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 129663440 : ReservePrivateRefCountEntry(void)
260 : {
261 : /* Already reserved (or freed), nothing to do */
262 129663440 : if (ReservedRefCountEntry != NULL)
263 121445710 : 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 20964130 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
273 : {
274 : PrivateRefCountEntry *res;
275 :
276 20624922 : res = &PrivateRefCountArray[i];
277 :
278 20624922 : if (res->buffer == InvalidBuffer)
279 : {
280 7878522 : ReservedRefCountEntry = res;
281 7878522 : 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 339208 : ReservedRefCountEntry =
300 339208 : &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 339208 : hashent = hash_search(PrivateRefCountHash,
307 339208 : &(ReservedRefCountEntry->buffer),
308 : HASH_ENTER,
309 : &found);
310 : Assert(!found);
311 339208 : hashent->refcount = ReservedRefCountEntry->refcount;
312 :
313 : /* clear the now free array slot */
314 339208 : ReservedRefCountEntry->buffer = InvalidBuffer;
315 339208 : ReservedRefCountEntry->refcount = 0;
316 :
317 339208 : PrivateRefCountOverflowed++;
318 : }
319 : }
320 :
321 : /*
322 : * Fill a previously reserved refcount entry.
323 : */
324 : static PrivateRefCountEntry *
325 118057786 : 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 118057786 : res = ReservedRefCountEntry;
334 118057786 : ReservedRefCountEntry = NULL;
335 :
336 : /* and fill it */
337 118057786 : res->buffer = buffer;
338 118057786 : res->refcount = 0;
339 :
340 118057786 : 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 291873992 : 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 1372758422 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
364 : {
365 1258748676 : res = &PrivateRefCountArray[i];
366 :
367 1258748676 : if (res->buffer == buffer)
368 177864246 : 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 114009746 : if (PrivateRefCountOverflowed == 0)
379 113206274 : return NULL;
380 :
381 803472 : res = hash_search(PrivateRefCountHash, &buffer, HASH_FIND, NULL);
382 :
383 803472 : if (res == NULL)
384 459708 : return NULL;
385 343764 : else if (!do_move)
386 : {
387 : /* caller doesn't want us to move the hash entry into the array */
388 302218 : 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 41546 : ReservePrivateRefCountEntry();
398 :
399 : /* Use up the reserved slot */
400 : Assert(ReservedRefCountEntry != NULL);
401 41546 : free = ReservedRefCountEntry;
402 41546 : ReservedRefCountEntry = NULL;
403 : Assert(free->buffer == InvalidBuffer);
404 :
405 : /* and fill it */
406 41546 : free->buffer = buffer;
407 41546 : free->refcount = res->refcount;
408 :
409 : /* delete from hashtable */
410 41546 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
411 : Assert(found);
412 : Assert(PrivateRefCountOverflowed > 0);
413 41546 : PrivateRefCountOverflowed--;
414 :
415 41546 : 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 6404312 : 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 6404312 : ref = GetPrivateRefCountEntry(buffer, false);
437 :
438 6404312 : if (ref == NULL)
439 8936 : return 0;
440 6395376 : 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 118057786 : ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref)
449 : {
450 : Assert(ref->refcount == 0);
451 :
452 118057786 : if (ref >= &PrivateRefCountArray[0] &&
453 : ref < &PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES])
454 : {
455 117760124 : 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 117760124 : ReservedRefCountEntry = ref;
463 : }
464 : else
465 : {
466 : bool found;
467 297662 : Buffer buffer = ref->buffer;
468 :
469 297662 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
470 : Assert(found);
471 : Assert(PrivateRefCountOverflowed > 0);
472 297662 : PrivateRefCountOverflowed--;
473 : }
474 118057786 : }
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 : static void PinBuffer_Locked(BufferDesc *buf);
517 : static void UnpinBuffer(BufferDesc *buf);
518 : static void UnpinBufferNoOwner(BufferDesc *buf);
519 : static void BufferSync(int flags);
520 : static uint32 WaitBufHdrUnlocked(BufferDesc *buf);
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 FlushBuffer(BufferDesc *buf, SMgrRelation reln,
537 : IOObject io_object, IOContext io_context);
538 : static void FindAndDropRelationBuffers(RelFileLocator rlocator,
539 : ForkNumber forkNum,
540 : BlockNumber nForkBlock,
541 : BlockNumber firstDelBlock);
542 : static void RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
543 : RelFileLocator dstlocator,
544 : ForkNumber forkNum, bool permanent);
545 : static void AtProcExit_Buffers(int code, Datum arg);
546 : static void CheckForBufferLeaks(void);
547 : #ifdef USE_ASSERT_CHECKING
548 : static void AssertNotCatalogBufferLock(LWLock *lock, LWLockMode mode,
549 : void *unused_context);
550 : #endif
551 : static int rlocator_comparator(const void *p1, const void *p2);
552 : static inline int buffertag_comparator(const BufferTag *ba, const BufferTag *bb);
553 : static inline int ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b);
554 : static int ts_ckpt_progress_comparator(Datum a, Datum b, void *arg);
555 :
556 :
557 : /*
558 : * Implementation of PrefetchBuffer() for shared buffers.
559 : */
560 : PrefetchBufferResult
561 60976 : PrefetchSharedBuffer(SMgrRelation smgr_reln,
562 : ForkNumber forkNum,
563 : BlockNumber blockNum)
564 : {
565 60976 : PrefetchBufferResult result = {InvalidBuffer, false};
566 : BufferTag newTag; /* identity of requested block */
567 : uint32 newHash; /* hash value for newTag */
568 : LWLock *newPartitionLock; /* buffer partition lock for it */
569 : int buf_id;
570 :
571 : Assert(BlockNumberIsValid(blockNum));
572 :
573 : /* create a tag so we can lookup the buffer */
574 60976 : InitBufferTag(&newTag, &smgr_reln->smgr_rlocator.locator,
575 : forkNum, blockNum);
576 :
577 : /* determine its hash code and partition lock ID */
578 60976 : newHash = BufTableHashCode(&newTag);
579 60976 : newPartitionLock = BufMappingPartitionLock(newHash);
580 :
581 : /* see if the block is in the buffer pool already */
582 60976 : LWLockAcquire(newPartitionLock, LW_SHARED);
583 60976 : buf_id = BufTableLookup(&newTag, newHash);
584 60976 : LWLockRelease(newPartitionLock);
585 :
586 : /* If not in buffers, initiate prefetch */
587 60976 : if (buf_id < 0)
588 : {
589 : #ifdef USE_PREFETCH
590 : /*
591 : * Try to initiate an asynchronous read. This returns false in
592 : * recovery if the relation file doesn't exist.
593 : */
594 34888 : if ((io_direct_flags & IO_DIRECT_DATA) == 0 &&
595 17220 : smgrprefetch(smgr_reln, forkNum, blockNum, 1))
596 : {
597 17220 : result.initiated_io = true;
598 : }
599 : #endif /* USE_PREFETCH */
600 : }
601 : else
602 : {
603 : /*
604 : * Report the buffer it was in at that time. The caller may be able
605 : * to avoid a buffer table lookup, but it's not pinned and it must be
606 : * rechecked!
607 : */
608 43308 : result.recent_buffer = buf_id + 1;
609 : }
610 :
611 : /*
612 : * If the block *is* in buffers, we do nothing. This is not really ideal:
613 : * the block might be just about to be evicted, which would be stupid
614 : * since we know we are going to need it soon. But the only easy answer
615 : * is to bump the usage_count, which does not seem like a great solution:
616 : * when the caller does ultimately touch the block, usage_count would get
617 : * bumped again, resulting in too much favoritism for blocks that are
618 : * involved in a prefetch sequence. A real fix would involve some
619 : * additional per-buffer state, and it's not clear that there's enough of
620 : * a problem to justify that.
621 : */
622 :
623 60976 : return result;
624 : }
625 :
626 : /*
627 : * PrefetchBuffer -- initiate asynchronous read of a block of a relation
628 : *
629 : * This is named by analogy to ReadBuffer but doesn't actually allocate a
630 : * buffer. Instead it tries to ensure that a future ReadBuffer for the given
631 : * block will not be delayed by the I/O. Prefetching is optional.
632 : *
633 : * There are three possible outcomes:
634 : *
635 : * 1. If the block is already cached, the result includes a valid buffer that
636 : * could be used by the caller to avoid the need for a later buffer lookup, but
637 : * it's not pinned, so the caller must recheck it.
638 : *
639 : * 2. If the kernel has been asked to initiate I/O, the initiated_io member is
640 : * true. Currently there is no way to know if the data was already cached by
641 : * the kernel and therefore didn't really initiate I/O, and no way to know when
642 : * the I/O completes other than using synchronous ReadBuffer().
643 : *
644 : * 3. Otherwise, the buffer wasn't already cached by PostgreSQL, and
645 : * USE_PREFETCH is not defined (this build doesn't support prefetching due to
646 : * lack of a kernel facility), direct I/O is enabled, or the underlying
647 : * relation file wasn't found and we are in recovery. (If the relation file
648 : * wasn't found and we are not in recovery, an error is raised).
649 : */
650 : PrefetchBufferResult
651 39786 : PrefetchBuffer(Relation reln, ForkNumber forkNum, BlockNumber blockNum)
652 : {
653 : Assert(RelationIsValid(reln));
654 : Assert(BlockNumberIsValid(blockNum));
655 :
656 39786 : if (RelationUsesLocalBuffers(reln))
657 : {
658 : /* see comments in ReadBufferExtended */
659 1566 : if (RELATION_IS_OTHER_TEMP(reln))
660 0 : ereport(ERROR,
661 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
662 : errmsg("cannot access temporary tables of other sessions")));
663 :
664 : /* pass it off to localbuf.c */
665 1566 : return PrefetchLocalBuffer(RelationGetSmgr(reln), forkNum, blockNum);
666 : }
667 : else
668 : {
669 : /* pass it to the shared buffer version */
670 38220 : return PrefetchSharedBuffer(RelationGetSmgr(reln), forkNum, blockNum);
671 : }
672 : }
673 :
674 : /*
675 : * ReadRecentBuffer -- try to pin a block in a recently observed buffer
676 : *
677 : * Compared to ReadBuffer(), this avoids a buffer mapping lookup when it's
678 : * successful. Return true if the buffer is valid and still has the expected
679 : * tag. In that case, the buffer is pinned and the usage count is bumped.
680 : */
681 : bool
682 8950 : ReadRecentBuffer(RelFileLocator rlocator, ForkNumber forkNum, BlockNumber blockNum,
683 : Buffer recent_buffer)
684 : {
685 : BufferDesc *bufHdr;
686 : BufferTag tag;
687 : uint32 buf_state;
688 : bool have_private_ref;
689 :
690 : Assert(BufferIsValid(recent_buffer));
691 :
692 8950 : ResourceOwnerEnlarge(CurrentResourceOwner);
693 8950 : ReservePrivateRefCountEntry();
694 8950 : InitBufferTag(&tag, &rlocator, forkNum, blockNum);
695 :
696 8950 : if (BufferIsLocal(recent_buffer))
697 : {
698 64 : int b = -recent_buffer - 1;
699 :
700 64 : bufHdr = GetLocalBufferDescriptor(b);
701 64 : buf_state = pg_atomic_read_u32(&bufHdr->state);
702 :
703 : /* Is it still valid and holding the right tag? */
704 64 : if ((buf_state & BM_VALID) && BufferTagsEqual(&tag, &bufHdr->tag))
705 : {
706 64 : PinLocalBuffer(bufHdr, true);
707 :
708 64 : pgBufferUsage.local_blks_hit++;
709 :
710 64 : return true;
711 : }
712 : }
713 : else
714 : {
715 8886 : bufHdr = GetBufferDescriptor(recent_buffer - 1);
716 8886 : have_private_ref = GetPrivateRefCount(recent_buffer) > 0;
717 :
718 : /*
719 : * Do we already have this buffer pinned with a private reference? If
720 : * so, it must be valid and it is safe to check the tag without
721 : * locking. If not, we have to lock the header first and then check.
722 : */
723 8886 : if (have_private_ref)
724 0 : buf_state = pg_atomic_read_u32(&bufHdr->state);
725 : else
726 8886 : buf_state = LockBufHdr(bufHdr);
727 :
728 8886 : if ((buf_state & BM_VALID) && BufferTagsEqual(&tag, &bufHdr->tag))
729 : {
730 : /*
731 : * It's now safe to pin the buffer. We can't pin first and ask
732 : * questions later, because it might confuse code paths like
733 : * InvalidateBuffer() if we pinned a random non-matching buffer.
734 : */
735 8768 : if (have_private_ref)
736 0 : PinBuffer(bufHdr, NULL); /* bump pin count */
737 : else
738 8768 : PinBuffer_Locked(bufHdr); /* pin for first time */
739 :
740 8768 : pgBufferUsage.shared_blks_hit++;
741 :
742 8768 : return true;
743 : }
744 :
745 : /* If we locked the header above, now unlock. */
746 118 : if (!have_private_ref)
747 118 : UnlockBufHdr(bufHdr, buf_state);
748 : }
749 :
750 118 : return false;
751 : }
752 :
753 : /*
754 : * ReadBuffer -- a shorthand for ReadBufferExtended, for reading from main
755 : * fork with RBM_NORMAL mode and default strategy.
756 : */
757 : Buffer
758 86116118 : ReadBuffer(Relation reln, BlockNumber blockNum)
759 : {
760 86116118 : return ReadBufferExtended(reln, MAIN_FORKNUM, blockNum, RBM_NORMAL, NULL);
761 : }
762 :
763 : /*
764 : * ReadBufferExtended -- returns a buffer containing the requested
765 : * block of the requested relation. If the blknum
766 : * requested is P_NEW, extend the relation file and
767 : * allocate a new block. (Caller is responsible for
768 : * ensuring that only one backend tries to extend a
769 : * relation at the same time!)
770 : *
771 : * Returns: the buffer number for the buffer containing
772 : * the block read. The returned buffer has been pinned.
773 : * Does not return on error --- elog's instead.
774 : *
775 : * Assume when this function is called, that reln has been opened already.
776 : *
777 : * In RBM_NORMAL mode, the page is read from disk, and the page header is
778 : * validated. An error is thrown if the page header is not valid. (But
779 : * note that an all-zero page is considered "valid"; see
780 : * PageIsVerified().)
781 : *
782 : * RBM_ZERO_ON_ERROR is like the normal mode, but if the page header is not
783 : * valid, the page is zeroed instead of throwing an error. This is intended
784 : * for non-critical data, where the caller is prepared to repair errors.
785 : *
786 : * In RBM_ZERO_AND_LOCK mode, if the page isn't in buffer cache already, it's
787 : * filled with zeros instead of reading it from disk. Useful when the caller
788 : * is going to fill the page from scratch, since this saves I/O and avoids
789 : * unnecessary failure if the page-on-disk has corrupt page headers.
790 : * The page is returned locked to ensure that the caller has a chance to
791 : * initialize the page before it's made visible to others.
792 : * Caution: do not use this mode to read a page that is beyond the relation's
793 : * current physical EOF; that is likely to cause problems in md.c when
794 : * the page is modified and written out. P_NEW is OK, though.
795 : *
796 : * RBM_ZERO_AND_CLEANUP_LOCK is the same as RBM_ZERO_AND_LOCK, but acquires
797 : * a cleanup-strength lock on the page.
798 : *
799 : * RBM_NORMAL_NO_LOG mode is treated the same as RBM_NORMAL here.
800 : *
801 : * If strategy is not NULL, a nondefault buffer access strategy is used.
802 : * See buffer/README for details.
803 : */
804 : inline Buffer
805 104171442 : ReadBufferExtended(Relation reln, ForkNumber forkNum, BlockNumber blockNum,
806 : ReadBufferMode mode, BufferAccessStrategy strategy)
807 : {
808 : Buffer buf;
809 :
810 : /*
811 : * Reject attempts to read non-local temporary relations; we would be
812 : * likely to get wrong data since we have no visibility into the owning
813 : * session's local buffers.
814 : */
815 104171442 : if (RELATION_IS_OTHER_TEMP(reln))
816 0 : ereport(ERROR,
817 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
818 : errmsg("cannot access temporary tables of other sessions")));
819 :
820 : /*
821 : * Read the buffer, and update pgstat counters to reflect a cache hit or
822 : * miss.
823 : */
824 104171442 : buf = ReadBuffer_common(reln, RelationGetSmgr(reln), 0,
825 : forkNum, blockNum, mode, strategy);
826 :
827 104171396 : return buf;
828 : }
829 :
830 :
831 : /*
832 : * ReadBufferWithoutRelcache -- like ReadBufferExtended, but doesn't require
833 : * a relcache entry for the relation.
834 : *
835 : * Pass permanent = true for a RELPERSISTENCE_PERMANENT relation, and
836 : * permanent = false for a RELPERSISTENCE_UNLOGGED relation. This function
837 : * cannot be used for temporary relations (and making that work might be
838 : * difficult, unless we only want to read temporary relations for our own
839 : * ProcNumber).
840 : */
841 : Buffer
842 11343910 : ReadBufferWithoutRelcache(RelFileLocator rlocator, ForkNumber forkNum,
843 : BlockNumber blockNum, ReadBufferMode mode,
844 : BufferAccessStrategy strategy, bool permanent)
845 : {
846 11343910 : SMgrRelation smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
847 :
848 11343910 : return ReadBuffer_common(NULL, smgr,
849 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
850 : forkNum, blockNum,
851 : mode, strategy);
852 : }
853 :
854 : /*
855 : * Convenience wrapper around ExtendBufferedRelBy() extending by one block.
856 : */
857 : Buffer
858 90422 : ExtendBufferedRel(BufferManagerRelation bmr,
859 : ForkNumber forkNum,
860 : BufferAccessStrategy strategy,
861 : uint32 flags)
862 : {
863 : Buffer buf;
864 90422 : uint32 extend_by = 1;
865 :
866 90422 : ExtendBufferedRelBy(bmr, forkNum, strategy, flags, extend_by,
867 : &buf, &extend_by);
868 :
869 90422 : return buf;
870 : }
871 :
872 : /*
873 : * Extend relation by multiple blocks.
874 : *
875 : * Tries to extend the relation by extend_by blocks. Depending on the
876 : * availability of resources the relation may end up being extended by a
877 : * smaller number of pages (unless an error is thrown, always by at least one
878 : * page). *extended_by is updated to the number of pages the relation has been
879 : * extended to.
880 : *
881 : * buffers needs to be an array that is at least extend_by long. Upon
882 : * completion, the first extend_by array elements will point to a pinned
883 : * buffer.
884 : *
885 : * If EB_LOCK_FIRST is part of flags, the first returned buffer is
886 : * locked. This is useful for callers that want a buffer that is guaranteed to
887 : * be empty.
888 : */
889 : BlockNumber
890 317774 : ExtendBufferedRelBy(BufferManagerRelation bmr,
891 : ForkNumber fork,
892 : BufferAccessStrategy strategy,
893 : uint32 flags,
894 : uint32 extend_by,
895 : Buffer *buffers,
896 : uint32 *extended_by)
897 : {
898 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
899 : Assert(bmr.smgr == NULL || bmr.relpersistence != 0);
900 : Assert(extend_by > 0);
901 :
902 317774 : if (bmr.smgr == NULL)
903 : {
904 317774 : bmr.smgr = RelationGetSmgr(bmr.rel);
905 317774 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
906 : }
907 :
908 317774 : return ExtendBufferedRelCommon(bmr, fork, strategy, flags,
909 : extend_by, InvalidBlockNumber,
910 : buffers, extended_by);
911 : }
912 :
913 : /*
914 : * Extend the relation so it is at least extend_to blocks large, return buffer
915 : * (extend_to - 1).
916 : *
917 : * This is useful for callers that want to write a specific page, regardless
918 : * of the current size of the relation (e.g. useful for visibilitymap and for
919 : * crash recovery).
920 : */
921 : Buffer
922 102528 : ExtendBufferedRelTo(BufferManagerRelation bmr,
923 : ForkNumber fork,
924 : BufferAccessStrategy strategy,
925 : uint32 flags,
926 : BlockNumber extend_to,
927 : ReadBufferMode mode)
928 : {
929 : BlockNumber current_size;
930 102528 : uint32 extended_by = 0;
931 102528 : Buffer buffer = InvalidBuffer;
932 : Buffer buffers[64];
933 :
934 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
935 : Assert(bmr.smgr == NULL || bmr.relpersistence != 0);
936 : Assert(extend_to != InvalidBlockNumber && extend_to > 0);
937 :
938 102528 : if (bmr.smgr == NULL)
939 : {
940 14024 : bmr.smgr = RelationGetSmgr(bmr.rel);
941 14024 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
942 : }
943 :
944 : /*
945 : * If desired, create the file if it doesn't exist. If
946 : * smgr_cached_nblocks[fork] is positive then it must exist, no need for
947 : * an smgrexists call.
948 : */
949 102528 : if ((flags & EB_CREATE_FORK_IF_NEEDED) &&
950 14024 : (bmr.smgr->smgr_cached_nblocks[fork] == 0 ||
951 34 : bmr.smgr->smgr_cached_nblocks[fork] == InvalidBlockNumber) &&
952 13990 : !smgrexists(bmr.smgr, fork))
953 : {
954 13964 : LockRelationForExtension(bmr.rel, ExclusiveLock);
955 :
956 : /* recheck, fork might have been created concurrently */
957 13964 : if (!smgrexists(bmr.smgr, fork))
958 13960 : smgrcreate(bmr.smgr, fork, flags & EB_PERFORMING_RECOVERY);
959 :
960 13964 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
961 : }
962 :
963 : /*
964 : * If requested, invalidate size cache, so that smgrnblocks asks the
965 : * kernel.
966 : */
967 102528 : if (flags & EB_CLEAR_SIZE_CACHE)
968 14024 : bmr.smgr->smgr_cached_nblocks[fork] = InvalidBlockNumber;
969 :
970 : /*
971 : * Estimate how many pages we'll need to extend by. This avoids acquiring
972 : * unnecessarily many victim buffers.
973 : */
974 102528 : current_size = smgrnblocks(bmr.smgr, fork);
975 :
976 : /*
977 : * Since no-one else can be looking at the page contents yet, there is no
978 : * difference between an exclusive lock and a cleanup-strength lock. Note
979 : * that we pass the original mode to ReadBuffer_common() below, when
980 : * falling back to reading the buffer to a concurrent relation extension.
981 : */
982 102528 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
983 87772 : flags |= EB_LOCK_TARGET;
984 :
985 209042 : while (current_size < extend_to)
986 : {
987 106514 : uint32 num_pages = lengthof(buffers);
988 : BlockNumber first_block;
989 :
990 106514 : if ((uint64) current_size + num_pages > extend_to)
991 106512 : num_pages = extend_to - current_size;
992 :
993 106514 : first_block = ExtendBufferedRelCommon(bmr, fork, strategy, flags,
994 : num_pages, extend_to,
995 : buffers, &extended_by);
996 :
997 106514 : current_size = first_block + extended_by;
998 : Assert(num_pages != 0 || current_size >= extend_to);
999 :
1000 227342 : for (uint32 i = 0; i < extended_by; i++)
1001 : {
1002 120828 : if (first_block + i != extend_to - 1)
1003 18314 : ReleaseBuffer(buffers[i]);
1004 : else
1005 102514 : buffer = buffers[i];
1006 : }
1007 : }
1008 :
1009 : /*
1010 : * It's possible that another backend concurrently extended the relation.
1011 : * In that case read the buffer.
1012 : *
1013 : * XXX: Should we control this via a flag?
1014 : */
1015 102528 : if (buffer == InvalidBuffer)
1016 : {
1017 : Assert(extended_by == 0);
1018 14 : buffer = ReadBuffer_common(bmr.rel, bmr.smgr, bmr.relpersistence,
1019 : fork, extend_to - 1, mode, strategy);
1020 : }
1021 :
1022 102528 : return buffer;
1023 : }
1024 :
1025 : /*
1026 : * Lock and optionally zero a buffer, as part of the implementation of
1027 : * RBM_ZERO_AND_LOCK or RBM_ZERO_AND_CLEANUP_LOCK. The buffer must be already
1028 : * pinned. If the buffer is not already valid, it is zeroed and made valid.
1029 : */
1030 : static void
1031 631866 : ZeroAndLockBuffer(Buffer buffer, ReadBufferMode mode, bool already_valid)
1032 : {
1033 : BufferDesc *bufHdr;
1034 : bool need_to_zero;
1035 631866 : bool isLocalBuf = BufferIsLocal(buffer);
1036 :
1037 : Assert(mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK);
1038 :
1039 631866 : if (already_valid)
1040 : {
1041 : /*
1042 : * If the caller already knew the buffer was valid, we can skip some
1043 : * header interaction. The caller just wants to lock the buffer.
1044 : */
1045 74336 : need_to_zero = false;
1046 : }
1047 557530 : else if (isLocalBuf)
1048 : {
1049 : /* Simple case for non-shared buffers. */
1050 48 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
1051 48 : need_to_zero = StartLocalBufferIO(bufHdr, true, false);
1052 : }
1053 : else
1054 : {
1055 : /*
1056 : * Take BM_IO_IN_PROGRESS, or discover that BM_VALID has been set
1057 : * concurrently. Even though we aren't doing I/O, that ensures that
1058 : * we don't zero a page that someone else has pinned. An exclusive
1059 : * content lock wouldn't be enough, because readers are allowed to
1060 : * drop the content lock after determining that a tuple is visible
1061 : * (see buffer access rules in README).
1062 : */
1063 557482 : bufHdr = GetBufferDescriptor(buffer - 1);
1064 557482 : need_to_zero = StartBufferIO(bufHdr, true, false);
1065 : }
1066 :
1067 631866 : if (need_to_zero)
1068 : {
1069 557530 : memset(BufferGetPage(buffer), 0, BLCKSZ);
1070 :
1071 : /*
1072 : * Grab the buffer content lock before marking the page as valid, to
1073 : * make sure that no other backend sees the zeroed page before the
1074 : * caller has had a chance to initialize it.
1075 : *
1076 : * Since no-one else can be looking at the page contents yet, there is
1077 : * no difference between an exclusive lock and a cleanup-strength
1078 : * lock. (Note that we cannot use LockBuffer() or
1079 : * LockBufferForCleanup() here, because they assert that the buffer is
1080 : * already valid.)
1081 : */
1082 557530 : if (!isLocalBuf)
1083 557482 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_EXCLUSIVE);
1084 :
1085 : /* Set BM_VALID, terminate IO, and wake up any waiters */
1086 557530 : if (isLocalBuf)
1087 48 : TerminateLocalBufferIO(bufHdr, false, BM_VALID, false);
1088 : else
1089 557482 : TerminateBufferIO(bufHdr, false, BM_VALID, true, false);
1090 : }
1091 74336 : else if (!isLocalBuf)
1092 : {
1093 : /*
1094 : * The buffer is valid, so we can't zero it. The caller still expects
1095 : * the page to be locked on return.
1096 : */
1097 74296 : if (mode == RBM_ZERO_AND_LOCK)
1098 74198 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1099 : else
1100 98 : LockBufferForCleanup(buffer);
1101 : }
1102 631866 : }
1103 :
1104 : /*
1105 : * Pin a buffer for a given block. *foundPtr is set to true if the block was
1106 : * already present, or false if more work is required to either read it in or
1107 : * zero it.
1108 : */
1109 : static pg_attribute_always_inline Buffer
1110 123859044 : PinBufferForBlock(Relation rel,
1111 : SMgrRelation smgr,
1112 : char persistence,
1113 : ForkNumber forkNum,
1114 : BlockNumber blockNum,
1115 : BufferAccessStrategy strategy,
1116 : bool *foundPtr)
1117 : {
1118 : BufferDesc *bufHdr;
1119 : IOContext io_context;
1120 : IOObject io_object;
1121 :
1122 : Assert(blockNum != P_NEW);
1123 :
1124 : /* Persistence should be set before */
1125 : Assert((persistence == RELPERSISTENCE_TEMP ||
1126 : persistence == RELPERSISTENCE_PERMANENT ||
1127 : persistence == RELPERSISTENCE_UNLOGGED));
1128 :
1129 123859044 : if (persistence == RELPERSISTENCE_TEMP)
1130 : {
1131 2554082 : io_context = IOCONTEXT_NORMAL;
1132 2554082 : io_object = IOOBJECT_TEMP_RELATION;
1133 : }
1134 : else
1135 : {
1136 121304962 : io_context = IOContextForStrategy(strategy);
1137 121304962 : io_object = IOOBJECT_RELATION;
1138 : }
1139 :
1140 : TRACE_POSTGRESQL_BUFFER_READ_START(forkNum, blockNum,
1141 : smgr->smgr_rlocator.locator.spcOid,
1142 : smgr->smgr_rlocator.locator.dbOid,
1143 : smgr->smgr_rlocator.locator.relNumber,
1144 : smgr->smgr_rlocator.backend);
1145 :
1146 123859044 : if (persistence == RELPERSISTENCE_TEMP)
1147 : {
1148 2554082 : bufHdr = LocalBufferAlloc(smgr, forkNum, blockNum, foundPtr);
1149 2554070 : if (*foundPtr)
1150 2537294 : pgBufferUsage.local_blks_hit++;
1151 : }
1152 : else
1153 : {
1154 121304962 : bufHdr = BufferAlloc(smgr, persistence, forkNum, blockNum,
1155 : strategy, foundPtr, io_context);
1156 121304962 : if (*foundPtr)
1157 117954104 : pgBufferUsage.shared_blks_hit++;
1158 : }
1159 123859032 : if (rel)
1160 : {
1161 : /*
1162 : * While pgBufferUsage's "read" counter isn't bumped unless we reach
1163 : * WaitReadBuffers() (so, not for hits, and not for buffers that are
1164 : * zeroed instead), the per-relation stats always count them.
1165 : */
1166 112052406 : pgstat_count_buffer_read(rel);
1167 112052406 : if (*foundPtr)
1168 109601116 : pgstat_count_buffer_hit(rel);
1169 : }
1170 123859032 : if (*foundPtr)
1171 : {
1172 120491398 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1173 120491398 : if (VacuumCostActive)
1174 6670936 : VacuumCostBalance += VacuumCostPageHit;
1175 :
1176 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forkNum, blockNum,
1177 : smgr->smgr_rlocator.locator.spcOid,
1178 : smgr->smgr_rlocator.locator.dbOid,
1179 : smgr->smgr_rlocator.locator.relNumber,
1180 : smgr->smgr_rlocator.backend,
1181 : true);
1182 : }
1183 :
1184 123859032 : return BufferDescriptorGetBuffer(bufHdr);
1185 : }
1186 :
1187 : /*
1188 : * ReadBuffer_common -- common logic for all ReadBuffer variants
1189 : *
1190 : * smgr is required, rel is optional unless using P_NEW.
1191 : */
1192 : static pg_attribute_always_inline Buffer
1193 115515366 : ReadBuffer_common(Relation rel, SMgrRelation smgr, char smgr_persistence,
1194 : ForkNumber forkNum,
1195 : BlockNumber blockNum, ReadBufferMode mode,
1196 : BufferAccessStrategy strategy)
1197 : {
1198 : ReadBuffersOperation operation;
1199 : Buffer buffer;
1200 : int flags;
1201 : char persistence;
1202 :
1203 : /*
1204 : * Backward compatibility path, most code should use ExtendBufferedRel()
1205 : * instead, as acquiring the extension lock inside ExtendBufferedRel()
1206 : * scales a lot better.
1207 : */
1208 115515366 : if (unlikely(blockNum == P_NEW))
1209 : {
1210 520 : uint32 flags = EB_SKIP_EXTENSION_LOCK;
1211 :
1212 : /*
1213 : * Since no-one else can be looking at the page contents yet, there is
1214 : * no difference between an exclusive lock and a cleanup-strength
1215 : * lock.
1216 : */
1217 520 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1218 0 : flags |= EB_LOCK_FIRST;
1219 :
1220 520 : return ExtendBufferedRel(BMR_REL(rel), forkNum, strategy, flags);
1221 : }
1222 :
1223 115514846 : if (rel)
1224 104170936 : persistence = rel->rd_rel->relpersistence;
1225 : else
1226 11343910 : persistence = smgr_persistence;
1227 :
1228 115514846 : if (unlikely(mode == RBM_ZERO_AND_CLEANUP_LOCK ||
1229 : mode == RBM_ZERO_AND_LOCK))
1230 : {
1231 : bool found;
1232 :
1233 631866 : buffer = PinBufferForBlock(rel, smgr, persistence,
1234 : forkNum, blockNum, strategy, &found);
1235 631866 : ZeroAndLockBuffer(buffer, mode, found);
1236 631866 : return buffer;
1237 : }
1238 :
1239 : /*
1240 : * Signal that we are going to immediately wait. If we're immediately
1241 : * waiting, there is no benefit in actually executing the IO
1242 : * asynchronously, it would just add dispatch overhead.
1243 : */
1244 114882980 : flags = READ_BUFFERS_SYNCHRONOUSLY;
1245 114882980 : if (mode == RBM_ZERO_ON_ERROR)
1246 3199628 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1247 114882980 : operation.smgr = smgr;
1248 114882980 : operation.rel = rel;
1249 114882980 : operation.persistence = persistence;
1250 114882980 : operation.forknum = forkNum;
1251 114882980 : operation.strategy = strategy;
1252 114882980 : if (StartReadBuffer(&operation,
1253 : &buffer,
1254 : blockNum,
1255 : flags))
1256 1374962 : WaitReadBuffers(&operation);
1257 :
1258 114882934 : return buffer;
1259 : }
1260 :
1261 : static pg_attribute_always_inline bool
1262 122897672 : StartReadBuffersImpl(ReadBuffersOperation *operation,
1263 : Buffer *buffers,
1264 : BlockNumber blockNum,
1265 : int *nblocks,
1266 : int flags,
1267 : bool allow_forwarding)
1268 : {
1269 122897672 : int actual_nblocks = *nblocks;
1270 122897672 : int maxcombine = 0;
1271 : bool did_start_io;
1272 :
1273 : Assert(*nblocks == 1 || allow_forwarding);
1274 : Assert(*nblocks > 0);
1275 : Assert(*nblocks <= MAX_IO_COMBINE_LIMIT);
1276 :
1277 125707776 : for (int i = 0; i < actual_nblocks; ++i)
1278 : {
1279 : bool found;
1280 :
1281 123230018 : if (allow_forwarding && buffers[i] != InvalidBuffer)
1282 2840 : {
1283 : BufferDesc *bufHdr;
1284 :
1285 : /*
1286 : * This is a buffer that was pinned by an earlier call to
1287 : * StartReadBuffers(), but couldn't be handled in one operation at
1288 : * that time. The operation was split, and the caller has passed
1289 : * an already pinned buffer back to us to handle the rest of the
1290 : * operation. It must continue at the expected block number.
1291 : */
1292 : Assert(BufferGetBlockNumber(buffers[i]) == blockNum + i);
1293 :
1294 : /*
1295 : * It might be an already valid buffer (a hit) that followed the
1296 : * final contiguous block of an earlier I/O (a miss) marking the
1297 : * end of it, or a buffer that some other backend has since made
1298 : * valid by performing the I/O for us, in which case we can handle
1299 : * it as a hit now. It is safe to check for a BM_VALID flag with
1300 : * a relaxed load, because we got a fresh view of it while pinning
1301 : * it in the previous call.
1302 : *
1303 : * On the other hand if we don't see BM_VALID yet, it must be an
1304 : * I/O that was split by the previous call and we need to try to
1305 : * start a new I/O from this block. We're also racing against any
1306 : * other backend that might start the I/O or even manage to mark
1307 : * it BM_VALID after this check, but StartBufferIO() will handle
1308 : * those cases.
