Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * localbuf.c
4 : * local buffer manager. Fast buffer manager for temporary tables,
5 : * which never need to be WAL-logged or checkpointed, etc.
6 : *
7 : * Portions Copyright (c) 1996-2017, PostgreSQL Global Development Group
8 : * Portions Copyright (c) 1994-5, Regents of the University of California
9 : *
10 : *
11 : * IDENTIFICATION
12 : * src/backend/storage/buffer/localbuf.c
13 : *
14 : *-------------------------------------------------------------------------
15 : */
16 : #include "postgres.h"
17 :
18 : #include "access/parallel.h"
19 : #include "catalog/catalog.h"
20 : #include "executor/instrument.h"
21 : #include "storage/buf_internals.h"
22 : #include "storage/bufmgr.h"
23 : #include "utils/guc.h"
24 : #include "utils/memutils.h"
25 : #include "utils/resowner_private.h"
26 :
27 :
28 : /*#define LBDEBUG*/
29 :
30 : /* entry for buffer lookup hashtable */
31 : typedef struct
32 : {
33 : BufferTag key; /* Tag of a disk page */
34 : int id; /* Associated local buffer's index */
35 : } LocalBufferLookupEnt;
36 :
37 : /* Note: this macro only works on local buffers, not shared ones! */
38 : #define LocalBufHdrGetBlock(bufHdr) \
39 : LocalBufferBlockPointers[-((bufHdr)->buf_id + 2)]
40 :
41 : int NLocBuffer = 0; /* until buffers are initialized */
42 :
43 : BufferDesc *LocalBufferDescriptors = NULL;
44 : Block *LocalBufferBlockPointers = NULL;
45 : int32 *LocalRefCount = NULL;
46 :
47 : static int nextFreeLocalBuf = 0;
48 :
49 : static HTAB *LocalBufHash = NULL;
50 :
51 :
52 : static void InitLocalBuffers(void);
53 : static Block GetLocalBufferStorage(void);
54 :
55 :
56 : /*
57 : * LocalPrefetchBuffer -
58 : * initiate asynchronous read of a block of a relation
59 : *
60 : * Do PrefetchBuffer's work for temporary relations.
61 : * No-op if prefetching isn't compiled in.
62 : */
63 : void
64 5 : LocalPrefetchBuffer(SMgrRelation smgr, ForkNumber forkNum,
65 : BlockNumber blockNum)
66 : {
67 : #ifdef USE_PREFETCH
68 : BufferTag newTag; /* identity of requested block */
69 : LocalBufferLookupEnt *hresult;
70 :
71 5 : INIT_BUFFERTAG(newTag, smgr->smgr_rnode.node, forkNum, blockNum);
72 :
73 : /* Initialize local buffers if first request in this session */
74 5 : if (LocalBufHash == NULL)
75 0 : InitLocalBuffers();
76 :
77 : /* See if the desired buffer already exists */
78 5 : hresult = (LocalBufferLookupEnt *)
79 5 : hash_search(LocalBufHash, (void *) &newTag, HASH_FIND, NULL);
80 :
81 5 : if (hresult)
82 : {
83 : /* Yes, so nothing to do */
84 10 : return;
85 : }
86 :
87 : /* Not in buffers, so initiate prefetch */
88 0 : smgrprefetch(smgr, forkNum, blockNum);
89 : #endif /* USE_PREFETCH */
90 : }
91 :
92 :
93 : /*
94 : * LocalBufferAlloc -
95 : * Find or create a local buffer for the given page of the given relation.
96 : *
97 : * API is similar to bufmgr.c's BufferAlloc, except that we do not need
98 : * to do any locking since this is all local. Also, IO_IN_PROGRESS
99 : * does not get set. Lastly, we support only default access strategy
100 : * (hence, usage_count is always advanced).
