Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * bufmgr.c
4 : * buffer manager interface routines
5 : *
6 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/storage/buffer/bufmgr.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : /*
16 : * Principal entry points:
17 : *
18 : * ReadBuffer() -- find or create a buffer holding the requested page,
19 : * and pin it so that no one can destroy it while this process
20 : * is using it.
21 : *
22 : * StartReadBuffer() -- as above, with separate wait step
23 : * StartReadBuffers() -- multiple block version
24 : * WaitReadBuffers() -- second step of above
25 : *
26 : * ReleaseBuffer() -- unpin a buffer
27 : *
28 : * MarkBufferDirty() -- mark a pinned buffer's contents as "dirty".
29 : * The disk write is delayed until buffer replacement or checkpoint.
30 : *
31 : * See also these files:
32 : * freelist.c -- chooses victim for buffer replacement
33 : * buf_table.c -- manages the buffer lookup table
34 : */
35 : #include "postgres.h"
36 :
37 : #include <sys/file.h>
38 : #include <unistd.h>
39 :
40 : #include "access/tableam.h"
41 : #include "access/xloginsert.h"
42 : #include "access/xlogutils.h"
43 : #ifdef USE_ASSERT_CHECKING
44 : #include "catalog/pg_tablespace_d.h"
45 : #endif
46 : #include "catalog/storage.h"
47 : #include "catalog/storage_xlog.h"
48 : #include "executor/instrument.h"
49 : #include "lib/binaryheap.h"
50 : #include "miscadmin.h"
51 : #include "pg_trace.h"
52 : #include "pgstat.h"
53 : #include "postmaster/bgwriter.h"
54 : #include "storage/aio.h"
55 : #include "storage/buf_internals.h"
56 : #include "storage/bufmgr.h"
57 : #include "storage/fd.h"
58 : #include "storage/ipc.h"
59 : #include "storage/lmgr.h"
60 : #include "storage/proc.h"
61 : #include "storage/proclist.h"
62 : #include "storage/read_stream.h"
63 : #include "storage/smgr.h"
64 : #include "storage/standby.h"
65 : #include "utils/memdebug.h"
66 : #include "utils/ps_status.h"
67 : #include "utils/rel.h"
68 : #include "utils/resowner.h"
69 : #include "utils/timestamp.h"
70 :
71 :
72 : /* Note: these two macros only work on shared buffers, not local ones! */
73 : #define BufHdrGetBlock(bufHdr) ((Block) (BufferBlocks + ((Size) (bufHdr)->buf_id) * BLCKSZ))
74 : #define BufferGetLSN(bufHdr) (PageGetLSN(BufHdrGetBlock(bufHdr)))
75 :
76 : /* Note: this macro only works on local buffers, not shared ones! */
77 : #define LocalBufHdrGetBlock(bufHdr) \
78 : LocalBufferBlockPointers[-((bufHdr)->buf_id + 2)]
79 :
80 : /* Bits in SyncOneBuffer's return value */
81 : #define BUF_WRITTEN 0x01
82 : #define BUF_REUSABLE 0x02
83 :
84 : #define RELS_BSEARCH_THRESHOLD 20
85 :
86 : /*
87 : * This is the size (in the number of blocks) above which we scan the
88 : * entire buffer pool to remove the buffers for all the pages of relation
89 : * being dropped. For the relations with size below this threshold, we find
90 : * the buffers by doing lookups in BufMapping table.
91 : */
92 : #define BUF_DROP_FULL_SCAN_THRESHOLD (uint64) (NBuffers / 32)
93 :
94 : /*
95 : * This is separated out from PrivateRefCountEntry to allow for copying all
96 : * the data members via struct assignment.
97 : */
98 : typedef struct PrivateRefCountData
99 : {
100 : /*
101 : * How many times has the buffer been pinned by this backend.
102 : */
103 : int32 refcount;
104 :
105 : /*
106 : * Is the buffer locked by this backend? BUFFER_LOCK_UNLOCK indicates that
107 : * the buffer is not locked.
108 : */
109 : BufferLockMode lockmode;
110 : } PrivateRefCountData;
111 :
112 : typedef struct PrivateRefCountEntry
113 : {
114 : /*
115 : * Note that this needs to be same as the entry's corresponding
116 : * PrivateRefCountArrayKeys[i], if the entry is stored in the array. We
117 : * store it in both places as this is used for the hashtable key and
118 : * because it is more convenient (passing around a PrivateRefCountEntry
119 : * suffices to identify the buffer) and faster (checking the keys array is
120 : * faster when checking many entries, checking the entry is faster if just
121 : * checking a single entry).
122 : */
123 : Buffer buffer;
124 :
125 : PrivateRefCountData data;
126 : } PrivateRefCountEntry;
127 :
128 : /* 64 bytes, about the size of a cache line on common systems */
129 : #define REFCOUNT_ARRAY_ENTRIES 8
130 :
131 : /*
132 : * Status of buffers to checkpoint for a particular tablespace, used
133 : * internally in BufferSync.
134 : */
135 : typedef struct CkptTsStatus
136 : {
137 : /* oid of the tablespace */
138 : Oid tsId;
139 :
140 : /*
141 : * Checkpoint progress for this tablespace. To make progress comparable
142 : * between tablespaces the progress is, for each tablespace, measured as a
143 : * number between 0 and the total number of to-be-checkpointed pages. Each
144 : * page checkpointed in this tablespace increments this space's progress
145 : * by progress_slice.
146 : */
147 : float8 progress;
148 : float8 progress_slice;
149 :
150 : /* number of to-be checkpointed pages in this tablespace */
151 : int num_to_scan;
152 : /* already processed pages in this tablespace */
153 : int num_scanned;
154 :
155 : /* current offset in CkptBufferIds for this tablespace */
156 : int index;
157 : } CkptTsStatus;
158 :
159 : /*
160 : * Type for array used to sort SMgrRelations
161 : *
162 : * FlushRelationsAllBuffers shares the same comparator function with
163 : * DropRelationsAllBuffers. Pointer to this struct and RelFileLocator must be
164 : * compatible.
165 : */
166 : typedef struct SMgrSortArray
167 : {
168 : RelFileLocator rlocator; /* This must be the first member */
169 : SMgrRelation srel;
170 : } SMgrSortArray;
171 :
172 : /* GUC variables */
173 : bool zero_damaged_pages = false;
174 : int bgwriter_lru_maxpages = 100;
175 : double bgwriter_lru_multiplier = 2.0;
176 : bool track_io_timing = false;
177 :
178 : /*
179 : * How many buffers PrefetchBuffer callers should try to stay ahead of their
180 : * ReadBuffer calls by. Zero means "never prefetch". This value is only used
181 : * for buffers not belonging to tablespaces that have their
182 : * effective_io_concurrency parameter set.
183 : */
184 : int effective_io_concurrency = DEFAULT_EFFECTIVE_IO_CONCURRENCY;
185 :
186 : /*
187 : * Like effective_io_concurrency, but used by maintenance code paths that might
188 : * benefit from a higher setting because they work on behalf of many sessions.
189 : * Overridden by the tablespace setting of the same name.
190 : */
191 : int maintenance_io_concurrency = DEFAULT_MAINTENANCE_IO_CONCURRENCY;
192 :
193 : /*
194 : * Limit on how many blocks should be handled in single I/O operations.
195 : * StartReadBuffers() callers should respect it, as should other operations
196 : * that call smgr APIs directly. It is computed as the minimum of underlying
197 : * GUCs io_combine_limit_guc and io_max_combine_limit.
198 : */
199 : int io_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
200 : int io_combine_limit_guc = DEFAULT_IO_COMBINE_LIMIT;
201 : int io_max_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
202 :
203 : /*
204 : * GUC variables about triggering kernel writeback for buffers written; OS
205 : * dependent defaults are set via the GUC mechanism.
206 : */
207 : int checkpoint_flush_after = DEFAULT_CHECKPOINT_FLUSH_AFTER;
208 : int bgwriter_flush_after = DEFAULT_BGWRITER_FLUSH_AFTER;
209 : int backend_flush_after = DEFAULT_BACKEND_FLUSH_AFTER;
210 :
211 : /* local state for LockBufferForCleanup */
212 : static BufferDesc *PinCountWaitBuf = NULL;
213 :
214 : /*
215 : * Backend-Private refcount management:
216 : *
217 : * Each buffer also has a private refcount that keeps track of the number of
218 : * times the buffer is pinned in the current process. This is so that the
219 : * shared refcount needs to be modified only once if a buffer is pinned more
220 : * than once by an individual backend. It's also used to check that no
221 : * buffers are still pinned at the end of transactions and when exiting. We
222 : * also use this mechanism to track whether this backend has a buffer locked,
223 : * and, if so, in what mode.
224 : *
225 : *
226 : * To avoid - as we used to - requiring an array with NBuffers entries to keep
227 : * track of local buffers, we use a small sequentially searched array
228 : * (PrivateRefCountArrayKeys, with the corresponding data stored in
229 : * PrivateRefCountArray) and an overflow hash table (PrivateRefCountHash) to
230 : * keep track of backend local pins.
231 : *
232 : * Until no more than REFCOUNT_ARRAY_ENTRIES buffers are pinned at once, all
233 : * refcounts are kept track of in the array; after that, new array entries
234 : * displace old ones into the hash table. That way a frequently used entry
235 : * can't get "stuck" in the hashtable while infrequent ones clog the array.
236 : *
237 : * Note that in most scenarios the number of pinned buffers will not exceed
238 : * REFCOUNT_ARRAY_ENTRIES.
239 : *
240 : *
241 : * To enter a buffer into the refcount tracking mechanism first reserve a free
242 : * entry using ReservePrivateRefCountEntry() and then later, if necessary,
243 : * fill it with NewPrivateRefCountEntry(). That split lets us avoid doing
244 : * memory allocations in NewPrivateRefCountEntry() which can be important
245 : * because in some scenarios it's called with a spinlock held...
246 : */
247 : static Buffer PrivateRefCountArrayKeys[REFCOUNT_ARRAY_ENTRIES];
248 : static struct PrivateRefCountEntry PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES];
249 : static HTAB *PrivateRefCountHash = NULL;
250 : static int32 PrivateRefCountOverflowed = 0;
251 : static uint32 PrivateRefCountClock = 0;
252 : static int ReservedRefCountSlot = -1;
253 : static int PrivateRefCountEntryLast = -1;
254 :
255 : static uint32 MaxProportionalPins;
256 :
257 : static void ReservePrivateRefCountEntry(void);
258 : static PrivateRefCountEntry *NewPrivateRefCountEntry(Buffer buffer);
259 : static PrivateRefCountEntry *GetPrivateRefCountEntry(Buffer buffer, bool do_move);
260 : static inline int32 GetPrivateRefCount(Buffer buffer);
261 : static void ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref);
262 :
263 : /* ResourceOwner callbacks to hold in-progress I/Os and buffer pins */
264 : static void ResOwnerReleaseBufferIO(Datum res);
265 : static char *ResOwnerPrintBufferIO(Datum res);
266 : static void ResOwnerReleaseBuffer(Datum res);
267 : static char *ResOwnerPrintBuffer(Datum res);
268 :
269 : const ResourceOwnerDesc buffer_io_resowner_desc =
270 : {
271 : .name = "buffer io",
272 : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
273 : .release_priority = RELEASE_PRIO_BUFFER_IOS,
274 : .ReleaseResource = ResOwnerReleaseBufferIO,
275 : .DebugPrint = ResOwnerPrintBufferIO
276 : };
277 :
278 : const ResourceOwnerDesc buffer_resowner_desc =
279 : {
280 : .name = "buffer",
281 : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
282 : .release_priority = RELEASE_PRIO_BUFFER_PINS,
283 : .ReleaseResource = ResOwnerReleaseBuffer,
284 : .DebugPrint = ResOwnerPrintBuffer
285 : };
286 :
287 : /*
288 : * Ensure that the PrivateRefCountArray has sufficient space to store one more
289 : * entry. This has to be called before using NewPrivateRefCountEntry() to fill
290 : * a new entry - but it's perfectly fine to not use a reserved entry.
291 : */
292 : static void
293 130778226 : ReservePrivateRefCountEntry(void)
294 : {
295 : /* Already reserved (or freed), nothing to do */
296 130778226 : if (ReservedRefCountSlot != -1)
297 122442172 : return;
298 :
299 : /*
300 : * First search for a free entry the array, that'll be sufficient in the
301 : * majority of cases.
302 : */
303 : {
304 : int i;
305 :
306 75024486 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
307 : {
308 66688432 : if (PrivateRefCountArrayKeys[i] == InvalidBuffer)
309 : {
310 48855114 : ReservedRefCountSlot = i;
311 :
312 : /*
313 : * We could return immediately, but iterating till the end of
314 : * the array allows compiler-autovectorization.
315 : */
316 : }
317 : }
318 :
319 8336054 : if (ReservedRefCountSlot != -1)
320 7969706 : return;
321 : }
322 :
323 : /*
324 : * No luck. All array entries are full. Move one array entry into the hash
325 : * table.
326 : */
327 : {
328 : /*
329 : * Move entry from the current clock position in the array into the
330 : * hashtable. Use that slot.
331 : */
332 : int victim_slot;
333 : PrivateRefCountEntry *victim_entry;
334 : PrivateRefCountEntry *hashent;
335 : bool found;
336 :
337 : /* select victim slot */
338 366348 : victim_slot = PrivateRefCountClock++ % REFCOUNT_ARRAY_ENTRIES;
339 366348 : victim_entry = &PrivateRefCountArray[victim_slot];
340 366348 : ReservedRefCountSlot = victim_slot;
341 :
342 : /* Better be used, otherwise we shouldn't get here. */
343 : Assert(PrivateRefCountArrayKeys[victim_slot] != InvalidBuffer);
344 : Assert(PrivateRefCountArray[victim_slot].buffer != InvalidBuffer);
345 : Assert(PrivateRefCountArrayKeys[victim_slot] == PrivateRefCountArray[victim_slot].buffer);
346 :
347 : /* enter victim array entry into hashtable */
348 366348 : hashent = hash_search(PrivateRefCountHash,
349 366348 : &PrivateRefCountArrayKeys[victim_slot],
350 : HASH_ENTER,
351 : &found);
352 : Assert(!found);
353 : /* move data from the entry in the array to the hash entry */
354 366348 : hashent->data = victim_entry->data;
355 :
356 : /* clear the now free array slot */
357 366348 : PrivateRefCountArrayKeys[victim_slot] = InvalidBuffer;
358 366348 : victim_entry->buffer = InvalidBuffer;
359 :
360 : /* clear the whole data member, just for future proofing */
361 366348 : memset(&victim_entry->data, 0, sizeof(victim_entry->data));
362 366348 : victim_entry->data.refcount = 0;
363 366348 : victim_entry->data.lockmode = BUFFER_LOCK_UNLOCK;
364 :
365 366348 : PrivateRefCountOverflowed++;
366 : }
367 : }
368 :
369 : /*
370 : * Fill a previously reserved refcount entry.
371 : */
372 : static PrivateRefCountEntry *
373 118571244 : NewPrivateRefCountEntry(Buffer buffer)
374 : {
375 : PrivateRefCountEntry *res;
376 :
377 : /* only allowed to be called when a reservation has been made */
378 : Assert(ReservedRefCountSlot != -1);
379 :
380 : /* use up the reserved entry */
381 118571244 : res = &PrivateRefCountArray[ReservedRefCountSlot];
382 :
383 : /* and fill it */
384 118571244 : PrivateRefCountArrayKeys[ReservedRefCountSlot] = buffer;
385 118571244 : res->buffer = buffer;
386 118571244 : res->data.refcount = 0;
387 118571244 : res->data.lockmode = BUFFER_LOCK_UNLOCK;
388 :
389 : /* update cache for the next lookup */
390 118571244 : PrivateRefCountEntryLast = ReservedRefCountSlot;
391 :
392 118571244 : ReservedRefCountSlot = -1;
393 :
394 118571244 : return res;
395 : }
396 :
397 : /*
398 : * Slow-path for GetPrivateRefCountEntry(). This is big enough to not be worth
399 : * inlining. This particularly seems to be true if the compiler is capable of
400 : * auto-vectorizing the code, as that imposes additional stack-alignment
401 : * requirements etc.
402 : */
403 : static pg_noinline PrivateRefCountEntry *
404 147105824 : GetPrivateRefCountEntrySlow(Buffer buffer, bool do_move)
405 : {
406 : PrivateRefCountEntry *res;
407 147105824 : int match = -1;
408 : int i;
409 :
410 : /*
411 : * First search for references in the array, that'll be sufficient in the
412 : * majority of cases.
413 : */
414 1323952416 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
415 : {
416 1176846592 : if (PrivateRefCountArrayKeys[i] == buffer)
417 : {
418 32615310 : match = i;
419 : /* see ReservePrivateRefCountEntry() for why we don't return */
420 : }
421 : }
422 :
423 147105824 : if (likely(match != -1))
424 : {
425 : /* update cache for the next lookup */
426 32615310 : PrivateRefCountEntryLast = match;
427 :
428 32615310 : return &PrivateRefCountArray[match];
429 : }
430 :
431 : /*
432 : * By here we know that the buffer, if already pinned, isn't residing in
433 : * the array.
434 : *
435 : * Only look up the buffer in the hashtable if we've previously overflowed
436 : * into it.
437 : */
438 114490514 : if (PrivateRefCountOverflowed == 0)
439 113592028 : return NULL;
440 :
441 898486 : res = hash_search(PrivateRefCountHash, &buffer, HASH_FIND, NULL);
442 :
443 898486 : if (res == NULL)
444 442170 : return NULL;
445 456316 : else if (!do_move)
446 : {
447 : /* caller doesn't want us to move the hash entry into the array */
448 262322 : return res;
449 : }
450 : else
451 : {
452 : /* move buffer from hashtable into the free array slot */
453 : bool found;
454 : PrivateRefCountEntry *free;
455 :
456 : /* Ensure there's a free array slot */
457 193994 : ReservePrivateRefCountEntry();
458 :
459 : /* Use up the reserved slot */
460 : Assert(ReservedRefCountSlot != -1);
461 193994 : free = &PrivateRefCountArray[ReservedRefCountSlot];
462 : Assert(PrivateRefCountArrayKeys[ReservedRefCountSlot] == free->buffer);
463 : Assert(free->buffer == InvalidBuffer);
464 :
465 : /* and fill it */
466 193994 : free->buffer = buffer;
467 193994 : free->data = res->data;
468 193994 : PrivateRefCountArrayKeys[ReservedRefCountSlot] = buffer;
469 : /* update cache for the next lookup */
470 193994 : PrivateRefCountEntryLast = match;
471 :
472 193994 : ReservedRefCountSlot = -1;
473 :
474 :
475 : /* delete from hashtable */
476 193994 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
477 : Assert(found);
478 : Assert(PrivateRefCountOverflowed > 0);
479 193994 : PrivateRefCountOverflowed--;
480 :
481 193994 : return free;
482 : }
483 : }
484 :
485 : /*
486 : * Return the PrivateRefCount entry for the passed buffer.
487 : *
488 : * Returns NULL if a buffer doesn't have a refcount entry. Otherwise, if
489 : * do_move is true, and the entry resides in the hashtable the entry is
490 : * optimized for frequent access by moving it to the array.
491 : */
492 : static inline PrivateRefCountEntry *
493 624120410 : GetPrivateRefCountEntry(Buffer buffer, bool do_move)
494 : {
495 : Assert(BufferIsValid(buffer));
496 : Assert(!BufferIsLocal(buffer));
497 :
498 : /*
499 : * It's very common to look up the same buffer repeatedly. To make that
500 : * fast, we have a one-entry cache.
501 : *
502 : * In contrast to the loop in GetPrivateRefCountEntrySlow(), here it
503 : * faster to check PrivateRefCountArray[].buffer, as in the case of a hit
504 : * fewer addresses are computed and fewer cachelines are accessed. Whereas
505 : * in GetPrivateRefCountEntrySlow()'s case, checking
506 : * PrivateRefCountArrayKeys saves a lot of memory accesses.
507 : */
508 624120410 : if (likely(PrivateRefCountEntryLast != -1) &&
509 623994768 : likely(PrivateRefCountArray[PrivateRefCountEntryLast].buffer == buffer))
510 : {
511 477014586 : return &PrivateRefCountArray[PrivateRefCountEntryLast];
512 : }
513 :
514 : /*
515 : * The code for the cached lookup is small enough to be worth inlining
516 : * into the caller. In the miss case however, that empirically doesn't
517 : * seem worth it.
518 : */
519 147105824 : return GetPrivateRefCountEntrySlow(buffer, do_move);
520 : }
521 :
522 : /*
523 : * Returns how many times the passed buffer is pinned by this backend.
524 : *
525 : * Only works for shared memory buffers!
526 : */
527 : static inline int32
528 5557426 : GetPrivateRefCount(Buffer buffer)
529 : {
530 : PrivateRefCountEntry *ref;
531 :
532 : Assert(BufferIsValid(buffer));
533 : Assert(!BufferIsLocal(buffer));
534 :
535 : /*
536 : * Not moving the entry - that's ok for the current users, but we might
537 : * want to change this one day.
538 : */
539 5557426 : ref = GetPrivateRefCountEntry(buffer, false);
540 :
541 5557426 : if (ref == NULL)
542 68 : return 0;
543 5557358 : return ref->data.refcount;
544 : }
545 :
546 : /*
547 : * Release resources used to track the reference count of a buffer which we no
548 : * longer have pinned and don't want to pin again immediately.
549 : */
550 : static void
551 118571244 : ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref)
552 : {
553 : Assert(ref->data.refcount == 0);
554 : Assert(ref->data.lockmode == BUFFER_LOCK_UNLOCK);
555 :
556 118571244 : if (ref >= &PrivateRefCountArray[0] &&
557 : ref < &PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES])
558 : {
559 118398890 : ref->buffer = InvalidBuffer;
560 118398890 : PrivateRefCountArrayKeys[ref - PrivateRefCountArray] = InvalidBuffer;
561 :
562 :
563 : /*
564 : * Mark the just used entry as reserved - in many scenarios that
565 : * allows us to avoid ever having to search the array/hash for free
566 : * entries.
567 : */
568 118398890 : ReservedRefCountSlot = ref - PrivateRefCountArray;
569 : }
570 : else
571 : {
572 : bool found;
573 172354 : Buffer buffer = ref->buffer;
574 :
575 172354 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
576 : Assert(found);
577 : Assert(PrivateRefCountOverflowed > 0);
578 172354 : PrivateRefCountOverflowed--;
579 : }
580 118571244 : }
581 :
582 : /*
583 : * BufferIsPinned
584 : * True iff the buffer is pinned (also checks for valid buffer number).
585 : *
586 : * NOTE: what we check here is that *this* backend holds a pin on
587 : * the buffer. We do not care whether some other backend does.
588 : */
589 : #define BufferIsPinned(bufnum) \
590 : ( \
591 : !BufferIsValid(bufnum) ? \
592 : false \
593 : : \
594 : BufferIsLocal(bufnum) ? \
595 : (LocalRefCount[-(bufnum) - 1] > 0) \
596 : : \
597 : (GetPrivateRefCount(bufnum) > 0) \
598 : )
599 :
600 :
601 : static Buffer ReadBuffer_common(Relation rel,
602 : SMgrRelation smgr, char smgr_persistence,
603 : ForkNumber forkNum, BlockNumber blockNum,
604 : ReadBufferMode mode, BufferAccessStrategy strategy);
605 : static BlockNumber ExtendBufferedRelCommon(BufferManagerRelation bmr,
606 : ForkNumber fork,
607 : BufferAccessStrategy strategy,
608 : uint32 flags,
609 : uint32 extend_by,
610 : BlockNumber extend_upto,
611 : Buffer *buffers,
612 : uint32 *extended_by);
613 : static BlockNumber ExtendBufferedRelShared(BufferManagerRelation bmr,
614 : ForkNumber fork,
615 : BufferAccessStrategy strategy,
616 : uint32 flags,
617 : uint32 extend_by,
618 : BlockNumber extend_upto,
619 : Buffer *buffers,
620 : uint32 *extended_by);
621 : static bool PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
622 : bool skip_if_not_valid);
623 : static void PinBuffer_Locked(BufferDesc *buf);
624 : static void UnpinBuffer(BufferDesc *buf);
625 : static void UnpinBufferNoOwner(BufferDesc *buf);
626 : static void BufferSync(int flags);
627 : static int SyncOneBuffer(int buf_id, bool skip_recently_used,
628 : WritebackContext *wb_context);
629 : static void WaitIO(BufferDesc *buf);
630 : static void AbortBufferIO(Buffer buffer);
631 : static void shared_buffer_write_error_callback(void *arg);
632 : static void local_buffer_write_error_callback(void *arg);
633 : static inline BufferDesc *BufferAlloc(SMgrRelation smgr,
634 : char relpersistence,
635 : ForkNumber forkNum,
636 : BlockNumber blockNum,
637 : BufferAccessStrategy strategy,
638 : bool *foundPtr, IOContext io_context);
639 : static bool AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress);
640 : static void CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete);
641 : static Buffer GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context);
642 : static void FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
643 : IOObject io_object, IOContext io_context);
644 : static void FlushBuffer(BufferDesc *buf, SMgrRelation reln,
645 : IOObject io_object, IOContext io_context);
646 : static void FindAndDropRelationBuffers(RelFileLocator rlocator,
647 : ForkNumber forkNum,
648 : BlockNumber nForkBlock,
649 : BlockNumber firstDelBlock);
650 : static void RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
651 : RelFileLocator dstlocator,
652 : ForkNumber forkNum, bool permanent);
653 : static void AtProcExit_Buffers(int code, Datum arg);
654 : static void CheckForBufferLeaks(void);
655 : #ifdef USE_ASSERT_CHECKING
656 : static void AssertNotCatalogBufferLock(Buffer buffer, BufferLockMode mode);
657 : #endif
658 : static int rlocator_comparator(const void *p1, const void *p2);
659 : static inline int buffertag_comparator(const BufferTag *ba, const BufferTag *bb);
660 : static inline int ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b);
661 : static int ts_ckpt_progress_comparator(Datum a, Datum b, void *arg);
662 :
663 : static void BufferLockAcquire(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode);
664 : static void BufferLockUnlock(Buffer buffer, BufferDesc *buf_hdr);
665 : static bool BufferLockConditional(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode);
666 : static bool BufferLockHeldByMeInMode(BufferDesc *buf_hdr, BufferLockMode mode);
667 : static bool BufferLockHeldByMe(BufferDesc *buf_hdr);
668 : static inline void BufferLockDisown(Buffer buffer, BufferDesc *buf_hdr);
669 : static inline int BufferLockDisownInternal(Buffer buffer, BufferDesc *buf_hdr);
670 : static inline bool BufferLockAttempt(BufferDesc *buf_hdr, BufferLockMode mode);
671 : static void BufferLockQueueSelf(BufferDesc *buf_hdr, BufferLockMode mode);
672 : static void BufferLockDequeueSelf(BufferDesc *buf_hdr);
673 : static void BufferLockWakeup(BufferDesc *buf_hdr, bool unlocked);
674 : static void BufferLockProcessRelease(BufferDesc *buf_hdr, BufferLockMode mode, uint64 lockstate);
675 : static inline uint64 BufferLockReleaseSub(BufferLockMode mode);
676 :
677 :
678 : /*
679 : * Implementation of PrefetchBuffer() for shared buffers.
680 : */
681 : PrefetchBufferResult
682 63736 : PrefetchSharedBuffer(SMgrRelation smgr_reln,
683 : ForkNumber forkNum,
684 : BlockNumber blockNum)
685 : {
686 63736 : PrefetchBufferResult result = {InvalidBuffer, false};
687 : BufferTag newTag; /* identity of requested block */
688 : uint32 newHash; /* hash value for newTag */
689 : LWLock *newPartitionLock; /* buffer partition lock for it */
690 : int buf_id;
691 :
692 : Assert(BlockNumberIsValid(blockNum));
693 :
694 : /* create a tag so we can lookup the buffer */
695 63736 : InitBufferTag(&newTag, &smgr_reln->smgr_rlocator.locator,
696 : forkNum, blockNum);
697 :
698 : /* determine its hash code and partition lock ID */
699 63736 : newHash = BufTableHashCode(&newTag);
700 63736 : newPartitionLock = BufMappingPartitionLock(newHash);
701 :
702 : /* see if the block is in the buffer pool already */
703 63736 : LWLockAcquire(newPartitionLock, LW_SHARED);
704 63736 : buf_id = BufTableLookup(&newTag, newHash);
705 63736 : LWLockRelease(newPartitionLock);
706 :
707 : /* If not in buffers, initiate prefetch */
708 63736 : if (buf_id < 0)
709 : {
710 : #ifdef USE_PREFETCH
711 : /*
712 : * Try to initiate an asynchronous read. This returns false in
713 : * recovery if the relation file doesn't exist.
714 : */
715 35164 : if ((io_direct_flags & IO_DIRECT_DATA) == 0 &&
716 17358 : smgrprefetch(smgr_reln, forkNum, blockNum, 1))
717 : {
718 17358 : result.initiated_io = true;
719 : }
720 : #endif /* USE_PREFETCH */
721 : }
722 : else
723 : {
724 : /*
725 : * Report the buffer it was in at that time. The caller may be able
726 : * to avoid a buffer table lookup, but it's not pinned and it must be
727 : * rechecked!
728 : */
729 45930 : result.recent_buffer = buf_id + 1;
730 : }
731 :
732 : /*
733 : * If the block *is* in buffers, we do nothing. This is not really ideal:
734 : * the block might be just about to be evicted, which would be stupid
735 : * since we know we are going to need it soon. But the only easy answer
736 : * is to bump the usage_count, which does not seem like a great solution:
737 : * when the caller does ultimately touch the block, usage_count would get
738 : * bumped again, resulting in too much favoritism for blocks that are
739 : * involved in a prefetch sequence. A real fix would involve some
740 : * additional per-buffer state, and it's not clear that there's enough of
741 : * a problem to justify that.
742 : */
743 :
744 63736 : return result;
745 : }
746 :
747 : /*
748 : * PrefetchBuffer -- initiate asynchronous read of a block of a relation
749 : *
750 : * This is named by analogy to ReadBuffer but doesn't actually allocate a
751 : * buffer. Instead it tries to ensure that a future ReadBuffer for the given
752 : * block will not be delayed by the I/O. Prefetching is optional.
753 : *
754 : * There are three possible outcomes:
755 : *
756 : * 1. If the block is already cached, the result includes a valid buffer that
757 : * could be used by the caller to avoid the need for a later buffer lookup, but
758 : * it's not pinned, so the caller must recheck it.
759 : *
760 : * 2. If the kernel has been asked to initiate I/O, the initiated_io member is
761 : * true. Currently there is no way to know if the data was already cached by
762 : * the kernel and therefore didn't really initiate I/O, and no way to know when
763 : * the I/O completes other than using synchronous ReadBuffer().
764 : *
765 : * 3. Otherwise, the buffer wasn't already cached by PostgreSQL, and
766 : * USE_PREFETCH is not defined (this build doesn't support prefetching due to
767 : * lack of a kernel facility), direct I/O is enabled, or the underlying
768 : * relation file wasn't found and we are in recovery. (If the relation file
769 : * wasn't found and we are not in recovery, an error is raised).
770 : */
771 : PrefetchBufferResult
772 42184 : PrefetchBuffer(Relation reln, ForkNumber forkNum, BlockNumber blockNum)
773 : {
774 : Assert(RelationIsValid(reln));
775 : Assert(BlockNumberIsValid(blockNum));
776 :
777 42184 : if (RelationUsesLocalBuffers(reln))
778 : {
779 : /* see comments in ReadBufferExtended */
780 1566 : if (RELATION_IS_OTHER_TEMP(reln))
781 0 : ereport(ERROR,
782 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
783 : errmsg("cannot access temporary tables of other sessions")));
784 :
785 : /* pass it off to localbuf.c */
786 1566 : return PrefetchLocalBuffer(RelationGetSmgr(reln), forkNum, blockNum);
787 : }
788 : else
789 : {
790 : /* pass it to the shared buffer version */
791 40618 : return PrefetchSharedBuffer(RelationGetSmgr(reln), forkNum, blockNum);
792 : }
793 : }
794 :
795 : /*
796 : * ReadRecentBuffer -- try to pin a block in a recently observed buffer
797 : *
798 : * Compared to ReadBuffer(), this avoids a buffer mapping lookup when it's
799 : * successful. Return true if the buffer is valid and still has the expected
800 : * tag. In that case, the buffer is pinned and the usage count is bumped.
801 : */
802 : bool
803 9272 : ReadRecentBuffer(RelFileLocator rlocator, ForkNumber forkNum, BlockNumber blockNum,
804 : Buffer recent_buffer)
805 : {
806 : BufferDesc *bufHdr;
807 : BufferTag tag;
808 : uint64 buf_state;
809 :
810 : Assert(BufferIsValid(recent_buffer));
811 :
812 9272 : ResourceOwnerEnlarge(CurrentResourceOwner);
813 9272 : ReservePrivateRefCountEntry();
814 9272 : InitBufferTag(&tag, &rlocator, forkNum, blockNum);
815 :
816 9272 : if (BufferIsLocal(recent_buffer))
817 : {
818 64 : int b = -recent_buffer - 1;
819 :
820 64 : bufHdr = GetLocalBufferDescriptor(b);
821 64 : buf_state = pg_atomic_read_u64(&bufHdr->state);
822 :
823 : /* Is it still valid and holding the right tag? */
824 64 : if ((buf_state & BM_VALID) && BufferTagsEqual(&tag, &bufHdr->tag))
825 : {
826 64 : PinLocalBuffer(bufHdr, true);
827 :
828 64 : pgBufferUsage.local_blks_hit++;
829 :
830 64 : return true;
831 : }
832 : }
833 : else
834 : {
835 9208 : bufHdr = GetBufferDescriptor(recent_buffer - 1);
836 :
837 : /*
838 : * Is it still valid and holding the right tag? We do an unlocked tag
839 : * comparison first, to make it unlikely that we'll increment the
840 : * usage counter of the wrong buffer, if someone calls us with a very
841 : * out of date recent_buffer. Then we'll check it again if we get the
842 : * pin.
843 : */
844 18332 : if (BufferTagsEqual(&tag, &bufHdr->tag) &&
845 9124 : PinBuffer(bufHdr, NULL, true))
846 : {
847 9112 : if (BufferTagsEqual(&tag, &bufHdr->tag))
848 : {
849 9112 : pgBufferUsage.shared_blks_hit++;
850 9112 : return true;
851 : }
852 0 : UnpinBuffer(bufHdr);
853 : }
854 : }
855 :
856 96 : return false;
857 : }
858 :
859 : /*
860 : * ReadBuffer -- a shorthand for ReadBufferExtended, for reading from main
861 : * fork with RBM_NORMAL mode and default strategy.
862 : */
863 : Buffer
864 87404562 : ReadBuffer(Relation reln, BlockNumber blockNum)
865 : {
866 87404562 : return ReadBufferExtended(reln, MAIN_FORKNUM, blockNum, RBM_NORMAL, NULL);
867 : }
868 :
869 : /*
870 : * ReadBufferExtended -- returns a buffer containing the requested
871 : * block of the requested relation. If the blknum
872 : * requested is P_NEW, extend the relation file and
873 : * allocate a new block. (Caller is responsible for
874 : * ensuring that only one backend tries to extend a
875 : * relation at the same time!)
876 : *
877 : * Returns: the buffer number for the buffer containing
878 : * the block read. The returned buffer has been pinned.
879 : * Does not return on error --- elog's instead.
880 : *
881 : * Assume when this function is called, that reln has been opened already.
882 : *
883 : * In RBM_NORMAL mode, the page is read from disk, and the page header is
884 : * validated. An error is thrown if the page header is not valid. (But
885 : * note that an all-zero page is considered "valid"; see
886 : * PageIsVerified().)
887 : *
888 : * RBM_ZERO_ON_ERROR is like the normal mode, but if the page header is not
889 : * valid, the page is zeroed instead of throwing an error. This is intended
890 : * for non-critical data, where the caller is prepared to repair errors.
891 : *
892 : * In RBM_ZERO_AND_LOCK mode, if the page isn't in buffer cache already, it's
893 : * filled with zeros instead of reading it from disk. Useful when the caller
894 : * is going to fill the page from scratch, since this saves I/O and avoids
895 : * unnecessary failure if the page-on-disk has corrupt page headers.
896 : * The page is returned locked to ensure that the caller has a chance to
897 : * initialize the page before it's made visible to others.
898 : * Caution: do not use this mode to read a page that is beyond the relation's
899 : * current physical EOF; that is likely to cause problems in md.c when
900 : * the page is modified and written out. P_NEW is OK, though.
901 : *
902 : * RBM_ZERO_AND_CLEANUP_LOCK is the same as RBM_ZERO_AND_LOCK, but acquires
903 : * a cleanup-strength lock on the page.
904 : *
905 : * RBM_NORMAL_NO_LOG mode is treated the same as RBM_NORMAL here.
906 : *
907 : * If strategy is not NULL, a nondefault buffer access strategy is used.
908 : * See buffer/README for details.
909 : */
910 : inline Buffer
911 104954412 : ReadBufferExtended(Relation reln, ForkNumber forkNum, BlockNumber blockNum,
912 : ReadBufferMode mode, BufferAccessStrategy strategy)
913 : {
914 : Buffer buf;
915 :
916 : /*
917 : * Reject attempts to read non-local temporary relations; we would be
918 : * likely to get wrong data since we have no visibility into the owning
919 : * session's local buffers.
920 : */
921 104954412 : if (RELATION_IS_OTHER_TEMP(reln))
922 0 : ereport(ERROR,
923 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
924 : errmsg("cannot access temporary tables of other sessions")));
925 :
926 : /*
927 : * Read the buffer, and update pgstat counters to reflect a cache hit or
928 : * miss.
929 : */
930 104954412 : buf = ReadBuffer_common(reln, RelationGetSmgr(reln), 0,
931 : forkNum, blockNum, mode, strategy);
932 :
933 104954366 : return buf;
934 : }
935 :
936 :
937 : /*
938 : * ReadBufferWithoutRelcache -- like ReadBufferExtended, but doesn't require
939 : * a relcache entry for the relation.
940 : *
941 : * Pass permanent = true for a RELPERSISTENCE_PERMANENT relation, and
942 : * permanent = false for a RELPERSISTENCE_UNLOGGED relation. This function
943 : * cannot be used for temporary relations (and making that work might be
944 : * difficult, unless we only want to read temporary relations for our own
945 : * ProcNumber).
946 : */
947 : Buffer
948 11519770 : ReadBufferWithoutRelcache(RelFileLocator rlocator, ForkNumber forkNum,
949 : BlockNumber blockNum, ReadBufferMode mode,
950 : BufferAccessStrategy strategy, bool permanent)
951 : {
952 11519770 : SMgrRelation smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
953 :
954 11519770 : return ReadBuffer_common(NULL, smgr,
955 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
956 : forkNum, blockNum,
957 : mode, strategy);
958 : }
959 :
960 : /*
961 : * Convenience wrapper around ExtendBufferedRelBy() extending by one block.
962 : */
963 : Buffer
964 92228 : ExtendBufferedRel(BufferManagerRelation bmr,
965 : ForkNumber forkNum,
966 : BufferAccessStrategy strategy,
967 : uint32 flags)
968 : {
969 : Buffer buf;
970 92228 : uint32 extend_by = 1;
971 :
972 92228 : ExtendBufferedRelBy(bmr, forkNum, strategy, flags, extend_by,
973 : &buf, &extend_by);
974 :
975 92228 : return buf;
976 : }
977 :
978 : /*
979 : * Extend relation by multiple blocks.
980 : *
981 : * Tries to extend the relation by extend_by blocks. Depending on the
982 : * availability of resources the relation may end up being extended by a
983 : * smaller number of pages (unless an error is thrown, always by at least one
984 : * page). *extended_by is updated to the number of pages the relation has been
985 : * extended to.
986 : *
987 : * buffers needs to be an array that is at least extend_by long. Upon
988 : * completion, the first extend_by array elements will point to a pinned
989 : * buffer.
990 : *
991 : * If EB_LOCK_FIRST is part of flags, the first returned buffer is
992 : * locked. This is useful for callers that want a buffer that is guaranteed to
993 : * be empty.
994 : */
995 : BlockNumber
996 323910 : ExtendBufferedRelBy(BufferManagerRelation bmr,
997 : ForkNumber fork,
998 : BufferAccessStrategy strategy,
999 : uint32 flags,
1000 : uint32 extend_by,
1001 : Buffer *buffers,
1002 : uint32 *extended_by)
1003 : {
1004 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
1005 : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
1006 : Assert(extend_by > 0);
1007 :
1008 323910 : if (bmr.relpersistence == '\0')
1009 323910 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
1010 :
1011 323910 : return ExtendBufferedRelCommon(bmr, fork, strategy, flags,
1012 : extend_by, InvalidBlockNumber,
1013 : buffers, extended_by);
1014 : }
1015 :
1016 : /*
1017 : * Extend the relation so it is at least extend_to blocks large, return buffer
1018 : * (extend_to - 1).
1019 : *
1020 : * This is useful for callers that want to write a specific page, regardless
1021 : * of the current size of the relation (e.g. useful for visibilitymap and for
1022 : * crash recovery).
1023 : */
1024 : Buffer
1025 104100 : ExtendBufferedRelTo(BufferManagerRelation bmr,
1026 : ForkNumber fork,
1027 : BufferAccessStrategy strategy,
1028 : uint32 flags,
1029 : BlockNumber extend_to,
1030 : ReadBufferMode mode)
1031 : {
1032 : BlockNumber current_size;
1033 104100 : uint32 extended_by = 0;
1034 104100 : Buffer buffer = InvalidBuffer;
1035 : Buffer buffers[64];
1036 :
1037 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
1038 : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
1039 : Assert(extend_to != InvalidBlockNumber && extend_to > 0);
1040 :
1041 104100 : if (bmr.relpersistence == '\0')
1042 14328 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
1043 :
1044 : /*
1045 : * If desired, create the file if it doesn't exist. If
1046 : * smgr_cached_nblocks[fork] is positive then it must exist, no need for
1047 : * an smgrexists call.
1048 : */
1049 104100 : if ((flags & EB_CREATE_FORK_IF_NEEDED) &&
1050 14328 : (BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == 0 ||
1051 40 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == InvalidBlockNumber) &&
1052 14288 : !smgrexists(BMR_GET_SMGR(bmr), fork))
1053 : {
1054 14258 : LockRelationForExtension(bmr.rel, ExclusiveLock);
1055 :
1056 : /* recheck, fork might have been created concurrently */
1057 14258 : if (!smgrexists(BMR_GET_SMGR(bmr), fork))
1058 14252 : smgrcreate(BMR_GET_SMGR(bmr), fork, flags & EB_PERFORMING_RECOVERY);
1059 :
1060 14258 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
1061 : }
1062 :
1063 : /*
1064 : * If requested, invalidate size cache, so that smgrnblocks asks the
1065 : * kernel.
1066 : */
1067 104100 : if (flags & EB_CLEAR_SIZE_CACHE)
1068 14328 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
1069 :
1070 : /*
1071 : * Estimate how many pages we'll need to extend by. This avoids acquiring
1072 : * unnecessarily many victim buffers.
