTENCE_TEMP)
1878 0 : pgBufferUsage.local_blks_hit += 1;
1879 : else
1880 15084 : pgBufferUsage.shared_blks_hit += 1;
1881 :
1882 15084 : if (operation->rel)
1883 15084 : pgstat_count_buffer_hit(operation->rel);
1884 :
1885 15084 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1886 :
1887 15084 : if (VacuumCostActive)
1888 0 : VacuumCostBalance += VacuumCostPageHit;
1889 : }
1890 : else
1891 : {
1892 : instr_time io_start;
1893 :
1894 : /* We found a buffer that we need to read in. */
1895 : Assert(io_buffers[0] == buffers[nblocks_done]);
1896 2462212 : io_pages[0] = BufferGetBlock(buffers[nblocks_done]);
1897 2462212 : io_buffers_len = 1;
1898 :
1899 : /*
1900 : * How many neighboring-on-disk blocks can we scatter-read into other
1901 : * buffers at the same time? In this case we don't wait if we see an
1902 : * I/O already in progress. We already set BM_IO_IN_PROGRESS for the
1903 : * head block, so we should get on with that I/O as soon as possible.
1904 : */
1905 2794178 : for (int i = nblocks_done + 1; i < operation->nblocks; i++)
1906 : {
1907 331966 : if (!ReadBuffersCanStartIO(buffers[i], true))
1908 0 : break;
1909 : /* Must be consecutive block numbers. */
1910 : Assert(BufferGetBlockNumber(buffers[i - 1]) ==
1911 : BufferGetBlockNumber(buffers[i]) - 1);
1912 : Assert(io_buffers[io_buffers_len] == buffers[i]);
1913 :
1914 331966 : io_pages[io_buffers_len++] = BufferGetBlock(buffers[i]);
1915 : }
1916 :
1917 : /* get a reference to wait for in WaitReadBuffers() */
1918 2462212 : pgaio_io_get_wref(ioh, &operation->io_wref);
1919 :
1920 : /* provide the list of buffers to the completion callbacks */
1921 2462212 : pgaio_io_set_handle_data_32(ioh, (uint32 *) io_buffers, io_buffers_len);
1922 :
1923 2462212 : pgaio_io_register_callbacks(ioh,
1924 : persistence == RELPERSISTENCE_TEMP ?
1925 : PGAIO_HCB_LOCAL_BUFFER_READV :
1926 : PGAIO_HCB_SHARED_BUFFER_READV,
1927 : flags);
1928 :
1929 2462212 : pgaio_io_set_flag(ioh, ioh_flags);
1930 :
1931 : /* ---
1932 : * Even though we're trying to issue IO asynchronously, track the time
1933 : * in smgrstartreadv():
1934 : * - if io_method == IOMETHOD_SYNC, we will always perform the IO
1935 : * immediately
1936 : * - the io method might not support the IO (e.g. worker IO for a temp
1937 : * table)
1938 : * ---
1939 : */
1940 2462212 : io_start = pgstat_prepare_io_time(track_io_timing);
1941 2462212 : smgrstartreadv(ioh, operation->smgr, forknum,
1942 : blocknum + nblocks_done,
1943 : io_pages, io_buffers_len);
1944 2462182 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1945 2462182 : io_start, 1, io_buffers_len * BLCKSZ);
1946 :
1947 2462182 : if (persistence == RELPERSISTENCE_TEMP)
1948 3568 : pgBufferUsage.local_blks_read += io_buffers_len;
1949 : else
1950 2458614 : pgBufferUsage.shared_blks_read += io_buffers_len;
1951 :
1952 : /*
1953 : * Track vacuum cost when issuing IO, not after waiting for it.
1954 : * Otherwise we could end up issuing a lot of IO in a short timespan,
1955 : * despite a low cost limit.
1956 : */
1957 2462182 : if (VacuumCostActive)
1958 31152 : VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
1959 :
1960 2462182 : *nblocks_progress = io_buffers_len;
1961 2462182 : did_start_io = true;
1962 : }
1963 :
1964 2477266 : return did_start_io;
1965 : }
1966 :
1967 : /*
1968 : * BufferAlloc -- subroutine for PinBufferForBlock. Handles lookup of a shared
1969 : * buffer. If no buffer exists already, selects a replacement victim and
1970 : * evicts the old page, but does NOT read in new page.
1971 : *
1972 : * "strategy" can be a buffer replacement strategy object, or NULL for
1973 : * the default strategy. The selected buffer's usage_count is advanced when
1974 : * using the default strategy, but otherwise possibly not (see PinBuffer).
1975 : *
1976 : * The returned buffer is pinned and is already marked as holding the
1977 : * desired page. If it already did have the desired page, *foundPtr is
1978 : * set true. Otherwise, *foundPtr is set false.
1979 : *
1980 : * io_context is passed as an output parameter to avoid calling
1981 : * IOContextForStrategy() when there is a shared buffers hit and no IO
1982 : * statistics need be captured.
1983 : *
1984 : * No locks are held either at entry or exit.
1985 : */
1986 : static pg_attribute_always_inline BufferDesc *
1987 121360172 : BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
1988 : BlockNumber blockNum,
1989 : BufferAccessStrategy strategy,
1990 : bool *foundPtr, IOContext io_context)
1991 : {
1992 : BufferTag newTag; /* identity of requested block */
1993 : uint32 newHash; /* hash value for newTag */
1994 : LWLock *newPartitionLock; /* buffer partition lock for it */
1995 : int existing_buf_id;
1996 : Buffer victim_buffer;
1997 : BufferDesc *victim_buf_hdr;
1998 : uint32 victim_buf_state;
1999 :
2000 : /* Make sure we will have room to remember the buffer pin */
2001 121360172 : ResourceOwnerEnlarge(CurrentResourceOwner);
2002 121360172 : ReservePrivateRefCountEntry();
2003 :
2004 : /* create a tag so we can lookup the buffer */
2005 121360172 : InitBufferTag(&newTag, &smgr->smgr_rlocator.locator, forkNum, blockNum);
2006 :
2007 : /* determine its hash code and partition lock ID */
2008 121360172 : newHash = BufTableHashCode(&newTag);
2009 121360172 : newPartitionLock = BufMappingPartitionLock(newHash);
2010 :
2011 : /* see if the block is in the buffer pool already */
2012 121360172 : LWLockAcquire(newPartitionLock, LW_SHARED);
2013 121360172 : existing_buf_id = BufTableLookup(&newTag, newHash);
2014 121360172 : if (existing_buf_id >= 0)
2015 : {
2016 : BufferDesc *buf;
2017 : bool valid;
2018 :
2019 : /*
2020 : * Found it. Now, pin the buffer so no one can steal it from the
2021 : * buffer pool, and check to see if the correct data has been loaded
2022 : * into the buffer.
2023 : */
2024 118024422 : buf = GetBufferDescriptor(existing_buf_id);
2025 :
2026 118024422 : valid = PinBuffer(buf, strategy, false);
2027 :
2028 : /* Can release the mapping lock as soon as we've pinned it */
2029 118024422 : LWLockRelease(newPartitionLock);
2030 :
2031 118024422 : *foundPtr = true;
2032 :
2033 118024422 : if (!valid)
2034 : {
2035 : /*
2036 : * We can only get here if (a) someone else is still reading in
2037 : * the page, (b) a previous read attempt failed, or (c) someone
2038 : * called StartReadBuffers() but not yet WaitReadBuffers().
2039 : */
2040 13022 : *foundPtr = false;
2041 : }
2042 :
2043 118024422 : return buf;
2044 : }
2045 :
2046 : /*
2047 : * Didn't find it in the buffer pool. We'll have to initialize a new
2048 : * buffer. Remember to unlock the mapping lock while doing the work.
2049 : */
2050 3335750 : LWLockRelease(newPartitionLock);
2051 :
2052 : /*
2053 : * Acquire a victim buffer. Somebody else might try to do the same, we
2054 : * don't hold any conflicting locks. If so we'll have to undo our work
2055 : * later.
2056 : */
2057 3335750 : victim_buffer = GetVictimBuffer(strategy, io_context);
2058 3335750 : victim_buf_hdr = GetBufferDescriptor(victim_buffer - 1);
2059 :
2060 : /*
2061 : * Try to make a hashtable entry for the buffer under its new tag. If
2062 : * somebody else inserted another buffer for the tag, we'll release the
2063 : * victim buffer we acquired and use the already inserted one.
2064 : */
2065 3335750 : LWLockAcquire(newPartitionLock, LW_EXCLUSIVE);
2066 3335750 : existing_buf_id = BufTableInsert(&newTag, newHash, victim_buf_hdr->buf_id);
2067 3335750 : if (existing_buf_id >= 0)
2068 : {
2069 : BufferDesc *existing_buf_hdr;
2070 : bool valid;
2071 :
2072 : /*
2073 : * Got a collision. Someone has already done what we were about to do.
2074 : * We'll just handle this as if it were found in the buffer pool in
2075 : * the first place. First, give up the buffer we were planning to
2076 : * use.
2077 : *
2078 : * We could do this after releasing the partition lock, but then we'd
2079 : * have to call ResourceOwnerEnlarge() & ReservePrivateRefCountEntry()
2080 : * before acquiring the lock, for the rare case of such a collision.
2081 : */
2082 2830 : UnpinBuffer(victim_buf_hdr);
2083 :
2084 : /* remaining code should match code at top of routine */
2085 :
2086 2830 : existing_buf_hdr = GetBufferDescriptor(existing_buf_id);
2087 :
2088 2830 : valid = PinBuffer(existing_buf_hdr, strategy, false);
2089 :
2090 : /* Can release the mapping lock as soon as we've pinned it */
2091 2830 : LWLockRelease(newPartitionLock);
2092 :
2093 2830 : *foundPtr = true;
2094 :
2095 2830 : if (!valid)
2096 : {
2097 : /*
2098 : * We can only get here if (a) someone else is still reading in
2099 : * the page, (b) a previous read attempt failed, or (c) someone
2100 : * called StartReadBuffers() but not yet WaitReadBuffers().
2101 : */
2102 2158 : *foundPtr = false;
2103 : }
2104 :
2105 2830 : return existing_buf_hdr;
2106 : }
2107 :
2108 : /*
2109 : * Need to lock the buffer header too in order to change its tag.
2110 : */
2111 3332920 : victim_buf_state = LockBufHdr(victim_buf_hdr);
2112 :
2113 : /* some sanity checks while we hold the buffer header lock */
2114 : Assert(BUF_STATE_GET_REFCOUNT(victim_buf_state) == 1);
2115 : Assert(!(victim_buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY | BM_IO_IN_PROGRESS)));
2116 :
2117 3332920 : victim_buf_hdr->tag = newTag;
2118 :
2119 : /*
2120 : * Make sure BM_PERMANENT is set for buffers that must be written at every
2121 : * checkpoint. Unlogged buffers only need to be written at shutdown
2122 : * checkpoints, except for their "init" forks, which need to be treated
2123 : * just like permanent relations.
2124 : */
2125 3332920 : victim_buf_state |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2126 3332920 : if (relpersistence == RELPERSISTENCE_PERMANENT || forkNum == INIT_FORKNUM)
2127 3332238 : victim_buf_state |= BM_PERMANENT;
2128 :
2129 3332920 : UnlockBufHdr(victim_buf_hdr, victim_buf_state);
2130 :
2131 3332920 : LWLockRelease(newPartitionLock);
2132 :
2133 : /*
2134 : * Buffer contents are currently invalid.
2135 : */
2136 3332920 : *foundPtr = false;
2137 :
2138 3332920 : return victim_buf_hdr;
2139 : }
2140 :
2141 : /*
2142 : * InvalidateBuffer -- mark a shared buffer invalid.
2143 : *
2144 : * The buffer header spinlock must be held at entry. We drop it before
2145 : * returning. (This is sane because the caller must have locked the
2146 : * buffer in order to be sure it should be dropped.)
2147 : *
2148 : * This is used only in contexts such as dropping a relation. We assume
2149 : * that no other backend could possibly be interested in using the page,
2150 : * so the only reason the buffer might be pinned is if someone else is
2151 : * trying to write it out. We have to let them finish before we can
2152 : * reclaim the buffer.
2153 : *
2154 : * The buffer could get reclaimed by someone else while we are waiting
2155 : * to acquire the necessary locks; if so, don't mess it up.
2156 : */
2157 : static void
2158 210724 : InvalidateBuffer(BufferDesc *buf)
2159 : {
2160 : BufferTag oldTag;
2161 : uint32 oldHash; /* hash value for oldTag */
2162 : LWLock *oldPartitionLock; /* buffer partition lock for it */
2163 : uint32 oldFlags;
2164 : uint32 buf_state;
2165 :
2166 : /* Save the original buffer tag before dropping the spinlock */
2167 210724 : oldTag = buf->tag;
2168 :
2169 210724 : buf_state = pg_atomic_read_u32(&buf->state);
2170 : Assert(buf_state & BM_LOCKED);
2171 210724 : UnlockBufHdr(buf, buf_state);
2172 :
2173 : /*
2174 : * Need to compute the old tag's hashcode and partition lock ID. XXX is it
2175 : * worth storing the hashcode in BufferDesc so we need not recompute it
2176 : * here? Probably not.
2177 : */
2178 210724 : oldHash = BufTableHashCode(&oldTag);
2179 210724 : oldPartitionLock = BufMappingPartitionLock(oldHash);
2180 :
2181 210726 : retry:
2182 :
2183 : /*
2184 : * Acquire exclusive mapping lock in preparation for changing the buffer's
2185 : * association.
2186 : */
2187 210726 : LWLockAcquire(oldPartitionLock, LW_EXCLUSIVE);
2188 :
2189 : /* Re-lock the buffer header */
2190 210726 : buf_state = LockBufHdr(buf);
2191 :
2192 : /* If it's changed while we were waiting for lock, do nothing */
2193 210726 : if (!BufferTagsEqual(&buf->tag, &oldTag))
2194 : {
2195 2 : UnlockBufHdr(buf, buf_state);
2196 2 : LWLockRelease(oldPartitionLock);
2197 2 : return;
2198 : }
2199 :
2200 : /*
2201 : * We assume the reason for it to be pinned is that either we were
2202 : * asynchronously reading the page in before erroring out or someone else
2203 : * is flushing the page out. Wait for the IO to finish. (This could be
2204 : * an infinite loop if the refcount is messed up... it would be nice to
2205 : * time out after awhile, but there seems no way to be sure how many loops
2206 : * may be needed. Note that if the other guy has pinned the buffer but
2207 : * not yet done StartBufferIO, WaitIO will fall through and we'll
2208 : * effectively be busy-looping here.)
2209 : */
2210 210724 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 0)
2211 : {
2212 2 : UnlockBufHdr(buf, buf_state);
2213 2 : LWLockRelease(oldPartitionLock);
2214 : /* safety check: should definitely not be our *own* pin */
2215 2 : if (GetPrivateRefCount(BufferDescriptorGetBuffer(buf)) > 0)
2216 0 : elog(ERROR, "buffer is pinned in InvalidateBuffer");
2217 2 : WaitIO(buf);
2218 2 : goto retry;
2219 : }
2220 :
2221 : /*
2222 : * Clear out the buffer's tag and flags. We must do this to ensure that
2223 : * linear scans of the buffer array don't think the buffer is valid.
2224 : */
2225 210722 : oldFlags = buf_state & BUF_FLAG_MASK;
2226 210722 : ClearBufferTag(&buf->tag);
2227 210722 : buf_state &= ~(BUF_FLAG_MASK | BUF_USAGECOUNT_MASK);
2228 210722 : UnlockBufHdr(buf, buf_state);
2229 :
2230 : /*
2231 : * Remove the buffer from the lookup hashtable, if it was in there.
2232 : */
2233 210722 : if (oldFlags & BM_TAG_VALID)
2234 210722 : BufTableDelete(&oldTag, oldHash);
2235 :
2236 : /*
2237 : * Done with mapping lock.
2238 : */
2239 210722 : LWLockRelease(oldPartitionLock);
2240 : }
2241 :
2242 : /*
2243 : * Helper routine for GetVictimBuffer()
2244 : *
2245 : * Needs to be called on a buffer with a valid tag, pinned, but without the
2246 : * buffer header spinlock held.
2247 : *
2248 : * Returns true if the buffer can be reused, in which case the buffer is only
2249 : * pinned by this backend and marked as invalid, false otherwise.
2250 : */
2251 : static bool
2252 2340056 : InvalidateVictimBuffer(BufferDesc *buf_hdr)
2253 : {
2254 : uint32 buf_state;
2255 : uint32 hash;
2256 : LWLock *partition_lock;
2257 : BufferTag tag;
2258 :
2259 : Assert(GetPrivateRefCount(BufferDescriptorGetBuffer(buf_hdr)) == 1);
2260 :
2261 : /* have buffer pinned, so it's safe to read tag without lock */
2262 2340056 : tag = buf_hdr->tag;
2263 :
2264 2340056 : hash = BufTableHashCode(&tag);
2265 2340056 : partition_lock = BufMappingPartitionLock(hash);
2266 :
2267 2340056 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2268 :
2269 : /* lock the buffer header */
2270 2340056 : buf_state = LockBufHdr(buf_hdr);
2271 :
2272 : /*
2273 : * We have the buffer pinned nobody else should have been able to unset
2274 : * this concurrently.
2275 : */
2276 : Assert(buf_state & BM_TAG_VALID);
2277 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2278 : Assert(BufferTagsEqual(&buf_hdr->tag, &tag));
2279 :
2280 : /*
2281 : * If somebody else pinned the buffer since, or even worse, dirtied it,
2282 : * give up on this buffer: It's clearly in use.
2283 : */
2284 2340056 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 1 || (buf_state & BM_DIRTY))
2285 : {
2286 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2287 :
2288 850 : UnlockBufHdr(buf_hdr, buf_state);
2289 850 : LWLockRelease(partition_lock);
2290 :
2291 850 : return false;
2292 : }
2293 :
2294 : /*
2295 : * Clear out the buffer's tag and flags and usagecount. This is not
2296 : * strictly required, as BM_TAG_VALID/BM_VALID needs to be checked before
2297 : * doing anything with the buffer. But currently it's beneficial, as the
2298 : * cheaper pre-check for several linear scans of shared buffers use the
2299 : * tag (see e.g. FlushDatabaseBuffers()).
2300 : */
2301 2339206 : ClearBufferTag(&buf_hdr->tag);
2302 2339206 : buf_state &= ~(BUF_FLAG_MASK | BUF_USAGECOUNT_MASK);
2303 2339206 : UnlockBufHdr(buf_hdr, buf_state);
2304 :
2305 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2306 :
2307 : /* finally delete buffer from the buffer mapping table */
2308 2339206 : BufTableDelete(&tag, hash);
2309 :
2310 2339206 : LWLockRelease(partition_lock);
2311 :
2312 : Assert(!(buf_state & (BM_DIRTY | BM_VALID | BM_TAG_VALID)));
2313 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2314 : Assert(BUF_STATE_GET_REFCOUNT(pg_atomic_read_u32(&buf_hdr->state)) > 0);
2315 :
2316 2339206 : return true;
2317 : }
2318 :
2319 : static Buffer
2320 3781374 : GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context)
2321 : {
2322 : BufferDesc *buf_hdr;
2323 : Buffer buf;
2324 : uint32 buf_state;
2325 : bool from_ring;
2326 :
2327 : /*
2328 : * Ensure, before we pin a victim buffer, that there's a free refcount
2329 : * entry and resource owner slot for the pin.
2330 : */
2331 3781374 : ReservePrivateRefCountEntry();
2332 3781374 : ResourceOwnerEnlarge(CurrentResourceOwner);
2333 :
2334 : /* we return here if a prospective victim buffer gets used concurrently */
2335 12216 : again:
2336 :
2337 : /*
2338 : * Select a victim buffer. The buffer is returned pinned and owned by
2339 : * this backend.
2340 : */
2341 3793590 : buf_hdr = StrategyGetBuffer(strategy, &buf_state, &from_ring);
2342 3793590 : buf = BufferDescriptorGetBuffer(buf_hdr);
2343 :
2344 : /*
2345 : * We shouldn't have any other pins for this buffer.
2346 : */
2347 3793590 : CheckBufferIsPinnedOnce(buf);
2348 :
2349 : /*
2350 : * If the buffer was dirty, try to write it out. There is a race
2351 : * condition here, in that someone might dirty it after we released the
2352 : * buffer header lock above, or even while we are writing it out (since
2353 : * our share-lock won't prevent hint-bit updates). We will recheck the
2354 : * dirty bit after re-locking the buffer header.
2355 : */
2356 3793590 : if (buf_state & BM_DIRTY)
2357 : {
2358 : LWLock *content_lock;
2359 :
2360 : Assert(buf_state & BM_TAG_VALID);
2361 : Assert(buf_state & BM_VALID);
2362 :
2363 : /*
2364 : * We need a share-lock on the buffer contents to write it out (else
2365 : * we might write invalid data, eg because someone else is compacting
2366 : * the page contents while we write). We must use a conditional lock
2367 : * acquisition here to avoid deadlock. Even though the buffer was not
2368 : * pinned (and therefore surely not locked) when StrategyGetBuffer
2369 : * returned it, someone else could have pinned and exclusive-locked it
2370 : * by the time we get here. If we try to get the lock unconditionally,
2371 : * we'd block waiting for them; if they later block waiting for us,
2372 : * deadlock ensues. (This has been observed to happen when two
2373 : * backends are both trying to split btree index pages, and the second
2374 : * one just happens to be trying to split the page the first one got
2375 : * from StrategyGetBuffer.)
2376 : */
2377 521296 : content_lock = BufferDescriptorGetContentLock(buf_hdr);
2378 521296 : if (!LWLockConditionalAcquire(content_lock, LW_SHARED))
2379 : {
2380 : /*
2381 : * Someone else has locked the buffer, so give it up and loop back
2382 : * to get another one.
2383 : */
2384 0 : UnpinBuffer(buf_hdr);
2385 0 : goto again;
2386 : }
2387 :
2388 : /*
2389 : * If using a nondefault strategy, and writing the buffer would
2390 : * require a WAL flush, let the strategy decide whether to go ahead
2391 : * and write/reuse the buffer or to choose another victim. We need a
2392 : * lock to inspect the page LSN, so this can't be done inside
2393 : * StrategyGetBuffer.
2394 : */
2395 521296 : if (strategy != NULL)
2396 : {
2397 : XLogRecPtr lsn;
2398 :
2399 : /* Read the LSN while holding buffer header lock */
2400 149764 : buf_state = LockBufHdr(buf_hdr);
2401 149764 : lsn = BufferGetLSN(buf_hdr);
2402 149764 : UnlockBufHdr(buf_hdr, buf_state);
2403 :
2404 149764 : if (XLogNeedsFlush(lsn)
2405 17354 : && StrategyRejectBuffer(strategy, buf_hdr, from_ring))
2406 : {
2407 11366 : LWLockRelease(content_lock);
2408 11366 : UnpinBuffer(buf_hdr);
2409 11366 : goto again;
2410 : }
2411 : }
2412 :
2413 : /* OK, do the I/O */
2414 509930 : FlushBuffer(buf_hdr, NULL, IOOBJECT_RELATION, io_context);
2415 509930 : LWLockRelease(content_lock);
2416 :
2417 509930 : ScheduleBufferTagForWriteback(&BackendWritebackContext, io_context,
2418 : &buf_hdr->tag);
2419 : }
2420 :
2421 :
2422 3782224 : if (buf_state & BM_VALID)
2423 : {
2424 : /*
2425 : * When a BufferAccessStrategy is in use, blocks evicted from shared
2426 : * buffers are counted as IOOP_EVICT in the corresponding context
2427 : * (e.g. IOCONTEXT_BULKWRITE). Shared buffers are evicted by a
2428 : * strategy in two cases: 1) while initially claiming buffers for the
2429 : * strategy ring 2) to replace an existing strategy ring buffer
2430 : * because it is pinned or in use and cannot be reused.
2431 : *
2432 : * Blocks evicted from buffers already in the strategy ring are
2433 : * counted as IOOP_REUSE in the corresponding strategy context.
2434 : *
2435 : * At this point, we can accurately count evictions and reuses,
2436 : * because we have successfully claimed the valid buffer. Previously,
2437 : * we may have been forced to release the buffer due to concurrent
2438 : * pinners or erroring out.
2439 : */
2440 2335778 : pgstat_count_io_op(IOOBJECT_RELATION, io_context,
2441 2335778 : from_ring ? IOOP_REUSE : IOOP_EVICT, 1, 0);
2442 : }
2443 :
2444 : /*
2445 : * If the buffer has an entry in the buffer mapping table, delete it. This
2446 : * can fail because another backend could have pinned or dirtied the
2447 : * buffer.
2448 : */
2449 3782224 : if ((buf_state & BM_TAG_VALID) && !InvalidateVictimBuffer(buf_hdr))
2450 : {
2451 850 : UnpinBuffer(buf_hdr);
2452 850 : goto again;
2453 : }
2454 :
2455 : /* a final set of sanity checks */
2456 : #ifdef USE_ASSERT_CHECKING
2457 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
2458 :
2459 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2460 : Assert(!(buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY)));
2461 :
2462 : CheckBufferIsPinnedOnce(buf);
2463 : #endif
2464 :
2465 3781374 : return buf;
2466 : }
2467 :
2468 : /*
2469 : * Return the maximum number of buffers that a backend should try to pin once,
2470 : * to avoid exceeding its fair share. This is the highest value that
2471 : * GetAdditionalPinLimit() could ever return. Note that it may be zero on a
2472 : * system with a very small buffer pool relative to max_connections.
2473 : */
2474 : uint32
2475 1297958 : GetPinLimit(void)
2476 : {
2477 1297958 : return MaxProportionalPins;
2478 : }
2479 :
2480 : /*
2481 : * Return the maximum number of additional buffers that this backend should
2482 : * pin if it wants to stay under the per-backend limit, considering the number
2483 : * of buffers it has already pinned. Unlike LimitAdditionalPins(), the limit
2484 : * return by this function can be zero.
2485 : */
2486 : uint32
2487 7821486 : GetAdditionalPinLimit(void)
2488 : {
2489 : uint32 estimated_pins_held;
2490 :
2491 : /*
2492 : * We get the number of "overflowed" pins for free, but don't know the
2493 : * number of pins in PrivateRefCountArray. The cost of calculating that
2494 : * exactly doesn't seem worth it, so just assume the max.
2495 : */
2496 7821486 : estimated_pins_held = PrivateRefCountOverflowed + REFCOUNT_ARRAY_ENTRIES;
2497 :
2498 : /* Is this backend already holding more than its fair share? */
2499 7821486 : if (estimated_pins_held > MaxProportionalPins)
2500 2372002 : return 0;
2501 :
2502 5449484 : return MaxProportionalPins - estimated_pins_held;
2503 : }
2504 :
2505 : /*
2506 : * Limit the number of pins a batch operation may additionally acquire, to
2507 : * avoid running out of pinnable buffers.
2508 : *
2509 : * One additional pin is always allowed, on the assumption that the operation
2510 : * requires at least one to make progress.
2511 : */
2512 : void
2513 403194 : LimitAdditionalPins(uint32 *additional_pins)
2514 : {
2515 : uint32 limit;
2516 :
2517 403194 : if (*additional_pins <= 1)
2518 383212 : return;
2519 :
2520 19982 : limit = GetAdditionalPinLimit();
2521 19982 : limit = Max(limit, 1);
2522 19982 : if (limit < *additional_pins)
2523 11026 : *additional_pins = limit;
2524 : }
2525 :
2526 : /*
2527 : * Logic shared between ExtendBufferedRelBy(), ExtendBufferedRelTo(). Just to
2528 : * avoid duplicating the tracing and relpersistence related logic.
