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 88253886 : ReservePrivateRefCountEntry(void)
310 : {
311 : /* Already reserved (or freed), nothing to do */
312 88253886 : if (ReservedRefCountSlot != -1)
313 83086147 : 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 46509651 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
323 : {
324 41341912 : if (PrivateRefCountArrayKeys[i] == InvalidBuffer)
325 : {
326 30115762 : 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 5167739 : if (ReservedRefCountSlot != -1)
336 4968412 : 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 199327 : victim_slot = PrivateRefCountClock++ % REFCOUNT_ARRAY_ENTRIES;
355 199327 : victim_entry = &PrivateRefCountArray[victim_slot];
356 199327 : 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 199327 : 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 199327 : hashent->data = victim_entry->data;
370 :
371 : /* clear the now free array slot */
372 199327 : PrivateRefCountArrayKeys[victim_slot] = InvalidBuffer;
373 199327 : victim_entry->buffer = InvalidBuffer;
374 :
375 : /* clear the whole data member, just for future proofing */
376 199327 : memset(&victim_entry->data, 0, sizeof(victim_entry->data));
377 199327 : victim_entry->data.refcount = 0;
378 199327 : victim_entry->data.lockmode = BUFFER_LOCK_UNLOCK;
379 :
380 199327 : PrivateRefCountOverflowed++;
381 : }
382 : }
383 :
384 : /*
385 : * Fill a previously reserved refcount entry.
386 : */
387 : static PrivateRefCountEntry *
388 76449728 : 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 76449728 : res = &PrivateRefCountArray[ReservedRefCountSlot];
397 :
398 : /* and fill it */
399 76449728 : PrivateRefCountArrayKeys[ReservedRefCountSlot] = buffer;
400 76449728 : res->buffer = buffer;
401 76449728 : res->data.refcount = 0;
402 76449728 : res->data.lockmode = BUFFER_LOCK_UNLOCK;
403 :
404 : /* update cache for the next lookup */
405 76449728 : PrivateRefCountEntryLast = ReservedRefCountSlot;
406 :
407 76449728 : ReservedRefCountSlot = -1;
408 :
409 76449728 : 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 106712626 : GetPrivateRefCountEntrySlow(Buffer buffer, bool do_move)
420 : {
421 : PrivateRefCountEntry *res;
422 106712626 : 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 960413634 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
430 : {
431 853701008 : if (PrivateRefCountArrayKeys[i] == buffer)
432 : {
433 32622567 : match = i;
434 : /* see ReservePrivateRefCountEntry() for why we don't return */
435 : }
436 : }
437 :
438 106712626 : if (likely(match != -1))
439 : {
440 : /* update cache for the next lookup */
441 32622567 : PrivateRefCountEntryLast = match;
442 :
443 32622567 : 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 74090059 : if (PrivateRefCountOverflowed == 0)
454 73419163 : return NULL;
455 :
456 670896 : res = refcount_lookup(PrivateRefCountHash, buffer);
457 :
458 670896 : if (res == NULL)
459 416964 : return NULL;
460 253932 : else if (!do_move)
461 : {
462 : /* caller doesn't want us to move the hash entry into the array */
463 152898 : 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 101034 : data = res->data;
473 101034 : refcount_delete_item(PrivateRefCountHash, res);
474 : Assert(PrivateRefCountOverflowed > 0);
475 101034 : PrivateRefCountOverflowed--;
476 :
477 : /* Ensure there's a free array slot */
478 101034 : ReservePrivateRefCountEntry();
479 :
480 : /* Use up the reserved slot */
481 : Assert(ReservedRefCountSlot != -1);
482 101034 : free = &PrivateRefCountArray[ReservedRefCountSlot];
483 : Assert(PrivateRefCountArrayKeys[ReservedRefCountSlot] == free->buffer);
484 : Assert(free->buffer == InvalidBuffer);
485 :
486 : /* and fill it */
487 101034 : free->buffer = buffer;
488 101034 : free->data = data;
489 101034 : PrivateRefCountArrayKeys[ReservedRefCountSlot] = buffer;
490 : /* update cache for the next lookup */
491 101034 : PrivateRefCountEntryLast = ReservedRefCountSlot;
492 :
493 101034 : ReservedRefCountSlot = -1;
494 :
495 101034 : 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 439347174 : 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 439347174 : if (likely(PrivateRefCountEntryLast != -1) &&
523 439331415 : likely(PrivateRefCountArray[PrivateRefCountEntryLast].buffer == buffer))
524 : {
525 332634548 : 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 106712626 : 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 3258696 : 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 3258696 : ref = GetPrivateRefCountEntry(buffer, false);
554 :
555 3258696 : if (ref == NULL)
556 20801 : return 0;
557 3237895 : 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 76449728 : ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref)
566 : {
567 : Assert(ref->data.refcount == 0);
568 : Assert(ref->data.lockmode == BUFFER_LOCK_UNLOCK);
569 :
570 76449728 : if (ref >= &PrivateRefCountArray[0] &&
571 : ref < &PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES])
572 : {
573 76351435 : ref->buffer = InvalidBuffer;
574 76351435 : 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 76351435 : ReservedRefCountSlot = ref - PrivateRefCountArray;
583 : }
584 : else
585 : {
586 98293 : refcount_delete_item(PrivateRefCountHash, ref);
587 : Assert(PrivateRefCountOverflowed > 0);
588 98293 : PrivateRefCountOverflowed--;
589 : }
590 76449728 : }
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 unlocked);
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 39483 : PrefetchSharedBuffer(SMgrRelation smgr_reln,
698 : ForkNumber forkNum,
699 : BlockNumber blockNum)
700 : {
701 39483 : 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 39483 : InitBufferTag(&newTag, &smgr_reln->smgr_rlocator.locator,
711 : forkNum, blockNum);
712 :
713 : /* determine its hash code and partition lock ID */
714 39483 : newHash = BufTableHashCode(&newTag);
715 39483 : newPartitionLock = BufMappingPartitionLock(newHash);
716 :
717 : /* see if the block is in the buffer pool already */
718 39483 : LWLockAcquire(newPartitionLock, LW_SHARED);
719 39483 : buf_id = BufTableLookup(&newTag, newHash);
720 39483 : LWLockRelease(newPartitionLock);
721 :
722 : /* If not in buffers, initiate prefetch */
723 39483 : 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 18893 : if ((io_direct_flags & IO_DIRECT_DATA) == 0 &&
731 9320 : smgrprefetch(smgr_reln, forkNum, blockNum, 1))
732 : {
733 9320 : 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 29910 : 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 39483 : 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 28522 : PrefetchBuffer(Relation reln, ForkNumber forkNum, BlockNumber blockNum)
788 : {
789 : Assert(RelationIsValid(reln));
790 : Assert(BlockNumberIsValid(blockNum));
791 :
792 28522 : if (RelationUsesLocalBuffers(reln))
793 : {
794 : /* see comments in ReadBufferExtended */
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 27165 : 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 5124 : 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 5124 : ResourceOwnerEnlarge(CurrentResourceOwner);
828 5124 : ReservePrivateRefCountEntry();
829 5124 : InitBufferTag(&tag, &rlocator, forkNum, blockNum);
830 :
831 5124 : 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 4984 : 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 9935 : if (BufferTagsEqual(&tag, &bufHdr->tag) &&
860 4951 : PinBuffer(bufHdr, NULL, true))
861 : {
862 4945 : if (BufferTagsEqual(&tag, &bufHdr->tag))
863 : {
864 4945 : pgBufferUsage.shared_blks_hit++;
865 4945 : return true;
866 : }
867 0 : UnpinBuffer(bufHdr);
868 : }
869 : }
870 :
871 39 : 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 62452821 : ReadBuffer(Relation reln, BlockNumber blockNum)
880 : {
881 62452821 : 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 73676180 : ReadBufferExtended(Relation reln, ForkNumber forkNum, BlockNumber blockNum,
927 : ReadBufferMode mode, BufferAccessStrategy strategy)
928 : {
929 : Buffer buf;
930 :
931 : /*
932 : * Reject attempts to read non-local temporary relations; we would be
933 : * likely to get wrong data since we have no visibility into the owning
934 : * session's local buffers.
935 : */
936 73676180 : if (RELATION_IS_OTHER_TEMP(reln))
937 0 : ereport(ERROR,
938 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
939 : errmsg("cannot access temporary tables of other sessions")));
940 :
941 : /*
942 : * Read the buffer, and update pgstat counters to reflect a cache hit or
943 : * miss.
944 : */
945 73676180 : buf = ReadBuffer_common(reln, RelationGetSmgr(reln), 0,
946 : forkNum, blockNum, mode, strategy);
947 :
948 73676157 : return buf;
949 : }
950 :
951 :
952 : /*
953 : * ReadBufferWithoutRelcache -- like ReadBufferExtended, but doesn't require
954 : * a relcache entry for the relation.
955 : *
956 : * Pass permanent = true for a RELPERSISTENCE_PERMANENT relation, and
957 : * permanent = false for a RELPERSISTENCE_UNLOGGED relation. This function
958 : * cannot be used for temporary relations (and making that work might be
959 : * difficult, unless we only want to read temporary relations for our own
960 : * ProcNumber).
961 : */
962 : Buffer
963 5954054 : ReadBufferWithoutRelcache(RelFileLocator rlocator, ForkNumber forkNum,
964 : BlockNumber blockNum, ReadBufferMode mode,
965 : BufferAccessStrategy strategy, bool permanent)
966 : {
967 5954054 : SMgrRelation smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
968 :
969 5954054 : return ReadBuffer_common(NULL, smgr,
970 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
971 : forkNum, blockNum,
972 : mode, strategy);
973 : }
974 :
975 : /*
976 : * Convenience wrapper around ExtendBufferedRelBy() extending by one block.
977 : */
978 : Buffer
979 56749 : ExtendBufferedRel(BufferManagerRelation bmr,
980 : ForkNumber forkNum,
981 : BufferAccessStrategy strategy,
982 : uint32 flags)
983 : {
984 : Buffer buf;
985 56749 : uint32 extend_by = 1;
986 :
987 56749 : ExtendBufferedRelBy(bmr, forkNum, strategy, flags, extend_by,
988 : &buf, &extend_by);
989 :
990 56749 : return buf;
991 : }
992 :
993 : /*
994 : * Extend relation by multiple blocks.
995 : *
996 : * Tries to extend the relation by extend_by blocks. Depending on the
997 : * availability of resources the relation may end up being extended by a
998 : * smaller number of pages (unless an error is thrown, always by at least one
999 : * page). *extended_by is updated to the number of pages the relation has been
1000 : * extended to.
1001 : *
1002 : * buffers needs to be an array that is at least extend_by long. Upon
1003 : * completion, the first extend_by array elements will point to a pinned
1004 : * buffer.
1005 : *
1006 : * If EB_LOCK_FIRST is part of flags, the first returned buffer is
1007 : * locked. This is useful for callers that want a buffer that is guaranteed to
1008 : * be empty.
1009 : */
1010 : BlockNumber
1011 210121 : ExtendBufferedRelBy(BufferManagerRelation bmr,
1012 : ForkNumber fork,
1013 : BufferAccessStrategy strategy,
1014 : uint32 flags,
1015 : uint32 extend_by,
1016 : Buffer *buffers,
1017 : uint32 *extended_by)
1018 : {
1019 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
1020 : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
1021 : Assert(extend_by > 0);
1022 :
1023 210121 : if (bmr.relpersistence == '\0')
1024 210121 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
1025 :
1026 210121 : return ExtendBufferedRelCommon(bmr, fork, strategy, flags,
1027 : extend_by, InvalidBlockNumber,
1028 : buffers, extended_by);
1029 : }
1030 :
1031 : /*
1032 : * Extend the relation so it is at least extend_to blocks large, return buffer
1033 : * (extend_to - 1).
1034 : *
1035 : * This is useful for callers that want to write a specific page, regardless
1036 : * of the current size of the relation (e.g. useful for visibilitymap and for
1037 : * crash recovery).
1038 : */
1039 : Buffer
1040 57587 : ExtendBufferedRelTo(BufferManagerRelation bmr,
1041 : ForkNumber fork,
1042 : BufferAccessStrategy strategy,
1043 : uint32 flags,
1044 : BlockNumber extend_to,
1045 : ReadBufferMode mode)
1046 : {
1047 : BlockNumber current_size;
1048 57587 : uint32 extended_by = 0;
1049 57587 : Buffer buffer = InvalidBuffer;
1050 : Buffer buffers[64];
1051 :
1052 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
1053 : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
1054 : Assert(extend_to != InvalidBlockNumber && extend_to > 0);
1055 :
1056 57587 : if (bmr.relpersistence == '\0')
1057 9400 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
1058 :
1059 : /*
1060 : * If desired, create the file if it doesn't exist. If
1061 : * smgr_cached_nblocks[fork] is positive then it must exist, no need for
1062 : * an smgrexists call.
1063 : */
1064 57587 : if ((flags & EB_CREATE_FORK_IF_NEEDED) &&
1065 9400 : (BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == 0 ||
1066 28 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == InvalidBlockNumber) &&
1067 9372 : !smgrexists(BMR_GET_SMGR(bmr), fork))
1068 : {
1069 9342 : LockRelationForExtension(bmr.rel, ExclusiveLock);
1070 :
1071 : /* recheck, fork might have been created concurrently */
1072 9342 : if (!smgrexists(BMR_GET_SMGR(bmr), fork))
1073 9338 : smgrcreate(BMR_GET_SMGR(bmr), fork, flags & EB_PERFORMING_RECOVERY);
1074 :
1075 9342 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
1076 : }
1077 :
1078 : /*
1079 : * If requested, invalidate size cache, so that smgrnblocks asks the
1080 : * kernel.
1081 : */
1082 57587 : if (flags & EB_CLEAR_SIZE_CACHE)
1083 9400 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
1084 :
1085 : /*
1086 : * Estimate how many pages we'll need to extend by. This avoids acquiring
1087 : * unnecessarily many victim buffers.
1088 : */
1089 57587 : current_size = smgrnblocks(BMR_GET_SMGR(bmr), fork);
1090 :
1091 : /*
1092 : * Since no-one else can be looking at the page contents yet, there is no
1093 : * difference between an exclusive lock and a cleanup-strength lock. Note
1094 : * that we pass the original mode to ReadBuffer_common() below, when
1095 : * falling back to reading the buffer to a concurrent relation extension.
1096 : */
1097 57587 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1098 47794 : flags |= EB_LOCK_TARGET;
1099 :
1100 117355 : while (current_size < extend_to)
1101 : {
1102 59768 : uint32 num_pages = lengthof(buffers);
1103 : BlockNumber first_block;
1104 :
1105 59768 : if ((uint64) current_size + num_pages > extend_to)
1106 59702 : num_pages = extend_to - current_size;
1107 :
1108 59768 : first_block = ExtendBufferedRelCommon(bmr, fork, strategy, flags,
1109 : num_pages, extend_to,
1110 : buffers, &extended_by);
1111 :
1112 59768 : current_size = first_block + extended_by;
1113 : Assert(num_pages != 0 || current_size >= extend_to);
1114 :
1115 128264 : for (uint32 i = 0; i < extended_by; i++)
1116 : {
1117 68496 : if (first_block + i != extend_to - 1)
1118 10917 : ReleaseBuffer(buffers[i]);
1119 : else
1120 57579 : buffer = buffers[i];
1121 : }
1122 : }
1123 :
1124 : /*
1125 : * It's possible that another backend concurrently extended the relation.
1126 : * In that case read the buffer.
1127 : *
1128 : * XXX: Should we control this via a flag?
1129 : */
1130 57587 : if (buffer == InvalidBuffer)
1131 : {
1132 : Assert(extended_by == 0);
1133 8 : buffer = ReadBuffer_common(bmr.rel, BMR_GET_SMGR(bmr), bmr.relpersistence,
1134 : fork, extend_to - 1, mode, strategy);
1135 : }
1136 :
1137 57587 : return buffer;
1138 : }
1139 :
1140 : /*
1141 : * Lock and optionally zero a buffer, as part of the implementation of
1142 : * RBM_ZERO_AND_LOCK or RBM_ZERO_AND_CLEANUP_LOCK. The buffer must be already
1143 : * pinned. If the buffer is not already valid, it is zeroed and made valid.
1144 : */
1145 : static void
1146 361122 : ZeroAndLockBuffer(Buffer buffer, ReadBufferMode mode, bool already_valid)
1147 : {
1148 : BufferDesc *bufHdr;
1149 : bool need_to_zero;
1150 361122 : bool isLocalBuf = BufferIsLocal(buffer);
1151 : StartBufferIOResult sbres;
1152 :
1153 : Assert(mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK);
1154 :
1155 361122 : if (already_valid)
1156 : {
1157 : /*
1158 : * If the caller already knew the buffer was valid, we can skip some
1159 : * header interaction. The caller just wants to lock the buffer.
1160 : */
1161 38232 : need_to_zero = false;
1162 : }
1163 : else
1164 : {
1165 322890 : if (isLocalBuf)
1166 : {
1167 : /* Simple case for non-shared buffers. */
1168 30 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
1169 30 : sbres = StartLocalBufferIO(bufHdr, true, true, NULL);
1170 : }
1171 : else
1172 : {
1173 : /*
1174 : * Take BM_IO_IN_PROGRESS, or discover that BM_VALID has been set
1175 : * concurrently. Even though we aren't doing I/O, that ensures
1176 : * that we don't zero a page that someone else has pinned. An
1177 : * exclusive content lock wouldn't be enough, because readers are
1178 : * allowed to drop the content lock after determining that a tuple
1179 : * is visible (see buffer access rules in README).
1180 : */
1181 322860 : bufHdr = GetBufferDescriptor(buffer - 1);
1182 322860 : sbres = StartSharedBufferIO(bufHdr, true, true, NULL);
1183 : }
1184 :
1185 : Assert(sbres != BUFFER_IO_IN_PROGRESS);
1186 322890 : need_to_zero = sbres == BUFFER_IO_READY_FOR_IO;
1187 : }
1188 :
1189 361122 : if (need_to_zero)
1190 : {
1191 322890 : memset(BufferGetPage(buffer), 0, BLCKSZ);
1192 :
1193 : /*
1194 : * Grab the buffer content lock before marking the page as valid, to
1195 : * make sure that no other backend sees the zeroed page before the
1196 : * caller has had a chance to initialize it.
1197 : *
1198 : * Since no-one else can be looking at the page contents yet, there is
1199 : * no difference between an exclusive lock and a cleanup-strength
1200 : * lock. (Note that we cannot use LockBuffer() or
1201 : * LockBufferForCleanup() here, because they assert that the buffer is
1202 : * already valid.)
1203 : */
1204 322890 : if (!isLocalBuf)
1205 322860 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1206 :
1207 : /* Set BM_VALID, terminate IO, and wake up any waiters */
1208 322890 : if (isLocalBuf)
1209 30 : TerminateLocalBufferIO(bufHdr, false, BM_VALID, false);
1210 : else
1211 322860 : TerminateBufferIO(bufHdr, false, BM_VALID, true, false);
1212 : }
1213 38232 : else if (!isLocalBuf)
1214 : {
1215 : /*
1216 : * The buffer is valid, so we can't zero it. The caller still expects
1217 : * the page to be locked on return.
1218 : */
1219 38212 : if (mode == RBM_ZERO_AND_LOCK)
1220 38176 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1221 : else
1222 36 : LockBufferForCleanup(buffer);
1223 : }
1224 361122 : }
1225 :
1226 : /*
1227 : * Pin a buffer for a given block. *foundPtr is set to true if the block was
1228 : * already present, or false if more work is required to either read it in or
1229 : * zero it.
1230 : */
1231 : static pg_attribute_always_inline Buffer
1232 84756079 : PinBufferForBlock(Relation rel,
1233 : SMgrRelation smgr,
1234 : char persistence,
1235 : ForkNumber forkNum,
1236 : BlockNumber blockNum,
1237 : BufferAccessStrategy strategy,
1238 : IOObject io_object,
1239 : IOContext io_context,
1240 : bool *foundPtr)
1241 : {
1242 : BufferDesc *bufHdr;
1243 :
1244 : Assert(blockNum != P_NEW);
1245 :
1246 : /* Persistence should be set before */
1247 : Assert((persistence == RELPERSISTENCE_TEMP ||
1248 : persistence == RELPERSISTENCE_PERMANENT ||
1249 : persistence == RELPERSISTENCE_UNLOGGED));
1250 :
1251 : TRACE_POSTGRESQL_BUFFER_READ_START(forkNum, blockNum,
1252 : smgr->smgr_rlocator.locator.spcOid,
1253 : smgr->smgr_rlocator.locator.dbOid,
1254 : smgr->smgr_rlocator.locator.relNumber,
1255 : smgr->smgr_rlocator.backend);
1256 :
1257 84756079 : if (persistence == RELPERSISTENCE_TEMP)
1258 1648452 : bufHdr = LocalBufferAlloc(smgr, forkNum, blockNum, foundPtr);
1259 : else
1260 83107627 : bufHdr = BufferAlloc(smgr, persistence, forkNum, blockNum,
1261 : strategy, foundPtr, io_context);
1262 :
1263 84756071 : if (*foundPtr)
1264 82764213 : TrackBufferHit(io_object, io_context, rel, persistence, smgr, forkNum, blockNum);
1265 :
1266 84756071 : if (rel)
1267 : {
1268 : /*
1269 : * While pgBufferUsage's "read" counter isn't bumped unless we reach
1270 : * WaitReadBuffers() (so, not for hits, and not for buffers that are
1271 : * zeroed instead), the per-relation stats always count them.
1272 : */
1273 78533993 : pgstat_count_buffer_read(rel);
1274 : }
1275 :
1276 84756071 : return BufferDescriptorGetBuffer(bufHdr);
1277 : }
1278 :
1279 : /*
1280 : * ReadBuffer_common -- common logic for all ReadBuffer variants
1281 : *
1282 : * smgr is required, rel is optional unless using P_NEW.
1283 : */
1284 : static pg_attribute_always_inline Buffer
1285 79630242 : ReadBuffer_common(Relation rel, SMgrRelation smgr, char smgr_persistence,
1286 : ForkNumber forkNum,
1287 : BlockNumber blockNum, ReadBufferMode mode,
1288 : BufferAccessStrategy strategy)
1289 : {
1290 : ReadBuffersOperation operation;
1291 : Buffer buffer;
1292 : int flags;
1293 : char persistence;
1294 :
1295 : /*
1296 : * Backward compatibility path, most code should use ExtendBufferedRel()
1297 : * instead, as acquiring the extension lock inside ExtendBufferedRel()
1298 : * scales a lot better.
1299 : */
1300 79630242 : if (unlikely(blockNum == P_NEW))
1301 : {
1302 322 : uint32 flags = EB_SKIP_EXTENSION_LOCK;
1303 :
1304 : /*
1305 : * Since no-one else can be looking at the page contents yet, there is
1306 : * no difference between an exclusive lock and a cleanup-strength
1307 : * lock.
1308 : */
1309 322 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1310 0 : flags |= EB_LOCK_FIRST;
1311 :
1312 322 : return ExtendBufferedRel(BMR_REL(rel), forkNum, strategy, flags);
1313 : }
1314 :
1315 79629920 : if (rel)
1316 73675866 : persistence = rel->rd_rel->relpersistence;
1317 : else
1318 5954054 : persistence = smgr_persistence;
1319 :
1320 79629920 : if (unlikely(mode == RBM_ZERO_AND_CLEANUP_LOCK ||
1321 : mode == RBM_ZERO_AND_LOCK))
1322 : {
1323 : bool found;
1324 : IOContext io_context;
1325 : IOObject io_object;
1326 :
1327 361122 : if (persistence == RELPERSISTENCE_TEMP)
1328 : {
1329 50 : io_context = IOCONTEXT_NORMAL;
1330 50 : io_object = IOOBJECT_TEMP_RELATION;
1331 : }
1332 : else
1333 : {
1334 361072 : io_context = IOContextForStrategy(strategy);
1335 361072 : io_object = IOOBJECT_RELATION;
1336 : }
1337 :
1338 361122 : buffer = PinBufferForBlock(rel, smgr, persistence,
1339 : forkNum, blockNum, strategy,
1340 : io_object, io_context, &found);
1341 361122 : ZeroAndLockBuffer(buffer, mode, found);
1342 361122 : return buffer;
1343 : }
1344 :
1345 : /*
1346 : * Signal that we are going to immediately wait. If we're immediately
1347 : * waiting, there is no benefit in actually executing the IO
1348 : * asynchronously, it would just add dispatch overhead.
1349 : */
1350 79268798 : flags = READ_BUFFERS_SYNCHRONOUSLY;
1351 79268798 : if (mode == RBM_ZERO_ON_ERROR)
1352 1906675 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1353 79268798 : operation.smgr = smgr;
1354 79268798 : operation.rel = rel;
1355 79268798 : operation.persistence = persistence;
1356 79268798 : operation.forknum = forkNum;
1357 79268798 : operation.strategy = strategy;
1358 79268798 : if (StartReadBuffer(&operation,
1359 : &buffer,
1360 : blockNum,
1361 : flags))
1362 775630 : WaitReadBuffers(&operation);
1363 :
1364 79268775 : return buffer;
1365 : }
1366 :
1367 : static pg_attribute_always_inline bool
1368 84207276 : StartReadBuffersImpl(ReadBuffersOperation *operation,
1369 : Buffer *buffers,
1370 : BlockNumber blockNum,
1371 : int *nblocks,
1372 : int flags,
1373 : bool allow_forwarding)
1374 : {
1375 84207276 : int actual_nblocks = *nblocks;
1376 84207276 : int maxcombine = 0;
1377 : bool did_start_io;
1378 : IOContext io_context;
1379 : IOObject io_object;
1380 :
1381 : Assert(*nblocks == 1 || allow_forwarding);
1382 : Assert(*nblocks > 0);
1383 : Assert(*nblocks <= MAX_IO_COMBINE_LIMIT);
1384 :
1385 84207276 : if (operation->persistence == RELPERSISTENCE_TEMP)
1386 : {
1387 1640206 : io_context = IOCONTEXT_NORMAL;
1388 1640206 : io_object = IOOBJECT_TEMP_RELATION;
1389 : }
1390 : else
1391 : {
1392 82567070 : io_context = IOContextForStrategy(operation->strategy);
1393 82567070 : io_object = IOOBJECT_RELATION;
1394 : }
1395 :
1396 85876258 : for (int i = 0; i < actual_nblocks; ++i)
1397 : {
1398 : bool found;
1399 :
1400 84397230 : if (allow_forwarding && buffers[i] != InvalidBuffer)
1401 2273 : {
1402 : BufferDesc *bufHdr;
1403 :
1404 : /*
1405 : * This is a buffer that was pinned by an earlier call to
1406 : * StartReadBuffers(), but couldn't be handled in one operation at
1407 : * that time. The operation was split, and the caller has passed
1408 : * an already pinned buffer back to us to handle the rest of the
1409 : * operation. It must continue at the expected block number.
1410 : */
1411 : Assert(BufferGetBlockNumber(buffers[i]) == blockNum + i);
1412 :
1413 : /*
1414 : * It might be an already valid buffer (a hit) that followed the
1415 : * final contiguous block of an earlier I/O (a miss) marking the
1416 : * end of it, or a buffer that some other backend has since made
1417 : * valid by performing the I/O for us, in which case we can handle
1418 : * it as a hit now. It is safe to check for a BM_VALID flag with
1419 : * a relaxed load, because we got a fresh view of it while pinning
1420 : * it in the previous call.
1421 : *
1422 : * On the other hand if we don't see BM_VALID yet, it must be an
1423 : * I/O that was split by the previous call and we need to try to
1424 : * start a new I/O from this block. We're also racing against any
1425 : * other backend that might start the I/O or even manage to mark
1426 : * it BM_VALID after this check, but StartBufferIO() will handle
1427 : * those cases.
1428 : */
1429 2273 : if (BufferIsLocal(buffers[i]))
1430 16 : bufHdr = GetLocalBufferDescriptor(-buffers[i] - 1);
1431 : else
1432 2257 : bufHdr = GetBufferDescriptor(buffers[i] - 1);
1433 : Assert(pg_atomic_read_u64(&bufHdr->state) & BM_TAG_VALID);
1434 2273 : found = pg_atomic_read_u64(&bufHdr->state) & BM_VALID;
1435 : }
1436 : else
1437 : {
1438 84394949 : buffers[i] = PinBufferForBlock(operation->rel,
1439 : operation->smgr,
1440 84394957 : operation->persistence,
1441 : operation->forknum,
1442 : blockNum + i,
1443 : operation->strategy,
1444 : io_object, io_context,
1445 : &found);
1446 : }
1447 :
1448 84397222 : if (found)
1449 : {
1450 : /*
1451 : * We have a hit. If it's the first block in the requested range,
1452 : * we can return it immediately and report that WaitReadBuffers()
1453 : * does not need to be called. If the initial value of *nblocks
1454 : * was larger, the caller will have to call again for the rest.
1455 : */
1456 82728240 : if (i == 0)
1457 : {
1458 82725979 : *nblocks = 1;
1459 :
1460 : #ifdef USE_ASSERT_CHECKING
1461 :
1462 : /*
1463 : * Initialize enough of ReadBuffersOperation to make
1464 : * CheckReadBuffersOperation() work. Outside of assertions
1465 : * that's not necessary when no IO is issued.
1466 : */
1467 : operation->buffers = buffers;
1468 : operation->blocknum = blockNum;
1469 : operation->nblocks = 1;
1470 : operation->nblocks_done = 1;
1471 : CheckReadBuffersOperation(operation, true);
1472 : #endif
1473 82725979 : return false;
1474 : }
1475 :
1476 : /*
1477 : * Otherwise we already have an I/O to perform, but this block
1478 : * can't be included as it is already valid. Split the I/O here.
1479 : * There may or may not be more blocks requiring I/O after this
1480 : * one, we haven't checked, but they can't be contiguous with this
1481 : * one in the way. We'll leave this buffer pinned, forwarding it
1482 : * to the next call, avoiding the need to unpin it here and re-pin
1483 : * it in the next call.
1484 : */
1485 2261 : actual_nblocks = i;
1486 2261 : break;
1487 : }
1488 : else
1489 : {
1490 : /*
1491 : * Check how many blocks we can cover with the same IO. The smgr
1492 : * implementation might e.g. be limited due to a segment boundary.
1493 : */
1494 1668982 : if (i == 0 && actual_nblocks > 1)
1495 : {
1496 38450 : maxcombine = smgrmaxcombine(operation->smgr,
1497 : operation->forknum,
1498 : blockNum);
1499 38450 : if (unlikely(maxcombine < actual_nblocks))
1500 : {
1501 0 : elog(DEBUG2, "limiting nblocks at %u from %u to %u",
1502 : blockNum, actual_nblocks, maxcombine);
1503 0 : actual_nblocks = maxcombine;
1504 : }
1505 : }
1506 : }
1507 : }
1508 1481289 : *nblocks = actual_nblocks;
1509 :
1510 : /* Populate information needed for I/O. */
1511 1481289 : operation->buffers = buffers;
1512 1481289 : operation->blocknum = blockNum;
1513 1481289 : operation->flags = flags;
1514 1481289 : operation->nblocks = actual_nblocks;
1515 1481289 : operation->nblocks_done = 0;
1516 1481289 : pgaio_wref_clear(&operation->io_wref);
1517 :
1518 : /*
1519 : * When using AIO, start the IO in the background. If not, issue prefetch
1520 : * requests if desired by the caller.
1521 : *
1522 : * The reason we have a dedicated path for IOMETHOD_SYNC here is to
1523 : * de-risk the introduction of AIO somewhat. It's a large architectural
1524 : * change, with lots of chances for unanticipated performance effects.
1525 : *
1526 : * Use of IOMETHOD_SYNC already leads to not actually performing IO
1527 : * asynchronously, but without the check here we'd execute IO earlier than
1528 : * we used to. Eventually this IOMETHOD_SYNC specific path should go away.
1529 : */
1530 1481289 : if (io_method != IOMETHOD_SYNC)
1531 : {
1532 : /*
1533 : * Try to start IO asynchronously. It's possible that no IO needs to
1534 : * be started, if another backend already performed the IO.
1535 : *
1536 : * Note that if an IO is started, it might not cover the entire
1537 : * requested range, e.g. because an intermediary block has been read
1538 : * in by another backend. In that case any "trailing" buffers we
1539 : * already pinned above will be "forwarded" by read_stream.c to the
1540 : * next call to StartReadBuffers().
1541 : *
1542 : * This is signalled to the caller by decrementing *nblocks *and*
1543 : * reducing operation->nblocks. The latter is done here, but not below
1544 : * WaitReadBuffers(), as in WaitReadBuffers() we can't "shorten" the
1545 : * overall read size anymore, we need to retry until done in its
1546 : * entirety or until failed.
1547 : */
1548 1479915 : did_start_io = AsyncReadBuffers(operation, nblocks);
1549 :
1550 1479900 : operation->nblocks = *nblocks;
1551 : }
1552 : else
1553 : {
1554 1374 : operation->flags |= READ_BUFFERS_SYNCHRONOUSLY;
1555 :
1556 1374 : if (flags & READ_BUFFERS_ISSUE_ADVICE)
1557 : {
1558 : /*
1559 : * In theory we should only do this if PinBufferForBlock() had to
1560 : * allocate new buffers above. That way, if two calls to
1561 : * StartReadBuffers() were made for the same blocks before
1562 : * WaitReadBuffers(), only the first would issue the advice.
1563 : * That'd be a better simulation of true asynchronous I/O, which
1564 : * would only start the I/O once, but isn't done here for
1565 : * simplicity.
1566 : */
1567 19 : smgrprefetch(operation->smgr,
1568 : operation->forknum,
1569 : blockNum,
1570 : actual_nblocks);
1571 : }
1572 :
1573 : /*
1574 : * Indicate that WaitReadBuffers() should be called. WaitReadBuffers()
1575 : * will initiate the necessary IO.
1576 : */
1577 1374 : did_start_io = true;
1578 : }
1579 :
1580 1481274 : CheckReadBuffersOperation(operation, !did_start_io);
1581 :
1582 1481274 : return did_start_io;
1583 : }
1584 :
1585 : /*
1586 : * Begin reading a range of blocks beginning at blockNum and extending for
1587 : * *nblocks. *nblocks and the buffers array are in/out parameters. On entry,
1588 : * the buffers elements covered by *nblocks must hold either InvalidBuffer or
1589 : * buffers forwarded by an earlier call to StartReadBuffers() that was split
1590 : * and is now being continued. On return, *nblocks holds the number of blocks
1591 : * accepted by this operation. If it is less than the original number then
1592 : * this operation has been split, but buffer elements up to the original
1593 : * requested size may hold forwarded buffers to be used for a continuing
1594 : * operation. The caller must either start a new I/O beginning at the block
1595 : * immediately following the blocks accepted by this call and pass those
1596 : * buffers back in, or release them if it chooses not to. It shouldn't make
1597 : * any other use of or assumptions about forwarded buffers.
1598 : *
1599 : * If false is returned, no I/O is necessary and the buffers covered by
1600 : * *nblocks on exit are valid and ready to be accessed. If true is returned,
1601 : * an I/O has been started, and WaitReadBuffers() must be called with the same
1602 : * operation object before the buffers covered by *nblocks on exit can be
1603 : * accessed. Along with the operation object, the caller-supplied array of
1604 : * buffers must remain valid until WaitReadBuffers() is called, and any
1605 : * forwarded buffers must also be preserved for a continuing call unless
1606 : * they are explicitly released.
1607 : */
1608 : bool
1609 2109473 : StartReadBuffers(ReadBuffersOperation *operation,
1610 : Buffer *buffers,
1611 : BlockNumber blockNum,
1612 : int *nblocks,
1613 : int flags)
1614 : {
1615 2109473 : return StartReadBuffersImpl(operation, buffers, blockNum, nblocks, flags,
1616 : true /* expect forwarded buffers */ );
1617 : }
1618 :
1619 : /*
1620 : * Single block version of the StartReadBuffers(). This might save a few
1621 : * instructions when called from another translation unit, because it is
1622 : * specialized for nblocks == 1.
1623 : *
1624 : * This version does not support "forwarded" buffers: they cannot be created
1625 : * by reading only one block and *buffer is ignored on entry.
1626 : */
1627 : bool
1628 82097803 : StartReadBuffer(ReadBuffersOperation *operation,
1629 : Buffer *buffer,
1630 : BlockNumber blocknum,
1631 : int flags)
1632 : {
1633 82097803 : int nblocks = 1;
1634 : bool result;
1635 :
1636 82097803 : result = StartReadBuffersImpl(operation, buffer, blocknum, &nblocks, flags,
1637 : false /* single block, no forwarding */ );
1638 : Assert(nblocks == 1); /* single block can't be short */
1639 :
1640 82097788 : return result;
1641 : }
1642 :
1643 : /*
1644 : * Perform sanity checks on the ReadBuffersOperation.
1645 : */
1646 : static void
1647 4443402 : CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete)
1648 : {
1649 : #ifdef USE_ASSERT_CHECKING
1650 : Assert(operation->nblocks_done <= operation->nblocks);
1651 : Assert(!is_complete || operation->nblocks == operation->nblocks_done);
1652 :
1653 : for (int i = 0; i < operation->nblocks; i++)
1654 : {
1655 : Buffer buffer = operation->buffers[i];
1656 : BufferDesc *buf_hdr = BufferIsLocal(buffer) ?
1657 : GetLocalBufferDescriptor(-buffer - 1) :
1658 : GetBufferDescriptor(buffer - 1);
1659 :
1660 : Assert(BufferGetBlockNumber(buffer) == operation->blocknum + i);
1661 : Assert(pg_atomic_read_u64(&buf_hdr->state) & BM_TAG_VALID);
1662 :
1663 : if (i < operation->nblocks_done)
1664 : Assert(pg_atomic_read_u64(&buf_hdr->state) & BM_VALID);
1665 : }
1666 : #endif
1667 4443402 : }
1668 :
1669 : /*
1670 : * We track various stats related to buffer hits. Because this is done in a
1671 : * few separate places, this helper exists for convenience.
1672 : */
1673 : static pg_attribute_always_inline void
1674 82768082 : TrackBufferHit(IOObject io_object, IOContext io_context,
1675 : Relation rel, char persistence, SMgrRelation smgr,
1676 : ForkNumber forknum, BlockNumber blocknum)
1677 : {
1678 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forknum,
1679 : blocknum,
1680 : smgr->smgr_rlocator.locator.spcOid,
1681 : smgr->smgr_rlocator.locator.dbOid,
1682 : smgr->smgr_rlocator.locator.relNumber,
1683 : smgr->smgr_rlocator.backend,
1684 : true);
1685 :
1686 82768082 : if (persistence == RELPERSISTENCE_TEMP)
1687 1637410 : pgBufferUsage.local_blks_hit += 1;
1688 : else
1689 81130672 : pgBufferUsage.shared_blks_hit += 1;
1690 :
1691 82768082 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1692 :
1693 82768082 : if (VacuumCostActive)
1694 2680649 : VacuumCostBalance += VacuumCostPageHit;
1695 :
1696 82768082 : if (rel)
1697 77072939 : pgstat_count_buffer_hit(rel);
1698 82768082 : }
1699 :
1700 : /*
1701 : * Helper for WaitReadBuffers() that processes the results of a readv
1702 : * operation, raising an error if necessary.
1703 : */
1704 : static void
1705 1477123 : ProcessReadBuffersResult(ReadBuffersOperation *operation)
1706 : {
1707 1477123 : PgAioReturn *aio_ret = &operation->io_return;
1708 1477123 : PgAioResultStatus rs = aio_ret->result.status;
1709 1477123 : int newly_read_blocks = 0;
1710 :
1711 : Assert(pgaio_wref_valid(&operation->io_wref));
1712 : Assert(aio_ret->result.status != PGAIO_RS_UNKNOWN);
1713 :
1714 : /*
1715 : * SMGR reports the number of blocks successfully read as the result of
1716 : * the IO operation. Thus we can simply add that to ->nblocks_done.
1717 : */
1718 :
1719 1477123 : if (likely(rs != PGAIO_RS_ERROR))
1720 1477094 : newly_read_blocks = aio_ret->result.result;
1721 :
1722 1477123 : if (rs == PGAIO_RS_ERROR || rs == PGAIO_RS_WARNING)
1723 45 : pgaio_result_report(aio_ret->result, &aio_ret->target_data,
1724 : rs == PGAIO_RS_ERROR ? ERROR : WARNING);
1725 1477078 : else if (aio_ret->result.status == PGAIO_RS_PARTIAL)
1726 : {
1727 : /*
1728 : * We'll retry, so we just emit a debug message to the server log (or
1729 : * not even that in prod scenarios).
1730 : */
1731 109 : pgaio_result_report(aio_ret->result, &aio_ret->target_data, DEBUG1);
1732 109 : elog(DEBUG3, "partial read, will retry");
1733 : }
1734 :
1735 : Assert(newly_read_blocks > 0);
1736 : Assert(newly_read_blocks <= MAX_IO_COMBINE_LIMIT);
1737 :
1738 1477094 : operation->nblocks_done += newly_read_blocks;
1739 :
1740 : Assert(operation->nblocks_done <= operation->nblocks);
1741 1477094 : }
1742 :
1743 : /*
1744 : * Wait for the IO operation initiated by StartReadBuffers() et al to
1745 : * complete.
1746 : *
1747 : * Returns true if we needed to wait for the IO operation, false otherwise.
