Age Owner TLA Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * bufmgr.c
4 : * buffer manager interface routines
5 : *
6 : * Portions Copyright (c) 1996-2023, 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 : * ReleaseBuffer() -- unpin a buffer
23 : *
24 : * MarkBufferDirty() -- mark a pinned buffer's contents as "dirty".
25 : * The disk write is delayed until buffer replacement or checkpoint.
26 : *
27 : * See also these files:
28 : * freelist.c -- chooses victim for buffer replacement
29 : * buf_table.c -- manages the buffer lookup table
30 : */
31 : #include "postgres.h"
32 :
33 : #include <sys/file.h>
34 : #include <unistd.h>
35 :
36 : #include "access/tableam.h"
37 : #include "access/xloginsert.h"
38 : #include "access/xlogutils.h"
39 : #include "catalog/catalog.h"
40 : #include "catalog/storage.h"
41 : #include "catalog/storage_xlog.h"
42 : #include "executor/instrument.h"
43 : #include "lib/binaryheap.h"
44 : #include "miscadmin.h"
45 : #include "pg_trace.h"
46 : #include "pgstat.h"
47 : #include "postmaster/bgwriter.h"
48 : #include "storage/buf_internals.h"
49 : #include "storage/bufmgr.h"
50 : #include "storage/ipc.h"
51 : #include "storage/lmgr.h"
52 : #include "storage/proc.h"
53 : #include "storage/smgr.h"
54 : #include "storage/standby.h"
55 : #include "utils/memdebug.h"
56 : #include "utils/ps_status.h"
57 : #include "utils/rel.h"
58 : #include "utils/resowner_private.h"
59 : #include "utils/timestamp.h"
60 :
61 :
62 : /* Note: these two macros only work on shared buffers, not local ones! */
63 : #define BufHdrGetBlock(bufHdr) ((Block) (BufferBlocks + ((Size) (bufHdr)->buf_id) * BLCKSZ))
64 : #define BufferGetLSN(bufHdr) (PageGetLSN(BufHdrGetBlock(bufHdr)))
65 :
66 : /* Note: this macro only works on local buffers, not shared ones! */
67 : #define LocalBufHdrGetBlock(bufHdr) \
68 : LocalBufferBlockPointers[-((bufHdr)->buf_id + 2)]
69 :
70 : /* Bits in SyncOneBuffer's return value */
71 : #define BUF_WRITTEN 0x01
72 : #define BUF_REUSABLE 0x02
73 :
74 : #define RELS_BSEARCH_THRESHOLD 20
75 :
76 : /*
77 : * This is the size (in the number of blocks) above which we scan the
78 : * entire buffer pool to remove the buffers for all the pages of relation
79 : * being dropped. For the relations with size below this threshold, we find
80 : * the buffers by doing lookups in BufMapping table.
81 : */
82 : #define BUF_DROP_FULL_SCAN_THRESHOLD (uint64) (NBuffers / 32)
83 :
84 : typedef struct PrivateRefCountEntry
85 : {
86 : Buffer buffer;
87 : int32 refcount;
88 : } PrivateRefCountEntry;
89 :
90 : /* 64 bytes, about the size of a cache line on common systems */
91 : #define REFCOUNT_ARRAY_ENTRIES 8
92 :
93 : /*
94 : * Status of buffers to checkpoint for a particular tablespace, used
95 : * internally in BufferSync.
96 : */
97 : typedef struct CkptTsStatus
98 : {
99 : /* oid of the tablespace */
100 : Oid tsId;
101 :
102 : /*
103 : * Checkpoint progress for this tablespace. To make progress comparable
104 : * between tablespaces the progress is, for each tablespace, measured as a
105 : * number between 0 and the total number of to-be-checkpointed pages. Each
106 : * page checkpointed in this tablespace increments this space's progress
107 : * by progress_slice.
108 : */
109 : float8 progress;
110 : float8 progress_slice;
111 :
112 : /* number of to-be checkpointed pages in this tablespace */
113 : int num_to_scan;
114 : /* already processed pages in this tablespace */
115 : int num_scanned;
116 :
117 : /* current offset in CkptBufferIds for this tablespace */
118 : int index;
119 : } CkptTsStatus;
120 :
121 : /*
122 : * Type for array used to sort SMgrRelations
123 : *
124 : * FlushRelationsAllBuffers shares the same comparator function with
125 : * DropRelationsAllBuffers. Pointer to this struct and RelFileLocator must be
126 : * compatible.
127 : */
128 : typedef struct SMgrSortArray
129 : {
130 : RelFileLocator rlocator; /* This must be the first member */
131 : SMgrRelation srel;
132 : } SMgrSortArray;
133 :
134 : /* GUC variables */
135 : bool zero_damaged_pages = false;
136 : int bgwriter_lru_maxpages = 100;
137 : double bgwriter_lru_multiplier = 2.0;
138 : bool track_io_timing = false;
139 :
140 : /*
141 : * How many buffers PrefetchBuffer callers should try to stay ahead of their
142 : * ReadBuffer calls by. Zero means "never prefetch". This value is only used
143 : * for buffers not belonging to tablespaces that have their
144 : * effective_io_concurrency parameter set.
145 : */
146 : int effective_io_concurrency = DEFAULT_EFFECTIVE_IO_CONCURRENCY;
147 :
148 : /*
149 : * Like effective_io_concurrency, but used by maintenance code paths that might
150 : * benefit from a higher setting because they work on behalf of many sessions.
151 : * Overridden by the tablespace setting of the same name.
152 : */
153 : int maintenance_io_concurrency = DEFAULT_MAINTENANCE_IO_CONCURRENCY;
154 :
155 : /*
156 : * GUC variables about triggering kernel writeback for buffers written; OS
157 : * dependent defaults are set via the GUC mechanism.
158 : */
159 : int checkpoint_flush_after = DEFAULT_CHECKPOINT_FLUSH_AFTER;
160 : int bgwriter_flush_after = DEFAULT_BGWRITER_FLUSH_AFTER;
161 : int backend_flush_after = DEFAULT_BACKEND_FLUSH_AFTER;
162 :
163 : /* local state for LockBufferForCleanup */
164 : static BufferDesc *PinCountWaitBuf = NULL;
165 :
166 : /*
167 : * Backend-Private refcount management:
168 : *
169 : * Each buffer also has a private refcount that keeps track of the number of
170 : * times the buffer is pinned in the current process. This is so that the
171 : * shared refcount needs to be modified only once if a buffer is pinned more
172 : * than once by an individual backend. It's also used to check that no buffers
173 : * are still pinned at the end of transactions and when exiting.
174 : *
175 : *
176 : * To avoid - as we used to - requiring an array with NBuffers entries to keep
177 : * track of local buffers, we use a small sequentially searched array
178 : * (PrivateRefCountArray) and an overflow hash table (PrivateRefCountHash) to
179 : * keep track of backend local pins.
180 : *
181 : * Until no more than REFCOUNT_ARRAY_ENTRIES buffers are pinned at once, all
182 : * refcounts are kept track of in the array; after that, new array entries
183 : * displace old ones into the hash table. That way a frequently used entry
184 : * can't get "stuck" in the hashtable while infrequent ones clog the array.
185 : *
186 : * Note that in most scenarios the number of pinned buffers will not exceed
187 : * REFCOUNT_ARRAY_ENTRIES.
188 : *
189 : *
190 : * To enter a buffer into the refcount tracking mechanism first reserve a free
191 : * entry using ReservePrivateRefCountEntry() and then later, if necessary,
192 : * fill it with NewPrivateRefCountEntry(). That split lets us avoid doing
193 : * memory allocations in NewPrivateRefCountEntry() which can be important
194 : * because in some scenarios it's called with a spinlock held...
195 : */
196 : static struct PrivateRefCountEntry PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES];
197 : static HTAB *PrivateRefCountHash = NULL;
198 : static int32 PrivateRefCountOverflowed = 0;
199 : static uint32 PrivateRefCountClock = 0;
200 : static PrivateRefCountEntry *ReservedRefCountEntry = NULL;
201 :
202 : static void ReservePrivateRefCountEntry(void);
203 : static PrivateRefCountEntry *NewPrivateRefCountEntry(Buffer buffer);
204 : static PrivateRefCountEntry *GetPrivateRefCountEntry(Buffer buffer, bool do_move);
205 : static inline int32 GetPrivateRefCount(Buffer buffer);
206 : static void ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref);
207 :
208 : /*
209 : * Ensure that the PrivateRefCountArray has sufficient space to store one more
210 : * entry. This has to be called before using NewPrivateRefCountEntry() to fill
2993 heikki.linnakangas 211 ECB : * a new entry - but it's perfectly fine to not use a reserved entry.
212 : */
213 : static void
3002 andres 214 CBC 69663558 : ReservePrivateRefCountEntry(void)
3002 andres 215 ECB : {
216 : /* Already reserved (or freed), nothing to do */
3002 andres 217 GIC 69663558 : if (ReservedRefCountEntry != NULL)
218 66674250 : return;
219 :
220 : /*
221 : * First search for a free entry the array, that'll be sufficient in the
222 : * majority of cases.
223 : */
3002 andres 224 ECB : {
225 : int i;
226 :
3002 andres 227 GIC 6434020 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
3002 andres 228 ECB : {
229 : PrivateRefCountEntry *res;
230 :
3002 andres 231 GIC 6387812 : res = &PrivateRefCountArray[i];
3002 andres 232 ECB :
3002 andres 233 CBC 6387812 : if (res->buffer == InvalidBuffer)
234 : {
3002 andres 235 GIC 2943100 : ReservedRefCountEntry = res;
236 2943100 : return;
237 : }
238 : }
239 : }
240 :
241 : /*
242 : * No luck. All array entries are full. Move one array entry into the hash
243 : * table.
244 : */
245 : {
246 : /*
247 : * Move entry from the current clock position in the array into the
248 : * hashtable. Use that slot.
249 : */
250 : PrivateRefCountEntry *hashent;
2878 bruce 251 ECB : bool found;
3002 andres 252 :
253 : /* select victim slot */
2878 bruce 254 GIC 46208 : ReservedRefCountEntry =
3002 andres 255 CBC 46208 : &PrivateRefCountArray[PrivateRefCountClock++ % REFCOUNT_ARRAY_ENTRIES];
256 :
257 : /* Better be used, otherwise we shouldn't get here. */
258 46208 : Assert(ReservedRefCountEntry->buffer != InvalidBuffer);
3002 andres 259 ECB :
260 : /* enter victim array entry into hashtable */
3002 andres 261 GIC 46208 : hashent = hash_search(PrivateRefCountHash,
62 peter 262 GNC 46208 : &(ReservedRefCountEntry->buffer),
3002 andres 263 ECB : HASH_ENTER,
264 : &found);
3002 andres 265 GIC 46208 : Assert(!found);
3002 andres 266 CBC 46208 : hashent->refcount = ReservedRefCountEntry->refcount;
3002 andres 267 ECB :
268 : /* clear the now free array slot */
3002 andres 269 CBC 46208 : ReservedRefCountEntry->buffer = InvalidBuffer;
3002 andres 270 GIC 46208 : ReservedRefCountEntry->refcount = 0;
271 :
272 46208 : PrivateRefCountOverflowed++;
273 : }
274 : }
275 :
276 : /*
3002 andres 277 ECB : * Fill a previously reserved refcount entry.
278 : */
279 : static PrivateRefCountEntry *
3002 andres 280 GIC 68472110 : NewPrivateRefCountEntry(Buffer buffer)
281 : {
3002 andres 282 ECB : PrivateRefCountEntry *res;
283 :
284 : /* only allowed to be called when a reservation has been made */
3002 andres 285 CBC 68472110 : Assert(ReservedRefCountEntry != NULL);
3002 andres 286 ECB :
287 : /* use up the reserved entry */
3002 andres 288 GIC 68472110 : res = ReservedRefCountEntry;
3002 andres 289 CBC 68472110 : ReservedRefCountEntry = NULL;
3002 andres 290 ECB :
291 : /* and fill it */
3002 andres 292 CBC 68472110 : res->buffer = buffer;
3002 andres 293 GIC 68472110 : res->refcount = 0;
294 :
295 68472110 : return res;
296 : }
297 :
298 : /*
299 : * Return the PrivateRefCount entry for the passed buffer.
300 : *
301 : * Returns NULL if a buffer doesn't have a refcount entry. Otherwise, if
302 : * do_move is true, and the entry resides in the hashtable the entry is
3002 andres 303 ECB : * optimized for frequent access by moving it to the array.
304 : */
305 : static PrivateRefCountEntry *
3002 andres 306 GIC 574705670 : GetPrivateRefCountEntry(Buffer buffer, bool do_move)
307 : {
3144 andres 308 ECB : PrivateRefCountEntry *res;
309 : int i;
310 :
3144 andres 311 GIC 574705670 : Assert(BufferIsValid(buffer));
312 574705670 : Assert(!BufferIsLocal(buffer));
313 :
314 : /*
3144 andres 315 ECB : * First search for references in the array, that'll be sufficient in the
316 : * majority of cases.
317 : */
3144 andres 318 GIC 1583180876 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
3144 andres 319 ECB : {
3144 andres 320 CBC 1514073662 : res = &PrivateRefCountArray[i];
321 :
3144 andres 322 GIC 1514073662 : if (res->buffer == buffer)
323 505598456 : return res;
324 : }
325 :
326 : /*
327 : * By here we know that the buffer, if already pinned, isn't residing in
328 : * the array.
329 : *
3002 andres 330 ECB : * Only look up the buffer in the hashtable if we've previously overflowed
331 : * into it.
332 : */
3002 andres 333 CBC 69107214 : if (PrivateRefCountOverflowed == 0)
3002 andres 334 GIC 68758477 : return NULL;
3144 andres 335 ECB :
62 peter 336 GNC 348737 : res = hash_search(PrivateRefCountHash, &buffer, HASH_FIND, NULL);
3002 andres 337 ECB :
3002 andres 338 GIC 348737 : if (res == NULL)
339 137246 : return NULL;
340 211491 : else if (!do_move)
341 : {
342 : /* caller doesn't want us to move the hash entry into the array */
343 211254 : return res;
344 : }
345 : else
3144 andres 346 ECB : {
347 : /* move buffer from hashtable into the free array slot */
348 : bool found;
3002 349 : PrivateRefCountEntry *free;
3144 350 :
3002 351 : /* Ensure there's a free array slot */
3002 andres 352 CBC 237 : ReservePrivateRefCountEntry();
353 :
354 : /* Use up the reserved slot */
355 237 : Assert(ReservedRefCountEntry != NULL);
356 237 : free = ReservedRefCountEntry;
3002 andres 357 GIC 237 : ReservedRefCountEntry = NULL;
358 237 : Assert(free->buffer == InvalidBuffer);
3144 andres 359 ECB :
3002 360 : /* and fill it */
3002 andres 361 CBC 237 : free->buffer = buffer;
362 237 : free->refcount = res->refcount;
363 :
3002 andres 364 ECB : /* delete from hashtable */
62 peter 365 GNC 237 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
3002 andres 366 GIC 237 : Assert(found);
367 237 : Assert(PrivateRefCountOverflowed > 0);
368 237 : PrivateRefCountOverflowed--;
369 :
370 237 : return free;
3144 andres 371 ECB : }
372 : }
373 :
374 : /*
375 : * Returns how many times the passed buffer is pinned by this backend.
376 : *
377 : * Only works for shared memory buffers!
378 : */
379 : static inline int32
3144 andres 380 GIC 411969948 : GetPrivateRefCount(Buffer buffer)
381 : {
3144 andres 382 ECB : PrivateRefCountEntry *ref;
383 :
3144 andres 384 CBC 411969948 : Assert(BufferIsValid(buffer));
385 411969948 : Assert(!BufferIsLocal(buffer));
3144 andres 386 ECB :
387 : /*
388 : * Not moving the entry - that's ok for the current users, but we might
389 : * want to change this one day.
390 : */
3002 andres 391 GIC 411969948 : ref = GetPrivateRefCountEntry(buffer, false);
392 :
3144 393 411969948 : if (ref == NULL)
3144 andres 394 CBC 423613 : return 0;
3144 andres 395 GIC 411546335 : return ref->refcount;
3144 andres 396 ECB : }
397 :
398 : /*
399 : * Release resources used to track the reference count of a buffer which we no
400 : * longer have pinned and don't want to pin again immediately.
401 : */
402 : static void
3144 andres 403 GIC 68472110 : ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref)
404 : {
405 68472110 : Assert(ref->refcount == 0);
406 :
407 68472110 : if (ref >= &PrivateRefCountArray[0] &&
3144 andres 408 ECB : ref < &PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES])
409 : {
3144 andres 410 GIC 68426139 : ref->buffer = InvalidBuffer;
411 :
412 : /*
3002 andres 413 ECB : * Mark the just used entry as reserved - in many scenarios that
414 : * allows us to avoid ever having to search the array/hash for free
415 : * entries.
416 : */
3002 andres 417 CBC 68426139 : ReservedRefCountEntry = ref;
3144 andres 418 ECB : }
419 : else
420 : {
421 : bool found;
2878 bruce 422 GIC 45971 : Buffer buffer = ref->buffer;
423 :
62 peter 424 GNC 45971 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
3144 andres 425 GIC 45971 : Assert(found);
426 45971 : Assert(PrivateRefCountOverflowed > 0);
427 45971 : PrivateRefCountOverflowed--;
428 : }
429 68472110 : }
430 :
431 : /*
432 : * BufferIsPinned
433 : * True iff the buffer is pinned (also checks for valid buffer number).
434 : *
435 : * NOTE: what we check here is that *this* backend holds a pin on
436 : * the buffer. We do not care whether some other backend does.
437 : */
438 : #define BufferIsPinned(bufnum) \
439 : ( \
440 : !BufferIsValid(bufnum) ? \
441 : false \
442 : : \
443 : BufferIsLocal(bufnum) ? \
444 : (LocalRefCount[-(bufnum) - 1] > 0) \
445 : : \
446 : (GetPrivateRefCount(bufnum) > 0) \
447 : )
448 :
449 :
450 : static Buffer ReadBuffer_common(SMgrRelation smgr, char relpersistence,
451 : ForkNumber forkNum, BlockNumber blockNum,
452 : ReadBufferMode mode, BufferAccessStrategy strategy,
453 : bool *hit);
454 : static BlockNumber ExtendBufferedRelCommon(ExtendBufferedWhat eb,
455 : ForkNumber fork,
456 : BufferAccessStrategy strategy,
457 : uint32 flags,
458 : uint32 extend_by,
459 : BlockNumber extend_upto,
460 : Buffer *buffers,
461 : uint32 *extended_by);
462 : static BlockNumber ExtendBufferedRelShared(ExtendBufferedWhat eb,
463 : ForkNumber fork,
464 : BufferAccessStrategy strategy,
465 : uint32 flags,
466 : uint32 extend_by,
467 : BlockNumber extend_upto,
468 : Buffer *buffers,
469 : uint32 *extended_by);
470 : static bool PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy);
471 : static void PinBuffer_Locked(BufferDesc *buf);
472 : static void UnpinBuffer(BufferDesc *buf);
473 : static void BufferSync(int flags);
474 : static uint32 WaitBufHdrUnlocked(BufferDesc *buf);
475 : static int SyncOneBuffer(int buf_id, bool skip_recently_used,
476 : WritebackContext *wb_context);
477 : static void WaitIO(BufferDesc *buf);
478 : static bool StartBufferIO(BufferDesc *buf, bool forInput);
479 : static void TerminateBufferIO(BufferDesc *buf, bool clear_dirty,
480 : uint32 set_flag_bits);
481 : static void shared_buffer_write_error_callback(void *arg);
482 : static void local_buffer_write_error_callback(void *arg);
483 : static BufferDesc *BufferAlloc(SMgrRelation smgr,
484 : char relpersistence,
485 : ForkNumber forkNum,
486 : BlockNumber blockNum,
487 : BufferAccessStrategy strategy,
488 : bool *foundPtr, IOContext io_context);
489 : static Buffer GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context);
490 : static void FlushBuffer(BufferDesc *buf, SMgrRelation reln,
491 : IOObject io_object, IOContext io_context);
492 : static void FindAndDropRelationBuffers(RelFileLocator rlocator,
493 : ForkNumber forkNum,
494 : BlockNumber nForkBlock,
495 : BlockNumber firstDelBlock);
496 : static void RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
497 : RelFileLocator dstlocator,
498 : ForkNumber forkNum, bool permanent);
499 : static void AtProcExit_Buffers(int code, Datum arg);
500 : static void CheckForBufferLeaks(void);
501 : static int rlocator_comparator(const void *p1, const void *p2);
502 : static inline int buffertag_comparator(const BufferTag *ba, const BufferTag *bb);
503 : static inline int ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b);
504 : static int ts_ckpt_progress_comparator(Datum a, Datum b, void *arg);
505 :
506 :
507 : /*
508 : * Implementation of PrefetchBuffer() for shared buffers.
509 : */
510 : PrefetchBufferResult
1096 tmunro 511 878497 : PrefetchSharedBuffer(SMgrRelation smgr_reln,
512 : ForkNumber forkNum,
513 : BlockNumber blockNum)
514 : {
515 878497 : PrefetchBufferResult result = {InvalidBuffer, false};
516 : BufferTag newTag; /* identity of requested block */
1096 tmunro 517 ECB : uint32 newHash; /* hash value for newTag */
518 : LWLock *newPartitionLock; /* buffer partition lock for it */
519 : int buf_id;
520 :
1096 tmunro 521 CBC 878497 : Assert(BlockNumberIsValid(blockNum));
522 :
523 : /* create a tag so we can lookup the buffer */
256 rhaas 524 GNC 878497 : InitBufferTag(&newTag, &smgr_reln->smgr_rlocator.locator,
525 : forkNum, blockNum);
526 :
1096 tmunro 527 ECB : /* determine its hash code and partition lock ID */
1096 tmunro 528 GIC 878497 : newHash = BufTableHashCode(&newTag);
529 878497 : newPartitionLock = BufMappingPartitionLock(newHash);
1096 tmunro 530 ECB :
531 : /* see if the block is in the buffer pool already */
1096 tmunro 532 GIC 878497 : LWLockAcquire(newPartitionLock, LW_SHARED);
533 878497 : buf_id = BufTableLookup(&newTag, newHash);
1096 tmunro 534 CBC 878497 : LWLockRelease(newPartitionLock);
1096 tmunro 535 ECB :
536 : /* If not in buffers, initiate prefetch */
1096 tmunro 537 GIC 878497 : if (buf_id < 0)
1096 tmunro 538 ECB : {
539 : #ifdef USE_PREFETCH
540 : /*
541 : * Try to initiate an asynchronous read. This returns false in
542 : * recovery if the relation file doesn't exist.
543 : */
1 tmunro 544 GNC 424364 : if ((io_direct_flags & IO_DIRECT_DATA) == 0 &&
545 212103 : smgrprefetch(smgr_reln, forkNum, blockNum))
546 : {
1096 tmunro 547 GIC 212103 : result.initiated_io = true;
548 : }
549 : #endif /* USE_PREFETCH */
550 : }
551 : else
552 : {
1096 tmunro 553 ECB : /*
554 : * Report the buffer it was in at that time. The caller may be able
555 : * to avoid a buffer table lookup, but it's not pinned and it must be
556 : * rechecked!
557 : */
1096 tmunro 558 GIC 666236 : result.recent_buffer = buf_id + 1;
559 : }
560 :
561 : /*
562 : * If the block *is* in buffers, we do nothing. This is not really ideal:
563 : * the block might be just about to be evicted, which would be stupid
564 : * since we know we are going to need it soon. But the only easy answer
565 : * is to bump the usage_count, which does not seem like a great solution:
566 : * when the caller does ultimately touch the block, usage_count would get
1096 tmunro 567 ECB : * bumped again, resulting in too much favoritism for blocks that are
568 : * involved in a prefetch sequence. A real fix would involve some
569 : * additional per-buffer state, and it's not clear that there's enough of
570 : * a problem to justify that.
571 : */
572 :
1096 tmunro 573 GIC 878497 : return result;
574 : }
575 :
576 : /*
577 : * PrefetchBuffer -- initiate asynchronous read of a block of a relation
578 : *
579 : * This is named by analogy to ReadBuffer but doesn't actually allocate a
580 : * buffer. Instead it tries to ensure that a future ReadBuffer for the given
581 : * block will not be delayed by the I/O. Prefetching is optional.
1096 tmunro 582 ECB : *
583 : * There are three possible outcomes:
584 : *
585 : * 1. If the block is already cached, the result includes a valid buffer that
586 : * could be used by the caller to avoid the need for a later buffer lookup, but
587 : * it's not pinned, so the caller must recheck it.
588 : *
589 : * 2. If the kernel has been asked to initiate I/O, the initiated_io member is
590 : * true. Currently there is no way to know if the data was already cached by
591 : * the kernel and therefore didn't really initiate I/O, and no way to know when
592 : * the I/O completes other than using synchronous ReadBuffer().
593 : *
594 : * 3. Otherwise, the buffer wasn't already cached by PostgreSQL, and
595 : * USE_PREFETCH is not defined (this build doesn't support prefetching due to
596 : * lack of a kernel facility), direct I/O is enabled, or the underlying
597 : * relation file wasn't found and we are in recovery. (If the relation file
598 : * wasn't found and we are not in recovery, an error is raised).
599 : */
600 : PrefetchBufferResult
5200 tgl 601 GIC 420321 : PrefetchBuffer(Relation reln, ForkNumber forkNum, BlockNumber blockNum)
602 : {
603 420321 : Assert(RelationIsValid(reln));
604 420321 : Assert(BlockNumberIsValid(blockNum));
605 :
4500 rhaas 606 420321 : if (RelationUsesLocalBuffers(reln))
607 : {
608 : /* see comments in ReadBufferExtended */
5122 tgl 609 6244 : if (RELATION_IS_OTHER_TEMP(reln))
5122 tgl 610 LBC 0 : ereport(ERROR,
611 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2118 tgl 612 ECB : errmsg("cannot access temporary tables of other sessions")));
5122 613 :
614 : /* pass it off to localbuf.c */
636 tgl 615 CBC 6244 : return PrefetchLocalBuffer(RelationGetSmgr(reln), forkNum, blockNum);
616 : }
617 : else
5200 tgl 618 ECB : {
1096 tmunro 619 EUB : /* pass it to the shared buffer version */
636 tgl 620 GIC 414077 : return PrefetchSharedBuffer(RelationGetSmgr(reln), forkNum, blockNum);
621 : }
622 : }
623 :
731 tmunro 624 ECB : /*
625 : * ReadRecentBuffer -- try to pin a block in a recently observed buffer
626 : *
627 : * Compared to ReadBuffer(), this avoids a buffer mapping lookup when it's
628 : * successful. Return true if the buffer is valid and still has the expected
629 : * tag. In that case, the buffer is pinned and the usage count is bumped.
630 : */
631 : bool
277 rhaas 632 GNC 423615 : ReadRecentBuffer(RelFileLocator rlocator, ForkNumber forkNum, BlockNumber blockNum,
633 : Buffer recent_buffer)
634 : {
635 : BufferDesc *bufHdr;
636 : BufferTag tag;
637 : uint32 buf_state;
638 : bool have_private_ref;
639 :
731 tmunro 640 GIC 423615 : Assert(BufferIsValid(recent_buffer));
731 tmunro 641 ECB :
731 tmunro 642 GIC 423615 : ResourceOwnerEnlargeBuffers(CurrentResourceOwner);
643 423615 : ReservePrivateRefCountEntry();
256 rhaas 644 GNC 423615 : InitBufferTag(&tag, &rlocator, forkNum, blockNum);
645 :
731 tmunro 646 GIC 423615 : if (BufferIsLocal(recent_buffer))
647 : {
258 heikki.linnakangas 648 UIC 0 : int b = -recent_buffer - 1;
258 heikki.linnakangas 649 ECB :
258 heikki.linnakangas 650 UIC 0 : bufHdr = GetLocalBufferDescriptor(b);
731 tmunro 651 LBC 0 : buf_state = pg_atomic_read_u32(&bufHdr->state);
731 tmunro 652 ECB :
653 : /* Is it still valid and holding the right tag? */
256 rhaas 654 UNC 0 : if ((buf_state & BM_VALID) && BufferTagsEqual(&tag, &bufHdr->tag))
731 tmunro 655 ECB : {
4 andres 656 UNC 0 : PinLocalBuffer(bufHdr, true);
657 :
367 tmunro 658 UIC 0 : pgBufferUsage.local_blks_hit++;
659 :
731 660 0 : return true;
731 tmunro 661 ECB : }
662 : }
663 : else
664 : {
731 tmunro 665 GIC 423615 : bufHdr = GetBufferDescriptor(recent_buffer - 1);
666 423615 : have_private_ref = GetPrivateRefCount(recent_buffer) > 0;
667 :
668 : /*
731 tmunro 669 ECB : * Do we already have this buffer pinned with a private reference? If
670 : * so, it must be valid and it is safe to check the tag without
671 : * locking. If not, we have to lock the header first and then check.
672 : */
731 tmunro 673 GIC 423615 : if (have_private_ref)
731 tmunro 674 CBC 4 : buf_state = pg_atomic_read_u32(&bufHdr->state);
675 : else
731 tmunro 676 GIC 423611 : buf_state = LockBufHdr(bufHdr);
677 :
256 rhaas 678 GNC 423615 : if ((buf_state & BM_VALID) && BufferTagsEqual(&tag, &bufHdr->tag))
679 : {
680 : /*
731 tmunro 681 ECB : * It's now safe to pin the buffer. We can't pin first and ask
332 tgl 682 EUB : * questions later, because it might confuse code paths like
683 : * InvalidateBuffer() if we pinned a random non-matching buffer.
731 tmunro 684 ECB : */
731 tmunro 685 GIC 422006 : if (have_private_ref)
731 tmunro 686 LBC 0 : PinBuffer(bufHdr, NULL); /* bump pin count */
687 : else
731 tmunro 688 CBC 422006 : PinBuffer_Locked(bufHdr); /* pin for first time */
689 :
367 tmunro 690 GIC 422006 : pgBufferUsage.shared_blks_hit++;
691 :
731 tmunro 692 CBC 422006 : return true;
731 tmunro 693 ECB : }
694 :
695 : /* If we locked the header above, now unlock. */
731 tmunro 696 CBC 1609 : if (!have_private_ref)
731 tmunro 697 GIC 1605 : UnlockBufHdr(bufHdr, buf_state);
698 : }
699 :
700 1609 : return false;
701 : }
702 :
703 : /*
5273 heikki.linnakangas 704 ECB : * ReadBuffer -- a shorthand for ReadBufferExtended, for reading from main
705 : * fork with RBM_NORMAL mode and default strategy.
706 : */
707 : Buffer
5273 heikki.linnakangas 708 GIC 50153010 : ReadBuffer(Relation reln, BlockNumber blockNum)
709 : {
710 50153010 : return ReadBufferExtended(reln, MAIN_FORKNUM, blockNum, RBM_NORMAL, NULL);
711 : }
712 :
713 : /*
714 : * ReadBufferExtended -- returns a buffer containing the requested
715 : * block of the requested relation. If the blknum
716 : * requested is P_NEW, extend the relation file and
717 : * allocate a new block. (Caller is responsible for
718 : * ensuring that only one backend tries to extend a
719 : * relation at the same time!)
720 : *
721 : * Returns: the buffer number for the buffer containing
722 : * the block read. The returned buffer has been pinned.
723 : * Does not return on error --- elog's instead.
724 : *
725 : * Assume when this function is called, that reln has been opened already.
726 : *
727 : * In RBM_NORMAL mode, the page is read from disk, and the page header is
728 : * validated. An error is thrown if the page header is not valid. (But
729 : * note that an all-zero page is considered "valid"; see
730 : * PageIsVerifiedExtended().)
731 : *
732 : * RBM_ZERO_ON_ERROR is like the normal mode, but if the page header is not
733 : * valid, the page is zeroed instead of throwing an error. This is intended
734 : * for non-critical data, where the caller is prepared to repair errors.
735 : *
736 : * In RBM_ZERO_AND_LOCK mode, if the page isn't in buffer cache already, it's
737 : * filled with zeros instead of reading it from disk. Useful when the caller
738 : * is going to fill the page from scratch, since this saves I/O and avoids
739 : * unnecessary failure if the page-on-disk has corrupt page headers.
740 : * The page is returned locked to ensure that the caller has a chance to
741 : * initialize the page before it's made visible to others.
742 : * Caution: do not use this mode to read a page that is beyond the relation's
743 : * current physical EOF; that is likely to cause problems in md.c when
744 : * the page is modified and written out. P_NEW is OK, though.
745 : *
746 : * RBM_ZERO_AND_CLEANUP_LOCK is the same as RBM_ZERO_AND_LOCK, but acquires
747 : * a cleanup-strength lock on the page.
748 : *
749 : * RBM_NORMAL_NO_LOG mode is treated the same as RBM_NORMAL here.
750 : *
5273 heikki.linnakangas 751 ECB : * If strategy is not NULL, a nondefault buffer access strategy is used.