2529 : */
2530 : static BlockNumber
2531 425990 : ExtendBufferedRelCommon(BufferManagerRelation bmr,
2532 : ForkNumber fork,
2533 : BufferAccessStrategy strategy,
2534 : uint32 flags,
2535 : uint32 extend_by,
2536 : BlockNumber extend_upto,
2537 : Buffer *buffers,
2538 : uint32 *extended_by)
2539 : {
2540 : BlockNumber first_block;
2541 :
2542 : TRACE_POSTGRESQL_BUFFER_EXTEND_START(fork,
2543 : bmr.smgr->smgr_rlocator.locator.spcOid,
2544 : bmr.smgr->smgr_rlocator.locator.dbOid,
2545 : bmr.smgr->smgr_rlocator.locator.relNumber,
2546 : bmr.smgr->smgr_rlocator.backend,
2547 : extend_by);
2548 :
2549 425990 : if (bmr.relpersistence == RELPERSISTENCE_TEMP)
2550 22796 : first_block = ExtendBufferedRelLocal(bmr, fork, flags,
2551 : extend_by, extend_upto,
2552 : buffers, &extend_by);
2553 : else
2554 403194 : first_block = ExtendBufferedRelShared(bmr, fork, strategy, flags,
2555 : extend_by, extend_upto,
2556 : buffers, &extend_by);
2557 425990 : *extended_by = extend_by;
2558 :
2559 : TRACE_POSTGRESQL_BUFFER_EXTEND_DONE(fork,
2560 : bmr.smgr->smgr_rlocator.locator.spcOid,
2561 : bmr.smgr->smgr_rlocator.locator.dbOid,
2562 : bmr.smgr->smgr_rlocator.locator.relNumber,
2563 : bmr.smgr->smgr_rlocator.backend,
2564 : *extended_by,
2565 : first_block);
2566 :
2567 425990 : return first_block;
2568 : }
2569 :
2570 : /*
2571 : * Implementation of ExtendBufferedRelBy() and ExtendBufferedRelTo() for
2572 : * shared buffers.
2573 : */
2574 : static BlockNumber
2575 403194 : ExtendBufferedRelShared(BufferManagerRelation bmr,
2576 : ForkNumber fork,
2577 : BufferAccessStrategy strategy,
2578 : uint32 flags,
2579 : uint32 extend_by,
2580 : BlockNumber extend_upto,
2581 : Buffer *buffers,
2582 : uint32 *extended_by)
2583 : {
2584 : BlockNumber first_block;
2585 403194 : IOContext io_context = IOContextForStrategy(strategy);
2586 : instr_time io_start;
2587 :
2588 403194 : LimitAdditionalPins(&extend_by);
2589 :
2590 : /*
2591 : * Acquire victim buffers for extension without holding extension lock.
2592 : * Writing out victim buffers is the most expensive part of extending the
2593 : * relation, particularly when doing so requires WAL flushes. Zeroing out
2594 : * the buffers is also quite expensive, so do that before holding the
2595 : * extension lock as well.
2596 : *
2597 : * These pages are pinned by us and not valid. While we hold the pin they
2598 : * can't be acquired as victim buffers by another backend.
2599 : */
2600 848818 : for (uint32 i = 0; i < extend_by; i++)
2601 : {
2602 : Block buf_block;
2603 :
2604 445624 : buffers[i] = GetVictimBuffer(strategy, io_context);
2605 445624 : buf_block = BufHdrGetBlock(GetBufferDescriptor(buffers[i] - 1));
2606 :
2607 : /* new buffers are zero-filled */
2608 445624 : MemSet(buf_block, 0, BLCKSZ);
2609 : }
2610 :
2611 : /*
2612 : * Lock relation against concurrent extensions, unless requested not to.
2613 : *
2614 : * We use the same extension lock for all forks. That's unnecessarily
2615 : * restrictive, but currently extensions for forks don't happen often
2616 : * enough to make it worth locking more granularly.
2617 : *
2618 : * Note that another backend might have extended the relation by the time
2619 : * we get the lock.
2620 : */
2621 403194 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2622 300038 : LockRelationForExtension(bmr.rel, ExclusiveLock);
2623 :
2624 : /*
2625 : * If requested, invalidate size cache, so that smgrnblocks asks the
2626 : * kernel.
2627 : */
2628 403194 : if (flags & EB_CLEAR_SIZE_CACHE)
2629 15452 : bmr.smgr->smgr_cached_nblocks[fork] = InvalidBlockNumber;
2630 :
2631 403194 : first_block = smgrnblocks(bmr.smgr, fork);
2632 :
2633 : /*
2634 : * Now that we have the accurate relation size, check if the caller wants
2635 : * us to extend to only up to a specific size. If there were concurrent
2636 : * extensions, we might have acquired too many buffers and need to release
2637 : * them.
2638 : */
2639 403194 : if (extend_upto != InvalidBlockNumber)
2640 : {
2641 106774 : uint32 orig_extend_by = extend_by;
2642 :
2643 106774 : if (first_block > extend_upto)
2644 0 : extend_by = 0;
2645 106774 : else if ((uint64) first_block + extend_by > extend_upto)
2646 26 : extend_by = extend_upto - first_block;
2647 :
2648 106846 : for (uint32 i = extend_by; i < orig_extend_by; i++)
2649 : {
2650 72 : BufferDesc *buf_hdr = GetBufferDescriptor(buffers[i] - 1);
2651 :
2652 72 : UnpinBuffer(buf_hdr);
2653 : }
2654 :
2655 106774 : if (extend_by == 0)
2656 : {
2657 26 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2658 26 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2659 26 : *extended_by = extend_by;
2660 26 : return first_block;
2661 : }
2662 : }
2663 :
2664 : /* Fail if relation is already at maximum possible length */
2665 403168 : if ((uint64) first_block + extend_by >= MaxBlockNumber)
2666 0 : ereport(ERROR,
2667 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
2668 : errmsg("cannot extend relation %s beyond %u blocks",
2669 : relpath(bmr.smgr->smgr_rlocator, fork).str,
2670 : MaxBlockNumber)));
2671 :
2672 : /*
2673 : * Insert buffers into buffer table, mark as IO_IN_PROGRESS.
2674 : *
2675 : * This needs to happen before we extend the relation, because as soon as
2676 : * we do, other backends can start to read in those pages.
2677 : */
2678 848720 : for (uint32 i = 0; i < extend_by; i++)
2679 : {
2680 445552 : Buffer victim_buf = buffers[i];
2681 445552 : BufferDesc *victim_buf_hdr = GetBufferDescriptor(victim_buf - 1);
2682 : BufferTag tag;
2683 : uint32 hash;
2684 : LWLock *partition_lock;
2685 : int existing_id;
2686 :
2687 : /* in case we need to pin an existing buffer below */
2688 445552 : ResourceOwnerEnlarge(CurrentResourceOwner);
2689 445552 : ReservePrivateRefCountEntry();
2690 :
2691 445552 : InitBufferTag(&tag, &bmr.smgr->smgr_rlocator.locator, fork, first_block + i);
2692 445552 : hash = BufTableHashCode(&tag);
2693 445552 : partition_lock = BufMappingPartitionLock(hash);
2694 :
2695 445552 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2696 :
2697 445552 : existing_id = BufTableInsert(&tag, hash, victim_buf_hdr->buf_id);
2698 :
2699 : /*
2700 : * We get here only in the corner case where we are trying to extend
2701 : * the relation but we found a pre-existing buffer. This can happen
2702 : * because a prior attempt at extending the relation failed, and
2703 : * because mdread doesn't complain about reads beyond EOF (when
2704 : * zero_damaged_pages is ON) and so a previous attempt to read a block
2705 : * beyond EOF could have left a "valid" zero-filled buffer.
2706 : *
2707 : * This has also been observed when relation was overwritten by
2708 : * external process. Since the legitimate cases should always have
2709 : * left a zero-filled buffer, complain if not PageIsNew.
2710 : */
2711 445552 : if (existing_id >= 0)
2712 : {
2713 0 : BufferDesc *existing_hdr = GetBufferDescriptor(existing_id);
2714 : Block buf_block;
2715 : bool valid;
2716 :
2717 : /*
2718 : * Pin the existing buffer before releasing the partition lock,
2719 : * preventing it from being evicted.
2720 : */
2721 0 : valid = PinBuffer(existing_hdr, strategy, false);
2722 :
2723 0 : LWLockRelease(partition_lock);
2724 0 : UnpinBuffer(victim_buf_hdr);
2725 :
2726 0 : buffers[i] = BufferDescriptorGetBuffer(existing_hdr);
2727 0 : buf_block = BufHdrGetBlock(existing_hdr);
2728 :
2729 0 : if (valid && !PageIsNew((Page) buf_block))
2730 0 : ereport(ERROR,
2731 : (errmsg("unexpected data beyond EOF in block %u of relation \"%s\"",
2732 : existing_hdr->tag.blockNum,
2733 : relpath(bmr.smgr->smgr_rlocator, fork).str)));
2734 :
2735 : /*
2736 : * We *must* do smgr[zero]extend before succeeding, else the page
2737 : * will not be reserved by the kernel, and the next P_NEW call
2738 : * will decide to return the same page. Clear the BM_VALID bit,
2739 : * do StartBufferIO() and proceed.
2740 : *
2741 : * Loop to handle the very small possibility that someone re-sets
2742 : * BM_VALID between our clearing it and StartBufferIO inspecting
2743 : * it.
2744 : */
2745 : do
2746 : {
2747 0 : uint32 buf_state = LockBufHdr(existing_hdr);
2748 :
2749 0 : buf_state &= ~BM_VALID;
2750 0 : UnlockBufHdr(existing_hdr, buf_state);
2751 0 : } while (!StartBufferIO(existing_hdr, true, false));
2752 : }
2753 : else
2754 : {
2755 : uint32 buf_state;
2756 :
2757 445552 : buf_state = LockBufHdr(victim_buf_hdr);
2758 :
2759 : /* some sanity checks while we hold the buffer header lock */
2760 : Assert(!(buf_state & (BM_VALID | BM_TAG_VALID | BM_DIRTY | BM_JUST_DIRTIED)));
2761 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2762 :
2763 445552 : victim_buf_hdr->tag = tag;
2764 :
2765 445552 : buf_state |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2766 445552 : if (bmr.relpersistence == RELPERSISTENCE_PERMANENT || fork == INIT_FORKNUM)
2767 434998 : buf_state |= BM_PERMANENT;
2768 :
2769 445552 : UnlockBufHdr(victim_buf_hdr, buf_state);
2770 :
2771 445552 : LWLockRelease(partition_lock);
2772 :
2773 : /* XXX: could combine the locked operations in it with the above */
2774 445552 : StartBufferIO(victim_buf_hdr, true, false);
2775 : }
2776 : }
2777 :
2778 403168 : io_start = pgstat_prepare_io_time(track_io_timing);
2779 :
2780 : /*
2781 : * Note: if smgrzeroextend fails, we will end up with buffers that are
2782 : * allocated but not marked BM_VALID. The next relation extension will
2783 : * still select the same block number (because the relation didn't get any
2784 : * longer on disk) and so future attempts to extend the relation will find
2785 : * the same buffers (if they have not been recycled) but come right back
2786 : * here to try smgrzeroextend again.
2787 : *
2788 : * We don't need to set checksum for all-zero pages.
2789 : */
2790 403168 : smgrzeroextend(bmr.smgr, fork, first_block, extend_by, false);
2791 :
2792 : /*
2793 : * Release the file-extension lock; it's now OK for someone else to extend
2794 : * the relation some more.
2795 : *
2796 : * We remove IO_IN_PROGRESS after this, as waking up waiting backends can
2797 : * take noticeable time.
2798 : */
2799 403168 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2800 300012 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2801 :
2802 403168 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context, IOOP_EXTEND,
2803 403168 : io_start, 1, extend_by * BLCKSZ);
2804 :
2805 : /* Set BM_VALID, terminate IO, and wake up any waiters */
2806 848720 : for (uint32 i = 0; i < extend_by; i++)
2807 : {
2808 445552 : Buffer buf = buffers[i];
2809 445552 : BufferDesc *buf_hdr = GetBufferDescriptor(buf - 1);
2810 445552 : bool lock = false;
2811 :
2812 445552 : if (flags & EB_LOCK_FIRST && i == 0)
2813 295888 : lock = true;
2814 149664 : else if (flags & EB_LOCK_TARGET)
2815 : {
2816 : Assert(extend_upto != InvalidBlockNumber);
2817 89194 : if (first_block + i + 1 == extend_upto)
2818 88004 : lock = true;
2819 : }
2820 :
2821 445552 : if (lock)
2822 383892 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
2823 :
2824 445552 : TerminateBufferIO(buf_hdr, false, BM_VALID, true, false);
2825 : }
2826 :
2827 403168 : pgBufferUsage.shared_blks_written += extend_by;
2828 :
2829 403168 : *extended_by = extend_by;
2830 :
2831 403168 : return first_block;
2832 : }
2833 :
2834 : /*
2835 : * BufferIsLockedByMe
2836 : *
2837 : * Checks if this backend has the buffer locked in any mode.
2838 : *
2839 : * Buffer must be pinned.
2840 : */
2841 : bool
2842 0 : BufferIsLockedByMe(Buffer buffer)
2843 : {
2844 : BufferDesc *bufHdr;
2845 :
2846 : Assert(BufferIsPinned(buffer));
2847 :
2848 0 : if (BufferIsLocal(buffer))
2849 : {
2850 : /* Content locks are not maintained for local buffers. */
2851 0 : return true;
2852 : }
2853 : else
2854 : {
2855 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2856 0 : return LWLockHeldByMe(BufferDescriptorGetContentLock(bufHdr));
2857 : }
2858 : }
2859 :
2860 : /*
2861 : * BufferIsLockedByMeInMode
2862 : *
2863 : * Checks if this backend has the buffer locked in the specified mode.
2864 : *
2865 : * Buffer must be pinned.
2866 : */
2867 : bool
2868 0 : BufferIsLockedByMeInMode(Buffer buffer, int mode)
2869 : {
2870 : BufferDesc *bufHdr;
2871 :
2872 : Assert(BufferIsPinned(buffer));
2873 :
2874 0 : if (BufferIsLocal(buffer))
2875 : {
2876 : /* Content locks are not maintained for local buffers. */
2877 0 : return true;
2878 : }
2879 : else
2880 : {
2881 : LWLockMode lw_mode;
2882 :
2883 0 : switch (mode)
2884 : {
2885 0 : case BUFFER_LOCK_EXCLUSIVE:
2886 0 : lw_mode = LW_EXCLUSIVE;
2887 0 : break;
2888 0 : case BUFFER_LOCK_SHARE:
2889 0 : lw_mode = LW_SHARED;
2890 0 : break;
2891 0 : default:
2892 0 : pg_unreachable();
2893 : }
2894 :
2895 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2896 0 : return LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
2897 : lw_mode);
2898 : }
2899 : }
2900 :
2901 : /*
2902 : * BufferIsDirty
2903 : *
2904 : * Checks if buffer is already dirty.
2905 : *
2906 : * Buffer must be pinned and exclusive-locked. (Without an exclusive lock,
2907 : * the result may be stale before it's returned.)
2908 : */
2909 : bool
2910 0 : BufferIsDirty(Buffer buffer)
2911 : {
2912 : BufferDesc *bufHdr;
2913 :
2914 : Assert(BufferIsPinned(buffer));
2915 :
2916 0 : if (BufferIsLocal(buffer))
2917 : {
2918 0 : int bufid = -buffer - 1;
2919 :
2920 0 : bufHdr = GetLocalBufferDescriptor(bufid);
2921 : /* Content locks are not maintained for local buffers. */
2922 : }
2923 : else
2924 : {
2925 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2926 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
2927 : }
2928 :
2929 0 : return pg_atomic_read_u32(&bufHdr->state) & BM_DIRTY;
2930 : }
2931 :
2932 : /*
2933 : * MarkBufferDirty
2934 : *
2935 : * Marks buffer contents as dirty (actual write happens later).
2936 : *
2937 : * Buffer must be pinned and exclusive-locked. (If caller does not hold
2938 : * exclusive lock, then somebody could be in process of writing the buffer,
2939 : * leading to risk of bad data written to disk.)
2940 : */
2941 : void
2942 42960716 : MarkBufferDirty(Buffer buffer)
2943 : {
2944 : BufferDesc *bufHdr;
2945 : uint32 buf_state;
2946 : uint32 old_buf_state;
2947 :
2948 42960716 : if (!BufferIsValid(buffer))
2949 0 : elog(ERROR, "bad buffer ID: %d", buffer);
2950 :
2951 42960716 : if (BufferIsLocal(buffer))
2952 : {
2953 2444470 : MarkLocalBufferDirty(buffer);
2954 2444470 : return;
2955 : }
2956 :
2957 40516246 : bufHdr = GetBufferDescriptor(buffer - 1);
2958 :
2959 : Assert(BufferIsPinned(buffer));
2960 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
2961 :
2962 40516246 : old_buf_state = pg_atomic_read_u32(&bufHdr->state);
2963 : for (;;)
2964 : {
2965 40516436 : if (old_buf_state & BM_LOCKED)
2966 24 : old_buf_state = WaitBufHdrUnlocked(bufHdr);
2967 :
2968 40516436 : buf_state = old_buf_state;
2969 :
2970 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2971 40516436 : buf_state |= BM_DIRTY | BM_JUST_DIRTIED;
2972 :
2973 40516436 : if (pg_atomic_compare_exchange_u32(&bufHdr->state, &old_buf_state,
2974 : buf_state))
2975 40516246 : break;
2976 : }
2977 :
2978 : /*
2979 : * If the buffer was not dirty already, do vacuum accounting.
2980 : */
2981 40516246 : if (!(old_buf_state & BM_DIRTY))
2982 : {
2983 1280916 : pgBufferUsage.shared_blks_dirtied++;
2984 1280916 : if (VacuumCostActive)
2985 10986 : VacuumCostBalance += VacuumCostPageDirty;
2986 : }
2987 : }
2988 :
2989 : /*
2990 : * ReleaseAndReadBuffer -- combine ReleaseBuffer() and ReadBuffer()
2991 : *
2992 : * Formerly, this saved one cycle of acquiring/releasing the BufMgrLock
2993 : * compared to calling the two routines separately. Now it's mainly just
2994 : * a convenience function. However, if the passed buffer is valid and
2995 : * already contains the desired block, we just return it as-is; and that
2996 : * does save considerable work compared to a full release and reacquire.
2997 : *
2998 : * Note: it is OK to pass buffer == InvalidBuffer, indicating that no old
2999 : * buffer actually needs to be released. This case is the same as ReadBuffer,
3000 : * but can save some tests in the caller.
3001 : */
3002 : Buffer
3003 57748838 : ReleaseAndReadBuffer(Buffer buffer,
3004 : Relation relation,
3005 : BlockNumber blockNum)
3006 : {
3007 57748838 : ForkNumber forkNum = MAIN_FORKNUM;
3008 : BufferDesc *bufHdr;
3009 :
3010 57748838 : if (BufferIsValid(buffer))
3011 : {
3012 : Assert(BufferIsPinned(buffer));
3013 34538752 : if (BufferIsLocal(buffer))
3014 : {
3015 73728 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
3016 80772 : if (bufHdr->tag.blockNum == blockNum &&
3017 14088 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3018 7044 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3019 7044 : return buffer;
3020 66684 : UnpinLocalBuffer(buffer);
3021 : }
3022 : else
3023 : {
3024 34465024 : bufHdr = GetBufferDescriptor(buffer - 1);
3025 : /* we have pin, so it's ok to examine tag without spinlock */
3026 46336618 : if (bufHdr->tag.blockNum == blockNum &&
3027 23743188 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3028 11871594 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3029 11871594 : return buffer;
3030 22593430 : UnpinBuffer(bufHdr);
3031 : }
3032 : }
3033 :
3034 45870200 : return ReadBuffer(relation, blockNum);
3035 : }
3036 :
3037 : /*
3038 : * PinBuffer -- make buffer unavailable for replacement.
3039 : *
3040 : * For the default access strategy, the buffer's usage_count is incremented
3041 : * when we first pin it; for other strategies we just make sure the usage_count
3042 : * isn't zero. (The idea of the latter is that we don't want synchronized
3043 : * heap scans to inflate the count, but we need it to not be zero to discourage
3044 : * other backends from stealing buffers from our ring. As long as we cycle
3045 : * through the ring faster than the global clock-sweep cycles, buffers in
3046 : * our ring won't be chosen as victims for replacement by other backends.)
3047 : *
3048 : * This should be applied only to shared buffers, never local ones.
3049 : *
3050 : * Since buffers are pinned/unpinned very frequently, pin buffers without
3051 : * taking the buffer header lock; instead update the state variable in loop of
3052 : * CAS operations. Hopefully it's just a single CAS.
3053 : *
3054 : * Note that ResourceOwnerEnlarge() and ReservePrivateRefCountEntry()
3055 : * must have been done already.
3056 : *
3057 : * Returns true if buffer is BM_VALID, else false. This provision allows
3058 : * some callers to avoid an extra spinlock cycle. If skip_if_not_valid is
3059 : * true, then a false return value also indicates that the buffer was
3060 : * (recently) invalid and has not been pinned.
3061 : */
3062 : static bool
3063 118036288 : PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
3064 : bool skip_if_not_valid)
3065 : {
3066 118036288 : Buffer b = BufferDescriptorGetBuffer(buf);
3067 : bool result;
3068 : PrivateRefCountEntry *ref;
3069 :
3070 : Assert(!BufferIsLocal(b));
3071 : Assert(ReservedRefCountEntry != NULL);
3072 :
3073 118036288 : ref = GetPrivateRefCountEntry(b, true);
3074 :
3075 118036288 : if (ref == NULL)
3076 : {
3077 : uint32 buf_state;
3078 : uint32 old_buf_state;
3079 :
3080 113713324 : old_buf_state = pg_atomic_read_u32(&buf->state);
3081 : for (;;)
3082 : {
3083 113748426 : if (unlikely(skip_if_not_valid && !(old_buf_state & BM_VALID)))
3084 12 : return false;
3085 :
3086 113748414 : if (old_buf_state & BM_LOCKED)
3087 1584 : old_buf_state = WaitBufHdrUnlocked(buf);
3088 :
3089 113748414 : buf_state = old_buf_state;
3090 :
3091 : /* increase refcount */
3092 113748414 : buf_state += BUF_REFCOUNT_ONE;
3093 :
3094 113748414 : if (strategy == NULL)
3095 : {
3096 : /* Default case: increase usagecount unless already max. */
3097 112193772 : if (BUF_STATE_GET_USAGECOUNT(buf_state) < BM_MAX_USAGE_COUNT)
3098 6519816 : buf_state += BUF_USAGECOUNT_ONE;
3099 : }
3100 : else
3101 : {
3102 : /*
3103 : * Ring buffers shouldn't evict others from pool. Thus we
3104 : * don't make usagecount more than 1.
3105 : */
3106 1554642 : if (BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3107 75354 : buf_state += BUF_USAGECOUNT_ONE;
3108 : }
3109 :
3110 113748414 : if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
3111 : buf_state))
3112 : {
3113 113713312 : result = (buf_state & BM_VALID) != 0;
3114 :
3115 113713312 : TrackNewBufferPin(b);
3116 113713312 : break;
3117 : }
3118 : }
3119 : }
3120 : else
3121 : {
3122 : /*
3123 : * If we previously pinned the buffer, it is likely to be valid, but
3124 : * it may not be if StartReadBuffers() was called and
3125 : * WaitReadBuffers() hasn't been called yet. We'll check by loading
3126 : * the flags without locking. This is racy, but it's OK to return
3127 : * false spuriously: when WaitReadBuffers() calls StartBufferIO(),
3128 : * it'll see that it's now valid.
3129 : *
3130 : * Note: We deliberately avoid a Valgrind client request here.
3131 : * Individual access methods can optionally superimpose buffer page
3132 : * client requests on top of our client requests to enforce that
3133 : * buffers are only accessed while locked (and pinned). It's possible
3134 : * that the buffer page is legitimately non-accessible here. We
3135 : * cannot meddle with that.
3136 : */
3137 4322964 : result = (pg_atomic_read_u32(&buf->state) & BM_VALID) != 0;
3138 :
3139 : Assert(ref->refcount > 0);
3140 4322964 : ref->refcount++;
3141 4322964 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
3142 : }
3143 :
3144 118036276 : return result;
3145 : }
3146 :
3147 : /*
3148 : * PinBuffer_Locked -- as above, but caller already locked the buffer header.
3149 : * The spinlock is released before return.
3150 : *
3151 : * As this function is called with the spinlock held, the caller has to
3152 : * previously call ReservePrivateRefCountEntry() and
3153 : * ResourceOwnerEnlarge(CurrentResourceOwner);
3154 : *
3155 : * Currently, no callers of this function want to modify the buffer's
3156 : * usage_count at all, so there's no need for a strategy parameter.
3157 : * Also we don't bother with a BM_VALID test (the caller could check that for
3158 : * itself).
3159 : *
3160 : * Also all callers only ever use this function when it's known that the
3161 : * buffer can't have a preexisting pin by this backend. That allows us to skip
3162 : * searching the private refcount array & hash, which is a boon, because the
3163 : * spinlock is still held.
3164 : *
3165 : * Note: use of this routine is frequently mandatory, not just an optimization
3166 : * to save a spin lock/unlock cycle, because we need to pin a buffer before
3167 : * its state can change under us.
3168 : */
3169 : static void
3170 594946 : PinBuffer_Locked(BufferDesc *buf)
3171 : {
3172 : uint32 buf_state;
3173 :
3174 : /*
3175 : * As explained, We don't expect any preexisting pins. That allows us to
3176 : * manipulate the PrivateRefCount after releasing the spinlock
3177 : */
3178 : Assert(GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf), false) == NULL);
3179 :
3180 : /*
3181 : * Since we hold the buffer spinlock, we can update the buffer state and
3182 : * release the lock in one operation.
3183 : */
3184 594946 : buf_state = pg_atomic_read_u32(&buf->state);
3185 : Assert(buf_state & BM_LOCKED);
3186 594946 : buf_state += BUF_REFCOUNT_ONE;
3187 594946 : UnlockBufHdr(buf, buf_state);
3188 :
3189 594946 : TrackNewBufferPin(BufferDescriptorGetBuffer(buf));
3190 594946 : }
3191 :
3192 : /*
3193 : * Support for waking up another backend that is waiting for the cleanup lock
3194 : * to be released using BM_PIN_COUNT_WAITER.
3195 : *
3196 : * See LockBufferForCleanup().
3197 : *
3198 : * Expected to be called just after releasing a buffer pin (in a BufferDesc,
3199 : * not just reducing the backend-local pincount for the buffer).
3200 : */
3201 : static void
3202 156 : WakePinCountWaiter(BufferDesc *buf)
3203 : {
3204 : /*
3205 : * Acquire the buffer header lock, re-check that there's a waiter. Another
3206 : * backend could have unpinned this buffer, and already woken up the
3207 : * waiter.
3208 : *
3209 : * There's no danger of the buffer being replaced after we unpinned it
3210 : * above, as it's pinned by the waiter. The waiter removes
3211 : * BM_PIN_COUNT_WAITER if it stops waiting for a reason other than this
3212 : * backend waking it up.
3213 : */
3214 156 : uint32 buf_state = LockBufHdr(buf);
3215 :
3216 156 : if ((buf_state & BM_PIN_COUNT_WAITER) &&
3217 156 : BUF_STATE_GET_REFCOUNT(buf_state) == 1)
3218 156 : {
3219 : /* we just released the last pin other than the waiter's */
3220 156 : int wait_backend_pgprocno = buf->wait_backend_pgprocno;
3221 :
3222 156 : buf_state &= ~BM_PIN_COUNT_WAITER;
3223 156 : UnlockBufHdr(buf, buf_state);
3224 156 : ProcSendSignal(wait_backend_pgprocno);
3225 : }
3226 : else
3227 0 : UnlockBufHdr(buf, buf_state);
3228 156 : }
3229 :
3230 : /*
3231 : * UnpinBuffer -- make buffer available for replacement.
3232 : *
3233 : * This should be applied only to shared buffers, never local ones. This
3234 : * always adjusts CurrentResourceOwner.
3235 : */
3236 : static void
3237 145013400 : UnpinBuffer(BufferDesc *buf)
3238 : {
3239 145013400 : Buffer b = BufferDescriptorGetBuffer(buf);
3240 :
3241 145013400 : ResourceOwnerForgetBuffer(CurrentResourceOwner, b);
3242 145013400 : UnpinBufferNoOwner(buf);
3243 145013400 : }
3244 :
3245 : static void
3246 145022400 : UnpinBufferNoOwner(BufferDesc *buf)
3247 : {
3248 : PrivateRefCountEntry *ref;
3249 145022400 : Buffer b = BufferDescriptorGetBuffer(buf);
3250 :
3251 : Assert(!BufferIsLocal(b));
3252 :
3253 : /* not moving as we're likely deleting it soon anyway */
3254 145022400 : ref = GetPrivateRefCountEntry(b, false);
3255 : Assert(ref != NULL);
3256 : Assert(ref->refcount > 0);
3257 145022400 : ref->refcount--;
3258 145022400 : if (ref->refcount == 0)
3259 : {
3260 : uint32 buf_state;
3261 : uint32 old_buf_state;
3262 :
3263 : /*
3264 : * Mark buffer non-accessible to Valgrind.
