Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * bufmgr.c
4 : * buffer manager interface routines
5 : *
6 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/storage/buffer/bufmgr.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : /*
16 : * Principal entry points:
17 : *
18 : * ReadBuffer() -- find or create a buffer holding the requested page,
19 : * and pin it so that no one can destroy it while this process
20 : * is using it.
21 : *
22 : * StartReadBuffer() -- as above, with separate wait step
23 : * StartReadBuffers() -- multiple block version
24 : * WaitReadBuffers() -- second step of above
25 : *
26 : * ReleaseBuffer() -- unpin a buffer
27 : *
28 : * MarkBufferDirty() -- mark a pinned buffer's contents as "dirty".
29 : * The disk write is delayed until buffer replacement or checkpoint.
30 : *
31 : * See also these files:
32 : * freelist.c -- chooses victim for buffer replacement
33 : * buf_table.c -- manages the buffer lookup table
34 : */
35 : #include "postgres.h"
36 :
37 : #include <sys/file.h>
38 : #include <unistd.h>
39 :
40 : #include "access/tableam.h"
41 : #include "access/xloginsert.h"
42 : #include "access/xlogutils.h"
43 : #ifdef USE_ASSERT_CHECKING
44 : #include "catalog/pg_tablespace_d.h"
45 : #endif
46 : #include "catalog/storage.h"
47 : #include "catalog/storage_xlog.h"
48 : #include "executor/instrument.h"
49 : #include "lib/binaryheap.h"
50 : #include "miscadmin.h"
51 : #include "pg_trace.h"
52 : #include "pgstat.h"
53 : #include "postmaster/bgwriter.h"
54 : #include "storage/aio.h"
55 : #include "storage/buf_internals.h"
56 : #include "storage/bufmgr.h"
57 : #include "storage/fd.h"
58 : #include "storage/ipc.h"
59 : #include "storage/lmgr.h"
60 : #include "storage/proc.h"
61 : #include "storage/read_stream.h"
62 : #include "storage/smgr.h"
63 : #include "storage/standby.h"
64 : #include "utils/memdebug.h"
65 : #include "utils/ps_status.h"
66 : #include "utils/rel.h"
67 : #include "utils/resowner.h"
68 : #include "utils/timestamp.h"
69 :
70 :
71 : /* Note: these two macros only work on shared buffers, not local ones! */
72 : #define BufHdrGetBlock(bufHdr) ((Block) (BufferBlocks + ((Size) (bufHdr)->buf_id) * BLCKSZ))
73 : #define BufferGetLSN(bufHdr) (PageGetLSN(BufHdrGetBlock(bufHdr)))
74 :
75 : /* Note: this macro only works on local buffers, not shared ones! */
76 : #define LocalBufHdrGetBlock(bufHdr) \
77 : LocalBufferBlockPointers[-((bufHdr)->buf_id + 2)]
78 :
79 : /* Bits in SyncOneBuffer's return value */
80 : #define BUF_WRITTEN 0x01
81 : #define BUF_REUSABLE 0x02
82 :
83 : #define RELS_BSEARCH_THRESHOLD 20
84 :
85 : /*
86 : * This is the size (in the number of blocks) above which we scan the
87 : * entire buffer pool to remove the buffers for all the pages of relation
88 : * being dropped. For the relations with size below this threshold, we find
89 : * the buffers by doing lookups in BufMapping table.
90 : */
91 : #define BUF_DROP_FULL_SCAN_THRESHOLD (uint64) (NBuffers / 32)
92 :
93 : /*
94 : * This is separated out from PrivateRefCountEntry to allow for copying all
95 : * the data members via struct assignment.
96 : */
97 : typedef struct PrivateRefCountData
98 : {
99 : /*
100 : * How many times has the buffer been pinned by this backend.
101 : */
102 : int32 refcount;
103 : } PrivateRefCountData;
104 :
105 : typedef struct PrivateRefCountEntry
106 : {
107 : /*
108 : * Note that this needs to be same as the entry's corresponding
109 : * PrivateRefCountArrayKeys[i], if the entry is stored in the array. We
110 : * store it in both places as this is used for the hashtable key and
111 : * because it is more convenient (passing around a PrivateRefCountEntry
112 : * suffices to identify the buffer) and faster (checking the keys array is
113 : * faster when checking many entries, checking the entry is faster if just
114 : * checking a single entry).
115 : */
116 : Buffer buffer;
117 :
118 : PrivateRefCountData data;
119 : } PrivateRefCountEntry;
120 :
121 : /* 64 bytes, about the size of a cache line on common systems */
122 : #define REFCOUNT_ARRAY_ENTRIES 8
123 :
124 : /*
125 : * Status of buffers to checkpoint for a particular tablespace, used
126 : * internally in BufferSync.
127 : */
128 : typedef struct CkptTsStatus
129 : {
130 : /* oid of the tablespace */
131 : Oid tsId;
132 :
133 : /*
134 : * Checkpoint progress for this tablespace. To make progress comparable
135 : * between tablespaces the progress is, for each tablespace, measured as a
136 : * number between 0 and the total number of to-be-checkpointed pages. Each
137 : * page checkpointed in this tablespace increments this space's progress
138 : * by progress_slice.
139 : */
140 : float8 progress;
141 : float8 progress_slice;
142 :
143 : /* number of to-be checkpointed pages in this tablespace */
144 : int num_to_scan;
145 : /* already processed pages in this tablespace */
146 : int num_scanned;
147 :
148 : /* current offset in CkptBufferIds for this tablespace */
149 : int index;
150 : } CkptTsStatus;
151 :
152 : /*
153 : * Type for array used to sort SMgrRelations
154 : *
155 : * FlushRelationsAllBuffers shares the same comparator function with
156 : * DropRelationsAllBuffers. Pointer to this struct and RelFileLocator must be
157 : * compatible.
158 : */
159 : typedef struct SMgrSortArray
160 : {
161 : RelFileLocator rlocator; /* This must be the first member */
162 : SMgrRelation srel;
163 : } SMgrSortArray;
164 :
165 : /* GUC variables */
166 : bool zero_damaged_pages = false;
167 : int bgwriter_lru_maxpages = 100;
168 : double bgwriter_lru_multiplier = 2.0;
169 : bool track_io_timing = false;
170 :
171 : /*
172 : * How many buffers PrefetchBuffer callers should try to stay ahead of their
173 : * ReadBuffer calls by. Zero means "never prefetch". This value is only used
174 : * for buffers not belonging to tablespaces that have their
175 : * effective_io_concurrency parameter set.
176 : */
177 : int effective_io_concurrency = DEFAULT_EFFECTIVE_IO_CONCURRENCY;
178 :
179 : /*
180 : * Like effective_io_concurrency, but used by maintenance code paths that might
181 : * benefit from a higher setting because they work on behalf of many sessions.
182 : * Overridden by the tablespace setting of the same name.
183 : */
184 : int maintenance_io_concurrency = DEFAULT_MAINTENANCE_IO_CONCURRENCY;
185 :
186 : /*
187 : * Limit on how many blocks should be handled in single I/O operations.
188 : * StartReadBuffers() callers should respect it, as should other operations
189 : * that call smgr APIs directly. It is computed as the minimum of underlying
190 : * GUCs io_combine_limit_guc and io_max_combine_limit.
191 : */
192 : int io_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
193 : int io_combine_limit_guc = DEFAULT_IO_COMBINE_LIMIT;
194 : int io_max_combine_limit = DEFAULT_IO_COMBINE_LIMIT;
195 :
196 : /*
197 : * GUC variables about triggering kernel writeback for buffers written; OS
198 : * dependent defaults are set via the GUC mechanism.
199 : */
200 : int checkpoint_flush_after = DEFAULT_CHECKPOINT_FLUSH_AFTER;
201 : int bgwriter_flush_after = DEFAULT_BGWRITER_FLUSH_AFTER;
202 : int backend_flush_after = DEFAULT_BACKEND_FLUSH_AFTER;
203 :
204 : /* local state for LockBufferForCleanup */
205 : static BufferDesc *PinCountWaitBuf = NULL;
206 :
207 : /*
208 : * Backend-Private refcount management:
209 : *
210 : * Each buffer also has a private refcount that keeps track of the number of
211 : * times the buffer is pinned in the current process. This is so that the
212 : * shared refcount needs to be modified only once if a buffer is pinned more
213 : * than once by an individual backend. It's also used to check that no buffers
214 : * are still pinned at the end of transactions and when exiting.
215 : *
216 : *
217 : * To avoid - as we used to - requiring an array with NBuffers entries to keep
218 : * track of local buffers, we use a small sequentially searched array
219 : * (PrivateRefCountArrayKeys, with the corresponding data stored in
220 : * PrivateRefCountArray) and an overflow hash table (PrivateRefCountHash) to
221 : * keep track of backend local pins.
222 : *
223 : * Until no more than REFCOUNT_ARRAY_ENTRIES buffers are pinned at once, all
224 : * refcounts are kept track of in the array; after that, new array entries
225 : * displace old ones into the hash table. That way a frequently used entry
226 : * can't get "stuck" in the hashtable while infrequent ones clog the array.
227 : *
228 : * Note that in most scenarios the number of pinned buffers will not exceed
229 : * REFCOUNT_ARRAY_ENTRIES.
230 : *
231 : *
232 : * To enter a buffer into the refcount tracking mechanism first reserve a free
233 : * entry using ReservePrivateRefCountEntry() and then later, if necessary,
234 : * fill it with NewPrivateRefCountEntry(). That split lets us avoid doing
235 : * memory allocations in NewPrivateRefCountEntry() which can be important
236 : * because in some scenarios it's called with a spinlock held...
237 : */
238 : static Buffer PrivateRefCountArrayKeys[REFCOUNT_ARRAY_ENTRIES];
239 : static struct PrivateRefCountEntry PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES];
240 : static HTAB *PrivateRefCountHash = NULL;
241 : static int32 PrivateRefCountOverflowed = 0;
242 : static uint32 PrivateRefCountClock = 0;
243 : static int ReservedRefCountSlot = -1;
244 : static int PrivateRefCountEntryLast = -1;
245 :
246 : static uint32 MaxProportionalPins;
247 :
248 : static void ReservePrivateRefCountEntry(void);
249 : static PrivateRefCountEntry *NewPrivateRefCountEntry(Buffer buffer);
250 : static PrivateRefCountEntry *GetPrivateRefCountEntry(Buffer buffer, bool do_move);
251 : static inline int32 GetPrivateRefCount(Buffer buffer);
252 : static void ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref);
253 :
254 : /* ResourceOwner callbacks to hold in-progress I/Os and buffer pins */
255 : static void ResOwnerReleaseBufferIO(Datum res);
256 : static char *ResOwnerPrintBufferIO(Datum res);
257 : static void ResOwnerReleaseBufferPin(Datum res);
258 : static char *ResOwnerPrintBufferPin(Datum res);
259 :
260 : const ResourceOwnerDesc buffer_io_resowner_desc =
261 : {
262 : .name = "buffer io",
263 : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
264 : .release_priority = RELEASE_PRIO_BUFFER_IOS,
265 : .ReleaseResource = ResOwnerReleaseBufferIO,
266 : .DebugPrint = ResOwnerPrintBufferIO
267 : };
268 :
269 : const ResourceOwnerDesc buffer_pin_resowner_desc =
270 : {
271 : .name = "buffer pin",
272 : .release_phase = RESOURCE_RELEASE_BEFORE_LOCKS,
273 : .release_priority = RELEASE_PRIO_BUFFER_PINS,
274 : .ReleaseResource = ResOwnerReleaseBufferPin,
275 : .DebugPrint = ResOwnerPrintBufferPin
276 : };
277 :
278 : /*
279 : * Ensure that the PrivateRefCountArray has sufficient space to store one more
280 : * entry. This has to be called before using NewPrivateRefCountEntry() to fill
281 : * a new entry - but it's perfectly fine to not use a reserved entry.
282 : */
283 : static void
284 131222210 : ReservePrivateRefCountEntry(void)
285 : {
286 : /* Already reserved (or freed), nothing to do */
287 131222210 : if (ReservedRefCountSlot != -1)
288 122975710 : return;
289 :
290 : /*
291 : * First search for a free entry the array, that'll be sufficient in the
292 : * majority of cases.
293 : */
294 : {
295 : int i;
296 :
297 74218500 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
298 : {
299 65972000 : if (PrivateRefCountArrayKeys[i] == InvalidBuffer)
300 : {
301 49015008 : ReservedRefCountSlot = i;
302 :
303 : /*
304 : * We could return immediately, but iterating till the end of
305 : * the array allows compiler-autovectorization.
306 : */
307 : }
308 : }
309 :
310 8246500 : if (ReservedRefCountSlot != -1)
311 7908882 : return;
312 : }
313 :
314 : /*
315 : * No luck. All array entries are full. Move one array entry into the hash
316 : * table.
317 : */
318 : {
319 : /*
320 : * Move entry from the current clock position in the array into the
321 : * hashtable. Use that slot.
322 : */
323 : int victim_slot;
324 : PrivateRefCountEntry *victim_entry;
325 : PrivateRefCountEntry *hashent;
326 : bool found;
327 :
328 : /* select victim slot */
329 337618 : victim_slot = PrivateRefCountClock++ % REFCOUNT_ARRAY_ENTRIES;
330 337618 : victim_entry = &PrivateRefCountArray[victim_slot];
331 337618 : ReservedRefCountSlot = victim_slot;
332 :
333 : /* Better be used, otherwise we shouldn't get here. */
334 : Assert(PrivateRefCountArrayKeys[victim_slot] != InvalidBuffer);
335 : Assert(PrivateRefCountArray[victim_slot].buffer != InvalidBuffer);
336 : Assert(PrivateRefCountArrayKeys[victim_slot] == PrivateRefCountArray[victim_slot].buffer);
337 :
338 : /* enter victim array entry into hashtable */
339 337618 : hashent = hash_search(PrivateRefCountHash,
340 337618 : &PrivateRefCountArrayKeys[victim_slot],
341 : HASH_ENTER,
342 : &found);
343 : Assert(!found);
344 : /* move data from the entry in the array to the hash entry */
345 337618 : hashent->data = victim_entry->data;
346 :
347 : /* clear the now free array slot */
348 337618 : PrivateRefCountArrayKeys[victim_slot] = InvalidBuffer;
349 337618 : victim_entry->buffer = InvalidBuffer;
350 :
351 : /* clear the whole data member, just for future proofing */
352 337618 : memset(&victim_entry->data, 0, sizeof(victim_entry->data));
353 337618 : victim_entry->data.refcount = 0;
354 :
355 337618 : PrivateRefCountOverflowed++;
356 : }
357 : }
358 :
359 : /*
360 : * Fill a previously reserved refcount entry.
361 : */
362 : static PrivateRefCountEntry *
363 119126280 : NewPrivateRefCountEntry(Buffer buffer)
364 : {
365 : PrivateRefCountEntry *res;
366 :
367 : /* only allowed to be called when a reservation has been made */
368 : Assert(ReservedRefCountSlot != -1);
369 :
370 : /* use up the reserved entry */
371 119126280 : res = &PrivateRefCountArray[ReservedRefCountSlot];
372 :
373 : /* and fill it */
374 119126280 : PrivateRefCountArrayKeys[ReservedRefCountSlot] = buffer;
375 119126280 : res->buffer = buffer;
376 119126280 : res->data.refcount = 0;
377 :
378 : /* update cache for the next lookup */
379 119126280 : PrivateRefCountEntryLast = ReservedRefCountSlot;
380 :
381 119126280 : ReservedRefCountSlot = -1;
382 :
383 119126280 : return res;
384 : }
385 :
386 : /*
387 : * Slow-path for GetPrivateRefCountEntry(). This is big enough to not be worth
388 : * inlining. This particularly seems to be true if the compiler is capable of
389 : * auto-vectorizing the code, as that imposes additional stack-alignment
390 : * requirements etc.
391 : */
392 : static pg_noinline PrivateRefCountEntry *
393 139154306 : GetPrivateRefCountEntrySlow(Buffer buffer, bool do_move)
394 : {
395 : PrivateRefCountEntry *res;
396 139154306 : int match = -1;
397 : int i;
398 :
399 : /*
400 : * First search for references in the array, that'll be sufficient in the
401 : * majority of cases.
402 : */
403 1252388754 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
404 : {
405 1113234448 : if (PrivateRefCountArrayKeys[i] == buffer)
406 : {
407 24216866 : match = i;
408 : /* see ReservePrivateRefCountEntry() for why we don't return */
409 : }
410 : }
411 :
412 139154306 : if (likely(match != -1))
413 : {
414 : /* update cache for the next lookup */
415 24216866 : PrivateRefCountEntryLast = match;
416 :
417 24216866 : return &PrivateRefCountArray[match];
418 : }
419 :
420 : /*
421 : * By here we know that the buffer, if already pinned, isn't residing in
422 : * the array.
423 : *
424 : * Only look up the buffer in the hashtable if we've previously overflowed
425 : * into it.
426 : */
427 114937440 : if (PrivateRefCountOverflowed == 0)
428 114184446 : return NULL;
429 :
430 752994 : res = hash_search(PrivateRefCountHash, &buffer, HASH_FIND, NULL);
431 :
432 752994 : if (res == NULL)
433 411226 : return NULL;
434 341768 : else if (!do_move)
435 : {
436 : /* caller doesn't want us to move the hash entry into the array */
437 298992 : return res;
438 : }
439 : else
440 : {
441 : /* move buffer from hashtable into the free array slot */
442 : bool found;
443 : PrivateRefCountEntry *free;
444 :
445 : /* Ensure there's a free array slot */
446 42776 : ReservePrivateRefCountEntry();
447 :
448 : /* Use up the reserved slot */
449 : Assert(ReservedRefCountSlot != -1);
450 42776 : free = &PrivateRefCountArray[ReservedRefCountSlot];
451 : Assert(PrivateRefCountArrayKeys[ReservedRefCountSlot] == free->buffer);
452 : Assert(free->buffer == InvalidBuffer);
453 :
454 : /* and fill it */
455 42776 : free->buffer = buffer;
456 42776 : free->data = res->data;
457 42776 : PrivateRefCountArrayKeys[ReservedRefCountSlot] = buffer;
458 : /* update cache for the next lookup */
459 42776 : PrivateRefCountEntryLast = match;
460 :
461 42776 : ReservedRefCountSlot = -1;
462 :
463 :
464 : /* delete from hashtable */
465 42776 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
466 : Assert(found);
467 : Assert(PrivateRefCountOverflowed > 0);
468 42776 : PrivateRefCountOverflowed--;
469 :
470 42776 : return free;
471 : }
472 : }
473 :
474 : /*
475 : * Return the PrivateRefCount entry for the passed buffer.
476 : *
477 : * Returns NULL if a buffer doesn't have a refcount entry. Otherwise, if
478 : * do_move is true, and the entry resides in the hashtable the entry is
479 : * optimized for frequent access by moving it to the array.
480 : */
481 : static inline PrivateRefCountEntry *
482 294672228 : GetPrivateRefCountEntry(Buffer buffer, bool do_move)
483 : {
484 : Assert(BufferIsValid(buffer));
485 : Assert(!BufferIsLocal(buffer));
486 :
487 : /*
488 : * It's very common to look up the same buffer repeatedly. To make that
489 : * fast, we have a one-entry cache.
490 : *
491 : * In contrast to the loop in GetPrivateRefCountEntrySlow(), here it
492 : * faster to check PrivateRefCountArray[].buffer, as in the case of a hit
493 : * fewer addresses are computed and fewer cachelines are accessed. Whereas
494 : * in GetPrivateRefCountEntrySlow()'s case, checking
495 : * PrivateRefCountArrayKeys saves a lot of memory accesses.
496 : */
497 294672228 : if (likely(PrivateRefCountEntryLast != -1) &&
498 294599978 : likely(PrivateRefCountArray[PrivateRefCountEntryLast].buffer == buffer))
499 : {
500 155517922 : return &PrivateRefCountArray[PrivateRefCountEntryLast];
501 : }
502 :
503 : /*
504 : * The code for the cached lookup is small enough to be worth inlining
505 : * into the caller. In the miss case however, that empirically doesn't
506 : * seem worth it.
507 : */
508 139154306 : return GetPrivateRefCountEntrySlow(buffer, do_move);
509 : }
510 :
511 : /*
512 : * Returns how many times the passed buffer is pinned by this backend.
513 : *
514 : * Only works for shared memory buffers!
515 : */
516 : static inline int32
517 5611002 : GetPrivateRefCount(Buffer buffer)
518 : {
519 : PrivateRefCountEntry *ref;
520 :
521 : Assert(BufferIsValid(buffer));
522 : Assert(!BufferIsLocal(buffer));
523 :
524 : /*
525 : * Not moving the entry - that's ok for the current users, but we might
526 : * want to change this one day.
527 : */
528 5611002 : ref = GetPrivateRefCountEntry(buffer, false);
529 :
530 5611002 : if (ref == NULL)
531 54 : return 0;
532 5610948 : return ref->data.refcount;
533 : }
534 :
535 : /*
536 : * Release resources used to track the reference count of a buffer which we no
537 : * longer have pinned and don't want to pin again immediately.
538 : */
539 : static void
540 119126280 : ForgetPrivateRefCountEntry(PrivateRefCountEntry *ref)
541 : {
542 : Assert(ref->data.refcount == 0);
543 :
544 119126280 : if (ref >= &PrivateRefCountArray[0] &&
545 : ref < &PrivateRefCountArray[REFCOUNT_ARRAY_ENTRIES])
546 : {
547 118831438 : ref->buffer = InvalidBuffer;
548 118831438 : PrivateRefCountArrayKeys[ref - PrivateRefCountArray] = InvalidBuffer;
549 :
550 :
551 : /*
552 : * Mark the just used entry as reserved - in many scenarios that
553 : * allows us to avoid ever having to search the array/hash for free
554 : * entries.
555 : */
556 118831438 : ReservedRefCountSlot = ref - PrivateRefCountArray;
557 : }
558 : else
559 : {
560 : bool found;
561 294842 : Buffer buffer = ref->buffer;
562 :
563 294842 : hash_search(PrivateRefCountHash, &buffer, HASH_REMOVE, &found);
564 : Assert(found);
565 : Assert(PrivateRefCountOverflowed > 0);
566 294842 : PrivateRefCountOverflowed--;
567 : }
568 119126280 : }
569 :
570 : /*
571 : * BufferIsPinned
572 : * True iff the buffer is pinned (also checks for valid buffer number).
573 : *
574 : * NOTE: what we check here is that *this* backend holds a pin on
575 : * the buffer. We do not care whether some other backend does.
576 : */
577 : #define BufferIsPinned(bufnum) \
578 : ( \
579 : !BufferIsValid(bufnum) ? \
580 : false \
581 : : \
582 : BufferIsLocal(bufnum) ? \
583 : (LocalRefCount[-(bufnum) - 1] > 0) \
584 : : \
585 : (GetPrivateRefCount(bufnum) > 0) \
586 : )
587 :
588 :
589 : static Buffer ReadBuffer_common(Relation rel,
590 : SMgrRelation smgr, char smgr_persistence,
591 : ForkNumber forkNum, BlockNumber blockNum,
592 : ReadBufferMode mode, BufferAccessStrategy strategy);
593 : static BlockNumber ExtendBufferedRelCommon(BufferManagerRelation bmr,
594 : ForkNumber fork,
595 : BufferAccessStrategy strategy,
596 : uint32 flags,
597 : uint32 extend_by,
598 : BlockNumber extend_upto,
599 : Buffer *buffers,
600 : uint32 *extended_by);
601 : static BlockNumber ExtendBufferedRelShared(BufferManagerRelation bmr,
602 : ForkNumber fork,
603 : BufferAccessStrategy strategy,
604 : uint32 flags,
605 : uint32 extend_by,
606 : BlockNumber extend_upto,
607 : Buffer *buffers,
608 : uint32 *extended_by);
609 : static bool PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
610 : bool skip_if_not_valid);
611 : static void PinBuffer_Locked(BufferDesc *buf);
612 : static void UnpinBuffer(BufferDesc *buf);
613 : static void UnpinBufferNoOwner(BufferDesc *buf);
614 : static void BufferSync(int flags);
615 : static int SyncOneBuffer(int buf_id, bool skip_recently_used,
616 : WritebackContext *wb_context);
617 : static void WaitIO(BufferDesc *buf);
618 : static void AbortBufferIO(Buffer buffer);
619 : static void shared_buffer_write_error_callback(void *arg);
620 : static void local_buffer_write_error_callback(void *arg);
621 : static inline BufferDesc *BufferAlloc(SMgrRelation smgr,
622 : char relpersistence,
623 : ForkNumber forkNum,
624 : BlockNumber blockNum,
625 : BufferAccessStrategy strategy,
626 : bool *foundPtr, IOContext io_context);
627 : static bool AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress);
628 : static void CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete);
629 : static Buffer GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context);
630 : static void FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
631 : IOObject io_object, IOContext io_context);
632 : static void FlushBuffer(BufferDesc *buf, SMgrRelation reln,
633 : IOObject io_object, IOContext io_context);
634 : static void FindAndDropRelationBuffers(RelFileLocator rlocator,
635 : ForkNumber forkNum,
636 : BlockNumber nForkBlock,
637 : BlockNumber firstDelBlock);
638 : static void RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
639 : RelFileLocator dstlocator,
640 : ForkNumber forkNum, bool permanent);
641 : static void AtProcExit_Buffers(int code, Datum arg);
642 : static void CheckForBufferLeaks(void);
643 : #ifdef USE_ASSERT_CHECKING
644 : static void AssertNotCatalogBufferLock(LWLock *lock, LWLockMode mode,
645 : void *unused_context);
646 : #endif
647 : static int rlocator_comparator(const void *p1, const void *p2);
648 : static inline int buffertag_comparator(const BufferTag *ba, const BufferTag *bb);
649 : static inline int ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b);
650 : static int ts_ckpt_progress_comparator(Datum a, Datum b, void *arg);
651 :
652 :
653 : /*
654 : * Implementation of PrefetchBuffer() for shared buffers.
655 : */
656 : PrefetchBufferResult
657 65458 : PrefetchSharedBuffer(SMgrRelation smgr_reln,
658 : ForkNumber forkNum,
659 : BlockNumber blockNum)
660 : {
661 65458 : PrefetchBufferResult result = {InvalidBuffer, false};
662 : BufferTag newTag; /* identity of requested block */
663 : uint32 newHash; /* hash value for newTag */
664 : LWLock *newPartitionLock; /* buffer partition lock for it */
665 : int buf_id;
666 :
667 : Assert(BlockNumberIsValid(blockNum));
668 :
669 : /* create a tag so we can lookup the buffer */
670 65458 : InitBufferTag(&newTag, &smgr_reln->smgr_rlocator.locator,
671 : forkNum, blockNum);
672 :
673 : /* determine its hash code and partition lock ID */
674 65458 : newHash = BufTableHashCode(&newTag);
675 65458 : newPartitionLock = BufMappingPartitionLock(newHash);
676 :
677 : /* see if the block is in the buffer pool already */
678 65458 : LWLockAcquire(newPartitionLock, LW_SHARED);
679 65458 : buf_id = BufTableLookup(&newTag, newHash);
680 65458 : LWLockRelease(newPartitionLock);
681 :
682 : /* If not in buffers, initiate prefetch */
683 65458 : if (buf_id < 0)
684 : {
685 : #ifdef USE_PREFETCH
686 : /*
687 : * Try to initiate an asynchronous read. This returns false in
688 : * recovery if the relation file doesn't exist.
689 : */
690 34844 : if ((io_direct_flags & IO_DIRECT_DATA) == 0 &&
691 17198 : smgrprefetch(smgr_reln, forkNum, blockNum, 1))
692 : {
693 17198 : result.initiated_io = true;
694 : }
695 : #endif /* USE_PREFETCH */
696 : }
697 : else
698 : {
699 : /*
700 : * Report the buffer it was in at that time. The caller may be able
701 : * to avoid a buffer table lookup, but it's not pinned and it must be
702 : * rechecked!
703 : */
704 47812 : result.recent_buffer = buf_id + 1;
705 : }
706 :
707 : /*
708 : * If the block *is* in buffers, we do nothing. This is not really ideal:
709 : * the block might be just about to be evicted, which would be stupid
710 : * since we know we are going to need it soon. But the only easy answer
711 : * is to bump the usage_count, which does not seem like a great solution:
712 : * when the caller does ultimately touch the block, usage_count would get
713 : * bumped again, resulting in too much favoritism for blocks that are
714 : * involved in a prefetch sequence. A real fix would involve some
715 : * additional per-buffer state, and it's not clear that there's enough of
716 : * a problem to justify that.
717 : */
718 :
719 65458 : return result;
720 : }
721 :
722 : /*
723 : * PrefetchBuffer -- initiate asynchronous read of a block of a relation
724 : *
725 : * This is named by analogy to ReadBuffer but doesn't actually allocate a
726 : * buffer. Instead it tries to ensure that a future ReadBuffer for the given
727 : * block will not be delayed by the I/O. Prefetching is optional.
728 : *
729 : * There are three possible outcomes:
730 : *
731 : * 1. If the block is already cached, the result includes a valid buffer that
732 : * could be used by the caller to avoid the need for a later buffer lookup, but
733 : * it's not pinned, so the caller must recheck it.
734 : *
735 : * 2. If the kernel has been asked to initiate I/O, the initiated_io member is
736 : * true. Currently there is no way to know if the data was already cached by
737 : * the kernel and therefore didn't really initiate I/O, and no way to know when
738 : * the I/O completes other than using synchronous ReadBuffer().
739 : *
740 : * 3. Otherwise, the buffer wasn't already cached by PostgreSQL, and
741 : * USE_PREFETCH is not defined (this build doesn't support prefetching due to
742 : * lack of a kernel facility), direct I/O is enabled, or the underlying
743 : * relation file wasn't found and we are in recovery. (If the relation file
744 : * wasn't found and we are not in recovery, an error is raised).
745 : */
746 : PrefetchBufferResult
747 43884 : PrefetchBuffer(Relation reln, ForkNumber forkNum, BlockNumber blockNum)
748 : {
749 : Assert(RelationIsValid(reln));
750 : Assert(BlockNumberIsValid(blockNum));
751 :
752 43884 : if (RelationUsesLocalBuffers(reln))
753 : {
754 : /* see comments in ReadBufferExtended */
755 1566 : if (RELATION_IS_OTHER_TEMP(reln))
756 0 : ereport(ERROR,
757 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
758 : errmsg("cannot access temporary tables of other sessions")));
759 :
760 : /* pass it off to localbuf.c */
761 1566 : return PrefetchLocalBuffer(RelationGetSmgr(reln), forkNum, blockNum);
762 : }
763 : else
764 : {
765 : /* pass it to the shared buffer version */
766 42318 : return PrefetchSharedBuffer(RelationGetSmgr(reln), forkNum, blockNum);
767 : }
768 : }
769 :
770 : /*
771 : * ReadRecentBuffer -- try to pin a block in a recently observed buffer
772 : *
773 : * Compared to ReadBuffer(), this avoids a buffer mapping lookup when it's
774 : * successful. Return true if the buffer is valid and still has the expected
775 : * tag. In that case, the buffer is pinned and the usage count is bumped.
776 : */
777 : bool
778 9352 : ReadRecentBuffer(RelFileLocator rlocator, ForkNumber forkNum, BlockNumber blockNum,
779 : Buffer recent_buffer)
780 : {
781 : BufferDesc *bufHdr;
782 : BufferTag tag;
783 : uint32 buf_state;
784 :
785 : Assert(BufferIsValid(recent_buffer));
786 :
787 9352 : ResourceOwnerEnlarge(CurrentResourceOwner);
788 9352 : ReservePrivateRefCountEntry();
789 9352 : InitBufferTag(&tag, &rlocator, forkNum, blockNum);
790 :
791 9352 : if (BufferIsLocal(recent_buffer))
792 : {
793 64 : int b = -recent_buffer - 1;
794 :
795 64 : bufHdr = GetLocalBufferDescriptor(b);
796 64 : buf_state = pg_atomic_read_u32(&bufHdr->state);
797 :
798 : /* Is it still valid and holding the right tag? */
799 64 : if ((buf_state & BM_VALID) && BufferTagsEqual(&tag, &bufHdr->tag))
800 : {
801 64 : PinLocalBuffer(bufHdr, true);
802 :
803 64 : pgBufferUsage.local_blks_hit++;
804 :
805 64 : return true;
806 : }
807 : }
808 : else
809 : {
810 9288 : bufHdr = GetBufferDescriptor(recent_buffer - 1);
811 :
812 : /*
813 : * Is it still valid and holding the right tag? We do an unlocked tag
814 : * comparison first, to make it unlikely that we'll increment the
815 : * usage counter of the wrong buffer, if someone calls us with a very
816 : * out of date recent_buffer. Then we'll check it again if we get the
817 : * pin.
818 : */
819 18502 : if (BufferTagsEqual(&tag, &bufHdr->tag) &&
820 9214 : PinBuffer(bufHdr, NULL, true))
821 : {
822 9202 : if (BufferTagsEqual(&tag, &bufHdr->tag))
823 : {
824 9202 : pgBufferUsage.shared_blks_hit++;
825 9202 : return true;
826 : }
827 0 : UnpinBuffer(bufHdr);
828 : }
829 : }
830 :
831 86 : return false;
832 : }
833 :
834 : /*
835 : * ReadBuffer -- a shorthand for ReadBufferExtended, for reading from main
836 : * fork with RBM_NORMAL mode and default strategy.
837 : */
838 : Buffer
839 88041432 : ReadBuffer(Relation reln, BlockNumber blockNum)
840 : {
841 88041432 : return ReadBufferExtended(reln, MAIN_FORKNUM, blockNum, RBM_NORMAL, NULL);
842 : }
843 :
844 : /*
845 : * ReadBufferExtended -- returns a buffer containing the requested
846 : * block of the requested relation. If the blknum
847 : * requested is P_NEW, extend the relation file and
848 : * allocate a new block. (Caller is responsible for
849 : * ensuring that only one backend tries to extend a
850 : * relation at the same time!)
851 : *
852 : * Returns: the buffer number for the buffer containing
853 : * the block read. The returned buffer has been pinned.
854 : * Does not return on error --- elog's instead.
855 : *
856 : * Assume when this function is called, that reln has been opened already.
857 : *
858 : * In RBM_NORMAL mode, the page is read from disk, and the page header is
859 : * validated. An error is thrown if the page header is not valid. (But
860 : * note that an all-zero page is considered "valid"; see
861 : * PageIsVerified().)
862 : *
863 : * RBM_ZERO_ON_ERROR is like the normal mode, but if the page header is not
864 : * valid, the page is zeroed instead of throwing an error. This is intended
865 : * for non-critical data, where the caller is prepared to repair errors.
866 : *
867 : * In RBM_ZERO_AND_LOCK mode, if the page isn't in buffer cache already, it's
868 : * filled with zeros instead of reading it from disk. Useful when the caller
869 : * is going to fill the page from scratch, since this saves I/O and avoids
870 : * unnecessary failure if the page-on-disk has corrupt page headers.
871 : * The page is returned locked to ensure that the caller has a chance to
872 : * initialize the page before it's made visible to others.
873 : * Caution: do not use this mode to read a page that is beyond the relation's
874 : * current physical EOF; that is likely to cause problems in md.c when
875 : * the page is modified and written out. P_NEW is OK, though.
876 : *
877 : * RBM_ZERO_AND_CLEANUP_LOCK is the same as RBM_ZERO_AND_LOCK, but acquires
878 : * a cleanup-strength lock on the page.
879 : *
880 : * RBM_NORMAL_NO_LOG mode is treated the same as RBM_NORMAL here.
881 : *
882 : * If strategy is not NULL, a nondefault buffer access strategy is used.
883 : * See buffer/README for details.
884 : */
885 : inline Buffer
886 105671692 : ReadBufferExtended(Relation reln, ForkNumber forkNum, BlockNumber blockNum,
887 : ReadBufferMode mode, BufferAccessStrategy strategy)
888 : {
889 : Buffer buf;
890 :
891 : /*
892 : * Reject attempts to read non-local temporary relations; we would be
893 : * likely to get wrong data since we have no visibility into the owning
894 : * session's local buffers.
895 : */
896 105671692 : if (RELATION_IS_OTHER_TEMP(reln))
897 0 : ereport(ERROR,
898 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
899 : errmsg("cannot access temporary tables of other sessions")));
900 :
901 : /*
902 : * Read the buffer, and update pgstat counters to reflect a cache hit or
903 : * miss.
904 : */
905 105671692 : buf = ReadBuffer_common(reln, RelationGetSmgr(reln), 0,
906 : forkNum, blockNum, mode, strategy);
907 :
908 105671646 : return buf;
909 : }
910 :
911 :
912 : /*
913 : * ReadBufferWithoutRelcache -- like ReadBufferExtended, but doesn't require
914 : * a relcache entry for the relation.
915 : *
916 : * Pass permanent = true for a RELPERSISTENCE_PERMANENT relation, and
917 : * permanent = false for a RELPERSISTENCE_UNLOGGED relation. This function
918 : * cannot be used for temporary relations (and making that work might be
919 : * difficult, unless we only want to read temporary relations for our own
920 : * ProcNumber).
921 : */
922 : Buffer
923 11519628 : ReadBufferWithoutRelcache(RelFileLocator rlocator, ForkNumber forkNum,
924 : BlockNumber blockNum, ReadBufferMode mode,
925 : BufferAccessStrategy strategy, bool permanent)
926 : {
927 11519628 : SMgrRelation smgr = smgropen(rlocator, INVALID_PROC_NUMBER);
928 :
929 11519628 : return ReadBuffer_common(NULL, smgr,
930 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
931 : forkNum, blockNum,
932 : mode, strategy);
933 : }
934 :
935 : /*
936 : * Convenience wrapper around ExtendBufferedRelBy() extending by one block.
937 : */
938 : Buffer
939 91622 : ExtendBufferedRel(BufferManagerRelation bmr,
940 : ForkNumber forkNum,
941 : BufferAccessStrategy strategy,
942 : uint32 flags)
943 : {
944 : Buffer buf;
945 91622 : uint32 extend_by = 1;
946 :
947 91622 : ExtendBufferedRelBy(bmr, forkNum, strategy, flags, extend_by,
948 : &buf, &extend_by);
949 :
950 91622 : return buf;
951 : }
952 :
953 : /*
954 : * Extend relation by multiple blocks.
955 : *
956 : * Tries to extend the relation by extend_by blocks. Depending on the
957 : * availability of resources the relation may end up being extended by a
958 : * smaller number of pages (unless an error is thrown, always by at least one
959 : * page). *extended_by is updated to the number of pages the relation has been
960 : * extended to.
961 : *
962 : * buffers needs to be an array that is at least extend_by long. Upon
963 : * completion, the first extend_by array elements will point to a pinned
964 : * buffer.
965 : *
966 : * If EB_LOCK_FIRST is part of flags, the first returned buffer is
967 : * locked. This is useful for callers that want a buffer that is guaranteed to
968 : * be empty.
969 : */
970 : BlockNumber
971 320486 : ExtendBufferedRelBy(BufferManagerRelation bmr,
972 : ForkNumber fork,
973 : BufferAccessStrategy strategy,
974 : uint32 flags,
975 : uint32 extend_by,
976 : Buffer *buffers,
977 : uint32 *extended_by)
978 : {
979 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
980 : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
981 : Assert(extend_by > 0);
982 :
983 320486 : if (bmr.relpersistence == '\0')
984 320486 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
985 :
986 320486 : return ExtendBufferedRelCommon(bmr, fork, strategy, flags,
987 : extend_by, InvalidBlockNumber,
988 : buffers, extended_by);
989 : }
990 :
991 : /*
992 : * Extend the relation so it is at least extend_to blocks large, return buffer
993 : * (extend_to - 1).
994 : *
995 : * This is useful for callers that want to write a specific page, regardless
996 : * of the current size of the relation (e.g. useful for visibilitymap and for
997 : * crash recovery).
