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 "common/hashfn.h"
49 : #include "executor/instrument.h"
50 : #include "lib/binaryheap.h"
51 : #include "miscadmin.h"
52 : #include "pg_trace.h"
53 : #include "pgstat.h"
54 : #include "postmaster/bgwriter.h"
55 : #include "storage/aio.h"
56 : #include "storage/buf_internals.h"
57 : #include "storage/bufmgr.h"
58 : #include "storage/fd.h"
59 : #include "storage/ipc.h"
60 : #include "storage/lmgr.h"
61 : #include "storage/proc.h"
62 : #include "storage/proclist.h"
63 : #include "storage/procsignal.h"
64 : #include "storage/read_stream.h"
65 : #include "storage/smgr.h"
66 : #include "storage/standby.h"
67 : #include "utils/memdebug.h"
68 : #include "utils/ps_status.h"
69 : #include "utils/rel.h"
70 : #include "utils/resowner.h"
71 : #include "utils/timestamp.h"
72 : #include "utils/wait_event.h"
73 :
74 :
75 : /* Note: these two macros only work on shared buffers, not local ones! */
76 : #define BufHdrGetBlock(bufHdr) ((Block) (BufferBlocks + ((Size) (bufHdr)->buf_id) * BLCKSZ))
77 : #define BufferGetLSN(bufHdr) (PageGetLSN(BufHdrGetBlock(bufHdr)))
78 :
79 : /* Note: this macro only works on local buffers, not shared ones! */
80 : #define LocalBufHdrGetBlock(bufHdr) \
81 : LocalBufferBlockPointers[-((bufHdr)->buf_id + 2)]
82 :
83 : /* Bits in SyncOneBuffer's return value */
84 : #define BUF_WRITTEN 0x01
85 : #define BUF_REUSABLE 0x02
86 :
87 : #define RELS_BSEARCH_THRESHOLD 20
88 :
89 : /*
90 : * This is the size (in the number of blocks) above which we scan the
91 : * entire buffer pool to remove the buffers for all the pages of relation
92 : * being dropped. For the relations with size below this threshold, we find
93 : * the buffers by doing lookups in BufMapping table.
94 : */
95 : #define BUF_DROP_FULL_SCAN_THRESHOLD (uint64) (NBuffers / 32)
96 :
97 : /*
98 : * This is separated out from PrivateRefCountEntry to allow for copying all
99 : * the data members via struct assignment.
100 : */
101 : typedef struct PrivateRefCountData
102 : {
103 : /*
104 : * How many times has the buffer been pinned by this backend.
105 : */
106 : int32 refcount;
107 :
108 : /*
109 : * Is the buffer locked by this backend? BUFFER_LOCK_UNLOCK indicates that
110 : * the buffer is not locked.
111 : */
112 : BufferLockMode lockmode;
113 : } PrivateRefCountData;
114 :
115 : typedef struct PrivateRefCountEntry
116 : {
117 : /*
118 : * Note that this needs to be same as the entry's corresponding
119 : * PrivateRefCountArrayKeys[i], if the entry is stored in the array. We
120 : * store it in both places as this is used for the hashtable key and
121 : * because it is more convenient (passing around a PrivateRefCountEntry
122 : * suffices to identify the buffer) and faster (checking the keys array is
123 : * faster when checking many entries, checking the entry is faster if just
124 : * checking a single entry).
125 : */
126 : Buffer buffer;
127 :
128 : char status;
129 :
130 : PrivateRefCountData data;
131 : } PrivateRefCountEntry;
132 :
133 : #define SH_PREFIX refcount
134 : #define SH_ELEMENT_TYPE PrivateRefCountEntry
135 : #define SH_KEY_TYPE Buffer
136 : #define SH_KEY buffer
137 : #define SH_HASH_KEY(tb, key) murmurhash32((uint32) (key))
138 : #define SH_EQUAL(tb, a, b) ((a) == (b))
139 : #define SH_SCOPE static inline
140 : #define SH_DECLARE
141 : #define SH_DEFINE
142 : #include "lib/simplehash.h"
143 :
144 : /* 64 bytes, about the size of a cache line on common systems */
145 : #define REFCOUNT_ARRAY_ENTRIES 8
146 :
147 : /*
148 : * Status of buffers to checkpoint for a particular tablespace, used
149 : * internally in BufferSync.
150 : */
151 : typedef struct CkptTsStatus
152 : {
153 : /* oid of the tablespace */
154 : Oid tsId;
155 :
156 : /*
157 : * Checkpoint progress for this tablespace. To make progress comparable
158 : * between tablespaces the progress is, for each tablespace, measured as a
159 : * number between 0 and the total number of to-be-checkpointed pages. Each
160 : * page checkpointed in this tablespace increments this space's progress
161 : * by progress_slice.
162 : */
163 : float8 progress;
164 : float8 progress_slice;
165 :
166 : /* number of to-be checkpointed pages in this tablespace */
167 : int num_to_scan;
168 : /* already processed pages in this tablespace */
169 : int num_scanned;
170 :
171 : /* current offset in CkptBufferIds for this tablespace */
172 : int index;
173 : } CkptTsStatus;
174 :
175 : /*
176 : * Type for array used to sort SMgrRelations
177 : *
178 : * FlushRelationsAllBuffers shares the same comparator function with
179 : * DropRelationsAllBuffers. Pointer to this struct and RelFileLocator must be
180 : * compatible.
181 : */
182 : typedef struct SMgrSortArray
183 : {
184 : RelFileLocator rlocator; /* This must be the first member */
185 : SMgrRelation srel;
186 : } SMgrSortArray;
187 :
188 : /* GUC variables */
189 : bool zero_damaged_pages = false;
190 : int bgwriter_lru_maxpages = 100;
191 : double bgwriter_lru_multiplier = 2.0;
192 : bool track_io_timing = false;
193 :
194 : /*
195 : * How many buffers PrefetchBuffer callers should try to stay ahead of their
196 : * ReadBuffer calls by. Zero means "never prefetch". This value is only used
197 : * for buffers not belonging to tablespaces that have their
198 : * effective_io_concurrency parameter set.
199 : */
200 : int effective_io_concurrency = DEFAULT_EFFECTIVE_IO_CONCURRENCY;
201 :
202 : /*
203 : * Like effective_io_concurrency, but used by maintenance code paths that might
204 : * benefit from a higher setting because they work on behalf of many sessions.
205 : * Overridden by the tablespace setting of the same name.
206 : */
207 : int maintenance_io_concurrency = DEFAULT_MAINTENANCE_IO_CONCURRENCY;
208 :
209 : /*
210 : * Limit on how many blocks should be handled in single I/O operations.
211 : * StartReadBuffers() callers should respect it, as should other operations
212 : * that call smgr APIs directly. It is computed as the minimum of underlying
213 : * GUCs io_combine_limit_guc and io_max_combine_limit.
214 : */
215 : int io_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
216 : int io_combine_limit_guc = DEFAULT_IO_COMBINE_LIMIT;
217 : int io_max_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
218 :
219 : /*
220 : * GUC variables about triggering kernel writeback for buffers written; OS
221 : * dependent defaults are set via the GUC mechanism.
222 : */
223 : int checkpoint_flush_after = DEFAULT_CHECKPOINT_FLUSH_AFTER;
224 : int bgwriter_flush_after = DEFAULT_BGWRITER_FLUSH_AFTER;
225 : int backend_flush_after = DEFAULT_BACKEND_FLUSH_AFTER;
226 :
227 : /* local state for LockBufferForCleanup */
228 : static BufferDesc *PinCountWaitBuf = NULL;
229 :
230 : /*
231 : * Backend-Private refcount management:
232 : *
233 : * Each buffer also has a private refcount that keeps track of the number of
234 : * times the buffer is pinned in the current process. This is so that the
235 : * shared refcount needs to be modified only once if a buffer is pinned more
236 : * than once by an individual backend. It's also used to check that no
237 : * buffers are still pinned at the end of transactions and when exiting. We
238 : * also use this mechanism to track whether this backend has a buffer locked,
239 : * and, if so, in what mode.
240 : *
241 : *
242 : * To avoid - as we used to - requiring an array with NBuffers entries to keep
243 : * track of local buffers, we use a small sequentially searched array
244 : * (PrivateRefCountArrayKeys, with the corresponding data stored in
245 : * PrivateRefCountArray) and an overflow hash table (PrivateRefCountHash) to
246 : * keep track of backend local pins.
247 : *
248 : * Until no more than REFCOUNT_ARRAY_ENTRIES buffers are pinned at once, all
249 : * refcounts are kept track of in the array; after that, new array entries
250 : * displace old ones into the hash table. That way a frequently used entry
251 : * can't get "stuck" in the hashtable while infrequent ones clog the array.
252 : *
253 : * Note that in most scenarios the number of pinned buffers will not exceed
254 : * REFCOUNT_ARRAY_ENTRIES.
255 : *
256 : *
257 : * To enter a buffer into the refcount tracking mechanism first reserve a free
258 : * entry using ReservePrivateRefCountEntry() and then later, if necessary,
259 : * fill it with NewPrivateRefCountEntry(). That split lets us avoid doing
260 : * memory allocations in NewPrivateRefCountEntry() which can be important
261 : * because in some scenarios it's called with a spinlock held...
262 : */
263 : static Buffer PrivateRefCountArrayKeys[REFCOUNT_ARRAY_ENTRIES];
264 : static struct PrivateRefCountEntry PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES];
265 : static refcount_hash *PrivateRefCountHash = NULL;
266 : static int32 PrivateRefCountOverflowed = 0;
267 : static uint32 PrivateRefCountClock = 0;
268 : static int ReservedRefCountSlot = -1;
269 : static int PrivateRefCountEntryLast = -1;
270 :
271 : static uint32 MaxProportionalPins;
272 :
273 : static void ReservePrivateRefCountEntry(void);
274 : static PrivateRefCountEntry *NewPrivateRefCountEntry(Buffer buffer);
275 : static PrivateRefCountEntry *GetPrivateRefCountEntry(Buffer buffer, bool do_move);
276 : static inline int32 GetPrivateRefCount(Buffer buffer);
277 : static void ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref);
278 :
279 : /* ResourceOwner callbacks to hold in-progress I/Os and buffer pins */
280 : static void ResOwnerReleaseBufferIO(Datum res);
281 : static char *ResOwnerPrintBufferIO(Datum res);
282 : static void ResOwnerReleaseBuffer(Datum res);
283 : static char *ResOwnerPrintBuffer(Datum res);
284 :
285 : const ResourceOwnerDesc buffer_io_resowner_desc =
286 : {
287 : .name = "buffer io",
288 : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
289 : .release_priority = RELEASE_PRIO_BUFFER_IOS,
290 : .ReleaseResource = ResOwnerReleaseBufferIO,
291 : .DebugPrint = ResOwnerPrintBufferIO
292 : };
293 :
294 : const ResourceOwnerDesc buffer_resowner_desc =
295 : {
296 : .name = "buffer",
297 : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
298 : .release_priority = RELEASE_PRIO_BUFFER_PINS,
299 : .ReleaseResource = ResOwnerReleaseBuffer,
300 : .DebugPrint = ResOwnerPrintBuffer
301 : };
302 :
303 : /*
304 : * Ensure that the PrivateRefCountArray has sufficient space to store one more
305 : * entry. This has to be called before using NewPrivateRefCountEntry() to fill
306 : * a new entry - but it's perfectly fine to not use a reserved entry.
307 : */
308 : static void
309 89281193 : ReservePrivateRefCountEntry(void)
310 : {
311 : /* Already reserved (or freed), nothing to do */
312 89281193 : if (ReservedRefCountSlot != -1)
313 84068831 : return;
314 :
315 : /*
316 : * First search for a free entry the array, that'll be sufficient in the
317 : * majority of cases.
318 : */
319 : {
320 : int i;
321 :
322 46911258 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
323 : {
324 41698896 : if (PrivateRefCountArrayKeys[i] == InvalidBuffer)
325 : {
326 30557413 : ReservedRefCountSlot = i;
327 :
328 : /*
329 : * We could return immediately, but iterating till the end of
330 : * the array allows compiler-autovectorization.
331 : */
332 : }
333 : }
334 :
335 5212362 : if (ReservedRefCountSlot != -1)
336 5026827 : return;
337 : }
338 :
339 : /*
340 : * No luck. All array entries are full. Move one array entry into the hash
341 : * table.
342 : */
343 : {
344 : /*
345 : * Move entry from the current clock position in the array into the
346 : * hashtable. Use that slot.
347 : */
348 : int victim_slot;
349 : PrivateRefCountEntry *victim_entry;
350 : PrivateRefCountEntry *hashent;
351 : bool found;
352 :
353 : /* select victim slot */
354 185535 : victim_slot = PrivateRefCountClock++ % REFCOUNT_ARRAY_ENTRIES;
355 185535 : victim_entry = &PrivateRefCountArray[victim_slot];
356 185535 : ReservedRefCountSlot = victim_slot;
357 :
358 : /* Better be used, otherwise we shouldn't get here. */
359 : Assert(PrivateRefCountArrayKeys[victim_slot] != InvalidBuffer);
360 : Assert(PrivateRefCountArray[victim_slot].buffer != InvalidBuffer);
361 : Assert(PrivateRefCountArrayKeys[victim_slot] == PrivateRefCountArray[victim_slot].buffer);
362 :
363 : /* enter victim array entry into hashtable */
364 185535 : hashent = refcount_insert(PrivateRefCountHash,
365 : PrivateRefCountArrayKeys[victim_slot],
366 : &found);
367 : Assert(!found);
368 : /* move data from the entry in the array to the hash entry */
369 185535 : hashent->data = victim_entry->data;
370 :
371 : /* clear the now free array slot */
372 185535 : PrivateRefCountArrayKeys[victim_slot] = InvalidBuffer;
373 185535 : victim_entry->buffer = InvalidBuffer;
374 :
375 : /* clear the whole data member, just for future proofing */
376 185535 : memset(&victim_entry->data, 0, sizeof(victim_entry->data));
377 185535 : victim_entry->data.refcount = 0;
378 185535 : victim_entry->data.lockmode = BUFFER_LOCK_UNLOCK;
379 :
380 185535 : PrivateRefCountOverflowed++;
381 : }
382 : }
383 :
384 : /*
385 : * Fill a previously reserved refcount entry.
386 : */
387 : static PrivateRefCountEntry *
388 77576725 : NewPrivateRefCountEntry(Buffer buffer)
389 : {
390 : PrivateRefCountEntry *res;
391 :
392 : /* only allowed to be called when a reservation has been made */
393 : Assert(ReservedRefCountSlot != -1);
394 :
395 : /* use up the reserved entry */
396 77576725 : res = &PrivateRefCountArray[ReservedRefCountSlot];
397 :
398 : /* and fill it */
399 77576725 : PrivateRefCountArrayKeys[ReservedRefCountSlot] = buffer;
400 77576725 : res->buffer = buffer;
401 77576725 : res->data.refcount = 0;
402 77576725 : res->data.lockmode = BUFFER_LOCK_UNLOCK;
403 :
404 : /* update cache for the next lookup */
405 77576725 : PrivateRefCountEntryLast = ReservedRefCountSlot;
406 :
407 77576725 : ReservedRefCountSlot = -1;
408 :
409 77576725 : return res;
410 : }
411 :
412 : /*
413 : * Slow-path for GetPrivateRefCountEntry(). This is big enough to not be worth
414 : * inlining. This particularly seems to be true if the compiler is capable of
415 : * auto-vectorizing the code, as that imposes additional stack-alignment
416 : * requirements etc.
417 : */
418 : static pg_noinline PrivateRefCountEntry *
419 107751171 : GetPrivateRefCountEntrySlow(Buffer buffer, bool do_move)
420 : {
421 : PrivateRefCountEntry *res;
422 107751171 : int match = -1;
423 : int i;
424 :
425 : /*
426 : * First search for references in the array, that'll be sufficient in the
427 : * majority of cases.
428 : */
429 969760539 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
430 : {
431 862009368 : if (PrivateRefCountArrayKeys[i] == buffer)
432 : {
433 32586974 : match = i;
434 : /* see ReservePrivateRefCountEntry() for why we don't return */
435 : }
436 : }
437 :
438 107751171 : if (likely(match != -1))
439 : {
440 : /* update cache for the next lookup */
441 32586974 : PrivateRefCountEntryLast = match;
442 :
443 32586974 : return &PrivateRefCountArray[match];
444 : }
445 :
446 : /*
447 : * By here we know that the buffer, if already pinned, isn't residing in
448 : * the array.
449 : *
450 : * Only look up the buffer in the hashtable if we've previously overflowed
451 : * into it.
452 : */
453 75164197 : if (PrivateRefCountOverflowed == 0)
454 74557414 : return NULL;
455 :
456 606783 : res = refcount_lookup(PrivateRefCountHash, buffer);
457 :
458 606783 : if (res == NULL)
459 373219 : return NULL;
460 233564 : else if (!do_move)
461 : {
462 : /* caller doesn't want us to move the hash entry into the array */
463 136813 : return res;
464 : }
465 : else
466 : {
467 : /* move buffer from hashtable into the free array slot */
468 : PrivateRefCountEntry *free;
469 : PrivateRefCountData data;
470 :
471 : /* Save data and delete from hashtable while res is still valid */
472 96751 : data = res->data;
473 96751 : refcount_delete_item(PrivateRefCountHash, res);
474 : Assert(PrivateRefCountOverflowed > 0);
475 96751 : PrivateRefCountOverflowed--;
476 :
477 : /* Ensure there's a free array slot */
478 96751 : ReservePrivateRefCountEntry();
479 :
480 : /* Use up the reserved slot */
481 : Assert(ReservedRefCountSlot != -1);
482 96751 : free = &PrivateRefCountArray[ReservedRefCountSlot];
483 : Assert(PrivateRefCountArrayKeys[ReservedRefCountSlot] == free->buffer);
484 : Assert(free->buffer == InvalidBuffer);
485 :
486 : /* and fill it */
487 96751 : free->buffer = buffer;
488 96751 : free->data = data;
489 96751 : PrivateRefCountArrayKeys[ReservedRefCountSlot] = buffer;
490 : /* update cache for the next lookup */
491 96751 : PrivateRefCountEntryLast = ReservedRefCountSlot;
492 :
493 96751 : ReservedRefCountSlot = -1;
494 :
495 96751 : return free;
496 : }
497 : }
498 :
499 : /*
500 : * Return the PrivateRefCount entry for the passed buffer.
501 : *
502 : * Returns NULL if a buffer doesn't have a refcount entry. Otherwise, if
503 : * do_move is true, and the entry resides in the hashtable the entry is
504 : * optimized for frequent access by moving it to the array.
505 : */
506 : static inline PrivateRefCountEntry *
507 444725449 : GetPrivateRefCountEntry(Buffer buffer, bool do_move)
508 : {
509 : Assert(BufferIsValid(buffer));
510 : Assert(!BufferIsLocal(buffer));
511 :
512 : /*
513 : * It's very common to look up the same buffer repeatedly. To make that
514 : * fast, we have a one-entry cache.
515 : *
516 : * In contrast to the loop in GetPrivateRefCountEntrySlow(), here it
517 : * faster to check PrivateRefCountArray[].buffer, as in the case of a hit
518 : * fewer addresses are computed and fewer cachelines are accessed. Whereas
519 : * in GetPrivateRefCountEntrySlow()'s case, checking
520 : * PrivateRefCountArrayKeys saves a lot of memory accesses.
521 : */
522 444725449 : if (likely(PrivateRefCountEntryLast != -1) &&
523 444709362 : likely(PrivateRefCountArray[PrivateRefCountEntryLast].buffer == buffer))
524 : {
525 336974278 : return &PrivateRefCountArray[PrivateRefCountEntryLast];
526 : }
527 :
528 : /*
529 : * The code for the cached lookup is small enough to be worth inlining
530 : * into the caller. In the miss case however, that empirically doesn't
531 : * seem worth it.
532 : */
533 107751171 : return GetPrivateRefCountEntrySlow(buffer, do_move);
534 : }
535 :
536 : /*
537 : * Returns how many times the passed buffer is pinned by this backend.
538 : *
539 : * Only works for shared memory buffers!
540 : */
541 : static inline int32
542 3393776 : GetPrivateRefCount(Buffer buffer)
543 : {
544 : PrivateRefCountEntry *ref;
545 :
546 : Assert(BufferIsValid(buffer));
547 : Assert(!BufferIsLocal(buffer));
548 :
549 : /*
550 : * Not moving the entry - that's ok for the current users, but we might
551 : * want to change this one day.
552 : */
553 3393776 : ref = GetPrivateRefCountEntry(buffer, false);
554 :
555 3393776 : if (ref == NULL)
556 31 : return 0;
557 3393745 : return ref->data.refcount;
558 : }
559 :
560 : /*
561 : * Release resources used to track the reference count of a buffer which we no
562 : * longer have pinned and don't want to pin again immediately.
563 : */
564 : static void
565 77576725 : ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref)
566 : {
567 : Assert(ref->data.refcount == 0);
568 : Assert(ref->data.lockmode == BUFFER_LOCK_UNLOCK);
569 :
570 77576725 : if (ref >= &PrivateRefCountArray[0] &&
571 : ref < &PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES])
572 : {
573 77487941 : ref->buffer = InvalidBuffer;
574 77487941 : PrivateRefCountArrayKeys[ref - PrivateRefCountArray] = InvalidBuffer;
575 :
576 :
577 : /*
578 : * Mark the just used entry as reserved - in many scenarios that
579 : * allows us to avoid ever having to search the array/hash for free
580 : * entries.
581 : */
582 77487941 : ReservedRefCountSlot = ref - PrivateRefCountArray;
583 : }
584 : else
585 : {
586 88784 : refcount_delete_item(PrivateRefCountHash, ref);
587 : Assert(PrivateRefCountOverflowed > 0);
588 88784 : PrivateRefCountOverflowed--;
589 : }
590 77576725 : }
591 :
592 : /*
593 : * BufferIsPinned
594 : * True iff the buffer is pinned (also checks for valid buffer number).
595 : *
596 : * NOTE: what we check here is that *this* backend holds a pin on
597 : * the buffer. We do not care whether some other backend does.
598 : */
599 : #define BufferIsPinned(bufnum) \
600 : ( \
601 : !BufferIsValid(bufnum) ? \
602 : false \
603 : : \
604 : BufferIsLocal(bufnum) ? \
605 : (LocalRefCount[-(bufnum) - 1] > 0) \
606 : : \
607 : (GetPrivateRefCount(bufnum) > 0) \
608 : )
609 :
610 :
611 : static Buffer ReadBuffer_common(Relation rel,
612 : SMgrRelation smgr, char smgr_persistence,
613 : ForkNumber forkNum, BlockNumber blockNum,
614 : ReadBufferMode mode, BufferAccessStrategy strategy);
615 : static BlockNumber ExtendBufferedRelCommon(BufferManagerRelation bmr,
616 : ForkNumber fork,
617 : BufferAccessStrategy strategy,
618 : uint32 flags,
619 : uint32 extend_by,
620 : BlockNumber extend_upto,
621 : Buffer *buffers,
622 : uint32 *extended_by);
623 : static BlockNumber ExtendBufferedRelShared(BufferManagerRelation bmr,
624 : ForkNumber fork,
625 : BufferAccessStrategy strategy,
626 : uint32 flags,
627 : uint32 extend_by,
628 : BlockNumber extend_upto,
629 : Buffer *buffers,
630 : uint32 *extended_by);
631 : static bool PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
632 : bool skip_if_not_valid);
633 : static void PinBuffer_Locked(BufferDesc *buf);
634 : static void UnpinBuffer(BufferDesc *buf);
635 : static void UnpinBufferNoOwner(BufferDesc *buf);
636 : static void BufferSync(int flags);
637 : static int SyncOneBuffer(int buf_id, bool skip_recently_used,
638 : WritebackContext *wb_context);
639 : static void WaitIO(BufferDesc *buf);
640 : static void AbortBufferIO(Buffer buffer);
641 : static void shared_buffer_write_error_callback(void *arg);
642 : static void local_buffer_write_error_callback(void *arg);
643 : static inline BufferDesc *BufferAlloc(SMgrRelation smgr,
644 : char relpersistence,
645 : ForkNumber forkNum,
646 : BlockNumber blockNum,
647 : BufferAccessStrategy strategy,
648 : bool *foundPtr, IOContext io_context);
649 : static bool AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress);
650 : static void CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete);
651 :
652 : static pg_attribute_always_inline void TrackBufferHit(IOObject io_object,
653 : IOContext io_context,
654 : Relation rel, char persistence, SMgrRelation smgr,
655 : ForkNumber forknum, BlockNumber blocknum);
656 : static Buffer GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context);
657 : static void FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
658 : IOObject io_object, IOContext io_context);
659 : static void FlushBuffer(BufferDesc *buf, SMgrRelation reln,
660 : IOObject io_object, IOContext io_context);
661 : static void FindAndDropRelationBuffers(RelFileLocator rlocator,
662 : ForkNumber forkNum,
663 : BlockNumber nForkBlock,
664 : BlockNumber firstDelBlock);
665 : static void RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
666 : RelFileLocator dstlocator,
667 : ForkNumber forkNum, bool permanent);
668 : static void AtProcExit_Buffers(int code, Datum arg);
669 : static void CheckForBufferLeaks(void);
670 : #ifdef USE_ASSERT_CHECKING
671 : static void AssertNotCatalogBufferLock(Buffer buffer, BufferLockMode mode);
672 : #endif
673 : static int rlocator_comparator(const void *p1, const void *p2);
674 : static inline int buffertag_comparator(const BufferTag *ba, const BufferTag *bb);
675 : static inline int ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b);
676 : static int ts_ckpt_progress_comparator(Datum a, Datum b, void *arg);
677 :
678 : static void BufferLockAcquire(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode);
679 : static void BufferLockUnlock(Buffer buffer, BufferDesc *buf_hdr);
680 : static bool BufferLockConditional(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode);
681 : static bool BufferLockHeldByMeInMode(BufferDesc *buf_hdr, BufferLockMode mode);
682 : static bool BufferLockHeldByMe(BufferDesc *buf_hdr);
683 : static inline void BufferLockDisown(Buffer buffer, BufferDesc *buf_hdr);
684 : static inline int BufferLockDisownInternal(Buffer buffer, BufferDesc *buf_hdr);
685 : static inline bool BufferLockAttempt(BufferDesc *buf_hdr, BufferLockMode mode);
686 : static void BufferLockQueueSelf(BufferDesc *buf_hdr, BufferLockMode mode);
687 : static void BufferLockDequeueSelf(BufferDesc *buf_hdr);
688 : static void BufferLockWakeup(BufferDesc *buf_hdr, bool wake_exclusive);
689 : static void BufferLockProcessRelease(BufferDesc *buf_hdr, BufferLockMode mode, uint64 lockstate);
690 : static inline uint64 BufferLockReleaseSub(BufferLockMode mode);
691 :
692 :
693 : /*
694 : * Implementation of PrefetchBuffer() for shared buffers.
695 : */
696 : PrefetchBufferResult
697 40453 : PrefetchSharedBuffer(SMgrRelation smgr_reln,
698 : ForkNumber forkNum,
699 : BlockNumber blockNum)
700 : {
701 40453 : PrefetchBufferResult result = {InvalidBuffer, false};
702 : BufferTag newTag; /* identity of requested block */
703 : uint32 newHash; /* hash value for newTag */
704 : LWLock *newPartitionLock; /* buffer partition lock for it */
705 : int buf_id;
706 :
707 : Assert(BlockNumberIsValid(blockNum));
708 :
709 : /* create a tag so we can lookup the buffer */
710 40453 : InitBufferTag(&newTag, &smgr_reln->smgr_rlocator.locator,
711 : forkNum, blockNum);
712 :
713 : /* determine its hash code and partition lock ID */
714 40453 : newHash = BufTableHashCode(&newTag);
715 40453 : newPartitionLock = BufMappingPartitionLock(newHash);
716 :
717 : /* see if the block is in the buffer pool already */
718 40453 : LWLockAcquire(newPartitionLock, LW_SHARED);
719 40453 : buf_id = BufTableLookup(&newTag, newHash);
720 40453 : LWLockRelease(newPartitionLock);
721 :
722 : /* If not in buffers, initiate prefetch */
723 40453 : if (buf_id < 0)
724 : {
725 : #ifdef USE_PREFETCH
726 : /*
727 : * Try to initiate an asynchronous read. This returns false in
728 : * recovery if the relation file doesn't exist.
729 : */
730 19125 : if ((io_direct_flags & IO_DIRECT_DATA) == 0 &&
731 9436 : smgrprefetch(smgr_reln, forkNum, blockNum, 1))
732 : {
733 9436 : result.initiated_io = true;
734 : }
735 : #endif /* USE_PREFETCH */
736 : }
737 : else
738 : {
739 : /*
740 : * Report the buffer it was in at that time. The caller may be able
741 : * to avoid a buffer table lookup, but it's not pinned and it must be
742 : * rechecked!
743 : */
744 30764 : result.recent_buffer = buf_id + 1;
745 : }
746 :
747 : /*
748 : * If the block *is* in buffers, we do nothing. This is not really ideal:
749 : * the block might be just about to be evicted, which would be stupid
750 : * since we know we are going to need it soon. But the only easy answer
751 : * is to bump the usage_count, which does not seem like a great solution:
752 : * when the caller does ultimately touch the block, usage_count would get
753 : * bumped again, resulting in too much favoritism for blocks that are
754 : * involved in a prefetch sequence. A real fix would involve some
755 : * additional per-buffer state, and it's not clear that there's enough of
756 : * a problem to justify that.
757 : */
758 :
759 40453 : return result;
760 : }
761 :
762 : /*
763 : * PrefetchBuffer -- initiate asynchronous read of a block of a relation
764 : *
765 : * This is named by analogy to ReadBuffer but doesn't actually allocate a
766 : * buffer. Instead it tries to ensure that a future ReadBuffer for the given
767 : * block will not be delayed by the I/O. Prefetching is optional.
768 : *
769 : * There are three possible outcomes:
770 : *
771 : * 1. If the block is already cached, the result includes a valid buffer that
772 : * could be used by the caller to avoid the need for a later buffer lookup, but
773 : * it's not pinned, so the caller must recheck it.
774 : *
775 : * 2. If the kernel has been asked to initiate I/O, the initiated_io member is
776 : * true. Currently there is no way to know if the data was already cached by
777 : * the kernel and therefore didn't really initiate I/O, and no way to know when
778 : * the I/O completes other than using synchronous ReadBuffer().
779 : *
780 : * 3. Otherwise, the buffer wasn't already cached by PostgreSQL, and
781 : * USE_PREFETCH is not defined (this build doesn't support prefetching due to
782 : * lack of a kernel facility), direct I/O is enabled, or the underlying
783 : * relation file wasn't found and we are in recovery. (If the relation file
784 : * wasn't found and we are not in recovery, an error is raised).
785 : */
786 : PrefetchBufferResult
787 29368 : PrefetchBuffer(Relation reln, ForkNumber forkNum, BlockNumber blockNum)
788 : {
789 : Assert(RelationIsValid(reln));
790 : Assert(BlockNumberIsValid(blockNum));
791 :
792 29368 : if (RelationUsesLocalBuffers(reln))
793 : {
794 : /* see comments in ReadBuffer_common */
795 1357 : if (RELATION_IS_OTHER_TEMP(reln))
796 0 : ereport(ERROR,
797 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
798 : errmsg("cannot access temporary tables of other sessions")));
799 :
800 : /* pass it off to localbuf.c */
801 1357 : return PrefetchLocalBuffer(RelationGetSmgr(reln), forkNum, blockNum);
802 : }
803 : else
804 : {
805 : /* pass it to the shared buffer version */
806 28011 : return PrefetchSharedBuffer(RelationGetSmgr(reln), forkNum, blockNum);
807 : }
808 : }
809 :
810 : /*
811 : * ReadRecentBuffer -- try to pin a block in a recently observed buffer
812 : *
813 : * Compared to ReadBuffer(), this avoids a buffer mapping lookup when it's
814 : * successful. Return true if the buffer is valid and still has the expected
815 : * tag. In that case, the buffer is pinned and the usage count is bumped.
816 : */
817 : bool
818 5093 : ReadRecentBuffer(RelFileLocator rlocator, ForkNumber forkNum, BlockNumber blockNum,
819 : Buffer recent_buffer)
820 : {
821 : BufferDesc *bufHdr;
822 : BufferTag tag;
823 : uint64 buf_state;
824 :
825 : Assert(BufferIsValid(recent_buffer));
826 :
827 5093 : ResourceOwnerEnlarge(CurrentResourceOwner);
828 5093 : ReservePrivateRefCountEntry();
829 5093 : InitBufferTag(&tag, &rlocator, forkNum, blockNum);
830 :
831 5093 : if (BufferIsLocal(recent_buffer))
832 : {
833 140 : int b = -recent_buffer - 1;
834 :
835 140 : bufHdr = GetLocalBufferDescriptor(b);
836 140 : buf_state = pg_atomic_read_u64(&bufHdr->state);
837 :
838 : /* Is it still valid and holding the right tag? */
839 140 : if ((buf_state & BM_VALID) && BufferTagsEqual(&tag, &bufHdr->tag))
840 : {
841 140 : PinLocalBuffer(bufHdr, true);
842 :
843 140 : pgBufferUsage.local_blks_hit++;
844 :
845 140 : return true;
846 : }
847 : }
848 : else
849 : {
850 4953 : bufHdr = GetBufferDescriptor(recent_buffer - 1);
851 :
852 : /*
853 : * Is it still valid and holding the right tag? We do an unlocked tag
854 : * comparison first, to make it unlikely that we'll increment the
855 : * usage counter of the wrong buffer, if someone calls us with a very
856 : * out of date recent_buffer. Then we'll check it again if we get the
857 : * pin.
858 : */
859 9866 : if (BufferTagsEqual(&tag, &bufHdr->tag) &&
860 4913 : PinBuffer(bufHdr, NULL, true))
861 : {
862 4907 : if (BufferTagsEqual(&tag, &bufHdr->tag))
863 : {
864 4907 : pgBufferUsage.shared_blks_hit++;
865 4907 : return true;
866 : }
867 0 : UnpinBuffer(bufHdr);
868 : }
869 : }
870 :
871 46 : return false;
872 : }
873 :
874 : /*
875 : * ReadBuffer -- a shorthand for ReadBufferExtended, for reading from main
876 : * fork with RBM_NORMAL mode and default strategy.
877 : */
878 : Buffer
879 63412345 : ReadBuffer(Relation reln, BlockNumber blockNum)
880 : {
881 63412345 : return ReadBufferExtended(reln, MAIN_FORKNUM, blockNum, RBM_NORMAL, NULL);
882 : }
883 :
884 : /*
885 : * ReadBufferExtended -- returns a buffer containing the requested
886 : * block of the requested relation. If the blknum
887 : * requested is P_NEW, extend the relation file and
888 : * allocate a new block. (Caller is responsible for
889 : * ensuring that only one backend tries to extend a
890 : * relation at the same time!)
891 : *
892 : * Returns: the buffer number for the buffer containing
893 : * the block read. The returned buffer has been pinned.
894 : * Does not return on error --- elog's instead.
895 : *
896 : * Assume when this function is called, that reln has been opened already.
897 : *
898 : * In RBM_NORMAL mode, the page is read from disk, and the page header is
899 : * validated. An error is thrown if the page header is not valid. (But
900 : * note that an all-zero page is considered "valid"; see
901 : * PageIsVerified().)
902 : *
903 : * RBM_ZERO_ON_ERROR is like the normal mode, but if the page header is not
904 : * valid, the page is zeroed instead of throwing an error. This is intended
905 : * for non-critical data, where the caller is prepared to repair errors.
906 : *
907 : * In RBM_ZERO_AND_LOCK mode, if the page isn't in buffer cache already, it's
908 : * filled with zeros instead of reading it from disk. Useful when the caller
909 : * is going to fill the page from scratch, since this saves I/O and avoids
910 : * unnecessary failure if the page-on-disk has corrupt page headers.
911 : * The page is returned locked to ensure that the caller has a chance to
912 : * initialize the page before it's made visible to others.
913 : * Caution: do not use this mode to read a page that is beyond the relation's
914 : * current physical EOF; that is likely to cause problems in md.c when
915 : * the page is modified and written out. P_NEW is OK, though.
916 : *
917 : * RBM_ZERO_AND_CLEANUP_LOCK is the same as RBM_ZERO_AND_LOCK, but acquires
918 : * a cleanup-strength lock on the page.
919 : *
920 : * RBM_NORMAL_NO_LOG mode is treated the same as RBM_NORMAL here.
921 : *
922 : * If strategy is not NULL, a nondefault buffer access strategy is used.
923 : * See buffer/README for details.
924 : */
925 : inline Buffer
926 74836197 : ReadBufferExtended(Relation reln, ForkNumber forkNum, BlockNumber blockNum,
927 : ReadBufferMode mode, BufferAccessStrategy strategy)
928 : {
929 : Buffer buf;
930 :
931 : /*
932 : * Read the buffer, and update pgstat counters to reflect a cache hit or
933 : * miss. The other-session temp-relation check is enforced by
934 : * ReadBuffer_common().
935 : */
936 74836197 : buf = ReadBuffer_common(reln, RelationGetSmgr(reln), 0,
937 : forkNum, blockNum, mode, strategy);
938 :
939 74836172 : return buf;
940 : }
941 :
942 :
943 : /*
944 : * ReadBufferWithoutRelcache -- like ReadBufferExtended, but doesn't require
945 : * a relcache entry for the relation.
946 : *
947 : * Pass permanent = true for a RELPERSISTENCE_PERMANENT relation, and
948 : * permanent = false for a RELPERSISTENCE_UNLOGGED relation. This function
949 : * cannot be used for temporary relations (and making that work might be
950 : * difficult, unless we only want to read temporary relations for our own
951 : * ProcNumber).
952 : */
953 : Buffer
954 5971431 : ReadBufferWithoutRelcache(RelFileLocator rlocator, ForkNumber forkNum,
955 : BlockNumber blockNum, ReadBufferMode mode,
956 : BufferAccessStrategy strategy, bool permanent)
957 : {
958 5971431 : SMgrRelation smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
959 :
960 5971431 : return ReadBuffer_common(NULL, smgr,
961 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
962 : forkNum, blockNum,
963 : mode, strategy);
964 : }
965 :
966 : /*
967 : * Convenience wrapper around ExtendBufferedRelBy() extending by one block.
968 : */
969 : Buffer
970 56737 : ExtendBufferedRel(BufferManagerRelation bmr,
971 : ForkNumber forkNum,
972 : BufferAccessStrategy strategy,
973 : uint32 flags)
974 : {
975 : Buffer buf;
976 56737 : uint32 extend_by = 1;
977 :
978 56737 : ExtendBufferedRelBy(bmr, forkNum, strategy, flags, extend_by,
979 : &buf, &extend_by);
980 :
981 56737 : return buf;
982 : }
983 :
984 : /*
985 : * Extend relation by multiple blocks.
986 : *
987 : * Tries to extend the relation by extend_by blocks. Depending on the
988 : * availability of resources the relation may end up being extended by a
989 : * smaller number of pages (unless an error is thrown, always by at least one
990 : * page). *extended_by is updated to the number of pages the relation has been
991 : * extended to.
992 : *
993 : * buffers needs to be an array that is at least extend_by long. Upon
994 : * completion, the first extend_by array elements will point to a pinned
995 : * buffer.
996 : *
997 : * If EB_LOCK_FIRST is part of flags, the first returned buffer is
998 : * locked. This is useful for callers that want a buffer that is guaranteed to
999 : * be empty.
1000 : */
1001 : BlockNumber
1002 209519 : ExtendBufferedRelBy(BufferManagerRelation bmr,
1003 : ForkNumber fork,
1004 : BufferAccessStrategy strategy,
1005 : uint32 flags,
1006 : uint32 extend_by,
1007 : Buffer *buffers,
1008 : uint32 *extended_by)
1009 : {
1010 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
1011 : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
1012 : Assert(extend_by > 0);
1013 :
1014 209519 : if (bmr.relpersistence == '\0')
1015 209519 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
1016 :
1017 209519 : return ExtendBufferedRelCommon(bmr, fork, strategy, flags,
1018 : extend_by, InvalidBlockNumber,
1019 : buffers, extended_by);
1020 : }
1021 :
1022 : /*
1023 : * Extend the relation so it is at least extend_to blocks large, return buffer
1024 : * (extend_to - 1).
1025 : *
1026 : * This is useful for callers that want to write a specific page, regardless
1027 : * of the current size of the relation (e.g. useful for visibilitymap and for
1028 : * crash recovery).
1029 : */
1030 : Buffer
1031 55706 : ExtendBufferedRelTo(BufferManagerRelation bmr,
1032 : ForkNumber fork,
1033 : BufferAccessStrategy strategy,
1034 : uint32 flags,
1035 : BlockNumber extend_to,
1036 : ReadBufferMode mode)
1037 : {
1038 : BlockNumber current_size;
1039 55706 : uint32 extended_by = 0;
1040 55706 : Buffer buffer = InvalidBuffer;
1041 : Buffer buffers[64];
1042 :
1043 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
1044 : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
1045 : Assert(extend_to != InvalidBlockNumber && extend_to > 0);
1046 :
1047 55706 : if (bmr.relpersistence == '\0')
1048 9437 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
1049 :
1050 : /*
1051 : * If desired, create the file if it doesn't exist. If
1052 : * smgr_cached_nblocks[fork] is positive then it must exist, no need for
1053 : * an smgrexists call.
1054 : */
1055 55706 : if ((flags & EB_CREATE_FORK_IF_NEEDED) &&
1056 9437 : (BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == 0 ||
1057 30 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == InvalidBlockNumber) &&
1058 9407 : !smgrexists(BMR_GET_SMGR(bmr), fork))
1059 : {
1060 9378 : LockRelationForExtension(bmr.rel, ExclusiveLock);
1061 :
1062 : /* recheck, fork might have been created concurrently */
1063 9378 : if (!smgrexists(BMR_GET_SMGR(bmr), fork))
1064 9377 : smgrcreate(BMR_GET_SMGR(bmr), fork, flags & EB_PERFORMING_RECOVERY);
1065 :
1066 9378 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
1067 : }
1068 :
1069 : /*
1070 : * If requested, invalidate size cache, so that smgrnblocks asks the
1071 : * kernel.
1072 : */
1073 55706 : if (flags & EB_CLEAR_SIZE_CACHE)
1074 9437 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
1075 :
1076 : /*
1077 : * Estimate how many pages we'll need to extend by. This avoids acquiring
1078 : * unnecessarily many victim buffers.
1079 : */
1080 55706 : current_size = smgrnblocks(BMR_GET_SMGR(bmr), fork);
1081 :
1082 : /*
1083 : * Since no-one else can be looking at the page contents yet, there is no
1084 : * difference between an exclusive lock and a cleanup-strength lock. Note
1085 : * that we pass the original mode to ReadBuffer_common() below, when
1086 : * falling back to reading the buffer to a concurrent relation extension.
1087 : */
1088 55706 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1089 45869 : flags |= EB_LOCK_TARGET;
1090 :
1091 113617 : while (current_size < extend_to)
1092 : {
1093 57911 : uint32 num_pages = lengthof(buffers);
1094 : BlockNumber first_block;
1095 :
1096 57911 : if ((uint64) current_size + num_pages > extend_to)
1097 57845 : num_pages = extend_to - current_size;
1098 :
1099 57911 : first_block = ExtendBufferedRelCommon(bmr, fork, strategy, flags,
1100 : num_pages, extend_to,
1101 : buffers, &extended_by);
1102 :
1103 57911 : current_size = first_block + extended_by;
1104 : Assert(num_pages != 0 || current_size >= extend_to);
1105 :
1106 124610 : for (uint32 i = 0; i < extended_by; i++)
1107 : {
1108 66699 : if (first_block + i != extend_to - 1)
1109 10995 : ReleaseBuffer(buffers[i]);
1110 : else
1111 55704 : buffer = buffers[i];
1112 : }
1113 : }
1114 :
1115 : /*
1116 : * It's possible that another backend concurrently extended the relation.
1117 : * In that case read the buffer.
1118 : *
1119 : * XXX: Should we control this via a flag?
1120 : */
1121 55706 : if (buffer == InvalidBuffer)
1122 : {
1123 : Assert(extended_by == 0);
1124 2 : buffer = ReadBuffer_common(bmr.rel, BMR_GET_SMGR(bmr), bmr.relpersistence,
1125 : fork, extend_to - 1, mode, strategy);
1126 : }
1127 :
1128 55706 : return buffer;
1129 : }
1130 :
1131 : /*
1132 : * Lock and optionally zero a buffer, as part of the implementation of
1133 : * RBM_ZERO_AND_LOCK or RBM_ZERO_AND_CLEANUP_LOCK. The buffer must be already
1134 : * pinned. If the buffer is not already valid, it is zeroed and made valid.
1135 : */
1136 : static void
1137 350111 : ZeroAndLockBuffer(Buffer buffer, ReadBufferMode mode, bool already_valid)
1138 : {
1139 : BufferDesc *bufHdr;
1140 : bool need_to_zero;
1141 350111 : bool isLocalBuf = BufferIsLocal(buffer);
1142 : StartBufferIOResult sbres;
1143 :
1144 : Assert(mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK);
1145 :
1146 350111 : if (already_valid)
1147 : {
1148 : /*
1149 : * If the caller already knew the buffer was valid, we can skip some
1150 : * header interaction. The caller just wants to lock the buffer.