3265 : *
3266 : * Note that the buffer may have already been marked non-accessible
3267 : * within access method code that enforces that buffers are only
3268 : * accessed while a buffer lock is held.
3269 : */
3270 : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
3271 :
3272 : /*
3273 : * I'd better not still hold the buffer content lock. Can't use
3274 : * BufferIsLockedByMe(), as that asserts the buffer is pinned.
3275 : */
3276 : Assert(!LWLockHeldByMe(BufferDescriptorGetContentLock(buf)));
3277 :
3278 : /*
3279 : * Decrement the shared reference count.
3280 : *
3281 : * Since buffer spinlock holder can update status using just write,
3282 : * it's not safe to use atomic decrement here; thus use a CAS loop.
3283 : */
3284 118101848 : old_buf_state = pg_atomic_read_u32(&buf->state);
3285 : for (;;)
3286 : {
3287 118132284 : if (old_buf_state & BM_LOCKED)
3288 1826 : old_buf_state = WaitBufHdrUnlocked(buf);
3289 :
3290 118132284 : buf_state = old_buf_state;
3291 :
3292 118132284 : buf_state -= BUF_REFCOUNT_ONE;
3293 :
3294 118132284 : if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
3295 : buf_state))
3296 118101848 : break;
3297 : }
3298 :
3299 : /* Support LockBufferForCleanup() */
3300 118101848 : if (buf_state & BM_PIN_COUNT_WAITER)
3301 156 : WakePinCountWaiter(buf);
3302 :
3303 118101848 : ForgetPrivateRefCountEntry(ref);
3304 : }
3305 145022400 : }
3306 :
3307 : /*
3308 : * Set up backend-local tracking of a buffer pinned the first time by this
3309 : * backend.
3310 : */
3311 : inline void
3312 118101848 : TrackNewBufferPin(Buffer buf)
3313 : {
3314 : PrivateRefCountEntry *ref;
3315 :
3316 118101848 : ref = NewPrivateRefCountEntry(buf);
3317 118101848 : ref->refcount++;
3318 :
3319 118101848 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buf);
3320 :
3321 : /*
3322 : * This is the first pin for this page by this backend, mark its page as
3323 : * defined to valgrind. While the page contents might not actually be
3324 : * valid yet, we don't currently guarantee that such pages are marked
3325 : * undefined or non-accessible.
3326 : *
3327 : * It's not necessarily the prettiest to do this here, but otherwise we'd
3328 : * need this block of code in multiple places.
3329 : */
3330 : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(GetBufferDescriptor(buf - 1)),
3331 : BLCKSZ);
3332 118101848 : }
3333 :
3334 : #define ST_SORT sort_checkpoint_bufferids
3335 : #define ST_ELEMENT_TYPE CkptSortItem
3336 : #define ST_COMPARE(a, b) ckpt_buforder_comparator(a, b)
3337 : #define ST_SCOPE static
3338 : #define ST_DEFINE
3339 : #include "lib/sort_template.h"
3340 :
3341 : /*
3342 : * BufferSync -- Write out all dirty buffers in the pool.
3343 : *
3344 : * This is called at checkpoint time to write out all dirty shared buffers.
3345 : * The checkpoint request flags should be passed in. If CHECKPOINT_FAST is
3346 : * set, we disable delays between writes; if CHECKPOINT_IS_SHUTDOWN,
3347 : * CHECKPOINT_END_OF_RECOVERY or CHECKPOINT_FLUSH_UNLOGGED is set, we write
3348 : * even unlogged buffers, which are otherwise skipped. The remaining flags
3349 : * currently have no effect here.
3350 : */
3351 : static void
3352 3430 : BufferSync(int flags)
3353 : {
3354 : uint32 buf_state;
3355 : int buf_id;
3356 : int num_to_scan;
3357 : int num_spaces;
3358 : int num_processed;
3359 : int num_written;
3360 3430 : CkptTsStatus *per_ts_stat = NULL;
3361 : Oid last_tsid;
3362 : binaryheap *ts_heap;
3363 : int i;
3364 3430 : uint32 mask = BM_DIRTY;
3365 : WritebackContext wb_context;
3366 :
3367 : /*
3368 : * Unless this is a shutdown checkpoint or we have been explicitly told,
3369 : * we write only permanent, dirty buffers. But at shutdown or end of
3370 : * recovery, we write all dirty buffers.
3371 : */
3372 3430 : if (!((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
3373 : CHECKPOINT_FLUSH_UNLOGGED))))
3374 1926 : mask |= BM_PERMANENT;
3375 :
3376 : /*
3377 : * Loop over all buffers, and mark the ones that need to be written with
3378 : * BM_CHECKPOINT_NEEDED. Count them as we go (num_to_scan), so that we
3379 : * can estimate how much work needs to be done.
3380 : *
3381 : * This allows us to write only those pages that were dirty when the
3382 : * checkpoint began, and not those that get dirtied while it proceeds.
3383 : * Whenever a page with BM_CHECKPOINT_NEEDED is written out, either by us
3384 : * later in this function, or by normal backends or the bgwriter cleaning
3385 : * scan, the flag is cleared. Any buffer dirtied after this point won't
3386 : * have the flag set.
3387 : *
3388 : * Note that if we fail to write some buffer, we may leave buffers with
3389 : * BM_CHECKPOINT_NEEDED still set. This is OK since any such buffer would
3390 : * certainly need to be written for the next checkpoint attempt, too.
3391 : */
3392 3430 : num_to_scan = 0;
3393 23948038 : for (buf_id = 0; buf_id < NBuffers; buf_id++)
3394 : {
3395 23944608 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3396 :
3397 : /*
3398 : * Header spinlock is enough to examine BM_DIRTY, see comment in
3399 : * SyncOneBuffer.
3400 : */
3401 23944608 : buf_state = LockBufHdr(bufHdr);
3402 :
3403 23944608 : if ((buf_state & mask) == mask)
3404 : {
3405 : CkptSortItem *item;
3406 :
3407 584496 : buf_state |= BM_CHECKPOINT_NEEDED;
3408 :
3409 584496 : item = &CkptBufferIds[num_to_scan++];
3410 584496 : item->buf_id = buf_id;
3411 584496 : item->tsId = bufHdr->tag.spcOid;
3412 584496 : item->relNumber = BufTagGetRelNumber(&bufHdr->tag);
3413 584496 : item->forkNum = BufTagGetForkNum(&bufHdr->tag);
3414 584496 : item->blockNum = bufHdr->tag.blockNum;
3415 : }
3416 :
3417 23944608 : UnlockBufHdr(bufHdr, buf_state);
3418 :
3419 : /* Check for barrier events in case NBuffers is large. */
3420 23944608 : if (ProcSignalBarrierPending)
3421 0 : ProcessProcSignalBarrier();
3422 : }
3423 :
3424 3430 : if (num_to_scan == 0)
3425 1322 : return; /* nothing to do */
3426 :
3427 2108 : WritebackContextInit(&wb_context, &checkpoint_flush_after);
3428 :
3429 : TRACE_POSTGRESQL_BUFFER_SYNC_START(NBuffers, num_to_scan);
3430 :
3431 : /*
3432 : * Sort buffers that need to be written to reduce the likelihood of random
3433 : * IO. The sorting is also important for the implementation of balancing
3434 : * writes between tablespaces. Without balancing writes we'd potentially
3435 : * end up writing to the tablespaces one-by-one; possibly overloading the
3436 : * underlying system.
3437 : */
3438 2108 : sort_checkpoint_bufferids(CkptBufferIds, num_to_scan);
3439 :
3440 2108 : num_spaces = 0;
3441 :
3442 : /*
3443 : * Allocate progress status for each tablespace with buffers that need to
3444 : * be flushed. This requires the to-be-flushed array to be sorted.
3445 : */
3446 2108 : last_tsid = InvalidOid;
3447 586604 : for (i = 0; i < num_to_scan; i++)
3448 : {
3449 : CkptTsStatus *s;
3450 : Oid cur_tsid;
3451 :
3452 584496 : cur_tsid = CkptBufferIds[i].tsId;
3453 :
3454 : /*
3455 : * Grow array of per-tablespace status structs, every time a new
3456 : * tablespace is found.
3457 : */
3458 584496 : if (last_tsid == InvalidOid || last_tsid != cur_tsid)
3459 3164 : {
3460 : Size sz;
3461 :
3462 3164 : num_spaces++;
3463 :
3464 : /*
3465 : * Not worth adding grow-by-power-of-2 logic here - even with a
3466 : * few hundred tablespaces this should be fine.
3467 : */
3468 3164 : sz = sizeof(CkptTsStatus) * num_spaces;
3469 :
3470 3164 : if (per_ts_stat == NULL)
3471 2108 : per_ts_stat = (CkptTsStatus *) palloc(sz);
3472 : else
3473 1056 : per_ts_stat = (CkptTsStatus *) repalloc(per_ts_stat, sz);
3474 :
3475 3164 : s = &per_ts_stat[num_spaces - 1];
3476 3164 : memset(s, 0, sizeof(*s));
3477 3164 : s->tsId = cur_tsid;
3478 :
3479 : /*
3480 : * The first buffer in this tablespace. As CkptBufferIds is sorted
3481 : * by tablespace all (s->num_to_scan) buffers in this tablespace
3482 : * will follow afterwards.
3483 : */
3484 3164 : s->index = i;
3485 :
3486 : /*
3487 : * progress_slice will be determined once we know how many buffers
3488 : * are in each tablespace, i.e. after this loop.
3489 : */
3490 :
3491 3164 : last_tsid = cur_tsid;
3492 : }
3493 : else
3494 : {
3495 581332 : s = &per_ts_stat[num_spaces - 1];
3496 : }
3497 :
3498 584496 : s->num_to_scan++;
3499 :
3500 : /* Check for barrier events. */
3501 584496 : if (ProcSignalBarrierPending)
3502 0 : ProcessProcSignalBarrier();
3503 : }
3504 :
3505 : Assert(num_spaces > 0);
3506 :
3507 : /*
3508 : * Build a min-heap over the write-progress in the individual tablespaces,
3509 : * and compute how large a portion of the total progress a single
3510 : * processed buffer is.
3511 : */
3512 2108 : ts_heap = binaryheap_allocate(num_spaces,
3513 : ts_ckpt_progress_comparator,
3514 : NULL);
3515 :
3516 5272 : for (i = 0; i < num_spaces; i++)
3517 : {
3518 3164 : CkptTsStatus *ts_stat = &per_ts_stat[i];
3519 :
3520 3164 : ts_stat->progress_slice = (float8) num_to_scan / ts_stat->num_to_scan;
3521 :
3522 3164 : binaryheap_add_unordered(ts_heap, PointerGetDatum(ts_stat));
3523 : }
3524 :
3525 2108 : binaryheap_build(ts_heap);
3526 :
3527 : /*
3528 : * Iterate through to-be-checkpointed buffers and write the ones (still)
3529 : * marked with BM_CHECKPOINT_NEEDED. The writes are balanced between
3530 : * tablespaces; otherwise the sorting would lead to only one tablespace
3531 : * receiving writes at a time, making inefficient use of the hardware.
3532 : */
3533 2108 : num_processed = 0;
3534 2108 : num_written = 0;
3535 586604 : while (!binaryheap_empty(ts_heap))
3536 : {
3537 584496 : BufferDesc *bufHdr = NULL;
3538 : CkptTsStatus *ts_stat = (CkptTsStatus *)
3539 584496 : DatumGetPointer(binaryheap_first(ts_heap));
3540 :
3541 584496 : buf_id = CkptBufferIds[ts_stat->index].buf_id;
3542 : Assert(buf_id != -1);
3543 :
3544 584496 : bufHdr = GetBufferDescriptor(buf_id);
3545 :
3546 584496 : num_processed++;
3547 :
3548 : /*
3549 : * We don't need to acquire the lock here, because we're only looking
3550 : * at a single bit. It's possible that someone else writes the buffer
3551 : * and clears the flag right after we check, but that doesn't matter
3552 : * since SyncOneBuffer will then do nothing. However, there is a
3553 : * further race condition: it's conceivable that between the time we
3554 : * examine the bit here and the time SyncOneBuffer acquires the lock,
3555 : * someone else not only wrote the buffer but replaced it with another
3556 : * page and dirtied it. In that improbable case, SyncOneBuffer will
3557 : * write the buffer though we didn't need to. It doesn't seem worth
3558 : * guarding against this, though.
3559 : */
3560 584496 : if (pg_atomic_read_u32(&bufHdr->state) & BM_CHECKPOINT_NEEDED)
3561 : {
3562 547172 : if (SyncOneBuffer(buf_id, false, &wb_context) & BUF_WRITTEN)
3563 : {
3564 : TRACE_POSTGRESQL_BUFFER_SYNC_WRITTEN(buf_id);
3565 547172 : PendingCheckpointerStats.buffers_written++;
3566 547172 : num_written++;
3567 : }
3568 : }
3569 :
3570 : /*
3571 : * Measure progress independent of actually having to flush the buffer
3572 : * - otherwise writing become unbalanced.
3573 : */
3574 584496 : ts_stat->progress += ts_stat->progress_slice;
3575 584496 : ts_stat->num_scanned++;
3576 584496 : ts_stat->index++;
3577 :
3578 : /* Have all the buffers from the tablespace been processed? */
3579 584496 : if (ts_stat->num_scanned == ts_stat->num_to_scan)
3580 : {
3581 3164 : binaryheap_remove_first(ts_heap);
3582 : }
3583 : else
3584 : {
3585 : /* update heap with the new progress */
3586 581332 : binaryheap_replace_first(ts_heap, PointerGetDatum(ts_stat));
3587 : }
3588 :
3589 : /*
3590 : * Sleep to throttle our I/O rate.
3591 : *
3592 : * (This will check for barrier events even if it doesn't sleep.)
3593 : */
3594 584496 : CheckpointWriteDelay(flags, (double) num_processed / num_to_scan);
3595 : }
3596 :
3597 : /*
3598 : * Issue all pending flushes. Only checkpointer calls BufferSync(), so
3599 : * IOContext will always be IOCONTEXT_NORMAL.
3600 : */
3601 2108 : IssuePendingWritebacks(&wb_context, IOCONTEXT_NORMAL);
3602 :
3603 2108 : pfree(per_ts_stat);
3604 2108 : per_ts_stat = NULL;
3605 2108 : binaryheap_free(ts_heap);
3606 :
3607 : /*
3608 : * Update checkpoint statistics. As noted above, this doesn't include
3609 : * buffers written by other backends or bgwriter scan.
3610 : */
3611 2108 : CheckpointStats.ckpt_bufs_written += num_written;
3612 :
3613 : TRACE_POSTGRESQL_BUFFER_SYNC_DONE(NBuffers, num_written, num_to_scan);
3614 : }
3615 :
3616 : /*
3617 : * BgBufferSync -- Write out some dirty buffers in the pool.
3618 : *
3619 : * This is called periodically by the background writer process.
3620 : *
3621 : * Returns true if it's appropriate for the bgwriter process to go into
3622 : * low-power hibernation mode. (This happens if the strategy clock-sweep
3623 : * has been "lapped" and no buffer allocations have occurred recently,
3624 : * or if the bgwriter has been effectively disabled by setting
3625 : * bgwriter_lru_maxpages to 0.)
3626 : */
3627 : bool
3628 21972 : BgBufferSync(WritebackContext *wb_context)
3629 : {
3630 : /* info obtained from freelist.c */
3631 : int strategy_buf_id;
3632 : uint32 strategy_passes;
3633 : uint32 recent_alloc;
3634 :
3635 : /*
3636 : * Information saved between calls so we can determine the strategy
3637 : * point's advance rate and avoid scanning already-cleaned buffers.
3638 : */
3639 : static bool saved_info_valid = false;
3640 : static int prev_strategy_buf_id;
3641 : static uint32 prev_strategy_passes;
3642 : static int next_to_clean;
3643 : static uint32 next_passes;
3644 :
3645 : /* Moving averages of allocation rate and clean-buffer density */
3646 : static float smoothed_alloc = 0;
3647 : static float smoothed_density = 10.0;
3648 :
3649 : /* Potentially these could be tunables, but for now, not */
3650 21972 : float smoothing_samples = 16;
3651 21972 : float scan_whole_pool_milliseconds = 120000.0;
3652 :
3653 : /* Used to compute how far we scan ahead */
3654 : long strategy_delta;
3655 : int bufs_to_lap;
3656 : int bufs_ahead;
3657 : float scans_per_alloc;
3658 : int reusable_buffers_est;
3659 : int upcoming_alloc_est;
3660 : int min_scan_buffers;
3661 :
3662 : /* Variables for the scanning loop proper */
3663 : int num_to_scan;
3664 : int num_written;
3665 : int reusable_buffers;
3666 :
3667 : /* Variables for final smoothed_density update */
3668 : long new_strategy_delta;
3669 : uint32 new_recent_alloc;
3670 :
3671 : /*
3672 : * Find out where the clock-sweep currently is, and how many buffer
3673 : * allocations have happened since our last call.
3674 : */
3675 21972 : strategy_buf_id = StrategySyncStart(&strategy_passes, &recent_alloc);
3676 :
3677 : /* Report buffer alloc counts to pgstat */
3678 21972 : PendingBgWriterStats.buf_alloc += recent_alloc;
3679 :
3680 : /*
3681 : * If we're not running the LRU scan, just stop after doing the stats
3682 : * stuff. We mark the saved state invalid so that we can recover sanely
3683 : * if LRU scan is turned back on later.
3684 : */
3685 21972 : if (bgwriter_lru_maxpages <= 0)
3686 : {
3687 78 : saved_info_valid = false;
3688 78 : return true;
3689 : }
3690 :
3691 : /*
3692 : * Compute strategy_delta = how many buffers have been scanned by the
3693 : * clock-sweep since last time. If first time through, assume none. Then
3694 : * see if we are still ahead of the clock-sweep, and if so, how many
3695 : * buffers we could scan before we'd catch up with it and "lap" it. Note:
3696 : * weird-looking coding of xxx_passes comparisons are to avoid bogus
3697 : * behavior when the passes counts wrap around.
3698 : */
3699 21894 : if (saved_info_valid)
3700 : {
3701 20836 : int32 passes_delta = strategy_passes - prev_strategy_passes;
3702 :
3703 20836 : strategy_delta = strategy_buf_id - prev_strategy_buf_id;
3704 20836 : strategy_delta += (long) passes_delta * NBuffers;
3705 :
3706 : Assert(strategy_delta >= 0);
3707 :
3708 20836 : if ((int32) (next_passes - strategy_passes) > 0)
3709 : {
3710 : /* we're one pass ahead of the strategy point */
3711 3588 : bufs_to_lap = strategy_buf_id - next_to_clean;
3712 : #ifdef BGW_DEBUG
3713 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3714 : next_passes, next_to_clean,
3715 : strategy_passes, strategy_buf_id,
3716 : strategy_delta, bufs_to_lap);
3717 : #endif
3718 : }
3719 17248 : else if (next_passes == strategy_passes &&
3720 13552 : next_to_clean >= strategy_buf_id)
3721 : {
3722 : /* on same pass, but ahead or at least not behind */
3723 12264 : bufs_to_lap = NBuffers - (next_to_clean - strategy_buf_id);
3724 : #ifdef BGW_DEBUG
3725 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3726 : next_passes, next_to_clean,
3727 : strategy_passes, strategy_buf_id,
3728 : strategy_delta, bufs_to_lap);
3729 : #endif
3730 : }
3731 : else
3732 : {
3733 : /*
3734 : * We're behind, so skip forward to the strategy point and start
3735 : * cleaning from there.
3736 : */
3737 : #ifdef BGW_DEBUG
3738 : elog(DEBUG2, "bgwriter behind: bgw %u-%u strategy %u-%u delta=%ld",
3739 : next_passes, next_to_clean,
3740 : strategy_passes, strategy_buf_id,
3741 : strategy_delta);
3742 : #endif
3743 4984 : next_to_clean = strategy_buf_id;
3744 4984 : next_passes = strategy_passes;
3745 4984 : bufs_to_lap = NBuffers;
3746 : }
3747 : }
3748 : else
3749 : {
3750 : /*
3751 : * Initializing at startup or after LRU scanning had been off. Always
3752 : * start at the strategy point.
3753 : */
3754 : #ifdef BGW_DEBUG
3755 : elog(DEBUG2, "bgwriter initializing: strategy %u-%u",
3756 : strategy_passes, strategy_buf_id);
3757 : #endif
3758 1058 : strategy_delta = 0;
3759 1058 : next_to_clean = strategy_buf_id;
3760 1058 : next_passes = strategy_passes;
3761 1058 : bufs_to_lap = NBuffers;
3762 : }
3763 :
3764 : /* Update saved info for next time */
3765 21894 : prev_strategy_buf_id = strategy_buf_id;
3766 21894 : prev_strategy_passes = strategy_passes;
3767 21894 : saved_info_valid = true;
3768 :
3769 : /*
3770 : * Compute how many buffers had to be scanned for each new allocation, ie,
3771 : * 1/density of reusable buffers, and track a moving average of that.
3772 : *
3773 : * If the strategy point didn't move, we don't update the density estimate
3774 : */
3775 21894 : if (strategy_delta > 0 && recent_alloc > 0)
3776 : {
3777 11296 : scans_per_alloc = (float) strategy_delta / (float) recent_alloc;
3778 11296 : smoothed_density += (scans_per_alloc - smoothed_density) /
3779 : smoothing_samples;
3780 : }
3781 :
3782 : /*
3783 : * Estimate how many reusable buffers there are between the current
3784 : * strategy point and where we've scanned ahead to, based on the smoothed
3785 : * density estimate.
3786 : */
3787 21894 : bufs_ahead = NBuffers - bufs_to_lap;
3788 21894 : reusable_buffers_est = (float) bufs_ahead / smoothed_density;
3789 :
3790 : /*
3791 : * Track a moving average of recent buffer allocations. Here, rather than
3792 : * a true average we want a fast-attack, slow-decline behavior: we
3793 : * immediately follow any increase.
3794 : */
3795 21894 : if (smoothed_alloc <= (float) recent_alloc)
3796 6538 : smoothed_alloc = recent_alloc;
3797 : else
3798 15356 : smoothed_alloc += ((float) recent_alloc - smoothed_alloc) /
3799 : smoothing_samples;
3800 :
3801 : /* Scale the estimate by a GUC to allow more aggressive tuning. */
3802 21894 : upcoming_alloc_est = (int) (smoothed_alloc * bgwriter_lru_multiplier);
3803 :
3804 : /*
3805 : * If recent_alloc remains at zero for many cycles, smoothed_alloc will
3806 : * eventually underflow to zero, and the underflows produce annoying
3807 : * kernel warnings on some platforms. Once upcoming_alloc_est has gone to
3808 : * zero, there's no point in tracking smaller and smaller values of
3809 : * smoothed_alloc, so just reset it to exactly zero to avoid this
3810 : * syndrome. It will pop back up as soon as recent_alloc increases.
3811 : */
3812 21894 : if (upcoming_alloc_est == 0)
3813 4080 : smoothed_alloc = 0;
3814 :
3815 : /*
3816 : * Even in cases where there's been little or no buffer allocation
3817 : * activity, we want to make a small amount of progress through the buffer
3818 : * cache so that as many reusable buffers as possible are clean after an
3819 : * idle period.
3820 : *
3821 : * (scan_whole_pool_milliseconds / BgWriterDelay) computes how many times
3822 : * the BGW will be called during the scan_whole_pool time; slice the
3823 : * buffer pool into that many sections.
3824 : */
3825 21894 : min_scan_buffers = (int) (NBuffers / (scan_whole_pool_milliseconds / BgWriterDelay));
3826 :
3827 21894 : if (upcoming_alloc_est < (min_scan_buffers + reusable_buffers_est))
3828 : {
3829 : #ifdef BGW_DEBUG
3830 : elog(DEBUG2, "bgwriter: alloc_est=%d too small, using min=%d + reusable_est=%d",
3831 : upcoming_alloc_est, min_scan_buffers, reusable_buffers_est);
3832 : #endif
3833 10572 : upcoming_alloc_est = min_scan_buffers + reusable_buffers_est;
3834 : }
3835 :
3836 : /*
3837 : * Now write out dirty reusable buffers, working forward from the
3838 : * next_to_clean point, until we have lapped the strategy scan, or cleaned
3839 : * enough buffers to match our estimate of the next cycle's allocation
3840 : * requirements, or hit the bgwriter_lru_maxpages limit.
3841 : */
3842 :
3843 21894 : num_to_scan = bufs_to_lap;
3844 21894 : num_written = 0;
3845 21894 : reusable_buffers = reusable_buffers_est;
3846 :
3847 : /* Execute the LRU scan */
3848 3349564 : while (num_to_scan > 0 && reusable_buffers < upcoming_alloc_est)
3849 : {
3850 3327674 : int sync_state = SyncOneBuffer(next_to_clean, true,
3851 : wb_context);
3852 :
3853 3327674 : if (++next_to_clean >= NBuffers)
3854 : {
3855 4466 : next_to_clean = 0;
3856 4466 : next_passes++;
3857 : }
3858 3327674 : num_to_scan--;
3859 :
3860 3327674 : if (sync_state & BUF_WRITTEN)
3861 : {
3862 35202 : reusable_buffers++;
3863 35202 : if (++num_written >= bgwriter_lru_maxpages)
3864 : {
3865 4 : PendingBgWriterStats.maxwritten_clean++;
3866 4 : break;
3867 : }
3868 : }
3869 3292472 : else if (sync_state & BUF_REUSABLE)
3870 2661326 : reusable_buffers++;
3871 : }
3872 :
3873 21894 : PendingBgWriterStats.buf_written_clean += num_written;
3874 :
3875 : #ifdef BGW_DEBUG
3876 : 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",
3877 : recent_alloc, smoothed_alloc, strategy_delta, bufs_ahead,
3878 : smoothed_density, reusable_buffers_est, upcoming_alloc_est,
3879 : bufs_to_lap - num_to_scan,
3880 : num_written,
3881 : reusable_buffers - reusable_buffers_est);
3882 : #endif
3883 :
3884 : /*
3885 : * Consider the above scan as being like a new allocation scan.
3886 : * Characterize its density and update the smoothed one based on it. This
3887 : * effectively halves the moving average period in cases where both the
3888 : * strategy and the background writer are doing some useful scanning,
3889 : * which is helpful because a long memory isn't as desirable on the
3890 : * density estimates.
3891 : */
3892 21894 : new_strategy_delta = bufs_to_lap - num_to_scan;
3893 21894 : new_recent_alloc = reusable_buffers - reusable_buffers_est;
3894 21894 : if (new_strategy_delta > 0 && new_recent_alloc > 0)
3895 : {
3896 18154 : scans_per_alloc = (float) new_strategy_delta / (float) new_recent_alloc;
3897 18154 : smoothed_density += (scans_per_alloc - smoothed_density) /
3898 : smoothing_samples;
3899 :
3900 : #ifdef BGW_DEBUG
3901 : elog(DEBUG2, "bgwriter: cleaner density alloc=%u scan=%ld density=%.2f new smoothed=%.2f",
3902 : new_recent_alloc, new_strategy_delta,
3903 : scans_per_alloc, smoothed_density);
3904 : #endif
3905 : }
3906 :
3907 : /* Return true if OK to hibernate */
3908 21894 : return (bufs_to_lap == 0 && recent_alloc == 0);
3909 : }
3910 :
3911 : /*
3912 : * SyncOneBuffer -- process a single buffer during syncing.
3913 : *
3914 : * If skip_recently_used is true, we don't write currently-pinned buffers, nor
3915 : * buffers marked recently used, as these are not replacement candidates.
3916 : *
3917 : * Returns a bitmask containing the following flag bits:
3918 : * BUF_WRITTEN: we wrote the buffer.
3919 : * BUF_REUSABLE: buffer is available for replacement, ie, it has
3920 : * pin count 0 and usage count 0.
3921 : *
3922 : * (BUF_WRITTEN could be set in error if FlushBuffer finds the buffer clean
3923 : * after locking it, but we don't care all that much.)