1309 : */
1310 2840 : if (BufferIsLocal(buffers[i]))
1311 4 : bufHdr = GetLocalBufferDescriptor(-buffers[i] - 1);
1312 : else
1313 2836 : bufHdr = GetBufferDescriptor(buffers[i] - 1);
1314 : Assert(pg_atomic_read_u32(&bufHdr->state) & BM_TAG_VALID);
1315 2840 : found = pg_atomic_read_u32(&bufHdr->state) & BM_VALID;
1316 : }
1317 : else
1318 : {
1319 123227166 : buffers[i] = PinBufferForBlock(operation->rel,
1320 123227178 : operation->smgr,
1321 123227178 : operation->persistence,
1322 : operation->forknum,
1323 : blockNum + i,
1324 : operation->strategy,
1325 : &found);
1326 : }
1327 :
1328 123230006 : if (found)
1329 : {
1330 : /*
1331 : * We have a hit. If it's the first block in the requested range,
1332 : * we can return it immediately and report that WaitReadBuffers()
1333 : * does not need to be called. If the initial value of *nblocks
1334 : * was larger, the caller will have to call again for the rest.
1335 : */
1336 120419902 : if (i == 0)
1337 : {
1338 120417058 : *nblocks = 1;
1339 :
1340 : #ifdef USE_ASSERT_CHECKING
1341 :
1342 : /*
1343 : * Initialize enough of ReadBuffersOperation to make
1344 : * CheckReadBuffersOperation() work. Outside of assertions
1345 : * that's not necessary when no IO is issued.
1346 : */
1347 : operation->buffers = buffers;
1348 : operation->blocknum = blockNum;
1349 : operation->nblocks = 1;
1350 : operation->nblocks_done = 1;
1351 : CheckReadBuffersOperation(operation, true);
1352 : #endif
1353 120417058 : return false;
1354 : }
1355 :
1356 : /*
1357 : * Otherwise we already have an I/O to perform, but this block
1358 : * can't be included as it is already valid. Split the I/O here.
1359 : * There may or may not be more blocks requiring I/O after this
1360 : * one, we haven't checked, but they can't be contiguous with this
1361 : * one in the way. We'll leave this buffer pinned, forwarding it
1362 : * to the next call, avoiding the need to unpin it here and re-pin
1363 : * it in the next call.
1364 : */
1365 2844 : actual_nblocks = i;
1366 2844 : break;
1367 : }
1368 : else
1369 : {
1370 : /*
1371 : * Check how many blocks we can cover with the same IO. The smgr
1372 : * implementation might e.g. be limited due to a segment boundary.
1373 : */
1374 2810104 : if (i == 0 && actual_nblocks > 1)
1375 : {
1376 66798 : maxcombine = smgrmaxcombine(operation->smgr,
1377 : operation->forknum,
1378 : blockNum);
1379 66798 : if (unlikely(maxcombine < actual_nblocks))
1380 : {
1381 0 : elog(DEBUG2, "limiting nblocks at %u from %u to %u",
1382 : blockNum, actual_nblocks, maxcombine);
1383 0 : actual_nblocks = maxcombine;
1384 : }
1385 : }
1386 : }
1387 : }
1388 2480602 : *nblocks = actual_nblocks;
1389 :
1390 : /* Populate information needed for I/O. */
1391 2480602 : operation->buffers = buffers;
1392 2480602 : operation->blocknum = blockNum;
1393 2480602 : operation->flags = flags;
1394 2480602 : operation->nblocks = actual_nblocks;
1395 2480602 : operation->nblocks_done = 0;
1396 2480602 : pgaio_wref_clear(&operation->io_wref);
1397 :
1398 : /*
1399 : * When using AIO, start the IO in the background. If not, issue prefetch
1400 : * requests if desired by the caller.
1401 : *
1402 : * The reason we have a dedicated path for IOMETHOD_SYNC here is to
1403 : * de-risk the introduction of AIO somewhat. It's a large architectural
1404 : * change, with lots of chances for unanticipated performance effects.
1405 : *
1406 : * Use of IOMETHOD_SYNC already leads to not actually performing IO
1407 : * asynchronously, but without the check here we'd execute IO earlier than
1408 : * we used to. Eventually this IOMETHOD_SYNC specific path should go away.
1409 : */
1410 2480602 : if (io_method != IOMETHOD_SYNC)
1411 : {
1412 : /*
1413 : * Try to start IO asynchronously. It's possible that no IO needs to
1414 : * be started, if another backend already performed the IO.
1415 : *
1416 : * Note that if an IO is started, it might not cover the entire
1417 : * requested range, e.g. because an intermediary block has been read
1418 : * in by another backend. In that case any "trailing" buffers we
1419 : * already pinned above will be "forwarded" by read_stream.c to the
1420 : * next call to StartReadBuffers().
1421 : *
1422 : * This is signalled to the caller by decrementing *nblocks *and*
1423 : * reducing operation->nblocks. The latter is done here, but not below
1424 : * WaitReadBuffers(), as in WaitReadBuffers() we can't "shorten" the
1425 : * overall read size anymore, we need to retry until done in its
1426 : * entirety or until failed.
1427 : */
1428 2478442 : did_start_io = AsyncReadBuffers(operation, nblocks);
1429 :
1430 2478412 : operation->nblocks = *nblocks;
1431 : }
1432 : else
1433 : {
1434 2160 : operation->flags |= READ_BUFFERS_SYNCHRONOUSLY;
1435 :
1436 2160 : if (flags & READ_BUFFERS_ISSUE_ADVICE)
1437 : {
1438 : /*
1439 : * In theory we should only do this if PinBufferForBlock() had to
1440 : * allocate new buffers above. That way, if two calls to
1441 : * StartReadBuffers() were made for the same blocks before
1442 : * WaitReadBuffers(), only the first would issue the advice.
1443 : * That'd be a better simulation of true asynchronous I/O, which
1444 : * would only start the I/O once, but isn't done here for
1445 : * simplicity.
1446 : */
1447 4 : smgrprefetch(operation->smgr,
1448 : operation->forknum,
1449 : blockNum,
1450 : actual_nblocks);
1451 : }
1452 :
1453 : /*
1454 : * Indicate that WaitReadBuffers() should be called. WaitReadBuffers()
1455 : * will initiate the necessary IO.
1456 : */
1457 2160 : did_start_io = true;
1458 : }
1459 :
1460 2480572 : CheckReadBuffersOperation(operation, !did_start_io);
1461 :
1462 2480572 : return did_start_io;
1463 : }
1464 :
1465 : /*
1466 : * Begin reading a range of blocks beginning at blockNum and extending for
1467 : * *nblocks. *nblocks and the buffers array are in/out parameters. On entry,
1468 : * the buffers elements covered by *nblocks must hold either InvalidBuffer or
1469 : * buffers forwarded by an earlier call to StartReadBuffers() that was split
1470 : * and is now being continued. On return, *nblocks holds the number of blocks
1471 : * accepted by this operation. If it is less than the original number then
1472 : * this operation has been split, but buffer elements up to the original
1473 : * requested size may hold forwarded buffers to be used for a continuing
1474 : * operation. The caller must either start a new I/O beginning at the block
1475 : * immediately following the blocks accepted by this call and pass those
1476 : * buffers back in, or release them if it chooses not to. It shouldn't make
1477 : * any other use of or assumptions about forwarded buffers.
1478 : *
1479 : * If false is returned, no I/O is necessary and the buffers covered by
1480 : * *nblocks on exit are valid and ready to be accessed. If true is returned,
1481 : * an I/O has been started, and WaitReadBuffers() must be called with the same
1482 : * operation object before the buffers covered by *nblocks on exit can be
1483 : * accessed. Along with the operation object, the caller-supplied array of
1484 : * buffers must remain valid until WaitReadBuffers() is called, and any
1485 : * forwarded buffers must also be preserved for a continuing call unless
1486 : * they are explicitly released.
1487 : */
1488 : bool
1489 3990428 : StartReadBuffers(ReadBuffersOperation *operation,
1490 : Buffer *buffers,
1491 : BlockNumber blockNum,
1492 : int *nblocks,
1493 : int flags)
1494 : {
1495 3990428 : return StartReadBuffersImpl(operation, buffers, blockNum, nblocks, flags,
1496 : true /* expect forwarded buffers */ );
1497 : }
1498 :
1499 : /*
1500 : * Single block version of the StartReadBuffers(). This might save a few
1501 : * instructions when called from another translation unit, because it is
1502 : * specialized for nblocks == 1.
1503 : *
1504 : * This version does not support "forwarded" buffers: they cannot be created
1505 : * by reading only one block and *buffer is ignored on entry.
1506 : */
1507 : bool
1508 118907244 : StartReadBuffer(ReadBuffersOperation *operation,
1509 : Buffer *buffer,
1510 : BlockNumber blocknum,
1511 : int flags)
1512 : {
1513 118907244 : int nblocks = 1;
1514 : bool result;
1515 :
1516 118907244 : result = StartReadBuffersImpl(operation, buffer, blocknum, &nblocks, flags,
1517 : false /* single block, no forwarding */ );
1518 : Assert(nblocks == 1); /* single block can't be short */
1519 :
1520 118907214 : return result;
1521 : }
1522 :
1523 : /*
1524 : * Perform sanity checks on the ReadBuffersOperation.
1525 : */
1526 : static void
1527 7409898 : CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete)
1528 : {
1529 : #ifdef USE_ASSERT_CHECKING
1530 : Assert(operation->nblocks_done <= operation->nblocks);
1531 : Assert(!is_complete || operation->nblocks == operation->nblocks_done);
1532 :
1533 : for (int i = 0; i < operation->nblocks; i++)
1534 : {
1535 : Buffer buffer = operation->buffers[i];
1536 : BufferDesc *buf_hdr = BufferIsLocal(buffer) ?
1537 : GetLocalBufferDescriptor(-buffer - 1) :
1538 : GetBufferDescriptor(buffer - 1);
1539 :
1540 : Assert(BufferGetBlockNumber(buffer) == operation->blocknum + i);
1541 : Assert(pg_atomic_read_u32(&buf_hdr->state) & BM_TAG_VALID);
1542 :
1543 : if (i < operation->nblocks_done)
1544 : Assert(pg_atomic_read_u32(&buf_hdr->state) & BM_VALID);
1545 : }
1546 : #endif
1547 7409898 : }
1548 :
1549 : /* helper for ReadBuffersCanStartIO(), to avoid repetition */
1550 : static inline bool
1551 2810140 : ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
1552 : {
1553 2810140 : if (BufferIsLocal(buffer))
1554 16728 : return StartLocalBufferIO(GetLocalBufferDescriptor(-buffer - 1),
1555 : true, nowait);
1556 : else
1557 2793412 : return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
1558 : }
1559 :
1560 : /*
1561 : * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
1562 : */
1563 : static inline bool
1564 2810140 : ReadBuffersCanStartIO(Buffer buffer, bool nowait)
1565 : {
1566 : /*
1567 : * If this backend currently has staged IO, we need to submit the pending
1568 : * IO before waiting for the right to issue IO, to avoid the potential for
1569 : * deadlocks (and, more commonly, unnecessary delays for other backends).
1570 : */
1571 2810140 : if (!nowait && pgaio_have_staged())
1572 : {
1573 1140 : if (ReadBuffersCanStartIOOnce(buffer, true))
1574 1140 : return true;
1575 :
1576 : /*
1577 : * Unfortunately StartBufferIO() returning false doesn't allow to
1578 : * distinguish between the buffer already being valid and IO already
1579 : * being in progress. Since IO already being in progress is quite
1580 : * rare, this approach seems fine.
1581 : */
1582 0 : pgaio_submit_staged();
1583 : }
1584 :
1585 2809000 : return ReadBuffersCanStartIOOnce(buffer, nowait);
1586 : }
1587 :
1588 : /*
1589 : * Helper for WaitReadBuffers() that processes the results of a readv
1590 : * operation, raising an error if necessary.
1591 : */
1592 : static void
1593 2463622 : ProcessReadBuffersResult(ReadBuffersOperation *operation)
1594 : {
1595 2463622 : PgAioReturn *aio_ret = &operation->io_return;
1596 2463622 : PgAioResultStatus rs = aio_ret->result.status;
1597 2463622 : int newly_read_blocks = 0;
1598 :
1599 : Assert(pgaio_wref_valid(&operation->io_wref));
1600 : Assert(aio_ret->result.status != PGAIO_RS_UNKNOWN);
1601 :
1602 : /*
1603 : * SMGR reports the number of blocks successfully read as the result of
1604 : * the IO operation. Thus we can simply add that to ->nblocks_done.
1605 : */
1606 :
1607 2463622 : if (likely(rs != PGAIO_RS_ERROR))
1608 2463564 : newly_read_blocks = aio_ret->result.result;
1609 :
1610 2463622 : if (rs == PGAIO_RS_ERROR || rs == PGAIO_RS_WARNING)
1611 90 : pgaio_result_report(aio_ret->result, &aio_ret->target_data,
1612 : rs == PGAIO_RS_ERROR ? ERROR : WARNING);
1613 2463532 : else if (aio_ret->result.status == PGAIO_RS_PARTIAL)
1614 : {
1615 : /*
1616 : * We'll retry, so we just emit a debug message to the server log (or
1617 : * not even that in prod scenarios).
1618 : */
1619 20 : pgaio_result_report(aio_ret->result, &aio_ret->target_data, DEBUG1);
1620 20 : elog(DEBUG3, "partial read, will retry");
1621 : }
1622 :
1623 : Assert(newly_read_blocks > 0);
1624 : Assert(newly_read_blocks <= MAX_IO_COMBINE_LIMIT);
1625 :
1626 2463564 : operation->nblocks_done += newly_read_blocks;
1627 :
1628 : Assert(operation->nblocks_done <= operation->nblocks);
1629 2463564 : }
1630 :
1631 : void
1632 2463602 : WaitReadBuffers(ReadBuffersOperation *operation)
1633 : {
1634 2463602 : PgAioReturn *aio_ret = &operation->io_return;
1635 : IOContext io_context;
1636 : IOObject io_object;
1637 :
1638 2463602 : if (operation->persistence == RELPERSISTENCE_TEMP)
1639 : {
1640 2974 : io_context = IOCONTEXT_NORMAL;
1641 2974 : io_object = IOOBJECT_TEMP_RELATION;
1642 : }
1643 : else
1644 : {
1645 2460628 : io_context = IOContextForStrategy(operation->strategy);
1646 2460628 : io_object = IOOBJECT_RELATION;
1647 : }
1648 :
1649 : /*
1650 : * If we get here without an IO operation having been issued, the
1651 : * io_method == IOMETHOD_SYNC path must have been used. Otherwise the
1652 : * caller should not have called WaitReadBuffers().
1653 : *
1654 : * In the case of IOMETHOD_SYNC, we start - as we used to before the
1655 : * introducing of AIO - the IO in WaitReadBuffers(). This is done as part
1656 : * of the retry logic below, no extra code is required.
1657 : *
1658 : * This path is expected to eventually go away.
1659 : */
1660 2463602 : if (!pgaio_wref_valid(&operation->io_wref) && io_method != IOMETHOD_SYNC)
1661 0 : elog(ERROR, "waiting for read operation that didn't read");
1662 :
1663 : /*
1664 : * To handle partial reads, and IOMETHOD_SYNC, we re-issue IO until we're
1665 : * done. We may need multiple retries, not just because we could get
1666 : * multiple partial reads, but also because some of the remaining
1667 : * to-be-read buffers may have been read in by other backends, limiting
1668 : * the IO size.
1669 : */
1670 : while (true)
1671 2180 : {
1672 : int ignored_nblocks_progress;
1673 :
1674 2465782 : CheckReadBuffersOperation(operation, false);
1675 :
1676 : /*
1677 : * If there is an IO associated with the operation, we may need to
1678 : * wait for it.
1679 : */
1680 2465782 : if (pgaio_wref_valid(&operation->io_wref))
1681 : {
1682 : /*
1683 : * Track the time spent waiting for the IO to complete. As
1684 : * tracking a wait even if we don't actually need to wait
1685 : *
1686 : * a) is not cheap, due to the timestamping overhead
1687 : *
1688 : * b) reports some time as waiting, even if we never waited
1689 : *
1690 : * we first check if we already know the IO is complete.
1691 : */
1692 2463622 : if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
1693 1072484 : !pgaio_wref_check_done(&operation->io_wref))
1694 : {
1695 333288 : instr_time io_start = pgstat_prepare_io_time(track_io_timing);
1696 :
1697 333288 : pgaio_wref_wait(&operation->io_wref);
1698 :
1699 : /*
1700 : * The IO operation itself was already counted earlier, in
1701 : * AsyncReadBuffers(), this just accounts for the wait time.
1702 : */
1703 333288 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1704 : io_start, 0, 0);
1705 : }
1706 : else
1707 : {
1708 : Assert(pgaio_wref_check_done(&operation->io_wref));
1709 : }
1710 :
1711 : /*
1712 : * We now are sure the IO completed. Check the results. This
1713 : * includes reporting on errors if there were any.
1714 : */
1715 2463622 : ProcessReadBuffersResult(operation);
1716 : }
1717 :
1718 : /*
1719 : * Most of the time, the one IO we already started, will read in
1720 : * everything. But we need to deal with partial reads and buffers not
1721 : * needing IO anymore.
1722 : */
1723 2465724 : if (operation->nblocks_done == operation->nblocks)
1724 2463544 : break;
1725 :
1726 2180 : CHECK_FOR_INTERRUPTS();
1727 :
1728 : /*
1729 : * This may only complete the IO partially, either because some
1730 : * buffers were already valid, or because of a partial read.
1731 : *
1732 : * NB: In contrast to after the AsyncReadBuffers() call in
1733 : * StartReadBuffers(), we do *not* reduce
1734 : * ReadBuffersOperation->nblocks here, callers expect the full
1735 : * operation to be completed at this point (as more operations may
1736 : * have been queued).
1737 : */
1738 2180 : AsyncReadBuffers(operation, &ignored_nblocks_progress);
1739 : }
1740 :
1741 2463544 : CheckReadBuffersOperation(operation, true);
1742 :
1743 : /* NB: READ_DONE tracepoint was already executed in completion callback */
1744 2463544 : }
1745 :
1746 : /*
1747 : * Initiate IO for the ReadBuffersOperation
1748 : *
1749 : * This function only starts a single IO at a time. The size of the IO may be
1750 : * limited to below the to-be-read blocks, if one of the buffers has
1751 : * concurrently been read in. If the first to-be-read buffer is already valid,
1752 : * no IO will be issued.
1753 : *
1754 : * To support retries after partial reads, the first operation->nblocks_done
1755 : * buffers are skipped.
1756 : *
1757 : * On return *nblocks_progress is updated to reflect the number of buffers
1758 : * affected by the call. If the first buffer is valid, *nblocks_progress is
1759 : * set to 1 and operation->nblocks_done is incremented.
1760 : *
1761 : * Returns true if IO was initiated, false if no IO was necessary.
1762 : */
1763 : static bool
1764 2480622 : AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
1765 : {
1766 2480622 : Buffer *buffers = &operation->buffers[0];
1767 2480622 : int flags = operation->flags;
1768 2480622 : BlockNumber blocknum = operation->blocknum;
1769 2480622 : ForkNumber forknum = operation->forknum;
1770 2480622 : char persistence = operation->persistence;
1771 2480622 : int16 nblocks_done = operation->nblocks_done;
1772 2480622 : Buffer *io_buffers = &operation->buffers[nblocks_done];
1773 2480622 : int io_buffers_len = 0;
1774 : PgAioHandle *ioh;
1775 2480622 : uint32 ioh_flags = 0;
1776 : void *io_pages[MAX_IO_COMBINE_LIMIT];
1777 : IOContext io_context;
1778 : IOObject io_object;
1779 : bool did_start_io;
1780 :
1781 : /*
1782 : * When this IO is executed synchronously, either because the caller will
1783 : * immediately block waiting for the IO or because IOMETHOD_SYNC is used,
1784 : * the AIO subsystem needs to know.
1785 : */
1786 2480622 : if (flags & READ_BUFFERS_SYNCHRONOUSLY)
1787 1385794 : ioh_flags |= PGAIO_HF_SYNCHRONOUS;
1788 :
1789 2480622 : if (persistence == RELPERSISTENCE_TEMP)
1790 : {
1791 3562 : io_context = IOCONTEXT_NORMAL;
1792 3562 : io_object = IOOBJECT_TEMP_RELATION;
1793 3562 : ioh_flags |= PGAIO_HF_REFERENCES_LOCAL;
1794 : }
1795 : else
1796 : {
1797 2477060 : io_context = IOContextForStrategy(operation->strategy);
1798 2477060 : io_object = IOOBJECT_RELATION;
1799 : }
1800 :
1801 : /*
1802 : * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
1803 : * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
1804 : * set globally, but on a per-session basis. The completion callback,
1805 : * which may be run in other processes, e.g. in IO workers, may have a
1806 : * different value of the zero_damaged_pages GUC.
1807 : *
1808 : * XXX: We probably should eventually use a different flag for
1809 : * zero_damaged_pages, so we can report different log levels / error codes
1810 : * for zero_damaged_pages and ZERO_ON_ERROR.
1811 : */
1812 2480622 : if (zero_damaged_pages)
1813 32 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1814 :
1815 : /*
1816 : * For the same reason as with zero_damaged_pages we need to use this
1817 : * backend's ignore_checksum_failure value.
1818 : */
1819 2480622 : if (ignore_checksum_failure)
1820 16 : flags |= READ_BUFFERS_IGNORE_CHECKSUM_FAILURES;
1821 :
1822 :
1823 : /*
1824 : * To be allowed to report stats in the local completion callback we need
1825 : * to prepare to report stats now. This ensures we can safely report the
1826 : * checksum failure even in a critical section.
1827 : */
1828 2480622 : pgstat_prepare_report_checksum_failure(operation->smgr->smgr_rlocator.locator.dbOid);
1829 :
1830 : /*
1831 : * Get IO handle before ReadBuffersCanStartIO(), as pgaio_io_acquire()
1832 : * might block, which we don't want after setting IO_IN_PROGRESS.
1833 : *
1834 : * If we need to wait for IO before we can get a handle, submit
1835 : * already-staged IO first, so that other backends don't need to wait.
1836 : * There wouldn't be a deadlock risk, as pgaio_io_acquire() just needs to
1837 : * wait for already submitted IO, which doesn't require additional locks,
1838 : * but it could still cause undesirable waits.
1839 : *
1840 : * A secondary benefit is that this would allow us to measure the time in
1841 : * pgaio_io_acquire() without causing undue timer overhead in the common,
1842 : * non-blocking, case. However, currently the pgstats infrastructure
1843 : * doesn't really allow that, as it a) asserts that an operation can't
1844 : * have time without operations b) doesn't have an API to report
1845 : * "accumulated" time.
1846 : */
1847 2480622 : ioh = pgaio_io_acquire_nb(CurrentResourceOwner, &operation->io_return);
1848 2480622 : if (unlikely(!ioh))
1849 : {
1850 6706 : pgaio_submit_staged();
1851 :
1852 6706 : ioh = pgaio_io_acquire(CurrentResourceOwner, &operation->io_return);
1853 : }
1854 :
1855 : /*
1856 : * Check if we can start IO on the first to-be-read buffer.
1857 : *
1858 : * If an I/O is already in progress in another backend, we want to wait
1859 : * for the outcome: either done, or something went wrong and we will
1860 : * retry.
1861 : */
1862 2480622 : if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
1863 : {
1864 : /*
1865 : * Someone else has already completed this block, we're done.
1866 : *
1867 : * When IO is necessary, ->nblocks_done is updated in
1868 : * ProcessReadBuffersResult(), but that is not called if no IO is
1869 : * necessary. Thus update here.
1870 : */
1871 16376 : operation->nblocks_done += 1;
1872 16376 : *nblocks_progress = 1;
1873 :
1874 16376 : pgaio_io_release(ioh);
1875 16376 : pgaio_wref_clear(&operation->io_wref);
1876 16376 : did_start_io = false;
1877 :
1878 : /*
1879 : * Report and track this as a 'hit' for this backend, even though it
1880 : * must have started out as a miss in PinBufferForBlock(). The other
1881 : * backend will track this as a 'read'.
1882 : */
1883 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forknum, blocknum + operation->nblocks_done,
1884 : operation->smgr->smgr_rlocator.locator.spcOid,
1885 : operation->smgr->smgr_rlocator.locator.dbOid,
1886 : operation->smgr->smgr_rlocator.locator.relNumber,
1887 : operation->smgr->smgr_rlocator.backend,
1888 : true);
1889 :
1890 16376 : if (persistence == RELPERSISTENCE_TEMP)
1891 0 : pgBufferUsage.local_blks_hit += 1;
1892 : else
1893 16376 : pgBufferUsage.shared_blks_hit += 1;
1894 :
1895 16376 : if (operation->rel)
1896 16376 : pgstat_count_buffer_hit(operation->rel);
1897 :
1898 16376 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1899 :
1900 16376 : if (VacuumCostActive)
1901 22 : VacuumCostBalance += VacuumCostPageHit;
1902 : }
1903 : else
1904 : {
1905 : instr_time io_start;
1906 :
1907 : /* We found a buffer that we need to read in. */
1908 : Assert(io_buffers[0] == buffers[nblocks_done]);
1909 2464246 : io_pages[0] = BufferGetBlock(buffers[nblocks_done]);
1910 2464246 : io_buffers_len = 1;
1911 :
1912 : /*
1913 : * How many neighboring-on-disk blocks can we scatter-read into other
1914 : * buffers at the same time? In this case we don't wait if we see an
1915 : * I/O already in progress. We already set BM_IO_IN_PROGRESS for the
1916 : * head block, so we should get on with that I/O as soon as possible.
1917 : */
1918 2793764 : for (int i = nblocks_done + 1; i < operation->nblocks; i++)
1919 : {
1920 329518 : if (!ReadBuffersCanStartIO(buffers[i], true))
1921 0 : break;
1922 : /* Must be consecutive block numbers. */
1923 : Assert(BufferGetBlockNumber(buffers[i - 1]) ==
1924 : BufferGetBlockNumber(buffers[i]) - 1);
1925 : Assert(io_buffers[io_buffers_len] == buffers[i]);
1926 :
1927 329518 : io_pages[io_buffers_len++] = BufferGetBlock(buffers[i]);
1928 : }
1929 :
1930 : /* get a reference to wait for in WaitReadBuffers() */
1931 2464246 : pgaio_io_get_wref(ioh, &operation->io_wref);
1932 :
1933 : /* provide the list of buffers to the completion callbacks */
1934 2464246 : pgaio_io_set_handle_data_32(ioh, (uint32 *) io_buffers, io_buffers_len);
1935 :
1936 2464246 : pgaio_io_register_callbacks(ioh,
1937 : persistence == RELPERSISTENCE_TEMP ?
1938 : PGAIO_HCB_LOCAL_BUFFER_READV :
1939 : PGAIO_HCB_SHARED_BUFFER_READV,
1940 : flags);
1941 :
1942 2464246 : pgaio_io_set_flag(ioh, ioh_flags);
1943 :
1944 : /* ---
1945 : * Even though we're trying to issue IO asynchronously, track the time
1946 : * in smgrstartreadv():
1947 : * - if io_method == IOMETHOD_SYNC, we will always perform the IO
1948 : * immediately
1949 : * - the io method might not support the IO (e.g. worker IO for a temp
1950 : * table)
1951 : * ---
1952 : */
1953 2464246 : io_start = pgstat_prepare_io_time(track_io_timing);
1954 2464246 : smgrstartreadv(ioh, operation->smgr, forknum,
1955 : blocknum + nblocks_done,
1956 : io_pages, io_buffers_len);
1957 2464216 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1958 2464216 : io_start, 1, io_buffers_len * BLCKSZ);
1959 :
1960 2464216 : if (persistence == RELPERSISTENCE_TEMP)
1961 3562 : pgBufferUsage.local_blks_read += io_buffers_len;
1962 : else
1963 2460654 : pgBufferUsage.shared_blks_read += io_buffers_len;
1964 :
1965 : /*
1966 : * Track vacuum cost when issuing IO, not after waiting for it.
1967 : * Otherwise we could end up issuing a lot of IO in a short timespan,
1968 : * despite a low cost limit.
1969 : */
1970 2464216 : if (VacuumCostActive)
1971 48782 : VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
1972 :
1973 2464216 : *nblocks_progress = io_buffers_len;
1974 2464216 : did_start_io = true;
1975 : }
1976 :
1977 2480592 : return did_start_io;
1978 : }
1979 :
1980 : /*
1981 : * BufferAlloc -- subroutine for PinBufferForBlock. Handles lookup of a shared
1982 : * buffer. If no buffer exists already, selects a replacement victim and
1983 : * evicts the old page, but does NOT read in new page.
1984 : *
1985 : * "strategy" can be a buffer replacement strategy object, or NULL for
1986 : * the default strategy. The selected buffer's usage_count is advanced when
1987 : * using the default strategy, but otherwise possibly not (see PinBuffer).
1988 : *
1989 : * The returned buffer is pinned and is already marked as holding the
1990 : * desired page. If it already did have the desired page, *foundPtr is
1991 : * set true. Otherwise, *foundPtr is set false.
1992 : *
1993 : * io_context is passed as an output parameter to avoid calling
1994 : * IOContextForStrategy() when there is a shared buffers hit and no IO
1995 : * statistics need be captured.
1996 : *
1997 : * No locks are held either at entry or exit.
1998 : */
1999 : static pg_attribute_always_inline BufferDesc *
2000 121304962 : BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
2001 : BlockNumber blockNum,
2002 : BufferAccessStrategy strategy,
2003 : bool *foundPtr, IOContext io_context)
2004 : {
2005 : BufferTag newTag; /* identity of requested block */
2006 : uint32 newHash; /* hash value for newTag */
2007 : LWLock *newPartitionLock; /* buffer partition lock for it */
2008 : int existing_buf_id;
2009 : Buffer victim_buffer;
2010 : BufferDesc *victim_buf_hdr;
2011 : uint32 victim_buf_state;
2012 :
2013 : /* Make sure we will have room to remember the buffer pin */
2014 121304962 : ResourceOwnerEnlarge(CurrentResourceOwner);
2015 121304962 : ReservePrivateRefCountEntry();
2016 :
2017 : /* create a tag so we can lookup the buffer */
2018 121304962 : InitBufferTag(&newTag, &smgr->smgr_rlocator.locator, forkNum, blockNum);
2019 :
2020 : /* determine its hash code and partition lock ID */
2021 121304962 : newHash = BufTableHashCode(&newTag);
2022 121304962 : newPartitionLock = BufMappingPartitionLock(newHash);
2023 :
2024 : /* see if the block is in the buffer pool already */
2025 121304962 : LWLockAcquire(newPartitionLock, LW_SHARED);
2026 121304962 : existing_buf_id = BufTableLookup(&newTag, newHash);
2027 121304962 : if (existing_buf_id >= 0)
2028 : {
2029 : BufferDesc *buf;
2030 : bool valid;
2031 :
2032 : /*
2033 : * Found it. Now, pin the buffer so no one can steal it from the
2034 : * buffer pool, and check to see if the correct data has been loaded
2035 : * into the buffer.
2036 : */
2037 117967368 : buf = GetBufferDescriptor(existing_buf_id);
2038 :
2039 117967368 : valid = PinBuffer(buf, strategy);
2040 :
2041 : /* Can release the mapping lock as soon as we've pinned it */
2042 117967368 : LWLockRelease(newPartitionLock);
2043 :
2044 117967368 : *foundPtr = true;
2045 :
2046 117967368 : if (!valid)
2047 : {
2048 : /*
2049 : * We can only get here if (a) someone else is still reading in
2050 : * the page, (b) a previous read attempt failed, or (c) someone
2051 : * called StartReadBuffers() but not yet WaitReadBuffers().
2052 : */
2053 14062 : *foundPtr = false;
2054 : }
2055 :
2056 117967368 : return buf;
2057 : }
2058 :
2059 : /*
2060 : * Didn't find it in the buffer pool. We'll have to initialize a new
2061 : * buffer. Remember to unlock the mapping lock while doing the work.
2062 : */
2063 3337594 : LWLockRelease(newPartitionLock);
2064 :
2065 : /*
2066 : * Acquire a victim buffer. Somebody else might try to do the same, we
2067 : * don't hold any conflicting locks. If so we'll have to undo our work
2068 : * later.
2069 : */
2070 3337594 : victim_buffer = GetVictimBuffer(strategy, io_context);
2071 3337594 : victim_buf_hdr = GetBufferDescriptor(victim_buffer - 1);
2072 :
2073 : /*
2074 : * Try to make a hashtable entry for the buffer under its new tag. If
2075 : * somebody else inserted another buffer for the tag, we'll release the
2076 : * victim buffer we acquired and use the already inserted one.
2077 : */
2078 3337594 : LWLockAcquire(newPartitionLock, LW_EXCLUSIVE);
2079 3337594 : existing_buf_id = BufTableInsert(&newTag, newHash, victim_buf_hdr->buf_id);
2080 3337594 : if (existing_buf_id >= 0)
2081 : {
2082 : BufferDesc *existing_buf_hdr;
2083 : bool valid;
2084 :
2085 : /*
2086 : * Got a collision. Someone has already done what we were about to do.
2087 : * We'll just handle this as if it were found in the buffer pool in
2088 : * the first place. First, give up the buffer we were planning to
2089 : * use.
2090 : *
2091 : * We could do this after releasing the partition lock, but then we'd
2092 : * have to call ResourceOwnerEnlarge() & ReservePrivateRefCountEntry()
2093 : * before acquiring the lock, for the rare case of such a collision.
2094 : */
2095 3208 : UnpinBuffer(victim_buf_hdr);
2096 :
2097 : /* remaining code should match code at top of routine */
2098 :
2099 3208 : existing_buf_hdr = GetBufferDescriptor(existing_buf_id);
2100 :
2101 3208 : valid = PinBuffer(existing_buf_hdr, strategy);
2102 :
2103 : /* Can release the mapping lock as soon as we've pinned it */
2104 3208 : LWLockRelease(newPartitionLock);
2105 :
2106 3208 : *foundPtr = true;
2107 :
2108 3208 : if (!valid)
2109 : {
2110 : /*
2111 : * We can only get here if (a) someone else is still reading in
2112 : * the page, (b) a previous read attempt failed, or (c) someone
2113 : * called StartReadBuffers() but not yet WaitReadBuffers().
2114 : */
2115 2410 : *foundPtr = false;
2116 : }
2117 :
2118 3208 : return existing_buf_hdr;
2119 : }
2120 :
2121 : /*
2122 : * Need to lock the buffer header too in order to change its tag.
2123 : */
2124 3334386 : victim_buf_state = LockBufHdr(victim_buf_hdr);
2125 :
2126 : /* some sanity checks while we hold the buffer header lock */
2127 : Assert(BUF_STATE_GET_REFCOUNT(victim_buf_state) == 1);
2128 : Assert(!(victim_buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY | BM_IO_IN_PROGRESS)));
2129 :
2130 3334386 : victim_buf_hdr->tag = newTag;
2131 :
2132 : /*
2133 : * Make sure BM_PERMANENT is set for buffers that must be written at every
2134 : * checkpoint. Unlogged buffers only need to be written at shutdown
2135 : * checkpoints, except for their "init" forks, which need to be treated
2136 : * just like permanent relations.
2137 : */
2138 3334386 : victim_buf_state |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2139 3334386 : if (relpersistence == RELPERSISTENCE_PERMANENT || forkNum == INIT_FORKNUM)
2140 3333702 : victim_buf_state |= BM_PERMANENT;
2141 :
2142 3334386 : UnlockBufHdr(victim_buf_hdr, victim_buf_state);
2143 :
2144 3334386 : LWLockRelease(newPartitionLock);
2145 :
2146 : /*
2147 : * Buffer contents are currently invalid.