1748 : */
1749 : bool
1750 1480335 : WaitReadBuffers(ReadBuffersOperation *operation)
1751 : {
1752 1480335 : PgAioReturn *aio_ret = &operation->io_return;
1753 : IOContext io_context;
1754 : IOObject io_object;
1755 1480335 : bool needed_wait = false;
1756 :
1757 1480335 : if (operation->persistence == RELPERSISTENCE_TEMP)
1758 : {
1759 2418 : io_context = IOCONTEXT_NORMAL;
1760 2418 : io_object = IOOBJECT_TEMP_RELATION;
1761 : }
1762 : else
1763 : {
1764 1477917 : io_context = IOContextForStrategy(operation->strategy);
1765 1477917 : io_object = IOOBJECT_RELATION;
1766 : }
1767 :
1768 : /*
1769 : * If we get here without an IO operation having been issued, the
1770 : * io_method == IOMETHOD_SYNC path must have been used. Otherwise the
1771 : * caller should not have called WaitReadBuffers().
1772 : *
1773 : * In the case of IOMETHOD_SYNC, we start - as we used to before the
1774 : * introducing of AIO - the IO in WaitReadBuffers(). This is done as part
1775 : * of the retry logic below, no extra code is required.
1776 : *
1777 : * This path is expected to eventually go away.
1778 : */
1779 1480335 : if (!pgaio_wref_valid(&operation->io_wref) && io_method != IOMETHOD_SYNC)
1780 0 : elog(ERROR, "waiting for read operation that didn't read");
1781 :
1782 : /*
1783 : * To handle partial reads, and IOMETHOD_SYNC, we re-issue IO until we're
1784 : * done. We may need multiple retries, not just because we could get
1785 : * multiple partial reads, but also because some of the remaining
1786 : * to-be-read buffers may have been read in by other backends, limiting
1787 : * the IO size.
1788 : */
1789 : while (true)
1790 1488 : {
1791 : int ignored_nblocks_progress;
1792 :
1793 1481823 : CheckReadBuffersOperation(operation, false);
1794 :
1795 : /*
1796 : * If there is an IO associated with the operation, we may need to
1797 : * wait for it.
1798 : */
1799 1481823 : if (pgaio_wref_valid(&operation->io_wref))
1800 : {
1801 : /*
1802 : * Track the time spent waiting for the IO to complete. As
1803 : * tracking a wait even if we don't actually need to wait
1804 : *
1805 : * a) is not cheap, due to the timestamping overhead
1806 : *
1807 : * b) reports some time as waiting, even if we never waited
1808 : *
1809 : * we first check if we already know the IO is complete.
1810 : *
1811 : * Note that operation->io_return is uninitialized for foreign IO,
1812 : * so we cannot use the cheaper PGAIO_RS_UNKNOWN pre-check.
1813 : */
1814 1480440 : if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
1815 677595 : !pgaio_wref_check_done(&operation->io_wref))
1816 : {
1817 327642 : instr_time io_start = pgstat_prepare_io_time(track_io_timing);
1818 :
1819 327642 : pgaio_wref_wait(&operation->io_wref);
1820 327641 : needed_wait = true;
1821 :
1822 : /*
1823 : * The IO operation itself was already counted earlier, in
1824 : * AsyncReadBuffers(), this just accounts for the wait time.
1825 : */
1826 327641 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1827 : io_start, 0, 0);
1828 : }
1829 : else
1830 : {
1831 : Assert(pgaio_wref_check_done(&operation->io_wref));
1832 : }
1833 :
1834 1480439 : if (unlikely(operation->foreign_io))
1835 : {
1836 3316 : Buffer buffer = operation->buffers[operation->nblocks_done];
1837 3316 : BufferDesc *desc = BufferIsLocal(buffer) ?
1838 3316 : GetLocalBufferDescriptor(-buffer - 1) :
1839 3316 : GetBufferDescriptor(buffer - 1);
1840 3316 : uint64 buf_state = pg_atomic_read_u64(&desc->state);
1841 :
1842 3316 : if (buf_state & BM_VALID)
1843 : {
1844 3314 : BlockNumber blocknum = operation->blocknum + operation->nblocks_done;
1845 :
1846 3314 : operation->nblocks_done += 1;
1847 : Assert(operation->nblocks_done <= operation->nblocks);
1848 :
1849 : /*
1850 : * Track this as a 'hit' for this backend. The backend
1851 : * performing the IO will track it as a 'read'.
1852 : */
1853 3314 : TrackBufferHit(io_object, io_context,
1854 3314 : operation->rel, operation->persistence,
1855 : operation->smgr, operation->forknum,
1856 : blocknum);
1857 : }
1858 :
1859 : /*
1860 : * If the foreign IO failed and left the buffer invalid,
1861 : * nblocks_done is not incremented. The retry loop below will
1862 : * call AsyncReadBuffers() which will attempt the IO itself.
1863 : */
1864 : }
1865 : else
1866 : {
1867 : /*
1868 : * We now are sure the IO completed. Check the results. This
1869 : * includes reporting on errors if there were any.
1870 : */
1871 1477123 : ProcessReadBuffersResult(operation);
1872 : }
1873 : }
1874 :
1875 : /*
1876 : * Most of the time, the one IO we already started, will read in
1877 : * everything. But we need to deal with partial reads and buffers not
1878 : * needing IO anymore.
1879 : */
1880 1481793 : if (operation->nblocks_done == operation->nblocks)
1881 1480305 : break;
1882 :
1883 1488 : CHECK_FOR_INTERRUPTS();
1884 :
1885 : /*
1886 : * If the IO completed only partially, we need to perform additional
1887 : * work, consider that a form of having had to wait.
1888 : */
1889 1488 : needed_wait = true;
1890 :
1891 : /*
1892 : * This may only complete the IO partially, either because some
1893 : * buffers were already valid, or because of a partial read.
1894 : *
1895 : * NB: In contrast to after the AsyncReadBuffers() call in
1896 : * StartReadBuffers(), we do *not* reduce
1897 : * ReadBuffersOperation->nblocks here, callers expect the full
1898 : * operation to be completed at this point (as more operations may
1899 : * have been queued).
1900 : */
1901 1488 : AsyncReadBuffers(operation, &ignored_nblocks_progress);
1902 : }
1903 :
1904 1480305 : CheckReadBuffersOperation(operation, true);
1905 :
1906 : /* NB: READ_DONE tracepoint was already executed in completion callback */
1907 1480305 : return needed_wait;
1908 : }
1909 :
1910 : /*
1911 : * Initiate IO for the ReadBuffersOperation
1912 : *
1913 : * This function only starts a single IO at a time. The size of the IO may be
1914 : * limited to below the to-be-read blocks, if one of the buffers has
1915 : * concurrently been read in. If the first to-be-read buffer is already valid,
1916 : * no IO will be issued.
1917 : *
1918 : * To support retries after partial reads, the first operation->nblocks_done
1919 : * buffers are skipped.
1920 : *
1921 : * On return *nblocks_progress is updated to reflect the number of buffers
1922 : * affected by the call. If the first buffer is valid, *nblocks_progress is
1923 : * set to 1 and operation->nblocks_done is incremented.
1924 : *
1925 : * Returns true if IO was initiated or is already in progress (foreign IO),
1926 : * false if the buffer was already valid.
1927 : */
1928 : static bool
1929 1481403 : AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
1930 : {
1931 1481403 : Buffer *buffers = &operation->buffers[0];
1932 1481403 : int flags = operation->flags;
1933 1481403 : ForkNumber forknum = operation->forknum;
1934 1481403 : char persistence = operation->persistence;
1935 1481403 : int16 nblocks_done = operation->nblocks_done;
1936 1481403 : BlockNumber blocknum = operation->blocknum + nblocks_done;
1937 1481403 : Buffer *io_buffers = &operation->buffers[nblocks_done];
1938 1481403 : int io_buffers_len = 0;
1939 : PgAioHandle *ioh;
1940 1481403 : uint32 ioh_flags = 0;
1941 : void *io_pages[MAX_IO_COMBINE_LIMIT];
1942 : IOContext io_context;
1943 : IOObject io_object;
1944 : instr_time io_start;
1945 : StartBufferIOResult status;
1946 :
1947 1481403 : if (persistence == RELPERSISTENCE_TEMP)
1948 : {
1949 2810 : io_context = IOCONTEXT_NORMAL;
1950 2810 : io_object = IOOBJECT_TEMP_RELATION;
1951 : }
1952 : else
1953 : {
1954 1478593 : io_context = IOContextForStrategy(operation->strategy);
1955 1478593 : io_object = IOOBJECT_RELATION;
1956 : }
1957 :
1958 : /*
1959 : * When this IO is executed synchronously, either because the caller will
1960 : * immediately block waiting for the IO or because IOMETHOD_SYNC is used,
1961 : * the AIO subsystem needs to know.
1962 : */
1963 1481403 : if (flags & READ_BUFFERS_SYNCHRONOUSLY)
1964 792923 : ioh_flags |= PGAIO_HF_SYNCHRONOUS;
1965 :
1966 1481403 : if (persistence == RELPERSISTENCE_TEMP)
1967 2810 : ioh_flags |= PGAIO_HF_REFERENCES_LOCAL;
1968 :
1969 : /*
1970 : * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
1971 : * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
1972 : * set globally, but on a per-session basis. The completion callback,
1973 : * which may be run in other processes, e.g. in IO workers, may have a
1974 : * different value of the zero_damaged_pages GUC.
1975 : *
1976 : * XXX: We probably should eventually use a different flag for
1977 : * zero_damaged_pages, so we can report different log levels / error codes
1978 : * for zero_damaged_pages and ZERO_ON_ERROR.
1979 : */
1980 1481403 : if (zero_damaged_pages)
1981 16 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1982 :
1983 : /*
1984 : * For the same reason as with zero_damaged_pages we need to use this
1985 : * backend's ignore_checksum_failure value.
1986 : */
1987 1481403 : if (ignore_checksum_failure)
1988 8 : flags |= READ_BUFFERS_IGNORE_CHECKSUM_FAILURES;
1989 :
1990 :
1991 : /*
1992 : * To be allowed to report stats in the local completion callback we need
1993 : * to prepare to report stats now. This ensures we can safely report the
1994 : * checksum failure even in a critical section.
1995 : */
1996 1481403 : pgstat_prepare_report_checksum_failure(operation->smgr->smgr_rlocator.locator.dbOid);
1997 :
1998 : /*
1999 : * We must get an IO handle before StartBufferIO(), as pgaio_io_acquire()
2000 : * might block, which we don't want after setting IO_IN_PROGRESS. If we
2001 : * don't need to do the IO, we'll release the handle.
2002 : *
2003 : * If we need to wait for IO before we can get a handle, submit
2004 : * already-staged IO first, so that other backends don't need to wait.
2005 : * There wouldn't be a deadlock risk, as pgaio_io_acquire() just needs to
2006 : * wait for already submitted IO, which doesn't require additional locks,
2007 : * but it could still cause undesirable waits.
2008 : *
2009 : * A secondary benefit is that this would allow us to measure the time in
2010 : * pgaio_io_acquire() without causing undue timer overhead in the common,
2011 : * non-blocking, case. However, currently the pgstats infrastructure
2012 : * doesn't really allow that, as it a) asserts that an operation can't
2013 : * have time without operations b) doesn't have an API to report
2014 : * "accumulated" time.
2015 : */
2016 1481403 : ioh = pgaio_io_acquire_nb(CurrentResourceOwner, &operation->io_return);
2017 1481403 : if (unlikely(!ioh))
2018 : {
2019 3150 : pgaio_submit_staged();
2020 3150 : ioh = pgaio_io_acquire(CurrentResourceOwner, &operation->io_return);
2021 : }
2022 :
2023 1481403 : operation->foreign_io = false;
2024 1481403 : pgaio_wref_clear(&operation->io_wref);
2025 :
2026 : /*
2027 : * Try to start IO on the first buffer in a new run of blocks. If AIO is
2028 : * in progress, be it in this backend or another backend, we just
2029 : * associate the wait reference with the operation and wait in
2030 : * WaitReadBuffers(). This turns out to be important for performance in
2031 : * two workloads:
2032 : *
2033 : * 1) A read stream that has to read the same block multiple times within
2034 : * the readahead distance. This can happen e.g. for the table accesses of
2035 : * an index scan.
2036 : *
2037 : * 2) Concurrent scans by multiple backends on the same relation.
2038 : *
2039 : * If we were to synchronously wait for the in-progress IO, we'd not be
2040 : * able to keep enough I/O in flight.
2041 : *
2042 : * If we do find there is ongoing I/O for the buffer, we set up a 1-block
2043 : * ReadBuffersOperation that WaitReadBuffers then can wait on.
2044 : *
2045 : * It's possible that another backend has started IO on the buffer but not
2046 : * yet set its wait reference. In this case, we have no choice but to wait
2047 : * for either the wait reference to be valid or the IO to be done.
2048 : */
2049 1481403 : status = StartBufferIO(buffers[nblocks_done], true, true,
2050 : &operation->io_wref);
2051 1481403 : if (status != BUFFER_IO_READY_FOR_IO)
2052 : {
2053 3868 : pgaio_io_release(ioh);
2054 3868 : *nblocks_progress = 1;
2055 3868 : if (status == BUFFER_IO_ALREADY_DONE)
2056 : {
2057 : /*
2058 : * Someone has already completed this block, we're done.
2059 : *
2060 : * When IO is necessary, ->nblocks_done is updated in
2061 : * ProcessReadBuffersResult(), but that is not called if no IO is
2062 : * necessary. Thus update here.
2063 : */
2064 552 : operation->nblocks_done += 1;
2065 : Assert(operation->nblocks_done <= operation->nblocks);
2066 :
2067 : Assert(!pgaio_wref_valid(&operation->io_wref));
2068 :
2069 : /*
2070 : * Report and track this as a 'hit' for this backend, even though
2071 : * it must have started out as a miss in PinBufferForBlock(). The
2072 : * other backend will track this as a 'read'.
2073 : */
2074 552 : TrackBufferHit(io_object, io_context,
2075 552 : operation->rel, operation->persistence,
2076 : operation->smgr, operation->forknum,
2077 : blocknum);
2078 552 : return false;
2079 : }
2080 :
2081 : /* The IO is already in-progress */
2082 : Assert(status == BUFFER_IO_IN_PROGRESS);
2083 : Assert(pgaio_wref_valid(&operation->io_wref));
2084 3316 : operation->foreign_io = true;
2085 :
2086 3316 : return true;
2087 : }
2088 :
2089 : Assert(io_buffers[0] == buffers[nblocks_done]);
2090 1477535 : io_pages[0] = BufferGetBlock(buffers[nblocks_done]);
2091 1477535 : io_buffers_len = 1;
2092 :
2093 : /*
2094 : * NB: As little code as possible should be added between the
2095 : * StartBufferIO() above, the further StartBufferIO()s below and the
2096 : * smgrstartreadv(), as some of the buffers are now marked as
2097 : * IO_IN_PROGRESS and will thus cause other backends to wait.
2098 : */
2099 :
2100 : /*
2101 : * How many neighboring-on-disk blocks can we scatter-read into other
2102 : * buffers at the same time? In this case we don't wait if we see an I/O
2103 : * already in progress (see comment above).
2104 : */
2105 1665861 : for (int i = nblocks_done + 1; i < operation->nblocks; i++)
2106 : {
2107 : /* Must be consecutive block numbers. */
2108 : Assert(BufferGetBlockNumber(buffers[i - 1]) ==
2109 : BufferGetBlockNumber(buffers[i]) - 1);
2110 :
2111 188330 : status = StartBufferIO(buffers[i], true, false, NULL);
2112 188330 : if (status != BUFFER_IO_READY_FOR_IO)
2113 4 : break;
2114 :
2115 : Assert(io_buffers[io_buffers_len] == buffers[i]);
2116 :
2117 188326 : io_pages[io_buffers_len++] = BufferGetBlock(buffers[i]);
2118 : }
2119 :
2120 : /* get a reference to wait for in WaitReadBuffers() */
2121 1477535 : pgaio_io_get_wref(ioh, &operation->io_wref);
2122 :
2123 : /* provide the list of buffers to the completion callbacks */
2124 1477535 : pgaio_io_set_handle_data_32(ioh, (uint32 *) io_buffers, io_buffers_len);
2125 :
2126 1477535 : pgaio_io_register_callbacks(ioh,
2127 : persistence == RELPERSISTENCE_TEMP ?
2128 : PGAIO_HCB_LOCAL_BUFFER_READV :
2129 : PGAIO_HCB_SHARED_BUFFER_READV,
2130 : flags);
2131 :
2132 1477535 : pgaio_io_set_flag(ioh, ioh_flags);
2133 :
2134 : /* ---
2135 : * Even though we're trying to issue IO asynchronously, track the time
2136 : * in smgrstartreadv():
2137 : * - if io_method == IOMETHOD_SYNC, we will always perform the IO
2138 : * immediately
2139 : * - the io method might not support the IO (e.g. worker IO for a temp
2140 : * table)
2141 : * ---
2142 : */
2143 1477535 : io_start = pgstat_prepare_io_time(track_io_timing);
2144 1477535 : smgrstartreadv(ioh, operation->smgr, forknum,
2145 : blocknum,
2146 : io_pages, io_buffers_len);
2147 1477520 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
2148 1477520 : io_start, 1, io_buffers_len * BLCKSZ);
2149 :
2150 1477520 : if (persistence == RELPERSISTENCE_TEMP)
2151 2808 : pgBufferUsage.local_blks_read += io_buffers_len;
2152 : else
2153 1474712 : pgBufferUsage.shared_blks_read += io_buffers_len;
2154 :
2155 : /*
2156 : * Track vacuum cost when issuing IO, not after waiting for it. Otherwise
2157 : * we could end up issuing a lot of IO in a short timespan, despite a low
2158 : * cost limit.
2159 : */
2160 1477520 : if (VacuumCostActive)
2161 21623 : VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
2162 :
2163 1477520 : *nblocks_progress = io_buffers_len;
2164 :
2165 1477520 : return true;
2166 : }
2167 :
2168 : /*
2169 : * BufferAlloc -- subroutine for PinBufferForBlock. Handles lookup of a shared
2170 : * buffer. If no buffer exists already, selects a replacement victim and
2171 : * evicts the old page, but does NOT read in new page.
2172 : *
2173 : * "strategy" can be a buffer replacement strategy object, or NULL for
2174 : * the default strategy. The selected buffer's usage_count is advanced when
2175 : * using the default strategy, but otherwise possibly not (see PinBuffer).
2176 : *
2177 : * The returned buffer is pinned and is already marked as holding the
2178 : * desired page. If it already did have the desired page, *foundPtr is
2179 : * set true. Otherwise, *foundPtr is set false.
2180 : *
2181 : * io_context is passed as an output parameter to avoid calling
2182 : * IOContextForStrategy() when there is a shared buffers hit and no IO
2183 : * statistics need be captured.
2184 : *
2185 : * No locks are held either at entry or exit.
2186 : */
2187 : static pg_attribute_always_inline BufferDesc *
2188 83107630 : BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
2189 : BlockNumber blockNum,
2190 : BufferAccessStrategy strategy,
2191 : bool *foundPtr, IOContext io_context)
2192 : {
2193 : BufferTag newTag; /* identity of requested block */
2194 : uint32 newHash; /* hash value for newTag */
2195 : LWLock *newPartitionLock; /* buffer partition lock for it */
2196 : int existing_buf_id;
2197 : Buffer victim_buffer;
2198 : BufferDesc *victim_buf_hdr;
2199 : uint64 victim_buf_state;
2200 83107630 : uint64 set_bits = 0;
2201 :
2202 : /* Make sure we will have room to remember the buffer pin */
2203 83107630 : ResourceOwnerEnlarge(CurrentResourceOwner);
2204 83107630 : ReservePrivateRefCountEntry();
2205 :
2206 : /* create a tag so we can lookup the buffer */
2207 83107630 : InitBufferTag(&newTag, &smgr->smgr_rlocator.locator, forkNum, blockNum);
2208 :
2209 : /* determine its hash code and partition lock ID */
2210 83107630 : newHash = BufTableHashCode(&newTag);
2211 83107630 : newPartitionLock = BufMappingPartitionLock(newHash);
2212 :
2213 : /* see if the block is in the buffer pool already */
2214 83107630 : LWLockAcquire(newPartitionLock, LW_SHARED);
2215 83107630 : existing_buf_id = BufTableLookup(&newTag, newHash);
2216 83107630 : if (existing_buf_id >= 0)
2217 : {
2218 : BufferDesc *buf;
2219 : bool valid;
2220 :
2221 : /*
2222 : * Found it. Now, pin the buffer so no one can steal it from the
2223 : * buffer pool, and check to see if the correct data has been loaded
2224 : * into the buffer.
2225 : */
2226 81129700 : buf = GetBufferDescriptor(existing_buf_id);
2227 :
2228 81129700 : valid = PinBuffer(buf, strategy, false);
2229 :
2230 : /* Can release the mapping lock as soon as we've pinned it */
2231 81129700 : LWLockRelease(newPartitionLock);
2232 :
2233 81129700 : *foundPtr = true;
2234 :
2235 81129700 : if (!valid)
2236 : {
2237 : /*
2238 : * We can only get here if (a) someone else is still reading in
2239 : * the page, (b) a previous read attempt failed, or (c) someone
2240 : * called StartReadBuffers() but not yet WaitReadBuffers().
2241 : */
2242 3202 : *foundPtr = false;
2243 : }
2244 :
2245 81129700 : return buf;
2246 : }
2247 :
2248 : /*
2249 : * Didn't find it in the buffer pool. We'll have to initialize a new
2250 : * buffer. Remember to unlock the mapping lock while doing the work.
2251 : */
2252 1977930 : LWLockRelease(newPartitionLock);
2253 :
2254 : /*
2255 : * Acquire a victim buffer. Somebody else might try to do the same, we
2256 : * don't hold any conflicting locks. If so we'll have to undo our work
2257 : * later.
2258 : */
2259 1977930 : victim_buffer = GetVictimBuffer(strategy, io_context);
2260 1977930 : victim_buf_hdr = GetBufferDescriptor(victim_buffer - 1);
2261 :
2262 : /*
2263 : * Try to make a hashtable entry for the buffer under its new tag. If
2264 : * somebody else inserted another buffer for the tag, we'll release the
2265 : * victim buffer we acquired and use the already inserted one.
2266 : */
2267 1977930 : LWLockAcquire(newPartitionLock, LW_EXCLUSIVE);
2268 1977930 : existing_buf_id = BufTableInsert(&newTag, newHash, victim_buf_hdr->buf_id);
2269 1977930 : if (existing_buf_id >= 0)
2270 : {
2271 : BufferDesc *existing_buf_hdr;
2272 : bool valid;
2273 :
2274 : /*
2275 : * Got a collision. Someone has already done what we were about to do.
2276 : * We'll just handle this as if it were found in the buffer pool in
2277 : * the first place. First, give up the buffer we were planning to
2278 : * use.
2279 : *
2280 : * We could do this after releasing the partition lock, but then we'd
2281 : * have to call ResourceOwnerEnlarge() & ReservePrivateRefCountEntry()
2282 : * before acquiring the lock, for the rare case of such a collision.
2283 : */
2284 1024 : UnpinBuffer(victim_buf_hdr);
2285 :
2286 : /* remaining code should match code at top of routine */
2287 :
2288 1024 : existing_buf_hdr = GetBufferDescriptor(existing_buf_id);
2289 :
2290 1024 : valid = PinBuffer(existing_buf_hdr, strategy, false);
2291 :
2292 : /* Can release the mapping lock as soon as we've pinned it */
2293 1024 : LWLockRelease(newPartitionLock);
2294 :
2295 1024 : *foundPtr = true;
2296 :
2297 1024 : if (!valid)
2298 : {
2299 : /*
2300 : * We can only get here if (a) someone else is still reading in
2301 : * the page, (b) a previous read attempt failed, or (c) someone
2302 : * called StartReadBuffers() but not yet WaitReadBuffers().
2303 : */
2304 714 : *foundPtr = false;
2305 : }
2306 :
2307 1024 : return existing_buf_hdr;
2308 : }
2309 :
2310 : /*
2311 : * Need to lock the buffer header too in order to change its tag.
2312 : */
2313 1976906 : victim_buf_state = LockBufHdr(victim_buf_hdr);
2314 :
2315 : /* some sanity checks while we hold the buffer header lock */
2316 : Assert(BUF_STATE_GET_REFCOUNT(victim_buf_state) == 1);
2317 : Assert(!(victim_buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY | BM_IO_IN_PROGRESS)));
2318 :
2319 1976906 : victim_buf_hdr->tag = newTag;
2320 :
2321 : /*
2322 : * Make sure BM_PERMANENT is set for buffers that must be written at every
2323 : * checkpoint. Unlogged buffers only need to be written at shutdown
2324 : * checkpoints, except for their "init" forks, which need to be treated
2325 : * just like permanent relations.
2326 : */
2327 1976906 : set_bits |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2328 1976906 : if (relpersistence == RELPERSISTENCE_PERMANENT || forkNum == INIT_FORKNUM)
2329 1976547 : set_bits |= BM_PERMANENT;
2330 :
2331 1976906 : UnlockBufHdrExt(victim_buf_hdr, victim_buf_state,
2332 : set_bits, 0, 0);
2333 :
2334 1976906 : LWLockRelease(newPartitionLock);
2335 :
2336 : /*
2337 : * Buffer contents are currently invalid.
2338 : */
2339 1976906 : *foundPtr = false;
2340 :
2341 1976906 : return victim_buf_hdr;
2342 : }
2343 :
2344 : /*
2345 : * InvalidateBuffer -- mark a shared buffer invalid.
2346 : *
2347 : * The buffer header spinlock must be held at entry. We drop it before
2348 : * returning. (This is sane because the caller must have locked the
2349 : * buffer in order to be sure it should be dropped.)
2350 : *
2351 : * This is used only in contexts such as dropping a relation. We assume
2352 : * that no other backend could possibly be interested in using the page,
2353 : * so the only reason the buffer might be pinned is if someone else is
2354 : * trying to write it out. We have to let them finish before we can
2355 : * reclaim the buffer.
2356 : *
2357 : * The buffer could get reclaimed by someone else while we are waiting
2358 : * to acquire the necessary locks; if so, don't mess it up.
2359 : */
2360 : static void
2361 133472 : InvalidateBuffer(BufferDesc *buf)
2362 : {
2363 : BufferTag oldTag;
2364 : uint32 oldHash; /* hash value for oldTag */
2365 : LWLock *oldPartitionLock; /* buffer partition lock for it */
2366 : uint32 oldFlags;
2367 : uint64 buf_state;
2368 :
2369 : /* Save the original buffer tag before dropping the spinlock */
2370 133472 : oldTag = buf->tag;
2371 :
2372 133472 : UnlockBufHdr(buf);
2373 :
2374 : /*
2375 : * Need to compute the old tag's hashcode and partition lock ID. XXX is it
2376 : * worth storing the hashcode in BufferDesc so we need not recompute it
2377 : * here? Probably not.
2378 : */
2379 133472 : oldHash = BufTableHashCode(&oldTag);
2380 133472 : oldPartitionLock = BufMappingPartitionLock(oldHash);
2381 :
2382 154243 : retry:
2383 :
2384 : /*
2385 : * Acquire exclusive mapping lock in preparation for changing the buffer's
2386 : * association.
2387 : */
2388 154243 : LWLockAcquire(oldPartitionLock, LW_EXCLUSIVE);
2389 :
2390 : /* Re-lock the buffer header */
2391 154243 : buf_state = LockBufHdr(buf);
2392 :
2393 : /* If it's changed while we were waiting for lock, do nothing */
2394 154243 : if (!BufferTagsEqual(&buf->tag, &oldTag))
2395 : {
2396 3 : UnlockBufHdr(buf);
2397 3 : LWLockRelease(oldPartitionLock);
2398 3 : return;
2399 : }
2400 :
2401 : /*
2402 : * We assume the reason for it to be pinned is that either we were
2403 : * asynchronously reading the page in before erroring out or someone else
2404 : * is flushing the page out. Wait for the IO to finish. (This could be
2405 : * an infinite loop if the refcount is messed up... it would be nice to
2406 : * time out after awhile, but there seems no way to be sure how many loops
2407 : * may be needed. Note that if the other guy has pinned the buffer but
2408 : * not yet done StartBufferIO, WaitIO will fall through and we'll
2409 : * effectively be busy-looping here.)
2410 : */
2411 154240 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 0)
2412 : {
2413 20771 : UnlockBufHdr(buf);
2414 20771 : LWLockRelease(oldPartitionLock);
2415 : /* safety check: should definitely not be our *own* pin */
2416 20771 : if (GetPrivateRefCount(BufferDescriptorGetBuffer(buf)) > 0)
2417 0 : elog(ERROR, "buffer is pinned in InvalidateBuffer");
2418 20771 : WaitIO(buf);
2419 20771 : goto retry;
2420 : }
2421 :
2422 : /*
2423 : * An invalidated buffer should not have any backends waiting to lock the
2424 : * buffer, therefore BM_LOCK_WAKE_IN_PROGRESS should not be set.
2425 : */
2426 : Assert(!(buf_state & BM_LOCK_WAKE_IN_PROGRESS));
2427 :
2428 : /*
2429 : * Clear out the buffer's tag and flags. We must do this to ensure that
2430 : * linear scans of the buffer array don't think the buffer is valid.
2431 : */
2432 133469 : oldFlags = buf_state & BUF_FLAG_MASK;
2433 133469 : ClearBufferTag(&buf->tag);
2434 :
2435 133469 : UnlockBufHdrExt(buf, buf_state,
2436 : 0,
2437 : BUF_FLAG_MASK | BUF_USAGECOUNT_MASK,
2438 : 0);
2439 :
2440 : /*
2441 : * Remove the buffer from the lookup hashtable, if it was in there.
2442 : */
2443 133469 : if (oldFlags & BM_TAG_VALID)
2444 133469 : BufTableDelete(&oldTag, oldHash);
2445 :
2446 : /*
2447 : * Done with mapping lock.
2448 : */
2449 133469 : LWLockRelease(oldPartitionLock);
2450 : }
2451 :
2452 : /*
2453 : * Helper routine for GetVictimBuffer()
2454 : *
2455 : * Needs to be called on a buffer with a valid tag, pinned, but without the
2456 : * buffer header spinlock held.
2457 : *
2458 : * Returns true if the buffer can be reused, in which case the buffer is only
2459 : * pinned by this backend and marked as invalid, false otherwise.
2460 : */
2461 : static bool
2462 1411481 : InvalidateVictimBuffer(BufferDesc *buf_hdr)
2463 : {
2464 : uint64 buf_state;
2465 : uint32 hash;
2466 : LWLock *partition_lock;
2467 : BufferTag tag;
2468 :
2469 : Assert(GetPrivateRefCount(BufferDescriptorGetBuffer(buf_hdr)) == 1);
2470 :
2471 : /* have buffer pinned, so it's safe to read tag without lock */
2472 1411481 : tag = buf_hdr->tag;
2473 :
2474 1411481 : hash = BufTableHashCode(&tag);
2475 1411481 : partition_lock = BufMappingPartitionLock(hash);
2476 :
2477 1411481 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2478 :
2479 : /* lock the buffer header */
2480 1411481 : buf_state = LockBufHdr(buf_hdr);
2481 :
2482 : /*
2483 : * We have the buffer pinned nobody else should have been able to unset
2484 : * this concurrently.
2485 : */
2486 : Assert(buf_state & BM_TAG_VALID);
2487 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2488 : Assert(BufferTagsEqual(&buf_hdr->tag, &tag));
2489 :
2490 : /*
2491 : * If somebody else pinned the buffer since, or even worse, dirtied it,
2492 : * give up on this buffer: It's clearly in use.
2493 : */
2494 1411481 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 1 || (buf_state & BM_DIRTY))
2495 : {
2496 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2497 :
2498 448 : UnlockBufHdr(buf_hdr);
2499 448 : LWLockRelease(partition_lock);
2500 :
2501 448 : return false;
2502 : }
2503 :
2504 : /*
2505 : * An invalidated buffer should not have any backends waiting to lock the
2506 : * buffer, therefore BM_LOCK_WAKE_IN_PROGRESS should not be set.
2507 : */
2508 : Assert(!(buf_state & BM_LOCK_WAKE_IN_PROGRESS));
2509 :
2510 : /*
2511 : * Clear out the buffer's tag and flags and usagecount. This is not
2512 : * strictly required, as BM_TAG_VALID/BM_VALID needs to be checked before
2513 : * doing anything with the buffer. But currently it's beneficial, as the
2514 : * cheaper pre-check for several linear scans of shared buffers use the
2515 : * tag (see e.g. FlushDatabaseBuffers()).
2516 : */
2517 1411033 : ClearBufferTag(&buf_hdr->tag);
2518 1411033 : UnlockBufHdrExt(buf_hdr, buf_state,
2519 : 0,
2520 : BUF_FLAG_MASK | BUF_USAGECOUNT_MASK,
2521 : 0);
2522 :
2523 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2524 :
2525 : /* finally delete buffer from the buffer mapping table */
2526 1411033 : BufTableDelete(&tag, hash);
2527 :
2528 1411033 : LWLockRelease(partition_lock);
2529 :
2530 1411033 : buf_state = pg_atomic_read_u64(&buf_hdr->state);
2531 : Assert(!(buf_state & (BM_DIRTY | BM_VALID | BM_TAG_VALID)));
2532 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2533 : Assert(BUF_STATE_GET_REFCOUNT(pg_atomic_read_u64(&buf_hdr->state)) > 0);
2534 :
2535 1411033 : return true;
2536 : }
2537 :
2538 : static Buffer
2539 2259691 : GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context)
2540 : {
2541 : BufferDesc *buf_hdr;
2542 : Buffer buf;
2543 : uint64 buf_state;
2544 : bool from_ring;
2545 :
2546 : /*
2547 : * Ensure, before we pin a victim buffer, that there's a free refcount
2548 : * entry and resource owner slot for the pin.
2549 : */
2550 2259691 : ReservePrivateRefCountEntry();
2551 2259691 : ResourceOwnerEnlarge(CurrentResourceOwner);
2552 :
2553 : /* we return here if a prospective victim buffer gets used concurrently */
2554 22833 : again:
2555 :
2556 : /*
2557 : * Select a victim buffer. The buffer is returned pinned and owned by
2558 : * this backend.
2559 : */
2560 2282524 : buf_hdr = StrategyGetBuffer(strategy, &buf_state, &from_ring);
2561 2282524 : buf = BufferDescriptorGetBuffer(buf_hdr);
2562 :
2563 : /*
2564 : * We shouldn't have any other pins for this buffer.
2565 : */
2566 2282524 : CheckBufferIsPinnedOnce(buf);
2567 :
2568 : /*
2569 : * If the buffer was dirty, try to write it out. There is a race
2570 : * condition here, another backend could dirty the buffer between
2571 : * StrategyGetBuffer() checking that it is not in use and invalidating the
2572 : * buffer below. That's addressed by InvalidateVictimBuffer() verifying
2573 : * that the buffer is not dirty.
2574 : */
2575 2282524 : if (buf_state & BM_DIRTY)
2576 : {
2577 : Assert(buf_state & BM_TAG_VALID);
2578 : Assert(buf_state & BM_VALID);
2579 :
2580 : /*
2581 : * We need a share-exclusive lock on the buffer contents to write it
2582 : * out (else we might write invalid data, eg because someone else is
2583 : * compacting the page contents while we write). We must use a
2584 : * conditional lock acquisition here to avoid deadlock. Even though
2585 : * the buffer was not pinned (and therefore surely not locked) when
2586 : * StrategyGetBuffer returned it, someone else could have pinned and
2587 : * (share-)exclusive-locked it by the time we get here. If we try to
2588 : * get the lock unconditionally, we'd block waiting for them; if they
2589 : * later block waiting for us, deadlock ensues. (This has been
2590 : * observed to happen when two backends are both trying to split btree
2591 : * index pages, and the second one just happens to be trying to split
2592 : * the page the first one got from StrategyGetBuffer.)
2593 : */
2594 376199 : if (!BufferLockConditional(buf, buf_hdr, BUFFER_LOCK_SHARE_EXCLUSIVE))
2595 : {
2596 : /*
2597 : * Someone else has locked the buffer, so give it up and loop back
2598 : * to get another one.
2599 : */
2600 0 : UnpinBuffer(buf_hdr);
2601 0 : goto again;
2602 : }
2603 :
2604 : /*
2605 : * If using a nondefault strategy, and this victim came from the
2606 : * strategy ring, let the strategy decide whether to reject it when
2607 : * reusing it would require a WAL flush. This only applies to
2608 : * permanent buffers; unlogged buffers can have fake LSNs, so
2609 : * XLogNeedsFlush() is not meaningful for them.
2610 : *
2611 : * We need to hold the content lock in at least share-exclusive mode
2612 : * to safely inspect the page LSN, so this couldn't have been done
2613 : * inside StrategyGetBuffer().
2614 : */
2615 376199 : if (strategy && from_ring &&
2616 190792 : buf_state & BM_PERMANENT &&
2617 122100 : XLogNeedsFlush(BufferGetLSN(buf_hdr)) &&
2618 26704 : StrategyRejectBuffer(strategy, buf_hdr, from_ring))
2619 : {
2620 22385 : UnlockReleaseBuffer(buf);
2621 22385 : goto again;
2622 : }
2623 :
2624 : /* OK, do the I/O */
2625 353814 : FlushBuffer(buf_hdr, NULL, IOOBJECT_RELATION, io_context);
2626 353814 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
2627 :
2628 353814 : ScheduleBufferTagForWriteback(&BackendWritebackContext, io_context,
2629 : &buf_hdr->tag);
2630 : }
2631 :
2632 :
2633 2260139 : if (buf_state & BM_VALID)
2634 : {
2635 : /*
2636 : * When a BufferAccessStrategy is in use, blocks evicted from shared
2637 : * buffers are counted as IOOP_EVICT in the corresponding context
2638 : * (e.g. IOCONTEXT_BULKWRITE). Shared buffers are evicted by a
2639 : * strategy in two cases: 1) while initially claiming buffers for the
2640 : * strategy ring 2) to replace an existing strategy ring buffer
2641 : * because it is pinned or in use and cannot be reused.
2642 : *
2643 : * Blocks evicted from buffers already in the strategy ring are
2644 : * counted as IOOP_REUSE in the corresponding strategy context.
2645 : *
2646 : * At this point, we can accurately count evictions and reuses,
2647 : * because we have successfully claimed the valid buffer. Previously,
2648 : * we may have been forced to release the buffer due to concurrent
2649 : * pinners or erroring out.
2650 : */
2651 1408941 : pgstat_count_io_op(IOOBJECT_RELATION, io_context,
2652 1408941 : from_ring ? IOOP_REUSE : IOOP_EVICT, 1, 0);
2653 : }
2654 :
2655 : /*
2656 : * If the buffer has an entry in the buffer mapping table, delete it. This
2657 : * can fail because another backend could have pinned or dirtied the
2658 : * buffer.
2659 : */
2660 2260139 : if ((buf_state & BM_TAG_VALID) && !InvalidateVictimBuffer(buf_hdr))
2661 : {
2662 448 : UnpinBuffer(buf_hdr);
2663 448 : goto again;
2664 : }
2665 :
2666 : /* a final set of sanity checks */
2667 : #ifdef USE_ASSERT_CHECKING
2668 : buf_state = pg_atomic_read_u64(&buf_hdr->state);
2669 :
2670 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2671 : Assert(!(buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY)));
2672 :
2673 : CheckBufferIsPinnedOnce(buf);
2674 : #endif
2675 :
2676 2259691 : return buf;
2677 : }
2678 :
2679 : /*
2680 : * Return the maximum number of buffers that a backend should try to pin once,
2681 : * to avoid exceeding its fair share. This is the highest value that
2682 : * GetAdditionalPinLimit() could ever return. Note that it may be zero on a
2683 : * system with a very small buffer pool relative to max_connections.
2684 : */
2685 : uint32
2686 766982 : GetPinLimit(void)
2687 : {
2688 766982 : return MaxProportionalPins;
2689 : }
2690 :
2691 : /*
2692 : * Return the maximum number of additional buffers that this backend should
2693 : * pin if it wants to stay under the per-backend limit, considering the number
2694 : * of buffers it has already pinned. Unlike LimitAdditionalPins(), the limit
2695 : * return by this function can be zero.
2696 : */
2697 : uint32
2698 4201476 : GetAdditionalPinLimit(void)
2699 : {
2700 : uint32 estimated_pins_held;
2701 :
2702 : /*
2703 : * We get the number of "overflowed" pins for free, but don't know the
2704 : * number of pins in PrivateRefCountArray. The cost of calculating that
2705 : * exactly doesn't seem worth it, so just assume the max.
2706 : */
2707 4201476 : estimated_pins_held = PrivateRefCountOverflowed + REFCOUNT_ARRAY_ENTRIES;
2708 :
2709 : /* Is this backend already holding more than its fair share? */
2710 4201476 : if (estimated_pins_held > MaxProportionalPins)
2711 1491685 : return 0;
2712 :
2713 2709791 : return MaxProportionalPins - estimated_pins_held;
2714 : }
2715 :
2716 : /*
2717 : * Limit the number of pins a batch operation may additionally acquire, to
2718 : * avoid running out of pinnable buffers.
2719 : *
2720 : * One additional pin is always allowed, on the assumption that the operation
2721 : * requires at least one to make progress.
2722 : */
2723 : void
2724 254977 : LimitAdditionalPins(uint32 *additional_pins)
2725 : {
2726 : uint32 limit;
2727 :
2728 254977 : if (*additional_pins <= 1)
2729 239523 : return;
2730 :
2731 15454 : limit = GetAdditionalPinLimit();
2732 15454 : limit = Max(limit, 1);
2733 15454 : if (limit < *additional_pins)
2734 10001 : *additional_pins = limit;
2735 : }
2736 :
2737 : /*
2738 : * Logic shared between ExtendBufferedRelBy(), ExtendBufferedRelTo(). Just to
2739 : * avoid duplicating the tracing and relpersistence related logic.