998 : */
999 : Buffer
1000 103688 : ExtendBufferedRelTo(BufferManagerRelation bmr,
1001 : ForkNumber fork,
1002 : BufferAccessStrategy strategy,
1003 : uint32 flags,
1004 : BlockNumber extend_to,
1005 : ReadBufferMode mode)
1006 : {
1007 : BlockNumber current_size;
1008 103688 : uint32 extended_by = 0;
1009 103688 : Buffer buffer = InvalidBuffer;
1010 : Buffer buffers[64];
1011 :
1012 : Assert((bmr.rel != NULL) != (bmr.smgr != NULL));
1013 : Assert(bmr.smgr == NULL || bmr.relpersistence != '\0');
1014 : Assert(extend_to != InvalidBlockNumber && extend_to > 0);
1015 :
1016 103688 : if (bmr.relpersistence == '\0')
1017 14416 : bmr.relpersistence = bmr.rel->rd_rel->relpersistence;
1018 :
1019 : /*
1020 : * If desired, create the file if it doesn't exist. If
1021 : * smgr_cached_nblocks[fork] is positive then it must exist, no need for
1022 : * an smgrexists call.
1023 : */
1024 103688 : if ((flags & EB_CREATE_FORK_IF_NEEDED) &&
1025 14416 : (BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == 0 ||
1026 48 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] == InvalidBlockNumber) &&
1027 14368 : !smgrexists(BMR_GET_SMGR(bmr), fork))
1028 : {
1029 14322 : LockRelationForExtension(bmr.rel, ExclusiveLock);
1030 :
1031 : /* recheck, fork might have been created concurrently */
1032 14322 : if (!smgrexists(BMR_GET_SMGR(bmr), fork))
1033 14316 : smgrcreate(BMR_GET_SMGR(bmr), fork, flags & EB_PERFORMING_RECOVERY);
1034 :
1035 14322 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
1036 : }
1037 :
1038 : /*
1039 : * If requested, invalidate size cache, so that smgrnblocks asks the
1040 : * kernel.
1041 : */
1042 103688 : if (flags & EB_CLEAR_SIZE_CACHE)
1043 14416 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
1044 :
1045 : /*
1046 : * Estimate how many pages we'll need to extend by. This avoids acquiring
1047 : * unnecessarily many victim buffers.
1048 : */
1049 103688 : current_size = smgrnblocks(BMR_GET_SMGR(bmr), fork);
1050 :
1051 : /*
1052 : * Since no-one else can be looking at the page contents yet, there is no
1053 : * difference between an exclusive lock and a cleanup-strength lock. Note
1054 : * that we pass the original mode to ReadBuffer_common() below, when
1055 : * falling back to reading the buffer to a concurrent relation extension.
1056 : */
1057 103688 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1058 88528 : flags |= EB_LOCK_TARGET;
1059 :
1060 211402 : while (current_size < extend_to)
1061 : {
1062 107714 : uint32 num_pages = lengthof(buffers);
1063 : BlockNumber first_block;
1064 :
1065 107714 : if ((uint64) current_size + num_pages > extend_to)
1066 107712 : num_pages = extend_to - current_size;
1067 :
1068 107714 : first_block = ExtendBufferedRelCommon(bmr, fork, strategy, flags,
1069 : num_pages, extend_to,
1070 : buffers, &extended_by);
1071 :
1072 107714 : current_size = first_block + extended_by;
1073 : Assert(num_pages != 0 || current_size >= extend_to);
1074 :
1075 230126 : for (uint32 i = 0; i < extended_by; i++)
1076 : {
1077 122412 : if (first_block + i != extend_to - 1)
1078 18750 : ReleaseBuffer(buffers[i]);
1079 : else
1080 103662 : buffer = buffers[i];
1081 : }
1082 : }
1083 :
1084 : /*
1085 : * It's possible that another backend concurrently extended the relation.
1086 : * In that case read the buffer.
1087 : *
1088 : * XXX: Should we control this via a flag?
1089 : */
1090 103688 : if (buffer == InvalidBuffer)
1091 : {
1092 : Assert(extended_by == 0);
1093 26 : buffer = ReadBuffer_common(bmr.rel, BMR_GET_SMGR(bmr), bmr.relpersistence,
1094 : fork, extend_to - 1, mode, strategy);
1095 : }
1096 :
1097 103688 : return buffer;
1098 : }
1099 :
1100 : /*
1101 : * Lock and optionally zero a buffer, as part of the implementation of
1102 : * RBM_ZERO_AND_LOCK or RBM_ZERO_AND_CLEANUP_LOCK. The buffer must be already
1103 : * pinned. If the buffer is not already valid, it is zeroed and made valid.
1104 : */
1105 : static void
1106 642178 : ZeroAndLockBuffer(Buffer buffer, ReadBufferMode mode, bool already_valid)
1107 : {
1108 : BufferDesc *bufHdr;
1109 : bool need_to_zero;
1110 642178 : bool isLocalBuf = BufferIsLocal(buffer);
1111 :
1112 : Assert(mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK);
1113 :
1114 642178 : if (already_valid)
1115 : {
1116 : /*
1117 : * If the caller already knew the buffer was valid, we can skip some
1118 : * header interaction. The caller just wants to lock the buffer.
1119 : */
1120 74638 : need_to_zero = false;
1121 : }
1122 567540 : else if (isLocalBuf)
1123 : {
1124 : /* Simple case for non-shared buffers. */
1125 48 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
1126 48 : need_to_zero = StartLocalBufferIO(bufHdr, true, false);
1127 : }
1128 : else
1129 : {
1130 : /*
1131 : * Take BM_IO_IN_PROGRESS, or discover that BM_VALID has been set
1132 : * concurrently. Even though we aren't doing I/O, that ensures that
1133 : * we don't zero a page that someone else has pinned. An exclusive
1134 : * content lock wouldn't be enough, because readers are allowed to
1135 : * drop the content lock after determining that a tuple is visible
1136 : * (see buffer access rules in README).
1137 : */
1138 567492 : bufHdr = GetBufferDescriptor(buffer - 1);
1139 567492 : need_to_zero = StartBufferIO(bufHdr, true, false);
1140 : }
1141 :
1142 642178 : if (need_to_zero)
1143 : {
1144 567540 : memset(BufferGetPage(buffer), 0, BLCKSZ);
1145 :
1146 : /*
1147 : * Grab the buffer content lock before marking the page as valid, to
1148 : * make sure that no other backend sees the zeroed page before the
1149 : * caller has had a chance to initialize it.
1150 : *
1151 : * Since no-one else can be looking at the page contents yet, there is
1152 : * no difference between an exclusive lock and a cleanup-strength
1153 : * lock. (Note that we cannot use LockBuffer() or
1154 : * LockBufferForCleanup() here, because they assert that the buffer is
1155 : * already valid.)
1156 : */
1157 567540 : if (!isLocalBuf)
1158 567492 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1159 :
1160 : /* Set BM_VALID, terminate IO, and wake up any waiters */
1161 567540 : if (isLocalBuf)
1162 48 : TerminateLocalBufferIO(bufHdr, false, BM_VALID, false);
1163 : else
1164 567492 : TerminateBufferIO(bufHdr, false, BM_VALID, true, false);
1165 : }
1166 74638 : else if (!isLocalBuf)
1167 : {
1168 : /*
1169 : * The buffer is valid, so we can't zero it. The caller still expects
1170 : * the page to be locked on return.
1171 : */
1172 74598 : if (mode == RBM_ZERO_AND_LOCK)
1173 74492 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
1174 : else
1175 106 : LockBufferForCleanup(buffer);
1176 : }
1177 642178 : }
1178 :
1179 : /*
1180 : * Pin a buffer for a given block. *foundPtr is set to true if the block was
1181 : * already present, or false if more work is required to either read it in or
1182 : * zero it.
1183 : */
1184 : static pg_attribute_always_inline Buffer
1185 125130836 : PinBufferForBlock(Relation rel,
1186 : SMgrRelation smgr,
1187 : char persistence,
1188 : ForkNumber forkNum,
1189 : BlockNumber blockNum,
1190 : BufferAccessStrategy strategy,
1191 : bool *foundPtr)
1192 : {
1193 : BufferDesc *bufHdr;
1194 : IOContext io_context;
1195 : IOObject io_object;
1196 :
1197 : Assert(blockNum != P_NEW);
1198 :
1199 : /* Persistence should be set before */
1200 : Assert((persistence == RELPERSISTENCE_TEMP ||
1201 : persistence == RELPERSISTENCE_PERMANENT ||
1202 : persistence == RELPERSISTENCE_UNLOGGED));
1203 :
1204 125130836 : if (persistence == RELPERSISTENCE_TEMP)
1205 : {
1206 2554504 : io_context = IOCONTEXT_NORMAL;
1207 2554504 : io_object = IOOBJECT_TEMP_RELATION;
1208 : }
1209 : else
1210 : {
1211 122576332 : io_context = IOContextForStrategy(strategy);
1212 122576332 : io_object = IOOBJECT_RELATION;
1213 : }
1214 :
1215 : TRACE_POSTGRESQL_BUFFER_READ_START(forkNum, blockNum,
1216 : smgr->smgr_rlocator.locator.spcOid,
1217 : smgr->smgr_rlocator.locator.dbOid,
1218 : smgr->smgr_rlocator.locator.relNumber,
1219 : smgr->smgr_rlocator.backend);
1220 :
1221 125130836 : if (persistence == RELPERSISTENCE_TEMP)
1222 : {
1223 2554504 : bufHdr = LocalBufferAlloc(smgr, forkNum, blockNum, foundPtr);
1224 2554492 : if (*foundPtr)
1225 2537710 : pgBufferUsage.local_blks_hit++;
1226 : }
1227 : else
1228 : {
1229 122576332 : bufHdr = BufferAlloc(smgr, persistence, forkNum, blockNum,
1230 : strategy, foundPtr, io_context);
1231 122576332 : if (*foundPtr)
1232 119106110 : pgBufferUsage.shared_blks_hit++;
1233 : }
1234 125130824 : if (rel)
1235 : {
1236 : /*
1237 : * While pgBufferUsage's "read" counter isn't bumped unless we reach
1238 : * WaitReadBuffers() (so, not for hits, and not for buffers that are
1239 : * zeroed instead), the per-relation stats always count them.
1240 : */
1241 113138062 : pgstat_count_buffer_read(rel);
1242 113138062 : if (*foundPtr)
1243 110590618 : pgstat_count_buffer_hit(rel);
1244 : }
1245 125130824 : if (*foundPtr)
1246 : {
1247 121643820 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1248 121643820 : if (VacuumCostActive)
1249 4746842 : VacuumCostBalance += VacuumCostPageHit;
1250 :
1251 : TRACE_POSTGRESQL_BUFFER_READ_DONE(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 : true);
1257 : }
1258 :
1259 125130824 : return BufferDescriptorGetBuffer(bufHdr);
1260 : }
1261 :
1262 : /*
1263 : * ReadBuffer_common -- common logic for all ReadBuffer variants
1264 : *
1265 : * smgr is required, rel is optional unless using P_NEW.
1266 : */
1267 : static pg_attribute_always_inline Buffer
1268 117191346 : ReadBuffer_common(Relation rel, SMgrRelation smgr, char smgr_persistence,
1269 : ForkNumber forkNum,
1270 : BlockNumber blockNum, ReadBufferMode mode,
1271 : BufferAccessStrategy strategy)
1272 : {
1273 : ReadBuffersOperation operation;
1274 : Buffer buffer;
1275 : int flags;
1276 : char persistence;
1277 :
1278 : /*
1279 : * Backward compatibility path, most code should use ExtendBufferedRel()
1280 : * instead, as acquiring the extension lock inside ExtendBufferedRel()
1281 : * scales a lot better.
1282 : */
1283 117191346 : if (unlikely(blockNum == P_NEW))
1284 : {
1285 522 : uint32 flags = EB_SKIP_EXTENSION_LOCK;
1286 :
1287 : /*
1288 : * Since no-one else can be looking at the page contents yet, there is
1289 : * no difference between an exclusive lock and a cleanup-strength
1290 : * lock.
1291 : */
1292 522 : if (mode == RBM_ZERO_AND_LOCK || mode == RBM_ZERO_AND_CLEANUP_LOCK)
1293 0 : flags |= EB_LOCK_FIRST;
1294 :
1295 522 : return ExtendBufferedRel(BMR_REL(rel), forkNum, strategy, flags);
1296 : }
1297 :
1298 117190824 : if (rel)
1299 105671196 : persistence = rel->rd_rel->relpersistence;
1300 : else
1301 11519628 : persistence = smgr_persistence;
1302 :
1303 117190824 : if (unlikely(mode == RBM_ZERO_AND_CLEANUP_LOCK ||
1304 : mode == RBM_ZERO_AND_LOCK))
1305 : {
1306 : bool found;
1307 :
1308 642178 : buffer = PinBufferForBlock(rel, smgr, persistence,
1309 : forkNum, blockNum, strategy, &found);
1310 642178 : ZeroAndLockBuffer(buffer, mode, found);
1311 642178 : return buffer;
1312 : }
1313 :
1314 : /*
1315 : * Signal that we are going to immediately wait. If we're immediately
1316 : * waiting, there is no benefit in actually executing the IO
1317 : * asynchronously, it would just add dispatch overhead.
1318 : */
1319 116548646 : flags = READ_BUFFERS_SYNCHRONOUSLY;
1320 116548646 : if (mode == RBM_ZERO_ON_ERROR)
1321 2629072 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1322 116548646 : operation.smgr = smgr;
1323 116548646 : operation.rel = rel;
1324 116548646 : operation.persistence = persistence;
1325 116548646 : operation.forknum = forkNum;
1326 116548646 : operation.strategy = strategy;
1327 116548646 : if (StartReadBuffer(&operation,
1328 : &buffer,
1329 : blockNum,
1330 : flags))
1331 1459994 : WaitReadBuffers(&operation);
1332 :
1333 116548600 : return buffer;
1334 : }
1335 :
1336 : static pg_attribute_always_inline bool
1337 124148594 : StartReadBuffersImpl(ReadBuffersOperation *operation,
1338 : Buffer *buffers,
1339 : BlockNumber blockNum,
1340 : int *nblocks,
1341 : int flags,
1342 : bool allow_forwarding)
1343 : {
1344 124148594 : int actual_nblocks = *nblocks;
1345 124148594 : int maxcombine = 0;
1346 : bool did_start_io;
1347 :
1348 : Assert(*nblocks == 1 || allow_forwarding);
1349 : Assert(*nblocks > 0);
1350 : Assert(*nblocks <= MAX_IO_COMBINE_LIMIT);
1351 :
1352 127068058 : for (int i = 0; i < actual_nblocks; ++i)
1353 : {
1354 : bool found;
1355 :
1356 124491892 : if (allow_forwarding && buffers[i] != InvalidBuffer)
1357 3234 : {
1358 : BufferDesc *bufHdr;
1359 :
1360 : /*
1361 : * This is a buffer that was pinned by an earlier call to
1362 : * StartReadBuffers(), but couldn't be handled in one operation at
1363 : * that time. The operation was split, and the caller has passed
1364 : * an already pinned buffer back to us to handle the rest of the
1365 : * operation. It must continue at the expected block number.
1366 : */
1367 : Assert(BufferGetBlockNumber(buffers[i]) == blockNum + i);
1368 :
1369 : /*
1370 : * It might be an already valid buffer (a hit) that followed the
1371 : * final contiguous block of an earlier I/O (a miss) marking the
1372 : * end of it, or a buffer that some other backend has since made
1373 : * valid by performing the I/O for us, in which case we can handle
1374 : * it as a hit now. It is safe to check for a BM_VALID flag with
1375 : * a relaxed load, because we got a fresh view of it while pinning
1376 : * it in the previous call.
1377 : *
1378 : * On the other hand if we don't see BM_VALID yet, it must be an
1379 : * I/O that was split by the previous call and we need to try to
1380 : * start a new I/O from this block. We're also racing against any
1381 : * other backend that might start the I/O or even manage to mark
1382 : * it BM_VALID after this check, but StartBufferIO() will handle
1383 : * those cases.
1384 : */
1385 3234 : if (BufferIsLocal(buffers[i]))
1386 4 : bufHdr = GetLocalBufferDescriptor(-buffers[i] - 1);
1387 : else
1388 3230 : bufHdr = GetBufferDescriptor(buffers[i] - 1);
1389 : Assert(pg_atomic_read_u32(&bufHdr->state) & BM_TAG_VALID);
1390 3234 : found = pg_atomic_read_u32(&bufHdr->state) & BM_VALID;
1391 : }
1392 : else
1393 : {
1394 124488646 : buffers[i] = PinBufferForBlock(operation->rel,
1395 : operation->smgr,
1396 124488658 : operation->persistence,
1397 : operation->forknum,
1398 : blockNum + i,
1399 : operation->strategy,
1400 : &found);
1401 : }
1402 :
1403 124491880 : if (found)
1404 : {
1405 : /*
1406 : * We have a hit. If it's the first block in the requested range,
1407 : * we can return it immediately and report that WaitReadBuffers()
1408 : * does not need to be called. If the initial value of *nblocks
1409 : * was larger, the caller will have to call again for the rest.
1410 : */
1411 121572416 : if (i == 0)
1412 : {
1413 121569178 : *nblocks = 1;
1414 :
1415 : #ifdef USE_ASSERT_CHECKING
1416 :
1417 : /*
1418 : * Initialize enough of ReadBuffersOperation to make
1419 : * CheckReadBuffersOperation() work. Outside of assertions
1420 : * that's not necessary when no IO is issued.
1421 : */
1422 : operation->buffers = buffers;
1423 : operation->blocknum = blockNum;
1424 : operation->nblocks = 1;
1425 : operation->nblocks_done = 1;
1426 : CheckReadBuffersOperation(operation, true);
1427 : #endif
1428 121569178 : return false;
1429 : }
1430 :
1431 : /*
1432 : * Otherwise we already have an I/O to perform, but this block
1433 : * can't be included as it is already valid. Split the I/O here.
1434 : * There may or may not be more blocks requiring I/O after this
1435 : * one, we haven't checked, but they can't be contiguous with this
1436 : * one in the way. We'll leave this buffer pinned, forwarding it
1437 : * to the next call, avoiding the need to unpin it here and re-pin
1438 : * it in the next call.
1439 : */
1440 3238 : actual_nblocks = i;
1441 3238 : break;
1442 : }
1443 : else
1444 : {
1445 : /*
1446 : * Check how many blocks we can cover with the same IO. The smgr
1447 : * implementation might e.g. be limited due to a segment boundary.
1448 : */
1449 2919464 : if (i == 0 && actual_nblocks > 1)
1450 : {
1451 69490 : maxcombine = smgrmaxcombine(operation->smgr,
1452 : operation->forknum,
1453 : blockNum);
1454 69490 : if (unlikely(maxcombine < actual_nblocks))
1455 : {
1456 0 : elog(DEBUG2, "limiting nblocks at %u from %u to %u",
1457 : blockNum, actual_nblocks, maxcombine);
1458 0 : actual_nblocks = maxcombine;
1459 : }
1460 : }
1461 : }
1462 : }
1463 2579404 : *nblocks = actual_nblocks;
1464 :
1465 : /* Populate information needed for I/O. */
1466 2579404 : operation->buffers = buffers;
1467 2579404 : operation->blocknum = blockNum;
1468 2579404 : operation->flags = flags;
1469 2579404 : operation->nblocks = actual_nblocks;
1470 2579404 : operation->nblocks_done = 0;
1471 2579404 : pgaio_wref_clear(&operation->io_wref);
1472 :
1473 : /*
1474 : * When using AIO, start the IO in the background. If not, issue prefetch
1475 : * requests if desired by the caller.
1476 : *
1477 : * The reason we have a dedicated path for IOMETHOD_SYNC here is to
1478 : * de-risk the introduction of AIO somewhat. It's a large architectural
1479 : * change, with lots of chances for unanticipated performance effects.
1480 : *
1481 : * Use of IOMETHOD_SYNC already leads to not actually performing IO
1482 : * asynchronously, but without the check here we'd execute IO earlier than
1483 : * we used to. Eventually this IOMETHOD_SYNC specific path should go away.
1484 : */
1485 2579404 : if (io_method != IOMETHOD_SYNC)
1486 : {
1487 : /*
1488 : * Try to start IO asynchronously. It's possible that no IO needs to
1489 : * be started, if another backend already performed the IO.
1490 : *
1491 : * Note that if an IO is started, it might not cover the entire
1492 : * requested range, e.g. because an intermediary block has been read
1493 : * in by another backend. In that case any "trailing" buffers we
1494 : * already pinned above will be "forwarded" by read_stream.c to the
1495 : * next call to StartReadBuffers().
1496 : *
1497 : * This is signalled to the caller by decrementing *nblocks *and*
1498 : * reducing operation->nblocks. The latter is done here, but not below
1499 : * WaitReadBuffers(), as in WaitReadBuffers() we can't "shorten" the
1500 : * overall read size anymore, we need to retry until done in its
1501 : * entirety or until failed.
1502 : */
1503 2577252 : did_start_io = AsyncReadBuffers(operation, nblocks);
1504 :
1505 2577222 : operation->nblocks = *nblocks;
1506 : }
1507 : else
1508 : {
1509 2152 : operation->flags |= READ_BUFFERS_SYNCHRONOUSLY;
1510 :
1511 2152 : if (flags & READ_BUFFERS_ISSUE_ADVICE)
1512 : {
1513 : /*
1514 : * In theory we should only do this if PinBufferForBlock() had to
1515 : * allocate new buffers above. That way, if two calls to
1516 : * StartReadBuffers() were made for the same blocks before
1517 : * WaitReadBuffers(), only the first would issue the advice.
1518 : * That'd be a better simulation of true asynchronous I/O, which
1519 : * would only start the I/O once, but isn't done here for
1520 : * simplicity.
1521 : */
1522 4 : smgrprefetch(operation->smgr,
1523 : operation->forknum,
1524 : blockNum,
1525 : actual_nblocks);
1526 : }
1527 :
1528 : /*
1529 : * Indicate that WaitReadBuffers() should be called. WaitReadBuffers()
1530 : * will initiate the necessary IO.
1531 : */
1532 2152 : did_start_io = true;
1533 : }
1534 :
1535 2579374 : CheckReadBuffersOperation(operation, !did_start_io);
1536 :
1537 2579374 : return did_start_io;
1538 : }
1539 :
1540 : /*
1541 : * Begin reading a range of blocks beginning at blockNum and extending for
1542 : * *nblocks. *nblocks and the buffers array are in/out parameters. On entry,
1543 : * the buffers elements covered by *nblocks must hold either InvalidBuffer or
1544 : * buffers forwarded by an earlier call to StartReadBuffers() that was split
1545 : * and is now being continued. On return, *nblocks holds the number of blocks
1546 : * accepted by this operation. If it is less than the original number then
1547 : * this operation has been split, but buffer elements up to the original
1548 : * requested size may hold forwarded buffers to be used for a continuing
1549 : * operation. The caller must either start a new I/O beginning at the block
1550 : * immediately following the blocks accepted by this call and pass those
1551 : * buffers back in, or release them if it chooses not to. It shouldn't make
1552 : * any other use of or assumptions about forwarded buffers.
1553 : *
1554 : * If false is returned, no I/O is necessary and the buffers covered by
1555 : * *nblocks on exit are valid and ready to be accessed. If true is returned,
1556 : * an I/O has been started, and WaitReadBuffers() must be called with the same
1557 : * operation object before the buffers covered by *nblocks on exit can be
1558 : * accessed. Along with the operation object, the caller-supplied array of
1559 : * buffers must remain valid until WaitReadBuffers() is called, and any
1560 : * forwarded buffers must also be preserved for a continuing call unless
1561 : * they are explicitly released.
1562 : */
1563 : bool
1564 3573928 : StartReadBuffers(ReadBuffersOperation *operation,
1565 : Buffer *buffers,
1566 : BlockNumber blockNum,
1567 : int *nblocks,
1568 : int flags)
1569 : {
1570 3573928 : return StartReadBuffersImpl(operation, buffers, blockNum, nblocks, flags,
1571 : true /* expect forwarded buffers */ );
1572 : }
1573 :
1574 : /*
1575 : * Single block version of the StartReadBuffers(). This might save a few
1576 : * instructions when called from another translation unit, because it is
1577 : * specialized for nblocks == 1.
1578 : *
1579 : * This version does not support "forwarded" buffers: they cannot be created
1580 : * by reading only one block and *buffer is ignored on entry.
1581 : */
1582 : bool
1583 120574666 : StartReadBuffer(ReadBuffersOperation *operation,
1584 : Buffer *buffer,
1585 : BlockNumber blocknum,
1586 : int flags)
1587 : {
1588 120574666 : int nblocks = 1;
1589 : bool result;
1590 :
1591 120574666 : result = StartReadBuffersImpl(operation, buffer, blocknum, &nblocks, flags,
1592 : false /* single block, no forwarding */ );
1593 : Assert(nblocks == 1); /* single block can't be short */
1594 :
1595 120574636 : return result;
1596 : }
1597 :
1598 : /*
1599 : * Perform sanity checks on the ReadBuffersOperation.
1600 : */
1601 : static void
1602 7707392 : CheckReadBuffersOperation(ReadBuffersOperation *operation, bool is_complete)
1603 : {
1604 : #ifdef USE_ASSERT_CHECKING
1605 : Assert(operation->nblocks_done <= operation->nblocks);
1606 : Assert(!is_complete || operation->nblocks == operation->nblocks_done);
1607 :
1608 : for (int i = 0; i < operation->nblocks; i++)
1609 : {
1610 : Buffer buffer = operation->buffers[i];
1611 : BufferDesc *buf_hdr = BufferIsLocal(buffer) ?
1612 : GetLocalBufferDescriptor(-buffer - 1) :
1613 : GetBufferDescriptor(buffer - 1);
1614 :
1615 : Assert(BufferGetBlockNumber(buffer) == operation->blocknum + i);
1616 : Assert(pg_atomic_read_u32(&buf_hdr->state) & BM_TAG_VALID);
1617 :
1618 : if (i < operation->nblocks_done)
1619 : Assert(pg_atomic_read_u32(&buf_hdr->state) & BM_VALID);
1620 : }
1621 : #endif
1622 7707392 : }
1623 :
1624 : /* helper for ReadBuffersCanStartIO(), to avoid repetition */
1625 : static inline bool
1626 2919500 : ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
1627 : {
1628 2919500 : if (BufferIsLocal(buffer))
1629 16734 : return StartLocalBufferIO(GetLocalBufferDescriptor(-buffer - 1),
1630 : true, nowait);
1631 : else
1632 2902766 : return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
1633 : }
1634 :
1635 : /*
1636 : * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
1637 : */
1638 : static inline bool
1639 2919500 : ReadBuffersCanStartIO(Buffer buffer, bool nowait)
1640 : {
1641 : /*
1642 : * If this backend currently has staged IO, we need to submit the pending
1643 : * IO before waiting for the right to issue IO, to avoid the potential for
1644 : * deadlocks (and, more commonly, unnecessary delays for other backends).
1645 : */
1646 2919500 : if (!nowait && pgaio_have_staged())
1647 : {
1648 1238 : if (ReadBuffersCanStartIOOnce(buffer, true))
1649 1238 : return true;
1650 :
1651 : /*
1652 : * Unfortunately StartBufferIO() returning false doesn't allow to
1653 : * distinguish between the buffer already being valid and IO already
1654 : * being in progress. Since IO already being in progress is quite
1655 : * rare, this approach seems fine.
1656 : */
1657 0 : pgaio_submit_staged();
1658 : }
1659 :
1660 2918262 : return ReadBuffersCanStartIOOnce(buffer, nowait);
1661 : }
1662 :
1663 : /*
1664 : * Helper for WaitReadBuffers() that processes the results of a readv
1665 : * operation, raising an error if necessary.
1666 : */
1667 : static void
1668 2562972 : ProcessReadBuffersResult(ReadBuffersOperation *operation)
1669 : {
1670 2562972 : PgAioReturn *aio_ret = &operation->io_return;
1671 2562972 : PgAioResultStatus rs = aio_ret->result.status;
1672 2562972 : int newly_read_blocks = 0;
1673 :
1674 : Assert(pgaio_wref_valid(&operation->io_wref));
1675 : Assert(aio_ret->result.status != PGAIO_RS_UNKNOWN);
1676 :
1677 : /*
1678 : * SMGR reports the number of blocks successfully read as the result of
1679 : * the IO operation. Thus we can simply add that to ->nblocks_done.
1680 : */
1681 :
1682 2562972 : if (likely(rs != PGAIO_RS_ERROR))
1683 2562914 : newly_read_blocks = aio_ret->result.result;
1684 :
1685 2562972 : if (rs == PGAIO_RS_ERROR || rs == PGAIO_RS_WARNING)
1686 90 : pgaio_result_report(aio_ret->result, &aio_ret->target_data,
1687 : rs == PGAIO_RS_ERROR ? ERROR : WARNING);
1688 2562882 : else if (aio_ret->result.status == PGAIO_RS_PARTIAL)
1689 : {
1690 : /*
1691 : * We'll retry, so we just emit a debug message to the server log (or
1692 : * not even that in prod scenarios).
1693 : */
1694 20 : pgaio_result_report(aio_ret->result, &aio_ret->target_data, DEBUG1);
1695 20 : elog(DEBUG3, "partial read, will retry");
1696 : }
1697 :
1698 : Assert(newly_read_blocks > 0);
1699 : Assert(newly_read_blocks <= MAX_IO_COMBINE_LIMIT);
1700 :
1701 2562914 : operation->nblocks_done += newly_read_blocks;
1702 :
1703 : Assert(operation->nblocks_done <= operation->nblocks);
1704 2562914 : }
1705 :
1706 : void
1707 2562952 : WaitReadBuffers(ReadBuffersOperation *operation)
1708 : {
1709 2562952 : PgAioReturn *aio_ret = &operation->io_return;
1710 : IOContext io_context;
1711 : IOObject io_object;
1712 :
1713 2562952 : if (operation->persistence == RELPERSISTENCE_TEMP)
1714 : {
1715 2980 : io_context = IOCONTEXT_NORMAL;
1716 2980 : io_object = IOOBJECT_TEMP_RELATION;
1717 : }
1718 : else
1719 : {
1720 2559972 : io_context = IOContextForStrategy(operation->strategy);
1721 2559972 : io_object = IOOBJECT_RELATION;
1722 : }
1723 :
1724 : /*
1725 : * If we get here without an IO operation having been issued, the
1726 : * io_method == IOMETHOD_SYNC path must have been used. Otherwise the
1727 : * caller should not have called WaitReadBuffers().
1728 : *
1729 : * In the case of IOMETHOD_SYNC, we start - as we used to before the
1730 : * introducing of AIO - the IO in WaitReadBuffers(). This is done as part
1731 : * of the retry logic below, no extra code is required.
1732 : *
1733 : * This path is expected to eventually go away.
1734 : */
1735 2562952 : if (!pgaio_wref_valid(&operation->io_wref) && io_method != IOMETHOD_SYNC)
1736 0 : elog(ERROR, "waiting for read operation that didn't read");
1737 :
1738 : /*
1739 : * To handle partial reads, and IOMETHOD_SYNC, we re-issue IO until we're
1740 : * done. We may need multiple retries, not just because we could get
1741 : * multiple partial reads, but also because some of the remaining
1742 : * to-be-read buffers may have been read in by other backends, limiting
1743 : * the IO size.
1744 : */
1745 : while (true)
1746 2172 : {
1747 : int ignored_nblocks_progress;
1748 :
1749 2565124 : CheckReadBuffersOperation(operation, false);
1750 :
1751 : /*
1752 : * If there is an IO associated with the operation, we may need to
1753 : * wait for it.
1754 : */
1755 2565124 : if (pgaio_wref_valid(&operation->io_wref))
1756 : {
1757 : /*
1758 : * Track the time spent waiting for the IO to complete. As
1759 : * tracking a wait even if we don't actually need to wait
1760 : *
1761 : * a) is not cheap, due to the timestamping overhead
1762 : *
1763 : * b) reports some time as waiting, even if we never waited
1764 : *
1765 : * we first check if we already know the IO is complete.
1766 : */
1767 2562972 : if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
1768 1085962 : !pgaio_wref_check_done(&operation->io_wref))
1769 : {
1770 359432 : instr_time io_start = pgstat_prepare_io_time(track_io_timing);
1771 :
1772 359432 : pgaio_wref_wait(&operation->io_wref);
1773 :
1774 : /*
1775 : * The IO operation itself was already counted earlier, in
1776 : * AsyncReadBuffers(), this just accounts for the wait time.
1777 : */
1778 359432 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
1779 : io_start, 0, 0);
1780 : }
1781 : else
1782 : {
1783 : Assert(pgaio_wref_check_done(&operation->io_wref));
1784 : }
1785 :
1786 : /*
1787 : * We now are sure the IO completed. Check the results. This
1788 : * includes reporting on errors if there were any.
1789 : */
1790 2562972 : ProcessReadBuffersResult(operation);
1791 : }
1792 :
1793 : /*
1794 : * Most of the time, the one IO we already started, will read in
1795 : * everything. But we need to deal with partial reads and buffers not
1796 : * needing IO anymore.
1797 : */
1798 2565066 : if (operation->nblocks_done == operation->nblocks)
1799 2562894 : break;
1800 :
1801 2172 : CHECK_FOR_INTERRUPTS();
1802 :
1803 : /*
1804 : * This may only complete the IO partially, either because some
1805 : * buffers were already valid, or because of a partial read.
1806 : *
1807 : * NB: In contrast to after the AsyncReadBuffers() call in
1808 : * StartReadBuffers(), we do *not* reduce
1809 : * ReadBuffersOperation->nblocks here, callers expect the full
1810 : * operation to be completed at this point (as more operations may
1811 : * have been queued).
1812 : */
1813 2172 : AsyncReadBuffers(operation, &ignored_nblocks_progress);
1814 : }
1815 :
1816 2562894 : CheckReadBuffersOperation(operation, true);
1817 :
1818 : /* NB: READ_DONE tracepoint was already executed in completion callback */
1819 2562894 : }
1820 :
1821 : /*
1822 : * Initiate IO for the ReadBuffersOperation
1823 : *
1824 : * This function only starts a single IO at a time. The size of the IO may be
1825 : * limited to below the to-be-read blocks, if one of the buffers has
1826 : * concurrently been read in. If the first to-be-read buffer is already valid,
1827 : * no IO will be issued.
1828 : *
1829 : * To support retries after partial reads, the first operation->nblocks_done
1830 : * buffers are skipped.
1831 : *
1832 : * On return *nblocks_progress is updated to reflect the number of buffers
1833 : * affected by the call. If the first buffer is valid, *nblocks_progress is
1834 : * set to 1 and operation->nblocks_done is incremented.
1835 : *
1836 : * Returns true if IO was initiated, false if no IO was necessary.
1837 : */
1838 : static bool
1839 2579424 : AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
1840 : {
1841 2579424 : Buffer *buffers = &operation->buffers[0];
1842 2579424 : int flags = operation->flags;
1843 2579424 : BlockNumber blocknum = operation->blocknum;
1844 2579424 : ForkNumber forknum = operation->forknum;
1845 2579424 : char persistence = operation->persistence;
1846 2579424 : int16 nblocks_done = operation->nblocks_done;
1847 2579424 : Buffer *io_buffers = &operation->buffers[nblocks_done];
1848 2579424 : int io_buffers_len = 0;
1849 : PgAioHandle *ioh;
1850 2579424 : uint32 ioh_flags = 0;
1851 : void *io_pages[MAX_IO_COMBINE_LIMIT];
1852 : IOContext io_context;
1853 : IOObject io_object;
1854 : bool did_start_io;
1855 :
1856 : /*
1857 : * When this IO is executed synchronously, either because the caller will
1858 : * immediately block waiting for the IO or because IOMETHOD_SYNC is used,
1859 : * the AIO subsystem needs to know.
1860 : */
1861 2579424 : if (flags & READ_BUFFERS_SYNCHRONOUSLY)
1862 1470072 : ioh_flags |= PGAIO_HF_SYNCHRONOUS;
1863 :
1864 2579424 : if (persistence == RELPERSISTENCE_TEMP)
1865 : {
1866 3568 : io_context = IOCONTEXT_NORMAL;
1867 3568 : io_object = IOOBJECT_TEMP_RELATION;
1868 3568 : ioh_flags |= PGAIO_HF_REFERENCES_LOCAL;
1869 : }
1870 : else
1871 : {
1872 2575856 : io_context = IOContextForStrategy(operation->strategy);
1873 2575856 : io_object = IOOBJECT_RELATION;
1874 : }
1875 :
1876 : /*
1877 : * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
1878 : * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
1879 : * set globally, but on a per-session basis. The completion callback,
1880 : * which may be run in other processes, e.g. in IO workers, may have a
1881 : * different value of the zero_damaged_pages GUC.
1882 : *
1883 : * XXX: We probably should eventually use a different flag for
1884 : * zero_damaged_pages, so we can report different log levels / error codes
1885 : * for zero_damaged_pages and ZERO_ON_ERROR.
1886 : */
1887 2579424 : if (zero_damaged_pages)
1888 32 : flags |= READ_BUFFERS_ZERO_ON_ERROR;
1889 :
1890 : /*
1891 : * For the same reason as with zero_damaged_pages we need to use this
1892 : * backend's ignore_checksum_failure value.
1893 : */
1894 2579424 : if (ignore_checksum_failure)
1895 16 : flags |= READ_BUFFERS_IGNORE_CHECKSUM_FAILURES;
1896 :
1897 :
1898 : /*
1899 : * To be allowed to report stats in the local completion callback we need
1900 : * to prepare to report stats now. This ensures we can safely report the
1901 : * checksum failure even in a critical section.
1902 : */
1903 2579424 : pgstat_prepare_report_checksum_failure(operation->smgr->smgr_rlocator.locator.dbOid);
1904 :
1905 : /*
1906 : * Get IO handle before ReadBuffersCanStartIO(), as pgaio_io_acquire()
1907 : * might block, which we don't want after setting IO_IN_PROGRESS.
1908 : *
1909 : * If we need to wait for IO before we can get a handle, submit
1910 : * already-staged IO first, so that other backends don't need to wait.
1911 : * There wouldn't be a deadlock risk, as pgaio_io_acquire() just needs to
1912 : * wait for already submitted IO, which doesn't require additional locks,
1913 : * but it could still cause undesirable waits.
1914 : *
1915 : * A secondary benefit is that this would allow us to measure the time in
1916 : * pgaio_io_acquire() without causing undue timer overhead in the common,
1917 : * non-blocking, case. However, currently the pgstats infrastructure
1918 : * doesn't really allow that, as it a) asserts that an operation can't
1919 : * have time without operations b) doesn't have an API to report
1920 : * "accumulated" time.
1921 : */
1922 2579424 : ioh = pgaio_io_acquire_nb(CurrentResourceOwner, &operation->io_return);
1923 2579424 : if (unlikely(!ioh))
1924 : {
1925 7382 : pgaio_submit_staged();
1926 :
1927 7382 : ioh = pgaio_io_acquire(CurrentResourceOwner, &operation->io_return);
1928 : }
1929 :
1930 : /*
1931 : * Check if we can start IO on the first to-be-read buffer.
1932 : *
1933 : * If an I/O is already in progress in another backend, we want to wait
1934 : * for the outcome: either done, or something went wrong and we will
1935 : * retry.
1936 : */
1937 2579424 : if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
1938 : {
1939 : /*
1940 : * Someone else has already completed this block, we're done.
1941 : *
1942 : * When IO is necessary, ->nblocks_done is updated in
1943 : * ProcessReadBuffersResult(), but that is not called if no IO is
1944 : * necessary. Thus update here.