2148 : */
2149 3334386 : *foundPtr = false;
2150 :
2151 3334386 : return victim_buf_hdr;
2152 : }
2153 :
2154 : /*
2155 : * InvalidateBuffer -- mark a shared buffer invalid.
2156 : *
2157 : * The buffer header spinlock must be held at entry. We drop it before
2158 : * returning. (This is sane because the caller must have locked the
2159 : * buffer in order to be sure it should be dropped.)
2160 : *
2161 : * This is used only in contexts such as dropping a relation. We assume
2162 : * that no other backend could possibly be interested in using the page,
2163 : * so the only reason the buffer might be pinned is if someone else is
2164 : * trying to write it out. We have to let them finish before we can
2165 : * reclaim the buffer.
2166 : *
2167 : * The buffer could get reclaimed by someone else while we are waiting
2168 : * to acquire the necessary locks; if so, don't mess it up.
2169 : */
2170 : static void
2171 210392 : InvalidateBuffer(BufferDesc *buf)
2172 : {
2173 : BufferTag oldTag;
2174 : uint32 oldHash; /* hash value for oldTag */
2175 : LWLock *oldPartitionLock; /* buffer partition lock for it */
2176 : uint32 oldFlags;
2177 : uint32 buf_state;
2178 :
2179 : /* Save the original buffer tag before dropping the spinlock */
2180 210392 : oldTag = buf->tag;
2181 :
2182 210392 : buf_state = pg_atomic_read_u32(&buf->state);
2183 : Assert(buf_state & BM_LOCKED);
2184 210392 : UnlockBufHdr(buf, buf_state);
2185 :
2186 : /*
2187 : * Need to compute the old tag's hashcode and partition lock ID. XXX is it
2188 : * worth storing the hashcode in BufferDesc so we need not recompute it
2189 : * here? Probably not.
2190 : */
2191 210392 : oldHash = BufTableHashCode(&oldTag);
2192 210392 : oldPartitionLock = BufMappingPartitionLock(oldHash);
2193 :
2194 210394 : retry:
2195 :
2196 : /*
2197 : * Acquire exclusive mapping lock in preparation for changing the buffer's
2198 : * association.
2199 : */
2200 210394 : LWLockAcquire(oldPartitionLock, LW_EXCLUSIVE);
2201 :
2202 : /* Re-lock the buffer header */
2203 210394 : buf_state = LockBufHdr(buf);
2204 :
2205 : /* If it's changed while we were waiting for lock, do nothing */
2206 210394 : if (!BufferTagsEqual(&buf->tag, &oldTag))
2207 : {
2208 2 : UnlockBufHdr(buf, buf_state);
2209 2 : LWLockRelease(oldPartitionLock);
2210 2 : return;
2211 : }
2212 :
2213 : /*
2214 : * We assume the reason for it to be pinned is that either we were
2215 : * asynchronously reading the page in before erroring out or someone else
2216 : * is flushing the page out. Wait for the IO to finish. (This could be
2217 : * an infinite loop if the refcount is messed up... it would be nice to
2218 : * time out after awhile, but there seems no way to be sure how many loops
2219 : * may be needed. Note that if the other guy has pinned the buffer but
2220 : * not yet done StartBufferIO, WaitIO will fall through and we'll
2221 : * effectively be busy-looping here.)
2222 : */
2223 210392 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 0)
2224 : {
2225 2 : UnlockBufHdr(buf, buf_state);
2226 2 : LWLockRelease(oldPartitionLock);
2227 : /* safety check: should definitely not be our *own* pin */
2228 2 : if (GetPrivateRefCount(BufferDescriptorGetBuffer(buf)) > 0)
2229 0 : elog(ERROR, "buffer is pinned in InvalidateBuffer");
2230 2 : WaitIO(buf);
2231 2 : goto retry;
2232 : }
2233 :
2234 : /*
2235 : * Clear out the buffer's tag and flags. We must do this to ensure that
2236 : * linear scans of the buffer array don't think the buffer is valid.
2237 : */
2238 210390 : oldFlags = buf_state & BUF_FLAG_MASK;
2239 210390 : ClearBufferTag(&buf->tag);
2240 210390 : buf_state &= ~(BUF_FLAG_MASK | BUF_USAGECOUNT_MASK);
2241 210390 : UnlockBufHdr(buf, buf_state);
2242 :
2243 : /*
2244 : * Remove the buffer from the lookup hashtable, if it was in there.
2245 : */
2246 210390 : if (oldFlags & BM_TAG_VALID)
2247 210390 : BufTableDelete(&oldTag, oldHash);
2248 :
2249 : /*
2250 : * Done with mapping lock.
2251 : */
2252 210390 : LWLockRelease(oldPartitionLock);
2253 : }
2254 :
2255 : /*
2256 : * Helper routine for GetVictimBuffer()
2257 : *
2258 : * Needs to be called on a buffer with a valid tag, pinned, but without the
2259 : * buffer header spinlock held.
2260 : *
2261 : * Returns true if the buffer can be reused, in which case the buffer is only
2262 : * pinned by this backend and marked as invalid, false otherwise.
2263 : */
2264 : static bool
2265 2348824 : InvalidateVictimBuffer(BufferDesc *buf_hdr)
2266 : {
2267 : uint32 buf_state;
2268 : uint32 hash;
2269 : LWLock *partition_lock;
2270 : BufferTag tag;
2271 :
2272 : Assert(GetPrivateRefCount(BufferDescriptorGetBuffer(buf_hdr)) == 1);
2273 :
2274 : /* have buffer pinned, so it's safe to read tag without lock */
2275 2348824 : tag = buf_hdr->tag;
2276 :
2277 2348824 : hash = BufTableHashCode(&tag);
2278 2348824 : partition_lock = BufMappingPartitionLock(hash);
2279 :
2280 2348824 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2281 :
2282 : /* lock the buffer header */
2283 2348824 : buf_state = LockBufHdr(buf_hdr);
2284 :
2285 : /*
2286 : * We have the buffer pinned nobody else should have been able to unset
2287 : * this concurrently.
2288 : */
2289 : Assert(buf_state & BM_TAG_VALID);
2290 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2291 : Assert(BufferTagsEqual(&buf_hdr->tag, &tag));
2292 :
2293 : /*
2294 : * If somebody else pinned the buffer since, or even worse, dirtied it,
2295 : * give up on this buffer: It's clearly in use.
2296 : */
2297 2348824 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 1 || (buf_state & BM_DIRTY))
2298 : {
2299 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2300 :
2301 964 : UnlockBufHdr(buf_hdr, buf_state);
2302 964 : LWLockRelease(partition_lock);
2303 :
2304 964 : return false;
2305 : }
2306 :
2307 : /*
2308 : * Clear out the buffer's tag and flags and usagecount. This is not
2309 : * strictly required, as BM_TAG_VALID/BM_VALID needs to be checked before
2310 : * doing anything with the buffer. But currently it's beneficial, as the
2311 : * cheaper pre-check for several linear scans of shared buffers use the
2312 : * tag (see e.g. FlushDatabaseBuffers()).
2313 : */
2314 2347860 : ClearBufferTag(&buf_hdr->tag);
2315 2347860 : buf_state &= ~(BUF_FLAG_MASK | BUF_USAGECOUNT_MASK);
2316 2347860 : UnlockBufHdr(buf_hdr, buf_state);
2317 :
2318 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2319 :
2320 : /* finally delete buffer from the buffer mapping table */
2321 2347860 : BufTableDelete(&tag, hash);
2322 :
2323 2347860 : LWLockRelease(partition_lock);
2324 :
2325 : Assert(!(buf_state & (BM_DIRTY | BM_VALID | BM_TAG_VALID)));
2326 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2327 : Assert(BUF_STATE_GET_REFCOUNT(pg_atomic_read_u32(&buf_hdr->state)) > 0);
2328 :
2329 2347860 : return true;
2330 : }
2331 :
2332 : static Buffer
2333 3781342 : GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context)
2334 : {
2335 : BufferDesc *buf_hdr;
2336 : Buffer buf;
2337 : uint32 buf_state;
2338 : bool from_ring;
2339 :
2340 : /*
2341 : * Ensure, while the spinlock's not yet held, that there's a free refcount
2342 : * entry, and a resource owner slot for the pin.
2343 : */
2344 3781342 : ReservePrivateRefCountEntry();
2345 3781342 : ResourceOwnerEnlarge(CurrentResourceOwner);
2346 :
2347 : /* we return here if a prospective victim buffer gets used concurrently */
2348 12424 : again:
2349 :
2350 : /*
2351 : * Select a victim buffer. The buffer is returned with its header
2352 : * spinlock still held!
2353 : */
2354 3793766 : buf_hdr = StrategyGetBuffer(strategy, &buf_state, &from_ring);
2355 3793766 : buf = BufferDescriptorGetBuffer(buf_hdr);
2356 :
2357 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 0);
2358 :
2359 : /* Pin the buffer and then release the buffer spinlock */
2360 3793766 : PinBuffer_Locked(buf_hdr);
2361 :
2362 : /*
2363 : * We shouldn't have any other pins for this buffer.
2364 : */
2365 3793766 : CheckBufferIsPinnedOnce(buf);
2366 :
2367 : /*
2368 : * If the buffer was dirty, try to write it out. There is a race
2369 : * condition here, in that someone might dirty it after we released the
2370 : * buffer header lock above, or even while we are writing it out (since
2371 : * our share-lock won't prevent hint-bit updates). We will recheck the
2372 : * dirty bit after re-locking the buffer header.
2373 : */
2374 3793766 : if (buf_state & BM_DIRTY)
2375 : {
2376 : LWLock *content_lock;
2377 :
2378 : Assert(buf_state & BM_TAG_VALID);
2379 : Assert(buf_state & BM_VALID);
2380 :
2381 : /*
2382 : * We need a share-lock on the buffer contents to write it out (else
2383 : * we might write invalid data, eg because someone else is compacting
2384 : * the page contents while we write). We must use a conditional lock
2385 : * acquisition here to avoid deadlock. Even though the buffer was not
2386 : * pinned (and therefore surely not locked) when StrategyGetBuffer
2387 : * returned it, someone else could have pinned and exclusive-locked it
2388 : * by the time we get here. If we try to get the lock unconditionally,
2389 : * we'd block waiting for them; if they later block waiting for us,
2390 : * deadlock ensues. (This has been observed to happen when two
2391 : * backends are both trying to split btree index pages, and the second
2392 : * one just happens to be trying to split the page the first one got
2393 : * from StrategyGetBuffer.)
2394 : */
2395 531570 : content_lock = BufferDescriptorGetContentLock(buf_hdr);
2396 531570 : if (!LWLockConditionalAcquire(content_lock, LW_SHARED))
2397 : {
2398 : /*
2399 : * Someone else has locked the buffer, so give it up and loop back
2400 : * to get another one.
2401 : */
2402 0 : UnpinBuffer(buf_hdr);
2403 0 : goto again;
2404 : }
2405 :
2406 : /*
2407 : * If using a nondefault strategy, and writing the buffer would
2408 : * require a WAL flush, let the strategy decide whether to go ahead
2409 : * and write/reuse the buffer or to choose another victim. We need a
2410 : * lock to inspect the page LSN, so this can't be done inside
2411 : * StrategyGetBuffer.
2412 : */
2413 531570 : if (strategy != NULL)
2414 : {
2415 : XLogRecPtr lsn;
2416 :
2417 : /* Read the LSN while holding buffer header lock */
2418 156364 : buf_state = LockBufHdr(buf_hdr);
2419 156364 : lsn = BufferGetLSN(buf_hdr);
2420 156364 : UnlockBufHdr(buf_hdr, buf_state);
2421 :
2422 156364 : if (XLogNeedsFlush(lsn)
2423 18178 : && StrategyRejectBuffer(strategy, buf_hdr, from_ring))
2424 : {
2425 11460 : LWLockRelease(content_lock);
2426 11460 : UnpinBuffer(buf_hdr);
2427 11460 : goto again;
2428 : }
2429 : }
2430 :
2431 : /* OK, do the I/O */
2432 520110 : FlushBuffer(buf_hdr, NULL, IOOBJECT_RELATION, io_context);
2433 520110 : LWLockRelease(content_lock);
2434 :
2435 520110 : ScheduleBufferTagForWriteback(&BackendWritebackContext, io_context,
2436 : &buf_hdr->tag);
2437 : }
2438 :
2439 :
2440 3782306 : if (buf_state & BM_VALID)
2441 : {
2442 : /*
2443 : * When a BufferAccessStrategy is in use, blocks evicted from shared
2444 : * buffers are counted as IOOP_EVICT in the corresponding context
2445 : * (e.g. IOCONTEXT_BULKWRITE). Shared buffers are evicted by a
2446 : * strategy in two cases: 1) while initially claiming buffers for the
2447 : * strategy ring 2) to replace an existing strategy ring buffer
2448 : * because it is pinned or in use and cannot be reused.
2449 : *
2450 : * Blocks evicted from buffers already in the strategy ring are
2451 : * counted as IOOP_REUSE in the corresponding strategy context.
2452 : *
2453 : * At this point, we can accurately count evictions and reuses,
2454 : * because we have successfully claimed the valid buffer. Previously,
2455 : * we may have been forced to release the buffer due to concurrent
2456 : * pinners or erroring out.
2457 : */
2458 2344548 : pgstat_count_io_op(IOOBJECT_RELATION, io_context,
2459 2344548 : from_ring ? IOOP_REUSE : IOOP_EVICT, 1, 0);
2460 : }
2461 :
2462 : /*
2463 : * If the buffer has an entry in the buffer mapping table, delete it. This
2464 : * can fail because another backend could have pinned or dirtied the
2465 : * buffer.
2466 : */
2467 3782306 : if ((buf_state & BM_TAG_VALID) && !InvalidateVictimBuffer(buf_hdr))
2468 : {
2469 964 : UnpinBuffer(buf_hdr);
2470 964 : goto again;
2471 : }
2472 :
2473 : /* a final set of sanity checks */
2474 : #ifdef USE_ASSERT_CHECKING
2475 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
2476 :
2477 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2478 : Assert(!(buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY)));
2479 :
2480 : CheckBufferIsPinnedOnce(buf);
2481 : #endif
2482 :
2483 3781342 : return buf;
2484 : }
2485 :
2486 : /*
2487 : * Return the maximum number of buffers that a backend should try to pin once,
2488 : * to avoid exceeding its fair share. This is the highest value that
2489 : * GetAdditionalPinLimit() could ever return. Note that it may be zero on a
2490 : * system with a very small buffer pool relative to max_connections.
2491 : */
2492 : uint32
2493 1307342 : GetPinLimit(void)
2494 : {
2495 1307342 : return MaxProportionalPins;
2496 : }
2497 :
2498 : /*
2499 : * Return the maximum number of additional buffers that this backend should
2500 : * pin if it wants to stay under the per-backend limit, considering the number
2501 : * of buffers it has already pinned. Unlike LimitAdditionalPins(), the limit
2502 : * return by this function can be zero.
2503 : */
2504 : uint32
2505 7958136 : GetAdditionalPinLimit(void)
2506 : {
2507 : uint32 estimated_pins_held;
2508 :
2509 : /*
2510 : * We get the number of "overflowed" pins for free, but don't know the
2511 : * number of pins in PrivateRefCountArray. The cost of calculating that
2512 : * exactly doesn't seem worth it, so just assume the max.
2513 : */
2514 7958136 : estimated_pins_held = PrivateRefCountOverflowed + REFCOUNT_ARRAY_ENTRIES;
2515 :
2516 : /* Is this backend already holding more than its fair share? */
2517 7958136 : if (estimated_pins_held > MaxProportionalPins)
2518 2368312 : return 0;
2519 :
2520 5589824 : return MaxProportionalPins - estimated_pins_held;
2521 : }
2522 :
2523 : /*
2524 : * Limit the number of pins a batch operation may additionally acquire, to
2525 : * avoid running out of pinnable buffers.
2526 : *
2527 : * One additional pin is always allowed, on the assumption that the operation
2528 : * requires at least one to make progress.
2529 : */
2530 : void
2531 401478 : LimitAdditionalPins(uint32 *additional_pins)
2532 : {
2533 : uint32 limit;
2534 :
2535 401478 : if (*additional_pins <= 1)
2536 381886 : return;
2537 :
2538 19592 : limit = GetAdditionalPinLimit();
2539 19592 : limit = Max(limit, 1);
2540 19592 : if (limit < *additional_pins)
2541 10720 : *additional_pins = limit;
2542 : }
2543 :
2544 : /*
2545 : * Logic shared between ExtendBufferedRelBy(), ExtendBufferedRelTo(). Just to
2546 : * avoid duplicating the tracing and relpersistence related logic.
2547 : */
2548 : static BlockNumber
2549 424288 : ExtendBufferedRelCommon(BufferManagerRelation bmr,
2550 : ForkNumber fork,
2551 : BufferAccessStrategy strategy,
2552 : uint32 flags,
2553 : uint32 extend_by,
2554 : BlockNumber extend_upto,
2555 : Buffer *buffers,
2556 : uint32 *extended_by)
2557 : {
2558 : BlockNumber first_block;
2559 :
2560 : TRACE_POSTGRESQL_BUFFER_EXTEND_START(fork,
2561 : bmr.smgr->smgr_rlocator.locator.spcOid,
2562 : bmr.smgr->smgr_rlocator.locator.dbOid,
2563 : bmr.smgr->smgr_rlocator.locator.relNumber,
2564 : bmr.smgr->smgr_rlocator.backend,
2565 : extend_by);
2566 :
2567 424288 : if (bmr.relpersistence == RELPERSISTENCE_TEMP)
2568 22810 : first_block = ExtendBufferedRelLocal(bmr, fork, flags,
2569 : extend_by, extend_upto,
2570 : buffers, &extend_by);
2571 : else
2572 401478 : first_block = ExtendBufferedRelShared(bmr, fork, strategy, flags,
2573 : extend_by, extend_upto,
2574 : buffers, &extend_by);
2575 424288 : *extended_by = extend_by;
2576 :
2577 : TRACE_POSTGRESQL_BUFFER_EXTEND_DONE(fork,
2578 : bmr.smgr->smgr_rlocator.locator.spcOid,
2579 : bmr.smgr->smgr_rlocator.locator.dbOid,
2580 : bmr.smgr->smgr_rlocator.locator.relNumber,
2581 : bmr.smgr->smgr_rlocator.backend,
2582 : *extended_by,
2583 : first_block);
2584 :
2585 424288 : return first_block;
2586 : }
2587 :
2588 : /*
2589 : * Implementation of ExtendBufferedRelBy() and ExtendBufferedRelTo() for
2590 : * shared buffers.
2591 : */
2592 : static BlockNumber
2593 401478 : ExtendBufferedRelShared(BufferManagerRelation bmr,
2594 : ForkNumber fork,
2595 : BufferAccessStrategy strategy,
2596 : uint32 flags,
2597 : uint32 extend_by,
2598 : BlockNumber extend_upto,
2599 : Buffer *buffers,
2600 : uint32 *extended_by)
2601 : {
2602 : BlockNumber first_block;
2603 401478 : IOContext io_context = IOContextForStrategy(strategy);
2604 : instr_time io_start;
2605 :
2606 401478 : LimitAdditionalPins(&extend_by);
2607 :
2608 : /*
2609 : * Acquire victim buffers for extension without holding extension lock.
2610 : * Writing out victim buffers is the most expensive part of extending the
2611 : * relation, particularly when doing so requires WAL flushes. Zeroing out
2612 : * the buffers is also quite expensive, so do that before holding the
2613 : * extension lock as well.
2614 : *
2615 : * These pages are pinned by us and not valid. While we hold the pin they
2616 : * can't be acquired as victim buffers by another backend.
2617 : */
2618 845226 : for (uint32 i = 0; i < extend_by; i++)
2619 : {
2620 : Block buf_block;
2621 :
2622 443748 : buffers[i] = GetVictimBuffer(strategy, io_context);
2623 443748 : buf_block = BufHdrGetBlock(GetBufferDescriptor(buffers[i] - 1));
2624 :
2625 : /* new buffers are zero-filled */
2626 443748 : MemSet(buf_block, 0, BLCKSZ);
2627 : }
2628 :
2629 : /*
2630 : * Lock relation against concurrent extensions, unless requested not to.
2631 : *
2632 : * We use the same extension lock for all forks. That's unnecessarily
2633 : * restrictive, but currently extensions for forks don't happen often
2634 : * enough to make it worth locking more granularly.
2635 : *
2636 : * Note that another backend might have extended the relation by the time
2637 : * we get the lock.
2638 : */
2639 401478 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2640 298884 : LockRelationForExtension(bmr.rel, ExclusiveLock);
2641 :
2642 : /*
2643 : * If requested, invalidate size cache, so that smgrnblocks asks the
2644 : * kernel.
2645 : */
2646 401478 : if (flags & EB_CLEAR_SIZE_CACHE)
2647 15360 : bmr.smgr->smgr_cached_nblocks[fork] = InvalidBlockNumber;
2648 :
2649 401478 : first_block = smgrnblocks(bmr.smgr, fork);
2650 :
2651 : /*
2652 : * Now that we have the accurate relation size, check if the caller wants
2653 : * us to extend to only up to a specific size. If there were concurrent
2654 : * extensions, we might have acquired too many buffers and need to release
2655 : * them.
2656 : */
2657 401478 : if (extend_upto != InvalidBlockNumber)
2658 : {
2659 106182 : uint32 orig_extend_by = extend_by;
2660 :
2661 106182 : if (first_block > extend_upto)
2662 0 : extend_by = 0;
2663 106182 : else if ((uint64) first_block + extend_by > extend_upto)
2664 14 : extend_by = extend_upto - first_block;
2665 :
2666 106224 : for (uint32 i = extend_by; i < orig_extend_by; i++)
2667 : {
2668 42 : BufferDesc *buf_hdr = GetBufferDescriptor(buffers[i] - 1);
2669 :
2670 42 : UnpinBuffer(buf_hdr);
2671 : }
2672 :
2673 106182 : if (extend_by == 0)
2674 : {
2675 14 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2676 14 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2677 14 : *extended_by = extend_by;
2678 14 : return first_block;
2679 : }
2680 : }
2681 :
2682 : /* Fail if relation is already at maximum possible length */
2683 401464 : if ((uint64) first_block + extend_by >= MaxBlockNumber)
2684 0 : ereport(ERROR,
2685 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
2686 : errmsg("cannot extend relation %s beyond %u blocks",
2687 : relpath(bmr.smgr->smgr_rlocator, fork).str,
2688 : MaxBlockNumber)));
2689 :
2690 : /*
2691 : * Insert buffers into buffer table, mark as IO_IN_PROGRESS.
2692 : *
2693 : * This needs to happen before we extend the relation, because as soon as
2694 : * we do, other backends can start to read in those pages.
2695 : */
2696 845170 : for (uint32 i = 0; i < extend_by; i++)
2697 : {
2698 443706 : Buffer victim_buf = buffers[i];
2699 443706 : BufferDesc *victim_buf_hdr = GetBufferDescriptor(victim_buf - 1);
2700 : BufferTag tag;
2701 : uint32 hash;
2702 : LWLock *partition_lock;
2703 : int existing_id;
2704 :
2705 : /* in case we need to pin an existing buffer below */
2706 443706 : ResourceOwnerEnlarge(CurrentResourceOwner);
2707 443706 : ReservePrivateRefCountEntry();
2708 :
2709 443706 : InitBufferTag(&tag, &bmr.smgr->smgr_rlocator.locator, fork, first_block + i);
2710 443706 : hash = BufTableHashCode(&tag);
2711 443706 : partition_lock = BufMappingPartitionLock(hash);
2712 :
2713 443706 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2714 :
2715 443706 : existing_id = BufTableInsert(&tag, hash, victim_buf_hdr->buf_id);
2716 :
2717 : /*
2718 : * We get here only in the corner case where we are trying to extend
2719 : * the relation but we found a pre-existing buffer. This can happen
2720 : * because a prior attempt at extending the relation failed, and
2721 : * because mdread doesn't complain about reads beyond EOF (when
2722 : * zero_damaged_pages is ON) and so a previous attempt to read a block
2723 : * beyond EOF could have left a "valid" zero-filled buffer.
2724 : *
2725 : * This has also been observed when relation was overwritten by
2726 : * external process. Since the legitimate cases should always have
2727 : * left a zero-filled buffer, complain if not PageIsNew.
2728 : */
2729 443706 : if (existing_id >= 0)
2730 : {
2731 0 : BufferDesc *existing_hdr = GetBufferDescriptor(existing_id);
2732 : Block buf_block;
2733 : bool valid;
2734 :
2735 : /*
2736 : * Pin the existing buffer before releasing the partition lock,
2737 : * preventing it from being evicted.
2738 : */
2739 0 : valid = PinBuffer(existing_hdr, strategy);
2740 :
2741 0 : LWLockRelease(partition_lock);
2742 0 : UnpinBuffer(victim_buf_hdr);
2743 :
2744 0 : buffers[i] = BufferDescriptorGetBuffer(existing_hdr);
2745 0 : buf_block = BufHdrGetBlock(existing_hdr);
2746 :
2747 0 : if (valid && !PageIsNew((Page) buf_block))
2748 0 : ereport(ERROR,
2749 : (errmsg("unexpected data beyond EOF in block %u of relation \"%s\"",
2750 : existing_hdr->tag.blockNum,
2751 : relpath(bmr.smgr->smgr_rlocator, fork).str)));
2752 :
2753 : /*
2754 : * We *must* do smgr[zero]extend before succeeding, else the page
2755 : * will not be reserved by the kernel, and the next P_NEW call
2756 : * will decide to return the same page. Clear the BM_VALID bit,
2757 : * do StartBufferIO() and proceed.
2758 : *
2759 : * Loop to handle the very small possibility that someone re-sets
2760 : * BM_VALID between our clearing it and StartBufferIO inspecting
2761 : * it.
2762 : */
2763 : do
2764 : {
2765 0 : uint32 buf_state = LockBufHdr(existing_hdr);
2766 :
2767 0 : buf_state &= ~BM_VALID;
2768 0 : UnlockBufHdr(existing_hdr, buf_state);
2769 0 : } while (!StartBufferIO(existing_hdr, true, false));
2770 : }
2771 : else
2772 : {
2773 : uint32 buf_state;
2774 :
2775 443706 : buf_state = LockBufHdr(victim_buf_hdr);
2776 :
2777 : /* some sanity checks while we hold the buffer header lock */
2778 : Assert(!(buf_state & (BM_VALID | BM_TAG_VALID | BM_DIRTY | BM_JUST_DIRTIED)));
2779 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2780 :
2781 443706 : victim_buf_hdr->tag = tag;
2782 :
2783 443706 : buf_state |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2784 443706 : if (bmr.relpersistence == RELPERSISTENCE_PERMANENT || fork == INIT_FORKNUM)
2785 433156 : buf_state |= BM_PERMANENT;
2786 :
2787 443706 : UnlockBufHdr(victim_buf_hdr, buf_state);
2788 :
2789 443706 : LWLockRelease(partition_lock);
2790 :
2791 : /* XXX: could combine the locked operations in it with the above */
2792 443706 : StartBufferIO(victim_buf_hdr, true, false);
2793 : }
2794 : }
2795 :
2796 401464 : io_start = pgstat_prepare_io_time(track_io_timing);
2797 :
2798 : /*
2799 : * Note: if smgrzeroextend fails, we will end up with buffers that are
2800 : * allocated but not marked BM_VALID. The next relation extension will
2801 : * still select the same block number (because the relation didn't get any
2802 : * longer on disk) and so future attempts to extend the relation will find
2803 : * the same buffers (if they have not been recycled) but come right back
2804 : * here to try smgrzeroextend again.
2805 : *
2806 : * We don't need to set checksum for all-zero pages.
2807 : */
2808 401464 : smgrzeroextend(bmr.smgr, fork, first_block, extend_by, false);
2809 :
2810 : /*
2811 : * Release the file-extension lock; it's now OK for someone else to extend
2812 : * the relation some more.
2813 : *
2814 : * We remove IO_IN_PROGRESS after this, as waking up waiting backends can
2815 : * take noticeable time.
2816 : */
2817 401464 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2818 298870 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2819 :
2820 401464 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context, IOOP_EXTEND,
2821 401464 : io_start, 1, extend_by * BLCKSZ);
2822 :
2823 : /* Set BM_VALID, terminate IO, and wake up any waiters */
2824 845170 : for (uint32 i = 0; i < extend_by; i++)
2825 : {
2826 443706 : Buffer buf = buffers[i];
2827 443706 : BufferDesc *buf_hdr = GetBufferDescriptor(buf - 1);
2828 443706 : bool lock = false;
2829 :
2830 443706 : if (flags & EB_LOCK_FIRST && i == 0)
2831 294764 : lock = true;
2832 148942 : else if (flags & EB_LOCK_TARGET)
2833 : {
2834 : Assert(extend_upto != InvalidBlockNumber);
2835 88694 : if (first_block + i + 1 == extend_upto)
2836 87772 : lock = true;
2837 : }
2838 :
2839 443706 : if (lock)
2840 382536 : LWLockAcquire(BufferDescriptorGetContentLock(buf_hdr), LW_EXCLUSIVE);
2841 :
2842 443706 : TerminateBufferIO(buf_hdr, false, BM_VALID, true, false);
2843 : }
2844 :
2845 401464 : pgBufferUsage.shared_blks_written += extend_by;
2846 :
2847 401464 : *extended_by = extend_by;
2848 :
2849 401464 : return first_block;
2850 : }
2851 :
2852 : /*
2853 : * BufferIsExclusiveLocked
2854 : *
2855 : * Checks if buffer is exclusive-locked.
2856 : *
2857 : * Buffer must be pinned.
2858 : */
2859 : bool
2860 0 : BufferIsExclusiveLocked(Buffer buffer)
2861 : {
2862 : BufferDesc *bufHdr;
2863 :
2864 : Assert(BufferIsPinned(buffer));
2865 :
2866 0 : if (BufferIsLocal(buffer))
2867 : {
2868 : /* Content locks are not maintained for local buffers. */
2869 0 : return true;
2870 : }
2871 : else
2872 : {
2873 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2874 0 : return LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
2875 : LW_EXCLUSIVE);
2876 : }
2877 : }
2878 :
2879 : /*
2880 : * BufferIsDirty
2881 : *
2882 : * Checks if buffer is already dirty.
2883 : *
2884 : * Buffer must be pinned and exclusive-locked. (Without an exclusive lock,
2885 : * the result may be stale before it's returned.)
2886 : */
2887 : bool
2888 0 : BufferIsDirty(Buffer buffer)
2889 : {
2890 : BufferDesc *bufHdr;
2891 :
2892 : Assert(BufferIsPinned(buffer));
2893 :
2894 0 : if (BufferIsLocal(buffer))
2895 : {
2896 0 : int bufid = -buffer - 1;
2897 :
2898 0 : bufHdr = GetLocalBufferDescriptor(bufid);
2899 : /* Content locks are not maintained for local buffers. */
2900 : }
2901 : else
2902 : {
2903 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2904 : Assert(LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
2905 : LW_EXCLUSIVE));
2906 : }
2907 :
2908 0 : return pg_atomic_read_u32(&bufHdr->state) & BM_DIRTY;
2909 : }
2910 :
2911 : /*
2912 : * MarkBufferDirty
2913 : *
2914 : * Marks buffer contents as dirty (actual write happens later).
2915 : *
2916 : * Buffer must be pinned and exclusive-locked. (If caller does not hold
2917 : * exclusive lock, then somebody could be in process of writing the buffer,
2918 : * leading to risk of bad data written to disk.)
2919 : */
2920 : void
2921 42820912 : MarkBufferDirty(Buffer buffer)
2922 : {
2923 : BufferDesc *bufHdr;
2924 : uint32 buf_state;
2925 : uint32 old_buf_state;
2926 :
2927 42820912 : if (!BufferIsValid(buffer))
2928 0 : elog(ERROR, "bad buffer ID: %d", buffer);
2929 :
2930 42820912 : if (BufferIsLocal(buffer))
2931 : {
2932 2444464 : MarkLocalBufferDirty(buffer);
2933 2444464 : return;
2934 : }
2935 :
2936 40376448 : bufHdr = GetBufferDescriptor(buffer - 1);
2937 :
2938 : Assert(BufferIsPinned(buffer));
2939 : Assert(LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
2940 : LW_EXCLUSIVE));
2941 :
2942 40376448 : old_buf_state = pg_atomic_read_u32(&bufHdr->state);
2943 : for (;;)
2944 : {
2945 40376734 : if (old_buf_state & BM_LOCKED)
2946 50 : old_buf_state = WaitBufHdrUnlocked(bufHdr);
2947 :
2948 40376734 : buf_state = old_buf_state;
2949 :
2950 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2951 40376734 : buf_state |= BM_DIRTY | BM_JUST_DIRTIED;
2952 :
2953 40376734 : if (pg_atomic_compare_exchange_u32(&bufHdr->state, &old_buf_state,
2954 : buf_state))
2955 40376448 : break;
2956 : }
2957 :
2958 : /*
2959 : * If the buffer was not dirty already, do vacuum accounting.
2960 : */
2961 40376448 : if (!(old_buf_state & BM_DIRTY))
2962 : {
2963 1292434 : pgBufferUsage.shared_blks_dirtied++;
2964 1292434 : if (VacuumCostActive)
2965 13548 : VacuumCostBalance += VacuumCostPageDirty;
2966 : }
2967 : }
2968 :
2969 : /*
2970 : * ReleaseAndReadBuffer -- combine ReleaseBuffer() and ReadBuffer()
2971 : *
2972 : * Formerly, this saved one cycle of acquiring/releasing the BufMgrLock
2973 : * compared to calling the two routines separately. Now it's mainly just
2974 : * a convenience function. However, if the passed buffer is valid and
2975 : * already contains the desired block, we just return it as-is; and that
2976 : * does save considerable work compared to a full release and reacquire.
2977 : *
2978 : * Note: it is OK to pass buffer == InvalidBuffer, indicating that no old
2979 : * buffer actually needs to be released. This case is the same as ReadBuffer,
2980 : * but can save some tests in the caller.
2981 : */
2982 : Buffer
2983 57575568 : ReleaseAndReadBuffer(Buffer buffer,
2984 : Relation relation,
2985 : BlockNumber blockNum)
2986 : {
2987 57575568 : ForkNumber forkNum = MAIN_FORKNUM;
2988 : BufferDesc *bufHdr;
2989 :
2990 57575568 : if (BufferIsValid(buffer))
2991 : {
2992 : Assert(BufferIsPinned(buffer));
2993 34376622 : if (BufferIsLocal(buffer))
2994 : {
2995 73728 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
2996 80772 : if (bufHdr->tag.blockNum == blockNum &&
2997 14088 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
2998 7044 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
2999 7044 : return buffer;
3000 66684 : UnpinLocalBuffer(buffer);
3001 : }
3002 : else
3003 : {
3004 34302894 : bufHdr = GetBufferDescriptor(buffer - 1);
3005 : /* we have pin, so it's ok to examine tag without spinlock */
3006 46093648 : if (bufHdr->tag.blockNum == blockNum &&
3007 23581508 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3008 11790754 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3009 11790754 : return buffer;
3010 22512140 : UnpinBuffer(bufHdr);
3011 : }
3012 : }
3013 :
3014 45777770 : return ReadBuffer(relation, blockNum);
3015 : }
3016 :
3017 : /*
3018 : * PinBuffer -- make buffer unavailable for replacement.
3019 : *
3020 : * For the default access strategy, the buffer's usage_count is incremented
3021 : * when we first pin it; for other strategies we just make sure the usage_count
3022 : * isn't zero. (The idea of the latter is that we don't want synchronized
3023 : * heap scans to inflate the count, but we need it to not be zero to discourage
3024 : * other backends from stealing buffers from our ring. As long as we cycle
3025 : * through the ring faster than the global clock-sweep cycles, buffers in
3026 : * our ring won't be chosen as victims for replacement by other backends.)