101 : */
102 : BufferDesc *
103 37373 : LocalBufferAlloc(SMgrRelation smgr, ForkNumber forkNum, BlockNumber blockNum,
104 : bool *foundPtr)
105 : {
106 : BufferTag newTag; /* identity of requested block */
107 : LocalBufferLookupEnt *hresult;
108 : BufferDesc *bufHdr;
109 : int b;
110 : int trycounter;
111 : bool found;
112 : uint32 buf_state;
113 :
114 37373 : INIT_BUFFERTAG(newTag, smgr->smgr_rnode.node, forkNum, blockNum);
115 :
116 : /* Initialize local buffers if first request in this session */
117 37373 : if (LocalBufHash == NULL)
118 47 : InitLocalBuffers();
119 :
120 : /* See if the desired buffer already exists */
121 37373 : hresult = (LocalBufferLookupEnt *)
122 37373 : hash_search(LocalBufHash, (void *) &newTag, HASH_FIND, NULL);
123 :
124 37373 : if (hresult)
125 : {
126 36028 : b = hresult->id;
127 36028 : bufHdr = GetLocalBufferDescriptor(b);
128 36028 : Assert(BUFFERTAGS_EQUAL(bufHdr->tag, newTag));
129 : #ifdef LBDEBUG
130 : fprintf(stderr, "LB ALLOC (%u,%d,%d) %d\n",
131 : smgr->smgr_rnode.node.relNode, forkNum, blockNum, -b - 1);
132 : #endif
133 36028 : buf_state = pg_atomic_read_u32(&bufHdr->state);
134 :
135 : /* this part is equivalent to PinBuffer for a shared buffer */
136 36028 : if (LocalRefCount[b] == 0)
137 : {
138 35106 : if (BUF_STATE_GET_USAGECOUNT(buf_state) < BM_MAX_USAGE_COUNT)
139 : {
140 4518 : buf_state += BUF_USAGECOUNT_ONE;
141 4518 : pg_atomic_unlocked_write_u32(&bufHdr->state, buf_state);
142 : }
143 : }
144 36028 : LocalRefCount[b]++;
145 36028 : ResourceOwnerRememberBuffer(CurrentResourceOwner,
146 36028 : BufferDescriptorGetBuffer(bufHdr));
147 36028 : if (buf_state & BM_VALID)
148 36028 : *foundPtr = TRUE;
149 : else
150 : {
151 : /* Previous read attempt must have failed; try again */
152 0 : *foundPtr = FALSE;
153 : }
154 36028 : return bufHdr;
155 : }
156 :
157 : #ifdef LBDEBUG
158 : fprintf(stderr, "LB ALLOC (%u,%d,%d) %d\n",
159 : smgr->smgr_rnode.node.relNode, forkNum, blockNum,
160 : -nextFreeLocalBuf - 1);
161 : #endif
162 :
163 : /*
164 : * Need to get a new buffer. We use a clock sweep algorithm (essentially
165 : * the same as what freelist.c does now...)
166 : */
167 1345 : trycounter = NLocBuffer;
168 : for (;;)
169 : {
170 1345 : b = nextFreeLocalBuf;
171 :
172 1345 : if (++nextFreeLocalBuf >= NLocBuffer)
173 0 : nextFreeLocalBuf = 0;
174 :
175 1345 : bufHdr = GetLocalBufferDescriptor(b);
176 :
177 1345 : if (LocalRefCount[b] == 0)
178 : {
179 1345 : buf_state = pg_atomic_read_u32(&bufHdr->state);
180 :
181 1345 : if (BUF_STATE_GET_USAGECOUNT(buf_state) > 0)
182 : {
183 0 : buf_state -= BUF_USAGECOUNT_ONE;
184 0 : pg_atomic_unlocked_write_u32(&bufHdr->state, buf_state);
185 0 : trycounter = NLocBuffer;
186 : }
187 : else
188 : {
189 : /* Found a usable buffer */
190 1345 : LocalRefCount[b]++;
191 1345 : ResourceOwnerRememberBuffer(CurrentResourceOwner,
192 1345 : BufferDescriptorGetBuffer(bufHdr));
193 1345 : break;
194 : }
195 : }
196 0 : else if (--trycounter == 0)
197 0 : ereport(ERROR,
198 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
199 : errmsg("no empty local buffer available")));
200 0 : }
201 :
202 : /*
203 : * this buffer is not referenced but it might still be dirty. if that's
204 : * the case, write it out before reusing it!