1151 : */
1152 37949 : need_to_zero = false;
1153 : }
1154 : else
1155 : {
1156 312162 : if (isLocalBuf)
1157 : {
1158 : /* Simple case for non-shared buffers. */
1159 30 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
1160 30 : sbres = StartLocalBufferIO(bufHdr, true, true, NULL);
1161 : }
1162 : else
1163 : {
1164 : /*
1165 : * Take BM_IO_IN_PROGRESS, or discover that BM_VALID has been set
1166 : * concurrently. Even though we aren't doing I/O, that ensures
1167 : * that we don't zero a page that someone else has pinned. An
1168 : * exclusive content lock wouldn't be enough, because readers are
1169 : * allowed to drop the content lock after determining that a tuple
1170 : * is visible (see buffer access rules in README).
1171 : */
1172 312132 : bufHdr = GetBufferDescriptor(buffer - 1);
1173 312132 : sbres = StartSharedBufferIO(bufHdr, true, true, NULL);
1174 : }
1175 :
1176 : Assert(sbres != BUFFER_IO_IN_PROGRESS);
1177 312162 : need_to_zero = sbres == BUFFER_IO_READY_FOR_IO;
1178 : }
1179 :
1180 350111 : if (need_to_zero)
1181 : {
1182 312162 : memset(BufferGetPage(buffer), 0, BLCKSZ);
1183 :
1184 : /*
1185 : * Grab the buffer content lock before marking the page as valid, to
1186 : * make sure that no other backend sees the zeroed page before the
1187 : * caller has had a chance to initialize it.
1188 : *
1189 : * Since no-one else can be looking at the page contents yet, there is
1190 : * no difference between an exclusive lock and a cleanup-strength
1191 : * lock. (Note that we cannot use LockBuffer() or
1192 : * LockBufferForCleanup() here, because they assert that the buffer is
1193 : * already valid.)
1194 : */
1195 312162 : if (!isLocalBuf)
1196 312132 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1197 :
1198 : /* Set BM_VALID, terminate IO, and wake up any waiters */
1199 312162 : if (isLocalBuf)
1200 30 : TerminateLocalBufferIO(bufHdr, false, BM_VALID, false);
1201 : else
1202 312132 : TerminateBufferIO(bufHdr, false, BM_VALID, true, false);
1203 : }
1204 37949 : else if (!isLocalBuf)
1205 : {
1206 : /*
1207 : * The buffer is valid, so we can't zero it. The caller still expects
1208 : * the page to be locked on return.
1209 : */
1210 37929 : if (mode == RBM_ZERO_AND_LOCK)
1211 37864 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1212 : else
1213 65 : LockBufferForCleanup(buffer);
1214 : }
1215 350111 : }
1216 :
1217 : /*
1218 : * Pin a buffer for a given block. *foundPtr is set to true if the block was
1219 : * already present, or false if more work is required to either read it in or
1220 : * zero it.
1221 : */
1222 : static pg_attribute_always_inline Buffer
1223 85913820 : PinBufferForBlock(Relation rel,
1224 : SMgrRelation smgr,
1225 : char persistence,
1226 : ForkNumber forkNum,
1227 : BlockNumber blockNum,
1228 : BufferAccessStrategy strategy,
1229 : IOObject io_object,
1230 : IOContext io_context,
1231 : bool *foundPtr)
1232 : {
1233 : BufferDesc *bufHdr;
1234 :
1235 : Assert(blockNum != P_NEW);
1236 :
1237 : /* Persistence should be set before */
1238 : Assert((persistence == RELPERSISTENCE_TEMP ||
1239 : persistence == RELPERSISTENCE_PERMANENT ||
1240 : persistence == RELPERSISTENCE_UNLOGGED));
1241 :
1242 : TRACE_POSTGRESQL_BUFFER_READ_START(forkNum, blockNum,
1243 : smgr->smgr_rlocator.locator.spcOid,
1244 : smgr->smgr_rlocator.locator.dbOid,
1245 : smgr->smgr_rlocator.locator.relNumber,
1246 : smgr->smgr_rlocator.backend);
1247 :
1248 85913820 : if (persistence == RELPERSISTENCE_TEMP)
1249 1648607 : bufHdr = LocalBufferAlloc(smgr, forkNum, blockNum, foundPtr);
1250 : else
1251 84265213 : bufHdr = BufferAlloc(smgr, persistence, forkNum, blockNum,
1252 : strategy, foundPtr, io_context);
1253 :
1254 85913812 : if (*foundPtr)
1255 83912241 : TrackBufferHit(io_object, io_context, rel, persistence, smgr, forkNum, blockNum);
1256 :
1257 85913812 : if (rel)
1258 : {
1259 : /*
1260 : * While pgBufferUsage's "read" counter isn't bumped unless we reach
1261 : * WaitReadBuffers() (so, not for hits, and not for buffers that are
1262 : * zeroed instead), the per-relation stats always count them.
1263 : */
1264 79689419 : pgstat_count_buffer_read(rel);
1265 : }
1266 :
1267 85913812 : return BufferDescriptorGetBuffer(bufHdr);
1268 : }
1269 :
1270 : /*
1271 : * ReadBuffer_common -- common logic for all ReadBuffer variants
1272 : *
1273 : * smgr is required, rel is optional unless using P_NEW.
1274 : */
1275 : static pg_attribute_always_inline Buffer
1276 80807630 : ReadBuffer_common(Relation rel, SMgrRelation smgr, char smgr_persistence,
1277 : ForkNumber forkNum,
1278 : BlockNumber blockNum, ReadBufferMode mode,
1279 : BufferAccessStrategy strategy)
1280 : {
1281 : ReadBuffersOperation operation;
1282 : Buffer buffer;
1283 : int flags;
1284 : char persistence;
1285 :
1286 : /*
1287 : * Reject attempts to read non-local temporary relations; we would be
1288 : * likely to get wrong data since we have no visibility into the owning
1289 : * session's local buffers. This is the canonical place for the check,
1290 : * covering the ReadBufferExtended() entry point and any other caller that
1291 : * supplies a Relation.
1292 : */
1293 80807630 : if (rel && RELATION_IS_OTHER_TEMP(rel))
1294 2 : ereport(ERROR,
1295 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1296 : errmsg("cannot access temporary tables of other sessions")));
1297 :
1298 : /*
1299 : * Backward compatibility path, most code should use ExtendBufferedRel()
1300 : * instead, as acquiring the extension lock inside ExtendBufferedRel()
1301 : * scales a lot better.
1302 : */
1303 80807628 : if (unlikely(blockNum == P_NEW))
1304 : {
1305 322 : uint32 flags = EB_SKIP_EXTENSION_LOCK;
1306 :
1307 : /*
1308 : * Since no-one else can be looking at the page contents yet, there is
1309 : * no difference between an exclusive lock and a cleanup-strength
1310 : * lock.
1311 : */
1312 322 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1313 0 : flags |= EB_LOCK_FIRST;
1314 :
1315 322 : return ExtendBufferedRel(BMR_REL(rel), forkNum, strategy, flags);
1316 : }
1317 :
1318 80807306 : if (rel)
1319 74835875 : persistence = rel->rd_rel->relpersistence;
1320 : else
1321 5971431 : persistence = smgr_persistence;
1322 :
1323 80807306 : if (unlikely(mode == RBM_ZERO_AND_CLEANUP_LOCK ||
1324 : mode == RBM_ZERO_AND_LOCK))
1325 : {
1326 : bool found;
1327 : IOContext io_context;
1328 : IOObject io_object;
1329 :
1330 350111 : if (persistence == RELPERSISTENCE_TEMP)
1331 : {
1332 50 : io_context = IOCONTEXT_NORMAL;
1333 50 : io_object = IOOBJECT_TEMP_RELATION;
1334 : }
1335 : else
1336 : {
1337 350061 : io_context = IOContextForStrategy(strategy);
1338 350061 : io_object = IOOBJECT_RELATION;
1339 : }
1340 :
1341 350111 : buffer = PinBufferForBlock(rel, smgr, persistence,
1342 : forkNum, blockNum, strategy,
1343 : io_object, io_context, &found);
1344 350111 : ZeroAndLockBuffer(buffer, mode, found);
1345 350111 : return buffer;
1346 : }
1347 :
1348 : /*
1349 : * Signal that we are going to immediately wait. If we're immediately
1350 : * waiting, there is no benefit in actually executing the IO
1351 : * asynchronously, it would just add dispatch overhead.
1352 : */
1353 80457195 : flags = READ_BUFFERS_SYNCHRONOUSLY;
1354 80457195 : if (mode == RBM_ZERO_ON_ERROR)
1355 2002390 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1356 80457195 : operation.smgr = smgr;
1357 80457195 : operation.rel = rel;
1358 80457195 : operation.persistence = persistence;
1359 80457195 : operation.forknum = forkNum;
1360 80457195 : operation.strategy = strategy;
1361 80457195 : if (StartReadBuffer(&operation,
1362 : &buffer,
1363 : blockNum,
1364 : flags))
1365 790953 : WaitReadBuffers(&operation);
1366 :
1367 80457172 : return buffer;
1368 : }
1369 :
1370 : static pg_attribute_always_inline bool
1371 85383351 : StartReadBuffersImpl(ReadBuffersOperation *operation,
1372 : Buffer *buffers,
1373 : BlockNumber blockNum,
1374 : int *nblocks,
1375 : int flags,
1376 : bool allow_forwarding)
1377 : {
1378 85383351 : int actual_nblocks = *nblocks;
1379 85383351 : int maxcombine = 0;
1380 : bool did_start_io;
1381 : IOContext io_context;
1382 : IOObject io_object;
1383 :
1384 : Assert(*nblocks == 1 || allow_forwarding);
1385 : Assert(*nblocks > 0);
1386 : Assert(*nblocks <= MAX_IO_COMBINE_LIMIT);
1387 :
1388 : /* see comments in ReadBuffer_common */
1389 85383351 : if (operation->rel && RELATION_IS_OTHER_TEMP(operation->rel))
1390 0 : ereport(ERROR,
1391 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1392 : errmsg("cannot access temporary tables of other sessions")));
1393 :
1394 85383351 : if (operation->persistence == RELPERSISTENCE_TEMP)
1395 : {
1396 1640361 : io_context = IOCONTEXT_NORMAL;
1397 1640361 : io_object = IOOBJECT_TEMP_RELATION;
1398 : }
1399 : else
1400 : {
1401 83742990 : io_context = IOContextForStrategy(operation->strategy);
1402 83742990 : io_object = IOOBJECT_RELATION;
1403 : }
1404 :
1405 87072774 : for (int i = 0; i < actual_nblocks; ++i)
1406 : {
1407 : bool found;
1408 :
1409 85566079 : if (allow_forwarding && buffers[i] != InvalidBuffer)
1410 2370 : {
1411 : BufferDesc *bufHdr;
1412 :
1413 : /*
1414 : * This is a buffer that was pinned by an earlier call to
1415 : * StartReadBuffers(), but couldn't be handled in one operation at
1416 : * that time. The operation was split, and the caller has passed
1417 : * an already pinned buffer back to us to handle the rest of the
1418 : * operation. It must continue at the expected block number.
1419 : */
1420 : Assert(BufferGetBlockNumber(buffers[i]) == blockNum + i);
1421 :
1422 : /*
1423 : * It might be an already valid buffer (a hit) that followed the
1424 : * final contiguous block of an earlier I/O (a miss) marking the
1425 : * end of it, or a buffer that some other backend has since made
1426 : * valid by performing the I/O for us, in which case we can handle
1427 : * it as a hit now. It is safe to check for a BM_VALID flag with
1428 : * a relaxed load, because we got a fresh view of it while pinning
1429 : * it in the previous call.
1430 : *
1431 : * On the other hand if we don't see BM_VALID yet, it must be an
1432 : * I/O that was split by the previous call and we need to try to
1433 : * start a new I/O from this block. We're also racing against any
1434 : * other backend that might start the I/O or even manage to mark
1435 : * it BM_VALID after this check, but StartBufferIO() will handle
1436 : * those cases.
1437 : */
1438 2370 : if (BufferIsLocal(buffers[i]))
1439 16 : bufHdr = GetLocalBufferDescriptor(-buffers[i] - 1);
1440 : else
1441 2354 : bufHdr = GetBufferDescriptor(buffers[i] - 1);
1442 : Assert(pg_atomic_read_u64(&bufHdr->state) & BM_TAG_VALID);
1443 2370 : found = pg_atomic_read_u64(&bufHdr->state) & BM_VALID;
1444 : }
1445 : else
1446 : {
1447 85563701 : buffers[i] = PinBufferForBlock(operation->rel,
1448 : operation->smgr,
1449 85563709 : operation->persistence,
1450 : operation->forknum,
1451 : blockNum + i,
1452 : operation->strategy,
1453 : io_object, io_context,
1454 : &found);
1455 : }
1456 :
1457 85566071 : if (found)
1458 : {
1459 : /*
1460 : * We have a hit. If it's the first block in the requested range,
1461 : * we can return it immediately and report that WaitReadBuffers()
1462 : * does not need to be called. If the initial value of *nblocks
1463 : * was larger, the caller will have to call again for the rest.
1464 : */
1465 83876648 : if (i == 0)
1466 : {
1467 83874290 : *nblocks = 1;
1468 :
1469 : #ifdef USE_ASSERT_CHECKING
1470 :
1471 : /*
1472 : * Initialize enough of ReadBuffersOperation to make
1473 : * CheckReadBuffersOperation() work. Outside of assertions
1474 : * that's not necessary when no IO is issued.
1475 : */
1476 : operation->buffers = buffers;
1477 : operation->blocknum = blockNum;
1478 : operation->nblocks = 1;
1479 : operation->nblocks_done = 1;
1480 : CheckReadBuffersOperation(operation, true);
1481 : #endif
1482 83874290 : return false;
1483 : }
1484 :
1485 : /*
1486 : * Otherwise we already have an I/O to perform, but this block
1487 : * can't be included as it is already valid. Split the I/O here.
1488 : * There may or may not be more blocks requiring I/O after this
1489 : * one, we haven't checked, but they can't be contiguous with this
1490 : * one in the way. We'll leave this buffer pinned, forwarding it
1491 : * to the next call, avoiding the need to unpin it here and re-pin
1492 : * it in the next call.
1493 : */
1494 2358 : actual_nblocks = i;
1495 2358 : break;
1496 : }
1497 : else
1498 : {
1499 : /*
1500 : * Check how many blocks we can cover with the same IO. The smgr
1501 : * implementation might e.g. be limited due to a segment boundary.
1502 : */
1503 1689423 : if (i == 0 && actual_nblocks > 1)
1504 : {
1505 37486 : maxcombine = smgrmaxcombine(operation->smgr,
1506 : operation->forknum,
1507 : blockNum);
1508 37486 : if (unlikely(maxcombine < actual_nblocks))
1509 : {
1510 0 : elog(DEBUG2, "limiting nblocks at %u from %u to %u",
1511 : blockNum, actual_nblocks, maxcombine);
1512 0 : actual_nblocks = maxcombine;
1513 : }
1514 : }
1515 : }
1516 : }
1517 1509053 : *nblocks = actual_nblocks;
1518 :
1519 : /* Populate information needed for I/O. */
1520 1509053 : operation->buffers = buffers;
1521 1509053 : operation->blocknum = blockNum;
1522 1509053 : operation->flags = flags;
1523 1509053 : operation->nblocks = actual_nblocks;
1524 1509053 : operation->nblocks_done = 0;
1525 1509053 : pgaio_wref_clear(&operation->io_wref);
1526 :
1527 : /*
1528 : * When using AIO, start the IO in the background. If not, issue prefetch
1529 : * requests if desired by the caller.
1530 : *
1531 : * The reason we have a dedicated path for IOMETHOD_SYNC here is to
1532 : * de-risk the introduction of AIO somewhat. It's a large architectural
1533 : * change, with lots of chances for unanticipated performance effects.
1534 : *
1535 : * Use of IOMETHOD_SYNC already leads to not actually performing IO
1536 : * asynchronously, but without the check here we'd execute IO earlier than
1537 : * we used to. Eventually this IOMETHOD_SYNC specific path should go away.
1538 : */
1539 1509053 : if (io_method != IOMETHOD_SYNC)
1540 : {
1541 : /*
1542 : * Try to start IO asynchronously. It's possible that no IO needs to
1543 : * be started, if another backend already performed the IO.
1544 : *
1545 : * Note that if an IO is started, it might not cover the entire
1546 : * requested range, e.g. because an intermediary block has been read
1547 : * in by another backend. In that case any "trailing" buffers we
1548 : * already pinned above will be "forwarded" by read_stream.c to the
1549 : * next call to StartReadBuffers().
1550 : *
1551 : * This is signalled to the caller by decrementing *nblocks *and*
1552 : * reducing operation->nblocks. The latter is done here, but not below
1553 : * WaitReadBuffers(), as in WaitReadBuffers() we can't "shorten" the
1554 : * overall read size anymore, we need to retry until done in its
1555 : * entirety or until failed.
1556 : */
1557 1507669 : did_start_io = AsyncReadBuffers(operation, nblocks);
1558 :
1559 1507654 : operation->nblocks = *nblocks;
1560 : }
1561 : else
1562 : {
1563 1384 : operation->flags |= READ_BUFFERS_SYNCHRONOUSLY;
1564 :
1565 1384 : if (flags & READ_BUFFERS_ISSUE_ADVICE)
1566 : {
1567 : /*
1568 : * In theory we should only do this if PinBufferForBlock() had to
1569 : * allocate new buffers above. That way, if two calls to
1570 : * StartReadBuffers() were made for the same blocks before
1571 : * WaitReadBuffers(), only the first would issue the advice.
1572 : * That'd be a better simulation of true asynchronous I/O, which
1573 : * would only start the I/O once, but isn't done here for
1574 : * simplicity.
1575 : */
1576 19 : smgrprefetch(operation->smgr,
1577 : operation->forknum,
1578 : blockNum,
1579 : actual_nblocks);
1580 : }
1581 :
1582 : /*
1583 : * Indicate that WaitReadBuffers() should be called. WaitReadBuffers()
1584 : * will initiate the necessary IO.
1585 : */
1586 1384 : did_start_io = true;
1587 : }
1588 :
1589 1509038 : CheckReadBuffersOperation(operation, !did_start_io);
1590 :
1591 1509038 : return did_start_io;
1592 : }
1593 :
1594 : /*
1595 : * Begin reading a range of blocks beginning at blockNum and extending for
1596 : * *nblocks. *nblocks and the buffers array are in/out parameters. On entry,
1597 : * the buffers elements covered by *nblocks must hold either InvalidBuffer or
1598 : * buffers forwarded by an earlier call to StartReadBuffers() that was split
1599 : * and is now being continued. On return, *nblocks holds the number of blocks
1600 : * accepted by this operation. If it is less than the original number then
1601 : * this operation has been split, but buffer elements up to the original
1602 : * requested size may hold forwarded buffers to be used for a continuing
1603 : * operation. The caller must either start a new I/O beginning at the block
1604 : * immediately following the blocks accepted by this call and pass those
1605 : * buffers back in, or release them if it chooses not to. It shouldn't make
1606 : * any other use of or assumptions about forwarded buffers.
1607 : *
1608 : * If false is returned, no I/O is necessary and the buffers covered by
1609 : * *nblocks on exit are valid and ready to be accessed. If true is returned,
1610 : * an I/O has been started, and WaitReadBuffers() must be called with the same
1611 : * operation object before the buffers covered by *nblocks on exit can be
1612 : * accessed. Along with the operation object, the caller-supplied array of
1613 : * buffers must remain valid until WaitReadBuffers() is called, and any
1614 : * forwarded buffers must also be preserved for a continuing call unless
1615 : * they are explicitly released.
1616 : */
1617 : bool
1618 2209036 : StartReadBuffers(ReadBuffersOperation *operation,
1619 : Buffer *buffers,
1620 : BlockNumber blockNum,
1621 : int *nblocks,
1622 : int flags)
1623 : {
1624 2209036 : return StartReadBuffersImpl(operation, buffers, blockNum, nblocks, flags,
1625 : true /* expect forwarded buffers */ );
1626 : }
1627 :
1628 : /*
1629 : * Single block version of the StartReadBuffers(). This might save a few
1630 : * instructions when called from another translation unit, because it is
1631 : * specialized for nblocks == 1.
1632 : *
1633 : * This version does not support "forwarded" buffers: they cannot be created
1634 : * by reading only one block and *buffer is ignored on entry.
1635 : */
1636 : bool
1637 83174315 : StartReadBuffer(ReadBuffersOperation *operation,
1638 : Buffer *buffer,
1639 : BlockNumber blocknum,
1640 : int flags)
1641 : {
1642 83174315 : int nblocks = 1;
1643 : bool result;
1644 :
1645 83174315 : result = StartReadBuffersImpl(operation, buffer, blocknum, &nblocks, flags,
1646 : false /* single block, no forwarding */ );
1647 : Assert(nblocks == 1); /* single block can't be short */
1648 :
1649 83174300 : return result;
1650 : }
1651 :
1652 : /*
1653 : * Perform sanity checks on the ReadBuffersOperation.
1654 : */
1655 : static void
1656 4526964 : CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete)
1657 : {
1658 : #ifdef USE_ASSERT_CHECKING
1659 : Assert(operation->nblocks_done <= operation->nblocks);
1660 : Assert(!is_complete || operation->nblocks == operation->nblocks_done);
1661 :
1662 : for (int i = 0; i < operation->nblocks; i++)
1663 : {
1664 : Buffer buffer = operation->buffers[i];
1665 : BufferDesc *buf_hdr = BufferIsLocal(buffer) ?
1666 : GetLocalBufferDescriptor(-buffer - 1) :
1667 : GetBufferDescriptor(buffer - 1);
1668 :
1669 : Assert(BufferGetBlockNumber(buffer) == operation->blocknum + i);
1670 : Assert(pg_atomic_read_u64(&buf_hdr->state) & BM_TAG_VALID);
1671 :
1672 : if (i < operation->nblocks_done)
1673 : Assert(pg_atomic_read_u64(&buf_hdr->state) & BM_VALID);
1674 : }
1675 : #endif
1676 4526964 : }
1677 :
1678 : /*
1679 : * We track various stats related to buffer hits. Because this is done in a
1680 : * few separate places, this helper exists for convenience.
1681 : */
1682 : static pg_attribute_always_inline void
1683 83914793 : TrackBufferHit(IOObject io_object, IOContext io_context,
1684 : Relation rel, char persistence, SMgrRelation smgr,
1685 : ForkNumber forknum, BlockNumber blocknum)
1686 : {
1687 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forknum,
1688 : blocknum,
1689 : smgr->smgr_rlocator.locator.spcOid,
1690 : smgr->smgr_rlocator.locator.dbOid,
1691 : smgr->smgr_rlocator.locator.relNumber,
1692 : smgr->smgr_rlocator.backend,
1693 : true);
1694 :
1695 83914793 : if (persistence == RELPERSISTENCE_TEMP)
1696 1637554 : pgBufferUsage.local_blks_hit += 1;
1697 : else
1698 82277239 : pgBufferUsage.shared_blks_hit += 1;
1699 :
1700 83914793 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1701 :
1702 83914793 : if (VacuumCostActive)
1703 2987951 : VacuumCostBalance += VacuumCostPageHit;
1704 :
1705 83914793 : if (rel)
1706 78200794 : pgstat_count_buffer_hit(rel);
1707 83914793 : }
1708 :
1709 : /*
1710 : * Helper for WaitReadBuffers() that processes the results of a readv
1711 : * operation, raising an error if necessary.
1712 : */
1713 : static void
1714 1506201 : ProcessReadBuffersResult(ReadBuffersOperation *operation)
1715 : {
1716 1506201 : PgAioReturn *aio_ret = &operation->io_return;
1717 1506201 : PgAioResultStatus rs = aio_ret->result.status;
1718 1506201 : int newly_read_blocks = 0;
1719 :
1720 : Assert(pgaio_wref_valid(&operation->io_wref));
1721 : Assert(aio_ret->result.status != PGAIO_RS_UNKNOWN);
1722 :
1723 : /*
1724 : * SMGR reports the number of blocks successfully read as the result of
1725 : * the IO operation. Thus we can simply add that to ->nblocks_done.
1726 : */
1727 :
1728 1506201 : if (likely(rs != PGAIO_RS_ERROR))
1729 1506172 : newly_read_blocks = aio_ret->result.result;
1730 :
1731 1506201 : if (rs == PGAIO_RS_ERROR || rs == PGAIO_RS_WARNING)
1732 45 : pgaio_result_report(aio_ret->result, &aio_ret->target_data,
1733 : rs == PGAIO_RS_ERROR ? ERROR : WARNING);
1734 1506156 : else if (aio_ret->result.status == PGAIO_RS_PARTIAL)
1735 : {
1736 : /*
1737 : * We'll retry, so we just emit a debug message to the server log (or
1738 : * not even that in prod scenarios).
1739 : */
1740 109 : pgaio_result_report(aio_ret->result, &aio_ret->target_data, DEBUG1);
1741 109 : elog(DEBUG3, "partial read, will retry");
1742 : }
1743 :
1744 : Assert(newly_read_blocks > 0);
1745 : Assert(newly_read_blocks <= MAX_IO_COMBINE_LIMIT);
1746 :
1747 1506172 : operation->nblocks_done += newly_read_blocks;
1748 :
1749 : Assert(operation->nblocks_done <= operation->nblocks);
1750 1506172 : }
1751 :
1752 : /*
1753 : * Wait for the IO operation initiated by StartReadBuffers() et al to
1754 : * complete.
1755 : *
1756 : * Returns true if we needed to wait for the IO operation, false otherwise.
1757 : */
1758 : bool
1759 1508229 : WaitReadBuffers(ReadBuffersOperation *operation)
1760 : {
1761 1508229 : PgAioReturn *aio_ret = &operation->io_return;
1762 : IOContext io_context;
1763 : IOObject io_object;
1764 1508229 : bool needed_wait = false;
1765 :
1766 1508229 : if (operation->persistence == RELPERSISTENCE_TEMP)
1767 : {
1768 2429 : io_context = IOCONTEXT_NORMAL;
1769 2429 : io_object = IOOBJECT_TEMP_RELATION;
1770 : }
1771 : else
1772 : {
1773 1505800 : io_context = IOContextForStrategy(operation->strategy);
1774 1505800 : io_object = IOOBJECT_RELATION;
1775 : }
1776 :
1777 : /*
1778 : * If we get here without an IO operation having been issued, the
1779 : * io_method == IOMETHOD_SYNC path must have been used. Otherwise the
1780 : * caller should not have called WaitReadBuffers().
1781 : *
1782 : * In the case of IOMETHOD_SYNC, we start - as we used to before the
1783 : * introducing of AIO - the IO in WaitReadBuffers(). This is done as part
1784 : * of the retry logic below, no extra code is required.
1785 : *
1786 : * This path is expected to eventually go away.
1787 : */
1788 1508229 : if (!pgaio_wref_valid(&operation->io_wref) && io_method != IOMETHOD_SYNC)
1789 0 : elog(ERROR, "waiting for read operation that didn't read");
1790 :
1791 : /*
1792 : * To handle partial reads, and IOMETHOD_SYNC, we re-issue IO until we're
1793 : * done. We may need multiple retries, not just because we could get
1794 : * multiple partial reads, but also because some of the remaining
1795 : * to-be-read buffers may have been read in by other backends, limiting
1796 : * the IO size.
1797 : */
1798 : while (true)
1799 1498 : {
1800 : int ignored_nblocks_progress;
1801 :
1802 1509727 : CheckReadBuffersOperation(operation, false);
1803 :
1804 : /*
1805 : * If there is an IO associated with the operation, we may need to
1806 : * wait for it.
1807 : */
1808 1509727 : if (pgaio_wref_valid(&operation->io_wref))
1809 : {
1810 : /*
1811 : * Track the time spent waiting for the IO to complete. As
1812 : * tracking a wait even if we don't actually need to wait
1813 : *
1814 : * a) is not cheap, due to the timestamping overhead
1815 : *
1816 : * b) reports some time as waiting, even if we never waited
1817 : *
1818 : * we first check if we already know the IO is complete.
1819 : *
1820 : * Note that operation->io_return is uninitialized for foreign IO,
1821 : * so we cannot use the cheaper PGAIO_RS_UNKNOWN pre-check.
1822 : */
1823 1508334 : if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
1824 691108 : !pgaio_wref_check_done(&operation->io_wref))
1825 : {
1826 388087 : instr_time io_start = pgstat_prepare_io_time(track_io_timing);
1827 :
1828 388087 : pgaio_wref_wait(&operation->io_wref);
1829 388086 : needed_wait = true;
1830 :
1831 : /*
1832 : * The IO operation itself was already counted earlier, in
1833 : * AsyncReadBuffers(), this just accounts for the wait time.
1834 : */
1835 388086 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1836 : io_start, 0, 0);
1837 : }
1838 : else
1839 : {
1840 : Assert(pgaio_wref_check_done(&operation->io_wref));
1841 : }
1842 :
1843 1508333 : if (unlikely(operation->foreign_io))
1844 : {
1845 2132 : Buffer buffer = operation->buffers[operation->nblocks_done];
1846 2132 : BufferDesc *desc = BufferIsLocal(buffer) ?
1847 2132 : GetLocalBufferDescriptor(-buffer - 1) :
1848 2132 : GetBufferDescriptor(buffer - 1);
1849 2132 : uint64 buf_state = pg_atomic_read_u64(&desc->state);
1850 :
1851 2132 : if (buf_state & BM_VALID)
1852 : {
1853 2130 : BlockNumber blocknum = operation->blocknum + operation->nblocks_done;
1854 :
1855 2130 : operation->nblocks_done += 1;
1856 : Assert(operation->nblocks_done <= operation->nblocks);
1857 :
1858 : /*
1859 : * Track this as a 'hit' for this backend. The backend
1860 : * performing the IO will track it as a 'read'.
1861 : */
1862 2130 : TrackBufferHit(io_object, io_context,
1863 2130 : operation->rel, operation->persistence,
1864 : operation->smgr, operation->forknum,
1865 : blocknum);
1866 : }
1867 :
1868 : /*
1869 : * If the foreign IO failed and left the buffer invalid,
1870 : * nblocks_done is not incremented. The retry loop below will
1871 : * call AsyncReadBuffers() which will attempt the IO itself.
1872 : */
1873 : }
1874 : else
1875 : {
1876 : /*
1877 : * We now are sure the IO completed. Check the results. This
1878 : * includes reporting on errors if there were any.
1879 : */
1880 1506201 : ProcessReadBuffersResult(operation);
1881 : }
1882 : }
1883 :
1884 : /*
1885 : * Most of the time, the one IO we already started, will read in
1886 : * everything. But we need to deal with partial reads and buffers not
1887 : * needing IO anymore.
1888 : */
1889 1509697 : if (operation->nblocks_done == operation->nblocks)
1890 1508199 : break;
1891 :
1892 1498 : CHECK_FOR_INTERRUPTS();
1893 :
1894 : /*
1895 : * If the IO completed only partially, we need to perform additional
1896 : * work, consider that a form of having had to wait.
1897 : */
1898 1498 : needed_wait = true;
1899 :
1900 : /*
1901 : * This may only complete the IO partially, either because some
1902 : * buffers were already valid, or because of a partial read.
1903 : *
1904 : * NB: In contrast to after the AsyncReadBuffers() call in
1905 : * StartReadBuffers(), we do *not* reduce
1906 : * ReadBuffersOperation->nblocks here, callers expect the full
1907 : * operation to be completed at this point (as more operations may
1908 : * have been queued).
1909 : */
1910 1498 : AsyncReadBuffers(operation, &ignored_nblocks_progress);
1911 : }
1912 :
1913 1508199 : CheckReadBuffersOperation(operation, true);
1914 :
1915 : /* NB: READ_DONE tracepoint was already executed in completion callback */
1916 1508199 : return needed_wait;
1917 : }
1918 :
1919 : /*
1920 : * Initiate IO for the ReadBuffersOperation
1921 : *
1922 : * This function only starts a single IO at a time. The size of the IO may be
1923 : * limited to below the to-be-read blocks, if one of the buffers has
1924 : * concurrently been read in. If the first to-be-read buffer is already valid,
1925 : * no IO will be issued.
1926 : *
1927 : * To support retries after partial reads, the first operation->nblocks_done
1928 : * buffers are skipped.
1929 : *
1930 : * On return *nblocks_progress is updated to reflect the number of buffers
1931 : * affected by the call. If the first buffer is valid, *nblocks_progress is
1932 : * set to 1 and operation->nblocks_done is incremented.
1933 : *
1934 : * Returns true if IO was initiated or is already in progress (foreign IO),
1935 : * false if the buffer was already valid.
1936 : */
1937 : static bool
1938 1509173 : AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
1939 : {
1940 1509173 : Buffer *buffers = &operation->buffers[0];
1941 1509173 : int flags = operation->flags;
1942 1509173 : ForkNumber forknum = operation->forknum;
1943 1509173 : char persistence = operation->persistence;
1944 1509173 : int16 nblocks_done = operation->nblocks_done;
1945 1509173 : BlockNumber blocknum = operation->blocknum + nblocks_done;
1946 1509173 : Buffer *io_buffers = &operation->buffers[nblocks_done];
1947 1509173 : int io_buffers_len = 0;
1948 : PgAioHandle *ioh;
1949 1509173 : uint32 ioh_flags = 0;
1950 : void *io_pages[MAX_IO_COMBINE_LIMIT];
1951 : IOContext io_context;
1952 : IOObject io_object;
1953 : instr_time io_start;
1954 : StartBufferIOResult status;
1955 :
1956 1509173 : if (persistence == RELPERSISTENCE_TEMP)
1957 : {
1958 2821 : io_context = IOCONTEXT_NORMAL;
1959 2821 : io_object = IOOBJECT_TEMP_RELATION;
1960 : }
1961 : else
1962 : {
1963 1506352 : io_context = IOContextForStrategy(operation->strategy);
1964 1506352 : io_object = IOOBJECT_RELATION;
1965 : }
1966 :
1967 : /*
1968 : * When this IO is executed synchronously, either because the caller will
1969 : * immediately block waiting for the IO or because IOMETHOD_SYNC is used,
1970 : * the AIO subsystem needs to know.
1971 : */
1972 1509173 : if (flags & READ_BUFFERS_SYNCHRONOUSLY)
1973 809116 : ioh_flags |= PGAIO_HF_SYNCHRONOUS;
1974 :
1975 1509173 : if (persistence == RELPERSISTENCE_TEMP)
1976 2821 : ioh_flags |= PGAIO_HF_REFERENCES_LOCAL;
1977 :
1978 : /*
1979 : * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
1980 : * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
1981 : * set globally, but on a per-session basis. The completion callback,
1982 : * which may be run in other processes, e.g. in IO workers, may have a
1983 : * different value of the zero_damaged_pages GUC.
1984 : *
1985 : * XXX: We probably should eventually use a different flag for
1986 : * zero_damaged_pages, so we can report different log levels / error codes
1987 : * for zero_damaged_pages and ZERO_ON_ERROR.
1988 : */
1989 1509173 : if (zero_damaged_pages)
1990 16 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1991 :
1992 : /*
1993 : * For the same reason as with zero_damaged_pages we need to use this
1994 : * backend's ignore_checksum_failure value.
1995 : */
1996 1509173 : if (ignore_checksum_failure)
1997 8 : flags |= READ_BUFFERS_IGNORE_CHECKSUM_FAILURES;
1998 :
1999 :
2000 : /*
2001 : * To be allowed to report stats in the local completion callback we need
2002 : * to prepare to report stats now. This ensures we can safely report the
2003 : * checksum failure even in a critical section.
2004 : */
2005 1509173 : pgstat_prepare_report_checksum_failure(operation->smgr->smgr_rlocator.locator.dbOid);
2006 :
2007 : /*
2008 : * We must get an IO handle before StartBufferIO(), as pgaio_io_acquire()
2009 : * might block, which we don't want after setting IO_IN_PROGRESS. If we
2010 : * don't need to do the IO, we'll release the handle.
2011 : *
2012 : * If we need to wait for IO before we can get a handle, submit
2013 : * already-staged IO first, so that other backends don't need to wait.
2014 : * There wouldn't be a deadlock risk, as pgaio_io_acquire() just needs to
2015 : * wait for already submitted IO, which doesn't require additional locks,
2016 : * but it could still cause undesirable waits.
2017 : *
2018 : * A secondary benefit is that this would allow us to measure the time in
2019 : * pgaio_io_acquire() without causing undue timer overhead in the common,
2020 : * non-blocking, case. However, currently the pgstats infrastructure
2021 : * doesn't really allow that, as it a) asserts that an operation can't
2022 : * have time without operations b) doesn't have an API to report
2023 : * "accumulated" time.
2024 : */
2025 1509173 : ioh = pgaio_io_acquire_nb(CurrentResourceOwner, &operation->io_return);
2026 1509173 : if (unlikely(!ioh))
2027 : {
2028 3404 : pgaio_submit_staged();
2029 3404 : ioh = pgaio_io_acquire(CurrentResourceOwner, &operation->io_return);
2030 : }
2031 :
2032 1509173 : operation->foreign_io = false;
2033 1509173 : pgaio_wref_clear(&operation->io_wref);
2034 :
2035 : /*
2036 : * Try to start IO on the first buffer in a new run of blocks. If AIO is
2037 : * in progress, be it in this backend or another backend, we just
2038 : * associate the wait reference with the operation and wait in
2039 : * WaitReadBuffers(). This turns out to be important for performance in
2040 : * two workloads:
2041 : *
2042 : * 1) A read stream that has to read the same block multiple times within
2043 : * the readahead distance. This can happen e.g. for the table accesses of
2044 : * an index scan.
2045 : *
2046 : * 2) Concurrent scans by multiple backends on the same relation.
2047 : *
2048 : * If we were to synchronously wait for the in-progress IO, we'd not be
2049 : * able to keep enough I/O in flight.
2050 : *
2051 : * If we do find there is ongoing I/O for the buffer, we set up a 1-block
2052 : * ReadBuffersOperation that WaitReadBuffers then can wait on.
2053 : *
2054 : * It's possible that another backend has started IO on the buffer but not
2055 : * yet set its wait reference. In this case, we have no choice but to wait
2056 : * for either the wait reference to be valid or the IO to be done.
2057 : */
2058 1509173 : status = StartBufferIO(buffers[nblocks_done], true, true,
2059 : &operation->io_wref);
2060 1509173 : if (status != BUFFER_IO_READY_FOR_IO)
2061 : {
2062 2554 : pgaio_io_release(ioh);
2063 2554 : *nblocks_progress = 1;
2064 2554 : if (status == BUFFER_IO_ALREADY_DONE)
2065 : {
2066 : /*
2067 : * Someone has already completed this block, we're done.
2068 : *
2069 : * When IO is necessary, ->nblocks_done is updated in
2070 : * ProcessReadBuffersResult(), but that is not called if no IO is
2071 : * necessary. Thus update here.
2072 : */
2073 422 : operation->nblocks_done += 1;
2074 : Assert(operation->nblocks_done <= operation->nblocks);
2075 :
2076 : Assert(!pgaio_wref_valid(&operation->io_wref));
2077 :
2078 : /*
2079 : * Report and track this as a 'hit' for this backend, even though
2080 : * it must have started out as a miss in PinBufferForBlock(). The
2081 : * other backend will track this as a 'read'.
2082 : */
2083 422 : TrackBufferHit(io_object, io_context,
2084 422 : operation->rel, operation->persistence,
2085 : operation->smgr, operation->forknum,
2086 : blocknum);
2087 422 : return false;
2088 : }
2089 :
2090 : /* The IO is already in-progress */
2091 : Assert(status == BUFFER_IO_IN_PROGRESS);
2092 : Assert(pgaio_wref_valid(&operation->io_wref));
2093 2132 : operation->foreign_io = true;
2094 :
2095 2132 : return true;
2096 : }
2097 :
2098 : Assert(io_buffers[0] == buffers[nblocks_done]);
2099 1506619 : io_pages[0] = BufferGetBlock(buffers[nblocks_done]);
2100 1506613 : io_buffers_len = 1;
2101 :
2102 : /*
2103 : * NB: As little code as possible should be added between the
2104 : * StartBufferIO() above, the further StartBufferIO()s below and the
2105 : * smgrstartreadv(), as some of the buffers are now marked as
2106 : * IO_IN_PROGRESS and will thus cause other backends to wait.
2107 : */
2108 :
2109 : /*
2110 : * How many neighboring-on-disk blocks can we scatter-read into other
2111 : * buffers at the same time? In this case we don't wait if we see an I/O
2112 : * already in progress (see comment above).
2113 : */
2114 1687616 : for (int i = nblocks_done + 1; i < operation->nblocks; i++)
2115 : {
2116 : /* Must be consecutive block numbers. */
2117 : Assert(BufferGetBlockNumber(buffers[i - 1]) ==
2118 : BufferGetBlockNumber(buffers[i]) - 1);
2119 :
2120 181007 : status = StartBufferIO(buffers[i], true, false, NULL);
2121 181007 : if (status != BUFFER_IO_READY_FOR_IO)
2122 4 : break;
2123 :
2124 : Assert(io_buffers[io_buffers_len] == buffers[i]);
2125 :
2126 181003 : io_pages[io_buffers_len++] = BufferGetBlock(buffers[i]);
2127 : }
2128 :
2129 : /* get a reference to wait for in WaitReadBuffers() */
2130 1506613 : pgaio_io_get_wref(ioh, &operation->io_wref);
2131 :
2132 : /* provide the list of buffers to the completion callbacks */
2133 1506613 : pgaio_io_set_handle_data_32(ioh, (uint32 *) io_buffers, io_buffers_len);
2134 :
2135 1506613 : pgaio_io_register_callbacks(ioh,
2136 : persistence == RELPERSISTENCE_TEMP ?
2137 : PGAIO_HCB_LOCAL_BUFFER_READV :
2138 : PGAIO_HCB_SHARED_BUFFER_READV,
2139 : flags);
2140 :
2141 1506613 : pgaio_io_set_flag(ioh, ioh_flags);
2142 :
2143 : /* ---
2144 : * Even though we're trying to issue IO asynchronously, track the time
2145 : * in smgrstartreadv():
2146 : * - if io_method == IOMETHOD_SYNC, we will always perform the IO
2147 : * immediately
2148 : * - the io method might not support the IO (e.g. worker IO for a temp
2149 : * table)
2150 : * ---
2151 : */
2152 1506613 : io_start = pgstat_prepare_io_time(track_io_timing);
2153 1506613 : smgrstartreadv(ioh, operation->smgr, forknum,
2154 : blocknum,
2155 : io_pages, io_buffers_len);
2156 1506598 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
2157 1506598 : io_start, 1, io_buffers_len * BLCKSZ);
2158 :
2159 1506598 : if (persistence == RELPERSISTENCE_TEMP)
2160 2819 : pgBufferUsage.local_blks_read += io_buffers_len;
2161 : else
2162 1503779 : pgBufferUsage.shared_blks_read += io_buffers_len;
2163 :
2164 : /*
2165 : * Track vacuum cost when issuing IO, not after waiting for it. Otherwise
2166 : * we could end up issuing a lot of IO in a short timespan, despite a low
2167 : * cost limit.
2168 : */
2169 1506598 : if (VacuumCostActive)
2170 20925 : VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
2171 :
2172 1506598 : *nblocks_progress = io_buffers_len;
2173 :
2174 1506598 : return true;
2175 : }
2176 :
2177 : /*
2178 : * BufferAlloc -- subroutine for PinBufferForBlock. Handles lookup of a shared
2179 : * buffer. If no buffer exists already, selects a replacement victim and
2180 : * evicts the old page, but does NOT read in new page.
2181 : *
2182 : * "strategy" can be a buffer replacement strategy object, or NULL for
2183 : * the default strategy. The selected buffer's usage_count is advanced when
2184 : * using the default strategy, but otherwise possibly not (see PinBuffer).
2185 : *
2186 : * The returned buffer is pinned and is already marked as holding the
2187 : * desired page. If it already did have the desired page, *foundPtr is
2188 : * set true. Otherwise, *foundPtr is set false.
2189 : *
2190 : * io_context is passed as an output parameter to avoid calling
2191 : * IOContextForStrategy() when there is a shared buffers hit and no IO
2192 : * statistics need be captured.
2193 : *
2194 : * No locks are held either at entry or exit.
2195 : */
2196 : static pg_attribute_always_inline BufferDesc *
2197 84265305 : BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
2198 : BlockNumber blockNum,
2199 : BufferAccessStrategy strategy,
2200 : bool *foundPtr, IOContext io_context)
2201 : {
2202 : BufferTag newTag; /* identity of requested block */
2203 : uint32 newHash; /* hash value for newTag */
2204 : LWLock *newPartitionLock; /* buffer partition lock for it */
2205 : int existing_buf_id;
2206 : Buffer victim_buffer;
2207 : BufferDesc *victim_buf_hdr;
2208 : uint64 victim_buf_state;
2209 84265305 : uint64 set_bits = 0;
2210 :
2211 : /* Make sure we will have room to remember the buffer pin */
2212 84265305 : ResourceOwnerEnlarge(CurrentResourceOwner);
2213 84265305 : ReservePrivateRefCountEntry();
2214 :
2215 : /* create a tag so we can lookup the buffer */
2216 84265305 : InitBufferTag(&newTag, &smgr->smgr_rlocator.locator, forkNum, blockNum);
2217 :
2218 : /* determine its hash code and partition lock ID */
2219 84265305 : newHash = BufTableHashCode(&newTag);
2220 84265305 : newPartitionLock = BufMappingPartitionLock(newHash);
2221 :
2222 : /* see if the block is in the buffer pool already */
2223 84265305 : LWLockAcquire(newPartitionLock, LW_SHARED);
2224 84265305 : existing_buf_id = BufTableLookup(&newTag, newHash);
2225 84265305 : if (existing_buf_id >= 0)
2226 : {
2227 : BufferDesc *buf;
2228 : bool valid;
2229 :
2230 : /*
2231 : * Found it. Now, pin the buffer so no one can steal it from the
2232 : * buffer pool, and check to see if the correct data has been loaded
2233 : * into the buffer.