1073 : */
1074 104100 : current_size = smgrnblocks(BMR_GET_SMGR(bmr), fork);
1075 :
1076 : /*
1077 : * Since no-one else can be looking at the page contents yet, there is no
1078 : * difference between an exclusive lock and a cleanup-strength lock. Note
1079 : * that we pass the original mode to ReadBuffer_common() below, when
1080 : * falling back to reading the buffer to a concurrent relation extension.
1081 : */
1082 104100 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1083 89018 : flags |= EB_LOCK_TARGET;
1084 :
1085 212258 : while (current_size < extend_to)
1086 : {
1087 108158 : uint32 num_pages = lengthof(buffers);
1088 : BlockNumber first_block;
1089 :
1090 108158 : if ((uint64) current_size + num_pages > extend_to)
1091 108156 : num_pages = extend_to - current_size;
1092 :
1093 108158 : first_block = ExtendBufferedRelCommon(bmr, fork, strategy, flags,
1094 : num_pages, extend_to,
1095 : buffers, &extended_by);
1096 :
1097 108158 : current_size = first_block + extended_by;
1098 : Assert(num_pages != 0 || current_size >= extend_to);
1099 :
1100 230956 : for (uint32 i = 0; i < extended_by; i++)
1101 : {
1102 122798 : if (first_block + i != extend_to - 1)
1103 18714 : ReleaseBuffer(buffers[i]);
1104 : else
1105 104084 : buffer = buffers[i];
1106 : }
1107 : }
1108 :
1109 : /*
1110 : * It's possible that another backend concurrently extended the relation.
1111 : * In that case read the buffer.
1112 : *
1113 : * XXX: Should we control this via a flag?
1114 : */
1115 104100 : if (buffer == InvalidBuffer)
1116 : {
1117 : Assert(extended_by == 0);
1118 16 : buffer = ReadBuffer_common(bmr.rel, BMR_GET_SMGR(bmr), bmr.relpersistence,
1119 : fork, extend_to - 1, mode, strategy);
1120 : }
1121 :
1122 104100 : return buffer;
1123 : }
1124 :
1125 : /*
1126 : * Lock and optionally zero a buffer, as part of the implementation of
1127 : * RBM_ZERO_AND_LOCK or RBM_ZERO_AND_CLEANUP_LOCK. The buffer must be already
1128 : * pinned. If the buffer is not already valid, it is zeroed and made valid.
1129 : */
1130 : static void
1131 645678 : ZeroAndLockBuffer(Buffer buffer, ReadBufferMode mode, bool already_valid)
1132 : {
1133 : BufferDesc *bufHdr;
1134 : bool need_to_zero;
1135 645678 : bool isLocalBuf = BufferIsLocal(buffer);
1136 :
1137 : Assert(mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK);
1138 :
1139 645678 : if (already_valid)
1140 : {
1141 : /*
1142 : * If the caller already knew the buffer was valid, we can skip some
1143 : * header interaction. The caller just wants to lock the buffer.
1144 : */
1145 75314 : need_to_zero = false;
1146 : }
1147 570364 : else if (isLocalBuf)
1148 : {
1149 : /* Simple case for non-shared buffers. */
1150 48 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
1151 48 : need_to_zero = StartLocalBufferIO(bufHdr, true, false);
1152 : }
1153 : else
1154 : {
1155 : /*
1156 : * Take BM_IO_IN_PROGRESS, or discover that BM_VALID has been set
1157 : * concurrently. Even though we aren't doing I/O, that ensures that
1158 : * we don't zero a page that someone else has pinned. An exclusive
1159 : * content lock wouldn't be enough, because readers are allowed to
1160 : * drop the content lock after determining that a tuple is visible
1161 : * (see buffer access rules in README).
1162 : */
1163 570316 : bufHdr = GetBufferDescriptor(buffer - 1);
1164 570316 : need_to_zero = StartBufferIO(bufHdr, true, false);
1165 : }
1166 :
1167 645678 : if (need_to_zero)
1168 : {
1169 570364 : memset(BufferGetPage(buffer), 0, BLCKSZ);
1170 :
1171 : /*
1172 : * Grab the buffer content lock before marking the page as valid, to
1173 : * make sure that no other backend sees the zeroed page before the
1174 : * caller has had a chance to initialize it.
1175 : *
1176 : * Since no-one else can be looking at the page contents yet, there is
1177 : * no difference between an exclusive lock and a cleanup-strength
1178 : * lock. (Note that we cannot use LockBuffer() or
1179 : * LockBufferForCleanup() here, because they assert that the buffer is
1180 : * already valid.)
1181 : */
1182 570364 : if (!isLocalBuf)
1183 570316 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1184 :
1185 : /* Set BM_VALID, terminate IO, and wake up any waiters */
1186 570364 : if (isLocalBuf)
1187 48 : TerminateLocalBufferIO(bufHdr, false, BM_VALID, false);
1188 : else
1189 570316 : TerminateBufferIO(bufHdr, false, BM_VALID, true, false);
1190 : }
1191 75314 : else if (!isLocalBuf)
1192 : {
1193 : /*
1194 : * The buffer is valid, so we can't zero it. The caller still expects
1195 : * the page to be locked on return.
1196 : */
1197 75274 : if (mode == RBM_ZERO_AND_LOCK)
1198 75162 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1199 : else
1200 112 : LockBufferForCleanup(buffer);
1201 : }
1202 645678 : }
1203 :
1204 : /*
1205 : * Pin a buffer for a given block. *foundPtr is set to true if the block was
1206 : * already present, or false if more work is required to either read it in or
1207 : * zero it.
1208 : */
1209 : static pg_attribute_always_inline Buffer
1210 124571896 : PinBufferForBlock(Relation rel,
1211 : SMgrRelation smgr,
1212 : char persistence,
1213 : ForkNumber forkNum,
1214 : BlockNumber blockNum,
1215 : BufferAccessStrategy strategy,
1216 : bool *foundPtr)
1217 : {
1218 : BufferDesc *bufHdr;
1219 : IOContext io_context;
1220 : IOObject io_object;
1221 :
1222 : Assert(blockNum != P_NEW);
1223 :
1224 : /* Persistence should be set before */
1225 : Assert((persistence == RELPERSISTENCE_TEMP ||
1226 : persistence == RELPERSISTENCE_PERMANENT ||
1227 : persistence == RELPERSISTENCE_UNLOGGED));
1228 :
1229 124571896 : if (persistence == RELPERSISTENCE_TEMP)
1230 : {
1231 2554536 : io_context = IOCONTEXT_NORMAL;
1232 2554536 : io_object = IOOBJECT_TEMP_RELATION;
1233 : }
1234 : else
1235 : {
1236 122017360 : io_context = IOContextForStrategy(strategy);
1237 122017360 : io_object = IOOBJECT_RELATION;
1238 : }
1239 :
1240 : TRACE_POSTGRESQL_BUFFER_READ_START(forkNum, blockNum,
1241 : smgr->smgr_rlocator.locator.spcOid,
1242 : smgr->smgr_rlocator.locator.dbOid,
1243 : smgr->smgr_rlocator.locator.relNumber,
1244 : smgr->smgr_rlocator.backend);
1245 :
1246 124571896 : if (persistence == RELPERSISTENCE_TEMP)
1247 : {
1248 2554536 : bufHdr = LocalBufferAlloc(smgr, forkNum, blockNum, foundPtr);
1249 2554524 : if (*foundPtr)
1250 2537742 : pgBufferUsage.local_blks_hit++;
1251 : }
1252 : else
1253 : {
1254 122017360 : bufHdr = BufferAlloc(smgr, persistence, forkNum, blockNum,
1255 : strategy, foundPtr, io_context);
1256 122017360 : if (*foundPtr)
1257 118550284 : pgBufferUsage.shared_blks_hit++;
1258 : }
1259 124571884 : if (rel)
1260 : {
1261 : /*
1262 : * While pgBufferUsage's "read" counter isn't bumped unless we reach
1263 : * WaitReadBuffers() (so, not for hits, and not for buffers that are
1264 : * zeroed instead), the per-relation stats always count them.
1265 : */
1266 112575212 : pgstat_count_buffer_read(rel);
1267 112575212 : if (*foundPtr)
1268 110034178 : pgstat_count_buffer_hit(rel);
1269 : }
1270 124571884 : if (*foundPtr)
1271 : {
1272 121088026 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1273 121088026 : if (VacuumCostActive)
1274 4277330 : VacuumCostBalance += VacuumCostPageHit;
1275 :
1276 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forkNum, blockNum,
1277 : smgr->smgr_rlocator.locator.spcOid,
1278 : smgr->smgr_rlocator.locator.dbOid,
1279 : smgr->smgr_rlocator.locator.relNumber,
1280 : smgr->smgr_rlocator.backend,
1281 : true);
1282 : }
1283 :
1284 124571884 : return BufferDescriptorGetBuffer(bufHdr);
1285 : }
1286 :
1287 : /*
1288 : * ReadBuffer_common -- common logic for all ReadBuffer variants
1289 : *
1290 : * smgr is required, rel is optional unless using P_NEW.
1291 : */
1292 : static pg_attribute_always_inline Buffer
1293 116474204 : ReadBuffer_common(Relation rel, SMgrRelation smgr, char smgr_persistence,
1294 : ForkNumber forkNum,
1295 : BlockNumber blockNum, ReadBufferMode mode,
1296 : BufferAccessStrategy strategy)
1297 : {
1298 : ReadBuffersOperation operation;
1299 : Buffer buffer;
1300 : int flags;
1301 : char persistence;
1302 :
1303 : /*
1304 : * Backward compatibility path, most code should use ExtendBufferedRel()
1305 : * instead, as acquiring the extension lock inside ExtendBufferedRel()
1306 : * scales a lot better.
1307 : */
1308 116474204 : if (unlikely(blockNum == P_NEW))
1309 : {
1310 522 : uint32 flags = EB_SKIP_EXTENSION_LOCK;
1311 :
1312 : /*
1313 : * Since no-one else can be looking at the page contents yet, there is
1314 : * no difference between an exclusive lock and a cleanup-strength
1315 : * lock.
1316 : */
1317 522 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1318 0 : flags |= EB_LOCK_FIRST;
1319 :
1320 522 : return ExtendBufferedRel(BMR_REL(rel), forkNum, strategy, flags);
1321 : }
1322 :
1323 116473682 : if (rel)
1324 104953906 : persistence = rel->rd_rel->relpersistence;
1325 : else
1326 11519776 : persistence = smgr_persistence;
1327 :
1328 116473682 : if (unlikely(mode == RBM_ZERO_AND_CLEANUP_LOCK ||
1329 : mode == RBM_ZERO_AND_LOCK))
1330 : {
1331 : bool found;
1332 :
1333 645678 : buffer = PinBufferForBlock(rel, smgr, persistence,
1334 : forkNum, blockNum, strategy, &found);
1335 645678 : ZeroAndLockBuffer(buffer, mode, found);
1336 645678 : return buffer;
1337 : }
1338 :
1339 : /*
1340 : * Signal that we are going to immediately wait. If we're immediately
1341 : * waiting, there is no benefit in actually executing the IO
1342 : * asynchronously, it would just add dispatch overhead.
1343 : */
1344 115828004 : flags = READ_BUFFERS_SYNCHRONOUSLY;
1345 115828004 : if (mode == RBM_ZERO_ON_ERROR)
1346 2540130 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1347 115828004 : operation.smgr = smgr;
1348 115828004 : operation.rel = rel;
1349 115828004 : operation.persistence = persistence;
1350 115828004 : operation.forknum = forkNum;
1351 115828004 : operation.strategy = strategy;
1352 115828004 : if (StartReadBuffer(&operation,
1353 : &buffer,
1354 : blockNum,
1355 : flags))
1356 1466180 : WaitReadBuffers(&operation);
1357 :
1358 115827952 : return buffer;
1359 : }
1360 :
1361 : static pg_attribute_always_inline bool
1362 123581076 : StartReadBuffersImpl(ReadBuffersOperation *operation,
1363 : Buffer *buffers,
1364 : BlockNumber blockNum,
1365 : int *nblocks,
1366 : int flags,
1367 : bool allow_forwarding)
1368 : {
1369 123581076 : int actual_nblocks = *nblocks;
1370 123581076 : int maxcombine = 0;
1371 : bool did_start_io;
1372 :
1373 : Assert(*nblocks == 1 || allow_forwarding);
1374 : Assert(*nblocks > 0);
1375 : Assert(*nblocks <= MAX_IO_COMBINE_LIMIT);
1376 :
1377 126494570 : for (int i = 0; i < actual_nblocks; ++i)
1378 : {
1379 : bool found;
1380 :
1381 123929448 : if (allow_forwarding && buffers[i] != InvalidBuffer)
1382 3224 : {
1383 : BufferDesc *bufHdr;
1384 :
1385 : /*
1386 : * This is a buffer that was pinned by an earlier call to
1387 : * StartReadBuffers(), but couldn't be handled in one operation at
1388 : * that time. The operation was split, and the caller has passed
1389 : * an already pinned buffer back to us to handle the rest of the
1390 : * operation. It must continue at the expected block number.
1391 : */
1392 : Assert(BufferGetBlockNumber(buffers[i]) == blockNum + i);
1393 :
1394 : /*
1395 : * It might be an already valid buffer (a hit) that followed the
1396 : * final contiguous block of an earlier I/O (a miss) marking the
1397 : * end of it, or a buffer that some other backend has since made
1398 : * valid by performing the I/O for us, in which case we can handle
1399 : * it as a hit now. It is safe to check for a BM_VALID flag with
1400 : * a relaxed load, because we got a fresh view of it while pinning
1401 : * it in the previous call.
1402 : *
1403 : * On the other hand if we don't see BM_VALID yet, it must be an
1404 : * I/O that was split by the previous call and we need to try to
1405 : * start a new I/O from this block. We're also racing against any
1406 : * other backend that might start the I/O or even manage to mark
1407 : * it BM_VALID after this check, but StartBufferIO() will handle
1408 : * those cases.
1409 : */
1410 3224 : if (BufferIsLocal(buffers[i]))
1411 4 : bufHdr = GetLocalBufferDescriptor(-buffers[i] - 1);
1412 : else
1413 3220 : bufHdr = GetBufferDescriptor(buffers[i] - 1);
1414 : Assert(pg_atomic_read_u64(&bufHdr->state) & BM_TAG_VALID);
1415 3224 : found = pg_atomic_read_u64(&bufHdr->state) & BM_VALID;
1416 : }
1417 : else
1418 : {
1419 123926212 : buffers[i] = PinBufferForBlock(operation->rel,
1420 : operation->smgr,
1421 123926224 : operation->persistence,
1422 : operation->forknum,
1423 : blockNum + i,
1424 : operation->strategy,
1425 : &found);
1426 : }
1427 :
1428 123929436 : if (found)
1429 : {
1430 : /*
1431 : * We have a hit. If it's the first block in the requested range,
1432 : * we can return it immediately and report that WaitReadBuffers()
1433 : * does not need to be called. If the initial value of *nblocks
1434 : * was larger, the caller will have to call again for the rest.
1435 : */
1436 121015942 : if (i == 0)
1437 : {
1438 121012714 : *nblocks = 1;
1439 :
1440 : #ifdef USE_ASSERT_CHECKING
1441 :
1442 : /*
1443 : * Initialize enough of ReadBuffersOperation to make
1444 : * CheckReadBuffersOperation() work. Outside of assertions
1445 : * that's not necessary when no IO is issued.
1446 : */
1447 : operation->buffers = buffers;
1448 : operation->blocknum = blockNum;
1449 : operation->nblocks = 1;
1450 : operation->nblocks_done = 1;
1451 : CheckReadBuffersOperation(operation, true);
1452 : #endif
1453 121012714 : return false;
1454 : }
1455 :
1456 : /*
1457 : * Otherwise we already have an I/O to perform, but this block
1458 : * can't be included as it is already valid. Split the I/O here.
1459 : * There may or may not be more blocks requiring I/O after this
1460 : * one, we haven't checked, but they can't be contiguous with this
1461 : * one in the way. We'll leave this buffer pinned, forwarding it
1462 : * to the next call, avoiding the need to unpin it here and re-pin
1463 : * it in the next call.
1464 : */
1465 3228 : actual_nblocks = i;
1466 3228 : break;
1467 : }
1468 : else
1469 : {
1470 : /*
1471 : * Check how many blocks we can cover with the same IO. The smgr
1472 : * implementation might e.g. be limited due to a segment boundary.
1473 : */
1474 2913494 : if (i == 0 && actual_nblocks > 1)
1475 : {
1476 70026 : maxcombine = smgrmaxcombine(operation->smgr,
1477 : operation->forknum,
1478 : blockNum);
1479 70026 : if (unlikely(maxcombine < actual_nblocks))
1480 : {
1481 0 : elog(DEBUG2, "limiting nblocks at %u from %u to %u",
1482 : blockNum, actual_nblocks, maxcombine);
1483 0 : actual_nblocks = maxcombine;
1484 : }
1485 : }
1486 : }
1487 : }
1488 2568350 : *nblocks = actual_nblocks;
1489 :
1490 : /* Populate information needed for I/O. */
1491 2568350 : operation->buffers = buffers;
1492 2568350 : operation->blocknum = blockNum;
1493 2568350 : operation->flags = flags;
1494 2568350 : operation->nblocks = actual_nblocks;
1495 2568350 : operation->nblocks_done = 0;
1496 2568350 : pgaio_wref_clear(&operation->io_wref);
1497 :
1498 : /*
1499 : * When using AIO, start the IO in the background. If not, issue prefetch
1500 : * requests if desired by the caller.
1501 : *
1502 : * The reason we have a dedicated path for IOMETHOD_SYNC here is to
1503 : * de-risk the introduction of AIO somewhat. It's a large architectural
1504 : * change, with lots of chances for unanticipated performance effects.
1505 : *
1506 : * Use of IOMETHOD_SYNC already leads to not actually performing IO
1507 : * asynchronously, but without the check here we'd execute IO earlier than
1508 : * we used to. Eventually this IOMETHOD_SYNC specific path should go away.
1509 : */
1510 2568350 : if (io_method != IOMETHOD_SYNC)
1511 : {
1512 : /*
1513 : * Try to start IO asynchronously. It's possible that no IO needs to
1514 : * be started, if another backend already performed the IO.
1515 : *
1516 : * Note that if an IO is started, it might not cover the entire
1517 : * requested range, e.g. because an intermediary block has been read
1518 : * in by another backend. In that case any "trailing" buffers we
1519 : * already pinned above will be "forwarded" by read_stream.c to the
1520 : * next call to StartReadBuffers().
1521 : *
1522 : * This is signalled to the caller by decrementing *nblocks *and*
1523 : * reducing operation->nblocks. The latter is done here, but not below
1524 : * WaitReadBuffers(), as in WaitReadBuffers() we can't "shorten" the
1525 : * overall read size anymore, we need to retry until done in its
1526 : * entirety or until failed.
1527 : */
1528 2566194 : did_start_io = AsyncReadBuffers(operation, nblocks);
1529 :
1530 2566164 : operation->nblocks = *nblocks;
1531 : }
1532 : else
1533 : {
1534 2156 : operation->flags |= READ_BUFFERS_SYNCHRONOUSLY;
1535 :
1536 2156 : if (flags & READ_BUFFERS_ISSUE_ADVICE)
1537 : {
1538 : /*
1539 : * In theory we should only do this if PinBufferForBlock() had to
1540 : * allocate new buffers above. That way, if two calls to
1541 : * StartReadBuffers() were made for the same blocks before
1542 : * WaitReadBuffers(), only the first would issue the advice.
1543 : * That'd be a better simulation of true asynchronous I/O, which
1544 : * would only start the I/O once, but isn't done here for
1545 : * simplicity.
1546 : */
1547 4 : smgrprefetch(operation->smgr,
1548 : operation->forknum,
1549 : blockNum,
1550 : actual_nblocks);
1551 : }
1552 :
1553 : /*
1554 : * Indicate that WaitReadBuffers() should be called. WaitReadBuffers()
1555 : * will initiate the necessary IO.
1556 : */
1557 2156 : did_start_io = true;
1558 : }
1559 :
1560 2568320 : CheckReadBuffersOperation(operation, !did_start_io);
1561 :
1562 2568320 : return did_start_io;
1563 : }
1564 :
1565 : /*
1566 : * Begin reading a range of blocks beginning at blockNum and extending for
1567 : * *nblocks. *nblocks and the buffers array are in/out parameters. On entry,
1568 : * the buffers elements covered by *nblocks must hold either InvalidBuffer or
1569 : * buffers forwarded by an earlier call to StartReadBuffers() that was split
1570 : * and is now being continued. On return, *nblocks holds the number of blocks
1571 : * accepted by this operation. If it is less than the original number then
1572 : * this operation has been split, but buffer elements up to the original
1573 : * requested size may hold forwarded buffers to be used for a continuing
1574 : * operation. The caller must either start a new I/O beginning at the block
1575 : * immediately following the blocks accepted by this call and pass those
1576 : * buffers back in, or release them if it chooses not to. It shouldn't make
1577 : * any other use of or assumptions about forwarded buffers.
1578 : *
1579 : * If false is returned, no I/O is necessary and the buffers covered by
1580 : * *nblocks on exit are valid and ready to be accessed. If true is returned,
1581 : * an I/O has been started, and WaitReadBuffers() must be called with the same
1582 : * operation object before the buffers covered by *nblocks on exit can be
1583 : * accessed. Along with the operation object, the caller-supplied array of
1584 : * buffers must remain valid until WaitReadBuffers() is called, and any
1585 : * forwarded buffers must also be preserved for a continuing call unless
1586 : * they are explicitly released.
1587 : */
1588 : bool
1589 3524704 : StartReadBuffers(ReadBuffersOperation *operation,
1590 : Buffer *buffers,
1591 : BlockNumber blockNum,
1592 : int *nblocks,
1593 : int flags)
1594 : {
1595 3524704 : return StartReadBuffersImpl(operation, buffers, blockNum, nblocks, flags,
1596 : true /* expect forwarded buffers */ );
1597 : }
1598 :
1599 : /*
1600 : * Single block version of the StartReadBuffers(). This might save a few
1601 : * instructions when called from another translation unit, because it is
1602 : * specialized for nblocks == 1.
1603 : *
1604 : * This version does not support "forwarded" buffers: they cannot be created
1605 : * by reading only one block and *buffer is ignored on entry.
1606 : */
1607 : bool
1608 120056372 : StartReadBuffer(ReadBuffersOperation *operation,
1609 : Buffer *buffer,
1610 : BlockNumber blocknum,
1611 : int flags)
1612 : {
1613 120056372 : int nblocks = 1;
1614 : bool result;
1615 :
1616 120056372 : result = StartReadBuffersImpl(operation, buffer, blocknum, &nblocks, flags,
1617 : false /* single block, no forwarding */ );
1618 : Assert(nblocks == 1); /* single block can't be short */
1619 :
1620 120056342 : return result;
1621 : }
1622 :
1623 : /*
1624 : * Perform sanity checks on the ReadBuffersOperation.
1625 : */
1626 : static void
1627 7684602 : CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete)
1628 : {
1629 : #ifdef USE_ASSERT_CHECKING
1630 : Assert(operation->nblocks_done <= operation->nblocks);
1631 : Assert(!is_complete || operation->nblocks == operation->nblocks_done);
1632 :
1633 : for (int i = 0; i < operation->nblocks; i++)
1634 : {
1635 : Buffer buffer = operation->buffers[i];
1636 : BufferDesc *buf_hdr = BufferIsLocal(buffer) ?
1637 : GetLocalBufferDescriptor(-buffer - 1) :
1638 : GetBufferDescriptor(buffer - 1);
1639 :
1640 : Assert(BufferGetBlockNumber(buffer) == operation->blocknum + i);
1641 : Assert(pg_atomic_read_u64(&buf_hdr->state) & BM_TAG_VALID);
1642 :
1643 : if (i < operation->nblocks_done)
1644 : Assert(pg_atomic_read_u64(&buf_hdr->state) & BM_VALID);
1645 : }
1646 : #endif
1647 7684602 : }
1648 :
1649 : /* helper for ReadBuffersCanStartIO(), to avoid repetition */
1650 : static inline bool
1651 2913530 : ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
1652 : {
1653 2913530 : if (BufferIsLocal(buffer))
1654 16734 : return StartLocalBufferIO(GetLocalBufferDescriptor(-buffer - 1),
1655 : true, nowait);
1656 : else
1657 2896796 : return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
1658 : }
1659 :
1660 : /*
1661 : * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
1662 : */
1663 : static inline bool
1664 2913530 : ReadBuffersCanStartIO(Buffer buffer, bool nowait)
1665 : {
1666 : /*
1667 : * If this backend currently has staged IO, we need to submit the pending
1668 : * IO before waiting for the right to issue IO, to avoid the potential for
1669 : * deadlocks (and, more commonly, unnecessary delays for other backends).
1670 : */
1671 2913530 : if (!nowait && pgaio_have_staged())
1672 : {
1673 1230 : if (ReadBuffersCanStartIOOnce(buffer, true))
1674 1230 : return true;
1675 :
1676 : /*
1677 : * Unfortunately StartBufferIO() returning false doesn't allow to
1678 : * distinguish between the buffer already being valid and IO already
1679 : * being in progress. Since IO already being in progress is quite
1680 : * rare, this approach seems fine.
1681 : */
1682 0 : pgaio_submit_staged();
1683 : }
1684 :
1685 2912300 : return ReadBuffersCanStartIOOnce(buffer, nowait);
1686 : }
1687 :
1688 : /*
1689 : * Helper for WaitReadBuffers() that processes the results of a readv
1690 : * operation, raising an error if necessary.
1691 : */
1692 : static void
1693 2557102 : ProcessReadBuffersResult(ReadBuffersOperation *operation)
1694 : {
1695 2557102 : PgAioReturn *aio_ret = &operation->io_return;
1696 2557102 : PgAioResultStatus rs = aio_ret->result.status;
1697 2557102 : int newly_read_blocks = 0;
1698 :
1699 : Assert(pgaio_wref_valid(&operation->io_wref));
1700 : Assert(aio_ret->result.status != PGAIO_RS_UNKNOWN);
1701 :
1702 : /*
1703 : * SMGR reports the number of blocks successfully read as the result of
1704 : * the IO operation. Thus we can simply add that to ->nblocks_done.
1705 : */
1706 :
1707 2557102 : if (likely(rs != PGAIO_RS_ERROR))
1708 2557044 : newly_read_blocks = aio_ret->result.result;
1709 :
1710 2557102 : if (rs == PGAIO_RS_ERROR || rs == PGAIO_RS_WARNING)
1711 90 : pgaio_result_report(aio_ret->result, &aio_ret->target_data,
1712 : rs == PGAIO_RS_ERROR ? ERROR : WARNING);
1713 2557012 : else if (aio_ret->result.status == PGAIO_RS_PARTIAL)
1714 : {
1715 : /*
1716 : * We'll retry, so we just emit a debug message to the server log (or
1717 : * not even that in prod scenarios).
1718 : */
1719 20 : pgaio_result_report(aio_ret->result, &aio_ret->target_data, DEBUG1);
1720 20 : elog(DEBUG3, "partial read, will retry");
1721 : }
1722 :
1723 : Assert(newly_read_blocks > 0);
1724 : Assert(newly_read_blocks <= MAX_IO_COMBINE_LIMIT);
1725 :
1726 2557044 : operation->nblocks_done += newly_read_blocks;
1727 :
1728 : Assert(operation->nblocks_done <= operation->nblocks);
1729 2557044 : }
1730 :
1731 : void
1732 2557082 : WaitReadBuffers(ReadBuffersOperation *operation)
1733 : {
1734 2557082 : PgAioReturn *aio_ret = &operation->io_return;
1735 : IOContext io_context;
1736 : IOObject io_object;
1737 :
1738 2557082 : if (operation->persistence == RELPERSISTENCE_TEMP)
1739 : {
1740 2980 : io_context = IOCONTEXT_NORMAL;
1741 2980 : io_object = IOOBJECT_TEMP_RELATION;
1742 : }
1743 : else
1744 : {
1745 2554102 : io_context = IOContextForStrategy(operation->strategy);
1746 2554102 : io_object = IOOBJECT_RELATION;
1747 : }
1748 :
1749 : /*
1750 : * If we get here without an IO operation having been issued, the
1751 : * io_method == IOMETHOD_SYNC path must have been used. Otherwise the
1752 : * caller should not have called WaitReadBuffers().
1753 : *
1754 : * In the case of IOMETHOD_SYNC, we start - as we used to before the
1755 : * introducing of AIO - the IO in WaitReadBuffers(). This is done as part
1756 : * of the retry logic below, no extra code is required.
1757 : *
1758 : * This path is expected to eventually go away.
1759 : */
1760 2557082 : if (!pgaio_wref_valid(&operation->io_wref) && io_method != IOMETHOD_SYNC)
1761 0 : elog(ERROR, "waiting for read operation that didn't read");
1762 :
1763 : /*
1764 : * To handle partial reads, and IOMETHOD_SYNC, we re-issue IO until we're
1765 : * done. We may need multiple retries, not just because we could get
1766 : * multiple partial reads, but also because some of the remaining
1767 : * to-be-read buffers may have been read in by other backends, limiting
1768 : * the IO size.
1769 : */
1770 : while (true)
1771 2176 : {
1772 : int ignored_nblocks_progress;
1773 :
1774 2559258 : CheckReadBuffersOperation(operation, false);
1775 :
1776 : /*
1777 : * If there is an IO associated with the operation, we may need to
1778 : * wait for it.
1779 : */
1780 2559258 : if (pgaio_wref_valid(&operation->io_wref))
1781 : {
1782 : /*
1783 : * Track the time spent waiting for the IO to complete. As
1784 : * tracking a wait even if we don't actually need to wait
1785 : *
1786 : * a) is not cheap, due to the timestamping overhead
1787 : *
1788 : * b) reports some time as waiting, even if we never waited
1789 : *
1790 : * we first check if we already know the IO is complete.
1791 : */
1792 2557102 : if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
1793 1074084 : !pgaio_wref_check_done(&operation->io_wref))
1794 : {
1795 322160 : instr_time io_start = pgstat_prepare_io_time(track_io_timing);
1796 :
1797 322160 : pgaio_wref_wait(&operation->io_wref);
1798 :
1799 : /*
1800 : * The IO operation itself was already counted earlier, in
1801 : * AsyncReadBuffers(), this just accounts for the wait time.
1802 : */
1803 322160 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1804 : io_start, 0, 0);
1805 : }
1806 : else
1807 : {
1808 : Assert(pgaio_wref_check_done(&operation->io_wref));
1809 : }
1810 :
1811 : /*
1812 : * We now are sure the IO completed. Check the results. This
1813 : * includes reporting on errors if there were any.
1814 : */
1815 2557102 : ProcessReadBuffersResult(operation);
1816 : }
1817 :
1818 : /*
1819 : * Most of the time, the one IO we already started, will read in
1820 : * everything. But we need to deal with partial reads and buffers not
1821 : * needing IO anymore.
1822 : */
1823 2559200 : if (operation->nblocks_done == operation->nblocks)
1824 2557024 : break;
1825 :
1826 2176 : CHECK_FOR_INTERRUPTS();
1827 :
1828 : /*
1829 : * This may only complete the IO partially, either because some
1830 : * buffers were already valid, or because of a partial read.
1831 : *
1832 : * NB: In contrast to after the AsyncReadBuffers() call in
1833 : * StartReadBuffers(), we do *not* reduce
1834 : * ReadBuffersOperation->nblocks here, callers expect the full
1835 : * operation to be completed at this point (as more operations may
1836 : * have been queued).
1837 : */
1838 2176 : AsyncReadBuffers(operation, &ignored_nblocks_progress);
1839 : }
1840 :
1841 2557024 : CheckReadBuffersOperation(operation, true);
1842 :
1843 : /* NB: READ_DONE tracepoint was already executed in completion callback */
1844 2557024 : }
1845 :
1846 : /*
1847 : * Initiate IO for the ReadBuffersOperation
1848 : *
1849 : * This function only starts a single IO at a time. The size of the IO may be
1850 : * limited to below the to-be-read blocks, if one of the buffers has
1851 : * concurrently been read in. If the first to-be-read buffer is already valid,
1852 : * no IO will be issued.
1853 : *
1854 : * To support retries after partial reads, the first operation->nblocks_done
1855 : * buffers are skipped.
1856 : *
1857 : * On return *nblocks_progress is updated to reflect the number of buffers
1858 : * affected by the call. If the first buffer is valid, *nblocks_progress is
1859 : * set to 1 and operation->nblocks_done is incremented.
1860 : *
1861 : * Returns true if IO was initiated, false if no IO was necessary.
1862 : */
1863 : static bool
1864 2568370 : AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
1865 : {
1866 2568370 : Buffer *buffers = &operation->buffers[0];
1867 2568370 : int flags = operation->flags;
1868 2568370 : BlockNumber blocknum = operation->blocknum;
1869 2568370 : ForkNumber forknum = operation->forknum;
1870 2568370 : char persistence = operation->persistence;
1871 2568370 : int16 nblocks_done = operation->nblocks_done;
1872 2568370 : Buffer *io_buffers = &operation->buffers[nblocks_done];
1873 2568370 : int io_buffers_len = 0;
1874 : PgAioHandle *ioh;
1875 2568370 : uint32 ioh_flags = 0;
1876 : void *io_pages[MAX_IO_COMBINE_LIMIT];
1877 : IOContext io_context;
1878 : IOObject io_object;
1879 : bool did_start_io;
1880 :
1881 : /*
1882 : * When this IO is executed synchronously, either because the caller will
1883 : * immediately block waiting for the IO or because IOMETHOD_SYNC is used,
1884 : * the AIO subsystem needs to know.
1885 : */
1886 2568370 : if (flags & READ_BUFFERS_SYNCHRONOUSLY)
1887 1473094 : ioh_flags |= PGAIO_HF_SYNCHRONOUS;
1888 :
1889 2568370 : if (persistence == RELPERSISTENCE_TEMP)
1890 : {
1891 3568 : io_context = IOCONTEXT_NORMAL;
1892 3568 : io_object = IOOBJECT_TEMP_RELATION;
1893 3568 : ioh_flags |= PGAIO_HF_REFERENCES_LOCAL;
1894 : }
1895 : else
1896 : {
1897 2564802 : io_context = IOContextForStrategy(operation->strategy);
1898 2564802 : io_object = IOOBJECT_RELATION;
1899 : }
1900 :
1901 : /*
1902 : * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
1903 : * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
1904 : * set globally, but on a per-session basis. The completion callback,
1905 : * which may be run in other processes, e.g. in IO workers, may have a
1906 : * different value of the zero_damaged_pages GUC.
1907 : *
1908 : * XXX: We probably should eventually use a different flag for
1909 : * zero_damaged_pages, so we can report different log levels / error codes
1910 : * for zero_damaged_pages and ZERO_ON_ERROR.
1911 : */
1912 2568370 : if (zero_damaged_pages)
1913 32 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1914 :
1915 : /*
1916 : * For the same reason as with zero_damaged_pages we need to use this
1917 : * backend's ignore_checksum_failure value.
1918 : */
1919 2568370 : if (ignore_checksum_failure)
1920 16 : flags |= READ_BUFFERS_IGNORE_CHECKSUM_FAILURES;
1921 :
1922 :
1923 : /*
1924 : * To be allowed to report stats in the local completion callback we need
1925 : * to prepare to report stats now. This ensures we can safely report the
1926 : * checksum failure even in a critical section.
1927 : */
1928 2568370 : pgstat_prepare_report_checksum_failure(operation->smgr->smgr_rlocator.locator.dbOid);
1929 :
1930 : /*
1931 : * Get IO handle before ReadBuffersCanStartIO(), as pgaio_io_acquire()
1932 : * might block, which we don't want after setting IO_IN_PROGRESS.
1933 : *
1934 : * If we need to wait for IO before we can get a handle, submit
1935 : * already-staged IO first, so that other backends don't need to wait.
1936 : * There wouldn't be a deadlock risk, as pgaio_io_acquire() just needs to
1937 : * wait for already submitted IO, which doesn't require additional locks,
1938 : * but it could still cause undesirable waits.
1939 : *
1940 : * A secondary benefit is that this would allow us to measure the time in
1941 : * pgaio_io_acquire() without causing undue timer overhead in the common,
1942 : * non-blocking, case. However, currently the pgstats infrastructure
1943 : * doesn't really allow that, as it a) asserts that an operation can't
1944 : * have time without operations b) doesn't have an API to report
1945 : * "accumulated" time.
1946 : */
1947 2568370 : ioh = pgaio_io_acquire_nb(CurrentResourceOwner, &operation->io_return);
1948 2568370 : if (unlikely(!ioh))
1949 : {
1950 7188 : pgaio_submit_staged();
1951 :
1952 7188 : ioh = pgaio_io_acquire(CurrentResourceOwner, &operation->io_return);
1953 : }
1954 :
1955 : /*
1956 : * Check if we can start IO on the first to-be-read buffer.
1957 : *
1958 : * If an I/O is already in progress in another backend, we want to wait
1959 : * for the outcome: either done, or something went wrong and we will
1960 : * retry.
1961 : */
1962 2568370 : if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
1963 : {
1964 : /*
1965 : * Someone else has already completed this block, we're done.
1966 : *
1967 : * When IO is necessary, ->nblocks_done is updated in
1968 : * ProcessReadBuffersResult(), but that is not called if no IO is
1969 : * necessary. Thus update here.
1970 : */
1971 10644 : operation->nblocks_done += 1;
1972 10644 : *nblocks_progress = 1;
1973 :
1974 10644 : pgaio_io_release(ioh);
1975 10644 : pgaio_wref_clear(&operation->io_wref);
1976 10644 : did_start_io = false;
1977 :
1978 : /*
1979 : * Report and track this as a 'hit' for this backend, even though it
1980 : * must have started out as a miss in PinBufferForBlock(). The other
1981 : * backend will track this as a 'read'.
1982 : */
1983 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forknum, blocknum + operation->nblocks_done,
1984 : operation->smgr->smgr_rlocator.locator.spcOid,
1985 : operation->smgr->smgr_rlocator.locator.dbOid,
1986 : operation->smgr->smgr_rlocator.locator.relNumber,
1987 : operation->smgr->smgr_rlocator.backend,
1988 : true);
1989 :
1990 10644 : if (persistence == RELPERSISTENCE_TEMP)
1991 0 : pgBufferUsage.local_blks_hit += 1;
1992 : else
1993 10644 : pgBufferUsage.shared_blks_hit += 1;
1994 :
1995 10644 : if (operation->rel)
1996 10644 : pgstat_count_buffer_hit(operation->rel);
1997 :
1998 10644 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1999 :
2000 10644 : if (VacuumCostActive)
2001 16 : VacuumCostBalance += VacuumCostPageHit;
2002 : }
2003 : else
2004 : {
2005 : instr_time io_start;
2006 :
2007 : /* We found a buffer that we need to read in. */
2008 : Assert(io_buffers[0] == buffers[nblocks_done]);
2009 2557726 : io_pages[0] = BufferGetBlock(buffers[nblocks_done]);
2010 2557726 : io_buffers_len = 1;
2011 :
2012 : /*
2013 : * How many neighboring-on-disk blocks can we scatter-read into other
2014 : * buffers at the same time? In this case we don't wait if we see an
2015 : * I/O already in progress. We already set BM_IO_IN_PROGRESS for the
2016 : * head block, so we should get on with that I/O as soon as possible.
2017 : */
2018 2902886 : for (int i = nblocks_done + 1; i < operation->nblocks; i++)
2019 : {
2020 345160 : if (!ReadBuffersCanStartIO(buffers[i], true))
2021 0 : break;
2022 : /* Must be consecutive block numbers. */
2023 : Assert(BufferGetBlockNumber(buffers[i - 1]) ==
2024 : BufferGetBlockNumber(buffers[i]) - 1);
2025 : Assert(io_buffers[io_buffers_len] == buffers[i]);
2026 :
2027 345160 : io_pages[io_buffers_len++] = BufferGetBlock(buffers[i]);
2028 : }
2029 :
2030 : /* get a reference to wait for in WaitReadBuffers() */
2031 2557726 : pgaio_io_get_wref(ioh, &operation->io_wref);
2032 :
2033 : /* provide the list of buffers to the completion callbacks */
2034 2557726 : pgaio_io_set_handle_data_32(ioh, (uint32 *) io_buffers, io_buffers_len);
2035 :
2036 2557726 : pgaio_io_register_callbacks(ioh,
2037 : persistence == RELPERSISTENCE_TEMP ?
2038 : PGAIO_HCB_LOCAL_BUFFER_READV :
2039 : PGAIO_HCB_SHARED_BUFFER_READV,
2040 : flags);
2041 :
2042 2557726 : pgaio_io_set_flag(ioh, ioh_flags);
2043 :
2044 : /* ---
2045 : * Even though we're trying to issue IO asynchronously, track the time
2046 : * in smgrstartreadv():
2047 : * - if io_method == IOMETHOD_SYNC, we will always perform the IO
2048 : * immediately
2049 : * - the io method might not support the IO (e.g. worker IO for a temp
2050 : * table)
2051 : * ---
2052 : */
2053 2557726 : io_start = pgstat_prepare_io_time(track_io_timing);
2054 2557726 : smgrstartreadv(ioh, operation->smgr, forknum,
2055 : blocknum + nblocks_done,
2056 : io_pages, io_buffers_len);
2057 2557696 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
2058 2557696 : io_start, 1, io_buffers_len * BLCKSZ);
2059 :
2060 2557696 : if (persistence == RELPERSISTENCE_TEMP)
2061 3568 : pgBufferUsage.local_blks_read += io_buffers_len;
2062 : else
2063 2554128 : pgBufferUsage.shared_blks_read += io_buffers_len;
2064 :
2065 : /*
2066 : * Track vacuum cost when issuing IO, not after waiting for it.
2067 : * Otherwise we could end up issuing a lot of IO in a short timespan,
2068 : * despite a low cost limit.
2069 : */
2070 2557696 : if (VacuumCostActive)
2071 58294 : VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
2072 :
2073 2557696 : *nblocks_progress = io_buffers_len;
2074 2557696 : did_start_io = true;
2075 : }
2076 :
2077 2568340 : return did_start_io;
2078 : }
2079 :
2080 : /*
2081 : * BufferAlloc -- subroutine for PinBufferForBlock. Handles lookup of a shared
2082 : * buffer. If no buffer exists already, selects a replacement victim and
2083 : * evicts the old page, but does NOT read in new page.
2084 : *
2085 : * "strategy" can be a buffer replacement strategy object, or NULL for
2086 : * the default strategy. The selected buffer's usage_count is advanced when
2087 : * using the default strategy, but otherwise possibly not (see PinBuffer).
2088 : *
2089 : * The returned buffer is pinned and is already marked as holding the
2090 : * desired page. If it already did have the desired page, *foundPtr is
2091 : * set true. Otherwise, *foundPtr is set false.
2092 : *
2093 : * io_context is passed as an output parameter to avoid calling
2094 : * IOContextForStrategy() when there is a shared buffers hit and no IO
2095 : * statistics need be captured.