1945 : */
1946 15822 : operation->nblocks_done += 1;
1947 15822 : *nblocks_progress = 1;
1948 :
1949 15822 : pgaio_io_release(ioh);
1950 15822 : pgaio_wref_clear(&operation->io_wref);
1951 15822 : did_start_io = false;
1952 :
1953 : /*
1954 : * Report and track this as a 'hit' for this backend, even though it
1955 : * must have started out as a miss in PinBufferForBlock(). The other
1956 : * backend will track this as a 'read'.
1957 : */
1958 : TRACE_POSTGRESQL_BUFFER_READ_DONE(forknum, blocknum + operation->nblocks_done,
1959 : operation->smgr->smgr_rlocator.locator.spcOid,
1960 : operation->smgr->smgr_rlocator.locator.dbOid,
1961 : operation->smgr->smgr_rlocator.locator.relNumber,
1962 : operation->smgr->smgr_rlocator.backend,
1963 : true);
1964 :
1965 15822 : if (persistence == RELPERSISTENCE_TEMP)
1966 0 : pgBufferUsage.local_blks_hit += 1;
1967 : else
1968 15822 : pgBufferUsage.shared_blks_hit += 1;
1969 :
1970 15822 : if (operation->rel)
1971 15822 : pgstat_count_buffer_hit(operation->rel);
1972 :
1973 15822 : pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
1974 :
1975 15822 : if (VacuumCostActive)
1976 108 : VacuumCostBalance += VacuumCostPageHit;
1977 : }
1978 : else
1979 : {
1980 : instr_time io_start;
1981 :
1982 : /* We found a buffer that we need to read in. */
1983 : Assert(io_buffers[0] == buffers[nblocks_done]);
1984 2563602 : io_pages[0] = BufferGetBlock(buffers[nblocks_done]);
1985 2563602 : io_buffers_len = 1;
1986 :
1987 : /*
1988 : * How many neighboring-on-disk blocks can we scatter-read into other
1989 : * buffers at the same time? In this case we don't wait if we see an
1990 : * I/O already in progress. We already set BM_IO_IN_PROGRESS for the
1991 : * head block, so we should get on with that I/O as soon as possible.
1992 : */
1993 2903678 : for (int i = nblocks_done + 1; i < operation->nblocks; i++)
1994 : {
1995 340076 : if (!ReadBuffersCanStartIO(buffers[i], true))
1996 0 : break;
1997 : /* Must be consecutive block numbers. */
1998 : Assert(BufferGetBlockNumber(buffers[i - 1]) ==
1999 : BufferGetBlockNumber(buffers[i]) - 1);
2000 : Assert(io_buffers[io_buffers_len] == buffers[i]);
2001 :
2002 340076 : io_pages[io_buffers_len++] = BufferGetBlock(buffers[i]);
2003 : }
2004 :
2005 : /* get a reference to wait for in WaitReadBuffers() */
2006 2563602 : pgaio_io_get_wref(ioh, &operation->io_wref);
2007 :
2008 : /* provide the list of buffers to the completion callbacks */
2009 2563602 : pgaio_io_set_handle_data_32(ioh, (uint32 *) io_buffers, io_buffers_len);
2010 :
2011 2563602 : pgaio_io_register_callbacks(ioh,
2012 : persistence == RELPERSISTENCE_TEMP ?
2013 : PGAIO_HCB_LOCAL_BUFFER_READV :
2014 : PGAIO_HCB_SHARED_BUFFER_READV,
2015 : flags);
2016 :
2017 2563602 : pgaio_io_set_flag(ioh, ioh_flags);
2018 :
2019 : /* ---
2020 : * Even though we're trying to issue IO asynchronously, track the time
2021 : * in smgrstartreadv():
2022 : * - if io_method == IOMETHOD_SYNC, we will always perform the IO
2023 : * immediately
2024 : * - the io method might not support the IO (e.g. worker IO for a temp
2025 : * table)
2026 : * ---
2027 : */
2028 2563602 : io_start = pgstat_prepare_io_time(track_io_timing);
2029 2563602 : smgrstartreadv(ioh, operation->smgr, forknum,
2030 : blocknum + nblocks_done,
2031 : io_pages, io_buffers_len);
2032 2563572 : pgstat_count_io_op_time(io_object, io_context, IOOP_READ,
2033 2563572 : io_start, 1, io_buffers_len * BLCKSZ);
2034 :
2035 2563572 : if (persistence == RELPERSISTENCE_TEMP)
2036 3568 : pgBufferUsage.local_blks_read += io_buffers_len;
2037 : else
2038 2560004 : pgBufferUsage.shared_blks_read += io_buffers_len;
2039 :
2040 : /*
2041 : * Track vacuum cost when issuing IO, not after waiting for it.
2042 : * Otherwise we could end up issuing a lot of IO in a short timespan,
2043 : * despite a low cost limit.
2044 : */
2045 2563572 : if (VacuumCostActive)
2046 60098 : VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
2047 :
2048 2563572 : *nblocks_progress = io_buffers_len;
2049 2563572 : did_start_io = true;
2050 : }
2051 :
2052 2579394 : return did_start_io;
2053 : }
2054 :
2055 : /*
2056 : * BufferAlloc -- subroutine for PinBufferForBlock. Handles lookup of a shared
2057 : * buffer. If no buffer exists already, selects a replacement victim and
2058 : * evicts the old page, but does NOT read in new page.
2059 : *
2060 : * "strategy" can be a buffer replacement strategy object, or NULL for
2061 : * the default strategy. The selected buffer's usage_count is advanced when
2062 : * using the default strategy, but otherwise possibly not (see PinBuffer).
2063 : *
2064 : * The returned buffer is pinned and is already marked as holding the
2065 : * desired page. If it already did have the desired page, *foundPtr is
2066 : * set true. Otherwise, *foundPtr is set false.
2067 : *
2068 : * io_context is passed as an output parameter to avoid calling
2069 : * IOContextForStrategy() when there is a shared buffers hit and no IO
2070 : * statistics need be captured.
2071 : *
2072 : * No locks are held either at entry or exit.
2073 : */
2074 : static pg_attribute_always_inline BufferDesc *
2075 122576332 : BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
2076 : BlockNumber blockNum,
2077 : BufferAccessStrategy strategy,
2078 : bool *foundPtr, IOContext io_context)
2079 : {
2080 : BufferTag newTag; /* identity of requested block */
2081 : uint32 newHash; /* hash value for newTag */
2082 : LWLock *newPartitionLock; /* buffer partition lock for it */
2083 : int existing_buf_id;
2084 : Buffer victim_buffer;
2085 : BufferDesc *victim_buf_hdr;
2086 : uint32 victim_buf_state;
2087 122576332 : uint32 set_bits = 0;
2088 :
2089 : /* Make sure we will have room to remember the buffer pin */
2090 122576332 : ResourceOwnerEnlarge(CurrentResourceOwner);
2091 122576332 : ReservePrivateRefCountEntry();
2092 :
2093 : /* create a tag so we can lookup the buffer */
2094 122576332 : InitBufferTag(&newTag, &smgr->smgr_rlocator.locator, forkNum, blockNum);
2095 :
2096 : /* determine its hash code and partition lock ID */
2097 122576332 : newHash = BufTableHashCode(&newTag);
2098 122576332 : newPartitionLock = BufMappingPartitionLock(newHash);
2099 :
2100 : /* see if the block is in the buffer pool already */
2101 122576332 : LWLockAcquire(newPartitionLock, LW_SHARED);
2102 122576332 : existing_buf_id = BufTableLookup(&newTag, newHash);
2103 122576332 : if (existing_buf_id >= 0)
2104 : {
2105 : BufferDesc *buf;
2106 : bool valid;
2107 :
2108 : /*
2109 : * Found it. Now, pin the buffer so no one can steal it from the
2110 : * buffer pool, and check to see if the correct data has been loaded
2111 : * into the buffer.
2112 : */
2113 119118740 : buf = GetBufferDescriptor(existing_buf_id);
2114 :
2115 119118740 : valid = PinBuffer(buf, strategy, false);
2116 :
2117 : /* Can release the mapping lock as soon as we've pinned it */
2118 119118740 : LWLockRelease(newPartitionLock);
2119 :
2120 119118740 : *foundPtr = true;
2121 :
2122 119118740 : if (!valid)
2123 : {
2124 : /*
2125 : * We can only get here if (a) someone else is still reading in
2126 : * the page, (b) a previous read attempt failed, or (c) someone
2127 : * called StartReadBuffers() but not yet WaitReadBuffers().
2128 : */
2129 13458 : *foundPtr = false;
2130 : }
2131 :
2132 119118740 : return buf;
2133 : }
2134 :
2135 : /*
2136 : * Didn't find it in the buffer pool. We'll have to initialize a new
2137 : * buffer. Remember to unlock the mapping lock while doing the work.
2138 : */
2139 3457592 : LWLockRelease(newPartitionLock);
2140 :
2141 : /*
2142 : * Acquire a victim buffer. Somebody else might try to do the same, we
2143 : * don't hold any conflicting locks. If so we'll have to undo our work
2144 : * later.
2145 : */
2146 3457592 : victim_buffer = GetVictimBuffer(strategy, io_context);
2147 3457592 : victim_buf_hdr = GetBufferDescriptor(victim_buffer - 1);
2148 :
2149 : /*
2150 : * Try to make a hashtable entry for the buffer under its new tag. If
2151 : * somebody else inserted another buffer for the tag, we'll release the
2152 : * victim buffer we acquired and use the already inserted one.
2153 : */
2154 3457592 : LWLockAcquire(newPartitionLock, LW_EXCLUSIVE);
2155 3457592 : existing_buf_id = BufTableInsert(&newTag, newHash, victim_buf_hdr->buf_id);
2156 3457592 : if (existing_buf_id >= 0)
2157 : {
2158 : BufferDesc *existing_buf_hdr;
2159 : bool valid;
2160 :
2161 : /*
2162 : * Got a collision. Someone has already done what we were about to do.
2163 : * We'll just handle this as if it were found in the buffer pool in
2164 : * the first place. First, give up the buffer we were planning to
2165 : * use.
2166 : *
2167 : * We could do this after releasing the partition lock, but then we'd
2168 : * have to call ResourceOwnerEnlarge() & ReservePrivateRefCountEntry()
2169 : * before acquiring the lock, for the rare case of such a collision.
2170 : */
2171 3288 : UnpinBuffer(victim_buf_hdr);
2172 :
2173 : /* remaining code should match code at top of routine */
2174 :
2175 3288 : existing_buf_hdr = GetBufferDescriptor(existing_buf_id);
2176 :
2177 3288 : valid = PinBuffer(existing_buf_hdr, strategy, false);
2178 :
2179 : /* Can release the mapping lock as soon as we've pinned it */
2180 3288 : LWLockRelease(newPartitionLock);
2181 :
2182 3288 : *foundPtr = true;
2183 :
2184 3288 : if (!valid)
2185 : {
2186 : /*
2187 : * We can only get here if (a) someone else is still reading in
2188 : * the page, (b) a previous read attempt failed, or (c) someone
2189 : * called StartReadBuffers() but not yet WaitReadBuffers().
2190 : */
2191 2460 : *foundPtr = false;
2192 : }
2193 :
2194 3288 : return existing_buf_hdr;
2195 : }
2196 :
2197 : /*
2198 : * Need to lock the buffer header too in order to change its tag.
2199 : */
2200 3454304 : victim_buf_state = LockBufHdr(victim_buf_hdr);
2201 :
2202 : /* some sanity checks while we hold the buffer header lock */
2203 : Assert(BUF_STATE_GET_REFCOUNT(victim_buf_state) == 1);
2204 : Assert(!(victim_buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY | BM_IO_IN_PROGRESS)));
2205 :
2206 3454304 : victim_buf_hdr->tag = newTag;
2207 :
2208 : /*
2209 : * Make sure BM_PERMANENT is set for buffers that must be written at every
2210 : * checkpoint. Unlogged buffers only need to be written at shutdown
2211 : * checkpoints, except for their "init" forks, which need to be treated
2212 : * just like permanent relations.
2213 : */
2214 3454304 : set_bits |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2215 3454304 : if (relpersistence == RELPERSISTENCE_PERMANENT || forkNum == INIT_FORKNUM)
2216 3453592 : set_bits |= BM_PERMANENT;
2217 :
2218 3454304 : UnlockBufHdrExt(victim_buf_hdr, victim_buf_state,
2219 : set_bits, 0, 0);
2220 :
2221 3454304 : LWLockRelease(newPartitionLock);
2222 :
2223 : /*
2224 : * Buffer contents are currently invalid.
2225 : */
2226 3454304 : *foundPtr = false;
2227 :
2228 3454304 : return victim_buf_hdr;
2229 : }
2230 :
2231 : /*
2232 : * InvalidateBuffer -- mark a shared buffer invalid.
2233 : *
2234 : * The buffer header spinlock must be held at entry. We drop it before
2235 : * returning. (This is sane because the caller must have locked the
2236 : * buffer in order to be sure it should be dropped.)
2237 : *
2238 : * This is used only in contexts such as dropping a relation. We assume
2239 : * that no other backend could possibly be interested in using the page,
2240 : * so the only reason the buffer might be pinned is if someone else is
2241 : * trying to write it out. We have to let them finish before we can
2242 : * reclaim the buffer.
2243 : *
2244 : * The buffer could get reclaimed by someone else while we are waiting
2245 : * to acquire the necessary locks; if so, don't mess it up.
2246 : */
2247 : static void
2248 212380 : InvalidateBuffer(BufferDesc *buf)
2249 : {
2250 : BufferTag oldTag;
2251 : uint32 oldHash; /* hash value for oldTag */
2252 : LWLock *oldPartitionLock; /* buffer partition lock for it */
2253 : uint32 oldFlags;
2254 : uint32 buf_state;
2255 :
2256 : /* Save the original buffer tag before dropping the spinlock */
2257 212380 : oldTag = buf->tag;
2258 :
2259 212380 : UnlockBufHdr(buf);
2260 :
2261 : /*
2262 : * Need to compute the old tag's hashcode and partition lock ID. XXX is it
2263 : * worth storing the hashcode in BufferDesc so we need not recompute it
2264 : * here? Probably not.
2265 : */
2266 212380 : oldHash = BufTableHashCode(&oldTag);
2267 212380 : oldPartitionLock = BufMappingPartitionLock(oldHash);
2268 :
2269 212386 : retry:
2270 :
2271 : /*
2272 : * Acquire exclusive mapping lock in preparation for changing the buffer's
2273 : * association.
2274 : */
2275 212386 : LWLockAcquire(oldPartitionLock, LW_EXCLUSIVE);
2276 :
2277 : /* Re-lock the buffer header */
2278 212386 : buf_state = LockBufHdr(buf);
2279 :
2280 : /* If it's changed while we were waiting for lock, do nothing */
2281 212386 : if (!BufferTagsEqual(&buf->tag, &oldTag))
2282 : {
2283 6 : UnlockBufHdr(buf);
2284 6 : LWLockRelease(oldPartitionLock);
2285 6 : return;
2286 : }
2287 :
2288 : /*
2289 : * We assume the reason for it to be pinned is that either we were
2290 : * asynchronously reading the page in before erroring out or someone else
2291 : * is flushing the page out. Wait for the IO to finish. (This could be
2292 : * an infinite loop if the refcount is messed up... it would be nice to
2293 : * time out after awhile, but there seems no way to be sure how many loops
2294 : * may be needed. Note that if the other guy has pinned the buffer but
2295 : * not yet done StartBufferIO, WaitIO will fall through and we'll
2296 : * effectively be busy-looping here.)
2297 : */
2298 212380 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 0)
2299 : {
2300 6 : UnlockBufHdr(buf);
2301 6 : LWLockRelease(oldPartitionLock);
2302 : /* safety check: should definitely not be our *own* pin */
2303 6 : if (GetPrivateRefCount(BufferDescriptorGetBuffer(buf)) > 0)
2304 0 : elog(ERROR, "buffer is pinned in InvalidateBuffer");
2305 6 : WaitIO(buf);
2306 6 : goto retry;
2307 : }
2308 :
2309 : /*
2310 : * Clear out the buffer's tag and flags. We must do this to ensure that
2311 : * linear scans of the buffer array don't think the buffer is valid.
2312 : */
2313 212374 : oldFlags = buf_state & BUF_FLAG_MASK;
2314 212374 : ClearBufferTag(&buf->tag);
2315 :
2316 212374 : UnlockBufHdrExt(buf, buf_state,
2317 : 0,
2318 : BUF_FLAG_MASK | BUF_USAGECOUNT_MASK,
2319 : 0);
2320 :
2321 : /*
2322 : * Remove the buffer from the lookup hashtable, if it was in there.
2323 : */
2324 212374 : if (oldFlags & BM_TAG_VALID)
2325 212374 : BufTableDelete(&oldTag, oldHash);
2326 :
2327 : /*
2328 : * Done with mapping lock.
2329 : */
2330 212374 : LWLockRelease(oldPartitionLock);
2331 : }
2332 :
2333 : /*
2334 : * Helper routine for GetVictimBuffer()
2335 : *
2336 : * Needs to be called on a buffer with a valid tag, pinned, but without the
2337 : * buffer header spinlock held.
2338 : *
2339 : * Returns true if the buffer can be reused, in which case the buffer is only
2340 : * pinned by this backend and marked as invalid, false otherwise.
2341 : */
2342 : static bool
2343 2433424 : InvalidateVictimBuffer(BufferDesc *buf_hdr)
2344 : {
2345 : uint32 buf_state;
2346 : uint32 hash;
2347 : LWLock *partition_lock;
2348 : BufferTag tag;
2349 :
2350 : Assert(GetPrivateRefCount(BufferDescriptorGetBuffer(buf_hdr)) == 1);
2351 :
2352 : /* have buffer pinned, so it's safe to read tag without lock */
2353 2433424 : tag = buf_hdr->tag;
2354 :
2355 2433424 : hash = BufTableHashCode(&tag);
2356 2433424 : partition_lock = BufMappingPartitionLock(hash);
2357 :
2358 2433424 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2359 :
2360 : /* lock the buffer header */
2361 2433424 : buf_state = LockBufHdr(buf_hdr);
2362 :
2363 : /*
2364 : * We have the buffer pinned nobody else should have been able to unset
2365 : * this concurrently.
2366 : */
2367 : Assert(buf_state & BM_TAG_VALID);
2368 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2369 : Assert(BufferTagsEqual(&buf_hdr->tag, &tag));
2370 :
2371 : /*
2372 : * If somebody else pinned the buffer since, or even worse, dirtied it,
2373 : * give up on this buffer: It's clearly in use.
2374 : */
2375 2433424 : if (BUF_STATE_GET_REFCOUNT(buf_state) != 1 || (buf_state & BM_DIRTY))
2376 : {
2377 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2378 :
2379 872 : UnlockBufHdr(buf_hdr);
2380 872 : LWLockRelease(partition_lock);
2381 :
2382 872 : return false;
2383 : }
2384 :
2385 : /*
2386 : * Clear out the buffer's tag and flags and usagecount. This is not
2387 : * strictly required, as BM_TAG_VALID/BM_VALID needs to be checked before
2388 : * doing anything with the buffer. But currently it's beneficial, as the
2389 : * cheaper pre-check for several linear scans of shared buffers use the
2390 : * tag (see e.g. FlushDatabaseBuffers()).
2391 : */
2392 2432552 : ClearBufferTag(&buf_hdr->tag);
2393 2432552 : UnlockBufHdrExt(buf_hdr, buf_state,
2394 : 0,
2395 : BUF_FLAG_MASK | BUF_USAGECOUNT_MASK,
2396 : 0);
2397 :
2398 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2399 :
2400 : /* finally delete buffer from the buffer mapping table */
2401 2432552 : BufTableDelete(&tag, hash);
2402 :
2403 2432552 : LWLockRelease(partition_lock);
2404 :
2405 2432552 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
2406 : Assert(!(buf_state & (BM_DIRTY | BM_VALID | BM_TAG_VALID)));
2407 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
2408 : Assert(BUF_STATE_GET_REFCOUNT(pg_atomic_read_u32(&buf_hdr->state)) > 0);
2409 :
2410 2432552 : return true;
2411 : }
2412 :
2413 : static Buffer
2414 3905718 : GetVictimBuffer(BufferAccessStrategy strategy, IOContext io_context)
2415 : {
2416 : BufferDesc *buf_hdr;
2417 : Buffer buf;
2418 : uint32 buf_state;
2419 : bool from_ring;
2420 :
2421 : /*
2422 : * Ensure, before we pin a victim buffer, that there's a free refcount
2423 : * entry and resource owner slot for the pin.
2424 : */
2425 3905718 : ReservePrivateRefCountEntry();
2426 3905718 : ResourceOwnerEnlarge(CurrentResourceOwner);
2427 :
2428 : /* we return here if a prospective victim buffer gets used concurrently */
2429 11778 : again:
2430 :
2431 : /*
2432 : * Select a victim buffer. The buffer is returned pinned and owned by
2433 : * this backend.
2434 : */
2435 3917496 : buf_hdr = StrategyGetBuffer(strategy, &buf_state, &from_ring);
2436 3917496 : buf = BufferDescriptorGetBuffer(buf_hdr);
2437 :
2438 : /*
2439 : * We shouldn't have any other pins for this buffer.
2440 : */
2441 3917496 : CheckBufferIsPinnedOnce(buf);
2442 :
2443 : /*
2444 : * If the buffer was dirty, try to write it out. There is a race
2445 : * condition here, in that someone might dirty it after we released the
2446 : * buffer header lock above, or even while we are writing it out (since
2447 : * our share-lock won't prevent hint-bit updates). We will recheck the
2448 : * dirty bit after re-locking the buffer header.
2449 : */
2450 3917496 : if (buf_state & BM_DIRTY)
2451 : {
2452 : LWLock *content_lock;
2453 :
2454 : Assert(buf_state & BM_TAG_VALID);
2455 : Assert(buf_state & BM_VALID);
2456 :
2457 : /*
2458 : * We need a share-lock on the buffer contents to write it out (else
2459 : * we might write invalid data, eg because someone else is compacting
2460 : * the page contents while we write). We must use a conditional lock
2461 : * acquisition here to avoid deadlock. Even though the buffer was not
2462 : * pinned (and therefore surely not locked) when StrategyGetBuffer
2463 : * returned it, someone else could have pinned and exclusive-locked it
2464 : * by the time we get here. If we try to get the lock unconditionally,
2465 : * we'd block waiting for them; if they later block waiting for us,
2466 : * deadlock ensues. (This has been observed to happen when two
2467 : * backends are both trying to split btree index pages, and the second
2468 : * one just happens to be trying to split the page the first one got
2469 : * from StrategyGetBuffer.)
2470 : */
2471 546382 : content_lock = BufferDescriptorGetContentLock(buf_hdr);
2472 546382 : if (!LWLockConditionalAcquire(content_lock, LW_SHARED))
2473 : {
2474 : /*
2475 : * Someone else has locked the buffer, so give it up and loop back
2476 : * to get another one.
2477 : */
2478 0 : UnpinBuffer(buf_hdr);
2479 0 : goto again;
2480 : }
2481 :
2482 : /*
2483 : * If using a nondefault strategy, and writing the buffer would
2484 : * require a WAL flush, let the strategy decide whether to go ahead
2485 : * and write/reuse the buffer or to choose another victim. We need a
2486 : * lock to inspect the page LSN, so this can't be done inside
2487 : * StrategyGetBuffer.
2488 : */
2489 546382 : if (strategy != NULL)
2490 : {
2491 : XLogRecPtr lsn;
2492 :
2493 : /* Read the LSN while holding buffer header lock */
2494 156608 : buf_state = LockBufHdr(buf_hdr);
2495 156608 : lsn = BufferGetLSN(buf_hdr);
2496 156608 : UnlockBufHdr(buf_hdr);
2497 :
2498 156608 : if (XLogNeedsFlush(lsn)
2499 17352 : && StrategyRejectBuffer(strategy, buf_hdr, from_ring))
2500 : {
2501 10906 : LWLockRelease(content_lock);
2502 10906 : UnpinBuffer(buf_hdr);
2503 10906 : goto again;
2504 : }
2505 : }
2506 :
2507 : /* OK, do the I/O */
2508 535476 : FlushBuffer(buf_hdr, NULL, IOOBJECT_RELATION, io_context);
2509 535476 : LWLockRelease(content_lock);
2510 :
2511 535476 : ScheduleBufferTagForWriteback(&BackendWritebackContext, io_context,
2512 : &buf_hdr->tag);
2513 : }
2514 :
2515 :
2516 3906590 : if (buf_state & BM_VALID)
2517 : {
2518 : /*
2519 : * When a BufferAccessStrategy is in use, blocks evicted from shared
2520 : * buffers are counted as IOOP_EVICT in the corresponding context
2521 : * (e.g. IOCONTEXT_BULKWRITE). Shared buffers are evicted by a
2522 : * strategy in two cases: 1) while initially claiming buffers for the
2523 : * strategy ring 2) to replace an existing strategy ring buffer
2524 : * because it is pinned or in use and cannot be reused.
2525 : *
2526 : * Blocks evicted from buffers already in the strategy ring are
2527 : * counted as IOOP_REUSE in the corresponding strategy context.
2528 : *
2529 : * At this point, we can accurately count evictions and reuses,
2530 : * because we have successfully claimed the valid buffer. Previously,
2531 : * we may have been forced to release the buffer due to concurrent
2532 : * pinners or erroring out.
2533 : */
2534 2429138 : pgstat_count_io_op(IOOBJECT_RELATION, io_context,
2535 2429138 : from_ring ? IOOP_REUSE : IOOP_EVICT, 1, 0);
2536 : }
2537 :
2538 : /*
2539 : * If the buffer has an entry in the buffer mapping table, delete it. This
2540 : * can fail because another backend could have pinned or dirtied the
2541 : * buffer.
2542 : */
2543 3906590 : if ((buf_state & BM_TAG_VALID) && !InvalidateVictimBuffer(buf_hdr))
2544 : {
2545 872 : UnpinBuffer(buf_hdr);
2546 872 : goto again;
2547 : }
2548 :
2549 : /* a final set of sanity checks */
2550 : #ifdef USE_ASSERT_CHECKING
2551 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
2552 :
2553 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2554 : Assert(!(buf_state & (BM_TAG_VALID | BM_VALID | BM_DIRTY)));
2555 :
2556 : CheckBufferIsPinnedOnce(buf);
2557 : #endif
2558 :
2559 3905718 : return buf;
2560 : }
2561 :
2562 : /*
2563 : * Return the maximum number of buffers that a backend should try to pin once,
2564 : * to avoid exceeding its fair share. This is the highest value that
2565 : * GetAdditionalPinLimit() could ever return. Note that it may be zero on a
2566 : * system with a very small buffer pool relative to max_connections.
2567 : */
2568 : uint32
2569 1289060 : GetPinLimit(void)
2570 : {
2571 1289060 : return MaxProportionalPins;
2572 : }
2573 :
2574 : /*
2575 : * Return the maximum number of additional buffers that this backend should
2576 : * pin if it wants to stay under the per-backend limit, considering the number
2577 : * of buffers it has already pinned. Unlike LimitAdditionalPins(), the limit
2578 : * return by this function can be zero.
2579 : */
2580 : uint32
2581 7124170 : GetAdditionalPinLimit(void)
2582 : {
2583 : uint32 estimated_pins_held;
2584 :
2585 : /*
2586 : * We get the number of "overflowed" pins for free, but don't know the
2587 : * number of pins in PrivateRefCountArray. The cost of calculating that
2588 : * exactly doesn't seem worth it, so just assume the max.
2589 : */
2590 7124170 : estimated_pins_held = PrivateRefCountOverflowed + REFCOUNT_ARRAY_ENTRIES;
2591 :
2592 : /* Is this backend already holding more than its fair share? */
2593 7124170 : if (estimated_pins_held > MaxProportionalPins)
2594 2394344 : return 0;
2595 :
2596 4729826 : return MaxProportionalPins - estimated_pins_held;
2597 : }
2598 :
2599 : /*
2600 : * Limit the number of pins a batch operation may additionally acquire, to
2601 : * avoid running out of pinnable buffers.
2602 : *
2603 : * One additional pin is always allowed, on the assumption that the operation
2604 : * requires at least one to make progress.
2605 : */
2606 : void
2607 405344 : LimitAdditionalPins(uint32 *additional_pins)
2608 : {
2609 : uint32 limit;
2610 :
2611 405344 : if (*additional_pins <= 1)
2612 385406 : return;
2613 :
2614 19938 : limit = GetAdditionalPinLimit();
2615 19938 : limit = Max(limit, 1);
2616 19938 : if (limit < *additional_pins)
2617 10764 : *additional_pins = limit;
2618 : }
2619 :
2620 : /*
2621 : * Logic shared between ExtendBufferedRelBy(), ExtendBufferedRelTo(). Just to
2622 : * avoid duplicating the tracing and relpersistence related logic.
2623 : */
2624 : static BlockNumber
2625 428200 : ExtendBufferedRelCommon(BufferManagerRelation bmr,
2626 : ForkNumber fork,
2627 : BufferAccessStrategy strategy,
2628 : uint32 flags,
2629 : uint32 extend_by,
2630 : BlockNumber extend_upto,
2631 : Buffer *buffers,
2632 : uint32 *extended_by)
2633 : {
2634 : BlockNumber first_block;
2635 :
2636 : TRACE_POSTGRESQL_BUFFER_EXTEND_START(fork,
2637 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2638 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2639 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2640 : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2641 : extend_by);
2642 :
2643 428200 : if (bmr.relpersistence == RELPERSISTENCE_TEMP)
2644 22856 : first_block = ExtendBufferedRelLocal(bmr, fork, flags,
2645 : extend_by, extend_upto,
2646 : buffers, &extend_by);
2647 : else
2648 405344 : first_block = ExtendBufferedRelShared(bmr, fork, strategy, flags,
2649 : extend_by, extend_upto,
2650 : buffers, &extend_by);
2651 428200 : *extended_by = extend_by;
2652 :
2653 : TRACE_POSTGRESQL_BUFFER_EXTEND_DONE(fork,
2654 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.spcOid,
2655 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.dbOid,
2656 : BMR_GET_SMGR(bmr)->smgr_rlocator.locator.relNumber,
2657 : BMR_GET_SMGR(bmr)->smgr_rlocator.backend,
2658 : *extended_by,
2659 : first_block);
2660 :
2661 428200 : return first_block;
2662 : }
2663 :
2664 : /*
2665 : * Implementation of ExtendBufferedRelBy() and ExtendBufferedRelTo() for
2666 : * shared buffers.
2667 : */
2668 : static BlockNumber
2669 405344 : ExtendBufferedRelShared(BufferManagerRelation bmr,
2670 : ForkNumber fork,
2671 : BufferAccessStrategy strategy,
2672 : uint32 flags,
2673 : uint32 extend_by,
2674 : BlockNumber extend_upto,
2675 : Buffer *buffers,
2676 : uint32 *extended_by)
2677 : {
2678 : BlockNumber first_block;
2679 405344 : IOContext io_context = IOContextForStrategy(strategy);
2680 : instr_time io_start;
2681 :
2682 405344 : LimitAdditionalPins(&extend_by);
2683 :
2684 : /*
2685 : * Acquire victim buffers for extension without holding extension lock.
2686 : * Writing out victim buffers is the most expensive part of extending the
2687 : * relation, particularly when doing so requires WAL flushes. Zeroing out
2688 : * the buffers is also quite expensive, so do that before holding the
2689 : * extension lock as well.
2690 : *
2691 : * These pages are pinned by us and not valid. While we hold the pin they
2692 : * can't be acquired as victim buffers by another backend.
2693 : */
2694 853470 : for (uint32 i = 0; i < extend_by; i++)
2695 : {
2696 : Block buf_block;
2697 :
2698 448126 : buffers[i] = GetVictimBuffer(strategy, io_context);
2699 448126 : buf_block = BufHdrGetBlock(GetBufferDescriptor(buffers[i] - 1));
2700 :
2701 : /* new buffers are zero-filled */
2702 448126 : MemSet(buf_block, 0, BLCKSZ);
2703 : }
2704 :
2705 : /*
2706 : * Lock relation against concurrent extensions, unless requested not to.
2707 : *
2708 : * We use the same extension lock for all forks. That's unnecessarily
2709 : * restrictive, but currently extensions for forks don't happen often
2710 : * enough to make it worth locking more granularly.
2711 : *
2712 : * Note that another backend might have extended the relation by the time
2713 : * we get the lock.
2714 : */
2715 405344 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2716 301648 : LockRelationForExtension(bmr.rel, ExclusiveLock);
2717 :
2718 : /*
2719 : * If requested, invalidate size cache, so that smgrnblocks asks the
2720 : * kernel.
2721 : */
2722 405344 : if (flags & EB_CLEAR_SIZE_CACHE)
2723 15796 : BMR_GET_SMGR(bmr)->smgr_cached_nblocks[fork] = InvalidBlockNumber;
2724 :
2725 405344 : first_block = smgrnblocks(BMR_GET_SMGR(bmr), fork);
2726 :
2727 : /*
2728 : * Now that we have the accurate relation size, check if the caller wants
2729 : * us to extend to only up to a specific size. If there were concurrent
2730 : * extensions, we might have acquired too many buffers and need to release
2731 : * them.
2732 : */
2733 405344 : if (extend_upto != InvalidBlockNumber)
2734 : {
2735 107382 : uint32 orig_extend_by = extend_by;
2736 :
2737 107382 : if (first_block > extend_upto)
2738 0 : extend_by = 0;
2739 107382 : else if ((uint64) first_block + extend_by > extend_upto)
2740 22 : extend_by = extend_upto - first_block;
2741 :
2742 107432 : for (uint32 i = extend_by; i < orig_extend_by; i++)
2743 : {
2744 50 : BufferDesc *buf_hdr = GetBufferDescriptor(buffers[i] - 1);
2745 :
2746 50 : UnpinBuffer(buf_hdr);
2747 : }
2748 :
2749 107382 : if (extend_by == 0)
2750 : {
2751 22 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2752 22 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2753 22 : *extended_by = extend_by;
2754 22 : return first_block;
2755 : }
2756 : }
2757 :
2758 : /* Fail if relation is already at maximum possible length */
2759 405322 : if ((uint64) first_block + extend_by >= MaxBlockNumber)
2760 0 : ereport(ERROR,
2761 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
2762 : errmsg("cannot extend relation %s beyond %u blocks",
2763 : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str,
2764 : MaxBlockNumber)));
2765 :
2766 : /*
2767 : * Insert buffers into buffer table, mark as IO_IN_PROGRESS.
2768 : *
2769 : * This needs to happen before we extend the relation, because as soon as
2770 : * we do, other backends can start to read in those pages.
2771 : */
2772 853398 : for (uint32 i = 0; i < extend_by; i++)
2773 : {
2774 448076 : Buffer victim_buf = buffers[i];
2775 448076 : BufferDesc *victim_buf_hdr = GetBufferDescriptor(victim_buf - 1);
2776 : BufferTag tag;
2777 : uint32 hash;
2778 : LWLock *partition_lock;
2779 : int existing_id;
2780 :
2781 : /* in case we need to pin an existing buffer below */
2782 448076 : ResourceOwnerEnlarge(CurrentResourceOwner);
2783 448076 : ReservePrivateRefCountEntry();
2784 :
2785 448076 : InitBufferTag(&tag, &BMR_GET_SMGR(bmr)->smgr_rlocator.locator, fork,
2786 : first_block + i);
2787 448076 : hash = BufTableHashCode(&tag);
2788 448076 : partition_lock = BufMappingPartitionLock(hash);
2789 :
2790 448076 : LWLockAcquire(partition_lock, LW_EXCLUSIVE);
2791 :
2792 448076 : existing_id = BufTableInsert(&tag, hash, victim_buf_hdr->buf_id);
2793 :
2794 : /*
2795 : * We get here only in the corner case where we are trying to extend
2796 : * the relation but we found a pre-existing buffer. This can happen
2797 : * because a prior attempt at extending the relation failed, and
2798 : * because mdread doesn't complain about reads beyond EOF (when
2799 : * zero_damaged_pages is ON) and so a previous attempt to read a block
2800 : * beyond EOF could have left a "valid" zero-filled buffer.
2801 : *
2802 : * This has also been observed when relation was overwritten by
2803 : * external process. Since the legitimate cases should always have
2804 : * left a zero-filled buffer, complain if not PageIsNew.
2805 : */
2806 448076 : if (existing_id >= 0)
2807 : {
2808 0 : BufferDesc *existing_hdr = GetBufferDescriptor(existing_id);
2809 : Block buf_block;
2810 : bool valid;
2811 :
2812 : /*
2813 : * Pin the existing buffer before releasing the partition lock,
2814 : * preventing it from being evicted.
2815 : */
2816 0 : valid = PinBuffer(existing_hdr, strategy, false);
2817 :
2818 0 : LWLockRelease(partition_lock);
2819 0 : UnpinBuffer(victim_buf_hdr);
2820 :
2821 0 : buffers[i] = BufferDescriptorGetBuffer(existing_hdr);
2822 0 : buf_block = BufHdrGetBlock(existing_hdr);
2823 :
2824 0 : if (valid && !PageIsNew((Page) buf_block))
2825 0 : ereport(ERROR,
2826 : (errmsg("unexpected data beyond EOF in block %u of relation \"%s\"",
2827 : existing_hdr->tag.blockNum,
2828 : relpath(BMR_GET_SMGR(bmr)->smgr_rlocator, fork).str)));
2829 :
2830 : /*
2831 : * We *must* do smgr[zero]extend before succeeding, else the page
2832 : * will not be reserved by the kernel, and the next P_NEW call
2833 : * will decide to return the same page. Clear the BM_VALID bit,
2834 : * do StartBufferIO() and proceed.
2835 : *
2836 : * Loop to handle the very small possibility that someone re-sets
2837 : * BM_VALID between our clearing it and StartBufferIO inspecting
2838 : * it.
2839 : */
2840 : do
2841 : {
2842 0 : pg_atomic_fetch_and_u32(&existing_hdr->state, ~BM_VALID);
2843 0 : } while (!StartBufferIO(existing_hdr, true, false));
2844 : }
2845 : else
2846 : {
2847 : uint32 buf_state;
2848 448076 : uint32 set_bits = 0;
2849 :
2850 448076 : buf_state = LockBufHdr(victim_buf_hdr);
2851 :
2852 : /* some sanity checks while we hold the buffer header lock */
2853 : Assert(!(buf_state & (BM_VALID | BM_TAG_VALID | BM_DIRTY | BM_JUST_DIRTIED)));
2854 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) == 1);
2855 :
2856 448076 : victim_buf_hdr->tag = tag;
2857 :
2858 448076 : set_bits |= BM_TAG_VALID | BUF_USAGECOUNT_ONE;
2859 448076 : if (bmr.relpersistence == RELPERSISTENCE_PERMANENT || fork == INIT_FORKNUM)
2860 437422 : set_bits |= BM_PERMANENT;
2861 :
2862 448076 : UnlockBufHdrExt(victim_buf_hdr, buf_state,
2863 : set_bits, 0,
2864 : 0);
2865 :
2866 448076 : LWLockRelease(partition_lock);
2867 :
2868 : /* XXX: could combine the locked operations in it with the above */
2869 448076 : StartBufferIO(victim_buf_hdr, true, false);
2870 : }
2871 : }
2872 :
2873 405322 : io_start = pgstat_prepare_io_time(track_io_timing);
2874 :
2875 : /*
2876 : * Note: if smgrzeroextend fails, we will end up with buffers that are
2877 : * allocated but not marked BM_VALID. The next relation extension will
2878 : * still select the same block number (because the relation didn't get any
2879 : * longer on disk) and so future attempts to extend the relation will find
2880 : * the same buffers (if they have not been recycled) but come right back
2881 : * here to try smgrzeroextend again.
2882 : *
2883 : * We don't need to set checksum for all-zero pages.
2884 : */
2885 405322 : smgrzeroextend(BMR_GET_SMGR(bmr), fork, first_block, extend_by, false);
2886 :
2887 : /*
2888 : * Release the file-extension lock; it's now OK for someone else to extend
2889 : * the relation some more.