752 : * See buffer/README for details.
753 : */
754 : Buffer
5273 heikki.linnakangas 755 GIC 67680696 : ReadBufferExtended(Relation reln, ForkNumber forkNum, BlockNumber blockNum,
756 : ReadBufferMode mode, BufferAccessStrategy strategy)
757 : {
758 : bool hit;
759 : Buffer buf;
760 :
761 : /*
5050 bruce 762 ECB : * Reject attempts to read non-local temporary relations; we would be
5050 bruce 763 EUB : * likely to get wrong data since we have no visibility into the owning
764 : * session's local buffers.
765 : */
5122 tgl 766 GIC 67680696 : if (RELATION_IS_OTHER_TEMP(reln))
5122 tgl 767 UIC 0 : ereport(ERROR,
768 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
769 : errmsg("cannot access temporary tables of other sessions")));
770 :
5414 heikki.linnakangas 771 ECB : /*
5050 bruce 772 : * Read the buffer, and update pgstat counters to reflect a cache hit or
773 : * miss.
5414 heikki.linnakangas 774 : */
5414 heikki.linnakangas 775 CBC 67680696 : pgstat_count_buffer_read(reln);
636 tgl 776 67680696 : buf = ReadBuffer_common(RelationGetSmgr(reln), reln->rd_rel->relpersistence,
777 : forkNum, blockNum, mode, strategy, &hit);
5414 heikki.linnakangas 778 GIC 67680683 : if (hit)
779 66529542 : pgstat_count_buffer_hit(reln);
780 67680683 : return buf;
781 : }
782 :
783 :
784 : /*
785 : * ReadBufferWithoutRelcache -- like ReadBufferExtended, but doesn't require
786 : * a relcache entry for the relation.
787 : *
788 : * Pass permanent = true for a RELPERSISTENCE_PERMANENT relation, and
789 : * permanent = false for a RELPERSISTENCE_UNLOGGED relation. This function
790 : * cannot be used for temporary relations (and making that work might be
376 rhaas 791 ECB : * difficult, unless we only want to read temporary relations for our own
792 : * BackendId).
793 : */
794 : Buffer
277 rhaas 795 GNC 2815054 : ReadBufferWithoutRelcache(RelFileLocator rlocator, ForkNumber forkNum,
796 : BlockNumber blockNum, ReadBufferMode mode,
376 rhaas 797 ECB : BufferAccessStrategy strategy, bool permanent)
798 : {
5050 bruce 799 : bool hit;
800 :
277 rhaas 801 GNC 2815054 : SMgrRelation smgr = smgropen(rlocator, InvalidBackendId);
802 :
376 rhaas 803 GIC 2815054 : return ReadBuffer_common(smgr, permanent ? RELPERSISTENCE_PERMANENT :
804 : RELPERSISTENCE_UNLOGGED, forkNum, blockNum,
805 : mode, strategy, &hit);
806 : }
807 :
808 : /*
809 : * Convenience wrapper around ExtendBufferedRelBy() extending by one block.
810 : */
811 : Buffer
4 andres 812 GNC 62012 : ExtendBufferedRel(ExtendBufferedWhat eb,
813 : ForkNumber forkNum,
814 : BufferAccessStrategy strategy,
815 : uint32 flags)
816 : {
817 : Buffer buf;
818 62012 : uint32 extend_by = 1;
819 :
820 62012 : ExtendBufferedRelBy(eb, forkNum, strategy, flags, extend_by,
821 : &buf, &extend_by);
822 :
823 62012 : return buf;
824 : }
825 :
826 : /*
827 : * Extend relation by multiple blocks.
828 : *
829 : * Tries to extend the relation by extend_by blocks. Depending on the
830 : * availability of resources the relation may end up being extended by a
831 : * smaller number of pages (unless an error is thrown, always by at least one
832 : * page). *extended_by is updated to the number of pages the relation has been
833 : * extended to.
834 : *
835 : * buffers needs to be an array that is at least extend_by long. Upon
836 : * completion, the first extend_by array elements will point to a pinned
837 : * buffer.
838 : *
839 : * If EB_LOCK_FIRST is part of flags, the first returned buffer is
840 : * locked. This is useful for callers that want a buffer that is guaranteed to
841 : * be empty.
842 : */
843 : BlockNumber
844 277364 : ExtendBufferedRelBy(ExtendBufferedWhat eb,
845 : ForkNumber fork,
846 : BufferAccessStrategy strategy,
847 : uint32 flags,
848 : uint32 extend_by,
849 : Buffer *buffers,
850 : uint32 *extended_by)
851 : {
852 277364 : Assert((eb.rel != NULL) != (eb.smgr != NULL));
853 277364 : Assert(eb.smgr == NULL || eb.relpersistence != 0);
854 277364 : Assert(extend_by > 0);
855 :
856 277364 : if (eb.smgr == NULL)
857 : {
858 277177 : eb.smgr = RelationGetSmgr(eb.rel);
859 277177 : eb.relpersistence = eb.rel->rd_rel->relpersistence;
860 : }
861 :
862 277364 : return ExtendBufferedRelCommon(eb, fork, strategy, flags,
863 : extend_by, InvalidBlockNumber,
864 : buffers, extended_by);
865 : }
866 :
867 : /*
868 : * Extend the relation so it is at least extend_to blocks large, return buffer
869 : * (extend_to - 1).
870 : *
871 : * This is useful for callers that want to write a specific page, regardless
872 : * of the current size of the relation (e.g. useful for visibilitymap and for
873 : * crash recovery).
874 : */
875 : Buffer
876 64243 : ExtendBufferedRelTo(ExtendBufferedWhat eb,
877 : ForkNumber fork,
878 : BufferAccessStrategy strategy,
879 : uint32 flags,
880 : BlockNumber extend_to,
881 : ReadBufferMode mode)
882 : {
883 : BlockNumber current_size;
884 64243 : uint32 extended_by = 0;
885 64243 : Buffer buffer = InvalidBuffer;
886 : Buffer buffers[64];
887 :
888 64243 : Assert((eb.rel != NULL) != (eb.smgr != NULL));
889 64243 : Assert(eb.smgr == NULL || eb.relpersistence != 0);
890 64243 : Assert(extend_to != InvalidBlockNumber && extend_to > 0);
891 64243 : Assert(mode == RBM_NORMAL || mode == RBM_ZERO_ON_ERROR ||
892 : mode == RBM_ZERO_AND_LOCK);
893 :
894 64243 : if (eb.smgr == NULL)
895 : {
896 27870 : eb.smgr = RelationGetSmgr(eb.rel);
897 27870 : eb.relpersistence = eb.rel->rd_rel->relpersistence;
898 : }
899 :
900 : /*
901 : * If desired, create the file if it doesn't exist. If
902 : * smgr_cached_nblocks[fork] is positive then it must exist, no need for
903 : * an smgrexists call.
904 : */
905 64243 : if ((flags & EB_CREATE_FORK_IF_NEEDED) &&
906 27870 : (eb.smgr->smgr_cached_nblocks[fork] == 0 ||
907 10 : eb.smgr->smgr_cached_nblocks[fork] == InvalidBlockNumber) &&
908 27860 : !smgrexists(eb.smgr, fork))
909 : {
910 27855 : LockRelationForExtension(eb.rel, ExclusiveLock);
911 :
912 : /* could have been closed while waiting for lock */
913 27855 : if (eb.rel)
914 27855 : eb.smgr = RelationGetSmgr(eb.rel);
915 :
916 : /* recheck, fork might have been created concurrently */
917 27855 : if (!smgrexists(eb.smgr, fork))
918 27854 : smgrcreate(eb.smgr, fork, flags & EB_PERFORMING_RECOVERY);
919 :
920 27855 : UnlockRelationForExtension(eb.rel, ExclusiveLock);
921 : }
922 :
923 : /*
924 : * If requested, invalidate size cache, so that smgrnblocks asks the
925 : * kernel.
926 : */
927 64243 : if (flags & EB_CLEAR_SIZE_CACHE)
928 27870 : eb.smgr->smgr_cached_nblocks[fork] = InvalidBlockNumber;
929 :
930 : /*
931 : * Estimate how many pages we'll need to extend by. This avoids acquiring
932 : * unnecessarily many victim buffers.
933 : */
934 64243 : current_size = smgrnblocks(eb.smgr, fork);
935 :
936 64243 : if (mode == RBM_ZERO_AND_LOCK)
937 36056 : flags |= EB_LOCK_TARGET;
938 :
939 130664 : while (current_size < extend_to)
940 : {
941 66421 : uint32 num_pages = lengthof(buffers);
942 : BlockNumber first_block;
943 :
944 66421 : if ((uint64) current_size + num_pages > extend_to)
945 66355 : num_pages = extend_to - current_size;
946 :
947 66421 : first_block = ExtendBufferedRelCommon(eb, fork, strategy, flags,
948 : num_pages, extend_to,
949 : buffers, &extended_by);
950 :
951 66421 : current_size = first_block + extended_by;
952 66421 : Assert(current_size <= extend_to);
953 66421 : Assert(num_pages != 0 || current_size >= extend_to);
954 :
955 160982 : for (int i = 0; i < extended_by; i++)
956 : {
957 94561 : if (first_block + i != extend_to - 1)
958 30319 : ReleaseBuffer(buffers[i]);
959 : else
960 64242 : buffer = buffers[i];
961 : }
962 : }
963 :
964 : /*
965 : * It's possible that another backend concurrently extended the relation.
966 : * In that case read the buffer.
967 : *
968 : * XXX: Should we control this via a flag?
969 : */
970 64243 : if (buffer == InvalidBuffer)
971 : {
972 : bool hit;
973 :
974 1 : Assert(extended_by == 0);
975 1 : buffer = ReadBuffer_common(eb.smgr, eb.relpersistence,
976 : fork, extend_to - 1, mode, strategy,
977 : &hit);
978 : }
979 :
980 64243 : return buffer;
981 : }
5273 heikki.linnakangas 982 ECB :
983 : /*
984 : * ReadBuffer_common -- common logic for all ReadBuffer variants
985 : *
986 : * *hit is set to true if the request was satisfied from shared buffer cache.
987 : */
5414 988 : static Buffer
4484 rhaas 989 GIC 70495751 : ReadBuffer_common(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
5273 heikki.linnakangas 990 ECB : BlockNumber blockNum, ReadBufferMode mode,
991 : BufferAccessStrategy strategy, bool *hit)
992 : {
2701 rhaas 993 : BufferDesc *bufHdr;
994 : Block bufBlock;
995 : bool found;
996 : IOContext io_context;
997 : IOObject io_object;
4615 rhaas 998 GIC 70495751 : bool isLocalBuf = SmgrIsTemp(smgr);
999 :
5414 heikki.linnakangas 1000 70495751 : *hit = false;
1001 :
1002 : /*
1003 : * Backward compatibility path, most code should use ExtendBufferedRel()
1004 : * instead, as acquiring the extension lock inside ExtendBufferedRel()
1005 : * scales a lot better.
1006 : */
4 andres 1007 GNC 70495751 : if (unlikely(blockNum == P_NEW))
1008 : {
1009 187 : uint32 flags = EB_SKIP_EXTENSION_LOCK;
1010 :
1011 187 : Assert(mode == RBM_NORMAL ||
1012 : mode == RBM_ZERO_AND_LOCK ||
1013 : mode == RBM_ZERO_ON_ERROR);
1014 :
1015 187 : if (mode == RBM_ZERO_AND_LOCK)
4 andres 1016 UNC 0 : flags |= EB_LOCK_FIRST;
1017 :
4 andres 1018 GNC 187 : return ExtendBufferedRel(EB_SMGR(smgr, relpersistence),
1019 : forkNum, strategy, flags);
1020 : }
1021 :
1022 : /* Make sure we will have room to remember the buffer pin */
6610 tgl 1023 GIC 70495564 : ResourceOwnerEnlargeBuffers(CurrentResourceOwner);
1024 :
1025 : TRACE_POSTGRESQL_BUFFER_READ_START(forkNum, blockNum,
1026 : smgr->smgr_rlocator.locator.spcOid,
1027 : smgr->smgr_rlocator.locator.dbOid,
1028 : smgr->smgr_rlocator.locator.relNumber,
1029 : smgr->smgr_rlocator.backend);
1030 :
9345 bruce 1031 CBC 70495564 : if (isLocalBuf)
1032 : {
1033 : /*
1034 : * We do not use a BufferAccessStrategy for I/O of temporary tables.
1035 : * However, in some cases, the "strategy" may not be NULL, so we can't
1036 : * rely on IOContextForStrategy() to set the right IOContext for us.
1037 : * This may happen in cases like CREATE TEMPORARY TABLE AS...
1038 : */
10 andres 1039 GNC 1247393 : io_context = IOCONTEXT_NORMAL;
1040 1247393 : io_object = IOOBJECT_TEMP_RELATION;
10 andres 1041 CBC 1247393 : bufHdr = LocalBufferAlloc(smgr, forkNum, blockNum, &found);
9345 bruce 1042 1247393 : if (found)
4863 rhaas 1043 GIC 1243603 : pgBufferUsage.local_blks_hit++;
1593 tmunro 1044 3790 : else if (mode == RBM_NORMAL || mode == RBM_NORMAL_NO_LOG ||
1045 : mode == RBM_ZERO_ON_ERROR)
4863 rhaas 1046 3790 : pgBufferUsage.local_blks_read++;
1047 : }
1048 : else
1049 : {
1050 : /*
1051 : * lookup the buffer. IO_IN_PROGRESS is set if the requested block is
1052 : * not currently in memory.
1053 : */
10 andres 1054 GNC 69248171 : io_context = IOContextForStrategy(strategy);
1055 69248171 : io_object = IOOBJECT_RELATION;
4484 rhaas 1056 GIC 69248171 : bufHdr = BufferAlloc(smgr, relpersistence, forkNum, blockNum,
1057 : strategy, &found, io_context);
9345 bruce 1058 69248171 : if (found)
4863 rhaas 1059 CBC 67748583 : pgBufferUsage.shared_blks_hit++;
1593 tmunro 1060 GIC 1499588 : else if (mode == RBM_NORMAL || mode == RBM_NORMAL_NO_LOG ||
1061 : mode == RBM_ZERO_ON_ERROR)
4863 rhaas 1062 1306022 : pgBufferUsage.shared_blks_read++;
1063 : }
1064 :
6610 tgl 1065 ECB : /* At this point we do NOT hold any locks. */
8002 1066 :
1067 : /* if it was already in the buffer pool, we're done */
9345 bruce 1068 GIC 70495564 : if (found)
9345 bruce 1069 ECB : {
1070 : /* Just need to update stats before we exit */
4 andres 1071 GNC 68992186 : *hit = true;
1072 68992186 : VacuumPageHit++;
1073 68992186 : pgstat_count_io_op(io_object, io_context, IOOP_HIT);
3069 heikki.linnakangas 1074 ECB :
4 andres 1075 GNC 68992186 : if (VacuumCostActive)
1076 53710 : VacuumCostBalance += VacuumCostPageHit;
6302 tgl 1077 ECB :
1078 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forkNum, blockNum,
1079 : smgr->smgr_rlocator.locator.spcOid,
1080 : smgr->smgr_rlocator.locator.dbOid,
1081 : smgr->smgr_rlocator.locator.relNumber,
1082 : smgr->smgr_rlocator.backend,
1083 : found);
1084 :
1085 : /*
1086 : * In RBM_ZERO_AND_LOCK mode the caller expects the page to be locked
1087 : * on return.
1088 : */
4 andres 1089 GNC 68992186 : if (!isLocalBuf)
1090 : {
1091 67748583 : if (mode == RBM_ZERO_AND_LOCK)
1092 30153 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr),
1093 : LW_EXCLUSIVE);
1094 67718430 : else if (mode == RBM_ZERO_AND_CLEANUP_LOCK)
1095 19 : LockBufferForCleanup(BufferDescriptorGetBuffer(bufHdr));
1096 : }
1097 :
1098 68992186 : return BufferDescriptorGetBuffer(bufHdr);
1099 : }
9345 bruce 1100 ECB :
9770 scrappy 1101 : /*
1102 : * if we have gotten to this point, we have allocated a buffer for the
1103 : * page but its contents are not yet valid. IO_IN_PROGRESS is set for it,
1104 : * if it's a shared buffer.
1105 : */
2555 andres 1106 GIC 1503378 : Assert(!(pg_atomic_read_u32(&bufHdr->state) & BM_VALID)); /* spinlock not needed */
1107 :
10 andres 1108 CBC 1503378 : bufBlock = isLocalBuf ? LocalBufHdrGetBlock(bufHdr) : BufHdrGetBlock(bufHdr);
1109 :
1110 : /*
1111 : * Read in the page, unless the caller intends to overwrite it and just
1112 : * wants us to allocate a buffer.
1113 : */
4 andres 1114 GNC 1503378 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
6610 tgl 1115 GIC 193566 : MemSet((char *) bufBlock, 0, BLCKSZ);
1116 : else
1117 : {
2 andres 1118 GNC 1309812 : instr_time io_start = pgstat_prepare_io_time();
1119 :
4 1120 1309812 : smgrread(smgr, forkNum, blockNum, bufBlock);
1121 :
2 1122 1309799 : pgstat_count_io_op_time(io_object, io_context,
1123 : IOOP_READ, io_start, 1);
1124 :
1125 : /* check for garbage data */
4 1126 1309799 : if (!PageIsVerifiedExtended((Page) bufBlock, blockNum,
1127 : PIV_LOG_WARNING | PIV_REPORT_STAT))
1128 : {
4 andres 1129 UNC 0 : if (mode == RBM_ZERO_ON_ERROR || zero_damaged_pages)
1130 : {
1131 0 : ereport(WARNING,
1132 : (errcode(ERRCODE_DATA_CORRUPTED),
1133 : errmsg("invalid page in block %u of relation %s; zeroing out page",
1134 : blockNum,
1135 : relpath(smgr->smgr_rlocator, forkNum))));
1136 0 : MemSet((char *) bufBlock, 0, BLCKSZ);
1137 : }
1138 : else
1139 0 : ereport(ERROR,
1140 : (errcode(ERRCODE_DATA_CORRUPTED),
1141 : errmsg("invalid page in block %u of relation %s",
1142 : blockNum,
1143 : relpath(smgr->smgr_rlocator, forkNum))));
1144 : }
1145 : }
1146 :
3069 heikki.linnakangas 1147 ECB : /*
1148 : * In RBM_ZERO_AND_LOCK mode, grab the buffer content lock before marking
1149 : * the page as valid, to make sure that no other backend sees the zeroed
1150 : * page before the caller has had a chance to initialize it.
1151 : *
1152 : * Since no-one else can be looking at the page contents yet, there is no
2878 bruce 1153 : * difference between an exclusive lock and a cleanup-strength lock. (Note
1154 : * that we cannot use LockBuffer() or LockBufferForCleanup() here, because
1155 : * they assert that the buffer is already valid.)
1156 : */
3069 heikki.linnakangas 1157 GIC 1503365 : if ((mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK) &&
1158 193566 : !isLocalBuf)
1159 : {
2672 rhaas 1160 193566 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_EXCLUSIVE);
3069 heikki.linnakangas 1161 ECB : }
1162 :
9345 bruce 1163 GIC 1503365 : if (isLocalBuf)
8004 tgl 1164 ECB : {
6927 1165 : /* Only need to adjust flags */
2555 andres 1166 CBC 3790 : uint32 buf_state = pg_atomic_read_u32(&bufHdr->state);
1167 :
1168 3790 : buf_state |= BM_VALID;
2375 1169 3790 : pg_atomic_unlocked_write_u32(&bufHdr->state, buf_state);
1170 : }
1171 : else
1172 : {
1173 : /* Set BM_VALID, terminate IO, and wake up any waiters */
6610 tgl 1174 GIC 1499575 : TerminateBufferIO(bufHdr, false, BM_VALID);
1175 : }
1176 :
4153 alvherre 1177 1503365 : VacuumPageMiss++;
6927 tgl 1178 1503365 : if (VacuumCostActive)
1179 616 : VacuumCostBalance += VacuumCostPageMiss;
1180 :
1181 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forkNum, blockNum,
1182 : smgr->smgr_rlocator.locator.spcOid,
1183 : smgr->smgr_rlocator.locator.dbOid,
1184 : smgr->smgr_rlocator.locator.relNumber,
1185 : smgr->smgr_rlocator.backend,
5142 tgl 1186 ECB : found);
5364 alvherre 1187 :
8986 bruce 1188 GIC 1503365 : return BufferDescriptorGetBuffer(bufHdr);
1189 : }
9770 scrappy 1190 ECB :
1191 : /*
1192 : * BufferAlloc -- subroutine for ReadBuffer. Handles lookup of a shared
1193 : * buffer. If no buffer exists already, selects a replacement
1194 : * victim and evicts the old page, but does NOT read in new page.
1195 : *
1196 : * "strategy" can be a buffer replacement strategy object, or NULL for
1197 : * the default strategy. The selected buffer's usage_count is advanced when
5793 tgl 1198 : * using the default strategy, but otherwise possibly not (see PinBuffer).
1199 : *
6927 1200 : * The returned buffer is pinned and is already marked as holding the
1201 : * desired page. If it already did have the desired page, *foundPtr is
1202 : * set true. Otherwise, *foundPtr is set false and the buffer is marked
1203 : * as IO_IN_PROGRESS; ReadBuffer will now need to do I/O to fill it.
1204 : *
1205 : * *foundPtr is actually redundant with the buffer's BM_VALID flag, but
1206 : * we keep it for simplicity in ReadBuffer.
9770 scrappy 1207 : *
1208 : * io_context is passed as an output parameter to avoid calling
1209 : * IOContextForStrategy() when there is a shared buffers hit and no IO
1210 : * statistics need be captured.
1211 : *
1212 : * No locks are held either at entry or exit.
1213 : */
2701 rhaas 1214 : static BufferDesc *
4484 rhaas 1215 GIC 69248171 : BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
9345 bruce 1216 ECB : BlockNumber blockNum,
1217 : BufferAccessStrategy strategy,
1218 : bool *foundPtr, IOContext io_context)
1219 : {
1220 : BufferTag newTag; /* identity of requested block */
1221 : uint32 newHash; /* hash value for newTag */
2118 tgl 1222 : LWLock *newPartitionLock; /* buffer partition lock for it */
1223 : int existing_buf_id;
1224 : Buffer victim_buffer;
1225 : BufferDesc *victim_buf_hdr;
1226 : uint32 victim_buf_state;
1227 :
6929 tgl 1228 EUB : /* create a tag so we can lookup the buffer */
256 rhaas 1229 GNC 69248171 : InitBufferTag(&newTag, &smgr->smgr_rlocator.locator, forkNum, blockNum);
1230 :
6104 tgl 1231 EUB : /* determine its hash code and partition lock ID */
6104 tgl 1232 GIC 69248171 : newHash = BufTableHashCode(&newTag);
1233 69248171 : newPartitionLock = BufMappingPartitionLock(newHash);
1234 :
1235 : /* see if the block is in the buffer pool already */
1236 69248171 : LWLockAcquire(newPartitionLock, LW_SHARED);
4 andres 1237 GNC 69248171 : existing_buf_id = BufTableLookup(&newTag, newHash);
1238 69248171 : if (existing_buf_id >= 0)
1239 : {
1240 : BufferDesc *buf;
1241 : bool valid;
1242 :
1243 : /*
1244 : * Found it. Now, pin the buffer so no one can steal it from the
1245 : * buffer pool, and check to see if the correct data has been loaded
1246 : * into the buffer.
1247 : */
1248 67747784 : buf = GetBufferDescriptor(existing_buf_id);
1249 :
5793 tgl 1250 GIC 67747784 : valid = PinBuffer(buf, strategy);
1251 :
6610 tgl 1252 ECB : /* Can release the mapping lock as soon as we've pinned it */
6104 tgl 1253 CBC 67747784 : LWLockRelease(newPartitionLock);
1254 :
2062 peter_e 1255 67747784 : *foundPtr = true;
1256 :
6610 tgl 1257 GIC 67747784 : if (!valid)
8483 inoue 1258 ECB : {
1259 : /*
1260 : * We can only get here if (a) someone else is still reading in
6385 bruce 1261 : * the page, or (b) a previous read attempt failed. We have to
1262 : * wait for any active read attempt to finish, and then set up our
1263 : * own read attempt if the page is still not BM_VALID.
1264 : * StartBufferIO does it all.
1265 : */
6610 tgl 1266 GIC 1269 : if (StartBufferIO(buf, true))
1267 : {
1268 : /*
6385 bruce 1269 ECB : * If we get here, previous attempts to read the buffer must
1270 : * have failed ... but we shall bravely try again.
1271 : */
2062 peter_e 1272 CBC 19 : *foundPtr = false;
6927 tgl 1273 ECB : }
9345 bruce 1274 : }
1275 :
8986 bruce 1276 GIC 67747784 : return buf;
1277 : }
1278 :
1279 : /*
1280 : * Didn't find it in the buffer pool. We'll have to initialize a new
1281 : * buffer. Remember to unlock the mapping lock while doing the work.
1282 : */
6104 tgl 1283 CBC 1500387 : LWLockRelease(newPartitionLock);
1284 :
1285 : /*
1286 : * Acquire a victim buffer. Somebody else might try to do the same, we
1287 : * don't hold any conflicting locks. If so we'll have to undo our work
1288 : * later.
1289 : */
4 andres 1290 GNC 1500387 : victim_buffer = GetVictimBuffer(strategy, io_context);
1291 1500387 : victim_buf_hdr = GetBufferDescriptor(victim_buffer - 1);
1292 :
1293 : /*
1294 : * Try to make a hashtable entry for the buffer under its new tag. If
1295 : * somebody else inserted another buffer for the tag, we'll release the
1296 : * victim buffer we acquired and use the already inserted one.
1297 : */
1298 1500387 : LWLockAcquire(newPartitionLock, LW_EXCLUSIVE);
1299 1500387 : existing_buf_id = BufTableInsert(&newTag, newHash, victim_buf_hdr->buf_id);
1300 1500387 : if (existing_buf_id >= 0)
1301 : {
1302 : BufferDesc *existing_buf_hdr;
1303 : bool valid;
1304 :
1305 : /*
1306 : * Got a collision. Someone has already done what we were about to do.
1307 : * We'll just handle this as if it were found in the buffer pool in
1308 : * the first place. First, give up the buffer we were planning to
1309 : * use.
1310 : *
1311 : * We could do this after releasing the partition lock, but then we'd
1312 : * have to call ResourceOwnerEnlargeBuffers() &
1313 : * ReservePrivateRefCountEntry() before acquiring the lock, for the
1314 : * rare case of such a collision.
1315 : */
1316 809 : UnpinBuffer(victim_buf_hdr);
1317 :
1318 : /*
1319 : * The victim buffer we acquired peviously is clean and unused, let it
1320 : * be found again quickly
1321 : */
1322 809 : StrategyFreeBuffer(victim_buf_hdr);
5793 tgl 1323 ECB :
1324 : /* remaining code should match code at top of routine */
5226 bruce 1325 :
4 andres 1326 GNC 809 : existing_buf_hdr = GetBufferDescriptor(existing_buf_id);
1327 :
1328 809 : valid = PinBuffer(existing_buf_hdr, strategy);
1329 :
1330 : /* Can release the mapping lock as soon as we've pinned it */
4 andres 1331 GIC 809 : LWLockRelease(newPartitionLock);
1332 :
4 andres 1333 GNC 809 : *foundPtr = true;
1334 :
1335 809 : if (!valid)
1336 : {
1337 : /*
1338 : * We can only get here if (a) someone else is still reading in
1339 : * the page, or (b) a previous read attempt failed. We have to
1340 : * wait for any active read attempt to finish, and then set up our
1341 : * own read attempt if the page is still not BM_VALID.
1342 : * StartBufferIO does it all.
1343 : */
1344 509 : if (StartBufferIO(existing_buf_hdr, true))
1345 : {
1346 : /*
1347 : * If we get here, previous attempts to read the buffer must
1348 : * have failed ... but we shall bravely try again.
1349 : */
1350 3 : *foundPtr = false;
1351 : }
1352 : }
1353 :
1354 809 : return existing_buf_hdr;
1355 : }
1356 :
1357 : /*
1358 : * Need to lock the buffer header too in order to change its tag.
1359 : */
1360 1499578 : victim_buf_state = LockBufHdr(victim_buf_hdr);
1361 :
1362 : /* some sanity checks while we hold the buffer header lock */
1363 1499578 : Assert(BUF_STATE_GET_REFCOUNT(victim_buf_state) == 1);
1364 1499578 : Assert(!(victim_buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY | BM_IO_IN_PROGRESS)));
1365 :
1366 1499578 : victim_buf_hdr->tag = newTag;
1367 :
1368 : /*
1369 : * Make sure BM_PERMANENT is set for buffers that must be written at every
1370 : * checkpoint. Unlogged buffers only need to be written at shutdown
1371 : * checkpoints, except for their "init" forks, which need to be treated
1372 : * just like permanent relations.
1373 : */
1374 1499578 : victim_buf_state |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2217 rhaas 1375 CBC 1499578 : if (relpersistence == RELPERSISTENCE_PERMANENT || forkNum == INIT_FORKNUM)
4 andres 1376 GNC 1499550 : victim_buf_state |= BM_PERMANENT;
1377 :
1378 1499578 : UnlockBufHdr(victim_buf_hdr, victim_buf_state);
1379 :
6104 tgl 1380 CBC 1499578 : LWLockRelease(newPartitionLock);
7947 tgl 1381 ECB :
9345 bruce 1382 : /*
1383 : * Buffer contents are currently invalid. Try to obtain the right to
1384 : * start I/O. If StartBufferIO returns false, then someone else managed
1385 : * to read it before we did, so there's nothing left for BufferAlloc() to
1386 : * do.
1387 : */
4 andres 1388 GNC 1499578 : if (StartBufferIO(victim_buf_hdr, true))
2062 peter_e 1389 GIC 1499566 : *foundPtr = false;
1390 : else
1391 12 : *foundPtr = true;
1392 :
4 andres 1393 GNC 1499578 : return victim_buf_hdr;
9770 scrappy 1394 ECB : }
1395 :
6610 tgl 1396 : /*
1397 : * InvalidateBuffer -- mark a shared buffer invalid and return it to the
1398 : * freelist.
1399 : *
6610 tgl 1400 EUB : * The buffer header spinlock must be held at entry. We drop it before
6610 tgl 1401 ECB : * returning. (This is sane because the caller must have locked the
1402 : * buffer in order to be sure it should be dropped.)
1403 : *
1404 : * This is used only in contexts such as dropping a relation. We assume
1405 : * that no other backend could possibly be interested in using the page,
1406 : * so the only reason the buffer might be pinned is if someone else is
1407 : * trying to write it out. We have to let them finish before we can
1408 : * reclaim the buffer.
1409 : *
1410 : * The buffer could get reclaimed by someone else while we are waiting
1411 : * to acquire the necessary locks; if so, don't mess it up.
1412 : */
1413 : static void
2701 rhaas 1414 GIC 83771 : InvalidateBuffer(BufferDesc *buf)
1415 : {
1416 : BufferTag oldTag;
6104 tgl 1417 ECB : uint32 oldHash; /* hash value for oldTag */
2118 1418 : LWLock *oldPartitionLock; /* buffer partition lock for it */
1419 : uint32 oldFlags;
1420 : uint32 buf_state;
1421 :
1422 : /* Save the original buffer tag before dropping the spinlock */
6610 tgl 1423 CBC 83771 : oldTag = buf->tag;
1424 :
2555 andres 1425 GIC 83771 : buf_state = pg_atomic_read_u32(&buf->state);
1426 83771 : Assert(buf_state & BM_LOCKED);
1427 83771 : UnlockBufHdr(buf, buf_state);
6610 tgl 1428 ECB :
1429 : /*
1430 : * Need to compute the old tag's hashcode and partition lock ID. XXX is it
1431 : * worth storing the hashcode in BufferDesc so we need not recompute it
1432 : * here? Probably not.
1433 : */
6104 tgl 1434 GIC 83771 : oldHash = BufTableHashCode(&oldTag);
1435 83771 : oldPartitionLock = BufMappingPartitionLock(oldHash);
1436 :
6610 1437 83773 : retry:
1438 :
1439 : /*
1440 : * Acquire exclusive mapping lock in preparation for changing the buffer's
6385 bruce 1441 ECB : * association.
1442 : */
6104 tgl 1443 GIC 83773 : LWLockAcquire(oldPartitionLock, LW_EXCLUSIVE);
1444 :
1445 : /* Re-lock the buffer header */
2555 andres 1446 83773 : buf_state = LockBufHdr(buf);
1447 :
6610 tgl 1448 ECB : /* If it's changed while we were waiting for lock, do nothing */
256 rhaas 1449 GNC 83773 : if (!BufferTagsEqual(&buf->tag, &oldTag))
1450 : {
2555 andres 1451 CBC 2 : UnlockBufHdr(buf, buf_state);
6104 tgl 1452 GIC 2 : LWLockRelease(oldPartitionLock);
6610 tgl 1453 CBC 2 : return;
6610 tgl 1454 ECB : }
1455 :
1456 : /*
1457 : * We assume the only reason for it to be pinned is that someone else is
1458 : * flushing the page out. Wait for them to finish. (This could be an
6385 bruce 1459 : * infinite loop if the refcount is messed up... it would be nice to time
1460 : * out after awhile, but there seems no way to be sure how many loops may
1461 : * be needed. Note that if the other guy has pinned the buffer but not
1462 : * yet done StartBufferIO, WaitIO will fall through and we'll effectively
1463 : * be busy-looping here.)