2234 : */
2235 82276790 : buf = GetBufferDescriptor(existing_buf_id);
2236 :
2237 82276790 : valid = PinBuffer(buf, strategy, false);
2238 :
2239 : /* Can release the mapping lock as soon as we've pinned it */
2240 82276790 : LWLockRelease(newPartitionLock);
2241 :
2242 82276790 : *foundPtr = true;
2243 :
2244 82276790 : if (!valid)
2245 : {
2246 : /*
2247 : * We can only get here if (a) someone else is still reading in
2248 : * the page, (b) a previous read attempt failed, or (c) someone
2249 : * called StartReadBuffers() but not yet WaitReadBuffers().
2250 : */
2251 2314 : *foundPtr = false;
2252 : }
2253 :
2254 82276790 : return buf;
2255 : }
2256 :
2257 : /*
2258 : * Didn't find it in the buffer pool. We'll have to initialize a new
2259 : * buffer. Remember to unlock the mapping lock while doing the work.
2260 : */
2261 1988515 : LWLockRelease(newPartitionLock);
2262 :
2263 : /*
2264 : * Acquire a victim buffer. Somebody else might try to do the same, we
2265 : * don't hold any conflicting locks. If so we'll have to undo our work
2266 : * later.
2267 : */
2268 1988515 : victim_buffer = GetVictimBuffer(strategy, io_context);
2269 1988515 : victim_buf_hdr = GetBufferDescriptor(victim_buffer - 1);
2270 :
2271 : /*
2272 : * Try to make a hashtable entry for the buffer under its new tag. If
2273 : * somebody else inserted another buffer for the tag, we'll release the
2274 : * victim buffer we acquired and use the already inserted one.
2275 : */
2276 1988515 : LWLockAcquire(newPartitionLock, LW_EXCLUSIVE);
2277 1988515 : existing_buf_id = BufTableInsert(&newTag, newHash, victim_buf_hdr->buf_id);
2278 1988515 : if (existing_buf_id >= 0)
2279 : {
2280 : BufferDesc *existing_buf_hdr;
2281 : bool valid;
2282 :
2283 : /*
2284 : * Got a collision. Someone has already done what we were about to do.
2285 : * We'll just handle this as if it were found in the buffer pool in
2286 : * the first place. First, give up the buffer we were planning to
2287 : * use.
2288 : *
2289 : * We could do this after releasing the partition lock, but then we'd
2290 : * have to call ResourceOwnerEnlarge() & ReservePrivateRefCountEntry()
2291 : * before acquiring the lock, for the rare case of such a collision.
2292 : */
2293 587 : UnpinBuffer(victim_buf_hdr);
2294 :
2295 : /* remaining code should match code at top of routine */
2296 :
2297 587 : existing_buf_hdr = GetBufferDescriptor(existing_buf_id);
2298 :
2299 587 : valid = PinBuffer(existing_buf_hdr, strategy, false);
2300 :
2301 : /* Can release the mapping lock as soon as we've pinned it */
2302 587 : LWLockRelease(newPartitionLock);
2303 :
2304 587 : *foundPtr = true;
2305 :
2306 587 : if (!valid)
2307 : {
2308 : /*
2309 : * We can only get here if (a) someone else is still reading in
2310 : * the page, (b) a previous read attempt failed, or (c) someone
2311 : * called StartReadBuffers() but not yet WaitReadBuffers().
2312 : */
2313 288 : *foundPtr = false;
2314 : }
2315 :
2316 587 : return existing_buf_hdr;
2317 : }
2318 :
2319 : /*
2320 : * Need to lock the buffer header too in order to change its tag.
2321 : */
2322 1987928 : victim_buf_state = LockBufHdr(victim_buf_hdr);
2323 :
2324 : /* some sanity checks while we hold the buffer header lock */
2325 : Assert(BUF_STATE_GET_REFCOUNT(victim_buf_state) == 1);
2326 : Assert(!(victim_buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY | BM_IO_IN_PROGRESS)));
2327 :
2328 1987928 : victim_buf_hdr->tag = newTag;
2329 :
2330 : /*
2331 : * Make sure BM_PERMANENT is set for buffers that must be written at every
2332 : * checkpoint. Unlogged buffers only need to be written at shutdown
2333 : * checkpoints, except for their "init" forks, which need to be treated
2334 : * just like permanent relations.
2335 : */
2336 1987928 : set_bits |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2337 1987928 : if (relpersistence == RELPERSISTENCE_PERMANENT || forkNum == INIT_FORKNUM)
2338 1987525 : set_bits |= BM_PERMANENT;
2339 :
2340 1987928 : UnlockBufHdrExt(victim_buf_hdr, victim_buf_state,
2341 : set_bits, 0, 0);
2342 :
2343 1987928 : LWLockRelease(newPartitionLock);
2344 :
2345 : /*
2346 : * Buffer contents are currently invalid.
2347 : */
2348 1987928 : *foundPtr = false;
2349 :
2350 1987928 : return victim_buf_hdr;
2351 : }
2352 :
2353 : /*
2354 : * InvalidateBuffer -- mark a shared buffer invalid.
2355 : *
2356 : * The buffer header spinlock must be held at entry. We drop it before
2357 : * returning. (This is sane because the caller must have locked the
2358 : * buffer in order to be sure it should be dropped.)
2359 : *
2360 : * This is used only in contexts such as dropping a relation. We assume
2361 : * that no other backend could possibly be interested in using the page,
2362 : * so the only reason the buffer might be pinned is if someone else is
2363 : * trying to write it out. We have to let them finish before we can
2364 : * reclaim the buffer.
2365 : *
2366 : * The buffer could get reclaimed by someone else while we are waiting
2367 : * to acquire the necessary locks; if so, don't mess it up.
2368 : */
2369 : static void
2370 126281 : InvalidateBuffer(BufferDesc *buf)
2371 : {
2372 : BufferTag oldTag;
2373 : uint32 oldHash; /* hash value for oldTag */
2374 : LWLock *oldPartitionLock; /* buffer partition lock for it */
2375 : uint32 oldFlags;
2376 : uint64 buf_state;
2377 :
2378 : /* Save the original buffer tag before dropping the spinlock */
2379 126281 : oldTag = buf->tag;
2380 :
2381 126281 : UnlockBufHdr(buf);
2382 :
2383 : /*
2384 : * Need to compute the old tag's hashcode and partition lock ID. XXX is it
2385 : * worth storing the hashcode in BufferDesc so we need not recompute it
2386 : * here? Probably not.
2387 : */
2388 126281 : oldHash = BufTableHashCode(&oldTag);
2389 126281 : oldPartitionLock = BufMappingPartitionLock(oldHash);
2390 :
2391 126282 : retry:
2392 :
2393 : /*
2394 : * Acquire exclusive mapping lock in preparation for changing the buffer's
2395 : * association.
2396 : */
2397 126282 : LWLockAcquire(oldPartitionLock, LW_EXCLUSIVE);
2398 :
2399 : /* Re-lock the buffer header */
2400 126282 : buf_state = LockBufHdr(buf);
2401 :
2402 : /* If it's changed while we were waiting for lock, do nothing */
2403 126282 : if (!BufferTagsEqual(&buf->tag, &oldTag))
2404 : {
2405 0 : UnlockBufHdr(buf);
2406 0 : LWLockRelease(oldPartitionLock);
2407 0 : return;
2408 : }
2409 :
2410 : /*
2411 : * We assume the reason for it to be pinned is that either we were
2412 : * asynchronously reading the page in before erroring out or someone else
2413 : * is flushing the page out. Wait for the IO to finish. (This could be
2414 : * an infinite loop if the refcount is messed up... it would be nice to
2415 : * time out after awhile, but there seems no way to be sure how many loops
2416 : * may be needed. Note that if the other guy has pinned the buffer but
2417 : * not yet done StartBufferIO, WaitIO will fall through and we'll
2418 : * effectively be busy-looping here.)
2419 : */
2420 126282 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 0)
2421 : {
2422 1 : UnlockBufHdr(buf);
2423 1 : LWLockRelease(oldPartitionLock);
2424 : /* safety check: should definitely not be our *own* pin */
2425 1 : if (GetPrivateRefCount(BufferDescriptorGetBuffer(buf)) > 0)
2426 0 : elog(ERROR, "buffer is pinned in InvalidateBuffer");
2427 1 : WaitIO(buf);
2428 1 : goto retry;
2429 : }
2430 :
2431 : /*
2432 : * An invalidated buffer should not have any backends waiting to lock the
2433 : * buffer, therefore BM_LOCK_WAKE_IN_PROGRESS should not be set.
2434 : */
2435 : Assert(!(buf_state & BM_LOCK_WAKE_IN_PROGRESS));
2436 :
2437 : /*
2438 : * Clear out the buffer's tag and flags. We must do this to ensure that
2439 : * linear scans of the buffer array don't think the buffer is valid.
2440 : */
2441 126281 : oldFlags = buf_state & BUF_FLAG_MASK;
2442 126281 : ClearBufferTag(&buf->tag);
2443 :
2444 126281 : UnlockBufHdrExt(buf, buf_state,
2445 : 0,
2446 : BUF_FLAG_MASK | BUF_USAGECOUNT_MASK,
2447 : 0);
2448 :
2449 : /*
2450 : * Remove the buffer from the lookup hashtable, if it was in there.
2451 : */
2452 126281 : if (oldFlags & BM_TAG_VALID)
2453 126281 : BufTableDelete(&oldTag, oldHash);
2454 :
2455 : /*
2456 : * Done with mapping lock.
2457 : */
2458 126281 : LWLockRelease(oldPartitionLock);
2459 : }
2460 :
2461 : /*
2462 : * Helper routine for GetVictimBuffer()
2463 : *
2464 : * Needs to be called on a buffer with a valid tag, pinned, but without the
2465 : * buffer header spinlock held.
2466 : *
2467 : * Returns true if the buffer can be reused, in which case the buffer is only
2468 : * pinned by this backend and marked as invalid, false otherwise.
2469 : */
2470 : static bool
2471 1443438 : InvalidateVictimBuffer(BufferDesc *buf_hdr)
2472 : {
2473 : uint64 buf_state;
2474 : uint32 hash;
2475 : LWLock *partition_lock;
2476 : BufferTag tag;
2477 :
2478 : Assert(GetPrivateRefCount(BufferDescriptorGetBuffer(buf_hdr)) == 1);
2479 :
2480 : /* have buffer pinned, so it's safe to read tag without lock */
2481 1443438 : tag = buf_hdr->tag;
2482 :
2483 1443438 : hash = BufTableHashCode(&tag);
2484 1443438 : partition_lock = BufMappingPartitionLock(hash);
2485 :
2486 1443438 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2487 :
2488 : /* lock the buffer header */
2489 1443438 : buf_state = LockBufHdr(buf_hdr);
2490 :
2491 : /*
2492 : * We have the buffer pinned nobody else should have been able to unset
2493 : * this concurrently.
2494 : */
2495 : Assert(buf_state & BM_TAG_VALID);
2496 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2497 : Assert(BufferTagsEqual(&buf_hdr->tag, &tag));
2498 :
2499 : /*
2500 : * If somebody else pinned the buffer since, or even worse, dirtied it,
2501 : * give up on this buffer: It's clearly in use.
2502 : */
2503 1443438 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 1 || (buf_state & BM_DIRTY))
2504 : {
2505 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2506 :
2507 549 : UnlockBufHdr(buf_hdr);
2508 549 : LWLockRelease(partition_lock);
2509 :
2510 549 : return false;
2511 : }
2512 :
2513 : /*
2514 : * An invalidated buffer should not have any backends waiting to lock the
2515 : * buffer, therefore BM_LOCK_WAKE_IN_PROGRESS should not be set.
2516 : */
2517 : Assert(!(buf_state & BM_LOCK_WAKE_IN_PROGRESS));
2518 :
2519 : /*
2520 : * Clear out the buffer's tag and flags and usagecount. This is not
2521 : * strictly required, as BM_TAG_VALID/BM_VALID needs to be checked before
2522 : * doing anything with the buffer. But currently it's beneficial, as the
2523 : * cheaper pre-check for several linear scans of shared buffers use the
2524 : * tag (see e.g. FlushDatabaseBuffers()).
2525 : */
2526 1442889 : ClearBufferTag(&buf_hdr->tag);
2527 1442889 : UnlockBufHdrExt(buf_hdr, buf_state,
2528 : 0,
2529 : BUF_FLAG_MASK | BUF_USAGECOUNT_MASK,
2530 : 0);
2531 :
2532 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2533 :
2534 : /* finally delete buffer from the buffer mapping table */
2535 1442889 : BufTableDelete(&tag, hash);
2536 :
2537 1442889 : LWLockRelease(partition_lock);
2538 :
2539 1442889 : buf_state = pg_atomic_read_u64(&buf_hdr->state);
2540 : Assert(!(buf_state & (BM_DIRTY | BM_VALID | BM_TAG_VALID)));
2541 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2542 : Assert(BUF_STATE_GET_REFCOUNT(pg_atomic_read_u64(&buf_hdr->state)) > 0);
2543 :
2544 1442889 : return true;
2545 : }
2546 :
2547 : static Buffer
2548 2267885 : GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context)
2549 : {
2550 : BufferDesc *buf_hdr;
2551 : Buffer buf;
2552 : uint64 buf_state;
2553 : bool from_ring;
2554 :
2555 : /*
2556 : * Ensure, before we pin a victim buffer, that there's a free refcount
2557 : * entry and resource owner slot for the pin.
2558 : */
2559 2267885 : ReservePrivateRefCountEntry();
2560 2267885 : ResourceOwnerEnlarge(CurrentResourceOwner);
2561 :
2562 : /* we return here if a prospective victim buffer gets used concurrently */
2563 23091 : again:
2564 :
2565 : /*
2566 : * Select a victim buffer. The buffer is returned pinned and owned by
2567 : * this backend.
2568 : */
2569 2290976 : buf_hdr = StrategyGetBuffer(strategy, &buf_state, &from_ring);
2570 2290976 : buf = BufferDescriptorGetBuffer(buf_hdr);
2571 :
2572 : /*
2573 : * We shouldn't have any other pins for this buffer.
2574 : */
2575 2290976 : CheckBufferIsPinnedOnce(buf);
2576 :
2577 : /*
2578 : * If the buffer was dirty, try to write it out. There is a race
2579 : * condition here, another backend could dirty the buffer between
2580 : * StrategyGetBuffer() checking that it is not in use and invalidating the
2581 : * buffer below. That's addressed by InvalidateVictimBuffer() verifying
2582 : * that the buffer is not dirty.
2583 : */
2584 2290976 : if (buf_state & BM_DIRTY)
2585 : {
2586 : Assert(buf_state & BM_TAG_VALID);
2587 : Assert(buf_state & BM_VALID);
2588 :
2589 : /*
2590 : * We need a share-exclusive lock on the buffer contents to write it
2591 : * out (else we might write invalid data, eg because someone else is
2592 : * compacting the page contents while we write). We must use a
2593 : * conditional lock acquisition here to avoid deadlock. Even though
2594 : * the buffer was not pinned (and therefore surely not locked) when
2595 : * StrategyGetBuffer returned it, someone else could have pinned and
2596 : * (share-)exclusive-locked it by the time we get here. If we try to
2597 : * get the lock unconditionally, we'd block waiting for them; if they
2598 : * later block waiting for us, deadlock ensues. (This has been
2599 : * observed to happen when two backends are both trying to split btree
2600 : * index pages, and the second one just happens to be trying to split
2601 : * the page the first one got from StrategyGetBuffer.)
2602 : */
2603 375835 : if (!BufferLockConditional(buf, buf_hdr, BUFFER_LOCK_SHARE_EXCLUSIVE))
2604 : {
2605 : /*
2606 : * Someone else has locked the buffer, so give it up and loop back
2607 : * to get another one.
2608 : */
2609 0 : UnpinBuffer(buf_hdr);
2610 0 : goto again;
2611 : }
2612 :
2613 : /*
2614 : * If using a nondefault strategy, and this victim came from the
2615 : * strategy ring, let the strategy decide whether to reject it when
2616 : * reusing it would require a WAL flush. This only applies to
2617 : * permanent buffers; unlogged buffers can have fake LSNs, so
2618 : * XLogNeedsFlush() is not meaningful for them.
2619 : *
2620 : * We need to hold the content lock in at least share-exclusive mode
2621 : * to safely inspect the page LSN, so this couldn't have been done
2622 : * inside StrategyGetBuffer().
2623 : */
2624 375835 : if (strategy && from_ring &&
2625 191252 : buf_state & BM_PERMANENT &&
2626 122653 : XLogNeedsFlush(BufferGetLSN(buf_hdr)) &&
2627 27045 : StrategyRejectBuffer(strategy, buf_hdr, from_ring))
2628 : {
2629 22542 : UnlockReleaseBuffer(buf);
2630 22542 : goto again;
2631 : }
2632 :
2633 : /* OK, do the I/O */
2634 353293 : FlushBuffer(buf_hdr, NULL, IOOBJECT_RELATION, io_context);
2635 353293 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
2636 :
2637 353293 : ScheduleBufferTagForWriteback(&BackendWritebackContext, io_context,
2638 : &buf_hdr->tag);
2639 : }
2640 :
2641 :
2642 2268434 : if (buf_state & BM_VALID)
2643 : {
2644 : /*
2645 : * When a BufferAccessStrategy is in use, blocks evicted from shared
2646 : * buffers are counted as IOOP_EVICT in the corresponding context
2647 : * (e.g. IOCONTEXT_BULKWRITE). Shared buffers are evicted by a
2648 : * strategy in two cases: 1) while initially claiming buffers for the
2649 : * strategy ring 2) to replace an existing strategy ring buffer
2650 : * because it is pinned or in use and cannot be reused.
2651 : *
2652 : * Blocks evicted from buffers already in the strategy ring are
2653 : * counted as IOOP_REUSE in the corresponding strategy context.
2654 : *
2655 : * At this point, we can accurately count evictions and reuses,
2656 : * because we have successfully claimed the valid buffer. Previously,
2657 : * we may have been forced to release the buffer due to concurrent
2658 : * pinners or erroring out.
2659 : */
2660 1440929 : pgstat_count_io_op(IOOBJECT_RELATION, io_context,
2661 1440929 : from_ring ? IOOP_REUSE : IOOP_EVICT, 1, 0);
2662 : }
2663 :
2664 : /*
2665 : * If the buffer has an entry in the buffer mapping table, delete it. This
2666 : * can fail because another backend could have pinned or dirtied the
2667 : * buffer.
2668 : */
2669 2268434 : if ((buf_state & BM_TAG_VALID) && !InvalidateVictimBuffer(buf_hdr))
2670 : {
2671 549 : UnpinBuffer(buf_hdr);
2672 549 : goto again;
2673 : }
2674 :
2675 : /* a final set of sanity checks */
2676 : #ifdef USE_ASSERT_CHECKING
2677 : buf_state = pg_atomic_read_u64(&buf_hdr->state);
2678 :
2679 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2680 : Assert(!(buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY)));
2681 :
2682 : CheckBufferIsPinnedOnce(buf);
2683 : #endif
2684 :
2685 2267885 : return buf;
2686 : }
2687 :
2688 : /*
2689 : * Return the maximum number of buffers that a backend should try to pin once,
2690 : * to avoid exceeding its fair share. This is the highest value that
2691 : * GetAdditionalPinLimit() could ever return. Note that it may be zero on a
2692 : * system with a very small buffer pool relative to max_connections.
2693 : */
2694 : uint32
2695 769294 : GetPinLimit(void)
2696 : {
2697 769294 : return MaxProportionalPins;
2698 : }
2699 :
2700 : /*
2701 : * Return the maximum number of additional buffers that this backend should
2702 : * pin if it wants to stay under the per-backend limit, considering the number
2703 : * of buffers it has already pinned. Unlike LimitAdditionalPins(), the limit
2704 : * return by this function can be zero.
2705 : */
2706 : uint32
2707 4400704 : GetAdditionalPinLimit(void)
2708 : {
2709 : uint32 estimated_pins_held;
2710 :
2711 : /*
2712 : * We get the number of "overflowed" pins for free, but don't know the
2713 : * number of pins in PrivateRefCountArray. The cost of calculating that
2714 : * exactly doesn't seem worth it, so just assume the max.
2715 : */
2716 4400704 : estimated_pins_held = PrivateRefCountOverflowed + REFCOUNT_ARRAY_ENTRIES;
2717 :
2718 : /* Is this backend already holding more than its fair share? */
2719 4400704 : if (estimated_pins_held > MaxProportionalPins)
2720 1526760 : return 0;
2721 :
2722 2873944 : return MaxProportionalPins - estimated_pins_held;
2723 : }
2724 :
2725 : /*
2726 : * Limit the number of pins a batch operation may additionally acquire, to
2727 : * avoid running out of pinnable buffers.
2728 : *
2729 : * One additional pin is always allowed, on the assumption that the operation
2730 : * requires at least one to make progress.
2731 : */
2732 : void
2733 252469 : LimitAdditionalPins(uint32 *additional_pins)
2734 : {
2735 : uint32 limit;
2736 :
2737 252469 : if (*additional_pins <= 1)
2738 236891 : return;
2739 :
2740 15578 : limit = GetAdditionalPinLimit();
2741 15578 : limit = Max(limit, 1);
2742 15578 : if (limit < *additional_pins)
2743 10044 : *additional_pins = limit;
2744 : }
2745 :
2746 : /*
2747 : * Logic shared between ExtendBufferedRelBy(), ExtendBufferedRelTo(). Just to
2748 : * avoid duplicating the tracing and relpersistence related logic.
2749 : */
2750 : static BlockNumber
2751 267430 : ExtendBufferedRelCommon(BufferManagerRelation bmr,
2752 : ForkNumber fork,
2753 : BufferAccessStrategy strategy,
2754 : uint32 flags,
2755 : uint32 extend_by,
2756 : BlockNumber extend_upto,
2757 : Buffer *buffers,
2758 : uint32 *extended_by)
2759 : {
2760 : BlockNumber first_block;
2761 :
2762 : TRACE_POSTGRESQL_BUFFER_EXTEND_START(fork,
2763 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2764 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2765 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2766 : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2767 : extend_by);
2768 :
2769 267430 : if (bmr.relpersistence == RELPERSISTENCE_TEMP)
2770 14961 : first_block = ExtendBufferedRelLocal(bmr, fork, flags,
2771 : extend_by, extend_upto,
2772 : buffers, &extend_by);
2773 : else
2774 252469 : first_block = ExtendBufferedRelShared(bmr, fork, strategy, flags,
2775 : extend_by, extend_upto,
2776 : buffers, &extend_by);
2777 267430 : *extended_by = extend_by;
2778 :
2779 : TRACE_POSTGRESQL_BUFFER_EXTEND_DONE(fork,
2780 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2781 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2782 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2783 : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2784 : *extended_by,
2785 : first_block);
2786 :
2787 267430 : return first_block;
2788 : }
2789 :
2790 : /*
2791 : * Implementation of ExtendBufferedRelBy() and ExtendBufferedRelTo() for
2792 : * shared buffers.
2793 : */
2794 : static BlockNumber
2795 252469 : ExtendBufferedRelShared(BufferManagerRelation bmr,
2796 : ForkNumber fork,
2797 : BufferAccessStrategy strategy,
2798 : uint32 flags,
2799 : uint32 extend_by,
2800 : BlockNumber extend_upto,
2801 : Buffer *buffers,
2802 : uint32 *extended_by)
2803 : {
2804 : BlockNumber first_block;
2805 252469 : IOContext io_context = IOContextForStrategy(strategy);
2806 : instr_time io_start;
2807 :
2808 252469 : LimitAdditionalPins(&extend_by);
2809 :
2810 : /*
2811 : * Acquire victim buffers for extension without holding extension lock.
2812 : * Writing out victim buffers is the most expensive part of extending the
2813 : * relation, particularly when doing so requires WAL flushes. Zeroing out
2814 : * the buffers is also quite expensive, so do that before holding the
2815 : * extension lock as well.
2816 : *
2817 : * These pages are pinned by us and not valid. While we hold the pin they
2818 : * can't be acquired as victim buffers by another backend.
2819 : */
2820 531839 : for (uint32 i = 0; i < extend_by; i++)
2821 : {
2822 : Block buf_block;
2823 :
2824 279370 : buffers[i] = GetVictimBuffer(strategy, io_context);
2825 279370 : buf_block = BufHdrGetBlock(GetBufferDescriptor(buffers[i] - 1));
2826 :
2827 : /* new buffers are zero-filled */
2828 279370 : MemSet(buf_block, 0, BLCKSZ);
2829 : }
2830 :
2831 : /*
2832 : * Lock relation against concurrent extensions, unless requested not to.
2833 : *
2834 : * We use the same extension lock for all forks. That's unnecessarily
2835 : * restrictive, but currently extensions for forks don't happen often
2836 : * enough to make it worth locking more granularly.
2837 : *
2838 : * Note that another backend might have extended the relation by the time
2839 : * we get the lock.
2840 : */
2841 252469 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2842 197048 : LockRelationForExtension(bmr.rel, ExclusiveLock);
2843 :
2844 : /*
2845 : * If requested, invalidate size cache, so that smgrnblocks asks the
2846 : * kernel.
2847 : */
2848 252469 : if (flags & EB_CLEAR_SIZE_CACHE)
2849 9950 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
2850 :
2851 252469 : first_block = smgrnblocks(BMR_GET_SMGR(bmr), fork);
2852 :
2853 : /*
2854 : * Now that we have the accurate relation size, check if the caller wants
2855 : * us to extend to only up to a specific size. If there were concurrent
2856 : * extensions, we might have acquired too many buffers and need to release
2857 : * them.
2858 : */
2859 252469 : if (extend_upto != InvalidBlockNumber)
2860 : {
2861 57539 : uint32 orig_extend_by = extend_by;
2862 :
2863 57539 : if (first_block > extend_upto)
2864 0 : extend_by = 0;
2865 57539 : else if ((uint64) first_block + extend_by > extend_upto)
2866 2 : extend_by = extend_upto - first_block;
2867 :
2868 57543 : for (uint32 i = extend_by; i < orig_extend_by; i++)
2869 : {
2870 4 : BufferDesc *buf_hdr = GetBufferDescriptor(buffers[i] - 1);
2871 :
2872 4 : UnpinBuffer(buf_hdr);
2873 : }
2874 :
2875 57539 : if (extend_by == 0)
2876 : {
2877 2 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2878 2 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2879 2 : *extended_by = extend_by;
2880 2 : return first_block;
2881 : }
2882 : }
2883 :
2884 : /* Fail if relation is already at maximum possible length */
2885 252467 : if ((uint64) first_block + extend_by >= MaxBlockNumber)
2886 0 : ereport(ERROR,
2887 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
2888 : errmsg("cannot extend relation %s beyond %u blocks",
2889 : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str,
2890 : MaxBlockNumber)));
2891 :
2892 : /*
2893 : * Insert buffers into buffer table, mark as IO_IN_PROGRESS.
2894 : *
2895 : * This needs to happen before we extend the relation, because as soon as
2896 : * we do, other backends can start to read in those pages.
2897 : */
2898 531833 : for (uint32 i = 0; i < extend_by; i++)
2899 : {
2900 279366 : Buffer victim_buf = buffers[i];
2901 279366 : BufferDesc *victim_buf_hdr = GetBufferDescriptor(victim_buf - 1);
2902 : BufferTag tag;
2903 : uint32 hash;
2904 : LWLock *partition_lock;
2905 : int existing_id;
2906 :
2907 : /* in case we need to pin an existing buffer below */
2908 279366 : ResourceOwnerEnlarge(CurrentResourceOwner);
2909 279366 : ReservePrivateRefCountEntry();
2910 :
2911 279366 : InitBufferTag(&tag, &BMR_GET_SMGR(bmr)->smgr_rlocator.locator, fork,
2912 : first_block + i);
2913 279366 : hash = BufTableHashCode(&tag);
2914 279366 : partition_lock = BufMappingPartitionLock(hash);
2915 :
2916 279366 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2917 :
2918 279366 : existing_id = BufTableInsert(&tag, hash, victim_buf_hdr->buf_id);
2919 :
2920 : /*
2921 : * We get here only in the corner case where we are trying to extend
2922 : * the relation but we found a pre-existing buffer. This can happen
2923 : * because a prior attempt at extending the relation failed, and
2924 : * because mdread doesn't complain about reads beyond EOF (when
2925 : * zero_damaged_pages is ON) and so a previous attempt to read a block
2926 : * beyond EOF could have left a "valid" zero-filled buffer.
2927 : *
2928 : * This has also been observed when relation was overwritten by
2929 : * external process. Since the legitimate cases should always have
2930 : * left a zero-filled buffer, complain if not PageIsNew.
2931 : */
2932 279366 : if (existing_id >= 0)
2933 : {
2934 0 : BufferDesc *existing_hdr = GetBufferDescriptor(existing_id);
2935 : Block buf_block;
2936 : bool valid;
2937 :
2938 : /*
2939 : * Pin the existing buffer before releasing the partition lock,
2940 : * preventing it from being evicted.
2941 : */
2942 0 : valid = PinBuffer(existing_hdr, strategy, false);
2943 :
2944 0 : LWLockRelease(partition_lock);
2945 0 : UnpinBuffer(victim_buf_hdr);
2946 :
2947 0 : buffers[i] = BufferDescriptorGetBuffer(existing_hdr);
2948 0 : buf_block = BufHdrGetBlock(existing_hdr);
2949 :
2950 0 : if (valid && !PageIsNew((Page) buf_block))
2951 0 : ereport(ERROR,
2952 : (errmsg("unexpected data beyond EOF in block %u of relation \"%s\"",
2953 : existing_hdr->tag.blockNum,
2954 : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str)));
2955 :
2956 : /*
2957 : * We *must* do smgr[zero]extend before succeeding, else the page
2958 : * will not be reserved by the kernel, and the next P_NEW call
2959 : * will decide to return the same page. Clear the BM_VALID bit,
2960 : * do StartSharedBufferIO() and proceed.
2961 : *
2962 : * Loop to handle the very small possibility that someone re-sets
2963 : * BM_VALID between our clearing it and StartSharedBufferIO
2964 : * inspecting it.
2965 : */
2966 : while (true)
2967 0 : {
2968 : StartBufferIOResult sbres;
2969 :
2970 0 : pg_atomic_fetch_and_u64(&existing_hdr->state, ~BM_VALID);
2971 :
2972 0 : sbres = StartSharedBufferIO(existing_hdr, true, true, NULL);
2973 :
2974 0 : if (sbres != BUFFER_IO_ALREADY_DONE)
2975 0 : break;
2976 : }
2977 : }
2978 : else
2979 : {
2980 : uint64 buf_state;
2981 279366 : uint64 set_bits = 0;
2982 :
2983 279366 : buf_state = LockBufHdr(victim_buf_hdr);
2984 :
2985 : /* some sanity checks while we hold the buffer header lock */
2986 : Assert(!(buf_state & (BM_VALID | BM_TAG_VALID | BM_DIRTY)));
2987 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2988 :
2989 279366 : victim_buf_hdr->tag = tag;
2990 :
2991 279366 : set_bits |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2992 279366 : if (bmr.relpersistence == RELPERSISTENCE_PERMANENT || fork == INIT_FORKNUM)
2993 273942 : set_bits |= BM_PERMANENT;
2994 :
2995 279366 : UnlockBufHdrExt(victim_buf_hdr, buf_state,
2996 : set_bits, 0,
2997 : 0);
2998 :
2999 279366 : LWLockRelease(partition_lock);
3000 :
3001 : /* XXX: could combine the locked operations in it with the above */
3002 279366 : StartSharedBufferIO(victim_buf_hdr, true, true, NULL);
3003 : }
3004 : }
3005 :
3006 252467 : io_start = pgstat_prepare_io_time(track_io_timing);
3007 :
3008 : /*
3009 : * Note: if smgrzeroextend fails, we will end up with buffers that are
3010 : * allocated but not marked BM_VALID. The next relation extension will
3011 : * still select the same block number (because the relation didn't get any
3012 : * longer on disk) and so future attempts to extend the relation will find
3013 : * the same buffers (if they have not been recycled) but come right back
3014 : * here to try smgrzeroextend again.
3015 : *
3016 : * We don't need to set checksum for all-zero pages.
3017 : */
3018 252467 : smgrzeroextend(BMR_GET_SMGR(bmr), fork, first_block, extend_by, false);
3019 :
3020 : /*
3021 : * Release the file-extension lock; it's now OK for someone else to extend
3022 : * the relation some more.
3023 : *
3024 : * We remove IO_IN_PROGRESS after this, as waking up waiting backends can
3025 : * take noticeable time.
3026 : */
3027 252467 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
3028 197046 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
3029 :
3030 252467 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context, IOOP_EXTEND,
3031 252467 : io_start, 1, extend_by * BLCKSZ);
3032 :
3033 : /* Set BM_VALID, terminate IO, and wake up any waiters */
3034 531833 : for (uint32 i = 0; i < extend_by; i++)
3035 : {
3036 279366 : Buffer buf = buffers[i];
3037 279366 : BufferDesc *buf_hdr = GetBufferDescriptor(buf - 1);
3038 279366 : bool lock = false;
3039 :
3040 279366 : if (flags & EB_LOCK_FIRST && i == 0)
3041 194602 : lock = true;
3042 84764 : else if (flags & EB_LOCK_TARGET)
3043 : {
3044 : Assert(extend_upto != InvalidBlockNumber);
3045 46433 : if (first_block + i + 1 == extend_upto)
3046 45869 : lock = true;
3047 : }
3048 :
3049 279366 : if (lock)
3050 240471 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
3051 :
3052 279366 : TerminateBufferIO(buf_hdr, false, BM_VALID, true, false);
3053 : }
3054 :
3055 252467 : pgBufferUsage.shared_blks_written += extend_by;
3056 :
3057 252467 : *extended_by = extend_by;
3058 :
3059 252467 : return first_block;
3060 : }
3061 :
3062 : /*
3063 : * BufferIsLockedByMe
3064 : *
3065 : * Checks if this backend has the buffer locked in any mode.
3066 : *
3067 : * Buffer must be pinned.
3068 : */
3069 : bool
3070 0 : BufferIsLockedByMe(Buffer buffer)
3071 : {
3072 : BufferDesc *bufHdr;
3073 :
3074 : Assert(BufferIsPinned(buffer));
3075 :
3076 0 : if (BufferIsLocal(buffer))
3077 : {
3078 : /* Content locks are not maintained for local buffers. */
3079 0 : return true;
3080 : }
3081 : else
3082 : {
3083 0 : bufHdr = GetBufferDescriptor(buffer - 1);
3084 0 : return BufferLockHeldByMe(bufHdr);
3085 : }
3086 : }
3087 :
3088 : /*
3089 : * BufferIsLockedByMeInMode
3090 : *
3091 : * Checks if this backend has the buffer locked in the specified mode.
3092 : *
3093 : * Buffer must be pinned.
3094 : */
3095 : bool
3096 0 : BufferIsLockedByMeInMode(Buffer buffer, BufferLockMode mode)
3097 : {
3098 : BufferDesc *bufHdr;
3099 :
3100 : Assert(BufferIsPinned(buffer));
3101 :
3102 0 : if (BufferIsLocal(buffer))
3103 : {
3104 : /* Content locks are not maintained for local buffers. */
3105 0 : return true;
3106 : }
3107 : else
3108 : {
3109 0 : bufHdr = GetBufferDescriptor(buffer - 1);
3110 0 : return BufferLockHeldByMeInMode(bufHdr, mode);
3111 : }
3112 : }
3113 :
3114 : /*
3115 : * BufferIsDirty
3116 : *
3117 : * Checks if buffer is already dirty.
3118 : *
3119 : * Buffer must be pinned and [share-]exclusive-locked. (Without such a lock,
3120 : * the result may be stale before it's returned.)
3121 : */
3122 : bool
3123 28923 : BufferIsDirty(Buffer buffer)
3124 : {
3125 : BufferDesc *bufHdr;
3126 :
3127 : Assert(BufferIsPinned(buffer));
3128 :
3129 28923 : if (BufferIsLocal(buffer))
3130 : {
3131 9523 : int bufid = -buffer - 1;
3132 :
3133 9523 : bufHdr = GetLocalBufferDescriptor(bufid);
3134 : /* Content locks are not maintained for local buffers. */
3135 : }
3136 : else
3137 : {
3138 19400 : bufHdr = GetBufferDescriptor(buffer - 1);
3139 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_SHARE_EXCLUSIVE) ||
3140 : BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
3141 : }
3142 :
3143 28923 : return pg_atomic_read_u64(&bufHdr->state) & BM_DIRTY;
3144 : }
3145 :
3146 : /*
3147 : * MarkBufferDirty
3148 : *
3149 : * Marks buffer contents as dirty (actual write happens later).
3150 : *
3151 : * Buffer must be pinned and exclusive-locked. (If caller does not hold
3152 : * exclusive lock, then somebody could be in process of writing the buffer,
3153 : * leading to risk of bad data written to disk.)
3154 : */
3155 : void
3156 34532696 : MarkBufferDirty(Buffer buffer)
3157 : {
3158 : BufferDesc *bufHdr;
3159 : uint64 buf_state;
3160 : uint64 old_buf_state;
3161 :
3162 34532696 : if (!BufferIsValid(buffer))
3163 0 : elog(ERROR, "bad buffer ID: %d", buffer);
3164 :
3165 34532696 : if (BufferIsLocal(buffer))
3166 : {
3167 1580083 : MarkLocalBufferDirty(buffer);
3168 1580083 : return;
3169 : }
3170 :
3171 32952613 : bufHdr = GetBufferDescriptor(buffer - 1);
3172 :
3173 : Assert(BufferIsPinned(buffer));
3174 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
3175 :
3176 : /*
3177 : * NB: We have to wait for the buffer header spinlock to be not held, as
3178 : * TerminateBufferIO() relies on the spinlock.
3179 : */
3180 32952613 : old_buf_state = pg_atomic_read_u64(&bufHdr->state);
3181 : for (;;)
3182 : {
3183 32952735 : if (old_buf_state & BM_LOCKED)
3184 208 : old_buf_state = WaitBufHdrUnlocked(bufHdr);
3185 :
3186 32952735 : buf_state = old_buf_state;
3187 :
3188 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
3189 32952735 : buf_state |= BM_DIRTY;
3190 :
3191 32952735 : if (pg_atomic_compare_exchange_u64(&bufHdr->state, &old_buf_state,
3192 : buf_state))
3193 32952613 : break;
3194 : }
3195 :
3196 : /*
3197 : * If the buffer was not dirty already, do vacuum accounting.
3198 : */
3199 32952613 : if (!(old_buf_state & BM_DIRTY))
3200 : {
3201 802266 : pgBufferUsage.shared_blks_dirtied++;
3202 802266 : if (VacuumCostActive)
3203 10009 : VacuumCostBalance += VacuumCostPageDirty;
3204 : }
3205 : }
3206 :
3207 : /*
3208 : * ReleaseAndReadBuffer -- combine ReleaseBuffer() and ReadBuffer()
3209 : *
3210 : * Formerly, this saved one cycle of acquiring/releasing the BufMgrLock
3211 : * compared to calling the two routines separately. Now it's mainly just
3212 : * a convenience function. However, if the passed buffer is valid and
3213 : * already contains the desired block, we just return it as-is; and that
3214 : * does save considerable work compared to a full release and reacquire.
3215 : *
3216 : * Note: it is OK to pass buffer == InvalidBuffer, indicating that no old
3217 : * buffer actually needs to be released. This case is the same as ReadBuffer,
3218 : * but can save some tests in the caller.
3219 : */
3220 : Buffer
3221 3646 : ReleaseAndReadBuffer(Buffer buffer,
3222 : Relation relation,
3223 : BlockNumber blockNum)
3224 : {
3225 3646 : ForkNumber forkNum = MAIN_FORKNUM;
3226 : BufferDesc *bufHdr;
3227 :
3228 3646 : if (BufferIsValid(buffer))
3229 : {
3230 : Assert(BufferIsPinned(buffer));
3231 3646 : if (BufferIsLocal(buffer))
3232 : {
3233 50 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
3234 50 : if (bufHdr->tag.blockNum == blockNum &&
3235 0 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3236 0 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3237 0 : return buffer;
3238 50 : UnpinLocalBuffer(buffer);
3239 : }
3240 : else
3241 : {
3242 3596 : bufHdr = GetBufferDescriptor(buffer - 1);
3243 : /* we have pin, so it's ok to examine tag without spinlock */
3244 3596 : if (bufHdr->tag.blockNum == blockNum &&
3245 0 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3246 0 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3247 0 : return buffer;
3248 3596 : UnpinBuffer(bufHdr);
3249 : }
3250 : }
3251 :
3252 3646 : return ReadBuffer(relation, blockNum);
3253 : }
3254 :
3255 : /*
3256 : * PinBuffer -- make buffer unavailable for replacement.
3257 : *
3258 : * For the default access strategy, the buffer's usage_count is incremented
3259 : * when we first pin it; for other strategies we just make sure the usage_count
3260 : * isn't zero. (The idea of the latter is that we don't want synchronized
3261 : * heap scans to inflate the count, but we need it to not be zero to discourage
3262 : * other backends from stealing buffers from our ring. As long as we cycle
3263 : * through the ring faster than the global clock-sweep cycles, buffers in
3264 : * our ring won't be chosen as victims for replacement by other backends.)
3265 : *
3266 : * This should be applied only to shared buffers, never local ones.
3267 : *
3268 : * Since buffers are pinned/unpinned very frequently, pin buffers without
3269 : * taking the buffer header lock; instead update the state variable in loop of
3270 : * CAS operations. Hopefully it's just a single CAS.
3271 : *
3272 : * Note that ResourceOwnerEnlarge() and ReservePrivateRefCountEntry()
3273 : * must have been done already.
3274 : *
3275 : * Returns true if buffer is BM_VALID, else false. This provision allows
3276 : * some callers to avoid an extra spinlock cycle. If skip_if_not_valid is
3277 : * true, then a false return value also indicates that the buffer was
3278 : * (recently) invalid and has not been pinned.
3279 : */
3280 : static bool
3281 82282290 : PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
3282 : bool skip_if_not_valid)
3283 : {
3284 82282290 : Buffer b = BufferDescriptorGetBuffer(buf);
3285 : bool result;
3286 : PrivateRefCountEntry *ref;
3287 :
3288 : Assert(!BufferIsLocal(b));
3289 : Assert(ReservedRefCountSlot != -1);
3290 :
3291 82282290 : ref = GetPrivateRefCountEntry(b, true);
3292 :
3293 82282290 : if (ref == NULL)
3294 : {
3295 : uint64 buf_state;
3296 : uint64 old_buf_state;
3297 :
3298 74930688 : old_buf_state = pg_atomic_read_u64(&buf->state);
3299 : for (;;)
3300 : {
3301 74940685 : if (unlikely(skip_if_not_valid && !(old_buf_state & BM_VALID)))
3302 6 : return false;
3303 :
3304 : /*
3305 : * We're not allowed to increase the refcount while the buffer
3306 : * header spinlock is held. Wait for the lock to be released.
3307 : */
3308 74940679 : if (unlikely(old_buf_state & BM_LOCKED))
3309 : {
3310 76 : old_buf_state = WaitBufHdrUnlocked(buf);
3311 :
3312 : /* perform checks at the top of the loop again */
3313 76 : continue;
3314 : }
3315 :
3316 74940603 : buf_state = old_buf_state;
3317 :
3318 : /* increase refcount */
3319 74940603 : buf_state += BUF_REFCOUNT_ONE;
3320 :
3321 74940603 : if (strategy == NULL)
3322 : {
3323 : /* Default case: increase usagecount unless already max. */
3324 74081292 : if (BUF_STATE_GET_USAGECOUNT(buf_state) < BM_MAX_USAGE_COUNT)
3325 3805483 : buf_state += BUF_USAGECOUNT_ONE;
3326 : }
3327 : else
3328 : {
3329 : /*
3330 : * Ring buffers shouldn't evict others from pool. Thus we
3331 : * don't make usagecount more than 1.
3332 : */
3333 859311 : if (BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3334 41268 : buf_state += BUF_USAGECOUNT_ONE;
3335 : }
3336 :
3337 74940603 : if (pg_atomic_compare_exchange_u64(&buf->state, &old_buf_state,
3338 : buf_state))
3339 : {
3340 74930682 : result = (buf_state & BM_VALID) != 0;
3341 :
3342 74930682 : TrackNewBufferPin(b);
3343 74930682 : break;
3344 : }
3345 : }
3346 : }
3347 : else
3348 : {
3349 : /*
3350 : * If we previously pinned the buffer, it is likely to be valid, but
3351 : * it may not be if StartReadBuffers() was called and
3352 : * WaitReadBuffers() hasn't been called yet. We'll check by loading
3353 : * the flags without locking. This is racy, but it's OK to return
3354 : * false spuriously: when WaitReadBuffers() calls StartBufferIO(),
3355 : * it'll see that it's now valid.