2890 : *
2891 : * We remove IO_IN_PROGRESS after this, as waking up waiting backends can
2892 : * take noticeable time.
2893 : */
2894 405322 : if (!(flags & EB_SKIP_EXTENSION_LOCK))
2895 301626 : UnlockRelationForExtension(bmr.rel, ExclusiveLock);
2896 :
2897 405322 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context, IOOP_EXTEND,
2898 405322 : io_start, 1, extend_by * BLCKSZ);
2899 :
2900 : /* Set BM_VALID, terminate IO, and wake up any waiters */
2901 853398 : for (uint32 i = 0; i < extend_by; i++)
2902 : {
2903 448076 : Buffer buf = buffers[i];
2904 448076 : BufferDesc *buf_hdr = GetBufferDescriptor(buf - 1);
2905 448076 : bool lock = false;
2906 :
2907 448076 : if (flags & EB_LOCK_FIRST && i == 0)
2908 297428 : lock = true;
2909 150648 : else if (flags & EB_LOCK_TARGET)
2910 : {
2911 : Assert(extend_upto != InvalidBlockNumber);
2912 89430 : if (first_block + i + 1 == extend_upto)
2913 88528 : lock = true;
2914 : }
2915 :
2916 448076 : if (lock)
2917 385956 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
2918 :
2919 448076 : TerminateBufferIO(buf_hdr, false, BM_VALID, true, false);
2920 : }
2921 :
2922 405322 : pgBufferUsage.shared_blks_written += extend_by;
2923 :
2924 405322 : *extended_by = extend_by;
2925 :
2926 405322 : return first_block;
2927 : }
2928 :
2929 : /*
2930 : * BufferIsLockedByMe
2931 : *
2932 : * Checks if this backend has the buffer locked in any mode.
2933 : *
2934 : * Buffer must be pinned.
2935 : */
2936 : bool
2937 0 : BufferIsLockedByMe(Buffer buffer)
2938 : {
2939 : BufferDesc *bufHdr;
2940 :
2941 : Assert(BufferIsPinned(buffer));
2942 :
2943 0 : if (BufferIsLocal(buffer))
2944 : {
2945 : /* Content locks are not maintained for local buffers. */
2946 0 : return true;
2947 : }
2948 : else
2949 : {
2950 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2951 0 : return LWLockHeldByMe(BufferDescriptorGetContentLock(bufHdr));
2952 : }
2953 : }
2954 :
2955 : /*
2956 : * BufferIsLockedByMeInMode
2957 : *
2958 : * Checks if this backend has the buffer locked in the specified mode.
2959 : *
2960 : * Buffer must be pinned.
2961 : */
2962 : bool
2963 0 : BufferIsLockedByMeInMode(Buffer buffer, BufferLockMode mode)
2964 : {
2965 : BufferDesc *bufHdr;
2966 :
2967 : Assert(BufferIsPinned(buffer));
2968 :
2969 0 : if (BufferIsLocal(buffer))
2970 : {
2971 : /* Content locks are not maintained for local buffers. */
2972 0 : return true;
2973 : }
2974 : else
2975 : {
2976 : LWLockMode lw_mode;
2977 :
2978 0 : switch (mode)
2979 : {
2980 0 : case BUFFER_LOCK_EXCLUSIVE:
2981 0 : lw_mode = LW_EXCLUSIVE;
2982 0 : break;
2983 0 : case BUFFER_LOCK_SHARE:
2984 0 : lw_mode = LW_SHARED;
2985 0 : break;
2986 0 : default:
2987 0 : pg_unreachable();
2988 : }
2989 :
2990 0 : bufHdr = GetBufferDescriptor(buffer - 1);
2991 0 : return LWLockHeldByMeInMode(BufferDescriptorGetContentLock(bufHdr),
2992 : lw_mode);
2993 : }
2994 : }
2995 :
2996 : /*
2997 : * BufferIsDirty
2998 : *
2999 : * Checks if buffer is already dirty.
3000 : *
3001 : * Buffer must be pinned and exclusive-locked. (Without an exclusive lock,
3002 : * the result may be stale before it's returned.)
3003 : */
3004 : bool
3005 0 : BufferIsDirty(Buffer buffer)
3006 : {
3007 : BufferDesc *bufHdr;
3008 :
3009 : Assert(BufferIsPinned(buffer));
3010 :
3011 0 : if (BufferIsLocal(buffer))
3012 : {
3013 0 : int bufid = -buffer - 1;
3014 :
3015 0 : bufHdr = GetLocalBufferDescriptor(bufid);
3016 : /* Content locks are not maintained for local buffers. */
3017 : }
3018 : else
3019 : {
3020 0 : bufHdr = GetBufferDescriptor(buffer - 1);
3021 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
3022 : }
3023 :
3024 0 : return pg_atomic_read_u32(&bufHdr->state) & BM_DIRTY;
3025 : }
3026 :
3027 : /*
3028 : * MarkBufferDirty
3029 : *
3030 : * Marks buffer contents as dirty (actual write happens later).
3031 : *
3032 : * Buffer must be pinned and exclusive-locked. (If caller does not hold
3033 : * exclusive lock, then somebody could be in process of writing the buffer,
3034 : * leading to risk of bad data written to disk.)
3035 : */
3036 : void
3037 43593960 : MarkBufferDirty(Buffer buffer)
3038 : {
3039 : BufferDesc *bufHdr;
3040 : uint32 buf_state;
3041 : uint32 old_buf_state;
3042 :
3043 43593960 : if (!BufferIsValid(buffer))
3044 0 : elog(ERROR, "bad buffer ID: %d", buffer);
3045 :
3046 43593960 : if (BufferIsLocal(buffer))
3047 : {
3048 2444634 : MarkLocalBufferDirty(buffer);
3049 2444634 : return;
3050 : }
3051 :
3052 41149326 : bufHdr = GetBufferDescriptor(buffer - 1);
3053 :
3054 : Assert(BufferIsPinned(buffer));
3055 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
3056 :
3057 : /*
3058 : * NB: We have to wait for the buffer header spinlock to be not held, as
3059 : * TerminateBufferIO() relies on the spinlock.
3060 : */
3061 41149326 : old_buf_state = pg_atomic_read_u32(&bufHdr->state);
3062 : for (;;)
3063 : {
3064 41149556 : if (old_buf_state & BM_LOCKED)
3065 10 : old_buf_state = WaitBufHdrUnlocked(bufHdr);
3066 :
3067 41149556 : buf_state = old_buf_state;
3068 :
3069 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
3070 41149556 : buf_state |= BM_DIRTY | BM_JUST_DIRTIED;
3071 :
3072 41149556 : if (pg_atomic_compare_exchange_u32(&bufHdr->state, &old_buf_state,
3073 : buf_state))
3074 41149326 : break;
3075 : }
3076 :
3077 : /*
3078 : * If the buffer was not dirty already, do vacuum accounting.
3079 : */
3080 41149326 : if (!(old_buf_state & BM_DIRTY))
3081 : {
3082 1321466 : pgBufferUsage.shared_blks_dirtied++;
3083 1321466 : if (VacuumCostActive)
3084 19300 : VacuumCostBalance += VacuumCostPageDirty;
3085 : }
3086 : }
3087 :
3088 : /*
3089 : * ReleaseAndReadBuffer -- combine ReleaseBuffer() and ReadBuffer()
3090 : *
3091 : * Formerly, this saved one cycle of acquiring/releasing the BufMgrLock
3092 : * compared to calling the two routines separately. Now it's mainly just
3093 : * a convenience function. However, if the passed buffer is valid and
3094 : * already contains the desired block, we just return it as-is; and that
3095 : * does save considerable work compared to a full release and reacquire.
3096 : *
3097 : * Note: it is OK to pass buffer == InvalidBuffer, indicating that no old
3098 : * buffer actually needs to be released. This case is the same as ReadBuffer,
3099 : * but can save some tests in the caller.
3100 : */
3101 : Buffer
3102 58919056 : ReleaseAndReadBuffer(Buffer buffer,
3103 : Relation relation,
3104 : BlockNumber blockNum)
3105 : {
3106 58919056 : ForkNumber forkNum = MAIN_FORKNUM;
3107 : BufferDesc *bufHdr;
3108 :
3109 58919056 : if (BufferIsValid(buffer))
3110 : {
3111 : Assert(BufferIsPinned(buffer));
3112 35504972 : if (BufferIsLocal(buffer))
3113 : {
3114 73728 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
3115 80772 : if (bufHdr->tag.blockNum == blockNum &&
3116 14088 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3117 7044 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3118 7044 : return buffer;
3119 66684 : UnpinLocalBuffer(buffer);
3120 : }
3121 : else
3122 : {
3123 35431244 : bufHdr = GetBufferDescriptor(buffer - 1);
3124 : /* we have pin, so it's ok to examine tag without spinlock */
3125 47508648 : if (bufHdr->tag.blockNum == blockNum &&
3126 24154808 : BufTagMatchesRelFileLocator(&bufHdr->tag, &relation->rd_locator) &&
3127 12077404 : BufTagGetForkNum(&bufHdr->tag) == forkNum)
3128 12077404 : return buffer;
3129 23353840 : UnpinBuffer(bufHdr);
3130 : }
3131 : }
3132 :
3133 46834608 : return ReadBuffer(relation, blockNum);
3134 : }
3135 :
3136 : /*
3137 : * PinBuffer -- make buffer unavailable for replacement.
3138 : *
3139 : * For the default access strategy, the buffer's usage_count is incremented
3140 : * when we first pin it; for other strategies we just make sure the usage_count
3141 : * isn't zero. (The idea of the latter is that we don't want synchronized
3142 : * heap scans to inflate the count, but we need it to not be zero to discourage
3143 : * other backends from stealing buffers from our ring. As long as we cycle
3144 : * through the ring faster than the global clock-sweep cycles, buffers in
3145 : * our ring won't be chosen as victims for replacement by other backends.)
3146 : *
3147 : * This should be applied only to shared buffers, never local ones.
3148 : *
3149 : * Since buffers are pinned/unpinned very frequently, pin buffers without
3150 : * taking the buffer header lock; instead update the state variable in loop of
3151 : * CAS operations. Hopefully it's just a single CAS.
3152 : *
3153 : * Note that ResourceOwnerEnlarge() and ReservePrivateRefCountEntry()
3154 : * must have been done already.
3155 : *
3156 : * Returns true if buffer is BM_VALID, else false. This provision allows
3157 : * some callers to avoid an extra spinlock cycle. If skip_if_not_valid is
3158 : * true, then a false return value also indicates that the buffer was
3159 : * (recently) invalid and has not been pinned.
3160 : */
3161 : static bool
3162 119131242 : PinBuffer(BufferDesc *buf, BufferAccessStrategy strategy,
3163 : bool skip_if_not_valid)
3164 : {
3165 119131242 : Buffer b = BufferDescriptorGetBuffer(buf);
3166 : bool result;
3167 : PrivateRefCountEntry *ref;
3168 :
3169 : Assert(!BufferIsLocal(b));
3170 : Assert(ReservedRefCountSlot != -1);
3171 :
3172 119131242 : ref = GetPrivateRefCountEntry(b, true);
3173 :
3174 119131242 : if (ref == NULL)
3175 : {
3176 : uint32 buf_state;
3177 : uint32 old_buf_state;
3178 :
3179 114595618 : old_buf_state = pg_atomic_read_u32(&buf->state);
3180 : for (;;)
3181 : {
3182 114635590 : if (unlikely(skip_if_not_valid && !(old_buf_state & BM_VALID)))
3183 12 : return false;
3184 :
3185 : /*
3186 : * We're not allowed to increase the refcount while the buffer
3187 : * header spinlock is held. Wait for the lock to be released.
3188 : */
3189 114635578 : if (old_buf_state & BM_LOCKED)
3190 1226 : old_buf_state = WaitBufHdrUnlocked(buf);
3191 :
3192 114635578 : buf_state = old_buf_state;
3193 :
3194 : /* increase refcount */
3195 114635578 : buf_state += BUF_REFCOUNT_ONE;
3196 :
3197 114635578 : if (strategy == NULL)
3198 : {
3199 : /* Default case: increase usagecount unless already max. */
3200 113267570 : if (BUF_STATE_GET_USAGECOUNT(buf_state) < BM_MAX_USAGE_COUNT)
3201 6799782 : buf_state += BUF_USAGECOUNT_ONE;
3202 : }
3203 : else
3204 : {
3205 : /*
3206 : * Ring buffers shouldn't evict others from pool. Thus we
3207 : * don't make usagecount more than 1.
3208 : */
3209 1368008 : if (BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
3210 71932 : buf_state += BUF_USAGECOUNT_ONE;
3211 : }
3212 :
3213 114635578 : if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
3214 : buf_state))
3215 : {
3216 114595606 : result = (buf_state & BM_VALID) != 0;
3217 :
3218 114595606 : TrackNewBufferPin(b);
3219 114595606 : break;
3220 : }
3221 : }
3222 : }
3223 : else
3224 : {
3225 : /*
3226 : * If we previously pinned the buffer, it is likely to be valid, but
3227 : * it may not be if StartReadBuffers() was called and
3228 : * WaitReadBuffers() hasn't been called yet. We'll check by loading
3229 : * the flags without locking. This is racy, but it's OK to return
3230 : * false spuriously: when WaitReadBuffers() calls StartBufferIO(),
3231 : * it'll see that it's now valid.
3232 : *
3233 : * Note: We deliberately avoid a Valgrind client request here.
3234 : * Individual access methods can optionally superimpose buffer page
3235 : * client requests on top of our client requests to enforce that
3236 : * buffers are only accessed while locked (and pinned). It's possible
3237 : * that the buffer page is legitimately non-accessible here. We
3238 : * cannot meddle with that.
3239 : */
3240 4535624 : result = (pg_atomic_read_u32(&buf->state) & BM_VALID) != 0;
3241 :
3242 : Assert(ref->data.refcount > 0);
3243 4535624 : ref->data.refcount++;
3244 4535624 : ResourceOwnerRememberBuffer(CurrentResourceOwner, b);
3245 : }
3246 :
3247 119131230 : return result;
3248 : }
3249 :
3250 : /*
3251 : * PinBuffer_Locked -- as above, but caller already locked the buffer header.
3252 : * The spinlock is released before return.
3253 : *
3254 : * As this function is called with the spinlock held, the caller has to
3255 : * previously call ReservePrivateRefCountEntry() and
3256 : * ResourceOwnerEnlarge(CurrentResourceOwner);
3257 : *
3258 : * Currently, no callers of this function want to modify the buffer's
3259 : * usage_count at all, so there's no need for a strategy parameter.
3260 : * Also we don't bother with a BM_VALID test (the caller could check that for
3261 : * itself).
3262 : *
3263 : * Also all callers only ever use this function when it's known that the
3264 : * buffer can't have a preexisting pin by this backend. That allows us to skip
3265 : * searching the private refcount array & hash, which is a boon, because the
3266 : * spinlock is still held.
3267 : *
3268 : * Note: use of this routine is frequently mandatory, not just an optimization
3269 : * to save a spin lock/unlock cycle, because we need to pin a buffer before
3270 : * its state can change under us.
3271 : */
3272 : static void
3273 613178 : PinBuffer_Locked(BufferDesc *buf)
3274 : {
3275 : uint32 old_buf_state;
3276 :
3277 : /*
3278 : * As explained, We don't expect any preexisting pins. That allows us to
3279 : * manipulate the PrivateRefCount after releasing the spinlock
3280 : */
3281 : Assert(GetPrivateRefCountEntry(BufferDescriptorGetBuffer(buf), false) == NULL);
3282 :
3283 : /*
3284 : * Since we hold the buffer spinlock, we can update the buffer state and
3285 : * release the lock in one operation.
3286 : */
3287 613178 : old_buf_state = pg_atomic_read_u32(&buf->state);
3288 :
3289 613178 : UnlockBufHdrExt(buf, old_buf_state,
3290 : 0, 0, 1);
3291 :
3292 613178 : TrackNewBufferPin(BufferDescriptorGetBuffer(buf));
3293 613178 : }
3294 :
3295 : /*
3296 : * Support for waking up another backend that is waiting for the cleanup lock
3297 : * to be released using BM_PIN_COUNT_WAITER.
3298 : *
3299 : * See LockBufferForCleanup().
3300 : *
3301 : * Expected to be called just after releasing a buffer pin (in a BufferDesc,
3302 : * not just reducing the backend-local pincount for the buffer).
3303 : */
3304 : static void
3305 184 : WakePinCountWaiter(BufferDesc *buf)
3306 : {
3307 : /*
3308 : * Acquire the buffer header lock, re-check that there's a waiter. Another
3309 : * backend could have unpinned this buffer, and already woken up the
3310 : * waiter.
3311 : *
3312 : * There's no danger of the buffer being replaced after we unpinned it
3313 : * above, as it's pinned by the waiter. The waiter removes
3314 : * BM_PIN_COUNT_WAITER if it stops waiting for a reason other than this
3315 : * backend waking it up.
3316 : */
3317 184 : uint32 buf_state = LockBufHdr(buf);
3318 :
3319 184 : if ((buf_state & BM_PIN_COUNT_WAITER) &&
3320 184 : BUF_STATE_GET_REFCOUNT(buf_state) == 1)
3321 184 : {
3322 : /* we just released the last pin other than the waiter's */
3323 184 : int wait_backend_pgprocno = buf->wait_backend_pgprocno;
3324 :
3325 184 : UnlockBufHdrExt(buf, buf_state,
3326 : 0, BM_PIN_COUNT_WAITER,
3327 : 0);
3328 184 : ProcSendSignal(wait_backend_pgprocno);
3329 : }
3330 : else
3331 0 : UnlockBufHdr(buf);
3332 184 : }
3333 :
3334 : /*
3335 : * UnpinBuffer -- make buffer available for replacement.
3336 : *
3337 : * This should be applied only to shared buffers, never local ones. This
3338 : * always adjusts CurrentResourceOwner.
3339 : */
3340 : static void
3341 146786806 : UnpinBuffer(BufferDesc *buf)
3342 : {
3343 146786806 : Buffer b = BufferDescriptorGetBuffer(buf);
3344 :
3345 146786806 : ResourceOwnerForgetBuffer(CurrentResourceOwner, b);
3346 146786806 : UnpinBufferNoOwner(buf);
3347 146786806 : }
3348 :
3349 : static void
3350 146795944 : UnpinBufferNoOwner(BufferDesc *buf)
3351 : {
3352 : PrivateRefCountEntry *ref;
3353 146795944 : Buffer b = BufferDescriptorGetBuffer(buf);
3354 :
3355 : Assert(!BufferIsLocal(b));
3356 :
3357 : /* not moving as we're likely deleting it soon anyway */
3358 146795944 : ref = GetPrivateRefCountEntry(b, false);
3359 : Assert(ref != NULL);
3360 : Assert(ref->data.refcount > 0);
3361 146795944 : ref->data.refcount--;
3362 146795944 : if (ref->data.refcount == 0)
3363 : {
3364 : uint32 old_buf_state;
3365 :
3366 : /*
3367 : * Mark buffer non-accessible to Valgrind.
3368 : *
3369 : * Note that the buffer may have already been marked non-accessible
3370 : * within access method code that enforces that buffers are only
3371 : * accessed while a buffer lock is held.
3372 : */
3373 : VALGRIND_MAKE_MEM_NOACCESS(BufHdrGetBlock(buf), BLCKSZ);
3374 :
3375 : /*
3376 : * I'd better not still hold the buffer content lock. Can't use
3377 : * BufferIsLockedByMe(), as that asserts the buffer is pinned.
3378 : */
3379 : Assert(!LWLockHeldByMe(BufferDescriptorGetContentLock(buf)));
3380 :
3381 : /* decrement the shared reference count */
3382 119126280 : old_buf_state = pg_atomic_fetch_sub_u32(&buf->state, BUF_REFCOUNT_ONE);
3383 :
3384 : /* Support LockBufferForCleanup() */
3385 119126280 : if (old_buf_state & BM_PIN_COUNT_WAITER)
3386 184 : WakePinCountWaiter(buf);
3387 :
3388 119126280 : ForgetPrivateRefCountEntry(ref);
3389 : }
3390 146795944 : }
3391 :
3392 : /*
3393 : * Set up backend-local tracking of a buffer pinned the first time by this
3394 : * backend.
3395 : */
3396 : inline void
3397 119126280 : TrackNewBufferPin(Buffer buf)
3398 : {
3399 : PrivateRefCountEntry *ref;
3400 :
3401 119126280 : ref = NewPrivateRefCountEntry(buf);
3402 119126280 : ref->data.refcount++;
3403 :
3404 119126280 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buf);
3405 :
3406 : /*
3407 : * This is the first pin for this page by this backend, mark its page as
3408 : * defined to valgrind. While the page contents might not actually be
3409 : * valid yet, we don't currently guarantee that such pages are marked
3410 : * undefined or non-accessible.
3411 : *
3412 : * It's not necessarily the prettiest to do this here, but otherwise we'd
3413 : * need this block of code in multiple places.
3414 : */
3415 : VALGRIND_MAKE_MEM_DEFINED(BufHdrGetBlock(GetBufferDescriptor(buf - 1)),
3416 : BLCKSZ);
3417 119126280 : }
3418 :
3419 : #define ST_SORT sort_checkpoint_bufferids
3420 : #define ST_ELEMENT_TYPE CkptSortItem
3421 : #define ST_COMPARE(a, b) ckpt_buforder_comparator(a, b)
3422 : #define ST_SCOPE static
3423 : #define ST_DEFINE
3424 : #include "lib/sort_template.h"
3425 :
3426 : /*
3427 : * BufferSync -- Write out all dirty buffers in the pool.
3428 : *
3429 : * This is called at checkpoint time to write out all dirty shared buffers.
3430 : * The checkpoint request flags should be passed in. If CHECKPOINT_FAST is
3431 : * set, we disable delays between writes; if CHECKPOINT_IS_SHUTDOWN,
3432 : * CHECKPOINT_END_OF_RECOVERY or CHECKPOINT_FLUSH_UNLOGGED is set, we write
3433 : * even unlogged buffers, which are otherwise skipped. The remaining flags
3434 : * currently have no effect here.
3435 : */
3436 : static void
3437 3552 : BufferSync(int flags)
3438 : {
3439 : uint32 buf_state;
3440 : int buf_id;
3441 : int num_to_scan;
3442 : int num_spaces;
3443 : int num_processed;
3444 : int num_written;
3445 3552 : CkptTsStatus *per_ts_stat = NULL;
3446 : Oid last_tsid;
3447 : binaryheap *ts_heap;
3448 : int i;
3449 3552 : uint32 mask = BM_DIRTY;
3450 : WritebackContext wb_context;
3451 :
3452 : /*
3453 : * Unless this is a shutdown checkpoint or we have been explicitly told,
3454 : * we write only permanent, dirty buffers. But at shutdown or end of
3455 : * recovery, we write all dirty buffers.
3456 : */
3457 3552 : if (!((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_END_OF_RECOVERY |
3458 : CHECKPOINT_FLUSH_UNLOGGED))))
3459 1980 : mask |= BM_PERMANENT;
3460 :
3461 : /*
3462 : * Loop over all buffers, and mark the ones that need to be written with
3463 : * BM_CHECKPOINT_NEEDED. Count them as we go (num_to_scan), so that we
3464 : * can estimate how much work needs to be done.
3465 : *
3466 : * This allows us to write only those pages that were dirty when the
3467 : * checkpoint began, and not those that get dirtied while it proceeds.
3468 : * Whenever a page with BM_CHECKPOINT_NEEDED is written out, either by us
3469 : * later in this function, or by normal backends or the bgwriter cleaning
3470 : * scan, the flag is cleared. Any buffer dirtied after this point won't
3471 : * have the flag set.
3472 : *
3473 : * Note that if we fail to write some buffer, we may leave buffers with
3474 : * BM_CHECKPOINT_NEEDED still set. This is OK since any such buffer would
3475 : * certainly need to be written for the next checkpoint attempt, too.
3476 : */
3477 3552 : num_to_scan = 0;
3478 24744064 : for (buf_id = 0; buf_id < NBuffers; buf_id++)
3479 : {
3480 24740512 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
3481 24740512 : uint32 set_bits = 0;
3482 :
3483 : /*
3484 : * Header spinlock is enough to examine BM_DIRTY, see comment in
3485 : * SyncOneBuffer.
3486 : */
3487 24740512 : buf_state = LockBufHdr(bufHdr);
3488 :
3489 24740512 : if ((buf_state & mask) == mask)
3490 : {
3491 : CkptSortItem *item;
3492 :
3493 590996 : set_bits = BM_CHECKPOINT_NEEDED;
3494 :
3495 590996 : item = &CkptBufferIds[num_to_scan++];
3496 590996 : item->buf_id = buf_id;
3497 590996 : item->tsId = bufHdr->tag.spcOid;
3498 590996 : item->relNumber = BufTagGetRelNumber(&bufHdr->tag);
3499 590996 : item->forkNum = BufTagGetForkNum(&bufHdr->tag);
3500 590996 : item->blockNum = bufHdr->tag.blockNum;
3501 : }
3502 :
3503 24740512 : UnlockBufHdrExt(bufHdr, buf_state,
3504 : set_bits, 0,
3505 : 0);
3506 :
3507 : /* Check for barrier events in case NBuffers is large. */
3508 24740512 : if (ProcSignalBarrierPending)
3509 0 : ProcessProcSignalBarrier();
3510 : }
3511 :
3512 3552 : if (num_to_scan == 0)
3513 1380 : return; /* nothing to do */
3514 :
3515 2172 : WritebackContextInit(&wb_context, &checkpoint_flush_after);
3516 :
3517 : TRACE_POSTGRESQL_BUFFER_SYNC_START(NBuffers, num_to_scan);
3518 :
3519 : /*
3520 : * Sort buffers that need to be written to reduce the likelihood of random
3521 : * IO. The sorting is also important for the implementation of balancing
3522 : * writes between tablespaces. Without balancing writes we'd potentially
3523 : * end up writing to the tablespaces one-by-one; possibly overloading the
3524 : * underlying system.
3525 : */
3526 2172 : sort_checkpoint_bufferids(CkptBufferIds, num_to_scan);
3527 :
3528 2172 : num_spaces = 0;
3529 :
3530 : /*
3531 : * Allocate progress status for each tablespace with buffers that need to
3532 : * be flushed. This requires the to-be-flushed array to be sorted.
3533 : */
3534 2172 : last_tsid = InvalidOid;
3535 593168 : for (i = 0; i < num_to_scan; i++)
3536 : {
3537 : CkptTsStatus *s;
3538 : Oid cur_tsid;
3539 :
3540 590996 : cur_tsid = CkptBufferIds[i].tsId;
3541 :
3542 : /*
3543 : * Grow array of per-tablespace status structs, every time a new
3544 : * tablespace is found.
3545 : */
3546 590996 : if (last_tsid == InvalidOid || last_tsid != cur_tsid)
3547 3292 : {
3548 : Size sz;
3549 :
3550 3292 : num_spaces++;
3551 :
3552 : /*
3553 : * Not worth adding grow-by-power-of-2 logic here - even with a
3554 : * few hundred tablespaces this should be fine.
3555 : */
3556 3292 : sz = sizeof(CkptTsStatus) * num_spaces;
3557 :
3558 3292 : if (per_ts_stat == NULL)
3559 2172 : per_ts_stat = (CkptTsStatus *) palloc(sz);
3560 : else
3561 1120 : per_ts_stat = (CkptTsStatus *) repalloc(per_ts_stat, sz);
3562 :
3563 3292 : s = &per_ts_stat[num_spaces - 1];
3564 3292 : memset(s, 0, sizeof(*s));
3565 3292 : s->tsId = cur_tsid;
3566 :
3567 : /*
3568 : * The first buffer in this tablespace. As CkptBufferIds is sorted
3569 : * by tablespace all (s->num_to_scan) buffers in this tablespace
3570 : * will follow afterwards.
3571 : */
3572 3292 : s->index = i;
3573 :
3574 : /*
3575 : * progress_slice will be determined once we know how many buffers
3576 : * are in each tablespace, i.e. after this loop.
3577 : */
3578 :
3579 3292 : last_tsid = cur_tsid;
3580 : }
3581 : else
3582 : {
3583 587704 : s = &per_ts_stat[num_spaces - 1];
3584 : }
3585 :
3586 590996 : s->num_to_scan++;
3587 :
3588 : /* Check for barrier events. */
3589 590996 : if (ProcSignalBarrierPending)
3590 0 : ProcessProcSignalBarrier();
3591 : }
3592 :
3593 : Assert(num_spaces > 0);
3594 :
3595 : /*
3596 : * Build a min-heap over the write-progress in the individual tablespaces,
3597 : * and compute how large a portion of the total progress a single
3598 : * processed buffer is.
3599 : */
3600 2172 : ts_heap = binaryheap_allocate(num_spaces,
3601 : ts_ckpt_progress_comparator,
3602 : NULL);
3603 :
3604 5464 : for (i = 0; i < num_spaces; i++)
3605 : {
3606 3292 : CkptTsStatus *ts_stat = &per_ts_stat[i];
3607 :
3608 3292 : ts_stat->progress_slice = (float8) num_to_scan / ts_stat->num_to_scan;
3609 :
3610 3292 : binaryheap_add_unordered(ts_heap, PointerGetDatum(ts_stat));
3611 : }
3612 :
3613 2172 : binaryheap_build(ts_heap);
3614 :
3615 : /*
3616 : * Iterate through to-be-checkpointed buffers and write the ones (still)
3617 : * marked with BM_CHECKPOINT_NEEDED. The writes are balanced between
3618 : * tablespaces; otherwise the sorting would lead to only one tablespace
3619 : * receiving writes at a time, making inefficient use of the hardware.
3620 : */
3621 2172 : num_processed = 0;
3622 2172 : num_written = 0;
3623 593168 : while (!binaryheap_empty(ts_heap))
3624 : {
3625 590996 : BufferDesc *bufHdr = NULL;
3626 : CkptTsStatus *ts_stat = (CkptTsStatus *)
3627 590996 : DatumGetPointer(binaryheap_first(ts_heap));
3628 :
3629 590996 : buf_id = CkptBufferIds[ts_stat->index].buf_id;
3630 : Assert(buf_id != -1);
3631 :
3632 590996 : bufHdr = GetBufferDescriptor(buf_id);
3633 :
3634 590996 : num_processed++;
3635 :
3636 : /*
3637 : * We don't need to acquire the lock here, because we're only looking
3638 : * at a single bit. It's possible that someone else writes the buffer
3639 : * and clears the flag right after we check, but that doesn't matter
3640 : * since SyncOneBuffer will then do nothing. However, there is a
3641 : * further race condition: it's conceivable that between the time we
3642 : * examine the bit here and the time SyncOneBuffer acquires the lock,
3643 : * someone else not only wrote the buffer but replaced it with another
3644 : * page and dirtied it. In that improbable case, SyncOneBuffer will
3645 : * write the buffer though we didn't need to. It doesn't seem worth
3646 : * guarding against this, though.
3647 : */
3648 590996 : if (pg_atomic_read_u32(&bufHdr->state) & BM_CHECKPOINT_NEEDED)
3649 : {
3650 553406 : if (SyncOneBuffer(buf_id, false, &wb_context) & BUF_WRITTEN)
3651 : {
3652 : TRACE_POSTGRESQL_BUFFER_SYNC_WRITTEN(buf_id);
3653 553406 : PendingCheckpointerStats.buffers_written++;
3654 553406 : num_written++;
3655 : }
3656 : }
3657 :
3658 : /*
3659 : * Measure progress independent of actually having to flush the buffer
3660 : * - otherwise writing become unbalanced.
3661 : */
3662 590996 : ts_stat->progress += ts_stat->progress_slice;
3663 590996 : ts_stat->num_scanned++;
3664 590996 : ts_stat->index++;
3665 :
3666 : /* Have all the buffers from the tablespace been processed? */
3667 590996 : if (ts_stat->num_scanned == ts_stat->num_to_scan)
3668 : {
3669 3292 : binaryheap_remove_first(ts_heap);
3670 : }
3671 : else
3672 : {
3673 : /* update heap with the new progress */
3674 587704 : binaryheap_replace_first(ts_heap, PointerGetDatum(ts_stat));
3675 : }
3676 :
3677 : /*
3678 : * Sleep to throttle our I/O rate.
3679 : *
3680 : * (This will check for barrier events even if it doesn't sleep.)
3681 : */
3682 590996 : CheckpointWriteDelay(flags, (double) num_processed / num_to_scan);
3683 : }
3684 :
3685 : /*
3686 : * Issue all pending flushes. Only checkpointer calls BufferSync(), so
3687 : * IOContext will always be IOCONTEXT_NORMAL.
3688 : */
3689 2172 : IssuePendingWritebacks(&wb_context, IOCONTEXT_NORMAL);
3690 :
3691 2172 : pfree(per_ts_stat);
3692 2172 : per_ts_stat = NULL;
3693 2172 : binaryheap_free(ts_heap);
3694 :
3695 : /*
3696 : * Update checkpoint statistics. As noted above, this doesn't include
3697 : * buffers written by other backends or bgwriter scan.
3698 : */
3699 2172 : CheckpointStats.ckpt_bufs_written += num_written;
3700 :
3701 : TRACE_POSTGRESQL_BUFFER_SYNC_DONE(NBuffers, num_written, num_to_scan);
3702 : }
3703 :
3704 : /*
3705 : * BgBufferSync -- Write out some dirty buffers in the pool.
3706 : *
3707 : * This is called periodically by the background writer process.
3708 : *
3709 : * Returns true if it's appropriate for the bgwriter process to go into
3710 : * low-power hibernation mode. (This happens if the strategy clock-sweep
3711 : * has been "lapped" and no buffer allocations have occurred recently,
3712 : * or if the bgwriter has been effectively disabled by setting
3713 : * bgwriter_lru_maxpages to 0.)
3714 : */
3715 : bool
3716 28170 : BgBufferSync(WritebackContext *wb_context)
3717 : {
3718 : /* info obtained from freelist.c */
3719 : int strategy_buf_id;
3720 : uint32 strategy_passes;
3721 : uint32 recent_alloc;
3722 :
3723 : /*
3724 : * Information saved between calls so we can determine the strategy
3725 : * point's advance rate and avoid scanning already-cleaned buffers.
3726 : */
3727 : static bool saved_info_valid = false;
3728 : static int prev_strategy_buf_id;
3729 : static uint32 prev_strategy_passes;
3730 : static int next_to_clean;
3731 : static uint32 next_passes;
3732 :
3733 : /* Moving averages of allocation rate and clean-buffer density */
3734 : static float smoothed_alloc = 0;
3735 : static float smoothed_density = 10.0;
3736 :
3737 : /* Potentially these could be tunables, but for now, not */
3738 28170 : float smoothing_samples = 16;
3739 28170 : float scan_whole_pool_milliseconds = 120000.0;
3740 :
3741 : /* Used to compute how far we scan ahead */
3742 : long strategy_delta;
3743 : int bufs_to_lap;
3744 : int bufs_ahead;
3745 : float scans_per_alloc;
3746 : int reusable_buffers_est;
3747 : int upcoming_alloc_est;
3748 : int min_scan_buffers;
3749 :
3750 : /* Variables for the scanning loop proper */
3751 : int num_to_scan;
3752 : int num_written;
3753 : int reusable_buffers;
3754 :
3755 : /* Variables for final smoothed_density update */
3756 : long new_strategy_delta;
3757 : uint32 new_recent_alloc;
3758 :
3759 : /*
3760 : * Find out where the clock-sweep currently is, and how many buffer
3761 : * allocations have happened since our last call.
3762 : */
3763 28170 : strategy_buf_id = StrategySyncStart(&strategy_passes, &recent_alloc);
3764 :
3765 : /* Report buffer alloc counts to pgstat */
3766 28170 : PendingBgWriterStats.buf_alloc += recent_alloc;
3767 :
3768 : /*
3769 : * If we're not running the LRU scan, just stop after doing the stats
3770 : * stuff. We mark the saved state invalid so that we can recover sanely
3771 : * if LRU scan is turned back on later.
3772 : */
3773 28170 : if (bgwriter_lru_maxpages <= 0)
3774 : {
3775 96 : saved_info_valid = false;
3776 96 : return true;
3777 : }
3778 :
3779 : /*
3780 : * Compute strategy_delta = how many buffers have been scanned by the
3781 : * clock-sweep since last time. If first time through, assume none. Then
3782 : * see if we are still ahead of the clock-sweep, and if so, how many
3783 : * buffers we could scan before we'd catch up with it and "lap" it. Note:
3784 : * weird-looking coding of xxx_passes comparisons are to avoid bogus
3785 : * behavior when the passes counts wrap around.
3786 : */
3787 28074 : if (saved_info_valid)
3788 : {
3789 26952 : int32 passes_delta = strategy_passes - prev_strategy_passes;
3790 :
3791 26952 : strategy_delta = strategy_buf_id - prev_strategy_buf_id;
3792 26952 : strategy_delta += (long) passes_delta * NBuffers;
3793 :
3794 : Assert(strategy_delta >= 0);
3795 :
3796 26952 : if ((int32) (next_passes - strategy_passes) > 0)
3797 : {
3798 : /* we're one pass ahead of the strategy point */
3799 6404 : bufs_to_lap = strategy_buf_id - next_to_clean;
3800 : #ifdef BGW_DEBUG
3801 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3802 : next_passes, next_to_clean,
3803 : strategy_passes, strategy_buf_id,
3804 : strategy_delta, bufs_to_lap);
3805 : #endif
3806 : }
3807 20548 : else if (next_passes == strategy_passes &&
3808 15768 : next_to_clean >= strategy_buf_id)
3809 : {
3810 : /* on same pass, but ahead or at least not behind */
3811 14194 : bufs_to_lap = NBuffers - (next_to_clean - strategy_buf_id);
3812 : #ifdef BGW_DEBUG
3813 : elog(DEBUG2, "bgwriter ahead: bgw %u-%u strategy %u-%u delta=%ld lap=%d",
3814 : next_passes, next_to_clean,
3815 : strategy_passes, strategy_buf_id,
3816 : strategy_delta, bufs_to_lap);
3817 : #endif
3818 : }
3819 : else
3820 : {
3821 : /*
3822 : * We're behind, so skip forward to the strategy point and start
3823 : * cleaning from there.
3824 : */
3825 : #ifdef BGW_DEBUG
3826 : elog(DEBUG2, "bgwriter behind: bgw %u-%u strategy %u-%u delta=%ld",
3827 : next_passes, next_to_clean,
3828 : strategy_passes, strategy_buf_id,
3829 : strategy_delta);
3830 : #endif
3831 6354 : next_to_clean = strategy_buf_id;
3832 6354 : next_passes = strategy_passes;
3833 6354 : bufs_to_lap = NBuffers;
3834 : }
3835 : }
3836 : else
3837 : {
3838 : /*
3839 : * Initializing at startup or after LRU scanning had been off. Always
3840 : * start at the strategy point.
3841 : */
3842 : #ifdef BGW_DEBUG
3843 : elog(DEBUG2, "bgwriter initializing: strategy %u-%u",
3844 : strategy_passes, strategy_buf_id);
3845 : #endif
3846 1122 : strategy_delta = 0;
3847 1122 : next_to_clean = strategy_buf_id;
3848 1122 : next_passes = strategy_passes;
3849 1122 : bufs_to_lap = NBuffers;
3850 : }
3851 :
3852 : /* Update saved info for next time */
3853 28074 : prev_strategy_buf_id = strategy_buf_id;
3854 28074 : prev_strategy_passes = strategy_passes;
3855 28074 : saved_info_valid = true;
3856 :
3857 : /*
3858 : * Compute how many buffers had to be scanned for each new allocation, ie,
3859 : * 1/density of reusable buffers, and track a moving average of that.