1464 : */
2555 andres 1465 CBC 83771 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 0)
6610 tgl 1466 ECB : {
2555 andres 1467 CBC 2 : UnlockBufHdr(buf, buf_state);
6104 tgl 1468 GIC 2 : LWLockRelease(oldPartitionLock);
1469 : /* safety check: should definitely not be our *own* pin */
2797 andres 1470 2 : if (GetPrivateRefCount(BufferDescriptorGetBuffer(buf)) > 0)
6596 tgl 1471 UIC 0 : elog(ERROR, "buffer is pinned in InvalidateBuffer");
6610 tgl 1472 GIC 2 : WaitIO(buf);
6610 tgl 1473 CBC 2 : goto retry;
1474 : }
6610 tgl 1475 ECB :
1476 : /*
6385 bruce 1477 : * Clear out the buffer's tag and flags. We must do this to ensure that
1478 : * linear scans of the buffer array don't think the buffer is valid.
1479 : */
2555 andres 1480 CBC 83769 : oldFlags = buf_state & BUF_FLAG_MASK;
256 rhaas 1481 GNC 83769 : ClearBufferTag(&buf->tag);
2555 andres 1482 GIC 83769 : buf_state &= ~(BUF_FLAG_MASK | BUF_USAGECOUNT_MASK);
1483 83769 : UnlockBufHdr(buf, buf_state);
1484 :
1485 : /*
1486 : * Remove the buffer from the lookup hashtable, if it was in there.
1487 : */
6610 tgl 1488 83769 : if (oldFlags & BM_TAG_VALID)
6104 1489 83769 : BufTableDelete(&oldTag, oldHash);
6610 tgl 1490 ECB :
1491 : /*
6013 1492 : * Done with mapping lock.
1493 : */
6104 tgl 1494 CBC 83769 : LWLockRelease(oldPartitionLock);
1495 :
1496 : /*
6610 tgl 1497 ECB : * Insert the buffer at the head of the list of free buffers.
1498 : */
5793 tgl 1499 CBC 83769 : StrategyFreeBuffer(buf);
1500 : }
6610 tgl 1501 ECB :
1502 : /*
1503 : * Helper routine for GetVictimBuffer()
1504 : *
1505 : * Needs to be called on a buffer with a valid tag, pinned, but without the
1506 : * buffer header spinlock held.
1507 : *
1508 : * Returns true if the buffer can be reused, in which case the buffer is only
1509 : * pinned by this backend and marked as invalid, false otherwise.
1510 : */
1511 : static bool
4 andres 1512 GNC 1052237 : InvalidateVictimBuffer(BufferDesc *buf_hdr)
1513 : {
1514 : uint32 buf_state;
1515 : uint32 hash;
1516 : LWLock *partition_lock;
1517 : BufferTag tag;
1518 :
1519 1052237 : Assert(GetPrivateRefCount(BufferDescriptorGetBuffer(buf_hdr)) == 1);
1520 :
1521 : /* have buffer pinned, so it's safe to read tag without lock */
1522 1052237 : tag = buf_hdr->tag;
1523 :
1524 1052237 : hash = BufTableHashCode(&tag);
1525 1052237 : partition_lock = BufMappingPartitionLock(hash);
1526 :
1527 1052237 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
1528 :
1529 : /* lock the buffer header */
1530 1052237 : buf_state = LockBufHdr(buf_hdr);
1531 :
1532 : /*
1533 : * We have the buffer pinned nobody else should have been able to unset
1534 : * this concurrently.
1535 : */
1536 1052237 : Assert(buf_state & BM_TAG_VALID);
1537 1052237 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
1538 1052237 : Assert(BufferTagsEqual(&buf_hdr->tag, &tag));
1539 :
1540 : /*
1541 : * If somebody else pinned the buffer since, or even worse, dirtied it,
1542 : * give up on this buffer: It's clearly in use.
1543 : */
1544 1052237 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 1 || (buf_state & BM_DIRTY))
1545 : {
1546 308 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
1547 :
1548 308 : UnlockBufHdr(buf_hdr, buf_state);
1549 308 : LWLockRelease(partition_lock);
1550 :
1551 308 : return false;
1552 : }
1553 :
1554 : /*
1555 : * Clear out the buffer's tag and flags and usagecount. This is not
1556 : * strictly required, as BM_TAG_VALID/BM_VALID needs to be checked before
1557 : * doing anything with the buffer. But currently it's beneficial, as the
1558 : * cheaper pre-check for several linear scans of shared buffers use the
1559 : * tag (see e.g. FlushDatabaseBuffers()).
1560 : */
1561 1051929 : ClearBufferTag(&buf_hdr->tag);
1562 1051929 : buf_state &= ~(BUF_FLAG_MASK | BUF_USAGECOUNT_MASK);
1563 1051929 : UnlockBufHdr(buf_hdr, buf_state);
1564 :
1565 1051929 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
1566 :
1567 : /* finally delete buffer from the buffer mapping table */
1568 1051929 : BufTableDelete(&tag, hash);
1569 :
1570 1051929 : LWLockRelease(partition_lock);
1571 :
1572 1051929 : Assert(!(buf_state & (BM_DIRTY | BM_VALID | BM_TAG_VALID)));
1573 1051929 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
1574 1051929 : Assert(BUF_STATE_GET_REFCOUNT(pg_atomic_read_u32(&buf_hdr->state)) > 0);
1575 :
1576 1051929 : return true;
1577 : }
1578 :
1579 : static Buffer
1580 1870075 : GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context)
1581 : {
1582 : BufferDesc *buf_hdr;
1583 : Buffer buf;
1584 : uint32 buf_state;
1585 : bool from_ring;
1586 :
1587 : /*
1588 : * Ensure, while the spinlock's not yet held, that there's a free refcount
1589 : * entry.
1590 : */
1591 1870075 : ReservePrivateRefCountEntry();
1592 1870075 : ResourceOwnerEnlargeBuffers(CurrentResourceOwner);
1593 :
1594 : /* we return here if a prospective victim buffer gets used concurrently */
1595 8428 : again:
1596 :
1597 : /*
1598 : * Select a victim buffer. The buffer is returned with its header
1599 : * spinlock still held!
1600 : */
1601 1878503 : buf_hdr = StrategyGetBuffer(strategy, &buf_state, &from_ring);
1602 1878503 : buf = BufferDescriptorGetBuffer(buf_hdr);
1603 :
1604 1878503 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 0);
1605 :
1606 : /* Pin the buffer and then release the buffer spinlock */
1607 1878503 : PinBuffer_Locked(buf_hdr);
1608 :
1609 : /*
1610 : * We shouldn't have any other pins for this buffer.
1611 : */
1612 1878503 : CheckBufferIsPinnedOnce(buf);
1613 :
1614 : /*
1615 : * If the buffer was dirty, try to write it out. There is a race
1616 : * condition here, in that someone might dirty it after we released the
1617 : * buffer header lock above, or even while we are writing it out (since
1618 : * our share-lock won't prevent hint-bit updates). We will recheck the
1619 : * dirty bit after re-locking the buffer header.
1620 : */
1621 1878503 : if (buf_state & BM_DIRTY)
1622 : {
1623 : LWLock *content_lock;
1624 :
1625 242096 : Assert(buf_state & BM_TAG_VALID);
1626 242096 : Assert(buf_state & BM_VALID);
1627 :
1628 : /*
1629 : * We need a share-lock on the buffer contents to write it out (else
1630 : * we might write invalid data, eg because someone else is compacting
1631 : * the page contents while we write). We must use a conditional lock
1632 : * acquisition here to avoid deadlock. Even though the buffer was not
1633 : * pinned (and therefore surely not locked) when StrategyGetBuffer
1634 : * returned it, someone else could have pinned and exclusive-locked it
1635 : * by the time we get here. If we try to get the lock unconditionally,
1636 : * we'd block waiting for them; if they later block waiting for us,
1637 : * deadlock ensues. (This has been observed to happen when two
1638 : * backends are both trying to split btree index pages, and the second
1639 : * one just happens to be trying to split the page the first one got
1640 : * from StrategyGetBuffer.)
1641 : */
1642 242096 : content_lock = BufferDescriptorGetContentLock(buf_hdr);
1643 242096 : if (!LWLockConditionalAcquire(content_lock, LW_SHARED))
1644 : {
1645 : /*
1646 : * Someone else has locked the buffer, so give it up and loop back
1647 : * to get another one.
1648 : */
4 andres 1649 UNC 0 : UnpinBuffer(buf_hdr);
1650 0 : goto again;
1651 : }
1652 :
1653 : /*
1654 : * If using a nondefault strategy, and writing the buffer would
1655 : * require a WAL flush, let the strategy decide whether to go ahead
1656 : * and write/reuse the buffer or to choose another victim. We need a
1657 : * lock to inspect the page LSN, so this can't be done inside
1658 : * StrategyGetBuffer.
1659 : */
4 andres 1660 GNC 242096 : if (strategy != NULL)
1661 : {
1662 : XLogRecPtr lsn;
1663 :
1664 : /* Read the LSN while holding buffer header lock */
1665 65674 : buf_state = LockBufHdr(buf_hdr);
1666 65674 : lsn = BufferGetLSN(buf_hdr);
1667 65674 : UnlockBufHdr(buf_hdr, buf_state);
1668 :
1669 65674 : if (XLogNeedsFlush(lsn)
1670 10275 : && StrategyRejectBuffer(strategy, buf_hdr, from_ring))
1671 : {
1672 8120 : LWLockRelease(content_lock);
1673 8120 : UnpinBuffer(buf_hdr);
1674 8120 : goto again;
1675 : }
1676 : }
1677 :
1678 : /* OK, do the I/O */
1679 233976 : FlushBuffer(buf_hdr, NULL, IOOBJECT_RELATION, io_context);
1680 233976 : LWLockRelease(content_lock);
1681 :
1682 233976 : ScheduleBufferTagForWriteback(&BackendWritebackContext,
1683 : &buf_hdr->tag);
1684 : }
1685 :
1686 :
1687 1870383 : if (buf_state & BM_VALID)
1688 : {
1689 : /*
1690 : * When a BufferAccessStrategy is in use, blocks evicted from shared
1691 : * buffers are counted as IOOP_EVICT in the corresponding context
1692 : * (e.g. IOCONTEXT_BULKWRITE). Shared buffers are evicted by a
1693 : * strategy in two cases: 1) while initially claiming buffers for the
1694 : * strategy ring 2) to replace an existing strategy ring buffer
1695 : * because it is pinned or in use and cannot be reused.
1696 : *
1697 : * Blocks evicted from buffers already in the strategy ring are
1698 : * counted as IOOP_REUSE in the corresponding strategy context.
1699 : *
1700 : * At this point, we can accurately count evictions and reuses,
1701 : * because we have successfully claimed the valid buffer. Previously,
1702 : * we may have been forced to release the buffer due to concurrent
1703 : * pinners or erroring out.
1704 : */
1705 1052237 : pgstat_count_io_op(IOOBJECT_RELATION, io_context,
1706 1052237 : from_ring ? IOOP_REUSE : IOOP_EVICT);
1707 : }
1708 :
1709 : /*
1710 : * If the buffer has an entry in the buffer mapping table, delete it. This
1711 : * can fail because another backend could have pinned or dirtied the
1712 : * buffer.
1713 : */
1714 1870383 : if ((buf_state & BM_TAG_VALID) && !InvalidateVictimBuffer(buf_hdr))
1715 : {
1716 308 : UnpinBuffer(buf_hdr);
1717 308 : goto again;
1718 : }
1719 :
1720 : /* a final set of sanity checks */
1721 : #ifdef USE_ASSERT_CHECKING
1722 1870075 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
1723 :
1724 1870075 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
1725 1870075 : Assert(!(buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY)));
1726 :
1727 1870075 : CheckBufferIsPinnedOnce(buf);
1728 : #endif
1729 :
1730 1870075 : return buf;
1731 : }
1732 :
1733 : /*
1734 : * Limit the number of pins a batch operation may additionally acquire, to
1735 : * avoid running out of pinnable buffers.
1736 : *
1737 : * One additional pin is always allowed, as otherwise the operation likely
1738 : * cannot be performed at all.
1739 : *
1740 : * The number of allowed pins for a backend is computed based on
1741 : * shared_buffers and the maximum number of connections possible. That's very
1742 : * pessimistic, but outside of toy-sized shared_buffers it should allow
1743 : * sufficient pins.
1744 : */
1745 : static void
1746 332890 : LimitAdditionalPins(uint32 *additional_pins)
1747 : {
1748 : uint32 max_backends;
1749 : int max_proportional_pins;
1750 :
1751 332890 : if (*additional_pins <= 1)
1752 314544 : return;
1753 :
1754 18346 : max_backends = MaxBackends + NUM_AUXILIARY_PROCS;
1755 18346 : max_proportional_pins = NBuffers / max_backends;
1756 :
1757 : /*
1758 : * Subtract the approximate number of buffers already pinned by this
1759 : * backend. We get the number of "overflowed" pins for free, but don't
1760 : * know the number of pins in PrivateRefCountArray. The cost of
1761 : * calculating that exactly doesn't seem worth it, so just assume the max.
1762 : */
1763 18346 : max_proportional_pins -= PrivateRefCountOverflowed + REFCOUNT_ARRAY_ENTRIES;
1764 :
1765 18346 : if (max_proportional_pins < 0)
1766 3236 : max_proportional_pins = 1;
1767 :
1768 18346 : if (*additional_pins > max_proportional_pins)
1769 3236 : *additional_pins = max_proportional_pins;
1770 : }
1771 :
1772 : /*
1773 : * Logic shared between ExtendBufferedRelBy(), ExtendBufferedRelTo(). Just to
1774 : * avoid duplicating the tracing and relpersistence related logic.
1775 : */
1776 : static BlockNumber
1777 343785 : ExtendBufferedRelCommon(ExtendBufferedWhat eb,
1778 : ForkNumber fork,
1779 : BufferAccessStrategy strategy,
1780 : uint32 flags,
1781 : uint32 extend_by,
1782 : BlockNumber extend_upto,
1783 : Buffer *buffers,
1784 : uint32 *extended_by)
1785 : {
1786 : BlockNumber first_block;
1787 :
1788 : TRACE_POSTGRESQL_BUFFER_EXTEND_START(fork,
1789 : eb.smgr->smgr_rlocator.locator.spcOid,
1790 : eb.smgr->smgr_rlocator.locator.dbOid,
1791 : eb.smgr->smgr_rlocator.locator.relNumber,
1792 : eb.smgr->smgr_rlocator.backend,
1793 : extend_by);
1794 :
1795 343785 : if (eb.relpersistence == RELPERSISTENCE_TEMP)
1796 10895 : first_block = ExtendBufferedRelLocal(eb, fork, flags,
1797 : extend_by, extend_upto,
1798 : buffers, &extend_by);
1799 : else
1800 332890 : first_block = ExtendBufferedRelShared(eb, fork, strategy, flags,
1801 : extend_by, extend_upto,
1802 : buffers, &extend_by);
1803 343785 : *extended_by = extend_by;
1804 :
1805 : TRACE_POSTGRESQL_BUFFER_EXTEND_DONE(fork,
1806 : eb.smgr->smgr_rlocator.locator.spcOid,
1807 : eb.smgr->smgr_rlocator.locator.dbOid,
1808 : eb.smgr->smgr_rlocator.locator.relNumber,
1809 : eb.smgr->smgr_rlocator.backend,
1810 : *extended_by,
1811 : first_block);
1812 :
1813 343785 : return first_block;
1814 : }
1815 :
1816 : /*
1817 : * Implementation of ExtendBufferedRelBy() and ExtendBufferedRelTo() for
1818 : * shared buffers.
1819 : */
1820 : static BlockNumber
1821 332890 : ExtendBufferedRelShared(ExtendBufferedWhat eb,
1822 : ForkNumber fork,
1823 : BufferAccessStrategy strategy,
1824 : uint32 flags,
1825 : uint32 extend_by,
1826 : BlockNumber extend_upto,
1827 : Buffer *buffers,
1828 : uint32 *extended_by)
1829 : {
1830 : BlockNumber first_block;
1831 332890 : IOContext io_context = IOContextForStrategy(strategy);
1832 : instr_time io_start;
1833 :
1834 332890 : LimitAdditionalPins(&extend_by);
1835 :
1836 : /*
1837 : * Acquire victim buffers for extension without holding extension lock.
1838 : * Writing out victim buffers is the most expensive part of extending the
1839 : * relation, particularly when doing so requires WAL flushes. Zeroing out
1840 : * the buffers is also quite expensive, so do that before holding the
1841 : * extension lock as well.
1842 : *
1843 : * These pages are pinned by us and not valid. While we hold the pin they
1844 : * can't be acquired as victim buffers by another backend.
1845 : */
1846 702578 : for (uint32 i = 0; i < extend_by; i++)
1847 : {
1848 : Block buf_block;
1849 :
1850 369688 : buffers[i] = GetVictimBuffer(strategy, io_context);
1851 369688 : buf_block = BufHdrGetBlock(GetBufferDescriptor(buffers[i] - 1));
1852 :
1853 : /* new buffers are zero-filled */
1854 369688 : MemSet((char *) buf_block, 0, BLCKSZ);
1855 : }
1856 :
1857 : /* in case we need to pin an existing buffer below */
1858 332890 : ResourceOwnerEnlargeBuffers(CurrentResourceOwner);
1859 :
1860 : /*
1861 : * Lock relation against concurrent extensions, unless requested not to.
1862 : *
1863 : * We use the same extension lock for all forks. That's unnecessarily
1864 : * restrictive, but currently extensions for forks don't happen often
1865 : * enough to make it worth locking more granularly.
1866 : *
1867 : * Note that another backend might have extended the relation by the time
1868 : * we get the lock.
1869 : */
1870 332890 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
1871 : {
1872 289641 : LockRelationForExtension(eb.rel, ExclusiveLock);
1873 289641 : if (eb.rel)
1874 289641 : eb.smgr = RelationGetSmgr(eb.rel);
1875 : }
1876 :
1877 : /*
1878 : * If requested, invalidate size cache, so that smgrnblocks asks the
1879 : * kernel.
1880 : */
1881 332890 : if (flags & EB_CLEAR_SIZE_CACHE)
1882 28427 : eb.smgr->smgr_cached_nblocks[fork] = InvalidBlockNumber;
1883 :
1884 332890 : first_block = smgrnblocks(eb.smgr, fork);
1885 :
1886 : /*
1887 : * Now that we have the accurate relation size, check if the caller wants
1888 : * us to extend to only up to a specific size. If there were concurrent
1889 : * extensions, we might have acquired too many buffers and need to release
1890 : * them.
1891 : */
1892 332890 : if (extend_upto != InvalidBlockNumber)
1893 : {
1894 66293 : uint32 orig_extend_by = extend_by;
1895 :
1896 66293 : if (first_block > extend_upto)
4 andres 1897 UNC 0 : extend_by = 0;
4 andres 1898 GNC 66293 : else if ((uint64) first_block + extend_by > extend_upto)
4 andres 1899 UNC 0 : extend_by = extend_upto - first_block;
1900 :
4 andres 1901 GNC 66293 : for (uint32 i = extend_by; i < orig_extend_by; i++)
1902 : {
4 andres 1903 UNC 0 : BufferDesc *buf_hdr = GetBufferDescriptor(buffers[i] - 1);
1904 :
1905 : /*
1906 : * The victim buffer we acquired peviously is clean and unused,
1907 : * let it be found again quickly
1908 : */
1909 0 : StrategyFreeBuffer(buf_hdr);
1910 0 : UnpinBuffer(buf_hdr);
1911 : }
1912 :
4 andres 1913 GNC 66293 : if (extend_by == 0)
1914 : {
4 andres 1915 UNC 0 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
1916 0 : UnlockRelationForExtension(eb.rel, ExclusiveLock);
1917 0 : *extended_by = extend_by;
1918 0 : return first_block;
1919 : }
1920 : }
1921 :
1922 : /* Fail if relation is already at maximum possible length */
4 andres 1923 GNC 332890 : if ((uint64) first_block + extend_by >= MaxBlockNumber)
4 andres 1924 UNC 0 : ereport(ERROR,
1925 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1926 : errmsg("cannot extend relation %s beyond %u blocks",
1927 : relpath(eb.smgr->smgr_rlocator, fork),
1928 : MaxBlockNumber)));
1929 :
1930 : /*
1931 : * Insert buffers into buffer table, mark as IO_IN_PROGRESS.
1932 : *
1933 : * This needs to happen before we extend the relation, because as soon as
1934 : * we do, other backends can start to read in those pages.
1935 : */
4 andres 1936 GNC 702578 : for (int i = 0; i < extend_by; i++)
1937 : {
1938 369688 : Buffer victim_buf = buffers[i];
1939 369688 : BufferDesc *victim_buf_hdr = GetBufferDescriptor(victim_buf - 1);
1940 : BufferTag tag;
1941 : uint32 hash;
1942 : LWLock *partition_lock;
1943 : int existing_id;
1944 :
1945 369688 : InitBufferTag(&tag, &eb.smgr->smgr_rlocator.locator, fork, first_block + i);
1946 369688 : hash = BufTableHashCode(&tag);
1947 369688 : partition_lock = BufMappingPartitionLock(hash);
1948 :
1949 369688 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
1950 :
1951 369688 : existing_id = BufTableInsert(&tag, hash, victim_buf_hdr->buf_id);
1952 :
1953 : /*
1954 : * We get here only in the corner case where we are trying to extend
1955 : * the relation but we found a pre-existing buffer. This can happen
1956 : * because a prior attempt at extending the relation failed, and
1957 : * because mdread doesn't complain about reads beyond EOF (when
1958 : * zero_damaged_pages is ON) and so a previous attempt to read a block
1959 : * beyond EOF could have left a "valid" zero-filled buffer.
1960 : * Unfortunately, we have also seen this case occurring because of
1961 : * buggy Linux kernels that sometimes return an lseek(SEEK_END) result
1962 : * that doesn't account for a recent write. In that situation, the
1963 : * pre-existing buffer would contain valid data that we don't want to
1964 : * overwrite. Since the legitimate cases should always have left a
1965 : * zero-filled buffer, complain if not PageIsNew.
1966 : */
1967 369688 : if (existing_id >= 0)
1968 : {
4 andres 1969 UNC 0 : BufferDesc *existing_hdr = GetBufferDescriptor(existing_id);
1970 : Block buf_block;
1971 : bool valid;
1972 :
1973 : /*
1974 : * Pin the existing buffer before releasing the partition lock,
1975 : * preventing it from being evicted.
1976 : */
1977 0 : valid = PinBuffer(existing_hdr, strategy);
1978 :
1979 0 : LWLockRelease(partition_lock);
1980 :
1981 : /*
1982 : * The victim buffer we acquired peviously is clean and unused,
1983 : * let it be found again quickly
1984 : */
1985 0 : StrategyFreeBuffer(victim_buf_hdr);
1986 0 : UnpinBuffer(victim_buf_hdr);
1987 :
1988 0 : buffers[i] = BufferDescriptorGetBuffer(existing_hdr);
1989 0 : buf_block = BufHdrGetBlock(existing_hdr);
1990 :
1991 0 : if (valid && !PageIsNew((Page) buf_block))
1992 0 : ereport(ERROR,
1993 : (errmsg("unexpected data beyond EOF in block %u of relation %s",
1994 : existing_hdr->tag.blockNum, relpath(eb.smgr->smgr_rlocator, fork)),
1995 : errhint("This has been seen to occur with buggy kernels; consider updating your system.")));
1996 :
1997 : /*
1998 : * We *must* do smgr[zero]extend before succeeding, else the page
1999 : * will not be reserved by the kernel, and the next P_NEW call
2000 : * will decide to return the same page. Clear the BM_VALID bit,
2001 : * do StartBufferIO() and proceed.
2002 : *
2003 : * Loop to handle the very small possibility that someone re-sets
2004 : * BM_VALID between our clearing it and StartBufferIO inspecting
2005 : * it.
2006 : */
2007 : do
2008 : {
2009 0 : uint32 buf_state = LockBufHdr(existing_hdr);
2010 :
2011 0 : buf_state &= ~BM_VALID;
2012 0 : UnlockBufHdr(existing_hdr, buf_state);
2013 0 : } while (!StartBufferIO(existing_hdr, true));
2014 : }
2015 : else
2016 : {
2017 : uint32 buf_state;
2018 :
4 andres 2019 GNC 369688 : buf_state = LockBufHdr(victim_buf_hdr);
2020 :
2021 : /* some sanity checks while we hold the buffer header lock */
2022 369688 : Assert(!(buf_state & (BM_VALID | BM_TAG_VALID | BM_DIRTY | BM_JUST_DIRTIED)));
2023 369688 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2024 :
2025 369688 : victim_buf_hdr->tag = tag;
2026 :
2027 369688 : buf_state |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2028 369688 : if (eb.relpersistence == RELPERSISTENCE_PERMANENT || fork == INIT_FORKNUM)
2029 365635 : buf_state |= BM_PERMANENT;
2030 :
2031 369688 : UnlockBufHdr(victim_buf_hdr, buf_state);
2032 :
2033 369688 : LWLockRelease(partition_lock);
2034 :
2035 : /* XXX: could combine the locked operations in it with the above */
2036 369688 : StartBufferIO(victim_buf_hdr, true);
2037 : }
2038 : }
2039 :
2 2040 332890 : io_start = pgstat_prepare_io_time();
2041 :
2042 : /*
2043 : * Note: if smgzerorextend fails, we will end up with buffers that are
2044 : * allocated but not marked BM_VALID. The next relation extension will
2045 : * still select the same block number (because the relation didn't get any
2046 : * longer on disk) and so future attempts to extend the relation will find
2047 : * the same buffers (if they have not been recycled) but come right back
2048 : * here to try smgrzeroextend again.
2049 : *
2050 : * We don't need to set checksum for all-zero pages.
2051 : */
4 2052 332890 : smgrzeroextend(eb.smgr, fork, first_block, extend_by, false);
2053 :
2054 : /*
2055 : * Release the file-extension lock; it's now OK for someone else to extend
2056 : * the relation some more.
2057 : *
2058 : * We remove IO_IN_PROGRESS after this, as waking up waiting backends can
2059 : * take noticeable time.
2060 : */
2061 332890 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2062 289641 : UnlockRelationForExtension(eb.rel, ExclusiveLock);
2063 :
2 2064 332890 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context, IOOP_EXTEND,
2065 : io_start, extend_by);
2066 :
2067 : /* Set BM_VALID, terminate IO, and wake up any waiters */
4 2068 702578 : for (int i = 0; i < extend_by; i++)
2069 : {
2070 369688 : Buffer buf = buffers[i];
2071 369688 : BufferDesc *buf_hdr = GetBufferDescriptor(buf - 1);
2072 369688 : bool lock = false;
2073 :
2074 369688 : if (flags & EB_LOCK_FIRST && i == 0)
2075 266410 : lock = true;
2076 103278 : else if (flags & EB_LOCK_TARGET)
2077 : {
2078 36935 : Assert(extend_upto != InvalidBlockNumber);
2079 36935 : if (first_block + i + 1 == extend_upto)
2080 36056 : lock = true;
2081 : }
2082 :
2083 369688 : if (lock)
2084 302466 : LWLockAcquire(BufferDescriptorGetContentLock(buf_hdr), LW_EXCLUSIVE);
2085 :
2086 369688 : TerminateBufferIO(buf_hdr, false, BM_VALID);
2087 : }
2088 :
2089 332890 : pgBufferUsage.shared_blks_written += extend_by;
2090 :
2091 332890 : *extended_by = extend_by;
2092 :
2093 332890 : return first_block;
2094 : }
2095 :
9770 scrappy 2096 ECB : /*
6218 tgl 2097 : * MarkBufferDirty
2098 : *
2099 : * Marks buffer contents as dirty (actual write happens later).
2100 : *
2101 : * Buffer must be pinned and exclusive-locked. (If caller does not hold
2102 : * exclusive lock, then somebody could be in process of writing the buffer,
2103 : * leading to risk of bad data written to disk.)
2104 : */
2105 : void
6218 tgl 2106 GIC 29096899 : MarkBufferDirty(Buffer buffer)
2107 : {
2108 : BufferDesc *bufHdr;
2109 : uint32 buf_state;
2110 : uint32 old_buf_state;
2111 :
6750 2112 29096899 : if (!BufferIsValid(buffer))
4312 peter_e 2113 UIC 0 : elog(ERROR, "bad buffer ID: %d", buffer);
6750 tgl 2114 ECB :
8409 tgl 2115 CBC 29096899 : if (BufferIsLocal(buffer))
2116 : {
6218 tgl 2117 GIC 1138590 : MarkLocalBufferDirty(buffer);
7603 bruce 2118 CBC 1138590 : return;
2119 : }
2120 :
2992 andres 2121 GIC 27958309 : bufHdr = GetBufferDescriptor(buffer - 1);
2122 :
3144 2123 27958309 : Assert(BufferIsPinned(buffer));
2407 simon 2124 CBC 27958309 : Assert(LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
2407 simon 2125 ECB : LW_EXCLUSIVE));
2126 :
2555 andres 2127 CBC 27958309 : old_buf_state = pg_atomic_read_u32(&bufHdr->state);
2128 : for (;;)
2129 : {
2130 27958389 : if (old_buf_state & BM_LOCKED)
2555 andres 2131 GIC 65 : old_buf_state = WaitBufHdrUnlocked(bufHdr);
2132 :
2133 27958389 : buf_state = old_buf_state;
2134 :
2555 andres 2135 CBC 27958389 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2555 andres 2136 GIC 27958389 : buf_state |= BM_DIRTY | BM_JUST_DIRTIED;
2137 :
2138 27958389 : if (pg_atomic_compare_exchange_u32(&bufHdr->state, &old_buf_state,
2139 : buf_state))
2140 27958309 : break;
2141 : }
2142 :
2143 : /*
3987 tgl 2144 ECB : * If the buffer was not dirty already, do vacuum accounting.
2145 : */
2555 andres 2146 GIC 27958309 : if (!(old_buf_state & BM_DIRTY))
2147 : {
4153 alvherre 2148 CBC 743896 : VacuumPageDirty++;
4064 rhaas 2149 743896 : pgBufferUsage.shared_blks_dirtied++;
4153 alvherre 2150 GIC 743896 : if (VacuumCostActive)
2151 1244 : VacuumCostBalance += VacuumCostPageDirty;
2152 : }
2153 : }
2154 :
2155 : /*
2156 : * ReleaseAndReadBuffer -- combine ReleaseBuffer() and ReadBuffer()
2157 : *
2158 : * Formerly, this saved one cycle of acquiring/releasing the BufMgrLock
2159 : * compared to calling the two routines separately. Now it's mainly just
2160 : * a convenience function. However, if the passed buffer is valid and
2161 : * already contains the desired block, we just return it as-is; and that
2162 : * does save considerable work compared to a full release and reacquire.
2163 : *
2164 : * Note: it is OK to pass buffer == InvalidBuffer, indicating that no old
7954 tgl 2165 ECB : * buffer actually needs to be released. This case is the same as ReadBuffer,
2166 : * but can save some tests in the caller.
2167 : */
2168 : Buffer
9770 scrappy 2169 GIC 33565066 : ReleaseAndReadBuffer(Buffer buffer,
2170 : Relation relation,
2171 : BlockNumber blockNum)
9770 scrappy 2172 EUB : {
5050 bruce 2173 GBC 33565066 : ForkNumber forkNum = MAIN_FORKNUM;
2174 : BufferDesc *bufHdr;
2175 :
8002 tgl 2176 GIC 33565066 : if (BufferIsValid(buffer))
2177 : {
3144 andres 2178 22743215 : Assert(BufferIsPinned(buffer));
8002 tgl 2179 22743215 : if (BufferIsLocal(buffer))
2180 : {
2992 andres 2181 105134 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
7974 tgl 2182 107078 : if (bufHdr->tag.blockNum == blockNum &&
228 rhaas 2183 GNC 3888 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
2184 1944 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
7974 tgl 2185 GIC 1944 : return buffer;
4 andres 2186 GNC 103190 : UnpinLocalBuffer(buffer);
8002 tgl 2187 ECB : }
2188 : else
9345 bruce 2189 : {
2992 andres 2190 GIC 22638081 : bufHdr = GetBufferDescriptor(buffer - 1);
6610 tgl 2191 ECB : /* we have pin, so it's ok to examine tag without spinlock */
7974 tgl 2192 CBC 30582933 : if (bufHdr->tag.blockNum == blockNum &&
228 rhaas 2193 GNC 15889704 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
2194 7944852 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
7974 tgl 2195 CBC 7944852 : return buffer;
191 michael 2196 GNC 14693229 : UnpinBuffer(bufHdr);
2197 : }
2198 : }
2199 :
6610 tgl 2200 GIC 25618270 : return ReadBuffer(relation, blockNum);
9770 scrappy 2201 ECB : }
2202 :
2203 : /*
6929 tgl 2204 : * PinBuffer -- make buffer unavailable for replacement.