205 : */
206 1345 : if (buf_state & BM_DIRTY)
207 : {
208 : SMgrRelation oreln;
209 0 : Page localpage = (char *) LocalBufHdrGetBlock(bufHdr);
210 :
211 : /* Find smgr relation for buffer */
212 0 : oreln = smgropen(bufHdr->tag.rnode, MyBackendId);
213 :
214 0 : PageSetChecksumInplace(localpage, bufHdr->tag.blockNum);
215 :
216 : /* And write... */
217 0 : smgrwrite(oreln,
218 : bufHdr->tag.forkNum,
219 : bufHdr->tag.blockNum,
220 : localpage,
221 : false);
222 :
223 : /* Mark not-dirty now in case we error out below */
224 0 : buf_state &= ~BM_DIRTY;
225 0 : pg_atomic_unlocked_write_u32(&bufHdr->state, buf_state);
226 :
227 0 : pgBufferUsage.local_blks_written++;
228 : }
229 :
230 : /*
231 : * lazy memory allocation: allocate space on first use of a buffer.
232 : */
233 1345 : if (LocalBufHdrGetBlock(bufHdr) == NULL)
234 : {
235 : /* Set pointer for use by BufferGetBlock() macro */
236 1345 : LocalBufHdrGetBlock(bufHdr) = GetLocalBufferStorage();
237 : }
238 :
239 : /*
240 : * Update the hash table: remove old entry, if any, and make new one.
241 : */
242 1345 : if (buf_state & BM_TAG_VALID)
243 : {
244 0 : hresult = (LocalBufferLookupEnt *)
245 0 : hash_search(LocalBufHash, (void *) &bufHdr->tag,
246 : HASH_REMOVE, NULL);
247 0 : if (!hresult) /* shouldn't happen */
248 0 : elog(ERROR, "local buffer hash table corrupted");
249 : /* mark buffer invalid just in case hash insert fails */
250 0 : CLEAR_BUFFERTAG(bufHdr->tag);
251 0 : buf_state &= ~(BM_VALID | BM_TAG_VALID);
252 0 : pg_atomic_unlocked_write_u32(&bufHdr->state, buf_state);
253 : }
254 :
255 1345 : hresult = (LocalBufferLookupEnt *)
256 1345 : hash_search(LocalBufHash, (void *) &newTag, HASH_ENTER, &found);
257 1345 : if (found) /* shouldn't happen */
258 0 : elog(ERROR, "local buffer hash table corrupted");
259 1345 : hresult->id = b;
260 :
261 : /*
262 : * it's all ours now.
263 : */
264 1345 : bufHdr->tag = newTag;
265 1345 : buf_state &= ~(BM_VALID | BM_DIRTY | BM_JUST_DIRTIED | BM_IO_ERROR);
266 1345 : buf_state |= BM_TAG_VALID;
267 1345 : buf_state &= ~BUF_USAGECOUNT_MASK;
268 1345 : buf_state += BUF_USAGECOUNT_ONE;
269 1345 : pg_atomic_unlocked_write_u32(&bufHdr->state, buf_state);
270 :
271 1345 : *foundPtr = FALSE;
272 1345 : return bufHdr;
273 : }
274 :
275 : /*
276 : * MarkLocalBufferDirty -
277 : * mark a local buffer dirty
278 : */
279 : void
280 63177 : MarkLocalBufferDirty(Buffer buffer)
281 : {
282 : int bufid;
283 : BufferDesc *bufHdr;
284 : uint32 buf_state;
285 :
286 63177 : Assert(BufferIsLocal(buffer));
287 :
288 : #ifdef LBDEBUG
289 : fprintf(stderr, "LB DIRTY %d\n", buffer);
290 : #endif
291 :
292 63177 : bufid = -(buffer + 1);
293 :
294 63177 : Assert(LocalRefCount[bufid] > 0);
295 :
296 63177 : bufHdr = GetLocalBufferDescriptor(bufid);
297 :
298 63177 : buf_state = pg_atomic_read_u32(&bufHdr->state);
299 :
300 63177 : if (!(buf_state & BM_DIRTY))
301 1267 : pgBufferUsage.local_blks_dirtied++;
302 :
303 63177 : buf_state |= BM_DIRTY;
304 :
305 63177 : pg_atomic_unlocked_write_u32(&bufHdr->state, buf_state);
306 63177 : }
307 :
308 : /*
309 : * DropRelFileNodeLocalBuffers
310 : * This function removes from the buffer pool all the pages of the
311 : * specified relation that have block numbers >= firstDelBlock.