2740 : */
2741 : static BlockNumber
2742 269889 : ExtendBufferedRelCommon(BufferManagerRelation bmr,
2743 : ForkNumber fork,
2744 : BufferAccessStrategy strategy,
2745 : uint32 flags,
2746 : uint32 extend_by,
2747 : BlockNumber extend_upto,
2748 : Buffer *buffers,
2749 : uint32 *extended_by)
2750 : {
2751 : BlockNumber first_block;
2752 :
2753 : TRACE_POSTGRESQL_BUFFER_EXTEND_START(fork,
2754 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2755 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2756 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2757 : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2758 : extend_by);
2759 :
2760 269889 : if (bmr.relpersistence == RELPERSISTENCE_TEMP)
2761 14912 : first_block = ExtendBufferedRelLocal(bmr, fork, flags,
2762 : extend_by, extend_upto,
2763 : buffers, &extend_by);
2764 : else
2765 254977 : first_block = ExtendBufferedRelShared(bmr, fork, strategy, flags,
2766 : extend_by, extend_upto,
2767 : buffers, &extend_by);
2768 269889 : *extended_by = extend_by;
2769 :
2770 : TRACE_POSTGRESQL_BUFFER_EXTEND_DONE(fork,
2771 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2772 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2773 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2774 : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2775 : *extended_by,
2776 : first_block);
2777 :
2778 269889 : return first_block;
2779 : }
2780 :
2781 : /*
2782 : * Implementation of ExtendBufferedRelBy() and ExtendBufferedRelTo() for
2783 : * shared buffers.
2784 : */
2785 : static BlockNumber
2786 254977 : ExtendBufferedRelShared(BufferManagerRelation bmr,
2787 : ForkNumber fork,
2788 : BufferAccessStrategy strategy,
2789 : uint32 flags,
2790 : uint32 extend_by,
2791 : BlockNumber extend_upto,
2792 : Buffer *buffers,
2793 : uint32 *extended_by)
2794 : {
2795 : BlockNumber first_block;
2796 254977 : IOContext io_context = IOContextForStrategy(strategy);
2797 : instr_time io_start;
2798 :
2799 254977 : LimitAdditionalPins(&extend_by);
2800 :
2801 : /*
2802 : * Acquire victim buffers for extension without holding extension lock.
2803 : * Writing out victim buffers is the most expensive part of extending the
2804 : * relation, particularly when doing so requires WAL flushes. Zeroing out
2805 : * the buffers is also quite expensive, so do that before holding the
2806 : * extension lock as well.
2807 : *
2808 : * These pages are pinned by us and not valid. While we hold the pin they
2809 : * can't be acquired as victim buffers by another backend.
2810 : */
2811 536738 : for (uint32 i = 0; i < extend_by; i++)
2812 : {
2813 : Block buf_block;
2814 :
2815 281761 : buffers[i] = GetVictimBuffer(strategy, io_context);
2816 281761 : buf_block = BufHdrGetBlock(GetBufferDescriptor(buffers[i] - 1));
2817 :
2818 : /* new buffers are zero-filled */
2819 281761 : MemSet(buf_block, 0, BLCKSZ);
2820 : }
2821 :
2822 : /*
2823 : * Lock relation against concurrent extensions, unless requested not to.
2824 : *
2825 : * We use the same extension lock for all forks. That's unnecessarily
2826 : * restrictive, but currently extensions for forks don't happen often
2827 : * enough to make it worth locking more granularly.
2828 : *
2829 : * Note that another backend might have extended the relation by the time
2830 : * we get the lock.
2831 : */
2832 254977 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2833 197671 : LockRelationForExtension(bmr.rel, ExclusiveLock);
2834 :
2835 : /*
2836 : * If requested, invalidate size cache, so that smgrnblocks asks the
2837 : * kernel.
2838 : */
2839 254977 : if (flags & EB_CLEAR_SIZE_CACHE)
2840 9907 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
2841 :
2842 254977 : first_block = smgrnblocks(BMR_GET_SMGR(bmr), fork);
2843 :
2844 : /*
2845 : * Now that we have the accurate relation size, check if the caller wants
2846 : * us to extend to only up to a specific size. If there were concurrent
2847 : * extensions, we might have acquired too many buffers and need to release
2848 : * them.
2849 : */
2850 254977 : if (extend_upto != InvalidBlockNumber)
2851 : {
2852 59398 : uint32 orig_extend_by = extend_by;
2853 :
2854 59398 : if (first_block > extend_upto)
2855 0 : extend_by = 0;
2856 59398 : else if ((uint64) first_block + extend_by > extend_upto)
2857 8 : extend_by = extend_upto - first_block;
2858 :
2859 59420 : for (uint32 i = extend_by; i < orig_extend_by; i++)
2860 : {
2861 22 : BufferDesc *buf_hdr = GetBufferDescriptor(buffers[i] - 1);
2862 :
2863 22 : UnpinBuffer(buf_hdr);
2864 : }
2865 :
2866 59398 : if (extend_by == 0)
2867 : {
2868 8 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2869 8 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2870 8 : *extended_by = extend_by;
2871 8 : return first_block;
2872 : }
2873 : }
2874 :
2875 : /* Fail if relation is already at maximum possible length */
2876 254969 : if ((uint64) first_block + extend_by >= MaxBlockNumber)
2877 0 : ereport(ERROR,
2878 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
2879 : errmsg("cannot extend relation %s beyond %u blocks",
2880 : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str,
2881 : MaxBlockNumber)));
2882 :
2883 : /*
2884 : * Insert buffers into buffer table, mark as IO_IN_PROGRESS.
2885 : *
2886 : * This needs to happen before we extend the relation, because as soon as
2887 : * we do, other backends can start to read in those pages.
2888 : */
2889 536708 : for (uint32 i = 0; i < extend_by; i++)
2890 : {
2891 281739 : Buffer victim_buf = buffers[i];
2892 281739 : BufferDesc *victim_buf_hdr = GetBufferDescriptor(victim_buf - 1);
2893 : BufferTag tag;
2894 : uint32 hash;
2895 : LWLock *partition_lock;
2896 : int existing_id;
2897 :
2898 : /* in case we need to pin an existing buffer below */
2899 281739 : ResourceOwnerEnlarge(CurrentResourceOwner);
2900 281739 : ReservePrivateRefCountEntry();
2901 :
2902 281739 : InitBufferTag(&tag, &BMR_GET_SMGR(bmr)->smgr_rlocator.locator, fork,
2903 : first_block + i);
2904 281739 : hash = BufTableHashCode(&tag);
2905 281739 : partition_lock = BufMappingPartitionLock(hash);
2906 :
2907 281739 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2908 :
2909 281739 : existing_id = BufTableInsert(&tag, hash, victim_buf_hdr->buf_id);
2910 :
2911 : /*
2912 : * We get here only in the corner case where we are trying to extend
2913 : * the relation but we found a pre-existing buffer. This can happen
2914 : * because a prior attempt at extending the relation failed, and
2915 : * because mdread doesn't complain about reads beyond EOF (when
2916 : * zero_damaged_pages is ON) and so a previous attempt to read a block
2917 : * beyond EOF could have left a "valid" zero-filled buffer.
2918 : *
2919 : * This has also been observed when relation was overwritten by
2920 : * external process. Since the legitimate cases should always have
2921 : * left a zero-filled buffer, complain if not PageIsNew.
2922 : */
2923 281739 : if (existing_id >= 0)
2924 : {
2925 0 : BufferDesc *existing_hdr = GetBufferDescriptor(existing_id);
2926 : Block buf_block;
2927 : bool valid;
2928 :
2929 : /*
2930 : * Pin the existing buffer before releasing the partition lock,
2931 : * preventing it from being evicted.
2932 : */
2933 0 : valid = PinBuffer(existing_hdr, strategy, false);
2934 :
2935 0 : LWLockRelease(partition_lock);
2936 0 : UnpinBuffer(victim_buf_hdr);
2937 :
2938 0 : buffers[i] = BufferDescriptorGetBuffer(existing_hdr);
2939 0 : buf_block = BufHdrGetBlock(existing_hdr);
2940 :
2941 0 : if (valid && !PageIsNew((Page) buf_block))
2942 0 : ereport(ERROR,
2943 : (errmsg("unexpected data beyond EOF in block %u of relation \"%s\"",
2944 : existing_hdr->tag.blockNum,
2945 : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str)));
2946 :
2947 : /*
2948 : * We *must* do smgr[zero]extend before succeeding, else the page
2949 : * will not be reserved by the kernel, and the next P_NEW call
2950 : * will decide to return the same page. Clear the BM_VALID bit,
2951 : * do StartSharedBufferIO() and proceed.
2952 : *
2953 : * Loop to handle the very small possibility that someone re-sets
2954 : * BM_VALID between our clearing it and StartSharedBufferIO
2955 : * inspecting it.
2956 : */
2957 : while (true)
2958 0 : {
2959 : StartBufferIOResult sbres;
2960 :
2961 0 : pg_atomic_fetch_and_u64(&existing_hdr->state, ~BM_VALID);
2962 :
2963 0 : sbres = StartSharedBufferIO(existing_hdr, true, true, NULL);
2964 :
2965 0 : if (sbres != BUFFER_IO_ALREADY_DONE)
2966 0 : break;
2967 : }
2968 : }
2969 : else
2970 : {
2971 : uint64 buf_state;
2972 281739 : uint64 set_bits = 0;
2973 :
2974 281739 : buf_state = LockBufHdr(victim_buf_hdr);
2975 :
2976 : /* some sanity checks while we hold the buffer header lock */
2977 : Assert(!(buf_state & (BM_VALID | BM_TAG_VALID | BM_DIRTY)));
2978 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2979 :
2980 281739 : victim_buf_hdr->tag = tag;
2981 :
2982 281739 : set_bits |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2983 281739 : if (bmr.relpersistence == RELPERSISTENCE_PERMANENT || fork == INIT_FORKNUM)
2984 276354 : set_bits |= BM_PERMANENT;
2985 :
2986 281739 : UnlockBufHdrExt(victim_buf_hdr, buf_state,
2987 : set_bits, 0,
2988 : 0);
2989 :
2990 281739 : LWLockRelease(partition_lock);
2991 :
2992 : /* XXX: could combine the locked operations in it with the above */
2993 281739 : StartSharedBufferIO(victim_buf_hdr, true, true, NULL);
2994 : }
2995 : }
2996 :
2997 254969 : io_start = pgstat_prepare_io_time(track_io_timing);
2998 :
2999 : /*
3000 : * Note: if smgrzeroextend fails, we will end up with buffers that are
3001 : * allocated but not marked BM_VALID. The next relation extension will
3002 : * still select the same block number (because the relation didn't get any
3003 : * longer on disk) and so future attempts to extend the relation will find
3004 : * the same buffers (if they have not been recycled) but come right back
3005 : * here to try smgrzeroextend again.
3006 : *
3007 : * We don't need to set checksum for all-zero pages.
3008 : */
3009 254969 : smgrzeroextend(BMR_GET_SMGR(bmr), fork, first_block, extend_by, false);
3010 :
3011 : /*
3012 : * Release the file-extension lock; it's now OK for someone else to extend
3013 : * the relation some more.
3014 : *
3015 : * We remove IO_IN_PROGRESS after this, as waking up waiting backends can
3016 : * take noticeable time.
3017 : */
3018 254969 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
3019 197663 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
3020 :
3021 254969 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context, IOOP_EXTEND,
3022 254969 : io_start, 1, extend_by * BLCKSZ);
3023 :
3024 : /* Set BM_VALID, terminate IO, and wake up any waiters */
3025 536708 : for (uint32 i = 0; i < extend_by; i++)
3026 : {
3027 281739 : Buffer buf = buffers[i];
3028 281739 : BufferDesc *buf_hdr = GetBufferDescriptor(buf - 1);
3029 281739 : bool lock = false;
3030 :
3031 281739 : if (flags & EB_LOCK_FIRST && i == 0)
3032 195251 : lock = true;
3033 86488 : else if (flags & EB_LOCK_TARGET)
3034 : {
3035 : Assert(extend_upto != InvalidBlockNumber);
3036 48352 : if (first_block + i + 1 == extend_upto)
3037 47794 : lock = true;
3038 : }
3039 :
3040 281739 : if (lock)
3041 243045 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
3042 :
3043 281739 : TerminateBufferIO(buf_hdr, false, BM_VALID, true, false);
3044 : }
3045 :
3046 254969 : pgBufferUsage.shared_blks_written += extend_by;
3047 :
3048 254969 : *extended_by = extend_by;
3049 :
3050 254969 : return first_block;
3051 : }
3052 :
3053 : /*
3054 : * BufferIsLockedByMe
3055 : *
3056 : * Checks if this backend has the buffer locked in any mode.
3057 : *
3058 : * Buffer must be pinned.
3059 : */
3060 : bool
3061 0 : BufferIsLockedByMe(Buffer buffer)
3062 : {
3063 : BufferDesc *bufHdr;
3064 :
3065 : Assert(BufferIsPinned(buffer));
3066 :
3067 0 : if (BufferIsLocal(buffer))
3068 : {
3069 : /* Content locks are not maintained for local buffers. */
3070 0 : return true;
3071 : }
3072 : else
3073 : {
3074 0 : bufHdr = GetBufferDescriptor(buffer - 1);
3075 0 : return BufferLockHeldByMe(bufHdr);
3076 : }
3077 : }
3078 :
3079 : /*
3080 : * BufferIsLockedByMeInMode
3081 : *
3082 : * Checks if this backend has the buffer locked in the specified mode.
3083 : *
3084 : * Buffer must be pinned.
3085 : */
3086 : bool
3087 0 : BufferIsLockedByMeInMode(Buffer buffer, BufferLockMode mode)
3088 : {
3089 : BufferDesc *bufHdr;
3090 :
3091 : Assert(BufferIsPinned(buffer));
3092 :
3093 0 : if (BufferIsLocal(buffer))
3094 : {
3095 : /* Content locks are not maintained for local buffers. */
3096 0 : return true;
3097 : }
3098 : else
3099 : {
3100 0 : bufHdr = GetBufferDescriptor(buffer - 1);
3101 0 : return BufferLockHeldByMeInMode(bufHdr, mode);
3102 : }
3103 : }
3104 :
3105 : /*
3106 : * BufferIsDirty
3107 : *
3108 : * Checks if buffer is already dirty.
3109 : *
3110 : * Buffer must be pinned and [share-]exclusive-locked. (Without such a lock,
3111 : * the result may be stale before it's returned.)
3112 : */
3113 : bool
3114 29093 : BufferIsDirty(Buffer buffer)
3115 : {
3116 : BufferDesc *bufHdr;
3117 :
3118 : Assert(BufferIsPinned(buffer));
3119 :
3120 29093 : if (BufferIsLocal(buffer))
3121 : {
3122 9521 : int bufid = -buffer - 1;
3123 :
3124 9521 : bufHdr = GetLocalBufferDescriptor(bufid);
3125 : /* Content locks are not maintained for local buffers. */
3126 : }
3127 : else
3128 : {
3129 19572 : bufHdr = GetBufferDescriptor(buffer - 1);
3130 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_SHARE_EXCLUSIVE) ||
3131 : BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
3132 : }
3133 :
3134 29093 : return pg_atomic_read_u64(&bufHdr->state) & BM_DIRTY;
3135 : }
3136 :
3137 : /*
3138 : * MarkBufferDirty
3139 : *
3140 : * Marks buffer contents as dirty (actual write happens later).
3141 : *
3142 : * Buffer must be pinned and exclusive-locked. (If caller does not hold
3143 : * exclusive lock, then somebody could be in process of writing the buffer,
3144 : * leading to risk of bad data written to disk.)
3145 : */
3146 : void
3147 34013343 : MarkBufferDirty(Buffer buffer)
3148 : {
3149 : BufferDesc *bufHdr;
3150 : uint64 buf_state;
3151 : uint64 old_buf_state;
3152 :
3153 34013343 : if (!BufferIsValid(buffer))
3154 0 : elog(ERROR, "bad buffer ID: %d", buffer);
3155 :
3156 34013343 : if (BufferIsLocal(buffer))
3157 : {
3158 1579744 : MarkLocalBufferDirty(buffer);
3159 1579744 : return;
3160 : }
3161 :
3162 32433599 : bufHdr = GetBufferDescriptor(buffer - 1);
3163 :
3164 : Assert(BufferIsPinned(buffer));
3165 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
3166 :
3167 : /*
3168 : * NB: We have to wait for the buffer header spinlock to be not held, as
3169 : * TerminateBufferIO() relies on the spinlock.
3170 : */
3171 32433599 : old_buf_state = pg_atomic_read_u64(&bufHdr->state);
3172 : for (;;)
3173 : {
3174 32433976 : if (old_buf_state & BM_LOCKED)
3175 541 : old_buf_state = WaitBufHdrUnlocked(bufHdr);
3176 :
3177 32433976 : buf_state = old_buf_state;
3178 :
3179 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
3180 32433976 : buf_state |= BM_DIRTY;
3181 :
3182 32433976 : if (pg_atomic_compare_exchange_u64(&bufHdr->state, &old_buf_state,
3183 : buf_state))
3184 32433599 : break;
3185 : }
3186 :
3187 : /*
3188 : * If the buffer was not dirty already, do vacuum accounting.
3189 : */
3190 32433599 : if (!(old_buf_state & BM_DIRTY))
3191 : {
3192 807326 : pgBufferUsage.shared_blks_dirtied++;
3193 807326 : if (VacuumCostActive)
3194 7003 : VacuumCostBalance += VacuumCostPageDirty;
3195 : }
3196 : }
3197 :
3198 : /*
3199 : * ReleaseAndReadBuffer -- combine ReleaseBuffer() and ReadBuffer()
3200 : *
3201 : * Formerly, this saved one cycle of acquiring/releasing the BufMgrLock
3202 : * compared to calling the two routines separately. Now it's mainly just
3203 : * a convenience function. However, if the passed buffer is valid and
3204 : * already contains the desired block, we just return it as-is; and that
3205 : * does save considerable work compared to a full release and reacquire.
3206 : *
3207 : * Note: it is OK to pass buffer == InvalidBuffer, indicating that no old
3208 : * buffer actually needs to be released. This case is the same as ReadBuffer,
3209 : * but can save some tests in the caller.
3210 : */
3211 : Buffer
3212 3646 : ReleaseAndReadBuffer(Buffer buffer,
3213 : Relation relation,
3214 : BlockNumber blockNum)
3215 : {
3216 3646 : ForkNumber forkNum = MAIN_FORKNUM;
3217 : BufferDesc *bufHdr;
3218 :
3219 3646 : if (BufferIsValid(buffer))
3220 : {
3221 : Assert(BufferIsPinned(buffer));
3222 3646 : if (BufferIsLocal(buffer))
3223 : {
3224 50 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
3225 50 : if (bufHdr->tag.blockNum == blockNum &&
3226 0 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3227 0 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3228 0 : return buffer;
3229 50 : UnpinLocalBuffer(buffer);
3230 : }
3231 : else
3232 : {
3233 3596 : bufHdr = GetBufferDescriptor(buffer - 1);
3234 : /* we have pin, so it's ok to examine tag without spinlock */
3235 3596 : if (bufHdr->tag.blockNum == blockNum &&
3236 0 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3237 0 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3238 0 : return buffer;
3239 3596 : UnpinBuffer(bufHdr);
3240 : }
3241 : }
3242 :
3243 3646 : return ReadBuffer(relation, blockNum);
3244 : }
3245 :
3246 : /*
3247 : * PinBuffer -- make buffer unavailable for replacement.
3248 : *
3249 : * For the default access strategy, the buffer's usage_count is incremented
3250 : * when we first pin it; for other strategies we just make sure the usage_count
3251 : * isn't zero. (The idea of the latter is that we don't want synchronized
3252 : * heap scans to inflate the count, but we need it to not be zero to discourage
3253 : * other backends from stealing buffers from our ring. As long as we cycle
3254 : * through the ring faster than the global clock-sweep cycles, buffers in
3255 : * our ring won't be chosen as victims for replacement by other backends.)
3256 : *
3257 : * This should be applied only to shared buffers, never local ones.
3258 : *
3259 : * Since buffers are pinned/unpinned very frequently, pin buffers without
3260 : * taking the buffer header lock; instead update the state variable in loop of
3261 : * CAS operations. Hopefully it's just a single CAS.
3262 : *
3263 : * Note that ResourceOwnerEnlarge() and ReservePrivateRefCountEntry()
3264 : * must have been done already.
3265 : *
3266 : * Returns true if buffer is BM_VALID, else false. This provision allows
3267 : * some callers to avoid an extra spinlock cycle. If skip_if_not_valid is
3268 : * true, then a false return value also indicates that the buffer was
3269 : * (recently) invalid and has not been pinned.
3270 : */
3271 : static bool
3272 81135675 : PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
3273 : bool skip_if_not_valid)
3274 : {
3275 81135675 : Buffer b = BufferDescriptorGetBuffer(buf);
3276 : bool result;
3277 : PrivateRefCountEntry *ref;
3278 :
3279 : Assert(!BufferIsLocal(b));
3280 : Assert(ReservedRefCountSlot != -1);
3281 :
3282 81135675 : ref = GetPrivateRefCountEntry(b, true);
3283 :
3284 81135675 : if (ref == NULL)
3285 : {
3286 : uint64 buf_state;
3287 : uint64 old_buf_state;
3288 :
3289 73815329 : old_buf_state = pg_atomic_read_u64(&buf->state);
3290 : for (;;)
3291 : {
3292 73836856 : if (unlikely(skip_if_not_valid && !(old_buf_state & BM_VALID)))
3293 6 : return false;
3294 :
3295 : /*
3296 : * We're not allowed to increase the refcount while the buffer
3297 : * header spinlock is held. Wait for the lock to be released.
3298 : */
3299 73836850 : if (unlikely(old_buf_state & BM_LOCKED))
3300 : {
3301 290 : old_buf_state = WaitBufHdrUnlocked(buf);
3302 :
3303 : /* perform checks at the top of the loop again */
3304 290 : continue;
3305 : }
3306 :
3307 73836560 : buf_state = old_buf_state;
3308 :
3309 : /* increase refcount */
3310 73836560 : buf_state += BUF_REFCOUNT_ONE;
3311 :
3312 73836560 : if (strategy == NULL)
3313 : {
3314 : /* Default case: increase usagecount unless already max. */
3315 73070550 : if (BUF_STATE_GET_USAGECOUNT(buf_state) < BM_MAX_USAGE_COUNT)
3316 3772112 : buf_state += BUF_USAGECOUNT_ONE;
3317 : }
3318 : else
3319 : {
3320 : /*
3321 : * Ring buffers shouldn't evict others from pool. Thus we
3322 : * don't make usagecount more than 1.
3323 : */
3324 766010 : if (BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3325 38255 : buf_state += BUF_USAGECOUNT_ONE;
3326 : }
3327 :
3328 73836560 : if (pg_atomic_compare_exchange_u64(&buf->state, &old_buf_state,
3329 : buf_state))
3330 : {
3331 73815323 : result = (buf_state & BM_VALID) != 0;
3332 :
3333 73815323 : TrackNewBufferPin(b);
3334 73815323 : break;
3335 : }
3336 : }
3337 : }
3338 : else
3339 : {
3340 : /*
3341 : * If we previously pinned the buffer, it is likely to be valid, but
3342 : * it may not be if StartReadBuffers() was called and
3343 : * WaitReadBuffers() hasn't been called yet. We'll check by loading
3344 : * the flags without locking. This is racy, but it's OK to return
3345 : * false spuriously: when WaitReadBuffers() calls StartBufferIO(),
3346 : * it'll see that it's now valid.
3347 : *
3348 : * Note: We deliberately avoid a Valgrind client request here.
3349 : * Individual access methods can optionally superimpose buffer page
3350 : * client requests on top of our client requests to enforce that
3351 : * buffers are only accessed while locked (and pinned). It's possible
3352 : * that the buffer page is legitimately non-accessible here. We
3353 : * cannot meddle with that.
3354 : */
3355 7320346 : result = (pg_atomic_read_u64(&buf->state) & BM_VALID) != 0;
3356 :
3357 : Assert(ref->data.refcount > 0);
3358 7320346 : ref->data.refcount++;
3359 7320346 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
3360 : }
3361 :
3362 81135669 : return result;
3363 : }
3364 :
3365 : /*
3366 : * PinBuffer_Locked -- as above, but caller already locked the buffer header.
3367 : * The spinlock is released before return.
3368 : *
3369 : * As this function is called with the spinlock held, the caller has to
3370 : * previously call ReservePrivateRefCountEntry() and
3371 : * ResourceOwnerEnlarge(CurrentResourceOwner);
3372 : *
3373 : * Currently, no callers of this function want to modify the buffer's
3374 : * usage_count at all, so there's no need for a strategy parameter.
3375 : * Also we don't bother with a BM_VALID test (the caller could check that for
3376 : * itself).
3377 : *
3378 : * Also all callers only ever use this function when it's known that the
3379 : * buffer can't have a preexisting pin by this backend. That allows us to skip
3380 : * searching the private refcount array & hash, which is a boon, because the
3381 : * spinlock is still held.
3382 : *
3383 : * Note: use of this routine is frequently mandatory, not just an optimization
3384 : * to save a spin lock/unlock cycle, because we need to pin a buffer before
3385 : * its state can change under us.
3386 : */
3387 : static void
3388 351884 : PinBuffer_Locked(BufferDesc *buf)
3389 : {
3390 : uint64 old_buf_state;
3391 :
3392 : /*
3393 : * As explained, We don't expect any preexisting pins. That allows us to
3394 : * manipulate the PrivateRefCount after releasing the spinlock
3395 : */
3396 : Assert(GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf), false) == NULL);
3397 :
3398 : /*
3399 : * Since we hold the buffer spinlock, we can update the buffer state and
3400 : * release the lock in one operation.
3401 : */
3402 351884 : old_buf_state = pg_atomic_read_u64(&buf->state);
3403 :
3404 351884 : UnlockBufHdrExt(buf, old_buf_state,
3405 : 0, 0, 1);
3406 :
3407 351884 : TrackNewBufferPin(BufferDescriptorGetBuffer(buf));
3408 351884 : }
3409 :
3410 : /*
3411 : * Support for waking up another backend that is waiting for the cleanup lock
3412 : * to be released using BM_PIN_COUNT_WAITER.
3413 : *
3414 : * See LockBufferForCleanup().
3415 : *
3416 : * Expected to be called just after releasing a buffer pin (in a BufferDesc,
3417 : * not just reducing the backend-local pincount for the buffer).
3418 : */
3419 : static void
3420 10 : WakePinCountWaiter(BufferDesc *buf)
3421 : {
3422 : /*
3423 : * Acquire the buffer header lock, re-check that there's a waiter. Another
3424 : * backend could have unpinned this buffer, and already woken up the
3425 : * waiter.
3426 : *
3427 : * There's no danger of the buffer being replaced after we unpinned it
3428 : * above, as it's pinned by the waiter. The waiter removes
3429 : * BM_PIN_COUNT_WAITER if it stops waiting for a reason other than this
3430 : * backend waking it up.
3431 : */
3432 10 : uint64 buf_state = LockBufHdr(buf);
3433 :
3434 10 : if ((buf_state & BM_PIN_COUNT_WAITER) &&
3435 10 : BUF_STATE_GET_REFCOUNT(buf_state) == 1)
3436 10 : {
3437 : /* we just released the last pin other than the waiter's */
3438 10 : int wait_backend_pgprocno = buf->wait_backend_pgprocno;
3439 :
3440 10 : UnlockBufHdrExt(buf, buf_state,
3441 : 0, BM_PIN_COUNT_WAITER,
3442 : 0);
3443 10 : ProcSendSignal(wait_backend_pgprocno);
3444 : }
3445 : else
3446 0 : UnlockBufHdr(buf);
3447 10 : }
3448 :
3449 : /*
3450 : * UnpinBuffer -- make buffer available for replacement.
3451 : *
3452 : * This should be applied only to shared buffers, never local ones. This
3453 : * always adjusts CurrentResourceOwner.
3454 : */
3455 : static void
3456 50024067 : UnpinBuffer(BufferDesc *buf)
3457 : {
3458 50024067 : Buffer b = BufferDescriptorGetBuffer(buf);
3459 :
3460 50024067 : ResourceOwnerForgetBuffer(CurrentResourceOwner, b);
3461 50024067 : UnpinBufferNoOwner(buf);
3462 50024067 : }
3463 :
3464 : static void
3465 50030608 : UnpinBufferNoOwner(BufferDesc *buf)
3466 : {
3467 : PrivateRefCountEntry *ref;
3468 50030608 : Buffer b = BufferDescriptorGetBuffer(buf);
3469 :
3470 : Assert(!BufferIsLocal(b));
3471 :
3472 : /* not moving as we're likely deleting it soon anyway */
3473 50030608 : ref = GetPrivateRefCountEntry(b, false);
3474 : Assert(ref != NULL);
3475 : Assert(ref->data.refcount > 0);
3476 50030608 : ref->data.refcount--;
3477 50030608 : if (ref->data.refcount == 0)
3478 : {
3479 : uint64 old_buf_state;
3480 :
3481 : /*
3482 : * Mark buffer non-accessible to Valgrind.
3483 : *
3484 : * Note that the buffer may have already been marked non-accessible
3485 : * within access method code that enforces that buffers are only
3486 : * accessed while a buffer lock is held.
3487 : */
3488 : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
3489 :
3490 : /*
3491 : * I'd better not still hold the buffer content lock. Can't use
3492 : * BufferIsLockedByMe(), as that asserts the buffer is pinned.
3493 : */
3494 : Assert(!BufferLockHeldByMe(buf));
3495 :
3496 : /* decrement the shared reference count */
3497 30239717 : old_buf_state = pg_atomic_fetch_sub_u64(&buf->state, BUF_REFCOUNT_ONE);
3498 :
3499 : /* Support LockBufferForCleanup() */
3500 30239717 : if (old_buf_state & BM_PIN_COUNT_WAITER)
3501 10 : WakePinCountWaiter(buf);
3502 :
3503 30239717 : ForgetPrivateRefCountEntry(ref);
3504 : }
3505 50030608 : }
3506 :
3507 : /*
3508 : * Set up backend-local tracking of a buffer pinned the first time by this
3509 : * backend.
3510 : */
3511 : inline void
3512 76449731 : TrackNewBufferPin(Buffer buf)
3513 : {
3514 : PrivateRefCountEntry *ref;
3515 :
3516 76449731 : ref = NewPrivateRefCountEntry(buf);
3517 76449731 : ref->data.refcount++;
3518 :
3519 76449731 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buf);
3520 :
3521 : /*
3522 : * This is the first pin for this page by this backend, mark its page as
3523 : * defined to valgrind. While the page contents might not actually be
3524 : * valid yet, we don't currently guarantee that such pages are marked
3525 : * undefined or non-accessible.
3526 : *
3527 : * It's not necessarily the prettiest to do this here, but otherwise we'd
3528 : * need this block of code in multiple places.
3529 : */
3530 : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(GetBufferDescriptor(buf - 1)),
3531 : BLCKSZ);
3532 76449731 : }
3533 :
3534 : #define ST_SORT sort_checkpoint_bufferids
3535 : #define ST_ELEMENT_TYPE CkptSortItem
3536 : #define ST_COMPARE(a, b) ckpt_buforder_comparator(a, b)
3537 : #define ST_SCOPE static
3538 : #define ST_DEFINE
3539 : #include "lib/sort_template.h"
3540 :
3541 : /*
3542 : * BufferSync -- Write out all dirty buffers in the pool.
3543 : *
3544 : * This is called at checkpoint time to write out all dirty shared buffers.
3545 : * The checkpoint request flags should be passed in. If CHECKPOINT_FAST is
3546 : * set, we disable delays between writes; if CHECKPOINT_IS_SHUTDOWN,
3547 : * CHECKPOINT_END_OF_RECOVERY or CHECKPOINT_FLUSH_UNLOGGED is set, we write
3548 : * even unlogged buffers, which are otherwise skipped. The remaining flags
3549 : * currently have no effect here.
3550 : */
3551 : static void
3552 1931 : BufferSync(int flags)
3553 : {
3554 : uint64 buf_state;
3555 : int buf_id;
3556 : int num_to_scan;
3557 : int num_spaces;
3558 : int num_processed;
3559 : int num_written;
3560 1931 : CkptTsStatus *per_ts_stat = NULL;
3561 : Oid last_tsid;
3562 : binaryheap *ts_heap;
3563 : int i;
3564 1931 : uint64 mask = BM_DIRTY;
3565 : WritebackContext wb_context;
3566 :
3567 : /*
3568 : * Unless this is a shutdown checkpoint or we have been explicitly told,
3569 : * we write only permanent, dirty buffers. But at shutdown or end of
3570 : * recovery, we write all dirty buffers.
3571 : */
3572 1931 : if (!((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
3573 : CHECKPOINT_FLUSH_UNLOGGED))))
3574 1056 : mask |= BM_PERMANENT;
3575 :
3576 : /*
3577 : * Loop over all buffers, and mark the ones that need to be written with
3578 : * BM_CHECKPOINT_NEEDED. Count them as we go (num_to_scan), so that we
3579 : * can estimate how much work needs to be done.
3580 : *
3581 : * This allows us to write only those pages that were dirty when the
3582 : * checkpoint began, and not those that get dirtied while it proceeds.
3583 : * Whenever a page with BM_CHECKPOINT_NEEDED is written out, either by us
3584 : * later in this function, or by normal backends or the bgwriter cleaning
3585 : * scan, the flag is cleared. Any buffer dirtied after this point won't
3586 : * have the flag set.
3587 : *
3588 : * Note that if we fail to write some buffer, we may leave buffers with
3589 : * BM_CHECKPOINT_NEEDED still set. This is OK since any such buffer would
3590 : * certainly need to be written for the next checkpoint attempt, too.
3591 : */
3592 1931 : num_to_scan = 0;
3593 14033883 : for (buf_id = 0; buf_id < NBuffers; buf_id++)
3594 : {
3595 14031952 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3596 14031952 : uint64 set_bits = 0;
3597 :
3598 : /*
3599 : * Header spinlock is enough to examine BM_DIRTY, see comment in
3600 : * SyncOneBuffer.
3601 : */
3602 14031952 : buf_state = LockBufHdr(bufHdr);
3603 :
3604 14031952 : if ((buf_state & mask) == mask)
3605 : {
3606 : CkptSortItem *item;
3607 :
3608 340638 : set_bits = BM_CHECKPOINT_NEEDED;
3609 :
3610 340638 : item = &CkptBufferIds[num_to_scan++];
3611 340638 : item->buf_id = buf_id;
3612 340638 : item->tsId = bufHdr->tag.spcOid;
3613 340638 : item->relNumber = BufTagGetRelNumber(&bufHdr->tag);
3614 340638 : item->forkNum = BufTagGetForkNum(&bufHdr->tag);
3615 340638 : item->blockNum = bufHdr->tag.blockNum;
3616 : }
3617 :
3618 14031952 : UnlockBufHdrExt(bufHdr, buf_state,
3619 : set_bits, 0,
3620 : 0);
3621 :
3622 : /* Check for barrier events in case NBuffers is large. */
3623 14031952 : if (ProcSignalBarrierPending)
3624 2 : ProcessProcSignalBarrier();
3625 : }
3626 :
3627 1931 : if (num_to_scan == 0)
3628 737 : return; /* nothing to do */
3629 :
3630 1194 : WritebackContextInit(&wb_context, &checkpoint_flush_after);
3631 :
3632 : TRACE_POSTGRESQL_BUFFER_SYNC_START(NBuffers, num_to_scan);
3633 :
3634 : /*
3635 : * Sort buffers that need to be written to reduce the likelihood of random
3636 : * IO. The sorting is also important for the implementation of balancing
3637 : * writes between tablespaces. Without balancing writes we'd potentially
3638 : * end up writing to the tablespaces one-by-one; possibly overloading the
3639 : * underlying system.
3640 : */
3641 1194 : sort_checkpoint_bufferids(CkptBufferIds, num_to_scan);
3642 :
3643 1194 : num_spaces = 0;
3644 :
3645 : /*
3646 : * Allocate progress status for each tablespace with buffers that need to
3647 : * be flushed. This requires the to-be-flushed array to be sorted.
3648 : */
3649 1194 : last_tsid = InvalidOid;
3650 341832 : for (i = 0; i < num_to_scan; i++)
3651 : {
3652 : CkptTsStatus *s;
3653 : Oid cur_tsid;
3654 :
3655 340638 : cur_tsid = CkptBufferIds[i].tsId;
3656 :
3657 : /*
3658 : * Grow array of per-tablespace status structs, every time a new
3659 : * tablespace is found.
3660 : */
3661 340638 : if (last_tsid == InvalidOid || last_tsid != cur_tsid)
3662 1829 : {
3663 : Size sz;
3664 :
3665 1829 : num_spaces++;
3666 :
3667 : /*
3668 : * Not worth adding grow-by-power-of-2 logic here - even with a
3669 : * few hundred tablespaces this should be fine.
3670 : */
3671 1829 : sz = sizeof(CkptTsStatus) * num_spaces;
3672 :
3673 1829 : if (per_ts_stat == NULL)
3674 1194 : per_ts_stat = (CkptTsStatus *) palloc(sz);
3675 : else
3676 635 : per_ts_stat = (CkptTsStatus *) repalloc(per_ts_stat, sz);
3677 :
3678 1829 : s = &per_ts_stat[num_spaces - 1];
3679 1829 : memset(s, 0, sizeof(*s));
3680 1829 : s->tsId = cur_tsid;
3681 :
3682 : /*
3683 : * The first buffer in this tablespace. As CkptBufferIds is sorted
3684 : * by tablespace all (s->num_to_scan) buffers in this tablespace
3685 : * will follow afterwards.
3686 : */
3687 1829 : s->index = i;
3688 :
3689 : /*
3690 : * progress_slice will be determined once we know how many buffers
3691 : * are in each tablespace, i.e. after this loop.
3692 : */
3693 :
3694 1829 : last_tsid = cur_tsid;
3695 : }
3696 : else
3697 : {
3698 338809 : s = &per_ts_stat[num_spaces - 1];
3699 : }
3700 :
3701 340638 : s->num_to_scan++;
3702 :
3703 : /* Check for barrier events. */
3704 340638 : if (ProcSignalBarrierPending)
3705 0 : ProcessProcSignalBarrier();
3706 : }
3707 :
3708 : Assert(num_spaces > 0);
3709 :
3710 : /*
3711 : * Build a min-heap over the write-progress in the individual tablespaces,
3712 : * and compute how large a portion of the total progress a single
3713 : * processed buffer is.
3714 : */
3715 1194 : ts_heap = binaryheap_allocate(num_spaces,
3716 : ts_ckpt_progress_comparator,
3717 : NULL);
3718 :
3719 3023 : for (i = 0; i < num_spaces; i++)
3720 : {
3721 1829 : CkptTsStatus *ts_stat = &per_ts_stat[i];
3722 :
3723 1829 : ts_stat->progress_slice = (float8) num_to_scan / ts_stat->num_to_scan;
3724 :
3725 1829 : binaryheap_add_unordered(ts_heap, PointerGetDatum(ts_stat));
3726 : }
3727 :
3728 1194 : binaryheap_build(ts_heap);
3729 :
3730 : /*
3731 : * Iterate through to-be-checkpointed buffers and write the ones (still)
3732 : * marked with BM_CHECKPOINT_NEEDED. The writes are balanced between
3733 : * tablespaces; otherwise the sorting would lead to only one tablespace
3734 : * receiving writes at a time, making inefficient use of the hardware.
3735 : */
3736 1194 : num_processed = 0;
3737 1194 : num_written = 0;
3738 341832 : while (!binaryheap_empty(ts_heap))
3739 : {
3740 340638 : BufferDesc *bufHdr = NULL;
3741 : CkptTsStatus *ts_stat = (CkptTsStatus *)
3742 340638 : DatumGetPointer(binaryheap_first(ts_heap));
3743 :
3744 340638 : buf_id = CkptBufferIds[ts_stat->index].buf_id;
3745 : Assert(buf_id != -1);
3746 :
3747 340638 : bufHdr = GetBufferDescriptor(buf_id);
3748 :
3749 340638 : num_processed++;
3750 :
3751 : /*
3752 : * We don't need to acquire the lock here, because we're only looking
3753 : * at a single bit. It's possible that someone else writes the buffer
3754 : * and clears the flag right after we check, but that doesn't matter
3755 : * since SyncOneBuffer will then do nothing. However, there is a
3756 : * further race condition: it's conceivable that between the time we
3757 : * examine the bit here and the time SyncOneBuffer acquires the lock,
3758 : * someone else not only wrote the buffer but replaced it with another
3759 : * page and dirtied it. In that improbable case, SyncOneBuffer will
3760 : * write the buffer though we didn't need to. It doesn't seem worth
3761 : * guarding against this, though.
3762 : */
3763 340638 : if (pg_atomic_read_u64(&bufHdr->state) & BM_CHECKPOINT_NEEDED)
3764 : {
3765 317426 : if (SyncOneBuffer(buf_id, false, &wb_context) & BUF_WRITTEN)
3766 : {
3767 : TRACE_POSTGRESQL_BUFFER_SYNC_WRITTEN(buf_id);
3768 317426 : PendingCheckpointerStats.buffers_written++;
3769 317426 : num_written++;
3770 : }
3771 : }
3772 :
3773 : /*
3774 : * Measure progress independent of actually having to flush the buffer
3775 : * - otherwise writing become unbalanced.