3027 : *
3028 : * This should be applied only to shared buffers, never local ones.
3029 : *
3030 : * Since buffers are pinned/unpinned very frequently, pin buffers without
3031 : * taking the buffer header lock; instead update the state variable in loop of
3032 : * CAS operations. Hopefully it's just a single CAS.
3033 : *
3034 : * Note that ResourceOwnerEnlarge() and ReservePrivateRefCountEntry()
3035 : * must have been done already.
3036 : *
3037 : * Returns true if buffer is BM_VALID, else false. This provision allows
3038 : * some callers to avoid an extra spinlock cycle.
3039 : */
3040 : static bool
3041 117970576 : PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy)
3042 : {
3043 117970576 : Buffer b = BufferDescriptorGetBuffer(buf);
3044 : bool result;
3045 : PrivateRefCountEntry *ref;
3046 :
3047 : Assert(!BufferIsLocal(b));
3048 : Assert(ReservedRefCountEntry != NULL);
3049 :
3050 117970576 : ref = GetPrivateRefCountEntry(b, true);
3051 :
3052 117970576 : if (ref == NULL)
3053 : {
3054 : uint32 buf_state;
3055 : uint32 old_buf_state;
3056 :
3057 113657046 : ref = NewPrivateRefCountEntry(b);
3058 :
3059 113657046 : old_buf_state = pg_atomic_read_u32(&buf->state);
3060 : for (;;)
3061 : {
3062 113701608 : if (old_buf_state & BM_LOCKED)
3063 1750 : old_buf_state = WaitBufHdrUnlocked(buf);
3064 :
3065 113701608 : buf_state = old_buf_state;
3066 :
3067 : /* increase refcount */
3068 113701608 : buf_state += BUF_REFCOUNT_ONE;
3069 :
3070 113701608 : if (strategy == NULL)
3071 : {
3072 : /* Default case: increase usagecount unless already max. */
3073 112030192 : if (BUF_STATE_GET_USAGECOUNT(buf_state) < BM_MAX_USAGE_COUNT)
3074 6527450 : buf_state += BUF_USAGECOUNT_ONE;
3075 : }
3076 : else
3077 : {
3078 : /*
3079 : * Ring buffers shouldn't evict others from pool. Thus we
3080 : * don't make usagecount more than 1.
3081 : */
3082 1671416 : if (BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3083 68184 : buf_state += BUF_USAGECOUNT_ONE;
3084 : }
3085 :
3086 113701608 : if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
3087 : buf_state))
3088 : {
3089 113657046 : result = (buf_state & BM_VALID) != 0;
3090 :
3091 : /*
3092 : * Assume that we acquired a buffer pin for the purposes of
3093 : * Valgrind buffer client checks (even in !result case) to
3094 : * keep things simple. Buffers that are unsafe to access are
3095 : * not generally guaranteed to be marked undefined or
3096 : * non-accessible in any case.
3097 : */
3098 : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(buf), BLCKSZ);
3099 113657046 : break;
3100 : }
3101 : }
3102 : }
3103 : else
3104 : {
3105 : /*
3106 : * If we previously pinned the buffer, it is likely to be valid, but
3107 : * it may not be if StartReadBuffers() was called and
3108 : * WaitReadBuffers() hasn't been called yet. We'll check by loading
3109 : * the flags without locking. This is racy, but it's OK to return
3110 : * false spuriously: when WaitReadBuffers() calls StartBufferIO(),
3111 : * it'll see that it's now valid.
3112 : *
3113 : * Note: We deliberately avoid a Valgrind client request here.
3114 : * Individual access methods can optionally superimpose buffer page
3115 : * client requests on top of our client requests to enforce that
3116 : * buffers are only accessed while locked (and pinned). It's possible
3117 : * that the buffer page is legitimately non-accessible here. We
3118 : * cannot meddle with that.
3119 : */
3120 4313530 : result = (pg_atomic_read_u32(&buf->state) & BM_VALID) != 0;
3121 : }
3122 :
3123 117970576 : ref->refcount++;
3124 : Assert(ref->refcount > 0);
3125 117970576 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
3126 117970576 : return result;
3127 : }
3128 :
3129 : /*
3130 : * PinBuffer_Locked -- as above, but caller already locked the buffer header.
3131 : * The spinlock is released before return.
3132 : *
3133 : * As this function is called with the spinlock held, the caller has to
3134 : * previously call ReservePrivateRefCountEntry() and
3135 : * ResourceOwnerEnlarge(CurrentResourceOwner);
3136 : *
3137 : * Currently, no callers of this function want to modify the buffer's
3138 : * usage_count at all, so there's no need for a strategy parameter.
3139 : * Also we don't bother with a BM_VALID test (the caller could check that for
3140 : * itself).
3141 : *
3142 : * Also all callers only ever use this function when it's known that the
3143 : * buffer can't have a preexisting pin by this backend. That allows us to skip
3144 : * searching the private refcount array & hash, which is a boon, because the
3145 : * spinlock is still held.
3146 : *
3147 : * Note: use of this routine is frequently mandatory, not just an optimization
3148 : * to save a spin lock/unlock cycle, because we need to pin a buffer before
3149 : * its state can change under us.
3150 : */
3151 : static void
3152 4400740 : PinBuffer_Locked(BufferDesc *buf)
3153 : {
3154 : Buffer b;
3155 : PrivateRefCountEntry *ref;
3156 : uint32 buf_state;
3157 :
3158 : /*
3159 : * As explained, We don't expect any preexisting pins. That allows us to
3160 : * manipulate the PrivateRefCount after releasing the spinlock
3161 : */
3162 : Assert(GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf), false) == NULL);
3163 :
3164 : /*
3165 : * Buffer can't have a preexisting pin, so mark its page as defined to
3166 : * Valgrind (this is similar to the PinBuffer() case where the backend
3167 : * doesn't already have a buffer pin)
3168 : */
3169 : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(buf), BLCKSZ);
3170 :
3171 : /*
3172 : * Since we hold the buffer spinlock, we can update the buffer state and
3173 : * release the lock in one operation.
3174 : */
3175 4400740 : buf_state = pg_atomic_read_u32(&buf->state);
3176 : Assert(buf_state & BM_LOCKED);
3177 4400740 : buf_state += BUF_REFCOUNT_ONE;
3178 4400740 : UnlockBufHdr(buf, buf_state);
3179 :
3180 4400740 : b = BufferDescriptorGetBuffer(buf);
3181 :
3182 4400740 : ref = NewPrivateRefCountEntry(b);
3183 4400740 : ref->refcount++;
3184 :
3185 4400740 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
3186 4400740 : }
3187 :
3188 : /*
3189 : * Support for waking up another backend that is waiting for the cleanup lock
3190 : * to be released using BM_PIN_COUNT_WAITER.
3191 : *
3192 : * See LockBufferForCleanup().
3193 : *
3194 : * Expected to be called just after releasing a buffer pin (in a BufferDesc,
3195 : * not just reducing the backend-local pincount for the buffer).
3196 : */
3197 : static void
3198 106 : WakePinCountWaiter(BufferDesc *buf)
3199 : {
3200 : /*
3201 : * Acquire the buffer header lock, re-check that there's a waiter. Another
3202 : * backend could have unpinned this buffer, and already woken up the
3203 : * waiter.
3204 : *
3205 : * There's no danger of the buffer being replaced after we unpinned it
3206 : * above, as it's pinned by the waiter. The waiter removes
3207 : * BM_PIN_COUNT_WAITER if it stops waiting for a reason other than this
3208 : * backend waking it up.
3209 : */
3210 106 : uint32 buf_state = LockBufHdr(buf);
3211 :
3212 106 : if ((buf_state & BM_PIN_COUNT_WAITER) &&
3213 106 : BUF_STATE_GET_REFCOUNT(buf_state) == 1)
3214 106 : {
3215 : /* we just released the last pin other than the waiter's */
3216 106 : int wait_backend_pgprocno = buf->wait_backend_pgprocno;
3217 :
3218 106 : buf_state &= ~BM_PIN_COUNT_WAITER;
3219 106 : UnlockBufHdr(buf, buf_state);
3220 106 : ProcSendSignal(wait_backend_pgprocno);
3221 : }
3222 : else
3223 0 : UnlockBufHdr(buf, buf_state);
3224 106 : }
3225 :
3226 : /*
3227 : * UnpinBuffer -- make buffer available for replacement.
3228 : *
3229 : * This should be applied only to shared buffers, never local ones. This
3230 : * always adjusts CurrentResourceOwner.
3231 : */
3232 : static void
3233 144926368 : UnpinBuffer(BufferDesc *buf)
3234 : {
3235 144926368 : Buffer b = BufferDescriptorGetBuffer(buf);
3236 :
3237 144926368 : ResourceOwnerForgetBuffer(CurrentResourceOwner, b);
3238 144926368 : UnpinBufferNoOwner(buf);
3239 144926368 : }
3240 :
3241 : static void
3242 144935210 : UnpinBufferNoOwner(BufferDesc *buf)
3243 : {
3244 : PrivateRefCountEntry *ref;
3245 144935210 : Buffer b = BufferDescriptorGetBuffer(buf);
3246 :
3247 : Assert(!BufferIsLocal(b));
3248 :
3249 : /* not moving as we're likely deleting it soon anyway */
3250 144935210 : ref = GetPrivateRefCountEntry(b, false);
3251 : Assert(ref != NULL);
3252 : Assert(ref->refcount > 0);
3253 144935210 : ref->refcount--;
3254 144935210 : if (ref->refcount == 0)
3255 : {
3256 : uint32 buf_state;
3257 : uint32 old_buf_state;
3258 :
3259 : /*
3260 : * Mark buffer non-accessible to Valgrind.
3261 : *
3262 : * Note that the buffer may have already been marked non-accessible
3263 : * within access method code that enforces that buffers are only
3264 : * accessed while a buffer lock is held.
3265 : */
3266 : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
3267 :
3268 : /* I'd better not still hold the buffer content lock */
3269 : Assert(!LWLockHeldByMe(BufferDescriptorGetContentLock(buf)));
3270 :
3271 : /*
3272 : * Decrement the shared reference count.
3273 : *
3274 : * Since buffer spinlock holder can update status using just write,
3275 : * it's not safe to use atomic decrement here; thus use a CAS loop.
3276 : */
3277 118057786 : old_buf_state = pg_atomic_read_u32(&buf->state);
3278 : for (;;)
3279 : {
3280 118095808 : if (old_buf_state & BM_LOCKED)
3281 1588 : old_buf_state = WaitBufHdrUnlocked(buf);
3282 :
3283 118095808 : buf_state = old_buf_state;
3284 :
3285 118095808 : buf_state -= BUF_REFCOUNT_ONE;
3286 :
3287 118095808 : if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
3288 : buf_state))
3289 118057786 : break;
3290 : }
3291 :
3292 : /* Support LockBufferForCleanup() */
3293 118057786 : if (buf_state & BM_PIN_COUNT_WAITER)
3294 106 : WakePinCountWaiter(buf);
3295 :
3296 118057786 : ForgetPrivateRefCountEntry(ref);
3297 : }
3298 144935210 : }
3299 :
3300 : #define ST_SORT sort_checkpoint_bufferids
3301 : #define ST_ELEMENT_TYPE CkptSortItem
3302 : #define ST_COMPARE(a, b) ckpt_buforder_comparator(a, b)
3303 : #define ST_SCOPE static
3304 : #define ST_DEFINE
3305 : #include "lib/sort_template.h"
3306 :
3307 : /*
3308 : * BufferSync -- Write out all dirty buffers in the pool.
3309 : *
3310 : * This is called at checkpoint time to write out all dirty shared buffers.
3311 : * The checkpoint request flags should be passed in. If CHECKPOINT_FAST is
3312 : * set, we disable delays between writes; if CHECKPOINT_IS_SHUTDOWN,
3313 : * CHECKPOINT_END_OF_RECOVERY or CHECKPOINT_FLUSH_UNLOGGED is set, we write
3314 : * even unlogged buffers, which are otherwise skipped. The remaining flags
3315 : * currently have no effect here.
3316 : */
3317 : static void
3318 3414 : BufferSync(int flags)
3319 : {
3320 : uint32 buf_state;
3321 : int buf_id;
3322 : int num_to_scan;
3323 : int num_spaces;
3324 : int num_processed;
3325 : int num_written;
3326 3414 : CkptTsStatus *per_ts_stat = NULL;
3327 : Oid last_tsid;
3328 : binaryheap *ts_heap;
3329 : int i;
3330 3414 : int mask = BM_DIRTY;
3331 : WritebackContext wb_context;
3332 :
3333 : /*
3334 : * Unless this is a shutdown checkpoint or we have been explicitly told,
3335 : * we write only permanent, dirty buffers. But at shutdown or end of
3336 : * recovery, we write all dirty buffers.
3337 : */
3338 3414 : if (!((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
3339 : CHECKPOINT_FLUSH_UNLOGGED))))
3340 1918 : mask |= BM_PERMANENT;
3341 :
3342 : /*
3343 : * Loop over all buffers, and mark the ones that need to be written with
3344 : * BM_CHECKPOINT_NEEDED. Count them as we go (num_to_scan), so that we
3345 : * can estimate how much work needs to be done.
3346 : *
3347 : * This allows us to write only those pages that were dirty when the
3348 : * checkpoint began, and not those that get dirtied while it proceeds.
3349 : * Whenever a page with BM_CHECKPOINT_NEEDED is written out, either by us
3350 : * later in this function, or by normal backends or the bgwriter cleaning
3351 : * scan, the flag is cleared. Any buffer dirtied after this point won't
3352 : * have the flag set.
3353 : *
3354 : * Note that if we fail to write some buffer, we may leave buffers with
3355 : * BM_CHECKPOINT_NEEDED still set. This is OK since any such buffer would
3356 : * certainly need to be written for the next checkpoint attempt, too.
3357 : */
3358 3414 : num_to_scan = 0;
3359 23783414 : for (buf_id = 0; buf_id < NBuffers; buf_id++)
3360 : {
3361 23780000 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3362 :
3363 : /*
3364 : * Header spinlock is enough to examine BM_DIRTY, see comment in
3365 : * SyncOneBuffer.
3366 : */
3367 23780000 : buf_state = LockBufHdr(bufHdr);
3368 :
3369 23780000 : if ((buf_state & mask) == mask)
3370 : {
3371 : CkptSortItem *item;
3372 :
3373 578296 : buf_state |= BM_CHECKPOINT_NEEDED;
3374 :
3375 578296 : item = &CkptBufferIds[num_to_scan++];
3376 578296 : item->buf_id = buf_id;
3377 578296 : item->tsId = bufHdr->tag.spcOid;
3378 578296 : item->relNumber = BufTagGetRelNumber(&bufHdr->tag);
3379 578296 : item->forkNum = BufTagGetForkNum(&bufHdr->tag);
3380 578296 : item->blockNum = bufHdr->tag.blockNum;
3381 : }
3382 :
3383 23780000 : UnlockBufHdr(bufHdr, buf_state);
3384 :
3385 : /* Check for barrier events in case NBuffers is large. */
3386 23780000 : if (ProcSignalBarrierPending)
3387 0 : ProcessProcSignalBarrier();
3388 : }
3389 :
3390 3414 : if (num_to_scan == 0)
3391 1328 : return; /* nothing to do */
3392 :
3393 2086 : WritebackContextInit(&wb_context, &checkpoint_flush_after);
3394 :
3395 : TRACE_POSTGRESQL_BUFFER_SYNC_START(NBuffers, num_to_scan);
3396 :
3397 : /*
3398 : * Sort buffers that need to be written to reduce the likelihood of random
3399 : * IO. The sorting is also important for the implementation of balancing
3400 : * writes between tablespaces. Without balancing writes we'd potentially
3401 : * end up writing to the tablespaces one-by-one; possibly overloading the
3402 : * underlying system.
3403 : */
3404 2086 : sort_checkpoint_bufferids(CkptBufferIds, num_to_scan);
3405 :
3406 2086 : num_spaces = 0;
3407 :
3408 : /*
3409 : * Allocate progress status for each tablespace with buffers that need to
3410 : * be flushed. This requires the to-be-flushed array to be sorted.
3411 : */
3412 2086 : last_tsid = InvalidOid;
3413 580382 : for (i = 0; i < num_to_scan; i++)
3414 : {
3415 : CkptTsStatus *s;
3416 : Oid cur_tsid;
3417 :
3418 578296 : cur_tsid = CkptBufferIds[i].tsId;
3419 :
3420 : /*
3421 : * Grow array of per-tablespace status structs, every time a new
3422 : * tablespace is found.
3423 : */
3424 578296 : if (last_tsid == InvalidOid || last_tsid != cur_tsid)
3425 3132 : {
3426 : Size sz;
3427 :
3428 3132 : num_spaces++;
3429 :
3430 : /*
3431 : * Not worth adding grow-by-power-of-2 logic here - even with a
3432 : * few hundred tablespaces this should be fine.
3433 : */
3434 3132 : sz = sizeof(CkptTsStatus) * num_spaces;
3435 :
3436 3132 : if (per_ts_stat == NULL)
3437 2086 : per_ts_stat = (CkptTsStatus *) palloc(sz);
3438 : else
3439 1046 : per_ts_stat = (CkptTsStatus *) repalloc(per_ts_stat, sz);
3440 :
3441 3132 : s = &per_ts_stat[num_spaces - 1];
3442 3132 : memset(s, 0, sizeof(*s));
3443 3132 : s->tsId = cur_tsid;
3444 :
3445 : /*
3446 : * The first buffer in this tablespace. As CkptBufferIds is sorted
3447 : * by tablespace all (s->num_to_scan) buffers in this tablespace
3448 : * will follow afterwards.
3449 : */
3450 3132 : s->index = i;
3451 :
3452 : /*
3453 : * progress_slice will be determined once we know how many buffers
3454 : * are in each tablespace, i.e. after this loop.
3455 : */
3456 :
3457 3132 : last_tsid = cur_tsid;
3458 : }
3459 : else
3460 : {
3461 575164 : s = &per_ts_stat[num_spaces - 1];
3462 : }
3463 :
3464 578296 : s->num_to_scan++;
3465 :
3466 : /* Check for barrier events. */
3467 578296 : if (ProcSignalBarrierPending)
3468 0 : ProcessProcSignalBarrier();
3469 : }
3470 :
3471 : Assert(num_spaces > 0);
3472 :
3473 : /*
3474 : * Build a min-heap over the write-progress in the individual tablespaces,
3475 : * and compute how large a portion of the total progress a single
3476 : * processed buffer is.
3477 : */
3478 2086 : ts_heap = binaryheap_allocate(num_spaces,
3479 : ts_ckpt_progress_comparator,
3480 : NULL);
3481 :
3482 5218 : for (i = 0; i < num_spaces; i++)
3483 : {
3484 3132 : CkptTsStatus *ts_stat = &per_ts_stat[i];
3485 :
3486 3132 : ts_stat->progress_slice = (float8) num_to_scan / ts_stat->num_to_scan;
3487 :
3488 3132 : binaryheap_add_unordered(ts_heap, PointerGetDatum(ts_stat));
3489 : }
3490 :
3491 2086 : binaryheap_build(ts_heap);
3492 :
3493 : /*
3494 : * Iterate through to-be-checkpointed buffers and write the ones (still)
3495 : * marked with BM_CHECKPOINT_NEEDED. The writes are balanced between
3496 : * tablespaces; otherwise the sorting would lead to only one tablespace
3497 : * receiving writes at a time, making inefficient use of the hardware.
3498 : */
3499 2086 : num_processed = 0;
3500 2086 : num_written = 0;
3501 580382 : while (!binaryheap_empty(ts_heap))
3502 : {
3503 578296 : BufferDesc *bufHdr = NULL;
3504 : CkptTsStatus *ts_stat = (CkptTsStatus *)
3505 578296 : DatumGetPointer(binaryheap_first(ts_heap));
3506 :
3507 578296 : buf_id = CkptBufferIds[ts_stat->index].buf_id;
3508 : Assert(buf_id != -1);
3509 :
3510 578296 : bufHdr = GetBufferDescriptor(buf_id);
3511 :
3512 578296 : num_processed++;
3513 :
3514 : /*
3515 : * We don't need to acquire the lock here, because we're only looking
3516 : * at a single bit. It's possible that someone else writes the buffer
3517 : * and clears the flag right after we check, but that doesn't matter
3518 : * since SyncOneBuffer will then do nothing. However, there is a
3519 : * further race condition: it's conceivable that between the time we
3520 : * examine the bit here and the time SyncOneBuffer acquires the lock,
3521 : * someone else not only wrote the buffer but replaced it with another
3522 : * page and dirtied it. In that improbable case, SyncOneBuffer will
3523 : * write the buffer though we didn't need to. It doesn't seem worth
3524 : * guarding against this, though.
3525 : */
3526 578296 : if (pg_atomic_read_u32(&bufHdr->state) & BM_CHECKPOINT_NEEDED)
3527 : {
3528 539178 : if (SyncOneBuffer(buf_id, false, &wb_context) & BUF_WRITTEN)
3529 : {
3530 : TRACE_POSTGRESQL_BUFFER_SYNC_WRITTEN(buf_id);
3531 539178 : PendingCheckpointerStats.buffers_written++;
3532 539178 : num_written++;
3533 : }
3534 : }
3535 :
3536 : /*
3537 : * Measure progress independent of actually having to flush the buffer
3538 : * - otherwise writing become unbalanced.
3539 : */
3540 578296 : ts_stat->progress += ts_stat->progress_slice;
3541 578296 : ts_stat->num_scanned++;
3542 578296 : ts_stat->index++;
3543 :
3544 : /* Have all the buffers from the tablespace been processed? */
3545 578296 : if (ts_stat->num_scanned == ts_stat->num_to_scan)
3546 : {
3547 3132 : binaryheap_remove_first(ts_heap);
3548 : }
3549 : else
3550 : {
3551 : /* update heap with the new progress */
3552 575164 : binaryheap_replace_first(ts_heap, PointerGetDatum(ts_stat));
3553 : }
3554 :
3555 : /*
3556 : * Sleep to throttle our I/O rate.
3557 : *
3558 : * (This will check for barrier events even if it doesn't sleep.)
3559 : */
3560 578296 : CheckpointWriteDelay(flags, (double) num_processed / num_to_scan);
3561 : }
3562 :
3563 : /*
3564 : * Issue all pending flushes. Only checkpointer calls BufferSync(), so
3565 : * IOContext will always be IOCONTEXT_NORMAL.
3566 : */
3567 2086 : IssuePendingWritebacks(&wb_context, IOCONTEXT_NORMAL);
3568 :
3569 2086 : pfree(per_ts_stat);
3570 2086 : per_ts_stat = NULL;
3571 2086 : binaryheap_free(ts_heap);
3572 :
3573 : /*
3574 : * Update checkpoint statistics. As noted above, this doesn't include
3575 : * buffers written by other backends or bgwriter scan.
3576 : */
3577 2086 : CheckpointStats.ckpt_bufs_written += num_written;
3578 :
3579 : TRACE_POSTGRESQL_BUFFER_SYNC_DONE(NBuffers, num_written, num_to_scan);
3580 : }
3581 :
3582 : /*
3583 : * BgBufferSync -- Write out some dirty buffers in the pool.
3584 : *
3585 : * This is called periodically by the background writer process.
3586 : *
3587 : * Returns true if it's appropriate for the bgwriter process to go into
3588 : * low-power hibernation mode. (This happens if the strategy clock-sweep
3589 : * has been "lapped" and no buffer allocations have occurred recently,
3590 : * or if the bgwriter has been effectively disabled by setting
3591 : * bgwriter_lru_maxpages to 0.)
3592 : */
3593 : bool
3594 24236 : BgBufferSync(WritebackContext *wb_context)
3595 : {
3596 : /* info obtained from freelist.c */
3597 : int strategy_buf_id;
3598 : uint32 strategy_passes;
3599 : uint32 recent_alloc;
3600 :
3601 : /*
3602 : * Information saved between calls so we can determine the strategy
3603 : * point's advance rate and avoid scanning already-cleaned buffers.
3604 : */
3605 : static bool saved_info_valid = false;
3606 : static int prev_strategy_buf_id;
3607 : static uint32 prev_strategy_passes;
3608 : static int next_to_clean;
3609 : static uint32 next_passes;
3610 :
3611 : /* Moving averages of allocation rate and clean-buffer density */
3612 : static float smoothed_alloc = 0;
3613 : static float smoothed_density = 10.0;
3614 :
3615 : /* Potentially these could be tunables, but for now, not */
3616 24236 : float smoothing_samples = 16;
3617 24236 : float scan_whole_pool_milliseconds = 120000.0;
3618 :
3619 : /* Used to compute how far we scan ahead */
3620 : long strategy_delta;
3621 : int bufs_to_lap;
3622 : int bufs_ahead;
3623 : float scans_per_alloc;
3624 : int reusable_buffers_est;
3625 : int upcoming_alloc_est;
3626 : int min_scan_buffers;
3627 :
3628 : /* Variables for the scanning loop proper */
3629 : int num_to_scan;
3630 : int num_written;
3631 : int reusable_buffers;
3632 :
3633 : /* Variables for final smoothed_density update */
3634 : long new_strategy_delta;
3635 : uint32 new_recent_alloc;
3636 :
3637 : /*
3638 : * Find out where the clock-sweep currently is, and how many buffer
3639 : * allocations have happened since our last call.
3640 : */
3641 24236 : strategy_buf_id = StrategySyncStart(&strategy_passes, &recent_alloc);
3642 :
3643 : /* Report buffer alloc counts to pgstat */
3644 24236 : PendingBgWriterStats.buf_alloc += recent_alloc;
3645 :
3646 : /*
3647 : * If we're not running the LRU scan, just stop after doing the stats
3648 : * stuff. We mark the saved state invalid so that we can recover sanely
3649 : * if LRU scan is turned back on later.
3650 : */
3651 24236 : if (bgwriter_lru_maxpages <= 0)
3652 : {
3653 76 : saved_info_valid = false;
3654 76 : return true;
3655 : }
3656 :
3657 : /*
3658 : * Compute strategy_delta = how many buffers have been scanned by the
3659 : * clock-sweep since last time. If first time through, assume none. Then
3660 : * see if we are still ahead of the clock-sweep, and if so, how many
3661 : * buffers we could scan before we'd catch up with it and "lap" it. Note:
3662 : * weird-looking coding of xxx_passes comparisons are to avoid bogus
3663 : * behavior when the passes counts wrap around.
3664 : */
3665 24160 : if (saved_info_valid)
3666 : {
3667 23108 : int32 passes_delta = strategy_passes - prev_strategy_passes;
3668 :
3669 23108 : strategy_delta = strategy_buf_id - prev_strategy_buf_id;
3670 23108 : strategy_delta += (long) passes_delta * NBuffers;
3671 :
3672 : Assert(strategy_delta >= 0);
3673 :
3674 23108 : if ((int32) (next_passes - strategy_passes) > 0)
3675 : {
3676 : /* we're one pass ahead of the strategy point */
3677 4520 : bufs_to_lap = strategy_buf_id - next_to_clean;
3678 : #ifdef BGW_DEBUG
3679 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3680 : next_passes, next_to_clean,
3681 : strategy_passes, strategy_buf_id,
3682 : strategy_delta, bufs_to_lap);
3683 : #endif
3684 : }
3685 18588 : else if (next_passes == strategy_passes &&
3686 14244 : next_to_clean >= strategy_buf_id)
3687 : {
3688 : /* on same pass, but ahead or at least not behind */
3689 12794 : bufs_to_lap = NBuffers - (next_to_clean - strategy_buf_id);
3690 : #ifdef BGW_DEBUG
3691 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3692 : next_passes, next_to_clean,
3693 : strategy_passes, strategy_buf_id,
3694 : strategy_delta, bufs_to_lap);
3695 : #endif
3696 : }
3697 : else
3698 : {
3699 : /*
3700 : * We're behind, so skip forward to the strategy point and start
3701 : * cleaning from there.
3702 : */
3703 : #ifdef BGW_DEBUG
3704 : elog(DEBUG2, "bgwriter behind: bgw %u-%u strategy %u-%u delta=%ld",
3705 : next_passes, next_to_clean,
3706 : strategy_passes, strategy_buf_id,
3707 : strategy_delta);
3708 : #endif
3709 5794 : next_to_clean = strategy_buf_id;
3710 5794 : next_passes = strategy_passes;
3711 5794 : bufs_to_lap = NBuffers;
3712 : }
3713 : }
3714 : else
3715 : {
3716 : /*
3717 : * Initializing at startup or after LRU scanning had been off. Always
3718 : * start at the strategy point.
3719 : */
3720 : #ifdef BGW_DEBUG
3721 : elog(DEBUG2, "bgwriter initializing: strategy %u-%u",
3722 : strategy_passes, strategy_buf_id);
3723 : #endif
3724 1052 : strategy_delta = 0;
3725 1052 : next_to_clean = strategy_buf_id;
3726 1052 : next_passes = strategy_passes;
3727 1052 : bufs_to_lap = NBuffers;
3728 : }
3729 :
3730 : /* Update saved info for next time */
3731 24160 : prev_strategy_buf_id = strategy_buf_id;
3732 24160 : prev_strategy_passes = strategy_passes;
3733 24160 : saved_info_valid = true;
3734 :
3735 : /*
3736 : * Compute how many buffers had to be scanned for each new allocation, ie,
3737 : * 1/density of reusable buffers, and track a moving average of that.
3738 : *
3739 : * If the strategy point didn't move, we don't update the density estimate
3740 : */
3741 24160 : if (strategy_delta > 0 && recent_alloc > 0)
3742 : {
3743 12924 : scans_per_alloc = (float) strategy_delta / (float) recent_alloc;
3744 12924 : smoothed_density += (scans_per_alloc - smoothed_density) /
3745 : smoothing_samples;
3746 : }
3747 :
3748 : /*
3749 : * Estimate how many reusable buffers there are between the current
3750 : * strategy point and where we've scanned ahead to, based on the smoothed
3751 : * density estimate.
3752 : */
3753 24160 : bufs_ahead = NBuffers - bufs_to_lap;
3754 24160 : reusable_buffers_est = (float) bufs_ahead / smoothed_density;
3755 :
3756 : /*
3757 : * Track a moving average of recent buffer allocations. Here, rather than
3758 : * a true average we want a fast-attack, slow-decline behavior: we
3759 : * immediately follow any increase.
3760 : */
3761 24160 : if (smoothed_alloc <= (float) recent_alloc)
3762 6504 : smoothed_alloc = recent_alloc;
3763 : else
3764 17656 : smoothed_alloc += ((float) recent_alloc - smoothed_alloc) /
3765 : smoothing_samples;
3766 :
3767 : /* Scale the estimate by a GUC to allow more aggressive tuning. */
3768 24160 : upcoming_alloc_est = (int) (smoothed_alloc * bgwriter_lru_multiplier);
3769 :
3770 : /*
3771 : * If recent_alloc remains at zero for many cycles, smoothed_alloc will
3772 : * eventually underflow to zero, and the underflows produce annoying
3773 : * kernel warnings on some platforms. Once upcoming_alloc_est has gone to
3774 : * zero, there's no point in tracking smaller and smaller values of
3775 : * smoothed_alloc, so just reset it to exactly zero to avoid this
3776 : * syndrome. It will pop back up as soon as recent_alloc increases.
3777 : */
3778 24160 : if (upcoming_alloc_est == 0)
3779 3842 : smoothed_alloc = 0;
3780 :
3781 : /*
3782 : * Even in cases where there's been little or no buffer allocation
3783 : * activity, we want to make a small amount of progress through the buffer
3784 : * cache so that as many reusable buffers as possible are clean after an
3785 : * idle period.
3786 : *
3787 : * (scan_whole_pool_milliseconds / BgWriterDelay) computes how many times
3788 : * the BGW will be called during the scan_whole_pool time; slice the
3789 : * buffer pool into that many sections.
3790 : */
3791 24160 : min_scan_buffers = (int) (NBuffers / (scan_whole_pool_milliseconds / BgWriterDelay));
3792 :
3793 24160 : if (upcoming_alloc_est < (min_scan_buffers + reusable_buffers_est))
3794 : {
3795 : #ifdef BGW_DEBUG
3796 : elog(DEBUG2, "bgwriter: alloc_est=%d too small, using min=%d + reusable_est=%d",
3797 : upcoming_alloc_est, min_scan_buffers, reusable_buffers_est);
3798 : #endif
3799 10734 : upcoming_alloc_est = min_scan_buffers + reusable_buffers_est;
3800 : }
3801 :
3802 : /*
3803 : * Now write out dirty reusable buffers, working forward from the
3804 : * next_to_clean point, until we have lapped the strategy scan, or cleaned
3805 : * enough buffers to match our estimate of the next cycle's allocation
3806 : * requirements, or hit the bgwriter_lru_maxpages limit.
3807 : */
3808 :
3809 24160 : num_to_scan = bufs_to_lap;
3810 24160 : num_written = 0;
3811 24160 : reusable_buffers = reusable_buffers_est;
3812 :
3813 : /* Execute the LRU scan */
3814 3553460 : while (num_to_scan > 0 && reusable_buffers < upcoming_alloc_est)
3815 : {
3816 3529304 : int sync_state = SyncOneBuffer(next_to_clean, true,
3817 : wb_context);
3818 :
3819 3529304 : if (++next_to_clean >= NBuffers)
3820 : {
3821 5466 : next_to_clean = 0;
3822 5466 : next_passes++;
3823 : }
3824 3529304 : num_to_scan--;
3825 :
3826 3529304 : if (sync_state & BUF_WRITTEN)
3827 : {
3828 46410 : reusable_buffers++;
3829 46410 : if (++num_written >= bgwriter_lru_maxpages)
3830 : {
3831 4 : PendingBgWriterStats.maxwritten_clean++;
3832 4 : break;
3833 : }
3834 : }
3835 3482894 : else if (sync_state & BUF_REUSABLE)
3836 2725620 : reusable_buffers++;
3837 : }
3838 :
3839 24160 : PendingBgWriterStats.buf_written_clean += num_written;
3840 :
3841 : #ifdef BGW_DEBUG
3842 : 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",
3843 : recent_alloc, smoothed_alloc, strategy_delta, bufs_ahead,
3844 : smoothed_density, reusable_buffers_est, upcoming_alloc_est,
3845 : bufs_to_lap - num_to_scan,
3846 : num_written,
3847 : reusable_buffers - reusable_buffers_est);
3848 : #endif
3849 :
3850 : /*
3851 : * Consider the above scan as being like a new allocation scan.
3852 : * Characterize its density and update the smoothed one based on it. This
3853 : * effectively halves the moving average period in cases where both the
3854 : * strategy and the background writer are doing some useful scanning,
3855 : * which is helpful because a long memory isn't as desirable on the
3856 : * density estimates.
3857 : */
3858 24160 : new_strategy_delta = bufs_to_lap - num_to_scan;
3859 24160 : new_recent_alloc = reusable_buffers - reusable_buffers_est;
3860 24160 : if (new_strategy_delta > 0 && new_recent_alloc > 0)
3861 : {
3862 19694 : scans_per_alloc = (float) new_strategy_delta / (float) new_recent_alloc;
3863 19694 : smoothed_density += (scans_per_alloc - smoothed_density) /
3864 : smoothing_samples;
3865 :
3866 : #ifdef BGW_DEBUG
3867 : elog(DEBUG2, "bgwriter: cleaner density alloc=%u scan=%ld density=%.2f new smoothed=%.2f",
3868 : new_recent_alloc, new_strategy_delta,
3869 : scans_per_alloc, smoothed_density);
3870 : #endif
3871 : }
3872 :
3873 : /* Return true if OK to hibernate */
3874 24160 : return (bufs_to_lap == 0 && recent_alloc == 0);
3875 : }
3876 :
3877 : /*
3878 : * SyncOneBuffer -- process a single buffer during syncing.