3924 : */
3925 : static int
3926 3874846 : SyncOneBuffer(int buf_id, bool skip_recently_used, WritebackContext *wb_context)
3927 : {
3928 3874846 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3929 3874846 : int result = 0;
3930 : uint32 buf_state;
3931 : BufferTag tag;
3932 :
3933 : /* Make sure we can handle the pin */
3934 3874846 : ReservePrivateRefCountEntry();
3935 3874846 : ResourceOwnerEnlarge(CurrentResourceOwner);
3936 :
3937 : /*
3938 : * Check whether buffer needs writing.
3939 : *
3940 : * We can make this check without taking the buffer content lock so long
3941 : * as we mark pages dirty in access methods *before* logging changes with
3942 : * XLogInsert(): if someone marks the buffer dirty just after our check we
3943 : * don't worry because our checkpoint.redo points before log record for
3944 : * upcoming changes and so we are not required to write such dirty buffer.
3945 : */
3946 3874846 : buf_state = LockBufHdr(bufHdr);
3947 :
3948 3874846 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 0 &&
3949 3867960 : BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3950 : {
3951 2700388 : result |= BUF_REUSABLE;
3952 : }
3953 1174458 : else if (skip_recently_used)
3954 : {
3955 : /* Caller told us not to write recently-used buffers */
3956 631146 : UnlockBufHdr(bufHdr, buf_state);
3957 631146 : return result;
3958 : }
3959 :
3960 3243700 : if (!(buf_state & BM_VALID) || !(buf_state & BM_DIRTY))
3961 : {
3962 : /* It's clean, so nothing to do */
3963 2661326 : UnlockBufHdr(bufHdr, buf_state);
3964 2661326 : return result;
3965 : }
3966 :
3967 : /*
3968 : * Pin it, share-lock it, write it. (FlushBuffer will do nothing if the
3969 : * buffer is clean by the time we've locked it.)
3970 : */
3971 582374 : PinBuffer_Locked(bufHdr);
3972 :
3973 582374 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
3974 :
3975 582374 : tag = bufHdr->tag;
3976 :
3977 582374 : UnpinBuffer(bufHdr);
3978 :
3979 : /*
3980 : * SyncOneBuffer() is only called by checkpointer and bgwriter, so
3981 : * IOContext will always be IOCONTEXT_NORMAL.
3982 : */
3983 582374 : ScheduleBufferTagForWriteback(wb_context, IOCONTEXT_NORMAL, &tag);
3984 :
3985 582374 : return result | BUF_WRITTEN;
3986 : }
3987 :
3988 : /*
3989 : * AtEOXact_Buffers - clean up at end of transaction.
3990 : *
3991 : * As of PostgreSQL 8.0, buffer pins should get released by the
3992 : * ResourceOwner mechanism. This routine is just a debugging
3993 : * cross-check that no pins remain.
3994 : */
3995 : void
3996 1100036 : AtEOXact_Buffers(bool isCommit)
3997 : {
3998 1100036 : CheckForBufferLeaks();
3999 :
4000 1100036 : AtEOXact_LocalBuffers(isCommit);
4001 :
4002 : Assert(PrivateRefCountOverflowed == 0);
4003 1100036 : }
4004 :
4005 : /*
4006 : * Initialize access to shared buffer pool
4007 : *
4008 : * This is called during backend startup (whether standalone or under the
4009 : * postmaster). It sets up for this backend's access to the already-existing
4010 : * buffer pool.
4011 : */
4012 : void
4013 44438 : InitBufferManagerAccess(void)
4014 : {
4015 : HASHCTL hash_ctl;
4016 :
4017 : /*
4018 : * An advisory limit on the number of pins each backend should hold, based
4019 : * on shared_buffers and the maximum number of connections possible.
4020 : * That's very pessimistic, but outside toy-sized shared_buffers it should
4021 : * allow plenty of pins. LimitAdditionalPins() and
4022 : * GetAdditionalPinLimit() can be used to check the remaining balance.
4023 : */
4024 44438 : MaxProportionalPins = NBuffers / (MaxBackends + NUM_AUXILIARY_PROCS);
4025 :
4026 44438 : memset(&PrivateRefCountArray, 0, sizeof(PrivateRefCountArray));
4027 :
4028 44438 : hash_ctl.keysize = sizeof(int32);
4029 44438 : hash_ctl.entrysize = sizeof(PrivateRefCountEntry);
4030 :
4031 44438 : PrivateRefCountHash = hash_create("PrivateRefCount", 100, &hash_ctl,
4032 : HASH_ELEM | HASH_BLOBS);
4033 :
4034 : /*
4035 : * AtProcExit_Buffers needs LWLock access, and thereby has to be called at
4036 : * the corresponding phase of backend shutdown.
4037 : */
4038 : Assert(MyProc != NULL);
4039 44438 : on_shmem_exit(AtProcExit_Buffers, 0);
4040 44438 : }
4041 :
4042 : /*
4043 : * During backend exit, ensure that we released all shared-buffer locks and
4044 : * assert that we have no remaining pins.
4045 : */
4046 : static void
4047 44438 : AtProcExit_Buffers(int code, Datum arg)
4048 : {
4049 44438 : UnlockBuffers();
4050 :
4051 44438 : CheckForBufferLeaks();
4052 :
4053 : /* localbuf.c needs a chance too */
4054 44438 : AtProcExit_LocalBuffers();
4055 44438 : }
4056 :
4057 : /*
4058 : * CheckForBufferLeaks - ensure this backend holds no buffer pins
4059 : *
4060 : * As of PostgreSQL 8.0, buffer pins should get released by the
4061 : * ResourceOwner mechanism. This routine is just a debugging
4062 : * cross-check that no pins remain.
4063 : */
4064 : static void
4065 1144474 : CheckForBufferLeaks(void)
4066 : {
4067 : #ifdef USE_ASSERT_CHECKING
4068 : int RefCountErrors = 0;
4069 : PrivateRefCountEntry *res;
4070 : int i;
4071 : char *s;
4072 :
4073 : /* check the array */
4074 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4075 : {
4076 : res = &PrivateRefCountArray[i];
4077 :
4078 : if (res->buffer != InvalidBuffer)
4079 : {
4080 : s = DebugPrintBufferRefcount(res->buffer);
4081 : elog(WARNING, "buffer refcount leak: %s", s);
4082 : pfree(s);
4083 :
4084 : RefCountErrors++;
4085 : }
4086 : }
4087 :
4088 : /* if necessary search the hash */
4089 : if (PrivateRefCountOverflowed)
4090 : {
4091 : HASH_SEQ_STATUS hstat;
4092 :
4093 : hash_seq_init(&hstat, PrivateRefCountHash);
4094 : while ((res = (PrivateRefCountEntry *) hash_seq_search(&hstat)) != NULL)
4095 : {
4096 : s = DebugPrintBufferRefcount(res->buffer);
4097 : elog(WARNING, "buffer refcount leak: %s", s);
4098 : pfree(s);
4099 : RefCountErrors++;
4100 : }
4101 : }
4102 :
4103 : Assert(RefCountErrors == 0);
4104 : #endif
4105 1144474 : }
4106 :
4107 : #ifdef USE_ASSERT_CHECKING
4108 : /*
4109 : * Check for exclusive-locked catalog buffers. This is the core of
4110 : * AssertCouldGetRelation().
4111 : *
4112 : * A backend would self-deadlock on LWLocks if the catalog scan read the
4113 : * exclusive-locked buffer. The main threat is exclusive-locked buffers of
4114 : * catalogs used in relcache, because a catcache search on any catalog may
4115 : * build that catalog's relcache entry. We don't have an inventory of
4116 : * catalogs relcache uses, so just check buffers of most catalogs.
4117 : *
4118 : * It's better to minimize waits while holding an exclusive buffer lock, so it
4119 : * would be nice to broaden this check not to be catalog-specific. However,
4120 : * bttextcmp() accesses pg_collation, and non-core opclasses might similarly
4121 : * read tables. That is deadlock-free as long as there's no loop in the
4122 : * dependency graph: modifying table A may cause an opclass to read table B,
4123 : * but it must not cause a read of table A.
4124 : */
4125 : void
4126 : AssertBufferLocksPermitCatalogRead(void)
4127 : {
4128 : ForEachLWLockHeldByMe(AssertNotCatalogBufferLock, NULL);
4129 : }
4130 :
4131 : static void
4132 : AssertNotCatalogBufferLock(LWLock *lock, LWLockMode mode,
4133 : void *unused_context)
4134 : {
4135 : BufferDesc *bufHdr;
4136 : BufferTag tag;
4137 : Oid relid;
4138 :
4139 : if (mode != LW_EXCLUSIVE)
4140 : return;
4141 :
4142 : if (!((BufferDescPadded *) lock > BufferDescriptors &&
4143 : (BufferDescPadded *) lock < BufferDescriptors + NBuffers))
4144 : return; /* not a buffer lock */
4145 :
4146 : bufHdr = (BufferDesc *)
4147 : ((char *) lock - offsetof(BufferDesc, content_lock));
4148 : tag = bufHdr->tag;
4149 :
4150 : /*
4151 : * This relNumber==relid assumption holds until a catalog experiences
4152 : * VACUUM FULL or similar. After a command like that, relNumber will be
4153 : * in the normal (non-catalog) range, and we lose the ability to detect
4154 : * hazardous access to that catalog. Calling RelidByRelfilenumber() would
4155 : * close that gap, but RelidByRelfilenumber() might then deadlock with a
4156 : * held lock.
4157 : */
4158 : relid = tag.relNumber;
4159 :
4160 : if (IsCatalogTextUniqueIndexOid(relid)) /* see comments at the callee */
4161 : return;
4162 :
4163 : Assert(!IsCatalogRelationOid(relid));
4164 : }
4165 : #endif
4166 :
4167 :
4168 : /*
4169 : * Helper routine to issue warnings when a buffer is unexpectedly pinned
4170 : */
4171 : char *
4172 80 : DebugPrintBufferRefcount(Buffer buffer)
4173 : {
4174 : BufferDesc *buf;
4175 : int32 loccount;
4176 : char *result;
4177 : ProcNumber backend;
4178 : uint32 buf_state;
4179 :
4180 : Assert(BufferIsValid(buffer));
4181 80 : if (BufferIsLocal(buffer))
4182 : {
4183 32 : buf = GetLocalBufferDescriptor(-buffer - 1);
4184 32 : loccount = LocalRefCount[-buffer - 1];
4185 32 : backend = MyProcNumber;
4186 : }
4187 : else
4188 : {
4189 48 : buf = GetBufferDescriptor(buffer - 1);
4190 48 : loccount = GetPrivateRefCount(buffer);
4191 48 : backend = INVALID_PROC_NUMBER;
4192 : }
4193 :
4194 : /* theoretically we should lock the bufhdr here */
4195 80 : buf_state = pg_atomic_read_u32(&buf->state);
4196 :
4197 80 : result = psprintf("[%03d] (rel=%s, blockNum=%u, flags=0x%x, refcount=%u %d)",
4198 : buffer,
4199 80 : relpathbackend(BufTagGetRelFileLocator(&buf->tag), backend,
4200 : BufTagGetForkNum(&buf->tag)).str,
4201 : buf->tag.blockNum, buf_state & BUF_FLAG_MASK,
4202 : BUF_STATE_GET_REFCOUNT(buf_state), loccount);
4203 80 : return result;
4204 : }
4205 :
4206 : /*
4207 : * CheckPointBuffers
4208 : *
4209 : * Flush all dirty blocks in buffer pool to disk at checkpoint time.
4210 : *
4211 : * Note: temporary relations do not participate in checkpoints, so they don't
4212 : * need to be flushed.
4213 : */
4214 : void
4215 3430 : CheckPointBuffers(int flags)
4216 : {
4217 3430 : BufferSync(flags);
4218 3430 : }
4219 :
4220 : /*
4221 : * BufferGetBlockNumber
4222 : * Returns the block number associated with a buffer.
4223 : *
4224 : * Note:
4225 : * Assumes that the buffer is valid and pinned, else the
4226 : * value may be obsolete immediately...
4227 : */
4228 : BlockNumber
4229 100543124 : BufferGetBlockNumber(Buffer buffer)
4230 : {
4231 : BufferDesc *bufHdr;
4232 :
4233 : Assert(BufferIsPinned(buffer));
4234 :
4235 100543124 : if (BufferIsLocal(buffer))
4236 3809354 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4237 : else
4238 96733770 : bufHdr = GetBufferDescriptor(buffer - 1);
4239 :
4240 : /* pinned, so OK to read tag without spinlock */
4241 100543124 : return bufHdr->tag.blockNum;
4242 : }
4243 :
4244 : /*
4245 : * BufferGetTag
4246 : * Returns the relfilelocator, fork number and block number associated with
4247 : * a buffer.
4248 : */
4249 : void
4250 30114838 : BufferGetTag(Buffer buffer, RelFileLocator *rlocator, ForkNumber *forknum,
4251 : BlockNumber *blknum)
4252 : {
4253 : BufferDesc *bufHdr;
4254 :
4255 : /* Do the same checks as BufferGetBlockNumber. */
4256 : Assert(BufferIsPinned(buffer));
4257 :
4258 30114838 : if (BufferIsLocal(buffer))
4259 0 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4260 : else
4261 30114838 : bufHdr = GetBufferDescriptor(buffer - 1);
4262 :
4263 : /* pinned, so OK to read tag without spinlock */
4264 30114838 : *rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
4265 30114838 : *forknum = BufTagGetForkNum(&bufHdr->tag);
4266 30114838 : *blknum = bufHdr->tag.blockNum;
4267 30114838 : }
4268 :
4269 : /*
4270 : * FlushBuffer
4271 : * Physically write out a shared buffer.
4272 : *
4273 : * NOTE: this actually just passes the buffer contents to the kernel; the
4274 : * real write to disk won't happen until the kernel feels like it. This
4275 : * is okay from our point of view since we can redo the changes from WAL.
4276 : * However, we will need to force the changes to disk via fsync before
4277 : * we can checkpoint WAL.
4278 : *
4279 : * The caller must hold a pin on the buffer and have share-locked the
4280 : * buffer contents. (Note: a share-lock does not prevent updates of
4281 : * hint bits in the buffer, so the page could change while the write
4282 : * is in progress, but we assume that that will not invalidate the data
4283 : * written.)
4284 : *
4285 : * If the caller has an smgr reference for the buffer's relation, pass it
4286 : * as the second parameter. If not, pass NULL.
4287 : */
4288 : static void
4289 1102696 : FlushBuffer(BufferDesc *buf, SMgrRelation reln, IOObject io_object,
4290 : IOContext io_context)
4291 : {
4292 : XLogRecPtr recptr;
4293 : ErrorContextCallback errcallback;
4294 : instr_time io_start;
4295 : Block bufBlock;
4296 : char *bufToWrite;
4297 : uint32 buf_state;
4298 :
4299 : /*
4300 : * Try to start an I/O operation. If StartBufferIO returns false, then
4301 : * someone else flushed the buffer before we could, so we need not do
4302 : * anything.
4303 : */
4304 1102696 : if (!StartBufferIO(buf, false, false))
4305 10 : return;
4306 :
4307 : /* Setup error traceback support for ereport() */
4308 1102686 : errcallback.callback = shared_buffer_write_error_callback;
4309 1102686 : errcallback.arg = buf;
4310 1102686 : errcallback.previous = error_context_stack;
4311 1102686 : error_context_stack = &errcallback;
4312 :
4313 : /* Find smgr relation for buffer */
4314 1102686 : if (reln == NULL)
4315 1094412 : reln = smgropen(BufTagGetRelFileLocator(&buf->tag), INVALID_PROC_NUMBER);
4316 :
4317 : TRACE_POSTGRESQL_BUFFER_FLUSH_START(BufTagGetForkNum(&buf->tag),
4318 : buf->tag.blockNum,
4319 : reln->smgr_rlocator.locator.spcOid,
4320 : reln->smgr_rlocator.locator.dbOid,
4321 : reln->smgr_rlocator.locator.relNumber);
4322 :
4323 1102686 : buf_state = LockBufHdr(buf);
4324 :
4325 : /*
4326 : * Run PageGetLSN while holding header lock, since we don't have the
4327 : * buffer locked exclusively in all cases.
4328 : */
4329 1102686 : recptr = BufferGetLSN(buf);
4330 :
4331 : /* To check if block content changes while flushing. - vadim 01/17/97 */
4332 1102686 : buf_state &= ~BM_JUST_DIRTIED;
4333 1102686 : UnlockBufHdr(buf, buf_state);
4334 :
4335 : /*
4336 : * Force XLOG flush up to buffer's LSN. This implements the basic WAL
4337 : * rule that log updates must hit disk before any of the data-file changes
4338 : * they describe do.
4339 : *
4340 : * However, this rule does not apply to unlogged relations, which will be
4341 : * lost after a crash anyway. Most unlogged relation pages do not bear
4342 : * LSNs since we never emit WAL records for them, and therefore flushing
4343 : * up through the buffer LSN would be useless, but harmless. However,
4344 : * GiST indexes use LSNs internally to track page-splits, and therefore
4345 : * unlogged GiST pages bear "fake" LSNs generated by
4346 : * GetFakeLSNForUnloggedRel. It is unlikely but possible that the fake
4347 : * LSN counter could advance past the WAL insertion point; and if it did
4348 : * happen, attempting to flush WAL through that location would fail, with
4349 : * disastrous system-wide consequences. To make sure that can't happen,
4350 : * skip the flush if the buffer isn't permanent.
4351 : */
4352 1102686 : if (buf_state & BM_PERMANENT)
4353 1099134 : XLogFlush(recptr);
4354 :
4355 : /*
4356 : * Now it's safe to write the buffer to disk. Note that no one else should
4357 : * have been able to write it, while we were busy with log flushing,
4358 : * because we got the exclusive right to perform I/O by setting the
4359 : * BM_IO_IN_PROGRESS bit.
4360 : */
4361 1102686 : bufBlock = BufHdrGetBlock(buf);
4362 :
4363 : /*
4364 : * Update page checksum if desired. Since we have only shared lock on the
4365 : * buffer, other processes might be updating hint bits in it, so we must
4366 : * copy the page to private storage if we do checksumming.
4367 : */
4368 1102686 : bufToWrite = PageSetChecksumCopy((Page) bufBlock, buf->tag.blockNum);
4369 :
4370 1102686 : io_start = pgstat_prepare_io_time(track_io_timing);
4371 :
4372 : /*
4373 : * bufToWrite is either the shared buffer or a copy, as appropriate.
4374 : */
4375 1102686 : smgrwrite(reln,
4376 1102686 : BufTagGetForkNum(&buf->tag),
4377 : buf->tag.blockNum,
4378 : bufToWrite,
4379 : false);
4380 :
4381 : /*
4382 : * When a strategy is in use, only flushes of dirty buffers already in the
4383 : * strategy ring are counted as strategy writes (IOCONTEXT
4384 : * [BULKREAD|BULKWRITE|VACUUM] IOOP_WRITE) for the purpose of IO
4385 : * statistics tracking.
4386 : *
4387 : * If a shared buffer initially added to the ring must be flushed before
4388 : * being used, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE.
4389 : *
4390 : * If a shared buffer which was added to the ring later because the
4391 : * current strategy buffer is pinned or in use or because all strategy
4392 : * buffers were dirty and rejected (for BAS_BULKREAD operations only)
4393 : * requires flushing, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE
4394 : * (from_ring will be false).
4395 : *
4396 : * When a strategy is not in use, the write can only be a "regular" write
4397 : * of a dirty shared buffer (IOCONTEXT_NORMAL IOOP_WRITE).
4398 : */
4399 1102686 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
4400 : IOOP_WRITE, io_start, 1, BLCKSZ);
4401 :
4402 1102686 : pgBufferUsage.shared_blks_written++;
4403 :
4404 : /*
4405 : * Mark the buffer as clean (unless BM_JUST_DIRTIED has become set) and
4406 : * end the BM_IO_IN_PROGRESS state.
4407 : */
4408 1102686 : TerminateBufferIO(buf, true, 0, true, false);
4409 :
4410 : TRACE_POSTGRESQL_BUFFER_FLUSH_DONE(BufTagGetForkNum(&buf->tag),
4411 : buf->tag.blockNum,
4412 : reln->smgr_rlocator.locator.spcOid,
4413 : reln->smgr_rlocator.locator.dbOid,
4414 : reln->smgr_rlocator.locator.relNumber);
4415 :
4416 : /* Pop the error context stack */
4417 1102686 : error_context_stack = errcallback.previous;
4418 : }
4419 :
4420 : /*
4421 : * Convenience wrapper around FlushBuffer() that locks/unlocks the buffer
4422 : * before/after calling FlushBuffer().
4423 : */
4424 : static void
4425 592608 : FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
4426 : IOObject io_object, IOContext io_context)
4427 : {
4428 592608 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_SHARED);
4429 592608 : FlushBuffer(buf, reln, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
4430 592608 : LWLockRelease(BufferDescriptorGetContentLock(buf));
4431 592608 : }
4432 :
4433 : /*
4434 : * RelationGetNumberOfBlocksInFork
4435 : * Determines the current number of pages in the specified relation fork.
4436 : *
4437 : * Note that the accuracy of the result will depend on the details of the
4438 : * relation's storage. For builtin AMs it'll be accurate, but for external AMs
4439 : * it might not be.
4440 : */
4441 : BlockNumber
4442 3743956 : RelationGetNumberOfBlocksInFork(Relation relation, ForkNumber forkNum)
4443 : {
4444 3743956 : if (RELKIND_HAS_TABLE_AM(relation->rd_rel->relkind))
4445 : {
4446 : /*
4447 : * Not every table AM uses BLCKSZ wide fixed size blocks. Therefore
4448 : * tableam returns the size in bytes - but for the purpose of this
4449 : * routine, we want the number of blocks. Therefore divide, rounding
4450 : * up.
4451 : */
4452 : uint64 szbytes;
4453 :
4454 2847052 : szbytes = table_relation_size(relation, forkNum);
4455 :
4456 2847014 : return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
4457 : }
4458 896904 : else if (RELKIND_HAS_STORAGE(relation->rd_rel->relkind))
4459 : {
4460 896904 : return smgrnblocks(RelationGetSmgr(relation), forkNum);
4461 : }
4462 : else
4463 : Assert(false);
4464 :
4465 0 : return 0; /* keep compiler quiet */
4466 : }
4467 :
4468 : /*
4469 : * BufferIsPermanent
4470 : * Determines whether a buffer will potentially still be around after
4471 : * a crash. Caller must hold a buffer pin.
4472 : */
4473 : bool
4474 19126068 : BufferIsPermanent(Buffer buffer)
4475 : {
4476 : BufferDesc *bufHdr;
4477 :
4478 : /* Local buffers are used only for temp relations. */
4479 19126068 : if (BufferIsLocal(buffer))
4480 1253940 : return false;
4481 :
4482 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4483 : Assert(BufferIsValid(buffer));
4484 : Assert(BufferIsPinned(buffer));
4485 :
4486 : /*
4487 : * BM_PERMANENT can't be changed while we hold a pin on the buffer, so we
4488 : * need not bother with the buffer header spinlock. Even if someone else
4489 : * changes the buffer header state while we're doing this, the state is
4490 : * changed atomically, so we'll read the old value or the new value, but
4491 : * not random garbage.
4492 : */
4493 17872128 : bufHdr = GetBufferDescriptor(buffer - 1);
4494 17872128 : return (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT) != 0;
4495 : }
4496 :
4497 : /*
4498 : * BufferGetLSNAtomic
4499 : * Retrieves the LSN of the buffer atomically using a buffer header lock.
4500 : * This is necessary for some callers who may not have an exclusive lock
4501 : * on the buffer.
4502 : */
4503 : XLogRecPtr
4504 14196788 : BufferGetLSNAtomic(Buffer buffer)
4505 : {
4506 14196788 : char *page = BufferGetPage(buffer);
4507 : BufferDesc *bufHdr;
4508 : XLogRecPtr lsn;
4509 : uint32 buf_state;
4510 :
4511 : /*
4512 : * If we don't need locking for correctness, fastpath out.
4513 : */
4514 14196788 : if (!XLogHintBitIsNeeded() || BufferIsLocal(buffer))
4515 475456 : return PageGetLSN(page);
4516 :
4517 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4518 : Assert(BufferIsValid(buffer));
4519 : Assert(BufferIsPinned(buffer));
4520 :
4521 13721332 : bufHdr = GetBufferDescriptor(buffer - 1);
4522 13721332 : buf_state = LockBufHdr(bufHdr);
4523 13721332 : lsn = PageGetLSN(page);
4524 13721332 : UnlockBufHdr(bufHdr, buf_state);
4525 :
4526 13721332 : return lsn;
4527 : }
4528 :
4529 : /* ---------------------------------------------------------------------
4530 : * DropRelationBuffers
4531 : *
4532 : * This function removes from the buffer pool all the pages of the
4533 : * specified relation forks that have block numbers >= firstDelBlock.
4534 : * (In particular, with firstDelBlock = 0, all pages are removed.)
4535 : * Dirty pages are simply dropped, without bothering to write them
4536 : * out first. Therefore, this is NOT rollback-able, and so should be
4537 : * used only with extreme caution!
4538 : *
4539 : * Currently, this is called only from smgr.c when the underlying file
4540 : * is about to be deleted or truncated (firstDelBlock is needed for
4541 : * the truncation case). The data in the affected pages would therefore
4542 : * be deleted momentarily anyway, and there is no point in writing it.
4543 : * It is the responsibility of higher-level code to ensure that the
4544 : * deletion or truncation does not lose any data that could be needed
4545 : * later. It is also the responsibility of higher-level code to ensure
4546 : * that no other process could be trying to load more pages of the
4547 : * relation into buffers.
4548 : * --------------------------------------------------------------------
4549 : */
4550 : void
4551 1252 : DropRelationBuffers(SMgrRelation smgr_reln, ForkNumber *forkNum,
4552 : int nforks, BlockNumber *firstDelBlock)
4553 : {
4554 : int i;
4555 : int j;
4556 : RelFileLocatorBackend rlocator;
4557 : BlockNumber nForkBlock[MAX_FORKNUM];
4558 1252 : uint64 nBlocksToInvalidate = 0;
4559 :
4560 1252 : rlocator = smgr_reln->smgr_rlocator;
4561 :
4562 : /* If it's a local relation, it's localbuf.c's problem. */
4563 1252 : if (RelFileLocatorBackendIsTemp(rlocator))
4564 : {
4565 748 : if (rlocator.backend == MyProcNumber)
4566 748 : DropRelationLocalBuffers(rlocator.locator, forkNum, nforks,
4567 : firstDelBlock);
4568 :
4569 820 : return;
4570 : }
4571 :
4572 : /*
4573 : * To remove all the pages of the specified relation forks from the buffer
4574 : * pool, we need to scan the entire buffer pool but we can optimize it by
4575 : * finding the buffers from BufMapping table provided we know the exact
4576 : * size of each fork of the relation. The exact size is required to ensure
4577 : * that we don't leave any buffer for the relation being dropped as
4578 : * otherwise the background writer or checkpointer can lead to a PANIC
4579 : * error while flushing buffers corresponding to files that don't exist.
4580 : *
4581 : * To know the exact size, we rely on the size cached for each fork by us
4582 : * during recovery which limits the optimization to recovery and on
4583 : * standbys but we can easily extend it once we have shared cache for
4584 : * relation size.
4585 : *
4586 : * In recovery, we cache the value returned by the first lseek(SEEK_END)
4587 : * and the future writes keeps the cached value up-to-date. See
4588 : * smgrextend. It is possible that the value of the first lseek is smaller
4589 : * than the actual number of existing blocks in the file due to buggy
4590 : * Linux kernels that might not have accounted for the recent write. But
4591 : * that should be fine because there must not be any buffers after that
4592 : * file size.