3356 : *
3357 : * Note: We deliberately avoid a Valgrind client request here.
3358 : * Individual access methods can optionally superimpose buffer page
3359 : * client requests on top of our client requests to enforce that
3360 : * buffers are only accessed while locked (and pinned). It's possible
3361 : * that the buffer page is legitimately non-accessible here. We
3362 : * cannot meddle with that.
3363 : */
3364 7351602 : result = (pg_atomic_read_u64(&buf->state) & BM_VALID) != 0;
3365 :
3366 : Assert(ref->data.refcount > 0);
3367 7351602 : ref->data.refcount++;
3368 7351602 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
3369 : }
3370 :
3371 82282284 : return result;
3372 : }
3373 :
3374 : /*
3375 : * PinBuffer_Locked -- as above, but caller already locked the buffer header.
3376 : * The spinlock is released before return.
3377 : *
3378 : * As this function is called with the spinlock held, the caller has to
3379 : * previously call ReservePrivateRefCountEntry() and
3380 : * ResourceOwnerEnlarge(CurrentResourceOwner);
3381 : *
3382 : * Currently, no callers of this function want to modify the buffer's
3383 : * usage_count at all, so there's no need for a strategy parameter.
3384 : * Also we don't bother with a BM_VALID test (the caller could check that for
3385 : * itself).
3386 : *
3387 : * Also all callers only ever use this function when it's known that the
3388 : * buffer can't have a preexisting pin by this backend. That allows us to skip
3389 : * searching the private refcount array & hash, which is a boon, because the
3390 : * spinlock is still held.
3391 : *
3392 : * Note: use of this routine is frequently mandatory, not just an optimization
3393 : * to save a spin lock/unlock cycle, because we need to pin a buffer before
3394 : * its state can change under us.
3395 : */
3396 : static void
3397 355159 : PinBuffer_Locked(BufferDesc *buf)
3398 : {
3399 : uint64 old_buf_state;
3400 :
3401 : /*
3402 : * As explained, We don't expect any preexisting pins. That allows us to
3403 : * manipulate the PrivateRefCount after releasing the spinlock
3404 : */
3405 : Assert(GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf), false) == NULL);
3406 :
3407 : /*
3408 : * Since we hold the buffer spinlock, we can update the buffer state and
3409 : * release the lock in one operation.
3410 : */
3411 355159 : old_buf_state = pg_atomic_read_u64(&buf->state);
3412 :
3413 355159 : UnlockBufHdrExt(buf, old_buf_state,
3414 : 0, 0, 1);
3415 :
3416 355159 : TrackNewBufferPin(BufferDescriptorGetBuffer(buf));
3417 355159 : }
3418 :
3419 : /*
3420 : * Support for waking up another backend that is waiting for the cleanup lock
3421 : * to be released using BM_PIN_COUNT_WAITER.
3422 : *
3423 : * See LockBufferForCleanup().
3424 : *
3425 : * Expected to be called just after releasing a buffer pin (in a BufferDesc,
3426 : * not just reducing the backend-local pincount for the buffer).
3427 : */
3428 : static void
3429 98 : WakePinCountWaiter(BufferDesc *buf)
3430 : {
3431 : /*
3432 : * Acquire the buffer header lock, re-check that there's a waiter. Another
3433 : * backend could have unpinned this buffer, and already woken up the
3434 : * waiter.
3435 : *
3436 : * There's no danger of the buffer being replaced after we unpinned it
3437 : * above, as it's pinned by the waiter. The waiter removes
3438 : * BM_PIN_COUNT_WAITER if it stops waiting for a reason other than this
3439 : * backend waking it up.
3440 : */
3441 98 : uint64 buf_state = LockBufHdr(buf);
3442 :
3443 98 : if ((buf_state & BM_PIN_COUNT_WAITER) &&
3444 98 : BUF_STATE_GET_REFCOUNT(buf_state) == 1)
3445 94 : {
3446 : /* we just released the last pin other than the waiter's */
3447 94 : int wait_backend_pgprocno = buf->wait_backend_pgprocno;
3448 :
3449 94 : UnlockBufHdrExt(buf, buf_state,
3450 : 0, BM_PIN_COUNT_WAITER,
3451 : 0);
3452 94 : ProcSendSignal(wait_backend_pgprocno);
3453 : }
3454 : else
3455 4 : UnlockBufHdr(buf);
3456 98 : }
3457 :
3458 : /*
3459 : * UnpinBuffer -- make buffer available for replacement.
3460 : *
3461 : * This should be applied only to shared buffers, never local ones. This
3462 : * always adjusts CurrentResourceOwner.
3463 : */
3464 : static void
3465 50495498 : UnpinBuffer(BufferDesc *buf)
3466 : {
3467 50495498 : Buffer b = BufferDescriptorGetBuffer(buf);
3468 :
3469 50495498 : ResourceOwnerForgetBuffer(CurrentResourceOwner, b);
3470 50495498 : UnpinBufferNoOwner(buf);
3471 50495498 : }
3472 :
3473 : static void
3474 50502185 : UnpinBufferNoOwner(BufferDesc *buf)
3475 : {
3476 : PrivateRefCountEntry *ref;
3477 50502185 : Buffer b = BufferDescriptorGetBuffer(buf);
3478 :
3479 : Assert(!BufferIsLocal(b));
3480 :
3481 : /* not moving as we're likely deleting it soon anyway */
3482 50502185 : ref = GetPrivateRefCountEntry(b, false);
3483 : Assert(ref != NULL);
3484 : Assert(ref->data.refcount > 0);
3485 50502185 : ref->data.refcount--;
3486 50502185 : if (ref->data.refcount == 0)
3487 : {
3488 : uint64 old_buf_state;
3489 :
3490 : /*
3491 : * Mark buffer non-accessible to Valgrind.
3492 : *
3493 : * Note that the buffer may have already been marked non-accessible
3494 : * within access method code that enforces that buffers are only
3495 : * accessed while a buffer lock is held.
3496 : */
3497 : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
3498 :
3499 : /*
3500 : * I'd better not still hold the buffer content lock. Can't use
3501 : * BufferIsLockedByMe(), as that asserts the buffer is pinned.
3502 : */
3503 : Assert(!BufferLockHeldByMe(buf));
3504 :
3505 : /* decrement the shared reference count */
3506 30506569 : old_buf_state = pg_atomic_fetch_sub_u64(&buf->state, BUF_REFCOUNT_ONE);
3507 :
3508 : /* Support LockBufferForCleanup() */
3509 30506569 : if (old_buf_state & BM_PIN_COUNT_WAITER)
3510 92 : WakePinCountWaiter(buf);
3511 :
3512 30506569 : ForgetPrivateRefCountEntry(ref);
3513 : }
3514 50502185 : }
3515 :
3516 : /*
3517 : * Set up backend-local tracking of a buffer pinned the first time by this
3518 : * backend.
3519 : */
3520 : inline void
3521 77576817 : TrackNewBufferPin(Buffer buf)
3522 : {
3523 : PrivateRefCountEntry *ref;
3524 :
3525 77576817 : ref = NewPrivateRefCountEntry(buf);
3526 77576817 : ref->data.refcount++;
3527 :
3528 77576817 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buf);
3529 :
3530 : /*
3531 : * This is the first pin for this page by this backend, mark its page as
3532 : * defined to valgrind. While the page contents might not actually be
3533 : * valid yet, we don't currently guarantee that such pages are marked
3534 : * undefined or non-accessible.
3535 : *
3536 : * It's not necessarily the prettiest to do this here, but otherwise we'd
3537 : * need this block of code in multiple places.
3538 : */
3539 : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(GetBufferDescriptor(buf - 1)),
3540 : BLCKSZ);
3541 77576817 : }
3542 :
3543 : #define ST_SORT sort_checkpoint_bufferids
3544 : #define ST_ELEMENT_TYPE CkptSortItem
3545 : #define ST_COMPARE(a, b) ckpt_buforder_comparator(a, b)
3546 : #define ST_SCOPE static
3547 : #define ST_DEFINE
3548 : #include "lib/sort_template.h"
3549 :
3550 : /*
3551 : * BufferSync -- Write out all dirty buffers in the pool.
3552 : *
3553 : * This is called at checkpoint time to write out all dirty shared buffers.
3554 : * The checkpoint request flags should be passed in. If CHECKPOINT_FAST is
3555 : * set, we disable delays between writes; if CHECKPOINT_IS_SHUTDOWN,
3556 : * CHECKPOINT_END_OF_RECOVERY or CHECKPOINT_FLUSH_UNLOGGED is set, we write
3557 : * even unlogged buffers, which are otherwise skipped. The remaining flags
3558 : * currently have no effect here.
3559 : */
3560 : static void
3561 1944 : BufferSync(int flags)
3562 : {
3563 : uint64 buf_state;
3564 : int buf_id;
3565 : int num_to_scan;
3566 : int num_spaces;
3567 : int num_processed;
3568 : int num_written;
3569 1944 : CkptTsStatus *per_ts_stat = NULL;
3570 : Oid last_tsid;
3571 : binaryheap *ts_heap;
3572 : int i;
3573 1944 : uint64 mask = BM_DIRTY;
3574 : WritebackContext wb_context;
3575 :
3576 : /*
3577 : * Unless this is a shutdown checkpoint or we have been explicitly told,
3578 : * we write only permanent, dirty buffers. But at shutdown or end of
3579 : * recovery, we write all dirty buffers.
3580 : */
3581 1944 : if (!((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
3582 : CHECKPOINT_FLUSH_UNLOGGED))))
3583 1062 : mask |= BM_PERMANENT;
3584 :
3585 : /*
3586 : * Loop over all buffers, and mark the ones that need to be written with
3587 : * BM_CHECKPOINT_NEEDED. Count them as we go (num_to_scan), so that we
3588 : * can estimate how much work needs to be done.
3589 : *
3590 : * This allows us to write only those pages that were dirty when the
3591 : * checkpoint began, and not those that get dirtied while it proceeds.
3592 : * Whenever a page with BM_CHECKPOINT_NEEDED is written out, either by us
3593 : * later in this function, or by normal backends or the bgwriter cleaning
3594 : * scan, the flag is cleared. Any buffer dirtied after this point won't
3595 : * have the flag set.
3596 : *
3597 : * Note that if we fail to write some buffer, we may leave buffers with
3598 : * BM_CHECKPOINT_NEEDED still set. This is OK since any such buffer would
3599 : * certainly need to be written for the next checkpoint attempt, too.
3600 : */
3601 1944 : num_to_scan = 0;
3602 13938024 : for (buf_id = 0; buf_id < NBuffers; buf_id++)
3603 : {
3604 13936080 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3605 13936080 : uint64 set_bits = 0;
3606 :
3607 : /*
3608 : * Header spinlock is enough to examine BM_DIRTY, see comment in
3609 : * SyncOneBuffer.
3610 : */
3611 13936080 : buf_state = LockBufHdr(bufHdr);
3612 :
3613 13936080 : if ((buf_state & mask) == mask)
3614 : {
3615 : CkptSortItem *item;
3616 :
3617 336711 : set_bits = BM_CHECKPOINT_NEEDED;
3618 :
3619 336711 : item = &CkptBufferIds[num_to_scan++];
3620 336711 : item->buf_id = buf_id;
3621 336711 : item->tsId = bufHdr->tag.spcOid;
3622 336711 : item->relNumber = BufTagGetRelNumber(&bufHdr->tag);
3623 336711 : item->forkNum = BufTagGetForkNum(&bufHdr->tag);
3624 336711 : item->blockNum = bufHdr->tag.blockNum;
3625 : }
3626 :
3627 13936080 : UnlockBufHdrExt(bufHdr, buf_state,
3628 : set_bits, 0,
3629 : 0);
3630 :
3631 : /* Check for barrier events in case NBuffers is large. */
3632 13936080 : if (ProcSignalBarrierPending)
3633 0 : ProcessProcSignalBarrier();
3634 : }
3635 :
3636 1944 : if (num_to_scan == 0)
3637 730 : return; /* nothing to do */
3638 :
3639 1214 : WritebackContextInit(&wb_context, &checkpoint_flush_after);
3640 :
3641 : TRACE_POSTGRESQL_BUFFER_SYNC_START(NBuffers, num_to_scan);
3642 :
3643 : /*
3644 : * Sort buffers that need to be written to reduce the likelihood of random
3645 : * IO. The sorting is also important for the implementation of balancing
3646 : * writes between tablespaces. Without balancing writes we'd potentially
3647 : * end up writing to the tablespaces one-by-one; possibly overloading the
3648 : * underlying system.
3649 : */
3650 1214 : sort_checkpoint_bufferids(CkptBufferIds, num_to_scan);
3651 :
3652 1214 : num_spaces = 0;
3653 :
3654 : /*
3655 : * Allocate progress status for each tablespace with buffers that need to
3656 : * be flushed. This requires the to-be-flushed array to be sorted.
3657 : */
3658 1214 : last_tsid = InvalidOid;
3659 337925 : for (i = 0; i < num_to_scan; i++)
3660 : {
3661 : CkptTsStatus *s;
3662 : Oid cur_tsid;
3663 :
3664 336711 : cur_tsid = CkptBufferIds[i].tsId;
3665 :
3666 : /*
3667 : * Grow array of per-tablespace status structs, every time a new
3668 : * tablespace is found.
3669 : */
3670 336711 : if (last_tsid == InvalidOid || last_tsid != cur_tsid)
3671 1826 : {
3672 : Size sz;
3673 :
3674 1826 : num_spaces++;
3675 :
3676 : /*
3677 : * Not worth adding grow-by-power-of-2 logic here - even with a
3678 : * few hundred tablespaces this should be fine.
3679 : */
3680 1826 : sz = sizeof(CkptTsStatus) * num_spaces;
3681 :
3682 1826 : if (per_ts_stat == NULL)
3683 1214 : per_ts_stat = (CkptTsStatus *) palloc(sz);
3684 : else
3685 612 : per_ts_stat = (CkptTsStatus *) repalloc(per_ts_stat, sz);
3686 :
3687 1826 : s = &per_ts_stat[num_spaces - 1];
3688 1826 : memset(s, 0, sizeof(*s));
3689 1826 : s->tsId = cur_tsid;
3690 :
3691 : /*
3692 : * The first buffer in this tablespace. As CkptBufferIds is sorted
3693 : * by tablespace all (s->num_to_scan) buffers in this tablespace
3694 : * will follow afterwards.
3695 : */
3696 1826 : s->index = i;
3697 :
3698 : /*
3699 : * progress_slice will be determined once we know how many buffers
3700 : * are in each tablespace, i.e. after this loop.
3701 : */
3702 :
3703 1826 : last_tsid = cur_tsid;
3704 : }
3705 : else
3706 : {
3707 334885 : s = &per_ts_stat[num_spaces - 1];
3708 : }
3709 :
3710 336711 : s->num_to_scan++;
3711 :
3712 : /* Check for barrier events. */
3713 336711 : if (ProcSignalBarrierPending)
3714 0 : ProcessProcSignalBarrier();
3715 : }
3716 :
3717 : Assert(num_spaces > 0);
3718 :
3719 : /*
3720 : * Build a min-heap over the write-progress in the individual tablespaces,
3721 : * and compute how large a portion of the total progress a single
3722 : * processed buffer is.
3723 : */
3724 1214 : ts_heap = binaryheap_allocate(num_spaces,
3725 : ts_ckpt_progress_comparator,
3726 : NULL);
3727 :
3728 3040 : for (i = 0; i < num_spaces; i++)
3729 : {
3730 1826 : CkptTsStatus *ts_stat = &per_ts_stat[i];
3731 :
3732 1826 : ts_stat->progress_slice = (float8) num_to_scan / ts_stat->num_to_scan;
3733 :
3734 1826 : binaryheap_add_unordered(ts_heap, PointerGetDatum(ts_stat));
3735 : }
3736 :
3737 1214 : binaryheap_build(ts_heap);
3738 :
3739 : /*
3740 : * Iterate through to-be-checkpointed buffers and write the ones (still)
3741 : * marked with BM_CHECKPOINT_NEEDED. The writes are balanced between
3742 : * tablespaces; otherwise the sorting would lead to only one tablespace
3743 : * receiving writes at a time, making inefficient use of the hardware.
3744 : */
3745 1214 : num_processed = 0;
3746 1214 : num_written = 0;
3747 337925 : while (!binaryheap_empty(ts_heap))
3748 : {
3749 336711 : BufferDesc *bufHdr = NULL;
3750 : CkptTsStatus *ts_stat = (CkptTsStatus *)
3751 336711 : DatumGetPointer(binaryheap_first(ts_heap));
3752 :
3753 336711 : buf_id = CkptBufferIds[ts_stat->index].buf_id;
3754 : Assert(buf_id != -1);
3755 :
3756 336711 : bufHdr = GetBufferDescriptor(buf_id);
3757 :
3758 336711 : num_processed++;
3759 :
3760 : /*
3761 : * We don't need to acquire the lock here, because we're only looking
3762 : * at a single bit. It's possible that someone else writes the buffer
3763 : * and clears the flag right after we check, but that doesn't matter
3764 : * since SyncOneBuffer will then do nothing. However, there is a
3765 : * further race condition: it's conceivable that between the time we
3766 : * examine the bit here and the time SyncOneBuffer acquires the lock,
3767 : * someone else not only wrote the buffer but replaced it with another
3768 : * page and dirtied it. In that improbable case, SyncOneBuffer will
3769 : * write the buffer though we didn't need to. It doesn't seem worth
3770 : * guarding against this, though.
3771 : */
3772 336711 : if (pg_atomic_read_u64(&bufHdr->state) & BM_CHECKPOINT_NEEDED)
3773 : {
3774 313924 : if (SyncOneBuffer(buf_id, false, &wb_context) & BUF_WRITTEN)
3775 : {
3776 : TRACE_POSTGRESQL_BUFFER_SYNC_WRITTEN(buf_id);
3777 313924 : PendingCheckpointerStats.buffers_written++;
3778 313924 : num_written++;
3779 : }
3780 : }
3781 :
3782 : /*
3783 : * Measure progress independent of actually having to flush the buffer
3784 : * - otherwise writing become unbalanced.
3785 : */
3786 336711 : ts_stat->progress += ts_stat->progress_slice;
3787 336711 : ts_stat->num_scanned++;
3788 336711 : ts_stat->index++;
3789 :
3790 : /* Have all the buffers from the tablespace been processed? */
3791 336711 : if (ts_stat->num_scanned == ts_stat->num_to_scan)
3792 : {
3793 1826 : binaryheap_remove_first(ts_heap);
3794 : }
3795 : else
3796 : {
3797 : /* update heap with the new progress */
3798 334885 : binaryheap_replace_first(ts_heap, PointerGetDatum(ts_stat));
3799 : }
3800 :
3801 : /*
3802 : * Sleep to throttle our I/O rate.
3803 : *
3804 : * (This will check for barrier events even if it doesn't sleep.)
3805 : */
3806 336711 : CheckpointWriteDelay(flags, (double) num_processed / num_to_scan);
3807 : }
3808 :
3809 : /*
3810 : * Issue all pending flushes. Only checkpointer calls BufferSync(), so
3811 : * IOContext will always be IOCONTEXT_NORMAL.
3812 : */
3813 1214 : IssuePendingWritebacks(&wb_context, IOCONTEXT_NORMAL);
3814 :
3815 1214 : pfree(per_ts_stat);
3816 1214 : per_ts_stat = NULL;
3817 1214 : binaryheap_free(ts_heap);
3818 :
3819 : /*
3820 : * Update checkpoint statistics. As noted above, this doesn't include
3821 : * buffers written by other backends or bgwriter scan.
3822 : */
3823 1214 : CheckpointStats.ckpt_bufs_written += num_written;
3824 :
3825 : TRACE_POSTGRESQL_BUFFER_SYNC_DONE(NBuffers, num_written, num_to_scan);
3826 : }
3827 :
3828 : /*
3829 : * BgBufferSync -- Write out some dirty buffers in the pool.
3830 : *
3831 : * This is called periodically by the background writer process.
3832 : *
3833 : * Returns true if it's appropriate for the bgwriter process to go into
3834 : * low-power hibernation mode. (This happens if the strategy clock-sweep
3835 : * has been "lapped" and no buffer allocations have occurred recently,
3836 : * or if the bgwriter has been effectively disabled by setting
3837 : * bgwriter_lru_maxpages to 0.)
3838 : */
3839 : bool
3840 15371 : BgBufferSync(WritebackContext *wb_context)
3841 : {
3842 : /* info obtained from freelist.c */
3843 : int strategy_buf_id;
3844 : uint32 strategy_passes;
3845 : uint32 recent_alloc;
3846 :
3847 : /*
3848 : * Information saved between calls so we can determine the strategy
3849 : * point's advance rate and avoid scanning already-cleaned buffers.
3850 : */
3851 : static bool saved_info_valid = false;
3852 : static int prev_strategy_buf_id;
3853 : static uint32 prev_strategy_passes;
3854 : static int next_to_clean;
3855 : static uint32 next_passes;
3856 :
3857 : /* Moving averages of allocation rate and clean-buffer density */
3858 : static float smoothed_alloc = 0;
3859 : static float smoothed_density = 10.0;
3860 :
3861 : /* Potentially these could be tunables, but for now, not */
3862 15371 : float smoothing_samples = 16;
3863 15371 : float scan_whole_pool_milliseconds = 120000.0;
3864 :
3865 : /* Used to compute how far we scan ahead */
3866 : long strategy_delta;
3867 : int bufs_to_lap;
3868 : int bufs_ahead;
3869 : float scans_per_alloc;
3870 : int reusable_buffers_est;
3871 : int upcoming_alloc_est;
3872 : int min_scan_buffers;
3873 :
3874 : /* Variables for the scanning loop proper */
3875 : int num_to_scan;
3876 : int num_written;
3877 : int reusable_buffers;
3878 :
3879 : /* Variables for final smoothed_density update */
3880 : long new_strategy_delta;
3881 : uint32 new_recent_alloc;
3882 :
3883 : /*
3884 : * Find out where the clock-sweep currently is, and how many buffer
3885 : * allocations have happened since our last call.
3886 : */
3887 15371 : strategy_buf_id = StrategySyncStart(&strategy_passes, &recent_alloc);
3888 :
3889 : /* Report buffer alloc counts to pgstat */
3890 15371 : PendingBgWriterStats.buf_alloc += recent_alloc;
3891 :
3892 : /*
3893 : * If we're not running the LRU scan, just stop after doing the stats
3894 : * stuff. We mark the saved state invalid so that we can recover sanely
3895 : * if LRU scan is turned back on later.
3896 : */
3897 15371 : if (bgwriter_lru_maxpages <= 0)
3898 : {
3899 42 : saved_info_valid = false;
3900 42 : return true;
3901 : }
3902 :
3903 : /*
3904 : * Compute strategy_delta = how many buffers have been scanned by the
3905 : * clock-sweep since last time. If first time through, assume none. Then
3906 : * see if we are still ahead of the clock-sweep, and if so, how many
3907 : * buffers we could scan before we'd catch up with it and "lap" it. Note:
3908 : * weird-looking coding of xxx_passes comparisons are to avoid bogus
3909 : * behavior when the passes counts wrap around.
3910 : */
3911 15329 : if (saved_info_valid)
3912 : {
3913 14698 : int32 passes_delta = strategy_passes - prev_strategy_passes;
3914 :
3915 14698 : strategy_delta = strategy_buf_id - prev_strategy_buf_id;
3916 14698 : strategy_delta += (long) passes_delta * NBuffers;
3917 :
3918 : Assert(strategy_delta >= 0);
3919 :
3920 14698 : if ((int32) (next_passes - strategy_passes) > 0)
3921 : {
3922 : /* we're one pass ahead of the strategy point */
3923 2255 : bufs_to_lap = strategy_buf_id - next_to_clean;
3924 : #ifdef BGW_DEBUG
3925 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3926 : next_passes, next_to_clean,
3927 : strategy_passes, strategy_buf_id,
3928 : strategy_delta, bufs_to_lap);
3929 : #endif
3930 : }
3931 12443 : else if (next_passes == strategy_passes &&
3932 9294 : next_to_clean >= strategy_buf_id)
3933 : {
3934 : /* on same pass, but ahead or at least not behind */
3935 8346 : bufs_to_lap = NBuffers - (next_to_clean - strategy_buf_id);
3936 : #ifdef BGW_DEBUG
3937 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3938 : next_passes, next_to_clean,
3939 : strategy_passes, strategy_buf_id,
3940 : strategy_delta, bufs_to_lap);
3941 : #endif
3942 : }
3943 : else
3944 : {
3945 : /*
3946 : * We're behind, so skip forward to the strategy point and start
3947 : * cleaning from there.
3948 : */
3949 : #ifdef BGW_DEBUG
3950 : elog(DEBUG2, "bgwriter behind: bgw %u-%u strategy %u-%u delta=%ld",
3951 : next_passes, next_to_clean,
3952 : strategy_passes, strategy_buf_id,
3953 : strategy_delta);
3954 : #endif
3955 4097 : next_to_clean = strategy_buf_id;
3956 4097 : next_passes = strategy_passes;
3957 4097 : bufs_to_lap = NBuffers;
3958 : }
3959 : }
3960 : else
3961 : {
3962 : /*
3963 : * Initializing at startup or after LRU scanning had been off. Always
3964 : * start at the strategy point.
3965 : */
3966 : #ifdef BGW_DEBUG
3967 : elog(DEBUG2, "bgwriter initializing: strategy %u-%u",
3968 : strategy_passes, strategy_buf_id);
3969 : #endif
3970 631 : strategy_delta = 0;
3971 631 : next_to_clean = strategy_buf_id;
3972 631 : next_passes = strategy_passes;
3973 631 : bufs_to_lap = NBuffers;
3974 : }
3975 :
3976 : /* Update saved info for next time */
3977 15329 : prev_strategy_buf_id = strategy_buf_id;
3978 15329 : prev_strategy_passes = strategy_passes;
3979 15329 : saved_info_valid = true;
3980 :
3981 : /*
3982 : * Compute how many buffers had to be scanned for each new allocation, ie,
3983 : * 1/density of reusable buffers, and track a moving average of that.
3984 : *
3985 : * If the strategy point didn't move, we don't update the density estimate
3986 : */
3987 15329 : if (strategy_delta > 0 && recent_alloc > 0)
3988 : {
3989 8953 : scans_per_alloc = (float) strategy_delta / (float) recent_alloc;
3990 8953 : smoothed_density += (scans_per_alloc - smoothed_density) /
3991 : smoothing_samples;
3992 : }
3993 :
3994 : /*
3995 : * Estimate how many reusable buffers there are between the current
3996 : * strategy point and where we've scanned ahead to, based on the smoothed
3997 : * density estimate.
3998 : */
3999 15329 : bufs_ahead = NBuffers - bufs_to_lap;
4000 15329 : reusable_buffers_est = (float) bufs_ahead / smoothed_density;
4001 :
4002 : /*
4003 : * Track a moving average of recent buffer allocations. Here, rather than
4004 : * a true average we want a fast-attack, slow-decline behavior: we
4005 : * immediately follow any increase.
4006 : */
4007 15329 : if (smoothed_alloc <= (float) recent_alloc)
4008 4136 : smoothed_alloc = recent_alloc;
4009 : else
4010 11193 : smoothed_alloc += ((float) recent_alloc - smoothed_alloc) /
4011 : smoothing_samples;
4012 :
4013 : /* Scale the estimate by a GUC to allow more aggressive tuning. */
4014 15329 : upcoming_alloc_est = (int) (smoothed_alloc * bgwriter_lru_multiplier);
4015 :
4016 : /*
4017 : * If recent_alloc remains at zero for many cycles, smoothed_alloc will
4018 : * eventually underflow to zero, and the underflows produce annoying
4019 : * kernel warnings on some platforms. Once upcoming_alloc_est has gone to
4020 : * zero, there's no point in tracking smaller and smaller values of
4021 : * smoothed_alloc, so just reset it to exactly zero to avoid this
4022 : * syndrome. It will pop back up as soon as recent_alloc increases.
4023 : */
4024 15329 : if (upcoming_alloc_est == 0)
4025 2355 : smoothed_alloc = 0;
4026 :
4027 : /*
4028 : * Even in cases where there's been little or no buffer allocation
4029 : * activity, we want to make a small amount of progress through the buffer
4030 : * cache so that as many reusable buffers as possible are clean after an
4031 : * idle period.
4032 : *
4033 : * (scan_whole_pool_milliseconds / BgWriterDelay) computes how many times
4034 : * the BGW will be called during the scan_whole_pool time; slice the
4035 : * buffer pool into that many sections.
4036 : */
4037 15329 : min_scan_buffers = (int) (NBuffers / (scan_whole_pool_milliseconds / BgWriterDelay));
4038 :
4039 15329 : if (upcoming_alloc_est < (min_scan_buffers + reusable_buffers_est))
4040 : {
4041 : #ifdef BGW_DEBUG
4042 : elog(DEBUG2, "bgwriter: alloc_est=%d too small, using min=%d + reusable_est=%d",
4043 : upcoming_alloc_est, min_scan_buffers, reusable_buffers_est);
4044 : #endif
4045 7343 : upcoming_alloc_est = min_scan_buffers + reusable_buffers_est;
4046 : }
4047 :
4048 : /*
4049 : * Now write out dirty reusable buffers, working forward from the
4050 : * next_to_clean point, until we have lapped the strategy scan, or cleaned
4051 : * enough buffers to match our estimate of the next cycle's allocation
4052 : * requirements, or hit the bgwriter_lru_maxpages limit.
4053 : */
4054 :
4055 15329 : num_to_scan = bufs_to_lap;
4056 15329 : num_written = 0;
4057 15329 : reusable_buffers = reusable_buffers_est;
4058 :
4059 : /* Execute the LRU scan */
4060 2062704 : while (num_to_scan > 0 && reusable_buffers < upcoming_alloc_est)
4061 : {
4062 2047375 : int sync_state = SyncOneBuffer(next_to_clean, true,
4063 : wb_context);
4064 :
4065 2047375 : if (++next_to_clean >= NBuffers)
4066 : {
4067 3970 : next_to_clean = 0;
4068 3970 : next_passes++;
4069 : }
4070 2047375 : num_to_scan--;
4071 :
4072 2047375 : if (sync_state & BUF_WRITTEN)
4073 : {
4074 36248 : reusable_buffers++;
4075 36248 : if (++num_written >= bgwriter_lru_maxpages)
4076 : {
4077 0 : PendingBgWriterStats.maxwritten_clean++;
4078 0 : break;
4079 : }
4080 : }
4081 2011127 : else if (sync_state & BUF_REUSABLE)
4082 1511125 : reusable_buffers++;
4083 : }
4084 :
4085 15329 : PendingBgWriterStats.buf_written_clean += num_written;
4086 :
4087 : #ifdef BGW_DEBUG
4088 : 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",
4089 : recent_alloc, smoothed_alloc, strategy_delta, bufs_ahead,
4090 : smoothed_density, reusable_buffers_est, upcoming_alloc_est,
4091 : bufs_to_lap - num_to_scan,
4092 : num_written,
4093 : reusable_buffers - reusable_buffers_est);
4094 : #endif
4095 :
4096 : /*
4097 : * Consider the above scan as being like a new allocation scan.
4098 : * Characterize its density and update the smoothed one based on it. This
4099 : * effectively halves the moving average period in cases where both the
4100 : * strategy and the background writer are doing some useful scanning,
4101 : * which is helpful because a long memory isn't as desirable on the
4102 : * density estimates.
4103 : */
4104 15329 : new_strategy_delta = bufs_to_lap - num_to_scan;
4105 15329 : new_recent_alloc = reusable_buffers - reusable_buffers_est;
4106 15329 : if (new_strategy_delta > 0 && new_recent_alloc > 0)
4107 : {
4108 13451 : scans_per_alloc = (float) new_strategy_delta / (float) new_recent_alloc;
4109 13451 : smoothed_density += (scans_per_alloc - smoothed_density) /
4110 : smoothing_samples;
4111 :
4112 : #ifdef BGW_DEBUG
4113 : elog(DEBUG2, "bgwriter: cleaner density alloc=%u scan=%ld density=%.2f new smoothed=%.2f",
4114 : new_recent_alloc, new_strategy_delta,
4115 : scans_per_alloc, smoothed_density);
4116 : #endif
4117 : }
4118 :
4119 : /* Return true if OK to hibernate */
4120 15329 : return (bufs_to_lap == 0 && recent_alloc == 0);
4121 : }
4122 :
4123 : /*
4124 : * SyncOneBuffer -- process a single buffer during syncing.
4125 : *
4126 : * If skip_recently_used is true, we don't write currently-pinned buffers, nor
4127 : * buffers marked recently used, as these are not replacement candidates.
4128 : *
4129 : * Returns a bitmask containing the following flag bits:
4130 : * BUF_WRITTEN: we wrote the buffer.
4131 : * BUF_REUSABLE: buffer is available for replacement, ie, it has
4132 : * pin count 0 and usage count 0.
4133 : *
4134 : * (BUF_WRITTEN could be set in error if FlushBuffer finds the buffer clean
4135 : * after locking it, but we don't care all that much.)
4136 : */
4137 : static int
4138 2361299 : SyncOneBuffer(int buf_id, bool skip_recently_used, WritebackContext *wb_context)
4139 : {
4140 2361299 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
4141 2361299 : int result = 0;
4142 : uint64 buf_state;
4143 : BufferTag tag;
4144 :
4145 : /* Make sure we can handle the pin */
4146 2361299 : ReservePrivateRefCountEntry();
4147 2361299 : ResourceOwnerEnlarge(CurrentResourceOwner);
4148 :
4149 : /*
4150 : * Check whether buffer needs writing.
4151 : *
4152 : * We can make this check without taking the buffer content lock so long
4153 : * as we mark pages dirty in access methods *before* logging changes with
4154 : * XLogInsert(): if someone marks the buffer dirty just after our check we
4155 : * don't worry because our checkpoint.redo points before log record for
4156 : * upcoming changes and so we are not required to write such dirty buffer.
4157 : */
4158 2361299 : buf_state = LockBufHdr(bufHdr);
4159 :
4160 2361299 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 0 &&
4161 2355573 : BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
4162 : {
4163 1549328 : result |= BUF_REUSABLE;
4164 : }
4165 811971 : else if (skip_recently_used)
4166 : {
4167 : /* Caller told us not to write recently-used buffers */
4168 500002 : UnlockBufHdr(bufHdr);
4169 500002 : return result;
4170 : }
4171 :
4172 1861297 : if (!(buf_state & BM_VALID) || !(buf_state & BM_DIRTY))
4173 : {
4174 : /* It's clean, so nothing to do */
4175 1511125 : UnlockBufHdr(bufHdr);
4176 1511125 : return result;
4177 : }
4178 :
4179 : /*
4180 : * Pin it, share-exclusive-lock it, write it. (FlushBuffer will do
4181 : * nothing if the buffer is clean by the time we've locked it.)
4182 : */
4183 350172 : PinBuffer_Locked(bufHdr);
4184 :
4185 350172 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
4186 :
4187 350172 : tag = bufHdr->tag;
4188 :
4189 350172 : UnpinBuffer(bufHdr);
4190 :
4191 : /*
4192 : * SyncOneBuffer() is only called by checkpointer and bgwriter, so
4193 : * IOContext will always be IOCONTEXT_NORMAL.
4194 : */
4195 350172 : ScheduleBufferTagForWriteback(wb_context, IOCONTEXT_NORMAL, &tag);
4196 :
4197 350172 : return result | BUF_WRITTEN;
4198 : }
4199 :
4200 : /*
4201 : * AtEOXact_Buffers - clean up at end of transaction.
4202 : *
4203 : * As of PostgreSQL 8.0, buffer pins should get released by the
4204 : * ResourceOwner mechanism. This routine is just a debugging
4205 : * cross-check that no pins remain.
4206 : */
4207 : void
4208 646694 : AtEOXact_Buffers(bool isCommit)
4209 : {
4210 646694 : CheckForBufferLeaks();
4211 :
4212 646694 : AtEOXact_LocalBuffers(isCommit);
4213 :
4214 : Assert(PrivateRefCountOverflowed == 0);
4215 646694 : }
4216 :
4217 : /*
4218 : * Initialize access to shared buffer pool
4219 : *
4220 : * This is called during backend startup (whether standalone or under the
4221 : * postmaster). It sets up for this backend's access to the already-existing
4222 : * buffer pool.
4223 : */
4224 : void
4225 24430 : InitBufferManagerAccess(void)
4226 : {
4227 : /*
4228 : * An advisory limit on the number of pins each backend should hold, based
4229 : * on shared_buffers and the maximum number of connections possible.
4230 : * That's very pessimistic, but outside toy-sized shared_buffers it should
4231 : * allow plenty of pins. LimitAdditionalPins() and
4232 : * GetAdditionalPinLimit() can be used to check the remaining balance.
4233 : */
4234 24430 : MaxProportionalPins = NBuffers / (MaxBackends + NUM_AUXILIARY_PROCS);
4235 :
4236 24430 : memset(&PrivateRefCountArray, 0, sizeof(PrivateRefCountArray));
4237 24430 : memset(&PrivateRefCountArrayKeys, 0, sizeof(PrivateRefCountArrayKeys));
4238 :
4239 24430 : PrivateRefCountHash = refcount_create(CurrentMemoryContext, 100, NULL);
4240 :
4241 : /*
4242 : * AtProcExit_Buffers needs LWLock access, and thereby has to be called at
4243 : * the corresponding phase of backend shutdown.
4244 : */
4245 : Assert(MyProc != NULL);
4246 24430 : on_shmem_exit(AtProcExit_Buffers, 0);
4247 24430 : }
4248 :
4249 : /*
4250 : * During backend exit, ensure that we released all shared-buffer locks and
4251 : * assert that we have no remaining pins.
4252 : */
4253 : static void
4254 24430 : AtProcExit_Buffers(int code, Datum arg)
4255 : {
4256 24430 : UnlockBuffers();
4257 :
4258 24430 : CheckForBufferLeaks();
4259 :
4260 : /* localbuf.c needs a chance too */
4261 24430 : AtProcExit_LocalBuffers();
4262 24430 : }
4263 :
4264 : /*
4265 : * CheckForBufferLeaks - ensure this backend holds no buffer pins
4266 : *
4267 : * As of PostgreSQL 8.0, buffer pins should get released by the
4268 : * ResourceOwner mechanism. This routine is just a debugging
4269 : * cross-check that no pins remain.
4270 : */
4271 : static void
4272 671124 : CheckForBufferLeaks(void)
4273 : {
4274 : #ifdef USE_ASSERT_CHECKING
4275 : int RefCountErrors = 0;
4276 : PrivateRefCountEntry *res;
4277 : int i;
4278 : char *s;
4279 :
4280 : /* check the array */
4281 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4282 : {
4283 : if (PrivateRefCountArrayKeys[i] != InvalidBuffer)
4284 : {
4285 : res = &PrivateRefCountArray[i];
4286 :
4287 : s = DebugPrintBufferRefcount(res->buffer);
4288 : elog(WARNING, "buffer refcount leak: %s", s);
4289 : pfree(s);
4290 :
4291 : RefCountErrors++;
4292 : }
4293 : }
4294 :
4295 : /* if necessary search the hash */
4296 : if (PrivateRefCountOverflowed)
4297 : {
4298 : refcount_iterator iter;
4299 :
4300 : refcount_start_iterate(PrivateRefCountHash, &iter);
4301 : while ((res = refcount_iterate(PrivateRefCountHash, &iter)) != NULL)
4302 : {
4303 : s = DebugPrintBufferRefcount(res->buffer);
4304 : elog(WARNING, "buffer refcount leak: %s", s);
4305 : pfree(s);
4306 : RefCountErrors++;
4307 : }
4308 : }
4309 :
4310 : Assert(RefCountErrors == 0);
4311 : #endif
4312 671124 : }
4313 :
4314 : #ifdef USE_ASSERT_CHECKING
4315 : /*
4316 : * Check for exclusive-locked catalog buffers. This is the core of
4317 : * AssertCouldGetRelation().
4318 : *
4319 : * A backend would self-deadlock on the content lock if the catalog scan read
4320 : * the exclusive-locked buffer. The main threat is exclusive-locked buffers
4321 : * of catalogs used in relcache, because a catcache search on any catalog may
4322 : * build that catalog's relcache entry. We don't have an inventory of
4323 : * catalogs relcache uses, so just check buffers of most catalogs.
4324 : *
4325 : * It's better to minimize waits while holding an exclusive buffer lock, so it
4326 : * would be nice to broaden this check not to be catalog-specific. However,
4327 : * bttextcmp() accesses pg_collation, and non-core opclasses might similarly
4328 : * read tables. That is deadlock-free as long as there's no loop in the
4329 : * dependency graph: modifying table A may cause an opclass to read table B,
4330 : * but it must not cause a read of table A.
4331 : */
4332 : void
4333 : AssertBufferLocksPermitCatalogRead(void)
4334 : {
4335 : PrivateRefCountEntry *res;
4336 :
4337 : /* check the array */
4338 : for (int i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4339 : {
4340 : if (PrivateRefCountArrayKeys[i] != InvalidBuffer)
4341 : {
4342 : res = &PrivateRefCountArray[i];
4343 :
4344 : if (res->buffer == InvalidBuffer)
4345 : continue;
4346 :
4347 : AssertNotCatalogBufferLock(res->buffer, res->data.lockmode);
4348 : }
4349 : }
4350 :
4351 : /* if necessary search the hash */
4352 : if (PrivateRefCountOverflowed)
4353 : {
4354 : refcount_iterator iter;
4355 :
4356 : refcount_start_iterate(PrivateRefCountHash, &iter);
4357 : while ((res = refcount_iterate(PrivateRefCountHash, &iter)) != NULL)
4358 : {
4359 : AssertNotCatalogBufferLock(res->buffer, res->data.lockmode);
4360 : }
4361 : }
4362 : }
4363 :
4364 : static void
4365 : AssertNotCatalogBufferLock(Buffer buffer, BufferLockMode mode)
4366 : {
4367 : BufferDesc *bufHdr = GetBufferDescriptor(buffer - 1);
4368 : BufferTag tag;
4369 : Oid relid;
4370 :
4371 : if (mode != BUFFER_LOCK_EXCLUSIVE)
4372 : return;
4373 :
4374 : tag = bufHdr->tag;
4375 :
4376 : /*
4377 : * This relNumber==relid assumption holds until a catalog experiences
4378 : * VACUUM FULL or similar. After a command like that, relNumber will be
4379 : * in the normal (non-catalog) range, and we lose the ability to detect
4380 : * hazardous access to that catalog. Calling RelidByRelfilenumber() would
4381 : * close that gap, but RelidByRelfilenumber() might then deadlock with a
4382 : * held lock.
4383 : */
4384 : relid = tag.relNumber;
4385 :
4386 : if (IsCatalogTextUniqueIndexOid(relid)) /* see comments at the callee */
4387 : return;
4388 :
4389 : Assert(!IsCatalogRelationOid(relid));
4390 : }
4391 : #endif
4392 :
4393 :
4394 : /*
4395 : * Helper routine to issue warnings when a buffer is unexpectedly pinned
4396 : */
4397 : char *
4398 46 : DebugPrintBufferRefcount(Buffer buffer)
4399 : {
4400 : BufferDesc *buf;
4401 : int32 loccount;
4402 : char *result;
4403 : ProcNumber backend;
4404 : uint64 buf_state;
4405 :
4406 : Assert(BufferIsValid(buffer));
4407 46 : if (BufferIsLocal(buffer))
4408 : {
4409 16 : buf = GetLocalBufferDescriptor(-buffer - 1);
4410 16 : loccount = LocalRefCount[-buffer - 1];
4411 16 : backend = MyProcNumber;
4412 : }
4413 : else
4414 : {
4415 30 : buf = GetBufferDescriptor(buffer - 1);
4416 30 : loccount = GetPrivateRefCount(buffer);
4417 30 : backend = INVALID_PROC_NUMBER;
4418 : }
4419 :
4420 : /* theoretically we should lock the bufHdr here */
4421 46 : buf_state = pg_atomic_read_u64(&buf->state);
4422 :
4423 46 : result = psprintf("[%03d] (rel=%s, blockNum=%u, flags=0x%" PRIx64 ", refcount=%u %d)",
4424 : buffer,
4425 46 : relpathbackend(BufTagGetRelFileLocator(&buf->tag), backend,
4426 : BufTagGetForkNum(&buf->tag)).str,
4427 : buf->tag.blockNum, buf_state & BUF_FLAG_MASK,
4428 : BUF_STATE_GET_REFCOUNT(buf_state), loccount);
4429 46 : return result;
4430 : }
4431 :
4432 : /*
4433 : * CheckPointBuffers
4434 : *
4435 : * Flush all dirty blocks in buffer pool to disk at checkpoint time.
4436 : *
4437 : * Note: temporary relations do not participate in checkpoints, so they don't
4438 : * need to be flushed.
4439 : */
4440 : void
4441 1944 : CheckPointBuffers(int flags)
4442 : {
4443 1944 : BufferSync(flags);
4444 1944 : }
4445 :
4446 : /*
4447 : * BufferGetBlockNumber
4448 : * Returns the block number associated with a buffer.
4449 : *
4450 : * Note:
4451 : * Assumes that the buffer is valid and pinned, else the
4452 : * value may be obsolete immediately...