3776 : */
3777 340638 : ts_stat->progress += ts_stat->progress_slice;
3778 340638 : ts_stat->num_scanned++;
3779 340638 : ts_stat->index++;
3780 :
3781 : /* Have all the buffers from the tablespace been processed? */
3782 340638 : if (ts_stat->num_scanned == ts_stat->num_to_scan)
3783 : {
3784 1829 : binaryheap_remove_first(ts_heap);
3785 : }
3786 : else
3787 : {
3788 : /* update heap with the new progress */
3789 338809 : binaryheap_replace_first(ts_heap, PointerGetDatum(ts_stat));
3790 : }
3791 :
3792 : /*
3793 : * Sleep to throttle our I/O rate.
3794 : *
3795 : * (This will check for barrier events even if it doesn't sleep.)
3796 : */
3797 340638 : CheckpointWriteDelay(flags, (double) num_processed / num_to_scan);
3798 : }
3799 :
3800 : /*
3801 : * Issue all pending flushes. Only checkpointer calls BufferSync(), so
3802 : * IOContext will always be IOCONTEXT_NORMAL.
3803 : */
3804 1194 : IssuePendingWritebacks(&wb_context, IOCONTEXT_NORMAL);
3805 :
3806 1194 : pfree(per_ts_stat);
3807 1194 : per_ts_stat = NULL;
3808 1194 : binaryheap_free(ts_heap);
3809 :
3810 : /*
3811 : * Update checkpoint statistics. As noted above, this doesn't include
3812 : * buffers written by other backends or bgwriter scan.
3813 : */
3814 1194 : CheckpointStats.ckpt_bufs_written += num_written;
3815 :
3816 : TRACE_POSTGRESQL_BUFFER_SYNC_DONE(NBuffers, num_written, num_to_scan);
3817 : }
3818 :
3819 : /*
3820 : * BgBufferSync -- Write out some dirty buffers in the pool.
3821 : *
3822 : * This is called periodically by the background writer process.
3823 : *
3824 : * Returns true if it's appropriate for the bgwriter process to go into
3825 : * low-power hibernation mode. (This happens if the strategy clock-sweep
3826 : * has been "lapped" and no buffer allocations have occurred recently,
3827 : * or if the bgwriter has been effectively disabled by setting
3828 : * bgwriter_lru_maxpages to 0.)
3829 : */
3830 : bool
3831 14504 : BgBufferSync(WritebackContext *wb_context)
3832 : {
3833 : /* info obtained from freelist.c */
3834 : int strategy_buf_id;
3835 : uint32 strategy_passes;
3836 : uint32 recent_alloc;
3837 :
3838 : /*
3839 : * Information saved between calls so we can determine the strategy
3840 : * point's advance rate and avoid scanning already-cleaned buffers.
3841 : */
3842 : static bool saved_info_valid = false;
3843 : static int prev_strategy_buf_id;
3844 : static uint32 prev_strategy_passes;
3845 : static int next_to_clean;
3846 : static uint32 next_passes;
3847 :
3848 : /* Moving averages of allocation rate and clean-buffer density */
3849 : static float smoothed_alloc = 0;
3850 : static float smoothed_density = 10.0;
3851 :
3852 : /* Potentially these could be tunables, but for now, not */
3853 14504 : float smoothing_samples = 16;
3854 14504 : float scan_whole_pool_milliseconds = 120000.0;
3855 :
3856 : /* Used to compute how far we scan ahead */
3857 : long strategy_delta;
3858 : int bufs_to_lap;
3859 : int bufs_ahead;
3860 : float scans_per_alloc;
3861 : int reusable_buffers_est;
3862 : int upcoming_alloc_est;
3863 : int min_scan_buffers;
3864 :
3865 : /* Variables for the scanning loop proper */
3866 : int num_to_scan;
3867 : int num_written;
3868 : int reusable_buffers;
3869 :
3870 : /* Variables for final smoothed_density update */
3871 : long new_strategy_delta;
3872 : uint32 new_recent_alloc;
3873 :
3874 : /*
3875 : * Find out where the clock-sweep currently is, and how many buffer
3876 : * allocations have happened since our last call.
3877 : */
3878 14504 : strategy_buf_id = StrategySyncStart(&strategy_passes, &recent_alloc);
3879 :
3880 : /* Report buffer alloc counts to pgstat */
3881 14504 : PendingBgWriterStats.buf_alloc += recent_alloc;
3882 :
3883 : /*
3884 : * If we're not running the LRU scan, just stop after doing the stats
3885 : * stuff. We mark the saved state invalid so that we can recover sanely
3886 : * if LRU scan is turned back on later.
3887 : */
3888 14504 : if (bgwriter_lru_maxpages <= 0)
3889 : {
3890 40 : saved_info_valid = false;
3891 40 : return true;
3892 : }
3893 :
3894 : /*
3895 : * Compute strategy_delta = how many buffers have been scanned by the
3896 : * clock-sweep since last time. If first time through, assume none. Then
3897 : * see if we are still ahead of the clock-sweep, and if so, how many
3898 : * buffers we could scan before we'd catch up with it and "lap" it. Note:
3899 : * weird-looking coding of xxx_passes comparisons are to avoid bogus
3900 : * behavior when the passes counts wrap around.
3901 : */
3902 14464 : if (saved_info_valid)
3903 : {
3904 13837 : int32 passes_delta = strategy_passes - prev_strategy_passes;
3905 :
3906 13837 : strategy_delta = strategy_buf_id - prev_strategy_buf_id;
3907 13837 : strategy_delta += (long) passes_delta * NBuffers;
3908 :
3909 : Assert(strategy_delta >= 0);
3910 :
3911 13837 : if ((int32) (next_passes - strategy_passes) > 0)
3912 : {
3913 : /* we're one pass ahead of the strategy point */
3914 2345 : bufs_to_lap = strategy_buf_id - next_to_clean;
3915 : #ifdef BGW_DEBUG
3916 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3917 : next_passes, next_to_clean,
3918 : strategy_passes, strategy_buf_id,
3919 : strategy_delta, bufs_to_lap);
3920 : #endif
3921 : }
3922 11492 : else if (next_passes == strategy_passes &&
3923 8918 : next_to_clean >= strategy_buf_id)
3924 : {
3925 : /* on same pass, but ahead or at least not behind */
3926 7984 : bufs_to_lap = NBuffers - (next_to_clean - strategy_buf_id);
3927 : #ifdef BGW_DEBUG
3928 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3929 : next_passes, next_to_clean,
3930 : strategy_passes, strategy_buf_id,
3931 : strategy_delta, bufs_to_lap);
3932 : #endif
3933 : }
3934 : else
3935 : {
3936 : /*
3937 : * We're behind, so skip forward to the strategy point and start
3938 : * cleaning from there.
3939 : */
3940 : #ifdef BGW_DEBUG
3941 : elog(DEBUG2, "bgwriter behind: bgw %u-%u strategy %u-%u delta=%ld",
3942 : next_passes, next_to_clean,
3943 : strategy_passes, strategy_buf_id,
3944 : strategy_delta);
3945 : #endif
3946 3508 : next_to_clean = strategy_buf_id;
3947 3508 : next_passes = strategy_passes;
3948 3508 : bufs_to_lap = NBuffers;
3949 : }
3950 : }
3951 : else
3952 : {
3953 : /*
3954 : * Initializing at startup or after LRU scanning had been off. Always
3955 : * start at the strategy point.
3956 : */
3957 : #ifdef BGW_DEBUG
3958 : elog(DEBUG2, "bgwriter initializing: strategy %u-%u",
3959 : strategy_passes, strategy_buf_id);
3960 : #endif
3961 627 : strategy_delta = 0;
3962 627 : next_to_clean = strategy_buf_id;
3963 627 : next_passes = strategy_passes;
3964 627 : bufs_to_lap = NBuffers;
3965 : }
3966 :
3967 : /* Update saved info for next time */
3968 14464 : prev_strategy_buf_id = strategy_buf_id;
3969 14464 : prev_strategy_passes = strategy_passes;
3970 14464 : saved_info_valid = true;
3971 :
3972 : /*
3973 : * Compute how many buffers had to be scanned for each new allocation, ie,
3974 : * 1/density of reusable buffers, and track a moving average of that.
3975 : *
3976 : * If the strategy point didn't move, we don't update the density estimate
3977 : */
3978 14464 : if (strategy_delta > 0 && recent_alloc > 0)
3979 : {
3980 7824 : scans_per_alloc = (float) strategy_delta / (float) recent_alloc;
3981 7824 : smoothed_density += (scans_per_alloc - smoothed_density) /
3982 : smoothing_samples;
3983 : }
3984 :
3985 : /*
3986 : * Estimate how many reusable buffers there are between the current
3987 : * strategy point and where we've scanned ahead to, based on the smoothed
3988 : * density estimate.
3989 : */
3990 14464 : bufs_ahead = NBuffers - bufs_to_lap;
3991 14464 : reusable_buffers_est = (float) bufs_ahead / smoothed_density;
3992 :
3993 : /*
3994 : * Track a moving average of recent buffer allocations. Here, rather than
3995 : * a true average we want a fast-attack, slow-decline behavior: we
3996 : * immediately follow any increase.
3997 : */
3998 14464 : if (smoothed_alloc <= (float) recent_alloc)
3999 3858 : smoothed_alloc = recent_alloc;
4000 : else
4001 10606 : smoothed_alloc += ((float) recent_alloc - smoothed_alloc) /
4002 : smoothing_samples;
4003 :
4004 : /* Scale the estimate by a GUC to allow more aggressive tuning. */
4005 14464 : upcoming_alloc_est = (int) (smoothed_alloc * bgwriter_lru_multiplier);
4006 :
4007 : /*
4008 : * If recent_alloc remains at zero for many cycles, smoothed_alloc will
4009 : * eventually underflow to zero, and the underflows produce annoying
4010 : * kernel warnings on some platforms. Once upcoming_alloc_est has gone to
4011 : * zero, there's no point in tracking smaller and smaller values of
4012 : * smoothed_alloc, so just reset it to exactly zero to avoid this
4013 : * syndrome. It will pop back up as soon as recent_alloc increases.
4014 : */
4015 14464 : if (upcoming_alloc_est == 0)
4016 2281 : smoothed_alloc = 0;
4017 :
4018 : /*
4019 : * Even in cases where there's been little or no buffer allocation
4020 : * activity, we want to make a small amount of progress through the buffer
4021 : * cache so that as many reusable buffers as possible are clean after an
4022 : * idle period.
4023 : *
4024 : * (scan_whole_pool_milliseconds / BgWriterDelay) computes how many times
4025 : * the BGW will be called during the scan_whole_pool time; slice the
4026 : * buffer pool into that many sections.
4027 : */
4028 14464 : min_scan_buffers = (int) (NBuffers / (scan_whole_pool_milliseconds / BgWriterDelay));
4029 :
4030 14464 : if (upcoming_alloc_est < (min_scan_buffers + reusable_buffers_est))
4031 : {
4032 : #ifdef BGW_DEBUG
4033 : elog(DEBUG2, "bgwriter: alloc_est=%d too small, using min=%d + reusable_est=%d",
4034 : upcoming_alloc_est, min_scan_buffers, reusable_buffers_est);
4035 : #endif
4036 6928 : upcoming_alloc_est = min_scan_buffers + reusable_buffers_est;
4037 : }
4038 :
4039 : /*
4040 : * Now write out dirty reusable buffers, working forward from the
4041 : * next_to_clean point, until we have lapped the strategy scan, or cleaned
4042 : * enough buffers to match our estimate of the next cycle's allocation
4043 : * requirements, or hit the bgwriter_lru_maxpages limit.
4044 : */
4045 :
4046 14464 : num_to_scan = bufs_to_lap;
4047 14464 : num_written = 0;
4048 14464 : reusable_buffers = reusable_buffers_est;
4049 :
4050 : /* Execute the LRU scan */
4051 2187891 : while (num_to_scan > 0 && reusable_buffers < upcoming_alloc_est)
4052 : {
4053 2173428 : int sync_state = SyncOneBuffer(next_to_clean, true,
4054 : wb_context);
4055 :
4056 2173428 : if (++next_to_clean >= NBuffers)
4057 : {
4058 3356 : next_to_clean = 0;
4059 3356 : next_passes++;
4060 : }
4061 2173428 : num_to_scan--;
4062 :
4063 2173428 : if (sync_state & BUF_WRITTEN)
4064 : {
4065 27665 : reusable_buffers++;
4066 27665 : if (++num_written >= bgwriter_lru_maxpages)
4067 : {
4068 1 : PendingBgWriterStats.maxwritten_clean++;
4069 1 : break;
4070 : }
4071 : }
4072 2145763 : else if (sync_state & BUF_REUSABLE)
4073 1652171 : reusable_buffers++;
4074 : }
4075 :
4076 14464 : PendingBgWriterStats.buf_written_clean += num_written;
4077 :
4078 : #ifdef BGW_DEBUG
4079 : 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",
4080 : recent_alloc, smoothed_alloc, strategy_delta, bufs_ahead,
4081 : smoothed_density, reusable_buffers_est, upcoming_alloc_est,
4082 : bufs_to_lap - num_to_scan,
4083 : num_written,
4084 : reusable_buffers - reusable_buffers_est);
4085 : #endif
4086 :
4087 : /*
4088 : * Consider the above scan as being like a new allocation scan.
4089 : * Characterize its density and update the smoothed one based on it. This
4090 : * effectively halves the moving average period in cases where both the
4091 : * strategy and the background writer are doing some useful scanning,
4092 : * which is helpful because a long memory isn't as desirable on the
4093 : * density estimates.
4094 : */
4095 14464 : new_strategy_delta = bufs_to_lap - num_to_scan;
4096 14464 : new_recent_alloc = reusable_buffers - reusable_buffers_est;
4097 14464 : if (new_strategy_delta > 0 && new_recent_alloc > 0)
4098 : {
4099 12321 : scans_per_alloc = (float) new_strategy_delta / (float) new_recent_alloc;
4100 12321 : smoothed_density += (scans_per_alloc - smoothed_density) /
4101 : smoothing_samples;
4102 :
4103 : #ifdef BGW_DEBUG
4104 : elog(DEBUG2, "bgwriter: cleaner density alloc=%u scan=%ld density=%.2f new smoothed=%.2f",
4105 : new_recent_alloc, new_strategy_delta,
4106 : scans_per_alloc, smoothed_density);
4107 : #endif
4108 : }
4109 :
4110 : /* Return true if OK to hibernate */
4111 14464 : return (bufs_to_lap == 0 && recent_alloc == 0);
4112 : }
4113 :
4114 : /*
4115 : * SyncOneBuffer -- process a single buffer during syncing.
4116 : *
4117 : * If skip_recently_used is true, we don't write currently-pinned buffers, nor
4118 : * buffers marked recently used, as these are not replacement candidates.
4119 : *
4120 : * Returns a bitmask containing the following flag bits:
4121 : * BUF_WRITTEN: we wrote the buffer.
4122 : * BUF_REUSABLE: buffer is available for replacement, ie, it has
4123 : * pin count 0 and usage count 0.
4124 : *
4125 : * (BUF_WRITTEN could be set in error if FlushBuffer finds the buffer clean
4126 : * after locking it, but we don't care all that much.)
4127 : */
4128 : static int
4129 2490854 : SyncOneBuffer(int buf_id, bool skip_recently_used, WritebackContext *wb_context)
4130 : {
4131 2490854 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
4132 2490854 : int result = 0;
4133 : uint64 buf_state;
4134 : BufferTag tag;
4135 :
4136 : /* Make sure we can handle the pin */
4137 2490854 : ReservePrivateRefCountEntry();
4138 2490854 : ResourceOwnerEnlarge(CurrentResourceOwner);
4139 :
4140 : /*
4141 : * Check whether buffer needs writing.
4142 : *
4143 : * We can make this check without taking the buffer content lock so long
4144 : * as we mark pages dirty in access methods *before* logging changes with
4145 : * XLogInsert(): if someone marks the buffer dirty just after our check we
4146 : * don't worry because our checkpoint.redo points before log record for
4147 : * upcoming changes and so we are not required to write such dirty buffer.
4148 : */
4149 2490854 : buf_state = LockBufHdr(bufHdr);
4150 :
4151 2490854 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 0 &&
4152 2487173 : BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
4153 : {
4154 1682017 : result |= BUF_REUSABLE;
4155 : }
4156 808837 : else if (skip_recently_used)
4157 : {
4158 : /* Caller told us not to write recently-used buffers */
4159 493592 : UnlockBufHdr(bufHdr);
4160 493592 : return result;
4161 : }
4162 :
4163 1997262 : if (!(buf_state & BM_VALID) || !(buf_state & BM_DIRTY))
4164 : {
4165 : /* It's clean, so nothing to do */
4166 1652171 : UnlockBufHdr(bufHdr);
4167 1652171 : return result;
4168 : }
4169 :
4170 : /*
4171 : * Pin it, share-exclusive-lock it, write it. (FlushBuffer will do
4172 : * nothing if the buffer is clean by the time we've locked it.)
4173 : */
4174 345091 : PinBuffer_Locked(bufHdr);
4175 :
4176 345091 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
4177 :
4178 345091 : tag = bufHdr->tag;
4179 :
4180 345091 : UnpinBuffer(bufHdr);
4181 :
4182 : /*
4183 : * SyncOneBuffer() is only called by checkpointer and bgwriter, so
4184 : * IOContext will always be IOCONTEXT_NORMAL.
4185 : */
4186 345091 : ScheduleBufferTagForWriteback(wb_context, IOCONTEXT_NORMAL, &tag);
4187 :
4188 345091 : return result | BUF_WRITTEN;
4189 : }
4190 :
4191 : /*
4192 : * AtEOXact_Buffers - clean up at end of transaction.
4193 : *
4194 : * As of PostgreSQL 8.0, buffer pins should get released by the
4195 : * ResourceOwner mechanism. This routine is just a debugging
4196 : * cross-check that no pins remain.
4197 : */
4198 : void
4199 629828 : AtEOXact_Buffers(bool isCommit)
4200 : {
4201 629828 : CheckForBufferLeaks();
4202 :
4203 629828 : AtEOXact_LocalBuffers(isCommit);
4204 :
4205 : Assert(PrivateRefCountOverflowed == 0);
4206 629828 : }
4207 :
4208 : /*
4209 : * Initialize access to shared buffer pool
4210 : *
4211 : * This is called during backend startup (whether standalone or under the
4212 : * postmaster). It sets up for this backend's access to the already-existing
4213 : * buffer pool.
4214 : */
4215 : void
4216 24763 : InitBufferManagerAccess(void)
4217 : {
4218 : /*
4219 : * An advisory limit on the number of pins each backend should hold, based
4220 : * on shared_buffers and the maximum number of connections possible.
4221 : * That's very pessimistic, but outside toy-sized shared_buffers it should
4222 : * allow plenty of pins. LimitAdditionalPins() and
4223 : * GetAdditionalPinLimit() can be used to check the remaining balance.
4224 : */
4225 24763 : MaxProportionalPins = NBuffers / (MaxBackends + NUM_AUXILIARY_PROCS);
4226 :
4227 24763 : memset(&PrivateRefCountArray, 0, sizeof(PrivateRefCountArray));
4228 24763 : memset(&PrivateRefCountArrayKeys, 0, sizeof(PrivateRefCountArrayKeys));
4229 :
4230 24763 : PrivateRefCountHash = refcount_create(CurrentMemoryContext, 100, NULL);
4231 :
4232 : /*
4233 : * AtProcExit_Buffers needs LWLock access, and thereby has to be called at
4234 : * the corresponding phase of backend shutdown.
4235 : */
4236 : Assert(MyProc != NULL);
4237 24763 : on_shmem_exit(AtProcExit_Buffers, 0);
4238 24763 : }
4239 :
4240 : /*
4241 : * During backend exit, ensure that we released all shared-buffer locks and
4242 : * assert that we have no remaining pins.
4243 : */
4244 : static void
4245 24763 : AtProcExit_Buffers(int code, Datum arg)
4246 : {
4247 24763 : UnlockBuffers();
4248 :
4249 24763 : CheckForBufferLeaks();
4250 :
4251 : /* localbuf.c needs a chance too */
4252 24763 : AtProcExit_LocalBuffers();
4253 24763 : }
4254 :
4255 : /*
4256 : * CheckForBufferLeaks - ensure this backend holds no buffer pins
4257 : *
4258 : * As of PostgreSQL 8.0, buffer pins should get released by the
4259 : * ResourceOwner mechanism. This routine is just a debugging
4260 : * cross-check that no pins remain.
4261 : */
4262 : static void
4263 654591 : CheckForBufferLeaks(void)
4264 : {
4265 : #ifdef USE_ASSERT_CHECKING
4266 : int RefCountErrors = 0;
4267 : PrivateRefCountEntry *res;
4268 : int i;
4269 : char *s;
4270 :
4271 : /* check the array */
4272 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4273 : {
4274 : if (PrivateRefCountArrayKeys[i] != InvalidBuffer)
4275 : {
4276 : res = &PrivateRefCountArray[i];
4277 :
4278 : s = DebugPrintBufferRefcount(res->buffer);
4279 : elog(WARNING, "buffer refcount leak: %s", s);
4280 : pfree(s);
4281 :
4282 : RefCountErrors++;
4283 : }
4284 : }
4285 :
4286 : /* if necessary search the hash */
4287 : if (PrivateRefCountOverflowed)
4288 : {
4289 : refcount_iterator iter;
4290 :
4291 : refcount_start_iterate(PrivateRefCountHash, &iter);
4292 : while ((res = refcount_iterate(PrivateRefCountHash, &iter)) != NULL)
4293 : {
4294 : s = DebugPrintBufferRefcount(res->buffer);
4295 : elog(WARNING, "buffer refcount leak: %s", s);
4296 : pfree(s);
4297 : RefCountErrors++;
4298 : }
4299 : }
4300 :
4301 : Assert(RefCountErrors == 0);
4302 : #endif
4303 654591 : }
4304 :
4305 : #ifdef USE_ASSERT_CHECKING
4306 : /*
4307 : * Check for exclusive-locked catalog buffers. This is the core of
4308 : * AssertCouldGetRelation().
4309 : *
4310 : * A backend would self-deadlock on the content lock if the catalog scan read
4311 : * the exclusive-locked buffer. The main threat is exclusive-locked buffers
4312 : * of catalogs used in relcache, because a catcache search on any catalog may
4313 : * build that catalog's relcache entry. We don't have an inventory of
4314 : * catalogs relcache uses, so just check buffers of most catalogs.
4315 : *
4316 : * It's better to minimize waits while holding an exclusive buffer lock, so it
4317 : * would be nice to broaden this check not to be catalog-specific. However,
4318 : * bttextcmp() accesses pg_collation, and non-core opclasses might similarly
4319 : * read tables. That is deadlock-free as long as there's no loop in the
4320 : * dependency graph: modifying table A may cause an opclass to read table B,
4321 : * but it must not cause a read of table A.
4322 : */
4323 : void
4324 : AssertBufferLocksPermitCatalogRead(void)
4325 : {
4326 : PrivateRefCountEntry *res;
4327 :
4328 : /* check the array */
4329 : for (int i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4330 : {
4331 : if (PrivateRefCountArrayKeys[i] != InvalidBuffer)
4332 : {
4333 : res = &PrivateRefCountArray[i];
4334 :
4335 : if (res->buffer == InvalidBuffer)
4336 : continue;
4337 :
4338 : AssertNotCatalogBufferLock(res->buffer, res->data.lockmode);
4339 : }
4340 : }
4341 :
4342 : /* if necessary search the hash */
4343 : if (PrivateRefCountOverflowed)
4344 : {
4345 : refcount_iterator iter;
4346 :
4347 : refcount_start_iterate(PrivateRefCountHash, &iter);
4348 : while ((res = refcount_iterate(PrivateRefCountHash, &iter)) != NULL)
4349 : {
4350 : AssertNotCatalogBufferLock(res->buffer, res->data.lockmode);
4351 : }
4352 : }
4353 : }
4354 :
4355 : static void
4356 : AssertNotCatalogBufferLock(Buffer buffer, BufferLockMode mode)
4357 : {
4358 : BufferDesc *bufHdr = GetBufferDescriptor(buffer - 1);
4359 : BufferTag tag;
4360 : Oid relid;
4361 :
4362 : if (mode != BUFFER_LOCK_EXCLUSIVE)
4363 : return;
4364 :
4365 : tag = bufHdr->tag;
4366 :
4367 : /*
4368 : * This relNumber==relid assumption holds until a catalog experiences
4369 : * VACUUM FULL or similar. After a command like that, relNumber will be
4370 : * in the normal (non-catalog) range, and we lose the ability to detect
4371 : * hazardous access to that catalog. Calling RelidByRelfilenumber() would
4372 : * close that gap, but RelidByRelfilenumber() might then deadlock with a
4373 : * held lock.
4374 : */
4375 : relid = tag.relNumber;
4376 :
4377 : if (IsCatalogTextUniqueIndexOid(relid)) /* see comments at the callee */
4378 : return;
4379 :
4380 : Assert(!IsCatalogRelationOid(relid));
4381 : }
4382 : #endif
4383 :
4384 :
4385 : /*
4386 : * Helper routine to issue warnings when a buffer is unexpectedly pinned
4387 : */
4388 : char *
4389 46 : DebugPrintBufferRefcount(Buffer buffer)
4390 : {
4391 : BufferDesc *buf;
4392 : int32 loccount;
4393 : char *result;
4394 : ProcNumber backend;
4395 : uint64 buf_state;
4396 :
4397 : Assert(BufferIsValid(buffer));
4398 46 : if (BufferIsLocal(buffer))
4399 : {
4400 16 : buf = GetLocalBufferDescriptor(-buffer - 1);
4401 16 : loccount = LocalRefCount[-buffer - 1];
4402 16 : backend = MyProcNumber;
4403 : }
4404 : else
4405 : {
4406 30 : buf = GetBufferDescriptor(buffer - 1);
4407 30 : loccount = GetPrivateRefCount(buffer);
4408 30 : backend = INVALID_PROC_NUMBER;
4409 : }
4410 :
4411 : /* theoretically we should lock the bufHdr here */
4412 46 : buf_state = pg_atomic_read_u64(&buf->state);
4413 :
4414 46 : result = psprintf("[%03d] (rel=%s, blockNum=%u, flags=0x%" PRIx64 ", refcount=%u %d)",
4415 : buffer,
4416 46 : relpathbackend(BufTagGetRelFileLocator(&buf->tag), backend,
4417 : BufTagGetForkNum(&buf->tag)).str,
4418 : buf->tag.blockNum, buf_state & BUF_FLAG_MASK,
4419 : BUF_STATE_GET_REFCOUNT(buf_state), loccount);
4420 46 : return result;
4421 : }
4422 :
4423 : /*
4424 : * CheckPointBuffers
4425 : *
4426 : * Flush all dirty blocks in buffer pool to disk at checkpoint time.
4427 : *
4428 : * Note: temporary relations do not participate in checkpoints, so they don't
4429 : * need to be flushed.
4430 : */
4431 : void
4432 1931 : CheckPointBuffers(int flags)
4433 : {
4434 1931 : BufferSync(flags);
4435 1931 : }
4436 :
4437 : /*
4438 : * BufferGetBlockNumber
4439 : * Returns the block number associated with a buffer.
4440 : *
4441 : * Note:
4442 : * Assumes that the buffer is valid and pinned, else the
4443 : * value may be obsolete immediately...
4444 : */
4445 : BlockNumber
4446 79200578 : BufferGetBlockNumber(Buffer buffer)
4447 : {
4448 : BufferDesc *bufHdr;
4449 :
4450 : Assert(BufferIsPinned(buffer));
4451 :
4452 79200578 : if (BufferIsLocal(buffer))
4453 2561335 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4454 : else
4455 76639243 : bufHdr = GetBufferDescriptor(buffer - 1);
4456 :
4457 : /* pinned, so OK to read tag without spinlock */
4458 79200578 : return bufHdr->tag.blockNum;
4459 : }
4460 :
4461 : /*
4462 : * BufferGetTag
4463 : * Returns the relfilelocator, fork number and block number associated with
4464 : * a buffer.
4465 : */
4466 : void
4467 26655788 : BufferGetTag(Buffer buffer, RelFileLocator *rlocator, ForkNumber *forknum,
4468 : BlockNumber *blknum)
4469 : {
4470 : BufferDesc *bufHdr;
4471 :
4472 : /* Do the same checks as BufferGetBlockNumber. */
4473 : Assert(BufferIsPinned(buffer));
4474 :
4475 26655788 : if (BufferIsLocal(buffer))
4476 0 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4477 : else
4478 26655788 : bufHdr = GetBufferDescriptor(buffer - 1);
4479 :
4480 : /* pinned, so OK to read tag without spinlock */
4481 26655788 : *rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
4482 26655788 : *forknum = BufTagGetForkNum(&bufHdr->tag);
4483 26655788 : *blknum = bufHdr->tag.blockNum;
4484 26655788 : }
4485 :
4486 : /*
4487 : * FlushBuffer
4488 : * Physically write out a shared buffer.
4489 : *
4490 : * NOTE: this actually just passes the buffer contents to the kernel; the
4491 : * real write to disk won't happen until the kernel feels like it. This
4492 : * is okay from our point of view since we can redo the changes from WAL.
4493 : * However, we will need to force the changes to disk via fsync before
4494 : * we can checkpoint WAL.
4495 : *
4496 : * The caller must hold a pin on the buffer and have
4497 : * (share-)exclusively-locked the buffer contents.
4498 : *
4499 : * If the caller has an smgr reference for the buffer's relation, pass it
4500 : * as the second parameter. If not, pass NULL.
4501 : */
4502 : static void
4503 704281 : FlushBuffer(BufferDesc *buf, SMgrRelation reln, IOObject io_object,
4504 : IOContext io_context)
4505 : {
4506 : XLogRecPtr recptr;
4507 : ErrorContextCallback errcallback;
4508 : instr_time io_start;
4509 : Block bufBlock;
4510 :
4511 : Assert(BufferLockHeldByMeInMode(buf, BUFFER_LOCK_EXCLUSIVE) ||
4512 : BufferLockHeldByMeInMode(buf, BUFFER_LOCK_SHARE_EXCLUSIVE));
4513 :
4514 : /*
4515 : * Try to start an I/O operation. If StartBufferIO returns false, then
4516 : * someone else flushed the buffer before we could, so we need not do
4517 : * anything.
4518 : */
4519 704281 : if (StartSharedBufferIO(buf, false, true, NULL) == BUFFER_IO_ALREADY_DONE)
4520 19 : return;
4521 :
4522 : /* Setup error traceback support for ereport() */
4523 704262 : errcallback.callback = shared_buffer_write_error_callback;
4524 704262 : errcallback.arg = buf;
4525 704262 : errcallback.previous = error_context_stack;
4526 704262 : error_context_stack = &errcallback;
4527 :
4528 : /* Find smgr relation for buffer */
4529 704262 : if (reln == NULL)
4530 700065 : reln = smgropen(BufTagGetRelFileLocator(&buf->tag), INVALID_PROC_NUMBER);
4531 :
4532 : TRACE_POSTGRESQL_BUFFER_FLUSH_START(BufTagGetForkNum(&buf->tag),
4533 : buf->tag.blockNum,
4534 : reln->smgr_rlocator.locator.spcOid,
4535 : reln->smgr_rlocator.locator.dbOid,
4536 : reln->smgr_rlocator.locator.relNumber);
4537 :
4538 : /*
4539 : * As we hold at least a share-exclusive lock on the buffer, the LSN
4540 : * cannot change during the flush (and thus can't be torn).
4541 : */
4542 704262 : recptr = BufferGetLSN(buf);
4543 :
4544 : /*
4545 : * Force XLOG flush up to buffer's LSN. This implements the basic WAL
4546 : * rule that log updates must hit disk before any of the data-file changes
4547 : * they describe do.
4548 : *
4549 : * However, this rule does not apply to unlogged relations, which will be
4550 : * lost after a crash anyway. Most unlogged relation pages do not bear
4551 : * LSNs since we never emit WAL records for them, and therefore flushing
4552 : * up through the buffer LSN would be useless, but harmless. However,
4553 : * some index AMs use LSNs internally to detect concurrent page
4554 : * modifications, and therefore unlogged index pages bear "fake" LSNs
4555 : * generated by XLogGetFakeLSN. It is unlikely but possible that the fake
4556 : * LSN counter could advance past the WAL insertion point; and if it did
4557 : * happen, attempting to flush WAL through that location would fail, with
4558 : * disastrous system-wide consequences. To make sure that can't happen,
4559 : * skip the flush if the buffer isn't permanent.
4560 : */
4561 704262 : if (pg_atomic_read_u64(&buf->state) & BM_PERMANENT)
4562 702432 : XLogFlush(recptr);
4563 :
4564 : /*
4565 : * Now it's safe to write the buffer to disk. Note that no one else should
4566 : * have been able to write it, while we were busy with log flushing,
4567 : * because we got the exclusive right to perform I/O by setting the
4568 : * BM_IO_IN_PROGRESS bit.
4569 : */
4570 704262 : bufBlock = BufHdrGetBlock(buf);
4571 :
4572 : /* Update page checksum if desired. */
4573 704262 : PageSetChecksum((Page) bufBlock, buf->tag.blockNum);
4574 :
4575 704262 : io_start = pgstat_prepare_io_time(track_io_timing);
4576 :
4577 704262 : smgrwrite(reln,
4578 704262 : BufTagGetForkNum(&buf->tag),
4579 : buf->tag.blockNum,
4580 : bufBlock,
4581 : false);
4582 :
4583 : /*
4584 : * When a strategy is in use, only flushes of dirty buffers already in the
4585 : * strategy ring are counted as strategy writes (IOCONTEXT
4586 : * [BULKREAD|BULKWRITE|VACUUM] IOOP_WRITE) for the purpose of IO
4587 : * statistics tracking.
4588 : *
4589 : * If a shared buffer initially added to the ring must be flushed before
4590 : * being used, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE.
4591 : *
4592 : * If a shared buffer which was added to the ring later because the
4593 : * current strategy buffer is pinned or in use or because all strategy
4594 : * buffers were dirty and rejected (for BAS_BULKREAD operations only)
4595 : * requires flushing, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE
4596 : * (from_ring will be false).
4597 : *
4598 : * When a strategy is not in use, the write can only be a "regular" write
4599 : * of a dirty shared buffer (IOCONTEXT_NORMAL IOOP_WRITE).
4600 : */
4601 704262 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
4602 : IOOP_WRITE, io_start, 1, BLCKSZ);
4603 :
4604 704262 : pgBufferUsage.shared_blks_written++;
4605 :
4606 : /*
4607 : * Mark the buffer as clean and end the BM_IO_IN_PROGRESS state.
4608 : */
4609 704262 : TerminateBufferIO(buf, true, 0, true, false);
4610 :
4611 : TRACE_POSTGRESQL_BUFFER_FLUSH_DONE(BufTagGetForkNum(&buf->tag),
4612 : buf->tag.blockNum,
4613 : reln->smgr_rlocator.locator.spcOid,
4614 : reln->smgr_rlocator.locator.dbOid,
4615 : reln->smgr_rlocator.locator.relNumber);
4616 :
4617 : /* Pop the error context stack */
4618 704262 : error_context_stack = errcallback.previous;
4619 : }
4620 :
4621 : /*
4622 : * Convenience wrapper around FlushBuffer() that locks/unlocks the buffer
4623 : * before/after calling FlushBuffer().
4624 : */
4625 : static void
4626 350386 : FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
4627 : IOObject io_object, IOContext io_context)
4628 : {
4629 350386 : Buffer buffer = BufferDescriptorGetBuffer(buf);
4630 :
4631 350386 : BufferLockAcquire(buffer, buf, BUFFER_LOCK_SHARE_EXCLUSIVE);
4632 350386 : FlushBuffer(buf, reln, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
4633 350386 : BufferLockUnlock(buffer, buf);
4634 350386 : }
4635 :
4636 : /*
4637 : * RelationGetNumberOfBlocksInFork
4638 : * Determines the current number of pages in the specified relation fork.
4639 : *
4640 : * Note that the accuracy of the result will depend on the details of the
4641 : * relation's storage. For builtin AMs it'll be accurate, but for external AMs
4642 : * it might not be.
4643 : */
4644 : BlockNumber
4645 2614734 : RelationGetNumberOfBlocksInFork(Relation relation, ForkNumber forkNum)
4646 : {
4647 2614734 : if (RELKIND_HAS_TABLE_AM(relation->rd_rel->relkind))
4648 : {
4649 : /*
4650 : * Not every table AM uses BLCKSZ wide fixed size blocks. Therefore
4651 : * tableam returns the size in bytes - but for the purpose of this
4652 : * routine, we want the number of blocks. Therefore divide, rounding
4653 : * up.
4654 : */
4655 : uint64 szbytes;
4656 :
4657 1928496 : szbytes = table_relation_size(relation, forkNum);
4658 :
4659 1928477 : return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
4660 : }
4661 686238 : else if (RELKIND_HAS_STORAGE(relation->rd_rel->relkind))
4662 : {
4663 686238 : return smgrnblocks(RelationGetSmgr(relation), forkNum);
4664 : }
4665 : else
4666 : Assert(false);
4667 :
4668 0 : return 0; /* keep compiler quiet */
4669 : }
4670 :
4671 : /*
4672 : * BufferIsPermanent
4673 : * Determines whether a buffer will potentially still be around after
4674 : * a crash. Caller must hold a buffer pin.
4675 : */
4676 : bool
4677 17141488 : BufferIsPermanent(Buffer buffer)
4678 : {
4679 : BufferDesc *bufHdr;
4680 :
4681 : /* Local buffers are used only for temp relations. */
4682 17141488 : if (BufferIsLocal(buffer))
4683 827660 : return false;
4684 :
4685 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4686 : Assert(BufferIsValid(buffer));
4687 : Assert(BufferIsPinned(buffer));
4688 :
4689 : /*
4690 : * BM_PERMANENT can't be changed while we hold a pin on the buffer, so we
4691 : * need not bother with the buffer header spinlock. Even if someone else
4692 : * changes the buffer header state while we're doing this, the state is
4693 : * changed atomically, so we'll read the old value or the new value, but
4694 : * not random garbage.
4695 : */
4696 16313828 : bufHdr = GetBufferDescriptor(buffer - 1);
4697 16313828 : return (pg_atomic_read_u64(&bufHdr->state) & BM_PERMANENT) != 0;
4698 : }
4699 :
4700 : /*
4701 : * BufferGetLSNAtomic
4702 : * Retrieves the LSN of the buffer atomically.
4703 : *
4704 : * This is necessary for some callers who may only hold a share lock on
4705 : * the buffer. A share lock allows a concurrent backend to set hint bits
4706 : * on the page, which in turn may require a WAL record to be emitted.
4707 : *
4708 : * On platforms with 8 byte atomic reads/writes, we don't need to do any
4709 : * additional locking. On platforms not supporting such 8 byte atomic
4710 : * reads/writes, we need to actually take the header lock.
4711 : */
4712 : XLogRecPtr
4713 8786931 : BufferGetLSNAtomic(Buffer buffer)
4714 : {
4715 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4716 : Assert(BufferIsValid(buffer));
4717 : Assert(BufferIsPinned(buffer));
4718 :
4719 : #ifdef PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY
4720 8786931 : return PageGetLSN(BufferGetPage(buffer));
4721 : #else
4722 : {
4723 : char *page = BufferGetPage(buffer);
4724 : BufferDesc *bufHdr;
4725 : XLogRecPtr lsn;
4726 :
4727 : /*
4728 : * If we don't need locking for correctness, fastpath out.
4729 : */
4730 : if (!XLogHintBitIsNeeded() || BufferIsLocal(buffer))
4731 : return PageGetLSN(page);
4732 :
4733 : bufHdr = GetBufferDescriptor(buffer - 1);
4734 : LockBufHdr(bufHdr);
4735 : lsn = PageGetLSN(page);
4736 : UnlockBufHdr(bufHdr);
4737 :
4738 : return lsn;
4739 : }
4740 : #endif
4741 : }
4742 :
4743 : /* ---------------------------------------------------------------------
4744 : * DropRelationBuffers
4745 : *
4746 : * This function removes from the buffer pool all the pages of the
4747 : * specified relation forks that have block numbers >= firstDelBlock.
4748 : * (In particular, with firstDelBlock = 0, all pages are removed.)
4749 : * Dirty pages are simply dropped, without bothering to write them
4750 : * out first. Therefore, this is NOT rollback-able, and so should be
4751 : * used only with extreme caution!
4752 : *
4753 : * Currently, this is called only from smgr.c when the underlying file
4754 : * is about to be deleted or truncated (firstDelBlock is needed for
4755 : * the truncation case). The data in the affected pages would therefore
4756 : * be deleted momentarily anyway, and there is no point in writing it.
4757 : * It is the responsibility of higher-level code to ensure that the
4758 : * deletion or truncation does not lose any data that could be needed
4759 : * later. It is also the responsibility of higher-level code to ensure
4760 : * that no other process could be trying to load more pages of the
4761 : * relation into buffers.