2096 : *
2097 : * No locks are held either at entry or exit.
2098 : */
2099 : static pg_attribute_always_inline BufferDesc *
2100 122017366 : BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
2101 : BlockNumber blockNum,
2102 : BufferAccessStrategy strategy,
2103 : bool *foundPtr, IOContext io_context)
2104 : {
2105 : BufferTag newTag; /* identity of requested block */
2106 : uint32 newHash; /* hash value for newTag */
2107 : LWLock *newPartitionLock; /* buffer partition lock for it */
2108 : int existing_buf_id;
2109 : Buffer victim_buffer;
2110 : BufferDesc *victim_buf_hdr;
2111 : uint64 victim_buf_state;
2112 122017366 : uint64 set_bits = 0;
2113 :
2114 : /* Make sure we will have room to remember the buffer pin */
2115 122017366 : ResourceOwnerEnlarge(CurrentResourceOwner);
2116 122017366 : ReservePrivateRefCountEntry();
2117 :
2118 : /* create a tag so we can lookup the buffer */
2119 122017366 : InitBufferTag(&newTag, &smgr->smgr_rlocator.locator, forkNum, blockNum);
2120 :
2121 : /* determine its hash code and partition lock ID */
2122 122017366 : newHash = BufTableHashCode(&newTag);
2123 122017366 : newPartitionLock = BufMappingPartitionLock(newHash);
2124 :
2125 : /* see if the block is in the buffer pool already */
2126 122017366 : LWLockAcquire(newPartitionLock, LW_SHARED);
2127 122017366 : existing_buf_id = BufTableLookup(&newTag, newHash);
2128 122017366 : if (existing_buf_id >= 0)
2129 : {
2130 : BufferDesc *buf;
2131 : bool valid;
2132 :
2133 : /*
2134 : * Found it. Now, pin the buffer so no one can steal it from the
2135 : * buffer pool, and check to see if the correct data has been loaded
2136 : * into the buffer.
2137 : */
2138 118558368 : buf = GetBufferDescriptor(existing_buf_id);
2139 :
2140 118558368 : valid = PinBuffer(buf, strategy, false);
2141 :
2142 : /* Can release the mapping lock as soon as we've pinned it */
2143 118558368 : LWLockRelease(newPartitionLock);
2144 :
2145 118558368 : *foundPtr = true;
2146 :
2147 118558368 : if (!valid)
2148 : {
2149 : /*
2150 : * We can only get here if (a) someone else is still reading in
2151 : * the page, (b) a previous read attempt failed, or (c) someone
2152 : * called StartReadBuffers() but not yet WaitReadBuffers().
2153 : */
2154 8866 : *foundPtr = false;
2155 : }
2156 :
2157 118558368 : return buf;
2158 : }
2159 :
2160 : /*
2161 : * Didn't find it in the buffer pool. We'll have to initialize a new
2162 : * buffer. Remember to unlock the mapping lock while doing the work.
2163 : */
2164 3458998 : LWLockRelease(newPartitionLock);
2165 :
2166 : /*
2167 : * Acquire a victim buffer. Somebody else might try to do the same, we
2168 : * don't hold any conflicting locks. If so we'll have to undo our work
2169 : * later.
2170 : */
2171 3458998 : victim_buffer = GetVictimBuffer(strategy, io_context);
2172 3458998 : victim_buf_hdr = GetBufferDescriptor(victim_buffer - 1);
2173 :
2174 : /*
2175 : * Try to make a hashtable entry for the buffer under its new tag. If
2176 : * somebody else inserted another buffer for the tag, we'll release the
2177 : * victim buffer we acquired and use the already inserted one.
2178 : */
2179 3458998 : LWLockAcquire(newPartitionLock, LW_EXCLUSIVE);
2180 3458998 : existing_buf_id = BufTableInsert(&newTag, newHash, victim_buf_hdr->buf_id);
2181 3458998 : if (existing_buf_id >= 0)
2182 : {
2183 : BufferDesc *existing_buf_hdr;
2184 : bool valid;
2185 :
2186 : /*
2187 : * Got a collision. Someone has already done what we were about to do.
2188 : * We'll just handle this as if it were found in the buffer pool in
2189 : * the first place. First, give up the buffer we were planning to
2190 : * use.
2191 : *
2192 : * We could do this after releasing the partition lock, but then we'd
2193 : * have to call ResourceOwnerEnlarge() & ReservePrivateRefCountEntry()
2194 : * before acquiring the lock, for the rare case of such a collision.
2195 : */
2196 2662 : UnpinBuffer(victim_buf_hdr);
2197 :
2198 : /* remaining code should match code at top of routine */
2199 :
2200 2662 : existing_buf_hdr = GetBufferDescriptor(existing_buf_id);
2201 :
2202 2662 : valid = PinBuffer(existing_buf_hdr, strategy, false);
2203 :
2204 : /* Can release the mapping lock as soon as we've pinned it */
2205 2662 : LWLockRelease(newPartitionLock);
2206 :
2207 2662 : *foundPtr = true;
2208 :
2209 2662 : if (!valid)
2210 : {
2211 : /*
2212 : * We can only get here if (a) someone else is still reading in
2213 : * the page, (b) a previous read attempt failed, or (c) someone
2214 : * called StartReadBuffers() but not yet WaitReadBuffers().
2215 : */
2216 1874 : *foundPtr = false;
2217 : }
2218 :
2219 2662 : return existing_buf_hdr;
2220 : }
2221 :
2222 : /*
2223 : * Need to lock the buffer header too in order to change its tag.
2224 : */
2225 3456336 : victim_buf_state = LockBufHdr(victim_buf_hdr);
2226 :
2227 : /* some sanity checks while we hold the buffer header lock */
2228 : Assert(BUF_STATE_GET_REFCOUNT(victim_buf_state) == 1);
2229 : Assert(!(victim_buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY | BM_IO_IN_PROGRESS)));
2230 :
2231 3456336 : victim_buf_hdr->tag = newTag;
2232 :
2233 : /*
2234 : * Make sure BM_PERMANENT is set for buffers that must be written at every
2235 : * checkpoint. Unlogged buffers only need to be written at shutdown
2236 : * checkpoints, except for their "init" forks, which need to be treated
2237 : * just like permanent relations.
2238 : */
2239 3456336 : set_bits |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2240 3456336 : if (relpersistence == RELPERSISTENCE_PERMANENT || forkNum == INIT_FORKNUM)
2241 3455630 : set_bits |= BM_PERMANENT;
2242 :
2243 3456336 : UnlockBufHdrExt(victim_buf_hdr, victim_buf_state,
2244 : set_bits, 0, 0);
2245 :
2246 3456336 : LWLockRelease(newPartitionLock);
2247 :
2248 : /*
2249 : * Buffer contents are currently invalid.
2250 : */
2251 3456336 : *foundPtr = false;
2252 :
2253 3456336 : return victim_buf_hdr;
2254 : }
2255 :
2256 : /*
2257 : * InvalidateBuffer -- mark a shared buffer invalid.
2258 : *
2259 : * The buffer header spinlock must be held at entry. We drop it before
2260 : * returning. (This is sane because the caller must have locked the
2261 : * buffer in order to be sure it should be dropped.)
2262 : *
2263 : * This is used only in contexts such as dropping a relation. We assume
2264 : * that no other backend could possibly be interested in using the page,
2265 : * so the only reason the buffer might be pinned is if someone else is
2266 : * trying to write it out. We have to let them finish before we can
2267 : * reclaim the buffer.
2268 : *
2269 : * The buffer could get reclaimed by someone else while we are waiting
2270 : * to acquire the necessary locks; if so, don't mess it up.
2271 : */
2272 : static void
2273 211470 : InvalidateBuffer(BufferDesc *buf)
2274 : {
2275 : BufferTag oldTag;
2276 : uint32 oldHash; /* hash value for oldTag */
2277 : LWLock *oldPartitionLock; /* buffer partition lock for it */
2278 : uint32 oldFlags;
2279 : uint64 buf_state;
2280 :
2281 : /* Save the original buffer tag before dropping the spinlock */
2282 211470 : oldTag = buf->tag;
2283 :
2284 211470 : UnlockBufHdr(buf);
2285 :
2286 : /*
2287 : * Need to compute the old tag's hashcode and partition lock ID. XXX is it
2288 : * worth storing the hashcode in BufferDesc so we need not recompute it
2289 : * here? Probably not.
2290 : */
2291 211470 : oldHash = BufTableHashCode(&oldTag);
2292 211470 : oldPartitionLock = BufMappingPartitionLock(oldHash);
2293 :
2294 211490 : retry:
2295 :
2296 : /*
2297 : * Acquire exclusive mapping lock in preparation for changing the buffer's
2298 : * association.
2299 : */
2300 211490 : LWLockAcquire(oldPartitionLock, LW_EXCLUSIVE);
2301 :
2302 : /* Re-lock the buffer header */
2303 211490 : buf_state = LockBufHdr(buf);
2304 :
2305 : /* If it's changed while we were waiting for lock, do nothing */
2306 211490 : if (!BufferTagsEqual(&buf->tag, &oldTag))
2307 : {
2308 8 : UnlockBufHdr(buf);
2309 8 : LWLockRelease(oldPartitionLock);
2310 8 : return;
2311 : }
2312 :
2313 : /*
2314 : * We assume the reason for it to be pinned is that either we were
2315 : * asynchronously reading the page in before erroring out or someone else
2316 : * is flushing the page out. Wait for the IO to finish. (This could be
2317 : * an infinite loop if the refcount is messed up... it would be nice to
2318 : * time out after awhile, but there seems no way to be sure how many loops
2319 : * may be needed. Note that if the other guy has pinned the buffer but
2320 : * not yet done StartBufferIO, WaitIO will fall through and we'll
2321 : * effectively be busy-looping here.)
2322 : */
2323 211482 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 0)
2324 : {
2325 20 : UnlockBufHdr(buf);
2326 20 : LWLockRelease(oldPartitionLock);
2327 : /* safety check: should definitely not be our *own* pin */
2328 20 : if (GetPrivateRefCount(BufferDescriptorGetBuffer(buf)) > 0)
2329 0 : elog(ERROR, "buffer is pinned in InvalidateBuffer");
2330 20 : WaitIO(buf);
2331 20 : goto retry;
2332 : }
2333 :
2334 : /*
2335 : * An invalidated buffer should not have any backends waiting to lock the
2336 : * buffer, therefore BM_LOCK_WAKE_IN_PROGRESS should not be set.
2337 : */
2338 : Assert(!(buf_state & BM_LOCK_WAKE_IN_PROGRESS));
2339 :
2340 : /*
2341 : * Clear out the buffer's tag and flags. We must do this to ensure that
2342 : * linear scans of the buffer array don't think the buffer is valid.
2343 : */
2344 211462 : oldFlags = buf_state & BUF_FLAG_MASK;
2345 211462 : ClearBufferTag(&buf->tag);
2346 :
2347 211462 : UnlockBufHdrExt(buf, buf_state,
2348 : 0,
2349 : BUF_FLAG_MASK | BUF_USAGECOUNT_MASK,
2350 : 0);
2351 :
2352 : /*
2353 : * Remove the buffer from the lookup hashtable, if it was in there.
2354 : */
2355 211462 : if (oldFlags & BM_TAG_VALID)
2356 211462 : BufTableDelete(&oldTag, oldHash);
2357 :
2358 : /*
2359 : * Done with mapping lock.
2360 : */
2361 211462 : LWLockRelease(oldPartitionLock);
2362 : }
2363 :
2364 : /*
2365 : * Helper routine for GetVictimBuffer()
2366 : *
2367 : * Needs to be called on a buffer with a valid tag, pinned, but without the
2368 : * buffer header spinlock held.
2369 : *
2370 : * Returns true if the buffer can be reused, in which case the buffer is only
2371 : * pinned by this backend and marked as invalid, false otherwise.
2372 : */
2373 : static bool
2374 2425434 : InvalidateVictimBuffer(BufferDesc *buf_hdr)
2375 : {
2376 : uint64 buf_state;
2377 : uint32 hash;
2378 : LWLock *partition_lock;
2379 : BufferTag tag;
2380 :
2381 : Assert(GetPrivateRefCount(BufferDescriptorGetBuffer(buf_hdr)) == 1);
2382 :
2383 : /* have buffer pinned, so it's safe to read tag without lock */
2384 2425434 : tag = buf_hdr->tag;
2385 :
2386 2425434 : hash = BufTableHashCode(&tag);
2387 2425434 : partition_lock = BufMappingPartitionLock(hash);
2388 :
2389 2425434 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2390 :
2391 : /* lock the buffer header */
2392 2425434 : buf_state = LockBufHdr(buf_hdr);
2393 :
2394 : /*
2395 : * We have the buffer pinned nobody else should have been able to unset
2396 : * this concurrently.
2397 : */
2398 : Assert(buf_state & BM_TAG_VALID);
2399 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2400 : Assert(BufferTagsEqual(&buf_hdr->tag, &tag));
2401 :
2402 : /*
2403 : * If somebody else pinned the buffer since, or even worse, dirtied it,
2404 : * give up on this buffer: It's clearly in use.
2405 : */
2406 2425434 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 1 || (buf_state & BM_DIRTY))
2407 : {
2408 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2409 :
2410 966 : UnlockBufHdr(buf_hdr);
2411 966 : LWLockRelease(partition_lock);
2412 :
2413 966 : return false;
2414 : }
2415 :
2416 : /*
2417 : * An invalidated buffer should not have any backends waiting to lock the
2418 : * buffer, therefore BM_LOCK_WAKE_IN_PROGRESS should not be set.
2419 : */
2420 : Assert(!(buf_state & BM_LOCK_WAKE_IN_PROGRESS));
2421 :
2422 : /*
2423 : * Clear out the buffer's tag and flags and usagecount. This is not
2424 : * strictly required, as BM_TAG_VALID/BM_VALID needs to be checked before
2425 : * doing anything with the buffer. But currently it's beneficial, as the
2426 : * cheaper pre-check for several linear scans of shared buffers use the
2427 : * tag (see e.g. FlushDatabaseBuffers()).
2428 : */
2429 2424468 : ClearBufferTag(&buf_hdr->tag);
2430 2424468 : UnlockBufHdrExt(buf_hdr, buf_state,
2431 : 0,
2432 : BUF_FLAG_MASK | BUF_USAGECOUNT_MASK,
2433 : 0);
2434 :
2435 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2436 :
2437 : /* finally delete buffer from the buffer mapping table */
2438 2424468 : BufTableDelete(&tag, hash);
2439 :
2440 2424468 : LWLockRelease(partition_lock);
2441 :
2442 2424468 : buf_state = pg_atomic_read_u64(&buf_hdr->state);
2443 : Assert(!(buf_state & (BM_DIRTY | BM_VALID | BM_TAG_VALID)));
2444 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2445 : Assert(BUF_STATE_GET_REFCOUNT(pg_atomic_read_u64(&buf_hdr->state)) > 0);
2446 :
2447 2424468 : return true;
2448 : }
2449 :
2450 : static Buffer
2451 3910912 : GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context)
2452 : {
2453 : BufferDesc *buf_hdr;
2454 : Buffer buf;
2455 : uint64 buf_state;
2456 : bool from_ring;
2457 :
2458 : /*
2459 : * Ensure, before we pin a victim buffer, that there's a free refcount
2460 : * entry and resource owner slot for the pin.
2461 : */
2462 3910912 : ReservePrivateRefCountEntry();
2463 3910912 : ResourceOwnerEnlarge(CurrentResourceOwner);
2464 :
2465 : /* we return here if a prospective victim buffer gets used concurrently */
2466 11632 : again:
2467 :
2468 : /*
2469 : * Select a victim buffer. The buffer is returned pinned and owned by
2470 : * this backend.
2471 : */
2472 3922544 : buf_hdr = StrategyGetBuffer(strategy, &buf_state, &from_ring);
2473 3922544 : buf = BufferDescriptorGetBuffer(buf_hdr);
2474 :
2475 : /*
2476 : * We shouldn't have any other pins for this buffer.
2477 : */
2478 3922544 : CheckBufferIsPinnedOnce(buf);
2479 :
2480 : /*
2481 : * If the buffer was dirty, try to write it out. There is a race
2482 : * condition here, in that someone might dirty it after we released the
2483 : * buffer header lock above, or even while we are writing it out (since
2484 : * our share-lock won't prevent hint-bit updates). We will recheck the
2485 : * dirty bit after re-locking the buffer header.
2486 : */
2487 3922544 : if (buf_state & BM_DIRTY)
2488 : {
2489 : Assert(buf_state & BM_TAG_VALID);
2490 : Assert(buf_state & BM_VALID);
2491 :
2492 : /*
2493 : * We need a share-lock on the buffer contents to write it out (else
2494 : * we might write invalid data, eg because someone else is compacting
2495 : * the page contents while we write). We must use a conditional lock
2496 : * acquisition here to avoid deadlock. Even though the buffer was not
2497 : * pinned (and therefore surely not locked) when StrategyGetBuffer
2498 : * returned it, someone else could have pinned and exclusive-locked it
2499 : * by the time we get here. If we try to get the lock unconditionally,
2500 : * we'd block waiting for them; if they later block waiting for us,
2501 : * deadlock ensues. (This has been observed to happen when two
2502 : * backends are both trying to split btree index pages, and the second
2503 : * one just happens to be trying to split the page the first one got
2504 : * from StrategyGetBuffer.)
2505 : */
2506 543818 : if (!BufferLockConditional(buf, buf_hdr, BUFFER_LOCK_SHARE))
2507 : {
2508 : /*
2509 : * Someone else has locked the buffer, so give it up and loop back
2510 : * to get another one.
2511 : */
2512 0 : UnpinBuffer(buf_hdr);
2513 0 : goto again;
2514 : }
2515 :
2516 : /*
2517 : * If using a nondefault strategy, and writing the buffer would
2518 : * require a WAL flush, let the strategy decide whether to go ahead
2519 : * and write/reuse the buffer or to choose another victim. We need a
2520 : * lock to inspect the page LSN, so this can't be done inside
2521 : * StrategyGetBuffer.
2522 : */
2523 543818 : if (strategy != NULL)
2524 : {
2525 : XLogRecPtr lsn;
2526 :
2527 : /* Read the LSN while holding buffer header lock */
2528 152742 : buf_state = LockBufHdr(buf_hdr);
2529 152742 : lsn = BufferGetLSN(buf_hdr);
2530 152742 : UnlockBufHdr(buf_hdr);
2531 :
2532 152742 : if (XLogNeedsFlush(lsn)
2533 17264 : && StrategyRejectBuffer(strategy, buf_hdr, from_ring))
2534 : {
2535 10666 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
2536 10666 : UnpinBuffer(buf_hdr);
2537 10666 : goto again;
2538 : }
2539 : }
2540 :
2541 : /* OK, do the I/O */
2542 533152 : FlushBuffer(buf_hdr, NULL, IOOBJECT_RELATION, io_context);
2543 533152 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
2544 :
2545 533152 : ScheduleBufferTagForWriteback(&BackendWritebackContext, io_context,
2546 : &buf_hdr->tag);
2547 : }
2548 :
2549 :
2550 3911878 : if (buf_state & BM_VALID)
2551 : {
2552 : /*
2553 : * When a BufferAccessStrategy is in use, blocks evicted from shared
2554 : * buffers are counted as IOOP_EVICT in the corresponding context
2555 : * (e.g. IOCONTEXT_BULKWRITE). Shared buffers are evicted by a
2556 : * strategy in two cases: 1) while initially claiming buffers for the
2557 : * strategy ring 2) to replace an existing strategy ring buffer
2558 : * because it is pinned or in use and cannot be reused.
2559 : *
2560 : * Blocks evicted from buffers already in the strategy ring are
2561 : * counted as IOOP_REUSE in the corresponding strategy context.
2562 : *
2563 : * At this point, we can accurately count evictions and reuses,
2564 : * because we have successfully claimed the valid buffer. Previously,
2565 : * we may have been forced to release the buffer due to concurrent
2566 : * pinners or erroring out.
2567 : */
2568 2421148 : pgstat_count_io_op(IOOBJECT_RELATION, io_context,
2569 2421148 : from_ring ? IOOP_REUSE : IOOP_EVICT, 1, 0);
2570 : }
2571 :
2572 : /*
2573 : * If the buffer has an entry in the buffer mapping table, delete it. This
2574 : * can fail because another backend could have pinned or dirtied the
2575 : * buffer.
2576 : */
2577 3911878 : if ((buf_state & BM_TAG_VALID) && !InvalidateVictimBuffer(buf_hdr))
2578 : {
2579 966 : UnpinBuffer(buf_hdr);
2580 966 : goto again;
2581 : }
2582 :
2583 : /* a final set of sanity checks */
2584 : #ifdef USE_ASSERT_CHECKING
2585 : buf_state = pg_atomic_read_u64(&buf_hdr->state);
2586 :
2587 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2588 : Assert(!(buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY)));
2589 :
2590 : CheckBufferIsPinnedOnce(buf);
2591 : #endif
2592 :
2593 3910912 : return buf;
2594 : }
2595 :
2596 : /*
2597 : * Return the maximum number of buffers that a backend should try to pin once,
2598 : * to avoid exceeding its fair share. This is the highest value that
2599 : * GetAdditionalPinLimit() could ever return. Note that it may be zero on a
2600 : * system with a very small buffer pool relative to max_connections.
2601 : */
2602 : uint32
2603 1265604 : GetPinLimit(void)
2604 : {
2605 1265604 : return MaxProportionalPins;
2606 : }
2607 :
2608 : /*
2609 : * Return the maximum number of additional buffers that this backend should
2610 : * pin if it wants to stay under the per-backend limit, considering the number
2611 : * of buffers it has already pinned. Unlike LimitAdditionalPins(), the limit
2612 : * return by this function can be zero.
2613 : */
2614 : uint32
2615 7027020 : GetAdditionalPinLimit(void)
2616 : {
2617 : uint32 estimated_pins_held;
2618 :
2619 : /*
2620 : * We get the number of "overflowed" pins for free, but don't know the
2621 : * number of pins in PrivateRefCountArray. The cost of calculating that
2622 : * exactly doesn't seem worth it, so just assume the max.
2623 : */
2624 7027020 : estimated_pins_held = PrivateRefCountOverflowed + REFCOUNT_ARRAY_ENTRIES;
2625 :
2626 : /* Is this backend already holding more than its fair share? */
2627 7027020 : if (estimated_pins_held > MaxProportionalPins)
2628 2378138 : return 0;
2629 :
2630 4648882 : return MaxProportionalPins - estimated_pins_held;
2631 : }
2632 :
2633 : /*
2634 : * Limit the number of pins a batch operation may additionally acquire, to
2635 : * avoid running out of pinnable buffers.
2636 : *
2637 : * One additional pin is always allowed, on the assumption that the operation
2638 : * requires at least one to make progress.
2639 : */
2640 : void
2641 409180 : LimitAdditionalPins(uint32 *additional_pins)
2642 : {
2643 : uint32 limit;
2644 :
2645 409180 : if (*additional_pins <= 1)
2646 389248 : return;
2647 :
2648 19932 : limit = GetAdditionalPinLimit();
2649 19932 : limit = Max(limit, 1);
2650 19932 : if (limit < *additional_pins)
2651 10786 : *additional_pins = limit;
2652 : }
2653 :
2654 : /*
2655 : * Logic shared between ExtendBufferedRelBy(), ExtendBufferedRelTo(). Just to
2656 : * avoid duplicating the tracing and relpersistence related logic.
2657 : */
2658 : static BlockNumber
2659 432068 : ExtendBufferedRelCommon(BufferManagerRelation bmr,
2660 : ForkNumber fork,
2661 : BufferAccessStrategy strategy,
2662 : uint32 flags,
2663 : uint32 extend_by,
2664 : BlockNumber extend_upto,
2665 : Buffer *buffers,
2666 : uint32 *extended_by)
2667 : {
2668 : BlockNumber first_block;
2669 :
2670 : TRACE_POSTGRESQL_BUFFER_EXTEND_START(fork,
2671 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2672 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2673 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2674 : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2675 : extend_by);
2676 :
2677 432068 : if (bmr.relpersistence == RELPERSISTENCE_TEMP)
2678 22888 : first_block = ExtendBufferedRelLocal(bmr, fork, flags,
2679 : extend_by, extend_upto,
2680 : buffers, &extend_by);
2681 : else
2682 409180 : first_block = ExtendBufferedRelShared(bmr, fork, strategy, flags,
2683 : extend_by, extend_upto,
2684 : buffers, &extend_by);
2685 432068 : *extended_by = extend_by;
2686 :
2687 : TRACE_POSTGRESQL_BUFFER_EXTEND_DONE(fork,
2688 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2689 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2690 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2691 : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2692 : *extended_by,
2693 : first_block);
2694 :
2695 432068 : return first_block;
2696 : }
2697 :
2698 : /*
2699 : * Implementation of ExtendBufferedRelBy() and ExtendBufferedRelTo() for
2700 : * shared buffers.
2701 : */
2702 : static BlockNumber
2703 409180 : ExtendBufferedRelShared(BufferManagerRelation bmr,
2704 : ForkNumber fork,
2705 : BufferAccessStrategy strategy,
2706 : uint32 flags,
2707 : uint32 extend_by,
2708 : BlockNumber extend_upto,
2709 : Buffer *buffers,
2710 : uint32 *extended_by)
2711 : {
2712 : BlockNumber first_block;
2713 409180 : IOContext io_context = IOContextForStrategy(strategy);
2714 : instr_time io_start;
2715 :
2716 409180 : LimitAdditionalPins(&extend_by);
2717 :
2718 : /*
2719 : * Acquire victim buffers for extension without holding extension lock.
2720 : * Writing out victim buffers is the most expensive part of extending the
2721 : * relation, particularly when doing so requires WAL flushes. Zeroing out
2722 : * the buffers is also quite expensive, so do that before holding the
2723 : * extension lock as well.
2724 : *
2725 : * These pages are pinned by us and not valid. While we hold the pin they
2726 : * can't be acquired as victim buffers by another backend.
2727 : */
2728 861094 : for (uint32 i = 0; i < extend_by; i++)
2729 : {
2730 : Block buf_block;
2731 :
2732 451914 : buffers[i] = GetVictimBuffer(strategy, io_context);
2733 451914 : buf_block = BufHdrGetBlock(GetBufferDescriptor(buffers[i] - 1));
2734 :
2735 : /* new buffers are zero-filled */
2736 451914 : MemSet(buf_block, 0, BLCKSZ);
2737 : }
2738 :
2739 : /*
2740 : * Lock relation against concurrent extensions, unless requested not to.
2741 : *
2742 : * We use the same extension lock for all forks. That's unnecessarily
2743 : * restrictive, but currently extensions for forks don't happen often
2744 : * enough to make it worth locking more granularly.
2745 : *
2746 : * Note that another backend might have extended the relation by the time
2747 : * we get the lock.
2748 : */
2749 409180 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2750 304976 : LockRelationForExtension(bmr.rel, ExclusiveLock);
2751 :
2752 : /*
2753 : * If requested, invalidate size cache, so that smgrnblocks asks the
2754 : * kernel.
2755 : */
2756 409180 : if (flags & EB_CLEAR_SIZE_CACHE)
2757 15728 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
2758 :
2759 409180 : first_block = smgrnblocks(BMR_GET_SMGR(bmr), fork);
2760 :
2761 : /*
2762 : * Now that we have the accurate relation size, check if the caller wants
2763 : * us to extend to only up to a specific size. If there were concurrent
2764 : * extensions, we might have acquired too many buffers and need to release
2765 : * them.
2766 : */
2767 409180 : if (extend_upto != InvalidBlockNumber)
2768 : {
2769 107826 : uint32 orig_extend_by = extend_by;
2770 :
2771 107826 : if (first_block > extend_upto)
2772 0 : extend_by = 0;
2773 107826 : else if ((uint64) first_block + extend_by > extend_upto)
2774 16 : extend_by = extend_upto - first_block;
2775 :
2776 107870 : for (uint32 i = extend_by; i < orig_extend_by; i++)
2777 : {
2778 44 : BufferDesc *buf_hdr = GetBufferDescriptor(buffers[i] - 1);
2779 :
2780 44 : UnpinBuffer(buf_hdr);
2781 : }
2782 :
2783 107826 : if (extend_by == 0)
2784 : {
2785 16 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2786 16 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2787 16 : *extended_by = extend_by;
2788 16 : return first_block;
2789 : }
2790 : }
2791 :
2792 : /* Fail if relation is already at maximum possible length */
2793 409164 : if ((uint64) first_block + extend_by >= MaxBlockNumber)
2794 0 : ereport(ERROR,
2795 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
2796 : errmsg("cannot extend relation %s beyond %u blocks",
2797 : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str,
2798 : MaxBlockNumber)));
2799 :
2800 : /*
2801 : * Insert buffers into buffer table, mark as IO_IN_PROGRESS.
2802 : *
2803 : * This needs to happen before we extend the relation, because as soon as
2804 : * we do, other backends can start to read in those pages.
2805 : */
2806 861034 : for (uint32 i = 0; i < extend_by; i++)
2807 : {
2808 451870 : Buffer victim_buf = buffers[i];
2809 451870 : BufferDesc *victim_buf_hdr = GetBufferDescriptor(victim_buf - 1);
2810 : BufferTag tag;
2811 : uint32 hash;
2812 : LWLock *partition_lock;
2813 : int existing_id;
2814 :
2815 : /* in case we need to pin an existing buffer below */
2816 451870 : ResourceOwnerEnlarge(CurrentResourceOwner);
2817 451870 : ReservePrivateRefCountEntry();
2818 :
2819 451870 : InitBufferTag(&tag, &BMR_GET_SMGR(bmr)->smgr_rlocator.locator, fork,
2820 : first_block + i);
2821 451870 : hash = BufTableHashCode(&tag);
2822 451870 : partition_lock = BufMappingPartitionLock(hash);
2823 :
2824 451870 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2825 :
2826 451870 : existing_id = BufTableInsert(&tag, hash, victim_buf_hdr->buf_id);
2827 :
2828 : /*
2829 : * We get here only in the corner case where we are trying to extend
2830 : * the relation but we found a pre-existing buffer. This can happen
2831 : * because a prior attempt at extending the relation failed, and
2832 : * because mdread doesn't complain about reads beyond EOF (when
2833 : * zero_damaged_pages is ON) and so a previous attempt to read a block
2834 : * beyond EOF could have left a "valid" zero-filled buffer.
2835 : *
2836 : * This has also been observed when relation was overwritten by
2837 : * external process. Since the legitimate cases should always have
2838 : * left a zero-filled buffer, complain if not PageIsNew.
2839 : */
2840 451870 : if (existing_id >= 0)
2841 : {
2842 0 : BufferDesc *existing_hdr = GetBufferDescriptor(existing_id);
2843 : Block buf_block;
2844 : bool valid;
2845 :
2846 : /*
2847 : * Pin the existing buffer before releasing the partition lock,
2848 : * preventing it from being evicted.
2849 : */
2850 0 : valid = PinBuffer(existing_hdr, strategy, false);
2851 :
2852 0 : LWLockRelease(partition_lock);
2853 0 : UnpinBuffer(victim_buf_hdr);
2854 :
2855 0 : buffers[i] = BufferDescriptorGetBuffer(existing_hdr);
2856 0 : buf_block = BufHdrGetBlock(existing_hdr);
2857 :
2858 0 : if (valid && !PageIsNew((Page) buf_block))
2859 0 : ereport(ERROR,
2860 : (errmsg("unexpected data beyond EOF in block %u of relation \"%s\"",
2861 : existing_hdr->tag.blockNum,
2862 : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str)));
2863 :
2864 : /*
2865 : * We *must* do smgr[zero]extend before succeeding, else the page
2866 : * will not be reserved by the kernel, and the next P_NEW call
2867 : * will decide to return the same page. Clear the BM_VALID bit,
2868 : * do StartBufferIO() and proceed.
2869 : *
2870 : * Loop to handle the very small possibility that someone re-sets
2871 : * BM_VALID between our clearing it and StartBufferIO inspecting
2872 : * it.
2873 : */
2874 : do
2875 : {
2876 0 : pg_atomic_fetch_and_u64(&existing_hdr->state, ~BM_VALID);
2877 0 : } while (!StartBufferIO(existing_hdr, true, false));
2878 : }
2879 : else
2880 : {
2881 : uint64 buf_state;
2882 451870 : uint64 set_bits = 0;
2883 :
2884 451870 : buf_state = LockBufHdr(victim_buf_hdr);
2885 :
2886 : /* some sanity checks while we hold the buffer header lock */
2887 : Assert(!(buf_state & (BM_VALID | BM_TAG_VALID | BM_DIRTY | BM_JUST_DIRTIED)));
2888 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2889 :
2890 451870 : victim_buf_hdr->tag = tag;
2891 :
2892 451870 : set_bits |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2893 451870 : if (bmr.relpersistence == RELPERSISTENCE_PERMANENT || fork == INIT_FORKNUM)
2894 441224 : set_bits |= BM_PERMANENT;
2895 :
2896 451870 : UnlockBufHdrExt(victim_buf_hdr, buf_state,
2897 : set_bits, 0,
2898 : 0);
2899 :
2900 451870 : LWLockRelease(partition_lock);
2901 :
2902 : /* XXX: could combine the locked operations in it with the above */
2903 451870 : StartBufferIO(victim_buf_hdr, true, false);
2904 : }
2905 : }
2906 :
2907 409164 : io_start = pgstat_prepare_io_time(track_io_timing);
2908 :
2909 : /*
2910 : * Note: if smgrzeroextend fails, we will end up with buffers that are
2911 : * allocated but not marked BM_VALID. The next relation extension will
2912 : * still select the same block number (because the relation didn't get any
2913 : * longer on disk) and so future attempts to extend the relation will find
2914 : * the same buffers (if they have not been recycled) but come right back
2915 : * here to try smgrzeroextend again.
2916 : *
2917 : * We don't need to set checksum for all-zero pages.
2918 : */
2919 409164 : smgrzeroextend(BMR_GET_SMGR(bmr), fork, first_block, extend_by, false);
2920 :
2921 : /*
2922 : * Release the file-extension lock; it's now OK for someone else to extend
2923 : * the relation some more.
2924 : *
2925 : * We remove IO_IN_PROGRESS after this, as waking up waiting backends can
2926 : * take noticeable time.
2927 : */
2928 409164 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2929 304960 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2930 :
2931 409164 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context, IOOP_EXTEND,
2932 409164 : io_start, 1, extend_by * BLCKSZ);
2933 :
2934 : /* Set BM_VALID, terminate IO, and wake up any waiters */
2935 861034 : for (uint32 i = 0; i < extend_by; i++)
2936 : {
2937 451870 : Buffer buf = buffers[i];
2938 451870 : BufferDesc *buf_hdr = GetBufferDescriptor(buf - 1);
2939 451870 : bool lock = false;
2940 :
2941 451870 : if (flags & EB_LOCK_FIRST && i == 0)
2942 300820 : lock = true;
2943 151050 : else if (flags & EB_LOCK_TARGET)
2944 : {
2945 : Assert(extend_upto != InvalidBlockNumber);
2946 89908 : if (first_block + i + 1 == extend_upto)
2947 89018 : lock = true;
2948 : }
2949 :
2950 451870 : if (lock)
2951 389838 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
2952 :
2953 451870 : TerminateBufferIO(buf_hdr, false, BM_VALID, true, false);
2954 : }
2955 :
2956 409164 : pgBufferUsage.shared_blks_written += extend_by;
2957 :
2958 409164 : *extended_by = extend_by;
2959 :
2960 409164 : return first_block;
2961 : }
2962 :
2963 : /*
2964 : * BufferIsLockedByMe
2965 : *
2966 : * Checks if this backend has the buffer locked in any mode.
2967 : *
2968 : * Buffer must be pinned.
2969 : */
2970 : bool
2971 0 : BufferIsLockedByMe(Buffer buffer)
2972 : {
2973 : BufferDesc *bufHdr;
2974 :
2975 : Assert(BufferIsPinned(buffer));
2976 :
2977 0 : if (BufferIsLocal(buffer))
2978 : {
2979 : /* Content locks are not maintained for local buffers. */
2980 0 : return true;
2981 : }
2982 : else
2983 : {
2984 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2985 0 : return BufferLockHeldByMe(bufHdr);
2986 : }
2987 : }
2988 :
2989 : /*
2990 : * BufferIsLockedByMeInMode
2991 : *
2992 : * Checks if this backend has the buffer locked in the specified mode.
2993 : *
2994 : * Buffer must be pinned.
2995 : */
2996 : bool
2997 0 : BufferIsLockedByMeInMode(Buffer buffer, BufferLockMode mode)
2998 : {
2999 : BufferDesc *bufHdr;
3000 :
3001 : Assert(BufferIsPinned(buffer));
3002 :
3003 0 : if (BufferIsLocal(buffer))
3004 : {
3005 : /* Content locks are not maintained for local buffers. */
3006 0 : return true;
3007 : }
3008 : else
3009 : {
3010 0 : bufHdr = GetBufferDescriptor(buffer - 1);
3011 0 : return BufferLockHeldByMeInMode(bufHdr, mode);
3012 : }
3013 : }
3014 :
3015 : /*
3016 : * BufferIsDirty
3017 : *
3018 : * Checks if buffer is already dirty.
3019 : *
3020 : * Buffer must be pinned and exclusive-locked. (Without an exclusive lock,
3021 : * the result may be stale before it's returned.)
3022 : */
3023 : bool
3024 0 : BufferIsDirty(Buffer buffer)
3025 : {
3026 : BufferDesc *bufHdr;
3027 :
3028 : Assert(BufferIsPinned(buffer));
3029 :
3030 0 : if (BufferIsLocal(buffer))
3031 : {
3032 0 : int bufid = -buffer - 1;
3033 :
3034 0 : bufHdr = GetLocalBufferDescriptor(bufid);
3035 : /* Content locks are not maintained for local buffers. */
3036 : }
3037 : else
3038 : {
3039 0 : bufHdr = GetBufferDescriptor(buffer - 1);
3040 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
3041 : }
3042 :
3043 0 : return pg_atomic_read_u64(&bufHdr->state) & BM_DIRTY;
3044 : }
3045 :
3046 : /*
3047 : * MarkBufferDirty
3048 : *
3049 : * Marks buffer contents as dirty (actual write happens later).
3050 : *
3051 : * Buffer must be pinned and exclusive-locked. (If caller does not hold
3052 : * exclusive lock, then somebody could be in process of writing the buffer,
3053 : * leading to risk of bad data written to disk.)
3054 : */
3055 : void
3056 43595016 : MarkBufferDirty(Buffer buffer)
3057 : {
3058 : BufferDesc *bufHdr;
3059 : uint64 buf_state;
3060 : uint64 old_buf_state;
3061 :
3062 43595016 : if (!BufferIsValid(buffer))
3063 0 : elog(ERROR, "bad buffer ID: %d", buffer);
3064 :
3065 43595016 : if (BufferIsLocal(buffer))
3066 : {
3067 2444718 : MarkLocalBufferDirty(buffer);
3068 2444718 : return;
3069 : }
3070 :
3071 41150298 : bufHdr = GetBufferDescriptor(buffer - 1);
3072 :
3073 : Assert(BufferIsPinned(buffer));
3074 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
3075 :
3076 : /*
3077 : * NB: We have to wait for the buffer header spinlock to be not held, as
3078 : * TerminateBufferIO() relies on the spinlock.
3079 : */
3080 41150298 : old_buf_state = pg_atomic_read_u64(&bufHdr->state);
3081 : for (;;)
3082 : {
3083 41150650 : if (old_buf_state & BM_LOCKED)
3084 532 : old_buf_state = WaitBufHdrUnlocked(bufHdr);
3085 :
3086 41150650 : buf_state = old_buf_state;
3087 :
3088 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
3089 41150650 : buf_state |= BM_DIRTY | BM_JUST_DIRTIED;
3090 :
3091 41150650 : if (pg_atomic_compare_exchange_u64(&bufHdr->state, &old_buf_state,
3092 : buf_state))
3093 41150298 : break;
3094 : }
3095 :
3096 : /*
3097 : * If the buffer was not dirty already, do vacuum accounting.
3098 : */
3099 41150298 : if (!(old_buf_state & BM_DIRTY))
3100 : {
3101 1321872 : pgBufferUsage.shared_blks_dirtied++;
3102 1321872 : if (VacuumCostActive)
3103 13634 : VacuumCostBalance += VacuumCostPageDirty;
3104 : }
3105 : }
3106 :
3107 : /*
3108 : * ReleaseAndReadBuffer -- combine ReleaseBuffer() and ReadBuffer()
3109 : *
3110 : * Formerly, this saved one cycle of acquiring/releasing the BufMgrLock
3111 : * compared to calling the two routines separately. Now it's mainly just
3112 : * a convenience function. However, if the passed buffer is valid and
3113 : * already contains the desired block, we just return it as-is; and that
3114 : * does save considerable work compared to a full release and reacquire.
3115 : *
3116 : * Note: it is OK to pass buffer == InvalidBuffer, indicating that no old
3117 : * buffer actually needs to be released. This case is the same as ReadBuffer,
3118 : * but can save some tests in the caller.
3119 : */
3120 : Buffer
3121 58486720 : ReleaseAndReadBuffer(Buffer buffer,
3122 : Relation relation,
3123 : BlockNumber blockNum)
3124 : {
3125 58486720 : ForkNumber forkNum = MAIN_FORKNUM;
3126 : BufferDesc *bufHdr;
3127 :
3128 58486720 : if (BufferIsValid(buffer))
3129 : {
3130 : Assert(BufferIsPinned(buffer));
3131 35306412 : if (BufferIsLocal(buffer))
3132 : {
3133 73728 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
3134 80772 : if (bufHdr->tag.blockNum == blockNum &&
3135 14088 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3136 7044 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3137 7044 : return buffer;
3138 66684 : UnpinLocalBuffer(buffer);
3139 : }
3140 : else
3141 : {
3142 35232684 : bufHdr = GetBufferDescriptor(buffer - 1);
3143 : /* we have pin, so it's ok to examine tag without spinlock */
3144 47249138 : if (bufHdr->tag.blockNum == blockNum &&
3145 24032908 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3146 12016454 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3147 12016454 : return buffer;
3148 23216230 : UnpinBuffer(bufHdr);
3149 : }
3150 : }
3151 :
3152 46463222 : return ReadBuffer(relation, blockNum);
3153 : }
3154 :
3155 : /*
3156 : * PinBuffer -- make buffer unavailable for replacement.
3157 : *
3158 : * For the default access strategy, the buffer's usage_count is incremented
3159 : * when we first pin it; for other strategies we just make sure the usage_count
3160 : * isn't zero. (The idea of the latter is that we don't want synchronized
3161 : * heap scans to inflate the count, but we need it to not be zero to discourage
3162 : * other backends from stealing buffers from our ring. As long as we cycle
3163 : * through the ring faster than the global clock-sweep cycles, buffers in
3164 : * our ring won't be chosen as victims for replacement by other backends.)
3165 : *
3166 : * This should be applied only to shared buffers, never local ones.
3167 : *
3168 : * Since buffers are pinned/unpinned very frequently, pin buffers without
3169 : * taking the buffer header lock; instead update the state variable in loop of
3170 : * CAS operations. Hopefully it's just a single CAS.