4593 : */
4594 678 : for (i = 0; i < nforks; i++)
4595 : {
4596 : /* Get the number of blocks for a relation's fork */
4597 582 : nForkBlock[i] = smgrnblocks_cached(smgr_reln, forkNum[i]);
4598 :
4599 582 : if (nForkBlock[i] == InvalidBlockNumber)
4600 : {
4601 408 : nBlocksToInvalidate = InvalidBlockNumber;
4602 408 : break;
4603 : }
4604 :
4605 : /* calculate the number of blocks to be invalidated */
4606 174 : nBlocksToInvalidate += (nForkBlock[i] - firstDelBlock[i]);
4607 : }
4608 :
4609 : /*
4610 : * We apply the optimization iff the total number of blocks to invalidate
4611 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4612 : */
4613 504 : if (BlockNumberIsValid(nBlocksToInvalidate) &&
4614 96 : nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4615 : {
4616 192 : for (j = 0; j < nforks; j++)
4617 120 : FindAndDropRelationBuffers(rlocator.locator, forkNum[j],
4618 120 : nForkBlock[j], firstDelBlock[j]);
4619 72 : return;
4620 : }
4621 :
4622 5745328 : for (i = 0; i < NBuffers; i++)
4623 : {
4624 5744896 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4625 : uint32 buf_state;
4626 :
4627 : /*
4628 : * We can make this a tad faster by prechecking the buffer tag before
4629 : * we attempt to lock the buffer; this saves a lot of lock
4630 : * acquisitions in typical cases. It should be safe because the
4631 : * caller must have AccessExclusiveLock on the relation, or some other
4632 : * reason to be certain that no one is loading new pages of the rel
4633 : * into the buffer pool. (Otherwise we might well miss such pages
4634 : * entirely.) Therefore, while the tag might be changing while we
4635 : * look at it, it can't be changing *to* a value we care about, only
4636 : * *away* from such a value. So false negatives are impossible, and
4637 : * false positives are safe because we'll recheck after getting the
4638 : * buffer lock.
4639 : *
4640 : * We could check forkNum and blockNum as well as the rlocator, but
4641 : * the incremental win from doing so seems small.
4642 : */
4643 5744896 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator))
4644 5731018 : continue;
4645 :
4646 13878 : buf_state = LockBufHdr(bufHdr);
4647 :
4648 35964 : for (j = 0; j < nforks; j++)
4649 : {
4650 25162 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator) &&
4651 25162 : BufTagGetForkNum(&bufHdr->tag) == forkNum[j] &&
4652 13682 : bufHdr->tag.blockNum >= firstDelBlock[j])
4653 : {
4654 3076 : InvalidateBuffer(bufHdr); /* releases spinlock */
4655 3076 : break;
4656 : }
4657 : }
4658 13878 : if (j >= nforks)
4659 10802 : UnlockBufHdr(bufHdr, buf_state);
4660 : }
4661 : }
4662 :
4663 : /* ---------------------------------------------------------------------
4664 : * DropRelationsAllBuffers
4665 : *
4666 : * This function removes from the buffer pool all the pages of all
4667 : * forks of the specified relations. It's equivalent to calling
4668 : * DropRelationBuffers once per fork per relation with firstDelBlock = 0.
4669 : * --------------------------------------------------------------------
4670 : */
4671 : void
4672 27176 : DropRelationsAllBuffers(SMgrRelation *smgr_reln, int nlocators)
4673 : {
4674 : int i;
4675 27176 : int n = 0;
4676 : SMgrRelation *rels;
4677 : BlockNumber (*block)[MAX_FORKNUM + 1];
4678 27176 : uint64 nBlocksToInvalidate = 0;
4679 : RelFileLocator *locators;
4680 27176 : bool cached = true;
4681 : bool use_bsearch;
4682 :
4683 27176 : if (nlocators == 0)
4684 0 : return;
4685 :
4686 27176 : rels = palloc(sizeof(SMgrRelation) * nlocators); /* non-local relations */
4687 :
4688 : /* If it's a local relation, it's localbuf.c's problem. */
4689 118948 : for (i = 0; i < nlocators; i++)
4690 : {
4691 91772 : if (RelFileLocatorBackendIsTemp(smgr_reln[i]->smgr_rlocator))
4692 : {
4693 6348 : if (smgr_reln[i]->smgr_rlocator.backend == MyProcNumber)
4694 6348 : DropRelationAllLocalBuffers(smgr_reln[i]->smgr_rlocator.locator);
4695 : }
4696 : else
4697 85424 : rels[n++] = smgr_reln[i];
4698 : }
4699 :
4700 : /*
4701 : * If there are no non-local relations, then we're done. Release the
4702 : * memory and return.
4703 : */
4704 27176 : if (n == 0)
4705 : {
4706 1654 : pfree(rels);
4707 1654 : return;
4708 : }
4709 :
4710 : /*
4711 : * This is used to remember the number of blocks for all the relations
4712 : * forks.
4713 : */
4714 : block = (BlockNumber (*)[MAX_FORKNUM + 1])
4715 25522 : palloc(sizeof(BlockNumber) * n * (MAX_FORKNUM + 1));
4716 :
4717 : /*
4718 : * We can avoid scanning the entire buffer pool if we know the exact size
4719 : * of each of the given relation forks. See DropRelationBuffers.
4720 : */
4721 53366 : for (i = 0; i < n && cached; i++)
4722 : {
4723 43256 : for (int j = 0; j <= MAX_FORKNUM; j++)
4724 : {
4725 : /* Get the number of blocks for a relation's fork. */
4726 39434 : block[i][j] = smgrnblocks_cached(rels[i], j);
4727 :
4728 : /* We need to only consider the relation forks that exists. */
4729 39434 : if (block[i][j] == InvalidBlockNumber)
4730 : {
4731 35278 : if (!smgrexists(rels[i], j))
4732 11256 : continue;
4733 24022 : cached = false;
4734 24022 : break;
4735 : }
4736 :
4737 : /* calculate the total number of blocks to be invalidated */
4738 4156 : nBlocksToInvalidate += block[i][j];
4739 : }
4740 : }
4741 :
4742 : /*
4743 : * We apply the optimization iff the total number of blocks to invalidate
4744 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4745 : */
4746 25522 : if (cached && nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4747 : {
4748 2520 : for (i = 0; i < n; i++)
4749 : {
4750 6940 : for (int j = 0; j <= MAX_FORKNUM; j++)
4751 : {
4752 : /* ignore relation forks that doesn't exist */
4753 5552 : if (!BlockNumberIsValid(block[i][j]))
4754 4146 : continue;
4755 :
4756 : /* drop all the buffers for a particular relation fork */
4757 1406 : FindAndDropRelationBuffers(rels[i]->smgr_rlocator.locator,
4758 1406 : j, block[i][j], 0);
4759 : }
4760 : }
4761 :
4762 1132 : pfree(block);
4763 1132 : pfree(rels);
4764 1132 : return;
4765 : }
4766 :
4767 24390 : pfree(block);
4768 24390 : locators = palloc(sizeof(RelFileLocator) * n); /* non-local relations */
4769 108426 : for (i = 0; i < n; i++)
4770 84036 : locators[i] = rels[i]->smgr_rlocator.locator;
4771 :
4772 : /*
4773 : * For low number of relations to drop just use a simple walk through, to
4774 : * save the bsearch overhead. The threshold to use is rather a guess than
4775 : * an exactly determined value, as it depends on many factors (CPU and RAM
4776 : * speeds, amount of shared buffers etc.).
4777 : */
4778 24390 : use_bsearch = n > RELS_BSEARCH_THRESHOLD;
4779 :
4780 : /* sort the list of rlocators if necessary */
4781 24390 : if (use_bsearch)
4782 344 : qsort(locators, n, sizeof(RelFileLocator), rlocator_comparator);
4783 :
4784 264380230 : for (i = 0; i < NBuffers; i++)
4785 : {
4786 264355840 : RelFileLocator *rlocator = NULL;
4787 264355840 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4788 : uint32 buf_state;
4789 :
4790 : /*
4791 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4792 : * saves some cycles.
4793 : */
4794 :
4795 264355840 : if (!use_bsearch)
4796 : {
4797 : int j;
4798 :
4799 1053447270 : for (j = 0; j < n; j++)
4800 : {
4801 792954134 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &locators[j]))
4802 : {
4803 177328 : rlocator = &locators[j];
4804 177328 : break;
4805 : }
4806 : }
4807 : }
4808 : else
4809 : {
4810 : RelFileLocator locator;
4811 :
4812 3685376 : locator = BufTagGetRelFileLocator(&bufHdr->tag);
4813 3685376 : rlocator = bsearch(&locator,
4814 : locators, n, sizeof(RelFileLocator),
4815 : rlocator_comparator);
4816 : }
4817 :
4818 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
4819 264355840 : if (rlocator == NULL)
4820 264175012 : continue;
4821 :
4822 180828 : buf_state = LockBufHdr(bufHdr);
4823 180828 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, rlocator))
4824 180828 : InvalidateBuffer(bufHdr); /* releases spinlock */
4825 : else
4826 0 : UnlockBufHdr(bufHdr, buf_state);
4827 : }
4828 :
4829 24390 : pfree(locators);
4830 24390 : pfree(rels);
4831 : }
4832 :
4833 : /* ---------------------------------------------------------------------
4834 : * FindAndDropRelationBuffers
4835 : *
4836 : * This function performs look up in BufMapping table and removes from the
4837 : * buffer pool all the pages of the specified relation fork that has block
4838 : * number >= firstDelBlock. (In particular, with firstDelBlock = 0, all
4839 : * pages are removed.)
4840 : * --------------------------------------------------------------------
4841 : */
4842 : static void
4843 1526 : FindAndDropRelationBuffers(RelFileLocator rlocator, ForkNumber forkNum,
4844 : BlockNumber nForkBlock,
4845 : BlockNumber firstDelBlock)
4846 : {
4847 : BlockNumber curBlock;
4848 :
4849 3722 : for (curBlock = firstDelBlock; curBlock < nForkBlock; curBlock++)
4850 : {
4851 : uint32 bufHash; /* hash value for tag */
4852 : BufferTag bufTag; /* identity of requested block */
4853 : LWLock *bufPartitionLock; /* buffer partition lock for it */
4854 : int buf_id;
4855 : BufferDesc *bufHdr;
4856 : uint32 buf_state;
4857 :
4858 : /* create a tag so we can lookup the buffer */
4859 2196 : InitBufferTag(&bufTag, &rlocator, forkNum, curBlock);
4860 :
4861 : /* determine its hash code and partition lock ID */
4862 2196 : bufHash = BufTableHashCode(&bufTag);
4863 2196 : bufPartitionLock = BufMappingPartitionLock(bufHash);
4864 :
4865 : /* Check that it is in the buffer pool. If not, do nothing. */
4866 2196 : LWLockAcquire(bufPartitionLock, LW_SHARED);
4867 2196 : buf_id = BufTableLookup(&bufTag, bufHash);
4868 2196 : LWLockRelease(bufPartitionLock);
4869 :
4870 2196 : if (buf_id < 0)
4871 188 : continue;
4872 :
4873 2008 : bufHdr = GetBufferDescriptor(buf_id);
4874 :
4875 : /*
4876 : * We need to lock the buffer header and recheck if the buffer is
4877 : * still associated with the same block because the buffer could be
4878 : * evicted by some other backend loading blocks for a different
4879 : * relation after we release lock on the BufMapping table.
4880 : */
4881 2008 : buf_state = LockBufHdr(bufHdr);
4882 :
4883 4016 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator) &&
4884 2008 : BufTagGetForkNum(&bufHdr->tag) == forkNum &&
4885 2008 : bufHdr->tag.blockNum >= firstDelBlock)
4886 2008 : InvalidateBuffer(bufHdr); /* releases spinlock */
4887 : else
4888 0 : UnlockBufHdr(bufHdr, buf_state);
4889 : }
4890 1526 : }
4891 :
4892 : /* ---------------------------------------------------------------------
4893 : * DropDatabaseBuffers
4894 : *
4895 : * This function removes all the buffers in the buffer cache for a
4896 : * particular database. Dirty pages are simply dropped, without
4897 : * bothering to write them out first. This is used when we destroy a
4898 : * database, to avoid trying to flush data to disk when the directory
4899 : * tree no longer exists. Implementation is pretty similar to
4900 : * DropRelationBuffers() which is for destroying just one relation.
4901 : * --------------------------------------------------------------------
4902 : */
4903 : void
4904 144 : DropDatabaseBuffers(Oid dbid)
4905 : {
4906 : int i;
4907 :
4908 : /*
4909 : * We needn't consider local buffers, since by assumption the target
4910 : * database isn't our own.
4911 : */
4912 :
4913 961424 : for (i = 0; i < NBuffers; i++)
4914 : {
4915 961280 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4916 : uint32 buf_state;
4917 :
4918 : /*
4919 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4920 : * saves some cycles.
4921 : */
4922 961280 : if (bufHdr->tag.dbOid != dbid)
4923 936468 : continue;
4924 :
4925 24812 : buf_state = LockBufHdr(bufHdr);
4926 24812 : if (bufHdr->tag.dbOid == dbid)
4927 24812 : InvalidateBuffer(bufHdr); /* releases spinlock */
4928 : else
4929 0 : UnlockBufHdr(bufHdr, buf_state);
4930 : }
4931 144 : }
4932 :
4933 : /* ---------------------------------------------------------------------
4934 : * FlushRelationBuffers
4935 : *
4936 : * This function writes all dirty pages of a relation out to disk
4937 : * (or more accurately, out to kernel disk buffers), ensuring that the
4938 : * kernel has an up-to-date view of the relation.
4939 : *
4940 : * Generally, the caller should be holding AccessExclusiveLock on the
4941 : * target relation to ensure that no other backend is busy dirtying
4942 : * more blocks of the relation; the effects can't be expected to last
4943 : * after the lock is released.
4944 : *
4945 : * XXX currently it sequentially searches the buffer pool, should be
4946 : * changed to more clever ways of searching. This routine is not
4947 : * used in any performance-critical code paths, so it's not worth
4948 : * adding additional overhead to normal paths to make it go faster.
4949 : * --------------------------------------------------------------------
4950 : */
4951 : void
4952 276 : FlushRelationBuffers(Relation rel)
4953 : {
4954 : int i;
4955 : BufferDesc *bufHdr;
4956 276 : SMgrRelation srel = RelationGetSmgr(rel);
4957 :
4958 276 : if (RelationUsesLocalBuffers(rel))
4959 : {
4960 1818 : for (i = 0; i < NLocBuffer; i++)
4961 : {
4962 : uint32 buf_state;
4963 :
4964 1800 : bufHdr = GetLocalBufferDescriptor(i);
4965 1800 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
4966 600 : ((buf_state = pg_atomic_read_u32(&bufHdr->state)) &
4967 : (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
4968 : {
4969 : ErrorContextCallback errcallback;
4970 :
4971 : /* Setup error traceback support for ereport() */
4972 600 : errcallback.callback = local_buffer_write_error_callback;
4973 600 : errcallback.arg = bufHdr;
4974 600 : errcallback.previous = error_context_stack;
4975 600 : error_context_stack = &errcallback;
4976 :
4977 : /* Make sure we can handle the pin */
4978 600 : ReservePrivateRefCountEntry();
4979 600 : ResourceOwnerEnlarge(CurrentResourceOwner);
4980 :
4981 : /*
4982 : * Pin/unpin mostly to make valgrind work, but it also seems
4983 : * like the right thing to do.
4984 : */
4985 600 : PinLocalBuffer(bufHdr, false);
4986 :
4987 :
4988 600 : FlushLocalBuffer(bufHdr, srel);
4989 :
4990 600 : UnpinLocalBuffer(BufferDescriptorGetBuffer(bufHdr));
4991 :
4992 : /* Pop the error context stack */
4993 600 : error_context_stack = errcallback.previous;
4994 : }
4995 : }
4996 :
4997 18 : return;
4998 : }
4999 :
5000 3024386 : for (i = 0; i < NBuffers; i++)
5001 : {
5002 : uint32 buf_state;
5003 :
5004 3024128 : bufHdr = GetBufferDescriptor(i);
5005 :
5006 : /*
5007 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5008 : * saves some cycles.
5009 : */
5010 3024128 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator))
5011 3023682 : continue;
5012 :
5013 : /* Make sure we can handle the pin */
5014 446 : ReservePrivateRefCountEntry();
5015 446 : ResourceOwnerEnlarge(CurrentResourceOwner);
5016 :
5017 446 : buf_state = LockBufHdr(bufHdr);
5018 446 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
5019 446 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5020 : {
5021 360 : PinBuffer_Locked(bufHdr);
5022 360 : FlushUnlockedBuffer(bufHdr, srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5023 360 : UnpinBuffer(bufHdr);
5024 : }
5025 : else
5026 86 : UnlockBufHdr(bufHdr, buf_state);
5027 : }
5028 : }
5029 :
5030 : /* ---------------------------------------------------------------------
5031 : * FlushRelationsAllBuffers
5032 : *
5033 : * This function flushes out of the buffer pool all the pages of all
5034 : * forks of the specified smgr relations. It's equivalent to calling
5035 : * FlushRelationBuffers once per relation. The relations are assumed not
5036 : * to use local buffers.
5037 : * --------------------------------------------------------------------
5038 : */
5039 : void
5040 24 : FlushRelationsAllBuffers(SMgrRelation *smgrs, int nrels)
5041 : {
5042 : int i;
5043 : SMgrSortArray *srels;
5044 : bool use_bsearch;
5045 :
5046 24 : if (nrels == 0)
5047 0 : return;
5048 :
5049 : /* fill-in array for qsort */
5050 24 : srels = palloc(sizeof(SMgrSortArray) * nrels);
5051 :
5052 48 : for (i = 0; i < nrels; i++)
5053 : {
5054 : Assert(!RelFileLocatorBackendIsTemp(smgrs[i]->smgr_rlocator));
5055 :
5056 24 : srels[i].rlocator = smgrs[i]->smgr_rlocator.locator;
5057 24 : srels[i].srel = smgrs[i];
5058 : }
5059 :
5060 : /*
5061 : * Save the bsearch overhead for low number of relations to sync. See
5062 : * DropRelationsAllBuffers for details.
5063 : */
5064 24 : use_bsearch = nrels > RELS_BSEARCH_THRESHOLD;
5065 :
5066 : /* sort the list of SMgrRelations if necessary */
5067 24 : if (use_bsearch)
5068 0 : qsort(srels, nrels, sizeof(SMgrSortArray), rlocator_comparator);
5069 :
5070 393240 : for (i = 0; i < NBuffers; i++)
5071 : {
5072 393216 : SMgrSortArray *srelent = NULL;
5073 393216 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5074 : uint32 buf_state;
5075 :
5076 : /*
5077 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5078 : * saves some cycles.
5079 : */
5080 :
5081 393216 : if (!use_bsearch)
5082 : {
5083 : int j;
5084 :
5085 777604 : for (j = 0; j < nrels; j++)
5086 : {
5087 393216 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srels[j].rlocator))
5088 : {
5089 8828 : srelent = &srels[j];
5090 8828 : break;
5091 : }
5092 : }
5093 : }
5094 : else
5095 : {
5096 : RelFileLocator rlocator;
5097 :
5098 0 : rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
5099 0 : srelent = bsearch(&rlocator,
5100 : srels, nrels, sizeof(SMgrSortArray),
5101 : rlocator_comparator);
5102 : }
5103 :
5104 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
5105 393216 : if (srelent == NULL)
5106 384388 : continue;
5107 :
5108 : /* Make sure we can handle the pin */
5109 8828 : ReservePrivateRefCountEntry();
5110 8828 : ResourceOwnerEnlarge(CurrentResourceOwner);
5111 :
5112 8828 : buf_state = LockBufHdr(bufHdr);
5113 8828 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srelent->rlocator) &&
5114 8828 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5115 : {
5116 7914 : PinBuffer_Locked(bufHdr);
5117 7914 : FlushUnlockedBuffer(bufHdr, srelent->srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5118 7914 : UnpinBuffer(bufHdr);
5119 : }
5120 : else
5121 914 : UnlockBufHdr(bufHdr, buf_state);
5122 : }
5123 :
5124 24 : pfree(srels);
5125 : }
5126 :
5127 : /* ---------------------------------------------------------------------
5128 : * RelationCopyStorageUsingBuffer
5129 : *
5130 : * Copy fork's data using bufmgr. Same as RelationCopyStorage but instead
5131 : * of using smgrread and smgrextend this will copy using bufmgr APIs.
5132 : *
5133 : * Refer comments atop CreateAndCopyRelationData() for details about
5134 : * 'permanent' parameter.
5135 : * --------------------------------------------------------------------
5136 : */
5137 : static void
5138 147856 : RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
5139 : RelFileLocator dstlocator,
5140 : ForkNumber forkNum, bool permanent)
5141 : {
5142 : Buffer srcBuf;
5143 : Buffer dstBuf;
5144 : Page srcPage;
5145 : Page dstPage;
5146 : bool use_wal;
5147 : BlockNumber nblocks;
5148 : BlockNumber blkno;
5149 : PGIOAlignedBlock buf;
5150 : BufferAccessStrategy bstrategy_src;
5151 : BufferAccessStrategy bstrategy_dst;
5152 : BlockRangeReadStreamPrivate p;
5153 : ReadStream *src_stream;
5154 : SMgrRelation src_smgr;
5155 :
5156 : /*
5157 : * In general, we want to write WAL whenever wal_level > 'minimal', but we
5158 : * can skip it when copying any fork of an unlogged relation other than
5159 : * the init fork.
5160 : */
5161 147856 : use_wal = XLogIsNeeded() && (permanent || forkNum == INIT_FORKNUM);
5162 :
5163 : /* Get number of blocks in the source relation. */
5164 147856 : nblocks = smgrnblocks(smgropen(srclocator, INVALID_PROC_NUMBER),
5165 : forkNum);
5166 :
5167 : /* Nothing to copy; just return. */
5168 147856 : if (nblocks == 0)
5169 25790 : return;
5170 :
5171 : /*
5172 : * Bulk extend the destination relation of the same size as the source
5173 : * relation before starting to copy block by block.
5174 : */
5175 122066 : memset(buf.data, 0, BLCKSZ);
5176 122066 : smgrextend(smgropen(dstlocator, INVALID_PROC_NUMBER), forkNum, nblocks - 1,
5177 : buf.data, true);
5178 :
5179 : /* This is a bulk operation, so use buffer access strategies. */
5180 122066 : bstrategy_src = GetAccessStrategy(BAS_BULKREAD);
5181 122066 : bstrategy_dst = GetAccessStrategy(BAS_BULKWRITE);
5182 :
5183 : /* Initialize streaming read */
5184 122066 : p.current_blocknum = 0;
5185 122066 : p.last_exclusive = nblocks;
5186 122066 : src_smgr = smgropen(srclocator, INVALID_PROC_NUMBER);
5187 :
5188 : /*
5189 : * It is safe to use batchmode as block_range_read_stream_cb takes no
5190 : * locks.
5191 : */
5192 122066 : src_stream = read_stream_begin_smgr_relation(READ_STREAM_FULL |
5193 : READ_STREAM_USE_BATCHING,
5194 : bstrategy_src,
5195 : src_smgr,
5196 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
5197 : forkNum,
5198 : block_range_read_stream_cb,
5199 : &p,
5200 : 0);
5201 :
5202 : /* Iterate over each block of the source relation file. */
5203 588534 : for (blkno = 0; blkno < nblocks; blkno++)
5204 : {
5205 466472 : CHECK_FOR_INTERRUPTS();
5206 :
5207 : /* Read block from source relation. */
5208 466472 : srcBuf = read_stream_next_buffer(src_stream, NULL);
5209 466468 : LockBuffer(srcBuf, BUFFER_LOCK_SHARE);
5210 466468 : srcPage = BufferGetPage(srcBuf);
5211 :
5212 466468 : dstBuf = ReadBufferWithoutRelcache(dstlocator, forkNum,
5213 : BufferGetBlockNumber(srcBuf),
5214 : RBM_ZERO_AND_LOCK, bstrategy_dst,
5215 : permanent);
5216 466468 : dstPage = BufferGetPage(dstBuf);
5217 :
5218 466468 : START_CRIT_SECTION();
5219 :
5220 : /* Copy page data from the source to the destination. */
5221 466468 : memcpy(dstPage, srcPage, BLCKSZ);
5222 466468 : MarkBufferDirty(dstBuf);
5223 :
5224 : /* WAL-log the copied page. */
5225 466468 : if (use_wal)
5226 257740 : log_newpage_buffer(dstBuf, true);
5227 :
5228 466468 : END_CRIT_SECTION();
5229 :
5230 466468 : UnlockReleaseBuffer(dstBuf);
5231 466468 : UnlockReleaseBuffer(srcBuf);
5232 : }
5233 : Assert(read_stream_next_buffer(src_stream, NULL) == InvalidBuffer);
5234 122062 : read_stream_end(src_stream);
5235 :
5236 122062 : FreeAccessStrategy(bstrategy_src);
5237 122062 : FreeAccessStrategy(bstrategy_dst);
5238 : }
5239 :
5240 : /* ---------------------------------------------------------------------
5241 : * CreateAndCopyRelationData
5242 : *
5243 : * Create destination relation storage and copy all forks from the
5244 : * source relation to the destination.
5245 : *
5246 : * Pass permanent as true for permanent relations and false for
5247 : * unlogged relations. Currently this API is not supported for
5248 : * temporary relations.
5249 : * --------------------------------------------------------------------
5250 : */
5251 : void
5252 111144 : CreateAndCopyRelationData(RelFileLocator src_rlocator,
5253 : RelFileLocator dst_rlocator, bool permanent)
5254 : {
5255 : char relpersistence;
5256 : SMgrRelation src_rel;
5257 : SMgrRelation dst_rel;
5258 :
5259 : /* Set the relpersistence. */
5260 111144 : relpersistence = permanent ?
5261 : RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED;
5262 :
5263 111144 : src_rel = smgropen(src_rlocator, INVALID_PROC_NUMBER);
5264 111144 : dst_rel = smgropen(dst_rlocator, INVALID_PROC_NUMBER);
5265 :
5266 : /*
5267 : * Create and copy all forks of the relation. During create database we
5268 : * have a separate cleanup mechanism which deletes complete database
5269 : * directory. Therefore, each individual relation doesn't need to be
5270 : * registered for cleanup.
5271 : */
5272 111144 : RelationCreateStorage(dst_rlocator, relpersistence, false);
5273 :
5274 : /* copy main fork. */
5275 111144 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, MAIN_FORKNUM,
5276 : permanent);
5277 :
5278 : /* copy those extra forks that exist */
5279 111140 : for (ForkNumber forkNum = MAIN_FORKNUM + 1;
5280 444560 : forkNum <= MAX_FORKNUM; forkNum++)
5281 : {
5282 333420 : if (smgrexists(src_rel, forkNum))
5283 : {
5284 36712 : smgrcreate(dst_rel, forkNum, false);
5285 :
5286 : /*
5287 : * WAL log creation if the relation is persistent, or this is the
5288 : * init fork of an unlogged relation.
5289 : */
5290 36712 : if (permanent || forkNum == INIT_FORKNUM)
5291 36712 : log_smgrcreate(&dst_rlocator, forkNum);
5292 :
5293 : /* Copy a fork's data, block by block. */
5294 36712 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, forkNum,
5295 : permanent);
5296 : }
5297 : }
5298 111140 : }
5299 :
5300 : /* ---------------------------------------------------------------------
5301 : * FlushDatabaseBuffers
5302 : *
5303 : * This function writes all dirty pages of a database out to disk
5304 : * (or more accurately, out to kernel disk buffers), ensuring that the
5305 : * kernel has an up-to-date view of the database.
5306 : *
5307 : * Generally, the caller should be holding an appropriate lock to ensure
5308 : * no other backend is active in the target database; otherwise more
5309 : * pages could get dirtied.
5310 : *
5311 : * Note we don't worry about flushing any pages of temporary relations.
5312 : * It's assumed these wouldn't be interesting.
5313 : * --------------------------------------------------------------------
5314 : */
5315 : void
5316 10 : FlushDatabaseBuffers(Oid dbid)
5317 : {
5318 : int i;
5319 : BufferDesc *bufHdr;
5320 :
5321 1290 : for (i = 0; i < NBuffers; i++)
5322 : {
5323 : uint32 buf_state;
5324 :
5325 1280 : bufHdr = GetBufferDescriptor(i);
5326 :
5327 : /*
5328 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5329 : * saves some cycles.
5330 : */
5331 1280 : if (bufHdr->tag.dbOid != dbid)
5332 1018 : continue;
5333 :
5334 : /* Make sure we can handle the pin */
5335 262 : ReservePrivateRefCountEntry();
5336 262 : ResourceOwnerEnlarge(CurrentResourceOwner);
5337 :
5338 262 : buf_state = LockBufHdr(bufHdr);
5339 262 : if (bufHdr->tag.dbOid == dbid &&
5340 262 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5341 : {
5342 20 : PinBuffer_Locked(bufHdr);
5343 20 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5344 20 : UnpinBuffer(bufHdr);
5345 : }
5346 : else
5347 242 : UnlockBufHdr(bufHdr, buf_state);
5348 : }
5349 10 : }
5350 :
5351 : /*
5352 : * Flush a previously, shared or exclusively, locked and pinned buffer to the
5353 : * OS.