4762 : * --------------------------------------------------------------------
4763 : */
4764 : void
4765 805 : DropRelationBuffers(SMgrRelation smgr_reln, ForkNumber *forkNum,
4766 : int nforks, BlockNumber *firstDelBlock)
4767 : {
4768 : int i;
4769 : int j;
4770 : RelFileLocatorBackend rlocator;
4771 : BlockNumber nForkBlock[MAX_FORKNUM];
4772 805 : uint64 nBlocksToInvalidate = 0;
4773 :
4774 805 : rlocator = smgr_reln->smgr_rlocator;
4775 :
4776 : /* If it's a local relation, it's localbuf.c's problem. */
4777 805 : if (RelFileLocatorBackendIsTemp(rlocator))
4778 : {
4779 498 : if (rlocator.backend == MyProcNumber)
4780 498 : DropRelationLocalBuffers(rlocator.locator, forkNum, nforks,
4781 : firstDelBlock);
4782 :
4783 541 : return;
4784 : }
4785 :
4786 : /*
4787 : * To remove all the pages of the specified relation forks from the buffer
4788 : * pool, we need to scan the entire buffer pool but we can optimize it by
4789 : * finding the buffers from BufMapping table provided we know the exact
4790 : * size of each fork of the relation. The exact size is required to ensure
4791 : * that we don't leave any buffer for the relation being dropped as
4792 : * otherwise the background writer or checkpointer can lead to a PANIC
4793 : * error while flushing buffers corresponding to files that don't exist.
4794 : *
4795 : * To know the exact size, we rely on the size cached for each fork by us
4796 : * during recovery which limits the optimization to recovery and on
4797 : * standbys but we can easily extend it once we have shared cache for
4798 : * relation size.
4799 : *
4800 : * In recovery, we cache the value returned by the first lseek(SEEK_END)
4801 : * and the future writes keeps the cached value up-to-date. See
4802 : * smgrextend. It is possible that the value of the first lseek is smaller
4803 : * than the actual number of existing blocks in the file due to buggy
4804 : * Linux kernels that might not have accounted for the recent write. But
4805 : * that should be fine because there must not be any buffers after that
4806 : * file size.
4807 : */
4808 413 : for (i = 0; i < nforks; i++)
4809 : {
4810 : /* Get the number of blocks for a relation's fork */
4811 358 : nForkBlock[i] = smgrnblocks_cached(smgr_reln, forkNum[i]);
4812 :
4813 358 : if (nForkBlock[i] == InvalidBlockNumber)
4814 : {
4815 252 : nBlocksToInvalidate = InvalidBlockNumber;
4816 252 : break;
4817 : }
4818 :
4819 : /* calculate the number of blocks to be invalidated */
4820 106 : nBlocksToInvalidate += (nForkBlock[i] - firstDelBlock[i]);
4821 : }
4822 :
4823 : /*
4824 : * We apply the optimization iff the total number of blocks to invalidate
4825 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4826 : */
4827 307 : if (BlockNumberIsValid(nBlocksToInvalidate) &&
4828 55 : nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4829 : {
4830 120 : for (j = 0; j < nforks; j++)
4831 77 : FindAndDropRelationBuffers(rlocator.locator, forkNum[j],
4832 77 : nForkBlock[j], firstDelBlock[j]);
4833 43 : return;
4834 : }
4835 :
4836 3529096 : for (i = 0; i < NBuffers; i++)
4837 : {
4838 3528832 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4839 :
4840 : /*
4841 : * We can make this a tad faster by prechecking the buffer tag before
4842 : * we attempt to lock the buffer; this saves a lot of lock
4843 : * acquisitions in typical cases. It should be safe because the
4844 : * caller must have AccessExclusiveLock on the relation, or some other
4845 : * reason to be certain that no one is loading new pages of the rel
4846 : * into the buffer pool. (Otherwise we might well miss such pages
4847 : * entirely.) Therefore, while the tag might be changing while we
4848 : * look at it, it can't be changing *to* a value we care about, only
4849 : * *away* from such a value. So false negatives are impossible, and
4850 : * false positives are safe because we'll recheck after getting the
4851 : * buffer lock.
4852 : *
4853 : * We could check forkNum and blockNum as well as the rlocator, but
4854 : * the incremental win from doing so seems small.
4855 : */
4856 3528832 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator))
4857 3520287 : continue;
4858 :
4859 8545 : LockBufHdr(bufHdr);
4860 :
4861 21978 : for (j = 0; j < nforks; j++)
4862 : {
4863 15438 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator) &&
4864 15438 : BufTagGetForkNum(&bufHdr->tag) == forkNum[j] &&
4865 8437 : bufHdr->tag.blockNum >= firstDelBlock[j])
4866 : {
4867 2005 : InvalidateBuffer(bufHdr); /* releases spinlock */
4868 2005 : break;
4869 : }
4870 : }
4871 8545 : if (j >= nforks)
4872 6540 : UnlockBufHdr(bufHdr);
4873 : }
4874 : }
4875 :
4876 : /* ---------------------------------------------------------------------
4877 : * DropRelationsAllBuffers
4878 : *
4879 : * This function removes from the buffer pool all the pages of all
4880 : * forks of the specified relations. It's equivalent to calling
4881 : * DropRelationBuffers once per fork per relation with firstDelBlock = 0.
4882 : * --------------------------------------------------------------------
4883 : */
4884 : void
4885 17752 : DropRelationsAllBuffers(SMgrRelation *smgr_reln, int nlocators)
4886 : {
4887 : int i;
4888 17752 : int n = 0;
4889 : SMgrRelation *rels;
4890 : BlockNumber (*block)[MAX_FORKNUM + 1];
4891 17752 : uint64 nBlocksToInvalidate = 0;
4892 : RelFileLocator *locators;
4893 17752 : bool cached = true;
4894 : bool use_bsearch;
4895 :
4896 17752 : if (nlocators == 0)
4897 0 : return;
4898 :
4899 17752 : rels = palloc_array(SMgrRelation, nlocators); /* non-local relations */
4900 :
4901 : /* If it's a local relation, it's localbuf.c's problem. */
4902 77973 : for (i = 0; i < nlocators; i++)
4903 : {
4904 60221 : if (RelFileLocatorBackendIsTemp(smgr_reln[i]->smgr_rlocator))
4905 : {
4906 4374 : if (smgr_reln[i]->smgr_rlocator.backend == MyProcNumber)
4907 4374 : DropRelationAllLocalBuffers(smgr_reln[i]->smgr_rlocator.locator);
4908 : }
4909 : else
4910 55847 : rels[n++] = smgr_reln[i];
4911 : }
4912 :
4913 : /*
4914 : * If there are no non-local relations, then we're done. Release the
4915 : * memory and return.
4916 : */
4917 17752 : if (n == 0)
4918 : {
4919 1164 : pfree(rels);
4920 1164 : return;
4921 : }
4922 :
4923 : /*
4924 : * This is used to remember the number of blocks for all the relations
4925 : * forks.
4926 : */
4927 : block = (BlockNumber (*)[MAX_FORKNUM + 1])
4928 16588 : palloc(sizeof(BlockNumber) * n * (MAX_FORKNUM + 1));
4929 :
4930 : /*
4931 : * We can avoid scanning the entire buffer pool if we know the exact size
4932 : * of each of the given relation forks. See DropRelationBuffers.
4933 : */
4934 34534 : for (i = 0; i < n && cached; i++)
4935 : {
4936 26803 : for (int j = 0; j <= MAX_FORKNUM; j++)
4937 : {
4938 : /* Get the number of blocks for a relation's fork. */
4939 24604 : block[i][j] = smgrnblocks_cached(rels[i], j);
4940 :
4941 : /* We need to only consider the relation forks that exists. */
4942 24604 : if (block[i][j] == InvalidBlockNumber)
4943 : {
4944 22213 : if (!smgrexists(rels[i], j))
4945 6466 : continue;
4946 15747 : cached = false;
4947 15747 : break;
4948 : }
4949 :
4950 : /* calculate the total number of blocks to be invalidated */
4951 2391 : nBlocksToInvalidate += block[i][j];
4952 : }
4953 : }
4954 :
4955 : /*
4956 : * We apply the optimization iff the total number of blocks to invalidate
4957 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4958 : */
4959 16588 : if (cached && nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4960 : {
4961 1396 : for (i = 0; i < n; i++)
4962 : {
4963 3850 : for (int j = 0; j <= MAX_FORKNUM; j++)
4964 : {
4965 : /* ignore relation forks that doesn't exist */
4966 3080 : if (!BlockNumberIsValid(block[i][j]))
4967 2301 : continue;
4968 :
4969 : /* drop all the buffers for a particular relation fork */
4970 779 : FindAndDropRelationBuffers(rels[i]->smgr_rlocator.locator,
4971 779 : j, block[i][j], 0);
4972 : }
4973 : }
4974 :
4975 626 : pfree(block);
4976 626 : pfree(rels);
4977 626 : return;
4978 : }
4979 :
4980 15962 : pfree(block);
4981 15962 : locators = palloc_array(RelFileLocator, n); /* non-local relations */
4982 71039 : for (i = 0; i < n; i++)
4983 55077 : locators[i] = rels[i]->smgr_rlocator.locator;
4984 :
4985 : /*
4986 : * For low number of relations to drop just use a simple walk through, to
4987 : * save the bsearch overhead. The threshold to use is rather a guess than
4988 : * an exactly determined value, as it depends on many factors (CPU and RAM
4989 : * speeds, amount of shared buffers etc.).
4990 : */
4991 15962 : use_bsearch = n > RELS_BSEARCH_THRESHOLD;
4992 :
4993 : /* sort the list of rlocators if necessary */
4994 15962 : if (use_bsearch)
4995 218 : qsort(locators, n, sizeof(RelFileLocator), rlocator_comparator);
4996 :
4997 186613466 : for (i = 0; i < NBuffers; i++)
4998 : {
4999 186597504 : RelFileLocator *rlocator = NULL;
5000 186597504 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5001 :
5002 : /*
5003 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5004 : * saves some cycles.
5005 : */
5006 :
5007 186597504 : if (!use_bsearch)
5008 : {
5009 : int j;
5010 :
5011 740289455 : for (j = 0; j < n; j++)
5012 : {
5013 556297285 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &locators[j]))
5014 : {
5015 106518 : rlocator = &locators[j];
5016 106518 : break;
5017 : }
5018 : }
5019 : }
5020 : else
5021 : {
5022 : RelFileLocator locator;
5023 :
5024 2498816 : locator = BufTagGetRelFileLocator(&bufHdr->tag);
5025 2498816 : rlocator = bsearch(&locator,
5026 : locators, n, sizeof(RelFileLocator),
5027 : rlocator_comparator);
5028 : }
5029 :
5030 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
5031 186597504 : if (rlocator == NULL)
5032 186488822 : continue;
5033 :
5034 108682 : LockBufHdr(bufHdr);
5035 108682 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, rlocator))
5036 108682 : InvalidateBuffer(bufHdr); /* releases spinlock */
5037 : else
5038 0 : UnlockBufHdr(bufHdr);
5039 : }
5040 :
5041 15962 : pfree(locators);
5042 15962 : pfree(rels);
5043 : }
5044 :
5045 : /* ---------------------------------------------------------------------
5046 : * FindAndDropRelationBuffers
5047 : *
5048 : * This function performs look up in BufMapping table and removes from the
5049 : * buffer pool all the pages of the specified relation fork that has block
5050 : * number >= firstDelBlock. (In particular, with firstDelBlock = 0, all
5051 : * pages are removed.)
5052 : * --------------------------------------------------------------------
5053 : */
5054 : static void
5055 856 : FindAndDropRelationBuffers(RelFileLocator rlocator, ForkNumber forkNum,
5056 : BlockNumber nForkBlock,
5057 : BlockNumber firstDelBlock)
5058 : {
5059 : BlockNumber curBlock;
5060 :
5061 2056 : for (curBlock = firstDelBlock; curBlock < nForkBlock; curBlock++)
5062 : {
5063 : uint32 bufHash; /* hash value for tag */
5064 : BufferTag bufTag; /* identity of requested block */
5065 : LWLock *bufPartitionLock; /* buffer partition lock for it */
5066 : int buf_id;
5067 : BufferDesc *bufHdr;
5068 :
5069 : /* create a tag so we can lookup the buffer */
5070 1200 : InitBufferTag(&bufTag, &rlocator, forkNum, curBlock);
5071 :
5072 : /* determine its hash code and partition lock ID */
5073 1200 : bufHash = BufTableHashCode(&bufTag);
5074 1200 : bufPartitionLock = BufMappingPartitionLock(bufHash);
5075 :
5076 : /* Check that it is in the buffer pool. If not, do nothing. */
5077 1200 : LWLockAcquire(bufPartitionLock, LW_SHARED);
5078 1200 : buf_id = BufTableLookup(&bufTag, bufHash);
5079 1200 : LWLockRelease(bufPartitionLock);
5080 :
5081 1200 : if (buf_id < 0)
5082 143 : continue;
5083 :
5084 1057 : bufHdr = GetBufferDescriptor(buf_id);
5085 :
5086 : /*
5087 : * We need to lock the buffer header and recheck if the buffer is
5088 : * still associated with the same block because the buffer could be
5089 : * evicted by some other backend loading blocks for a different
5090 : * relation after we release lock on the BufMapping table.
5091 : */
5092 1057 : LockBufHdr(bufHdr);
5093 :
5094 2114 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator) &&
5095 1057 : BufTagGetForkNum(&bufHdr->tag) == forkNum &&
5096 1057 : bufHdr->tag.blockNum >= firstDelBlock)
5097 1057 : InvalidateBuffer(bufHdr); /* releases spinlock */
5098 : else
5099 0 : UnlockBufHdr(bufHdr);
5100 : }
5101 856 : }
5102 :
5103 : /* ---------------------------------------------------------------------
5104 : * DropDatabaseBuffers
5105 : *
5106 : * This function removes all the buffers in the buffer cache for a
5107 : * particular database. Dirty pages are simply dropped, without
5108 : * bothering to write them out first. This is used when we destroy a
5109 : * database, to avoid trying to flush data to disk when the directory
5110 : * tree no longer exists. Implementation is pretty similar to
5111 : * DropRelationBuffers() which is for destroying just one relation.
5112 : * --------------------------------------------------------------------
5113 : */
5114 : void
5115 91 : DropDatabaseBuffers(Oid dbid)
5116 : {
5117 : int i;
5118 :
5119 : /*
5120 : * We needn't consider local buffers, since by assumption the target
5121 : * database isn't our own.
5122 : */
5123 :
5124 727003 : for (i = 0; i < NBuffers; i++)
5125 : {
5126 726912 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5127 :
5128 : /*
5129 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5130 : * saves some cycles.
5131 : */
5132 726912 : if (bufHdr->tag.dbOid != dbid)
5133 705184 : continue;
5134 :
5135 21728 : LockBufHdr(bufHdr);
5136 21728 : if (bufHdr->tag.dbOid == dbid)
5137 21728 : InvalidateBuffer(bufHdr); /* releases spinlock */
5138 : else
5139 0 : UnlockBufHdr(bufHdr);
5140 : }
5141 91 : }
5142 :
5143 : /* ---------------------------------------------------------------------
5144 : * FlushRelationBuffers
5145 : *
5146 : * This function writes all dirty pages of a relation out to disk
5147 : * (or more accurately, out to kernel disk buffers), ensuring that the
5148 : * kernel has an up-to-date view of the relation.
5149 : *
5150 : * Generally, the caller should be holding AccessExclusiveLock on the
5151 : * target relation to ensure that no other backend is busy dirtying
5152 : * more blocks of the relation; the effects can't be expected to last
5153 : * after the lock is released.
5154 : *
5155 : * XXX currently it sequentially searches the buffer pool, should be
5156 : * changed to more clever ways of searching. This routine is not
5157 : * used in any performance-critical code paths, so it's not worth
5158 : * adding additional overhead to normal paths to make it go faster.
5159 : * --------------------------------------------------------------------
5160 : */
5161 : void
5162 174 : FlushRelationBuffers(Relation rel)
5163 : {
5164 : int i;
5165 : BufferDesc *bufHdr;
5166 174 : SMgrRelation srel = RelationGetSmgr(rel);
5167 :
5168 174 : if (RelationUsesLocalBuffers(rel))
5169 : {
5170 1212 : for (i = 0; i < NLocBuffer; i++)
5171 : {
5172 : uint64 buf_state;
5173 :
5174 1200 : bufHdr = GetLocalBufferDescriptor(i);
5175 1200 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
5176 400 : ((buf_state = pg_atomic_read_u64(&bufHdr->state)) &
5177 : (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5178 : {
5179 : ErrorContextCallback errcallback;
5180 :
5181 : /* Setup error traceback support for ereport() */
5182 392 : errcallback.callback = local_buffer_write_error_callback;
5183 392 : errcallback.arg = bufHdr;
5184 392 : errcallback.previous = error_context_stack;
5185 392 : error_context_stack = &errcallback;
5186 :
5187 : /* Make sure we can handle the pin */
5188 392 : ReservePrivateRefCountEntry();
5189 392 : ResourceOwnerEnlarge(CurrentResourceOwner);
5190 :
5191 : /*
5192 : * Pin/unpin mostly to make valgrind work, but it also seems
5193 : * like the right thing to do.
5194 : */
5195 392 : PinLocalBuffer(bufHdr, false);
5196 :
5197 :
5198 392 : FlushLocalBuffer(bufHdr, srel);
5199 :
5200 392 : UnpinLocalBuffer(BufferDescriptorGetBuffer(bufHdr));
5201 :
5202 : /* Pop the error context stack */
5203 392 : error_context_stack = errcallback.previous;
5204 : }
5205 : }
5206 :
5207 12 : return;
5208 : }
5209 :
5210 2052898 : for (i = 0; i < NBuffers; i++)
5211 : {
5212 : uint64 buf_state;
5213 :
5214 2052736 : bufHdr = GetBufferDescriptor(i);
5215 :
5216 : /*
5217 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5218 : * saves some cycles.
5219 : */
5220 2052736 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator))
5221 2052476 : continue;
5222 :
5223 : /* Make sure we can handle the pin */
5224 260 : ReservePrivateRefCountEntry();
5225 260 : ResourceOwnerEnlarge(CurrentResourceOwner);
5226 :
5227 260 : buf_state = LockBufHdr(bufHdr);
5228 260 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
5229 260 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5230 : {
5231 215 : PinBuffer_Locked(bufHdr);
5232 215 : FlushUnlockedBuffer(bufHdr, srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5233 215 : UnpinBuffer(bufHdr);
5234 : }
5235 : else
5236 45 : UnlockBufHdr(bufHdr);
5237 : }
5238 : }
5239 :
5240 : /* ---------------------------------------------------------------------
5241 : * FlushRelationsAllBuffers
5242 : *
5243 : * This function flushes out of the buffer pool all the pages of all
5244 : * forks of the specified smgr relations. It's equivalent to calling
5245 : * FlushRelationBuffers once per relation. The relations are assumed not
5246 : * to use local buffers.
5247 : * --------------------------------------------------------------------
5248 : */
5249 : void
5250 20 : FlushRelationsAllBuffers(SMgrRelation *smgrs, int nrels)
5251 : {
5252 : int i;
5253 : SMgrSortArray *srels;
5254 : bool use_bsearch;
5255 :
5256 20 : if (nrels == 0)
5257 0 : return;
5258 :
5259 : /* fill-in array for qsort */
5260 20 : srels = palloc_array(SMgrSortArray, nrels);
5261 :
5262 59 : for (i = 0; i < nrels; i++)
5263 : {
5264 : Assert(!RelFileLocatorBackendIsTemp(smgrs[i]->smgr_rlocator));
5265 :
5266 39 : srels[i].rlocator = smgrs[i]->smgr_rlocator.locator;
5267 39 : srels[i].srel = smgrs[i];
5268 : }
5269 :
5270 : /*
5271 : * Save the bsearch overhead for low number of relations to sync. See
5272 : * DropRelationsAllBuffers for details.
5273 : */
5274 20 : use_bsearch = nrels > RELS_BSEARCH_THRESHOLD;
5275 :
5276 : /* sort the list of SMgrRelations if necessary */
5277 20 : if (use_bsearch)
5278 0 : qsort(srels, nrels, sizeof(SMgrSortArray), rlocator_comparator);
5279 :
5280 327700 : for (i = 0; i < NBuffers; i++)
5281 : {
5282 327680 : SMgrSortArray *srelent = NULL;
5283 327680 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5284 : uint64 buf_state;
5285 :
5286 : /*
5287 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5288 : * saves some cycles.
5289 : */
5290 :
5291 327680 : if (!use_bsearch)
5292 : {
5293 : int j;
5294 :
5295 962164 : for (j = 0; j < nrels; j++)
5296 : {
5297 638925 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srels[j].rlocator))
5298 : {
5299 4441 : srelent = &srels[j];
5300 4441 : break;
5301 : }
5302 : }
5303 : }
5304 : else
5305 : {
5306 : RelFileLocator rlocator;
5307 :
5308 0 : rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
5309 0 : srelent = bsearch(&rlocator,
5310 : srels, nrels, sizeof(SMgrSortArray),
5311 : rlocator_comparator);
5312 : }
5313 :
5314 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
5315 327680 : if (srelent == NULL)
5316 323239 : continue;
5317 :
5318 : /* Make sure we can handle the pin */
5319 4441 : ReservePrivateRefCountEntry();
5320 4441 : ResourceOwnerEnlarge(CurrentResourceOwner);
5321 :
5322 4441 : buf_state = LockBufHdr(bufHdr);
5323 4441 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srelent->rlocator) &&
5324 4441 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5325 : {
5326 3982 : PinBuffer_Locked(bufHdr);
5327 3982 : FlushUnlockedBuffer(bufHdr, srelent->srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5328 3982 : UnpinBuffer(bufHdr);
5329 : }
5330 : else
5331 459 : UnlockBufHdr(bufHdr);
5332 : }
5333 :
5334 20 : pfree(srels);
5335 : }
5336 :
5337 : /* ---------------------------------------------------------------------
5338 : * RelationCopyStorageUsingBuffer
5339 : *
5340 : * Copy fork's data using bufmgr. Same as RelationCopyStorage but instead
5341 : * of using smgrread and smgrextend this will copy using bufmgr APIs.
5342 : *
5343 : * Refer comments atop CreateAndCopyRelationData() for details about
5344 : * 'permanent' parameter.
5345 : * --------------------------------------------------------------------
5346 : */
5347 : static void
5348 88118 : RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
5349 : RelFileLocator dstlocator,
5350 : ForkNumber forkNum, bool permanent)
5351 : {
5352 : Buffer srcBuf;
5353 : Buffer dstBuf;
5354 : Page srcPage;
5355 : Page dstPage;
5356 : bool use_wal;
5357 : BlockNumber nblocks;
5358 : BlockNumber blkno;
5359 : PGIOAlignedBlock buf;
5360 : BufferAccessStrategy bstrategy_src;
5361 : BufferAccessStrategy bstrategy_dst;
5362 : BlockRangeReadStreamPrivate p;
5363 : ReadStream *src_stream;
5364 : SMgrRelation src_smgr;
5365 :
5366 : /*
5367 : * In general, we want to write WAL whenever wal_level > 'minimal', but we
5368 : * can skip it when copying any fork of an unlogged relation other than
5369 : * the init fork.
5370 : */
5371 88118 : use_wal = XLogIsNeeded() && (permanent || forkNum == INIT_FORKNUM);
5372 :
5373 : /* Get number of blocks in the source relation. */
5374 88118 : nblocks = smgrnblocks(smgropen(srclocator, INVALID_PROC_NUMBER),
5375 : forkNum);
5376 :
5377 : /* Nothing to copy; just return. */
5378 88118 : if (nblocks == 0)
5379 16346 : return;
5380 :
5381 : /*
5382 : * Bulk extend the destination relation of the same size as the source
5383 : * relation before starting to copy block by block.
5384 : */
5385 71772 : memset(buf.data, 0, BLCKSZ);
5386 71772 : smgrextend(smgropen(dstlocator, INVALID_PROC_NUMBER), forkNum, nblocks - 1,
5387 : buf.data, true);
5388 :
5389 : /* This is a bulk operation, so use buffer access strategies. */
5390 71772 : bstrategy_src = GetAccessStrategy(BAS_BULKREAD);
5391 71772 : bstrategy_dst = GetAccessStrategy(BAS_BULKWRITE);
5392 :
5393 : /* Initialize streaming read */
5394 71772 : p.current_blocknum = 0;
5395 71772 : p.last_exclusive = nblocks;
5396 71772 : src_smgr = smgropen(srclocator, INVALID_PROC_NUMBER);
5397 :
5398 : /*
5399 : * It is safe to use batchmode as block_range_read_stream_cb takes no
5400 : * locks.
5401 : */
5402 71772 : src_stream = read_stream_begin_smgr_relation(READ_STREAM_FULL |
5403 : READ_STREAM_USE_BATCHING,
5404 : bstrategy_src,
5405 : src_smgr,
5406 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
5407 : forkNum,
5408 : block_range_read_stream_cb,
5409 : &p,
5410 : 0);
5411 :
5412 : /* Iterate over each block of the source relation file. */
5413 339792 : for (blkno = 0; blkno < nblocks; blkno++)
5414 : {
5415 268022 : CHECK_FOR_INTERRUPTS();
5416 :
5417 : /* Read block from source relation. */
5418 268022 : srcBuf = read_stream_next_buffer(src_stream, NULL);
5419 268020 : LockBuffer(srcBuf, BUFFER_LOCK_SHARE);
5420 268020 : srcPage = BufferGetPage(srcBuf);
5421 :
5422 268020 : dstBuf = ReadBufferWithoutRelcache(dstlocator, forkNum,
5423 : BufferGetBlockNumber(srcBuf),
5424 : RBM_ZERO_AND_LOCK, bstrategy_dst,
5425 : permanent);
5426 268020 : dstPage = BufferGetPage(dstBuf);
5427 :
5428 268020 : START_CRIT_SECTION();
5429 :
5430 : /* Copy page data from the source to the destination. */
5431 268020 : memcpy(dstPage, srcPage, BLCKSZ);
5432 268020 : MarkBufferDirty(dstBuf);
5433 :
5434 : /* WAL-log the copied page. */
5435 268020 : if (use_wal)
5436 150970 : log_newpage_buffer(dstBuf, true);
5437 :
5438 268020 : END_CRIT_SECTION();
5439 :
5440 268020 : UnlockReleaseBuffer(dstBuf);
5441 268020 : UnlockReleaseBuffer(srcBuf);
5442 : }
5443 : Assert(read_stream_next_buffer(src_stream, NULL) == InvalidBuffer);
5444 71770 : read_stream_end(src_stream);
5445 :
5446 71770 : FreeAccessStrategy(bstrategy_src);
5447 71770 : FreeAccessStrategy(bstrategy_dst);
5448 : }
5449 :
5450 : /* ---------------------------------------------------------------------
5451 : * CreateAndCopyRelationData
5452 : *
5453 : * Create destination relation storage and copy all forks from the
5454 : * source relation to the destination.
5455 : *
5456 : * Pass permanent as true for permanent relations and false for
5457 : * unlogged relations. Currently this API is not supported for
5458 : * temporary relations.
5459 : * --------------------------------------------------------------------
5460 : */
5461 : void
5462 67616 : CreateAndCopyRelationData(RelFileLocator src_rlocator,
5463 : RelFileLocator dst_rlocator, bool permanent)
5464 : {
5465 : char relpersistence;
5466 : SMgrRelation src_rel;
5467 : SMgrRelation dst_rel;
5468 :
5469 : /* Set the relpersistence. */
5470 67616 : relpersistence = permanent ?
5471 : RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED;
5472 :
5473 67616 : src_rel = smgropen(src_rlocator, INVALID_PROC_NUMBER);
5474 67616 : dst_rel = smgropen(dst_rlocator, INVALID_PROC_NUMBER);
5475 :
5476 : /*
5477 : * Create and copy all forks of the relation. During create database we
5478 : * have a separate cleanup mechanism which deletes complete database
5479 : * directory. Therefore, each individual relation doesn't need to be
5480 : * registered for cleanup.
5481 : */
5482 67616 : RelationCreateStorage(dst_rlocator, relpersistence, false);
5483 :
5484 : /* copy main fork. */
5485 67616 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, MAIN_FORKNUM,
5486 : permanent);
5487 :
5488 : /* copy those extra forks that exist */
5489 67614 : for (ForkNumber forkNum = MAIN_FORKNUM + 1;
5490 270456 : forkNum <= MAX_FORKNUM; forkNum++)
5491 : {
5492 202842 : if (smgrexists(src_rel, forkNum))
5493 : {
5494 20502 : smgrcreate(dst_rel, forkNum, false);
5495 :
5496 : /*
5497 : * WAL log creation if the relation is persistent, or this is the
5498 : * init fork of an unlogged relation.
5499 : */
5500 20502 : if (permanent || forkNum == INIT_FORKNUM)
5501 20502 : log_smgrcreate(&dst_rlocator, forkNum);
5502 :
5503 : /* Copy a fork's data, block by block. */
5504 20502 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, forkNum,
5505 : permanent);
5506 : }
5507 : }
5508 67614 : }
5509 :
5510 : /* ---------------------------------------------------------------------
5511 : * FlushDatabaseBuffers
5512 : *
5513 : * This function writes all dirty pages of a database out to disk
5514 : * (or more accurately, out to kernel disk buffers), ensuring that the
5515 : * kernel has an up-to-date view of the database.
5516 : *
5517 : * Generally, the caller should be holding an appropriate lock to ensure
5518 : * no other backend is active in the target database; otherwise more
5519 : * pages could get dirtied.
5520 : *
5521 : * Note we don't worry about flushing any pages of temporary relations.
5522 : * It's assumed these wouldn't be interesting.
5523 : * --------------------------------------------------------------------
5524 : */
5525 : void
5526 5 : FlushDatabaseBuffers(Oid dbid)
5527 : {
5528 : int i;
5529 : BufferDesc *bufHdr;
5530 :
5531 645 : for (i = 0; i < NBuffers; i++)
5532 : {
5533 : uint64 buf_state;
5534 :
5535 640 : bufHdr = GetBufferDescriptor(i);
5536 :
5537 : /*
5538 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5539 : * saves some cycles.
5540 : */
5541 640 : if (bufHdr->tag.dbOid != dbid)
5542 492 : continue;
5543 :
5544 : /* Make sure we can handle the pin */
5545 148 : ReservePrivateRefCountEntry();
5546 148 : ResourceOwnerEnlarge(CurrentResourceOwner);
5547 :
5548 148 : buf_state = LockBufHdr(bufHdr);
5549 148 : if (bufHdr->tag.dbOid == dbid &&
5550 148 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5551 : {
5552 21 : PinBuffer_Locked(bufHdr);
5553 21 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5554 21 : UnpinBuffer(bufHdr);
5555 : }
5556 : else
5557 127 : UnlockBufHdr(bufHdr);
5558 : }
5559 5 : }
5560 :
5561 : /*
5562 : * Flush a previously, share-exclusively or exclusively, locked and pinned
5563 : * buffer to the OS.
5564 : */
5565 : void
5566 81 : FlushOneBuffer(Buffer buffer)
5567 : {
5568 : BufferDesc *bufHdr;
5569 :
5570 : /* currently not needed, but no fundamental reason not to support */
5571 : Assert(!BufferIsLocal(buffer));
5572 :
5573 : Assert(BufferIsPinned(buffer));
5574 :
5575 81 : bufHdr = GetBufferDescriptor(buffer - 1);
5576 :
5577 : Assert(BufferIsLockedByMe(buffer));
5578 :
5579 81 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5580 81 : }
5581 :
5582 : /*
5583 : * ReleaseBuffer -- release the pin on a buffer
5584 : */
5585 : void
5586 50419066 : ReleaseBuffer(Buffer buffer)
5587 : {
5588 50419066 : if (!BufferIsValid(buffer))
5589 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5590 :
5591 50419066 : if (BufferIsLocal(buffer))
5592 751973 : UnpinLocalBuffer(buffer);
5593 : else
5594 49667093 : UnpinBuffer(GetBufferDescriptor(buffer - 1));
5595 50419066 : }
5596 :
5597 : /*
5598 : * UnlockReleaseBuffer -- release the content lock and pin on a buffer
5599 : *
5600 : * This is just a, more efficient, shorthand for a common combination.
5601 : */
5602 : void
5603 49965694 : UnlockReleaseBuffer(Buffer buffer)
5604 : {
5605 : int mode;
5606 : BufferDesc *buf;
5607 : PrivateRefCountEntry *ref;
5608 : uint64 sub;
5609 : uint64 lockstate;
5610 :
5611 : Assert(BufferIsPinned(buffer));
5612 :
5613 49965694 : if (BufferIsLocal(buffer))
5614 : {
5615 1381532 : UnpinLocalBuffer(buffer);
5616 1381532 : return;
5617 : }
5618 :
5619 48584162 : ResourceOwnerForgetBuffer(CurrentResourceOwner, buffer);
5620 :
5621 48584162 : buf = GetBufferDescriptor(buffer - 1);
5622 :
5623 48584162 : mode = BufferLockDisownInternal(buffer, buf);
5624 :
5625 : /* compute state modification for lock release */
5626 48584162 : sub = BufferLockReleaseSub(mode);
5627 :
5628 : /* compute state modification for pin release */
5629 48584162 : ref = GetPrivateRefCountEntry(buffer, false);
5630 : Assert(ref != NULL);
5631 : Assert(ref->data.refcount > 0);
5632 48584162 : ref->data.refcount--;
5633 :
5634 : /* no more backend local pins, reduce shared pin count */
5635 48584162 : if (likely(ref->data.refcount == 0))
5636 : {
5637 : /* See comment in UnpinBufferNoOwner() */
5638 : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
5639 :
5640 46210014 : sub |= BUF_REFCOUNT_ONE;
5641 46210014 : ForgetPrivateRefCountEntry(ref);
5642 : }
5643 :
5644 : /* perform the lock and pin release in one atomic op */
5645 48584162 : lockstate = pg_atomic_sub_fetch_u64(&buf->state, sub);
5646 :
5647 : /* wake up waiters for the lock */
5648 48584162 : BufferLockProcessRelease(buf, mode, lockstate);
5649 :
5650 : /* wake up waiter for the pin release */
5651 48584162 : if (lockstate & BM_PIN_COUNT_WAITER)
5652 0 : WakePinCountWaiter(buf);
5653 :
5654 : /*
5655 : * Now okay to allow cancel/die interrupts again, which were held when the
5656 : * lock was acquired.
5657 : */
5658 48584162 : RESUME_INTERRUPTS();
5659 : }
5660 :
5661 : /*
5662 : * IncrBufferRefCount
5663 : * Increment the pin count on a buffer that we have *already* pinned
5664 : * at least once.
5665 : *
5666 : * This function cannot be used on a buffer we do not have pinned,
5667 : * because it doesn't change the shared buffer state.
5668 : */
5669 : void
5670 15314371 : IncrBufferRefCount(Buffer buffer)
5671 : {
5672 : Assert(BufferIsPinned(buffer));
5673 15314371 : ResourceOwnerEnlarge(CurrentResourceOwner);
5674 15314371 : if (BufferIsLocal(buffer))
5675 469678 : LocalRefCount[-buffer - 1]++;
5676 : else
5677 : {
5678 : PrivateRefCountEntry *ref;
5679 :
5680 14844693 : ref = GetPrivateRefCountEntry(buffer, true);
5681 : Assert(ref != NULL);
5682 14844693 : ref->data.refcount++;
5683 : }
5684 15314371 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buffer);
5685 15314371 : }
5686 :
5687 : /*
5688 : * Shared-buffer only helper for MarkBufferDirtyHint() and
5689 : * BufferSetHintBits16().
5690 : *
5691 : * This is separated out because it turns out that the repeated checks for
5692 : * local buffers, repeated GetBufferDescriptor() and repeated reading of the
5693 : * buffer's state sufficiently hurts the performance of BufferSetHintBits16().
5694 : */
5695 : static inline void
5696 15308201 : MarkSharedBufferDirtyHint(Buffer buffer, BufferDesc *bufHdr, uint64 lockstate,
5697 : bool buffer_std)
5698 : {
5699 15308201 : Page page = BufferGetPage(buffer);
5700 :
5701 : Assert(GetPrivateRefCount(buffer) > 0);
5702 :
5703 : /* here, either share-exclusive or exclusive lock is OK */
5704 : Assert(BufferLockHeldByMeInMode(bufHdr, BUFFER_LOCK_EXCLUSIVE) ||
5705 : BufferLockHeldByMeInMode(bufHdr, BUFFER_LOCK_SHARE_EXCLUSIVE));
5706 :
5707 : /*
5708 : * This routine might get called many times on the same page, if we are
5709 : * making the first scan after commit of an xact that added/deleted many
5710 : * tuples. So, be as quick as we can if the buffer is already dirty.
5711 : *
5712 : * As we are holding (at least) a share-exclusive lock, nobody could have
5713 : * cleaned or dirtied the page concurrently, so we can just rely on the
5714 : * previously fetched value here without any danger of races.
5715 : */
5716 15308201 : if (unlikely(!(lockstate & BM_DIRTY)))
5717 : {
5718 675096 : XLogRecPtr lsn = InvalidXLogRecPtr;
5719 675096 : bool wal_log = false;
5720 : uint64 buf_state;
5721 :
5722 : /*
5723 : * If we need to protect hint bit updates from torn writes, WAL-log a
5724 : * full page image of the page. This full page image is only necessary
5725 : * if the hint bit update is the first change to the page since the
5726 : * last checkpoint.
5727 : *
5728 : * We don't check full_page_writes here because that logic is included
5729 : * when we call XLogInsert() since the value changes dynamically.
5730 : */
5731 675096 : if (XLogHintBitIsNeeded() && (lockstate & BM_PERMANENT))
5732 : {
5733 : /*
5734 : * If we must not write WAL, due to a relfilelocator-specific
5735 : * condition or being in recovery, don't dirty the page. We can
5736 : * set the hint, just not dirty the page as a result so the hint
5737 : * is lost when we evict the page or shutdown.
5738 : *
5739 : * See src/backend/storage/page/README for longer discussion.
5740 : */
5741 761868 : if (RecoveryInProgress() ||
5742 88754 : RelFileLocatorSkippingWAL(BufTagGetRelFileLocator(&bufHdr->tag)))
5743 587147 : return;
5744 :
5745 85967 : wal_log = true;
5746 : }
5747 :
5748 : /*
5749 : * We must mark the page dirty before we emit the WAL record, as per
5750 : * the usual rules, to ensure that BufferSync()/SyncOneBuffer() try to
5751 : * flush the buffer, even if we haven't inserted the WAL record yet.
5752 : * As we hold at least a share-exclusive lock, checkpoints will wait
5753 : * for this backend to be done with the buffer before continuing. If
5754 : * we did it the other way round, a checkpoint could start between
5755 : * writing the WAL record and marking the buffer dirty.
5756 : */
5757 87949 : buf_state = LockBufHdr(bufHdr);
5758 :
5759 : /*
5760 : * It should not be possible for the buffer to already be dirty, see
5761 : * comment above.
5762 : */
5763 : Assert(!(buf_state & BM_DIRTY));
5764 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5765 87949 : UnlockBufHdrExt(bufHdr, buf_state,
5766 : BM_DIRTY,
5767 : 0, 0);
5768 :
5769 : /*
5770 : * If the block is already dirty because we either made a change or
5771 : * set a hint already, then we don't need to write a full page image.
5772 : * Note that aggressive cleaning of blocks dirtied by hint bit setting
5773 : * would increase the call rate. Bulk setting of hint bits would
5774 : * reduce the call rate...
5775 : */
5776 87949 : if (wal_log)
5777 85967 : lsn = XLogSaveBufferForHint(buffer, buffer_std);
5778 :
5779 87949 : if (XLogRecPtrIsValid(lsn))
5780 : {
5781 : /*
5782 : * Set the page LSN if we wrote a backup block. To allow backends
5783 : * that only hold a share lock on the buffer to read the LSN in a
5784 : * tear-free manner, we set the page LSN while holding the buffer
5785 : * header lock. This allows any reader of an LSN who holds only a
5786 : * share lock to also obtain a buffer header lock before using
5787 : * PageGetLSN() to read the LSN in a tear free way. This is done
5788 : * in BufferGetLSNAtomic().
5789 : *
5790 : * If checksums are enabled, you might think we should reset the
5791 : * checksum here. That will happen when the page is written
5792 : * sometime later in this checkpoint cycle.
5793 : */
5794 59516 : buf_state = LockBufHdr(bufHdr);
5795 59516 : PageSetLSN(page, lsn);
5796 59516 : UnlockBufHdr(bufHdr);
5797 : }
5798 :
5799 87949 : pgBufferUsage.shared_blks_dirtied++;
5800 87949 : if (VacuumCostActive)
5801 1141 : VacuumCostBalance += VacuumCostPageDirty;
5802 : }
5803 : }
5804 :
5805 : /*
5806 : * MarkBufferDirtyHint
5807 : *
5808 : * Mark a buffer dirty for non-critical changes.
5809 : *
5810 : * This is essentially the same as MarkBufferDirty, except:
5811 : *
5812 : * 1. The caller does not write WAL; so if checksums are enabled, we may need
5813 : * to write an XLOG_FPI_FOR_HINT WAL record to protect against torn pages.
5814 : * 2. The caller might have only a share-exclusive-lock instead of an
5815 : * exclusive-lock on the buffer's content lock.
5816 : * 3. This function does not guarantee that the buffer is always marked dirty
5817 : * (it e.g. can't always on a hot standby), so it cannot be used for
5818 : * important changes.
5819 : */
5820 : inline void
5821 734964 : MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
5822 : {
5823 : BufferDesc *bufHdr;
5824 :
5825 734964 : bufHdr = GetBufferDescriptor(buffer - 1);
5826 :
5827 734964 : if (!BufferIsValid(buffer))
5828 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5829 :
5830 734964 : if (BufferIsLocal(buffer))
5831 : {
5832 22023 : MarkLocalBufferDirty(buffer);
5833 22023 : return;
5834 : }
5835 :
5836 712941 : MarkSharedBufferDirtyHint(buffer, bufHdr,
5837 712941 : pg_atomic_read_u64(&bufHdr->state),
5838 : buffer_std);
5839 : }
5840 :
5841 : /*
5842 : * Release buffer content locks for shared buffers.