4453 : */
4454 : BlockNumber
4455 80584374 : BufferGetBlockNumber(Buffer buffer)
4456 : {
4457 : BufferDesc *bufHdr;
4458 :
4459 : Assert(BufferIsPinned(buffer));
4460 :
4461 80584374 : if (BufferIsLocal(buffer))
4462 2562142 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4463 : else
4464 78022232 : bufHdr = GetBufferDescriptor(buffer - 1);
4465 :
4466 : /* pinned, so OK to read tag without spinlock */
4467 80584374 : return bufHdr->tag.blockNum;
4468 : }
4469 :
4470 : /*
4471 : * BufferGetTag
4472 : * Returns the relfilelocator, fork number and block number associated with
4473 : * a buffer.
4474 : */
4475 : void
4476 27487978 : BufferGetTag(Buffer buffer, RelFileLocator *rlocator, ForkNumber *forknum,
4477 : BlockNumber *blknum)
4478 : {
4479 : BufferDesc *bufHdr;
4480 :
4481 : /* Do the same checks as BufferGetBlockNumber. */
4482 : Assert(BufferIsPinned(buffer));
4483 :
4484 27487978 : if (BufferIsLocal(buffer))
4485 0 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4486 : else
4487 27487978 : bufHdr = GetBufferDescriptor(buffer - 1);
4488 :
4489 : /* pinned, so OK to read tag without spinlock */
4490 27487978 : *rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
4491 27487978 : *forknum = BufTagGetForkNum(&bufHdr->tag);
4492 27487978 : *blknum = bufHdr->tag.blockNum;
4493 27487978 : }
4494 :
4495 : /*
4496 : * FlushBuffer
4497 : * Physically write out a shared buffer.
4498 : *
4499 : * NOTE: this actually just passes the buffer contents to the kernel; the
4500 : * real write to disk won't happen until the kernel feels like it. This
4501 : * is okay from our point of view since we can redo the changes from WAL.
4502 : * However, we will need to force the changes to disk via fsync before
4503 : * we can checkpoint WAL.
4504 : *
4505 : * The caller must hold a pin on the buffer and have
4506 : * (share-)exclusively-locked the buffer contents.
4507 : *
4508 : * If the caller has an smgr reference for the buffer's relation, pass it
4509 : * as the second parameter. If not, pass NULL.
4510 : */
4511 : static void
4512 707095 : FlushBuffer(BufferDesc *buf, SMgrRelation reln, IOObject io_object,
4513 : IOContext io_context)
4514 : {
4515 : XLogRecPtr recptr;
4516 : ErrorContextCallback errcallback;
4517 : instr_time io_start;
4518 : Block bufBlock;
4519 :
4520 : Assert(BufferLockHeldByMeInMode(buf, BUFFER_LOCK_EXCLUSIVE) ||
4521 : BufferLockHeldByMeInMode(buf, BUFFER_LOCK_SHARE_EXCLUSIVE));
4522 :
4523 : /*
4524 : * Try to start an I/O operation. If StartBufferIO returns false, then
4525 : * someone else flushed the buffer before we could, so we need not do
4526 : * anything.
4527 : */
4528 707095 : if (StartSharedBufferIO(buf, false, true, NULL) == BUFFER_IO_ALREADY_DONE)
4529 11 : return;
4530 :
4531 : /* Setup error traceback support for ereport() */
4532 707084 : errcallback.callback = shared_buffer_write_error_callback;
4533 707084 : errcallback.arg = buf;
4534 707084 : errcallback.previous = error_context_stack;
4535 707084 : error_context_stack = &errcallback;
4536 :
4537 : /* Find smgr relation for buffer */
4538 707084 : if (reln == NULL)
4539 704648 : reln = smgropen(BufTagGetRelFileLocator(&buf->tag), INVALID_PROC_NUMBER);
4540 :
4541 : TRACE_POSTGRESQL_BUFFER_FLUSH_START(BufTagGetForkNum(&buf->tag),
4542 : buf->tag.blockNum,
4543 : reln->smgr_rlocator.locator.spcOid,
4544 : reln->smgr_rlocator.locator.dbOid,
4545 : reln->smgr_rlocator.locator.relNumber);
4546 :
4547 : /*
4548 : * As we hold at least a share-exclusive lock on the buffer, the LSN
4549 : * cannot change during the flush (and thus can't be torn).
4550 : */
4551 707084 : recptr = BufferGetLSN(buf);
4552 :
4553 : /*
4554 : * Force XLOG flush up to buffer's LSN. This implements the basic WAL
4555 : * rule that log updates must hit disk before any of the data-file changes
4556 : * they describe do.
4557 : *
4558 : * However, this rule does not apply to unlogged relations, which will be
4559 : * lost after a crash anyway. Most unlogged relation pages do not bear
4560 : * LSNs since we never emit WAL records for them, and therefore flushing
4561 : * up through the buffer LSN would be useless, but harmless. However,
4562 : * some index AMs use LSNs internally to detect concurrent page
4563 : * modifications, and therefore unlogged index pages bear "fake" LSNs
4564 : * generated by XLogGetFakeLSN. It is unlikely but possible that the fake
4565 : * LSN counter could advance past the WAL insertion point; and if it did
4566 : * happen, attempting to flush WAL through that location would fail, with
4567 : * disastrous system-wide consequences. To make sure that can't happen,
4568 : * skip the flush if the buffer isn't permanent.
4569 : */
4570 707084 : if (pg_atomic_read_u64(&buf->state) & BM_PERMANENT)
4571 705192 : XLogFlush(recptr);
4572 :
4573 : /*
4574 : * Now it's safe to write the buffer to disk. Note that no one else should
4575 : * have been able to write it, while we were busy with log flushing,
4576 : * because we got the exclusive right to perform I/O by setting the
4577 : * BM_IO_IN_PROGRESS bit.
4578 : */
4579 707084 : bufBlock = BufHdrGetBlock(buf);
4580 :
4581 : /* Update page checksum if desired. */
4582 707084 : PageSetChecksum((Page) bufBlock, buf->tag.blockNum);
4583 :
4584 707084 : io_start = pgstat_prepare_io_time(track_io_timing);
4585 :
4586 707084 : smgrwrite(reln,
4587 707084 : BufTagGetForkNum(&buf->tag),
4588 : buf->tag.blockNum,
4589 : bufBlock,
4590 : false);
4591 :
4592 : /*
4593 : * When a strategy is in use, only flushes of dirty buffers already in the
4594 : * strategy ring are counted as strategy writes (IOCONTEXT
4595 : * [BULKREAD|BULKWRITE|VACUUM] IOOP_WRITE) for the purpose of IO
4596 : * statistics tracking.
4597 : *
4598 : * If a shared buffer initially added to the ring must be flushed before
4599 : * being used, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE.
4600 : *
4601 : * If a shared buffer which was added to the ring later because the
4602 : * current strategy buffer is pinned or in use or because all strategy
4603 : * buffers were dirty and rejected (for BAS_BULKREAD operations only)
4604 : * requires flushing, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE
4605 : * (from_ring will be false).
4606 : *
4607 : * When a strategy is not in use, the write can only be a "regular" write
4608 : * of a dirty shared buffer (IOCONTEXT_NORMAL IOOP_WRITE).
4609 : */
4610 707084 : pgstat_count_io_op_time(io_object, io_context,
4611 : IOOP_WRITE, io_start, 1, BLCKSZ);
4612 :
4613 707084 : pgBufferUsage.shared_blks_written++;
4614 :
4615 : /*
4616 : * Mark the buffer as clean and end the BM_IO_IN_PROGRESS state.
4617 : */
4618 707084 : TerminateBufferIO(buf, true, 0, true, false);
4619 :
4620 : TRACE_POSTGRESQL_BUFFER_FLUSH_DONE(BufTagGetForkNum(&buf->tag),
4621 : buf->tag.blockNum,
4622 : reln->smgr_rlocator.locator.spcOid,
4623 : reln->smgr_rlocator.locator.dbOid,
4624 : reln->smgr_rlocator.locator.relNumber);
4625 :
4626 : /* Pop the error context stack */
4627 707084 : error_context_stack = errcallback.previous;
4628 : }
4629 :
4630 : /*
4631 : * Convenience wrapper around FlushBuffer() that locks/unlocks the buffer
4632 : * before/after calling FlushBuffer().
4633 : */
4634 : static void
4635 353690 : FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
4636 : IOObject io_object, IOContext io_context)
4637 : {
4638 353690 : Buffer buffer = BufferDescriptorGetBuffer(buf);
4639 :
4640 353690 : BufferLockAcquire(buffer, buf, BUFFER_LOCK_SHARE_EXCLUSIVE);
4641 353690 : FlushBuffer(buf, reln, io_object, io_context);
4642 353690 : BufferLockUnlock(buffer, buf);
4643 353690 : }
4644 :
4645 : /*
4646 : * RelationGetNumberOfBlocksInFork
4647 : * Determines the current number of pages in the specified relation fork.
4648 : *
4649 : * Note that the accuracy of the result will depend on the details of the
4650 : * relation's storage. For builtin AMs it'll be accurate, but for external AMs
4651 : * it might not be.
4652 : */
4653 : BlockNumber
4654 2367042 : RelationGetNumberOfBlocksInFork(Relation relation, ForkNumber forkNum)
4655 : {
4656 2367042 : if (RELKIND_HAS_TABLE_AM(relation->rd_rel->relkind))
4657 : {
4658 : /*
4659 : * Not every table AM uses BLCKSZ wide fixed size blocks. Therefore
4660 : * tableam returns the size in bytes - but for the purpose of this
4661 : * routine, we want the number of blocks. Therefore divide, rounding
4662 : * up.
4663 : */
4664 : uint64 szbytes;
4665 :
4666 1698345 : szbytes = table_relation_size(relation, forkNum);
4667 :
4668 1698326 : return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
4669 : }
4670 668697 : else if (RELKIND_HAS_STORAGE(relation->rd_rel->relkind))
4671 : {
4672 668697 : return smgrnblocks(RelationGetSmgr(relation), forkNum);
4673 : }
4674 : else
4675 : Assert(false);
4676 :
4677 0 : return 0; /* keep compiler quiet */
4678 : }
4679 :
4680 : /*
4681 : * BufferIsPermanent
4682 : * Determines whether a buffer will potentially still be around after
4683 : * a crash. Caller must hold a buffer pin.
4684 : */
4685 : bool
4686 17317466 : BufferIsPermanent(Buffer buffer)
4687 : {
4688 : BufferDesc *bufHdr;
4689 :
4690 : /* Local buffers are used only for temp relations. */
4691 17317466 : if (BufferIsLocal(buffer))
4692 828419 : return false;
4693 :
4694 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4695 : Assert(BufferIsValid(buffer));
4696 : Assert(BufferIsPinned(buffer));
4697 :
4698 : /*
4699 : * BM_PERMANENT can't be changed while we hold a pin on the buffer, so we
4700 : * need not bother with the buffer header spinlock. Even if someone else
4701 : * changes the buffer header state while we're doing this, the state is
4702 : * changed atomically, so we'll read the old value or the new value, but
4703 : * not random garbage.
4704 : */
4705 16489047 : bufHdr = GetBufferDescriptor(buffer - 1);
4706 16489047 : return (pg_atomic_read_u64(&bufHdr->state) & BM_PERMANENT) != 0;
4707 : }
4708 :
4709 : /*
4710 : * BufferGetLSNAtomic
4711 : * Retrieves the LSN of the buffer atomically.
4712 : *
4713 : * This is necessary for some callers who may only hold a share lock on
4714 : * the buffer. A share lock allows a concurrent backend to set hint bits
4715 : * on the page, which in turn may require a WAL record to be emitted.
4716 : *
4717 : * On platforms with 8 byte atomic reads/writes, we don't need to do any
4718 : * additional locking. On platforms not supporting such 8 byte atomic
4719 : * reads/writes, we need to actually take the header lock.
4720 : */
4721 : XLogRecPtr
4722 9011107 : BufferGetLSNAtomic(Buffer buffer)
4723 : {
4724 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4725 : Assert(BufferIsValid(buffer));
4726 : Assert(BufferIsPinned(buffer));
4727 :
4728 : #ifdef PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY
4729 9011107 : return PageGetLSN(BufferGetPage(buffer));
4730 : #else
4731 : {
4732 : char *page = BufferGetPage(buffer);
4733 : BufferDesc *bufHdr;
4734 : XLogRecPtr lsn;
4735 :
4736 : /*
4737 : * If we don't need locking for correctness, fastpath out.
4738 : */
4739 : if (!XLogHintBitIsNeeded() || BufferIsLocal(buffer))
4740 : return PageGetLSN(page);
4741 :
4742 : bufHdr = GetBufferDescriptor(buffer - 1);
4743 : LockBufHdr(bufHdr);
4744 : lsn = PageGetLSN(page);
4745 : UnlockBufHdr(bufHdr);
4746 :
4747 : return lsn;
4748 : }
4749 : #endif
4750 : }
4751 :
4752 : /* ---------------------------------------------------------------------
4753 : * DropRelationBuffers
4754 : *
4755 : * This function removes from the buffer pool all the pages of the
4756 : * specified relation forks that have block numbers >= firstDelBlock.
4757 : * (In particular, with firstDelBlock = 0, all pages are removed.)
4758 : * Dirty pages are simply dropped, without bothering to write them
4759 : * out first. Therefore, this is NOT rollback-able, and so should be
4760 : * used only with extreme caution!
4761 : *
4762 : * Currently, this is called only from smgr.c when the underlying file
4763 : * is about to be deleted or truncated (firstDelBlock is needed for
4764 : * the truncation case). The data in the affected pages would therefore
4765 : * be deleted momentarily anyway, and there is no point in writing it.
4766 : * It is the responsibility of higher-level code to ensure that the
4767 : * deletion or truncation does not lose any data that could be needed
4768 : * later. It is also the responsibility of higher-level code to ensure
4769 : * that no other process could be trying to load more pages of the
4770 : * relation into buffers.
4771 : * --------------------------------------------------------------------
4772 : */
4773 : void
4774 803 : DropRelationBuffers(SMgrRelation smgr_reln, ForkNumber *forkNum,
4775 : int nforks, BlockNumber *firstDelBlock)
4776 : {
4777 : int i;
4778 : int j;
4779 : RelFileLocatorBackend rlocator;
4780 : BlockNumber nForkBlock[MAX_FORKNUM];
4781 803 : uint64 nBlocksToInvalidate = 0;
4782 :
4783 803 : rlocator = smgr_reln->smgr_rlocator;
4784 :
4785 : /* If it's a local relation, it's localbuf.c's problem. */
4786 803 : if (RelFileLocatorBackendIsTemp(rlocator))
4787 : {
4788 498 : if (rlocator.backend == MyProcNumber)
4789 498 : DropRelationLocalBuffers(rlocator.locator, forkNum, nforks,
4790 : firstDelBlock);
4791 :
4792 538 : return;
4793 : }
4794 :
4795 : /*
4796 : * To remove all the pages of the specified relation forks from the buffer
4797 : * pool, we need to scan the entire buffer pool but we can optimize it by
4798 : * finding the buffers from BufMapping table provided we know the exact
4799 : * size of each fork of the relation. The exact size is required to ensure
4800 : * that we don't leave any buffer for the relation being dropped as
4801 : * otherwise the background writer or checkpointer can lead to a PANIC
4802 : * error while flushing buffers corresponding to files that don't exist.
4803 : *
4804 : * To know the exact size, we rely on the size cached for each fork by us
4805 : * during recovery which limits the optimization to recovery and on
4806 : * standbys but we can easily extend it once we have shared cache for
4807 : * relation size.
4808 : *
4809 : * In recovery, we cache the value returned by the first lseek(SEEK_END)
4810 : * and the future writes keeps the cached value up-to-date. See
4811 : * smgrextend. It is possible that the value of the first lseek is smaller
4812 : * than the actual number of existing blocks in the file due to buggy
4813 : * Linux kernels that might not have accounted for the recent write. But
4814 : * that should be fine because there must not be any buffers after that
4815 : * file size.
4816 : */
4817 403 : for (i = 0; i < nforks; i++)
4818 : {
4819 : /* Get the number of blocks for a relation's fork */
4820 352 : nForkBlock[i] = smgrnblocks_cached(smgr_reln, forkNum[i]);
4821 :
4822 352 : if (nForkBlock[i] == InvalidBlockNumber)
4823 : {
4824 254 : nBlocksToInvalidate = InvalidBlockNumber;
4825 254 : break;
4826 : }
4827 :
4828 : /* calculate the number of blocks to be invalidated */
4829 98 : nBlocksToInvalidate += (nForkBlock[i] - firstDelBlock[i]);
4830 : }
4831 :
4832 : /*
4833 : * We apply the optimization iff the total number of blocks to invalidate
4834 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4835 : */
4836 305 : if (BlockNumberIsValid(nBlocksToInvalidate) &&
4837 51 : nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4838 : {
4839 112 : for (j = 0; j < nforks; j++)
4840 72 : FindAndDropRelationBuffers(rlocator.locator, forkNum[j],
4841 72 : nForkBlock[j], firstDelBlock[j]);
4842 40 : return;
4843 : }
4844 :
4845 3626761 : for (i = 0; i < NBuffers; i++)
4846 : {
4847 3626496 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4848 :
4849 : /*
4850 : * We can make this a tad faster by prechecking the buffer tag before
4851 : * we attempt to lock the buffer; this saves a lot of lock
4852 : * acquisitions in typical cases. It should be safe because the
4853 : * caller must have AccessExclusiveLock on the relation, or some other
4854 : * reason to be certain that no one is loading new pages of the rel
4855 : * into the buffer pool. (Otherwise we might well miss such pages
4856 : * entirely.) Therefore, while the tag might be changing while we
4857 : * look at it, it can't be changing *to* a value we care about, only
4858 : * *away* from such a value. So false negatives are impossible, and
4859 : * false positives are safe because we'll recheck after getting the
4860 : * buffer lock.
4861 : *
4862 : * We could check forkNum and blockNum as well as the rlocator, but
4863 : * the incremental win from doing so seems small.
4864 : */
4865 3626496 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator))
4866 3615996 : continue;
4867 :
4868 10500 : LockBufHdr(bufHdr);
4869 :
4870 26257 : for (j = 0; j < nforks; j++)
4871 : {
4872 18554 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator) &&
4873 18554 : BufTagGetForkNum(&bufHdr->tag) == forkNum[j] &&
4874 10378 : bufHdr->tag.blockNum >= firstDelBlock[j])
4875 : {
4876 2797 : InvalidateBuffer(bufHdr); /* releases spinlock */
4877 2797 : break;
4878 : }
4879 : }
4880 10500 : if (j >= nforks)
4881 7703 : UnlockBufHdr(bufHdr);
4882 : }
4883 : }
4884 :
4885 : /* ---------------------------------------------------------------------
4886 : * DropRelationsAllBuffers
4887 : *
4888 : * This function removes from the buffer pool all the pages of all
4889 : * forks of the specified relations. It's equivalent to calling
4890 : * DropRelationBuffers once per fork per relation with firstDelBlock = 0.
4891 : * --------------------------------------------------------------------
4892 : */
4893 : void
4894 18110 : DropRelationsAllBuffers(SMgrRelation *smgr_reln, int nlocators)
4895 : {
4896 : int i;
4897 18110 : int n = 0;
4898 : SMgrRelation *rels;
4899 : BlockNumber (*block)[MAX_FORKNUM + 1];
4900 18110 : uint64 nBlocksToInvalidate = 0;
4901 : RelFileLocator *locators;
4902 18110 : bool cached = true;
4903 : bool use_bsearch;
4904 :
4905 18110 : if (nlocators == 0)
4906 0 : return;
4907 :
4908 18110 : rels = palloc_array(SMgrRelation, nlocators); /* non-local relations */
4909 :
4910 : /* If it's a local relation, it's localbuf.c's problem. */
4911 79670 : for (i = 0; i < nlocators; i++)
4912 : {
4913 61560 : if (RelFileLocatorBackendIsTemp(smgr_reln[i]->smgr_rlocator))
4914 : {
4915 4426 : if (smgr_reln[i]->smgr_rlocator.backend == MyProcNumber)
4916 4424 : DropRelationAllLocalBuffers(smgr_reln[i]->smgr_rlocator.locator);
4917 : }
4918 : else
4919 57134 : rels[n++] = smgr_reln[i];
4920 : }
4921 :
4922 : /*
4923 : * If there are no non-local relations, then we're done. Release the
4924 : * memory and return.
4925 : */
4926 18110 : if (n == 0)
4927 : {
4928 1189 : pfree(rels);
4929 1189 : return;
4930 : }
4931 :
4932 : /*
4933 : * This is used to remember the number of blocks for all the relations
4934 : * forks.
4935 : */
4936 : block = (BlockNumber (*)[MAX_FORKNUM + 1])
4937 16921 : palloc(sizeof(BlockNumber) * n * (MAX_FORKNUM + 1));
4938 :
4939 : /*
4940 : * We can avoid scanning the entire buffer pool if we know the exact size
4941 : * of each of the given relation forks. See DropRelationBuffers.
4942 : */
4943 35266 : for (i = 0; i < n && cached; i++)
4944 : {
4945 27575 : for (int j = 0; j <= MAX_FORKNUM; j++)
4946 : {
4947 : /* Get the number of blocks for a relation's fork. */
4948 25284 : block[i][j] = smgrnblocks_cached(rels[i], j);
4949 :
4950 : /* We need to only consider the relation forks that exists. */
4951 25284 : if (block[i][j] == InvalidBlockNumber)
4952 : {
4953 22796 : if (!smgrexists(rels[i], j))
4954 6742 : continue;
4955 16054 : cached = false;
4956 16054 : break;
4957 : }
4958 :
4959 : /* calculate the total number of blocks to be invalidated */
4960 2488 : nBlocksToInvalidate += block[i][j];
4961 : }
4962 : }
4963 :
4964 : /*
4965 : * We apply the optimization iff the total number of blocks to invalidate
4966 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4967 : */
4968 16921 : if (cached && nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4969 : {
4970 1444 : for (i = 0; i < n; i++)
4971 : {
4972 3990 : for (int j = 0; j <= MAX_FORKNUM; j++)
4973 : {
4974 : /* ignore relation forks that doesn't exist */
4975 3192 : if (!BlockNumberIsValid(block[i][j]))
4976 2385 : continue;
4977 :
4978 : /* drop all the buffers for a particular relation fork */
4979 807 : FindAndDropRelationBuffers(rels[i]->smgr_rlocator.locator,
4980 807 : j, block[i][j], 0);
4981 : }
4982 : }
4983 :
4984 646 : pfree(block);
4985 646 : pfree(rels);
4986 646 : return;
4987 : }
4988 :
4989 16275 : pfree(block);
4990 16275 : locators = palloc_array(RelFileLocator, n); /* non-local relations */
4991 72611 : for (i = 0; i < n; i++)
4992 56336 : locators[i] = rels[i]->smgr_rlocator.locator;
4993 :
4994 : /*
4995 : * For low number of relations to drop just use a simple walk through, to
4996 : * save the bsearch overhead. The threshold to use is rather a guess than
4997 : * an exactly determined value, as it depends on many factors (CPU and RAM
4998 : * speeds, amount of shared buffers etc.).
4999 : */
5000 16275 : use_bsearch = n > RELS_BSEARCH_THRESHOLD;
5001 :
5002 : /* sort the list of rlocators if necessary */
5003 16275 : if (use_bsearch)
5004 218 : qsort(locators, n, sizeof(RelFileLocator), rlocator_comparator);
5005 :
5006 189823763 : for (i = 0; i < NBuffers; i++)
5007 : {
5008 189807488 : RelFileLocator *rlocator = NULL;
5009 189807488 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5010 :
5011 : /*
5012 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5013 : * saves some cycles.
5014 : */
5015 :
5016 189807488 : if (!use_bsearch)
5017 : {
5018 : int j;
5019 :
5020 756943142 : for (j = 0; j < n; j++)
5021 : {
5022 569740390 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &locators[j]))
5023 : {
5024 105920 : rlocator = &locators[j];
5025 105920 : break;
5026 : }
5027 : }
5028 : }
5029 : else
5030 : {
5031 : RelFileLocator locator;
5032 :
5033 2498816 : locator = BufTagGetRelFileLocator(&bufHdr->tag);
5034 2498816 : rlocator = bsearch(&locator,
5035 : locators, n, sizeof(RelFileLocator),
5036 : rlocator_comparator);
5037 : }
5038 :
5039 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
5040 189807488 : if (rlocator == NULL)
5041 189699801 : continue;
5042 :
5043 107687 : LockBufHdr(bufHdr);
5044 107687 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, rlocator))
5045 107687 : InvalidateBuffer(bufHdr); /* releases spinlock */
5046 : else
5047 0 : UnlockBufHdr(bufHdr);
5048 : }
5049 :
5050 16275 : pfree(locators);
5051 16275 : pfree(rels);
5052 : }
5053 :
5054 : /* ---------------------------------------------------------------------
5055 : * FindAndDropRelationBuffers
5056 : *
5057 : * This function performs look up in BufMapping table and removes from the
5058 : * buffer pool all the pages of the specified relation fork that has block
5059 : * number >= firstDelBlock. (In particular, with firstDelBlock = 0, all
5060 : * pages are removed.)
5061 : * --------------------------------------------------------------------
5062 : */
5063 : static void
5064 879 : FindAndDropRelationBuffers(RelFileLocator rlocator, ForkNumber forkNum,
5065 : BlockNumber nForkBlock,
5066 : BlockNumber firstDelBlock)
5067 : {
5068 : BlockNumber curBlock;
5069 :
5070 2107 : for (curBlock = firstDelBlock; curBlock < nForkBlock; curBlock++)
5071 : {
5072 : uint32 bufHash; /* hash value for tag */
5073 : BufferTag bufTag; /* identity of requested block */
5074 : LWLock *bufPartitionLock; /* buffer partition lock for it */
5075 : int buf_id;
5076 : BufferDesc *bufHdr;
5077 :
5078 : /* create a tag so we can lookup the buffer */
5079 1228 : InitBufferTag(&bufTag, &rlocator, forkNum, curBlock);
5080 :
5081 : /* determine its hash code and partition lock ID */
5082 1228 : bufHash = BufTableHashCode(&bufTag);
5083 1228 : bufPartitionLock = BufMappingPartitionLock(bufHash);
5084 :
5085 : /* Check that it is in the buffer pool. If not, do nothing. */
5086 1228 : LWLockAcquire(bufPartitionLock, LW_SHARED);
5087 1228 : buf_id = BufTableLookup(&bufTag, bufHash);
5088 1228 : LWLockRelease(bufPartitionLock);
5089 :
5090 1228 : if (buf_id < 0)
5091 112 : continue;
5092 :
5093 1116 : bufHdr = GetBufferDescriptor(buf_id);
5094 :
5095 : /*
5096 : * We need to lock the buffer header and recheck if the buffer is
5097 : * still associated with the same block because the buffer could be
5098 : * evicted by some other backend loading blocks for a different
5099 : * relation after we release lock on the BufMapping table.
5100 : */
5101 1116 : LockBufHdr(bufHdr);
5102 :
5103 2232 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator) &&
5104 1116 : BufTagGetForkNum(&bufHdr->tag) == forkNum &&
5105 1116 : bufHdr->tag.blockNum >= firstDelBlock)
5106 1116 : InvalidateBuffer(bufHdr); /* releases spinlock */
5107 : else
5108 0 : UnlockBufHdr(bufHdr);
5109 : }
5110 879 : }
5111 :
5112 : /* ---------------------------------------------------------------------
5113 : * DropDatabaseBuffers
5114 : *
5115 : * This function removes all the buffers in the buffer cache for a
5116 : * particular database. Dirty pages are simply dropped, without
5117 : * bothering to write them out first. This is used when we destroy a
5118 : * database, to avoid trying to flush data to disk when the directory
5119 : * tree no longer exists. Implementation is pretty similar to
5120 : * DropRelationBuffers() which is for destroying just one relation.
5121 : * --------------------------------------------------------------------
5122 : */
5123 : void
5124 81 : DropDatabaseBuffers(Oid dbid)
5125 : {
5126 : int i;
5127 :
5128 : /*
5129 : * We needn't consider local buffers, since by assumption the target
5130 : * database isn't our own.
5131 : */
5132 :
5133 611921 : for (i = 0; i < NBuffers; i++)
5134 : {
5135 611840 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5136 :
5137 : /*
5138 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5139 : * saves some cycles.
5140 : */
5141 611840 : if (bufHdr->tag.dbOid != dbid)
5142 597159 : continue;
5143 :
5144 14681 : LockBufHdr(bufHdr);
5145 14681 : if (bufHdr->tag.dbOid == dbid)
5146 14681 : InvalidateBuffer(bufHdr); /* releases spinlock */
5147 : else
5148 0 : UnlockBufHdr(bufHdr);
5149 : }
5150 81 : }
5151 :
5152 : /* ---------------------------------------------------------------------
5153 : * FlushRelationBuffers
5154 : *
5155 : * This function writes all dirty pages of a relation out to disk
5156 : * (or more accurately, out to kernel disk buffers), ensuring that the
5157 : * kernel has an up-to-date view of the relation.
5158 : *
5159 : * Generally, the caller should be holding AccessExclusiveLock on the
5160 : * target relation to ensure that no other backend is busy dirtying
5161 : * more blocks of the relation; the effects can't be expected to last
5162 : * after the lock is released.
5163 : *
5164 : * XXX currently it sequentially searches the buffer pool, should be
5165 : * changed to more clever ways of searching. This routine is not
5166 : * used in any performance-critical code paths, so it's not worth
5167 : * adding additional overhead to normal paths to make it go faster.
5168 : * --------------------------------------------------------------------
5169 : */
5170 : void
5171 168 : FlushRelationBuffers(Relation rel)
5172 : {
5173 : int i;
5174 : BufferDesc *bufHdr;
5175 168 : SMgrRelation srel = RelationGetSmgr(rel);
5176 :
5177 168 : if (RelationUsesLocalBuffers(rel))
5178 : {
5179 1212 : for (i = 0; i < NLocBuffer; i++)
5180 : {
5181 : uint64 buf_state;
5182 :
5183 1200 : bufHdr = GetLocalBufferDescriptor(i);
5184 1200 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
5185 400 : ((buf_state = pg_atomic_read_u64(&bufHdr->state)) &
5186 : (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5187 : {
5188 : ErrorContextCallback errcallback;
5189 :
5190 : /* Setup error traceback support for ereport() */
5191 392 : errcallback.callback = local_buffer_write_error_callback;
5192 392 : errcallback.arg = bufHdr;
5193 392 : errcallback.previous = error_context_stack;
5194 392 : error_context_stack = &errcallback;
5195 :
5196 : /* Make sure we can handle the pin */
5197 392 : ReservePrivateRefCountEntry();
5198 392 : ResourceOwnerEnlarge(CurrentResourceOwner);
5199 :
5200 : /*
5201 : * Pin/unpin mostly to make valgrind work, but it also seems
5202 : * like the right thing to do.
5203 : */
5204 392 : PinLocalBuffer(bufHdr, false);
5205 :
5206 :
5207 392 : FlushLocalBuffer(bufHdr, srel);
5208 :
5209 392 : UnpinLocalBuffer(BufferDescriptorGetBuffer(bufHdr));
5210 :
5211 : /* Pop the error context stack */
5212 392 : error_context_stack = errcallback.previous;
5213 : }
5214 : }
5215 :
5216 12 : return;
5217 : }
5218 :
5219 1938332 : for (i = 0; i < NBuffers; i++)
5220 : {
5221 : uint64 buf_state;
5222 :
5223 1938176 : bufHdr = GetBufferDescriptor(i);
5224 :
5225 : /*
5226 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5227 : * saves some cycles.
5228 : */
5229 1938176 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator))
5230 1937929 : continue;
5231 :
5232 : /* Make sure we can handle the pin */
5233 247 : ReservePrivateRefCountEntry();
5234 247 : ResourceOwnerEnlarge(CurrentResourceOwner);
5235 :
5236 247 : buf_state = LockBufHdr(bufHdr);
5237 247 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
5238 247 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5239 : {
5240 205 : PinBuffer_Locked(bufHdr);
5241 205 : FlushUnlockedBuffer(bufHdr, srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5242 205 : UnpinBuffer(bufHdr);
5243 : }
5244 : else
5245 42 : UnlockBufHdr(bufHdr);
5246 : }
5247 : }
5248 :
5249 : /* ---------------------------------------------------------------------
5250 : * FlushRelationsAllBuffers
5251 : *
5252 : * This function flushes out of the buffer pool all the pages of all
5253 : * forks of the specified smgr relations. It's equivalent to calling
5254 : * FlushRelationBuffers once per relation. The relations are assumed not
5255 : * to use local buffers.
5256 : * --------------------------------------------------------------------
5257 : */
5258 : void
5259 4 : FlushRelationsAllBuffers(SMgrRelation *smgrs, int nrels)
5260 : {
5261 : int i;
5262 : SMgrSortArray *srels;
5263 : bool use_bsearch;
5264 :
5265 4 : if (nrels == 0)
5266 0 : return;
5267 :
5268 : /* fill-in array for qsort */
5269 4 : srels = palloc_array(SMgrSortArray, nrels);
5270 :
5271 8 : for (i = 0; i < nrels; i++)
5272 : {
5273 : Assert(!RelFileLocatorBackendIsTemp(smgrs[i]->smgr_rlocator));
5274 :
5275 4 : srels[i].rlocator = smgrs[i]->smgr_rlocator.locator;
5276 4 : srels[i].srel = smgrs[i];
5277 : }
5278 :
5279 : /*
5280 : * Save the bsearch overhead for low number of relations to sync. See
5281 : * DropRelationsAllBuffers for details.
5282 : */
5283 4 : use_bsearch = nrels > RELS_BSEARCH_THRESHOLD;
5284 :
5285 : /* sort the list of SMgrRelations if necessary */
5286 4 : if (use_bsearch)
5287 0 : qsort(srels, nrels, sizeof(SMgrSortArray), rlocator_comparator);
5288 :
5289 65540 : for (i = 0; i < NBuffers; i++)
5290 : {
5291 65536 : SMgrSortArray *srelent = NULL;
5292 65536 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5293 : uint64 buf_state;
5294 :
5295 : /*
5296 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5297 : * saves some cycles.
5298 : */
5299 :
5300 65536 : if (!use_bsearch)
5301 : {
5302 : int j;
5303 :
5304 128807 : for (j = 0; j < nrels; j++)
5305 : {
5306 65536 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srels[j].rlocator))
5307 : {
5308 2265 : srelent = &srels[j];
5309 2265 : break;
5310 : }
5311 : }
5312 : }
5313 : else
5314 : {
5315 : RelFileLocator rlocator;
5316 :
5317 0 : rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
5318 0 : srelent = bsearch(&rlocator,
5319 : srels, nrels, sizeof(SMgrSortArray),
5320 : rlocator_comparator);
5321 : }
5322 :
5323 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
5324 65536 : if (srelent == NULL)
5325 63271 : continue;
5326 :
5327 : /* Make sure we can handle the pin */
5328 2265 : ReservePrivateRefCountEntry();
5329 2265 : ResourceOwnerEnlarge(CurrentResourceOwner);
5330 :
5331 2265 : buf_state = LockBufHdr(bufHdr);
5332 2265 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srelent->rlocator) &&
5333 2265 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5334 : {
5335 2231 : PinBuffer_Locked(bufHdr);
5336 2231 : FlushUnlockedBuffer(bufHdr, srelent->srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5337 2231 : UnpinBuffer(bufHdr);
5338 : }
5339 : else
5340 34 : UnlockBufHdr(bufHdr);
5341 : }
5342 :
5343 4 : pfree(srels);
5344 : }
5345 :
5346 : /* ---------------------------------------------------------------------
5347 : * RelationCopyStorageUsingBuffer
5348 : *
5349 : * Copy fork's data using bufmgr. Same as RelationCopyStorage but instead
5350 : * of using smgrread and smgrextend this will copy using bufmgr APIs.
5351 : *
5352 : * Refer comments atop CreateAndCopyRelationData() for details about
5353 : * 'permanent' parameter.
5354 : * --------------------------------------------------------------------
5355 : */
5356 : static void
5357 83348 : RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
5358 : RelFileLocator dstlocator,
5359 : ForkNumber forkNum, bool permanent)
5360 : {
5361 : Buffer srcBuf;
5362 : Buffer dstBuf;
5363 : Page srcPage;
5364 : Page dstPage;
5365 : bool use_wal;
5366 : BlockNumber nblocks;
5367 : BlockNumber blkno;
5368 : PGIOAlignedBlock buf;
5369 : BufferAccessStrategy bstrategy_src;
5370 : BufferAccessStrategy bstrategy_dst;
5371 : BlockRangeReadStreamPrivate p;
5372 : ReadStream *src_stream;
5373 : SMgrRelation src_smgr;
5374 :
5375 : /*
5376 : * In general, we want to write WAL whenever wal_level > 'minimal', but we
5377 : * can skip it when copying any fork of an unlogged relation other than
5378 : * the init fork.
5379 : */
5380 83348 : use_wal = XLogIsNeeded() && (permanent || forkNum == INIT_FORKNUM);
5381 :
5382 : /* Get number of blocks in the source relation. */
5383 83348 : nblocks = smgrnblocks(smgropen(srclocator, INVALID_PROC_NUMBER),
5384 : forkNum);
5385 :
5386 : /* Nothing to copy; just return. */
5387 83348 : if (nblocks == 0)
5388 15461 : return;
5389 :
5390 : /*
5391 : * Bulk extend the destination relation of the same size as the source
5392 : * relation before starting to copy block by block.
5393 : */
5394 67887 : memset(buf.data, 0, BLCKSZ);
5395 67887 : smgrextend(smgropen(dstlocator, INVALID_PROC_NUMBER), forkNum, nblocks - 1,
5396 : buf.data, true);
5397 :
5398 : /* This is a bulk operation, so use buffer access strategies. */
5399 67887 : bstrategy_src = GetAccessStrategy(BAS_BULKREAD);
5400 67887 : bstrategy_dst = GetAccessStrategy(BAS_BULKWRITE);
5401 :
5402 : /* Initialize streaming read */
5403 67887 : p.current_blocknum = 0;
5404 67887 : p.last_exclusive = nblocks;
5405 67887 : src_smgr = smgropen(srclocator, INVALID_PROC_NUMBER);
5406 :
5407 : /*
5408 : * It is safe to use batchmode as block_range_read_stream_cb takes no
5409 : * locks.
5410 : */
5411 67887 : src_stream = read_stream_begin_smgr_relation(READ_STREAM_FULL |
5412 : READ_STREAM_USE_BATCHING,
5413 : bstrategy_src,
5414 : src_smgr,
5415 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
5416 : forkNum,
5417 : block_range_read_stream_cb,
5418 : &p,
5419 : 0);
5420 :
5421 : /* Iterate over each block of the source relation file. */
5422 320845 : for (blkno = 0; blkno < nblocks; blkno++)
5423 : {
5424 252960 : CHECK_FOR_INTERRUPTS();
5425 :
5426 : /* Read block from source relation. */
5427 252960 : srcBuf = read_stream_next_buffer(src_stream, NULL);
5428 252958 : LockBuffer(srcBuf, BUFFER_LOCK_SHARE);
5429 252958 : srcPage = BufferGetPage(srcBuf);
5430 :
5431 252958 : dstBuf = ReadBufferWithoutRelcache(dstlocator, forkNum,
5432 : BufferGetBlockNumber(srcBuf),
5433 : RBM_ZERO_AND_LOCK, bstrategy_dst,
5434 : permanent);
5435 252958 : dstPage = BufferGetPage(dstBuf);
5436 :
5437 252958 : START_CRIT_SECTION();
5438 :
5439 : /* Copy page data from the source to the destination. */
5440 252958 : memcpy(dstPage, srcPage, BLCKSZ);
5441 252958 : MarkBufferDirty(dstBuf);
5442 :
5443 : /* WAL-log the copied page. */
5444 252958 : if (use_wal)
5445 147740 : log_newpage_buffer(dstBuf, true);
5446 :
5447 252958 : END_CRIT_SECTION();
5448 :
5449 252958 : UnlockReleaseBuffer(dstBuf);
5450 252958 : UnlockReleaseBuffer(srcBuf);
5451 : }
5452 : Assert(read_stream_next_buffer(src_stream, NULL) == InvalidBuffer);
5453 67885 : read_stream_end(src_stream);
5454 :
5455 67885 : FreeAccessStrategy(bstrategy_src);
5456 67885 : FreeAccessStrategy(bstrategy_dst);
5457 : }
5458 :
5459 : /* ---------------------------------------------------------------------
5460 : * CreateAndCopyRelationData
5461 : *
5462 : * Create destination relation storage and copy all forks from the
5463 : * source relation to the destination.
5464 : *
5465 : * Pass permanent as true for permanent relations and false for
5466 : * unlogged relations. Currently this API is not supported for
5467 : * temporary relations.
5468 : * --------------------------------------------------------------------
5469 : */
5470 : void
5471 63956 : CreateAndCopyRelationData(RelFileLocator src_rlocator,
5472 : RelFileLocator dst_rlocator, bool permanent)
5473 : {
5474 : char relpersistence;
5475 : SMgrRelation src_rel;
5476 : SMgrRelation dst_rel;
5477 :
5478 : /* Set the relpersistence. */
5479 63956 : relpersistence = permanent ?
5480 : RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED;
5481 :
5482 63956 : src_rel = smgropen(src_rlocator, INVALID_PROC_NUMBER);
5483 63956 : dst_rel = smgropen(dst_rlocator, INVALID_PROC_NUMBER);
5484 :
5485 : /*
5486 : * Create and copy all forks of the relation. During create database we
5487 : * have a separate cleanup mechanism which deletes complete database
5488 : * directory. Therefore, each individual relation doesn't need to be
5489 : * registered for cleanup.
5490 : */
5491 63956 : RelationCreateStorage(dst_rlocator, relpersistence, false);
5492 :
5493 : /* copy main fork. */
5494 63956 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, MAIN_FORKNUM,
5495 : permanent);
5496 :
5497 : /* copy those extra forks that exist */
5498 63954 : for (ForkNumber forkNum = MAIN_FORKNUM + 1;
5499 255816 : forkNum <= MAX_FORKNUM; forkNum++)
5500 : {
5501 191862 : if (smgrexists(src_rel, forkNum))
5502 : {
5503 19392 : smgrcreate(dst_rel, forkNum, false);
5504 :
5505 : /*
5506 : * WAL log creation if the relation is persistent, or this is the
5507 : * init fork of an unlogged relation.
5508 : */
5509 19392 : if (permanent || forkNum == INIT_FORKNUM)
5510 19392 : log_smgrcreate(&dst_rlocator, forkNum);
5511 :
5512 : /* Copy a fork's data, block by block. */
5513 19392 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, forkNum,
5514 : permanent);
5515 : }
5516 : }
5517 63954 : }
5518 :
5519 : /* ---------------------------------------------------------------------
5520 : * FlushDatabaseBuffers
5521 : *
5522 : * This function writes all dirty pages of a database out to disk
5523 : * (or more accurately, out to kernel disk buffers), ensuring that the
5524 : * kernel has an up-to-date view of the database.
5525 : *
5526 : * Generally, the caller should be holding an appropriate lock to ensure
5527 : * no other backend is active in the target database; otherwise more
5528 : * pages could get dirtied.
5529 : *
5530 : * Note we don't worry about flushing any pages of temporary relations.
5531 : * It's assumed these wouldn't be interesting.
5532 : * --------------------------------------------------------------------
5533 : */
5534 : void
5535 5 : FlushDatabaseBuffers(Oid dbid)
5536 : {
5537 : int i;
5538 : BufferDesc *bufHdr;
5539 :
5540 645 : for (i = 0; i < NBuffers; i++)
5541 : {
5542 : uint64 buf_state;
5543 :
5544 640 : bufHdr = GetBufferDescriptor(i);
5545 :
5546 : /*
5547 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5548 : * saves some cycles.
5549 : */
5550 640 : if (bufHdr->tag.dbOid != dbid)
5551 497 : continue;
5552 :
5553 : /* Make sure we can handle the pin */
5554 143 : ReservePrivateRefCountEntry();
5555 143 : ResourceOwnerEnlarge(CurrentResourceOwner);
5556 :
5557 143 : buf_state = LockBufHdr(bufHdr);
5558 143 : if (bufHdr->tag.dbOid == dbid &&
5559 143 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5560 : {
5561 7 : PinBuffer_Locked(bufHdr);
5562 7 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5563 7 : UnpinBuffer(bufHdr);
5564 : }
5565 : else
5566 136 : UnlockBufHdr(bufHdr);
5567 : }
5568 5 : }
5569 :
5570 : /*
5571 : * Flush a previously, share-exclusively or exclusively, locked and pinned
5572 : * buffer to the OS.
5573 : */
5574 : void
5575 112 : FlushOneBuffer(Buffer buffer)
5576 : {
5577 : BufferDesc *bufHdr;
5578 :
5579 : /* currently not needed, but no fundamental reason not to support */
5580 : Assert(!BufferIsLocal(buffer));
5581 :
5582 : Assert(BufferIsPinned(buffer));
5583 :
5584 112 : bufHdr = GetBufferDescriptor(buffer - 1);
5585 :
5586 : Assert(BufferIsLockedByMe(buffer));
5587 :
5588 112 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5589 112 : }
5590 :
5591 : /*
5592 : * ReleaseBuffer -- release the pin on a buffer
5593 : */
5594 : void
5595 50887727 : ReleaseBuffer(Buffer buffer)
5596 : {
5597 50887727 : if (!BufferIsValid(buffer))
5598 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5599 :
5600 50887727 : if (BufferIsLocal(buffer))
5601 752124 : UnpinLocalBuffer(buffer);
5602 : else
5603 50135603 : UnpinBuffer(GetBufferDescriptor(buffer - 1));
5604 50887727 : }
5605 :
5606 : /*
5607 : * UnlockReleaseBuffer -- release the content lock and pin on a buffer
5608 : *
5609 : * This is just a, more efficient, shorthand for a common combination.