3879 : *
3880 : * If skip_recently_used is true, we don't write currently-pinned buffers, nor
3881 : * buffers marked recently used, as these are not replacement candidates.
3882 : *
3883 : * Returns a bitmask containing the following flag bits:
3884 : * BUF_WRITTEN: we wrote the buffer.
3885 : * BUF_REUSABLE: buffer is available for replacement, ie, it has
3886 : * pin count 0 and usage count 0.
3887 : *
3888 : * (BUF_WRITTEN could be set in error if FlushBuffer finds the buffer clean
3889 : * after locking it, but we don't care all that much.)
3890 : */
3891 : static int
3892 4068482 : SyncOneBuffer(int buf_id, bool skip_recently_used, WritebackContext *wb_context)
3893 : {
3894 4068482 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3895 4068482 : int result = 0;
3896 : uint32 buf_state;
3897 : BufferTag tag;
3898 :
3899 : /* Make sure we can handle the pin */
3900 4068482 : ReservePrivateRefCountEntry();
3901 4068482 : ResourceOwnerEnlarge(CurrentResourceOwner);
3902 :
3903 : /*
3904 : * Check whether buffer needs writing.
3905 : *
3906 : * We can make this check without taking the buffer content lock so long
3907 : * as we mark pages dirty in access methods *before* logging changes with
3908 : * XLogInsert(): if someone marks the buffer dirty just after our check we
3909 : * don't worry because our checkpoint.redo points before log record for
3910 : * upcoming changes and so we are not required to write such dirty buffer.
3911 : */
3912 4068482 : buf_state = LockBufHdr(bufHdr);
3913 :
3914 4068482 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 0 &&
3915 4059204 : BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3916 : {
3917 2776142 : result |= BUF_REUSABLE;
3918 : }
3919 1292340 : else if (skip_recently_used)
3920 : {
3921 : /* Caller told us not to write recently-used buffers */
3922 757274 : UnlockBufHdr(bufHdr, buf_state);
3923 757274 : return result;
3924 : }
3925 :
3926 3311208 : if (!(buf_state & BM_VALID) || !(buf_state & BM_DIRTY))
3927 : {
3928 : /* It's clean, so nothing to do */
3929 2725620 : UnlockBufHdr(bufHdr, buf_state);
3930 2725620 : return result;
3931 : }
3932 :
3933 : /*
3934 : * Pin it, share-lock it, write it. (FlushBuffer will do nothing if the
3935 : * buffer is clean by the time we've locked it.)
3936 : */
3937 585588 : PinBuffer_Locked(bufHdr);
3938 585588 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_SHARED);
3939 :
3940 585588 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
3941 :
3942 585588 : LWLockRelease(BufferDescriptorGetContentLock(bufHdr));
3943 :
3944 585588 : tag = bufHdr->tag;
3945 :
3946 585588 : UnpinBuffer(bufHdr);
3947 :
3948 : /*
3949 : * SyncOneBuffer() is only called by checkpointer and bgwriter, so
3950 : * IOContext will always be IOCONTEXT_NORMAL.
3951 : */
3952 585588 : ScheduleBufferTagForWriteback(wb_context, IOCONTEXT_NORMAL, &tag);
3953 :
3954 585588 : return result | BUF_WRITTEN;
3955 : }
3956 :
3957 : /*
3958 : * AtEOXact_Buffers - clean up at end of transaction.
3959 : *
3960 : * As of PostgreSQL 8.0, buffer pins should get released by the
3961 : * ResourceOwner mechanism. This routine is just a debugging
3962 : * cross-check that no pins remain.
3963 : */
3964 : void
3965 1142280 : AtEOXact_Buffers(bool isCommit)
3966 : {
3967 1142280 : CheckForBufferLeaks();
3968 :
3969 1142280 : AtEOXact_LocalBuffers(isCommit);
3970 :
3971 : Assert(PrivateRefCountOverflowed == 0);
3972 1142280 : }
3973 :
3974 : /*
3975 : * Initialize access to shared buffer pool
3976 : *
3977 : * This is called during backend startup (whether standalone or under the
3978 : * postmaster). It sets up for this backend's access to the already-existing
3979 : * buffer pool.
3980 : */
3981 : void
3982 44504 : InitBufferManagerAccess(void)
3983 : {
3984 : HASHCTL hash_ctl;
3985 :
3986 : /*
3987 : * An advisory limit on the number of pins each backend should hold, based
3988 : * on shared_buffers and the maximum number of connections possible.
3989 : * That's very pessimistic, but outside toy-sized shared_buffers it should
3990 : * allow plenty of pins. LimitAdditionalPins() and
3991 : * GetAdditionalPinLimit() can be used to check the remaining balance.
3992 : */
3993 44504 : MaxProportionalPins = NBuffers / (MaxBackends + NUM_AUXILIARY_PROCS);
3994 :
3995 44504 : memset(&PrivateRefCountArray, 0, sizeof(PrivateRefCountArray));
3996 :
3997 44504 : hash_ctl.keysize = sizeof(int32);
3998 44504 : hash_ctl.entrysize = sizeof(PrivateRefCountEntry);
3999 :
4000 44504 : PrivateRefCountHash = hash_create("PrivateRefCount", 100, &hash_ctl,
4001 : HASH_ELEM | HASH_BLOBS);
4002 :
4003 : /*
4004 : * AtProcExit_Buffers needs LWLock access, and thereby has to be called at
4005 : * the corresponding phase of backend shutdown.
4006 : */
4007 : Assert(MyProc != NULL);
4008 44504 : on_shmem_exit(AtProcExit_Buffers, 0);
4009 44504 : }
4010 :
4011 : /*
4012 : * During backend exit, ensure that we released all shared-buffer locks and
4013 : * assert that we have no remaining pins.
4014 : */
4015 : static void
4016 44506 : AtProcExit_Buffers(int code, Datum arg)
4017 : {
4018 44506 : UnlockBuffers();
4019 :
4020 44506 : CheckForBufferLeaks();
4021 :
4022 : /* localbuf.c needs a chance too */
4023 44506 : AtProcExit_LocalBuffers();
4024 44506 : }
4025 :
4026 : /*
4027 : * CheckForBufferLeaks - ensure this backend holds no buffer pins
4028 : *
4029 : * As of PostgreSQL 8.0, buffer pins should get released by the
4030 : * ResourceOwner mechanism. This routine is just a debugging
4031 : * cross-check that no pins remain.
4032 : */
4033 : static void
4034 1186790 : CheckForBufferLeaks(void)
4035 : {
4036 : #ifdef USE_ASSERT_CHECKING
4037 : int RefCountErrors = 0;
4038 : PrivateRefCountEntry *res;
4039 : int i;
4040 : char *s;
4041 :
4042 : /* check the array */
4043 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4044 : {
4045 : res = &PrivateRefCountArray[i];
4046 :
4047 : if (res->buffer != InvalidBuffer)
4048 : {
4049 : s = DebugPrintBufferRefcount(res->buffer);
4050 : elog(WARNING, "buffer refcount leak: %s", s);
4051 : pfree(s);
4052 :
4053 : RefCountErrors++;
4054 : }
4055 : }
4056 :
4057 : /* if necessary search the hash */
4058 : if (PrivateRefCountOverflowed)
4059 : {
4060 : HASH_SEQ_STATUS hstat;
4061 :
4062 : hash_seq_init(&hstat, PrivateRefCountHash);
4063 : while ((res = (PrivateRefCountEntry *) hash_seq_search(&hstat)) != NULL)
4064 : {
4065 : s = DebugPrintBufferRefcount(res->buffer);
4066 : elog(WARNING, "buffer refcount leak: %s", s);
4067 : pfree(s);
4068 : RefCountErrors++;
4069 : }
4070 : }
4071 :
4072 : Assert(RefCountErrors == 0);
4073 : #endif
4074 1186790 : }
4075 :
4076 : #ifdef USE_ASSERT_CHECKING
4077 : /*
4078 : * Check for exclusive-locked catalog buffers. This is the core of
4079 : * AssertCouldGetRelation().
4080 : *
4081 : * A backend would self-deadlock on LWLocks if the catalog scan read the
4082 : * exclusive-locked buffer. The main threat is exclusive-locked buffers of
4083 : * catalogs used in relcache, because a catcache search on any catalog may
4084 : * build that catalog's relcache entry. We don't have an inventory of
4085 : * catalogs relcache uses, so just check buffers of most catalogs.
4086 : *
4087 : * It's better to minimize waits while holding an exclusive buffer lock, so it
4088 : * would be nice to broaden this check not to be catalog-specific. However,
4089 : * bttextcmp() accesses pg_collation, and non-core opclasses might similarly
4090 : * read tables. That is deadlock-free as long as there's no loop in the
4091 : * dependency graph: modifying table A may cause an opclass to read table B,
4092 : * but it must not cause a read of table A.
4093 : */
4094 : void
4095 : AssertBufferLocksPermitCatalogRead(void)
4096 : {
4097 : ForEachLWLockHeldByMe(AssertNotCatalogBufferLock, NULL);
4098 : }
4099 :
4100 : static void
4101 : AssertNotCatalogBufferLock(LWLock *lock, LWLockMode mode,
4102 : void *unused_context)
4103 : {
4104 : BufferDesc *bufHdr;
4105 : BufferTag tag;
4106 : Oid relid;
4107 :
4108 : if (mode != LW_EXCLUSIVE)
4109 : return;
4110 :
4111 : if (!((BufferDescPadded *) lock > BufferDescriptors &&
4112 : (BufferDescPadded *) lock < BufferDescriptors + NBuffers))
4113 : return; /* not a buffer lock */
4114 :
4115 : bufHdr = (BufferDesc *)
4116 : ((char *) lock - offsetof(BufferDesc, content_lock));
4117 : tag = bufHdr->tag;
4118 :
4119 : /*
4120 : * This relNumber==relid assumption holds until a catalog experiences
4121 : * VACUUM FULL or similar. After a command like that, relNumber will be
4122 : * in the normal (non-catalog) range, and we lose the ability to detect
4123 : * hazardous access to that catalog. Calling RelidByRelfilenumber() would
4124 : * close that gap, but RelidByRelfilenumber() might then deadlock with a
4125 : * held lock.
4126 : */
4127 : relid = tag.relNumber;
4128 :
4129 : if (IsCatalogTextUniqueIndexOid(relid)) /* see comments at the callee */
4130 : return;
4131 :
4132 : Assert(!IsCatalogRelationOid(relid));
4133 : }
4134 : #endif
4135 :
4136 :
4137 : /*
4138 : * Helper routine to issue warnings when a buffer is unexpectedly pinned
4139 : */
4140 : char *
4141 80 : DebugPrintBufferRefcount(Buffer buffer)
4142 : {
4143 : BufferDesc *buf;
4144 : int32 loccount;
4145 : char *result;
4146 : ProcNumber backend;
4147 : uint32 buf_state;
4148 :
4149 : Assert(BufferIsValid(buffer));
4150 80 : if (BufferIsLocal(buffer))
4151 : {
4152 32 : buf = GetLocalBufferDescriptor(-buffer - 1);
4153 32 : loccount = LocalRefCount[-buffer - 1];
4154 32 : backend = MyProcNumber;
4155 : }
4156 : else
4157 : {
4158 48 : buf = GetBufferDescriptor(buffer - 1);
4159 48 : loccount = GetPrivateRefCount(buffer);
4160 48 : backend = INVALID_PROC_NUMBER;
4161 : }
4162 :
4163 : /* theoretically we should lock the bufhdr here */
4164 80 : buf_state = pg_atomic_read_u32(&buf->state);
4165 :
4166 80 : result = psprintf("[%03d] (rel=%s, blockNum=%u, flags=0x%x, refcount=%u %d)",
4167 : buffer,
4168 80 : relpathbackend(BufTagGetRelFileLocator(&buf->tag), backend,
4169 : BufTagGetForkNum(&buf->tag)).str,
4170 : buf->tag.blockNum, buf_state & BUF_FLAG_MASK,
4171 : BUF_STATE_GET_REFCOUNT(buf_state), loccount);
4172 80 : return result;
4173 : }
4174 :
4175 : /*
4176 : * CheckPointBuffers
4177 : *
4178 : * Flush all dirty blocks in buffer pool to disk at checkpoint time.
4179 : *
4180 : * Note: temporary relations do not participate in checkpoints, so they don't
4181 : * need to be flushed.
4182 : */
4183 : void
4184 3414 : CheckPointBuffers(int flags)
4185 : {
4186 3414 : BufferSync(flags);
4187 3414 : }
4188 :
4189 : /*
4190 : * BufferGetBlockNumber
4191 : * Returns the block number associated with a buffer.
4192 : *
4193 : * Note:
4194 : * Assumes that the buffer is valid and pinned, else the
4195 : * value may be obsolete immediately...
4196 : */
4197 : BlockNumber
4198 100812846 : BufferGetBlockNumber(Buffer buffer)
4199 : {
4200 : BufferDesc *bufHdr;
4201 :
4202 : Assert(BufferIsPinned(buffer));
4203 :
4204 100812846 : if (BufferIsLocal(buffer))
4205 3807692 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4206 : else
4207 97005154 : bufHdr = GetBufferDescriptor(buffer - 1);
4208 :
4209 : /* pinned, so OK to read tag without spinlock */
4210 100812846 : return bufHdr->tag.blockNum;
4211 : }
4212 :
4213 : /*
4214 : * BufferGetTag
4215 : * Returns the relfilelocator, fork number and block number associated with
4216 : * a buffer.
4217 : */
4218 : void
4219 29998000 : BufferGetTag(Buffer buffer, RelFileLocator *rlocator, ForkNumber *forknum,
4220 : BlockNumber *blknum)
4221 : {
4222 : BufferDesc *bufHdr;
4223 :
4224 : /* Do the same checks as BufferGetBlockNumber. */
4225 : Assert(BufferIsPinned(buffer));
4226 :
4227 29998000 : if (BufferIsLocal(buffer))
4228 0 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4229 : else
4230 29998000 : bufHdr = GetBufferDescriptor(buffer - 1);
4231 :
4232 : /* pinned, so OK to read tag without spinlock */
4233 29998000 : *rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
4234 29998000 : *forknum = BufTagGetForkNum(&bufHdr->tag);
4235 29998000 : *blknum = bufHdr->tag.blockNum;
4236 29998000 : }
4237 :
4238 : /*
4239 : * FlushBuffer
4240 : * Physically write out a shared buffer.
4241 : *
4242 : * NOTE: this actually just passes the buffer contents to the kernel; the
4243 : * real write to disk won't happen until the kernel feels like it. This
4244 : * is okay from our point of view since we can redo the changes from WAL.
4245 : * However, we will need to force the changes to disk via fsync before
4246 : * we can checkpoint WAL.
4247 : *
4248 : * The caller must hold a pin on the buffer and have share-locked the
4249 : * buffer contents. (Note: a share-lock does not prevent updates of
4250 : * hint bits in the buffer, so the page could change while the write
4251 : * is in progress, but we assume that that will not invalidate the data
4252 : * written.)
4253 : *
4254 : * If the caller has an smgr reference for the buffer's relation, pass it
4255 : * as the second parameter. If not, pass NULL.
4256 : */
4257 : static void
4258 1116136 : FlushBuffer(BufferDesc *buf, SMgrRelation reln, IOObject io_object,
4259 : IOContext io_context)
4260 : {
4261 : XLogRecPtr recptr;
4262 : ErrorContextCallback errcallback;
4263 : instr_time io_start;
4264 : Block bufBlock;
4265 : char *bufToWrite;
4266 : uint32 buf_state;
4267 :
4268 : /*
4269 : * Try to start an I/O operation. If StartBufferIO returns false, then
4270 : * someone else flushed the buffer before we could, so we need not do
4271 : * anything.
4272 : */
4273 1116136 : if (!StartBufferIO(buf, false, false))
4274 18 : return;
4275 :
4276 : /* Setup error traceback support for ereport() */
4277 1116118 : errcallback.callback = shared_buffer_write_error_callback;
4278 1116118 : errcallback.arg = buf;
4279 1116118 : errcallback.previous = error_context_stack;
4280 1116118 : error_context_stack = &errcallback;
4281 :
4282 : /* Find smgr relation for buffer */
4283 1116118 : if (reln == NULL)
4284 1107844 : reln = smgropen(BufTagGetRelFileLocator(&buf->tag), INVALID_PROC_NUMBER);
4285 :
4286 : TRACE_POSTGRESQL_BUFFER_FLUSH_START(BufTagGetForkNum(&buf->tag),
4287 : buf->tag.blockNum,
4288 : reln->smgr_rlocator.locator.spcOid,
4289 : reln->smgr_rlocator.locator.dbOid,
4290 : reln->smgr_rlocator.locator.relNumber);
4291 :
4292 1116118 : buf_state = LockBufHdr(buf);
4293 :
4294 : /*
4295 : * Run PageGetLSN while holding header lock, since we don't have the
4296 : * buffer locked exclusively in all cases.
4297 : */
4298 1116118 : recptr = BufferGetLSN(buf);
4299 :
4300 : /* To check if block content changes while flushing. - vadim 01/17/97 */
4301 1116118 : buf_state &= ~BM_JUST_DIRTIED;
4302 1116118 : UnlockBufHdr(buf, buf_state);
4303 :
4304 : /*
4305 : * Force XLOG flush up to buffer's LSN. This implements the basic WAL
4306 : * rule that log updates must hit disk before any of the data-file changes
4307 : * they describe do.
4308 : *
4309 : * However, this rule does not apply to unlogged relations, which will be
4310 : * lost after a crash anyway. Most unlogged relation pages do not bear
4311 : * LSNs since we never emit WAL records for them, and therefore flushing
4312 : * up through the buffer LSN would be useless, but harmless. However,
4313 : * GiST indexes use LSNs internally to track page-splits, and therefore
4314 : * unlogged GiST pages bear "fake" LSNs generated by
4315 : * GetFakeLSNForUnloggedRel. It is unlikely but possible that the fake
4316 : * LSN counter could advance past the WAL insertion point; and if it did
4317 : * happen, attempting to flush WAL through that location would fail, with
4318 : * disastrous system-wide consequences. To make sure that can't happen,
4319 : * skip the flush if the buffer isn't permanent.
4320 : */
4321 1116118 : if (buf_state & BM_PERMANENT)
4322 1112566 : XLogFlush(recptr);
4323 :
4324 : /*
4325 : * Now it's safe to write the buffer to disk. Note that no one else should
4326 : * have been able to write it, while we were busy with log flushing,
4327 : * because we got the exclusive right to perform I/O by setting the
4328 : * BM_IO_IN_PROGRESS bit.
4329 : */
4330 1116118 : bufBlock = BufHdrGetBlock(buf);
4331 :
4332 : /*
4333 : * Update page checksum if desired. Since we have only shared lock on the
4334 : * buffer, other processes might be updating hint bits in it, so we must
4335 : * copy the page to private storage if we do checksumming.
4336 : */
4337 1116118 : bufToWrite = PageSetChecksumCopy((Page) bufBlock, buf->tag.blockNum);
4338 :
4339 1116118 : io_start = pgstat_prepare_io_time(track_io_timing);
4340 :
4341 : /*
4342 : * bufToWrite is either the shared buffer or a copy, as appropriate.
4343 : */
4344 1116118 : smgrwrite(reln,
4345 1116118 : BufTagGetForkNum(&buf->tag),
4346 : buf->tag.blockNum,
4347 : bufToWrite,
4348 : false);
4349 :
4350 : /*
4351 : * When a strategy is in use, only flushes of dirty buffers already in the
4352 : * strategy ring are counted as strategy writes (IOCONTEXT
4353 : * [BULKREAD|BULKWRITE|VACUUM] IOOP_WRITE) for the purpose of IO
4354 : * statistics tracking.
4355 : *
4356 : * If a shared buffer initially added to the ring must be flushed before
4357 : * being used, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE.
4358 : *
4359 : * If a shared buffer which was added to the ring later because the
4360 : * current strategy buffer is pinned or in use or because all strategy
4361 : * buffers were dirty and rejected (for BAS_BULKREAD operations only)
4362 : * requires flushing, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE
4363 : * (from_ring will be false).
4364 : *
4365 : * When a strategy is not in use, the write can only be a "regular" write
4366 : * of a dirty shared buffer (IOCONTEXT_NORMAL IOOP_WRITE).
4367 : */
4368 1116118 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
4369 : IOOP_WRITE, io_start, 1, BLCKSZ);
4370 :
4371 1116118 : pgBufferUsage.shared_blks_written++;
4372 :
4373 : /*
4374 : * Mark the buffer as clean (unless BM_JUST_DIRTIED has become set) and
4375 : * end the BM_IO_IN_PROGRESS state.
4376 : */
4377 1116118 : TerminateBufferIO(buf, true, 0, true, false);
4378 :
4379 : TRACE_POSTGRESQL_BUFFER_FLUSH_DONE(BufTagGetForkNum(&buf->tag),
4380 : buf->tag.blockNum,
4381 : reln->smgr_rlocator.locator.spcOid,
4382 : reln->smgr_rlocator.locator.dbOid,
4383 : reln->smgr_rlocator.locator.relNumber);
4384 :
4385 : /* Pop the error context stack */
4386 1116118 : error_context_stack = errcallback.previous;
4387 : }
4388 :
4389 : /*
4390 : * RelationGetNumberOfBlocksInFork
4391 : * Determines the current number of pages in the specified relation fork.
4392 : *
4393 : * Note that the accuracy of the result will depend on the details of the
4394 : * relation's storage. For builtin AMs it'll be accurate, but for external AMs
4395 : * it might not be.
4396 : */
4397 : BlockNumber
4398 3696602 : RelationGetNumberOfBlocksInFork(Relation relation, ForkNumber forkNum)
4399 : {
4400 3696602 : if (RELKIND_HAS_TABLE_AM(relation->rd_rel->relkind))
4401 : {
4402 : /*
4403 : * Not every table AM uses BLCKSZ wide fixed size blocks. Therefore
4404 : * tableam returns the size in bytes - but for the purpose of this
4405 : * routine, we want the number of blocks. Therefore divide, rounding
4406 : * up.
4407 : */
4408 : uint64 szbytes;
4409 :
4410 2822178 : szbytes = table_relation_size(relation, forkNum);
4411 :
4412 2822140 : return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
4413 : }
4414 874424 : else if (RELKIND_HAS_STORAGE(relation->rd_rel->relkind))
4415 : {
4416 874424 : return smgrnblocks(RelationGetSmgr(relation), forkNum);
4417 : }
4418 : else
4419 : Assert(false);
4420 :
4421 0 : return 0; /* keep compiler quiet */
4422 : }
4423 :
4424 : /*
4425 : * BufferIsPermanent
4426 : * Determines whether a buffer will potentially still be around after
4427 : * a crash. Caller must hold a buffer pin.
4428 : */
4429 : bool
4430 18809034 : BufferIsPermanent(Buffer buffer)
4431 : {
4432 : BufferDesc *bufHdr;
4433 :
4434 : /* Local buffers are used only for temp relations. */
4435 18809034 : if (BufferIsLocal(buffer))
4436 1253868 : return false;
4437 :
4438 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4439 : Assert(BufferIsValid(buffer));
4440 : Assert(BufferIsPinned(buffer));
4441 :
4442 : /*
4443 : * BM_PERMANENT can't be changed while we hold a pin on the buffer, so we
4444 : * need not bother with the buffer header spinlock. Even if someone else
4445 : * changes the buffer header state while we're doing this, the state is
4446 : * changed atomically, so we'll read the old value or the new value, but
4447 : * not random garbage.
4448 : */
4449 17555166 : bufHdr = GetBufferDescriptor(buffer - 1);
4450 17555166 : return (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT) != 0;
4451 : }
4452 :
4453 : /*
4454 : * BufferGetLSNAtomic
4455 : * Retrieves the LSN of the buffer atomically using a buffer header lock.
4456 : * This is necessary for some callers who may not have an exclusive lock
4457 : * on the buffer.
4458 : */
4459 : XLogRecPtr
4460 14178354 : BufferGetLSNAtomic(Buffer buffer)
4461 : {
4462 14178354 : char *page = BufferGetPage(buffer);
4463 : BufferDesc *bufHdr;
4464 : XLogRecPtr lsn;
4465 : uint32 buf_state;
4466 :
4467 : /*
4468 : * If we don't need locking for correctness, fastpath out.
4469 : */
4470 14178354 : if (!XLogHintBitIsNeeded() || BufferIsLocal(buffer))
4471 477480 : return PageGetLSN(page);
4472 :
4473 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4474 : Assert(BufferIsValid(buffer));
4475 : Assert(BufferIsPinned(buffer));
4476 :
4477 13700874 : bufHdr = GetBufferDescriptor(buffer - 1);
4478 13700874 : buf_state = LockBufHdr(bufHdr);
4479 13700874 : lsn = PageGetLSN(page);
4480 13700874 : UnlockBufHdr(bufHdr, buf_state);
4481 :
4482 13700874 : return lsn;
4483 : }
4484 :
4485 : /* ---------------------------------------------------------------------
4486 : * DropRelationBuffers
4487 : *
4488 : * This function removes from the buffer pool all the pages of the
4489 : * specified relation forks that have block numbers >= firstDelBlock.
4490 : * (In particular, with firstDelBlock = 0, all pages are removed.)
4491 : * Dirty pages are simply dropped, without bothering to write them
4492 : * out first. Therefore, this is NOT rollback-able, and so should be
4493 : * used only with extreme caution!
4494 : *
4495 : * Currently, this is called only from smgr.c when the underlying file
4496 : * is about to be deleted or truncated (firstDelBlock is needed for
4497 : * the truncation case). The data in the affected pages would therefore
4498 : * be deleted momentarily anyway, and there is no point in writing it.
4499 : * It is the responsibility of higher-level code to ensure that the
4500 : * deletion or truncation does not lose any data that could be needed
4501 : * later. It is also the responsibility of higher-level code to ensure
4502 : * that no other process could be trying to load more pages of the
4503 : * relation into buffers.
4504 : * --------------------------------------------------------------------
4505 : */
4506 : void
4507 1266 : DropRelationBuffers(SMgrRelation smgr_reln, ForkNumber *forkNum,
4508 : int nforks, BlockNumber *firstDelBlock)
4509 : {
4510 : int i;
4511 : int j;
4512 : RelFileLocatorBackend rlocator;
4513 : BlockNumber nForkBlock[MAX_FORKNUM];
4514 1266 : uint64 nBlocksToInvalidate = 0;
4515 :
4516 1266 : rlocator = smgr_reln->smgr_rlocator;
4517 :
4518 : /* If it's a local relation, it's localbuf.c's problem. */
4519 1266 : if (RelFileLocatorBackendIsTemp(rlocator))
4520 : {
4521 748 : if (rlocator.backend == MyProcNumber)
4522 748 : DropRelationLocalBuffers(rlocator.locator, forkNum, nforks,
4523 : firstDelBlock);
4524 :
4525 830 : return;
4526 : }
4527 :
4528 : /*
4529 : * To remove all the pages of the specified relation forks from the buffer
4530 : * pool, we need to scan the entire buffer pool but we can optimize it by
4531 : * finding the buffers from BufMapping table provided we know the exact
4532 : * size of each fork of the relation. The exact size is required to ensure
4533 : * that we don't leave any buffer for the relation being dropped as
4534 : * otherwise the background writer or checkpointer can lead to a PANIC
4535 : * error while flushing buffers corresponding to files that don't exist.
4536 : *
4537 : * To know the exact size, we rely on the size cached for each fork by us
4538 : * during recovery which limits the optimization to recovery and on
4539 : * standbys but we can easily extend it once we have shared cache for
4540 : * relation size.
4541 : *
4542 : * In recovery, we cache the value returned by the first lseek(SEEK_END)
4543 : * and the future writes keeps the cached value up-to-date. See
4544 : * smgrextend. It is possible that the value of the first lseek is smaller
4545 : * than the actual number of existing blocks in the file due to buggy
4546 : * Linux kernels that might not have accounted for the recent write. But
4547 : * that should be fine because there must not be any buffers after that
4548 : * file size.
4549 : */
4550 718 : for (i = 0; i < nforks; i++)
4551 : {
4552 : /* Get the number of blocks for a relation's fork */
4553 610 : nForkBlock[i] = smgrnblocks_cached(smgr_reln, forkNum[i]);
4554 :
4555 610 : if (nForkBlock[i] == InvalidBlockNumber)
4556 : {
4557 410 : nBlocksToInvalidate = InvalidBlockNumber;
4558 410 : break;
4559 : }
4560 :
4561 : /* calculate the number of blocks to be invalidated */
4562 200 : nBlocksToInvalidate += (nForkBlock[i] - firstDelBlock[i]);
4563 : }
4564 :
4565 : /*
4566 : * We apply the optimization iff the total number of blocks to invalidate
4567 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4568 : */
4569 518 : if (BlockNumberIsValid(nBlocksToInvalidate) &&
4570 108 : nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4571 : {
4572 222 : for (j = 0; j < nforks; j++)
4573 140 : FindAndDropRelationBuffers(rlocator.locator, forkNum[j],
4574 140 : nForkBlock[j], firstDelBlock[j]);
4575 82 : return;
4576 : }
4577 :
4578 5550772 : for (i = 0; i < NBuffers; i++)
4579 : {
4580 5550336 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4581 : uint32 buf_state;
4582 :
4583 : /*
4584 : * We can make this a tad faster by prechecking the buffer tag before
4585 : * we attempt to lock the buffer; this saves a lot of lock
4586 : * acquisitions in typical cases. It should be safe because the
4587 : * caller must have AccessExclusiveLock on the relation, or some other
4588 : * reason to be certain that no one is loading new pages of the rel
4589 : * into the buffer pool. (Otherwise we might well miss such pages
4590 : * entirely.) Therefore, while the tag might be changing while we
4591 : * look at it, it can't be changing *to* a value we care about, only
4592 : * *away* from such a value. So false negatives are impossible, and
4593 : * false positives are safe because we'll recheck after getting the
4594 : * buffer lock.
4595 : *
4596 : * We could check forkNum and blockNum as well as the rlocator, but
4597 : * the incremental win from doing so seems small.
4598 : */
4599 5550336 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator))
4600 5536494 : continue;
4601 :
4602 13842 : buf_state = LockBufHdr(bufHdr);
4603 :
4604 35780 : for (j = 0; j < nforks; j++)
4605 : {
4606 25064 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator) &&
4607 25064 : BufTagGetForkNum(&bufHdr->tag) == forkNum[j] &&
4608 13648 : bufHdr->tag.blockNum >= firstDelBlock[j])
4609 : {
4610 3126 : InvalidateBuffer(bufHdr); /* releases spinlock */
4611 3126 : break;
4612 : }
4613 : }
4614 13842 : if (j >= nforks)
4615 10716 : UnlockBufHdr(bufHdr, buf_state);
4616 : }
4617 : }
4618 :
4619 : /* ---------------------------------------------------------------------
4620 : * DropRelationsAllBuffers
4621 : *
4622 : * This function removes from the buffer pool all the pages of all
4623 : * forks of the specified relations. It's equivalent to calling
4624 : * DropRelationBuffers once per fork per relation with firstDelBlock = 0.
4625 : * --------------------------------------------------------------------
4626 : */
4627 : void
4628 26974 : DropRelationsAllBuffers(SMgrRelation *smgr_reln, int nlocators)
4629 : {
4630 : int i;
4631 26974 : int n = 0;
4632 : SMgrRelation *rels;
4633 : BlockNumber (*block)[MAX_FORKNUM + 1];
4634 26974 : uint64 nBlocksToInvalidate = 0;
4635 : RelFileLocator *locators;
4636 26974 : bool cached = true;
4637 : bool use_bsearch;
4638 :
4639 26974 : if (nlocators == 0)
4640 0 : return;
4641 :
4642 26974 : rels = palloc(sizeof(SMgrRelation) * nlocators); /* non-local relations */
4643 :
4644 : /* If it's a local relation, it's localbuf.c's problem. */
4645 117928 : for (i = 0; i < nlocators; i++)
4646 : {
4647 90954 : if (RelFileLocatorBackendIsTemp(smgr_reln[i]->smgr_rlocator))
4648 : {
4649 6300 : if (smgr_reln[i]->smgr_rlocator.backend == MyProcNumber)
4650 6300 : DropRelationAllLocalBuffers(smgr_reln[i]->smgr_rlocator.locator);
4651 : }
4652 : else
4653 84654 : rels[n++] = smgr_reln[i];
4654 : }
4655 :
4656 : /*
4657 : * If there are no non-local relations, then we're done. Release the
4658 : * memory and return.
4659 : */
4660 26974 : if (n == 0)
4661 : {
4662 1630 : pfree(rels);
4663 1630 : return;
4664 : }
4665 :
4666 : /*
4667 : * This is used to remember the number of blocks for all the relations
4668 : * forks.
4669 : */
4670 : block = (BlockNumber (*)[MAX_FORKNUM + 1])
4671 25344 : palloc(sizeof(BlockNumber) * n * (MAX_FORKNUM + 1));
4672 :
4673 : /*
4674 : * We can avoid scanning the entire buffer pool if we know the exact size
4675 : * of each of the given relation forks. See DropRelationBuffers.
4676 : */
4677 53022 : for (i = 0; i < n && cached; i++)
4678 : {
4679 43070 : for (int j = 0; j <= MAX_FORKNUM; j++)
4680 : {
4681 : /* Get the number of blocks for a relation's fork. */
4682 39252 : block[i][j] = smgrnblocks_cached(rels[i], j);
4683 :
4684 : /* We need to only consider the relation forks that exists. */
4685 39252 : if (block[i][j] == InvalidBlockNumber)
4686 : {
4687 35084 : if (!smgrexists(rels[i], j))
4688 11224 : continue;
4689 23860 : cached = false;
4690 23860 : break;
4691 : }
4692 :
4693 : /* calculate the total number of blocks to be invalidated */
4694 4168 : nBlocksToInvalidate += block[i][j];
4695 : }
4696 : }
4697 :
4698 : /*
4699 : * We apply the optimization iff the total number of blocks to invalidate
4700 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4701 : */
4702 25344 : if (cached && nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4703 : {
4704 2454 : for (i = 0; i < n; i++)
4705 : {
4706 6750 : for (int j = 0; j <= MAX_FORKNUM; j++)
4707 : {
4708 : /* ignore relation forks that doesn't exist */
4709 5400 : if (!BlockNumberIsValid(block[i][j]))
4710 4032 : continue;
4711 :
4712 : /* drop all the buffers for a particular relation fork */
4713 1368 : FindAndDropRelationBuffers(rels[i]->smgr_rlocator.locator,
4714 1368 : j, block[i][j], 0);
4715 : }
4716 : }
4717 :
4718 1104 : pfree(block);
4719 1104 : pfree(rels);
4720 1104 : return;
4721 : }
4722 :
4723 24240 : pfree(block);
4724 24240 : locators = palloc(sizeof(RelFileLocator) * n); /* non-local relations */
4725 107544 : for (i = 0; i < n; i++)
4726 83304 : locators[i] = rels[i]->smgr_rlocator.locator;
4727 :
4728 : /*
4729 : * For low number of relations to drop just use a simple walk through, to
4730 : * save the bsearch overhead. The threshold to use is rather a guess than
4731 : * an exactly determined value, as it depends on many factors (CPU and RAM
4732 : * speeds, amount of shared buffers etc.).