5843 : *
5844 : * Used to clean up after errors.
5845 : *
5846 : * Currently, we can expect that resource owner cleanup, via
5847 : * ResOwnerReleaseBufferPin(), took care of releasing buffer content locks per
5848 : * se; the only thing we need to deal with here is clearing any PIN_COUNT
5849 : * request that was in progress.
5850 : */
5851 : void
5852 66091 : UnlockBuffers(void)
5853 : {
5854 66091 : BufferDesc *buf = PinCountWaitBuf;
5855 :
5856 66091 : if (buf)
5857 : {
5858 : uint64 buf_state;
5859 0 : uint64 unset_bits = 0;
5860 :
5861 0 : buf_state = LockBufHdr(buf);
5862 :
5863 : /*
5864 : * Don't complain if flag bit not set; it could have been reset but we
5865 : * got a cancel/die interrupt before getting the signal.
5866 : */
5867 0 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5868 0 : buf->wait_backend_pgprocno == MyProcNumber)
5869 0 : unset_bits = BM_PIN_COUNT_WAITER;
5870 :
5871 0 : UnlockBufHdrExt(buf, buf_state,
5872 : 0, unset_bits,
5873 : 0);
5874 :
5875 0 : PinCountWaitBuf = NULL;
5876 : }
5877 66091 : }
5878 :
5879 : /*
5880 : * Acquire the buffer content lock in the specified mode
5881 : *
5882 : * If the lock is not available, sleep until it is.
5883 : *
5884 : * Side effect: cancel/die interrupts are held off until lock release.
5885 : *
5886 : * This uses almost the same locking approach as lwlock.c's
5887 : * LWLockAcquire(). See documentation at the top of lwlock.c for a more
5888 : * detailed discussion.
5889 : *
5890 : * The reason that this, and most of the other BufferLock* functions, get both
5891 : * the Buffer and BufferDesc* as parameters, is that looking up one from the
5892 : * other repeatedly shows up noticeably in profiles.
5893 : *
5894 : * Callers should provide a constant for mode, for more efficient code
5895 : * generation.
5896 : */
5897 : static inline void
5898 111173174 : BufferLockAcquire(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode)
5899 : {
5900 : PrivateRefCountEntry *entry;
5901 111173174 : int extraWaits = 0;
5902 :
5903 : /*
5904 : * Get reference to the refcount entry before we hold the lock, it seems
5905 : * better to do before holding the lock.
5906 : */
5907 111173174 : entry = GetPrivateRefCountEntry(buffer, true);
5908 :
5909 : /*
5910 : * We better not already hold a lock on the buffer.
5911 : */
5912 : Assert(entry->data.lockmode == BUFFER_LOCK_UNLOCK);
5913 :
5914 : /*
5915 : * Lock out cancel/die interrupts until we exit the code section protected
5916 : * by the content lock. This ensures that interrupts will not interfere
5917 : * with manipulations of data structures in shared memory.
5918 : */
5919 111173174 : HOLD_INTERRUPTS();
5920 :
5921 : for (;;)
5922 23388 : {
5923 111196562 : uint32 wait_event = 0; /* initialized to avoid compiler warning */
5924 : bool mustwait;
5925 :
5926 : /*
5927 : * Try to grab the lock the first time, we're not in the waitqueue
5928 : * yet/anymore.
5929 : */
5930 111196562 : mustwait = BufferLockAttempt(buf_hdr, mode);
5931 :
5932 111196562 : if (likely(!mustwait))
5933 : {
5934 111171361 : break;
5935 : }
5936 :
5937 : /*
5938 : * Ok, at this point we couldn't grab the lock on the first try. We
5939 : * cannot simply queue ourselves to the end of the list and wait to be
5940 : * woken up because by now the lock could long have been released.
5941 : * Instead add us to the queue and try to grab the lock again. If we
5942 : * succeed we need to revert the queuing and be happy, otherwise we
5943 : * recheck the lock. If we still couldn't grab it, we know that the
5944 : * other locker will see our queue entries when releasing since they
5945 : * existed before we checked for the lock.
5946 : */
5947 :
5948 : /* add to the queue */
5949 25201 : BufferLockQueueSelf(buf_hdr, mode);
5950 :
5951 : /* we're now guaranteed to be woken up if necessary */
5952 25201 : mustwait = BufferLockAttempt(buf_hdr, mode);
5953 :
5954 : /* ok, grabbed the lock the second time round, need to undo queueing */
5955 25201 : if (!mustwait)
5956 : {
5957 1813 : BufferLockDequeueSelf(buf_hdr);
5958 1813 : break;
5959 : }
5960 :
5961 23388 : switch (mode)
5962 : {
5963 12896 : case BUFFER_LOCK_EXCLUSIVE:
5964 12896 : wait_event = WAIT_EVENT_BUFFER_EXCLUSIVE;
5965 12896 : break;
5966 114 : case BUFFER_LOCK_SHARE_EXCLUSIVE:
5967 114 : wait_event = WAIT_EVENT_BUFFER_SHARE_EXCLUSIVE;
5968 114 : break;
5969 10378 : case BUFFER_LOCK_SHARE:
5970 10378 : wait_event = WAIT_EVENT_BUFFER_SHARED;
5971 10378 : break;
5972 : case BUFFER_LOCK_UNLOCK:
5973 : pg_unreachable();
5974 :
5975 : }
5976 23388 : pgstat_report_wait_start(wait_event);
5977 :
5978 : /*
5979 : * Wait until awakened.
5980 : *
5981 : * It is possible that we get awakened for a reason other than being
5982 : * signaled by BufferLockWakeup(). If so, loop back and wait again.
5983 : * Once we've gotten the lock, re-increment the sema by the number of
5984 : * additional signals received.
5985 : */
5986 : for (;;)
5987 : {
5988 23388 : PGSemaphoreLock(MyProc->sem);
5989 23388 : if (MyProc->lwWaiting == LW_WS_NOT_WAITING)
5990 23388 : break;
5991 0 : extraWaits++;
5992 : }
5993 :
5994 23388 : pgstat_report_wait_end();
5995 :
5996 : /* Retrying, allow BufferLockRelease to release waiters again. */
5997 23388 : pg_atomic_fetch_and_u64(&buf_hdr->state, ~BM_LOCK_WAKE_IN_PROGRESS);
5998 : }
5999 :
6000 : /* Remember that we now hold this lock */
6001 111173174 : entry->data.lockmode = mode;
6002 :
6003 : /*
6004 : * Fix the process wait semaphore's count for any absorbed wakeups.
6005 : */
6006 111173174 : while (unlikely(extraWaits-- > 0))
6007 0 : PGSemaphoreUnlock(MyProc->sem);
6008 111173174 : }
6009 :
6010 : /*
6011 : * Release a previously acquired buffer content lock.
6012 : */
6013 : static void
6014 64656069 : BufferLockUnlock(Buffer buffer, BufferDesc *buf_hdr)
6015 : {
6016 : BufferLockMode mode;
6017 : uint64 oldstate;
6018 : uint64 sub;
6019 :
6020 64656069 : mode = BufferLockDisownInternal(buffer, buf_hdr);
6021 :
6022 : /*
6023 : * Release my hold on lock, after that it can immediately be acquired by
6024 : * others, even if we still have to wakeup other waiters.
6025 : */
6026 64656069 : sub = BufferLockReleaseSub(mode);
6027 :
6028 64656069 : oldstate = pg_atomic_sub_fetch_u64(&buf_hdr->state, sub);
6029 :
6030 64656069 : BufferLockProcessRelease(buf_hdr, mode, oldstate);
6031 :
6032 : /*
6033 : * Now okay to allow cancel/die interrupts.
6034 : */
6035 64656069 : RESUME_INTERRUPTS();
6036 64656069 : }
6037 :
6038 :
6039 : /*
6040 : * Acquire the content lock for the buffer, but only if we don't have to wait.
6041 : *
6042 : * It is allowed to try to conditionally acquire a lock on a buffer that this
6043 : * backend has already locked, but the lock acquisition will always fail, even
6044 : * if the new lock acquisition does not conflict with an already held lock
6045 : * (e.g. two share locks). This is because we currently do not have space to
6046 : * track multiple lock ownerships of the same buffer within one backend. That
6047 : * is ok for the current uses of BufferLockConditional().
6048 : */
6049 : static bool
6050 2069008 : BufferLockConditional(Buffer buffer, BufferDesc *buf_hdr, BufferLockMode mode)
6051 : {
6052 2069008 : PrivateRefCountEntry *entry = GetPrivateRefCountEntry(buffer, true);
6053 : bool mustwait;
6054 :
6055 : /*
6056 : * As described above, if we're trying to lock a buffer this backend
6057 : * already has locked, return false, independent of the existing and
6058 : * desired lock level.
6059 : */
6060 2069008 : if (entry->data.lockmode != BUFFER_LOCK_UNLOCK)
6061 0 : return false;
6062 :
6063 : /*
6064 : * Lock out cancel/die interrupts until we exit the code section protected
6065 : * by the content lock. This ensures that interrupts will not interfere
6066 : * with manipulations of data structures in shared memory.
6067 : */
6068 2069008 : HOLD_INTERRUPTS();
6069 :
6070 : /* Check for the lock */
6071 2069008 : mustwait = BufferLockAttempt(buf_hdr, mode);
6072 :
6073 2069008 : if (mustwait)
6074 : {
6075 : /* Failed to get lock, so release interrupt holdoff */
6076 1951 : RESUME_INTERRUPTS();
6077 : }
6078 : else
6079 : {
6080 2067057 : entry->data.lockmode = mode;
6081 : }
6082 :
6083 2069008 : return !mustwait;
6084 : }
6085 :
6086 : /*
6087 : * Internal function that tries to atomically acquire the content lock in the
6088 : * passed in mode.
6089 : *
6090 : * This function will not block waiting for a lock to become free - that's the
6091 : * caller's job.
6092 : *
6093 : * Similar to LWLockAttemptLock().
6094 : */
6095 : static inline bool
6096 113290771 : BufferLockAttempt(BufferDesc *buf_hdr, BufferLockMode mode)
6097 : {
6098 : uint64 old_state;
6099 :
6100 : /*
6101 : * Read once outside the loop, later iterations will get the newer value
6102 : * via compare & exchange.
6103 : */
6104 113290771 : old_state = pg_atomic_read_u64(&buf_hdr->state);
6105 :
6106 : /* loop until we've determined whether we could acquire the lock or not */
6107 : while (true)
6108 16773 : {
6109 : uint64 desired_state;
6110 : bool lock_free;
6111 :
6112 113307544 : desired_state = old_state;
6113 :
6114 113307544 : if (mode == BUFFER_LOCK_EXCLUSIVE)
6115 : {
6116 38210074 : lock_free = (old_state & BM_LOCK_MASK) == 0;
6117 38210074 : if (lock_free)
6118 38180852 : desired_state += BM_LOCK_VAL_EXCLUSIVE;
6119 : }
6120 75097470 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6121 : {
6122 726817 : lock_free = (old_state & (BM_LOCK_VAL_EXCLUSIVE | BM_LOCK_VAL_SHARE_EXCLUSIVE)) == 0;
6123 726817 : if (lock_free)
6124 726586 : desired_state += BM_LOCK_VAL_SHARE_EXCLUSIVE;
6125 : }
6126 : else
6127 : {
6128 74370653 : lock_free = (old_state & BM_LOCK_VAL_EXCLUSIVE) == 0;
6129 74370653 : if (lock_free)
6130 74349021 : desired_state += BM_LOCK_VAL_SHARED;
6131 : }
6132 :
6133 : /*
6134 : * Attempt to swap in the state we are expecting. If we didn't see
6135 : * lock to be free, that's just the old value. If we saw it as free,
6136 : * we'll attempt to mark it acquired. The reason that we always swap
6137 : * in the value is that this doubles as a memory barrier. We could try
6138 : * to be smarter and only swap in values if we saw the lock as free,
6139 : * but benchmark haven't shown it as beneficial so far.
6140 : *
6141 : * Retry if the value changed since we last looked at it.
6142 : */
6143 113307544 : if (likely(pg_atomic_compare_exchange_u64(&buf_hdr->state,
6144 : &old_state, desired_state)))
6145 : {
6146 113290771 : if (lock_free)
6147 : {
6148 : /* Great! Got the lock. */
6149 113240231 : return false;
6150 : }
6151 : else
6152 50540 : return true; /* somebody else has the lock */
6153 : }
6154 : }
6155 :
6156 : pg_unreachable();
6157 : }
6158 :
6159 : /*
6160 : * Add ourselves to the end of the content lock's wait queue.
6161 : */
6162 : static void
6163 25201 : BufferLockQueueSelf(BufferDesc *buf_hdr, BufferLockMode mode)
6164 : {
6165 : /*
6166 : * If we don't have a PGPROC structure, there's no way to wait. This
6167 : * should never occur, since MyProc should only be null during shared
6168 : * memory initialization.
6169 : */
6170 25201 : if (MyProc == NULL)
6171 0 : elog(PANIC, "cannot wait without a PGPROC structure");
6172 :
6173 25201 : if (MyProc->lwWaiting != LW_WS_NOT_WAITING)
6174 0 : elog(PANIC, "queueing for lock while waiting on another one");
6175 :
6176 25201 : LockBufHdr(buf_hdr);
6177 :
6178 : /* setting the flag is protected by the spinlock */
6179 25201 : pg_atomic_fetch_or_u64(&buf_hdr->state, BM_LOCK_HAS_WAITERS);
6180 :
6181 : /*
6182 : * These are currently used both for lwlocks and buffer content locks,
6183 : * which is acceptable, although not pretty, because a backend can't wait
6184 : * for both types of locks at the same time.
6185 : */
6186 25201 : MyProc->lwWaiting = LW_WS_WAITING;
6187 25201 : MyProc->lwWaitMode = mode;
6188 :
6189 25201 : proclist_push_tail(&buf_hdr->lock_waiters, MyProcNumber, lwWaitLink);
6190 :
6191 : /* Can release the mutex now */
6192 25201 : UnlockBufHdr(buf_hdr);
6193 25201 : }
6194 :
6195 : /*
6196 : * Remove ourselves from the waitlist.
6197 : *
6198 : * This is used if we queued ourselves because we thought we needed to sleep
6199 : * but, after further checking, we discovered that we don't actually need to
6200 : * do so.
6201 : */
6202 : static void
6203 1813 : BufferLockDequeueSelf(BufferDesc *buf_hdr)
6204 : {
6205 : bool on_waitlist;
6206 :
6207 1813 : LockBufHdr(buf_hdr);
6208 :
6209 1813 : on_waitlist = MyProc->lwWaiting == LW_WS_WAITING;
6210 1813 : if (on_waitlist)
6211 1333 : proclist_delete(&buf_hdr->lock_waiters, MyProcNumber, lwWaitLink);
6212 :
6213 1813 : if (proclist_is_empty(&buf_hdr->lock_waiters) &&
6214 1751 : (pg_atomic_read_u64(&buf_hdr->state) & BM_LOCK_HAS_WAITERS) != 0)
6215 : {
6216 1274 : pg_atomic_fetch_and_u64(&buf_hdr->state, ~BM_LOCK_HAS_WAITERS);
6217 : }
6218 :
6219 : /* XXX: combine with fetch_and above? */
6220 1813 : UnlockBufHdr(buf_hdr);
6221 :
6222 : /* clear waiting state again, nice for debugging */
6223 1813 : if (on_waitlist)
6224 1333 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
6225 : else
6226 : {
6227 480 : int extraWaits = 0;
6228 :
6229 :
6230 : /*
6231 : * Somebody else dequeued us and has or will wake us up. Deal with the
6232 : * superfluous absorption of a wakeup.
6233 : */
6234 :
6235 : /*
6236 : * Clear BM_LOCK_WAKE_IN_PROGRESS if somebody woke us before we
6237 : * removed ourselves - they'll have set it.
6238 : */
6239 480 : pg_atomic_fetch_and_u64(&buf_hdr->state, ~BM_LOCK_WAKE_IN_PROGRESS);
6240 :
6241 : /*
6242 : * Now wait for the scheduled wakeup, otherwise our ->lwWaiting would
6243 : * get reset at some inconvenient point later. Most of the time this
6244 : * will immediately return.
6245 : */
6246 : for (;;)
6247 : {
6248 480 : PGSemaphoreLock(MyProc->sem);
6249 480 : if (MyProc->lwWaiting == LW_WS_NOT_WAITING)
6250 480 : break;
6251 0 : extraWaits++;
6252 : }
6253 :
6254 : /*
6255 : * Fix the process wait semaphore's count for any absorbed wakeups.
6256 : */
6257 480 : while (extraWaits-- > 0)
6258 0 : PGSemaphoreUnlock(MyProc->sem);
6259 : }
6260 1813 : }
6261 :
6262 : /*
6263 : * Stop treating lock as held by current backend.
6264 : *
6265 : * After calling this function it's the callers responsibility to ensure that
6266 : * the lock gets released, even in case of an error. This only is desirable if
6267 : * the lock is going to be released in a different process than the process
6268 : * that acquired it.
6269 : */
6270 : static inline void
6271 0 : BufferLockDisown(Buffer buffer, BufferDesc *buf_hdr)
6272 : {
6273 0 : BufferLockDisownInternal(buffer, buf_hdr);
6274 0 : RESUME_INTERRUPTS();
6275 0 : }
6276 :
6277 : /*
6278 : * Stop treating lock as held by current backend.
6279 : *
6280 : * This is the code that can be shared between actually releasing a lock
6281 : * (BufferLockUnlock()) and just not tracking ownership of the lock anymore
6282 : * without releasing the lock (BufferLockDisown()).
6283 : */
6284 : static inline int
6285 113240231 : BufferLockDisownInternal(Buffer buffer, BufferDesc *buf_hdr)
6286 : {
6287 : BufferLockMode mode;
6288 : PrivateRefCountEntry *ref;
6289 :
6290 113240231 : ref = GetPrivateRefCountEntry(buffer, false);
6291 113240231 : if (ref == NULL)
6292 0 : elog(ERROR, "lock %d is not held", buffer);
6293 113240231 : mode = ref->data.lockmode;
6294 113240231 : ref->data.lockmode = BUFFER_LOCK_UNLOCK;
6295 :
6296 113240231 : return mode;
6297 : }
6298 :
6299 : /*
6300 : * Wakeup all the lockers that currently have a chance to acquire the lock.
6301 : *
6302 : * wake_exclusive indicates whether exclusive lock waiters should be woken up.
6303 : */
6304 : static void
6305 23158 : BufferLockWakeup(BufferDesc *buf_hdr, bool wake_exclusive)
6306 : {
6307 23158 : bool new_wake_in_progress = false;
6308 23158 : bool wake_share_exclusive = true;
6309 : proclist_head wakeup;
6310 : proclist_mutable_iter iter;
6311 :
6312 23158 : proclist_init(&wakeup);
6313 :
6314 : /* lock wait list while collecting backends to wake up */
6315 23158 : LockBufHdr(buf_hdr);
6316 :
6317 34564 : proclist_foreach_modify(iter, &buf_hdr->lock_waiters, lwWaitLink)
6318 : {
6319 24620 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
6320 :
6321 : /*
6322 : * Already woke up a conflicting lock, so skip over this wait list
6323 : * entry.
6324 : */
6325 24620 : if (!wake_exclusive && waiter->lwWaitMode == BUFFER_LOCK_EXCLUSIVE)
6326 747 : continue;
6327 23873 : if (!wake_share_exclusive && waiter->lwWaitMode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6328 0 : continue;
6329 :
6330 23873 : proclist_delete(&buf_hdr->lock_waiters, iter.cur, lwWaitLink);
6331 23873 : proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
6332 :
6333 : /*
6334 : * Prevent additional wakeups until retryer gets to run. Backends that
6335 : * are just waiting for the lock to become free don't retry
6336 : * automatically.
6337 : */
6338 23873 : new_wake_in_progress = true;
6339 :
6340 : /*
6341 : * Signal that the process isn't on the wait list anymore. This allows
6342 : * BufferLockDequeueSelf() to remove itself from the waitlist with a
6343 : * proclist_delete(), rather than having to check if it has been
6344 : * removed from the list.
6345 : */
6346 : Assert(waiter->lwWaiting == LW_WS_WAITING);
6347 23873 : waiter->lwWaiting = LW_WS_PENDING_WAKEUP;
6348 :
6349 : /*
6350 : * Don't wakeup further waiters after waking a conflicting waiter.
6351 : */
6352 23873 : if (waiter->lwWaitMode == BUFFER_LOCK_SHARE)
6353 : {
6354 : /*
6355 : * Share locks conflict with exclusive locks.
6356 : */
6357 10539 : wake_exclusive = false;
6358 : }
6359 13334 : else if (waiter->lwWaitMode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6360 : {
6361 : /*
6362 : * Share-exclusive locks conflict with share-exclusive and
6363 : * exclusive locks.
6364 : */
6365 120 : wake_exclusive = false;
6366 120 : wake_share_exclusive = false;
6367 : }
6368 13214 : else if (waiter->lwWaitMode == BUFFER_LOCK_EXCLUSIVE)
6369 : {
6370 : /*
6371 : * Exclusive locks conflict with all other locks, there's no point
6372 : * in waking up anybody else.
6373 : */
6374 13214 : break;
6375 : }
6376 : }
6377 :
6378 : Assert(proclist_is_empty(&wakeup) || pg_atomic_read_u64(&buf_hdr->state) & BM_LOCK_HAS_WAITERS);
6379 :
6380 : /* unset required flags, and release lock, in one fell swoop */
6381 : {
6382 : uint64 old_state;
6383 : uint64 desired_state;
6384 :
6385 23158 : old_state = pg_atomic_read_u64(&buf_hdr->state);
6386 : while (true)
6387 : {
6388 23191 : desired_state = old_state;
6389 :
6390 : /* compute desired flags */
6391 :
6392 23191 : if (new_wake_in_progress)
6393 22830 : desired_state |= BM_LOCK_WAKE_IN_PROGRESS;
6394 : else
6395 361 : desired_state &= ~BM_LOCK_WAKE_IN_PROGRESS;
6396 :
6397 23191 : if (proclist_is_empty(&buf_hdr->lock_waiters))
6398 20550 : desired_state &= ~BM_LOCK_HAS_WAITERS;
6399 :
6400 23191 : desired_state &= ~BM_LOCKED; /* release lock */
6401 :
6402 23191 : if (pg_atomic_compare_exchange_u64(&buf_hdr->state, &old_state,
6403 : desired_state))
6404 23158 : break;
6405 : }
6406 : }
6407 :
6408 : /* Awaken any waiters I removed from the queue. */
6409 47031 : proclist_foreach_modify(iter, &wakeup, lwWaitLink)
6410 : {
6411 23873 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
6412 :
6413 23873 : proclist_delete(&wakeup, iter.cur, lwWaitLink);
6414 :
6415 : /*
6416 : * Guarantee that lwWaiting being unset only becomes visible once the
6417 : * unlink from the link has completed. Otherwise the target backend
6418 : * could be woken up for other reason and enqueue for a new lock - if
6419 : * that happens before the list unlink happens, the list would end up
6420 : * being corrupted.
6421 : *
6422 : * The barrier pairs with the LockBufHdr() when enqueuing for another
6423 : * lock.
6424 : */
6425 23873 : pg_write_barrier();
6426 23873 : waiter->lwWaiting = LW_WS_NOT_WAITING;
6427 23873 : PGSemaphoreUnlock(waiter->sem);
6428 : }
6429 23158 : }
6430 :
6431 : /*
6432 : * Compute subtraction from buffer state for a release of a held lock in
6433 : * `mode`.
6434 : *
6435 : * This is separated from BufferLockUnlock() as we want to combine the lock
6436 : * release with other atomic operations when possible, leading to the lock
6437 : * release being done in multiple places, each needing to compute what to
6438 : * subtract from the lock state.
6439 : */
6440 : static inline uint64
6441 113240231 : BufferLockReleaseSub(BufferLockMode mode)
6442 : {
6443 : /*
6444 : * Turns out that a switch() leads gcc to generate sufficiently worse code
6445 : * for this to show up in profiles...
6446 : */
6447 113240231 : if (mode == BUFFER_LOCK_EXCLUSIVE)
6448 38180345 : return BM_LOCK_VAL_EXCLUSIVE;
6449 75059886 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6450 4094380 : return BM_LOCK_VAL_SHARE_EXCLUSIVE;
6451 : else
6452 : {
6453 : Assert(mode == BUFFER_LOCK_SHARE);
6454 70965506 : return BM_LOCK_VAL_SHARED;
6455 : }
6456 :
6457 : return 0; /* keep compiler quiet */
6458 : }
6459 :
6460 : /*
6461 : * Handle work that needs to be done after releasing a lock that was held in
6462 : * `mode`, where `lockstate` is the result of the atomic operation modifying
6463 : * the state variable.
6464 : *
6465 : * This is separated from BufferLockUnlock() as we want to combine the lock
6466 : * release with other atomic operations when possible, leading to the lock
6467 : * release being done in multiple places.
6468 : */
6469 : static void
6470 113240231 : BufferLockProcessRelease(BufferDesc *buf_hdr, BufferLockMode mode, uint64 lockstate)
6471 : {
6472 113240231 : bool check_waiters = false;
6473 113240231 : bool wake_exclusive = false;
6474 :
6475 : /* nobody else can have that kind of lock */
6476 : Assert(!(lockstate & BM_LOCK_VAL_EXCLUSIVE));
6477 :
6478 : /*
6479 : * If we're still waiting for backends to get scheduled, don't wake them
6480 : * up again. Otherwise check if we need to look through the waitqueue to
6481 : * wake other backends.
6482 : */
6483 113240231 : if ((lockstate & BM_LOCK_HAS_WAITERS) &&
6484 95448 : !(lockstate & BM_LOCK_WAKE_IN_PROGRESS))
6485 : {
6486 47744 : if ((lockstate & BM_LOCK_MASK) == 0)
6487 : {
6488 : /*
6489 : * We released a lock and the lock was, in that moment, free. We
6490 : * therefore can wake waiters for any kind of lock.
6491 : */
6492 23158 : check_waiters = true;
6493 23158 : wake_exclusive = true;
6494 : }
6495 24586 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6496 : {
6497 : /*
6498 : * We released the lock, but another backend still holds a lock.
6499 : * We can't have released an exclusive lock, as there couldn't
6500 : * have been other lock holders. If we released a share lock, no
6501 : * waiters need to be woken up, as there must be other share
6502 : * lockers. However, if we held a share-exclusive lock, another
6503 : * backend now could acquire a share-exclusive lock.
6504 : */
6505 0 : check_waiters = true;
6506 0 : wake_exclusive = false;
6507 : }
6508 : }
6509 :
6510 : /*
6511 : * As waking up waiters requires the spinlock to be acquired, only do so
6512 : * if necessary.
6513 : */
6514 113240231 : if (check_waiters)
6515 23158 : BufferLockWakeup(buf_hdr, wake_exclusive);
6516 113240231 : }
6517 :
6518 : /*
6519 : * BufferLockHeldByMeInMode - test whether my process holds the content lock
6520 : * in the specified mode
6521 : *
6522 : * This is meant as debug support only.
6523 : */
6524 : static bool
6525 0 : BufferLockHeldByMeInMode(BufferDesc *buf_hdr, BufferLockMode mode)
6526 : {
6527 : PrivateRefCountEntry *entry =
6528 0 : GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf_hdr), false);
6529 :
6530 0 : if (!entry)
6531 0 : return false;
6532 : else
6533 0 : return entry->data.lockmode == mode;
6534 : }
6535 :
6536 : /*
6537 : * BufferLockHeldByMe - test whether my process holds the content lock in any
6538 : * mode
6539 : *
6540 : * This is meant as debug support only.
6541 : */
6542 : static bool
6543 0 : BufferLockHeldByMe(BufferDesc *buf_hdr)
6544 : {
6545 : PrivateRefCountEntry *entry =
6546 0 : GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf_hdr), false);
6547 :
6548 0 : if (!entry)
6549 0 : return false;
6550 : else
6551 0 : return entry->data.lockmode != BUFFER_LOCK_UNLOCK;
6552 : }
6553 :
6554 : /*
6555 : * Release the content lock for the buffer.
6556 : */
6557 : void
6558 69504876 : UnlockBuffer(Buffer buffer)
6559 : {
6560 : BufferDesc *buf_hdr;
6561 :
6562 : Assert(BufferIsPinned(buffer));
6563 69504876 : if (BufferIsLocal(buffer))
6564 5199321 : return; /* local buffers need no lock */
6565 :
6566 64305555 : buf_hdr = GetBufferDescriptor(buffer - 1);
6567 64305555 : BufferLockUnlock(buffer, buf_hdr);
6568 : }
6569 :
6570 : /*
6571 : * Acquire the content_lock for the buffer.
6572 : */
6573 : void
6574 117291946 : LockBufferInternal(Buffer buffer, BufferLockMode mode)
6575 : {
6576 : BufferDesc *buf_hdr;
6577 :
6578 : /*
6579 : * We can't wait if we haven't got a PGPROC. This should only occur
6580 : * during bootstrap or shared memory initialization. Put an Assert here
6581 : * to catch unsafe coding practices.
6582 : */
6583 : Assert(!(MyProc == NULL && IsUnderPostmaster));
6584 :
6585 : /* handled in LockBuffer() wrapper */
6586 : Assert(mode != BUFFER_LOCK_UNLOCK);
6587 :
6588 : Assert(BufferIsPinned(buffer));
6589 117291946 : if (BufferIsLocal(buffer))
6590 6469174 : return; /* local buffers need no lock */
6591 :
6592 110822772 : buf_hdr = GetBufferDescriptor(buffer - 1);
6593 :
6594 : /*
6595 : * Test the most frequent lock modes first. While a switch (mode) would be
6596 : * nice, at least gcc generates considerably worse code for it.
6597 : *
6598 : * Call BufferLockAcquire() with a constant argument for mode, to generate
6599 : * more efficient code for the different lock modes.
6600 : */
6601 110822772 : if (mode == BUFFER_LOCK_SHARE)
6602 74333301 : BufferLockAcquire(buffer, buf_hdr, BUFFER_LOCK_SHARE);
6603 36489471 : else if (mode == BUFFER_LOCK_EXCLUSIVE)
6604 36489471 : BufferLockAcquire(buffer, buf_hdr, BUFFER_LOCK_EXCLUSIVE);
6605 0 : else if (mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6606 0 : BufferLockAcquire(buffer, buf_hdr, BUFFER_LOCK_SHARE_EXCLUSIVE);
6607 : else
6608 0 : elog(ERROR, "unrecognized buffer lock mode: %d", mode);
6609 : }
6610 :
6611 : /*
6612 : * Acquire the content_lock for the buffer, but only if we don't have to wait.
6613 : *
6614 : * This assumes the caller wants BUFFER_LOCK_EXCLUSIVE mode.
6615 : */
6616 : bool
6617 1779047 : ConditionalLockBuffer(Buffer buffer)
6618 : {
6619 : BufferDesc *buf;
6620 :
6621 : Assert(BufferIsPinned(buffer));
6622 1779047 : if (BufferIsLocal(buffer))
6623 86238 : return true; /* act as though we got it */
6624 :
6625 1692809 : buf = GetBufferDescriptor(buffer - 1);
6626 :
6627 1692809 : return BufferLockConditional(buffer, buf, BUFFER_LOCK_EXCLUSIVE);
6628 : }
6629 :
6630 : /*
6631 : * Verify that this backend is pinning the buffer exactly once.
6632 : *
6633 : * NOTE: Like in BufferIsPinned(), what we check here is that *this* backend
6634 : * holds a pin on the buffer. We do not care whether some other backend does.
6635 : */
6636 : void
6637 2712280 : CheckBufferIsPinnedOnce(Buffer buffer)
6638 : {
6639 2712280 : if (BufferIsLocal(buffer))
6640 : {
6641 1049 : if (LocalRefCount[-buffer - 1] != 1)
6642 0 : elog(ERROR, "incorrect local pin count: %d",
6643 : LocalRefCount[-buffer - 1]);
6644 : }
6645 : else
6646 : {
6647 2711231 : if (GetPrivateRefCount(buffer) != 1)
6648 0 : elog(ERROR, "incorrect local pin count: %d",
6649 : GetPrivateRefCount(buffer));
6650 : }
6651 2712280 : }
6652 :
6653 : /*
6654 : * LockBufferForCleanup - lock a buffer in preparation for deleting items
6655 : *
6656 : * Items may be deleted from a disk page only when the caller (a) holds an
6657 : * exclusive lock on the buffer and (b) has observed that no other backend
6658 : * holds a pin on the buffer. If there is a pin, then the other backend
6659 : * might have a pointer into the buffer (for example, a heapscan reference
6660 : * to an item --- see README for more details). It's OK if a pin is added
6661 : * after the cleanup starts, however; the newly-arrived backend will be
6662 : * unable to look at the page until we release the exclusive lock.
6663 : *
6664 : * To implement this protocol, a would-be deleter must pin the buffer and
6665 : * then call LockBufferForCleanup(). LockBufferForCleanup() is similar to
6666 : * LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE), except that it loops until
6667 : * it has successfully observed pin count = 1.
6668 : */
6669 : void
6670 26459 : LockBufferForCleanup(Buffer buffer)
6671 : {
6672 : BufferDesc *bufHdr;
6673 26459 : TimestampTz waitStart = 0;
6674 26459 : bool waiting = false;
6675 26459 : bool logged_recovery_conflict = false;
6676 :
6677 : Assert(BufferIsPinned(buffer));
6678 : Assert(PinCountWaitBuf == NULL);
6679 :
6680 26459 : CheckBufferIsPinnedOnce(buffer);
6681 :
6682 : /*
6683 : * We do not yet need to be worried about in-progress AIOs holding a pin,
6684 : * as we, so far, only support doing reads via AIO and this function can
6685 : * only be called once the buffer is valid (i.e. no read can be in
6686 : * flight).
6687 : */
6688 :
6689 : /* Nobody else to wait for */
6690 26459 : if (BufferIsLocal(buffer))
6691 18 : return;
6692 :
6693 26441 : bufHdr = GetBufferDescriptor(buffer - 1);
6694 :
6695 : for (;;)
6696 16 : {
6697 : uint64 buf_state;
6698 26457 : uint64 unset_bits = 0;
6699 :
6700 : /* Try to acquire lock */
6701 26457 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6702 26457 : buf_state = LockBufHdr(bufHdr);
6703 :
6704 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6705 26457 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
6706 : {
6707 : /* Successfully acquired exclusive lock with pincount 1 */
6708 26441 : UnlockBufHdr(bufHdr);
6709 :
6710 : /*
6711 : * Emit the log message if recovery conflict on buffer pin was
6712 : * resolved but the startup process waited longer than
6713 : * deadlock_timeout for it.
6714 : */
6715 26441 : if (logged_recovery_conflict)
6716 2 : LogRecoveryConflict(RECOVERY_CONFLICT_BUFFERPIN,
6717 : waitStart, GetCurrentTimestamp(),
6718 : NULL, false);
6719 :
6720 26441 : if (waiting)
6721 : {
6722 : /* reset ps display to remove the suffix if we added one */
6723 2 : set_ps_display_remove_suffix();
6724 2 : waiting = false;
6725 : }
6726 26441 : return;
6727 : }
6728 : /* Failed, so mark myself as waiting for pincount 1 */
6729 16 : if (buf_state & BM_PIN_COUNT_WAITER)
6730 : {
6731 0 : UnlockBufHdr(bufHdr);
6732 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6733 0 : elog(ERROR, "multiple backends attempting to wait for pincount 1");
6734 : }
6735 16 : bufHdr->wait_backend_pgprocno = MyProcNumber;
6736 16 : PinCountWaitBuf = bufHdr;
6737 16 : UnlockBufHdrExt(bufHdr, buf_state,
6738 : BM_PIN_COUNT_WAITER, 0,
6739 : 0);
6740 16 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6741 :
6742 : /* Wait to be signaled by UnpinBuffer() */
6743 16 : if (InHotStandby)
6744 : {
6745 8 : if (!waiting)
6746 : {
6747 : /* adjust the process title to indicate that it's waiting */
6748 2 : set_ps_display_suffix("waiting");
6749 2 : waiting = true;
6750 : }
6751 :
6752 : /*
6753 : * Emit the log message if the startup process is waiting longer
6754 : * than deadlock_timeout for recovery conflict on buffer pin.
6755 : *
6756 : * Skip this if first time through because the startup process has
6757 : * not started waiting yet in this case. So, the wait start
6758 : * timestamp is set after this logic.
6759 : */
6760 8 : if (waitStart != 0 && !logged_recovery_conflict)
6761 : {
6762 3 : TimestampTz now = GetCurrentTimestamp();
6763 :
6764 3 : if (TimestampDifferenceExceeds(waitStart, now,
6765 : DeadlockTimeout))
6766 : {
6767 2 : LogRecoveryConflict(RECOVERY_CONFLICT_BUFFERPIN,
6768 : waitStart, now, NULL, true);
6769 2 : logged_recovery_conflict = true;
6770 : }
6771 : }
6772 :
6773 : /*
6774 : * Set the wait start timestamp if logging is enabled and first
6775 : * time through.
6776 : */
6777 8 : if (log_recovery_conflict_waits && waitStart == 0)
6778 2 : waitStart = GetCurrentTimestamp();
6779 :
6780 : /* Publish the bufid that Startup process waits on */
6781 8 : SetStartupBufferPinWaitBufId(buffer - 1);
6782 : /* Set alarm and then wait to be signaled by UnpinBuffer() */
6783 8 : ResolveRecoveryConflictWithBufferPin();
6784 : /* Reset the published bufid */
6785 8 : SetStartupBufferPinWaitBufId(-1);
6786 : }
6787 : else
6788 8 : ProcWaitForSignal(WAIT_EVENT_BUFFER_CLEANUP);
6789 :
6790 : /*
6791 : * Remove flag marking us as waiter. Normally this will not be set
6792 : * anymore, but ProcWaitForSignal() can return for other signals as
6793 : * well. We take care to only reset the flag if we're the waiter, as
6794 : * theoretically another backend could have started waiting. That's
6795 : * impossible with the current usages due to table level locking, but
6796 : * better be safe.
6797 : */
6798 16 : buf_state = LockBufHdr(bufHdr);
6799 16 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
6800 6 : bufHdr->wait_backend_pgprocno == MyProcNumber)
6801 6 : unset_bits |= BM_PIN_COUNT_WAITER;
6802 :
6803 16 : UnlockBufHdrExt(bufHdr, buf_state,
6804 : 0, unset_bits,
6805 : 0);
6806 :
6807 16 : PinCountWaitBuf = NULL;
6808 : /* Loop back and try again */
6809 : }
6810 : }
6811 :
6812 : /*
6813 : * Check called from ProcessRecoveryConflictInterrupts() when Startup process
6814 : * requests cancellation of all pin holders that are blocking it.
6815 : */
6816 : bool
6817 3 : HoldingBufferPinThatDelaysRecovery(void)
6818 : {
6819 3 : int bufid = GetStartupBufferPinWaitBufId();
6820 :
6821 : /*
6822 : * If we get woken slowly then it's possible that the Startup process was
6823 : * already woken by other backends before we got here. Also possible that
6824 : * we get here by multiple interrupts or interrupts at inappropriate
6825 : * times, so make sure we do nothing if the bufid is not set.
6826 : */
6827 3 : if (bufid < 0)
6828 1 : return false;
6829 :
6830 2 : if (GetPrivateRefCount(bufid + 1) > 0)
6831 2 : return true;
6832 :
6833 0 : return false;
6834 : }
6835 :
6836 : /*
6837 : * ConditionalLockBufferForCleanup - as above, but don't wait to get the lock
6838 : *
6839 : * We won't loop, but just check once to see if the pin count is OK. If
6840 : * not, return false with no lock held.
6841 : */
6842 : bool
6843 536337 : ConditionalLockBufferForCleanup(Buffer buffer)
6844 : {
6845 : BufferDesc *bufHdr;
6846 : uint64 buf_state,
6847 : refcount;
6848 :
6849 : Assert(BufferIsValid(buffer));
6850 :
6851 : /* see AIO related comment in LockBufferForCleanup() */
6852 :
6853 536337 : if (BufferIsLocal(buffer))
6854 : {
6855 12371 : refcount = LocalRefCount[-buffer - 1];
6856 : /* There should be exactly one pin */
6857 : Assert(refcount > 0);
6858 12371 : if (refcount != 1)
6859 1540 : return false;
6860 : /* Nobody else to wait for */
6861 10831 : return true;
6862 : }
6863 :
6864 : /* There should be exactly one local pin */
6865 523966 : refcount = GetPrivateRefCount(buffer);
6866 : Assert(refcount);
6867 523966 : if (refcount != 1)
6868 307 : return false;
6869 :
6870 : /* Try to acquire lock */
6871 523659 : if (!ConditionalLockBuffer(buffer))
6872 68 : return false;
6873 :
6874 523591 : bufHdr = GetBufferDescriptor(buffer - 1);
6875 523591 : buf_state = LockBufHdr(bufHdr);
6876 523591 : refcount = BUF_STATE_GET_REFCOUNT(buf_state);
6877 :
6878 : Assert(refcount > 0);
6879 523591 : if (refcount == 1)
6880 : {
6881 : /* Successfully acquired exclusive lock with pincount 1 */
6882 523419 : UnlockBufHdr(bufHdr);
6883 523419 : return true;
6884 : }
6885 :
6886 : /* Failed, so release the lock */
6887 172 : UnlockBufHdr(bufHdr);
6888 172 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6889 172 : return false;
6890 : }
6891 :
6892 : /*
6893 : * IsBufferCleanupOK - as above, but we already have the lock
6894 : *
6895 : * Check whether it's OK to perform cleanup on a buffer we've already
6896 : * locked. If we observe that the pin count is 1, our exclusive lock
6897 : * happens to be a cleanup lock, and we can proceed with anything that
6898 : * would have been allowable had we sought a cleanup lock originally.