2205 : *
2206 : * For the default access strategy, the buffer's usage_count is incremented
2207 : * when we first pin it; for other strategies we just make sure the usage_count
2208 : * isn't zero. (The idea of the latter is that we don't want synchronized
5793 2209 : * heap scans to inflate the count, but we need it to not be zero to discourage
2210 : * other backends from stealing buffers from our ring. As long as we cycle
2211 : * through the ring faster than the global clock-sweep cycles, buffers in
2212 : * our ring won't be chosen as victims for replacement by other backends.)
2213 : *
2214 : * This should be applied only to shared buffers, never local ones.
2215 : *
2216 : * Since buffers are pinned/unpinned very frequently, pin buffers without
2217 : * taking the buffer header lock; instead update the state variable in loop of
2218 : * CAS operations. Hopefully it's just a single CAS.
2219 : *
2220 : * Note that ResourceOwnerEnlargeBuffers must have been done already.
2221 : *
2222 : * Returns true if buffer is BM_VALID, else false. This provision allows
2223 : * some callers to avoid an extra spinlock cycle.
2224 : */
2225 : static bool
2701 rhaas 2226 GIC 67748593 : PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy)
6610 tgl 2227 ECB : {
2797 andres 2228 CBC 67748593 : Buffer b = BufferDescriptorGetBuffer(buf);
2229 : bool result;
2230 : PrivateRefCountEntry *ref;
2231 :
4 andres 2232 GNC 67748593 : Assert(!BufferIsLocal(b));
2233 :
2797 andres 2234 GIC 67748593 : ref = GetPrivateRefCountEntry(b, true);
2235 :
3002 2236 67748593 : if (ref == NULL)
2237 : {
2555 andres 2238 ECB : uint32 buf_state;
2239 : uint32 old_buf_state;
2240 :
3002 andres 2241 CBC 65679782 : ReservePrivateRefCountEntry();
2797 andres 2242 GIC 65679782 : ref = NewPrivateRefCountEntry(b);
2243 :
2555 2244 65679782 : old_buf_state = pg_atomic_read_u32(&buf->state);
2245 : for (;;)
5793 tgl 2246 ECB : {
2555 andres 2247 GIC 65691662 : if (old_buf_state & BM_LOCKED)
2555 andres 2248 CBC 503 : old_buf_state = WaitBufHdrUnlocked(buf);
2555 andres 2249 ECB :
2555 andres 2250 GIC 65691662 : buf_state = old_buf_state;
2555 andres 2251 ECB :
2252 : /* increase refcount */
2555 andres 2253 GIC 65691662 : buf_state += BUF_REFCOUNT_ONE;
2555 andres 2254 ECB :
2211 teodor 2255 GIC 65691662 : if (strategy == NULL)
2256 : {
2257 : /* Default case: increase usagecount unless already max. */
2258 65132340 : if (BUF_STATE_GET_USAGECOUNT(buf_state) < BM_MAX_USAGE_COUNT)
2259 3299577 : buf_state += BUF_USAGECOUNT_ONE;
2260 : }
2261 : else
2262 : {
2263 : /*
2264 : * Ring buffers shouldn't evict others from pool. Thus we
2265 : * don't make usagecount more than 1.
2266 : */
2267 559322 : if (BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
2268 24540 : buf_state += BUF_USAGECOUNT_ONE;
2269 : }
2555 andres 2270 ECB :
2555 andres 2271 GIC 65691662 : if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
2272 : buf_state))
2273 : {
2274 65679782 : result = (buf_state & BM_VALID) != 0;
996 pg 2275 ECB :
2276 : /*
2277 : * Assume that we acquired a buffer pin for the purposes of
994 2278 : * Valgrind buffer client checks (even in !result case) to
2279 : * keep things simple. Buffers that are unsafe to access are
2280 : * not generally guaranteed to be marked undefined or
2281 : * non-accessible in any case.
2282 : */
2283 : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(buf), BLCKSZ);
2555 andres 2284 GIC 65679782 : break;
2285 : }
2286 : }
6610 tgl 2287 ECB : }
2288 : else
2289 : {
992 pg 2290 : /*
2291 : * If we previously pinned the buffer, it must surely be valid.
2292 : *
2293 : * Note: We deliberately avoid a Valgrind client request here.
2294 : * Individual access methods can optionally superimpose buffer page
2295 : * client requests on top of our client requests to enforce that
2296 : * buffers are only accessed while locked (and pinned). It's possible
2297 : * that the buffer page is legitimately non-accessible here. We
2298 : * cannot meddle with that.
2299 : */
6610 tgl 2300 GIC 2068811 : result = true;
6610 tgl 2301 ECB : }
2302 :
3144 andres 2303 GIC 67748593 : ref->refcount++;
2304 67748593 : Assert(ref->refcount > 0);
2797 2305 67748593 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
6610 tgl 2306 67748593 : return result;
2307 : }
2308 :
2309 : /*
2310 : * PinBuffer_Locked -- as above, but caller already locked the buffer header.
2311 : * The spinlock is released before return.
2312 : *
2313 : * As this function is called with the spinlock held, the caller has to
2314 : * previously call ReservePrivateRefCountEntry().
2315 : *
2316 : * Currently, no callers of this function want to modify the buffer's
2317 : * usage_count at all, so there's no need for a strategy parameter.
2318 : * Also we don't bother with a BM_VALID test (the caller could check that for
5793 tgl 2319 ECB : * itself).
2320 : *
2321 : * Also all callers only ever use this function when it's known that the
2322 : * buffer can't have a preexisting pin by this backend. That allows us to skip
2323 : * searching the private refcount array & hash, which is a boon, because the
3002 andres 2324 : * spinlock is still held.
2325 : *
2326 : * Note: use of this routine is frequently mandatory, not just an optimization
6610 tgl 2327 : * to save a spin lock/unlock cycle, because we need to pin a buffer before
2328 : * its state can change under us.
2329 : */
2330 : static void
2701 rhaas 2331 GIC 2792328 : PinBuffer_Locked(BufferDesc *buf)
2332 : {
2333 : Buffer b;
2334 : PrivateRefCountEntry *ref;
2335 : uint32 buf_state;
2336 :
3002 andres 2337 ECB : /*
2338 : * As explained, We don't expect any preexisting pins. That allows us to
2339 : * manipulate the PrivateRefCount after releasing the spinlock
2340 : */
2797 andres 2341 GIC 2792328 : Assert(GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf), false) == NULL);
2342 :
2343 : /*
2344 : * Buffer can't have a preexisting pin, so mark its page as defined to
996 pg 2345 ECB : * Valgrind (this is similar to the PinBuffer() case where the backend
2346 : * doesn't already have a buffer pin)
2347 : */
2348 : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(buf), BLCKSZ);
2349 :
2350 : /*
2351 : * Since we hold the buffer spinlock, we can update the buffer state and
2352 : * release the lock in one operation.
2353 : */
2555 andres 2354 GIC 2792328 : buf_state = pg_atomic_read_u32(&buf->state);
2555 andres 2355 CBC 2792328 : Assert(buf_state & BM_LOCKED);
2555 andres 2356 GIC 2792328 : buf_state += BUF_REFCOUNT_ONE;
2357 2792328 : UnlockBufHdr(buf, buf_state);
3002 andres 2358 ECB :
2797 andres 2359 GIC 2792328 : b = BufferDescriptorGetBuffer(buf);
2360 :
2361 2792328 : ref = NewPrivateRefCountEntry(b);
3144 2362 2792328 : ref->refcount++;
2363 :
2797 2364 2792328 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
6929 tgl 2365 2792328 : }
2366 :
2367 : /*
2368 : * UnpinBuffer -- make buffer available for replacement.
2369 : *
2370 : * This should be applied only to shared buffers, never local ones. This
2371 : * always adjusts CurrentResourceOwner.
6929 tgl 2372 ECB : */
2373 : static void
191 michael 2374 GNC 81367861 : UnpinBuffer(BufferDesc *buf)
2375 : {
3144 andres 2376 ECB : PrivateRefCountEntry *ref;
2797 andres 2377 GIC 81367861 : Buffer b = BufferDescriptorGetBuffer(buf);
2378 :
4 andres 2379 GNC 81367861 : Assert(!BufferIsLocal(b));
2380 :
2381 : /* not moving as we're likely deleting it soon anyway */
2797 andres 2382 CBC 81367861 : ref = GetPrivateRefCountEntry(b, false);
3144 andres 2383 GIC 81367861 : Assert(ref != NULL);
2384 :
191 michael 2385 GNC 81367861 : ResourceOwnerForgetBuffer(CurrentResourceOwner, b);
2386 :
3144 andres 2387 GIC 81367861 : Assert(ref->refcount > 0);
2388 81367861 : ref->refcount--;
2389 81367861 : if (ref->refcount == 0)
2390 : {
2391 : uint32 buf_state;
2392 : uint32 old_buf_state;
2555 andres 2393 ECB :
2394 : /*
992 pg 2395 : * Mark buffer non-accessible to Valgrind.
2396 : *
2397 : * Note that the buffer may have already been marked non-accessible
2398 : * within access method code that enforces that buffers are only
2399 : * accessed while a buffer lock is held.
2400 : */
2401 : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
2402 :
2403 : /* I'd better not still hold the buffer content lock */
2672 rhaas 2404 CBC 68472110 : Assert(!LWLockHeldByMe(BufferDescriptorGetContentLock(buf)));
6929 tgl 2405 ECB :
2406 : /*
2555 andres 2407 : * Decrement the shared reference count.
2408 : *
2409 : * Since buffer spinlock holder can update status using just write,
2410 : * it's not safe to use atomic decrement here; thus use a CAS loop.
2411 : */
2555 andres 2412 GIC 68472110 : old_buf_state = pg_atomic_read_u32(&buf->state);
2413 : for (;;)
2414 : {
2555 andres 2415 CBC 68491196 : if (old_buf_state & BM_LOCKED)
2555 andres 2416 GIC 465 : old_buf_state = WaitBufHdrUnlocked(buf);
2555 andres 2417 ECB :
2555 andres 2418 GIC 68491196 : buf_state = old_buf_state;
2555 andres 2419 ECB :
2555 andres 2420 GBC 68491196 : buf_state -= BUF_REFCOUNT_ONE;
6610 tgl 2421 ECB :
2555 andres 2422 GBC 68491196 : if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
2423 : buf_state))
2555 andres 2424 CBC 68472110 : break;
2425 : }
6610 tgl 2426 EUB :
2427 : /* Support LockBufferForCleanup() */
2555 andres 2428 GIC 68472110 : if (buf_state & BM_PIN_COUNT_WAITER)
2429 : {
2430 : /*
2431 : * Acquire the buffer header lock, re-check that there's a waiter.
2555 andres 2432 EUB : * Another backend could have unpinned this buffer, and already
2433 : * woken up the waiter. There's no danger of the buffer being
2434 : * replaced after we unpinned it above, as it's pinned by the
2435 : * waiter.
2555 andres 2436 ECB : */
2555 andres 2437 GIC 2 : buf_state = LockBufHdr(buf);
6610 tgl 2438 EUB :
2555 andres 2439 GBC 2 : if ((buf_state & BM_PIN_COUNT_WAITER) &&
2440 2 : BUF_STATE_GET_REFCOUNT(buf_state) == 1)
2441 2 : {
2442 : /* we just released the last pin other than the waiter's */
479 tmunro 2443 GIC 2 : int wait_backend_pgprocno = buf->wait_backend_pgprocno;
2444 :
2555 andres 2445 2 : buf_state &= ~BM_PIN_COUNT_WAITER;
2555 andres 2446 CBC 2 : UnlockBufHdr(buf, buf_state);
479 tmunro 2447 GBC 2 : ProcSendSignal(wait_backend_pgprocno);
2448 : }
2449 : else
2555 andres 2450 UIC 0 : UnlockBufHdr(buf, buf_state);
2451 : }
3144 andres 2452 GIC 68472110 : ForgetPrivateRefCountEntry(ref);
2453 : }
6929 tgl 2454 81367861 : }
2455 :
2456 : #define ST_SORT sort_checkpoint_bufferids
2457 : #define ST_ELEMENT_TYPE CkptSortItem
2458 : #define ST_COMPARE(a, b) ckpt_buforder_comparator(a, b)
758 tmunro 2459 ECB : #define ST_SCOPE static
2460 : #define ST_DEFINE
2461 : #include <lib/sort_template.h>
2462 :
2463 : /*
2464 : * BufferSync -- Write out all dirty buffers in the pool.
2465 : *
2466 : * This is called at checkpoint time to write out all dirty shared buffers.
2467 : * The checkpoint request flags should be passed in. If CHECKPOINT_IMMEDIATE
3093 andres 2468 : * is set, we disable delays between writes; if CHECKPOINT_IS_SHUTDOWN,
2469 : * CHECKPOINT_END_OF_RECOVERY or CHECKPOINT_FLUSH_ALL is set, we write even
2470 : * unlogged buffers, which are otherwise skipped. The remaining flags
2471 : * currently have no effect here.
8400 tgl 2472 : */
2473 : static void
5764 tgl 2474 CBC 2363 : BufferSync(int flags)
2475 : {
2476 : uint32 buf_state;
2477 : int buf_id;
2478 : int num_to_scan;
2479 : int num_spaces;
2480 : int num_processed;
2481 : int num_written;
2606 andres 2482 GIC 2363 : CkptTsStatus *per_ts_stat = NULL;
2483 : Oid last_tsid;
2484 : binaryheap *ts_heap;
2485 : int i;
4484 rhaas 2486 2363 : int mask = BM_DIRTY;
2487 : WritebackContext wb_context;
2488 :
2489 : /* Make sure we can handle the pin inside SyncOneBuffer */
5764 tgl 2490 CBC 2363 : ResourceOwnerEnlargeBuffers(CurrentResourceOwner);
2491 :
4484 rhaas 2492 EUB : /*
2493 : * Unless this is a shutdown checkpoint or we have been explicitly told,
2494 : * we write only permanent, dirty buffers. But at shutdown or end of
2495 : * recovery, we write all dirty buffers.
2496 : */
3093 andres 2497 GIC 2363 : if (!((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
2498 : CHECKPOINT_FLUSH_ALL))))
4111 rhaas 2499 774 : mask |= BM_PERMANENT;
4484 rhaas 2500 EUB :
2501 : /*
5764 tgl 2502 : * Loop over all buffers, and mark the ones that need to be written with
2503 : * BM_CHECKPOINT_NEEDED. Count them as we go (num_to_scan), so that we
2504 : * can estimate how much work needs to be done.
2505 : *
2506 : * This allows us to write only those pages that were dirty when the
2507 : * checkpoint began, and not those that get dirtied while it proceeds.
2508 : * Whenever a page with BM_CHECKPOINT_NEEDED is written out, either by us
2509 : * later in this function, or by normal backends or the bgwriter cleaning
2510 : * scan, the flag is cleared. Any buffer dirtied after this point won't
2511 : * have the flag set.
2512 : *
2513 : * Note that if we fail to write some buffer, we may leave buffers with
3260 bruce 2514 : * BM_CHECKPOINT_NEEDED still set. This is OK since any such buffer would
5624 2515 : * certainly need to be written for the next checkpoint attempt, too.
2516 : */
2606 andres 2517 GIC 2363 : num_to_scan = 0;
5764 tgl 2518 34165771 : for (buf_id = 0; buf_id < NBuffers; buf_id++)
2519 : {
2701 rhaas 2520 34163408 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
2521 :
2522 : /*
2523 : * Header spinlock is enough to examine BM_DIRTY, see comment in
2524 : * SyncOneBuffer.
2525 : */
2555 andres 2526 34163408 : buf_state = LockBufHdr(bufHdr);
2527 :
2528 34163408 : if ((buf_state & mask) == mask)
2529 : {
2530 : CkptSortItem *item;
2531 :
2555 andres 2532 GBC 477090 : buf_state |= BM_CHECKPOINT_NEEDED;
2533 :
2606 2534 477090 : item = &CkptBufferIds[num_to_scan++];
2535 477090 : item->buf_id = buf_id;
228 rhaas 2536 GNC 477090 : item->tsId = bufHdr->tag.spcOid;
2537 477090 : item->relNumber = BufTagGetRelNumber(&bufHdr->tag);
2538 477090 : item->forkNum = BufTagGetForkNum(&bufHdr->tag);
2606 andres 2539 GIC 477090 : item->blockNum = bufHdr->tag.blockNum;
2540 : }
2541 :
2555 andres 2542 CBC 34163408 : UnlockBufHdr(bufHdr, buf_state);
2543 :
2544 : /* Check for barrier events in case NBuffers is large. */
1207 rhaas 2545 34163408 : if (ProcSignalBarrierPending)
1207 rhaas 2546 LBC 0 : ProcessProcSignalBarrier();
2547 : }
5764 tgl 2548 ECB :
2606 andres 2549 GIC 2363 : if (num_to_scan == 0)
5764 tgl 2550 CBC 767 : return; /* nothing to do */
7087 JanWieck 2551 ECB :
2606 andres 2552 CBC 1596 : WritebackContextInit(&wb_context, &checkpoint_flush_after);
2553 :
2606 andres 2554 ECB : TRACE_POSTGRESQL_BUFFER_SYNC_START(NBuffers, num_to_scan);
2555 :
7081 JanWieck 2556 : /*
2557 : * Sort buffers that need to be written to reduce the likelihood of random
2558 : * IO. The sorting is also important for the implementation of balancing
2606 andres 2559 : * writes between tablespaces. Without balancing writes we'd potentially
2560 : * end up writing to the tablespaces one-by-one; possibly overloading the
2561 : * underlying system.
2562 : */
758 tmunro 2563 CBC 1596 : sort_checkpoint_bufferids(CkptBufferIds, num_to_scan);
2564 :
2606 andres 2565 GIC 1596 : num_spaces = 0;
2566 :
2567 : /*
2568 : * Allocate progress status for each tablespace with buffers that need to
2569 : * be flushed. This requires the to-be-flushed array to be sorted.
2570 : */
2571 1596 : last_tsid = InvalidOid;
2572 478686 : for (i = 0; i < num_to_scan; i++)
2573 : {
2574 : CkptTsStatus *s;
2606 andres 2575 ECB : Oid cur_tsid;
2576 :
2606 andres 2577 GIC 477090 : cur_tsid = CkptBufferIds[i].tsId;
2578 :
2579 : /*
2580 : * Grow array of per-tablespace status structs, every time a new
2581 : * tablespace is found.
2582 : */
2583 477090 : if (last_tsid == InvalidOid || last_tsid != cur_tsid)
2606 andres 2584 CBC 2677 : {
2606 andres 2585 ECB : Size sz;
2586 :
2606 andres 2587 CBC 2677 : num_spaces++;
2588 :
2589 : /*
2590 : * Not worth adding grow-by-power-of-2 logic here - even with a
2606 andres 2591 ECB : * few hundred tablespaces this should be fine.
2592 : */
2606 andres 2593 CBC 2677 : sz = sizeof(CkptTsStatus) * num_spaces;
2606 andres 2594 ECB :
2606 andres 2595 CBC 2677 : if (per_ts_stat == NULL)
2606 andres 2596 GIC 1596 : per_ts_stat = (CkptTsStatus *) palloc(sz);
2606 andres 2597 ECB : else
2606 andres 2598 CBC 1081 : per_ts_stat = (CkptTsStatus *) repalloc(per_ts_stat, sz);
2606 andres 2599 ECB :
2606 andres 2600 GIC 2677 : s = &per_ts_stat[num_spaces - 1];
2606 andres 2601 CBC 2677 : memset(s, 0, sizeof(*s));
2602 2677 : s->tsId = cur_tsid;
2606 andres 2603 ECB :
2604 : /*
2605 : * The first buffer in this tablespace. As CkptBufferIds is sorted
2606 : * by tablespace all (s->num_to_scan) buffers in this tablespace
2607 : * will follow afterwards.
2608 : */
2606 andres 2609 CBC 2677 : s->index = i;
2610 :
2611 : /*
2606 andres 2612 ECB : * progress_slice will be determined once we know how many buffers
2613 : * are in each tablespace, i.e. after this loop.
2614 : */
2615 :
2606 andres 2616 CBC 2677 : last_tsid = cur_tsid;
2617 : }
2618 : else
2619 : {
2606 andres 2620 GIC 474413 : s = &per_ts_stat[num_spaces - 1];
2621 : }
2622 :
2623 477090 : s->num_to_scan++;
2624 :
2625 : /* Check for barrier events. */
1207 rhaas 2626 477090 : if (ProcSignalBarrierPending)
1207 rhaas 2627 UIC 0 : ProcessProcSignalBarrier();
2628 : }
2606 andres 2629 ECB :
2606 andres 2630 GIC 1596 : Assert(num_spaces > 0);
2631 :
2632 : /*
2633 : * Build a min-heap over the write-progress in the individual tablespaces,
2634 : * and compute how large a portion of the total progress a single
2606 andres 2635 ECB : * processed buffer is.
2606 andres 2636 EUB : */
2606 andres 2637 GIC 1596 : ts_heap = binaryheap_allocate(num_spaces,
2606 andres 2638 ECB : ts_ckpt_progress_comparator,
2639 : NULL);
2640 :
2606 andres 2641 CBC 4273 : for (i = 0; i < num_spaces; i++)
2642 : {
2606 andres 2643 GIC 2677 : CkptTsStatus *ts_stat = &per_ts_stat[i];
2606 andres 2644 ECB :
2606 andres 2645 GIC 2677 : ts_stat->progress_slice = (float8) num_to_scan / ts_stat->num_to_scan;
2606 andres 2646 ECB :
2606 andres 2647 CBC 2677 : binaryheap_add_unordered(ts_heap, PointerGetDatum(ts_stat));
2648 : }
2649 :
2650 1596 : binaryheap_build(ts_heap);
2651 :
2652 : /*
2606 andres 2653 ECB : * Iterate through to-be-checkpointed buffers and write the ones (still)
2654 : * marked with BM_CHECKPOINT_NEEDED. The writes are balanced between
2655 : * tablespaces; otherwise the sorting would lead to only one tablespace
2656 : * receiving writes at a time, making inefficient use of the hardware.
2657 : */
2606 andres 2658 CBC 1596 : num_processed = 0;
5764 tgl 2659 1596 : num_written = 0;
2606 andres 2660 GIC 478686 : while (!binaryheap_empty(ts_heap))
7081 JanWieck 2661 ECB : {
2606 andres 2662 GIC 477090 : BufferDesc *bufHdr = NULL;
2606 andres 2663 ECB : CkptTsStatus *ts_stat = (CkptTsStatus *)
2606 andres 2664 GIC 477090 : DatumGetPointer(binaryheap_first(ts_heap));
2665 :
2666 477090 : buf_id = CkptBufferIds[ts_stat->index].buf_id;
2667 477090 : Assert(buf_id != -1);
2668 :
2606 andres 2669 CBC 477090 : bufHdr = GetBufferDescriptor(buf_id);
2670 :
2671 477090 : num_processed++;
5854 magnus 2672 ECB :
5764 tgl 2673 : /*
2674 : * We don't need to acquire the lock here, because we're only looking
2675 : * at a single bit. It's possible that someone else writes the buffer
2676 : * and clears the flag right after we check, but that doesn't matter
2677 : * since SyncOneBuffer will then do nothing. However, there is a
2678 : * further race condition: it's conceivable that between the time we
2679 : * examine the bit here and the time SyncOneBuffer acquires the lock,
2680 : * someone else not only wrote the buffer but replaced it with another
2681 : * page and dirtied it. In that improbable case, SyncOneBuffer will
2682 : * write the buffer though we didn't need to. It doesn't seem worth
2683 : * guarding against this, though.
2684 : */
2555 andres 2685 GIC 477090 : if (pg_atomic_read_u32(&bufHdr->state) & BM_CHECKPOINT_NEEDED)
2686 : {
2606 2687 474825 : if (SyncOneBuffer(buf_id, false, &wb_context) & BUF_WRITTEN)
2688 : {
2689 : TRACE_POSTGRESQL_BUFFER_SYNC_WRITTEN(buf_id);
368 2690 474825 : PendingCheckpointerStats.buf_written_checkpoints++;
5764 tgl 2691 474825 : num_written++;
2606 andres 2692 ECB : }
2693 : }
2694 :
2695 : /*
2501 stark 2696 : * Measure progress independent of actually having to flush the buffer
2697 : * - otherwise writing become unbalanced.
2698 : */
2606 andres 2699 CBC 477090 : ts_stat->progress += ts_stat->progress_slice;
2606 andres 2700 GIC 477090 : ts_stat->num_scanned++;
2606 andres 2701 CBC 477090 : ts_stat->index++;
5764 tgl 2702 ECB :
2703 : /* Have all the buffers from the tablespace been processed? */
2606 andres 2704 CBC 477090 : if (ts_stat->num_scanned == ts_stat->num_to_scan)
2606 andres 2705 ECB : {
2606 andres 2706 CBC 2677 : binaryheap_remove_first(ts_heap);
2606 andres 2707 ECB : }
2708 : else
2709 : {
2710 : /* update heap with the new progress */
2606 andres 2711 GIC 474413 : binaryheap_replace_first(ts_heap, PointerGetDatum(ts_stat));
2712 : }
5764 tgl 2713 ECB :
2714 : /*
2606 andres 2715 : * Sleep to throttle our I/O rate.
1207 rhaas 2716 : *
2717 : * (This will check for barrier events even if it doesn't sleep.)
2606 andres 2718 : */
2606 andres 2719 CBC 477090 : CheckpointWriteDelay(flags, (double) num_processed / num_to_scan);
2720 : }
2721 :
2722 : /* issue all pending flushes */
2723 1596 : IssuePendingWritebacks(&wb_context);
2724 :
2606 andres 2725 GIC 1596 : pfree(per_ts_stat);
2726 1596 : per_ts_stat = NULL;
2727 1596 : binaryheap_free(ts_heap);
2728 :
2729 : /*
2730 : * Update checkpoint statistics. As noted above, this doesn't include
2731 : * buffers written by other backends or bgwriter scan.
2732 : */
5762 tgl 2733 1596 : CheckpointStats.ckpt_bufs_written += num_written;
2734 :
2735 : TRACE_POSTGRESQL_BUFFER_SYNC_DONE(NBuffers, num_written, num_to_scan);
2736 : }
2737 :
2738 : /*
2739 : * BgBufferSync -- Write out some dirty buffers in the pool.
2740 : *
2741 : * This is called periodically by the background writer process.
2742 : *
2743 : * Returns true if it's appropriate for the bgwriter process to go into
2744 : * low-power hibernation mode. (This happens if the strategy clock sweep
2745 : * has been "lapped" and no buffer allocations have occurred recently,
2746 : * or if the bgwriter has been effectively disabled by setting
2747 : * bgwriter_lru_maxpages to 0.)
2748 : */
4091 heikki.linnakangas 2749 ECB : bool
2606 andres 2750 GIC 10438 : BgBufferSync(WritebackContext *wb_context)
6610 tgl 2751 ECB : {
2752 : /* info obtained from freelist.c */
2753 : int strategy_buf_id;
2754 : uint32 strategy_passes;
5675 2755 : uint32 recent_alloc;
2756 :
2757 : /*
2758 : * Information saved between calls so we can determine the strategy
2759 : * point's advance rate and avoid scanning already-cleaned buffers.
2760 : */
2761 : static bool saved_info_valid = false;
2762 : static int prev_strategy_buf_id;
2763 : static uint32 prev_strategy_passes;
2764 : static int next_to_clean;
2765 : static uint32 next_passes;
2766 :
2767 : /* Moving averages of allocation rate and clean-buffer density */
2768 : static float smoothed_alloc = 0;
2769 : static float smoothed_density = 10.0;
2770 :
2771 : /* Potentially these could be tunables, but for now, not */
5675 tgl 2772 GIC 10438 : float smoothing_samples = 16;
5675 tgl 2773 CBC 10438 : float scan_whole_pool_milliseconds = 120000.0;
2774 :
2775 : /* Used to compute how far we scan ahead */
5675 tgl 2776 ECB : long strategy_delta;
2777 : int bufs_to_lap;
2778 : int bufs_ahead;
2779 : float scans_per_alloc;
2780 : int reusable_buffers_est;
2781 : int upcoming_alloc_est;
2782 : int min_scan_buffers;
2783 :
2784 : /* Variables for the scanning loop proper */
2785 : int num_to_scan;
2786 : int num_written;
2787 : int reusable_buffers;
2788 :
2789 : /* Variables for final smoothed_density update */
3987 2790 : long new_strategy_delta;
2791 : uint32 new_recent_alloc;
2792 :
2793 : /*
5624 bruce 2794 : * Find out where the freelist clock sweep currently is, and how many
2795 : * buffer allocations have happened since our last call.
2796 : */
5675 tgl 2797 CBC 10438 : strategy_buf_id = StrategySyncStart(&strategy_passes, &recent_alloc);
2798 :
2799 : /* Report buffer alloc counts to pgstat */
368 andres 2800 GIC 10438 : PendingBgWriterStats.buf_alloc += recent_alloc;
2801 :
2802 : /*
2803 : * If we're not running the LRU scan, just stop after doing the stats
2804 : * stuff. We mark the saved state invalid so that we can recover sanely
2805 : * if LRU scan is turned back on later.
2806 : */
5675 tgl 2807 CBC 10438 : if (bgwriter_lru_maxpages <= 0)
2808 : {
5675 tgl 2809 UIC 0 : saved_info_valid = false;
4091 heikki.linnakangas 2810 0 : return true;
2811 : }
2812 :
2813 : /*
2814 : * Compute strategy_delta = how many buffers have been scanned by the
2815 : * clock sweep since last time. If first time through, assume none. Then
2816 : * see if we are still ahead of the clock sweep, and if so, how many
2817 : * buffers we could scan before we'd catch up with it and "lap" it. Note:
2818 : * weird-looking coding of xxx_passes comparisons are to avoid bogus
2819 : * behavior when the passes counts wrap around.
2820 : */
5675 tgl 2821 GIC 10438 : if (saved_info_valid)
2822 : {
5624 bruce 2823 CBC 10083 : int32 passes_delta = strategy_passes - prev_strategy_passes;
2824 :
5675 tgl 2825 GIC 10083 : strategy_delta = strategy_buf_id - prev_strategy_buf_id;
2118 tgl 2826 CBC 10083 : strategy_delta += (long) passes_delta * NBuffers;
5624 bruce 2827 ECB :
5675 tgl 2828 CBC 10083 : Assert(strategy_delta >= 0);
5675 tgl 2829 ECB :
5675 tgl 2830 GIC 10083 : if ((int32) (next_passes - strategy_passes) > 0)
2831 : {
2832 : /* we're one pass ahead of the strategy point */
2833 2017 : bufs_to_lap = strategy_buf_id - next_to_clean;
2834 : #ifdef BGW_DEBUG
2835 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
2836 : next_passes, next_to_clean,
2837 : strategy_passes, strategy_buf_id,
2838 : strategy_delta, bufs_to_lap);
2839 : #endif
2840 : }
2841 8066 : else if (next_passes == strategy_passes &&
2842 6768 : next_to_clean >= strategy_buf_id)
2843 : {
2844 : /* on same pass, but ahead or at least not behind */
2845 6574 : bufs_to_lap = NBuffers - (next_to_clean - strategy_buf_id);
2846 : #ifdef BGW_DEBUG
2847 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
2848 : next_passes, next_to_clean,
2849 : strategy_passes, strategy_buf_id,
2850 : strategy_delta, bufs_to_lap);
2851 : #endif
2852 : }
2853 : else
5675 tgl 2854 ECB : {
2855 : /*
2856 : * We're behind, so skip forward to the strategy point and start
2857 : * cleaning from there.
2858 : */
2859 : #ifdef BGW_DEBUG
2860 : elog(DEBUG2, "bgwriter behind: bgw %u-%u strategy %u-%u delta=%ld",
2861 : next_passes, next_to_clean,
2862 : strategy_passes, strategy_buf_id,
2863 : strategy_delta);
2864 : #endif
5675 tgl 2865 GIC 1492 : next_to_clean = strategy_buf_id;
2866 1492 : next_passes = strategy_passes;
2867 1492 : bufs_to_lap = NBuffers;
2868 : }
2869 : }
2870 : else
2871 : {
2872 : /*
2873 : * Initializing at startup or after LRU scanning had been off. Always
2874 : * start at the strategy point.