5354 : */
5355 : void
5356 158 : FlushOneBuffer(Buffer buffer)
5357 : {
5358 : BufferDesc *bufHdr;
5359 :
5360 : /* currently not needed, but no fundamental reason not to support */
5361 : Assert(!BufferIsLocal(buffer));
5362 :
5363 : Assert(BufferIsPinned(buffer));
5364 :
5365 158 : bufHdr = GetBufferDescriptor(buffer - 1);
5366 :
5367 : Assert(BufferIsLockedByMe(buffer));
5368 :
5369 158 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5370 158 : }
5371 :
5372 : /*
5373 : * ReleaseBuffer -- release the pin on a buffer
5374 : */
5375 : void
5376 125030106 : ReleaseBuffer(Buffer buffer)
5377 : {
5378 125030106 : if (!BufferIsValid(buffer))
5379 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5380 :
5381 125030106 : if (BufferIsLocal(buffer))
5382 3220200 : UnpinLocalBuffer(buffer);
5383 : else
5384 121809906 : UnpinBuffer(GetBufferDescriptor(buffer - 1));
5385 125030106 : }
5386 :
5387 : /*
5388 : * UnlockReleaseBuffer -- release the content lock and pin on a buffer
5389 : *
5390 : * This is just a shorthand for a common combination.
5391 : */
5392 : void
5393 37928024 : UnlockReleaseBuffer(Buffer buffer)
5394 : {
5395 37928024 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5396 37928024 : ReleaseBuffer(buffer);
5397 37928024 : }
5398 :
5399 : /*
5400 : * IncrBufferRefCount
5401 : * Increment the pin count on a buffer that we have *already* pinned
5402 : * at least once.
5403 : *
5404 : * This function cannot be used on a buffer we do not have pinned,
5405 : * because it doesn't change the shared buffer state.
5406 : */
5407 : void
5408 23306954 : IncrBufferRefCount(Buffer buffer)
5409 : {
5410 : Assert(BufferIsPinned(buffer));
5411 23306954 : ResourceOwnerEnlarge(CurrentResourceOwner);
5412 23306954 : if (BufferIsLocal(buffer))
5413 709366 : LocalRefCount[-buffer - 1]++;
5414 : else
5415 : {
5416 : PrivateRefCountEntry *ref;
5417 :
5418 22597588 : ref = GetPrivateRefCountEntry(buffer, true);
5419 : Assert(ref != NULL);
5420 22597588 : ref->refcount++;
5421 : }
5422 23306954 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buffer);
5423 23306954 : }
5424 :
5425 : /*
5426 : * MarkBufferDirtyHint
5427 : *
5428 : * Mark a buffer dirty for non-critical changes.
5429 : *
5430 : * This is essentially the same as MarkBufferDirty, except:
5431 : *
5432 : * 1. The caller does not write WAL; so if checksums are enabled, we may need
5433 : * to write an XLOG_FPI_FOR_HINT WAL record to protect against torn pages.
5434 : * 2. The caller might have only share-lock instead of exclusive-lock on the
5435 : * buffer's content lock.
5436 : * 3. This function does not guarantee that the buffer is always marked dirty
5437 : * (due to a race condition), so it cannot be used for important changes.
5438 : */
5439 : void
5440 20060118 : MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
5441 : {
5442 : BufferDesc *bufHdr;
5443 20060118 : Page page = BufferGetPage(buffer);
5444 :
5445 20060118 : if (!BufferIsValid(buffer))
5446 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5447 :
5448 20060118 : if (BufferIsLocal(buffer))
5449 : {
5450 1270290 : MarkLocalBufferDirty(buffer);
5451 1270290 : return;
5452 : }
5453 :
5454 18789828 : bufHdr = GetBufferDescriptor(buffer - 1);
5455 :
5456 : Assert(GetPrivateRefCount(buffer) > 0);
5457 : /* here, either share or exclusive lock is OK */
5458 : Assert(BufferIsLockedByMe(buffer));
5459 :
5460 : /*
5461 : * This routine might get called many times on the same page, if we are
5462 : * making the first scan after commit of an xact that added/deleted many
5463 : * tuples. So, be as quick as we can if the buffer is already dirty. We
5464 : * do this by not acquiring spinlock if it looks like the status bits are
5465 : * already set. Since we make this test unlocked, there's a chance we
5466 : * might fail to notice that the flags have just been cleared, and failed
5467 : * to reset them, due to memory-ordering issues. But since this function
5468 : * is only intended to be used in cases where failing to write out the
5469 : * data would be harmless anyway, it doesn't really matter.
5470 : */
5471 18789828 : if ((pg_atomic_read_u32(&bufHdr->state) & (BM_DIRTY | BM_JUST_DIRTIED)) !=
5472 : (BM_DIRTY | BM_JUST_DIRTIED))
5473 : {
5474 1867938 : XLogRecPtr lsn = InvalidXLogRecPtr;
5475 1867938 : bool dirtied = false;
5476 1867938 : bool delayChkptFlags = false;
5477 : uint32 buf_state;
5478 :
5479 : /*
5480 : * If we need to protect hint bit updates from torn writes, WAL-log a
5481 : * full page image of the page. This full page image is only necessary
5482 : * if the hint bit update is the first change to the page since the
5483 : * last checkpoint.
5484 : *
5485 : * We don't check full_page_writes here because that logic is included
5486 : * when we call XLogInsert() since the value changes dynamically.
5487 : */
5488 3733688 : if (XLogHintBitIsNeeded() &&
5489 1865750 : (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT))
5490 : {
5491 : /*
5492 : * If we must not write WAL, due to a relfilelocator-specific
5493 : * condition or being in recovery, don't dirty the page. We can
5494 : * set the hint, just not dirty the page as a result so the hint
5495 : * is lost when we evict the page or shutdown.
5496 : *
5497 : * See src/backend/storage/page/README for longer discussion.
5498 : */
5499 1987600 : if (RecoveryInProgress() ||
5500 121914 : RelFileLocatorSkippingWAL(BufTagGetRelFileLocator(&bufHdr->tag)))
5501 1749458 : return;
5502 :
5503 : /*
5504 : * If the block is already dirty because we either made a change
5505 : * or set a hint already, then we don't need to write a full page
5506 : * image. Note that aggressive cleaning of blocks dirtied by hint
5507 : * bit setting would increase the call rate. Bulk setting of hint
5508 : * bits would reduce the call rate...
5509 : *
5510 : * We must issue the WAL record before we mark the buffer dirty.
5511 : * Otherwise we might write the page before we write the WAL. That
5512 : * causes a race condition, since a checkpoint might occur between
5513 : * writing the WAL record and marking the buffer dirty. We solve
5514 : * that with a kluge, but one that is already in use during
5515 : * transaction commit to prevent race conditions. Basically, we
5516 : * simply prevent the checkpoint WAL record from being written
5517 : * until we have marked the buffer dirty. We don't start the
5518 : * checkpoint flush until we have marked dirty, so our checkpoint
5519 : * must flush the change to disk successfully or the checkpoint
5520 : * never gets written, so crash recovery will fix.
5521 : *
5522 : * It's possible we may enter here without an xid, so it is
5523 : * essential that CreateCheckPoint waits for virtual transactions
5524 : * rather than full transactionids.
5525 : */
5526 : Assert((MyProc->delayChkptFlags & DELAY_CHKPT_START) == 0);
5527 116228 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
5528 116228 : delayChkptFlags = true;
5529 116228 : lsn = XLogSaveBufferForHint(buffer, buffer_std);
5530 : }
5531 :
5532 118480 : buf_state = LockBufHdr(bufHdr);
5533 :
5534 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5535 :
5536 118480 : if (!(buf_state & BM_DIRTY))
5537 : {
5538 118394 : dirtied = true; /* Means "will be dirtied by this action" */
5539 :
5540 : /*
5541 : * Set the page LSN if we wrote a backup block. We aren't supposed
5542 : * to set this when only holding a share lock but as long as we
5543 : * serialise it somehow we're OK. We choose to set LSN while
5544 : * holding the buffer header lock, which causes any reader of an
5545 : * LSN who holds only a share lock to also obtain a buffer header
5546 : * lock before using PageGetLSN(), which is enforced in
5547 : * BufferGetLSNAtomic().
5548 : *
5549 : * If checksums are enabled, you might think we should reset the
5550 : * checksum here. That will happen when the page is written
5551 : * sometime later in this checkpoint cycle.
5552 : */
5553 118394 : if (!XLogRecPtrIsInvalid(lsn))
5554 60972 : PageSetLSN(page, lsn);
5555 : }
5556 :
5557 118480 : buf_state |= BM_DIRTY | BM_JUST_DIRTIED;
5558 118480 : UnlockBufHdr(bufHdr, buf_state);
5559 :
5560 118480 : if (delayChkptFlags)
5561 116228 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
5562 :
5563 118480 : if (dirtied)
5564 : {
5565 118394 : pgBufferUsage.shared_blks_dirtied++;
5566 118394 : if (VacuumCostActive)
5567 2168 : VacuumCostBalance += VacuumCostPageDirty;
5568 : }
5569 : }
5570 : }
5571 :
5572 : /*
5573 : * Release buffer content locks for shared buffers.
5574 : *
5575 : * Used to clean up after errors.
5576 : *
5577 : * Currently, we can expect that lwlock.c's LWLockReleaseAll() took care
5578 : * of releasing buffer content locks per se; the only thing we need to deal
5579 : * with here is clearing any PIN_COUNT request that was in progress.
5580 : */
5581 : void
5582 103964 : UnlockBuffers(void)
5583 : {
5584 103964 : BufferDesc *buf = PinCountWaitBuf;
5585 :
5586 103964 : if (buf)
5587 : {
5588 : uint32 buf_state;
5589 :
5590 0 : buf_state = LockBufHdr(buf);
5591 :
5592 : /*
5593 : * Don't complain if flag bit not set; it could have been reset but we
5594 : * got a cancel/die interrupt before getting the signal.
5595 : */
5596 0 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5597 0 : buf->wait_backend_pgprocno == MyProcNumber)
5598 0 : buf_state &= ~BM_PIN_COUNT_WAITER;
5599 :
5600 0 : UnlockBufHdr(buf, buf_state);
5601 :
5602 0 : PinCountWaitBuf = NULL;
5603 : }
5604 103964 : }
5605 :
5606 : /*
5607 : * Acquire or release the content_lock for the buffer.
5608 : */
5609 : void
5610 346027842 : LockBuffer(Buffer buffer, int mode)
5611 : {
5612 : BufferDesc *buf;
5613 :
5614 : Assert(BufferIsPinned(buffer));
5615 346027842 : if (BufferIsLocal(buffer))
5616 19800110 : return; /* local buffers need no lock */
5617 :
5618 326227732 : buf = GetBufferDescriptor(buffer - 1);
5619 :
5620 326227732 : if (mode == BUFFER_LOCK_UNLOCK)
5621 164572406 : LWLockRelease(BufferDescriptorGetContentLock(buf));
5622 161655326 : else if (mode == BUFFER_LOCK_SHARE)
5623 114278302 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_SHARED);
5624 47377024 : else if (mode == BUFFER_LOCK_EXCLUSIVE)
5625 47377024 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_EXCLUSIVE);
5626 : else
5627 0 : elog(ERROR, "unrecognized buffer lock mode: %d", mode);
5628 : }
5629 :
5630 : /*
5631 : * Acquire the content_lock for the buffer, but only if we don't have to wait.
5632 : *
5633 : * This assumes the caller wants BUFFER_LOCK_EXCLUSIVE mode.
5634 : */
5635 : bool
5636 3047648 : ConditionalLockBuffer(Buffer buffer)
5637 : {
5638 : BufferDesc *buf;
5639 :
5640 : Assert(BufferIsPinned(buffer));
5641 3047648 : if (BufferIsLocal(buffer))
5642 129264 : return true; /* act as though we got it */
5643 :
5644 2918384 : buf = GetBufferDescriptor(buffer - 1);
5645 :
5646 2918384 : return LWLockConditionalAcquire(BufferDescriptorGetContentLock(buf),
5647 : LW_EXCLUSIVE);
5648 : }
5649 :
5650 : /*
5651 : * Verify that this backend is pinning the buffer exactly once.
5652 : *
5653 : * NOTE: Like in BufferIsPinned(), what we check here is that *this* backend
5654 : * holds a pin on the buffer. We do not care whether some other backend does.
5655 : */
5656 : void
5657 4964516 : CheckBufferIsPinnedOnce(Buffer buffer)
5658 : {
5659 4964516 : if (BufferIsLocal(buffer))
5660 : {
5661 1578 : if (LocalRefCount[-buffer - 1] != 1)
5662 0 : elog(ERROR, "incorrect local pin count: %d",
5663 : LocalRefCount[-buffer - 1]);
5664 : }
5665 : else
5666 : {
5667 4962938 : if (GetPrivateRefCount(buffer) != 1)
5668 0 : elog(ERROR, "incorrect local pin count: %d",
5669 : GetPrivateRefCount(buffer));
5670 : }
5671 4964516 : }
5672 :
5673 : /*
5674 : * LockBufferForCleanup - lock a buffer in preparation for deleting items
5675 : *
5676 : * Items may be deleted from a disk page only when the caller (a) holds an
5677 : * exclusive lock on the buffer and (b) has observed that no other backend
5678 : * holds a pin on the buffer. If there is a pin, then the other backend
5679 : * might have a pointer into the buffer (for example, a heapscan reference
5680 : * to an item --- see README for more details). It's OK if a pin is added
5681 : * after the cleanup starts, however; the newly-arrived backend will be
5682 : * unable to look at the page until we release the exclusive lock.
5683 : *
5684 : * To implement this protocol, a would-be deleter must pin the buffer and
5685 : * then call LockBufferForCleanup(). LockBufferForCleanup() is similar to
5686 : * LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE), except that it loops until
5687 : * it has successfully observed pin count = 1.
5688 : */
5689 : void
5690 41610 : LockBufferForCleanup(Buffer buffer)
5691 : {
5692 : BufferDesc *bufHdr;
5693 41610 : TimestampTz waitStart = 0;
5694 41610 : bool waiting = false;
5695 41610 : bool logged_recovery_conflict = false;
5696 :
5697 : Assert(BufferIsPinned(buffer));
5698 : Assert(PinCountWaitBuf == NULL);
5699 :
5700 41610 : CheckBufferIsPinnedOnce(buffer);
5701 :
5702 : /*
5703 : * We do not yet need to be worried about in-progress AIOs holding a pin,
5704 : * as we, so far, only support doing reads via AIO and this function can
5705 : * only be called once the buffer is valid (i.e. no read can be in
5706 : * flight).
5707 : */
5708 :
5709 : /* Nobody else to wait for */
5710 41610 : if (BufferIsLocal(buffer))
5711 32 : return;
5712 :
5713 41578 : bufHdr = GetBufferDescriptor(buffer - 1);
5714 :
5715 : for (;;)
5716 166 : {
5717 : uint32 buf_state;
5718 :
5719 : /* Try to acquire lock */
5720 41744 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
5721 41744 : buf_state = LockBufHdr(bufHdr);
5722 :
5723 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5724 41744 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
5725 : {
5726 : /* Successfully acquired exclusive lock with pincount 1 */
5727 41578 : UnlockBufHdr(bufHdr, buf_state);
5728 :
5729 : /*
5730 : * Emit the log message if recovery conflict on buffer pin was
5731 : * resolved but the startup process waited longer than
5732 : * deadlock_timeout for it.
5733 : */
5734 41578 : if (logged_recovery_conflict)
5735 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
5736 : waitStart, GetCurrentTimestamp(),
5737 : NULL, false);
5738 :
5739 41578 : if (waiting)
5740 : {
5741 : /* reset ps display to remove the suffix if we added one */
5742 4 : set_ps_display_remove_suffix();
5743 4 : waiting = false;
5744 : }
5745 41578 : return;
5746 : }
5747 : /* Failed, so mark myself as waiting for pincount 1 */
5748 166 : if (buf_state & BM_PIN_COUNT_WAITER)
5749 : {
5750 0 : UnlockBufHdr(bufHdr, buf_state);
5751 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5752 0 : elog(ERROR, "multiple backends attempting to wait for pincount 1");
5753 : }
5754 166 : bufHdr->wait_backend_pgprocno = MyProcNumber;
5755 166 : PinCountWaitBuf = bufHdr;
5756 166 : buf_state |= BM_PIN_COUNT_WAITER;
5757 166 : UnlockBufHdr(bufHdr, buf_state);
5758 166 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5759 :
5760 : /* Wait to be signaled by UnpinBuffer() */
5761 166 : if (InHotStandby)
5762 : {
5763 20 : if (!waiting)
5764 : {
5765 : /* adjust the process title to indicate that it's waiting */
5766 4 : set_ps_display_suffix("waiting");
5767 4 : waiting = true;
5768 : }
5769 :
5770 : /*
5771 : * Emit the log message if the startup process is waiting longer
5772 : * than deadlock_timeout for recovery conflict on buffer pin.
5773 : *
5774 : * Skip this if first time through because the startup process has
5775 : * not started waiting yet in this case. So, the wait start
5776 : * timestamp is set after this logic.
5777 : */
5778 20 : if (waitStart != 0 && !logged_recovery_conflict)
5779 : {
5780 6 : TimestampTz now = GetCurrentTimestamp();
5781 :
5782 6 : if (TimestampDifferenceExceeds(waitStart, now,
5783 : DeadlockTimeout))
5784 : {
5785 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
5786 : waitStart, now, NULL, true);
5787 4 : logged_recovery_conflict = true;
5788 : }
5789 : }
5790 :
5791 : /*
5792 : * Set the wait start timestamp if logging is enabled and first
5793 : * time through.
5794 : */
5795 20 : if (log_recovery_conflict_waits && waitStart == 0)
5796 4 : waitStart = GetCurrentTimestamp();
5797 :
5798 : /* Publish the bufid that Startup process waits on */
5799 20 : SetStartupBufferPinWaitBufId(buffer - 1);
5800 : /* Set alarm and then wait to be signaled by UnpinBuffer() */
5801 20 : ResolveRecoveryConflictWithBufferPin();
5802 : /* Reset the published bufid */
5803 20 : SetStartupBufferPinWaitBufId(-1);
5804 : }
5805 : else
5806 146 : ProcWaitForSignal(WAIT_EVENT_BUFFER_PIN);
5807 :
5808 : /*
5809 : * Remove flag marking us as waiter. Normally this will not be set
5810 : * anymore, but ProcWaitForSignal() can return for other signals as
5811 : * well. We take care to only reset the flag if we're the waiter, as
5812 : * theoretically another backend could have started waiting. That's
5813 : * impossible with the current usages due to table level locking, but
5814 : * better be safe.
5815 : */
5816 166 : buf_state = LockBufHdr(bufHdr);
5817 166 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5818 16 : bufHdr->wait_backend_pgprocno == MyProcNumber)
5819 16 : buf_state &= ~BM_PIN_COUNT_WAITER;
5820 166 : UnlockBufHdr(bufHdr, buf_state);
5821 :
5822 166 : PinCountWaitBuf = NULL;
5823 : /* Loop back and try again */
5824 : }
5825 : }
5826 :
5827 : /*
5828 : * Check called from ProcessRecoveryConflictInterrupts() when Startup process
5829 : * requests cancellation of all pin holders that are blocking it.
5830 : */
5831 : bool
5832 8 : HoldingBufferPinThatDelaysRecovery(void)
5833 : {
5834 8 : int bufid = GetStartupBufferPinWaitBufId();
5835 :
5836 : /*
5837 : * If we get woken slowly then it's possible that the Startup process was
5838 : * already woken by other backends before we got here. Also possible that
5839 : * we get here by multiple interrupts or interrupts at inappropriate
5840 : * times, so make sure we do nothing if the bufid is not set.
5841 : */
5842 8 : if (bufid < 0)
5843 4 : return false;
5844 :
5845 4 : if (GetPrivateRefCount(bufid + 1) > 0)
5846 4 : return true;
5847 :
5848 0 : return false;
5849 : }
5850 :
5851 : /*
5852 : * ConditionalLockBufferForCleanup - as above, but don't wait to get the lock
5853 : *
5854 : * We won't loop, but just check once to see if the pin count is OK. If
5855 : * not, return false with no lock held.
5856 : */
5857 : bool
5858 1214402 : ConditionalLockBufferForCleanup(Buffer buffer)
5859 : {
5860 : BufferDesc *bufHdr;
5861 : uint32 buf_state,
5862 : refcount;
5863 :
5864 : Assert(BufferIsValid(buffer));
5865 :
5866 : /* see AIO related comment in LockBufferForCleanup() */
5867 :
5868 1214402 : if (BufferIsLocal(buffer))
5869 : {
5870 1608 : refcount = LocalRefCount[-buffer - 1];
5871 : /* There should be exactly one pin */
5872 : Assert(refcount > 0);
5873 1608 : if (refcount != 1)
5874 42 : return false;
5875 : /* Nobody else to wait for */
5876 1566 : return true;
5877 : }
5878 :
5879 : /* There should be exactly one local pin */
5880 1212794 : refcount = GetPrivateRefCount(buffer);
5881 : Assert(refcount);
5882 1212794 : if (refcount != 1)
5883 524 : return false;
5884 :
5885 : /* Try to acquire lock */
5886 1212270 : if (!ConditionalLockBuffer(buffer))
5887 110 : return false;
5888 :
5889 1212160 : bufHdr = GetBufferDescriptor(buffer - 1);
5890 1212160 : buf_state = LockBufHdr(bufHdr);
5891 1212160 : refcount = BUF_STATE_GET_REFCOUNT(buf_state);
5892 :
5893 : Assert(refcount > 0);
5894 1212160 : if (refcount == 1)
5895 : {
5896 : /* Successfully acquired exclusive lock with pincount 1 */
5897 1211434 : UnlockBufHdr(bufHdr, buf_state);
5898 1211434 : return true;
5899 : }
5900 :
5901 : /* Failed, so release the lock */
5902 726 : UnlockBufHdr(bufHdr, buf_state);
5903 726 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5904 726 : return false;
5905 : }
5906 :
5907 : /*
5908 : * IsBufferCleanupOK - as above, but we already have the lock
5909 : *
5910 : * Check whether it's OK to perform cleanup on a buffer we've already
5911 : * locked. If we observe that the pin count is 1, our exclusive lock
5912 : * happens to be a cleanup lock, and we can proceed with anything that
5913 : * would have been allowable had we sought a cleanup lock originally.
5914 : */
5915 : bool
5916 4060 : IsBufferCleanupOK(Buffer buffer)
5917 : {
5918 : BufferDesc *bufHdr;
5919 : uint32 buf_state;
5920 :
5921 : Assert(BufferIsValid(buffer));
5922 :
5923 : /* see AIO related comment in LockBufferForCleanup() */
5924 :
5925 4060 : if (BufferIsLocal(buffer))
5926 : {
5927 : /* There should be exactly one pin */
5928 0 : if (LocalRefCount[-buffer - 1] != 1)
5929 0 : return false;
5930 : /* Nobody else to wait for */
5931 0 : return true;
5932 : }
5933 :
5934 : /* There should be exactly one local pin */
5935 4060 : if (GetPrivateRefCount(buffer) != 1)
5936 0 : return false;
5937 :
5938 4060 : bufHdr = GetBufferDescriptor(buffer - 1);
5939 :
5940 : /* caller must hold exclusive lock on buffer */
5941 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
5942 :
5943 4060 : buf_state = LockBufHdr(bufHdr);
5944 :
5945 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5946 4060 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
5947 : {
5948 : /* pincount is OK. */
5949 4060 : UnlockBufHdr(bufHdr, buf_state);
5950 4060 : return true;
5951 : }
5952 :
5953 0 : UnlockBufHdr(bufHdr, buf_state);
5954 0 : return false;
5955 : }
5956 :
5957 :
5958 : /*
5959 : * Functions for buffer I/O handling
5960 : *
5961 : * Also note that these are used only for shared buffers, not local ones.
5962 : */
5963 :
5964 : /*
5965 : * WaitIO -- Block until the IO_IN_PROGRESS flag on 'buf' is cleared.
5966 : */
5967 : static void
5968 14522 : WaitIO(BufferDesc *buf)
5969 : {
5970 14522 : ConditionVariable *cv = BufferDescriptorGetIOCV(buf);
5971 :
5972 14522 : ConditionVariablePrepareToSleep(cv);
5973 : for (;;)
5974 14422 : {
5975 : uint32 buf_state;
5976 : PgAioWaitRef iow;
5977 :
5978 : /*
5979 : * It may not be necessary to acquire the spinlock to check the flag
5980 : * here, but since this test is essential for correctness, we'd better
5981 : * play it safe.
5982 : */
5983 28944 : buf_state = LockBufHdr(buf);
5984 :
5985 : /*
5986 : * Copy the wait reference while holding the spinlock. This protects
5987 : * against a concurrent TerminateBufferIO() in another backend from
5988 : * clearing the wref while it's being read.
5989 : */
5990 28944 : iow = buf->io_wref;
5991 28944 : UnlockBufHdr(buf, buf_state);
5992 :
5993 : /* no IO in progress, we don't need to wait */
5994 28944 : if (!(buf_state & BM_IO_IN_PROGRESS))
5995 14522 : break;
5996 :
5997 : /*
5998 : * The buffer has asynchronous IO in progress, wait for it to
5999 : * complete.
6000 : */
6001 14422 : if (pgaio_wref_valid(&iow))
6002 : {
6003 12032 : pgaio_wref_wait(&iow);
6004 :
6005 : /*
6006 : * The AIO subsystem internally uses condition variables and thus
6007 : * might remove this backend from the BufferDesc's CV. While that
6008 : * wouldn't cause a correctness issue (the first CV sleep just
6009 : * immediately returns if not already registered), it seems worth
6010 : * avoiding unnecessary loop iterations, given that we take care
6011 : * to do so at the start of the function.
6012 : */
6013 12032 : ConditionVariablePrepareToSleep(cv);
6014 12032 : continue;
6015 : }
6016 :
6017 : /* wait on BufferDesc->cv, e.g. for concurrent synchronous IO */
6018 2390 : ConditionVariableSleep(cv, WAIT_EVENT_BUFFER_IO);
6019 : }
6020 14522 : ConditionVariableCancelSleep();
6021 14522 : }
6022 :
6023 : /*
6024 : * StartBufferIO: begin I/O on this buffer
6025 : * (Assumptions)
6026 : * My process is executing no IO on this buffer
6027 : * The buffer is Pinned
6028 : *
6029 : * In some scenarios multiple backends could attempt the same I/O operation
6030 : * concurrently. If someone else has already started I/O on this buffer then
6031 : * we will wait for completion of the IO using WaitIO().
6032 : *
6033 : * Input operations are only attempted on buffers that are not BM_VALID,
6034 : * and output operations only on buffers that are BM_VALID and BM_DIRTY,
6035 : * so we can always tell if the work is already done.
6036 : *
6037 : * Returns true if we successfully marked the buffer as I/O busy,
6038 : * false if someone else already did the work.
6039 : *
6040 : * If nowait is true, then we don't wait for an I/O to be finished by another
6041 : * backend. In that case, false indicates either that the I/O was already
6042 : * finished, or is still in progress. This is useful for callers that want to
6043 : * find out if they can perform the I/O as part of a larger operation, without
6044 : * waiting for the answer or distinguishing the reasons why not.