3171 : *
3172 : * Note that ResourceOwnerEnlarge() and ReservePrivateRefCountEntry()
3173 : * must have been done already.
3174 : *
3175 : * Returns true if buffer is BM_VALID, else false. This provision allows
3176 : * some callers to avoid an extra spinlock cycle. If skip_if_not_valid is
3177 : * true, then a false return value also indicates that the buffer was
3178 : * (recently) invalid and has not been pinned.
3179 : */
3180 : static bool
3181 118570154 : PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
3182 : bool skip_if_not_valid)
3183 : {
3184 118570154 : Buffer b = BufferDescriptorGetBuffer(buf);
3185 : bool result;
3186 : PrivateRefCountEntry *ref;
3187 :
3188 : Assert(!BufferIsLocal(b));
3189 : Assert(ReservedRefCountSlot != -1);
3190 :
3191 118570154 : ref = GetPrivateRefCountEntry(b, true);
3192 :
3193 118570154 : if (ref == NULL)
3194 : {
3195 : uint64 buf_state;
3196 : uint64 old_buf_state;
3197 :
3198 114034136 : old_buf_state = pg_atomic_read_u64(&buf->state);
3199 : for (;;)
3200 : {
3201 114082768 : if (unlikely(skip_if_not_valid && !(old_buf_state & BM_VALID)))
3202 12 : return false;
3203 :
3204 : /*
3205 : * We're not allowed to increase the refcount while the buffer
3206 : * header spinlock is held. Wait for the lock to be released.
3207 : */
3208 114082756 : if (old_buf_state & BM_LOCKED)
3209 1070 : old_buf_state = WaitBufHdrUnlocked(buf);
3210 :
3211 114082756 : buf_state = old_buf_state;
3212 :
3213 : /* increase refcount */
3214 114082756 : buf_state += BUF_REFCOUNT_ONE;
3215 :
3216 114082756 : if (strategy == NULL)
3217 : {
3218 : /* Default case: increase usagecount unless already max. */
3219 112760646 : if (BUF_STATE_GET_USAGECOUNT(buf_state) < BM_MAX_USAGE_COUNT)
3220 6837432 : buf_state += BUF_USAGECOUNT_ONE;
3221 : }
3222 : else
3223 : {
3224 : /*
3225 : * Ring buffers shouldn't evict others from pool. Thus we
3226 : * don't make usagecount more than 1.
3227 : */
3228 1322110 : if (BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3229 71132 : buf_state += BUF_USAGECOUNT_ONE;
3230 : }
3231 :
3232 114082756 : if (pg_atomic_compare_exchange_u64(&buf->state, &old_buf_state,
3233 : buf_state))
3234 : {
3235 114034124 : result = (buf_state & BM_VALID) != 0;
3236 :
3237 114034124 : TrackNewBufferPin(b);
3238 114034124 : break;
3239 : }
3240 : }
3241 : }
3242 : else
3243 : {
3244 : /*
3245 : * If we previously pinned the buffer, it is likely to be valid, but
3246 : * it may not be if StartReadBuffers() was called and
3247 : * WaitReadBuffers() hasn't been called yet. We'll check by loading
3248 : * the flags without locking. This is racy, but it's OK to return
3249 : * false spuriously: when WaitReadBuffers() calls StartBufferIO(),
3250 : * it'll see that it's now valid.
3251 : *
3252 : * Note: We deliberately avoid a Valgrind client request here.
3253 : * Individual access methods can optionally superimpose buffer page
3254 : * client requests on top of our client requests to enforce that
3255 : * buffers are only accessed while locked (and pinned). It's possible
3256 : * that the buffer page is legitimately non-accessible here. We
3257 : * cannot meddle with that.
3258 : */
3259 4536018 : result = (pg_atomic_read_u64(&buf->state) & BM_VALID) != 0;
3260 :
3261 : Assert(ref->data.refcount > 0);
3262 4536018 : ref->data.refcount++;
3263 4536018 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
3264 : }
3265 :
3266 118570142 : return result;
3267 : }
3268 :
3269 : /*
3270 : * PinBuffer_Locked -- as above, but caller already locked the buffer header.
3271 : * The spinlock is released before return.
3272 : *
3273 : * As this function is called with the spinlock held, the caller has to
3274 : * previously call ReservePrivateRefCountEntry() and
3275 : * ResourceOwnerEnlarge(CurrentResourceOwner);
3276 : *
3277 : * Currently, no callers of this function want to modify the buffer's
3278 : * usage_count at all, so there's no need for a strategy parameter.
3279 : * Also we don't bother with a BM_VALID test (the caller could check that for
3280 : * itself).
3281 : *
3282 : * Also all callers only ever use this function when it's known that the
3283 : * buffer can't have a preexisting pin by this backend. That allows us to skip
3284 : * searching the private refcount array & hash, which is a boon, because the
3285 : * spinlock is still held.
3286 : *
3287 : * Note: use of this routine is frequently mandatory, not just an optimization
3288 : * to save a spin lock/unlock cycle, because we need to pin a buffer before
3289 : * its state can change under us.
3290 : */
3291 : static void
3292 614582 : PinBuffer_Locked(BufferDesc *buf)
3293 : {
3294 : uint64 old_buf_state;
3295 :
3296 : /*
3297 : * As explained, We don't expect any preexisting pins. That allows us to
3298 : * manipulate the PrivateRefCount after releasing the spinlock
3299 : */
3300 : Assert(GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf), false) == NULL);
3301 :
3302 : /*
3303 : * Since we hold the buffer spinlock, we can update the buffer state and
3304 : * release the lock in one operation.
3305 : */
3306 614582 : old_buf_state = pg_atomic_read_u64(&buf->state);
3307 :
3308 614582 : UnlockBufHdrExt(buf, old_buf_state,
3309 : 0, 0, 1);
3310 :
3311 614582 : TrackNewBufferPin(BufferDescriptorGetBuffer(buf));
3312 614582 : }
3313 :
3314 : /*
3315 : * Support for waking up another backend that is waiting for the cleanup lock
3316 : * to be released using BM_PIN_COUNT_WAITER.
3317 : *
3318 : * See LockBufferForCleanup().
3319 : *
3320 : * Expected to be called just after releasing a buffer pin (in a BufferDesc,
3321 : * not just reducing the backend-local pincount for the buffer).
3322 : */
3323 : static void
3324 184 : WakePinCountWaiter(BufferDesc *buf)
3325 : {
3326 : /*
3327 : * Acquire the buffer header lock, re-check that there's a waiter. Another
3328 : * backend could have unpinned this buffer, and already woken up the
3329 : * waiter.
3330 : *
3331 : * There's no danger of the buffer being replaced after we unpinned it
3332 : * above, as it's pinned by the waiter. The waiter removes
3333 : * BM_PIN_COUNT_WAITER if it stops waiting for a reason other than this
3334 : * backend waking it up.
3335 : */
3336 184 : uint64 buf_state = LockBufHdr(buf);
3337 :
3338 184 : if ((buf_state & BM_PIN_COUNT_WAITER) &&
3339 184 : BUF_STATE_GET_REFCOUNT(buf_state) == 1)
3340 184 : {
3341 : /* we just released the last pin other than the waiter's */
3342 184 : int wait_backend_pgprocno = buf->wait_backend_pgprocno;
3343 :
3344 184 : UnlockBufHdrExt(buf, buf_state,
3345 : 0, BM_PIN_COUNT_WAITER,
3346 : 0);
3347 184 : ProcSendSignal(wait_backend_pgprocno);
3348 : }
3349 : else
3350 0 : UnlockBufHdr(buf);
3351 184 : }
3352 :
3353 : /*
3354 : * UnpinBuffer -- make buffer available for replacement.
3355 : *
3356 : * This should be applied only to shared buffers, never local ones. This
3357 : * always adjusts CurrentResourceOwner.
3358 : */
3359 : static void
3360 145961598 : UnpinBuffer(BufferDesc *buf)
3361 : {
3362 145961598 : Buffer b = BufferDescriptorGetBuffer(buf);
3363 :
3364 145961598 : ResourceOwnerForgetBuffer(CurrentResourceOwner, b);
3365 145961598 : UnpinBufferNoOwner(buf);
3366 145961598 : }
3367 :
3368 : static void
3369 145970702 : UnpinBufferNoOwner(BufferDesc *buf)
3370 : {
3371 : PrivateRefCountEntry *ref;
3372 145970702 : Buffer b = BufferDescriptorGetBuffer(buf);
3373 :
3374 : Assert(!BufferIsLocal(b));
3375 :
3376 : /* not moving as we're likely deleting it soon anyway */
3377 145970702 : ref = GetPrivateRefCountEntry(b, false);
3378 : Assert(ref != NULL);
3379 : Assert(ref->data.refcount > 0);
3380 145970702 : ref->data.refcount--;
3381 145970702 : if (ref->data.refcount == 0)
3382 : {
3383 : uint64 old_buf_state;
3384 :
3385 : /*
3386 : * Mark buffer non-accessible to Valgrind.
3387 : *
3388 : * Note that the buffer may have already been marked non-accessible
3389 : * within access method code that enforces that buffers are only
3390 : * accessed while a buffer lock is held.
3391 : */
3392 : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
3393 :
3394 : /*
3395 : * I'd better not still hold the buffer content lock. Can't use
3396 : * BufferIsLockedByMe(), as that asserts the buffer is pinned.
3397 : */
3398 : Assert(!BufferLockHeldByMe(buf));
3399 :
3400 : /* decrement the shared reference count */
3401 118571250 : old_buf_state = pg_atomic_fetch_sub_u64(&buf->state, BUF_REFCOUNT_ONE);
3402 :
3403 : /* Support LockBufferForCleanup() */
3404 118571250 : if (old_buf_state & BM_PIN_COUNT_WAITER)
3405 184 : WakePinCountWaiter(buf);
3406 :
3407 118571250 : ForgetPrivateRefCountEntry(ref);
3408 : }
3409 145970702 : }
3410 :
3411 : /*
3412 : * Set up backend-local tracking of a buffer pinned the first time by this
3413 : * backend.
3414 : */
3415 : inline void
3416 118571250 : TrackNewBufferPin(Buffer buf)
3417 : {
3418 : PrivateRefCountEntry *ref;
3419 :
3420 118571250 : ref = NewPrivateRefCountEntry(buf);
3421 118571250 : ref->data.refcount++;
3422 :
3423 118571250 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buf);
3424 :
3425 : /*
3426 : * This is the first pin for this page by this backend, mark its page as
3427 : * defined to valgrind. While the page contents might not actually be
3428 : * valid yet, we don't currently guarantee that such pages are marked
3429 : * undefined or non-accessible.
3430 : *
3431 : * It's not necessarily the prettiest to do this here, but otherwise we'd
3432 : * need this block of code in multiple places.
3433 : */
3434 : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(GetBufferDescriptor(buf - 1)),
3435 : BLCKSZ);
3436 118571250 : }
3437 :
3438 : #define ST_SORT sort_checkpoint_bufferids
3439 : #define ST_ELEMENT_TYPE CkptSortItem
3440 : #define ST_COMPARE(a, b) ckpt_buforder_comparator(a, b)
3441 : #define ST_SCOPE static
3442 : #define ST_DEFINE
3443 : #include "lib/sort_template.h"
3444 :
3445 : /*
3446 : * BufferSync -- Write out all dirty buffers in the pool.
3447 : *
3448 : * This is called at checkpoint time to write out all dirty shared buffers.
3449 : * The checkpoint request flags should be passed in. If CHECKPOINT_FAST is
3450 : * set, we disable delays between writes; if CHECKPOINT_IS_SHUTDOWN,
3451 : * CHECKPOINT_END_OF_RECOVERY or CHECKPOINT_FLUSH_UNLOGGED is set, we write
3452 : * even unlogged buffers, which are otherwise skipped. The remaining flags
3453 : * currently have no effect here.
3454 : */
3455 : static void
3456 3560 : BufferSync(int flags)
3457 : {
3458 : uint64 buf_state;
3459 : int buf_id;
3460 : int num_to_scan;
3461 : int num_spaces;
3462 : int num_processed;
3463 : int num_written;
3464 3560 : CkptTsStatus *per_ts_stat = NULL;
3465 : Oid last_tsid;
3466 : binaryheap *ts_heap;
3467 : int i;
3468 3560 : uint64 mask = BM_DIRTY;
3469 : WritebackContext wb_context;
3470 :
3471 : /*
3472 : * Unless this is a shutdown checkpoint or we have been explicitly told,
3473 : * we write only permanent, dirty buffers. But at shutdown or end of
3474 : * recovery, we write all dirty buffers.
3475 : */
3476 3560 : if (!((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
3477 : CHECKPOINT_FLUSH_UNLOGGED))))
3478 1988 : mask |= BM_PERMANENT;
3479 :
3480 : /*
3481 : * Loop over all buffers, and mark the ones that need to be written with
3482 : * BM_CHECKPOINT_NEEDED. Count them as we go (num_to_scan), so that we
3483 : * can estimate how much work needs to be done.
3484 : *
3485 : * This allows us to write only those pages that were dirty when the
3486 : * checkpoint began, and not those that get dirtied while it proceeds.
3487 : * Whenever a page with BM_CHECKPOINT_NEEDED is written out, either by us
3488 : * later in this function, or by normal backends or the bgwriter cleaning
3489 : * scan, the flag is cleared. Any buffer dirtied after this point won't
3490 : * have the flag set.
3491 : *
3492 : * Note that if we fail to write some buffer, we may leave buffers with
3493 : * BM_CHECKPOINT_NEEDED still set. This is OK since any such buffer would
3494 : * certainly need to be written for the next checkpoint attempt, too.
3495 : */
3496 3560 : num_to_scan = 0;
3497 24810120 : for (buf_id = 0; buf_id < NBuffers; buf_id++)
3498 : {
3499 24806560 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3500 24806560 : uint64 set_bits = 0;
3501 :
3502 : /*
3503 : * Header spinlock is enough to examine BM_DIRTY, see comment in
3504 : * SyncOneBuffer.
3505 : */
3506 24806560 : buf_state = LockBufHdr(bufHdr);
3507 :
3508 24806560 : if ((buf_state & mask) == mask)
3509 : {
3510 : CkptSortItem *item;
3511 :
3512 594318 : set_bits = BM_CHECKPOINT_NEEDED;
3513 :
3514 594318 : item = &CkptBufferIds[num_to_scan++];
3515 594318 : item->buf_id = buf_id;
3516 594318 : item->tsId = bufHdr->tag.spcOid;
3517 594318 : item->relNumber = BufTagGetRelNumber(&bufHdr->tag);
3518 594318 : item->forkNum = BufTagGetForkNum(&bufHdr->tag);
3519 594318 : item->blockNum = bufHdr->tag.blockNum;
3520 : }
3521 :
3522 24806560 : UnlockBufHdrExt(bufHdr, buf_state,
3523 : set_bits, 0,
3524 : 0);
3525 :
3526 : /* Check for barrier events in case NBuffers is large. */
3527 24806560 : if (ProcSignalBarrierPending)
3528 0 : ProcessProcSignalBarrier();
3529 : }
3530 :
3531 3560 : if (num_to_scan == 0)
3532 1370 : return; /* nothing to do */
3533 :
3534 2190 : WritebackContextInit(&wb_context, &checkpoint_flush_after);
3535 :
3536 : TRACE_POSTGRESQL_BUFFER_SYNC_START(NBuffers, num_to_scan);
3537 :
3538 : /*
3539 : * Sort buffers that need to be written to reduce the likelihood of random
3540 : * IO. The sorting is also important for the implementation of balancing
3541 : * writes between tablespaces. Without balancing writes we'd potentially
3542 : * end up writing to the tablespaces one-by-one; possibly overloading the
3543 : * underlying system.
3544 : */
3545 2190 : sort_checkpoint_bufferids(CkptBufferIds, num_to_scan);
3546 :
3547 2190 : num_spaces = 0;
3548 :
3549 : /*
3550 : * Allocate progress status for each tablespace with buffers that need to
3551 : * be flushed. This requires the to-be-flushed array to be sorted.
3552 : */
3553 2190 : last_tsid = InvalidOid;
3554 596508 : for (i = 0; i < num_to_scan; i++)
3555 : {
3556 : CkptTsStatus *s;
3557 : Oid cur_tsid;
3558 :
3559 594318 : cur_tsid = CkptBufferIds[i].tsId;
3560 :
3561 : /*
3562 : * Grow array of per-tablespace status structs, every time a new
3563 : * tablespace is found.
3564 : */
3565 594318 : if (last_tsid == InvalidOid || last_tsid != cur_tsid)
3566 3340 : {
3567 : Size sz;
3568 :
3569 3340 : num_spaces++;
3570 :
3571 : /*
3572 : * Not worth adding grow-by-power-of-2 logic here - even with a
3573 : * few hundred tablespaces this should be fine.
3574 : */
3575 3340 : sz = sizeof(CkptTsStatus) * num_spaces;
3576 :
3577 3340 : if (per_ts_stat == NULL)
3578 2190 : per_ts_stat = (CkptTsStatus *) palloc(sz);
3579 : else
3580 1150 : per_ts_stat = (CkptTsStatus *) repalloc(per_ts_stat, sz);
3581 :
3582 3340 : s = &per_ts_stat[num_spaces - 1];
3583 3340 : memset(s, 0, sizeof(*s));
3584 3340 : s->tsId = cur_tsid;
3585 :
3586 : /*
3587 : * The first buffer in this tablespace. As CkptBufferIds is sorted
3588 : * by tablespace all (s->num_to_scan) buffers in this tablespace
3589 : * will follow afterwards.
3590 : */
3591 3340 : s->index = i;
3592 :
3593 : /*
3594 : * progress_slice will be determined once we know how many buffers
3595 : * are in each tablespace, i.e. after this loop.
3596 : */
3597 :
3598 3340 : last_tsid = cur_tsid;
3599 : }
3600 : else
3601 : {
3602 590978 : s = &per_ts_stat[num_spaces - 1];
3603 : }
3604 :
3605 594318 : s->num_to_scan++;
3606 :
3607 : /* Check for barrier events. */
3608 594318 : if (ProcSignalBarrierPending)
3609 0 : ProcessProcSignalBarrier();
3610 : }
3611 :
3612 : Assert(num_spaces > 0);
3613 :
3614 : /*
3615 : * Build a min-heap over the write-progress in the individual tablespaces,
3616 : * and compute how large a portion of the total progress a single
3617 : * processed buffer is.
3618 : */
3619 2190 : ts_heap = binaryheap_allocate(num_spaces,
3620 : ts_ckpt_progress_comparator,
3621 : NULL);
3622 :
3623 5530 : for (i = 0; i < num_spaces; i++)
3624 : {
3625 3340 : CkptTsStatus *ts_stat = &per_ts_stat[i];
3626 :
3627 3340 : ts_stat->progress_slice = (float8) num_to_scan / ts_stat->num_to_scan;
3628 :
3629 3340 : binaryheap_add_unordered(ts_heap, PointerGetDatum(ts_stat));
3630 : }
3631 :
3632 2190 : binaryheap_build(ts_heap);
3633 :
3634 : /*
3635 : * Iterate through to-be-checkpointed buffers and write the ones (still)
3636 : * marked with BM_CHECKPOINT_NEEDED. The writes are balanced between
3637 : * tablespaces; otherwise the sorting would lead to only one tablespace
3638 : * receiving writes at a time, making inefficient use of the hardware.
3639 : */
3640 2190 : num_processed = 0;
3641 2190 : num_written = 0;
3642 596508 : while (!binaryheap_empty(ts_heap))
3643 : {
3644 594318 : BufferDesc *bufHdr = NULL;
3645 : CkptTsStatus *ts_stat = (CkptTsStatus *)
3646 594318 : DatumGetPointer(binaryheap_first(ts_heap));
3647 :
3648 594318 : buf_id = CkptBufferIds[ts_stat->index].buf_id;
3649 : Assert(buf_id != -1);
3650 :
3651 594318 : bufHdr = GetBufferDescriptor(buf_id);
3652 :
3653 594318 : num_processed++;
3654 :
3655 : /*
3656 : * We don't need to acquire the lock here, because we're only looking
3657 : * at a single bit. It's possible that someone else writes the buffer
3658 : * and clears the flag right after we check, but that doesn't matter
3659 : * since SyncOneBuffer will then do nothing. However, there is a
3660 : * further race condition: it's conceivable that between the time we
3661 : * examine the bit here and the time SyncOneBuffer acquires the lock,
3662 : * someone else not only wrote the buffer but replaced it with another
3663 : * page and dirtied it. In that improbable case, SyncOneBuffer will
3664 : * write the buffer though we didn't need to. It doesn't seem worth
3665 : * guarding against this, though.
3666 : */
3667 594318 : if (pg_atomic_read_u64(&bufHdr->state) & BM_CHECKPOINT_NEEDED)
3668 : {
3669 555870 : if (SyncOneBuffer(buf_id, false, &wb_context) & BUF_WRITTEN)
3670 : {
3671 : TRACE_POSTGRESQL_BUFFER_SYNC_WRITTEN(buf_id);
3672 555870 : PendingCheckpointerStats.buffers_written++;
3673 555870 : num_written++;
3674 : }
3675 : }
3676 :
3677 : /*
3678 : * Measure progress independent of actually having to flush the buffer
3679 : * - otherwise writing become unbalanced.
3680 : */
3681 594318 : ts_stat->progress += ts_stat->progress_slice;
3682 594318 : ts_stat->num_scanned++;
3683 594318 : ts_stat->index++;
3684 :
3685 : /* Have all the buffers from the tablespace been processed? */
3686 594318 : if (ts_stat->num_scanned == ts_stat->num_to_scan)
3687 : {
3688 3340 : binaryheap_remove_first(ts_heap);
3689 : }
3690 : else
3691 : {
3692 : /* update heap with the new progress */
3693 590978 : binaryheap_replace_first(ts_heap, PointerGetDatum(ts_stat));
3694 : }
3695 :
3696 : /*
3697 : * Sleep to throttle our I/O rate.
3698 : *
3699 : * (This will check for barrier events even if it doesn't sleep.)
3700 : */
3701 594318 : CheckpointWriteDelay(flags, (double) num_processed / num_to_scan);
3702 : }
3703 :
3704 : /*
3705 : * Issue all pending flushes. Only checkpointer calls BufferSync(), so
3706 : * IOContext will always be IOCONTEXT_NORMAL.
3707 : */
3708 2190 : IssuePendingWritebacks(&wb_context, IOCONTEXT_NORMAL);
3709 :
3710 2190 : pfree(per_ts_stat);
3711 2190 : per_ts_stat = NULL;
3712 2190 : binaryheap_free(ts_heap);
3713 :
3714 : /*
3715 : * Update checkpoint statistics. As noted above, this doesn't include
3716 : * buffers written by other backends or bgwriter scan.
3717 : */
3718 2190 : CheckpointStats.ckpt_bufs_written += num_written;
3719 :
3720 : TRACE_POSTGRESQL_BUFFER_SYNC_DONE(NBuffers, num_written, num_to_scan);
3721 : }
3722 :
3723 : /*
3724 : * BgBufferSync -- Write out some dirty buffers in the pool.
3725 : *
3726 : * This is called periodically by the background writer process.
3727 : *
3728 : * Returns true if it's appropriate for the bgwriter process to go into
3729 : * low-power hibernation mode. (This happens if the strategy clock-sweep
3730 : * has been "lapped" and no buffer allocations have occurred recently,
3731 : * or if the bgwriter has been effectively disabled by setting
3732 : * bgwriter_lru_maxpages to 0.)
3733 : */
3734 : bool
3735 27200 : BgBufferSync(WritebackContext *wb_context)
3736 : {
3737 : /* info obtained from freelist.c */
3738 : int strategy_buf_id;
3739 : uint32 strategy_passes;
3740 : uint32 recent_alloc;
3741 :
3742 : /*
3743 : * Information saved between calls so we can determine the strategy
3744 : * point's advance rate and avoid scanning already-cleaned buffers.
3745 : */
3746 : static bool saved_info_valid = false;
3747 : static int prev_strategy_buf_id;
3748 : static uint32 prev_strategy_passes;
3749 : static int next_to_clean;
3750 : static uint32 next_passes;
3751 :
3752 : /* Moving averages of allocation rate and clean-buffer density */
3753 : static float smoothed_alloc = 0;
3754 : static float smoothed_density = 10.0;
3755 :
3756 : /* Potentially these could be tunables, but for now, not */
3757 27200 : float smoothing_samples = 16;
3758 27200 : float scan_whole_pool_milliseconds = 120000.0;
3759 :
3760 : /* Used to compute how far we scan ahead */
3761 : long strategy_delta;
3762 : int bufs_to_lap;
3763 : int bufs_ahead;
3764 : float scans_per_alloc;
3765 : int reusable_buffers_est;
3766 : int upcoming_alloc_est;
3767 : int min_scan_buffers;
3768 :
3769 : /* Variables for the scanning loop proper */
3770 : int num_to_scan;
3771 : int num_written;
3772 : int reusable_buffers;
3773 :
3774 : /* Variables for final smoothed_density update */
3775 : long new_strategy_delta;
3776 : uint32 new_recent_alloc;
3777 :
3778 : /*
3779 : * Find out where the clock-sweep currently is, and how many buffer
3780 : * allocations have happened since our last call.
3781 : */
3782 27200 : strategy_buf_id = StrategySyncStart(&strategy_passes, &recent_alloc);
3783 :
3784 : /* Report buffer alloc counts to pgstat */
3785 27200 : PendingBgWriterStats.buf_alloc += recent_alloc;
3786 :
3787 : /*
3788 : * If we're not running the LRU scan, just stop after doing the stats
3789 : * stuff. We mark the saved state invalid so that we can recover sanely
3790 : * if LRU scan is turned back on later.
3791 : */
3792 27200 : if (bgwriter_lru_maxpages <= 0)
3793 : {
3794 86 : saved_info_valid = false;
3795 86 : return true;
3796 : }
3797 :
3798 : /*
3799 : * Compute strategy_delta = how many buffers have been scanned by the
3800 : * clock-sweep since last time. If first time through, assume none. Then
3801 : * see if we are still ahead of the clock-sweep, and if so, how many
3802 : * buffers we could scan before we'd catch up with it and "lap" it. Note:
3803 : * weird-looking coding of xxx_passes comparisons are to avoid bogus
3804 : * behavior when the passes counts wrap around.
3805 : */
3806 27114 : if (saved_info_valid)
3807 : {
3808 25984 : int32 passes_delta = strategy_passes - prev_strategy_passes;
3809 :
3810 25984 : strategy_delta = strategy_buf_id - prev_strategy_buf_id;
3811 25984 : strategy_delta += (long) passes_delta * NBuffers;
3812 :
3813 : Assert(strategy_delta >= 0);
3814 :
3815 25984 : if ((int32) (next_passes - strategy_passes) > 0)
3816 : {
3817 : /* we're one pass ahead of the strategy point */
3818 6024 : bufs_to_lap = strategy_buf_id - next_to_clean;
3819 : #ifdef BGW_DEBUG
3820 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3821 : next_passes, next_to_clean,
3822 : strategy_passes, strategy_buf_id,
3823 : strategy_delta, bufs_to_lap);
3824 : #endif
3825 : }
3826 19960 : else if (next_passes == strategy_passes &&
3827 15332 : next_to_clean >= strategy_buf_id)
3828 : {
3829 : /* on same pass, but ahead or at least not behind */
3830 13826 : bufs_to_lap = NBuffers - (next_to_clean - strategy_buf_id);
3831 : #ifdef BGW_DEBUG
3832 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3833 : next_passes, next_to_clean,
3834 : strategy_passes, strategy_buf_id,
3835 : strategy_delta, bufs_to_lap);
3836 : #endif
3837 : }
3838 : else
3839 : {
3840 : /*
3841 : * We're behind, so skip forward to the strategy point and start
3842 : * cleaning from there.
3843 : */
3844 : #ifdef BGW_DEBUG
3845 : elog(DEBUG2, "bgwriter behind: bgw %u-%u strategy %u-%u delta=%ld",
3846 : next_passes, next_to_clean,
3847 : strategy_passes, strategy_buf_id,
3848 : strategy_delta);
3849 : #endif
3850 6134 : next_to_clean = strategy_buf_id;
3851 6134 : next_passes = strategy_passes;
3852 6134 : bufs_to_lap = NBuffers;
3853 : }
3854 : }
3855 : else
3856 : {
3857 : /*
3858 : * Initializing at startup or after LRU scanning had been off. Always
3859 : * start at the strategy point.
3860 : */
3861 : #ifdef BGW_DEBUG
3862 : elog(DEBUG2, "bgwriter initializing: strategy %u-%u",
3863 : strategy_passes, strategy_buf_id);
3864 : #endif
3865 1130 : strategy_delta = 0;
3866 1130 : next_to_clean = strategy_buf_id;
3867 1130 : next_passes = strategy_passes;
3868 1130 : bufs_to_lap = NBuffers;
3869 : }
3870 :
3871 : /* Update saved info for next time */
3872 27114 : prev_strategy_buf_id = strategy_buf_id;
3873 27114 : prev_strategy_passes = strategy_passes;
3874 27114 : saved_info_valid = true;
3875 :
3876 : /*
3877 : * Compute how many buffers had to be scanned for each new allocation, ie,
3878 : * 1/density of reusable buffers, and track a moving average of that.
3879 : *
3880 : * If the strategy point didn't move, we don't update the density estimate
3881 : */
3882 27114 : if (strategy_delta > 0 && recent_alloc > 0)
3883 : {
3884 14060 : scans_per_alloc = (float) strategy_delta / (float) recent_alloc;
3885 14060 : smoothed_density += (scans_per_alloc - smoothed_density) /
3886 : smoothing_samples;
3887 : }
3888 :
3889 : /*
3890 : * Estimate how many reusable buffers there are between the current
3891 : * strategy point and where we've scanned ahead to, based on the smoothed
3892 : * density estimate.
3893 : */
3894 27114 : bufs_ahead = NBuffers - bufs_to_lap;
3895 27114 : reusable_buffers_est = (float) bufs_ahead / smoothed_density;
3896 :
3897 : /*
3898 : * Track a moving average of recent buffer allocations. Here, rather than
3899 : * a true average we want a fast-attack, slow-decline behavior: we
3900 : * immediately follow any increase.
3901 : */
3902 27114 : if (smoothed_alloc <= (float) recent_alloc)
3903 7412 : smoothed_alloc = recent_alloc;
3904 : else
3905 19702 : smoothed_alloc += ((float) recent_alloc - smoothed_alloc) /
3906 : smoothing_samples;
3907 :
3908 : /* Scale the estimate by a GUC to allow more aggressive tuning. */
3909 27114 : upcoming_alloc_est = (int) (smoothed_alloc * bgwriter_lru_multiplier);
3910 :
3911 : /*
3912 : * If recent_alloc remains at zero for many cycles, smoothed_alloc will
3913 : * eventually underflow to zero, and the underflows produce annoying
3914 : * kernel warnings on some platforms. Once upcoming_alloc_est has gone to
3915 : * zero, there's no point in tracking smaller and smaller values of
3916 : * smoothed_alloc, so just reset it to exactly zero to avoid this
3917 : * syndrome. It will pop back up as soon as recent_alloc increases.
3918 : */
3919 27114 : if (upcoming_alloc_est == 0)
3920 4556 : smoothed_alloc = 0;
3921 :
3922 : /*
3923 : * Even in cases where there's been little or no buffer allocation
3924 : * activity, we want to make a small amount of progress through the buffer
3925 : * cache so that as many reusable buffers as possible are clean after an
3926 : * idle period.
3927 : *
3928 : * (scan_whole_pool_milliseconds / BgWriterDelay) computes how many times
3929 : * the BGW will be called during the scan_whole_pool time; slice the
3930 : * buffer pool into that many sections.
3931 : */
3932 27114 : min_scan_buffers = (int) (NBuffers / (scan_whole_pool_milliseconds / BgWriterDelay));
3933 :
3934 27114 : if (upcoming_alloc_est < (min_scan_buffers + reusable_buffers_est))
3935 : {
3936 : #ifdef BGW_DEBUG
3937 : elog(DEBUG2, "bgwriter: alloc_est=%d too small, using min=%d + reusable_est=%d",
3938 : upcoming_alloc_est, min_scan_buffers, reusable_buffers_est);
3939 : #endif
3940 12088 : upcoming_alloc_est = min_scan_buffers + reusable_buffers_est;
3941 : }
3942 :
3943 : /*
3944 : * Now write out dirty reusable buffers, working forward from the
3945 : * next_to_clean point, until we have lapped the strategy scan, or cleaned
3946 : * enough buffers to match our estimate of the next cycle's allocation
3947 : * requirements, or hit the bgwriter_lru_maxpages limit.
3948 : */
3949 :
3950 27114 : num_to_scan = bufs_to_lap;
3951 27114 : num_written = 0;
3952 27114 : reusable_buffers = reusable_buffers_est;
3953 :
3954 : /* Execute the LRU scan */
3955 3655808 : while (num_to_scan > 0 && reusable_buffers < upcoming_alloc_est)
3956 : {
3957 3628696 : int sync_state = SyncOneBuffer(next_to_clean, true,
3958 : wb_context);
3959 :
3960 3628696 : if (++next_to_clean >= NBuffers)
3961 : {
3962 5874 : next_to_clean = 0;
3963 5874 : next_passes++;
3964 : }
3965 3628696 : num_to_scan--;
3966 :
3967 3628696 : if (sync_state & BUF_WRITTEN)
3968 : {
3969 49522 : reusable_buffers++;
3970 49522 : if (++num_written >= bgwriter_lru_maxpages)
3971 : {
3972 2 : PendingBgWriterStats.maxwritten_clean++;
3973 2 : break;
3974 : }
3975 : }
3976 3579174 : else if (sync_state & BUF_REUSABLE)
3977 2782018 : reusable_buffers++;
3978 : }
3979 :
3980 27114 : PendingBgWriterStats.buf_written_clean += num_written;
3981 :
3982 : #ifdef BGW_DEBUG
3983 : 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",
3984 : recent_alloc, smoothed_alloc, strategy_delta, bufs_ahead,
3985 : smoothed_density, reusable_buffers_est, upcoming_alloc_est,
3986 : bufs_to_lap - num_to_scan,
3987 : num_written,
3988 : reusable_buffers - reusable_buffers_est);
3989 : #endif
3990 :
3991 : /*
3992 : * Consider the above scan as being like a new allocation scan.
3993 : * Characterize its density and update the smoothed one based on it. This
3994 : * effectively halves the moving average period in cases where both the
3995 : * strategy and the background writer are doing some useful scanning,
3996 : * which is helpful because a long memory isn't as desirable on the
3997 : * density estimates.
3998 : */
3999 27114 : new_strategy_delta = bufs_to_lap - num_to_scan;
4000 27114 : new_recent_alloc = reusable_buffers - reusable_buffers_est;
4001 27114 : if (new_strategy_delta > 0 && new_recent_alloc > 0)
4002 : {
4003 21564 : scans_per_alloc = (float) new_strategy_delta / (float) new_recent_alloc;
4004 21564 : smoothed_density += (scans_per_alloc - smoothed_density) /
4005 : smoothing_samples;
4006 :
4007 : #ifdef BGW_DEBUG
4008 : elog(DEBUG2, "bgwriter: cleaner density alloc=%u scan=%ld density=%.2f new smoothed=%.2f",
4009 : new_recent_alloc, new_strategy_delta,
4010 : scans_per_alloc, smoothed_density);
4011 : #endif
4012 : }
4013 :
4014 : /* Return true if OK to hibernate */
4015 27114 : return (bufs_to_lap == 0 && recent_alloc == 0);
4016 : }
4017 :
4018 : /*
4019 : * SyncOneBuffer -- process a single buffer during syncing.
4020 : *
4021 : * If skip_recently_used is true, we don't write currently-pinned buffers, nor
4022 : * buffers marked recently used, as these are not replacement candidates.
4023 : *
4024 : * Returns a bitmask containing the following flag bits:
4025 : * BUF_WRITTEN: we wrote the buffer.
4026 : * BUF_REUSABLE: buffer is available for replacement, ie, it has
4027 : * pin count 0 and usage count 0.
4028 : *
4029 : * (BUF_WRITTEN could be set in error if FlushBuffer finds the buffer clean
4030 : * after locking it, but we don't care all that much.)
4031 : */
4032 : static int
4033 4184566 : SyncOneBuffer(int buf_id, bool skip_recently_used, WritebackContext *wb_context)
4034 : {
4035 4184566 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
4036 4184566 : int result = 0;
4037 : uint64 buf_state;
4038 : BufferTag tag;
4039 :
4040 : /* Make sure we can handle the pin */
4041 4184566 : ReservePrivateRefCountEntry();
4042 4184566 : ResourceOwnerEnlarge(CurrentResourceOwner);
4043 :
4044 : /*
4045 : * Check whether buffer needs writing.
4046 : *
4047 : * We can make this check without taking the buffer content lock so long
4048 : * as we mark pages dirty in access methods *before* logging changes with
4049 : * XLogInsert(): if someone marks the buffer dirty just after our check we
4050 : * don't worry because our checkpoint.redo points before log record for
4051 : * upcoming changes and so we are not required to write such dirty buffer.
4052 : */
4053 4184566 : buf_state = LockBufHdr(bufHdr);
4054 :
4055 4184566 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 0 &&
4056 4174632 : BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
4057 : {
4058 2835200 : result |= BUF_REUSABLE;
4059 : }
4060 1349366 : else if (skip_recently_used)
4061 : {
4062 : /* Caller told us not to write recently-used buffers */
4063 797156 : UnlockBufHdr(bufHdr);
4064 797156 : return result;
4065 : }
4066 :
4067 3387410 : if (!(buf_state & BM_VALID) || !(buf_state & BM_DIRTY))
4068 : {
4069 : /* It's clean, so nothing to do */
4070 2782018 : UnlockBufHdr(bufHdr);
4071 2782018 : return result;
4072 : }
4073 :
4074 : /*
4075 : * Pin it, share-lock it, write it. (FlushBuffer will do nothing if the
4076 : * buffer is clean by the time we've locked it.)
4077 : */
4078 605392 : PinBuffer_Locked(bufHdr);
4079 :
4080 605392 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
4081 :
4082 605392 : tag = bufHdr->tag;
4083 :
4084 605392 : UnpinBuffer(bufHdr);
4085 :
4086 : /*
4087 : * SyncOneBuffer() is only called by checkpointer and bgwriter, so
4088 : * IOContext will always be IOCONTEXT_NORMAL.
4089 : */
4090 605392 : ScheduleBufferTagForWriteback(wb_context, IOCONTEXT_NORMAL, &tag);
4091 :
4092 605392 : return result | BUF_WRITTEN;
4093 : }
4094 :
4095 : /*
4096 : * AtEOXact_Buffers - clean up at end of transaction.
4097 : *
4098 : * As of PostgreSQL 8.0, buffer pins should get released by the
4099 : * ResourceOwner mechanism. This routine is just a debugging
4100 : * cross-check that no pins remain.
4101 : */
4102 : void
4103 1009290 : AtEOXact_Buffers(bool isCommit)
4104 : {
4105 1009290 : CheckForBufferLeaks();
4106 :
4107 1009290 : AtEOXact_LocalBuffers(isCommit);
4108 :
4109 : Assert(PrivateRefCountOverflowed == 0);
4110 1009290 : }
4111 :
4112 : /*
4113 : * Initialize access to shared buffer pool
4114 : *
4115 : * This is called during backend startup (whether standalone or under the
4116 : * postmaster). It sets up for this backend's access to the already-existing
4117 : * buffer pool.
4118 : */
4119 : void
4120 45448 : InitBufferManagerAccess(void)
4121 : {
4122 : HASHCTL hash_ctl;
4123 :
4124 : /*
4125 : * An advisory limit on the number of pins each backend should hold, based
4126 : * on shared_buffers and the maximum number of connections possible.
4127 : * That's very pessimistic, but outside toy-sized shared_buffers it should
4128 : * allow plenty of pins. LimitAdditionalPins() and
4129 : * GetAdditionalPinLimit() can be used to check the remaining balance.
4130 : */
4131 45448 : MaxProportionalPins = NBuffers / (MaxBackends + NUM_AUXILIARY_PROCS);
4132 :
4133 45448 : memset(&PrivateRefCountArray, 0, sizeof(PrivateRefCountArray));
4134 45448 : memset(&PrivateRefCountArrayKeys, 0, sizeof(PrivateRefCountArrayKeys));
4135 :
4136 45448 : hash_ctl.keysize = sizeof(Buffer);
4137 45448 : hash_ctl.entrysize = sizeof(PrivateRefCountEntry);
4138 :
4139 45448 : PrivateRefCountHash = hash_create("PrivateRefCount", 100, &hash_ctl,
4140 : HASH_ELEM | HASH_BLOBS);
4141 :
4142 : /*
4143 : * AtProcExit_Buffers needs LWLock access, and thereby has to be called at
4144 : * the corresponding phase of backend shutdown.
4145 : */
4146 : Assert(MyProc != NULL);
4147 45448 : on_shmem_exit(AtProcExit_Buffers, 0);
4148 45448 : }
4149 :
4150 : /*
4151 : * During backend exit, ensure that we released all shared-buffer locks and
4152 : * assert that we have no remaining pins.
4153 : */
4154 : static void
4155 45448 : AtProcExit_Buffers(int code, Datum arg)
4156 : {
4157 45448 : UnlockBuffers();
4158 :
4159 45448 : CheckForBufferLeaks();
4160 :
4161 : /* localbuf.c needs a chance too */
4162 45448 : AtProcExit_LocalBuffers();
4163 45448 : }
4164 :
4165 : /*
4166 : * CheckForBufferLeaks - ensure this backend holds no buffer pins
4167 : *
4168 : * As of PostgreSQL 8.0, buffer pins should get released by the
4169 : * ResourceOwner mechanism. This routine is just a debugging
4170 : * cross-check that no pins remain.
4171 : */
4172 : static void
4173 1054738 : CheckForBufferLeaks(void)
4174 : {
4175 : #ifdef USE_ASSERT_CHECKING
4176 : int RefCountErrors = 0;
4177 : PrivateRefCountEntry *res;
4178 : int i;
4179 : char *s;
4180 :
4181 : /* check the array */
4182 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4183 : {
4184 : if (PrivateRefCountArrayKeys[i] != InvalidBuffer)
4185 : {
4186 : res = &PrivateRefCountArray[i];
4187 :
4188 : s = DebugPrintBufferRefcount(res->buffer);
4189 : elog(WARNING, "buffer refcount leak: %s", s);
4190 : pfree(s);
4191 :
4192 : RefCountErrors++;
4193 : }
4194 : }
4195 :
4196 : /* if necessary search the hash */
4197 : if (PrivateRefCountOverflowed)
4198 : {
4199 : HASH_SEQ_STATUS hstat;
4200 :
4201 : hash_seq_init(&hstat, PrivateRefCountHash);
4202 : while ((res = (PrivateRefCountEntry *) hash_seq_search(&hstat)) != NULL)
4203 : {
4204 : s = DebugPrintBufferRefcount(res->buffer);
4205 : elog(WARNING, "buffer refcount leak: %s", s);
4206 : pfree(s);
4207 : RefCountErrors++;
4208 : }
4209 : }
4210 :
4211 : Assert(RefCountErrors == 0);
4212 : #endif
4213 1054738 : }
4214 :
4215 : #ifdef USE_ASSERT_CHECKING
4216 : /*
4217 : * Check for exclusive-locked catalog buffers. This is the core of
4218 : * AssertCouldGetRelation().