3860 : *
3861 : * If the strategy point didn't move, we don't update the density estimate
3862 : */
3863 28074 : if (strategy_delta > 0 && recent_alloc > 0)
3864 : {
3865 14722 : scans_per_alloc = (float) strategy_delta / (float) recent_alloc;
3866 14722 : smoothed_density += (scans_per_alloc - smoothed_density) /
3867 : smoothing_samples;
3868 : }
3869 :
3870 : /*
3871 : * Estimate how many reusable buffers there are between the current
3872 : * strategy point and where we've scanned ahead to, based on the smoothed
3873 : * density estimate.
3874 : */
3875 28074 : bufs_ahead = NBuffers - bufs_to_lap;
3876 28074 : reusable_buffers_est = (float) bufs_ahead / smoothed_density;
3877 :
3878 : /*
3879 : * Track a moving average of recent buffer allocations. Here, rather than
3880 : * a true average we want a fast-attack, slow-decline behavior: we
3881 : * immediately follow any increase.
3882 : */
3883 28074 : if (smoothed_alloc <= (float) recent_alloc)
3884 7332 : smoothed_alloc = recent_alloc;
3885 : else
3886 20742 : smoothed_alloc += ((float) recent_alloc - smoothed_alloc) /
3887 : smoothing_samples;
3888 :
3889 : /* Scale the estimate by a GUC to allow more aggressive tuning. */
3890 28074 : upcoming_alloc_est = (int) (smoothed_alloc * bgwriter_lru_multiplier);
3891 :
3892 : /*
3893 : * If recent_alloc remains at zero for many cycles, smoothed_alloc will
3894 : * eventually underflow to zero, and the underflows produce annoying
3895 : * kernel warnings on some platforms. Once upcoming_alloc_est has gone to
3896 : * zero, there's no point in tracking smaller and smaller values of
3897 : * smoothed_alloc, so just reset it to exactly zero to avoid this
3898 : * syndrome. It will pop back up as soon as recent_alloc increases.
3899 : */
3900 28074 : if (upcoming_alloc_est == 0)
3901 4354 : smoothed_alloc = 0;
3902 :
3903 : /*
3904 : * Even in cases where there's been little or no buffer allocation
3905 : * activity, we want to make a small amount of progress through the buffer
3906 : * cache so that as many reusable buffers as possible are clean after an
3907 : * idle period.
3908 : *
3909 : * (scan_whole_pool_milliseconds / BgWriterDelay) computes how many times
3910 : * the BGW will be called during the scan_whole_pool time; slice the
3911 : * buffer pool into that many sections.
3912 : */
3913 28074 : min_scan_buffers = (int) (NBuffers / (scan_whole_pool_milliseconds / BgWriterDelay));
3914 :
3915 28074 : if (upcoming_alloc_est < (min_scan_buffers + reusable_buffers_est))
3916 : {
3917 : #ifdef BGW_DEBUG
3918 : elog(DEBUG2, "bgwriter: alloc_est=%d too small, using min=%d + reusable_est=%d",
3919 : upcoming_alloc_est, min_scan_buffers, reusable_buffers_est);
3920 : #endif
3921 12924 : upcoming_alloc_est = min_scan_buffers + reusable_buffers_est;
3922 : }
3923 :
3924 : /*
3925 : * Now write out dirty reusable buffers, working forward from the
3926 : * next_to_clean point, until we have lapped the strategy scan, or cleaned
3927 : * enough buffers to match our estimate of the next cycle's allocation
3928 : * requirements, or hit the bgwriter_lru_maxpages limit.
3929 : */
3930 :
3931 28074 : num_to_scan = bufs_to_lap;
3932 28074 : num_written = 0;
3933 28074 : reusable_buffers = reusable_buffers_est;
3934 :
3935 : /* Execute the LRU scan */
3936 3704476 : while (num_to_scan > 0 && reusable_buffers < upcoming_alloc_est)
3937 : {
3938 3676406 : int sync_state = SyncOneBuffer(next_to_clean, true,
3939 : wb_context);
3940 :
3941 3676406 : if (++next_to_clean >= NBuffers)
3942 : {
3943 6082 : next_to_clean = 0;
3944 6082 : next_passes++;
3945 : }
3946 3676406 : num_to_scan--;
3947 :
3948 3676406 : if (sync_state & BUF_WRITTEN)
3949 : {
3950 50612 : reusable_buffers++;
3951 50612 : if (++num_written >= bgwriter_lru_maxpages)
3952 : {
3953 4 : PendingBgWriterStats.maxwritten_clean++;
3954 4 : break;
3955 : }
3956 : }
3957 3625794 : else if (sync_state & BUF_REUSABLE)
3958 2823024 : reusable_buffers++;
3959 : }
3960 :
3961 28074 : PendingBgWriterStats.buf_written_clean += num_written;
3962 :
3963 : #ifdef BGW_DEBUG
3964 : 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",
3965 : recent_alloc, smoothed_alloc, strategy_delta, bufs_ahead,
3966 : smoothed_density, reusable_buffers_est, upcoming_alloc_est,
3967 : bufs_to_lap - num_to_scan,
3968 : num_written,
3969 : reusable_buffers - reusable_buffers_est);
3970 : #endif
3971 :
3972 : /*
3973 : * Consider the above scan as being like a new allocation scan.
3974 : * Characterize its density and update the smoothed one based on it. This
3975 : * effectively halves the moving average period in cases where both the
3976 : * strategy and the background writer are doing some useful scanning,
3977 : * which is helpful because a long memory isn't as desirable on the
3978 : * density estimates.
3979 : */
3980 28074 : new_strategy_delta = bufs_to_lap - num_to_scan;
3981 28074 : new_recent_alloc = reusable_buffers - reusable_buffers_est;
3982 28074 : if (new_strategy_delta > 0 && new_recent_alloc > 0)
3983 : {
3984 22434 : scans_per_alloc = (float) new_strategy_delta / (float) new_recent_alloc;
3985 22434 : smoothed_density += (scans_per_alloc - smoothed_density) /
3986 : smoothing_samples;
3987 :
3988 : #ifdef BGW_DEBUG
3989 : elog(DEBUG2, "bgwriter: cleaner density alloc=%u scan=%ld density=%.2f new smoothed=%.2f",
3990 : new_recent_alloc, new_strategy_delta,
3991 : scans_per_alloc, smoothed_density);
3992 : #endif
3993 : }
3994 :
3995 : /* Return true if OK to hibernate */
3996 28074 : return (bufs_to_lap == 0 && recent_alloc == 0);
3997 : }
3998 :
3999 : /*
4000 : * SyncOneBuffer -- process a single buffer during syncing.
4001 : *
4002 : * If skip_recently_used is true, we don't write currently-pinned buffers, nor
4003 : * buffers marked recently used, as these are not replacement candidates.
4004 : *
4005 : * Returns a bitmask containing the following flag bits:
4006 : * BUF_WRITTEN: we wrote the buffer.
4007 : * BUF_REUSABLE: buffer is available for replacement, ie, it has
4008 : * pin count 0 and usage count 0.
4009 : *
4010 : * (BUF_WRITTEN could be set in error if FlushBuffer finds the buffer clean
4011 : * after locking it, but we don't care all that much.)
4012 : */
4013 : static int
4014 4229812 : SyncOneBuffer(int buf_id, bool skip_recently_used, WritebackContext *wb_context)
4015 : {
4016 4229812 : BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
4017 4229812 : int result = 0;
4018 : uint32 buf_state;
4019 : BufferTag tag;
4020 :
4021 : /* Make sure we can handle the pin */
4022 4229812 : ReservePrivateRefCountEntry();
4023 4229812 : ResourceOwnerEnlarge(CurrentResourceOwner);
4024 :
4025 : /*
4026 : * Check whether buffer needs writing.
4027 : *
4028 : * We can make this check without taking the buffer content lock so long
4029 : * as we mark pages dirty in access methods *before* logging changes with
4030 : * XLogInsert(): if someone marks the buffer dirty just after our check we
4031 : * don't worry because our checkpoint.redo points before log record for
4032 : * upcoming changes and so we are not required to write such dirty buffer.
4033 : */
4034 4229812 : buf_state = LockBufHdr(bufHdr);
4035 :
4036 4229812 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 0 &&
4037 4219284 : BUF_STATE_GET_USAGECOUNT(buf_state) == 0)
4038 : {
4039 2877386 : result |= BUF_REUSABLE;
4040 : }
4041 1352426 : else if (skip_recently_used)
4042 : {
4043 : /* Caller told us not to write recently-used buffers */
4044 802770 : UnlockBufHdr(bufHdr);
4045 802770 : return result;
4046 : }
4047 :
4048 3427042 : if (!(buf_state & BM_VALID) || !(buf_state & BM_DIRTY))
4049 : {
4050 : /* It's clean, so nothing to do */
4051 2823024 : UnlockBufHdr(bufHdr);
4052 2823024 : return result;
4053 : }
4054 :
4055 : /*
4056 : * Pin it, share-lock it, write it. (FlushBuffer will do nothing if the
4057 : * buffer is clean by the time we've locked it.)
4058 : */
4059 604018 : PinBuffer_Locked(bufHdr);
4060 :
4061 604018 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
4062 :
4063 604018 : tag = bufHdr->tag;
4064 :
4065 604018 : UnpinBuffer(bufHdr);
4066 :
4067 : /*
4068 : * SyncOneBuffer() is only called by checkpointer and bgwriter, so
4069 : * IOContext will always be IOCONTEXT_NORMAL.
4070 : */
4071 604018 : ScheduleBufferTagForWriteback(wb_context, IOCONTEXT_NORMAL, &tag);
4072 :
4073 604018 : return result | BUF_WRITTEN;
4074 : }
4075 :
4076 : /*
4077 : * AtEOXact_Buffers - clean up at end of transaction.
4078 : *
4079 : * As of PostgreSQL 8.0, buffer pins should get released by the
4080 : * ResourceOwner mechanism. This routine is just a debugging
4081 : * cross-check that no pins remain.
4082 : */
4083 : void
4084 1041938 : AtEOXact_Buffers(bool isCommit)
4085 : {
4086 1041938 : CheckForBufferLeaks();
4087 :
4088 1041938 : AtEOXact_LocalBuffers(isCommit);
4089 :
4090 : Assert(PrivateRefCountOverflowed == 0);
4091 1041938 : }
4092 :
4093 : /*
4094 : * Initialize access to shared buffer pool
4095 : *
4096 : * This is called during backend startup (whether standalone or under the
4097 : * postmaster). It sets up for this backend's access to the already-existing
4098 : * buffer pool.
4099 : */
4100 : void
4101 45876 : InitBufferManagerAccess(void)
4102 : {
4103 : HASHCTL hash_ctl;
4104 :
4105 : /*
4106 : * An advisory limit on the number of pins each backend should hold, based
4107 : * on shared_buffers and the maximum number of connections possible.
4108 : * That's very pessimistic, but outside toy-sized shared_buffers it should
4109 : * allow plenty of pins. LimitAdditionalPins() and
4110 : * GetAdditionalPinLimit() can be used to check the remaining balance.
4111 : */
4112 45876 : MaxProportionalPins = NBuffers / (MaxBackends + NUM_AUXILIARY_PROCS);
4113 :
4114 45876 : memset(&PrivateRefCountArray, 0, sizeof(PrivateRefCountArray));
4115 45876 : memset(&PrivateRefCountArrayKeys, 0, sizeof(PrivateRefCountArrayKeys));
4116 :
4117 45876 : hash_ctl.keysize = sizeof(Buffer);
4118 45876 : hash_ctl.entrysize = sizeof(PrivateRefCountEntry);
4119 :
4120 45876 : PrivateRefCountHash = hash_create("PrivateRefCount", 100, &hash_ctl,
4121 : HASH_ELEM | HASH_BLOBS);
4122 :
4123 : /*
4124 : * AtProcExit_Buffers needs LWLock access, and thereby has to be called at
4125 : * the corresponding phase of backend shutdown.
4126 : */
4127 : Assert(MyProc != NULL);
4128 45876 : on_shmem_exit(AtProcExit_Buffers, 0);
4129 45876 : }
4130 :
4131 : /*
4132 : * During backend exit, ensure that we released all shared-buffer locks and
4133 : * assert that we have no remaining pins.
4134 : */
4135 : static void
4136 45876 : AtProcExit_Buffers(int code, Datum arg)
4137 : {
4138 45876 : UnlockBuffers();
4139 :
4140 45876 : CheckForBufferLeaks();
4141 :
4142 : /* localbuf.c needs a chance too */
4143 45876 : AtProcExit_LocalBuffers();
4144 45876 : }
4145 :
4146 : /*
4147 : * CheckForBufferLeaks - ensure this backend holds no buffer pins
4148 : *
4149 : * As of PostgreSQL 8.0, buffer pins should get released by the
4150 : * ResourceOwner mechanism. This routine is just a debugging
4151 : * cross-check that no pins remain.
4152 : */
4153 : static void
4154 1087814 : CheckForBufferLeaks(void)
4155 : {
4156 : #ifdef USE_ASSERT_CHECKING
4157 : int RefCountErrors = 0;
4158 : PrivateRefCountEntry *res;
4159 : int i;
4160 : char *s;
4161 :
4162 : /* check the array */
4163 : for (i = 0; i < REFCOUNT_ARRAY_ENTRIES; i++)
4164 : {
4165 : if (PrivateRefCountArrayKeys[i] != InvalidBuffer)
4166 : {
4167 : res = &PrivateRefCountArray[i];
4168 :
4169 : s = DebugPrintBufferRefcount(res->buffer);
4170 : elog(WARNING, "buffer refcount leak: %s", s);
4171 : pfree(s);
4172 :
4173 : RefCountErrors++;
4174 : }
4175 : }
4176 :
4177 : /* if necessary search the hash */
4178 : if (PrivateRefCountOverflowed)
4179 : {
4180 : HASH_SEQ_STATUS hstat;
4181 :
4182 : hash_seq_init(&hstat, PrivateRefCountHash);
4183 : while ((res = (PrivateRefCountEntry *) hash_seq_search(&hstat)) != NULL)
4184 : {
4185 : s = DebugPrintBufferRefcount(res->buffer);
4186 : elog(WARNING, "buffer refcount leak: %s", s);
4187 : pfree(s);
4188 : RefCountErrors++;
4189 : }
4190 : }
4191 :
4192 : Assert(RefCountErrors == 0);
4193 : #endif
4194 1087814 : }
4195 :
4196 : #ifdef USE_ASSERT_CHECKING
4197 : /*
4198 : * Check for exclusive-locked catalog buffers. This is the core of
4199 : * AssertCouldGetRelation().
4200 : *
4201 : * A backend would self-deadlock on LWLocks if the catalog scan read the
4202 : * exclusive-locked buffer. The main threat is exclusive-locked buffers of
4203 : * catalogs used in relcache, because a catcache search on any catalog may
4204 : * build that catalog's relcache entry. We don't have an inventory of
4205 : * catalogs relcache uses, so just check buffers of most catalogs.
4206 : *
4207 : * It's better to minimize waits while holding an exclusive buffer lock, so it
4208 : * would be nice to broaden this check not to be catalog-specific. However,
4209 : * bttextcmp() accesses pg_collation, and non-core opclasses might similarly
4210 : * read tables. That is deadlock-free as long as there's no loop in the
4211 : * dependency graph: modifying table A may cause an opclass to read table B,
4212 : * but it must not cause a read of table A.
4213 : */
4214 : void
4215 : AssertBufferLocksPermitCatalogRead(void)
4216 : {
4217 : ForEachLWLockHeldByMe(AssertNotCatalogBufferLock, NULL);
4218 : }
4219 :
4220 : static void
4221 : AssertNotCatalogBufferLock(LWLock *lock, LWLockMode mode,
4222 : void *unused_context)
4223 : {
4224 : BufferDesc *bufHdr;
4225 : BufferTag tag;
4226 : Oid relid;
4227 :
4228 : if (mode != LW_EXCLUSIVE)
4229 : return;
4230 :
4231 : if (!((BufferDescPadded *) lock > BufferDescriptors &&
4232 : (BufferDescPadded *) lock < BufferDescriptors + NBuffers))
4233 : return; /* not a buffer lock */
4234 :
4235 : bufHdr = (BufferDesc *)
4236 : ((char *) lock - offsetof(BufferDesc, content_lock));
4237 : tag = bufHdr->tag;
4238 :
4239 : /*
4240 : * This relNumber==relid assumption holds until a catalog experiences
4241 : * VACUUM FULL or similar. After a command like that, relNumber will be
4242 : * in the normal (non-catalog) range, and we lose the ability to detect
4243 : * hazardous access to that catalog. Calling RelidByRelfilenumber() would
4244 : * close that gap, but RelidByRelfilenumber() might then deadlock with a
4245 : * held lock.
4246 : */
4247 : relid = tag.relNumber;
4248 :
4249 : if (IsCatalogTextUniqueIndexOid(relid)) /* see comments at the callee */
4250 : return;
4251 :
4252 : Assert(!IsCatalogRelationOid(relid));
4253 : }
4254 : #endif
4255 :
4256 :
4257 : /*
4258 : * Helper routine to issue warnings when a buffer is unexpectedly pinned
4259 : */
4260 : char *
4261 80 : DebugPrintBufferRefcount(Buffer buffer)
4262 : {
4263 : BufferDesc *buf;
4264 : int32 loccount;
4265 : char *result;
4266 : ProcNumber backend;
4267 : uint32 buf_state;
4268 :
4269 : Assert(BufferIsValid(buffer));
4270 80 : if (BufferIsLocal(buffer))
4271 : {
4272 32 : buf = GetLocalBufferDescriptor(-buffer - 1);
4273 32 : loccount = LocalRefCount[-buffer - 1];
4274 32 : backend = MyProcNumber;
4275 : }
4276 : else
4277 : {
4278 48 : buf = GetBufferDescriptor(buffer - 1);
4279 48 : loccount = GetPrivateRefCount(buffer);
4280 48 : backend = INVALID_PROC_NUMBER;
4281 : }
4282 :
4283 : /* theoretically we should lock the bufHdr here */
4284 80 : buf_state = pg_atomic_read_u32(&buf->state);
4285 :
4286 80 : result = psprintf("[%03d] (rel=%s, blockNum=%u, flags=0x%x, refcount=%u %d)",
4287 : buffer,
4288 80 : relpathbackend(BufTagGetRelFileLocator(&buf->tag), backend,
4289 : BufTagGetForkNum(&buf->tag)).str,
4290 : buf->tag.blockNum, buf_state & BUF_FLAG_MASK,
4291 : BUF_STATE_GET_REFCOUNT(buf_state), loccount);
4292 80 : return result;
4293 : }
4294 :
4295 : /*
4296 : * CheckPointBuffers
4297 : *
4298 : * Flush all dirty blocks in buffer pool to disk at checkpoint time.
4299 : *
4300 : * Note: temporary relations do not participate in checkpoints, so they don't
4301 : * need to be flushed.
4302 : */
4303 : void
4304 3552 : CheckPointBuffers(int flags)
4305 : {
4306 3552 : BufferSync(flags);
4307 3552 : }
4308 :
4309 : /*
4310 : * BufferGetBlockNumber
4311 : * Returns the block number associated with a buffer.
4312 : *
4313 : * Note:
4314 : * Assumes that the buffer is valid and pinned, else the
4315 : * value may be obsolete immediately...
4316 : */
4317 : BlockNumber
4318 100433424 : BufferGetBlockNumber(Buffer buffer)
4319 : {
4320 : BufferDesc *bufHdr;
4321 :
4322 : Assert(BufferIsPinned(buffer));
4323 :
4324 100433424 : if (BufferIsLocal(buffer))
4325 3808076 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4326 : else
4327 96625348 : bufHdr = GetBufferDescriptor(buffer - 1);
4328 :
4329 : /* pinned, so OK to read tag without spinlock */
4330 100433424 : return bufHdr->tag.blockNum;
4331 : }
4332 :
4333 : /*
4334 : * BufferGetTag
4335 : * Returns the relfilelocator, fork number and block number associated with
4336 : * a buffer.
4337 : */
4338 : void
4339 31744730 : BufferGetTag(Buffer buffer, RelFileLocator *rlocator, ForkNumber *forknum,
4340 : BlockNumber *blknum)
4341 : {
4342 : BufferDesc *bufHdr;
4343 :
4344 : /* Do the same checks as BufferGetBlockNumber. */
4345 : Assert(BufferIsPinned(buffer));
4346 :
4347 31744730 : if (BufferIsLocal(buffer))
4348 0 : bufHdr = GetLocalBufferDescriptor(-buffer - 1);
4349 : else
4350 31744730 : bufHdr = GetBufferDescriptor(buffer - 1);
4351 :
4352 : /* pinned, so OK to read tag without spinlock */
4353 31744730 : *rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
4354 31744730 : *forknum = BufTagGetForkNum(&bufHdr->tag);
4355 31744730 : *blknum = bufHdr->tag.blockNum;
4356 31744730 : }
4357 :
4358 : /*
4359 : * FlushBuffer
4360 : * Physically write out a shared buffer.
4361 : *
4362 : * NOTE: this actually just passes the buffer contents to the kernel; the
4363 : * real write to disk won't happen until the kernel feels like it. This
4364 : * is okay from our point of view since we can redo the changes from WAL.
4365 : * However, we will need to force the changes to disk via fsync before
4366 : * we can checkpoint WAL.
4367 : *
4368 : * The caller must hold a pin on the buffer and have share-locked the
4369 : * buffer contents. (Note: a share-lock does not prevent updates of
4370 : * hint bits in the buffer, so the page could change while the write
4371 : * is in progress, but we assume that that will not invalidate the data
4372 : * written.)
4373 : *
4374 : * If the caller has an smgr reference for the buffer's relation, pass it
4375 : * as the second parameter. If not, pass NULL.
4376 : */
4377 : static void
4378 1146404 : FlushBuffer(BufferDesc *buf, SMgrRelation reln, IOObject io_object,
4379 : IOContext io_context)
4380 : {
4381 : XLogRecPtr recptr;
4382 : ErrorContextCallback errcallback;
4383 : instr_time io_start;
4384 : Block bufBlock;
4385 : char *bufToWrite;
4386 : uint32 buf_state;
4387 :
4388 : /*
4389 : * Try to start an I/O operation. If StartBufferIO returns false, then
4390 : * someone else flushed the buffer before we could, so we need not do
4391 : * anything.
4392 : */
4393 1146404 : if (!StartBufferIO(buf, false, false))
4394 20 : return;
4395 :
4396 : /* Setup error traceback support for ereport() */
4397 1146384 : errcallback.callback = shared_buffer_write_error_callback;
4398 1146384 : errcallback.arg = buf;
4399 1146384 : errcallback.previous = error_context_stack;
4400 1146384 : error_context_stack = &errcallback;
4401 :
4402 : /* Find smgr relation for buffer */
4403 1146384 : if (reln == NULL)
4404 1141580 : reln = smgropen(BufTagGetRelFileLocator(&buf->tag), INVALID_PROC_NUMBER);
4405 :
4406 : TRACE_POSTGRESQL_BUFFER_FLUSH_START(BufTagGetForkNum(&buf->tag),
4407 : buf->tag.blockNum,
4408 : reln->smgr_rlocator.locator.spcOid,
4409 : reln->smgr_rlocator.locator.dbOid,
4410 : reln->smgr_rlocator.locator.relNumber);
4411 :
4412 1146384 : buf_state = LockBufHdr(buf);
4413 :
4414 : /*
4415 : * Run PageGetLSN while holding header lock, since we don't have the
4416 : * buffer locked exclusively in all cases.
4417 : */
4418 1146384 : recptr = BufferGetLSN(buf);
4419 :
4420 : /* To check if block content changes while flushing. - vadim 01/17/97 */
4421 1146384 : UnlockBufHdrExt(buf, buf_state,
4422 : 0, BM_JUST_DIRTIED,
4423 : 0);
4424 :
4425 : /*
4426 : * Force XLOG flush up to buffer's LSN. This implements the basic WAL
4427 : * rule that log updates must hit disk before any of the data-file changes
4428 : * they describe do.
4429 : *
4430 : * However, this rule does not apply to unlogged relations, which will be
4431 : * lost after a crash anyway. Most unlogged relation pages do not bear
4432 : * LSNs since we never emit WAL records for them, and therefore flushing
4433 : * up through the buffer LSN would be useless, but harmless. However,
4434 : * GiST indexes use LSNs internally to track page-splits, and therefore
4435 : * unlogged GiST pages bear "fake" LSNs generated by
4436 : * GetFakeLSNForUnloggedRel. It is unlikely but possible that the fake
4437 : * LSN counter could advance past the WAL insertion point; and if it did
4438 : * happen, attempting to flush WAL through that location would fail, with
4439 : * disastrous system-wide consequences. To make sure that can't happen,
4440 : * skip the flush if the buffer isn't permanent.
4441 : */
4442 1146384 : if (buf_state & BM_PERMANENT)
4443 1142784 : XLogFlush(recptr);
4444 :
4445 : /*
4446 : * Now it's safe to write the buffer to disk. Note that no one else should
4447 : * have been able to write it, while we were busy with log flushing,
4448 : * because we got the exclusive right to perform I/O by setting the
4449 : * BM_IO_IN_PROGRESS bit.
4450 : */
4451 1146384 : bufBlock = BufHdrGetBlock(buf);
4452 :
4453 : /*
4454 : * Update page checksum if desired. Since we have only shared lock on the
4455 : * buffer, other processes might be updating hint bits in it, so we must
4456 : * copy the page to private storage if we do checksumming.
4457 : */
4458 1146384 : bufToWrite = PageSetChecksumCopy((Page) bufBlock, buf->tag.blockNum);
4459 :
4460 1146384 : io_start = pgstat_prepare_io_time(track_io_timing);
4461 :
4462 : /*
4463 : * bufToWrite is either the shared buffer or a copy, as appropriate.
4464 : */
4465 1146384 : smgrwrite(reln,
4466 1146384 : BufTagGetForkNum(&buf->tag),
4467 : buf->tag.blockNum,
4468 : bufToWrite,
4469 : false);
4470 :
4471 : /*
4472 : * When a strategy is in use, only flushes of dirty buffers already in the
4473 : * strategy ring are counted as strategy writes (IOCONTEXT
4474 : * [BULKREAD|BULKWRITE|VACUUM] IOOP_WRITE) for the purpose of IO
4475 : * statistics tracking.
4476 : *
4477 : * If a shared buffer initially added to the ring must be flushed before
4478 : * being used, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE.
4479 : *
4480 : * If a shared buffer which was added to the ring later because the
4481 : * current strategy buffer is pinned or in use or because all strategy
4482 : * buffers were dirty and rejected (for BAS_BULKREAD operations only)
4483 : * requires flushing, this is counted as an IOCONTEXT_NORMAL IOOP_WRITE
4484 : * (from_ring will be false).
4485 : *
4486 : * When a strategy is not in use, the write can only be a "regular" write
4487 : * of a dirty shared buffer (IOCONTEXT_NORMAL IOOP_WRITE).
4488 : */
4489 1146384 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
4490 : IOOP_WRITE, io_start, 1, BLCKSZ);
4491 :
4492 1146384 : pgBufferUsage.shared_blks_written++;
4493 :
4494 : /*
4495 : * Mark the buffer as clean (unless BM_JUST_DIRTIED has become set) and
4496 : * end the BM_IO_IN_PROGRESS state.
4497 : */
4498 1146384 : TerminateBufferIO(buf, true, 0, true, false);
4499 :
4500 : TRACE_POSTGRESQL_BUFFER_FLUSH_DONE(BufTagGetForkNum(&buf->tag),
4501 : buf->tag.blockNum,
4502 : reln->smgr_rlocator.locator.spcOid,
4503 : reln->smgr_rlocator.locator.dbOid,
4504 : reln->smgr_rlocator.locator.relNumber);
4505 :
4506 : /* Pop the error context stack */
4507 1146384 : error_context_stack = errcallback.previous;
4508 : }
4509 :
4510 : /*
4511 : * Convenience wrapper around FlushBuffer() that locks/unlocks the buffer
4512 : * before/after calling FlushBuffer().
4513 : */
4514 : static void
4515 610768 : FlushUnlockedBuffer(BufferDesc *buf, SMgrRelation reln,
4516 : IOObject io_object, IOContext io_context)
4517 : {
4518 610768 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_SHARED);
4519 610768 : FlushBuffer(buf, reln, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
4520 610768 : LWLockRelease(BufferDescriptorGetContentLock(buf));
4521 610768 : }
4522 :
4523 : /*
4524 : * RelationGetNumberOfBlocksInFork
4525 : * Determines the current number of pages in the specified relation fork.
4526 : *
4527 : * Note that the accuracy of the result will depend on the details of the
4528 : * relation's storage. For builtin AMs it'll be accurate, but for external AMs
4529 : * it might not be.
4530 : */
4531 : BlockNumber
4532 3804432 : RelationGetNumberOfBlocksInFork(Relation relation, ForkNumber forkNum)
4533 : {
4534 3804432 : if (RELKIND_HAS_TABLE_AM(relation->rd_rel->relkind))
4535 : {
4536 : /*
4537 : * Not every table AM uses BLCKSZ wide fixed size blocks. Therefore
4538 : * tableam returns the size in bytes - but for the purpose of this
4539 : * routine, we want the number of blocks. Therefore divide, rounding
4540 : * up.
4541 : */
4542 : uint64 szbytes;
4543 :
4544 2880548 : szbytes = table_relation_size(relation, forkNum);
4545 :
4546 2880510 : return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
4547 : }
4548 923884 : else if (RELKIND_HAS_STORAGE(relation->rd_rel->relkind))
4549 : {
4550 923884 : return smgrnblocks(RelationGetSmgr(relation), forkNum);
4551 : }
4552 : else
4553 : Assert(false);
4554 :
4555 0 : return 0; /* keep compiler quiet */
4556 : }
4557 :
4558 : /*
4559 : * BufferIsPermanent
4560 : * Determines whether a buffer will potentially still be around after
4561 : * a crash. Caller must hold a buffer pin.
4562 : */
4563 : bool
4564 19142154 : BufferIsPermanent(Buffer buffer)
4565 : {
4566 : BufferDesc *bufHdr;
4567 :
4568 : /* Local buffers are used only for temp relations. */
4569 19142154 : if (BufferIsLocal(buffer))
4570 1253962 : return false;
4571 :
4572 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4573 : Assert(BufferIsValid(buffer));
4574 : Assert(BufferIsPinned(buffer));
4575 :
4576 : /*
4577 : * BM_PERMANENT can't be changed while we hold a pin on the buffer, so we
4578 : * need not bother with the buffer header spinlock. Even if someone else
4579 : * changes the buffer header state while we're doing this, the state is
4580 : * changed atomically, so we'll read the old value or the new value, but
4581 : * not random garbage.
4582 : */
4583 17888192 : bufHdr = GetBufferDescriptor(buffer - 1);
4584 17888192 : return (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT) != 0;
4585 : }
4586 :
4587 : /*
4588 : * BufferGetLSNAtomic
4589 : * Retrieves the LSN of the buffer atomically using a buffer header lock.
4590 : * This is necessary for some callers who may not have an exclusive lock
4591 : * on the buffer.
4592 : */
4593 : XLogRecPtr
4594 14338788 : BufferGetLSNAtomic(Buffer buffer)
4595 : {
4596 14338788 : char *page = BufferGetPage(buffer);
4597 : BufferDesc *bufHdr;
4598 : XLogRecPtr lsn;
4599 :
4600 : /*
4601 : * If we don't need locking for correctness, fastpath out.
4602 : */
4603 14338788 : if (!XLogHintBitIsNeeded() || BufferIsLocal(buffer))
4604 478244 : return PageGetLSN(page);
4605 :
4606 : /* Make sure we've got a real buffer, and that we hold a pin on it. */
4607 : Assert(BufferIsValid(buffer));
4608 : Assert(BufferIsPinned(buffer));
4609 :
4610 13860544 : bufHdr = GetBufferDescriptor(buffer - 1);
4611 13860544 : LockBufHdr(bufHdr);
4612 13860544 : lsn = PageGetLSN(page);
4613 13860544 : UnlockBufHdr(bufHdr);
4614 :
4615 13860544 : return lsn;
4616 : }
4617 :
4618 : /* ---------------------------------------------------------------------
4619 : * DropRelationBuffers
4620 : *
4621 : * This function removes from the buffer pool all the pages of the
4622 : * specified relation forks that have block numbers >= firstDelBlock.
4623 : * (In particular, with firstDelBlock = 0, all pages are removed.)
4624 : * Dirty pages are simply dropped, without bothering to write them
4625 : * out first. Therefore, this is NOT rollback-able, and so should be
4626 : * used only with extreme caution!
4627 : *
4628 : * Currently, this is called only from smgr.c when the underlying file
4629 : * is about to be deleted or truncated (firstDelBlock is needed for
4630 : * the truncation case). The data in the affected pages would therefore
4631 : * be deleted momentarily anyway, and there is no point in writing it.
4632 : * It is the responsibility of higher-level code to ensure that the
4633 : * deletion or truncation does not lose any data that could be needed
4634 : * later. It is also the responsibility of higher-level code to ensure
4635 : * that no other process could be trying to load more pages of the
4636 : * relation into buffers.
4637 : * --------------------------------------------------------------------
4638 : */
4639 : void
4640 1288 : DropRelationBuffers(SMgrRelation smgr_reln, ForkNumber *forkNum,
4641 : int nforks, BlockNumber *firstDelBlock)
4642 : {
4643 : int i;
4644 : int j;
4645 : RelFileLocatorBackend rlocator;
4646 : BlockNumber nForkBlock[MAX_FORKNUM];
4647 1288 : uint64 nBlocksToInvalidate = 0;
4648 :
4649 1288 : rlocator = smgr_reln->smgr_rlocator;
4650 :
4651 : /* If it's a local relation, it's localbuf.c's problem. */
4652 1288 : if (RelFileLocatorBackendIsTemp(rlocator))
4653 : {
4654 748 : if (rlocator.backend == MyProcNumber)
4655 748 : DropRelationLocalBuffers(rlocator.locator, forkNum, nforks,
4656 : firstDelBlock);
4657 :
4658 828 : return;
4659 : }
4660 :
4661 : /*
4662 : * To remove all the pages of the specified relation forks from the buffer
4663 : * pool, we need to scan the entire buffer pool but we can optimize it by
4664 : * finding the buffers from BufMapping table provided we know the exact
4665 : * size of each fork of the relation. The exact size is required to ensure
4666 : * that we don't leave any buffer for the relation being dropped as
4667 : * otherwise the background writer or checkpointer can lead to a PANIC
4668 : * error while flushing buffers corresponding to files that don't exist.
4669 : *
4670 : * To know the exact size, we rely on the size cached for each fork by us
4671 : * during recovery which limits the optimization to recovery and on
4672 : * standbys but we can easily extend it once we have shared cache for
4673 : * relation size.
4674 : *
4675 : * In recovery, we cache the value returned by the first lseek(SEEK_END)
4676 : * and the future writes keeps the cached value up-to-date. See
4677 : * smgrextend. It is possible that the value of the first lseek is smaller
4678 : * than the actual number of existing blocks in the file due to buggy
4679 : * Linux kernels that might not have accounted for the recent write. But
4680 : * that should be fine because there must not be any buffers after that
4681 : * file size.
4682 : */
4683 730 : for (i = 0; i < nforks; i++)
4684 : {
4685 : /* Get the number of blocks for a relation's fork */
4686 628 : nForkBlock[i] = smgrnblocks_cached(smgr_reln, forkNum[i]);
4687 :
4688 628 : if (nForkBlock[i] == InvalidBlockNumber)
4689 : {
4690 438 : nBlocksToInvalidate = InvalidBlockNumber;
4691 438 : break;
4692 : }
4693 :
4694 : /* calculate the number of blocks to be invalidated */
4695 190 : nBlocksToInvalidate += (nForkBlock[i] - firstDelBlock[i]);
4696 : }
4697 :
4698 : /*
4699 : * We apply the optimization iff the total number of blocks to invalidate
4700 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4701 : */
4702 540 : if (BlockNumberIsValid(nBlocksToInvalidate) &&
4703 102 : nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4704 : {
4705 222 : for (j = 0; j < nforks; j++)
4706 142 : FindAndDropRelationBuffers(rlocator.locator, forkNum[j],
4707 142 : nForkBlock[j], firstDelBlock[j]);
4708 80 : return;
4709 : }
4710 :
4711 6139084 : for (i = 0; i < NBuffers; i++)
4712 : {
4713 6138624 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4714 :
4715 : /*
4716 : * We can make this a tad faster by prechecking the buffer tag before
4717 : * we attempt to lock the buffer; this saves a lot of lock
4718 : * acquisitions in typical cases. It should be safe because the
4719 : * caller must have AccessExclusiveLock on the relation, or some other
4720 : * reason to be certain that no one is loading new pages of the rel
4721 : * into the buffer pool. (Otherwise we might well miss such pages
4722 : * entirely.) Therefore, while the tag might be changing while we
4723 : * look at it, it can't be changing *to* a value we care about, only
4724 : * *away* from such a value. So false negatives are impossible, and
4725 : * false positives are safe because we'll recheck after getting the
4726 : * buffer lock.
4727 : *
4728 : * We could check forkNum and blockNum as well as the rlocator, but
4729 : * the incremental win from doing so seems small.
4730 : */
4731 6138624 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator))
4732 6120496 : continue;
4733 :
4734 18128 : LockBufHdr(bufHdr);
4735 :
4736 45186 : for (j = 0; j < nforks; j++)
4737 : {
4738 31910 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator.locator) &&
4739 31910 : BufTagGetForkNum(&bufHdr->tag) == forkNum[j] &&
4740 17910 : bufHdr->tag.blockNum >= firstDelBlock[j])
4741 : {
4742 4852 : InvalidateBuffer(bufHdr); /* releases spinlock */
4743 4852 : break;
4744 : }
4745 : }
4746 18128 : if (j >= nforks)
4747 13276 : UnlockBufHdr(bufHdr);
4748 : }
4749 : }
4750 :
4751 : /* ---------------------------------------------------------------------
4752 : * DropRelationsAllBuffers
4753 : *
4754 : * This function removes from the buffer pool all the pages of all
4755 : * forks of the specified relations. It's equivalent to calling
4756 : * DropRelationBuffers once per fork per relation with firstDelBlock = 0.
4757 : * --------------------------------------------------------------------
4758 : */
4759 : void
4760 28558 : DropRelationsAllBuffers(SMgrRelation *smgr_reln, int nlocators)
4761 : {
4762 : int i;
4763 28558 : int n = 0;
4764 : SMgrRelation *rels;
4765 : BlockNumber (*block)[MAX_FORKNUM + 1];
4766 28558 : uint64 nBlocksToInvalidate = 0;
4767 : RelFileLocator *locators;
4768 28558 : bool cached = true;
4769 : bool use_bsearch;
4770 :
4771 28558 : if (nlocators == 0)
4772 0 : return;
4773 :
4774 28558 : rels = palloc_array(SMgrRelation, nlocators); /* non-local relations */
4775 :
4776 : /* If it's a local relation, it's localbuf.c's problem. */
4777 124820 : for (i = 0; i < nlocators; i++)
4778 : {
4779 96262 : if (RelFileLocatorBackendIsTemp(smgr_reln[i]->smgr_rlocator))
4780 : {
4781 6510 : if (smgr_reln[i]->smgr_rlocator.backend == MyProcNumber)
4782 6510 : DropRelationAllLocalBuffers(smgr_reln[i]->smgr_rlocator.locator);
4783 : }
4784 : else
4785 89752 : rels[n++] = smgr_reln[i];
4786 : }
4787 :
4788 : /*
4789 : * If there are no non-local relations, then we're done. Release the
4790 : * memory and return.
4791 : */
4792 28558 : if (n == 0)
4793 : {
4794 1704 : pfree(rels);
4795 1704 : return;
4796 : }
4797 :
4798 : /*
4799 : * This is used to remember the number of blocks for all the relations
4800 : * forks.