4733 : */
4734 24240 : use_bsearch = n > RELS_BSEARCH_THRESHOLD;
4735 :
4736 : /* sort the list of rlocators if necessary */
4737 24240 : if (use_bsearch)
4738 334 : qsort(locators, n, sizeof(RelFileLocator), rlocator_comparator);
4739 :
4740 262085040 : for (i = 0; i < NBuffers; i++)
4741 : {
4742 262060800 : RelFileLocator *rlocator = NULL;
4743 262060800 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4744 : uint32 buf_state;
4745 :
4746 : /*
4747 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4748 : * saves some cycles.
4749 : */
4750 :
4751 262060800 : if (!use_bsearch)
4752 : {
4753 : int j;
4754 :
4755 1043973036 : for (j = 0; j < n; j++)
4756 : {
4757 785610496 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &locators[j]))
4758 : {
4759 176724 : rlocator = &locators[j];
4760 176724 : break;
4761 : }
4762 : }
4763 : }
4764 : else
4765 : {
4766 : RelFileLocator locator;
4767 :
4768 3521536 : locator = BufTagGetRelFileLocator(&bufHdr->tag);
4769 3521536 : rlocator = bsearch(&locator,
4770 : locators, n, sizeof(RelFileLocator),
4771 : rlocator_comparator);
4772 : }
4773 :
4774 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
4775 262060800 : if (rlocator == NULL)
4776 261880608 : continue;
4777 :
4778 180192 : buf_state = LockBufHdr(bufHdr);
4779 180192 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, rlocator))
4780 180192 : InvalidateBuffer(bufHdr); /* releases spinlock */
4781 : else
4782 0 : UnlockBufHdr(bufHdr, buf_state);
4783 : }
4784 :
4785 24240 : pfree(locators);
4786 24240 : pfree(rels);
4787 : }
4788 :
4789 : /* ---------------------------------------------------------------------
4790 : * FindAndDropRelationBuffers
4791 : *
4792 : * This function performs look up in BufMapping table and removes from the
4793 : * buffer pool all the pages of the specified relation fork that has block
4794 : * number >= firstDelBlock. (In particular, with firstDelBlock = 0, all
4795 : * pages are removed.)
4796 : * --------------------------------------------------------------------
4797 : */
4798 : static void
4799 1508 : FindAndDropRelationBuffers(RelFileLocator rlocator, ForkNumber forkNum,
4800 : BlockNumber nForkBlock,
4801 : BlockNumber firstDelBlock)
4802 : {
4803 : BlockNumber curBlock;
4804 :
4805 3674 : for (curBlock = firstDelBlock; curBlock < nForkBlock; curBlock++)
4806 : {
4807 : uint32 bufHash; /* hash value for tag */
4808 : BufferTag bufTag; /* identity of requested block */
4809 : LWLock *bufPartitionLock; /* buffer partition lock for it */
4810 : int buf_id;
4811 : BufferDesc *bufHdr;
4812 : uint32 buf_state;
4813 :
4814 : /* create a tag so we can lookup the buffer */
4815 2166 : InitBufferTag(&bufTag, &rlocator, forkNum, curBlock);
4816 :
4817 : /* determine its hash code and partition lock ID */
4818 2166 : bufHash = BufTableHashCode(&bufTag);
4819 2166 : bufPartitionLock = BufMappingPartitionLock(bufHash);
4820 :
4821 : /* Check that it is in the buffer pool. If not, do nothing. */
4822 2166 : LWLockAcquire(bufPartitionLock, LW_SHARED);
4823 2166 : buf_id = BufTableLookup(&bufTag, bufHash);
4824 2166 : LWLockRelease(bufPartitionLock);
4825 :
4826 2166 : if (buf_id < 0)
4827 176 : continue;
4828 :
4829 1990 : bufHdr = GetBufferDescriptor(buf_id);
4830 :
4831 : /*
4832 : * We need to lock the buffer header and recheck if the buffer is
4833 : * still associated with the same block because the buffer could be
4834 : * evicted by some other backend loading blocks for a different
4835 : * relation after we release lock on the BufMapping table.
4836 : */
4837 1990 : buf_state = LockBufHdr(bufHdr);
4838 :
4839 3980 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator) &&
4840 1990 : BufTagGetForkNum(&bufHdr->tag) == forkNum &&
4841 1990 : bufHdr->tag.blockNum >= firstDelBlock)
4842 1990 : InvalidateBuffer(bufHdr); /* releases spinlock */
4843 : else
4844 0 : UnlockBufHdr(bufHdr, buf_state);
4845 : }
4846 1508 : }
4847 :
4848 : /* ---------------------------------------------------------------------
4849 : * DropDatabaseBuffers
4850 : *
4851 : * This function removes all the buffers in the buffer cache for a
4852 : * particular database. Dirty pages are simply dropped, without
4853 : * bothering to write them out first. This is used when we destroy a
4854 : * database, to avoid trying to flush data to disk when the directory
4855 : * tree no longer exists. Implementation is pretty similar to
4856 : * DropRelationBuffers() which is for destroying just one relation.
4857 : * --------------------------------------------------------------------
4858 : */
4859 : void
4860 136 : DropDatabaseBuffers(Oid dbid)
4861 : {
4862 : int i;
4863 :
4864 : /*
4865 : * We needn't consider local buffers, since by assumption the target
4866 : * database isn't our own.
4867 : */
4868 :
4869 895368 : for (i = 0; i < NBuffers; i++)
4870 : {
4871 895232 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4872 : uint32 buf_state;
4873 :
4874 : /*
4875 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4876 : * saves some cycles.
4877 : */
4878 895232 : if (bufHdr->tag.dbOid != dbid)
4879 870148 : continue;
4880 :
4881 25084 : buf_state = LockBufHdr(bufHdr);
4882 25084 : if (bufHdr->tag.dbOid == dbid)
4883 25084 : InvalidateBuffer(bufHdr); /* releases spinlock */
4884 : else
4885 0 : UnlockBufHdr(bufHdr, buf_state);
4886 : }
4887 136 : }
4888 :
4889 : /* ---------------------------------------------------------------------
4890 : * FlushRelationBuffers
4891 : *
4892 : * This function writes all dirty pages of a relation out to disk
4893 : * (or more accurately, out to kernel disk buffers), ensuring that the
4894 : * kernel has an up-to-date view of the relation.
4895 : *
4896 : * Generally, the caller should be holding AccessExclusiveLock on the
4897 : * target relation to ensure that no other backend is busy dirtying
4898 : * more blocks of the relation; the effects can't be expected to last
4899 : * after the lock is released.
4900 : *
4901 : * XXX currently it sequentially searches the buffer pool, should be
4902 : * changed to more clever ways of searching. This routine is not
4903 : * used in any performance-critical code paths, so it's not worth
4904 : * adding additional overhead to normal paths to make it go faster.
4905 : * --------------------------------------------------------------------
4906 : */
4907 : void
4908 276 : FlushRelationBuffers(Relation rel)
4909 : {
4910 : int i;
4911 : BufferDesc *bufHdr;
4912 276 : SMgrRelation srel = RelationGetSmgr(rel);
4913 :
4914 276 : if (RelationUsesLocalBuffers(rel))
4915 : {
4916 1818 : for (i = 0; i < NLocBuffer; i++)
4917 : {
4918 : uint32 buf_state;
4919 :
4920 1800 : bufHdr = GetLocalBufferDescriptor(i);
4921 1800 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
4922 600 : ((buf_state = pg_atomic_read_u32(&bufHdr->state)) &
4923 : (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
4924 : {
4925 : ErrorContextCallback errcallback;
4926 :
4927 : /* Setup error traceback support for ereport() */
4928 600 : errcallback.callback = local_buffer_write_error_callback;
4929 600 : errcallback.arg = bufHdr;
4930 600 : errcallback.previous = error_context_stack;
4931 600 : error_context_stack = &errcallback;
4932 :
4933 : /* Make sure we can handle the pin */
4934 600 : ReservePrivateRefCountEntry();
4935 600 : ResourceOwnerEnlarge(CurrentResourceOwner);
4936 :
4937 : /*
4938 : * Pin/unpin mostly to make valgrind work, but it also seems
4939 : * like the right thing to do.
4940 : */
4941 600 : PinLocalBuffer(bufHdr, false);
4942 :
4943 :
4944 600 : FlushLocalBuffer(bufHdr, srel);
4945 :
4946 600 : UnpinLocalBuffer(BufferDescriptorGetBuffer(bufHdr));
4947 :
4948 : /* Pop the error context stack */
4949 600 : error_context_stack = errcallback.previous;
4950 : }
4951 : }
4952 :
4953 18 : return;
4954 : }
4955 :
4956 3024386 : for (i = 0; i < NBuffers; i++)
4957 : {
4958 : uint32 buf_state;
4959 :
4960 3024128 : bufHdr = GetBufferDescriptor(i);
4961 :
4962 : /*
4963 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4964 : * saves some cycles.
4965 : */
4966 3024128 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator))
4967 3023682 : continue;
4968 :
4969 : /* Make sure we can handle the pin */
4970 446 : ReservePrivateRefCountEntry();
4971 446 : ResourceOwnerEnlarge(CurrentResourceOwner);
4972 :
4973 446 : buf_state = LockBufHdr(bufHdr);
4974 446 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
4975 446 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
4976 : {
4977 360 : PinBuffer_Locked(bufHdr);
4978 360 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_SHARED);
4979 360 : FlushBuffer(bufHdr, srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
4980 360 : LWLockRelease(BufferDescriptorGetContentLock(bufHdr));
4981 360 : UnpinBuffer(bufHdr);
4982 : }
4983 : else
4984 86 : UnlockBufHdr(bufHdr, buf_state);
4985 : }
4986 : }
4987 :
4988 : /* ---------------------------------------------------------------------
4989 : * FlushRelationsAllBuffers
4990 : *
4991 : * This function flushes out of the buffer pool all the pages of all
4992 : * forks of the specified smgr relations. It's equivalent to calling
4993 : * FlushRelationBuffers once per relation. The relations are assumed not
4994 : * to use local buffers.
4995 : * --------------------------------------------------------------------
4996 : */
4997 : void
4998 24 : FlushRelationsAllBuffers(SMgrRelation *smgrs, int nrels)
4999 : {
5000 : int i;
5001 : SMgrSortArray *srels;
5002 : bool use_bsearch;
5003 :
5004 24 : if (nrels == 0)
5005 0 : return;
5006 :
5007 : /* fill-in array for qsort */
5008 24 : srels = palloc(sizeof(SMgrSortArray) * nrels);
5009 :
5010 48 : for (i = 0; i < nrels; i++)
5011 : {
5012 : Assert(!RelFileLocatorBackendIsTemp(smgrs[i]->smgr_rlocator));
5013 :
5014 24 : srels[i].rlocator = smgrs[i]->smgr_rlocator.locator;
5015 24 : srels[i].srel = smgrs[i];
5016 : }
5017 :
5018 : /*
5019 : * Save the bsearch overhead for low number of relations to sync. See
5020 : * DropRelationsAllBuffers for details.
5021 : */
5022 24 : use_bsearch = nrels > RELS_BSEARCH_THRESHOLD;
5023 :
5024 : /* sort the list of SMgrRelations if necessary */
5025 24 : if (use_bsearch)
5026 0 : qsort(srels, nrels, sizeof(SMgrSortArray), rlocator_comparator);
5027 :
5028 393240 : for (i = 0; i < NBuffers; i++)
5029 : {
5030 393216 : SMgrSortArray *srelent = NULL;
5031 393216 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5032 : uint32 buf_state;
5033 :
5034 : /*
5035 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5036 : * saves some cycles.
5037 : */
5038 :
5039 393216 : if (!use_bsearch)
5040 : {
5041 : int j;
5042 :
5043 777604 : for (j = 0; j < nrels; j++)
5044 : {
5045 393216 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srels[j].rlocator))
5046 : {
5047 8828 : srelent = &srels[j];
5048 8828 : break;
5049 : }
5050 : }
5051 : }
5052 : else
5053 : {
5054 : RelFileLocator rlocator;
5055 :
5056 0 : rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
5057 0 : srelent = bsearch(&rlocator,
5058 : srels, nrels, sizeof(SMgrSortArray),
5059 : rlocator_comparator);
5060 : }
5061 :
5062 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
5063 393216 : if (srelent == NULL)
5064 384388 : continue;
5065 :
5066 : /* Make sure we can handle the pin */
5067 8828 : ReservePrivateRefCountEntry();
5068 8828 : ResourceOwnerEnlarge(CurrentResourceOwner);
5069 :
5070 8828 : buf_state = LockBufHdr(bufHdr);
5071 8828 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srelent->rlocator) &&
5072 8828 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5073 : {
5074 7914 : PinBuffer_Locked(bufHdr);
5075 7914 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_SHARED);
5076 7914 : FlushBuffer(bufHdr, srelent->srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5077 7914 : LWLockRelease(BufferDescriptorGetContentLock(bufHdr));
5078 7914 : UnpinBuffer(bufHdr);
5079 : }
5080 : else
5081 914 : UnlockBufHdr(bufHdr, buf_state);
5082 : }
5083 :
5084 24 : pfree(srels);
5085 : }
5086 :
5087 : /* ---------------------------------------------------------------------
5088 : * RelationCopyStorageUsingBuffer
5089 : *
5090 : * Copy fork's data using bufmgr. Same as RelationCopyStorage but instead
5091 : * of using smgrread and smgrextend this will copy using bufmgr APIs.
5092 : *
5093 : * Refer comments atop CreateAndCopyRelationData() for details about
5094 : * 'permanent' parameter.
5095 : * --------------------------------------------------------------------
5096 : */
5097 : static void
5098 146664 : RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
5099 : RelFileLocator dstlocator,
5100 : ForkNumber forkNum, bool permanent)
5101 : {
5102 : Buffer srcBuf;
5103 : Buffer dstBuf;
5104 : Page srcPage;
5105 : Page dstPage;
5106 : bool use_wal;
5107 : BlockNumber nblocks;
5108 : BlockNumber blkno;
5109 : PGIOAlignedBlock buf;
5110 : BufferAccessStrategy bstrategy_src;
5111 : BufferAccessStrategy bstrategy_dst;
5112 : BlockRangeReadStreamPrivate p;
5113 : ReadStream *src_stream;
5114 : SMgrRelation src_smgr;
5115 :
5116 : /*
5117 : * In general, we want to write WAL whenever wal_level > 'minimal', but we
5118 : * can skip it when copying any fork of an unlogged relation other than
5119 : * the init fork.
5120 : */
5121 146664 : use_wal = XLogIsNeeded() && (permanent || forkNum == INIT_FORKNUM);
5122 :
5123 : /* Get number of blocks in the source relation. */
5124 146664 : nblocks = smgrnblocks(smgropen(srclocator, INVALID_PROC_NUMBER),
5125 : forkNum);
5126 :
5127 : /* Nothing to copy; just return. */
5128 146664 : if (nblocks == 0)
5129 25582 : return;
5130 :
5131 : /*
5132 : * Bulk extend the destination relation of the same size as the source
5133 : * relation before starting to copy block by block.
5134 : */
5135 121082 : memset(buf.data, 0, BLCKSZ);
5136 121082 : smgrextend(smgropen(dstlocator, INVALID_PROC_NUMBER), forkNum, nblocks - 1,
5137 : buf.data, true);
5138 :
5139 : /* This is a bulk operation, so use buffer access strategies. */
5140 121082 : bstrategy_src = GetAccessStrategy(BAS_BULKREAD);
5141 121082 : bstrategy_dst = GetAccessStrategy(BAS_BULKWRITE);
5142 :
5143 : /* Initialize streaming read */
5144 121082 : p.current_blocknum = 0;
5145 121082 : p.last_exclusive = nblocks;
5146 121082 : src_smgr = smgropen(srclocator, INVALID_PROC_NUMBER);
5147 :
5148 : /*
5149 : * It is safe to use batchmode as block_range_read_stream_cb takes no
5150 : * locks.
5151 : */
5152 121082 : src_stream = read_stream_begin_smgr_relation(READ_STREAM_FULL |
5153 : READ_STREAM_USE_BATCHING,
5154 : bstrategy_src,
5155 : src_smgr,
5156 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
5157 : forkNum,
5158 : block_range_read_stream_cb,
5159 : &p,
5160 : 0);
5161 :
5162 : /* Iterate over each block of the source relation file. */
5163 583790 : for (blkno = 0; blkno < nblocks; blkno++)
5164 : {
5165 462712 : CHECK_FOR_INTERRUPTS();
5166 :
5167 : /* Read block from source relation. */
5168 462712 : srcBuf = read_stream_next_buffer(src_stream, NULL);
5169 462708 : LockBuffer(srcBuf, BUFFER_LOCK_SHARE);
5170 462708 : srcPage = BufferGetPage(srcBuf);
5171 :
5172 462708 : dstBuf = ReadBufferWithoutRelcache(dstlocator, forkNum,
5173 : BufferGetBlockNumber(srcBuf),
5174 : RBM_ZERO_AND_LOCK, bstrategy_dst,
5175 : permanent);
5176 462708 : dstPage = BufferGetPage(dstBuf);
5177 :
5178 462708 : START_CRIT_SECTION();
5179 :
5180 : /* Copy page data from the source to the destination. */
5181 462708 : memcpy(dstPage, srcPage, BLCKSZ);
5182 462708 : MarkBufferDirty(dstBuf);
5183 :
5184 : /* WAL-log the copied page. */
5185 462708 : if (use_wal)
5186 253980 : log_newpage_buffer(dstBuf, true);
5187 :
5188 462708 : END_CRIT_SECTION();
5189 :
5190 462708 : UnlockReleaseBuffer(dstBuf);
5191 462708 : UnlockReleaseBuffer(srcBuf);
5192 : }
5193 : Assert(read_stream_next_buffer(src_stream, NULL) == InvalidBuffer);
5194 121078 : read_stream_end(src_stream);
5195 :
5196 121078 : FreeAccessStrategy(bstrategy_src);
5197 121078 : FreeAccessStrategy(bstrategy_dst);
5198 : }
5199 :
5200 : /* ---------------------------------------------------------------------
5201 : * CreateAndCopyRelationData
5202 : *
5203 : * Create destination relation storage and copy all forks from the
5204 : * source relation to the destination.
5205 : *
5206 : * Pass permanent as true for permanent relations and false for
5207 : * unlogged relations. Currently this API is not supported for
5208 : * temporary relations.
5209 : * --------------------------------------------------------------------
5210 : */
5211 : void
5212 110248 : CreateAndCopyRelationData(RelFileLocator src_rlocator,
5213 : RelFileLocator dst_rlocator, bool permanent)
5214 : {
5215 : char relpersistence;
5216 : SMgrRelation src_rel;
5217 : SMgrRelation dst_rel;
5218 :
5219 : /* Set the relpersistence. */
5220 110248 : relpersistence = permanent ?
5221 : RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED;
5222 :
5223 110248 : src_rel = smgropen(src_rlocator, INVALID_PROC_NUMBER);
5224 110248 : dst_rel = smgropen(dst_rlocator, INVALID_PROC_NUMBER);
5225 :
5226 : /*
5227 : * Create and copy all forks of the relation. During create database we
5228 : * have a separate cleanup mechanism which deletes complete database
5229 : * directory. Therefore, each individual relation doesn't need to be
5230 : * registered for cleanup.
5231 : */
5232 110248 : RelationCreateStorage(dst_rlocator, relpersistence, false);
5233 :
5234 : /* copy main fork. */
5235 110248 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, MAIN_FORKNUM,
5236 : permanent);
5237 :
5238 : /* copy those extra forks that exist */
5239 110244 : for (ForkNumber forkNum = MAIN_FORKNUM + 1;
5240 440976 : forkNum <= MAX_FORKNUM; forkNum++)
5241 : {
5242 330732 : if (smgrexists(src_rel, forkNum))
5243 : {
5244 36416 : smgrcreate(dst_rel, forkNum, false);
5245 :
5246 : /*
5247 : * WAL log creation if the relation is persistent, or this is the
5248 : * init fork of an unlogged relation.
5249 : */
5250 36416 : if (permanent || forkNum == INIT_FORKNUM)
5251 36416 : log_smgrcreate(&dst_rlocator, forkNum);
5252 :
5253 : /* Copy a fork's data, block by block. */
5254 36416 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, forkNum,
5255 : permanent);
5256 : }
5257 : }
5258 110244 : }
5259 :
5260 : /* ---------------------------------------------------------------------
5261 : * FlushDatabaseBuffers
5262 : *
5263 : * This function writes all dirty pages of a database out to disk
5264 : * (or more accurately, out to kernel disk buffers), ensuring that the
5265 : * kernel has an up-to-date view of the database.
5266 : *
5267 : * Generally, the caller should be holding an appropriate lock to ensure
5268 : * no other backend is active in the target database; otherwise more
5269 : * pages could get dirtied.
5270 : *
5271 : * Note we don't worry about flushing any pages of temporary relations.
5272 : * It's assumed these wouldn't be interesting.
5273 : * --------------------------------------------------------------------
5274 : */
5275 : void
5276 8 : FlushDatabaseBuffers(Oid dbid)
5277 : {
5278 : int i;
5279 : BufferDesc *bufHdr;
5280 :
5281 1032 : for (i = 0; i < NBuffers; i++)
5282 : {
5283 : uint32 buf_state;
5284 :
5285 1024 : bufHdr = GetBufferDescriptor(i);
5286 :
5287 : /*
5288 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5289 : * saves some cycles.
5290 : */
5291 1024 : if (bufHdr->tag.dbOid != dbid)
5292 722 : continue;
5293 :
5294 : /* Make sure we can handle the pin */
5295 302 : ReservePrivateRefCountEntry();
5296 302 : ResourceOwnerEnlarge(CurrentResourceOwner);
5297 :
5298 302 : buf_state = LockBufHdr(bufHdr);
5299 302 : if (bufHdr->tag.dbOid == dbid &&
5300 302 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5301 : {
5302 68 : PinBuffer_Locked(bufHdr);
5303 68 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_SHARED);
5304 68 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5305 68 : LWLockRelease(BufferDescriptorGetContentLock(bufHdr));
5306 68 : UnpinBuffer(bufHdr);
5307 : }
5308 : else
5309 234 : UnlockBufHdr(bufHdr, buf_state);
5310 : }
5311 8 : }
5312 :
5313 : /*
5314 : * Flush a previously, shared or exclusively, locked and pinned buffer to the
5315 : * OS.
5316 : */
5317 : void
5318 158 : FlushOneBuffer(Buffer buffer)
5319 : {
5320 : BufferDesc *bufHdr;
5321 :
5322 : /* currently not needed, but no fundamental reason not to support */
5323 : Assert(!BufferIsLocal(buffer));
5324 :
5325 : Assert(BufferIsPinned(buffer));
5326 :
5327 158 : bufHdr = GetBufferDescriptor(buffer - 1);
5328 :
5329 : Assert(LWLockHeldByMe(BufferDescriptorGetContentLock(bufHdr)));
5330 :
5331 158 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5332 158 : }
5333 :
5334 : /*
5335 : * ReleaseBuffer -- release the pin on a buffer
5336 : */
5337 : void
5338 125020932 : ReleaseBuffer(Buffer buffer)
5339 : {
5340 125020932 : if (!BufferIsValid(buffer))
5341 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5342 :
5343 125020932 : if (BufferIsLocal(buffer))
5344 3220056 : UnpinLocalBuffer(buffer);
5345 : else
5346 121800876 : UnpinBuffer(GetBufferDescriptor(buffer - 1));
5347 125020932 : }
5348 :
5349 : /*
5350 : * UnlockReleaseBuffer -- release the content lock and pin on a buffer
5351 : *
5352 : * This is just a shorthand for a common combination.
5353 : */
5354 : void
5355 37996930 : UnlockReleaseBuffer(Buffer buffer)
5356 : {
5357 37996930 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5358 37996930 : ReleaseBuffer(buffer);
5359 37996930 : }
5360 :
5361 : /*
5362 : * IncrBufferRefCount
5363 : * Increment the pin count on a buffer that we have *already* pinned
5364 : * at least once.
5365 : *
5366 : * This function cannot be used on a buffer we do not have pinned,
5367 : * because it doesn't change the shared buffer state.
5368 : */
5369 : void
5370 23273244 : IncrBufferRefCount(Buffer buffer)
5371 : {
5372 : Assert(BufferIsPinned(buffer));
5373 23273244 : ResourceOwnerEnlarge(CurrentResourceOwner);
5374 23273244 : if (BufferIsLocal(buffer))
5375 709186 : LocalRefCount[-buffer - 1]++;
5376 : else
5377 : {
5378 : PrivateRefCountEntry *ref;
5379 :
5380 22564058 : ref = GetPrivateRefCountEntry(buffer, true);
5381 : Assert(ref != NULL);
5382 22564058 : ref->refcount++;
5383 : }
5384 23273244 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buffer);
5385 23273244 : }
5386 :
5387 : /*
5388 : * MarkBufferDirtyHint
5389 : *
5390 : * Mark a buffer dirty for non-critical changes.
5391 : *
5392 : * This is essentially the same as MarkBufferDirty, except:
5393 : *
5394 : * 1. The caller does not write WAL; so if checksums are enabled, we may need
5395 : * to write an XLOG_FPI_FOR_HINT WAL record to protect against torn pages.
5396 : * 2. The caller might have only share-lock instead of exclusive-lock on the
5397 : * buffer's content lock.
5398 : * 3. This function does not guarantee that the buffer is always marked dirty
5399 : * (due to a race condition), so it cannot be used for important changes.
5400 : */
5401 : void
5402 19751290 : MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
5403 : {
5404 : BufferDesc *bufHdr;
5405 19751290 : Page page = BufferGetPage(buffer);
5406 :
5407 19751290 : if (!BufferIsValid(buffer))
5408 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5409 :
5410 19751290 : if (BufferIsLocal(buffer))
5411 : {
5412 1270218 : MarkLocalBufferDirty(buffer);
5413 1270218 : return;
5414 : }
5415 :
5416 18481072 : bufHdr = GetBufferDescriptor(buffer - 1);
5417 :
5418 : Assert(GetPrivateRefCount(buffer) > 0);
5419 : /* here, either share or exclusive lock is OK */
5420 : Assert(LWLockHeldByMe(BufferDescriptorGetContentLock(bufHdr)));
5421 :
5422 : /*
5423 : * This routine might get called many times on the same page, if we are
5424 : * making the first scan after commit of an xact that added/deleted many
5425 : * tuples. So, be as quick as we can if the buffer is already dirty. We
5426 : * do this by not acquiring spinlock if it looks like the status bits are
5427 : * already set. Since we make this test unlocked, there's a chance we
5428 : * might fail to notice that the flags have just been cleared, and failed
5429 : * to reset them, due to memory-ordering issues. But since this function
5430 : * is only intended to be used in cases where failing to write out the
5431 : * data would be harmless anyway, it doesn't really matter.
5432 : */
5433 18481072 : if ((pg_atomic_read_u32(&bufHdr->state) & (BM_DIRTY | BM_JUST_DIRTIED)) !=
5434 : (BM_DIRTY | BM_JUST_DIRTIED))
5435 : {
5436 1611388 : XLogRecPtr lsn = InvalidXLogRecPtr;
5437 1611388 : bool dirtied = false;
5438 1611388 : bool delayChkptFlags = false;
5439 : uint32 buf_state;
5440 :
5441 : /*
5442 : * If we need to protect hint bit updates from torn writes, WAL-log a
5443 : * full page image of the page. This full page image is only necessary
5444 : * if the hint bit update is the first change to the page since the
5445 : * last checkpoint.
5446 : *
5447 : * We don't check full_page_writes here because that logic is included
5448 : * when we call XLogInsert() since the value changes dynamically.
5449 : */
5450 3220584 : if (XLogHintBitIsNeeded() &&
5451 1609196 : (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT))
5452 : {
5453 : /*
5454 : * If we must not write WAL, due to a relfilelocator-specific
5455 : * condition or being in recovery, don't dirty the page. We can
5456 : * set the hint, just not dirty the page as a result so the hint
5457 : * is lost when we evict the page or shutdown.
5458 : *
5459 : * See src/backend/storage/page/README for longer discussion.
5460 : */
5461 1732586 : if (RecoveryInProgress() ||
5462 123454 : RelFileLocatorSkippingWAL(BufTagGetRelFileLocator(&bufHdr->tag)))
5463 1491364 : return;
5464 :
5465 : /*
5466 : * If the block is already dirty because we either made a change
5467 : * or set a hint already, then we don't need to write a full page
5468 : * image. Note that aggressive cleaning of blocks dirtied by hint
5469 : * bit setting would increase the call rate. Bulk setting of hint
5470 : * bits would reduce the call rate...
5471 : *
5472 : * We must issue the WAL record before we mark the buffer dirty.
5473 : * Otherwise we might write the page before we write the WAL. That
5474 : * causes a race condition, since a checkpoint might occur between
5475 : * writing the WAL record and marking the buffer dirty. We solve
5476 : * that with a kluge, but one that is already in use during
5477 : * transaction commit to prevent race conditions. Basically, we
5478 : * simply prevent the checkpoint WAL record from being written
5479 : * until we have marked the buffer dirty. We don't start the
5480 : * checkpoint flush until we have marked dirty, so our checkpoint
5481 : * must flush the change to disk successfully or the checkpoint
5482 : * never gets written, so crash recovery will fix.
5483 : *
5484 : * It's possible we may enter here without an xid, so it is
5485 : * essential that CreateCheckPoint waits for virtual transactions
5486 : * rather than full transactionids.
5487 : */
5488 : Assert((MyProc->delayChkptFlags & DELAY_CHKPT_START) == 0);
5489 117768 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
5490 117768 : delayChkptFlags = true;
5491 117768 : lsn = XLogSaveBufferForHint(buffer, buffer_std);
5492 : }
5493 :
5494 120024 : buf_state = LockBufHdr(bufHdr);
5495 :
5496 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5497 :
5498 120024 : if (!(buf_state & BM_DIRTY))
5499 : {
5500 119938 : dirtied = true; /* Means "will be dirtied by this action" */
5501 :
5502 : /*
5503 : * Set the page LSN if we wrote a backup block. We aren't supposed
5504 : * to set this when only holding a share lock but as long as we
5505 : * serialise it somehow we're OK. We choose to set LSN while
5506 : * holding the buffer header lock, which causes any reader of an
5507 : * LSN who holds only a share lock to also obtain a buffer header
5508 : * lock before using PageGetLSN(), which is enforced in
5509 : * BufferGetLSNAtomic().
5510 : *
5511 : * If checksums are enabled, you might think we should reset the
5512 : * checksum here. That will happen when the page is written
5513 : * sometime later in this checkpoint cycle.
5514 : */
5515 119938 : if (!XLogRecPtrIsInvalid(lsn))
5516 62566 : PageSetLSN(page, lsn);
5517 : }
5518 :
5519 120024 : buf_state |= BM_DIRTY | BM_JUST_DIRTIED;
5520 120024 : UnlockBufHdr(bufHdr, buf_state);
5521 :
5522 120024 : if (delayChkptFlags)
5523 117768 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
5524 :
5525 120024 : if (dirtied)
5526 : {
5527 119938 : pgBufferUsage.shared_blks_dirtied++;
5528 119938 : if (VacuumCostActive)
5529 3030 : VacuumCostBalance += VacuumCostPageDirty;
5530 : }
5531 : }
5532 : }
5533 :
5534 : /*
5535 : * Release buffer content locks for shared buffers.
5536 : *
5537 : * Used to clean up after errors.
5538 : *
5539 : * Currently, we can expect that lwlock.c's LWLockReleaseAll() took care
5540 : * of releasing buffer content locks per se; the only thing we need to deal
5541 : * with here is clearing any PIN_COUNT request that was in progress.
5542 : */
5543 : void
5544 103750 : UnlockBuffers(void)
5545 : {
5546 103750 : BufferDesc *buf = PinCountWaitBuf;
5547 :
5548 103750 : if (buf)
5549 : {
5550 : uint32 buf_state;
5551 :
5552 0 : buf_state = LockBufHdr(buf);
5553 :
5554 : /*
5555 : * Don't complain if flag bit not set; it could have been reset but we
5556 : * got a cancel/die interrupt before getting the signal.
5557 : */
5558 0 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5559 0 : buf->wait_backend_pgprocno == MyProcNumber)
5560 0 : buf_state &= ~BM_PIN_COUNT_WAITER;
5561 :
5562 0 : UnlockBufHdr(buf, buf_state);
5563 :
5564 0 : PinCountWaitBuf = NULL;
5565 : }
5566 103750 : }
5567 :
5568 : /*
5569 : * Acquire or release the content_lock for the buffer.
5570 : */
5571 : void
5572 344533154 : LockBuffer(Buffer buffer, int mode)
5573 : {
5574 : BufferDesc *buf;
5575 :
5576 : Assert(BufferIsPinned(buffer));
5577 344533154 : if (BufferIsLocal(buffer))
5578 19799688 : return; /* local buffers need no lock */
5579 :
5580 324733466 : buf = GetBufferDescriptor(buffer - 1);
5581 :
5582 324733466 : if (mode == BUFFER_LOCK_UNLOCK)
5583 164350716 : LWLockRelease(BufferDescriptorGetContentLock(buf));
5584 160382750 : else if (mode == BUFFER_LOCK_SHARE)
5585 114022384 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_SHARED);
5586 46360366 : else if (mode == BUFFER_LOCK_EXCLUSIVE)
5587 46360366 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_EXCLUSIVE);
5588 : else
5589 0 : elog(ERROR, "unrecognized buffer lock mode: %d", mode);
5590 : }
5591 :
5592 : /*
5593 : * Acquire the content_lock for the buffer, but only if we don't have to wait.
5594 : *
5595 : * This assumes the caller wants BUFFER_LOCK_EXCLUSIVE mode.
5596 : */
5597 : bool
5598 3168926 : ConditionalLockBuffer(Buffer buffer)
5599 : {
5600 : BufferDesc *buf;
5601 :
5602 : Assert(BufferIsPinned(buffer));
5603 3168926 : if (BufferIsLocal(buffer))
5604 129308 : return true; /* act as though we got it */
5605 :
5606 3039618 : buf = GetBufferDescriptor(buffer - 1);
5607 :
5608 3039618 : return LWLockConditionalAcquire(BufferDescriptorGetContentLock(buf),
5609 : LW_EXCLUSIVE);
5610 : }
5611 :
5612 : /*
5613 : * Verify that this backend is pinning the buffer exactly once.
5614 : *
5615 : * NOTE: Like in BufferIsPinned(), what we check here is that *this* backend
5616 : * holds a pin on the buffer. We do not care whether some other backend does.
5617 : */
5618 : void
5619 5074718 : CheckBufferIsPinnedOnce(Buffer buffer)
5620 : {
5621 5074718 : if (BufferIsLocal(buffer))
5622 : {
5623 1578 : if (LocalRefCount[-buffer - 1] != 1)
5624 0 : elog(ERROR, "incorrect local pin count: %d",
5625 : LocalRefCount[-buffer - 1]);
5626 : }
5627 : else
5628 : {
5629 5073140 : if (GetPrivateRefCount(buffer) != 1)
5630 0 : elog(ERROR, "incorrect local pin count: %d",
5631 : GetPrivateRefCount(buffer));
5632 : }
5633 5074718 : }
5634 :
5635 : /*
5636 : * LockBufferForCleanup - lock a buffer in preparation for deleting items
5637 : *
5638 : * Items may be deleted from a disk page only when the caller (a) holds an
5639 : * exclusive lock on the buffer and (b) has observed that no other backend
5640 : * holds a pin on the buffer. If there is a pin, then the other backend
5641 : * might have a pointer into the buffer (for example, a heapscan reference
5642 : * to an item --- see README for more details). It's OK if a pin is added
5643 : * after the cleanup starts, however; the newly-arrived backend will be
5644 : * unable to look at the page until we release the exclusive lock.