4219 : *
4220 : * A backend would self-deadlock on the content lock if the catalog scan read
4221 : * the exclusive-locked buffer. The main threat is exclusive-locked buffers
4222 : * of catalogs used in relcache, because a catcache search on any catalog may
4223 : * build that catalog's relcache entry. We don't have an inventory of
4224 : * catalogs relcache uses, so just check buffers of most catalogs.
4225 : *
4226 : * It's better to minimize waits while holding an exclusive buffer lock, so it
4227 : * would be nice to broaden this check not to be catalog-specific. However,
4228 : * bttextcmp() accesses pg_collation, and non-core opclasses might similarly
4229 : * read tables. That is deadlock-free as long as there's no loop in the
4230 : * dependency graph: modifying table A may cause an opclass to read table B,
4231 : * but it must not cause a read of table A.
4232 : */
4233 : void
4234 : AssertBufferLocksPermitCatalogRead(void)
4235 : {
4236 : PrivateRefCountEntry *res;
4237 :
4238 : /* check the array */
4239 : for (int i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4240 : {
4241 : if (PrivateRefCountArrayKeys[i] != InvalidBuffer)
4242 : {
4243 : res = &PrivateRefCountArray[i];
4244 :
4245 : if (res->buffer == InvalidBuffer)
4246 : continue;
4247 :
4248 : AssertNotCatalogBufferLock(res->buffer, res->data.lockmode);
4249 : }
4250 : }
4251 :
4252 : /* if necessary search the hash */
4253 : if (PrivateRefCountOverflowed)
4254 : {
4255 : HASH_SEQ_STATUS hstat;
4256 :
4257 : hash_seq_init(&hstat, PrivateRefCountHash);
4258 : while ((res = (PrivateRefCountEntry *) hash_seq_search(&hstat)) != NULL)
4259 : {
4260 : AssertNotCatalogBufferLock(res->buffer, res->data.lockmode);
4261 : }
4262 : }
4263 : }
4264 :
4265 : static void
4266 : AssertNotCatalogBufferLock(Buffer buffer, BufferLockMode mode)
4267 : {
4268 : BufferDesc *bufHdr = GetBufferDescriptor(buffer - 1);
4269 : BufferTag tag;
4270 : Oid relid;
4271 :
4272 : if (mode != BUFFER_LOCK_EXCLUSIVE)
4273 : return;
4274 :
4275 : tag = bufHdr->tag;
4276 :
4277 : /*
4278 : * This relNumber==relid assumption holds until a catalog experiences
4279 : * VACUUM FULL or similar. After a command like that, relNumber will be
4280 : * in the normal (non-catalog) range, and we lose the ability to detect
4281 : * hazardous access to that catalog. Calling RelidByRelfilenumber() would
4282 : * close that gap, but RelidByRelfilenumber() might then deadlock with a
4283 : * held lock.
4284 : */
4285 : relid = tag.relNumber;
4286 :
4287 : if (IsCatalogTextUniqueIndexOid(relid)) /* see comments at the callee */
4288 : return;
4289 :
4290 : Assert(!IsCatalogRelationOid(relid));
4291 : }
4292 : #endif
4293 :
4294 :
4295 : /*
4296 : * Helper routine to issue warnings when a buffer is unexpectedly pinned
4297 : */
4298 : char *
4299 80 : DebugPrintBufferRefcount(Buffer buffer)
4300 : {
4301 : BufferDesc *buf;
4302 : int32 loccount;
4303 : char *result;
4304 : ProcNumber backend;
4305 : uint64 buf_state;
4306 :
4307 : Assert(BufferIsValid(buffer));
4308 80 : if (BufferIsLocal(buffer))
4309 : {
4310 32 : buf = GetLocalBufferDescriptor(-buffer - 1);
4311 32 : loccount = LocalRefCount[-buffer - 1];
4312 32 : backend = MyProcNumber;
4313 : }
4314 : else
4315 : {
4316 48 : buf = GetBufferDescriptor(buffer - 1);
4317 48 : loccount = GetPrivateRefCount(buffer);
4318 48 : backend = INVALID_PROC_NUMBER;
4319 : }
4320 :
4321 : /* theoretically we should lock the bufHdr here */
4322 80 : buf_state = pg_atomic_read_u64(&buf->state);
4323 :
4324 80 : result = psprintf("[%03d] (rel=%s, blockNum=%u, flags=0x%" PRIx64 ", refcount=%u %d)",
4325 : buffer,
4326 80 : relpathbackend(BufTagGetRelFileLocator(&buf->tag), backend,
4327 : BufTagGetForkNum(&buf->tag)).str,
4328 : buf->tag.blockNum, buf_state & BUF_FLAG_MASK,
4329 : BUF_STATE_GET_REFCOUNT(buf_state), loccount);
4330 80 : return result;
4331 : }
4332 :
4333 : /*
4334 : * CheckPointBuffers
4335 : *
4336 : * Flush all dirty blocks in buffer pool to disk at checkpoint time.
4337 : *
4338 : * Note: temporary relations do not participate in checkpoints, so they don't
4339 : * need to be flushed.
4340 : */
4341 : void
4342 3560 : CheckPointBuffers(int flags)
4343 : {
4344 3560 : BufferSync(flags);
4345 3560 : }
4346 :
4347 : /*
4348 : * BufferGetBlockNumber
4349 : * Returns the block number associated with a buffer.
4350 : *
4351 : * Note:
4352 : * Assumes that the buffer is valid and pinned, else the
4353 : * value may be obsolete immediately...
4354 : */
4355 : BlockNumber
4356 100432388 : BufferGetBlockNumber(Buffer buffer)
4357 : {
4358 : BufferDesc *bufHdr;
4359 :
4360 : Assert(BufferIsPinned(buffer));
4361 :
4362 100432388 : if (BufferIsLocal(buffer))
4363 3810458 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4364 : else
4365 96621930 : bufHdr = GetBufferDescriptor(buffer - 1);
4366 :
4367 : /* pinned, so OK to read tag without spinlock */
4368 100432388 : return bufHdr->tag.blockNum;
4369 : }
4370 :
4371 : /*
4372 : * BufferGetTag
4373 : * Returns the relfilelocator, fork number and block number associated with
4374 : * a buffer.
4375 : */
4376 : void
4377 31742324 : BufferGetTag(Buffer buffer, RelFileLocator *rlocator, ForkNumber *forknum,
4378 : BlockNumber *blknum)
4379 : {
4380 : BufferDesc *bufHdr;
4381 :
4382 : /* Do the same checks as BufferGetBlockNumber. */
4383 : Assert(BufferIsPinned(buffer));
4384 :
4385 31742324 : if (BufferIsLocal(buffer))
4386 0 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4387 : else
4388 31742324 : bufHdr = GetBufferDescriptor(buffer - 1);
4389 :
4390 : /* pinned, so OK to read tag without spinlock */
4391 31742324 : *rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
4392 31742324 : *forknum = BufTagGetForkNum(&bufHdr->tag);
4393 31742324 : *blknum = bufHdr->tag.blockNum;
4394 31742324 : }
4395 :
4396 : /*
4397 : * FlushBuffer
4398 : * Physically write out a shared buffer.
4399 : *
4400 : * NOTE: this actually just passes the buffer contents to the kernel; the
4401 : * real write to disk won't happen until the kernel feels like it. This
4402 : * is okay from our point of view since we can redo the changes from WAL.
4403 : * However, we will need to force the changes to disk via fsync before
4404 : * we can checkpoint WAL.
4405 : *
4406 : * The caller must hold a pin on the buffer and have share-locked the
4407 : * buffer contents. (Note: a share-lock does not prevent updates of
4408 : * hint bits in the buffer, so the page could change while the write
4409 : * is in progress, but we assume that that will not invalidate the data
4410 : * written.)
4411 : *
4412 : * If the caller has an smgr reference for the buffer's relation, pass it
4413 : * as the second parameter. If not, pass NULL.
4414 : */
4415 : static void
4416 1145482 : FlushBuffer(BufferDesc *buf, SMgrRelation reln, IOObject io_object,
4417 : IOContext io_context)
4418 : {
4419 : XLogRecPtr recptr;
4420 : ErrorContextCallback errcallback;
4421 : instr_time io_start;
4422 : Block bufBlock;
4423 : char *bufToWrite;
4424 : uint64 buf_state;
4425 :
4426 : /*
4427 : * Try to start an I/O operation. If StartBufferIO returns false, then
4428 : * someone else flushed the buffer before we could, so we need not do
4429 : * anything.
4430 : */
4431 1145482 : if (!StartBufferIO(buf, false, false))
4432 30 : return;
4433 :
4434 : /* Setup error traceback support for ereport() */
4435 1145452 : errcallback.callback = shared_buffer_write_error_callback;
4436 1145452 : errcallback.arg = buf;
4437 1145452 : errcallback.previous = error_context_stack;
4438 1145452 : error_context_stack = &errcallback;
4439 :
4440 : /* Find smgr relation for buffer */
4441 1145452 : if (reln == NULL)
4442 1140652 : reln = smgropen(BufTagGetRelFileLocator(&buf->tag), INVALID_PROC_NUMBER);
4443 :
4444 : TRACE_POSTGRESQL_BUFFER_FLUSH_START(BufTagGetForkNum(&buf->tag),
4445 : buf->tag.blockNum,
4446 : reln->smgr_rlocator.locator.spcOid,
4447 : reln->smgr_rlocator.locator.dbOid,
4448 : reln->smgr_rlocator.locator.relNumber);
4449 :
4450 1145452 : buf_state = LockBufHdr(buf);
4451 :
4452 : /*
4453 : * Run PageGetLSN while holding header lock, since we don't have the
4454 : * buffer locked exclusively in all cases.
4455 : */
4456 1145452 : recptr = BufferGetLSN(buf);
4457 :
4458 : /* To check if block content changes while flushing. - vadim 01/17/97 */
4459 1145452 : UnlockBufHdrExt(buf, buf_state,
4460 : 0, BM_JUST_DIRTIED,
4461 : 0);
4462 :
4463 : /*
4464 : * Force XLOG flush up to buffer's LSN. This implements the basic WAL
4465 : * rule that log updates must hit disk before any of the data-file changes
4466 : * they describe do.
4467 : *
4468 : * However, this rule does not apply to unlogged relations, which will be
4469 : * lost after a crash anyway. Most unlogged relation pages do not bear
4470 : * LSNs since we never emit WAL records for them, and therefore flushing
4471 : * up through the buffer LSN would be useless, but harmless. However,
4472 : * GiST indexes use LSNs internally to track page-splits, and therefore
4473 : * unlogged GiST pages bear "fake" LSNs generated by
4474 : * GetFakeLSNForUnloggedRel. It is unlikely but possible that the fake
4475 : * LSN counter could advance past the WAL insertion point; and if it did
4476 : * happen, attempting to flush WAL through that location would fail, with
4477 : * disastrous system-wide consequences. To make sure that can't happen,
4478 : * skip the flush if the buffer isn't permanent.
4479 : */
4480 1145452 : if (buf_state & BM_PERMANENT)
4481 1141854 : XLogFlush(recptr);
4482 :
4483 : /*
4484 : * Now it's safe to write the buffer to disk. Note that no one else should
4485 : * have been able to write it, while we were busy with log flushing,
4486 : * because we got the exclusive right to perform I/O by setting the
4487 : * BM_IO_IN_PROGRESS bit.
4488 : */
4489 1145452 : bufBlock = BufHdrGetBlock(buf);
4490 :
4491 : /*
4492 : * Update page checksum if desired. Since we have only shared lock on the
4493 : * buffer, other processes might be updating hint bits in it, so we must
4494 : * copy the page to private storage if we do checksumming.
4495 : */
4496 1145452 : bufToWrite = PageSetChecksumCopy((Page) bufBlock, buf->tag.blockNum);
4497 :
4498 1145452 : io_start = pgstat_prepare_io_time(track_io_timing);
4499 :
4500 : /*
4501 : * bufToWrite is either the shared buffer or a copy, as appropriate.
4502 : */
4503 1145452 : smgrwrite(reln,
4504 1145452 : BufTagGetForkNum(&buf->tag),
4505 : buf->tag.blockNum,
4506 : bufToWrite,
4507 : false);
4508 :
4509 : /*
4510 : * When a strategy is in use, only flushes of dirty buffers already in the
4511 : * strategy ring are counted as strategy writes (IOCONTEXT
4512 : * [BULKREAD|BULKWRITE|VACUUM] IOOP_WRITE) for the purpose of IO
4513 : * statistics tracking.
4514 : *
4515 : * If a shared buffer initially added to the ring must be flushed before
4516 : * being used, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE.
4517 : *
4518 : * If a shared buffer which was added to the ring later because the
4519 : * current strategy buffer is pinned or in use or because all strategy
4520 : * buffers were dirty and rejected (for BAS_BULKREAD operations only)
4521 : * requires flushing, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE
4522 : * (from_ring will be false).
4523 : *
4524 : * When a strategy is not in use, the write can only be a "regular" write
4525 : * of a dirty shared buffer (IOCONTEXT_NORMAL IOOP_WRITE).
4526 : */
4527 1145452 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
4528 : IOOP_WRITE, io_start, 1, BLCKSZ);
4529 :
4530 1145452 : pgBufferUsage.shared_blks_written++;
4531 :
4532 : /*
4533 : * Mark the buffer as clean (unless BM_JUST_DIRTIED has become set) and
4534 : * end the BM_IO_IN_PROGRESS state.
4535 : */
4536 1145452 : TerminateBufferIO(buf, true, 0, true, false);
4537 :
4538 : TRACE_POSTGRESQL_BUFFER_FLUSH_DONE(BufTagGetForkNum(&buf->tag),
4539 : buf->tag.blockNum,
4540 : reln->smgr_rlocator.locator.spcOid,
4541 : reln->smgr_rlocator.locator.dbOid,
4542 : reln->smgr_rlocator.locator.relNumber);
4543 :
4544 : /* Pop the error context stack */
4545 1145452 : error_context_stack = errcallback.previous;
4546 : }
4547 :
4548 : /*
4549 : * Convenience wrapper around FlushBuffer() that locks/unlocks the buffer
4550 : * before/after calling FlushBuffer().
4551 : */
4552 : static void
4553 612172 : FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
4554 : IOObject io_object, IOContext io_context)
4555 : {
4556 612172 : Buffer buffer = BufferDescriptorGetBuffer(buf);
4557 :
4558 612172 : BufferLockAcquire(buffer, buf, BUFFER_LOCK_SHARE);
4559 612172 : FlushBuffer(buf, reln, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
4560 612172 : BufferLockUnlock(buffer, buf);
4561 612172 : }
4562 :
4563 : /*
4564 : * RelationGetNumberOfBlocksInFork
4565 : * Determines the current number of pages in the specified relation fork.
4566 : *
4567 : * Note that the accuracy of the result will depend on the details of the
4568 : * relation's storage. For builtin AMs it'll be accurate, but for external AMs
4569 : * it might not be.
4570 : */
4571 : BlockNumber
4572 3787196 : RelationGetNumberOfBlocksInFork(Relation relation, ForkNumber forkNum)
4573 : {
4574 3787196 : if (RELKIND_HAS_TABLE_AM(relation->rd_rel->relkind))
4575 : {
4576 : /*
4577 : * Not every table AM uses BLCKSZ wide fixed size blocks. Therefore
4578 : * tableam returns the size in bytes - but for the purpose of this
4579 : * routine, we want the number of blocks. Therefore divide, rounding
4580 : * up.
4581 : */
4582 : uint64 szbytes;
4583 :
4584 2859128 : szbytes = table_relation_size(relation, forkNum);
4585 :
4586 2859090 : return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
4587 : }
4588 928068 : else if (RELKIND_HAS_STORAGE(relation->rd_rel->relkind))
4589 : {
4590 928068 : return smgrnblocks(RelationGetSmgr(relation), forkNum);
4591 : }
4592 : else
4593 : Assert(false);
4594 :
4595 0 : return 0; /* keep compiler quiet */
4596 : }
4597 :
4598 : /*
4599 : * BufferIsPermanent
4600 : * Determines whether a buffer will potentially still be around after
4601 : * a crash. Caller must hold a buffer pin.
4602 : */
4603 : bool
4604 19351374 : BufferIsPermanent(Buffer buffer)
4605 : {
4606 : BufferDesc *bufHdr;
4607 :
4608 : /* Local buffers are used only for temp relations. */
4609 19351374 : if (BufferIsLocal(buffer))
4610 1253980 : return false;
4611 :
4612 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4613 : Assert(BufferIsValid(buffer));
4614 : Assert(BufferIsPinned(buffer));
4615 :
4616 : /*
4617 : * BM_PERMANENT can't be changed while we hold a pin on the buffer, so we
4618 : * need not bother with the buffer header spinlock. Even if someone else
4619 : * changes the buffer header state while we're doing this, the state is
4620 : * changed atomically, so we'll read the old value or the new value, but
4621 : * not random garbage.
4622 : */
4623 18097394 : bufHdr = GetBufferDescriptor(buffer - 1);
4624 18097394 : return (pg_atomic_read_u64(&bufHdr->state) & BM_PERMANENT) != 0;
4625 : }
4626 :
4627 : /*
4628 : * BufferGetLSNAtomic
4629 : * Retrieves the LSN of the buffer atomically using a buffer header lock.
4630 : * This is necessary for some callers who may not have an exclusive lock
4631 : * on the buffer.
4632 : */
4633 : XLogRecPtr
4634 14140502 : BufferGetLSNAtomic(Buffer buffer)
4635 : {
4636 14140502 : char *page = BufferGetPage(buffer);
4637 : BufferDesc *bufHdr;
4638 : XLogRecPtr lsn;
4639 :
4640 : /*
4641 : * If we don't need locking for correctness, fastpath out.
4642 : */
4643 14140502 : if (!XLogHintBitIsNeeded() || BufferIsLocal(buffer))
4644 478336 : return PageGetLSN(page);
4645 :
4646 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4647 : Assert(BufferIsValid(buffer));
4648 : Assert(BufferIsPinned(buffer));
4649 :
4650 13662166 : bufHdr = GetBufferDescriptor(buffer - 1);
4651 13662166 : LockBufHdr(bufHdr);
4652 13662166 : lsn = PageGetLSN(page);
4653 13662166 : UnlockBufHdr(bufHdr);
4654 :
4655 13662166 : return lsn;
4656 : }
4657 :
4658 : /* ---------------------------------------------------------------------
4659 : * DropRelationBuffers
4660 : *
4661 : * This function removes from the buffer pool all the pages of the
4662 : * specified relation forks that have block numbers >= firstDelBlock.
4663 : * (In particular, with firstDelBlock = 0, all pages are removed.)
4664 : * Dirty pages are simply dropped, without bothering to write them
4665 : * out first. Therefore, this is NOT rollback-able, and so should be
4666 : * used only with extreme caution!
4667 : *
4668 : * Currently, this is called only from smgr.c when the underlying file
4669 : * is about to be deleted or truncated (firstDelBlock is needed for
4670 : * the truncation case). The data in the affected pages would therefore
4671 : * be deleted momentarily anyway, and there is no point in writing it.
4672 : * It is the responsibility of higher-level code to ensure that the
4673 : * deletion or truncation does not lose any data that could be needed
4674 : * later. It is also the responsibility of higher-level code to ensure
4675 : * that no other process could be trying to load more pages of the
4676 : * relation into buffers.
4677 : * --------------------------------------------------------------------
4678 : */
4679 : void
4680 1236 : DropRelationBuffers(SMgrRelation smgr_reln, ForkNumber *forkNum,
4681 : int nforks, BlockNumber *firstDelBlock)
4682 : {
4683 : int i;
4684 : int j;
4685 : RelFileLocatorBackend rlocator;
4686 : BlockNumber nForkBlock[MAX_FORKNUM];
4687 1236 : uint64 nBlocksToInvalidate = 0;
4688 :
4689 1236 : rlocator = smgr_reln->smgr_rlocator;
4690 :
4691 : /* If it's a local relation, it's localbuf.c's problem. */
4692 1236 : if (RelFileLocatorBackendIsTemp(rlocator))
4693 : {
4694 748 : if (rlocator.backend == MyProcNumber)
4695 748 : DropRelationLocalBuffers(rlocator.locator, forkNum, nforks,
4696 : firstDelBlock);
4697 :
4698 826 : return;
4699 : }
4700 :
4701 : /*
4702 : * To remove all the pages of the specified relation forks from the buffer
4703 : * pool, we need to scan the entire buffer pool but we can optimize it by
4704 : * finding the buffers from BufMapping table provided we know the exact
4705 : * size of each fork of the relation. The exact size is required to ensure
4706 : * that we don't leave any buffer for the relation being dropped as
4707 : * otherwise the background writer or checkpointer can lead to a PANIC
4708 : * error while flushing buffers corresponding to files that don't exist.
4709 : *
4710 : * To know the exact size, we rely on the size cached for each fork by us
4711 : * during recovery which limits the optimization to recovery and on
4712 : * standbys but we can easily extend it once we have shared cache for
4713 : * relation size.
4714 : *
4715 : * In recovery, we cache the value returned by the first lseek(SEEK_END)
4716 : * and the future writes keeps the cached value up-to-date. See
4717 : * smgrextend. It is possible that the value of the first lseek is smaller
4718 : * than the actual number of existing blocks in the file due to buggy
4719 : * Linux kernels that might not have accounted for the recent write. But
4720 : * that should be fine because there must not be any buffers after that
4721 : * file size.
4722 : */
4723 672 : for (i = 0; i < nforks; i++)
4724 : {
4725 : /* Get the number of blocks for a relation's fork */
4726 572 : nForkBlock[i] = smgrnblocks_cached(smgr_reln, forkNum[i]);
4727 :
4728 572 : if (nForkBlock[i] == InvalidBlockNumber)
4729 : {
4730 388 : nBlocksToInvalidate = InvalidBlockNumber;
4731 388 : break;
4732 : }
4733 :
4734 : /* calculate the number of blocks to be invalidated */
4735 184 : nBlocksToInvalidate += (nForkBlock[i] - firstDelBlock[i]);
4736 : }
4737 :
4738 : /*
4739 : * We apply the optimization iff the total number of blocks to invalidate
4740 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4741 : */
4742 488 : if (BlockNumberIsValid(nBlocksToInvalidate) &&
4743 100 : nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4744 : {
4745 210 : for (j = 0; j < nforks; j++)
4746 132 : FindAndDropRelationBuffers(rlocator.locator, forkNum[j],
4747 132 : nForkBlock[j], firstDelBlock[j]);
4748 78 : return;
4749 : }
4750 :
4751 5319834 : for (i = 0; i < NBuffers; i++)
4752 : {
4753 5319424 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4754 :
4755 : /*
4756 : * We can make this a tad faster by prechecking the buffer tag before
4757 : * we attempt to lock the buffer; this saves a lot of lock
4758 : * acquisitions in typical cases. It should be safe because the
4759 : * caller must have AccessExclusiveLock on the relation, or some other
4760 : * reason to be certain that no one is loading new pages of the rel
4761 : * into the buffer pool. (Otherwise we might well miss such pages
4762 : * entirely.) Therefore, while the tag might be changing while we
4763 : * look at it, it can't be changing *to* a value we care about, only
4764 : * *away* from such a value. So false negatives are impossible, and
4765 : * false positives are safe because we'll recheck after getting the
4766 : * buffer lock.
4767 : *
4768 : * We could check forkNum and blockNum as well as the rlocator, but
4769 : * the incremental win from doing so seems small.
4770 : */
4771 5319424 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator))
4772 5306030 : continue;
4773 :
4774 13394 : LockBufHdr(bufHdr);
4775 :
4776 34404 : for (j = 0; j < nforks; j++)
4777 : {
4778 24164 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator) &&
4779 24164 : BufTagGetForkNum(&bufHdr->tag) == forkNum[j] &&
4780 13224 : bufHdr->tag.blockNum >= firstDelBlock[j])
4781 : {
4782 3154 : InvalidateBuffer(bufHdr); /* releases spinlock */
4783 3154 : break;
4784 : }
4785 : }
4786 13394 : if (j >= nforks)
4787 10240 : UnlockBufHdr(bufHdr);
4788 : }
4789 : }
4790 :
4791 : /* ---------------------------------------------------------------------
4792 : * DropRelationsAllBuffers
4793 : *
4794 : * This function removes from the buffer pool all the pages of all
4795 : * forks of the specified relations. It's equivalent to calling
4796 : * DropRelationBuffers once per fork per relation with firstDelBlock = 0.
4797 : * --------------------------------------------------------------------
4798 : */
4799 : void
4800 28648 : DropRelationsAllBuffers(SMgrRelation *smgr_reln, int nlocators)
4801 : {
4802 : int i;
4803 28648 : int n = 0;
4804 : SMgrRelation *rels;
4805 : BlockNumber (*block)[MAX_FORKNUM + 1];
4806 28648 : uint64 nBlocksToInvalidate = 0;
4807 : RelFileLocator *locators;
4808 28648 : bool cached = true;
4809 : bool use_bsearch;
4810 :
4811 28648 : if (nlocators == 0)
4812 0 : return;
4813 :
4814 28648 : rels = palloc_array(SMgrRelation, nlocators); /* non-local relations */
4815 :
4816 : /* If it's a local relation, it's localbuf.c's problem. */
4817 125132 : for (i = 0; i < nlocators; i++)
4818 : {
4819 96484 : if (RelFileLocatorBackendIsTemp(smgr_reln[i]->smgr_rlocator))
4820 : {
4821 6522 : if (smgr_reln[i]->smgr_rlocator.backend == MyProcNumber)
4822 6522 : DropRelationAllLocalBuffers(smgr_reln[i]->smgr_rlocator.locator);
4823 : }
4824 : else
4825 89962 : rels[n++] = smgr_reln[i];
4826 : }
4827 :
4828 : /*
4829 : * If there are no non-local relations, then we're done. Release the
4830 : * memory and return.
4831 : */
4832 28648 : if (n == 0)
4833 : {
4834 1716 : pfree(rels);
4835 1716 : return;
4836 : }
4837 :
4838 : /*
4839 : * This is used to remember the number of blocks for all the relations
4840 : * forks.
4841 : */
4842 : block = (BlockNumber (*)[MAX_FORKNUM + 1])
4843 26932 : palloc(sizeof(BlockNumber) * n * (MAX_FORKNUM + 1));
4844 :
4845 : /*
4846 : * We can avoid scanning the entire buffer pool if we know the exact size
4847 : * of each of the given relation forks. See DropRelationBuffers.
4848 : */
4849 56444 : for (i = 0; i < n && cached; i++)
4850 : {
4851 46480 : for (int j = 0; j <= MAX_FORKNUM; j++)
4852 : {
4853 : /* Get the number of blocks for a relation's fork. */
4854 42264 : block[i][j] = smgrnblocks_cached(rels[i], j);
4855 :
4856 : /* We need to only consider the relation forks that exists. */
4857 42264 : if (block[i][j] == InvalidBlockNumber)
4858 : {
4859 37698 : if (!smgrexists(rels[i], j))
4860 12402 : continue;
4861 25296 : cached = false;
4862 25296 : break;
4863 : }
4864 :
4865 : /* calculate the total number of blocks to be invalidated */
4866 4566 : nBlocksToInvalidate += block[i][j];
4867 : }
4868 : }
4869 :
4870 : /*
4871 : * We apply the optimization iff the total number of blocks to invalidate
4872 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4873 : */
4874 26932 : if (cached && nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4875 : {
4876 2726 : for (i = 0; i < n; i++)
4877 : {
4878 7520 : for (int j = 0; j <= MAX_FORKNUM; j++)
4879 : {
4880 : /* ignore relation forks that doesn't exist */
4881 6016 : if (!BlockNumberIsValid(block[i][j]))
4882 4494 : continue;
4883 :
4884 : /* drop all the buffers for a particular relation fork */
4885 1522 : FindAndDropRelationBuffers(rels[i]->smgr_rlocator.locator,
4886 1522 : j, block[i][j], 0);
4887 : }
4888 : }
4889 :
4890 1222 : pfree(block);
4891 1222 : pfree(rels);
4892 1222 : return;
4893 : }
4894 :
4895 25710 : pfree(block);
4896 25710 : locators = palloc_array(RelFileLocator, n); /* non-local relations */
4897 114168 : for (i = 0; i < n; i++)
4898 88458 : locators[i] = rels[i]->smgr_rlocator.locator;
4899 :
4900 : /*
4901 : * For low number of relations to drop just use a simple walk through, to
4902 : * save the bsearch overhead. The threshold to use is rather a guess than
4903 : * an exactly determined value, as it depends on many factors (CPU and RAM
4904 : * speeds, amount of shared buffers etc.).
4905 : */
4906 25710 : use_bsearch = n > RELS_BSEARCH_THRESHOLD;
4907 :
4908 : /* sort the list of rlocators if necessary */
4909 25710 : if (use_bsearch)
4910 348 : qsort(locators, n, sizeof(RelFileLocator), rlocator_comparator);
4911 :
4912 277425262 : for (i = 0; i < NBuffers; i++)
4913 : {
4914 277399552 : RelFileLocator *rlocator = NULL;
4915 277399552 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4916 :
4917 : /*
4918 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4919 : * saves some cycles.
4920 : */
4921 :
4922 277399552 : if (!use_bsearch)
4923 : {
4924 : int j;
4925 :
4926 1110223766 : for (j = 0; j < n; j++)
4927 : {
4928 836750502 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &locators[j]))
4929 : {
4930 175376 : rlocator = &locators[j];
4931 175376 : break;
4932 : }
4933 : }
4934 : }
4935 : else
4936 : {
4937 : RelFileLocator locator;
4938 :
4939 3750912 : locator = BufTagGetRelFileLocator(&bufHdr->tag);
4940 3750912 : rlocator = bsearch(&locator,
4941 : locators, n, sizeof(RelFileLocator),
4942 : rlocator_comparator);
4943 : }
4944 :
4945 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
4946 277399552 : if (rlocator == NULL)
4947 277221210 : continue;
4948 :
4949 178342 : LockBufHdr(bufHdr);
4950 178342 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, rlocator))
4951 178342 : InvalidateBuffer(bufHdr); /* releases spinlock */
4952 : else
4953 0 : UnlockBufHdr(bufHdr);
4954 : }
4955 :
4956 25710 : pfree(locators);
4957 25710 : pfree(rels);
4958 : }
4959 :
4960 : /* ---------------------------------------------------------------------
4961 : * FindAndDropRelationBuffers
4962 : *
4963 : * This function performs look up in BufMapping table and removes from the
4964 : * buffer pool all the pages of the specified relation fork that has block
4965 : * number >= firstDelBlock. (In particular, with firstDelBlock = 0, all
4966 : * pages are removed.)
4967 : * --------------------------------------------------------------------
4968 : */
4969 : static void
4970 1654 : FindAndDropRelationBuffers(RelFileLocator rlocator, ForkNumber forkNum,
4971 : BlockNumber nForkBlock,
4972 : BlockNumber firstDelBlock)
4973 : {
4974 : BlockNumber curBlock;
4975 :
4976 3982 : for (curBlock = firstDelBlock; curBlock < nForkBlock; curBlock++)
4977 : {
4978 : uint32 bufHash; /* hash value for tag */
4979 : BufferTag bufTag; /* identity of requested block */
4980 : LWLock *bufPartitionLock; /* buffer partition lock for it */
4981 : int buf_id;
4982 : BufferDesc *bufHdr;
4983 :
4984 : /* create a tag so we can lookup the buffer */
4985 2328 : InitBufferTag(&bufTag, &rlocator, forkNum, curBlock);
4986 :
4987 : /* determine its hash code and partition lock ID */
4988 2328 : bufHash = BufTableHashCode(&bufTag);
4989 2328 : bufPartitionLock = BufMappingPartitionLock(bufHash);
4990 :
4991 : /* Check that it is in the buffer pool. If not, do nothing. */
4992 2328 : LWLockAcquire(bufPartitionLock, LW_SHARED);
4993 2328 : buf_id = BufTableLookup(&bufTag, bufHash);
4994 2328 : LWLockRelease(bufPartitionLock);
4995 :
4996 2328 : if (buf_id < 0)
4997 272 : continue;
4998 :
4999 2056 : bufHdr = GetBufferDescriptor(buf_id);
5000 :
5001 : /*
5002 : * We need to lock the buffer header and recheck if the buffer is
5003 : * still associated with the same block because the buffer could be
5004 : * evicted by some other backend loading blocks for a different
5005 : * relation after we release lock on the BufMapping table.
5006 : */
5007 2056 : LockBufHdr(bufHdr);
5008 :
5009 4112 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator) &&
5010 2056 : BufTagGetForkNum(&bufHdr->tag) == forkNum &&
5011 2056 : bufHdr->tag.blockNum >= firstDelBlock)
5012 2056 : InvalidateBuffer(bufHdr); /* releases spinlock */
5013 : else
5014 0 : UnlockBufHdr(bufHdr);
5015 : }
5016 1654 : }
5017 :
5018 : /* ---------------------------------------------------------------------
5019 : * DropDatabaseBuffers
5020 : *
5021 : * This function removes all the buffers in the buffer cache for a
5022 : * particular database. Dirty pages are simply dropped, without
5023 : * bothering to write them out first. This is used when we destroy a
5024 : * database, to avoid trying to flush data to disk when the directory
5025 : * tree no longer exists. Implementation is pretty similar to
5026 : * DropRelationBuffers() which is for destroying just one relation.
5027 : * --------------------------------------------------------------------
5028 : */
5029 : void
5030 152 : DropDatabaseBuffers(Oid dbid)
5031 : {
5032 : int i;
5033 :
5034 : /*
5035 : * We needn't consider local buffers, since by assumption the target
5036 : * database isn't our own.
5037 : */
5038 :
5039 1092504 : for (i = 0; i < NBuffers; i++)
5040 : {
5041 1092352 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5042 :
5043 : /*
5044 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5045 : * saves some cycles.
5046 : */
5047 1092352 : if (bufHdr->tag.dbOid != dbid)
5048 1064434 : continue;
5049 :
5050 27918 : LockBufHdr(bufHdr);
5051 27918 : if (bufHdr->tag.dbOid == dbid)
5052 27918 : InvalidateBuffer(bufHdr); /* releases spinlock */
5053 : else
5054 0 : UnlockBufHdr(bufHdr);
5055 : }
5056 152 : }
5057 :
5058 : /* ---------------------------------------------------------------------
5059 : * FlushRelationBuffers
5060 : *
5061 : * This function writes all dirty pages of a relation out to disk
5062 : * (or more accurately, out to kernel disk buffers), ensuring that the
5063 : * kernel has an up-to-date view of the relation.
5064 : *
5065 : * Generally, the caller should be holding AccessExclusiveLock on the
5066 : * target relation to ensure that no other backend is busy dirtying
5067 : * more blocks of the relation; the effects can't be expected to last
5068 : * after the lock is released.
5069 : *
5070 : * XXX currently it sequentially searches the buffer pool, should be
5071 : * changed to more clever ways of searching. This routine is not
5072 : * used in any performance-critical code paths, so it's not worth
5073 : * adding additional overhead to normal paths to make it go faster.
5074 : * --------------------------------------------------------------------
5075 : */
5076 : void
5077 274 : FlushRelationBuffers(Relation rel)
5078 : {
5079 : int i;
5080 : BufferDesc *bufHdr;
5081 274 : SMgrRelation srel = RelationGetSmgr(rel);
5082 :
5083 274 : if (RelationUsesLocalBuffers(rel))
5084 : {
5085 1818 : for (i = 0; i < NLocBuffer; i++)
5086 : {
5087 : uint64 buf_state;
5088 :
5089 1800 : bufHdr = GetLocalBufferDescriptor(i);
5090 1800 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
5091 600 : ((buf_state = pg_atomic_read_u64(&bufHdr->state)) &
5092 : (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5093 : {
5094 : ErrorContextCallback errcallback;
5095 :
5096 : /* Setup error traceback support for ereport() */
5097 600 : errcallback.callback = local_buffer_write_error_callback;
5098 600 : errcallback.arg = bufHdr;
5099 600 : errcallback.previous = error_context_stack;
5100 600 : error_context_stack = &errcallback;
5101 :
5102 : /* Make sure we can handle the pin */
5103 600 : ReservePrivateRefCountEntry();
5104 600 : ResourceOwnerEnlarge(CurrentResourceOwner);
5105 :
5106 : /*
5107 : * Pin/unpin mostly to make valgrind work, but it also seems
5108 : * like the right thing to do.
5109 : */
5110 600 : PinLocalBuffer(bufHdr, false);
5111 :
5112 :
5113 600 : FlushLocalBuffer(bufHdr, srel);
5114 :
5115 600 : UnpinLocalBuffer(BufferDescriptorGetBuffer(bufHdr));
5116 :
5117 : /* Pop the error context stack */
5118 600 : error_context_stack = errcallback.previous;
5119 : }
5120 : }
5121 :
5122 18 : return;
5123 : }
5124 :
5125 2991616 : for (i = 0; i < NBuffers; i++)
5126 : {
5127 : uint64 buf_state;
5128 :
5129 2991360 : bufHdr = GetBufferDescriptor(i);
5130 :
5131 : /*
5132 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5133 : * saves some cycles.
5134 : */
5135 2991360 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator))
5136 2990942 : continue;
5137 :
5138 : /* Make sure we can handle the pin */
5139 418 : ReservePrivateRefCountEntry();
5140 418 : ResourceOwnerEnlarge(CurrentResourceOwner);
5141 :
5142 418 : buf_state = LockBufHdr(bufHdr);
5143 418 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
5144 418 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5145 : {
5146 338 : PinBuffer_Locked(bufHdr);
5147 338 : FlushUnlockedBuffer(bufHdr, srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5148 338 : UnpinBuffer(bufHdr);
5149 : }
5150 : else
5151 80 : UnlockBufHdr(bufHdr);
5152 : }
5153 : }
5154 :
5155 : /* ---------------------------------------------------------------------
5156 : * FlushRelationsAllBuffers
5157 : *
5158 : * This function flushes out of the buffer pool all the pages of all
5159 : * forks of the specified smgr relations. It's equivalent to calling
5160 : * FlushRelationBuffers once per relation. The relations are assumed not
5161 : * to use local buffers.
5162 : * --------------------------------------------------------------------
5163 : */
5164 : void
5165 8 : FlushRelationsAllBuffers(SMgrRelation *smgrs, int nrels)
5166 : {
5167 : int i;
5168 : SMgrSortArray *srels;
5169 : bool use_bsearch;
5170 :
5171 8 : if (nrels == 0)
5172 0 : return;
5173 :
5174 : /* fill-in array for qsort */
5175 8 : srels = palloc_array(SMgrSortArray, nrels);
5176 :
5177 16 : for (i = 0; i < nrels; i++)
5178 : {
5179 : Assert(!RelFileLocatorBackendIsTemp(smgrs[i]->smgr_rlocator));
5180 :
5181 8 : srels[i].rlocator = smgrs[i]->smgr_rlocator.locator;
5182 8 : srels[i].srel = smgrs[i];
5183 : }
5184 :
5185 : /*
5186 : * Save the bsearch overhead for low number of relations to sync. See
5187 : * DropRelationsAllBuffers for details.
5188 : */
5189 8 : use_bsearch = nrels > RELS_BSEARCH_THRESHOLD;
5190 :
5191 : /* sort the list of SMgrRelations if necessary */
5192 8 : if (use_bsearch)
5193 0 : qsort(srels, nrels, sizeof(SMgrSortArray), rlocator_comparator);
5194 :
5195 131080 : for (i = 0; i < NBuffers; i++)
5196 : {
5197 131072 : SMgrSortArray *srelent = NULL;
5198 131072 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5199 : uint64 buf_state;
5200 :
5201 : /*
5202 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5203 : * saves some cycles.
5204 : */
5205 :
5206 131072 : if (!use_bsearch)
5207 : {
5208 : int j;
5209 :
5210 257614 : for (j = 0; j < nrels; j++)
5211 : {
5212 131072 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srels[j].rlocator))
5213 : {
5214 4530 : srelent = &srels[j];
5215 4530 : break;
5216 : }
5217 : }
5218 : }
5219 : else
5220 : {
5221 : RelFileLocator rlocator;
5222 :
5223 0 : rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
5224 0 : srelent = bsearch(&rlocator,
5225 : srels, nrels, sizeof(SMgrSortArray),
5226 : rlocator_comparator);
5227 : }
5228 :
5229 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
5230 131072 : if (srelent == NULL)
5231 126542 : continue;
5232 :
5233 : /* Make sure we can handle the pin */
5234 4530 : ReservePrivateRefCountEntry();
5235 4530 : ResourceOwnerEnlarge(CurrentResourceOwner);
5236 :
5237 4530 : buf_state = LockBufHdr(bufHdr);
5238 4530 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srelent->rlocator) &&
5239 4530 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5240 : {
5241 4462 : PinBuffer_Locked(bufHdr);
5242 4462 : FlushUnlockedBuffer(bufHdr, srelent->srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5243 4462 : UnpinBuffer(bufHdr);
5244 : }
5245 : else
5246 68 : UnlockBufHdr(bufHdr);
5247 : }
5248 :
5249 8 : pfree(srels);
5250 : }
5251 :
5252 : /* ---------------------------------------------------------------------
5253 : * RelationCopyStorageUsingBuffer
5254 : *
5255 : * Copy fork's data using bufmgr. Same as RelationCopyStorage but instead
5256 : * of using smgrread and smgrextend this will copy using bufmgr APIs.
5257 : *
5258 : * Refer comments atop CreateAndCopyRelationData() for details about
5259 : * 'permanent' parameter.
5260 : * --------------------------------------------------------------------
5261 : */
5262 : static void
5263 150852 : RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
5264 : RelFileLocator dstlocator,
5265 : ForkNumber forkNum, bool permanent)
5266 : {
5267 : Buffer srcBuf;
5268 : Buffer dstBuf;
5269 : Page srcPage;
5270 : Page dstPage;
5271 : bool use_wal;
5272 : BlockNumber nblocks;
5273 : BlockNumber blkno;
5274 : PGIOAlignedBlock buf;
5275 : BufferAccessStrategy bstrategy_src;
5276 : BufferAccessStrategy bstrategy_dst;
5277 : BlockRangeReadStreamPrivate p;
5278 : ReadStream *src_stream;
5279 : SMgrRelation src_smgr;
5280 :
5281 : /*
5282 : * In general, we want to write WAL whenever wal_level > 'minimal', but we
5283 : * can skip it when copying any fork of an unlogged relation other than
5284 : * the init fork.
5285 : */
5286 150852 : use_wal = XLogIsNeeded() && (permanent || forkNum == INIT_FORKNUM);
5287 :
5288 : /* Get number of blocks in the source relation. */
5289 150852 : nblocks = smgrnblocks(smgropen(srclocator, INVALID_PROC_NUMBER),
5290 : forkNum);
5291 :
5292 : /* Nothing to copy; just return. */
5293 150852 : if (nblocks == 0)
5294 26318 : return;
5295 :
5296 : /*
5297 : * Bulk extend the destination relation of the same size as the source
5298 : * relation before starting to copy block by block.