4801 : */
4802 : block = (BlockNumber (*)[MAX_FORKNUM + 1])
4803 26854 : palloc(sizeof(BlockNumber) * n * (MAX_FORKNUM + 1));
4804 :
4805 : /*
4806 : * We can avoid scanning the entire buffer pool if we know the exact size
4807 : * of each of the given relation forks. See DropRelationBuffers.
4808 : */
4809 56264 : for (i = 0; i < n && cached; i++)
4810 : {
4811 46170 : for (int j = 0; j <= MAX_FORKNUM; j++)
4812 : {
4813 : /* Get the number of blocks for a relation's fork. */
4814 42010 : block[i][j] = smgrnblocks_cached(rels[i], j);
4815 :
4816 : /* We need to only consider the relation forks that exists. */
4817 42010 : if (block[i][j] == InvalidBlockNumber)
4818 : {
4819 37530 : if (!smgrexists(rels[i], j))
4820 12280 : continue;
4821 25250 : cached = false;
4822 25250 : break;
4823 : }
4824 :
4825 : /* calculate the total number of blocks to be invalidated */
4826 4480 : nBlocksToInvalidate += block[i][j];
4827 : }
4828 : }
4829 :
4830 : /*
4831 : * We apply the optimization iff the total number of blocks to invalidate
4832 : * is below the BUF_DROP_FULL_SCAN_THRESHOLD.
4833 : */
4834 26854 : if (cached && nBlocksToInvalidate < BUF_DROP_FULL_SCAN_THRESHOLD)
4835 : {
4836 2658 : for (i = 0; i < n; i++)
4837 : {
4838 7320 : for (int j = 0; j <= MAX_FORKNUM; j++)
4839 : {
4840 : /* ignore relation forks that doesn't exist */
4841 5856 : if (!BlockNumberIsValid(block[i][j]))
4842 4374 : continue;
4843 :
4844 : /* drop all the buffers for a particular relation fork */
4845 1482 : FindAndDropRelationBuffers(rels[i]->smgr_rlocator.locator,
4846 1482 : j, block[i][j], 0);
4847 : }
4848 : }
4849 :
4850 1194 : pfree(block);
4851 1194 : pfree(rels);
4852 1194 : return;
4853 : }
4854 :
4855 25660 : pfree(block);
4856 25660 : locators = palloc_array(RelFileLocator, n); /* non-local relations */
4857 113948 : for (i = 0; i < n; i++)
4858 88288 : locators[i] = rels[i]->smgr_rlocator.locator;
4859 :
4860 : /*
4861 : * For low number of relations to drop just use a simple walk through, to
4862 : * save the bsearch overhead. The threshold to use is rather a guess than
4863 : * an exactly determined value, as it depends on many factors (CPU and RAM
4864 : * speeds, amount of shared buffers etc.).
4865 : */
4866 25660 : use_bsearch = n > RELS_BSEARCH_THRESHOLD;
4867 :
4868 : /* sort the list of rlocators if necessary */
4869 25660 : if (use_bsearch)
4870 346 : qsort(locators, n, sizeof(RelFileLocator), rlocator_comparator);
4871 :
4872 276410940 : for (i = 0; i < NBuffers; i++)
4873 : {
4874 276385280 : RelFileLocator *rlocator = NULL;
4875 276385280 : BufferDesc *bufHdr = GetBufferDescriptor(i);
4876 :
4877 : /*
4878 : * As in DropRelationBuffers, an unlocked precheck should be safe and
4879 : * saves some cycles.
4880 : */
4881 :
4882 276385280 : if (!use_bsearch)
4883 : {
4884 : int j;
4885 :
4886 1107375382 : for (j = 0; j < n; j++)
4887 : {
4888 834882866 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &locators[j]))
4889 : {
4890 174620 : rlocator = &locators[j];
4891 174620 : break;
4892 : }
4893 : }
4894 : }
4895 : else
4896 : {
4897 : RelFileLocator locator;
4898 :
4899 3718144 : locator = BufTagGetRelFileLocator(&bufHdr->tag);
4900 3718144 : rlocator = bsearch(&locator,
4901 : locators, n, sizeof(RelFileLocator),
4902 : rlocator_comparator);
4903 : }
4904 :
4905 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
4906 276385280 : if (rlocator == NULL)
4907 276207622 : continue;
4908 :
4909 177658 : LockBufHdr(bufHdr);
4910 177658 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, rlocator))
4911 177658 : InvalidateBuffer(bufHdr); /* releases spinlock */
4912 : else
4913 0 : UnlockBufHdr(bufHdr);
4914 : }
4915 :
4916 25660 : pfree(locators);
4917 25660 : pfree(rels);
4918 : }
4919 :
4920 : /* ---------------------------------------------------------------------
4921 : * FindAndDropRelationBuffers
4922 : *
4923 : * This function performs look up in BufMapping table and removes from the
4924 : * buffer pool all the pages of the specified relation fork that has block
4925 : * number >= firstDelBlock. (In particular, with firstDelBlock = 0, all
4926 : * pages are removed.)
4927 : * --------------------------------------------------------------------
4928 : */
4929 : static void
4930 1624 : FindAndDropRelationBuffers(RelFileLocator rlocator, ForkNumber forkNum,
4931 : BlockNumber nForkBlock,
4932 : BlockNumber firstDelBlock)
4933 : {
4934 : BlockNumber curBlock;
4935 :
4936 3920 : for (curBlock = firstDelBlock; curBlock < nForkBlock; curBlock++)
4937 : {
4938 : uint32 bufHash; /* hash value for tag */
4939 : BufferTag bufTag; /* identity of requested block */
4940 : LWLock *bufPartitionLock; /* buffer partition lock for it */
4941 : int buf_id;
4942 : BufferDesc *bufHdr;
4943 :
4944 : /* create a tag so we can lookup the buffer */
4945 2296 : InitBufferTag(&bufTag, &rlocator, forkNum, curBlock);
4946 :
4947 : /* determine its hash code and partition lock ID */
4948 2296 : bufHash = BufTableHashCode(&bufTag);
4949 2296 : bufPartitionLock = BufMappingPartitionLock(bufHash);
4950 :
4951 : /* Check that it is in the buffer pool. If not, do nothing. */
4952 2296 : LWLockAcquire(bufPartitionLock, LW_SHARED);
4953 2296 : buf_id = BufTableLookup(&bufTag, bufHash);
4954 2296 : LWLockRelease(bufPartitionLock);
4955 :
4956 2296 : if (buf_id < 0)
4957 230 : continue;
4958 :
4959 2066 : bufHdr = GetBufferDescriptor(buf_id);
4960 :
4961 : /*
4962 : * We need to lock the buffer header and recheck if the buffer is
4963 : * still associated with the same block because the buffer could be
4964 : * evicted by some other backend loading blocks for a different
4965 : * relation after we release lock on the BufMapping table.
4966 : */
4967 2066 : LockBufHdr(bufHdr);
4968 :
4969 4132 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rlocator) &&
4970 2066 : BufTagGetForkNum(&bufHdr->tag) == forkNum &&
4971 2066 : bufHdr->tag.blockNum >= firstDelBlock)
4972 2066 : InvalidateBuffer(bufHdr); /* releases spinlock */
4973 : else
4974 0 : UnlockBufHdr(bufHdr);
4975 : }
4976 1624 : }
4977 :
4978 : /* ---------------------------------------------------------------------
4979 : * DropDatabaseBuffers
4980 : *
4981 : * This function removes all the buffers in the buffer cache for a
4982 : * particular database. Dirty pages are simply dropped, without
4983 : * bothering to write them out first. This is used when we destroy a
4984 : * database, to avoid trying to flush data to disk when the directory
4985 : * tree no longer exists. Implementation is pretty similar to
4986 : * DropRelationBuffers() which is for destroying just one relation.
4987 : * --------------------------------------------------------------------
4988 : */
4989 : void
4990 152 : DropDatabaseBuffers(Oid dbid)
4991 : {
4992 : int i;
4993 :
4994 : /*
4995 : * We needn't consider local buffers, since by assumption the target
4996 : * database isn't our own.
4997 : */
4998 :
4999 1092504 : for (i = 0; i < NBuffers; i++)
5000 : {
5001 1092352 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5002 :
5003 : /*
5004 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5005 : * saves some cycles.
5006 : */
5007 1092352 : if (bufHdr->tag.dbOid != dbid)
5008 1064548 : continue;
5009 :
5010 27804 : LockBufHdr(bufHdr);
5011 27804 : if (bufHdr->tag.dbOid == dbid)
5012 27804 : InvalidateBuffer(bufHdr); /* releases spinlock */
5013 : else
5014 0 : UnlockBufHdr(bufHdr);
5015 : }
5016 152 : }
5017 :
5018 : /* ---------------------------------------------------------------------
5019 : * FlushRelationBuffers
5020 : *
5021 : * This function writes all dirty pages of a relation out to disk
5022 : * (or more accurately, out to kernel disk buffers), ensuring that the
5023 : * kernel has an up-to-date view of the relation.
5024 : *
5025 : * Generally, the caller should be holding AccessExclusiveLock on the
5026 : * target relation to ensure that no other backend is busy dirtying
5027 : * more blocks of the relation; the effects can't be expected to last
5028 : * after the lock is released.
5029 : *
5030 : * XXX currently it sequentially searches the buffer pool, should be
5031 : * changed to more clever ways of searching. This routine is not
5032 : * used in any performance-critical code paths, so it's not worth
5033 : * adding additional overhead to normal paths to make it go faster.
5034 : * --------------------------------------------------------------------
5035 : */
5036 : void
5037 276 : FlushRelationBuffers(Relation rel)
5038 : {
5039 : int i;
5040 : BufferDesc *bufHdr;
5041 276 : SMgrRelation srel = RelationGetSmgr(rel);
5042 :
5043 276 : if (RelationUsesLocalBuffers(rel))
5044 : {
5045 1818 : for (i = 0; i < NLocBuffer; i++)
5046 : {
5047 : uint32 buf_state;
5048 :
5049 1800 : bufHdr = GetLocalBufferDescriptor(i);
5050 1800 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
5051 600 : ((buf_state = pg_atomic_read_u32(&bufHdr->state)) &
5052 : (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5053 : {
5054 : ErrorContextCallback errcallback;
5055 :
5056 : /* Setup error traceback support for ereport() */
5057 600 : errcallback.callback = local_buffer_write_error_callback;
5058 600 : errcallback.arg = bufHdr;
5059 600 : errcallback.previous = error_context_stack;
5060 600 : error_context_stack = &errcallback;
5061 :
5062 : /* Make sure we can handle the pin */
5063 600 : ReservePrivateRefCountEntry();
5064 600 : ResourceOwnerEnlarge(CurrentResourceOwner);
5065 :
5066 : /*
5067 : * Pin/unpin mostly to make valgrind work, but it also seems
5068 : * like the right thing to do.
5069 : */
5070 600 : PinLocalBuffer(bufHdr, false);
5071 :
5072 :
5073 600 : FlushLocalBuffer(bufHdr, srel);
5074 :
5075 600 : UnpinLocalBuffer(BufferDescriptorGetBuffer(bufHdr));
5076 :
5077 : /* Pop the error context stack */
5078 600 : error_context_stack = errcallback.previous;
5079 : }
5080 : }
5081 :
5082 18 : return;
5083 : }
5084 :
5085 3024386 : for (i = 0; i < NBuffers; i++)
5086 : {
5087 : uint32 buf_state;
5088 :
5089 3024128 : bufHdr = GetBufferDescriptor(i);
5090 :
5091 : /*
5092 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5093 : * saves some cycles.
5094 : */
5095 3024128 : if (!BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator))
5096 3023706 : continue;
5097 :
5098 : /* Make sure we can handle the pin */
5099 422 : ReservePrivateRefCountEntry();
5100 422 : ResourceOwnerEnlarge(CurrentResourceOwner);
5101 :
5102 422 : buf_state = LockBufHdr(bufHdr);
5103 422 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &rel->rd_locator) &&
5104 422 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5105 : {
5106 342 : PinBuffer_Locked(bufHdr);
5107 342 : FlushUnlockedBuffer(bufHdr, srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5108 342 : UnpinBuffer(bufHdr);
5109 : }
5110 : else
5111 80 : UnlockBufHdr(bufHdr);
5112 : }
5113 : }
5114 :
5115 : /* ---------------------------------------------------------------------
5116 : * FlushRelationsAllBuffers
5117 : *
5118 : * This function flushes out of the buffer pool all the pages of all
5119 : * forks of the specified smgr relations. It's equivalent to calling
5120 : * FlushRelationBuffers once per relation. The relations are assumed not
5121 : * to use local buffers.
5122 : * --------------------------------------------------------------------
5123 : */
5124 : void
5125 8 : FlushRelationsAllBuffers(SMgrRelation *smgrs, int nrels)
5126 : {
5127 : int i;
5128 : SMgrSortArray *srels;
5129 : bool use_bsearch;
5130 :
5131 8 : if (nrels == 0)
5132 0 : return;
5133 :
5134 : /* fill-in array for qsort */
5135 8 : srels = palloc_array(SMgrSortArray, nrels);
5136 :
5137 16 : for (i = 0; i < nrels; i++)
5138 : {
5139 : Assert(!RelFileLocatorBackendIsTemp(smgrs[i]->smgr_rlocator));
5140 :
5141 8 : srels[i].rlocator = smgrs[i]->smgr_rlocator.locator;
5142 8 : srels[i].srel = smgrs[i];
5143 : }
5144 :
5145 : /*
5146 : * Save the bsearch overhead for low number of relations to sync. See
5147 : * DropRelationsAllBuffers for details.
5148 : */
5149 8 : use_bsearch = nrels > RELS_BSEARCH_THRESHOLD;
5150 :
5151 : /* sort the list of SMgrRelations if necessary */
5152 8 : if (use_bsearch)
5153 0 : qsort(srels, nrels, sizeof(SMgrSortArray), rlocator_comparator);
5154 :
5155 131080 : for (i = 0; i < NBuffers; i++)
5156 : {
5157 131072 : SMgrSortArray *srelent = NULL;
5158 131072 : BufferDesc *bufHdr = GetBufferDescriptor(i);
5159 : uint32 buf_state;
5160 :
5161 : /*
5162 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5163 : * saves some cycles.
5164 : */
5165 :
5166 131072 : if (!use_bsearch)
5167 : {
5168 : int j;
5169 :
5170 257614 : for (j = 0; j < nrels; j++)
5171 : {
5172 131072 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srels[j].rlocator))
5173 : {
5174 4530 : srelent = &srels[j];
5175 4530 : break;
5176 : }
5177 : }
5178 : }
5179 : else
5180 : {
5181 : RelFileLocator rlocator;
5182 :
5183 0 : rlocator = BufTagGetRelFileLocator(&bufHdr->tag);
5184 0 : srelent = bsearch(&rlocator,
5185 : srels, nrels, sizeof(SMgrSortArray),
5186 : rlocator_comparator);
5187 : }
5188 :
5189 : /* buffer doesn't belong to any of the given relfilelocators; skip it */
5190 131072 : if (srelent == NULL)
5191 126542 : continue;
5192 :
5193 : /* Make sure we can handle the pin */
5194 4530 : ReservePrivateRefCountEntry();
5195 4530 : ResourceOwnerEnlarge(CurrentResourceOwner);
5196 :
5197 4530 : buf_state = LockBufHdr(bufHdr);
5198 4530 : if (BufTagMatchesRelFileLocator(&bufHdr->tag, &srelent->rlocator) &&
5199 4530 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5200 : {
5201 4462 : PinBuffer_Locked(bufHdr);
5202 4462 : FlushUnlockedBuffer(bufHdr, srelent->srel, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5203 4462 : UnpinBuffer(bufHdr);
5204 : }
5205 : else
5206 68 : UnlockBufHdr(bufHdr);
5207 : }
5208 :
5209 8 : pfree(srels);
5210 : }
5211 :
5212 : /* ---------------------------------------------------------------------
5213 : * RelationCopyStorageUsingBuffer
5214 : *
5215 : * Copy fork's data using bufmgr. Same as RelationCopyStorage but instead
5216 : * of using smgrread and smgrextend this will copy using bufmgr APIs.
5217 : *
5218 : * Refer comments atop CreateAndCopyRelationData() for details about
5219 : * 'permanent' parameter.
5220 : * --------------------------------------------------------------------
5221 : */
5222 : static void
5223 149660 : RelationCopyStorageUsingBuffer(RelFileLocator srclocator,
5224 : RelFileLocator dstlocator,
5225 : ForkNumber forkNum, bool permanent)
5226 : {
5227 : Buffer srcBuf;
5228 : Buffer dstBuf;
5229 : Page srcPage;
5230 : Page dstPage;
5231 : bool use_wal;
5232 : BlockNumber nblocks;
5233 : BlockNumber blkno;
5234 : PGIOAlignedBlock buf;
5235 : BufferAccessStrategy bstrategy_src;
5236 : BufferAccessStrategy bstrategy_dst;
5237 : BlockRangeReadStreamPrivate p;
5238 : ReadStream *src_stream;
5239 : SMgrRelation src_smgr;
5240 :
5241 : /*
5242 : * In general, we want to write WAL whenever wal_level > 'minimal', but we
5243 : * can skip it when copying any fork of an unlogged relation other than
5244 : * the init fork.
5245 : */
5246 149660 : use_wal = XLogIsNeeded() && (permanent || forkNum == INIT_FORKNUM);
5247 :
5248 : /* Get number of blocks in the source relation. */
5249 149660 : nblocks = smgrnblocks(smgropen(srclocator, INVALID_PROC_NUMBER),
5250 : forkNum);
5251 :
5252 : /* Nothing to copy; just return. */
5253 149660 : if (nblocks == 0)
5254 26110 : return;
5255 :
5256 : /*
5257 : * Bulk extend the destination relation of the same size as the source
5258 : * relation before starting to copy block by block.
5259 : */
5260 123550 : memset(buf.data, 0, BLCKSZ);
5261 123550 : smgrextend(smgropen(dstlocator, INVALID_PROC_NUMBER), forkNum, nblocks - 1,
5262 : buf.data, true);
5263 :
5264 : /* This is a bulk operation, so use buffer access strategies. */
5265 123550 : bstrategy_src = GetAccessStrategy(BAS_BULKREAD);
5266 123550 : bstrategy_dst = GetAccessStrategy(BAS_BULKWRITE);
5267 :
5268 : /* Initialize streaming read */
5269 123550 : p.current_blocknum = 0;
5270 123550 : p.last_exclusive = nblocks;
5271 123550 : src_smgr = smgropen(srclocator, INVALID_PROC_NUMBER);
5272 :
5273 : /*
5274 : * It is safe to use batchmode as block_range_read_stream_cb takes no
5275 : * locks.
5276 : */
5277 123550 : src_stream = read_stream_begin_smgr_relation(READ_STREAM_FULL |
5278 : READ_STREAM_USE_BATCHING,
5279 : bstrategy_src,
5280 : src_smgr,
5281 : permanent ? RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED,
5282 : forkNum,
5283 : block_range_read_stream_cb,
5284 : &p,
5285 : 0);
5286 :
5287 : /* Iterate over each block of the source relation file. */
5288 596676 : for (blkno = 0; blkno < nblocks; blkno++)
5289 : {
5290 473130 : CHECK_FOR_INTERRUPTS();
5291 :
5292 : /* Read block from source relation. */
5293 473130 : srcBuf = read_stream_next_buffer(src_stream, NULL);
5294 473126 : LockBuffer(srcBuf, BUFFER_LOCK_SHARE);
5295 473126 : srcPage = BufferGetPage(srcBuf);
5296 :
5297 473126 : dstBuf = ReadBufferWithoutRelcache(dstlocator, forkNum,
5298 : BufferGetBlockNumber(srcBuf),
5299 : RBM_ZERO_AND_LOCK, bstrategy_dst,
5300 : permanent);
5301 473126 : dstPage = BufferGetPage(dstBuf);
5302 :
5303 473126 : START_CRIT_SECTION();
5304 :
5305 : /* Copy page data from the source to the destination. */
5306 473126 : memcpy(dstPage, srcPage, BLCKSZ);
5307 473126 : MarkBufferDirty(dstBuf);
5308 :
5309 : /* WAL-log the copied page. */
5310 473126 : if (use_wal)
5311 273374 : log_newpage_buffer(dstBuf, true);
5312 :
5313 473126 : END_CRIT_SECTION();
5314 :
5315 473126 : UnlockReleaseBuffer(dstBuf);
5316 473126 : UnlockReleaseBuffer(srcBuf);
5317 : }
5318 : Assert(read_stream_next_buffer(src_stream, NULL) == InvalidBuffer);
5319 123546 : read_stream_end(src_stream);
5320 :
5321 123546 : FreeAccessStrategy(bstrategy_src);
5322 123546 : FreeAccessStrategy(bstrategy_dst);
5323 : }
5324 :
5325 : /* ---------------------------------------------------------------------
5326 : * CreateAndCopyRelationData
5327 : *
5328 : * Create destination relation storage and copy all forks from the
5329 : * source relation to the destination.
5330 : *
5331 : * Pass permanent as true for permanent relations and false for
5332 : * unlogged relations. Currently this API is not supported for
5333 : * temporary relations.
5334 : * --------------------------------------------------------------------
5335 : */
5336 : void
5337 112504 : CreateAndCopyRelationData(RelFileLocator src_rlocator,
5338 : RelFileLocator dst_rlocator, bool permanent)
5339 : {
5340 : char relpersistence;
5341 : SMgrRelation src_rel;
5342 : SMgrRelation dst_rel;
5343 :
5344 : /* Set the relpersistence. */
5345 112504 : relpersistence = permanent ?
5346 : RELPERSISTENCE_PERMANENT : RELPERSISTENCE_UNLOGGED;
5347 :
5348 112504 : src_rel = smgropen(src_rlocator, INVALID_PROC_NUMBER);
5349 112504 : dst_rel = smgropen(dst_rlocator, INVALID_PROC_NUMBER);
5350 :
5351 : /*
5352 : * Create and copy all forks of the relation. During create database we
5353 : * have a separate cleanup mechanism which deletes complete database
5354 : * directory. Therefore, each individual relation doesn't need to be
5355 : * registered for cleanup.
5356 : */
5357 112504 : RelationCreateStorage(dst_rlocator, relpersistence, false);
5358 :
5359 : /* copy main fork. */
5360 112504 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, MAIN_FORKNUM,
5361 : permanent);
5362 :
5363 : /* copy those extra forks that exist */
5364 112500 : for (ForkNumber forkNum = MAIN_FORKNUM + 1;
5365 450000 : forkNum <= MAX_FORKNUM; forkNum++)
5366 : {
5367 337500 : if (smgrexists(src_rel, forkNum))
5368 : {
5369 37156 : smgrcreate(dst_rel, forkNum, false);
5370 :
5371 : /*
5372 : * WAL log creation if the relation is persistent, or this is the
5373 : * init fork of an unlogged relation.
5374 : */
5375 37156 : if (permanent || forkNum == INIT_FORKNUM)
5376 37156 : log_smgrcreate(&dst_rlocator, forkNum);
5377 :
5378 : /* Copy a fork's data, block by block. */
5379 37156 : RelationCopyStorageUsingBuffer(src_rlocator, dst_rlocator, forkNum,
5380 : permanent);
5381 : }
5382 : }
5383 112500 : }
5384 :
5385 : /* ---------------------------------------------------------------------
5386 : * FlushDatabaseBuffers
5387 : *
5388 : * This function writes all dirty pages of a database out to disk
5389 : * (or more accurately, out to kernel disk buffers), ensuring that the
5390 : * kernel has an up-to-date view of the database.
5391 : *
5392 : * Generally, the caller should be holding an appropriate lock to ensure
5393 : * no other backend is active in the target database; otherwise more
5394 : * pages could get dirtied.
5395 : *
5396 : * Note we don't worry about flushing any pages of temporary relations.
5397 : * It's assumed these wouldn't be interesting.
5398 : * --------------------------------------------------------------------
5399 : */
5400 : void
5401 10 : FlushDatabaseBuffers(Oid dbid)
5402 : {
5403 : int i;
5404 : BufferDesc *bufHdr;
5405 :
5406 1290 : for (i = 0; i < NBuffers; i++)
5407 : {
5408 : uint32 buf_state;
5409 :
5410 1280 : bufHdr = GetBufferDescriptor(i);
5411 :
5412 : /*
5413 : * As in DropRelationBuffers, an unlocked precheck should be safe and
5414 : * saves some cycles.
5415 : */
5416 1280 : if (bufHdr->tag.dbOid != dbid)
5417 1046 : continue;
5418 :
5419 : /* Make sure we can handle the pin */
5420 234 : ReservePrivateRefCountEntry();
5421 234 : ResourceOwnerEnlarge(CurrentResourceOwner);
5422 :
5423 234 : buf_state = LockBufHdr(bufHdr);
5424 234 : if (bufHdr->tag.dbOid == dbid &&
5425 234 : (buf_state & (BM_VALID | BM_DIRTY)) == (BM_VALID | BM_DIRTY))
5426 : {
5427 0 : PinBuffer_Locked(bufHdr);
5428 0 : FlushUnlockedBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5429 0 : UnpinBuffer(bufHdr);
5430 : }
5431 : else
5432 234 : UnlockBufHdr(bufHdr);
5433 : }
5434 10 : }
5435 :
5436 : /*
5437 : * Flush a previously, shared or exclusively, locked and pinned buffer to the
5438 : * OS.
5439 : */
5440 : void
5441 160 : FlushOneBuffer(Buffer buffer)
5442 : {
5443 : BufferDesc *bufHdr;
5444 :
5445 : /* currently not needed, but no fundamental reason not to support */
5446 : Assert(!BufferIsLocal(buffer));
5447 :
5448 : Assert(BufferIsPinned(buffer));
5449 :
5450 160 : bufHdr = GetBufferDescriptor(buffer - 1);
5451 :
5452 : Assert(BufferIsLockedByMe(buffer));
5453 :
5454 160 : FlushBuffer(bufHdr, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
5455 160 : }
5456 :
5457 : /*
5458 : * ReleaseBuffer -- release the pin on a buffer
5459 : */
5460 : void
5461 126025316 : ReleaseBuffer(Buffer buffer)
5462 : {
5463 126025316 : if (!BufferIsValid(buffer))
5464 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5465 :
5466 126025316 : if (BufferIsLocal(buffer))
5467 3220644 : UnpinLocalBuffer(buffer);
5468 : else
5469 122804672 : UnpinBuffer(GetBufferDescriptor(buffer - 1));
5470 126025316 : }
5471 :
5472 : /*
5473 : * UnlockReleaseBuffer -- release the content lock and pin on a buffer
5474 : *
5475 : * This is just a shorthand for a common combination.
5476 : */
5477 : void
5478 37455472 : UnlockReleaseBuffer(Buffer buffer)
5479 : {
5480 37455472 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5481 37455472 : ReleaseBuffer(buffer);
5482 37455472 : }
5483 :
5484 : /*
5485 : * IncrBufferRefCount
5486 : * Increment the pin count on a buffer that we have *already* pinned
5487 : * at least once.
5488 : *
5489 : * This function cannot be used on a buffer we do not have pinned,
5490 : * because it doesn't change the shared buffer state.
5491 : */
5492 : void
5493 23843442 : IncrBufferRefCount(Buffer buffer)
5494 : {
5495 : Assert(BufferIsPinned(buffer));
5496 23843442 : ResourceOwnerEnlarge(CurrentResourceOwner);
5497 23843442 : if (BufferIsLocal(buffer))
5498 709402 : LocalRefCount[-buffer - 1]++;
5499 : else
5500 : {
5501 : PrivateRefCountEntry *ref;
5502 :
5503 23134040 : ref = GetPrivateRefCountEntry(buffer, true);
5504 : Assert(ref != NULL);
5505 23134040 : ref->data.refcount++;
5506 : }
5507 23843442 : ResourceOwnerRememberBuffer(CurrentResourceOwner, buffer);
5508 23843442 : }
5509 :
5510 : /*
5511 : * MarkBufferDirtyHint
5512 : *
5513 : * Mark a buffer dirty for non-critical changes.
5514 : *
5515 : * This is essentially the same as MarkBufferDirty, except:
5516 : *
5517 : * 1. The caller does not write WAL; so if checksums are enabled, we may need
5518 : * to write an XLOG_FPI_FOR_HINT WAL record to protect against torn pages.
5519 : * 2. The caller might have only share-lock instead of exclusive-lock on the
5520 : * buffer's content lock.
5521 : * 3. This function does not guarantee that the buffer is always marked dirty
5522 : * (due to a race condition), so it cannot be used for important changes.
5523 : */
5524 : void
5525 20137566 : MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
5526 : {
5527 : BufferDesc *bufHdr;
5528 20137566 : Page page = BufferGetPage(buffer);
5529 :
5530 20137566 : if (!BufferIsValid(buffer))
5531 0 : elog(ERROR, "bad buffer ID: %d", buffer);
5532 :
5533 20137566 : if (BufferIsLocal(buffer))
5534 : {
5535 1270312 : MarkLocalBufferDirty(buffer);
5536 1270312 : return;
5537 : }
5538 :
5539 18867254 : bufHdr = GetBufferDescriptor(buffer - 1);
5540 :
5541 : Assert(GetPrivateRefCount(buffer) > 0);
5542 : /* here, either share or exclusive lock is OK */
5543 : Assert(BufferIsLockedByMe(buffer));
5544 :
5545 : /*
5546 : * This routine might get called many times on the same page, if we are
5547 : * making the first scan after commit of an xact that added/deleted many
5548 : * tuples. So, be as quick as we can if the buffer is already dirty. We
5549 : * do this by not acquiring spinlock if it looks like the status bits are
5550 : * already set. Since we make this test unlocked, there's a chance we
5551 : * might fail to notice that the flags have just been cleared, and failed
5552 : * to reset them, due to memory-ordering issues. But since this function
5553 : * is only intended to be used in cases where failing to write out the
5554 : * data would be harmless anyway, it doesn't really matter.
5555 : */
5556 18867254 : if ((pg_atomic_read_u32(&bufHdr->state) & (BM_DIRTY | BM_JUST_DIRTIED)) !=
5557 : (BM_DIRTY | BM_JUST_DIRTIED))
5558 : {
5559 1817506 : XLogRecPtr lsn = InvalidXLogRecPtr;
5560 1817506 : bool dirtied = false;
5561 1817506 : bool delayChkptFlags = false;
5562 : uint32 buf_state;
5563 :
5564 : /*
5565 : * If we need to protect hint bit updates from torn writes, WAL-log a
5566 : * full page image of the page. This full page image is only necessary
5567 : * if the hint bit update is the first change to the page since the
5568 : * last checkpoint.
5569 : *
5570 : * We don't check full_page_writes here because that logic is included
5571 : * when we call XLogInsert() since the value changes dynamically.
5572 : */
5573 3632840 : if (XLogHintBitIsNeeded() &&
5574 1815334 : (pg_atomic_read_u32(&bufHdr->state) & BM_PERMANENT))
5575 : {
5576 : /*
5577 : * If we must not write WAL, due to a relfilelocator-specific
5578 : * condition or being in recovery, don't dirty the page. We can
5579 : * set the hint, just not dirty the page as a result so the hint
5580 : * is lost when we evict the page or shutdown.
5581 : *
5582 : * See src/backend/storage/page/README for longer discussion.
5583 : */
5584 1939730 : if (RecoveryInProgress() ||
5585 124460 : RelFileLocatorSkippingWAL(BufTagGetRelFileLocator(&bufHdr->tag)))
5586 1693306 : return;
5587 :
5588 : /*
5589 : * If the block is already dirty because we either made a change
5590 : * or set a hint already, then we don't need to write a full page
5591 : * image. Note that aggressive cleaning of blocks dirtied by hint
5592 : * bit setting would increase the call rate. Bulk setting of hint
5593 : * bits would reduce the call rate...
5594 : *
5595 : * We must issue the WAL record before we mark the buffer dirty.
5596 : * Otherwise we might write the page before we write the WAL. That
5597 : * causes a race condition, since a checkpoint might occur between
5598 : * writing the WAL record and marking the buffer dirty. We solve
5599 : * that with a kluge, but one that is already in use during
5600 : * transaction commit to prevent race conditions. Basically, we
5601 : * simply prevent the checkpoint WAL record from being written
5602 : * until we have marked the buffer dirty. We don't start the
5603 : * checkpoint flush until we have marked dirty, so our checkpoint
5604 : * must flush the change to disk successfully or the checkpoint
5605 : * never gets written, so crash recovery will fix.
5606 : *
5607 : * It's possible we may enter here without an xid, so it is
5608 : * essential that CreateCheckPoint waits for virtual transactions
5609 : * rather than full transactionids.
5610 : */
5611 : Assert((MyProc->delayChkptFlags & DELAY_CHKPT_START) == 0);
5612 121964 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
5613 121964 : delayChkptFlags = true;
5614 121964 : lsn = XLogSaveBufferForHint(buffer, buffer_std);
5615 : }
5616 :
5617 124200 : buf_state = LockBufHdr(bufHdr);
5618 :
5619 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5620 :
5621 124200 : if (!(buf_state & BM_DIRTY))
5622 : {
5623 124130 : dirtied = true; /* Means "will be dirtied by this action" */
5624 :
5625 : /*
5626 : * Set the page LSN if we wrote a backup block. We aren't supposed
5627 : * to set this when only holding a share lock but as long as we
5628 : * serialise it somehow we're OK. We choose to set LSN while
5629 : * holding the buffer header lock, which causes any reader of an
5630 : * LSN who holds only a share lock to also obtain a buffer header
5631 : * lock before using PageGetLSN(), which is enforced in
5632 : * BufferGetLSNAtomic().
5633 : *
5634 : * If checksums are enabled, you might think we should reset the
5635 : * checksum here. That will happen when the page is written
5636 : * sometime later in this checkpoint cycle.
5637 : */
5638 124130 : if (XLogRecPtrIsValid(lsn))
5639 62632 : PageSetLSN(page, lsn);
5640 : }
5641 :
5642 124200 : UnlockBufHdrExt(bufHdr, buf_state,
5643 : BM_DIRTY | BM_JUST_DIRTIED,
5644 : 0, 0);
5645 :
5646 124200 : if (delayChkptFlags)
5647 121964 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
5648 :
5649 124200 : if (dirtied)
5650 : {
5651 124130 : pgBufferUsage.shared_blks_dirtied++;
5652 124130 : if (VacuumCostActive)
5653 5644 : VacuumCostBalance += VacuumCostPageDirty;
5654 : }
5655 : }
5656 : }
5657 :
5658 : /*
5659 : * Release buffer content locks for shared buffers.
5660 : *
5661 : * Used to clean up after errors.
5662 : *
5663 : * Currently, we can expect that lwlock.c's LWLockReleaseAll() took care
5664 : * of releasing buffer content locks per se; the only thing we need to deal
5665 : * with here is clearing any PIN_COUNT request that was in progress.
5666 : */
5667 : void
5668 107538 : UnlockBuffers(void)
5669 : {
5670 107538 : BufferDesc *buf = PinCountWaitBuf;
5671 :
5672 107538 : if (buf)
5673 : {
5674 : uint32 buf_state;
5675 0 : uint32 unset_bits = 0;
5676 :
5677 0 : buf_state = LockBufHdr(buf);
5678 :
5679 : /*
5680 : * Don't complain if flag bit not set; it could have been reset but we
5681 : * got a cancel/die interrupt before getting the signal.
5682 : */
5683 0 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5684 0 : buf->wait_backend_pgprocno == MyProcNumber)
5685 0 : unset_bits = BM_PIN_COUNT_WAITER;
5686 :
5687 0 : UnlockBufHdrExt(buf, buf_state,
5688 : 0, unset_bits,
5689 : 0);
5690 :
5691 0 : PinCountWaitBuf = NULL;
5692 : }
5693 107538 : }
5694 :
5695 : /*
5696 : * Acquire or release the content_lock for the buffer.
5697 : */
5698 : void
5699 346754324 : LockBuffer(Buffer buffer, BufferLockMode mode)
5700 : {
5701 : BufferDesc *buf;
5702 :
5703 : Assert(BufferIsPinned(buffer));
5704 346754324 : if (BufferIsLocal(buffer))
5705 19800544 : return; /* local buffers need no lock */
5706 :
5707 326953780 : buf = GetBufferDescriptor(buffer - 1);
5708 :
5709 326953780 : if (mode == BUFFER_LOCK_UNLOCK)
5710 164762216 : LWLockRelease(BufferDescriptorGetContentLock(buf));
5711 162191564 : else if (mode == BUFFER_LOCK_SHARE)
5712 114432898 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_SHARED);
5713 47758666 : else if (mode == BUFFER_LOCK_EXCLUSIVE)
5714 47758666 : LWLockAcquire(BufferDescriptorGetContentLock(buf), LW_EXCLUSIVE);
5715 : else
5716 0 : elog(ERROR, "unrecognized buffer lock mode: %d", mode);
5717 : }
5718 :
5719 : /*
5720 : * Acquire the content_lock for the buffer, but only if we don't have to wait.
5721 : *
5722 : * This assumes the caller wants BUFFER_LOCK_EXCLUSIVE mode.
5723 : */
5724 : bool
5725 2715428 : ConditionalLockBuffer(Buffer buffer)
5726 : {
5727 : BufferDesc *buf;
5728 :
5729 : Assert(BufferIsPinned(buffer));
5730 2715428 : if (BufferIsLocal(buffer))
5731 129578 : return true; /* act as though we got it */
5732 :
5733 2585850 : buf = GetBufferDescriptor(buffer - 1);
5734 :
5735 2585850 : return LWLockConditionalAcquire(BufferDescriptorGetContentLock(buf),
5736 : LW_EXCLUSIVE);
5737 : }
5738 :
5739 : /*
5740 : * Verify that this backend is pinning the buffer exactly once.
5741 : *
5742 : * NOTE: Like in BufferIsPinned(), what we check here is that *this* backend
5743 : * holds a pin on the buffer. We do not care whether some other backend does.
5744 : */
5745 : void
5746 4744946 : CheckBufferIsPinnedOnce(Buffer buffer)
5747 : {
5748 4744946 : if (BufferIsLocal(buffer))
5749 : {
5750 1578 : if (LocalRefCount[-buffer - 1] != 1)
5751 0 : elog(ERROR, "incorrect local pin count: %d",
5752 : LocalRefCount[-buffer - 1]);
5753 : }
5754 : else
5755 : {
5756 4743368 : if (GetPrivateRefCount(buffer) != 1)
5757 0 : elog(ERROR, "incorrect local pin count: %d",
5758 : GetPrivateRefCount(buffer));
5759 : }
5760 4744946 : }
5761 :
5762 : /*
5763 : * LockBufferForCleanup - lock a buffer in preparation for deleting items
5764 : *
5765 : * Items may be deleted from a disk page only when the caller (a) holds an
5766 : * exclusive lock on the buffer and (b) has observed that no other backend
5767 : * holds a pin on the buffer. If there is a pin, then the other backend
5768 : * might have a pointer into the buffer (for example, a heapscan reference
5769 : * to an item --- see README for more details). It's OK if a pin is added
5770 : * after the cleanup starts, however; the newly-arrived backend will be
5771 : * unable to look at the page until we release the exclusive lock.
5772 : *
5773 : * To implement this protocol, a would-be deleter must pin the buffer and
5774 : * then call LockBufferForCleanup(). LockBufferForCleanup() is similar to
5775 : * LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE), except that it loops until
5776 : * it has successfully observed pin count = 1.
5777 : */
5778 : void
5779 49394 : LockBufferForCleanup(Buffer buffer)
5780 : {
5781 : BufferDesc *bufHdr;
5782 49394 : TimestampTz waitStart = 0;
5783 49394 : bool waiting = false;
5784 49394 : bool logged_recovery_conflict = false;
5785 :
5786 : Assert(BufferIsPinned(buffer));
5787 : Assert(PinCountWaitBuf == NULL);
5788 :
5789 49394 : CheckBufferIsPinnedOnce(buffer);
5790 :
5791 : /*
5792 : * We do not yet need to be worried about in-progress AIOs holding a pin,
5793 : * as we, so far, only support doing reads via AIO and this function can
5794 : * only be called once the buffer is valid (i.e. no read can be in
5795 : * flight).