2875 : */
2876 : #ifdef BGW_DEBUG
5675 tgl 2877 ECB : elog(DEBUG2, "bgwriter initializing: strategy %u-%u",
2878 : strategy_passes, strategy_buf_id);
2879 : #endif
5675 tgl 2880 CBC 355 : strategy_delta = 0;
5675 tgl 2881 GIC 355 : next_to_clean = strategy_buf_id;
5675 tgl 2882 CBC 355 : next_passes = strategy_passes;
5675 tgl 2883 GIC 355 : bufs_to_lap = NBuffers;
5675 tgl 2884 ECB : }
2885 :
2886 : /* Update saved info for next time */
5675 tgl 2887 CBC 10438 : prev_strategy_buf_id = strategy_buf_id;
2888 10438 : prev_strategy_passes = strategy_passes;
5675 tgl 2889 GIC 10438 : saved_info_valid = true;
2890 :
2891 : /*
2892 : * Compute how many buffers had to be scanned for each new allocation, ie,
2893 : * 1/density of reusable buffers, and track a moving average of that.
2894 : *
2895 : * If the strategy point didn't move, we don't update the density estimate
2896 : */
5675 tgl 2897 CBC 10438 : if (strategy_delta > 0 && recent_alloc > 0)
2898 : {
5675 tgl 2899 GIC 2053 : scans_per_alloc = (float) strategy_delta / (float) recent_alloc;
5675 tgl 2900 CBC 2053 : smoothed_density += (scans_per_alloc - smoothed_density) /
2901 : smoothing_samples;
5675 tgl 2902 ECB : }
2903 :
2904 : /*
2905 : * Estimate how many reusable buffers there are between the current
5624 bruce 2906 : * strategy point and where we've scanned ahead to, based on the smoothed
2907 : * density estimate.
5675 tgl 2908 : */
5675 tgl 2909 GIC 10438 : bufs_ahead = NBuffers - bufs_to_lap;
5675 tgl 2910 CBC 10438 : reusable_buffers_est = (float) bufs_ahead / smoothed_density;
6929 tgl 2911 ECB :
5675 2912 : /*
2913 : * Track a moving average of recent buffer allocations. Here, rather than
2914 : * a true average we want a fast-attack, slow-decline behavior: we
2915 : * immediately follow any increase.
2916 : */
5675 tgl 2917 GIC 10438 : if (smoothed_alloc <= (float) recent_alloc)
2918 2468 : smoothed_alloc = recent_alloc;
2919 : else
2920 7970 : smoothed_alloc += ((float) recent_alloc - smoothed_alloc) /
2921 : smoothing_samples;
2922 :
2923 : /* Scale the estimate by a GUC to allow more aggressive tuning. */
4159 2924 10438 : upcoming_alloc_est = (int) (smoothed_alloc * bgwriter_lru_multiplier);
2925 :
2926 : /*
4159 tgl 2927 ECB : * If recent_alloc remains at zero for many cycles, smoothed_alloc will
2928 : * eventually underflow to zero, and the underflows produce annoying
2929 : * kernel warnings on some platforms. Once upcoming_alloc_est has gone to
2930 : * zero, there's no point in tracking smaller and smaller values of
2931 : * smoothed_alloc, so just reset it to exactly zero to avoid this
2932 : * syndrome. It will pop back up as soon as recent_alloc increases.
2933 : */
4159 tgl 2934 GIC 10438 : if (upcoming_alloc_est == 0)
4159 tgl 2935 CBC 1384 : smoothed_alloc = 0;
2936 :
2937 : /*
5675 tgl 2938 ECB : * Even in cases where there's been little or no buffer allocation
2939 : * activity, we want to make a small amount of progress through the buffer
2940 : * cache so that as many reusable buffers as possible are clean after an
5624 bruce 2941 : * idle period.
2942 : *
2943 : * (scan_whole_pool_milliseconds / BgWriterDelay) computes how many times
2944 : * the BGW will be called during the scan_whole_pool time; slice the
2945 : * buffer pool into that many sections.
2946 : */
5675 tgl 2947 CBC 10438 : min_scan_buffers = (int) (NBuffers / (scan_whole_pool_milliseconds / BgWriterDelay));
2948 :
5675 tgl 2949 GIC 10438 : if (upcoming_alloc_est < (min_scan_buffers + reusable_buffers_est))
2950 : {
5675 tgl 2951 ECB : #ifdef BGW_DEBUG
2952 : elog(DEBUG2, "bgwriter: alloc_est=%d too small, using min=%d + reusable_est=%d",
2953 : upcoming_alloc_est, min_scan_buffers, reusable_buffers_est);
2954 : #endif
5675 tgl 2955 GIC 5960 : upcoming_alloc_est = min_scan_buffers + reusable_buffers_est;
2956 : }
2957 :
2958 : /*
2959 : * Now write out dirty reusable buffers, working forward from the
5624 bruce 2960 ECB : * next_to_clean point, until we have lapped the strategy scan, or cleaned
2961 : * enough buffers to match our estimate of the next cycle's allocation
2962 : * requirements, or hit the bgwriter_lru_maxpages limit.
5675 tgl 2963 : */
2964 :
2965 : /* Make sure we can handle the pin inside SyncOneBuffer */
5675 tgl 2966 CBC 10438 : ResourceOwnerEnlargeBuffers(CurrentResourceOwner);
2967 :
2968 10438 : num_to_scan = bufs_to_lap;
2969 10438 : num_written = 0;
2970 10438 : reusable_buffers = reusable_buffers_est;
2971 :
2972 : /* Execute the LRU scan */
5675 tgl 2973 GBC 1221474 : while (num_to_scan > 0 && reusable_buffers < upcoming_alloc_est)
2974 : {
2555 andres 2975 CBC 1211036 : int sync_state = SyncOneBuffer(next_to_clean, true,
2976 : wb_context);
5675 tgl 2977 ECB :
5675 tgl 2978 GIC 1211036 : if (++next_to_clean >= NBuffers)
2979 : {
2980 1758 : next_to_clean = 0;
2981 1758 : next_passes++;
2982 : }
2983 1211036 : num_to_scan--;
2984 :
2555 andres 2985 1211036 : if (sync_state & BUF_WRITTEN)
2986 : {
5675 tgl 2987 13485 : reusable_buffers++;
2988 13485 : if (++num_written >= bgwriter_lru_maxpages)
2989 : {
368 andres 2990 UIC 0 : PendingBgWriterStats.maxwritten_clean++;
5675 tgl 2991 0 : break;
2992 : }
2993 : }
2555 andres 2994 GIC 1197551 : else if (sync_state & BUF_REUSABLE)
5675 tgl 2995 906569 : reusable_buffers++;
2996 : }
5675 tgl 2997 ECB :
368 andres 2998 GIC 10438 : PendingBgWriterStats.buf_written_clean += num_written;
2999 :
3000 : #ifdef BGW_DEBUG
3001 : 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",
3002 : recent_alloc, smoothed_alloc, strategy_delta, bufs_ahead,
3003 : smoothed_density, reusable_buffers_est, upcoming_alloc_est,
3004 : bufs_to_lap - num_to_scan,
5675 tgl 3005 ECB : num_written,
3006 : reusable_buffers - reusable_buffers_est);
3007 : #endif
3008 :
3009 : /*
3010 : * Consider the above scan as being like a new allocation scan.
3011 : * Characterize its density and update the smoothed one based on it. This
3012 : * effectively halves the moving average period in cases where both the
5624 bruce 3013 : * strategy and the background writer are doing some useful scanning,
3014 : * which is helpful because a long memory isn't as desirable on the
3015 : * density estimates.
3016 : */
3987 tgl 3017 GIC 10438 : new_strategy_delta = bufs_to_lap - num_to_scan;
3018 10438 : new_recent_alloc = reusable_buffers - reusable_buffers_est;
3019 10438 : if (new_strategy_delta > 0 && new_recent_alloc > 0)
5675 tgl 3020 ECB : {
3987 tgl 3021 GIC 7699 : scans_per_alloc = (float) new_strategy_delta / (float) new_recent_alloc;
5675 tgl 3022 CBC 7699 : smoothed_density += (scans_per_alloc - smoothed_density) /
3023 : smoothing_samples;
3024 :
3025 : #ifdef BGW_DEBUG
3026 : elog(DEBUG2, "bgwriter: cleaner density alloc=%u scan=%ld density=%.2f new smoothed=%.2f",
3027 : new_recent_alloc, new_strategy_delta,
3028 : scans_per_alloc, smoothed_density);
3029 : #endif
3030 : }
3031 :
3032 : /* Return true if OK to hibernate */
3987 tgl 3033 GIC 10438 : return (bufs_to_lap == 0 && recent_alloc == 0);
3034 : }
3035 :
3036 : /*
3037 : * SyncOneBuffer -- process a single buffer during syncing.
3038 : *
3039 : * If skip_recently_used is true, we don't write currently-pinned buffers, nor
6610 tgl 3040 ECB : * buffers marked recently used, as these are not replacement candidates.
8539 3041 : *
3042 : * Returns a bitmask containing the following flag bits:
5675 3043 : * BUF_WRITTEN: we wrote the buffer.
3044 : * BUF_REUSABLE: buffer is available for replacement, ie, it has
3045 : * pin count 0 and usage count 0.
3046 : *
3047 : * (BUF_WRITTEN could be set in error if FlushBuffer finds the buffer clean
3048 : * after locking it, but we don't care all that much.)
6610 3049 : *
3050 : * Note: caller must have done ResourceOwnerEnlargeBuffers.
9770 scrappy 3051 : */
3052 : static int
2606 andres 3053 GIC 1685861 : SyncOneBuffer(int buf_id, bool skip_recently_used, WritebackContext *wb_context)
3054 : {
2701 rhaas 3055 CBC 1685861 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
5624 bruce 3056 GIC 1685861 : int result = 0;
2555 andres 3057 ECB : uint32 buf_state;
2606 3058 : BufferTag tag;
6610 tgl 3059 :
3002 andres 3060 CBC 1685861 : ReservePrivateRefCountEntry();
3002 andres 3061 ECB :
8483 inoue 3062 : /*
3063 : * Check whether buffer needs writing.
3064 : *
6347 bruce 3065 : * We can make this check without taking the buffer content lock so long
3066 : * as we mark pages dirty in access methods *before* logging changes with
3067 : * XLogInsert(): if someone marks the buffer dirty just after our check we
6385 3068 : * don't worry because our checkpoint.redo points before log record for
6385 bruce 3069 EUB : * upcoming changes and so we are not required to write such dirty buffer.
3070 : */
2555 andres 3071 GIC 1685861 : buf_state = LockBufHdr(bufHdr);
5675 tgl 3072 ECB :
2555 andres 3073 CBC 1685861 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 0 &&
2555 andres 3074 GIC 1685075 : BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
2555 andres 3075 ECB : {
5675 tgl 3076 GIC 920525 : result |= BUF_REUSABLE;
3077 : }
3078 765336 : else if (skip_recently_used)
3079 : {
3080 : /* Caller told us not to write recently-used buffers */
2555 andres 3081 290982 : UnlockBufHdr(bufHdr, buf_state);
5675 tgl 3082 290982 : return result;
3083 : }
3084 :
2555 andres 3085 1394879 : if (!(buf_state & BM_VALID) || !(buf_state & BM_DIRTY))
6610 tgl 3086 ECB : {
3087 : /* It's clean, so nothing to do */
2555 andres 3088 CBC 906569 : UnlockBufHdr(bufHdr, buf_state);
5675 tgl 3089 GIC 906569 : return result;
3090 : }
3091 :
3092 : /*
3093 : * Pin it, share-lock it, write it. (FlushBuffer will do nothing if the
6385 bruce 3094 ECB : * buffer is clean by the time we've locked it.)
6610 tgl 3095 : */
6610 tgl 3096 GIC 488310 : PinBuffer_Locked(bufHdr);
2672 rhaas 3097 488310 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_SHARED);
3098 :
59 andres 3099 GNC 488310 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
6610 tgl 3100 ECB :
2672 rhaas 3101 GIC 488310 : LWLockRelease(BufferDescriptorGetContentLock(bufHdr));
3102 :
2606 andres 3103 488310 : tag = bufHdr->tag;
3104 :
191 michael 3105 GNC 488310 : UnpinBuffer(bufHdr);
6610 tgl 3106 ECB :
2606 andres 3107 CBC 488310 : ScheduleBufferTagForWriteback(wb_context, &tag);
3108 :
5675 tgl 3109 GIC 488310 : return result | BUF_WRITTEN;
9770 scrappy 3110 ECB : }
3111 :
3112 : /*
3113 : * AtEOXact_Buffers - clean up at end of transaction.
3114 : *
3115 : * As of PostgreSQL 8.0, buffer pins should get released by the
3144 andres 3116 : * ResourceOwner mechanism. This routine is just a debugging
3117 : * cross-check that no pins remain.
9770 scrappy 3118 : */
3119 : void
7551 tgl 3120 GIC 486167 : AtEOXact_Buffers(bool isCommit)
9770 scrappy 3121 ECB : {
3215 andres 3122 GIC 486167 : CheckForBufferLeaks();
6749 tgl 3123 ECB :
6749 tgl 3124 CBC 486167 : AtEOXact_LocalBuffers(isCommit);
3144 andres 3125 ECB :
3144 andres 3126 GIC 486167 : Assert(PrivateRefCountOverflowed == 0);
3127 486167 : }
3128 :
3129 : /*
3130 : * Initialize access to shared buffer pool
3131 : *
3144 andres 3132 ECB : * This is called during backend startup (whether standalone or under the
3133 : * postmaster). It sets up for this backend's access to the already-existing
3134 : * buffer pool.
3135 : */
3136 : void
3144 andres 3137 GIC 13291 : InitBufferPoolAccess(void)
3138 : {
3144 andres 3139 ECB : HASHCTL hash_ctl;
3140 :
3144 andres 3141 GIC 13291 : memset(&PrivateRefCountArray, 0, sizeof(PrivateRefCountArray));
3142 :
3144 andres 3143 CBC 13291 : hash_ctl.keysize = sizeof(int32);
2604 andres 3144 GIC 13291 : hash_ctl.entrysize = sizeof(PrivateRefCountEntry);
3145 :
3144 andres 3146 CBC 13291 : PrivateRefCountHash = hash_create("PrivateRefCount", 100, &hash_ctl,
3147 : HASH_ELEM | HASH_BLOBS);
3148 :
612 andres 3149 ECB : /*
612 andres 3150 EUB : * AtProcExit_Buffers needs LWLock access, and thereby has to be called at
3151 : * the corresponding phase of backend shutdown.
3152 : */
612 andres 3153 CBC 13291 : Assert(MyProc != NULL);
6453 tgl 3154 GIC 13291 : on_shmem_exit(AtProcExit_Buffers, 0);
3155 13291 : }
3156 :
3157 : /*
3158 : * During backend exit, ensure that we released all shared-buffer locks and
3159 : * assert that we have no remaining pins.
6453 tgl 3160 ECB : */
3161 : static void
6453 tgl 3162 GIC 13291 : AtProcExit_Buffers(int code, Datum arg)
3163 : {
6749 3164 13291 : UnlockBuffers();
6749 tgl 3165 ECB :
3215 andres 3166 GIC 13291 : CheckForBufferLeaks();
3215 andres 3167 ECB :
3168 : /* localbuf.c needs a chance too */
3215 andres 3169 CBC 13291 : AtProcExit_LocalBuffers();
3215 andres 3170 GIC 13291 : }
3171 :
3215 andres 3172 ECB : /*
3173 : * CheckForBufferLeaks - ensure this backend holds no buffer pins
3174 : *
3175 : * As of PostgreSQL 8.0, buffer pins should get released by the
3176 : * ResourceOwner mechanism. This routine is just a debugging
3177 : * cross-check that no pins remain.
3178 : */
3179 : static void
3215 andres 3180 CBC 499458 : CheckForBufferLeaks(void)
3215 andres 3181 ECB : {
4518 rhaas 3182 : #ifdef USE_ASSERT_CHECKING
3215 andres 3183 GIC 499458 : int RefCountErrors = 0;
3144 andres 3184 ECB : PrivateRefCountEntry *res;
3185 : int i;
3186 :
3187 : /* check the array */
3144 andres 3188 CBC 4495122 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
3144 andres 3189 ECB : {
3144 andres 3190 GIC 3995664 : res = &PrivateRefCountArray[i];
3144 andres 3191 ECB :
3144 andres 3192 GIC 3995664 : if (res->buffer != InvalidBuffer)
3144 andres 3193 ECB : {
3144 andres 3194 UIC 0 : PrintBufferLeakWarning(res->buffer);
3195 0 : RefCountErrors++;
3196 : }
3197 : }
3198 :
3199 : /* if necessary search the hash */
3144 andres 3200 GIC 499458 : if (PrivateRefCountOverflowed)
3201 : {
3202 : HASH_SEQ_STATUS hstat;
3203 :
3144 andres 3204 UIC 0 : hash_seq_init(&hstat, PrivateRefCountHash);
3205 0 : while ((res = (PrivateRefCountEntry *) hash_seq_search(&hstat)) != NULL)
3206 : {
3144 andres 3207 LBC 0 : PrintBufferLeakWarning(res->buffer);
3215 andres 3208 UIC 0 : RefCountErrors++;
6840 tgl 3209 ECB : }
3210 : }
3211 :
3215 andres 3212 CBC 499458 : Assert(RefCountErrors == 0);
4518 rhaas 3213 ECB : #endif
6749 tgl 3214 GIC 499458 : }
3215 :
3216 : /*
3217 : * Helper routine to issue warnings when a buffer is unexpectedly pinned
3218 : */
3219 : void
6749 tgl 3220 UIC 0 : PrintBufferLeakWarning(Buffer buffer)
6749 tgl 3221 ECB : {
2701 rhaas 3222 : BufferDesc *buf;
6749 tgl 3223 : int32 loccount;
3224 : char *path;
3225 : BackendId backend;
2555 andres 3226 : uint32 buf_state;
3227 :
6749 tgl 3228 LBC 0 : Assert(BufferIsValid(buffer));
6749 tgl 3229 UIC 0 : if (BufferIsLocal(buffer))
3230 : {
2992 andres 3231 0 : buf = GetLocalBufferDescriptor(-buffer - 1);
6749 tgl 3232 0 : loccount = LocalRefCount[-buffer - 1];
4622 rhaas 3233 LBC 0 : backend = MyBackendId;
3234 : }
3235 : else
3236 : {
2992 andres 3237 UIC 0 : buf = GetBufferDescriptor(buffer - 1);
3144 3238 0 : loccount = GetPrivateRefCount(buffer);
4622 rhaas 3239 0 : backend = InvalidBackendId;
3240 : }
6749 tgl 3241 ECB :
3242 : /* theoretically we should lock the bufhdr here */
228 rhaas 3243 UNC 0 : path = relpathbackend(BufTagGetRelFileLocator(&buf->tag), backend,
3244 : BufTagGetForkNum(&buf->tag));
2555 andres 3245 UIC 0 : buf_state = pg_atomic_read_u32(&buf->state);
6749 tgl 3246 LBC 0 : elog(WARNING,
3247 : "buffer refcount leak: [%03d] "
5262 heikki.linnakangas 3248 ECB : "(rel=%s, blockNum=%u, flags=0x%x, refcount=%u %d)",
3249 : buffer, path,
2555 andres 3250 : buf->tag.blockNum, buf_state & BUF_FLAG_MASK,
3251 : BUF_STATE_GET_REFCOUNT(buf_state), loccount);
5262 heikki.linnakangas 3252 UIC 0 : pfree(path);
6856 tgl 3253 0 : }
3254 :
3255 : /*
5764 tgl 3256 ECB : * CheckPointBuffers
3257 : *
3258 : * Flush all dirty blocks in buffer pool to disk at checkpoint time.
3259 : *
3260 : * Note: temporary relations do not participate in checkpoints, so they don't
3261 : * need to be flushed.
3262 : */
3263 : void
5764 tgl 3264 GIC 2363 : CheckPointBuffers(int flags)
3265 : {
3266 2363 : BufferSync(flags);
8165 vadim4o 3267 2363 : }
3268 :
3269 : /*
3270 : * BufferGetBlockNumber
3271 : * Returns the block number associated with a buffer.
3272 : *
3273 : * Note:
3274 : * Assumes that the buffer is valid and pinned, else the
3275 : * value may be obsolete immediately...
3276 : */
3277 : BlockNumber
9770 scrappy 3278 104052285 : BufferGetBlockNumber(Buffer buffer)
3279 : {
3280 : BufferDesc *bufHdr;
3281 :
7664 bruce 3282 104052285 : Assert(BufferIsPinned(buffer));
3283 :
9345 3284 104052285 : if (BufferIsLocal(buffer))
2992 andres 3285 CBC 2225860 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
9345 bruce 3286 ECB : else
2992 andres 3287 GIC 101826425 : bufHdr = GetBufferDescriptor(buffer - 1);
3288 :
3289 : /* pinned, so OK to read tag without spinlock */
6610 tgl 3290 104052285 : return bufHdr->tag.blockNum;
3291 : }
3292 :
3293 : /*
3294 : * BufferGetTag
3295 : * Returns the relfilelocator, fork number and block number associated with
3296 : * a buffer.
3297 : */
3298 : void
277 rhaas 3299 GNC 23082246 : BufferGetTag(Buffer buffer, RelFileLocator *rlocator, ForkNumber *forknum,
3300 : BlockNumber *blknum)
3301 : {
3302 : BufferDesc *bufHdr;
3303 :
3304 : /* Do the same checks as BufferGetBlockNumber. */
5354 heikki.linnakangas 3305 GIC 23082246 : Assert(BufferIsPinned(buffer));
3306 :
6929 tgl 3307 23082246 : if (BufferIsLocal(buffer))
2992 andres 3308 UIC 0 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
3309 : else
2992 andres 3310 CBC 23082246 : bufHdr = GetBufferDescriptor(buffer - 1);
3311 :
3312 : /* pinned, so OK to read tag without spinlock */
228 rhaas 3313 GNC 23082246 : *rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
3314 23082246 : *forknum = BufTagGetForkNum(&bufHdr->tag);
5354 heikki.linnakangas 3315 GIC 23082246 : *blknum = bufHdr->tag.blockNum;
6929 tgl 3316 23082246 : }
3317 :
3318 : /*
3319 : * FlushBuffer
6927 tgl 3320 ECB : * Physically write out a shared buffer.
3321 : *
6927 tgl 3322 EUB : * NOTE: this actually just passes the buffer contents to the kernel; the
3323 : * real write to disk won't happen until the kernel feels like it. This
3324 : * is okay from our point of view since we can redo the changes from WAL.
3325 : * However, we will need to force the changes to disk via fsync before
3326 : * we can checkpoint WAL.
3327 : *
3328 : * The caller must hold a pin on the buffer and have share-locked the
3329 : * buffer contents. (Note: a share-lock does not prevent updates of
3330 : * hint bits in the buffer, so the page could change while the write
3331 : * is in progress, but we assume that that will not invalidate the data
3332 : * written.)
3333 : *
6927 tgl 3334 ECB : * If the caller has an smgr reference for the buffer's relation, pass it
3335 : * as the second parameter. If not, pass NULL.
9770 scrappy 3336 : */
3337 : static void
59 andres 3338 GNC 725820 : FlushBuffer(BufferDesc *buf, SMgrRelation reln, IOObject io_object,
3339 : IOContext io_context)
9770 scrappy 3340 ECB : {
3341 : XLogRecPtr recptr;
3800 heikki.linnakangas 3342 : ErrorContextCallback errcallback;
3343 : instr_time io_start;
3344 : Block bufBlock;
3345 : char *bufToWrite;
2555 andres 3346 : uint32 buf_state;
3347 :
3348 : /*
3349 : * Try to start an I/O operation. If StartBufferIO returns false, then
3350 : * someone else flushed the buffer before we could, so we need not do
3351 : * anything.
3352 : */
6610 tgl 3353 GIC 725820 : if (!StartBufferIO(buf, false))
6610 tgl 3354 LBC 0 : return;
8231 tgl 3355 ECB :
3356 : /* Setup error traceback support for ereport() */
3800 heikki.linnakangas 3357 GIC 725820 : errcallback.callback = shared_buffer_write_error_callback;
3800 heikki.linnakangas 3358 CBC 725820 : errcallback.arg = (void *) buf;
3800 heikki.linnakangas 3359 GIC 725820 : errcallback.previous = error_context_stack;
3360 725820 : error_context_stack = &errcallback;
3361 :
3362 : /* Find smgr relation for buffer */
6927 tgl 3363 725820 : if (reln == NULL)
228 rhaas 3364 GNC 722320 : reln = smgropen(BufTagGetRelFileLocator(&buf->tag), InvalidBackendId);
3365 :
3366 : TRACE_POSTGRESQL_BUFFER_FLUSH_START(BufTagGetForkNum(&buf->tag),
3367 : buf->tag.blockNum,
3368 : reln->smgr_rlocator.locator.spcOid,
3369 : reln->smgr_rlocator.locator.dbOid,
3370 : reln->smgr_rlocator.locator.relNumber);
3371 :
2555 andres 3372 GIC 725820 : buf_state = LockBufHdr(buf);
3373 :
3374 : /*
3375 : * Run PageGetLSN while holding header lock, since we don't have the
3376 : * buffer locked exclusively in all cases.
3377 : */
3670 simon 3378 CBC 725820 : recptr = BufferGetLSN(buf);
3670 simon 3379 ECB :
3380 : /* To check if block content changes while flushing. - vadim 01/17/97 */
2555 andres 3381 GIC 725820 : buf_state &= ~BM_JUST_DIRTIED;
3382 725820 : UnlockBufHdr(buf, buf_state);
3383 :
3384 : /*
3385 : * Force XLOG flush up to buffer's LSN. This implements the basic WAL
3386 : * rule that log updates must hit disk before any of the data-file changes
3387 : * they describe do.
3388 : *
3389 : * However, this rule does not apply to unlogged relations, which will be
3390 : * lost after a crash anyway. Most unlogged relation pages do not bear
3391 : * LSNs since we never emit WAL records for them, and therefore flushing
3392 : * up through the buffer LSN would be useless, but harmless. However,
3602 bruce 3393 ECB : * GiST indexes use LSNs internally to track page-splits, and therefore
3394 : * unlogged GiST pages bear "fake" LSNs generated by
3395 : * GetFakeLSNForUnloggedRel. It is unlikely but possible that the fake
3396 : * LSN counter could advance past the WAL insertion point; and if it did
3397 : * happen, attempting to flush WAL through that location would fail, with
3398 : * disastrous system-wide consequences. To make sure that can't happen,
3399 : * skip the flush if the buffer isn't permanent.
6927 tgl 3400 : */
2555 andres 3401 CBC 725820 : if (buf_state & BM_PERMANENT)
3709 heikki.linnakangas 3402 723799 : XLogFlush(recptr);
3403 :
3404 : /*
3405 : * Now it's safe to write buffer to disk. Note that no one else should
3406 : * have been able to write it while we were busy with log flushing because
3407 : * only one process at a time can set the BM_IO_IN_PROGRESS bit.
3408 : */
3670 simon 3409 GIC 725820 : bufBlock = BufHdrGetBlock(buf);
3670 simon 3410 ECB :
3411 : /*
3587 tgl 3412 : * Update page checksum if desired. Since we have only shared lock on the
3413 : * buffer, other processes might be updating hint bits in it, so we must
3414 : * copy the page to private storage if we do checksumming.
3415 : */
3670 simon 3416 GIC 725820 : bufToWrite = PageSetChecksumCopy((Page) bufBlock, buf->tag.blockNum);
3417 :
2 andres 3418 GNC 725820 : io_start = pgstat_prepare_io_time();
3419 :
3420 : /*
3670 simon 3421 ECB : * bufToWrite is either the shared buffer or a copy, as appropriate.
3422 : */
6998 tgl 3423 GIC 725820 : smgrwrite(reln,
228 rhaas 3424 GNC 725820 : BufTagGetForkNum(&buf->tag),
3425 : buf->tag.blockNum,
3426 : bufToWrite,
3427 : false);
3428 :
3429 : /*
3430 : * When a strategy is in use, only flushes of dirty buffers already in the
3431 : * strategy ring are counted as strategy writes (IOCONTEXT
3432 : * [BULKREAD|BULKWRITE|VACUUM] IOOP_WRITE) for the purpose of IO
3433 : * statistics tracking.
3434 : *
3435 : * If a shared buffer initially added to the ring must be flushed before
3436 : * being used, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE.
3437 : *
3438 : * If a shared buffer which was added to the ring later because the
3439 : * current strategy buffer is pinned or in use or because all strategy
3440 : * buffers were dirty and rejected (for BAS_BULKREAD operations only)
3441 : * requires flushing, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE
3442 : * (from_ring will be false).
3443 : *
3444 : * When a strategy is not in use, the write can only be a "regular" write
3445 : * of a dirty shared buffer (IOCONTEXT_NORMAL IOOP_WRITE).
3446 : */
2 andres 3447 725820 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
3448 : IOOP_WRITE, io_start, 1);
4030 rhaas 3449 ECB :
4863 rhaas 3450 GIC 725820 : pgBufferUsage.shared_blks_written++;
3451 :
3452 : /*
3453 : * Mark the buffer as clean (unless BM_JUST_DIRTIED has become set) and
3454 : * end the BM_IO_IN_PROGRESS state.
3455 : */
6610 tgl 3456 725820 : TerminateBufferIO(buf, true, 0);
3457 :
3458 : TRACE_POSTGRESQL_BUFFER_FLUSH_DONE(BufTagGetForkNum(&buf->tag),
5131 tgl 3459 ECB : buf->tag.blockNum,
3460 : reln->smgr_rlocator.locator.spcOid,
3461 : reln->smgr_rlocator.locator.dbOid,
3462 : reln->smgr_rlocator.locator.relNumber);
3463 :
3464 : /* Pop the error context stack */
3800 heikki.linnakangas 3465 GIC 725820 : error_context_stack = errcallback.previous;
3466 : }
3467 :
3468 : /*
3469 : * RelationGetNumberOfBlocksInFork
3470 : * Determines the current number of pages in the specified relation fork.
3471 : *
1423 andres 3472 ECB : * Note that the accuracy of the result will depend on the details of the
3473 : * relation's storage. For builtin AMs it'll be accurate, but for external AMs
3474 : * it might not be.
3475 : */
3476 : BlockNumber
4484 rhaas 3477 GIC 2074269 : RelationGetNumberOfBlocksInFork(Relation relation, ForkNumber forkNum)
3478 : {
492 peter 3479 2074269 : if (RELKIND_HAS_TABLE_AM(relation->rd_rel->relkind))
1423 andres 3480 ECB : {
3481 : /*
3482 : * Not every table AM uses BLCKSZ wide fixed size blocks. Therefore
3483 : * tableam returns the size in bytes - but for the purpose of this
3484 : * routine, we want the number of blocks. Therefore divide, rounding
3485 : * up.
3486 : */
3487 : uint64 szbytes;
3488 :
492 peter 3489 GIC 1614501 : szbytes = table_relation_size(relation, forkNum);
3490 :
492 peter 3491 CBC 1614483 : return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
3492 : }
3493 459768 : else if (RELKIND_HAS_STORAGE(relation->rd_rel->relkind))
492 peter 3494 ECB : {
332 tgl 3495 CBC 459768 : return smgrnblocks(RelationGetSmgr(relation), forkNum);
3496 : }
3497 : else
492 peter 3498 LBC 0 : Assert(false);
3499 :
1423 andres 3500 ECB : return 0; /* keep compiler quiet */
3501 : }
3502 :
4181 rhaas 3503 : /*
3504 : * BufferIsPermanent
3505 : * Determines whether a buffer will potentially still be around after
3506 : * a crash. Caller must hold a buffer pin.
3507 : */
3508 : bool
4181 rhaas 3509 GIC 15775629 : BufferIsPermanent(Buffer buffer)
4181 rhaas 3510 ECB : {
3511 : BufferDesc *bufHdr;
3512 :
3513 : /* Local buffers are used only for temp relations. */
4181 rhaas 3514 GIC 15775629 : if (BufferIsLocal(buffer))
4181 rhaas 3515 GBC 672453 : return false;
4181 rhaas 3516 EUB :
3517 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4181 rhaas 3518 GIC 15103176 : Assert(BufferIsValid(buffer));
4181 rhaas 3519 CBC 15103176 : Assert(BufferIsPinned(buffer));
4181 rhaas 3520 ECB :
3521 : /*
3522 : * BM_PERMANENT can't be changed while we hold a pin on the buffer, so we
3955 bruce 3523 : * need not bother with the buffer header spinlock. Even if someone else
3524 : * changes the buffer header state while we're doing this, the state is
3525 : * changed atomically, so we'll read the old value or the new value, but
3526 : * not random garbage.
3527 : */
2992 andres 3528 GIC 15103176 : bufHdr = GetBufferDescriptor(buffer - 1);
2555 3529 15103176 : return (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT) != 0;
3530 : }
3531 :
3532 : /*
3533 : * BufferGetLSNAtomic
3534 : * Retrieves the LSN of the buffer atomically using a buffer header lock.
3535 : * This is necessary for some callers who may not have an exclusive lock
3536 : * on the buffer.