5610 : */
5611 : void
5612 50837220 : UnlockReleaseBuffer(Buffer buffer)
5613 : {
5614 : int mode;
5615 : BufferDesc *buf;
5616 : PrivateRefCountEntry *ref;
5617 : uint64 sub;
5618 : uint64 lockstate;
5619 :
5620 : Assert(BufferIsPinned(buffer));
5621 :
5622 50837220 : if (BufferIsLocal(buffer))
5623 : {
5624 1381810 : UnpinLocalBuffer(buffer);
5625 1381810 : return;
5626 : }
5627 :
5628 49455410 : ResourceOwnerForgetBuffer(CurrentResourceOwner, buffer);
5629 :
5630 49455410 : buf = GetBufferDescriptor(buffer - 1);
5631 :
5632 49455410 : mode = BufferLockDisownInternal(buffer, buf);
5633 :
5634 : /* compute state modification for lock release */
5635 49455410 : sub = BufferLockReleaseSub(mode);
5636 :
5637 : /* compute state modification for pin release */
5638 49455410 : ref = GetPrivateRefCountEntry(buffer, false);
5639 : Assert(ref != NULL);
5640 : Assert(ref->data.refcount > 0);
5641 49455410 : ref->data.refcount--;
5642 :
5643 : /* no more backend local pins, reduce shared pin count */
5644 49455410 : if (likely(ref->data.refcount == 0))
5645 : {
5646 : /* See comment in UnpinBufferNoOwner() */
5647 : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
5648 :
5649 47070248 : sub |= BUF_REFCOUNT_ONE;
5650 47070248 : ForgetPrivateRefCountEntry(ref);
5651 : }
5652 :
5653 : /* perform the lock and pin release in one atomic op */
5654 49455410 : lockstate = pg_atomic_sub_fetch_u64(&buf->state, sub);
5655 :
5656 : /* wake up waiters for the lock */
5657 49455410 : BufferLockProcessRelease(buf, mode, lockstate);
5658 :
5659 : /* wake up waiter for the pin release */
5660 49455410 : if (lockstate & BM_PIN_COUNT_WAITER)
5661 6 : WakePinCountWaiter(buf);
5662 :
5663 : /*
5664 : * Now okay to allow cancel/die interrupts again, which were held when the
5665 : * lock was acquired.
5666 : */
5667 49455410 : RESUME_INTERRUPTS();
5668 : }
5669 :
5670 : /*
5671 : * IncrBufferRefCount
5672 : * Increment the pin count on a buffer that we have *already* pinned
5673 : * at least once.
5674 : *
5675 : * This function cannot be used on a buffer we do not have pinned,
5676 : * because it doesn't change the shared buffer state.
5677 : */
5678 : void
5679 15499086 : IncrBufferRefCount(Buffer buffer)
5680 : {
5681 : Assert(BufferIsPinned(buffer));
5682 15499086 : ResourceOwnerEnlarge(CurrentResourceOwner);
5683 15499086 : if (BufferIsLocal(buffer))
5684 469910 : LocalRefCount[-buffer - 1]++;
5685 : else
5686 : {
5687 : PrivateRefCountEntry *ref;
5688 :
5689 15029176 : ref = GetPrivateRefCountEntry(buffer, true);
5690 : Assert(ref != NULL);
5691 15029176 : ref->data.refcount++;
5692 : }
5693 15499086 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buffer);
5694 15499086 : }
5695 :
5696 : /*
5697 : * Shared-buffer only helper for MarkBufferDirtyHint() and
5698 : * BufferSetHintBits16().
5699 : *
5700 : * This is separated out because it turns out that the repeated checks for
5701 : * local buffers, repeated GetBufferDescriptor() and repeated reading of the
5702 : * buffer's state sufficiently hurts the performance of BufferSetHintBits16().
5703 : */
5704 : static inline void
5705 15116389 : MarkSharedBufferDirtyHint(Buffer buffer, BufferDesc *bufHdr, uint64 lockstate,
5706 : bool buffer_std)
5707 : {
5708 15116389 : Page page = BufferGetPage(buffer);
5709 :
5710 : Assert(GetPrivateRefCount(buffer) > 0);
5711 :
5712 : /* here, either share-exclusive or exclusive lock is OK */
5713 : Assert(BufferLockHeldByMeInMode(bufHdr, BUFFER_LOCK_EXCLUSIVE) ||
5714 : BufferLockHeldByMeInMode(bufHdr, BUFFER_LOCK_SHARE_EXCLUSIVE));
5715 :
5716 : /*
5717 : * This routine might get called many times on the same page, if we are
5718 : * making the first scan after commit of an xact that added/deleted many
5719 : * tuples. So, be as quick as we can if the buffer is already dirty.
5720 : *
5721 : * As we are holding (at least) a share-exclusive lock, nobody could have
5722 : * cleaned or dirtied the page concurrently, so we can just rely on the
5723 : * previously fetched value here without any danger of races.
5724 : */
5725 15116389 : if (unlikely(!(lockstate & BM_DIRTY)))
5726 : {
5727 427559 : XLogRecPtr lsn = InvalidXLogRecPtr;
5728 427559 : bool wal_log = false;
5729 : uint64 buf_state;
5730 :
5731 : /*
5732 : * If we need to protect hint bit updates from torn writes, WAL-log a
5733 : * full page image of the page. This full page image is only necessary
5734 : * if the hint bit update is the first change to the page since the
5735 : * last checkpoint.
5736 : *
5737 : * We don't check full_page_writes here because that logic is included
5738 : * when we call XLogInsert() since the value changes dynamically.
5739 : */
5740 427559 : if (XLogHintBitIsNeeded() && (lockstate & BM_PERMANENT))
5741 : {
5742 : /*
5743 : * If we must not write WAL, due to a relfilelocator-specific
5744 : * condition or being in recovery, don't dirty the page. We can
5745 : * set the hint, just not dirty the page as a result so the hint
5746 : * is lost when we evict the page or shutdown.
5747 : *
5748 : * See src/backend/storage/page/README for longer discussion.
5749 : */
5750 512872 : if (RecoveryInProgress() ||
5751 87303 : RelFileLocatorSkippingWAL(BufTagGetRelFileLocator(&bufHdr->tag)))
5752 339514 : return;
5753 :
5754 86055 : wal_log = true;
5755 : }
5756 :
5757 : /*
5758 : * We must mark the page dirty before we emit the WAL record, as per
5759 : * the usual rules, to ensure that BufferSync()/SyncOneBuffer() try to
5760 : * flush the buffer, even if we haven't inserted the WAL record yet.
5761 : * As we hold at least a share-exclusive lock, checkpoints will wait
5762 : * for this backend to be done with the buffer before continuing. If
5763 : * we did it the other way round, a checkpoint could start between
5764 : * writing the WAL record and marking the buffer dirty.
5765 : */
5766 88045 : buf_state = LockBufHdr(bufHdr);
5767 :
5768 : /*
5769 : * It should not be possible for the buffer to already be dirty, see
5770 : * comment above.
5771 : */
5772 : Assert(!(buf_state & BM_DIRTY));
5773 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5774 88045 : UnlockBufHdrExt(bufHdr, buf_state,
5775 : BM_DIRTY,
5776 : 0, 0);
5777 :
5778 : /*
5779 : * If the block is already dirty because we either made a change or
5780 : * set a hint already, then we don't need to write a full page image.
5781 : * Note that aggressive cleaning of blocks dirtied by hint bit setting
5782 : * would increase the call rate. Bulk setting of hint bits would
5783 : * reduce the call rate...
5784 : */
5785 88045 : if (wal_log)
5786 86055 : lsn = XLogSaveBufferForHint(buffer, buffer_std);
5787 :
5788 88045 : if (XLogRecPtrIsValid(lsn))
5789 : {
5790 : /*
5791 : * Set the page LSN if we wrote a backup block. To allow backends
5792 : * that only hold a share lock on the buffer to read the LSN in a
5793 : * tear-free manner, we set the page LSN while holding the buffer
5794 : * header lock. This allows any reader of an LSN who holds only a
5795 : * share lock to also obtain a buffer header lock before using
5796 : * PageGetLSN() to read the LSN in a tear free way. This is done
5797 : * in BufferGetLSNAtomic().
5798 : *
5799 : * If checksums are enabled, you might think we should reset the
5800 : * checksum here. That will happen when the page is written
5801 : * sometime later in this checkpoint cycle.
5802 : */
5803 58871 : buf_state = LockBufHdr(bufHdr);
5804 58871 : PageSetLSN(page, lsn);
5805 58871 : UnlockBufHdr(bufHdr);
5806 : }
5807 :
5808 88045 : pgBufferUsage.shared_blks_dirtied++;
5809 88045 : if (VacuumCostActive)
5810 2392 : VacuumCostBalance += VacuumCostPageDirty;
5811 : }
5812 : }
5813 :
5814 : /*
5815 : * MarkBufferDirtyHint
5816 : *
5817 : * Mark a buffer dirty for non-critical changes.
5818 : *
5819 : * This is essentially the same as MarkBufferDirty, except:
5820 : *
5821 : * 1. The caller does not write WAL; so if checksums are enabled, we may need
5822 : * to write an XLOG_FPI_FOR_HINT WAL record to protect against torn pages.
5823 : * 2. The caller might have only a share-exclusive-lock instead of an
5824 : * exclusive-lock on the buffer's content lock.
5825 : * 3. This function does not guarantee that the buffer is always marked dirty
5826 : * (it e.g. can't always on a hot standby), so it cannot be used for
5827 : * important changes.
5828 : */
5829 : inline void
5830 442483 : MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
5831 : {
5832 : BufferDesc *bufHdr;
5833 :
5834 442483 : if (!BufferIsValid(buffer))
5835 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5836 :
5837 442483 : if (BufferIsLocal(buffer))
5838 : {
5839 22047 : MarkLocalBufferDirty(buffer);
5840 22047 : return;
5841 : }
5842 :
5843 420436 : bufHdr = GetBufferDescriptor(buffer - 1);
5844 :
5845 420436 : MarkSharedBufferDirtyHint(buffer, bufHdr,
5846 420436 : pg_atomic_read_u64(&bufHdr->state),
5847 : buffer_std);
5848 : }
5849 :
5850 : /*
5851 : * Release buffer content locks for shared buffers.
5852 : *
5853 : * Used to clean up after errors.
5854 : *
5855 : * Currently, we can expect that resource owner cleanup, via
5856 : * ResOwnerReleaseBuffer(), took care of releasing buffer content locks per
5857 : * se; the only thing we need to deal with here is clearing any PIN_COUNT
5858 : * request that was in progress.
5859 : */
5860 : void
5861 65305 : UnlockBuffers(void)
5862 : {
5863 65305 : BufferDesc *buf = PinCountWaitBuf;
5864 :
5865 65305 : if (buf)
5866 : {
5867 : uint64 buf_state;
5868 0 : uint64 unset_bits = 0;
5869 :
5870 0 : buf_state = LockBufHdr(buf);
5871 :
5872 : /*
5873 : * Don't complain if flag bit not set; it could have been reset but we
5874 : * got a cancel/die interrupt before getting the signal.
5875 : */
5876 0 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5877 0 : buf->wait_backend_pgprocno == MyProcNumber)
5878 0 : unset_bits = BM_PIN_COUNT_WAITER;
5879 :
5880 0 : UnlockBufHdrExt(buf, buf_state,
5881 : 0, unset_bits,
5882 : 0);
5883 :
5884 0 : PinCountWaitBuf = NULL;
5885 : }
5886 65305 : }
5887 :
5888 : /*
5889 : * Acquire the buffer content lock in the specified mode
5890 : *
5891 : * If the lock is not available, sleep until it is.
5892 : *
5893 : * Side effect: cancel/die interrupts are held off until lock release.
5894 : *
5895 : * This uses almost the same locking approach as lwlock.c's
5896 : * LWLockAcquire(). See documentation at the top of lwlock.c for a more
5897 : * detailed discussion.
5898 : *
5899 : * The reason that this, and most of the other BufferLock* functions, get both
5900 : * the Buffer and BufferDesc* as parameters, is that looking up one from the
5901 : * other repeatedly shows up noticeably in profiles.
5902 : *
5903 : * Callers should provide a constant for mode, for more efficient code
5904 : * generation.
5905 : */
5906 : static inline void
5907 112322958 : BufferLockAcquire(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode)
5908 : {
5909 : PrivateRefCountEntry *entry;
5910 112322958 : int extraWaits = 0;
5911 :
5912 : /*
5913 : * Get reference to the refcount entry before we hold the lock, it seems
5914 : * better to do before holding the lock.
5915 : */
5916 112322958 : entry = GetPrivateRefCountEntry(buffer, true);
5917 :
5918 : /*
5919 : * We better not already hold a lock on the buffer.
5920 : */
5921 : Assert(entry->data.lockmode == BUFFER_LOCK_UNLOCK);
5922 :
5923 : /*
5924 : * Lock out cancel/die interrupts until we exit the code section protected
5925 : * by the content lock. This ensures that interrupts will not interfere
5926 : * with manipulations of data structures in shared memory.
5927 : */
5928 112322958 : HOLD_INTERRUPTS();
5929 :
5930 : for (;;)
5931 19213 : {
5932 112342171 : uint32 wait_event = 0; /* initialized to avoid compiler warning */
5933 : bool mustwait;
5934 :
5935 : /*
5936 : * Try to grab the lock the first time, we're not in the waitqueue
5937 : * yet/anymore.
5938 : */
5939 112342171 : mustwait = BufferLockAttempt(buf_hdr, mode);
5940 :
5941 112342171 : if (likely(!mustwait))
5942 : {
5943 112322208 : break;
5944 : }
5945 :
5946 : /*
5947 : * Ok, at this point we couldn't grab the lock on the first try. We
5948 : * cannot simply queue ourselves to the end of the list and wait to be
5949 : * woken up because by now the lock could long have been released.
5950 : * Instead add us to the queue and try to grab the lock again. If we
5951 : * succeed we need to revert the queuing and be happy, otherwise we
5952 : * recheck the lock. If we still couldn't grab it, we know that the
5953 : * other locker will see our queue entries when releasing since they
5954 : * existed before we checked for the lock.
5955 : */
5956 :
5957 : /* add to the queue */
5958 19963 : BufferLockQueueSelf(buf_hdr, mode);
5959 :
5960 : /* we're now guaranteed to be woken up if necessary */
5961 19963 : mustwait = BufferLockAttempt(buf_hdr, mode);
5962 :
5963 : /* ok, grabbed the lock the second time round, need to undo queueing */
5964 19963 : if (!mustwait)
5965 : {
5966 750 : BufferLockDequeueSelf(buf_hdr);
5967 750 : break;
5968 : }
5969 :
5970 19213 : switch (mode)
5971 : {
5972 10846 : case BUFFER_LOCK_EXCLUSIVE:
5973 10846 : wait_event = WAIT_EVENT_BUFFER_EXCLUSIVE;
5974 10846 : break;
5975 100 : case BUFFER_LOCK_SHARE_EXCLUSIVE:
5976 100 : wait_event = WAIT_EVENT_BUFFER_SHARE_EXCLUSIVE;
5977 100 : break;
5978 8267 : case BUFFER_LOCK_SHARE:
5979 8267 : wait_event = WAIT_EVENT_BUFFER_SHARED;
5980 8267 : break;
5981 : case BUFFER_LOCK_UNLOCK:
5982 : pg_unreachable();
5983 :
5984 : }
5985 19213 : pgstat_report_wait_start(wait_event);
5986 :
5987 : /*
5988 : * Wait until awakened.
5989 : *
5990 : * It is possible that we get awakened for a reason other than being
5991 : * signaled by BufferLockWakeup(). If so, loop back and wait again.
5992 : * Once we've gotten the lock, re-increment the sema by the number of
5993 : * additional signals received.
5994 : */
5995 : for (;;)
5996 : {
5997 19213 : PGSemaphoreLock(MyProc->sem);
5998 19213 : if (MyProc->lwWaiting == LW_WS_NOT_WAITING)
5999 19213 : break;
6000 0 : extraWaits++;
6001 : }
6002 :
6003 19213 : pgstat_report_wait_end();
6004 :
6005 : /* Retrying, allow BufferLockReleaseSub to release waiters again. */
6006 19213 : pg_atomic_fetch_and_u64(&buf_hdr->state, ~BM_LOCK_WAKE_IN_PROGRESS);
6007 : }
6008 :
6009 : /* Remember that we now hold this lock */
6010 112322958 : entry->data.lockmode = mode;
6011 :
6012 : /*
6013 : * Fix the process wait semaphore's count for any absorbed wakeups.
6014 : */
6015 112322958 : while (unlikely(extraWaits-- > 0))
6016 0 : PGSemaphoreUnlock(MyProc->sem);
6017 112322958 : }
6018 :
6019 : /*
6020 : * Release a previously acquired buffer content lock.
6021 : */
6022 : static void
6023 65013067 : BufferLockUnlock(Buffer buffer, BufferDesc *buf_hdr)
6024 : {
6025 : BufferLockMode mode;
6026 : uint64 oldstate;
6027 : uint64 sub;
6028 :
6029 65013067 : mode = BufferLockDisownInternal(buffer, buf_hdr);
6030 :
6031 : /*
6032 : * Release my hold on lock, after that it can immediately be acquired by
6033 : * others, even if we still have to wakeup other waiters.
6034 : */
6035 65013067 : sub = BufferLockReleaseSub(mode);
6036 :
6037 65013067 : oldstate = pg_atomic_sub_fetch_u64(&buf_hdr->state, sub);
6038 :
6039 65013067 : BufferLockProcessRelease(buf_hdr, mode, oldstate);
6040 :
6041 : /*
6042 : * Now okay to allow cancel/die interrupts.
6043 : */
6044 65013067 : RESUME_INTERRUPTS();
6045 65013067 : }
6046 :
6047 :
6048 : /*
6049 : * Acquire the content lock for the buffer, but only if we don't have to wait.
6050 : *
6051 : * It is allowed to try to conditionally acquire a lock on a buffer that this
6052 : * backend has already locked, but the lock acquisition will always fail, even
6053 : * if the new lock acquisition does not conflict with an already held lock
6054 : * (e.g. two share locks). This is because we currently do not have space to
6055 : * track multiple lock ownerships of the same buffer within one backend. That
6056 : * is ok for the current uses of BufferLockConditional().
6057 : */
6058 : static bool
6059 2146130 : BufferLockConditional(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode)
6060 : {
6061 2146130 : PrivateRefCountEntry *entry = GetPrivateRefCountEntry(buffer, true);
6062 : bool mustwait;
6063 :
6064 : /*
6065 : * As described above, if we're trying to lock a buffer this backend
6066 : * already has locked, return false, independent of the existing and
6067 : * desired lock level.
6068 : */
6069 2146130 : if (entry->data.lockmode != BUFFER_LOCK_UNLOCK)
6070 0 : return false;
6071 :
6072 : /*
6073 : * Lock out cancel/die interrupts until we exit the code section protected
6074 : * by the content lock. This ensures that interrupts will not interfere
6075 : * with manipulations of data structures in shared memory.
6076 : */
6077 2146130 : HOLD_INTERRUPTS();
6078 :
6079 : /* Check for the lock */
6080 2146130 : mustwait = BufferLockAttempt(buf_hdr, mode);
6081 :
6082 2146130 : if (mustwait)
6083 : {
6084 : /* Failed to get lock, so release interrupt holdoff */
6085 611 : RESUME_INTERRUPTS();
6086 : }
6087 : else
6088 : {
6089 2145519 : entry->data.lockmode = mode;
6090 : }
6091 :
6092 2146130 : return !mustwait;
6093 : }
6094 :
6095 : /*
6096 : * Internal function that tries to atomically acquire the content lock in the
6097 : * passed in mode.
6098 : *
6099 : * This function will not block waiting for a lock to become free - that's the
6100 : * caller's job.
6101 : *
6102 : * Similar to LWLockAttemptLock().
6103 : */
6104 : static inline bool
6105 114508264 : BufferLockAttempt(BufferDesc *buf_hdr, BufferLockMode mode)
6106 : {
6107 : uint64 old_state;
6108 :
6109 : /*
6110 : * Read once outside the loop, later iterations will get the newer value
6111 : * via compare & exchange.
6112 : */
6113 114508264 : old_state = pg_atomic_read_u64(&buf_hdr->state);
6114 :
6115 : /* loop until we've determined whether we could acquire the lock or not */
6116 : while (true)
6117 7845 : {
6118 : uint64 desired_state;
6119 : bool lock_free;
6120 :
6121 114516109 : desired_state = old_state;
6122 :
6123 114516109 : if (mode == BUFFER_LOCK_EXCLUSIVE)
6124 : {
6125 38613870 : lock_free = (old_state & BM_LOCK_MASK) == 0;
6126 38613870 : if (lock_free)
6127 38590975 : desired_state += BM_LOCK_VAL_EXCLUSIVE;
6128 : }
6129 75902239 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6130 : {
6131 729726 : lock_free = (old_state & (BM_LOCK_VAL_EXCLUSIVE | BM_LOCK_VAL_SHARE_EXCLUSIVE)) == 0;
6132 729726 : if (lock_free)
6133 729526 : desired_state += BM_LOCK_VAL_SHARE_EXCLUSIVE;
6134 : }
6135 : else
6136 : {
6137 75172513 : lock_free = (old_state & BM_LOCK_VAL_EXCLUSIVE) == 0;
6138 75172513 : if (lock_free)
6139 75155618 : desired_state += BM_LOCK_VAL_SHARED;
6140 : }
6141 :
6142 : /*
6143 : * Attempt to swap in the state we are expecting. If we didn't see
6144 : * lock to be free, that's just the old value. If we saw it as free,
6145 : * we'll attempt to mark it acquired. The reason that we always swap
6146 : * in the value is that this doubles as a memory barrier. We could try
6147 : * to be smarter and only swap in values if we saw the lock as free,
6148 : * but benchmark haven't shown it as beneficial so far.
6149 : *
6150 : * Retry if the value changed since we last looked at it.
6151 : */
6152 114516109 : if (likely(pg_atomic_compare_exchange_u64(&buf_hdr->state,
6153 : &old_state, desired_state)))
6154 : {
6155 114508264 : if (lock_free)
6156 : {
6157 : /* Great! Got the lock. */
6158 114468477 : return false;
6159 : }
6160 : else
6161 39787 : return true; /* somebody else has the lock */
6162 : }
6163 : }
6164 :
6165 : pg_unreachable();
6166 : }
6167 :
6168 : /*
6169 : * Add ourselves to the end of the content lock's wait queue.
6170 : */
6171 : static void
6172 19963 : BufferLockQueueSelf(BufferDesc *buf_hdr, BufferLockMode mode)
6173 : {
6174 : /*
6175 : * If we don't have a PGPROC structure, there's no way to wait. This
6176 : * should never occur, since MyProc should only be null during shared
6177 : * memory initialization.
6178 : */
6179 19963 : if (MyProc == NULL)
6180 0 : elog(PANIC, "cannot wait without a PGPROC structure");
6181 :
6182 19963 : if (MyProc->lwWaiting != LW_WS_NOT_WAITING)
6183 0 : elog(PANIC, "queueing for lock while waiting on another one");
6184 :
6185 19963 : LockBufHdr(buf_hdr);
6186 :
6187 : /* setting the flag is protected by the spinlock */
6188 19963 : pg_atomic_fetch_or_u64(&buf_hdr->state, BM_LOCK_HAS_WAITERS);
6189 :
6190 : /*
6191 : * These are currently used both for lwlocks and buffer content locks,
6192 : * which is acceptable, although not pretty, because a backend can't wait
6193 : * for both types of locks at the same time.
6194 : */
6195 19963 : MyProc->lwWaiting = LW_WS_WAITING;
6196 19963 : MyProc->lwWaitMode = mode;
6197 :
6198 19963 : proclist_push_tail(&buf_hdr->lock_waiters, MyProcNumber, lwWaitLink);
6199 :
6200 : /* Can release the mutex now */
6201 19963 : UnlockBufHdr(buf_hdr);
6202 19963 : }
6203 :
6204 : /*
6205 : * Remove ourselves from the waitlist.
6206 : *
6207 : * This is used if we queued ourselves because we thought we needed to sleep
6208 : * but, after further checking, we discovered that we don't actually need to
6209 : * do so.
6210 : */
6211 : static void
6212 750 : BufferLockDequeueSelf(BufferDesc *buf_hdr)
6213 : {
6214 : bool on_waitlist;
6215 :
6216 750 : LockBufHdr(buf_hdr);
6217 :
6218 750 : on_waitlist = MyProc->lwWaiting == LW_WS_WAITING;
6219 750 : if (on_waitlist)
6220 648 : proclist_delete(&buf_hdr->lock_waiters, MyProcNumber, lwWaitLink);
6221 :
6222 750 : if (proclist_is_empty(&buf_hdr->lock_waiters) &&
6223 729 : (pg_atomic_read_u64(&buf_hdr->state) & BM_LOCK_HAS_WAITERS) != 0)
6224 : {
6225 627 : pg_atomic_fetch_and_u64(&buf_hdr->state, ~BM_LOCK_HAS_WAITERS);
6226 : }
6227 :
6228 : /* XXX: combine with fetch_and above? */
6229 750 : UnlockBufHdr(buf_hdr);
6230 :
6231 : /* clear waiting state again, nice for debugging */
6232 750 : if (on_waitlist)
6233 648 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
6234 : else
6235 : {
6236 102 : int extraWaits = 0;
6237 :
6238 :
6239 : /*
6240 : * Somebody else dequeued us and has or will wake us up. Deal with the
6241 : * superfluous absorption of a wakeup.
6242 : */
6243 :
6244 : /*
6245 : * Clear BM_LOCK_WAKE_IN_PROGRESS if somebody woke us before we
6246 : * removed ourselves - they'll have set it.
6247 : */
6248 102 : pg_atomic_fetch_and_u64(&buf_hdr->state, ~BM_LOCK_WAKE_IN_PROGRESS);
6249 :
6250 : /*
6251 : * Now wait for the scheduled wakeup, otherwise our ->lwWaiting would
6252 : * get reset at some inconvenient point later. Most of the time this
6253 : * will immediately return.
6254 : */
6255 : for (;;)
6256 : {
6257 102 : PGSemaphoreLock(MyProc->sem);
6258 102 : if (MyProc->lwWaiting == LW_WS_NOT_WAITING)
6259 102 : break;
6260 0 : extraWaits++;
6261 : }
6262 :
6263 : /*
6264 : * Fix the process wait semaphore's count for any absorbed wakeups.
6265 : */
6266 102 : while (extraWaits-- > 0)
6267 0 : PGSemaphoreUnlock(MyProc->sem);
6268 : }
6269 750 : }
6270 :
6271 : /*
6272 : * Stop treating lock as held by current backend.
6273 : *
6274 : * After calling this function it's the callers responsibility to ensure that
6275 : * the lock gets released, even in case of an error. This only is desirable if
6276 : * the lock is going to be released in a different process than the process
6277 : * that acquired it.
6278 : */
6279 : static inline void
6280 0 : BufferLockDisown(Buffer buffer, BufferDesc *buf_hdr)
6281 : {
6282 0 : BufferLockDisownInternal(buffer, buf_hdr);
6283 0 : RESUME_INTERRUPTS();
6284 0 : }
6285 :
6286 : /*
6287 : * Stop treating lock as held by current backend.
6288 : *
6289 : * This is the code that can be shared between actually releasing a lock
6290 : * (BufferLockUnlock()) and just not tracking ownership of the lock anymore
6291 : * without releasing the lock (BufferLockDisown()).
6292 : */
6293 : static inline int
6294 114468477 : BufferLockDisownInternal(Buffer buffer, BufferDesc *buf_hdr)
6295 : {
6296 : BufferLockMode mode;
6297 : PrivateRefCountEntry *ref;
6298 :
6299 114468477 : ref = GetPrivateRefCountEntry(buffer, false);
6300 114468477 : if (ref == NULL)
6301 0 : elog(ERROR, "lock %d is not held", buffer);
6302 114468477 : mode = ref->data.lockmode;
6303 114468477 : ref->data.lockmode = BUFFER_LOCK_UNLOCK;
6304 :
6305 114468477 : return mode;
6306 : }
6307 :
6308 : /*
6309 : * Wakeup all the lockers that currently have a chance to acquire the lock.
6310 : *
6311 : * wake_exclusive indicates whether exclusive lock waiters should be woken up.
6312 : */
6313 : static void
6314 17980 : BufferLockWakeup(BufferDesc *buf_hdr, bool wake_exclusive)
6315 : {
6316 17980 : bool new_wake_in_progress = false;
6317 17980 : bool wake_share_exclusive = true;
6318 : proclist_head wakeup;
6319 : proclist_mutable_iter iter;
6320 :
6321 17980 : proclist_init(&wakeup);
6322 :
6323 : /* lock wait list while collecting backends to wake up */
6324 17980 : LockBufHdr(buf_hdr);
6325 :
6326 27520 : proclist_foreach_modify(iter, &buf_hdr->lock_waiters, lwWaitLink)
6327 : {
6328 20456 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
6329 :
6330 : /*
6331 : * Already woke up a conflicting lock, so skip over this wait list
6332 : * entry.
6333 : */
6334 20456 : if (!wake_exclusive && waiter->lwWaitMode == BUFFER_LOCK_EXCLUSIVE)
6335 1138 : continue;
6336 19318 : if (!wake_share_exclusive && waiter->lwWaitMode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6337 0 : continue;
6338 :
6339 19318 : proclist_delete(&buf_hdr->lock_waiters, iter.cur, lwWaitLink);
6340 19318 : proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
6341 :
6342 : /*
6343 : * Prevent additional wakeups until retryer gets to run. Backends that
6344 : * are just waiting for the lock to become free don't retry
6345 : * automatically.
6346 : */
6347 19318 : new_wake_in_progress = true;
6348 :
6349 : /*
6350 : * Signal that the process isn't on the wait list anymore. This allows
6351 : * BufferLockDequeueSelf() to remove itself from the waitlist with a
6352 : * proclist_delete(), rather than having to check if it has been
6353 : * removed from the list.
6354 : */
6355 : Assert(waiter->lwWaiting == LW_WS_WAITING);
6356 19318 : waiter->lwWaiting = LW_WS_PENDING_WAKEUP;
6357 :
6358 : /*
6359 : * Don't wakeup further waiters after waking a conflicting waiter.
6360 : */
6361 19318 : if (waiter->lwWaitMode == BUFFER_LOCK_SHARE)
6362 : {
6363 : /*
6364 : * Share locks conflict with exclusive locks.
6365 : */
6366 8296 : wake_exclusive = false;
6367 : }
6368 11022 : else if (waiter->lwWaitMode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6369 : {
6370 : /*
6371 : * Share-exclusive locks conflict with share-exclusive and
6372 : * exclusive locks.
6373 : */
6374 106 : wake_exclusive = false;
6375 106 : wake_share_exclusive = false;
6376 : }
6377 10916 : else if (waiter->lwWaitMode == BUFFER_LOCK_EXCLUSIVE)
6378 : {
6379 : /*
6380 : * Exclusive locks conflict with all other locks, there's no point
6381 : * in waking up anybody else.
6382 : */
6383 10916 : break;
6384 : }
6385 : }
6386 :
6387 : Assert(proclist_is_empty(&wakeup) || pg_atomic_read_u64(&buf_hdr->state) & BM_LOCK_HAS_WAITERS);
6388 :
6389 : /* unset required flags, and release lock, in one fell swoop */
6390 : {
6391 : uint64 old_state;
6392 : uint64 desired_state;
6393 :
6394 17980 : old_state = pg_atomic_read_u64(&buf_hdr->state);
6395 : while (true)
6396 : {
6397 17987 : desired_state = old_state;
6398 :
6399 : /* compute desired flags */
6400 :
6401 17987 : if (new_wake_in_progress)
6402 17785 : desired_state |= BM_LOCK_WAKE_IN_PROGRESS;
6403 : else
6404 202 : desired_state &= ~BM_LOCK_WAKE_IN_PROGRESS;
6405 :
6406 17987 : if (proclist_is_empty(&buf_hdr->lock_waiters))
6407 14491 : desired_state &= ~BM_LOCK_HAS_WAITERS;
6408 :
6409 17987 : desired_state &= ~BM_LOCKED; /* release lock */
6410 :
6411 17987 : if (pg_atomic_compare_exchange_u64(&buf_hdr->state, &old_state,
6412 : desired_state))
6413 17980 : break;
6414 : }
6415 : }
6416 :
6417 : /* Awaken any waiters I removed from the queue. */
6418 37298 : proclist_foreach_modify(iter, &wakeup, lwWaitLink)
6419 : {
6420 19318 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
6421 :
6422 19318 : proclist_delete(&wakeup, iter.cur, lwWaitLink);
6423 :
6424 : /*
6425 : * Guarantee that lwWaiting being unset only becomes visible once the
6426 : * unlink from the link has completed. Otherwise the target backend
6427 : * could be woken up for other reason and enqueue for a new lock - if
6428 : * that happens before the list unlink happens, the list would end up
6429 : * being corrupted.
6430 : *
6431 : * The barrier pairs with the LockBufHdr() when enqueuing for another
6432 : * lock.
6433 : */
6434 19318 : pg_write_barrier();
6435 19318 : waiter->lwWaiting = LW_WS_NOT_WAITING;
6436 19318 : PGSemaphoreUnlock(waiter->sem);
6437 : }
6438 17980 : }
6439 :
6440 : /*
6441 : * Compute subtraction from buffer state for a release of a held lock in
6442 : * `mode`.
6443 : *
6444 : * This is separated from BufferLockUnlock() as we want to combine the lock
6445 : * release with other atomic operations when possible, leading to the lock
6446 : * release being done in multiple places, each needing to compute what to
6447 : * subtract from the lock state.
6448 : */
6449 : static inline uint64
6450 114468477 : BufferLockReleaseSub(BufferLockMode mode)
6451 : {
6452 : /*
6453 : * Turns out that a switch() leads gcc to generate sufficiently worse code
6454 : * for this to show up in profiles...
6455 : */
6456 114468477 : if (mode == BUFFER_LOCK_EXCLUSIVE)
6457 38590814 : return BM_LOCK_VAL_EXCLUSIVE;
6458 75877663 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6459 4139945 : return BM_LOCK_VAL_SHARE_EXCLUSIVE;
6460 : else
6461 : {
6462 : Assert(mode == BUFFER_LOCK_SHARE);
6463 71737718 : return BM_LOCK_VAL_SHARED;
6464 : }
6465 :
6466 : return 0; /* keep compiler quiet */
6467 : }
6468 :
6469 : /*
6470 : * Handle work that needs to be done after releasing a lock that was held in
6471 : * `mode`, where `lockstate` is the result of the atomic operation modifying
6472 : * the state variable.
6473 : *
6474 : * This is separated from BufferLockUnlock() as we want to combine the lock
6475 : * release with other atomic operations when possible, leading to the lock
6476 : * release being done in multiple places.
6477 : */
6478 : static void
6479 114468477 : BufferLockProcessRelease(BufferDesc *buf_hdr, BufferLockMode mode, uint64 lockstate)
6480 : {
6481 114468477 : bool check_waiters = false;
6482 114468477 : bool wake_exclusive = false;
6483 :
6484 : /* nobody else can have that kind of lock */
6485 : Assert(!(lockstate & BM_LOCK_VAL_EXCLUSIVE));
6486 :
6487 : /*
6488 : * If we're still waiting for backends to get scheduled, don't wake them
6489 : * up again. Otherwise check if we need to look through the waitqueue to
6490 : * wake other backends.
6491 : */
6492 114468477 : if ((lockstate & BM_LOCK_HAS_WAITERS) &&
6493 78323 : !(lockstate & BM_LOCK_WAKE_IN_PROGRESS))
6494 : {
6495 36208 : if ((lockstate & BM_LOCK_MASK) == 0)
6496 : {
6497 : /*
6498 : * We released a lock and the lock was, in that moment, free. We
6499 : * therefore can wake waiters for any kind of lock.
6500 : */
6501 17979 : check_waiters = true;
6502 17979 : wake_exclusive = true;
6503 : }
6504 18229 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6505 : {
6506 : /*
6507 : * We released the lock, but another backend still holds a lock.
6508 : * We can't have released an exclusive lock, as there couldn't
6509 : * have been other lock holders. If we released a share lock, no
6510 : * waiters need to be woken up, as there must be other share
6511 : * lockers. However, if we held a share-exclusive lock, another
6512 : * backend now could acquire a share-exclusive lock.
6513 : */
6514 1 : check_waiters = true;
6515 1 : wake_exclusive = false;
6516 : }
6517 : }
6518 :
6519 : /*
6520 : * As waking up waiters requires the spinlock to be acquired, only do so
6521 : * if necessary.
6522 : */
6523 114468477 : if (check_waiters)
6524 17980 : BufferLockWakeup(buf_hdr, wake_exclusive);
6525 114468477 : }
6526 :
6527 : /*
6528 : * BufferLockHeldByMeInMode - test whether my process holds the content lock
6529 : * in the specified mode
6530 : *
6531 : * This is meant as debug support only.
6532 : */
6533 : static bool
6534 0 : BufferLockHeldByMeInMode(BufferDesc *buf_hdr, BufferLockMode mode)
6535 : {
6536 : PrivateRefCountEntry *entry =
6537 0 : GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf_hdr), false);
6538 :
6539 0 : if (!entry)
6540 0 : return false;
6541 : else
6542 0 : return entry->data.lockmode == mode;
6543 : }
6544 :
6545 : /*
6546 : * BufferLockHeldByMe - test whether my process holds the content lock in any
6547 : * mode
6548 : *
6549 : * This is meant as debug support only.
6550 : */
6551 : static bool
6552 0 : BufferLockHeldByMe(BufferDesc *buf_hdr)
6553 : {
6554 : PrivateRefCountEntry *entry =
6555 0 : GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf_hdr), false);
6556 :
6557 0 : if (!entry)
6558 0 : return false;
6559 : else
6560 0 : return entry->data.lockmode != BUFFER_LOCK_UNLOCK;
6561 : }
6562 :
6563 : /*
6564 : * Release the content lock for the buffer.
6565 : */
6566 : void
6567 69859846 : UnlockBuffer(Buffer buffer)
6568 : {
6569 : BufferDesc *buf_hdr;
6570 :
6571 : Assert(BufferIsPinned(buffer));
6572 69859846 : if (BufferIsLocal(buffer))
6573 5200597 : return; /* local buffers need no lock */
6574 :
6575 64659249 : buf_hdr = GetBufferDescriptor(buffer - 1);
6576 64659249 : BufferLockUnlock(buffer, buf_hdr);
6577 : }
6578 :
6579 : /*
6580 : * Acquire the content_lock for the buffer.
6581 : */
6582 : void
6583 118440009 : LockBufferInternal(Buffer buffer, BufferLockMode mode)
6584 : {
6585 : BufferDesc *buf_hdr;
6586 :
6587 : /*
6588 : * We can't wait if we haven't got a PGPROC. This should only occur
6589 : * during bootstrap or shared memory initialization. Put an Assert here
6590 : * to catch unsafe coding practices.
6591 : */
6592 : Assert(!(MyProc == NULL && IsUnderPostmaster));
6593 :
6594 : /* handled in LockBuffer() wrapper */
6595 : Assert(mode != BUFFER_LOCK_UNLOCK);
6596 :
6597 : Assert(BufferIsPinned(buffer));
6598 118440009 : if (BufferIsLocal(buffer))
6599 6470757 : return; /* local buffers need no lock */
6600 :
6601 111969252 : buf_hdr = GetBufferDescriptor(buffer - 1);
6602 :
6603 : /*
6604 : * Test the most frequent lock modes first. While a switch (mode) would be
6605 : * nice, at least gcc generates considerably worse code for it.
6606 : *
6607 : * Call BufferLockAcquire() with a constant argument for mode, to generate
6608 : * more efficient code for the different lock modes.
6609 : */
6610 111969252 : if (mode == BUFFER_LOCK_SHARE)
6611 75148138 : BufferLockAcquire(buffer, buf_hdr, BUFFER_LOCK_SHARE);
6612 36821114 : else if (mode == BUFFER_LOCK_EXCLUSIVE)
6613 36821114 : BufferLockAcquire(buffer, buf_hdr, BUFFER_LOCK_EXCLUSIVE);
6614 0 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6615 0 : BufferLockAcquire(buffer, buf_hdr, BUFFER_LOCK_SHARE_EXCLUSIVE);
6616 : else
6617 0 : elog(ERROR, "unrecognized buffer lock mode: %d", mode);
6618 : }
6619 :
6620 : /*
6621 : * Acquire the content_lock for the buffer, but only if we don't have to wait.
6622 : *
6623 : * This assumes the caller wants BUFFER_LOCK_EXCLUSIVE mode.
6624 : */
6625 : bool
6626 1856455 : ConditionalLockBuffer(Buffer buffer)
6627 : {
6628 : BufferDesc *buf;
6629 :
6630 : Assert(BufferIsPinned(buffer));
6631 1856455 : if (BufferIsLocal(buffer))
6632 86160 : return true; /* act as though we got it */
6633 :
6634 1770295 : buf = GetBufferDescriptor(buffer - 1);
6635 :
6636 1770295 : return BufferLockConditional(buffer, buf, BUFFER_LOCK_EXCLUSIVE);
6637 : }
6638 :
6639 : /*
6640 : * Verify that this backend is pinning the buffer exactly once.
6641 : *
6642 : * NOTE: Like in BufferIsPinned(), what we check here is that *this* backend
6643 : * holds a pin on the buffer. We do not care whether some other backend does.
6644 : */
6645 : void
6646 2795068 : CheckBufferIsPinnedOnce(Buffer buffer)
6647 : {
6648 2795068 : if (BufferIsLocal(buffer))
6649 : {
6650 1049 : if (LocalRefCount[-buffer - 1] != 1)
6651 0 : elog(ERROR, "incorrect local pin count: %d",
6652 : LocalRefCount[-buffer - 1]);
6653 : }
6654 : else
6655 : {
6656 2794019 : if (GetPrivateRefCount(buffer) != 1)
6657 0 : elog(ERROR, "incorrect local pin count: %d",
6658 : GetPrivateRefCount(buffer));
6659 : }
6660 2795068 : }
6661 :
6662 : /*
6663 : * LockBufferForCleanup - lock a buffer in preparation for deleting items
6664 : *
6665 : * Items may be deleted from a disk page only when the caller (a) holds an
6666 : * exclusive lock on the buffer and (b) has observed that no other backend
6667 : * holds a pin on the buffer. If there is a pin, then the other backend
6668 : * might have a pointer into the buffer (for example, a heapscan reference
6669 : * to an item --- see README for more details). It's OK if a pin is added
6670 : * after the cleanup starts, however; the newly-arrived backend will be
6671 : * unable to look at the page until we release the exclusive lock.
6672 : *
6673 : * To implement this protocol, a would-be deleter must pin the buffer and
6674 : * then call LockBufferForCleanup(). LockBufferForCleanup() is similar to
6675 : * LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE), except that it loops until
6676 : * it has successfully observed pin count = 1.
6677 : */
6678 : void
6679 28980 : LockBufferForCleanup(Buffer buffer)
6680 : {
6681 : BufferDesc *bufHdr;
6682 28980 : TimestampTz waitStart = 0;
6683 28980 : bool waiting = false;
6684 28980 : bool logged_recovery_conflict = false;
6685 :
6686 : Assert(BufferIsPinned(buffer));
6687 : Assert(PinCountWaitBuf == NULL);
6688 :
6689 28980 : CheckBufferIsPinnedOnce(buffer);
6690 :
6691 : /*
6692 : * We do not yet need to be worried about in-progress AIOs holding a pin,
6693 : * as we, so far, only support doing reads via AIO and this function can
6694 : * only be called once the buffer is valid (i.e. no read can be in
6695 : * flight).
6696 : */
6697 :
6698 : /* Nobody else to wait for */
6699 28980 : if (BufferIsLocal(buffer))
6700 18 : return;
6701 :
6702 28962 : bufHdr = GetBufferDescriptor(buffer - 1);
6703 :
6704 : for (;;)
6705 101 : {
6706 : uint64 buf_state;
6707 29063 : uint64 unset_bits = 0;
6708 :
6709 : /* Try to acquire lock */
6710 29063 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6711 29063 : buf_state = LockBufHdr(bufHdr);
6712 :
6713 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6714 29063 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
6715 : {
6716 : /* Successfully acquired exclusive lock with pincount 1 */
6717 28962 : UnlockBufHdr(bufHdr);
6718 28962 : goto cleanup_lock_acquired;
6719 : }
6720 : /* Failed, so mark myself as waiting for pincount 1 */
6721 101 : if (buf_state & BM_PIN_COUNT_WAITER)
6722 : {
6723 0 : UnlockBufHdr(bufHdr);
6724 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6725 0 : elog(ERROR, "multiple backends attempting to wait for pincount 1");
6726 : }
6727 101 : bufHdr->wait_backend_pgprocno = MyProcNumber;
6728 101 : PinCountWaitBuf = bufHdr;
6729 :
6730 : /*
6731 : * Publish BM_PIN_COUNT_WAITER while retaining the buffer header lock.