6899 : */
6900 : bool
6901 2696 : IsBufferCleanupOK(Buffer buffer)
6902 : {
6903 : BufferDesc *bufHdr;
6904 : uint64 buf_state;
6905 :
6906 : Assert(BufferIsValid(buffer));
6907 :
6908 : /* see AIO related comment in LockBufferForCleanup() */
6909 :
6910 2696 : if (BufferIsLocal(buffer))
6911 : {
6912 : /* There should be exactly one pin */
6913 0 : if (LocalRefCount[-buffer - 1] != 1)
6914 0 : return false;
6915 : /* Nobody else to wait for */
6916 0 : return true;
6917 : }
6918 :
6919 : /* There should be exactly one local pin */
6920 2696 : if (GetPrivateRefCount(buffer) != 1)
6921 0 : return false;
6922 :
6923 2696 : bufHdr = GetBufferDescriptor(buffer - 1);
6924 :
6925 : /* caller must hold exclusive lock on buffer */
6926 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
6927 :
6928 2696 : buf_state = LockBufHdr(bufHdr);
6929 :
6930 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6931 2696 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
6932 : {
6933 : /* pincount is OK. */
6934 2696 : UnlockBufHdr(bufHdr);
6935 2696 : return true;
6936 : }
6937 :
6938 0 : UnlockBufHdr(bufHdr);
6939 0 : return false;
6940 : }
6941 :
6942 : /*
6943 : * Helper for BufferBeginSetHintBits() and BufferSetHintBits16().
6944 : *
6945 : * This checks if the current lock mode already suffices to allow hint bits
6946 : * being set and, if not, whether the current lock can be upgraded.
6947 : *
6948 : * Updates *lockstate when returning true.
6949 : */
6950 : static inline bool
6951 15004529 : SharedBufferBeginSetHintBits(Buffer buffer, BufferDesc *buf_hdr, uint64 *lockstate)
6952 : {
6953 : uint64 old_state;
6954 : PrivateRefCountEntry *ref;
6955 : BufferLockMode mode;
6956 :
6957 15004529 : ref = GetPrivateRefCountEntry(buffer, true);
6958 :
6959 15004529 : if (ref == NULL)
6960 0 : elog(ERROR, "buffer is not pinned");
6961 :
6962 15004529 : mode = ref->data.lockmode;
6963 15004529 : if (mode == BUFFER_LOCK_UNLOCK)
6964 0 : elog(ERROR, "buffer is not locked");
6965 :
6966 : /* we're done if we are already holding a sufficient lock level */
6967 15004529 : if (mode == BUFFER_LOCK_EXCLUSIVE || mode == BUFFER_LOCK_SHARE_EXCLUSIVE)
6968 : {
6969 11636587 : *lockstate = pg_atomic_read_u64(&buf_hdr->state);
6970 11636587 : return true;
6971 : }
6972 :
6973 : /*
6974 : * We are only holding a share lock right now, try to upgrade it to
6975 : * SHARE_EXCLUSIVE.
6976 : */
6977 : Assert(mode == BUFFER_LOCK_SHARE);
6978 :
6979 3367942 : old_state = pg_atomic_read_u64(&buf_hdr->state);
6980 : while (true)
6981 4 : {
6982 : uint64 desired_state;
6983 :
6984 3367946 : desired_state = old_state;
6985 :
6986 : /*
6987 : * Can't upgrade if somebody else holds the lock in exclusive or
6988 : * share-exclusive mode.
6989 : */
6990 3367946 : if (unlikely((old_state & (BM_LOCK_VAL_EXCLUSIVE | BM_LOCK_VAL_SHARE_EXCLUSIVE)) != 0))
6991 : {
6992 136 : return false;
6993 : }
6994 :
6995 : /* currently held lock state */
6996 3367810 : desired_state -= BM_LOCK_VAL_SHARED;
6997 :
6998 : /* new lock level */
6999 3367810 : desired_state += BM_LOCK_VAL_SHARE_EXCLUSIVE;
7000 :
7001 3367810 : if (likely(pg_atomic_compare_exchange_u64(&buf_hdr->state,
7002 : &old_state, desired_state)))
7003 : {
7004 3367806 : ref->data.lockmode = BUFFER_LOCK_SHARE_EXCLUSIVE;
7005 3367806 : *lockstate = desired_state;
7006 :
7007 3367806 : return true;
7008 : }
7009 : }
7010 : }
7011 :
7012 : /*
7013 : * Try to acquire the right to set hint bits on the buffer.
7014 : *
7015 : * To be allowed to set hint bits, this backend needs to hold either a
7016 : * share-exclusive or an exclusive lock. In case this backend only holds a
7017 : * share lock, this function will try to upgrade the lock to
7018 : * share-exclusive. The caller is only allowed to set hint bits if true is
7019 : * returned.
7020 : *
7021 : * Once BufferBeginSetHintBits() has returned true, hint bits may be set
7022 : * without further calls to BufferBeginSetHintBits(), until the buffer is
7023 : * unlocked.
7024 : *
7025 : *
7026 : * Requiring a share-exclusive lock to set hint bits prevents setting hint
7027 : * bits on buffers that are currently being written out, which could corrupt
7028 : * the checksum on the page. Flushing buffers also requires a share-exclusive
7029 : * lock.
7030 : *
7031 : * Due to a lock >= share-exclusive being required to set hint bits, only one
7032 : * backend can set hint bits at a time. Allowing multiple backends to set hint
7033 : * bits would require more complicated locking: For setting hint bits we'd
7034 : * need to store the count of backends currently setting hint bits, for I/O we
7035 : * would need another lock-level conflicting with the hint-setting
7036 : * lock-level. Given that the share-exclusive lock for setting hint bits is
7037 : * only held for a short time, that backends often would just set the same
7038 : * hint bits and that the cost of occasionally not setting hint bits in hotly
7039 : * accessed pages is fairly low, this seems like an acceptable tradeoff.
7040 : */
7041 : bool
7042 411411 : BufferBeginSetHintBits(Buffer buffer)
7043 : {
7044 : BufferDesc *buf_hdr;
7045 : uint64 lockstate;
7046 :
7047 411411 : if (BufferIsLocal(buffer))
7048 : {
7049 : /*
7050 : * NB: Will need to check if there is a write in progress, once it is
7051 : * possible for writes to be done asynchronously.
7052 : */
7053 2378 : return true;
7054 : }
7055 :
7056 409033 : buf_hdr = GetBufferDescriptor(buffer - 1);
7057 :
7058 409033 : return SharedBufferBeginSetHintBits(buffer, buf_hdr, &lockstate);
7059 : }
7060 :
7061 : /*
7062 : * End a phase of setting hint bits on this buffer, started with
7063 : * BufferBeginSetHintBits().
7064 : *
7065 : * This would strictly speaking not be required (i.e. the caller could do
7066 : * MarkBufferDirtyHint() if so desired), but allows us to perform some sanity
7067 : * checks.
7068 : */
7069 : void
7070 411391 : BufferFinishSetHintBits(Buffer buffer, bool mark_dirty, bool buffer_std)
7071 : {
7072 : if (!BufferIsLocal(buffer))
7073 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_SHARE_EXCLUSIVE) ||
7074 : BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
7075 :
7076 411391 : if (mark_dirty)
7077 233328 : MarkBufferDirtyHint(buffer, buffer_std);
7078 411391 : }
7079 :
7080 : /*
7081 : * Try to set hint bits on a single 16bit value in a buffer.
7082 : *
7083 : * If hint bits are allowed to be set, set *ptr = val, try to mark the buffer
7084 : * dirty and return true. Otherwise false is returned.
7085 : *
7086 : * *ptr needs to be a pointer to memory within the buffer.
7087 : *
7088 : * This is a bit faster than BufferBeginSetHintBits() /
7089 : * BufferFinishSetHintBits() when setting hints once in a buffer, but slower
7090 : * than the former when setting hint bits multiple times in the same buffer.
7091 : */
7092 : bool
7093 15398564 : BufferSetHintBits16(uint16 *ptr, uint16 val, Buffer buffer)
7094 : {
7095 : BufferDesc *buf_hdr;
7096 : uint64 lockstate;
7097 : #ifdef USE_ASSERT_CHECKING
7098 : char *page;
7099 :
7100 : /* verify that the address is on the page */
7101 : page = BufferGetPage(buffer);
7102 : Assert((char *) ptr >= page && (char *) ptr < (page + BLCKSZ));
7103 : #endif
7104 :
7105 15398564 : if (BufferIsLocal(buffer))
7106 : {
7107 803068 : *ptr = val;
7108 :
7109 803068 : MarkLocalBufferDirty(buffer);
7110 :
7111 803068 : return true;
7112 : }
7113 :
7114 14595496 : buf_hdr = GetBufferDescriptor(buffer - 1);
7115 :
7116 14595496 : if (SharedBufferBeginSetHintBits(buffer, buf_hdr, &lockstate))
7117 : {
7118 14595380 : *ptr = val;
7119 :
7120 14595380 : MarkSharedBufferDirtyHint(buffer, buf_hdr, lockstate, true);
7121 :
7122 14595380 : return true;
7123 : }
7124 :
7125 116 : return false;
7126 : }
7127 :
7128 :
7129 : /*
7130 : * Functions for buffer I/O handling
7131 : *
7132 : * Also note that these are used only for shared buffers, not local ones.
7133 : */
7134 :
7135 : /*
7136 : * WaitIO -- Block until the IO_IN_PROGRESS flag on 'buf' is cleared.
7137 : */
7138 : static void
7139 21151 : WaitIO(BufferDesc *buf)
7140 : {
7141 21151 : ConditionVariable *cv = BufferDescriptorGetIOCV(buf);
7142 :
7143 : /*
7144 : * Should never end up here with unsubmitted IO, as no AIO unaware code
7145 : * may be used while in batch mode and AIO aware code needs to have
7146 : * submitted all staged IO to avoid deadlocks & slowness.
7147 : */
7148 : Assert(!pgaio_have_staged());
7149 :
7150 21151 : ConditionVariablePrepareToSleep(cv);
7151 : for (;;)
7152 385 : {
7153 : uint64 buf_state;
7154 : PgAioWaitRef iow;
7155 :
7156 : /*
7157 : * It may not be necessary to acquire the spinlock to check the flag
7158 : * here, but since this test is essential for correctness, we'd better
7159 : * play it safe.
7160 : */
7161 21536 : buf_state = LockBufHdr(buf);
7162 :
7163 : /*
7164 : * Copy the wait reference while holding the spinlock. This protects
7165 : * against a concurrent TerminateBufferIO() in another backend from
7166 : * clearing the wref while it's being read.
7167 : */
7168 21536 : iow = buf->io_wref;
7169 21536 : UnlockBufHdr(buf);
7170 :
7171 : /* no IO in progress, we don't need to wait */
7172 21536 : if (!(buf_state & BM_IO_IN_PROGRESS))
7173 21151 : break;
7174 :
7175 : /*
7176 : * The buffer has asynchronous IO in progress, wait for it to
7177 : * complete.
7178 : */
7179 385 : if (pgaio_wref_valid(&iow))
7180 : {
7181 73 : pgaio_wref_wait(&iow);
7182 :
7183 : /*
7184 : * The AIO subsystem internally uses condition variables and thus
7185 : * might remove this backend from the BufferDesc's CV. While that
7186 : * wouldn't cause a correctness issue (the first CV sleep just
7187 : * immediately returns if not already registered), it seems worth
7188 : * avoiding unnecessary loop iterations, given that we take care
7189 : * to do so at the start of the function.
7190 : */
7191 73 : ConditionVariablePrepareToSleep(cv);
7192 73 : continue;
7193 : }
7194 :
7195 : /* wait on BufferDesc->cv, e.g. for concurrent synchronous IO */
7196 312 : ConditionVariableSleep(cv, WAIT_EVENT_BUFFER_IO);
7197 : }
7198 21151 : ConditionVariableCancelSleep();
7199 21151 : }
7200 :
7201 : /*
7202 : * StartSharedBufferIO: begin I/O on this buffer
7203 : * (Assumptions)
7204 : * The buffer is Pinned
7205 : *
7206 : * In several scenarios the buffer may already be undergoing I/O in this or
7207 : * another backend. How to best handle that depends on the caller's
7208 : * situation. It might be appropriate to wait synchronously (e.g., because the
7209 : * buffer is about to be invalidated); wait asynchronously, using the buffer's
7210 : * IO wait reference (e.g., because the caller is doing readahead and doesn't
7211 : * need the buffer to be ready immediately); or to not wait at all (e.g.,
7212 : * because the caller is trying to combine IO for this buffer with another
7213 : * buffer).
7214 : *
7215 : * How and whether to wait is controlled by the wait and io_wref
7216 : * parameters. In detail:
7217 : *
7218 : * - If the caller passes a non-NULL io_wref and the buffer has an I/O wait
7219 : * reference, the *io_wref is set to the buffer's io_wref and
7220 : * BUFFER_IO_IN_PROGRESS is returned. This is done regardless of the wait
7221 : * parameter.
7222 : *
7223 : * - If the caller passes a NULL io_wref (i.e. the caller does not want to
7224 : * asynchronously wait for the completion of the IO), wait = false and the
7225 : * buffer is undergoing IO, BUFFER_IO_IN_PROGRESS is returned.
7226 : *
7227 : * - If wait = true and either the buffer does not have a wait reference,
7228 : * or the caller passes io_wref = NULL, WaitIO() is used to wait for the IO
7229 : * to complete. To avoid the potential of deadlocks and unnecessary delays,
7230 : * all staged I/O is submitted before waiting.
7231 : *
7232 : * Input operations are only attempted on buffers that are not BM_VALID, and
7233 : * output operations only on buffers that are BM_VALID and BM_DIRTY, so we can
7234 : * always tell if the work is already done. If no I/O is necessary,
7235 : * BUFFER_IO_ALREADY_DONE is returned.
7236 : *
7237 : * If we successfully marked the buffer as BM_IO_IN_PROGRESS,
7238 : * BUFFER_IO_READY_FOR_IO is returned.
7239 : */
7240 : StartBufferIOResult
7241 2967783 : StartSharedBufferIO(BufferDesc *buf, bool forInput, bool wait, PgAioWaitRef *io_wref)
7242 : {
7243 : uint64 buf_state;
7244 :
7245 2967783 : ResourceOwnerEnlarge(CurrentResourceOwner);
7246 :
7247 : for (;;)
7248 : {
7249 2968163 : buf_state = LockBufHdr(buf);
7250 :
7251 2968163 : if (!(buf_state & BM_IO_IN_PROGRESS))
7252 2964459 : break;
7253 :
7254 : /* Join the existing IO */
7255 3704 : if (io_wref != NULL && pgaio_wref_valid(&buf->io_wref))
7256 : {
7257 3316 : *io_wref = buf->io_wref;
7258 3316 : UnlockBufHdr(buf);
7259 :
7260 3316 : return BUFFER_IO_IN_PROGRESS;
7261 : }
7262 388 : else if (!wait)
7263 : {
7264 8 : UnlockBufHdr(buf);
7265 8 : return BUFFER_IO_IN_PROGRESS;
7266 : }
7267 : else
7268 : {
7269 : /*
7270 : * With wait = true, we always have to wait if the caller has
7271 : * passed io_wref = NULL.
7272 : *
7273 : * Even with io_wref != NULL, we have to wait if the buffer's wait
7274 : * ref is not valid but the IO is in progress, someone else
7275 : * started IO but hasn't set the wait ref yet. We have no choice
7276 : * but to wait until the IO completes.
7277 : */
7278 380 : UnlockBufHdr(buf);
7279 :
7280 : /*
7281 : * If this backend currently has staged IO, submit it before
7282 : * waiting for in-progress IO, to avoid potential deadlocks and
7283 : * unnecessary delays.
7284 : */
7285 380 : pgaio_submit_staged();
7286 :
7287 380 : WaitIO(buf);
7288 : }
7289 : }
7290 :
7291 : /* Once we get here, there is definitely no I/O active on this buffer */
7292 :
7293 : /* Check if someone else already did the I/O */
7294 2964459 : if (forInput ? (buf_state & BM_VALID) : !(buf_state & BM_DIRTY))
7295 : {
7296 571 : UnlockBufHdr(buf);
7297 571 : return BUFFER_IO_ALREADY_DONE;
7298 : }
7299 :
7300 : /*
7301 : * No IO in progress and not already done; we will start IO. It's possible
7302 : * that the IO was in progress but we're not done, because the IO errored
7303 : * out. We'll do the IO ourselves.
7304 : */
7305 2963888 : UnlockBufHdrExt(buf, buf_state,
7306 : BM_IO_IN_PROGRESS, 0,
7307 : 0);
7308 :
7309 2963888 : ResourceOwnerRememberBufferIO(CurrentResourceOwner,
7310 : BufferDescriptorGetBuffer(buf));
7311 :
7312 2963888 : return BUFFER_IO_READY_FOR_IO;
7313 : }
7314 :
7315 : /*
7316 : * Wrapper around StartSharedBufferIO / StartLocalBufferIO. Only to be used
7317 : * when the caller doesn't otherwise need to care about local vs shared. See
7318 : * StartSharedBufferIO() for details.
7319 : */
7320 : StartBufferIOResult
7321 1669786 : StartBufferIO(Buffer buffer, bool forInput, bool wait, PgAioWaitRef *io_wref)
7322 : {
7323 : BufferDesc *buf_hdr;
7324 :
7325 1669786 : if (BufferIsLocal(buffer))
7326 : {
7327 11006 : buf_hdr = GetLocalBufferDescriptor(-buffer - 1);
7328 :
7329 11006 : return StartLocalBufferIO(buf_hdr, forInput, wait, io_wref);
7330 : }
7331 : else
7332 : {
7333 1658780 : buf_hdr = GetBufferDescriptor(buffer - 1);
7334 :
7335 1658780 : return StartSharedBufferIO(buf_hdr, forInput, wait, io_wref);
7336 : }
7337 : }
7338 :
7339 : /*
7340 : * TerminateBufferIO: release a buffer we were doing I/O on
7341 : * (Assumptions)
7342 : * My process is executing IO for the buffer
7343 : * BM_IO_IN_PROGRESS bit is set for the buffer
7344 : * The buffer is Pinned
7345 : *
7346 : * If clear_dirty is true, we clear the buffer's BM_DIRTY flag. This is
7347 : * appropriate when terminating a successful write.
7348 : *
7349 : * set_flag_bits gets ORed into the buffer's flags. It must include
7350 : * BM_IO_ERROR in a failure case. For successful completion it could
7351 : * be 0, or BM_VALID if we just finished reading in the page.
7352 : *
7353 : * If forget_owner is true, we release the buffer I/O from the current
7354 : * resource owner. (forget_owner=false is used when the resource owner itself
7355 : * is being released)
7356 : */
7357 : void
7358 2791267 : TerminateBufferIO(BufferDesc *buf, bool clear_dirty, uint64 set_flag_bits,
7359 : bool forget_owner, bool release_aio)
7360 : {
7361 : uint64 buf_state;
7362 2791267 : uint64 unset_flag_bits = 0;
7363 2791267 : int refcount_change = 0;
7364 :
7365 2791267 : buf_state = LockBufHdr(buf);
7366 :
7367 : Assert(buf_state & BM_IO_IN_PROGRESS);
7368 2791267 : unset_flag_bits |= BM_IO_IN_PROGRESS;
7369 :
7370 : /* Clear earlier errors, if this IO failed, it'll be marked again */
7371 2791267 : unset_flag_bits |= BM_IO_ERROR;
7372 :
7373 2791267 : if (clear_dirty)
7374 704262 : unset_flag_bits |= BM_DIRTY | BM_CHECKPOINT_NEEDED;
7375 :
7376 2791267 : if (release_aio)
7377 : {
7378 : /* release ownership by the AIO subsystem */
7379 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
7380 1482378 : refcount_change = -1;
7381 1482378 : pgaio_wref_clear(&buf->io_wref);
7382 : }
7383 :
7384 2791267 : buf_state = UnlockBufHdrExt(buf, buf_state,
7385 : set_flag_bits, unset_flag_bits,
7386 : refcount_change);
7387 :
7388 2791267 : if (forget_owner)
7389 1308866 : ResourceOwnerForgetBufferIO(CurrentResourceOwner,
7390 : BufferDescriptorGetBuffer(buf));
7391 :
7392 2791267 : ConditionVariableBroadcast(BufferDescriptorGetIOCV(buf));
7393 :
7394 : /*
7395 : * Support LockBufferForCleanup()
7396 : *
7397 : * We may have just released the last pin other than the waiter's. In most
7398 : * cases, this backend holds another pin on the buffer. But, if, for
7399 : * example, this backend is completing an IO issued by another backend, it
7400 : * may be time to wake the waiter.
7401 : */
7402 2791267 : if (release_aio && (buf_state & BM_PIN_COUNT_WAITER))
7403 0 : WakePinCountWaiter(buf);
7404 2791267 : }
7405 :
7406 : /*
7407 : * AbortBufferIO: Clean up active buffer I/O after an error.
7408 : *
7409 : * All LWLocks & content locks we might have held have been released, but we
7410 : * haven't yet released buffer pins, so the buffer is still pinned.
7411 : *
7412 : * If I/O was in progress, we always set BM_IO_ERROR, even though it's
7413 : * possible the error condition wasn't related to the I/O.
7414 : *
7415 : * Note: this does not remove the buffer I/O from the resource owner.
7416 : * That's correct when we're releasing the whole resource owner, but
7417 : * beware if you use this in other contexts.
7418 : */
7419 : static void
7420 15 : AbortBufferIO(Buffer buffer)
7421 : {
7422 15 : BufferDesc *buf_hdr = GetBufferDescriptor(buffer - 1);
7423 : uint64 buf_state;
7424 :
7425 15 : buf_state = LockBufHdr(buf_hdr);
7426 : Assert(buf_state & (BM_IO_IN_PROGRESS | BM_TAG_VALID));
7427 :
7428 15 : if (!(buf_state & BM_VALID))
7429 : {
7430 : Assert(!(buf_state & BM_DIRTY));
7431 15 : UnlockBufHdr(buf_hdr);
7432 : }
7433 : else
7434 : {
7435 : Assert(buf_state & BM_DIRTY);
7436 0 : UnlockBufHdr(buf_hdr);
7437 :
7438 : /* Issue notice if this is not the first failure... */
7439 0 : if (buf_state & BM_IO_ERROR)
7440 : {
7441 : /* Buffer is pinned, so we can read tag without spinlock */
7442 0 : ereport(WARNING,
7443 : (errcode(ERRCODE_IO_ERROR),
7444 : errmsg("could not write block %u of %s",
7445 : buf_hdr->tag.blockNum,
7446 : relpathperm(BufTagGetRelFileLocator(&buf_hdr->tag),
7447 : BufTagGetForkNum(&buf_hdr->tag)).str),
7448 : errdetail("Multiple failures --- write error might be permanent.")));
7449 : }
7450 : }
7451 :
7452 15 : TerminateBufferIO(buf_hdr, false, BM_IO_ERROR, false, false);
7453 15 : }
7454 :
7455 : /*
7456 : * Error context callback for errors occurring during shared buffer writes.
7457 : */
7458 : static void
7459 41 : shared_buffer_write_error_callback(void *arg)
7460 : {
7461 41 : BufferDesc *bufHdr = (BufferDesc *) arg;
7462 :
7463 : /* Buffer is pinned, so we can read the tag without locking the spinlock */
7464 41 : if (bufHdr != NULL)
7465 82 : errcontext("writing block %u of relation \"%s\"",
7466 : bufHdr->tag.blockNum,
7467 41 : relpathperm(BufTagGetRelFileLocator(&bufHdr->tag),
7468 : BufTagGetForkNum(&bufHdr->tag)).str);
7469 41 : }
7470 :
7471 : /*
7472 : * Error context callback for errors occurring during local buffer writes.
7473 : */
7474 : static void
7475 0 : local_buffer_write_error_callback(void *arg)
7476 : {
7477 0 : BufferDesc *bufHdr = (BufferDesc *) arg;
7478 :
7479 0 : if (bufHdr != NULL)
7480 0 : errcontext("writing block %u of relation \"%s\"",
7481 : bufHdr->tag.blockNum,
7482 0 : relpathbackend(BufTagGetRelFileLocator(&bufHdr->tag),
7483 : MyProcNumber,
7484 : BufTagGetForkNum(&bufHdr->tag)).str);
7485 0 : }
7486 :
7487 : /*
7488 : * RelFileLocator qsort/bsearch comparator; see RelFileLocatorEquals.
7489 : */
7490 : static int
7491 13252787 : rlocator_comparator(const void *p1, const void *p2)
7492 : {
7493 13252787 : RelFileLocator n1 = *(const RelFileLocator *) p1;
7494 13252787 : RelFileLocator n2 = *(const RelFileLocator *) p2;
7495 :
7496 13252787 : if (n1.relNumber < n2.relNumber)
7497 13198394 : return -1;
7498 54393 : else if (n1.relNumber > n2.relNumber)
7499 52229 : return 1;
7500 :
7501 2164 : if (n1.dbOid < n2.dbOid)
7502 0 : return -1;
7503 2164 : else if (n1.dbOid > n2.dbOid)
7504 0 : return 1;
7505 :
7506 2164 : if (n1.spcOid < n2.spcOid)
7507 0 : return -1;
7508 2164 : else if (n1.spcOid > n2.spcOid)
7509 0 : return 1;
7510 : else
7511 2164 : return 0;
7512 : }
7513 :
7514 : /*
7515 : * Lock buffer header - set BM_LOCKED in buffer state.
7516 : */
7517 : uint64
7518 28795865 : LockBufHdr(BufferDesc *desc)
7519 : {
7520 : uint64 old_buf_state;
7521 :
7522 : Assert(!BufferIsLocal(BufferDescriptorGetBuffer(desc)));
7523 :
7524 : while (true)
7525 : {
7526 : /*
7527 : * Always try once to acquire the lock directly, without setting up
7528 : * the spin-delay infrastructure. The work necessary for that shows up
7529 : * in profiles and is rarely necessary.
7530 : */
7531 28796806 : old_buf_state = pg_atomic_fetch_or_u64(&desc->state, BM_LOCKED);
7532 28796806 : if (likely(!(old_buf_state & BM_LOCKED)))
7533 28795865 : break; /* got lock */
7534 :
7535 : /* and then spin without atomic operations until lock is released */
7536 : {
7537 : SpinDelayStatus delayStatus;
7538 :
7539 941 : init_local_spin_delay(&delayStatus);
7540 :
7541 8189 : while (old_buf_state & BM_LOCKED)
7542 : {
7543 7248 : perform_spin_delay(&delayStatus);
7544 7248 : old_buf_state = pg_atomic_read_u64(&desc->state);
7545 : }
7546 941 : finish_spin_delay(&delayStatus);
7547 : }
7548 :
7549 : /*
7550 : * Retry. The lock might obviously already be re-acquired by the time
7551 : * we're attempting to get it again.
7552 : */
7553 : }
7554 :
7555 28795865 : return old_buf_state | BM_LOCKED;
7556 : }
7557 :
7558 : /*
7559 : * Wait until the BM_LOCKED flag isn't set anymore and return the buffer's
7560 : * state at that point.
7561 : *
7562 : * Obviously the buffer could be locked by the time the value is returned, so
7563 : * this is primarily useful in CAS style loops.
7564 : */
7565 : pg_noinline uint64
7566 831 : WaitBufHdrUnlocked(BufferDesc *buf)
7567 : {
7568 : SpinDelayStatus delayStatus;
7569 : uint64 buf_state;
7570 :
7571 831 : init_local_spin_delay(&delayStatus);
7572 :
7573 831 : buf_state = pg_atomic_read_u64(&buf->state);
7574 :
7575 12969 : while (buf_state & BM_LOCKED)
7576 : {
7577 12138 : perform_spin_delay(&delayStatus);
7578 12138 : buf_state = pg_atomic_read_u64(&buf->state);
7579 : }
7580 :
7581 831 : finish_spin_delay(&delayStatus);
7582 :
7583 831 : return buf_state;
7584 : }
7585 :
7586 : /*
7587 : * BufferTag comparator.
7588 : */
7589 : static inline int
7590 0 : buffertag_comparator(const BufferTag *ba, const BufferTag *bb)
7591 : {
7592 : int ret;
7593 : RelFileLocator rlocatora;
7594 : RelFileLocator rlocatorb;
7595 :
7596 0 : rlocatora = BufTagGetRelFileLocator(ba);
7597 0 : rlocatorb = BufTagGetRelFileLocator(bb);
7598 :
7599 0 : ret = rlocator_comparator(&rlocatora, &rlocatorb);
7600 :
7601 0 : if (ret != 0)
7602 0 : return ret;
7603 :
7604 0 : if (BufTagGetForkNum(ba) < BufTagGetForkNum(bb))
7605 0 : return -1;
7606 0 : if (BufTagGetForkNum(ba) > BufTagGetForkNum(bb))
7607 0 : return 1;
7608 :
7609 0 : if (ba->blockNum < bb->blockNum)
7610 0 : return -1;
7611 0 : if (ba->blockNum > bb->blockNum)
7612 0 : return 1;
7613 :
7614 0 : return 0;
7615 : }
7616 :
7617 : /*
7618 : * Comparator determining the writeout order in a checkpoint.
7619 : *
7620 : * It is important that tablespaces are compared first, the logic balancing
7621 : * writes between tablespaces relies on it.
7622 : */
7623 : static inline int
7624 3462826 : ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b)
7625 : {
7626 : /* compare tablespace */
7627 3462826 : if (a->tsId < b->tsId)
7628 8049 : return -1;
7629 3454777 : else if (a->tsId > b->tsId)
7630 29959 : return 1;
7631 : /* compare relation */
7632 3424818 : if (a->relNumber < b->relNumber)
7633 969794 : return -1;
7634 2455024 : else if (a->relNumber > b->relNumber)
7635 939376 : return 1;
7636 : /* compare fork */
7637 1515648 : else if (a->forkNum < b->forkNum)
7638 65969 : return -1;
7639 1449679 : else if (a->forkNum > b->forkNum)
7640 70965 : return 1;
7641 : /* compare block number */
7642 1378714 : else if (a->blockNum < b->blockNum)
7643 671066 : return -1;
7644 707648 : else if (a->blockNum > b->blockNum)
7645 654438 : return 1;
7646 : /* equal page IDs are unlikely, but not impossible */
7647 53210 : return 0;
7648 : }
7649 :
7650 : /*
7651 : * Comparator for a Min-Heap over the per-tablespace checkpoint completion
7652 : * progress.
7653 : */
7654 : static int
7655 283889 : ts_ckpt_progress_comparator(Datum a, Datum b, void *arg)
7656 : {
7657 283889 : CkptTsStatus *sa = (CkptTsStatus *) DatumGetPointer(a);
7658 283889 : CkptTsStatus *sb = (CkptTsStatus *) DatumGetPointer(b);
7659 :
7660 : /* we want a min-heap, so return 1 for the a < b */
7661 283889 : if (sa->progress < sb->progress)
7662 258287 : return 1;
7663 25602 : else if (sa->progress == sb->progress)
7664 902 : return 0;
7665 : else
7666 24700 : return -1;
7667 : }
7668 :
7669 : /*
7670 : * Initialize a writeback context, discarding potential previous state.
7671 : *
7672 : * *max_pending is a pointer instead of an immediate value, so the coalesce
7673 : * limits can easily changed by the GUC mechanism, and so calling code does
7674 : * not have to check the current configuration. A value of 0 means that no
7675 : * writeback control will be performed.
7676 : */
7677 : void
7678 3058 : WritebackContextInit(WritebackContext *context, int *max_pending)
7679 : {
7680 : Assert(*max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
7681 :
7682 3058 : context->max_pending = max_pending;
7683 3058 : context->nr_pending = 0;
7684 3058 : }
7685 :
7686 : /*
7687 : * Add buffer to list of pending writeback requests.
7688 : */
7689 : void
7690 698905 : ScheduleBufferTagForWriteback(WritebackContext *wb_context, IOContext io_context,
7691 : BufferTag *tag)
7692 : {
7693 : PendingWriteback *pending;
7694 :
7695 : /*
7696 : * As pg_flush_data() doesn't do anything with fsync disabled, there's no
7697 : * point in tracking in that case.
7698 : */
7699 698905 : if (io_direct_flags & IO_DIRECT_DATA ||
7700 698380 : !enableFsync)
7701 698903 : return;
7702 :
7703 : /*
7704 : * Add buffer to the pending writeback array, unless writeback control is
7705 : * disabled.
7706 : */
7707 2 : if (*wb_context->max_pending > 0)
7708 : {
7709 : Assert(*wb_context->max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
7710 :
7711 0 : pending = &wb_context->pending_writebacks[wb_context->nr_pending++];
7712 :
7713 0 : pending->tag = *tag;
7714 : }
7715 :
7716 : /*
7717 : * Perform pending flushes if the writeback limit is exceeded. This
7718 : * includes the case where previously an item has been added, but control
7719 : * is now disabled.
7720 : */
7721 2 : if (wb_context->nr_pending >= *wb_context->max_pending)
7722 2 : IssuePendingWritebacks(wb_context, io_context);
7723 : }
7724 :
7725 : #define ST_SORT sort_pending_writebacks
7726 : #define ST_ELEMENT_TYPE PendingWriteback
7727 : #define ST_COMPARE(a, b) buffertag_comparator(&a->tag, &b->tag)
7728 : #define ST_SCOPE static
7729 : #define ST_DEFINE
7730 : #include "lib/sort_template.h"
7731 :
7732 : /*
7733 : * Issue all pending writeback requests, previously scheduled with
7734 : * ScheduleBufferTagForWriteback, to the OS.
7735 : *
7736 : * Because this is only used to improve the OSs IO scheduling we try to never
7737 : * error out - it's just a hint.
7738 : */
7739 : void
7740 1196 : IssuePendingWritebacks(WritebackContext *wb_context, IOContext io_context)
7741 : {
7742 : instr_time io_start;
7743 : int i;
7744 :
7745 1196 : if (wb_context->nr_pending == 0)
7746 1196 : return;
7747 :
7748 : /*
7749 : * Executing the writes in-order can make them a lot faster, and allows to
7750 : * merge writeback requests to consecutive blocks into larger writebacks.
7751 : */
7752 0 : sort_pending_writebacks(wb_context->pending_writebacks,
7753 0 : wb_context->nr_pending);
7754 :
7755 0 : io_start = pgstat_prepare_io_time(track_io_timing);
7756 :
7757 : /*
7758 : * Coalesce neighbouring writes, but nothing else. For that we iterate
7759 : * through the, now sorted, array of pending flushes, and look forward to
7760 : * find all neighbouring (or identical) writes.
7761 : */
7762 0 : for (i = 0; i < wb_context->nr_pending; i++)
7763 : {
7764 : PendingWriteback *cur;
7765 : PendingWriteback *next;
7766 : SMgrRelation reln;
7767 : int ahead;
7768 : BufferTag tag;
7769 : RelFileLocator currlocator;
7770 0 : Size nblocks = 1;
7771 :
7772 0 : cur = &wb_context->pending_writebacks[i];
7773 0 : tag = cur->tag;
7774 0 : currlocator = BufTagGetRelFileLocator(&tag);
7775 :
7776 : /*
7777 : * Peek ahead, into following writeback requests, to see if they can
7778 : * be combined with the current one.
7779 : */
7780 0 : for (ahead = 0; i + ahead + 1 < wb_context->nr_pending; ahead++)
7781 : {
7782 :
7783 0 : next = &wb_context->pending_writebacks[i + ahead + 1];
7784 :
7785 : /* different file, stop */
7786 0 : if (!RelFileLocatorEquals(currlocator,
7787 0 : BufTagGetRelFileLocator(&next->tag)) ||
7788 0 : BufTagGetForkNum(&cur->tag) != BufTagGetForkNum(&next->tag))
7789 : break;
7790 :
7791 : /* ok, block queued twice, skip */
7792 0 : if (cur->tag.blockNum == next->tag.blockNum)
7793 0 : continue;
7794 :
7795 : /* only merge consecutive writes */
7796 0 : if (cur->tag.blockNum + 1 != next->tag.blockNum)
7797 0 : break;
7798 :
7799 0 : nblocks++;
7800 0 : cur = next;
7801 : }
7802 :
7803 0 : i += ahead;
7804 :
7805 : /* and finally tell the kernel to write the data to storage */
7806 0 : reln = smgropen(currlocator, INVALID_PROC_NUMBER);
7807 0 : smgrwriteback(reln, BufTagGetForkNum(&tag), tag.blockNum, nblocks);
7808 : }
7809 :
7810 : /*
7811 : * Assume that writeback requests are only issued for buffers containing
7812 : * blocks of permanent relations.
7813 : */
7814 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
7815 0 : IOOP_WRITEBACK, io_start, wb_context->nr_pending, 0);
7816 :
7817 0 : wb_context->nr_pending = 0;
7818 : }
7819 :
7820 : /* ResourceOwner callbacks */
7821 :
7822 : static void
7823 15 : ResOwnerReleaseBufferIO(Datum res)
7824 : {
7825 15 : Buffer buffer = DatumGetInt32(res);
7826 :
7827 15 : AbortBufferIO(buffer);
7828 15 : }
7829 :
7830 : static char *
7831 0 : ResOwnerPrintBufferIO(Datum res)
7832 : {
7833 0 : Buffer buffer = DatumGetInt32(res);
7834 :
7835 0 : return psprintf("lost track of buffer IO on buffer %d", buffer);
7836 : }
7837 :
7838 : /*
7839 : * Release buffer as part of resource owner cleanup. This will only be called
7840 : * if the buffer is pinned. If this backend held the content lock at the time
7841 : * of the error we also need to release that (note that it is not possible to
7842 : * hold a content lock without a pin).
7843 : */
7844 : static void
7845 10525 : ResOwnerReleaseBuffer(Datum res)
7846 : {
7847 10525 : Buffer buffer = DatumGetInt32(res);
7848 :
7849 : /* Like ReleaseBuffer, but don't call ResourceOwnerForgetBuffer */
7850 10525 : if (!BufferIsValid(buffer))
7851 0 : elog(ERROR, "bad buffer ID: %d", buffer);
7852 :
7853 10525 : if (BufferIsLocal(buffer))
7854 3984 : UnpinLocalBufferNoOwner(buffer);
7855 : else
7856 : {
7857 : PrivateRefCountEntry *ref;
7858 :
7859 6541 : ref = GetPrivateRefCountEntry(buffer, false);
7860 :
7861 : /* not having a private refcount would imply resowner corruption */
7862 : Assert(ref != NULL);
7863 :
7864 : /*
7865 : * If the buffer was locked at the time of the resowner release,
7866 : * release the lock now. This should only happen after errors.
7867 : */
7868 6541 : if (ref->data.lockmode != BUFFER_LOCK_UNLOCK)
7869 : {
7870 112 : BufferDesc *buf = GetBufferDescriptor(buffer - 1);
7871 :
7872 112 : HOLD_INTERRUPTS(); /* match the upcoming RESUME_INTERRUPTS */
7873 112 : BufferLockUnlock(buffer, buf);
7874 : }
7875 :
7876 6541 : UnpinBufferNoOwner(GetBufferDescriptor(buffer - 1));
7877 : }
7878 10525 : }
7879 :
7880 : static char *
7881 0 : ResOwnerPrintBuffer(Datum res)
7882 : {
7883 0 : return DebugPrintBufferRefcount(DatumGetInt32(res));
7884 : }
7885 :
7886 : /*
7887 : * Helper function to evict unpinned buffer whose buffer header lock is
7888 : * already acquired.
7889 : */
7890 : static bool
7891 2540 : EvictUnpinnedBufferInternal(BufferDesc *desc, bool *buffer_flushed)
7892 : {
7893 : uint64 buf_state;
7894 : bool result;
7895 :
7896 2540 : *buffer_flushed = false;
7897 :
7898 2540 : buf_state = pg_atomic_read_u64(&(desc->state));
7899 : Assert(buf_state & BM_LOCKED);
7900 :
7901 2540 : if ((buf_state & BM_VALID) == 0)
7902 : {
7903 0 : UnlockBufHdr(desc);
7904 0 : return false;
7905 : }
7906 :
7907 : /* Check that it's not pinned already. */
7908 2540 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
7909 : {
7910 0 : UnlockBufHdr(desc);
7911 0 : return false;
7912 : }
7913 :
7914 2540 : PinBuffer_Locked(desc); /* releases spinlock */
7915 :
7916 : /* If it was dirty, try to clean it once. */
7917 2540 : if (buf_state & BM_DIRTY)
7918 : {
7919 1077 : FlushUnlockedBuffer(desc, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
7920 1077 : *buffer_flushed = true;
7921 : }
7922 :
7923 : /* This will return false if it becomes dirty or someone else pins it. */
7924 2540 : result = InvalidateVictimBuffer(desc);
7925 :
7926 2540 : UnpinBuffer(desc);
7927 :
7928 2540 : return result;
7929 : }
7930 :
7931 : /*
7932 : * Try to evict the current block in a shared buffer.
7933 : *
7934 : * This function is intended for testing/development use only!
7935 : *
7936 : * To succeed, the buffer must not be pinned on entry, so if the caller had a
7937 : * particular block in mind, it might already have been replaced by some other
7938 : * block by the time this function runs. It's also unpinned on return, so the
7939 : * buffer might be occupied again by the time control is returned, potentially
7940 : * even by the same block. This inherent raciness without other interlocking
7941 : * makes the function unsuitable for non-testing usage.