312 : * (In particular, with firstDelBlock = 0, all pages are removed.)
313 : * Dirty pages are simply dropped, without bothering to write them
314 : * out first. Therefore, this is NOT rollback-able, and so should be
315 : * used only with extreme caution!
316 : *
317 : * See DropRelFileNodeBuffers in bufmgr.c for more notes.
318 : */
319 : void
320 18 : DropRelFileNodeLocalBuffers(RelFileNode rnode, ForkNumber forkNum,
321 : BlockNumber firstDelBlock)
322 : {
323 : int i;
324 :
325 11282 : for (i = 0; i < NLocBuffer; i++)
326 : {
327 11264 : BufferDesc *bufHdr = GetLocalBufferDescriptor(i);
328 : LocalBufferLookupEnt *hresult;
329 : uint32 buf_state;
330 :
331 11264 : buf_state = pg_atomic_read_u32(&bufHdr->state);
332 :
333 11291 : if ((buf_state & BM_TAG_VALID) &&
334 33 : RelFileNodeEquals(bufHdr->tag.rnode, rnode) &&
335 6 : bufHdr->tag.forkNum == forkNum &&
336 3 : bufHdr->tag.blockNum >= firstDelBlock)
337 : {
338 3 : if (LocalRefCount[i] != 0)
339 0 : elog(ERROR, "block %u of %s is still referenced (local %u)",
340 : bufHdr->tag.blockNum,
341 : relpathbackend(bufHdr->tag.rnode, MyBackendId,
342 : bufHdr->tag.forkNum),
343 : LocalRefCount[i]);
344 : /* Remove entry from hashtable */
345 3 : hresult = (LocalBufferLookupEnt *)
346 3 : hash_search(LocalBufHash, (void *) &bufHdr->tag,
347 : HASH_REMOVE, NULL);
348 3 : if (!hresult) /* shouldn't happen */
349 0 : elog(ERROR, "local buffer hash table corrupted");
350 : /* Mark buffer invalid */
351 3 : CLEAR_BUFFERTAG(bufHdr->tag);
352 3 : buf_state &= ~BUF_FLAG_MASK;
353 3 : buf_state &= ~BUF_USAGECOUNT_MASK;
354 3 : pg_atomic_unlocked_write_u32(&bufHdr->state, buf_state);
355 : }
356 : }
357 18 : }
358 :
359 : /*
360 : * DropRelFileNodeAllLocalBuffers
361 : * This function removes from the buffer pool all pages of all forks
362 : * of the specified relation.
363 : *
364 : * See DropRelFileNodeAllBuffers in bufmgr.c for more notes.
365 : */
366 : void
367 605 : DropRelFileNodeAllLocalBuffers(RelFileNode rnode)
368 : {
369 : int i;
370 :
371 590429 : for (i = 0; i < NLocBuffer; i++)
372 : {
373 589824 : BufferDesc *bufHdr = GetLocalBufferDescriptor(i);
374 : LocalBufferLookupEnt *hresult;
375 : uint32 buf_state;
376 :
377 589824 : buf_state = pg_atomic_read_u32(&bufHdr->state);
378 :
379 617672 : if ((buf_state & BM_TAG_VALID) &&
380 29190 : RelFileNodeEquals(bufHdr->tag.rnode, rnode))
381 : {
382 1342 : if (LocalRefCount[i] != 0)
383 0 : elog(ERROR, "block %u of %s is still referenced (local %u)",
384 : bufHdr->tag.blockNum,
385 : relpathbackend(bufHdr->tag.rnode, MyBackendId,
386 : bufHdr->tag.forkNum),
387 : LocalRefCount[i]);
388 : /* Remove entry from hashtable */
389 1342 : hresult = (LocalBufferLookupEnt *)
390 1342 : hash_search(LocalBufHash, (void *) &bufHdr->tag,
391 : HASH_REMOVE, NULL);
392 1342 : if (!hresult) /* shouldn't happen */
393 0 : elog(ERROR, "local buffer hash table corrupted");
394 : /* Mark buffer invalid */
395 1342 : CLEAR_BUFFERTAG(bufHdr->tag);
396 1342 : buf_state &= ~BUF_FLAG_MASK;
397 1342 : buf_state &= ~BUF_USAGECOUNT_MASK;
398 1342 : pg_atomic_unlocked_write_u32(&bufHdr->state, buf_state);
399 : }
400 : }
401 605 : }
402 :
403 : /*
404 : * InitLocalBuffers -
405 : * init the local buffer cache. Since most queries (esp. multi-user ones)
406 : * don't involve local buffers, we delay allocating actual memory for the
407 : * buffers until we need them; just make the buffer headers here.