3537 : */
3538 : XLogRecPtr
3670 simon 3539 10063084 : BufferGetLSNAtomic(Buffer buffer)
3540 : {
2701 rhaas 3541 10063084 : BufferDesc *bufHdr = GetBufferDescriptor(buffer - 1);
2545 kgrittn 3542 CBC 10063084 : char *page = BufferGetPage(buffer);
3602 bruce 3543 ECB : XLogRecPtr lsn;
2555 andres 3544 : uint32 buf_state;
3545 :
3654 simon 3546 : /*
3547 : * If we don't need locking for correctness, fastpath out.
3548 : */
2844 heikki.linnakangas 3549 GIC 10063084 : if (!XLogHintBitIsNeeded() || BufferIsLocal(buffer))
3670 simon 3550 8369376 : return PageGetLSN(page);
3551 :
3552 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
3553 1693708 : Assert(BufferIsValid(buffer));
3554 1693708 : Assert(BufferIsPinned(buffer));
3555 :
2555 andres 3556 1693708 : buf_state = LockBufHdr(bufHdr);
3670 simon 3557 1693708 : lsn = PageGetLSN(page);
2555 andres 3558 CBC 1693708 : UnlockBufHdr(bufHdr, buf_state);
3559 :
3670 simon 3560 GIC 1693708 : return lsn;
3561 : }
3562 :
3563 : /* ---------------------------------------------------------------------
3564 : * DropRelationBuffers
3565 : *
3566 : * This function removes from the buffer pool all the pages of the
3567 : * specified relation forks that have block numbers >= firstDelBlock.
3568 : * (In particular, with firstDelBlock = 0, all pages are removed.)
3569 : * Dirty pages are simply dropped, without bothering to write them
3570 : * out first. Therefore, this is NOT rollback-able, and so should be
3571 : * used only with extreme caution!
3572 : *
3573 : * Currently, this is called only from smgr.c when the underlying file
3574 : * is about to be deleted or truncated (firstDelBlock is needed for
3575 : * the truncation case). The data in the affected pages would therefore
3576 : * be deleted momentarily anyway, and there is no point in writing it.
3577 : * It is the responsibility of higher-level code to ensure that the
6594 tgl 3578 ECB : * deletion or truncation does not lose any data that could be needed
3579 : * later. It is also the responsibility of higher-level code to ensure
3580 : * that no other process could be trying to load more pages of the
3581 : * relation into buffers.
3582 : * --------------------------------------------------------------------
3583 : */
3584 : void
271 rhaas 3585 GNC 494 : DropRelationBuffers(SMgrRelation smgr_reln, ForkNumber *forkNum,
3586 : int nforks, BlockNumber *firstDelBlock)
3587 : {
3588 : int i;
3589 : int j;
3590 : RelFileLocatorBackend rlocator;
3591 : BlockNumber nForkBlock[MAX_FORKNUM];
697 tgl 3592 GIC 494 : uint64 nBlocksToInvalidate = 0;
3593 :
277 rhaas 3594 GNC 494 : rlocator = smgr_reln->smgr_rlocator;
3595 :
3958 tgl 3596 ECB : /* If it's a local relation, it's localbuf.c's problem. */
277 rhaas 3597 GNC 494 : if (RelFileLocatorBackendIsTemp(rlocator))
9345 bruce 3598 ECB : {
277 rhaas 3599 GNC 329 : if (rlocator.backend == MyBackendId)
3600 : {
1293 fujii 3601 CBC 675 : for (j = 0; j < nforks; j++)
271 rhaas 3602 GNC 346 : DropRelationLocalBuffers(rlocator.locator, forkNum[j],
3603 346 : firstDelBlock[j]);
3604 : }
7551 tgl 3605 GIC 363 : return;
9770 scrappy 3606 ECB : }
9345 bruce 3607 :
3608 : /*
3609 : * To remove all the pages of the specified relation forks from the buffer
817 akapila 3610 : * pool, we need to scan the entire buffer pool but we can optimize it by
3611 : * finding the buffers from BufMapping table provided we know the exact
3612 : * size of each fork of the relation. The exact size is required to ensure
3613 : * that we don't leave any buffer for the relation being dropped as
3614 : * otherwise the background writer or checkpointer can lead to a PANIC
3615 : * error while flushing buffers corresponding to files that don't exist.
3616 : *
3617 : * To know the exact size, we rely on the size cached for each fork by us
3618 : * during recovery which limits the optimization to recovery and on
3619 : * standbys but we can easily extend it once we have shared cache for
3620 : * relation size.
3621 : *
3622 : * In recovery, we cache the value returned by the first lseek(SEEK_END)
3623 : * and the future writes keeps the cached value up-to-date. See
3624 : * smgrextend. It is possible that the value of the first lseek is smaller
3625 : * than the actual number of existing blocks in the file due to buggy
3626 : * Linux kernels that might not have accounted for the recent write. But
3627 : * that should be fine because there must not be any buffers after that
3628 : * file size.
3629 : */
817 akapila 3630 CBC 234 : for (i = 0; i < nforks; i++)
3631 : {
817 akapila 3632 ECB : /* Get the number of blocks for a relation's fork */
817 akapila 3633 GIC 195 : nForkBlock[i] = smgrnblocks_cached(smgr_reln, forkNum[i]);
817 akapila 3634 ECB :
817 akapila 3635 GIC 195 : if (nForkBlock[i] == InvalidBlockNumber)
3636 : {
3637 126 : nBlocksToInvalidate = InvalidBlockNumber;
3638 126 : break;
3639 : }
3640 :
3641 : /* calculate the number of blocks to be invalidated */
3642 69 : nBlocksToInvalidate += (nForkBlock[i] - firstDelBlock[i]);
3643 : }
3644 :
817 akapila 3645 ECB : /*
3646 : * We apply the optimization iff the total number of blocks to invalidate
3647 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
3648 : */
817 akapila 3649 CBC 165 : if (BlockNumberIsValid(nBlocksToInvalidate) &&
817 akapila 3650 GIC 39 : nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
817 akapila 3651 ECB : {
817 akapila 3652 CBC 91 : for (j = 0; j < nforks; j++)
271 rhaas 3653 GNC 57 : FindAndDropRelationBuffers(rlocator.locator, forkNum[j],
3654 57 : nForkBlock[j], firstDelBlock[j]);
817 akapila 3655 GIC 34 : return;
3656 : }
3657 :
6610 tgl 3658 1658755 : for (i = 0; i < NBuffers; i++)
3659 : {
2701 rhaas 3660 1658624 : BufferDesc *bufHdr = GetBufferDescriptor(i);
3661 : uint32 buf_state;
6352 tgl 3662 ECB :
3663 : /*
3664 : * We can make this a tad faster by prechecking the buffer tag before
3665 : * we attempt to lock the buffer; this saves a lot of lock
3958 3666 : * acquisitions in typical cases. It should be safe because the
3667 : * caller must have AccessExclusiveLock on the relation, or some other
3668 : * reason to be certain that no one is loading new pages of the rel
3669 : * into the buffer pool. (Otherwise we might well miss such pages
3670 : * entirely.) Therefore, while the tag might be changing while we
3671 : * look at it, it can't be changing *to* a value we care about, only
3672 : * *away* from such a value. So false negatives are impossible, and
3673 : * false positives are safe because we'll recheck after getting the
3674 : * buffer lock.
3675 : *
3676 : * We could check forkNum and blockNum as well as the rlocator, but
3677 : * the incremental win from doing so seems small.
3678 : */
228 rhaas 3679 GNC 1658624 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator))
3958 tgl 3680 CBC 1657013 : continue;
3681 :
2555 andres 3682 GIC 1611 : buf_state = LockBufHdr(bufHdr);
3683 :
1293 fujii 3684 3479 : for (j = 0; j < nforks; j++)
3685 : {
228 rhaas 3686 GNC 2648 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator) &&
3687 2648 : BufTagGetForkNum(&bufHdr->tag) == forkNum[j] &&
1293 fujii 3688 GIC 1562 : bufHdr->tag.blockNum >= firstDelBlock[j])
1293 fujii 3689 ECB : {
1060 tgl 3690 GIC 780 : InvalidateBuffer(bufHdr); /* releases spinlock */
1293 fujii 3691 CBC 780 : break;
3692 : }
3693 : }
3694 1611 : if (j >= nforks)
2555 andres 3695 831 : UnlockBufHdr(bufHdr, buf_state);
3696 : }
3697 : }
3698 :
3699 : /* ---------------------------------------------------------------------
3700 : * DropRelationsAllBuffers
3701 : *
3702 : * This function removes from the buffer pool all the pages of all
3703 : * forks of the specified relations. It's equivalent to calling
3704 : * DropRelationBuffers once per fork per relation with firstDelBlock = 0.
3705 : * --------------------------------------------------------------------
3706 : */
3958 tgl 3707 ECB : void
271 rhaas 3708 GNC 10663 : DropRelationsAllBuffers(SMgrRelation *smgr_reln, int nlocators)
3709 : {
3710 : int i;
816 akapila 3711 CBC 10663 : int n = 0;
3712 : SMgrRelation *rels;
816 akapila 3713 ECB : BlockNumber (*block)[MAX_FORKNUM + 1];
697 tgl 3714 GIC 10663 : uint64 nBlocksToInvalidate = 0;
3715 : RelFileLocator *locators;
816 akapila 3716 10663 : bool cached = true;
3734 alvherre 3717 EUB : bool use_bsearch;
3718 :
277 rhaas 3719 GNC 10663 : if (nlocators == 0)
3734 alvherre 3720 UIC 0 : return;
3721 :
277 rhaas 3722 GNC 10663 : rels = palloc(sizeof(SMgrRelation) * nlocators); /* non-local relations */
3958 tgl 3723 ECB :
3724 : /* If it's a local relation, it's localbuf.c's problem. */
277 rhaas 3725 GNC 48045 : for (i = 0; i < nlocators; i++)
3726 : {
3727 37382 : if (RelFileLocatorBackendIsTemp(smgr_reln[i]->smgr_rlocator))
3734 alvherre 3728 EUB : {
277 rhaas 3729 GNC 2808 : if (smgr_reln[i]->smgr_rlocator.backend == MyBackendId)
271 3730 2808 : DropRelationAllLocalBuffers(smgr_reln[i]->smgr_rlocator.locator);
3734 alvherre 3731 EUB : }
3732 : else
816 akapila 3733 GIC 34574 : rels[n++] = smgr_reln[i];
3734 : }
3734 alvherre 3735 ECB :
3736 : /*
3602 bruce 3737 : * If there are no non-local relations, then we're done. Release the
3738 : * memory and return.
3739 : */
3734 alvherre 3740 GIC 10663 : if (n == 0)
3741 : {
816 akapila 3742 700 : pfree(rels);
3958 tgl 3743 GBC 700 : return;
3744 : }
3745 :
3746 : /*
3747 : * This is used to remember the number of blocks for all the relations
3748 : * forks.
3749 : */
3750 : block = (BlockNumber (*)[MAX_FORKNUM + 1])
816 akapila 3751 9963 : palloc(sizeof(BlockNumber) * n * (MAX_FORKNUM + 1));
816 akapila 3752 EUB :
3753 : /*
3754 : * We can avoid scanning the entire buffer pool if we know the exact size
3755 : * of each of the given relation forks. See DropRelationBuffers.
3756 : */
816 akapila 3757 GIC 20986 : for (i = 0; i < n && cached; i++)
3758 : {
228 drowley 3759 GNC 18021 : for (int j = 0; j <= MAX_FORKNUM; j++)
816 akapila 3760 EUB : {
3761 : /* Get the number of blocks for a relation's fork. */
816 akapila 3762 GBC 16281 : block[i][j] = smgrnblocks_cached(rels[i], j);
3763 :
3764 : /* We need to only consider the relation forks that exists. */
816 akapila 3765 GIC 16281 : if (block[i][j] == InvalidBlockNumber)
816 akapila 3766 EUB : {
816 akapila 3767 GIC 14398 : if (!smgrexists(rels[i], j))
816 akapila 3768 GBC 5115 : continue;
3769 9283 : cached = false;
816 akapila 3770 GIC 9283 : break;
3771 : }
3772 :
3773 : /* calculate the total number of blocks to be invalidated */
3774 1883 : nBlocksToInvalidate += block[i][j];
816 akapila 3775 EUB : }
3776 : }
3777 :
3778 : /*
3779 : * We apply the optimization iff the total number of blocks to invalidate
3780 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
3781 : */
816 akapila 3782 GIC 9963 : if (cached && nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
3783 : {
3784 1107 : for (i = 0; i < n; i++)
3785 : {
228 drowley 3786 GNC 3050 : for (int j = 0; j <= MAX_FORKNUM; j++)
816 akapila 3787 ECB : {
3788 : /* ignore relation forks that doesn't exist */
816 akapila 3789 CBC 2440 : if (!BlockNumberIsValid(block[i][j]))
3790 1829 : continue;
3791 :
3792 : /* drop all the buffers for a particular relation fork */
271 rhaas 3793 GNC 611 : FindAndDropRelationBuffers(rels[i]->smgr_rlocator.locator,
3794 611 : j, block[i][j], 0);
3795 : }
3796 : }
3797 :
816 akapila 3798 GIC 497 : pfree(block);
3799 497 : pfree(rels);
3800 497 : return;
816 akapila 3801 ECB : }
3802 :
816 akapila 3803 GIC 9466 : pfree(block);
277 rhaas 3804 GNC 9466 : locators = palloc(sizeof(RelFileLocator) * n); /* non-local relations */
816 akapila 3805 CBC 43430 : for (i = 0; i < n; i++)
277 rhaas 3806 GNC 33964 : locators[i] = rels[i]->smgr_rlocator.locator;
816 akapila 3807 ECB :
3734 alvherre 3808 : /*
3809 : * For low number of relations to drop just use a simple walk through, to
3602 bruce 3810 : * save the bsearch overhead. The threshold to use is rather a guess than
3811 : * an exactly determined value, as it depends on many factors (CPU and RAM
3812 : * speeds, amount of shared buffers etc.).
3734 alvherre 3813 : */
1100 noah 3814 GIC 9466 : use_bsearch = n > RELS_BSEARCH_THRESHOLD;
3815 :
3816 : /* sort the list of rlocators if necessary */
3734 alvherre 3817 9466 : if (use_bsearch)
277 rhaas 3818 GNC 165 : pg_qsort(locators, n, sizeof(RelFileLocator), rlocator_comparator);
3819 :
3958 tgl 3820 GIC 102040826 : for (i = 0; i < NBuffers; i++)
3821 : {
277 rhaas 3822 GNC 102031360 : RelFileLocator *rlocator = NULL;
2701 rhaas 3823 GIC 102031360 : BufferDesc *bufHdr = GetBufferDescriptor(i);
3824 : uint32 buf_state;
3825 :
3826 : /*
3827 : * As in DropRelationBuffers, an unlocked precheck should be safe and
3828 : * saves some cycles.
3829 : */
3734 alvherre 3830 ECB :
3734 alvherre 3831 GBC 102031360 : if (!use_bsearch)
3832 : {
3602 bruce 3833 ECB : int j;
3834 :
3734 alvherre 3835 GIC 413532505 : for (j = 0; j < n; j++)
3734 alvherre 3836 ECB : {
228 rhaas 3837 GNC 313271388 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &locators[j]))
3734 alvherre 3838 ECB : {
277 rhaas 3839 GNC 74755 : rlocator = &locators[j];
3734 alvherre 3840 GIC 74755 : break;
3841 : }
3842 : }
3843 : }
3844 : else
3845 : {
3846 : RelFileLocator locator;
3847 :
228 rhaas 3848 GNC 1695488 : locator = BufTagGetRelFileLocator(&bufHdr->tag);
3849 1695488 : rlocator = bsearch((const void *) &(locator),
3850 : locators, n, sizeof(RelFileLocator),
3851 : rlocator_comparator);
3852 : }
3853 :
3854 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
277 3855 102031360 : if (rlocator == NULL)
3958 tgl 3856 GIC 101954940 : continue;
3857 :
2555 andres 3858 76420 : buf_state = LockBufHdr(bufHdr);
228 rhaas 3859 GNC 76420 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, rlocator))
3958 tgl 3860 GIC 76420 : InvalidateBuffer(bufHdr); /* releases spinlock */
3861 : else
2555 andres 3862 UIC 0 : UnlockBufHdr(bufHdr, buf_state);
3863 : }
3734 alvherre 3864 ECB :
277 rhaas 3865 GNC 9466 : pfree(locators);
816 akapila 3866 GIC 9466 : pfree(rels);
3867 : }
3868 :
3869 : /* ---------------------------------------------------------------------
3870 : * FindAndDropRelationBuffers
3871 : *
3872 : * This function performs look up in BufMapping table and removes from the
3873 : * buffer pool all the pages of the specified relation fork that has block
3874 : * number >= firstDelBlock. (In particular, with firstDelBlock = 0, all
3875 : * pages are removed.)
3876 : * --------------------------------------------------------------------
3877 : */
3878 : static void
271 rhaas 3879 GNC 668 : FindAndDropRelationBuffers(RelFileLocator rlocator, ForkNumber forkNum,
3880 : BlockNumber nForkBlock,
3881 : BlockNumber firstDelBlock)
3882 : {
817 akapila 3883 ECB : BlockNumber curBlock;
3884 :
817 akapila 3885 CBC 1612 : for (curBlock = firstDelBlock; curBlock < nForkBlock; curBlock++)
817 akapila 3886 ECB : {
3887 : uint32 bufHash; /* hash value for tag */
3888 : BufferTag bufTag; /* identity of requested block */
3889 : LWLock *bufPartitionLock; /* buffer partition lock for it */
3890 : int buf_id;
3891 : BufferDesc *bufHdr;
3892 : uint32 buf_state;
3893 :
3894 : /* create a tag so we can lookup the buffer */
256 rhaas 3895 GNC 944 : InitBufferTag(&bufTag, &rlocator, forkNum, curBlock);
3896 :
3897 : /* determine its hash code and partition lock ID */
817 akapila 3898 CBC 944 : bufHash = BufTableHashCode(&bufTag);
817 akapila 3899 GIC 944 : bufPartitionLock = BufMappingPartitionLock(bufHash);
3900 :
3901 : /* Check that it is in the buffer pool. If not, do nothing. */
3902 944 : LWLockAcquire(bufPartitionLock, LW_SHARED);
3903 944 : buf_id = BufTableLookup(&bufTag, bufHash);
817 akapila 3904 CBC 944 : LWLockRelease(bufPartitionLock);
3905 :
817 akapila 3906 GIC 944 : if (buf_id < 0)
817 akapila 3907 CBC 120 : continue;
817 akapila 3908 ECB :
817 akapila 3909 GIC 824 : bufHdr = GetBufferDescriptor(buf_id);
3910 :
3911 : /*
3912 : * We need to lock the buffer header and recheck if the buffer is
3913 : * still associated with the same block because the buffer could be
3914 : * evicted by some other backend loading blocks for a different
3915 : * relation after we release lock on the BufMapping table.
3916 : */
3917 824 : buf_state = LockBufHdr(bufHdr);
3918 :
228 rhaas 3919 GNC 1648 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator) &&
3920 824 : BufTagGetForkNum(&bufHdr->tag) == forkNum &&
817 akapila 3921 GIC 824 : bufHdr->tag.blockNum >= firstDelBlock)
3922 824 : InvalidateBuffer(bufHdr); /* releases spinlock */
3923 : else
817 akapila 3924 UIC 0 : UnlockBufHdr(bufHdr, buf_state);
3925 : }
817 akapila 3926 GIC 668 : }
817 akapila 3927 ECB :
9770 scrappy 3928 : /* ---------------------------------------------------------------------
3929 : * DropDatabaseBuffers
3930 : *
3931 : * This function removes all the buffers in the buffer cache for a
3932 : * particular database. Dirty pages are simply dropped, without
3933 : * bothering to write them out first. This is used when we destroy a
3934 : * database, to avoid trying to flush data to disk when the directory
3260 bruce 3935 : * tree no longer exists. Implementation is pretty similar to
3936 : * DropRelationBuffers() which is for destroying just one relation.
3937 : * --------------------------------------------------------------------
3938 : */
3939 : void
6220 tgl 3940 GIC 34 : DropDatabaseBuffers(Oid dbid)
3941 : {
9173 bruce 3942 ECB : int i;
3943 :
6610 tgl 3944 : /*
3945 : * We needn't consider local buffers, since by assumption the target
3946 : * database isn't our own.
3947 : */
3948 :
6610 tgl 3949 CBC 118178 : for (i = 0; i < NBuffers; i++)
9345 bruce 3950 ECB : {
2701 rhaas 3951 GIC 118144 : BufferDesc *bufHdr = GetBufferDescriptor(i);
3952 : uint32 buf_state;
3953 :
3954 : /*
3955 : * As in DropRelationBuffers, an unlocked precheck should be safe and
3956 : * saves some cycles.
3957 : */
228 rhaas 3958 GNC 118144 : if (bufHdr->tag.dbOid != dbid)
3958 tgl 3959 GIC 112397 : continue;
3960 :
2555 andres 3961 5747 : buf_state = LockBufHdr(bufHdr);
228 rhaas 3962 GNC 5747 : if (bufHdr->tag.dbOid == dbid)
6385 bruce 3963 GIC 5747 : InvalidateBuffer(bufHdr); /* releases spinlock */
3964 : else
2555 andres 3965 UIC 0 : UnlockBufHdr(bufHdr, buf_state);
3966 : }
9770 scrappy 3967 GIC 34 : }
3968 :
3969 : /* -----------------------------------------------------------------
3970 : * PrintBufferDescs
3971 : *
3972 : * this function prints all the buffer descriptors, for debugging
9345 bruce 3973 ECB : * use only.
3974 : * -----------------------------------------------------------------
3975 : */
7586 3976 : #ifdef NOT_USED
3977 : void
3978 : PrintBufferDescs(void)
3979 : {
3980 : int i;
3981 :
2992 andres 3982 : for (i = 0; i < NBuffers; ++i)
3983 : {
3984 : BufferDesc *buf = GetBufferDescriptor(i);
3985 : Buffer b = BufferDescriptorGetBuffer(buf);
3986 :
3987 : /* theoretically we should lock the bufhdr here */
3988 : elog(LOG,
3989 : "[%02d] (freeNext=%d, rel=%s, "
3990 : "blockNum=%u, flags=0x%x, refcount=%u %d)",
6610 tgl 3991 : i, buf->freeNext,
3992 : relpathbackend(BufTagGetRelFileLocator(&buf->tag),
3993 : InvalidBackendId, BufTagGetForkNum(&buf->tag)),
3994 : buf->tag.blockNum, buf->flags,
3995 : buf->refcount, GetPrivateRefCount(b));
3996 : }
3997 : }
3998 : #endif
3999 :
4000 : #ifdef NOT_USED
4001 : void
4002 : PrintPinnedBufs(void)
4003 : {
9344 bruce 4004 : int i;
4005 :
2992 andres 4006 : for (i = 0; i < NBuffers; ++i)
4007 : {
4008 : BufferDesc *buf = GetBufferDescriptor(i);
4009 : Buffer b = BufferDescriptorGetBuffer(buf);
4010 :
4011 : if (GetPrivateRefCount(b) > 0)
4012 : {
4013 : /* theoretically we should lock the bufhdr here */
4014 : elog(LOG,
4015 : "[%02d] (freeNext=%d, rel=%s, "
6926 neilc 4016 : "blockNum=%u, flags=0x%x, refcount=%u %d)",
4017 : i, buf->freeNext,
4018 : relpathperm(BufTagGetRelFileLocator(&buf->tag),
4019 : BufTagGetForkNum(&buf->tag)),
4020 : buf->tag.blockNum, buf->flags,
2797 andres 4021 : buf->refcount, GetPrivateRefCount(b));
4022 : }
9345 bruce 4023 : }
4024 : }
4025 : #endif
9770 scrappy 4026 EUB :
4027 : /* ---------------------------------------------------------------------
4028 : * FlushRelationBuffers
4029 : *
4030 : * This function writes all dirty pages of a relation out to disk
4031 : * (or more accurately, out to kernel disk buffers), ensuring that the
4032 : * kernel has an up-to-date view of the relation.
4033 : *
4034 : * Generally, the caller should be holding AccessExclusiveLock on the
4035 : * target relation to ensure that no other backend is busy dirtying
4036 : * more blocks of the relation; the effects can't be expected to last
6594 tgl 4037 ECB : * after the lock is released.
4038 : *
4039 : * XXX currently it sequentially searches the buffer pool, should be
4040 : * changed to more clever ways of searching. This routine is not
4041 : * used in any performance-critical code paths, so it's not worth
817 akapila 4042 : * adding additional overhead to normal paths to make it go faster.
9330 vadim4o 4043 : * --------------------------------------------------------------------
4044 : */
4045 : void
6594 tgl 4046 CBC 115 : FlushRelationBuffers(Relation rel)
9330 vadim4o 4047 ECB : {
4048 : int i;
4049 : BufferDesc *bufHdr;
4050 :
4500 rhaas 4051 GIC 115 : if (RelationUsesLocalBuffers(rel))
4052 : {
9330 vadim4o 4053 909 : for (i = 0; i < NLocBuffer; i++)
4054 : {
4055 : uint32 buf_state;
4056 : instr_time io_start;
2555 andres 4057 ECB :
2992 andres 4058 CBC 900 : bufHdr = GetLocalBufferDescriptor(i);
228 rhaas 4059 GNC 900 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
2555 andres 4060 GIC 300 : ((buf_state = pg_atomic_read_u32(&bufHdr->state)) &
4061 : (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
4062 : {
4063 : ErrorContextCallback errcallback;
4064 : Page localpage;
4065 :
3670 simon 4066 297 : localpage = (char *) LocalBufHdrGetBlock(bufHdr);
4067 :
6594 tgl 4068 ECB : /* Setup error traceback support for ereport() */
3800 heikki.linnakangas 4069 GIC 297 : errcallback.callback = local_buffer_write_error_callback;
3800 heikki.linnakangas 4070 CBC 297 : errcallback.arg = (void *) bufHdr;
4071 297 : errcallback.previous = error_context_stack;
3800 heikki.linnakangas 4072 GIC 297 : error_context_stack = &errcallback;
4073 :
3670 simon 4074 297 : PageSetChecksumInplace(localpage, bufHdr->tag.blockNum);
4075 :
2 andres 4076 GNC 297 : io_start = pgstat_prepare_io_time();
4077 :
636 tgl 4078 GIC 297 : smgrwrite(RelationGetSmgr(rel),
228 rhaas 4079 GNC 297 : BufTagGetForkNum(&bufHdr->tag),
6594 tgl 4080 ECB : bufHdr->tag.blockNum,
3670 simon 4081 : localpage,
4082 : false);
4083 :
2 andres 4084 GNC 297 : pgstat_count_io_op_time(IOOBJECT_TEMP_RELATION,
4085 : IOCONTEXT_NORMAL, IOOP_WRITE,
4086 : io_start, 1);
4087 :
2555 andres 4088 CBC 297 : buf_state &= ~(BM_DIRTY | BM_JUST_DIRTIED);
2375 4089 297 : pg_atomic_unlocked_write_u32(&bufHdr->state, buf_state);
4090 :
2 andres 4091 GNC 297 : pgBufferUsage.local_blks_written++;
4092 :
6594 tgl 4093 ECB : /* Pop the error context stack */
3800 heikki.linnakangas 4094 CBC 297 : error_context_stack = errcallback.previous;
9330 vadim4o 4095 ECB : }
4096 : }
7274 tgl 4097 :
6887 tgl 4098 GIC 9 : return;
4099 : }
4100 :
4101 : /* Make sure we can handle the pin inside the loop */
6840 4102 106 : ResourceOwnerEnlargeBuffers(CurrentResourceOwner);
4103 :
9330 vadim4o 4104 1249130 : for (i = 0; i < NBuffers; i++)
4105 : {
4106 : uint32 buf_state;
4107 :
2992 andres 4108 1249024 : bufHdr = GetBufferDescriptor(i);
4109 :
4110 : /*
4111 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4112 : * saves some cycles.
4113 : */
228 rhaas 4114 GNC 1249024 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator))
3958 tgl 4115 GIC 1248867 : continue;
4116 :
3002 andres 4117 157 : ReservePrivateRefCountEntry();
4118 :
2555 4119 157 : buf_state = LockBufHdr(bufHdr);
228 rhaas 4120 GNC 157 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
2555 andres 4121 GIC 157 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
9330 vadim4o 4122 ECB : {
6594 tgl 4123 GIC 133 : PinBuffer_Locked(bufHdr);
2672 rhaas 4124 133 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_SHARED);
59 andres 4125 GNC 133 : FlushBuffer(bufHdr, RelationGetSmgr(rel), IOOBJECT_RELATION, IOCONTEXT_NORMAL);
2672 rhaas 4126 GIC 133 : LWLockRelease(BufferDescriptorGetContentLock(bufHdr));
191 michael 4127 GNC 133 : UnpinBuffer(bufHdr);
4128 : }
6610 tgl 4129 ECB : else
2555 andres 4130 GIC 24 : UnlockBufHdr(bufHdr, buf_state);
9330 vadim4o 4131 ECB : }
4132 : }
4133 :
1100 noah 4134 : /* ---------------------------------------------------------------------
4135 : * FlushRelationsAllBuffers
4136 : *
4137 : * This function flushes out of the buffer pool all the pages of all
4138 : * forks of the specified smgr relations. It's equivalent to calling
4139 : * FlushRelationBuffers once per relation. The relations are assumed not
4140 : * to use local buffers.
4141 : * --------------------------------------------------------------------
4142 : */
4143 : void
1100 noah 4144 GIC 9 : FlushRelationsAllBuffers(SMgrRelation *smgrs, int nrels)
4145 : {
4146 : int i;
4147 : SMgrSortArray *srels;
4148 : bool use_bsearch;
4149 :
4150 9 : if (nrels == 0)
1100 noah 4151 UIC 0 : return;
4152 :
4153 : /* fill-in array for qsort */
1100 noah 4154 GIC 9 : srels = palloc(sizeof(SMgrSortArray) * nrels);
4155 :
4156 18 : for (i = 0; i < nrels; i++)
4157 : {
277 rhaas 4158 GNC 9 : Assert(!RelFileLocatorBackendIsTemp(smgrs[i]->smgr_rlocator));
4159 :
4160 9 : srels[i].rlocator = smgrs[i]->smgr_rlocator.locator;
1100 noah 4161 GIC 9 : srels[i].srel = smgrs[i];
4162 : }
4163 :
4164 : /*
4165 : * Save the bsearch overhead for low number of relations to sync. See
4166 : * DropRelationsAllBuffers for details.
1100 noah 4167 ECB : */
1100 noah 4168 GIC 9 : use_bsearch = nrels > RELS_BSEARCH_THRESHOLD;
4169 :
1100 noah 4170 ECB : /* sort the list of SMgrRelations if necessary */
1100 noah 4171 GIC 9 : if (use_bsearch)
277 rhaas 4172 UNC 0 : pg_qsort(srels, nrels, sizeof(SMgrSortArray), rlocator_comparator);
4173 :
1100 noah 4174 ECB : /* Make sure we can handle the pin inside the loop */
1100 noah 4175 CBC 9 : ResourceOwnerEnlargeBuffers(CurrentResourceOwner);
4176 :
1100 noah 4177 GIC 147465 : for (i = 0; i < NBuffers; i++)
4178 : {
1100 noah 4179 CBC 147456 : SMgrSortArray *srelent = NULL;
1100 noah 4180 GIC 147456 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4181 : uint32 buf_state;
4182 :
4183 : /*
4184 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4185 : * saves some cycles.
1100 noah 4186 ECB : */
4187 :
1100 noah 4188 GIC 147456 : if (!use_bsearch)
1100 noah 4189 ECB : {
4190 : int j;
4191 :
1100 noah 4192 CBC 291200 : for (j = 0; j < nrels; j++)
4193 : {
228 rhaas 4194 GNC 147456 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srels[j].rlocator))
1100 noah 4195 ECB : {
1100 noah 4196 GIC 3712 : srelent = &srels[j];
1100 noah 4197 CBC 3712 : break;
4198 : }
4199 : }
4200 : }
4201 : else
4202 : {
4203 : RelFileLocator rlocator;
4204 :
228 rhaas 4205 UNC 0 : rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
4206 0 : srelent = bsearch((const void *) &(rlocator),
4207 : srels, nrels, sizeof(SMgrSortArray),
4208 : rlocator_comparator);
4209 : }
4210 :
4211 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
1100 noah 4212 GIC 147456 : if (srelent == NULL)
4213 143744 : continue;
4214 :
4215 3712 : ReservePrivateRefCountEntry();
4216 :
4217 3712 : buf_state = LockBufHdr(bufHdr);
228 rhaas 4218 GNC 3712 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srelent->rlocator) &&
1100 noah 4219 CBC 3712 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
1100 noah 4220 ECB : {
1100 noah 4221 GIC 3367 : PinBuffer_Locked(bufHdr);
1100 noah 4222 CBC 3367 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_SHARED);
59 andres 4223 GNC 3367 : FlushBuffer(bufHdr, srelent->srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
1100 noah 4224 CBC 3367 : LWLockRelease(BufferDescriptorGetContentLock(bufHdr));
191 michael 4225 GNC 3367 : UnpinBuffer(bufHdr);
1100 noah 4226 ECB : }
4227 : else
1100 noah 4228 CBC 345 : UnlockBufHdr(bufHdr, buf_state);
4229 : }
1100 noah 4230 ECB :
1100 noah 4231 CBC 9 : pfree(srels);
4232 : }
4233 :
376 rhaas 4234 ECB : /* ---------------------------------------------------------------------
4235 : * RelationCopyStorageUsingBuffer
4236 : *
4237 : * Copy fork's data using bufmgr. Same as RelationCopyStorage but instead
4238 : * of using smgrread and smgrextend this will copy using bufmgr APIs.