5645 : *
5646 : * To implement this protocol, a would-be deleter must pin the buffer and
5647 : * then call LockBufferForCleanup(). LockBufferForCleanup() is similar to
5648 : * LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE), except that it loops until
5649 : * it has successfully observed pin count = 1.
5650 : */
5651 : void
5652 43920 : LockBufferForCleanup(Buffer buffer)
5653 : {
5654 : BufferDesc *bufHdr;
5655 43920 : TimestampTz waitStart = 0;
5656 43920 : bool waiting = false;
5657 43920 : bool logged_recovery_conflict = false;
5658 :
5659 : Assert(BufferIsPinned(buffer));
5660 : Assert(PinCountWaitBuf == NULL);
5661 :
5662 43920 : CheckBufferIsPinnedOnce(buffer);
5663 :
5664 : /*
5665 : * We do not yet need to be worried about in-progress AIOs holding a pin,
5666 : * as we, so far, only support doing reads via AIO and this function can
5667 : * only be called once the buffer is valid (i.e. no read can be in
5668 : * flight).
5669 : */
5670 :
5671 : /* Nobody else to wait for */
5672 43920 : if (BufferIsLocal(buffer))
5673 32 : return;
5674 :
5675 43888 : bufHdr = GetBufferDescriptor(buffer - 1);
5676 :
5677 : for (;;)
5678 120 : {
5679 : uint32 buf_state;
5680 :
5681 : /* Try to acquire lock */
5682 44008 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
5683 44008 : buf_state = LockBufHdr(bufHdr);
5684 :
5685 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5686 44008 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
5687 : {
5688 : /* Successfully acquired exclusive lock with pincount 1 */
5689 43888 : UnlockBufHdr(bufHdr, buf_state);
5690 :
5691 : /*
5692 : * Emit the log message if recovery conflict on buffer pin was
5693 : * resolved but the startup process waited longer than
5694 : * deadlock_timeout for it.
5695 : */
5696 43888 : if (logged_recovery_conflict)
5697 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
5698 : waitStart, GetCurrentTimestamp(),
5699 : NULL, false);
5700 :
5701 43888 : if (waiting)
5702 : {
5703 : /* reset ps display to remove the suffix if we added one */
5704 4 : set_ps_display_remove_suffix();
5705 4 : waiting = false;
5706 : }
5707 43888 : return;
5708 : }
5709 : /* Failed, so mark myself as waiting for pincount 1 */
5710 120 : if (buf_state & BM_PIN_COUNT_WAITER)
5711 : {
5712 0 : UnlockBufHdr(bufHdr, buf_state);
5713 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5714 0 : elog(ERROR, "multiple backends attempting to wait for pincount 1");
5715 : }
5716 120 : bufHdr->wait_backend_pgprocno = MyProcNumber;
5717 120 : PinCountWaitBuf = bufHdr;
5718 120 : buf_state |= BM_PIN_COUNT_WAITER;
5719 120 : UnlockBufHdr(bufHdr, buf_state);
5720 120 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5721 :
5722 : /* Wait to be signaled by UnpinBuffer() */
5723 120 : if (InHotStandby)
5724 : {
5725 18 : if (!waiting)
5726 : {
5727 : /* adjust the process title to indicate that it's waiting */
5728 4 : set_ps_display_suffix("waiting");
5729 4 : waiting = true;
5730 : }
5731 :
5732 : /*
5733 : * Emit the log message if the startup process is waiting longer
5734 : * than deadlock_timeout for recovery conflict on buffer pin.
5735 : *
5736 : * Skip this if first time through because the startup process has
5737 : * not started waiting yet in this case. So, the wait start
5738 : * timestamp is set after this logic.
5739 : */
5740 18 : if (waitStart != 0 && !logged_recovery_conflict)
5741 : {
5742 6 : TimestampTz now = GetCurrentTimestamp();
5743 :
5744 6 : if (TimestampDifferenceExceeds(waitStart, now,
5745 : DeadlockTimeout))
5746 : {
5747 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
5748 : waitStart, now, NULL, true);
5749 4 : logged_recovery_conflict = true;
5750 : }
5751 : }
5752 :
5753 : /*
5754 : * Set the wait start timestamp if logging is enabled and first
5755 : * time through.
5756 : */
5757 18 : if (log_recovery_conflict_waits && waitStart == 0)
5758 4 : waitStart = GetCurrentTimestamp();
5759 :
5760 : /* Publish the bufid that Startup process waits on */
5761 18 : SetStartupBufferPinWaitBufId(buffer - 1);
5762 : /* Set alarm and then wait to be signaled by UnpinBuffer() */
5763 18 : ResolveRecoveryConflictWithBufferPin();
5764 : /* Reset the published bufid */
5765 18 : SetStartupBufferPinWaitBufId(-1);
5766 : }
5767 : else
5768 102 : ProcWaitForSignal(WAIT_EVENT_BUFFER_PIN);
5769 :
5770 : /*
5771 : * Remove flag marking us as waiter. Normally this will not be set
5772 : * anymore, but ProcWaitForSignal() can return for other signals as
5773 : * well. We take care to only reset the flag if we're the waiter, as
5774 : * theoretically another backend could have started waiting. That's
5775 : * impossible with the current usages due to table level locking, but
5776 : * better be safe.
5777 : */
5778 120 : buf_state = LockBufHdr(bufHdr);
5779 120 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5780 14 : bufHdr->wait_backend_pgprocno == MyProcNumber)
5781 14 : buf_state &= ~BM_PIN_COUNT_WAITER;
5782 120 : UnlockBufHdr(bufHdr, buf_state);
5783 :
5784 120 : PinCountWaitBuf = NULL;
5785 : /* Loop back and try again */
5786 : }
5787 : }
5788 :
5789 : /*
5790 : * Check called from ProcessRecoveryConflictInterrupts() when Startup process
5791 : * requests cancellation of all pin holders that are blocking it.
5792 : */
5793 : bool
5794 6 : HoldingBufferPinThatDelaysRecovery(void)
5795 : {
5796 6 : int bufid = GetStartupBufferPinWaitBufId();
5797 :
5798 : /*
5799 : * If we get woken slowly then it's possible that the Startup process was
5800 : * already woken by other backends before we got here. Also possible that
5801 : * we get here by multiple interrupts or interrupts at inappropriate
5802 : * times, so make sure we do nothing if the bufid is not set.
5803 : */
5804 6 : if (bufid < 0)
5805 2 : return false;
5806 :
5807 4 : if (GetPrivateRefCount(bufid + 1) > 0)
5808 4 : return true;
5809 :
5810 0 : return false;
5811 : }
5812 :
5813 : /*
5814 : * ConditionalLockBufferForCleanup - as above, but don't wait to get the lock
5815 : *
5816 : * We won't loop, but just check once to see if the pin count is OK. If
5817 : * not, return false with no lock held.
5818 : */
5819 : bool
5820 1320676 : ConditionalLockBufferForCleanup(Buffer buffer)
5821 : {
5822 : BufferDesc *bufHdr;
5823 : uint32 buf_state,
5824 : refcount;
5825 :
5826 : Assert(BufferIsValid(buffer));
5827 :
5828 : /* see AIO related comment in LockBufferForCleanup() */
5829 :
5830 1320676 : if (BufferIsLocal(buffer))
5831 : {
5832 1608 : refcount = LocalRefCount[-buffer - 1];
5833 : /* There should be exactly one pin */
5834 : Assert(refcount > 0);
5835 1608 : if (refcount != 1)
5836 42 : return false;
5837 : /* Nobody else to wait for */
5838 1566 : return true;
5839 : }
5840 :
5841 : /* There should be exactly one local pin */
5842 1319068 : refcount = GetPrivateRefCount(buffer);
5843 : Assert(refcount);
5844 1319068 : if (refcount != 1)
5845 424 : return false;
5846 :
5847 : /* Try to acquire lock */
5848 1318644 : if (!ConditionalLockBuffer(buffer))
5849 78 : return false;
5850 :
5851 1318566 : bufHdr = GetBufferDescriptor(buffer - 1);
5852 1318566 : buf_state = LockBufHdr(bufHdr);
5853 1318566 : refcount = BUF_STATE_GET_REFCOUNT(buf_state);
5854 :
5855 : Assert(refcount > 0);
5856 1318566 : if (refcount == 1)
5857 : {
5858 : /* Successfully acquired exclusive lock with pincount 1 */
5859 1318042 : UnlockBufHdr(bufHdr, buf_state);
5860 1318042 : return true;
5861 : }
5862 :
5863 : /* Failed, so release the lock */
5864 524 : UnlockBufHdr(bufHdr, buf_state);
5865 524 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5866 524 : return false;
5867 : }
5868 :
5869 : /*
5870 : * IsBufferCleanupOK - as above, but we already have the lock
5871 : *
5872 : * Check whether it's OK to perform cleanup on a buffer we've already
5873 : * locked. If we observe that the pin count is 1, our exclusive lock
5874 : * happens to be a cleanup lock, and we can proceed with anything that
5875 : * would have been allowable had we sought a cleanup lock originally.
5876 : */
5877 : bool
5878 3300 : IsBufferCleanupOK(Buffer buffer)
5879 : {
5880 : BufferDesc *bufHdr;
5881 : uint32 buf_state;
5882 :
5883 : Assert(BufferIsValid(buffer));
5884 :
5885 : /* see AIO related comment in LockBufferForCleanup() */
5886 :
5887 3300 : if (BufferIsLocal(buffer))
5888 : {
5889 : /* There should be exactly one pin */
5890 0 : if (LocalRefCount[-buffer - 1] != 1)
5891 0 : return false;
5892 : /* Nobody else to wait for */
5893 0 : return true;
5894 : }
5895 :
5896 : /* There should be exactly one local pin */
5897 3300 : if (GetPrivateRefCount(buffer) != 1)
5898 0 : return false;
5899 :
5900 3300 : bufHdr = GetBufferDescriptor(buffer - 1);
5901 :
5902 : /* caller must hold exclusive lock on buffer */
5903 : Assert(LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
5904 : LW_EXCLUSIVE));
5905 :
5906 3300 : buf_state = LockBufHdr(bufHdr);
5907 :
5908 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5909 3300 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
5910 : {
5911 : /* pincount is OK. */
5912 3300 : UnlockBufHdr(bufHdr, buf_state);
5913 3300 : return true;
5914 : }
5915 :
5916 0 : UnlockBufHdr(bufHdr, buf_state);
5917 0 : return false;
5918 : }
5919 :
5920 :
5921 : /*
5922 : * Functions for buffer I/O handling
5923 : *
5924 : * Also note that these are used only for shared buffers, not local ones.
5925 : */
5926 :
5927 : /*
5928 : * WaitIO -- Block until the IO_IN_PROGRESS flag on 'buf' is cleared.
5929 : */
5930 : static void
5931 14920 : WaitIO(BufferDesc *buf)
5932 : {
5933 14920 : ConditionVariable *cv = BufferDescriptorGetIOCV(buf);
5934 :
5935 14920 : ConditionVariablePrepareToSleep(cv);
5936 : for (;;)
5937 14822 : {
5938 : uint32 buf_state;
5939 : PgAioWaitRef iow;
5940 :
5941 : /*
5942 : * It may not be necessary to acquire the spinlock to check the flag
5943 : * here, but since this test is essential for correctness, we'd better
5944 : * play it safe.
5945 : */
5946 29742 : buf_state = LockBufHdr(buf);
5947 :
5948 : /*
5949 : * Copy the wait reference while holding the spinlock. This protects
5950 : * against a concurrent TerminateBufferIO() in another backend from
5951 : * clearing the wref while it's being read.
5952 : */
5953 29742 : iow = buf->io_wref;
5954 29742 : UnlockBufHdr(buf, buf_state);
5955 :
5956 : /* no IO in progress, we don't need to wait */
5957 29742 : if (!(buf_state & BM_IO_IN_PROGRESS))
5958 14920 : break;
5959 :
5960 : /*
5961 : * The buffer has asynchronous IO in progress, wait for it to
5962 : * complete.
5963 : */
5964 14822 : if (pgaio_wref_valid(&iow))
5965 : {
5966 12332 : pgaio_wref_wait(&iow);
5967 :
5968 : /*
5969 : * The AIO subsystem internally uses condition variables and thus
5970 : * might remove this backend from the BufferDesc's CV. While that
5971 : * wouldn't cause a correctness issue (the first CV sleep just
5972 : * immediately returns if not already registered), it seems worth
5973 : * avoiding unnecessary loop iterations, given that we take care
5974 : * to do so at the start of the function.
5975 : */
5976 12332 : ConditionVariablePrepareToSleep(cv);
5977 12332 : continue;
5978 : }
5979 :
5980 : /* wait on BufferDesc->cv, e.g. for concurrent synchronous IO */
5981 2490 : ConditionVariableSleep(cv, WAIT_EVENT_BUFFER_IO);
5982 : }
5983 14920 : ConditionVariableCancelSleep();
5984 14920 : }
5985 :
5986 : /*
5987 : * StartBufferIO: begin I/O on this buffer
5988 : * (Assumptions)
5989 : * My process is executing no IO on this buffer
5990 : * The buffer is Pinned
5991 : *
5992 : * In some scenarios multiple backends could attempt the same I/O operation
5993 : * concurrently. If someone else has already started I/O on this buffer then
5994 : * we will wait for completion of the IO using WaitIO().
5995 : *
5996 : * Input operations are only attempted on buffers that are not BM_VALID,
5997 : * and output operations only on buffers that are BM_VALID and BM_DIRTY,
5998 : * so we can always tell if the work is already done.
5999 : *
6000 : * Returns true if we successfully marked the buffer as I/O busy,
6001 : * false if someone else already did the work.
6002 : *
6003 : * If nowait is true, then we don't wait for an I/O to be finished by another
6004 : * backend. In that case, false indicates either that the I/O was already
6005 : * finished, or is still in progress. This is useful for callers that want to
6006 : * find out if they can perform the I/O as part of a larger operation, without
6007 : * waiting for the answer or distinguishing the reasons why not.
6008 : */
6009 : bool
6010 4911060 : StartBufferIO(BufferDesc *buf, bool forInput, bool nowait)
6011 : {
6012 : uint32 buf_state;
6013 :
6014 4911060 : ResourceOwnerEnlarge(CurrentResourceOwner);
6015 :
6016 : for (;;)
6017 : {
6018 4925978 : buf_state = LockBufHdr(buf);
6019 :
6020 4925978 : if (!(buf_state & BM_IO_IN_PROGRESS))
6021 4911052 : break;
6022 14926 : UnlockBufHdr(buf, buf_state);
6023 14926 : if (nowait)
6024 8 : return false;
6025 14918 : WaitIO(buf);
6026 : }
6027 :
6028 : /* Once we get here, there is definitely no I/O active on this buffer */
6029 :
6030 : /* Check if someone else already did the I/O */
6031 4911052 : if (forInput ? (buf_state & BM_VALID) : !(buf_state & BM_DIRTY))
6032 : {
6033 16398 : UnlockBufHdr(buf, buf_state);
6034 16398 : return false;
6035 : }
6036 :
6037 4894654 : buf_state |= BM_IO_IN_PROGRESS;
6038 4894654 : UnlockBufHdr(buf, buf_state);
6039 :
6040 4894654 : ResourceOwnerRememberBufferIO(CurrentResourceOwner,
6041 : BufferDescriptorGetBuffer(buf));
6042 :
6043 4894654 : return true;
6044 : }
6045 :
6046 : /*
6047 : * TerminateBufferIO: release a buffer we were doing I/O on
6048 : * (Assumptions)
6049 : * My process is executing IO for the buffer
6050 : * BM_IO_IN_PROGRESS bit is set for the buffer
6051 : * The buffer is Pinned
6052 : *
6053 : * If clear_dirty is true and BM_JUST_DIRTIED is not set, we clear the
6054 : * buffer's BM_DIRTY flag. This is appropriate when terminating a
6055 : * successful write. The check on BM_JUST_DIRTIED is necessary to avoid
6056 : * marking the buffer clean if it was re-dirtied while we were writing.
6057 : *
6058 : * set_flag_bits gets ORed into the buffer's flags. It must include
6059 : * BM_IO_ERROR in a failure case. For successful completion it could
6060 : * be 0, or BM_VALID if we just finished reading in the page.
6061 : *
6062 : * If forget_owner is true, we release the buffer I/O from the current
6063 : * resource owner. (forget_owner=false is used when the resource owner itself
6064 : * is being released)
6065 : */
6066 : void
6067 4619434 : TerminateBufferIO(BufferDesc *buf, bool clear_dirty, uint32 set_flag_bits,
6068 : bool forget_owner, bool release_aio)
6069 : {
6070 : uint32 buf_state;
6071 :
6072 4619434 : buf_state = LockBufHdr(buf);
6073 :
6074 : Assert(buf_state & BM_IO_IN_PROGRESS);
6075 4619434 : buf_state &= ~BM_IO_IN_PROGRESS;
6076 :
6077 : /* Clear earlier errors, if this IO failed, it'll be marked again */
6078 4619434 : buf_state &= ~BM_IO_ERROR;
6079 :
6080 4619434 : if (clear_dirty && !(buf_state & BM_JUST_DIRTIED))
6081 1116044 : buf_state &= ~(BM_DIRTY | BM_CHECKPOINT_NEEDED);
6082 :
6083 4619434 : if (release_aio)
6084 : {
6085 : /* release ownership by the AIO subsystem */
6086 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6087 2502086 : buf_state -= BUF_REFCOUNT_ONE;
6088 2502086 : pgaio_wref_clear(&buf->io_wref);
6089 : }
6090 :
6091 4619434 : buf_state |= set_flag_bits;
6092 4619434 : UnlockBufHdr(buf, buf_state);
6093 :
6094 4619434 : if (forget_owner)
6095 2117306 : ResourceOwnerForgetBufferIO(CurrentResourceOwner,
6096 : BufferDescriptorGetBuffer(buf));
6097 :
6098 4619434 : ConditionVariableBroadcast(BufferDescriptorGetIOCV(buf));
6099 :
6100 : /*
6101 : * Support LockBufferForCleanup()
6102 : *
6103 : * We may have just released the last pin other than the waiter's. In most
6104 : * cases, this backend holds another pin on the buffer. But, if, for
6105 : * example, this backend is completing an IO issued by another backend, it
6106 : * may be time to wake the waiter.
6107 : */
6108 4619434 : if (release_aio && (buf_state & BM_PIN_COUNT_WAITER))
6109 0 : WakePinCountWaiter(buf);
6110 4619434 : }
6111 :
6112 : /*
6113 : * AbortBufferIO: Clean up active buffer I/O after an error.
6114 : *
6115 : * All LWLocks we might have held have been released,
6116 : * but we haven't yet released buffer pins, so the buffer is still pinned.
6117 : *
6118 : * If I/O was in progress, we always set BM_IO_ERROR, even though it's
6119 : * possible the error condition wasn't related to the I/O.
6120 : *
6121 : * Note: this does not remove the buffer I/O from the resource owner.
6122 : * That's correct when we're releasing the whole resource owner, but
6123 : * beware if you use this in other contexts.
6124 : */
6125 : static void
6126 30 : AbortBufferIO(Buffer buffer)
6127 : {
6128 30 : BufferDesc *buf_hdr = GetBufferDescriptor(buffer - 1);
6129 : uint32 buf_state;
6130 :
6131 30 : buf_state = LockBufHdr(buf_hdr);
6132 : Assert(buf_state & (BM_IO_IN_PROGRESS | BM_TAG_VALID));
6133 :
6134 30 : if (!(buf_state & BM_VALID))
6135 : {
6136 : Assert(!(buf_state & BM_DIRTY));
6137 30 : UnlockBufHdr(buf_hdr, buf_state);
6138 : }
6139 : else
6140 : {
6141 : Assert(buf_state & BM_DIRTY);
6142 0 : UnlockBufHdr(buf_hdr, buf_state);
6143 :
6144 : /* Issue notice if this is not the first failure... */
6145 0 : if (buf_state & BM_IO_ERROR)
6146 : {
6147 : /* Buffer is pinned, so we can read tag without spinlock */
6148 0 : ereport(WARNING,
6149 : (errcode(ERRCODE_IO_ERROR),
6150 : errmsg("could not write block %u of %s",
6151 : buf_hdr->tag.blockNum,
6152 : relpathperm(BufTagGetRelFileLocator(&buf_hdr->tag),
6153 : BufTagGetForkNum(&buf_hdr->tag)).str),
6154 : errdetail("Multiple failures --- write error might be permanent.")));
6155 : }
6156 : }
6157 :
6158 30 : TerminateBufferIO(buf_hdr, false, BM_IO_ERROR, false, false);
6159 30 : }
6160 :
6161 : /*
6162 : * Error context callback for errors occurring during shared buffer writes.
6163 : */
6164 : static void
6165 94 : shared_buffer_write_error_callback(void *arg)
6166 : {
6167 94 : BufferDesc *bufHdr = (BufferDesc *) arg;
6168 :
6169 : /* Buffer is pinned, so we can read the tag without locking the spinlock */
6170 94 : if (bufHdr != NULL)
6171 188 : errcontext("writing block %u of relation \"%s\"",
6172 : bufHdr->tag.blockNum,
6173 94 : relpathperm(BufTagGetRelFileLocator(&bufHdr->tag),
6174 : BufTagGetForkNum(&bufHdr->tag)).str);
6175 94 : }
6176 :
6177 : /*
6178 : * Error context callback for errors occurring during local buffer writes.
6179 : */
6180 : static void
6181 0 : local_buffer_write_error_callback(void *arg)
6182 : {
6183 0 : BufferDesc *bufHdr = (BufferDesc *) arg;
6184 :
6185 0 : if (bufHdr != NULL)
6186 0 : errcontext("writing block %u of relation \"%s\"",
6187 : bufHdr->tag.blockNum,
6188 0 : relpathbackend(BufTagGetRelFileLocator(&bufHdr->tag),
6189 : MyProcNumber,
6190 : BufTagGetForkNum(&bufHdr->tag)).str);
6191 0 : }
6192 :
6193 : /*
6194 : * RelFileLocator qsort/bsearch comparator; see RelFileLocatorEquals.
6195 : */
6196 : static int
6197 18430980 : rlocator_comparator(const void *p1, const void *p2)
6198 : {
6199 18430980 : RelFileLocator n1 = *(const RelFileLocator *) p1;
6200 18430980 : RelFileLocator n2 = *(const RelFileLocator *) p2;
6201 :
6202 18430980 : if (n1.relNumber < n2.relNumber)
6203 18358248 : return -1;
6204 72732 : else if (n1.relNumber > n2.relNumber)
6205 69264 : return 1;
6206 :
6207 3468 : if (n1.dbOid < n2.dbOid)
6208 0 : return -1;
6209 3468 : else if (n1.dbOid > n2.dbOid)
6210 0 : return 1;
6211 :
6212 3468 : if (n1.spcOid < n2.spcOid)
6213 0 : return -1;
6214 3468 : else if (n1.spcOid > n2.spcOid)
6215 0 : return 1;
6216 : else
6217 3468 : return 0;
6218 : }
6219 :
6220 : /*
6221 : * Lock buffer header - set BM_LOCKED in buffer state.
6222 : */
6223 : uint32
6224 73418542 : LockBufHdr(BufferDesc *desc)
6225 : {
6226 : SpinDelayStatus delayStatus;
6227 : uint32 old_buf_state;
6228 :
6229 : Assert(!BufferIsLocal(BufferDescriptorGetBuffer(desc)));
6230 :
6231 73418542 : init_local_spin_delay(&delayStatus);
6232 :
6233 : while (true)
6234 : {
6235 : /* set BM_LOCKED flag */
6236 73439558 : old_buf_state = pg_atomic_fetch_or_u32(&desc->state, BM_LOCKED);
6237 : /* if it wasn't set before we're OK */
6238 73439558 : if (!(old_buf_state & BM_LOCKED))
6239 73418542 : break;
6240 21016 : perform_spin_delay(&delayStatus);
6241 : }
6242 73418542 : finish_spin_delay(&delayStatus);
6243 73418542 : return old_buf_state | BM_LOCKED;
6244 : }
6245 :
6246 : /*
6247 : * Wait until the BM_LOCKED flag isn't set anymore and return the buffer's
6248 : * state at that point.
6249 : *
6250 : * Obviously the buffer could be locked by the time the value is returned, so
6251 : * this is primarily useful in CAS style loops.
6252 : */
6253 : static uint32
6254 3388 : WaitBufHdrUnlocked(BufferDesc *buf)
6255 : {
6256 : SpinDelayStatus delayStatus;
6257 : uint32 buf_state;
6258 :
6259 3388 : init_local_spin_delay(&delayStatus);
6260 :
6261 3388 : buf_state = pg_atomic_read_u32(&buf->state);
6262 :
6263 23196 : while (buf_state & BM_LOCKED)
6264 : {
6265 19808 : perform_spin_delay(&delayStatus);
6266 19808 : buf_state = pg_atomic_read_u32(&buf->state);
6267 : }
6268 :
6269 3388 : finish_spin_delay(&delayStatus);
6270 :
6271 3388 : return buf_state;
6272 : }
6273 :
6274 : /*
6275 : * BufferTag comparator.
6276 : */
6277 : static inline int
6278 0 : buffertag_comparator(const BufferTag *ba, const BufferTag *bb)
6279 : {
6280 : int ret;
6281 : RelFileLocator rlocatora;
6282 : RelFileLocator rlocatorb;
6283 :
6284 0 : rlocatora = BufTagGetRelFileLocator(ba);
6285 0 : rlocatorb = BufTagGetRelFileLocator(bb);
6286 :
6287 0 : ret = rlocator_comparator(&rlocatora, &rlocatorb);
6288 :
6289 0 : if (ret != 0)
6290 0 : return ret;
6291 :
6292 0 : if (BufTagGetForkNum(ba) < BufTagGetForkNum(bb))
6293 0 : return -1;
6294 0 : if (BufTagGetForkNum(ba) > BufTagGetForkNum(bb))
6295 0 : return 1;
6296 :
6297 0 : if (ba->blockNum < bb->blockNum)
6298 0 : return -1;
6299 0 : if (ba->blockNum > bb->blockNum)
6300 0 : return 1;
6301 :
6302 0 : return 0;
6303 : }
6304 :
6305 : /*
6306 : * Comparator determining the writeout order in a checkpoint.
6307 : *
6308 : * It is important that tablespaces are compared first, the logic balancing
6309 : * writes between tablespaces relies on it.
6310 : */
6311 : static inline int
6312 5918144 : ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b)
6313 : {
6314 : /* compare tablespace */
6315 5918144 : if (a->tsId < b->tsId)
6316 16704 : return -1;
6317 5901440 : else if (a->tsId > b->tsId)
6318 55210 : return 1;
6319 : /* compare relation */
6320 5846230 : if (a->relNumber < b->relNumber)
6321 1654884 : return -1;
6322 4191346 : else if (a->relNumber > b->relNumber)
6323 1573842 : return 1;
6324 : /* compare fork */
6325 2617504 : else if (a->forkNum < b->forkNum)
6326 109464 : return -1;
6327 2508040 : else if (a->forkNum > b->forkNum)
6328 115530 : return 1;
6329 : /* compare block number */
6330 2392510 : else if (a->blockNum < b->blockNum)
6331 1174116 : return -1;
6332 1218394 : else if (a->blockNum > b->blockNum)
6333 1145530 : return 1;
6334 : /* equal page IDs are unlikely, but not impossible */
6335 72864 : return 0;
6336 : }
6337 :
6338 : /*
6339 : * Comparator for a Min-Heap over the per-tablespace checkpoint completion
6340 : * progress.
6341 : */
6342 : static int
6343 481652 : ts_ckpt_progress_comparator(Datum a, Datum b, void *arg)
6344 : {
6345 481652 : CkptTsStatus *sa = (CkptTsStatus *) DatumGetPointer(a);
6346 481652 : CkptTsStatus *sb = (CkptTsStatus *) DatumGetPointer(b);
6347 :
6348 : /* we want a min-heap, so return 1 for the a < b */
6349 481652 : if (sa->progress < sb->progress)
6350 431730 : return 1;
6351 49922 : else if (sa->progress == sb->progress)
6352 1612 : return 0;
6353 : else
6354 48310 : return -1;
6355 : }
6356 :
6357 : /*
6358 : * Initialize a writeback context, discarding potential previous state.
6359 : *
6360 : * *max_pending is a pointer instead of an immediate value, so the coalesce
6361 : * limits can easily changed by the GUC mechanism, and so calling code does
6362 : * not have to check the current configuration. A value of 0 means that no
6363 : * writeback control will be performed.
6364 : */
6365 : void
6366 5324 : WritebackContextInit(WritebackContext *context, int *max_pending)
6367 : {
6368 : Assert(*max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
6369 :
6370 5324 : context->max_pending = max_pending;
6371 5324 : context->nr_pending = 0;
6372 5324 : }
6373 :
6374 : /*
6375 : * Add buffer to list of pending writeback requests.
6376 : */
6377 : void
6378 1105698 : ScheduleBufferTagForWriteback(WritebackContext *wb_context, IOContext io_context,
6379 : BufferTag *tag)
6380 : {
6381 : PendingWriteback *pending;
6382 :
6383 : /*
6384 : * As pg_flush_data() doesn't do anything with fsync disabled, there's no
6385 : * point in tracking in that case.
6386 : */
6387 1105698 : if (io_direct_flags & IO_DIRECT_DATA ||
6388 1104654 : !enableFsync)
6389 1105698 : return;
6390 :
6391 : /*
6392 : * Add buffer to the pending writeback array, unless writeback control is
6393 : * disabled.
6394 : */
6395 0 : if (*wb_context->max_pending > 0)
6396 : {
6397 : Assert(*wb_context->max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
6398 :
6399 0 : pending = &wb_context->pending_writebacks[wb_context->nr_pending++];
6400 :
6401 0 : pending->tag = *tag;
6402 : }
6403 :
6404 : /*
6405 : * Perform pending flushes if the writeback limit is exceeded. This
6406 : * includes the case where previously an item has been added, but control
6407 : * is now disabled.
6408 : */
6409 0 : if (wb_context->nr_pending >= *wb_context->max_pending)
6410 0 : IssuePendingWritebacks(wb_context, io_context);
6411 : }
6412 :
6413 : #define ST_SORT sort_pending_writebacks
6414 : #define ST_ELEMENT_TYPE PendingWriteback
6415 : #define ST_COMPARE(a, b) buffertag_comparator(&a->tag, &b->tag)
6416 : #define ST_SCOPE static
6417 : #define ST_DEFINE
6418 : #include "lib/sort_template.h"
6419 :
6420 : /*
6421 : * Issue all pending writeback requests, previously scheduled with
6422 : * ScheduleBufferTagForWriteback, to the OS.
6423 : *
6424 : * Because this is only used to improve the OSs IO scheduling we try to never
6425 : * error out - it's just a hint.
6426 : */
6427 : void
6428 2086 : IssuePendingWritebacks(WritebackContext *wb_context, IOContext io_context)
6429 : {
6430 : instr_time io_start;
6431 : int i;
6432 :
6433 2086 : if (wb_context->nr_pending == 0)
6434 2086 : return;
6435 :
6436 : /*
6437 : * Executing the writes in-order can make them a lot faster, and allows to
6438 : * merge writeback requests to consecutive blocks into larger writebacks.
6439 : */
6440 0 : sort_pending_writebacks(wb_context->pending_writebacks,
6441 0 : wb_context->nr_pending);
6442 :
6443 0 : io_start = pgstat_prepare_io_time(track_io_timing);
6444 :
6445 : /*
6446 : * Coalesce neighbouring writes, but nothing else. For that we iterate
6447 : * through the, now sorted, array of pending flushes, and look forward to
6448 : * find all neighbouring (or identical) writes.
6449 : */
6450 0 : for (i = 0; i < wb_context->nr_pending; i++)
6451 : {
6452 : PendingWriteback *cur;
6453 : PendingWriteback *next;
6454 : SMgrRelation reln;
6455 : int ahead;
6456 : BufferTag tag;
6457 : RelFileLocator currlocator;
6458 0 : Size nblocks = 1;
6459 :
6460 0 : cur = &wb_context->pending_writebacks[i];
6461 0 : tag = cur->tag;
6462 0 : currlocator = BufTagGetRelFileLocator(&tag);
6463 :
6464 : /*
6465 : * Peek ahead, into following writeback requests, to see if they can
6466 : * be combined with the current one.
6467 : */
6468 0 : for (ahead = 0; i + ahead + 1 < wb_context->nr_pending; ahead++)
6469 : {
6470 :
6471 0 : next = &wb_context->pending_writebacks[i + ahead + 1];
6472 :
6473 : /* different file, stop */
6474 0 : if (!RelFileLocatorEquals(currlocator,
6475 0 : BufTagGetRelFileLocator(&next->tag)) ||
6476 0 : BufTagGetForkNum(&cur->tag) != BufTagGetForkNum(&next->tag))
6477 : break;
6478 :
6479 : /* ok, block queued twice, skip */
6480 0 : if (cur->tag.blockNum == next->tag.blockNum)
6481 0 : continue;
6482 :
6483 : /* only merge consecutive writes */
6484 0 : if (cur->tag.blockNum + 1 != next->tag.blockNum)
6485 0 : break;
6486 :
6487 0 : nblocks++;
6488 0 : cur = next;
6489 : }
6490 :
6491 0 : i += ahead;
6492 :
6493 : /* and finally tell the kernel to write the data to storage */
6494 0 : reln = smgropen(currlocator, INVALID_PROC_NUMBER);
6495 0 : smgrwriteback(reln, BufTagGetForkNum(&tag), tag.blockNum, nblocks);
6496 : }
6497 :
6498 : /*
6499 : * Assume that writeback requests are only issued for buffers containing
6500 : * blocks of permanent relations.
6501 : */
6502 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
6503 0 : IOOP_WRITEBACK, io_start, wb_context->nr_pending, 0);
6504 :
6505 0 : wb_context->nr_pending = 0;
6506 : }
6507 :
6508 : /* ResourceOwner callbacks */
6509 :
6510 : static void
6511 30 : ResOwnerReleaseBufferIO(Datum res)
6512 : {
6513 30 : Buffer buffer = DatumGetInt32(res);
6514 :
6515 30 : AbortBufferIO(buffer);
6516 30 : }
6517 :
6518 : static char *
6519 0 : ResOwnerPrintBufferIO(Datum res)
6520 : {
6521 0 : Buffer buffer = DatumGetInt32(res);
6522 :
6523 0 : return psprintf("lost track of buffer IO on buffer %d", buffer);
6524 : }
6525 :
6526 : static void
6527 14812 : ResOwnerReleaseBufferPin(Datum res)
6528 : {
6529 14812 : Buffer buffer = DatumGetInt32(res);
6530 :
6531 : /* Like ReleaseBuffer, but don't call ResourceOwnerForgetBuffer */
6532 14812 : if (!BufferIsValid(buffer))
6533 0 : elog(ERROR, "bad buffer ID: %d", buffer);
6534 :
6535 14812 : if (BufferIsLocal(buffer))
6536 5970 : UnpinLocalBufferNoOwner(buffer);
6537 : else
6538 8842 : UnpinBufferNoOwner(GetBufferDescriptor(buffer - 1));
6539 14812 : }
6540 :
6541 : static char *
6542 0 : ResOwnerPrintBufferPin(Datum res)
6543 : {
6544 0 : return DebugPrintBufferRefcount(DatumGetInt32(res));
6545 : }
6546 :
6547 : /*
6548 : * Helper function to evict unpinned buffer whose buffer header lock is
6549 : * already acquired.