408 : */
409 : static void
410 47 : InitLocalBuffers(void)
411 : {
412 47 : int nbufs = num_temp_buffers;
413 : HASHCTL info;
414 : int i;
415 :
416 : /*
417 : * Parallel workers can't access data in temporary tables, because they
418 : * have no visibility into the local buffers of their leader. This is a
419 : * convenient, low-cost place to provide a backstop check for that. Note
420 : * that we don't wish to prevent a parallel worker from accessing catalog
421 : * metadata about a temp table, so checks at higher levels would be
422 : * inappropriate.
423 : */
424 47 : if (IsParallelWorker())
425 0 : ereport(ERROR,
426 : (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
427 : errmsg("cannot access temporary tables during a parallel operation")));
428 :
429 : /* Allocate and zero buffer headers and auxiliary arrays */
430 47 : LocalBufferDescriptors = (BufferDesc *) calloc(nbufs, sizeof(BufferDesc));
431 47 : LocalBufferBlockPointers = (Block *) calloc(nbufs, sizeof(Block));
432 47 : LocalRefCount = (int32 *) calloc(nbufs, sizeof(int32));
433 47 : if (!LocalBufferDescriptors || !LocalBufferBlockPointers || !LocalRefCount)
434 0 : ereport(FATAL,
435 : (errcode(ERRCODE_OUT_OF_MEMORY),
436 : errmsg("out of memory")));
437 :
438 47 : nextFreeLocalBuf = 0;
439 :
440 : /* initialize fields that need to start off nonzero */
441 48175 : for (i = 0; i < nbufs; i++)
442 : {
443 48128 : BufferDesc *buf = GetLocalBufferDescriptor(i);
444 :
445 : /*
446 : * negative to indicate local buffer. This is tricky: shared buffers
447 : * start with 0. We have to start with -2. (Note that the routine
448 : * BufferDescriptorGetBuffer adds 1 to buf_id so our first buffer id
449 : * is -1.)
450 : */
451 48128 : buf->buf_id = -i - 2;
452 :
453 : /*
454 : * Intentionally do not initialize the buffer's atomic variable
455 : * (besides zeroing the underlying memory above). That way we get
456 : * errors on platforms without atomics, if somebody (re-)introduces
457 : * atomic operations for local buffers.
458 : */
459 : }
460 :
461 : /* Create the lookup hash table */
462 47 : MemSet(&info, 0, sizeof(info));
463 47 : info.keysize = sizeof(BufferTag);
464 47 : info.entrysize = sizeof(LocalBufferLookupEnt);
465 :
466 47 : LocalBufHash = hash_create("Local Buffer Lookup Table",
467 : nbufs,
468 : &info,
469 : HASH_ELEM | HASH_BLOBS);
470 :
471 47 : if (!LocalBufHash)
472 0 : elog(ERROR, "could not initialize local buffer hash table");
473 :
474 : /* Initialization done, mark buffers allocated */
475 47 : NLocBuffer = nbufs;
476 47 : }
477 :
478 : /*
479 : * GetLocalBufferStorage - allocate memory for a local buffer
480 : *
481 : * The idea of this function is to aggregate our requests for storage
482 : * so that the memory manager doesn't see a whole lot of relatively small
483 : * requests. Since we'll never give back a local buffer once it's created
484 : * within a particular process, no point in burdening memmgr with separately
485 : * managed chunks.