5796 : */
5797 :
5798 : /* Nobody else to wait for */
5799 49394 : if (BufferIsLocal(buffer))
5800 32 : return;
5801 :
5802 49362 : bufHdr = GetBufferDescriptor(buffer - 1);
5803 :
5804 : for (;;)
5805 200 : {
5806 : uint32 buf_state;
5807 49562 : uint32 unset_bits = 0;
5808 :
5809 : /* Try to acquire lock */
5810 49562 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
5811 49562 : buf_state = LockBufHdr(bufHdr);
5812 :
5813 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
5814 49562 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
5815 : {
5816 : /* Successfully acquired exclusive lock with pincount 1 */
5817 49362 : UnlockBufHdr(bufHdr);
5818 :
5819 : /*
5820 : * Emit the log message if recovery conflict on buffer pin was
5821 : * resolved but the startup process waited longer than
5822 : * deadlock_timeout for it.
5823 : */
5824 49362 : if (logged_recovery_conflict)
5825 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
5826 : waitStart, GetCurrentTimestamp(),
5827 : NULL, false);
5828 :
5829 49362 : if (waiting)
5830 : {
5831 : /* reset ps display to remove the suffix if we added one */
5832 4 : set_ps_display_remove_suffix();
5833 4 : waiting = false;
5834 : }
5835 49362 : return;
5836 : }
5837 : /* Failed, so mark myself as waiting for pincount 1 */
5838 200 : if (buf_state & BM_PIN_COUNT_WAITER)
5839 : {
5840 0 : UnlockBufHdr(bufHdr);
5841 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5842 0 : elog(ERROR, "multiple backends attempting to wait for pincount 1");
5843 : }
5844 200 : bufHdr->wait_backend_pgprocno = MyProcNumber;
5845 200 : PinCountWaitBuf = bufHdr;
5846 200 : UnlockBufHdrExt(bufHdr, buf_state,
5847 : BM_PIN_COUNT_WAITER, 0,
5848 : 0);
5849 200 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5850 :
5851 : /* Wait to be signaled by UnpinBuffer() */
5852 200 : if (InHotStandby)
5853 : {
5854 20 : if (!waiting)
5855 : {
5856 : /* adjust the process title to indicate that it's waiting */
5857 4 : set_ps_display_suffix("waiting");
5858 4 : waiting = true;
5859 : }
5860 :
5861 : /*
5862 : * Emit the log message if the startup process is waiting longer
5863 : * than deadlock_timeout for recovery conflict on buffer pin.
5864 : *
5865 : * Skip this if first time through because the startup process has
5866 : * not started waiting yet in this case. So, the wait start
5867 : * timestamp is set after this logic.
5868 : */
5869 20 : if (waitStart != 0 && !logged_recovery_conflict)
5870 : {
5871 8 : TimestampTz now = GetCurrentTimestamp();
5872 :
5873 8 : if (TimestampDifferenceExceeds(waitStart, now,
5874 : DeadlockTimeout))
5875 : {
5876 4 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN,
5877 : waitStart, now, NULL, true);
5878 4 : logged_recovery_conflict = true;
5879 : }
5880 : }
5881 :
5882 : /*
5883 : * Set the wait start timestamp if logging is enabled and first
5884 : * time through.
5885 : */
5886 20 : if (log_recovery_conflict_waits && waitStart == 0)
5887 4 : waitStart = GetCurrentTimestamp();
5888 :
5889 : /* Publish the bufid that Startup process waits on */
5890 20 : SetStartupBufferPinWaitBufId(buffer - 1);
5891 : /* Set alarm and then wait to be signaled by UnpinBuffer() */
5892 20 : ResolveRecoveryConflictWithBufferPin();
5893 : /* Reset the published bufid */
5894 20 : SetStartupBufferPinWaitBufId(-1);
5895 : }
5896 : else
5897 180 : ProcWaitForSignal(WAIT_EVENT_BUFFER_CLEANUP);
5898 :
5899 : /*
5900 : * Remove flag marking us as waiter. Normally this will not be set
5901 : * anymore, but ProcWaitForSignal() can return for other signals as
5902 : * well. We take care to only reset the flag if we're the waiter, as
5903 : * theoretically another backend could have started waiting. That's
5904 : * impossible with the current usages due to table level locking, but
5905 : * better be safe.
5906 : */
5907 200 : buf_state = LockBufHdr(bufHdr);
5908 200 : if ((buf_state & BM_PIN_COUNT_WAITER) != 0 &&
5909 16 : bufHdr->wait_backend_pgprocno == MyProcNumber)
5910 16 : unset_bits |= BM_PIN_COUNT_WAITER;
5911 :
5912 200 : UnlockBufHdrExt(bufHdr, buf_state,
5913 : 0, unset_bits,
5914 : 0);
5915 :
5916 200 : PinCountWaitBuf = NULL;
5917 : /* Loop back and try again */
5918 : }
5919 : }
5920 :
5921 : /*
5922 : * Check called from ProcessRecoveryConflictInterrupts() when Startup process
5923 : * requests cancellation of all pin holders that are blocking it.
5924 : */
5925 : bool
5926 6 : HoldingBufferPinThatDelaysRecovery(void)
5927 : {
5928 6 : int bufid = GetStartupBufferPinWaitBufId();
5929 :
5930 : /*
5931 : * If we get woken slowly then it's possible that the Startup process was
5932 : * already woken by other backends before we got here. Also possible that
5933 : * we get here by multiple interrupts or interrupts at inappropriate
5934 : * times, so make sure we do nothing if the bufid is not set.
5935 : */
5936 6 : if (bufid < 0)
5937 2 : return false;
5938 :
5939 4 : if (GetPrivateRefCount(bufid + 1) > 0)
5940 4 : return true;
5941 :
5942 0 : return false;
5943 : }
5944 :
5945 : /*
5946 : * ConditionalLockBufferForCleanup - as above, but don't wait to get the lock
5947 : *
5948 : * We won't loop, but just check once to see if the pin count is OK. If
5949 : * not, return false with no lock held.
5950 : */
5951 : bool
5952 865870 : ConditionalLockBufferForCleanup(Buffer buffer)
5953 : {
5954 : BufferDesc *bufHdr;
5955 : uint32 buf_state,
5956 : refcount;
5957 :
5958 : Assert(BufferIsValid(buffer));
5959 :
5960 : /* see AIO related comment in LockBufferForCleanup() */
5961 :
5962 865870 : if (BufferIsLocal(buffer))
5963 : {
5964 1608 : refcount = LocalRefCount[-buffer - 1];
5965 : /* There should be exactly one pin */
5966 : Assert(refcount > 0);
5967 1608 : if (refcount != 1)
5968 42 : return false;
5969 : /* Nobody else to wait for */
5970 1566 : return true;
5971 : }
5972 :
5973 : /* There should be exactly one local pin */
5974 864262 : refcount = GetPrivateRefCount(buffer);
5975 : Assert(refcount);
5976 864262 : if (refcount != 1)
5977 462 : return false;
5978 :
5979 : /* Try to acquire lock */
5980 863800 : if (!ConditionalLockBuffer(buffer))
5981 88 : return false;
5982 :
5983 863712 : bufHdr = GetBufferDescriptor(buffer - 1);
5984 863712 : buf_state = LockBufHdr(bufHdr);
5985 863712 : refcount = BUF_STATE_GET_REFCOUNT(buf_state);
5986 :
5987 : Assert(refcount > 0);
5988 863712 : if (refcount == 1)
5989 : {
5990 : /* Successfully acquired exclusive lock with pincount 1 */
5991 863124 : UnlockBufHdr(bufHdr);
5992 863124 : return true;
5993 : }
5994 :
5995 : /* Failed, so release the lock */
5996 588 : UnlockBufHdr(bufHdr);
5997 588 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5998 588 : return false;
5999 : }
6000 :
6001 : /*
6002 : * IsBufferCleanupOK - as above, but we already have the lock
6003 : *
6004 : * Check whether it's OK to perform cleanup on a buffer we've already
6005 : * locked. If we observe that the pin count is 1, our exclusive lock
6006 : * happens to be a cleanup lock, and we can proceed with anything that
6007 : * would have been allowable had we sought a cleanup lock originally.
6008 : */
6009 : bool
6010 3314 : IsBufferCleanupOK(Buffer buffer)
6011 : {
6012 : BufferDesc *bufHdr;
6013 : uint32 buf_state;
6014 :
6015 : Assert(BufferIsValid(buffer));
6016 :
6017 : /* see AIO related comment in LockBufferForCleanup() */
6018 :
6019 3314 : if (BufferIsLocal(buffer))
6020 : {
6021 : /* There should be exactly one pin */
6022 0 : if (LocalRefCount[-buffer - 1] != 1)
6023 0 : return false;
6024 : /* Nobody else to wait for */
6025 0 : return true;
6026 : }
6027 :
6028 : /* There should be exactly one local pin */
6029 3314 : if (GetPrivateRefCount(buffer) != 1)
6030 0 : return false;
6031 :
6032 3314 : bufHdr = GetBufferDescriptor(buffer - 1);
6033 :
6034 : /* caller must hold exclusive lock on buffer */
6035 : Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE));
6036 :
6037 3314 : buf_state = LockBufHdr(bufHdr);
6038 :
6039 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6040 3314 : if (BUF_STATE_GET_REFCOUNT(buf_state) == 1)
6041 : {
6042 : /* pincount is OK. */
6043 3312 : UnlockBufHdr(bufHdr);
6044 3312 : return true;
6045 : }
6046 :
6047 2 : UnlockBufHdr(bufHdr);
6048 2 : return false;
6049 : }
6050 :
6051 :
6052 : /*
6053 : * Functions for buffer I/O handling
6054 : *
6055 : * Also note that these are used only for shared buffers, not local ones.
6056 : */
6057 :
6058 : /*
6059 : * WaitIO -- Block until the IO_IN_PROGRESS flag on 'buf' is cleared.
6060 : */
6061 : static void
6062 14354 : WaitIO(BufferDesc *buf)
6063 : {
6064 14354 : ConditionVariable *cv = BufferDescriptorGetIOCV(buf);
6065 :
6066 14354 : ConditionVariablePrepareToSleep(cv);
6067 : for (;;)
6068 14270 : {
6069 : uint32 buf_state;
6070 : PgAioWaitRef iow;
6071 :
6072 : /*
6073 : * It may not be necessary to acquire the spinlock to check the flag
6074 : * here, but since this test is essential for correctness, we'd better
6075 : * play it safe.
6076 : */
6077 28624 : buf_state = LockBufHdr(buf);
6078 :
6079 : /*
6080 : * Copy the wait reference while holding the spinlock. This protects
6081 : * against a concurrent TerminateBufferIO() in another backend from
6082 : * clearing the wref while it's being read.
6083 : */
6084 28624 : iow = buf->io_wref;
6085 28624 : UnlockBufHdr(buf);
6086 :
6087 : /* no IO in progress, we don't need to wait */
6088 28624 : if (!(buf_state & BM_IO_IN_PROGRESS))
6089 14354 : break;
6090 :
6091 : /*
6092 : * The buffer has asynchronous IO in progress, wait for it to
6093 : * complete.
6094 : */
6095 14270 : if (pgaio_wref_valid(&iow))
6096 : {
6097 12354 : pgaio_wref_wait(&iow);
6098 :
6099 : /*
6100 : * The AIO subsystem internally uses condition variables and thus
6101 : * might remove this backend from the BufferDesc's CV. While that
6102 : * wouldn't cause a correctness issue (the first CV sleep just
6103 : * immediately returns if not already registered), it seems worth
6104 : * avoiding unnecessary loop iterations, given that we take care
6105 : * to do so at the start of the function.
6106 : */
6107 12354 : ConditionVariablePrepareToSleep(cv);
6108 12354 : continue;
6109 : }
6110 :
6111 : /* wait on BufferDesc->cv, e.g. for concurrent synchronous IO */
6112 1916 : ConditionVariableSleep(cv, WAIT_EVENT_BUFFER_IO);
6113 : }
6114 14354 : ConditionVariableCancelSleep();
6115 14354 : }
6116 :
6117 : /*
6118 : * StartBufferIO: begin I/O on this buffer
6119 : * (Assumptions)
6120 : * My process is executing no IO on this buffer
6121 : * The buffer is Pinned
6122 : *
6123 : * In some scenarios multiple backends could attempt the same I/O operation
6124 : * concurrently. If someone else has already started I/O on this buffer then
6125 : * we will wait for completion of the IO using WaitIO().
6126 : *
6127 : * Input operations are only attempted on buffers that are not BM_VALID,
6128 : * and output operations only on buffers that are BM_VALID and BM_DIRTY,
6129 : * so we can always tell if the work is already done.
6130 : *
6131 : * Returns true if we successfully marked the buffer as I/O busy,
6132 : * false if someone else already did the work.
6133 : *
6134 : * If nowait is true, then we don't wait for an I/O to be finished by another
6135 : * backend. In that case, false indicates either that the I/O was already
6136 : * finished, or is still in progress. This is useful for callers that want to
6137 : * find out if they can perform the I/O as part of a larger operation, without
6138 : * waiting for the answer or distinguishing the reasons why not.
6139 : */
6140 : bool
6141 5064926 : StartBufferIO(BufferDesc *buf, bool forInput, bool nowait)
6142 : {
6143 : uint32 buf_state;
6144 :
6145 5064926 : ResourceOwnerEnlarge(CurrentResourceOwner);
6146 :
6147 : for (;;)
6148 : {
6149 5079274 : buf_state = LockBufHdr(buf);
6150 :
6151 5079274 : if (!(buf_state & BM_IO_IN_PROGRESS))
6152 5064918 : break;
6153 14356 : UnlockBufHdr(buf);
6154 14356 : if (nowait)
6155 8 : return false;
6156 14348 : WaitIO(buf);
6157 : }
6158 :
6159 : /* Once we get here, there is definitely no I/O active on this buffer */
6160 :
6161 : /* Check if someone else already did the I/O */
6162 5064918 : if (forInput ? (buf_state & BM_VALID) : !(buf_state & BM_DIRTY))
6163 : {
6164 15846 : UnlockBufHdr(buf);
6165 15846 : return false;
6166 : }
6167 :
6168 5049072 : UnlockBufHdrExt(buf, buf_state,
6169 : BM_IO_IN_PROGRESS, 0,
6170 : 0);
6171 :
6172 5049072 : ResourceOwnerRememberBufferIO(CurrentResourceOwner,
6173 : BufferDescriptorGetBuffer(buf));
6174 :
6175 5049072 : return true;
6176 : }
6177 :
6178 : /*
6179 : * TerminateBufferIO: release a buffer we were doing I/O on
6180 : * (Assumptions)
6181 : * My process is executing IO for the buffer
6182 : * BM_IO_IN_PROGRESS bit is set for the buffer
6183 : * The buffer is Pinned
6184 : *
6185 : * If clear_dirty is true and BM_JUST_DIRTIED is not set, we clear the
6186 : * buffer's BM_DIRTY flag. This is appropriate when terminating a
6187 : * successful write. The check on BM_JUST_DIRTIED is necessary to avoid
6188 : * marking the buffer clean if it was re-dirtied while we were writing.
6189 : *
6190 : * set_flag_bits gets ORed into the buffer's flags. It must include
6191 : * BM_IO_ERROR in a failure case. For successful completion it could
6192 : * be 0, or BM_VALID if we just finished reading in the page.
6193 : *
6194 : * If forget_owner is true, we release the buffer I/O from the current
6195 : * resource owner. (forget_owner=false is used when the resource owner itself
6196 : * is being released)
6197 : */
6198 : void
6199 4771594 : TerminateBufferIO(BufferDesc *buf, bool clear_dirty, uint32 set_flag_bits,
6200 : bool forget_owner, bool release_aio)
6201 : {
6202 : uint32 buf_state;
6203 4771594 : uint32 unset_flag_bits = 0;
6204 4771594 : int refcount_change = 0;
6205 :
6206 4771594 : buf_state = LockBufHdr(buf);
6207 :
6208 : Assert(buf_state & BM_IO_IN_PROGRESS);
6209 4771594 : unset_flag_bits |= BM_IO_IN_PROGRESS;
6210 :
6211 : /* Clear earlier errors, if this IO failed, it'll be marked again */
6212 4771594 : unset_flag_bits |= BM_IO_ERROR;
6213 :
6214 4771594 : if (clear_dirty && !(buf_state & BM_JUST_DIRTIED))
6215 1146318 : unset_flag_bits |= BM_DIRTY | BM_CHECKPOINT_NEEDED;
6216 :
6217 4771594 : if (release_aio)
6218 : {
6219 : /* release ownership by the AIO subsystem */
6220 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) > 0);
6221 2609600 : refcount_change = -1;
6222 2609600 : pgaio_wref_clear(&buf->io_wref);
6223 : }
6224 :
6225 4771594 : buf_state = UnlockBufHdrExt(buf, buf_state,
6226 : set_flag_bits, unset_flag_bits,
6227 : refcount_change);
6228 :
6229 4771594 : if (forget_owner)
6230 2161952 : ResourceOwnerForgetBufferIO(CurrentResourceOwner,
6231 : BufferDescriptorGetBuffer(buf));
6232 :
6233 4771594 : ConditionVariableBroadcast(BufferDescriptorGetIOCV(buf));
6234 :
6235 : /*
6236 : * Support LockBufferForCleanup()
6237 : *
6238 : * We may have just released the last pin other than the waiter's. In most
6239 : * cases, this backend holds another pin on the buffer. But, if, for
6240 : * example, this backend is completing an IO issued by another backend, it
6241 : * may be time to wake the waiter.
6242 : */
6243 4771594 : if (release_aio && (buf_state & BM_PIN_COUNT_WAITER))
6244 0 : WakePinCountWaiter(buf);
6245 4771594 : }
6246 :
6247 : /*
6248 : * AbortBufferIO: Clean up active buffer I/O after an error.
6249 : *
6250 : * All LWLocks we might have held have been released,
6251 : * but we haven't yet released buffer pins, so the buffer is still pinned.
6252 : *
6253 : * If I/O was in progress, we always set BM_IO_ERROR, even though it's
6254 : * possible the error condition wasn't related to the I/O.
6255 : *
6256 : * Note: this does not remove the buffer I/O from the resource owner.
6257 : * That's correct when we're releasing the whole resource owner, but
6258 : * beware if you use this in other contexts.
6259 : */
6260 : static void
6261 30 : AbortBufferIO(Buffer buffer)
6262 : {
6263 30 : BufferDesc *buf_hdr = GetBufferDescriptor(buffer - 1);
6264 : uint32 buf_state;
6265 :
6266 30 : buf_state = LockBufHdr(buf_hdr);
6267 : Assert(buf_state & (BM_IO_IN_PROGRESS | BM_TAG_VALID));
6268 :
6269 30 : if (!(buf_state & BM_VALID))
6270 : {
6271 : Assert(!(buf_state & BM_DIRTY));
6272 30 : UnlockBufHdr(buf_hdr);
6273 : }
6274 : else
6275 : {
6276 : Assert(buf_state & BM_DIRTY);
6277 0 : UnlockBufHdr(buf_hdr);
6278 :
6279 : /* Issue notice if this is not the first failure... */
6280 0 : if (buf_state & BM_IO_ERROR)
6281 : {
6282 : /* Buffer is pinned, so we can read tag without spinlock */
6283 0 : ereport(WARNING,
6284 : (errcode(ERRCODE_IO_ERROR),
6285 : errmsg("could not write block %u of %s",
6286 : buf_hdr->tag.blockNum,
6287 : relpathperm(BufTagGetRelFileLocator(&buf_hdr->tag),
6288 : BufTagGetForkNum(&buf_hdr->tag)).str),
6289 : errdetail("Multiple failures --- write error might be permanent.")));
6290 : }
6291 : }
6292 :
6293 30 : TerminateBufferIO(buf_hdr, false, BM_IO_ERROR, false, false);
6294 30 : }
6295 :
6296 : /*
6297 : * Error context callback for errors occurring during shared buffer writes.
6298 : */
6299 : static void
6300 102 : shared_buffer_write_error_callback(void *arg)
6301 : {
6302 102 : BufferDesc *bufHdr = (BufferDesc *) arg;
6303 :
6304 : /* Buffer is pinned, so we can read the tag without locking the spinlock */
6305 102 : if (bufHdr != NULL)
6306 204 : errcontext("writing block %u of relation \"%s\"",
6307 : bufHdr->tag.blockNum,
6308 102 : relpathperm(BufTagGetRelFileLocator(&bufHdr->tag),
6309 : BufTagGetForkNum(&bufHdr->tag)).str);
6310 102 : }
6311 :
6312 : /*
6313 : * Error context callback for errors occurring during local buffer writes.
6314 : */
6315 : static void
6316 0 : local_buffer_write_error_callback(void *arg)
6317 : {
6318 0 : BufferDesc *bufHdr = (BufferDesc *) arg;
6319 :
6320 0 : if (bufHdr != NULL)
6321 0 : errcontext("writing block %u of relation \"%s\"",
6322 : bufHdr->tag.blockNum,
6323 0 : relpathbackend(BufTagGetRelFileLocator(&bufHdr->tag),
6324 : MyProcNumber,
6325 : BufTagGetForkNum(&bufHdr->tag)).str);
6326 0 : }
6327 :
6328 : /*
6329 : * RelFileLocator qsort/bsearch comparator; see RelFileLocatorEquals.
6330 : */
6331 : static int
6332 19418486 : rlocator_comparator(const void *p1, const void *p2)
6333 : {
6334 19418486 : RelFileLocator n1 = *(const RelFileLocator *) p1;
6335 19418486 : RelFileLocator n2 = *(const RelFileLocator *) p2;
6336 :
6337 19418486 : if (n1.relNumber < n2.relNumber)
6338 19349636 : return -1;
6339 68850 : else if (n1.relNumber > n2.relNumber)
6340 65812 : return 1;
6341 :
6342 3038 : if (n1.dbOid < n2.dbOid)
6343 0 : return -1;
6344 3038 : else if (n1.dbOid > n2.dbOid)
6345 0 : return 1;
6346 :
6347 3038 : if (n1.spcOid < n2.spcOid)
6348 0 : return -1;
6349 3038 : else if (n1.spcOid > n2.spcOid)
6350 0 : return 1;
6351 : else
6352 3038 : return 0;
6353 : }
6354 :
6355 : /*
6356 : * Lock buffer header - set BM_LOCKED in buffer state.
6357 : */
6358 : uint32
6359 64954644 : LockBufHdr(BufferDesc *desc)
6360 : {
6361 : uint32 old_buf_state;
6362 :
6363 : Assert(!BufferIsLocal(BufferDescriptorGetBuffer(desc)));
6364 :
6365 : while (true)
6366 : {
6367 : /*
6368 : * Always try once to acquire the lock directly, without setting up
6369 : * the spin-delay infrastructure. The work necessary for that shows up
6370 : * in profiles and is rarely necessary.
6371 : */
6372 64958738 : old_buf_state = pg_atomic_fetch_or_u32(&desc->state, BM_LOCKED);
6373 64958738 : if (likely(!(old_buf_state & BM_LOCKED)))
6374 64954644 : break; /* got lock */
6375 :
6376 : /* and then spin without atomic operations until lock is released */
6377 : {
6378 : SpinDelayStatus delayStatus;
6379 :
6380 4094 : init_local_spin_delay(&delayStatus);
6381 :
6382 8200 : while (old_buf_state & BM_LOCKED)
6383 : {
6384 4106 : perform_spin_delay(&delayStatus);
6385 4106 : old_buf_state = pg_atomic_read_u32(&desc->state);
6386 : }
6387 4094 : finish_spin_delay(&delayStatus);
6388 : }
6389 :
6390 : /*
6391 : * Retry. The lock might obviously already be re-acquired by the time
6392 : * we're attempting to get it again.
6393 : */
6394 : }
6395 :
6396 64954644 : return old_buf_state | BM_LOCKED;
6397 : }
6398 :
6399 : /*
6400 : * Wait until the BM_LOCKED flag isn't set anymore and return the buffer's
6401 : * state at that point.
6402 : *
6403 : * Obviously the buffer could be locked by the time the value is returned, so
6404 : * this is primarily useful in CAS style loops.
6405 : */
6406 : pg_noinline uint32
6407 1236 : WaitBufHdrUnlocked(BufferDesc *buf)
6408 : {
6409 : SpinDelayStatus delayStatus;
6410 : uint32 buf_state;
6411 :
6412 1236 : init_local_spin_delay(&delayStatus);
6413 :
6414 1236 : buf_state = pg_atomic_read_u32(&buf->state);
6415 :
6416 134152 : while (buf_state & BM_LOCKED)
6417 : {
6418 132916 : perform_spin_delay(&delayStatus);
6419 132916 : buf_state = pg_atomic_read_u32(&buf->state);
6420 : }
6421 :
6422 1236 : finish_spin_delay(&delayStatus);
6423 :
6424 1236 : return buf_state;
6425 : }
6426 :
6427 : /*
6428 : * BufferTag comparator.
6429 : */
6430 : static inline int
6431 0 : buffertag_comparator(const BufferTag *ba, const BufferTag *bb)
6432 : {
6433 : int ret;
6434 : RelFileLocator rlocatora;
6435 : RelFileLocator rlocatorb;
6436 :
6437 0 : rlocatora = BufTagGetRelFileLocator(ba);
6438 0 : rlocatorb = BufTagGetRelFileLocator(bb);
6439 :
6440 0 : ret = rlocator_comparator(&rlocatora, &rlocatorb);
6441 :
6442 0 : if (ret != 0)
6443 0 : return ret;
6444 :
6445 0 : if (BufTagGetForkNum(ba) < BufTagGetForkNum(bb))
6446 0 : return -1;
6447 0 : if (BufTagGetForkNum(ba) > BufTagGetForkNum(bb))
6448 0 : return 1;
6449 :
6450 0 : if (ba->blockNum < bb->blockNum)
6451 0 : return -1;
6452 0 : if (ba->blockNum > bb->blockNum)
6453 0 : return 1;
6454 :
6455 0 : return 0;
6456 : }
6457 :
6458 : /*
6459 : * Comparator determining the writeout order in a checkpoint.
6460 : *
6461 : * It is important that tablespaces are compared first, the logic balancing
6462 : * writes between tablespaces relies on it.
6463 : */
6464 : static inline int
6465 6064122 : ckpt_buforder_comparator(const CkptSortItem *a, const CkptSortItem *b)
6466 : {
6467 : /* compare tablespace */
6468 6064122 : if (a->tsId < b->tsId)
6469 20132 : return -1;
6470 6043990 : else if (a->tsId > b->tsId)
6471 51978 : return 1;
6472 : /* compare relation */
6473 5992012 : if (a->relNumber < b->relNumber)
6474 1679230 : return -1;
6475 4312782 : else if (a->relNumber > b->relNumber)
6476 1636134 : return 1;
6477 : /* compare fork */
6478 2676648 : else if (a->forkNum < b->forkNum)
6479 119698 : return -1;
6480 2556950 : else if (a->forkNum > b->forkNum)
6481 121660 : return 1;
6482 : /* compare block number */
6483 2435290 : else if (a->blockNum < b->blockNum)
6484 1193614 : return -1;
6485 1241676 : else if (a->blockNum > b->blockNum)
6486 1168254 : return 1;
6487 : /* equal page IDs are unlikely, but not impossible */
6488 73422 : return 0;
6489 : }
6490 :
6491 : /*
6492 : * Comparator for a Min-Heap over the per-tablespace checkpoint completion
6493 : * progress.
6494 : */
6495 : static int
6496 492606 : ts_ckpt_progress_comparator(Datum a, Datum b, void *arg)
6497 : {
6498 492606 : CkptTsStatus *sa = (CkptTsStatus *) DatumGetPointer(a);
6499 492606 : CkptTsStatus *sb = (CkptTsStatus *) DatumGetPointer(b);
6500 :
6501 : /* we want a min-heap, so return 1 for the a < b */
6502 492606 : if (sa->progress < sb->progress)
6503 445886 : return 1;
6504 46720 : else if (sa->progress == sb->progress)
6505 1658 : return 0;
6506 : else
6507 45062 : return -1;
6508 : }
6509 :
6510 : /*
6511 : * Initialize a writeback context, discarding potential previous state.
6512 : *
6513 : * *max_pending is a pointer instead of an immediate value, so the coalesce
6514 : * limits can easily changed by the GUC mechanism, and so calling code does
6515 : * not have to check the current configuration. A value of 0 means that no
6516 : * writeback control will be performed.
6517 : */
6518 : void
6519 5578 : WritebackContextInit(WritebackContext *context, int *max_pending)
6520 : {
6521 : Assert(*max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
6522 :
6523 5578 : context->max_pending = max_pending;
6524 5578 : context->nr_pending = 0;
6525 5578 : }
6526 :
6527 : /*
6528 : * Add buffer to list of pending writeback requests.
6529 : */
6530 : void
6531 1139494 : ScheduleBufferTagForWriteback(WritebackContext *wb_context, IOContext io_context,
6532 : BufferTag *tag)
6533 : {
6534 : PendingWriteback *pending;
6535 :
6536 : /*
6537 : * As pg_flush_data() doesn't do anything with fsync disabled, there's no
6538 : * point in tracking in that case.
6539 : */
6540 1139494 : if (io_direct_flags & IO_DIRECT_DATA ||
6541 1138474 : !enableFsync)
6542 1139490 : return;
6543 :
6544 : /*
6545 : * Add buffer to the pending writeback array, unless writeback control is
6546 : * disabled.
6547 : */
6548 4 : if (*wb_context->max_pending > 0)
6549 : {
6550 : Assert(*wb_context->max_pending <= WRITEBACK_MAX_PENDING_FLUSHES);
6551 :
6552 0 : pending = &wb_context->pending_writebacks[wb_context->nr_pending++];
6553 :
6554 0 : pending->tag = *tag;
6555 : }
6556 :
6557 : /*
6558 : * Perform pending flushes if the writeback limit is exceeded. This
6559 : * includes the case where previously an item has been added, but control
6560 : * is now disabled.
6561 : */
6562 4 : if (wb_context->nr_pending >= *wb_context->max_pending)
6563 4 : IssuePendingWritebacks(wb_context, io_context);
6564 : }
6565 :
6566 : #define ST_SORT sort_pending_writebacks
6567 : #define ST_ELEMENT_TYPE PendingWriteback
6568 : #define ST_COMPARE(a, b) buffertag_comparator(&a->tag, &b->tag)
6569 : #define ST_SCOPE static
6570 : #define ST_DEFINE
6571 : #include "lib/sort_template.h"
6572 :
6573 : /*
6574 : * Issue all pending writeback requests, previously scheduled with
6575 : * ScheduleBufferTagForWriteback, to the OS.
6576 : *
6577 : * Because this is only used to improve the OSs IO scheduling we try to never
6578 : * error out - it's just a hint.
6579 : */
6580 : void
6581 2176 : IssuePendingWritebacks(WritebackContext *wb_context, IOContext io_context)
6582 : {
6583 : instr_time io_start;
6584 : int i;
6585 :
6586 2176 : if (wb_context->nr_pending == 0)
6587 2176 : return;
6588 :
6589 : /*
6590 : * Executing the writes in-order can make them a lot faster, and allows to
6591 : * merge writeback requests to consecutive blocks into larger writebacks.
6592 : */
6593 0 : sort_pending_writebacks(wb_context->pending_writebacks,
6594 0 : wb_context->nr_pending);
6595 :
6596 0 : io_start = pgstat_prepare_io_time(track_io_timing);
6597 :
6598 : /*
6599 : * Coalesce neighbouring writes, but nothing else. For that we iterate
6600 : * through the, now sorted, array of pending flushes, and look forward to
6601 : * find all neighbouring (or identical) writes.
6602 : */
6603 0 : for (i = 0; i < wb_context->nr_pending; i++)
6604 : {
6605 : PendingWriteback *cur;
6606 : PendingWriteback *next;
6607 : SMgrRelation reln;
6608 : int ahead;
6609 : BufferTag tag;
6610 : RelFileLocator currlocator;
6611 0 : Size nblocks = 1;
6612 :
6613 0 : cur = &wb_context->pending_writebacks[i];
6614 0 : tag = cur->tag;
6615 0 : currlocator = BufTagGetRelFileLocator(&tag);
6616 :
6617 : /*
6618 : * Peek ahead, into following writeback requests, to see if they can
6619 : * be combined with the current one.
6620 : */
6621 0 : for (ahead = 0; i + ahead + 1 < wb_context->nr_pending; ahead++)
6622 : {
6623 :
6624 0 : next = &wb_context->pending_writebacks[i + ahead + 1];
6625 :
6626 : /* different file, stop */
6627 0 : if (!RelFileLocatorEquals(currlocator,
6628 0 : BufTagGetRelFileLocator(&next->tag)) ||
6629 0 : BufTagGetForkNum(&cur->tag) != BufTagGetForkNum(&next->tag))
6630 : break;
6631 :
6632 : /* ok, block queued twice, skip */
6633 0 : if (cur->tag.blockNum == next->tag.blockNum)
6634 0 : continue;
6635 :
6636 : /* only merge consecutive writes */
6637 0 : if (cur->tag.blockNum + 1 != next->tag.blockNum)
6638 0 : break;
6639 :
6640 0 : nblocks++;
6641 0 : cur = next;
6642 : }
6643 :
6644 0 : i += ahead;
6645 :
6646 : /* and finally tell the kernel to write the data to storage */
6647 0 : reln = smgropen(currlocator, INVALID_PROC_NUMBER);
6648 0 : smgrwriteback(reln, BufTagGetForkNum(&tag), tag.blockNum, nblocks);
6649 : }
6650 :
6651 : /*
6652 : * Assume that writeback requests are only issued for buffers containing
6653 : * blocks of permanent relations.
6654 : */
6655 0 : pgstat_count_io_op_time(IOOBJECT_RELATION, io_context,
6656 0 : IOOP_WRITEBACK, io_start, wb_context->nr_pending, 0);
6657 :
6658 0 : wb_context->nr_pending = 0;
6659 : }
6660 :
6661 : /* ResourceOwner callbacks */
6662 :
6663 : static void
6664 30 : ResOwnerReleaseBufferIO(Datum res)
6665 : {
6666 30 : Buffer buffer = DatumGetInt32(res);
6667 :
6668 30 : AbortBufferIO(buffer);
6669 30 : }
6670 :
6671 : static char *
6672 0 : ResOwnerPrintBufferIO(Datum res)
6673 : {
6674 0 : Buffer buffer = DatumGetInt32(res);
6675 :
6676 0 : return psprintf("lost track of buffer IO on buffer %d", buffer);
6677 : }
6678 :
6679 : static void
6680 15204 : ResOwnerReleaseBufferPin(Datum res)
6681 : {
6682 15204 : Buffer buffer = DatumGetInt32(res);
6683 :
6684 : /* Like ReleaseBuffer, but don't call ResourceOwnerForgetBuffer */
6685 15204 : if (!BufferIsValid(buffer))
6686 0 : elog(ERROR, "bad buffer ID: %d", buffer);
6687 :
6688 15204 : if (BufferIsLocal(buffer))
6689 6066 : UnpinLocalBufferNoOwner(buffer);
6690 : else
6691 9138 : UnpinBufferNoOwner(GetBufferDescriptor(buffer - 1));
6692 15204 : }
6693 :
6694 : static char *
6695 0 : ResOwnerPrintBufferPin(Datum res)
6696 : {
6697 0 : return DebugPrintBufferRefcount(DatumGetInt32(res));
6698 : }
6699 :
6700 : /*
6701 : * Helper function to evict unpinned buffer whose buffer header lock is
6702 : * already acquired.
6703 : */
6704 : static bool
6705 4286 : EvictUnpinnedBufferInternal(BufferDesc *desc, bool *buffer_flushed)
6706 : {
6707 : uint32 buf_state;
6708 : bool result;
6709 :
6710 4286 : *buffer_flushed = false;
6711 :
6712 4286 : buf_state = pg_atomic_read_u32(&(desc->state));
6713 : Assert(buf_state & BM_LOCKED);
6714 :
6715 4286 : if ((buf_state & BM_VALID) == 0)
6716 : {
6717 0 : UnlockBufHdr(desc);
6718 0 : return false;
6719 : }
6720 :
6721 : /* Check that it's not pinned already. */
6722 4286 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
6723 : {
6724 0 : UnlockBufHdr(desc);
6725 0 : return false;
6726 : }
6727 :
6728 4286 : PinBuffer_Locked(desc); /* releases spinlock */
6729 :
6730 : /* If it was dirty, try to clean it once. */
6731 4286 : if (buf_state & BM_DIRTY)
6732 : {
6733 1946 : FlushUnlockedBuffer(desc, NULL, IOOBJECT_RELATION, IOCONTEXT_NORMAL);
6734 1946 : *buffer_flushed = true;
6735 : }
6736 :
6737 : /* This will return false if it becomes dirty or someone else pins it. */
6738 4286 : result = InvalidateVictimBuffer(desc);
6739 :
6740 4286 : UnpinBuffer(desc);
6741 :
6742 4286 : return result;
6743 : }
6744 :
6745 : /*
6746 : * Try to evict the current block in a shared buffer.
6747 : *
6748 : * This function is intended for testing/development use only!
6749 : *
6750 : * To succeed, the buffer must not be pinned on entry, so if the caller had a
6751 : * particular block in mind, it might already have been replaced by some other
6752 : * block by the time this function runs. It's also unpinned on return, so the
6753 : * buffer might be occupied again by the time control is returned, potentially
6754 : * even by the same block. This inherent raciness without other interlocking
6755 : * makes the function unsuitable for non-testing usage.
6756 : *
6757 : * *buffer_flushed is set to true if the buffer was dirty and has been
6758 : * flushed, false otherwise. However, *buffer_flushed=true does not
6759 : * necessarily mean that we flushed the buffer, it could have been flushed by
6760 : * someone else.
6761 : *
6762 : * Returns true if the buffer was valid and it has now been made invalid.
6763 : * Returns false if it wasn't valid, if it couldn't be evicted due to a pin,
6764 : * or if the buffer becomes dirty again while we're trying to write it out.
6765 : */
6766 : bool
6767 280 : EvictUnpinnedBuffer(Buffer buf, bool *buffer_flushed)
6768 : {
6769 : BufferDesc *desc;
6770 :
6771 : Assert(BufferIsValid(buf) && !BufferIsLocal(buf));
6772 :
6773 : /* Make sure we can pin the buffer. */
6774 280 : ResourceOwnerEnlarge(CurrentResourceOwner);
6775 280 : ReservePrivateRefCountEntry();
6776 :
6777 280 : desc = GetBufferDescriptor(buf - 1);
6778 280 : LockBufHdr(desc);
6779 :
6780 280 : return EvictUnpinnedBufferInternal(desc, buffer_flushed);
6781 : }
6782 :
6783 : /*
6784 : * Try to evict all the shared buffers.
6785 : *
6786 : * This function is intended for testing/development use only! See
6787 : * EvictUnpinnedBuffer().