4239 : *
4240 : * Refer comments atop CreateAndCopyRelationData() for details about
4241 : * 'permanent' parameter.
4242 : * --------------------------------------------------------------------
4243 : */
4244 : static void
240 rhaas 4245 GNC 54479 : RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
4246 : RelFileLocator dstlocator,
4247 : ForkNumber forkNum, bool permanent)
376 rhaas 4248 ECB : {
4249 : Buffer srcBuf;
4250 : Buffer dstBuf;
4251 : Page srcPage;
4252 : Page dstPage;
4253 : bool use_wal;
4254 : BlockNumber nblocks;
4255 : BlockNumber blkno;
4256 : PGIOAlignedBlock buf;
4257 : BufferAccessStrategy bstrategy_src;
4258 : BufferAccessStrategy bstrategy_dst;
4259 :
376 rhaas 4260 EUB : /*
4261 : * In general, we want to write WAL whenever wal_level > 'minimal', but we
332 tgl 4262 ECB : * can skip it when copying any fork of an unlogged relation other than
4263 : * the init fork.
4264 : */
376 rhaas 4265 CBC 54479 : use_wal = XLogIsNeeded() && (permanent || forkNum == INIT_FORKNUM);
4266 :
376 rhaas 4267 ECB : /* Get number of blocks in the source relation. */
240 rhaas 4268 GNC 54479 : nblocks = smgrnblocks(smgropen(srclocator, InvalidBackendId),
240 rhaas 4269 ECB : forkNum);
376 4270 :
4271 : /* Nothing to copy; just return. */
376 rhaas 4272 GIC 54479 : if (nblocks == 0)
376 rhaas 4273 CBC 9382 : return;
4274 :
4275 : /*
4276 : * Bulk extend the destination relation of the same size as the source
4277 : * relation before starting to copy block by block.
4278 : */
234 rhaas 4279 GIC 45097 : memset(buf.data, 0, BLCKSZ);
234 rhaas 4280 GNC 45097 : smgrextend(smgropen(dstlocator, InvalidBackendId), forkNum, nblocks - 1,
4281 : buf.data, true);
234 rhaas 4282 ECB :
376 4283 : /* This is a bulk operation, so use buffer access strategies. */
376 rhaas 4284 GIC 45097 : bstrategy_src = GetAccessStrategy(BAS_BULKREAD);
4285 45097 : bstrategy_dst = GetAccessStrategy(BAS_BULKWRITE);
4286 :
4287 : /* Iterate over each block of the source relation file. */
4288 217434 : for (blkno = 0; blkno < nblocks; blkno++)
4289 : {
4290 172337 : CHECK_FOR_INTERRUPTS();
376 rhaas 4291 ECB :
4292 : /* Read block from source relation. */
240 rhaas 4293 GNC 172337 : srcBuf = ReadBufferWithoutRelcache(srclocator, forkNum, blkno,
4294 : RBM_NORMAL, bstrategy_src,
4295 : permanent);
246 tgl 4296 GIC 172337 : LockBuffer(srcBuf, BUFFER_LOCK_SHARE);
376 rhaas 4297 CBC 172337 : srcPage = BufferGetPage(srcBuf);
4298 :
234 rhaas 4299 GNC 172337 : dstBuf = ReadBufferWithoutRelcache(dstlocator, forkNum, blkno,
4300 : RBM_ZERO_AND_LOCK, bstrategy_dst,
4301 : permanent);
246 tgl 4302 CBC 172337 : dstPage = BufferGetPage(dstBuf);
4303 :
376 rhaas 4304 GIC 172337 : START_CRIT_SECTION();
376 rhaas 4305 ECB :
4306 : /* Copy page data from the source to the destination. */
376 rhaas 4307 CBC 172337 : memcpy(dstPage, srcPage, BLCKSZ);
4308 172337 : MarkBufferDirty(dstBuf);
376 rhaas 4309 ECB :
4310 : /* WAL-log the copied page. */
376 rhaas 4311 GIC 172337 : if (use_wal)
4312 103988 : log_newpage_buffer(dstBuf, true);
4313 :
376 rhaas 4314 CBC 172337 : END_CRIT_SECTION();
4315 :
376 rhaas 4316 GIC 172337 : UnlockReleaseBuffer(dstBuf);
246 tgl 4317 172337 : UnlockReleaseBuffer(srcBuf);
4318 : }
4319 :
20 andres 4320 45097 : FreeAccessStrategy(bstrategy_src);
4321 45097 : FreeAccessStrategy(bstrategy_dst);
376 rhaas 4322 ECB : }
4323 :
4324 : /* ---------------------------------------------------------------------
4325 : * CreateAndCopyRelationData
4326 : *
4327 : * Create destination relation storage and copy all forks from the
4328 : * source relation to the destination.
4329 : *
4330 : * Pass permanent as true for permanent relations and false for
4331 : * unlogged relations. Currently this API is not supported for
4332 : * temporary relations.
4333 : * --------------------------------------------------------------------
4334 : */
4335 : void
277 rhaas 4336 GNC 40861 : CreateAndCopyRelationData(RelFileLocator src_rlocator,
4337 : RelFileLocator dst_rlocator, bool permanent)
376 rhaas 4338 ECB : {
4339 : RelFileLocatorBackend rlocator;
332 tgl 4340 : char relpersistence;
4341 :
4342 : /* Set the relpersistence. */
376 rhaas 4343 CBC 40861 : relpersistence = permanent ?
376 rhaas 4344 ECB : RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED;
4345 :
4346 : /*
4347 : * Create and copy all forks of the relation. During create database we
4348 : * have a separate cleanup mechanism which deletes complete database
4349 : * directory. Therefore, each individual relation doesn't need to be
4350 : * registered for cleanup.
4351 : */
277 rhaas 4352 GNC 40861 : RelationCreateStorage(dst_rlocator, relpersistence, false);
4353 :
376 rhaas 4354 ECB : /* copy main fork. */
240 rhaas 4355 GNC 40861 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, MAIN_FORKNUM,
4356 : permanent);
376 rhaas 4357 ECB :
4358 : /* copy those extra forks that exist */
376 rhaas 4359 GIC 40861 : for (ForkNumber forkNum = MAIN_FORKNUM + 1;
376 rhaas 4360 CBC 163444 : forkNum <= MAX_FORKNUM; forkNum++)
4361 : {
240 rhaas 4362 GNC 122583 : if (smgrexists(smgropen(src_rlocator, InvalidBackendId), forkNum))
376 rhaas 4363 ECB : {
240 rhaas 4364 GNC 13618 : smgrcreate(smgropen(dst_rlocator, InvalidBackendId), forkNum, false);
4365 :
4366 : /*
4367 : * WAL log creation if the relation is persistent, or this is the
4368 : * init fork of an unlogged relation.
4369 : */
376 rhaas 4370 GIC 13618 : if (permanent || forkNum == INIT_FORKNUM)
277 rhaas 4371 GNC 13618 : log_smgrcreate(&dst_rlocator, forkNum);
4372 :
4373 : /* Copy a fork's data, block by block. */
240 4374 13618 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, forkNum,
376 rhaas 4375 ECB : permanent);
4376 : }
4377 : }
4378 :
240 4379 : /* close source and destination smgr if exists. */
240 rhaas 4380 GNC 40861 : rlocator.backend = InvalidBackendId;
4381 :
4382 40861 : rlocator.locator = src_rlocator;
4383 40861 : smgrcloserellocator(rlocator);
4384 :
4385 40861 : rlocator.locator = dst_rlocator;
4386 40861 : smgrcloserellocator(rlocator);
376 rhaas 4387 GIC 40861 : }
376 rhaas 4388 ECB :
5764 tgl 4389 : /* ---------------------------------------------------------------------
4390 : * FlushDatabaseBuffers
4391 : *
4392 : * This function writes all dirty pages of a database out to disk
4393 : * (or more accurately, out to kernel disk buffers), ensuring that the
4394 : * kernel has an up-to-date view of the database.
4395 : *
4396 : * Generally, the caller should be holding an appropriate lock to ensure
4397 : * no other backend is active in the target database; otherwise more
4398 : * pages could get dirtied.
4399 : *
4400 : * Note we don't worry about flushing any pages of temporary relations.
4401 : * It's assumed these wouldn't be interesting.
5764 tgl 4402 EUB : * --------------------------------------------------------------------
4403 : */
4404 : void
5764 tgl 4405 CBC 3 : FlushDatabaseBuffers(Oid dbid)
5764 tgl 4406 ECB : {
4407 : int i;
4408 : BufferDesc *bufHdr;
4409 :
4410 : /* Make sure we can handle the pin inside the loop */
5764 tgl 4411 GIC 3 : ResourceOwnerEnlargeBuffers(CurrentResourceOwner);
4412 :
4413 387 : for (i = 0; i < NBuffers; i++)
4414 : {
4415 : uint32 buf_state;
4416 :
2992 andres 4417 384 : bufHdr = GetBufferDescriptor(i);
4418 :
3958 tgl 4419 ECB : /*
4420 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4421 : * saves some cycles.
4422 : */
228 rhaas 4423 GNC 384 : if (bufHdr->tag.dbOid != dbid)
3958 tgl 4424 GIC 265 : continue;
3958 tgl 4425 ECB :
3002 andres 4426 GIC 119 : ReservePrivateRefCountEntry();
4427 :
2555 4428 119 : buf_state = LockBufHdr(bufHdr);
228 rhaas 4429 GNC 119 : if (bufHdr->tag.dbOid == dbid &&
2555 andres 4430 GIC 119 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
4431 : {
5764 tgl 4432 9 : PinBuffer_Locked(bufHdr);
2672 rhaas 4433 9 : LWLockAcquire(BufferDescriptorGetContentLock(bufHdr), LW_SHARED);
59 andres 4434 GNC 9 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
2672 rhaas 4435 CBC 9 : LWLockRelease(BufferDescriptorGetContentLock(bufHdr));
191 michael 4436 GNC 9 : UnpinBuffer(bufHdr);
4437 : }
5764 tgl 4438 ECB : else
2555 andres 4439 CBC 110 : UnlockBufHdr(bufHdr, buf_state);
4440 : }
5764 tgl 4441 GIC 3 : }
5764 tgl 4442 ECB :
2677 andres 4443 : /*
4444 : * Flush a previously, shared or exclusively, locked and pinned buffer to the
4445 : * OS.
4446 : */
4447 : void
2677 andres 4448 GIC 25 : FlushOneBuffer(Buffer buffer)
2677 andres 4449 ECB : {
4450 : BufferDesc *bufHdr;
4451 :
4452 : /* currently not needed, but no fundamental reason not to support */
2677 andres 4453 GIC 25 : Assert(!BufferIsLocal(buffer));
4454 :
4455 25 : Assert(BufferIsPinned(buffer));
4456 :
2677 andres 4457 CBC 25 : bufHdr = GetBufferDescriptor(buffer - 1);
4458 :
2672 rhaas 4459 25 : Assert(LWLockHeldByMe(BufferDescriptorGetContentLock(bufHdr)));
2677 andres 4460 ECB :
59 andres 4461 GNC 25 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
2677 andres 4462 CBC 25 : }
4463 :
9770 scrappy 4464 EUB : /*
4465 : * ReleaseBuffer -- release the pin on a buffer
9770 scrappy 4466 ECB : */
4467 : void
7551 tgl 4468 GIC 67676152 : ReleaseBuffer(Buffer buffer)
4469 : {
6750 4470 67676152 : if (!BufferIsValid(buffer))
4312 peter_e 4471 UIC 0 : elog(ERROR, "bad buffer ID: %d", buffer);
4472 :
9345 bruce 4473 GIC 67676152 : if (BufferIsLocal(buffer))
4 andres 4474 GNC 1502576 : UnpinLocalBuffer(buffer);
4475 : else
4476 66173576 : UnpinBuffer(GetBufferDescriptor(buffer - 1));
9770 scrappy 4477 GIC 67676152 : }
4478 :
4479 : /*
4480 : * UnlockReleaseBuffer -- release the content lock and pin on a buffer
4481 : *
4482 : * This is just a shorthand for a common combination.
6218 tgl 4483 ECB : */
4484 : void
6218 tgl 4485 CBC 19887237 : UnlockReleaseBuffer(Buffer buffer)
4486 : {
6218 tgl 4487 GIC 19887237 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
4488 19887237 : ReleaseBuffer(buffer);
4489 19887237 : }
4490 :
4491 : /*
6840 tgl 4492 ECB : * IncrBufferRefCount
4493 : * Increment the pin count on a buffer that we have *already* pinned
4494 : * at least once.
4495 : *
4496 : * This function cannot be used on a buffer we do not have pinned,
6710 neilc 4497 : * because it doesn't change the shared buffer state.
4498 : */
6840 tgl 4499 EUB : void
6840 tgl 4500 GIC 11173807 : IncrBufferRefCount(Buffer buffer)
6840 tgl 4501 ECB : {
6710 neilc 4502 GIC 11173807 : Assert(BufferIsPinned(buffer));
6840 tgl 4503 11173807 : ResourceOwnerEnlargeBuffers(CurrentResourceOwner);
4504 11173807 : if (BufferIsLocal(buffer))
4505 346867 : LocalRefCount[-buffer - 1]++;
4506 : else
4507 : {
4508 : PrivateRefCountEntry *ref;
4509 :
3002 andres 4510 10826940 : ref = GetPrivateRefCountEntry(buffer, true);
3144 4511 10826940 : Assert(ref != NULL);
4512 10826940 : ref->refcount++;
4513 : }
1978 tgl 4514 11173807 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buffer);
6840 4515 11173807 : }
4516 :
4517 : /*
4518 : * MarkBufferDirtyHint
4519 : *
4520 : * Mark a buffer dirty for non-critical changes.
4521 : *
4522 : * This is essentially the same as MarkBufferDirty, except:
4523 : *
4524 : * 1. The caller does not write WAL; so if checksums are enabled, we may need
4525 : * to write an XLOG_FPI_FOR_HINT WAL record to protect against torn pages.
4526 : * 2. The caller might have only share-lock instead of exclusive-lock on the
4527 : * buffer's content lock.
4528 : * 3. This function does not guarantee that the buffer is always marked dirty
4529 : * (due to a race condition), so it cannot be used for important changes.
4530 : */
4531 : void
3583 jdavis 4532 16382378 : MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
4533 : {
4534 : BufferDesc *bufHdr;
2545 kgrittn 4535 16382378 : Page page = BufferGetPage(buffer);
4536 :
6750 tgl 4537 16382378 : if (!BufferIsValid(buffer))
4312 peter_e 4538 UIC 0 : elog(ERROR, "bad buffer ID: %d", buffer);
4539 :
8400 tgl 4540 GIC 16382378 : if (BufferIsLocal(buffer))
4541 : {
6218 4542 678465 : MarkLocalBufferDirty(buffer);
8400 4543 678465 : return;
4544 : }
4545 :
2992 andres 4546 15703913 : bufHdr = GetBufferDescriptor(buffer - 1);
4547 :
3144 4548 15703913 : Assert(GetPrivateRefCount(buffer) > 0);
4549 : /* here, either share or exclusive lock is OK */
2672 rhaas 4550 15703913 : Assert(LWLockHeldByMe(BufferDescriptorGetContentLock(bufHdr)));
4551 :
4552 : /*
4553 : * This routine might get called many times on the same page, if we are
4554 : * making the first scan after commit of an xact that added/deleted many
4555 : * tuples. So, be as quick as we can if the buffer is already dirty. We
4556 : * do this by not acquiring spinlock if it looks like the status bits are
4557 : * already set. Since we make this test unlocked, there's a chance we
4558 : * might fail to notice that the flags have just been cleared, and failed
4559 : * to reset them, due to memory-ordering issues. But since this function
4560 : * is only intended to be used in cases where failing to write out the
4561 : * data would be harmless anyway, it doesn't really matter.
4562 : */
2555 andres 4563 15703913 : if ((pg_atomic_read_u32(&bufHdr->state) & (BM_DIRTY | BM_JUST_DIRTIED)) !=
4564 : (BM_DIRTY | BM_JUST_DIRTIED))
4565 : {
3670 simon 4566 1716964 : XLogRecPtr lsn = InvalidXLogRecPtr;
4567 1716964 : bool dirtied = false;
366 rhaas 4568 1716964 : bool delayChkptFlags = false;
4569 : uint32 buf_state;
4570 :
4571 : /*
4572 : * If we need to protect hint bit updates from torn writes, WAL-log a
4573 : * full page image of the page. This full page image is only necessary
4574 : * if the hint bit update is the first change to the page since the
4575 : * last checkpoint.
4576 : *
4577 : * We don't check full_page_writes here because that logic is included
4578 : * when we call XLogInsert() since the value changes dynamically.
4579 : */
2555 andres 4580 CBC 3384330 : if (XLogHintBitIsNeeded() &&
2555 andres 4581 GIC 1667366 : (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT))
4582 : {
4583 : /*
4584 : * If we must not write WAL, due to a relfilelocator-specific
1100 noah 4585 ECB : * condition or being in recovery, don't dirty the page. We can
4586 : * set the hint, just not dirty the page as a result so the hint
4587 : * is lost when we evict the page or shutdown.
4588 : *
4589 : * See src/backend/storage/page/README for longer discussion.
4590 : */
1100 noah 4591 GIC 1716890 : if (RecoveryInProgress() ||
228 rhaas 4592 GNC 49527 : RelFileLocatorSkippingWAL(BufTagGetRelFileLocator(&bufHdr->tag)))
3670 simon 4593 CBC 1617836 : return;
3670 simon 4594 ECB :
4595 : /*
4596 : * If the block is already dirty because we either made a change
4597 : * or set a hint already, then we don't need to write a full page
4598 : * image. Note that aggressive cleaning of blocks dirtied by hint
4599 : * bit setting would increase the call rate. Bulk setting of hint
3602 bruce 4600 : * bits would reduce the call rate...
4601 : *
4602 : * We must issue the WAL record before we mark the buffer dirty.
4603 : * Otherwise we might write the page before we write the WAL. That
4604 : * causes a race condition, since a checkpoint might occur between
4605 : * writing the WAL record and marking the buffer dirty. We solve
4606 : * that with a kluge, but one that is already in use during
4607 : * transaction commit to prevent race conditions. Basically, we
4608 : * simply prevent the checkpoint WAL record from being written
4609 : * until we have marked the buffer dirty. We don't start the
4610 : * checkpoint flush until we have marked dirty, so our checkpoint
4611 : * must flush the change to disk successfully or the checkpoint
4612 : * never gets written, so crash recovery will fix.
3670 simon 4613 : *
4614 : * It's possible we may enter here without an xid, so it is
4615 : * essential that CreateCheckPoint waits for virtual transactions
4616 : * rather than full transactionids.
4617 : */
366 rhaas 4618 CBC 49527 : Assert((MyProc->delayChkptFlags & DELAY_CHKPT_START) == 0);
366 rhaas 4619 GIC 49527 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
4620 49527 : delayChkptFlags = true;
3583 jdavis 4621 49527 : lsn = XLogSaveBufferForHint(buffer, buffer_std);
3670 simon 4622 ECB : }
4623 :
2555 andres 4624 GIC 99128 : buf_state = LockBufHdr(bufHdr);
2555 andres 4625 ECB :
2555 andres 4626 GIC 99128 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
4627 :
2555 andres 4628 CBC 99128 : if (!(buf_state & BM_DIRTY))
4629 : {
3670 simon 4630 GIC 99112 : dirtied = true; /* Means "will be dirtied by this action" */
4631 :
3670 simon 4632 ECB : /*
4633 : * Set the page LSN if we wrote a backup block. We aren't supposed
4634 : * to set this when only holding a share lock but as long as we
4635 : * serialise it somehow we're OK. We choose to set LSN while
3602 bruce 4636 : * holding the buffer header lock, which causes any reader of an
4637 : * LSN who holds only a share lock to also obtain a buffer header
4638 : * lock before using PageGetLSN(), which is enforced in
4639 : * BufferGetLSNAtomic().
4640 : *
4641 : * If checksums are enabled, you might think we should reset the
3670 simon 4642 : * checksum here. That will happen when the page is written
4643 : * sometime later in this checkpoint cycle.
4644 : */
3670 simon 4645 GIC 99112 : if (!XLogRecPtrIsInvalid(lsn))
4646 5130 : PageSetLSN(page, lsn);
4647 : }
2555 andres 4648 ECB :
2555 andres 4649 CBC 99128 : buf_state |= BM_DIRTY | BM_JUST_DIRTIED;
2555 andres 4650 GIC 99128 : UnlockBufHdr(bufHdr, buf_state);
3670 simon 4651 ECB :
366 rhaas 4652 GIC 99128 : if (delayChkptFlags)
366 rhaas 4653 CBC 49527 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
3670 simon 4654 ECB :
3670 simon 4655 CBC 99128 : if (dirtied)
4656 : {
4153 alvherre 4657 99112 : VacuumPageDirty++;
3296 rhaas 4658 99112 : pgBufferUsage.shared_blks_dirtied++;
4153 alvherre 4659 99112 : if (VacuumCostActive)
4660 445 : VacuumCostBalance += VacuumCostPageDirty;
4153 alvherre 4661 ECB : }
4662 : }
4663 : }
8881 vadim4o 4664 :
4665 : /*
4666 : * Release buffer content locks for shared buffers.
4667 : *
4668 : * Used to clean up after errors.
4669 : *
4670 : * Currently, we can expect that lwlock.c's LWLockReleaseAll() took care
4671 : * of releasing buffer content locks per se; the only thing we need to deal
4672 : * with here is clearing any PIN_COUNT request that was in progress.
4673 : */
4674 : void
8122 tgl 4675 GIC 37890 : UnlockBuffers(void)
4676 : {
2701 rhaas 4677 37890 : BufferDesc *buf = PinCountWaitBuf;
8881 vadim4o 4678 ECB :
6749 tgl 4679 GIC 37890 : if (buf)
4680 : {
4681 : uint32 buf_state;
4682 :
2555 andres 4683 UIC 0 : buf_state = LockBufHdr(buf);
6749 tgl 4684 ECB :
7862 tgl 4685 EUB : /*
4686 : * Don't complain if flag bit not set; it could have been reset but we
4687 : * got a cancel/die interrupt before getting the signal.
7862 tgl 4688 ECB : */
2555 andres 4689 UIC 0 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
479 tmunro 4690 LBC 0 : buf->wait_backend_pgprocno == MyProc->pgprocno)
2555 andres 4691 UIC 0 : buf_state &= ~BM_PIN_COUNT_WAITER;
6610 tgl 4692 ECB :
2555 andres 4693 UIC 0 : UnlockBufHdr(buf, buf_state);
8167 tgl 4694 ECB :
6610 tgl 4695 LBC 0 : PinCountWaitBuf = NULL;
4696 : }
8881 vadim4o 4697 GIC 37890 : }
4698 :
4699 : /*
4700 : * Acquire or release the content_lock for the buffer.
4701 : */
8881 vadim4o 4702 ECB : void
8720 bruce 4703 GIC 196227071 : LockBuffer(Buffer buffer, int mode)
4704 : {
2701 rhaas 4705 ECB : BufferDesc *buf;
8881 vadim4o 4706 EUB :
993 pg 4707 GIC 196227071 : Assert(BufferIsPinned(buffer));
8881 vadim4o 4708 196227071 : if (BufferIsLocal(buffer))
6218 tgl 4709 CBC 9816707 : return; /* local buffers need no lock */
4710 :
2992 andres 4711 186410364 : buf = GetBufferDescriptor(buffer - 1);
4712 :
8881 vadim4o 4713 186410364 : if (mode == BUFFER_LOCK_UNLOCK)
2672 rhaas 4714 94045800 : LWLockRelease(BufferDescriptorGetContentLock(buf));
8881 vadim4o 4715 GIC 92364564 : else if (mode == BUFFER_LOCK_SHARE)
2672 rhaas 4716 64459683 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_SHARED);
8881 vadim4o 4717 27904881 : else if (mode == BUFFER_LOCK_EXCLUSIVE)
2672 rhaas 4718 27904881 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_EXCLUSIVE);
4719 : else
7199 tgl 4720 UIC 0 : elog(ERROR, "unrecognized buffer lock mode: %d", mode);
4721 : }
8483 inoue 4722 ECB :
4723 : /*
4724 : * Acquire the content_lock for the buffer, but only if we don't have to wait.
4725 : *
7182 tgl 4726 : * This assumes the caller wants BUFFER_LOCK_EXCLUSIVE mode.
4727 : */
4728 : bool
7182 tgl 4729 GIC 1220713 : ConditionalLockBuffer(Buffer buffer)
7182 tgl 4730 ECB : {
2701 rhaas 4731 : BufferDesc *buf;
4732 :
993 pg 4733 GIC 1220713 : Assert(BufferIsPinned(buffer));
7182 tgl 4734 1220713 : if (BufferIsLocal(buffer))
4735 64689 : return true; /* act as though we got it */
4736 :
2992 andres 4737 1156024 : buf = GetBufferDescriptor(buffer - 1);
4738 :
2672 rhaas 4739 GBC 1156024 : return LWLockConditionalAcquire(BufferDescriptorGetContentLock(buf),
2672 rhaas 4740 EUB : LW_EXCLUSIVE);
4741 : }
4742 :
4743 : /*
4744 : * Verify that this backend is pinning the buffer exactly once.
4745 : *
4746 : * NOTE: Like in BufferIsPinned(), what we check here is that *this* backend
4747 : * holds a pin on the buffer. We do not care whether some other backend does.
4748 : */
4749 : void
4 andres 4750 GNC 3785962 : CheckBufferIsPinnedOnce(Buffer buffer)
4751 : {
4752 3785962 : if (BufferIsLocal(buffer))
4753 : {
4754 16 : if (LocalRefCount[-buffer - 1] != 1)
4 andres 4755 UNC 0 : elog(ERROR, "incorrect local pin count: %d",
4756 : LocalRefCount[-buffer - 1]);
4757 : }
4758 : else
4759 : {
4 andres 4760 GNC 3785946 : if (GetPrivateRefCount(buffer) != 1)
4 andres 4761 UNC 0 : elog(ERROR, "incorrect local pin count: %d",
4762 : GetPrivateRefCount(buffer));
4763 : }
4 andres 4764 GNC 3785962 : }
4765 :
4766 : /*
4767 : * LockBufferForCleanup - lock a buffer in preparation for deleting items
4768 : *
7947 tgl 4769 ECB : * Items may be deleted from a disk page only when the caller (a) holds an
4770 : * exclusive lock on the buffer and (b) has observed that no other backend
4771 : * holds a pin on the buffer. If there is a pin, then the other backend
4772 : * might have a pointer into the buffer (for example, a heapscan reference
4773 : * to an item --- see README for more details). It's OK if a pin is added
4774 : * after the cleanup starts, however; the newly-arrived backend will be
4775 : * unable to look at the page until we release the exclusive lock.
4776 : *
4777 : * To implement this protocol, a would-be deleter must pin the buffer and
4778 : * then call LockBufferForCleanup(). LockBufferForCleanup() is similar to
4779 : * LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE), except that it loops until
4780 : * it has successfully observed pin count = 1.
4781 : */
4782 : void
7947 tgl 4783 GIC 37384 : LockBufferForCleanup(Buffer buffer)
4784 : {
2701 rhaas 4785 ECB : BufferDesc *bufHdr;
821 fujii 4786 GIC 37384 : TimestampTz waitStart = 0;
48 drowley 4787 GNC 37384 : bool waiting = false;
821 fujii 4788 CBC 37384 : bool logged_recovery_conflict = false;
4789 :
993 pg 4790 GIC 37384 : Assert(BufferIsPinned(buffer));
6749 tgl 4791 37384 : Assert(PinCountWaitBuf == NULL);
4792 :
4 andres 4793 GNC 37384 : CheckBufferIsPinnedOnce(buffer);
4 andres 4794 ECB :
4795 : /* Nobody else to wait for */
7947 tgl 4796 GNC 37384 : if (BufferIsLocal(buffer))
4797 16 : return;
4798 :
2992 andres 4799 GIC 37368 : bufHdr = GetBufferDescriptor(buffer - 1);
4800 :
4801 : for (;;)
7947 tgl 4802 11 : {
4803 : uint32 buf_state;
4804 :
4805 : /* Try to acquire lock */
4806 37379 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2555 andres 4807 37379 : buf_state = LockBufHdr(bufHdr);
4808 :
4809 37379 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
4810 37379 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
4811 : {
4812 : /* Successfully acquired exclusive lock with pincount 1 */
2555 andres 4813 CBC 37368 : UnlockBufHdr(bufHdr, buf_state);
4814 :
4815 : /*
816 fujii 4816 ECB : * Emit the log message if recovery conflict on buffer pin was
4817 : * resolved but the startup process waited longer than
4818 : * deadlock_timeout for it.
4819 : */
816 fujii 4820 CBC 37368 : if (logged_recovery_conflict)
4821 2 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
4822 : waitStart, GetCurrentTimestamp(),
4823 : NULL, false);
4824 :
48 drowley 4825 GNC 37368 : if (waiting)
1105 fujii 4826 ECB : {
4827 : /* reset ps display to remove the suffix if we added one */
48 drowley 4828 GNC 2 : set_ps_display_remove_suffix();
4829 2 : waiting = false;
4830 : }
7947 tgl 4831 GIC 37368 : return;
7947 tgl 4832 ECB : }
4833 : /* Failed, so mark myself as waiting for pincount 1 */
2555 andres 4834 GIC 11 : if (buf_state & BM_PIN_COUNT_WAITER)
4835 : {
2555 andres 4836 LBC 0 : UnlockBufHdr(bufHdr, buf_state);
7947 tgl 4837 UIC 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
7199 tgl 4838 LBC 0 : elog(ERROR, "multiple backends attempting to wait for pincount 1");
4839 : }
479 tmunro 4840 GIC 11 : bufHdr->wait_backend_pgprocno = MyProc->pgprocno;
6749 tgl 4841 CBC 11 : PinCountWaitBuf = bufHdr;
2555 andres 4842 GIC 11 : buf_state |= BM_PIN_COUNT_WAITER;
4843 11 : UnlockBufHdr(bufHdr, buf_state);
7947 tgl 4844 CBC 11 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
4824 simon 4845 ECB :
4846 : /* Wait to be signaled by UnpinBuffer() */
4824 simon 4847 CBC 11 : if (InHotStandby)
4848 : {
48 drowley 4849 GNC 11 : if (!waiting)
4850 : {
4851 : /* adjust the process title to indicate that it's waiting */
4852 2 : set_ps_display_suffix("waiting");
4853 2 : waiting = true;
4854 : }
1105 fujii 4855 ECB :
4856 : /*
821 4857 : * Emit the log message if the startup process is waiting longer
4858 : * than deadlock_timeout for recovery conflict on buffer pin.
4859 : *
4860 : * Skip this if first time through because the startup process has
4861 : * not started waiting yet in this case. So, the wait start
4862 : * timestamp is set after this logic.
4863 : */
821 fujii 4864 GIC 11 : if (waitStart != 0 && !logged_recovery_conflict)
4865 : {
4866 4 : TimestampTz now = GetCurrentTimestamp();
4867 :
4868 4 : if (TimestampDifferenceExceeds(waitStart, now,
4869 : DeadlockTimeout))
4870 : {
4871 2 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
4872 : waitStart, now, NULL, true);
4873 2 : logged_recovery_conflict = true;
4874 : }
4875 : }
4876 :
821 fujii 4877 ECB : /*
4878 : * Set the wait start timestamp if logging is enabled and first
4879 : * time through.
4880 : */
821 fujii 4881 GIC 11 : if (log_recovery_conflict_waits && waitStart == 0)
4882 2 : waitStart = GetCurrentTimestamp();
4883 :
4268 tgl 4884 ECB : /* Publish the bufid that Startup process waits on */
4824 simon 4885 GIC 11 : SetStartupBufferPinWaitBufId(buffer - 1);
4886 : /* Set alarm and then wait to be signaled by UnpinBuffer() */
4887 11 : ResolveRecoveryConflictWithBufferPin();
4888 : /* Reset the published bufid */
4889 11 : SetStartupBufferPinWaitBufId(-1);
4890 : }
4891 : else
2377 rhaas 4892 UIC 0 : ProcWaitForSignal(PG_WAIT_BUFFER_PIN);
2586 rhaas 4893 ECB :
4894 : /*
4895 : * Remove flag marking us as waiter. Normally this will not be set
2967 andres 4896 : * anymore, but ProcWaitForSignal() can return for other signals as
4897 : * well. We take care to only reset the flag if we're the waiter, as
4898 : * theoretically another backend could have started waiting. That's
4899 : * impossible with the current usages due to table level locking, but
4900 : * better be safe.