6732 : * The shared refcount can be decremented while BM_LOCKED is set, so
6733 : * use an atomic operation that preserves concurrent refcount changes.
6734 : */
6735 101 : pg_atomic_fetch_or_u64(&bufHdr->state, BM_PIN_COUNT_WAITER);
6736 :
6737 : /*
6738 : * Recheck the refcount after publishing the waiter flag, while shared
6739 : * refcount increments are still prevented by BM_LOCKED. If only our
6740 : * pin remains, the cleanup-lock condition has already been satisfied,
6741 : * so remove the waiter state and return without sleeping.
6742 : */
6743 101 : buf_state = pg_atomic_read_u64(&bufHdr->state);
6744 :
6745 101 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
6746 : {
6747 0 : UnlockBufHdrExt(bufHdr, buf_state,
6748 : 0, BM_PIN_COUNT_WAITER,
6749 : 0);
6750 0 : PinCountWaitBuf = NULL;
6751 0 : goto cleanup_lock_acquired;
6752 : }
6753 :
6754 101 : UnlockBufHdr(bufHdr);
6755 101 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6756 :
6757 : /* Wait to be signaled by UnpinBuffer() */
6758 101 : if (InHotStandby)
6759 : {
6760 9 : if (!waiting)
6761 : {
6762 : /* adjust the process title to indicate that it's waiting */
6763 2 : set_ps_display_suffix("waiting");
6764 2 : waiting = true;
6765 : }
6766 :
6767 : /*
6768 : * Emit the log message if the startup process is waiting longer
6769 : * than deadlock_timeout for recovery conflict on buffer pin.
6770 : *
6771 : * Skip this if first time through because the startup process has
6772 : * not started waiting yet in this case. So, the wait start
6773 : * timestamp is set after this logic.
6774 : */
6775 9 : if (waitStart != 0 && !logged_recovery_conflict)
6776 : {
6777 3 : TimestampTz now = GetCurrentTimestamp();
6778 :
6779 3 : if (TimestampDifferenceExceeds(waitStart, now,
6780 : DeadlockTimeout))
6781 : {
6782 2 : LogRecoveryConflict(RECOVERY_CONFLICT_BUFFERPIN,
6783 : waitStart, now, NULL, true);
6784 2 : logged_recovery_conflict = true;
6785 : }
6786 : }
6787 :
6788 : /*
6789 : * Set the wait start timestamp if logging is enabled and first
6790 : * time through.
6791 : */
6792 9 : if (log_recovery_conflict_waits && waitStart == 0)
6793 2 : waitStart = GetCurrentTimestamp();
6794 :
6795 : /* Publish the bufid that Startup process waits on */
6796 9 : SetStartupBufferPinWaitBufId(buffer - 1);
6797 : /* Set alarm and then wait to be signaled by UnpinBuffer() */
6798 9 : ResolveRecoveryConflictWithBufferPin();
6799 : /* Reset the published bufid */
6800 9 : SetStartupBufferPinWaitBufId(-1);
6801 : }
6802 : else
6803 92 : ProcWaitForSignal(WAIT_EVENT_BUFFER_CLEANUP);
6804 :
6805 : /*
6806 : * Remove flag marking us as waiter. Normally this will not be set
6807 : * anymore, but ProcWaitForSignal() can return for other signals as
6808 : * well. We take care to only reset the flag if we're the waiter, as
6809 : * theoretically another backend could have started waiting. That's
6810 : * impossible with the current usages due to table level locking, but
6811 : * better be safe.
6812 : */
6813 101 : buf_state = LockBufHdr(bufHdr);
6814 101 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
6815 7 : bufHdr->wait_backend_pgprocno == MyProcNumber)
6816 7 : unset_bits |= BM_PIN_COUNT_WAITER;
6817 :
6818 101 : UnlockBufHdrExt(bufHdr, buf_state,
6819 : 0, unset_bits,
6820 : 0);
6821 :
6822 101 : PinCountWaitBuf = NULL;
6823 : /* Loop back and try again */
6824 : }
6825 :
6826 28962 : cleanup_lock_acquired:
6827 :
6828 : /*
6829 : * Emit the log message if recovery conflict on buffer pin was resolved
6830 : * but the startup process waited longer than deadlock_timeout for it.
6831 : */
6832 28962 : if (logged_recovery_conflict)
6833 2 : LogRecoveryConflict(RECOVERY_CONFLICT_BUFFERPIN,
6834 : waitStart, GetCurrentTimestamp(),
6835 : NULL, false);
6836 :
6837 28962 : if (waiting)
6838 : {
6839 : /* reset ps display to remove the suffix if we added one */
6840 2 : set_ps_display_remove_suffix();
6841 2 : waiting = false;
6842 : }
6843 :
6844 28962 : return;
6845 : }
6846 :
6847 : /*
6848 : * Check called from ProcessRecoveryConflictInterrupts() when Startup process
6849 : * requests cancellation of all pin holders that are blocking it.
6850 : */
6851 : bool
6852 3 : HoldingBufferPinThatDelaysRecovery(void)
6853 : {
6854 3 : int bufid = GetStartupBufferPinWaitBufId();
6855 :
6856 : /*
6857 : * If we get woken slowly then it's possible that the Startup process was
6858 : * already woken by other backends before we got here. Also possible that
6859 : * we get here by multiple interrupts or interrupts at inappropriate
6860 : * times, so make sure we do nothing if the bufid is not set.
6861 : */
6862 3 : if (bufid < 0)
6863 1 : return false;
6864 :
6865 2 : if (GetPrivateRefCount(bufid + 1) > 0)
6866 2 : return true;
6867 :
6868 0 : return false;
6869 : }
6870 :
6871 : /*
6872 : * ConditionalLockBufferForCleanup - as above, but don't wait to get the lock
6873 : *
6874 : * We won't loop, but just check once to see if the pin count is OK. If
6875 : * not, return false with no lock held.
6876 : */
6877 : bool
6878 609398 : ConditionalLockBufferForCleanup(Buffer buffer)
6879 : {
6880 : BufferDesc *bufHdr;
6881 : uint64 buf_state,
6882 : refcount;
6883 :
6884 : Assert(BufferIsValid(buffer));
6885 :
6886 : /* see AIO related comment in LockBufferForCleanup() */
6887 :
6888 609398 : if (BufferIsLocal(buffer))
6889 : {
6890 12373 : refcount = LocalRefCount[-buffer - 1];
6891 : /* There should be exactly one pin */
6892 : Assert(refcount > 0);
6893 12373 : if (refcount != 1)
6894 1540 : return false;
6895 : /* Nobody else to wait for */
6896 10833 : return true;
6897 : }
6898 :
6899 : /* There should be exactly one local pin */
6900 597025 : refcount = GetPrivateRefCount(buffer);
6901 : Assert(refcount);
6902 597025 : if (refcount != 1)
6903 395 : return false;
6904 :
6905 : /* Try to acquire lock */
6906 596630 : if (!ConditionalLockBuffer(buffer))
6907 52 : return false;
6908 :
6909 596578 : bufHdr = GetBufferDescriptor(buffer - 1);
6910 596578 : buf_state = LockBufHdr(bufHdr);
6911 596578 : refcount = BUF_STATE_GET_REFCOUNT(buf_state);
6912 :
6913 : Assert(refcount > 0);
6914 596578 : if (refcount == 1)
6915 : {
6916 : /* Successfully acquired exclusive lock with pincount 1 */
6917 596127 : UnlockBufHdr(bufHdr);
6918 596127 : return true;
6919 : }
6920 :
6921 : /* Failed, so release the lock */
6922 451 : UnlockBufHdr(bufHdr);
6923 451 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6924 451 : return false;
6925 : }
6926 :
6927 : /*
6928 : * IsBufferCleanupOK - as above, but we already have the lock
6929 : *
6930 : * Check whether it's OK to perform cleanup on a buffer we've already
6931 : * locked. If we observe that the pin count is 1, our exclusive lock
6932 : * happens to be a cleanup lock, and we can proceed with anything that
6933 : * would have been allowable had we sought a cleanup lock originally.
6934 : */
6935 : bool
6936 2705 : IsBufferCleanupOK(Buffer buffer)
6937 : {
6938 : BufferDesc *bufHdr;
6939 : uint64 buf_state;
6940 :
6941 : Assert(BufferIsValid(buffer));
6942 :
6943 : /* see AIO related comment in LockBufferForCleanup() */
6944 :
6945 2705 : if (BufferIsLocal(buffer))
6946 : {
6947 : /* There should be exactly one pin */
6948 0 : if (LocalRefCount[-buffer - 1] != 1)
6949 0 : return false;
6950 : /* Nobody else to wait for */
6951 0 : return true;
6952 : }
6953 :
6954 : /* There should be exactly one local pin */
6955 2705 : if (GetPrivateRefCount(buffer) != 1)
6956 0 : return false;
6957 :
6958 2705 : bufHdr = GetBufferDescriptor(buffer - 1);
6959 :
6960 : /* caller must hold exclusive lock on buffer */
6961 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
6962 :
6963 2705 : buf_state = LockBufHdr(bufHdr);
6964 :
6965 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6966 2705 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
6967 : {
6968 : /* pincount is OK. */
6969 2705 : UnlockBufHdr(bufHdr);
6970 2705 : return true;
6971 : }
6972 :
6973 0 : UnlockBufHdr(bufHdr);
6974 0 : return false;
6975 : }
6976 :
6977 : /*
6978 : * Helper for BufferBeginSetHintBits() and BufferSetHintBits16().
6979 : *
6980 : * This checks if the current lock mode already suffices to allow hint bits
6981 : * being set and, if not, whether the current lock can be upgraded.
6982 : *
6983 : * Updates *lockstate when returning true.
6984 : */
6985 : static inline bool
6986 15118788 : SharedBufferBeginSetHintBits(Buffer buffer, BufferDesc *buf_hdr, uint64 *lockstate)
6987 : {
6988 : uint64 old_state;
6989 : PrivateRefCountEntry *ref;
6990 : BufferLockMode mode;
6991 :
6992 15118788 : ref = GetPrivateRefCountEntry(buffer, true);
6993 :
6994 15118788 : if (ref == NULL)
6995 0 : elog(ERROR, "buffer is not pinned");
6996 :
6997 15118788 : mode = ref->data.lockmode;
6998 15118788 : if (mode == BUFFER_LOCK_UNLOCK)
6999 0 : elog(ERROR, "buffer is not locked");
7000 :
7001 : /* we're done if we are already holding a sufficient lock level */
7002 15118788 : if (mode == BUFFER_LOCK_EXCLUSIVE || mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
7003 : {
7004 11708238 : *lockstate = pg_atomic_read_u64(&buf_hdr->state);
7005 11708238 : return true;
7006 : }
7007 :
7008 : /*
7009 : * We are only holding a share lock right now, try to upgrade it to
7010 : * SHARE_EXCLUSIVE.
7011 : */
7012 : Assert(mode == BUFFER_LOCK_SHARE);
7013 :
7014 3410550 : old_state = pg_atomic_read_u64(&buf_hdr->state);
7015 : while (true)
7016 5 : {
7017 : uint64 desired_state;
7018 :
7019 3410555 : desired_state = old_state;
7020 :
7021 : /*
7022 : * Can't upgrade if somebody else holds the lock in exclusive or
7023 : * share-exclusive mode.
7024 : */
7025 3410555 : if (unlikely((old_state & (BM_LOCK_VAL_EXCLUSIVE | BM_LOCK_VAL_SHARE_EXCLUSIVE)) != 0))
7026 : {
7027 130 : return false;
7028 : }
7029 :
7030 : /* currently held lock state */
7031 3410425 : desired_state -= BM_LOCK_VAL_SHARED;
7032 :
7033 : /* new lock level */
7034 3410425 : desired_state += BM_LOCK_VAL_SHARE_EXCLUSIVE;
7035 :
7036 3410425 : if (likely(pg_atomic_compare_exchange_u64(&buf_hdr->state,
7037 : &old_state, desired_state)))
7038 : {
7039 3410420 : ref->data.lockmode = BUFFER_LOCK_SHARE_EXCLUSIVE;
7040 3410420 : *lockstate = desired_state;
7041 :
7042 3410420 : return true;
7043 : }
7044 : }
7045 : }
7046 :
7047 : /*
7048 : * Try to acquire the right to set hint bits on the buffer.
7049 : *
7050 : * To be allowed to set hint bits, this backend needs to hold either a
7051 : * share-exclusive or an exclusive lock. In case this backend only holds a
7052 : * share lock, this function will try to upgrade the lock to
7053 : * share-exclusive. The caller is only allowed to set hint bits if true is
7054 : * returned.
7055 : *
7056 : * Once BufferBeginSetHintBits() has returned true, hint bits may be set
7057 : * without further calls to BufferBeginSetHintBits(), until the buffer is
7058 : * unlocked.
7059 : *
7060 : *
7061 : * Requiring a share-exclusive lock to set hint bits prevents setting hint
7062 : * bits on buffers that are currently being written out, which could corrupt
7063 : * the checksum on the page. Flushing buffers also requires a share-exclusive
7064 : * lock.
7065 : *
7066 : * Due to a lock >= share-exclusive being required to set hint bits, only one
7067 : * backend can set hint bits at a time. Allowing multiple backends to set hint
7068 : * bits would require more complicated locking: For setting hint bits we'd
7069 : * need to store the count of backends currently setting hint bits, for I/O we
7070 : * would need another lock-level conflicting with the hint-setting
7071 : * lock-level. Given that the share-exclusive lock for setting hint bits is
7072 : * only held for a short time, that backends often would just set the same
7073 : * hint bits and that the cost of occasionally not setting hint bits in hotly
7074 : * accessed pages is fairly low, this seems like an acceptable tradeoff.
7075 : */
7076 : bool
7077 425126 : BufferBeginSetHintBits(Buffer buffer)
7078 : {
7079 : BufferDesc *buf_hdr;
7080 : uint64 lockstate;
7081 :
7082 425126 : if (BufferIsLocal(buffer))
7083 : {
7084 : /*
7085 : * NB: Will need to check if there is a write in progress, once it is
7086 : * possible for writes to be done asynchronously.
7087 : */
7088 2400 : return true;
7089 : }
7090 :
7091 422726 : buf_hdr = GetBufferDescriptor(buffer - 1);
7092 :
7093 422726 : return SharedBufferBeginSetHintBits(buffer, buf_hdr, &lockstate);
7094 : }
7095 :
7096 : /*
7097 : * End a phase of setting hint bits on this buffer, started with
7098 : * BufferBeginSetHintBits().
7099 : *
7100 : * This would strictly speaking not be required (i.e. the caller could do
7101 : * MarkBufferDirtyHint() if so desired), but allows us to perform some sanity
7102 : * checks.
7103 : */
7104 : void
7105 425105 : BufferFinishSetHintBits(Buffer buffer, bool mark_dirty, bool buffer_std)
7106 : {
7107 : if (!BufferIsLocal(buffer))
7108 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_SHARE_EXCLUSIVE) ||
7109 : BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
7110 :
7111 425105 : if (mark_dirty)
7112 234172 : MarkBufferDirtyHint(buffer, buffer_std);
7113 425105 : }
7114 :
7115 : /*
7116 : * Try to set hint bits on a single 16bit value in a buffer.
7117 : *
7118 : * If hint bits are allowed to be set, set *ptr = val, try to mark the buffer
7119 : * dirty and return true. Otherwise false is returned.
7120 : *
7121 : * *ptr needs to be a pointer to memory within the buffer.
7122 : *
7123 : * This is a bit faster than BufferBeginSetHintBits() /
7124 : * BufferFinishSetHintBits() when setting hints once in a buffer, but slower
7125 : * than the former when setting hint bits multiple times in the same buffer.
7126 : */
7127 : bool
7128 15499723 : BufferSetHintBits16(uint16 *ptr, uint16 val, Buffer buffer)
7129 : {
7130 : BufferDesc *buf_hdr;
7131 : uint64 lockstate;
7132 : #ifdef USE_ASSERT_CHECKING
7133 : char *page;
7134 :
7135 : /* verify that the address is on the page */
7136 : page = BufferGetPage(buffer);
7137 : Assert((char *) ptr >= page && (char *) ptr < (page + BLCKSZ));
7138 : #endif
7139 :
7140 15499723 : if (BufferIsLocal(buffer))
7141 : {
7142 803661 : *ptr = val;
7143 :
7144 803661 : MarkLocalBufferDirty(buffer);
7145 :
7146 803661 : return true;
7147 : }
7148 :
7149 14696062 : buf_hdr = GetBufferDescriptor(buffer - 1);
7150 :
7151 14696062 : if (SharedBufferBeginSetHintBits(buffer, buf_hdr, &lockstate))
7152 : {
7153 14695953 : *ptr = val;
7154 :
7155 14695953 : MarkSharedBufferDirtyHint(buffer, buf_hdr, lockstate, true);
7156 :
7157 14695953 : return true;
7158 : }
7159 :
7160 109 : return false;
7161 : }
7162 :
7163 :
7164 : /*
7165 : * Functions for buffer I/O handling
7166 : *
7167 : * Also note that these are used only for shared buffers, not local ones.
7168 : */
7169 :
7170 : /*
7171 : * WaitIO -- Block until the IO_IN_PROGRESS flag on 'buf' is cleared.
7172 : */
7173 : static void
7174 196 : WaitIO(BufferDesc *buf)
7175 : {
7176 196 : ConditionVariable *cv = BufferDescriptorGetIOCV(buf);
7177 :
7178 : /*
7179 : * Should never end up here with unsubmitted IO, as no AIO unaware code
7180 : * may be used while in batch mode and AIO aware code needs to have
7181 : * submitted all staged IO to avoid deadlocks & slowness.
7182 : */
7183 : Assert(!pgaio_have_staged());
7184 :
7185 196 : ConditionVariablePrepareToSleep(cv);
7186 : for (;;)
7187 196 : {
7188 : uint64 buf_state;
7189 : PgAioWaitRef iow;
7190 :
7191 : /*
7192 : * It may not be necessary to acquire the spinlock to check the flag
7193 : * here, but since this test is essential for correctness, we'd better
7194 : * play it safe.
7195 : */
7196 392 : buf_state = LockBufHdr(buf);
7197 :
7198 : /*
7199 : * Copy the wait reference while holding the spinlock. This protects
7200 : * against a concurrent TerminateBufferIO() in another backend from
7201 : * clearing the wref while it's being read.
7202 : */
7203 392 : iow = buf->io_wref;
7204 392 : UnlockBufHdr(buf);
7205 :
7206 : /* no IO in progress, we don't need to wait */
7207 392 : if (!(buf_state & BM_IO_IN_PROGRESS))
7208 196 : break;
7209 :
7210 : /*
7211 : * The buffer has asynchronous IO in progress, wait for it to
7212 : * complete.
7213 : */
7214 196 : if (pgaio_wref_valid(&iow))
7215 : {
7216 31 : pgaio_wref_wait(&iow);
7217 :
7218 : /*
7219 : * The AIO subsystem internally uses condition variables and thus
7220 : * might remove this backend from the BufferDesc's CV. While that
7221 : * wouldn't cause a correctness issue (the first CV sleep just
7222 : * immediately returns if not already registered), it seems worth
7223 : * avoiding unnecessary loop iterations, given that we take care
7224 : * to do so at the start of the function.
7225 : */
7226 31 : ConditionVariablePrepareToSleep(cv);
7227 31 : continue;
7228 : }
7229 :
7230 : /* wait on BufferDesc->cv, e.g. for concurrent synchronous IO */
7231 165 : ConditionVariableSleep(cv, WAIT_EVENT_BUFFER_IO);
7232 : }
7233 196 : ConditionVariableCancelSleep();
7234 196 : }
7235 :
7236 : /*
7237 : * StartSharedBufferIO: begin I/O on this buffer
7238 : * (Assumptions)
7239 : * The buffer is Pinned
7240 : *
7241 : * In several scenarios the buffer may already be undergoing I/O in this or
7242 : * another backend. How to best handle that depends on the caller's
7243 : * situation. It might be appropriate to wait synchronously (e.g., because the
7244 : * buffer is about to be invalidated); wait asynchronously, using the buffer's
7245 : * IO wait reference (e.g., because the caller is doing readahead and doesn't
7246 : * need the buffer to be ready immediately); or to not wait at all (e.g.,
7247 : * because the caller is trying to combine IO for this buffer with another
7248 : * buffer).
7249 : *
7250 : * How and whether to wait is controlled by the wait and io_wref
7251 : * parameters. In detail:
7252 : *
7253 : * - If the caller passes a non-NULL io_wref and the buffer has an I/O wait
7254 : * reference, the *io_wref is set to the buffer's io_wref and
7255 : * BUFFER_IO_IN_PROGRESS is returned. This is done regardless of the wait
7256 : * parameter.
7257 : *
7258 : * - If the caller passes a NULL io_wref (i.e. the caller does not want to
7259 : * asynchronously wait for the completion of the IO), wait = false and the
7260 : * buffer is undergoing IO, BUFFER_IO_IN_PROGRESS is returned.
7261 : *
7262 : * - If wait = true and either the buffer does not have a wait reference,
7263 : * or the caller passes io_wref = NULL, WaitIO() is used to wait for the IO
7264 : * to complete. To avoid the potential of deadlocks and unnecessary delays,
7265 : * all staged I/O is submitted before waiting.
7266 : *
7267 : * Input operations are only attempted on buffers that are not BM_VALID, and
7268 : * output operations only on buffers that are BM_VALID and BM_DIRTY, so we can
7269 : * always tell if the work is already done. If no I/O is necessary,
7270 : * BUFFER_IO_ALREADY_DONE is returned.
7271 : *
7272 : * If we successfully marked the buffer as BM_IO_IN_PROGRESS,
7273 : * BUFFER_IO_READY_FOR_IO is returned.
7274 : */
7275 : StartBufferIOResult
7276 2977874 : StartSharedBufferIO(BufferDesc *buf, bool forInput, bool wait, PgAioWaitRef *io_wref)
7277 : {
7278 : uint64 buf_state;
7279 :
7280 2977874 : ResourceOwnerEnlarge(CurrentResourceOwner);
7281 :
7282 : for (;;)
7283 : {
7284 2978069 : buf_state = LockBufHdr(buf);
7285 :
7286 2978069 : if (!(buf_state & BM_IO_IN_PROGRESS))
7287 2975734 : break;
7288 :
7289 : /* Join the existing IO */
7290 2335 : if (io_wref != NULL && pgaio_wref_valid(&buf->io_wref))
7291 : {
7292 2132 : *io_wref = buf->io_wref;
7293 2132 : UnlockBufHdr(buf);
7294 :
7295 2132 : return BUFFER_IO_IN_PROGRESS;
7296 : }
7297 203 : else if (!wait)
7298 : {
7299 8 : UnlockBufHdr(buf);
7300 8 : return BUFFER_IO_IN_PROGRESS;
7301 : }
7302 : else
7303 : {
7304 : /*
7305 : * With wait = true, we always have to wait if the caller has
7306 : * passed io_wref = NULL.
7307 : *
7308 : * Even with io_wref != NULL, we have to wait if the buffer's wait
7309 : * ref is not valid but the IO is in progress, someone else
7310 : * started IO but hasn't set the wait ref yet. We have no choice
7311 : * but to wait until the IO completes.
7312 : */
7313 195 : UnlockBufHdr(buf);
7314 :
7315 : /*
7316 : * If this backend currently has staged IO, submit it before
7317 : * waiting for in-progress IO, to avoid potential deadlocks and
7318 : * unnecessary delays.
7319 : */
7320 195 : pgaio_submit_staged();
7321 :
7322 195 : WaitIO(buf);
7323 : }
7324 : }
7325 :
7326 : /* Once we get here, there is definitely no I/O active on this buffer */
7327 :
7328 : /* Check if someone else already did the I/O */
7329 2975734 : if (forInput ? (buf_state & BM_VALID) : !(buf_state & BM_DIRTY))
7330 : {
7331 433 : UnlockBufHdr(buf);
7332 433 : return BUFFER_IO_ALREADY_DONE;
7333 : }
7334 :
7335 : /*
7336 : * No IO in progress and not already done; we will start IO. It's possible
7337 : * that the IO was in progress but we're not done, because the IO errored
7338 : * out. We'll do the IO ourselves.
7339 : */
7340 2975301 : UnlockBufHdrExt(buf, buf_state,
7341 : BM_IO_IN_PROGRESS, 0,
7342 : 0);
7343 :
7344 2975301 : ResourceOwnerRememberBufferIO(CurrentResourceOwner,
7345 : BufferDescriptorGetBuffer(buf));
7346 :
7347 2975301 : return BUFFER_IO_READY_FOR_IO;
7348 : }
7349 :
7350 : /*
7351 : * Wrapper around StartSharedBufferIO / StartLocalBufferIO. Only to be used
7352 : * when the caller doesn't otherwise need to care about local vs shared. See
7353 : * StartSharedBufferIO() for details.
7354 : */
7355 : StartBufferIOResult
7356 1690180 : StartBufferIO(Buffer buffer, bool forInput, bool wait, PgAioWaitRef *io_wref)
7357 : {
7358 : BufferDesc *buf_hdr;
7359 :
7360 1690180 : if (BufferIsLocal(buffer))
7361 : {
7362 11017 : buf_hdr = GetLocalBufferDescriptor(-buffer - 1);
7363 :
7364 11017 : return StartLocalBufferIO(buf_hdr, forInput, wait, io_wref);
7365 : }
7366 : else
7367 : {
7368 1679163 : buf_hdr = GetBufferDescriptor(buffer - 1);
7369 :
7370 1679163 : return StartSharedBufferIO(buf_hdr, forInput, wait, io_wref);
7371 : }
7372 : }
7373 :
7374 : /*
7375 : * TerminateBufferIO: release a buffer we were doing I/O on
7376 : * (Assumptions)
7377 : * My process is executing IO for the buffer
7378 : * BM_IO_IN_PROGRESS bit is set for the buffer
7379 : * The buffer is Pinned
7380 : *
7381 : * If clear_dirty is true, we clear the buffer's BM_DIRTY flag. This is
7382 : * appropriate when terminating a successful write.
7383 : *
7384 : * set_flag_bits gets ORed into the buffer's flags. It must include
7385 : * BM_IO_ERROR in a failure case. For successful completion it could
7386 : * be 0, or BM_VALID if we just finished reading in the page.
7387 : *
7388 : * If forget_owner is true, we release the buffer I/O from the current
7389 : * resource owner. (forget_owner=false is used when the resource owner itself
7390 : * is being released)
7391 : */
7392 : void
7393 2804482 : TerminateBufferIO(BufferDesc *buf, bool clear_dirty, uint64 set_flag_bits,
7394 : bool forget_owner, bool release_aio)
7395 : {
7396 : uint64 buf_state;
7397 2804482 : uint64 unset_flag_bits = 0;
7398 2804482 : int refcount_change = 0;
7399 :
7400 2804482 : buf_state = LockBufHdr(buf);
7401 :
7402 : Assert(buf_state & BM_IO_IN_PROGRESS);
7403 2804482 : unset_flag_bits |= BM_IO_IN_PROGRESS;
7404 :
7405 : /* Clear earlier errors, if this IO failed, it'll be marked again */
7406 2804482 : unset_flag_bits |= BM_IO_ERROR;
7407 :
7408 2804482 : if (clear_dirty)
7409 707084 : unset_flag_bits |= BM_DIRTY | BM_CHECKPOINT_NEEDED;
7410 :
7411 2804482 : if (release_aio)
7412 : {
7413 : /* release ownership by the AIO subsystem */
7414 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
7415 1505877 : refcount_change = -1;
7416 1505877 : pgaio_wref_clear(&buf->io_wref);
7417 : }
7418 :
7419 2804482 : buf_state = UnlockBufHdrExt(buf, buf_state,
7420 : set_flag_bits, unset_flag_bits,
7421 : refcount_change);
7422 :
7423 2804482 : if (forget_owner)
7424 1298582 : ResourceOwnerForgetBufferIO(CurrentResourceOwner,
7425 : BufferDescriptorGetBuffer(buf));
7426 :
7427 2804482 : ConditionVariableBroadcast(BufferDescriptorGetIOCV(buf));
7428 :
7429 : /*
7430 : * Support LockBufferForCleanup()
7431 : *
7432 : * We may have just released the last pin other than the waiter's. In most
7433 : * cases, this backend holds another pin on the buffer. But, if, for
7434 : * example, this backend is completing an IO issued by another backend, it
7435 : * may be time to wake the waiter.
7436 : */
7437 2804482 : if (release_aio && (buf_state & BM_PIN_COUNT_WAITER))
7438 0 : WakePinCountWaiter(buf);
7439 2804482 : }
7440 :
7441 : /*
7442 : * AbortBufferIO: Clean up active buffer I/O after an error.
7443 : *
7444 : * All LWLocks & content locks we might have held have been released, but we
7445 : * haven't yet released buffer pins, so the buffer is still pinned.
7446 : *
7447 : * If I/O was in progress, we always set BM_IO_ERROR, even though it's
7448 : * possible the error condition wasn't related to the I/O.
7449 : *
7450 : * Note: this does not remove the buffer I/O from the resource owner.
7451 : * That's correct when we're releasing the whole resource owner, but
7452 : * beware if you use this in other contexts.
7453 : */
7454 : static void
7455 15 : AbortBufferIO(Buffer buffer)
7456 : {
7457 15 : BufferDesc *buf_hdr = GetBufferDescriptor(buffer - 1);
7458 : uint64 buf_state;
7459 :
7460 15 : buf_state = LockBufHdr(buf_hdr);
7461 : Assert(buf_state & (BM_IO_IN_PROGRESS | BM_TAG_VALID));
7462 :
7463 15 : if (!(buf_state & BM_VALID))
7464 : {
7465 : Assert(!(buf_state & BM_DIRTY));
7466 15 : UnlockBufHdr(buf_hdr);
7467 : }
7468 : else
7469 : {
7470 : Assert(buf_state & BM_DIRTY);
7471 0 : UnlockBufHdr(buf_hdr);
7472 :
7473 : /* Issue notice if this is not the first failure... */
7474 0 : if (buf_state & BM_IO_ERROR)
7475 : {
7476 : /* Buffer is pinned, so we can read tag without spinlock */
7477 0 : ereport(WARNING,
7478 : (errcode(ERRCODE_IO_ERROR),
7479 : errmsg("could not write block %u of %s",
7480 : buf_hdr->tag.blockNum,
7481 : relpathperm(BufTagGetRelFileLocator(&buf_hdr->tag),
7482 : BufTagGetForkNum(&buf_hdr->tag)).str),
7483 : errdetail("Multiple failures --- write error might be permanent.")));
7484 : }
7485 : }
7486 :
7487 15 : TerminateBufferIO(buf_hdr, false, BM_IO_ERROR, false, false);
7488 15 : }
7489 :
7490 : /*
7491 : * Error context callback for errors occurring during shared buffer writes.
7492 : */
7493 : static void
7494 42 : shared_buffer_write_error_callback(void *arg)
7495 : {
7496 42 : BufferDesc *bufHdr = (BufferDesc *) arg;
7497 :
7498 : /* Buffer is pinned, so we can read the tag without locking the spinlock */
7499 42 : if (bufHdr != NULL)
7500 84 : errcontext("writing block %u of relation \"%s\"",
7501 : bufHdr->tag.blockNum,
7502 42 : relpathperm(BufTagGetRelFileLocator(&bufHdr->tag),
7503 : BufTagGetForkNum(&bufHdr->tag)).str);
7504 42 : }
7505 :
7506 : /*
7507 : * Error context callback for errors occurring during local buffer writes.
7508 : */
7509 : static void
7510 0 : local_buffer_write_error_callback(void *arg)
7511 : {
7512 0 : BufferDesc *bufHdr = (BufferDesc *) arg;
7513 :
7514 0 : if (bufHdr != NULL)
7515 0 : errcontext("writing block %u of relation \"%s\"",
7516 : bufHdr->tag.blockNum,
7517 0 : relpathbackend(BufTagGetRelFileLocator(&bufHdr->tag),
7518 : MyProcNumber,
7519 : BufTagGetForkNum(&bufHdr->tag)).str);
7520 0 : }
7521 :
7522 : /*
7523 : * RelFileLocator qsort/bsearch comparator; see RelFileLocatorEquals.
7524 : */
7525 : static int
7526 13256914 : rlocator_comparator(const void *p1, const void *p2)
7527 : {
7528 13256914 : RelFileLocator n1 = *(const RelFileLocator *) p1;
7529 13256914 : RelFileLocator n2 = *(const RelFileLocator *) p2;
7530 :
7531 13256914 : if (n1.relNumber < n2.relNumber)
7532 13211677 : return -1;
7533 45237 : else if (n1.relNumber > n2.relNumber)
7534 43470 : return 1;
7535 :
7536 1767 : if (n1.dbOid < n2.dbOid)
7537 0 : return -1;
7538 1767 : else if (n1.dbOid > n2.dbOid)
7539 0 : return 1;
7540 :
7541 1767 : if (n1.spcOid < n2.spcOid)
7542 0 : return -1;
7543 1767 : else if (n1.spcOid > n2.spcOid)
7544 0 : return 1;
7545 : else
7546 1767 : return 0;
7547 : }
7548 :
7549 : /*
7550 : * Lock buffer header - set BM_LOCKED in buffer state.
7551 : */
7552 : uint64
7553 28678567 : LockBufHdr(BufferDesc *desc)
7554 : {
7555 : uint64 old_buf_state;
7556 :
7557 : Assert(!BufferIsLocal(BufferDescriptorGetBuffer(desc)));
7558 :
7559 : while (true)
7560 : {
7561 : /*
7562 : * Always try once to acquire the lock directly, without setting up
7563 : * the spin-delay infrastructure. The work necessary for that shows up
7564 : * in profiles and is rarely necessary.
7565 : */
7566 28678878 : old_buf_state = pg_atomic_fetch_or_u64(&desc->state, BM_LOCKED);
7567 28678878 : if (likely(!(old_buf_state & BM_LOCKED)))
7568 28678567 : break; /* got lock */
7569 :
7570 : /* and then spin without atomic operations until lock is released */
7571 : {
7572 : SpinDelayStatus delayStatus;
7573 :
7574 311 : init_local_spin_delay(&delayStatus);
7575 :
7576 885 : while (old_buf_state & BM_LOCKED)
7577 : {
7578 574 : perform_spin_delay(&delayStatus);
7579 574 : old_buf_state = pg_atomic_read_u64(&desc->state);
7580 : }
7581 311 : finish_spin_delay(&delayStatus);
7582 : }
7583 :
7584 : /*
7585 : * Retry. The lock might obviously already be re-acquired by the time
7586 : * we're attempting to get it again.
7587 : */
7588 : }
7589 :
7590 28678567 : return old_buf_state | BM_LOCKED;
7591 : }
7592 :
7593 : /*
7594 : * Wait until the BM_LOCKED flag isn't set anymore and return the buffer's
7595 : * state at that point.
7596 : *
7597 : * Obviously the buffer could be locked by the time the value is returned, so
7598 : * this is primarily useful in CAS style loops.
7599 : */
7600 : pg_noinline uint64
7601 284 : WaitBufHdrUnlocked(BufferDesc *buf)
7602 : {
7603 : SpinDelayStatus delayStatus;
7604 : uint64 buf_state;
7605 :
7606 284 : init_local_spin_delay(&delayStatus);
7607 :
7608 284 : buf_state = pg_atomic_read_u64(&buf->state);
7609 :
7610 368 : while (buf_state & BM_LOCKED)
7611 : {
7612 84 : perform_spin_delay(&delayStatus);
7613 84 : buf_state = pg_atomic_read_u64(&buf->state);
7614 : }
7615 :
7616 284 : finish_spin_delay(&delayStatus);
7617 :
7618 284 : return buf_state;
7619 : }
7620 :
7621 : /*
7622 : * BufferTag comparator.
7623 : */
7624 : static inline int
7625 0 : buffertag_comparator(const BufferTag *ba, const BufferTag *bb)
7626 : {
7627 : int ret;
7628 : RelFileLocator rlocatora;
7629 : RelFileLocator rlocatorb;
7630 :
7631 0 : rlocatora = BufTagGetRelFileLocator(ba);
7632 0 : rlocatorb = BufTagGetRelFileLocator(bb);
7633 :
7634 0 : ret = rlocator_comparator(&rlocatora, &rlocatorb);
7635 :
7636 0 : if (ret != 0)
7637 0 : return ret;
7638 :
7639 0 : if (BufTagGetForkNum(ba) < BufTagGetForkNum(bb))
7640 0 : return -1;
7641 0 : if (BufTagGetForkNum(ba) > BufTagGetForkNum(bb))
7642 0 : return 1;
7643 :
7644 0 : if (ba->blockNum < bb->blockNum)
7645 0 : return -1;
7646 0 : if (ba->blockNum > bb->blockNum)
7647 0 : return 1;
7648 :
7649 0 : return 0;
7650 : }
7651 :
7652 : /*
7653 : * Comparator determining the writeout order in a checkpoint.
7654 : *
7655 : * It is important that tablespaces are compared first, the logic balancing
7656 : * writes between tablespaces relies on it.
7657 : */
7658 : static inline int
7659 3458941 : ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b)
7660 : {
7661 : /* compare tablespace */
7662 3458941 : if (a->tsId < b->tsId)
7663 10508 : return -1;
7664 3448433 : else if (a->tsId > b->tsId)
7665 29338 : return 1;
7666 : /* compare relation */
7667 3419095 : if (a->relNumber < b->relNumber)
7668 978661 : return -1;
7669 2440434 : else if (a->relNumber > b->relNumber)
7670 917844 : return 1;
7671 : /* compare fork */
7672 1522590 : else if (a->forkNum < b->forkNum)
7673 63697 : return -1;
7674 1458893 : else if (a->forkNum > b->forkNum)
7675 69106 : return 1;
7676 : /* compare block number */
7677 1389787 : else if (a->blockNum < b->blockNum)
7678 685272 : return -1;
7679 704515 : else if (a->blockNum > b->blockNum)
7680 658285 : return 1;
7681 : /* equal page IDs are unlikely, but not impossible */
7682 46230 : return 0;
7683 : }
7684 :
7685 : /*
7686 : * Comparator for a Min-Heap over the per-tablespace checkpoint completion
7687 : * progress.
7688 : */
7689 : static int
7690 281194 : ts_ckpt_progress_comparator(Datum a, Datum b, void *arg)
7691 : {
7692 281194 : CkptTsStatus *sa = (CkptTsStatus *) DatumGetPointer(a);
7693 281194 : CkptTsStatus *sb = (CkptTsStatus *) DatumGetPointer(b);
7694 :
7695 : /* we want a min-heap, so return 1 for the a < b */
7696 281194 : if (sa->progress < sb->progress)
7697 253296 : return 1;
7698 27898 : else if (sa->progress == sb->progress)
7699 998 : return 0;
7700 : else
7701 26900 : return -1;
7702 : }
7703 :
7704 : /*
7705 : * Initialize a writeback context, discarding potential previous state.
7706 : *
7707 : * *max_pending is a pointer instead of an immediate value, so the coalesce
7708 : * limits can easily changed by the GUC mechanism, and so calling code does
7709 : * not have to check the current configuration. A value of 0 means that no
7710 : * writeback control will be performed.
7711 : */
7712 : void
7713 3095 : WritebackContextInit(WritebackContext *context, int *max_pending)
7714 : {
7715 : Assert(*max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
7716 :
7717 3095 : context->max_pending = max_pending;
7718 3095 : context->nr_pending = 0;
7719 3095 : }
7720 :
7721 : /*
7722 : * Add buffer to list of pending writeback requests.
7723 : */
7724 : void
7725 703465 : ScheduleBufferTagForWriteback(WritebackContext *wb_context, IOContext io_context,
7726 : BufferTag *tag)
7727 : {
7728 : PendingWriteback *pending;
7729 :
7730 : /*
7731 : * As pg_flush_data() doesn't do anything with fsync disabled, there's no
7732 : * point in tracking in that case.
7733 : */
7734 703465 : if (io_direct_flags & IO_DIRECT_DATA ||
7735 702985 : !enableFsync)
7736 703463 : return;
7737 :
7738 : /*
7739 : * Add buffer to the pending writeback array, unless writeback control is
7740 : * disabled.
7741 : */
7742 2 : if (*wb_context->max_pending > 0)
7743 : {
7744 : Assert(*wb_context->max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
7745 :
7746 0 : pending = &wb_context->pending_writebacks[wb_context->nr_pending++];
7747 :
7748 0 : pending->tag = *tag;
7749 : }
7750 :
7751 : /*
7752 : * Perform pending flushes if the writeback limit is exceeded. This
7753 : * includes the case where previously an item has been added, but control
7754 : * is now disabled.
7755 : */
7756 2 : if (wb_context->nr_pending >= *wb_context->max_pending)
7757 2 : IssuePendingWritebacks(wb_context, io_context);
7758 : }
7759 :
7760 : #define ST_SORT sort_pending_writebacks
7761 : #define ST_ELEMENT_TYPE PendingWriteback
7762 : #define ST_COMPARE(a, b) buffertag_comparator(&a->tag, &b->tag)
7763 : #define ST_SCOPE static
7764 : #define ST_DEFINE
7765 : #include "lib/sort_template.h"
7766 :
7767 : /*
7768 : * Issue all pending writeback requests, previously scheduled with
7769 : * ScheduleBufferTagForWriteback, to the OS.
7770 : *
7771 : * Because this is only used to improve the OSs IO scheduling we try to never
7772 : * error out - it's just a hint.
7773 : */
7774 : void
7775 1216 : IssuePendingWritebacks(WritebackContext *wb_context, IOContext io_context)
7776 : {
7777 : instr_time io_start;
7778 : int i;
7779 :
7780 1216 : if (wb_context->nr_pending == 0)
7781 1216 : return;
7782 :
7783 : /*
7784 : * Executing the writes in-order can make them a lot faster, and allows to
7785 : * merge writeback requests to consecutive blocks into larger writebacks.
7786 : */
7787 0 : sort_pending_writebacks(wb_context->pending_writebacks,
7788 0 : wb_context->nr_pending);
7789 :
7790 0 : io_start = pgstat_prepare_io_time(track_io_timing);
7791 :
7792 : /*
7793 : * Coalesce neighbouring writes, but nothing else. For that we iterate
7794 : * through the, now sorted, array of pending flushes, and look forward to
7795 : * find all neighbouring (or identical) writes.
7796 : */
7797 0 : for (i = 0; i < wb_context->nr_pending; i++)
7798 : {
7799 : PendingWriteback *cur;
7800 : PendingWriteback *next;
7801 : SMgrRelation reln;
7802 : int ahead;
7803 : BufferTag tag;
7804 : RelFileLocator currlocator;
7805 0 : Size nblocks = 1;
7806 :
7807 0 : cur = &wb_context->pending_writebacks[i];
7808 0 : tag = cur->tag;
7809 0 : currlocator = BufTagGetRelFileLocator(&tag);
7810 :
7811 : /*
7812 : * Peek ahead, into following writeback requests, to see if they can
7813 : * be combined with the current one.
7814 : */
7815 0 : for (ahead = 0; i + ahead + 1 < wb_context->nr_pending; ahead++)
7816 : {
7817 :
7818 0 : next = &wb_context->pending_writebacks[i + ahead + 1];
7819 :
7820 : /* different file, stop */
7821 0 : if (!RelFileLocatorEquals(currlocator,
7822 0 : BufTagGetRelFileLocator(&next->tag)) ||
7823 0 : BufTagGetForkNum(&cur->tag) != BufTagGetForkNum(&next->tag))
7824 : break;
7825 :
7826 : /* ok, block queued twice, skip */
7827 0 : if (cur->tag.blockNum == next->tag.blockNum)
7828 0 : continue;
7829 :
7830 : /* only merge consecutive writes */
7831 0 : if (cur->tag.blockNum + 1 != next->tag.blockNum)
7832 0 : break;
7833 :
7834 0 : nblocks++;
7835 0 : cur = next;
7836 : }
7837 :
7838 0 : i += ahead;
7839 :
7840 : /* and finally tell the kernel to write the data to storage */
7841 0 : reln = smgropen(currlocator, INVALID_PROC_NUMBER);
7842 0 : smgrwriteback(reln, BufTagGetForkNum(&tag), tag.blockNum, nblocks);
7843 : }
7844 :
7845 : /*
7846 : * Assume that writeback requests are only issued for buffers containing
7847 : * blocks of permanent relations.
7848 : */
7849 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
7850 0 : IOOP_WRITEBACK, io_start, wb_context->nr_pending, 0);
7851 :
7852 0 : wb_context->nr_pending = 0;
7853 : }
7854 :
7855 : /* ResourceOwner callbacks */
7856 :
7857 : static void
7858 15 : ResOwnerReleaseBufferIO(Datum res)
7859 : {
7860 15 : Buffer buffer = DatumGetInt32(res);
7861 :
7862 15 : AbortBufferIO(buffer);
7863 15 : }
7864 :
7865 : static char *
7866 0 : ResOwnerPrintBufferIO(Datum res)
7867 : {
7868 0 : Buffer buffer = DatumGetInt32(res);
7869 :
7870 0 : return psprintf("lost track of buffer IO on buffer %d", buffer);
7871 : }
7872 :
7873 : /*
7874 : * Release buffer as part of resource owner cleanup. This will only be called
7875 : * if the buffer is pinned. If this backend held the content lock at the time
7876 : * of the error we also need to release that (note that it is not possible to
7877 : * hold a content lock without a pin).