6788 : *
6789 : * The buffers_* parameters are mandatory and indicate the total count of
6790 : * buffers that:
6791 : * - buffers_evicted - were evicted
6792 : * - buffers_flushed - were flushed
6793 : * - buffers_skipped - could not be evicted
6794 : */
6795 : void
6796 2 : EvictAllUnpinnedBuffers(int32 *buffers_evicted, int32 *buffers_flushed,
6797 : int32 *buffers_skipped)
6798 : {
6799 2 : *buffers_evicted = 0;
6800 2 : *buffers_skipped = 0;
6801 2 : *buffers_flushed = 0;
6802 :
6803 32770 : for (int buf = 1; buf <= NBuffers; buf++)
6804 : {
6805 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
6806 : uint32 buf_state;
6807 : bool buffer_flushed;
6808 :
6809 32768 : CHECK_FOR_INTERRUPTS();
6810 :
6811 32768 : buf_state = pg_atomic_read_u32(&desc->state);
6812 32768 : if (!(buf_state & BM_VALID))
6813 28762 : continue;
6814 :
6815 4006 : ResourceOwnerEnlarge(CurrentResourceOwner);
6816 4006 : ReservePrivateRefCountEntry();
6817 :
6818 4006 : LockBufHdr(desc);
6819 :
6820 4006 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
6821 4006 : (*buffers_evicted)++;
6822 : else
6823 0 : (*buffers_skipped)++;
6824 :
6825 4006 : if (buffer_flushed)
6826 1908 : (*buffers_flushed)++;
6827 : }
6828 2 : }
6829 :
6830 : /*
6831 : * Try to evict all the shared buffers containing provided relation's pages.
6832 : *
6833 : * This function is intended for testing/development use only! See
6834 : * EvictUnpinnedBuffer().
6835 : *
6836 : * The caller must hold at least AccessShareLock on the relation to prevent
6837 : * the relation from being dropped.
6838 : *
6839 : * The buffers_* parameters are mandatory and indicate the total count of
6840 : * buffers that:
6841 : * - buffers_evicted - were evicted
6842 : * - buffers_flushed - were flushed
6843 : * - buffers_skipped - could not be evicted
6844 : */
6845 : void
6846 2 : EvictRelUnpinnedBuffers(Relation rel, int32 *buffers_evicted,
6847 : int32 *buffers_flushed, int32 *buffers_skipped)
6848 : {
6849 : Assert(!RelationUsesLocalBuffers(rel));
6850 :
6851 2 : *buffers_skipped = 0;
6852 2 : *buffers_evicted = 0;
6853 2 : *buffers_flushed = 0;
6854 :
6855 32770 : for (int buf = 1; buf <= NBuffers; buf++)
6856 : {
6857 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
6858 32768 : uint32 buf_state = pg_atomic_read_u32(&(desc->state));
6859 : bool buffer_flushed;
6860 :
6861 32768 : CHECK_FOR_INTERRUPTS();
6862 :
6863 : /* An unlocked precheck should be safe and saves some cycles. */
6864 32768 : if ((buf_state & BM_VALID) == 0 ||
6865 54 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
6866 32768 : continue;
6867 :
6868 : /* Make sure we can pin the buffer. */
6869 0 : ResourceOwnerEnlarge(CurrentResourceOwner);
6870 0 : ReservePrivateRefCountEntry();
6871 :
6872 0 : buf_state = LockBufHdr(desc);
6873 :
6874 : /* recheck, could have changed without the lock */
6875 0 : if ((buf_state & BM_VALID) == 0 ||
6876 0 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
6877 : {
6878 0 : UnlockBufHdr(desc);
6879 0 : continue;
6880 : }
6881 :
6882 0 : if (EvictUnpinnedBufferInternal(desc, &buffer_flushed))
6883 0 : (*buffers_evicted)++;
6884 : else
6885 0 : (*buffers_skipped)++;
6886 :
6887 0 : if (buffer_flushed)
6888 0 : (*buffers_flushed)++;
6889 : }
6890 2 : }
6891 :
6892 : /*
6893 : * Helper function to mark unpinned buffer dirty whose buffer header lock is
6894 : * already acquired.
6895 : */
6896 : static bool
6897 72 : MarkDirtyUnpinnedBufferInternal(Buffer buf, BufferDesc *desc,
6898 : bool *buffer_already_dirty)
6899 : {
6900 : uint32 buf_state;
6901 72 : bool result = false;
6902 :
6903 72 : *buffer_already_dirty = false;
6904 :
6905 72 : buf_state = pg_atomic_read_u32(&(desc->state));
6906 : Assert(buf_state & BM_LOCKED);
6907 :
6908 72 : if ((buf_state & BM_VALID) == 0)
6909 : {
6910 2 : UnlockBufHdr(desc);
6911 2 : return false;
6912 : }
6913 :
6914 : /* Check that it's not pinned already. */
6915 70 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
6916 : {
6917 0 : UnlockBufHdr(desc);
6918 0 : return false;
6919 : }
6920 :
6921 : /* Pin the buffer and then release the buffer spinlock */
6922 70 : PinBuffer_Locked(desc);
6923 :
6924 : /* If it was not already dirty, mark it as dirty. */
6925 70 : if (!(buf_state & BM_DIRTY))
6926 : {
6927 34 : LWLockAcquire(BufferDescriptorGetContentLock(desc), LW_EXCLUSIVE);
6928 34 : MarkBufferDirty(buf);
6929 34 : result = true;
6930 34 : LWLockRelease(BufferDescriptorGetContentLock(desc));
6931 : }
6932 : else
6933 36 : *buffer_already_dirty = true;
6934 :
6935 70 : UnpinBuffer(desc);
6936 :
6937 70 : return result;
6938 : }
6939 :
6940 : /*
6941 : * Try to mark the provided shared buffer as dirty.
6942 : *
6943 : * This function is intended for testing/development use only!
6944 : *
6945 : * Same as EvictUnpinnedBuffer() but with MarkBufferDirty() call inside.
6946 : *
6947 : * The buffer_already_dirty parameter is mandatory and indicate if the buffer
6948 : * could not be dirtied because it is already dirty.
6949 : *
6950 : * Returns true if the buffer has successfully been marked as dirty.
6951 : */
6952 : bool
6953 2 : MarkDirtyUnpinnedBuffer(Buffer buf, bool *buffer_already_dirty)
6954 : {
6955 : BufferDesc *desc;
6956 2 : bool buffer_dirtied = false;
6957 :
6958 : Assert(!BufferIsLocal(buf));
6959 :
6960 : /* Make sure we can pin the buffer. */
6961 2 : ResourceOwnerEnlarge(CurrentResourceOwner);
6962 2 : ReservePrivateRefCountEntry();
6963 :
6964 2 : desc = GetBufferDescriptor(buf - 1);
6965 2 : LockBufHdr(desc);
6966 :
6967 2 : buffer_dirtied = MarkDirtyUnpinnedBufferInternal(buf, desc, buffer_already_dirty);
6968 : /* Both can not be true at the same time */
6969 : Assert(!(buffer_dirtied && *buffer_already_dirty));
6970 :
6971 2 : return buffer_dirtied;
6972 : }
6973 :
6974 : /*
6975 : * Try to mark all the shared buffers containing provided relation's pages as
6976 : * dirty.
6977 : *
6978 : * This function is intended for testing/development use only! See
6979 : * MarkDirtyUnpinnedBuffer().
6980 : *
6981 : * The buffers_* parameters are mandatory and indicate the total count of
6982 : * buffers that:
6983 : * - buffers_dirtied - were dirtied
6984 : * - buffers_already_dirty - were already dirty
6985 : * - buffers_skipped - could not be dirtied because of a reason different
6986 : * than a buffer being already dirty.
6987 : */
6988 : void
6989 2 : MarkDirtyRelUnpinnedBuffers(Relation rel,
6990 : int32 *buffers_dirtied,
6991 : int32 *buffers_already_dirty,
6992 : int32 *buffers_skipped)
6993 : {
6994 : Assert(!RelationUsesLocalBuffers(rel));
6995 :
6996 2 : *buffers_dirtied = 0;
6997 2 : *buffers_already_dirty = 0;
6998 2 : *buffers_skipped = 0;
6999 :
7000 32770 : for (int buf = 1; buf <= NBuffers; buf++)
7001 : {
7002 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
7003 32768 : uint32 buf_state = pg_atomic_read_u32(&(desc->state));
7004 : bool buffer_already_dirty;
7005 :
7006 32768 : CHECK_FOR_INTERRUPTS();
7007 :
7008 : /* An unlocked precheck should be safe and saves some cycles. */
7009 32768 : if ((buf_state & BM_VALID) == 0 ||
7010 54 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
7011 32768 : continue;
7012 :
7013 : /* Make sure we can pin the buffer. */
7014 0 : ResourceOwnerEnlarge(CurrentResourceOwner);
7015 0 : ReservePrivateRefCountEntry();
7016 :
7017 0 : buf_state = LockBufHdr(desc);
7018 :
7019 : /* recheck, could have changed without the lock */
7020 0 : if ((buf_state & BM_VALID) == 0 ||
7021 0 : !BufTagMatchesRelFileLocator(&desc->tag, &rel->rd_locator))
7022 : {
7023 0 : UnlockBufHdr(desc);
7024 0 : continue;
7025 : }
7026 :
7027 0 : if (MarkDirtyUnpinnedBufferInternal(buf, desc, &buffer_already_dirty))
7028 0 : (*buffers_dirtied)++;
7029 0 : else if (buffer_already_dirty)
7030 0 : (*buffers_already_dirty)++;
7031 : else
7032 0 : (*buffers_skipped)++;
7033 : }
7034 2 : }
7035 :
7036 : /*
7037 : * Try to mark all the shared buffers as dirty.
7038 : *
7039 : * This function is intended for testing/development use only! See
7040 : * MarkDirtyUnpinnedBuffer().
7041 : *
7042 : * See MarkDirtyRelUnpinnedBuffers() above for details about the buffers_*
7043 : * parameters.
7044 : */
7045 : void
7046 2 : MarkDirtyAllUnpinnedBuffers(int32 *buffers_dirtied,
7047 : int32 *buffers_already_dirty,
7048 : int32 *buffers_skipped)
7049 : {
7050 2 : *buffers_dirtied = 0;
7051 2 : *buffers_already_dirty = 0;
7052 2 : *buffers_skipped = 0;
7053 :
7054 32770 : for (int buf = 1; buf <= NBuffers; buf++)
7055 : {
7056 32768 : BufferDesc *desc = GetBufferDescriptor(buf - 1);
7057 : uint32 buf_state;
7058 : bool buffer_already_dirty;
7059 :
7060 32768 : CHECK_FOR_INTERRUPTS();
7061 :
7062 32768 : buf_state = pg_atomic_read_u32(&desc->state);
7063 32768 : if (!(buf_state & BM_VALID))
7064 32698 : continue;
7065 :
7066 70 : ResourceOwnerEnlarge(CurrentResourceOwner);
7067 70 : ReservePrivateRefCountEntry();
7068 :
7069 70 : LockBufHdr(desc);
7070 :
7071 70 : if (MarkDirtyUnpinnedBufferInternal(buf, desc, &buffer_already_dirty))
7072 34 : (*buffers_dirtied)++;
7073 36 : else if (buffer_already_dirty)
7074 36 : (*buffers_already_dirty)++;
7075 : else
7076 0 : (*buffers_skipped)++;
7077 : }
7078 2 : }
7079 :
7080 : /*
7081 : * Generic implementation of the AIO handle staging callback for readv/writev
7082 : * on local/shared buffers.
7083 : *
7084 : * Each readv/writev can target multiple buffers. The buffers have already
7085 : * been registered with the IO handle.
7086 : *
7087 : * To make the IO ready for execution ("staging"), we need to ensure that the
7088 : * targeted buffers are in an appropriate state while the IO is ongoing. For
7089 : * that the AIO subsystem needs to have its own buffer pin, otherwise an error
7090 : * in this backend could lead to this backend's buffer pin being released as
7091 : * part of error handling, which in turn could lead to the buffer being
7092 : * replaced while IO is ongoing.
7093 : */
7094 : static pg_attribute_always_inline void
7095 2563716 : buffer_stage_common(PgAioHandle *ioh, bool is_write, bool is_temp)
7096 : {
7097 : uint64 *io_data;
7098 : uint8 handle_data_len;
7099 : PgAioWaitRef io_ref;
7100 2563716 : BufferTag first PG_USED_FOR_ASSERTS_ONLY = {0};
7101 :
7102 2563716 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
7103 :
7104 2563716 : pgaio_io_get_wref(ioh, &io_ref);
7105 :
7106 : /* iterate over all buffers affected by the vectored readv/writev */
7107 5467612 : for (int i = 0; i < handle_data_len; i++)
7108 : {
7109 2903896 : Buffer buffer = (Buffer) io_data[i];
7110 2903896 : BufferDesc *buf_hdr = is_temp ?
7111 16818 : GetLocalBufferDescriptor(-buffer - 1)
7112 2903896 : : GetBufferDescriptor(buffer - 1);
7113 : uint32 buf_state;
7114 :
7115 : /*
7116 : * Check that all the buffers are actually ones that could conceivably
7117 : * be done in one IO, i.e. are sequential. This is the last
7118 : * buffer-aware code before IO is actually executed and confusion
7119 : * about which buffers are targeted by IO can be hard to debug, making
7120 : * it worth doing extra-paranoid checks.
7121 : */
7122 2903896 : if (i == 0)
7123 2563716 : first = buf_hdr->tag;
7124 : else
7125 : {
7126 : Assert(buf_hdr->tag.relNumber == first.relNumber);
7127 : Assert(buf_hdr->tag.blockNum == first.blockNum + i);
7128 : }
7129 :
7130 2903896 : if (is_temp)
7131 16818 : buf_state = pg_atomic_read_u32(&buf_hdr->state);
7132 : else
7133 2887078 : buf_state = LockBufHdr(buf_hdr);
7134 :
7135 : /* verify the buffer is in the expected state */
7136 : Assert(buf_state & BM_TAG_VALID);
7137 : if (is_write)
7138 : {
7139 : Assert(buf_state & BM_VALID);
7140 : Assert(buf_state & BM_DIRTY);
7141 : }
7142 : else
7143 : {
7144 : Assert(!(buf_state & BM_VALID));
7145 : Assert(!(buf_state & BM_DIRTY));
7146 : }
7147 :
7148 : /* temp buffers don't use BM_IO_IN_PROGRESS */
7149 2903896 : if (!is_temp)
7150 : Assert(buf_state & BM_IO_IN_PROGRESS);
7151 :
7152 : Assert(BUF_STATE_GET_REFCOUNT(buf_state) >= 1);
7153 :
7154 : /*
7155 : * Reflect that the buffer is now owned by the AIO subsystem.
7156 : *
7157 : * For local buffers: This can't be done just via LocalRefCount, as
7158 : * one might initially think, as this backend could error out while
7159 : * AIO is still in progress, releasing all the pins by the backend
7160 : * itself.
7161 : *
7162 : * This pin is released again in TerminateBufferIO().
7163 : */
7164 2903896 : buf_hdr->io_wref = io_ref;
7165 :
7166 2903896 : if (is_temp)
7167 : {
7168 16818 : buf_state += BUF_REFCOUNT_ONE;
7169 16818 : pg_atomic_unlocked_write_u32(&buf_hdr->state, buf_state);
7170 : }
7171 : else
7172 2887078 : UnlockBufHdrExt(buf_hdr, buf_state, 0, 0, 1);
7173 :
7174 : /*
7175 : * Ensure the content lock that prevents buffer modifications while
7176 : * the buffer is being written out is not released early due to an
7177 : * error.
7178 : */
7179 2903896 : if (is_write && !is_temp)
7180 : {
7181 : LWLock *content_lock;
7182 :
7183 0 : content_lock = BufferDescriptorGetContentLock(buf_hdr);
7184 :
7185 : Assert(LWLockHeldByMe(content_lock));
7186 :
7187 : /*
7188 : * Lock is now owned by AIO subsystem.
7189 : */
7190 0 : LWLockDisown(content_lock);
7191 : }
7192 :
7193 : /*
7194 : * Stop tracking this buffer via the resowner - the AIO system now
7195 : * keeps track.
7196 : */
7197 2903896 : if (!is_temp)
7198 2887078 : ResourceOwnerForgetBufferIO(CurrentResourceOwner, buffer);
7199 : }
7200 2563716 : }
7201 :
7202 : /*
7203 : * Decode readv errors as encoded by buffer_readv_encode_error().
7204 : */
7205 : static inline void
7206 698 : buffer_readv_decode_error(PgAioResult result,
7207 : bool *zeroed_any,
7208 : bool *ignored_any,
7209 : uint8 *zeroed_or_error_count,
7210 : uint8 *checkfail_count,
7211 : uint8 *first_off)
7212 : {
7213 698 : uint32 rem_error = result.error_data;
7214 :
7215 : /* see static asserts in buffer_readv_encode_error */
7216 : #define READV_COUNT_BITS 7
7217 : #define READV_COUNT_MASK ((1 << READV_COUNT_BITS) - 1)
7218 :
7219 698 : *zeroed_any = rem_error & 1;
7220 698 : rem_error >>= 1;
7221 :
7222 698 : *ignored_any = rem_error & 1;
7223 698 : rem_error >>= 1;
7224 :
7225 698 : *zeroed_or_error_count = rem_error & READV_COUNT_MASK;
7226 698 : rem_error >>= READV_COUNT_BITS;
7227 :
7228 698 : *checkfail_count = rem_error & READV_COUNT_MASK;
7229 698 : rem_error >>= READV_COUNT_BITS;
7230 :
7231 698 : *first_off = rem_error & READV_COUNT_MASK;
7232 698 : rem_error >>= READV_COUNT_BITS;
7233 698 : }
7234 :
7235 : /*
7236 : * Helper to encode errors for buffer_readv_complete()
7237 : *
7238 : * Errors are encoded as follows:
7239 : * - bit 0 indicates whether any page was zeroed (1) or not (0)
7240 : * - bit 1 indicates whether any checksum failure was ignored (1) or not (0)
7241 : * - next READV_COUNT_BITS bits indicate the number of errored or zeroed pages
7242 : * - next READV_COUNT_BITS bits indicate the number of checksum failures
7243 : * - next READV_COUNT_BITS bits indicate the first offset of the first page
7244 : * that was errored or zeroed or, if no errors/zeroes, the first ignored
7245 : * checksum
7246 : */
7247 : static inline void
7248 384 : buffer_readv_encode_error(PgAioResult *result,
7249 : bool is_temp,
7250 : bool zeroed_any,
7251 : bool ignored_any,
7252 : uint8 error_count,
7253 : uint8 zeroed_count,
7254 : uint8 checkfail_count,
7255 : uint8 first_error_off,
7256 : uint8 first_zeroed_off,
7257 : uint8 first_ignored_off)
7258 : {
7259 :
7260 384 : uint8 shift = 0;
7261 384 : uint8 zeroed_or_error_count =
7262 : error_count > 0 ? error_count : zeroed_count;
7263 : uint8 first_off;
7264 :
7265 : StaticAssertDecl(PG_IOV_MAX <= 1 << READV_COUNT_BITS,
7266 : "PG_IOV_MAX is bigger than reserved space for error data");
7267 : StaticAssertDecl((1 + 1 + 3 * READV_COUNT_BITS) <= PGAIO_RESULT_ERROR_BITS,
7268 : "PGAIO_RESULT_ERROR_BITS is insufficient for buffer_readv");
7269 :
7270 : /*
7271 : * We only have space to encode one offset - but luckily that's good
7272 : * enough. If there is an error, the error is the interesting offset, same
7273 : * with a zeroed buffer vs an ignored buffer.
7274 : */
7275 384 : if (error_count > 0)
7276 188 : first_off = first_error_off;
7277 196 : else if (zeroed_count > 0)
7278 160 : first_off = first_zeroed_off;
7279 : else
7280 36 : first_off = first_ignored_off;
7281 :
7282 : Assert(!zeroed_any || error_count == 0);
7283 :
7284 384 : result->error_data = 0;
7285 :
7286 384 : result->error_data |= zeroed_any << shift;
7287 384 : shift += 1;
7288 :
7289 384 : result->error_data |= ignored_any << shift;
7290 384 : shift += 1;
7291 :
7292 384 : result->error_data |= ((uint32) zeroed_or_error_count) << shift;
7293 384 : shift += READV_COUNT_BITS;
7294 :
7295 384 : result->error_data |= ((uint32) checkfail_count) << shift;
7296 384 : shift += READV_COUNT_BITS;
7297 :
7298 384 : result->error_data |= ((uint32) first_off) << shift;
7299 384 : shift += READV_COUNT_BITS;
7300 :
7301 384 : result->id = is_temp ? PGAIO_HCB_LOCAL_BUFFER_READV :
7302 : PGAIO_HCB_SHARED_BUFFER_READV;
7303 :
7304 384 : if (error_count > 0)
7305 188 : result->status = PGAIO_RS_ERROR;
7306 : else
7307 196 : result->status = PGAIO_RS_WARNING;
7308 :
7309 : /*
7310 : * The encoding is complicated enough to warrant cross-checking it against
7311 : * the decode function.
7312 : */
7313 : #ifdef USE_ASSERT_CHECKING
7314 : {
7315 : bool zeroed_any_2,
7316 : ignored_any_2;
7317 : uint8 zeroed_or_error_count_2,
7318 : checkfail_count_2,
7319 : first_off_2;
7320 :
7321 : buffer_readv_decode_error(*result,
7322 : &zeroed_any_2, &ignored_any_2,
7323 : &zeroed_or_error_count_2,
7324 : &checkfail_count_2,
7325 : &first_off_2);
7326 : Assert(zeroed_any == zeroed_any_2);
7327 : Assert(ignored_any == ignored_any_2);
7328 : Assert(zeroed_or_error_count == zeroed_or_error_count_2);
7329 : Assert(checkfail_count == checkfail_count_2);
7330 : Assert(first_off == first_off_2);
7331 : }
7332 : #endif
7333 :
7334 : #undef READV_COUNT_BITS
7335 : #undef READV_COUNT_MASK
7336 384 : }
7337 :
7338 : /*
7339 : * Helper for AIO readv completion callbacks, supporting both shared and temp
7340 : * buffers. Gets called once for each buffer in a multi-page read.
7341 : */
7342 : static pg_attribute_always_inline void
7343 2626418 : buffer_readv_complete_one(PgAioTargetData *td, uint8 buf_off, Buffer buffer,
7344 : uint8 flags, bool failed, bool is_temp,
7345 : bool *buffer_invalid,
7346 : bool *failed_checksum,
7347 : bool *ignored_checksum,
7348 : bool *zeroed_buffer)
7349 : {
7350 2626418 : BufferDesc *buf_hdr = is_temp ?
7351 16818 : GetLocalBufferDescriptor(-buffer - 1)
7352 2626418 : : GetBufferDescriptor(buffer - 1);
7353 2626418 : BufferTag tag = buf_hdr->tag;
7354 2626418 : char *bufdata = BufferGetBlock(buffer);
7355 : uint32 set_flag_bits;
7356 : int piv_flags;
7357 :
7358 : /* check that the buffer is in the expected state for a read */
7359 : #ifdef USE_ASSERT_CHECKING
7360 : {
7361 : uint32 buf_state = pg_atomic_read_u32(&buf_hdr->state);
7362 :
7363 : Assert(buf_state & BM_TAG_VALID);
7364 : Assert(!(buf_state & BM_VALID));
7365 : /* temp buffers don't use BM_IO_IN_PROGRESS */
7366 : if (!is_temp)
7367 : Assert(buf_state & BM_IO_IN_PROGRESS);
7368 : Assert(!(buf_state & BM_DIRTY));
7369 : }
7370 : #endif
7371 :
7372 2626418 : *buffer_invalid = false;
7373 2626418 : *failed_checksum = false;
7374 2626418 : *ignored_checksum = false;
7375 2626418 : *zeroed_buffer = false;
7376 :
7377 : /*
7378 : * We ask PageIsVerified() to only log the message about checksum errors,
7379 : * as the completion might be run in any backend (or IO workers). We will
7380 : * report checksum errors in buffer_readv_report().
7381 : */
7382 2626418 : piv_flags = PIV_LOG_LOG;
7383 :
7384 : /* the local zero_damaged_pages may differ from the definer's */
7385 2626418 : if (flags & READ_BUFFERS_IGNORE_CHECKSUM_FAILURES)
7386 76 : piv_flags |= PIV_IGNORE_CHECKSUM_FAILURE;
7387 :
7388 : /* Check for garbage data. */
7389 2626418 : if (!failed)
7390 : {
7391 : /*
7392 : * If the buffer is not currently pinned by this backend, e.g. because
7393 : * we're completing this IO after an error, the buffer data will have
7394 : * been marked as inaccessible when the buffer was unpinned. The AIO
7395 : * subsystem holds a pin, but that doesn't prevent the buffer from
7396 : * having been marked as inaccessible. The completion might also be
7397 : * executed in a different process.
7398 : */
7399 : #ifdef USE_VALGRIND
7400 : if (!BufferIsPinned(buffer))
7401 : VALGRIND_MAKE_MEM_DEFINED(bufdata, BLCKSZ);
7402 : #endif
7403 :
7404 2626360 : if (!PageIsVerified((Page) bufdata, tag.blockNum, piv_flags,
7405 : failed_checksum))
7406 : {
7407 192 : if (flags & READ_BUFFERS_ZERO_ON_ERROR)
7408 : {
7409 92 : memset(bufdata, 0, BLCKSZ);
7410 92 : *zeroed_buffer = true;
7411 : }
7412 : else
7413 : {
7414 100 : *buffer_invalid = true;
7415 : /* mark buffer as having failed */
7416 100 : failed = true;
7417 : }
7418 : }
7419 2626168 : else if (*failed_checksum)
7420 24 : *ignored_checksum = true;
7421 :
7422 : /* undo what we did above */
7423 : #ifdef USE_VALGRIND
7424 : if (!BufferIsPinned(buffer))
7425 : VALGRIND_MAKE_MEM_NOACCESS(bufdata, BLCKSZ);
7426 : #endif
7427 :
7428 : /*
7429 : * Immediately log a message about the invalid page, but only to the
7430 : * server log. The reason to do so immediately is that this may be
7431 : * executed in a different backend than the one that originated the
7432 : * request. The reason to do so immediately is that the originator
7433 : * might not process the query result immediately (because it is busy
7434 : * doing another part of query processing) or at all (e.g. if it was
7435 : * cancelled or errored out due to another IO also failing). The
7436 : * definer of the IO will emit an ERROR or WARNING when processing the
7437 : * IO's results
7438 : *
7439 : * To avoid duplicating the code to emit these log messages, we reuse
7440 : * buffer_readv_report().
7441 : */
7442 2626360 : if (*buffer_invalid || *failed_checksum || *zeroed_buffer)
7443 : {
7444 216 : PgAioResult result_one = {0};
7445 :
7446 216 : buffer_readv_encode_error(&result_one, is_temp,
7447 216 : *zeroed_buffer,
7448 216 : *ignored_checksum,
7449 216 : *buffer_invalid,
7450 216 : *zeroed_buffer ? 1 : 0,
7451 216 : *failed_checksum ? 1 : 0,
7452 : buf_off, buf_off, buf_off);
7453 216 : pgaio_result_report(result_one, td, LOG_SERVER_ONLY);
7454 : }
7455 : }
7456 :
7457 : /* Terminate I/O and set BM_VALID. */
7458 2626418 : set_flag_bits = failed ? BM_IO_ERROR : BM_VALID;
7459 2626418 : if (is_temp)
7460 16818 : TerminateLocalBufferIO(buf_hdr, false, set_flag_bits, true);
7461 : else
7462 2609600 : TerminateBufferIO(buf_hdr, false, set_flag_bits, false, true);
7463 :
7464 : /*
7465 : * Call the BUFFER_READ_DONE tracepoint in the callback, even though the
7466 : * callback may not be executed in the same backend that called
7467 : * BUFFER_READ_START. The alternative would be to defer calling the
7468 : * tracepoint to a later point (e.g. the local completion callback for
7469 : * shared buffer reads), which seems even less helpful.
7470 : */
7471 : TRACE_POSTGRESQL_BUFFER_READ_DONE(tag.forkNum,
7472 : tag.blockNum,
7473 : tag.spcOid,
7474 : tag.dbOid,
7475 : tag.relNumber,
7476 : is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
7477 : false);
7478 2626418 : }
7479 :
7480 : /*
7481 : * Perform completion handling of a single AIO read. This read may cover
7482 : * multiple blocks / buffers.
7483 : *
7484 : * Shared between shared and local buffers, to reduce code duplication.
7485 : */
7486 : static pg_attribute_always_inline PgAioResult
7487 2353370 : buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7488 : uint8 cb_data, bool is_temp)
7489 : {
7490 2353370 : PgAioResult result = prior_result;
7491 2353370 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
7492 2353370 : uint8 first_error_off = 0;
7493 2353370 : uint8 first_zeroed_off = 0;
7494 2353370 : uint8 first_ignored_off = 0;
7495 2353370 : uint8 error_count = 0;
7496 2353370 : uint8 zeroed_count = 0;
7497 2353370 : uint8 ignored_count = 0;
7498 2353370 : uint8 checkfail_count = 0;
7499 : uint64 *io_data;
7500 : uint8 handle_data_len;
7501 :
7502 : if (is_temp)
7503 : {
7504 : Assert(td->smgr.is_temp);
7505 : Assert(pgaio_io_get_owner(ioh) == MyProcNumber);
7506 : }
7507 : else
7508 : Assert(!td->smgr.is_temp);
7509 :
7510 : /*
7511 : * Iterate over all the buffers affected by this IO and call the
7512 : * per-buffer completion function for each buffer.
7513 : */
7514 2353370 : io_data = pgaio_io_get_handle_data(ioh, &handle_data_len);
7515 4979788 : for (uint8 buf_off = 0; buf_off < handle_data_len; buf_off++)
7516 : {
7517 2626418 : Buffer buf = io_data[buf_off];
7518 : bool failed;
7519 2626418 : bool failed_verification = false;
7520 2626418 : bool failed_checksum = false;
7521 2626418 : bool zeroed_buffer = false;
7522 2626418 : bool ignored_checksum = false;
7523 :
7524 : Assert(BufferIsValid(buf));
7525 :
7526 : /*
7527 : * If the entire I/O failed on a lower-level, each buffer needs to be
7528 : * marked as failed. In case of a partial read, the first few buffers
7529 : * may be ok.
7530 : */
7531 2626418 : failed =
7532 2626418 : prior_result.status == PGAIO_RS_ERROR
7533 2626418 : || prior_result.result <= buf_off;
7534 :
7535 2626418 : buffer_readv_complete_one(td, buf_off, buf, cb_data, failed, is_temp,
7536 : &failed_verification,
7537 : &failed_checksum,
7538 : &ignored_checksum,
7539 : &zeroed_buffer);
7540 :
7541 : /*
7542 : * Track information about the number of different kinds of error
7543 : * conditions across all pages, as there can be multiple pages failing
7544 : * verification as part of one IO.
7545 : */
7546 2626418 : if (failed_verification && !zeroed_buffer && error_count++ == 0)
7547 88 : first_error_off = buf_off;
7548 2626418 : if (zeroed_buffer && zeroed_count++ == 0)
7549 68 : first_zeroed_off = buf_off;
7550 2626418 : if (ignored_checksum && ignored_count++ == 0)
7551 20 : first_ignored_off = buf_off;
7552 2626418 : if (failed_checksum)
7553 64 : checkfail_count++;
7554 : }
7555 :
7556 : /*
7557 : * If the smgr read succeeded [partially] and page verification failed for
7558 : * some of the pages, adjust the IO's result state appropriately.
7559 : */
7560 2353370 : if (prior_result.status != PGAIO_RS_ERROR &&
7561 2353264 : (error_count > 0 || ignored_count > 0 || zeroed_count > 0))
7562 : {
7563 168 : buffer_readv_encode_error(&result, is_temp,
7564 : zeroed_count > 0, ignored_count > 0,
7565 : error_count, zeroed_count, checkfail_count,
7566 : first_error_off, first_zeroed_off,
7567 : first_ignored_off);
7568 168 : pgaio_result_report(result, td, DEBUG1);
7569 : }
7570 :
7571 : /*
7572 : * For shared relations this reporting is done in
7573 : * shared_buffer_readv_complete_local().
7574 : */
7575 2353370 : if (is_temp && checkfail_count > 0)
7576 4 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
7577 : checkfail_count);
7578 :
7579 2353370 : return result;
7580 : }
7581 :
7582 : /*
7583 : * AIO error reporting callback for aio_shared_buffer_readv_cb and
7584 : * aio_local_buffer_readv_cb.
7585 : *
7586 : * The error is encoded / decoded in buffer_readv_encode_error() /
7587 : * buffer_readv_decode_error().
7588 : */
7589 : static void
7590 544 : buffer_readv_report(PgAioResult result, const PgAioTargetData *td,
7591 : int elevel)
7592 : {
7593 544 : int nblocks = td->smgr.nblocks;
7594 544 : BlockNumber first = td->smgr.blockNum;
7595 544 : BlockNumber last = first + nblocks - 1;
7596 544 : ProcNumber errProc =
7597 544 : td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER;
7598 : RelPathStr rpath =
7599 544 : relpathbackend(td->smgr.rlocator, errProc, td->smgr.forkNum);
7600 : bool zeroed_any,
7601 : ignored_any;
7602 : uint8 zeroed_or_error_count,
7603 : checkfail_count,
7604 : first_off;
7605 : uint8 affected_count;
7606 : const char *msg_one,
7607 : *msg_mult,
7608 : *det_mult,
7609 : *hint_mult;
7610 :
7611 544 : buffer_readv_decode_error(result, &zeroed_any, &ignored_any,
7612 : &zeroed_or_error_count,
7613 : &checkfail_count,
7614 : &first_off);
7615 :
7616 : /*
7617 : * Treat a read that had both zeroed buffers *and* ignored checksums as a
7618 : * special case, it's too irregular to be emitted the same way as the
7619 : * other cases.
7620 : */
7621 544 : if (zeroed_any && ignored_any)
7622 : {
7623 : Assert(zeroed_any && ignored_any);
7624 : Assert(nblocks > 1); /* same block can't be both zeroed and ignored */
7625 : Assert(result.status != PGAIO_RS_ERROR);
7626 8 : affected_count = zeroed_or_error_count;
7627 :
7628 8 : ereport(elevel,
7629 : errcode(ERRCODE_DATA_CORRUPTED),
7630 : errmsg("zeroing %u page(s) and ignoring %u checksum failure(s) among blocks %u..%u of relation \"%s\"",
7631 : affected_count, checkfail_count, first, last, rpath.str),
7632 : affected_count > 1 ?
7633 : errdetail("Block %u held the first zeroed page.",
7634 : first + first_off) : 0,
7635 : errhint_plural("See server log for details about the other %d invalid block.",
7636 : "See server log for details about the other %d invalid blocks.",
7637 : affected_count + checkfail_count - 1,
7638 : affected_count + checkfail_count - 1));
7639 8 : return;
7640 : }
7641 :
7642 : /*
7643 : * The other messages are highly repetitive. To avoid duplicating a long
7644 : * and complicated ereport(), gather the translated format strings
7645 : * separately and then do one common ereport.
7646 : */
7647 536 : if (result.status == PGAIO_RS_ERROR)
7648 : {
7649 : Assert(!zeroed_any); /* can't have invalid pages when zeroing them */
7650 272 : affected_count = zeroed_or_error_count;
7651 272 : msg_one = _("invalid page in block %u of relation \"%s\"");
7652 272 : msg_mult = _("%u invalid pages among blocks %u..%u of relation \"%s\"");
7653 272 : det_mult = _("Block %u held the first invalid page.");
7654 272 : hint_mult = _("See server log for the other %u invalid block(s).");
7655 : }
7656 264 : else if (zeroed_any && !ignored_any)
7657 : {
7658 216 : affected_count = zeroed_or_error_count;
7659 216 : msg_one = _("invalid page in block %u of relation \"%s\"; zeroing out page");
7660 216 : msg_mult = _("zeroing out %u invalid pages among blocks %u..%u of relation \"%s\"");
7661 216 : det_mult = _("Block %u held the first zeroed page.");
7662 216 : hint_mult = _("See server log for the other %u zeroed block(s).");
7663 : }
7664 48 : else if (!zeroed_any && ignored_any)
7665 : {
7666 48 : affected_count = checkfail_count;
7667 48 : msg_one = _("ignoring checksum failure in block %u of relation \"%s\"");
7668 48 : msg_mult = _("ignoring %u checksum failures among blocks %u..%u of relation \"%s\"");
7669 48 : det_mult = _("Block %u held the first ignored page.");
7670 48 : hint_mult = _("See server log for the other %u ignored block(s).");
7671 : }
7672 : else
7673 0 : pg_unreachable();
7674 :
7675 536 : ereport(elevel,
7676 : errcode(ERRCODE_DATA_CORRUPTED),
7677 : affected_count == 1 ?
7678 : errmsg_internal(msg_one, first + first_off, rpath.str) :
7679 : errmsg_internal(msg_mult, affected_count, first, last, rpath.str),
7680 : affected_count > 1 ? errdetail_internal(det_mult, first + first_off) : 0,
7681 : affected_count > 1 ? errhint_internal(hint_mult, affected_count - 1) : 0);
7682 : }
7683 :
7684 : static void
7685 2560104 : shared_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
7686 : {
7687 2560104 : buffer_stage_common(ioh, false, false);
7688 2560104 : }
7689 :
7690 : static PgAioResult
7691 2349758 : shared_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7692 : uint8 cb_data)
7693 : {
7694 2349758 : return buffer_readv_complete(ioh, prior_result, cb_data, false);
7695 : }
7696 :
7697 : /*
7698 : * We need a backend-local completion callback for shared buffers, to be able
7699 : * to report checksum errors correctly. Unfortunately that can only safely
7700 : * happen if the reporting backend has previously called
7701 : * pgstat_prepare_report_checksum_failure(), which we can only guarantee in
7702 : * the backend that started the IO. Hence this callback.
7703 : */
7704 : static PgAioResult
7705 2560104 : shared_buffer_readv_complete_local(PgAioHandle *ioh, PgAioResult prior_result,
7706 : uint8 cb_data)
7707 : {
7708 : bool zeroed_any,
7709 : ignored_any;
7710 : uint8 zeroed_or_error_count,
7711 : checkfail_count,
7712 : first_off;
7713 :
7714 2560104 : if (prior_result.status == PGAIO_RS_OK)
7715 2559950 : return prior_result;
7716 :
7717 154 : buffer_readv_decode_error(prior_result,
7718 : &zeroed_any,
7719 : &ignored_any,
7720 : &zeroed_or_error_count,
7721 : &checkfail_count,
7722 : &first_off);
7723 :
7724 154 : if (checkfail_count)
7725 : {
7726 48 : PgAioTargetData *td = pgaio_io_get_target_data(ioh);
7727 :
7728 48 : pgstat_report_checksum_failures_in_db(td->smgr.rlocator.dbOid,
7729 : checkfail_count);
7730 : }
7731 :
7732 154 : return prior_result;
7733 : }
7734 :
7735 : static void
7736 3612 : local_buffer_readv_stage(PgAioHandle *ioh, uint8 cb_data)
7737 : {
7738 3612 : buffer_stage_common(ioh, false, true);
7739 3612 : }
7740 :
7741 : static PgAioResult
7742 3612 : local_buffer_readv_complete(PgAioHandle *ioh, PgAioResult prior_result,
7743 : uint8 cb_data)
7744 : {
7745 3612 : return buffer_readv_complete(ioh, prior_result, cb_data, true);
7746 : }
7747 :
7748 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
7749 : const PgAioHandleCallbacks aio_shared_buffer_readv_cb = {
7750 : .stage = shared_buffer_readv_stage,
7751 : .complete_shared = shared_buffer_readv_complete,
7752 : /* need a local callback to report checksum failures */
7753 : .complete_local = shared_buffer_readv_complete_local,
7754 : .report = buffer_readv_report,
7755 : };
7756 :
7757 : /* readv callback is passed READ_BUFFERS_* flags as callback data */
7758 : const PgAioHandleCallbacks aio_local_buffer_readv_cb = {
7759 : .stage = local_buffer_readv_stage,
7760 :
7761 : /*
7762 : * Note that this, in contrast to the shared_buffers case, uses
7763 : * complete_local, as only the issuing backend has access to the required
7764 : * datastructures. This is important in case the IO completion may be
7765 : * consumed incidentally by another backend.
7766 : */
7767 : .complete_local = local_buffer_readv_complete,
7768 : .report = buffer_readv_report,
7769 : };
|