5299 : */
5300 124534 : memset(buf.data, 0, BLCKSZ);
5301 124534 : smgrextend(smgropen(dstlocator, INVALID_PROC_NUMBER), forkNum, nblocks - 1,
5302 : buf.data, true);
5303 :
5304 : /* This is a bulk operation, so use buffer access strategies. */
5305 124534 : bstrategy_src = GetAccessStrategy(BAS_BULKREAD);
5306 124534 : bstrategy_dst = GetAccessStrategy(BAS_BULKWRITE);
5307 :
5308 : /* Initialize streaming read */
5309 124534 : p.current_blocknum = 0;
5310 124534 : p.last_exclusive = nblocks;
5311 124534 : src_smgr = smgropen(srclocator, INVALID_PROC_NUMBER);
5312 :
5313 : /*
5314 : * It is safe to use batchmode as block_range_read_stream_cb takes no
5315 : * locks.
5316 : */
5317 124534 : src_stream = read_stream_begin_smgr_relation(READ_STREAM_FULL |
5318 : READ_STREAM_USE_BATCHING,
5319 : bstrategy_src,
5320 : src_smgr,
5321 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
5322 : forkNum,
5323 : block_range_read_stream_cb,
5324 : &p,
5325 : 0);
5326 :
5327 : /* Iterate over each block of the source relation file. */
5328 601428 : for (blkno = 0; blkno < nblocks; blkno++)
5329 : {
5330 476898 : CHECK_FOR_INTERRUPTS();
5331 :
5332 : /* Read block from source relation. */
5333 476898 : srcBuf = read_stream_next_buffer(src_stream, NULL);
5334 476894 : LockBuffer(srcBuf, BUFFER_LOCK_SHARE);
5335 476894 : srcPage = BufferGetPage(srcBuf);
5336 :
5337 476894 : dstBuf = ReadBufferWithoutRelcache(dstlocator, forkNum,
5338 : BufferGetBlockNumber(srcBuf),
5339 : RBM_ZERO_AND_LOCK, bstrategy_dst,
5340 : permanent);
5341 476894 : dstPage = BufferGetPage(dstBuf);
5342 :
5343 476894 : START_CRIT_SECTION();
5344 :
5345 : /* Copy page data from the source to the destination. */
5346 476894 : memcpy(dstPage, srcPage, BLCKSZ);
5347 476894 : MarkBufferDirty(dstBuf);
5348 :
5349 : /* WAL-log the copied page. */
5350 476894 : if (use_wal)
5351 275258 : log_newpage_buffer(dstBuf, true);
5352 :
5353 476894 : END_CRIT_SECTION();
5354 :
5355 476894 : UnlockReleaseBuffer(dstBuf);
5356 476894 : UnlockReleaseBuffer(srcBuf);
5357 : }
5358 : Assert(read_stream_next_buffer(src_stream, NULL) == InvalidBuffer);
5359 124530 : read_stream_end(src_stream);
5360 :
5361 124530 : FreeAccessStrategy(bstrategy_src);
5362 124530 : FreeAccessStrategy(bstrategy_dst);
5363 : }
5364 :
5365 : /* ---------------------------------------------------------------------
5366 : * CreateAndCopyRelationData
5367 : *
5368 : * Create destination relation storage and copy all forks from the
5369 : * source relation to the destination.
5370 : *
5371 : * Pass permanent as true for permanent relations and false for
5372 : * unlogged relations. Currently this API is not supported for
5373 : * temporary relations.
5374 : * --------------------------------------------------------------------
5375 : */
5376 : void
5377 113400 : CreateAndCopyRelationData(RelFileLocator src_rlocator,
5378 : RelFileLocator dst_rlocator, bool permanent)
5379 : {
5380 : char relpersistence;
5381 : SMgrRelation src_rel;
5382 : SMgrRelation dst_rel;
5383 :
5384 : /* Set the relpersistence. */
5385 113400 : relpersistence = permanent ?
5386 : RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED;
5387 :
5388 113400 : src_rel = smgropen(src_rlocator, INVALID_PROC_NUMBER);
5389 113400 : dst_rel = smgropen(dst_rlocator, INVALID_PROC_NUMBER);
5390 :
5391 : /*
5392 : * Create and copy all forks of the relation. During create database we
5393 : * have a separate cleanup mechanism which deletes complete database
5394 : * directory. Therefore, each individual relation doesn't need to be
5395 : * registered for cleanup.
5396 : */
5397 113400 : RelationCreateStorage(dst_rlocator, relpersistence, false);
5398 :
5399 : /* copy main fork. */
5400 113400 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, MAIN_FORKNUM,
5401 : permanent);
5402 :
5403 : /* copy those extra forks that exist */
5404 113396 : for (ForkNumber forkNum = MAIN_FORKNUM + 1;
5405 453584 : forkNum <= MAX_FORKNUM; forkNum++)
5406 : {
5407 340188 : if (smgrexists(src_rel, forkNum))
5408 : {
5409 37452 : smgrcreate(dst_rel, forkNum, false);
5410 :
5411 : /*
5412 : * WAL log creation if the relation is persistent, or this is the
5413 : * init fork of an unlogged relation.
5414 : */
5415 37452 : if (permanent || forkNum == INIT_FORKNUM)
5416 37452 : log_smgrcreate(&dst_rlocator, forkNum);
5417 :
5418 : /* Copy a fork's data, block by block. */
5419 37452 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, forkNum,
5420 : permanent);
5421 : }
5422 : }
5423 113396 : }
5424 :
5425 : /* ---------------------------------------------------------------------
5426 : * FlushDatabaseBuffers
5427 : *
5428 : * This function writes all dirty pages of a database out to disk
5429 : * (or more accurately, out to kernel disk buffers), ensuring that the
5430 : * kernel has an up-to-date view of the database.
5431 : *
5432 : * Generally, the caller should be holding an appropriate lock to ensure
5433 : * no other backend is active in the target database; otherwise more
5434 : * pages could get dirtied.
5435 : *
5436 : * Note we don't worry about flushing any pages of temporary relations.
5437 : * It's assumed these wouldn't be interesting.
5438 : * --------------------------------------------------------------------
5439 : */
5440 : void
5441 10 : FlushDatabaseBuffers(Oid dbid)
5442 : {
5443 : int i;
5444 : BufferDesc *bufHdr;
5445 :
5446 1290 : for (i = 0; i < NBuffers; i++)
5447 : {
5448 : uint64 buf_state;
5449 :
5450 1280 : bufHdr = GetBufferDescriptor(i);
5451 :
5452 : /*
5453 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5454 : * saves some cycles.
5455 : */
5456 1280 : if (bufHdr->tag.dbOid != dbid)
5457 934 : continue;
5458 :
5459 : /* Make sure we can handle the pin */
5460 346 : ReservePrivateRefCountEntry();
5461 346 : ResourceOwnerEnlarge(CurrentResourceOwner);
5462 :
5463 346 : buf_state = LockBufHdr(bufHdr);
5464 346 : if (bufHdr->tag.dbOid == dbid &&
5465 346 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5466 : {
5467 34 : PinBuffer_Locked(bufHdr);
5468 34 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5469 34 : UnpinBuffer(bufHdr);
5470 : }
5471 : else
5472 312 : UnlockBufHdr(bufHdr);
5473 : }
5474 10 : }
5475 :
5476 : /*
5477 : * Flush a previously, shared or exclusively, locked and pinned buffer to the
5478 : * OS.
5479 : */
5480 : void
5481 158 : FlushOneBuffer(Buffer buffer)
5482 : {
5483 : BufferDesc *bufHdr;
5484 :
5485 : /* currently not needed, but no fundamental reason not to support */
5486 : Assert(!BufferIsLocal(buffer));
5487 :
5488 : Assert(BufferIsPinned(buffer));
5489 :
5490 158 : bufHdr = GetBufferDescriptor(buffer - 1);
5491 :
5492 : Assert(BufferIsLockedByMe(buffer));
5493 :
5494 158 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5495 158 : }
5496 :
5497 : /*
5498 : * ReleaseBuffer -- release the pin on a buffer
5499 : */
5500 : void
5501 125337180 : ReleaseBuffer(Buffer buffer)
5502 : {
5503 125337180 : if (!BufferIsValid(buffer))
5504 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5505 :
5506 125337180 : if (BufferIsLocal(buffer))
5507 3220732 : UnpinLocalBuffer(buffer);
5508 : else
5509 122116448 : UnpinBuffer(GetBufferDescriptor(buffer - 1));
5510 125337180 : }
5511 :
5512 : /*
5513 : * UnlockReleaseBuffer -- release the content lock and pin on a buffer
5514 : *
5515 : * This is just a shorthand for a common combination.
5516 : */
5517 : void
5518 37398106 : UnlockReleaseBuffer(Buffer buffer)
5519 : {
5520 37398106 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5521 37398106 : ReleaseBuffer(buffer);
5522 37398106 : }
5523 :
5524 : /*
5525 : * IncrBufferRefCount
5526 : * Increment the pin count on a buffer that we have *already* pinned
5527 : * at least once.
5528 : *
5529 : * This function cannot be used on a buffer we do not have pinned,
5530 : * because it doesn't change the shared buffer state.
5531 : */
5532 : void
5533 23572860 : IncrBufferRefCount(Buffer buffer)
5534 : {
5535 : Assert(BufferIsPinned(buffer));
5536 23572860 : ResourceOwnerEnlarge(CurrentResourceOwner);
5537 23572860 : if (BufferIsLocal(buffer))
5538 709426 : LocalRefCount[-buffer - 1]++;
5539 : else
5540 : {
5541 : PrivateRefCountEntry *ref;
5542 :
5543 22863434 : ref = GetPrivateRefCountEntry(buffer, true);
5544 : Assert(ref != NULL);
5545 22863434 : ref->data.refcount++;
5546 : }
5547 23572860 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buffer);
5548 23572860 : }
5549 :
5550 : /*
5551 : * MarkBufferDirtyHint
5552 : *
5553 : * Mark a buffer dirty for non-critical changes.
5554 : *
5555 : * This is essentially the same as MarkBufferDirty, except:
5556 : *
5557 : * 1. The caller does not write WAL; so if checksums are enabled, we may need
5558 : * to write an XLOG_FPI_FOR_HINT WAL record to protect against torn pages.
5559 : * 2. The caller might have only share-lock instead of exclusive-lock on the
5560 : * buffer's content lock.
5561 : * 3. This function does not guarantee that the buffer is always marked dirty
5562 : * (due to a race condition), so it cannot be used for important changes.
5563 : */
5564 : void
5565 20324528 : MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
5566 : {
5567 : BufferDesc *bufHdr;
5568 20324528 : Page page = BufferGetPage(buffer);
5569 :
5570 20324528 : if (!BufferIsValid(buffer))
5571 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5572 :
5573 20324528 : if (BufferIsLocal(buffer))
5574 : {
5575 1270330 : MarkLocalBufferDirty(buffer);
5576 1270330 : return;
5577 : }
5578 :
5579 19054198 : bufHdr = GetBufferDescriptor(buffer - 1);
5580 :
5581 : Assert(GetPrivateRefCount(buffer) > 0);
5582 : /* here, either share or exclusive lock is OK */
5583 : Assert(BufferIsLockedByMe(buffer));
5584 :
5585 : /*
5586 : * This routine might get called many times on the same page, if we are
5587 : * making the first scan after commit of an xact that added/deleted many
5588 : * tuples. So, be as quick as we can if the buffer is already dirty. We
5589 : * do this by not acquiring spinlock if it looks like the status bits are
5590 : * already set. Since we make this test unlocked, there's a chance we
5591 : * might fail to notice that the flags have just been cleared, and failed
5592 : * to reset them, due to memory-ordering issues. But since this function
5593 : * is only intended to be used in cases where failing to write out the
5594 : * data would be harmless anyway, it doesn't really matter.
5595 : */
5596 19054198 : if ((pg_atomic_read_u64(&bufHdr->state) & (BM_DIRTY | BM_JUST_DIRTIED)) !=
5597 : (BM_DIRTY | BM_JUST_DIRTIED))
5598 : {
5599 1847578 : XLogRecPtr lsn = InvalidXLogRecPtr;
5600 1847578 : bool dirtied = false;
5601 1847578 : bool delayChkptFlags = false;
5602 : uint64 buf_state;
5603 :
5604 : /*
5605 : * If we need to protect hint bit updates from torn writes, WAL-log a
5606 : * full page image of the page. This full page image is only necessary
5607 : * if the hint bit update is the first change to the page since the
5608 : * last checkpoint.
5609 : *
5610 : * We don't check full_page_writes here because that logic is included
5611 : * when we call XLogInsert() since the value changes dynamically.
5612 : */
5613 1847578 : if (XLogHintBitIsNeeded() &&
5614 1845400 : (pg_atomic_read_u64(&bufHdr->state) & BM_PERMANENT))
5615 : {
5616 : /*
5617 : * If we must not write WAL, due to a relfilelocator-specific
5618 : * condition or being in recovery, don't dirty the page. We can
5619 : * set the hint, just not dirty the page as a result so the hint
5620 : * is lost when we evict the page or shutdown.
5621 : *
5622 : * See src/backend/storage/page/README for longer discussion.
5623 : */
5624 1970326 : if (RecoveryInProgress() ||
5625 124990 : RelFileLocatorSkippingWAL(BufTagGetRelFileLocator(&bufHdr->tag)))
5626 1722842 : return;
5627 :
5628 : /*
5629 : * If the block is already dirty because we either made a change
5630 : * or set a hint already, then we don't need to write a full page
5631 : * image. Note that aggressive cleaning of blocks dirtied by hint
5632 : * bit setting would increase the call rate. Bulk setting of hint
5633 : * bits would reduce the call rate...
5634 : *
5635 : * We must issue the WAL record before we mark the buffer dirty.
5636 : * Otherwise we might write the page before we write the WAL. That
5637 : * causes a race condition, since a checkpoint might occur between
5638 : * writing the WAL record and marking the buffer dirty. We solve
5639 : * that with a kluge, but one that is already in use during
5640 : * transaction commit to prevent race conditions. Basically, we
5641 : * simply prevent the checkpoint WAL record from being written
5642 : * until we have marked the buffer dirty. We don't start the
5643 : * checkpoint flush until we have marked dirty, so our checkpoint
5644 : * must flush the change to disk successfully or the checkpoint
5645 : * never gets written, so crash recovery will fix.
5646 : *
5647 : * It's possible we may enter here without an xid, so it is
5648 : * essential that CreateCheckPoint waits for virtual transactions
5649 : * rather than full transactionids.
5650 : */
5651 : Assert((MyProc->delayChkptFlags & DELAY_CHKPT_START) == 0);
5652 122494 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
5653 122494 : delayChkptFlags = true;
5654 122494 : lsn = XLogSaveBufferForHint(buffer, buffer_std);
5655 : }
5656 :
5657 124736 : buf_state = LockBufHdr(bufHdr);
5658 :
5659 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5660 :
5661 124736 : if (!(buf_state & BM_DIRTY))
5662 : {
5663 124642 : dirtied = true; /* Means "will be dirtied by this action" */
5664 :
5665 : /*
5666 : * Set the page LSN if we wrote a backup block. We aren't supposed
5667 : * to set this when only holding a share lock but as long as we
5668 : * serialise it somehow we're OK. We choose to set LSN while
5669 : * holding the buffer header lock, which causes any reader of an
5670 : * LSN who holds only a share lock to also obtain a buffer header
5671 : * lock before using PageGetLSN(), which is enforced in
5672 : * BufferGetLSNAtomic().
5673 : *
5674 : * If checksums are enabled, you might think we should reset the
5675 : * checksum here. That will happen when the page is written
5676 : * sometime later in this checkpoint cycle.
5677 : */
5678 124642 : if (XLogRecPtrIsValid(lsn))
5679 62916 : PageSetLSN(page, lsn);
5680 : }
5681 :
5682 124736 : UnlockBufHdrExt(bufHdr, buf_state,
5683 : BM_DIRTY | BM_JUST_DIRTIED,
5684 : 0, 0);
5685 :
5686 124736 : if (delayChkptFlags)
5687 122494 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
5688 :
5689 124736 : if (dirtied)
5690 : {
5691 124642 : pgBufferUsage.shared_blks_dirtied++;
5692 124642 : if (VacuumCostActive)
5693 4210 : VacuumCostBalance += VacuumCostPageDirty;
5694 : }
5695 : }
5696 : }
5697 :
5698 : /*
5699 : * Release buffer content locks for shared buffers.
5700 : *
5701 : * Used to clean up after errors.
5702 : *
5703 : * Currently, we can expect that resource owner cleanup, via
5704 : * ResOwnerReleaseBufferPin(), took care of releasing buffer content locks per
5705 : * se; the only thing we need to deal with here is clearing any PIN_COUNT
5706 : * request that was in progress.
5707 : */
5708 : void
5709 107092 : UnlockBuffers(void)
5710 : {
5711 107092 : BufferDesc *buf = PinCountWaitBuf;
5712 :
5713 107092 : if (buf)
5714 : {
5715 : uint64 buf_state;
5716 0 : uint64 unset_bits = 0;
5717 :
5718 0 : buf_state = LockBufHdr(buf);
5719 :
5720 : /*
5721 : * Don't complain if flag bit not set; it could have been reset but we
5722 : * got a cancel/die interrupt before getting the signal.
5723 : */
5724 0 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5725 0 : buf->wait_backend_pgprocno == MyProcNumber)
5726 0 : unset_bits = BM_PIN_COUNT_WAITER;
5727 :
5728 0 : UnlockBufHdrExt(buf, buf_state,
5729 : 0, unset_bits,
5730 : 0);
5731 :
5732 0 : PinCountWaitBuf = NULL;
5733 : }
5734 107092 : }
5735 :
5736 : /*
5737 : * Acquire the buffer content lock in the specified mode
5738 : *
5739 : * If the lock is not available, sleep until it is.
5740 : *
5741 : * Side effect: cancel/die interrupts are held off until lock release.
5742 : *
5743 : * This uses almost the same locking approach as lwlock.c's
5744 : * LWLockAcquire(). See documentation at the top of lwlock.c for a more
5745 : * detailed discussion.
5746 : *
5747 : * The reason that this, and most of the other BufferLock* functions, get both
5748 : * the Buffer and BufferDesc* as parameters, is that looking up one from the
5749 : * other repeatedly shows up noticeably in profiles.
5750 : *
5751 : * Callers should provide a constant for mode, for more efficient code
5752 : * generation.
5753 : */
5754 : static inline void
5755 162483468 : BufferLockAcquire(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode)
5756 : {
5757 : PrivateRefCountEntry *entry;
5758 162483468 : int extraWaits = 0;
5759 :
5760 : /*
5761 : * Get reference to the refcount entry before we hold the lock, it seems
5762 : * better to do before holding the lock.
5763 : */
5764 162483468 : entry = GetPrivateRefCountEntry(buffer, true);
5765 :
5766 : /*
5767 : * We better not already hold a lock on the buffer.
5768 : */
5769 : Assert(entry->data.lockmode == BUFFER_LOCK_UNLOCK);
5770 :
5771 : /*
5772 : * Lock out cancel/die interrupts until we exit the code section protected
5773 : * by the content lock. This ensures that interrupts will not interfere
5774 : * with manipulations of data structures in shared memory.
5775 : */
5776 162483468 : HOLD_INTERRUPTS();
5777 :
5778 : for (;;)
5779 47492 : {
5780 162530960 : uint32 wait_event = 0; /* initialized to avoid compiler warning */
5781 : bool mustwait;
5782 :
5783 : /*
5784 : * Try to grab the lock the first time, we're not in the waitqueue
5785 : * yet/anymore.
5786 : */
5787 162530960 : mustwait = BufferLockAttempt(buf_hdr, mode);
5788 :
5789 162530960 : if (likely(!mustwait))
5790 : {
5791 162481682 : break;
5792 : }
5793 :
5794 : /*
5795 : * Ok, at this point we couldn't grab the lock on the first try. We
5796 : * cannot simply queue ourselves to the end of the list and wait to be
5797 : * woken up because by now the lock could long have been released.
5798 : * Instead add us to the queue and try to grab the lock again. If we
5799 : * succeed we need to revert the queuing and be happy, otherwise we
5800 : * recheck the lock. If we still couldn't grab it, we know that the
5801 : * other locker will see our queue entries when releasing since they
5802 : * existed before we checked for the lock.
5803 : */
5804 :
5805 : /* add to the queue */
5806 49278 : BufferLockQueueSelf(buf_hdr, mode);
5807 :
5808 : /* we're now guaranteed to be woken up if necessary */
5809 49278 : mustwait = BufferLockAttempt(buf_hdr, mode);
5810 :
5811 : /* ok, grabbed the lock the second time round, need to undo queueing */
5812 49278 : if (!mustwait)
5813 : {
5814 1786 : BufferLockDequeueSelf(buf_hdr);
5815 1786 : break;
5816 : }
5817 :
5818 47492 : switch (mode)
5819 : {
5820 26980 : case BUFFER_LOCK_EXCLUSIVE:
5821 26980 : wait_event = WAIT_EVENT_BUFFER_EXCLUSIVE;
5822 26980 : break;
5823 0 : case BUFFER_LOCK_SHARE_EXCLUSIVE:
5824 0 : wait_event = WAIT_EVENT_BUFFER_SHARE_EXCLUSIVE;
5825 0 : break;
5826 20512 : case BUFFER_LOCK_SHARE:
5827 20512 : wait_event = WAIT_EVENT_BUFFER_SHARED;
5828 20512 : break;
5829 : case BUFFER_LOCK_UNLOCK:
5830 : pg_unreachable();
5831 :
5832 : }
5833 47492 : pgstat_report_wait_start(wait_event);
5834 :
5835 : /*
5836 : * Wait until awakened.
5837 : *
5838 : * It is possible that we get awakened for a reason other than being
5839 : * signaled by BufferLockWakeup(). If so, loop back and wait again.
5840 : * Once we've gotten the lock, re-increment the sema by the number of
5841 : * additional signals received.
5842 : */
5843 : for (;;)
5844 : {
5845 47492 : PGSemaphoreLock(MyProc->sem);
5846 47492 : if (MyProc->lwWaiting == LW_WS_NOT_WAITING)
5847 47492 : break;
5848 0 : extraWaits++;
5849 : }
5850 :
5851 47492 : pgstat_report_wait_end();
5852 :
5853 : /* Retrying, allow BufferLockRelease to release waiters again. */
5854 47492 : pg_atomic_fetch_and_u64(&buf_hdr->state, ~BM_LOCK_WAKE_IN_PROGRESS);
5855 : }
5856 :
5857 : /* Remember that we now hold this lock */
5858 162483468 : entry->data.lockmode = mode;
5859 :
5860 : /*
5861 : * Fix the process wait semaphore's count for any absorbed wakeups.
5862 : */
5863 162483468 : while (unlikely(extraWaits-- > 0))
5864 0 : PGSemaphoreUnlock(MyProc->sem);
5865 162483468 : }
5866 :
5867 : /*
5868 : * Release a previously acquired buffer content lock.
5869 : */
5870 : static void
5871 165565890 : BufferLockUnlock(Buffer buffer, BufferDesc *buf_hdr)
5872 : {
5873 : BufferLockMode mode;
5874 : uint64 oldstate;
5875 : uint64 sub;
5876 :
5877 165565890 : mode = BufferLockDisownInternal(buffer, buf_hdr);
5878 :
5879 : /*
5880 : * Release my hold on lock, after that it can immediately be acquired by
5881 : * others, even if we still have to wakeup other waiters.
5882 : */
5883 165565890 : sub = BufferLockReleaseSub(mode);
5884 :
5885 165565890 : oldstate = pg_atomic_sub_fetch_u64(&buf_hdr->state, sub);
5886 :
5887 165565890 : BufferLockProcessRelease(buf_hdr, mode, oldstate);
5888 :
5889 : /*
5890 : * Now okay to allow cancel/die interrupts.
5891 : */
5892 165565890 : RESUME_INTERRUPTS();
5893 165565890 : }
5894 :
5895 :
5896 : /*
5897 : * Acquire the content lock for the buffer, but only if we don't have to wait.
5898 : *
5899 : * It is allowed to try to conditionally acquire a lock on a buffer that this
5900 : * backend has already locked, but the lock acquisition will always fail, even
5901 : * if the new lock acquisition does not conflict with an already held lock
5902 : * (e.g. two share locks). This is because we currently do not have space to
5903 : * track multiple lock ownerships of the same buffer within one backend. That
5904 : * is ok for the current uses of BufferLockConditional().
5905 : */
5906 : static bool
5907 3100260 : BufferLockConditional(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode)
5908 : {
5909 3100260 : PrivateRefCountEntry *entry = GetPrivateRefCountEntry(buffer, true);
5910 : bool mustwait;
5911 :
5912 : /*
5913 : * As described above, if we're trying to lock a buffer this backend
5914 : * already has locked, return false, independent of the existing and
5915 : * desired lock level.
5916 : */
5917 3100260 : if (entry->data.lockmode != BUFFER_LOCK_UNLOCK)
5918 0 : return false;
5919 :
5920 : /*
5921 : * Lock out cancel/die interrupts until we exit the code section protected
5922 : * by the content lock. This ensures that interrupts will not interfere
5923 : * with manipulations of data structures in shared memory.
5924 : */
5925 3100260 : HOLD_INTERRUPTS();
5926 :
5927 : /* Check for the lock */
5928 3100260 : mustwait = BufferLockAttempt(buf_hdr, mode);
5929 :
5930 3100260 : if (mustwait)
5931 : {
5932 : /* Failed to get lock, so release interrupt holdoff */
5933 17838 : RESUME_INTERRUPTS();
5934 : }
5935 : else
5936 : {
5937 3082422 : entry->data.lockmode = mode;
5938 : }
5939 :
5940 3100260 : return !mustwait;
5941 : }
5942 :
5943 : /*
5944 : * Internal function that tries to atomically acquire the content lock in the
5945 : * passed in mode.
5946 : *
5947 : * This function will not block waiting for a lock to become free - that's the
5948 : * caller's job.
5949 : *
5950 : * Similar to LWLockAttemptLock().
5951 : */
5952 : static inline bool
5953 165680498 : BufferLockAttempt(BufferDesc *buf_hdr, BufferLockMode mode)
5954 : {
5955 : uint64 old_state;
5956 :
5957 : /*
5958 : * Read once outside the loop, later iterations will get the newer value
5959 : * via compare & exchange.
5960 : */
5961 165680498 : old_state = pg_atomic_read_u64(&buf_hdr->state);
5962 :
5963 : /* loop until we've determined whether we could acquire the lock or not */
5964 : while (true)
5965 51000 : {
5966 : uint64 desired_state;
5967 : bool lock_free;
5968 :
5969 165731498 : desired_state = old_state;
5970 :
5971 165731498 : if (mode == BUFFER_LOCK_EXCLUSIVE)
5972 : {
5973 50622062 : lock_free = (old_state & BM_LOCK_MASK) == 0;
5974 50622062 : if (lock_free)
5975 50548608 : desired_state += BM_LOCK_VAL_EXCLUSIVE;
5976 : }
5977 115109436 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
5978 : {
5979 0 : lock_free = (old_state & (BM_LOCK_VAL_EXCLUSIVE | BM_LOCK_VAL_SHARE_EXCLUSIVE)) == 0;
5980 0 : if (lock_free)
5981 0 : desired_state += BM_LOCK_VAL_SHARE_EXCLUSIVE;
5982 : }
5983 : else
5984 : {
5985 115109436 : lock_free = (old_state & BM_LOCK_VAL_EXCLUSIVE) == 0;
5986 115109436 : if (lock_free)
5987 115067532 : desired_state += BM_LOCK_VAL_SHARED;
5988 : }
5989 :
5990 : /*
5991 : * Attempt to swap in the state we are expecting. If we didn't see
5992 : * lock to be free, that's just the old value. If we saw it as free,
5993 : * we'll attempt to mark it acquired. The reason that we always swap
5994 : * in the value is that this doubles as a memory barrier. We could try
5995 : * to be smarter and only swap in values if we saw the lock as free,
5996 : * but benchmark haven't shown it as beneficial so far.
5997 : *
5998 : * Retry if the value changed since we last looked at it.
5999 : */
6000 165731498 : if (likely(pg_atomic_compare_exchange_u64(&buf_hdr->state,
6001 : &old_state, desired_state)))
6002 : {
6003 165680498 : if (lock_free)
6004 : {
6005 : /* Great! Got the lock. */
6006 165565890 : return false;
6007 : }
6008 : else
6009 114608 : return true; /* somebody else has the lock */
6010 : }
6011 : }
6012 :
6013 : pg_unreachable();
6014 : }
6015 :
6016 : /*
6017 : * Add ourselves to the end of the content lock's wait queue.
6018 : */
6019 : static void
6020 49278 : BufferLockQueueSelf(BufferDesc *buf_hdr, BufferLockMode mode)
6021 : {
6022 : /*
6023 : * If we don't have a PGPROC structure, there's no way to wait. This
6024 : * should never occur, since MyProc should only be null during shared
6025 : * memory initialization.
6026 : */
6027 49278 : if (MyProc == NULL)
6028 0 : elog(PANIC, "cannot wait without a PGPROC structure");
6029 :
6030 49278 : if (MyProc->lwWaiting != LW_WS_NOT_WAITING)
6031 0 : elog(PANIC, "queueing for lock while waiting on another one");
6032 :
6033 49278 : LockBufHdr(buf_hdr);
6034 :
6035 : /* setting the flag is protected by the spinlock */
6036 49278 : pg_atomic_fetch_or_u64(&buf_hdr->state, BM_LOCK_HAS_WAITERS);
6037 :
6038 : /*
6039 : * These are currently used both for lwlocks and buffer content locks,
6040 : * which is acceptable, although not pretty, because a backend can't wait
6041 : * for both types of locks at the same time.
6042 : */
6043 49278 : MyProc->lwWaiting = LW_WS_WAITING;
6044 49278 : MyProc->lwWaitMode = mode;
6045 :
6046 49278 : proclist_push_tail(&buf_hdr->lock_waiters, MyProcNumber, lwWaitLink);
6047 :
6048 : /* Can release the mutex now */
6049 49278 : UnlockBufHdr(buf_hdr);
6050 49278 : }
6051 :
6052 : /*
6053 : * Remove ourselves from the waitlist.
6054 : *
6055 : * This is used if we queued ourselves because we thought we needed to sleep
6056 : * but, after further checking, we discovered that we don't actually need to
6057 : * do so.
6058 : */
6059 : static void
6060 1786 : BufferLockDequeueSelf(BufferDesc *buf_hdr)
6061 : {
6062 : bool on_waitlist;
6063 :
6064 1786 : LockBufHdr(buf_hdr);
6065 :
6066 1786 : on_waitlist = MyProc->lwWaiting == LW_WS_WAITING;
6067 1786 : if (on_waitlist)
6068 1312 : proclist_delete(&buf_hdr->lock_waiters, MyProcNumber, lwWaitLink);
6069 :
6070 1786 : if (proclist_is_empty(&buf_hdr->lock_waiters) &&
6071 1722 : (pg_atomic_read_u64(&buf_hdr->state) & BM_LOCK_HAS_WAITERS) != 0)
6072 : {
6073 1250 : pg_atomic_fetch_and_u64(&buf_hdr->state, ~BM_LOCK_HAS_WAITERS);
6074 : }
6075 :
6076 : /* XXX: combine with fetch_and above? */
6077 1786 : UnlockBufHdr(buf_hdr);
6078 :
6079 : /* clear waiting state again, nice for debugging */
6080 1786 : if (on_waitlist)
6081 1312 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
6082 : else
6083 : {
6084 474 : int extraWaits = 0;
6085 :
6086 :
6087 : /*
6088 : * Somebody else dequeued us and has or will wake us up. Deal with the
6089 : * superfluous absorption of a wakeup.
6090 : */
6091 :
6092 : /*
6093 : * Clear BM_LOCK_WAKE_IN_PROGRESS if somebody woke us before we
6094 : * removed ourselves - they'll have set it.
6095 : */
6096 474 : pg_atomic_fetch_and_u64(&buf_hdr->state, ~BM_LOCK_WAKE_IN_PROGRESS);
6097 :
6098 : /*
6099 : * Now wait for the scheduled wakeup, otherwise our ->lwWaiting would
6100 : * get reset at some inconvenient point later. Most of the time this
6101 : * will immediately return.
6102 : */
6103 : for (;;)
6104 : {
6105 474 : PGSemaphoreLock(MyProc->sem);
6106 474 : if (MyProc->lwWaiting == LW_WS_NOT_WAITING)
6107 474 : break;
6108 0 : extraWaits++;
6109 : }
6110 :
6111 : /*
6112 : * Fix the process wait semaphore's count for any absorbed wakeups.
6113 : */
6114 474 : while (extraWaits-- > 0)
6115 0 : PGSemaphoreUnlock(MyProc->sem);
6116 : }
6117 1786 : }
6118 :
6119 : /*
6120 : * Stop treating lock as held by current backend.
6121 : *
6122 : * After calling this function it's the callers responsibility to ensure that
6123 : * the lock gets released, even in case of an error. This only is desirable if
6124 : * the lock is going to be released in a different process than the process
6125 : * that acquired it.
6126 : */
6127 : static inline void
6128 0 : BufferLockDisown(Buffer buffer, BufferDesc *buf_hdr)
6129 : {
6130 0 : BufferLockDisownInternal(buffer, buf_hdr);
6131 0 : RESUME_INTERRUPTS();
6132 0 : }
6133 :
6134 : /*
6135 : * Stop treating lock as held by current backend.
6136 : *
6137 : * This is the code that can be shared between actually releasing a lock
6138 : * (BufferLockUnlock()) and just not tracking ownership of the lock anymore
6139 : * without releasing the lock (BufferLockDisown()).
6140 : */
6141 : static inline int
6142 165565890 : BufferLockDisownInternal(Buffer buffer, BufferDesc *buf_hdr)
6143 : {
6144 : BufferLockMode mode;
6145 : PrivateRefCountEntry *ref;
6146 :
6147 165565890 : ref = GetPrivateRefCountEntry(buffer, false);
6148 165565890 : if (ref == NULL)
6149 0 : elog(ERROR, "lock %d is not held", buffer);
6150 165565890 : mode = ref->data.lockmode;
6151 165565890 : ref->data.lockmode = BUFFER_LOCK_UNLOCK;
6152 :
6153 165565890 : return mode;
6154 : }
6155 :
6156 : /*
6157 : * Wakeup all the lockers that currently have a chance to acquire the lock.
6158 : *
6159 : * wake_exclusive indicates whether exclusive lock waiters should be woken up.
6160 : */
6161 : static void
6162 45350 : BufferLockWakeup(BufferDesc *buf_hdr, bool wake_exclusive)
6163 : {
6164 45350 : bool new_wake_in_progress = false;
6165 45350 : bool wake_share_exclusive = true;
6166 : proclist_head wakeup;
6167 : proclist_mutable_iter iter;
6168 :
6169 45350 : proclist_init(&wakeup);
6170 :
6171 : /* lock wait list while collecting backends to wake up */
6172 45350 : LockBufHdr(buf_hdr);
6173 :
6174 68734 : proclist_foreach_modify(iter, &buf_hdr->lock_waiters, lwWaitLink)
6175 : {
6176 50626 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
6177 :
6178 : /*
6179 : * Already woke up a conflicting lock, so skip over this wait list
6180 : * entry.
6181 : */
6182 50626 : if (!wake_exclusive && waiter->lwWaitMode == BUFFER_LOCK_EXCLUSIVE)
6183 2660 : continue;
6184 47966 : if (!wake_share_exclusive && waiter->lwWaitMode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6185 0 : continue;
6186 :
6187 47966 : proclist_delete(&buf_hdr->lock_waiters, iter.cur, lwWaitLink);
6188 47966 : proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
6189 :
6190 : /*
6191 : * Prevent additional wakeups until retryer gets to run. Backends that
6192 : * are just waiting for the lock to become free don't retry
6193 : * automatically.
6194 : */
6195 47966 : new_wake_in_progress = true;
6196 :
6197 : /*
6198 : * Signal that the process isn't on the wait list anymore. This allows
6199 : * BufferLockDequeueSelf() to remove itself from the waitlist with a
6200 : * proclist_delete(), rather than having to check if it has been
6201 : * removed from the list.
6202 : */
6203 : Assert(waiter->lwWaiting == LW_WS_WAITING);
6204 47966 : waiter->lwWaiting = LW_WS_PENDING_WAKEUP;
6205 :
6206 : /*
6207 : * Don't wakeup further waiters after waking a conflicting waiter.
6208 : */
6209 47966 : if (waiter->lwWaitMode == BUFFER_LOCK_SHARE)
6210 : {
6211 : /*
6212 : * Share locks conflict with exclusive locks.
6213 : */
6214 20724 : wake_exclusive = false;
6215 : }
6216 27242 : else if (waiter->lwWaitMode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6217 : {
6218 : /*
6219 : * Share-exclusive locks conflict with share-exclusive and
6220 : * exclusive locks.
6221 : */
6222 0 : wake_exclusive = false;
6223 0 : wake_share_exclusive = false;
6224 : }
6225 27242 : else if (waiter->lwWaitMode == BUFFER_LOCK_EXCLUSIVE)
6226 : {
6227 : /*
6228 : * Exclusive locks conflict with all other locks, there's no point
6229 : * in waking up anybody else.
6230 : */
6231 27242 : break;
6232 : }
6233 : }
6234 :
6235 : Assert(proclist_is_empty(&wakeup) || pg_atomic_read_u64(&buf_hdr->state) & BM_LOCK_HAS_WAITERS);
6236 :
6237 : /* unset required flags, and release lock, in one fell swoop */
6238 : {
6239 : uint64 old_state;
6240 : uint64 desired_state;
6241 :
6242 45350 : old_state = pg_atomic_read_u64(&buf_hdr->state);
6243 : while (true)
6244 : {
6245 45372 : desired_state = old_state;
6246 :
6247 : /* compute desired flags */
6248 :
6249 45372 : if (new_wake_in_progress)
6250 45106 : desired_state |= BM_LOCK_WAKE_IN_PROGRESS;
6251 : else
6252 266 : desired_state &= ~BM_LOCK_WAKE_IN_PROGRESS;
6253 :
6254 45372 : if (proclist_is_empty(&buf_hdr->lock_waiters))
6255 36950 : desired_state &= ~BM_LOCK_HAS_WAITERS;
6256 :
6257 45372 : desired_state &= ~BM_LOCKED; /* release lock */
6258 :
6259 45372 : if (pg_atomic_compare_exchange_u64(&buf_hdr->state, &old_state,
6260 : desired_state))
6261 45350 : break;
6262 : }
6263 : }
6264 :
6265 : /* Awaken any waiters I removed from the queue. */
6266 93316 : proclist_foreach_modify(iter, &wakeup, lwWaitLink)
6267 : {
6268 47966 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
6269 :
6270 47966 : proclist_delete(&wakeup, iter.cur, lwWaitLink);
6271 :
6272 : /*
6273 : * Guarantee that lwWaiting being unset only becomes visible once the
6274 : * unlink from the link has completed. Otherwise the target backend
6275 : * could be woken up for other reason and enqueue for a new lock - if
6276 : * that happens before the list unlink happens, the list would end up
6277 : * being corrupted.
6278 : *
6279 : * The barrier pairs with the LockBufHdr() when enqueuing for another
6280 : * lock.
6281 : */
6282 47966 : pg_write_barrier();
6283 47966 : waiter->lwWaiting = LW_WS_NOT_WAITING;
6284 47966 : PGSemaphoreUnlock(waiter->sem);
6285 : }
6286 45350 : }
6287 :
6288 : /*
6289 : * Compute subtraction from buffer state for a release of a held lock in
6290 : * `mode`.
6291 : *
6292 : * This is separated from BufferLockUnlock() as we want to combine the lock
6293 : * release with other atomic operations when possible, leading to the lock
6294 : * release being done in multiple places, each needing to compute what to
6295 : * subtract from the lock state.
6296 : */
6297 : static inline uint64
6298 165565890 : BufferLockReleaseSub(BufferLockMode mode)
6299 : {
6300 : /*
6301 : * Turns out that a switch() leads gcc to generate sufficiently worse code
6302 : * for this to show up in profiles...
6303 : */
6304 165565890 : if (mode == BUFFER_LOCK_EXCLUSIVE)
6305 50547916 : return BM_LOCK_VAL_EXCLUSIVE;
6306 115017974 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6307 0 : return BM_LOCK_VAL_SHARE_EXCLUSIVE;
6308 : else
6309 : {
6310 : Assert(mode == BUFFER_LOCK_SHARE);
6311 115017974 : return BM_LOCK_VAL_SHARED;
6312 : }
6313 :
6314 : return 0; /* keep compiler quiet */
6315 : }
6316 :
6317 : /*
6318 : * Handle work that needs to be done after releasing a lock that was held in
6319 : * `mode`, where `lockstate` is the result of the atomic operation modifying
6320 : * the state variable.
6321 : *
6322 : * This is separated from BufferLockUnlock() as we want to combine the lock
6323 : * release with other atomic operations when possible, leading to the lock
6324 : * release being done in multiple places.
6325 : */
6326 : static void
6327 165565890 : BufferLockProcessRelease(BufferDesc *buf_hdr, BufferLockMode mode, uint64 lockstate)
6328 : {
6329 165565890 : bool check_waiters = false;
6330 165565890 : bool wake_exclusive = false;
6331 :
6332 : /* nobody else can have that kind of lock */
6333 : Assert(!(lockstate & BM_LOCK_VAL_EXCLUSIVE));
6334 :
6335 : /*
6336 : * If we're still waiting for backends to get scheduled, don't wake them
6337 : * up again. Otherwise check if we need to look through the waitqueue to
6338 : * wake other backends.
6339 : */
6340 165565890 : if ((lockstate & BM_LOCK_HAS_WAITERS) &&
6341 185330 : !(lockstate & BM_LOCK_WAKE_IN_PROGRESS))
6342 : {
6343 75150 : if ((lockstate & BM_LOCK_MASK) == 0)
6344 : {
6345 : /*
6346 : * We released a lock and the lock was, in that moment, free. We
6347 : * therefore can wake waiters for any kind of lock.
6348 : */
6349 45350 : check_waiters = true;
6350 45350 : wake_exclusive = true;
6351 : }
6352 29800 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6353 : {
6354 : /*
6355 : * We released the lock, but another backend still holds a lock.
6356 : * We can't have released an exclusive lock, as there couldn't
6357 : * have been other lock holders. If we released a share lock, no
6358 : * waiters need to be woken up, as there must be other share
6359 : * lockers. However, if we held a share-exclusive lock, another
6360 : * backend now could acquire a share-exclusive lock.
6361 : */
6362 0 : check_waiters = true;
6363 0 : wake_exclusive = false;
6364 : }
6365 : }
6366 :
6367 : /*
6368 : * As waking up waiters requires the spinlock to be acquired, only do so
6369 : * if necessary.