486 : */
487 : static Block
488 1345 : GetLocalBufferStorage(void)
489 : {
490 : static char *cur_block = NULL;
491 : static int next_buf_in_block = 0;
492 : static int num_bufs_in_block = 0;
493 : static int total_bufs_allocated = 0;
494 : static MemoryContext LocalBufferContext = NULL;
495 :
496 : char *this_buf;
497 :
498 1345 : Assert(total_bufs_allocated < NLocBuffer);
499 :
500 1345 : if (next_buf_in_block >= num_bufs_in_block)
501 : {
502 : /* Need to make a new request to memmgr */
503 : int num_bufs;
504 :
505 : /*
506 : * We allocate local buffers in a context of their own, so that the
507 : * space eaten for them is easily recognizable in MemoryContextStats
508 : * output. Create the context on first use.
509 : */
510 66 : if (LocalBufferContext == NULL)
511 47 : LocalBufferContext =
512 47 : AllocSetContextCreate(TopMemoryContext,
513 : "LocalBufferContext",
514 : ALLOCSET_DEFAULT_SIZES);
515 :
516 : /* Start with a 16-buffer request; subsequent ones double each time */
517 66 : num_bufs = Max(num_bufs_in_block * 2, 16);
518 : /* But not more than what we need for all remaining local bufs */
519 66 : num_bufs = Min(num_bufs, NLocBuffer - total_bufs_allocated);
520 : /* And don't overflow MaxAllocSize, either */
521 66 : num_bufs = Min(num_bufs, MaxAllocSize / BLCKSZ);
522 :
523 66 : cur_block = (char *) MemoryContextAlloc(LocalBufferContext,
524 66 : num_bufs * BLCKSZ);
525 66 : next_buf_in_block = 0;
526 66 : num_bufs_in_block = num_bufs;
527 : }
528 :
529 : /* Allocate next buffer in current memory block */
530 1345 : this_buf = cur_block + next_buf_in_block * BLCKSZ;
531 1345 : next_buf_in_block++;
532 1345 : total_bufs_allocated++;
533 :
534 1345 : return (Block) this_buf;
535 : }
536 :
537 : /*
538 : * CheckForLocalBufferLeaks - ensure this backend holds no local buffer pins
539 : *
540 : * This is just like CheckBufferLeaks(), but for local buffers.
541 : */
542 : static void
543 26560 : CheckForLocalBufferLeaks(void)
544 : {
545 : #ifdef USE_ASSERT_CHECKING
546 26560 : if (LocalRefCount)
547 : {
548 8495 : int RefCountErrors = 0;
549 : int i;
550 :
551 8707375 : for (i = 0; i < NLocBuffer; i++)
552 : {
553 8698880 : if (LocalRefCount[i] != 0)
554 : {
555 0 : Buffer b = -i - 1;
556 :
557 0 : PrintBufferLeakWarning(b);
558 0 : RefCountErrors++;
559 : }
560 : }
561 8495 : Assert(RefCountErrors == 0);
562 : }
563 : #endif
564 26560 : }
565 :
566 : /*
567 : * AtEOXact_LocalBuffers - clean up at end of transaction.
568 : *
569 : * This is just like AtEOXact_Buffers, but for local buffers.
570 : */
571 : void
572 26218 : AtEOXact_LocalBuffers(bool isCommit)
573 : {
574 26218 : CheckForLocalBufferLeaks();
575 26218 : }
576 :
577 : /*
578 : * AtProcExit_LocalBuffers - ensure we have dropped pins during backend exit.
579 : *
580 : * This is just like AtProcExit_Buffers, but for local buffers.
581 : */
582 : void
583 342 : AtProcExit_LocalBuffers(void)
584 : {
585 : /*
586 : * We shouldn't be holding any remaining pins; if we are, and assertions
587 : * aren't enabled, we'll fail later in DropRelFileNodeBuffers while trying
588 : * to drop the temp rels.
589 : */
590 342 : CheckForLocalBufferLeaks();
591 342 : }
|