7878 : */
7879 : static void
7880 10682 : ResOwnerReleaseBuffer(Datum res)
7881 : {
7882 10682 : Buffer buffer = DatumGetInt32(res);
7883 :
7884 : /* Like ReleaseBuffer, but don't call ResourceOwnerForgetBuffer */
7885 10682 : if (!BufferIsValid(buffer))
7886 0 : elog(ERROR, "bad buffer ID: %d", buffer);
7887 :
7888 10682 : if (BufferIsLocal(buffer))
7889 3995 : UnpinLocalBufferNoOwner(buffer);
7890 : else
7891 : {
7892 : PrivateRefCountEntry *ref;
7893 :
7894 6687 : ref = GetPrivateRefCountEntry(buffer, false);
7895 :
7896 : /* not having a private refcount would imply resowner corruption */
7897 : Assert(ref != NULL);
7898 :
7899 : /*
7900 : * If the buffer was locked at the time of the resowner release,
7901 : * release the lock now. This should only happen after errors.
7902 : */
7903 6687 : if (ref->data.lockmode != BUFFER_LOCK_UNLOCK)
7904 : {
7905 112 : BufferDesc *buf = GetBufferDescriptor(buffer - 1);
7906 :
7907 112 : HOLD_INTERRUPTS(); /* match the upcoming RESUME_INTERRUPTS */
7908 112 : BufferLockUnlock(buffer, buf);
7909 : }
7910 :
7911 6687 : UnpinBufferNoOwner(GetBufferDescriptor(buffer - 1));
7912 : }
7913 10682 : }
7914 :
7915 : static char *
7916 0 : ResOwnerPrintBuffer(Datum res)
7917 : {
7918 0 : return DebugPrintBufferRefcount(DatumGetInt32(res));
7919 : }
7920 :
7921 : /*
7922 : * Helper function to evict unpinned buffer whose buffer header lock is
7923 : * already acquired.
7924 : */
7925 : static bool
7926 2509 : EvictUnpinnedBufferInternal(BufferDesc *desc, bool *buffer_flushed)
7927 : {
7928 : uint64 buf_state;
7929 : bool result;
7930 :
7931 2509 : *buffer_flushed = false;
7932 :
7933 2509 : buf_state = pg_atomic_read_u64(&(desc->state));
7934 : Assert(buf_state & BM_LOCKED);
7935 :
7936 2509 : if ((buf_state & BM_VALID) == 0)
7937 : {
7938 0 : UnlockBufHdr(desc);
7939 0 : return false;
7940 : }
7941 :
7942 : /* Check that it's not pinned already. */
7943 2509 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
7944 : {
7945 0 : UnlockBufHdr(desc);
7946 0 : return false;
7947 : }
7948 :
7949 2509 : PinBuffer_Locked(desc); /* releases spinlock */
7950 :
7951 : /* If it was dirty, try to clean it once. */
7952 2509 : if (buf_state & BM_DIRTY)
7953 : {
7954 1075 : FlushUnlockedBuffer(desc, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
7955 1075 : *buffer_flushed = true;
7956 : }
7957 :
7958 : /* This will return false if it becomes dirty or someone else pins it. */
7959 2509 : result = InvalidateVictimBuffer(desc);
7960 :
7961 2509 : UnpinBuffer(desc);
7962 :
7963 2509 : return result;
7964 : }
7965 :
7966 : /*
7967 : * Try to evict the current block in a shared buffer.
7968 : *
7969 : * This function is intended for testing/development use only!
7970 : *
7971 : * To succeed, the buffer must not be pinned on entry, so if the caller had a
7972 : * particular block in mind, it might already have been replaced by some other
7973 : * block by the time this function runs. It's also unpinned on return, so the
7974 : * buffer might be occupied again by the time control is returned, potentially
7975 : * even by the same block. This inherent raciness without other interlocking
7976 : * makes the function unsuitable for non-testing usage.
7977 : *
7978 : * *buffer_flushed is set to true if the buffer was dirty and has been
7979 : * flushed, false otherwise. However, *buffer_flushed=true does not
7980 : * necessarily mean that we flushed the buffer, it could have been flushed by
7981 : * someone else.
7982 : *
7983 : * Returns true if the buffer was valid and it has now been made invalid.
7984 : * Returns false if it wasn't valid, if it couldn't be evicted due to a pin,
7985 : * or if the buffer becomes dirty again while we're trying to write it out.
7986 : */
7987 : bool
7988 137 : EvictUnpinnedBuffer(Buffer buf, bool *buffer_flushed)
7989 : {
7990 : BufferDesc *desc;
7991 :
7992 : Assert(BufferIsValid(buf) && !BufferIsLocal(buf));
7993 :
7994 : /* Make sure we can pin the buffer. */
7995 137 : ResourceOwnerEnlarge(CurrentResourceOwner);
7996 137 : ReservePrivateRefCountEntry();
7997 :
7998 137 : desc = GetBufferDescriptor(buf - 1);
7999 137 : LockBufHdr(desc);
8000 :
8001 137 : return EvictUnpinnedBufferInternal(desc, buffer_flushed);
8002 : }
8003 :
8004 : /*
8005 : * Try to evict all the shared buffers.
8006 : *
8007 : * This function is intended for testing/development use only! See
8008 : * EvictUnpinnedBuffer().
8009 : *
8010 : * The buffers_* parameters are mandatory and indicate the total count of
8011 : * buffers that:
8012 : * - buffers_evicted - were evicted
8013 : * - buffers_flushed - were flushed
8014 : * - buffers_skipped - could not be evicted
8015 : */
8016 : void
8017 1 : EvictAllUnpinnedBuffers(int32 *buffers_evicted, int32 *buffers_flushed,
8018 : int32 *buffers_skipped)
8019 : {
8020 1 : *buffers_evicted = 0;
8021 1 : *buffers_skipped = 0;
8022 1 : *buffers_flushed = 0;
8023 :
8024 16385 : for (int buf = 1; buf <= NBuffers; buf++)
8025 : {
8026 16384 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
8027 : uint64 buf_state;
8028 : bool buffer_flushed;
8029 :
8030 16384 : CHECK_FOR_INTERRUPTS();
8031 :
8032 16384 : buf_state = pg_atomic_read_u64(&desc->state);
8033 16384 : if (!(buf_state & BM_VALID))
8034 14358 : continue;
8035 :
8036 2026 : ResourceOwnerEnlarge(CurrentResourceOwner);
8037 2026 : ReservePrivateRefCountEntry();
8038 :
8039 2026 : LockBufHdr(desc);
8040 :
8041 2026 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
8042 2026 : (*buffers_evicted)++;
8043 : else
8044 0 : (*buffers_skipped)++;
8045 :
8046 2026 : if (buffer_flushed)
8047 978 : (*buffers_flushed)++;
8048 : }
8049 1 : }
8050 :
8051 : /*
8052 : * Try to evict all the shared buffers containing provided relation's pages.
8053 : *
8054 : * This function is intended for testing/development use only! See
8055 : * EvictUnpinnedBuffer().
8056 : *
8057 : * The caller must hold at least AccessShareLock on the relation to prevent
8058 : * the relation from being dropped.
8059 : *
8060 : * The buffers_* parameters are mandatory and indicate the total count of
8061 : * buffers that:
8062 : * - buffers_evicted - were evicted
8063 : * - buffers_flushed - were flushed
8064 : * - buffers_skipped - could not be evicted
8065 : */
8066 : void
8067 29 : EvictRelUnpinnedBuffers(Relation rel, int32 *buffers_evicted,
8068 : int32 *buffers_flushed, int32 *buffers_skipped)
8069 : {
8070 : Assert(!RelationUsesLocalBuffers(rel));
8071 :
8072 29 : *buffers_skipped = 0;
8073 29 : *buffers_evicted = 0;
8074 29 : *buffers_flushed = 0;
8075 :
8076 475165 : for (int buf = 1; buf <= NBuffers; buf++)
8077 : {
8078 475136 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
8079 475136 : uint64 buf_state = pg_atomic_read_u64(&(desc->state));
8080 : bool buffer_flushed;
8081 :
8082 475136 : CHECK_FOR_INTERRUPTS();
8083 :
8084 : /* An unlocked precheck should be safe and saves some cycles. */
8085 475136 : if ((buf_state & BM_VALID) == 0 ||
8086 62777 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
8087 474790 : continue;
8088 :
8089 : /* Make sure we can pin the buffer. */
8090 346 : ResourceOwnerEnlarge(CurrentResourceOwner);
8091 346 : ReservePrivateRefCountEntry();
8092 :
8093 346 : buf_state = LockBufHdr(desc);
8094 :
8095 : /* recheck, could have changed without the lock */
8096 346 : if ((buf_state & BM_VALID) == 0 ||
8097 346 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
8098 : {
8099 0 : UnlockBufHdr(desc);
8100 0 : continue;
8101 : }
8102 :
8103 346 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
8104 346 : (*buffers_evicted)++;
8105 : else
8106 0 : (*buffers_skipped)++;
8107 :
8108 346 : if (buffer_flushed)
8109 78 : (*buffers_flushed)++;
8110 : }
8111 29 : }
8112 :
8113 : /*
8114 : * Helper function to mark unpinned buffer dirty whose buffer header lock is
8115 : * already acquired.
8116 : */
8117 : static bool
8118 36 : MarkDirtyUnpinnedBufferInternal(Buffer buf, BufferDesc *desc,
8119 : bool *buffer_already_dirty)
8120 : {
8121 : uint64 buf_state;
8122 36 : bool result = false;
8123 :
8124 36 : *buffer_already_dirty = false;
8125 :
8126 36 : buf_state = pg_atomic_read_u64(&(desc->state));
8127 : Assert(buf_state & BM_LOCKED);
8128 :
8129 36 : if ((buf_state & BM_VALID) == 0)
8130 : {
8131 1 : UnlockBufHdr(desc);
8132 1 : return false;
8133 : }
8134 :
8135 : /* Check that it's not pinned already. */
8136 35 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
8137 : {
8138 0 : UnlockBufHdr(desc);
8139 0 : return false;
8140 : }
8141 :
8142 : /* Pin the buffer and then release the buffer spinlock */
8143 35 : PinBuffer_Locked(desc);
8144 :
8145 : /* If it was not already dirty, mark it as dirty. */
8146 35 : if (!(buf_state & BM_DIRTY))
8147 : {
8148 16 : BufferLockAcquire(buf, desc, BUFFER_LOCK_EXCLUSIVE);
8149 16 : MarkBufferDirty(buf);
8150 16 : result = true;
8151 16 : BufferLockUnlock(buf, desc);
8152 : }
8153 : else
8154 19 : *buffer_already_dirty = true;
8155 :
8156 35 : UnpinBuffer(desc);
8157 :
8158 35 : return result;
8159 : }
8160 :
8161 : /*
8162 : * Try to mark the provided shared buffer as dirty.
8163 : *
8164 : * This function is intended for testing/development use only!
8165 : *
8166 : * Same as EvictUnpinnedBuffer() but with MarkBufferDirty() call inside.
8167 : *
8168 : * The buffer_already_dirty parameter is mandatory and indicate if the buffer
8169 : * could not be dirtied because it is already dirty.
8170 : *
8171 : * Returns true if the buffer has successfully been marked as dirty.
8172 : */
8173 : bool
8174 1 : MarkDirtyUnpinnedBuffer(Buffer buf, bool *buffer_already_dirty)
8175 : {
8176 : BufferDesc *desc;
8177 1 : bool buffer_dirtied = false;
8178 :
8179 : Assert(!BufferIsLocal(buf));
8180 :
8181 : /* Make sure we can pin the buffer. */
8182 1 : ResourceOwnerEnlarge(CurrentResourceOwner);
8183 1 : ReservePrivateRefCountEntry();
8184 :
8185 1 : desc = GetBufferDescriptor(buf - 1);
8186 1 : LockBufHdr(desc);
8187 :
8188 1 : buffer_dirtied = MarkDirtyUnpinnedBufferInternal(buf, desc, buffer_already_dirty);
8189 : /* Both can not be true at the same time */
8190 : Assert(!(buffer_dirtied && *buffer_already_dirty));
8191 :
8192 1 : return buffer_dirtied;
8193 : }
8194 :
8195 : /*
8196 : * Try to mark all the shared buffers containing provided relation's pages as
8197 : * dirty.
8198 : *
8199 : * This function is intended for testing/development use only! See
8200 : * MarkDirtyUnpinnedBuffer().
8201 : *
8202 : * The buffers_* parameters are mandatory and indicate the total count of
8203 : * buffers that:
8204 : * - buffers_dirtied - were dirtied
8205 : * - buffers_already_dirty - were already dirty
8206 : * - buffers_skipped - could not be dirtied because of a reason different
8207 : * than a buffer being already dirty.
8208 : */
8209 : void
8210 1 : MarkDirtyRelUnpinnedBuffers(Relation rel,
8211 : int32 *buffers_dirtied,
8212 : int32 *buffers_already_dirty,
8213 : int32 *buffers_skipped)
8214 : {
8215 : Assert(!RelationUsesLocalBuffers(rel));
8216 :
8217 1 : *buffers_dirtied = 0;
8218 1 : *buffers_already_dirty = 0;
8219 1 : *buffers_skipped = 0;
8220 :
8221 16385 : for (int buf = 1; buf <= NBuffers; buf++)
8222 : {
8223 16384 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
8224 16384 : uint64 buf_state = pg_atomic_read_u64(&(desc->state));
8225 : bool buffer_already_dirty;
8226 :
8227 16384 : CHECK_FOR_INTERRUPTS();
8228 :
8229 : /* An unlocked precheck should be safe and saves some cycles. */
8230 16384 : if ((buf_state & BM_VALID) == 0 ||
8231 27 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
8232 16384 : continue;
8233 :
8234 : /* Make sure we can pin the buffer. */
8235 0 : ResourceOwnerEnlarge(CurrentResourceOwner);
8236 0 : ReservePrivateRefCountEntry();
8237 :
8238 0 : buf_state = LockBufHdr(desc);
8239 :
8240 : /* recheck, could have changed without the lock */
8241 0 : if ((buf_state & BM_VALID) == 0 ||
8242 0 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
8243 : {
8244 0 : UnlockBufHdr(desc);
8245 0 : continue;
8246 : }
8247 :
8248 0 : if (MarkDirtyUnpinnedBufferInternal(buf, desc, &buffer_already_dirty))
8249 0 : (*buffers_dirtied)++;
8250 0 : else if (buffer_already_dirty)
8251 0 : (*buffers_already_dirty)++;
8252 : else
8253 0 : (*buffers_skipped)++;
8254 : }
8255 1 : }
8256 :
8257 : /*
8258 : * Try to mark all the shared buffers as dirty.
8259 : *
8260 : * This function is intended for testing/development use only! See
8261 : * MarkDirtyUnpinnedBuffer().
8262 : *
8263 : * See MarkDirtyRelUnpinnedBuffers() above for details about the buffers_*
8264 : * parameters.
8265 : */
8266 : void
8267 1 : MarkDirtyAllUnpinnedBuffers(int32 *buffers_dirtied,
8268 : int32 *buffers_already_dirty,
8269 : int32 *buffers_skipped)
8270 : {
8271 1 : *buffers_dirtied = 0;
8272 1 : *buffers_already_dirty = 0;
8273 1 : *buffers_skipped = 0;
8274 :
8275 16385 : for (int buf = 1; buf <= NBuffers; buf++)
8276 : {
8277 16384 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
8278 : uint64 buf_state;
8279 : bool buffer_already_dirty;
8280 :
8281 16384 : CHECK_FOR_INTERRUPTS();
8282 :
8283 16384 : buf_state = pg_atomic_read_u64(&desc->state);
8284 16384 : if (!(buf_state & BM_VALID))
8285 16349 : continue;
8286 :
8287 35 : ResourceOwnerEnlarge(CurrentResourceOwner);
8288 35 : ReservePrivateRefCountEntry();
8289 :
8290 35 : LockBufHdr(desc);
8291 :
8292 35 : if (MarkDirtyUnpinnedBufferInternal(buf, desc, &buffer_already_dirty))
8293 16 : (*buffers_dirtied)++;
8294 19 : else if (buffer_already_dirty)
8295 19 : (*buffers_already_dirty)++;
8296 : else
8297 0 : (*buffers_skipped)++;
8298 : }
8299 1 : }
8300 :
8301 : /*
8302 : * Generic implementation of the AIO handle staging callback for readv/writev
8303 : * on local/shared buffers.
8304 : *
8305 : * Each readv/writev can target multiple buffers. The buffers have already
8306 : * been registered with the IO handle.
8307 : *
8308 : * To make the IO ready for execution ("staging"), we need to ensure that the
8309 : * targeted buffers are in an appropriate state while the IO is ongoing. For
8310 : * that the AIO subsystem needs to have its own buffer pin, otherwise an error
8311 : * in this backend could lead to this backend's buffer pin being released as
8312 : * part of error handling, which in turn could lead to the buffer being
8313 : * replaced while IO is ongoing.
8314 : */
8315 : static pg_attribute_always_inline void
8316 1506698 : buffer_stage_common(PgAioHandle *ioh, bool is_write, bool is_temp)
8317 : {
8318 : uint64 *io_data;
8319 : uint8 handle_data_len;
8320 : PgAioWaitRef io_ref;
8321 1506698 : BufferTag first PG_USED_FOR_ASSERTS_ONLY = {0};
8322 :
8323 1506698 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
8324 :
8325 1506698 : pgaio_io_get_wref(ioh, &io_ref);
8326 :
8327 : /* iterate over all buffers affected by the vectored readv/writev */
8328 3194457 : for (int i = 0; i < handle_data_len; i++)
8329 : {
8330 1687759 : Buffer buffer = (Buffer) io_data[i];
8331 1687759 : BufferDesc *buf_hdr = is_temp ?
8332 11063 : GetLocalBufferDescriptor(-buffer - 1)
8333 1687759 : : GetBufferDescriptor(buffer - 1);
8334 : uint64 buf_state;
8335 :
8336 : /*
8337 : * Check that all the buffers are actually ones that could conceivably
8338 : * be done in one IO, i.e. are sequential. This is the last
8339 : * buffer-aware code before IO is actually executed and confusion
8340 : * about which buffers are targeted by IO can be hard to debug, making
8341 : * it worth doing extra-paranoid checks.
8342 : */
8343 1687759 : if (i == 0)
8344 1506698 : first = buf_hdr->tag;
8345 : else
8346 : {
8347 : Assert(buf_hdr->tag.relNumber == first.relNumber);
8348 : Assert(buf_hdr->tag.blockNum == first.blockNum + i);
8349 : }
8350 :
8351 1687759 : if (is_temp)
8352 11063 : buf_state = pg_atomic_read_u64(&buf_hdr->state);
8353 : else
8354 1676696 : buf_state = LockBufHdr(buf_hdr);
8355 :
8356 : /* verify the buffer is in the expected state */
8357 : Assert(buf_state & BM_TAG_VALID);
8358 : if (is_write)
8359 : {
8360 : Assert(buf_state & BM_VALID);
8361 : Assert(buf_state & BM_DIRTY);
8362 : }
8363 : else
8364 : {
8365 : Assert(!(buf_state & BM_VALID));
8366 : Assert(!(buf_state & BM_DIRTY));
8367 : }
8368 :
8369 : /* temp buffers don't use BM_IO_IN_PROGRESS */
8370 1687759 : if (!is_temp)
8371 : Assert(buf_state & BM_IO_IN_PROGRESS);
8372 :
8373 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) >= 1);
8374 :
8375 : /*
8376 : * Reflect that the buffer is now owned by the AIO subsystem.
8377 : *
8378 : * For local buffers: This can't be done just via LocalRefCount, as
8379 : * one might initially think, as this backend could error out while
8380 : * AIO is still in progress, releasing all the pins by the backend
8381 : * itself.
8382 : *
8383 : * This pin is released again in TerminateBufferIO().
8384 : */
8385 1687759 : buf_hdr->io_wref = io_ref;
8386 :
8387 1687759 : if (is_temp)
8388 : {
8389 11063 : buf_state += BUF_REFCOUNT_ONE;
8390 11063 : pg_atomic_unlocked_write_u64(&buf_hdr->state, buf_state);
8391 : }
8392 : else
8393 1676696 : UnlockBufHdrExt(buf_hdr, buf_state, 0, 0, 1);
8394 :
8395 : /*
8396 : * Ensure the content lock that prevents buffer modifications while
8397 : * the buffer is being written out is not released early due to an
8398 : * error.
8399 : */
8400 1687759 : if (is_write && !is_temp)
8401 : {
8402 : Assert(BufferLockHeldByMe(buf_hdr));
8403 :
8404 : /*
8405 : * Lock is now owned by AIO subsystem.
8406 : */
8407 0 : BufferLockDisown(buffer, buf_hdr);
8408 : }
8409 :
8410 : /*
8411 : * Stop tracking this buffer via the resowner - the AIO system now
8412 : * keeps track.
8413 : */
8414 1687759 : if (!is_temp)
8415 1676696 : ResourceOwnerForgetBufferIO(CurrentResourceOwner, buffer);
8416 : }
8417 1506698 : }
8418 :
8419 : /*
8420 : * Decode readv errors as encoded by buffer_readv_encode_error().
8421 : */
8422 : static inline void
8423 450 : buffer_readv_decode_error(PgAioResult result,
8424 : bool *zeroed_any,
8425 : bool *ignored_any,
8426 : uint8 *zeroed_or_error_count,
8427 : uint8 *checkfail_count,
8428 : uint8 *first_off)
8429 : {
8430 450 : uint32 rem_error = result.error_data;
8431 :
8432 : /* see static asserts in buffer_readv_encode_error */
8433 : #define READV_COUNT_BITS 7
8434 : #define READV_COUNT_MASK ((1 << READV_COUNT_BITS) - 1)
8435 :
8436 450 : *zeroed_any = rem_error & 1;
8437 450 : rem_error >>= 1;
8438 :
8439 450 : *ignored_any = rem_error & 1;
8440 450 : rem_error >>= 1;
8441 :
8442 450 : *zeroed_or_error_count = rem_error & READV_COUNT_MASK;
8443 450 : rem_error >>= READV_COUNT_BITS;
8444 :
8445 450 : *checkfail_count = rem_error & READV_COUNT_MASK;
8446 450 : rem_error >>= READV_COUNT_BITS;
8447 :
8448 450 : *first_off = rem_error & READV_COUNT_MASK;
8449 450 : rem_error >>= READV_COUNT_BITS;
8450 450 : }
8451 :
8452 : /*
8453 : * Helper to encode errors for buffer_readv_complete()
8454 : *
8455 : * Errors are encoded as follows:
8456 : * - bit 0 indicates whether any page was zeroed (1) or not (0)
8457 : * - bit 1 indicates whether any checksum failure was ignored (1) or not (0)
8458 : * - next READV_COUNT_BITS bits indicate the number of errored or zeroed pages
8459 : * - next READV_COUNT_BITS bits indicate the number of checksum failures
8460 : * - next READV_COUNT_BITS bits indicate the first offset of the first page
8461 : * that was errored or zeroed or, if no errors/zeroes, the first ignored
8462 : * checksum
8463 : */
8464 : static inline void
8465 192 : buffer_readv_encode_error(PgAioResult *result,
8466 : bool is_temp,
8467 : bool zeroed_any,
8468 : bool ignored_any,
8469 : uint8 error_count,
8470 : uint8 zeroed_count,
8471 : uint8 checkfail_count,
8472 : uint8 first_error_off,
8473 : uint8 first_zeroed_off,
8474 : uint8 first_ignored_off)
8475 : {
8476 :
8477 192 : uint8 shift = 0;
8478 192 : uint8 zeroed_or_error_count =
8479 : error_count > 0 ? error_count : zeroed_count;
8480 : uint8 first_off;
8481 :
8482 : StaticAssertDecl(PG_IOV_MAX <= 1 << READV_COUNT_BITS,
8483 : "PG_IOV_MAX is bigger than reserved space for error data");
8484 : StaticAssertDecl((1 + 1 + 3 * READV_COUNT_BITS) <= PGAIO_RESULT_ERROR_BITS,
8485 : "PGAIO_RESULT_ERROR_BITS is insufficient for buffer_readv");
8486 :
8487 : /*
8488 : * We only have space to encode one offset - but luckily that's good
8489 : * enough. If there is an error, the error is the interesting offset, same
8490 : * with a zeroed buffer vs an ignored buffer.
8491 : */
8492 192 : if (error_count > 0)
8493 94 : first_off = first_error_off;
8494 98 : else if (zeroed_count > 0)
8495 80 : first_off = first_zeroed_off;
8496 : else
8497 18 : first_off = first_ignored_off;
8498 :
8499 : Assert(!zeroed_any || error_count == 0);
8500 :
8501 192 : result->error_data = 0;
8502 :
8503 192 : result->error_data |= zeroed_any << shift;
8504 192 : shift += 1;
8505 :
8506 192 : result->error_data |= ignored_any << shift;
8507 192 : shift += 1;
8508 :
8509 192 : result->error_data |= ((uint32) zeroed_or_error_count) << shift;
8510 192 : shift += READV_COUNT_BITS;
8511 :
8512 192 : result->error_data |= ((uint32) checkfail_count) << shift;
8513 192 : shift += READV_COUNT_BITS;
8514 :
8515 192 : result->error_data |= ((uint32) first_off) << shift;
8516 192 : shift += READV_COUNT_BITS;
8517 :
8518 192 : result->id = is_temp ? PGAIO_HCB_LOCAL_BUFFER_READV :
8519 : PGAIO_HCB_SHARED_BUFFER_READV;
8520 :
8521 192 : if (error_count > 0)
8522 94 : result->status = PGAIO_RS_ERROR;
8523 : else
8524 98 : result->status = PGAIO_RS_WARNING;
8525 :
8526 : /*
8527 : * The encoding is complicated enough to warrant cross-checking it against
8528 : * the decode function.
8529 : */
8530 : #ifdef USE_ASSERT_CHECKING
8531 : {
8532 : bool zeroed_any_2,
8533 : ignored_any_2;
8534 : uint8 zeroed_or_error_count_2,
8535 : checkfail_count_2,
8536 : first_off_2;
8537 :
8538 : buffer_readv_decode_error(*result,
8539 : &zeroed_any_2, &ignored_any_2,
8540 : &zeroed_or_error_count_2,
8541 : &checkfail_count_2,
8542 : &first_off_2);
8543 : Assert(zeroed_any == zeroed_any_2);
8544 : Assert(ignored_any == ignored_any_2);
8545 : Assert(zeroed_or_error_count == zeroed_or_error_count_2);
8546 : Assert(checkfail_count == checkfail_count_2);
8547 : Assert(first_off == first_off_2);
8548 : }
8549 : #endif
8550 :
8551 : #undef READV_COUNT_BITS
8552 : #undef READV_COUNT_MASK
8553 192 : }
8554 :
8555 : /*
8556 : * Helper for AIO readv completion callbacks, supporting both shared and temp
8557 : * buffers. Gets called once for each buffer in a multi-page read.
8558 : */
8559 : static pg_attribute_always_inline void
8560 1516940 : buffer_readv_complete_one(PgAioTargetData *td, uint8 buf_off, Buffer buffer,
8561 : uint8 flags, bool failed, bool is_temp,
8562 : bool *buffer_invalid,
8563 : bool *failed_checksum,
8564 : bool *ignored_checksum,
8565 : bool *zeroed_buffer)
8566 : {
8567 1516940 : BufferDesc *buf_hdr = is_temp ?
8568 11063 : GetLocalBufferDescriptor(-buffer - 1)
8569 1516940 : : GetBufferDescriptor(buffer - 1);
8570 1516940 : BufferTag tag = buf_hdr->tag;
8571 1516940 : char *bufdata = BufferGetBlock(buffer);
8572 : uint64 set_flag_bits;
8573 : int piv_flags;
8574 :
8575 : /* check that the buffer is in the expected state for a read */
8576 : #ifdef USE_ASSERT_CHECKING
8577 : {
8578 : uint64 buf_state = pg_atomic_read_u64(&buf_hdr->state);
8579 :
8580 : Assert(buf_state & BM_TAG_VALID);
8581 : Assert(!(buf_state & BM_VALID));
8582 : /* temp buffers don't use BM_IO_IN_PROGRESS */
8583 : if (!is_temp)
8584 : Assert(buf_state & BM_IO_IN_PROGRESS);
8585 : Assert(!(buf_state & BM_DIRTY));
8586 : }
8587 : #endif
8588 :
8589 1516940 : *buffer_invalid = false;
8590 1516940 : *failed_checksum = false;
8591 1516940 : *ignored_checksum = false;
8592 1516940 : *zeroed_buffer = false;
8593 :
8594 : /*
8595 : * We ask PageIsVerified() to only log the message about checksum errors,
8596 : * as the completion might be run in any backend (or IO workers). We will
8597 : * report checksum errors in buffer_readv_report().
8598 : */
8599 1516940 : piv_flags = PIV_LOG_LOG;
8600 :
8601 : /* the local zero_damaged_pages may differ from the definer's */
8602 1516940 : if (flags & READ_BUFFERS_IGNORE_CHECKSUM_FAILURES)
8603 38 : piv_flags |= PIV_IGNORE_CHECKSUM_FAILURE;
8604 :
8605 : /*
8606 : * If the buffers are marked for zero on error, we want to log that in
8607 : * case of a checksum failure.
8608 : */
8609 1516940 : if (flags & READ_BUFFERS_ZERO_ON_ERROR)
8610 47379 : piv_flags |= PIV_ZERO_BUFFERS_ON_ERROR;
8611 :
8612 : /* Check for garbage data. */
8613 1516940 : if (!failed)
8614 : {
8615 : /*
8616 : * If the buffer is not currently pinned by this backend, e.g. because
8617 : * we're completing this IO after an error, the buffer data will have
8618 : * been marked as inaccessible when the buffer was unpinned. The AIO
8619 : * subsystem holds a pin, but that doesn't prevent the buffer from
8620 : * having been marked as inaccessible. The completion might also be
8621 : * executed in a different process.
8622 : */
8623 : #ifdef USE_VALGRIND
8624 : if (!BufferIsPinned(buffer))
8625 : VALGRIND_MAKE_MEM_DEFINED(bufdata, BLCKSZ);
8626 : #endif
8627 :
8628 1516168 : if (!PageIsVerified((Page) bufdata, tag.blockNum, piv_flags,
8629 : failed_checksum))
8630 : {
8631 96 : if (flags & READ_BUFFERS_ZERO_ON_ERROR)
8632 : {
8633 46 : memset(bufdata, 0, BLCKSZ);
8634 46 : *zeroed_buffer = true;
8635 : }
8636 : else
8637 : {
8638 50 : *buffer_invalid = true;
8639 : /* mark buffer as having failed */
8640 50 : failed = true;
8641 : }
8642 : }
8643 1516072 : else if (*failed_checksum)
8644 12 : *ignored_checksum = true;
8645 :
8646 : /* undo what we did above */
8647 : #ifdef USE_VALGRIND
8648 : if (!BufferIsPinned(buffer))
8649 : VALGRIND_MAKE_MEM_NOACCESS(bufdata, BLCKSZ);
8650 : #endif
8651 :
8652 : /*
8653 : * Immediately log a message about the invalid page, but only to the
8654 : * server log. The reason to do so immediately is that this may be
8655 : * executed in a different backend than the one that originated the
8656 : * request. The reason to do so immediately is that the originator
8657 : * might not process the query result immediately (because it is busy
8658 : * doing another part of query processing) or at all (e.g. if it was
8659 : * cancelled or errored out due to another IO also failing). The
8660 : * definer of the IO will emit an ERROR or WARNING when processing the
8661 : * IO's results
8662 : *
8663 : * To avoid duplicating the code to emit these log messages, we reuse
8664 : * buffer_readv_report().
8665 : */
8666 1516168 : if (*buffer_invalid || *failed_checksum || *zeroed_buffer)
8667 : {
8668 108 : PgAioResult result_one = {0};
8669 :
8670 108 : buffer_readv_encode_error(&result_one, is_temp,
8671 108 : *zeroed_buffer,
8672 108 : *ignored_checksum,
8673 108 : *buffer_invalid,
8674 108 : *zeroed_buffer ? 1 : 0,
8675 108 : *failed_checksum ? 1 : 0,
8676 : buf_off, buf_off, buf_off);
8677 108 : pgaio_result_report(result_one, td, LOG_SERVER_ONLY);
8678 : }
8679 : }
8680 :
8681 : /* Terminate I/O and set BM_VALID. */
8682 1516940 : set_flag_bits = failed ? BM_IO_ERROR : BM_VALID;
8683 1516940 : if (is_temp)
8684 11063 : TerminateLocalBufferIO(buf_hdr, false, set_flag_bits, true);
8685 : else
8686 1505877 : TerminateBufferIO(buf_hdr, false, set_flag_bits, false, true);
8687 :
8688 : /*
8689 : * Call the BUFFER_READ_DONE tracepoint in the callback, even though the
8690 : * callback may not be executed in the same backend that called
8691 : * BUFFER_READ_START. The alternative would be to defer calling the
8692 : * tracepoint to a later point (e.g. the local completion callback for
8693 : * shared buffer reads), which seems even less helpful.
8694 : */
8695 : TRACE_POSTGRESQL_BUFFER_READ_DONE(tag.forkNum,
8696 : tag.blockNum,
8697 : tag.spcOid,
8698 : tag.dbOid,
8699 : tag.relNumber,
8700 : is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
8701 : false);
8702 1516940 : }
8703 :
8704 : /*
8705 : * Perform completion handling of a single AIO read. This read may cover
8706 : * multiple blocks / buffers.
8707 : *
8708 : * Shared between shared and local buffers, to reduce code duplication.
8709 : */
8710 : static pg_attribute_always_inline PgAioResult
8711 1370825 : buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
8712 : uint8 cb_data, bool is_temp)
8713 : {
8714 1370825 : PgAioResult result = prior_result;
8715 1370825 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
8716 1370825 : uint8 first_error_off = 0;
8717 1370825 : uint8 first_zeroed_off = 0;
8718 1370825 : uint8 first_ignored_off = 0;
8719 1370825 : uint8 error_count = 0;
8720 1370825 : uint8 zeroed_count = 0;
8721 1370825 : uint8 ignored_count = 0;
8722 1370825 : uint8 checkfail_count = 0;
8723 : uint64 *io_data;
8724 : uint8 handle_data_len;
8725 :
8726 : if (is_temp)
8727 : {
8728 : Assert(td->smgr.is_temp);
8729 : Assert(pgaio_io_get_owner(ioh) == MyProcNumber);
8730 : }
8731 : else
8732 : Assert(!td->smgr.is_temp);
8733 :
8734 : /*
8735 : * Iterate over all the buffers affected by this IO and call the
8736 : * per-buffer completion function for each buffer.
8737 : */
8738 1370825 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
8739 2887765 : for (uint8 buf_off = 0; buf_off < handle_data_len; buf_off++)
8740 : {
8741 1516940 : Buffer buf = io_data[buf_off];
8742 : bool failed;
8743 1516940 : bool failed_verification = false;
8744 1516940 : bool failed_checksum = false;
8745 1516940 : bool zeroed_buffer = false;
8746 1516940 : bool ignored_checksum = false;
8747 :
8748 : Assert(BufferIsValid(buf));
8749 :
8750 : /*
8751 : * If the entire I/O failed on a lower-level, each buffer needs to be
8752 : * marked as failed. In case of a partial read, the first few buffers
8753 : * may be ok.
8754 : */
8755 1516940 : failed =
8756 1516940 : prior_result.status == PGAIO_RS_ERROR
8757 1516940 : || prior_result.result <= buf_off;
8758 :
8759 1516940 : buffer_readv_complete_one(td, buf_off, buf, cb_data, failed, is_temp,
8760 : &failed_verification,
8761 : &failed_checksum,
8762 : &ignored_checksum,
8763 : &zeroed_buffer);
8764 :
8765 : /*
8766 : * Track information about the number of different kinds of error
8767 : * conditions across all pages, as there can be multiple pages failing
8768 : * verification as part of one IO.
8769 : */
8770 1516940 : if (failed_verification && !zeroed_buffer && error_count++ == 0)
8771 44 : first_error_off = buf_off;
8772 1516940 : if (zeroed_buffer && zeroed_count++ == 0)
8773 34 : first_zeroed_off = buf_off;
8774 1516940 : if (ignored_checksum && ignored_count++ == 0)
8775 10 : first_ignored_off = buf_off;
8776 1516940 : if (failed_checksum)
8777 32 : checkfail_count++;
8778 : }
8779 :
8780 : /*
8781 : * If the smgr read succeeded [partially] and page verification failed for
8782 : * some of the pages, adjust the IO's result state appropriately.
8783 : */
8784 1370825 : if (prior_result.status != PGAIO_RS_ERROR &&
8785 1370770 : (error_count > 0 || ignored_count > 0 || zeroed_count > 0))
8786 : {
8787 84 : buffer_readv_encode_error(&result, is_temp,
8788 : zeroed_count > 0, ignored_count > 0,
8789 : error_count, zeroed_count, checkfail_count,
8790 : first_error_off, first_zeroed_off,
8791 : first_ignored_off);
8792 84 : pgaio_result_report(result, td, DEBUG1);
8793 : }
8794 :
8795 : /*
8796 : * For shared relations this reporting is done in
8797 : * shared_buffer_readv_complete_local().
8798 : */
8799 1370825 : if (is_temp && checkfail_count > 0)
8800 2 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
8801 : checkfail_count);
8802 :
8803 1370825 : return result;
8804 : }
8805 :
8806 : /*
8807 : * AIO error reporting callback for aio_shared_buffer_readv_cb and
8808 : * aio_local_buffer_readv_cb.
8809 : *
8810 : * The error is encoded / decoded in buffer_readv_encode_error() /
8811 : * buffer_readv_decode_error().
8812 : */
8813 : static void
8814 272 : buffer_readv_report(PgAioResult result, const PgAioTargetData *td,
8815 : int elevel)
8816 : {
8817 272 : int nblocks = td->smgr.nblocks;
8818 272 : BlockNumber first = td->smgr.blockNum;
8819 272 : BlockNumber last = first + nblocks - 1;
8820 272 : ProcNumber errProc =
8821 272 : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER;
8822 : RelPathStr rpath =
8823 272 : relpathbackend(td->smgr.rlocator, errProc, td->smgr.forkNum);
8824 : bool zeroed_any,
8825 : ignored_any;
8826 : uint8 zeroed_or_error_count,
8827 : checkfail_count,
8828 : first_off;
8829 : uint8 affected_count;
8830 : const char *msg_one,
8831 : *msg_mult,
8832 : *det_mult,
8833 : *hint_mult;
8834 :
8835 272 : buffer_readv_decode_error(result, &zeroed_any, &ignored_any,
8836 : &zeroed_or_error_count,
8837 : &checkfail_count,
8838 : &first_off);
8839 :
8840 : /*
8841 : * Treat a read that had both zeroed buffers *and* ignored checksums as a
8842 : * special case, it's too irregular to be emitted the same way as the
8843 : * other cases.
8844 : */
8845 272 : if (zeroed_any && ignored_any)
8846 : {
8847 : Assert(zeroed_any && ignored_any);
8848 : Assert(nblocks > 1); /* same block can't be both zeroed and ignored */
8849 : Assert(result.status != PGAIO_RS_ERROR);
8850 4 : affected_count = zeroed_or_error_count;
8851 :
8852 4 : ereport(elevel,
8853 : errcode(ERRCODE_DATA_CORRUPTED),
8854 : errmsg("zeroing %u page(s) and ignoring %u checksum failure(s) among blocks %u..%u of relation \"%s\"",
8855 : affected_count, checkfail_count, first, last, rpath.str),
8856 : affected_count > 1 ?
8857 : errdetail("Block %u held the first zeroed page.",
8858 : first + first_off) : 0,
8859 : errhint_plural("See server log for details about the other %d invalid block.",
8860 : "See server log for details about the other %d invalid blocks.",
8861 : affected_count + checkfail_count - 1,
8862 : affected_count + checkfail_count - 1));
8863 4 : return;
8864 : }
8865 :
8866 : /*
8867 : * The other messages are highly repetitive. To avoid duplicating a long
8868 : * and complicated ereport(), gather the translated format strings
8869 : * separately and then do one common ereport.
8870 : */
8871 268 : if (result.status == PGAIO_RS_ERROR)
8872 : {
8873 : Assert(!zeroed_any); /* can't have invalid pages when zeroing them */
8874 136 : affected_count = zeroed_or_error_count;
8875 136 : msg_one = _("invalid page in block %u of relation \"%s\"");
8876 136 : msg_mult = _("%u invalid pages among blocks %u..%u of relation \"%s\"");
8877 136 : det_mult = _("Block %u held the first invalid page.");
8878 136 : hint_mult = _("See server log for the other %u invalid block(s).");
8879 : }
8880 132 : else if (zeroed_any && !ignored_any)
8881 : {
8882 108 : affected_count = zeroed_or_error_count;
8883 108 : msg_one = _("invalid page in block %u of relation \"%s\"; zeroing out page");
8884 108 : msg_mult = _("zeroing out %u invalid pages among blocks %u..%u of relation \"%s\"");
8885 108 : det_mult = _("Block %u held the first zeroed page.");
8886 108 : hint_mult = _("See server log for the other %u zeroed block(s).");
8887 : }
8888 24 : else if (!zeroed_any && ignored_any)
8889 : {
8890 24 : affected_count = checkfail_count;
8891 24 : msg_one = _("ignoring checksum failure in block %u of relation \"%s\"");
8892 24 : msg_mult = _("ignoring %u checksum failures among blocks %u..%u of relation \"%s\"");
8893 24 : det_mult = _("Block %u held the first ignored page.");
8894 24 : hint_mult = _("See server log for the other %u ignored block(s).");
8895 : }
8896 : else
8897 0 : pg_unreachable();
8898 :
8899 268 : ereport(elevel,
8900 : errcode(ERRCODE_DATA_CORRUPTED),
8901 : affected_count == 1 ?
8902 : errmsg_internal(msg_one, first + first_off, rpath.str) :
8903 : errmsg_internal(msg_mult, affected_count, first, last, rpath.str),
8904 : affected_count > 1 ? errdetail_internal(det_mult, first + first_off) : 0,
8905 : affected_count > 1 ? errhint_internal(hint_mult, affected_count - 1) : 0);
8906 : }
8907 :
8908 : static void
8909 1503851 : shared_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
8910 : {
8911 1503851 : buffer_stage_common(ioh, false, false);
8912 1503851 : }
8913 :
8914 : static PgAioResult
8915 1367978 : shared_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
8916 : uint8 cb_data)
8917 : {
8918 1367978 : return buffer_readv_complete(ioh, prior_result, cb_data, false);
8919 : }
8920 :
8921 : /*
8922 : * We need a backend-local completion callback for shared buffers, to be able
8923 : * to report checksum errors correctly. Unfortunately that can only safely
8924 : * happen if the reporting backend has previously called
8925 : * pgstat_prepare_report_checksum_failure(), which we can only guarantee in
8926 : * the backend that started the IO. Hence this callback.
8927 : */
8928 : static PgAioResult
8929 1503851 : shared_buffer_readv_complete_local(PgAioHandle *ioh, PgAioResult prior_result,
8930 : uint8 cb_data)
8931 : {
8932 : bool zeroed_any,
8933 : ignored_any;
8934 : uint8 zeroed_or_error_count,
8935 : checkfail_count,
8936 : first_off;
8937 :
8938 1503851 : if (prior_result.status == PGAIO_RS_OK)
8939 1503673 : return prior_result;
8940 :
8941 178 : buffer_readv_decode_error(prior_result,
8942 : &zeroed_any,
8943 : &ignored_any,
8944 : &zeroed_or_error_count,
8945 : &checkfail_count,
8946 : &first_off);
8947 :
8948 178 : if (checkfail_count)
8949 : {
8950 24 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
8951 :
8952 24 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
8953 : checkfail_count);
8954 : }
8955 :
8956 178 : return prior_result;
8957 : }
8958 :
8959 : static void
8960 2847 : local_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
8961 : {
8962 2847 : buffer_stage_common(ioh, false, true);
8963 2847 : }
8964 :
8965 : static PgAioResult
8966 2847 : local_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
8967 : uint8 cb_data)
8968 : {
8969 2847 : return buffer_readv_complete(ioh, prior_result, cb_data, true);
8970 : }
8971 :
8972 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
8973 : const PgAioHandleCallbacks aio_shared_buffer_readv_cb = {
8974 : .stage = shared_buffer_readv_stage,
8975 : .complete_shared = shared_buffer_readv_complete,
8976 : /* need a local callback to report checksum failures */
8977 : .complete_local = shared_buffer_readv_complete_local,
8978 : .report = buffer_readv_report,
8979 : };
8980 :
8981 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
8982 : const PgAioHandleCallbacks aio_local_buffer_readv_cb = {
8983 : .stage = local_buffer_readv_stage,
8984 :
8985 : /*
8986 : * Note that this, in contrast to the shared_buffers case, uses
8987 : * complete_local, as only the issuing backend has access to the required
8988 : * datastructures. This is important in case the IO completion may be
8989 : * consumed incidentally by another backend.
8990 : */
8991 : .complete_local = local_buffer_readv_complete,
8992 : .report = buffer_readv_report,
8993 : };
|