6370 : */
6371 165565890 : if (check_waiters)
6372 45350 : BufferLockWakeup(buf_hdr, wake_exclusive);
6373 165565890 : }
6374 :
6375 : /*
6376 : * BufferLockHeldByMeInMode - test whether my process holds the content lock
6377 : * in the specified mode
6378 : *
6379 : * This is meant as debug support only.
6380 : */
6381 : static bool
6382 0 : BufferLockHeldByMeInMode(BufferDesc *buf_hdr, BufferLockMode mode)
6383 : {
6384 : PrivateRefCountEntry *entry =
6385 0 : GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf_hdr), false);
6386 :
6387 0 : if (!entry)
6388 0 : return false;
6389 : else
6390 0 : return entry->data.lockmode == mode;
6391 : }
6392 :
6393 : /*
6394 : * BufferLockHeldByMe - test whether my process holds the content lock in any
6395 : * mode
6396 : *
6397 : * This is meant as debug support only.
6398 : */
6399 : static bool
6400 0 : BufferLockHeldByMe(BufferDesc *buf_hdr)
6401 : {
6402 : PrivateRefCountEntry *entry =
6403 0 : GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf_hdr), false);
6404 :
6405 0 : if (!entry)
6406 0 : return false;
6407 : else
6408 0 : return entry->data.lockmode != BUFFER_LOCK_UNLOCK;
6409 : }
6410 :
6411 : /*
6412 : * Release the content lock for the buffer.
6413 : */
6414 : void
6415 174930990 : UnlockBuffer(Buffer buffer)
6416 : {
6417 : BufferDesc *buf_hdr;
6418 :
6419 : Assert(BufferIsPinned(buffer));
6420 174930990 : if (BufferIsLocal(buffer))
6421 9977522 : return; /* local buffers need no lock */
6422 :
6423 164953468 : buf_hdr = GetBufferDescriptor(buffer - 1);
6424 164953468 : BufferLockUnlock(buffer, buf_hdr);
6425 : }
6426 :
6427 : /*
6428 : * Acquire the content_lock for the buffer.
6429 : */
6430 : void
6431 171695034 : LockBufferInternal(Buffer buffer, BufferLockMode mode)
6432 : {
6433 : BufferDesc *buf_hdr;
6434 :
6435 : /*
6436 : * We can't wait if we haven't got a PGPROC. This should only occur
6437 : * during bootstrap or shared memory initialization. Put an Assert here
6438 : * to catch unsafe coding practices.
6439 : */
6440 : Assert(!(MyProc == NULL && IsUnderPostmaster));
6441 :
6442 : /* handled in LockBuffer() wrapper */
6443 : Assert(mode != BUFFER_LOCK_UNLOCK);
6444 :
6445 : Assert(BufferIsPinned(buffer));
6446 171695034 : if (BufferIsLocal(buffer))
6447 9823772 : return; /* local buffers need no lock */
6448 :
6449 161871262 : buf_hdr = GetBufferDescriptor(buffer - 1);
6450 :
6451 : /*
6452 : * Test the most frequent lock modes first. While a switch (mode) would be
6453 : * nice, at least gcc generates considerably worse code for it.
6454 : *
6455 : * Call BufferLockAcquire() with a constant argument for mode, to generate
6456 : * more efficient code for the different lock modes.
6457 : */
6458 161871262 : if (mode == BUFFER_LOCK_SHARE)
6459 113861984 : BufferLockAcquire(buffer, buf_hdr, BUFFER_LOCK_SHARE);
6460 48009278 : else if (mode == BUFFER_LOCK_EXCLUSIVE)
6461 48009278 : BufferLockAcquire(buffer, buf_hdr, BUFFER_LOCK_EXCLUSIVE);
6462 0 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6463 0 : BufferLockAcquire(buffer, buf_hdr, BUFFER_LOCK_SHARE_EXCLUSIVE);
6464 : else
6465 0 : elog(ERROR, "unrecognized buffer lock mode: %d", mode);
6466 : }
6467 :
6468 : /*
6469 : * Acquire the content_lock for the buffer, but only if we don't have to wait.
6470 : *
6471 : * This assumes the caller wants BUFFER_LOCK_EXCLUSIVE mode.
6472 : */
6473 : bool
6474 2685950 : ConditionalLockBuffer(Buffer buffer)
6475 : {
6476 : BufferDesc *buf;
6477 :
6478 : Assert(BufferIsPinned(buffer));
6479 2685950 : if (BufferIsLocal(buffer))
6480 129508 : return true; /* act as though we got it */
6481 :
6482 2556442 : buf = GetBufferDescriptor(buffer - 1);
6483 :
6484 2556442 : return BufferLockConditional(buffer, buf, BUFFER_LOCK_EXCLUSIVE);
6485 : }
6486 :
6487 : /*
6488 : * Verify that this backend is pinning the buffer exactly once.
6489 : *
6490 : * NOTE: Like in BufferIsPinned(), what we check here is that *this* backend
6491 : * holds a pin on the buffer. We do not care whether some other backend does.
6492 : */
6493 : void
6494 4719362 : CheckBufferIsPinnedOnce(Buffer buffer)
6495 : {
6496 4719362 : if (BufferIsLocal(buffer))
6497 : {
6498 1578 : if (LocalRefCount[-buffer - 1] != 1)
6499 0 : elog(ERROR, "incorrect local pin count: %d",
6500 : LocalRefCount[-buffer - 1]);
6501 : }
6502 : else
6503 : {
6504 4717784 : if (GetPrivateRefCount(buffer) != 1)
6505 0 : elog(ERROR, "incorrect local pin count: %d",
6506 : GetPrivateRefCount(buffer));
6507 : }
6508 4719362 : }
6509 :
6510 : /*
6511 : * LockBufferForCleanup - lock a buffer in preparation for deleting items
6512 : *
6513 : * Items may be deleted from a disk page only when the caller (a) holds an
6514 : * exclusive lock on the buffer and (b) has observed that no other backend
6515 : * holds a pin on the buffer. If there is a pin, then the other backend
6516 : * might have a pointer into the buffer (for example, a heapscan reference
6517 : * to an item --- see README for more details). It's OK if a pin is added
6518 : * after the cleanup starts, however; the newly-arrived backend will be
6519 : * unable to look at the page until we release the exclusive lock.
6520 : *
6521 : * To implement this protocol, a would-be deleter must pin the buffer and
6522 : * then call LockBufferForCleanup(). LockBufferForCleanup() is similar to
6523 : * LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE), except that it loops until
6524 : * it has successfully observed pin count = 1.
6525 : */
6526 : void
6527 47108 : LockBufferForCleanup(Buffer buffer)
6528 : {
6529 : BufferDesc *bufHdr;
6530 47108 : TimestampTz waitStart = 0;
6531 47108 : bool waiting = false;
6532 47108 : bool logged_recovery_conflict = false;
6533 :
6534 : Assert(BufferIsPinned(buffer));
6535 : Assert(PinCountWaitBuf == NULL);
6536 :
6537 47108 : CheckBufferIsPinnedOnce(buffer);
6538 :
6539 : /*
6540 : * We do not yet need to be worried about in-progress AIOs holding a pin,
6541 : * as we, so far, only support doing reads via AIO and this function can
6542 : * only be called once the buffer is valid (i.e. no read can be in
6543 : * flight).
6544 : */
6545 :
6546 : /* Nobody else to wait for */
6547 47108 : if (BufferIsLocal(buffer))
6548 32 : return;
6549 :
6550 47076 : bufHdr = GetBufferDescriptor(buffer - 1);
6551 :
6552 : for (;;)
6553 198 : {
6554 : uint64 buf_state;
6555 47274 : uint64 unset_bits = 0;
6556 :
6557 : /* Try to acquire lock */
6558 47274 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6559 47274 : buf_state = LockBufHdr(bufHdr);
6560 :
6561 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6562 47274 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
6563 : {
6564 : /* Successfully acquired exclusive lock with pincount 1 */
6565 47076 : UnlockBufHdr(bufHdr);
6566 :
6567 : /*
6568 : * Emit the log message if recovery conflict on buffer pin was
6569 : * resolved but the startup process waited longer than
6570 : * deadlock_timeout for it.
6571 : */
6572 47076 : if (logged_recovery_conflict)
6573 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
6574 : waitStart, GetCurrentTimestamp(),
6575 : NULL, false);
6576 :
6577 47076 : if (waiting)
6578 : {
6579 : /* reset ps display to remove the suffix if we added one */
6580 4 : set_ps_display_remove_suffix();
6581 4 : waiting = false;
6582 : }
6583 47076 : return;
6584 : }
6585 : /* Failed, so mark myself as waiting for pincount 1 */
6586 198 : if (buf_state & BM_PIN_COUNT_WAITER)
6587 : {
6588 0 : UnlockBufHdr(bufHdr);
6589 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6590 0 : elog(ERROR, "multiple backends attempting to wait for pincount 1");
6591 : }
6592 198 : bufHdr->wait_backend_pgprocno = MyProcNumber;
6593 198 : PinCountWaitBuf = bufHdr;
6594 198 : UnlockBufHdrExt(bufHdr, buf_state,
6595 : BM_PIN_COUNT_WAITER, 0,
6596 : 0);
6597 198 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6598 :
6599 : /* Wait to be signaled by UnpinBuffer() */
6600 198 : if (InHotStandby)
6601 : {
6602 18 : if (!waiting)
6603 : {
6604 : /* adjust the process title to indicate that it's waiting */
6605 4 : set_ps_display_suffix("waiting");
6606 4 : waiting = true;
6607 : }
6608 :
6609 : /*
6610 : * Emit the log message if the startup process is waiting longer
6611 : * than deadlock_timeout for recovery conflict on buffer pin.
6612 : *
6613 : * Skip this if first time through because the startup process has
6614 : * not started waiting yet in this case. So, the wait start
6615 : * timestamp is set after this logic.
6616 : */
6617 18 : if (waitStart != 0 && !logged_recovery_conflict)
6618 : {
6619 6 : TimestampTz now = GetCurrentTimestamp();
6620 :
6621 6 : if (TimestampDifferenceExceeds(waitStart, now,
6622 : DeadlockTimeout))
6623 : {
6624 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
6625 : waitStart, now, NULL, true);
6626 4 : logged_recovery_conflict = true;
6627 : }
6628 : }
6629 :
6630 : /*
6631 : * Set the wait start timestamp if logging is enabled and first
6632 : * time through.
6633 : */
6634 18 : if (log_recovery_conflict_waits && waitStart == 0)
6635 4 : waitStart = GetCurrentTimestamp();
6636 :
6637 : /* Publish the bufid that Startup process waits on */
6638 18 : SetStartupBufferPinWaitBufId(buffer - 1);
6639 : /* Set alarm and then wait to be signaled by UnpinBuffer() */
6640 18 : ResolveRecoveryConflictWithBufferPin();
6641 : /* Reset the published bufid */
6642 18 : SetStartupBufferPinWaitBufId(-1);
6643 : }
6644 : else
6645 180 : ProcWaitForSignal(WAIT_EVENT_BUFFER_CLEANUP);
6646 :
6647 : /*
6648 : * Remove flag marking us as waiter. Normally this will not be set
6649 : * anymore, but ProcWaitForSignal() can return for other signals as
6650 : * well. We take care to only reset the flag if we're the waiter, as
6651 : * theoretically another backend could have started waiting. That's
6652 : * impossible with the current usages due to table level locking, but
6653 : * better be safe.
6654 : */
6655 198 : buf_state = LockBufHdr(bufHdr);
6656 198 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
6657 14 : bufHdr->wait_backend_pgprocno == MyProcNumber)
6658 14 : unset_bits |= BM_PIN_COUNT_WAITER;
6659 :
6660 198 : UnlockBufHdrExt(bufHdr, buf_state,
6661 : 0, unset_bits,
6662 : 0);
6663 :
6664 198 : PinCountWaitBuf = NULL;
6665 : /* Loop back and try again */
6666 : }
6667 : }
6668 :
6669 : /*
6670 : * Check called from ProcessRecoveryConflictInterrupts() when Startup process
6671 : * requests cancellation of all pin holders that are blocking it.
6672 : */
6673 : bool
6674 6 : HoldingBufferPinThatDelaysRecovery(void)
6675 : {
6676 6 : int bufid = GetStartupBufferPinWaitBufId();
6677 :
6678 : /*
6679 : * If we get woken slowly then it's possible that the Startup process was
6680 : * already woken by other backends before we got here. Also possible that
6681 : * we get here by multiple interrupts or interrupts at inappropriate
6682 : * times, so make sure we do nothing if the bufid is not set.
6683 : */
6684 6 : if (bufid < 0)
6685 2 : return false;
6686 :
6687 4 : if (GetPrivateRefCount(bufid + 1) > 0)
6688 4 : return true;
6689 :
6690 0 : return false;
6691 : }
6692 :
6693 : /*
6694 : * ConditionalLockBufferForCleanup - as above, but don't wait to get the lock
6695 : *
6696 : * We won't loop, but just check once to see if the pin count is OK. If
6697 : * not, return false with no lock held.
6698 : */
6699 : bool
6700 837864 : ConditionalLockBufferForCleanup(Buffer buffer)
6701 : {
6702 : BufferDesc *bufHdr;
6703 : uint64 buf_state,
6704 : refcount;
6705 :
6706 : Assert(BufferIsValid(buffer));
6707 :
6708 : /* see AIO related comment in LockBufferForCleanup() */
6709 :
6710 837864 : if (BufferIsLocal(buffer))
6711 : {
6712 1608 : refcount = LocalRefCount[-buffer - 1];
6713 : /* There should be exactly one pin */
6714 : Assert(refcount > 0);
6715 1608 : if (refcount != 1)
6716 42 : return false;
6717 : /* Nobody else to wait for */
6718 1566 : return true;
6719 : }
6720 :
6721 : /* There should be exactly one local pin */
6722 836256 : refcount = GetPrivateRefCount(buffer);
6723 : Assert(refcount);
6724 836256 : if (refcount != 1)
6725 568 : return false;
6726 :
6727 : /* Try to acquire lock */
6728 835688 : if (!ConditionalLockBuffer(buffer))
6729 62 : return false;
6730 :
6731 835626 : bufHdr = GetBufferDescriptor(buffer - 1);
6732 835626 : buf_state = LockBufHdr(bufHdr);
6733 835626 : refcount = BUF_STATE_GET_REFCOUNT(buf_state);
6734 :
6735 : Assert(refcount > 0);
6736 835626 : if (refcount == 1)
6737 : {
6738 : /* Successfully acquired exclusive lock with pincount 1 */
6739 834996 : UnlockBufHdr(bufHdr);
6740 834996 : return true;
6741 : }
6742 :
6743 : /* Failed, so release the lock */
6744 630 : UnlockBufHdr(bufHdr);
6745 630 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6746 630 : return false;
6747 : }
6748 :
6749 : /*
6750 : * IsBufferCleanupOK - as above, but we already have the lock
6751 : *
6752 : * Check whether it's OK to perform cleanup on a buffer we've already
6753 : * locked. If we observe that the pin count is 1, our exclusive lock
6754 : * happens to be a cleanup lock, and we can proceed with anything that
6755 : * would have been allowable had we sought a cleanup lock originally.
6756 : */
6757 : bool
6758 3314 : IsBufferCleanupOK(Buffer buffer)
6759 : {
6760 : BufferDesc *bufHdr;
6761 : uint64 buf_state;
6762 :
6763 : Assert(BufferIsValid(buffer));
6764 :
6765 : /* see AIO related comment in LockBufferForCleanup() */
6766 :
6767 3314 : if (BufferIsLocal(buffer))
6768 : {
6769 : /* There should be exactly one pin */
6770 0 : if (LocalRefCount[-buffer - 1] != 1)
6771 0 : return false;
6772 : /* Nobody else to wait for */
6773 0 : return true;
6774 : }
6775 :
6776 : /* There should be exactly one local pin */
6777 3314 : if (GetPrivateRefCount(buffer) != 1)
6778 0 : return false;
6779 :
6780 3314 : bufHdr = GetBufferDescriptor(buffer - 1);
6781 :
6782 : /* caller must hold exclusive lock on buffer */
6783 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
6784 :
6785 3314 : buf_state = LockBufHdr(bufHdr);
6786 :
6787 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6788 3314 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
6789 : {
6790 : /* pincount is OK. */
6791 3314 : UnlockBufHdr(bufHdr);
6792 3314 : return true;
6793 : }
6794 :
6795 0 : UnlockBufHdr(bufHdr);
6796 0 : return false;
6797 : }
6798 :
6799 :
6800 : /*
6801 : * Functions for buffer I/O handling
6802 : *
6803 : * Also note that these are used only for shared buffers, not local ones.
6804 : */
6805 :
6806 : /*
6807 : * WaitIO -- Block until the IO_IN_PROGRESS flag on 'buf' is cleared.
6808 : */
6809 : static void
6810 9418 : WaitIO(BufferDesc *buf)
6811 : {
6812 9418 : ConditionVariable *cv = BufferDescriptorGetIOCV(buf);
6813 :
6814 9418 : ConditionVariablePrepareToSleep(cv);
6815 : for (;;)
6816 9380 : {
6817 : uint64 buf_state;
6818 : PgAioWaitRef iow;
6819 :
6820 : /*
6821 : * It may not be necessary to acquire the spinlock to check the flag
6822 : * here, but since this test is essential for correctness, we'd better
6823 : * play it safe.
6824 : */
6825 18798 : buf_state = LockBufHdr(buf);
6826 :
6827 : /*
6828 : * Copy the wait reference while holding the spinlock. This protects
6829 : * against a concurrent TerminateBufferIO() in another backend from
6830 : * clearing the wref while it's being read.
6831 : */
6832 18798 : iow = buf->io_wref;
6833 18798 : UnlockBufHdr(buf);
6834 :
6835 : /* no IO in progress, we don't need to wait */
6836 18798 : if (!(buf_state & BM_IO_IN_PROGRESS))
6837 9418 : break;
6838 :
6839 : /*
6840 : * The buffer has asynchronous IO in progress, wait for it to
6841 : * complete.
6842 : */
6843 9380 : if (pgaio_wref_valid(&iow))
6844 : {
6845 7794 : pgaio_wref_wait(&iow);
6846 :
6847 : /*
6848 : * The AIO subsystem internally uses condition variables and thus
6849 : * might remove this backend from the BufferDesc's CV. While that
6850 : * wouldn't cause a correctness issue (the first CV sleep just
6851 : * immediately returns if not already registered), it seems worth
6852 : * avoiding unnecessary loop iterations, given that we take care
6853 : * to do so at the start of the function.
6854 : */
6855 7794 : ConditionVariablePrepareToSleep(cv);
6856 7794 : continue;
6857 : }
6858 :
6859 : /* wait on BufferDesc->cv, e.g. for concurrent synchronous IO */
6860 1586 : ConditionVariableSleep(cv, WAIT_EVENT_BUFFER_IO);
6861 : }
6862 9418 : ConditionVariableCancelSleep();
6863 9418 : }
6864 :
6865 : /*
6866 : * StartBufferIO: begin I/O on this buffer
6867 : * (Assumptions)
6868 : * My process is executing no IO on this buffer
6869 : * The buffer is Pinned
6870 : *
6871 : * In some scenarios multiple backends could attempt the same I/O operation
6872 : * concurrently. If someone else has already started I/O on this buffer then
6873 : * we will wait for completion of the IO using WaitIO().
6874 : *
6875 : * Input operations are only attempted on buffers that are not BM_VALID,
6876 : * and output operations only on buffers that are BM_VALID and BM_DIRTY,
6877 : * so we can always tell if the work is already done.
6878 : *
6879 : * Returns true if we successfully marked the buffer as I/O busy,
6880 : * false if someone else already did the work.
6881 : *
6882 : * If nowait is true, then we don't wait for an I/O to be finished by another
6883 : * backend. In that case, false indicates either that the I/O was already
6884 : * finished, or is still in progress. This is useful for callers that want to
6885 : * find out if they can perform the I/O as part of a larger operation, without
6886 : * waiting for the answer or distinguishing the reasons why not.
6887 : */
6888 : bool
6889 5064652 : StartBufferIO(BufferDesc *buf, bool forInput, bool nowait)
6890 : {
6891 : uint64 buf_state;
6892 :
6893 5064652 : ResourceOwnerEnlarge(CurrentResourceOwner);
6894 :
6895 : for (;;)
6896 : {
6897 5074050 : buf_state = LockBufHdr(buf);
6898 :
6899 5074050 : if (!(buf_state & BM_IO_IN_PROGRESS))
6900 5064644 : break;
6901 9406 : UnlockBufHdr(buf);
6902 9406 : if (nowait)
6903 8 : return false;
6904 9398 : WaitIO(buf);
6905 : }
6906 :
6907 : /* Once we get here, there is definitely no I/O active on this buffer */
6908 :
6909 : /* Check if someone else already did the I/O */
6910 5064644 : if (forInput ? (buf_state & BM_VALID) : !(buf_state & BM_DIRTY))
6911 : {
6912 10678 : UnlockBufHdr(buf);
6913 10678 : return false;
6914 : }
6915 :
6916 5053966 : UnlockBufHdrExt(buf, buf_state,
6917 : BM_IO_IN_PROGRESS, 0,
6918 : 0);
6919 :
6920 5053966 : ResourceOwnerRememberBufferIO(CurrentResourceOwner,
6921 : BufferDescriptorGetBuffer(buf));
6922 :
6923 5053966 : return true;
6924 : }
6925 :
6926 : /*
6927 : * TerminateBufferIO: release a buffer we were doing I/O on
6928 : * (Assumptions)
6929 : * My process is executing IO for the buffer
6930 : * BM_IO_IN_PROGRESS bit is set for the buffer
6931 : * The buffer is Pinned
6932 : *
6933 : * If clear_dirty is true and BM_JUST_DIRTIED is not set, we clear the
6934 : * buffer's BM_DIRTY flag. This is appropriate when terminating a
6935 : * successful write. The check on BM_JUST_DIRTIED is necessary to avoid
6936 : * marking the buffer clean if it was re-dirtied while we were writing.
6937 : *
6938 : * set_flag_bits gets ORed into the buffer's flags. It must include
6939 : * BM_IO_ERROR in a failure case. For successful completion it could
6940 : * be 0, or BM_VALID if we just finished reading in the page.
6941 : *
6942 : * If forget_owner is true, we release the buffer I/O from the current
6943 : * resource owner. (forget_owner=false is used when the resource owner itself
6944 : * is being released)
6945 : */
6946 : void
6947 4774186 : TerminateBufferIO(BufferDesc *buf, bool clear_dirty, uint64 set_flag_bits,
6948 : bool forget_owner, bool release_aio)
6949 : {
6950 : uint64 buf_state;
6951 4774186 : uint64 unset_flag_bits = 0;
6952 4774186 : int refcount_change = 0;
6953 :
6954 4774186 : buf_state = LockBufHdr(buf);
6955 :
6956 : Assert(buf_state & BM_IO_IN_PROGRESS);
6957 4774186 : unset_flag_bits |= BM_IO_IN_PROGRESS;
6958 :
6959 : /* Clear earlier errors, if this IO failed, it'll be marked again */
6960 4774186 : unset_flag_bits |= BM_IO_ERROR;
6961 :
6962 4774186 : if (clear_dirty && !(buf_state & BM_JUST_DIRTIED))
6963 1145376 : unset_flag_bits |= BM_DIRTY | BM_CHECKPOINT_NEEDED;
6964 :
6965 4774186 : if (release_aio)
6966 : {
6967 : /* release ownership by the AIO subsystem */
6968 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6969 2606506 : refcount_change = -1;
6970 2606506 : pgaio_wref_clear(&buf->io_wref);
6971 : }
6972 :
6973 4774186 : buf_state = UnlockBufHdrExt(buf, buf_state,
6974 : set_flag_bits, unset_flag_bits,
6975 : refcount_change);
6976 :
6977 4774186 : if (forget_owner)
6978 2167638 : ResourceOwnerForgetBufferIO(CurrentResourceOwner,
6979 : BufferDescriptorGetBuffer(buf));
6980 :
6981 4774186 : ConditionVariableBroadcast(BufferDescriptorGetIOCV(buf));
6982 :
6983 : /*
6984 : * Support LockBufferForCleanup()
6985 : *
6986 : * We may have just released the last pin other than the waiter's. In most
6987 : * cases, this backend holds another pin on the buffer. But, if, for
6988 : * example, this backend is completing an IO issued by another backend, it
6989 : * may be time to wake the waiter.
6990 : */
6991 4774186 : if (release_aio && (buf_state & BM_PIN_COUNT_WAITER))
6992 0 : WakePinCountWaiter(buf);
6993 4774186 : }
6994 :
6995 : /*
6996 : * AbortBufferIO: Clean up active buffer I/O after an error.
6997 : *
6998 : * All LWLocks & content locks we might have held have been released, but we
6999 : * haven't yet released buffer pins, so the buffer is still pinned.
7000 : *
7001 : * If I/O was in progress, we always set BM_IO_ERROR, even though it's
7002 : * possible the error condition wasn't related to the I/O.
7003 : *
7004 : * Note: this does not remove the buffer I/O from the resource owner.
7005 : * That's correct when we're releasing the whole resource owner, but
7006 : * beware if you use this in other contexts.
7007 : */
7008 : static void
7009 30 : AbortBufferIO(Buffer buffer)
7010 : {
7011 30 : BufferDesc *buf_hdr = GetBufferDescriptor(buffer - 1);
7012 : uint64 buf_state;
7013 :
7014 30 : buf_state = LockBufHdr(buf_hdr);
7015 : Assert(buf_state & (BM_IO_IN_PROGRESS | BM_TAG_VALID));
7016 :
7017 30 : if (!(buf_state & BM_VALID))
7018 : {
7019 : Assert(!(buf_state & BM_DIRTY));
7020 30 : UnlockBufHdr(buf_hdr);
7021 : }
7022 : else
7023 : {
7024 : Assert(buf_state & BM_DIRTY);
7025 0 : UnlockBufHdr(buf_hdr);
7026 :
7027 : /* Issue notice if this is not the first failure... */
7028 0 : if (buf_state & BM_IO_ERROR)
7029 : {
7030 : /* Buffer is pinned, so we can read tag without spinlock */
7031 0 : ereport(WARNING,
7032 : (errcode(ERRCODE_IO_ERROR),
7033 : errmsg("could not write block %u of %s",
7034 : buf_hdr->tag.blockNum,
7035 : relpathperm(BufTagGetRelFileLocator(&buf_hdr->tag),
7036 : BufTagGetForkNum(&buf_hdr->tag)).str),
7037 : errdetail("Multiple failures --- write error might be permanent.")));
7038 : }
7039 : }
7040 :
7041 30 : TerminateBufferIO(buf_hdr, false, BM_IO_ERROR, false, false);
7042 30 : }
7043 :
7044 : /*
7045 : * Error context callback for errors occurring during shared buffer writes.
7046 : */
7047 : static void
7048 76 : shared_buffer_write_error_callback(void *arg)
7049 : {
7050 76 : BufferDesc *bufHdr = (BufferDesc *) arg;
7051 :
7052 : /* Buffer is pinned, so we can read the tag without locking the spinlock */
7053 76 : if (bufHdr != NULL)
7054 152 : errcontext("writing block %u of relation \"%s\"",
7055 : bufHdr->tag.blockNum,
7056 76 : relpathperm(BufTagGetRelFileLocator(&bufHdr->tag),
7057 : BufTagGetForkNum(&bufHdr->tag)).str);
7058 76 : }
7059 :
7060 : /*
7061 : * Error context callback for errors occurring during local buffer writes.
7062 : */
7063 : static void
7064 0 : local_buffer_write_error_callback(void *arg)
7065 : {
7066 0 : BufferDesc *bufHdr = (BufferDesc *) arg;
7067 :
7068 0 : if (bufHdr != NULL)
7069 0 : errcontext("writing block %u of relation \"%s\"",
7070 : bufHdr->tag.blockNum,
7071 0 : relpathbackend(BufTagGetRelFileLocator(&bufHdr->tag),
7072 : MyProcNumber,
7073 : BufTagGetForkNum(&bufHdr->tag)).str);
7074 0 : }
7075 :
7076 : /*
7077 : * RelFileLocator qsort/bsearch comparator; see RelFileLocatorEquals.
7078 : */
7079 : static int
7080 19579604 : rlocator_comparator(const void *p1, const void *p2)
7081 : {
7082 19579604 : RelFileLocator n1 = *(const RelFileLocator *) p1;
7083 19579604 : RelFileLocator n2 = *(const RelFileLocator *) p2;
7084 :
7085 19579604 : if (n1.relNumber < n2.relNumber)
7086 19504014 : return -1;
7087 75590 : else if (n1.relNumber > n2.relNumber)
7088 72624 : return 1;
7089 :
7090 2966 : if (n1.dbOid < n2.dbOid)
7091 0 : return -1;
7092 2966 : else if (n1.dbOid > n2.dbOid)
7093 0 : return 1;
7094 :
7095 2966 : if (n1.spcOid < n2.spcOid)
7096 0 : return -1;
7097 2966 : else if (n1.spcOid > n2.spcOid)
7098 0 : return 1;
7099 : else
7100 2966 : return 0;
7101 : }
7102 :
7103 : /*
7104 : * Lock buffer header - set BM_LOCKED in buffer state.
7105 : */
7106 : uint64
7107 64818622 : LockBufHdr(BufferDesc *desc)
7108 : {
7109 : uint64 old_buf_state;
7110 :
7111 : Assert(!BufferIsLocal(BufferDescriptorGetBuffer(desc)));
7112 :
7113 : while (true)
7114 : {
7115 : /*
7116 : * Always try once to acquire the lock directly, without setting up
7117 : * the spin-delay infrastructure. The work necessary for that shows up
7118 : * in profiles and is rarely necessary.
7119 : */
7120 64823878 : old_buf_state = pg_atomic_fetch_or_u64(&desc->state, BM_LOCKED);
7121 64823878 : if (likely(!(old_buf_state & BM_LOCKED)))
7122 64818622 : break; /* got lock */
7123 :
7124 : /* and then spin without atomic operations until lock is released */
7125 : {
7126 : SpinDelayStatus delayStatus;
7127 :
7128 5256 : init_local_spin_delay(&delayStatus);
7129 :
7130 14564 : while (old_buf_state & BM_LOCKED)
7131 : {
7132 9308 : perform_spin_delay(&delayStatus);
7133 9308 : old_buf_state = pg_atomic_read_u64(&desc->state);
7134 : }
7135 5256 : finish_spin_delay(&delayStatus);
7136 : }
7137 :
7138 : /*
7139 : * Retry. The lock might obviously already be re-acquired by the time
7140 : * we're attempting to get it again.
7141 : */
7142 : }
7143 :
7144 64818622 : return old_buf_state | BM_LOCKED;
7145 : }
7146 :
7147 : /*
7148 : * Wait until the BM_LOCKED flag isn't set anymore and return the buffer's
7149 : * state at that point.
7150 : *
7151 : * Obviously the buffer could be locked by the time the value is returned, so
7152 : * this is primarily useful in CAS style loops.
7153 : */
7154 : pg_noinline uint64
7155 1608 : WaitBufHdrUnlocked(BufferDesc *buf)
7156 : {
7157 : SpinDelayStatus delayStatus;
7158 : uint64 buf_state;
7159 :
7160 1608 : init_local_spin_delay(&delayStatus);
7161 :
7162 1608 : buf_state = pg_atomic_read_u64(&buf->state);
7163 :
7164 86096 : while (buf_state & BM_LOCKED)
7165 : {
7166 84488 : perform_spin_delay(&delayStatus);
7167 84488 : buf_state = pg_atomic_read_u64(&buf->state);
7168 : }
7169 :
7170 1608 : finish_spin_delay(&delayStatus);
7171 :
7172 1608 : return buf_state;
7173 : }
7174 :
7175 : /*
7176 : * BufferTag comparator.
7177 : */
7178 : static inline int
7179 0 : buffertag_comparator(const BufferTag *ba, const BufferTag *bb)
7180 : {
7181 : int ret;
7182 : RelFileLocator rlocatora;
7183 : RelFileLocator rlocatorb;
7184 :
7185 0 : rlocatora = BufTagGetRelFileLocator(ba);
7186 0 : rlocatorb = BufTagGetRelFileLocator(bb);
7187 :
7188 0 : ret = rlocator_comparator(&rlocatora, &rlocatorb);
7189 :
7190 0 : if (ret != 0)
7191 0 : return ret;
7192 :
7193 0 : if (BufTagGetForkNum(ba) < BufTagGetForkNum(bb))
7194 0 : return -1;
7195 0 : if (BufTagGetForkNum(ba) > BufTagGetForkNum(bb))
7196 0 : return 1;
7197 :
7198 0 : if (ba->blockNum < bb->blockNum)
7199 0 : return -1;
7200 0 : if (ba->blockNum > bb->blockNum)
7201 0 : return 1;
7202 :
7203 0 : return 0;
7204 : }
7205 :
7206 : /*
7207 : * Comparator determining the writeout order in a checkpoint.
7208 : *
7209 : * It is important that tablespaces are compared first, the logic balancing
7210 : * writes between tablespaces relies on it.
7211 : */
7212 : static inline int
7213 6073580 : ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b)
7214 : {
7215 : /* compare tablespace */
7216 6073580 : if (a->tsId < b->tsId)
7217 20132 : return -1;
7218 6053448 : else if (a->tsId > b->tsId)
7219 53770 : return 1;
7220 : /* compare relation */
7221 5999678 : if (a->relNumber < b->relNumber)
7222 1675564 : return -1;
7223 4324114 : else if (a->relNumber > b->relNumber)
7224 1632094 : return 1;
7225 : /* compare fork */
7226 2692020 : else if (a->forkNum < b->forkNum)
7227 122324 : return -1;
7228 2569696 : else if (a->forkNum > b->forkNum)
7229 122228 : return 1;
7230 : /* compare block number */
7231 2447468 : else if (a->blockNum < b->blockNum)
7232 1198494 : return -1;
7233 1248974 : else if (a->blockNum > b->blockNum)
7234 1175736 : return 1;
7235 : /* equal page IDs are unlikely, but not impossible */
7236 73238 : return 0;
7237 : }
7238 :
7239 : /*
7240 : * Comparator for a Min-Heap over the per-tablespace checkpoint completion
7241 : * progress.
7242 : */
7243 : static int
7244 499176 : ts_ckpt_progress_comparator(Datum a, Datum b, void *arg)
7245 : {
7246 499176 : CkptTsStatus *sa = (CkptTsStatus *) DatumGetPointer(a);
7247 499176 : CkptTsStatus *sb = (CkptTsStatus *) DatumGetPointer(b);
7248 :
7249 : /* we want a min-heap, so return 1 for the a < b */
7250 499176 : if (sa->progress < sb->progress)
7251 449984 : return 1;
7252 49192 : else if (sa->progress == sb->progress)
7253 3116 : return 0;
7254 : else
7255 46076 : return -1;
7256 : }
7257 :
7258 : /*
7259 : * Initialize a writeback context, discarding potential previous state.
7260 : *
7261 : * *max_pending is a pointer instead of an immediate value, so the coalesce
7262 : * limits can easily changed by the GUC mechanism, and so calling code does
7263 : * not have to check the current configuration. A value of 0 means that no
7264 : * writeback control will be performed.
7265 : */
7266 : void
7267 5612 : WritebackContextInit(WritebackContext *context, int *max_pending)
7268 : {
7269 : Assert(*max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
7270 :
7271 5612 : context->max_pending = max_pending;
7272 5612 : context->nr_pending = 0;
7273 5612 : }
7274 :
7275 : /*
7276 : * Add buffer to list of pending writeback requests.
7277 : */
7278 : void
7279 1138544 : ScheduleBufferTagForWriteback(WritebackContext *wb_context, IOContext io_context,
7280 : BufferTag *tag)
7281 : {
7282 : PendingWriteback *pending;
7283 :
7284 : /*
7285 : * As pg_flush_data() doesn't do anything with fsync disabled, there's no
7286 : * point in tracking in that case.
7287 : */
7288 1138544 : if (io_direct_flags & IO_DIRECT_DATA ||
7289 1137500 : !enableFsync)
7290 1138536 : return;
7291 :
7292 : /*
7293 : * Add buffer to the pending writeback array, unless writeback control is
7294 : * disabled.
7295 : */
7296 8 : if (*wb_context->max_pending > 0)
7297 : {
7298 : Assert(*wb_context->max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
7299 :
7300 0 : pending = &wb_context->pending_writebacks[wb_context->nr_pending++];
7301 :
7302 0 : pending->tag = *tag;
7303 : }
7304 :
7305 : /*
7306 : * Perform pending flushes if the writeback limit is exceeded. This
7307 : * includes the case where previously an item has been added, but control
7308 : * is now disabled.
7309 : */
7310 8 : if (wb_context->nr_pending >= *wb_context->max_pending)
7311 8 : IssuePendingWritebacks(wb_context, io_context);
7312 : }
7313 :
7314 : #define ST_SORT sort_pending_writebacks
7315 : #define ST_ELEMENT_TYPE PendingWriteback
7316 : #define ST_COMPARE(a, b) buffertag_comparator(&a->tag, &b->tag)
7317 : #define ST_SCOPE static
7318 : #define ST_DEFINE
7319 : #include "lib/sort_template.h"
7320 :
7321 : /*
7322 : * Issue all pending writeback requests, previously scheduled with
7323 : * ScheduleBufferTagForWriteback, to the OS.
7324 : *
7325 : * Because this is only used to improve the OSs IO scheduling we try to never
7326 : * error out - it's just a hint.
7327 : */
7328 : void
7329 2198 : IssuePendingWritebacks(WritebackContext *wb_context, IOContext io_context)
7330 : {
7331 : instr_time io_start;
7332 : int i;
7333 :
7334 2198 : if (wb_context->nr_pending == 0)
7335 2198 : return;
7336 :
7337 : /*
7338 : * Executing the writes in-order can make them a lot faster, and allows to
7339 : * merge writeback requests to consecutive blocks into larger writebacks.
7340 : */
7341 0 : sort_pending_writebacks(wb_context->pending_writebacks,
7342 0 : wb_context->nr_pending);
7343 :
7344 0 : io_start = pgstat_prepare_io_time(track_io_timing);
7345 :
7346 : /*
7347 : * Coalesce neighbouring writes, but nothing else. For that we iterate
7348 : * through the, now sorted, array of pending flushes, and look forward to
7349 : * find all neighbouring (or identical) writes.
7350 : */
7351 0 : for (i = 0; i < wb_context->nr_pending; i++)
7352 : {
7353 : PendingWriteback *cur;
7354 : PendingWriteback *next;
7355 : SMgrRelation reln;
7356 : int ahead;
7357 : BufferTag tag;
7358 : RelFileLocator currlocator;
7359 0 : Size nblocks = 1;
7360 :
7361 0 : cur = &wb_context->pending_writebacks[i];
7362 0 : tag = cur->tag;
7363 0 : currlocator = BufTagGetRelFileLocator(&tag);
7364 :
7365 : /*
7366 : * Peek ahead, into following writeback requests, to see if they can
7367 : * be combined with the current one.
7368 : */
7369 0 : for (ahead = 0; i + ahead + 1 < wb_context->nr_pending; ahead++)
7370 : {
7371 :
7372 0 : next = &wb_context->pending_writebacks[i + ahead + 1];
7373 :
7374 : /* different file, stop */
7375 0 : if (!RelFileLocatorEquals(currlocator,
7376 0 : BufTagGetRelFileLocator(&next->tag)) ||
7377 0 : BufTagGetForkNum(&cur->tag) != BufTagGetForkNum(&next->tag))
7378 : break;
7379 :
7380 : /* ok, block queued twice, skip */
7381 0 : if (cur->tag.blockNum == next->tag.blockNum)
7382 0 : continue;
7383 :
7384 : /* only merge consecutive writes */
7385 0 : if (cur->tag.blockNum + 1 != next->tag.blockNum)
7386 0 : break;
7387 :
7388 0 : nblocks++;
7389 0 : cur = next;
7390 : }
7391 :
7392 0 : i += ahead;
7393 :
7394 : /* and finally tell the kernel to write the data to storage */
7395 0 : reln = smgropen(currlocator, INVALID_PROC_NUMBER);
7396 0 : smgrwriteback(reln, BufTagGetForkNum(&tag), tag.blockNum, nblocks);
7397 : }
7398 :
7399 : /*
7400 : * Assume that writeback requests are only issued for buffers containing
7401 : * blocks of permanent relations.
7402 : */
7403 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
7404 0 : IOOP_WRITEBACK, io_start, wb_context->nr_pending, 0);
7405 :
7406 0 : wb_context->nr_pending = 0;
7407 : }
7408 :
7409 : /* ResourceOwner callbacks */
7410 :
7411 : static void
7412 30 : ResOwnerReleaseBufferIO(Datum res)
7413 : {
7414 30 : Buffer buffer = DatumGetInt32(res);
7415 :
7416 30 : AbortBufferIO(buffer);
7417 30 : }
7418 :
7419 : static char *
7420 0 : ResOwnerPrintBufferIO(Datum res)
7421 : {
7422 0 : Buffer buffer = DatumGetInt32(res);
7423 :
7424 0 : return psprintf("lost track of buffer IO on buffer %d", buffer);
7425 : }
7426 :
7427 : /*
7428 : * Release buffer as part of resource owner cleanup. This will only be called
7429 : * if the buffer is pinned. If this backend held the content lock at the time
7430 : * of the error we also need to release that (note that it is not possible to
7431 : * hold a content lock without a pin).
7432 : */
7433 : static void
7434 15170 : ResOwnerReleaseBuffer(Datum res)
7435 : {
7436 15170 : Buffer buffer = DatumGetInt32(res);
7437 :
7438 : /* Like ReleaseBuffer, but don't call ResourceOwnerForgetBuffer */
7439 15170 : if (!BufferIsValid(buffer))
7440 0 : elog(ERROR, "bad buffer ID: %d", buffer);
7441 :
7442 15170 : if (BufferIsLocal(buffer))
7443 6066 : UnpinLocalBufferNoOwner(buffer);
7444 : else
7445 : {
7446 : PrivateRefCountEntry *ref;
7447 :
7448 9104 : ref = GetPrivateRefCountEntry(buffer, false);
7449 :
7450 : /* not having a private refcount would imply resowner corruption */
7451 : Assert(ref != NULL);
7452 :
7453 : /*
7454 : * If the buffer was locked at the time of the resowner release,
7455 : * release the lock now. This should only happen after errors.
7456 : */
7457 9104 : if (ref->data.lockmode != BUFFER_LOCK_UNLOCK)
7458 : {
7459 216 : BufferDesc *buf = GetBufferDescriptor(buffer - 1);
7460 :
7461 216 : HOLD_INTERRUPTS(); /* match the upcoming RESUME_INTERRUPTS */
7462 216 : BufferLockUnlock(buffer, buf);
7463 : }
7464 :
7465 9104 : UnpinBufferNoOwner(GetBufferDescriptor(buffer - 1));
7466 : }
7467 15170 : }
7468 :
7469 : static char *
7470 0 : ResOwnerPrintBuffer(Datum res)
7471 : {
7472 0 : return DebugPrintBufferRefcount(DatumGetInt32(res));
7473 : }
7474 :
7475 : /*
7476 : * Helper function to evict unpinned buffer whose buffer header lock is
7477 : * already acquired.