6045 : */
6046 : bool
6047 4896572 : StartBufferIO(BufferDesc *buf, bool forInput, bool nowait)
6048 : {
6049 : uint32 buf_state;
6050 :
6051 4896572 : ResourceOwnerEnlarge(CurrentResourceOwner);
6052 :
6053 : for (;;)
6054 : {
6055 4911092 : buf_state = LockBufHdr(buf);
6056 :
6057 4911092 : if (!(buf_state & BM_IO_IN_PROGRESS))
6058 4896564 : break;
6059 14528 : UnlockBufHdr(buf, buf_state);
6060 14528 : if (nowait)
6061 8 : return false;
6062 14520 : WaitIO(buf);
6063 : }
6064 :
6065 : /* Once we get here, there is definitely no I/O active on this buffer */
6066 :
6067 : /* Check if someone else already did the I/O */
6068 4896564 : if (forInput ? (buf_state & BM_VALID) : !(buf_state & BM_DIRTY))
6069 : {
6070 15098 : UnlockBufHdr(buf, buf_state);
6071 15098 : return false;
6072 : }
6073 :
6074 4881466 : buf_state |= BM_IO_IN_PROGRESS;
6075 4881466 : UnlockBufHdr(buf, buf_state);
6076 :
6077 4881466 : ResourceOwnerRememberBufferIO(CurrentResourceOwner,
6078 : BufferDescriptorGetBuffer(buf));
6079 :
6080 4881466 : return true;
6081 : }
6082 :
6083 : /*
6084 : * TerminateBufferIO: release a buffer we were doing I/O on
6085 : * (Assumptions)
6086 : * My process is executing IO for the buffer
6087 : * BM_IO_IN_PROGRESS bit is set for the buffer
6088 : * The buffer is Pinned
6089 : *
6090 : * If clear_dirty is true and BM_JUST_DIRTIED is not set, we clear the
6091 : * buffer's BM_DIRTY flag. This is appropriate when terminating a
6092 : * successful write. The check on BM_JUST_DIRTIED is necessary to avoid
6093 : * marking the buffer clean if it was re-dirtied while we were writing.
6094 : *
6095 : * set_flag_bits gets ORed into the buffer's flags. It must include
6096 : * BM_IO_ERROR in a failure case. For successful completion it could
6097 : * be 0, or BM_VALID if we just finished reading in the page.
6098 : *
6099 : * If forget_owner is true, we release the buffer I/O from the current
6100 : * resource owner. (forget_owner=false is used when the resource owner itself
6101 : * is being released)
6102 : */
6103 : void
6104 4607164 : TerminateBufferIO(BufferDesc *buf, bool clear_dirty, uint32 set_flag_bits,
6105 : bool forget_owner, bool release_aio)
6106 : {
6107 : uint32 buf_state;
6108 :
6109 4607164 : buf_state = LockBufHdr(buf);
6110 :
6111 : Assert(buf_state & BM_IO_IN_PROGRESS);
6112 4607164 : buf_state &= ~BM_IO_IN_PROGRESS;
6113 :
6114 : /* Clear earlier errors, if this IO failed, it'll be marked again */
6115 4607164 : buf_state &= ~BM_IO_ERROR;
6116 :
6117 4607164 : if (clear_dirty && !(buf_state & BM_JUST_DIRTIED))
6118 1102606 : buf_state &= ~(BM_DIRTY | BM_CHECKPOINT_NEEDED);
6119 :
6120 4607164 : if (release_aio)
6121 : {
6122 : /* release ownership by the AIO subsystem */
6123 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6124 2503276 : buf_state -= BUF_REFCOUNT_ONE;
6125 2503276 : pgaio_wref_clear(&buf->io_wref);
6126 : }
6127 :
6128 4607164 : buf_state |= set_flag_bits;
6129 4607164 : UnlockBufHdr(buf, buf_state);
6130 :
6131 4607164 : if (forget_owner)
6132 2103846 : ResourceOwnerForgetBufferIO(CurrentResourceOwner,
6133 : BufferDescriptorGetBuffer(buf));
6134 :
6135 4607164 : ConditionVariableBroadcast(BufferDescriptorGetIOCV(buf));
6136 :
6137 : /*
6138 : * Support LockBufferForCleanup()
6139 : *
6140 : * We may have just released the last pin other than the waiter's. In most
6141 : * cases, this backend holds another pin on the buffer. But, if, for
6142 : * example, this backend is completing an IO issued by another backend, it
6143 : * may be time to wake the waiter.
6144 : */
6145 4607164 : if (release_aio && (buf_state & BM_PIN_COUNT_WAITER))
6146 0 : WakePinCountWaiter(buf);
6147 4607164 : }
6148 :
6149 : /*
6150 : * AbortBufferIO: Clean up active buffer I/O after an error.
6151 : *
6152 : * All LWLocks we might have held have been released,
6153 : * but we haven't yet released buffer pins, so the buffer is still pinned.
6154 : *
6155 : * If I/O was in progress, we always set BM_IO_ERROR, even though it's
6156 : * possible the error condition wasn't related to the I/O.
6157 : *
6158 : * Note: this does not remove the buffer I/O from the resource owner.
6159 : * That's correct when we're releasing the whole resource owner, but
6160 : * beware if you use this in other contexts.
6161 : */
6162 : static void
6163 30 : AbortBufferIO(Buffer buffer)
6164 : {
6165 30 : BufferDesc *buf_hdr = GetBufferDescriptor(buffer - 1);
6166 : uint32 buf_state;
6167 :
6168 30 : buf_state = LockBufHdr(buf_hdr);
6169 : Assert(buf_state & (BM_IO_IN_PROGRESS | BM_TAG_VALID));
6170 :
6171 30 : if (!(buf_state & BM_VALID))
6172 : {
6173 : Assert(!(buf_state & BM_DIRTY));
6174 30 : UnlockBufHdr(buf_hdr, buf_state);
6175 : }
6176 : else
6177 : {
6178 : Assert(buf_state & BM_DIRTY);
6179 0 : UnlockBufHdr(buf_hdr, buf_state);
6180 :
6181 : /* Issue notice if this is not the first failure... */
6182 0 : if (buf_state & BM_IO_ERROR)
6183 : {
6184 : /* Buffer is pinned, so we can read tag without spinlock */
6185 0 : ereport(WARNING,
6186 : (errcode(ERRCODE_IO_ERROR),
6187 : errmsg("could not write block %u of %s",
6188 : buf_hdr->tag.blockNum,
6189 : relpathperm(BufTagGetRelFileLocator(&buf_hdr->tag),
6190 : BufTagGetForkNum(&buf_hdr->tag)).str),
6191 : errdetail("Multiple failures --- write error might be permanent.")));
6192 : }
6193 : }
6194 :
6195 30 : TerminateBufferIO(buf_hdr, false, BM_IO_ERROR, false, false);
6196 30 : }
6197 :
6198 : /*
6199 : * Error context callback for errors occurring during shared buffer writes.
6200 : */
6201 : static void
6202 70 : shared_buffer_write_error_callback(void *arg)
6203 : {
6204 70 : BufferDesc *bufHdr = (BufferDesc *) arg;
6205 :
6206 : /* Buffer is pinned, so we can read the tag without locking the spinlock */
6207 70 : if (bufHdr != NULL)
6208 140 : errcontext("writing block %u of relation \"%s\"",
6209 : bufHdr->tag.blockNum,
6210 70 : relpathperm(BufTagGetRelFileLocator(&bufHdr->tag),
6211 : BufTagGetForkNum(&bufHdr->tag)).str);
6212 70 : }
6213 :
6214 : /*
6215 : * Error context callback for errors occurring during local buffer writes.
6216 : */
6217 : static void
6218 0 : local_buffer_write_error_callback(void *arg)
6219 : {
6220 0 : BufferDesc *bufHdr = (BufferDesc *) arg;
6221 :
6222 0 : if (bufHdr != NULL)
6223 0 : errcontext("writing block %u of relation \"%s\"",
6224 : bufHdr->tag.blockNum,
6225 0 : relpathbackend(BufTagGetRelFileLocator(&bufHdr->tag),
6226 : MyProcNumber,
6227 : BufTagGetForkNum(&bufHdr->tag)).str);
6228 0 : }
6229 :
6230 : /*
6231 : * RelFileLocator qsort/bsearch comparator; see RelFileLocatorEquals.
6232 : */
6233 : static int
6234 19253842 : rlocator_comparator(const void *p1, const void *p2)
6235 : {
6236 19253842 : RelFileLocator n1 = *(const RelFileLocator *) p1;
6237 19253842 : RelFileLocator n2 = *(const RelFileLocator *) p2;
6238 :
6239 19253842 : if (n1.relNumber < n2.relNumber)
6240 19184198 : return -1;
6241 69644 : else if (n1.relNumber > n2.relNumber)
6242 66144 : return 1;
6243 :
6244 3500 : if (n1.dbOid < n2.dbOid)
6245 0 : return -1;
6246 3500 : else if (n1.dbOid > n2.dbOid)
6247 0 : return 1;
6248 :
6249 3500 : if (n1.spcOid < n2.spcOid)
6250 0 : return -1;
6251 3500 : else if (n1.spcOid > n2.spcOid)
6252 0 : return 1;
6253 : else
6254 3500 : return 0;
6255 : }
6256 :
6257 : /*
6258 : * Lock buffer header - set BM_LOCKED in buffer state.
6259 : */
6260 : uint32
6261 63223416 : LockBufHdr(BufferDesc *desc)
6262 : {
6263 : SpinDelayStatus delayStatus;
6264 : uint32 old_buf_state;
6265 :
6266 : Assert(!BufferIsLocal(BufferDescriptorGetBuffer(desc)));
6267 :
6268 63223416 : init_local_spin_delay(&delayStatus);
6269 :
6270 : while (true)
6271 : {
6272 : /* set BM_LOCKED flag */
6273 63231638 : old_buf_state = pg_atomic_fetch_or_u32(&desc->state, BM_LOCKED);
6274 : /* if it wasn't set before we're OK */
6275 63231638 : if (!(old_buf_state & BM_LOCKED))
6276 63223416 : break;
6277 8222 : perform_spin_delay(&delayStatus);
6278 : }
6279 63223416 : finish_spin_delay(&delayStatus);
6280 63223416 : return old_buf_state | BM_LOCKED;
6281 : }
6282 :
6283 : /*
6284 : * Wait until the BM_LOCKED flag isn't set anymore and return the buffer's
6285 : * state at that point.
6286 : *
6287 : * Obviously the buffer could be locked by the time the value is returned, so
6288 : * this is primarily useful in CAS style loops.
6289 : */
6290 : pg_noinline uint32
6291 3446 : WaitBufHdrUnlocked(BufferDesc *buf)
6292 : {
6293 : SpinDelayStatus delayStatus;
6294 : uint32 buf_state;
6295 :
6296 3446 : init_local_spin_delay(&delayStatus);
6297 :
6298 3446 : buf_state = pg_atomic_read_u32(&buf->state);
6299 :
6300 28518 : while (buf_state & BM_LOCKED)
6301 : {
6302 25072 : perform_spin_delay(&delayStatus);
6303 25072 : buf_state = pg_atomic_read_u32(&buf->state);
6304 : }
6305 :
6306 3446 : finish_spin_delay(&delayStatus);
6307 :
6308 3446 : return buf_state;
6309 : }
6310 :
6311 : /*
6312 : * BufferTag comparator.
6313 : */
6314 : static inline int
6315 0 : buffertag_comparator(const BufferTag *ba, const BufferTag *bb)
6316 : {
6317 : int ret;
6318 : RelFileLocator rlocatora;
6319 : RelFileLocator rlocatorb;
6320 :
6321 0 : rlocatora = BufTagGetRelFileLocator(ba);
6322 0 : rlocatorb = BufTagGetRelFileLocator(bb);
6323 :
6324 0 : ret = rlocator_comparator(&rlocatora, &rlocatorb);
6325 :
6326 0 : if (ret != 0)
6327 0 : return ret;
6328 :
6329 0 : if (BufTagGetForkNum(ba) < BufTagGetForkNum(bb))
6330 0 : return -1;
6331 0 : if (BufTagGetForkNum(ba) > BufTagGetForkNum(bb))
6332 0 : return 1;
6333 :
6334 0 : if (ba->blockNum < bb->blockNum)
6335 0 : return -1;
6336 0 : if (ba->blockNum > bb->blockNum)
6337 0 : return 1;
6338 :
6339 0 : return 0;
6340 : }
6341 :
6342 : /*
6343 : * Comparator determining the writeout order in a checkpoint.
6344 : *
6345 : * It is important that tablespaces are compared first, the logic balancing
6346 : * writes between tablespaces relies on it.
6347 : */
6348 : static inline int
6349 5999204 : ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b)
6350 : {
6351 : /* compare tablespace */
6352 5999204 : if (a->tsId < b->tsId)
6353 14180 : return -1;
6354 5985024 : else if (a->tsId > b->tsId)
6355 51998 : return 1;
6356 : /* compare relation */
6357 5933026 : if (a->relNumber < b->relNumber)
6358 1674354 : return -1;
6359 4258672 : else if (a->relNumber > b->relNumber)
6360 1606794 : return 1;
6361 : /* compare fork */
6362 2651878 : else if (a->forkNum < b->forkNum)
6363 110678 : return -1;
6364 2541200 : else if (a->forkNum > b->forkNum)
6365 119876 : return 1;
6366 : /* compare block number */
6367 2421324 : else if (a->blockNum < b->blockNum)
6368 1186286 : return -1;
6369 1235038 : else if (a->blockNum > b->blockNum)
6370 1161982 : return 1;
6371 : /* equal page IDs are unlikely, but not impossible */
6372 73056 : return 0;
6373 : }
6374 :
6375 : /*
6376 : * Comparator for a Min-Heap over the per-tablespace checkpoint completion
6377 : * progress.
6378 : */
6379 : static int
6380 488640 : ts_ckpt_progress_comparator(Datum a, Datum b, void *arg)
6381 : {
6382 488640 : CkptTsStatus *sa = (CkptTsStatus *) DatumGetPointer(a);
6383 488640 : CkptTsStatus *sb = (CkptTsStatus *) DatumGetPointer(b);
6384 :
6385 : /* we want a min-heap, so return 1 for the a < b */
6386 488640 : if (sa->progress < sb->progress)
6387 437748 : return 1;
6388 50892 : else if (sa->progress == sb->progress)
6389 2074 : return 0;
6390 : else
6391 48818 : return -1;
6392 : }
6393 :
6394 : /*
6395 : * Initialize a writeback context, discarding potential previous state.
6396 : *
6397 : * *max_pending is a pointer instead of an immediate value, so the coalesce
6398 : * limits can easily changed by the GUC mechanism, and so calling code does
6399 : * not have to check the current configuration. A value of 0 means that no
6400 : * writeback control will be performed.
6401 : */
6402 : void
6403 5358 : WritebackContextInit(WritebackContext *context, int *max_pending)
6404 : {
6405 : Assert(*max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
6406 :
6407 5358 : context->max_pending = max_pending;
6408 5358 : context->nr_pending = 0;
6409 5358 : }
6410 :
6411 : /*
6412 : * Add buffer to list of pending writeback requests.
6413 : */
6414 : void
6415 1092304 : ScheduleBufferTagForWriteback(WritebackContext *wb_context, IOContext io_context,
6416 : BufferTag *tag)
6417 : {
6418 : PendingWriteback *pending;
6419 :
6420 : /*
6421 : * As pg_flush_data() doesn't do anything with fsync disabled, there's no
6422 : * point in tracking in that case.
6423 : */
6424 1092304 : if (io_direct_flags & IO_DIRECT_DATA ||
6425 1091254 : !enableFsync)
6426 1092302 : return;
6427 :
6428 : /*
6429 : * Add buffer to the pending writeback array, unless writeback control is
6430 : * disabled.
6431 : */
6432 2 : if (*wb_context->max_pending > 0)
6433 : {
6434 : Assert(*wb_context->max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
6435 :
6436 0 : pending = &wb_context->pending_writebacks[wb_context->nr_pending++];
6437 :
6438 0 : pending->tag = *tag;
6439 : }
6440 :
6441 : /*
6442 : * Perform pending flushes if the writeback limit is exceeded. This
6443 : * includes the case where previously an item has been added, but control
6444 : * is now disabled.
6445 : */
6446 2 : if (wb_context->nr_pending >= *wb_context->max_pending)
6447 2 : IssuePendingWritebacks(wb_context, io_context);
6448 : }
6449 :
6450 : #define ST_SORT sort_pending_writebacks
6451 : #define ST_ELEMENT_TYPE PendingWriteback
6452 : #define ST_COMPARE(a, b) buffertag_comparator(&a->tag, &b->tag)
6453 : #define ST_SCOPE static
6454 : #define ST_DEFINE
6455 : #include "lib/sort_template.h"
6456 :
6457 : /*
6458 : * Issue all pending writeback requests, previously scheduled with
6459 : * ScheduleBufferTagForWriteback, to the OS.
6460 : *
6461 : * Because this is only used to improve the OSs IO scheduling we try to never
6462 : * error out - it's just a hint.
6463 : */
6464 : void
6465 2110 : IssuePendingWritebacks(WritebackContext *wb_context, IOContext io_context)
6466 : {
6467 : instr_time io_start;
6468 : int i;
6469 :
6470 2110 : if (wb_context->nr_pending == 0)
6471 2110 : return;
6472 :
6473 : /*
6474 : * Executing the writes in-order can make them a lot faster, and allows to
6475 : * merge writeback requests to consecutive blocks into larger writebacks.
6476 : */
6477 0 : sort_pending_writebacks(wb_context->pending_writebacks,
6478 0 : wb_context->nr_pending);
6479 :
6480 0 : io_start = pgstat_prepare_io_time(track_io_timing);
6481 :
6482 : /*
6483 : * Coalesce neighbouring writes, but nothing else. For that we iterate
6484 : * through the, now sorted, array of pending flushes, and look forward to
6485 : * find all neighbouring (or identical) writes.
6486 : */
6487 0 : for (i = 0; i < wb_context->nr_pending; i++)
6488 : {
6489 : PendingWriteback *cur;
6490 : PendingWriteback *next;
6491 : SMgrRelation reln;
6492 : int ahead;
6493 : BufferTag tag;
6494 : RelFileLocator currlocator;
6495 0 : Size nblocks = 1;
6496 :
6497 0 : cur = &wb_context->pending_writebacks[i];
6498 0 : tag = cur->tag;
6499 0 : currlocator = BufTagGetRelFileLocator(&tag);
6500 :
6501 : /*
6502 : * Peek ahead, into following writeback requests, to see if they can
6503 : * be combined with the current one.
6504 : */
6505 0 : for (ahead = 0; i + ahead + 1 < wb_context->nr_pending; ahead++)
6506 : {
6507 :
6508 0 : next = &wb_context->pending_writebacks[i + ahead + 1];
6509 :
6510 : /* different file, stop */
6511 0 : if (!RelFileLocatorEquals(currlocator,
6512 0 : BufTagGetRelFileLocator(&next->tag)) ||
6513 0 : BufTagGetForkNum(&cur->tag) != BufTagGetForkNum(&next->tag))
6514 : break;
6515 :
6516 : /* ok, block queued twice, skip */
6517 0 : if (cur->tag.blockNum == next->tag.blockNum)
6518 0 : continue;
6519 :
6520 : /* only merge consecutive writes */
6521 0 : if (cur->tag.blockNum + 1 != next->tag.blockNum)
6522 0 : break;
6523 :
6524 0 : nblocks++;
6525 0 : cur = next;
6526 : }
6527 :
6528 0 : i += ahead;
6529 :
6530 : /* and finally tell the kernel to write the data to storage */
6531 0 : reln = smgropen(currlocator, INVALID_PROC_NUMBER);
6532 0 : smgrwriteback(reln, BufTagGetForkNum(&tag), tag.blockNum, nblocks);
6533 : }
6534 :
6535 : /*
6536 : * Assume that writeback requests are only issued for buffers containing
6537 : * blocks of permanent relations.
6538 : */
6539 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
6540 0 : IOOP_WRITEBACK, io_start, wb_context->nr_pending, 0);
6541 :
6542 0 : wb_context->nr_pending = 0;
6543 : }
6544 :
6545 : /* ResourceOwner callbacks */
6546 :
6547 : static void
6548 30 : ResOwnerReleaseBufferIO(Datum res)
6549 : {
6550 30 : Buffer buffer = DatumGetInt32(res);
6551 :
6552 30 : AbortBufferIO(buffer);
6553 30 : }
6554 :
6555 : static char *
6556 0 : ResOwnerPrintBufferIO(Datum res)
6557 : {
6558 0 : Buffer buffer = DatumGetInt32(res);
6559 :
6560 0 : return psprintf("lost track of buffer IO on buffer %d", buffer);
6561 : }
6562 :
6563 : static void
6564 15066 : ResOwnerReleaseBufferPin(Datum res)
6565 : {
6566 15066 : Buffer buffer = DatumGetInt32(res);
6567 :
6568 : /* Like ReleaseBuffer, but don't call ResourceOwnerForgetBuffer */
6569 15066 : if (!BufferIsValid(buffer))
6570 0 : elog(ERROR, "bad buffer ID: %d", buffer);
6571 :
6572 15066 : if (BufferIsLocal(buffer))
6573 6066 : UnpinLocalBufferNoOwner(buffer);
6574 : else
6575 9000 : UnpinBufferNoOwner(GetBufferDescriptor(buffer - 1));
6576 15066 : }
6577 :
6578 : static char *
6579 0 : ResOwnerPrintBufferPin(Datum res)
6580 : {
6581 0 : return DebugPrintBufferRefcount(DatumGetInt32(res));
6582 : }
6583 :
6584 : /*
6585 : * Helper function to evict unpinned buffer whose buffer header lock is
6586 : * already acquired.
6587 : */
6588 : static bool
6589 4278 : EvictUnpinnedBufferInternal(BufferDesc *desc, bool *buffer_flushed)
6590 : {
6591 : uint32 buf_state;
6592 : bool result;
6593 :
6594 4278 : *buffer_flushed = false;
6595 :
6596 4278 : buf_state = pg_atomic_read_u32(&(desc->state));
6597 : Assert(buf_state & BM_LOCKED);
6598 :
6599 4278 : if ((buf_state & BM_VALID) == 0)
6600 : {
6601 0 : UnlockBufHdr(desc, buf_state);
6602 0 : return false;
6603 : }
6604 :
6605 : /* Check that it's not pinned already. */
6606 4278 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
6607 : {
6608 0 : UnlockBufHdr(desc, buf_state);
6609 0 : return false;
6610 : }
6611 :
6612 4278 : PinBuffer_Locked(desc); /* releases spinlock */
6613 :
6614 : /* If it was dirty, try to clean it once. */
6615 4278 : if (buf_state & BM_DIRTY)
6616 : {
6617 1940 : FlushUnlockedBuffer(desc, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
6618 1940 : *buffer_flushed = true;
6619 : }
6620 :
6621 : /* This will return false if it becomes dirty or someone else pins it. */
6622 4278 : result = InvalidateVictimBuffer(desc);
6623 :
6624 4278 : UnpinBuffer(desc);
6625 :
6626 4278 : return result;
6627 : }
6628 :
6629 : /*
6630 : * Try to evict the current block in a shared buffer.
6631 : *
6632 : * This function is intended for testing/development use only!
6633 : *
6634 : * To succeed, the buffer must not be pinned on entry, so if the caller had a
6635 : * particular block in mind, it might already have been replaced by some other
6636 : * block by the time this function runs. It's also unpinned on return, so the
6637 : * buffer might be occupied again by the time control is returned, potentially
6638 : * even by the same block. This inherent raciness without other interlocking
6639 : * makes the function unsuitable for non-testing usage.
6640 : *
6641 : * *buffer_flushed is set to true if the buffer was dirty and has been
6642 : * flushed, false otherwise. However, *buffer_flushed=true does not
6643 : * necessarily mean that we flushed the buffer, it could have been flushed by
6644 : * someone else.
6645 : *
6646 : * Returns true if the buffer was valid and it has now been made invalid.
6647 : * Returns false if it wasn't valid, if it couldn't be evicted due to a pin,
6648 : * or if the buffer becomes dirty again while we're trying to write it out.
6649 : */
6650 : bool
6651 280 : EvictUnpinnedBuffer(Buffer buf, bool *buffer_flushed)
6652 : {
6653 : BufferDesc *desc;
6654 :
6655 : Assert(BufferIsValid(buf) && !BufferIsLocal(buf));
6656 :
6657 : /* Make sure we can pin the buffer. */
6658 280 : ResourceOwnerEnlarge(CurrentResourceOwner);
6659 280 : ReservePrivateRefCountEntry();
6660 :
6661 280 : desc = GetBufferDescriptor(buf - 1);
6662 280 : LockBufHdr(desc);
6663 :
6664 280 : return EvictUnpinnedBufferInternal(desc, buffer_flushed);
6665 : }
6666 :
6667 : /*
6668 : * Try to evict all the shared buffers.
6669 : *
6670 : * This function is intended for testing/development use only! See
6671 : * EvictUnpinnedBuffer().
6672 : *
6673 : * The buffers_* parameters are mandatory and indicate the total count of
6674 : * buffers that:
6675 : * - buffers_evicted - were evicted
6676 : * - buffers_flushed - were flushed
6677 : * - buffers_skipped - could not be evicted
6678 : */
6679 : void
6680 2 : EvictAllUnpinnedBuffers(int32 *buffers_evicted, int32 *buffers_flushed,
6681 : int32 *buffers_skipped)
6682 : {
6683 2 : *buffers_evicted = 0;
6684 2 : *buffers_skipped = 0;
6685 2 : *buffers_flushed = 0;
6686 :
6687 32770 : for (int buf = 1; buf <= NBuffers; buf++)
6688 : {
6689 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
6690 : uint32 buf_state;
6691 : bool buffer_flushed;
6692 :
6693 32768 : buf_state = pg_atomic_read_u32(&desc->state);
6694 32768 : if (!(buf_state & BM_VALID))
6695 28770 : continue;
6696 :
6697 3998 : ResourceOwnerEnlarge(CurrentResourceOwner);
6698 3998 : ReservePrivateRefCountEntry();
6699 :
6700 3998 : LockBufHdr(desc);
6701 :
6702 3998 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
6703 3998 : (*buffers_evicted)++;
6704 : else
6705 0 : (*buffers_skipped)++;
6706 :
6707 3998 : if (buffer_flushed)
6708 1902 : (*buffers_flushed)++;
6709 : }
6710 2 : }
6711 :
6712 : /*
6713 : * Try to evict all the shared buffers containing provided relation's pages.
6714 : *
6715 : * This function is intended for testing/development use only! See
6716 : * EvictUnpinnedBuffer().
6717 : *
6718 : * The caller must hold at least AccessShareLock on the relation to prevent
6719 : * the relation from being dropped.
6720 : *
6721 : * The buffers_* parameters are mandatory and indicate the total count of
6722 : * buffers that:
6723 : * - buffers_evicted - were evicted
6724 : * - buffers_flushed - were flushed
6725 : * - buffers_skipped - could not be evicted
6726 : */
6727 : void
6728 2 : EvictRelUnpinnedBuffers(Relation rel, int32 *buffers_evicted,
6729 : int32 *buffers_flushed, int32 *buffers_skipped)
6730 : {
6731 : Assert(!RelationUsesLocalBuffers(rel));
6732 :
6733 2 : *buffers_skipped = 0;
6734 2 : *buffers_evicted = 0;
6735 2 : *buffers_flushed = 0;
6736 :
6737 32770 : for (int buf = 1; buf <= NBuffers; buf++)
6738 : {
6739 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
6740 32768 : uint32 buf_state = pg_atomic_read_u32(&(desc->state));
6741 : bool buffer_flushed;
6742 :
6743 : /* An unlocked precheck should be safe and saves some cycles. */
6744 32768 : if ((buf_state & BM_VALID) == 0 ||
6745 54 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
6746 32768 : continue;
6747 :
6748 : /* Make sure we can pin the buffer. */
6749 0 : ResourceOwnerEnlarge(CurrentResourceOwner);
6750 0 : ReservePrivateRefCountEntry();
6751 :
6752 0 : buf_state = LockBufHdr(desc);
6753 :
6754 : /* recheck, could have changed without the lock */
6755 0 : if ((buf_state & BM_VALID) == 0 ||
6756 0 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
6757 : {
6758 0 : UnlockBufHdr(desc, buf_state);
6759 0 : continue;
6760 : }
6761 :
6762 0 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
6763 0 : (*buffers_evicted)++;
6764 : else
6765 0 : (*buffers_skipped)++;
6766 :
6767 0 : if (buffer_flushed)
6768 0 : (*buffers_flushed)++;
6769 : }
6770 2 : }
6771 :
6772 : /*
6773 : * Generic implementation of the AIO handle staging callback for readv/writev
6774 : * on local/shared buffers.
6775 : *
6776 : * Each readv/writev can target multiple buffers. The buffers have already
6777 : * been registered with the IO handle.
6778 : *
6779 : * To make the IO ready for execution ("staging"), we need to ensure that the
6780 : * targeted buffers are in an appropriate state while the IO is ongoing. For
6781 : * that the AIO subsystem needs to have its own buffer pin, otherwise an error
6782 : * in this backend could lead to this backend's buffer pin being released as
6783 : * part of error handling, which in turn could lead to the buffer being
6784 : * replaced while IO is ongoing.