6550 : */
6551 : static bool
6552 4276 : EvictUnpinnedBufferInternal(BufferDesc *desc, bool *buffer_flushed)
6553 : {
6554 : uint32 buf_state;
6555 : bool result;
6556 :
6557 4276 : *buffer_flushed = false;
6558 :
6559 4276 : buf_state = pg_atomic_read_u32(&(desc->state));
6560 : Assert(buf_state & BM_LOCKED);
6561 :
6562 4276 : if ((buf_state & BM_VALID) == 0)
6563 : {
6564 0 : UnlockBufHdr(desc, buf_state);
6565 0 : return false;
6566 : }
6567 :
6568 : /* Check that it's not pinned already. */
6569 4276 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
6570 : {
6571 0 : UnlockBufHdr(desc, buf_state);
6572 0 : return false;
6573 : }
6574 :
6575 4276 : PinBuffer_Locked(desc); /* releases spinlock */
6576 :
6577 : /* If it was dirty, try to clean it once. */
6578 4276 : if (buf_state & BM_DIRTY)
6579 : {
6580 1938 : LWLockAcquire(BufferDescriptorGetContentLock(desc), LW_SHARED);
6581 1938 : FlushBuffer(desc, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
6582 1938 : *buffer_flushed = true;
6583 1938 : LWLockRelease(BufferDescriptorGetContentLock(desc));
6584 : }
6585 :
6586 : /* This will return false if it becomes dirty or someone else pins it. */
6587 4276 : result = InvalidateVictimBuffer(desc);
6588 :
6589 4276 : UnpinBuffer(desc);
6590 :
6591 4276 : return result;
6592 : }
6593 :
6594 : /*
6595 : * Try to evict the current block in a shared buffer.
6596 : *
6597 : * This function is intended for testing/development use only!
6598 : *
6599 : * To succeed, the buffer must not be pinned on entry, so if the caller had a
6600 : * particular block in mind, it might already have been replaced by some other
6601 : * block by the time this function runs. It's also unpinned on return, so the
6602 : * buffer might be occupied again by the time control is returned, potentially
6603 : * even by the same block. This inherent raciness without other interlocking
6604 : * makes the function unsuitable for non-testing usage.
6605 : *
6606 : * *buffer_flushed is set to true if the buffer was dirty and has been
6607 : * flushed, false otherwise. However, *buffer_flushed=true does not
6608 : * necessarily mean that we flushed the buffer, it could have been flushed by
6609 : * someone else.
6610 : *
6611 : * Returns true if the buffer was valid and it has now been made invalid.
6612 : * Returns false if it wasn't valid, if it couldn't be evicted due to a pin,
6613 : * or if the buffer becomes dirty again while we're trying to write it out.
6614 : */
6615 : bool
6616 280 : EvictUnpinnedBuffer(Buffer buf, bool *buffer_flushed)
6617 : {
6618 : BufferDesc *desc;
6619 :
6620 : Assert(BufferIsValid(buf) && !BufferIsLocal(buf));
6621 :
6622 : /* Make sure we can pin the buffer. */
6623 280 : ResourceOwnerEnlarge(CurrentResourceOwner);
6624 280 : ReservePrivateRefCountEntry();
6625 :
6626 280 : desc = GetBufferDescriptor(buf - 1);
6627 280 : LockBufHdr(desc);
6628 :
6629 280 : return EvictUnpinnedBufferInternal(desc, buffer_flushed);
6630 : }
6631 :
6632 : /*
6633 : * Try to evict all the shared buffers.
6634 : *
6635 : * This function is intended for testing/development use only! See
6636 : * EvictUnpinnedBuffer().
6637 : *
6638 : * The buffers_* parameters are mandatory and indicate the total count of
6639 : * buffers that:
6640 : * - buffers_evicted - were evicted
6641 : * - buffers_flushed - were flushed
6642 : * - buffers_skipped - could not be evicted
6643 : */
6644 : void
6645 2 : EvictAllUnpinnedBuffers(int32 *buffers_evicted, int32 *buffers_flushed,
6646 : int32 *buffers_skipped)
6647 : {
6648 2 : *buffers_evicted = 0;
6649 2 : *buffers_skipped = 0;
6650 2 : *buffers_flushed = 0;
6651 :
6652 32770 : for (int buf = 1; buf <= NBuffers; buf++)
6653 : {
6654 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
6655 : uint32 buf_state;
6656 : bool buffer_flushed;
6657 :
6658 32768 : buf_state = pg_atomic_read_u32(&desc->state);
6659 32768 : if (!(buf_state & BM_VALID))
6660 28772 : continue;
6661 :
6662 3996 : ResourceOwnerEnlarge(CurrentResourceOwner);
6663 3996 : ReservePrivateRefCountEntry();
6664 :
6665 3996 : LockBufHdr(desc);
6666 :
6667 3996 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
6668 3996 : (*buffers_evicted)++;
6669 : else
6670 0 : (*buffers_skipped)++;
6671 :
6672 3996 : if (buffer_flushed)
6673 1900 : (*buffers_flushed)++;
6674 : }
6675 2 : }
6676 :
6677 : /*
6678 : * Try to evict all the shared buffers containing provided relation's pages.
6679 : *
6680 : * This function is intended for testing/development use only! See
6681 : * EvictUnpinnedBuffer().
6682 : *
6683 : * The caller must hold at least AccessShareLock on the relation to prevent
6684 : * the relation from being dropped.
6685 : *
6686 : * The buffers_* parameters are mandatory and indicate the total count of
6687 : * buffers that:
6688 : * - buffers_evicted - were evicted
6689 : * - buffers_flushed - were flushed
6690 : * - buffers_skipped - could not be evicted
6691 : */
6692 : void
6693 2 : EvictRelUnpinnedBuffers(Relation rel, int32 *buffers_evicted,
6694 : int32 *buffers_flushed, int32 *buffers_skipped)
6695 : {
6696 : Assert(!RelationUsesLocalBuffers(rel));
6697 :
6698 2 : *buffers_skipped = 0;
6699 2 : *buffers_evicted = 0;
6700 2 : *buffers_flushed = 0;
6701 :
6702 32770 : for (int buf = 1; buf <= NBuffers; buf++)
6703 : {
6704 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
6705 32768 : uint32 buf_state = pg_atomic_read_u32(&(desc->state));
6706 : bool buffer_flushed;
6707 :
6708 : /* An unlocked precheck should be safe and saves some cycles. */
6709 32768 : if ((buf_state & BM_VALID) == 0 ||
6710 54 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
6711 32768 : continue;
6712 :
6713 : /* Make sure we can pin the buffer. */
6714 0 : ResourceOwnerEnlarge(CurrentResourceOwner);
6715 0 : ReservePrivateRefCountEntry();
6716 :
6717 0 : buf_state = LockBufHdr(desc);
6718 :
6719 : /* recheck, could have changed without the lock */
6720 0 : if ((buf_state & BM_VALID) == 0 ||
6721 0 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
6722 : {
6723 0 : UnlockBufHdr(desc, buf_state);
6724 0 : continue;
6725 : }
6726 :
6727 0 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
6728 0 : (*buffers_evicted)++;
6729 : else
6730 0 : (*buffers_skipped)++;
6731 :
6732 0 : if (buffer_flushed)
6733 0 : (*buffers_flushed)++;
6734 : }
6735 2 : }
6736 :
6737 : /*
6738 : * Generic implementation of the AIO handle staging callback for readv/writev
6739 : * on local/shared buffers.
6740 : *
6741 : * Each readv/writev can target multiple buffers. The buffers have already
6742 : * been registered with the IO handle.
6743 : *
6744 : * To make the IO ready for execution ("staging"), we need to ensure that the
6745 : * targeted buffers are in an appropriate state while the IO is ongoing. For
6746 : * that the AIO subsystem needs to have its own buffer pin, otherwise an error
6747 : * in this backend could lead to this backend's buffer pin being released as
6748 : * part of error handling, which in turn could lead to the buffer being
6749 : * replaced while IO is ongoing.
6750 : */
6751 : static pg_attribute_always_inline void
6752 2464496 : buffer_stage_common(PgAioHandle *ioh, bool is_write, bool is_temp)
6753 : {
6754 : uint64 *io_data;
6755 : uint8 handle_data_len;
6756 : PgAioWaitRef io_ref;
6757 2464496 : BufferTag first PG_USED_FOR_ASSERTS_ONLY = {0};
6758 :
6759 2464496 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
6760 :
6761 2464496 : pgaio_io_get_wref(ioh, &io_ref);
6762 :
6763 : /* iterate over all buffers affected by the vectored readv/writev */
6764 5258614 : for (int i = 0; i < handle_data_len; i++)
6765 : {
6766 2794118 : Buffer buffer = (Buffer) io_data[i];
6767 2794118 : BufferDesc *buf_hdr = is_temp ?
6768 16812 : GetLocalBufferDescriptor(-buffer - 1)
6769 2794118 : : GetBufferDescriptor(buffer - 1);
6770 : uint32 buf_state;
6771 :
6772 : /*
6773 : * Check that all the buffers are actually ones that could conceivably
6774 : * be done in one IO, i.e. are sequential. This is the last
6775 : * buffer-aware code before IO is actually executed and confusion
6776 : * about which buffers are targeted by IO can be hard to debug, making
6777 : * it worth doing extra-paranoid checks.
6778 : */
6779 2794118 : if (i == 0)
6780 2464496 : first = buf_hdr->tag;
6781 : else
6782 : {
6783 : Assert(buf_hdr->tag.relNumber == first.relNumber);
6784 : Assert(buf_hdr->tag.blockNum == first.blockNum + i);
6785 : }
6786 :
6787 2794118 : if (is_temp)
6788 16812 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
6789 : else
6790 2777306 : buf_state = LockBufHdr(buf_hdr);
6791 :
6792 : /* verify the buffer is in the expected state */
6793 : Assert(buf_state & BM_TAG_VALID);
6794 : if (is_write)
6795 : {
6796 : Assert(buf_state & BM_VALID);
6797 : Assert(buf_state & BM_DIRTY);
6798 : }
6799 : else
6800 : {
6801 : Assert(!(buf_state & BM_VALID));
6802 : Assert(!(buf_state & BM_DIRTY));
6803 : }
6804 :
6805 : /* temp buffers don't use BM_IO_IN_PROGRESS */
6806 2794118 : if (!is_temp)
6807 : Assert(buf_state & BM_IO_IN_PROGRESS);
6808 :
6809 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) >= 1);
6810 :
6811 : /*
6812 : * Reflect that the buffer is now owned by the AIO subsystem.
6813 : *
6814 : * For local buffers: This can't be done just via LocalRefCount, as
6815 : * one might initially think, as this backend could error out while
6816 : * AIO is still in progress, releasing all the pins by the backend
6817 : * itself.
6818 : *
6819 : * This pin is released again in TerminateBufferIO().
6820 : */
6821 2794118 : buf_state += BUF_REFCOUNT_ONE;
6822 2794118 : buf_hdr->io_wref = io_ref;
6823 :
6824 2794118 : if (is_temp)
6825 16812 : pg_atomic_unlocked_write_u32(&buf_hdr->state, buf_state);
6826 : else
6827 2777306 : UnlockBufHdr(buf_hdr, buf_state);
6828 :
6829 : /*
6830 : * Ensure the content lock that prevents buffer modifications while
6831 : * the buffer is being written out is not released early due to an
6832 : * error.
6833 : */
6834 2794118 : if (is_write && !is_temp)
6835 : {
6836 : LWLock *content_lock;
6837 :
6838 0 : content_lock = BufferDescriptorGetContentLock(buf_hdr);
6839 :
6840 : Assert(LWLockHeldByMe(content_lock));
6841 :
6842 : /*
6843 : * Lock is now owned by AIO subsystem.
6844 : */
6845 0 : LWLockDisown(content_lock);
6846 : }
6847 :
6848 : /*
6849 : * Stop tracking this buffer via the resowner - the AIO system now
6850 : * keeps track.
6851 : */
6852 2794118 : if (!is_temp)
6853 2777306 : ResourceOwnerForgetBufferIO(CurrentResourceOwner, buffer);
6854 : }
6855 2464496 : }
6856 :
6857 : /*
6858 : * Decode readv errors as encoded by buffer_readv_encode_error().
6859 : */
6860 : static inline void
6861 698 : buffer_readv_decode_error(PgAioResult result,
6862 : bool *zeroed_any,
6863 : bool *ignored_any,
6864 : uint8 *zeroed_or_error_count,
6865 : uint8 *checkfail_count,
6866 : uint8 *first_off)
6867 : {
6868 698 : uint32 rem_error = result.error_data;
6869 :
6870 : /* see static asserts in buffer_readv_encode_error */
6871 : #define READV_COUNT_BITS 7
6872 : #define READV_COUNT_MASK ((1 << READV_COUNT_BITS) - 1)
6873 :
6874 698 : *zeroed_any = rem_error & 1;
6875 698 : rem_error >>= 1;
6876 :
6877 698 : *ignored_any = rem_error & 1;
6878 698 : rem_error >>= 1;
6879 :
6880 698 : *zeroed_or_error_count = rem_error & READV_COUNT_MASK;
6881 698 : rem_error >>= READV_COUNT_BITS;
6882 :
6883 698 : *checkfail_count = rem_error & READV_COUNT_MASK;
6884 698 : rem_error >>= READV_COUNT_BITS;
6885 :
6886 698 : *first_off = rem_error & READV_COUNT_MASK;
6887 698 : rem_error >>= READV_COUNT_BITS;
6888 698 : }
6889 :
6890 : /*
6891 : * Helper to encode errors for buffer_readv_complete()
6892 : *
6893 : * Errors are encoded as follows:
6894 : * - bit 0 indicates whether any page was zeroed (1) or not (0)
6895 : * - bit 1 indicates whether any checksum failure was ignored (1) or not (0)
6896 : * - next READV_COUNT_BITS bits indicate the number of errored or zeroed pages
6897 : * - next READV_COUNT_BITS bits indicate the number of checksum failures
6898 : * - next READV_COUNT_BITS bits indicate the first offset of the first page
6899 : * that was errored or zeroed or, if no errors/zeroes, the first ignored
6900 : * checksum
6901 : */
6902 : static inline void
6903 384 : buffer_readv_encode_error(PgAioResult *result,
6904 : bool is_temp,
6905 : bool zeroed_any,
6906 : bool ignored_any,
6907 : uint8 error_count,
6908 : uint8 zeroed_count,
6909 : uint8 checkfail_count,
6910 : uint8 first_error_off,
6911 : uint8 first_zeroed_off,
6912 : uint8 first_ignored_off)
6913 : {
6914 :
6915 384 : uint8 shift = 0;
6916 384 : uint8 zeroed_or_error_count =
6917 : error_count > 0 ? error_count : zeroed_count;
6918 : uint8 first_off;
6919 :
6920 : StaticAssertStmt(PG_IOV_MAX <= 1 << READV_COUNT_BITS,
6921 : "PG_IOV_MAX is bigger than reserved space for error data");
6922 : StaticAssertStmt((1 + 1 + 3 * READV_COUNT_BITS) <= PGAIO_RESULT_ERROR_BITS,
6923 : "PGAIO_RESULT_ERROR_BITS is insufficient for buffer_readv");
6924 :
6925 : /*
6926 : * We only have space to encode one offset - but luckily that's good
6927 : * enough. If there is an error, the error is the interesting offset, same
6928 : * with a zeroed buffer vs an ignored buffer.
6929 : */
6930 384 : if (error_count > 0)
6931 188 : first_off = first_error_off;
6932 196 : else if (zeroed_count > 0)
6933 160 : first_off = first_zeroed_off;
6934 : else
6935 36 : first_off = first_ignored_off;
6936 :
6937 : Assert(!zeroed_any || error_count == 0);
6938 :
6939 384 : result->error_data = 0;
6940 :
6941 384 : result->error_data |= zeroed_any << shift;
6942 384 : shift += 1;
6943 :
6944 384 : result->error_data |= ignored_any << shift;
6945 384 : shift += 1;
6946 :
6947 384 : result->error_data |= ((uint32) zeroed_or_error_count) << shift;
6948 384 : shift += READV_COUNT_BITS;
6949 :
6950 384 : result->error_data |= ((uint32) checkfail_count) << shift;
6951 384 : shift += READV_COUNT_BITS;
6952 :
6953 384 : result->error_data |= ((uint32) first_off) << shift;
6954 384 : shift += READV_COUNT_BITS;
6955 :
6956 384 : result->id = is_temp ? PGAIO_HCB_LOCAL_BUFFER_READV :
6957 : PGAIO_HCB_SHARED_BUFFER_READV;
6958 :
6959 384 : if (error_count > 0)
6960 188 : result->status = PGAIO_RS_ERROR;
6961 : else
6962 196 : result->status = PGAIO_RS_WARNING;
6963 :
6964 : /*
6965 : * The encoding is complicated enough to warrant cross-checking it against
6966 : * the decode function.
6967 : */
6968 : #ifdef USE_ASSERT_CHECKING
6969 : {
6970 : bool zeroed_any_2,
6971 : ignored_any_2;
6972 : uint8 zeroed_or_error_count_2,
6973 : checkfail_count_2,
6974 : first_off_2;
6975 :
6976 : buffer_readv_decode_error(*result,
6977 : &zeroed_any_2, &ignored_any_2,
6978 : &zeroed_or_error_count_2,
6979 : &checkfail_count_2,
6980 : &first_off_2);
6981 : Assert(zeroed_any == zeroed_any_2);
6982 : Assert(ignored_any == ignored_any_2);
6983 : Assert(zeroed_or_error_count == zeroed_or_error_count_2);
6984 : Assert(checkfail_count == checkfail_count_2);
6985 : Assert(first_off == first_off_2);
6986 : }
6987 : #endif
6988 :
6989 : #undef READV_COUNT_BITS
6990 : #undef READV_COUNT_MASK
6991 384 : }
6992 :
6993 : /*
6994 : * Helper for AIO readv completion callbacks, supporting both shared and temp
6995 : * buffers. Gets called once for each buffer in a multi-page read.
6996 : */
6997 : static pg_attribute_always_inline void
6998 2518898 : buffer_readv_complete_one(PgAioTargetData *td, uint8 buf_off, Buffer buffer,
6999 : uint8 flags, bool failed, bool is_temp,
7000 : bool *buffer_invalid,
7001 : bool *failed_checksum,
7002 : bool *ignored_checksum,
7003 : bool *zeroed_buffer)
7004 : {
7005 2518898 : BufferDesc *buf_hdr = is_temp ?
7006 16812 : GetLocalBufferDescriptor(-buffer - 1)
7007 2518898 : : GetBufferDescriptor(buffer - 1);
7008 2518898 : BufferTag tag = buf_hdr->tag;
7009 2518898 : char *bufdata = BufferGetBlock(buffer);
7010 : uint32 set_flag_bits;
7011 : int piv_flags;
7012 :
7013 : /* check that the buffer is in the expected state for a read */
7014 : #ifdef USE_ASSERT_CHECKING
7015 : {
7016 : uint32 buf_state = pg_atomic_read_u32(&buf_hdr->state);
7017 :
7018 : Assert(buf_state & BM_TAG_VALID);
7019 : Assert(!(buf_state & BM_VALID));
7020 : /* temp buffers don't use BM_IO_IN_PROGRESS */
7021 : if (!is_temp)
7022 : Assert(buf_state & BM_IO_IN_PROGRESS);
7023 : Assert(!(buf_state & BM_DIRTY));
7024 : }
7025 : #endif
7026 :
7027 2518898 : *buffer_invalid = false;
7028 2518898 : *failed_checksum = false;
7029 2518898 : *ignored_checksum = false;
7030 2518898 : *zeroed_buffer = false;
7031 :
7032 : /*
7033 : * We ask PageIsVerified() to only log the message about checksum errors,
7034 : * as the completion might be run in any backend (or IO workers). We will
7035 : * report checksum errors in buffer_readv_report().
7036 : */
7037 2518898 : piv_flags = PIV_LOG_LOG;
7038 :
7039 : /* the local zero_damaged_pages may differ from the definer's */
7040 2518898 : if (flags & READ_BUFFERS_IGNORE_CHECKSUM_FAILURES)
7041 76 : piv_flags |= PIV_IGNORE_CHECKSUM_FAILURE;
7042 :
7043 : /* Check for garbage data. */
7044 2518898 : if (!failed)
7045 : {
7046 : /*
7047 : * If the buffer is not currently pinned by this backend, e.g. because
7048 : * we're completing this IO after an error, the buffer data will have
7049 : * been marked as inaccessible when the buffer was unpinned. The AIO
7050 : * subsystem holds a pin, but that doesn't prevent the buffer from
7051 : * having been marked as inaccessible. The completion might also be
7052 : * executed in a different process.
7053 : */
7054 : #ifdef USE_VALGRIND
7055 : if (!BufferIsPinned(buffer))
7056 : VALGRIND_MAKE_MEM_DEFINED(bufdata, BLCKSZ);
7057 : #endif
7058 :
7059 2518840 : if (!PageIsVerified((Page) bufdata, tag.blockNum, piv_flags,
7060 : failed_checksum))
7061 : {
7062 192 : if (flags & READ_BUFFERS_ZERO_ON_ERROR)
7063 : {
7064 92 : memset(bufdata, 0, BLCKSZ);
7065 92 : *zeroed_buffer = true;
7066 : }
7067 : else
7068 : {
7069 100 : *buffer_invalid = true;
7070 : /* mark buffer as having failed */
7071 100 : failed = true;
7072 : }
7073 : }
7074 2518648 : else if (*failed_checksum)
7075 24 : *ignored_checksum = true;
7076 :
7077 : /* undo what we did above */
7078 : #ifdef USE_VALGRIND
7079 : if (!BufferIsPinned(buffer))
7080 : VALGRIND_MAKE_MEM_NOACCESS(bufdata, BLCKSZ);
7081 : #endif
7082 :
7083 : /*
7084 : * Immediately log a message about the invalid page, but only to the
7085 : * server log. The reason to do so immediately is that this may be
7086 : * executed in a different backend than the one that originated the
7087 : * request. The reason to do so immediately is that the originator
7088 : * might not process the query result immediately (because it is busy
7089 : * doing another part of query processing) or at all (e.g. if it was
7090 : * cancelled or errored out due to another IO also failing). The
7091 : * definer of the IO will emit an ERROR or WARNING when processing the
7092 : * IO's results
7093 : *
7094 : * To avoid duplicating the code to emit these log messages, we reuse
7095 : * buffer_readv_report().
7096 : */
7097 2518840 : if (*buffer_invalid || *failed_checksum || *zeroed_buffer)
7098 : {
7099 216 : PgAioResult result_one = {0};
7100 :
7101 216 : buffer_readv_encode_error(&result_one, is_temp,
7102 216 : *zeroed_buffer,
7103 216 : *ignored_checksum,
7104 216 : *buffer_invalid,
7105 216 : *zeroed_buffer ? 1 : 0,
7106 216 : *failed_checksum ? 1 : 0,
7107 : buf_off, buf_off, buf_off);
7108 216 : pgaio_result_report(result_one, td, LOG_SERVER_ONLY);
7109 : }
7110 : }
7111 :
7112 : /* Terminate I/O and set BM_VALID. */
7113 2518898 : set_flag_bits = failed ? BM_IO_ERROR : BM_VALID;
7114 2518898 : if (is_temp)
7115 16812 : TerminateLocalBufferIO(buf_hdr, false, set_flag_bits, true);
7116 : else
7117 2502086 : TerminateBufferIO(buf_hdr, false, set_flag_bits, false, true);
7118 :
7119 : /*
7120 : * Call the BUFFER_READ_DONE tracepoint in the callback, even though the
7121 : * callback may not be executed in the same backend that called
7122 : * BUFFER_READ_START. The alternative would be to defer calling the
7123 : * tracepoint to a later point (e.g. the local completion callback for
7124 : * shared buffer reads), which seems even less helpful.
7125 : */
7126 : TRACE_POSTGRESQL_BUFFER_READ_DONE(tag.forkNum,
7127 : tag.blockNum,
7128 : tag.spcOid,
7129 : tag.dbOid,
7130 : tag.relNumber,
7131 : is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
7132 : false);
7133 2518898 : }
7134 :
7135 : /*
7136 : * Perform completion handling of a single AIO read. This read may cover
7137 : * multiple blocks / buffers.
7138 : *
7139 : * Shared between shared and local buffers, to reduce code duplication.
7140 : */
7141 : static pg_attribute_always_inline PgAioResult
7142 2255356 : buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7143 : uint8 cb_data, bool is_temp)
7144 : {
7145 2255356 : PgAioResult result = prior_result;
7146 2255356 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
7147 2255356 : uint8 first_error_off = 0;
7148 2255356 : uint8 first_zeroed_off = 0;
7149 2255356 : uint8 first_ignored_off = 0;
7150 2255356 : uint8 error_count = 0;
7151 2255356 : uint8 zeroed_count = 0;
7152 2255356 : uint8 ignored_count = 0;
7153 2255356 : uint8 checkfail_count = 0;
7154 : uint64 *io_data;
7155 : uint8 handle_data_len;
7156 :
7157 : if (is_temp)
7158 : {
7159 : Assert(td->smgr.is_temp);
7160 : Assert(pgaio_io_get_owner(ioh) == MyProcNumber);
7161 : }
7162 : else
7163 : Assert(!td->smgr.is_temp);
7164 :
7165 : /*
7166 : * Iterate over all the buffers affected by this IO and call the
7167 : * per-buffer completion function for each buffer.
7168 : */
7169 2255356 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
7170 4774254 : for (uint8 buf_off = 0; buf_off < handle_data_len; buf_off++)
7171 : {
7172 2518898 : Buffer buf = io_data[buf_off];
7173 : bool failed;
7174 2518898 : bool failed_verification = false;
7175 2518898 : bool failed_checksum = false;
7176 2518898 : bool zeroed_buffer = false;
7177 2518898 : bool ignored_checksum = false;
7178 :
7179 : Assert(BufferIsValid(buf));
7180 :
7181 : /*
7182 : * If the entire I/O failed on a lower-level, each buffer needs to be
7183 : * marked as failed. In case of a partial read, the first few buffers
7184 : * may be ok.
7185 : */
7186 2518898 : failed =
7187 2518898 : prior_result.status == PGAIO_RS_ERROR
7188 2518898 : || prior_result.result <= buf_off;
7189 :
7190 2518898 : buffer_readv_complete_one(td, buf_off, buf, cb_data, failed, is_temp,
7191 : &failed_verification,
7192 : &failed_checksum,
7193 : &ignored_checksum,
7194 : &zeroed_buffer);
7195 :
7196 : /*
7197 : * Track information about the number of different kinds of error
7198 : * conditions across all pages, as there can be multiple pages failing
7199 : * verification as part of one IO.
7200 : */
7201 2518898 : if (failed_verification && !zeroed_buffer && error_count++ == 0)
7202 88 : first_error_off = buf_off;
7203 2518898 : if (zeroed_buffer && zeroed_count++ == 0)
7204 68 : first_zeroed_off = buf_off;
7205 2518898 : if (ignored_checksum && ignored_count++ == 0)
7206 20 : first_ignored_off = buf_off;
7207 2518898 : if (failed_checksum)
7208 64 : checkfail_count++;
7209 : }
7210 :
7211 : /*
7212 : * If the smgr read succeeded [partially] and page verification failed for
7213 : * some of the pages, adjust the IO's result state appropriately.
7214 : */
7215 2255356 : if (prior_result.status != PGAIO_RS_ERROR &&
7216 2255250 : (error_count > 0 || ignored_count > 0 || zeroed_count > 0))
7217 : {
7218 168 : buffer_readv_encode_error(&result, is_temp,
7219 : zeroed_count > 0, ignored_count > 0,
7220 : error_count, zeroed_count, checkfail_count,
7221 : first_error_off, first_zeroed_off,
7222 : first_ignored_off);
7223 168 : pgaio_result_report(result, td, DEBUG1);
7224 : }
7225 :
7226 : /*
7227 : * For shared relations this reporting is done in
7228 : * shared_buffer_readv_complete_local().
7229 : */
7230 2255356 : if (is_temp && checkfail_count > 0)
7231 4 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
7232 : checkfail_count);
7233 :
7234 2255356 : return result;
7235 : }
7236 :
7237 : /*
7238 : * AIO error reporting callback for aio_shared_buffer_readv_cb and
7239 : * aio_local_buffer_readv_cb.
7240 : *
7241 : * The error is encoded / decoded in buffer_readv_encode_error() /
7242 : * buffer_readv_decode_error().
7243 : */
7244 : static void
7245 544 : buffer_readv_report(PgAioResult result, const PgAioTargetData *td,
7246 : int elevel)
7247 : {
7248 544 : int nblocks = td->smgr.nblocks;
7249 544 : BlockNumber first = td->smgr.blockNum;
7250 544 : BlockNumber last = first + nblocks - 1;
7251 544 : ProcNumber errProc =
7252 544 : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER;
7253 : RelPathStr rpath =
7254 544 : relpathbackend(td->smgr.rlocator, errProc, td->smgr.forkNum);
7255 : bool zeroed_any,
7256 : ignored_any;
7257 : uint8 zeroed_or_error_count,
7258 : checkfail_count,
7259 : first_off;
7260 : uint8 affected_count;
7261 : const char *msg_one,
7262 : *msg_mult,
7263 : *det_mult,
7264 : *hint_mult;
7265 :
7266 544 : buffer_readv_decode_error(result, &zeroed_any, &ignored_any,
7267 : &zeroed_or_error_count,
7268 : &checkfail_count,
7269 : &first_off);
7270 :
7271 : /*
7272 : * Treat a read that had both zeroed buffers *and* ignored checksums as a
7273 : * special case, it's too irregular to be emitted the same way as the
7274 : * other cases.
7275 : */
7276 544 : if (zeroed_any && ignored_any)
7277 : {
7278 : Assert(zeroed_any && ignored_any);
7279 : Assert(nblocks > 1); /* same block can't be both zeroed and ignored */
7280 : Assert(result.status != PGAIO_RS_ERROR);
7281 8 : affected_count = zeroed_or_error_count;
7282 :
7283 8 : ereport(elevel,
7284 : errcode(ERRCODE_DATA_CORRUPTED),
7285 : errmsg("zeroing %u page(s) and ignoring %u checksum failure(s) among blocks %u..%u of relation \"%s\"",
7286 : affected_count, checkfail_count, first, last, rpath.str),
7287 : affected_count > 1 ?
7288 : errdetail("Block %u held the first zeroed page.",
7289 : first + first_off) : 0,
7290 : errhint_plural("See server log for details about the other %d invalid block.",
7291 : "See server log for details about the other %d invalid blocks.",
7292 : affected_count + checkfail_count - 1,
7293 : affected_count + checkfail_count - 1));
7294 8 : return;
7295 : }
7296 :
7297 : /*
7298 : * The other messages are highly repetitive. To avoid duplicating a long
7299 : * and complicated ereport(), gather the translated format strings
7300 : * separately and then do one common ereport.
7301 : */
7302 536 : if (result.status == PGAIO_RS_ERROR)
7303 : {
7304 : Assert(!zeroed_any); /* can't have invalid pages when zeroing them */
7305 272 : affected_count = zeroed_or_error_count;
7306 272 : msg_one = _("invalid page in block %u of relation \"%s\"");
7307 272 : msg_mult = _("%u invalid pages among blocks %u..%u of relation \"%s\"");
7308 272 : det_mult = _("Block %u held the first invalid page.");
7309 272 : hint_mult = _("See server log for the other %u invalid block(s).");
7310 : }
7311 264 : else if (zeroed_any && !ignored_any)
7312 : {
7313 216 : affected_count = zeroed_or_error_count;
7314 216 : msg_one = _("invalid page in block %u of relation \"%s\"; zeroing out page");
7315 216 : msg_mult = _("zeroing out %u invalid pages among blocks %u..%u of relation \"%s\"");
7316 216 : det_mult = _("Block %u held the first zeroed page.");
7317 216 : hint_mult = _("See server log for the other %u zeroed block(s).");
7318 : }
7319 48 : else if (!zeroed_any && ignored_any)
7320 : {
7321 48 : affected_count = checkfail_count;
7322 48 : msg_one = _("ignoring checksum failure in block %u of relation \"%s\"");
7323 48 : msg_mult = _("ignoring %u checksum failures among blocks %u..%u of relation \"%s\"");
7324 48 : det_mult = _("Block %u held the first ignored page.");
7325 48 : hint_mult = _("See server log for the other %u ignored block(s).");
7326 : }
7327 : else
7328 0 : pg_unreachable();
7329 :
7330 536 : ereport(elevel,
7331 : errcode(ERRCODE_DATA_CORRUPTED),
7332 : affected_count == 1 ?
7333 : errmsg_internal(msg_one, first + first_off, rpath.str) :
7334 : errmsg_internal(msg_mult, affected_count, first, last, rpath.str),
7335 : affected_count > 1 ? errdetail_internal(det_mult, first + first_off) : 0,
7336 : affected_count > 1 ? errhint_internal(hint_mult, affected_count - 1) : 0);
7337 : }
7338 :
7339 : static void
7340 2460890 : shared_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
7341 : {
7342 2460890 : buffer_stage_common(ioh, false, false);
7343 2460890 : }
7344 :
7345 : static PgAioResult
7346 2251750 : shared_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7347 : uint8 cb_data)
7348 : {
7349 2251750 : return buffer_readv_complete(ioh, prior_result, cb_data, false);
7350 : }
7351 :
7352 : /*
7353 : * We need a backend-local completion callback for shared buffers, to be able
7354 : * to report checksum errors correctly. Unfortunately that can only safely
7355 : * happen if the reporting backend has previously called
7356 : * pgstat_prepare_report_checksum_failure(), which we can only guarantee in
7357 : * the backend that started the IO. Hence this callback.
7358 : */
7359 : static PgAioResult
7360 2460890 : shared_buffer_readv_complete_local(PgAioHandle *ioh, PgAioResult prior_result,
7361 : uint8 cb_data)
7362 : {
7363 : bool zeroed_any,
7364 : ignored_any;
7365 : uint8 zeroed_or_error_count,
7366 : checkfail_count,
7367 : first_off;
7368 :
7369 2460890 : if (prior_result.status == PGAIO_RS_OK)
7370 2460736 : return prior_result;
7371 :
7372 154 : buffer_readv_decode_error(prior_result,
7373 : &zeroed_any,
7374 : &ignored_any,
7375 : &zeroed_or_error_count,
7376 : &checkfail_count,
7377 : &first_off);
7378 :
7379 154 : if (checkfail_count)
7380 : {
7381 48 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
7382 :
7383 48 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
7384 : checkfail_count);
7385 : }
7386 :
7387 154 : return prior_result;
7388 : }
7389 :
7390 : static void
7391 3606 : local_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
7392 : {
7393 3606 : buffer_stage_common(ioh, false, true);
7394 3606 : }
7395 :
7396 : static PgAioResult
7397 3606 : local_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7398 : uint8 cb_data)
7399 : {
7400 3606 : return buffer_readv_complete(ioh, prior_result, cb_data, true);
7401 : }
7402 :
7403 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
7404 : const PgAioHandleCallbacks aio_shared_buffer_readv_cb = {
7405 : .stage = shared_buffer_readv_stage,
7406 : .complete_shared = shared_buffer_readv_complete,
7407 : /* need a local callback to report checksum failures */
7408 : .complete_local = shared_buffer_readv_complete_local,
7409 : .report = buffer_readv_report,
7410 : };
7411 :
7412 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
7413 : const PgAioHandleCallbacks aio_local_buffer_readv_cb = {
7414 : .stage = local_buffer_readv_stage,
7415 :
7416 : /*
7417 : * Note that this, in contrast to the shared_buffers case, uses
7418 : * complete_local, as only the issuing backend has access to the required
7419 : * datastructures. This is important in case the IO completion may be
7420 : * consumed incidentally by another backend.
7421 : */
7422 : .complete_local = local_buffer_readv_complete,
7423 : .report = buffer_readv_report,
7424 : };
|