7478 : */
7479 : static bool
7480 4286 : EvictUnpinnedBufferInternal(BufferDesc *desc, bool *buffer_flushed)
7481 : {
7482 : uint64 buf_state;
7483 : bool result;
7484 :
7485 4286 : *buffer_flushed = false;
7486 :
7487 4286 : buf_state = pg_atomic_read_u64(&(desc->state));
7488 : Assert(buf_state & BM_LOCKED);
7489 :
7490 4286 : if ((buf_state & BM_VALID) == 0)
7491 : {
7492 0 : UnlockBufHdr(desc);
7493 0 : return false;
7494 : }
7495 :
7496 : /* Check that it's not pinned already. */
7497 4286 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
7498 : {
7499 0 : UnlockBufHdr(desc);
7500 0 : return false;
7501 : }
7502 :
7503 4286 : PinBuffer_Locked(desc); /* releases spinlock */
7504 :
7505 : /* If it was dirty, try to clean it once. */
7506 4286 : if (buf_state & BM_DIRTY)
7507 : {
7508 1946 : FlushUnlockedBuffer(desc, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
7509 1946 : *buffer_flushed = true;
7510 : }
7511 :
7512 : /* This will return false if it becomes dirty or someone else pins it. */
7513 4286 : result = InvalidateVictimBuffer(desc);
7514 :
7515 4286 : UnpinBuffer(desc);
7516 :
7517 4286 : return result;
7518 : }
7519 :
7520 : /*
7521 : * Try to evict the current block in a shared buffer.
7522 : *
7523 : * This function is intended for testing/development use only!
7524 : *
7525 : * To succeed, the buffer must not be pinned on entry, so if the caller had a
7526 : * particular block in mind, it might already have been replaced by some other
7527 : * block by the time this function runs. It's also unpinned on return, so the
7528 : * buffer might be occupied again by the time control is returned, potentially
7529 : * even by the same block. This inherent raciness without other interlocking
7530 : * makes the function unsuitable for non-testing usage.
7531 : *
7532 : * *buffer_flushed is set to true if the buffer was dirty and has been
7533 : * flushed, false otherwise. However, *buffer_flushed=true does not
7534 : * necessarily mean that we flushed the buffer, it could have been flushed by
7535 : * someone else.
7536 : *
7537 : * Returns true if the buffer was valid and it has now been made invalid.
7538 : * Returns false if it wasn't valid, if it couldn't be evicted due to a pin,
7539 : * or if the buffer becomes dirty again while we're trying to write it out.
7540 : */
7541 : bool
7542 280 : EvictUnpinnedBuffer(Buffer buf, bool *buffer_flushed)
7543 : {
7544 : BufferDesc *desc;
7545 :
7546 : Assert(BufferIsValid(buf) && !BufferIsLocal(buf));
7547 :
7548 : /* Make sure we can pin the buffer. */
7549 280 : ResourceOwnerEnlarge(CurrentResourceOwner);
7550 280 : ReservePrivateRefCountEntry();
7551 :
7552 280 : desc = GetBufferDescriptor(buf - 1);
7553 280 : LockBufHdr(desc);
7554 :
7555 280 : return EvictUnpinnedBufferInternal(desc, buffer_flushed);
7556 : }
7557 :
7558 : /*
7559 : * Try to evict all the shared buffers.
7560 : *
7561 : * This function is intended for testing/development use only! See
7562 : * EvictUnpinnedBuffer().
7563 : *
7564 : * The buffers_* parameters are mandatory and indicate the total count of
7565 : * buffers that:
7566 : * - buffers_evicted - were evicted
7567 : * - buffers_flushed - were flushed
7568 : * - buffers_skipped - could not be evicted
7569 : */
7570 : void
7571 2 : EvictAllUnpinnedBuffers(int32 *buffers_evicted, int32 *buffers_flushed,
7572 : int32 *buffers_skipped)
7573 : {
7574 2 : *buffers_evicted = 0;
7575 2 : *buffers_skipped = 0;
7576 2 : *buffers_flushed = 0;
7577 :
7578 32770 : for (int buf = 1; buf <= NBuffers; buf++)
7579 : {
7580 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
7581 : uint64 buf_state;
7582 : bool buffer_flushed;
7583 :
7584 32768 : CHECK_FOR_INTERRUPTS();
7585 :
7586 32768 : buf_state = pg_atomic_read_u64(&desc->state);
7587 32768 : if (!(buf_state & BM_VALID))
7588 28762 : continue;
7589 :
7590 4006 : ResourceOwnerEnlarge(CurrentResourceOwner);
7591 4006 : ReservePrivateRefCountEntry();
7592 :
7593 4006 : LockBufHdr(desc);
7594 :
7595 4006 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
7596 4006 : (*buffers_evicted)++;
7597 : else
7598 0 : (*buffers_skipped)++;
7599 :
7600 4006 : if (buffer_flushed)
7601 1908 : (*buffers_flushed)++;
7602 : }
7603 2 : }
7604 :
7605 : /*
7606 : * Try to evict all the shared buffers containing provided relation's pages.
7607 : *
7608 : * This function is intended for testing/development use only! See
7609 : * EvictUnpinnedBuffer().
7610 : *
7611 : * The caller must hold at least AccessShareLock on the relation to prevent
7612 : * the relation from being dropped.
7613 : *
7614 : * The buffers_* parameters are mandatory and indicate the total count of
7615 : * buffers that:
7616 : * - buffers_evicted - were evicted
7617 : * - buffers_flushed - were flushed
7618 : * - buffers_skipped - could not be evicted
7619 : */
7620 : void
7621 2 : EvictRelUnpinnedBuffers(Relation rel, int32 *buffers_evicted,
7622 : int32 *buffers_flushed, int32 *buffers_skipped)
7623 : {
7624 : Assert(!RelationUsesLocalBuffers(rel));
7625 :
7626 2 : *buffers_skipped = 0;
7627 2 : *buffers_evicted = 0;
7628 2 : *buffers_flushed = 0;
7629 :
7630 32770 : for (int buf = 1; buf <= NBuffers; buf++)
7631 : {
7632 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
7633 32768 : uint64 buf_state = pg_atomic_read_u64(&(desc->state));
7634 : bool buffer_flushed;
7635 :
7636 32768 : CHECK_FOR_INTERRUPTS();
7637 :
7638 : /* An unlocked precheck should be safe and saves some cycles. */
7639 32768 : if ((buf_state & BM_VALID) == 0 ||
7640 54 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
7641 32768 : continue;
7642 :
7643 : /* Make sure we can pin the buffer. */
7644 0 : ResourceOwnerEnlarge(CurrentResourceOwner);
7645 0 : ReservePrivateRefCountEntry();
7646 :
7647 0 : buf_state = LockBufHdr(desc);
7648 :
7649 : /* recheck, could have changed without the lock */
7650 0 : if ((buf_state & BM_VALID) == 0 ||
7651 0 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
7652 : {
7653 0 : UnlockBufHdr(desc);
7654 0 : continue;
7655 : }
7656 :
7657 0 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
7658 0 : (*buffers_evicted)++;
7659 : else
7660 0 : (*buffers_skipped)++;
7661 :
7662 0 : if (buffer_flushed)
7663 0 : (*buffers_flushed)++;
7664 : }
7665 2 : }
7666 :
7667 : /*
7668 : * Helper function to mark unpinned buffer dirty whose buffer header lock is
7669 : * already acquired.
7670 : */
7671 : static bool
7672 72 : MarkDirtyUnpinnedBufferInternal(Buffer buf, BufferDesc *desc,
7673 : bool *buffer_already_dirty)
7674 : {
7675 : uint64 buf_state;
7676 72 : bool result = false;
7677 :
7678 72 : *buffer_already_dirty = false;
7679 :
7680 72 : buf_state = pg_atomic_read_u64(&(desc->state));
7681 : Assert(buf_state & BM_LOCKED);
7682 :
7683 72 : if ((buf_state & BM_VALID) == 0)
7684 : {
7685 2 : UnlockBufHdr(desc);
7686 2 : return false;
7687 : }
7688 :
7689 : /* Check that it's not pinned already. */
7690 70 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
7691 : {
7692 0 : UnlockBufHdr(desc);
7693 0 : return false;
7694 : }
7695 :
7696 : /* Pin the buffer and then release the buffer spinlock */
7697 70 : PinBuffer_Locked(desc);
7698 :
7699 : /* If it was not already dirty, mark it as dirty. */
7700 70 : if (!(buf_state & BM_DIRTY))
7701 : {
7702 34 : BufferLockAcquire(buf, desc, BUFFER_LOCK_EXCLUSIVE);
7703 34 : MarkBufferDirty(buf);
7704 34 : result = true;
7705 34 : BufferLockUnlock(buf, desc);
7706 : }
7707 : else
7708 36 : *buffer_already_dirty = true;
7709 :
7710 70 : UnpinBuffer(desc);
7711 :
7712 70 : return result;
7713 : }
7714 :
7715 : /*
7716 : * Try to mark the provided shared buffer as dirty.
7717 : *
7718 : * This function is intended for testing/development use only!
7719 : *
7720 : * Same as EvictUnpinnedBuffer() but with MarkBufferDirty() call inside.
7721 : *
7722 : * The buffer_already_dirty parameter is mandatory and indicate if the buffer
7723 : * could not be dirtied because it is already dirty.
7724 : *
7725 : * Returns true if the buffer has successfully been marked as dirty.
7726 : */
7727 : bool
7728 2 : MarkDirtyUnpinnedBuffer(Buffer buf, bool *buffer_already_dirty)
7729 : {
7730 : BufferDesc *desc;
7731 2 : bool buffer_dirtied = false;
7732 :
7733 : Assert(!BufferIsLocal(buf));
7734 :
7735 : /* Make sure we can pin the buffer. */
7736 2 : ResourceOwnerEnlarge(CurrentResourceOwner);
7737 2 : ReservePrivateRefCountEntry();
7738 :
7739 2 : desc = GetBufferDescriptor(buf - 1);
7740 2 : LockBufHdr(desc);
7741 :
7742 2 : buffer_dirtied = MarkDirtyUnpinnedBufferInternal(buf, desc, buffer_already_dirty);
7743 : /* Both can not be true at the same time */
7744 : Assert(!(buffer_dirtied && *buffer_already_dirty));
7745 :
7746 2 : return buffer_dirtied;
7747 : }
7748 :
7749 : /*
7750 : * Try to mark all the shared buffers containing provided relation's pages as
7751 : * dirty.
7752 : *
7753 : * This function is intended for testing/development use only! See
7754 : * MarkDirtyUnpinnedBuffer().
7755 : *
7756 : * The buffers_* parameters are mandatory and indicate the total count of
7757 : * buffers that:
7758 : * - buffers_dirtied - were dirtied
7759 : * - buffers_already_dirty - were already dirty
7760 : * - buffers_skipped - could not be dirtied because of a reason different
7761 : * than a buffer being already dirty.
7762 : */
7763 : void
7764 2 : MarkDirtyRelUnpinnedBuffers(Relation rel,
7765 : int32 *buffers_dirtied,
7766 : int32 *buffers_already_dirty,
7767 : int32 *buffers_skipped)
7768 : {
7769 : Assert(!RelationUsesLocalBuffers(rel));
7770 :
7771 2 : *buffers_dirtied = 0;
7772 2 : *buffers_already_dirty = 0;
7773 2 : *buffers_skipped = 0;
7774 :
7775 32770 : for (int buf = 1; buf <= NBuffers; buf++)
7776 : {
7777 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
7778 32768 : uint64 buf_state = pg_atomic_read_u64(&(desc->state));
7779 : bool buffer_already_dirty;
7780 :
7781 32768 : CHECK_FOR_INTERRUPTS();
7782 :
7783 : /* An unlocked precheck should be safe and saves some cycles. */
7784 32768 : if ((buf_state & BM_VALID) == 0 ||
7785 54 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
7786 32768 : continue;
7787 :
7788 : /* Make sure we can pin the buffer. */
7789 0 : ResourceOwnerEnlarge(CurrentResourceOwner);
7790 0 : ReservePrivateRefCountEntry();
7791 :
7792 0 : buf_state = LockBufHdr(desc);
7793 :
7794 : /* recheck, could have changed without the lock */
7795 0 : if ((buf_state & BM_VALID) == 0 ||
7796 0 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
7797 : {
7798 0 : UnlockBufHdr(desc);
7799 0 : continue;
7800 : }
7801 :
7802 0 : if (MarkDirtyUnpinnedBufferInternal(buf, desc, &buffer_already_dirty))
7803 0 : (*buffers_dirtied)++;
7804 0 : else if (buffer_already_dirty)
7805 0 : (*buffers_already_dirty)++;
7806 : else
7807 0 : (*buffers_skipped)++;
7808 : }
7809 2 : }
7810 :
7811 : /*
7812 : * Try to mark all the shared buffers as dirty.
7813 : *
7814 : * This function is intended for testing/development use only! See
7815 : * MarkDirtyUnpinnedBuffer().
7816 : *
7817 : * See MarkDirtyRelUnpinnedBuffers() above for details about the buffers_*
7818 : * parameters.
7819 : */
7820 : void
7821 2 : MarkDirtyAllUnpinnedBuffers(int32 *buffers_dirtied,
7822 : int32 *buffers_already_dirty,
7823 : int32 *buffers_skipped)
7824 : {
7825 2 : *buffers_dirtied = 0;
7826 2 : *buffers_already_dirty = 0;
7827 2 : *buffers_skipped = 0;
7828 :
7829 32770 : for (int buf = 1; buf <= NBuffers; buf++)
7830 : {
7831 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
7832 : uint64 buf_state;
7833 : bool buffer_already_dirty;
7834 :
7835 32768 : CHECK_FOR_INTERRUPTS();
7836 :
7837 32768 : buf_state = pg_atomic_read_u64(&desc->state);
7838 32768 : if (!(buf_state & BM_VALID))
7839 32698 : continue;
7840 :
7841 70 : ResourceOwnerEnlarge(CurrentResourceOwner);
7842 70 : ReservePrivateRefCountEntry();
7843 :
7844 70 : LockBufHdr(desc);
7845 :
7846 70 : if (MarkDirtyUnpinnedBufferInternal(buf, desc, &buffer_already_dirty))
7847 34 : (*buffers_dirtied)++;
7848 36 : else if (buffer_already_dirty)
7849 36 : (*buffers_already_dirty)++;
7850 : else
7851 0 : (*buffers_skipped)++;
7852 : }
7853 2 : }
7854 :
7855 : /*
7856 : * Generic implementation of the AIO handle staging callback for readv/writev
7857 : * on local/shared buffers.
7858 : *
7859 : * Each readv/writev can target multiple buffers. The buffers have already
7860 : * been registered with the IO handle.
7861 : *
7862 : * To make the IO ready for execution ("staging"), we need to ensure that the
7863 : * targeted buffers are in an appropriate state while the IO is ongoing. For
7864 : * that the AIO subsystem needs to have its own buffer pin, otherwise an error
7865 : * in this backend could lead to this backend's buffer pin being released as
7866 : * part of error handling, which in turn could lead to the buffer being
7867 : * replaced while IO is ongoing.
7868 : */
7869 : static pg_attribute_always_inline void
7870 2557840 : buffer_stage_common(PgAioHandle *ioh, bool is_write, bool is_temp)
7871 : {
7872 : uint64 *io_data;
7873 : uint8 handle_data_len;
7874 : PgAioWaitRef io_ref;
7875 2557840 : BufferTag first PG_USED_FOR_ASSERTS_ONLY = {0};
7876 :
7877 2557840 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
7878 :
7879 2557840 : pgaio_io_get_wref(ioh, &io_ref);
7880 :
7881 : /* iterate over all buffers affected by the vectored readv/writev */
7882 5460944 : for (int i = 0; i < handle_data_len; i++)
7883 : {
7884 2903104 : Buffer buffer = (Buffer) io_data[i];
7885 2903104 : BufferDesc *buf_hdr = is_temp ?
7886 16818 : GetLocalBufferDescriptor(-buffer - 1)
7887 2903104 : : GetBufferDescriptor(buffer - 1);
7888 : uint64 buf_state;
7889 :
7890 : /*
7891 : * Check that all the buffers are actually ones that could conceivably
7892 : * be done in one IO, i.e. are sequential. This is the last
7893 : * buffer-aware code before IO is actually executed and confusion
7894 : * about which buffers are targeted by IO can be hard to debug, making
7895 : * it worth doing extra-paranoid checks.
7896 : */
7897 2903104 : if (i == 0)
7898 2557840 : first = buf_hdr->tag;
7899 : else
7900 : {
7901 : Assert(buf_hdr->tag.relNumber == first.relNumber);
7902 : Assert(buf_hdr->tag.blockNum == first.blockNum + i);
7903 : }
7904 :
7905 2903104 : if (is_temp)
7906 16818 : buf_state = pg_atomic_read_u64(&buf_hdr->state);
7907 : else
7908 2886286 : buf_state = LockBufHdr(buf_hdr);
7909 :
7910 : /* verify the buffer is in the expected state */
7911 : Assert(buf_state & BM_TAG_VALID);
7912 : if (is_write)
7913 : {
7914 : Assert(buf_state & BM_VALID);
7915 : Assert(buf_state & BM_DIRTY);
7916 : }
7917 : else
7918 : {
7919 : Assert(!(buf_state & BM_VALID));
7920 : Assert(!(buf_state & BM_DIRTY));
7921 : }
7922 :
7923 : /* temp buffers don't use BM_IO_IN_PROGRESS */
7924 2903104 : if (!is_temp)
7925 : Assert(buf_state & BM_IO_IN_PROGRESS);
7926 :
7927 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) >= 1);
7928 :
7929 : /*
7930 : * Reflect that the buffer is now owned by the AIO subsystem.
7931 : *
7932 : * For local buffers: This can't be done just via LocalRefCount, as
7933 : * one might initially think, as this backend could error out while
7934 : * AIO is still in progress, releasing all the pins by the backend
7935 : * itself.
7936 : *
7937 : * This pin is released again in TerminateBufferIO().
7938 : */
7939 2903104 : buf_hdr->io_wref = io_ref;
7940 :
7941 2903104 : if (is_temp)
7942 : {
7943 16818 : buf_state += BUF_REFCOUNT_ONE;
7944 16818 : pg_atomic_unlocked_write_u64(&buf_hdr->state, buf_state);
7945 : }
7946 : else
7947 2886286 : UnlockBufHdrExt(buf_hdr, buf_state, 0, 0, 1);
7948 :
7949 : /*
7950 : * Ensure the content lock that prevents buffer modifications while
7951 : * the buffer is being written out is not released early due to an
7952 : * error.
7953 : */
7954 2903104 : if (is_write && !is_temp)
7955 : {
7956 : Assert(BufferLockHeldByMe(buf_hdr));
7957 :
7958 : /*
7959 : * Lock is now owned by AIO subsystem.
7960 : */
7961 0 : BufferLockDisown(buffer, buf_hdr);
7962 : }
7963 :
7964 : /*
7965 : * Stop tracking this buffer via the resowner - the AIO system now
7966 : * keeps track.
7967 : */
7968 2903104 : if (!is_temp)
7969 2886286 : ResourceOwnerForgetBufferIO(CurrentResourceOwner, buffer);
7970 : }
7971 2557840 : }
7972 :
7973 : /*
7974 : * Decode readv errors as encoded by buffer_readv_encode_error().
7975 : */
7976 : static inline void
7977 698 : buffer_readv_decode_error(PgAioResult result,
7978 : bool *zeroed_any,
7979 : bool *ignored_any,
7980 : uint8 *zeroed_or_error_count,
7981 : uint8 *checkfail_count,
7982 : uint8 *first_off)
7983 : {
7984 698 : uint32 rem_error = result.error_data;
7985 :
7986 : /* see static asserts in buffer_readv_encode_error */
7987 : #define READV_COUNT_BITS 7
7988 : #define READV_COUNT_MASK ((1 << READV_COUNT_BITS) - 1)
7989 :
7990 698 : *zeroed_any = rem_error & 1;
7991 698 : rem_error >>= 1;
7992 :
7993 698 : *ignored_any = rem_error & 1;
7994 698 : rem_error >>= 1;
7995 :
7996 698 : *zeroed_or_error_count = rem_error & READV_COUNT_MASK;
7997 698 : rem_error >>= READV_COUNT_BITS;
7998 :
7999 698 : *checkfail_count = rem_error & READV_COUNT_MASK;
8000 698 : rem_error >>= READV_COUNT_BITS;
8001 :
8002 698 : *first_off = rem_error & READV_COUNT_MASK;
8003 698 : rem_error >>= READV_COUNT_BITS;
8004 698 : }
8005 :
8006 : /*
8007 : * Helper to encode errors for buffer_readv_complete()
8008 : *
8009 : * Errors are encoded as follows:
8010 : * - bit 0 indicates whether any page was zeroed (1) or not (0)
8011 : * - bit 1 indicates whether any checksum failure was ignored (1) or not (0)
8012 : * - next READV_COUNT_BITS bits indicate the number of errored or zeroed pages
8013 : * - next READV_COUNT_BITS bits indicate the number of checksum failures
8014 : * - next READV_COUNT_BITS bits indicate the first offset of the first page
8015 : * that was errored or zeroed or, if no errors/zeroes, the first ignored
8016 : * checksum
8017 : */
8018 : static inline void
8019 384 : buffer_readv_encode_error(PgAioResult *result,
8020 : bool is_temp,
8021 : bool zeroed_any,
8022 : bool ignored_any,
8023 : uint8 error_count,
8024 : uint8 zeroed_count,
8025 : uint8 checkfail_count,
8026 : uint8 first_error_off,
8027 : uint8 first_zeroed_off,
8028 : uint8 first_ignored_off)
8029 : {
8030 :
8031 384 : uint8 shift = 0;
8032 384 : uint8 zeroed_or_error_count =
8033 : error_count > 0 ? error_count : zeroed_count;
8034 : uint8 first_off;
8035 :
8036 : StaticAssertDecl(PG_IOV_MAX <= 1 << READV_COUNT_BITS,
8037 : "PG_IOV_MAX is bigger than reserved space for error data");
8038 : StaticAssertDecl((1 + 1 + 3 * READV_COUNT_BITS) <= PGAIO_RESULT_ERROR_BITS,
8039 : "PGAIO_RESULT_ERROR_BITS is insufficient for buffer_readv");
8040 :
8041 : /*
8042 : * We only have space to encode one offset - but luckily that's good
8043 : * enough. If there is an error, the error is the interesting offset, same
8044 : * with a zeroed buffer vs an ignored buffer.
8045 : */
8046 384 : if (error_count > 0)
8047 188 : first_off = first_error_off;
8048 196 : else if (zeroed_count > 0)
8049 160 : first_off = first_zeroed_off;
8050 : else
8051 36 : first_off = first_ignored_off;
8052 :
8053 : Assert(!zeroed_any || error_count == 0);
8054 :
8055 384 : result->error_data = 0;
8056 :
8057 384 : result->error_data |= zeroed_any << shift;
8058 384 : shift += 1;
8059 :
8060 384 : result->error_data |= ignored_any << shift;
8061 384 : shift += 1;
8062 :
8063 384 : result->error_data |= ((uint32) zeroed_or_error_count) << shift;
8064 384 : shift += READV_COUNT_BITS;
8065 :
8066 384 : result->error_data |= ((uint32) checkfail_count) << shift;
8067 384 : shift += READV_COUNT_BITS;
8068 :
8069 384 : result->error_data |= ((uint32) first_off) << shift;
8070 384 : shift += READV_COUNT_BITS;
8071 :
8072 384 : result->id = is_temp ? PGAIO_HCB_LOCAL_BUFFER_READV :
8073 : PGAIO_HCB_SHARED_BUFFER_READV;
8074 :
8075 384 : if (error_count > 0)
8076 188 : result->status = PGAIO_RS_ERROR;
8077 : else
8078 196 : result->status = PGAIO_RS_WARNING;
8079 :
8080 : /*
8081 : * The encoding is complicated enough to warrant cross-checking it against
8082 : * the decode function.
8083 : */
8084 : #ifdef USE_ASSERT_CHECKING
8085 : {
8086 : bool zeroed_any_2,
8087 : ignored_any_2;
8088 : uint8 zeroed_or_error_count_2,
8089 : checkfail_count_2,
8090 : first_off_2;
8091 :
8092 : buffer_readv_decode_error(*result,
8093 : &zeroed_any_2, &ignored_any_2,
8094 : &zeroed_or_error_count_2,
8095 : &checkfail_count_2,
8096 : &first_off_2);
8097 : Assert(zeroed_any == zeroed_any_2);
8098 : Assert(ignored_any == ignored_any_2);
8099 : Assert(zeroed_or_error_count == zeroed_or_error_count_2);
8100 : Assert(checkfail_count == checkfail_count_2);
8101 : Assert(first_off == first_off_2);
8102 : }
8103 : #endif
8104 :
8105 : #undef READV_COUNT_BITS
8106 : #undef READV_COUNT_MASK
8107 384 : }
8108 :
8109 : /*
8110 : * Helper for AIO readv completion callbacks, supporting both shared and temp
8111 : * buffers. Gets called once for each buffer in a multi-page read.
8112 : */
8113 : static pg_attribute_always_inline void
8114 2623324 : buffer_readv_complete_one(PgAioTargetData *td, uint8 buf_off, Buffer buffer,
8115 : uint8 flags, bool failed, bool is_temp,
8116 : bool *buffer_invalid,
8117 : bool *failed_checksum,
8118 : bool *ignored_checksum,
8119 : bool *zeroed_buffer)
8120 : {
8121 2623324 : BufferDesc *buf_hdr = is_temp ?
8122 16818 : GetLocalBufferDescriptor(-buffer - 1)
8123 2623324 : : GetBufferDescriptor(buffer - 1);
8124 2623324 : BufferTag tag = buf_hdr->tag;
8125 2623324 : char *bufdata = BufferGetBlock(buffer);
8126 : uint64 set_flag_bits;
8127 : int piv_flags;
8128 :
8129 : /* check that the buffer is in the expected state for a read */
8130 : #ifdef USE_ASSERT_CHECKING
8131 : {
8132 : uint64 buf_state = pg_atomic_read_u64(&buf_hdr->state);
8133 :
8134 : Assert(buf_state & BM_TAG_VALID);
8135 : Assert(!(buf_state & BM_VALID));
8136 : /* temp buffers don't use BM_IO_IN_PROGRESS */
8137 : if (!is_temp)
8138 : Assert(buf_state & BM_IO_IN_PROGRESS);
8139 : Assert(!(buf_state & BM_DIRTY));
8140 : }
8141 : #endif
8142 :
8143 2623324 : *buffer_invalid = false;
8144 2623324 : *failed_checksum = false;
8145 2623324 : *ignored_checksum = false;
8146 2623324 : *zeroed_buffer = false;
8147 :
8148 : /*
8149 : * We ask PageIsVerified() to only log the message about checksum errors,
8150 : * as the completion might be run in any backend (or IO workers). We will
8151 : * report checksum errors in buffer_readv_report().
8152 : */
8153 2623324 : piv_flags = PIV_LOG_LOG;
8154 :
8155 : /* the local zero_damaged_pages may differ from the definer's */
8156 2623324 : if (flags & READ_BUFFERS_IGNORE_CHECKSUM_FAILURES)
8157 76 : piv_flags |= PIV_IGNORE_CHECKSUM_FAILURE;
8158 :
8159 : /* Check for garbage data. */
8160 2623324 : if (!failed)
8161 : {
8162 : /*
8163 : * If the buffer is not currently pinned by this backend, e.g. because
8164 : * we're completing this IO after an error, the buffer data will have
8165 : * been marked as inaccessible when the buffer was unpinned. The AIO
8166 : * subsystem holds a pin, but that doesn't prevent the buffer from
8167 : * having been marked as inaccessible. The completion might also be
8168 : * executed in a different process.
8169 : */
8170 : #ifdef USE_VALGRIND
8171 : if (!BufferIsPinned(buffer))
8172 : VALGRIND_MAKE_MEM_DEFINED(bufdata, BLCKSZ);
8173 : #endif
8174 :
8175 2623266 : if (!PageIsVerified((Page) bufdata, tag.blockNum, piv_flags,
8176 : failed_checksum))
8177 : {
8178 192 : if (flags & READ_BUFFERS_ZERO_ON_ERROR)
8179 : {
8180 92 : memset(bufdata, 0, BLCKSZ);
8181 92 : *zeroed_buffer = true;
8182 : }
8183 : else
8184 : {
8185 100 : *buffer_invalid = true;
8186 : /* mark buffer as having failed */
8187 100 : failed = true;
8188 : }
8189 : }
8190 2623074 : else if (*failed_checksum)
8191 24 : *ignored_checksum = true;
8192 :
8193 : /* undo what we did above */
8194 : #ifdef USE_VALGRIND
8195 : if (!BufferIsPinned(buffer))
8196 : VALGRIND_MAKE_MEM_NOACCESS(bufdata, BLCKSZ);
8197 : #endif
8198 :
8199 : /*
8200 : * Immediately log a message about the invalid page, but only to the
8201 : * server log. The reason to do so immediately is that this may be
8202 : * executed in a different backend than the one that originated the
8203 : * request. The reason to do so immediately is that the originator
8204 : * might not process the query result immediately (because it is busy
8205 : * doing another part of query processing) or at all (e.g. if it was
8206 : * cancelled or errored out due to another IO also failing). The
8207 : * definer of the IO will emit an ERROR or WARNING when processing the
8208 : * IO's results
8209 : *
8210 : * To avoid duplicating the code to emit these log messages, we reuse
8211 : * buffer_readv_report().
8212 : */
8213 2623266 : if (*buffer_invalid || *failed_checksum || *zeroed_buffer)
8214 : {
8215 216 : PgAioResult result_one = {0};
8216 :
8217 216 : buffer_readv_encode_error(&result_one, is_temp,
8218 216 : *zeroed_buffer,
8219 216 : *ignored_checksum,
8220 216 : *buffer_invalid,
8221 216 : *zeroed_buffer ? 1 : 0,
8222 216 : *failed_checksum ? 1 : 0,
8223 : buf_off, buf_off, buf_off);
8224 216 : pgaio_result_report(result_one, td, LOG_SERVER_ONLY);
8225 : }
8226 : }
8227 :
8228 : /* Terminate I/O and set BM_VALID. */
8229 2623324 : set_flag_bits = failed ? BM_IO_ERROR : BM_VALID;
8230 2623324 : if (is_temp)
8231 16818 : TerminateLocalBufferIO(buf_hdr, false, set_flag_bits, true);
8232 : else
8233 2606506 : TerminateBufferIO(buf_hdr, false, set_flag_bits, false, true);
8234 :
8235 : /*
8236 : * Call the BUFFER_READ_DONE tracepoint in the callback, even though the
8237 : * callback may not be executed in the same backend that called
8238 : * BUFFER_READ_START. The alternative would be to defer calling the
8239 : * tracepoint to a later point (e.g. the local completion callback for
8240 : * shared buffer reads), which seems even less helpful.
8241 : */
8242 : TRACE_POSTGRESQL_BUFFER_READ_DONE(tag.forkNum,
8243 : tag.blockNum,
8244 : tag.spcOid,
8245 : tag.dbOid,
8246 : tag.relNumber,
8247 : is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
8248 : false);
8249 2623324 : }
8250 :
8251 : /*
8252 : * Perform completion handling of a single AIO read. This read may cover
8253 : * multiple blocks / buffers.
8254 : *
8255 : * Shared between shared and local buffers, to reduce code duplication.
8256 : */
8257 : static pg_attribute_always_inline PgAioResult
8258 2346540 : buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
8259 : uint8 cb_data, bool is_temp)
8260 : {
8261 2346540 : PgAioResult result = prior_result;
8262 2346540 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
8263 2346540 : uint8 first_error_off = 0;
8264 2346540 : uint8 first_zeroed_off = 0;
8265 2346540 : uint8 first_ignored_off = 0;
8266 2346540 : uint8 error_count = 0;
8267 2346540 : uint8 zeroed_count = 0;
8268 2346540 : uint8 ignored_count = 0;
8269 2346540 : uint8 checkfail_count = 0;
8270 : uint64 *io_data;
8271 : uint8 handle_data_len;
8272 :
8273 : if (is_temp)
8274 : {
8275 : Assert(td->smgr.is_temp);
8276 : Assert(pgaio_io_get_owner(ioh) == MyProcNumber);
8277 : }
8278 : else
8279 : Assert(!td->smgr.is_temp);
8280 :
8281 : /*
8282 : * Iterate over all the buffers affected by this IO and call the
8283 : * per-buffer completion function for each buffer.
8284 : */
8285 2346540 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
8286 4969864 : for (uint8 buf_off = 0; buf_off < handle_data_len; buf_off++)
8287 : {
8288 2623324 : Buffer buf = io_data[buf_off];
8289 : bool failed;
8290 2623324 : bool failed_verification = false;
8291 2623324 : bool failed_checksum = false;
8292 2623324 : bool zeroed_buffer = false;
8293 2623324 : bool ignored_checksum = false;
8294 :
8295 : Assert(BufferIsValid(buf));
8296 :
8297 : /*
8298 : * If the entire I/O failed on a lower-level, each buffer needs to be
8299 : * marked as failed. In case of a partial read, the first few buffers
8300 : * may be ok.
8301 : */
8302 2623324 : failed =
8303 2623324 : prior_result.status == PGAIO_RS_ERROR
8304 2623324 : || prior_result.result <= buf_off;
8305 :
8306 2623324 : buffer_readv_complete_one(td, buf_off, buf, cb_data, failed, is_temp,
8307 : &failed_verification,
8308 : &failed_checksum,
8309 : &ignored_checksum,
8310 : &zeroed_buffer);
8311 :
8312 : /*
8313 : * Track information about the number of different kinds of error
8314 : * conditions across all pages, as there can be multiple pages failing
8315 : * verification as part of one IO.
8316 : */
8317 2623324 : if (failed_verification && !zeroed_buffer && error_count++ == 0)
8318 88 : first_error_off = buf_off;
8319 2623324 : if (zeroed_buffer && zeroed_count++ == 0)
8320 68 : first_zeroed_off = buf_off;
8321 2623324 : if (ignored_checksum && ignored_count++ == 0)
8322 20 : first_ignored_off = buf_off;
8323 2623324 : if (failed_checksum)
8324 64 : checkfail_count++;
8325 : }
8326 :
8327 : /*
8328 : * If the smgr read succeeded [partially] and page verification failed for
8329 : * some of the pages, adjust the IO's result state appropriately.
8330 : */
8331 2346540 : if (prior_result.status != PGAIO_RS_ERROR &&
8332 2346434 : (error_count > 0 || ignored_count > 0 || zeroed_count > 0))
8333 : {
8334 168 : buffer_readv_encode_error(&result, is_temp,
8335 : zeroed_count > 0, ignored_count > 0,
8336 : error_count, zeroed_count, checkfail_count,
8337 : first_error_off, first_zeroed_off,
8338 : first_ignored_off);
8339 168 : pgaio_result_report(result, td, DEBUG1);
8340 : }
8341 :
8342 : /*
8343 : * For shared relations this reporting is done in
8344 : * shared_buffer_readv_complete_local().
8345 : */
8346 2346540 : if (is_temp && checkfail_count > 0)
8347 4 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
8348 : checkfail_count);
8349 :
8350 2346540 : return result;
8351 : }
8352 :
8353 : /*
8354 : * AIO error reporting callback for aio_shared_buffer_readv_cb and
8355 : * aio_local_buffer_readv_cb.
8356 : *
8357 : * The error is encoded / decoded in buffer_readv_encode_error() /
8358 : * buffer_readv_decode_error().
8359 : */
8360 : static void
8361 544 : buffer_readv_report(PgAioResult result, const PgAioTargetData *td,
8362 : int elevel)
8363 : {
8364 544 : int nblocks = td->smgr.nblocks;
8365 544 : BlockNumber first = td->smgr.blockNum;
8366 544 : BlockNumber last = first + nblocks - 1;
8367 544 : ProcNumber errProc =
8368 544 : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER;
8369 : RelPathStr rpath =
8370 544 : relpathbackend(td->smgr.rlocator, errProc, td->smgr.forkNum);
8371 : bool zeroed_any,
8372 : ignored_any;
8373 : uint8 zeroed_or_error_count,
8374 : checkfail_count,
8375 : first_off;
8376 : uint8 affected_count;
8377 : const char *msg_one,
8378 : *msg_mult,
8379 : *det_mult,
8380 : *hint_mult;
8381 :
8382 544 : buffer_readv_decode_error(result, &zeroed_any, &ignored_any,
8383 : &zeroed_or_error_count,
8384 : &checkfail_count,
8385 : &first_off);
8386 :
8387 : /*
8388 : * Treat a read that had both zeroed buffers *and* ignored checksums as a
8389 : * special case, it's too irregular to be emitted the same way as the
8390 : * other cases.
8391 : */
8392 544 : if (zeroed_any && ignored_any)
8393 : {
8394 : Assert(zeroed_any && ignored_any);
8395 : Assert(nblocks > 1); /* same block can't be both zeroed and ignored */
8396 : Assert(result.status != PGAIO_RS_ERROR);
8397 8 : affected_count = zeroed_or_error_count;
8398 :
8399 8 : ereport(elevel,
8400 : errcode(ERRCODE_DATA_CORRUPTED),
8401 : errmsg("zeroing %u page(s) and ignoring %u checksum failure(s) among blocks %u..%u of relation \"%s\"",
8402 : affected_count, checkfail_count, first, last, rpath.str),
8403 : affected_count > 1 ?
8404 : errdetail("Block %u held the first zeroed page.",
8405 : first + first_off) : 0,
8406 : errhint_plural("See server log for details about the other %d invalid block.",
8407 : "See server log for details about the other %d invalid blocks.",
8408 : affected_count + checkfail_count - 1,
8409 : affected_count + checkfail_count - 1));
8410 8 : return;
8411 : }
8412 :
8413 : /*
8414 : * The other messages are highly repetitive. To avoid duplicating a long
8415 : * and complicated ereport(), gather the translated format strings
8416 : * separately and then do one common ereport.
8417 : */
8418 536 : if (result.status == PGAIO_RS_ERROR)
8419 : {
8420 : Assert(!zeroed_any); /* can't have invalid pages when zeroing them */
8421 272 : affected_count = zeroed_or_error_count;
8422 272 : msg_one = _("invalid page in block %u of relation \"%s\"");
8423 272 : msg_mult = _("%u invalid pages among blocks %u..%u of relation \"%s\"");
8424 272 : det_mult = _("Block %u held the first invalid page.");
8425 272 : hint_mult = _("See server log for the other %u invalid block(s).");
8426 : }
8427 264 : else if (zeroed_any && !ignored_any)
8428 : {
8429 216 : affected_count = zeroed_or_error_count;
8430 216 : msg_one = _("invalid page in block %u of relation \"%s\"; zeroing out page");
8431 216 : msg_mult = _("zeroing out %u invalid pages among blocks %u..%u of relation \"%s\"");
8432 216 : det_mult = _("Block %u held the first zeroed page.");
8433 216 : hint_mult = _("See server log for the other %u zeroed block(s).");
8434 : }
8435 48 : else if (!zeroed_any && ignored_any)
8436 : {
8437 48 : affected_count = checkfail_count;
8438 48 : msg_one = _("ignoring checksum failure in block %u of relation \"%s\"");
8439 48 : msg_mult = _("ignoring %u checksum failures among blocks %u..%u of relation \"%s\"");
8440 48 : det_mult = _("Block %u held the first ignored page.");
8441 48 : hint_mult = _("See server log for the other %u ignored block(s).");
8442 : }
8443 : else
8444 0 : pg_unreachable();
8445 :
8446 536 : ereport(elevel,
8447 : errcode(ERRCODE_DATA_CORRUPTED),
8448 : affected_count == 1 ?
8449 : errmsg_internal(msg_one, first + first_off, rpath.str) :
8450 : errmsg_internal(msg_mult, affected_count, first, last, rpath.str),
8451 : affected_count > 1 ? errdetail_internal(det_mult, first + first_off) : 0,
8452 : affected_count > 1 ? errhint_internal(hint_mult, affected_count - 1) : 0);
8453 : }
8454 :
8455 : static void
8456 2554228 : shared_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
8457 : {
8458 2554228 : buffer_stage_common(ioh, false, false);
8459 2554228 : }
8460 :
8461 : static PgAioResult
8462 2342928 : shared_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
8463 : uint8 cb_data)
8464 : {
8465 2342928 : return buffer_readv_complete(ioh, prior_result, cb_data, false);
8466 : }
8467 :
8468 : /*
8469 : * We need a backend-local completion callback for shared buffers, to be able
8470 : * to report checksum errors correctly. Unfortunately that can only safely
8471 : * happen if the reporting backend has previously called
8472 : * pgstat_prepare_report_checksum_failure(), which we can only guarantee in
8473 : * the backend that started the IO. Hence this callback.
8474 : */
8475 : static PgAioResult
8476 2554228 : shared_buffer_readv_complete_local(PgAioHandle *ioh, PgAioResult prior_result,
8477 : uint8 cb_data)
8478 : {
8479 : bool zeroed_any,
8480 : ignored_any;
8481 : uint8 zeroed_or_error_count,
8482 : checkfail_count,
8483 : first_off;
8484 :
8485 2554228 : if (prior_result.status == PGAIO_RS_OK)
8486 2554074 : return prior_result;
8487 :
8488 154 : buffer_readv_decode_error(prior_result,
8489 : &zeroed_any,
8490 : &ignored_any,
8491 : &zeroed_or_error_count,
8492 : &checkfail_count,
8493 : &first_off);
8494 :
8495 154 : if (checkfail_count)
8496 : {
8497 48 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
8498 :
8499 48 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
8500 : checkfail_count);
8501 : }
8502 :
8503 154 : return prior_result;
8504 : }
8505 :
8506 : static void
8507 3612 : local_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
8508 : {
8509 3612 : buffer_stage_common(ioh, false, true);
8510 3612 : }
8511 :
8512 : static PgAioResult
8513 3612 : local_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
8514 : uint8 cb_data)
8515 : {
8516 3612 : return buffer_readv_complete(ioh, prior_result, cb_data, true);
8517 : }
8518 :
8519 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
8520 : const PgAioHandleCallbacks aio_shared_buffer_readv_cb = {
8521 : .stage = shared_buffer_readv_stage,
8522 : .complete_shared = shared_buffer_readv_complete,
8523 : /* need a local callback to report checksum failures */
8524 : .complete_local = shared_buffer_readv_complete_local,
8525 : .report = buffer_readv_report,
8526 : };
8527 :
8528 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
8529 : const PgAioHandleCallbacks aio_local_buffer_readv_cb = {
8530 : .stage = local_buffer_readv_stage,
8531 :
8532 : /*
8533 : * Note that this, in contrast to the shared_buffers case, uses
8534 : * complete_local, as only the issuing backend has access to the required
8535 : * datastructures. This is important in case the IO completion may be
8536 : * consumed incidentally by another backend.
8537 : */
8538 : .complete_local = local_buffer_readv_complete,
8539 : .report = buffer_readv_report,
8540 : };
|