4901 : */
2555 andres 4902 GIC 11 : buf_state = LockBufHdr(bufHdr);
2555 andres 4903 CBC 11 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
479 tmunro 4904 GIC 9 : bufHdr->wait_backend_pgprocno == MyProc->pgprocno)
2555 andres 4905 CBC 9 : buf_state &= ~BM_PIN_COUNT_WAITER;
2555 andres 4906 GIC 11 : UnlockBufHdr(bufHdr, buf_state);
4907 :
6749 tgl 4908 11 : PinCountWaitBuf = NULL;
4909 : /* Loop back and try again */
4910 : }
7947 tgl 4911 ECB : }
4912 :
4913 : /*
4914 : * Check called from RecoveryConflictInterrupt handler when Startup
4302 peter_e 4915 : * process requests cancellation of all pin holders that are blocking it.
4916 : */
4917 : bool
4824 simon 4918 GIC 4 : HoldingBufferPinThatDelaysRecovery(void)
4919 : {
4790 bruce 4920 4 : int bufid = GetStartupBufferPinWaitBufId();
4824 simon 4921 ECB :
4922 : /*
4790 bruce 4923 : * If we get woken slowly then it's possible that the Startup process was
4924 : * already woken by other backends before we got here. Also possible that
4925 : * we get here by multiple interrupts or interrupts at inappropriate
4824 simon 4926 : * times, so make sure we do nothing if the bufid is not set.
4927 : */
4824 simon 4928 CBC 4 : if (bufid < 0)
4824 simon 4929 GIC 2 : return false;
4930 :
3144 andres 4931 2 : if (GetPrivateRefCount(bufid + 1) > 0)
4824 simon 4932 2 : return true;
4933 :
4824 simon 4934 UIC 0 : return false;
4935 : }
4936 :
4937 : /*
4938 : * ConditionalLockBufferForCleanup - as above, but don't wait to get the lock
4939 : *
4940 : * We won't loop, but just check once to see if the pin count is OK. If
4941 : * not, return false with no lock held.
4942 : */
4943 : bool
5680 tgl 4944 GIC 270333 : ConditionalLockBufferForCleanup(Buffer buffer)
4945 : {
2701 rhaas 4946 ECB : BufferDesc *bufHdr;
4947 : uint32 buf_state,
4948 : refcount;
4949 :
5680 tgl 4950 GIC 270333 : Assert(BufferIsValid(buffer));
4951 :
5680 tgl 4952 CBC 270333 : if (BufferIsLocal(buffer))
4953 : {
2555 andres 4954 782 : refcount = LocalRefCount[-buffer - 1];
4955 : /* There should be exactly one pin */
2555 andres 4956 GIC 782 : Assert(refcount > 0);
4957 782 : if (refcount != 1)
5680 tgl 4958 CBC 21 : return false;
4959 : /* Nobody else to wait for */
5680 tgl 4960 GIC 761 : return true;
4961 : }
4962 :
4963 : /* There should be exactly one local pin */
2555 andres 4964 CBC 269551 : refcount = GetPrivateRefCount(buffer);
4965 269551 : Assert(refcount);
2555 andres 4966 GIC 269551 : if (refcount != 1)
5680 tgl 4967 CBC 210 : return false;
4968 :
5680 tgl 4969 ECB : /* Try to acquire lock */
5680 tgl 4970 CBC 269341 : if (!ConditionalLockBuffer(buffer))
4971 28 : return false;
4972 :
2992 andres 4973 269313 : bufHdr = GetBufferDescriptor(buffer - 1);
2555 4974 269313 : buf_state = LockBufHdr(bufHdr);
4975 269313 : refcount = BUF_STATE_GET_REFCOUNT(buf_state);
2555 andres 4976 ECB :
2555 andres 4977 CBC 269313 : Assert(refcount > 0);
2555 andres 4978 GIC 269313 : if (refcount == 1)
4979 : {
5680 tgl 4980 ECB : /* Successfully acquired exclusive lock with pincount 1 */
2555 andres 4981 GIC 269275 : UnlockBufHdr(bufHdr, buf_state);
5680 tgl 4982 CBC 269275 : return true;
4983 : }
4984 :
4985 : /* Failed, so release the lock */
2555 andres 4986 GIC 38 : UnlockBufHdr(bufHdr, buf_state);
5680 tgl 4987 38 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
4988 38 : return false;
5680 tgl 4989 ECB : }
4990 :
4991 : /*
4992 : * IsBufferCleanupOK - as above, but we already have the lock
4993 : *
2347 rhaas 4994 : * Check whether it's OK to perform cleanup on a buffer we've already
4995 : * locked. If we observe that the pin count is 1, our exclusive lock
4996 : * happens to be a cleanup lock, and we can proceed with anything that
4997 : * would have been allowable had we sought a cleanup lock originally.
4998 : */
4999 : bool
2347 rhaas 5000 CBC 2016 : IsBufferCleanupOK(Buffer buffer)
5001 : {
2347 rhaas 5002 ECB : BufferDesc *bufHdr;
5003 : uint32 buf_state;
5004 :
2347 rhaas 5005 GIC 2016 : Assert(BufferIsValid(buffer));
5006 :
5007 2016 : if (BufferIsLocal(buffer))
5008 : {
2347 rhaas 5009 ECB : /* There should be exactly one pin */
2347 rhaas 5010 UIC 0 : if (LocalRefCount[-buffer - 1] != 1)
2347 rhaas 5011 LBC 0 : return false;
2347 rhaas 5012 EUB : /* Nobody else to wait for */
2347 rhaas 5013 UIC 0 : return true;
2347 rhaas 5014 ECB : }
5015 :
5016 : /* There should be exactly one local pin */
2347 rhaas 5017 CBC 2016 : if (GetPrivateRefCount(buffer) != 1)
2347 rhaas 5018 LBC 0 : return false;
5019 :
2347 rhaas 5020 GIC 2016 : bufHdr = GetBufferDescriptor(buffer - 1);
5021 :
5022 : /* caller must hold exclusive lock on buffer */
5023 2016 : Assert(LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
5024 : LW_EXCLUSIVE));
5025 :
2347 rhaas 5026 CBC 2016 : buf_state = LockBufHdr(bufHdr);
5027 :
5028 2016 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5029 2016 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
2347 rhaas 5030 ECB : {
5031 : /* pincount is OK. */
2347 rhaas 5032 GIC 2016 : UnlockBufHdr(bufHdr, buf_state);
5033 2016 : return true;
5034 : }
5035 :
2347 rhaas 5036 UIC 0 : UnlockBufHdr(bufHdr, buf_state);
5037 0 : return false;
5038 : }
5039 :
5040 :
8483 inoue 5041 ECB : /*
5042 : * Functions for buffer I/O handling
5043 : *
6610 tgl 5044 : * Note: We assume that nested buffer I/O never occurs.
759 tmunro 5045 : * i.e at most one BM_IO_IN_PROGRESS bit is set per proc.
6610 tgl 5046 : *
5047 : * Also note that these are used only for shared buffers, not local ones.
5048 : */
5049 :
5050 : /*
5051 : * WaitIO -- Block until the IO_IN_PROGRESS flag on 'buf' is cleared.
6749 5052 : */
6610 5053 : static void
2701 rhaas 5054 GIC 1696 : WaitIO(BufferDesc *buf)
6610 tgl 5055 ECB : {
759 tmunro 5056 CBC 1696 : ConditionVariable *cv = BufferDescriptorGetIOCV(buf);
5057 :
759 tmunro 5058 GIC 1696 : ConditionVariablePrepareToSleep(cv);
5059 : for (;;)
6610 tgl 5060 1623 : {
5061 : uint32 buf_state;
5062 :
5063 : /*
5064 : * It may not be necessary to acquire the spinlock to check the flag
5065 : * here, but since this test is essential for correctness, we'd better
5066 : * play it safe.
5067 : */
2555 andres 5068 3319 : buf_state = LockBufHdr(buf);
5069 3319 : UnlockBufHdr(buf, buf_state);
5070 :
5071 3319 : if (!(buf_state & BM_IO_IN_PROGRESS))
6610 tgl 5072 1696 : break;
759 tmunro 5073 CBC 1623 : ConditionVariableSleep(cv, WAIT_EVENT_BUFFER_IO);
5074 : }
759 tmunro 5075 GIC 1696 : ConditionVariableCancelSleep();
6610 tgl 5076 CBC 1696 : }
5077 :
8483 inoue 5078 ECB : /*
6610 tgl 5079 EUB : * StartBufferIO: begin I/O on this buffer
5080 : * (Assumptions)
8483 inoue 5081 ECB : * My process is executing no IO
5082 : * The buffer is Pinned
5083 : *
6610 tgl 5084 : * In some scenarios there are race conditions in which multiple backends
5085 : * could attempt the same I/O operation concurrently. If someone else
5086 : * has already started I/O on this buffer then we will block on the
759 tmunro 5087 : * I/O condition variable until he's done.
5088 : *
6610 tgl 5089 : * Input operations are only attempted on buffers that are not BM_VALID,
5090 : * and output operations only on buffers that are BM_VALID and BM_DIRTY,
5091 : * so we can always tell if the work is already done.
5092 : *
5093 : * Returns true if we successfully marked the buffer as I/O busy,
5094 : * false if someone else already did the work.
5095 : */
5096 : static bool
2701 rhaas 5097 GIC 2596864 : StartBufferIO(BufferDesc *buf, bool forInput)
5098 : {
5099 : uint32 buf_state;
5100 :
4 andres 5101 GNC 2596864 : ResourceOwnerEnlargeBufferIOs(CurrentResourceOwner);
5102 :
5103 : for (;;)
6610 tgl 5104 ECB : {
2555 andres 5105 GIC 2598558 : buf_state = LockBufHdr(buf);
5106 :
2555 andres 5107 CBC 2598558 : if (!(buf_state & BM_IO_IN_PROGRESS))
6610 tgl 5108 2596864 : break;
2555 andres 5109 1694 : UnlockBufHdr(buf, buf_state);
6610 tgl 5110 GIC 1694 : WaitIO(buf);
5111 : }
5112 :
5113 : /* Once we get here, there is definitely no I/O active on this buffer */
5114 :
2555 andres 5115 2596864 : if (forInput ? (buf_state & BM_VALID) : !(buf_state & BM_DIRTY))
5116 : {
5117 : /* someone else already did the I/O */
5118 1768 : UnlockBufHdr(buf, buf_state);
6610 tgl 5119 1768 : return false;
5120 : }
6610 tgl 5121 ECB :
2555 andres 5122 CBC 2595096 : buf_state |= BM_IO_IN_PROGRESS;
2555 andres 5123 GIC 2595096 : UnlockBufHdr(buf, buf_state);
5124 :
4 andres 5125 GNC 2595096 : ResourceOwnerRememberBufferIO(CurrentResourceOwner,
5126 : BufferDescriptorGetBuffer(buf));
5127 :
6610 tgl 5128 GIC 2595096 : return true;
5129 : }
5130 :
5131 : /*
6610 tgl 5132 ECB : * TerminateBufferIO: release a buffer we were doing I/O on
8483 inoue 5133 : * (Assumptions)
5134 : * My process is executing IO for the buffer
5135 : * BM_IO_IN_PROGRESS bit is set for the buffer
5136 : * The buffer is Pinned
5137 : *
5138 : * If clear_dirty is true and BM_JUST_DIRTIED is not set, we clear the
5139 : * buffer's BM_DIRTY flag. This is appropriate when terminating a
5140 : * successful write. The check on BM_JUST_DIRTIED is necessary to avoid
5141 : * marking the buffer clean if it was re-dirtied while we were writing.
5142 : *
5143 : * set_flag_bits gets ORed into the buffer's flags. It must include
5144 : * BM_IO_ERROR in a failure case. For successful completion it could
5145 : * be 0, or BM_VALID if we just finished reading in the page.
5146 : */
5147 : static void
2555 andres 5148 GIC 2595096 : TerminateBufferIO(BufferDesc *buf, bool clear_dirty, uint32 set_flag_bits)
5149 : {
5150 : uint32 buf_state;
5151 :
5152 2595096 : buf_state = LockBufHdr(buf);
5153 :
5154 2595096 : Assert(buf_state & BM_IO_IN_PROGRESS);
5155 :
5156 2595096 : buf_state &= ~(BM_IO_IN_PROGRESS | BM_IO_ERROR);
2555 andres 5157 CBC 2595096 : if (clear_dirty && !(buf_state & BM_JUST_DIRTIED))
5158 725804 : buf_state &= ~(BM_DIRTY | BM_CHECKPOINT_NEEDED);
6927 tgl 5159 ECB :
2555 andres 5160 CBC 2595096 : buf_state |= set_flag_bits;
2555 andres 5161 GIC 2595096 : UnlockBufHdr(buf, buf_state);
5162 :
4 andres 5163 GNC 2595096 : ResourceOwnerForgetBufferIO(CurrentResourceOwner,
5164 : BufferDescriptorGetBuffer(buf));
5165 :
759 tmunro 5166 CBC 2595096 : ConditionVariableBroadcast(BufferDescriptorGetIOCV(buf));
8483 inoue 5167 GIC 2595096 : }
8483 inoue 5168 ECB :
5169 : /*
5170 : * AbortBufferIO: Clean up active buffer I/O after an error.
5171 : *
5172 : * All LWLocks we might have held have been released,
5173 : * but we haven't yet released buffer pins, so the buffer is still pinned.
5174 : *
5175 : * If I/O was in progress, we always set BM_IO_ERROR, even though it's
5176 : * possible the error condition wasn't related to the I/O.
5177 : */
5178 : void
4 andres 5179 GNC 13 : AbortBufferIO(Buffer buf)
5180 : {
5181 13 : BufferDesc *buf_hdr = GetBufferDescriptor(buf - 1);
5182 : uint32 buf_state;
5183 :
5184 13 : buf_state = LockBufHdr(buf_hdr);
5185 13 : Assert(buf_state & (BM_IO_IN_PROGRESS | BM_TAG_VALID));
5186 :
5187 13 : if (!(buf_state & BM_VALID))
5188 : {
5189 13 : Assert(!(buf_state & BM_DIRTY));
5190 13 : UnlockBufHdr(buf_hdr, buf_state);
4 andres 5191 ECB : }
5192 : else
5193 : {
2 andres 5194 UNC 0 : Assert(buf_state & BM_DIRTY);
4 5195 0 : UnlockBufHdr(buf_hdr, buf_state);
5196 :
5197 : /* Issue notice if this is not the first failure... */
5198 0 : if (buf_state & BM_IO_ERROR)
5199 : {
5200 : /* Buffer is pinned, so we can read tag without spinlock */
5201 : char *path;
5202 :
5203 0 : path = relpathperm(BufTagGetRelFileLocator(&buf_hdr->tag),
5204 : BufTagGetForkNum(&buf_hdr->tag));
5205 0 : ereport(WARNING,
5206 : (errcode(ERRCODE_IO_ERROR),
5207 : errmsg("could not write block %u of %s",
5208 : buf_hdr->tag.blockNum, path),
5209 : errdetail("Multiple failures --- write error might be permanent.")));
5210 0 : pfree(path);
5211 : }
5212 : }
5213 :
4 andres 5214 GNC 13 : TerminateBufferIO(buf_hdr, false, BM_IO_ERROR);
8483 inoue 5215 GIC 13 : }
8206 vadim4o 5216 ECB :
5217 : /*
5218 : * Error context callback for errors occurring during shared buffer writes.
5219 : */
7274 tgl 5220 EUB : static void
4622 rhaas 5221 GIC 73 : shared_buffer_write_error_callback(void *arg)
5222 : {
2701 5223 73 : BufferDesc *bufHdr = (BufferDesc *) arg;
5224 :
5225 : /* Buffer is pinned, so we can read the tag without locking the spinlock */
7274 tgl 5226 GBC 73 : if (bufHdr != NULL)
5262 heikki.linnakangas 5227 EUB : {
228 rhaas 5228 GNC 73 : char *path = relpathperm(BufTagGetRelFileLocator(&bufHdr->tag),
5229 : BufTagGetForkNum(&bufHdr->tag));
5230 :
4622 rhaas 5231 GBC 73 : errcontext("writing block %u of relation %s",
5232 : bufHdr->tag.blockNum, path);
5233 73 : pfree(path);
5234 : }
4622 rhaas 5235 CBC 73 : }
5236 :
5237 : /*
5238 : * Error context callback for errors occurring during local buffer writes.
5239 : */
5240 : static void
4622 rhaas 5241 LBC 0 : local_buffer_write_error_callback(void *arg)
5242 : {
2701 rhaas 5243 UIC 0 : BufferDesc *bufHdr = (BufferDesc *) arg;
5244 :
4622 rhaas 5245 LBC 0 : if (bufHdr != NULL)
4622 rhaas 5246 ECB : {
228 rhaas 5247 UNC 0 : char *path = relpathbackend(BufTagGetRelFileLocator(&bufHdr->tag),
5248 : MyBackendId,
5249 : BufTagGetForkNum(&bufHdr->tag));
5050 bruce 5250 ECB :
5262 heikki.linnakangas 5251 UIC 0 : errcontext("writing block %u of relation %s",
5262 heikki.linnakangas 5252 ECB : bufHdr->tag.blockNum, path);
5262 heikki.linnakangas 5253 LBC 0 : pfree(path);
5262 heikki.linnakangas 5254 ECB : }
7274 tgl 5255 LBC 0 : }
3734 alvherre 5256 ECB :
5257 : /*
5258 : * RelFileLocator qsort/bsearch comparator; see RelFileLocatorEquals.
3734 alvherre 5259 EUB : */
5260 : static int
277 rhaas 5261 GNC 10899862 : rlocator_comparator(const void *p1, const void *p2)
5262 : {
5263 10899862 : RelFileLocator n1 = *(const RelFileLocator *) p1;
5264 10899862 : RelFileLocator n2 = *(const RelFileLocator *) p2;
5265 :
5266 10899862 : if (n1.relNumber < n2.relNumber)
3734 alvherre 5267 GIC 9145773 : return -1;
277 rhaas 5268 GNC 1754089 : else if (n1.relNumber > n2.relNumber)
3734 alvherre 5269 GIC 333597 : return 1;
5270 :
277 rhaas 5271 GNC 1420492 : if (n1.dbOid < n2.dbOid)
3734 alvherre 5272 CBC 37470 : return -1;
277 rhaas 5273 GNC 1383022 : else if (n1.dbOid > n2.dbOid)
3734 alvherre 5274 CBC 45334 : return 1;
5275 :
277 rhaas 5276 GNC 1337688 : if (n1.spcOid < n2.spcOid)
3734 alvherre 5277 UIC 0 : return -1;
277 rhaas 5278 GNC 1337688 : else if (n1.spcOid > n2.spcOid)
3734 alvherre 5279 UIC 0 : return 1;
5280 : else
3734 alvherre 5281 GIC 1337688 : return 0;
5282 : }
5283 :
5284 : /*
5285 : * Lock buffer header - set BM_LOCKED in buffer state.
5286 : */
5287 : uint32
2555 andres 5288 51777935 : LockBufHdr(BufferDesc *desc)
2555 andres 5289 ECB : {
5290 : SpinDelayStatus delayStatus;
5291 : uint32 old_buf_state;
5292 :
4 andres 5293 GNC 51777935 : Assert(!BufferIsLocal(BufferDescriptorGetBuffer(desc)));
5294 :
2551 andres 5295 CBC 51777935 : init_local_spin_delay(&delayStatus);
2551 andres 5296 EUB :
5297 : while (true)
5298 : {
5299 : /* set BM_LOCKED flag */
2555 andres 5300 GIC 51792397 : old_buf_state = pg_atomic_fetch_or_u32(&desc->state, BM_LOCKED);
2555 andres 5301 ECB : /* if it wasn't set before we're OK */
2555 andres 5302 GBC 51792397 : if (!(old_buf_state & BM_LOCKED))
2555 andres 5303 GIC 51777935 : break;
5304 14462 : perform_spin_delay(&delayStatus);
2555 andres 5305 ECB : }
2555 andres 5306 GIC 51777935 : finish_spin_delay(&delayStatus);
5307 51777935 : return old_buf_state | BM_LOCKED;
5308 : }
5309 :
5310 : /*
5311 : * Wait until the BM_LOCKED flag isn't set anymore and return the buffer's
5312 : * state at that point.
5313 : *
5314 : * Obviously the buffer could be locked by the time the value is returned, so
5315 : * this is primarily useful in CAS style loops.
5316 : */
5317 : static uint32
5318 1033 : WaitBufHdrUnlocked(BufferDesc *buf)
5319 : {
5320 : SpinDelayStatus delayStatus;
5321 : uint32 buf_state;
5322 :
2551 5323 1033 : init_local_spin_delay(&delayStatus);
2551 andres 5324 ECB :
2555 andres 5325 GIC 1033 : buf_state = pg_atomic_read_u32(&buf->state);
5326 :
2555 andres 5327 CBC 7839 : while (buf_state & BM_LOCKED)
2555 andres 5328 ECB : {
2555 andres 5329 CBC 6806 : perform_spin_delay(&delayStatus);
2555 andres 5330 GIC 6806 : buf_state = pg_atomic_read_u32(&buf->state);
2555 andres 5331 ECB : }
5332 :
2555 andres 5333 GIC 1033 : finish_spin_delay(&delayStatus);
2555 andres 5334 ECB :
2555 andres 5335 GIC 1033 : return buf_state;
5336 : }
2555 andres 5337 ECB :
2606 5338 : /*
5339 : * BufferTag comparator.
5340 : */
5341 : static inline int
758 tmunro 5342 GIC 2012103 : buffertag_comparator(const BufferTag *ba, const BufferTag *bb)
2606 andres 5343 ECB : {
5344 : int ret;
5345 : RelFileLocator rlocatora;
5346 : RelFileLocator rlocatorb;
5347 :
228 rhaas 5348 GNC 2012103 : rlocatora = BufTagGetRelFileLocator(ba);
5349 2012103 : rlocatorb = BufTagGetRelFileLocator(bb);
5350 :
5351 2012103 : ret = rlocator_comparator(&rlocatora, &rlocatorb);
2606 andres 5352 ECB :
2606 andres 5353 CBC 2012103 : if (ret != 0)
2606 andres 5354 GIC 676080 : return ret;
2606 andres 5355 ECB :
228 rhaas 5356 GNC 1336023 : if (BufTagGetForkNum(ba) < BufTagGetForkNum(bb))
2606 andres 5357 GIC 104693 : return -1;
228 rhaas 5358 GNC 1231330 : if (BufTagGetForkNum(ba) > BufTagGetForkNum(bb))
2606 andres 5359 CBC 65622 : return 1;
5360 :
2606 andres 5361 GIC 1165708 : if (ba->blockNum < bb->blockNum)
5362 767765 : return -1;
5363 397943 : if (ba->blockNum > bb->blockNum)
5364 397318 : return 1;
5365 :
2606 andres 5366 CBC 625 : return 0;
2606 andres 5367 ECB : }
5368 :
5369 : /*
5370 : * Comparator determining the writeout order in a checkpoint.
5371 : *
5372 : * It is important that tablespaces are compared first, the logic balancing
5373 : * writes between tablespaces relies on it.
5374 : */
758 tmunro 5375 : static inline int
758 tmunro 5376 GIC 4878998 : ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b)
2606 andres 5377 ECB : {
5378 : /* compare tablespace */
2606 andres 5379 GIC 4878998 : if (a->tsId < b->tsId)
2606 andres 5380 CBC 17450 : return -1;
2606 andres 5381 GIC 4861548 : else if (a->tsId > b->tsId)
2606 andres 5382 GBC 71952 : return 1;
2606 andres 5383 EUB : /* compare relation */
277 rhaas 5384 GNC 4789596 : if (a->relNumber < b->relNumber)
2606 andres 5385 GIC 1323526 : return -1;
277 rhaas 5386 GNC 3466070 : else if (a->relNumber > b->relNumber)
2606 andres 5387 CBC 1212862 : return 1;
2606 andres 5388 ECB : /* compare fork */
2606 andres 5389 CBC 2253208 : else if (a->forkNum < b->forkNum)
5390 115236 : return -1;
2606 andres 5391 GIC 2137972 : else if (a->forkNum > b->forkNum)
5392 112381 : return 1;
2606 andres 5393 ECB : /* compare block number */
2606 andres 5394 GIC 2025591 : else if (a->blockNum < b->blockNum)
2606 andres 5395 CBC 1018511 : return -1;
1915 tgl 5396 GIC 1007080 : else if (a->blockNum > b->blockNum)
2606 andres 5397 975421 : return 1;
1915 tgl 5398 ECB : /* equal page IDs are unlikely, but not impossible */
1915 tgl 5399 CBC 31659 : return 0;
5400 : }
5401 :
5402 : /*
5403 : * Comparator for a Min-Heap over the per-tablespace checkpoint completion
5404 : * progress.
5405 : */
5406 : static int
2606 andres 5407 GIC 426318 : ts_ckpt_progress_comparator(Datum a, Datum b, void *arg)
5408 : {
5409 426318 : CkptTsStatus *sa = (CkptTsStatus *) a;
2606 andres 5410 CBC 426318 : CkptTsStatus *sb = (CkptTsStatus *) b;
5411 :
2606 andres 5412 ECB : /* we want a min-heap, so return 1 for the a < b */
2606 andres 5413 GIC 426318 : if (sa->progress < sb->progress)
2606 andres 5414 CBC 398630 : return 1;
2606 andres 5415 GIC 27688 : else if (sa->progress == sb->progress)
5416 1458 : return 0;
2606 andres 5417 ECB : else
2606 andres 5418 GIC 26230 : return -1;
2606 andres 5419 ECB : }
5420 :
5421 : /*
5422 : * Initialize a writeback context, discarding potential previous state.
5423 : *
5424 : * *max_pending is a pointer instead of an immediate value, so the coalesce
5425 : * limits can easily changed by the GUC mechanism, and so calling code does
5426 : * not have to check the current configuration. A value of 0 means that no
2473 5427 : * writeback control will be performed.
2606 5428 : */
5429 : void
2606 andres 5430 GIC 3777 : WritebackContextInit(WritebackContext *context, int *max_pending)
2606 andres 5431 ECB : {
2606 andres 5432 GIC 3777 : Assert(*max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
2606 andres 5433 ECB :
2606 andres 5434 GIC 3777 : context->max_pending = max_pending;
2606 andres 5435 CBC 3777 : context->nr_pending = 0;
2606 andres 5436 GIC 3777 : }
5437 :
2606 andres 5438 EUB : /*
5439 : * Add buffer to list of pending writeback requests.
5440 : */
5441 : void
2606 andres 5442 GIC 722286 : ScheduleBufferTagForWriteback(WritebackContext *context, BufferTag *tag)
5443 : {
5444 : PendingWriteback *pending;
5445 :
1 tmunro 5446 GNC 722286 : if (io_direct_flags & IO_DIRECT_DATA)
5447 534 : return;
5448 :
5449 : /*
5450 : * Add buffer to the pending writeback array, unless writeback control is
2606 andres 5451 ECB : * disabled.
5452 : */
2606 andres 5453 CBC 721752 : if (*context->max_pending > 0)
2606 andres 5454 ECB : {
2606 andres 5455 CBC 488257 : Assert(*context->max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
5456 :
5457 488257 : pending = &context->pending_writebacks[context->nr_pending++];
5458 :
2606 andres 5459 GIC 488257 : pending->tag = *tag;
5460 : }
5461 :
5462 : /*
5463 : * Perform pending flushes if the writeback limit is exceeded. This
5464 : * includes the case where previously an item has been added, but control
5465 : * is now disabled.
5466 : */
2606 andres 5467 CBC 721752 : if (context->nr_pending >= *context->max_pending)
2606 andres 5468 GIC 247995 : IssuePendingWritebacks(context);
2606 andres 5469 ECB : }
5470 :
5471 : #define ST_SORT sort_pending_writebacks
5472 : #define ST_ELEMENT_TYPE PendingWriteback
5473 : #define ST_COMPARE(a, b) buffertag_comparator(&a->tag, &b->tag)
5474 : #define ST_SCOPE static
5475 : #define ST_DEFINE
5476 : #include <lib/sort_template.h>
758 tmunro 5477 :
2606 andres 5478 : /*
5479 : * Issue all pending writeback requests, previously scheduled with
5480 : * ScheduleBufferTagForWriteback, to the OS.
5481 : *
5482 : * Because this is only used to improve the OSs IO scheduling we try to never
2606 andres 5483 EUB : * error out - it's just a hint.
5484 : */
5485 : void
2606 andres 5486 GIC 249591 : IssuePendingWritebacks(WritebackContext *context)
5487 : {
5488 : int i;
5489 :
5490 249591 : if (context->nr_pending == 0)
5491 233806 : return;
5492 :
2606 andres 5493 ECB : /*
5494 : * Executing the writes in-order can make them a lot faster, and allows to
5495 : * merge writeback requests to consecutive blocks into larger writebacks.
5496 : */
758 tmunro 5497 GIC 15785 : sort_pending_writebacks(context->pending_writebacks, context->nr_pending);
5498 :
2606 andres 5499 ECB : /*
5500 : * Coalesce neighbouring writes, but nothing else. For that we iterate
5501 : * through the, now sorted, array of pending flushes, and look forward to
5502 : * find all neighbouring (or identical) writes.
5503 : */
2606 andres 5504 GIC 150084 : for (i = 0; i < context->nr_pending; i++)
2606 andres 5505 ECB : {
5506 : PendingWriteback *cur;
5507 : PendingWriteback *next;
5508 : SMgrRelation reln;
5509 : int ahead;
5510 : BufferTag tag;
5511 : RelFileLocator currlocator;
2606 andres 5512 GIC 134299 : Size nblocks = 1;
5513 :
2606 andres 5514 CBC 134299 : cur = &context->pending_writebacks[i];
5515 134299 : tag = cur->tag;
228 rhaas 5516 GNC 134299 : currlocator = BufTagGetRelFileLocator(&tag);
2606 andres 5517 ECB :
5518 : /*
5519 : * Peek ahead, into following writeback requests, to see if they can
5520 : * be combined with the current one.
5521 : */
2606 andres 5522 CBC 485991 : for (ahead = 0; i + ahead + 1 < context->nr_pending; ahead++)
5523 : {
5524 :
5525 470206 : next = &context->pending_writebacks[i + ahead + 1];
2606 andres 5526 ECB :
5527 : /* different file, stop */
228 rhaas 5528 GNC 470206 : if (!RelFileLocatorEquals(currlocator,
5529 394654 : BufTagGetRelFileLocator(&next->tag)) ||
5530 394654 : BufTagGetForkNum(&cur->tag) != BufTagGetForkNum(&next->tag))
2606 andres 5531 ECB : break;
5532 :
5533 : /* ok, block queued twice, skip */
2606 andres 5534 CBC 357539 : if (cur->tag.blockNum == next->tag.blockNum)
5535 550 : continue;
5536 :
5537 : /* only merge consecutive writes */
2606 andres 5538 GIC 356989 : if (cur->tag.blockNum + 1 != next->tag.blockNum)
2606 andres 5539 CBC 5847 : break;
2606 andres 5540 ECB :
2606 andres 5541 CBC 351142 : nblocks++;
2606 andres 5542 GIC 351142 : cur = next;
5543 : }
5544 :
5545 134299 : i += ahead;
5546 :
5547 : /* and finally tell the kernel to write the data to storage */
228 rhaas 5548 GNC 134299 : reln = smgropen(currlocator, InvalidBackendId);
5549 134299 : smgrwriteback(reln, BufTagGetForkNum(&tag), tag.blockNum, nblocks);
5550 : }
5551 :
2606 andres 5552 GIC 15785 : context->nr_pending = 0;
2606 andres 5553 ECB : }
5554 :
5555 :
5556 : /*
5557 : * Implement slower/larger portions of TestForOldSnapshot
2557 kgrittn 5558 : *
5559 : * Smaller/faster portions are put inline, but the entire set of logic is too
2544 5560 : * big for that.
5561 : */
5562 : void
2544 kgrittn 5563 GBC 651 : TestForOldSnapshot_impl(Snapshot snapshot, Relation relation)
2557 kgrittn 5564 EUB : {
2494 kgrittn 5565 GIC 651 : if (RelationAllowsEarlyPruning(relation)
2495 rhaas 5566 GBC 651 : && (snapshot)->whenTaken < GetOldSnapshotThresholdTimestamp())
2557 kgrittn 5567 GIC 3 : ereport(ERROR,
5568 : (errcode(ERRCODE_SNAPSHOT_TOO_OLD),
5569 : errmsg("snapshot too old")));
2557 kgrittn 5570 CBC 648 : }
|