7942 : *
7943 : * *buffer_flushed is set to true if the buffer was dirty and has been
7944 : * flushed, false otherwise. However, *buffer_flushed=true does not
7945 : * necessarily mean that we flushed the buffer, it could have been flushed by
7946 : * someone else.
7947 : *
7948 : * Returns true if the buffer was valid and it has now been made invalid.
7949 : * Returns false if it wasn't valid, if it couldn't be evicted due to a pin,
7950 : * or if the buffer becomes dirty again while we're trying to write it out.
7951 : */
7952 : bool
7953 137 : EvictUnpinnedBuffer(Buffer buf, bool *buffer_flushed)
7954 : {
7955 : BufferDesc *desc;
7956 :
7957 : Assert(BufferIsValid(buf) && !BufferIsLocal(buf));
7958 :
7959 : /* Make sure we can pin the buffer. */
7960 137 : ResourceOwnerEnlarge(CurrentResourceOwner);
7961 137 : ReservePrivateRefCountEntry();
7962 :
7963 137 : desc = GetBufferDescriptor(buf - 1);
7964 137 : LockBufHdr(desc);
7965 :
7966 137 : return EvictUnpinnedBufferInternal(desc, buffer_flushed);
7967 : }
7968 :
7969 : /*
7970 : * Try to evict all the shared buffers.
7971 : *
7972 : * This function is intended for testing/development use only! See
7973 : * EvictUnpinnedBuffer().
7974 : *
7975 : * The buffers_* parameters are mandatory and indicate the total count of
7976 : * buffers that:
7977 : * - buffers_evicted - were evicted
7978 : * - buffers_flushed - were flushed
7979 : * - buffers_skipped - could not be evicted
7980 : */
7981 : void
7982 1 : EvictAllUnpinnedBuffers(int32 *buffers_evicted, int32 *buffers_flushed,
7983 : int32 *buffers_skipped)
7984 : {
7985 1 : *buffers_evicted = 0;
7986 1 : *buffers_skipped = 0;
7987 1 : *buffers_flushed = 0;
7988 :
7989 16385 : for (int buf = 1; buf <= NBuffers; buf++)
7990 : {
7991 16384 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
7992 : uint64 buf_state;
7993 : bool buffer_flushed;
7994 :
7995 16384 : CHECK_FOR_INTERRUPTS();
7996 :
7997 16384 : buf_state = pg_atomic_read_u64(&desc->state);
7998 16384 : if (!(buf_state & BM_VALID))
7999 14327 : continue;
8000 :
8001 2057 : ResourceOwnerEnlarge(CurrentResourceOwner);
8002 2057 : ReservePrivateRefCountEntry();
8003 :
8004 2057 : LockBufHdr(desc);
8005 :
8006 2057 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
8007 2057 : (*buffers_evicted)++;
8008 : else
8009 0 : (*buffers_skipped)++;
8010 :
8011 2057 : if (buffer_flushed)
8012 980 : (*buffers_flushed)++;
8013 : }
8014 1 : }
8015 :
8016 : /*
8017 : * Try to evict all the shared buffers containing provided relation's pages.
8018 : *
8019 : * This function is intended for testing/development use only! See
8020 : * EvictUnpinnedBuffer().
8021 : *
8022 : * The caller must hold at least AccessShareLock on the relation to prevent
8023 : * the relation from being dropped.
8024 : *
8025 : * The buffers_* parameters are mandatory and indicate the total count of
8026 : * buffers that:
8027 : * - buffers_evicted - were evicted
8028 : * - buffers_flushed - were flushed
8029 : * - buffers_skipped - could not be evicted
8030 : */
8031 : void
8032 29 : EvictRelUnpinnedBuffers(Relation rel, int32 *buffers_evicted,
8033 : int32 *buffers_flushed, int32 *buffers_skipped)
8034 : {
8035 : Assert(!RelationUsesLocalBuffers(rel));
8036 :
8037 29 : *buffers_skipped = 0;
8038 29 : *buffers_evicted = 0;
8039 29 : *buffers_flushed = 0;
8040 :
8041 475165 : for (int buf = 1; buf <= NBuffers; buf++)
8042 : {
8043 475136 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
8044 475136 : uint64 buf_state = pg_atomic_read_u64(&(desc->state));
8045 : bool buffer_flushed;
8046 :
8047 475136 : CHECK_FOR_INTERRUPTS();
8048 :
8049 : /* An unlocked precheck should be safe and saves some cycles. */
8050 475136 : if ((buf_state & BM_VALID) == 0 ||
8051 62585 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
8052 474790 : continue;
8053 :
8054 : /* Make sure we can pin the buffer. */
8055 346 : ResourceOwnerEnlarge(CurrentResourceOwner);
8056 346 : ReservePrivateRefCountEntry();
8057 :
8058 346 : buf_state = LockBufHdr(desc);
8059 :
8060 : /* recheck, could have changed without the lock */
8061 346 : if ((buf_state & BM_VALID) == 0 ||
8062 346 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
8063 : {
8064 0 : UnlockBufHdr(desc);
8065 0 : continue;
8066 : }
8067 :
8068 346 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
8069 346 : (*buffers_evicted)++;
8070 : else
8071 0 : (*buffers_skipped)++;
8072 :
8073 346 : if (buffer_flushed)
8074 78 : (*buffers_flushed)++;
8075 : }
8076 29 : }
8077 :
8078 : /*
8079 : * Helper function to mark unpinned buffer dirty whose buffer header lock is
8080 : * already acquired.
8081 : */
8082 : static bool
8083 36 : MarkDirtyUnpinnedBufferInternal(Buffer buf, BufferDesc *desc,
8084 : bool *buffer_already_dirty)
8085 : {
8086 : uint64 buf_state;
8087 36 : bool result = false;
8088 :
8089 36 : *buffer_already_dirty = false;
8090 :
8091 36 : buf_state = pg_atomic_read_u64(&(desc->state));
8092 : Assert(buf_state & BM_LOCKED);
8093 :
8094 36 : if ((buf_state & BM_VALID) == 0)
8095 : {
8096 1 : UnlockBufHdr(desc);
8097 1 : return false;
8098 : }
8099 :
8100 : /* Check that it's not pinned already. */
8101 35 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
8102 : {
8103 0 : UnlockBufHdr(desc);
8104 0 : return false;
8105 : }
8106 :
8107 : /* Pin the buffer and then release the buffer spinlock */
8108 35 : PinBuffer_Locked(desc);
8109 :
8110 : /* If it was not already dirty, mark it as dirty. */
8111 35 : if (!(buf_state & BM_DIRTY))
8112 : {
8113 16 : BufferLockAcquire(buf, desc, BUFFER_LOCK_EXCLUSIVE);
8114 16 : MarkBufferDirty(buf);
8115 16 : result = true;
8116 16 : BufferLockUnlock(buf, desc);
8117 : }
8118 : else
8119 19 : *buffer_already_dirty = true;
8120 :
8121 35 : UnpinBuffer(desc);
8122 :
8123 35 : return result;
8124 : }
8125 :
8126 : /*
8127 : * Try to mark the provided shared buffer as dirty.
8128 : *
8129 : * This function is intended for testing/development use only!
8130 : *
8131 : * Same as EvictUnpinnedBuffer() but with MarkBufferDirty() call inside.
8132 : *
8133 : * The buffer_already_dirty parameter is mandatory and indicate if the buffer
8134 : * could not be dirtied because it is already dirty.
8135 : *
8136 : * Returns true if the buffer has successfully been marked as dirty.
8137 : */
8138 : bool
8139 1 : MarkDirtyUnpinnedBuffer(Buffer buf, bool *buffer_already_dirty)
8140 : {
8141 : BufferDesc *desc;
8142 1 : bool buffer_dirtied = false;
8143 :
8144 : Assert(!BufferIsLocal(buf));
8145 :
8146 : /* Make sure we can pin the buffer. */
8147 1 : ResourceOwnerEnlarge(CurrentResourceOwner);
8148 1 : ReservePrivateRefCountEntry();
8149 :
8150 1 : desc = GetBufferDescriptor(buf - 1);
8151 1 : LockBufHdr(desc);
8152 :
8153 1 : buffer_dirtied = MarkDirtyUnpinnedBufferInternal(buf, desc, buffer_already_dirty);
8154 : /* Both can not be true at the same time */
8155 : Assert(!(buffer_dirtied && *buffer_already_dirty));
8156 :
8157 1 : return buffer_dirtied;
8158 : }
8159 :
8160 : /*
8161 : * Try to mark all the shared buffers containing provided relation's pages as
8162 : * dirty.
8163 : *
8164 : * This function is intended for testing/development use only! See
8165 : * MarkDirtyUnpinnedBuffer().
8166 : *
8167 : * The buffers_* parameters are mandatory and indicate the total count of
8168 : * buffers that:
8169 : * - buffers_dirtied - were dirtied
8170 : * - buffers_already_dirty - were already dirty
8171 : * - buffers_skipped - could not be dirtied because of a reason different
8172 : * than a buffer being already dirty.
8173 : */
8174 : void
8175 1 : MarkDirtyRelUnpinnedBuffers(Relation rel,
8176 : int32 *buffers_dirtied,
8177 : int32 *buffers_already_dirty,
8178 : int32 *buffers_skipped)
8179 : {
8180 : Assert(!RelationUsesLocalBuffers(rel));
8181 :
8182 1 : *buffers_dirtied = 0;
8183 1 : *buffers_already_dirty = 0;
8184 1 : *buffers_skipped = 0;
8185 :
8186 16385 : for (int buf = 1; buf <= NBuffers; buf++)
8187 : {
8188 16384 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
8189 16384 : uint64 buf_state = pg_atomic_read_u64(&(desc->state));
8190 : bool buffer_already_dirty;
8191 :
8192 16384 : CHECK_FOR_INTERRUPTS();
8193 :
8194 : /* An unlocked precheck should be safe and saves some cycles. */
8195 16384 : if ((buf_state & BM_VALID) == 0 ||
8196 27 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
8197 16384 : continue;
8198 :
8199 : /* Make sure we can pin the buffer. */
8200 0 : ResourceOwnerEnlarge(CurrentResourceOwner);
8201 0 : ReservePrivateRefCountEntry();
8202 :
8203 0 : buf_state = LockBufHdr(desc);
8204 :
8205 : /* recheck, could have changed without the lock */
8206 0 : if ((buf_state & BM_VALID) == 0 ||
8207 0 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
8208 : {
8209 0 : UnlockBufHdr(desc);
8210 0 : continue;
8211 : }
8212 :
8213 0 : if (MarkDirtyUnpinnedBufferInternal(buf, desc, &buffer_already_dirty))
8214 0 : (*buffers_dirtied)++;
8215 0 : else if (buffer_already_dirty)
8216 0 : (*buffers_already_dirty)++;
8217 : else
8218 0 : (*buffers_skipped)++;
8219 : }
8220 1 : }
8221 :
8222 : /*
8223 : * Try to mark all the shared buffers as dirty.
8224 : *
8225 : * This function is intended for testing/development use only! See
8226 : * MarkDirtyUnpinnedBuffer().
8227 : *
8228 : * See MarkDirtyRelUnpinnedBuffers() above for details about the buffers_*
8229 : * parameters.
8230 : */
8231 : void
8232 1 : MarkDirtyAllUnpinnedBuffers(int32 *buffers_dirtied,
8233 : int32 *buffers_already_dirty,
8234 : int32 *buffers_skipped)
8235 : {
8236 1 : *buffers_dirtied = 0;
8237 1 : *buffers_already_dirty = 0;
8238 1 : *buffers_skipped = 0;
8239 :
8240 16385 : for (int buf = 1; buf <= NBuffers; buf++)
8241 : {
8242 16384 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
8243 : uint64 buf_state;
8244 : bool buffer_already_dirty;
8245 :
8246 16384 : CHECK_FOR_INTERRUPTS();
8247 :
8248 16384 : buf_state = pg_atomic_read_u64(&desc->state);
8249 16384 : if (!(buf_state & BM_VALID))
8250 16349 : continue;
8251 :
8252 35 : ResourceOwnerEnlarge(CurrentResourceOwner);
8253 35 : ReservePrivateRefCountEntry();
8254 :
8255 35 : LockBufHdr(desc);
8256 :
8257 35 : if (MarkDirtyUnpinnedBufferInternal(buf, desc, &buffer_already_dirty))
8258 16 : (*buffers_dirtied)++;
8259 19 : else if (buffer_already_dirty)
8260 19 : (*buffers_already_dirty)++;
8261 : else
8262 0 : (*buffers_skipped)++;
8263 : }
8264 1 : }
8265 :
8266 : /*
8267 : * Generic implementation of the AIO handle staging callback for readv/writev
8268 : * on local/shared buffers.
8269 : *
8270 : * Each readv/writev can target multiple buffers. The buffers have already
8271 : * been registered with the IO handle.
8272 : *
8273 : * To make the IO ready for execution ("staging"), we need to ensure that the
8274 : * targeted buffers are in an appropriate state while the IO is ongoing. For
8275 : * that the AIO subsystem needs to have its own buffer pin, otherwise an error
8276 : * in this backend could lead to this backend's buffer pin being released as
8277 : * part of error handling, which in turn could lead to the buffer being
8278 : * replaced while IO is ongoing.
8279 : */
8280 : static pg_attribute_always_inline void
8281 1477667 : buffer_stage_common(PgAioHandle *ioh, bool is_write, bool is_temp)
8282 : {
8283 : uint64 *io_data;
8284 : uint8 handle_data_len;
8285 : PgAioWaitRef io_ref;
8286 1477667 : BufferTag first PG_USED_FOR_ASSERTS_ONLY = {0};
8287 :
8288 1477667 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
8289 :
8290 1477667 : pgaio_io_get_wref(ioh, &io_ref);
8291 :
8292 : /* iterate over all buffers affected by the vectored readv/writev */
8293 3143718 : for (int i = 0; i < handle_data_len; i++)
8294 : {
8295 1666051 : Buffer buffer = (Buffer) io_data[i];
8296 1666051 : BufferDesc *buf_hdr = is_temp ?
8297 11052 : GetLocalBufferDescriptor(-buffer - 1)
8298 1666051 : : GetBufferDescriptor(buffer - 1);
8299 : uint64 buf_state;
8300 :
8301 : /*
8302 : * Check that all the buffers are actually ones that could conceivably
8303 : * be done in one IO, i.e. are sequential. This is the last
8304 : * buffer-aware code before IO is actually executed and confusion
8305 : * about which buffers are targeted by IO can be hard to debug, making
8306 : * it worth doing extra-paranoid checks.
8307 : */
8308 1666051 : if (i == 0)
8309 1477667 : first = buf_hdr->tag;
8310 : else
8311 : {
8312 : Assert(buf_hdr->tag.relNumber == first.relNumber);
8313 : Assert(buf_hdr->tag.blockNum == first.blockNum + i);
8314 : }
8315 :
8316 1666051 : if (is_temp)
8317 11052 : buf_state = pg_atomic_read_u64(&buf_hdr->state);
8318 : else
8319 1654999 : buf_state = LockBufHdr(buf_hdr);
8320 :
8321 : /* verify the buffer is in the expected state */
8322 : Assert(buf_state & BM_TAG_VALID);
8323 : if (is_write)
8324 : {
8325 : Assert(buf_state & BM_VALID);
8326 : Assert(buf_state & BM_DIRTY);
8327 : }
8328 : else
8329 : {
8330 : Assert(!(buf_state & BM_VALID));
8331 : Assert(!(buf_state & BM_DIRTY));
8332 : }
8333 :
8334 : /* temp buffers don't use BM_IO_IN_PROGRESS */
8335 1666051 : if (!is_temp)
8336 : Assert(buf_state & BM_IO_IN_PROGRESS);
8337 :
8338 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) >= 1);
8339 :
8340 : /*
8341 : * Reflect that the buffer is now owned by the AIO subsystem.
8342 : *
8343 : * For local buffers: This can't be done just via LocalRefCount, as
8344 : * one might initially think, as this backend could error out while
8345 : * AIO is still in progress, releasing all the pins by the backend
8346 : * itself.
8347 : *
8348 : * This pin is released again in TerminateBufferIO().
8349 : */
8350 1666051 : buf_hdr->io_wref = io_ref;
8351 :
8352 1666051 : if (is_temp)
8353 : {
8354 11052 : buf_state += BUF_REFCOUNT_ONE;
8355 11052 : pg_atomic_unlocked_write_u64(&buf_hdr->state, buf_state);
8356 : }
8357 : else
8358 1654999 : UnlockBufHdrExt(buf_hdr, buf_state, 0, 0, 1);
8359 :
8360 : /*
8361 : * Ensure the content lock that prevents buffer modifications while
8362 : * the buffer is being written out is not released early due to an
8363 : * error.
8364 : */
8365 1666051 : if (is_write && !is_temp)
8366 : {
8367 : Assert(BufferLockHeldByMe(buf_hdr));
8368 :
8369 : /*
8370 : * Lock is now owned by AIO subsystem.
8371 : */
8372 0 : BufferLockDisown(buffer, buf_hdr);
8373 : }
8374 :
8375 : /*
8376 : * Stop tracking this buffer via the resowner - the AIO system now
8377 : * keeps track.
8378 : */
8379 1666051 : if (!is_temp)
8380 1654999 : ResourceOwnerForgetBufferIO(CurrentResourceOwner, buffer);
8381 : }
8382 1477667 : }
8383 :
8384 : /*
8385 : * Decode readv errors as encoded by buffer_readv_encode_error().
8386 : */
8387 : static inline void
8388 450 : buffer_readv_decode_error(PgAioResult result,
8389 : bool *zeroed_any,
8390 : bool *ignored_any,
8391 : uint8 *zeroed_or_error_count,
8392 : uint8 *checkfail_count,
8393 : uint8 *first_off)
8394 : {
8395 450 : uint32 rem_error = result.error_data;
8396 :
8397 : /* see static asserts in buffer_readv_encode_error */
8398 : #define READV_COUNT_BITS 7
8399 : #define READV_COUNT_MASK ((1 << READV_COUNT_BITS) - 1)
8400 :
8401 450 : *zeroed_any = rem_error & 1;
8402 450 : rem_error >>= 1;
8403 :
8404 450 : *ignored_any = rem_error & 1;
8405 450 : rem_error >>= 1;
8406 :
8407 450 : *zeroed_or_error_count = rem_error & READV_COUNT_MASK;
8408 450 : rem_error >>= READV_COUNT_BITS;
8409 :
8410 450 : *checkfail_count = rem_error & READV_COUNT_MASK;
8411 450 : rem_error >>= READV_COUNT_BITS;
8412 :
8413 450 : *first_off = rem_error & READV_COUNT_MASK;
8414 450 : rem_error >>= READV_COUNT_BITS;
8415 450 : }
8416 :
8417 : /*
8418 : * Helper to encode errors for buffer_readv_complete()
8419 : *
8420 : * Errors are encoded as follows:
8421 : * - bit 0 indicates whether any page was zeroed (1) or not (0)
8422 : * - bit 1 indicates whether any checksum failure was ignored (1) or not (0)
8423 : * - next READV_COUNT_BITS bits indicate the number of errored or zeroed pages
8424 : * - next READV_COUNT_BITS bits indicate the number of checksum failures
8425 : * - next READV_COUNT_BITS bits indicate the first offset of the first page
8426 : * that was errored or zeroed or, if no errors/zeroes, the first ignored
8427 : * checksum
8428 : */
8429 : static inline void
8430 192 : buffer_readv_encode_error(PgAioResult *result,
8431 : bool is_temp,
8432 : bool zeroed_any,
8433 : bool ignored_any,
8434 : uint8 error_count,
8435 : uint8 zeroed_count,
8436 : uint8 checkfail_count,
8437 : uint8 first_error_off,
8438 : uint8 first_zeroed_off,
8439 : uint8 first_ignored_off)
8440 : {
8441 :
8442 192 : uint8 shift = 0;
8443 192 : uint8 zeroed_or_error_count =
8444 : error_count > 0 ? error_count : zeroed_count;
8445 : uint8 first_off;
8446 :
8447 : StaticAssertDecl(PG_IOV_MAX <= 1 << READV_COUNT_BITS,
8448 : "PG_IOV_MAX is bigger than reserved space for error data");
8449 : StaticAssertDecl((1 + 1 + 3 * READV_COUNT_BITS) <= PGAIO_RESULT_ERROR_BITS,
8450 : "PGAIO_RESULT_ERROR_BITS is insufficient for buffer_readv");
8451 :
8452 : /*
8453 : * We only have space to encode one offset - but luckily that's good
8454 : * enough. If there is an error, the error is the interesting offset, same
8455 : * with a zeroed buffer vs an ignored buffer.
8456 : */
8457 192 : if (error_count > 0)
8458 94 : first_off = first_error_off;
8459 98 : else if (zeroed_count > 0)
8460 80 : first_off = first_zeroed_off;
8461 : else
8462 18 : first_off = first_ignored_off;
8463 :
8464 : Assert(!zeroed_any || error_count == 0);
8465 :
8466 192 : result->error_data = 0;
8467 :
8468 192 : result->error_data |= zeroed_any << shift;
8469 192 : shift += 1;
8470 :
8471 192 : result->error_data |= ignored_any << shift;
8472 192 : shift += 1;
8473 :
8474 192 : result->error_data |= ((uint32) zeroed_or_error_count) << shift;
8475 192 : shift += READV_COUNT_BITS;
8476 :
8477 192 : result->error_data |= ((uint32) checkfail_count) << shift;
8478 192 : shift += READV_COUNT_BITS;
8479 :
8480 192 : result->error_data |= ((uint32) first_off) << shift;
8481 192 : shift += READV_COUNT_BITS;
8482 :
8483 192 : result->id = is_temp ? PGAIO_HCB_LOCAL_BUFFER_READV :
8484 : PGAIO_HCB_SHARED_BUFFER_READV;
8485 :
8486 192 : if (error_count > 0)
8487 94 : result->status = PGAIO_RS_ERROR;
8488 : else
8489 98 : result->status = PGAIO_RS_WARNING;
8490 :
8491 : /*
8492 : * The encoding is complicated enough to warrant cross-checking it against
8493 : * the decode function.
8494 : */
8495 : #ifdef USE_ASSERT_CHECKING
8496 : {
8497 : bool zeroed_any_2,
8498 : ignored_any_2;
8499 : uint8 zeroed_or_error_count_2,
8500 : checkfail_count_2,
8501 : first_off_2;
8502 :
8503 : buffer_readv_decode_error(*result,
8504 : &zeroed_any_2, &ignored_any_2,
8505 : &zeroed_or_error_count_2,
8506 : &checkfail_count_2,
8507 : &first_off_2);
8508 : Assert(zeroed_any == zeroed_any_2);
8509 : Assert(ignored_any == ignored_any_2);
8510 : Assert(zeroed_or_error_count == zeroed_or_error_count_2);
8511 : Assert(checkfail_count == checkfail_count_2);
8512 : Assert(first_off == first_off_2);
8513 : }
8514 : #endif
8515 :
8516 : #undef READV_COUNT_BITS
8517 : #undef READV_COUNT_MASK
8518 192 : }
8519 :
8520 : /*
8521 : * Helper for AIO readv completion callbacks, supporting both shared and temp
8522 : * buffers. Gets called once for each buffer in a multi-page read.
8523 : */
8524 : static pg_attribute_always_inline void
8525 1493430 : buffer_readv_complete_one(PgAioTargetData *td, uint8 buf_off, Buffer buffer,
8526 : uint8 flags, bool failed, bool is_temp,
8527 : bool *buffer_invalid,
8528 : bool *failed_checksum,
8529 : bool *ignored_checksum,
8530 : bool *zeroed_buffer)
8531 : {
8532 1493430 : BufferDesc *buf_hdr = is_temp ?
8533 11052 : GetLocalBufferDescriptor(-buffer - 1)
8534 1493430 : : GetBufferDescriptor(buffer - 1);
8535 1493430 : BufferTag tag = buf_hdr->tag;
8536 1493430 : char *bufdata = BufferGetBlock(buffer);
8537 : uint64 set_flag_bits;
8538 : int piv_flags;
8539 :
8540 : /* check that the buffer is in the expected state for a read */
8541 : #ifdef USE_ASSERT_CHECKING
8542 : {
8543 : uint64 buf_state = pg_atomic_read_u64(&buf_hdr->state);
8544 :
8545 : Assert(buf_state & BM_TAG_VALID);
8546 : Assert(!(buf_state & BM_VALID));
8547 : /* temp buffers don't use BM_IO_IN_PROGRESS */
8548 : if (!is_temp)
8549 : Assert(buf_state & BM_IO_IN_PROGRESS);
8550 : Assert(!(buf_state & BM_DIRTY));
8551 : }
8552 : #endif
8553 :
8554 1493430 : *buffer_invalid = false;
8555 1493430 : *failed_checksum = false;
8556 1493430 : *ignored_checksum = false;
8557 1493430 : *zeroed_buffer = false;
8558 :
8559 : /*
8560 : * We ask PageIsVerified() to only log the message about checksum errors,
8561 : * as the completion might be run in any backend (or IO workers). We will
8562 : * report checksum errors in buffer_readv_report().
8563 : */
8564 1493430 : piv_flags = PIV_LOG_LOG;
8565 :
8566 : /* the local zero_damaged_pages may differ from the definer's */
8567 1493430 : if (flags & READ_BUFFERS_IGNORE_CHECKSUM_FAILURES)
8568 38 : piv_flags |= PIV_IGNORE_CHECKSUM_FAILURE;
8569 :
8570 : /*
8571 : * If the buffers are marked for zero on error, we want to log that in
8572 : * case of a checksum failure.
8573 : */
8574 1493430 : if (flags & READ_BUFFERS_ZERO_ON_ERROR)
8575 46802 : piv_flags |= PIV_ZERO_BUFFERS_ON_ERROR;
8576 :
8577 : /* Check for garbage data. */
8578 1493430 : if (!failed)
8579 : {
8580 : /*
8581 : * If the buffer is not currently pinned by this backend, e.g. because
8582 : * we're completing this IO after an error, the buffer data will have
8583 : * been marked as inaccessible when the buffer was unpinned. The AIO
8584 : * subsystem holds a pin, but that doesn't prevent the buffer from
8585 : * having been marked as inaccessible. The completion might also be
8586 : * executed in a different process.
8587 : */
8588 : #ifdef USE_VALGRIND
8589 : if (!BufferIsPinned(buffer))
8590 : VALGRIND_MAKE_MEM_DEFINED(bufdata, BLCKSZ);
8591 : #endif
8592 :
8593 1492658 : if (!PageIsVerified((Page) bufdata, tag.blockNum, piv_flags,
8594 : failed_checksum))
8595 : {
8596 96 : if (flags & READ_BUFFERS_ZERO_ON_ERROR)
8597 : {
8598 46 : memset(bufdata, 0, BLCKSZ);
8599 46 : *zeroed_buffer = true;
8600 : }
8601 : else
8602 : {
8603 50 : *buffer_invalid = true;
8604 : /* mark buffer as having failed */
8605 50 : failed = true;
8606 : }
8607 : }
8608 1492562 : else if (*failed_checksum)
8609 12 : *ignored_checksum = true;
8610 :
8611 : /* undo what we did above */
8612 : #ifdef USE_VALGRIND
8613 : if (!BufferIsPinned(buffer))
8614 : VALGRIND_MAKE_MEM_NOACCESS(bufdata, BLCKSZ);
8615 : #endif
8616 :
8617 : /*
8618 : * Immediately log a message about the invalid page, but only to the
8619 : * server log. The reason to do so immediately is that this may be
8620 : * executed in a different backend than the one that originated the
8621 : * request. The reason to do so immediately is that the originator
8622 : * might not process the query result immediately (because it is busy
8623 : * doing another part of query processing) or at all (e.g. if it was
8624 : * cancelled or errored out due to another IO also failing). The
8625 : * definer of the IO will emit an ERROR or WARNING when processing the
8626 : * IO's results
8627 : *
8628 : * To avoid duplicating the code to emit these log messages, we reuse
8629 : * buffer_readv_report().
8630 : */
8631 1492658 : if (*buffer_invalid || *failed_checksum || *zeroed_buffer)
8632 : {
8633 108 : PgAioResult result_one = {0};
8634 :
8635 108 : buffer_readv_encode_error(&result_one, is_temp,
8636 108 : *zeroed_buffer,
8637 108 : *ignored_checksum,
8638 108 : *buffer_invalid,
8639 108 : *zeroed_buffer ? 1 : 0,
8640 108 : *failed_checksum ? 1 : 0,
8641 : buf_off, buf_off, buf_off);
8642 108 : pgaio_result_report(result_one, td, LOG_SERVER_ONLY);
8643 : }
8644 : }
8645 :
8646 : /* Terminate I/O and set BM_VALID. */
8647 1493430 : set_flag_bits = failed ? BM_IO_ERROR : BM_VALID;
8648 1493430 : if (is_temp)
8649 11052 : TerminateLocalBufferIO(buf_hdr, false, set_flag_bits, true);
8650 : else
8651 1482378 : TerminateBufferIO(buf_hdr, false, set_flag_bits, false, true);
8652 :
8653 : /*
8654 : * Call the BUFFER_READ_DONE tracepoint in the callback, even though the
8655 : * callback may not be executed in the same backend that called
8656 : * BUFFER_READ_START. The alternative would be to defer calling the
8657 : * tracepoint to a later point (e.g. the local completion callback for
8658 : * shared buffer reads), which seems even less helpful.
8659 : */
8660 : TRACE_POSTGRESQL_BUFFER_READ_DONE(tag.forkNum,
8661 : tag.blockNum,
8662 : tag.spcOid,
8663 : tag.dbOid,
8664 : tag.relNumber,
8665 : is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
8666 : false);
8667 1493430 : }
8668 :
8669 : /*
8670 : * Perform completion handling of a single AIO read. This read may cover
8671 : * multiple blocks / buffers.
8672 : *
8673 : * Shared between shared and local buffers, to reduce code duplication.
8674 : */
8675 : static pg_attribute_always_inline PgAioResult
8676 1342110 : buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
8677 : uint8 cb_data, bool is_temp)
8678 : {
8679 1342110 : PgAioResult result = prior_result;
8680 1342110 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
8681 1342110 : uint8 first_error_off = 0;
8682 1342110 : uint8 first_zeroed_off = 0;
8683 1342110 : uint8 first_ignored_off = 0;
8684 1342110 : uint8 error_count = 0;
8685 1342110 : uint8 zeroed_count = 0;
8686 1342110 : uint8 ignored_count = 0;
8687 1342110 : uint8 checkfail_count = 0;
8688 : uint64 *io_data;
8689 : uint8 handle_data_len;
8690 :
8691 : if (is_temp)
8692 : {
8693 : Assert(td->smgr.is_temp);
8694 : Assert(pgaio_io_get_owner(ioh) == MyProcNumber);
8695 : }
8696 : else
8697 : Assert(!td->smgr.is_temp);
8698 :
8699 : /*
8700 : * Iterate over all the buffers affected by this IO and call the
8701 : * per-buffer completion function for each buffer.
8702 : */
8703 1342110 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
8704 2835540 : for (uint8 buf_off = 0; buf_off < handle_data_len; buf_off++)
8705 : {
8706 1493430 : Buffer buf = io_data[buf_off];
8707 : bool failed;
8708 1493430 : bool failed_verification = false;
8709 1493430 : bool failed_checksum = false;
8710 1493430 : bool zeroed_buffer = false;
8711 1493430 : bool ignored_checksum = false;
8712 :
8713 : Assert(BufferIsValid(buf));
8714 :
8715 : /*
8716 : * If the entire I/O failed on a lower-level, each buffer needs to be
8717 : * marked as failed. In case of a partial read, the first few buffers
8718 : * may be ok.
8719 : */
8720 1493430 : failed =
8721 1493430 : prior_result.status == PGAIO_RS_ERROR
8722 1493430 : || prior_result.result <= buf_off;
8723 :
8724 1493430 : buffer_readv_complete_one(td, buf_off, buf, cb_data, failed, is_temp,
8725 : &failed_verification,
8726 : &failed_checksum,
8727 : &ignored_checksum,
8728 : &zeroed_buffer);
8729 :
8730 : /*
8731 : * Track information about the number of different kinds of error
8732 : * conditions across all pages, as there can be multiple pages failing
8733 : * verification as part of one IO.
8734 : */
8735 1493430 : if (failed_verification && !zeroed_buffer && error_count++ == 0)
8736 44 : first_error_off = buf_off;
8737 1493430 : if (zeroed_buffer && zeroed_count++ == 0)
8738 34 : first_zeroed_off = buf_off;
8739 1493430 : if (ignored_checksum && ignored_count++ == 0)
8740 10 : first_ignored_off = buf_off;
8741 1493430 : if (failed_checksum)
8742 32 : checkfail_count++;
8743 : }
8744 :
8745 : /*
8746 : * If the smgr read succeeded [partially] and page verification failed for
8747 : * some of the pages, adjust the IO's result state appropriately.
8748 : */
8749 1342110 : if (prior_result.status != PGAIO_RS_ERROR &&
8750 1342055 : (error_count > 0 || ignored_count > 0 || zeroed_count > 0))
8751 : {
8752 84 : buffer_readv_encode_error(&result, is_temp,
8753 : zeroed_count > 0, ignored_count > 0,
8754 : error_count, zeroed_count, checkfail_count,
8755 : first_error_off, first_zeroed_off,
8756 : first_ignored_off);
8757 84 : pgaio_result_report(result, td, DEBUG1);
8758 : }
8759 :
8760 : /*
8761 : * For shared relations this reporting is done in
8762 : * shared_buffer_readv_complete_local().
8763 : */
8764 1342110 : if (is_temp && checkfail_count > 0)
8765 2 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
8766 : checkfail_count);
8767 :
8768 1342110 : return result;
8769 : }
8770 :
8771 : /*
8772 : * AIO error reporting callback for aio_shared_buffer_readv_cb and
8773 : * aio_local_buffer_readv_cb.
8774 : *
8775 : * The error is encoded / decoded in buffer_readv_encode_error() /
8776 : * buffer_readv_decode_error().
8777 : */
8778 : static void
8779 272 : buffer_readv_report(PgAioResult result, const PgAioTargetData *td,
8780 : int elevel)
8781 : {
8782 272 : int nblocks = td->smgr.nblocks;
8783 272 : BlockNumber first = td->smgr.blockNum;
8784 272 : BlockNumber last = first + nblocks - 1;
8785 272 : ProcNumber errProc =
8786 272 : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER;
8787 : RelPathStr rpath =
8788 272 : relpathbackend(td->smgr.rlocator, errProc, td->smgr.forkNum);
8789 : bool zeroed_any,
8790 : ignored_any;
8791 : uint8 zeroed_or_error_count,
8792 : checkfail_count,
8793 : first_off;
8794 : uint8 affected_count;
8795 : const char *msg_one,
8796 : *msg_mult,
8797 : *det_mult,
8798 : *hint_mult;
8799 :
8800 272 : buffer_readv_decode_error(result, &zeroed_any, &ignored_any,
8801 : &zeroed_or_error_count,
8802 : &checkfail_count,
8803 : &first_off);
8804 :
8805 : /*
8806 : * Treat a read that had both zeroed buffers *and* ignored checksums as a
8807 : * special case, it's too irregular to be emitted the same way as the
8808 : * other cases.
8809 : */
8810 272 : if (zeroed_any && ignored_any)
8811 : {
8812 : Assert(zeroed_any && ignored_any);
8813 : Assert(nblocks > 1); /* same block can't be both zeroed and ignored */
8814 : Assert(result.status != PGAIO_RS_ERROR);
8815 4 : affected_count = zeroed_or_error_count;
8816 :
8817 4 : ereport(elevel,
8818 : errcode(ERRCODE_DATA_CORRUPTED),
8819 : errmsg("zeroing %u page(s) and ignoring %u checksum failure(s) among blocks %u..%u of relation \"%s\"",
8820 : affected_count, checkfail_count, first, last, rpath.str),
8821 : affected_count > 1 ?
8822 : errdetail("Block %u held the first zeroed page.",
8823 : first + first_off) : 0,
8824 : errhint_plural("See server log for details about the other %d invalid block.",
8825 : "See server log for details about the other %d invalid blocks.",
8826 : affected_count + checkfail_count - 1,
8827 : affected_count + checkfail_count - 1));
8828 4 : return;
8829 : }
8830 :
8831 : /*
8832 : * The other messages are highly repetitive. To avoid duplicating a long
8833 : * and complicated ereport(), gather the translated format strings
8834 : * separately and then do one common ereport.
8835 : */
8836 268 : if (result.status == PGAIO_RS_ERROR)
8837 : {
8838 : Assert(!zeroed_any); /* can't have invalid pages when zeroing them */
8839 136 : affected_count = zeroed_or_error_count;
8840 136 : msg_one = _("invalid page in block %u of relation \"%s\"");
8841 136 : msg_mult = _("%u invalid pages among blocks %u..%u of relation \"%s\"");
8842 136 : det_mult = _("Block %u held the first invalid page.");
8843 136 : hint_mult = _("See server log for the other %u invalid block(s).");
8844 : }
8845 132 : else if (zeroed_any && !ignored_any)
8846 : {
8847 108 : affected_count = zeroed_or_error_count;
8848 108 : msg_one = _("invalid page in block %u of relation \"%s\"; zeroing out page");
8849 108 : msg_mult = _("zeroing out %u invalid pages among blocks %u..%u of relation \"%s\"");
8850 108 : det_mult = _("Block %u held the first zeroed page.");
8851 108 : hint_mult = _("See server log for the other %u zeroed block(s).");
8852 : }
8853 24 : else if (!zeroed_any && ignored_any)
8854 : {
8855 24 : affected_count = checkfail_count;
8856 24 : msg_one = _("ignoring checksum failure in block %u of relation \"%s\"");
8857 24 : msg_mult = _("ignoring %u checksum failures among blocks %u..%u of relation \"%s\"");
8858 24 : det_mult = _("Block %u held the first ignored page.");
8859 24 : hint_mult = _("See server log for the other %u ignored block(s).");
8860 : }
8861 : else
8862 0 : pg_unreachable();
8863 :
8864 268 : ereport(elevel,
8865 : errcode(ERRCODE_DATA_CORRUPTED),
8866 : affected_count == 1 ?
8867 : errmsg_internal(msg_one, first + first_off, rpath.str) :
8868 : errmsg_internal(msg_mult, affected_count, first, last, rpath.str),
8869 : affected_count > 1 ? errdetail_internal(det_mult, first + first_off) : 0,
8870 : affected_count > 1 ? errhint_internal(hint_mult, affected_count - 1) : 0);
8871 : }
8872 :
8873 : static void
8874 1474831 : shared_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
8875 : {
8876 1474831 : buffer_stage_common(ioh, false, false);
8877 1474831 : }
8878 :
8879 : static PgAioResult
8880 1339274 : shared_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
8881 : uint8 cb_data)
8882 : {
8883 1339274 : return buffer_readv_complete(ioh, prior_result, cb_data, false);
8884 : }
8885 :
8886 : /*
8887 : * We need a backend-local completion callback for shared buffers, to be able
8888 : * to report checksum errors correctly. Unfortunately that can only safely
8889 : * happen if the reporting backend has previously called
8890 : * pgstat_prepare_report_checksum_failure(), which we can only guarantee in
8891 : * the backend that started the IO. Hence this callback.
8892 : */
8893 : static PgAioResult
8894 1474831 : shared_buffer_readv_complete_local(PgAioHandle *ioh, PgAioResult prior_result,
8895 : uint8 cb_data)
8896 : {
8897 : bool zeroed_any,
8898 : ignored_any;
8899 : uint8 zeroed_or_error_count,
8900 : checkfail_count,
8901 : first_off;
8902 :
8903 1474831 : if (prior_result.status == PGAIO_RS_OK)
8904 1474653 : return prior_result;
8905 :
8906 178 : buffer_readv_decode_error(prior_result,
8907 : &zeroed_any,
8908 : &ignored_any,
8909 : &zeroed_or_error_count,
8910 : &checkfail_count,
8911 : &first_off);
8912 :
8913 178 : if (checkfail_count)
8914 : {
8915 24 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
8916 :
8917 24 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
8918 : checkfail_count);
8919 : }
8920 :
8921 178 : return prior_result;
8922 : }
8923 :
8924 : static void
8925 2836 : local_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
8926 : {
8927 2836 : buffer_stage_common(ioh, false, true);
8928 2836 : }
8929 :
8930 : static PgAioResult
8931 2836 : local_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
8932 : uint8 cb_data)
8933 : {
8934 2836 : return buffer_readv_complete(ioh, prior_result, cb_data, true);
8935 : }
8936 :
8937 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
8938 : const PgAioHandleCallbacks aio_shared_buffer_readv_cb = {
8939 : .stage = shared_buffer_readv_stage,
8940 : .complete_shared = shared_buffer_readv_complete,
8941 : /* need a local callback to report checksum failures */
8942 : .complete_local = shared_buffer_readv_complete_local,
8943 : .report = buffer_readv_report,
8944 : };
8945 :
8946 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
8947 : const PgAioHandleCallbacks aio_local_buffer_readv_cb = {
8948 : .stage = local_buffer_readv_stage,
8949 :
8950 : /*
8951 : * Note that this, in contrast to the shared_buffers case, uses
8952 : * complete_local, as only the issuing backend has access to the required
8953 : * datastructures. This is important in case the IO completion may be
8954 : * consumed incidentally by another backend.
8955 : */
8956 : .complete_local = local_buffer_readv_complete,
8957 : .report = buffer_readv_report,
8958 : };
|