6785 : */
6786 : static pg_attribute_always_inline void
6787 2462326 : buffer_stage_common(PgAioHandle *ioh, bool is_write, bool is_temp)
6788 : {
6789 : uint64 *io_data;
6790 : uint8 handle_data_len;
6791 : PgAioWaitRef io_ref;
6792 2462326 : BufferTag first PG_USED_FOR_ASSERTS_ONLY = {0};
6793 :
6794 2462326 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
6795 :
6796 2462326 : pgaio_io_get_wref(ioh, &io_ref);
6797 :
6798 : /* iterate over all buffers affected by the vectored readv/writev */
6799 5256722 : for (int i = 0; i < handle_data_len; i++)
6800 : {
6801 2794396 : Buffer buffer = (Buffer) io_data[i];
6802 2794396 : BufferDesc *buf_hdr = is_temp ?
6803 16818 : GetLocalBufferDescriptor(-buffer - 1)
6804 2794396 : : GetBufferDescriptor(buffer - 1);
6805 : uint32 buf_state;
6806 :
6807 : /*
6808 : * Check that all the buffers are actually ones that could conceivably
6809 : * be done in one IO, i.e. are sequential. This is the last
6810 : * buffer-aware code before IO is actually executed and confusion
6811 : * about which buffers are targeted by IO can be hard to debug, making
6812 : * it worth doing extra-paranoid checks.
6813 : */
6814 2794396 : if (i == 0)
6815 2462326 : first = buf_hdr->tag;
6816 : else
6817 : {
6818 : Assert(buf_hdr->tag.relNumber == first.relNumber);
6819 : Assert(buf_hdr->tag.blockNum == first.blockNum + i);
6820 : }
6821 :
6822 2794396 : if (is_temp)
6823 16818 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
6824 : else
6825 2777578 : buf_state = LockBufHdr(buf_hdr);
6826 :
6827 : /* verify the buffer is in the expected state */
6828 : Assert(buf_state & BM_TAG_VALID);
6829 : if (is_write)
6830 : {
6831 : Assert(buf_state & BM_VALID);
6832 : Assert(buf_state & BM_DIRTY);
6833 : }
6834 : else
6835 : {
6836 : Assert(!(buf_state & BM_VALID));
6837 : Assert(!(buf_state & BM_DIRTY));
6838 : }
6839 :
6840 : /* temp buffers don't use BM_IO_IN_PROGRESS */
6841 2794396 : if (!is_temp)
6842 : Assert(buf_state & BM_IO_IN_PROGRESS);
6843 :
6844 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) >= 1);
6845 :
6846 : /*
6847 : * Reflect that the buffer is now owned by the AIO subsystem.
6848 : *
6849 : * For local buffers: This can't be done just via LocalRefCount, as
6850 : * one might initially think, as this backend could error out while
6851 : * AIO is still in progress, releasing all the pins by the backend
6852 : * itself.
6853 : *
6854 : * This pin is released again in TerminateBufferIO().
6855 : */
6856 2794396 : buf_state += BUF_REFCOUNT_ONE;
6857 2794396 : buf_hdr->io_wref = io_ref;
6858 :
6859 2794396 : if (is_temp)
6860 16818 : pg_atomic_unlocked_write_u32(&buf_hdr->state, buf_state);
6861 : else
6862 2777578 : UnlockBufHdr(buf_hdr, buf_state);
6863 :
6864 : /*
6865 : * Ensure the content lock that prevents buffer modifications while
6866 : * the buffer is being written out is not released early due to an
6867 : * error.
6868 : */
6869 2794396 : if (is_write && !is_temp)
6870 : {
6871 : LWLock *content_lock;
6872 :
6873 0 : content_lock = BufferDescriptorGetContentLock(buf_hdr);
6874 :
6875 : Assert(LWLockHeldByMe(content_lock));
6876 :
6877 : /*
6878 : * Lock is now owned by AIO subsystem.
6879 : */
6880 0 : LWLockDisown(content_lock);
6881 : }
6882 :
6883 : /*
6884 : * Stop tracking this buffer via the resowner - the AIO system now
6885 : * keeps track.
6886 : */
6887 2794396 : if (!is_temp)
6888 2777578 : ResourceOwnerForgetBufferIO(CurrentResourceOwner, buffer);
6889 : }
6890 2462326 : }
6891 :
6892 : /*
6893 : * Decode readv errors as encoded by buffer_readv_encode_error().
6894 : */
6895 : static inline void
6896 698 : buffer_readv_decode_error(PgAioResult result,
6897 : bool *zeroed_any,
6898 : bool *ignored_any,
6899 : uint8 *zeroed_or_error_count,
6900 : uint8 *checkfail_count,
6901 : uint8 *first_off)
6902 : {
6903 698 : uint32 rem_error = result.error_data;
6904 :
6905 : /* see static asserts in buffer_readv_encode_error */
6906 : #define READV_COUNT_BITS 7
6907 : #define READV_COUNT_MASK ((1 << READV_COUNT_BITS) - 1)
6908 :
6909 698 : *zeroed_any = rem_error & 1;
6910 698 : rem_error >>= 1;
6911 :
6912 698 : *ignored_any = rem_error & 1;
6913 698 : rem_error >>= 1;
6914 :
6915 698 : *zeroed_or_error_count = rem_error & READV_COUNT_MASK;
6916 698 : rem_error >>= READV_COUNT_BITS;
6917 :
6918 698 : *checkfail_count = rem_error & READV_COUNT_MASK;
6919 698 : rem_error >>= READV_COUNT_BITS;
6920 :
6921 698 : *first_off = rem_error & READV_COUNT_MASK;
6922 698 : rem_error >>= READV_COUNT_BITS;
6923 698 : }
6924 :
6925 : /*
6926 : * Helper to encode errors for buffer_readv_complete()
6927 : *
6928 : * Errors are encoded as follows:
6929 : * - bit 0 indicates whether any page was zeroed (1) or not (0)
6930 : * - bit 1 indicates whether any checksum failure was ignored (1) or not (0)
6931 : * - next READV_COUNT_BITS bits indicate the number of errored or zeroed pages
6932 : * - next READV_COUNT_BITS bits indicate the number of checksum failures
6933 : * - next READV_COUNT_BITS bits indicate the first offset of the first page
6934 : * that was errored or zeroed or, if no errors/zeroes, the first ignored
6935 : * checksum
6936 : */
6937 : static inline void
6938 384 : buffer_readv_encode_error(PgAioResult *result,
6939 : bool is_temp,
6940 : bool zeroed_any,
6941 : bool ignored_any,
6942 : uint8 error_count,
6943 : uint8 zeroed_count,
6944 : uint8 checkfail_count,
6945 : uint8 first_error_off,
6946 : uint8 first_zeroed_off,
6947 : uint8 first_ignored_off)
6948 : {
6949 :
6950 384 : uint8 shift = 0;
6951 384 : uint8 zeroed_or_error_count =
6952 : error_count > 0 ? error_count : zeroed_count;
6953 : uint8 first_off;
6954 :
6955 : StaticAssertStmt(PG_IOV_MAX <= 1 << READV_COUNT_BITS,
6956 : "PG_IOV_MAX is bigger than reserved space for error data");
6957 : StaticAssertStmt((1 + 1 + 3 * READV_COUNT_BITS) <= PGAIO_RESULT_ERROR_BITS,
6958 : "PGAIO_RESULT_ERROR_BITS is insufficient for buffer_readv");
6959 :
6960 : /*
6961 : * We only have space to encode one offset - but luckily that's good
6962 : * enough. If there is an error, the error is the interesting offset, same
6963 : * with a zeroed buffer vs an ignored buffer.
6964 : */
6965 384 : if (error_count > 0)
6966 188 : first_off = first_error_off;
6967 196 : else if (zeroed_count > 0)
6968 160 : first_off = first_zeroed_off;
6969 : else
6970 36 : first_off = first_ignored_off;
6971 :
6972 : Assert(!zeroed_any || error_count == 0);
6973 :
6974 384 : result->error_data = 0;
6975 :
6976 384 : result->error_data |= zeroed_any << shift;
6977 384 : shift += 1;
6978 :
6979 384 : result->error_data |= ignored_any << shift;
6980 384 : shift += 1;
6981 :
6982 384 : result->error_data |= ((uint32) zeroed_or_error_count) << shift;
6983 384 : shift += READV_COUNT_BITS;
6984 :
6985 384 : result->error_data |= ((uint32) checkfail_count) << shift;
6986 384 : shift += READV_COUNT_BITS;
6987 :
6988 384 : result->error_data |= ((uint32) first_off) << shift;
6989 384 : shift += READV_COUNT_BITS;
6990 :
6991 384 : result->id = is_temp ? PGAIO_HCB_LOCAL_BUFFER_READV :
6992 : PGAIO_HCB_SHARED_BUFFER_READV;
6993 :
6994 384 : if (error_count > 0)
6995 188 : result->status = PGAIO_RS_ERROR;
6996 : else
6997 196 : result->status = PGAIO_RS_WARNING;
6998 :
6999 : /*
7000 : * The encoding is complicated enough to warrant cross-checking it against
7001 : * the decode function.
7002 : */
7003 : #ifdef USE_ASSERT_CHECKING
7004 : {
7005 : bool zeroed_any_2,
7006 : ignored_any_2;
7007 : uint8 zeroed_or_error_count_2,
7008 : checkfail_count_2,
7009 : first_off_2;
7010 :
7011 : buffer_readv_decode_error(*result,
7012 : &zeroed_any_2, &ignored_any_2,
7013 : &zeroed_or_error_count_2,
7014 : &checkfail_count_2,
7015 : &first_off_2);
7016 : Assert(zeroed_any == zeroed_any_2);
7017 : Assert(ignored_any == ignored_any_2);
7018 : Assert(zeroed_or_error_count == zeroed_or_error_count_2);
7019 : Assert(checkfail_count == checkfail_count_2);
7020 : Assert(first_off == first_off_2);
7021 : }
7022 : #endif
7023 :
7024 : #undef READV_COUNT_BITS
7025 : #undef READV_COUNT_MASK
7026 384 : }
7027 :
7028 : /*
7029 : * Helper for AIO readv completion callbacks, supporting both shared and temp
7030 : * buffers. Gets called once for each buffer in a multi-page read.
7031 : */
7032 : static pg_attribute_always_inline void
7033 2520094 : buffer_readv_complete_one(PgAioTargetData *td, uint8 buf_off, Buffer buffer,
7034 : uint8 flags, bool failed, bool is_temp,
7035 : bool *buffer_invalid,
7036 : bool *failed_checksum,
7037 : bool *ignored_checksum,
7038 : bool *zeroed_buffer)
7039 : {
7040 2520094 : BufferDesc *buf_hdr = is_temp ?
7041 16818 : GetLocalBufferDescriptor(-buffer - 1)
7042 2520094 : : GetBufferDescriptor(buffer - 1);
7043 2520094 : BufferTag tag = buf_hdr->tag;
7044 2520094 : char *bufdata = BufferGetBlock(buffer);
7045 : uint32 set_flag_bits;
7046 : int piv_flags;
7047 :
7048 : /* check that the buffer is in the expected state for a read */
7049 : #ifdef USE_ASSERT_CHECKING
7050 : {
7051 : uint32 buf_state = pg_atomic_read_u32(&buf_hdr->state);
7052 :
7053 : Assert(buf_state & BM_TAG_VALID);
7054 : Assert(!(buf_state & BM_VALID));
7055 : /* temp buffers don't use BM_IO_IN_PROGRESS */
7056 : if (!is_temp)
7057 : Assert(buf_state & BM_IO_IN_PROGRESS);
7058 : Assert(!(buf_state & BM_DIRTY));
7059 : }
7060 : #endif
7061 :
7062 2520094 : *buffer_invalid = false;
7063 2520094 : *failed_checksum = false;
7064 2520094 : *ignored_checksum = false;
7065 2520094 : *zeroed_buffer = false;
7066 :
7067 : /*
7068 : * We ask PageIsVerified() to only log the message about checksum errors,
7069 : * as the completion might be run in any backend (or IO workers). We will
7070 : * report checksum errors in buffer_readv_report().
7071 : */
7072 2520094 : piv_flags = PIV_LOG_LOG;
7073 :
7074 : /* the local zero_damaged_pages may differ from the definer's */
7075 2520094 : if (flags & READ_BUFFERS_IGNORE_CHECKSUM_FAILURES)
7076 76 : piv_flags |= PIV_IGNORE_CHECKSUM_FAILURE;
7077 :
7078 : /* Check for garbage data. */
7079 2520094 : if (!failed)
7080 : {
7081 : /*
7082 : * If the buffer is not currently pinned by this backend, e.g. because
7083 : * we're completing this IO after an error, the buffer data will have
7084 : * been marked as inaccessible when the buffer was unpinned. The AIO
7085 : * subsystem holds a pin, but that doesn't prevent the buffer from
7086 : * having been marked as inaccessible. The completion might also be
7087 : * executed in a different process.
7088 : */
7089 : #ifdef USE_VALGRIND
7090 : if (!BufferIsPinned(buffer))
7091 : VALGRIND_MAKE_MEM_DEFINED(bufdata, BLCKSZ);
7092 : #endif
7093 :
7094 2520036 : if (!PageIsVerified((Page) bufdata, tag.blockNum, piv_flags,
7095 : failed_checksum))
7096 : {
7097 192 : if (flags & READ_BUFFERS_ZERO_ON_ERROR)
7098 : {
7099 92 : memset(bufdata, 0, BLCKSZ);
7100 92 : *zeroed_buffer = true;
7101 : }
7102 : else
7103 : {
7104 100 : *buffer_invalid = true;
7105 : /* mark buffer as having failed */
7106 100 : failed = true;
7107 : }
7108 : }
7109 2519844 : else if (*failed_checksum)
7110 24 : *ignored_checksum = true;
7111 :
7112 : /* undo what we did above */
7113 : #ifdef USE_VALGRIND
7114 : if (!BufferIsPinned(buffer))
7115 : VALGRIND_MAKE_MEM_NOACCESS(bufdata, BLCKSZ);
7116 : #endif
7117 :
7118 : /*
7119 : * Immediately log a message about the invalid page, but only to the
7120 : * server log. The reason to do so immediately is that this may be
7121 : * executed in a different backend than the one that originated the
7122 : * request. The reason to do so immediately is that the originator
7123 : * might not process the query result immediately (because it is busy
7124 : * doing another part of query processing) or at all (e.g. if it was
7125 : * cancelled or errored out due to another IO also failing). The
7126 : * definer of the IO will emit an ERROR or WARNING when processing the
7127 : * IO's results
7128 : *
7129 : * To avoid duplicating the code to emit these log messages, we reuse
7130 : * buffer_readv_report().
7131 : */
7132 2520036 : if (*buffer_invalid || *failed_checksum || *zeroed_buffer)
7133 : {
7134 216 : PgAioResult result_one = {0};
7135 :
7136 216 : buffer_readv_encode_error(&result_one, is_temp,
7137 216 : *zeroed_buffer,
7138 216 : *ignored_checksum,
7139 216 : *buffer_invalid,
7140 216 : *zeroed_buffer ? 1 : 0,
7141 216 : *failed_checksum ? 1 : 0,
7142 : buf_off, buf_off, buf_off);
7143 216 : pgaio_result_report(result_one, td, LOG_SERVER_ONLY);
7144 : }
7145 : }
7146 :
7147 : /* Terminate I/O and set BM_VALID. */
7148 2520094 : set_flag_bits = failed ? BM_IO_ERROR : BM_VALID;
7149 2520094 : if (is_temp)
7150 16818 : TerminateLocalBufferIO(buf_hdr, false, set_flag_bits, true);
7151 : else
7152 2503276 : TerminateBufferIO(buf_hdr, false, set_flag_bits, false, true);
7153 :
7154 : /*
7155 : * Call the BUFFER_READ_DONE tracepoint in the callback, even though the
7156 : * callback may not be executed in the same backend that called
7157 : * BUFFER_READ_START. The alternative would be to defer calling the
7158 : * tracepoint to a later point (e.g. the local completion callback for
7159 : * shared buffer reads), which seems even less helpful.
7160 : */
7161 : TRACE_POSTGRESQL_BUFFER_READ_DONE(tag.forkNum,
7162 : tag.blockNum,
7163 : tag.spcOid,
7164 : tag.dbOid,
7165 : tag.relNumber,
7166 : is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
7167 : false);
7168 2520094 : }
7169 :
7170 : /*
7171 : * Perform completion handling of a single AIO read. This read may cover
7172 : * multiple blocks / buffers.
7173 : *
7174 : * Shared between shared and local buffers, to reduce code duplication.
7175 : */
7176 : static pg_attribute_always_inline PgAioResult
7177 2253518 : buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7178 : uint8 cb_data, bool is_temp)
7179 : {
7180 2253518 : PgAioResult result = prior_result;
7181 2253518 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
7182 2253518 : uint8 first_error_off = 0;
7183 2253518 : uint8 first_zeroed_off = 0;
7184 2253518 : uint8 first_ignored_off = 0;
7185 2253518 : uint8 error_count = 0;
7186 2253518 : uint8 zeroed_count = 0;
7187 2253518 : uint8 ignored_count = 0;
7188 2253518 : uint8 checkfail_count = 0;
7189 : uint64 *io_data;
7190 : uint8 handle_data_len;
7191 :
7192 : if (is_temp)
7193 : {
7194 : Assert(td->smgr.is_temp);
7195 : Assert(pgaio_io_get_owner(ioh) == MyProcNumber);
7196 : }
7197 : else
7198 : Assert(!td->smgr.is_temp);
7199 :
7200 : /*
7201 : * Iterate over all the buffers affected by this IO and call the
7202 : * per-buffer completion function for each buffer.
7203 : */
7204 2253518 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
7205 4773612 : for (uint8 buf_off = 0; buf_off < handle_data_len; buf_off++)
7206 : {
7207 2520094 : Buffer buf = io_data[buf_off];
7208 : bool failed;
7209 2520094 : bool failed_verification = false;
7210 2520094 : bool failed_checksum = false;
7211 2520094 : bool zeroed_buffer = false;
7212 2520094 : bool ignored_checksum = false;
7213 :
7214 : Assert(BufferIsValid(buf));
7215 :
7216 : /*
7217 : * If the entire I/O failed on a lower-level, each buffer needs to be
7218 : * marked as failed. In case of a partial read, the first few buffers
7219 : * may be ok.
7220 : */
7221 2520094 : failed =
7222 2520094 : prior_result.status == PGAIO_RS_ERROR
7223 2520094 : || prior_result.result <= buf_off;
7224 :
7225 2520094 : buffer_readv_complete_one(td, buf_off, buf, cb_data, failed, is_temp,
7226 : &failed_verification,
7227 : &failed_checksum,
7228 : &ignored_checksum,
7229 : &zeroed_buffer);
7230 :
7231 : /*
7232 : * Track information about the number of different kinds of error
7233 : * conditions across all pages, as there can be multiple pages failing
7234 : * verification as part of one IO.
7235 : */
7236 2520094 : if (failed_verification && !zeroed_buffer && error_count++ == 0)
7237 88 : first_error_off = buf_off;
7238 2520094 : if (zeroed_buffer && zeroed_count++ == 0)
7239 68 : first_zeroed_off = buf_off;
7240 2520094 : if (ignored_checksum && ignored_count++ == 0)
7241 20 : first_ignored_off = buf_off;
7242 2520094 : if (failed_checksum)
7243 64 : checkfail_count++;
7244 : }
7245 :
7246 : /*
7247 : * If the smgr read succeeded [partially] and page verification failed for
7248 : * some of the pages, adjust the IO's result state appropriately.
7249 : */
7250 2253518 : if (prior_result.status != PGAIO_RS_ERROR &&
7251 2253412 : (error_count > 0 || ignored_count > 0 || zeroed_count > 0))
7252 : {
7253 168 : buffer_readv_encode_error(&result, is_temp,
7254 : zeroed_count > 0, ignored_count > 0,
7255 : error_count, zeroed_count, checkfail_count,
7256 : first_error_off, first_zeroed_off,
7257 : first_ignored_off);
7258 168 : pgaio_result_report(result, td, DEBUG1);
7259 : }
7260 :
7261 : /*
7262 : * For shared relations this reporting is done in
7263 : * shared_buffer_readv_complete_local().
7264 : */
7265 2253518 : if (is_temp && checkfail_count > 0)
7266 4 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
7267 : checkfail_count);
7268 :
7269 2253518 : return result;
7270 : }
7271 :
7272 : /*
7273 : * AIO error reporting callback for aio_shared_buffer_readv_cb and
7274 : * aio_local_buffer_readv_cb.
7275 : *
7276 : * The error is encoded / decoded in buffer_readv_encode_error() /
7277 : * buffer_readv_decode_error().
7278 : */
7279 : static void
7280 544 : buffer_readv_report(PgAioResult result, const PgAioTargetData *td,
7281 : int elevel)
7282 : {
7283 544 : int nblocks = td->smgr.nblocks;
7284 544 : BlockNumber first = td->smgr.blockNum;
7285 544 : BlockNumber last = first + nblocks - 1;
7286 544 : ProcNumber errProc =
7287 544 : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER;
7288 : RelPathStr rpath =
7289 544 : relpathbackend(td->smgr.rlocator, errProc, td->smgr.forkNum);
7290 : bool zeroed_any,
7291 : ignored_any;
7292 : uint8 zeroed_or_error_count,
7293 : checkfail_count,
7294 : first_off;
7295 : uint8 affected_count;
7296 : const char *msg_one,
7297 : *msg_mult,
7298 : *det_mult,
7299 : *hint_mult;
7300 :
7301 544 : buffer_readv_decode_error(result, &zeroed_any, &ignored_any,
7302 : &zeroed_or_error_count,
7303 : &checkfail_count,
7304 : &first_off);
7305 :
7306 : /*
7307 : * Treat a read that had both zeroed buffers *and* ignored checksums as a
7308 : * special case, it's too irregular to be emitted the same way as the
7309 : * other cases.
7310 : */
7311 544 : if (zeroed_any && ignored_any)
7312 : {
7313 : Assert(zeroed_any && ignored_any);
7314 : Assert(nblocks > 1); /* same block can't be both zeroed and ignored */
7315 : Assert(result.status != PGAIO_RS_ERROR);
7316 8 : affected_count = zeroed_or_error_count;
7317 :
7318 8 : ereport(elevel,
7319 : errcode(ERRCODE_DATA_CORRUPTED),
7320 : errmsg("zeroing %u page(s) and ignoring %u checksum failure(s) among blocks %u..%u of relation \"%s\"",
7321 : affected_count, checkfail_count, first, last, rpath.str),
7322 : affected_count > 1 ?
7323 : errdetail("Block %u held the first zeroed page.",
7324 : first + first_off) : 0,
7325 : errhint_plural("See server log for details about the other %d invalid block.",
7326 : "See server log for details about the other %d invalid blocks.",
7327 : affected_count + checkfail_count - 1,
7328 : affected_count + checkfail_count - 1));
7329 8 : return;
7330 : }
7331 :
7332 : /*
7333 : * The other messages are highly repetitive. To avoid duplicating a long
7334 : * and complicated ereport(), gather the translated format strings
7335 : * separately and then do one common ereport.
7336 : */
7337 536 : if (result.status == PGAIO_RS_ERROR)
7338 : {
7339 : Assert(!zeroed_any); /* can't have invalid pages when zeroing them */
7340 272 : affected_count = zeroed_or_error_count;
7341 272 : msg_one = _("invalid page in block %u of relation \"%s\"");
7342 272 : msg_mult = _("%u invalid pages among blocks %u..%u of relation \"%s\"");
7343 272 : det_mult = _("Block %u held the first invalid page.");
7344 272 : hint_mult = _("See server log for the other %u invalid block(s).");
7345 : }
7346 264 : else if (zeroed_any && !ignored_any)
7347 : {
7348 216 : affected_count = zeroed_or_error_count;
7349 216 : msg_one = _("invalid page in block %u of relation \"%s\"; zeroing out page");
7350 216 : msg_mult = _("zeroing out %u invalid pages among blocks %u..%u of relation \"%s\"");
7351 216 : det_mult = _("Block %u held the first zeroed page.");
7352 216 : hint_mult = _("See server log for the other %u zeroed block(s).");
7353 : }
7354 48 : else if (!zeroed_any && ignored_any)
7355 : {
7356 48 : affected_count = checkfail_count;
7357 48 : msg_one = _("ignoring checksum failure in block %u of relation \"%s\"");
7358 48 : msg_mult = _("ignoring %u checksum failures among blocks %u..%u of relation \"%s\"");
7359 48 : det_mult = _("Block %u held the first ignored page.");
7360 48 : hint_mult = _("See server log for the other %u ignored block(s).");
7361 : }
7362 : else
7363 0 : pg_unreachable();
7364 :
7365 536 : ereport(elevel,
7366 : errcode(ERRCODE_DATA_CORRUPTED),
7367 : affected_count == 1 ?
7368 : errmsg_internal(msg_one, first + first_off, rpath.str) :
7369 : errmsg_internal(msg_mult, affected_count, first, last, rpath.str),
7370 : affected_count > 1 ? errdetail_internal(det_mult, first + first_off) : 0,
7371 : affected_count > 1 ? errhint_internal(hint_mult, affected_count - 1) : 0);
7372 : }
7373 :
7374 : static void
7375 2458714 : shared_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
7376 : {
7377 2458714 : buffer_stage_common(ioh, false, false);
7378 2458714 : }
7379 :
7380 : static PgAioResult
7381 2249906 : shared_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7382 : uint8 cb_data)
7383 : {
7384 2249906 : return buffer_readv_complete(ioh, prior_result, cb_data, false);
7385 : }
7386 :
7387 : /*
7388 : * We need a backend-local completion callback for shared buffers, to be able
7389 : * to report checksum errors correctly. Unfortunately that can only safely
7390 : * happen if the reporting backend has previously called
7391 : * pgstat_prepare_report_checksum_failure(), which we can only guarantee in
7392 : * the backend that started the IO. Hence this callback.
7393 : */
7394 : static PgAioResult
7395 2458714 : shared_buffer_readv_complete_local(PgAioHandle *ioh, PgAioResult prior_result,
7396 : uint8 cb_data)
7397 : {
7398 : bool zeroed_any,
7399 : ignored_any;
7400 : uint8 zeroed_or_error_count,
7401 : checkfail_count,
7402 : first_off;
7403 :
7404 2458714 : if (prior_result.status == PGAIO_RS_OK)
7405 2458560 : return prior_result;
7406 :
7407 154 : buffer_readv_decode_error(prior_result,
7408 : &zeroed_any,
7409 : &ignored_any,
7410 : &zeroed_or_error_count,
7411 : &checkfail_count,
7412 : &first_off);
7413 :
7414 154 : if (checkfail_count)
7415 : {
7416 48 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
7417 :
7418 48 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
7419 : checkfail_count);
7420 : }
7421 :
7422 154 : return prior_result;
7423 : }
7424 :
7425 : static void
7426 3612 : local_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
7427 : {
7428 3612 : buffer_stage_common(ioh, false, true);
7429 3612 : }
7430 :
7431 : static PgAioResult
7432 3612 : local_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7433 : uint8 cb_data)
7434 : {
7435 3612 : return buffer_readv_complete(ioh, prior_result, cb_data, true);
7436 : }
7437 :
7438 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
7439 : const PgAioHandleCallbacks aio_shared_buffer_readv_cb = {
7440 : .stage = shared_buffer_readv_stage,
7441 : .complete_shared = shared_buffer_readv_complete,
7442 : /* need a local callback to report checksum failures */
7443 : .complete_local = shared_buffer_readv_complete_local,
7444 : .report = buffer_readv_report,
7445 : };
7446 :
7447 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
7448 : const PgAioHandleCallbacks aio_local_buffer_readv_cb = {
7449 : .stage = local_buffer_readv_stage,
7450 :
7451 : /*
7452 : * Note that this, in contrast to the shared_buffers case, uses
7453 : * complete_local, as only the issuing backend has access to the required
7454 : * datastructures. This is important in case the IO completion may be
7455 : * consumed incidentally by another backend.
7456 : */
7457 : .complete_local = local_buffer_readv_complete,
7458 : .report = buffer_readv_report,
7459 : };
|
